1 /*
2 * Copyright (c) 2005, 2016, 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 "libadt/vectset.hpp"
28 #include "opto/addnode.hpp"
29 #include "opto/arraycopynode.hpp"
30 #include "opto/callnode.hpp"
31 #include "opto/castnode.hpp"
32 #include "opto/cfgnode.hpp"
33 #include "opto/compile.hpp"
34 #include "opto/convertnode.hpp"
35 #include "opto/graphKit.hpp"
36 #include "opto/locknode.hpp"
37 #include "opto/loopnode.hpp"
38 #include "opto/macro.hpp"
39 #include "opto/memnode.hpp"
40 #include "opto/narrowptrnode.hpp"
41 #include "opto/node.hpp"
42 #include "opto/opaquenode.hpp"
43 #include "opto/phaseX.hpp"
44 #include "opto/rootnode.hpp"
45 #include "opto/runtime.hpp"
46 #include "opto/subnode.hpp"
47 #include "opto/type.hpp"
48 #include "runtime/sharedRuntime.hpp"
49
50
51 //
52 // Replace any references to "oldref" in inputs to "use" with "newref".
53 // Returns the number of replacements made.
54 //
55 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) {
56 int nreplacements = 0;
57 uint req = use->req();
58 for (uint j = 0; j < use->len(); j++) {
59 Node *uin = use->in(j);
60 if (uin == oldref) {
61 if (j < req)
62 use->set_req(j, newref);
63 else
64 use->set_prec(j, newref);
65 nreplacements++;
66 } else if (j >= req && uin == NULL) {
67 break;
68 }
69 }
70 return nreplacements;
71 }
72
73 void PhaseMacroExpand::copy_call_debug_info(CallNode *oldcall, CallNode * newcall) {
74 // Copy debug information and adjust JVMState information
75 uint old_dbg_start = oldcall->tf()->domain()->cnt();
76 uint new_dbg_start = newcall->tf()->domain()->cnt();
77 int jvms_adj = new_dbg_start - old_dbg_start;
78 assert (new_dbg_start == newcall->req(), "argument count mismatch");
79
80 // SafePointScalarObject node could be referenced several times in debug info.
81 // Use Dict to record cloned nodes.
82 Dict* sosn_map = new Dict(cmpkey,hashkey);
83 for (uint i = old_dbg_start; i < oldcall->req(); i++) {
84 Node* old_in = oldcall->in(i);
85 // Clone old SafePointScalarObjectNodes, adjusting their field contents.
86 if (old_in != NULL && old_in->is_SafePointScalarObject()) {
87 SafePointScalarObjectNode* old_sosn = old_in->as_SafePointScalarObject();
88 uint old_unique = C->unique();
89 Node* new_in = old_sosn->clone(sosn_map);
90 if (old_unique != C->unique()) { // New node?
91 new_in->set_req(0, C->root()); // reset control edge
92 new_in = transform_later(new_in); // Register new node.
93 }
94 old_in = new_in;
95 }
96 newcall->add_req(old_in);
97 }
98
99 // JVMS may be shared so clone it before we modify it
100 newcall->set_jvms(oldcall->jvms() != NULL ? oldcall->jvms()->clone_deep(C) : NULL);
101 for (JVMState *jvms = newcall->jvms(); jvms != NULL; jvms = jvms->caller()) {
102 jvms->set_map(newcall);
103 jvms->set_locoff(jvms->locoff()+jvms_adj);
104 jvms->set_stkoff(jvms->stkoff()+jvms_adj);
105 jvms->set_monoff(jvms->monoff()+jvms_adj);
106 jvms->set_scloff(jvms->scloff()+jvms_adj);
107 jvms->set_endoff(jvms->endoff()+jvms_adj);
108 }
109 }
110
111 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word, int mask, int bits, bool return_fast_path) {
112 Node* cmp;
113 if (mask != 0) {
114 Node* and_node = transform_later(new AndXNode(word, MakeConX(mask)));
115 cmp = transform_later(new CmpXNode(and_node, MakeConX(bits)));
116 } else {
117 cmp = word;
118 }
119 Node* bol = transform_later(new BoolNode(cmp, BoolTest::ne));
120 IfNode* iff = new IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN );
121 transform_later(iff);
122
123 // Fast path taken.
124 Node *fast_taken = transform_later(new IfFalseNode(iff));
125
126 // Fast path not-taken, i.e. slow path
127 Node *slow_taken = transform_later(new IfTrueNode(iff));
128
129 if (return_fast_path) {
130 region->init_req(edge, slow_taken); // Capture slow-control
131 return fast_taken;
132 } else {
133 region->init_req(edge, fast_taken); // Capture fast-control
134 return slow_taken;
135 }
136 }
137
138 //--------------------copy_predefined_input_for_runtime_call--------------------
139 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) {
140 // Set fixed predefined input arguments
141 call->init_req( TypeFunc::Control, ctrl );
142 call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) );
143 call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ?????
144 call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) );
145 call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) );
146 }
147
148 //------------------------------make_slow_call---------------------------------
149 CallNode* PhaseMacroExpand::make_slow_call(CallNode *oldcall, const TypeFunc* slow_call_type,
150 address slow_call, const char* leaf_name, Node* slow_path,
151 Node* parm0, Node* parm1, Node* parm2) {
152
153 // Slow-path call
154 CallNode *call = leaf_name
155 ? (CallNode*)new CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM )
156 : (CallNode*)new CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), oldcall->jvms()->bci(), TypeRawPtr::BOTTOM );
157
158 // Slow path call has no side-effects, uses few values
159 copy_predefined_input_for_runtime_call(slow_path, oldcall, call );
160 if (parm0 != NULL) call->init_req(TypeFunc::Parms+0, parm0);
161 if (parm1 != NULL) call->init_req(TypeFunc::Parms+1, parm1);
162 if (parm2 != NULL) call->init_req(TypeFunc::Parms+2, parm2);
163 copy_call_debug_info(oldcall, call);
164 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
165 _igvn.replace_node(oldcall, call);
166 transform_later(call);
167
168 return call;
169 }
170
171 void PhaseMacroExpand::extract_call_projections(CallNode *call) {
172 _fallthroughproj = NULL;
173 _fallthroughcatchproj = NULL;
174 _ioproj_fallthrough = NULL;
175 _ioproj_catchall = NULL;
176 _catchallcatchproj = NULL;
177 _memproj_fallthrough = NULL;
178 _memproj_catchall = NULL;
179 _resproj = NULL;
180 for (DUIterator_Fast imax, i = call->fast_outs(imax); i < imax; i++) {
181 ProjNode *pn = call->fast_out(i)->as_Proj();
182 switch (pn->_con) {
183 case TypeFunc::Control:
184 {
185 // For Control (fallthrough) and I_O (catch_all_index) we have CatchProj -> Catch -> Proj
186 _fallthroughproj = pn;
187 DUIterator_Fast jmax, j = pn->fast_outs(jmax);
188 const Node *cn = pn->fast_out(j);
189 if (cn->is_Catch()) {
190 ProjNode *cpn = NULL;
191 for (DUIterator_Fast kmax, k = cn->fast_outs(kmax); k < kmax; k++) {
192 cpn = cn->fast_out(k)->as_Proj();
193 assert(cpn->is_CatchProj(), "must be a CatchProjNode");
194 if (cpn->_con == CatchProjNode::fall_through_index)
195 _fallthroughcatchproj = cpn;
196 else {
197 assert(cpn->_con == CatchProjNode::catch_all_index, "must be correct index.");
198 _catchallcatchproj = cpn;
199 }
200 }
201 }
202 break;
203 }
204 case TypeFunc::I_O:
205 if (pn->_is_io_use)
206 _ioproj_catchall = pn;
207 else
208 _ioproj_fallthrough = pn;
209 break;
210 case TypeFunc::Memory:
211 if (pn->_is_io_use)
212 _memproj_catchall = pn;
213 else
214 _memproj_fallthrough = pn;
215 break;
216 case TypeFunc::Parms:
217 _resproj = pn;
218 break;
219 default:
220 assert(false, "unexpected projection from allocation node.");
221 }
222 }
223
224 }
225
226 // Eliminate a card mark sequence. p2x is a ConvP2XNode
227 void PhaseMacroExpand::eliminate_gc_barrier(Node* p2x) {
228 C2BarrierSetCodeGen *code_gen = Universe::heap()->barrier_set()->c2_code_gen();
229 code_gen->eliminate_gc_barrier(this, p2x);
230 }
231
232 // Search for a memory operation for the specified memory slice.
233 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) {
234 Node *orig_mem = mem;
235 Node *alloc_mem = alloc->in(TypeFunc::Memory);
236 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr();
237 while (true) {
238 if (mem == alloc_mem || mem == start_mem ) {
239 return mem; // hit one of our sentinels
240 } else if (mem->is_MergeMem()) {
241 mem = mem->as_MergeMem()->memory_at(alias_idx);
242 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) {
243 Node *in = mem->in(0);
244 // we can safely skip over safepoints, calls, locks and membars because we
245 // already know that the object is safe to eliminate.
246 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) {
247 return in;
248 } else if (in->is_Call()) {
249 CallNode *call = in->as_Call();
250 if (call->may_modify(tinst, phase)) {
251 assert(call->is_ArrayCopy(), "ArrayCopy is the only call node that doesn't make allocation escape");
252 if (call->as_ArrayCopy()->modifies(offset, offset, phase, false)) {
253 return in;
254 }
255 }
256 mem = in->in(TypeFunc::Memory);
257 } else if (in->is_MemBar()) {
258 ArrayCopyNode* ac = NULL;
259 if (ArrayCopyNode::may_modify(tinst, in->as_MemBar(), phase, ac)) {
260 assert(ac != NULL && ac->is_clonebasic(), "Only basic clone is a non escaping clone");
261 return ac;
262 }
263 mem = in->in(TypeFunc::Memory);
264 } else {
265 assert(false, "unexpected projection");
266 }
267 } else if (mem->is_Store()) {
268 const TypePtr* atype = mem->as_Store()->adr_type();
269 int adr_idx = phase->C->get_alias_index(atype);
270 if (adr_idx == alias_idx) {
271 assert(atype->isa_oopptr(), "address type must be oopptr");
272 int adr_offset = atype->offset();
273 uint adr_iid = atype->is_oopptr()->instance_id();
274 // Array elements references have the same alias_idx
275 // but different offset and different instance_id.
276 if (adr_offset == offset && adr_iid == alloc->_idx)
277 return mem;
278 } else {
279 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw");
280 }
281 mem = mem->in(MemNode::Memory);
282 } else if (mem->is_ClearArray()) {
283 if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) {
284 // Can not bypass initialization of the instance
285 // we are looking.
286 debug_only(intptr_t offset;)
287 assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity");
288 InitializeNode* init = alloc->as_Allocate()->initialization();
289 // We are looking for stored value, return Initialize node
290 // or memory edge from Allocate node.
291 if (init != NULL)
292 return init;
293 else
294 return alloc->in(TypeFunc::Memory); // It will produce zero value (see callers).
295 }
296 // Otherwise skip it (the call updated 'mem' value).
297 } else if (mem->Opcode() == Op_SCMemProj) {
298 mem = mem->in(0);
299 Node* adr = NULL;
300 if (mem->is_LoadStore()) {
301 adr = mem->in(MemNode::Address);
302 } else {
303 assert(mem->Opcode() == Op_EncodeISOArray ||
304 mem->Opcode() == Op_StrCompressedCopy, "sanity");
305 adr = mem->in(3); // Destination array
306 }
307 const TypePtr* atype = adr->bottom_type()->is_ptr();
308 int adr_idx = phase->C->get_alias_index(atype);
309 if (adr_idx == alias_idx) {
310 DEBUG_ONLY(mem->dump();)
311 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
312 return NULL;
313 }
314 mem = mem->in(MemNode::Memory);
315 } else if (mem->Opcode() == Op_StrInflatedCopy) {
316 Node* adr = mem->in(3); // Destination array
317 const TypePtr* atype = adr->bottom_type()->is_ptr();
318 int adr_idx = phase->C->get_alias_index(atype);
319 if (adr_idx == alias_idx) {
320 DEBUG_ONLY(mem->dump();)
321 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field");
322 return NULL;
323 }
324 mem = mem->in(MemNode::Memory);
325 } else {
326 return mem;
327 }
328 assert(mem != orig_mem, "dead memory loop");
329 }
330 }
331
332 // Generate loads from source of the arraycopy for fields of
333 // destination needed at a deoptimization point
334 Node* PhaseMacroExpand::make_arraycopy_load(ArrayCopyNode* ac, intptr_t offset, Node* ctl, Node* mem, BasicType ft, const Type *ftype, AllocateNode *alloc) {
335 BasicType bt = ft;
336 const Type *type = ftype;
337 if (ft == T_NARROWOOP) {
338 bt = T_OBJECT;
339 type = ftype->make_oopptr();
340 }
341 Node* res = NULL;
342 if (ac->is_clonebasic()) {
343 Node* base = ac->in(ArrayCopyNode::Src)->in(AddPNode::Base);
344 Node* adr = _igvn.transform(new AddPNode(base, base, MakeConX(offset)));
345 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset);
346 res = LoadNode::make(_igvn, ctl, mem, adr, adr_type, type, bt, MemNode::unordered, LoadNode::Pinned);
347 } else {
348 if (ac->modifies(offset, offset, &_igvn, true)) {
349 assert(ac->in(ArrayCopyNode::Dest) == alloc->result_cast(), "arraycopy destination should be allocation's result");
350 uint shift = exact_log2(type2aelembytes(bt));
351 Node* diff = _igvn.transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
352 #ifdef _LP64
353 diff = _igvn.transform(new ConvI2LNode(diff));
354 #endif
355 diff = _igvn.transform(new LShiftXNode(diff, intcon(shift)));
356
357 Node* off = _igvn.transform(new AddXNode(MakeConX(offset), diff));
358 Node* base = ac->in(ArrayCopyNode::Src);
359 Node* adr = _igvn.transform(new AddPNode(base, base, off));
360 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset);
361 res = LoadNode::make(_igvn, ctl, mem, adr, adr_type, type, bt, MemNode::unordered, LoadNode::Pinned);
362 }
363 }
364 if (res != NULL) {
365 res = _igvn.transform(res);
366 if (ftype->isa_narrowoop()) {
367 // PhaseMacroExpand::scalar_replacement adds DecodeN nodes
368 res = _igvn.transform(new EncodePNode(res, ftype));
369 }
370 return res;
371 }
372 return NULL;
373 }
374
375 //
376 // Given a Memory Phi, compute a value Phi containing the values from stores
377 // on the input paths.
378 // Note: this function is recursive, its depth is limited by the "level" argument
379 // Returns the computed Phi, or NULL if it cannot compute it.
380 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) {
381 assert(mem->is_Phi(), "sanity");
382 int alias_idx = C->get_alias_index(adr_t);
383 int offset = adr_t->offset();
384 int instance_id = adr_t->instance_id();
385
386 // Check if an appropriate value phi already exists.
387 Node* region = mem->in(0);
388 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) {
389 Node* phi = region->fast_out(k);
390 if (phi->is_Phi() && phi != mem &&
391 phi->as_Phi()->is_same_inst_field(phi_type, (int)mem->_idx, instance_id, alias_idx, offset)) {
392 return phi;
393 }
394 }
395 // Check if an appropriate new value phi already exists.
396 Node* new_phi = value_phis->find(mem->_idx);
397 if (new_phi != NULL)
398 return new_phi;
399
400 if (level <= 0) {
401 return NULL; // Give up: phi tree too deep
402 }
403 Node *start_mem = C->start()->proj_out(TypeFunc::Memory);
404 Node *alloc_mem = alloc->in(TypeFunc::Memory);
405
406 uint length = mem->req();
407 GrowableArray <Node *> values(length, length, NULL, false);
408
409 // create a new Phi for the value
410 PhiNode *phi = new PhiNode(mem->in(0), phi_type, NULL, mem->_idx, instance_id, alias_idx, offset);
411 transform_later(phi);
412 value_phis->push(phi, mem->_idx);
413
414 for (uint j = 1; j < length; j++) {
415 Node *in = mem->in(j);
416 if (in == NULL || in->is_top()) {
417 values.at_put(j, in);
418 } else {
419 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn);
420 if (val == start_mem || val == alloc_mem) {
421 // hit a sentinel, return appropriate 0 value
422 values.at_put(j, _igvn.zerocon(ft));
423 continue;
424 }
425 if (val->is_Initialize()) {
426 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
427 }
428 if (val == NULL) {
429 return NULL; // can't find a value on this path
430 }
431 if (val == mem) {
432 values.at_put(j, mem);
433 } else if (val->is_Store()) {
434 values.at_put(j, val->in(MemNode::ValueIn));
435 } else if(val->is_Proj() && val->in(0) == alloc) {
436 values.at_put(j, _igvn.zerocon(ft));
437 } else if (val->is_Phi()) {
438 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1);
439 if (val == NULL) {
440 return NULL;
441 }
442 values.at_put(j, val);
443 } else if (val->Opcode() == Op_SCMemProj) {
444 assert(val->in(0)->is_LoadStore() ||
445 val->in(0)->Opcode() == Op_EncodeISOArray ||
446 val->in(0)->Opcode() == Op_StrCompressedCopy, "sanity");
447 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
448 return NULL;
449 } else if (val->is_ArrayCopy()) {
450 Node* res = make_arraycopy_load(val->as_ArrayCopy(), offset, val->in(0), val->in(TypeFunc::Memory), ft, phi_type, alloc);
451 if (res == NULL) {
452 return NULL;
453 }
454 values.at_put(j, res);
455 } else {
456 #ifdef ASSERT
457 val->dump();
458 assert(false, "unknown node on this path");
459 #endif
460 return NULL; // unknown node on this path
461 }
462 }
463 }
464 // Set Phi's inputs
465 for (uint j = 1; j < length; j++) {
466 if (values.at(j) == mem) {
467 phi->init_req(j, phi);
468 } else {
469 phi->init_req(j, values.at(j));
470 }
471 }
472 return phi;
473 }
474
475 // Search the last value stored into the object's field.
476 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, Node *sfpt_ctl, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, AllocateNode *alloc) {
477 assert(adr_t->is_known_instance_field(), "instance required");
478 int instance_id = adr_t->instance_id();
479 assert((uint)instance_id == alloc->_idx, "wrong allocation");
480
481 int alias_idx = C->get_alias_index(adr_t);
482 int offset = adr_t->offset();
483 Node *start_mem = C->start()->proj_out(TypeFunc::Memory);
484 Node *alloc_ctrl = alloc->in(TypeFunc::Control);
485 Node *alloc_mem = alloc->in(TypeFunc::Memory);
486 Arena *a = Thread::current()->resource_area();
487 VectorSet visited(a);
488
489
490 bool done = sfpt_mem == alloc_mem;
491 Node *mem = sfpt_mem;
492 while (!done) {
493 if (visited.test_set(mem->_idx)) {
494 return NULL; // found a loop, give up
495 }
496 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn);
497 if (mem == start_mem || mem == alloc_mem) {
498 done = true; // hit a sentinel, return appropriate 0 value
499 } else if (mem->is_Initialize()) {
500 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
501 if (mem == NULL) {
502 done = true; // Something go wrong.
503 } else if (mem->is_Store()) {
504 const TypePtr* atype = mem->as_Store()->adr_type();
505 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice");
506 done = true;
507 }
508 } else if (mem->is_Store()) {
509 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr();
510 assert(atype != NULL, "address type must be oopptr");
511 assert(C->get_alias_index(atype) == alias_idx &&
512 atype->is_known_instance_field() && atype->offset() == offset &&
513 atype->instance_id() == instance_id, "store is correct memory slice");
514 done = true;
515 } else if (mem->is_Phi()) {
516 // try to find a phi's unique input
517 Node *unique_input = NULL;
518 Node *top = C->top();
519 for (uint i = 1; i < mem->req(); i++) {
520 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn);
521 if (n == NULL || n == top || n == mem) {
522 continue;
523 } else if (unique_input == NULL) {
524 unique_input = n;
525 } else if (unique_input != n) {
526 unique_input = top;
527 break;
528 }
529 }
530 if (unique_input != NULL && unique_input != top) {
531 mem = unique_input;
532 } else {
533 done = true;
534 }
535 } else if (mem->is_ArrayCopy()) {
536 done = true;
537 } else {
538 assert(false, "unexpected node");
539 }
540 }
541 if (mem != NULL) {
542 if (mem == start_mem || mem == alloc_mem) {
543 // hit a sentinel, return appropriate 0 value
544 return _igvn.zerocon(ft);
545 } else if (mem->is_Store()) {
546 return mem->in(MemNode::ValueIn);
547 } else if (mem->is_Phi()) {
548 // attempt to produce a Phi reflecting the values on the input paths of the Phi
549 Node_Stack value_phis(a, 8);
550 Node * phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit);
551 if (phi != NULL) {
552 return phi;
553 } else {
554 // Kill all new Phis
555 while(value_phis.is_nonempty()) {
556 Node* n = value_phis.node();
557 _igvn.replace_node(n, C->top());
558 value_phis.pop();
559 }
560 }
561 } else if (mem->is_ArrayCopy()) {
562 Node* ctl = mem->in(0);
563 Node* m = mem->in(TypeFunc::Memory);
564 if (sfpt_ctl->is_Proj() && sfpt_ctl->as_Proj()->is_uncommon_trap_proj(Deoptimization::Reason_none)) {
565 // pin the loads in the uncommon trap path
566 ctl = sfpt_ctl;
567 m = sfpt_mem;
568 }
569 return make_arraycopy_load(mem->as_ArrayCopy(), offset, ctl, m, ft, ftype, alloc);
570 }
571 }
572 // Something go wrong.
573 return NULL;
574 }
575
576 // Check the possibility of scalar replacement.
577 bool PhaseMacroExpand::can_eliminate_allocation(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
578 // Scan the uses of the allocation to check for anything that would
579 // prevent us from eliminating it.
580 NOT_PRODUCT( const char* fail_eliminate = NULL; )
581 DEBUG_ONLY( Node* disq_node = NULL; )
582 bool can_eliminate = true;
583
584 Node* res = alloc->result_cast();
585 const TypeOopPtr* res_type = NULL;
586 if (res == NULL) {
587 // All users were eliminated.
588 } else if (!res->is_CheckCastPP()) {
589 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";)
590 can_eliminate = false;
591 } else {
592 res_type = _igvn.type(res)->isa_oopptr();
593 if (res_type == NULL) {
594 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";)
595 can_eliminate = false;
596 } else if (res_type->isa_aryptr()) {
597 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1);
598 if (length < 0) {
599 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";)
600 can_eliminate = false;
601 }
602 }
603 }
604
605 if (can_eliminate && res != NULL) {
606 for (DUIterator_Fast jmax, j = res->fast_outs(jmax);
607 j < jmax && can_eliminate; j++) {
608 Node* use = res->fast_out(j);
609
610 if (use->is_AddP()) {
611 const TypePtr* addp_type = _igvn.type(use)->is_ptr();
612 int offset = addp_type->offset();
613
614 if (offset == Type::OffsetTop || offset == Type::OffsetBot) {
615 NOT_PRODUCT(fail_eliminate = "Undefined field referrence";)
616 can_eliminate = false;
617 break;
618 }
619 for (DUIterator_Fast kmax, k = use->fast_outs(kmax);
620 k < kmax && can_eliminate; k++) {
621 Node* n = use->fast_out(k);
622 if (!n->is_Store() && n->Opcode() != Op_CastP2X &&
623 !(n->is_ArrayCopy() &&
624 n->as_ArrayCopy()->is_clonebasic() &&
625 n->in(ArrayCopyNode::Dest) == use)) {
626 DEBUG_ONLY(disq_node = n;)
627 if (n->is_Load() || n->is_LoadStore()) {
628 NOT_PRODUCT(fail_eliminate = "Field load";)
629 } else {
630 NOT_PRODUCT(fail_eliminate = "Not store field referrence";)
631 }
632 can_eliminate = false;
633 }
634 }
635 } else if (use->is_ArrayCopy() &&
636 (use->as_ArrayCopy()->is_arraycopy_validated() ||
637 use->as_ArrayCopy()->is_copyof_validated() ||
638 use->as_ArrayCopy()->is_copyofrange_validated()) &&
639 use->in(ArrayCopyNode::Dest) == res) {
640 // ok to eliminate
641 } else if (use->is_SafePoint()) {
642 SafePointNode* sfpt = use->as_SafePoint();
643 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) {
644 // Object is passed as argument.
645 DEBUG_ONLY(disq_node = use;)
646 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";)
647 can_eliminate = false;
648 }
649 Node* sfptMem = sfpt->memory();
650 if (sfptMem == NULL || sfptMem->is_top()) {
651 DEBUG_ONLY(disq_node = use;)
652 NOT_PRODUCT(fail_eliminate = "NULL or TOP memory";)
653 can_eliminate = false;
654 } else {
655 safepoints.append_if_missing(sfpt);
656 }
657 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
658 if (use->is_Phi()) {
659 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
660 NOT_PRODUCT(fail_eliminate = "Object is return value";)
661 } else {
662 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
663 }
664 DEBUG_ONLY(disq_node = use;)
665 } else {
666 if (use->Opcode() == Op_Return) {
667 NOT_PRODUCT(fail_eliminate = "Object is return value";)
668 }else {
669 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
670 }
671 DEBUG_ONLY(disq_node = use;)
672 }
673 can_eliminate = false;
674 }
675 }
676 }
677
678 #ifndef PRODUCT
679 if (PrintEliminateAllocations) {
680 if (can_eliminate) {
681 tty->print("Scalar ");
682 if (res == NULL)
683 alloc->dump();
684 else
685 res->dump();
686 } else if (alloc->_is_scalar_replaceable) {
687 tty->print("NotScalar (%s)", fail_eliminate);
688 if (res == NULL)
689 alloc->dump();
690 else
691 res->dump();
692 #ifdef ASSERT
693 if (disq_node != NULL) {
694 tty->print(" >>>> ");
695 disq_node->dump();
696 }
697 #endif /*ASSERT*/
698 }
699 }
700 #endif
701 return can_eliminate;
702 }
703
704 // Do scalar replacement.
705 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
706 GrowableArray <SafePointNode *> safepoints_done;
707
708 ciKlass* klass = NULL;
709 ciInstanceKlass* iklass = NULL;
710 int nfields = 0;
711 int array_base = 0;
712 int element_size = 0;
713 BasicType basic_elem_type = T_ILLEGAL;
714 ciType* elem_type = NULL;
715
716 Node* res = alloc->result_cast();
717 assert(res == NULL || res->is_CheckCastPP(), "unexpected AllocateNode result");
718 const TypeOopPtr* res_type = NULL;
719 if (res != NULL) { // Could be NULL when there are no users
720 res_type = _igvn.type(res)->isa_oopptr();
721 }
722
723 if (res != NULL) {
724 klass = res_type->klass();
725 if (res_type->isa_instptr()) {
726 // find the fields of the class which will be needed for safepoint debug information
727 assert(klass->is_instance_klass(), "must be an instance klass.");
728 iklass = klass->as_instance_klass();
729 nfields = iklass->nof_nonstatic_fields();
730 } else {
731 // find the array's elements which will be needed for safepoint debug information
732 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
733 assert(klass->is_array_klass() && nfields >= 0, "must be an array klass.");
734 elem_type = klass->as_array_klass()->element_type();
735 basic_elem_type = elem_type->basic_type();
736 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
737 element_size = type2aelembytes(basic_elem_type);
738 }
739 }
740 //
741 // Process the safepoint uses
742 //
743 while (safepoints.length() > 0) {
744 SafePointNode* sfpt = safepoints.pop();
745 Node* mem = sfpt->memory();
746 Node* ctl = sfpt->control();
747 assert(sfpt->jvms() != NULL, "missed JVMS");
748 // Fields of scalar objs are referenced only at the end
749 // of regular debuginfo at the last (youngest) JVMS.
750 // Record relative start index.
751 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff());
752 SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type,
753 #ifdef ASSERT
754 alloc,
755 #endif
756 first_ind, nfields);
757 sobj->init_req(0, C->root());
758 transform_later(sobj);
759
760 // Scan object's fields adding an input to the safepoint for each field.
761 for (int j = 0; j < nfields; j++) {
762 intptr_t offset;
763 ciField* field = NULL;
764 if (iklass != NULL) {
765 field = iklass->nonstatic_field_at(j);
766 offset = field->offset();
767 elem_type = field->type();
768 basic_elem_type = field->layout_type();
769 } else {
770 offset = array_base + j * (intptr_t)element_size;
771 }
772
773 const Type *field_type;
774 // The next code is taken from Parse::do_get_xxx().
775 if (basic_elem_type == T_OBJECT || basic_elem_type == T_ARRAY) {
776 if (!elem_type->is_loaded()) {
777 field_type = TypeInstPtr::BOTTOM;
778 } else if (field != NULL && field->is_static_constant()) {
779 // This can happen if the constant oop is non-perm.
780 ciObject* con = field->constant_value().as_object();
781 // Do not "join" in the previous type; it doesn't add value,
782 // and may yield a vacuous result if the field is of interface type.
783 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr();
784 assert(field_type != NULL, "field singleton type must be consistent");
785 } else {
786 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
787 }
788 if (UseCompressedOops) {
789 field_type = field_type->make_narrowoop();
790 basic_elem_type = T_NARROWOOP;
791 }
792 } else {
793 field_type = Type::get_const_basic_type(basic_elem_type);
794 }
795
796 const TypeOopPtr *field_addr_type = res_type->add_offset(offset)->isa_oopptr();
797
798 Node *field_val = value_from_mem(mem, ctl, basic_elem_type, field_type, field_addr_type, alloc);
799 if (field_val == NULL) {
800 // We weren't able to find a value for this field,
801 // give up on eliminating this allocation.
802
803 // Remove any extra entries we added to the safepoint.
804 uint last = sfpt->req() - 1;
805 for (int k = 0; k < j; k++) {
806 sfpt->del_req(last--);
807 }
808 _igvn._worklist.push(sfpt);
809 // rollback processed safepoints
810 while (safepoints_done.length() > 0) {
811 SafePointNode* sfpt_done = safepoints_done.pop();
812 // remove any extra entries we added to the safepoint
813 last = sfpt_done->req() - 1;
814 for (int k = 0; k < nfields; k++) {
815 sfpt_done->del_req(last--);
816 }
817 JVMState *jvms = sfpt_done->jvms();
818 jvms->set_endoff(sfpt_done->req());
819 // Now make a pass over the debug information replacing any references
820 // to SafePointScalarObjectNode with the allocated object.
821 int start = jvms->debug_start();
822 int end = jvms->debug_end();
823 for (int i = start; i < end; i++) {
824 if (sfpt_done->in(i)->is_SafePointScalarObject()) {
825 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject();
826 if (scobj->first_index(jvms) == sfpt_done->req() &&
827 scobj->n_fields() == (uint)nfields) {
828 assert(scobj->alloc() == alloc, "sanity");
829 sfpt_done->set_req(i, res);
830 }
831 }
832 }
833 _igvn._worklist.push(sfpt_done);
834 }
835 #ifndef PRODUCT
836 if (PrintEliminateAllocations) {
837 if (field != NULL) {
838 tty->print("=== At SafePoint node %d can't find value of Field: ",
839 sfpt->_idx);
840 field->print();
841 int field_idx = C->get_alias_index(field_addr_type);
842 tty->print(" (alias_idx=%d)", field_idx);
843 } else { // Array's element
844 tty->print("=== At SafePoint node %d can't find value of array element [%d]",
845 sfpt->_idx, j);
846 }
847 tty->print(", which prevents elimination of: ");
848 if (res == NULL)
849 alloc->dump();
850 else
851 res->dump();
852 }
853 #endif
854 return false;
855 }
856 if (UseCompressedOops && field_type->isa_narrowoop()) {
857 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
858 // to be able scalar replace the allocation.
859 if (field_val->is_EncodeP()) {
860 field_val = field_val->in(1);
861 } else {
862 field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type()));
863 }
864 }
865 sfpt->add_req(field_val);
866 }
867 JVMState *jvms = sfpt->jvms();
868 jvms->set_endoff(sfpt->req());
869 // Now make a pass over the debug information replacing any references
870 // to the allocated object with "sobj"
871 int start = jvms->debug_start();
872 int end = jvms->debug_end();
873 sfpt->replace_edges_in_range(res, sobj, start, end);
874 _igvn._worklist.push(sfpt);
875 safepoints_done.append_if_missing(sfpt); // keep it for rollback
876 }
877 return true;
878 }
879
880 static void disconnect_projections(MultiNode* n, PhaseIterGVN& igvn) {
881 Node* ctl_proj = n->proj_out(TypeFunc::Control);
882 Node* mem_proj = n->proj_out(TypeFunc::Memory);
883 if (ctl_proj != NULL) {
884 igvn.replace_node(ctl_proj, n->in(0));
885 }
886 if (mem_proj != NULL) {
887 igvn.replace_node(mem_proj, n->in(TypeFunc::Memory));
888 }
889 }
890
891 // Process users of eliminated allocation.
892 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc) {
893 Node* res = alloc->result_cast();
894 if (res != NULL) {
895 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
896 Node *use = res->last_out(j);
897 uint oc1 = res->outcnt();
898
899 if (use->is_AddP()) {
900 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) {
901 Node *n = use->last_out(k);
902 uint oc2 = use->outcnt();
903 if (n->is_Store()) {
904 #ifdef ASSERT
905 // Verify that there is no dependent MemBarVolatile nodes,
906 // they should be removed during IGVN, see MemBarNode::Ideal().
907 for (DUIterator_Fast pmax, p = n->fast_outs(pmax);
908 p < pmax; p++) {
909 Node* mb = n->fast_out(p);
910 assert(mb->is_Initialize() || !mb->is_MemBar() ||
911 mb->req() <= MemBarNode::Precedent ||
912 mb->in(MemBarNode::Precedent) != n,
913 "MemBarVolatile should be eliminated for non-escaping object");
914 }
915 #endif
916 _igvn.replace_node(n, n->in(MemNode::Memory));
917 } else if (n->is_ArrayCopy()) {
918 // Disconnect ArrayCopy node
919 ArrayCopyNode* ac = n->as_ArrayCopy();
920 assert(ac->is_clonebasic(), "unexpected array copy kind");
921 Node* membar_after = ac->proj_out(TypeFunc::Control)->unique_ctrl_out();
922 disconnect_projections(ac, _igvn);
923 assert(alloc->in(0)->is_Proj() && alloc->in(0)->in(0)->Opcode() == Op_MemBarCPUOrder, "mem barrier expected before allocation");
924 Node* membar_before = alloc->in(0)->in(0);
925 disconnect_projections(membar_before->as_MemBar(), _igvn);
926 if (membar_after->is_MemBar()) {
927 disconnect_projections(membar_after->as_MemBar(), _igvn);
928 }
929 } else {
930 eliminate_gc_barrier(n);
931 }
932 k -= (oc2 - use->outcnt());
933 }
934 } else if (use->is_ArrayCopy()) {
935 // Disconnect ArrayCopy node
936 ArrayCopyNode* ac = use->as_ArrayCopy();
937 assert(ac->is_arraycopy_validated() ||
938 ac->is_copyof_validated() ||
939 ac->is_copyofrange_validated(), "unsupported");
940 CallProjections callprojs;
941 ac->extract_projections(&callprojs, true);
942
943 _igvn.replace_node(callprojs.fallthrough_ioproj, ac->in(TypeFunc::I_O));
944 _igvn.replace_node(callprojs.fallthrough_memproj, ac->in(TypeFunc::Memory));
945 _igvn.replace_node(callprojs.fallthrough_catchproj, ac->in(TypeFunc::Control));
946
947 // Set control to top. IGVN will remove the remaining projections
948 ac->set_req(0, top());
949 ac->replace_edge(res, top());
950
951 // Disconnect src right away: it can help find new
952 // opportunities for allocation elimination
953 Node* src = ac->in(ArrayCopyNode::Src);
954 ac->replace_edge(src, top());
955 if (src->outcnt() == 0) {
956 _igvn.remove_dead_node(src);
957 }
958
959 _igvn._worklist.push(ac);
960 } else {
961 eliminate_gc_barrier(use);
962 }
963 j -= (oc1 - res->outcnt());
964 }
965 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
966 _igvn.remove_dead_node(res);
967 }
968
969 //
970 // Process other users of allocation's projections
971 //
972 if (_resproj != NULL && _resproj->outcnt() != 0) {
973 // First disconnect stores captured by Initialize node.
974 // If Initialize node is eliminated first in the following code,
975 // it will kill such stores and DUIterator_Last will assert.
976 for (DUIterator_Fast jmax, j = _resproj->fast_outs(jmax); j < jmax; j++) {
977 Node *use = _resproj->fast_out(j);
978 if (use->is_AddP()) {
979 // raw memory addresses used only by the initialization
980 _igvn.replace_node(use, C->top());
981 --j; --jmax;
982 }
983 }
984 for (DUIterator_Last jmin, j = _resproj->last_outs(jmin); j >= jmin; ) {
985 Node *use = _resproj->last_out(j);
986 uint oc1 = _resproj->outcnt();
987 if (use->is_Initialize()) {
988 // Eliminate Initialize node.
989 InitializeNode *init = use->as_Initialize();
990 assert(init->outcnt() <= 2, "only a control and memory projection expected");
991 Node *ctrl_proj = init->proj_out(TypeFunc::Control);
992 if (ctrl_proj != NULL) {
993 assert(init->in(TypeFunc::Control) == _fallthroughcatchproj, "allocation control projection");
994 _igvn.replace_node(ctrl_proj, _fallthroughcatchproj);
995 }
996 Node *mem_proj = init->proj_out(TypeFunc::Memory);
997 if (mem_proj != NULL) {
998 Node *mem = init->in(TypeFunc::Memory);
999 #ifdef ASSERT
1000 if (mem->is_MergeMem()) {
1001 assert(mem->in(TypeFunc::Memory) == _memproj_fallthrough, "allocation memory projection");
1002 } else {
1003 assert(mem == _memproj_fallthrough, "allocation memory projection");
1004 }
1005 #endif
1006 _igvn.replace_node(mem_proj, mem);
1007 }
1008 } else {
1009 assert(false, "only Initialize or AddP expected");
1010 }
1011 j -= (oc1 - _resproj->outcnt());
1012 }
1013 }
1014 if (_fallthroughcatchproj != NULL) {
1015 _igvn.replace_node(_fallthroughcatchproj, alloc->in(TypeFunc::Control));
1016 }
1017 if (_memproj_fallthrough != NULL) {
1018 _igvn.replace_node(_memproj_fallthrough, alloc->in(TypeFunc::Memory));
1019 }
1020 if (_memproj_catchall != NULL) {
1021 _igvn.replace_node(_memproj_catchall, C->top());
1022 }
1023 if (_ioproj_fallthrough != NULL) {
1024 _igvn.replace_node(_ioproj_fallthrough, alloc->in(TypeFunc::I_O));
1025 }
1026 if (_ioproj_catchall != NULL) {
1027 _igvn.replace_node(_ioproj_catchall, C->top());
1028 }
1029 if (_catchallcatchproj != NULL) {
1030 _igvn.replace_node(_catchallcatchproj, C->top());
1031 }
1032 }
1033
1034 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) {
1035 // Don't do scalar replacement if the frame can be popped by JVMTI:
1036 // if reallocation fails during deoptimization we'll pop all
1037 // interpreter frames for this compiled frame and that won't play
1038 // nice with JVMTI popframe.
1039 if (!EliminateAllocations || JvmtiExport::can_pop_frame() || !alloc->_is_non_escaping) {
1040 return false;
1041 }
1042 Node* klass = alloc->in(AllocateNode::KlassNode);
1043 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr();
1044 Node* res = alloc->result_cast();
1045 // Eliminate boxing allocations which are not used
1046 // regardless scalar replacable status.
1047 bool boxing_alloc = C->eliminate_boxing() &&
1048 tklass->klass()->is_instance_klass() &&
1049 tklass->klass()->as_instance_klass()->is_box_klass();
1050 if (!alloc->_is_scalar_replaceable && (!boxing_alloc || (res != NULL))) {
1051 return false;
1052 }
1053
1054 extract_call_projections(alloc);
1055
1056 GrowableArray <SafePointNode *> safepoints;
1057 if (!can_eliminate_allocation(alloc, safepoints)) {
1058 return false;
1059 }
1060
1061 if (!alloc->_is_scalar_replaceable) {
1062 assert(res == NULL, "sanity");
1063 // We can only eliminate allocation if all debug info references
1064 // are already replaced with SafePointScalarObject because
1065 // we can't search for a fields value without instance_id.
1066 if (safepoints.length() > 0) {
1067 return false;
1068 }
1069 }
1070
1071 if (!scalar_replacement(alloc, safepoints)) {
1072 return false;
1073 }
1074
1075 CompileLog* log = C->log();
1076 if (log != NULL) {
1077 log->head("eliminate_allocation type='%d'",
1078 log->identify(tklass->klass()));
1079 JVMState* p = alloc->jvms();
1080 while (p != NULL) {
1081 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1082 p = p->caller();
1083 }
1084 log->tail("eliminate_allocation");
1085 }
1086
1087 process_users_of_allocation(alloc);
1088
1089 #ifndef PRODUCT
1090 if (PrintEliminateAllocations) {
1091 if (alloc->is_AllocateArray())
1092 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1093 else
1094 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1095 }
1096 #endif
1097
1098 return true;
1099 }
1100
1101 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) {
1102 // EA should remove all uses of non-escaping boxing node.
1103 if (!C->eliminate_boxing() || boxing->proj_out(TypeFunc::Parms) != NULL) {
1104 return false;
1105 }
1106
1107 assert(boxing->result_cast() == NULL, "unexpected boxing node result");
1108
1109 extract_call_projections(boxing);
1110
1111 const TypeTuple* r = boxing->tf()->range();
1112 assert(r->cnt() > TypeFunc::Parms, "sanity");
1113 const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr();
1114 assert(t != NULL, "sanity");
1115
1116 CompileLog* log = C->log();
1117 if (log != NULL) {
1118 log->head("eliminate_boxing type='%d'",
1119 log->identify(t->klass()));
1120 JVMState* p = boxing->jvms();
1121 while (p != NULL) {
1122 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1123 p = p->caller();
1124 }
1125 log->tail("eliminate_boxing");
1126 }
1127
1128 process_users_of_allocation(boxing);
1129
1130 #ifndef PRODUCT
1131 if (PrintEliminateAllocations) {
1132 tty->print("++++ Eliminated: %d ", boxing->_idx);
1133 boxing->method()->print_short_name(tty);
1134 tty->cr();
1135 }
1136 #endif
1137
1138 return true;
1139 }
1140
1141 //---------------------------set_eden_pointers-------------------------
1142 void PhaseMacroExpand::set_eden_pointers(Node* &eden_top_adr, Node* &eden_end_adr) {
1143 if (UseTLAB) { // Private allocation: load from TLS
1144 Node* thread = transform_later(new ThreadLocalNode());
1145 int tlab_top_offset = in_bytes(JavaThread::tlab_top_offset());
1146 int tlab_end_offset = in_bytes(JavaThread::tlab_end_offset());
1147 eden_top_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_top_offset);
1148 eden_end_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_end_offset);
1149 } else { // Shared allocation: load from globals
1150 CollectedHeap* ch = Universe::heap();
1151 address top_adr = (address)ch->top_addr();
1152 address end_adr = (address)ch->end_addr();
1153 eden_top_adr = makecon(TypeRawPtr::make(top_adr));
1154 eden_end_adr = basic_plus_adr(eden_top_adr, end_adr - top_adr);
1155 }
1156 }
1157
1158
1159 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) {
1160 Node* adr = basic_plus_adr(base, offset);
1161 const TypePtr* adr_type = adr->bottom_type()->is_ptr();
1162 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt, MemNode::unordered);
1163 transform_later(value);
1164 return value;
1165 }
1166
1167
1168 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) {
1169 Node* adr = basic_plus_adr(base, offset);
1170 mem = StoreNode::make(_igvn, ctl, mem, adr, NULL, value, bt, MemNode::unordered);
1171 transform_later(mem);
1172 return mem;
1173 }
1174
1175 //=============================================================================
1176 //
1177 // A L L O C A T I O N
1178 //
1179 // Allocation attempts to be fast in the case of frequent small objects.
1180 // It breaks down like this:
1181 //
1182 // 1) Size in doublewords is computed. This is a constant for objects and
1183 // variable for most arrays. Doubleword units are used to avoid size
1184 // overflow of huge doubleword arrays. We need doublewords in the end for
1185 // rounding.
1186 //
1187 // 2) Size is checked for being 'too large'. Too-large allocations will go
1188 // the slow path into the VM. The slow path can throw any required
1189 // exceptions, and does all the special checks for very large arrays. The
1190 // size test can constant-fold away for objects. For objects with
1191 // finalizers it constant-folds the otherway: you always go slow with
1192 // finalizers.
1193 //
1194 // 3) If NOT using TLABs, this is the contended loop-back point.
1195 // Load-Locked the heap top. If using TLABs normal-load the heap top.
1196 //
1197 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route.
1198 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish
1199 // "size*8" we always enter the VM, where "largish" is a constant picked small
1200 // enough that there's always space between the eden max and 4Gig (old space is
1201 // there so it's quite large) and large enough that the cost of entering the VM
1202 // is dwarfed by the cost to initialize the space.
1203 //
1204 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back
1205 // down. If contended, repeat at step 3. If using TLABs normal-store
1206 // adjusted heap top back down; there is no contention.
1207 //
1208 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark
1209 // fields.
1210 //
1211 // 7) Merge with the slow-path; cast the raw memory pointer to the correct
1212 // oop flavor.
1213 //
1214 //=============================================================================
1215 // FastAllocateSizeLimit value is in DOUBLEWORDS.
1216 // Allocations bigger than this always go the slow route.
1217 // This value must be small enough that allocation attempts that need to
1218 // trigger exceptions go the slow route. Also, it must be small enough so
1219 // that heap_top + size_in_bytes does not wrap around the 4Gig limit.
1220 //=============================================================================j//
1221 // %%% Here is an old comment from parseHelper.cpp; is it outdated?
1222 // The allocator will coalesce int->oop copies away. See comment in
1223 // coalesce.cpp about how this works. It depends critically on the exact
1224 // code shape produced here, so if you are changing this code shape
1225 // make sure the GC info for the heap-top is correct in and around the
1226 // slow-path call.
1227 //
1228
1229 void PhaseMacroExpand::expand_allocate_common(
1230 AllocateNode* alloc, // allocation node to be expanded
1231 Node* length, // array length for an array allocation
1232 const TypeFunc* slow_call_type, // Type of slow call
1233 address slow_call_address // Address of slow call
1234 )
1235 {
1236
1237 Node* ctrl = alloc->in(TypeFunc::Control);
1238 Node* mem = alloc->in(TypeFunc::Memory);
1239 Node* i_o = alloc->in(TypeFunc::I_O);
1240 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize);
1241 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1242 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest);
1243
1244 assert(ctrl != NULL, "must have control");
1245 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results.
1246 // they will not be used if "always_slow" is set
1247 enum { slow_result_path = 1, fast_result_path = 2 };
1248 Node *result_region = NULL;
1249 Node *result_phi_rawmem = NULL;
1250 Node *result_phi_rawoop = NULL;
1251 Node *result_phi_i_o = NULL;
1252
1253 // The initial slow comparison is a size check, the comparison
1254 // we want to do is a BoolTest::gt
1255 bool always_slow = false;
1256 int tv = _igvn.find_int_con(initial_slow_test, -1);
1257 if (tv >= 0) {
1258 always_slow = (tv == 1);
1259 initial_slow_test = NULL;
1260 } else {
1261 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn);
1262 }
1263
1264 if (C->env()->dtrace_alloc_probes() ||
1265 !UseTLAB && (!Universe::heap()->supports_inline_contig_alloc())) {
1266 // Force slow-path allocation
1267 always_slow = true;
1268 initial_slow_test = NULL;
1269 }
1270
1271
1272 enum { too_big_or_final_path = 1, need_gc_path = 2 };
1273 Node *slow_region = NULL;
1274 Node *toobig_false = ctrl;
1275
1276 assert (initial_slow_test == NULL || !always_slow, "arguments must be consistent");
1277 // generate the initial test if necessary
1278 if (initial_slow_test != NULL ) {
1279 slow_region = new RegionNode(3);
1280
1281 // Now make the initial failure test. Usually a too-big test but
1282 // might be a TRUE for finalizers or a fancy class check for
1283 // newInstance0.
1284 IfNode *toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
1285 transform_later(toobig_iff);
1286 // Plug the failing-too-big test into the slow-path region
1287 Node *toobig_true = new IfTrueNode( toobig_iff );
1288 transform_later(toobig_true);
1289 slow_region ->init_req( too_big_or_final_path, toobig_true );
1290 toobig_false = new IfFalseNode( toobig_iff );
1291 transform_later(toobig_false);
1292 } else { // No initial test, just fall into next case
1293 toobig_false = ctrl;
1294 debug_only(slow_region = NodeSentinel);
1295 }
1296
1297 Node *slow_mem = mem; // save the current memory state for slow path
1298 // generate the fast allocation code unless we know that the initial test will always go slow
1299 if (!always_slow) {
1300 // Fast path modifies only raw memory.
1301 if (mem->is_MergeMem()) {
1302 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1303 }
1304
1305 Node* eden_top_adr;
1306 Node* eden_end_adr;
1307
1308 set_eden_pointers(eden_top_adr, eden_end_adr);
1309
1310 // Load Eden::end. Loop invariant and hoisted.
1311 //
1312 // Note: We set the control input on "eden_end" and "old_eden_top" when using
1313 // a TLAB to work around a bug where these values were being moved across
1314 // a safepoint. These are not oops, so they cannot be include in the oop
1315 // map, but they can be changed by a GC. The proper way to fix this would
1316 // be to set the raw memory state when generating a SafepointNode. However
1317 // this will require extensive changes to the loop optimization in order to
1318 // prevent a degradation of the optimization.
1319 // See comment in memnode.hpp, around line 227 in class LoadPNode.
1320 Node *eden_end = make_load(ctrl, mem, eden_end_adr, 0, TypeRawPtr::BOTTOM, T_ADDRESS);
1321
1322 // allocate the Region and Phi nodes for the result
1323 result_region = new RegionNode(3);
1324 result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1325 result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM);
1326 result_phi_i_o = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch
1327
1328 // We need a Region for the loop-back contended case.
1329 enum { fall_in_path = 1, contended_loopback_path = 2 };
1330 Node *contended_region;
1331 Node *contended_phi_rawmem;
1332 if (UseTLAB) {
1333 contended_region = toobig_false;
1334 contended_phi_rawmem = mem;
1335 } else {
1336 contended_region = new RegionNode(3);
1337 contended_phi_rawmem = new PhiNode(contended_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1338 // Now handle the passing-too-big test. We fall into the contended
1339 // loop-back merge point.
1340 contended_region ->init_req(fall_in_path, toobig_false);
1341 contended_phi_rawmem->init_req(fall_in_path, mem);
1342 transform_later(contended_region);
1343 transform_later(contended_phi_rawmem);
1344 }
1345
1346 // Load(-locked) the heap top.
1347 // See note above concerning the control input when using a TLAB
1348 Node *old_eden_top = UseTLAB
1349 ? new LoadPNode (ctrl, contended_phi_rawmem, eden_top_adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, MemNode::unordered)
1350 : new LoadPLockedNode(contended_region, contended_phi_rawmem, eden_top_adr, MemNode::acquire);
1351
1352 transform_later(old_eden_top);
1353 // Add to heap top to get a new heap top
1354 Node *new_eden_top = new AddPNode(top(), old_eden_top, size_in_bytes);
1355 transform_later(new_eden_top);
1356 // Check for needing a GC; compare against heap end
1357 Node *needgc_cmp = new CmpPNode(new_eden_top, eden_end);
1358 transform_later(needgc_cmp);
1359 Node *needgc_bol = new BoolNode(needgc_cmp, BoolTest::ge);
1360 transform_later(needgc_bol);
1361 IfNode *needgc_iff = new IfNode(contended_region, needgc_bol, PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN);
1362 transform_later(needgc_iff);
1363
1364 // Plug the failing-heap-space-need-gc test into the slow-path region
1365 Node *needgc_true = new IfTrueNode(needgc_iff);
1366 transform_later(needgc_true);
1367 if (initial_slow_test) {
1368 slow_region->init_req(need_gc_path, needgc_true);
1369 // This completes all paths into the slow merge point
1370 transform_later(slow_region);
1371 } else { // No initial slow path needed!
1372 // Just fall from the need-GC path straight into the VM call.
1373 slow_region = needgc_true;
1374 }
1375 // No need for a GC. Setup for the Store-Conditional
1376 Node *needgc_false = new IfFalseNode(needgc_iff);
1377 transform_later(needgc_false);
1378
1379 // Grab regular I/O before optional prefetch may change it.
1380 // Slow-path does no I/O so just set it to the original I/O.
1381 result_phi_i_o->init_req(slow_result_path, i_o);
1382
1383 i_o = prefetch_allocation(i_o, needgc_false, contended_phi_rawmem,
1384 old_eden_top, new_eden_top, length);
1385
1386 // Name successful fast-path variables
1387 Node* fast_oop = old_eden_top;
1388 Node* fast_oop_ctrl;
1389 Node* fast_oop_rawmem;
1390
1391 // Store (-conditional) the modified eden top back down.
1392 // StorePConditional produces flags for a test PLUS a modified raw
1393 // memory state.
1394 if (UseTLAB) {
1395 Node* store_eden_top =
1396 new StorePNode(needgc_false, contended_phi_rawmem, eden_top_adr,
1397 TypeRawPtr::BOTTOM, new_eden_top, MemNode::unordered);
1398 transform_later(store_eden_top);
1399 fast_oop_ctrl = needgc_false; // No contention, so this is the fast path
1400 fast_oop_rawmem = store_eden_top;
1401 } else {
1402 Node* store_eden_top =
1403 new StorePConditionalNode(needgc_false, contended_phi_rawmem, eden_top_adr,
1404 new_eden_top, fast_oop/*old_eden_top*/);
1405 transform_later(store_eden_top);
1406 Node *contention_check = new BoolNode(store_eden_top, BoolTest::ne);
1407 transform_later(contention_check);
1408 store_eden_top = new SCMemProjNode(store_eden_top);
1409 transform_later(store_eden_top);
1410
1411 // If not using TLABs, check to see if there was contention.
1412 IfNode *contention_iff = new IfNode (needgc_false, contention_check, PROB_MIN, COUNT_UNKNOWN);
1413 transform_later(contention_iff);
1414 Node *contention_true = new IfTrueNode(contention_iff);
1415 transform_later(contention_true);
1416 // If contention, loopback and try again.
1417 contended_region->init_req(contended_loopback_path, contention_true);
1418 contended_phi_rawmem->init_req(contended_loopback_path, store_eden_top);
1419
1420 // Fast-path succeeded with no contention!
1421 Node *contention_false = new IfFalseNode(contention_iff);
1422 transform_later(contention_false);
1423 fast_oop_ctrl = contention_false;
1424
1425 // Bump total allocated bytes for this thread
1426 Node* thread = new ThreadLocalNode();
1427 transform_later(thread);
1428 Node* alloc_bytes_adr = basic_plus_adr(top()/*not oop*/, thread,
1429 in_bytes(JavaThread::allocated_bytes_offset()));
1430 Node* alloc_bytes = make_load(fast_oop_ctrl, store_eden_top, alloc_bytes_adr,
1431 0, TypeLong::LONG, T_LONG);
1432 #ifdef _LP64
1433 Node* alloc_size = size_in_bytes;
1434 #else
1435 Node* alloc_size = new ConvI2LNode(size_in_bytes);
1436 transform_later(alloc_size);
1437 #endif
1438 Node* new_alloc_bytes = new AddLNode(alloc_bytes, alloc_size);
1439 transform_later(new_alloc_bytes);
1440 fast_oop_rawmem = make_store(fast_oop_ctrl, store_eden_top, alloc_bytes_adr,
1441 0, new_alloc_bytes, T_LONG);
1442 }
1443
1444 InitializeNode* init = alloc->initialization();
1445 fast_oop_rawmem = initialize_object(alloc,
1446 fast_oop_ctrl, fast_oop_rawmem, fast_oop,
1447 klass_node, length, size_in_bytes);
1448
1449 // If initialization is performed by an array copy, any required
1450 // MemBarStoreStore was already added. If the object does not
1451 // escape no need for a MemBarStoreStore. If the object does not
1452 // escape in its initializer and memory barrier (MemBarStoreStore or
1453 // stronger) is already added at exit of initializer, also no need
1454 // for a MemBarStoreStore. Otherwise we need a MemBarStoreStore
1455 // so that stores that initialize this object can't be reordered
1456 // with a subsequent store that makes this object accessible by
1457 // other threads.
1458 // Other threads include java threads and JVM internal threads
1459 // (for example concurrent GC threads). Current concurrent GC
1460 // implementation: CMS and G1 will not scan newly created object,
1461 // so it's safe to skip storestore barrier when allocation does
1462 // not escape.
1463 if (!alloc->does_not_escape_thread() &&
1464 !alloc->is_allocation_MemBar_redundant() &&
1465 (init == NULL || !init->is_complete_with_arraycopy())) {
1466 if (init == NULL || init->req() < InitializeNode::RawStores) {
1467 // No InitializeNode or no stores captured by zeroing
1468 // elimination. Simply add the MemBarStoreStore after object
1469 // initialization.
1470 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1471 transform_later(mb);
1472
1473 mb->init_req(TypeFunc::Memory, fast_oop_rawmem);
1474 mb->init_req(TypeFunc::Control, fast_oop_ctrl);
1475 fast_oop_ctrl = new ProjNode(mb,TypeFunc::Control);
1476 transform_later(fast_oop_ctrl);
1477 fast_oop_rawmem = new ProjNode(mb,TypeFunc::Memory);
1478 transform_later(fast_oop_rawmem);
1479 } else {
1480 // Add the MemBarStoreStore after the InitializeNode so that
1481 // all stores performing the initialization that were moved
1482 // before the InitializeNode happen before the storestore
1483 // barrier.
1484
1485 Node* init_ctrl = init->proj_out(TypeFunc::Control);
1486 Node* init_mem = init->proj_out(TypeFunc::Memory);
1487
1488 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1489 transform_later(mb);
1490
1491 Node* ctrl = new ProjNode(init,TypeFunc::Control);
1492 transform_later(ctrl);
1493 Node* mem = new ProjNode(init,TypeFunc::Memory);
1494 transform_later(mem);
1495
1496 // The MemBarStoreStore depends on control and memory coming
1497 // from the InitializeNode
1498 mb->init_req(TypeFunc::Memory, mem);
1499 mb->init_req(TypeFunc::Control, ctrl);
1500
1501 ctrl = new ProjNode(mb,TypeFunc::Control);
1502 transform_later(ctrl);
1503 mem = new ProjNode(mb,TypeFunc::Memory);
1504 transform_later(mem);
1505
1506 // All nodes that depended on the InitializeNode for control
1507 // and memory must now depend on the MemBarNode that itself
1508 // depends on the InitializeNode
1509 if (init_ctrl != NULL) {
1510 _igvn.replace_node(init_ctrl, ctrl);
1511 }
1512 if (init_mem != NULL) {
1513 _igvn.replace_node(init_mem, mem);
1514 }
1515 }
1516 }
1517
1518 if (C->env()->dtrace_extended_probes()) {
1519 // Slow-path call
1520 int size = TypeFunc::Parms + 2;
1521 CallLeafNode *call = new CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(),
1522 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_object_alloc_base),
1523 "dtrace_object_alloc",
1524 TypeRawPtr::BOTTOM);
1525
1526 // Get base of thread-local storage area
1527 Node* thread = new ThreadLocalNode();
1528 transform_later(thread);
1529
1530 call->init_req(TypeFunc::Parms+0, thread);
1531 call->init_req(TypeFunc::Parms+1, fast_oop);
1532 call->init_req(TypeFunc::Control, fast_oop_ctrl);
1533 call->init_req(TypeFunc::I_O , top()); // does no i/o
1534 call->init_req(TypeFunc::Memory , fast_oop_rawmem);
1535 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1536 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1537 transform_later(call);
1538 fast_oop_ctrl = new ProjNode(call,TypeFunc::Control);
1539 transform_later(fast_oop_ctrl);
1540 fast_oop_rawmem = new ProjNode(call,TypeFunc::Memory);
1541 transform_later(fast_oop_rawmem);
1542 }
1543
1544 // Plug in the successful fast-path into the result merge point
1545 result_region ->init_req(fast_result_path, fast_oop_ctrl);
1546 result_phi_rawoop->init_req(fast_result_path, fast_oop);
1547 result_phi_i_o ->init_req(fast_result_path, i_o);
1548 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem);
1549 } else {
1550 slow_region = ctrl;
1551 result_phi_i_o = i_o; // Rename it to use in the following code.
1552 }
1553
1554 // Generate slow-path call
1555 CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address,
1556 OptoRuntime::stub_name(slow_call_address),
1557 alloc->jvms()->bci(),
1558 TypePtr::BOTTOM);
1559 call->init_req( TypeFunc::Control, slow_region );
1560 call->init_req( TypeFunc::I_O , top() ) ; // does no i/o
1561 call->init_req( TypeFunc::Memory , slow_mem ); // may gc ptrs
1562 call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) );
1563 call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) );
1564
1565 call->init_req(TypeFunc::Parms+0, klass_node);
1566 if (length != NULL) {
1567 call->init_req(TypeFunc::Parms+1, length);
1568 }
1569
1570 // Copy debug information and adjust JVMState information, then replace
1571 // allocate node with the call
1572 copy_call_debug_info((CallNode *) alloc, call);
1573 if (!always_slow) {
1574 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1575 } else {
1576 // Hook i_o projection to avoid its elimination during allocation
1577 // replacement (when only a slow call is generated).
1578 call->set_req(TypeFunc::I_O, result_phi_i_o);
1579 }
1580 _igvn.replace_node(alloc, call);
1581 transform_later(call);
1582
1583 // Identify the output projections from the allocate node and
1584 // adjust any references to them.
1585 // The control and io projections look like:
1586 //
1587 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl)
1588 // Allocate Catch
1589 // ^---Proj(io) <-------+ ^---CatchProj(io)
1590 //
1591 // We are interested in the CatchProj nodes.
1592 //
1593 extract_call_projections(call);
1594
1595 // An allocate node has separate memory projections for the uses on
1596 // the control and i_o paths. Replace the control memory projection with
1597 // result_phi_rawmem (unless we are only generating a slow call when
1598 // both memory projections are combined)
1599 if (!always_slow && _memproj_fallthrough != NULL) {
1600 for (DUIterator_Fast imax, i = _memproj_fallthrough->fast_outs(imax); i < imax; i++) {
1601 Node *use = _memproj_fallthrough->fast_out(i);
1602 _igvn.rehash_node_delayed(use);
1603 imax -= replace_input(use, _memproj_fallthrough, result_phi_rawmem);
1604 // back up iterator
1605 --i;
1606 }
1607 }
1608 // Now change uses of _memproj_catchall to use _memproj_fallthrough and delete
1609 // _memproj_catchall so we end up with a call that has only 1 memory projection.
1610 if (_memproj_catchall != NULL ) {
1611 if (_memproj_fallthrough == NULL) {
1612 _memproj_fallthrough = new ProjNode(call, TypeFunc::Memory);
1613 transform_later(_memproj_fallthrough);
1614 }
1615 for (DUIterator_Fast imax, i = _memproj_catchall->fast_outs(imax); i < imax; i++) {
1616 Node *use = _memproj_catchall->fast_out(i);
1617 _igvn.rehash_node_delayed(use);
1618 imax -= replace_input(use, _memproj_catchall, _memproj_fallthrough);
1619 // back up iterator
1620 --i;
1621 }
1622 assert(_memproj_catchall->outcnt() == 0, "all uses must be deleted");
1623 _igvn.remove_dead_node(_memproj_catchall);
1624 }
1625
1626 // An allocate node has separate i_o projections for the uses on the control
1627 // and i_o paths. Always replace the control i_o projection with result i_o
1628 // otherwise incoming i_o become dead when only a slow call is generated
1629 // (it is different from memory projections where both projections are
1630 // combined in such case).
1631 if (_ioproj_fallthrough != NULL) {
1632 for (DUIterator_Fast imax, i = _ioproj_fallthrough->fast_outs(imax); i < imax; i++) {
1633 Node *use = _ioproj_fallthrough->fast_out(i);
1634 _igvn.rehash_node_delayed(use);
1635 imax -= replace_input(use, _ioproj_fallthrough, result_phi_i_o);
1636 // back up iterator
1637 --i;
1638 }
1639 }
1640 // Now change uses of _ioproj_catchall to use _ioproj_fallthrough and delete
1641 // _ioproj_catchall so we end up with a call that has only 1 i_o projection.
1642 if (_ioproj_catchall != NULL ) {
1643 if (_ioproj_fallthrough == NULL) {
1644 _ioproj_fallthrough = new ProjNode(call, TypeFunc::I_O);
1645 transform_later(_ioproj_fallthrough);
1646 }
1647 for (DUIterator_Fast imax, i = _ioproj_catchall->fast_outs(imax); i < imax; i++) {
1648 Node *use = _ioproj_catchall->fast_out(i);
1649 _igvn.rehash_node_delayed(use);
1650 imax -= replace_input(use, _ioproj_catchall, _ioproj_fallthrough);
1651 // back up iterator
1652 --i;
1653 }
1654 assert(_ioproj_catchall->outcnt() == 0, "all uses must be deleted");
1655 _igvn.remove_dead_node(_ioproj_catchall);
1656 }
1657
1658 // if we generated only a slow call, we are done
1659 if (always_slow) {
1660 // Now we can unhook i_o.
1661 if (result_phi_i_o->outcnt() > 1) {
1662 call->set_req(TypeFunc::I_O, top());
1663 } else {
1664 assert(result_phi_i_o->unique_ctrl_out() == call, "");
1665 // Case of new array with negative size known during compilation.
1666 // AllocateArrayNode::Ideal() optimization disconnect unreachable
1667 // following code since call to runtime will throw exception.
1668 // As result there will be no users of i_o after the call.
1669 // Leave i_o attached to this call to avoid problems in preceding graph.
1670 }
1671 return;
1672 }
1673
1674
1675 if (_fallthroughcatchproj != NULL) {
1676 ctrl = _fallthroughcatchproj->clone();
1677 transform_later(ctrl);
1678 _igvn.replace_node(_fallthroughcatchproj, result_region);
1679 } else {
1680 ctrl = top();
1681 }
1682 Node *slow_result;
1683 if (_resproj == NULL) {
1684 // no uses of the allocation result
1685 slow_result = top();
1686 } else {
1687 slow_result = _resproj->clone();
1688 transform_later(slow_result);
1689 _igvn.replace_node(_resproj, result_phi_rawoop);
1690 }
1691
1692 // Plug slow-path into result merge point
1693 result_region ->init_req( slow_result_path, ctrl );
1694 result_phi_rawoop->init_req( slow_result_path, slow_result);
1695 result_phi_rawmem->init_req( slow_result_path, _memproj_fallthrough );
1696 transform_later(result_region);
1697 transform_later(result_phi_rawoop);
1698 transform_later(result_phi_rawmem);
1699 transform_later(result_phi_i_o);
1700 // This completes all paths into the result merge point
1701 }
1702
1703
1704 // Helper for PhaseMacroExpand::expand_allocate_common.
1705 // Initializes the newly-allocated storage.
1706 Node*
1707 PhaseMacroExpand::initialize_object(AllocateNode* alloc,
1708 Node* control, Node* rawmem, Node* object,
1709 Node* klass_node, Node* length,
1710 Node* size_in_bytes) {
1711 InitializeNode* init = alloc->initialization();
1712 // Store the klass & mark bits
1713 Node* mark_node = NULL;
1714 // For now only enable fast locking for non-array types
1715 if (UseBiasedLocking && (length == NULL)) {
1716 mark_node = make_load(control, rawmem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
1717 } else {
1718 mark_node = makecon(TypeRawPtr::make((address)markOopDesc::prototype()));
1719 }
1720 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS);
1721
1722 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
1723 int header_size = alloc->minimum_header_size(); // conservatively small
1724
1725 // Array length
1726 if (length != NULL) { // Arrays need length field
1727 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
1728 // conservatively small header size:
1729 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1730 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
1731 if (k->is_array_klass()) // we know the exact header size in most cases:
1732 header_size = Klass::layout_helper_header_size(k->layout_helper());
1733 }
1734
1735 // Clear the object body, if necessary.
1736 if (init == NULL) {
1737 // The init has somehow disappeared; be cautious and clear everything.
1738 //
1739 // This can happen if a node is allocated but an uncommon trap occurs
1740 // immediately. In this case, the Initialize gets associated with the
1741 // trap, and may be placed in a different (outer) loop, if the Allocate
1742 // is in a loop. If (this is rare) the inner loop gets unrolled, then
1743 // there can be two Allocates to one Initialize. The answer in all these
1744 // edge cases is safety first. It is always safe to clear immediately
1745 // within an Allocate, and then (maybe or maybe not) clear some more later.
1746 if (!(UseTLAB && ZeroTLAB)) {
1747 rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
1748 header_size, size_in_bytes,
1749 &_igvn);
1750 }
1751 } else {
1752 if (!init->is_complete()) {
1753 // Try to win by zeroing only what the init does not store.
1754 // We can also try to do some peephole optimizations,
1755 // such as combining some adjacent subword stores.
1756 rawmem = init->complete_stores(control, rawmem, object,
1757 header_size, size_in_bytes, &_igvn);
1758 }
1759 // We have no more use for this link, since the AllocateNode goes away:
1760 init->set_req(InitializeNode::RawAddress, top());
1761 // (If we keep the link, it just confuses the register allocator,
1762 // who thinks he sees a real use of the address by the membar.)
1763 }
1764
1765 return rawmem;
1766 }
1767
1768 // Generate prefetch instructions for next allocations.
1769 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false,
1770 Node*& contended_phi_rawmem,
1771 Node* old_eden_top, Node* new_eden_top,
1772 Node* length) {
1773 enum { fall_in_path = 1, pf_path = 2 };
1774 if( UseTLAB && AllocatePrefetchStyle == 2 ) {
1775 // Generate prefetch allocation with watermark check.
1776 // As an allocation hits the watermark, we will prefetch starting
1777 // at a "distance" away from watermark.
1778
1779 Node *pf_region = new RegionNode(3);
1780 Node *pf_phi_rawmem = new PhiNode( pf_region, Type::MEMORY,
1781 TypeRawPtr::BOTTOM );
1782 // I/O is used for Prefetch
1783 Node *pf_phi_abio = new PhiNode( pf_region, Type::ABIO );
1784
1785 Node *thread = new ThreadLocalNode();
1786 transform_later(thread);
1787
1788 Node *eden_pf_adr = new AddPNode( top()/*not oop*/, thread,
1789 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) );
1790 transform_later(eden_pf_adr);
1791
1792 Node *old_pf_wm = new LoadPNode(needgc_false,
1793 contended_phi_rawmem, eden_pf_adr,
1794 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM,
1795 MemNode::unordered);
1796 transform_later(old_pf_wm);
1797
1798 // check against new_eden_top
1799 Node *need_pf_cmp = new CmpPNode( new_eden_top, old_pf_wm );
1800 transform_later(need_pf_cmp);
1801 Node *need_pf_bol = new BoolNode( need_pf_cmp, BoolTest::ge );
1802 transform_later(need_pf_bol);
1803 IfNode *need_pf_iff = new IfNode( needgc_false, need_pf_bol,
1804 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN );
1805 transform_later(need_pf_iff);
1806
1807 // true node, add prefetchdistance
1808 Node *need_pf_true = new IfTrueNode( need_pf_iff );
1809 transform_later(need_pf_true);
1810
1811 Node *need_pf_false = new IfFalseNode( need_pf_iff );
1812 transform_later(need_pf_false);
1813
1814 Node *new_pf_wmt = new AddPNode( top(), old_pf_wm,
1815 _igvn.MakeConX(AllocatePrefetchDistance) );
1816 transform_later(new_pf_wmt );
1817 new_pf_wmt->set_req(0, need_pf_true);
1818
1819 Node *store_new_wmt = new StorePNode(need_pf_true,
1820 contended_phi_rawmem, eden_pf_adr,
1821 TypeRawPtr::BOTTOM, new_pf_wmt,
1822 MemNode::unordered);
1823 transform_later(store_new_wmt);
1824
1825 // adding prefetches
1826 pf_phi_abio->init_req( fall_in_path, i_o );
1827
1828 Node *prefetch_adr;
1829 Node *prefetch;
1830 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1831 uint step_size = AllocatePrefetchStepSize;
1832 uint distance = 0;
1833
1834 for ( uint i = 0; i < lines; i++ ) {
1835 prefetch_adr = new AddPNode( old_pf_wm, new_pf_wmt,
1836 _igvn.MakeConX(distance) );
1837 transform_later(prefetch_adr);
1838 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
1839 transform_later(prefetch);
1840 distance += step_size;
1841 i_o = prefetch;
1842 }
1843 pf_phi_abio->set_req( pf_path, i_o );
1844
1845 pf_region->init_req( fall_in_path, need_pf_false );
1846 pf_region->init_req( pf_path, need_pf_true );
1847
1848 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem );
1849 pf_phi_rawmem->init_req( pf_path, store_new_wmt );
1850
1851 transform_later(pf_region);
1852 transform_later(pf_phi_rawmem);
1853 transform_later(pf_phi_abio);
1854
1855 needgc_false = pf_region;
1856 contended_phi_rawmem = pf_phi_rawmem;
1857 i_o = pf_phi_abio;
1858 } else if( UseTLAB && AllocatePrefetchStyle == 3 ) {
1859 // Insert a prefetch instruction for each allocation.
1860 // This code is used to generate 1 prefetch instruction per cache line.
1861
1862 // Generate several prefetch instructions.
1863 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1864 uint step_size = AllocatePrefetchStepSize;
1865 uint distance = AllocatePrefetchDistance;
1866
1867 // Next cache address.
1868 Node *cache_adr = new AddPNode(old_eden_top, old_eden_top,
1869 _igvn.MakeConX(step_size + distance));
1870 transform_later(cache_adr);
1871 cache_adr = new CastP2XNode(needgc_false, cache_adr);
1872 transform_later(cache_adr);
1873 // Address is aligned to execute prefetch to the beginning of cache line size
1874 // (it is important when BIS instruction is used on SPARC as prefetch).
1875 Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1));
1876 cache_adr = new AndXNode(cache_adr, mask);
1877 transform_later(cache_adr);
1878 cache_adr = new CastX2PNode(cache_adr);
1879 transform_later(cache_adr);
1880
1881 // Prefetch
1882 Node *prefetch = new PrefetchAllocationNode( contended_phi_rawmem, cache_adr );
1883 prefetch->set_req(0, needgc_false);
1884 transform_later(prefetch);
1885 contended_phi_rawmem = prefetch;
1886 Node *prefetch_adr;
1887 distance = step_size;
1888 for ( uint i = 1; i < lines; i++ ) {
1889 prefetch_adr = new AddPNode( cache_adr, cache_adr,
1890 _igvn.MakeConX(distance) );
1891 transform_later(prefetch_adr);
1892 prefetch = new PrefetchAllocationNode( contended_phi_rawmem, prefetch_adr );
1893 transform_later(prefetch);
1894 distance += step_size;
1895 contended_phi_rawmem = prefetch;
1896 }
1897 } else if( AllocatePrefetchStyle > 0 ) {
1898 // Insert a prefetch for each allocation only on the fast-path
1899 Node *prefetch_adr;
1900 Node *prefetch;
1901 // Generate several prefetch instructions.
1902 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1903 uint step_size = AllocatePrefetchStepSize;
1904 uint distance = AllocatePrefetchDistance;
1905 for ( uint i = 0; i < lines; i++ ) {
1906 prefetch_adr = new AddPNode( old_eden_top, new_eden_top,
1907 _igvn.MakeConX(distance) );
1908 transform_later(prefetch_adr);
1909 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
1910 // Do not let it float too high, since if eden_top == eden_end,
1911 // both might be null.
1912 if( i == 0 ) { // Set control for first prefetch, next follows it
1913 prefetch->init_req(0, needgc_false);
1914 }
1915 transform_later(prefetch);
1916 distance += step_size;
1917 i_o = prefetch;
1918 }
1919 }
1920 return i_o;
1921 }
1922
1923
1924 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) {
1925 expand_allocate_common(alloc, NULL,
1926 OptoRuntime::new_instance_Type(),
1927 OptoRuntime::new_instance_Java());
1928 }
1929
1930 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) {
1931 Node* length = alloc->in(AllocateNode::ALength);
1932 InitializeNode* init = alloc->initialization();
1933 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1934 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
1935 address slow_call_address; // Address of slow call
1936 if (init != NULL && init->is_complete_with_arraycopy() &&
1937 k->is_type_array_klass()) {
1938 // Don't zero type array during slow allocation in VM since
1939 // it will be initialized later by arraycopy in compiled code.
1940 slow_call_address = OptoRuntime::new_array_nozero_Java();
1941 } else {
1942 slow_call_address = OptoRuntime::new_array_Java();
1943 }
1944 expand_allocate_common(alloc, length,
1945 OptoRuntime::new_array_Type(),
1946 slow_call_address);
1947 }
1948
1949 //-------------------mark_eliminated_box----------------------------------
1950 //
1951 // During EA obj may point to several objects but after few ideal graph
1952 // transformations (CCP) it may point to only one non escaping object
1953 // (but still using phi), corresponding locks and unlocks will be marked
1954 // for elimination. Later obj could be replaced with a new node (new phi)
1955 // and which does not have escape information. And later after some graph
1956 // reshape other locks and unlocks (which were not marked for elimination
1957 // before) are connected to this new obj (phi) but they still will not be
1958 // marked for elimination since new obj has no escape information.
1959 // Mark all associated (same box and obj) lock and unlock nodes for
1960 // elimination if some of them marked already.
1961 void PhaseMacroExpand::mark_eliminated_box(Node* oldbox, Node* obj) {
1962 if (oldbox->as_BoxLock()->is_eliminated())
1963 return; // This BoxLock node was processed already.
1964
1965 // New implementation (EliminateNestedLocks) has separate BoxLock
1966 // node for each locked region so mark all associated locks/unlocks as
1967 // eliminated even if different objects are referenced in one locked region
1968 // (for example, OSR compilation of nested loop inside locked scope).
1969 if (EliminateNestedLocks ||
1970 oldbox->as_BoxLock()->is_simple_lock_region(NULL, obj)) {
1971 // Box is used only in one lock region. Mark this box as eliminated.
1972 _igvn.hash_delete(oldbox);
1973 oldbox->as_BoxLock()->set_eliminated(); // This changes box's hash value
1974 _igvn.hash_insert(oldbox);
1975
1976 for (uint i = 0; i < oldbox->outcnt(); i++) {
1977 Node* u = oldbox->raw_out(i);
1978 if (u->is_AbstractLock() && !u->as_AbstractLock()->is_non_esc_obj()) {
1979 AbstractLockNode* alock = u->as_AbstractLock();
1980 // Check lock's box since box could be referenced by Lock's debug info.
1981 if (alock->box_node() == oldbox) {
1982 // Mark eliminated all related locks and unlocks.
1983 #ifdef ASSERT
1984 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc4");
1985 #endif
1986 alock->set_non_esc_obj();
1987 }
1988 }
1989 }
1990 return;
1991 }
1992
1993 // Create new "eliminated" BoxLock node and use it in monitor debug info
1994 // instead of oldbox for the same object.
1995 BoxLockNode* newbox = oldbox->clone()->as_BoxLock();
1996
1997 // Note: BoxLock node is marked eliminated only here and it is used
1998 // to indicate that all associated lock and unlock nodes are marked
1999 // for elimination.
2000 newbox->set_eliminated();
2001 transform_later(newbox);
2002
2003 // Replace old box node with new box for all users of the same object.
2004 for (uint i = 0; i < oldbox->outcnt();) {
2005 bool next_edge = true;
2006
2007 Node* u = oldbox->raw_out(i);
2008 if (u->is_AbstractLock()) {
2009 AbstractLockNode* alock = u->as_AbstractLock();
2010 if (alock->box_node() == oldbox && alock->obj_node()->eqv_uncast(obj)) {
2011 // Replace Box and mark eliminated all related locks and unlocks.
2012 #ifdef ASSERT
2013 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc5");
2014 #endif
2015 alock->set_non_esc_obj();
2016 _igvn.rehash_node_delayed(alock);
2017 alock->set_box_node(newbox);
2018 next_edge = false;
2019 }
2020 }
2021 if (u->is_FastLock() && u->as_FastLock()->obj_node()->eqv_uncast(obj)) {
2022 FastLockNode* flock = u->as_FastLock();
2023 assert(flock->box_node() == oldbox, "sanity");
2024 _igvn.rehash_node_delayed(flock);
2025 flock->set_box_node(newbox);
2026 next_edge = false;
2027 }
2028
2029 // Replace old box in monitor debug info.
2030 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) {
2031 SafePointNode* sfn = u->as_SafePoint();
2032 JVMState* youngest_jvms = sfn->jvms();
2033 int max_depth = youngest_jvms->depth();
2034 for (int depth = 1; depth <= max_depth; depth++) {
2035 JVMState* jvms = youngest_jvms->of_depth(depth);
2036 int num_mon = jvms->nof_monitors();
2037 // Loop over monitors
2038 for (int idx = 0; idx < num_mon; idx++) {
2039 Node* obj_node = sfn->monitor_obj(jvms, idx);
2040 Node* box_node = sfn->monitor_box(jvms, idx);
2041 if (box_node == oldbox && obj_node->eqv_uncast(obj)) {
2042 int j = jvms->monitor_box_offset(idx);
2043 _igvn.replace_input_of(u, j, newbox);
2044 next_edge = false;
2045 }
2046 }
2047 }
2048 }
2049 if (next_edge) i++;
2050 }
2051 }
2052
2053 //-----------------------mark_eliminated_locking_nodes-----------------------
2054 void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) {
2055 if (EliminateNestedLocks) {
2056 if (alock->is_nested()) {
2057 assert(alock->box_node()->as_BoxLock()->is_eliminated(), "sanity");
2058 return;
2059 } else if (!alock->is_non_esc_obj()) { // Not eliminated or coarsened
2060 // Only Lock node has JVMState needed here.
2061 // Not that preceding claim is documented anywhere else.
2062 if (alock->jvms() != NULL) {
2063 if (alock->as_Lock()->is_nested_lock_region()) {
2064 // Mark eliminated related nested locks and unlocks.
2065 Node* obj = alock->obj_node();
2066 BoxLockNode* box_node = alock->box_node()->as_BoxLock();
2067 assert(!box_node->is_eliminated(), "should not be marked yet");
2068 // Note: BoxLock node is marked eliminated only here
2069 // and it is used to indicate that all associated lock
2070 // and unlock nodes are marked for elimination.
2071 box_node->set_eliminated(); // Box's hash is always NO_HASH here
2072 for (uint i = 0; i < box_node->outcnt(); i++) {
2073 Node* u = box_node->raw_out(i);
2074 if (u->is_AbstractLock()) {
2075 alock = u->as_AbstractLock();
2076 if (alock->box_node() == box_node) {
2077 // Verify that this Box is referenced only by related locks.
2078 assert(alock->obj_node()->eqv_uncast(obj), "");
2079 // Mark all related locks and unlocks.
2080 #ifdef ASSERT
2081 alock->log_lock_optimization(C, "eliminate_lock_set_nested");
2082 #endif
2083 alock->set_nested();
2084 }
2085 }
2086 }
2087 } else {
2088 #ifdef ASSERT
2089 alock->log_lock_optimization(C, "eliminate_lock_NOT_nested_lock_region");
2090 if (C->log() != NULL)
2091 alock->as_Lock()->is_nested_lock_region(C); // rerun for debugging output
2092 #endif
2093 }
2094 }
2095 return;
2096 }
2097 // Process locks for non escaping object
2098 assert(alock->is_non_esc_obj(), "");
2099 } // EliminateNestedLocks
2100
2101 if (alock->is_non_esc_obj()) { // Lock is used for non escaping object
2102 // Look for all locks of this object and mark them and
2103 // corresponding BoxLock nodes as eliminated.
2104 Node* obj = alock->obj_node();
2105 for (uint j = 0; j < obj->outcnt(); j++) {
2106 Node* o = obj->raw_out(j);
2107 if (o->is_AbstractLock() &&
2108 o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) {
2109 alock = o->as_AbstractLock();
2110 Node* box = alock->box_node();
2111 // Replace old box node with new eliminated box for all users
2112 // of the same object and mark related locks as eliminated.
2113 mark_eliminated_box(box, obj);
2114 }
2115 }
2116 }
2117 }
2118
2119 // we have determined that this lock/unlock can be eliminated, we simply
2120 // eliminate the node without expanding it.
2121 //
2122 // Note: The membar's associated with the lock/unlock are currently not
2123 // eliminated. This should be investigated as a future enhancement.
2124 //
2125 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) {
2126
2127 if (!alock->is_eliminated()) {
2128 return false;
2129 }
2130 #ifdef ASSERT
2131 if (!alock->is_coarsened()) {
2132 // Check that new "eliminated" BoxLock node is created.
2133 BoxLockNode* oldbox = alock->box_node()->as_BoxLock();
2134 assert(oldbox->is_eliminated(), "should be done already");
2135 }
2136 #endif
2137
2138 alock->log_lock_optimization(C, "eliminate_lock");
2139
2140 #ifndef PRODUCT
2141 if (PrintEliminateLocks) {
2142 if (alock->is_Lock()) {
2143 tty->print_cr("++++ Eliminated: %d Lock", alock->_idx);
2144 } else {
2145 tty->print_cr("++++ Eliminated: %d Unlock", alock->_idx);
2146 }
2147 }
2148 #endif
2149
2150 Node* mem = alock->in(TypeFunc::Memory);
2151 Node* ctrl = alock->in(TypeFunc::Control);
2152
2153 extract_call_projections(alock);
2154 // There are 2 projections from the lock. The lock node will
2155 // be deleted when its last use is subsumed below.
2156 assert(alock->outcnt() == 2 &&
2157 _fallthroughproj != NULL &&
2158 _memproj_fallthrough != NULL,
2159 "Unexpected projections from Lock/Unlock");
2160
2161 Node* fallthroughproj = _fallthroughproj;
2162 Node* memproj_fallthrough = _memproj_fallthrough;
2163
2164 // The memory projection from a lock/unlock is RawMem
2165 // The input to a Lock is merged memory, so extract its RawMem input
2166 // (unless the MergeMem has been optimized away.)
2167 if (alock->is_Lock()) {
2168 // Seach for MemBarAcquireLock node and delete it also.
2169 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
2170 assert(membar != NULL && membar->Opcode() == Op_MemBarAcquireLock, "");
2171 Node* ctrlproj = membar->proj_out(TypeFunc::Control);
2172 Node* memproj = membar->proj_out(TypeFunc::Memory);
2173 _igvn.replace_node(ctrlproj, fallthroughproj);
2174 _igvn.replace_node(memproj, memproj_fallthrough);
2175
2176 // Delete FastLock node also if this Lock node is unique user
2177 // (a loop peeling may clone a Lock node).
2178 Node* flock = alock->as_Lock()->fastlock_node();
2179 if (flock->outcnt() == 1) {
2180 assert(flock->unique_out() == alock, "sanity");
2181 _igvn.replace_node(flock, top());
2182 }
2183 }
2184
2185 // Seach for MemBarReleaseLock node and delete it also.
2186 if (alock->is_Unlock() && ctrl != NULL && ctrl->is_Proj() &&
2187 ctrl->in(0)->is_MemBar()) {
2188 MemBarNode* membar = ctrl->in(0)->as_MemBar();
2189 assert(membar->Opcode() == Op_MemBarReleaseLock &&
2190 mem->is_Proj() && membar == mem->in(0), "");
2191 _igvn.replace_node(fallthroughproj, ctrl);
2192 _igvn.replace_node(memproj_fallthrough, mem);
2193 fallthroughproj = ctrl;
2194 memproj_fallthrough = mem;
2195 ctrl = membar->in(TypeFunc::Control);
2196 mem = membar->in(TypeFunc::Memory);
2197 }
2198
2199 _igvn.replace_node(fallthroughproj, ctrl);
2200 _igvn.replace_node(memproj_fallthrough, mem);
2201 return true;
2202 }
2203
2204
2205 //------------------------------expand_lock_node----------------------
2206 void PhaseMacroExpand::expand_lock_node(LockNode *lock) {
2207
2208 Node* ctrl = lock->in(TypeFunc::Control);
2209 Node* mem = lock->in(TypeFunc::Memory);
2210 Node* obj = lock->obj_node();
2211 Node* box = lock->box_node();
2212 Node* flock = lock->fastlock_node();
2213
2214 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2215
2216 // Make the merge point
2217 Node *region;
2218 Node *mem_phi;
2219 Node *slow_path;
2220
2221 if (UseOptoBiasInlining) {
2222 /*
2223 * See the full description in MacroAssembler::biased_locking_enter().
2224 *
2225 * if( (mark_word & biased_lock_mask) == biased_lock_pattern ) {
2226 * // The object is biased.
2227 * proto_node = klass->prototype_header;
2228 * o_node = thread | proto_node;
2229 * x_node = o_node ^ mark_word;
2230 * if( (x_node & ~age_mask) == 0 ) { // Biased to the current thread ?
2231 * // Done.
2232 * } else {
2233 * if( (x_node & biased_lock_mask) != 0 ) {
2234 * // The klass's prototype header is no longer biased.
2235 * cas(&mark_word, mark_word, proto_node)
2236 * goto cas_lock;
2237 * } else {
2238 * // The klass's prototype header is still biased.
2239 * if( (x_node & epoch_mask) != 0 ) { // Expired epoch?
2240 * old = mark_word;
2241 * new = o_node;
2242 * } else {
2243 * // Different thread or anonymous biased.
2244 * old = mark_word & (epoch_mask | age_mask | biased_lock_mask);
2245 * new = thread | old;
2246 * }
2247 * // Try to rebias.
2248 * if( cas(&mark_word, old, new) == 0 ) {
2249 * // Done.
2250 * } else {
2251 * goto slow_path; // Failed.
2252 * }
2253 * }
2254 * }
2255 * } else {
2256 * // The object is not biased.
2257 * cas_lock:
2258 * if( FastLock(obj) == 0 ) {
2259 * // Done.
2260 * } else {
2261 * slow_path:
2262 * OptoRuntime::complete_monitor_locking_Java(obj);
2263 * }
2264 * }
2265 */
2266
2267 region = new RegionNode(5);
2268 // create a Phi for the memory state
2269 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2270
2271 Node* fast_lock_region = new RegionNode(3);
2272 Node* fast_lock_mem_phi = new PhiNode( fast_lock_region, Type::MEMORY, TypeRawPtr::BOTTOM);
2273
2274 // First, check mark word for the biased lock pattern.
2275 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2276
2277 // Get fast path - mark word has the biased lock pattern.
2278 ctrl = opt_bits_test(ctrl, fast_lock_region, 1, mark_node,
2279 markOopDesc::biased_lock_mask_in_place,
2280 markOopDesc::biased_lock_pattern, true);
2281 // fast_lock_region->in(1) is set to slow path.
2282 fast_lock_mem_phi->init_req(1, mem);
2283
2284 // Now check that the lock is biased to the current thread and has
2285 // the same epoch and bias as Klass::_prototype_header.
2286
2287 // Special-case a fresh allocation to avoid building nodes:
2288 Node* klass_node = AllocateNode::Ideal_klass(obj, &_igvn);
2289 if (klass_node == NULL) {
2290 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
2291 klass_node = transform_later(LoadKlassNode::make(_igvn, NULL, mem, k_adr, _igvn.type(k_adr)->is_ptr()));
2292 #ifdef _LP64
2293 if (UseCompressedClassPointers && klass_node->is_DecodeNKlass()) {
2294 assert(klass_node->in(1)->Opcode() == Op_LoadNKlass, "sanity");
2295 klass_node->in(1)->init_req(0, ctrl);
2296 } else
2297 #endif
2298 klass_node->init_req(0, ctrl);
2299 }
2300 Node *proto_node = make_load(ctrl, mem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeX_X, TypeX_X->basic_type());
2301
2302 Node* thread = transform_later(new ThreadLocalNode());
2303 Node* cast_thread = transform_later(new CastP2XNode(ctrl, thread));
2304 Node* o_node = transform_later(new OrXNode(cast_thread, proto_node));
2305 Node* x_node = transform_later(new XorXNode(o_node, mark_node));
2306
2307 // Get slow path - mark word does NOT match the value.
2308 Node* not_biased_ctrl = opt_bits_test(ctrl, region, 3, x_node,
2309 (~markOopDesc::age_mask_in_place), 0);
2310 // region->in(3) is set to fast path - the object is biased to the current thread.
2311 mem_phi->init_req(3, mem);
2312
2313
2314 // Mark word does NOT match the value (thread | Klass::_prototype_header).
2315
2316
2317 // First, check biased pattern.
2318 // Get fast path - _prototype_header has the same biased lock pattern.
2319 ctrl = opt_bits_test(not_biased_ctrl, fast_lock_region, 2, x_node,
2320 markOopDesc::biased_lock_mask_in_place, 0, true);
2321
2322 not_biased_ctrl = fast_lock_region->in(2); // Slow path
2323 // fast_lock_region->in(2) - the prototype header is no longer biased
2324 // and we have to revoke the bias on this object.
2325 // We are going to try to reset the mark of this object to the prototype
2326 // value and fall through to the CAS-based locking scheme.
2327 Node* adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
2328 Node* cas = new StoreXConditionalNode(not_biased_ctrl, mem, adr,
2329 proto_node, mark_node);
2330 transform_later(cas);
2331 Node* proj = transform_later(new SCMemProjNode(cas));
2332 fast_lock_mem_phi->init_req(2, proj);
2333
2334
2335 // Second, check epoch bits.
2336 Node* rebiased_region = new RegionNode(3);
2337 Node* old_phi = new PhiNode( rebiased_region, TypeX_X);
2338 Node* new_phi = new PhiNode( rebiased_region, TypeX_X);
2339
2340 // Get slow path - mark word does NOT match epoch bits.
2341 Node* epoch_ctrl = opt_bits_test(ctrl, rebiased_region, 1, x_node,
2342 markOopDesc::epoch_mask_in_place, 0);
2343 // The epoch of the current bias is not valid, attempt to rebias the object
2344 // toward the current thread.
2345 rebiased_region->init_req(2, epoch_ctrl);
2346 old_phi->init_req(2, mark_node);
2347 new_phi->init_req(2, o_node);
2348
2349 // rebiased_region->in(1) is set to fast path.
2350 // The epoch of the current bias is still valid but we know
2351 // nothing about the owner; it might be set or it might be clear.
2352 Node* cmask = MakeConX(markOopDesc::biased_lock_mask_in_place |
2353 markOopDesc::age_mask_in_place |
2354 markOopDesc::epoch_mask_in_place);
2355 Node* old = transform_later(new AndXNode(mark_node, cmask));
2356 cast_thread = transform_later(new CastP2XNode(ctrl, thread));
2357 Node* new_mark = transform_later(new OrXNode(cast_thread, old));
2358 old_phi->init_req(1, old);
2359 new_phi->init_req(1, new_mark);
2360
2361 transform_later(rebiased_region);
2362 transform_later(old_phi);
2363 transform_later(new_phi);
2364
2365 // Try to acquire the bias of the object using an atomic operation.
2366 // If this fails we will go in to the runtime to revoke the object's bias.
2367 cas = new StoreXConditionalNode(rebiased_region, mem, adr, new_phi, old_phi);
2368 transform_later(cas);
2369 proj = transform_later(new SCMemProjNode(cas));
2370
2371 // Get slow path - Failed to CAS.
2372 not_biased_ctrl = opt_bits_test(rebiased_region, region, 4, cas, 0, 0);
2373 mem_phi->init_req(4, proj);
2374 // region->in(4) is set to fast path - the object is rebiased to the current thread.
2375
2376 // Failed to CAS.
2377 slow_path = new RegionNode(3);
2378 Node *slow_mem = new PhiNode( slow_path, Type::MEMORY, TypeRawPtr::BOTTOM);
2379
2380 slow_path->init_req(1, not_biased_ctrl); // Capture slow-control
2381 slow_mem->init_req(1, proj);
2382
2383 // Call CAS-based locking scheme (FastLock node).
2384
2385 transform_later(fast_lock_region);
2386 transform_later(fast_lock_mem_phi);
2387
2388 // Get slow path - FastLock failed to lock the object.
2389 ctrl = opt_bits_test(fast_lock_region, region, 2, flock, 0, 0);
2390 mem_phi->init_req(2, fast_lock_mem_phi);
2391 // region->in(2) is set to fast path - the object is locked to the current thread.
2392
2393 slow_path->init_req(2, ctrl); // Capture slow-control
2394 slow_mem->init_req(2, fast_lock_mem_phi);
2395
2396 transform_later(slow_path);
2397 transform_later(slow_mem);
2398 // Reset lock's memory edge.
2399 lock->set_req(TypeFunc::Memory, slow_mem);
2400
2401 } else {
2402 region = new RegionNode(3);
2403 // create a Phi for the memory state
2404 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2405
2406 // Optimize test; set region slot 2
2407 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0);
2408 mem_phi->init_req(2, mem);
2409 }
2410
2411 // Make slow path call
2412 CallNode *call = make_slow_call((CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(),
2413 OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path,
2414 obj, box, NULL);
2415
2416 extract_call_projections(call);
2417
2418 // Slow path can only throw asynchronous exceptions, which are always
2419 // de-opted. So the compiler thinks the slow-call can never throw an
2420 // exception. If it DOES throw an exception we would need the debug
2421 // info removed first (since if it throws there is no monitor).
2422 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
2423 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
2424
2425 // Capture slow path
2426 // disconnect fall-through projection from call and create a new one
2427 // hook up users of fall-through projection to region
2428 Node *slow_ctrl = _fallthroughproj->clone();
2429 transform_later(slow_ctrl);
2430 _igvn.hash_delete(_fallthroughproj);
2431 _fallthroughproj->disconnect_inputs(NULL, C);
2432 region->init_req(1, slow_ctrl);
2433 // region inputs are now complete
2434 transform_later(region);
2435 _igvn.replace_node(_fallthroughproj, region);
2436
2437 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory));
2438 mem_phi->init_req(1, memproj );
2439 transform_later(mem_phi);
2440 _igvn.replace_node(_memproj_fallthrough, mem_phi);
2441 }
2442
2443 //------------------------------expand_unlock_node----------------------
2444 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
2445
2446 Node* ctrl = unlock->in(TypeFunc::Control);
2447 Node* mem = unlock->in(TypeFunc::Memory);
2448 Node* obj = unlock->obj_node();
2449 Node* box = unlock->box_node();
2450
2451 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2452
2453 // No need for a null check on unlock
2454
2455 // Make the merge point
2456 Node *region;
2457 Node *mem_phi;
2458
2459 if (UseOptoBiasInlining) {
2460 // Check for biased locking unlock case, which is a no-op.
2461 // See the full description in MacroAssembler::biased_locking_exit().
2462 region = new RegionNode(4);
2463 // create a Phi for the memory state
2464 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2465 mem_phi->init_req(3, mem);
2466
2467 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2468 ctrl = opt_bits_test(ctrl, region, 3, mark_node,
2469 markOopDesc::biased_lock_mask_in_place,
2470 markOopDesc::biased_lock_pattern);
2471 } else {
2472 region = new RegionNode(3);
2473 // create a Phi for the memory state
2474 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2475 }
2476
2477 FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box );
2478 funlock = transform_later( funlock )->as_FastUnlock();
2479 // Optimize test; set region slot 2
2480 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0);
2481 Node *thread = transform_later(new ThreadLocalNode());
2482
2483 CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(),
2484 CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C),
2485 "complete_monitor_unlocking_C", slow_path, obj, box, thread);
2486
2487 extract_call_projections(call);
2488
2489 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
2490 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
2491
2492 // No exceptions for unlocking
2493 // Capture slow path
2494 // disconnect fall-through projection from call and create a new one
2495 // hook up users of fall-through projection to region
2496 Node *slow_ctrl = _fallthroughproj->clone();
2497 transform_later(slow_ctrl);
2498 _igvn.hash_delete(_fallthroughproj);
2499 _fallthroughproj->disconnect_inputs(NULL, C);
2500 region->init_req(1, slow_ctrl);
2501 // region inputs are now complete
2502 transform_later(region);
2503 _igvn.replace_node(_fallthroughproj, region);
2504
2505 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory) );
2506 mem_phi->init_req(1, memproj );
2507 mem_phi->init_req(2, mem);
2508 transform_later(mem_phi);
2509 _igvn.replace_node(_memproj_fallthrough, mem_phi);
2510 }
2511
2512 //---------------------------eliminate_macro_nodes----------------------
2513 // Eliminate scalar replaced allocations and associated locks.
2514 void PhaseMacroExpand::eliminate_macro_nodes() {
2515 if (C->macro_count() == 0)
2516 return;
2517
2518 // First, attempt to eliminate locks
2519 int cnt = C->macro_count();
2520 for (int i=0; i < cnt; i++) {
2521 Node *n = C->macro_node(i);
2522 if (n->is_AbstractLock()) { // Lock and Unlock nodes
2523 // Before elimination mark all associated (same box and obj)
2524 // lock and unlock nodes.
2525 mark_eliminated_locking_nodes(n->as_AbstractLock());
2526 }
2527 }
2528 bool progress = true;
2529 while (progress) {
2530 progress = false;
2531 for (int i = C->macro_count(); i > 0; i--) {
2532 Node * n = C->macro_node(i-1);
2533 bool success = false;
2534 debug_only(int old_macro_count = C->macro_count(););
2535 if (n->is_AbstractLock()) {
2536 success = eliminate_locking_node(n->as_AbstractLock());
2537 }
2538 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2539 progress = progress || success;
2540 }
2541 }
2542 // Next, attempt to eliminate allocations
2543 _has_locks = false;
2544 progress = true;
2545 while (progress) {
2546 progress = false;
2547 for (int i = C->macro_count(); i > 0; i--) {
2548 Node * n = C->macro_node(i-1);
2549 bool success = false;
2550 debug_only(int old_macro_count = C->macro_count(););
2551 switch (n->class_id()) {
2552 case Node::Class_Allocate:
2553 case Node::Class_AllocateArray:
2554 success = eliminate_allocate_node(n->as_Allocate());
2555 break;
2556 case Node::Class_CallStaticJava:
2557 success = eliminate_boxing_node(n->as_CallStaticJava());
2558 break;
2559 case Node::Class_Lock:
2560 case Node::Class_Unlock:
2561 assert(!n->as_AbstractLock()->is_eliminated(), "sanity");
2562 _has_locks = true;
2563 break;
2564 case Node::Class_ArrayCopy:
2565 break;
2566 default:
2567 assert(n->Opcode() == Op_LoopLimit ||
2568 n->Opcode() == Op_Opaque1 ||
2569 n->Opcode() == Op_Opaque2 ||
2570 n->Opcode() == Op_Opaque3, "unknown node type in macro list");
2571 }
2572 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2573 progress = progress || success;
2574 }
2575 }
2576 }
2577
2578 //------------------------------expand_macro_nodes----------------------
2579 // Returns true if a failure occurred.
2580 bool PhaseMacroExpand::expand_macro_nodes() {
2581 // Last attempt to eliminate macro nodes.
2582 eliminate_macro_nodes();
2583
2584 // Make sure expansion will not cause node limit to be exceeded.
2585 // Worst case is a macro node gets expanded into about 200 nodes.
2586 // Allow 50% more for optimization.
2587 if (C->check_node_count(C->macro_count() * 300, "out of nodes before macro expansion" ) )
2588 return true;
2589
2590 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations.
2591 bool progress = true;
2592 while (progress) {
2593 progress = false;
2594 for (int i = C->macro_count(); i > 0; i--) {
2595 Node * n = C->macro_node(i-1);
2596 bool success = false;
2597 debug_only(int old_macro_count = C->macro_count(););
2598 if (n->Opcode() == Op_LoopLimit) {
2599 // Remove it from macro list and put on IGVN worklist to optimize.
2600 C->remove_macro_node(n);
2601 _igvn._worklist.push(n);
2602 success = true;
2603 } else if (n->Opcode() == Op_CallStaticJava) {
2604 // Remove it from macro list and put on IGVN worklist to optimize.
2605 C->remove_macro_node(n);
2606 _igvn._worklist.push(n);
2607 success = true;
2608 } else if (n->Opcode() == Op_Opaque1 || n->Opcode() == Op_Opaque2) {
2609 _igvn.replace_node(n, n->in(1));
2610 success = true;
2611 #if INCLUDE_RTM_OPT
2612 } else if ((n->Opcode() == Op_Opaque3) && ((Opaque3Node*)n)->rtm_opt()) {
2613 assert(C->profile_rtm(), "should be used only in rtm deoptimization code");
2614 assert((n->outcnt() == 1) && n->unique_out()->is_Cmp(), "");
2615 Node* cmp = n->unique_out();
2616 #ifdef ASSERT
2617 // Validate graph.
2618 assert((cmp->outcnt() == 1) && cmp->unique_out()->is_Bool(), "");
2619 BoolNode* bol = cmp->unique_out()->as_Bool();
2620 assert((bol->outcnt() == 1) && bol->unique_out()->is_If() &&
2621 (bol->_test._test == BoolTest::ne), "");
2622 IfNode* ifn = bol->unique_out()->as_If();
2623 assert((ifn->outcnt() == 2) &&
2624 ifn->proj_out(1)->is_uncommon_trap_proj(Deoptimization::Reason_rtm_state_change) != NULL, "");
2625 #endif
2626 Node* repl = n->in(1);
2627 if (!_has_locks) {
2628 // Remove RTM state check if there are no locks in the code.
2629 // Replace input to compare the same value.
2630 repl = (cmp->in(1) == n) ? cmp->in(2) : cmp->in(1);
2631 }
2632 _igvn.replace_node(n, repl);
2633 success = true;
2634 #endif
2635 }
2636 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2637 progress = progress || success;
2638 }
2639 }
2640
2641 // expand arraycopy "macro" nodes first
2642 // For ReduceBulkZeroing, we must first process all arraycopy nodes
2643 // before the allocate nodes are expanded.
2644 int macro_idx = C->macro_count() - 1;
2645 while (macro_idx >= 0) {
2646 Node * n = C->macro_node(macro_idx);
2647 assert(n->is_macro(), "only macro nodes expected here");
2648 if (_igvn.type(n) == Type::TOP || n->in(0)->is_top() ) {
2649 // node is unreachable, so don't try to expand it
2650 C->remove_macro_node(n);
2651 } else if (n->is_ArrayCopy()){
2652 int macro_count = C->macro_count();
2653 expand_arraycopy_node(n->as_ArrayCopy());
2654 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
2655 }
2656 if (C->failing()) return true;
2657 macro_idx --;
2658 }
2659
2660 // expand "macro" nodes
2661 // nodes are removed from the macro list as they are processed
2662 while (C->macro_count() > 0) {
2663 int macro_count = C->macro_count();
2664 Node * n = C->macro_node(macro_count-1);
2665 assert(n->is_macro(), "only macro nodes expected here");
2666 if (_igvn.type(n) == Type::TOP || n->in(0)->is_top() ) {
2667 // node is unreachable, so don't try to expand it
2668 C->remove_macro_node(n);
2669 continue;
2670 }
2671 switch (n->class_id()) {
2672 case Node::Class_Allocate:
2673 expand_allocate(n->as_Allocate());
2674 break;
2675 case Node::Class_AllocateArray:
2676 expand_allocate_array(n->as_AllocateArray());
2677 break;
2678 case Node::Class_Lock:
2679 expand_lock_node(n->as_Lock());
2680 break;
2681 case Node::Class_Unlock:
2682 expand_unlock_node(n->as_Unlock());
2683 break;
2684 default:
2685 assert(false, "unknown node type in macro list");
2686 }
2687 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
2688 if (C->failing()) return true;
2689 }
2690
2691 _igvn.set_delay_transform(false);
2692 _igvn.optimize();
2693 if (C->failing()) return true;
2694 return false;
2695 }