1 /*
2 * Copyright (c) 2005, 2018, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "compiler/compileLog.hpp"
27 #include "gc/z/zGlobals.hpp"
28 #include "gc/z/zHeap.inline.hpp"
29 #include "libadt/vectset.hpp"
30 #include "opto/addnode.hpp"
31 #include "opto/arraycopynode.hpp"
32 #include "opto/callnode.hpp"
33 #include "opto/castnode.hpp"
34 #include "opto/cfgnode.hpp"
35 #include "opto/compile.hpp"
36 #include "opto/convertnode.hpp"
37 #include "opto/graphKit.hpp"
38 #include "opto/locknode.hpp"
39 #include "opto/loopnode.hpp"
40 #include "opto/macro.hpp"
41 #include "opto/memnode.hpp"
42 #include "opto/narrowptrnode.hpp"
43 #include "opto/node.hpp"
44 #include "opto/opaquenode.hpp"
45 #include "opto/phaseX.hpp"
46 #include "opto/rootnode.hpp"
47 #include "opto/runtime.hpp"
48 #include "opto/subnode.hpp"
49 #include "opto/type.hpp"
50 #include "runtime/sharedRuntime.hpp"
51
52
53 //
54 // Replace any references to "oldref" in inputs to "use" with "newref".
55 // Returns the number of replacements made.
56 //
57 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) {
58 int nreplacements = 0;
59 uint req = use->req();
60 for (uint j = 0; j < use->len(); j++) {
61 Node *uin = use->in(j);
62 if (uin == oldref) {
63 if (j < req)
64 use->set_req(j, newref);
65 else
66 use->set_prec(j, newref);
67 nreplacements++;
68 } else if (j >= req && uin == NULL) {
69 break;
70 }
71 }
72 return nreplacements;
73 }
74
75 void PhaseMacroExpand::copy_call_debug_info(CallNode *oldcall, CallNode * newcall) {
76 // Copy debug information and adjust JVMState information
77 uint old_dbg_start = oldcall->tf()->domain()->cnt();
78 uint new_dbg_start = newcall->tf()->domain()->cnt();
79 int jvms_adj = new_dbg_start - old_dbg_start;
80 assert (new_dbg_start == newcall->req(), "argument count mismatch");
81
82 // SafePointScalarObject node could be referenced several times in debug info.
83 // Use Dict to record cloned nodes.
84 Dict* sosn_map = new Dict(cmpkey,hashkey);
85 for (uint i = old_dbg_start; i < oldcall->req(); i++) {
86 Node* old_in = oldcall->in(i);
87 // Clone old SafePointScalarObjectNodes, adjusting their field contents.
88 if (old_in != NULL && old_in->is_SafePointScalarObject()) {
89 SafePointScalarObjectNode* old_sosn = old_in->as_SafePointScalarObject();
90 uint old_unique = C->unique();
91 Node* new_in = old_sosn->clone(sosn_map);
92 if (old_unique != C->unique()) { // New node?
93 new_in->set_req(0, C->root()); // reset control edge
94 new_in = transform_later(new_in); // Register new node.
95 }
96 old_in = new_in;
97 }
98 newcall->add_req(old_in);
99 }
100
101 // JVMS may be shared so clone it before we modify it
102 newcall->set_jvms(oldcall->jvms() != NULL ? oldcall->jvms()->clone_deep(C) : NULL);
103 for (JVMState *jvms = newcall->jvms(); jvms != NULL; jvms = jvms->caller()) {
104 jvms->set_map(newcall);
105 jvms->set_locoff(jvms->locoff()+jvms_adj);
106 jvms->set_stkoff(jvms->stkoff()+jvms_adj);
107 jvms->set_monoff(jvms->monoff()+jvms_adj);
108 jvms->set_scloff(jvms->scloff()+jvms_adj);
109 jvms->set_endoff(jvms->endoff()+jvms_adj);
110 }
111 }
112
113 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word, int mask, int bits, bool return_fast_path) {
114 Node* cmp;
115 if (mask != 0) {
116 Node* and_node = transform_later(new AndXNode(word, MakeConX(mask)));
117 cmp = transform_later(new CmpXNode(and_node, MakeConX(bits)));
118 } else {
119 cmp = word;
120 }
121 Node* bol = transform_later(new BoolNode(cmp, BoolTest::ne));
122 IfNode* iff = new IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN );
123 transform_later(iff);
124
125 // Fast path taken.
126 Node *fast_taken = transform_later(new IfFalseNode(iff));
127
128 // Fast path not-taken, i.e. slow path
129 Node *slow_taken = transform_later(new IfTrueNode(iff));
130
131 if (return_fast_path) {
132 region->init_req(edge, slow_taken); // Capture slow-control
133 return fast_taken;
134 } else {
135 region->init_req(edge, fast_taken); // Capture fast-control
136 return slow_taken;
137 }
138 }
139
140 //--------------------copy_predefined_input_for_runtime_call--------------------
141 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) {
142 // Set fixed predefined input arguments
143 call->init_req( TypeFunc::Control, ctrl );
144 call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) );
145 call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ?????
146 call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) );
147 call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) );
148 }
149
150 //------------------------------make_slow_call---------------------------------
151 CallNode* PhaseMacroExpand::make_slow_call(CallNode *oldcall, const TypeFunc* slow_call_type,
152 address slow_call, const char* leaf_name, Node* slow_path,
153 Node* parm0, Node* parm1, Node* parm2) {
154
155 // Slow-path call
156 CallNode *call = leaf_name
157 ? (CallNode*)new CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM )
158 : (CallNode*)new CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), oldcall->jvms()->bci(), TypeRawPtr::BOTTOM );
159
160 // Slow path call has no side-effects, uses few values
161 copy_predefined_input_for_runtime_call(slow_path, oldcall, call );
162 if (parm0 != NULL) call->init_req(TypeFunc::Parms+0, parm0);
163 if (parm1 != NULL) call->init_req(TypeFunc::Parms+1, parm1);
164 if (parm2 != NULL) call->init_req(TypeFunc::Parms+2, parm2);
165 copy_call_debug_info(oldcall, call);
166 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
167 _igvn.replace_node(oldcall, call);
168 transform_later(call);
169
170 return call;
171 }
172
173 void PhaseMacroExpand::extract_call_projections(CallNode *call) {
174 _fallthroughproj = NULL;
175 _fallthroughcatchproj = NULL;
176 _ioproj_fallthrough = NULL;
177 _ioproj_catchall = NULL;
178 _catchallcatchproj = NULL;
179 _memproj_fallthrough = NULL;
180 _memproj_catchall = NULL;
181 _resproj = NULL;
182 for (DUIterator_Fast imax, i = call->fast_outs(imax); i < imax; i++) {
183 ProjNode *pn = call->fast_out(i)->as_Proj();
184 switch (pn->_con) {
185 case TypeFunc::Control:
186 {
187 // For Control (fallthrough) and I_O (catch_all_index) we have CatchProj -> Catch -> Proj
188 _fallthroughproj = pn;
189 DUIterator_Fast jmax, j = pn->fast_outs(jmax);
190 const Node *cn = pn->fast_out(j);
191 if (cn->is_Catch()) {
192 ProjNode *cpn = NULL;
193 for (DUIterator_Fast kmax, k = cn->fast_outs(kmax); k < kmax; k++) {
194 cpn = cn->fast_out(k)->as_Proj();
195 assert(cpn->is_CatchProj(), "must be a CatchProjNode");
196 if (cpn->_con == CatchProjNode::fall_through_index)
197 _fallthroughcatchproj = cpn;
198 else {
199 assert(cpn->_con == CatchProjNode::catch_all_index, "must be correct index.");
200 _catchallcatchproj = cpn;
201 }
202 }
203 }
204 break;
205 }
206 case TypeFunc::I_O:
207 if (pn->_is_io_use)
208 _ioproj_catchall = pn;
209 else
210 _ioproj_fallthrough = pn;
211 break;
212 case TypeFunc::Memory:
213 if (pn->_is_io_use)
214 _memproj_catchall = pn;
215 else
216 _memproj_fallthrough = pn;
217 break;
218 case TypeFunc::Parms:
219 _resproj = pn;
220 break;
221 default:
222 assert(false, "unexpected projection from allocation node.");
223 }
224 }
225
226 }
227
228 // Eliminate a card mark sequence. p2x is a ConvP2XNode
229 void PhaseMacroExpand::eliminate_card_mark(Node* p2x) {
230 assert(p2x->Opcode() == Op_CastP2X, "ConvP2XNode required");
231 if (!UseG1GC) {
232 // vanilla/CMS post barrier
233 Node *shift = p2x->unique_out();
234 Node *addp = shift->unique_out();
235 for (DUIterator_Last jmin, j = addp->last_outs(jmin); j >= jmin; --j) {
236 Node *mem = addp->last_out(j);
237 if (UseCondCardMark && mem->is_Load()) {
238 assert(mem->Opcode() == Op_LoadB, "unexpected code shape");
239 // The load is checking if the card has been written so
240 // replace it with zero to fold the test.
241 _igvn.replace_node(mem, intcon(0));
242 continue;
243 }
244 assert(mem->is_Store(), "store required");
245 _igvn.replace_node(mem, mem->in(MemNode::Memory));
246 }
247 } else {
248 // G1 pre/post barriers
249 assert(p2x->outcnt() <= 2, "expects 1 or 2 users: Xor and URShift nodes");
250 // It could be only one user, URShift node, in Object.clone() intrinsic
251 // but the new allocation is passed to arraycopy stub and it could not
252 // be scalar replaced. So we don't check the case.
253
254 // An other case of only one user (Xor) is when the value check for NULL
255 // in G1 post barrier is folded after CCP so the code which used URShift
256 // is removed.
257
258 // Take Region node before eliminating post barrier since it also
259 // eliminates CastP2X node when it has only one user.
260 Node* this_region = p2x->in(0);
261 assert(this_region != NULL, "");
262
263 // Remove G1 post barrier.
264
265 // Search for CastP2X->Xor->URShift->Cmp path which
266 // checks if the store done to a different from the value's region.
267 // And replace Cmp with #0 (false) to collapse G1 post barrier.
268 Node* xorx = p2x->find_out_with(Op_XorX);
269 if (xorx != NULL) {
270 Node* shift = xorx->unique_out();
271 Node* cmpx = shift->unique_out();
272 assert(cmpx->is_Cmp() && cmpx->unique_out()->is_Bool() &&
273 cmpx->unique_out()->as_Bool()->_test._test == BoolTest::ne,
274 "missing region check in G1 post barrier");
275 _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ));
276
277 // Remove G1 pre barrier.
278
279 // Search "if (marking != 0)" check and set it to "false".
280 // There is no G1 pre barrier if previous stored value is NULL
281 // (for example, after initialization).
282 if (this_region->is_Region() && this_region->req() == 3) {
283 int ind = 1;
284 if (!this_region->in(ind)->is_IfFalse()) {
285 ind = 2;
286 }
287 if (this_region->in(ind)->is_IfFalse() &&
288 this_region->in(ind)->in(0)->Opcode() == Op_If) {
289 Node* bol = this_region->in(ind)->in(0)->in(1);
290 assert(bol->is_Bool(), "");
291 cmpx = bol->in(1);
292 if (bol->as_Bool()->_test._test == BoolTest::ne &&
293 cmpx->is_Cmp() && cmpx->in(2) == intcon(0) &&
294 cmpx->in(1)->is_Load()) {
295 Node* adr = cmpx->in(1)->as_Load()->in(MemNode::Address);
296 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() +
297 SATBMarkQueue::byte_offset_of_active());
298 if (adr->is_AddP() && adr->in(AddPNode::Base) == top() &&
299 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
300 adr->in(AddPNode::Offset) == MakeConX(marking_offset)) {
301 _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ));
302 }
303 }
304 }
305 }
306 } else {
307 assert(!GraphKit::use_ReduceInitialCardMarks(), "can only happen with card marking");
308 // This is a G1 post barrier emitted by the Object.clone() intrinsic.
309 // Search for the CastP2X->URShiftX->AddP->LoadB->Cmp path which checks if the card
310 // is marked as young_gen and replace the Cmp with 0 (false) to collapse the barrier.
311 Node* shift = p2x->find_out_with(Op_URShiftX);
312 assert(shift != NULL, "missing G1 post barrier");
313 Node* addp = shift->unique_out();
314 Node* load = addp->find_out_with(Op_LoadB);
315 assert(load != NULL, "missing G1 post barrier");
316 Node* cmpx = load->unique_out();
317 assert(cmpx->is_Cmp() && cmpx->unique_out()->is_Bool() &&
318 cmpx->unique_out()->as_Bool()->_test._test == BoolTest::ne,
319 "missing card value check in G1 post barrier");
320 _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ));
321 // There is no G1 pre barrier in this case
322 }
323 // Now CastP2X can be removed since it is used only on dead path
324 // which currently still alive until igvn optimize it.
325 assert(p2x->outcnt() == 0 || p2x->unique_out()->Opcode() == Op_URShiftX, "");
326 _igvn.replace_node(p2x, top());
327 }
328 }
329
330 // Search for a memory operation for the specified memory slice.
331 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) {
332 Node *orig_mem = mem;
333 Node *alloc_mem = alloc->in(TypeFunc::Memory);
334 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr();
335 while (true) {
336 if (mem == alloc_mem || mem == start_mem ) {
337 return mem; // hit one of our sentinels
338 } else if (mem->is_MergeMem()) {
339 mem = mem->as_MergeMem()->memory_at(alias_idx);
340 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) {
341 Node *in = mem->in(0);
342 // we can safely skip over safepoints, calls, locks and membars because we
343 // already know that the object is safe to eliminate.
344 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) {
345 return in;
346 } else if (in->is_Call()) {
347 CallNode *call = in->as_Call();
348 if (call->may_modify(tinst, phase)) {
349 assert(call->is_ArrayCopy(), "ArrayCopy is the only call node that doesn't make allocation escape");
350 if (call->as_ArrayCopy()->modifies(offset, offset, phase, false)) {
351 return in;
352 }
353 }
354 mem = in->in(TypeFunc::Memory);
355 } else if (in->is_MemBar()) {
356 ArrayCopyNode* ac = NULL;
357 if (ArrayCopyNode::may_modify(tinst, in->as_MemBar(), phase, ac)) {
358 assert(ac != NULL && ac->is_clonebasic(), "Only basic clone is a non escaping clone");
359 return ac;
360 }
361 mem = in->in(TypeFunc::Memory);
362 } else {
363 assert(false, "unexpected projection");
364 }
365 } else if (mem->is_Store()) {
366 const TypePtr* atype = mem->as_Store()->adr_type();
367 int adr_idx = phase->C->get_alias_index(atype);
368 if (adr_idx == alias_idx) {
369 assert(atype->isa_oopptr(), "address type must be oopptr");
370 int adr_offset = atype->offset();
371 uint adr_iid = atype->is_oopptr()->instance_id();
372 // Array elements references have the same alias_idx
373 // but different offset and different instance_id.
374 if (adr_offset == offset && adr_iid == alloc->_idx)
375 return mem;
376 } else {
377 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw");
378 }
379 mem = mem->in(MemNode::Memory);
380 } else if (mem->is_ClearArray()) {
381 if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) {
382 // Can not bypass initialization of the instance
383 // we are looking.
384 debug_only(intptr_t offset;)
385 assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity");
386 InitializeNode* init = alloc->as_Allocate()->initialization();
387 // We are looking for stored value, return Initialize node
388 // or memory edge from Allocate node.
389 if (init != NULL)
390 return init;
391 else
392 return alloc->in(TypeFunc::Memory); // It will produce zero value (see callers).
393 }
394 // Otherwise skip it (the call updated 'mem' value).
395 } else if (mem->Opcode() == Op_SCMemProj) {
396 mem = mem->in(0);
397 Node* adr = NULL;
398 if (mem->is_LoadStore()) {
399 adr = mem->in(MemNode::Address);
400 } else {
401 assert(mem->Opcode() == Op_EncodeISOArray ||
402 mem->Opcode() == Op_StrCompressedCopy, "sanity");
403 adr = mem->in(3); // Destination array
404 }
405 const TypePtr* atype = adr->bottom_type()->is_ptr();
406 int adr_idx = phase->C->get_alias_index(atype);
407 if (adr_idx == alias_idx) {
408 DEBUG_ONLY(mem->dump();)
409 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
410 return NULL;
411 }
412 mem = mem->in(MemNode::Memory);
413 } else if (mem->Opcode() == Op_StrInflatedCopy) {
414 Node* adr = mem->in(3); // Destination array
415 const TypePtr* atype = adr->bottom_type()->is_ptr();
416 int adr_idx = phase->C->get_alias_index(atype);
417 if (adr_idx == alias_idx) {
418 DEBUG_ONLY(mem->dump();)
419 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field");
420 return NULL;
421 }
422 mem = mem->in(MemNode::Memory);
423 } else {
424 return mem;
425 }
426 assert(mem != orig_mem, "dead memory loop");
427 }
428 }
429
430 // Generate loads from source of the arraycopy for fields of
431 // destination needed at a deoptimization point
432 Node* PhaseMacroExpand::make_arraycopy_load(ArrayCopyNode* ac, intptr_t offset, Node* ctl, Node* mem, BasicType ft, const Type *ftype, AllocateNode *alloc) {
433 BasicType bt = ft;
434 const Type *type = ftype;
435 if (ft == T_NARROWOOP) {
436 bt = T_OBJECT;
437 type = ftype->make_oopptr();
438 }
439 Node* res = NULL;
440 if (ac->is_clonebasic()) {
441 Node* base = ac->in(ArrayCopyNode::Src)->in(AddPNode::Base);
442 Node* adr = _igvn.transform(new AddPNode(base, base, MakeConX(offset)));
443 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset);
444 res = LoadNode::make(_igvn, ctl, mem, adr, adr_type, type, bt, MemNode::unordered, LoadNode::Pinned);
445 } else {
446 if (ac->modifies(offset, offset, &_igvn, true)) {
447 assert(ac->in(ArrayCopyNode::Dest) == alloc->result_cast(), "arraycopy destination should be allocation's result");
448 uint shift = exact_log2(type2aelembytes(bt));
449 Node* diff = _igvn.transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
450 #ifdef _LP64
451 diff = _igvn.transform(new ConvI2LNode(diff));
452 #endif
453 diff = _igvn.transform(new LShiftXNode(diff, intcon(shift)));
454
455 Node* off = _igvn.transform(new AddXNode(MakeConX(offset), diff));
456 Node* base = ac->in(ArrayCopyNode::Src);
457 Node* adr = _igvn.transform(new AddPNode(base, base, off));
458 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset);
459 res = LoadNode::make(_igvn, ctl, mem, adr, adr_type, type, bt, MemNode::unordered, LoadNode::Pinned);
460 }
461 }
462 if (res != NULL) {
463 res = _igvn.transform(res);
464 if (ftype->isa_narrowoop()) {
465 // PhaseMacroExpand::scalar_replacement adds DecodeN nodes
466 res = _igvn.transform(new EncodePNode(res, ftype));
467 }
468 return res;
469 }
470 return NULL;
471 }
472
473 //
474 // Given a Memory Phi, compute a value Phi containing the values from stores
475 // on the input paths.
476 // Note: this function is recursive, its depth is limited by the "level" argument
477 // Returns the computed Phi, or NULL if it cannot compute it.
478 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) {
479 assert(mem->is_Phi(), "sanity");
480 int alias_idx = C->get_alias_index(adr_t);
481 int offset = adr_t->offset();
482 int instance_id = adr_t->instance_id();
483
484 // Check if an appropriate value phi already exists.
485 Node* region = mem->in(0);
486 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) {
487 Node* phi = region->fast_out(k);
488 if (phi->is_Phi() && phi != mem &&
489 phi->as_Phi()->is_same_inst_field(phi_type, (int)mem->_idx, instance_id, alias_idx, offset)) {
490 return phi;
491 }
492 }
493 // Check if an appropriate new value phi already exists.
494 Node* new_phi = value_phis->find(mem->_idx);
495 if (new_phi != NULL)
496 return new_phi;
497
498 if (level <= 0) {
499 return NULL; // Give up: phi tree too deep
500 }
501 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
502 Node *alloc_mem = alloc->in(TypeFunc::Memory);
503
504 uint length = mem->req();
505 GrowableArray <Node *> values(length, length, NULL, false);
506
507 // create a new Phi for the value
508 PhiNode *phi = new PhiNode(mem->in(0), phi_type, NULL, mem->_idx, instance_id, alias_idx, offset);
509 transform_later(phi);
510 value_phis->push(phi, mem->_idx);
511
512 for (uint j = 1; j < length; j++) {
513 Node *in = mem->in(j);
514 if (in == NULL || in->is_top()) {
515 values.at_put(j, in);
516 } else {
517 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn);
518 if (val == start_mem || val == alloc_mem) {
519 // hit a sentinel, return appropriate 0 value
520 values.at_put(j, _igvn.zerocon(ft));
521 continue;
522 }
523 if (val->is_Initialize()) {
524 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
525 }
526 if (val == NULL) {
527 return NULL; // can't find a value on this path
528 }
529 if (val == mem) {
530 values.at_put(j, mem);
531 } else if (val->is_Store()) {
532 values.at_put(j, val->in(MemNode::ValueIn));
533 } else if(val->is_Proj() && val->in(0) == alloc) {
534 values.at_put(j, _igvn.zerocon(ft));
535 } else if (val->is_Phi()) {
536 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1);
537 if (val == NULL) {
538 return NULL;
539 }
540 values.at_put(j, val);
541 } else if (val->Opcode() == Op_SCMemProj) {
542 assert(val->in(0)->is_LoadStore() ||
543 val->in(0)->Opcode() == Op_EncodeISOArray ||
544 val->in(0)->Opcode() == Op_StrCompressedCopy, "sanity");
545 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
546 return NULL;
547 } else if (val->is_ArrayCopy()) {
548 Node* res = make_arraycopy_load(val->as_ArrayCopy(), offset, val->in(0), val->in(TypeFunc::Memory), ft, phi_type, alloc);
549 if (res == NULL) {
550 return NULL;
551 }
552 values.at_put(j, res);
553 } else {
554 #ifdef ASSERT
555 val->dump();
556 assert(false, "unknown node on this path");
557 #endif
558 return NULL; // unknown node on this path
559 }
560 }
561 }
562 // Set Phi's inputs
563 for (uint j = 1; j < length; j++) {
564 if (values.at(j) == mem) {
565 phi->init_req(j, phi);
566 } else {
567 phi->init_req(j, values.at(j));
568 }
569 }
570 return phi;
571 }
572
573 // Search the last value stored into the object's field.
574 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, Node *sfpt_ctl, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, AllocateNode *alloc) {
575 assert(adr_t->is_known_instance_field(), "instance required");
576 int instance_id = adr_t->instance_id();
577 assert((uint)instance_id == alloc->_idx, "wrong allocation");
578
579 int alias_idx = C->get_alias_index(adr_t);
580 int offset = adr_t->offset();
581 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
582 Node *alloc_ctrl = alloc->in(TypeFunc::Control);
583 Node *alloc_mem = alloc->in(TypeFunc::Memory);
584 Arena *a = Thread::current()->resource_area();
585 VectorSet visited(a);
586
587
588 bool done = sfpt_mem == alloc_mem;
589 Node *mem = sfpt_mem;
590 while (!done) {
591 if (visited.test_set(mem->_idx)) {
592 return NULL; // found a loop, give up
593 }
594 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn);
595 if (mem == start_mem || mem == alloc_mem) {
596 done = true; // hit a sentinel, return appropriate 0 value
597 } else if (mem->is_Initialize()) {
598 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
599 if (mem == NULL) {
600 done = true; // Something go wrong.
601 } else if (mem->is_Store()) {
602 const TypePtr* atype = mem->as_Store()->adr_type();
603 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice");
604 done = true;
605 }
606 } else if (mem->is_Store()) {
607 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr();
608 assert(atype != NULL, "address type must be oopptr");
609 assert(C->get_alias_index(atype) == alias_idx &&
610 atype->is_known_instance_field() && atype->offset() == offset &&
611 atype->instance_id() == instance_id, "store is correct memory slice");
612 done = true;
613 } else if (mem->is_Phi()) {
614 // try to find a phi's unique input
615 Node *unique_input = NULL;
616 Node *top = C->top();
617 for (uint i = 1; i < mem->req(); i++) {
618 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn);
619 if (n == NULL || n == top || n == mem) {
620 continue;
621 } else if (unique_input == NULL) {
622 unique_input = n;
623 } else if (unique_input != n) {
624 unique_input = top;
625 break;
626 }
627 }
628 if (unique_input != NULL && unique_input != top) {
629 mem = unique_input;
630 } else {
631 done = true;
632 }
633 } else if (mem->is_ArrayCopy()) {
634 done = true;
635 } else {
636 assert(false, "unexpected node");
637 }
638 }
639 if (mem != NULL) {
640 if (mem == start_mem || mem == alloc_mem) {
641 // hit a sentinel, return appropriate 0 value
642 return _igvn.zerocon(ft);
643 } else if (mem->is_Store()) {
644 return mem->in(MemNode::ValueIn);
645 } else if (mem->is_Phi()) {
646 // attempt to produce a Phi reflecting the values on the input paths of the Phi
647 Node_Stack value_phis(a, 8);
648 Node * phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit);
649 if (phi != NULL) {
650 return phi;
651 } else {
652 // Kill all new Phis
653 while(value_phis.is_nonempty()) {
654 Node* n = value_phis.node();
655 _igvn.replace_node(n, C->top());
656 value_phis.pop();
657 }
658 }
659 } else if (mem->is_ArrayCopy()) {
660 Node* ctl = mem->in(0);
661 Node* m = mem->in(TypeFunc::Memory);
662 if (sfpt_ctl->is_Proj() && sfpt_ctl->as_Proj()->is_uncommon_trap_proj(Deoptimization::Reason_none)) {
663 // pin the loads in the uncommon trap path
664 ctl = sfpt_ctl;
665 m = sfpt_mem;
666 }
667 return make_arraycopy_load(mem->as_ArrayCopy(), offset, ctl, m, ft, ftype, alloc);
668 }
669 }
670 // Something go wrong.
671 return NULL;
672 }
673
674 // Check the possibility of scalar replacement.
675 bool PhaseMacroExpand::can_eliminate_allocation(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
676 // Scan the uses of the allocation to check for anything that would
677 // prevent us from eliminating it.
678 NOT_PRODUCT( const char* fail_eliminate = NULL; )
679 DEBUG_ONLY( Node* disq_node = NULL; )
680 bool can_eliminate = true;
681
682 Node* res = alloc->result_cast();
683 const TypeOopPtr* res_type = NULL;
684 if (res == NULL) {
685 // All users were eliminated.
686 } else if (!res->is_CheckCastPP()) {
687 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";)
688 can_eliminate = false;
689 } else {
690 res_type = _igvn.type(res)->isa_oopptr();
691 if (res_type == NULL) {
692 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";)
693 can_eliminate = false;
694 } else if (res_type->isa_aryptr()) {
695 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1);
696 if (length < 0) {
697 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";)
698 can_eliminate = false;
699 }
700 }
701 }
702
703 if (can_eliminate && res != NULL) {
704 for (DUIterator_Fast jmax, j = res->fast_outs(jmax);
705 j < jmax && can_eliminate; j++) {
706 Node* use = res->fast_out(j);
707
708 if (use->is_AddP()) {
709 const TypePtr* addp_type = _igvn.type(use)->is_ptr();
710 int offset = addp_type->offset();
711
712 if (offset == Type::OffsetTop || offset == Type::OffsetBot) {
713 NOT_PRODUCT(fail_eliminate = "Undefined field referrence";)
714 can_eliminate = false;
715 break;
716 }
717 for (DUIterator_Fast kmax, k = use->fast_outs(kmax);
718 k < kmax && can_eliminate; k++) {
719 Node* n = use->fast_out(k);
720 if (!n->is_Store() && n->Opcode() != Op_CastP2X &&
721 !(n->is_ArrayCopy() &&
722 n->as_ArrayCopy()->is_clonebasic() &&
723 n->in(ArrayCopyNode::Dest) == use)) {
724 DEBUG_ONLY(disq_node = n;)
725 if (n->is_Load() || n->is_LoadStore()) {
726 NOT_PRODUCT(fail_eliminate = "Field load";)
727 } else {
728 NOT_PRODUCT(fail_eliminate = "Not store field referrence";)
729 }
730 can_eliminate = false;
731 }
732 }
733 } else if (use->is_ArrayCopy() &&
734 (use->as_ArrayCopy()->is_arraycopy_validated() ||
735 use->as_ArrayCopy()->is_copyof_validated() ||
736 use->as_ArrayCopy()->is_copyofrange_validated()) &&
737 use->in(ArrayCopyNode::Dest) == res) {
738 // ok to eliminate
739 } else if (use->is_SafePoint()) {
740 SafePointNode* sfpt = use->as_SafePoint();
741 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) {
742 // Object is passed as argument.
743 DEBUG_ONLY(disq_node = use;)
744 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";)
745 can_eliminate = false;
746 }
747 Node* sfptMem = sfpt->memory();
748 if (sfptMem == NULL || sfptMem->is_top()) {
749 DEBUG_ONLY(disq_node = use;)
750 NOT_PRODUCT(fail_eliminate = "NULL or TOP memory";)
751 can_eliminate = false;
752 } else {
753 safepoints.append_if_missing(sfpt);
754 }
755 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
756 if (use->is_Phi()) {
757 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
758 NOT_PRODUCT(fail_eliminate = "Object is return value";)
759 } else {
760 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
761 }
762 DEBUG_ONLY(disq_node = use;)
763 } else {
764 if (use->Opcode() == Op_Return) {
765 NOT_PRODUCT(fail_eliminate = "Object is return value";)
766 }else {
767 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
768 }
769 DEBUG_ONLY(disq_node = use;)
770 }
771 can_eliminate = false;
772 }
773 }
774 }
775
776 #ifndef PRODUCT
777 if (PrintEliminateAllocations) {
778 if (can_eliminate) {
779 tty->print("Scalar ");
780 if (res == NULL)
781 alloc->dump();
782 else
783 res->dump();
784 } else if (alloc->_is_scalar_replaceable) {
785 tty->print("NotScalar (%s)", fail_eliminate);
786 if (res == NULL)
787 alloc->dump();
788 else
789 res->dump();
790 #ifdef ASSERT
791 if (disq_node != NULL) {
792 tty->print(" >>>> ");
793 disq_node->dump();
794 }
795 #endif /*ASSERT*/
796 }
797 }
798 #endif
799 return can_eliminate;
800 }
801
802 // Do scalar replacement.
803 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
804 GrowableArray <SafePointNode *> safepoints_done;
805
806 ciKlass* klass = NULL;
807 ciInstanceKlass* iklass = NULL;
808 int nfields = 0;
809 int array_base = 0;
810 int element_size = 0;
811 BasicType basic_elem_type = T_ILLEGAL;
812 ciType* elem_type = NULL;
813
814 Node* res = alloc->result_cast();
815 assert(res == NULL || res->is_CheckCastPP(), "unexpected AllocateNode result");
816 const TypeOopPtr* res_type = NULL;
817 if (res != NULL) { // Could be NULL when there are no users
818 res_type = _igvn.type(res)->isa_oopptr();
819 }
820
821 if (res != NULL) {
822 klass = res_type->klass();
823 if (res_type->isa_instptr()) {
824 // find the fields of the class which will be needed for safepoint debug information
825 assert(klass->is_instance_klass(), "must be an instance klass.");
826 iklass = klass->as_instance_klass();
827 nfields = iklass->nof_nonstatic_fields();
828 } else {
829 // find the array's elements which will be needed for safepoint debug information
830 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
831 assert(klass->is_array_klass() && nfields >= 0, "must be an array klass.");
832 elem_type = klass->as_array_klass()->element_type();
833 basic_elem_type = elem_type->basic_type();
834 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
835 element_size = type2aelembytes(basic_elem_type);
836 }
837 }
838 //
839 // Process the safepoint uses
840 //
841 while (safepoints.length() > 0) {
842 SafePointNode* sfpt = safepoints.pop();
843 Node* mem = sfpt->memory();
844 Node* ctl = sfpt->control();
845 assert(sfpt->jvms() != NULL, "missed JVMS");
846 // Fields of scalar objs are referenced only at the end
847 // of regular debuginfo at the last (youngest) JVMS.
848 // Record relative start index.
849 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff());
850 SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type,
851 #ifdef ASSERT
852 alloc,
853 #endif
854 first_ind, nfields);
855 sobj->init_req(0, C->root());
856 transform_later(sobj);
857
858 // Scan object's fields adding an input to the safepoint for each field.
859 for (int j = 0; j < nfields; j++) {
860 intptr_t offset;
861 ciField* field = NULL;
862 if (iklass != NULL) {
863 field = iklass->nonstatic_field_at(j);
864 offset = field->offset();
865 elem_type = field->type();
866 basic_elem_type = field->layout_type();
867 } else {
868 offset = array_base + j * (intptr_t)element_size;
869 }
870
871 const Type *field_type;
872 // The next code is taken from Parse::do_get_xxx().
873 if (basic_elem_type == T_OBJECT || basic_elem_type == T_ARRAY) {
874 if (!elem_type->is_loaded()) {
875 field_type = TypeInstPtr::BOTTOM;
876 } else if (field != NULL && field->is_static_constant()) {
877 // This can happen if the constant oop is non-perm.
878 ciObject* con = field->constant_value().as_object();
879 // Do not "join" in the previous type; it doesn't add value,
880 // and may yield a vacuous result if the field is of interface type.
881 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr();
882 assert(field_type != NULL, "field singleton type must be consistent");
883 } else {
884 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
885 }
886 if (UseCompressedOops) {
887 field_type = field_type->make_narrowoop();
888 basic_elem_type = T_NARROWOOP;
889 }
890 } else {
891 field_type = Type::get_const_basic_type(basic_elem_type);
892 }
893
894 const TypeOopPtr *field_addr_type = res_type->add_offset(offset)->isa_oopptr();
895
896 Node *field_val = value_from_mem(mem, ctl, basic_elem_type, field_type, field_addr_type, alloc);
897 if (field_val == NULL) {
898 // We weren't able to find a value for this field,
899 // give up on eliminating this allocation.
900
901 // Remove any extra entries we added to the safepoint.
902 uint last = sfpt->req() - 1;
903 for (int k = 0; k < j; k++) {
904 sfpt->del_req(last--);
905 }
906 _igvn._worklist.push(sfpt);
907 // rollback processed safepoints
908 while (safepoints_done.length() > 0) {
909 SafePointNode* sfpt_done = safepoints_done.pop();
910 // remove any extra entries we added to the safepoint
911 last = sfpt_done->req() - 1;
912 for (int k = 0; k < nfields; k++) {
913 sfpt_done->del_req(last--);
914 }
915 JVMState *jvms = sfpt_done->jvms();
916 jvms->set_endoff(sfpt_done->req());
917 // Now make a pass over the debug information replacing any references
918 // to SafePointScalarObjectNode with the allocated object.
919 int start = jvms->debug_start();
920 int end = jvms->debug_end();
921 for (int i = start; i < end; i++) {
922 if (sfpt_done->in(i)->is_SafePointScalarObject()) {
923 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject();
924 if (scobj->first_index(jvms) == sfpt_done->req() &&
925 scobj->n_fields() == (uint)nfields) {
926 assert(scobj->alloc() == alloc, "sanity");
927 sfpt_done->set_req(i, res);
928 }
929 }
930 }
931 _igvn._worklist.push(sfpt_done);
932 }
933 #ifndef PRODUCT
934 if (PrintEliminateAllocations) {
935 if (field != NULL) {
936 tty->print("=== At SafePoint node %d can't find value of Field: ",
937 sfpt->_idx);
938 field->print();
939 int field_idx = C->get_alias_index(field_addr_type);
940 tty->print(" (alias_idx=%d)", field_idx);
941 } else { // Array's element
942 tty->print("=== At SafePoint node %d can't find value of array element [%d]",
943 sfpt->_idx, j);
944 }
945 tty->print(", which prevents elimination of: ");
946 if (res == NULL)
947 alloc->dump();
948 else
949 res->dump();
950 }
951 #endif
952 return false;
953 }
954 if (UseCompressedOops && field_type->isa_narrowoop()) {
955 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
956 // to be able scalar replace the allocation.
957 if (field_val->is_EncodeP()) {
958 field_val = field_val->in(1);
959 } else {
960 field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type()));
961 }
962 }
963 sfpt->add_req(field_val);
964 }
965 JVMState *jvms = sfpt->jvms();
966 jvms->set_endoff(sfpt->req());
967 // Now make a pass over the debug information replacing any references
968 // to the allocated object with "sobj"
969 int start = jvms->debug_start();
970 int end = jvms->debug_end();
971 sfpt->replace_edges_in_range(res, sobj, start, end);
972 _igvn._worklist.push(sfpt);
973 safepoints_done.append_if_missing(sfpt); // keep it for rollback
974 }
975 return true;
976 }
977
978 static void disconnect_projections(MultiNode* n, PhaseIterGVN& igvn) {
979 Node* ctl_proj = n->proj_out_or_null(TypeFunc::Control);
980 Node* mem_proj = n->proj_out_or_null(TypeFunc::Memory);
981 if (ctl_proj != NULL) {
982 igvn.replace_node(ctl_proj, n->in(0));
983 }
984 if (mem_proj != NULL) {
985 igvn.replace_node(mem_proj, n->in(TypeFunc::Memory));
986 }
987 }
988
989 // Process users of eliminated allocation.
990 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc) {
991 Node* res = alloc->result_cast();
992 if (res != NULL) {
993 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
994 Node *use = res->last_out(j);
995 uint oc1 = res->outcnt();
996
997 if (use->is_AddP()) {
998 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) {
999 Node *n = use->last_out(k);
1000 uint oc2 = use->outcnt();
1001 if (n->is_Store()) {
1002 #ifdef ASSERT
1003 // Verify that there is no dependent MemBarVolatile nodes,
1004 // they should be removed during IGVN, see MemBarNode::Ideal().
1005 for (DUIterator_Fast pmax, p = n->fast_outs(pmax);
1006 p < pmax; p++) {
1007 Node* mb = n->fast_out(p);
1008 assert(mb->is_Initialize() || !mb->is_MemBar() ||
1009 mb->req() <= MemBarNode::Precedent ||
1010 mb->in(MemBarNode::Precedent) != n,
1011 "MemBarVolatile should be eliminated for non-escaping object");
1012 }
1013 #endif
1014 _igvn.replace_node(n, n->in(MemNode::Memory));
1015 } else if (n->is_ArrayCopy()) {
1016 // Disconnect ArrayCopy node
1017 ArrayCopyNode* ac = n->as_ArrayCopy();
1018 assert(ac->is_clonebasic(), "unexpected array copy kind");
1019 Node* membar_after = ac->proj_out(TypeFunc::Control)->unique_ctrl_out();
1020 disconnect_projections(ac, _igvn);
1021 assert(alloc->in(0)->is_Proj() && alloc->in(0)->in(0)->Opcode() == Op_MemBarCPUOrder, "mem barrier expected before allocation");
1022 Node* membar_before = alloc->in(0)->in(0);
1023 disconnect_projections(membar_before->as_MemBar(), _igvn);
1024 if (membar_after->is_MemBar()) {
1025 disconnect_projections(membar_after->as_MemBar(), _igvn);
1026 }
1027 } else {
1028 eliminate_card_mark(n);
1029 }
1030 k -= (oc2 - use->outcnt());
1031 }
1032 } else if (use->is_ArrayCopy()) {
1033 // Disconnect ArrayCopy node
1034 ArrayCopyNode* ac = use->as_ArrayCopy();
1035 assert(ac->is_arraycopy_validated() ||
1036 ac->is_copyof_validated() ||
1037 ac->is_copyofrange_validated(), "unsupported");
1038 CallProjections callprojs;
1039 ac->extract_projections(&callprojs, true);
1040
1041 _igvn.replace_node(callprojs.fallthrough_ioproj, ac->in(TypeFunc::I_O));
1042 _igvn.replace_node(callprojs.fallthrough_memproj, ac->in(TypeFunc::Memory));
1043 _igvn.replace_node(callprojs.fallthrough_catchproj, ac->in(TypeFunc::Control));
1044
1045 // Set control to top. IGVN will remove the remaining projections
1046 ac->set_req(0, top());
1047 ac->replace_edge(res, top());
1048
1049 // Disconnect src right away: it can help find new
1050 // opportunities for allocation elimination
1051 Node* src = ac->in(ArrayCopyNode::Src);
1052 ac->replace_edge(src, top());
1053 // src can be top at this point if src and dest of the
1054 // arraycopy were the same
1055 if (src->outcnt() == 0 && !src->is_top()) {
1056 _igvn.remove_dead_node(src);
1057 }
1058
1059 _igvn._worklist.push(ac);
1060 } else {
1061 eliminate_card_mark(use);
1062 }
1063 j -= (oc1 - res->outcnt());
1064 }
1065 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
1066 _igvn.remove_dead_node(res);
1067 }
1068
1069 //
1070 // Process other users of allocation's projections
1071 //
1072 if (_resproj != NULL && _resproj->outcnt() != 0) {
1073 // First disconnect stores captured by Initialize node.
1074 // If Initialize node is eliminated first in the following code,
1075 // it will kill such stores and DUIterator_Last will assert.
1076 for (DUIterator_Fast jmax, j = _resproj->fast_outs(jmax); j < jmax; j++) {
1077 Node *use = _resproj->fast_out(j);
1078 if (use->is_AddP()) {
1079 // raw memory addresses used only by the initialization
1080 _igvn.replace_node(use, C->top());
1081 --j; --jmax;
1082 }
1083 }
1084 for (DUIterator_Last jmin, j = _resproj->last_outs(jmin); j >= jmin; ) {
1085 Node *use = _resproj->last_out(j);
1086 uint oc1 = _resproj->outcnt();
1087 if (use->is_Initialize()) {
1088 // Eliminate Initialize node.
1089 InitializeNode *init = use->as_Initialize();
1090 assert(init->outcnt() <= 2, "only a control and memory projection expected");
1091 Node *ctrl_proj = init->proj_out_or_null(TypeFunc::Control);
1092 if (ctrl_proj != NULL) {
1093 assert(init->in(TypeFunc::Control) == _fallthroughcatchproj, "allocation control projection");
1094 _igvn.replace_node(ctrl_proj, _fallthroughcatchproj);
1095 }
1096 Node *mem_proj = init->proj_out_or_null(TypeFunc::Memory);
1097 if (mem_proj != NULL) {
1098 Node *mem = init->in(TypeFunc::Memory);
1099 #ifdef ASSERT
1100 if (mem->is_MergeMem()) {
1101 assert(mem->in(TypeFunc::Memory) == _memproj_fallthrough, "allocation memory projection");
1102 } else {
1103 assert(mem == _memproj_fallthrough, "allocation memory projection");
1104 }
1105 #endif
1106 _igvn.replace_node(mem_proj, mem);
1107 }
1108 } else {
1109 assert(false, "only Initialize or AddP expected");
1110 }
1111 j -= (oc1 - _resproj->outcnt());
1112 }
1113 }
1114 if (_fallthroughcatchproj != NULL) {
1115 _igvn.replace_node(_fallthroughcatchproj, alloc->in(TypeFunc::Control));
1116 }
1117 if (_memproj_fallthrough != NULL) {
1118 _igvn.replace_node(_memproj_fallthrough, alloc->in(TypeFunc::Memory));
1119 }
1120 if (_memproj_catchall != NULL) {
1121 _igvn.replace_node(_memproj_catchall, C->top());
1122 }
1123 if (_ioproj_fallthrough != NULL) {
1124 _igvn.replace_node(_ioproj_fallthrough, alloc->in(TypeFunc::I_O));
1125 }
1126 if (_ioproj_catchall != NULL) {
1127 _igvn.replace_node(_ioproj_catchall, C->top());
1128 }
1129 if (_catchallcatchproj != NULL) {
1130 _igvn.replace_node(_catchallcatchproj, C->top());
1131 }
1132 }
1133
1134 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) {
1135 // Don't do scalar replacement if the frame can be popped by JVMTI:
1136 // if reallocation fails during deoptimization we'll pop all
1137 // interpreter frames for this compiled frame and that won't play
1138 // nice with JVMTI popframe.
1139 if (!EliminateAllocations || JvmtiExport::can_pop_frame() || !alloc->_is_non_escaping) {
1140 return false;
1141 }
1142 Node* klass = alloc->in(AllocateNode::KlassNode);
1143 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr();
1144 Node* res = alloc->result_cast();
1145 // Eliminate boxing allocations which are not used
1146 // regardless scalar replacable status.
1147 bool boxing_alloc = C->eliminate_boxing() &&
1148 tklass->klass()->is_instance_klass() &&
1149 tklass->klass()->as_instance_klass()->is_box_klass();
1150 if (!alloc->_is_scalar_replaceable && (!boxing_alloc || (res != NULL))) {
1151 return false;
1152 }
1153
1154 extract_call_projections(alloc);
1155
1156 GrowableArray <SafePointNode *> safepoints;
1157 if (!can_eliminate_allocation(alloc, safepoints)) {
1158 return false;
1159 }
1160
1161 if (!alloc->_is_scalar_replaceable) {
1162 assert(res == NULL, "sanity");
1163 // We can only eliminate allocation if all debug info references
1164 // are already replaced with SafePointScalarObject because
1165 // we can't search for a fields value without instance_id.
1166 if (safepoints.length() > 0) {
1167 return false;
1168 }
1169 }
1170
1171 if (!scalar_replacement(alloc, safepoints)) {
1172 return false;
1173 }
1174
1175 CompileLog* log = C->log();
1176 if (log != NULL) {
1177 log->head("eliminate_allocation type='%d'",
1178 log->identify(tklass->klass()));
1179 JVMState* p = alloc->jvms();
1180 while (p != NULL) {
1181 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1182 p = p->caller();
1183 }
1184 log->tail("eliminate_allocation");
1185 }
1186
1187 process_users_of_allocation(alloc);
1188
1189 #ifndef PRODUCT
1190 if (PrintEliminateAllocations) {
1191 if (alloc->is_AllocateArray())
1192 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1193 else
1194 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1195 }
1196 #endif
1197
1198 return true;
1199 }
1200
1201 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) {
1202 // EA should remove all uses of non-escaping boxing node.
1203 if (!C->eliminate_boxing() || boxing->proj_out_or_null(TypeFunc::Parms) != NULL) {
1204 return false;
1205 }
1206
1207 assert(boxing->result_cast() == NULL, "unexpected boxing node result");
1208
1209 extract_call_projections(boxing);
1210
1211 const TypeTuple* r = boxing->tf()->range();
1212 assert(r->cnt() > TypeFunc::Parms, "sanity");
1213 const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr();
1214 assert(t != NULL, "sanity");
1215
1216 CompileLog* log = C->log();
1217 if (log != NULL) {
1218 log->head("eliminate_boxing type='%d'",
1219 log->identify(t->klass()));
1220 JVMState* p = boxing->jvms();
1221 while (p != NULL) {
1222 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1223 p = p->caller();
1224 }
1225 log->tail("eliminate_boxing");
1226 }
1227
1228 process_users_of_allocation(boxing);
1229
1230 #ifndef PRODUCT
1231 if (PrintEliminateAllocations) {
1232 tty->print("++++ Eliminated: %d ", boxing->_idx);
1233 boxing->method()->print_short_name(tty);
1234 tty->cr();
1235 }
1236 #endif
1237
1238 return true;
1239 }
1240
1241 //---------------------------set_eden_pointers-------------------------
1242 void PhaseMacroExpand::set_eden_pointers(Node* &eden_top_adr, Node* &eden_end_adr) {
1243 if (UseTLAB) { // Private allocation: load from TLS
1244 Node* thread = transform_later(new ThreadLocalNode());
1245 int tlab_top_offset = in_bytes(JavaThread::tlab_top_offset());
1246 int tlab_end_offset = in_bytes(JavaThread::tlab_end_offset());
1247 eden_top_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_top_offset);
1248 eden_end_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_end_offset);
1249 } else { // Shared allocation: load from globals
1250 CollectedHeap* ch = Universe::heap();
1251 address top_adr = (address)ch->top_addr();
1252 address end_adr = (address)ch->end_addr();
1253 eden_top_adr = makecon(TypeRawPtr::make(top_adr));
1254 eden_end_adr = basic_plus_adr(eden_top_adr, end_adr - top_adr);
1255 }
1256 }
1257
1258
1259 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) {
1260 Node* adr = basic_plus_adr(base, offset);
1261 const TypePtr* adr_type = adr->bottom_type()->is_ptr();
1262 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt, MemNode::unordered);
1263 transform_later(value);
1264 return value;
1265 }
1266
1267
1268 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) {
1269 Node* adr = basic_plus_adr(base, offset);
1270 mem = StoreNode::make(_igvn, ctl, mem, adr, NULL, value, bt, MemNode::unordered);
1271 transform_later(mem);
1272 return mem;
1273 }
1274
1275 //=============================================================================
1276 //
1277 // A L L O C A T I O N
1278 //
1279 // Allocation attempts to be fast in the case of frequent small objects.
1280 // It breaks down like this:
1281 //
1282 // 1) Size in doublewords is computed. This is a constant for objects and
1283 // variable for most arrays. Doubleword units are used to avoid size
1284 // overflow of huge doubleword arrays. We need doublewords in the end for
1285 // rounding.
1286 //
1287 // 2) Size is checked for being 'too large'. Too-large allocations will go
1288 // the slow path into the VM. The slow path can throw any required
1289 // exceptions, and does all the special checks for very large arrays. The
1290 // size test can constant-fold away for objects. For objects with
1291 // finalizers it constant-folds the otherway: you always go slow with
1292 // finalizers.
1293 //
1294 // 3) If NOT using TLABs, this is the contended loop-back point.
1295 // Load-Locked the heap top. If using TLABs normal-load the heap top.
1296 //
1297 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route.
1298 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish
1299 // "size*8" we always enter the VM, where "largish" is a constant picked small
1300 // enough that there's always space between the eden max and 4Gig (old space is
1301 // there so it's quite large) and large enough that the cost of entering the VM
1302 // is dwarfed by the cost to initialize the space.
1303 //
1304 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back
1305 // down. If contended, repeat at step 3. If using TLABs normal-store
1306 // adjusted heap top back down; there is no contention.
1307 //
1308 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark
1309 // fields.
1310 //
1311 // 7) Merge with the slow-path; cast the raw memory pointer to the correct
1312 // oop flavor.
1313 //
1314 //=============================================================================
1315 // FastAllocateSizeLimit value is in DOUBLEWORDS.
1316 // Allocations bigger than this always go the slow route.
1317 // This value must be small enough that allocation attempts that need to
1318 // trigger exceptions go the slow route. Also, it must be small enough so
1319 // that heap_top + size_in_bytes does not wrap around the 4Gig limit.
1320 //=============================================================================j//
1321 // %%% Here is an old comment from parseHelper.cpp; is it outdated?
1322 // The allocator will coalesce int->oop copies away. See comment in
1323 // coalesce.cpp about how this works. It depends critically on the exact
1324 // code shape produced here, so if you are changing this code shape
1325 // make sure the GC info for the heap-top is correct in and around the
1326 // slow-path call.
1327 //
1328
1329 void PhaseMacroExpand::expand_allocate_common(
1330 AllocateNode* alloc, // allocation node to be expanded
1331 Node* length, // array length for an array allocation
1332 const TypeFunc* slow_call_type, // Type of slow call
1333 address slow_call_address // Address of slow call
1334 )
1335 {
1336
1337 Node* ctrl = alloc->in(TypeFunc::Control);
1338 Node* mem = alloc->in(TypeFunc::Memory);
1339 Node* i_o = alloc->in(TypeFunc::I_O);
1340 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize);
1341 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1342 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest);
1343
1344 assert(ctrl != NULL, "must have control");
1345 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results.
1346 // they will not be used if "always_slow" is set
1347 enum { slow_result_path = 1, fast_result_path = 2 };
1348 Node *result_region = NULL;
1349 Node *result_phi_rawmem = NULL;
1350 Node *result_phi_rawoop = NULL;
1351 Node *result_phi_i_o = NULL;
1352
1353 // The initial slow comparison is a size check, the comparison
1354 // we want to do is a BoolTest::gt
1355 bool always_slow = false;
1356 int tv = _igvn.find_int_con(initial_slow_test, -1);
1357 if (tv >= 0) {
1358 always_slow = (tv == 1);
1359 initial_slow_test = NULL;
1360 } else {
1361 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn);
1362 }
1363
1364 if (C->env()->dtrace_alloc_probes() ||
1365 (!UseTLAB && !Universe::heap()->supports_inline_contig_alloc())) {
1366 // Force slow-path allocation
1367 always_slow = true;
1368 initial_slow_test = NULL;
1369 }
1370
1371
1372 enum { too_big_or_final_path = 1, need_gc_path = 2 };
1373 Node *slow_region = NULL;
1374 Node *toobig_false = ctrl;
1375
1376 assert (initial_slow_test == NULL || !always_slow, "arguments must be consistent");
1377 // generate the initial test if necessary
1378 if (initial_slow_test != NULL ) {
1379 slow_region = new RegionNode(3);
1380
1381 // Now make the initial failure test. Usually a too-big test but
1382 // might be a TRUE for finalizers or a fancy class check for
1383 // newInstance0.
1384 IfNode *toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
1385 transform_later(toobig_iff);
1386 // Plug the failing-too-big test into the slow-path region
1387 Node *toobig_true = new IfTrueNode( toobig_iff );
1388 transform_later(toobig_true);
1389 slow_region ->init_req( too_big_or_final_path, toobig_true );
1390 toobig_false = new IfFalseNode( toobig_iff );
1391 transform_later(toobig_false);
1392 } else { // No initial test, just fall into next case
1393 toobig_false = ctrl;
1394 debug_only(slow_region = NodeSentinel);
1395 }
1396
1397 Node *slow_mem = mem; // save the current memory state for slow path
1398 // generate the fast allocation code unless we know that the initial test will always go slow
1399 if (!always_slow) {
1400 // Fast path modifies only raw memory.
1401 if (mem->is_MergeMem()) {
1402 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1403 }
1404
1405 Node* eden_top_adr;
1406 Node* eden_end_adr;
1407
1408 set_eden_pointers(eden_top_adr, eden_end_adr);
1409
1410 // Load Eden::end. Loop invariant and hoisted.
1411 //
1412 // Note: We set the control input on "eden_end" and "old_eden_top" when using
1413 // a TLAB to work around a bug where these values were being moved across
1414 // a safepoint. These are not oops, so they cannot be include in the oop
1415 // map, but they can be changed by a GC. The proper way to fix this would
1416 // be to set the raw memory state when generating a SafepointNode. However
1417 // this will require extensive changes to the loop optimization in order to
1418 // prevent a degradation of the optimization.
1419 // See comment in memnode.hpp, around line 227 in class LoadPNode.
1420 Node *eden_end = make_load(ctrl, mem, eden_end_adr, 0, TypeRawPtr::BOTTOM, T_ADDRESS);
1421
1422 // allocate the Region and Phi nodes for the result
1423 result_region = new RegionNode(3);
1424 result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1425 result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM);
1426 result_phi_i_o = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch
1427
1428 // We need a Region for the loop-back contended case.
1429 enum { fall_in_path = 1, contended_loopback_path = 2 };
1430 Node *contended_region;
1431 Node *contended_phi_rawmem;
1432 if (UseTLAB) {
1433 contended_region = toobig_false;
1434 contended_phi_rawmem = mem;
1435 } else {
1436 contended_region = new RegionNode(3);
1437 contended_phi_rawmem = new PhiNode(contended_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1438 // Now handle the passing-too-big test. We fall into the contended
1439 // loop-back merge point.
1440 contended_region ->init_req(fall_in_path, toobig_false);
1441 contended_phi_rawmem->init_req(fall_in_path, mem);
1442 transform_later(contended_region);
1443 transform_later(contended_phi_rawmem);
1444 }
1445
1446 // Load(-locked) the heap top.
1447 // See note above concerning the control input when using a TLAB
1448 Node *old_eden_top = UseTLAB
1449 ? new LoadPNode (ctrl, contended_phi_rawmem, eden_top_adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, MemNode::unordered)
1450 : new LoadPLockedNode(contended_region, contended_phi_rawmem, eden_top_adr, MemNode::acquire);
1451
1452 transform_later(old_eden_top);
1453 // Add to heap top to get a new heap top
1454 Node *new_eden_top = new AddPNode(top(), old_eden_top, size_in_bytes);
1455 transform_later(new_eden_top);
1456 // Check for needing a GC; compare against heap end
1457 Node *needgc_cmp = new CmpPNode(new_eden_top, eden_end);
1458 transform_later(needgc_cmp);
1459 Node *needgc_bol = new BoolNode(needgc_cmp, BoolTest::ge);
1460 transform_later(needgc_bol);
1461 IfNode *needgc_iff = new IfNode(contended_region, needgc_bol, PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN);
1462 transform_later(needgc_iff);
1463
1464 // Plug the failing-heap-space-need-gc test into the slow-path region
1465 Node *needgc_true = new IfTrueNode(needgc_iff);
1466 transform_later(needgc_true);
1467 if (initial_slow_test) {
1468 slow_region->init_req(need_gc_path, needgc_true);
1469 // This completes all paths into the slow merge point
1470 transform_later(slow_region);
1471 } else { // No initial slow path needed!
1472 // Just fall from the need-GC path straight into the VM call.
1473 slow_region = needgc_true;
1474 }
1475 // No need for a GC. Setup for the Store-Conditional
1476 Node *needgc_false = new IfFalseNode(needgc_iff);
1477 transform_later(needgc_false);
1478
1479 // Grab regular I/O before optional prefetch may change it.
1480 // Slow-path does no I/O so just set it to the original I/O.
1481 result_phi_i_o->init_req(slow_result_path, i_o);
1482
1483 i_o = prefetch_allocation(i_o, needgc_false, contended_phi_rawmem,
1484 old_eden_top, new_eden_top, length);
1485
1486 // Name successful fast-path variables
1487 Node* fast_oop = old_eden_top;
1488 Node* fast_oop_ctrl;
1489 Node* fast_oop_rawmem;
1490
1491 // Store (-conditional) the modified eden top back down.
1492 // StorePConditional produces flags for a test PLUS a modified raw
1493 // memory state.
1494 if (UseTLAB) {
1495 Node* store_eden_top =
1496 new StorePNode(needgc_false, contended_phi_rawmem, eden_top_adr,
1497 TypeRawPtr::BOTTOM, new_eden_top, MemNode::unordered);
1498 transform_later(store_eden_top);
1499 fast_oop_ctrl = needgc_false; // No contention, so this is the fast path
1500 fast_oop_rawmem = store_eden_top;
1501 } else {
1502 Node* store_eden_top =
1503 new StorePConditionalNode(needgc_false, contended_phi_rawmem, eden_top_adr,
1504 new_eden_top, fast_oop/*old_eden_top*/);
1505 transform_later(store_eden_top);
1506 Node *contention_check = new BoolNode(store_eden_top, BoolTest::ne);
1507 transform_later(contention_check);
1508 store_eden_top = new SCMemProjNode(store_eden_top);
1509 transform_later(store_eden_top);
1510
1511 // If not using TLABs, check to see if there was contention.
1512 IfNode *contention_iff = new IfNode (needgc_false, contention_check, PROB_MIN, COUNT_UNKNOWN);
1513 transform_later(contention_iff);
1514 Node *contention_true = new IfTrueNode(contention_iff);
1515 transform_later(contention_true);
1516 // If contention, loopback and try again.
1517 contended_region->init_req(contended_loopback_path, contention_true);
1518 contended_phi_rawmem->init_req(contended_loopback_path, store_eden_top);
1519
1520 // Fast-path succeeded with no contention!
1521 Node *contention_false = new IfFalseNode(contention_iff);
1522 transform_later(contention_false);
1523 fast_oop_ctrl = contention_false;
1524
1525 // Bump total allocated bytes for this thread
1526 Node* thread = new ThreadLocalNode();
1527 transform_later(thread);
1528 Node* alloc_bytes_adr = basic_plus_adr(top()/*not oop*/, thread,
1529 in_bytes(JavaThread::allocated_bytes_offset()));
1530 Node* alloc_bytes = make_load(fast_oop_ctrl, store_eden_top, alloc_bytes_adr,
1531 0, TypeLong::LONG, T_LONG);
1532 #ifdef _LP64
1533 Node* alloc_size = size_in_bytes;
1534 #else
1535 Node* alloc_size = new ConvI2LNode(size_in_bytes);
1536 transform_later(alloc_size);
1537 #endif
1538 Node* new_alloc_bytes = new AddLNode(alloc_bytes, alloc_size);
1539 transform_later(new_alloc_bytes);
1540 fast_oop_rawmem = make_store(fast_oop_ctrl, store_eden_top, alloc_bytes_adr,
1541 0, new_alloc_bytes, T_LONG);
1542 }
1543
1544 InitializeNode* init = alloc->initialization();
1545 fast_oop_rawmem = initialize_object(alloc,
1546 fast_oop_ctrl, fast_oop_rawmem, fast_oop,
1547 klass_node, length, size_in_bytes);
1548
1549 // If initialization is performed by an array copy, any required
1550 // MemBarStoreStore was already added. If the object does not
1551 // escape no need for a MemBarStoreStore. If the object does not
1552 // escape in its initializer and memory barrier (MemBarStoreStore or
1553 // stronger) is already added at exit of initializer, also no need
1554 // for a MemBarStoreStore. Otherwise we need a MemBarStoreStore
1555 // so that stores that initialize this object can't be reordered
1556 // with a subsequent store that makes this object accessible by
1557 // other threads.
1558 // Other threads include java threads and JVM internal threads
1559 // (for example concurrent GC threads). Current concurrent GC
1560 // implementation: CMS and G1 will not scan newly created object,
1561 // so it's safe to skip storestore barrier when allocation does
1562 // not escape.
1563 if (!alloc->does_not_escape_thread() &&
1564 !alloc->is_allocation_MemBar_redundant() &&
1565 (init == NULL || !init->is_complete_with_arraycopy())) {
1566 if (init == NULL || init->req() < InitializeNode::RawStores) {
1567 // No InitializeNode or no stores captured by zeroing
1568 // elimination. Simply add the MemBarStoreStore after object
1569 // initialization.
1570 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1571 transform_later(mb);
1572
1573 mb->init_req(TypeFunc::Memory, fast_oop_rawmem);
1574 mb->init_req(TypeFunc::Control, fast_oop_ctrl);
1575 fast_oop_ctrl = new ProjNode(mb,TypeFunc::Control);
1576 transform_later(fast_oop_ctrl);
1577 fast_oop_rawmem = new ProjNode(mb,TypeFunc::Memory);
1578 transform_later(fast_oop_rawmem);
1579 } else {
1580 // Add the MemBarStoreStore after the InitializeNode so that
1581 // all stores performing the initialization that were moved
1582 // before the InitializeNode happen before the storestore
1583 // barrier.
1584
1585 Node* init_ctrl = init->proj_out_or_null(TypeFunc::Control);
1586 Node* init_mem = init->proj_out_or_null(TypeFunc::Memory);
1587
1588 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1589 transform_later(mb);
1590
1591 Node* ctrl = new ProjNode(init,TypeFunc::Control);
1592 transform_later(ctrl);
1593 Node* mem = new ProjNode(init,TypeFunc::Memory);
1594 transform_later(mem);
1595
1596 // The MemBarStoreStore depends on control and memory coming
1597 // from the InitializeNode
1598 mb->init_req(TypeFunc::Memory, mem);
1599 mb->init_req(TypeFunc::Control, ctrl);
1600
1601 ctrl = new ProjNode(mb,TypeFunc::Control);
1602 transform_later(ctrl);
1603 mem = new ProjNode(mb,TypeFunc::Memory);
1604 transform_later(mem);
1605
1606 // All nodes that depended on the InitializeNode for control
1607 // and memory must now depend on the MemBarNode that itself
1608 // depends on the InitializeNode
1609 if (init_ctrl != NULL) {
1610 _igvn.replace_node(init_ctrl, ctrl);
1611 }
1612 if (init_mem != NULL) {
1613 _igvn.replace_node(init_mem, mem);
1614 }
1615 }
1616 }
1617
1618 if (C->env()->dtrace_extended_probes()) {
1619 // Slow-path call
1620 int size = TypeFunc::Parms + 2;
1621 CallLeafNode *call = new CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(),
1622 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_object_alloc_base),
1623 "dtrace_object_alloc",
1624 TypeRawPtr::BOTTOM);
1625
1626 // Get base of thread-local storage area
1627 Node* thread = new ThreadLocalNode();
1628 transform_later(thread);
1629
1630 call->init_req(TypeFunc::Parms+0, thread);
1631 call->init_req(TypeFunc::Parms+1, fast_oop);
1632 call->init_req(TypeFunc::Control, fast_oop_ctrl);
1633 call->init_req(TypeFunc::I_O , top()); // does no i/o
1634 call->init_req(TypeFunc::Memory , fast_oop_rawmem);
1635 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1636 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1637 transform_later(call);
1638 fast_oop_ctrl = new ProjNode(call,TypeFunc::Control);
1639 transform_later(fast_oop_ctrl);
1640 fast_oop_rawmem = new ProjNode(call,TypeFunc::Memory);
1641 transform_later(fast_oop_rawmem);
1642 }
1643
1644 // Plug in the successful fast-path into the result merge point
1645 result_region ->init_req(fast_result_path, fast_oop_ctrl);
1646 result_phi_rawoop->init_req(fast_result_path, fast_oop);
1647 result_phi_i_o ->init_req(fast_result_path, i_o);
1648 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem);
1649 } else {
1650 slow_region = ctrl;
1651 result_phi_i_o = i_o; // Rename it to use in the following code.
1652 }
1653
1654 // Generate slow-path call
1655 CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address,
1656 OptoRuntime::stub_name(slow_call_address),
1657 alloc->jvms()->bci(),
1658 TypePtr::BOTTOM);
1659 call->init_req( TypeFunc::Control, slow_region );
1660 call->init_req( TypeFunc::I_O , top() ) ; // does no i/o
1661 call->init_req( TypeFunc::Memory , slow_mem ); // may gc ptrs
1662 call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) );
1663 call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) );
1664
1665 call->init_req(TypeFunc::Parms+0, klass_node);
1666 if (length != NULL) {
1667 call->init_req(TypeFunc::Parms+1, length);
1668 }
1669
1670 // Copy debug information and adjust JVMState information, then replace
1671 // allocate node with the call
1672 copy_call_debug_info((CallNode *) alloc, call);
1673 if (!always_slow) {
1674 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1675 } else {
1676 // Hook i_o projection to avoid its elimination during allocation
1677 // replacement (when only a slow call is generated).
1678 call->set_req(TypeFunc::I_O, result_phi_i_o);
1679 }
1680 _igvn.replace_node(alloc, call);
1681 transform_later(call);
1682
1683 // Identify the output projections from the allocate node and
1684 // adjust any references to them.
1685 // The control and io projections look like:
1686 //
1687 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl)
1688 // Allocate Catch
1689 // ^---Proj(io) <-------+ ^---CatchProj(io)
1690 //
1691 // We are interested in the CatchProj nodes.
1692 //
1693 extract_call_projections(call);
1694
1695 // An allocate node has separate memory projections for the uses on
1696 // the control and i_o paths. Replace the control memory projection with
1697 // result_phi_rawmem (unless we are only generating a slow call when
1698 // both memory projections are combined)
1699 if (!always_slow && _memproj_fallthrough != NULL) {
1700 for (DUIterator_Fast imax, i = _memproj_fallthrough->fast_outs(imax); i < imax; i++) {
1701 Node *use = _memproj_fallthrough->fast_out(i);
1702 _igvn.rehash_node_delayed(use);
1703 imax -= replace_input(use, _memproj_fallthrough, result_phi_rawmem);
1704 // back up iterator
1705 --i;
1706 }
1707 }
1708 // Now change uses of _memproj_catchall to use _memproj_fallthrough and delete
1709 // _memproj_catchall so we end up with a call that has only 1 memory projection.
1710 if (_memproj_catchall != NULL ) {
1711 if (_memproj_fallthrough == NULL) {
1712 _memproj_fallthrough = new ProjNode(call, TypeFunc::Memory);
1713 transform_later(_memproj_fallthrough);
1714 }
1715 for (DUIterator_Fast imax, i = _memproj_catchall->fast_outs(imax); i < imax; i++) {
1716 Node *use = _memproj_catchall->fast_out(i);
1717 _igvn.rehash_node_delayed(use);
1718 imax -= replace_input(use, _memproj_catchall, _memproj_fallthrough);
1719 // back up iterator
1720 --i;
1721 }
1722 assert(_memproj_catchall->outcnt() == 0, "all uses must be deleted");
1723 _igvn.remove_dead_node(_memproj_catchall);
1724 }
1725
1726 // An allocate node has separate i_o projections for the uses on the control
1727 // and i_o paths. Always replace the control i_o projection with result i_o
1728 // otherwise incoming i_o become dead when only a slow call is generated
1729 // (it is different from memory projections where both projections are
1730 // combined in such case).
1731 if (_ioproj_fallthrough != NULL) {
1732 for (DUIterator_Fast imax, i = _ioproj_fallthrough->fast_outs(imax); i < imax; i++) {
1733 Node *use = _ioproj_fallthrough->fast_out(i);
1734 _igvn.rehash_node_delayed(use);
1735 imax -= replace_input(use, _ioproj_fallthrough, result_phi_i_o);
1736 // back up iterator
1737 --i;
1738 }
1739 }
1740 // Now change uses of _ioproj_catchall to use _ioproj_fallthrough and delete
1741 // _ioproj_catchall so we end up with a call that has only 1 i_o projection.
1742 if (_ioproj_catchall != NULL ) {
1743 if (_ioproj_fallthrough == NULL) {
1744 _ioproj_fallthrough = new ProjNode(call, TypeFunc::I_O);
1745 transform_later(_ioproj_fallthrough);
1746 }
1747 for (DUIterator_Fast imax, i = _ioproj_catchall->fast_outs(imax); i < imax; i++) {
1748 Node *use = _ioproj_catchall->fast_out(i);
1749 _igvn.rehash_node_delayed(use);
1750 imax -= replace_input(use, _ioproj_catchall, _ioproj_fallthrough);
1751 // back up iterator
1752 --i;
1753 }
1754 assert(_ioproj_catchall->outcnt() == 0, "all uses must be deleted");
1755 _igvn.remove_dead_node(_ioproj_catchall);
1756 }
1757
1758 // if we generated only a slow call, we are done
1759 if (always_slow) {
1760 // Now we can unhook i_o.
1761 if (result_phi_i_o->outcnt() > 1) {
1762 call->set_req(TypeFunc::I_O, top());
1763 } else {
1764 assert(result_phi_i_o->unique_ctrl_out() == call, "");
1765 // Case of new array with negative size known during compilation.
1766 // AllocateArrayNode::Ideal() optimization disconnect unreachable
1767 // following code since call to runtime will throw exception.
1768 // As result there will be no users of i_o after the call.
1769 // Leave i_o attached to this call to avoid problems in preceding graph.
1770 }
1771 return;
1772 }
1773
1774
1775 if (_fallthroughcatchproj != NULL) {
1776 ctrl = _fallthroughcatchproj->clone();
1777 transform_later(ctrl);
1778 _igvn.replace_node(_fallthroughcatchproj, result_region);
1779 } else {
1780 ctrl = top();
1781 }
1782 Node *slow_result;
1783 if (_resproj == NULL) {
1784 // no uses of the allocation result
1785 slow_result = top();
1786 } else {
1787 slow_result = _resproj->clone();
1788 transform_later(slow_result);
1789 _igvn.replace_node(_resproj, result_phi_rawoop);
1790 }
1791
1792 // Plug slow-path into result merge point
1793 result_region ->init_req( slow_result_path, ctrl );
1794 result_phi_rawoop->init_req( slow_result_path, slow_result);
1795 result_phi_rawmem->init_req( slow_result_path, _memproj_fallthrough );
1796 transform_later(result_region);
1797 transform_later(result_phi_rawoop);
1798 transform_later(result_phi_rawmem);
1799 transform_later(result_phi_i_o);
1800 // This completes all paths into the result merge point
1801 }
1802
1803
1804 // Helper for PhaseMacroExpand::expand_allocate_common.
1805 // Initializes the newly-allocated storage.
1806 Node*
1807 PhaseMacroExpand::initialize_object(AllocateNode* alloc,
1808 Node* control, Node* rawmem, Node* object,
1809 Node* klass_node, Node* length,
1810 Node* size_in_bytes) {
1811 InitializeNode* init = alloc->initialization();
1812 // Store the klass & mark bits
1813 Node* mark_node = NULL;
1814 // For now only enable fast locking for non-array types
1815 if (UseBiasedLocking && (length == NULL)) {
1816 mark_node = make_load(control, rawmem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
1817 } else {
1818 mark_node = makecon(TypeRawPtr::make((address)markOopDesc::prototype()));
1819 }
1820 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS);
1821
1822 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
1823 int header_size = alloc->minimum_header_size(); // conservatively small
1824
1825 // Array length
1826 if (length != NULL) { // Arrays need length field
1827 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
1828 // conservatively small header size:
1829 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1830 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
1831 if (k->is_array_klass()) // we know the exact header size in most cases:
1832 header_size = Klass::layout_helper_header_size(k->layout_helper());
1833 }
1834
1835 // Clear the object body, if necessary.
1836 if (init == NULL) {
1837 // The init has somehow disappeared; be cautious and clear everything.
1838 //
1839 // This can happen if a node is allocated but an uncommon trap occurs
1840 // immediately. In this case, the Initialize gets associated with the
1841 // trap, and may be placed in a different (outer) loop, if the Allocate
1842 // is in a loop. If (this is rare) the inner loop gets unrolled, then
1843 // there can be two Allocates to one Initialize. The answer in all these
1844 // edge cases is safety first. It is always safe to clear immediately
1845 // within an Allocate, and then (maybe or maybe not) clear some more later.
1846 if (!(UseTLAB && ZeroTLAB)) {
1847 rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
1848 header_size, size_in_bytes,
1849 &_igvn);
1850 }
1851 } else {
1852 if (!init->is_complete()) {
1853 // Try to win by zeroing only what the init does not store.
1854 // We can also try to do some peephole optimizations,
1855 // such as combining some adjacent subword stores.
1856 rawmem = init->complete_stores(control, rawmem, object,
1857 header_size, size_in_bytes, &_igvn);
1858 }
1859 // We have no more use for this link, since the AllocateNode goes away:
1860 init->set_req(InitializeNode::RawAddress, top());
1861 // (If we keep the link, it just confuses the register allocator,
1862 // who thinks he sees a real use of the address by the membar.)
1863 }
1864
1865 return rawmem;
1866 }
1867
1868 // Generate prefetch instructions for next allocations.
1869 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false,
1870 Node*& contended_phi_rawmem,
1871 Node* old_eden_top, Node* new_eden_top,
1872 Node* length) {
1873 enum { fall_in_path = 1, pf_path = 2 };
1874 if( UseTLAB && AllocatePrefetchStyle == 2 ) {
1875 // Generate prefetch allocation with watermark check.
1876 // As an allocation hits the watermark, we will prefetch starting
1877 // at a "distance" away from watermark.
1878
1879 Node *pf_region = new RegionNode(3);
1880 Node *pf_phi_rawmem = new PhiNode( pf_region, Type::MEMORY,
1881 TypeRawPtr::BOTTOM );
1882 // I/O is used for Prefetch
1883 Node *pf_phi_abio = new PhiNode( pf_region, Type::ABIO );
1884
1885 Node *thread = new ThreadLocalNode();
1886 transform_later(thread);
1887
1888 Node *eden_pf_adr = new AddPNode( top()/*not oop*/, thread,
1889 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) );
1890 transform_later(eden_pf_adr);
1891
1892 Node *old_pf_wm = new LoadPNode(needgc_false,
1893 contended_phi_rawmem, eden_pf_adr,
1894 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM,
1895 MemNode::unordered);
1896 transform_later(old_pf_wm);
1897
1898 // check against new_eden_top
1899 Node *need_pf_cmp = new CmpPNode( new_eden_top, old_pf_wm );
1900 transform_later(need_pf_cmp);
1901 Node *need_pf_bol = new BoolNode( need_pf_cmp, BoolTest::ge );
1902 transform_later(need_pf_bol);
1903 IfNode *need_pf_iff = new IfNode( needgc_false, need_pf_bol,
1904 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN );
1905 transform_later(need_pf_iff);
1906
1907 // true node, add prefetchdistance
1908 Node *need_pf_true = new IfTrueNode( need_pf_iff );
1909 transform_later(need_pf_true);
1910
1911 Node *need_pf_false = new IfFalseNode( need_pf_iff );
1912 transform_later(need_pf_false);
1913
1914 Node *new_pf_wmt = new AddPNode( top(), old_pf_wm,
1915 _igvn.MakeConX(AllocatePrefetchDistance) );
1916 transform_later(new_pf_wmt );
1917 new_pf_wmt->set_req(0, need_pf_true);
1918
1919 Node *store_new_wmt = new StorePNode(need_pf_true,
1920 contended_phi_rawmem, eden_pf_adr,
1921 TypeRawPtr::BOTTOM, new_pf_wmt,
1922 MemNode::unordered);
1923 transform_later(store_new_wmt);
1924
1925 // adding prefetches
1926 pf_phi_abio->init_req( fall_in_path, i_o );
1927
1928 Node *prefetch_adr;
1929 Node *prefetch;
1930 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1931 uint step_size = AllocatePrefetchStepSize;
1932 uint distance = 0;
1933
1934 for ( uint i = 0; i < lines; i++ ) {
1935 prefetch_adr = new AddPNode( old_pf_wm, new_pf_wmt,
1936 _igvn.MakeConX(distance) );
1937 transform_later(prefetch_adr);
1938 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
1939 transform_later(prefetch);
1940 distance += step_size;
1941 i_o = prefetch;
1942 }
1943 pf_phi_abio->set_req( pf_path, i_o );
1944
1945 pf_region->init_req( fall_in_path, need_pf_false );
1946 pf_region->init_req( pf_path, need_pf_true );
1947
1948 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem );
1949 pf_phi_rawmem->init_req( pf_path, store_new_wmt );
1950
1951 transform_later(pf_region);
1952 transform_later(pf_phi_rawmem);
1953 transform_later(pf_phi_abio);
1954
1955 needgc_false = pf_region;
1956 contended_phi_rawmem = pf_phi_rawmem;
1957 i_o = pf_phi_abio;
1958 } else if( UseTLAB && AllocatePrefetchStyle == 3 ) {
1959 // Insert a prefetch instruction for each allocation.
1960 // This code is used to generate 1 prefetch instruction per cache line.
1961
1962 // Generate several prefetch instructions.
1963 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1964 uint step_size = AllocatePrefetchStepSize;
1965 uint distance = AllocatePrefetchDistance;
1966
1967 // Next cache address.
1968 Node *cache_adr = new AddPNode(old_eden_top, old_eden_top,
1969 _igvn.MakeConX(step_size + distance));
1970 transform_later(cache_adr);
1971 cache_adr = new CastP2XNode(needgc_false, cache_adr);
1972 transform_later(cache_adr);
1973 // Address is aligned to execute prefetch to the beginning of cache line size
1974 // (it is important when BIS instruction is used on SPARC as prefetch).
1975 Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1));
1976 cache_adr = new AndXNode(cache_adr, mask);
1977 transform_later(cache_adr);
1978 cache_adr = new CastX2PNode(cache_adr);
1979 transform_later(cache_adr);
1980
1981 // Prefetch
1982 Node *prefetch = new PrefetchAllocationNode( contended_phi_rawmem, cache_adr );
1983 prefetch->set_req(0, needgc_false);
1984 transform_later(prefetch);
1985 contended_phi_rawmem = prefetch;
1986 Node *prefetch_adr;
1987 distance = step_size;
1988 for ( uint i = 1; i < lines; i++ ) {
1989 prefetch_adr = new AddPNode( cache_adr, cache_adr,
1990 _igvn.MakeConX(distance) );
1991 transform_later(prefetch_adr);
1992 prefetch = new PrefetchAllocationNode( contended_phi_rawmem, prefetch_adr );
1993 transform_later(prefetch);
1994 distance += step_size;
1995 contended_phi_rawmem = prefetch;
1996 }
1997 } else if( AllocatePrefetchStyle > 0 ) {
1998 // Insert a prefetch for each allocation only on the fast-path
1999 Node *prefetch_adr;
2000 Node *prefetch;
2001 // Generate several prefetch instructions.
2002 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
2003 uint step_size = AllocatePrefetchStepSize;
2004 uint distance = AllocatePrefetchDistance;
2005 for ( uint i = 0; i < lines; i++ ) {
2006 prefetch_adr = new AddPNode( old_eden_top, new_eden_top,
2007 _igvn.MakeConX(distance) );
2008 transform_later(prefetch_adr);
2009 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
2010 // Do not let it float too high, since if eden_top == eden_end,
2011 // both might be null.
2012 if( i == 0 ) { // Set control for first prefetch, next follows it
2013 prefetch->init_req(0, needgc_false);
2014 }
2015 transform_later(prefetch);
2016 distance += step_size;
2017 i_o = prefetch;
2018 }
2019 }
2020 return i_o;
2021 }
2022
2023
2024 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) {
2025 expand_allocate_common(alloc, NULL,
2026 OptoRuntime::new_instance_Type(),
2027 OptoRuntime::new_instance_Java());
2028 }
2029
2030 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) {
2031 Node* length = alloc->in(AllocateNode::ALength);
2032 InitializeNode* init = alloc->initialization();
2033 Node* klass_node = alloc->in(AllocateNode::KlassNode);
2034 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
2035 address slow_call_address; // Address of slow call
2036 if (init != NULL && init->is_complete_with_arraycopy() &&
2037 k->is_type_array_klass()) {
2038 // Don't zero type array during slow allocation in VM since
2039 // it will be initialized later by arraycopy in compiled code.
2040 slow_call_address = OptoRuntime::new_array_nozero_Java();
2041 } else {
2042 slow_call_address = OptoRuntime::new_array_Java();
2043 }
2044 expand_allocate_common(alloc, length,
2045 OptoRuntime::new_array_Type(),
2046 slow_call_address);
2047 }
2048
2049 //-------------------mark_eliminated_box----------------------------------
2050 //
2051 // During EA obj may point to several objects but after few ideal graph
2052 // transformations (CCP) it may point to only one non escaping object
2053 // (but still using phi), corresponding locks and unlocks will be marked
2054 // for elimination. Later obj could be replaced with a new node (new phi)
2055 // and which does not have escape information. And later after some graph
2056 // reshape other locks and unlocks (which were not marked for elimination
2057 // before) are connected to this new obj (phi) but they still will not be
2058 // marked for elimination since new obj has no escape information.
2059 // Mark all associated (same box and obj) lock and unlock nodes for
2060 // elimination if some of them marked already.
2061 void PhaseMacroExpand::mark_eliminated_box(Node* oldbox, Node* obj) {
2062 if (oldbox->as_BoxLock()->is_eliminated())
2063 return; // This BoxLock node was processed already.
2064
2065 // New implementation (EliminateNestedLocks) has separate BoxLock
2066 // node for each locked region so mark all associated locks/unlocks as
2067 // eliminated even if different objects are referenced in one locked region
2068 // (for example, OSR compilation of nested loop inside locked scope).
2069 if (EliminateNestedLocks ||
2070 oldbox->as_BoxLock()->is_simple_lock_region(NULL, obj)) {
2071 // Box is used only in one lock region. Mark this box as eliminated.
2072 _igvn.hash_delete(oldbox);
2073 oldbox->as_BoxLock()->set_eliminated(); // This changes box's hash value
2074 _igvn.hash_insert(oldbox);
2075
2076 for (uint i = 0; i < oldbox->outcnt(); i++) {
2077 Node* u = oldbox->raw_out(i);
2078 if (u->is_AbstractLock() && !u->as_AbstractLock()->is_non_esc_obj()) {
2079 AbstractLockNode* alock = u->as_AbstractLock();
2080 // Check lock's box since box could be referenced by Lock's debug info.
2081 if (alock->box_node() == oldbox) {
2082 // Mark eliminated all related locks and unlocks.
2083 #ifdef ASSERT
2084 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc4");
2085 #endif
2086 alock->set_non_esc_obj();
2087 }
2088 }
2089 }
2090 return;
2091 }
2092
2093 // Create new "eliminated" BoxLock node and use it in monitor debug info
2094 // instead of oldbox for the same object.
2095 BoxLockNode* newbox = oldbox->clone()->as_BoxLock();
2096
2097 // Note: BoxLock node is marked eliminated only here and it is used
2098 // to indicate that all associated lock and unlock nodes are marked
2099 // for elimination.
2100 newbox->set_eliminated();
2101 transform_later(newbox);
2102
2103 // Replace old box node with new box for all users of the same object.
2104 for (uint i = 0; i < oldbox->outcnt();) {
2105 bool next_edge = true;
2106
2107 Node* u = oldbox->raw_out(i);
2108 if (u->is_AbstractLock()) {
2109 AbstractLockNode* alock = u->as_AbstractLock();
2110 if (alock->box_node() == oldbox && alock->obj_node()->eqv_uncast(obj)) {
2111 // Replace Box and mark eliminated all related locks and unlocks.
2112 #ifdef ASSERT
2113 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc5");
2114 #endif
2115 alock->set_non_esc_obj();
2116 _igvn.rehash_node_delayed(alock);
2117 alock->set_box_node(newbox);
2118 next_edge = false;
2119 }
2120 }
2121 if (u->is_FastLock() && u->as_FastLock()->obj_node()->eqv_uncast(obj)) {
2122 FastLockNode* flock = u->as_FastLock();
2123 assert(flock->box_node() == oldbox, "sanity");
2124 _igvn.rehash_node_delayed(flock);
2125 flock->set_box_node(newbox);
2126 next_edge = false;
2127 }
2128
2129 // Replace old box in monitor debug info.
2130 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) {
2131 SafePointNode* sfn = u->as_SafePoint();
2132 JVMState* youngest_jvms = sfn->jvms();
2133 int max_depth = youngest_jvms->depth();
2134 for (int depth = 1; depth <= max_depth; depth++) {
2135 JVMState* jvms = youngest_jvms->of_depth(depth);
2136 int num_mon = jvms->nof_monitors();
2137 // Loop over monitors
2138 for (int idx = 0; idx < num_mon; idx++) {
2139 Node* obj_node = sfn->monitor_obj(jvms, idx);
2140 Node* box_node = sfn->monitor_box(jvms, idx);
2141 if (box_node == oldbox && obj_node->eqv_uncast(obj)) {
2142 int j = jvms->monitor_box_offset(idx);
2143 _igvn.replace_input_of(u, j, newbox);
2144 next_edge = false;
2145 }
2146 }
2147 }
2148 }
2149 if (next_edge) i++;
2150 }
2151 }
2152
2153 //-----------------------mark_eliminated_locking_nodes-----------------------
2154 void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) {
2155 if (EliminateNestedLocks) {
2156 if (alock->is_nested()) {
2157 assert(alock->box_node()->as_BoxLock()->is_eliminated(), "sanity");
2158 return;
2159 } else if (!alock->is_non_esc_obj()) { // Not eliminated or coarsened
2160 // Only Lock node has JVMState needed here.
2161 // Not that preceding claim is documented anywhere else.
2162 if (alock->jvms() != NULL) {
2163 if (alock->as_Lock()->is_nested_lock_region()) {
2164 // Mark eliminated related nested locks and unlocks.
2165 Node* obj = alock->obj_node();
2166 BoxLockNode* box_node = alock->box_node()->as_BoxLock();
2167 assert(!box_node->is_eliminated(), "should not be marked yet");
2168 // Note: BoxLock node is marked eliminated only here
2169 // and it is used to indicate that all associated lock
2170 // and unlock nodes are marked for elimination.
2171 box_node->set_eliminated(); // Box's hash is always NO_HASH here
2172 for (uint i = 0; i < box_node->outcnt(); i++) {
2173 Node* u = box_node->raw_out(i);
2174 if (u->is_AbstractLock()) {
2175 alock = u->as_AbstractLock();
2176 if (alock->box_node() == box_node) {
2177 // Verify that this Box is referenced only by related locks.
2178 assert(alock->obj_node()->eqv_uncast(obj), "");
2179 // Mark all related locks and unlocks.
2180 #ifdef ASSERT
2181 alock->log_lock_optimization(C, "eliminate_lock_set_nested");
2182 #endif
2183 alock->set_nested();
2184 }
2185 }
2186 }
2187 } else {
2188 #ifdef ASSERT
2189 alock->log_lock_optimization(C, "eliminate_lock_NOT_nested_lock_region");
2190 if (C->log() != NULL)
2191 alock->as_Lock()->is_nested_lock_region(C); // rerun for debugging output
2192 #endif
2193 }
2194 }
2195 return;
2196 }
2197 // Process locks for non escaping object
2198 assert(alock->is_non_esc_obj(), "");
2199 } // EliminateNestedLocks
2200
2201 if (alock->is_non_esc_obj()) { // Lock is used for non escaping object
2202 // Look for all locks of this object and mark them and
2203 // corresponding BoxLock nodes as eliminated.
2204 Node* obj = alock->obj_node();
2205 for (uint j = 0; j < obj->outcnt(); j++) {
2206 Node* o = obj->raw_out(j);
2207 if (o->is_AbstractLock() &&
2208 o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) {
2209 alock = o->as_AbstractLock();
2210 Node* box = alock->box_node();
2211 // Replace old box node with new eliminated box for all users
2212 // of the same object and mark related locks as eliminated.
2213 mark_eliminated_box(box, obj);
2214 }
2215 }
2216 }
2217 }
2218
2219 // we have determined that this lock/unlock can be eliminated, we simply
2220 // eliminate the node without expanding it.
2221 //
2222 // Note: The membar's associated with the lock/unlock are currently not
2223 // eliminated. This should be investigated as a future enhancement.
2224 //
2225 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) {
2226
2227 if (!alock->is_eliminated()) {
2228 return false;
2229 }
2230 #ifdef ASSERT
2231 if (!alock->is_coarsened()) {
2232 // Check that new "eliminated" BoxLock node is created.
2233 BoxLockNode* oldbox = alock->box_node()->as_BoxLock();
2234 assert(oldbox->is_eliminated(), "should be done already");
2235 }
2236 #endif
2237
2238 alock->log_lock_optimization(C, "eliminate_lock");
2239
2240 #ifndef PRODUCT
2241 if (PrintEliminateLocks) {
2242 if (alock->is_Lock()) {
2243 tty->print_cr("++++ Eliminated: %d Lock", alock->_idx);
2244 } else {
2245 tty->print_cr("++++ Eliminated: %d Unlock", alock->_idx);
2246 }
2247 }
2248 #endif
2249
2250 Node* mem = alock->in(TypeFunc::Memory);
2251 Node* ctrl = alock->in(TypeFunc::Control);
2252
2253 extract_call_projections(alock);
2254 // There are 2 projections from the lock. The lock node will
2255 // be deleted when its last use is subsumed below.
2256 assert(alock->outcnt() == 2 &&
2257 _fallthroughproj != NULL &&
2258 _memproj_fallthrough != NULL,
2259 "Unexpected projections from Lock/Unlock");
2260
2261 Node* fallthroughproj = _fallthroughproj;
2262 Node* memproj_fallthrough = _memproj_fallthrough;
2263
2264 // The memory projection from a lock/unlock is RawMem
2265 // The input to a Lock is merged memory, so extract its RawMem input
2266 // (unless the MergeMem has been optimized away.)
2267 if (alock->is_Lock()) {
2268 // Seach for MemBarAcquireLock node and delete it also.
2269 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
2270 assert(membar != NULL && membar->Opcode() == Op_MemBarAcquireLock, "");
2271 Node* ctrlproj = membar->proj_out(TypeFunc::Control);
2272 Node* memproj = membar->proj_out(TypeFunc::Memory);
2273 _igvn.replace_node(ctrlproj, fallthroughproj);
2274 _igvn.replace_node(memproj, memproj_fallthrough);
2275
2276 // Delete FastLock node also if this Lock node is unique user
2277 // (a loop peeling may clone a Lock node).
2278 Node* flock = alock->as_Lock()->fastlock_node();
2279 if (flock->outcnt() == 1) {
2280 assert(flock->unique_out() == alock, "sanity");
2281 _igvn.replace_node(flock, top());
2282 }
2283 }
2284
2285 // Seach for MemBarReleaseLock node and delete it also.
2286 if (alock->is_Unlock() && ctrl != NULL && ctrl->is_Proj() &&
2287 ctrl->in(0)->is_MemBar()) {
2288 MemBarNode* membar = ctrl->in(0)->as_MemBar();
2289 assert(membar->Opcode() == Op_MemBarReleaseLock &&
2290 mem->is_Proj() && membar == mem->in(0), "");
2291 _igvn.replace_node(fallthroughproj, ctrl);
2292 _igvn.replace_node(memproj_fallthrough, mem);
2293 fallthroughproj = ctrl;
2294 memproj_fallthrough = mem;
2295 ctrl = membar->in(TypeFunc::Control);
2296 mem = membar->in(TypeFunc::Memory);
2297 }
2298
2299 _igvn.replace_node(fallthroughproj, ctrl);
2300 _igvn.replace_node(memproj_fallthrough, mem);
2301 return true;
2302 }
2303
2304
2305 //------------------------------expand_lock_node----------------------
2306 void PhaseMacroExpand::expand_lock_node(LockNode *lock) {
2307
2308 Node* ctrl = lock->in(TypeFunc::Control);
2309 Node* mem = lock->in(TypeFunc::Memory);
2310 Node* obj = lock->obj_node();
2311 Node* box = lock->box_node();
2312 Node* flock = lock->fastlock_node();
2313
2314 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2315
2316 // Make the merge point
2317 Node *region;
2318 Node *mem_phi;
2319 Node *slow_path;
2320
2321 if (UseOptoBiasInlining) {
2322 /*
2323 * See the full description in MacroAssembler::biased_locking_enter().
2324 *
2325 * if( (mark_word & biased_lock_mask) == biased_lock_pattern ) {
2326 * // The object is biased.
2327 * proto_node = klass->prototype_header;
2328 * o_node = thread | proto_node;
2329 * x_node = o_node ^ mark_word;
2330 * if( (x_node & ~age_mask) == 0 ) { // Biased to the current thread ?
2331 * // Done.
2332 * } else {
2333 * if( (x_node & biased_lock_mask) != 0 ) {
2334 * // The klass's prototype header is no longer biased.
2335 * cas(&mark_word, mark_word, proto_node)
2336 * goto cas_lock;
2337 * } else {
2338 * // The klass's prototype header is still biased.
2339 * if( (x_node & epoch_mask) != 0 ) { // Expired epoch?
2340 * old = mark_word;
2341 * new = o_node;
2342 * } else {
2343 * // Different thread or anonymous biased.
2344 * old = mark_word & (epoch_mask | age_mask | biased_lock_mask);
2345 * new = thread | old;
2346 * }
2347 * // Try to rebias.
2348 * if( cas(&mark_word, old, new) == 0 ) {
2349 * // Done.
2350 * } else {
2351 * goto slow_path; // Failed.
2352 * }
2353 * }
2354 * }
2355 * } else {
2356 * // The object is not biased.
2357 * cas_lock:
2358 * if( FastLock(obj) == 0 ) {
2359 * // Done.
2360 * } else {
2361 * slow_path:
2362 * OptoRuntime::complete_monitor_locking_Java(obj);
2363 * }
2364 * }
2365 */
2366
2367 region = new RegionNode(5);
2368 // create a Phi for the memory state
2369 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2370
2371 Node* fast_lock_region = new RegionNode(3);
2372 Node* fast_lock_mem_phi = new PhiNode( fast_lock_region, Type::MEMORY, TypeRawPtr::BOTTOM);
2373
2374 // First, check mark word for the biased lock pattern.
2375 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2376
2377 // Get fast path - mark word has the biased lock pattern.
2378 ctrl = opt_bits_test(ctrl, fast_lock_region, 1, mark_node,
2379 markOopDesc::biased_lock_mask_in_place,
2380 markOopDesc::biased_lock_pattern, true);
2381 // fast_lock_region->in(1) is set to slow path.
2382 fast_lock_mem_phi->init_req(1, mem);
2383
2384 // Now check that the lock is biased to the current thread and has
2385 // the same epoch and bias as Klass::_prototype_header.
2386
2387 // Special-case a fresh allocation to avoid building nodes:
2388 Node* klass_node = AllocateNode::Ideal_klass(obj, &_igvn);
2389 if (klass_node == NULL) {
2390 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
2391 klass_node = transform_later(LoadKlassNode::make(_igvn, NULL, mem, k_adr, _igvn.type(k_adr)->is_ptr()));
2392 #ifdef _LP64
2393 if (UseCompressedClassPointers && klass_node->is_DecodeNKlass()) {
2394 assert(klass_node->in(1)->Opcode() == Op_LoadNKlass, "sanity");
2395 klass_node->in(1)->init_req(0, ctrl);
2396 } else
2397 #endif
2398 klass_node->init_req(0, ctrl);
2399 }
2400 Node *proto_node = make_load(ctrl, mem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeX_X, TypeX_X->basic_type());
2401
2402 Node* thread = transform_later(new ThreadLocalNode());
2403 Node* cast_thread = transform_later(new CastP2XNode(ctrl, thread));
2404 Node* o_node = transform_later(new OrXNode(cast_thread, proto_node));
2405 Node* x_node = transform_later(new XorXNode(o_node, mark_node));
2406
2407 // Get slow path - mark word does NOT match the value.
2408 Node* not_biased_ctrl = opt_bits_test(ctrl, region, 3, x_node,
2409 (~markOopDesc::age_mask_in_place), 0);
2410 // region->in(3) is set to fast path - the object is biased to the current thread.
2411 mem_phi->init_req(3, mem);
2412
2413
2414 // Mark word does NOT match the value (thread | Klass::_prototype_header).
2415
2416
2417 // First, check biased pattern.
2418 // Get fast path - _prototype_header has the same biased lock pattern.
2419 ctrl = opt_bits_test(not_biased_ctrl, fast_lock_region, 2, x_node,
2420 markOopDesc::biased_lock_mask_in_place, 0, true);
2421
2422 not_biased_ctrl = fast_lock_region->in(2); // Slow path
2423 // fast_lock_region->in(2) - the prototype header is no longer biased
2424 // and we have to revoke the bias on this object.
2425 // We are going to try to reset the mark of this object to the prototype
2426 // value and fall through to the CAS-based locking scheme.
2427 Node* adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
2428 Node* cas = new StoreXConditionalNode(not_biased_ctrl, mem, adr,
2429 proto_node, mark_node);
2430 transform_later(cas);
2431 Node* proj = transform_later(new SCMemProjNode(cas));
2432 fast_lock_mem_phi->init_req(2, proj);
2433
2434
2435 // Second, check epoch bits.
2436 Node* rebiased_region = new RegionNode(3);
2437 Node* old_phi = new PhiNode( rebiased_region, TypeX_X);
2438 Node* new_phi = new PhiNode( rebiased_region, TypeX_X);
2439
2440 // Get slow path - mark word does NOT match epoch bits.
2441 Node* epoch_ctrl = opt_bits_test(ctrl, rebiased_region, 1, x_node,
2442 markOopDesc::epoch_mask_in_place, 0);
2443 // The epoch of the current bias is not valid, attempt to rebias the object
2444 // toward the current thread.
2445 rebiased_region->init_req(2, epoch_ctrl);
2446 old_phi->init_req(2, mark_node);
2447 new_phi->init_req(2, o_node);
2448
2449 // rebiased_region->in(1) is set to fast path.
2450 // The epoch of the current bias is still valid but we know
2451 // nothing about the owner; it might be set or it might be clear.
2452 Node* cmask = MakeConX(markOopDesc::biased_lock_mask_in_place |
2453 markOopDesc::age_mask_in_place |
2454 markOopDesc::epoch_mask_in_place);
2455 Node* old = transform_later(new AndXNode(mark_node, cmask));
2456 cast_thread = transform_later(new CastP2XNode(ctrl, thread));
2457 Node* new_mark = transform_later(new OrXNode(cast_thread, old));
2458 old_phi->init_req(1, old);
2459 new_phi->init_req(1, new_mark);
2460
2461 transform_later(rebiased_region);
2462 transform_later(old_phi);
2463 transform_later(new_phi);
2464
2465 // Try to acquire the bias of the object using an atomic operation.
2466 // If this fails we will go in to the runtime to revoke the object's bias.
2467 cas = new StoreXConditionalNode(rebiased_region, mem, adr, new_phi, old_phi);
2468 transform_later(cas);
2469 proj = transform_later(new SCMemProjNode(cas));
2470
2471 // Get slow path - Failed to CAS.
2472 not_biased_ctrl = opt_bits_test(rebiased_region, region, 4, cas, 0, 0);
2473 mem_phi->init_req(4, proj);
2474 // region->in(4) is set to fast path - the object is rebiased to the current thread.
2475
2476 // Failed to CAS.
2477 slow_path = new RegionNode(3);
2478 Node *slow_mem = new PhiNode( slow_path, Type::MEMORY, TypeRawPtr::BOTTOM);
2479
2480 slow_path->init_req(1, not_biased_ctrl); // Capture slow-control
2481 slow_mem->init_req(1, proj);
2482
2483 // Call CAS-based locking scheme (FastLock node).
2484
2485 transform_later(fast_lock_region);
2486 transform_later(fast_lock_mem_phi);
2487
2488 // Get slow path - FastLock failed to lock the object.
2489 ctrl = opt_bits_test(fast_lock_region, region, 2, flock, 0, 0);
2490 mem_phi->init_req(2, fast_lock_mem_phi);
2491 // region->in(2) is set to fast path - the object is locked to the current thread.
2492
2493 slow_path->init_req(2, ctrl); // Capture slow-control
2494 slow_mem->init_req(2, fast_lock_mem_phi);
2495
2496 transform_later(slow_path);
2497 transform_later(slow_mem);
2498 // Reset lock's memory edge.
2499 lock->set_req(TypeFunc::Memory, slow_mem);
2500
2501 } else {
2502 region = new RegionNode(3);
2503 // create a Phi for the memory state
2504 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2505
2506 // Optimize test; set region slot 2
2507 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0);
2508 mem_phi->init_req(2, mem);
2509 }
2510
2511 // Make slow path call
2512 CallNode *call = make_slow_call((CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(),
2513 OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path,
2514 obj, box, NULL);
2515
2516 extract_call_projections(call);
2517
2518 // Slow path can only throw asynchronous exceptions, which are always
2519 // de-opted. So the compiler thinks the slow-call can never throw an
2520 // exception. If it DOES throw an exception we would need the debug
2521 // info removed first (since if it throws there is no monitor).
2522 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
2523 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
2524
2525 // Capture slow path
2526 // disconnect fall-through projection from call and create a new one
2527 // hook up users of fall-through projection to region
2528 Node *slow_ctrl = _fallthroughproj->clone();
2529 transform_later(slow_ctrl);
2530 _igvn.hash_delete(_fallthroughproj);
2531 _fallthroughproj->disconnect_inputs(NULL, C);
2532 region->init_req(1, slow_ctrl);
2533 // region inputs are now complete
2534 transform_later(region);
2535 _igvn.replace_node(_fallthroughproj, region);
2536
2537 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory));
2538 mem_phi->init_req(1, memproj );
2539 transform_later(mem_phi);
2540 _igvn.replace_node(_memproj_fallthrough, mem_phi);
2541 }
2542
2543 //------------------------------expand_unlock_node----------------------
2544 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
2545
2546 Node* ctrl = unlock->in(TypeFunc::Control);
2547 Node* mem = unlock->in(TypeFunc::Memory);
2548 Node* obj = unlock->obj_node();
2549 Node* box = unlock->box_node();
2550
2551 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2552
2553 // No need for a null check on unlock
2554
2555 // Make the merge point
2556 Node *region;
2557 Node *mem_phi;
2558
2559 if (UseOptoBiasInlining) {
2560 // Check for biased locking unlock case, which is a no-op.
2561 // See the full description in MacroAssembler::biased_locking_exit().
2562 region = new RegionNode(4);
2563 // create a Phi for the memory state
2564 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2565 mem_phi->init_req(3, mem);
2566
2567 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2568 ctrl = opt_bits_test(ctrl, region, 3, mark_node,
2569 markOopDesc::biased_lock_mask_in_place,
2570 markOopDesc::biased_lock_pattern);
2571 } else {
2572 region = new RegionNode(3);
2573 // create a Phi for the memory state
2574 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2575 }
2576
2577 FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box );
2578 funlock = transform_later( funlock )->as_FastUnlock();
2579 // Optimize test; set region slot 2
2580 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0);
2581 Node *thread = transform_later(new ThreadLocalNode());
2582
2583 CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(),
2584 CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C),
2585 "complete_monitor_unlocking_C", slow_path, obj, box, thread);
2586
2587 extract_call_projections(call);
2588
2589 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
2590 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
2591
2592 // No exceptions for unlocking
2593 // Capture slow path
2594 // disconnect fall-through projection from call and create a new one
2595 // hook up users of fall-through projection to region
2596 Node *slow_ctrl = _fallthroughproj->clone();
2597 transform_later(slow_ctrl);
2598 _igvn.hash_delete(_fallthroughproj);
2599 _fallthroughproj->disconnect_inputs(NULL, C);
2600 region->init_req(1, slow_ctrl);
2601 // region inputs are now complete
2602 transform_later(region);
2603 _igvn.replace_node(_fallthroughproj, region);
2604
2605 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory) );
2606 mem_phi->init_req(1, memproj );
2607 mem_phi->init_req(2, mem);
2608 transform_later(mem_phi);
2609 _igvn.replace_node(_memproj_fallthrough, mem_phi);
2610 }
2611
2612 void PhaseMacroExpand::expand_loadbarrier_node(LoadBarrierNode *barrier) {
2613 Node* in_ctrl = barrier->in(LoadBarrierNode::Control);
2614 Node* in_mem = barrier->in(LoadBarrierNode::Memory);
2615 Node* in_val = barrier->in(LoadBarrierNode::Oop);
2616 Node* in_adr = barrier->in(LoadBarrierNode::Address);
2617
2618 Node* out_ctrl = barrier->proj_out(LoadBarrierNode::Control);
2619 Node* out_res = barrier->proj_out(LoadBarrierNode::Oop);
2620
2621 if (VerifyLoadBarriers) {
2622 _igvn.replace_node(out_res, in_val);
2623 _igvn.replace_node(out_ctrl, in_ctrl);
2624 return;
2625 }
2626
2627 if (barrier->can_be_eliminated()) {
2628 // Clone and pin the load for this barrier below the dominating
2629 // barrier: the load cannot be allowed to float above the
2630 // dominating barrier
2631 Node* load = in_val;
2632 Node* decode = NULL;
2633 if (load->is_DecodeN()) {
2634 decode = load;
2635 load = load->in(1);
2636 }
2637 if (load->is_Load()) {
2638 Node* new_load = load->clone();
2639 Node* addp = new_load->in(MemNode::Address);
2640 assert(addp->is_AddP() || addp->is_Phi(), "bad address");
2641 Node* cast = new CastPPNode(addp, _igvn.type(addp), true);
2642 Node* ctrl = NULL;
2643 Node* similar = barrier->in(LoadBarrierNode::Similar);
2644 if (similar->is_Phi()) {
2645 // already expanded
2646 ctrl = similar->in(0);
2647 } else {
2648 assert(similar->is_Proj() && similar->in(0)->is_LoadBarrier(), "unexpected graph shape");
2649 ctrl = similar->in(0)->as_LoadBarrier()->proj_out(LoadBarrierNode::Control);
2650 }
2651 assert(ctrl != NULL, "bad control");
2652 cast->set_req(0, ctrl);
2653 _igvn.transform(cast);
2654 new_load->set_req(MemNode::Address, cast);
2655 _igvn.transform(new_load);
2656
2657 Node* new_in_val = new_load;
2658 if (decode != NULL) {
2659 new_in_val = decode->clone();
2660 new_in_val->set_req(1, new_load);
2661 _igvn.transform(new_in_val);
2662 }
2663
2664 _igvn.replace_node(out_res, new_in_val);
2665 _igvn.replace_node(out_ctrl, in_ctrl);
2666 return;
2667 }
2668 // cannot eliminate
2669 }
2670
2671 // There are two cases that require the basic loadbarrier
2672 // 1) When the writeback of a healed oop must be avoided (swap)
2673 // 2) When we must guarantee that no reload of is done (swap, cas, cmpx)
2674 if (!barrier->is_writeback()) {
2675 assert(!barrier->oop_reload_allowed(), "writeback barriers should be marked as requires oop");
2676 }
2677
2678 if (!barrier->oop_reload_allowed() || C->directive()->UseBasicLoadBarrierOption) {
2679 expand_loadbarrier_basic(barrier);
2680 } else {
2681 expand_loadbarrier_optimized(barrier);
2682 }
2683 }
2684
2685 // CMH - Recfactor out overlap between barrier variants
2686
2687 // Original basic loadbarrier using conventional arg passing
2688 void PhaseMacroExpand::expand_loadbarrier_basic(LoadBarrierNode *barrier)
2689 {
2690 Node* in_ctrl = barrier->in(LoadBarrierNode::Control);
2691 Node* in_mem = barrier->in(LoadBarrierNode::Memory);
2692 Node* in_val = barrier->in(LoadBarrierNode::Oop);
2693 Node* in_adr = barrier->in(LoadBarrierNode::Address);
2694
2695 Node* out_ctrl = barrier->proj_out(LoadBarrierNode::Control);
2696 Node* out_res = barrier->proj_out(LoadBarrierNode::Oop);
2697
2698 const uintptr_t in_cset_bit = 0x1;
2699 const unsigned char* in_cset_map = (unsigned char*)ZHeap::heap()->pagetable_addr();
2700 const int in_cset_shift = ZPageSizeMinShift;
2701
2702 const uintptr_t* addr_good_bit = &ZAddressGoodMask;
2703 const uintptr_t* addr_bad_bit = &ZAddressBadMask;
2704
2705 const uintptr_t remapped_bit = ZAddressMetadataRemapped;
2706 const uintptr_t offset_mask = ZAddressOffsetMask;
2707
2708 float unlikely = PROB_UNLIKELY(0.999);
2709 float likely = PROB_LIKELY(0.999);
2710 const Type* in_val_maybe_null_t = _igvn.type(in_val);
2711
2712 #ifdef SPARC
2713 Node* bad_mask = _igvn.transform(new AddrBadBitNode());
2714 #else
2715 Node* jthread = _igvn.transform(new ThreadLocalNode());
2716 Node* adr = basic_plus_adr(jthread, in_bytes(JavaThread::zaddress_bad_mask_offset()));
2717 Node* bad_mask = _igvn.transform(LoadNode::make(_igvn, in_ctrl, in_mem, adr, TypeRawPtr::BOTTOM, TypeX_X, TypeX_X->basic_type(), MemNode::unordered));
2718 #endif
2719 Node* cast = _igvn.transform(new CastP2XNode(in_ctrl, in_val));
2720 Node* obj_masked = _igvn.transform(new AndXNode(cast, bad_mask));
2721 Node* cmp = _igvn.transform(new CmpXNode(obj_masked, _igvn.zerocon(TypeX_X->basic_type())));
2722 Node *bol = _igvn.transform(new BoolNode(cmp, BoolTest::ne))->as_Bool();
2723 IfNode* iff = _igvn.transform(new IfNode(in_ctrl, bol, unlikely, COUNT_UNKNOWN))->as_If();
2724 Node* then = _igvn.transform(new IfTrueNode(iff));
2725 Node* elsen = _igvn.transform(new IfFalseNode(iff));
2726
2727 Node* result_region;
2728 Node* result_val;
2729
2730 result_region = new RegionNode(3);
2731 result_val = new PhiNode(result_region, TypeInstPtr::BOTTOM);
2732
2733 result_region->set_req(1, elsen);
2734 Node* res = _igvn.transform(new CastPPNode(in_val, in_val_maybe_null_t));
2735 res->init_req(0, elsen);
2736 result_val->set_req(1, res);
2737
2738 const TypeFunc *tf = OptoRuntime::load_barrier_Type();
2739 Node* call;
2740 if (barrier->is_weak()) {
2741 call = new CallLeafNode(tf, CAST_FROM_FN_PTR(address, SharedRuntime::z_load_barrier_on_weak_oop_field_preloaded), "load_barrier", TypeRawPtr::BOTTOM);
2742 } else {
2743 call = new CallLeafNode(tf, CAST_FROM_FN_PTR(address, SharedRuntime::z_load_barrier_on_oop_field_preloaded), "load_barrier", TypeRawPtr::BOTTOM);
2744 }
2745
2746 call->init_req(TypeFunc::Control, then);
2747 call->init_req(TypeFunc::I_O , top());
2748 call->init_req(TypeFunc::Memory , in_mem);
2749 call->init_req(TypeFunc::FramePtr, top());
2750 call->init_req(TypeFunc::ReturnAdr, top());
2751 call->init_req(TypeFunc::Parms+0, in_val);
2752 if (barrier->is_writeback()) {
2753 call->init_req(TypeFunc::Parms+1, in_adr);
2754 } else {
2755 // when slow path is called with a null adr, the healed oop will not be written back
2756 call->init_req(TypeFunc::Parms+1, _igvn.zerocon(T_OBJECT));
2757 }
2758 call = _igvn.transform(call);
2759
2760 Node* ctrl = _igvn.transform(new ProjNode(call, TypeFunc::Control));
2761 res = _igvn.transform(new ProjNode(call, TypeFunc::Parms));
2762 res = _igvn.transform(new CheckCastPPNode(ctrl, res, in_val_maybe_null_t));
2763
2764 result_region->set_req(2, ctrl);
2765 result_val->set_req(2, res);
2766
2767 result_region = _igvn.transform(result_region);
2768 result_val = _igvn.transform(result_val);
2769
2770 if ( out_ctrl != NULL ) { // added if cond
2771 _igvn.replace_node(out_ctrl, result_region);
2772 }
2773 _igvn.replace_node(out_res, result_val);
2774 }
2775
2776 // Low spill loadbarrier variant using stub specialized on register used
2777 void PhaseMacroExpand::expand_loadbarrier_optimized(LoadBarrierNode *barrier)
2778 {
2779 #ifdef PRINT_NODE_TRAVERSALS
2780 Node* preceding_barrier_node = barrier->in(LoadBarrierNode::Oop);
2781 #endif
2782
2783 Node* in_ctrl = barrier->in(LoadBarrierNode::Control);
2784 Node* in_mem = barrier->in(LoadBarrierNode::Memory);
2785 Node* in_val = barrier->in(LoadBarrierNode::Oop);
2786 Node* in_adr = barrier->in(LoadBarrierNode::Address);
2787
2788 Node* out_ctrl = barrier->proj_out(LoadBarrierNode::Control);
2789 Node* out_res = barrier->proj_out(LoadBarrierNode::Oop);
2790
2791 assert( barrier->in(LoadBarrierNode::Oop) != NULL, "oop to loadbarrier node cannot be null" );
2792
2793 #ifdef PRINT_NODE_TRAVERSALS
2794 tty->print( "\n\n\nBefore barrier optimization:\n" );
2795 traverse( barrier, out_ctrl, out_res, -1 );
2796
2797 tty->print( "\nBefore barrier optimization: preceding_barrier_node\n" );
2798 traverse( preceding_barrier_node, out_ctrl, out_res, -1 );
2799 #endif
2800
2801 const uintptr_t in_cset_bit = 0x1;
2802 const unsigned char* in_cset_map = (unsigned char*)ZHeap::heap()->pagetable_addr();
2803 const int in_cset_shift = ZPageSizeMinShift;
2804
2805 const uintptr_t* addr_good_bit = &ZAddressGoodMask;
2806 const uintptr_t* addr_bad_bit = &ZAddressBadMask;
2807
2808 const uintptr_t remapped_bit = ZAddressMetadataRemapped;
2809 const uintptr_t offset_mask = ZAddressOffsetMask;
2810
2811 float unlikely = PROB_UNLIKELY(0.999);
2812 float likely = PROB_LIKELY(0.999);
2813 const Type* in_val_not_null_t = _igvn.type(in_val);
2814
2815 Node* jthread = _igvn.transform(new ThreadLocalNode());
2816 Node* adr = basic_plus_adr(jthread, in_bytes(JavaThread::zaddress_bad_mask_offset()));
2817 Node* bad_mask = _igvn.transform(LoadNode::make(_igvn, in_ctrl, in_mem, adr, TypeRawPtr::BOTTOM, TypeX_X, TypeX_X->basic_type(), MemNode::unordered));
2818 Node* cast = _igvn.transform(new CastP2XNode(in_ctrl, in_val));
2819 Node* obj_masked = _igvn.transform(new AndXNode(cast, bad_mask));
2820 Node* cmp = _igvn.transform(new CmpXNode(obj_masked, _igvn.zerocon(TypeX_X->basic_type())));
2821 Node *bol = _igvn.transform(new BoolNode(cmp, BoolTest::ne))->as_Bool();
2822 IfNode* iff = _igvn.transform(new IfNode(in_ctrl, bol, unlikely, COUNT_UNKNOWN))->as_If();
2823 Node* then = _igvn.transform(new IfTrueNode(iff));
2824 Node* elsen = _igvn.transform(new IfFalseNode(iff));
2825
2826 Node* slow_path_surrogate;
2827 if (!barrier->is_weak()) {
2828 slow_path_surrogate = _igvn.transform( new LoadBarrierSlowRegNode(then, in_mem, in_adr, in_val->adr_type(), (const TypePtr*) in_val->bottom_type(), MemNode::unordered) );
2829 } else {
2830 slow_path_surrogate = _igvn.transform( new LoadBarrierWeakSlowRegNode(then, in_mem, in_adr, in_val->adr_type(), (const TypePtr*) in_val->bottom_type(), MemNode::unordered) );
2831 }
2832
2833 Node *new_loadp;
2834 new_loadp = slow_path_surrogate;
2835 // create the final region/phi pair to converge cntl/data paths to downstream code
2836 Node* result_region = _igvn.transform(new RegionNode(3));
2837 result_region->set_req(1, then);
2838 result_region->set_req(2, elsen);
2839
2840 Node* result_phi = _igvn.transform(new PhiNode(result_region, TypeInstPtr::BOTTOM));
2841 result_phi->set_req(1, new_loadp);
2842 result_phi->set_req(2, barrier->in(LoadBarrierNode::Oop));
2843
2844 // finally, connect the original outputs to the barrier region and phi to complete the expansion/substitution
2845 // _igvn.replace_node(out_ctrl, result_region);
2846 if ( out_ctrl != NULL ) { // added if cond
2847 _igvn.replace_node(out_ctrl, result_region);
2848 }
2849 _igvn.replace_node(out_res, result_phi);
2850
2851 assert(barrier->outcnt() == 0,"LoadBarrier macro node has non-null outputs after expansion!");
2852
2853 #ifdef PRINT_NODE_TRAVERSALS
2854 tty->print( "\nAfter barrier optimization: old out_ctrl\n" );
2855 traverse( out_ctrl, out_ctrl, out_res, -1 );
2856 tty->print( "\nAfter barrier optimization: old out_res\n" );
2857 traverse( out_res, out_ctrl, out_res, -1 );
2858 tty->print( "\nAfter barrier optimization: old barrier\n" );
2859 traverse( barrier, out_ctrl, out_res, -1 );
2860 tty->print( "\nAfter barrier optimization: preceding_barrier_node\n" );
2861 traverse( preceding_barrier_node, result_region, result_phi, -1 );
2862 #endif
2863
2864 return;
2865 }
2866
2867 //---------------------------eliminate_macro_nodes----------------------
2868 // Eliminate scalar replaced allocations and associated locks.
2869 void PhaseMacroExpand::eliminate_macro_nodes() {
2870 if (C->macro_count() == 0)
2871 return;
2872
2873 // First, attempt to eliminate locks
2874 int cnt = C->macro_count();
2875 for (int i=0; i < cnt; i++) {
2876 Node *n = C->macro_node(i);
2877 if (n->is_AbstractLock()) { // Lock and Unlock nodes
2878 // Before elimination mark all associated (same box and obj)
2879 // lock and unlock nodes.
2880 mark_eliminated_locking_nodes(n->as_AbstractLock());
2881 }
2882 }
2883 bool progress = true;
2884 while (progress) {
2885 progress = false;
2886 for (int i = C->macro_count(); i > 0; i--) {
2887 Node * n = C->macro_node(i-1);
2888 bool success = false;
2889 debug_only(int old_macro_count = C->macro_count(););
2890 if (n->is_AbstractLock()) {
2891 success = eliminate_locking_node(n->as_AbstractLock());
2892 }
2893 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2894 progress = progress || success;
2895 }
2896 }
2897 // Next, attempt to eliminate allocations
2898 _has_locks = false;
2899 progress = true;
2900 while (progress) {
2901 progress = false;
2902 for (int i = C->macro_count(); i > 0; i--) {
2903 Node * n = C->macro_node(i-1);
2904 bool success = false;
2905 debug_only(int old_macro_count = C->macro_count(););
2906 switch (n->class_id()) {
2907 case Node::Class_Allocate:
2908 case Node::Class_AllocateArray:
2909 success = eliminate_allocate_node(n->as_Allocate());
2910 break;
2911 case Node::Class_CallStaticJava:
2912 success = eliminate_boxing_node(n->as_CallStaticJava());
2913 break;
2914 case Node::Class_Lock:
2915 case Node::Class_Unlock:
2916 assert(!n->as_AbstractLock()->is_eliminated(), "sanity");
2917 _has_locks = true;
2918 break;
2919 case Node::Class_ArrayCopy:
2920 break;
2921 case Node::Class_OuterStripMinedLoop:
2922 break;
2923 default:
2924 assert(n->Opcode() == Op_LoopLimit ||
2925 n->Opcode() == Op_Opaque1 ||
2926 n->Opcode() == Op_Opaque2 ||
2927 n->Opcode() == Op_Opaque3 ||
2928 n->Opcode() == Op_LoadBarrier ||
2929 n->Opcode() == Op_Opaque4, "unknown node type in macro list");
2930 }
2931 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2932 progress = progress || success;
2933 }
2934 }
2935 }
2936
2937 //------------------------------expand_macro_nodes----------------------
2938 // Returns true if a failure occurred.
2939 bool PhaseMacroExpand::expand_macro_nodes() {
2940 // Last attempt to eliminate macro nodes.
2941 eliminate_macro_nodes();
2942
2943 // Make sure expansion will not cause node limit to be exceeded.
2944 // Worst case is a macro node gets expanded into about 200 nodes.
2945 // Allow 50% more for optimization.
2946 if (C->check_node_count(C->macro_count() * 300, "out of nodes before macro expansion" ) )
2947 return true;
2948
2949 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations.
2950 bool progress = true;
2951 while (progress) {
2952 progress = false;
2953 for (int i = C->macro_count(); i > 0; i--) {
2954 Node * n = C->macro_node(i-1);
2955 bool success = false;
2956 debug_only(int old_macro_count = C->macro_count(););
2957 if (n->Opcode() == Op_LoopLimit) {
2958 // Remove it from macro list and put on IGVN worklist to optimize.
2959 C->remove_macro_node(n);
2960 _igvn._worklist.push(n);
2961 success = true;
2962 } else if (n->Opcode() == Op_CallStaticJava) {
2963 // Remove it from macro list and put on IGVN worklist to optimize.
2964 C->remove_macro_node(n);
2965 _igvn._worklist.push(n);
2966 success = true;
2967 } else if (n->Opcode() == Op_Opaque1 || n->Opcode() == Op_Opaque2) {
2968 _igvn.replace_node(n, n->in(1));
2969 success = true;
2970 #if INCLUDE_RTM_OPT
2971 } else if ((n->Opcode() == Op_Opaque3) && ((Opaque3Node*)n)->rtm_opt()) {
2972 assert(C->profile_rtm(), "should be used only in rtm deoptimization code");
2973 assert((n->outcnt() == 1) && n->unique_out()->is_Cmp(), "");
2974 Node* cmp = n->unique_out();
2975 #ifdef ASSERT
2976 // Validate graph.
2977 assert((cmp->outcnt() == 1) && cmp->unique_out()->is_Bool(), "");
2978 BoolNode* bol = cmp->unique_out()->as_Bool();
2979 assert((bol->outcnt() == 1) && bol->unique_out()->is_If() &&
2980 (bol->_test._test == BoolTest::ne), "");
2981 IfNode* ifn = bol->unique_out()->as_If();
2982 assert((ifn->outcnt() == 2) &&
2983 ifn->proj_out(1)->is_uncommon_trap_proj(Deoptimization::Reason_rtm_state_change) != NULL, "");
2984 #endif
2985 Node* repl = n->in(1);
2986 if (!_has_locks) {
2987 // Remove RTM state check if there are no locks in the code.
2988 // Replace input to compare the same value.
2989 repl = (cmp->in(1) == n) ? cmp->in(2) : cmp->in(1);
2990 }
2991 _igvn.replace_node(n, repl);
2992 success = true;
2993 #endif
2994 } else if (n->Opcode() == Op_Opaque4) {
2995 _igvn.replace_node(n, n->in(2));
2996 success = true;
2997 } else if (n->Opcode() == Op_OuterStripMinedLoop) {
2998 n->as_OuterStripMinedLoop()->adjust_strip_mined_loop(&_igvn);
2999 C->remove_macro_node(n);
3000 success = true;
3001 }
3002 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
3003 progress = progress || success;
3004 }
3005 }
3006
3007 // expand arraycopy "macro" nodes first
3008 // For ReduceBulkZeroing, we must first process all arraycopy nodes
3009 // before the allocate nodes are expanded.
3010 int macro_idx = C->macro_count() - 1;
3011 while (macro_idx >= 0) {
3012 Node * n = C->macro_node(macro_idx);
3013 assert(n->is_macro(), "only macro nodes expected here");
3014 if (_igvn.type(n) == Type::TOP || (n->in(0) != NULL && n->in(0)->is_top())) {
3015 // node is unreachable, so don't try to expand it
3016 C->remove_macro_node(n);
3017 } else if (n->is_ArrayCopy()){
3018 int macro_count = C->macro_count();
3019 expand_arraycopy_node(n->as_ArrayCopy());
3020 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
3021 }
3022 if (C->failing()) return true;
3023 macro_idx --;
3024 }
3025
3026 // expand "macro" nodes
3027 // nodes are removed from the macro list as they are processed
3028 while (C->macro_count() > 0) {
3029 int macro_count = C->macro_count();
3030 Node * n = C->macro_node(macro_count-1);
3031 assert(n->is_macro(), "only macro nodes expected here");
3032 if (_igvn.type(n) == Type::TOP || (n->in(0) != NULL && n->in(0)->is_top())) {
3033 // node is unreachable, so don't try to expand it
3034 C->remove_macro_node(n);
3035 continue;
3036 }
3037 switch (n->class_id()) {
3038 case Node::Class_Allocate:
3039 expand_allocate(n->as_Allocate());
3040 break;
3041 case Node::Class_AllocateArray:
3042 expand_allocate_array(n->as_AllocateArray());
3043 break;
3044 case Node::Class_Lock:
3045 expand_lock_node(n->as_Lock());
3046 break;
3047 case Node::Class_Unlock:
3048 expand_unlock_node(n->as_Unlock());
3049 break;
3050 default:
3051 assert(false, "unknown node type in macro list");
3052 }
3053 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
3054 if (C->failing()) return true;
3055 }
3056
3057 _igvn.set_delay_transform(false);
3058 _igvn.optimize();
3059 if (C->failing()) return true;
3060 if (C->load_barrier_count() > 0) {
3061 #ifdef ASSERT
3062 C->verify_load_barriers(false);
3063 {
3064 ResourceMark rm;
3065 stringStream ss;
3066 C->method()->print_short_name(&ss);
3067 if (!strcmp(ss.as_string(), " EDU.purdue.cs.bloat.editor.Type::simple")) {
3068 C->root()->dump(9999);
3069 }
3070 }
3071 #endif
3072 _igvn.set_delay_transform(true);
3073 int skipped = 0;
3074 while (C->load_barrier_count() > skipped) {
3075 int load_barrier_count = C->load_barrier_count();
3076 LoadBarrierNode * n = C->load_barrier_node(load_barrier_count-1-skipped);
3077 if (_igvn.type(n) == Type::TOP || (n->in(0) != NULL && n->in(0)->is_top())) {
3078 // node is unreachable, so don't try to expand it
3079 C->remove_load_barrier_node(n);
3080 continue;
3081 }
3082 if (!n->can_be_eliminated()) {
3083 skipped++;
3084 continue;
3085 }
3086 expand_loadbarrier_node(n);
3087 assert(C->load_barrier_count() < load_barrier_count, "must have deleted a node from load barrier list");
3088 if (C->failing()) return true;
3089 }
3090 while (C->load_barrier_count() > 0) {
3091 int load_barrier_count = C->load_barrier_count();
3092 LoadBarrierNode * n = C->load_barrier_node(load_barrier_count-1);
3093 assert(!(_igvn.type(n) == Type::TOP || (n->in(0) != NULL && n->in(0)->is_top())), "should have been processed already");
3094 assert(!n->can_be_eliminated(), "should have been processed already");
3095 expand_loadbarrier_node(n);
3096 assert(C->load_barrier_count() < load_barrier_count, "must have deleted a node from load barrier list");
3097 if (C->failing()) return true;
3098 }
3099 _igvn.set_delay_transform(false);
3100 _igvn.optimize();
3101 if (C->failing()) return true;
3102 }
3103 return false;
3104 }