1 /*
2 * Copyright (c) 2005, 2018, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "compiler/compileLog.hpp"
27 #include "gc/shared/collectedHeap.inline.hpp"
28 #include "libadt/vectset.hpp"
29 #include "opto/addnode.hpp"
30 #include "opto/arraycopynode.hpp"
31 #include "opto/callnode.hpp"
32 #include "opto/castnode.hpp"
33 #include "opto/cfgnode.hpp"
34 #include "opto/compile.hpp"
35 #include "opto/convertnode.hpp"
36 #include "opto/graphKit.hpp"
37 #include "opto/locknode.hpp"
38 #include "opto/loopnode.hpp"
39 #include "opto/macro.hpp"
40 #include "opto/memnode.hpp"
41 #include "opto/narrowptrnode.hpp"
42 #include "opto/node.hpp"
43 #include "opto/opaquenode.hpp"
44 #include "opto/phaseX.hpp"
45 #include "opto/rootnode.hpp"
46 #include "opto/runtime.hpp"
47 #include "opto/subnode.hpp"
48 #include "opto/type.hpp"
49 #include "opto/valuetypenode.hpp"
50 #include "runtime/sharedRuntime.hpp"
51 #if INCLUDE_G1GC
52 #include "gc/g1/g1ThreadLocalData.hpp"
53 #endif // INCLUDE_G1GC
54
55
56 //
57 // Replace any references to "oldref" in inputs to "use" with "newref".
58 // Returns the number of replacements made.
59 //
60 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) {
61 int nreplacements = 0;
62 uint req = use->req();
63 for (uint j = 0; j < use->len(); j++) {
64 Node *uin = use->in(j);
65 if (uin == oldref) {
66 if (j < req)
67 use->set_req(j, newref);
68 else
69 use->set_prec(j, newref);
70 nreplacements++;
71 } else if (j >= req && uin == NULL) {
72 break;
73 }
74 }
75 return nreplacements;
76 }
77
78 void PhaseMacroExpand::copy_call_debug_info(CallNode *oldcall, CallNode * newcall) {
79 // Copy debug information and adjust JVMState information
80 uint old_dbg_start = oldcall->tf()->domain_sig()->cnt();
81 uint new_dbg_start = newcall->tf()->domain_sig()->cnt();
82 int jvms_adj = new_dbg_start - old_dbg_start;
83 assert (new_dbg_start == newcall->req(), "argument count mismatch");
84
85 // SafePointScalarObject node could be referenced several times in debug info.
86 // Use Dict to record cloned nodes.
87 Dict* sosn_map = new Dict(cmpkey,hashkey);
88 for (uint i = old_dbg_start; i < oldcall->req(); i++) {
89 Node* old_in = oldcall->in(i);
90 // Clone old SafePointScalarObjectNodes, adjusting their field contents.
91 if (old_in != NULL && old_in->is_SafePointScalarObject()) {
92 SafePointScalarObjectNode* old_sosn = old_in->as_SafePointScalarObject();
93 uint old_unique = C->unique();
94 Node* new_in = old_sosn->clone(sosn_map);
95 if (old_unique != C->unique()) { // New node?
96 new_in->set_req(0, C->root()); // reset control edge
97 new_in = transform_later(new_in); // Register new node.
98 }
99 old_in = new_in;
100 }
101 newcall->add_req(old_in);
102 }
103
104 // JVMS may be shared so clone it before we modify it
105 newcall->set_jvms(oldcall->jvms() != NULL ? oldcall->jvms()->clone_deep(C) : NULL);
106 for (JVMState *jvms = newcall->jvms(); jvms != NULL; jvms = jvms->caller()) {
107 jvms->set_map(newcall);
108 jvms->set_locoff(jvms->locoff()+jvms_adj);
109 jvms->set_stkoff(jvms->stkoff()+jvms_adj);
110 jvms->set_monoff(jvms->monoff()+jvms_adj);
111 jvms->set_scloff(jvms->scloff()+jvms_adj);
112 jvms->set_endoff(jvms->endoff()+jvms_adj);
113 }
114 }
115
116 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word, int mask, int bits, bool return_fast_path) {
117 Node* cmp;
118 if (mask != 0) {
119 Node* and_node = transform_later(new AndXNode(word, MakeConX(mask)));
120 cmp = transform_later(new CmpXNode(and_node, MakeConX(bits)));
121 } else {
122 cmp = word;
123 }
124 Node* bol = transform_later(new BoolNode(cmp, BoolTest::ne));
125 IfNode* iff = new IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN );
126 transform_later(iff);
127
128 // Fast path taken.
129 Node *fast_taken = transform_later(new IfFalseNode(iff));
130
131 // Fast path not-taken, i.e. slow path
132 Node *slow_taken = transform_later(new IfTrueNode(iff));
133
134 if (return_fast_path) {
135 region->init_req(edge, slow_taken); // Capture slow-control
136 return fast_taken;
137 } else {
138 region->init_req(edge, fast_taken); // Capture fast-control
139 return slow_taken;
140 }
141 }
142
143 //--------------------copy_predefined_input_for_runtime_call--------------------
144 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) {
145 // Set fixed predefined input arguments
146 call->init_req( TypeFunc::Control, ctrl );
147 call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) );
148 call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ?????
149 call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) );
150 call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) );
151 }
152
153 //------------------------------make_slow_call---------------------------------
154 CallNode* PhaseMacroExpand::make_slow_call(CallNode *oldcall, const TypeFunc* slow_call_type,
155 address slow_call, const char* leaf_name, Node* slow_path,
156 Node* parm0, Node* parm1, Node* parm2) {
157
158 // Slow-path call
159 CallNode *call = leaf_name
160 ? (CallNode*)new CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM )
161 : (CallNode*)new CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), oldcall->jvms()->bci(), TypeRawPtr::BOTTOM );
162
163 // Slow path call has no side-effects, uses few values
164 copy_predefined_input_for_runtime_call(slow_path, oldcall, call );
165 if (parm0 != NULL) call->init_req(TypeFunc::Parms+0, parm0);
166 if (parm1 != NULL) call->init_req(TypeFunc::Parms+1, parm1);
167 if (parm2 != NULL) call->init_req(TypeFunc::Parms+2, parm2);
168 copy_call_debug_info(oldcall, call);
169 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
170 _igvn.replace_node(oldcall, call);
171 transform_later(call);
172
173 return call;
174 }
175
176 void PhaseMacroExpand::extract_call_projections(CallNode *call) {
177 _fallthroughproj = NULL;
178 _fallthroughcatchproj = NULL;
179 _ioproj_fallthrough = NULL;
180 _ioproj_catchall = NULL;
181 _catchallcatchproj = NULL;
182 _memproj_fallthrough = NULL;
183 _memproj_catchall = NULL;
184 _resproj = NULL;
185 for (DUIterator_Fast imax, i = call->fast_outs(imax); i < imax; i++) {
186 ProjNode *pn = call->fast_out(i)->as_Proj();
187 switch (pn->_con) {
188 case TypeFunc::Control:
189 {
190 // For Control (fallthrough) and I_O (catch_all_index) we have CatchProj -> Catch -> Proj
191 _fallthroughproj = pn;
192 DUIterator_Fast jmax, j = pn->fast_outs(jmax);
193 const Node *cn = pn->fast_out(j);
194 if (cn->is_Catch()) {
195 ProjNode *cpn = NULL;
196 for (DUIterator_Fast kmax, k = cn->fast_outs(kmax); k < kmax; k++) {
197 cpn = cn->fast_out(k)->as_Proj();
198 assert(cpn->is_CatchProj(), "must be a CatchProjNode");
199 if (cpn->_con == CatchProjNode::fall_through_index)
200 _fallthroughcatchproj = cpn;
201 else {
202 assert(cpn->_con == CatchProjNode::catch_all_index, "must be correct index.");
203 _catchallcatchproj = cpn;
204 }
205 }
206 }
207 break;
208 }
209 case TypeFunc::I_O:
210 if (pn->_is_io_use)
211 _ioproj_catchall = pn;
212 else
213 _ioproj_fallthrough = pn;
214 break;
215 case TypeFunc::Memory:
216 if (pn->_is_io_use)
217 _memproj_catchall = pn;
218 else
219 _memproj_fallthrough = pn;
220 break;
221 case TypeFunc::Parms:
222 _resproj = pn;
223 break;
224 default:
225 assert(false, "unexpected projection from allocation node.");
226 }
227 }
228
229 }
230
231 // Eliminate a card mark sequence. p2x is a ConvP2XNode
232 void PhaseMacroExpand::eliminate_card_mark(Node* p2x) {
233 assert(p2x->Opcode() == Op_CastP2X, "ConvP2XNode required");
234 if (!UseG1GC) {
235 // vanilla/CMS post barrier
236 Node *shift = p2x->unique_out();
237 Node *addp = shift->unique_out();
238 for (DUIterator_Last jmin, j = addp->last_outs(jmin); j >= jmin; --j) {
239 Node *mem = addp->last_out(j);
240 if (UseCondCardMark && mem->is_Load()) {
241 assert(mem->Opcode() == Op_LoadB, "unexpected code shape");
242 // The load is checking if the card has been written so
243 // replace it with zero to fold the test.
244 _igvn.replace_node(mem, intcon(0));
245 continue;
246 }
247 assert(mem->is_Store(), "store required");
248 _igvn.replace_node(mem, mem->in(MemNode::Memory));
249 }
250 }
251 #if INCLUDE_G1GC
252 else {
253 // G1 pre/post barriers
254 assert(p2x->outcnt() <= 2, "expects 1 or 2 users: Xor and URShift nodes");
255 // It could be only one user, URShift node, in Object.clone() intrinsic
256 // but the new allocation is passed to arraycopy stub and it could not
257 // be scalar replaced. So we don't check the case.
258
259 // An other case of only one user (Xor) is when the value check for NULL
260 // in G1 post barrier is folded after CCP so the code which used URShift
261 // is removed.
262
263 // Take Region node before eliminating post barrier since it also
264 // eliminates CastP2X node when it has only one user.
265 Node* this_region = p2x->in(0);
266 assert(this_region != NULL, "");
267
268 // Remove G1 post barrier.
269
270 // Search for CastP2X->Xor->URShift->Cmp path which
271 // checks if the store done to a different from the value's region.
272 // And replace Cmp with #0 (false) to collapse G1 post barrier.
273 Node* xorx = p2x->find_out_with(Op_XorX);
274 if (xorx != NULL) {
275 Node* shift = xorx->unique_out();
276 Node* cmpx = shift->unique_out();
277 assert(cmpx->is_Cmp() && cmpx->unique_out()->is_Bool() &&
278 cmpx->unique_out()->as_Bool()->_test._test == BoolTest::ne,
279 "missing region check in G1 post barrier");
280 _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ));
281
282 // Remove G1 pre barrier.
283
284 // Search "if (marking != 0)" check and set it to "false".
285 // There is no G1 pre barrier if previous stored value is NULL
286 // (for example, after initialization).
287 if (this_region->is_Region() && this_region->req() == 3) {
288 int ind = 1;
289 if (!this_region->in(ind)->is_IfFalse()) {
290 ind = 2;
291 }
292 if (this_region->in(ind)->is_IfFalse() &&
293 this_region->in(ind)->in(0)->Opcode() == Op_If) {
294 Node* bol = this_region->in(ind)->in(0)->in(1);
295 assert(bol->is_Bool(), "");
296 cmpx = bol->in(1);
297 if (bol->as_Bool()->_test._test == BoolTest::ne &&
298 cmpx->is_Cmp() && cmpx->in(2) == intcon(0) &&
299 cmpx->in(1)->is_Load()) {
300 Node* adr = cmpx->in(1)->as_Load()->in(MemNode::Address);
301 const int marking_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_active_offset());
302 if (adr->is_AddP() && adr->in(AddPNode::Base) == top() &&
303 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
304 adr->in(AddPNode::Offset) == MakeConX(marking_offset)) {
305 _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ));
306 }
307 }
308 }
309 }
310 } else {
311 assert(!GraphKit::use_ReduceInitialCardMarks(), "can only happen with card marking");
312 // This is a G1 post barrier emitted by the Object.clone() intrinsic.
313 // Search for the CastP2X->URShiftX->AddP->LoadB->Cmp path which checks if the card
314 // is marked as young_gen and replace the Cmp with 0 (false) to collapse the barrier.
315 Node* shift = p2x->find_out_with(Op_URShiftX);
316 assert(shift != NULL, "missing G1 post barrier");
317 Node* addp = shift->unique_out();
318 Node* load = addp->find_out_with(Op_LoadB);
319 assert(load != NULL, "missing G1 post barrier");
320 Node* cmpx = load->unique_out();
321 assert(cmpx->is_Cmp() && cmpx->unique_out()->is_Bool() &&
322 cmpx->unique_out()->as_Bool()->_test._test == BoolTest::ne,
323 "missing card value check in G1 post barrier");
324 _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ));
325 // There is no G1 pre barrier in this case
326 }
327 // Now CastP2X can be removed since it is used only on dead path
328 // which currently still alive until igvn optimize it.
329 assert(p2x->outcnt() == 0 || p2x->unique_out()->Opcode() == Op_URShiftX, "");
330 _igvn.replace_node(p2x, top());
331 }
332 #endif // INCLUDE_G1GC
333 }
334
335 // Search for a memory operation for the specified memory slice.
336 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) {
337 Node *orig_mem = mem;
338 Node *alloc_mem = alloc->in(TypeFunc::Memory);
339 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr();
340 while (true) {
341 if (mem == alloc_mem || mem == start_mem ) {
342 return mem; // hit one of our sentinels
343 } else if (mem->is_MergeMem()) {
344 mem = mem->as_MergeMem()->memory_at(alias_idx);
345 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) {
346 Node *in = mem->in(0);
347 // we can safely skip over safepoints, calls, locks and membars because we
348 // already know that the object is safe to eliminate.
349 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) {
350 return in;
351 } else if (in->is_Call()) {
352 CallNode *call = in->as_Call();
353 if (call->may_modify(tinst, phase)) {
354 assert(call->is_ArrayCopy(), "ArrayCopy is the only call node that doesn't make allocation escape");
355 if (call->as_ArrayCopy()->modifies(offset, offset, phase, false)) {
356 return in;
357 }
358 }
359 mem = in->in(TypeFunc::Memory);
360 } else if (in->is_MemBar()) {
361 ArrayCopyNode* ac = NULL;
362 if (ArrayCopyNode::may_modify(tinst, in->as_MemBar(), phase, ac)) {
363 assert(ac != NULL && ac->is_clonebasic(), "Only basic clone is a non escaping clone");
364 return ac;
365 }
366 mem = in->in(TypeFunc::Memory);
367 } else {
368 assert(false, "unexpected projection");
369 }
370 } else if (mem->is_Store()) {
371 const TypePtr* atype = mem->as_Store()->adr_type();
372 int adr_idx = phase->C->get_alias_index(atype);
373 if (adr_idx == alias_idx) {
374 assert(atype->isa_oopptr(), "address type must be oopptr");
375 int adr_offset = atype->flattened_offset();
376 uint adr_iid = atype->is_oopptr()->instance_id();
377 // Array elements references have the same alias_idx
378 // but different offset and different instance_id.
379 if (adr_offset == offset && adr_iid == alloc->_idx)
380 return mem;
381 } else {
382 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw");
383 }
384 mem = mem->in(MemNode::Memory);
385 } else if (mem->is_ClearArray()) {
386 if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) {
387 // Can not bypass initialization of the instance
388 // we are looking.
389 debug_only(intptr_t offset;)
390 assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity");
391 InitializeNode* init = alloc->as_Allocate()->initialization();
392 // We are looking for stored value, return Initialize node
393 // or memory edge from Allocate node.
394 if (init != NULL)
395 return init;
396 else
397 return alloc->in(TypeFunc::Memory); // It will produce zero value (see callers).
398 }
399 // Otherwise skip it (the call updated 'mem' value).
400 } else if (mem->Opcode() == Op_SCMemProj) {
401 mem = mem->in(0);
402 Node* adr = NULL;
403 if (mem->is_LoadStore()) {
404 adr = mem->in(MemNode::Address);
405 } else {
406 assert(mem->Opcode() == Op_EncodeISOArray ||
407 mem->Opcode() == Op_StrCompressedCopy, "sanity");
408 adr = mem->in(3); // Destination array
409 }
410 const TypePtr* atype = adr->bottom_type()->is_ptr();
411 int adr_idx = phase->C->get_alias_index(atype);
412 if (adr_idx == alias_idx) {
413 DEBUG_ONLY(mem->dump();)
414 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
415 return NULL;
416 }
417 mem = mem->in(MemNode::Memory);
418 } else if (mem->Opcode() == Op_StrInflatedCopy) {
419 Node* adr = mem->in(3); // Destination array
420 const TypePtr* atype = adr->bottom_type()->is_ptr();
421 int adr_idx = phase->C->get_alias_index(atype);
422 if (adr_idx == alias_idx) {
423 DEBUG_ONLY(mem->dump();)
424 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field");
425 return NULL;
426 }
427 mem = mem->in(MemNode::Memory);
428 } else {
429 return mem;
430 }
431 assert(mem != orig_mem, "dead memory loop");
432 }
433 }
434
435 // Generate loads from source of the arraycopy for fields of
436 // destination needed at a deoptimization point
437 Node* PhaseMacroExpand::make_arraycopy_load(ArrayCopyNode* ac, intptr_t offset, Node* ctl, Node* mem, BasicType ft, const Type *ftype, AllocateNode *alloc) {
438 BasicType bt = ft;
439 const Type *type = ftype;
440 if (ft == T_NARROWOOP) {
441 bt = T_OBJECT;
442 type = ftype->make_oopptr();
443 }
444 Node* res = NULL;
445 if (ac->is_clonebasic()) {
446 Node* base = ac->in(ArrayCopyNode::Src)->in(AddPNode::Base);
447 Node* adr = _igvn.transform(new AddPNode(base, base, MakeConX(offset)));
448 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset);
449 res = LoadNode::make(_igvn, ctl, mem, adr, adr_type, type, bt, MemNode::unordered, LoadNode::Pinned);
450 } else {
451 if (ac->modifies(offset, offset, &_igvn, true)) {
452 assert(ac->in(ArrayCopyNode::Dest) == alloc->result_cast(), "arraycopy destination should be allocation's result");
453 uint shift = exact_log2(type2aelembytes(bt));
454 Node* diff = _igvn.transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
455 #ifdef _LP64
456 diff = _igvn.transform(new ConvI2LNode(diff));
457 #endif
458 diff = _igvn.transform(new LShiftXNode(diff, intcon(shift)));
459
460 Node* off = _igvn.transform(new AddXNode(MakeConX(offset), diff));
461 Node* base = ac->in(ArrayCopyNode::Src);
462 Node* adr = _igvn.transform(new AddPNode(base, base, off));
463 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset);
464 res = LoadNode::make(_igvn, ctl, mem, adr, adr_type, type, bt, MemNode::unordered, LoadNode::Pinned);
465 }
466 }
467 if (res != NULL) {
468 res = _igvn.transform(res);
469 if (ftype->isa_narrowoop()) {
470 // PhaseMacroExpand::scalar_replacement adds DecodeN nodes
471 res = _igvn.transform(new EncodePNode(res, ftype));
472 }
473 return res;
474 }
475 return NULL;
476 }
477
478 //
479 // Given a Memory Phi, compute a value Phi containing the values from stores
480 // on the input paths.
481 // Note: this function is recursive, its depth is limited by the "level" argument
482 // Returns the computed Phi, or NULL if it cannot compute it.
483 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) {
484 assert(mem->is_Phi(), "sanity");
485 int alias_idx = C->get_alias_index(adr_t);
486 int offset = adr_t->flattened_offset();
487 int instance_id = adr_t->instance_id();
488
489 // Check if an appropriate value phi already exists.
490 Node* region = mem->in(0);
491 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) {
492 Node* phi = region->fast_out(k);
493 if (phi->is_Phi() && phi != mem &&
494 phi->as_Phi()->is_same_inst_field(phi_type, (int)mem->_idx, instance_id, alias_idx, offset)) {
495 return phi;
496 }
497 }
498 // Check if an appropriate new value phi already exists.
499 Node* new_phi = value_phis->find(mem->_idx);
500 if (new_phi != NULL)
501 return new_phi;
502
503 if (level <= 0) {
504 return NULL; // Give up: phi tree too deep
505 }
506 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
507 Node *alloc_mem = alloc->in(TypeFunc::Memory);
508
509 uint length = mem->req();
510 GrowableArray <Node *> values(length, length, NULL, false);
511
512 // create a new Phi for the value
513 PhiNode *phi = new PhiNode(mem->in(0), phi_type, NULL, mem->_idx, instance_id, alias_idx, offset);
514 transform_later(phi);
515 value_phis->push(phi, mem->_idx);
516
517 for (uint j = 1; j < length; j++) {
518 Node *in = mem->in(j);
519 if (in == NULL || in->is_top()) {
520 values.at_put(j, in);
521 } else {
522 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn);
523 if (val == start_mem || val == alloc_mem) {
524 // hit a sentinel, return appropriate 0 value
525 values.at_put(j, _igvn.zerocon(ft));
526 continue;
527 }
528 if (val->is_Initialize()) {
529 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
530 }
531 if (val == NULL) {
532 return NULL; // can't find a value on this path
533 }
534 if (val == mem) {
535 values.at_put(j, mem);
536 } else if (val->is_Store()) {
537 values.at_put(j, val->in(MemNode::ValueIn));
538 } else if(val->is_Proj() && val->in(0) == alloc) {
539 values.at_put(j, _igvn.zerocon(ft));
540 } else if (val->is_Phi()) {
541 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1);
542 if (val == NULL) {
543 return NULL;
544 }
545 values.at_put(j, val);
546 } else if (val->Opcode() == Op_SCMemProj) {
547 assert(val->in(0)->is_LoadStore() ||
548 val->in(0)->Opcode() == Op_EncodeISOArray ||
549 val->in(0)->Opcode() == Op_StrCompressedCopy, "sanity");
550 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
551 return NULL;
552 } else if (val->is_ArrayCopy()) {
553 Node* res = make_arraycopy_load(val->as_ArrayCopy(), offset, val->in(0), val->in(TypeFunc::Memory), ft, phi_type, alloc);
554 if (res == NULL) {
555 return NULL;
556 }
557 values.at_put(j, res);
558 } else {
559 #ifdef ASSERT
560 val->dump();
561 assert(false, "unknown node on this path");
562 #endif
563 return NULL; // unknown node on this path
564 }
565 }
566 }
567 // Set Phi's inputs
568 for (uint j = 1; j < length; j++) {
569 if (values.at(j) == mem) {
570 phi->init_req(j, phi);
571 } else {
572 phi->init_req(j, values.at(j));
573 }
574 }
575 return phi;
576 }
577
578 // Search the last value stored into the object's field.
579 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, Node *sfpt_ctl, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, AllocateNode *alloc) {
580 assert(adr_t->is_known_instance_field(), "instance required");
581 int instance_id = adr_t->instance_id();
582 assert((uint)instance_id == alloc->_idx, "wrong allocation");
583
584 int alias_idx = C->get_alias_index(adr_t);
585 int offset = adr_t->flattened_offset();
586 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
587 Node *alloc_ctrl = alloc->in(TypeFunc::Control);
588 Node *alloc_mem = alloc->in(TypeFunc::Memory);
589 Arena *a = Thread::current()->resource_area();
590 VectorSet visited(a);
591
592 bool done = sfpt_mem == alloc_mem;
593 Node *mem = sfpt_mem;
594 while (!done) {
595 if (visited.test_set(mem->_idx)) {
596 return NULL; // found a loop, give up
597 }
598 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn);
599 if (mem == start_mem || mem == alloc_mem) {
600 done = true; // hit a sentinel, return appropriate 0 value
601 } else if (mem->is_Initialize()) {
602 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
603 if (mem == NULL) {
604 done = true; // Something went wrong.
605 } else if (mem->is_Store()) {
606 const TypePtr* atype = mem->as_Store()->adr_type();
607 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice");
608 done = true;
609 }
610 } else if (mem->is_Store()) {
611 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr();
612 assert(atype != NULL, "address type must be oopptr");
613 assert(C->get_alias_index(atype) == alias_idx &&
614 atype->is_known_instance_field() && atype->flattened_offset() == offset &&
615 atype->instance_id() == instance_id, "store is correct memory slice");
616 done = true;
617 } else if (mem->is_Phi()) {
618 // try to find a phi's unique input
619 Node *unique_input = NULL;
620 Node *top = C->top();
621 for (uint i = 1; i < mem->req(); i++) {
622 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn);
623 if (n == NULL || n == top || n == mem) {
624 continue;
625 } else if (unique_input == NULL) {
626 unique_input = n;
627 } else if (unique_input != n) {
628 unique_input = top;
629 break;
630 }
631 }
632 if (unique_input != NULL && unique_input != top) {
633 mem = unique_input;
634 } else {
635 done = true;
636 }
637 } else if (mem->is_ArrayCopy()) {
638 done = true;
639 } else {
640 assert(false, "unexpected node");
641 }
642 }
643 if (mem != NULL) {
644 if (mem == start_mem || mem == alloc_mem) {
645 // hit a sentinel, return appropriate 0 value
646 return _igvn.zerocon(ft);
647 } else if (mem->is_Store()) {
648 return mem->in(MemNode::ValueIn);
649 } else if (mem->is_Phi()) {
650 // attempt to produce a Phi reflecting the values on the input paths of the Phi
651 Node_Stack value_phis(a, 8);
652 Node * phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit);
653 if (phi != NULL) {
654 return phi;
655 } else {
656 // Kill all new Phis
657 while(value_phis.is_nonempty()) {
658 Node* n = value_phis.node();
659 _igvn.replace_node(n, C->top());
660 value_phis.pop();
661 }
662 }
663 } else if (mem->is_ArrayCopy()) {
664 Node* ctl = mem->in(0);
665 Node* m = mem->in(TypeFunc::Memory);
666 if (sfpt_ctl->is_Proj() && sfpt_ctl->as_Proj()->is_uncommon_trap_proj(Deoptimization::Reason_none)) {
667 // pin the loads in the uncommon trap path
668 ctl = sfpt_ctl;
669 m = sfpt_mem;
670 }
671 return make_arraycopy_load(mem->as_ArrayCopy(), offset, ctl, m, ft, ftype, alloc);
672 }
673 }
674 // Something went wrong.
675 return NULL;
676 }
677
678 // Search the last value stored into the value type's fields.
679 Node* PhaseMacroExpand::value_type_from_mem(Node* mem, Node* ctl, ciValueKlass* vk, const TypeAryPtr* adr_type, int offset, AllocateNode* alloc) {
680 // Subtract the offset of the first field to account for the missing oop header
681 offset -= vk->first_field_offset();
682 // Create a new ValueTypeNode and retrieve the field values from memory
683 ValueTypeNode* vt = ValueTypeNode::make_uninitialized(_igvn, vk)->as_ValueType();
684 for (int i = 0; i < vk->nof_declared_nonstatic_fields(); ++i) {
685 ciType* field_type = vt->field_type(i);
686 int field_offset = offset + vt->field_offset(i);
687 // Each value type field has its own memory slice
688 adr_type = adr_type->with_field_offset(field_offset);
689 Node* value = NULL;
690 if (vt->field_is_flattened(i)) {
691 value = value_type_from_mem(mem, ctl, field_type->as_value_klass(), adr_type, field_offset, alloc);
692 } else {
693 const Type* ft = Type::get_const_type(field_type);
694 BasicType bt = field_type->basic_type();
695 if (UseCompressedOops && !is_java_primitive(bt)) {
696 ft = ft->make_narrowoop();
697 bt = T_NARROWOOP;
698 }
699 value = value_from_mem(mem, ctl, bt, ft, adr_type, alloc);
700 if (value != NULL && ft->isa_narrowoop()) {
701 assert(UseCompressedOops, "unexpected narrow oop");
702 value = transform_later(new DecodeNNode(value, value->get_ptr_type()));
703 }
704 }
705 if (value != NULL) {
706 vt->set_field_value(i, value);
707 } else {
708 // We might have reached the TrackedInitializationLimit
709 return NULL;
710 }
711 }
712 return vt;
713 }
714
715 // Check the possibility of scalar replacement.
716 bool PhaseMacroExpand::can_eliminate_allocation(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
717 // Scan the uses of the allocation to check for anything that would
718 // prevent us from eliminating it.
719 NOT_PRODUCT( const char* fail_eliminate = NULL; )
720 DEBUG_ONLY( Node* disq_node = NULL; )
721 bool can_eliminate = true;
722
723 Node* res = alloc->result_cast();
724 const TypeOopPtr* res_type = NULL;
725 if (res == NULL) {
726 // All users were eliminated.
727 } else if (!res->is_CheckCastPP()) {
728 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";)
729 can_eliminate = false;
730 } else {
731 res_type = _igvn.type(res)->isa_oopptr();
732 if (res_type == NULL) {
733 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";)
734 can_eliminate = false;
735 } else if (res_type->isa_aryptr()) {
736 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1);
737 if (length < 0) {
738 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";)
739 can_eliminate = false;
740 }
741 }
742 }
743
744 if (can_eliminate && res != NULL) {
745 for (DUIterator_Fast jmax, j = res->fast_outs(jmax);
746 j < jmax && can_eliminate; j++) {
747 Node* use = res->fast_out(j);
748
749 if (use->is_AddP()) {
750 const TypePtr* addp_type = _igvn.type(use)->is_ptr();
751 int offset = addp_type->offset();
752
753 if (offset == Type::OffsetTop || offset == Type::OffsetBot) {
754 NOT_PRODUCT(fail_eliminate = "Undefined field referrence";)
755 can_eliminate = false;
756 break;
757 }
758 for (DUIterator_Fast kmax, k = use->fast_outs(kmax);
759 k < kmax && can_eliminate; k++) {
760 Node* n = use->fast_out(k);
761 if (!n->is_Store() && n->Opcode() != Op_CastP2X &&
762 !(n->is_ArrayCopy() &&
763 n->as_ArrayCopy()->is_clonebasic() &&
764 n->in(ArrayCopyNode::Dest) == use)) {
765 DEBUG_ONLY(disq_node = n;)
766 if (n->is_Load() || n->is_LoadStore()) {
767 NOT_PRODUCT(fail_eliminate = "Field load";)
768 } else {
769 NOT_PRODUCT(fail_eliminate = "Not store field reference";)
770 }
771 can_eliminate = false;
772 }
773 }
774 } else if (use->is_ArrayCopy() &&
775 (use->as_ArrayCopy()->is_arraycopy_validated() ||
776 use->as_ArrayCopy()->is_copyof_validated() ||
777 use->as_ArrayCopy()->is_copyofrange_validated()) &&
778 use->in(ArrayCopyNode::Dest) == res) {
779 // ok to eliminate
780 } else if (use->is_SafePoint()) {
781 SafePointNode* sfpt = use->as_SafePoint();
782 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) {
783 // Object is passed as argument.
784 DEBUG_ONLY(disq_node = use;)
785 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";)
786 can_eliminate = false;
787 }
788 Node* sfptMem = sfpt->memory();
789 if (sfptMem == NULL || sfptMem->is_top()) {
790 DEBUG_ONLY(disq_node = use;)
791 NOT_PRODUCT(fail_eliminate = "NULL or TOP memory";)
792 can_eliminate = false;
793 } else {
794 safepoints.append_if_missing(sfpt);
795 }
796 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
797 if (use->is_Phi()) {
798 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
799 NOT_PRODUCT(fail_eliminate = "Object is return value";)
800 } else {
801 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
802 }
803 DEBUG_ONLY(disq_node = use;)
804 } else {
805 if (use->Opcode() == Op_Return) {
806 NOT_PRODUCT(fail_eliminate = "Object is return value";)
807 } else {
808 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
809 }
810 DEBUG_ONLY(disq_node = use;)
811 }
812 can_eliminate = false;
813 }
814 }
815 }
816
817 #ifndef PRODUCT
818 if (PrintEliminateAllocations) {
819 if (can_eliminate) {
820 tty->print("Scalar ");
821 if (res == NULL)
822 alloc->dump();
823 else
824 res->dump();
825 } else if (alloc->_is_scalar_replaceable) {
826 tty->print("NotScalar (%s)", fail_eliminate);
827 if (res == NULL)
828 alloc->dump();
829 else
830 res->dump();
831 #ifdef ASSERT
832 if (disq_node != NULL) {
833 tty->print(" >>>> ");
834 disq_node->dump();
835 }
836 #endif /*ASSERT*/
837 }
838 }
839 #endif
840 return can_eliminate;
841 }
842
843 // Do scalar replacement.
844 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
845 GrowableArray <SafePointNode *> safepoints_done;
846
847 ciKlass* klass = NULL;
848 ciInstanceKlass* iklass = NULL;
849 int nfields = 0;
850 int array_base = 0;
851 int element_size = 0;
852 BasicType basic_elem_type = T_ILLEGAL;
853 ciType* elem_type = NULL;
854
855 Node* res = alloc->result_cast();
856 assert(res == NULL || res->is_CheckCastPP(), "unexpected AllocateNode result");
857 const TypeOopPtr* res_type = NULL;
858 if (res != NULL) { // Could be NULL when there are no users
859 res_type = _igvn.type(res)->isa_oopptr();
860 }
861
862 if (res != NULL) {
863 klass = res_type->klass();
864 // Value types are only allocated on demand
865 if (res_type->isa_instptr() || res_type->isa_valuetypeptr()) {
866 // find the fields of the class which will be needed for safepoint debug information
867 assert(klass->is_instance_klass(), "must be an instance klass.");
868 iklass = klass->as_instance_klass();
869 nfields = iklass->nof_nonstatic_fields();
870 } else {
871 // find the array's elements which will be needed for safepoint debug information
872 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
873 assert(klass->is_array_klass() && nfields >= 0, "must be an array klass.");
874 elem_type = klass->as_array_klass()->element_type();
875 basic_elem_type = elem_type->basic_type();
876 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
877 element_size = type2aelembytes(basic_elem_type);
878 if (klass->is_value_array_klass()) {
879 // Flattened value type array
880 element_size = klass->as_value_array_klass()->element_byte_size();
881 }
882 }
883 }
884 //
885 // Process the safepoint uses
886 //
887 Unique_Node_List value_worklist;
888 while (safepoints.length() > 0) {
889 SafePointNode* sfpt = safepoints.pop();
890 Node* mem = sfpt->memory();
891 Node* ctl = sfpt->control();
892 assert(sfpt->jvms() != NULL, "missed JVMS");
893 // Fields of scalar objs are referenced only at the end
894 // of regular debuginfo at the last (youngest) JVMS.
895 // Record relative start index.
896 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff());
897 SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type,
898 #ifdef ASSERT
899 alloc,
900 #endif
901 first_ind, nfields);
902 sobj->init_req(0, C->root());
903 transform_later(sobj);
904
905 // Scan object's fields adding an input to the safepoint for each field.
906 for (int j = 0; j < nfields; j++) {
907 intptr_t offset;
908 ciField* field = NULL;
909 if (iklass != NULL) {
910 field = iklass->nonstatic_field_at(j);
911 offset = field->offset();
912 elem_type = field->type();
913 basic_elem_type = field->layout_type();
914 assert(!field->is_flattened(), "flattened value type fields should not have safepoint uses");
915 } else {
916 offset = array_base + j * (intptr_t)element_size;
917 }
918
919 const Type *field_type;
920 // The next code is taken from Parse::do_get_xxx().
921 if (basic_elem_type == T_OBJECT || basic_elem_type == T_ARRAY) {
922 if (!elem_type->is_loaded()) {
923 field_type = TypeInstPtr::BOTTOM;
924 } else if (field != NULL && field->is_static_constant()) {
925 // This can happen if the constant oop is non-perm.
926 ciObject* con = field->constant_value().as_object();
927 // Do not "join" in the previous type; it doesn't add value,
928 // and may yield a vacuous result if the field is of interface type.
929 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr();
930 assert(field_type != NULL, "field singleton type must be consistent");
931 } else {
932 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
933 }
934 if (UseCompressedOops) {
935 field_type = field_type->make_narrowoop();
936 basic_elem_type = T_NARROWOOP;
937 }
938 } else {
939 field_type = Type::get_const_basic_type(basic_elem_type);
940 }
941
942 Node* field_val = NULL;
943 const TypeOopPtr* field_addr_type = res_type->add_offset(offset)->isa_oopptr();
944 if (klass->is_value_array_klass()) {
945 ciValueKlass* vk = elem_type->as_value_klass();
946 assert(vk->flatten_array(), "must be flattened");
947 field_val = value_type_from_mem(mem, ctl, vk, field_addr_type->isa_aryptr(), 0, alloc);
948 } else {
949 field_val = value_from_mem(mem, ctl, basic_elem_type, field_type, field_addr_type, alloc);
950 }
951 if (field_val == NULL) {
952 // We weren't able to find a value for this field,
953 // give up on eliminating this allocation.
954
955 // Remove any extra entries we added to the safepoint.
956 uint last = sfpt->req() - 1;
957 for (int k = 0; k < j; k++) {
958 sfpt->del_req(last--);
959 }
960 _igvn._worklist.push(sfpt);
961 // rollback processed safepoints
962 while (safepoints_done.length() > 0) {
963 SafePointNode* sfpt_done = safepoints_done.pop();
964 // remove any extra entries we added to the safepoint
965 last = sfpt_done->req() - 1;
966 for (int k = 0; k < nfields; k++) {
967 sfpt_done->del_req(last--);
968 }
969 JVMState *jvms = sfpt_done->jvms();
970 jvms->set_endoff(sfpt_done->req());
971 // Now make a pass over the debug information replacing any references
972 // to SafePointScalarObjectNode with the allocated object.
973 int start = jvms->debug_start();
974 int end = jvms->debug_end();
975 for (int i = start; i < end; i++) {
976 if (sfpt_done->in(i)->is_SafePointScalarObject()) {
977 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject();
978 if (scobj->first_index(jvms) == sfpt_done->req() &&
979 scobj->n_fields() == (uint)nfields) {
980 assert(scobj->alloc() == alloc, "sanity");
981 sfpt_done->set_req(i, res);
982 }
983 }
984 }
985 _igvn._worklist.push(sfpt_done);
986 }
987 #ifndef PRODUCT
988 if (PrintEliminateAllocations) {
989 if (field != NULL) {
990 tty->print("=== At SafePoint node %d can't find value of Field: ",
991 sfpt->_idx);
992 field->print();
993 int field_idx = C->get_alias_index(field_addr_type);
994 tty->print(" (alias_idx=%d)", field_idx);
995 } else { // Array's element
996 tty->print("=== At SafePoint node %d can't find value of array element [%d]",
997 sfpt->_idx, j);
998 }
999 tty->print(", which prevents elimination of: ");
1000 if (res == NULL)
1001 alloc->dump();
1002 else
1003 res->dump();
1004 }
1005 #endif
1006 return false;
1007 }
1008 if (UseCompressedOops && field_type->isa_narrowoop()) {
1009 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
1010 // to be able scalar replace the allocation.
1011 if (field_val->is_EncodeP()) {
1012 field_val = field_val->in(1);
1013 } else {
1014 field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type()));
1015 }
1016 } else if (field_val->is_ValueType()) {
1017 // Keep track of value types to scalarize them later
1018 value_worklist.push(field_val);
1019 }
1020 sfpt->add_req(field_val);
1021 }
1022 JVMState *jvms = sfpt->jvms();
1023 jvms->set_endoff(sfpt->req());
1024 // Now make a pass over the debug information replacing any references
1025 // to the allocated object with "sobj"
1026 int start = jvms->debug_start();
1027 int end = jvms->debug_end();
1028 sfpt->replace_edges_in_range(res, sobj, start, end);
1029 _igvn._worklist.push(sfpt);
1030 safepoints_done.append_if_missing(sfpt); // keep it for rollback
1031 }
1032 // Scalarize value types that were added to the safepoint
1033 for (uint i = 0; i < value_worklist.size(); ++i) {
1034 Node* vt = value_worklist.at(i);
1035 vt->as_ValueType()->make_scalar_in_safepoints(C->root(), &_igvn);
1036 }
1037 return true;
1038 }
1039
1040 static void disconnect_projections(MultiNode* n, PhaseIterGVN& igvn) {
1041 Node* ctl_proj = n->proj_out_or_null(TypeFunc::Control);
1042 Node* mem_proj = n->proj_out_or_null(TypeFunc::Memory);
1043 if (ctl_proj != NULL) {
1044 igvn.replace_node(ctl_proj, n->in(0));
1045 }
1046 if (mem_proj != NULL) {
1047 igvn.replace_node(mem_proj, n->in(TypeFunc::Memory));
1048 }
1049 }
1050
1051 // Process users of eliminated allocation.
1052 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc) {
1053 Node* res = alloc->result_cast();
1054 if (res != NULL) {
1055 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
1056 Node *use = res->last_out(j);
1057 uint oc1 = res->outcnt();
1058
1059 if (use->is_AddP()) {
1060 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) {
1061 Node *n = use->last_out(k);
1062 uint oc2 = use->outcnt();
1063 if (n->is_Store()) {
1064 #ifdef ASSERT
1065 // Verify that there is no dependent MemBarVolatile nodes,
1066 // they should be removed during IGVN, see MemBarNode::Ideal().
1067 for (DUIterator_Fast pmax, p = n->fast_outs(pmax);
1068 p < pmax; p++) {
1069 Node* mb = n->fast_out(p);
1070 assert(mb->is_Initialize() || !mb->is_MemBar() ||
1071 mb->req() <= MemBarNode::Precedent ||
1072 mb->in(MemBarNode::Precedent) != n,
1073 "MemBarVolatile should be eliminated for non-escaping object");
1074 }
1075 #endif
1076 _igvn.replace_node(n, n->in(MemNode::Memory));
1077 } else if (n->is_ArrayCopy()) {
1078 // Disconnect ArrayCopy node
1079 ArrayCopyNode* ac = n->as_ArrayCopy();
1080 assert(ac->is_clonebasic(), "unexpected array copy kind");
1081 Node* membar_after = ac->proj_out(TypeFunc::Control)->unique_ctrl_out();
1082 disconnect_projections(ac, _igvn);
1083 assert(alloc->in(0)->is_Proj() && alloc->in(0)->in(0)->Opcode() == Op_MemBarCPUOrder, "mem barrier expected before allocation");
1084 Node* membar_before = alloc->in(0)->in(0);
1085 disconnect_projections(membar_before->as_MemBar(), _igvn);
1086 if (membar_after->is_MemBar()) {
1087 disconnect_projections(membar_after->as_MemBar(), _igvn);
1088 }
1089 } else {
1090 eliminate_card_mark(n);
1091 }
1092 k -= (oc2 - use->outcnt());
1093 }
1094 } else if (use->is_ArrayCopy()) {
1095 // Disconnect ArrayCopy node
1096 ArrayCopyNode* ac = use->as_ArrayCopy();
1097 assert(ac->is_arraycopy_validated() ||
1098 ac->is_copyof_validated() ||
1099 ac->is_copyofrange_validated(), "unsupported");
1100 CallProjections* callprojs = ac->extract_projections(true);
1101
1102 _igvn.replace_node(callprojs->fallthrough_ioproj, ac->in(TypeFunc::I_O));
1103 _igvn.replace_node(callprojs->fallthrough_memproj, ac->in(TypeFunc::Memory));
1104 _igvn.replace_node(callprojs->fallthrough_catchproj, ac->in(TypeFunc::Control));
1105
1106 // Set control to top. IGVN will remove the remaining projections
1107 ac->set_req(0, top());
1108 ac->replace_edge(res, top());
1109
1110 // Disconnect src right away: it can help find new
1111 // opportunities for allocation elimination
1112 Node* src = ac->in(ArrayCopyNode::Src);
1113 ac->replace_edge(src, top());
1114 // src can be top at this point if src and dest of the
1115 // arraycopy were the same
1116 if (src->outcnt() == 0 && !src->is_top()) {
1117 _igvn.remove_dead_node(src);
1118 }
1119
1120 _igvn._worklist.push(ac);
1121 } else {
1122 eliminate_card_mark(use);
1123 }
1124 j -= (oc1 - res->outcnt());
1125 }
1126 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
1127 _igvn.remove_dead_node(res);
1128 }
1129
1130 //
1131 // Process other users of allocation's projections
1132 //
1133 if (_resproj != NULL && _resproj->outcnt() != 0) {
1134 // First disconnect stores captured by Initialize node.
1135 // If Initialize node is eliminated first in the following code,
1136 // it will kill such stores and DUIterator_Last will assert.
1137 for (DUIterator_Fast jmax, j = _resproj->fast_outs(jmax); j < jmax; j++) {
1138 Node *use = _resproj->fast_out(j);
1139 if (use->is_AddP()) {
1140 // raw memory addresses used only by the initialization
1141 _igvn.replace_node(use, C->top());
1142 --j; --jmax;
1143 }
1144 }
1145 for (DUIterator_Last jmin, j = _resproj->last_outs(jmin); j >= jmin; ) {
1146 Node *use = _resproj->last_out(j);
1147 uint oc1 = _resproj->outcnt();
1148 if (use->is_Initialize()) {
1149 // Eliminate Initialize node.
1150 InitializeNode *init = use->as_Initialize();
1151 assert(init->outcnt() <= 2, "only a control and memory projection expected");
1152 Node *ctrl_proj = init->proj_out_or_null(TypeFunc::Control);
1153 if (ctrl_proj != NULL) {
1154 assert(init->in(TypeFunc::Control) == _fallthroughcatchproj, "allocation control projection");
1155 _igvn.replace_node(ctrl_proj, _fallthroughcatchproj);
1156 }
1157 Node *mem_proj = init->proj_out_or_null(TypeFunc::Memory);
1158 if (mem_proj != NULL) {
1159 Node *mem = init->in(TypeFunc::Memory);
1160 #ifdef ASSERT
1161 if (mem->is_MergeMem()) {
1162 assert(mem->in(TypeFunc::Memory) == _memproj_fallthrough, "allocation memory projection");
1163 } else {
1164 assert(mem == _memproj_fallthrough, "allocation memory projection");
1165 }
1166 #endif
1167 _igvn.replace_node(mem_proj, mem);
1168 }
1169 } else {
1170 assert(false, "only Initialize or AddP expected");
1171 }
1172 j -= (oc1 - _resproj->outcnt());
1173 }
1174 }
1175 if (_fallthroughcatchproj != NULL) {
1176 _igvn.replace_node(_fallthroughcatchproj, alloc->in(TypeFunc::Control));
1177 }
1178 if (_memproj_fallthrough != NULL) {
1179 _igvn.replace_node(_memproj_fallthrough, alloc->in(TypeFunc::Memory));
1180 }
1181 if (_memproj_catchall != NULL) {
1182 _igvn.replace_node(_memproj_catchall, C->top());
1183 }
1184 if (_ioproj_fallthrough != NULL) {
1185 _igvn.replace_node(_ioproj_fallthrough, alloc->in(TypeFunc::I_O));
1186 }
1187 if (_ioproj_catchall != NULL) {
1188 _igvn.replace_node(_ioproj_catchall, C->top());
1189 }
1190 if (_catchallcatchproj != NULL) {
1191 _igvn.replace_node(_catchallcatchproj, C->top());
1192 }
1193 }
1194
1195 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) {
1196 // Don't do scalar replacement if the frame can be popped by JVMTI:
1197 // if reallocation fails during deoptimization we'll pop all
1198 // interpreter frames for this compiled frame and that won't play
1199 // nice with JVMTI popframe.
1200 if (!EliminateAllocations || JvmtiExport::can_pop_frame() || !alloc->_is_non_escaping) {
1201 return false;
1202 }
1203 Node* klass = alloc->in(AllocateNode::KlassNode);
1204 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr();
1205 Node* res = alloc->result_cast();
1206 // Eliminate boxing allocations which are not used
1207 // regardless scalar replacable status.
1208 bool boxing_alloc = C->eliminate_boxing() &&
1209 tklass->klass()->is_instance_klass() &&
1210 tklass->klass()->as_instance_klass()->is_box_klass();
1211 if (!alloc->_is_scalar_replaceable && (!boxing_alloc || (res != NULL))) {
1212 return false;
1213 }
1214
1215 extract_call_projections(alloc);
1216
1217 GrowableArray <SafePointNode *> safepoints;
1218 if (!can_eliminate_allocation(alloc, safepoints)) {
1219 return false;
1220 }
1221
1222 if (!alloc->_is_scalar_replaceable) {
1223 assert(res == NULL, "sanity");
1224 // We can only eliminate allocation if all debug info references
1225 // are already replaced with SafePointScalarObject because
1226 // we can't search for a fields value without instance_id.
1227 if (safepoints.length() > 0) {
1228 return false;
1229 }
1230 }
1231
1232 if (!scalar_replacement(alloc, safepoints)) {
1233 return false;
1234 }
1235
1236 CompileLog* log = C->log();
1237 if (log != NULL) {
1238 log->head("eliminate_allocation type='%d'",
1239 log->identify(tklass->klass()));
1240 JVMState* p = alloc->jvms();
1241 while (p != NULL) {
1242 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1243 p = p->caller();
1244 }
1245 log->tail("eliminate_allocation");
1246 }
1247
1248 process_users_of_allocation(alloc);
1249
1250 #ifndef PRODUCT
1251 if (PrintEliminateAllocations) {
1252 if (alloc->is_AllocateArray())
1253 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1254 else
1255 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1256 }
1257 #endif
1258
1259 return true;
1260 }
1261
1262 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) {
1263 // EA should remove all uses of non-escaping boxing node.
1264 if (!C->eliminate_boxing() || boxing->proj_out_or_null(TypeFunc::Parms) != NULL) {
1265 return false;
1266 }
1267
1268 assert(boxing->result_cast() == NULL, "unexpected boxing node result");
1269
1270 extract_call_projections(boxing);
1271
1272 const TypeTuple* r = boxing->tf()->range_sig();
1273 assert(r->cnt() > TypeFunc::Parms, "sanity");
1274 const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr();
1275 assert(t != NULL, "sanity");
1276
1277 CompileLog* log = C->log();
1278 if (log != NULL) {
1279 log->head("eliminate_boxing type='%d'",
1280 log->identify(t->klass()));
1281 JVMState* p = boxing->jvms();
1282 while (p != NULL) {
1283 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1284 p = p->caller();
1285 }
1286 log->tail("eliminate_boxing");
1287 }
1288
1289 process_users_of_allocation(boxing);
1290
1291 #ifndef PRODUCT
1292 if (PrintEliminateAllocations) {
1293 tty->print("++++ Eliminated: %d ", boxing->_idx);
1294 boxing->method()->print_short_name(tty);
1295 tty->cr();
1296 }
1297 #endif
1298
1299 return true;
1300 }
1301
1302 //---------------------------set_eden_pointers-------------------------
1303 void PhaseMacroExpand::set_eden_pointers(Node* &eden_top_adr, Node* &eden_end_adr) {
1304 if (UseTLAB) { // Private allocation: load from TLS
1305 Node* thread = transform_later(new ThreadLocalNode());
1306 int tlab_top_offset = in_bytes(JavaThread::tlab_top_offset());
1307 int tlab_end_offset = in_bytes(JavaThread::tlab_end_offset());
1308 eden_top_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_top_offset);
1309 eden_end_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_end_offset);
1310 } else { // Shared allocation: load from globals
1311 CollectedHeap* ch = Universe::heap();
1312 address top_adr = (address)ch->top_addr();
1313 address end_adr = (address)ch->end_addr();
1314 eden_top_adr = makecon(TypeRawPtr::make(top_adr));
1315 eden_end_adr = basic_plus_adr(eden_top_adr, end_adr - top_adr);
1316 }
1317 }
1318
1319
1320 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) {
1321 Node* adr = basic_plus_adr(base, offset);
1322 const TypePtr* adr_type = adr->bottom_type()->is_ptr();
1323 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt, MemNode::unordered);
1324 transform_later(value);
1325 return value;
1326 }
1327
1328
1329 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) {
1330 Node* adr = basic_plus_adr(base, offset);
1331 mem = StoreNode::make(_igvn, ctl, mem, adr, NULL, value, bt, MemNode::unordered);
1332 transform_later(mem);
1333 return mem;
1334 }
1335
1336 //=============================================================================
1337 //
1338 // A L L O C A T I O N
1339 //
1340 // Allocation attempts to be fast in the case of frequent small objects.
1341 // It breaks down like this:
1342 //
1343 // 1) Size in doublewords is computed. This is a constant for objects and
1344 // variable for most arrays. Doubleword units are used to avoid size
1345 // overflow of huge doubleword arrays. We need doublewords in the end for
1346 // rounding.
1347 //
1348 // 2) Size is checked for being 'too large'. Too-large allocations will go
1349 // the slow path into the VM. The slow path can throw any required
1350 // exceptions, and does all the special checks for very large arrays. The
1351 // size test can constant-fold away for objects. For objects with
1352 // finalizers it constant-folds the otherway: you always go slow with
1353 // finalizers.
1354 //
1355 // 3) If NOT using TLABs, this is the contended loop-back point.
1356 // Load-Locked the heap top. If using TLABs normal-load the heap top.
1357 //
1358 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route.
1359 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish
1360 // "size*8" we always enter the VM, where "largish" is a constant picked small
1361 // enough that there's always space between the eden max and 4Gig (old space is
1362 // there so it's quite large) and large enough that the cost of entering the VM
1363 // is dwarfed by the cost to initialize the space.
1364 //
1365 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back
1366 // down. If contended, repeat at step 3. If using TLABs normal-store
1367 // adjusted heap top back down; there is no contention.
1368 //
1369 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark
1370 // fields.
1371 //
1372 // 7) Merge with the slow-path; cast the raw memory pointer to the correct
1373 // oop flavor.
1374 //
1375 //=============================================================================
1376 // FastAllocateSizeLimit value is in DOUBLEWORDS.
1377 // Allocations bigger than this always go the slow route.
1378 // This value must be small enough that allocation attempts that need to
1379 // trigger exceptions go the slow route. Also, it must be small enough so
1380 // that heap_top + size_in_bytes does not wrap around the 4Gig limit.
1381 //=============================================================================j//
1382 // %%% Here is an old comment from parseHelper.cpp; is it outdated?
1383 // The allocator will coalesce int->oop copies away. See comment in
1384 // coalesce.cpp about how this works. It depends critically on the exact
1385 // code shape produced here, so if you are changing this code shape
1386 // make sure the GC info for the heap-top is correct in and around the
1387 // slow-path call.
1388 //
1389
1390 void PhaseMacroExpand::expand_allocate_common(
1391 AllocateNode* alloc, // allocation node to be expanded
1392 Node* length, // array length for an array allocation
1393 const TypeFunc* slow_call_type, // Type of slow call
1394 address slow_call_address // Address of slow call
1395 )
1396 {
1397
1398 Node* ctrl = alloc->in(TypeFunc::Control);
1399 Node* mem = alloc->in(TypeFunc::Memory);
1400 Node* i_o = alloc->in(TypeFunc::I_O);
1401 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize);
1402 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1403 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest);
1404
1405 assert(ctrl != NULL, "must have control");
1406 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results.
1407 // they will not be used if "always_slow" is set
1408 enum { slow_result_path = 1, fast_result_path = 2 };
1409 Node *result_region = NULL;
1410 Node *result_phi_rawmem = NULL;
1411 Node *result_phi_rawoop = NULL;
1412 Node *result_phi_i_o = NULL;
1413
1414 // The initial slow comparison is a size check, the comparison
1415 // we want to do is a BoolTest::gt
1416 bool always_slow = false;
1417 int tv = _igvn.find_int_con(initial_slow_test, -1);
1418 if (tv >= 0) {
1419 always_slow = (tv == 1);
1420 initial_slow_test = NULL;
1421 } else {
1422 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn);
1423 }
1424
1425 if (C->env()->dtrace_alloc_probes() ||
1426 (!UseTLAB && !Universe::heap()->supports_inline_contig_alloc())) {
1427 // Force slow-path allocation
1428 always_slow = true;
1429 initial_slow_test = NULL;
1430 }
1431
1432
1433 enum { too_big_or_final_path = 1, need_gc_path = 2 };
1434 Node *slow_region = NULL;
1435 Node *toobig_false = ctrl;
1436
1437 assert (initial_slow_test == NULL || !always_slow, "arguments must be consistent");
1438 // generate the initial test if necessary
1439 if (initial_slow_test != NULL ) {
1440 slow_region = new RegionNode(3);
1441
1442 // Now make the initial failure test. Usually a too-big test but
1443 // might be a TRUE for finalizers or a fancy class check for
1444 // newInstance0.
1445 IfNode *toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
1446 transform_later(toobig_iff);
1447 // Plug the failing-too-big test into the slow-path region
1448 Node *toobig_true = new IfTrueNode( toobig_iff );
1449 transform_later(toobig_true);
1450 slow_region ->init_req( too_big_or_final_path, toobig_true );
1451 toobig_false = new IfFalseNode( toobig_iff );
1452 transform_later(toobig_false);
1453 } else { // No initial test, just fall into next case
1454 toobig_false = ctrl;
1455 debug_only(slow_region = NodeSentinel);
1456 }
1457
1458 Node *slow_mem = mem; // save the current memory state for slow path
1459 // generate the fast allocation code unless we know that the initial test will always go slow
1460 if (!always_slow) {
1461 // Fast path modifies only raw memory.
1462 if (mem->is_MergeMem()) {
1463 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1464 }
1465
1466 Node* eden_top_adr;
1467 Node* eden_end_adr;
1468
1469 set_eden_pointers(eden_top_adr, eden_end_adr);
1470
1471 // Load Eden::end. Loop invariant and hoisted.
1472 //
1473 // Note: We set the control input on "eden_end" and "old_eden_top" when using
1474 // a TLAB to work around a bug where these values were being moved across
1475 // a safepoint. These are not oops, so they cannot be include in the oop
1476 // map, but they can be changed by a GC. The proper way to fix this would
1477 // be to set the raw memory state when generating a SafepointNode. However
1478 // this will require extensive changes to the loop optimization in order to
1479 // prevent a degradation of the optimization.
1480 // See comment in memnode.hpp, around line 227 in class LoadPNode.
1481 Node *eden_end = make_load(ctrl, mem, eden_end_adr, 0, TypeRawPtr::BOTTOM, T_ADDRESS);
1482
1483 // allocate the Region and Phi nodes for the result
1484 result_region = new RegionNode(3);
1485 result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1486 result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM);
1487 result_phi_i_o = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch
1488
1489 // We need a Region for the loop-back contended case.
1490 enum { fall_in_path = 1, contended_loopback_path = 2 };
1491 Node *contended_region;
1492 Node *contended_phi_rawmem;
1493 if (UseTLAB) {
1494 contended_region = toobig_false;
1495 contended_phi_rawmem = mem;
1496 } else {
1497 contended_region = new RegionNode(3);
1498 contended_phi_rawmem = new PhiNode(contended_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1499 // Now handle the passing-too-big test. We fall into the contended
1500 // loop-back merge point.
1501 contended_region ->init_req(fall_in_path, toobig_false);
1502 contended_phi_rawmem->init_req(fall_in_path, mem);
1503 transform_later(contended_region);
1504 transform_later(contended_phi_rawmem);
1505 }
1506
1507 // Load(-locked) the heap top.
1508 // See note above concerning the control input when using a TLAB
1509 Node *old_eden_top = UseTLAB
1510 ? new LoadPNode (ctrl, contended_phi_rawmem, eden_top_adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, MemNode::unordered)
1511 : new LoadPLockedNode(contended_region, contended_phi_rawmem, eden_top_adr, MemNode::acquire);
1512
1513 transform_later(old_eden_top);
1514 // Add to heap top to get a new heap top
1515 Node *new_eden_top = new AddPNode(top(), old_eden_top, size_in_bytes);
1516 transform_later(new_eden_top);
1517 // Check for needing a GC; compare against heap end
1518 Node *needgc_cmp = new CmpPNode(new_eden_top, eden_end);
1519 transform_later(needgc_cmp);
1520 Node *needgc_bol = new BoolNode(needgc_cmp, BoolTest::ge);
1521 transform_later(needgc_bol);
1522 IfNode *needgc_iff = new IfNode(contended_region, needgc_bol, PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN);
1523 transform_later(needgc_iff);
1524
1525 // Plug the failing-heap-space-need-gc test into the slow-path region
1526 Node *needgc_true = new IfTrueNode(needgc_iff);
1527 transform_later(needgc_true);
1528 if (initial_slow_test) {
1529 slow_region->init_req(need_gc_path, needgc_true);
1530 // This completes all paths into the slow merge point
1531 transform_later(slow_region);
1532 } else { // No initial slow path needed!
1533 // Just fall from the need-GC path straight into the VM call.
1534 slow_region = needgc_true;
1535 }
1536 // No need for a GC. Setup for the Store-Conditional
1537 Node *needgc_false = new IfFalseNode(needgc_iff);
1538 transform_later(needgc_false);
1539
1540 // Grab regular I/O before optional prefetch may change it.
1541 // Slow-path does no I/O so just set it to the original I/O.
1542 result_phi_i_o->init_req(slow_result_path, i_o);
1543
1544 i_o = prefetch_allocation(i_o, needgc_false, contended_phi_rawmem,
1545 old_eden_top, new_eden_top, length);
1546
1547 // Name successful fast-path variables
1548 Node* fast_oop = old_eden_top;
1549 Node* fast_oop_ctrl;
1550 Node* fast_oop_rawmem;
1551
1552 // Store (-conditional) the modified eden top back down.
1553 // StorePConditional produces flags for a test PLUS a modified raw
1554 // memory state.
1555 if (UseTLAB) {
1556 Node* store_eden_top =
1557 new StorePNode(needgc_false, contended_phi_rawmem, eden_top_adr,
1558 TypeRawPtr::BOTTOM, new_eden_top, MemNode::unordered);
1559 transform_later(store_eden_top);
1560 fast_oop_ctrl = needgc_false; // No contention, so this is the fast path
1561 fast_oop_rawmem = store_eden_top;
1562 } else {
1563 Node* store_eden_top =
1564 new StorePConditionalNode(needgc_false, contended_phi_rawmem, eden_top_adr,
1565 new_eden_top, fast_oop/*old_eden_top*/);
1566 transform_later(store_eden_top);
1567 Node *contention_check = new BoolNode(store_eden_top, BoolTest::ne);
1568 transform_later(contention_check);
1569 store_eden_top = new SCMemProjNode(store_eden_top);
1570 transform_later(store_eden_top);
1571
1572 // If not using TLABs, check to see if there was contention.
1573 IfNode *contention_iff = new IfNode (needgc_false, contention_check, PROB_MIN, COUNT_UNKNOWN);
1574 transform_later(contention_iff);
1575 Node *contention_true = new IfTrueNode(contention_iff);
1576 transform_later(contention_true);
1577 // If contention, loopback and try again.
1578 contended_region->init_req(contended_loopback_path, contention_true);
1579 contended_phi_rawmem->init_req(contended_loopback_path, store_eden_top);
1580
1581 // Fast-path succeeded with no contention!
1582 Node *contention_false = new IfFalseNode(contention_iff);
1583 transform_later(contention_false);
1584 fast_oop_ctrl = contention_false;
1585
1586 // Bump total allocated bytes for this thread
1587 Node* thread = new ThreadLocalNode();
1588 transform_later(thread);
1589 Node* alloc_bytes_adr = basic_plus_adr(top()/*not oop*/, thread,
1590 in_bytes(JavaThread::allocated_bytes_offset()));
1591 Node* alloc_bytes = make_load(fast_oop_ctrl, store_eden_top, alloc_bytes_adr,
1592 0, TypeLong::LONG, T_LONG);
1593 #ifdef _LP64
1594 Node* alloc_size = size_in_bytes;
1595 #else
1596 Node* alloc_size = new ConvI2LNode(size_in_bytes);
1597 transform_later(alloc_size);
1598 #endif
1599 Node* new_alloc_bytes = new AddLNode(alloc_bytes, alloc_size);
1600 transform_later(new_alloc_bytes);
1601 fast_oop_rawmem = make_store(fast_oop_ctrl, store_eden_top, alloc_bytes_adr,
1602 0, new_alloc_bytes, T_LONG);
1603 }
1604
1605 InitializeNode* init = alloc->initialization();
1606 fast_oop_rawmem = initialize_object(alloc,
1607 fast_oop_ctrl, fast_oop_rawmem, fast_oop,
1608 klass_node, length, size_in_bytes);
1609
1610 // If initialization is performed by an array copy, any required
1611 // MemBarStoreStore was already added. If the object does not
1612 // escape no need for a MemBarStoreStore. If the object does not
1613 // escape in its initializer and memory barrier (MemBarStoreStore or
1614 // stronger) is already added at exit of initializer, also no need
1615 // for a MemBarStoreStore. Otherwise we need a MemBarStoreStore
1616 // so that stores that initialize this object can't be reordered
1617 // with a subsequent store that makes this object accessible by
1618 // other threads.
1619 // Other threads include java threads and JVM internal threads
1620 // (for example concurrent GC threads). Current concurrent GC
1621 // implementation: CMS and G1 will not scan newly created object,
1622 // so it's safe to skip storestore barrier when allocation does
1623 // not escape.
1624 if (!alloc->does_not_escape_thread() &&
1625 !alloc->is_allocation_MemBar_redundant() &&
1626 (init == NULL || !init->is_complete_with_arraycopy())) {
1627 if (init == NULL || init->req() < InitializeNode::RawStores) {
1628 // No InitializeNode or no stores captured by zeroing
1629 // elimination. Simply add the MemBarStoreStore after object
1630 // initialization.
1631 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1632 transform_later(mb);
1633
1634 mb->init_req(TypeFunc::Memory, fast_oop_rawmem);
1635 mb->init_req(TypeFunc::Control, fast_oop_ctrl);
1636 fast_oop_ctrl = new ProjNode(mb,TypeFunc::Control);
1637 transform_later(fast_oop_ctrl);
1638 fast_oop_rawmem = new ProjNode(mb,TypeFunc::Memory);
1639 transform_later(fast_oop_rawmem);
1640 } else {
1641 // Add the MemBarStoreStore after the InitializeNode so that
1642 // all stores performing the initialization that were moved
1643 // before the InitializeNode happen before the storestore
1644 // barrier.
1645
1646 Node* init_ctrl = init->proj_out_or_null(TypeFunc::Control);
1647 Node* init_mem = init->proj_out_or_null(TypeFunc::Memory);
1648
1649 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1650 transform_later(mb);
1651
1652 Node* ctrl = new ProjNode(init,TypeFunc::Control);
1653 transform_later(ctrl);
1654 Node* mem = new ProjNode(init,TypeFunc::Memory);
1655 transform_later(mem);
1656
1657 // The MemBarStoreStore depends on control and memory coming
1658 // from the InitializeNode
1659 mb->init_req(TypeFunc::Memory, mem);
1660 mb->init_req(TypeFunc::Control, ctrl);
1661
1662 ctrl = new ProjNode(mb,TypeFunc::Control);
1663 transform_later(ctrl);
1664 mem = new ProjNode(mb,TypeFunc::Memory);
1665 transform_later(mem);
1666
1667 // All nodes that depended on the InitializeNode for control
1668 // and memory must now depend on the MemBarNode that itself
1669 // depends on the InitializeNode
1670 if (init_ctrl != NULL) {
1671 _igvn.replace_node(init_ctrl, ctrl);
1672 }
1673 if (init_mem != NULL) {
1674 _igvn.replace_node(init_mem, mem);
1675 }
1676 }
1677 }
1678
1679 if (C->env()->dtrace_extended_probes()) {
1680 // Slow-path call
1681 int size = TypeFunc::Parms + 2;
1682 CallLeafNode *call = new CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(),
1683 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_object_alloc_base),
1684 "dtrace_object_alloc",
1685 TypeRawPtr::BOTTOM);
1686
1687 // Get base of thread-local storage area
1688 Node* thread = new ThreadLocalNode();
1689 transform_later(thread);
1690
1691 call->init_req(TypeFunc::Parms+0, thread);
1692 call->init_req(TypeFunc::Parms+1, fast_oop);
1693 call->init_req(TypeFunc::Control, fast_oop_ctrl);
1694 call->init_req(TypeFunc::I_O , top()); // does no i/o
1695 call->init_req(TypeFunc::Memory , fast_oop_rawmem);
1696 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1697 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1698 transform_later(call);
1699 fast_oop_ctrl = new ProjNode(call,TypeFunc::Control);
1700 transform_later(fast_oop_ctrl);
1701 fast_oop_rawmem = new ProjNode(call,TypeFunc::Memory);
1702 transform_later(fast_oop_rawmem);
1703 }
1704
1705 // Plug in the successful fast-path into the result merge point
1706 result_region ->init_req(fast_result_path, fast_oop_ctrl);
1707 result_phi_rawoop->init_req(fast_result_path, fast_oop);
1708 result_phi_i_o ->init_req(fast_result_path, i_o);
1709 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem);
1710 } else {
1711 slow_region = ctrl;
1712 result_phi_i_o = i_o; // Rename it to use in the following code.
1713 }
1714
1715 // Generate slow-path call
1716 CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address,
1717 OptoRuntime::stub_name(slow_call_address),
1718 alloc->jvms()->bci(),
1719 TypePtr::BOTTOM);
1720 call->init_req( TypeFunc::Control, slow_region );
1721 call->init_req( TypeFunc::I_O , top() ) ; // does no i/o
1722 call->init_req( TypeFunc::Memory , slow_mem ); // may gc ptrs
1723 call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) );
1724 call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) );
1725
1726 call->init_req(TypeFunc::Parms+0, klass_node);
1727 if (length != NULL) {
1728 call->init_req(TypeFunc::Parms+1, length);
1729 }
1730
1731 // Copy debug information and adjust JVMState information, then replace
1732 // allocate node with the call
1733 copy_call_debug_info((CallNode *) alloc, call);
1734 if (!always_slow) {
1735 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1736 } else {
1737 // Hook i_o projection to avoid its elimination during allocation
1738 // replacement (when only a slow call is generated).
1739 call->set_req(TypeFunc::I_O, result_phi_i_o);
1740 }
1741 _igvn.replace_node(alloc, call);
1742 transform_later(call);
1743
1744 // Identify the output projections from the allocate node and
1745 // adjust any references to them.
1746 // The control and io projections look like:
1747 //
1748 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl)
1749 // Allocate Catch
1750 // ^---Proj(io) <-------+ ^---CatchProj(io)
1751 //
1752 // We are interested in the CatchProj nodes.
1753 //
1754 extract_call_projections(call);
1755
1756 // An allocate node has separate memory projections for the uses on
1757 // the control and i_o paths. Replace the control memory projection with
1758 // result_phi_rawmem (unless we are only generating a slow call when
1759 // both memory projections are combined)
1760 if (!always_slow && _memproj_fallthrough != NULL) {
1761 for (DUIterator_Fast imax, i = _memproj_fallthrough->fast_outs(imax); i < imax; i++) {
1762 Node *use = _memproj_fallthrough->fast_out(i);
1763 _igvn.rehash_node_delayed(use);
1764 imax -= replace_input(use, _memproj_fallthrough, result_phi_rawmem);
1765 // back up iterator
1766 --i;
1767 }
1768 }
1769 // Now change uses of _memproj_catchall to use _memproj_fallthrough and delete
1770 // _memproj_catchall so we end up with a call that has only 1 memory projection.
1771 if (_memproj_catchall != NULL ) {
1772 if (_memproj_fallthrough == NULL) {
1773 _memproj_fallthrough = new ProjNode(call, TypeFunc::Memory);
1774 transform_later(_memproj_fallthrough);
1775 }
1776 for (DUIterator_Fast imax, i = _memproj_catchall->fast_outs(imax); i < imax; i++) {
1777 Node *use = _memproj_catchall->fast_out(i);
1778 _igvn.rehash_node_delayed(use);
1779 imax -= replace_input(use, _memproj_catchall, _memproj_fallthrough);
1780 // back up iterator
1781 --i;
1782 }
1783 assert(_memproj_catchall->outcnt() == 0, "all uses must be deleted");
1784 _igvn.remove_dead_node(_memproj_catchall);
1785 }
1786
1787 // An allocate node has separate i_o projections for the uses on the control
1788 // and i_o paths. Always replace the control i_o projection with result i_o
1789 // otherwise incoming i_o become dead when only a slow call is generated
1790 // (it is different from memory projections where both projections are
1791 // combined in such case).
1792 if (_ioproj_fallthrough != NULL) {
1793 for (DUIterator_Fast imax, i = _ioproj_fallthrough->fast_outs(imax); i < imax; i++) {
1794 Node *use = _ioproj_fallthrough->fast_out(i);
1795 _igvn.rehash_node_delayed(use);
1796 imax -= replace_input(use, _ioproj_fallthrough, result_phi_i_o);
1797 // back up iterator
1798 --i;
1799 }
1800 }
1801 // Now change uses of _ioproj_catchall to use _ioproj_fallthrough and delete
1802 // _ioproj_catchall so we end up with a call that has only 1 i_o projection.
1803 if (_ioproj_catchall != NULL ) {
1804 if (_ioproj_fallthrough == NULL) {
1805 _ioproj_fallthrough = new ProjNode(call, TypeFunc::I_O);
1806 transform_later(_ioproj_fallthrough);
1807 }
1808 for (DUIterator_Fast imax, i = _ioproj_catchall->fast_outs(imax); i < imax; i++) {
1809 Node *use = _ioproj_catchall->fast_out(i);
1810 _igvn.rehash_node_delayed(use);
1811 imax -= replace_input(use, _ioproj_catchall, _ioproj_fallthrough);
1812 // back up iterator
1813 --i;
1814 }
1815 assert(_ioproj_catchall->outcnt() == 0, "all uses must be deleted");
1816 _igvn.remove_dead_node(_ioproj_catchall);
1817 }
1818
1819 // if we generated only a slow call, we are done
1820 if (always_slow) {
1821 // Now we can unhook i_o.
1822 if (result_phi_i_o->outcnt() > 1) {
1823 call->set_req(TypeFunc::I_O, top());
1824 } else {
1825 assert(result_phi_i_o->unique_ctrl_out() == call, "");
1826 // Case of new array with negative size known during compilation.
1827 // AllocateArrayNode::Ideal() optimization disconnect unreachable
1828 // following code since call to runtime will throw exception.
1829 // As result there will be no users of i_o after the call.
1830 // Leave i_o attached to this call to avoid problems in preceding graph.
1831 }
1832 return;
1833 }
1834
1835
1836 if (_fallthroughcatchproj != NULL) {
1837 ctrl = _fallthroughcatchproj->clone();
1838 transform_later(ctrl);
1839 _igvn.replace_node(_fallthroughcatchproj, result_region);
1840 } else {
1841 ctrl = top();
1842 }
1843 Node *slow_result;
1844 if (_resproj == NULL) {
1845 // no uses of the allocation result
1846 slow_result = top();
1847 } else {
1848 slow_result = _resproj->clone();
1849 transform_later(slow_result);
1850 _igvn.replace_node(_resproj, result_phi_rawoop);
1851 }
1852
1853 // Plug slow-path into result merge point
1854 result_region ->init_req( slow_result_path, ctrl );
1855 result_phi_rawoop->init_req( slow_result_path, slow_result);
1856 result_phi_rawmem->init_req( slow_result_path, _memproj_fallthrough );
1857 transform_later(result_region);
1858 transform_later(result_phi_rawoop);
1859 transform_later(result_phi_rawmem);
1860 transform_later(result_phi_i_o);
1861 // This completes all paths into the result merge point
1862 }
1863
1864
1865 // Helper for PhaseMacroExpand::expand_allocate_common.
1866 // Initializes the newly-allocated storage.
1867 Node*
1868 PhaseMacroExpand::initialize_object(AllocateNode* alloc,
1869 Node* control, Node* rawmem, Node* object,
1870 Node* klass_node, Node* length,
1871 Node* size_in_bytes) {
1872 InitializeNode* init = alloc->initialization();
1873 // Store the klass & mark bits
1874 Node* mark_node = NULL;
1875 // For now only enable fast locking for non-array types
1876 if (UseBiasedLocking && (length == NULL)) {
1877 mark_node = make_load(control, rawmem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
1878 } else {
1879 mark_node = makecon(TypeRawPtr::make((address)markOopDesc::prototype()));
1880 }
1881 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS);
1882
1883 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
1884 int header_size = alloc->minimum_header_size(); // conservatively small
1885
1886 // Array length
1887 if (length != NULL) { // Arrays need length field
1888 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
1889 // conservatively small header size:
1890 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1891 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
1892 if (k->is_array_klass()) // we know the exact header size in most cases:
1893 header_size = Klass::layout_helper_header_size(k->layout_helper());
1894 }
1895
1896 // Clear the object body, if necessary.
1897 if (init == NULL) {
1898 // The init has somehow disappeared; be cautious and clear everything.
1899 //
1900 // This can happen if a node is allocated but an uncommon trap occurs
1901 // immediately. In this case, the Initialize gets associated with the
1902 // trap, and may be placed in a different (outer) loop, if the Allocate
1903 // is in a loop. If (this is rare) the inner loop gets unrolled, then
1904 // there can be two Allocates to one Initialize. The answer in all these
1905 // edge cases is safety first. It is always safe to clear immediately
1906 // within an Allocate, and then (maybe or maybe not) clear some more later.
1907 if (!(UseTLAB && ZeroTLAB)) {
1908 rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
1909 header_size, size_in_bytes,
1910 &_igvn);
1911 }
1912 } else {
1913 if (!init->is_complete()) {
1914 // Try to win by zeroing only what the init does not store.
1915 // We can also try to do some peephole optimizations,
1916 // such as combining some adjacent subword stores.
1917 rawmem = init->complete_stores(control, rawmem, object,
1918 header_size, size_in_bytes, &_igvn);
1919 }
1920 // We have no more use for this link, since the AllocateNode goes away:
1921 init->set_req(InitializeNode::RawAddress, top());
1922 // (If we keep the link, it just confuses the register allocator,
1923 // who thinks he sees a real use of the address by the membar.)
1924 }
1925
1926 return rawmem;
1927 }
1928
1929 // Generate prefetch instructions for next allocations.
1930 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false,
1931 Node*& contended_phi_rawmem,
1932 Node* old_eden_top, Node* new_eden_top,
1933 Node* length) {
1934 enum { fall_in_path = 1, pf_path = 2 };
1935 if( UseTLAB && AllocatePrefetchStyle == 2 ) {
1936 // Generate prefetch allocation with watermark check.
1937 // As an allocation hits the watermark, we will prefetch starting
1938 // at a "distance" away from watermark.
1939
1940 Node *pf_region = new RegionNode(3);
1941 Node *pf_phi_rawmem = new PhiNode( pf_region, Type::MEMORY,
1942 TypeRawPtr::BOTTOM );
1943 // I/O is used for Prefetch
1944 Node *pf_phi_abio = new PhiNode( pf_region, Type::ABIO );
1945
1946 Node *thread = new ThreadLocalNode();
1947 transform_later(thread);
1948
1949 Node *eden_pf_adr = new AddPNode( top()/*not oop*/, thread,
1950 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) );
1951 transform_later(eden_pf_adr);
1952
1953 Node *old_pf_wm = new LoadPNode(needgc_false,
1954 contended_phi_rawmem, eden_pf_adr,
1955 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM,
1956 MemNode::unordered);
1957 transform_later(old_pf_wm);
1958
1959 // check against new_eden_top
1960 Node *need_pf_cmp = new CmpPNode( new_eden_top, old_pf_wm );
1961 transform_later(need_pf_cmp);
1962 Node *need_pf_bol = new BoolNode( need_pf_cmp, BoolTest::ge );
1963 transform_later(need_pf_bol);
1964 IfNode *need_pf_iff = new IfNode( needgc_false, need_pf_bol,
1965 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN );
1966 transform_later(need_pf_iff);
1967
1968 // true node, add prefetchdistance
1969 Node *need_pf_true = new IfTrueNode( need_pf_iff );
1970 transform_later(need_pf_true);
1971
1972 Node *need_pf_false = new IfFalseNode( need_pf_iff );
1973 transform_later(need_pf_false);
1974
1975 Node *new_pf_wmt = new AddPNode( top(), old_pf_wm,
1976 _igvn.MakeConX(AllocatePrefetchDistance) );
1977 transform_later(new_pf_wmt );
1978 new_pf_wmt->set_req(0, need_pf_true);
1979
1980 Node *store_new_wmt = new StorePNode(need_pf_true,
1981 contended_phi_rawmem, eden_pf_adr,
1982 TypeRawPtr::BOTTOM, new_pf_wmt,
1983 MemNode::unordered);
1984 transform_later(store_new_wmt);
1985
1986 // adding prefetches
1987 pf_phi_abio->init_req( fall_in_path, i_o );
1988
1989 Node *prefetch_adr;
1990 Node *prefetch;
1991 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1992 uint step_size = AllocatePrefetchStepSize;
1993 uint distance = 0;
1994
1995 for ( uint i = 0; i < lines; i++ ) {
1996 prefetch_adr = new AddPNode( old_pf_wm, new_pf_wmt,
1997 _igvn.MakeConX(distance) );
1998 transform_later(prefetch_adr);
1999 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
2000 transform_later(prefetch);
2001 distance += step_size;
2002 i_o = prefetch;
2003 }
2004 pf_phi_abio->set_req( pf_path, i_o );
2005
2006 pf_region->init_req( fall_in_path, need_pf_false );
2007 pf_region->init_req( pf_path, need_pf_true );
2008
2009 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem );
2010 pf_phi_rawmem->init_req( pf_path, store_new_wmt );
2011
2012 transform_later(pf_region);
2013 transform_later(pf_phi_rawmem);
2014 transform_later(pf_phi_abio);
2015
2016 needgc_false = pf_region;
2017 contended_phi_rawmem = pf_phi_rawmem;
2018 i_o = pf_phi_abio;
2019 } else if( UseTLAB && AllocatePrefetchStyle == 3 ) {
2020 // Insert a prefetch instruction for each allocation.
2021 // This code is used to generate 1 prefetch instruction per cache line.
2022
2023 // Generate several prefetch instructions.
2024 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
2025 uint step_size = AllocatePrefetchStepSize;
2026 uint distance = AllocatePrefetchDistance;
2027
2028 // Next cache address.
2029 Node *cache_adr = new AddPNode(old_eden_top, old_eden_top,
2030 _igvn.MakeConX(step_size + distance));
2031 transform_later(cache_adr);
2032 cache_adr = new CastP2XNode(needgc_false, cache_adr);
2033 transform_later(cache_adr);
2034 // Address is aligned to execute prefetch to the beginning of cache line size
2035 // (it is important when BIS instruction is used on SPARC as prefetch).
2036 Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1));
2037 cache_adr = new AndXNode(cache_adr, mask);
2038 transform_later(cache_adr);
2039 cache_adr = new CastX2PNode(cache_adr);
2040 transform_later(cache_adr);
2041
2042 // Prefetch
2043 Node *prefetch = new PrefetchAllocationNode( contended_phi_rawmem, cache_adr );
2044 prefetch->set_req(0, needgc_false);
2045 transform_later(prefetch);
2046 contended_phi_rawmem = prefetch;
2047 Node *prefetch_adr;
2048 distance = step_size;
2049 for ( uint i = 1; i < lines; i++ ) {
2050 prefetch_adr = new AddPNode( cache_adr, cache_adr,
2051 _igvn.MakeConX(distance) );
2052 transform_later(prefetch_adr);
2053 prefetch = new PrefetchAllocationNode( contended_phi_rawmem, prefetch_adr );
2054 transform_later(prefetch);
2055 distance += step_size;
2056 contended_phi_rawmem = prefetch;
2057 }
2058 } else if( AllocatePrefetchStyle > 0 ) {
2059 // Insert a prefetch for each allocation only on the fast-path
2060 Node *prefetch_adr;
2061 Node *prefetch;
2062 // Generate several prefetch instructions.
2063 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
2064 uint step_size = AllocatePrefetchStepSize;
2065 uint distance = AllocatePrefetchDistance;
2066 for ( uint i = 0; i < lines; i++ ) {
2067 prefetch_adr = new AddPNode( old_eden_top, new_eden_top,
2068 _igvn.MakeConX(distance) );
2069 transform_later(prefetch_adr);
2070 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
2071 // Do not let it float too high, since if eden_top == eden_end,
2072 // both might be null.
2073 if( i == 0 ) { // Set control for first prefetch, next follows it
2074 prefetch->init_req(0, needgc_false);
2075 }
2076 transform_later(prefetch);
2077 distance += step_size;
2078 i_o = prefetch;
2079 }
2080 }
2081 return i_o;
2082 }
2083
2084
2085 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) {
2086 expand_allocate_common(alloc, NULL,
2087 OptoRuntime::new_instance_Type(),
2088 OptoRuntime::new_instance_Java());
2089 }
2090
2091 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) {
2092 Node* length = alloc->in(AllocateNode::ALength);
2093 InitializeNode* init = alloc->initialization();
2094 Node* klass_node = alloc->in(AllocateNode::KlassNode);
2095 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
2096 address slow_call_address; // Address of slow call
2097 if (init != NULL && init->is_complete_with_arraycopy() &&
2098 k->is_type_array_klass()) {
2099 // Don't zero type array during slow allocation in VM since
2100 // it will be initialized later by arraycopy in compiled code.
2101 slow_call_address = OptoRuntime::new_array_nozero_Java();
2102 } else {
2103 slow_call_address = OptoRuntime::new_array_Java();
2104 }
2105 expand_allocate_common(alloc, length,
2106 OptoRuntime::new_array_Type(),
2107 slow_call_address);
2108 }
2109
2110 //-------------------mark_eliminated_box----------------------------------
2111 //
2112 // During EA obj may point to several objects but after few ideal graph
2113 // transformations (CCP) it may point to only one non escaping object
2114 // (but still using phi), corresponding locks and unlocks will be marked
2115 // for elimination. Later obj could be replaced with a new node (new phi)
2116 // and which does not have escape information. And later after some graph
2117 // reshape other locks and unlocks (which were not marked for elimination
2118 // before) are connected to this new obj (phi) but they still will not be
2119 // marked for elimination since new obj has no escape information.
2120 // Mark all associated (same box and obj) lock and unlock nodes for
2121 // elimination if some of them marked already.
2122 void PhaseMacroExpand::mark_eliminated_box(Node* oldbox, Node* obj) {
2123 if (oldbox->as_BoxLock()->is_eliminated())
2124 return; // This BoxLock node was processed already.
2125
2126 // New implementation (EliminateNestedLocks) has separate BoxLock
2127 // node for each locked region so mark all associated locks/unlocks as
2128 // eliminated even if different objects are referenced in one locked region
2129 // (for example, OSR compilation of nested loop inside locked scope).
2130 if (EliminateNestedLocks ||
2131 oldbox->as_BoxLock()->is_simple_lock_region(NULL, obj)) {
2132 // Box is used only in one lock region. Mark this box as eliminated.
2133 _igvn.hash_delete(oldbox);
2134 oldbox->as_BoxLock()->set_eliminated(); // This changes box's hash value
2135 _igvn.hash_insert(oldbox);
2136
2137 for (uint i = 0; i < oldbox->outcnt(); i++) {
2138 Node* u = oldbox->raw_out(i);
2139 if (u->is_AbstractLock() && !u->as_AbstractLock()->is_non_esc_obj()) {
2140 AbstractLockNode* alock = u->as_AbstractLock();
2141 // Check lock's box since box could be referenced by Lock's debug info.
2142 if (alock->box_node() == oldbox) {
2143 // Mark eliminated all related locks and unlocks.
2144 #ifdef ASSERT
2145 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc4");
2146 #endif
2147 alock->set_non_esc_obj();
2148 }
2149 }
2150 }
2151 return;
2152 }
2153
2154 // Create new "eliminated" BoxLock node and use it in monitor debug info
2155 // instead of oldbox for the same object.
2156 BoxLockNode* newbox = oldbox->clone()->as_BoxLock();
2157
2158 // Note: BoxLock node is marked eliminated only here and it is used
2159 // to indicate that all associated lock and unlock nodes are marked
2160 // for elimination.
2161 newbox->set_eliminated();
2162 transform_later(newbox);
2163
2164 // Replace old box node with new box for all users of the same object.
2165 for (uint i = 0; i < oldbox->outcnt();) {
2166 bool next_edge = true;
2167
2168 Node* u = oldbox->raw_out(i);
2169 if (u->is_AbstractLock()) {
2170 AbstractLockNode* alock = u->as_AbstractLock();
2171 if (alock->box_node() == oldbox && alock->obj_node()->eqv_uncast(obj)) {
2172 // Replace Box and mark eliminated all related locks and unlocks.
2173 #ifdef ASSERT
2174 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc5");
2175 #endif
2176 alock->set_non_esc_obj();
2177 _igvn.rehash_node_delayed(alock);
2178 alock->set_box_node(newbox);
2179 next_edge = false;
2180 }
2181 }
2182 if (u->is_FastLock() && u->as_FastLock()->obj_node()->eqv_uncast(obj)) {
2183 FastLockNode* flock = u->as_FastLock();
2184 assert(flock->box_node() == oldbox, "sanity");
2185 _igvn.rehash_node_delayed(flock);
2186 flock->set_box_node(newbox);
2187 next_edge = false;
2188 }
2189
2190 // Replace old box in monitor debug info.
2191 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) {
2192 SafePointNode* sfn = u->as_SafePoint();
2193 JVMState* youngest_jvms = sfn->jvms();
2194 int max_depth = youngest_jvms->depth();
2195 for (int depth = 1; depth <= max_depth; depth++) {
2196 JVMState* jvms = youngest_jvms->of_depth(depth);
2197 int num_mon = jvms->nof_monitors();
2198 // Loop over monitors
2199 for (int idx = 0; idx < num_mon; idx++) {
2200 Node* obj_node = sfn->monitor_obj(jvms, idx);
2201 Node* box_node = sfn->monitor_box(jvms, idx);
2202 if (box_node == oldbox && obj_node->eqv_uncast(obj)) {
2203 int j = jvms->monitor_box_offset(idx);
2204 _igvn.replace_input_of(u, j, newbox);
2205 next_edge = false;
2206 }
2207 }
2208 }
2209 }
2210 if (next_edge) i++;
2211 }
2212 }
2213
2214 //-----------------------mark_eliminated_locking_nodes-----------------------
2215 void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) {
2216 if (EliminateNestedLocks) {
2217 if (alock->is_nested()) {
2218 assert(alock->box_node()->as_BoxLock()->is_eliminated(), "sanity");
2219 return;
2220 } else if (!alock->is_non_esc_obj()) { // Not eliminated or coarsened
2221 // Only Lock node has JVMState needed here.
2222 // Not that preceding claim is documented anywhere else.
2223 if (alock->jvms() != NULL) {
2224 if (alock->as_Lock()->is_nested_lock_region()) {
2225 // Mark eliminated related nested locks and unlocks.
2226 Node* obj = alock->obj_node();
2227 BoxLockNode* box_node = alock->box_node()->as_BoxLock();
2228 assert(!box_node->is_eliminated(), "should not be marked yet");
2229 // Note: BoxLock node is marked eliminated only here
2230 // and it is used to indicate that all associated lock
2231 // and unlock nodes are marked for elimination.
2232 box_node->set_eliminated(); // Box's hash is always NO_HASH here
2233 for (uint i = 0; i < box_node->outcnt(); i++) {
2234 Node* u = box_node->raw_out(i);
2235 if (u->is_AbstractLock()) {
2236 alock = u->as_AbstractLock();
2237 if (alock->box_node() == box_node) {
2238 // Verify that this Box is referenced only by related locks.
2239 assert(alock->obj_node()->eqv_uncast(obj), "");
2240 // Mark all related locks and unlocks.
2241 #ifdef ASSERT
2242 alock->log_lock_optimization(C, "eliminate_lock_set_nested");
2243 #endif
2244 alock->set_nested();
2245 }
2246 }
2247 }
2248 } else {
2249 #ifdef ASSERT
2250 alock->log_lock_optimization(C, "eliminate_lock_NOT_nested_lock_region");
2251 if (C->log() != NULL)
2252 alock->as_Lock()->is_nested_lock_region(C); // rerun for debugging output
2253 #endif
2254 }
2255 }
2256 return;
2257 }
2258 // Process locks for non escaping object
2259 assert(alock->is_non_esc_obj(), "");
2260 } // EliminateNestedLocks
2261
2262 if (alock->is_non_esc_obj()) { // Lock is used for non escaping object
2263 // Look for all locks of this object and mark them and
2264 // corresponding BoxLock nodes as eliminated.
2265 Node* obj = alock->obj_node();
2266 for (uint j = 0; j < obj->outcnt(); j++) {
2267 Node* o = obj->raw_out(j);
2268 if (o->is_AbstractLock() &&
2269 o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) {
2270 alock = o->as_AbstractLock();
2271 Node* box = alock->box_node();
2272 // Replace old box node with new eliminated box for all users
2273 // of the same object and mark related locks as eliminated.
2274 mark_eliminated_box(box, obj);
2275 }
2276 }
2277 }
2278 }
2279
2280 // we have determined that this lock/unlock can be eliminated, we simply
2281 // eliminate the node without expanding it.
2282 //
2283 // Note: The membar's associated with the lock/unlock are currently not
2284 // eliminated. This should be investigated as a future enhancement.
2285 //
2286 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) {
2287
2288 if (!alock->is_eliminated()) {
2289 return false;
2290 }
2291 #ifdef ASSERT
2292 if (!alock->is_coarsened()) {
2293 // Check that new "eliminated" BoxLock node is created.
2294 BoxLockNode* oldbox = alock->box_node()->as_BoxLock();
2295 assert(oldbox->is_eliminated(), "should be done already");
2296 }
2297 #endif
2298
2299 alock->log_lock_optimization(C, "eliminate_lock");
2300
2301 #ifndef PRODUCT
2302 if (PrintEliminateLocks) {
2303 if (alock->is_Lock()) {
2304 tty->print_cr("++++ Eliminated: %d Lock", alock->_idx);
2305 } else {
2306 tty->print_cr("++++ Eliminated: %d Unlock", alock->_idx);
2307 }
2308 }
2309 #endif
2310
2311 Node* mem = alock->in(TypeFunc::Memory);
2312 Node* ctrl = alock->in(TypeFunc::Control);
2313
2314 extract_call_projections(alock);
2315 // There are 2 projections from the lock. The lock node will
2316 // be deleted when its last use is subsumed below.
2317 assert(alock->outcnt() == 2 &&
2318 _fallthroughproj != NULL &&
2319 _memproj_fallthrough != NULL,
2320 "Unexpected projections from Lock/Unlock");
2321
2322 Node* fallthroughproj = _fallthroughproj;
2323 Node* memproj_fallthrough = _memproj_fallthrough;
2324
2325 // The memory projection from a lock/unlock is RawMem
2326 // The input to a Lock is merged memory, so extract its RawMem input
2327 // (unless the MergeMem has been optimized away.)
2328 if (alock->is_Lock()) {
2329 // Seach for MemBarAcquireLock node and delete it also.
2330 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
2331 assert(membar != NULL && membar->Opcode() == Op_MemBarAcquireLock, "");
2332 Node* ctrlproj = membar->proj_out(TypeFunc::Control);
2333 Node* memproj = membar->proj_out(TypeFunc::Memory);
2334 _igvn.replace_node(ctrlproj, fallthroughproj);
2335 _igvn.replace_node(memproj, memproj_fallthrough);
2336
2337 // Delete FastLock node also if this Lock node is unique user
2338 // (a loop peeling may clone a Lock node).
2339 Node* flock = alock->as_Lock()->fastlock_node();
2340 if (flock->outcnt() == 1) {
2341 assert(flock->unique_out() == alock, "sanity");
2342 _igvn.replace_node(flock, top());
2343 }
2344 }
2345
2346 // Seach for MemBarReleaseLock node and delete it also.
2347 if (alock->is_Unlock() && ctrl != NULL && ctrl->is_Proj() &&
2348 ctrl->in(0)->is_MemBar()) {
2349 MemBarNode* membar = ctrl->in(0)->as_MemBar();
2350 assert(membar->Opcode() == Op_MemBarReleaseLock &&
2351 mem->is_Proj() && membar == mem->in(0), "");
2352 _igvn.replace_node(fallthroughproj, ctrl);
2353 _igvn.replace_node(memproj_fallthrough, mem);
2354 fallthroughproj = ctrl;
2355 memproj_fallthrough = mem;
2356 ctrl = membar->in(TypeFunc::Control);
2357 mem = membar->in(TypeFunc::Memory);
2358 }
2359
2360 _igvn.replace_node(fallthroughproj, ctrl);
2361 _igvn.replace_node(memproj_fallthrough, mem);
2362 return true;
2363 }
2364
2365
2366 //------------------------------expand_lock_node----------------------
2367 void PhaseMacroExpand::expand_lock_node(LockNode *lock) {
2368
2369 Node* ctrl = lock->in(TypeFunc::Control);
2370 Node* mem = lock->in(TypeFunc::Memory);
2371 Node* obj = lock->obj_node();
2372 Node* box = lock->box_node();
2373 Node* flock = lock->fastlock_node();
2374
2375 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2376
2377 // Make the merge point
2378 Node *region;
2379 Node *mem_phi;
2380 Node *slow_path;
2381
2382 if (UseOptoBiasInlining) {
2383 /*
2384 * See the full description in MacroAssembler::biased_locking_enter().
2385 *
2386 * if( (mark_word & biased_lock_mask) == biased_lock_pattern ) {
2387 * // The object is biased.
2388 * proto_node = klass->prototype_header;
2389 * o_node = thread | proto_node;
2390 * x_node = o_node ^ mark_word;
2391 * if( (x_node & ~age_mask) == 0 ) { // Biased to the current thread ?
2392 * // Done.
2393 * } else {
2394 * if( (x_node & biased_lock_mask) != 0 ) {
2395 * // The klass's prototype header is no longer biased.
2396 * cas(&mark_word, mark_word, proto_node)
2397 * goto cas_lock;
2398 * } else {
2399 * // The klass's prototype header is still biased.
2400 * if( (x_node & epoch_mask) != 0 ) { // Expired epoch?
2401 * old = mark_word;
2402 * new = o_node;
2403 * } else {
2404 * // Different thread or anonymous biased.
2405 * old = mark_word & (epoch_mask | age_mask | biased_lock_mask);
2406 * new = thread | old;
2407 * }
2408 * // Try to rebias.
2409 * if( cas(&mark_word, old, new) == 0 ) {
2410 * // Done.
2411 * } else {
2412 * goto slow_path; // Failed.
2413 * }
2414 * }
2415 * }
2416 * } else {
2417 * // The object is not biased.
2418 * cas_lock:
2419 * if( FastLock(obj) == 0 ) {
2420 * // Done.
2421 * } else {
2422 * slow_path:
2423 * OptoRuntime::complete_monitor_locking_Java(obj);
2424 * }
2425 * }
2426 */
2427
2428 region = new RegionNode(5);
2429 // create a Phi for the memory state
2430 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2431
2432 Node* fast_lock_region = new RegionNode(3);
2433 Node* fast_lock_mem_phi = new PhiNode( fast_lock_region, Type::MEMORY, TypeRawPtr::BOTTOM);
2434
2435 // First, check mark word for the biased lock pattern.
2436 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2437
2438 // Get fast path - mark word has the biased lock pattern.
2439 ctrl = opt_bits_test(ctrl, fast_lock_region, 1, mark_node,
2440 markOopDesc::biased_lock_mask_in_place,
2441 markOopDesc::biased_lock_pattern, true);
2442 // fast_lock_region->in(1) is set to slow path.
2443 fast_lock_mem_phi->init_req(1, mem);
2444
2445 // Now check that the lock is biased to the current thread and has
2446 // the same epoch and bias as Klass::_prototype_header.
2447
2448 // Special-case a fresh allocation to avoid building nodes:
2449 Node* klass_node = AllocateNode::Ideal_klass(obj, &_igvn);
2450 if (klass_node == NULL) {
2451 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
2452 klass_node = transform_later(LoadKlassNode::make(_igvn, NULL, mem, k_adr, _igvn.type(k_adr)->is_ptr()));
2453 #ifdef _LP64
2454 if (UseCompressedClassPointers && klass_node->is_DecodeNKlass()) {
2455 assert(klass_node->in(1)->Opcode() == Op_LoadNKlass, "sanity");
2456 klass_node->in(1)->init_req(0, ctrl);
2457 } else
2458 #endif
2459 klass_node->init_req(0, ctrl);
2460 }
2461 Node *proto_node = make_load(ctrl, mem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeX_X, TypeX_X->basic_type());
2462
2463 Node* thread = transform_later(new ThreadLocalNode());
2464 Node* cast_thread = transform_later(new CastP2XNode(ctrl, thread));
2465 Node* o_node = transform_later(new OrXNode(cast_thread, proto_node));
2466 Node* x_node = transform_later(new XorXNode(o_node, mark_node));
2467
2468 // Get slow path - mark word does NOT match the value.
2469 Node* not_biased_ctrl = opt_bits_test(ctrl, region, 3, x_node,
2470 (~markOopDesc::age_mask_in_place), 0);
2471 // region->in(3) is set to fast path - the object is biased to the current thread.
2472 mem_phi->init_req(3, mem);
2473
2474
2475 // Mark word does NOT match the value (thread | Klass::_prototype_header).
2476
2477
2478 // First, check biased pattern.
2479 // Get fast path - _prototype_header has the same biased lock pattern.
2480 ctrl = opt_bits_test(not_biased_ctrl, fast_lock_region, 2, x_node,
2481 markOopDesc::biased_lock_mask_in_place, 0, true);
2482
2483 not_biased_ctrl = fast_lock_region->in(2); // Slow path
2484 // fast_lock_region->in(2) - the prototype header is no longer biased
2485 // and we have to revoke the bias on this object.
2486 // We are going to try to reset the mark of this object to the prototype
2487 // value and fall through to the CAS-based locking scheme.
2488 Node* adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
2489 Node* cas = new StoreXConditionalNode(not_biased_ctrl, mem, adr,
2490 proto_node, mark_node);
2491 transform_later(cas);
2492 Node* proj = transform_later(new SCMemProjNode(cas));
2493 fast_lock_mem_phi->init_req(2, proj);
2494
2495
2496 // Second, check epoch bits.
2497 Node* rebiased_region = new RegionNode(3);
2498 Node* old_phi = new PhiNode( rebiased_region, TypeX_X);
2499 Node* new_phi = new PhiNode( rebiased_region, TypeX_X);
2500
2501 // Get slow path - mark word does NOT match epoch bits.
2502 Node* epoch_ctrl = opt_bits_test(ctrl, rebiased_region, 1, x_node,
2503 markOopDesc::epoch_mask_in_place, 0);
2504 // The epoch of the current bias is not valid, attempt to rebias the object
2505 // toward the current thread.
2506 rebiased_region->init_req(2, epoch_ctrl);
2507 old_phi->init_req(2, mark_node);
2508 new_phi->init_req(2, o_node);
2509
2510 // rebiased_region->in(1) is set to fast path.
2511 // The epoch of the current bias is still valid but we know
2512 // nothing about the owner; it might be set or it might be clear.
2513 Node* cmask = MakeConX(markOopDesc::biased_lock_mask_in_place |
2514 markOopDesc::age_mask_in_place |
2515 markOopDesc::epoch_mask_in_place);
2516 Node* old = transform_later(new AndXNode(mark_node, cmask));
2517 cast_thread = transform_later(new CastP2XNode(ctrl, thread));
2518 Node* new_mark = transform_later(new OrXNode(cast_thread, old));
2519 old_phi->init_req(1, old);
2520 new_phi->init_req(1, new_mark);
2521
2522 transform_later(rebiased_region);
2523 transform_later(old_phi);
2524 transform_later(new_phi);
2525
2526 // Try to acquire the bias of the object using an atomic operation.
2527 // If this fails we will go in to the runtime to revoke the object's bias.
2528 cas = new StoreXConditionalNode(rebiased_region, mem, adr, new_phi, old_phi);
2529 transform_later(cas);
2530 proj = transform_later(new SCMemProjNode(cas));
2531
2532 // Get slow path - Failed to CAS.
2533 not_biased_ctrl = opt_bits_test(rebiased_region, region, 4, cas, 0, 0);
2534 mem_phi->init_req(4, proj);
2535 // region->in(4) is set to fast path - the object is rebiased to the current thread.
2536
2537 // Failed to CAS.
2538 slow_path = new RegionNode(3);
2539 Node *slow_mem = new PhiNode( slow_path, Type::MEMORY, TypeRawPtr::BOTTOM);
2540
2541 slow_path->init_req(1, not_biased_ctrl); // Capture slow-control
2542 slow_mem->init_req(1, proj);
2543
2544 // Call CAS-based locking scheme (FastLock node).
2545
2546 transform_later(fast_lock_region);
2547 transform_later(fast_lock_mem_phi);
2548
2549 // Get slow path - FastLock failed to lock the object.
2550 ctrl = opt_bits_test(fast_lock_region, region, 2, flock, 0, 0);
2551 mem_phi->init_req(2, fast_lock_mem_phi);
2552 // region->in(2) is set to fast path - the object is locked to the current thread.
2553
2554 slow_path->init_req(2, ctrl); // Capture slow-control
2555 slow_mem->init_req(2, fast_lock_mem_phi);
2556
2557 transform_later(slow_path);
2558 transform_later(slow_mem);
2559 // Reset lock's memory edge.
2560 lock->set_req(TypeFunc::Memory, slow_mem);
2561
2562 } else {
2563 region = new RegionNode(3);
2564 // create a Phi for the memory state
2565 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2566
2567 // Optimize test; set region slot 2
2568 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0);
2569 mem_phi->init_req(2, mem);
2570 }
2571
2572 // Make slow path call
2573 CallNode *call = make_slow_call((CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(),
2574 OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path,
2575 obj, box, NULL);
2576
2577 extract_call_projections(call);
2578
2579 // Slow path can only throw asynchronous exceptions, which are always
2580 // de-opted. So the compiler thinks the slow-call can never throw an
2581 // exception. If it DOES throw an exception we would need the debug
2582 // info removed first (since if it throws there is no monitor).
2583 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
2584 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
2585
2586 // Capture slow path
2587 // disconnect fall-through projection from call and create a new one
2588 // hook up users of fall-through projection to region
2589 Node *slow_ctrl = _fallthroughproj->clone();
2590 transform_later(slow_ctrl);
2591 _igvn.hash_delete(_fallthroughproj);
2592 _fallthroughproj->disconnect_inputs(NULL, C);
2593 region->init_req(1, slow_ctrl);
2594 // region inputs are now complete
2595 transform_later(region);
2596 _igvn.replace_node(_fallthroughproj, region);
2597
2598 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory));
2599 mem_phi->init_req(1, memproj );
2600 transform_later(mem_phi);
2601 _igvn.replace_node(_memproj_fallthrough, mem_phi);
2602 }
2603
2604 //------------------------------expand_unlock_node----------------------
2605 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
2606
2607 Node* ctrl = unlock->in(TypeFunc::Control);
2608 Node* mem = unlock->in(TypeFunc::Memory);
2609 Node* obj = unlock->obj_node();
2610 Node* box = unlock->box_node();
2611
2612 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2613
2614 // No need for a null check on unlock
2615
2616 // Make the merge point
2617 Node *region;
2618 Node *mem_phi;
2619
2620 if (UseOptoBiasInlining) {
2621 // Check for biased locking unlock case, which is a no-op.
2622 // See the full description in MacroAssembler::biased_locking_exit().
2623 region = new RegionNode(4);
2624 // create a Phi for the memory state
2625 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2626 mem_phi->init_req(3, mem);
2627
2628 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2629 ctrl = opt_bits_test(ctrl, region, 3, mark_node,
2630 markOopDesc::biased_lock_mask_in_place,
2631 markOopDesc::biased_lock_pattern);
2632 } else {
2633 region = new RegionNode(3);
2634 // create a Phi for the memory state
2635 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2636 }
2637
2638 FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box );
2639 funlock = transform_later( funlock )->as_FastUnlock();
2640 // Optimize test; set region slot 2
2641 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0);
2642 Node *thread = transform_later(new ThreadLocalNode());
2643
2644 CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(),
2645 CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C),
2646 "complete_monitor_unlocking_C", slow_path, obj, box, thread);
2647
2648 extract_call_projections(call);
2649
2650 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
2651 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
2652
2653 // No exceptions for unlocking
2654 // Capture slow path
2655 // disconnect fall-through projection from call and create a new one
2656 // hook up users of fall-through projection to region
2657 Node *slow_ctrl = _fallthroughproj->clone();
2658 transform_later(slow_ctrl);
2659 _igvn.hash_delete(_fallthroughproj);
2660 _fallthroughproj->disconnect_inputs(NULL, C);
2661 region->init_req(1, slow_ctrl);
2662 // region inputs are now complete
2663 transform_later(region);
2664 _igvn.replace_node(_fallthroughproj, region);
2665
2666 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory) );
2667 mem_phi->init_req(1, memproj );
2668 mem_phi->init_req(2, mem);
2669 transform_later(mem_phi);
2670 _igvn.replace_node(_memproj_fallthrough, mem_phi);
2671 }
2672
2673 // A value type is returned from the call but we don't know its
2674 // type. Either we get a buffered value (and nothing needs to be done)
2675 // or one of the values being returned is the klass of the value type
2676 // and we need to allocate a value type instance of that type and
2677 // initialize it with other values being returned. In that case, we
2678 // first try a fast path allocation and initialize the value with the
2679 // value klass's pack handler or we fall back to a runtime call.
2680 void PhaseMacroExpand::expand_mh_intrinsic_return(CallStaticJavaNode* call) {
2681 Node* ret = call->proj_out(TypeFunc::Parms);
2682 if (ret == NULL) {
2683 return;
2684 }
2685 assert(ret->bottom_type()->is_valuetypeptr()->is__Value(), "unexpected return type from MH intrinsic");
2686 const TypeFunc* tf = call->_tf;
2687 const TypeTuple* domain = OptoRuntime::store_value_type_fields_Type()->domain_cc();
2688 const TypeFunc* new_tf = TypeFunc::make(tf->domain_sig(), tf->domain_cc(), tf->range_sig(), domain);
2689 call->_tf = new_tf;
2690 // Make sure the change of type is applied before projections are
2691 // processed by igvn
2692 _igvn.set_type(call, call->Value(&_igvn));
2693 _igvn.set_type(ret, ret->Value(&_igvn));
2694
2695 // Before any new projection is added:
2696 CallProjections* projs = call->extract_projections(true, true);
2697
2698 Node* ctl = new Node(1);
2699 Node* mem = new Node(1);
2700 Node* io = new Node(1);
2701 Node* ex_ctl = new Node(1);
2702 Node* ex_mem = new Node(1);
2703 Node* ex_io = new Node(1);
2704 Node* res = new Node(1);
2705
2706 Node* cast = transform_later(new CastP2XNode(ctl, res));
2707 Node* mask = MakeConX(0x1);
2708 Node* masked = transform_later(new AndXNode(cast, mask));
2709 Node* cmp = transform_later(new CmpXNode(masked, mask));
2710 Node* bol = transform_later(new BoolNode(cmp, BoolTest::eq));
2711 IfNode* allocation_iff = new IfNode(ctl, bol, PROB_MAX, COUNT_UNKNOWN);
2712 transform_later(allocation_iff);
2713 Node* allocation_ctl = transform_later(new IfTrueNode(allocation_iff));
2714 Node* no_allocation_ctl = transform_later(new IfFalseNode(allocation_iff));
2715
2716 Node* no_allocation_res = transform_later(new CheckCastPPNode(no_allocation_ctl, res, TypeValueTypePtr::NOTNULL));
2717
2718 Node* mask2 = MakeConX(-2);
2719 Node* masked2 = transform_later(new AndXNode(cast, mask2));
2720 Node* rawklassptr = transform_later(new CastX2PNode(masked2));
2721 Node* klass_node = transform_later(new CheckCastPPNode(allocation_ctl, rawklassptr, TypeKlassPtr::OBJECT_OR_NULL));
2722
2723 Node* slowpath_bol = NULL;
2724 Node* top_adr = NULL;
2725 Node* old_top = NULL;
2726 Node* new_top = NULL;
2727 if (UseTLAB) {
2728 Node* end_adr = NULL;
2729 set_eden_pointers(top_adr, end_adr);
2730 Node* end = make_load(ctl, mem, end_adr, 0, TypeRawPtr::BOTTOM, T_ADDRESS);
2731 old_top = new LoadPNode(ctl, mem, top_adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, MemNode::unordered);
2732 transform_later(old_top);
2733 Node* layout_val = make_load(NULL, mem, klass_node, in_bytes(Klass::layout_helper_offset()), TypeInt::INT, T_INT);
2734 Node* size_in_bytes = ConvI2X(layout_val);
2735 new_top = new AddPNode(top(), old_top, size_in_bytes);
2736 transform_later(new_top);
2737 Node* slowpath_cmp = new CmpPNode(new_top, end);
2738 transform_later(slowpath_cmp);
2739 slowpath_bol = new BoolNode(slowpath_cmp, BoolTest::ge);
2740 transform_later(slowpath_bol);
2741 } else {
2742 slowpath_bol = intcon(1);
2743 top_adr = top();
2744 old_top = top();
2745 new_top = top();
2746 }
2747 IfNode* slowpath_iff = new IfNode(allocation_ctl, slowpath_bol, PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN);
2748 transform_later(slowpath_iff);
2749
2750 Node* slowpath_true = new IfTrueNode(slowpath_iff);
2751 transform_later(slowpath_true);
2752
2753 CallStaticJavaNode* slow_call = new CallStaticJavaNode(OptoRuntime::store_value_type_fields_Type(),
2754 StubRoutines::store_value_type_fields_to_buf(),
2755 "store_value_type_fields",
2756 call->jvms()->bci(),
2757 TypePtr::BOTTOM);
2758 slow_call->init_req(TypeFunc::Control, slowpath_true);
2759 slow_call->init_req(TypeFunc::Memory, mem);
2760 slow_call->init_req(TypeFunc::I_O, io);
2761 slow_call->init_req(TypeFunc::FramePtr, call->in(TypeFunc::FramePtr));
2762 slow_call->init_req(TypeFunc::ReturnAdr, call->in(TypeFunc::ReturnAdr));
2763 slow_call->init_req(TypeFunc::Parms, res);
2764
2765 Node* slow_ctl = transform_later(new ProjNode(slow_call, TypeFunc::Control));
2766 Node* slow_mem = transform_later(new ProjNode(slow_call, TypeFunc::Memory));
2767 Node* slow_io = transform_later(new ProjNode(slow_call, TypeFunc::I_O));
2768 Node* slow_res = transform_later(new ProjNode(slow_call, TypeFunc::Parms));
2769 Node* slow_catc = transform_later(new CatchNode(slow_ctl, slow_io, 2));
2770 Node* slow_norm = transform_later(new CatchProjNode(slow_catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci));
2771 Node* slow_excp = transform_later(new CatchProjNode(slow_catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci));
2772
2773 Node* ex_r = new RegionNode(3);
2774 Node* ex_mem_phi = new PhiNode(ex_r, Type::MEMORY, TypePtr::BOTTOM);
2775 Node* ex_io_phi = new PhiNode(ex_r, Type::ABIO);
2776 ex_r->init_req(1, slow_excp);
2777 ex_mem_phi->init_req(1, slow_mem);
2778 ex_io_phi->init_req(1, slow_io);
2779 ex_r->init_req(2, ex_ctl);
2780 ex_mem_phi->init_req(2, ex_mem);
2781 ex_io_phi->init_req(2, ex_io);
2782
2783 transform_later(ex_r);
2784 transform_later(ex_mem_phi);
2785 transform_later(ex_io_phi);
2786
2787 Node* slowpath_false = new IfFalseNode(slowpath_iff);
2788 transform_later(slowpath_false);
2789 Node* rawmem = new StorePNode(slowpath_false, mem, top_adr, TypeRawPtr::BOTTOM, new_top, MemNode::unordered);
2790 transform_later(rawmem);
2791 Node* mark_node = NULL;
2792 // For now only enable fast locking for non-array types
2793 if (UseBiasedLocking) {
2794 mark_node = make_load(slowpath_false, rawmem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
2795 } else {
2796 mark_node = makecon(TypeRawPtr::make((address)markOopDesc::prototype()));
2797 }
2798 rawmem = make_store(slowpath_false, rawmem, old_top, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS);
2799 rawmem = make_store(slowpath_false, rawmem, old_top, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
2800 if (UseCompressedClassPointers) {
2801 rawmem = make_store(slowpath_false, rawmem, old_top, oopDesc::klass_gap_offset_in_bytes(), intcon(0), T_INT);
2802 }
2803 Node* pack_handler = make_load(slowpath_false, rawmem, klass_node, in_bytes(ValueKlass::pack_handler_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
2804
2805 CallLeafNoFPNode* handler_call = new CallLeafNoFPNode(OptoRuntime::pack_value_type_Type(),
2806 NULL,
2807 "pack handler",
2808 TypeRawPtr::BOTTOM);
2809 handler_call->init_req(TypeFunc::Control, slowpath_false);
2810 handler_call->init_req(TypeFunc::Memory, rawmem);
2811 handler_call->init_req(TypeFunc::I_O, top());
2812 handler_call->init_req(TypeFunc::FramePtr, call->in(TypeFunc::FramePtr));
2813 handler_call->init_req(TypeFunc::ReturnAdr, top());
2814 handler_call->init_req(TypeFunc::Parms, pack_handler);
2815 handler_call->init_req(TypeFunc::Parms+1, old_top);
2816
2817 // We don't know how many values are returned. This assumes the
2818 // worst case, that all available registers are used.
2819 for (uint i = TypeFunc::Parms+1; i < domain->cnt(); i++) {
2820 if (domain->field_at(i) == Type::HALF) {
2821 slow_call->init_req(i, top());
2822 handler_call->init_req(i+1, top());
2823 continue;
2824 }
2825 Node* proj = transform_later(new ProjNode(call, i));
2826 slow_call->init_req(i, proj);
2827 handler_call->init_req(i+1, proj);
2828 }
2829
2830 // We can safepoint at that new call
2831 copy_call_debug_info(call, slow_call);
2832 transform_later(slow_call);
2833 transform_later(handler_call);
2834
2835 Node* handler_ctl = transform_later(new ProjNode(handler_call, TypeFunc::Control));
2836 rawmem = transform_later(new ProjNode(handler_call, TypeFunc::Memory));
2837 Node* slowpath_false_res = transform_later(new ProjNode(handler_call, TypeFunc::Parms));
2838
2839 MergeMemNode* slowpath_false_mem = MergeMemNode::make(mem);
2840 slowpath_false_mem->set_memory_at(Compile::AliasIdxRaw, rawmem);
2841 transform_later(slowpath_false_mem);
2842
2843 Node* r = new RegionNode(4);
2844 Node* mem_phi = new PhiNode(r, Type::MEMORY, TypePtr::BOTTOM);
2845 Node* io_phi = new PhiNode(r, Type::ABIO);
2846 Node* res_phi = new PhiNode(r, ret->bottom_type());
2847
2848 r->init_req(1, no_allocation_ctl);
2849 mem_phi->init_req(1, mem);
2850 io_phi->init_req(1, io);
2851 res_phi->init_req(1, no_allocation_res);
2852 r->init_req(2, slow_norm);
2853 mem_phi->init_req(2, slow_mem);
2854 io_phi->init_req(2, slow_io);
2855 res_phi->init_req(2, slow_res);
2856 r->init_req(3, handler_ctl);
2857 mem_phi->init_req(3, slowpath_false_mem);
2858 io_phi->init_req(3, io);
2859 res_phi->init_req(3, slowpath_false_res);
2860
2861 transform_later(r);
2862 transform_later(mem_phi);
2863 transform_later(io_phi);
2864 transform_later(res_phi);
2865
2866 assert(projs->nb_resproj == 1, "unexpected number of results");
2867 _igvn.replace_in_uses(projs->fallthrough_catchproj, r);
2868 _igvn.replace_in_uses(projs->fallthrough_memproj, mem_phi);
2869 _igvn.replace_in_uses(projs->fallthrough_ioproj, io_phi);
2870 _igvn.replace_in_uses(projs->resproj[0], res_phi);
2871 _igvn.replace_in_uses(projs->catchall_catchproj, ex_r);
2872 _igvn.replace_in_uses(projs->catchall_memproj, ex_mem_phi);
2873 _igvn.replace_in_uses(projs->catchall_ioproj, ex_io_phi);
2874
2875 _igvn.replace_node(ctl, projs->fallthrough_catchproj);
2876 _igvn.replace_node(mem, projs->fallthrough_memproj);
2877 _igvn.replace_node(io, projs->fallthrough_ioproj);
2878 _igvn.replace_node(res, projs->resproj[0]);
2879 _igvn.replace_node(ex_ctl, projs->catchall_catchproj);
2880 _igvn.replace_node(ex_mem, projs->catchall_memproj);
2881 _igvn.replace_node(ex_io, projs->catchall_ioproj);
2882 }
2883
2884 //---------------------------eliminate_macro_nodes----------------------
2885 // Eliminate scalar replaced allocations and associated locks.
2886 void PhaseMacroExpand::eliminate_macro_nodes() {
2887 if (C->macro_count() == 0)
2888 return;
2889
2890 // First, attempt to eliminate locks
2891 int cnt = C->macro_count();
2892 for (int i=0; i < cnt; i++) {
2893 Node *n = C->macro_node(i);
2894 if (n->is_AbstractLock()) { // Lock and Unlock nodes
2895 // Before elimination mark all associated (same box and obj)
2896 // lock and unlock nodes.
2897 mark_eliminated_locking_nodes(n->as_AbstractLock());
2898 }
2899 }
2900 bool progress = true;
2901 while (progress) {
2902 progress = false;
2903 for (int i = C->macro_count(); i > 0; i--) {
2904 Node * n = C->macro_node(i-1);
2905 bool success = false;
2906 debug_only(int old_macro_count = C->macro_count(););
2907 if (n->is_AbstractLock()) {
2908 success = eliminate_locking_node(n->as_AbstractLock());
2909 }
2910 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2911 progress = progress || success;
2912 }
2913 }
2914 // Next, attempt to eliminate allocations
2915 _has_locks = false;
2916 progress = true;
2917 while (progress) {
2918 progress = false;
2919 for (int i = C->macro_count(); i > 0; i--) {
2920 Node * n = C->macro_node(i-1);
2921 bool success = false;
2922 debug_only(int old_macro_count = C->macro_count(););
2923 switch (n->class_id()) {
2924 case Node::Class_Allocate:
2925 case Node::Class_AllocateArray:
2926 success = eliminate_allocate_node(n->as_Allocate());
2927 break;
2928 case Node::Class_CallStaticJava: {
2929 CallStaticJavaNode* call = n->as_CallStaticJava();
2930 if (!call->method()->is_method_handle_intrinsic()) {
2931 success = eliminate_boxing_node(n->as_CallStaticJava());
2932 }
2933 break;
2934 }
2935 case Node::Class_Lock:
2936 case Node::Class_Unlock:
2937 assert(!n->as_AbstractLock()->is_eliminated(), "sanity");
2938 _has_locks = true;
2939 break;
2940 case Node::Class_ArrayCopy:
2941 break;
2942 case Node::Class_OuterStripMinedLoop:
2943 break;
2944 default:
2945 assert(n->Opcode() == Op_LoopLimit ||
2946 n->Opcode() == Op_Opaque1 ||
2947 n->Opcode() == Op_Opaque2 ||
2948 n->Opcode() == Op_Opaque3, "unknown node type in macro list");
2949 }
2950 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2951 progress = progress || success;
2952 }
2953 }
2954 }
2955
2956 //------------------------------expand_macro_nodes----------------------
2957 // Returns true if a failure occurred.
2958 bool PhaseMacroExpand::expand_macro_nodes() {
2959 // Last attempt to eliminate macro nodes.
2960 eliminate_macro_nodes();
2961
2962 // Make sure expansion will not cause node limit to be exceeded.
2963 // Worst case is a macro node gets expanded into about 200 nodes.
2964 // Allow 50% more for optimization.
2965 if (C->check_node_count(C->macro_count() * 300, "out of nodes before macro expansion" ) )
2966 return true;
2967
2968 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations.
2969 bool progress = true;
2970 while (progress) {
2971 progress = false;
2972 for (int i = C->macro_count(); i > 0; i--) {
2973 Node * n = C->macro_node(i-1);
2974 bool success = false;
2975 debug_only(int old_macro_count = C->macro_count(););
2976 if (n->Opcode() == Op_LoopLimit) {
2977 // Remove it from macro list and put on IGVN worklist to optimize.
2978 C->remove_macro_node(n);
2979 _igvn._worklist.push(n);
2980 success = true;
2981 } else if (n->Opcode() == Op_CallStaticJava) {
2982 CallStaticJavaNode* call = n->as_CallStaticJava();
2983 if (!call->method()->is_method_handle_intrinsic()) {
2984 // Remove it from macro list and put on IGVN worklist to optimize.
2985 C->remove_macro_node(n);
2986 _igvn._worklist.push(n);
2987 success = true;
2988 }
2989 } else if (n->Opcode() == Op_Opaque1 || n->Opcode() == Op_Opaque2) {
2990 _igvn.replace_node(n, n->in(1));
2991 success = true;
2992 #if INCLUDE_RTM_OPT
2993 } else if ((n->Opcode() == Op_Opaque3) && ((Opaque3Node*)n)->rtm_opt()) {
2994 assert(C->profile_rtm(), "should be used only in rtm deoptimization code");
2995 assert((n->outcnt() == 1) && n->unique_out()->is_Cmp(), "");
2996 Node* cmp = n->unique_out();
2997 #ifdef ASSERT
2998 // Validate graph.
2999 assert((cmp->outcnt() == 1) && cmp->unique_out()->is_Bool(), "");
3000 BoolNode* bol = cmp->unique_out()->as_Bool();
3001 assert((bol->outcnt() == 1) && bol->unique_out()->is_If() &&
3002 (bol->_test._test == BoolTest::ne), "");
3003 IfNode* ifn = bol->unique_out()->as_If();
3004 assert((ifn->outcnt() == 2) &&
3005 ifn->proj_out(1)->is_uncommon_trap_proj(Deoptimization::Reason_rtm_state_change) != NULL, "");
3006 #endif
3007 Node* repl = n->in(1);
3008 if (!_has_locks) {
3009 // Remove RTM state check if there are no locks in the code.
3010 // Replace input to compare the same value.
3011 repl = (cmp->in(1) == n) ? cmp->in(2) : cmp->in(1);
3012 }
3013 _igvn.replace_node(n, repl);
3014 success = true;
3015 #endif
3016 } else if (n->Opcode() == Op_OuterStripMinedLoop) {
3017 n->as_OuterStripMinedLoop()->adjust_strip_mined_loop(&_igvn);
3018 C->remove_macro_node(n);
3019 success = true;
3020 }
3021 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
3022 progress = progress || success;
3023 }
3024 }
3025
3026 // expand arraycopy "macro" nodes first
3027 // For ReduceBulkZeroing, we must first process all arraycopy nodes
3028 // before the allocate nodes are expanded.
3029 int macro_idx = C->macro_count() - 1;
3030 while (macro_idx >= 0) {
3031 Node * n = C->macro_node(macro_idx);
3032 assert(n->is_macro(), "only macro nodes expected here");
3033 if (_igvn.type(n) == Type::TOP || n->in(0)->is_top() ) {
3034 // node is unreachable, so don't try to expand it
3035 C->remove_macro_node(n);
3036 } else if (n->is_ArrayCopy()){
3037 int macro_count = C->macro_count();
3038 expand_arraycopy_node(n->as_ArrayCopy());
3039 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
3040 }
3041 if (C->failing()) return true;
3042 macro_idx --;
3043 }
3044
3045 // expand "macro" nodes
3046 // nodes are removed from the macro list as they are processed
3047 while (C->macro_count() > 0) {
3048 int macro_count = C->macro_count();
3049 Node * n = C->macro_node(macro_count-1);
3050 assert(n->is_macro(), "only macro nodes expected here");
3051 if (_igvn.type(n) == Type::TOP || n->in(0)->is_top() ) {
3052 // node is unreachable, so don't try to expand it
3053 C->remove_macro_node(n);
3054 continue;
3055 }
3056 switch (n->class_id()) {
3057 case Node::Class_Allocate:
3058 expand_allocate(n->as_Allocate());
3059 break;
3060 case Node::Class_AllocateArray:
3061 expand_allocate_array(n->as_AllocateArray());
3062 break;
3063 case Node::Class_Lock:
3064 expand_lock_node(n->as_Lock());
3065 break;
3066 case Node::Class_Unlock:
3067 expand_unlock_node(n->as_Unlock());
3068 break;
3069 case Node::Class_CallStaticJava:
3070 expand_mh_intrinsic_return(n->as_CallStaticJava());
3071 C->remove_macro_node(n);
3072 break;
3073 default:
3074 assert(false, "unknown node type in macro list");
3075 }
3076 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
3077 if (C->failing()) return true;
3078 }
3079
3080 _igvn.set_delay_transform(false);
3081 _igvn.optimize();
3082 if (C->failing()) return true;
3083 return false;
3084 }