1 /*
2 * Copyright (c) 2005, 2018, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "compiler/compileLog.hpp"
27 #include "gc/shared/collectedHeap.inline.hpp"
28 #include "libadt/vectset.hpp"
29 #include "memory/universe.hpp"
30 #include "opto/addnode.hpp"
31 #include "opto/arraycopynode.hpp"
32 #include "opto/callnode.hpp"
33 #include "opto/castnode.hpp"
34 #include "opto/cfgnode.hpp"
35 #include "opto/compile.hpp"
36 #include "opto/convertnode.hpp"
37 #include "opto/graphKit.hpp"
38 #include "opto/locknode.hpp"
39 #include "opto/loopnode.hpp"
40 #include "opto/macro.hpp"
41 #include "opto/memnode.hpp"
42 #include "opto/narrowptrnode.hpp"
43 #include "opto/node.hpp"
44 #include "opto/opaquenode.hpp"
45 #include "opto/phaseX.hpp"
46 #include "opto/rootnode.hpp"
47 #include "opto/runtime.hpp"
48 #include "opto/subnode.hpp"
49 #include "opto/type.hpp"
50 #include "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 //
replace_input(Node * use,Node * oldref,Node * newref)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
copy_call_debug_info(CallNode * oldcall,CallNode * newcall)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()->cnt();
85 uint new_dbg_start = newcall->tf()->domain()->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
opt_bits_test(Node * ctrl,Node * region,int edge,Node * word,int mask,int bits,bool return_fast_path)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--------------------
copy_predefined_input_for_runtime_call(Node * ctrl,CallNode * oldcall,CallNode * 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---------------------------------
make_slow_call(CallNode * oldcall,const TypeFunc * slow_call_type,address slow_call,const char * leaf_name,Node * slow_path,Node * parm0,Node * parm1,Node * parm2)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
extract_call_projections(CallNode * call)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
eliminate_gc_barrier(Node * p2x)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.
scan_mem_chain(Node * mem,int alias_idx,int offset,Node * start_mem,Node * alloc,PhaseGVN * phase)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->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
make_arraycopy_load(ArrayCopyNode * ac,intptr_t offset,Node * ctl,Node * mem,BasicType ft,const Type * ftype,AllocateNode * alloc)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 assert(ac->in(ArrayCopyNode::Src) != ac->in(ArrayCopyNode::Dest), "clone source equals destination");
352 Node* base = ac->in(ArrayCopyNode::Src)->in(AddPNode::Base);
353 Node* adr = _igvn.transform(new AddPNode(base, base, MakeConX(offset)));
354 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset);
355 res = LoadNode::make(_igvn, ctl, mem, adr, adr_type, type, bt, MemNode::unordered, LoadNode::UnknownControl);
356 } else {
357 if (ac->modifies(offset, offset, &_igvn, true)) {
358 assert(ac->in(ArrayCopyNode::Dest) == alloc->result_cast(), "arraycopy destination should be allocation's result");
359 uint shift = exact_log2(type2aelembytes(bt));
360 Node* src_pos = ac->in(ArrayCopyNode::SrcPos);
361 Node* dest_pos = ac->in(ArrayCopyNode::DestPos);
362 const TypeInt* src_pos_t = _igvn.type(src_pos)->is_int();
363 const TypeInt* dest_pos_t = _igvn.type(dest_pos)->is_int();
364
365 Node* adr = NULL;
366 const TypePtr* adr_type = NULL;
367 if (src_pos_t->is_con() && dest_pos_t->is_con()) {
368 intptr_t off = ((src_pos_t->get_con() - dest_pos_t->get_con()) << shift) + offset;
369 Node* base = ac->in(ArrayCopyNode::Src);
370 adr = _igvn.transform(new AddPNode(base, base, MakeConX(off)));
371 adr_type = _igvn.type(base)->is_ptr()->add_offset(off);
372 if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) {
373 // Don't emit a new load from src if src == dst but try to get the value from memory instead
374 return value_from_mem(ac->in(TypeFunc::Memory), ctl, ft, ftype, adr_type->isa_oopptr(), alloc);
375 }
376 } else {
377 Node* diff = _igvn.transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
378 #ifdef _LP64
379 diff = _igvn.transform(new ConvI2LNode(diff));
380 #endif
381 diff = _igvn.transform(new LShiftXNode(diff, intcon(shift)));
382
383 Node* off = _igvn.transform(new AddXNode(MakeConX(offset), diff));
384 Node* base = ac->in(ArrayCopyNode::Src);
385 adr = _igvn.transform(new AddPNode(base, base, off));
386 adr_type = _igvn.type(base)->is_ptr()->add_offset(Type::OffsetBot);
387 if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) {
388 // Non constant offset in the array: we can't statically
389 // determine the value
390 return NULL;
391 }
392 }
393 res = LoadNode::make(_igvn, ctl, mem, adr, adr_type, type, bt, MemNode::unordered, LoadNode::UnknownControl);
394 }
395 }
396 if (res != NULL) {
397 res = _igvn.transform(res);
398 if (ftype->isa_narrowoop()) {
399 // PhaseMacroExpand::scalar_replacement adds DecodeN nodes
400 res = _igvn.transform(new EncodePNode(res, ftype));
401 }
402 return res;
403 }
404 return NULL;
405 }
406
407 //
408 // Given a Memory Phi, compute a value Phi containing the values from stores
409 // on the input paths.
410 // Note: this function is recursive, its depth is limited by the "level" argument
411 // Returns the computed Phi, or NULL if it cannot compute it.
value_from_mem_phi(Node * mem,BasicType ft,const Type * phi_type,const TypeOopPtr * adr_t,AllocateNode * alloc,Node_Stack * value_phis,int level)412 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) {
413 assert(mem->is_Phi(), "sanity");
414 int alias_idx = C->get_alias_index(adr_t);
415 int offset = adr_t->offset();
416 int instance_id = adr_t->instance_id();
417
418 // Check if an appropriate value phi already exists.
419 Node* region = mem->in(0);
420 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) {
421 Node* phi = region->fast_out(k);
422 if (phi->is_Phi() && phi != mem &&
423 phi->as_Phi()->is_same_inst_field(phi_type, (int)mem->_idx, instance_id, alias_idx, offset)) {
424 return phi;
425 }
426 }
427 // Check if an appropriate new value phi already exists.
428 Node* new_phi = value_phis->find(mem->_idx);
429 if (new_phi != NULL)
430 return new_phi;
431
432 if (level <= 0) {
433 return NULL; // Give up: phi tree too deep
434 }
435 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
436 Node *alloc_mem = alloc->in(TypeFunc::Memory);
437
438 uint length = mem->req();
439 GrowableArray <Node *> values(length, length, NULL, false);
440
441 // create a new Phi for the value
442 PhiNode *phi = new PhiNode(mem->in(0), phi_type, NULL, mem->_idx, instance_id, alias_idx, offset);
443 transform_later(phi);
444 value_phis->push(phi, mem->_idx);
445
446 for (uint j = 1; j < length; j++) {
447 Node *in = mem->in(j);
448 if (in == NULL || in->is_top()) {
449 values.at_put(j, in);
450 } else {
451 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn);
452 if (val == start_mem || val == alloc_mem) {
453 // hit a sentinel, return appropriate 0 value
454 values.at_put(j, _igvn.zerocon(ft));
455 continue;
456 }
457 if (val->is_Initialize()) {
458 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
459 }
460 if (val == NULL) {
461 return NULL; // can't find a value on this path
462 }
463 if (val == mem) {
464 values.at_put(j, mem);
465 } else if (val->is_Store()) {
466 Node* n = val->in(MemNode::ValueIn);
467 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
468 n = bs->step_over_gc_barrier(n);
469 values.at_put(j, n);
470 } else if(val->is_Proj() && val->in(0) == alloc) {
471 values.at_put(j, _igvn.zerocon(ft));
472 } else if (val->is_Phi()) {
473 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1);
474 if (val == NULL) {
475 return NULL;
476 }
477 values.at_put(j, val);
478 } else if (val->Opcode() == Op_SCMemProj) {
479 assert(val->in(0)->is_LoadStore() ||
480 val->in(0)->Opcode() == Op_EncodeISOArray ||
481 val->in(0)->Opcode() == Op_StrCompressedCopy, "sanity");
482 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
483 return NULL;
484 } else if (val->is_ArrayCopy()) {
485 Node* res = make_arraycopy_load(val->as_ArrayCopy(), offset, val->in(0), val->in(TypeFunc::Memory), ft, phi_type, alloc);
486 if (res == NULL) {
487 return NULL;
488 }
489 values.at_put(j, res);
490 } else {
491 #ifdef ASSERT
492 val->dump();
493 assert(false, "unknown node on this path");
494 #endif
495 return NULL; // unknown node on this path
496 }
497 }
498 }
499 // Set Phi's inputs
500 for (uint j = 1; j < length; j++) {
501 if (values.at(j) == mem) {
502 phi->init_req(j, phi);
503 } else {
504 phi->init_req(j, values.at(j));
505 }
506 }
507 return phi;
508 }
509
510 // Search the last value stored into the object's field.
value_from_mem(Node * sfpt_mem,Node * sfpt_ctl,BasicType ft,const Type * ftype,const TypeOopPtr * adr_t,AllocateNode * alloc)511 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, Node *sfpt_ctl, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, AllocateNode *alloc) {
512 assert(adr_t->is_known_instance_field(), "instance required");
513 int instance_id = adr_t->instance_id();
514 assert((uint)instance_id == alloc->_idx, "wrong allocation");
515
516 int alias_idx = C->get_alias_index(adr_t);
517 int offset = adr_t->offset();
518 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
519 Node *alloc_ctrl = alloc->in(TypeFunc::Control);
520 Node *alloc_mem = alloc->in(TypeFunc::Memory);
521 Arena *a = Thread::current()->resource_area();
522 VectorSet visited(a);
523
524 bool done = sfpt_mem == alloc_mem;
525 Node *mem = sfpt_mem;
526 while (!done) {
527 if (visited.test_set(mem->_idx)) {
528 return NULL; // found a loop, give up
529 }
530 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn);
531 if (mem == start_mem || mem == alloc_mem) {
532 done = true; // hit a sentinel, return appropriate 0 value
533 } else if (mem->is_Initialize()) {
534 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
535 if (mem == NULL) {
536 done = true; // Something go wrong.
537 } else if (mem->is_Store()) {
538 const TypePtr* atype = mem->as_Store()->adr_type();
539 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice");
540 done = true;
541 }
542 } else if (mem->is_Store()) {
543 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr();
544 assert(atype != NULL, "address type must be oopptr");
545 assert(C->get_alias_index(atype) == alias_idx &&
546 atype->is_known_instance_field() && atype->offset() == offset &&
547 atype->instance_id() == instance_id, "store is correct memory slice");
548 done = true;
549 } else if (mem->is_Phi()) {
550 // try to find a phi's unique input
551 Node *unique_input = NULL;
552 Node *top = C->top();
553 for (uint i = 1; i < mem->req(); i++) {
554 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn);
555 if (n == NULL || n == top || n == mem) {
556 continue;
557 } else if (unique_input == NULL) {
558 unique_input = n;
559 } else if (unique_input != n) {
560 unique_input = top;
561 break;
562 }
563 }
564 if (unique_input != NULL && unique_input != top) {
565 mem = unique_input;
566 } else {
567 done = true;
568 }
569 } else if (mem->is_ArrayCopy()) {
570 done = true;
571 } else {
572 assert(false, "unexpected node");
573 }
574 }
575 if (mem != NULL) {
576 if (mem == start_mem || mem == alloc_mem) {
577 // hit a sentinel, return appropriate 0 value
578 return _igvn.zerocon(ft);
579 } else if (mem->is_Store()) {
580 Node* n = mem->in(MemNode::ValueIn);
581 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
582 n = bs->step_over_gc_barrier(n);
583 return n;
584 } else if (mem->is_Phi()) {
585 // attempt to produce a Phi reflecting the values on the input paths of the Phi
586 Node_Stack value_phis(a, 8);
587 Node * phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit);
588 if (phi != NULL) {
589 return phi;
590 } else {
591 // Kill all new Phis
592 while(value_phis.is_nonempty()) {
593 Node* n = value_phis.node();
594 _igvn.replace_node(n, C->top());
595 value_phis.pop();
596 }
597 }
598 } else if (mem->is_ArrayCopy()) {
599 Node* ctl = mem->in(0);
600 Node* m = mem->in(TypeFunc::Memory);
601 if (sfpt_ctl->is_Proj() && sfpt_ctl->as_Proj()->is_uncommon_trap_proj(Deoptimization::Reason_none)) {
602 // pin the loads in the uncommon trap path
603 ctl = sfpt_ctl;
604 m = sfpt_mem;
605 }
606 return make_arraycopy_load(mem->as_ArrayCopy(), offset, ctl, m, ft, ftype, alloc);
607 }
608 }
609 // Something go wrong.
610 return NULL;
611 }
612
613 // Check the possibility of scalar replacement.
can_eliminate_allocation(AllocateNode * alloc,GrowableArray<SafePointNode * > & safepoints)614 bool PhaseMacroExpand::can_eliminate_allocation(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
615 // Scan the uses of the allocation to check for anything that would
616 // prevent us from eliminating it.
617 NOT_PRODUCT( const char* fail_eliminate = NULL; )
618 DEBUG_ONLY( Node* disq_node = NULL; )
619 bool can_eliminate = true;
620
621 Node* res = alloc->result_cast();
622 const TypeOopPtr* res_type = NULL;
623 if (res == NULL) {
624 // All users were eliminated.
625 } else if (!res->is_CheckCastPP()) {
626 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";)
627 can_eliminate = false;
628 } else {
629 res_type = _igvn.type(res)->isa_oopptr();
630 if (res_type == NULL) {
631 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";)
632 can_eliminate = false;
633 } else if (res_type->isa_aryptr()) {
634 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1);
635 if (length < 0) {
636 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";)
637 can_eliminate = false;
638 }
639 }
640 }
641
642 if (can_eliminate && res != NULL) {
643 for (DUIterator_Fast jmax, j = res->fast_outs(jmax);
644 j < jmax && can_eliminate; j++) {
645 Node* use = res->fast_out(j);
646
647 if (use->is_AddP()) {
648 const TypePtr* addp_type = _igvn.type(use)->is_ptr();
649 int offset = addp_type->offset();
650
651 if (offset == Type::OffsetTop || offset == Type::OffsetBot) {
652 NOT_PRODUCT(fail_eliminate = "Undefined field referrence";)
653 can_eliminate = false;
654 break;
655 }
656 for (DUIterator_Fast kmax, k = use->fast_outs(kmax);
657 k < kmax && can_eliminate; k++) {
658 Node* n = use->fast_out(k);
659 if (!n->is_Store() && n->Opcode() != Op_CastP2X &&
660 SHENANDOAHGC_ONLY((!UseShenandoahGC || !ShenandoahBarrierSetC2::is_shenandoah_wb_pre_call(n)) &&)
661 !(n->is_ArrayCopy() &&
662 n->as_ArrayCopy()->is_clonebasic() &&
663 n->in(ArrayCopyNode::Dest) == use)) {
664 DEBUG_ONLY(disq_node = n;)
665 if (n->is_Load() || n->is_LoadStore()) {
666 NOT_PRODUCT(fail_eliminate = "Field load";)
667 } else {
668 NOT_PRODUCT(fail_eliminate = "Not store field referrence";)
669 }
670 can_eliminate = false;
671 }
672 }
673 } else if (use->is_ArrayCopy() &&
674 (use->as_ArrayCopy()->is_arraycopy_validated() ||
675 use->as_ArrayCopy()->is_copyof_validated() ||
676 use->as_ArrayCopy()->is_copyofrange_validated()) &&
677 use->in(ArrayCopyNode::Dest) == res) {
678 // ok to eliminate
679 } else if (use->is_SafePoint()) {
680 SafePointNode* sfpt = use->as_SafePoint();
681 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) {
682 // Object is passed as argument.
683 DEBUG_ONLY(disq_node = use;)
684 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";)
685 can_eliminate = false;
686 }
687 Node* sfptMem = sfpt->memory();
688 if (sfptMem == NULL || sfptMem->is_top()) {
689 DEBUG_ONLY(disq_node = use;)
690 NOT_PRODUCT(fail_eliminate = "NULL or TOP memory";)
691 can_eliminate = false;
692 } else {
693 safepoints.append_if_missing(sfpt);
694 }
695 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
696 if (use->is_Phi()) {
697 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
698 NOT_PRODUCT(fail_eliminate = "Object is return value";)
699 } else {
700 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
701 }
702 DEBUG_ONLY(disq_node = use;)
703 } else {
704 if (use->Opcode() == Op_Return) {
705 NOT_PRODUCT(fail_eliminate = "Object is return value";)
706 }else {
707 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
708 }
709 DEBUG_ONLY(disq_node = use;)
710 }
711 can_eliminate = false;
712 }
713 }
714 }
715
716 #ifndef PRODUCT
717 if (PrintEliminateAllocations) {
718 if (can_eliminate) {
719 tty->print("Scalar ");
720 if (res == NULL)
721 alloc->dump();
722 else
723 res->dump();
724 } else if (alloc->_is_scalar_replaceable) {
725 tty->print("NotScalar (%s)", fail_eliminate);
726 if (res == NULL)
727 alloc->dump();
728 else
729 res->dump();
730 #ifdef ASSERT
731 if (disq_node != NULL) {
732 tty->print(" >>>> ");
733 disq_node->dump();
734 }
735 #endif /*ASSERT*/
736 }
737 }
738 #endif
739 return can_eliminate;
740 }
741
742 // Do scalar replacement.
scalar_replacement(AllocateNode * alloc,GrowableArray<SafePointNode * > & safepoints)743 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
744 GrowableArray <SafePointNode *> safepoints_done;
745
746 ciKlass* klass = NULL;
747 ciInstanceKlass* iklass = NULL;
748 int nfields = 0;
749 int array_base = 0;
750 int element_size = 0;
751 BasicType basic_elem_type = T_ILLEGAL;
752 ciType* elem_type = NULL;
753
754 Node* res = alloc->result_cast();
755 assert(res == NULL || res->is_CheckCastPP(), "unexpected AllocateNode result");
756 const TypeOopPtr* res_type = NULL;
757 if (res != NULL) { // Could be NULL when there are no users
758 res_type = _igvn.type(res)->isa_oopptr();
759 }
760
761 if (res != NULL) {
762 klass = res_type->klass();
763 if (res_type->isa_instptr()) {
764 // find the fields of the class which will be needed for safepoint debug information
765 assert(klass->is_instance_klass(), "must be an instance klass.");
766 iklass = klass->as_instance_klass();
767 nfields = iklass->nof_nonstatic_fields();
768 } else {
769 // find the array's elements which will be needed for safepoint debug information
770 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
771 assert(klass->is_array_klass() && nfields >= 0, "must be an array klass.");
772 elem_type = klass->as_array_klass()->element_type();
773 basic_elem_type = elem_type->basic_type();
774 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
775 element_size = type2aelembytes(basic_elem_type);
776 }
777 }
778 //
779 // Process the safepoint uses
780 //
781 while (safepoints.length() > 0) {
782 SafePointNode* sfpt = safepoints.pop();
783 Node* mem = sfpt->memory();
784 Node* ctl = sfpt->control();
785 assert(sfpt->jvms() != NULL, "missed JVMS");
786 // Fields of scalar objs are referenced only at the end
787 // of regular debuginfo at the last (youngest) JVMS.
788 // Record relative start index.
789 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff());
790 SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type,
791 #ifdef ASSERT
792 alloc,
793 #endif
794 first_ind, nfields);
795 sobj->init_req(0, C->root());
796 transform_later(sobj);
797
798 // Scan object's fields adding an input to the safepoint for each field.
799 for (int j = 0; j < nfields; j++) {
800 intptr_t offset;
801 ciField* field = NULL;
802 if (iklass != NULL) {
803 field = iklass->nonstatic_field_at(j);
804 offset = field->offset();
805 elem_type = field->type();
806 basic_elem_type = field->layout_type();
807 } else {
808 offset = array_base + j * (intptr_t)element_size;
809 }
810
811 const Type *field_type;
812 // The next code is taken from Parse::do_get_xxx().
813 if (basic_elem_type == T_OBJECT || basic_elem_type == T_ARRAY) {
814 if (!elem_type->is_loaded()) {
815 field_type = TypeInstPtr::BOTTOM;
816 } else if (field != NULL && field->is_static_constant()) {
817 // This can happen if the constant oop is non-perm.
818 ciObject* con = field->constant_value().as_object();
819 // Do not "join" in the previous type; it doesn't add value,
820 // and may yield a vacuous result if the field is of interface type.
821 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr();
822 assert(field_type != NULL, "field singleton type must be consistent");
823 } else {
824 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
825 }
826 if (UseCompressedOops) {
827 field_type = field_type->make_narrowoop();
828 basic_elem_type = T_NARROWOOP;
829 }
830 } else {
831 field_type = Type::get_const_basic_type(basic_elem_type);
832 }
833
834 const TypeOopPtr *field_addr_type = res_type->add_offset(offset)->isa_oopptr();
835
836 Node *field_val = value_from_mem(mem, ctl, basic_elem_type, field_type, field_addr_type, alloc);
837 if (field_val == NULL) {
838 // We weren't able to find a value for this field,
839 // give up on eliminating this allocation.
840
841 // Remove any extra entries we added to the safepoint.
842 uint last = sfpt->req() - 1;
843 for (int k = 0; k < j; k++) {
844 sfpt->del_req(last--);
845 }
846 _igvn._worklist.push(sfpt);
847 // rollback processed safepoints
848 while (safepoints_done.length() > 0) {
849 SafePointNode* sfpt_done = safepoints_done.pop();
850 // remove any extra entries we added to the safepoint
851 last = sfpt_done->req() - 1;
852 for (int k = 0; k < nfields; k++) {
853 sfpt_done->del_req(last--);
854 }
855 JVMState *jvms = sfpt_done->jvms();
856 jvms->set_endoff(sfpt_done->req());
857 // Now make a pass over the debug information replacing any references
858 // to SafePointScalarObjectNode with the allocated object.
859 int start = jvms->debug_start();
860 int end = jvms->debug_end();
861 for (int i = start; i < end; i++) {
862 if (sfpt_done->in(i)->is_SafePointScalarObject()) {
863 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject();
864 if (scobj->first_index(jvms) == sfpt_done->req() &&
865 scobj->n_fields() == (uint)nfields) {
866 assert(scobj->alloc() == alloc, "sanity");
867 sfpt_done->set_req(i, res);
868 }
869 }
870 }
871 _igvn._worklist.push(sfpt_done);
872 }
873 #ifndef PRODUCT
874 if (PrintEliminateAllocations) {
875 if (field != NULL) {
876 tty->print("=== At SafePoint node %d can't find value of Field: ",
877 sfpt->_idx);
878 field->print();
879 int field_idx = C->get_alias_index(field_addr_type);
880 tty->print(" (alias_idx=%d)", field_idx);
881 } else { // Array's element
882 tty->print("=== At SafePoint node %d can't find value of array element [%d]",
883 sfpt->_idx, j);
884 }
885 tty->print(", which prevents elimination of: ");
886 if (res == NULL)
887 alloc->dump();
888 else
889 res->dump();
890 }
891 #endif
892 return false;
893 }
894 if (UseCompressedOops && field_type->isa_narrowoop()) {
895 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
896 // to be able scalar replace the allocation.
897 if (field_val->is_EncodeP()) {
898 field_val = field_val->in(1);
899 } else {
900 field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type()));
901 }
902 }
903 sfpt->add_req(field_val);
904 }
905 JVMState *jvms = sfpt->jvms();
906 jvms->set_endoff(sfpt->req());
907 // Now make a pass over the debug information replacing any references
908 // to the allocated object with "sobj"
909 int start = jvms->debug_start();
910 int end = jvms->debug_end();
911 sfpt->replace_edges_in_range(res, sobj, start, end);
912 _igvn._worklist.push(sfpt);
913 safepoints_done.append_if_missing(sfpt); // keep it for rollback
914 }
915 return true;
916 }
917
disconnect_projections(MultiNode * n,PhaseIterGVN & igvn)918 static void disconnect_projections(MultiNode* n, PhaseIterGVN& igvn) {
919 Node* ctl_proj = n->proj_out_or_null(TypeFunc::Control);
920 Node* mem_proj = n->proj_out_or_null(TypeFunc::Memory);
921 if (ctl_proj != NULL) {
922 igvn.replace_node(ctl_proj, n->in(0));
923 }
924 if (mem_proj != NULL) {
925 igvn.replace_node(mem_proj, n->in(TypeFunc::Memory));
926 }
927 }
928
929 // Process users of eliminated allocation.
process_users_of_allocation(CallNode * alloc)930 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc) {
931 Node* res = alloc->result_cast();
932 if (res != NULL) {
933 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
934 Node *use = res->last_out(j);
935 uint oc1 = res->outcnt();
936
937 if (use->is_AddP()) {
938 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) {
939 Node *n = use->last_out(k);
940 uint oc2 = use->outcnt();
941 if (n->is_Store()) {
942 #ifdef ASSERT
943 // Verify that there is no dependent MemBarVolatile nodes,
944 // they should be removed during IGVN, see MemBarNode::Ideal().
945 for (DUIterator_Fast pmax, p = n->fast_outs(pmax);
946 p < pmax; p++) {
947 Node* mb = n->fast_out(p);
948 assert(mb->is_Initialize() || !mb->is_MemBar() ||
949 mb->req() <= MemBarNode::Precedent ||
950 mb->in(MemBarNode::Precedent) != n,
951 "MemBarVolatile should be eliminated for non-escaping object");
952 }
953 #endif
954 _igvn.replace_node(n, n->in(MemNode::Memory));
955 } else if (n->is_ArrayCopy()) {
956 // Disconnect ArrayCopy node
957 ArrayCopyNode* ac = n->as_ArrayCopy();
958 assert(ac->is_clonebasic(), "unexpected array copy kind");
959 Node* membar_after = ac->proj_out(TypeFunc::Control)->unique_ctrl_out();
960 disconnect_projections(ac, _igvn);
961 assert(alloc->in(0)->is_Proj() && alloc->in(0)->in(0)->Opcode() == Op_MemBarCPUOrder, "mem barrier expected before allocation");
962 Node* membar_before = alloc->in(0)->in(0);
963 disconnect_projections(membar_before->as_MemBar(), _igvn);
964 if (membar_after->is_MemBar()) {
965 disconnect_projections(membar_after->as_MemBar(), _igvn);
966 }
967 } else {
968 eliminate_gc_barrier(n);
969 }
970 k -= (oc2 - use->outcnt());
971 }
972 _igvn.remove_dead_node(use);
973 } else if (use->is_ArrayCopy()) {
974 // Disconnect ArrayCopy node
975 ArrayCopyNode* ac = use->as_ArrayCopy();
976 assert(ac->is_arraycopy_validated() ||
977 ac->is_copyof_validated() ||
978 ac->is_copyofrange_validated(), "unsupported");
979 CallProjections callprojs;
980 ac->extract_projections(&callprojs, true);
981
982 _igvn.replace_node(callprojs.fallthrough_ioproj, ac->in(TypeFunc::I_O));
983 _igvn.replace_node(callprojs.fallthrough_memproj, ac->in(TypeFunc::Memory));
984 _igvn.replace_node(callprojs.fallthrough_catchproj, ac->in(TypeFunc::Control));
985
986 // Set control to top. IGVN will remove the remaining projections
987 ac->set_req(0, top());
988 ac->replace_edge(res, top());
989
990 // Disconnect src right away: it can help find new
991 // opportunities for allocation elimination
992 Node* src = ac->in(ArrayCopyNode::Src);
993 ac->replace_edge(src, top());
994 // src can be top at this point if src and dest of the
995 // arraycopy were the same
996 if (src->outcnt() == 0 && !src->is_top()) {
997 _igvn.remove_dead_node(src);
998 }
999
1000 _igvn._worklist.push(ac);
1001 } else {
1002 eliminate_gc_barrier(use);
1003 }
1004 j -= (oc1 - res->outcnt());
1005 }
1006 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
1007 _igvn.remove_dead_node(res);
1008 }
1009
1010 //
1011 // Process other users of allocation's projections
1012 //
1013 if (_resproj != NULL && _resproj->outcnt() != 0) {
1014 // First disconnect stores captured by Initialize node.
1015 // If Initialize node is eliminated first in the following code,
1016 // it will kill such stores and DUIterator_Last will assert.
1017 for (DUIterator_Fast jmax, j = _resproj->fast_outs(jmax); j < jmax; j++) {
1018 Node *use = _resproj->fast_out(j);
1019 if (use->is_AddP()) {
1020 // raw memory addresses used only by the initialization
1021 _igvn.replace_node(use, C->top());
1022 --j; --jmax;
1023 }
1024 }
1025 for (DUIterator_Last jmin, j = _resproj->last_outs(jmin); j >= jmin; ) {
1026 Node *use = _resproj->last_out(j);
1027 uint oc1 = _resproj->outcnt();
1028 if (use->is_Initialize()) {
1029 // Eliminate Initialize node.
1030 InitializeNode *init = use->as_Initialize();
1031 assert(init->outcnt() <= 2, "only a control and memory projection expected");
1032 Node *ctrl_proj = init->proj_out_or_null(TypeFunc::Control);
1033 if (ctrl_proj != NULL) {
1034 _igvn.replace_node(ctrl_proj, init->in(TypeFunc::Control));
1035 #ifdef ASSERT
1036 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1037 Node* tmp = init->in(TypeFunc::Control);
1038 while (bs->is_gc_barrier_node(tmp)) {
1039 Node* tmp2 = bs->step_over_gc_barrier_ctrl(tmp);
1040 assert(tmp != tmp2, "Must make progress");
1041 tmp = tmp2;
1042 }
1043 assert(tmp == _fallthroughcatchproj, "allocation control projection");
1044 #endif
1045 }
1046 Node *mem_proj = init->proj_out_or_null(TypeFunc::Memory);
1047 if (mem_proj != NULL) {
1048 Node *mem = init->in(TypeFunc::Memory);
1049 #ifdef ASSERT
1050 if (mem->is_MergeMem()) {
1051 assert(mem->in(TypeFunc::Memory) == _memproj_fallthrough, "allocation memory projection");
1052 } else {
1053 assert(mem == _memproj_fallthrough, "allocation memory projection");
1054 }
1055 #endif
1056 _igvn.replace_node(mem_proj, mem);
1057 }
1058 } else {
1059 assert(false, "only Initialize or AddP expected");
1060 }
1061 j -= (oc1 - _resproj->outcnt());
1062 }
1063 }
1064 if (_fallthroughcatchproj != NULL) {
1065 _igvn.replace_node(_fallthroughcatchproj, alloc->in(TypeFunc::Control));
1066 }
1067 if (_memproj_fallthrough != NULL) {
1068 _igvn.replace_node(_memproj_fallthrough, alloc->in(TypeFunc::Memory));
1069 }
1070 if (_memproj_catchall != NULL) {
1071 _igvn.replace_node(_memproj_catchall, C->top());
1072 }
1073 if (_ioproj_fallthrough != NULL) {
1074 _igvn.replace_node(_ioproj_fallthrough, alloc->in(TypeFunc::I_O));
1075 }
1076 if (_ioproj_catchall != NULL) {
1077 _igvn.replace_node(_ioproj_catchall, C->top());
1078 }
1079 if (_catchallcatchproj != NULL) {
1080 _igvn.replace_node(_catchallcatchproj, C->top());
1081 }
1082 }
1083
eliminate_allocate_node(AllocateNode * alloc)1084 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) {
1085 // Don't do scalar replacement if the frame can be popped by JVMTI:
1086 // if reallocation fails during deoptimization we'll pop all
1087 // interpreter frames for this compiled frame and that won't play
1088 // nice with JVMTI popframe.
1089 if (!EliminateAllocations || JvmtiExport::can_pop_frame() || !alloc->_is_non_escaping) {
1090 return false;
1091 }
1092 Node* klass = alloc->in(AllocateNode::KlassNode);
1093 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr();
1094 Node* res = alloc->result_cast();
1095 // Eliminate boxing allocations which are not used
1096 // regardless scalar replacable status.
1097 bool boxing_alloc = C->eliminate_boxing() &&
1098 tklass->klass()->is_instance_klass() &&
1099 tklass->klass()->as_instance_klass()->is_box_klass();
1100 if (!alloc->_is_scalar_replaceable && (!boxing_alloc || (res != NULL))) {
1101 return false;
1102 }
1103
1104 extract_call_projections(alloc);
1105
1106 GrowableArray <SafePointNode *> safepoints;
1107 if (!can_eliminate_allocation(alloc, safepoints)) {
1108 return false;
1109 }
1110
1111 if (!alloc->_is_scalar_replaceable) {
1112 assert(res == NULL, "sanity");
1113 // We can only eliminate allocation if all debug info references
1114 // are already replaced with SafePointScalarObject because
1115 // we can't search for a fields value without instance_id.
1116 if (safepoints.length() > 0) {
1117 return false;
1118 }
1119 }
1120
1121 if (!scalar_replacement(alloc, safepoints)) {
1122 return false;
1123 }
1124
1125 CompileLog* log = C->log();
1126 if (log != NULL) {
1127 log->head("eliminate_allocation type='%d'",
1128 log->identify(tklass->klass()));
1129 JVMState* p = alloc->jvms();
1130 while (p != NULL) {
1131 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1132 p = p->caller();
1133 }
1134 log->tail("eliminate_allocation");
1135 }
1136
1137 process_users_of_allocation(alloc);
1138
1139 #ifndef PRODUCT
1140 if (PrintEliminateAllocations) {
1141 if (alloc->is_AllocateArray())
1142 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1143 else
1144 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1145 }
1146 #endif
1147
1148 return true;
1149 }
1150
eliminate_boxing_node(CallStaticJavaNode * boxing)1151 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) {
1152 // EA should remove all uses of non-escaping boxing node.
1153 if (!C->eliminate_boxing() || boxing->proj_out_or_null(TypeFunc::Parms) != NULL) {
1154 return false;
1155 }
1156
1157 assert(boxing->result_cast() == NULL, "unexpected boxing node result");
1158
1159 extract_call_projections(boxing);
1160
1161 const TypeTuple* r = boxing->tf()->range();
1162 assert(r->cnt() > TypeFunc::Parms, "sanity");
1163 const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr();
1164 assert(t != NULL, "sanity");
1165
1166 CompileLog* log = C->log();
1167 if (log != NULL) {
1168 log->head("eliminate_boxing type='%d'",
1169 log->identify(t->klass()));
1170 JVMState* p = boxing->jvms();
1171 while (p != NULL) {
1172 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1173 p = p->caller();
1174 }
1175 log->tail("eliminate_boxing");
1176 }
1177
1178 process_users_of_allocation(boxing);
1179
1180 #ifndef PRODUCT
1181 if (PrintEliminateAllocations) {
1182 tty->print("++++ Eliminated: %d ", boxing->_idx);
1183 boxing->method()->print_short_name(tty);
1184 tty->cr();
1185 }
1186 #endif
1187
1188 return true;
1189 }
1190
1191 //---------------------------set_eden_pointers-------------------------
set_eden_pointers(Node * & eden_top_adr,Node * & eden_end_adr)1192 void PhaseMacroExpand::set_eden_pointers(Node* &eden_top_adr, Node* &eden_end_adr) {
1193 if (UseTLAB) { // Private allocation: load from TLS
1194 Node* thread = transform_later(new ThreadLocalNode());
1195 int tlab_top_offset = in_bytes(JavaThread::tlab_top_offset());
1196 int tlab_end_offset = in_bytes(JavaThread::tlab_end_offset());
1197 eden_top_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_top_offset);
1198 eden_end_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_end_offset);
1199 } else { // Shared allocation: load from globals
1200 CollectedHeap* ch = Universe::heap();
1201 address top_adr = (address)ch->top_addr();
1202 address end_adr = (address)ch->end_addr();
1203 eden_top_adr = makecon(TypeRawPtr::make(top_adr));
1204 eden_end_adr = basic_plus_adr(eden_top_adr, end_adr - top_adr);
1205 }
1206 }
1207
1208
make_load(Node * ctl,Node * mem,Node * base,int offset,const Type * value_type,BasicType bt)1209 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) {
1210 Node* adr = basic_plus_adr(base, offset);
1211 const TypePtr* adr_type = adr->bottom_type()->is_ptr();
1212 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt, MemNode::unordered);
1213 transform_later(value);
1214 return value;
1215 }
1216
1217
make_store(Node * ctl,Node * mem,Node * base,int offset,Node * value,BasicType bt)1218 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) {
1219 Node* adr = basic_plus_adr(base, offset);
1220 mem = StoreNode::make(_igvn, ctl, mem, adr, NULL, value, bt, MemNode::unordered);
1221 transform_later(mem);
1222 return mem;
1223 }
1224
1225 //=============================================================================
1226 //
1227 // A L L O C A T I O N
1228 //
1229 // Allocation attempts to be fast in the case of frequent small objects.
1230 // It breaks down like this:
1231 //
1232 // 1) Size in doublewords is computed. This is a constant for objects and
1233 // variable for most arrays. Doubleword units are used to avoid size
1234 // overflow of huge doubleword arrays. We need doublewords in the end for
1235 // rounding.
1236 //
1237 // 2) Size is checked for being 'too large'. Too-large allocations will go
1238 // the slow path into the VM. The slow path can throw any required
1239 // exceptions, and does all the special checks for very large arrays. The
1240 // size test can constant-fold away for objects. For objects with
1241 // finalizers it constant-folds the otherway: you always go slow with
1242 // finalizers.
1243 //
1244 // 3) If NOT using TLABs, this is the contended loop-back point.
1245 // Load-Locked the heap top. If using TLABs normal-load the heap top.
1246 //
1247 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route.
1248 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish
1249 // "size*8" we always enter the VM, where "largish" is a constant picked small
1250 // enough that there's always space between the eden max and 4Gig (old space is
1251 // there so it's quite large) and large enough that the cost of entering the VM
1252 // is dwarfed by the cost to initialize the space.
1253 //
1254 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back
1255 // down. If contended, repeat at step 3. If using TLABs normal-store
1256 // adjusted heap top back down; there is no contention.
1257 //
1258 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark
1259 // fields.
1260 //
1261 // 7) Merge with the slow-path; cast the raw memory pointer to the correct
1262 // oop flavor.
1263 //
1264 //=============================================================================
1265 // FastAllocateSizeLimit value is in DOUBLEWORDS.
1266 // Allocations bigger than this always go the slow route.
1267 // This value must be small enough that allocation attempts that need to
1268 // trigger exceptions go the slow route. Also, it must be small enough so
1269 // that heap_top + size_in_bytes does not wrap around the 4Gig limit.
1270 //=============================================================================j//
1271 // %%% Here is an old comment from parseHelper.cpp; is it outdated?
1272 // The allocator will coalesce int->oop copies away. See comment in
1273 // coalesce.cpp about how this works. It depends critically on the exact
1274 // code shape produced here, so if you are changing this code shape
1275 // make sure the GC info for the heap-top is correct in and around the
1276 // slow-path call.
1277 //
1278
expand_allocate_common(AllocateNode * alloc,Node * length,const TypeFunc * slow_call_type,address slow_call_address)1279 void PhaseMacroExpand::expand_allocate_common(
1280 AllocateNode* alloc, // allocation node to be expanded
1281 Node* length, // array length for an array allocation
1282 const TypeFunc* slow_call_type, // Type of slow call
1283 address slow_call_address // Address of slow call
1284 )
1285 {
1286
1287 Node* ctrl = alloc->in(TypeFunc::Control);
1288 Node* mem = alloc->in(TypeFunc::Memory);
1289 Node* i_o = alloc->in(TypeFunc::I_O);
1290 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize);
1291 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1292 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest);
1293
1294 assert(ctrl != NULL, "must have control");
1295 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results.
1296 // they will not be used if "always_slow" is set
1297 enum { slow_result_path = 1, fast_result_path = 2 };
1298 Node *result_region = NULL;
1299 Node *result_phi_rawmem = NULL;
1300 Node *result_phi_rawoop = NULL;
1301 Node *result_phi_i_o = NULL;
1302
1303 // The initial slow comparison is a size check, the comparison
1304 // we want to do is a BoolTest::gt
1305 bool always_slow = false;
1306 int tv = _igvn.find_int_con(initial_slow_test, -1);
1307 if (tv >= 0) {
1308 always_slow = (tv == 1);
1309 initial_slow_test = NULL;
1310 } else {
1311 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn);
1312 }
1313
1314 if (C->env()->dtrace_alloc_probes() ||
1315 (!UseTLAB && !Universe::heap()->supports_inline_contig_alloc())) {
1316 // Force slow-path allocation
1317 always_slow = true;
1318 initial_slow_test = NULL;
1319 }
1320
1321
1322 enum { too_big_or_final_path = 1, need_gc_path = 2 };
1323 Node *slow_region = NULL;
1324 Node *toobig_false = ctrl;
1325
1326 assert (initial_slow_test == NULL || !always_slow, "arguments must be consistent");
1327 // generate the initial test if necessary
1328 if (initial_slow_test != NULL ) {
1329 slow_region = new RegionNode(3);
1330
1331 // Now make the initial failure test. Usually a too-big test but
1332 // might be a TRUE for finalizers or a fancy class check for
1333 // newInstance0.
1334 IfNode *toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
1335 transform_later(toobig_iff);
1336 // Plug the failing-too-big test into the slow-path region
1337 Node *toobig_true = new IfTrueNode( toobig_iff );
1338 transform_later(toobig_true);
1339 slow_region ->init_req( too_big_or_final_path, toobig_true );
1340 toobig_false = new IfFalseNode( toobig_iff );
1341 transform_later(toobig_false);
1342 } else { // No initial test, just fall into next case
1343 toobig_false = ctrl;
1344 debug_only(slow_region = NodeSentinel);
1345 }
1346
1347 Node *slow_mem = mem; // save the current memory state for slow path
1348 // generate the fast allocation code unless we know that the initial test will always go slow
1349 if (!always_slow) {
1350 // Fast path modifies only raw memory.
1351 if (mem->is_MergeMem()) {
1352 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1353 }
1354
1355 // allocate the Region and Phi nodes for the result
1356 result_region = new RegionNode(3);
1357 result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1358 result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM);
1359 result_phi_i_o = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch
1360
1361 // Grab regular I/O before optional prefetch may change it.
1362 // Slow-path does no I/O so just set it to the original I/O.
1363 result_phi_i_o->init_req(slow_result_path, i_o);
1364
1365 Node* needgc_ctrl = NULL;
1366 // Name successful fast-path variables
1367 Node* fast_oop_ctrl;
1368 Node* fast_oop_rawmem;
1369
1370 intx prefetch_lines = length != NULL ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1371
1372 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1373 Node* fast_oop = bs->obj_allocate(this, ctrl, mem, toobig_false, size_in_bytes, i_o, needgc_ctrl,
1374 fast_oop_ctrl, fast_oop_rawmem,
1375 prefetch_lines);
1376
1377 if (initial_slow_test) {
1378 slow_region->init_req(need_gc_path, needgc_ctrl);
1379 // This completes all paths into the slow merge point
1380 transform_later(slow_region);
1381 } else { // No initial slow path needed!
1382 // Just fall from the need-GC path straight into the VM call.
1383 slow_region = needgc_ctrl;
1384 }
1385
1386 InitializeNode* init = alloc->initialization();
1387 fast_oop_rawmem = initialize_object(alloc,
1388 fast_oop_ctrl, fast_oop_rawmem, fast_oop,
1389 klass_node, length, size_in_bytes);
1390
1391 // If initialization is performed by an array copy, any required
1392 // MemBarStoreStore was already added. If the object does not
1393 // escape no need for a MemBarStoreStore. If the object does not
1394 // escape in its initializer and memory barrier (MemBarStoreStore or
1395 // stronger) is already added at exit of initializer, also no need
1396 // for a MemBarStoreStore. Otherwise we need a MemBarStoreStore
1397 // so that stores that initialize this object can't be reordered
1398 // with a subsequent store that makes this object accessible by
1399 // other threads.
1400 // Other threads include java threads and JVM internal threads
1401 // (for example concurrent GC threads). Current concurrent GC
1402 // implementation: CMS and G1 will not scan newly created object,
1403 // so it's safe to skip storestore barrier when allocation does
1404 // not escape.
1405 if (!alloc->does_not_escape_thread() &&
1406 !alloc->is_allocation_MemBar_redundant() &&
1407 (init == NULL || !init->is_complete_with_arraycopy())) {
1408 if (init == NULL || init->req() < InitializeNode::RawStores) {
1409 // No InitializeNode or no stores captured by zeroing
1410 // elimination. Simply add the MemBarStoreStore after object
1411 // initialization.
1412 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1413 transform_later(mb);
1414
1415 mb->init_req(TypeFunc::Memory, fast_oop_rawmem);
1416 mb->init_req(TypeFunc::Control, fast_oop_ctrl);
1417 fast_oop_ctrl = new ProjNode(mb,TypeFunc::Control);
1418 transform_later(fast_oop_ctrl);
1419 fast_oop_rawmem = new ProjNode(mb,TypeFunc::Memory);
1420 transform_later(fast_oop_rawmem);
1421 } else {
1422 // Add the MemBarStoreStore after the InitializeNode so that
1423 // all stores performing the initialization that were moved
1424 // before the InitializeNode happen before the storestore
1425 // barrier.
1426
1427 Node* init_ctrl = init->proj_out_or_null(TypeFunc::Control);
1428 Node* init_mem = init->proj_out_or_null(TypeFunc::Memory);
1429
1430 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1431 transform_later(mb);
1432
1433 Node* ctrl = new ProjNode(init,TypeFunc::Control);
1434 transform_later(ctrl);
1435 Node* mem = new ProjNode(init,TypeFunc::Memory);
1436 transform_later(mem);
1437
1438 // The MemBarStoreStore depends on control and memory coming
1439 // from the InitializeNode
1440 mb->init_req(TypeFunc::Memory, mem);
1441 mb->init_req(TypeFunc::Control, ctrl);
1442
1443 ctrl = new ProjNode(mb,TypeFunc::Control);
1444 transform_later(ctrl);
1445 mem = new ProjNode(mb,TypeFunc::Memory);
1446 transform_later(mem);
1447
1448 // All nodes that depended on the InitializeNode for control
1449 // and memory must now depend on the MemBarNode that itself
1450 // depends on the InitializeNode
1451 if (init_ctrl != NULL) {
1452 _igvn.replace_node(init_ctrl, ctrl);
1453 }
1454 if (init_mem != NULL) {
1455 _igvn.replace_node(init_mem, mem);
1456 }
1457 }
1458 }
1459
1460 if (C->env()->dtrace_extended_probes()) {
1461 // Slow-path call
1462 int size = TypeFunc::Parms + 2;
1463 CallLeafNode *call = new CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(),
1464 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_object_alloc_base),
1465 "dtrace_object_alloc",
1466 TypeRawPtr::BOTTOM);
1467
1468 // Get base of thread-local storage area
1469 Node* thread = new ThreadLocalNode();
1470 transform_later(thread);
1471
1472 call->init_req(TypeFunc::Parms+0, thread);
1473 call->init_req(TypeFunc::Parms+1, fast_oop);
1474 call->init_req(TypeFunc::Control, fast_oop_ctrl);
1475 call->init_req(TypeFunc::I_O , top()); // does no i/o
1476 call->init_req(TypeFunc::Memory , fast_oop_rawmem);
1477 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1478 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1479 transform_later(call);
1480 fast_oop_ctrl = new ProjNode(call,TypeFunc::Control);
1481 transform_later(fast_oop_ctrl);
1482 fast_oop_rawmem = new ProjNode(call,TypeFunc::Memory);
1483 transform_later(fast_oop_rawmem);
1484 }
1485
1486 // Plug in the successful fast-path into the result merge point
1487 result_region ->init_req(fast_result_path, fast_oop_ctrl);
1488 result_phi_rawoop->init_req(fast_result_path, fast_oop);
1489 result_phi_i_o ->init_req(fast_result_path, i_o);
1490 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem);
1491 } else {
1492 slow_region = ctrl;
1493 result_phi_i_o = i_o; // Rename it to use in the following code.
1494 }
1495
1496 // Generate slow-path call
1497 CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address,
1498 OptoRuntime::stub_name(slow_call_address),
1499 alloc->jvms()->bci(),
1500 TypePtr::BOTTOM);
1501 call->init_req( TypeFunc::Control, slow_region );
1502 call->init_req( TypeFunc::I_O , top() ) ; // does no i/o
1503 call->init_req( TypeFunc::Memory , slow_mem ); // may gc ptrs
1504 call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) );
1505 call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) );
1506
1507 call->init_req(TypeFunc::Parms+0, klass_node);
1508 if (length != NULL) {
1509 call->init_req(TypeFunc::Parms+1, length);
1510 }
1511
1512 // Copy debug information and adjust JVMState information, then replace
1513 // allocate node with the call
1514 copy_call_debug_info((CallNode *) alloc, call);
1515 if (!always_slow) {
1516 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1517 } else {
1518 // Hook i_o projection to avoid its elimination during allocation
1519 // replacement (when only a slow call is generated).
1520 call->set_req(TypeFunc::I_O, result_phi_i_o);
1521 }
1522 _igvn.replace_node(alloc, call);
1523 transform_later(call);
1524
1525 // Identify the output projections from the allocate node and
1526 // adjust any references to them.
1527 // The control and io projections look like:
1528 //
1529 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl)
1530 // Allocate Catch
1531 // ^---Proj(io) <-------+ ^---CatchProj(io)
1532 //
1533 // We are interested in the CatchProj nodes.
1534 //
1535 extract_call_projections(call);
1536
1537 // An allocate node has separate memory projections for the uses on
1538 // the control and i_o paths. Replace the control memory projection with
1539 // result_phi_rawmem (unless we are only generating a slow call when
1540 // both memory projections are combined)
1541 if (!always_slow && _memproj_fallthrough != NULL) {
1542 for (DUIterator_Fast imax, i = _memproj_fallthrough->fast_outs(imax); i < imax; i++) {
1543 Node *use = _memproj_fallthrough->fast_out(i);
1544 _igvn.rehash_node_delayed(use);
1545 imax -= replace_input(use, _memproj_fallthrough, result_phi_rawmem);
1546 // back up iterator
1547 --i;
1548 }
1549 }
1550 // Now change uses of _memproj_catchall to use _memproj_fallthrough and delete
1551 // _memproj_catchall so we end up with a call that has only 1 memory projection.
1552 if (_memproj_catchall != NULL ) {
1553 if (_memproj_fallthrough == NULL) {
1554 _memproj_fallthrough = new ProjNode(call, TypeFunc::Memory);
1555 transform_later(_memproj_fallthrough);
1556 }
1557 for (DUIterator_Fast imax, i = _memproj_catchall->fast_outs(imax); i < imax; i++) {
1558 Node *use = _memproj_catchall->fast_out(i);
1559 _igvn.rehash_node_delayed(use);
1560 imax -= replace_input(use, _memproj_catchall, _memproj_fallthrough);
1561 // back up iterator
1562 --i;
1563 }
1564 assert(_memproj_catchall->outcnt() == 0, "all uses must be deleted");
1565 _igvn.remove_dead_node(_memproj_catchall);
1566 }
1567
1568 // An allocate node has separate i_o projections for the uses on the control
1569 // and i_o paths. Always replace the control i_o projection with result i_o
1570 // otherwise incoming i_o become dead when only a slow call is generated
1571 // (it is different from memory projections where both projections are
1572 // combined in such case).
1573 if (_ioproj_fallthrough != NULL) {
1574 for (DUIterator_Fast imax, i = _ioproj_fallthrough->fast_outs(imax); i < imax; i++) {
1575 Node *use = _ioproj_fallthrough->fast_out(i);
1576 _igvn.rehash_node_delayed(use);
1577 imax -= replace_input(use, _ioproj_fallthrough, result_phi_i_o);
1578 // back up iterator
1579 --i;
1580 }
1581 }
1582 // Now change uses of _ioproj_catchall to use _ioproj_fallthrough and delete
1583 // _ioproj_catchall so we end up with a call that has only 1 i_o projection.
1584 if (_ioproj_catchall != NULL ) {
1585 if (_ioproj_fallthrough == NULL) {
1586 _ioproj_fallthrough = new ProjNode(call, TypeFunc::I_O);
1587 transform_later(_ioproj_fallthrough);
1588 }
1589 for (DUIterator_Fast imax, i = _ioproj_catchall->fast_outs(imax); i < imax; i++) {
1590 Node *use = _ioproj_catchall->fast_out(i);
1591 _igvn.rehash_node_delayed(use);
1592 imax -= replace_input(use, _ioproj_catchall, _ioproj_fallthrough);
1593 // back up iterator
1594 --i;
1595 }
1596 assert(_ioproj_catchall->outcnt() == 0, "all uses must be deleted");
1597 _igvn.remove_dead_node(_ioproj_catchall);
1598 }
1599
1600 // if we generated only a slow call, we are done
1601 if (always_slow) {
1602 // Now we can unhook i_o.
1603 if (result_phi_i_o->outcnt() > 1) {
1604 call->set_req(TypeFunc::I_O, top());
1605 } else {
1606 assert(result_phi_i_o->unique_ctrl_out() == call, "");
1607 // Case of new array with negative size known during compilation.
1608 // AllocateArrayNode::Ideal() optimization disconnect unreachable
1609 // following code since call to runtime will throw exception.
1610 // As result there will be no users of i_o after the call.
1611 // Leave i_o attached to this call to avoid problems in preceding graph.
1612 }
1613 return;
1614 }
1615
1616
1617 if (_fallthroughcatchproj != NULL) {
1618 ctrl = _fallthroughcatchproj->clone();
1619 transform_later(ctrl);
1620 _igvn.replace_node(_fallthroughcatchproj, result_region);
1621 } else {
1622 ctrl = top();
1623 }
1624 Node *slow_result;
1625 if (_resproj == NULL) {
1626 // no uses of the allocation result
1627 slow_result = top();
1628 } else {
1629 slow_result = _resproj->clone();
1630 transform_later(slow_result);
1631 _igvn.replace_node(_resproj, result_phi_rawoop);
1632 }
1633
1634 // Plug slow-path into result merge point
1635 result_region ->init_req( slow_result_path, ctrl );
1636 result_phi_rawoop->init_req( slow_result_path, slow_result);
1637 result_phi_rawmem->init_req( slow_result_path, _memproj_fallthrough );
1638 transform_later(result_region);
1639 transform_later(result_phi_rawoop);
1640 transform_later(result_phi_rawmem);
1641 transform_later(result_phi_i_o);
1642 // This completes all paths into the result merge point
1643 }
1644
1645
1646 // Helper for PhaseMacroExpand::expand_allocate_common.
1647 // Initializes the newly-allocated storage.
1648 Node*
initialize_object(AllocateNode * alloc,Node * control,Node * rawmem,Node * object,Node * klass_node,Node * length,Node * size_in_bytes)1649 PhaseMacroExpand::initialize_object(AllocateNode* alloc,
1650 Node* control, Node* rawmem, Node* object,
1651 Node* klass_node, Node* length,
1652 Node* size_in_bytes) {
1653 InitializeNode* init = alloc->initialization();
1654 // Store the klass & mark bits
1655 Node* mark_node = NULL;
1656 // For now only enable fast locking for non-array types
1657 if (UseBiasedLocking && (length == NULL)) {
1658 mark_node = make_load(control, rawmem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
1659 } else {
1660 mark_node = makecon(TypeRawPtr::make((address)markOopDesc::prototype()));
1661 }
1662 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS);
1663
1664 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
1665 int header_size = alloc->minimum_header_size(); // conservatively small
1666
1667 // Array length
1668 if (length != NULL) { // Arrays need length field
1669 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
1670 // conservatively small header size:
1671 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1672 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
1673 if (k->is_array_klass()) // we know the exact header size in most cases:
1674 header_size = Klass::layout_helper_header_size(k->layout_helper());
1675 }
1676
1677 // Clear the object body, if necessary.
1678 if (init == NULL) {
1679 // The init has somehow disappeared; be cautious and clear everything.
1680 //
1681 // This can happen if a node is allocated but an uncommon trap occurs
1682 // immediately. In this case, the Initialize gets associated with the
1683 // trap, and may be placed in a different (outer) loop, if the Allocate
1684 // is in a loop. If (this is rare) the inner loop gets unrolled, then
1685 // there can be two Allocates to one Initialize. The answer in all these
1686 // edge cases is safety first. It is always safe to clear immediately
1687 // within an Allocate, and then (maybe or maybe not) clear some more later.
1688 if (!(UseTLAB && ZeroTLAB)) {
1689 rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
1690 header_size, size_in_bytes,
1691 &_igvn);
1692 }
1693 } else {
1694 if (!init->is_complete()) {
1695 // Try to win by zeroing only what the init does not store.
1696 // We can also try to do some peephole optimizations,
1697 // such as combining some adjacent subword stores.
1698 rawmem = init->complete_stores(control, rawmem, object,
1699 header_size, size_in_bytes, &_igvn);
1700 }
1701 // We have no more use for this link, since the AllocateNode goes away:
1702 init->set_req(InitializeNode::RawAddress, top());
1703 // (If we keep the link, it just confuses the register allocator,
1704 // who thinks he sees a real use of the address by the membar.)
1705 }
1706
1707 return rawmem;
1708 }
1709
1710 // Generate prefetch instructions for next allocations.
prefetch_allocation(Node * i_o,Node * & needgc_false,Node * & contended_phi_rawmem,Node * old_eden_top,Node * new_eden_top,intx lines)1711 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false,
1712 Node*& contended_phi_rawmem,
1713 Node* old_eden_top, Node* new_eden_top,
1714 intx lines) {
1715 enum { fall_in_path = 1, pf_path = 2 };
1716 if( UseTLAB && AllocatePrefetchStyle == 2 ) {
1717 // Generate prefetch allocation with watermark check.
1718 // As an allocation hits the watermark, we will prefetch starting
1719 // at a "distance" away from watermark.
1720
1721 Node *pf_region = new RegionNode(3);
1722 Node *pf_phi_rawmem = new PhiNode( pf_region, Type::MEMORY,
1723 TypeRawPtr::BOTTOM );
1724 // I/O is used for Prefetch
1725 Node *pf_phi_abio = new PhiNode( pf_region, Type::ABIO );
1726
1727 Node *thread = new ThreadLocalNode();
1728 transform_later(thread);
1729
1730 Node *eden_pf_adr = new AddPNode( top()/*not oop*/, thread,
1731 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) );
1732 transform_later(eden_pf_adr);
1733
1734 Node *old_pf_wm = new LoadPNode(needgc_false,
1735 contended_phi_rawmem, eden_pf_adr,
1736 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM,
1737 MemNode::unordered);
1738 transform_later(old_pf_wm);
1739
1740 // check against new_eden_top
1741 Node *need_pf_cmp = new CmpPNode( new_eden_top, old_pf_wm );
1742 transform_later(need_pf_cmp);
1743 Node *need_pf_bol = new BoolNode( need_pf_cmp, BoolTest::ge );
1744 transform_later(need_pf_bol);
1745 IfNode *need_pf_iff = new IfNode( needgc_false, need_pf_bol,
1746 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN );
1747 transform_later(need_pf_iff);
1748
1749 // true node, add prefetchdistance
1750 Node *need_pf_true = new IfTrueNode( need_pf_iff );
1751 transform_later(need_pf_true);
1752
1753 Node *need_pf_false = new IfFalseNode( need_pf_iff );
1754 transform_later(need_pf_false);
1755
1756 Node *new_pf_wmt = new AddPNode( top(), old_pf_wm,
1757 _igvn.MakeConX(AllocatePrefetchDistance) );
1758 transform_later(new_pf_wmt );
1759 new_pf_wmt->set_req(0, need_pf_true);
1760
1761 Node *store_new_wmt = new StorePNode(need_pf_true,
1762 contended_phi_rawmem, eden_pf_adr,
1763 TypeRawPtr::BOTTOM, new_pf_wmt,
1764 MemNode::unordered);
1765 transform_later(store_new_wmt);
1766
1767 // adding prefetches
1768 pf_phi_abio->init_req( fall_in_path, i_o );
1769
1770 Node *prefetch_adr;
1771 Node *prefetch;
1772 uint step_size = AllocatePrefetchStepSize;
1773 uint distance = 0;
1774
1775 for ( intx i = 0; i < lines; i++ ) {
1776 prefetch_adr = new AddPNode( old_pf_wm, new_pf_wmt,
1777 _igvn.MakeConX(distance) );
1778 transform_later(prefetch_adr);
1779 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
1780 transform_later(prefetch);
1781 distance += step_size;
1782 i_o = prefetch;
1783 }
1784 pf_phi_abio->set_req( pf_path, i_o );
1785
1786 pf_region->init_req( fall_in_path, need_pf_false );
1787 pf_region->init_req( pf_path, need_pf_true );
1788
1789 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem );
1790 pf_phi_rawmem->init_req( pf_path, store_new_wmt );
1791
1792 transform_later(pf_region);
1793 transform_later(pf_phi_rawmem);
1794 transform_later(pf_phi_abio);
1795
1796 needgc_false = pf_region;
1797 contended_phi_rawmem = pf_phi_rawmem;
1798 i_o = pf_phi_abio;
1799 } else if( UseTLAB && AllocatePrefetchStyle == 3 ) {
1800 // Insert a prefetch instruction for each allocation.
1801 // This code is used to generate 1 prefetch instruction per cache line.
1802
1803 // Generate several prefetch instructions.
1804 uint step_size = AllocatePrefetchStepSize;
1805 uint distance = AllocatePrefetchDistance;
1806
1807 // Next cache address.
1808 Node *cache_adr = new AddPNode(old_eden_top, old_eden_top,
1809 _igvn.MakeConX(step_size + distance));
1810 transform_later(cache_adr);
1811 cache_adr = new CastP2XNode(needgc_false, cache_adr);
1812 transform_later(cache_adr);
1813 // Address is aligned to execute prefetch to the beginning of cache line size
1814 // (it is important when BIS instruction is used on SPARC as prefetch).
1815 Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1));
1816 cache_adr = new AndXNode(cache_adr, mask);
1817 transform_later(cache_adr);
1818 cache_adr = new CastX2PNode(cache_adr);
1819 transform_later(cache_adr);
1820
1821 // Prefetch
1822 Node *prefetch = new PrefetchAllocationNode( contended_phi_rawmem, cache_adr );
1823 prefetch->set_req(0, needgc_false);
1824 transform_later(prefetch);
1825 contended_phi_rawmem = prefetch;
1826 Node *prefetch_adr;
1827 distance = step_size;
1828 for ( intx i = 1; i < lines; i++ ) {
1829 prefetch_adr = new AddPNode( cache_adr, cache_adr,
1830 _igvn.MakeConX(distance) );
1831 transform_later(prefetch_adr);
1832 prefetch = new PrefetchAllocationNode( contended_phi_rawmem, prefetch_adr );
1833 transform_later(prefetch);
1834 distance += step_size;
1835 contended_phi_rawmem = prefetch;
1836 }
1837 } else if( AllocatePrefetchStyle > 0 ) {
1838 // Insert a prefetch for each allocation only on the fast-path
1839 Node *prefetch_adr;
1840 Node *prefetch;
1841 // Generate several prefetch instructions.
1842 uint step_size = AllocatePrefetchStepSize;
1843 uint distance = AllocatePrefetchDistance;
1844 for ( intx i = 0; i < lines; i++ ) {
1845 prefetch_adr = new AddPNode( old_eden_top, new_eden_top,
1846 _igvn.MakeConX(distance) );
1847 transform_later(prefetch_adr);
1848 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
1849 // Do not let it float too high, since if eden_top == eden_end,
1850 // both might be null.
1851 if( i == 0 ) { // Set control for first prefetch, next follows it
1852 prefetch->init_req(0, needgc_false);
1853 }
1854 transform_later(prefetch);
1855 distance += step_size;
1856 i_o = prefetch;
1857 }
1858 }
1859 return i_o;
1860 }
1861
1862
expand_allocate(AllocateNode * alloc)1863 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) {
1864 expand_allocate_common(alloc, NULL,
1865 OptoRuntime::new_instance_Type(),
1866 OptoRuntime::new_instance_Java());
1867 }
1868
expand_allocate_array(AllocateArrayNode * alloc)1869 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) {
1870 Node* length = alloc->in(AllocateNode::ALength);
1871 InitializeNode* init = alloc->initialization();
1872 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1873 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
1874 address slow_call_address; // Address of slow call
1875 if (init != NULL && init->is_complete_with_arraycopy() &&
1876 k->is_type_array_klass()) {
1877 // Don't zero type array during slow allocation in VM since
1878 // it will be initialized later by arraycopy in compiled code.
1879 slow_call_address = OptoRuntime::new_array_nozero_Java();
1880 } else {
1881 slow_call_address = OptoRuntime::new_array_Java();
1882 }
1883 expand_allocate_common(alloc, length,
1884 OptoRuntime::new_array_Type(),
1885 slow_call_address);
1886 }
1887
1888 //-------------------mark_eliminated_box----------------------------------
1889 //
1890 // During EA obj may point to several objects but after few ideal graph
1891 // transformations (CCP) it may point to only one non escaping object
1892 // (but still using phi), corresponding locks and unlocks will be marked
1893 // for elimination. Later obj could be replaced with a new node (new phi)
1894 // and which does not have escape information. And later after some graph
1895 // reshape other locks and unlocks (which were not marked for elimination
1896 // before) are connected to this new obj (phi) but they still will not be
1897 // marked for elimination since new obj has no escape information.
1898 // Mark all associated (same box and obj) lock and unlock nodes for
1899 // elimination if some of them marked already.
mark_eliminated_box(Node * oldbox,Node * obj)1900 void PhaseMacroExpand::mark_eliminated_box(Node* oldbox, Node* obj) {
1901 if (oldbox->as_BoxLock()->is_eliminated())
1902 return; // This BoxLock node was processed already.
1903
1904 // New implementation (EliminateNestedLocks) has separate BoxLock
1905 // node for each locked region so mark all associated locks/unlocks as
1906 // eliminated even if different objects are referenced in one locked region
1907 // (for example, OSR compilation of nested loop inside locked scope).
1908 if (EliminateNestedLocks ||
1909 oldbox->as_BoxLock()->is_simple_lock_region(NULL, obj)) {
1910 // Box is used only in one lock region. Mark this box as eliminated.
1911 _igvn.hash_delete(oldbox);
1912 oldbox->as_BoxLock()->set_eliminated(); // This changes box's hash value
1913 _igvn.hash_insert(oldbox);
1914
1915 for (uint i = 0; i < oldbox->outcnt(); i++) {
1916 Node* u = oldbox->raw_out(i);
1917 if (u->is_AbstractLock() && !u->as_AbstractLock()->is_non_esc_obj()) {
1918 AbstractLockNode* alock = u->as_AbstractLock();
1919 // Check lock's box since box could be referenced by Lock's debug info.
1920 if (alock->box_node() == oldbox) {
1921 // Mark eliminated all related locks and unlocks.
1922 #ifdef ASSERT
1923 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc4");
1924 #endif
1925 alock->set_non_esc_obj();
1926 }
1927 }
1928 }
1929 return;
1930 }
1931
1932 // Create new "eliminated" BoxLock node and use it in monitor debug info
1933 // instead of oldbox for the same object.
1934 BoxLockNode* newbox = oldbox->clone()->as_BoxLock();
1935
1936 // Note: BoxLock node is marked eliminated only here and it is used
1937 // to indicate that all associated lock and unlock nodes are marked
1938 // for elimination.
1939 newbox->set_eliminated();
1940 transform_later(newbox);
1941
1942 // Replace old box node with new box for all users of the same object.
1943 for (uint i = 0; i < oldbox->outcnt();) {
1944 bool next_edge = true;
1945
1946 Node* u = oldbox->raw_out(i);
1947 if (u->is_AbstractLock()) {
1948 AbstractLockNode* alock = u->as_AbstractLock();
1949 if (alock->box_node() == oldbox && alock->obj_node()->eqv_uncast(obj)) {
1950 // Replace Box and mark eliminated all related locks and unlocks.
1951 #ifdef ASSERT
1952 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc5");
1953 #endif
1954 alock->set_non_esc_obj();
1955 _igvn.rehash_node_delayed(alock);
1956 alock->set_box_node(newbox);
1957 next_edge = false;
1958 }
1959 }
1960 if (u->is_FastLock() && u->as_FastLock()->obj_node()->eqv_uncast(obj)) {
1961 FastLockNode* flock = u->as_FastLock();
1962 assert(flock->box_node() == oldbox, "sanity");
1963 _igvn.rehash_node_delayed(flock);
1964 flock->set_box_node(newbox);
1965 next_edge = false;
1966 }
1967
1968 // Replace old box in monitor debug info.
1969 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) {
1970 SafePointNode* sfn = u->as_SafePoint();
1971 JVMState* youngest_jvms = sfn->jvms();
1972 int max_depth = youngest_jvms->depth();
1973 for (int depth = 1; depth <= max_depth; depth++) {
1974 JVMState* jvms = youngest_jvms->of_depth(depth);
1975 int num_mon = jvms->nof_monitors();
1976 // Loop over monitors
1977 for (int idx = 0; idx < num_mon; idx++) {
1978 Node* obj_node = sfn->monitor_obj(jvms, idx);
1979 Node* box_node = sfn->monitor_box(jvms, idx);
1980 if (box_node == oldbox && obj_node->eqv_uncast(obj)) {
1981 int j = jvms->monitor_box_offset(idx);
1982 _igvn.replace_input_of(u, j, newbox);
1983 next_edge = false;
1984 }
1985 }
1986 }
1987 }
1988 if (next_edge) i++;
1989 }
1990 }
1991
1992 //-----------------------mark_eliminated_locking_nodes-----------------------
mark_eliminated_locking_nodes(AbstractLockNode * alock)1993 void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) {
1994 if (EliminateNestedLocks) {
1995 if (alock->is_nested()) {
1996 assert(alock->box_node()->as_BoxLock()->is_eliminated(), "sanity");
1997 return;
1998 } else if (!alock->is_non_esc_obj()) { // Not eliminated or coarsened
1999 // Only Lock node has JVMState needed here.
2000 // Not that preceding claim is documented anywhere else.
2001 if (alock->jvms() != NULL) {
2002 if (alock->as_Lock()->is_nested_lock_region()) {
2003 // Mark eliminated related nested locks and unlocks.
2004 Node* obj = alock->obj_node();
2005 BoxLockNode* box_node = alock->box_node()->as_BoxLock();
2006 assert(!box_node->is_eliminated(), "should not be marked yet");
2007 // Note: BoxLock node is marked eliminated only here
2008 // and it is used to indicate that all associated lock
2009 // and unlock nodes are marked for elimination.
2010 box_node->set_eliminated(); // Box's hash is always NO_HASH here
2011 for (uint i = 0; i < box_node->outcnt(); i++) {
2012 Node* u = box_node->raw_out(i);
2013 if (u->is_AbstractLock()) {
2014 alock = u->as_AbstractLock();
2015 if (alock->box_node() == box_node) {
2016 // Verify that this Box is referenced only by related locks.
2017 assert(alock->obj_node()->eqv_uncast(obj), "");
2018 // Mark all related locks and unlocks.
2019 #ifdef ASSERT
2020 alock->log_lock_optimization(C, "eliminate_lock_set_nested");
2021 #endif
2022 alock->set_nested();
2023 }
2024 }
2025 }
2026 } else {
2027 #ifdef ASSERT
2028 alock->log_lock_optimization(C, "eliminate_lock_NOT_nested_lock_region");
2029 if (C->log() != NULL)
2030 alock->as_Lock()->is_nested_lock_region(C); // rerun for debugging output
2031 #endif
2032 }
2033 }
2034 return;
2035 }
2036 // Process locks for non escaping object
2037 assert(alock->is_non_esc_obj(), "");
2038 } // EliminateNestedLocks
2039
2040 if (alock->is_non_esc_obj()) { // Lock is used for non escaping object
2041 // Look for all locks of this object and mark them and
2042 // corresponding BoxLock nodes as eliminated.
2043 Node* obj = alock->obj_node();
2044 for (uint j = 0; j < obj->outcnt(); j++) {
2045 Node* o = obj->raw_out(j);
2046 if (o->is_AbstractLock() &&
2047 o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) {
2048 alock = o->as_AbstractLock();
2049 Node* box = alock->box_node();
2050 // Replace old box node with new eliminated box for all users
2051 // of the same object and mark related locks as eliminated.
2052 mark_eliminated_box(box, obj);
2053 }
2054 }
2055 }
2056 }
2057
2058 // we have determined that this lock/unlock can be eliminated, we simply
2059 // eliminate the node without expanding it.
2060 //
2061 // Note: The membar's associated with the lock/unlock are currently not
2062 // eliminated. This should be investigated as a future enhancement.
2063 //
eliminate_locking_node(AbstractLockNode * alock)2064 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) {
2065
2066 if (!alock->is_eliminated()) {
2067 return false;
2068 }
2069 #ifdef ASSERT
2070 if (!alock->is_coarsened()) {
2071 // Check that new "eliminated" BoxLock node is created.
2072 BoxLockNode* oldbox = alock->box_node()->as_BoxLock();
2073 assert(oldbox->is_eliminated(), "should be done already");
2074 }
2075 #endif
2076
2077 alock->log_lock_optimization(C, "eliminate_lock");
2078
2079 #ifndef PRODUCT
2080 if (PrintEliminateLocks) {
2081 if (alock->is_Lock()) {
2082 tty->print_cr("++++ Eliminated: %d Lock", alock->_idx);
2083 } else {
2084 tty->print_cr("++++ Eliminated: %d Unlock", alock->_idx);
2085 }
2086 }
2087 #endif
2088
2089 Node* mem = alock->in(TypeFunc::Memory);
2090 Node* ctrl = alock->in(TypeFunc::Control);
2091 guarantee(ctrl != NULL, "missing control projection, cannot replace_node() with NULL");
2092
2093 extract_call_projections(alock);
2094 // There are 2 projections from the lock. The lock node will
2095 // be deleted when its last use is subsumed below.
2096 assert(alock->outcnt() == 2 &&
2097 _fallthroughproj != NULL &&
2098 _memproj_fallthrough != NULL,
2099 "Unexpected projections from Lock/Unlock");
2100
2101 Node* fallthroughproj = _fallthroughproj;
2102 Node* memproj_fallthrough = _memproj_fallthrough;
2103
2104 // The memory projection from a lock/unlock is RawMem
2105 // The input to a Lock is merged memory, so extract its RawMem input
2106 // (unless the MergeMem has been optimized away.)
2107 if (alock->is_Lock()) {
2108 // Seach for MemBarAcquireLock node and delete it also.
2109 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
2110 assert(membar != NULL && membar->Opcode() == Op_MemBarAcquireLock, "");
2111 Node* ctrlproj = membar->proj_out(TypeFunc::Control);
2112 Node* memproj = membar->proj_out(TypeFunc::Memory);
2113 _igvn.replace_node(ctrlproj, fallthroughproj);
2114 _igvn.replace_node(memproj, memproj_fallthrough);
2115
2116 // Delete FastLock node also if this Lock node is unique user
2117 // (a loop peeling may clone a Lock node).
2118 Node* flock = alock->as_Lock()->fastlock_node();
2119 if (flock->outcnt() == 1) {
2120 assert(flock->unique_out() == alock, "sanity");
2121 _igvn.replace_node(flock, top());
2122 }
2123 }
2124
2125 // Seach for MemBarReleaseLock node and delete it also.
2126 if (alock->is_Unlock() && ctrl->is_Proj() && ctrl->in(0)->is_MemBar()) {
2127 MemBarNode* membar = ctrl->in(0)->as_MemBar();
2128 assert(membar->Opcode() == Op_MemBarReleaseLock &&
2129 mem->is_Proj() && membar == mem->in(0), "");
2130 _igvn.replace_node(fallthroughproj, ctrl);
2131 _igvn.replace_node(memproj_fallthrough, mem);
2132 fallthroughproj = ctrl;
2133 memproj_fallthrough = mem;
2134 ctrl = membar->in(TypeFunc::Control);
2135 mem = membar->in(TypeFunc::Memory);
2136 }
2137
2138 _igvn.replace_node(fallthroughproj, ctrl);
2139 _igvn.replace_node(memproj_fallthrough, mem);
2140 return true;
2141 }
2142
2143
2144 //------------------------------expand_lock_node----------------------
expand_lock_node(LockNode * lock)2145 void PhaseMacroExpand::expand_lock_node(LockNode *lock) {
2146
2147 Node* ctrl = lock->in(TypeFunc::Control);
2148 Node* mem = lock->in(TypeFunc::Memory);
2149 Node* obj = lock->obj_node();
2150 Node* box = lock->box_node();
2151 Node* flock = lock->fastlock_node();
2152
2153 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2154
2155 // Make the merge point
2156 Node *region;
2157 Node *mem_phi;
2158 Node *slow_path;
2159
2160 if (UseOptoBiasInlining) {
2161 /*
2162 * See the full description in MacroAssembler::biased_locking_enter().
2163 *
2164 * if( (mark_word & biased_lock_mask) == biased_lock_pattern ) {
2165 * // The object is biased.
2166 * proto_node = klass->prototype_header;
2167 * o_node = thread | proto_node;
2168 * x_node = o_node ^ mark_word;
2169 * if( (x_node & ~age_mask) == 0 ) { // Biased to the current thread ?
2170 * // Done.
2171 * } else {
2172 * if( (x_node & biased_lock_mask) != 0 ) {
2173 * // The klass's prototype header is no longer biased.
2174 * cas(&mark_word, mark_word, proto_node)
2175 * goto cas_lock;
2176 * } else {
2177 * // The klass's prototype header is still biased.
2178 * if( (x_node & epoch_mask) != 0 ) { // Expired epoch?
2179 * old = mark_word;
2180 * new = o_node;
2181 * } else {
2182 * // Different thread or anonymous biased.
2183 * old = mark_word & (epoch_mask | age_mask | biased_lock_mask);
2184 * new = thread | old;
2185 * }
2186 * // Try to rebias.
2187 * if( cas(&mark_word, old, new) == 0 ) {
2188 * // Done.
2189 * } else {
2190 * goto slow_path; // Failed.
2191 * }
2192 * }
2193 * }
2194 * } else {
2195 * // The object is not biased.
2196 * cas_lock:
2197 * if( FastLock(obj) == 0 ) {
2198 * // Done.
2199 * } else {
2200 * slow_path:
2201 * OptoRuntime::complete_monitor_locking_Java(obj);
2202 * }
2203 * }
2204 */
2205
2206 region = new RegionNode(5);
2207 // create a Phi for the memory state
2208 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2209
2210 Node* fast_lock_region = new RegionNode(3);
2211 Node* fast_lock_mem_phi = new PhiNode( fast_lock_region, Type::MEMORY, TypeRawPtr::BOTTOM);
2212
2213 // First, check mark word for the biased lock pattern.
2214 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2215
2216 // Get fast path - mark word has the biased lock pattern.
2217 ctrl = opt_bits_test(ctrl, fast_lock_region, 1, mark_node,
2218 markOopDesc::biased_lock_mask_in_place,
2219 markOopDesc::biased_lock_pattern, true);
2220 // fast_lock_region->in(1) is set to slow path.
2221 fast_lock_mem_phi->init_req(1, mem);
2222
2223 // Now check that the lock is biased to the current thread and has
2224 // the same epoch and bias as Klass::_prototype_header.
2225
2226 // Special-case a fresh allocation to avoid building nodes:
2227 Node* klass_node = AllocateNode::Ideal_klass(obj, &_igvn);
2228 if (klass_node == NULL) {
2229 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
2230 klass_node = transform_later(LoadKlassNode::make(_igvn, NULL, mem, k_adr, _igvn.type(k_adr)->is_ptr()));
2231 #ifdef _LP64
2232 if (UseCompressedClassPointers && klass_node->is_DecodeNKlass()) {
2233 assert(klass_node->in(1)->Opcode() == Op_LoadNKlass, "sanity");
2234 klass_node->in(1)->init_req(0, ctrl);
2235 } else
2236 #endif
2237 klass_node->init_req(0, ctrl);
2238 }
2239 Node *proto_node = make_load(ctrl, mem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeX_X, TypeX_X->basic_type());
2240
2241 Node* thread = transform_later(new ThreadLocalNode());
2242 Node* cast_thread = transform_later(new CastP2XNode(ctrl, thread));
2243 Node* o_node = transform_later(new OrXNode(cast_thread, proto_node));
2244 Node* x_node = transform_later(new XorXNode(o_node, mark_node));
2245
2246 // Get slow path - mark word does NOT match the value.
2247 Node* not_biased_ctrl = opt_bits_test(ctrl, region, 3, x_node,
2248 (~markOopDesc::age_mask_in_place), 0);
2249 // region->in(3) is set to fast path - the object is biased to the current thread.
2250 mem_phi->init_req(3, mem);
2251
2252
2253 // Mark word does NOT match the value (thread | Klass::_prototype_header).
2254
2255
2256 // First, check biased pattern.
2257 // Get fast path - _prototype_header has the same biased lock pattern.
2258 ctrl = opt_bits_test(not_biased_ctrl, fast_lock_region, 2, x_node,
2259 markOopDesc::biased_lock_mask_in_place, 0, true);
2260
2261 not_biased_ctrl = fast_lock_region->in(2); // Slow path
2262 // fast_lock_region->in(2) - the prototype header is no longer biased
2263 // and we have to revoke the bias on this object.
2264 // We are going to try to reset the mark of this object to the prototype
2265 // value and fall through to the CAS-based locking scheme.
2266 Node* adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
2267 Node* cas = new StoreXConditionalNode(not_biased_ctrl, mem, adr,
2268 proto_node, mark_node);
2269 transform_later(cas);
2270 Node* proj = transform_later(new SCMemProjNode(cas));
2271 fast_lock_mem_phi->init_req(2, proj);
2272
2273
2274 // Second, check epoch bits.
2275 Node* rebiased_region = new RegionNode(3);
2276 Node* old_phi = new PhiNode( rebiased_region, TypeX_X);
2277 Node* new_phi = new PhiNode( rebiased_region, TypeX_X);
2278
2279 // Get slow path - mark word does NOT match epoch bits.
2280 Node* epoch_ctrl = opt_bits_test(ctrl, rebiased_region, 1, x_node,
2281 markOopDesc::epoch_mask_in_place, 0);
2282 // The epoch of the current bias is not valid, attempt to rebias the object
2283 // toward the current thread.
2284 rebiased_region->init_req(2, epoch_ctrl);
2285 old_phi->init_req(2, mark_node);
2286 new_phi->init_req(2, o_node);
2287
2288 // rebiased_region->in(1) is set to fast path.
2289 // The epoch of the current bias is still valid but we know
2290 // nothing about the owner; it might be set or it might be clear.
2291 Node* cmask = MakeConX(markOopDesc::biased_lock_mask_in_place |
2292 markOopDesc::age_mask_in_place |
2293 markOopDesc::epoch_mask_in_place);
2294 Node* old = transform_later(new AndXNode(mark_node, cmask));
2295 cast_thread = transform_later(new CastP2XNode(ctrl, thread));
2296 Node* new_mark = transform_later(new OrXNode(cast_thread, old));
2297 old_phi->init_req(1, old);
2298 new_phi->init_req(1, new_mark);
2299
2300 transform_later(rebiased_region);
2301 transform_later(old_phi);
2302 transform_later(new_phi);
2303
2304 // Try to acquire the bias of the object using an atomic operation.
2305 // If this fails we will go in to the runtime to revoke the object's bias.
2306 cas = new StoreXConditionalNode(rebiased_region, mem, adr, new_phi, old_phi);
2307 transform_later(cas);
2308 proj = transform_later(new SCMemProjNode(cas));
2309
2310 // Get slow path - Failed to CAS.
2311 not_biased_ctrl = opt_bits_test(rebiased_region, region, 4, cas, 0, 0);
2312 mem_phi->init_req(4, proj);
2313 // region->in(4) is set to fast path - the object is rebiased to the current thread.
2314
2315 // Failed to CAS.
2316 slow_path = new RegionNode(3);
2317 Node *slow_mem = new PhiNode( slow_path, Type::MEMORY, TypeRawPtr::BOTTOM);
2318
2319 slow_path->init_req(1, not_biased_ctrl); // Capture slow-control
2320 slow_mem->init_req(1, proj);
2321
2322 // Call CAS-based locking scheme (FastLock node).
2323
2324 transform_later(fast_lock_region);
2325 transform_later(fast_lock_mem_phi);
2326
2327 // Get slow path - FastLock failed to lock the object.
2328 ctrl = opt_bits_test(fast_lock_region, region, 2, flock, 0, 0);
2329 mem_phi->init_req(2, fast_lock_mem_phi);
2330 // region->in(2) is set to fast path - the object is locked to the current thread.
2331
2332 slow_path->init_req(2, ctrl); // Capture slow-control
2333 slow_mem->init_req(2, fast_lock_mem_phi);
2334
2335 transform_later(slow_path);
2336 transform_later(slow_mem);
2337 // Reset lock's memory edge.
2338 lock->set_req(TypeFunc::Memory, slow_mem);
2339
2340 } else {
2341 region = new RegionNode(3);
2342 // create a Phi for the memory state
2343 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2344
2345 // Optimize test; set region slot 2
2346 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0);
2347 mem_phi->init_req(2, mem);
2348 }
2349
2350 // Make slow path call
2351 CallNode *call = make_slow_call((CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(),
2352 OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path,
2353 obj, box, NULL);
2354
2355 extract_call_projections(call);
2356
2357 // Slow path can only throw asynchronous exceptions, which are always
2358 // de-opted. So the compiler thinks the slow-call can never throw an
2359 // exception. If it DOES throw an exception we would need the debug
2360 // info removed first (since if it throws there is no monitor).
2361 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
2362 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
2363
2364 // Capture slow path
2365 // disconnect fall-through projection from call and create a new one
2366 // hook up users of fall-through projection to region
2367 Node *slow_ctrl = _fallthroughproj->clone();
2368 transform_later(slow_ctrl);
2369 _igvn.hash_delete(_fallthroughproj);
2370 _fallthroughproj->disconnect_inputs(NULL, C);
2371 region->init_req(1, slow_ctrl);
2372 // region inputs are now complete
2373 transform_later(region);
2374 _igvn.replace_node(_fallthroughproj, region);
2375
2376 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory));
2377 mem_phi->init_req(1, memproj );
2378 transform_later(mem_phi);
2379 _igvn.replace_node(_memproj_fallthrough, mem_phi);
2380 }
2381
2382 //------------------------------expand_unlock_node----------------------
expand_unlock_node(UnlockNode * unlock)2383 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
2384
2385 Node* ctrl = unlock->in(TypeFunc::Control);
2386 Node* mem = unlock->in(TypeFunc::Memory);
2387 Node* obj = unlock->obj_node();
2388 Node* box = unlock->box_node();
2389
2390 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2391
2392 // No need for a null check on unlock
2393
2394 // Make the merge point
2395 Node *region;
2396 Node *mem_phi;
2397
2398 if (UseOptoBiasInlining) {
2399 // Check for biased locking unlock case, which is a no-op.
2400 // See the full description in MacroAssembler::biased_locking_exit().
2401 region = new RegionNode(4);
2402 // create a Phi for the memory state
2403 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2404 mem_phi->init_req(3, mem);
2405
2406 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2407 ctrl = opt_bits_test(ctrl, region, 3, mark_node,
2408 markOopDesc::biased_lock_mask_in_place,
2409 markOopDesc::biased_lock_pattern);
2410 } else {
2411 region = new RegionNode(3);
2412 // create a Phi for the memory state
2413 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2414 }
2415
2416 FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box );
2417 funlock = transform_later( funlock )->as_FastUnlock();
2418 // Optimize test; set region slot 2
2419 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0);
2420 Node *thread = transform_later(new ThreadLocalNode());
2421
2422 CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(),
2423 CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C),
2424 "complete_monitor_unlocking_C", slow_path, obj, box, thread);
2425
2426 extract_call_projections(call);
2427
2428 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
2429 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
2430
2431 // No exceptions for unlocking
2432 // Capture slow path
2433 // disconnect fall-through projection from call and create a new one
2434 // hook up users of fall-through projection to region
2435 Node *slow_ctrl = _fallthroughproj->clone();
2436 transform_later(slow_ctrl);
2437 _igvn.hash_delete(_fallthroughproj);
2438 _fallthroughproj->disconnect_inputs(NULL, C);
2439 region->init_req(1, slow_ctrl);
2440 // region inputs are now complete
2441 transform_later(region);
2442 _igvn.replace_node(_fallthroughproj, region);
2443
2444 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory) );
2445 mem_phi->init_req(1, memproj );
2446 mem_phi->init_req(2, mem);
2447 transform_later(mem_phi);
2448 _igvn.replace_node(_memproj_fallthrough, mem_phi);
2449 }
2450
2451 //---------------------------eliminate_macro_nodes----------------------
2452 // Eliminate scalar replaced allocations and associated locks.
eliminate_macro_nodes()2453 void PhaseMacroExpand::eliminate_macro_nodes() {
2454 if (C->macro_count() == 0)
2455 return;
2456
2457 // First, attempt to eliminate locks
2458 int cnt = C->macro_count();
2459 for (int i=0; i < cnt; i++) {
2460 Node *n = C->macro_node(i);
2461 if (n->is_AbstractLock()) { // Lock and Unlock nodes
2462 // Before elimination mark all associated (same box and obj)
2463 // lock and unlock nodes.
2464 mark_eliminated_locking_nodes(n->as_AbstractLock());
2465 }
2466 }
2467 bool progress = true;
2468 while (progress) {
2469 progress = false;
2470 for (int i = C->macro_count(); i > 0; i--) {
2471 Node * n = C->macro_node(i-1);
2472 bool success = false;
2473 debug_only(int old_macro_count = C->macro_count(););
2474 if (n->is_AbstractLock()) {
2475 success = eliminate_locking_node(n->as_AbstractLock());
2476 }
2477 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2478 progress = progress || success;
2479 }
2480 }
2481 // Next, attempt to eliminate allocations
2482 _has_locks = false;
2483 progress = true;
2484 while (progress) {
2485 progress = false;
2486 for (int i = C->macro_count(); i > 0; i--) {
2487 Node * n = C->macro_node(i-1);
2488 bool success = false;
2489 debug_only(int old_macro_count = C->macro_count(););
2490 switch (n->class_id()) {
2491 case Node::Class_Allocate:
2492 case Node::Class_AllocateArray:
2493 success = eliminate_allocate_node(n->as_Allocate());
2494 break;
2495 case Node::Class_CallStaticJava:
2496 success = eliminate_boxing_node(n->as_CallStaticJava());
2497 break;
2498 case Node::Class_Lock:
2499 case Node::Class_Unlock:
2500 assert(!n->as_AbstractLock()->is_eliminated(), "sanity");
2501 _has_locks = true;
2502 break;
2503 case Node::Class_ArrayCopy:
2504 break;
2505 case Node::Class_OuterStripMinedLoop:
2506 break;
2507 default:
2508 assert(n->Opcode() == Op_LoopLimit ||
2509 n->Opcode() == Op_Opaque1 ||
2510 n->Opcode() == Op_Opaque2 ||
2511 n->Opcode() == Op_Opaque3 ||
2512 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(n),
2513 "unknown node type in macro list");
2514 }
2515 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2516 progress = progress || success;
2517 }
2518 }
2519 }
2520
2521 //------------------------------expand_macro_nodes----------------------
2522 // Returns true if a failure occurred.
expand_macro_nodes()2523 bool PhaseMacroExpand::expand_macro_nodes() {
2524 // Last attempt to eliminate macro nodes.
2525 eliminate_macro_nodes();
2526
2527 // Make sure expansion will not cause node limit to be exceeded.
2528 // Worst case is a macro node gets expanded into about 200 nodes.
2529 // Allow 50% more for optimization.
2530 if (C->check_node_count(C->macro_count() * 300, "out of nodes before macro expansion" ) )
2531 return true;
2532
2533 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations.
2534 bool progress = true;
2535 while (progress) {
2536 progress = false;
2537 for (int i = C->macro_count(); i > 0; i--) {
2538 Node * n = C->macro_node(i-1);
2539 bool success = false;
2540 debug_only(int old_macro_count = C->macro_count(););
2541 if (n->Opcode() == Op_LoopLimit) {
2542 // Remove it from macro list and put on IGVN worklist to optimize.
2543 C->remove_macro_node(n);
2544 _igvn._worklist.push(n);
2545 success = true;
2546 } else if (n->Opcode() == Op_CallStaticJava) {
2547 // Remove it from macro list and put on IGVN worklist to optimize.
2548 C->remove_macro_node(n);
2549 _igvn._worklist.push(n);
2550 success = true;
2551 } else if (n->Opcode() == Op_Opaque1 || n->Opcode() == Op_Opaque2) {
2552 _igvn.replace_node(n, n->in(1));
2553 success = true;
2554 #if INCLUDE_RTM_OPT
2555 } else if ((n->Opcode() == Op_Opaque3) && ((Opaque3Node*)n)->rtm_opt()) {
2556 assert(C->profile_rtm(), "should be used only in rtm deoptimization code");
2557 assert((n->outcnt() == 1) && n->unique_out()->is_Cmp(), "");
2558 Node* cmp = n->unique_out();
2559 #ifdef ASSERT
2560 // Validate graph.
2561 assert((cmp->outcnt() == 1) && cmp->unique_out()->is_Bool(), "");
2562 BoolNode* bol = cmp->unique_out()->as_Bool();
2563 assert((bol->outcnt() == 1) && bol->unique_out()->is_If() &&
2564 (bol->_test._test == BoolTest::ne), "");
2565 IfNode* ifn = bol->unique_out()->as_If();
2566 assert((ifn->outcnt() == 2) &&
2567 ifn->proj_out(1)->is_uncommon_trap_proj(Deoptimization::Reason_rtm_state_change) != NULL, "");
2568 #endif
2569 Node* repl = n->in(1);
2570 if (!_has_locks) {
2571 // Remove RTM state check if there are no locks in the code.
2572 // Replace input to compare the same value.
2573 repl = (cmp->in(1) == n) ? cmp->in(2) : cmp->in(1);
2574 }
2575 _igvn.replace_node(n, repl);
2576 success = true;
2577 #endif
2578 } else if (n->Opcode() == Op_OuterStripMinedLoop) {
2579 n->as_OuterStripMinedLoop()->adjust_strip_mined_loop(&_igvn);
2580 C->remove_macro_node(n);
2581 success = true;
2582 }
2583 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2584 progress = progress || success;
2585 }
2586 }
2587
2588 // expand arraycopy "macro" nodes first
2589 // For ReduceBulkZeroing, we must first process all arraycopy nodes
2590 // before the allocate nodes are expanded.
2591 int macro_idx = C->macro_count() - 1;
2592 while (macro_idx >= 0) {
2593 Node * n = C->macro_node(macro_idx);
2594 assert(n->is_macro(), "only macro nodes expected here");
2595 if (_igvn.type(n) == Type::TOP || (n->in(0) != NULL && n->in(0)->is_top())) {
2596 // node is unreachable, so don't try to expand it
2597 C->remove_macro_node(n);
2598 } else if (n->is_ArrayCopy()){
2599 int macro_count = C->macro_count();
2600 expand_arraycopy_node(n->as_ArrayCopy());
2601 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
2602 }
2603 if (C->failing()) return true;
2604 macro_idx --;
2605 }
2606
2607 // expand "macro" nodes
2608 // nodes are removed from the macro list as they are processed
2609 while (C->macro_count() > 0) {
2610 int macro_count = C->macro_count();
2611 Node * n = C->macro_node(macro_count-1);
2612 assert(n->is_macro(), "only macro nodes expected here");
2613 if (_igvn.type(n) == Type::TOP || (n->in(0) != NULL && n->in(0)->is_top())) {
2614 // node is unreachable, so don't try to expand it
2615 C->remove_macro_node(n);
2616 continue;
2617 }
2618 switch (n->class_id()) {
2619 case Node::Class_Allocate:
2620 expand_allocate(n->as_Allocate());
2621 break;
2622 case Node::Class_AllocateArray:
2623 expand_allocate_array(n->as_AllocateArray());
2624 break;
2625 case Node::Class_Lock:
2626 expand_lock_node(n->as_Lock());
2627 break;
2628 case Node::Class_Unlock:
2629 expand_unlock_node(n->as_Unlock());
2630 break;
2631 default:
2632 assert(false, "unknown node type in macro list");
2633 }
2634 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
2635 if (C->failing()) return true;
2636 }
2637
2638 _igvn.set_delay_transform(false);
2639 _igvn.optimize();
2640 if (C->failing()) return true;
2641 return false;
2642 }
2643