1 /*
2 * Copyright (c) 2005, 2018, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "compiler/compileLog.hpp"
27 #include "libadt/vectset.hpp"
28 #include "opto/addnode.hpp"
29 #include "opto/callnode.hpp"
30 #include "opto/cfgnode.hpp"
31 #include "opto/compile.hpp"
32 #include "opto/connode.hpp"
33 #include "opto/locknode.hpp"
34 #include "opto/loopnode.hpp"
35 #include "opto/macro.hpp"
36 #include "opto/memnode.hpp"
37 #include "opto/node.hpp"
38 #include "opto/phaseX.hpp"
39 #include "opto/rootnode.hpp"
40 #include "opto/runtime.hpp"
41 #include "opto/subnode.hpp"
42 #include "opto/type.hpp"
43 #include "runtime/sharedRuntime.hpp"
44
45
46 //
47 // Replace any references to "oldref" in inputs to "use" with "newref".
48 // Returns the number of replacements made.
49 //
replace_input(Node * use,Node * oldref,Node * newref)50 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) {
51 int nreplacements = 0;
52 uint req = use->req();
53 for (uint j = 0; j < use->len(); j++) {
54 Node *uin = use->in(j);
55 if (uin == oldref) {
56 if (j < req)
57 use->set_req(j, newref);
58 else
59 use->set_prec(j, newref);
60 nreplacements++;
61 } else if (j >= req && uin == NULL) {
62 break;
63 }
64 }
65 return nreplacements;
66 }
67
copy_call_debug_info(CallNode * oldcall,CallNode * newcall)68 void PhaseMacroExpand::copy_call_debug_info(CallNode *oldcall, CallNode * newcall) {
69 // Copy debug information and adjust JVMState information
70 uint old_dbg_start = oldcall->tf()->domain()->cnt();
71 uint new_dbg_start = newcall->tf()->domain()->cnt();
72 int jvms_adj = new_dbg_start - old_dbg_start;
73 assert (new_dbg_start == newcall->req(), "argument count mismatch");
74
75 // SafePointScalarObject node could be referenced several times in debug info.
76 // Use Dict to record cloned nodes.
77 Dict* sosn_map = new Dict(cmpkey,hashkey);
78 for (uint i = old_dbg_start; i < oldcall->req(); i++) {
79 Node* old_in = oldcall->in(i);
80 // Clone old SafePointScalarObjectNodes, adjusting their field contents.
81 if (old_in != NULL && old_in->is_SafePointScalarObject()) {
82 SafePointScalarObjectNode* old_sosn = old_in->as_SafePointScalarObject();
83 uint old_unique = C->unique();
84 Node* new_in = old_sosn->clone(sosn_map);
85 if (old_unique != C->unique()) { // New node?
86 new_in->set_req(0, C->root()); // reset control edge
87 new_in = transform_later(new_in); // Register new node.
88 }
89 old_in = new_in;
90 }
91 newcall->add_req(old_in);
92 }
93
94 newcall->set_jvms(oldcall->jvms());
95 for (JVMState *jvms = newcall->jvms(); jvms != NULL; jvms = jvms->caller()) {
96 jvms->set_map(newcall);
97 jvms->set_locoff(jvms->locoff()+jvms_adj);
98 jvms->set_stkoff(jvms->stkoff()+jvms_adj);
99 jvms->set_monoff(jvms->monoff()+jvms_adj);
100 jvms->set_scloff(jvms->scloff()+jvms_adj);
101 jvms->set_endoff(jvms->endoff()+jvms_adj);
102 }
103 }
104
opt_bits_test(Node * ctrl,Node * region,int edge,Node * word,int mask,int bits,bool return_fast_path)105 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word, int mask, int bits, bool return_fast_path) {
106 Node* cmp;
107 if (mask != 0) {
108 Node* and_node = transform_later(new (C) AndXNode(word, MakeConX(mask)));
109 cmp = transform_later(new (C) CmpXNode(and_node, MakeConX(bits)));
110 } else {
111 cmp = word;
112 }
113 Node* bol = transform_later(new (C) BoolNode(cmp, BoolTest::ne));
114 IfNode* iff = new (C) IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN );
115 transform_later(iff);
116
117 // Fast path taken.
118 Node *fast_taken = transform_later( new (C) IfFalseNode(iff) );
119
120 // Fast path not-taken, i.e. slow path
121 Node *slow_taken = transform_later( new (C) IfTrueNode(iff) );
122
123 if (return_fast_path) {
124 region->init_req(edge, slow_taken); // Capture slow-control
125 return fast_taken;
126 } else {
127 region->init_req(edge, fast_taken); // Capture fast-control
128 return slow_taken;
129 }
130 }
131
132 //--------------------copy_predefined_input_for_runtime_call--------------------
copy_predefined_input_for_runtime_call(Node * ctrl,CallNode * oldcall,CallNode * call)133 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) {
134 // Set fixed predefined input arguments
135 call->init_req( TypeFunc::Control, ctrl );
136 call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) );
137 call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ?????
138 call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) );
139 call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) );
140 }
141
142 //------------------------------make_slow_call---------------------------------
make_slow_call(CallNode * oldcall,const TypeFunc * slow_call_type,address slow_call,const char * leaf_name,Node * slow_path,Node * parm0,Node * parm1)143 CallNode* PhaseMacroExpand::make_slow_call(CallNode *oldcall, const TypeFunc* slow_call_type, address slow_call, const char* leaf_name, Node* slow_path, Node* parm0, Node* parm1) {
144
145 // Slow-path call
146 CallNode *call = leaf_name
147 ? (CallNode*)new (C) CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM )
148 : (CallNode*)new (C) CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), oldcall->jvms()->bci(), TypeRawPtr::BOTTOM );
149
150 // Slow path call has no side-effects, uses few values
151 copy_predefined_input_for_runtime_call(slow_path, oldcall, call );
152 if (parm0 != NULL) call->init_req(TypeFunc::Parms+0, parm0);
153 if (parm1 != NULL) call->init_req(TypeFunc::Parms+1, parm1);
154 copy_call_debug_info(oldcall, call);
155 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
156 _igvn.replace_node(oldcall, call);
157 transform_later(call);
158
159 return call;
160 }
161
extract_call_projections(CallNode * call)162 void PhaseMacroExpand::extract_call_projections(CallNode *call) {
163 _fallthroughproj = NULL;
164 _fallthroughcatchproj = NULL;
165 _ioproj_fallthrough = NULL;
166 _ioproj_catchall = NULL;
167 _catchallcatchproj = NULL;
168 _memproj_fallthrough = NULL;
169 _memproj_catchall = NULL;
170 _resproj = NULL;
171 for (DUIterator_Fast imax, i = call->fast_outs(imax); i < imax; i++) {
172 ProjNode *pn = call->fast_out(i)->as_Proj();
173 switch (pn->_con) {
174 case TypeFunc::Control:
175 {
176 // For Control (fallthrough) and I_O (catch_all_index) we have CatchProj -> Catch -> Proj
177 _fallthroughproj = pn;
178 DUIterator_Fast jmax, j = pn->fast_outs(jmax);
179 const Node *cn = pn->fast_out(j);
180 if (cn->is_Catch()) {
181 ProjNode *cpn = NULL;
182 for (DUIterator_Fast kmax, k = cn->fast_outs(kmax); k < kmax; k++) {
183 cpn = cn->fast_out(k)->as_Proj();
184 assert(cpn->is_CatchProj(), "must be a CatchProjNode");
185 if (cpn->_con == CatchProjNode::fall_through_index)
186 _fallthroughcatchproj = cpn;
187 else {
188 assert(cpn->_con == CatchProjNode::catch_all_index, "must be correct index.");
189 _catchallcatchproj = cpn;
190 }
191 }
192 }
193 break;
194 }
195 case TypeFunc::I_O:
196 if (pn->_is_io_use)
197 _ioproj_catchall = pn;
198 else
199 _ioproj_fallthrough = pn;
200 break;
201 case TypeFunc::Memory:
202 if (pn->_is_io_use)
203 _memproj_catchall = pn;
204 else
205 _memproj_fallthrough = pn;
206 break;
207 case TypeFunc::Parms:
208 _resproj = pn;
209 break;
210 default:
211 assert(false, "unexpected projection from allocation node.");
212 }
213 }
214
215 }
216
217 // Eliminate a card mark sequence. p2x is a ConvP2XNode
eliminate_card_mark(Node * p2x)218 void PhaseMacroExpand::eliminate_card_mark(Node* p2x) {
219 assert(p2x->Opcode() == Op_CastP2X, "ConvP2XNode required");
220 if (!UseG1GC) {
221 // vanilla/CMS post barrier
222 Node *shift = p2x->unique_out();
223 Node *addp = shift->unique_out();
224 for (DUIterator_Last jmin, j = addp->last_outs(jmin); j >= jmin; --j) {
225 Node *mem = addp->last_out(j);
226 if (UseCondCardMark && mem->is_Load()) {
227 assert(mem->Opcode() == Op_LoadB, "unexpected code shape");
228 // The load is checking if the card has been written so
229 // replace it with zero to fold the test.
230 _igvn.replace_node(mem, intcon(0));
231 continue;
232 }
233 assert(mem->is_Store(), "store required");
234 _igvn.replace_node(mem, mem->in(MemNode::Memory));
235 }
236 } else {
237 // G1 pre/post barriers
238 assert(p2x->outcnt() <= 2, "expects 1 or 2 users: Xor and URShift nodes");
239 // It could be only one user, URShift node, in Object.clone() instrinsic
240 // but the new allocation is passed to arraycopy stub and it could not
241 // be scalar replaced. So we don't check the case.
242
243 // An other case of only one user (Xor) is when the value check for NULL
244 // in G1 post barrier is folded after CCP so the code which used URShift
245 // is removed.
246
247 // Take Region node before eliminating post barrier since it also
248 // eliminates CastP2X node when it has only one user.
249 Node* this_region = p2x->in(0);
250 assert(this_region != NULL, "");
251
252 // Remove G1 post barrier.
253
254 // Search for CastP2X->Xor->URShift->Cmp path which
255 // checks if the store done to a different from the value's region.
256 // And replace Cmp with #0 (false) to collapse G1 post barrier.
257 Node* xorx = NULL;
258 for (DUIterator_Fast imax, i = p2x->fast_outs(imax); i < imax; i++) {
259 Node* u = p2x->fast_out(i);
260 if (u->Opcode() == Op_XorX) {
261 xorx = u;
262 break;
263 }
264 }
265 assert(xorx != NULL, "missing G1 post barrier");
266 Node* shift = xorx->unique_out();
267 Node* cmpx = shift->unique_out();
268 assert(cmpx->is_Cmp() && cmpx->unique_out()->is_Bool() &&
269 cmpx->unique_out()->as_Bool()->_test._test == BoolTest::ne,
270 "missing region check in G1 post barrier");
271 _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ));
272
273 // Remove G1 pre barrier.
274
275 // Search "if (marking != 0)" check and set it to "false".
276 // There is no G1 pre barrier if previous stored value is NULL
277 // (for example, after initialization).
278 if (this_region->is_Region() && this_region->req() == 3) {
279 int ind = 1;
280 if (!this_region->in(ind)->is_IfFalse()) {
281 ind = 2;
282 }
283 if (this_region->in(ind)->is_IfFalse()) {
284 Node* bol = this_region->in(ind)->in(0)->in(1);
285 assert(bol->is_Bool(), "");
286 cmpx = bol->in(1);
287 if (bol->as_Bool()->_test._test == BoolTest::ne &&
288 cmpx->is_Cmp() && cmpx->in(2) == intcon(0) &&
289 cmpx->in(1)->is_Load()) {
290 Node* adr = cmpx->in(1)->as_Load()->in(MemNode::Address);
291 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() +
292 PtrQueue::byte_offset_of_active());
293 if (adr->is_AddP() && adr->in(AddPNode::Base) == top() &&
294 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
295 adr->in(AddPNode::Offset) == MakeConX(marking_offset)) {
296 _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ));
297 }
298 }
299 }
300 }
301 // Now CastP2X can be removed since it is used only on dead path
302 // which currently still alive until igvn optimize it.
303 assert(p2x->outcnt() == 0 || p2x->unique_out()->Opcode() == Op_URShiftX, "");
304 _igvn.replace_node(p2x, top());
305 }
306 }
307
308 // Search for a memory operation for the specified memory slice.
scan_mem_chain(Node * mem,int alias_idx,int offset,Node * start_mem,Node * alloc,PhaseGVN * phase)309 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) {
310 Node *orig_mem = mem;
311 Node *alloc_mem = alloc->in(TypeFunc::Memory);
312 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr();
313 while (true) {
314 if (mem == alloc_mem || mem == start_mem ) {
315 return mem; // hit one of our sentinels
316 } else if (mem->is_MergeMem()) {
317 mem = mem->as_MergeMem()->memory_at(alias_idx);
318 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) {
319 Node *in = mem->in(0);
320 // we can safely skip over safepoints, calls, locks and membars because we
321 // already know that the object is safe to eliminate.
322 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) {
323 return in;
324 } else if (in->is_Call()) {
325 CallNode *call = in->as_Call();
326 if (!call->may_modify(tinst, phase)) {
327 mem = call->in(TypeFunc::Memory);
328 }
329 mem = in->in(TypeFunc::Memory);
330 } else if (in->is_MemBar()) {
331 mem = in->in(TypeFunc::Memory);
332 } else {
333 assert(false, "unexpected projection");
334 }
335 } else if (mem->is_Store()) {
336 const TypePtr* atype = mem->as_Store()->adr_type();
337 int adr_idx = Compile::current()->get_alias_index(atype);
338 if (adr_idx == alias_idx) {
339 assert(atype->isa_oopptr(), "address type must be oopptr");
340 int adr_offset = atype->offset();
341 uint adr_iid = atype->is_oopptr()->instance_id();
342 // Array elements references have the same alias_idx
343 // but different offset and different instance_id.
344 if (adr_offset == offset && adr_iid == alloc->_idx)
345 return mem;
346 } else {
347 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw");
348 }
349 mem = mem->in(MemNode::Memory);
350 } else if (mem->is_ClearArray()) {
351 if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) {
352 // Can not bypass initialization of the instance
353 // we are looking.
354 debug_only(intptr_t offset;)
355 assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity");
356 InitializeNode* init = alloc->as_Allocate()->initialization();
357 // We are looking for stored value, return Initialize node
358 // or memory edge from Allocate node.
359 if (init != NULL)
360 return init;
361 else
362 return alloc->in(TypeFunc::Memory); // It will produce zero value (see callers).
363 }
364 // Otherwise skip it (the call updated 'mem' value).
365 } else if (mem->Opcode() == Op_SCMemProj) {
366 mem = mem->in(0);
367 Node* adr = NULL;
368 if (mem->is_LoadStore()) {
369 adr = mem->in(MemNode::Address);
370 } else {
371 assert(mem->Opcode() == Op_EncodeISOArray, "sanity");
372 adr = mem->in(3); // Destination array
373 }
374 const TypePtr* atype = adr->bottom_type()->is_ptr();
375 int adr_idx = Compile::current()->get_alias_index(atype);
376 if (adr_idx == alias_idx) {
377 assert(false, "Object is not scalar replaceable if a LoadStore node access its field");
378 return NULL;
379 }
380 mem = mem->in(MemNode::Memory);
381 } else {
382 return mem;
383 }
384 assert(mem != orig_mem, "dead memory loop");
385 }
386 }
387
388 //
389 // Given a Memory Phi, compute a value Phi containing the values from stores
390 // on the input paths.
391 // Note: this function is recursive, its depth is limied by the "level" argument
392 // Returns the computed Phi, or NULL if it cannot compute it.
value_from_mem_phi(Node * mem,BasicType ft,const Type * phi_type,const TypeOopPtr * adr_t,Node * alloc,Node_Stack * value_phis,int level)393 Node *PhaseMacroExpand::value_from_mem_phi(Node *mem, BasicType ft, const Type *phi_type, const TypeOopPtr *adr_t, Node *alloc, Node_Stack *value_phis, int level) {
394 assert(mem->is_Phi(), "sanity");
395 int alias_idx = C->get_alias_index(adr_t);
396 int offset = adr_t->offset();
397 int instance_id = adr_t->instance_id();
398
399 // Check if an appropriate value phi already exists.
400 Node* region = mem->in(0);
401 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) {
402 Node* phi = region->fast_out(k);
403 if (phi->is_Phi() && phi != mem &&
404 phi->as_Phi()->is_same_inst_field(phi_type, (int)mem->_idx, instance_id, alias_idx, offset)) {
405 return phi;
406 }
407 }
408 // Check if an appropriate new value phi already exists.
409 Node* new_phi = value_phis->find(mem->_idx);
410 if (new_phi != NULL)
411 return new_phi;
412
413 if (level <= 0) {
414 return NULL; // Give up: phi tree too deep
415 }
416 Node *start_mem = C->start()->proj_out(TypeFunc::Memory);
417 Node *alloc_mem = alloc->in(TypeFunc::Memory);
418
419 uint length = mem->req();
420 GrowableArray <Node *> values(length, length, NULL, false);
421
422 // create a new Phi for the value
423 PhiNode *phi = new (C) PhiNode(mem->in(0), phi_type, NULL, mem->_idx, instance_id, alias_idx, offset);
424 transform_later(phi);
425 value_phis->push(phi, mem->_idx);
426
427 for (uint j = 1; j < length; j++) {
428 Node *in = mem->in(j);
429 if (in == NULL || in->is_top()) {
430 values.at_put(j, in);
431 } else {
432 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn);
433 if (val == start_mem || val == alloc_mem) {
434 // hit a sentinel, return appropriate 0 value
435 values.at_put(j, _igvn.zerocon(ft));
436 continue;
437 }
438 if (val->is_Initialize()) {
439 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
440 }
441 if (val == NULL) {
442 return NULL; // can't find a value on this path
443 }
444 if (val == mem) {
445 values.at_put(j, mem);
446 } else if (val->is_Store()) {
447 values.at_put(j, val->in(MemNode::ValueIn));
448 } else if(val->is_Proj() && val->in(0) == alloc) {
449 values.at_put(j, _igvn.zerocon(ft));
450 } else if (val->is_Phi()) {
451 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1);
452 if (val == NULL) {
453 return NULL;
454 }
455 values.at_put(j, val);
456 } else if (val->Opcode() == Op_SCMemProj) {
457 assert(val->in(0)->is_LoadStore() || val->in(0)->Opcode() == Op_EncodeISOArray, "sanity");
458 assert(false, "Object is not scalar replaceable if a LoadStore node access its field");
459 return NULL;
460 } else {
461 #ifdef ASSERT
462 val->dump();
463 assert(false, "unknown node on this path");
464 #endif
465 return NULL; // unknown node on this path
466 }
467 }
468 }
469 // Set Phi's inputs
470 for (uint j = 1; j < length; j++) {
471 if (values.at(j) == mem) {
472 phi->init_req(j, phi);
473 } else {
474 phi->init_req(j, values.at(j));
475 }
476 }
477 return phi;
478 }
479
480 // Search the last value stored into the object's field.
value_from_mem(Node * sfpt_mem,BasicType ft,const Type * ftype,const TypeOopPtr * adr_t,Node * alloc)481 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, Node *alloc) {
482 assert(adr_t->is_known_instance_field(), "instance required");
483 int instance_id = adr_t->instance_id();
484 assert((uint)instance_id == alloc->_idx, "wrong allocation");
485
486 int alias_idx = C->get_alias_index(adr_t);
487 int offset = adr_t->offset();
488 Node *start_mem = C->start()->proj_out(TypeFunc::Memory);
489 Node *alloc_ctrl = alloc->in(TypeFunc::Control);
490 Node *alloc_mem = alloc->in(TypeFunc::Memory);
491 Arena *a = Thread::current()->resource_area();
492 VectorSet visited(a);
493
494
495 bool done = sfpt_mem == alloc_mem;
496 Node *mem = sfpt_mem;
497 while (!done) {
498 if (visited.test_set(mem->_idx)) {
499 return NULL; // found a loop, give up
500 }
501 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn);
502 if (mem == start_mem || mem == alloc_mem) {
503 done = true; // hit a sentinel, return appropriate 0 value
504 } else if (mem->is_Initialize()) {
505 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
506 if (mem == NULL) {
507 done = true; // Something go wrong.
508 } else if (mem->is_Store()) {
509 const TypePtr* atype = mem->as_Store()->adr_type();
510 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice");
511 done = true;
512 }
513 } else if (mem->is_Store()) {
514 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr();
515 assert(atype != NULL, "address type must be oopptr");
516 assert(C->get_alias_index(atype) == alias_idx &&
517 atype->is_known_instance_field() && atype->offset() == offset &&
518 atype->instance_id() == instance_id, "store is correct memory slice");
519 done = true;
520 } else if (mem->is_Phi()) {
521 // try to find a phi's unique input
522 Node *unique_input = NULL;
523 Node *top = C->top();
524 for (uint i = 1; i < mem->req(); i++) {
525 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn);
526 if (n == NULL || n == top || n == mem) {
527 continue;
528 } else if (unique_input == NULL) {
529 unique_input = n;
530 } else if (unique_input != n) {
531 unique_input = top;
532 break;
533 }
534 }
535 if (unique_input != NULL && unique_input != top) {
536 mem = unique_input;
537 } else {
538 done = true;
539 }
540 } else {
541 assert(false, "unexpected node");
542 }
543 }
544 if (mem != NULL) {
545 if (mem == start_mem || mem == alloc_mem) {
546 // hit a sentinel, return appropriate 0 value
547 return _igvn.zerocon(ft);
548 } else if (mem->is_Store()) {
549 return mem->in(MemNode::ValueIn);
550 } else if (mem->is_Phi()) {
551 // attempt to produce a Phi reflecting the values on the input paths of the Phi
552 Node_Stack value_phis(a, 8);
553 Node * phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit);
554 if (phi != NULL) {
555 return phi;
556 } else {
557 // Kill all new Phis
558 while(value_phis.is_nonempty()) {
559 Node* n = value_phis.node();
560 _igvn.replace_node(n, C->top());
561 value_phis.pop();
562 }
563 }
564 }
565 }
566 // Something go wrong.
567 return NULL;
568 }
569
570 // Check the possibility of scalar replacement.
can_eliminate_allocation(AllocateNode * alloc,GrowableArray<SafePointNode * > & safepoints)571 bool PhaseMacroExpand::can_eliminate_allocation(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
572 // Scan the uses of the allocation to check for anything that would
573 // prevent us from eliminating it.
574 NOT_PRODUCT( const char* fail_eliminate = NULL; )
575 DEBUG_ONLY( Node* disq_node = NULL; )
576 bool can_eliminate = true;
577
578 Node* res = alloc->result_cast();
579 const TypeOopPtr* res_type = NULL;
580 if (res == NULL) {
581 // All users were eliminated.
582 } else if (!res->is_CheckCastPP()) {
583 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";)
584 can_eliminate = false;
585 } else {
586 res_type = _igvn.type(res)->isa_oopptr();
587 if (res_type == NULL) {
588 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";)
589 can_eliminate = false;
590 } else if (res_type->isa_aryptr()) {
591 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1);
592 if (length < 0) {
593 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";)
594 can_eliminate = false;
595 }
596 }
597 }
598
599 if (can_eliminate && res != NULL) {
600 for (DUIterator_Fast jmax, j = res->fast_outs(jmax);
601 j < jmax && can_eliminate; j++) {
602 Node* use = res->fast_out(j);
603
604 if (use->is_AddP()) {
605 const TypePtr* addp_type = _igvn.type(use)->is_ptr();
606 int offset = addp_type->offset();
607
608 if (offset == Type::OffsetTop || offset == Type::OffsetBot) {
609 NOT_PRODUCT(fail_eliminate = "Undefined field referrence";)
610 can_eliminate = false;
611 break;
612 }
613 for (DUIterator_Fast kmax, k = use->fast_outs(kmax);
614 k < kmax && can_eliminate; k++) {
615 Node* n = use->fast_out(k);
616 if (!n->is_Store() && n->Opcode() != Op_CastP2X) {
617 DEBUG_ONLY(disq_node = n;)
618 if (n->is_Load() || n->is_LoadStore()) {
619 NOT_PRODUCT(fail_eliminate = "Field load";)
620 } else {
621 NOT_PRODUCT(fail_eliminate = "Not store field referrence";)
622 }
623 can_eliminate = false;
624 }
625 }
626 } else if (use->is_SafePoint()) {
627 SafePointNode* sfpt = use->as_SafePoint();
628 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) {
629 // Object is passed as argument.
630 DEBUG_ONLY(disq_node = use;)
631 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";)
632 can_eliminate = false;
633 }
634 Node* sfptMem = sfpt->memory();
635 if (sfptMem == NULL || sfptMem->is_top()) {
636 DEBUG_ONLY(disq_node = use;)
637 NOT_PRODUCT(fail_eliminate = "NULL or TOP memory";)
638 can_eliminate = false;
639 } else {
640 safepoints.append_if_missing(sfpt);
641 }
642 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
643 if (use->is_Phi()) {
644 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
645 NOT_PRODUCT(fail_eliminate = "Object is return value";)
646 } else {
647 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
648 }
649 DEBUG_ONLY(disq_node = use;)
650 } else {
651 if (use->Opcode() == Op_Return) {
652 NOT_PRODUCT(fail_eliminate = "Object is return value";)
653 }else {
654 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
655 }
656 DEBUG_ONLY(disq_node = use;)
657 }
658 can_eliminate = false;
659 }
660 }
661 }
662
663 #ifndef PRODUCT
664 if (PrintEliminateAllocations) {
665 if (can_eliminate) {
666 tty->print("Scalar ");
667 if (res == NULL)
668 alloc->dump();
669 else
670 res->dump();
671 } else if (alloc->_is_scalar_replaceable) {
672 tty->print("NotScalar (%s)", fail_eliminate);
673 if (res == NULL)
674 alloc->dump();
675 else
676 res->dump();
677 #ifdef ASSERT
678 if (disq_node != NULL) {
679 tty->print(" >>>> ");
680 disq_node->dump();
681 }
682 #endif /*ASSERT*/
683 }
684 }
685 #endif
686 return can_eliminate;
687 }
688
689 // Do scalar replacement.
scalar_replacement(AllocateNode * alloc,GrowableArray<SafePointNode * > & safepoints)690 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
691 GrowableArray <SafePointNode *> safepoints_done;
692
693 ciKlass* klass = NULL;
694 ciInstanceKlass* iklass = NULL;
695 int nfields = 0;
696 int array_base = 0;
697 int element_size = 0;
698 BasicType basic_elem_type = T_ILLEGAL;
699 ciType* elem_type = NULL;
700
701 Node* res = alloc->result_cast();
702 assert(res == NULL || res->is_CheckCastPP(), "unexpected AllocateNode result");
703 const TypeOopPtr* res_type = NULL;
704 if (res != NULL) { // Could be NULL when there are no users
705 res_type = _igvn.type(res)->isa_oopptr();
706 }
707
708 if (res != NULL) {
709 klass = res_type->klass();
710 if (res_type->isa_instptr()) {
711 // find the fields of the class which will be needed for safepoint debug information
712 assert(klass->is_instance_klass(), "must be an instance klass.");
713 iklass = klass->as_instance_klass();
714 nfields = iklass->nof_nonstatic_fields();
715 } else {
716 // find the array's elements which will be needed for safepoint debug information
717 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
718 assert(klass->is_array_klass() && nfields >= 0, "must be an array klass.");
719 elem_type = klass->as_array_klass()->element_type();
720 basic_elem_type = elem_type->basic_type();
721 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
722 element_size = type2aelembytes(basic_elem_type);
723 }
724 }
725 //
726 // Process the safepoint uses
727 //
728 while (safepoints.length() > 0) {
729 SafePointNode* sfpt = safepoints.pop();
730 Node* mem = sfpt->memory();
731 assert(sfpt->jvms() != NULL, "missed JVMS");
732 // Fields of scalar objs are referenced only at the end
733 // of regular debuginfo at the last (youngest) JVMS.
734 // Record relative start index.
735 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff());
736 SafePointScalarObjectNode* sobj = new (C) SafePointScalarObjectNode(res_type,
737 #ifdef ASSERT
738 alloc,
739 #endif
740 first_ind, nfields);
741 sobj->init_req(0, C->root());
742 transform_later(sobj);
743
744 // Scan object's fields adding an input to the safepoint for each field.
745 for (int j = 0; j < nfields; j++) {
746 intptr_t offset;
747 ciField* field = NULL;
748 if (iklass != NULL) {
749 field = iklass->nonstatic_field_at(j);
750 offset = field->offset();
751 elem_type = field->type();
752 basic_elem_type = field->layout_type();
753 } else {
754 offset = array_base + j * (intptr_t)element_size;
755 }
756
757 const Type *field_type;
758 // The next code is taken from Parse::do_get_xxx().
759 if (basic_elem_type == T_OBJECT || basic_elem_type == T_ARRAY) {
760 if (!elem_type->is_loaded()) {
761 field_type = TypeInstPtr::BOTTOM;
762 } else if (field != NULL && field->is_constant() && field->is_static()) {
763 // This can happen if the constant oop is non-perm.
764 ciObject* con = field->constant_value().as_object();
765 // Do not "join" in the previous type; it doesn't add value,
766 // and may yield a vacuous result if the field is of interface type.
767 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr();
768 assert(field_type != NULL, "field singleton type must be consistent");
769 } else {
770 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
771 }
772 if (UseCompressedOops) {
773 field_type = field_type->make_narrowoop();
774 basic_elem_type = T_NARROWOOP;
775 }
776 } else {
777 field_type = Type::get_const_basic_type(basic_elem_type);
778 }
779
780 const TypeOopPtr *field_addr_type = res_type->add_offset(offset)->isa_oopptr();
781
782 Node *field_val = value_from_mem(mem, basic_elem_type, field_type, field_addr_type, alloc);
783 if (field_val == NULL) {
784 // We weren't able to find a value for this field,
785 // give up on eliminating this allocation.
786
787 // Remove any extra entries we added to the safepoint.
788 uint last = sfpt->req() - 1;
789 for (int k = 0; k < j; k++) {
790 sfpt->del_req(last--);
791 }
792 // rollback processed safepoints
793 while (safepoints_done.length() > 0) {
794 SafePointNode* sfpt_done = safepoints_done.pop();
795 // remove any extra entries we added to the safepoint
796 last = sfpt_done->req() - 1;
797 for (int k = 0; k < nfields; k++) {
798 sfpt_done->del_req(last--);
799 }
800 JVMState *jvms = sfpt_done->jvms();
801 jvms->set_endoff(sfpt_done->req());
802 // Now make a pass over the debug information replacing any references
803 // to SafePointScalarObjectNode with the allocated object.
804 int start = jvms->debug_start();
805 int end = jvms->debug_end();
806 for (int i = start; i < end; i++) {
807 if (sfpt_done->in(i)->is_SafePointScalarObject()) {
808 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject();
809 if (scobj->first_index(jvms) == sfpt_done->req() &&
810 scobj->n_fields() == (uint)nfields) {
811 assert(scobj->alloc() == alloc, "sanity");
812 sfpt_done->set_req(i, res);
813 }
814 }
815 }
816 }
817 #ifndef PRODUCT
818 if (PrintEliminateAllocations) {
819 if (field != NULL) {
820 tty->print("=== At SafePoint node %d can't find value of Field: ",
821 sfpt->_idx);
822 field->print();
823 int field_idx = C->get_alias_index(field_addr_type);
824 tty->print(" (alias_idx=%d)", field_idx);
825 } else { // Array's element
826 tty->print("=== At SafePoint node %d can't find value of array element [%d]",
827 sfpt->_idx, j);
828 }
829 tty->print(", which prevents elimination of: ");
830 if (res == NULL)
831 alloc->dump();
832 else
833 res->dump();
834 }
835 #endif
836 return false;
837 }
838 if (UseCompressedOops && field_type->isa_narrowoop()) {
839 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
840 // to be able scalar replace the allocation.
841 if (field_val->is_EncodeP()) {
842 field_val = field_val->in(1);
843 } else {
844 field_val = transform_later(new (C) DecodeNNode(field_val, field_val->get_ptr_type()));
845 }
846 }
847 sfpt->add_req(field_val);
848 }
849 JVMState *jvms = sfpt->jvms();
850 jvms->set_endoff(sfpt->req());
851 // Now make a pass over the debug information replacing any references
852 // to the allocated object with "sobj"
853 int start = jvms->debug_start();
854 int end = jvms->debug_end();
855 sfpt->replace_edges_in_range(res, sobj, start, end);
856 safepoints_done.append_if_missing(sfpt); // keep it for rollback
857 }
858 return true;
859 }
860
861 // Process users of eliminated allocation.
process_users_of_allocation(CallNode * alloc)862 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc) {
863 Node* res = alloc->result_cast();
864 if (res != NULL) {
865 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
866 Node *use = res->last_out(j);
867 uint oc1 = res->outcnt();
868
869 if (use->is_AddP()) {
870 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) {
871 Node *n = use->last_out(k);
872 uint oc2 = use->outcnt();
873 if (n->is_Store()) {
874 #ifdef ASSERT
875 // Verify that there is no dependent MemBarVolatile nodes,
876 // they should be removed during IGVN, see MemBarNode::Ideal().
877 for (DUIterator_Fast pmax, p = n->fast_outs(pmax);
878 p < pmax; p++) {
879 Node* mb = n->fast_out(p);
880 assert(mb->is_Initialize() || !mb->is_MemBar() ||
881 mb->req() <= MemBarNode::Precedent ||
882 mb->in(MemBarNode::Precedent) != n,
883 "MemBarVolatile should be eliminated for non-escaping object");
884 }
885 #endif
886 _igvn.replace_node(n, n->in(MemNode::Memory));
887 } else {
888 eliminate_card_mark(n);
889 }
890 k -= (oc2 - use->outcnt());
891 }
892 } else {
893 eliminate_card_mark(use);
894 }
895 j -= (oc1 - res->outcnt());
896 }
897 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
898 _igvn.remove_dead_node(res);
899 }
900
901 //
902 // Process other users of allocation's projections
903 //
904 if (_resproj != NULL && _resproj->outcnt() != 0) {
905 // First disconnect stores captured by Initialize node.
906 // If Initialize node is eliminated first in the following code,
907 // it will kill such stores and DUIterator_Last will assert.
908 for (DUIterator_Fast jmax, j = _resproj->fast_outs(jmax); j < jmax; j++) {
909 Node *use = _resproj->fast_out(j);
910 if (use->is_AddP()) {
911 // raw memory addresses used only by the initialization
912 _igvn.replace_node(use, C->top());
913 --j; --jmax;
914 }
915 }
916 for (DUIterator_Last jmin, j = _resproj->last_outs(jmin); j >= jmin; ) {
917 Node *use = _resproj->last_out(j);
918 uint oc1 = _resproj->outcnt();
919 if (use->is_Initialize()) {
920 // Eliminate Initialize node.
921 InitializeNode *init = use->as_Initialize();
922 assert(init->outcnt() <= 2, "only a control and memory projection expected");
923 Node *ctrl_proj = init->proj_out(TypeFunc::Control);
924 if (ctrl_proj != NULL) {
925 assert(init->in(TypeFunc::Control) == _fallthroughcatchproj, "allocation control projection");
926 _igvn.replace_node(ctrl_proj, _fallthroughcatchproj);
927 }
928 Node *mem_proj = init->proj_out(TypeFunc::Memory);
929 if (mem_proj != NULL) {
930 Node *mem = init->in(TypeFunc::Memory);
931 #ifdef ASSERT
932 if (mem->is_MergeMem()) {
933 assert(mem->in(TypeFunc::Memory) == _memproj_fallthrough, "allocation memory projection");
934 } else {
935 assert(mem == _memproj_fallthrough, "allocation memory projection");
936 }
937 #endif
938 _igvn.replace_node(mem_proj, mem);
939 }
940 } else {
941 assert(false, "only Initialize or AddP expected");
942 }
943 j -= (oc1 - _resproj->outcnt());
944 }
945 }
946 if (_fallthroughcatchproj != NULL) {
947 _igvn.replace_node(_fallthroughcatchproj, alloc->in(TypeFunc::Control));
948 }
949 if (_memproj_fallthrough != NULL) {
950 _igvn.replace_node(_memproj_fallthrough, alloc->in(TypeFunc::Memory));
951 }
952 if (_memproj_catchall != NULL) {
953 _igvn.replace_node(_memproj_catchall, C->top());
954 }
955 if (_ioproj_fallthrough != NULL) {
956 _igvn.replace_node(_ioproj_fallthrough, alloc->in(TypeFunc::I_O));
957 }
958 if (_ioproj_catchall != NULL) {
959 _igvn.replace_node(_ioproj_catchall, C->top());
960 }
961 if (_catchallcatchproj != NULL) {
962 _igvn.replace_node(_catchallcatchproj, C->top());
963 }
964 }
965
eliminate_allocate_node(AllocateNode * alloc)966 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) {
967 // Don't do scalar replacement if the frame can be popped by JVMTI:
968 // if reallocation fails during deoptimization we'll pop all
969 // interpreter frames for this compiled frame and that won't play
970 // nice with JVMTI popframe.
971 if (!EliminateAllocations || JvmtiExport::can_pop_frame() || !alloc->_is_non_escaping) {
972 return false;
973 }
974 Node* klass = alloc->in(AllocateNode::KlassNode);
975 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr();
976 Node* res = alloc->result_cast();
977 // Eliminate boxing allocations which are not used
978 // regardless scalar replacable status.
979 bool boxing_alloc = C->eliminate_boxing() &&
980 tklass->klass()->is_instance_klass() &&
981 tklass->klass()->as_instance_klass()->is_box_klass();
982 if (!alloc->_is_scalar_replaceable && (!boxing_alloc || (res != NULL))) {
983 return false;
984 }
985
986 extract_call_projections(alloc);
987
988 GrowableArray <SafePointNode *> safepoints;
989 if (!can_eliminate_allocation(alloc, safepoints)) {
990 return false;
991 }
992
993 if (!alloc->_is_scalar_replaceable) {
994 assert(res == NULL, "sanity");
995 // We can only eliminate allocation if all debug info references
996 // are already replaced with SafePointScalarObject because
997 // we can't search for a fields value without instance_id.
998 if (safepoints.length() > 0) {
999 return false;
1000 }
1001 }
1002
1003 if (!scalar_replacement(alloc, safepoints)) {
1004 return false;
1005 }
1006
1007 CompileLog* log = C->log();
1008 if (log != NULL) {
1009 log->head("eliminate_allocation type='%d'",
1010 log->identify(tklass->klass()));
1011 JVMState* p = alloc->jvms();
1012 while (p != NULL) {
1013 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1014 p = p->caller();
1015 }
1016 log->tail("eliminate_allocation");
1017 }
1018
1019 process_users_of_allocation(alloc);
1020
1021 #ifndef PRODUCT
1022 if (PrintEliminateAllocations) {
1023 if (alloc->is_AllocateArray())
1024 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1025 else
1026 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1027 }
1028 #endif
1029
1030 return true;
1031 }
1032
eliminate_boxing_node(CallStaticJavaNode * boxing)1033 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) {
1034 // EA should remove all uses of non-escaping boxing node.
1035 if (!C->eliminate_boxing() || boxing->proj_out(TypeFunc::Parms) != NULL) {
1036 return false;
1037 }
1038
1039 assert(boxing->result_cast() == NULL, "unexpected boxing node result");
1040
1041 extract_call_projections(boxing);
1042
1043 const TypeTuple* r = boxing->tf()->range();
1044 assert(r->cnt() > TypeFunc::Parms, "sanity");
1045 const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr();
1046 assert(t != NULL, "sanity");
1047
1048 CompileLog* log = C->log();
1049 if (log != NULL) {
1050 log->head("eliminate_boxing type='%d'",
1051 log->identify(t->klass()));
1052 JVMState* p = boxing->jvms();
1053 while (p != NULL) {
1054 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1055 p = p->caller();
1056 }
1057 log->tail("eliminate_boxing");
1058 }
1059
1060 process_users_of_allocation(boxing);
1061
1062 #ifndef PRODUCT
1063 if (PrintEliminateAllocations) {
1064 tty->print("++++ Eliminated: %d ", boxing->_idx);
1065 boxing->method()->print_short_name(tty);
1066 tty->cr();
1067 }
1068 #endif
1069
1070 return true;
1071 }
1072
1073 //---------------------------set_eden_pointers-------------------------
set_eden_pointers(Node * & eden_top_adr,Node * & eden_end_adr)1074 void PhaseMacroExpand::set_eden_pointers(Node* &eden_top_adr, Node* &eden_end_adr) {
1075 if (UseTLAB) { // Private allocation: load from TLS
1076 Node* thread = transform_later(new (C) ThreadLocalNode());
1077 int tlab_top_offset = in_bytes(JavaThread::tlab_top_offset());
1078 int tlab_end_offset = in_bytes(JavaThread::tlab_end_offset());
1079 eden_top_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_top_offset);
1080 eden_end_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_end_offset);
1081 } else { // Shared allocation: load from globals
1082 CollectedHeap* ch = Universe::heap();
1083 address top_adr = (address)ch->top_addr();
1084 address end_adr = (address)ch->end_addr();
1085 eden_top_adr = makecon(TypeRawPtr::make(top_adr));
1086 eden_end_adr = basic_plus_adr(eden_top_adr, end_adr - top_adr);
1087 }
1088 }
1089
1090
make_load(Node * ctl,Node * mem,Node * base,int offset,const Type * value_type,BasicType bt)1091 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) {
1092 Node* adr = basic_plus_adr(base, offset);
1093 const TypePtr* adr_type = adr->bottom_type()->is_ptr();
1094 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt, MemNode::unordered);
1095 transform_later(value);
1096 return value;
1097 }
1098
1099
make_store(Node * ctl,Node * mem,Node * base,int offset,Node * value,BasicType bt)1100 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) {
1101 Node* adr = basic_plus_adr(base, offset);
1102 mem = StoreNode::make(_igvn, ctl, mem, adr, NULL, value, bt, MemNode::unordered);
1103 transform_later(mem);
1104 return mem;
1105 }
1106
1107 //=============================================================================
1108 //
1109 // A L L O C A T I O N
1110 //
1111 // Allocation attempts to be fast in the case of frequent small objects.
1112 // It breaks down like this:
1113 //
1114 // 1) Size in doublewords is computed. This is a constant for objects and
1115 // variable for most arrays. Doubleword units are used to avoid size
1116 // overflow of huge doubleword arrays. We need doublewords in the end for
1117 // rounding.
1118 //
1119 // 2) Size is checked for being 'too large'. Too-large allocations will go
1120 // the slow path into the VM. The slow path can throw any required
1121 // exceptions, and does all the special checks for very large arrays. The
1122 // size test can constant-fold away for objects. For objects with
1123 // finalizers it constant-folds the otherway: you always go slow with
1124 // finalizers.
1125 //
1126 // 3) If NOT using TLABs, this is the contended loop-back point.
1127 // Load-Locked the heap top. If using TLABs normal-load the heap top.
1128 //
1129 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route.
1130 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish
1131 // "size*8" we always enter the VM, where "largish" is a constant picked small
1132 // enough that there's always space between the eden max and 4Gig (old space is
1133 // there so it's quite large) and large enough that the cost of entering the VM
1134 // is dwarfed by the cost to initialize the space.
1135 //
1136 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back
1137 // down. If contended, repeat at step 3. If using TLABs normal-store
1138 // adjusted heap top back down; there is no contention.
1139 //
1140 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark
1141 // fields.
1142 //
1143 // 7) Merge with the slow-path; cast the raw memory pointer to the correct
1144 // oop flavor.
1145 //
1146 //=============================================================================
1147 // FastAllocateSizeLimit value is in DOUBLEWORDS.
1148 // Allocations bigger than this always go the slow route.
1149 // This value must be small enough that allocation attempts that need to
1150 // trigger exceptions go the slow route. Also, it must be small enough so
1151 // that heap_top + size_in_bytes does not wrap around the 4Gig limit.
1152 //=============================================================================j//
1153 // %%% Here is an old comment from parseHelper.cpp; is it outdated?
1154 // The allocator will coalesce int->oop copies away. See comment in
1155 // coalesce.cpp about how this works. It depends critically on the exact
1156 // code shape produced here, so if you are changing this code shape
1157 // make sure the GC info for the heap-top is correct in and around the
1158 // slow-path call.
1159 //
1160
expand_allocate_common(AllocateNode * alloc,Node * length,const TypeFunc * slow_call_type,address slow_call_address)1161 void PhaseMacroExpand::expand_allocate_common(
1162 AllocateNode* alloc, // allocation node to be expanded
1163 Node* length, // array length for an array allocation
1164 const TypeFunc* slow_call_type, // Type of slow call
1165 address slow_call_address // Address of slow call
1166 )
1167 {
1168
1169 Node* ctrl = alloc->in(TypeFunc::Control);
1170 Node* mem = alloc->in(TypeFunc::Memory);
1171 Node* i_o = alloc->in(TypeFunc::I_O);
1172 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize);
1173 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1174 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest);
1175
1176 assert(ctrl != NULL, "must have control");
1177 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results.
1178 // they will not be used if "always_slow" is set
1179 enum { slow_result_path = 1, fast_result_path = 2 };
1180 Node *result_region = NULL;
1181 Node *result_phi_rawmem = NULL;
1182 Node *result_phi_rawoop = NULL;
1183 Node *result_phi_i_o = NULL;
1184
1185 // The initial slow comparison is a size check, the comparison
1186 // we want to do is a BoolTest::gt
1187 bool always_slow = false;
1188 int tv = _igvn.find_int_con(initial_slow_test, -1);
1189 if (tv >= 0) {
1190 always_slow = (tv == 1);
1191 initial_slow_test = NULL;
1192 } else {
1193 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn);
1194 }
1195
1196 if (C->env()->dtrace_alloc_probes() ||
1197 !UseTLAB && (!Universe::heap()->supports_inline_contig_alloc() ||
1198 (UseConcMarkSweepGC && CMSIncrementalMode))) {
1199 // Force slow-path allocation
1200 always_slow = true;
1201 initial_slow_test = NULL;
1202 }
1203
1204
1205 enum { too_big_or_final_path = 1, need_gc_path = 2 };
1206 Node *slow_region = NULL;
1207 Node *toobig_false = ctrl;
1208
1209 assert (initial_slow_test == NULL || !always_slow, "arguments must be consistent");
1210 // generate the initial test if necessary
1211 if (initial_slow_test != NULL ) {
1212 slow_region = new (C) RegionNode(3);
1213
1214 // Now make the initial failure test. Usually a too-big test but
1215 // might be a TRUE for finalizers or a fancy class check for
1216 // newInstance0.
1217 IfNode *toobig_iff = new (C) IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
1218 transform_later(toobig_iff);
1219 // Plug the failing-too-big test into the slow-path region
1220 Node *toobig_true = new (C) IfTrueNode( toobig_iff );
1221 transform_later(toobig_true);
1222 slow_region ->init_req( too_big_or_final_path, toobig_true );
1223 toobig_false = new (C) IfFalseNode( toobig_iff );
1224 transform_later(toobig_false);
1225 } else { // No initial test, just fall into next case
1226 toobig_false = ctrl;
1227 debug_only(slow_region = NodeSentinel);
1228 }
1229
1230 Node *slow_mem = mem; // save the current memory state for slow path
1231 // generate the fast allocation code unless we know that the initial test will always go slow
1232 if (!always_slow) {
1233 // Fast path modifies only raw memory.
1234 if (mem->is_MergeMem()) {
1235 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1236 }
1237
1238 Node* eden_top_adr;
1239 Node* eden_end_adr;
1240
1241 set_eden_pointers(eden_top_adr, eden_end_adr);
1242
1243 // Load Eden::end. Loop invariant and hoisted.
1244 //
1245 // Note: We set the control input on "eden_end" and "old_eden_top" when using
1246 // a TLAB to work around a bug where these values were being moved across
1247 // a safepoint. These are not oops, so they cannot be include in the oop
1248 // map, but they can be changed by a GC. The proper way to fix this would
1249 // be to set the raw memory state when generating a SafepointNode. However
1250 // this will require extensive changes to the loop optimization in order to
1251 // prevent a degradation of the optimization.
1252 // See comment in memnode.hpp, around line 227 in class LoadPNode.
1253 Node *eden_end = make_load(ctrl, mem, eden_end_adr, 0, TypeRawPtr::BOTTOM, T_ADDRESS);
1254
1255 // allocate the Region and Phi nodes for the result
1256 result_region = new (C) RegionNode(3);
1257 result_phi_rawmem = new (C) PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1258 result_phi_rawoop = new (C) PhiNode(result_region, TypeRawPtr::BOTTOM);
1259 result_phi_i_o = new (C) PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch
1260
1261 // We need a Region for the loop-back contended case.
1262 enum { fall_in_path = 1, contended_loopback_path = 2 };
1263 Node *contended_region;
1264 Node *contended_phi_rawmem;
1265 if (UseTLAB) {
1266 contended_region = toobig_false;
1267 contended_phi_rawmem = mem;
1268 } else {
1269 contended_region = new (C) RegionNode(3);
1270 contended_phi_rawmem = new (C) PhiNode(contended_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1271 // Now handle the passing-too-big test. We fall into the contended
1272 // loop-back merge point.
1273 contended_region ->init_req(fall_in_path, toobig_false);
1274 contended_phi_rawmem->init_req(fall_in_path, mem);
1275 transform_later(contended_region);
1276 transform_later(contended_phi_rawmem);
1277 }
1278
1279 // Load(-locked) the heap top.
1280 // See note above concerning the control input when using a TLAB
1281 Node *old_eden_top = UseTLAB
1282 ? new (C) LoadPNode (ctrl, contended_phi_rawmem, eden_top_adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, MemNode::unordered)
1283 : new (C) LoadPLockedNode(contended_region, contended_phi_rawmem, eden_top_adr, MemNode::acquire);
1284
1285 transform_later(old_eden_top);
1286 // Add to heap top to get a new heap top
1287 Node *new_eden_top = new (C) AddPNode(top(), old_eden_top, size_in_bytes);
1288 transform_later(new_eden_top);
1289 // Check for needing a GC; compare against heap end
1290 Node *needgc_cmp = new (C) CmpPNode(new_eden_top, eden_end);
1291 transform_later(needgc_cmp);
1292 Node *needgc_bol = new (C) BoolNode(needgc_cmp, BoolTest::ge);
1293 transform_later(needgc_bol);
1294 IfNode *needgc_iff = new (C) IfNode(contended_region, needgc_bol, PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN);
1295 transform_later(needgc_iff);
1296
1297 // Plug the failing-heap-space-need-gc test into the slow-path region
1298 Node *needgc_true = new (C) IfTrueNode(needgc_iff);
1299 transform_later(needgc_true);
1300 if (initial_slow_test) {
1301 slow_region->init_req(need_gc_path, needgc_true);
1302 // This completes all paths into the slow merge point
1303 transform_later(slow_region);
1304 } else { // No initial slow path needed!
1305 // Just fall from the need-GC path straight into the VM call.
1306 slow_region = needgc_true;
1307 }
1308 // No need for a GC. Setup for the Store-Conditional
1309 Node *needgc_false = new (C) IfFalseNode(needgc_iff);
1310 transform_later(needgc_false);
1311
1312 // Grab regular I/O before optional prefetch may change it.
1313 // Slow-path does no I/O so just set it to the original I/O.
1314 result_phi_i_o->init_req(slow_result_path, i_o);
1315
1316 i_o = prefetch_allocation(i_o, needgc_false, contended_phi_rawmem,
1317 old_eden_top, new_eden_top, length);
1318
1319 // Name successful fast-path variables
1320 Node* fast_oop = old_eden_top;
1321 Node* fast_oop_ctrl;
1322 Node* fast_oop_rawmem;
1323
1324 // Store (-conditional) the modified eden top back down.
1325 // StorePConditional produces flags for a test PLUS a modified raw
1326 // memory state.
1327 if (UseTLAB) {
1328 Node* store_eden_top =
1329 new (C) StorePNode(needgc_false, contended_phi_rawmem, eden_top_adr,
1330 TypeRawPtr::BOTTOM, new_eden_top, MemNode::unordered);
1331 transform_later(store_eden_top);
1332 fast_oop_ctrl = needgc_false; // No contention, so this is the fast path
1333 fast_oop_rawmem = store_eden_top;
1334 } else {
1335 Node* store_eden_top =
1336 new (C) StorePConditionalNode(needgc_false, contended_phi_rawmem, eden_top_adr,
1337 new_eden_top, fast_oop/*old_eden_top*/);
1338 transform_later(store_eden_top);
1339 Node *contention_check = new (C) BoolNode(store_eden_top, BoolTest::ne);
1340 transform_later(contention_check);
1341 store_eden_top = new (C) SCMemProjNode(store_eden_top);
1342 transform_later(store_eden_top);
1343
1344 // If not using TLABs, check to see if there was contention.
1345 IfNode *contention_iff = new (C) IfNode (needgc_false, contention_check, PROB_MIN, COUNT_UNKNOWN);
1346 transform_later(contention_iff);
1347 Node *contention_true = new (C) IfTrueNode(contention_iff);
1348 transform_later(contention_true);
1349 // If contention, loopback and try again.
1350 contended_region->init_req(contended_loopback_path, contention_true);
1351 contended_phi_rawmem->init_req(contended_loopback_path, store_eden_top);
1352
1353 // Fast-path succeeded with no contention!
1354 Node *contention_false = new (C) IfFalseNode(contention_iff);
1355 transform_later(contention_false);
1356 fast_oop_ctrl = contention_false;
1357
1358 // Bump total allocated bytes for this thread
1359 Node* thread = new (C) ThreadLocalNode();
1360 transform_later(thread);
1361 Node* alloc_bytes_adr = basic_plus_adr(top()/*not oop*/, thread,
1362 in_bytes(JavaThread::allocated_bytes_offset()));
1363 Node* alloc_bytes = make_load(fast_oop_ctrl, store_eden_top, alloc_bytes_adr,
1364 0, TypeLong::LONG, T_LONG);
1365 #ifdef _LP64
1366 Node* alloc_size = size_in_bytes;
1367 #else
1368 Node* alloc_size = new (C) ConvI2LNode(size_in_bytes);
1369 transform_later(alloc_size);
1370 #endif
1371 Node* new_alloc_bytes = new (C) AddLNode(alloc_bytes, alloc_size);
1372 transform_later(new_alloc_bytes);
1373 fast_oop_rawmem = make_store(fast_oop_ctrl, store_eden_top, alloc_bytes_adr,
1374 0, new_alloc_bytes, T_LONG);
1375 }
1376
1377 InitializeNode* init = alloc->initialization();
1378 fast_oop_rawmem = initialize_object(alloc,
1379 fast_oop_ctrl, fast_oop_rawmem, fast_oop,
1380 klass_node, length, size_in_bytes);
1381
1382 // If initialization is performed by an array copy, any required
1383 // MemBarStoreStore was already added. If the object does not
1384 // escape no need for a MemBarStoreStore. Otherwise we need a
1385 // MemBarStoreStore so that stores that initialize this object
1386 // can't be reordered with a subsequent store that makes this
1387 // object accessible by other threads.
1388 #ifndef AARCH64
1389 if (init == NULL || (!init->is_complete_with_arraycopy() && !init->does_not_escape())) {
1390 #else
1391 if (!alloc->does_not_escape_thread() &&
1392 (init == NULL || !init->is_complete_with_arraycopy())) {
1393 #endif
1394 if (init == NULL || init->req() < InitializeNode::RawStores) {
1395 // No InitializeNode or no stores captured by zeroing
1396 // elimination. Simply add the MemBarStoreStore after object
1397 // initialization.
1398 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1399 transform_later(mb);
1400
1401 mb->init_req(TypeFunc::Memory, fast_oop_rawmem);
1402 mb->init_req(TypeFunc::Control, fast_oop_ctrl);
1403 fast_oop_ctrl = new (C) ProjNode(mb,TypeFunc::Control);
1404 transform_later(fast_oop_ctrl);
1405 fast_oop_rawmem = new (C) ProjNode(mb,TypeFunc::Memory);
1406 transform_later(fast_oop_rawmem);
1407 } else {
1408 // Add the MemBarStoreStore after the InitializeNode so that
1409 // all stores performing the initialization that were moved
1410 // before the InitializeNode happen before the storestore
1411 // barrier.
1412
1413 Node* init_ctrl = init->proj_out(TypeFunc::Control);
1414 Node* init_mem = init->proj_out(TypeFunc::Memory);
1415
1416 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1417 transform_later(mb);
1418
1419 Node* ctrl = new (C) ProjNode(init,TypeFunc::Control);
1420 transform_later(ctrl);
1421 Node* mem = new (C) ProjNode(init,TypeFunc::Memory);
1422 transform_later(mem);
1423
1424 // The MemBarStoreStore depends on control and memory coming
1425 // from the InitializeNode
1426 mb->init_req(TypeFunc::Memory, mem);
1427 mb->init_req(TypeFunc::Control, ctrl);
1428
1429 ctrl = new (C) ProjNode(mb,TypeFunc::Control);
1430 transform_later(ctrl);
1431 mem = new (C) ProjNode(mb,TypeFunc::Memory);
1432 transform_later(mem);
1433
1434 // All nodes that depended on the InitializeNode for control
1435 // and memory must now depend on the MemBarNode that itself
1436 // depends on the InitializeNode
1437 _igvn.replace_node(init_ctrl, ctrl);
1438 _igvn.replace_node(init_mem, mem);
1439 }
1440 }
1441
1442 if (C->env()->dtrace_extended_probes()) {
1443 // Slow-path call
1444 int size = TypeFunc::Parms + 2;
1445 CallLeafNode *call = new (C) CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(),
1446 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_object_alloc_base),
1447 "dtrace_object_alloc",
1448 TypeRawPtr::BOTTOM);
1449
1450 // Get base of thread-local storage area
1451 Node* thread = new (C) ThreadLocalNode();
1452 transform_later(thread);
1453
1454 call->init_req(TypeFunc::Parms+0, thread);
1455 call->init_req(TypeFunc::Parms+1, fast_oop);
1456 call->init_req(TypeFunc::Control, fast_oop_ctrl);
1457 call->init_req(TypeFunc::I_O , top()); // does no i/o
1458 call->init_req(TypeFunc::Memory , fast_oop_rawmem);
1459 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1460 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1461 transform_later(call);
1462 fast_oop_ctrl = new (C) ProjNode(call,TypeFunc::Control);
1463 transform_later(fast_oop_ctrl);
1464 fast_oop_rawmem = new (C) ProjNode(call,TypeFunc::Memory);
1465 transform_later(fast_oop_rawmem);
1466 }
1467
1468 // Plug in the successful fast-path into the result merge point
1469 result_region ->init_req(fast_result_path, fast_oop_ctrl);
1470 result_phi_rawoop->init_req(fast_result_path, fast_oop);
1471 result_phi_i_o ->init_req(fast_result_path, i_o);
1472 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem);
1473 } else {
1474 slow_region = ctrl;
1475 result_phi_i_o = i_o; // Rename it to use in the following code.
1476 }
1477
1478 // Generate slow-path call
1479 CallNode *call = new (C) CallStaticJavaNode(slow_call_type, slow_call_address,
1480 OptoRuntime::stub_name(slow_call_address),
1481 alloc->jvms()->bci(),
1482 TypePtr::BOTTOM);
1483 call->init_req( TypeFunc::Control, slow_region );
1484 call->init_req( TypeFunc::I_O , top() ) ; // does no i/o
1485 call->init_req( TypeFunc::Memory , slow_mem ); // may gc ptrs
1486 call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) );
1487 call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) );
1488
1489 call->init_req(TypeFunc::Parms+0, klass_node);
1490 if (length != NULL) {
1491 call->init_req(TypeFunc::Parms+1, length);
1492 }
1493
1494 // Copy debug information and adjust JVMState information, then replace
1495 // allocate node with the call
1496 copy_call_debug_info((CallNode *) alloc, call);
1497 if (!always_slow) {
1498 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1499 } else {
1500 // Hook i_o projection to avoid its elimination during allocation
1501 // replacement (when only a slow call is generated).
1502 call->set_req(TypeFunc::I_O, result_phi_i_o);
1503 }
1504 _igvn.replace_node(alloc, call);
1505 transform_later(call);
1506
1507 // Identify the output projections from the allocate node and
1508 // adjust any references to them.
1509 // The control and io projections look like:
1510 //
1511 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl)
1512 // Allocate Catch
1513 // ^---Proj(io) <-------+ ^---CatchProj(io)
1514 //
1515 // We are interested in the CatchProj nodes.
1516 //
1517 extract_call_projections(call);
1518
1519 // An allocate node has separate memory projections for the uses on
1520 // the control and i_o paths. Replace the control memory projection with
1521 // result_phi_rawmem (unless we are only generating a slow call when
1522 // both memory projections are combined)
1523 if (!always_slow && _memproj_fallthrough != NULL) {
1524 for (DUIterator_Fast imax, i = _memproj_fallthrough->fast_outs(imax); i < imax; i++) {
1525 Node *use = _memproj_fallthrough->fast_out(i);
1526 _igvn.rehash_node_delayed(use);
1527 imax -= replace_input(use, _memproj_fallthrough, result_phi_rawmem);
1528 // back up iterator
1529 --i;
1530 }
1531 }
1532 // Now change uses of _memproj_catchall to use _memproj_fallthrough and delete
1533 // _memproj_catchall so we end up with a call that has only 1 memory projection.
1534 if (_memproj_catchall != NULL ) {
1535 if (_memproj_fallthrough == NULL) {
1536 _memproj_fallthrough = new (C) ProjNode(call, TypeFunc::Memory);
1537 transform_later(_memproj_fallthrough);
1538 }
1539 for (DUIterator_Fast imax, i = _memproj_catchall->fast_outs(imax); i < imax; i++) {
1540 Node *use = _memproj_catchall->fast_out(i);
1541 _igvn.rehash_node_delayed(use);
1542 imax -= replace_input(use, _memproj_catchall, _memproj_fallthrough);
1543 // back up iterator
1544 --i;
1545 }
1546 assert(_memproj_catchall->outcnt() == 0, "all uses must be deleted");
1547 _igvn.remove_dead_node(_memproj_catchall);
1548 }
1549
1550 // An allocate node has separate i_o projections for the uses on the control
1551 // and i_o paths. Always replace the control i_o projection with result i_o
1552 // otherwise incoming i_o become dead when only a slow call is generated
1553 // (it is different from memory projections where both projections are
1554 // combined in such case).
1555 if (_ioproj_fallthrough != NULL) {
1556 for (DUIterator_Fast imax, i = _ioproj_fallthrough->fast_outs(imax); i < imax; i++) {
1557 Node *use = _ioproj_fallthrough->fast_out(i);
1558 _igvn.rehash_node_delayed(use);
1559 imax -= replace_input(use, _ioproj_fallthrough, result_phi_i_o);
1560 // back up iterator
1561 --i;
1562 }
1563 }
1564 // Now change uses of _ioproj_catchall to use _ioproj_fallthrough and delete
1565 // _ioproj_catchall so we end up with a call that has only 1 i_o projection.
1566 if (_ioproj_catchall != NULL ) {
1567 if (_ioproj_fallthrough == NULL) {
1568 _ioproj_fallthrough = new (C) ProjNode(call, TypeFunc::I_O);
1569 transform_later(_ioproj_fallthrough);
1570 }
1571 for (DUIterator_Fast imax, i = _ioproj_catchall->fast_outs(imax); i < imax; i++) {
1572 Node *use = _ioproj_catchall->fast_out(i);
1573 _igvn.rehash_node_delayed(use);
1574 imax -= replace_input(use, _ioproj_catchall, _ioproj_fallthrough);
1575 // back up iterator
1576 --i;
1577 }
1578 assert(_ioproj_catchall->outcnt() == 0, "all uses must be deleted");
1579 _igvn.remove_dead_node(_ioproj_catchall);
1580 }
1581
1582 // if we generated only a slow call, we are done
1583 if (always_slow) {
1584 // Now we can unhook i_o.
1585 if (result_phi_i_o->outcnt() > 1) {
1586 call->set_req(TypeFunc::I_O, top());
1587 } else {
1588 assert(result_phi_i_o->unique_ctrl_out() == call, "");
1589 // Case of new array with negative size known during compilation.
1590 // AllocateArrayNode::Ideal() optimization disconnect unreachable
1591 // following code since call to runtime will throw exception.
1592 // As result there will be no users of i_o after the call.
1593 // Leave i_o attached to this call to avoid problems in preceding graph.
1594 }
1595 return;
1596 }
1597
1598
1599 if (_fallthroughcatchproj != NULL) {
1600 ctrl = _fallthroughcatchproj->clone();
1601 transform_later(ctrl);
1602 _igvn.replace_node(_fallthroughcatchproj, result_region);
1603 } else {
1604 ctrl = top();
1605 }
1606 Node *slow_result;
1607 if (_resproj == NULL) {
1608 // no uses of the allocation result
1609 slow_result = top();
1610 } else {
1611 slow_result = _resproj->clone();
1612 transform_later(slow_result);
1613 _igvn.replace_node(_resproj, result_phi_rawoop);
1614 }
1615
1616 // Plug slow-path into result merge point
1617 result_region ->init_req( slow_result_path, ctrl );
1618 result_phi_rawoop->init_req( slow_result_path, slow_result);
1619 result_phi_rawmem->init_req( slow_result_path, _memproj_fallthrough );
1620 transform_later(result_region);
1621 transform_later(result_phi_rawoop);
1622 transform_later(result_phi_rawmem);
1623 transform_later(result_phi_i_o);
1624 // This completes all paths into the result merge point
1625 }
1626
1627
1628 // Helper for PhaseMacroExpand::expand_allocate_common.
1629 // Initializes the newly-allocated storage.
1630 Node*
1631 PhaseMacroExpand::initialize_object(AllocateNode* alloc,
1632 Node* control, Node* rawmem, Node* object,
1633 Node* klass_node, Node* length,
1634 Node* size_in_bytes) {
1635 InitializeNode* init = alloc->initialization();
1636 // Store the klass & mark bits
1637 Node* mark_node = NULL;
1638 // For now only enable fast locking for non-array types
1639 if (UseBiasedLocking && (length == NULL)) {
1640 mark_node = make_load(control, rawmem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
1641 } else {
1642 mark_node = makecon(TypeRawPtr::make((address)markOopDesc::prototype()));
1643 }
1644 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS);
1645
1646 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
1647 int header_size = alloc->minimum_header_size(); // conservatively small
1648
1649 // Array length
1650 if (length != NULL) { // Arrays need length field
1651 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
1652 // conservatively small header size:
1653 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1654 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
1655 if (k->is_array_klass()) // we know the exact header size in most cases:
1656 header_size = Klass::layout_helper_header_size(k->layout_helper());
1657 }
1658
1659 // Clear the object body, if necessary.
1660 if (init == NULL) {
1661 // The init has somehow disappeared; be cautious and clear everything.
1662 //
1663 // This can happen if a node is allocated but an uncommon trap occurs
1664 // immediately. In this case, the Initialize gets associated with the
1665 // trap, and may be placed in a different (outer) loop, if the Allocate
1666 // is in a loop. If (this is rare) the inner loop gets unrolled, then
1667 // there can be two Allocates to one Initialize. The answer in all these
1668 // edge cases is safety first. It is always safe to clear immediately
1669 // within an Allocate, and then (maybe or maybe not) clear some more later.
1670 if (!ZeroTLAB)
1671 rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
1672 header_size, size_in_bytes,
1673 &_igvn);
1674 } else {
1675 if (!init->is_complete()) {
1676 // Try to win by zeroing only what the init does not store.
1677 // We can also try to do some peephole optimizations,
1678 // such as combining some adjacent subword stores.
1679 rawmem = init->complete_stores(control, rawmem, object,
1680 header_size, size_in_bytes, &_igvn);
1681 }
1682 // We have no more use for this link, since the AllocateNode goes away:
1683 init->set_req(InitializeNode::RawAddress, top());
1684 // (If we keep the link, it just confuses the register allocator,
1685 // who thinks he sees a real use of the address by the membar.)
1686 }
1687
1688 return rawmem;
1689 }
1690
1691 // Generate prefetch instructions for next allocations.
1692 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false,
1693 Node*& contended_phi_rawmem,
1694 Node* old_eden_top, Node* new_eden_top,
1695 Node* length) {
1696 enum { fall_in_path = 1, pf_path = 2 };
1697 if( UseTLAB && AllocatePrefetchStyle == 2 ) {
1698 // Generate prefetch allocation with watermark check.
1699 // As an allocation hits the watermark, we will prefetch starting
1700 // at a "distance" away from watermark.
1701
1702 Node *pf_region = new (C) RegionNode(3);
1703 Node *pf_phi_rawmem = new (C) PhiNode( pf_region, Type::MEMORY,
1704 TypeRawPtr::BOTTOM );
1705 // I/O is used for Prefetch
1706 Node *pf_phi_abio = new (C) PhiNode( pf_region, Type::ABIO );
1707
1708 Node *thread = new (C) ThreadLocalNode();
1709 transform_later(thread);
1710
1711 Node *eden_pf_adr = new (C) AddPNode( top()/*not oop*/, thread,
1712 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) );
1713 transform_later(eden_pf_adr);
1714
1715 Node *old_pf_wm = new (C) LoadPNode(needgc_false,
1716 contended_phi_rawmem, eden_pf_adr,
1717 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM,
1718 MemNode::unordered);
1719 transform_later(old_pf_wm);
1720
1721 // check against new_eden_top
1722 Node *need_pf_cmp = new (C) CmpPNode( new_eden_top, old_pf_wm );
1723 transform_later(need_pf_cmp);
1724 Node *need_pf_bol = new (C) BoolNode( need_pf_cmp, BoolTest::ge );
1725 transform_later(need_pf_bol);
1726 IfNode *need_pf_iff = new (C) IfNode( needgc_false, need_pf_bol,
1727 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN );
1728 transform_later(need_pf_iff);
1729
1730 // true node, add prefetchdistance
1731 Node *need_pf_true = new (C) IfTrueNode( need_pf_iff );
1732 transform_later(need_pf_true);
1733
1734 Node *need_pf_false = new (C) IfFalseNode( need_pf_iff );
1735 transform_later(need_pf_false);
1736
1737 Node *new_pf_wmt = new (C) AddPNode( top(), old_pf_wm,
1738 _igvn.MakeConX(AllocatePrefetchDistance) );
1739 transform_later(new_pf_wmt );
1740 new_pf_wmt->set_req(0, need_pf_true);
1741
1742 Node *store_new_wmt = new (C) StorePNode(need_pf_true,
1743 contended_phi_rawmem, eden_pf_adr,
1744 TypeRawPtr::BOTTOM, new_pf_wmt,
1745 MemNode::unordered);
1746 transform_later(store_new_wmt);
1747
1748 // adding prefetches
1749 pf_phi_abio->init_req( fall_in_path, i_o );
1750
1751 Node *prefetch_adr;
1752 Node *prefetch;
1753 uint lines = AllocatePrefetchDistance / AllocatePrefetchStepSize;
1754 uint step_size = AllocatePrefetchStepSize;
1755 uint distance = 0;
1756
1757 for ( uint i = 0; i < lines; i++ ) {
1758 prefetch_adr = new (C) AddPNode( old_pf_wm, new_pf_wmt,
1759 _igvn.MakeConX(distance) );
1760 transform_later(prefetch_adr);
1761 prefetch = new (C) PrefetchAllocationNode( i_o, prefetch_adr );
1762 transform_later(prefetch);
1763 distance += step_size;
1764 i_o = prefetch;
1765 }
1766 pf_phi_abio->set_req( pf_path, i_o );
1767
1768 pf_region->init_req( fall_in_path, need_pf_false );
1769 pf_region->init_req( pf_path, need_pf_true );
1770
1771 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem );
1772 pf_phi_rawmem->init_req( pf_path, store_new_wmt );
1773
1774 transform_later(pf_region);
1775 transform_later(pf_phi_rawmem);
1776 transform_later(pf_phi_abio);
1777
1778 needgc_false = pf_region;
1779 contended_phi_rawmem = pf_phi_rawmem;
1780 i_o = pf_phi_abio;
1781 } else if( UseTLAB && AllocatePrefetchStyle == 3 ) {
1782 // Insert a prefetch for each allocation.
1783 // This code is used to generate 1 prefetch instruction per cache line.
1784 Node *pf_region = new (C) RegionNode(3);
1785 Node *pf_phi_rawmem = new (C) PhiNode( pf_region, Type::MEMORY,
1786 TypeRawPtr::BOTTOM );
1787
1788 // Generate several prefetch instructions.
1789 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1790 uint step_size = AllocatePrefetchStepSize;
1791 uint distance = AllocatePrefetchDistance;
1792
1793 // Next cache address.
1794 Node *cache_adr = new (C) AddPNode(old_eden_top, old_eden_top,
1795 _igvn.MakeConX(distance));
1796 transform_later(cache_adr);
1797 cache_adr = new (C) CastP2XNode(needgc_false, cache_adr);
1798 transform_later(cache_adr);
1799 // Address is aligned to execute prefetch to the beginning of cache line size
1800 // (it is important when BIS instruction is used on SPARC as prefetch).
1801 Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1));
1802 cache_adr = new (C) AndXNode(cache_adr, mask);
1803 transform_later(cache_adr);
1804 cache_adr = new (C) CastX2PNode(cache_adr);
1805 transform_later(cache_adr);
1806
1807 // Prefetch
1808 Node *prefetch = new (C) PrefetchAllocationNode( contended_phi_rawmem, cache_adr );
1809 prefetch->set_req(0, needgc_false);
1810 transform_later(prefetch);
1811 contended_phi_rawmem = prefetch;
1812 Node *prefetch_adr;
1813 distance = step_size;
1814 for ( uint i = 1; i < lines; i++ ) {
1815 prefetch_adr = new (C) AddPNode( cache_adr, cache_adr,
1816 _igvn.MakeConX(distance) );
1817 transform_later(prefetch_adr);
1818 prefetch = new (C) PrefetchAllocationNode( contended_phi_rawmem, prefetch_adr );
1819 transform_later(prefetch);
1820 distance += step_size;
1821 contended_phi_rawmem = prefetch;
1822 }
1823 } else if( AllocatePrefetchStyle > 0 ) {
1824 // Insert a prefetch for each allocation only on the fast-path
1825 Node *prefetch_adr;
1826 Node *prefetch;
1827 // Generate several prefetch instructions.
1828 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1829 uint step_size = AllocatePrefetchStepSize;
1830 uint distance = AllocatePrefetchDistance;
1831 for ( uint i = 0; i < lines; i++ ) {
1832 prefetch_adr = new (C) AddPNode( old_eden_top, new_eden_top,
1833 _igvn.MakeConX(distance) );
1834 transform_later(prefetch_adr);
1835 prefetch = new (C) PrefetchAllocationNode( i_o, prefetch_adr );
1836 // Do not let it float too high, since if eden_top == eden_end,
1837 // both might be null.
1838 if( i == 0 ) { // Set control for first prefetch, next follows it
1839 prefetch->init_req(0, needgc_false);
1840 }
1841 transform_later(prefetch);
1842 distance += step_size;
1843 i_o = prefetch;
1844 }
1845 }
1846 return i_o;
1847 }
1848
1849
1850 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) {
1851 expand_allocate_common(alloc, NULL,
1852 OptoRuntime::new_instance_Type(),
1853 OptoRuntime::new_instance_Java());
1854 }
1855
1856 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) {
1857 Node* length = alloc->in(AllocateNode::ALength);
1858 InitializeNode* init = alloc->initialization();
1859 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1860 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
1861 address slow_call_address; // Address of slow call
1862 if (init != NULL && init->is_complete_with_arraycopy() &&
1863 k->is_type_array_klass()) {
1864 // Don't zero type array during slow allocation in VM since
1865 // it will be initialized later by arraycopy in compiled code.
1866 slow_call_address = OptoRuntime::new_array_nozero_Java();
1867 } else {
1868 slow_call_address = OptoRuntime::new_array_Java();
1869 }
1870 expand_allocate_common(alloc, length,
1871 OptoRuntime::new_array_Type(),
1872 slow_call_address);
1873 }
1874
1875 //-------------------mark_eliminated_box----------------------------------
1876 //
1877 // During EA obj may point to several objects but after few ideal graph
1878 // transformations (CCP) it may point to only one non escaping object
1879 // (but still using phi), corresponding locks and unlocks will be marked
1880 // for elimination. Later obj could be replaced with a new node (new phi)
1881 // and which does not have escape information. And later after some graph
1882 // reshape other locks and unlocks (which were not marked for elimination
1883 // before) are connected to this new obj (phi) but they still will not be
1884 // marked for elimination since new obj has no escape information.
1885 // Mark all associated (same box and obj) lock and unlock nodes for
1886 // elimination if some of them marked already.
1887 void PhaseMacroExpand::mark_eliminated_box(Node* oldbox, Node* obj) {
1888 if (oldbox->as_BoxLock()->is_eliminated())
1889 return; // This BoxLock node was processed already.
1890
1891 // New implementation (EliminateNestedLocks) has separate BoxLock
1892 // node for each locked region so mark all associated locks/unlocks as
1893 // eliminated even if different objects are referenced in one locked region
1894 // (for example, OSR compilation of nested loop inside locked scope).
1895 if (EliminateNestedLocks ||
1896 oldbox->as_BoxLock()->is_simple_lock_region(NULL, obj)) {
1897 // Box is used only in one lock region. Mark this box as eliminated.
1898 _igvn.hash_delete(oldbox);
1899 oldbox->as_BoxLock()->set_eliminated(); // This changes box's hash value
1900 _igvn.hash_insert(oldbox);
1901
1902 for (uint i = 0; i < oldbox->outcnt(); i++) {
1903 Node* u = oldbox->raw_out(i);
1904 if (u->is_AbstractLock() && !u->as_AbstractLock()->is_non_esc_obj()) {
1905 AbstractLockNode* alock = u->as_AbstractLock();
1906 // Check lock's box since box could be referenced by Lock's debug info.
1907 if (alock->box_node() == oldbox) {
1908 // Mark eliminated all related locks and unlocks.
1909 #ifdef ASSERT
1910 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc4");
1911 #endif
1912 alock->set_non_esc_obj();
1913 }
1914 }
1915 }
1916 return;
1917 }
1918
1919 // Create new "eliminated" BoxLock node and use it in monitor debug info
1920 // instead of oldbox for the same object.
1921 BoxLockNode* newbox = oldbox->clone()->as_BoxLock();
1922
1923 // Note: BoxLock node is marked eliminated only here and it is used
1924 // to indicate that all associated lock and unlock nodes are marked
1925 // for elimination.
1926 newbox->set_eliminated();
1927 transform_later(newbox);
1928
1929 // Replace old box node with new box for all users of the same object.
1930 for (uint i = 0; i < oldbox->outcnt();) {
1931 bool next_edge = true;
1932
1933 Node* u = oldbox->raw_out(i);
1934 if (u->is_AbstractLock()) {
1935 AbstractLockNode* alock = u->as_AbstractLock();
1936 if (alock->box_node() == oldbox && alock->obj_node()->eqv_uncast(obj)) {
1937 // Replace Box and mark eliminated all related locks and unlocks.
1938 #ifdef ASSERT
1939 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc5");
1940 #endif
1941 alock->set_non_esc_obj();
1942 _igvn.rehash_node_delayed(alock);
1943 alock->set_box_node(newbox);
1944 next_edge = false;
1945 }
1946 }
1947 if (u->is_FastLock() && u->as_FastLock()->obj_node()->eqv_uncast(obj)) {
1948 FastLockNode* flock = u->as_FastLock();
1949 assert(flock->box_node() == oldbox, "sanity");
1950 _igvn.rehash_node_delayed(flock);
1951 flock->set_box_node(newbox);
1952 next_edge = false;
1953 }
1954
1955 // Replace old box in monitor debug info.
1956 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) {
1957 SafePointNode* sfn = u->as_SafePoint();
1958 JVMState* youngest_jvms = sfn->jvms();
1959 int max_depth = youngest_jvms->depth();
1960 for (int depth = 1; depth <= max_depth; depth++) {
1961 JVMState* jvms = youngest_jvms->of_depth(depth);
1962 int num_mon = jvms->nof_monitors();
1963 // Loop over monitors
1964 for (int idx = 0; idx < num_mon; idx++) {
1965 Node* obj_node = sfn->monitor_obj(jvms, idx);
1966 Node* box_node = sfn->monitor_box(jvms, idx);
1967 if (box_node == oldbox && obj_node->eqv_uncast(obj)) {
1968 int j = jvms->monitor_box_offset(idx);
1969 _igvn.replace_input_of(u, j, newbox);
1970 next_edge = false;
1971 }
1972 }
1973 }
1974 }
1975 if (next_edge) i++;
1976 }
1977 }
1978
1979 //-----------------------mark_eliminated_locking_nodes-----------------------
1980 void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) {
1981 if (EliminateNestedLocks) {
1982 if (alock->is_nested()) {
1983 assert(alock->box_node()->as_BoxLock()->is_eliminated(), "sanity");
1984 return;
1985 } else if (!alock->is_non_esc_obj()) { // Not eliminated or coarsened
1986 // Only Lock node has JVMState needed here.
1987 // Not that preceding claim is documented anywhere else.
1988 if (alock->jvms() != NULL) {
1989 if (alock->as_Lock()->is_nested_lock_region()) {
1990 // Mark eliminated related nested locks and unlocks.
1991 Node* obj = alock->obj_node();
1992 BoxLockNode* box_node = alock->box_node()->as_BoxLock();
1993 assert(!box_node->is_eliminated(), "should not be marked yet");
1994 // Note: BoxLock node is marked eliminated only here
1995 // and it is used to indicate that all associated lock
1996 // and unlock nodes are marked for elimination.
1997 box_node->set_eliminated(); // Box's hash is always NO_HASH here
1998 for (uint i = 0; i < box_node->outcnt(); i++) {
1999 Node* u = box_node->raw_out(i);
2000 if (u->is_AbstractLock()) {
2001 alock = u->as_AbstractLock();
2002 if (alock->box_node() == box_node) {
2003 // Verify that this Box is referenced only by related locks.
2004 assert(alock->obj_node()->eqv_uncast(obj), "");
2005 // Mark all related locks and unlocks.
2006 #ifdef ASSERT
2007 alock->log_lock_optimization(C, "eliminate_lock_set_nested");
2008 #endif
2009 alock->set_nested();
2010 }
2011 }
2012 }
2013 } else {
2014 #ifdef ASSERT
2015 alock->log_lock_optimization(C, "eliminate_lock_NOT_nested_lock_region");
2016 if (C->log() != NULL)
2017 alock->as_Lock()->is_nested_lock_region(C); // rerun for debugging output
2018 #endif
2019 }
2020 }
2021 return;
2022 }
2023 // Process locks for non escaping object
2024 assert(alock->is_non_esc_obj(), "");
2025 } // EliminateNestedLocks
2026
2027 if (alock->is_non_esc_obj()) { // Lock is used for non escaping object
2028 // Look for all locks of this object and mark them and
2029 // corresponding BoxLock nodes as eliminated.
2030 Node* obj = alock->obj_node();
2031 for (uint j = 0; j < obj->outcnt(); j++) {
2032 Node* o = obj->raw_out(j);
2033 if (o->is_AbstractLock() &&
2034 o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) {
2035 alock = o->as_AbstractLock();
2036 Node* box = alock->box_node();
2037 // Replace old box node with new eliminated box for all users
2038 // of the same object and mark related locks as eliminated.
2039 mark_eliminated_box(box, obj);
2040 }
2041 }
2042 }
2043 }
2044
2045 // we have determined that this lock/unlock can be eliminated, we simply
2046 // eliminate the node without expanding it.
2047 //
2048 // Note: The membar's associated with the lock/unlock are currently not
2049 // eliminated. This should be investigated as a future enhancement.
2050 //
2051 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) {
2052
2053 if (!alock->is_eliminated()) {
2054 return false;
2055 }
2056 #ifdef ASSERT
2057 if (!alock->is_coarsened()) {
2058 // Check that new "eliminated" BoxLock node is created.
2059 BoxLockNode* oldbox = alock->box_node()->as_BoxLock();
2060 assert(oldbox->is_eliminated(), "should be done already");
2061 }
2062 #endif
2063
2064 alock->log_lock_optimization(C, "eliminate_lock");
2065
2066 #ifndef PRODUCT
2067 if (PrintEliminateLocks) {
2068 if (alock->is_Lock()) {
2069 tty->print_cr("++++ Eliminated: %d Lock", alock->_idx);
2070 } else {
2071 tty->print_cr("++++ Eliminated: %d Unlock", alock->_idx);
2072 }
2073 }
2074 #endif
2075
2076 Node* mem = alock->in(TypeFunc::Memory);
2077 Node* ctrl = alock->in(TypeFunc::Control);
2078
2079 extract_call_projections(alock);
2080 // There are 2 projections from the lock. The lock node will
2081 // be deleted when its last use is subsumed below.
2082 assert(alock->outcnt() == 2 &&
2083 _fallthroughproj != NULL &&
2084 _memproj_fallthrough != NULL,
2085 "Unexpected projections from Lock/Unlock");
2086
2087 Node* fallthroughproj = _fallthroughproj;
2088 Node* memproj_fallthrough = _memproj_fallthrough;
2089
2090 // The memory projection from a lock/unlock is RawMem
2091 // The input to a Lock is merged memory, so extract its RawMem input
2092 // (unless the MergeMem has been optimized away.)
2093 if (alock->is_Lock()) {
2094 // Seach for MemBarAcquireLock node and delete it also.
2095 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
2096 assert(membar != NULL && membar->Opcode() == Op_MemBarAcquireLock, "");
2097 Node* ctrlproj = membar->proj_out(TypeFunc::Control);
2098 Node* memproj = membar->proj_out(TypeFunc::Memory);
2099 _igvn.replace_node(ctrlproj, fallthroughproj);
2100 _igvn.replace_node(memproj, memproj_fallthrough);
2101
2102 // Delete FastLock node also if this Lock node is unique user
2103 // (a loop peeling may clone a Lock node).
2104 Node* flock = alock->as_Lock()->fastlock_node();
2105 if (flock->outcnt() == 1) {
2106 assert(flock->unique_out() == alock, "sanity");
2107 _igvn.replace_node(flock, top());
2108 }
2109 }
2110
2111 // Seach for MemBarReleaseLock node and delete it also.
2112 if (alock->is_Unlock() && ctrl != NULL && ctrl->is_Proj() &&
2113 ctrl->in(0)->is_MemBar()) {
2114 MemBarNode* membar = ctrl->in(0)->as_MemBar();
2115 assert(membar->Opcode() == Op_MemBarReleaseLock &&
2116 mem->is_Proj() && membar == mem->in(0), "");
2117 _igvn.replace_node(fallthroughproj, ctrl);
2118 _igvn.replace_node(memproj_fallthrough, mem);
2119 fallthroughproj = ctrl;
2120 memproj_fallthrough = mem;
2121 ctrl = membar->in(TypeFunc::Control);
2122 mem = membar->in(TypeFunc::Memory);
2123 }
2124
2125 _igvn.replace_node(fallthroughproj, ctrl);
2126 _igvn.replace_node(memproj_fallthrough, mem);
2127 return true;
2128 }
2129
2130
2131 //------------------------------expand_lock_node----------------------
2132 void PhaseMacroExpand::expand_lock_node(LockNode *lock) {
2133
2134 Node* ctrl = lock->in(TypeFunc::Control);
2135 Node* mem = lock->in(TypeFunc::Memory);
2136 Node* obj = lock->obj_node();
2137 Node* box = lock->box_node();
2138 Node* flock = lock->fastlock_node();
2139
2140 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2141
2142 // Make the merge point
2143 Node *region;
2144 Node *mem_phi;
2145 Node *slow_path;
2146
2147 if (UseOptoBiasInlining) {
2148 /*
2149 * See the full description in MacroAssembler::biased_locking_enter().
2150 *
2151 * if( (mark_word & biased_lock_mask) == biased_lock_pattern ) {
2152 * // The object is biased.
2153 * proto_node = klass->prototype_header;
2154 * o_node = thread | proto_node;
2155 * x_node = o_node ^ mark_word;
2156 * if( (x_node & ~age_mask) == 0 ) { // Biased to the current thread ?
2157 * // Done.
2158 * } else {
2159 * if( (x_node & biased_lock_mask) != 0 ) {
2160 * // The klass's prototype header is no longer biased.
2161 * cas(&mark_word, mark_word, proto_node)
2162 * goto cas_lock;
2163 * } else {
2164 * // The klass's prototype header is still biased.
2165 * if( (x_node & epoch_mask) != 0 ) { // Expired epoch?
2166 * old = mark_word;
2167 * new = o_node;
2168 * } else {
2169 * // Different thread or anonymous biased.
2170 * old = mark_word & (epoch_mask | age_mask | biased_lock_mask);
2171 * new = thread | old;
2172 * }
2173 * // Try to rebias.
2174 * if( cas(&mark_word, old, new) == 0 ) {
2175 * // Done.
2176 * } else {
2177 * goto slow_path; // Failed.
2178 * }
2179 * }
2180 * }
2181 * } else {
2182 * // The object is not biased.
2183 * cas_lock:
2184 * if( FastLock(obj) == 0 ) {
2185 * // Done.
2186 * } else {
2187 * slow_path:
2188 * OptoRuntime::complete_monitor_locking_Java(obj);
2189 * }
2190 * }
2191 */
2192
2193 region = new (C) RegionNode(5);
2194 // create a Phi for the memory state
2195 mem_phi = new (C) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2196
2197 Node* fast_lock_region = new (C) RegionNode(3);
2198 Node* fast_lock_mem_phi = new (C) PhiNode( fast_lock_region, Type::MEMORY, TypeRawPtr::BOTTOM);
2199
2200 // First, check mark word for the biased lock pattern.
2201 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2202
2203 // Get fast path - mark word has the biased lock pattern.
2204 ctrl = opt_bits_test(ctrl, fast_lock_region, 1, mark_node,
2205 markOopDesc::biased_lock_mask_in_place,
2206 markOopDesc::biased_lock_pattern, true);
2207 // fast_lock_region->in(1) is set to slow path.
2208 fast_lock_mem_phi->init_req(1, mem);
2209
2210 // Now check that the lock is biased to the current thread and has
2211 // the same epoch and bias as Klass::_prototype_header.
2212
2213 // Special-case a fresh allocation to avoid building nodes:
2214 Node* klass_node = AllocateNode::Ideal_klass(obj, &_igvn);
2215 if (klass_node == NULL) {
2216 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
2217 klass_node = transform_later(LoadKlassNode::make(_igvn, NULL, mem, k_adr, _igvn.type(k_adr)->is_ptr()));
2218 #ifdef _LP64
2219 if (UseCompressedClassPointers && klass_node->is_DecodeNKlass()) {
2220 assert(klass_node->in(1)->Opcode() == Op_LoadNKlass, "sanity");
2221 klass_node->in(1)->init_req(0, ctrl);
2222 } else
2223 #endif
2224 klass_node->init_req(0, ctrl);
2225 }
2226 Node *proto_node = make_load(ctrl, mem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeX_X, TypeX_X->basic_type());
2227
2228 Node* thread = transform_later(new (C) ThreadLocalNode());
2229 Node* cast_thread = transform_later(new (C) CastP2XNode(ctrl, thread));
2230 Node* o_node = transform_later(new (C) OrXNode(cast_thread, proto_node));
2231 Node* x_node = transform_later(new (C) XorXNode(o_node, mark_node));
2232
2233 // Get slow path - mark word does NOT match the value.
2234 Node* not_biased_ctrl = opt_bits_test(ctrl, region, 3, x_node,
2235 (~markOopDesc::age_mask_in_place), 0);
2236 // region->in(3) is set to fast path - the object is biased to the current thread.
2237 mem_phi->init_req(3, mem);
2238
2239
2240 // Mark word does NOT match the value (thread | Klass::_prototype_header).
2241
2242
2243 // First, check biased pattern.
2244 // Get fast path - _prototype_header has the same biased lock pattern.
2245 ctrl = opt_bits_test(not_biased_ctrl, fast_lock_region, 2, x_node,
2246 markOopDesc::biased_lock_mask_in_place, 0, true);
2247
2248 not_biased_ctrl = fast_lock_region->in(2); // Slow path
2249 // fast_lock_region->in(2) - the prototype header is no longer biased
2250 // and we have to revoke the bias on this object.
2251 // We are going to try to reset the mark of this object to the prototype
2252 // value and fall through to the CAS-based locking scheme.
2253 Node* adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
2254 Node* cas = new (C) StoreXConditionalNode(not_biased_ctrl, mem, adr,
2255 proto_node, mark_node);
2256 transform_later(cas);
2257 Node* proj = transform_later( new (C) SCMemProjNode(cas));
2258 fast_lock_mem_phi->init_req(2, proj);
2259
2260
2261 // Second, check epoch bits.
2262 Node* rebiased_region = new (C) RegionNode(3);
2263 Node* old_phi = new (C) PhiNode( rebiased_region, TypeX_X);
2264 Node* new_phi = new (C) PhiNode( rebiased_region, TypeX_X);
2265
2266 // Get slow path - mark word does NOT match epoch bits.
2267 Node* epoch_ctrl = opt_bits_test(ctrl, rebiased_region, 1, x_node,
2268 markOopDesc::epoch_mask_in_place, 0);
2269 // The epoch of the current bias is not valid, attempt to rebias the object
2270 // toward the current thread.
2271 rebiased_region->init_req(2, epoch_ctrl);
2272 old_phi->init_req(2, mark_node);
2273 new_phi->init_req(2, o_node);
2274
2275 // rebiased_region->in(1) is set to fast path.
2276 // The epoch of the current bias is still valid but we know
2277 // nothing about the owner; it might be set or it might be clear.
2278 Node* cmask = MakeConX(markOopDesc::biased_lock_mask_in_place |
2279 markOopDesc::age_mask_in_place |
2280 markOopDesc::epoch_mask_in_place);
2281 Node* old = transform_later(new (C) AndXNode(mark_node, cmask));
2282 cast_thread = transform_later(new (C) CastP2XNode(ctrl, thread));
2283 Node* new_mark = transform_later(new (C) OrXNode(cast_thread, old));
2284 old_phi->init_req(1, old);
2285 new_phi->init_req(1, new_mark);
2286
2287 transform_later(rebiased_region);
2288 transform_later(old_phi);
2289 transform_later(new_phi);
2290
2291 // Try to acquire the bias of the object using an atomic operation.
2292 // If this fails we will go in to the runtime to revoke the object's bias.
2293 cas = new (C) StoreXConditionalNode(rebiased_region, mem, adr,
2294 new_phi, old_phi);
2295 transform_later(cas);
2296 proj = transform_later( new (C) SCMemProjNode(cas));
2297
2298 // Get slow path - Failed to CAS.
2299 not_biased_ctrl = opt_bits_test(rebiased_region, region, 4, cas, 0, 0);
2300 mem_phi->init_req(4, proj);
2301 // region->in(4) is set to fast path - the object is rebiased to the current thread.
2302
2303 // Failed to CAS.
2304 slow_path = new (C) RegionNode(3);
2305 Node *slow_mem = new (C) PhiNode( slow_path, Type::MEMORY, TypeRawPtr::BOTTOM);
2306
2307 slow_path->init_req(1, not_biased_ctrl); // Capture slow-control
2308 slow_mem->init_req(1, proj);
2309
2310 // Call CAS-based locking scheme (FastLock node).
2311
2312 transform_later(fast_lock_region);
2313 transform_later(fast_lock_mem_phi);
2314
2315 // Get slow path - FastLock failed to lock the object.
2316 ctrl = opt_bits_test(fast_lock_region, region, 2, flock, 0, 0);
2317 mem_phi->init_req(2, fast_lock_mem_phi);
2318 // region->in(2) is set to fast path - the object is locked to the current thread.
2319
2320 slow_path->init_req(2, ctrl); // Capture slow-control
2321 slow_mem->init_req(2, fast_lock_mem_phi);
2322
2323 transform_later(slow_path);
2324 transform_later(slow_mem);
2325 // Reset lock's memory edge.
2326 lock->set_req(TypeFunc::Memory, slow_mem);
2327
2328 } else {
2329 region = new (C) RegionNode(3);
2330 // create a Phi for the memory state
2331 mem_phi = new (C) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2332
2333 // Optimize test; set region slot 2
2334 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0);
2335 mem_phi->init_req(2, mem);
2336 }
2337
2338 // Make slow path call
2339 CallNode *call = make_slow_call( (CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(), OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path, obj, box );
2340
2341 extract_call_projections(call);
2342
2343 // Slow path can only throw asynchronous exceptions, which are always
2344 // de-opted. So the compiler thinks the slow-call can never throw an
2345 // exception. If it DOES throw an exception we would need the debug
2346 // info removed first (since if it throws there is no monitor).
2347 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
2348 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
2349
2350 // Capture slow path
2351 // disconnect fall-through projection from call and create a new one
2352 // hook up users of fall-through projection to region
2353 Node *slow_ctrl = _fallthroughproj->clone();
2354 transform_later(slow_ctrl);
2355 _igvn.hash_delete(_fallthroughproj);
2356 _fallthroughproj->disconnect_inputs(NULL, C);
2357 region->init_req(1, slow_ctrl);
2358 // region inputs are now complete
2359 transform_later(region);
2360 _igvn.replace_node(_fallthroughproj, region);
2361
2362 Node *memproj = transform_later( new(C) ProjNode(call, TypeFunc::Memory) );
2363 mem_phi->init_req(1, memproj );
2364 transform_later(mem_phi);
2365 _igvn.replace_node(_memproj_fallthrough, mem_phi);
2366 }
2367
2368 //------------------------------expand_unlock_node----------------------
2369 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
2370
2371 Node* ctrl = unlock->in(TypeFunc::Control);
2372 Node* mem = unlock->in(TypeFunc::Memory);
2373 Node* obj = unlock->obj_node();
2374 Node* box = unlock->box_node();
2375
2376 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2377
2378 // No need for a null check on unlock
2379
2380 // Make the merge point
2381 Node *region;
2382 Node *mem_phi;
2383
2384 if (UseOptoBiasInlining) {
2385 // Check for biased locking unlock case, which is a no-op.
2386 // See the full description in MacroAssembler::biased_locking_exit().
2387 region = new (C) RegionNode(4);
2388 // create a Phi for the memory state
2389 mem_phi = new (C) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2390 mem_phi->init_req(3, mem);
2391
2392 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2393 ctrl = opt_bits_test(ctrl, region, 3, mark_node,
2394 markOopDesc::biased_lock_mask_in_place,
2395 markOopDesc::biased_lock_pattern);
2396 } else {
2397 region = new (C) RegionNode(3);
2398 // create a Phi for the memory state
2399 mem_phi = new (C) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2400 }
2401
2402 FastUnlockNode *funlock = new (C) FastUnlockNode( ctrl, obj, box );
2403 funlock = transform_later( funlock )->as_FastUnlock();
2404 // Optimize test; set region slot 2
2405 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0);
2406
2407 CallNode *call = make_slow_call( (CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C), "complete_monitor_unlocking_C", slow_path, obj, box );
2408
2409 extract_call_projections(call);
2410
2411 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
2412 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
2413
2414 // No exceptions for unlocking
2415 // Capture slow path
2416 // disconnect fall-through projection from call and create a new one
2417 // hook up users of fall-through projection to region
2418 Node *slow_ctrl = _fallthroughproj->clone();
2419 transform_later(slow_ctrl);
2420 _igvn.hash_delete(_fallthroughproj);
2421 _fallthroughproj->disconnect_inputs(NULL, C);
2422 region->init_req(1, slow_ctrl);
2423 // region inputs are now complete
2424 transform_later(region);
2425 _igvn.replace_node(_fallthroughproj, region);
2426
2427 Node *memproj = transform_later( new(C) ProjNode(call, TypeFunc::Memory) );
2428 mem_phi->init_req(1, memproj );
2429 mem_phi->init_req(2, mem);
2430 transform_later(mem_phi);
2431 _igvn.replace_node(_memproj_fallthrough, mem_phi);
2432 }
2433
2434 //---------------------------eliminate_macro_nodes----------------------
2435 // Eliminate scalar replaced allocations and associated locks.
2436 void PhaseMacroExpand::eliminate_macro_nodes() {
2437 if (C->macro_count() == 0)
2438 return;
2439
2440 // First, attempt to eliminate locks
2441 int cnt = C->macro_count();
2442 for (int i=0; i < cnt; i++) {
2443 Node *n = C->macro_node(i);
2444 if (n->is_AbstractLock()) { // Lock and Unlock nodes
2445 // Before elimination mark all associated (same box and obj)
2446 // lock and unlock nodes.
2447 mark_eliminated_locking_nodes(n->as_AbstractLock());
2448 }
2449 }
2450 bool progress = true;
2451 while (progress) {
2452 progress = false;
2453 for (int i = C->macro_count(); i > 0; i--) {
2454 Node * n = C->macro_node(i-1);
2455 bool success = false;
2456 debug_only(int old_macro_count = C->macro_count(););
2457 if (n->is_AbstractLock()) {
2458 success = eliminate_locking_node(n->as_AbstractLock());
2459 }
2460 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2461 progress = progress || success;
2462 }
2463 }
2464 // Next, attempt to eliminate allocations
2465 _has_locks = false;
2466 progress = true;
2467 while (progress) {
2468 progress = false;
2469 for (int i = C->macro_count(); i > 0; i--) {
2470 Node * n = C->macro_node(i-1);
2471 bool success = false;
2472 debug_only(int old_macro_count = C->macro_count(););
2473 switch (n->class_id()) {
2474 case Node::Class_Allocate:
2475 case Node::Class_AllocateArray:
2476 success = eliminate_allocate_node(n->as_Allocate());
2477 break;
2478 case Node::Class_CallStaticJava:
2479 success = eliminate_boxing_node(n->as_CallStaticJava());
2480 break;
2481 case Node::Class_Lock:
2482 case Node::Class_Unlock:
2483 assert(!n->as_AbstractLock()->is_eliminated(), "sanity");
2484 _has_locks = true;
2485 break;
2486 default:
2487 assert(n->Opcode() == Op_LoopLimit ||
2488 n->Opcode() == Op_Opaque1 ||
2489 n->Opcode() == Op_Opaque2 ||
2490 n->Opcode() == Op_Opaque3, "unknown node type in macro list");
2491 }
2492 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2493 progress = progress || success;
2494 }
2495 }
2496 }
2497
2498 //------------------------------expand_macro_nodes----------------------
2499 // Returns true if a failure occurred.
2500 bool PhaseMacroExpand::expand_macro_nodes() {
2501 // Last attempt to eliminate macro nodes.
2502 eliminate_macro_nodes();
2503
2504 // Make sure expansion will not cause node limit to be exceeded.
2505 // Worst case is a macro node gets expanded into about 50 nodes.
2506 // Allow 50% more for optimization.
2507 if (C->check_node_count(C->macro_count() * 75, "out of nodes before macro expansion" ) )
2508 return true;
2509
2510 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations.
2511 bool progress = true;
2512 while (progress) {
2513 progress = false;
2514 for (int i = C->macro_count(); i > 0; i--) {
2515 Node * n = C->macro_node(i-1);
2516 bool success = false;
2517 debug_only(int old_macro_count = C->macro_count(););
2518 if (n->Opcode() == Op_LoopLimit) {
2519 // Remove it from macro list and put on IGVN worklist to optimize.
2520 C->remove_macro_node(n);
2521 _igvn._worklist.push(n);
2522 success = true;
2523 } else if (n->Opcode() == Op_CallStaticJava) {
2524 // Remove it from macro list and put on IGVN worklist to optimize.
2525 C->remove_macro_node(n);
2526 _igvn._worklist.push(n);
2527 success = true;
2528 } else if (n->Opcode() == Op_Opaque1 || n->Opcode() == Op_Opaque2) {
2529 _igvn.replace_node(n, n->in(1));
2530 success = true;
2531 #if INCLUDE_RTM_OPT
2532 } else if ((n->Opcode() == Op_Opaque3) && ((Opaque3Node*)n)->rtm_opt()) {
2533 assert(C->profile_rtm(), "should be used only in rtm deoptimization code");
2534 assert((n->outcnt() == 1) && n->unique_out()->is_Cmp(), "");
2535 Node* cmp = n->unique_out();
2536 #ifdef ASSERT
2537 // Validate graph.
2538 assert((cmp->outcnt() == 1) && cmp->unique_out()->is_Bool(), "");
2539 BoolNode* bol = cmp->unique_out()->as_Bool();
2540 assert((bol->outcnt() == 1) && bol->unique_out()->is_If() &&
2541 (bol->_test._test == BoolTest::ne), "");
2542 IfNode* ifn = bol->unique_out()->as_If();
2543 assert((ifn->outcnt() == 2) &&
2544 ifn->proj_out(1)->is_uncommon_trap_proj(Deoptimization::Reason_rtm_state_change), "");
2545 #endif
2546 Node* repl = n->in(1);
2547 if (!_has_locks) {
2548 // Remove RTM state check if there are no locks in the code.
2549 // Replace input to compare the same value.
2550 repl = (cmp->in(1) == n) ? cmp->in(2) : cmp->in(1);
2551 }
2552 _igvn.replace_node(n, repl);
2553 success = true;
2554 #endif
2555 }
2556 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2557 progress = progress || success;
2558 }
2559 }
2560
2561 // expand "macro" nodes
2562 // nodes are removed from the macro list as they are processed
2563 while (C->macro_count() > 0) {
2564 int macro_count = C->macro_count();
2565 Node * n = C->macro_node(macro_count-1);
2566 assert(n->is_macro(), "only macro nodes expected here");
2567 if (_igvn.type(n) == Type::TOP || n->in(0)->is_top() ) {
2568 // node is unreachable, so don't try to expand it
2569 C->remove_macro_node(n);
2570 continue;
2571 }
2572 switch (n->class_id()) {
2573 case Node::Class_Allocate:
2574 expand_allocate(n->as_Allocate());
2575 break;
2576 case Node::Class_AllocateArray:
2577 expand_allocate_array(n->as_AllocateArray());
2578 break;
2579 case Node::Class_Lock:
2580 expand_lock_node(n->as_Lock());
2581 break;
2582 case Node::Class_Unlock:
2583 expand_unlock_node(n->as_Unlock());
2584 break;
2585 default:
2586 assert(false, "unknown node type in macro list");
2587 }
2588 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
2589 if (C->failing()) return true;
2590 }
2591
2592 _igvn.set_delay_transform(false);
2593 _igvn.optimize();
2594 if (C->failing()) return true;
2595 return false;
2596 }
2597