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