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