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 "libadt/vectset.hpp"
28 #include "opto/addnode.hpp"
29 #include "opto/arraycopynode.hpp"
30 #include "opto/callnode.hpp"
31 #include "opto/castnode.hpp"
32 #include "opto/cfgnode.hpp"
33 #include "opto/compile.hpp"
34 #include "opto/convertnode.hpp"
35 #include "opto/graphKit.hpp"
36 #include "opto/locknode.hpp"
37 #include "opto/loopnode.hpp"
38 #include "opto/macro.hpp"
39 #include "opto/memnode.hpp"
40 #include "opto/narrowptrnode.hpp"
41 #include "opto/node.hpp"
42 #include "opto/opaquenode.hpp"
43 #include "opto/phaseX.hpp"
44 #include "opto/rootnode.hpp"
45 #include "opto/runtime.hpp"
46 #include "opto/subnode.hpp"
47 #include "opto/type.hpp"
48 #include "opto/valuetypenode.hpp"
49 #include "runtime/sharedRuntime.hpp"
50
51
52 //
53 // Replace any references to "oldref" in inputs to "use" with "newref".
54 // Returns the number of replacements made.
55 //
56 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) {
57 int nreplacements = 0;
58 uint req = use->req();
59 for (uint j = 0; j < use->len(); j++) {
60 Node *uin = use->in(j);
61 if (uin == oldref) {
62 if (j < req)
63 use->set_req(j, newref);
64 else
65 use->set_prec(j, newref);
66 nreplacements++;
67 } else if (j >= req && uin == NULL) {
68 break;
69 }
70 }
71 return nreplacements;
72 }
73
74 void PhaseMacroExpand::copy_call_debug_info(CallNode *oldcall, CallNode * newcall) {
75 // Copy debug information and adjust JVMState information
76 uint old_dbg_start = oldcall->tf()->domain_sig()->cnt();
77 uint new_dbg_start = newcall->tf()->domain_sig()->cnt();
78 int jvms_adj = new_dbg_start - old_dbg_start;
79 assert (new_dbg_start == newcall->req(), "argument count mismatch");
80
81 // SafePointScalarObject node could be referenced several times in debug info.
82 // Use Dict to record cloned nodes.
83 Dict* sosn_map = new Dict(cmpkey,hashkey);
84 for (uint i = old_dbg_start; i < oldcall->req(); i++) {
85 Node* old_in = oldcall->in(i);
86 // Clone old SafePointScalarObjectNodes, adjusting their field contents.
87 if (old_in != NULL && old_in->is_SafePointScalarObject()) {
88 SafePointScalarObjectNode* old_sosn = old_in->as_SafePointScalarObject();
89 uint old_unique = C->unique();
90 Node* new_in = old_sosn->clone(sosn_map);
91 if (old_unique != C->unique()) { // New node?
92 new_in->set_req(0, C->root()); // reset control edge
93 new_in = transform_later(new_in); // Register new node.
94 }
95 old_in = new_in;
96 }
97 newcall->add_req(old_in);
98 }
99
100 // JVMS may be shared so clone it before we modify it
101 newcall->set_jvms(oldcall->jvms() != NULL ? oldcall->jvms()->clone_deep(C) : NULL);
102 for (JVMState *jvms = newcall->jvms(); jvms != NULL; jvms = jvms->caller()) {
103 jvms->set_map(newcall);
104 jvms->set_locoff(jvms->locoff()+jvms_adj);
105 jvms->set_stkoff(jvms->stkoff()+jvms_adj);
106 jvms->set_monoff(jvms->monoff()+jvms_adj);
107 jvms->set_scloff(jvms->scloff()+jvms_adj);
108 jvms->set_endoff(jvms->endoff()+jvms_adj);
109 }
110 }
111
112 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word, int mask, int bits, bool return_fast_path) {
113 Node* cmp;
114 if (mask != 0) {
115 Node* and_node = transform_later(new AndXNode(word, MakeConX(mask)));
116 cmp = transform_later(new CmpXNode(and_node, MakeConX(bits)));
117 } else {
118 cmp = word;
119 }
120 Node* bol = transform_later(new BoolNode(cmp, BoolTest::ne));
121 IfNode* iff = new IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN );
122 transform_later(iff);
123
124 // Fast path taken.
125 Node *fast_taken = transform_later(new IfFalseNode(iff));
126
127 // Fast path not-taken, i.e. slow path
128 Node *slow_taken = transform_later(new IfTrueNode(iff));
129
130 if (return_fast_path) {
131 region->init_req(edge, slow_taken); // Capture slow-control
132 return fast_taken;
133 } else {
134 region->init_req(edge, fast_taken); // Capture fast-control
135 return slow_taken;
136 }
137 }
138
139 //--------------------copy_predefined_input_for_runtime_call--------------------
140 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) {
141 // Set fixed predefined input arguments
142 call->init_req( TypeFunc::Control, ctrl );
143 call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) );
144 call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ?????
145 call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) );
146 call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) );
147 }
148
149 //------------------------------make_slow_call---------------------------------
150 CallNode* PhaseMacroExpand::make_slow_call(CallNode *oldcall, const TypeFunc* slow_call_type,
151 address slow_call, const char* leaf_name, Node* slow_path,
152 Node* parm0, Node* parm1, Node* parm2) {
153
154 // Slow-path call
155 CallNode *call = leaf_name
156 ? (CallNode*)new CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM )
157 : (CallNode*)new CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), oldcall->jvms()->bci(), TypeRawPtr::BOTTOM );
158
159 // Slow path call has no side-effects, uses few values
160 copy_predefined_input_for_runtime_call(slow_path, oldcall, call );
161 if (parm0 != NULL) call->init_req(TypeFunc::Parms+0, parm0);
162 if (parm1 != NULL) call->init_req(TypeFunc::Parms+1, parm1);
163 if (parm2 != NULL) call->init_req(TypeFunc::Parms+2, parm2);
164 copy_call_debug_info(oldcall, call);
165 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
166 _igvn.replace_node(oldcall, call);
167 transform_later(call);
168
169 return call;
170 }
171
172 void PhaseMacroExpand::extract_call_projections(CallNode *call) {
173 _fallthroughproj = NULL;
174 _fallthroughcatchproj = NULL;
175 _ioproj_fallthrough = NULL;
176 _ioproj_catchall = NULL;
177 _catchallcatchproj = NULL;
178 _memproj_fallthrough = NULL;
179 _memproj_catchall = NULL;
180 _resproj = NULL;
181 for (DUIterator_Fast imax, i = call->fast_outs(imax); i < imax; i++) {
182 ProjNode *pn = call->fast_out(i)->as_Proj();
183 switch (pn->_con) {
184 case TypeFunc::Control:
185 {
186 // For Control (fallthrough) and I_O (catch_all_index) we have CatchProj -> Catch -> Proj
187 _fallthroughproj = pn;
188 DUIterator_Fast jmax, j = pn->fast_outs(jmax);
189 const Node *cn = pn->fast_out(j);
190 if (cn->is_Catch()) {
191 ProjNode *cpn = NULL;
192 for (DUIterator_Fast kmax, k = cn->fast_outs(kmax); k < kmax; k++) {
193 cpn = cn->fast_out(k)->as_Proj();
194 assert(cpn->is_CatchProj(), "must be a CatchProjNode");
195 if (cpn->_con == CatchProjNode::fall_through_index)
196 _fallthroughcatchproj = cpn;
197 else {
198 assert(cpn->_con == CatchProjNode::catch_all_index, "must be correct index.");
199 _catchallcatchproj = cpn;
200 }
201 }
202 }
203 break;
204 }
205 case TypeFunc::I_O:
206 if (pn->_is_io_use)
207 _ioproj_catchall = pn;
208 else
209 _ioproj_fallthrough = pn;
210 break;
211 case TypeFunc::Memory:
212 if (pn->_is_io_use)
213 _memproj_catchall = pn;
214 else
215 _memproj_fallthrough = pn;
216 break;
217 case TypeFunc::Parms:
218 _resproj = pn;
219 break;
220 default:
221 assert(false, "unexpected projection from allocation node.");
222 }
223 }
224
225 }
226
227 // Eliminate a card mark sequence. p2x is a ConvP2XNode
228 void PhaseMacroExpand::eliminate_card_mark(Node* p2x) {
229 assert(p2x->Opcode() == Op_CastP2X, "ConvP2XNode required");
230 if (!UseG1GC) {
231 // vanilla/CMS post barrier
232 Node *shift = p2x->unique_out();
233 Node *addp = shift->unique_out();
234 for (DUIterator_Last jmin, j = addp->last_outs(jmin); j >= jmin; --j) {
235 Node *mem = addp->last_out(j);
236 if (UseCondCardMark && mem->is_Load()) {
237 assert(mem->Opcode() == Op_LoadB, "unexpected code shape");
238 // The load is checking if the card has been written so
239 // replace it with zero to fold the test.
240 _igvn.replace_node(mem, intcon(0));
241 continue;
242 }
243 assert(mem->is_Store(), "store required");
244 _igvn.replace_node(mem, mem->in(MemNode::Memory));
245 }
246 } else {
247 // G1 pre/post barriers
248 assert(p2x->outcnt() <= 2, "expects 1 or 2 users: Xor and URShift nodes");
249 // It could be only one user, URShift node, in Object.clone() intrinsic
250 // but the new allocation is passed to arraycopy stub and it could not
251 // be scalar replaced. So we don't check the case.
252
253 // An other case of only one user (Xor) is when the value check for NULL
254 // in G1 post barrier is folded after CCP so the code which used URShift
255 // is removed.
256
257 // Take Region node before eliminating post barrier since it also
258 // eliminates CastP2X node when it has only one user.
259 Node* this_region = p2x->in(0);
260 assert(this_region != NULL, "");
261
262 // Remove G1 post barrier.
263
264 // Search for CastP2X->Xor->URShift->Cmp path which
265 // checks if the store done to a different from the value's region.
266 // And replace Cmp with #0 (false) to collapse G1 post barrier.
267 Node* xorx = p2x->find_out_with(Op_XorX);
268 if (xorx != NULL) {
269 Node* shift = xorx->unique_out();
270 Node* cmpx = shift->unique_out();
271 assert(cmpx->is_Cmp() && cmpx->unique_out()->is_Bool() &&
272 cmpx->unique_out()->as_Bool()->_test._test == BoolTest::ne,
273 "missing region check in G1 post barrier");
274 _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ));
275
276 // Remove G1 pre barrier.
277
278 // Search "if (marking != 0)" check and set it to "false".
279 // There is no G1 pre barrier if previous stored value is NULL
280 // (for example, after initialization).
281 if (this_region->is_Region() && this_region->req() == 3) {
282 int ind = 1;
283 if (!this_region->in(ind)->is_IfFalse()) {
284 ind = 2;
285 }
286 if (this_region->in(ind)->is_IfFalse() &&
287 this_region->in(ind)->in(0)->Opcode() == Op_If) {
288 Node* bol = this_region->in(ind)->in(0)->in(1);
289 assert(bol->is_Bool(), "");
290 cmpx = bol->in(1);
291 if (bol->as_Bool()->_test._test == BoolTest::ne &&
292 cmpx->is_Cmp() && cmpx->in(2) == intcon(0) &&
293 cmpx->in(1)->is_Load()) {
294 Node* adr = cmpx->in(1)->as_Load()->in(MemNode::Address);
295 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() +
296 SATBMarkQueue::byte_offset_of_active());
297 if (adr->is_AddP() && adr->in(AddPNode::Base) == top() &&
298 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
299 adr->in(AddPNode::Offset) == MakeConX(marking_offset)) {
300 _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ));
301 }
302 }
303 }
304 }
305 } else {
306 assert(!GraphKit::use_ReduceInitialCardMarks(), "can only happen with card marking");
307 // This is a G1 post barrier emitted by the Object.clone() intrinsic.
308 // Search for the CastP2X->URShiftX->AddP->LoadB->Cmp path which checks if the card
309 // is marked as young_gen and replace the Cmp with 0 (false) to collapse the barrier.
310 Node* shift = p2x->find_out_with(Op_URShiftX);
311 assert(shift != NULL, "missing G1 post barrier");
312 Node* addp = shift->unique_out();
313 Node* load = addp->find_out_with(Op_LoadB);
314 assert(load != NULL, "missing G1 post barrier");
315 Node* cmpx = load->unique_out();
316 assert(cmpx->is_Cmp() && cmpx->unique_out()->is_Bool() &&
317 cmpx->unique_out()->as_Bool()->_test._test == BoolTest::ne,
318 "missing card value check in G1 post barrier");
319 _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ));
320 // There is no G1 pre barrier in this case
321 }
322 // Now CastP2X can be removed since it is used only on dead path
323 // which currently still alive until igvn optimize it.
324 assert(p2x->outcnt() == 0 || p2x->unique_out()->Opcode() == Op_URShiftX, "");
325 _igvn.replace_node(p2x, top());
326 }
327 }
328
329 // Search for a memory operation for the specified memory slice.
330 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) {
331 Node *orig_mem = mem;
332 Node *alloc_mem = alloc->in(TypeFunc::Memory);
333 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr();
334 while (true) {
335 if (mem == alloc_mem || mem == start_mem ) {
336 return mem; // hit one of our sentinels
337 } else if (mem->is_MergeMem()) {
338 mem = mem->as_MergeMem()->memory_at(alias_idx);
339 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) {
340 Node *in = mem->in(0);
341 // we can safely skip over safepoints, calls, locks and membars because we
342 // already know that the object is safe to eliminate.
343 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) {
344 return in;
345 } else if (in->is_Call()) {
346 CallNode *call = in->as_Call();
347 if (call->may_modify(tinst, phase)) {
348 assert(call->is_ArrayCopy(), "ArrayCopy is the only call node that doesn't make allocation escape");
349 if (call->as_ArrayCopy()->modifies(offset, offset, phase, false)) {
350 return in;
351 }
352 }
353 mem = in->in(TypeFunc::Memory);
354 } else if (in->is_MemBar()) {
355 ArrayCopyNode* ac = NULL;
356 if (ArrayCopyNode::may_modify(tinst, in->as_MemBar(), phase, ac)) {
357 assert(ac != NULL && ac->is_clonebasic(), "Only basic clone is a non escaping clone");
358 return ac;
359 }
360 mem = in->in(TypeFunc::Memory);
361 } else {
362 assert(false, "unexpected projection");
363 }
364 } else if (mem->is_Store()) {
365 const TypePtr* atype = mem->as_Store()->adr_type();
366 int adr_idx = phase->C->get_alias_index(atype);
367 if (adr_idx == alias_idx) {
368 assert(atype->isa_oopptr(), "address type must be oopptr");
369 int adr_offset = atype->flattened_offset();
370 uint adr_iid = atype->is_oopptr()->instance_id();
371 // Array elements references have the same alias_idx
372 // but different offset and different instance_id.
373 if (adr_offset == offset && adr_iid == alloc->_idx)
374 return mem;
375 } else {
376 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw");
377 }
378 mem = mem->in(MemNode::Memory);
379 } else if (mem->is_ClearArray()) {
380 if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) {
381 // Can not bypass initialization of the instance
382 // we are looking.
383 debug_only(intptr_t offset;)
384 assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity");
385 InitializeNode* init = alloc->as_Allocate()->initialization();
386 // We are looking for stored value, return Initialize node
387 // or memory edge from Allocate node.
388 if (init != NULL)
389 return init;
390 else
391 return alloc->in(TypeFunc::Memory); // It will produce zero value (see callers).
392 }
393 // Otherwise skip it (the call updated 'mem' value).
394 } else if (mem->Opcode() == Op_SCMemProj) {
395 mem = mem->in(0);
396 Node* adr = NULL;
397 if (mem->is_LoadStore()) {
398 adr = mem->in(MemNode::Address);
399 } else {
400 assert(mem->Opcode() == Op_EncodeISOArray ||
401 mem->Opcode() == Op_StrCompressedCopy, "sanity");
402 adr = mem->in(3); // Destination array
403 }
404 const TypePtr* atype = adr->bottom_type()->is_ptr();
405 int adr_idx = phase->C->get_alias_index(atype);
406 if (adr_idx == alias_idx) {
407 DEBUG_ONLY(mem->dump();)
408 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
409 return NULL;
410 }
411 mem = mem->in(MemNode::Memory);
412 } else if (mem->Opcode() == Op_StrInflatedCopy) {
413 Node* adr = mem->in(3); // Destination array
414 const TypePtr* atype = adr->bottom_type()->is_ptr();
415 int adr_idx = phase->C->get_alias_index(atype);
416 if (adr_idx == alias_idx) {
417 DEBUG_ONLY(mem->dump();)
418 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field");
419 return NULL;
420 }
421 mem = mem->in(MemNode::Memory);
422 } else {
423 return mem;
424 }
425 assert(mem != orig_mem, "dead memory loop");
426 }
427 }
428
429 // Generate loads from source of the arraycopy for fields of
430 // destination needed at a deoptimization point
431 Node* PhaseMacroExpand::make_arraycopy_load(ArrayCopyNode* ac, intptr_t offset, Node* ctl, Node* mem, BasicType ft, const Type *ftype, AllocateNode *alloc) {
432 BasicType bt = ft;
433 const Type *type = ftype;
434 if (ft == T_NARROWOOP) {
435 bt = T_OBJECT;
436 type = ftype->make_oopptr();
437 }
438 Node* res = NULL;
439 if (ac->is_clonebasic()) {
440 Node* base = ac->in(ArrayCopyNode::Src)->in(AddPNode::Base);
441 Node* adr = _igvn.transform(new AddPNode(base, base, MakeConX(offset)));
442 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset);
443 res = LoadNode::make(_igvn, ctl, mem, adr, adr_type, type, bt, MemNode::unordered, LoadNode::Pinned);
444 } else {
445 if (ac->modifies(offset, offset, &_igvn, true)) {
446 assert(ac->in(ArrayCopyNode::Dest) == alloc->result_cast(), "arraycopy destination should be allocation's result");
447 uint shift = exact_log2(type2aelembytes(bt));
448 Node* diff = _igvn.transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
449 #ifdef _LP64
450 diff = _igvn.transform(new ConvI2LNode(diff));
451 #endif
452 diff = _igvn.transform(new LShiftXNode(diff, intcon(shift)));
453
454 Node* off = _igvn.transform(new AddXNode(MakeConX(offset), diff));
455 Node* base = ac->in(ArrayCopyNode::Src);
456 Node* adr = _igvn.transform(new AddPNode(base, base, off));
457 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset);
458 res = LoadNode::make(_igvn, ctl, mem, adr, adr_type, type, bt, MemNode::unordered, LoadNode::Pinned);
459 }
460 }
461 if (res != NULL) {
462 res = _igvn.transform(res);
463 if (ftype->isa_narrowoop()) {
464 // PhaseMacroExpand::scalar_replacement adds DecodeN nodes
465 res = _igvn.transform(new EncodePNode(res, ftype));
466 }
467 return res;
468 }
469 return NULL;
470 }
471
472 //
473 // Given a Memory Phi, compute a value Phi containing the values from stores
474 // on the input paths.
475 // Note: this function is recursive, its depth is limited by the "level" argument
476 // Returns the computed Phi, or NULL if it cannot compute it.
477 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) {
478 assert(mem->is_Phi(), "sanity");
479 int alias_idx = C->get_alias_index(adr_t);
480 int offset = adr_t->flattened_offset();
481 int instance_id = adr_t->instance_id();
482
483 // Check if an appropriate value phi already exists.
484 Node* region = mem->in(0);
485 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) {
486 Node* phi = region->fast_out(k);
487 if (phi->is_Phi() && phi != mem &&
488 phi->as_Phi()->is_same_inst_field(phi_type, (int)mem->_idx, instance_id, alias_idx, offset)) {
489 return phi;
490 }
491 }
492 // Check if an appropriate new value phi already exists.
493 Node* new_phi = value_phis->find(mem->_idx);
494 if (new_phi != NULL)
495 return new_phi;
496
497 if (level <= 0) {
498 return NULL; // Give up: phi tree too deep
499 }
500 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
501 Node *alloc_mem = alloc->in(TypeFunc::Memory);
502
503 uint length = mem->req();
504 GrowableArray <Node *> values(length, length, NULL, false);
505
506 // create a new Phi for the value
507 PhiNode *phi = new PhiNode(mem->in(0), phi_type, NULL, mem->_idx, instance_id, alias_idx, offset);
508 transform_later(phi);
509 value_phis->push(phi, mem->_idx);
510
511 for (uint j = 1; j < length; j++) {
512 Node *in = mem->in(j);
513 if (in == NULL || in->is_top()) {
514 values.at_put(j, in);
515 } else {
516 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn);
517 if (val == start_mem || val == alloc_mem) {
518 // hit a sentinel, return appropriate 0 value
519 values.at_put(j, _igvn.zerocon(ft));
520 continue;
521 }
522 if (val->is_Initialize()) {
523 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
524 }
525 if (val == NULL) {
526 return NULL; // can't find a value on this path
527 }
528 if (val == mem) {
529 values.at_put(j, mem);
530 } else if (val->is_Store()) {
531 values.at_put(j, val->in(MemNode::ValueIn));
532 } else if(val->is_Proj() && val->in(0) == alloc) {
533 values.at_put(j, _igvn.zerocon(ft));
534 } else if (val->is_Phi()) {
535 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1);
536 if (val == NULL) {
537 return NULL;
538 }
539 values.at_put(j, val);
540 } else if (val->Opcode() == Op_SCMemProj) {
541 assert(val->in(0)->is_LoadStore() ||
542 val->in(0)->Opcode() == Op_EncodeISOArray ||
543 val->in(0)->Opcode() == Op_StrCompressedCopy, "sanity");
544 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
545 return NULL;
546 } else if (val->is_ArrayCopy()) {
547 Node* res = make_arraycopy_load(val->as_ArrayCopy(), offset, val->in(0), val->in(TypeFunc::Memory), ft, phi_type, alloc);
548 if (res == NULL) {
549 return NULL;
550 }
551 values.at_put(j, res);
552 } else {
553 #ifdef ASSERT
554 val->dump();
555 assert(false, "unknown node on this path");
556 #endif
557 return NULL; // unknown node on this path
558 }
559 }
560 }
561 // Set Phi's inputs
562 for (uint j = 1; j < length; j++) {
563 if (values.at(j) == mem) {
564 phi->init_req(j, phi);
565 } else {
566 phi->init_req(j, values.at(j));
567 }
568 }
569 return phi;
570 }
571
572 // Search the last value stored into the object's field.
573 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, Node *sfpt_ctl, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, AllocateNode *alloc) {
574 assert(adr_t->is_known_instance_field(), "instance required");
575 int instance_id = adr_t->instance_id();
576 assert((uint)instance_id == alloc->_idx, "wrong allocation");
577
578 int alias_idx = C->get_alias_index(adr_t);
579 int offset = adr_t->flattened_offset();
580 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
581 Node *alloc_ctrl = alloc->in(TypeFunc::Control);
582 Node *alloc_mem = alloc->in(TypeFunc::Memory);
583 Arena *a = Thread::current()->resource_area();
584 VectorSet visited(a);
585
586 bool done = sfpt_mem == alloc_mem;
587 Node *mem = sfpt_mem;
588 while (!done) {
589 if (visited.test_set(mem->_idx)) {
590 return NULL; // found a loop, give up
591 }
592 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn);
593 if (mem == start_mem || mem == alloc_mem) {
594 done = true; // hit a sentinel, return appropriate 0 value
595 } else if (mem->is_Initialize()) {
596 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
597 if (mem == NULL) {
598 done = true; // Something went wrong.
599 } else if (mem->is_Store()) {
600 const TypePtr* atype = mem->as_Store()->adr_type();
601 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice");
602 done = true;
603 }
604 } else if (mem->is_Store()) {
605 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr();
606 assert(atype != NULL, "address type must be oopptr");
607 assert(C->get_alias_index(atype) == alias_idx &&
608 atype->is_known_instance_field() && atype->flattened_offset() == offset &&
609 atype->instance_id() == instance_id, "store is correct memory slice");
610 done = true;
611 } else if (mem->is_Phi()) {
612 // try to find a phi's unique input
613 Node *unique_input = NULL;
614 Node *top = C->top();
615 for (uint i = 1; i < mem->req(); i++) {
616 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn);
617 if (n == NULL || n == top || n == mem) {
618 continue;
619 } else if (unique_input == NULL) {
620 unique_input = n;
621 } else if (unique_input != n) {
622 unique_input = top;
623 break;
624 }
625 }
626 if (unique_input != NULL && unique_input != top) {
627 mem = unique_input;
628 } else {
629 done = true;
630 }
631 } else if (mem->is_ArrayCopy()) {
632 done = true;
633 } else {
634 assert(false, "unexpected node");
635 }
636 }
637 if (mem != NULL) {
638 if (mem == start_mem || mem == alloc_mem) {
639 // hit a sentinel, return appropriate 0 value
640 return _igvn.zerocon(ft);
641 } else if (mem->is_Store()) {
642 return mem->in(MemNode::ValueIn);
643 } else if (mem->is_Phi()) {
644 // attempt to produce a Phi reflecting the values on the input paths of the Phi
645 Node_Stack value_phis(a, 8);
646 Node * phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit);
647 if (phi != NULL) {
648 return phi;
649 } else {
650 // Kill all new Phis
651 while(value_phis.is_nonempty()) {
652 Node* n = value_phis.node();
653 _igvn.replace_node(n, C->top());
654 value_phis.pop();
655 }
656 }
657 } else if (mem->is_ArrayCopy()) {
658 Node* ctl = mem->in(0);
659 Node* m = mem->in(TypeFunc::Memory);
660 if (sfpt_ctl->is_Proj() && sfpt_ctl->as_Proj()->is_uncommon_trap_proj(Deoptimization::Reason_none)) {
661 // pin the loads in the uncommon trap path
662 ctl = sfpt_ctl;
663 m = sfpt_mem;
664 }
665 return make_arraycopy_load(mem->as_ArrayCopy(), offset, ctl, m, ft, ftype, alloc);
666 }
667 }
668 // Something went wrong.
669 return NULL;
670 }
671
672 // Search the last value stored into the value type's fields.
673 Node* PhaseMacroExpand::value_type_from_mem(Node* mem, Node* ctl, ciValueKlass* vk, const TypeAryPtr* adr_type, int offset, AllocateNode* alloc) {
674 // Subtract the offset of the first field to account for the missing oop header
675 offset -= vk->first_field_offset();
676 // Create a new ValueTypeNode and retrieve the field values from memory
677 ValueTypeNode* vt = ValueTypeNode::make_uninitialized(_igvn, vk)->as_ValueType();
678 for (int i = 0; i < vk->nof_declared_nonstatic_fields(); ++i) {
679 ciType* field_type = vt->field_type(i);
680 int field_offset = offset + vt->field_offset(i);
681 // Each value type field has its own memory slice
682 adr_type = adr_type->with_field_offset(field_offset);
683 Node* value = NULL;
684 if (vt->field_is_flattened(i)) {
685 value = value_type_from_mem(mem, ctl, field_type->as_value_klass(), adr_type, field_offset, alloc);
686 } else {
687 const Type* ft = Type::get_const_type(field_type);
688 BasicType bt = field_type->basic_type();
689 if (UseCompressedOops && !is_java_primitive(bt)) {
690 ft = ft->make_narrowoop();
691 bt = T_NARROWOOP;
692 }
693 value = value_from_mem(mem, ctl, bt, ft, adr_type, alloc);
694 assert(value != NULL, "field value should not be null");
695 if (ft->isa_narrowoop()) {
696 assert(UseCompressedOops, "unexpected narrow oop");
697 value = transform_later(new DecodeNNode(value, value->get_ptr_type()));
698 }
699 }
700 vt->set_field_value(i, value);
701 }
702 return vt;
703 }
704
705 // Check the possibility of scalar replacement.
706 bool PhaseMacroExpand::can_eliminate_allocation(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
707 // Scan the uses of the allocation to check for anything that would
708 // prevent us from eliminating it.
709 NOT_PRODUCT( const char* fail_eliminate = NULL; )
710 DEBUG_ONLY( Node* disq_node = NULL; )
711 bool can_eliminate = true;
712
713 Node* res = alloc->result_cast();
714 const TypeOopPtr* res_type = NULL;
715 if (res == NULL) {
716 // All users were eliminated.
717 } else if (!res->is_CheckCastPP()) {
718 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";)
719 can_eliminate = false;
720 } else {
721 res_type = _igvn.type(res)->isa_oopptr();
722 if (res_type == NULL) {
723 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";)
724 can_eliminate = false;
725 } else if (res_type->isa_aryptr()) {
726 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1);
727 if (length < 0) {
728 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";)
729 can_eliminate = false;
730 }
731 }
732 }
733
734 if (can_eliminate && res != NULL) {
735 for (DUIterator_Fast jmax, j = res->fast_outs(jmax);
736 j < jmax && can_eliminate; j++) {
737 Node* use = res->fast_out(j);
738
739 if (use->is_AddP()) {
740 const TypePtr* addp_type = _igvn.type(use)->is_ptr();
741 int offset = addp_type->offset();
742
743 if (offset == Type::OffsetTop || offset == Type::OffsetBot) {
744 NOT_PRODUCT(fail_eliminate = "Undefined field referrence";)
745 can_eliminate = false;
746 break;
747 }
748 for (DUIterator_Fast kmax, k = use->fast_outs(kmax);
749 k < kmax && can_eliminate; k++) {
750 Node* n = use->fast_out(k);
751 if (!n->is_Store() && n->Opcode() != Op_CastP2X &&
752 !(n->is_ArrayCopy() &&
753 n->as_ArrayCopy()->is_clonebasic() &&
754 n->in(ArrayCopyNode::Dest) == use)) {
755 DEBUG_ONLY(disq_node = n;)
756 if (n->is_Load() || n->is_LoadStore()) {
757 NOT_PRODUCT(fail_eliminate = "Field load";)
758 } else {
759 NOT_PRODUCT(fail_eliminate = "Not store field reference";)
760 }
761 can_eliminate = false;
762 }
763 }
764 } else if (use->is_ArrayCopy() &&
765 (use->as_ArrayCopy()->is_arraycopy_validated() ||
766 use->as_ArrayCopy()->is_copyof_validated() ||
767 use->as_ArrayCopy()->is_copyofrange_validated()) &&
768 use->in(ArrayCopyNode::Dest) == res) {
769 // ok to eliminate
770 } else if (use->is_SafePoint()) {
771 SafePointNode* sfpt = use->as_SafePoint();
772 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) {
773 // Object is passed as argument.
774 DEBUG_ONLY(disq_node = use;)
775 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";)
776 can_eliminate = false;
777 }
778 Node* sfptMem = sfpt->memory();
779 if (sfptMem == NULL || sfptMem->is_top()) {
780 DEBUG_ONLY(disq_node = use;)
781 NOT_PRODUCT(fail_eliminate = "NULL or TOP memory";)
782 can_eliminate = false;
783 } else {
784 safepoints.append_if_missing(sfpt);
785 }
786 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
787 if (use->is_Phi()) {
788 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
789 NOT_PRODUCT(fail_eliminate = "Object is return value";)
790 } else {
791 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
792 }
793 DEBUG_ONLY(disq_node = use;)
794 } else {
795 if (use->Opcode() == Op_Return) {
796 NOT_PRODUCT(fail_eliminate = "Object is return value";)
797 } else {
798 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
799 }
800 DEBUG_ONLY(disq_node = use;)
801 }
802 can_eliminate = false;
803 }
804 }
805 }
806
807 #ifndef PRODUCT
808 if (PrintEliminateAllocations) {
809 if (can_eliminate) {
810 tty->print("Scalar ");
811 if (res == NULL)
812 alloc->dump();
813 else
814 res->dump();
815 } else if (alloc->_is_scalar_replaceable) {
816 tty->print("NotScalar (%s)", fail_eliminate);
817 if (res == NULL)
818 alloc->dump();
819 else
820 res->dump();
821 #ifdef ASSERT
822 if (disq_node != NULL) {
823 tty->print(" >>>> ");
824 disq_node->dump();
825 }
826 #endif /*ASSERT*/
827 }
828 }
829 #endif
830 return can_eliminate;
831 }
832
833 // Do scalar replacement.
834 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
835 GrowableArray <SafePointNode *> safepoints_done;
836
837 ciKlass* klass = NULL;
838 ciInstanceKlass* iklass = NULL;
839 int nfields = 0;
840 int array_base = 0;
841 int element_size = 0;
842 BasicType basic_elem_type = T_ILLEGAL;
843 ciType* elem_type = NULL;
844
845 Node* res = alloc->result_cast();
846 assert(res == NULL || res->is_CheckCastPP(), "unexpected AllocateNode result");
847 const TypeOopPtr* res_type = NULL;
848 if (res != NULL) { // Could be NULL when there are no users
849 res_type = _igvn.type(res)->isa_oopptr();
850 }
851
852 if (res != NULL) {
853 klass = res_type->klass();
854 // Value types are only allocated on demand
855 if (res_type->isa_instptr() || res_type->isa_valuetypeptr()) {
856 // find the fields of the class which will be needed for safepoint debug information
857 assert(klass->is_instance_klass(), "must be an instance klass.");
858 iklass = klass->as_instance_klass();
859 nfields = iklass->nof_nonstatic_fields();
860 } else {
861 // find the array's elements which will be needed for safepoint debug information
862 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
863 assert(klass->is_array_klass() && nfields >= 0, "must be an array klass.");
864 elem_type = klass->as_array_klass()->element_type();
865 basic_elem_type = elem_type->basic_type();
866 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
867 element_size = type2aelembytes(basic_elem_type);
868 if (klass->is_value_array_klass()) {
869 // Flattened value type array
870 element_size = klass->as_value_array_klass()->element_byte_size();
871 }
872 }
873 }
874 //
875 // Process the safepoint uses
876 //
877 Unique_Node_List value_worklist;
878 while (safepoints.length() > 0) {
879 SafePointNode* sfpt = safepoints.pop();
880 Node* mem = sfpt->memory();
881 Node* ctl = sfpt->control();
882 assert(sfpt->jvms() != NULL, "missed JVMS");
883 // Fields of scalar objs are referenced only at the end
884 // of regular debuginfo at the last (youngest) JVMS.
885 // Record relative start index.
886 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff());
887 SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type,
888 #ifdef ASSERT
889 alloc,
890 #endif
891 first_ind, nfields);
892 sobj->init_req(0, C->root());
893 transform_later(sobj);
894
895 // Scan object's fields adding an input to the safepoint for each field.
896 for (int j = 0; j < nfields; j++) {
897 intptr_t offset;
898 ciField* field = NULL;
899 if (iklass != NULL) {
900 field = iklass->nonstatic_field_at(j);
901 offset = field->offset();
902 elem_type = field->type();
903 basic_elem_type = field->layout_type();
904 assert(!field->is_flattened(), "flattened value type fields should not have safepoint uses");
905 } else {
906 offset = array_base + j * (intptr_t)element_size;
907 }
908
909 const Type *field_type;
910 // The next code is taken from Parse::do_get_xxx().
911 if (basic_elem_type == T_OBJECT || basic_elem_type == T_ARRAY) {
912 if (!elem_type->is_loaded()) {
913 field_type = TypeInstPtr::BOTTOM;
914 } else if (field != NULL && field->is_static_constant()) {
915 // This can happen if the constant oop is non-perm.
916 ciObject* con = field->constant_value().as_object();
917 // Do not "join" in the previous type; it doesn't add value,
918 // and may yield a vacuous result if the field is of interface type.
919 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr();
920 assert(field_type != NULL, "field singleton type must be consistent");
921 } else {
922 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
923 }
924 if (UseCompressedOops) {
925 field_type = field_type->make_narrowoop();
926 basic_elem_type = T_NARROWOOP;
927 }
928 } else {
929 field_type = Type::get_const_basic_type(basic_elem_type);
930 }
931
932 Node* field_val = NULL;
933 const TypeOopPtr* field_addr_type = res_type->add_offset(offset)->isa_oopptr();
934 if (klass->is_value_array_klass()) {
935 ciValueKlass* vk = elem_type->as_value_klass();
936 assert(vk->flatten_array(), "must be flattened");
937 field_val = value_type_from_mem(mem, ctl, vk, field_addr_type->isa_aryptr(), 0, alloc);
938 } else {
939 field_val = value_from_mem(mem, ctl, basic_elem_type, field_type, field_addr_type, alloc);
940 }
941 if (field_val == NULL) {
942 // We weren't able to find a value for this field,
943 // give up on eliminating this allocation.
944
945 // Remove any extra entries we added to the safepoint.
946 uint last = sfpt->req() - 1;
947 for (int k = 0; k < j; k++) {
948 sfpt->del_req(last--);
949 }
950 _igvn._worklist.push(sfpt);
951 // rollback processed safepoints
952 while (safepoints_done.length() > 0) {
953 SafePointNode* sfpt_done = safepoints_done.pop();
954 // remove any extra entries we added to the safepoint
955 last = sfpt_done->req() - 1;
956 for (int k = 0; k < nfields; k++) {
957 sfpt_done->del_req(last--);
958 }
959 JVMState *jvms = sfpt_done->jvms();
960 jvms->set_endoff(sfpt_done->req());
961 // Now make a pass over the debug information replacing any references
962 // to SafePointScalarObjectNode with the allocated object.
963 int start = jvms->debug_start();
964 int end = jvms->debug_end();
965 for (int i = start; i < end; i++) {
966 if (sfpt_done->in(i)->is_SafePointScalarObject()) {
967 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject();
968 if (scobj->first_index(jvms) == sfpt_done->req() &&
969 scobj->n_fields() == (uint)nfields) {
970 assert(scobj->alloc() == alloc, "sanity");
971 sfpt_done->set_req(i, res);
972 }
973 }
974 }
975 _igvn._worklist.push(sfpt_done);
976 }
977 #ifndef PRODUCT
978 if (PrintEliminateAllocations) {
979 if (field != NULL) {
980 tty->print("=== At SafePoint node %d can't find value of Field: ",
981 sfpt->_idx);
982 field->print();
983 int field_idx = C->get_alias_index(field_addr_type);
984 tty->print(" (alias_idx=%d)", field_idx);
985 } else { // Array's element
986 tty->print("=== At SafePoint node %d can't find value of array element [%d]",
987 sfpt->_idx, j);
988 }
989 tty->print(", which prevents elimination of: ");
990 if (res == NULL)
991 alloc->dump();
992 else
993 res->dump();
994 }
995 #endif
996 return false;
997 }
998 if (UseCompressedOops && field_type->isa_narrowoop()) {
999 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
1000 // to be able scalar replace the allocation.
1001 if (field_val->is_EncodeP()) {
1002 field_val = field_val->in(1);
1003 } else {
1004 field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type()));
1005 }
1006 } else if (field_val->is_ValueType()) {
1007 // Keep track of value types to scalarize them later
1008 value_worklist.push(field_val);
1009 }
1010 sfpt->add_req(field_val);
1011 }
1012 JVMState *jvms = sfpt->jvms();
1013 jvms->set_endoff(sfpt->req());
1014 // Now make a pass over the debug information replacing any references
1015 // to the allocated object with "sobj"
1016 int start = jvms->debug_start();
1017 int end = jvms->debug_end();
1018 sfpt->replace_edges_in_range(res, sobj, start, end);
1019 _igvn._worklist.push(sfpt);
1020 safepoints_done.append_if_missing(sfpt); // keep it for rollback
1021 }
1022 // Scalarize value types that were added to the safepoint
1023 for (uint i = 0; i < value_worklist.size(); ++i) {
1024 Node* vt = value_worklist.at(i);
1025 vt->as_ValueType()->make_scalar_in_safepoints(C->root(), &_igvn);
1026 }
1027 return true;
1028 }
1029
1030 static void disconnect_projections(MultiNode* n, PhaseIterGVN& igvn) {
1031 Node* ctl_proj = n->proj_out_or_null(TypeFunc::Control);
1032 Node* mem_proj = n->proj_out_or_null(TypeFunc::Memory);
1033 if (ctl_proj != NULL) {
1034 igvn.replace_node(ctl_proj, n->in(0));
1035 }
1036 if (mem_proj != NULL) {
1037 igvn.replace_node(mem_proj, n->in(TypeFunc::Memory));
1038 }
1039 }
1040
1041 // Process users of eliminated allocation.
1042 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc) {
1043 Node* res = alloc->result_cast();
1044 if (res != NULL) {
1045 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
1046 Node *use = res->last_out(j);
1047 uint oc1 = res->outcnt();
1048
1049 if (use->is_AddP()) {
1050 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) {
1051 Node *n = use->last_out(k);
1052 uint oc2 = use->outcnt();
1053 if (n->is_Store()) {
1054 #ifdef ASSERT
1055 // Verify that there is no dependent MemBarVolatile nodes,
1056 // they should be removed during IGVN, see MemBarNode::Ideal().
1057 for (DUIterator_Fast pmax, p = n->fast_outs(pmax);
1058 p < pmax; p++) {
1059 Node* mb = n->fast_out(p);
1060 assert(mb->is_Initialize() || !mb->is_MemBar() ||
1061 mb->req() <= MemBarNode::Precedent ||
1062 mb->in(MemBarNode::Precedent) != n,
1063 "MemBarVolatile should be eliminated for non-escaping object");
1064 }
1065 #endif
1066 _igvn.replace_node(n, n->in(MemNode::Memory));
1067 } else if (n->is_ArrayCopy()) {
1068 // Disconnect ArrayCopy node
1069 ArrayCopyNode* ac = n->as_ArrayCopy();
1070 assert(ac->is_clonebasic(), "unexpected array copy kind");
1071 Node* membar_after = ac->proj_out(TypeFunc::Control)->unique_ctrl_out();
1072 disconnect_projections(ac, _igvn);
1073 assert(alloc->in(0)->is_Proj() && alloc->in(0)->in(0)->Opcode() == Op_MemBarCPUOrder, "mem barrier expected before allocation");
1074 Node* membar_before = alloc->in(0)->in(0);
1075 disconnect_projections(membar_before->as_MemBar(), _igvn);
1076 if (membar_after->is_MemBar()) {
1077 disconnect_projections(membar_after->as_MemBar(), _igvn);
1078 }
1079 } else {
1080 eliminate_card_mark(n);
1081 }
1082 k -= (oc2 - use->outcnt());
1083 }
1084 } else if (use->is_ArrayCopy()) {
1085 // Disconnect ArrayCopy node
1086 ArrayCopyNode* ac = use->as_ArrayCopy();
1087 assert(ac->is_arraycopy_validated() ||
1088 ac->is_copyof_validated() ||
1089 ac->is_copyofrange_validated(), "unsupported");
1090 CallProjections* callprojs = ac->extract_projections(true);
1091
1092 _igvn.replace_node(callprojs->fallthrough_ioproj, ac->in(TypeFunc::I_O));
1093 _igvn.replace_node(callprojs->fallthrough_memproj, ac->in(TypeFunc::Memory));
1094 _igvn.replace_node(callprojs->fallthrough_catchproj, ac->in(TypeFunc::Control));
1095
1096 // Set control to top. IGVN will remove the remaining projections
1097 ac->set_req(0, top());
1098 ac->replace_edge(res, top());
1099
1100 // Disconnect src right away: it can help find new
1101 // opportunities for allocation elimination
1102 Node* src = ac->in(ArrayCopyNode::Src);
1103 ac->replace_edge(src, top());
1104 // src can be top at this point if src and dest of the
1105 // arraycopy were the same
1106 if (src->outcnt() == 0 && !src->is_top()) {
1107 _igvn.remove_dead_node(src);
1108 }
1109
1110 _igvn._worklist.push(ac);
1111 } else {
1112 eliminate_card_mark(use);
1113 }
1114 j -= (oc1 - res->outcnt());
1115 }
1116 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
1117 _igvn.remove_dead_node(res);
1118 }
1119
1120 //
1121 // Process other users of allocation's projections
1122 //
1123 if (_resproj != NULL && _resproj->outcnt() != 0) {
1124 // First disconnect stores captured by Initialize node.
1125 // If Initialize node is eliminated first in the following code,
1126 // it will kill such stores and DUIterator_Last will assert.
1127 for (DUIterator_Fast jmax, j = _resproj->fast_outs(jmax); j < jmax; j++) {
1128 Node *use = _resproj->fast_out(j);
1129 if (use->is_AddP()) {
1130 // raw memory addresses used only by the initialization
1131 _igvn.replace_node(use, C->top());
1132 --j; --jmax;
1133 }
1134 }
1135 for (DUIterator_Last jmin, j = _resproj->last_outs(jmin); j >= jmin; ) {
1136 Node *use = _resproj->last_out(j);
1137 uint oc1 = _resproj->outcnt();
1138 if (use->is_Initialize()) {
1139 // Eliminate Initialize node.
1140 InitializeNode *init = use->as_Initialize();
1141 assert(init->outcnt() <= 2, "only a control and memory projection expected");
1142 Node *ctrl_proj = init->proj_out_or_null(TypeFunc::Control);
1143 if (ctrl_proj != NULL) {
1144 assert(init->in(TypeFunc::Control) == _fallthroughcatchproj, "allocation control projection");
1145 _igvn.replace_node(ctrl_proj, _fallthroughcatchproj);
1146 }
1147 Node *mem_proj = init->proj_out_or_null(TypeFunc::Memory);
1148 if (mem_proj != NULL) {
1149 Node *mem = init->in(TypeFunc::Memory);
1150 #ifdef ASSERT
1151 if (mem->is_MergeMem()) {
1152 assert(mem->in(TypeFunc::Memory) == _memproj_fallthrough, "allocation memory projection");
1153 } else {
1154 assert(mem == _memproj_fallthrough, "allocation memory projection");
1155 }
1156 #endif
1157 _igvn.replace_node(mem_proj, mem);
1158 }
1159 } else {
1160 assert(false, "only Initialize or AddP expected");
1161 }
1162 j -= (oc1 - _resproj->outcnt());
1163 }
1164 }
1165 if (_fallthroughcatchproj != NULL) {
1166 _igvn.replace_node(_fallthroughcatchproj, alloc->in(TypeFunc::Control));
1167 }
1168 if (_memproj_fallthrough != NULL) {
1169 _igvn.replace_node(_memproj_fallthrough, alloc->in(TypeFunc::Memory));
1170 }
1171 if (_memproj_catchall != NULL) {
1172 _igvn.replace_node(_memproj_catchall, C->top());
1173 }
1174 if (_ioproj_fallthrough != NULL) {
1175 _igvn.replace_node(_ioproj_fallthrough, alloc->in(TypeFunc::I_O));
1176 }
1177 if (_ioproj_catchall != NULL) {
1178 _igvn.replace_node(_ioproj_catchall, C->top());
1179 }
1180 if (_catchallcatchproj != NULL) {
1181 _igvn.replace_node(_catchallcatchproj, C->top());
1182 }
1183 }
1184
1185 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) {
1186 // Don't do scalar replacement if the frame can be popped by JVMTI:
1187 // if reallocation fails during deoptimization we'll pop all
1188 // interpreter frames for this compiled frame and that won't play
1189 // nice with JVMTI popframe.
1190 if (!EliminateAllocations || JvmtiExport::can_pop_frame() || !alloc->_is_non_escaping) {
1191 return false;
1192 }
1193 Node* klass = alloc->in(AllocateNode::KlassNode);
1194 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr();
1195 Node* res = alloc->result_cast();
1196 // Eliminate boxing allocations which are not used
1197 // regardless scalar replacable status.
1198 bool boxing_alloc = C->eliminate_boxing() &&
1199 tklass->klass()->is_instance_klass() &&
1200 tklass->klass()->as_instance_klass()->is_box_klass();
1201 if (!alloc->_is_scalar_replaceable && (!boxing_alloc || (res != NULL))) {
1202 return false;
1203 }
1204
1205 extract_call_projections(alloc);
1206
1207 GrowableArray <SafePointNode *> safepoints;
1208 if (!can_eliminate_allocation(alloc, safepoints)) {
1209 return false;
1210 }
1211
1212 if (!alloc->_is_scalar_replaceable) {
1213 assert(res == NULL, "sanity");
1214 // We can only eliminate allocation if all debug info references
1215 // are already replaced with SafePointScalarObject because
1216 // we can't search for a fields value without instance_id.
1217 if (safepoints.length() > 0) {
1218 return false;
1219 }
1220 }
1221
1222 if (!scalar_replacement(alloc, safepoints)) {
1223 return false;
1224 }
1225
1226 CompileLog* log = C->log();
1227 if (log != NULL) {
1228 log->head("eliminate_allocation type='%d'",
1229 log->identify(tklass->klass()));
1230 JVMState* p = alloc->jvms();
1231 while (p != NULL) {
1232 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1233 p = p->caller();
1234 }
1235 log->tail("eliminate_allocation");
1236 }
1237
1238 process_users_of_allocation(alloc);
1239
1240 #ifndef PRODUCT
1241 if (PrintEliminateAllocations) {
1242 if (alloc->is_AllocateArray())
1243 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1244 else
1245 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1246 }
1247 #endif
1248
1249 return true;
1250 }
1251
1252 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) {
1253 // EA should remove all uses of non-escaping boxing node.
1254 if (!C->eliminate_boxing() || boxing->proj_out_or_null(TypeFunc::Parms) != NULL) {
1255 return false;
1256 }
1257
1258 assert(boxing->result_cast() == NULL, "unexpected boxing node result");
1259
1260 extract_call_projections(boxing);
1261
1262 const TypeTuple* r = boxing->tf()->range_sig();
1263 assert(r->cnt() > TypeFunc::Parms, "sanity");
1264 const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr();
1265 assert(t != NULL, "sanity");
1266
1267 CompileLog* log = C->log();
1268 if (log != NULL) {
1269 log->head("eliminate_boxing type='%d'",
1270 log->identify(t->klass()));
1271 JVMState* p = boxing->jvms();
1272 while (p != NULL) {
1273 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1274 p = p->caller();
1275 }
1276 log->tail("eliminate_boxing");
1277 }
1278
1279 process_users_of_allocation(boxing);
1280
1281 #ifndef PRODUCT
1282 if (PrintEliminateAllocations) {
1283 tty->print("++++ Eliminated: %d ", boxing->_idx);
1284 boxing->method()->print_short_name(tty);
1285 tty->cr();
1286 }
1287 #endif
1288
1289 return true;
1290 }
1291
1292 //---------------------------set_eden_pointers-------------------------
1293 void PhaseMacroExpand::set_eden_pointers(Node* &eden_top_adr, Node* &eden_end_adr) {
1294 if (UseTLAB) { // Private allocation: load from TLS
1295 Node* thread = transform_later(new ThreadLocalNode());
1296 int tlab_top_offset = in_bytes(JavaThread::tlab_top_offset());
1297 int tlab_end_offset = in_bytes(JavaThread::tlab_end_offset());
1298 eden_top_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_top_offset);
1299 eden_end_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_end_offset);
1300 } else { // Shared allocation: load from globals
1301 CollectedHeap* ch = Universe::heap();
1302 address top_adr = (address)ch->top_addr();
1303 address end_adr = (address)ch->end_addr();
1304 eden_top_adr = makecon(TypeRawPtr::make(top_adr));
1305 eden_end_adr = basic_plus_adr(eden_top_adr, end_adr - top_adr);
1306 }
1307 }
1308
1309
1310 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) {
1311 Node* adr = basic_plus_adr(base, offset);
1312 const TypePtr* adr_type = adr->bottom_type()->is_ptr();
1313 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt, MemNode::unordered);
1314 transform_later(value);
1315 return value;
1316 }
1317
1318
1319 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) {
1320 Node* adr = basic_plus_adr(base, offset);
1321 mem = StoreNode::make(_igvn, ctl, mem, adr, NULL, value, bt, MemNode::unordered);
1322 transform_later(mem);
1323 return mem;
1324 }
1325
1326 //=============================================================================
1327 //
1328 // A L L O C A T I O N
1329 //
1330 // Allocation attempts to be fast in the case of frequent small objects.
1331 // It breaks down like this:
1332 //
1333 // 1) Size in doublewords is computed. This is a constant for objects and
1334 // variable for most arrays. Doubleword units are used to avoid size
1335 // overflow of huge doubleword arrays. We need doublewords in the end for
1336 // rounding.
1337 //
1338 // 2) Size is checked for being 'too large'. Too-large allocations will go
1339 // the slow path into the VM. The slow path can throw any required
1340 // exceptions, and does all the special checks for very large arrays. The
1341 // size test can constant-fold away for objects. For objects with
1342 // finalizers it constant-folds the otherway: you always go slow with
1343 // finalizers.
1344 //
1345 // 3) If NOT using TLABs, this is the contended loop-back point.
1346 // Load-Locked the heap top. If using TLABs normal-load the heap top.
1347 //
1348 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route.
1349 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish
1350 // "size*8" we always enter the VM, where "largish" is a constant picked small
1351 // enough that there's always space between the eden max and 4Gig (old space is
1352 // there so it's quite large) and large enough that the cost of entering the VM
1353 // is dwarfed by the cost to initialize the space.
1354 //
1355 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back
1356 // down. If contended, repeat at step 3. If using TLABs normal-store
1357 // adjusted heap top back down; there is no contention.
1358 //
1359 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark
1360 // fields.
1361 //
1362 // 7) Merge with the slow-path; cast the raw memory pointer to the correct
1363 // oop flavor.
1364 //
1365 //=============================================================================
1366 // FastAllocateSizeLimit value is in DOUBLEWORDS.
1367 // Allocations bigger than this always go the slow route.
1368 // This value must be small enough that allocation attempts that need to
1369 // trigger exceptions go the slow route. Also, it must be small enough so
1370 // that heap_top + size_in_bytes does not wrap around the 4Gig limit.
1371 //=============================================================================j//
1372 // %%% Here is an old comment from parseHelper.cpp; is it outdated?
1373 // The allocator will coalesce int->oop copies away. See comment in
1374 // coalesce.cpp about how this works. It depends critically on the exact
1375 // code shape produced here, so if you are changing this code shape
1376 // make sure the GC info for the heap-top is correct in and around the
1377 // slow-path call.
1378 //
1379
1380 void PhaseMacroExpand::expand_allocate_common(
1381 AllocateNode* alloc, // allocation node to be expanded
1382 Node* length, // array length for an array allocation
1383 const TypeFunc* slow_call_type, // Type of slow call
1384 address slow_call_address // Address of slow call
1385 )
1386 {
1387
1388 Node* ctrl = alloc->in(TypeFunc::Control);
1389 Node* mem = alloc->in(TypeFunc::Memory);
1390 Node* i_o = alloc->in(TypeFunc::I_O);
1391 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize);
1392 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1393 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest);
1394
1395 assert(ctrl != NULL, "must have control");
1396 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results.
1397 // they will not be used if "always_slow" is set
1398 enum { slow_result_path = 1, fast_result_path = 2 };
1399 Node *result_region = NULL;
1400 Node *result_phi_rawmem = NULL;
1401 Node *result_phi_rawoop = NULL;
1402 Node *result_phi_i_o = NULL;
1403
1404 // The initial slow comparison is a size check, the comparison
1405 // we want to do is a BoolTest::gt
1406 bool always_slow = false;
1407 int tv = _igvn.find_int_con(initial_slow_test, -1);
1408 if (tv >= 0) {
1409 always_slow = (tv == 1);
1410 initial_slow_test = NULL;
1411 } else {
1412 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn);
1413 }
1414
1415 if (C->env()->dtrace_alloc_probes() ||
1416 (!UseTLAB && !Universe::heap()->supports_inline_contig_alloc())) {
1417 // Force slow-path allocation
1418 always_slow = true;
1419 initial_slow_test = NULL;
1420 }
1421
1422
1423 enum { too_big_or_final_path = 1, need_gc_path = 2 };
1424 Node *slow_region = NULL;
1425 Node *toobig_false = ctrl;
1426
1427 assert (initial_slow_test == NULL || !always_slow, "arguments must be consistent");
1428 // generate the initial test if necessary
1429 if (initial_slow_test != NULL ) {
1430 slow_region = new RegionNode(3);
1431
1432 // Now make the initial failure test. Usually a too-big test but
1433 // might be a TRUE for finalizers or a fancy class check for
1434 // newInstance0.
1435 IfNode *toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
1436 transform_later(toobig_iff);
1437 // Plug the failing-too-big test into the slow-path region
1438 Node *toobig_true = new IfTrueNode( toobig_iff );
1439 transform_later(toobig_true);
1440 slow_region ->init_req( too_big_or_final_path, toobig_true );
1441 toobig_false = new IfFalseNode( toobig_iff );
1442 transform_later(toobig_false);
1443 } else { // No initial test, just fall into next case
1444 toobig_false = ctrl;
1445 debug_only(slow_region = NodeSentinel);
1446 }
1447
1448 Node *slow_mem = mem; // save the current memory state for slow path
1449 // generate the fast allocation code unless we know that the initial test will always go slow
1450 if (!always_slow) {
1451 // Fast path modifies only raw memory.
1452 if (mem->is_MergeMem()) {
1453 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1454 }
1455
1456 Node* eden_top_adr;
1457 Node* eden_end_adr;
1458
1459 set_eden_pointers(eden_top_adr, eden_end_adr);
1460
1461 // Load Eden::end. Loop invariant and hoisted.
1462 //
1463 // Note: We set the control input on "eden_end" and "old_eden_top" when using
1464 // a TLAB to work around a bug where these values were being moved across
1465 // a safepoint. These are not oops, so they cannot be include in the oop
1466 // map, but they can be changed by a GC. The proper way to fix this would
1467 // be to set the raw memory state when generating a SafepointNode. However
1468 // this will require extensive changes to the loop optimization in order to
1469 // prevent a degradation of the optimization.
1470 // See comment in memnode.hpp, around line 227 in class LoadPNode.
1471 Node *eden_end = make_load(ctrl, mem, eden_end_adr, 0, TypeRawPtr::BOTTOM, T_ADDRESS);
1472
1473 // allocate the Region and Phi nodes for the result
1474 result_region = new RegionNode(3);
1475 result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1476 result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM);
1477 result_phi_i_o = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch
1478
1479 // We need a Region for the loop-back contended case.
1480 enum { fall_in_path = 1, contended_loopback_path = 2 };
1481 Node *contended_region;
1482 Node *contended_phi_rawmem;
1483 if (UseTLAB) {
1484 contended_region = toobig_false;
1485 contended_phi_rawmem = mem;
1486 } else {
1487 contended_region = new RegionNode(3);
1488 contended_phi_rawmem = new PhiNode(contended_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1489 // Now handle the passing-too-big test. We fall into the contended
1490 // loop-back merge point.
1491 contended_region ->init_req(fall_in_path, toobig_false);
1492 contended_phi_rawmem->init_req(fall_in_path, mem);
1493 transform_later(contended_region);
1494 transform_later(contended_phi_rawmem);
1495 }
1496
1497 // Load(-locked) the heap top.
1498 // See note above concerning the control input when using a TLAB
1499 Node *old_eden_top = UseTLAB
1500 ? new LoadPNode (ctrl, contended_phi_rawmem, eden_top_adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, MemNode::unordered)
1501 : new LoadPLockedNode(contended_region, contended_phi_rawmem, eden_top_adr, MemNode::acquire);
1502
1503 transform_later(old_eden_top);
1504 // Add to heap top to get a new heap top
1505 Node *new_eden_top = new AddPNode(top(), old_eden_top, size_in_bytes);
1506 transform_later(new_eden_top);
1507 // Check for needing a GC; compare against heap end
1508 Node *needgc_cmp = new CmpPNode(new_eden_top, eden_end);
1509 transform_later(needgc_cmp);
1510 Node *needgc_bol = new BoolNode(needgc_cmp, BoolTest::ge);
1511 transform_later(needgc_bol);
1512 IfNode *needgc_iff = new IfNode(contended_region, needgc_bol, PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN);
1513 transform_later(needgc_iff);
1514
1515 // Plug the failing-heap-space-need-gc test into the slow-path region
1516 Node *needgc_true = new IfTrueNode(needgc_iff);
1517 transform_later(needgc_true);
1518 if (initial_slow_test) {
1519 slow_region->init_req(need_gc_path, needgc_true);
1520 // This completes all paths into the slow merge point
1521 transform_later(slow_region);
1522 } else { // No initial slow path needed!
1523 // Just fall from the need-GC path straight into the VM call.
1524 slow_region = needgc_true;
1525 }
1526 // No need for a GC. Setup for the Store-Conditional
1527 Node *needgc_false = new IfFalseNode(needgc_iff);
1528 transform_later(needgc_false);
1529
1530 // Grab regular I/O before optional prefetch may change it.
1531 // Slow-path does no I/O so just set it to the original I/O.
1532 result_phi_i_o->init_req(slow_result_path, i_o);
1533
1534 i_o = prefetch_allocation(i_o, needgc_false, contended_phi_rawmem,
1535 old_eden_top, new_eden_top, length);
1536
1537 // Name successful fast-path variables
1538 Node* fast_oop = old_eden_top;
1539 Node* fast_oop_ctrl;
1540 Node* fast_oop_rawmem;
1541
1542 // Store (-conditional) the modified eden top back down.
1543 // StorePConditional produces flags for a test PLUS a modified raw
1544 // memory state.
1545 if (UseTLAB) {
1546 Node* store_eden_top =
1547 new StorePNode(needgc_false, contended_phi_rawmem, eden_top_adr,
1548 TypeRawPtr::BOTTOM, new_eden_top, MemNode::unordered);
1549 transform_later(store_eden_top);
1550 fast_oop_ctrl = needgc_false; // No contention, so this is the fast path
1551 fast_oop_rawmem = store_eden_top;
1552 } else {
1553 Node* store_eden_top =
1554 new StorePConditionalNode(needgc_false, contended_phi_rawmem, eden_top_adr,
1555 new_eden_top, fast_oop/*old_eden_top*/);
1556 transform_later(store_eden_top);
1557 Node *contention_check = new BoolNode(store_eden_top, BoolTest::ne);
1558 transform_later(contention_check);
1559 store_eden_top = new SCMemProjNode(store_eden_top);
1560 transform_later(store_eden_top);
1561
1562 // If not using TLABs, check to see if there was contention.
1563 IfNode *contention_iff = new IfNode (needgc_false, contention_check, PROB_MIN, COUNT_UNKNOWN);
1564 transform_later(contention_iff);
1565 Node *contention_true = new IfTrueNode(contention_iff);
1566 transform_later(contention_true);
1567 // If contention, loopback and try again.
1568 contended_region->init_req(contended_loopback_path, contention_true);
1569 contended_phi_rawmem->init_req(contended_loopback_path, store_eden_top);
1570
1571 // Fast-path succeeded with no contention!
1572 Node *contention_false = new IfFalseNode(contention_iff);
1573 transform_later(contention_false);
1574 fast_oop_ctrl = contention_false;
1575
1576 // Bump total allocated bytes for this thread
1577 Node* thread = new ThreadLocalNode();
1578 transform_later(thread);
1579 Node* alloc_bytes_adr = basic_plus_adr(top()/*not oop*/, thread,
1580 in_bytes(JavaThread::allocated_bytes_offset()));
1581 Node* alloc_bytes = make_load(fast_oop_ctrl, store_eden_top, alloc_bytes_adr,
1582 0, TypeLong::LONG, T_LONG);
1583 #ifdef _LP64
1584 Node* alloc_size = size_in_bytes;
1585 #else
1586 Node* alloc_size = new ConvI2LNode(size_in_bytes);
1587 transform_later(alloc_size);
1588 #endif
1589 Node* new_alloc_bytes = new AddLNode(alloc_bytes, alloc_size);
1590 transform_later(new_alloc_bytes);
1591 fast_oop_rawmem = make_store(fast_oop_ctrl, store_eden_top, alloc_bytes_adr,
1592 0, new_alloc_bytes, T_LONG);
1593 }
1594
1595 InitializeNode* init = alloc->initialization();
1596 fast_oop_rawmem = initialize_object(alloc,
1597 fast_oop_ctrl, fast_oop_rawmem, fast_oop,
1598 klass_node, length, size_in_bytes);
1599
1600 // If initialization is performed by an array copy, any required
1601 // MemBarStoreStore was already added. If the object does not
1602 // escape no need for a MemBarStoreStore. If the object does not
1603 // escape in its initializer and memory barrier (MemBarStoreStore or
1604 // stronger) is already added at exit of initializer, also no need
1605 // for a MemBarStoreStore. Otherwise we need a MemBarStoreStore
1606 // so that stores that initialize this object can't be reordered
1607 // with a subsequent store that makes this object accessible by
1608 // other threads.
1609 // Other threads include java threads and JVM internal threads
1610 // (for example concurrent GC threads). Current concurrent GC
1611 // implementation: CMS and G1 will not scan newly created object,
1612 // so it's safe to skip storestore barrier when allocation does
1613 // not escape.
1614 if (!alloc->does_not_escape_thread() &&
1615 !alloc->is_allocation_MemBar_redundant() &&
1616 (init == NULL || !init->is_complete_with_arraycopy())) {
1617 if (init == NULL || init->req() < InitializeNode::RawStores) {
1618 // No InitializeNode or no stores captured by zeroing
1619 // elimination. Simply add the MemBarStoreStore after object
1620 // initialization.
1621 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1622 transform_later(mb);
1623
1624 mb->init_req(TypeFunc::Memory, fast_oop_rawmem);
1625 mb->init_req(TypeFunc::Control, fast_oop_ctrl);
1626 fast_oop_ctrl = new ProjNode(mb,TypeFunc::Control);
1627 transform_later(fast_oop_ctrl);
1628 fast_oop_rawmem = new ProjNode(mb,TypeFunc::Memory);
1629 transform_later(fast_oop_rawmem);
1630 } else {
1631 // Add the MemBarStoreStore after the InitializeNode so that
1632 // all stores performing the initialization that were moved
1633 // before the InitializeNode happen before the storestore
1634 // barrier.
1635
1636 Node* init_ctrl = init->proj_out_or_null(TypeFunc::Control);
1637 Node* init_mem = init->proj_out_or_null(TypeFunc::Memory);
1638
1639 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1640 transform_later(mb);
1641
1642 Node* ctrl = new ProjNode(init,TypeFunc::Control);
1643 transform_later(ctrl);
1644 Node* mem = new ProjNode(init,TypeFunc::Memory);
1645 transform_later(mem);
1646
1647 // The MemBarStoreStore depends on control and memory coming
1648 // from the InitializeNode
1649 mb->init_req(TypeFunc::Memory, mem);
1650 mb->init_req(TypeFunc::Control, ctrl);
1651
1652 ctrl = new ProjNode(mb,TypeFunc::Control);
1653 transform_later(ctrl);
1654 mem = new ProjNode(mb,TypeFunc::Memory);
1655 transform_later(mem);
1656
1657 // All nodes that depended on the InitializeNode for control
1658 // and memory must now depend on the MemBarNode that itself
1659 // depends on the InitializeNode
1660 if (init_ctrl != NULL) {
1661 _igvn.replace_node(init_ctrl, ctrl);
1662 }
1663 if (init_mem != NULL) {
1664 _igvn.replace_node(init_mem, mem);
1665 }
1666 }
1667 }
1668
1669 if (C->env()->dtrace_extended_probes()) {
1670 // Slow-path call
1671 int size = TypeFunc::Parms + 2;
1672 CallLeafNode *call = new CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(),
1673 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_object_alloc_base),
1674 "dtrace_object_alloc",
1675 TypeRawPtr::BOTTOM);
1676
1677 // Get base of thread-local storage area
1678 Node* thread = new ThreadLocalNode();
1679 transform_later(thread);
1680
1681 call->init_req(TypeFunc::Parms+0, thread);
1682 call->init_req(TypeFunc::Parms+1, fast_oop);
1683 call->init_req(TypeFunc::Control, fast_oop_ctrl);
1684 call->init_req(TypeFunc::I_O , top()); // does no i/o
1685 call->init_req(TypeFunc::Memory , fast_oop_rawmem);
1686 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1687 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1688 transform_later(call);
1689 fast_oop_ctrl = new ProjNode(call,TypeFunc::Control);
1690 transform_later(fast_oop_ctrl);
1691 fast_oop_rawmem = new ProjNode(call,TypeFunc::Memory);
1692 transform_later(fast_oop_rawmem);
1693 }
1694
1695 // Plug in the successful fast-path into the result merge point
1696 result_region ->init_req(fast_result_path, fast_oop_ctrl);
1697 result_phi_rawoop->init_req(fast_result_path, fast_oop);
1698 result_phi_i_o ->init_req(fast_result_path, i_o);
1699 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem);
1700 } else {
1701 slow_region = ctrl;
1702 result_phi_i_o = i_o; // Rename it to use in the following code.
1703 }
1704
1705 // Generate slow-path call
1706 CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address,
1707 OptoRuntime::stub_name(slow_call_address),
1708 alloc->jvms()->bci(),
1709 TypePtr::BOTTOM);
1710 call->init_req( TypeFunc::Control, slow_region );
1711 call->init_req( TypeFunc::I_O , top() ) ; // does no i/o
1712 call->init_req( TypeFunc::Memory , slow_mem ); // may gc ptrs
1713 call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) );
1714 call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) );
1715
1716 call->init_req(TypeFunc::Parms+0, klass_node);
1717 if (length != NULL) {
1718 call->init_req(TypeFunc::Parms+1, length);
1719 }
1720
1721 // Copy debug information and adjust JVMState information, then replace
1722 // allocate node with the call
1723 copy_call_debug_info((CallNode *) alloc, call);
1724 if (!always_slow) {
1725 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1726 } else {
1727 // Hook i_o projection to avoid its elimination during allocation
1728 // replacement (when only a slow call is generated).
1729 call->set_req(TypeFunc::I_O, result_phi_i_o);
1730 }
1731 _igvn.replace_node(alloc, call);
1732 transform_later(call);
1733
1734 // Identify the output projections from the allocate node and
1735 // adjust any references to them.
1736 // The control and io projections look like:
1737 //
1738 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl)
1739 // Allocate Catch
1740 // ^---Proj(io) <-------+ ^---CatchProj(io)
1741 //
1742 // We are interested in the CatchProj nodes.
1743 //
1744 extract_call_projections(call);
1745
1746 // An allocate node has separate memory projections for the uses on
1747 // the control and i_o paths. Replace the control memory projection with
1748 // result_phi_rawmem (unless we are only generating a slow call when
1749 // both memory projections are combined)
1750 if (!always_slow && _memproj_fallthrough != NULL) {
1751 for (DUIterator_Fast imax, i = _memproj_fallthrough->fast_outs(imax); i < imax; i++) {
1752 Node *use = _memproj_fallthrough->fast_out(i);
1753 _igvn.rehash_node_delayed(use);
1754 imax -= replace_input(use, _memproj_fallthrough, result_phi_rawmem);
1755 // back up iterator
1756 --i;
1757 }
1758 }
1759 // Now change uses of _memproj_catchall to use _memproj_fallthrough and delete
1760 // _memproj_catchall so we end up with a call that has only 1 memory projection.
1761 if (_memproj_catchall != NULL ) {
1762 if (_memproj_fallthrough == NULL) {
1763 _memproj_fallthrough = new ProjNode(call, TypeFunc::Memory);
1764 transform_later(_memproj_fallthrough);
1765 }
1766 for (DUIterator_Fast imax, i = _memproj_catchall->fast_outs(imax); i < imax; i++) {
1767 Node *use = _memproj_catchall->fast_out(i);
1768 _igvn.rehash_node_delayed(use);
1769 imax -= replace_input(use, _memproj_catchall, _memproj_fallthrough);
1770 // back up iterator
1771 --i;
1772 }
1773 assert(_memproj_catchall->outcnt() == 0, "all uses must be deleted");
1774 _igvn.remove_dead_node(_memproj_catchall);
1775 }
1776
1777 // An allocate node has separate i_o projections for the uses on the control
1778 // and i_o paths. Always replace the control i_o projection with result i_o
1779 // otherwise incoming i_o become dead when only a slow call is generated
1780 // (it is different from memory projections where both projections are
1781 // combined in such case).
1782 if (_ioproj_fallthrough != NULL) {
1783 for (DUIterator_Fast imax, i = _ioproj_fallthrough->fast_outs(imax); i < imax; i++) {
1784 Node *use = _ioproj_fallthrough->fast_out(i);
1785 _igvn.rehash_node_delayed(use);
1786 imax -= replace_input(use, _ioproj_fallthrough, result_phi_i_o);
1787 // back up iterator
1788 --i;
1789 }
1790 }
1791 // Now change uses of _ioproj_catchall to use _ioproj_fallthrough and delete
1792 // _ioproj_catchall so we end up with a call that has only 1 i_o projection.
1793 if (_ioproj_catchall != NULL ) {
1794 if (_ioproj_fallthrough == NULL) {
1795 _ioproj_fallthrough = new ProjNode(call, TypeFunc::I_O);
1796 transform_later(_ioproj_fallthrough);
1797 }
1798 for (DUIterator_Fast imax, i = _ioproj_catchall->fast_outs(imax); i < imax; i++) {
1799 Node *use = _ioproj_catchall->fast_out(i);
1800 _igvn.rehash_node_delayed(use);
1801 imax -= replace_input(use, _ioproj_catchall, _ioproj_fallthrough);
1802 // back up iterator
1803 --i;
1804 }
1805 assert(_ioproj_catchall->outcnt() == 0, "all uses must be deleted");
1806 _igvn.remove_dead_node(_ioproj_catchall);
1807 }
1808
1809 // if we generated only a slow call, we are done
1810 if (always_slow) {
1811 // Now we can unhook i_o.
1812 if (result_phi_i_o->outcnt() > 1) {
1813 call->set_req(TypeFunc::I_O, top());
1814 } else {
1815 assert(result_phi_i_o->unique_ctrl_out() == call, "");
1816 // Case of new array with negative size known during compilation.
1817 // AllocateArrayNode::Ideal() optimization disconnect unreachable
1818 // following code since call to runtime will throw exception.
1819 // As result there will be no users of i_o after the call.
1820 // Leave i_o attached to this call to avoid problems in preceding graph.
1821 }
1822 return;
1823 }
1824
1825
1826 if (_fallthroughcatchproj != NULL) {
1827 ctrl = _fallthroughcatchproj->clone();
1828 transform_later(ctrl);
1829 _igvn.replace_node(_fallthroughcatchproj, result_region);
1830 } else {
1831 ctrl = top();
1832 }
1833 Node *slow_result;
1834 if (_resproj == NULL) {
1835 // no uses of the allocation result
1836 slow_result = top();
1837 } else {
1838 slow_result = _resproj->clone();
1839 transform_later(slow_result);
1840 _igvn.replace_node(_resproj, result_phi_rawoop);
1841 }
1842
1843 // Plug slow-path into result merge point
1844 result_region ->init_req( slow_result_path, ctrl );
1845 result_phi_rawoop->init_req( slow_result_path, slow_result);
1846 result_phi_rawmem->init_req( slow_result_path, _memproj_fallthrough );
1847 transform_later(result_region);
1848 transform_later(result_phi_rawoop);
1849 transform_later(result_phi_rawmem);
1850 transform_later(result_phi_i_o);
1851 // This completes all paths into the result merge point
1852 }
1853
1854
1855 // Helper for PhaseMacroExpand::expand_allocate_common.
1856 // Initializes the newly-allocated storage.
1857 Node*
1858 PhaseMacroExpand::initialize_object(AllocateNode* alloc,
1859 Node* control, Node* rawmem, Node* object,
1860 Node* klass_node, Node* length,
1861 Node* size_in_bytes) {
1862 InitializeNode* init = alloc->initialization();
1863 // Store the klass & mark bits
1864 Node* mark_node = NULL;
1865 // For now only enable fast locking for non-array types
1866 if (UseBiasedLocking && (length == NULL)) {
1867 mark_node = make_load(control, rawmem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
1868 } else {
1869 mark_node = makecon(TypeRawPtr::make((address)markOopDesc::prototype()));
1870 }
1871 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS);
1872
1873 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
1874 int header_size = alloc->minimum_header_size(); // conservatively small
1875
1876 // Array length
1877 if (length != NULL) { // Arrays need length field
1878 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
1879 // conservatively small header size:
1880 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1881 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
1882 if (k->is_array_klass()) // we know the exact header size in most cases:
1883 header_size = Klass::layout_helper_header_size(k->layout_helper());
1884 }
1885
1886 // Clear the object body, if necessary.
1887 if (init == NULL) {
1888 // The init has somehow disappeared; be cautious and clear everything.
1889 //
1890 // This can happen if a node is allocated but an uncommon trap occurs
1891 // immediately. In this case, the Initialize gets associated with the
1892 // trap, and may be placed in a different (outer) loop, if the Allocate
1893 // is in a loop. If (this is rare) the inner loop gets unrolled, then
1894 // there can be two Allocates to one Initialize. The answer in all these
1895 // edge cases is safety first. It is always safe to clear immediately
1896 // within an Allocate, and then (maybe or maybe not) clear some more later.
1897 if (!(UseTLAB && ZeroTLAB)) {
1898 rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
1899 header_size, size_in_bytes,
1900 &_igvn);
1901 }
1902 } else {
1903 if (!init->is_complete()) {
1904 // Try to win by zeroing only what the init does not store.
1905 // We can also try to do some peephole optimizations,
1906 // such as combining some adjacent subword stores.
1907 rawmem = init->complete_stores(control, rawmem, object,
1908 header_size, size_in_bytes, &_igvn);
1909 }
1910 // We have no more use for this link, since the AllocateNode goes away:
1911 init->set_req(InitializeNode::RawAddress, top());
1912 // (If we keep the link, it just confuses the register allocator,
1913 // who thinks he sees a real use of the address by the membar.)
1914 }
1915
1916 return rawmem;
1917 }
1918
1919 // Generate prefetch instructions for next allocations.
1920 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false,
1921 Node*& contended_phi_rawmem,
1922 Node* old_eden_top, Node* new_eden_top,
1923 Node* length) {
1924 enum { fall_in_path = 1, pf_path = 2 };
1925 if( UseTLAB && AllocatePrefetchStyle == 2 ) {
1926 // Generate prefetch allocation with watermark check.
1927 // As an allocation hits the watermark, we will prefetch starting
1928 // at a "distance" away from watermark.
1929
1930 Node *pf_region = new RegionNode(3);
1931 Node *pf_phi_rawmem = new PhiNode( pf_region, Type::MEMORY,
1932 TypeRawPtr::BOTTOM );
1933 // I/O is used for Prefetch
1934 Node *pf_phi_abio = new PhiNode( pf_region, Type::ABIO );
1935
1936 Node *thread = new ThreadLocalNode();
1937 transform_later(thread);
1938
1939 Node *eden_pf_adr = new AddPNode( top()/*not oop*/, thread,
1940 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) );
1941 transform_later(eden_pf_adr);
1942
1943 Node *old_pf_wm = new LoadPNode(needgc_false,
1944 contended_phi_rawmem, eden_pf_adr,
1945 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM,
1946 MemNode::unordered);
1947 transform_later(old_pf_wm);
1948
1949 // check against new_eden_top
1950 Node *need_pf_cmp = new CmpPNode( new_eden_top, old_pf_wm );
1951 transform_later(need_pf_cmp);
1952 Node *need_pf_bol = new BoolNode( need_pf_cmp, BoolTest::ge );
1953 transform_later(need_pf_bol);
1954 IfNode *need_pf_iff = new IfNode( needgc_false, need_pf_bol,
1955 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN );
1956 transform_later(need_pf_iff);
1957
1958 // true node, add prefetchdistance
1959 Node *need_pf_true = new IfTrueNode( need_pf_iff );
1960 transform_later(need_pf_true);
1961
1962 Node *need_pf_false = new IfFalseNode( need_pf_iff );
1963 transform_later(need_pf_false);
1964
1965 Node *new_pf_wmt = new AddPNode( top(), old_pf_wm,
1966 _igvn.MakeConX(AllocatePrefetchDistance) );
1967 transform_later(new_pf_wmt );
1968 new_pf_wmt->set_req(0, need_pf_true);
1969
1970 Node *store_new_wmt = new StorePNode(need_pf_true,
1971 contended_phi_rawmem, eden_pf_adr,
1972 TypeRawPtr::BOTTOM, new_pf_wmt,
1973 MemNode::unordered);
1974 transform_later(store_new_wmt);
1975
1976 // adding prefetches
1977 pf_phi_abio->init_req( fall_in_path, i_o );
1978
1979 Node *prefetch_adr;
1980 Node *prefetch;
1981 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1982 uint step_size = AllocatePrefetchStepSize;
1983 uint distance = 0;
1984
1985 for ( uint i = 0; i < lines; i++ ) {
1986 prefetch_adr = new AddPNode( old_pf_wm, new_pf_wmt,
1987 _igvn.MakeConX(distance) );
1988 transform_later(prefetch_adr);
1989 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
1990 transform_later(prefetch);
1991 distance += step_size;
1992 i_o = prefetch;
1993 }
1994 pf_phi_abio->set_req( pf_path, i_o );
1995
1996 pf_region->init_req( fall_in_path, need_pf_false );
1997 pf_region->init_req( pf_path, need_pf_true );
1998
1999 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem );
2000 pf_phi_rawmem->init_req( pf_path, store_new_wmt );
2001
2002 transform_later(pf_region);
2003 transform_later(pf_phi_rawmem);
2004 transform_later(pf_phi_abio);
2005
2006 needgc_false = pf_region;
2007 contended_phi_rawmem = pf_phi_rawmem;
2008 i_o = pf_phi_abio;
2009 } else if( UseTLAB && AllocatePrefetchStyle == 3 ) {
2010 // Insert a prefetch instruction for each allocation.
2011 // This code is used to generate 1 prefetch instruction per cache line.
2012
2013 // Generate several prefetch instructions.
2014 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
2015 uint step_size = AllocatePrefetchStepSize;
2016 uint distance = AllocatePrefetchDistance;
2017
2018 // Next cache address.
2019 Node *cache_adr = new AddPNode(old_eden_top, old_eden_top,
2020 _igvn.MakeConX(step_size + distance));
2021 transform_later(cache_adr);
2022 cache_adr = new CastP2XNode(needgc_false, cache_adr);
2023 transform_later(cache_adr);
2024 // Address is aligned to execute prefetch to the beginning of cache line size
2025 // (it is important when BIS instruction is used on SPARC as prefetch).
2026 Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1));
2027 cache_adr = new AndXNode(cache_adr, mask);
2028 transform_later(cache_adr);
2029 cache_adr = new CastX2PNode(cache_adr);
2030 transform_later(cache_adr);
2031
2032 // Prefetch
2033 Node *prefetch = new PrefetchAllocationNode( contended_phi_rawmem, cache_adr );
2034 prefetch->set_req(0, needgc_false);
2035 transform_later(prefetch);
2036 contended_phi_rawmem = prefetch;
2037 Node *prefetch_adr;
2038 distance = step_size;
2039 for ( uint i = 1; i < lines; i++ ) {
2040 prefetch_adr = new AddPNode( cache_adr, cache_adr,
2041 _igvn.MakeConX(distance) );
2042 transform_later(prefetch_adr);
2043 prefetch = new PrefetchAllocationNode( contended_phi_rawmem, prefetch_adr );
2044 transform_later(prefetch);
2045 distance += step_size;
2046 contended_phi_rawmem = prefetch;
2047 }
2048 } else if( AllocatePrefetchStyle > 0 ) {
2049 // Insert a prefetch for each allocation only on the fast-path
2050 Node *prefetch_adr;
2051 Node *prefetch;
2052 // Generate several prefetch instructions.
2053 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
2054 uint step_size = AllocatePrefetchStepSize;
2055 uint distance = AllocatePrefetchDistance;
2056 for ( uint i = 0; i < lines; i++ ) {
2057 prefetch_adr = new AddPNode( old_eden_top, new_eden_top,
2058 _igvn.MakeConX(distance) );
2059 transform_later(prefetch_adr);
2060 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
2061 // Do not let it float too high, since if eden_top == eden_end,
2062 // both might be null.
2063 if( i == 0 ) { // Set control for first prefetch, next follows it
2064 prefetch->init_req(0, needgc_false);
2065 }
2066 transform_later(prefetch);
2067 distance += step_size;
2068 i_o = prefetch;
2069 }
2070 }
2071 return i_o;
2072 }
2073
2074
2075 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) {
2076 expand_allocate_common(alloc, NULL,
2077 OptoRuntime::new_instance_Type(),
2078 OptoRuntime::new_instance_Java());
2079 }
2080
2081 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) {
2082 Node* length = alloc->in(AllocateNode::ALength);
2083 InitializeNode* init = alloc->initialization();
2084 Node* klass_node = alloc->in(AllocateNode::KlassNode);
2085 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
2086 address slow_call_address; // Address of slow call
2087 if (init != NULL && init->is_complete_with_arraycopy() &&
2088 k->is_type_array_klass()) {
2089 // Don't zero type array during slow allocation in VM since
2090 // it will be initialized later by arraycopy in compiled code.
2091 slow_call_address = OptoRuntime::new_array_nozero_Java();
2092 } else {
2093 slow_call_address = OptoRuntime::new_array_Java();
2094 }
2095 expand_allocate_common(alloc, length,
2096 OptoRuntime::new_array_Type(),
2097 slow_call_address);
2098 }
2099
2100 //-------------------mark_eliminated_box----------------------------------
2101 //
2102 // During EA obj may point to several objects but after few ideal graph
2103 // transformations (CCP) it may point to only one non escaping object
2104 // (but still using phi), corresponding locks and unlocks will be marked
2105 // for elimination. Later obj could be replaced with a new node (new phi)
2106 // and which does not have escape information. And later after some graph
2107 // reshape other locks and unlocks (which were not marked for elimination
2108 // before) are connected to this new obj (phi) but they still will not be
2109 // marked for elimination since new obj has no escape information.
2110 // Mark all associated (same box and obj) lock and unlock nodes for
2111 // elimination if some of them marked already.
2112 void PhaseMacroExpand::mark_eliminated_box(Node* oldbox, Node* obj) {
2113 if (oldbox->as_BoxLock()->is_eliminated())
2114 return; // This BoxLock node was processed already.
2115
2116 // New implementation (EliminateNestedLocks) has separate BoxLock
2117 // node for each locked region so mark all associated locks/unlocks as
2118 // eliminated even if different objects are referenced in one locked region
2119 // (for example, OSR compilation of nested loop inside locked scope).
2120 if (EliminateNestedLocks ||
2121 oldbox->as_BoxLock()->is_simple_lock_region(NULL, obj)) {
2122 // Box is used only in one lock region. Mark this box as eliminated.
2123 _igvn.hash_delete(oldbox);
2124 oldbox->as_BoxLock()->set_eliminated(); // This changes box's hash value
2125 _igvn.hash_insert(oldbox);
2126
2127 for (uint i = 0; i < oldbox->outcnt(); i++) {
2128 Node* u = oldbox->raw_out(i);
2129 if (u->is_AbstractLock() && !u->as_AbstractLock()->is_non_esc_obj()) {
2130 AbstractLockNode* alock = u->as_AbstractLock();
2131 // Check lock's box since box could be referenced by Lock's debug info.
2132 if (alock->box_node() == oldbox) {
2133 // Mark eliminated all related locks and unlocks.
2134 #ifdef ASSERT
2135 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc4");
2136 #endif
2137 alock->set_non_esc_obj();
2138 }
2139 }
2140 }
2141 return;
2142 }
2143
2144 // Create new "eliminated" BoxLock node and use it in monitor debug info
2145 // instead of oldbox for the same object.
2146 BoxLockNode* newbox = oldbox->clone()->as_BoxLock();
2147
2148 // Note: BoxLock node is marked eliminated only here and it is used
2149 // to indicate that all associated lock and unlock nodes are marked
2150 // for elimination.
2151 newbox->set_eliminated();
2152 transform_later(newbox);
2153
2154 // Replace old box node with new box for all users of the same object.
2155 for (uint i = 0; i < oldbox->outcnt();) {
2156 bool next_edge = true;
2157
2158 Node* u = oldbox->raw_out(i);
2159 if (u->is_AbstractLock()) {
2160 AbstractLockNode* alock = u->as_AbstractLock();
2161 if (alock->box_node() == oldbox && alock->obj_node()->eqv_uncast(obj)) {
2162 // Replace Box and mark eliminated all related locks and unlocks.
2163 #ifdef ASSERT
2164 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc5");
2165 #endif
2166 alock->set_non_esc_obj();
2167 _igvn.rehash_node_delayed(alock);
2168 alock->set_box_node(newbox);
2169 next_edge = false;
2170 }
2171 }
2172 if (u->is_FastLock() && u->as_FastLock()->obj_node()->eqv_uncast(obj)) {
2173 FastLockNode* flock = u->as_FastLock();
2174 assert(flock->box_node() == oldbox, "sanity");
2175 _igvn.rehash_node_delayed(flock);
2176 flock->set_box_node(newbox);
2177 next_edge = false;
2178 }
2179
2180 // Replace old box in monitor debug info.
2181 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) {
2182 SafePointNode* sfn = u->as_SafePoint();
2183 JVMState* youngest_jvms = sfn->jvms();
2184 int max_depth = youngest_jvms->depth();
2185 for (int depth = 1; depth <= max_depth; depth++) {
2186 JVMState* jvms = youngest_jvms->of_depth(depth);
2187 int num_mon = jvms->nof_monitors();
2188 // Loop over monitors
2189 for (int idx = 0; idx < num_mon; idx++) {
2190 Node* obj_node = sfn->monitor_obj(jvms, idx);
2191 Node* box_node = sfn->monitor_box(jvms, idx);
2192 if (box_node == oldbox && obj_node->eqv_uncast(obj)) {
2193 int j = jvms->monitor_box_offset(idx);
2194 _igvn.replace_input_of(u, j, newbox);
2195 next_edge = false;
2196 }
2197 }
2198 }
2199 }
2200 if (next_edge) i++;
2201 }
2202 }
2203
2204 //-----------------------mark_eliminated_locking_nodes-----------------------
2205 void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) {
2206 if (EliminateNestedLocks) {
2207 if (alock->is_nested()) {
2208 assert(alock->box_node()->as_BoxLock()->is_eliminated(), "sanity");
2209 return;
2210 } else if (!alock->is_non_esc_obj()) { // Not eliminated or coarsened
2211 // Only Lock node has JVMState needed here.
2212 // Not that preceding claim is documented anywhere else.
2213 if (alock->jvms() != NULL) {
2214 if (alock->as_Lock()->is_nested_lock_region()) {
2215 // Mark eliminated related nested locks and unlocks.
2216 Node* obj = alock->obj_node();
2217 BoxLockNode* box_node = alock->box_node()->as_BoxLock();
2218 assert(!box_node->is_eliminated(), "should not be marked yet");
2219 // Note: BoxLock node is marked eliminated only here
2220 // and it is used to indicate that all associated lock
2221 // and unlock nodes are marked for elimination.
2222 box_node->set_eliminated(); // Box's hash is always NO_HASH here
2223 for (uint i = 0; i < box_node->outcnt(); i++) {
2224 Node* u = box_node->raw_out(i);
2225 if (u->is_AbstractLock()) {
2226 alock = u->as_AbstractLock();
2227 if (alock->box_node() == box_node) {
2228 // Verify that this Box is referenced only by related locks.
2229 assert(alock->obj_node()->eqv_uncast(obj), "");
2230 // Mark all related locks and unlocks.
2231 #ifdef ASSERT
2232 alock->log_lock_optimization(C, "eliminate_lock_set_nested");
2233 #endif
2234 alock->set_nested();
2235 }
2236 }
2237 }
2238 } else {
2239 #ifdef ASSERT
2240 alock->log_lock_optimization(C, "eliminate_lock_NOT_nested_lock_region");
2241 if (C->log() != NULL)
2242 alock->as_Lock()->is_nested_lock_region(C); // rerun for debugging output
2243 #endif
2244 }
2245 }
2246 return;
2247 }
2248 // Process locks for non escaping object
2249 assert(alock->is_non_esc_obj(), "");
2250 } // EliminateNestedLocks
2251
2252 if (alock->is_non_esc_obj()) { // Lock is used for non escaping object
2253 // Look for all locks of this object and mark them and
2254 // corresponding BoxLock nodes as eliminated.
2255 Node* obj = alock->obj_node();
2256 for (uint j = 0; j < obj->outcnt(); j++) {
2257 Node* o = obj->raw_out(j);
2258 if (o->is_AbstractLock() &&
2259 o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) {
2260 alock = o->as_AbstractLock();
2261 Node* box = alock->box_node();
2262 // Replace old box node with new eliminated box for all users
2263 // of the same object and mark related locks as eliminated.
2264 mark_eliminated_box(box, obj);
2265 }
2266 }
2267 }
2268 }
2269
2270 // we have determined that this lock/unlock can be eliminated, we simply
2271 // eliminate the node without expanding it.
2272 //
2273 // Note: The membar's associated with the lock/unlock are currently not
2274 // eliminated. This should be investigated as a future enhancement.
2275 //
2276 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) {
2277
2278 if (!alock->is_eliminated()) {
2279 return false;
2280 }
2281 #ifdef ASSERT
2282 if (!alock->is_coarsened()) {
2283 // Check that new "eliminated" BoxLock node is created.
2284 BoxLockNode* oldbox = alock->box_node()->as_BoxLock();
2285 assert(oldbox->is_eliminated(), "should be done already");
2286 }
2287 #endif
2288
2289 alock->log_lock_optimization(C, "eliminate_lock");
2290
2291 #ifndef PRODUCT
2292 if (PrintEliminateLocks) {
2293 if (alock->is_Lock()) {
2294 tty->print_cr("++++ Eliminated: %d Lock", alock->_idx);
2295 } else {
2296 tty->print_cr("++++ Eliminated: %d Unlock", alock->_idx);
2297 }
2298 }
2299 #endif
2300
2301 Node* mem = alock->in(TypeFunc::Memory);
2302 Node* ctrl = alock->in(TypeFunc::Control);
2303
2304 extract_call_projections(alock);
2305 // There are 2 projections from the lock. The lock node will
2306 // be deleted when its last use is subsumed below.
2307 assert(alock->outcnt() == 2 &&
2308 _fallthroughproj != NULL &&
2309 _memproj_fallthrough != NULL,
2310 "Unexpected projections from Lock/Unlock");
2311
2312 Node* fallthroughproj = _fallthroughproj;
2313 Node* memproj_fallthrough = _memproj_fallthrough;
2314
2315 // The memory projection from a lock/unlock is RawMem
2316 // The input to a Lock is merged memory, so extract its RawMem input
2317 // (unless the MergeMem has been optimized away.)
2318 if (alock->is_Lock()) {
2319 // Seach for MemBarAcquireLock node and delete it also.
2320 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
2321 assert(membar != NULL && membar->Opcode() == Op_MemBarAcquireLock, "");
2322 Node* ctrlproj = membar->proj_out(TypeFunc::Control);
2323 Node* memproj = membar->proj_out(TypeFunc::Memory);
2324 _igvn.replace_node(ctrlproj, fallthroughproj);
2325 _igvn.replace_node(memproj, memproj_fallthrough);
2326
2327 // Delete FastLock node also if this Lock node is unique user
2328 // (a loop peeling may clone a Lock node).
2329 Node* flock = alock->as_Lock()->fastlock_node();
2330 if (flock->outcnt() == 1) {
2331 assert(flock->unique_out() == alock, "sanity");
2332 _igvn.replace_node(flock, top());
2333 }
2334 }
2335
2336 // Seach for MemBarReleaseLock node and delete it also.
2337 if (alock->is_Unlock() && ctrl != NULL && ctrl->is_Proj() &&
2338 ctrl->in(0)->is_MemBar()) {
2339 MemBarNode* membar = ctrl->in(0)->as_MemBar();
2340 assert(membar->Opcode() == Op_MemBarReleaseLock &&
2341 mem->is_Proj() && membar == mem->in(0), "");
2342 _igvn.replace_node(fallthroughproj, ctrl);
2343 _igvn.replace_node(memproj_fallthrough, mem);
2344 fallthroughproj = ctrl;
2345 memproj_fallthrough = mem;
2346 ctrl = membar->in(TypeFunc::Control);
2347 mem = membar->in(TypeFunc::Memory);
2348 }
2349
2350 _igvn.replace_node(fallthroughproj, ctrl);
2351 _igvn.replace_node(memproj_fallthrough, mem);
2352 return true;
2353 }
2354
2355
2356 //------------------------------expand_lock_node----------------------
2357 void PhaseMacroExpand::expand_lock_node(LockNode *lock) {
2358
2359 Node* ctrl = lock->in(TypeFunc::Control);
2360 Node* mem = lock->in(TypeFunc::Memory);
2361 Node* obj = lock->obj_node();
2362 Node* box = lock->box_node();
2363 Node* flock = lock->fastlock_node();
2364
2365 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2366
2367 // Make the merge point
2368 Node *region;
2369 Node *mem_phi;
2370 Node *slow_path;
2371
2372 if (UseOptoBiasInlining) {
2373 /*
2374 * See the full description in MacroAssembler::biased_locking_enter().
2375 *
2376 * if( (mark_word & biased_lock_mask) == biased_lock_pattern ) {
2377 * // The object is biased.
2378 * proto_node = klass->prototype_header;
2379 * o_node = thread | proto_node;
2380 * x_node = o_node ^ mark_word;
2381 * if( (x_node & ~age_mask) == 0 ) { // Biased to the current thread ?
2382 * // Done.
2383 * } else {
2384 * if( (x_node & biased_lock_mask) != 0 ) {
2385 * // The klass's prototype header is no longer biased.
2386 * cas(&mark_word, mark_word, proto_node)
2387 * goto cas_lock;
2388 * } else {
2389 * // The klass's prototype header is still biased.
2390 * if( (x_node & epoch_mask) != 0 ) { // Expired epoch?
2391 * old = mark_word;
2392 * new = o_node;
2393 * } else {
2394 * // Different thread or anonymous biased.
2395 * old = mark_word & (epoch_mask | age_mask | biased_lock_mask);
2396 * new = thread | old;
2397 * }
2398 * // Try to rebias.
2399 * if( cas(&mark_word, old, new) == 0 ) {
2400 * // Done.
2401 * } else {
2402 * goto slow_path; // Failed.
2403 * }
2404 * }
2405 * }
2406 * } else {
2407 * // The object is not biased.
2408 * cas_lock:
2409 * if( FastLock(obj) == 0 ) {
2410 * // Done.
2411 * } else {
2412 * slow_path:
2413 * OptoRuntime::complete_monitor_locking_Java(obj);
2414 * }
2415 * }
2416 */
2417
2418 region = new RegionNode(5);
2419 // create a Phi for the memory state
2420 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2421
2422 Node* fast_lock_region = new RegionNode(3);
2423 Node* fast_lock_mem_phi = new PhiNode( fast_lock_region, Type::MEMORY, TypeRawPtr::BOTTOM);
2424
2425 // First, check mark word for the biased lock pattern.
2426 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2427
2428 // Get fast path - mark word has the biased lock pattern.
2429 ctrl = opt_bits_test(ctrl, fast_lock_region, 1, mark_node,
2430 markOopDesc::biased_lock_mask_in_place,
2431 markOopDesc::biased_lock_pattern, true);
2432 // fast_lock_region->in(1) is set to slow path.
2433 fast_lock_mem_phi->init_req(1, mem);
2434
2435 // Now check that the lock is biased to the current thread and has
2436 // the same epoch and bias as Klass::_prototype_header.
2437
2438 // Special-case a fresh allocation to avoid building nodes:
2439 Node* klass_node = AllocateNode::Ideal_klass(obj, &_igvn);
2440 if (klass_node == NULL) {
2441 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
2442 klass_node = transform_later(LoadKlassNode::make(_igvn, NULL, mem, k_adr, _igvn.type(k_adr)->is_ptr()));
2443 #ifdef _LP64
2444 if (UseCompressedClassPointers && klass_node->is_DecodeNKlass()) {
2445 assert(klass_node->in(1)->Opcode() == Op_LoadNKlass, "sanity");
2446 klass_node->in(1)->init_req(0, ctrl);
2447 } else
2448 #endif
2449 klass_node->init_req(0, ctrl);
2450 }
2451 Node *proto_node = make_load(ctrl, mem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeX_X, TypeX_X->basic_type());
2452
2453 Node* thread = transform_later(new ThreadLocalNode());
2454 Node* cast_thread = transform_later(new CastP2XNode(ctrl, thread));
2455 Node* o_node = transform_later(new OrXNode(cast_thread, proto_node));
2456 Node* x_node = transform_later(new XorXNode(o_node, mark_node));
2457
2458 // Get slow path - mark word does NOT match the value.
2459 Node* not_biased_ctrl = opt_bits_test(ctrl, region, 3, x_node,
2460 (~markOopDesc::age_mask_in_place), 0);
2461 // region->in(3) is set to fast path - the object is biased to the current thread.
2462 mem_phi->init_req(3, mem);
2463
2464
2465 // Mark word does NOT match the value (thread | Klass::_prototype_header).
2466
2467
2468 // First, check biased pattern.
2469 // Get fast path - _prototype_header has the same biased lock pattern.
2470 ctrl = opt_bits_test(not_biased_ctrl, fast_lock_region, 2, x_node,
2471 markOopDesc::biased_lock_mask_in_place, 0, true);
2472
2473 not_biased_ctrl = fast_lock_region->in(2); // Slow path
2474 // fast_lock_region->in(2) - the prototype header is no longer biased
2475 // and we have to revoke the bias on this object.
2476 // We are going to try to reset the mark of this object to the prototype
2477 // value and fall through to the CAS-based locking scheme.
2478 Node* adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
2479 Node* cas = new StoreXConditionalNode(not_biased_ctrl, mem, adr,
2480 proto_node, mark_node);
2481 transform_later(cas);
2482 Node* proj = transform_later(new SCMemProjNode(cas));
2483 fast_lock_mem_phi->init_req(2, proj);
2484
2485
2486 // Second, check epoch bits.
2487 Node* rebiased_region = new RegionNode(3);
2488 Node* old_phi = new PhiNode( rebiased_region, TypeX_X);
2489 Node* new_phi = new PhiNode( rebiased_region, TypeX_X);
2490
2491 // Get slow path - mark word does NOT match epoch bits.
2492 Node* epoch_ctrl = opt_bits_test(ctrl, rebiased_region, 1, x_node,
2493 markOopDesc::epoch_mask_in_place, 0);
2494 // The epoch of the current bias is not valid, attempt to rebias the object
2495 // toward the current thread.
2496 rebiased_region->init_req(2, epoch_ctrl);
2497 old_phi->init_req(2, mark_node);
2498 new_phi->init_req(2, o_node);
2499
2500 // rebiased_region->in(1) is set to fast path.
2501 // The epoch of the current bias is still valid but we know
2502 // nothing about the owner; it might be set or it might be clear.
2503 Node* cmask = MakeConX(markOopDesc::biased_lock_mask_in_place |
2504 markOopDesc::age_mask_in_place |
2505 markOopDesc::epoch_mask_in_place);
2506 Node* old = transform_later(new AndXNode(mark_node, cmask));
2507 cast_thread = transform_later(new CastP2XNode(ctrl, thread));
2508 Node* new_mark = transform_later(new OrXNode(cast_thread, old));
2509 old_phi->init_req(1, old);
2510 new_phi->init_req(1, new_mark);
2511
2512 transform_later(rebiased_region);
2513 transform_later(old_phi);
2514 transform_later(new_phi);
2515
2516 // Try to acquire the bias of the object using an atomic operation.
2517 // If this fails we will go in to the runtime to revoke the object's bias.
2518 cas = new StoreXConditionalNode(rebiased_region, mem, adr, new_phi, old_phi);
2519 transform_later(cas);
2520 proj = transform_later(new SCMemProjNode(cas));
2521
2522 // Get slow path - Failed to CAS.
2523 not_biased_ctrl = opt_bits_test(rebiased_region, region, 4, cas, 0, 0);
2524 mem_phi->init_req(4, proj);
2525 // region->in(4) is set to fast path - the object is rebiased to the current thread.
2526
2527 // Failed to CAS.
2528 slow_path = new RegionNode(3);
2529 Node *slow_mem = new PhiNode( slow_path, Type::MEMORY, TypeRawPtr::BOTTOM);
2530
2531 slow_path->init_req(1, not_biased_ctrl); // Capture slow-control
2532 slow_mem->init_req(1, proj);
2533
2534 // Call CAS-based locking scheme (FastLock node).
2535
2536 transform_later(fast_lock_region);
2537 transform_later(fast_lock_mem_phi);
2538
2539 // Get slow path - FastLock failed to lock the object.
2540 ctrl = opt_bits_test(fast_lock_region, region, 2, flock, 0, 0);
2541 mem_phi->init_req(2, fast_lock_mem_phi);
2542 // region->in(2) is set to fast path - the object is locked to the current thread.
2543
2544 slow_path->init_req(2, ctrl); // Capture slow-control
2545 slow_mem->init_req(2, fast_lock_mem_phi);
2546
2547 transform_later(slow_path);
2548 transform_later(slow_mem);
2549 // Reset lock's memory edge.
2550 lock->set_req(TypeFunc::Memory, slow_mem);
2551
2552 } else {
2553 region = new RegionNode(3);
2554 // create a Phi for the memory state
2555 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2556
2557 // Optimize test; set region slot 2
2558 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0);
2559 mem_phi->init_req(2, mem);
2560 }
2561
2562 // Make slow path call
2563 CallNode *call = make_slow_call((CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(),
2564 OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path,
2565 obj, box, NULL);
2566
2567 extract_call_projections(call);
2568
2569 // Slow path can only throw asynchronous exceptions, which are always
2570 // de-opted. So the compiler thinks the slow-call can never throw an
2571 // exception. If it DOES throw an exception we would need the debug
2572 // info removed first (since if it throws there is no monitor).
2573 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
2574 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
2575
2576 // Capture slow path
2577 // disconnect fall-through projection from call and create a new one
2578 // hook up users of fall-through projection to region
2579 Node *slow_ctrl = _fallthroughproj->clone();
2580 transform_later(slow_ctrl);
2581 _igvn.hash_delete(_fallthroughproj);
2582 _fallthroughproj->disconnect_inputs(NULL, C);
2583 region->init_req(1, slow_ctrl);
2584 // region inputs are now complete
2585 transform_later(region);
2586 _igvn.replace_node(_fallthroughproj, region);
2587
2588 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory));
2589 mem_phi->init_req(1, memproj );
2590 transform_later(mem_phi);
2591 _igvn.replace_node(_memproj_fallthrough, mem_phi);
2592 }
2593
2594 //------------------------------expand_unlock_node----------------------
2595 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
2596
2597 Node* ctrl = unlock->in(TypeFunc::Control);
2598 Node* mem = unlock->in(TypeFunc::Memory);
2599 Node* obj = unlock->obj_node();
2600 Node* box = unlock->box_node();
2601
2602 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2603
2604 // No need for a null check on unlock
2605
2606 // Make the merge point
2607 Node *region;
2608 Node *mem_phi;
2609
2610 if (UseOptoBiasInlining) {
2611 // Check for biased locking unlock case, which is a no-op.
2612 // See the full description in MacroAssembler::biased_locking_exit().
2613 region = new RegionNode(4);
2614 // create a Phi for the memory state
2615 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2616 mem_phi->init_req(3, mem);
2617
2618 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2619 ctrl = opt_bits_test(ctrl, region, 3, mark_node,
2620 markOopDesc::biased_lock_mask_in_place,
2621 markOopDesc::biased_lock_pattern);
2622 } else {
2623 region = new RegionNode(3);
2624 // create a Phi for the memory state
2625 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2626 }
2627
2628 FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box );
2629 funlock = transform_later( funlock )->as_FastUnlock();
2630 // Optimize test; set region slot 2
2631 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0);
2632 Node *thread = transform_later(new ThreadLocalNode());
2633
2634 CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(),
2635 CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C),
2636 "complete_monitor_unlocking_C", slow_path, obj, box, thread);
2637
2638 extract_call_projections(call);
2639
2640 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
2641 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
2642
2643 // No exceptions for unlocking
2644 // Capture slow path
2645 // disconnect fall-through projection from call and create a new one
2646 // hook up users of fall-through projection to region
2647 Node *slow_ctrl = _fallthroughproj->clone();
2648 transform_later(slow_ctrl);
2649 _igvn.hash_delete(_fallthroughproj);
2650 _fallthroughproj->disconnect_inputs(NULL, C);
2651 region->init_req(1, slow_ctrl);
2652 // region inputs are now complete
2653 transform_later(region);
2654 _igvn.replace_node(_fallthroughproj, region);
2655
2656 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory) );
2657 mem_phi->init_req(1, memproj );
2658 mem_phi->init_req(2, mem);
2659 transform_later(mem_phi);
2660 _igvn.replace_node(_memproj_fallthrough, mem_phi);
2661 }
2662
2663 // A value type is returned from the call but we don't know its
2664 // type. Either we get a buffered value (and nothing needs to be done)
2665 // or one of the values being returned is the klass of the value type
2666 // and we need to allocate a value type instance of that type and
2667 // initialize it with other values being returned. In that case, we
2668 // first try a fast path allocation and initialize the value with the
2669 // value klass's pack handler or we fall back to a runtime call.
2670 void PhaseMacroExpand::expand_mh_intrinsic_return(CallStaticJavaNode* call) {
2671 Node* ret = call->proj_out(TypeFunc::Parms);
2672 if (ret == NULL) {
2673 return;
2674 }
2675 assert(ret->bottom_type()->is_valuetypeptr()->is__Value(), "unexpected return type from MH intrinsic");
2676 const TypeFunc* tf = call->_tf;
2677 const TypeTuple* domain = OptoRuntime::store_value_type_fields_Type()->domain_cc();
2678 const TypeFunc* new_tf = TypeFunc::make(tf->domain_sig(), tf->domain_cc(), tf->range_sig(), domain);
2679 call->_tf = new_tf;
2680 // Make sure the change of type is applied before projections are
2681 // processed by igvn
2682 _igvn.set_type(call, call->Value(&_igvn));
2683 _igvn.set_type(ret, ret->Value(&_igvn));
2684
2685 // Before any new projection is added:
2686 CallProjections* projs = call->extract_projections(true, true);
2687
2688 Node* ctl = new Node(1);
2689 Node* mem = new Node(1);
2690 Node* io = new Node(1);
2691 Node* ex_ctl = new Node(1);
2692 Node* ex_mem = new Node(1);
2693 Node* ex_io = new Node(1);
2694 Node* res = new Node(1);
2695
2696 Node* cast = transform_later(new CastP2XNode(ctl, res));
2697 Node* mask = MakeConX(0x1);
2698 Node* masked = transform_later(new AndXNode(cast, mask));
2699 Node* cmp = transform_later(new CmpXNode(masked, mask));
2700 Node* bol = transform_later(new BoolNode(cmp, BoolTest::eq));
2701 IfNode* allocation_iff = new IfNode(ctl, bol, PROB_MAX, COUNT_UNKNOWN);
2702 transform_later(allocation_iff);
2703 Node* allocation_ctl = transform_later(new IfTrueNode(allocation_iff));
2704 Node* no_allocation_ctl = transform_later(new IfFalseNode(allocation_iff));
2705
2706 Node* no_allocation_res = transform_later(new CheckCastPPNode(no_allocation_ctl, res, TypeValueTypePtr::NOTNULL));
2707
2708 Node* mask2 = MakeConX(-2);
2709 Node* masked2 = transform_later(new AndXNode(cast, mask2));
2710 Node* rawklassptr = transform_later(new CastX2PNode(masked2));
2711 Node* klass_node = transform_later(new CheckCastPPNode(allocation_ctl, rawklassptr, TypeKlassPtr::OBJECT_OR_NULL));
2712
2713 Node* slowpath_bol = NULL;
2714 Node* top_adr = NULL;
2715 Node* old_top = NULL;
2716 Node* new_top = NULL;
2717 if (UseTLAB) {
2718 Node* end_adr = NULL;
2719 set_eden_pointers(top_adr, end_adr);
2720 Node* end = make_load(ctl, mem, end_adr, 0, TypeRawPtr::BOTTOM, T_ADDRESS);
2721 old_top = new LoadPNode(ctl, mem, top_adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, MemNode::unordered);
2722 transform_later(old_top);
2723 Node* layout_val = make_load(NULL, mem, klass_node, in_bytes(Klass::layout_helper_offset()), TypeInt::INT, T_INT);
2724 Node* size_in_bytes = ConvI2X(layout_val);
2725 new_top = new AddPNode(top(), old_top, size_in_bytes);
2726 transform_later(new_top);
2727 Node* slowpath_cmp = new CmpPNode(new_top, end);
2728 transform_later(slowpath_cmp);
2729 slowpath_bol = new BoolNode(slowpath_cmp, BoolTest::ge);
2730 transform_later(slowpath_bol);
2731 } else {
2732 slowpath_bol = intcon(1);
2733 top_adr = top();
2734 old_top = top();
2735 new_top = top();
2736 }
2737 IfNode* slowpath_iff = new IfNode(allocation_ctl, slowpath_bol, PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN);
2738 transform_later(slowpath_iff);
2739
2740 Node* slowpath_true = new IfTrueNode(slowpath_iff);
2741 transform_later(slowpath_true);
2742
2743 CallStaticJavaNode* slow_call = new CallStaticJavaNode(OptoRuntime::store_value_type_fields_Type(),
2744 StubRoutines::store_value_type_fields_to_buf(),
2745 "store_value_type_fields",
2746 call->jvms()->bci(),
2747 TypePtr::BOTTOM);
2748 slow_call->init_req(TypeFunc::Control, slowpath_true);
2749 slow_call->init_req(TypeFunc::Memory, mem);
2750 slow_call->init_req(TypeFunc::I_O, io);
2751 slow_call->init_req(TypeFunc::FramePtr, call->in(TypeFunc::FramePtr));
2752 slow_call->init_req(TypeFunc::ReturnAdr, call->in(TypeFunc::ReturnAdr));
2753 slow_call->init_req(TypeFunc::Parms, res);
2754
2755 Node* slow_ctl = transform_later(new ProjNode(slow_call, TypeFunc::Control));
2756 Node* slow_mem = transform_later(new ProjNode(slow_call, TypeFunc::Memory));
2757 Node* slow_io = transform_later(new ProjNode(slow_call, TypeFunc::I_O));
2758 Node* slow_res = transform_later(new ProjNode(slow_call, TypeFunc::Parms));
2759 Node* slow_catc = transform_later(new CatchNode(slow_ctl, slow_io, 2));
2760 Node* slow_norm = transform_later(new CatchProjNode(slow_catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci));
2761 Node* slow_excp = transform_later(new CatchProjNode(slow_catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci));
2762
2763 Node* ex_r = new RegionNode(3);
2764 Node* ex_mem_phi = new PhiNode(ex_r, Type::MEMORY, TypePtr::BOTTOM);
2765 Node* ex_io_phi = new PhiNode(ex_r, Type::ABIO);
2766 ex_r->init_req(1, slow_excp);
2767 ex_mem_phi->init_req(1, slow_mem);
2768 ex_io_phi->init_req(1, slow_io);
2769 ex_r->init_req(2, ex_ctl);
2770 ex_mem_phi->init_req(2, ex_mem);
2771 ex_io_phi->init_req(2, ex_io);
2772
2773 transform_later(ex_r);
2774 transform_later(ex_mem_phi);
2775 transform_later(ex_io_phi);
2776
2777 Node* slowpath_false = new IfFalseNode(slowpath_iff);
2778 transform_later(slowpath_false);
2779 Node* rawmem = new StorePNode(slowpath_false, mem, top_adr, TypeRawPtr::BOTTOM, new_top, MemNode::unordered);
2780 transform_later(rawmem);
2781 Node* mark_node = NULL;
2782 // For now only enable fast locking for non-array types
2783 if (UseBiasedLocking) {
2784 mark_node = make_load(slowpath_false, rawmem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
2785 } else {
2786 mark_node = makecon(TypeRawPtr::make((address)markOopDesc::prototype()));
2787 }
2788 rawmem = make_store(slowpath_false, rawmem, old_top, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS);
2789 rawmem = make_store(slowpath_false, rawmem, old_top, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
2790 if (UseCompressedClassPointers) {
2791 rawmem = make_store(slowpath_false, rawmem, old_top, oopDesc::klass_gap_offset_in_bytes(), intcon(0), T_INT);
2792 }
2793 Node* pack_handler = make_load(slowpath_false, rawmem, klass_node, in_bytes(ValueKlass::pack_handler_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
2794
2795 CallLeafNoFPNode* handler_call = new CallLeafNoFPNode(OptoRuntime::pack_value_type_Type(),
2796 NULL,
2797 "pack handler",
2798 TypeRawPtr::BOTTOM);
2799 handler_call->init_req(TypeFunc::Control, slowpath_false);
2800 handler_call->init_req(TypeFunc::Memory, rawmem);
2801 handler_call->init_req(TypeFunc::I_O, top());
2802 handler_call->init_req(TypeFunc::FramePtr, call->in(TypeFunc::FramePtr));
2803 handler_call->init_req(TypeFunc::ReturnAdr, top());
2804 handler_call->init_req(TypeFunc::Parms, pack_handler);
2805 handler_call->init_req(TypeFunc::Parms+1, old_top);
2806
2807 // We don't know how many values are returned. This assumes the
2808 // worst case, that all available registers are used.
2809 for (uint i = TypeFunc::Parms+1; i < domain->cnt(); i++) {
2810 if (domain->field_at(i) == Type::HALF) {
2811 slow_call->init_req(i, top());
2812 handler_call->init_req(i+1, top());
2813 continue;
2814 }
2815 Node* proj = transform_later(new ProjNode(call, i));
2816 slow_call->init_req(i, proj);
2817 handler_call->init_req(i+1, proj);
2818 }
2819
2820 // We can safepoint at that new call
2821 copy_call_debug_info(call, slow_call);
2822 transform_later(slow_call);
2823 transform_later(handler_call);
2824
2825 Node* handler_ctl = transform_later(new ProjNode(handler_call, TypeFunc::Control));
2826 rawmem = transform_later(new ProjNode(handler_call, TypeFunc::Memory));
2827 Node* slowpath_false_res = transform_later(new ProjNode(handler_call, TypeFunc::Parms));
2828
2829 MergeMemNode* slowpath_false_mem = MergeMemNode::make(mem);
2830 slowpath_false_mem->set_memory_at(Compile::AliasIdxRaw, rawmem);
2831 transform_later(slowpath_false_mem);
2832
2833 Node* r = new RegionNode(4);
2834 Node* mem_phi = new PhiNode(r, Type::MEMORY, TypePtr::BOTTOM);
2835 Node* io_phi = new PhiNode(r, Type::ABIO);
2836 Node* res_phi = new PhiNode(r, ret->bottom_type());
2837
2838 r->init_req(1, no_allocation_ctl);
2839 mem_phi->init_req(1, mem);
2840 io_phi->init_req(1, io);
2841 res_phi->init_req(1, no_allocation_res);
2842 r->init_req(2, slow_norm);
2843 mem_phi->init_req(2, slow_mem);
2844 io_phi->init_req(2, slow_io);
2845 res_phi->init_req(2, slow_res);
2846 r->init_req(3, handler_ctl);
2847 mem_phi->init_req(3, slowpath_false_mem);
2848 io_phi->init_req(3, io);
2849 res_phi->init_req(3, slowpath_false_res);
2850
2851 transform_later(r);
2852 transform_later(mem_phi);
2853 transform_later(io_phi);
2854 transform_later(res_phi);
2855
2856 assert(projs->nb_resproj == 1, "unexpected number of results");
2857 _igvn.replace_in_uses(projs->fallthrough_catchproj, r);
2858 _igvn.replace_in_uses(projs->fallthrough_memproj, mem_phi);
2859 _igvn.replace_in_uses(projs->fallthrough_ioproj, io_phi);
2860 _igvn.replace_in_uses(projs->resproj[0], res_phi);
2861 _igvn.replace_in_uses(projs->catchall_catchproj, ex_r);
2862 _igvn.replace_in_uses(projs->catchall_memproj, ex_mem_phi);
2863 _igvn.replace_in_uses(projs->catchall_ioproj, ex_io_phi);
2864
2865 _igvn.replace_node(ctl, projs->fallthrough_catchproj);
2866 _igvn.replace_node(mem, projs->fallthrough_memproj);
2867 _igvn.replace_node(io, projs->fallthrough_ioproj);
2868 _igvn.replace_node(res, projs->resproj[0]);
2869 _igvn.replace_node(ex_ctl, projs->catchall_catchproj);
2870 _igvn.replace_node(ex_mem, projs->catchall_memproj);
2871 _igvn.replace_node(ex_io, projs->catchall_ioproj);
2872 }
2873
2874 //---------------------------eliminate_macro_nodes----------------------
2875 // Eliminate scalar replaced allocations and associated locks.
2876 void PhaseMacroExpand::eliminate_macro_nodes() {
2877 if (C->macro_count() == 0)
2878 return;
2879
2880 // First, attempt to eliminate locks
2881 int cnt = C->macro_count();
2882 for (int i=0; i < cnt; i++) {
2883 Node *n = C->macro_node(i);
2884 if (n->is_AbstractLock()) { // Lock and Unlock nodes
2885 // Before elimination mark all associated (same box and obj)
2886 // lock and unlock nodes.
2887 mark_eliminated_locking_nodes(n->as_AbstractLock());
2888 }
2889 }
2890 bool progress = true;
2891 while (progress) {
2892 progress = false;
2893 for (int i = C->macro_count(); i > 0; i--) {
2894 Node * n = C->macro_node(i-1);
2895 bool success = false;
2896 debug_only(int old_macro_count = C->macro_count(););
2897 if (n->is_AbstractLock()) {
2898 success = eliminate_locking_node(n->as_AbstractLock());
2899 }
2900 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2901 progress = progress || success;
2902 }
2903 }
2904 // Next, attempt to eliminate allocations
2905 _has_locks = false;
2906 progress = true;
2907 while (progress) {
2908 progress = false;
2909 for (int i = C->macro_count(); i > 0; i--) {
2910 Node * n = C->macro_node(i-1);
2911 bool success = false;
2912 debug_only(int old_macro_count = C->macro_count(););
2913 switch (n->class_id()) {
2914 case Node::Class_Allocate:
2915 case Node::Class_AllocateArray:
2916 success = eliminate_allocate_node(n->as_Allocate());
2917 break;
2918 case Node::Class_CallStaticJava: {
2919 CallStaticJavaNode* call = n->as_CallStaticJava();
2920 if (!call->method()->is_method_handle_intrinsic()) {
2921 success = eliminate_boxing_node(n->as_CallStaticJava());
2922 }
2923 break;
2924 }
2925 case Node::Class_Lock:
2926 case Node::Class_Unlock:
2927 assert(!n->as_AbstractLock()->is_eliminated(), "sanity");
2928 _has_locks = true;
2929 break;
2930 case Node::Class_ArrayCopy:
2931 break;
2932 case Node::Class_OuterStripMinedLoop:
2933 break;
2934 default:
2935 assert(n->Opcode() == Op_LoopLimit ||
2936 n->Opcode() == Op_Opaque1 ||
2937 n->Opcode() == Op_Opaque2 ||
2938 n->Opcode() == Op_Opaque3 ||
2939 n->Opcode() == Op_Opaque4, "unknown node type in macro list");
2940 }
2941 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2942 progress = progress || success;
2943 }
2944 }
2945 }
2946
2947 //------------------------------expand_macro_nodes----------------------
2948 // Returns true if a failure occurred.
2949 bool PhaseMacroExpand::expand_macro_nodes() {
2950 // Last attempt to eliminate macro nodes.
2951 eliminate_macro_nodes();
2952
2953 // Make sure expansion will not cause node limit to be exceeded.
2954 // Worst case is a macro node gets expanded into about 200 nodes.
2955 // Allow 50% more for optimization.
2956 if (C->check_node_count(C->macro_count() * 300, "out of nodes before macro expansion" ) )
2957 return true;
2958
2959 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations.
2960 bool progress = true;
2961 while (progress) {
2962 progress = false;
2963 for (int i = C->macro_count(); i > 0; i--) {
2964 Node * n = C->macro_node(i-1);
2965 bool success = false;
2966 debug_only(int old_macro_count = C->macro_count(););
2967 if (n->Opcode() == Op_LoopLimit) {
2968 // Remove it from macro list and put on IGVN worklist to optimize.
2969 C->remove_macro_node(n);
2970 _igvn._worklist.push(n);
2971 success = true;
2972 } else if (n->Opcode() == Op_CallStaticJava) {
2973 CallStaticJavaNode* call = n->as_CallStaticJava();
2974 if (!call->method()->is_method_handle_intrinsic()) {
2975 // Remove it from macro list and put on IGVN worklist to optimize.
2976 C->remove_macro_node(n);
2977 _igvn._worklist.push(n);
2978 success = true;
2979 }
2980 } else if (n->Opcode() == Op_Opaque1 || n->Opcode() == Op_Opaque2) {
2981 _igvn.replace_node(n, n->in(1));
2982 success = true;
2983 #if INCLUDE_RTM_OPT
2984 } else if ((n->Opcode() == Op_Opaque3) && ((Opaque3Node*)n)->rtm_opt()) {
2985 assert(C->profile_rtm(), "should be used only in rtm deoptimization code");
2986 assert((n->outcnt() == 1) && n->unique_out()->is_Cmp(), "");
2987 Node* cmp = n->unique_out();
2988 #ifdef ASSERT
2989 // Validate graph.
2990 assert((cmp->outcnt() == 1) && cmp->unique_out()->is_Bool(), "");
2991 BoolNode* bol = cmp->unique_out()->as_Bool();
2992 assert((bol->outcnt() == 1) && bol->unique_out()->is_If() &&
2993 (bol->_test._test == BoolTest::ne), "");
2994 IfNode* ifn = bol->unique_out()->as_If();
2995 assert((ifn->outcnt() == 2) &&
2996 ifn->proj_out(1)->is_uncommon_trap_proj(Deoptimization::Reason_rtm_state_change) != NULL, "");
2997 #endif
2998 Node* repl = n->in(1);
2999 if (!_has_locks) {
3000 // Remove RTM state check if there are no locks in the code.
3001 // Replace input to compare the same value.
3002 repl = (cmp->in(1) == n) ? cmp->in(2) : cmp->in(1);
3003 }
3004 _igvn.replace_node(n, repl);
3005 success = true;
3006 #endif
3007 } else if (n->Opcode() == Op_Opaque4) {
3008 _igvn.replace_node(n, n->in(2));
3009 success = true;
3010 } else if (n->Opcode() == Op_OuterStripMinedLoop) {
3011 n->as_OuterStripMinedLoop()->adjust_strip_mined_loop(&_igvn);
3012 C->remove_macro_node(n);
3013 success = true;
3014 }
3015 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
3016 progress = progress || success;
3017 }
3018 }
3019
3020 // expand arraycopy "macro" nodes first
3021 // For ReduceBulkZeroing, we must first process all arraycopy nodes
3022 // before the allocate nodes are expanded.
3023 int macro_idx = C->macro_count() - 1;
3024 while (macro_idx >= 0) {
3025 Node * n = C->macro_node(macro_idx);
3026 assert(n->is_macro(), "only macro nodes expected here");
3027 if (_igvn.type(n) == Type::TOP || n->in(0)->is_top() ) {
3028 // node is unreachable, so don't try to expand it
3029 C->remove_macro_node(n);
3030 } else if (n->is_ArrayCopy()){
3031 int macro_count = C->macro_count();
3032 expand_arraycopy_node(n->as_ArrayCopy());
3033 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
3034 }
3035 if (C->failing()) return true;
3036 macro_idx --;
3037 }
3038
3039 // expand "macro" nodes
3040 // nodes are removed from the macro list as they are processed
3041 while (C->macro_count() > 0) {
3042 int macro_count = C->macro_count();
3043 Node * n = C->macro_node(macro_count-1);
3044 assert(n->is_macro(), "only macro nodes expected here");
3045 if (_igvn.type(n) == Type::TOP || n->in(0)->is_top() ) {
3046 // node is unreachable, so don't try to expand it
3047 C->remove_macro_node(n);
3048 continue;
3049 }
3050 switch (n->class_id()) {
3051 case Node::Class_Allocate:
3052 expand_allocate(n->as_Allocate());
3053 break;
3054 case Node::Class_AllocateArray:
3055 expand_allocate_array(n->as_AllocateArray());
3056 break;
3057 case Node::Class_Lock:
3058 expand_lock_node(n->as_Lock());
3059 break;
3060 case Node::Class_Unlock:
3061 expand_unlock_node(n->as_Unlock());
3062 break;
3063 case Node::Class_CallStaticJava:
3064 expand_mh_intrinsic_return(n->as_CallStaticJava());
3065 C->remove_macro_node(n);
3066 break;
3067 default:
3068 assert(false, "unknown node type in macro list");
3069 }
3070 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
3071 if (C->failing()) return true;
3072 }
3073
3074 _igvn.set_delay_transform(false);
3075 _igvn.optimize();
3076 if (C->failing()) return true;
3077 return false;
3078 }