1 /*
2 * Copyright (c) 2005, 2018, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "compiler/compileLog.hpp"
27 #include "gc/shared/collectedHeap.inline.hpp"
28 #include "libadt/vectset.hpp"
29 #include "opto/addnode.hpp"
30 #include "opto/arraycopynode.hpp"
31 #include "opto/callnode.hpp"
32 #include "opto/castnode.hpp"
33 #include "opto/cfgnode.hpp"
34 #include "opto/compile.hpp"
35 #include "opto/convertnode.hpp"
36 #include "opto/graphKit.hpp"
37 #include "opto/locknode.hpp"
38 #include "opto/loopnode.hpp"
39 #include "opto/macro.hpp"
40 #include "opto/memnode.hpp"
41 #include "opto/narrowptrnode.hpp"
42 #include "opto/node.hpp"
43 #include "opto/opaquenode.hpp"
44 #include "opto/phaseX.hpp"
45 #include "opto/rootnode.hpp"
46 #include "opto/runtime.hpp"
47 #include "opto/subnode.hpp"
48 #include "opto/type.hpp"
49 #include "opto/valuetypenode.hpp"
50 #include "runtime/sharedRuntime.hpp"
51 #if INCLUDE_G1GC
52 #include "gc/g1/g1ThreadLocalData.hpp"
53 #endif // INCLUDE_G1GC
54
55
56 //
57 // Replace any references to "oldref" in inputs to "use" with "newref".
58 // Returns the number of replacements made.
59 //
60 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) {
61 int nreplacements = 0;
62 uint req = use->req();
63 for (uint j = 0; j < use->len(); j++) {
64 Node *uin = use->in(j);
65 if (uin == oldref) {
66 if (j < req)
67 use->set_req(j, newref);
68 else
69 use->set_prec(j, newref);
70 nreplacements++;
71 } else if (j >= req && uin == NULL) {
72 break;
73 }
74 }
75 return nreplacements;
76 }
77
78 void PhaseMacroExpand::copy_call_debug_info(CallNode *oldcall, CallNode * newcall) {
79 // Copy debug information and adjust JVMState information
80 uint old_dbg_start = oldcall->tf()->domain_sig()->cnt();
81 uint new_dbg_start = newcall->tf()->domain_sig()->cnt();
82 int jvms_adj = new_dbg_start - old_dbg_start;
83 assert (new_dbg_start == newcall->req(), "argument count mismatch");
84
85 // SafePointScalarObject node could be referenced several times in debug info.
86 // Use Dict to record cloned nodes.
87 Dict* sosn_map = new Dict(cmpkey,hashkey);
88 for (uint i = old_dbg_start; i < oldcall->req(); i++) {
89 Node* old_in = oldcall->in(i);
90 // Clone old SafePointScalarObjectNodes, adjusting their field contents.
91 if (old_in != NULL && old_in->is_SafePointScalarObject()) {
92 SafePointScalarObjectNode* old_sosn = old_in->as_SafePointScalarObject();
93 uint old_unique = C->unique();
94 Node* new_in = old_sosn->clone(sosn_map);
95 if (old_unique != C->unique()) { // New node?
96 new_in->set_req(0, C->root()); // reset control edge
97 new_in = transform_later(new_in); // Register new node.
98 }
99 old_in = new_in;
100 }
101 newcall->add_req(old_in);
102 }
103
104 // JVMS may be shared so clone it before we modify it
105 newcall->set_jvms(oldcall->jvms() != NULL ? oldcall->jvms()->clone_deep(C) : NULL);
106 for (JVMState *jvms = newcall->jvms(); jvms != NULL; jvms = jvms->caller()) {
107 jvms->set_map(newcall);
108 jvms->set_locoff(jvms->locoff()+jvms_adj);
109 jvms->set_stkoff(jvms->stkoff()+jvms_adj);
110 jvms->set_monoff(jvms->monoff()+jvms_adj);
111 jvms->set_scloff(jvms->scloff()+jvms_adj);
112 jvms->set_endoff(jvms->endoff()+jvms_adj);
113 }
114 }
115
116 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word, int mask, int bits, bool return_fast_path) {
117 Node* cmp;
118 if (mask != 0) {
119 Node* and_node = transform_later(new AndXNode(word, MakeConX(mask)));
120 cmp = transform_later(new CmpXNode(and_node, MakeConX(bits)));
121 } else {
122 cmp = word;
123 }
124 Node* bol = transform_later(new BoolNode(cmp, BoolTest::ne));
125 IfNode* iff = new IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN );
126 transform_later(iff);
127
128 // Fast path taken.
129 Node *fast_taken = transform_later(new IfFalseNode(iff));
130
131 // Fast path not-taken, i.e. slow path
132 Node *slow_taken = transform_later(new IfTrueNode(iff));
133
134 if (return_fast_path) {
135 region->init_req(edge, slow_taken); // Capture slow-control
136 return fast_taken;
137 } else {
138 region->init_req(edge, fast_taken); // Capture fast-control
139 return slow_taken;
140 }
141 }
142
143 //--------------------copy_predefined_input_for_runtime_call--------------------
144 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) {
145 // Set fixed predefined input arguments
146 call->init_req( TypeFunc::Control, ctrl );
147 call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) );
148 call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ?????
149 call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) );
150 call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) );
151 }
152
153 //------------------------------make_slow_call---------------------------------
154 CallNode* PhaseMacroExpand::make_slow_call(CallNode *oldcall, const TypeFunc* slow_call_type,
155 address slow_call, const char* leaf_name, Node* slow_path,
156 Node* parm0, Node* parm1, Node* parm2) {
157
158 // Slow-path call
159 CallNode *call = leaf_name
160 ? (CallNode*)new CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM )
161 : (CallNode*)new CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), oldcall->jvms()->bci(), TypeRawPtr::BOTTOM );
162
163 // Slow path call has no side-effects, uses few values
164 copy_predefined_input_for_runtime_call(slow_path, oldcall, call );
165 if (parm0 != NULL) call->init_req(TypeFunc::Parms+0, parm0);
166 if (parm1 != NULL) call->init_req(TypeFunc::Parms+1, parm1);
167 if (parm2 != NULL) call->init_req(TypeFunc::Parms+2, parm2);
168 copy_call_debug_info(oldcall, call);
169 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
170 _igvn.replace_node(oldcall, call);
171 transform_later(call);
172
173 return call;
174 }
175
176 void PhaseMacroExpand::extract_call_projections(CallNode *call) {
177 _fallthroughproj = NULL;
178 _fallthroughcatchproj = NULL;
179 _ioproj_fallthrough = NULL;
180 _ioproj_catchall = NULL;
181 _catchallcatchproj = NULL;
182 _memproj_fallthrough = NULL;
183 _memproj_catchall = NULL;
184 _resproj = NULL;
185 for (DUIterator_Fast imax, i = call->fast_outs(imax); i < imax; i++) {
186 ProjNode *pn = call->fast_out(i)->as_Proj();
187 switch (pn->_con) {
188 case TypeFunc::Control:
189 {
190 // For Control (fallthrough) and I_O (catch_all_index) we have CatchProj -> Catch -> Proj
191 _fallthroughproj = pn;
192 DUIterator_Fast jmax, j = pn->fast_outs(jmax);
193 const Node *cn = pn->fast_out(j);
194 if (cn->is_Catch()) {
195 ProjNode *cpn = NULL;
196 for (DUIterator_Fast kmax, k = cn->fast_outs(kmax); k < kmax; k++) {
197 cpn = cn->fast_out(k)->as_Proj();
198 assert(cpn->is_CatchProj(), "must be a CatchProjNode");
199 if (cpn->_con == CatchProjNode::fall_through_index)
200 _fallthroughcatchproj = cpn;
201 else {
202 assert(cpn->_con == CatchProjNode::catch_all_index, "must be correct index.");
203 _catchallcatchproj = cpn;
204 }
205 }
206 }
207 break;
208 }
209 case TypeFunc::I_O:
210 if (pn->_is_io_use)
211 _ioproj_catchall = pn;
212 else
213 _ioproj_fallthrough = pn;
214 break;
215 case TypeFunc::Memory:
216 if (pn->_is_io_use)
217 _memproj_catchall = pn;
218 else
219 _memproj_fallthrough = pn;
220 break;
221 case TypeFunc::Parms:
222 _resproj = pn;
223 break;
224 default:
225 assert(false, "unexpected projection from allocation node.");
226 }
227 }
228
229 }
230
231 void PhaseMacroExpand::eliminate_gc_barrier(Node* p2x) {
232 BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2();
233 bs->eliminate_gc_barrier(this, p2x);
234 }
235
236 // Search for a memory operation for the specified memory slice.
237 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) {
238 Node *orig_mem = mem;
239 Node *alloc_mem = alloc->in(TypeFunc::Memory);
240 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr();
241 while (true) {
242 if (mem == alloc_mem || mem == start_mem ) {
243 return mem; // hit one of our sentinels
244 } else if (mem->is_MergeMem()) {
245 mem = mem->as_MergeMem()->memory_at(alias_idx);
246 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) {
247 Node *in = mem->in(0);
248 // we can safely skip over safepoints, calls, locks and membars because we
249 // already know that the object is safe to eliminate.
250 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) {
251 return in;
252 } else if (in->is_Call()) {
253 CallNode *call = in->as_Call();
254 if (call->may_modify(tinst, phase)) {
255 assert(call->is_ArrayCopy(), "ArrayCopy is the only call node that doesn't make allocation escape");
256 if (call->as_ArrayCopy()->modifies(offset, offset, phase, false)) {
257 return in;
258 }
259 }
260 mem = in->in(TypeFunc::Memory);
261 } else if (in->is_MemBar()) {
262 ArrayCopyNode* ac = NULL;
263 if (ArrayCopyNode::may_modify(tinst, in->as_MemBar(), phase, ac)) {
264 assert(ac != NULL && ac->is_clonebasic(), "Only basic clone is a non escaping clone");
265 return ac;
266 }
267 mem = in->in(TypeFunc::Memory);
268 } else {
269 assert(false, "unexpected projection");
270 }
271 } else if (mem->is_Store()) {
272 const TypePtr* atype = mem->as_Store()->adr_type();
273 int adr_idx = phase->C->get_alias_index(atype);
274 if (adr_idx == alias_idx) {
275 assert(atype->isa_oopptr(), "address type must be oopptr");
276 int adr_offset = atype->flattened_offset();
277 uint adr_iid = atype->is_oopptr()->instance_id();
278 // Array elements references have the same alias_idx
279 // but different offset and different instance_id.
280 if (adr_offset == offset && adr_iid == alloc->_idx)
281 return mem;
282 } else {
283 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw");
284 }
285 mem = mem->in(MemNode::Memory);
286 } else if (mem->is_ClearArray()) {
287 if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) {
288 // Can not bypass initialization of the instance
289 // we are looking.
290 debug_only(intptr_t offset;)
291 assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity");
292 InitializeNode* init = alloc->as_Allocate()->initialization();
293 // We are looking for stored value, return Initialize node
294 // or memory edge from Allocate node.
295 if (init != NULL)
296 return init;
297 else
298 return alloc->in(TypeFunc::Memory); // It will produce zero value (see callers).
299 }
300 // Otherwise skip it (the call updated 'mem' value).
301 } else if (mem->Opcode() == Op_SCMemProj) {
302 mem = mem->in(0);
303 Node* adr = NULL;
304 if (mem->is_LoadStore()) {
305 adr = mem->in(MemNode::Address);
306 } else {
307 assert(mem->Opcode() == Op_EncodeISOArray ||
308 mem->Opcode() == Op_StrCompressedCopy, "sanity");
309 adr = mem->in(3); // Destination array
310 }
311 const TypePtr* atype = adr->bottom_type()->is_ptr();
312 int adr_idx = phase->C->get_alias_index(atype);
313 if (adr_idx == alias_idx) {
314 DEBUG_ONLY(mem->dump();)
315 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
316 return NULL;
317 }
318 mem = mem->in(MemNode::Memory);
319 } else if (mem->Opcode() == Op_StrInflatedCopy) {
320 Node* adr = mem->in(3); // Destination array
321 const TypePtr* atype = adr->bottom_type()->is_ptr();
322 int adr_idx = phase->C->get_alias_index(atype);
323 if (adr_idx == alias_idx) {
324 DEBUG_ONLY(mem->dump();)
325 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field");
326 return NULL;
327 }
328 mem = mem->in(MemNode::Memory);
329 } else {
330 return mem;
331 }
332 assert(mem != orig_mem, "dead memory loop");
333 }
334 }
335
336 // Generate loads from source of the arraycopy for fields of
337 // destination needed at a deoptimization point
338 Node* PhaseMacroExpand::make_arraycopy_load(ArrayCopyNode* ac, intptr_t offset, Node* ctl, Node* mem, BasicType ft, const Type *ftype, AllocateNode *alloc) {
339 BasicType bt = ft;
340 const Type *type = ftype;
341 if (ft == T_NARROWOOP) {
342 bt = T_OBJECT;
343 type = ftype->make_oopptr();
344 }
345 Node* res = NULL;
346 if (ac->is_clonebasic()) {
347 Node* base = ac->in(ArrayCopyNode::Src)->in(AddPNode::Base);
348 Node* adr = _igvn.transform(new AddPNode(base, base, MakeConX(offset)));
349 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset);
350 res = LoadNode::make(_igvn, ctl, mem, adr, adr_type, type, bt, MemNode::unordered, LoadNode::Pinned);
351 } else {
352 if (ac->modifies(offset, offset, &_igvn, true)) {
353 assert(ac->in(ArrayCopyNode::Dest) == alloc->result_cast(), "arraycopy destination should be allocation's result");
354 uint shift = exact_log2(type2aelembytes(bt));
355 Node* diff = _igvn.transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
356 #ifdef _LP64
357 diff = _igvn.transform(new ConvI2LNode(diff));
358 #endif
359 diff = _igvn.transform(new LShiftXNode(diff, intcon(shift)));
360
361 Node* off = _igvn.transform(new AddXNode(MakeConX(offset), diff));
362 Node* base = ac->in(ArrayCopyNode::Src);
363 Node* adr = _igvn.transform(new AddPNode(base, base, off));
364 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset);
365 res = LoadNode::make(_igvn, ctl, mem, adr, adr_type, type, bt, MemNode::unordered, LoadNode::Pinned);
366 }
367 }
368 if (res != NULL) {
369 res = _igvn.transform(res);
370 if (ftype->isa_narrowoop()) {
371 // PhaseMacroExpand::scalar_replacement adds DecodeN nodes
372 res = _igvn.transform(new EncodePNode(res, ftype));
373 }
374 return res;
375 }
376 return NULL;
377 }
378
379 //
380 // Given a Memory Phi, compute a value Phi containing the values from stores
381 // on the input paths.
382 // Note: this function is recursive, its depth is limited by the "level" argument
383 // Returns the computed Phi, or NULL if it cannot compute it.
384 Node *PhaseMacroExpand::value_from_mem_phi(Node *mem, BasicType ft, const Type *phi_type, const TypeOopPtr *adr_t, AllocateNode *alloc, Node_Stack *value_phis, int level) {
385 assert(mem->is_Phi(), "sanity");
386 int alias_idx = C->get_alias_index(adr_t);
387 int offset = adr_t->flattened_offset();
388 int instance_id = adr_t->instance_id();
389
390 // Check if an appropriate value phi already exists.
391 Node* region = mem->in(0);
392 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) {
393 Node* phi = region->fast_out(k);
394 if (phi->is_Phi() && phi != mem &&
395 phi->as_Phi()->is_same_inst_field(phi_type, (int)mem->_idx, instance_id, alias_idx, offset)) {
396 return phi;
397 }
398 }
399 // Check if an appropriate new value phi already exists.
400 Node* new_phi = value_phis->find(mem->_idx);
401 if (new_phi != NULL)
402 return new_phi;
403
404 if (level <= 0) {
405 return NULL; // Give up: phi tree too deep
406 }
407 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
408 Node *alloc_mem = alloc->in(TypeFunc::Memory);
409
410 uint length = mem->req();
411 GrowableArray <Node *> values(length, length, NULL, false);
412
413 // create a new Phi for the value
414 PhiNode *phi = new PhiNode(mem->in(0), phi_type, NULL, mem->_idx, instance_id, alias_idx, offset);
415 transform_later(phi);
416 value_phis->push(phi, mem->_idx);
417
418 for (uint j = 1; j < length; j++) {
419 Node *in = mem->in(j);
420 if (in == NULL || in->is_top()) {
421 values.at_put(j, in);
422 } else {
423 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn);
424 if (val == start_mem || val == alloc_mem) {
425 // hit a sentinel, return appropriate 0 value
426 values.at_put(j, _igvn.zerocon(ft));
427 continue;
428 }
429 if (val->is_Initialize()) {
430 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
431 }
432 if (val == NULL) {
433 return NULL; // can't find a value on this path
434 }
435 if (val == mem) {
436 values.at_put(j, mem);
437 } else if (val->is_Store()) {
438 values.at_put(j, val->in(MemNode::ValueIn));
439 } else if(val->is_Proj() && val->in(0) == alloc) {
440 values.at_put(j, _igvn.zerocon(ft));
441 } else if (val->is_Phi()) {
442 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1);
443 if (val == NULL) {
444 return NULL;
445 }
446 values.at_put(j, val);
447 } else if (val->Opcode() == Op_SCMemProj) {
448 assert(val->in(0)->is_LoadStore() ||
449 val->in(0)->Opcode() == Op_EncodeISOArray ||
450 val->in(0)->Opcode() == Op_StrCompressedCopy, "sanity");
451 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
452 return NULL;
453 } else if (val->is_ArrayCopy()) {
454 Node* res = make_arraycopy_load(val->as_ArrayCopy(), offset, val->in(0), val->in(TypeFunc::Memory), ft, phi_type, alloc);
455 if (res == NULL) {
456 return NULL;
457 }
458 values.at_put(j, res);
459 } else {
460 #ifdef ASSERT
461 val->dump();
462 assert(false, "unknown node on this path");
463 #endif
464 return NULL; // unknown node on this path
465 }
466 }
467 }
468 // Set Phi's inputs
469 for (uint j = 1; j < length; j++) {
470 if (values.at(j) == mem) {
471 phi->init_req(j, phi);
472 } else {
473 phi->init_req(j, values.at(j));
474 }
475 }
476 return phi;
477 }
478
479 // Search the last value stored into the object's field.
480 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, Node *sfpt_ctl, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, AllocateNode *alloc) {
481 assert(adr_t->is_known_instance_field(), "instance required");
482 int instance_id = adr_t->instance_id();
483 assert((uint)instance_id == alloc->_idx, "wrong allocation");
484
485 int alias_idx = C->get_alias_index(adr_t);
486 int offset = adr_t->flattened_offset();
487 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
488 Node *alloc_ctrl = alloc->in(TypeFunc::Control);
489 Node *alloc_mem = alloc->in(TypeFunc::Memory);
490 Arena *a = Thread::current()->resource_area();
491 VectorSet visited(a);
492
493 bool done = sfpt_mem == alloc_mem;
494 Node *mem = sfpt_mem;
495 while (!done) {
496 if (visited.test_set(mem->_idx)) {
497 return NULL; // found a loop, give up
498 }
499 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn);
500 if (mem == start_mem || mem == alloc_mem) {
501 done = true; // hit a sentinel, return appropriate 0 value
502 } else if (mem->is_Initialize()) {
503 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
504 if (mem == NULL) {
505 done = true; // Something went wrong.
506 } else if (mem->is_Store()) {
507 const TypePtr* atype = mem->as_Store()->adr_type();
508 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice");
509 done = true;
510 }
511 } else if (mem->is_Store()) {
512 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr();
513 assert(atype != NULL, "address type must be oopptr");
514 assert(C->get_alias_index(atype) == alias_idx &&
515 atype->is_known_instance_field() && atype->flattened_offset() == offset &&
516 atype->instance_id() == instance_id, "store is correct memory slice");
517 done = true;
518 } else if (mem->is_Phi()) {
519 // try to find a phi's unique input
520 Node *unique_input = NULL;
521 Node *top = C->top();
522 for (uint i = 1; i < mem->req(); i++) {
523 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn);
524 if (n == NULL || n == top || n == mem) {
525 continue;
526 } else if (unique_input == NULL) {
527 unique_input = n;
528 } else if (unique_input != n) {
529 unique_input = top;
530 break;
531 }
532 }
533 if (unique_input != NULL && unique_input != top) {
534 mem = unique_input;
535 } else {
536 done = true;
537 }
538 } else if (mem->is_ArrayCopy()) {
539 done = true;
540 } else {
541 assert(false, "unexpected node");
542 }
543 }
544 if (mem != NULL) {
545 if (mem == start_mem || mem == alloc_mem) {
546 // hit a sentinel, return appropriate 0 value
547 Node* default_value = alloc->in(AllocateNode::DefaultValue);
548 if (default_value != NULL) {
549 return default_value;
550 }
551 assert(alloc->in(AllocateNode::RawDefaultValue) == NULL, "default value may not be null");
552 return _igvn.zerocon(ft);
553 } else if (mem->is_Store()) {
554 return mem->in(MemNode::ValueIn);
555 } else if (mem->is_Phi()) {
556 // attempt to produce a Phi reflecting the values on the input paths of the Phi
557 Node_Stack value_phis(a, 8);
558 Node * phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit);
559 if (phi != NULL) {
560 return phi;
561 } else {
562 // Kill all new Phis
563 while(value_phis.is_nonempty()) {
564 Node* n = value_phis.node();
565 _igvn.replace_node(n, C->top());
566 value_phis.pop();
567 }
568 }
569 } else if (mem->is_ArrayCopy()) {
570 Node* ctl = mem->in(0);
571 Node* m = mem->in(TypeFunc::Memory);
572 if (sfpt_ctl->is_Proj() && sfpt_ctl->as_Proj()->is_uncommon_trap_proj(Deoptimization::Reason_none)) {
573 // pin the loads in the uncommon trap path
574 ctl = sfpt_ctl;
575 m = sfpt_mem;
576 }
577 return make_arraycopy_load(mem->as_ArrayCopy(), offset, ctl, m, ft, ftype, alloc);
578 }
579 }
580 // Something went wrong.
581 return NULL;
582 }
583
584 // Search the last value stored into the value type's fields.
585 Node* PhaseMacroExpand::value_type_from_mem(Node* mem, Node* ctl, ciValueKlass* vk, const TypeAryPtr* adr_type, int offset, AllocateNode* alloc) {
586 // Subtract the offset of the first field to account for the missing oop header
587 offset -= vk->first_field_offset();
588 // Create a new ValueTypeNode and retrieve the field values from memory
589 ValueTypeNode* vt = ValueTypeNode::make_uninitialized(_igvn, vk)->as_ValueType();
590 for (int i = 0; i < vk->nof_declared_nonstatic_fields(); ++i) {
591 ciType* field_type = vt->field_type(i);
592 int field_offset = offset + vt->field_offset(i);
593 // Each value type field has its own memory slice
594 adr_type = adr_type->with_field_offset(field_offset);
595 Node* value = NULL;
596 if (vt->field_is_flattened(i)) {
597 value = value_type_from_mem(mem, ctl, field_type->as_value_klass(), adr_type, field_offset, alloc);
598 } else {
599 const Type* ft = Type::get_const_type(field_type);
600 BasicType bt = field_type->basic_type();
601 if (UseCompressedOops && !is_java_primitive(bt)) {
602 ft = ft->make_narrowoop();
603 bt = T_NARROWOOP;
604 }
605 value = value_from_mem(mem, ctl, bt, ft, adr_type, alloc);
606 if (value != NULL && ft->isa_narrowoop()) {
607 assert(UseCompressedOops, "unexpected narrow oop");
608 value = transform_later(new DecodeNNode(value, value->get_ptr_type()));
609 }
610 /*
611 // TODO attempt scalarization here?
612 if (vt->field_is_flattenable(i)) {
613 // Loading a non-flattened but flattenable value type from memory
614 if (field_type->as_value_klass()->is_scalarizable()) {
615 value = ValueTypeNode::make_from_oop(kit, value, ft->as_value_klass());
616 } else {
617 value = kit->filter_null(value, true, ft->as_value_klass());
618 }
619 }
620 */
621 }
622 if (value != NULL) {
623 vt->set_field_value(i, value);
624 } else {
625 // We might have reached the TrackedInitializationLimit
626 return NULL;
627 }
628 }
629 return vt;
630 }
631
632 // Check the possibility of scalar replacement.
633 bool PhaseMacroExpand::can_eliminate_allocation(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
634 // Scan the uses of the allocation to check for anything that would
635 // prevent us from eliminating it.
636 NOT_PRODUCT( const char* fail_eliminate = NULL; )
637 DEBUG_ONLY( Node* disq_node = NULL; )
638 bool can_eliminate = true;
639
640 Node* res = alloc->result_cast();
641 const TypeOopPtr* res_type = NULL;
642 if (res == NULL) {
643 // All users were eliminated.
644 } else if (!res->is_CheckCastPP()) {
645 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";)
646 can_eliminate = false;
647 } else {
648 res_type = _igvn.type(res)->isa_oopptr();
649 if (res_type == NULL) {
650 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";)
651 can_eliminate = false;
652 } else if (res_type->isa_aryptr()) {
653 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1);
654 if (length < 0) {
655 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";)
656 can_eliminate = false;
657 }
658 }
659 }
660
661 if (can_eliminate && res != NULL) {
662 for (DUIterator_Fast jmax, j = res->fast_outs(jmax);
663 j < jmax && can_eliminate; j++) {
664 Node* use = res->fast_out(j);
665
666 if (use->is_AddP()) {
667 const TypePtr* addp_type = _igvn.type(use)->is_ptr();
668 int offset = addp_type->offset();
669
670 if (offset == Type::OffsetTop || offset == Type::OffsetBot) {
671 NOT_PRODUCT(fail_eliminate = "Undefined field referrence";)
672 can_eliminate = false;
673 break;
674 }
675 for (DUIterator_Fast kmax, k = use->fast_outs(kmax);
676 k < kmax && can_eliminate; k++) {
677 Node* n = use->fast_out(k);
678 if (!n->is_Store() && n->Opcode() != Op_CastP2X &&
679 !(n->is_ArrayCopy() &&
680 n->as_ArrayCopy()->is_clonebasic() &&
681 n->in(ArrayCopyNode::Dest) == use)) {
682 DEBUG_ONLY(disq_node = n;)
683 if (n->is_Load() || n->is_LoadStore()) {
684 NOT_PRODUCT(fail_eliminate = "Field load";)
685 } else {
686 NOT_PRODUCT(fail_eliminate = "Not store field reference";)
687 }
688 can_eliminate = false;
689 }
690 }
691 } else if (use->is_ArrayCopy() &&
692 (use->as_ArrayCopy()->is_arraycopy_validated() ||
693 use->as_ArrayCopy()->is_copyof_validated() ||
694 use->as_ArrayCopy()->is_copyofrange_validated()) &&
695 use->in(ArrayCopyNode::Dest) == res) {
696 // ok to eliminate
697 } else if (use->is_SafePoint()) {
698 SafePointNode* sfpt = use->as_SafePoint();
699 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) {
700 // Object is passed as argument.
701 DEBUG_ONLY(disq_node = use;)
702 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";)
703 can_eliminate = false;
704 }
705 Node* sfptMem = sfpt->memory();
706 if (sfptMem == NULL || sfptMem->is_top()) {
707 DEBUG_ONLY(disq_node = use;)
708 NOT_PRODUCT(fail_eliminate = "NULL or TOP memory";)
709 can_eliminate = false;
710 } else {
711 safepoints.append_if_missing(sfpt);
712 }
713 } else if (use->is_ValueType() && use->isa_ValueType()->get_oop() == res) {
714 // ok to eliminate
715 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
716 if (use->is_Phi()) {
717 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
718 NOT_PRODUCT(fail_eliminate = "Object is return value";)
719 } else {
720 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
721 }
722 DEBUG_ONLY(disq_node = use;)
723 } else {
724 if (use->Opcode() == Op_Return) {
725 NOT_PRODUCT(fail_eliminate = "Object is return value";)
726 } else {
727 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
728 }
729 DEBUG_ONLY(disq_node = use;)
730 }
731 can_eliminate = false;
732 } else {
733 assert(use->Opcode() == Op_CastP2X, "should be");
734 assert(!use->has_out_with(Op_OrL), "should have been removed because oop is never null");
735 }
736 }
737 }
738
739 #ifndef PRODUCT
740 if (PrintEliminateAllocations) {
741 if (can_eliminate) {
742 tty->print("Scalar ");
743 if (res == NULL)
744 alloc->dump();
745 else
746 res->dump();
747 } else if (alloc->_is_scalar_replaceable) {
748 tty->print("NotScalar (%s)", fail_eliminate);
749 if (res == NULL)
750 alloc->dump();
751 else
752 res->dump();
753 #ifdef ASSERT
754 if (disq_node != NULL) {
755 tty->print(" >>>> ");
756 disq_node->dump();
757 }
758 #endif /*ASSERT*/
759 }
760 }
761 #endif
762 return can_eliminate;
763 }
764
765 // Do scalar replacement.
766 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
767 GrowableArray <SafePointNode *> safepoints_done;
768
769 ciKlass* klass = NULL;
770 ciInstanceKlass* iklass = NULL;
771 int nfields = 0;
772 int array_base = 0;
773 int element_size = 0;
774 BasicType basic_elem_type = T_ILLEGAL;
775 ciType* elem_type = NULL;
776
777 Node* res = alloc->result_cast();
778 assert(res == NULL || res->is_CheckCastPP(), "unexpected AllocateNode result");
779 const TypeOopPtr* res_type = NULL;
780 if (res != NULL) { // Could be NULL when there are no users
781 res_type = _igvn.type(res)->isa_oopptr();
782 }
783
784 if (res != NULL) {
785 klass = res_type->klass();
786 if (res_type->isa_instptr()) {
787 // find the fields of the class which will be needed for safepoint debug information
788 assert(klass->is_instance_klass(), "must be an instance klass.");
789 iklass = klass->as_instance_klass();
790 nfields = iklass->nof_nonstatic_fields();
791 } else {
792 // find the array's elements which will be needed for safepoint debug information
793 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
794 assert(klass->is_array_klass() && nfields >= 0, "must be an array klass.");
795 elem_type = klass->as_array_klass()->element_type();
796 basic_elem_type = elem_type->basic_type();
797 if (elem_type->is_valuetype()) {
798 ciValueKlass* vk = elem_type->as_value_klass();
799 if (!vk->flatten_array()) {
800 assert(basic_elem_type == T_VALUETYPE, "unexpected element basic type");
801 basic_elem_type = T_OBJECT;
802 }
803 }
804 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
805 element_size = type2aelembytes(basic_elem_type);
806 if (klass->is_value_array_klass()) {
807 // Flattened value type array
808 element_size = klass->as_value_array_klass()->element_byte_size();
809 }
810 }
811 }
812 //
813 // Process the safepoint uses
814 //
815 Unique_Node_List value_worklist;
816 while (safepoints.length() > 0) {
817 SafePointNode* sfpt = safepoints.pop();
818 Node* mem = sfpt->memory();
819 Node* ctl = sfpt->control();
820 assert(sfpt->jvms() != NULL, "missed JVMS");
821 // Fields of scalar objs are referenced only at the end
822 // of regular debuginfo at the last (youngest) JVMS.
823 // Record relative start index.
824 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff());
825 SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type,
826 #ifdef ASSERT
827 alloc,
828 #endif
829 first_ind, nfields);
830 sobj->init_req(0, C->root());
831 transform_later(sobj);
832
833 // Scan object's fields adding an input to the safepoint for each field.
834 for (int j = 0; j < nfields; j++) {
835 intptr_t offset;
836 ciField* field = NULL;
837 if (iklass != NULL) {
838 field = iklass->nonstatic_field_at(j);
839 offset = field->offset();
840 elem_type = field->type();
841 basic_elem_type = field->layout_type();
842 assert(!field->is_flattened(), "flattened value type fields should not have safepoint uses");
843 } else {
844 offset = array_base + j * (intptr_t)element_size;
845 }
846
847 const Type *field_type;
848 // The next code is taken from Parse::do_get_xxx().
849 if (basic_elem_type == T_OBJECT || basic_elem_type == T_ARRAY) {
850 if (!elem_type->is_loaded()) {
851 field_type = TypeInstPtr::BOTTOM;
852 } else if (field != NULL && field->is_static_constant()) {
853 // This can happen if the constant oop is non-perm.
854 ciObject* con = field->constant_value().as_object();
855 // Do not "join" in the previous type; it doesn't add value,
856 // and may yield a vacuous result if the field is of interface type.
857 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr();
858 assert(field_type != NULL, "field singleton type must be consistent");
859 } else {
860 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
861 }
862 if (UseCompressedOops) {
863 field_type = field_type->make_narrowoop();
864 basic_elem_type = T_NARROWOOP;
865 }
866 } else {
867 field_type = Type::get_const_basic_type(basic_elem_type);
868 }
869
870 Node* field_val = NULL;
871 const TypeOopPtr* field_addr_type = res_type->add_offset(offset)->isa_oopptr();
872 if (klass->is_value_array_klass()) {
873 ciValueKlass* vk = elem_type->as_value_klass();
874 assert(vk->flatten_array(), "must be flattened");
875 field_val = value_type_from_mem(mem, ctl, vk, field_addr_type->isa_aryptr(), 0, alloc);
876 } else {
877 field_val = value_from_mem(mem, ctl, basic_elem_type, field_type, field_addr_type, alloc);
878 }
879 if (field_val == NULL) {
880 // We weren't able to find a value for this field,
881 // give up on eliminating this allocation.
882
883 // Remove any extra entries we added to the safepoint.
884 uint last = sfpt->req() - 1;
885 for (int k = 0; k < j; k++) {
886 sfpt->del_req(last--);
887 }
888 _igvn._worklist.push(sfpt);
889 // rollback processed safepoints
890 while (safepoints_done.length() > 0) {
891 SafePointNode* sfpt_done = safepoints_done.pop();
892 // remove any extra entries we added to the safepoint
893 last = sfpt_done->req() - 1;
894 for (int k = 0; k < nfields; k++) {
895 sfpt_done->del_req(last--);
896 }
897 JVMState *jvms = sfpt_done->jvms();
898 jvms->set_endoff(sfpt_done->req());
899 // Now make a pass over the debug information replacing any references
900 // to SafePointScalarObjectNode with the allocated object.
901 int start = jvms->debug_start();
902 int end = jvms->debug_end();
903 for (int i = start; i < end; i++) {
904 if (sfpt_done->in(i)->is_SafePointScalarObject()) {
905 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject();
906 if (scobj->first_index(jvms) == sfpt_done->req() &&
907 scobj->n_fields() == (uint)nfields) {
908 assert(scobj->alloc() == alloc, "sanity");
909 sfpt_done->set_req(i, res);
910 }
911 }
912 }
913 _igvn._worklist.push(sfpt_done);
914 }
915 #ifndef PRODUCT
916 if (PrintEliminateAllocations) {
917 if (field != NULL) {
918 tty->print("=== At SafePoint node %d can't find value of Field: ",
919 sfpt->_idx);
920 field->print();
921 int field_idx = C->get_alias_index(field_addr_type);
922 tty->print(" (alias_idx=%d)", field_idx);
923 } else { // Array's element
924 tty->print("=== At SafePoint node %d can't find value of array element [%d]",
925 sfpt->_idx, j);
926 }
927 tty->print(", which prevents elimination of: ");
928 if (res == NULL)
929 alloc->dump();
930 else
931 res->dump();
932 }
933 #endif
934 return false;
935 }
936 if (UseCompressedOops && field_type->isa_narrowoop()) {
937 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
938 // to be able scalar replace the allocation.
939 if (field_val->is_EncodeP()) {
940 field_val = field_val->in(1);
941 } else {
942 field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type()));
943 }
944 } else if (field_val->is_ValueType()) {
945 // Keep track of value types to scalarize them later
946 value_worklist.push(field_val);
947 }
948 sfpt->add_req(field_val);
949 }
950 JVMState *jvms = sfpt->jvms();
951 jvms->set_endoff(sfpt->req());
952 // Now make a pass over the debug information replacing any references
953 // to the allocated object with "sobj"
954 int start = jvms->debug_start();
955 int end = jvms->debug_end();
956 sfpt->replace_edges_in_range(res, sobj, start, end);
957 _igvn._worklist.push(sfpt);
958 safepoints_done.append_if_missing(sfpt); // keep it for rollback
959 }
960 // Scalarize value types that were added to the safepoint
961 for (uint i = 0; i < value_worklist.size(); ++i) {
962 Node* vt = value_worklist.at(i);
963 vt->as_ValueType()->make_scalar_in_safepoints(C->root(), &_igvn);
964 }
965 return true;
966 }
967
968 static void disconnect_projections(MultiNode* n, PhaseIterGVN& igvn) {
969 Node* ctl_proj = n->proj_out_or_null(TypeFunc::Control);
970 Node* mem_proj = n->proj_out_or_null(TypeFunc::Memory);
971 if (ctl_proj != NULL) {
972 igvn.replace_node(ctl_proj, n->in(0));
973 }
974 if (mem_proj != NULL) {
975 igvn.replace_node(mem_proj, n->in(TypeFunc::Memory));
976 }
977 }
978
979 // Process users of eliminated allocation.
980 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc) {
981 Node* res = alloc->result_cast();
982 if (res != NULL) {
983 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
984 Node *use = res->last_out(j);
985 uint oc1 = res->outcnt();
986
987 if (use->is_AddP()) {
988 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) {
989 Node *n = use->last_out(k);
990 uint oc2 = use->outcnt();
991 if (n->is_Store()) {
992 #ifdef ASSERT
993 // Verify that there is no dependent MemBarVolatile nodes,
994 // they should be removed during IGVN, see MemBarNode::Ideal().
995 for (DUIterator_Fast pmax, p = n->fast_outs(pmax);
996 p < pmax; p++) {
997 Node* mb = n->fast_out(p);
998 assert(mb->is_Initialize() || !mb->is_MemBar() ||
999 mb->req() <= MemBarNode::Precedent ||
1000 mb->in(MemBarNode::Precedent) != n,
1001 "MemBarVolatile should be eliminated for non-escaping object");
1002 }
1003 #endif
1004 _igvn.replace_node(n, n->in(MemNode::Memory));
1005 } else if (n->is_ArrayCopy()) {
1006 // Disconnect ArrayCopy node
1007 ArrayCopyNode* ac = n->as_ArrayCopy();
1008 assert(ac->is_clonebasic(), "unexpected array copy kind");
1009 Node* membar_after = ac->proj_out(TypeFunc::Control)->unique_ctrl_out();
1010 disconnect_projections(ac, _igvn);
1011 assert(alloc->in(0)->is_Proj() && alloc->in(0)->in(0)->Opcode() == Op_MemBarCPUOrder, "mem barrier expected before allocation");
1012 Node* membar_before = alloc->in(0)->in(0);
1013 disconnect_projections(membar_before->as_MemBar(), _igvn);
1014 if (membar_after->is_MemBar()) {
1015 disconnect_projections(membar_after->as_MemBar(), _igvn);
1016 }
1017 } else {
1018 eliminate_gc_barrier(n);
1019 }
1020 k -= (oc2 - use->outcnt());
1021 }
1022 } else if (use->is_ArrayCopy()) {
1023 // Disconnect ArrayCopy node
1024 ArrayCopyNode* ac = use->as_ArrayCopy();
1025 assert(ac->is_arraycopy_validated() ||
1026 ac->is_copyof_validated() ||
1027 ac->is_copyofrange_validated(), "unsupported");
1028 CallProjections* callprojs = ac->extract_projections(true);
1029
1030 _igvn.replace_node(callprojs->fallthrough_ioproj, ac->in(TypeFunc::I_O));
1031 _igvn.replace_node(callprojs->fallthrough_memproj, ac->in(TypeFunc::Memory));
1032 _igvn.replace_node(callprojs->fallthrough_catchproj, ac->in(TypeFunc::Control));
1033
1034 // Set control to top. IGVN will remove the remaining projections
1035 ac->set_req(0, top());
1036 ac->replace_edge(res, top());
1037
1038 // Disconnect src right away: it can help find new
1039 // opportunities for allocation elimination
1040 Node* src = ac->in(ArrayCopyNode::Src);
1041 ac->replace_edge(src, top());
1042 // src can be top at this point if src and dest of the
1043 // arraycopy were the same
1044 if (src->outcnt() == 0 && !src->is_top()) {
1045 _igvn.remove_dead_node(src);
1046 }
1047
1048 _igvn._worklist.push(ac);
1049 } else if (use->is_ValueType()) {
1050 assert(use->isa_ValueType()->get_oop() == res, "unexpected value type use");
1051 _igvn.rehash_node_delayed(use);
1052 use->isa_ValueType()->set_oop(_igvn.zerocon(T_VALUETYPE));
1053 } else {
1054 eliminate_gc_barrier(use);
1055 }
1056 j -= (oc1 - res->outcnt());
1057 }
1058 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
1059 _igvn.remove_dead_node(res);
1060 }
1061
1062 //
1063 // Process other users of allocation's projections
1064 //
1065 if (_resproj != NULL && _resproj->outcnt() != 0) {
1066 // First disconnect stores captured by Initialize node.
1067 // If Initialize node is eliminated first in the following code,
1068 // it will kill such stores and DUIterator_Last will assert.
1069 for (DUIterator_Fast jmax, j = _resproj->fast_outs(jmax); j < jmax; j++) {
1070 Node *use = _resproj->fast_out(j);
1071 if (use->is_AddP()) {
1072 // raw memory addresses used only by the initialization
1073 _igvn.replace_node(use, C->top());
1074 --j; --jmax;
1075 }
1076 }
1077 for (DUIterator_Last jmin, j = _resproj->last_outs(jmin); j >= jmin; ) {
1078 Node *use = _resproj->last_out(j);
1079 uint oc1 = _resproj->outcnt();
1080 if (use->is_Initialize()) {
1081 // Eliminate Initialize node.
1082 InitializeNode *init = use->as_Initialize();
1083 assert(init->outcnt() <= 2, "only a control and memory projection expected");
1084 Node *ctrl_proj = init->proj_out_or_null(TypeFunc::Control);
1085 if (ctrl_proj != NULL) {
1086 assert(init->in(TypeFunc::Control) == _fallthroughcatchproj, "allocation control projection");
1087 _igvn.replace_node(ctrl_proj, _fallthroughcatchproj);
1088 }
1089 Node *mem_proj = init->proj_out_or_null(TypeFunc::Memory);
1090 if (mem_proj != NULL) {
1091 Node *mem = init->in(TypeFunc::Memory);
1092 #ifdef ASSERT
1093 if (mem->is_MergeMem()) {
1094 assert(mem->in(TypeFunc::Memory) == _memproj_fallthrough, "allocation memory projection");
1095 } else {
1096 assert(mem == _memproj_fallthrough, "allocation memory projection");
1097 }
1098 #endif
1099 _igvn.replace_node(mem_proj, mem);
1100 }
1101 } else {
1102 assert(false, "only Initialize or AddP expected");
1103 }
1104 j -= (oc1 - _resproj->outcnt());
1105 }
1106 }
1107 if (_fallthroughcatchproj != NULL) {
1108 _igvn.replace_node(_fallthroughcatchproj, alloc->in(TypeFunc::Control));
1109 }
1110 if (_memproj_fallthrough != NULL) {
1111 _igvn.replace_node(_memproj_fallthrough, alloc->in(TypeFunc::Memory));
1112 }
1113 if (_memproj_catchall != NULL) {
1114 _igvn.replace_node(_memproj_catchall, C->top());
1115 }
1116 if (_ioproj_fallthrough != NULL) {
1117 _igvn.replace_node(_ioproj_fallthrough, alloc->in(TypeFunc::I_O));
1118 }
1119 if (_ioproj_catchall != NULL) {
1120 _igvn.replace_node(_ioproj_catchall, C->top());
1121 }
1122 if (_catchallcatchproj != NULL) {
1123 _igvn.replace_node(_catchallcatchproj, C->top());
1124 }
1125 }
1126
1127 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) {
1128 // Don't do scalar replacement if the frame can be popped by JVMTI:
1129 // if reallocation fails during deoptimization we'll pop all
1130 // interpreter frames for this compiled frame and that won't play
1131 // nice with JVMTI popframe.
1132 if (!EliminateAllocations || JvmtiExport::can_pop_frame() || !alloc->_is_non_escaping) {
1133 return false;
1134 }
1135 Node* klass = alloc->in(AllocateNode::KlassNode);
1136 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr();
1137 Node* res = alloc->result_cast();
1138 // Eliminate boxing allocations which are not used
1139 // regardless scalar replacable status.
1140 bool boxing_alloc = C->eliminate_boxing() &&
1141 tklass->klass()->is_instance_klass() &&
1142 tklass->klass()->as_instance_klass()->is_box_klass();
1143 if (!alloc->_is_scalar_replaceable && (!boxing_alloc || (res != NULL))) {
1144 return false;
1145 }
1146
1147 extract_call_projections(alloc);
1148
1149 GrowableArray <SafePointNode *> safepoints;
1150 if (!can_eliminate_allocation(alloc, safepoints)) {
1151 return false;
1152 }
1153
1154 if (!alloc->_is_scalar_replaceable) {
1155 assert(res == NULL, "sanity");
1156 // We can only eliminate allocation if all debug info references
1157 // are already replaced with SafePointScalarObject because
1158 // we can't search for a fields value without instance_id.
1159 if (safepoints.length() > 0) {
1160 return false;
1161 }
1162 }
1163
1164 if (!scalar_replacement(alloc, safepoints)) {
1165 return false;
1166 }
1167
1168 CompileLog* log = C->log();
1169 if (log != NULL) {
1170 log->head("eliminate_allocation type='%d'",
1171 log->identify(tklass->klass()));
1172 JVMState* p = alloc->jvms();
1173 while (p != NULL) {
1174 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1175 p = p->caller();
1176 }
1177 log->tail("eliminate_allocation");
1178 }
1179
1180 process_users_of_allocation(alloc);
1181
1182 #ifndef PRODUCT
1183 if (PrintEliminateAllocations) {
1184 if (alloc->is_AllocateArray())
1185 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1186 else
1187 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1188 }
1189 #endif
1190
1191 return true;
1192 }
1193
1194 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) {
1195 // EA should remove all uses of non-escaping boxing node.
1196 if (!C->eliminate_boxing() || boxing->proj_out_or_null(TypeFunc::Parms) != NULL) {
1197 return false;
1198 }
1199
1200 assert(boxing->result_cast() == NULL, "unexpected boxing node result");
1201
1202 extract_call_projections(boxing);
1203
1204 const TypeTuple* r = boxing->tf()->range_sig();
1205 assert(r->cnt() > TypeFunc::Parms, "sanity");
1206 const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr();
1207 assert(t != NULL, "sanity");
1208
1209 CompileLog* log = C->log();
1210 if (log != NULL) {
1211 log->head("eliminate_boxing type='%d'",
1212 log->identify(t->klass()));
1213 JVMState* p = boxing->jvms();
1214 while (p != NULL) {
1215 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1216 p = p->caller();
1217 }
1218 log->tail("eliminate_boxing");
1219 }
1220
1221 process_users_of_allocation(boxing);
1222
1223 #ifndef PRODUCT
1224 if (PrintEliminateAllocations) {
1225 tty->print("++++ Eliminated: %d ", boxing->_idx);
1226 boxing->method()->print_short_name(tty);
1227 tty->cr();
1228 }
1229 #endif
1230
1231 return true;
1232 }
1233
1234 //---------------------------set_eden_pointers-------------------------
1235 void PhaseMacroExpand::set_eden_pointers(Node* &eden_top_adr, Node* &eden_end_adr) {
1236 if (UseTLAB) { // Private allocation: load from TLS
1237 Node* thread = transform_later(new ThreadLocalNode());
1238 int tlab_top_offset = in_bytes(JavaThread::tlab_top_offset());
1239 int tlab_end_offset = in_bytes(JavaThread::tlab_end_offset());
1240 eden_top_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_top_offset);
1241 eden_end_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_end_offset);
1242 } else { // Shared allocation: load from globals
1243 CollectedHeap* ch = Universe::heap();
1244 address top_adr = (address)ch->top_addr();
1245 address end_adr = (address)ch->end_addr();
1246 eden_top_adr = makecon(TypeRawPtr::make(top_adr));
1247 eden_end_adr = basic_plus_adr(eden_top_adr, end_adr - top_adr);
1248 }
1249 }
1250
1251
1252 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) {
1253 Node* adr = basic_plus_adr(base, offset);
1254 const TypePtr* adr_type = adr->bottom_type()->is_ptr();
1255 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt, MemNode::unordered);
1256 transform_later(value);
1257 return value;
1258 }
1259
1260
1261 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) {
1262 Node* adr = basic_plus_adr(base, offset);
1263 mem = StoreNode::make(_igvn, ctl, mem, adr, NULL, value, bt, MemNode::unordered);
1264 transform_later(mem);
1265 return mem;
1266 }
1267
1268 //=============================================================================
1269 //
1270 // A L L O C A T I O N
1271 //
1272 // Allocation attempts to be fast in the case of frequent small objects.
1273 // It breaks down like this:
1274 //
1275 // 1) Size in doublewords is computed. This is a constant for objects and
1276 // variable for most arrays. Doubleword units are used to avoid size
1277 // overflow of huge doubleword arrays. We need doublewords in the end for
1278 // rounding.
1279 //
1280 // 2) Size is checked for being 'too large'. Too-large allocations will go
1281 // the slow path into the VM. The slow path can throw any required
1282 // exceptions, and does all the special checks for very large arrays. The
1283 // size test can constant-fold away for objects. For objects with
1284 // finalizers it constant-folds the otherway: you always go slow with
1285 // finalizers.
1286 //
1287 // 3) If NOT using TLABs, this is the contended loop-back point.
1288 // Load-Locked the heap top. If using TLABs normal-load the heap top.
1289 //
1290 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route.
1291 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish
1292 // "size*8" we always enter the VM, where "largish" is a constant picked small
1293 // enough that there's always space between the eden max and 4Gig (old space is
1294 // there so it's quite large) and large enough that the cost of entering the VM
1295 // is dwarfed by the cost to initialize the space.
1296 //
1297 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back
1298 // down. If contended, repeat at step 3. If using TLABs normal-store
1299 // adjusted heap top back down; there is no contention.
1300 //
1301 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark
1302 // fields.
1303 //
1304 // 7) Merge with the slow-path; cast the raw memory pointer to the correct
1305 // oop flavor.
1306 //
1307 //=============================================================================
1308 // FastAllocateSizeLimit value is in DOUBLEWORDS.
1309 // Allocations bigger than this always go the slow route.
1310 // This value must be small enough that allocation attempts that need to
1311 // trigger exceptions go the slow route. Also, it must be small enough so
1312 // that heap_top + size_in_bytes does not wrap around the 4Gig limit.
1313 //=============================================================================j//
1314 // %%% Here is an old comment from parseHelper.cpp; is it outdated?
1315 // The allocator will coalesce int->oop copies away. See comment in
1316 // coalesce.cpp about how this works. It depends critically on the exact
1317 // code shape produced here, so if you are changing this code shape
1318 // make sure the GC info for the heap-top is correct in and around the
1319 // slow-path call.
1320 //
1321
1322 void PhaseMacroExpand::expand_allocate_common(
1323 AllocateNode* alloc, // allocation node to be expanded
1324 Node* length, // array length for an array allocation
1325 const TypeFunc* slow_call_type, // Type of slow call
1326 address slow_call_address // Address of slow call
1327 )
1328 {
1329
1330 Node* ctrl = alloc->in(TypeFunc::Control);
1331 Node* mem = alloc->in(TypeFunc::Memory);
1332 Node* i_o = alloc->in(TypeFunc::I_O);
1333 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize);
1334 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1335 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest);
1336
1337 assert(ctrl != NULL, "must have control");
1338 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results.
1339 // they will not be used if "always_slow" is set
1340 enum { slow_result_path = 1, fast_result_path = 2 };
1341 Node *result_region = NULL;
1342 Node *result_phi_rawmem = NULL;
1343 Node *result_phi_rawoop = NULL;
1344 Node *result_phi_i_o = NULL;
1345
1346 // The initial slow comparison is a size check, the comparison
1347 // we want to do is a BoolTest::gt
1348 bool always_slow = false;
1349 int tv = _igvn.find_int_con(initial_slow_test, -1);
1350 if (tv >= 0) {
1351 always_slow = (tv == 1);
1352 initial_slow_test = NULL;
1353 } else {
1354 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn);
1355 }
1356
1357 if (C->env()->dtrace_alloc_probes() ||
1358 (!UseTLAB && !Universe::heap()->supports_inline_contig_alloc())) {
1359 // Force slow-path allocation
1360 always_slow = true;
1361 initial_slow_test = NULL;
1362 }
1363
1364 Node *slow_region = NULL;
1365 Node *toobig_false = ctrl;
1366
1367 assert (initial_slow_test == NULL || !always_slow, "arguments must be consistent");
1368 // generate the initial test if necessary
1369 if (initial_slow_test != NULL ) {
1370 if (slow_region == NULL) {
1371 slow_region = new RegionNode(1);
1372 }
1373 // Now make the initial failure test. Usually a too-big test but
1374 // might be a TRUE for finalizers or a fancy class check for
1375 // newInstance0.
1376 IfNode* toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
1377 transform_later(toobig_iff);
1378 // Plug the failing-too-big test into the slow-path region
1379 Node* toobig_true = new IfTrueNode(toobig_iff);
1380 transform_later(toobig_true);
1381 slow_region ->add_req(toobig_true);
1382 toobig_false = new IfFalseNode(toobig_iff);
1383 transform_later(toobig_false);
1384 } else { // No initial test, just fall into next case
1385 toobig_false = ctrl;
1386 }
1387
1388 Node *slow_mem = mem; // save the current memory state for slow path
1389 // generate the fast allocation code unless we know that the initial test will always go slow
1390 if (!always_slow) {
1391 // Fast path modifies only raw memory.
1392 if (mem->is_MergeMem()) {
1393 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1394 }
1395
1396 Node* eden_top_adr;
1397 Node* eden_end_adr;
1398
1399 set_eden_pointers(eden_top_adr, eden_end_adr);
1400
1401 // Load Eden::end. Loop invariant and hoisted.
1402 //
1403 // Note: We set the control input on "eden_end" and "old_eden_top" when using
1404 // a TLAB to work around a bug where these values were being moved across
1405 // a safepoint. These are not oops, so they cannot be include in the oop
1406 // map, but they can be changed by a GC. The proper way to fix this would
1407 // be to set the raw memory state when generating a SafepointNode. However
1408 // this will require extensive changes to the loop optimization in order to
1409 // prevent a degradation of the optimization.
1410 // See comment in memnode.hpp, around line 227 in class LoadPNode.
1411 Node *eden_end = make_load(ctrl, mem, eden_end_adr, 0, TypeRawPtr::BOTTOM, T_ADDRESS);
1412
1413 // allocate the Region and Phi nodes for the result
1414 result_region = new RegionNode(3);
1415 result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1416 result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM);
1417 result_phi_i_o = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch
1418
1419 // We need a Region for the loop-back contended case.
1420 enum { fall_in_path = 1, contended_loopback_path = 2 };
1421 Node *contended_region;
1422 Node *contended_phi_rawmem;
1423 if (UseTLAB) {
1424 contended_region = toobig_false;
1425 contended_phi_rawmem = mem;
1426 } else {
1427 contended_region = new RegionNode(3);
1428 contended_phi_rawmem = new PhiNode(contended_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1429 // Now handle the passing-too-big test. We fall into the contended
1430 // loop-back merge point.
1431 contended_region ->init_req(fall_in_path, toobig_false);
1432 contended_phi_rawmem->init_req(fall_in_path, mem);
1433 transform_later(contended_region);
1434 transform_later(contended_phi_rawmem);
1435 }
1436
1437 // Load(-locked) the heap top.
1438 // See note above concerning the control input when using a TLAB
1439 Node *old_eden_top = UseTLAB
1440 ? new LoadPNode (ctrl, contended_phi_rawmem, eden_top_adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, MemNode::unordered)
1441 : new LoadPLockedNode(contended_region, contended_phi_rawmem, eden_top_adr, MemNode::acquire);
1442
1443 transform_later(old_eden_top);
1444 // Add to heap top to get a new heap top
1445 Node *new_eden_top = new AddPNode(top(), old_eden_top, size_in_bytes);
1446 transform_later(new_eden_top);
1447 // Check for needing a GC; compare against heap end
1448 Node *needgc_cmp = new CmpPNode(new_eden_top, eden_end);
1449 transform_later(needgc_cmp);
1450 Node *needgc_bol = new BoolNode(needgc_cmp, BoolTest::ge);
1451 transform_later(needgc_bol);
1452 IfNode *needgc_iff = new IfNode(contended_region, needgc_bol, PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN);
1453 transform_later(needgc_iff);
1454
1455 // Plug the failing-heap-space-need-gc test into the slow-path region
1456 Node *needgc_true = new IfTrueNode(needgc_iff);
1457 transform_later(needgc_true);
1458 if (slow_region != NULL) {
1459 slow_region->add_req(needgc_true);
1460 // This completes all paths into the slow merge point
1461 transform_later(slow_region);
1462 } else {
1463 // Just fall from the need-GC path straight into the VM call.
1464 slow_region = needgc_true;
1465 }
1466 // No need for a GC. Setup for the Store-Conditional
1467 Node *needgc_false = new IfFalseNode(needgc_iff);
1468 transform_later(needgc_false);
1469
1470 // Grab regular I/O before optional prefetch may change it.
1471 // Slow-path does no I/O so just set it to the original I/O.
1472 result_phi_i_o->init_req(slow_result_path, i_o);
1473
1474 i_o = prefetch_allocation(i_o, needgc_false, contended_phi_rawmem,
1475 old_eden_top, new_eden_top, length);
1476
1477 // Name successful fast-path variables
1478 Node* fast_oop = old_eden_top;
1479 Node* fast_oop_ctrl;
1480 Node* fast_oop_rawmem;
1481
1482 // Store (-conditional) the modified eden top back down.
1483 // StorePConditional produces flags for a test PLUS a modified raw
1484 // memory state.
1485 if (UseTLAB) {
1486 Node* store_eden_top =
1487 new StorePNode(needgc_false, contended_phi_rawmem, eden_top_adr,
1488 TypeRawPtr::BOTTOM, new_eden_top, MemNode::unordered);
1489 transform_later(store_eden_top);
1490 fast_oop_ctrl = needgc_false; // No contention, so this is the fast path
1491 fast_oop_rawmem = store_eden_top;
1492 } else {
1493 Node* store_eden_top =
1494 new StorePConditionalNode(needgc_false, contended_phi_rawmem, eden_top_adr,
1495 new_eden_top, fast_oop/*old_eden_top*/);
1496 transform_later(store_eden_top);
1497 Node *contention_check = new BoolNode(store_eden_top, BoolTest::ne);
1498 transform_later(contention_check);
1499 store_eden_top = new SCMemProjNode(store_eden_top);
1500 transform_later(store_eden_top);
1501
1502 // If not using TLABs, check to see if there was contention.
1503 IfNode *contention_iff = new IfNode (needgc_false, contention_check, PROB_MIN, COUNT_UNKNOWN);
1504 transform_later(contention_iff);
1505 Node *contention_true = new IfTrueNode(contention_iff);
1506 transform_later(contention_true);
1507 // If contention, loopback and try again.
1508 contended_region->init_req(contended_loopback_path, contention_true);
1509 contended_phi_rawmem->init_req(contended_loopback_path, store_eden_top);
1510
1511 // Fast-path succeeded with no contention!
1512 Node *contention_false = new IfFalseNode(contention_iff);
1513 transform_later(contention_false);
1514 fast_oop_ctrl = contention_false;
1515
1516 // Bump total allocated bytes for this thread
1517 Node* thread = new ThreadLocalNode();
1518 transform_later(thread);
1519 Node* alloc_bytes_adr = basic_plus_adr(top()/*not oop*/, thread,
1520 in_bytes(JavaThread::allocated_bytes_offset()));
1521 Node* alloc_bytes = make_load(fast_oop_ctrl, store_eden_top, alloc_bytes_adr,
1522 0, TypeLong::LONG, T_LONG);
1523 #ifdef _LP64
1524 Node* alloc_size = size_in_bytes;
1525 #else
1526 Node* alloc_size = new ConvI2LNode(size_in_bytes);
1527 transform_later(alloc_size);
1528 #endif
1529 Node* new_alloc_bytes = new AddLNode(alloc_bytes, alloc_size);
1530 transform_later(new_alloc_bytes);
1531 fast_oop_rawmem = make_store(fast_oop_ctrl, store_eden_top, alloc_bytes_adr,
1532 0, new_alloc_bytes, T_LONG);
1533 }
1534
1535 InitializeNode* init = alloc->initialization();
1536 fast_oop_rawmem = initialize_object(alloc,
1537 fast_oop_ctrl, fast_oop_rawmem, fast_oop,
1538 klass_node, length, size_in_bytes);
1539
1540 // If initialization is performed by an array copy, any required
1541 // MemBarStoreStore was already added. If the object does not
1542 // escape no need for a MemBarStoreStore. If the object does not
1543 // escape in its initializer and memory barrier (MemBarStoreStore or
1544 // stronger) is already added at exit of initializer, also no need
1545 // for a MemBarStoreStore. Otherwise we need a MemBarStoreStore
1546 // so that stores that initialize this object can't be reordered
1547 // with a subsequent store that makes this object accessible by
1548 // other threads.
1549 // Other threads include java threads and JVM internal threads
1550 // (for example concurrent GC threads). Current concurrent GC
1551 // implementation: CMS and G1 will not scan newly created object,
1552 // so it's safe to skip storestore barrier when allocation does
1553 // not escape.
1554 if (!alloc->does_not_escape_thread() &&
1555 !alloc->is_allocation_MemBar_redundant() &&
1556 (init == NULL || !init->is_complete_with_arraycopy())) {
1557 if (init == NULL || init->req() < InitializeNode::RawStores) {
1558 // No InitializeNode or no stores captured by zeroing
1559 // elimination. Simply add the MemBarStoreStore after object
1560 // initialization.
1561 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1562 transform_later(mb);
1563
1564 mb->init_req(TypeFunc::Memory, fast_oop_rawmem);
1565 mb->init_req(TypeFunc::Control, fast_oop_ctrl);
1566 fast_oop_ctrl = new ProjNode(mb,TypeFunc::Control);
1567 transform_later(fast_oop_ctrl);
1568 fast_oop_rawmem = new ProjNode(mb,TypeFunc::Memory);
1569 transform_later(fast_oop_rawmem);
1570 } else {
1571 // Add the MemBarStoreStore after the InitializeNode so that
1572 // all stores performing the initialization that were moved
1573 // before the InitializeNode happen before the storestore
1574 // barrier.
1575
1576 Node* init_ctrl = init->proj_out_or_null(TypeFunc::Control);
1577 Node* init_mem = init->proj_out_or_null(TypeFunc::Memory);
1578
1579 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1580 transform_later(mb);
1581
1582 Node* ctrl = new ProjNode(init,TypeFunc::Control);
1583 transform_later(ctrl);
1584 Node* mem = new ProjNode(init,TypeFunc::Memory);
1585 transform_later(mem);
1586
1587 // The MemBarStoreStore depends on control and memory coming
1588 // from the InitializeNode
1589 mb->init_req(TypeFunc::Memory, mem);
1590 mb->init_req(TypeFunc::Control, ctrl);
1591
1592 ctrl = new ProjNode(mb,TypeFunc::Control);
1593 transform_later(ctrl);
1594 mem = new ProjNode(mb,TypeFunc::Memory);
1595 transform_later(mem);
1596
1597 // All nodes that depended on the InitializeNode for control
1598 // and memory must now depend on the MemBarNode that itself
1599 // depends on the InitializeNode
1600 if (init_ctrl != NULL) {
1601 _igvn.replace_node(init_ctrl, ctrl);
1602 }
1603 if (init_mem != NULL) {
1604 _igvn.replace_node(init_mem, mem);
1605 }
1606 }
1607 }
1608
1609 if (C->env()->dtrace_extended_probes()) {
1610 // Slow-path call
1611 int size = TypeFunc::Parms + 2;
1612 CallLeafNode *call = new CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(),
1613 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_object_alloc_base),
1614 "dtrace_object_alloc",
1615 TypeRawPtr::BOTTOM);
1616
1617 // Get base of thread-local storage area
1618 Node* thread = new ThreadLocalNode();
1619 transform_later(thread);
1620
1621 call->init_req(TypeFunc::Parms+0, thread);
1622 call->init_req(TypeFunc::Parms+1, fast_oop);
1623 call->init_req(TypeFunc::Control, fast_oop_ctrl);
1624 call->init_req(TypeFunc::I_O , top()); // does no i/o
1625 call->init_req(TypeFunc::Memory , fast_oop_rawmem);
1626 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1627 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1628 transform_later(call);
1629 fast_oop_ctrl = new ProjNode(call,TypeFunc::Control);
1630 transform_later(fast_oop_ctrl);
1631 fast_oop_rawmem = new ProjNode(call,TypeFunc::Memory);
1632 transform_later(fast_oop_rawmem);
1633 }
1634
1635 // Plug in the successful fast-path into the result merge point
1636 result_region ->init_req(fast_result_path, fast_oop_ctrl);
1637 result_phi_rawoop->init_req(fast_result_path, fast_oop);
1638 result_phi_i_o ->init_req(fast_result_path, i_o);
1639 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem);
1640 } else {
1641 slow_region = ctrl;
1642 result_phi_i_o = i_o; // Rename it to use in the following code.
1643 }
1644
1645 // Generate slow-path call
1646 CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address,
1647 OptoRuntime::stub_name(slow_call_address),
1648 alloc->jvms()->bci(),
1649 TypePtr::BOTTOM);
1650 call->init_req( TypeFunc::Control, slow_region );
1651 call->init_req( TypeFunc::I_O , top() ) ; // does no i/o
1652 call->init_req( TypeFunc::Memory , slow_mem ); // may gc ptrs
1653 call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) );
1654 call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) );
1655
1656 call->init_req(TypeFunc::Parms+0, klass_node);
1657 if (length != NULL) {
1658 call->init_req(TypeFunc::Parms+1, length);
1659 }
1660
1661 // Copy debug information and adjust JVMState information, then replace
1662 // allocate node with the call
1663 copy_call_debug_info((CallNode *) alloc, call);
1664 if (!always_slow) {
1665 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1666 } else {
1667 // Hook i_o projection to avoid its elimination during allocation
1668 // replacement (when only a slow call is generated).
1669 call->set_req(TypeFunc::I_O, result_phi_i_o);
1670 }
1671 _igvn.replace_node(alloc, call);
1672 transform_later(call);
1673
1674 // Identify the output projections from the allocate node and
1675 // adjust any references to them.
1676 // The control and io projections look like:
1677 //
1678 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl)
1679 // Allocate Catch
1680 // ^---Proj(io) <-------+ ^---CatchProj(io)
1681 //
1682 // We are interested in the CatchProj nodes.
1683 //
1684 extract_call_projections(call);
1685
1686 // An allocate node has separate memory projections for the uses on
1687 // the control and i_o paths. Replace the control memory projection with
1688 // result_phi_rawmem (unless we are only generating a slow call when
1689 // both memory projections are combined)
1690 if (!always_slow && _memproj_fallthrough != NULL) {
1691 for (DUIterator_Fast imax, i = _memproj_fallthrough->fast_outs(imax); i < imax; i++) {
1692 Node *use = _memproj_fallthrough->fast_out(i);
1693 _igvn.rehash_node_delayed(use);
1694 imax -= replace_input(use, _memproj_fallthrough, result_phi_rawmem);
1695 // back up iterator
1696 --i;
1697 }
1698 }
1699 // Now change uses of _memproj_catchall to use _memproj_fallthrough and delete
1700 // _memproj_catchall so we end up with a call that has only 1 memory projection.
1701 if (_memproj_catchall != NULL ) {
1702 if (_memproj_fallthrough == NULL) {
1703 _memproj_fallthrough = new ProjNode(call, TypeFunc::Memory);
1704 transform_later(_memproj_fallthrough);
1705 }
1706 for (DUIterator_Fast imax, i = _memproj_catchall->fast_outs(imax); i < imax; i++) {
1707 Node *use = _memproj_catchall->fast_out(i);
1708 _igvn.rehash_node_delayed(use);
1709 imax -= replace_input(use, _memproj_catchall, _memproj_fallthrough);
1710 // back up iterator
1711 --i;
1712 }
1713 assert(_memproj_catchall->outcnt() == 0, "all uses must be deleted");
1714 _igvn.remove_dead_node(_memproj_catchall);
1715 }
1716
1717 // An allocate node has separate i_o projections for the uses on the control
1718 // and i_o paths. Always replace the control i_o projection with result i_o
1719 // otherwise incoming i_o become dead when only a slow call is generated
1720 // (it is different from memory projections where both projections are
1721 // combined in such case).
1722 if (_ioproj_fallthrough != NULL) {
1723 for (DUIterator_Fast imax, i = _ioproj_fallthrough->fast_outs(imax); i < imax; i++) {
1724 Node *use = _ioproj_fallthrough->fast_out(i);
1725 _igvn.rehash_node_delayed(use);
1726 imax -= replace_input(use, _ioproj_fallthrough, result_phi_i_o);
1727 // back up iterator
1728 --i;
1729 }
1730 }
1731 // Now change uses of _ioproj_catchall to use _ioproj_fallthrough and delete
1732 // _ioproj_catchall so we end up with a call that has only 1 i_o projection.
1733 if (_ioproj_catchall != NULL ) {
1734 if (_ioproj_fallthrough == NULL) {
1735 _ioproj_fallthrough = new ProjNode(call, TypeFunc::I_O);
1736 transform_later(_ioproj_fallthrough);
1737 }
1738 for (DUIterator_Fast imax, i = _ioproj_catchall->fast_outs(imax); i < imax; i++) {
1739 Node *use = _ioproj_catchall->fast_out(i);
1740 _igvn.rehash_node_delayed(use);
1741 imax -= replace_input(use, _ioproj_catchall, _ioproj_fallthrough);
1742 // back up iterator
1743 --i;
1744 }
1745 assert(_ioproj_catchall->outcnt() == 0, "all uses must be deleted");
1746 _igvn.remove_dead_node(_ioproj_catchall);
1747 }
1748
1749 // if we generated only a slow call, we are done
1750 if (always_slow) {
1751 // Now we can unhook i_o.
1752 if (result_phi_i_o->outcnt() > 1) {
1753 call->set_req(TypeFunc::I_O, top());
1754 } else {
1755 assert(result_phi_i_o->unique_ctrl_out() == call, "");
1756 // Case of new array with negative size known during compilation.
1757 // AllocateArrayNode::Ideal() optimization disconnect unreachable
1758 // following code since call to runtime will throw exception.
1759 // As result there will be no users of i_o after the call.
1760 // Leave i_o attached to this call to avoid problems in preceding graph.
1761 }
1762 return;
1763 }
1764
1765
1766 if (_fallthroughcatchproj != NULL) {
1767 ctrl = _fallthroughcatchproj->clone();
1768 transform_later(ctrl);
1769 _igvn.replace_node(_fallthroughcatchproj, result_region);
1770 } else {
1771 ctrl = top();
1772 }
1773 Node *slow_result;
1774 if (_resproj == NULL) {
1775 // no uses of the allocation result
1776 slow_result = top();
1777 } else {
1778 slow_result = _resproj->clone();
1779 transform_later(slow_result);
1780 _igvn.replace_node(_resproj, result_phi_rawoop);
1781 }
1782
1783 // Plug slow-path into result merge point
1784 result_region ->init_req( slow_result_path, ctrl );
1785 result_phi_rawoop->init_req( slow_result_path, slow_result);
1786 result_phi_rawmem->init_req( slow_result_path, _memproj_fallthrough );
1787 transform_later(result_region);
1788 transform_later(result_phi_rawoop);
1789 transform_later(result_phi_rawmem);
1790 transform_later(result_phi_i_o);
1791 // This completes all paths into the result merge point
1792 }
1793
1794
1795 // Helper for PhaseMacroExpand::expand_allocate_common.
1796 // Initializes the newly-allocated storage.
1797 Node* PhaseMacroExpand::initialize_object(AllocateNode* alloc,
1798 Node* control, Node* rawmem, Node* object,
1799 Node* klass_node, Node* length,
1800 Node* size_in_bytes) {
1801 InitializeNode* init = alloc->initialization();
1802 // Store the klass & mark bits
1803 Node* mark_node = NULL;
1804 // For now only enable fast locking for non-array types
1805 if ((EnableValhalla || UseBiasedLocking) && length == NULL) {
1806 mark_node = make_load(control, rawmem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
1807 } else {
1808 mark_node = makecon(TypeRawPtr::make((address)markOopDesc::prototype()));
1809 }
1810 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS);
1811
1812 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
1813 int header_size = alloc->minimum_header_size(); // conservatively small
1814
1815 // Array length
1816 if (length != NULL) { // Arrays need length field
1817 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
1818 // conservatively small header size:
1819 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1820 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
1821 if (k->is_array_klass()) // we know the exact header size in most cases:
1822 header_size = Klass::layout_helper_header_size(k->layout_helper());
1823 }
1824
1825 // Clear the object body, if necessary.
1826 if (init == NULL) {
1827 // The init has somehow disappeared; be cautious and clear everything.
1828 //
1829 // This can happen if a node is allocated but an uncommon trap occurs
1830 // immediately. In this case, the Initialize gets associated with the
1831 // trap, and may be placed in a different (outer) loop, if the Allocate
1832 // is in a loop. If (this is rare) the inner loop gets unrolled, then
1833 // there can be two Allocates to one Initialize. The answer in all these
1834 // edge cases is safety first. It is always safe to clear immediately
1835 // within an Allocate, and then (maybe or maybe not) clear some more later.
1836 if (!(UseTLAB && ZeroTLAB)) {
1837 rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
1838 alloc->in(AllocateNode::DefaultValue),
1839 alloc->in(AllocateNode::RawDefaultValue),
1840 header_size, size_in_bytes,
1841 &_igvn);
1842 }
1843 } else {
1844 if (!init->is_complete()) {
1845 // Try to win by zeroing only what the init does not store.
1846 // We can also try to do some peephole optimizations,
1847 // such as combining some adjacent subword stores.
1848 rawmem = init->complete_stores(control, rawmem, object,
1849 header_size, size_in_bytes, &_igvn);
1850 }
1851 // We have no more use for this link, since the AllocateNode goes away:
1852 init->set_req(InitializeNode::RawAddress, top());
1853 // (If we keep the link, it just confuses the register allocator,
1854 // who thinks he sees a real use of the address by the membar.)
1855 }
1856
1857 return rawmem;
1858 }
1859
1860 // Generate prefetch instructions for next allocations.
1861 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false,
1862 Node*& contended_phi_rawmem,
1863 Node* old_eden_top, Node* new_eden_top,
1864 Node* length) {
1865 enum { fall_in_path = 1, pf_path = 2 };
1866 if( UseTLAB && AllocatePrefetchStyle == 2 ) {
1867 // Generate prefetch allocation with watermark check.
1868 // As an allocation hits the watermark, we will prefetch starting
1869 // at a "distance" away from watermark.
1870
1871 Node *pf_region = new RegionNode(3);
1872 Node *pf_phi_rawmem = new PhiNode( pf_region, Type::MEMORY,
1873 TypeRawPtr::BOTTOM );
1874 // I/O is used for Prefetch
1875 Node *pf_phi_abio = new PhiNode( pf_region, Type::ABIO );
1876
1877 Node *thread = new ThreadLocalNode();
1878 transform_later(thread);
1879
1880 Node *eden_pf_adr = new AddPNode( top()/*not oop*/, thread,
1881 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) );
1882 transform_later(eden_pf_adr);
1883
1884 Node *old_pf_wm = new LoadPNode(needgc_false,
1885 contended_phi_rawmem, eden_pf_adr,
1886 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM,
1887 MemNode::unordered);
1888 transform_later(old_pf_wm);
1889
1890 // check against new_eden_top
1891 Node *need_pf_cmp = new CmpPNode( new_eden_top, old_pf_wm );
1892 transform_later(need_pf_cmp);
1893 Node *need_pf_bol = new BoolNode( need_pf_cmp, BoolTest::ge );
1894 transform_later(need_pf_bol);
1895 IfNode *need_pf_iff = new IfNode( needgc_false, need_pf_bol,
1896 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN );
1897 transform_later(need_pf_iff);
1898
1899 // true node, add prefetchdistance
1900 Node *need_pf_true = new IfTrueNode( need_pf_iff );
1901 transform_later(need_pf_true);
1902
1903 Node *need_pf_false = new IfFalseNode( need_pf_iff );
1904 transform_later(need_pf_false);
1905
1906 Node *new_pf_wmt = new AddPNode( top(), old_pf_wm,
1907 _igvn.MakeConX(AllocatePrefetchDistance) );
1908 transform_later(new_pf_wmt );
1909 new_pf_wmt->set_req(0, need_pf_true);
1910
1911 Node *store_new_wmt = new StorePNode(need_pf_true,
1912 contended_phi_rawmem, eden_pf_adr,
1913 TypeRawPtr::BOTTOM, new_pf_wmt,
1914 MemNode::unordered);
1915 transform_later(store_new_wmt);
1916
1917 // adding prefetches
1918 pf_phi_abio->init_req( fall_in_path, i_o );
1919
1920 Node *prefetch_adr;
1921 Node *prefetch;
1922 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1923 uint step_size = AllocatePrefetchStepSize;
1924 uint distance = 0;
1925
1926 for ( uint i = 0; i < lines; i++ ) {
1927 prefetch_adr = new AddPNode( old_pf_wm, new_pf_wmt,
1928 _igvn.MakeConX(distance) );
1929 transform_later(prefetch_adr);
1930 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
1931 transform_later(prefetch);
1932 distance += step_size;
1933 i_o = prefetch;
1934 }
1935 pf_phi_abio->set_req( pf_path, i_o );
1936
1937 pf_region->init_req( fall_in_path, need_pf_false );
1938 pf_region->init_req( pf_path, need_pf_true );
1939
1940 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem );
1941 pf_phi_rawmem->init_req( pf_path, store_new_wmt );
1942
1943 transform_later(pf_region);
1944 transform_later(pf_phi_rawmem);
1945 transform_later(pf_phi_abio);
1946
1947 needgc_false = pf_region;
1948 contended_phi_rawmem = pf_phi_rawmem;
1949 i_o = pf_phi_abio;
1950 } else if( UseTLAB && AllocatePrefetchStyle == 3 ) {
1951 // Insert a prefetch instruction for each allocation.
1952 // This code is used to generate 1 prefetch instruction per cache line.
1953
1954 // Generate several prefetch instructions.
1955 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1956 uint step_size = AllocatePrefetchStepSize;
1957 uint distance = AllocatePrefetchDistance;
1958
1959 // Next cache address.
1960 Node *cache_adr = new AddPNode(old_eden_top, old_eden_top,
1961 _igvn.MakeConX(step_size + distance));
1962 transform_later(cache_adr);
1963 cache_adr = new CastP2XNode(needgc_false, cache_adr);
1964 transform_later(cache_adr);
1965 // Address is aligned to execute prefetch to the beginning of cache line size
1966 // (it is important when BIS instruction is used on SPARC as prefetch).
1967 Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1));
1968 cache_adr = new AndXNode(cache_adr, mask);
1969 transform_later(cache_adr);
1970 cache_adr = new CastX2PNode(cache_adr);
1971 transform_later(cache_adr);
1972
1973 // Prefetch
1974 Node *prefetch = new PrefetchAllocationNode( contended_phi_rawmem, cache_adr );
1975 prefetch->set_req(0, needgc_false);
1976 transform_later(prefetch);
1977 contended_phi_rawmem = prefetch;
1978 Node *prefetch_adr;
1979 distance = step_size;
1980 for ( uint i = 1; i < lines; i++ ) {
1981 prefetch_adr = new AddPNode( cache_adr, cache_adr,
1982 _igvn.MakeConX(distance) );
1983 transform_later(prefetch_adr);
1984 prefetch = new PrefetchAllocationNode( contended_phi_rawmem, prefetch_adr );
1985 transform_later(prefetch);
1986 distance += step_size;
1987 contended_phi_rawmem = prefetch;
1988 }
1989 } else if( AllocatePrefetchStyle > 0 ) {
1990 // Insert a prefetch for each allocation only on the fast-path
1991 Node *prefetch_adr;
1992 Node *prefetch;
1993 // Generate several prefetch instructions.
1994 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1995 uint step_size = AllocatePrefetchStepSize;
1996 uint distance = AllocatePrefetchDistance;
1997 for ( uint i = 0; i < lines; i++ ) {
1998 prefetch_adr = new AddPNode( old_eden_top, new_eden_top,
1999 _igvn.MakeConX(distance) );
2000 transform_later(prefetch_adr);
2001 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
2002 // Do not let it float too high, since if eden_top == eden_end,
2003 // both might be null.
2004 if( i == 0 ) { // Set control for first prefetch, next follows it
2005 prefetch->init_req(0, needgc_false);
2006 }
2007 transform_later(prefetch);
2008 distance += step_size;
2009 i_o = prefetch;
2010 }
2011 }
2012 return i_o;
2013 }
2014
2015
2016 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) {
2017 expand_allocate_common(alloc, NULL,
2018 OptoRuntime::new_instance_Type(),
2019 OptoRuntime::new_instance_Java());
2020 }
2021
2022 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) {
2023 Node* length = alloc->in(AllocateNode::ALength);
2024 InitializeNode* init = alloc->initialization();
2025 Node* klass_node = alloc->in(AllocateNode::KlassNode);
2026 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
2027 address slow_call_address; // Address of slow call
2028 if (init != NULL && init->is_complete_with_arraycopy() &&
2029 k->is_type_array_klass()) {
2030 // Don't zero type array during slow allocation in VM since
2031 // it will be initialized later by arraycopy in compiled code.
2032 slow_call_address = OptoRuntime::new_array_nozero_Java();
2033 } else {
2034 slow_call_address = OptoRuntime::new_array_Java();
2035 }
2036 expand_allocate_common(alloc, length,
2037 OptoRuntime::new_array_Type(),
2038 slow_call_address);
2039 }
2040
2041 //-------------------mark_eliminated_box----------------------------------
2042 //
2043 // During EA obj may point to several objects but after few ideal graph
2044 // transformations (CCP) it may point to only one non escaping object
2045 // (but still using phi), corresponding locks and unlocks will be marked
2046 // for elimination. Later obj could be replaced with a new node (new phi)
2047 // and which does not have escape information. And later after some graph
2048 // reshape other locks and unlocks (which were not marked for elimination
2049 // before) are connected to this new obj (phi) but they still will not be
2050 // marked for elimination since new obj has no escape information.
2051 // Mark all associated (same box and obj) lock and unlock nodes for
2052 // elimination if some of them marked already.
2053 void PhaseMacroExpand::mark_eliminated_box(Node* oldbox, Node* obj) {
2054 if (oldbox->as_BoxLock()->is_eliminated())
2055 return; // This BoxLock node was processed already.
2056
2057 // New implementation (EliminateNestedLocks) has separate BoxLock
2058 // node for each locked region so mark all associated locks/unlocks as
2059 // eliminated even if different objects are referenced in one locked region
2060 // (for example, OSR compilation of nested loop inside locked scope).
2061 if (EliminateNestedLocks ||
2062 oldbox->as_BoxLock()->is_simple_lock_region(NULL, obj)) {
2063 // Box is used only in one lock region. Mark this box as eliminated.
2064 _igvn.hash_delete(oldbox);
2065 oldbox->as_BoxLock()->set_eliminated(); // This changes box's hash value
2066 _igvn.hash_insert(oldbox);
2067
2068 for (uint i = 0; i < oldbox->outcnt(); i++) {
2069 Node* u = oldbox->raw_out(i);
2070 if (u->is_AbstractLock() && !u->as_AbstractLock()->is_non_esc_obj()) {
2071 AbstractLockNode* alock = u->as_AbstractLock();
2072 // Check lock's box since box could be referenced by Lock's debug info.
2073 if (alock->box_node() == oldbox) {
2074 // Mark eliminated all related locks and unlocks.
2075 #ifdef ASSERT
2076 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc4");
2077 #endif
2078 alock->set_non_esc_obj();
2079 }
2080 }
2081 }
2082 return;
2083 }
2084
2085 // Create new "eliminated" BoxLock node and use it in monitor debug info
2086 // instead of oldbox for the same object.
2087 BoxLockNode* newbox = oldbox->clone()->as_BoxLock();
2088
2089 // Note: BoxLock node is marked eliminated only here and it is used
2090 // to indicate that all associated lock and unlock nodes are marked
2091 // for elimination.
2092 newbox->set_eliminated();
2093 transform_later(newbox);
2094
2095 // Replace old box node with new box for all users of the same object.
2096 for (uint i = 0; i < oldbox->outcnt();) {
2097 bool next_edge = true;
2098
2099 Node* u = oldbox->raw_out(i);
2100 if (u->is_AbstractLock()) {
2101 AbstractLockNode* alock = u->as_AbstractLock();
2102 if (alock->box_node() == oldbox && alock->obj_node()->eqv_uncast(obj)) {
2103 // Replace Box and mark eliminated all related locks and unlocks.
2104 #ifdef ASSERT
2105 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc5");
2106 #endif
2107 alock->set_non_esc_obj();
2108 _igvn.rehash_node_delayed(alock);
2109 alock->set_box_node(newbox);
2110 next_edge = false;
2111 }
2112 }
2113 if (u->is_FastLock() && u->as_FastLock()->obj_node()->eqv_uncast(obj)) {
2114 FastLockNode* flock = u->as_FastLock();
2115 assert(flock->box_node() == oldbox, "sanity");
2116 _igvn.rehash_node_delayed(flock);
2117 flock->set_box_node(newbox);
2118 next_edge = false;
2119 }
2120
2121 // Replace old box in monitor debug info.
2122 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) {
2123 SafePointNode* sfn = u->as_SafePoint();
2124 JVMState* youngest_jvms = sfn->jvms();
2125 int max_depth = youngest_jvms->depth();
2126 for (int depth = 1; depth <= max_depth; depth++) {
2127 JVMState* jvms = youngest_jvms->of_depth(depth);
2128 int num_mon = jvms->nof_monitors();
2129 // Loop over monitors
2130 for (int idx = 0; idx < num_mon; idx++) {
2131 Node* obj_node = sfn->monitor_obj(jvms, idx);
2132 Node* box_node = sfn->monitor_box(jvms, idx);
2133 if (box_node == oldbox && obj_node->eqv_uncast(obj)) {
2134 int j = jvms->monitor_box_offset(idx);
2135 _igvn.replace_input_of(u, j, newbox);
2136 next_edge = false;
2137 }
2138 }
2139 }
2140 }
2141 if (next_edge) i++;
2142 }
2143 }
2144
2145 //-----------------------mark_eliminated_locking_nodes-----------------------
2146 void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) {
2147 if (EliminateNestedLocks) {
2148 if (alock->is_nested()) {
2149 assert(alock->box_node()->as_BoxLock()->is_eliminated(), "sanity");
2150 return;
2151 } else if (!alock->is_non_esc_obj()) { // Not eliminated or coarsened
2152 // Only Lock node has JVMState needed here.
2153 // Not that preceding claim is documented anywhere else.
2154 if (alock->jvms() != NULL) {
2155 if (alock->as_Lock()->is_nested_lock_region()) {
2156 // Mark eliminated related nested locks and unlocks.
2157 Node* obj = alock->obj_node();
2158 BoxLockNode* box_node = alock->box_node()->as_BoxLock();
2159 assert(!box_node->is_eliminated(), "should not be marked yet");
2160 // Note: BoxLock node is marked eliminated only here
2161 // and it is used to indicate that all associated lock
2162 // and unlock nodes are marked for elimination.
2163 box_node->set_eliminated(); // Box's hash is always NO_HASH here
2164 for (uint i = 0; i < box_node->outcnt(); i++) {
2165 Node* u = box_node->raw_out(i);
2166 if (u->is_AbstractLock()) {
2167 alock = u->as_AbstractLock();
2168 if (alock->box_node() == box_node) {
2169 // Verify that this Box is referenced only by related locks.
2170 assert(alock->obj_node()->eqv_uncast(obj), "");
2171 // Mark all related locks and unlocks.
2172 #ifdef ASSERT
2173 alock->log_lock_optimization(C, "eliminate_lock_set_nested");
2174 #endif
2175 alock->set_nested();
2176 }
2177 }
2178 }
2179 } else {
2180 #ifdef ASSERT
2181 alock->log_lock_optimization(C, "eliminate_lock_NOT_nested_lock_region");
2182 if (C->log() != NULL)
2183 alock->as_Lock()->is_nested_lock_region(C); // rerun for debugging output
2184 #endif
2185 }
2186 }
2187 return;
2188 }
2189 // Process locks for non escaping object
2190 assert(alock->is_non_esc_obj(), "");
2191 } // EliminateNestedLocks
2192
2193 if (alock->is_non_esc_obj()) { // Lock is used for non escaping object
2194 // Look for all locks of this object and mark them and
2195 // corresponding BoxLock nodes as eliminated.
2196 Node* obj = alock->obj_node();
2197 for (uint j = 0; j < obj->outcnt(); j++) {
2198 Node* o = obj->raw_out(j);
2199 if (o->is_AbstractLock() &&
2200 o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) {
2201 alock = o->as_AbstractLock();
2202 Node* box = alock->box_node();
2203 // Replace old box node with new eliminated box for all users
2204 // of the same object and mark related locks as eliminated.
2205 mark_eliminated_box(box, obj);
2206 }
2207 }
2208 }
2209 }
2210
2211 // we have determined that this lock/unlock can be eliminated, we simply
2212 // eliminate the node without expanding it.
2213 //
2214 // Note: The membar's associated with the lock/unlock are currently not
2215 // eliminated. This should be investigated as a future enhancement.
2216 //
2217 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) {
2218
2219 if (!alock->is_eliminated()) {
2220 return false;
2221 }
2222 #ifdef ASSERT
2223 if (!alock->is_coarsened()) {
2224 // Check that new "eliminated" BoxLock node is created.
2225 BoxLockNode* oldbox = alock->box_node()->as_BoxLock();
2226 assert(oldbox->is_eliminated(), "should be done already");
2227 }
2228 #endif
2229
2230 alock->log_lock_optimization(C, "eliminate_lock");
2231
2232 #ifndef PRODUCT
2233 if (PrintEliminateLocks) {
2234 if (alock->is_Lock()) {
2235 tty->print_cr("++++ Eliminated: %d Lock", alock->_idx);
2236 } else {
2237 tty->print_cr("++++ Eliminated: %d Unlock", alock->_idx);
2238 }
2239 }
2240 #endif
2241
2242 Node* mem = alock->in(TypeFunc::Memory);
2243 Node* ctrl = alock->in(TypeFunc::Control);
2244 guarantee(ctrl != NULL, "missing control projection, cannot replace_node() with NULL");
2245
2246 extract_call_projections(alock);
2247 // There are 2 projections from the lock. The lock node will
2248 // be deleted when its last use is subsumed below.
2249 assert(alock->outcnt() == 2 &&
2250 _fallthroughproj != NULL &&
2251 _memproj_fallthrough != NULL,
2252 "Unexpected projections from Lock/Unlock");
2253
2254 Node* fallthroughproj = _fallthroughproj;
2255 Node* memproj_fallthrough = _memproj_fallthrough;
2256
2257 // The memory projection from a lock/unlock is RawMem
2258 // The input to a Lock is merged memory, so extract its RawMem input
2259 // (unless the MergeMem has been optimized away.)
2260 if (alock->is_Lock()) {
2261 // Seach for MemBarAcquireLock node and delete it also.
2262 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
2263 assert(membar != NULL && membar->Opcode() == Op_MemBarAcquireLock, "");
2264 Node* ctrlproj = membar->proj_out(TypeFunc::Control);
2265 Node* memproj = membar->proj_out(TypeFunc::Memory);
2266 _igvn.replace_node(ctrlproj, fallthroughproj);
2267 _igvn.replace_node(memproj, memproj_fallthrough);
2268
2269 // Delete FastLock node also if this Lock node is unique user
2270 // (a loop peeling may clone a Lock node).
2271 Node* flock = alock->as_Lock()->fastlock_node();
2272 if (flock->outcnt() == 1) {
2273 assert(flock->unique_out() == alock, "sanity");
2274 _igvn.replace_node(flock, top());
2275 }
2276 }
2277
2278 // Seach for MemBarReleaseLock node and delete it also.
2279 if (alock->is_Unlock() && ctrl->is_Proj() && ctrl->in(0)->is_MemBar()) {
2280 MemBarNode* membar = ctrl->in(0)->as_MemBar();
2281 assert(membar->Opcode() == Op_MemBarReleaseLock &&
2282 mem->is_Proj() && membar == mem->in(0), "");
2283 _igvn.replace_node(fallthroughproj, ctrl);
2284 _igvn.replace_node(memproj_fallthrough, mem);
2285 fallthroughproj = ctrl;
2286 memproj_fallthrough = mem;
2287 ctrl = membar->in(TypeFunc::Control);
2288 mem = membar->in(TypeFunc::Memory);
2289 }
2290
2291 _igvn.replace_node(fallthroughproj, ctrl);
2292 _igvn.replace_node(memproj_fallthrough, mem);
2293 return true;
2294 }
2295
2296
2297 //------------------------------expand_lock_node----------------------
2298 void PhaseMacroExpand::expand_lock_node(LockNode *lock) {
2299
2300 Node* ctrl = lock->in(TypeFunc::Control);
2301 Node* mem = lock->in(TypeFunc::Memory);
2302 Node* obj = lock->obj_node();
2303 Node* box = lock->box_node();
2304 Node* flock = lock->fastlock_node();
2305
2306 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2307
2308 // Make the merge point
2309 Node *region;
2310 Node *mem_phi;
2311 Node *slow_path;
2312
2313 if (UseOptoBiasInlining) {
2314 /*
2315 * See the full description in MacroAssembler::biased_locking_enter().
2316 *
2317 * if( (mark_word & biased_lock_mask) == biased_lock_pattern ) {
2318 * // The object is biased.
2319 * proto_node = klass->prototype_header;
2320 * o_node = thread | proto_node;
2321 * x_node = o_node ^ mark_word;
2322 * if( (x_node & ~age_mask) == 0 ) { // Biased to the current thread ?
2323 * // Done.
2324 * } else {
2325 * if( (x_node & biased_lock_mask) != 0 ) {
2326 * // The klass's prototype header is no longer biased.
2327 * cas(&mark_word, mark_word, proto_node)
2328 * goto cas_lock;
2329 * } else {
2330 * // The klass's prototype header is still biased.
2331 * if( (x_node & epoch_mask) != 0 ) { // Expired epoch?
2332 * old = mark_word;
2333 * new = o_node;
2334 * } else {
2335 * // Different thread or anonymous biased.
2336 * old = mark_word & (epoch_mask | age_mask | biased_lock_mask);
2337 * new = thread | old;
2338 * }
2339 * // Try to rebias.
2340 * if( cas(&mark_word, old, new) == 0 ) {
2341 * // Done.
2342 * } else {
2343 * goto slow_path; // Failed.
2344 * }
2345 * }
2346 * }
2347 * } else {
2348 * // The object is not biased.
2349 * cas_lock:
2350 * if( FastLock(obj) == 0 ) {
2351 * // Done.
2352 * } else {
2353 * slow_path:
2354 * OptoRuntime::complete_monitor_locking_Java(obj);
2355 * }
2356 * }
2357 */
2358
2359 region = new RegionNode(5);
2360 // create a Phi for the memory state
2361 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2362
2363 Node* fast_lock_region = new RegionNode(3);
2364 Node* fast_lock_mem_phi = new PhiNode( fast_lock_region, Type::MEMORY, TypeRawPtr::BOTTOM);
2365
2366 // First, check mark word for the biased lock pattern.
2367 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2368
2369 // Get fast path - mark word has the biased lock pattern.
2370 ctrl = opt_bits_test(ctrl, fast_lock_region, 1, mark_node,
2371 markOopDesc::biased_lock_mask_in_place,
2372 markOopDesc::biased_lock_pattern, true);
2373 // fast_lock_region->in(1) is set to slow path.
2374 fast_lock_mem_phi->init_req(1, mem);
2375
2376 // Now check that the lock is biased to the current thread and has
2377 // the same epoch and bias as Klass::_prototype_header.
2378
2379 // Special-case a fresh allocation to avoid building nodes:
2380 Node* klass_node = AllocateNode::Ideal_klass(obj, &_igvn);
2381 if (klass_node == NULL) {
2382 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
2383 klass_node = transform_later(LoadKlassNode::make(_igvn, NULL, mem, k_adr, _igvn.type(k_adr)->is_ptr()));
2384 #ifdef _LP64
2385 if (UseCompressedClassPointers && klass_node->is_DecodeNKlass()) {
2386 assert(klass_node->in(1)->Opcode() == Op_LoadNKlass, "sanity");
2387 klass_node->in(1)->init_req(0, ctrl);
2388 } else
2389 #endif
2390 klass_node->init_req(0, ctrl);
2391 }
2392 Node *proto_node = make_load(ctrl, mem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeX_X, TypeX_X->basic_type());
2393
2394 Node* thread = transform_later(new ThreadLocalNode());
2395 Node* cast_thread = transform_later(new CastP2XNode(ctrl, thread));
2396 Node* o_node = transform_later(new OrXNode(cast_thread, proto_node));
2397 Node* x_node = transform_later(new XorXNode(o_node, mark_node));
2398
2399 // Get slow path - mark word does NOT match the value.
2400 Node* not_biased_ctrl = opt_bits_test(ctrl, region, 3, x_node,
2401 (~markOopDesc::age_mask_in_place), 0);
2402 // region->in(3) is set to fast path - the object is biased to the current thread.
2403 mem_phi->init_req(3, mem);
2404
2405
2406 // Mark word does NOT match the value (thread | Klass::_prototype_header).
2407
2408
2409 // First, check biased pattern.
2410 // Get fast path - _prototype_header has the same biased lock pattern.
2411 ctrl = opt_bits_test(not_biased_ctrl, fast_lock_region, 2, x_node,
2412 markOopDesc::biased_lock_mask_in_place, 0, true);
2413
2414 not_biased_ctrl = fast_lock_region->in(2); // Slow path
2415 // fast_lock_region->in(2) - the prototype header is no longer biased
2416 // and we have to revoke the bias on this object.
2417 // We are going to try to reset the mark of this object to the prototype
2418 // value and fall through to the CAS-based locking scheme.
2419 Node* adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
2420 Node* cas = new StoreXConditionalNode(not_biased_ctrl, mem, adr,
2421 proto_node, mark_node);
2422 transform_later(cas);
2423 Node* proj = transform_later(new SCMemProjNode(cas));
2424 fast_lock_mem_phi->init_req(2, proj);
2425
2426
2427 // Second, check epoch bits.
2428 Node* rebiased_region = new RegionNode(3);
2429 Node* old_phi = new PhiNode( rebiased_region, TypeX_X);
2430 Node* new_phi = new PhiNode( rebiased_region, TypeX_X);
2431
2432 // Get slow path - mark word does NOT match epoch bits.
2433 Node* epoch_ctrl = opt_bits_test(ctrl, rebiased_region, 1, x_node,
2434 markOopDesc::epoch_mask_in_place, 0);
2435 // The epoch of the current bias is not valid, attempt to rebias the object
2436 // toward the current thread.
2437 rebiased_region->init_req(2, epoch_ctrl);
2438 old_phi->init_req(2, mark_node);
2439 new_phi->init_req(2, o_node);
2440
2441 // rebiased_region->in(1) is set to fast path.
2442 // The epoch of the current bias is still valid but we know
2443 // nothing about the owner; it might be set or it might be clear.
2444 Node* cmask = MakeConX(markOopDesc::biased_lock_mask_in_place |
2445 markOopDesc::age_mask_in_place |
2446 markOopDesc::epoch_mask_in_place);
2447 Node* old = transform_later(new AndXNode(mark_node, cmask));
2448 cast_thread = transform_later(new CastP2XNode(ctrl, thread));
2449 Node* new_mark = transform_later(new OrXNode(cast_thread, old));
2450 old_phi->init_req(1, old);
2451 new_phi->init_req(1, new_mark);
2452
2453 transform_later(rebiased_region);
2454 transform_later(old_phi);
2455 transform_later(new_phi);
2456
2457 // Try to acquire the bias of the object using an atomic operation.
2458 // If this fails we will go in to the runtime to revoke the object's bias.
2459 cas = new StoreXConditionalNode(rebiased_region, mem, adr, new_phi, old_phi);
2460 transform_later(cas);
2461 proj = transform_later(new SCMemProjNode(cas));
2462
2463 // Get slow path - Failed to CAS.
2464 not_biased_ctrl = opt_bits_test(rebiased_region, region, 4, cas, 0, 0);
2465 mem_phi->init_req(4, proj);
2466 // region->in(4) is set to fast path - the object is rebiased to the current thread.
2467
2468 // Failed to CAS.
2469 slow_path = new RegionNode(3);
2470 Node *slow_mem = new PhiNode( slow_path, Type::MEMORY, TypeRawPtr::BOTTOM);
2471
2472 slow_path->init_req(1, not_biased_ctrl); // Capture slow-control
2473 slow_mem->init_req(1, proj);
2474
2475 // Call CAS-based locking scheme (FastLock node).
2476
2477 transform_later(fast_lock_region);
2478 transform_later(fast_lock_mem_phi);
2479
2480 // Get slow path - FastLock failed to lock the object.
2481 ctrl = opt_bits_test(fast_lock_region, region, 2, flock, 0, 0);
2482 mem_phi->init_req(2, fast_lock_mem_phi);
2483 // region->in(2) is set to fast path - the object is locked to the current thread.
2484
2485 slow_path->init_req(2, ctrl); // Capture slow-control
2486 slow_mem->init_req(2, fast_lock_mem_phi);
2487
2488 transform_later(slow_path);
2489 transform_later(slow_mem);
2490 // Reset lock's memory edge.
2491 lock->set_req(TypeFunc::Memory, slow_mem);
2492
2493 } else {
2494 region = new RegionNode(3);
2495 // create a Phi for the memory state
2496 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2497
2498 // Optimize test; set region slot 2
2499 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0);
2500 mem_phi->init_req(2, mem);
2501 }
2502
2503 // Make slow path call
2504 CallNode *call = make_slow_call((CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(),
2505 OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path,
2506 obj, box, NULL);
2507
2508 extract_call_projections(call);
2509
2510 // Slow path can only throw asynchronous exceptions, which are always
2511 // de-opted. So the compiler thinks the slow-call can never throw an
2512 // exception. If it DOES throw an exception we would need the debug
2513 // info removed first (since if it throws there is no monitor).
2514 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
2515 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
2516
2517 // Capture slow path
2518 // disconnect fall-through projection from call and create a new one
2519 // hook up users of fall-through projection to region
2520 Node *slow_ctrl = _fallthroughproj->clone();
2521 transform_later(slow_ctrl);
2522 _igvn.hash_delete(_fallthroughproj);
2523 _fallthroughproj->disconnect_inputs(NULL, C);
2524 region->init_req(1, slow_ctrl);
2525 // region inputs are now complete
2526 transform_later(region);
2527 _igvn.replace_node(_fallthroughproj, region);
2528
2529 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory));
2530 mem_phi->init_req(1, memproj );
2531 transform_later(mem_phi);
2532 _igvn.replace_node(_memproj_fallthrough, mem_phi);
2533 }
2534
2535 //------------------------------expand_unlock_node----------------------
2536 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
2537
2538 Node* ctrl = unlock->in(TypeFunc::Control);
2539 Node* mem = unlock->in(TypeFunc::Memory);
2540 Node* obj = unlock->obj_node();
2541 Node* box = unlock->box_node();
2542
2543 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2544
2545 // No need for a null check on unlock
2546
2547 // Make the merge point
2548 Node *region;
2549 Node *mem_phi;
2550
2551 if (UseOptoBiasInlining) {
2552 // Check for biased locking unlock case, which is a no-op.
2553 // See the full description in MacroAssembler::biased_locking_exit().
2554 region = new RegionNode(4);
2555 // create a Phi for the memory state
2556 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2557 mem_phi->init_req(3, mem);
2558
2559 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2560 ctrl = opt_bits_test(ctrl, region, 3, mark_node,
2561 markOopDesc::biased_lock_mask_in_place,
2562 markOopDesc::biased_lock_pattern);
2563 } else {
2564 region = new RegionNode(3);
2565 // create a Phi for the memory state
2566 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2567 }
2568
2569 FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box );
2570 funlock = transform_later( funlock )->as_FastUnlock();
2571 // Optimize test; set region slot 2
2572 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0);
2573 Node *thread = transform_later(new ThreadLocalNode());
2574
2575 CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(),
2576 CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C),
2577 "complete_monitor_unlocking_C", slow_path, obj, box, thread);
2578
2579 extract_call_projections(call);
2580
2581 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
2582 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
2583
2584 // No exceptions for unlocking
2585 // Capture slow path
2586 // disconnect fall-through projection from call and create a new one
2587 // hook up users of fall-through projection to region
2588 Node *slow_ctrl = _fallthroughproj->clone();
2589 transform_later(slow_ctrl);
2590 _igvn.hash_delete(_fallthroughproj);
2591 _fallthroughproj->disconnect_inputs(NULL, C);
2592 region->init_req(1, slow_ctrl);
2593 // region inputs are now complete
2594 transform_later(region);
2595 _igvn.replace_node(_fallthroughproj, region);
2596
2597 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory) );
2598 mem_phi->init_req(1, memproj );
2599 mem_phi->init_req(2, mem);
2600 transform_later(mem_phi);
2601 _igvn.replace_node(_memproj_fallthrough, mem_phi);
2602 }
2603
2604 // A value type is returned from the call but we don't know its
2605 // type. Either we get a buffered value (and nothing needs to be done)
2606 // or one of the values being returned is the klass of the value type
2607 // and we need to allocate a value type instance of that type and
2608 // initialize it with other values being returned. In that case, we
2609 // first try a fast path allocation and initialize the value with the
2610 // value klass's pack handler or we fall back to a runtime call.
2611 void PhaseMacroExpand::expand_mh_intrinsic_return(CallStaticJavaNode* call) {
2612 Node* ret = call->proj_out(TypeFunc::Parms);
2613 if (ret == NULL) {
2614 return;
2615 }
2616 // TODO fix this with the calling convention changes
2617 //assert(ret->bottom_type()->is_valuetypeptr()->is__Value(), "unexpected return type from MH intrinsic");
2618 const TypeFunc* tf = call->_tf;
2619 const TypeTuple* domain = OptoRuntime::store_value_type_fields_Type()->domain_cc();
2620 const TypeFunc* new_tf = TypeFunc::make(tf->domain_sig(), tf->domain_cc(), tf->range_sig(), domain);
2621 call->_tf = new_tf;
2622 // Make sure the change of type is applied before projections are
2623 // processed by igvn
2624 _igvn.set_type(call, call->Value(&_igvn));
2625 _igvn.set_type(ret, ret->Value(&_igvn));
2626
2627 // Before any new projection is added:
2628 CallProjections* projs = call->extract_projections(true, true);
2629
2630 Node* ctl = new Node(1);
2631 Node* mem = new Node(1);
2632 Node* io = new Node(1);
2633 Node* ex_ctl = new Node(1);
2634 Node* ex_mem = new Node(1);
2635 Node* ex_io = new Node(1);
2636 Node* res = new Node(1);
2637
2638 Node* cast = transform_later(new CastP2XNode(ctl, res));
2639 Node* mask = MakeConX(0x1);
2640 Node* masked = transform_later(new AndXNode(cast, mask));
2641 Node* cmp = transform_later(new CmpXNode(masked, mask));
2642 Node* bol = transform_later(new BoolNode(cmp, BoolTest::eq));
2643 IfNode* allocation_iff = new IfNode(ctl, bol, PROB_MAX, COUNT_UNKNOWN);
2644 transform_later(allocation_iff);
2645 Node* allocation_ctl = transform_later(new IfTrueNode(allocation_iff));
2646 Node* no_allocation_ctl = transform_later(new IfFalseNode(allocation_iff));
2647
2648 Node* no_allocation_res = transform_later(new CheckCastPPNode(no_allocation_ctl, res, TypeInstPtr::NOTNULL));
2649
2650 Node* mask2 = MakeConX(-2);
2651 Node* masked2 = transform_later(new AndXNode(cast, mask2));
2652 Node* rawklassptr = transform_later(new CastX2PNode(masked2));
2653 Node* klass_node = transform_later(new CheckCastPPNode(allocation_ctl, rawklassptr, TypeKlassPtr::OBJECT_OR_NULL));
2654
2655 Node* slowpath_bol = NULL;
2656 Node* top_adr = NULL;
2657 Node* old_top = NULL;
2658 Node* new_top = NULL;
2659 if (UseTLAB) {
2660 Node* end_adr = NULL;
2661 set_eden_pointers(top_adr, end_adr);
2662 Node* end = make_load(ctl, mem, end_adr, 0, TypeRawPtr::BOTTOM, T_ADDRESS);
2663 old_top = new LoadPNode(ctl, mem, top_adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, MemNode::unordered);
2664 transform_later(old_top);
2665 Node* layout_val = make_load(NULL, mem, klass_node, in_bytes(Klass::layout_helper_offset()), TypeInt::INT, T_INT);
2666 Node* size_in_bytes = ConvI2X(layout_val);
2667 new_top = new AddPNode(top(), old_top, size_in_bytes);
2668 transform_later(new_top);
2669 Node* slowpath_cmp = new CmpPNode(new_top, end);
2670 transform_later(slowpath_cmp);
2671 slowpath_bol = new BoolNode(slowpath_cmp, BoolTest::ge);
2672 transform_later(slowpath_bol);
2673 } else {
2674 slowpath_bol = intcon(1);
2675 top_adr = top();
2676 old_top = top();
2677 new_top = top();
2678 }
2679 IfNode* slowpath_iff = new IfNode(allocation_ctl, slowpath_bol, PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN);
2680 transform_later(slowpath_iff);
2681
2682 Node* slowpath_true = new IfTrueNode(slowpath_iff);
2683 transform_later(slowpath_true);
2684
2685 CallStaticJavaNode* slow_call = new CallStaticJavaNode(OptoRuntime::store_value_type_fields_Type(),
2686 StubRoutines::store_value_type_fields_to_buf(),
2687 "store_value_type_fields",
2688 call->jvms()->bci(),
2689 TypePtr::BOTTOM);
2690 slow_call->init_req(TypeFunc::Control, slowpath_true);
2691 slow_call->init_req(TypeFunc::Memory, mem);
2692 slow_call->init_req(TypeFunc::I_O, io);
2693 slow_call->init_req(TypeFunc::FramePtr, call->in(TypeFunc::FramePtr));
2694 slow_call->init_req(TypeFunc::ReturnAdr, call->in(TypeFunc::ReturnAdr));
2695 slow_call->init_req(TypeFunc::Parms, res);
2696
2697 Node* slow_ctl = transform_later(new ProjNode(slow_call, TypeFunc::Control));
2698 Node* slow_mem = transform_later(new ProjNode(slow_call, TypeFunc::Memory));
2699 Node* slow_io = transform_later(new ProjNode(slow_call, TypeFunc::I_O));
2700 Node* slow_res = transform_later(new ProjNode(slow_call, TypeFunc::Parms));
2701 Node* slow_catc = transform_later(new CatchNode(slow_ctl, slow_io, 2));
2702 Node* slow_norm = transform_later(new CatchProjNode(slow_catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci));
2703 Node* slow_excp = transform_later(new CatchProjNode(slow_catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci));
2704
2705 Node* ex_r = new RegionNode(3);
2706 Node* ex_mem_phi = new PhiNode(ex_r, Type::MEMORY, TypePtr::BOTTOM);
2707 Node* ex_io_phi = new PhiNode(ex_r, Type::ABIO);
2708 ex_r->init_req(1, slow_excp);
2709 ex_mem_phi->init_req(1, slow_mem);
2710 ex_io_phi->init_req(1, slow_io);
2711 ex_r->init_req(2, ex_ctl);
2712 ex_mem_phi->init_req(2, ex_mem);
2713 ex_io_phi->init_req(2, ex_io);
2714
2715 transform_later(ex_r);
2716 transform_later(ex_mem_phi);
2717 transform_later(ex_io_phi);
2718
2719 Node* slowpath_false = new IfFalseNode(slowpath_iff);
2720 transform_later(slowpath_false);
2721 Node* rawmem = new StorePNode(slowpath_false, mem, top_adr, TypeRawPtr::BOTTOM, new_top, MemNode::unordered);
2722 transform_later(rawmem);
2723 Node* mark_node = makecon(TypeRawPtr::make((address)markOopDesc::always_locked_prototype()));
2724 rawmem = make_store(slowpath_false, rawmem, old_top, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS);
2725 rawmem = make_store(slowpath_false, rawmem, old_top, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
2726 if (UseCompressedClassPointers) {
2727 rawmem = make_store(slowpath_false, rawmem, old_top, oopDesc::klass_gap_offset_in_bytes(), intcon(0), T_INT);
2728 }
2729 Node* pack_handler = make_load(slowpath_false, rawmem, klass_node, in_bytes(ValueKlass::pack_handler_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
2730
2731 CallLeafNoFPNode* handler_call = new CallLeafNoFPNode(OptoRuntime::pack_value_type_Type(),
2732 NULL,
2733 "pack handler",
2734 TypeRawPtr::BOTTOM);
2735 handler_call->init_req(TypeFunc::Control, slowpath_false);
2736 handler_call->init_req(TypeFunc::Memory, rawmem);
2737 handler_call->init_req(TypeFunc::I_O, top());
2738 handler_call->init_req(TypeFunc::FramePtr, call->in(TypeFunc::FramePtr));
2739 handler_call->init_req(TypeFunc::ReturnAdr, top());
2740 handler_call->init_req(TypeFunc::Parms, pack_handler);
2741 handler_call->init_req(TypeFunc::Parms+1, old_top);
2742
2743 // We don't know how many values are returned. This assumes the
2744 // worst case, that all available registers are used.
2745 for (uint i = TypeFunc::Parms+1; i < domain->cnt(); i++) {
2746 if (domain->field_at(i) == Type::HALF) {
2747 slow_call->init_req(i, top());
2748 handler_call->init_req(i+1, top());
2749 continue;
2750 }
2751 Node* proj = transform_later(new ProjNode(call, i));
2752 slow_call->init_req(i, proj);
2753 handler_call->init_req(i+1, proj);
2754 }
2755
2756 // We can safepoint at that new call
2757 copy_call_debug_info(call, slow_call);
2758 transform_later(slow_call);
2759 transform_later(handler_call);
2760
2761 Node* handler_ctl = transform_later(new ProjNode(handler_call, TypeFunc::Control));
2762 rawmem = transform_later(new ProjNode(handler_call, TypeFunc::Memory));
2763 Node* slowpath_false_res = transform_later(new ProjNode(handler_call, TypeFunc::Parms));
2764
2765 MergeMemNode* slowpath_false_mem = MergeMemNode::make(mem);
2766 slowpath_false_mem->set_memory_at(Compile::AliasIdxRaw, rawmem);
2767 transform_later(slowpath_false_mem);
2768
2769 Node* r = new RegionNode(4);
2770 Node* mem_phi = new PhiNode(r, Type::MEMORY, TypePtr::BOTTOM);
2771 Node* io_phi = new PhiNode(r, Type::ABIO);
2772 Node* res_phi = new PhiNode(r, ret->bottom_type());
2773
2774 r->init_req(1, no_allocation_ctl);
2775 mem_phi->init_req(1, mem);
2776 io_phi->init_req(1, io);
2777 res_phi->init_req(1, no_allocation_res);
2778 r->init_req(2, slow_norm);
2779 mem_phi->init_req(2, slow_mem);
2780 io_phi->init_req(2, slow_io);
2781 res_phi->init_req(2, slow_res);
2782 r->init_req(3, handler_ctl);
2783 mem_phi->init_req(3, slowpath_false_mem);
2784 io_phi->init_req(3, io);
2785 res_phi->init_req(3, slowpath_false_res);
2786
2787 transform_later(r);
2788 transform_later(mem_phi);
2789 transform_later(io_phi);
2790 transform_later(res_phi);
2791
2792 assert(projs->nb_resproj == 1, "unexpected number of results");
2793 _igvn.replace_in_uses(projs->fallthrough_catchproj, r);
2794 _igvn.replace_in_uses(projs->fallthrough_memproj, mem_phi);
2795 _igvn.replace_in_uses(projs->fallthrough_ioproj, io_phi);
2796 _igvn.replace_in_uses(projs->resproj[0], res_phi);
2797 _igvn.replace_in_uses(projs->catchall_catchproj, ex_r);
2798 _igvn.replace_in_uses(projs->catchall_memproj, ex_mem_phi);
2799 _igvn.replace_in_uses(projs->catchall_ioproj, ex_io_phi);
2800
2801 _igvn.replace_node(ctl, projs->fallthrough_catchproj);
2802 _igvn.replace_node(mem, projs->fallthrough_memproj);
2803 _igvn.replace_node(io, projs->fallthrough_ioproj);
2804 _igvn.replace_node(res, projs->resproj[0]);
2805 _igvn.replace_node(ex_ctl, projs->catchall_catchproj);
2806 _igvn.replace_node(ex_mem, projs->catchall_memproj);
2807 _igvn.replace_node(ex_io, projs->catchall_ioproj);
2808 }
2809
2810 //---------------------------eliminate_macro_nodes----------------------
2811 // Eliminate scalar replaced allocations and associated locks.
2812 void PhaseMacroExpand::eliminate_macro_nodes() {
2813 if (C->macro_count() == 0)
2814 return;
2815
2816 // First, attempt to eliminate locks
2817 int cnt = C->macro_count();
2818 for (int i=0; i < cnt; i++) {
2819 Node *n = C->macro_node(i);
2820 if (n->is_AbstractLock()) { // Lock and Unlock nodes
2821 // Before elimination mark all associated (same box and obj)
2822 // lock and unlock nodes.
2823 mark_eliminated_locking_nodes(n->as_AbstractLock());
2824 }
2825 }
2826 bool progress = true;
2827 while (progress) {
2828 progress = false;
2829 for (int i = C->macro_count(); i > 0; i--) {
2830 Node * n = C->macro_node(i-1);
2831 bool success = false;
2832 debug_only(int old_macro_count = C->macro_count(););
2833 if (n->is_AbstractLock()) {
2834 success = eliminate_locking_node(n->as_AbstractLock());
2835 }
2836 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2837 progress = progress || success;
2838 }
2839 }
2840 // Next, attempt to eliminate allocations
2841 _has_locks = false;
2842 progress = true;
2843 while (progress) {
2844 progress = false;
2845 for (int i = C->macro_count(); i > 0; i--) {
2846 Node * n = C->macro_node(i-1);
2847 bool success = false;
2848 debug_only(int old_macro_count = C->macro_count(););
2849 switch (n->class_id()) {
2850 case Node::Class_Allocate:
2851 case Node::Class_AllocateArray:
2852 success = eliminate_allocate_node(n->as_Allocate());
2853 break;
2854 case Node::Class_CallStaticJava: {
2855 CallStaticJavaNode* call = n->as_CallStaticJava();
2856 if (!call->method()->is_method_handle_intrinsic()) {
2857 success = eliminate_boxing_node(n->as_CallStaticJava());
2858 }
2859 break;
2860 }
2861 case Node::Class_Lock:
2862 case Node::Class_Unlock:
2863 assert(!n->as_AbstractLock()->is_eliminated(), "sanity");
2864 _has_locks = true;
2865 break;
2866 case Node::Class_ArrayCopy:
2867 break;
2868 case Node::Class_OuterStripMinedLoop:
2869 break;
2870 default:
2871 assert(n->Opcode() == Op_LoopLimit ||
2872 n->Opcode() == Op_Opaque1 ||
2873 n->Opcode() == Op_Opaque2 ||
2874 n->Opcode() == Op_Opaque3 ||
2875 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(n),
2876 "unknown node type in macro list");
2877 }
2878 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2879 progress = progress || success;
2880 }
2881 }
2882 }
2883
2884 //------------------------------expand_macro_nodes----------------------
2885 // Returns true if a failure occurred.
2886 bool PhaseMacroExpand::expand_macro_nodes() {
2887 // Last attempt to eliminate macro nodes.
2888 eliminate_macro_nodes();
2889
2890 // Make sure expansion will not cause node limit to be exceeded.
2891 // Worst case is a macro node gets expanded into about 200 nodes.
2892 // Allow 50% more for optimization.
2893 if (C->check_node_count(C->macro_count() * 300, "out of nodes before macro expansion" ) )
2894 return true;
2895
2896 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations.
2897 bool progress = true;
2898 while (progress) {
2899 progress = false;
2900 for (int i = C->macro_count(); i > 0; i--) {
2901 Node * n = C->macro_node(i-1);
2902 bool success = false;
2903 debug_only(int old_macro_count = C->macro_count(););
2904 if (n->Opcode() == Op_LoopLimit) {
2905 // Remove it from macro list and put on IGVN worklist to optimize.
2906 C->remove_macro_node(n);
2907 _igvn._worklist.push(n);
2908 success = true;
2909 } else if (n->Opcode() == Op_CallStaticJava) {
2910 CallStaticJavaNode* call = n->as_CallStaticJava();
2911 if (!call->method()->is_method_handle_intrinsic()) {
2912 // Remove it from macro list and put on IGVN worklist to optimize.
2913 C->remove_macro_node(n);
2914 _igvn._worklist.push(n);
2915 success = true;
2916 }
2917 } else if (n->Opcode() == Op_Opaque1 || n->Opcode() == Op_Opaque2) {
2918 _igvn.replace_node(n, n->in(1));
2919 success = true;
2920 #if INCLUDE_RTM_OPT
2921 } else if ((n->Opcode() == Op_Opaque3) && ((Opaque3Node*)n)->rtm_opt()) {
2922 assert(C->profile_rtm(), "should be used only in rtm deoptimization code");
2923 assert((n->outcnt() == 1) && n->unique_out()->is_Cmp(), "");
2924 Node* cmp = n->unique_out();
2925 #ifdef ASSERT
2926 // Validate graph.
2927 assert((cmp->outcnt() == 1) && cmp->unique_out()->is_Bool(), "");
2928 BoolNode* bol = cmp->unique_out()->as_Bool();
2929 assert((bol->outcnt() == 1) && bol->unique_out()->is_If() &&
2930 (bol->_test._test == BoolTest::ne), "");
2931 IfNode* ifn = bol->unique_out()->as_If();
2932 assert((ifn->outcnt() == 2) &&
2933 ifn->proj_out(1)->is_uncommon_trap_proj(Deoptimization::Reason_rtm_state_change) != NULL, "");
2934 #endif
2935 Node* repl = n->in(1);
2936 if (!_has_locks) {
2937 // Remove RTM state check if there are no locks in the code.
2938 // Replace input to compare the same value.
2939 repl = (cmp->in(1) == n) ? cmp->in(2) : cmp->in(1);
2940 }
2941 _igvn.replace_node(n, repl);
2942 success = true;
2943 #endif
2944 } else if (n->Opcode() == Op_OuterStripMinedLoop) {
2945 n->as_OuterStripMinedLoop()->adjust_strip_mined_loop(&_igvn);
2946 C->remove_macro_node(n);
2947 success = true;
2948 }
2949 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2950 progress = progress || success;
2951 }
2952 }
2953
2954 // expand arraycopy "macro" nodes first
2955 // For ReduceBulkZeroing, we must first process all arraycopy nodes
2956 // before the allocate nodes are expanded.
2957 int macro_idx = C->macro_count() - 1;
2958 while (macro_idx >= 0) {
2959 Node * n = C->macro_node(macro_idx);
2960 assert(n->is_macro(), "only macro nodes expected here");
2961 if (_igvn.type(n) == Type::TOP || (n->in(0) != NULL && n->in(0)->is_top())) {
2962 // node is unreachable, so don't try to expand it
2963 C->remove_macro_node(n);
2964 } else if (n->is_ArrayCopy()){
2965 int macro_count = C->macro_count();
2966 expand_arraycopy_node(n->as_ArrayCopy());
2967 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
2968 }
2969 if (C->failing()) return true;
2970 macro_idx --;
2971 }
2972
2973 // expand "macro" nodes
2974 // nodes are removed from the macro list as they are processed
2975 while (C->macro_count() > 0) {
2976 int macro_count = C->macro_count();
2977 Node * n = C->macro_node(macro_count-1);
2978 assert(n->is_macro(), "only macro nodes expected here");
2979 if (_igvn.type(n) == Type::TOP || (n->in(0) != NULL && n->in(0)->is_top())) {
2980 // node is unreachable, so don't try to expand it
2981 C->remove_macro_node(n);
2982 continue;
2983 }
2984 switch (n->class_id()) {
2985 case Node::Class_Allocate:
2986 expand_allocate(n->as_Allocate());
2987 break;
2988 case Node::Class_AllocateArray:
2989 expand_allocate_array(n->as_AllocateArray());
2990 break;
2991 case Node::Class_Lock:
2992 expand_lock_node(n->as_Lock());
2993 break;
2994 case Node::Class_Unlock:
2995 expand_unlock_node(n->as_Unlock());
2996 break;
2997 case Node::Class_CallStaticJava:
2998 expand_mh_intrinsic_return(n->as_CallStaticJava());
2999 C->remove_macro_node(n);
3000 break;
3001 default:
3002 assert(false, "unknown node type in macro list");
3003 }
3004 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
3005 if (C->failing()) return true;
3006 }
3007
3008 _igvn.set_delay_transform(false);
3009 _igvn.optimize();
3010 if (C->failing()) return true;
3011 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
3012 return bs->expand_macro_nodes(this);
3013 }