Thu, 20 Sep 2012 16:49:17 +0200
7023898: Intrinsify AtomicLongFieldUpdater.getAndIncrement()
Summary: use shorter instruction sequences for atomic add and atomic exchange when possible.
Reviewed-by: kvn, jrose
1 /*
2 * Copyright (c) 2005, 2012, 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 */
25 #include "precompiled.hpp"
26 #include "compiler/compileLog.hpp"
27 #include "libadt/vectset.hpp"
28 #include "opto/addnode.hpp"
29 #include "opto/callnode.hpp"
30 #include "opto/cfgnode.hpp"
31 #include "opto/compile.hpp"
32 #include "opto/connode.hpp"
33 #include "opto/locknode.hpp"
34 #include "opto/loopnode.hpp"
35 #include "opto/macro.hpp"
36 #include "opto/memnode.hpp"
37 #include "opto/node.hpp"
38 #include "opto/phaseX.hpp"
39 #include "opto/rootnode.hpp"
40 #include "opto/runtime.hpp"
41 #include "opto/subnode.hpp"
42 #include "opto/type.hpp"
43 #include "runtime/sharedRuntime.hpp"
46 //
47 // Replace any references to "oldref" in inputs to "use" with "newref".
48 // Returns the number of replacements made.
49 //
50 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) {
51 int nreplacements = 0;
52 uint req = use->req();
53 for (uint j = 0; j < use->len(); j++) {
54 Node *uin = use->in(j);
55 if (uin == oldref) {
56 if (j < req)
57 use->set_req(j, newref);
58 else
59 use->set_prec(j, newref);
60 nreplacements++;
61 } else if (j >= req && uin == NULL) {
62 break;
63 }
64 }
65 return nreplacements;
66 }
68 void PhaseMacroExpand::copy_call_debug_info(CallNode *oldcall, CallNode * newcall) {
69 // Copy debug information and adjust JVMState information
70 uint old_dbg_start = oldcall->tf()->domain()->cnt();
71 uint new_dbg_start = newcall->tf()->domain()->cnt();
72 int jvms_adj = new_dbg_start - old_dbg_start;
73 assert (new_dbg_start == newcall->req(), "argument count mismatch");
75 Dict* sosn_map = new Dict(cmpkey,hashkey);
76 for (uint i = old_dbg_start; i < oldcall->req(); i++) {
77 Node* old_in = oldcall->in(i);
78 // Clone old SafePointScalarObjectNodes, adjusting their field contents.
79 if (old_in != NULL && old_in->is_SafePointScalarObject()) {
80 SafePointScalarObjectNode* old_sosn = old_in->as_SafePointScalarObject();
81 uint old_unique = C->unique();
82 Node* new_in = old_sosn->clone(jvms_adj, sosn_map);
83 if (old_unique != C->unique()) {
84 new_in->set_req(0, C->root()); // reset control edge
85 new_in = transform_later(new_in); // Register new node.
86 }
87 old_in = new_in;
88 }
89 newcall->add_req(old_in);
90 }
92 newcall->set_jvms(oldcall->jvms());
93 for (JVMState *jvms = newcall->jvms(); jvms != NULL; jvms = jvms->caller()) {
94 jvms->set_map(newcall);
95 jvms->set_locoff(jvms->locoff()+jvms_adj);
96 jvms->set_stkoff(jvms->stkoff()+jvms_adj);
97 jvms->set_monoff(jvms->monoff()+jvms_adj);
98 jvms->set_scloff(jvms->scloff()+jvms_adj);
99 jvms->set_endoff(jvms->endoff()+jvms_adj);
100 }
101 }
103 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word, int mask, int bits, bool return_fast_path) {
104 Node* cmp;
105 if (mask != 0) {
106 Node* and_node = transform_later(new (C, 3) AndXNode(word, MakeConX(mask)));
107 cmp = transform_later(new (C, 3) CmpXNode(and_node, MakeConX(bits)));
108 } else {
109 cmp = word;
110 }
111 Node* bol = transform_later(new (C, 2) BoolNode(cmp, BoolTest::ne));
112 IfNode* iff = new (C, 2) IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN );
113 transform_later(iff);
115 // Fast path taken.
116 Node *fast_taken = transform_later( new (C, 1) IfFalseNode(iff) );
118 // Fast path not-taken, i.e. slow path
119 Node *slow_taken = transform_later( new (C, 1) IfTrueNode(iff) );
121 if (return_fast_path) {
122 region->init_req(edge, slow_taken); // Capture slow-control
123 return fast_taken;
124 } else {
125 region->init_req(edge, fast_taken); // Capture fast-control
126 return slow_taken;
127 }
128 }
130 //--------------------copy_predefined_input_for_runtime_call--------------------
131 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) {
132 // Set fixed predefined input arguments
133 call->init_req( TypeFunc::Control, ctrl );
134 call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) );
135 call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ?????
136 call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) );
137 call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) );
138 }
140 //------------------------------make_slow_call---------------------------------
141 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) {
143 // Slow-path call
144 int size = slow_call_type->domain()->cnt();
145 CallNode *call = leaf_name
146 ? (CallNode*)new (C, size) CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM )
147 : (CallNode*)new (C, size) CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), oldcall->jvms()->bci(), TypeRawPtr::BOTTOM );
149 // Slow path call has no side-effects, uses few values
150 copy_predefined_input_for_runtime_call(slow_path, oldcall, call );
151 if (parm0 != NULL) call->init_req(TypeFunc::Parms+0, parm0);
152 if (parm1 != NULL) call->init_req(TypeFunc::Parms+1, parm1);
153 copy_call_debug_info(oldcall, call);
154 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
155 _igvn.replace_node(oldcall, call);
156 transform_later(call);
158 return call;
159 }
161 void PhaseMacroExpand::extract_call_projections(CallNode *call) {
162 _fallthroughproj = NULL;
163 _fallthroughcatchproj = NULL;
164 _ioproj_fallthrough = NULL;
165 _ioproj_catchall = NULL;
166 _catchallcatchproj = NULL;
167 _memproj_fallthrough = NULL;
168 _memproj_catchall = NULL;
169 _resproj = NULL;
170 for (DUIterator_Fast imax, i = call->fast_outs(imax); i < imax; i++) {
171 ProjNode *pn = call->fast_out(i)->as_Proj();
172 switch (pn->_con) {
173 case TypeFunc::Control:
174 {
175 // For Control (fallthrough) and I_O (catch_all_index) we have CatchProj -> Catch -> Proj
176 _fallthroughproj = pn;
177 DUIterator_Fast jmax, j = pn->fast_outs(jmax);
178 const Node *cn = pn->fast_out(j);
179 if (cn->is_Catch()) {
180 ProjNode *cpn = NULL;
181 for (DUIterator_Fast kmax, k = cn->fast_outs(kmax); k < kmax; k++) {
182 cpn = cn->fast_out(k)->as_Proj();
183 assert(cpn->is_CatchProj(), "must be a CatchProjNode");
184 if (cpn->_con == CatchProjNode::fall_through_index)
185 _fallthroughcatchproj = cpn;
186 else {
187 assert(cpn->_con == CatchProjNode::catch_all_index, "must be correct index.");
188 _catchallcatchproj = cpn;
189 }
190 }
191 }
192 break;
193 }
194 case TypeFunc::I_O:
195 if (pn->_is_io_use)
196 _ioproj_catchall = pn;
197 else
198 _ioproj_fallthrough = pn;
199 break;
200 case TypeFunc::Memory:
201 if (pn->_is_io_use)
202 _memproj_catchall = pn;
203 else
204 _memproj_fallthrough = pn;
205 break;
206 case TypeFunc::Parms:
207 _resproj = pn;
208 break;
209 default:
210 assert(false, "unexpected projection from allocation node.");
211 }
212 }
214 }
216 // Eliminate a card mark sequence. p2x is a ConvP2XNode
217 void PhaseMacroExpand::eliminate_card_mark(Node* p2x) {
218 assert(p2x->Opcode() == Op_CastP2X, "ConvP2XNode required");
219 if (!UseG1GC) {
220 // vanilla/CMS post barrier
221 Node *shift = p2x->unique_out();
222 Node *addp = shift->unique_out();
223 for (DUIterator_Last jmin, j = addp->last_outs(jmin); j >= jmin; --j) {
224 Node *mem = addp->last_out(j);
225 if (UseCondCardMark && mem->is_Load()) {
226 assert(mem->Opcode() == Op_LoadB, "unexpected code shape");
227 // The load is checking if the card has been written so
228 // replace it with zero to fold the test.
229 _igvn.replace_node(mem, intcon(0));
230 continue;
231 }
232 assert(mem->is_Store(), "store required");
233 _igvn.replace_node(mem, mem->in(MemNode::Memory));
234 }
235 } else {
236 // G1 pre/post barriers
237 assert(p2x->outcnt() <= 2, "expects 1 or 2 users: Xor and URShift nodes");
238 // It could be only one user, URShift node, in Object.clone() instrinsic
239 // but the new allocation is passed to arraycopy stub and it could not
240 // be scalar replaced. So we don't check the case.
242 // An other case of only one user (Xor) is when the value check for NULL
243 // in G1 post barrier is folded after CCP so the code which used URShift
244 // is removed.
246 // Take Region node before eliminating post barrier since it also
247 // eliminates CastP2X node when it has only one user.
248 Node* this_region = p2x->in(0);
249 assert(this_region != NULL, "");
251 // Remove G1 post barrier.
253 // Search for CastP2X->Xor->URShift->Cmp path which
254 // checks if the store done to a different from the value's region.
255 // And replace Cmp with #0 (false) to collapse G1 post barrier.
256 Node* xorx = NULL;
257 for (DUIterator_Fast imax, i = p2x->fast_outs(imax); i < imax; i++) {
258 Node* u = p2x->fast_out(i);
259 if (u->Opcode() == Op_XorX) {
260 xorx = u;
261 break;
262 }
263 }
264 assert(xorx != NULL, "missing G1 post barrier");
265 Node* shift = xorx->unique_out();
266 Node* cmpx = shift->unique_out();
267 assert(cmpx->is_Cmp() && cmpx->unique_out()->is_Bool() &&
268 cmpx->unique_out()->as_Bool()->_test._test == BoolTest::ne,
269 "missing region check in G1 post barrier");
270 _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ));
272 // Remove G1 pre barrier.
274 // Search "if (marking != 0)" check and set it to "false".
275 // There is no G1 pre barrier if previous stored value is NULL
276 // (for example, after initialization).
277 if (this_region->is_Region() && this_region->req() == 3) {
278 int ind = 1;
279 if (!this_region->in(ind)->is_IfFalse()) {
280 ind = 2;
281 }
282 if (this_region->in(ind)->is_IfFalse()) {
283 Node* bol = this_region->in(ind)->in(0)->in(1);
284 assert(bol->is_Bool(), "");
285 cmpx = bol->in(1);
286 if (bol->as_Bool()->_test._test == BoolTest::ne &&
287 cmpx->is_Cmp() && cmpx->in(2) == intcon(0) &&
288 cmpx->in(1)->is_Load()) {
289 Node* adr = cmpx->in(1)->as_Load()->in(MemNode::Address);
290 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() +
291 PtrQueue::byte_offset_of_active());
292 if (adr->is_AddP() && adr->in(AddPNode::Base) == top() &&
293 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
294 adr->in(AddPNode::Offset) == MakeConX(marking_offset)) {
295 _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ));
296 }
297 }
298 }
299 }
300 // Now CastP2X can be removed since it is used only on dead path
301 // which currently still alive until igvn optimize it.
302 assert(p2x->outcnt() == 0 || p2x->unique_out()->Opcode() == Op_URShiftX, "");
303 _igvn.replace_node(p2x, top());
304 }
305 }
307 // Search for a memory operation for the specified memory slice.
308 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) {
309 Node *orig_mem = mem;
310 Node *alloc_mem = alloc->in(TypeFunc::Memory);
311 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr();
312 while (true) {
313 if (mem == alloc_mem || mem == start_mem ) {
314 return mem; // hit one of our sentinels
315 } else if (mem->is_MergeMem()) {
316 mem = mem->as_MergeMem()->memory_at(alias_idx);
317 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) {
318 Node *in = mem->in(0);
319 // we can safely skip over safepoints, calls, locks and membars because we
320 // already know that the object is safe to eliminate.
321 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) {
322 return in;
323 } else if (in->is_Call()) {
324 CallNode *call = in->as_Call();
325 if (!call->may_modify(tinst, phase)) {
326 mem = call->in(TypeFunc::Memory);
327 }
328 mem = in->in(TypeFunc::Memory);
329 } else if (in->is_MemBar()) {
330 mem = in->in(TypeFunc::Memory);
331 } else {
332 assert(false, "unexpected projection");
333 }
334 } else if (mem->is_Store()) {
335 const TypePtr* atype = mem->as_Store()->adr_type();
336 int adr_idx = Compile::current()->get_alias_index(atype);
337 if (adr_idx == alias_idx) {
338 assert(atype->isa_oopptr(), "address type must be oopptr");
339 int adr_offset = atype->offset();
340 uint adr_iid = atype->is_oopptr()->instance_id();
341 // Array elements references have the same alias_idx
342 // but different offset and different instance_id.
343 if (adr_offset == offset && adr_iid == alloc->_idx)
344 return mem;
345 } else {
346 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw");
347 }
348 mem = mem->in(MemNode::Memory);
349 } else if (mem->is_ClearArray()) {
350 if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) {
351 // Can not bypass initialization of the instance
352 // we are looking.
353 debug_only(intptr_t offset;)
354 assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity");
355 InitializeNode* init = alloc->as_Allocate()->initialization();
356 // We are looking for stored value, return Initialize node
357 // or memory edge from Allocate node.
358 if (init != NULL)
359 return init;
360 else
361 return alloc->in(TypeFunc::Memory); // It will produce zero value (see callers).
362 }
363 // Otherwise skip it (the call updated 'mem' value).
364 } else if (mem->Opcode() == Op_SCMemProj) {
365 assert(mem->in(0)->is_LoadStore(), "sanity");
366 const TypePtr* atype = mem->in(0)->in(MemNode::Address)->bottom_type()->is_ptr();
367 int adr_idx = Compile::current()->get_alias_index(atype);
368 if (adr_idx == alias_idx) {
369 assert(false, "Object is not scalar replaceable if a LoadStore node access its field");
370 return NULL;
371 }
372 mem = mem->in(0)->in(MemNode::Memory);
373 } else {
374 return mem;
375 }
376 assert(mem != orig_mem, "dead memory loop");
377 }
378 }
380 //
381 // Given a Memory Phi, compute a value Phi containing the values from stores
382 // on the input paths.
383 // Note: this function is recursive, its depth is limied by the "level" argument
384 // Returns the computed Phi, or NULL if it cannot compute it.
385 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) {
386 assert(mem->is_Phi(), "sanity");
387 int alias_idx = C->get_alias_index(adr_t);
388 int offset = adr_t->offset();
389 int instance_id = adr_t->instance_id();
391 // Check if an appropriate value phi already exists.
392 Node* region = mem->in(0);
393 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) {
394 Node* phi = region->fast_out(k);
395 if (phi->is_Phi() && phi != mem &&
396 phi->as_Phi()->is_same_inst_field(phi_type, instance_id, alias_idx, offset)) {
397 return phi;
398 }
399 }
400 // Check if an appropriate new value phi already exists.
401 Node* new_phi = value_phis->find(mem->_idx);
402 if (new_phi != NULL)
403 return new_phi;
405 if (level <= 0) {
406 return NULL; // Give up: phi tree too deep
407 }
408 Node *start_mem = C->start()->proj_out(TypeFunc::Memory);
409 Node *alloc_mem = alloc->in(TypeFunc::Memory);
411 uint length = mem->req();
412 GrowableArray <Node *> values(length, length, NULL, false);
414 // create a new Phi for the value
415 PhiNode *phi = new (C, length) PhiNode(mem->in(0), phi_type, NULL, instance_id, alias_idx, offset);
416 transform_later(phi);
417 value_phis->push(phi, mem->_idx);
419 for (uint j = 1; j < length; j++) {
420 Node *in = mem->in(j);
421 if (in == NULL || in->is_top()) {
422 values.at_put(j, in);
423 } else {
424 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn);
425 if (val == start_mem || val == alloc_mem) {
426 // hit a sentinel, return appropriate 0 value
427 values.at_put(j, _igvn.zerocon(ft));
428 continue;
429 }
430 if (val->is_Initialize()) {
431 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
432 }
433 if (val == NULL) {
434 return NULL; // can't find a value on this path
435 }
436 if (val == mem) {
437 values.at_put(j, mem);
438 } else if (val->is_Store()) {
439 values.at_put(j, val->in(MemNode::ValueIn));
440 } else if(val->is_Proj() && val->in(0) == alloc) {
441 values.at_put(j, _igvn.zerocon(ft));
442 } else if (val->is_Phi()) {
443 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1);
444 if (val == NULL) {
445 return NULL;
446 }
447 values.at_put(j, val);
448 } else if (val->Opcode() == Op_SCMemProj) {
449 assert(val->in(0)->is_LoadStore(), "sanity");
450 assert(false, "Object is not scalar replaceable if a LoadStore node access its field");
451 return NULL;
452 } else {
453 #ifdef ASSERT
454 val->dump();
455 assert(false, "unknown node on this path");
456 #endif
457 return NULL; // unknown node on this path
458 }
459 }
460 }
461 // Set Phi's inputs
462 for (uint j = 1; j < length; j++) {
463 if (values.at(j) == mem) {
464 phi->init_req(j, phi);
465 } else {
466 phi->init_req(j, values.at(j));
467 }
468 }
469 return phi;
470 }
472 // Search the last value stored into the object's field.
473 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, Node *alloc) {
474 assert(adr_t->is_known_instance_field(), "instance required");
475 int instance_id = adr_t->instance_id();
476 assert((uint)instance_id == alloc->_idx, "wrong allocation");
478 int alias_idx = C->get_alias_index(adr_t);
479 int offset = adr_t->offset();
480 Node *start_mem = C->start()->proj_out(TypeFunc::Memory);
481 Node *alloc_ctrl = alloc->in(TypeFunc::Control);
482 Node *alloc_mem = alloc->in(TypeFunc::Memory);
483 Arena *a = Thread::current()->resource_area();
484 VectorSet visited(a);
487 bool done = sfpt_mem == alloc_mem;
488 Node *mem = sfpt_mem;
489 while (!done) {
490 if (visited.test_set(mem->_idx)) {
491 return NULL; // found a loop, give up
492 }
493 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn);
494 if (mem == start_mem || mem == alloc_mem) {
495 done = true; // hit a sentinel, return appropriate 0 value
496 } else if (mem->is_Initialize()) {
497 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
498 if (mem == NULL) {
499 done = true; // Something go wrong.
500 } else if (mem->is_Store()) {
501 const TypePtr* atype = mem->as_Store()->adr_type();
502 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice");
503 done = true;
504 }
505 } else if (mem->is_Store()) {
506 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr();
507 assert(atype != NULL, "address type must be oopptr");
508 assert(C->get_alias_index(atype) == alias_idx &&
509 atype->is_known_instance_field() && atype->offset() == offset &&
510 atype->instance_id() == instance_id, "store is correct memory slice");
511 done = true;
512 } else if (mem->is_Phi()) {
513 // try to find a phi's unique input
514 Node *unique_input = NULL;
515 Node *top = C->top();
516 for (uint i = 1; i < mem->req(); i++) {
517 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn);
518 if (n == NULL || n == top || n == mem) {
519 continue;
520 } else if (unique_input == NULL) {
521 unique_input = n;
522 } else if (unique_input != n) {
523 unique_input = top;
524 break;
525 }
526 }
527 if (unique_input != NULL && unique_input != top) {
528 mem = unique_input;
529 } else {
530 done = true;
531 }
532 } else {
533 assert(false, "unexpected node");
534 }
535 }
536 if (mem != NULL) {
537 if (mem == start_mem || mem == alloc_mem) {
538 // hit a sentinel, return appropriate 0 value
539 return _igvn.zerocon(ft);
540 } else if (mem->is_Store()) {
541 return mem->in(MemNode::ValueIn);
542 } else if (mem->is_Phi()) {
543 // attempt to produce a Phi reflecting the values on the input paths of the Phi
544 Node_Stack value_phis(a, 8);
545 Node * phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit);
546 if (phi != NULL) {
547 return phi;
548 } else {
549 // Kill all new Phis
550 while(value_phis.is_nonempty()) {
551 Node* n = value_phis.node();
552 _igvn.replace_node(n, C->top());
553 value_phis.pop();
554 }
555 }
556 }
557 }
558 // Something go wrong.
559 return NULL;
560 }
562 // Check the possibility of scalar replacement.
563 bool PhaseMacroExpand::can_eliminate_allocation(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
564 // Scan the uses of the allocation to check for anything that would
565 // prevent us from eliminating it.
566 NOT_PRODUCT( const char* fail_eliminate = NULL; )
567 DEBUG_ONLY( Node* disq_node = NULL; )
568 bool can_eliminate = true;
570 Node* res = alloc->result_cast();
571 const TypeOopPtr* res_type = NULL;
572 if (res == NULL) {
573 // All users were eliminated.
574 } else if (!res->is_CheckCastPP()) {
575 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";)
576 can_eliminate = false;
577 } else {
578 res_type = _igvn.type(res)->isa_oopptr();
579 if (res_type == NULL) {
580 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";)
581 can_eliminate = false;
582 } else if (res_type->isa_aryptr()) {
583 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1);
584 if (length < 0) {
585 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";)
586 can_eliminate = false;
587 }
588 }
589 }
591 if (can_eliminate && res != NULL) {
592 for (DUIterator_Fast jmax, j = res->fast_outs(jmax);
593 j < jmax && can_eliminate; j++) {
594 Node* use = res->fast_out(j);
596 if (use->is_AddP()) {
597 const TypePtr* addp_type = _igvn.type(use)->is_ptr();
598 int offset = addp_type->offset();
600 if (offset == Type::OffsetTop || offset == Type::OffsetBot) {
601 NOT_PRODUCT(fail_eliminate = "Undefined field referrence";)
602 can_eliminate = false;
603 break;
604 }
605 for (DUIterator_Fast kmax, k = use->fast_outs(kmax);
606 k < kmax && can_eliminate; k++) {
607 Node* n = use->fast_out(k);
608 if (!n->is_Store() && n->Opcode() != Op_CastP2X) {
609 DEBUG_ONLY(disq_node = n;)
610 if (n->is_Load() || n->is_LoadStore()) {
611 NOT_PRODUCT(fail_eliminate = "Field load";)
612 } else {
613 NOT_PRODUCT(fail_eliminate = "Not store field referrence";)
614 }
615 can_eliminate = false;
616 }
617 }
618 } else if (use->is_SafePoint()) {
619 SafePointNode* sfpt = use->as_SafePoint();
620 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) {
621 // Object is passed as argument.
622 DEBUG_ONLY(disq_node = use;)
623 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";)
624 can_eliminate = false;
625 }
626 Node* sfptMem = sfpt->memory();
627 if (sfptMem == NULL || sfptMem->is_top()) {
628 DEBUG_ONLY(disq_node = use;)
629 NOT_PRODUCT(fail_eliminate = "NULL or TOP memory";)
630 can_eliminate = false;
631 } else {
632 safepoints.append_if_missing(sfpt);
633 }
634 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
635 if (use->is_Phi()) {
636 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
637 NOT_PRODUCT(fail_eliminate = "Object is return value";)
638 } else {
639 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
640 }
641 DEBUG_ONLY(disq_node = use;)
642 } else {
643 if (use->Opcode() == Op_Return) {
644 NOT_PRODUCT(fail_eliminate = "Object is return value";)
645 }else {
646 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
647 }
648 DEBUG_ONLY(disq_node = use;)
649 }
650 can_eliminate = false;
651 }
652 }
653 }
655 #ifndef PRODUCT
656 if (PrintEliminateAllocations) {
657 if (can_eliminate) {
658 tty->print("Scalar ");
659 if (res == NULL)
660 alloc->dump();
661 else
662 res->dump();
663 } else {
664 tty->print("NotScalar (%s)", fail_eliminate);
665 if (res == NULL)
666 alloc->dump();
667 else
668 res->dump();
669 #ifdef ASSERT
670 if (disq_node != NULL) {
671 tty->print(" >>>> ");
672 disq_node->dump();
673 }
674 #endif /*ASSERT*/
675 }
676 }
677 #endif
678 return can_eliminate;
679 }
681 // Do scalar replacement.
682 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
683 GrowableArray <SafePointNode *> safepoints_done;
685 ciKlass* klass = NULL;
686 ciInstanceKlass* iklass = NULL;
687 int nfields = 0;
688 int array_base;
689 int element_size;
690 BasicType basic_elem_type;
691 ciType* elem_type;
693 Node* res = alloc->result_cast();
694 const TypeOopPtr* res_type = NULL;
695 if (res != NULL) { // Could be NULL when there are no users
696 res_type = _igvn.type(res)->isa_oopptr();
697 }
699 if (res != NULL) {
700 klass = res_type->klass();
701 if (res_type->isa_instptr()) {
702 // find the fields of the class which will be needed for safepoint debug information
703 assert(klass->is_instance_klass(), "must be an instance klass.");
704 iklass = klass->as_instance_klass();
705 nfields = iklass->nof_nonstatic_fields();
706 } else {
707 // find the array's elements which will be needed for safepoint debug information
708 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
709 assert(klass->is_array_klass() && nfields >= 0, "must be an array klass.");
710 elem_type = klass->as_array_klass()->element_type();
711 basic_elem_type = elem_type->basic_type();
712 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
713 element_size = type2aelembytes(basic_elem_type);
714 }
715 }
716 //
717 // Process the safepoint uses
718 //
719 while (safepoints.length() > 0) {
720 SafePointNode* sfpt = safepoints.pop();
721 Node* mem = sfpt->memory();
722 uint first_ind = sfpt->req();
723 SafePointScalarObjectNode* sobj = new (C, 1) SafePointScalarObjectNode(res_type,
724 #ifdef ASSERT
725 alloc,
726 #endif
727 first_ind, nfields);
728 sobj->init_req(0, C->root());
729 transform_later(sobj);
731 // Scan object's fields adding an input to the safepoint for each field.
732 for (int j = 0; j < nfields; j++) {
733 intptr_t offset;
734 ciField* field = NULL;
735 if (iklass != NULL) {
736 field = iklass->nonstatic_field_at(j);
737 offset = field->offset();
738 elem_type = field->type();
739 basic_elem_type = field->layout_type();
740 } else {
741 offset = array_base + j * (intptr_t)element_size;
742 }
744 const Type *field_type;
745 // The next code is taken from Parse::do_get_xxx().
746 if (basic_elem_type == T_OBJECT || basic_elem_type == T_ARRAY) {
747 if (!elem_type->is_loaded()) {
748 field_type = TypeInstPtr::BOTTOM;
749 } else if (field != NULL && field->is_constant() && field->is_static()) {
750 // This can happen if the constant oop is non-perm.
751 ciObject* con = field->constant_value().as_object();
752 // Do not "join" in the previous type; it doesn't add value,
753 // and may yield a vacuous result if the field is of interface type.
754 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr();
755 assert(field_type != NULL, "field singleton type must be consistent");
756 } else {
757 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
758 }
759 if (UseCompressedOops) {
760 field_type = field_type->make_narrowoop();
761 basic_elem_type = T_NARROWOOP;
762 }
763 } else {
764 field_type = Type::get_const_basic_type(basic_elem_type);
765 }
767 const TypeOopPtr *field_addr_type = res_type->add_offset(offset)->isa_oopptr();
769 Node *field_val = value_from_mem(mem, basic_elem_type, field_type, field_addr_type, alloc);
770 if (field_val == NULL) {
771 // We weren't able to find a value for this field,
772 // give up on eliminating this allocation.
774 // Remove any extra entries we added to the safepoint.
775 uint last = sfpt->req() - 1;
776 for (int k = 0; k < j; k++) {
777 sfpt->del_req(last--);
778 }
779 // rollback processed safepoints
780 while (safepoints_done.length() > 0) {
781 SafePointNode* sfpt_done = safepoints_done.pop();
782 // remove any extra entries we added to the safepoint
783 last = sfpt_done->req() - 1;
784 for (int k = 0; k < nfields; k++) {
785 sfpt_done->del_req(last--);
786 }
787 JVMState *jvms = sfpt_done->jvms();
788 jvms->set_endoff(sfpt_done->req());
789 // Now make a pass over the debug information replacing any references
790 // to SafePointScalarObjectNode with the allocated object.
791 int start = jvms->debug_start();
792 int end = jvms->debug_end();
793 for (int i = start; i < end; i++) {
794 if (sfpt_done->in(i)->is_SafePointScalarObject()) {
795 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject();
796 if (scobj->first_index() == sfpt_done->req() &&
797 scobj->n_fields() == (uint)nfields) {
798 assert(scobj->alloc() == alloc, "sanity");
799 sfpt_done->set_req(i, res);
800 }
801 }
802 }
803 }
804 #ifndef PRODUCT
805 if (PrintEliminateAllocations) {
806 if (field != NULL) {
807 tty->print("=== At SafePoint node %d can't find value of Field: ",
808 sfpt->_idx);
809 field->print();
810 int field_idx = C->get_alias_index(field_addr_type);
811 tty->print(" (alias_idx=%d)", field_idx);
812 } else { // Array's element
813 tty->print("=== At SafePoint node %d can't find value of array element [%d]",
814 sfpt->_idx, j);
815 }
816 tty->print(", which prevents elimination of: ");
817 if (res == NULL)
818 alloc->dump();
819 else
820 res->dump();
821 }
822 #endif
823 return false;
824 }
825 if (UseCompressedOops && field_type->isa_narrowoop()) {
826 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
827 // to be able scalar replace the allocation.
828 if (field_val->is_EncodeP()) {
829 field_val = field_val->in(1);
830 } else {
831 field_val = transform_later(new (C, 2) DecodeNNode(field_val, field_val->bottom_type()->make_ptr()));
832 }
833 }
834 sfpt->add_req(field_val);
835 }
836 JVMState *jvms = sfpt->jvms();
837 jvms->set_endoff(sfpt->req());
838 // Now make a pass over the debug information replacing any references
839 // to the allocated object with "sobj"
840 int start = jvms->debug_start();
841 int end = jvms->debug_end();
842 for (int i = start; i < end; i++) {
843 if (sfpt->in(i) == res) {
844 sfpt->set_req(i, sobj);
845 }
846 }
847 safepoints_done.append_if_missing(sfpt); // keep it for rollback
848 }
849 return true;
850 }
852 // Process users of eliminated allocation.
853 void PhaseMacroExpand::process_users_of_allocation(AllocateNode *alloc) {
854 Node* res = alloc->result_cast();
855 if (res != NULL) {
856 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
857 Node *use = res->last_out(j);
858 uint oc1 = res->outcnt();
860 if (use->is_AddP()) {
861 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) {
862 Node *n = use->last_out(k);
863 uint oc2 = use->outcnt();
864 if (n->is_Store()) {
865 #ifdef ASSERT
866 // Verify that there is no dependent MemBarVolatile nodes,
867 // they should be removed during IGVN, see MemBarNode::Ideal().
868 for (DUIterator_Fast pmax, p = n->fast_outs(pmax);
869 p < pmax; p++) {
870 Node* mb = n->fast_out(p);
871 assert(mb->is_Initialize() || !mb->is_MemBar() ||
872 mb->req() <= MemBarNode::Precedent ||
873 mb->in(MemBarNode::Precedent) != n,
874 "MemBarVolatile should be eliminated for non-escaping object");
875 }
876 #endif
877 _igvn.replace_node(n, n->in(MemNode::Memory));
878 } else {
879 eliminate_card_mark(n);
880 }
881 k -= (oc2 - use->outcnt());
882 }
883 } else {
884 eliminate_card_mark(use);
885 }
886 j -= (oc1 - res->outcnt());
887 }
888 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
889 _igvn.remove_dead_node(res);
890 }
892 //
893 // Process other users of allocation's projections
894 //
895 if (_resproj != NULL && _resproj->outcnt() != 0) {
896 for (DUIterator_Last jmin, j = _resproj->last_outs(jmin); j >= jmin; ) {
897 Node *use = _resproj->last_out(j);
898 uint oc1 = _resproj->outcnt();
899 if (use->is_Initialize()) {
900 // Eliminate Initialize node.
901 InitializeNode *init = use->as_Initialize();
902 assert(init->outcnt() <= 2, "only a control and memory projection expected");
903 Node *ctrl_proj = init->proj_out(TypeFunc::Control);
904 if (ctrl_proj != NULL) {
905 assert(init->in(TypeFunc::Control) == _fallthroughcatchproj, "allocation control projection");
906 _igvn.replace_node(ctrl_proj, _fallthroughcatchproj);
907 }
908 Node *mem_proj = init->proj_out(TypeFunc::Memory);
909 if (mem_proj != NULL) {
910 Node *mem = init->in(TypeFunc::Memory);
911 #ifdef ASSERT
912 if (mem->is_MergeMem()) {
913 assert(mem->in(TypeFunc::Memory) == _memproj_fallthrough, "allocation memory projection");
914 } else {
915 assert(mem == _memproj_fallthrough, "allocation memory projection");
916 }
917 #endif
918 _igvn.replace_node(mem_proj, mem);
919 }
920 } else if (use->is_AddP()) {
921 // raw memory addresses used only by the initialization
922 _igvn.replace_node(use, C->top());
923 } else {
924 assert(false, "only Initialize or AddP expected");
925 }
926 j -= (oc1 - _resproj->outcnt());
927 }
928 }
929 if (_fallthroughcatchproj != NULL) {
930 _igvn.replace_node(_fallthroughcatchproj, alloc->in(TypeFunc::Control));
931 }
932 if (_memproj_fallthrough != NULL) {
933 _igvn.replace_node(_memproj_fallthrough, alloc->in(TypeFunc::Memory));
934 }
935 if (_memproj_catchall != NULL) {
936 _igvn.replace_node(_memproj_catchall, C->top());
937 }
938 if (_ioproj_fallthrough != NULL) {
939 _igvn.replace_node(_ioproj_fallthrough, alloc->in(TypeFunc::I_O));
940 }
941 if (_ioproj_catchall != NULL) {
942 _igvn.replace_node(_ioproj_catchall, C->top());
943 }
944 if (_catchallcatchproj != NULL) {
945 _igvn.replace_node(_catchallcatchproj, C->top());
946 }
947 }
949 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) {
951 if (!EliminateAllocations || !alloc->_is_scalar_replaceable) {
952 return false;
953 }
955 extract_call_projections(alloc);
957 GrowableArray <SafePointNode *> safepoints;
958 if (!can_eliminate_allocation(alloc, safepoints)) {
959 return false;
960 }
962 if (!scalar_replacement(alloc, safepoints)) {
963 return false;
964 }
966 CompileLog* log = C->log();
967 if (log != NULL) {
968 Node* klass = alloc->in(AllocateNode::KlassNode);
969 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr();
970 log->head("eliminate_allocation type='%d'",
971 log->identify(tklass->klass()));
972 JVMState* p = alloc->jvms();
973 while (p != NULL) {
974 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
975 p = p->caller();
976 }
977 log->tail("eliminate_allocation");
978 }
980 process_users_of_allocation(alloc);
982 #ifndef PRODUCT
983 if (PrintEliminateAllocations) {
984 if (alloc->is_AllocateArray())
985 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
986 else
987 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
988 }
989 #endif
991 return true;
992 }
995 //---------------------------set_eden_pointers-------------------------
996 void PhaseMacroExpand::set_eden_pointers(Node* &eden_top_adr, Node* &eden_end_adr) {
997 if (UseTLAB) { // Private allocation: load from TLS
998 Node* thread = transform_later(new (C, 1) ThreadLocalNode());
999 int tlab_top_offset = in_bytes(JavaThread::tlab_top_offset());
1000 int tlab_end_offset = in_bytes(JavaThread::tlab_end_offset());
1001 eden_top_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_top_offset);
1002 eden_end_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_end_offset);
1003 } else { // Shared allocation: load from globals
1004 CollectedHeap* ch = Universe::heap();
1005 address top_adr = (address)ch->top_addr();
1006 address end_adr = (address)ch->end_addr();
1007 eden_top_adr = makecon(TypeRawPtr::make(top_adr));
1008 eden_end_adr = basic_plus_adr(eden_top_adr, end_adr - top_adr);
1009 }
1010 }
1013 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) {
1014 Node* adr = basic_plus_adr(base, offset);
1015 const TypePtr* adr_type = adr->bottom_type()->is_ptr();
1016 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt);
1017 transform_later(value);
1018 return value;
1019 }
1022 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) {
1023 Node* adr = basic_plus_adr(base, offset);
1024 mem = StoreNode::make(_igvn, ctl, mem, adr, NULL, value, bt);
1025 transform_later(mem);
1026 return mem;
1027 }
1029 //=============================================================================
1030 //
1031 // A L L O C A T I O N
1032 //
1033 // Allocation attempts to be fast in the case of frequent small objects.
1034 // It breaks down like this:
1035 //
1036 // 1) Size in doublewords is computed. This is a constant for objects and
1037 // variable for most arrays. Doubleword units are used to avoid size
1038 // overflow of huge doubleword arrays. We need doublewords in the end for
1039 // rounding.
1040 //
1041 // 2) Size is checked for being 'too large'. Too-large allocations will go
1042 // the slow path into the VM. The slow path can throw any required
1043 // exceptions, and does all the special checks for very large arrays. The
1044 // size test can constant-fold away for objects. For objects with
1045 // finalizers it constant-folds the otherway: you always go slow with
1046 // finalizers.
1047 //
1048 // 3) If NOT using TLABs, this is the contended loop-back point.
1049 // Load-Locked the heap top. If using TLABs normal-load the heap top.
1050 //
1051 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route.
1052 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish
1053 // "size*8" we always enter the VM, where "largish" is a constant picked small
1054 // enough that there's always space between the eden max and 4Gig (old space is
1055 // there so it's quite large) and large enough that the cost of entering the VM
1056 // is dwarfed by the cost to initialize the space.
1057 //
1058 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back
1059 // down. If contended, repeat at step 3. If using TLABs normal-store
1060 // adjusted heap top back down; there is no contention.
1061 //
1062 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark
1063 // fields.
1064 //
1065 // 7) Merge with the slow-path; cast the raw memory pointer to the correct
1066 // oop flavor.
1067 //
1068 //=============================================================================
1069 // FastAllocateSizeLimit value is in DOUBLEWORDS.
1070 // Allocations bigger than this always go the slow route.
1071 // This value must be small enough that allocation attempts that need to
1072 // trigger exceptions go the slow route. Also, it must be small enough so
1073 // that heap_top + size_in_bytes does not wrap around the 4Gig limit.
1074 //=============================================================================j//
1075 // %%% Here is an old comment from parseHelper.cpp; is it outdated?
1076 // The allocator will coalesce int->oop copies away. See comment in
1077 // coalesce.cpp about how this works. It depends critically on the exact
1078 // code shape produced here, so if you are changing this code shape
1079 // make sure the GC info for the heap-top is correct in and around the
1080 // slow-path call.
1081 //
1083 void PhaseMacroExpand::expand_allocate_common(
1084 AllocateNode* alloc, // allocation node to be expanded
1085 Node* length, // array length for an array allocation
1086 const TypeFunc* slow_call_type, // Type of slow call
1087 address slow_call_address // Address of slow call
1088 )
1089 {
1091 Node* ctrl = alloc->in(TypeFunc::Control);
1092 Node* mem = alloc->in(TypeFunc::Memory);
1093 Node* i_o = alloc->in(TypeFunc::I_O);
1094 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize);
1095 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1096 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest);
1098 Node* storestore = alloc->storestore();
1099 if (storestore != NULL) {
1100 // Break this link that is no longer useful and confuses register allocation
1101 storestore->set_req(MemBarNode::Precedent, top());
1102 }
1104 assert(ctrl != NULL, "must have control");
1105 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results.
1106 // they will not be used if "always_slow" is set
1107 enum { slow_result_path = 1, fast_result_path = 2 };
1108 Node *result_region;
1109 Node *result_phi_rawmem;
1110 Node *result_phi_rawoop;
1111 Node *result_phi_i_o;
1113 // The initial slow comparison is a size check, the comparison
1114 // we want to do is a BoolTest::gt
1115 bool always_slow = false;
1116 int tv = _igvn.find_int_con(initial_slow_test, -1);
1117 if (tv >= 0) {
1118 always_slow = (tv == 1);
1119 initial_slow_test = NULL;
1120 } else {
1121 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn);
1122 }
1124 if (C->env()->dtrace_alloc_probes() ||
1125 !UseTLAB && (!Universe::heap()->supports_inline_contig_alloc() ||
1126 (UseConcMarkSweepGC && CMSIncrementalMode))) {
1127 // Force slow-path allocation
1128 always_slow = true;
1129 initial_slow_test = NULL;
1130 }
1133 enum { too_big_or_final_path = 1, need_gc_path = 2 };
1134 Node *slow_region = NULL;
1135 Node *toobig_false = ctrl;
1137 assert (initial_slow_test == NULL || !always_slow, "arguments must be consistent");
1138 // generate the initial test if necessary
1139 if (initial_slow_test != NULL ) {
1140 slow_region = new (C, 3) RegionNode(3);
1142 // Now make the initial failure test. Usually a too-big test but
1143 // might be a TRUE for finalizers or a fancy class check for
1144 // newInstance0.
1145 IfNode *toobig_iff = new (C, 2) IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
1146 transform_later(toobig_iff);
1147 // Plug the failing-too-big test into the slow-path region
1148 Node *toobig_true = new (C, 1) IfTrueNode( toobig_iff );
1149 transform_later(toobig_true);
1150 slow_region ->init_req( too_big_or_final_path, toobig_true );
1151 toobig_false = new (C, 1) IfFalseNode( toobig_iff );
1152 transform_later(toobig_false);
1153 } else { // No initial test, just fall into next case
1154 toobig_false = ctrl;
1155 debug_only(slow_region = NodeSentinel);
1156 }
1158 Node *slow_mem = mem; // save the current memory state for slow path
1159 // generate the fast allocation code unless we know that the initial test will always go slow
1160 if (!always_slow) {
1161 // Fast path modifies only raw memory.
1162 if (mem->is_MergeMem()) {
1163 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1164 }
1166 Node* eden_top_adr;
1167 Node* eden_end_adr;
1169 set_eden_pointers(eden_top_adr, eden_end_adr);
1171 // Load Eden::end. Loop invariant and hoisted.
1172 //
1173 // Note: We set the control input on "eden_end" and "old_eden_top" when using
1174 // a TLAB to work around a bug where these values were being moved across
1175 // a safepoint. These are not oops, so they cannot be include in the oop
1176 // map, but they can be changed by a GC. The proper way to fix this would
1177 // be to set the raw memory state when generating a SafepointNode. However
1178 // this will require extensive changes to the loop optimization in order to
1179 // prevent a degradation of the optimization.
1180 // See comment in memnode.hpp, around line 227 in class LoadPNode.
1181 Node *eden_end = make_load(ctrl, mem, eden_end_adr, 0, TypeRawPtr::BOTTOM, T_ADDRESS);
1183 // allocate the Region and Phi nodes for the result
1184 result_region = new (C, 3) RegionNode(3);
1185 result_phi_rawmem = new (C, 3) PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1186 result_phi_rawoop = new (C, 3) PhiNode(result_region, TypeRawPtr::BOTTOM);
1187 result_phi_i_o = new (C, 3) PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch
1189 // We need a Region for the loop-back contended case.
1190 enum { fall_in_path = 1, contended_loopback_path = 2 };
1191 Node *contended_region;
1192 Node *contended_phi_rawmem;
1193 if (UseTLAB) {
1194 contended_region = toobig_false;
1195 contended_phi_rawmem = mem;
1196 } else {
1197 contended_region = new (C, 3) RegionNode(3);
1198 contended_phi_rawmem = new (C, 3) PhiNode(contended_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1199 // Now handle the passing-too-big test. We fall into the contended
1200 // loop-back merge point.
1201 contended_region ->init_req(fall_in_path, toobig_false);
1202 contended_phi_rawmem->init_req(fall_in_path, mem);
1203 transform_later(contended_region);
1204 transform_later(contended_phi_rawmem);
1205 }
1207 // Load(-locked) the heap top.
1208 // See note above concerning the control input when using a TLAB
1209 Node *old_eden_top = UseTLAB
1210 ? new (C, 3) LoadPNode (ctrl, contended_phi_rawmem, eden_top_adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM)
1211 : new (C, 3) LoadPLockedNode(contended_region, contended_phi_rawmem, eden_top_adr);
1213 transform_later(old_eden_top);
1214 // Add to heap top to get a new heap top
1215 Node *new_eden_top = new (C, 4) AddPNode(top(), old_eden_top, size_in_bytes);
1216 transform_later(new_eden_top);
1217 // Check for needing a GC; compare against heap end
1218 Node *needgc_cmp = new (C, 3) CmpPNode(new_eden_top, eden_end);
1219 transform_later(needgc_cmp);
1220 Node *needgc_bol = new (C, 2) BoolNode(needgc_cmp, BoolTest::ge);
1221 transform_later(needgc_bol);
1222 IfNode *needgc_iff = new (C, 2) IfNode(contended_region, needgc_bol, PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN);
1223 transform_later(needgc_iff);
1225 // Plug the failing-heap-space-need-gc test into the slow-path region
1226 Node *needgc_true = new (C, 1) IfTrueNode(needgc_iff);
1227 transform_later(needgc_true);
1228 if (initial_slow_test) {
1229 slow_region->init_req(need_gc_path, needgc_true);
1230 // This completes all paths into the slow merge point
1231 transform_later(slow_region);
1232 } else { // No initial slow path needed!
1233 // Just fall from the need-GC path straight into the VM call.
1234 slow_region = needgc_true;
1235 }
1236 // No need for a GC. Setup for the Store-Conditional
1237 Node *needgc_false = new (C, 1) IfFalseNode(needgc_iff);
1238 transform_later(needgc_false);
1240 // Grab regular I/O before optional prefetch may change it.
1241 // Slow-path does no I/O so just set it to the original I/O.
1242 result_phi_i_o->init_req(slow_result_path, i_o);
1244 i_o = prefetch_allocation(i_o, needgc_false, contended_phi_rawmem,
1245 old_eden_top, new_eden_top, length);
1247 // Name successful fast-path variables
1248 Node* fast_oop = old_eden_top;
1249 Node* fast_oop_ctrl;
1250 Node* fast_oop_rawmem;
1252 // Store (-conditional) the modified eden top back down.
1253 // StorePConditional produces flags for a test PLUS a modified raw
1254 // memory state.
1255 if (UseTLAB) {
1256 Node* store_eden_top =
1257 new (C, 4) StorePNode(needgc_false, contended_phi_rawmem, eden_top_adr,
1258 TypeRawPtr::BOTTOM, new_eden_top);
1259 transform_later(store_eden_top);
1260 fast_oop_ctrl = needgc_false; // No contention, so this is the fast path
1261 fast_oop_rawmem = store_eden_top;
1262 } else {
1263 Node* store_eden_top =
1264 new (C, 5) StorePConditionalNode(needgc_false, contended_phi_rawmem, eden_top_adr,
1265 new_eden_top, fast_oop/*old_eden_top*/);
1266 transform_later(store_eden_top);
1267 Node *contention_check = new (C, 2) BoolNode(store_eden_top, BoolTest::ne);
1268 transform_later(contention_check);
1269 store_eden_top = new (C, 1) SCMemProjNode(store_eden_top);
1270 transform_later(store_eden_top);
1272 // If not using TLABs, check to see if there was contention.
1273 IfNode *contention_iff = new (C, 2) IfNode (needgc_false, contention_check, PROB_MIN, COUNT_UNKNOWN);
1274 transform_later(contention_iff);
1275 Node *contention_true = new (C, 1) IfTrueNode(contention_iff);
1276 transform_later(contention_true);
1277 // If contention, loopback and try again.
1278 contended_region->init_req(contended_loopback_path, contention_true);
1279 contended_phi_rawmem->init_req(contended_loopback_path, store_eden_top);
1281 // Fast-path succeeded with no contention!
1282 Node *contention_false = new (C, 1) IfFalseNode(contention_iff);
1283 transform_later(contention_false);
1284 fast_oop_ctrl = contention_false;
1286 // Bump total allocated bytes for this thread
1287 Node* thread = new (C, 1) ThreadLocalNode();
1288 transform_later(thread);
1289 Node* alloc_bytes_adr = basic_plus_adr(top()/*not oop*/, thread,
1290 in_bytes(JavaThread::allocated_bytes_offset()));
1291 Node* alloc_bytes = make_load(fast_oop_ctrl, store_eden_top, alloc_bytes_adr,
1292 0, TypeLong::LONG, T_LONG);
1293 #ifdef _LP64
1294 Node* alloc_size = size_in_bytes;
1295 #else
1296 Node* alloc_size = new (C, 2) ConvI2LNode(size_in_bytes);
1297 transform_later(alloc_size);
1298 #endif
1299 Node* new_alloc_bytes = new (C, 3) AddLNode(alloc_bytes, alloc_size);
1300 transform_later(new_alloc_bytes);
1301 fast_oop_rawmem = make_store(fast_oop_ctrl, store_eden_top, alloc_bytes_adr,
1302 0, new_alloc_bytes, T_LONG);
1303 }
1305 InitializeNode* init = alloc->initialization();
1306 fast_oop_rawmem = initialize_object(alloc,
1307 fast_oop_ctrl, fast_oop_rawmem, fast_oop,
1308 klass_node, length, size_in_bytes);
1310 // If initialization is performed by an array copy, any required
1311 // MemBarStoreStore was already added. If the object does not
1312 // escape no need for a MemBarStoreStore. Otherwise we need a
1313 // MemBarStoreStore so that stores that initialize this object
1314 // can't be reordered with a subsequent store that makes this
1315 // object accessible by other threads.
1316 if (init == NULL || (!init->is_complete_with_arraycopy() && !init->does_not_escape())) {
1317 if (init == NULL || init->req() < InitializeNode::RawStores) {
1318 // No InitializeNode or no stores captured by zeroing
1319 // elimination. Simply add the MemBarStoreStore after object
1320 // initialization.
1321 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot, fast_oop_rawmem);
1322 transform_later(mb);
1324 mb->init_req(TypeFunc::Memory, fast_oop_rawmem);
1325 mb->init_req(TypeFunc::Control, fast_oop_ctrl);
1326 fast_oop_ctrl = new (C, 1) ProjNode(mb,TypeFunc::Control);
1327 transform_later(fast_oop_ctrl);
1328 fast_oop_rawmem = new (C, 1) ProjNode(mb,TypeFunc::Memory);
1329 transform_later(fast_oop_rawmem);
1330 } else {
1331 // Add the MemBarStoreStore after the InitializeNode so that
1332 // all stores performing the initialization that were moved
1333 // before the InitializeNode happen before the storestore
1334 // barrier.
1336 Node* init_ctrl = init->proj_out(TypeFunc::Control);
1337 Node* init_mem = init->proj_out(TypeFunc::Memory);
1339 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1340 transform_later(mb);
1342 Node* ctrl = new (C, 1) ProjNode(init,TypeFunc::Control);
1343 transform_later(ctrl);
1344 Node* mem = new (C, 1) ProjNode(init,TypeFunc::Memory);
1345 transform_later(mem);
1347 // The MemBarStoreStore depends on control and memory coming
1348 // from the InitializeNode
1349 mb->init_req(TypeFunc::Memory, mem);
1350 mb->init_req(TypeFunc::Control, ctrl);
1352 ctrl = new (C, 1) ProjNode(mb,TypeFunc::Control);
1353 transform_later(ctrl);
1354 mem = new (C, 1) ProjNode(mb,TypeFunc::Memory);
1355 transform_later(mem);
1357 // All nodes that depended on the InitializeNode for control
1358 // and memory must now depend on the MemBarNode that itself
1359 // depends on the InitializeNode
1360 _igvn.replace_node(init_ctrl, ctrl);
1361 _igvn.replace_node(init_mem, mem);
1362 }
1363 }
1365 if (C->env()->dtrace_extended_probes()) {
1366 // Slow-path call
1367 int size = TypeFunc::Parms + 2;
1368 CallLeafNode *call = new (C, size) CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(),
1369 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_object_alloc_base),
1370 "dtrace_object_alloc",
1371 TypeRawPtr::BOTTOM);
1373 // Get base of thread-local storage area
1374 Node* thread = new (C, 1) ThreadLocalNode();
1375 transform_later(thread);
1377 call->init_req(TypeFunc::Parms+0, thread);
1378 call->init_req(TypeFunc::Parms+1, fast_oop);
1379 call->init_req(TypeFunc::Control, fast_oop_ctrl);
1380 call->init_req(TypeFunc::I_O , top()); // does no i/o
1381 call->init_req(TypeFunc::Memory , fast_oop_rawmem);
1382 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1383 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1384 transform_later(call);
1385 fast_oop_ctrl = new (C, 1) ProjNode(call,TypeFunc::Control);
1386 transform_later(fast_oop_ctrl);
1387 fast_oop_rawmem = new (C, 1) ProjNode(call,TypeFunc::Memory);
1388 transform_later(fast_oop_rawmem);
1389 }
1391 // Plug in the successful fast-path into the result merge point
1392 result_region ->init_req(fast_result_path, fast_oop_ctrl);
1393 result_phi_rawoop->init_req(fast_result_path, fast_oop);
1394 result_phi_i_o ->init_req(fast_result_path, i_o);
1395 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem);
1396 } else {
1397 slow_region = ctrl;
1398 result_phi_i_o = i_o; // Rename it to use in the following code.
1399 }
1401 // Generate slow-path call
1402 CallNode *call = new (C, slow_call_type->domain()->cnt())
1403 CallStaticJavaNode(slow_call_type, slow_call_address,
1404 OptoRuntime::stub_name(slow_call_address),
1405 alloc->jvms()->bci(),
1406 TypePtr::BOTTOM);
1407 call->init_req( TypeFunc::Control, slow_region );
1408 call->init_req( TypeFunc::I_O , top() ) ; // does no i/o
1409 call->init_req( TypeFunc::Memory , slow_mem ); // may gc ptrs
1410 call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) );
1411 call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) );
1413 call->init_req(TypeFunc::Parms+0, klass_node);
1414 if (length != NULL) {
1415 call->init_req(TypeFunc::Parms+1, length);
1416 }
1418 // Copy debug information and adjust JVMState information, then replace
1419 // allocate node with the call
1420 copy_call_debug_info((CallNode *) alloc, call);
1421 if (!always_slow) {
1422 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1423 } else {
1424 // Hook i_o projection to avoid its elimination during allocation
1425 // replacement (when only a slow call is generated).
1426 call->set_req(TypeFunc::I_O, result_phi_i_o);
1427 }
1428 _igvn.replace_node(alloc, call);
1429 transform_later(call);
1431 // Identify the output projections from the allocate node and
1432 // adjust any references to them.
1433 // The control and io projections look like:
1434 //
1435 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl)
1436 // Allocate Catch
1437 // ^---Proj(io) <-------+ ^---CatchProj(io)
1438 //
1439 // We are interested in the CatchProj nodes.
1440 //
1441 extract_call_projections(call);
1443 // An allocate node has separate memory projections for the uses on
1444 // the control and i_o paths. Replace the control memory projection with
1445 // result_phi_rawmem (unless we are only generating a slow call when
1446 // both memory projections are combined)
1447 if (!always_slow && _memproj_fallthrough != NULL) {
1448 for (DUIterator_Fast imax, i = _memproj_fallthrough->fast_outs(imax); i < imax; i++) {
1449 Node *use = _memproj_fallthrough->fast_out(i);
1450 _igvn.rehash_node_delayed(use);
1451 imax -= replace_input(use, _memproj_fallthrough, result_phi_rawmem);
1452 // back up iterator
1453 --i;
1454 }
1455 }
1456 // Now change uses of _memproj_catchall to use _memproj_fallthrough and delete
1457 // _memproj_catchall so we end up with a call that has only 1 memory projection.
1458 if (_memproj_catchall != NULL ) {
1459 if (_memproj_fallthrough == NULL) {
1460 _memproj_fallthrough = new (C, 1) ProjNode(call, TypeFunc::Memory);
1461 transform_later(_memproj_fallthrough);
1462 }
1463 for (DUIterator_Fast imax, i = _memproj_catchall->fast_outs(imax); i < imax; i++) {
1464 Node *use = _memproj_catchall->fast_out(i);
1465 _igvn.rehash_node_delayed(use);
1466 imax -= replace_input(use, _memproj_catchall, _memproj_fallthrough);
1467 // back up iterator
1468 --i;
1469 }
1470 assert(_memproj_catchall->outcnt() == 0, "all uses must be deleted");
1471 _igvn.remove_dead_node(_memproj_catchall);
1472 }
1474 // An allocate node has separate i_o projections for the uses on the control
1475 // and i_o paths. Always replace the control i_o projection with result i_o
1476 // otherwise incoming i_o become dead when only a slow call is generated
1477 // (it is different from memory projections where both projections are
1478 // combined in such case).
1479 if (_ioproj_fallthrough != NULL) {
1480 for (DUIterator_Fast imax, i = _ioproj_fallthrough->fast_outs(imax); i < imax; i++) {
1481 Node *use = _ioproj_fallthrough->fast_out(i);
1482 _igvn.rehash_node_delayed(use);
1483 imax -= replace_input(use, _ioproj_fallthrough, result_phi_i_o);
1484 // back up iterator
1485 --i;
1486 }
1487 }
1488 // Now change uses of _ioproj_catchall to use _ioproj_fallthrough and delete
1489 // _ioproj_catchall so we end up with a call that has only 1 i_o projection.
1490 if (_ioproj_catchall != NULL ) {
1491 if (_ioproj_fallthrough == NULL) {
1492 _ioproj_fallthrough = new (C, 1) ProjNode(call, TypeFunc::I_O);
1493 transform_later(_ioproj_fallthrough);
1494 }
1495 for (DUIterator_Fast imax, i = _ioproj_catchall->fast_outs(imax); i < imax; i++) {
1496 Node *use = _ioproj_catchall->fast_out(i);
1497 _igvn.rehash_node_delayed(use);
1498 imax -= replace_input(use, _ioproj_catchall, _ioproj_fallthrough);
1499 // back up iterator
1500 --i;
1501 }
1502 assert(_ioproj_catchall->outcnt() == 0, "all uses must be deleted");
1503 _igvn.remove_dead_node(_ioproj_catchall);
1504 }
1506 // if we generated only a slow call, we are done
1507 if (always_slow) {
1508 // Now we can unhook i_o.
1509 if (result_phi_i_o->outcnt() > 1) {
1510 call->set_req(TypeFunc::I_O, top());
1511 } else {
1512 assert(result_phi_i_o->unique_ctrl_out() == call, "");
1513 // Case of new array with negative size known during compilation.
1514 // AllocateArrayNode::Ideal() optimization disconnect unreachable
1515 // following code since call to runtime will throw exception.
1516 // As result there will be no users of i_o after the call.
1517 // Leave i_o attached to this call to avoid problems in preceding graph.
1518 }
1519 return;
1520 }
1523 if (_fallthroughcatchproj != NULL) {
1524 ctrl = _fallthroughcatchproj->clone();
1525 transform_later(ctrl);
1526 _igvn.replace_node(_fallthroughcatchproj, result_region);
1527 } else {
1528 ctrl = top();
1529 }
1530 Node *slow_result;
1531 if (_resproj == NULL) {
1532 // no uses of the allocation result
1533 slow_result = top();
1534 } else {
1535 slow_result = _resproj->clone();
1536 transform_later(slow_result);
1537 _igvn.replace_node(_resproj, result_phi_rawoop);
1538 }
1540 // Plug slow-path into result merge point
1541 result_region ->init_req( slow_result_path, ctrl );
1542 result_phi_rawoop->init_req( slow_result_path, slow_result);
1543 result_phi_rawmem->init_req( slow_result_path, _memproj_fallthrough );
1544 transform_later(result_region);
1545 transform_later(result_phi_rawoop);
1546 transform_later(result_phi_rawmem);
1547 transform_later(result_phi_i_o);
1548 // This completes all paths into the result merge point
1549 }
1552 // Helper for PhaseMacroExpand::expand_allocate_common.
1553 // Initializes the newly-allocated storage.
1554 Node*
1555 PhaseMacroExpand::initialize_object(AllocateNode* alloc,
1556 Node* control, Node* rawmem, Node* object,
1557 Node* klass_node, Node* length,
1558 Node* size_in_bytes) {
1559 InitializeNode* init = alloc->initialization();
1560 // Store the klass & mark bits
1561 Node* mark_node = NULL;
1562 // For now only enable fast locking for non-array types
1563 if (UseBiasedLocking && (length == NULL)) {
1564 mark_node = make_load(control, rawmem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
1565 } else {
1566 mark_node = makecon(TypeRawPtr::make((address)markOopDesc::prototype()));
1567 }
1568 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS);
1570 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
1571 int header_size = alloc->minimum_header_size(); // conservatively small
1573 // Array length
1574 if (length != NULL) { // Arrays need length field
1575 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
1576 // conservatively small header size:
1577 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1578 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
1579 if (k->is_array_klass()) // we know the exact header size in most cases:
1580 header_size = Klass::layout_helper_header_size(k->layout_helper());
1581 }
1583 // Clear the object body, if necessary.
1584 if (init == NULL) {
1585 // The init has somehow disappeared; be cautious and clear everything.
1586 //
1587 // This can happen if a node is allocated but an uncommon trap occurs
1588 // immediately. In this case, the Initialize gets associated with the
1589 // trap, and may be placed in a different (outer) loop, if the Allocate
1590 // is in a loop. If (this is rare) the inner loop gets unrolled, then
1591 // there can be two Allocates to one Initialize. The answer in all these
1592 // edge cases is safety first. It is always safe to clear immediately
1593 // within an Allocate, and then (maybe or maybe not) clear some more later.
1594 if (!ZeroTLAB)
1595 rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
1596 header_size, size_in_bytes,
1597 &_igvn);
1598 } else {
1599 if (!init->is_complete()) {
1600 // Try to win by zeroing only what the init does not store.
1601 // We can also try to do some peephole optimizations,
1602 // such as combining some adjacent subword stores.
1603 rawmem = init->complete_stores(control, rawmem, object,
1604 header_size, size_in_bytes, &_igvn);
1605 }
1606 // We have no more use for this link, since the AllocateNode goes away:
1607 init->set_req(InitializeNode::RawAddress, top());
1608 // (If we keep the link, it just confuses the register allocator,
1609 // who thinks he sees a real use of the address by the membar.)
1610 }
1612 return rawmem;
1613 }
1615 // Generate prefetch instructions for next allocations.
1616 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false,
1617 Node*& contended_phi_rawmem,
1618 Node* old_eden_top, Node* new_eden_top,
1619 Node* length) {
1620 enum { fall_in_path = 1, pf_path = 2 };
1621 if( UseTLAB && AllocatePrefetchStyle == 2 ) {
1622 // Generate prefetch allocation with watermark check.
1623 // As an allocation hits the watermark, we will prefetch starting
1624 // at a "distance" away from watermark.
1626 Node *pf_region = new (C, 3) RegionNode(3);
1627 Node *pf_phi_rawmem = new (C, 3) PhiNode( pf_region, Type::MEMORY,
1628 TypeRawPtr::BOTTOM );
1629 // I/O is used for Prefetch
1630 Node *pf_phi_abio = new (C, 3) PhiNode( pf_region, Type::ABIO );
1632 Node *thread = new (C, 1) ThreadLocalNode();
1633 transform_later(thread);
1635 Node *eden_pf_adr = new (C, 4) AddPNode( top()/*not oop*/, thread,
1636 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) );
1637 transform_later(eden_pf_adr);
1639 Node *old_pf_wm = new (C, 3) LoadPNode( needgc_false,
1640 contended_phi_rawmem, eden_pf_adr,
1641 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM );
1642 transform_later(old_pf_wm);
1644 // check against new_eden_top
1645 Node *need_pf_cmp = new (C, 3) CmpPNode( new_eden_top, old_pf_wm );
1646 transform_later(need_pf_cmp);
1647 Node *need_pf_bol = new (C, 2) BoolNode( need_pf_cmp, BoolTest::ge );
1648 transform_later(need_pf_bol);
1649 IfNode *need_pf_iff = new (C, 2) IfNode( needgc_false, need_pf_bol,
1650 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN );
1651 transform_later(need_pf_iff);
1653 // true node, add prefetchdistance
1654 Node *need_pf_true = new (C, 1) IfTrueNode( need_pf_iff );
1655 transform_later(need_pf_true);
1657 Node *need_pf_false = new (C, 1) IfFalseNode( need_pf_iff );
1658 transform_later(need_pf_false);
1660 Node *new_pf_wmt = new (C, 4) AddPNode( top(), old_pf_wm,
1661 _igvn.MakeConX(AllocatePrefetchDistance) );
1662 transform_later(new_pf_wmt );
1663 new_pf_wmt->set_req(0, need_pf_true);
1665 Node *store_new_wmt = new (C, 4) StorePNode( need_pf_true,
1666 contended_phi_rawmem, eden_pf_adr,
1667 TypeRawPtr::BOTTOM, new_pf_wmt );
1668 transform_later(store_new_wmt);
1670 // adding prefetches
1671 pf_phi_abio->init_req( fall_in_path, i_o );
1673 Node *prefetch_adr;
1674 Node *prefetch;
1675 uint lines = AllocatePrefetchDistance / AllocatePrefetchStepSize;
1676 uint step_size = AllocatePrefetchStepSize;
1677 uint distance = 0;
1679 for ( uint i = 0; i < lines; i++ ) {
1680 prefetch_adr = new (C, 4) AddPNode( old_pf_wm, new_pf_wmt,
1681 _igvn.MakeConX(distance) );
1682 transform_later(prefetch_adr);
1683 prefetch = new (C, 3) PrefetchAllocationNode( i_o, prefetch_adr );
1684 transform_later(prefetch);
1685 distance += step_size;
1686 i_o = prefetch;
1687 }
1688 pf_phi_abio->set_req( pf_path, i_o );
1690 pf_region->init_req( fall_in_path, need_pf_false );
1691 pf_region->init_req( pf_path, need_pf_true );
1693 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem );
1694 pf_phi_rawmem->init_req( pf_path, store_new_wmt );
1696 transform_later(pf_region);
1697 transform_later(pf_phi_rawmem);
1698 transform_later(pf_phi_abio);
1700 needgc_false = pf_region;
1701 contended_phi_rawmem = pf_phi_rawmem;
1702 i_o = pf_phi_abio;
1703 } else if( UseTLAB && AllocatePrefetchStyle == 3 ) {
1704 // Insert a prefetch for each allocation.
1705 // This code is used for Sparc with BIS.
1706 Node *pf_region = new (C, 3) RegionNode(3);
1707 Node *pf_phi_rawmem = new (C, 3) PhiNode( pf_region, Type::MEMORY,
1708 TypeRawPtr::BOTTOM );
1710 // Generate several prefetch instructions.
1711 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1712 uint step_size = AllocatePrefetchStepSize;
1713 uint distance = AllocatePrefetchDistance;
1715 // Next cache address.
1716 Node *cache_adr = new (C, 4) AddPNode(old_eden_top, old_eden_top,
1717 _igvn.MakeConX(distance));
1718 transform_later(cache_adr);
1719 cache_adr = new (C, 2) CastP2XNode(needgc_false, cache_adr);
1720 transform_later(cache_adr);
1721 Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1));
1722 cache_adr = new (C, 3) AndXNode(cache_adr, mask);
1723 transform_later(cache_adr);
1724 cache_adr = new (C, 2) CastX2PNode(cache_adr);
1725 transform_later(cache_adr);
1727 // Prefetch
1728 Node *prefetch = new (C, 3) PrefetchAllocationNode( contended_phi_rawmem, cache_adr );
1729 prefetch->set_req(0, needgc_false);
1730 transform_later(prefetch);
1731 contended_phi_rawmem = prefetch;
1732 Node *prefetch_adr;
1733 distance = step_size;
1734 for ( uint i = 1; i < lines; i++ ) {
1735 prefetch_adr = new (C, 4) AddPNode( cache_adr, cache_adr,
1736 _igvn.MakeConX(distance) );
1737 transform_later(prefetch_adr);
1738 prefetch = new (C, 3) PrefetchAllocationNode( contended_phi_rawmem, prefetch_adr );
1739 transform_later(prefetch);
1740 distance += step_size;
1741 contended_phi_rawmem = prefetch;
1742 }
1743 } else if( AllocatePrefetchStyle > 0 ) {
1744 // Insert a prefetch for each allocation only on the fast-path
1745 Node *prefetch_adr;
1746 Node *prefetch;
1747 // Generate several prefetch instructions.
1748 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1749 uint step_size = AllocatePrefetchStepSize;
1750 uint distance = AllocatePrefetchDistance;
1751 for ( uint i = 0; i < lines; i++ ) {
1752 prefetch_adr = new (C, 4) AddPNode( old_eden_top, new_eden_top,
1753 _igvn.MakeConX(distance) );
1754 transform_later(prefetch_adr);
1755 prefetch = new (C, 3) PrefetchAllocationNode( i_o, prefetch_adr );
1756 // Do not let it float too high, since if eden_top == eden_end,
1757 // both might be null.
1758 if( i == 0 ) { // Set control for first prefetch, next follows it
1759 prefetch->init_req(0, needgc_false);
1760 }
1761 transform_later(prefetch);
1762 distance += step_size;
1763 i_o = prefetch;
1764 }
1765 }
1766 return i_o;
1767 }
1770 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) {
1771 expand_allocate_common(alloc, NULL,
1772 OptoRuntime::new_instance_Type(),
1773 OptoRuntime::new_instance_Java());
1774 }
1776 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) {
1777 Node* length = alloc->in(AllocateNode::ALength);
1778 InitializeNode* init = alloc->initialization();
1779 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1780 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
1781 address slow_call_address; // Address of slow call
1782 if (init != NULL && init->is_complete_with_arraycopy() &&
1783 k->is_type_array_klass()) {
1784 // Don't zero type array during slow allocation in VM since
1785 // it will be initialized later by arraycopy in compiled code.
1786 slow_call_address = OptoRuntime::new_array_nozero_Java();
1787 } else {
1788 slow_call_address = OptoRuntime::new_array_Java();
1789 }
1790 expand_allocate_common(alloc, length,
1791 OptoRuntime::new_array_Type(),
1792 slow_call_address);
1793 }
1795 //-------------------mark_eliminated_box----------------------------------
1796 //
1797 // During EA obj may point to several objects but after few ideal graph
1798 // transformations (CCP) it may point to only one non escaping object
1799 // (but still using phi), corresponding locks and unlocks will be marked
1800 // for elimination. Later obj could be replaced with a new node (new phi)
1801 // and which does not have escape information. And later after some graph
1802 // reshape other locks and unlocks (which were not marked for elimination
1803 // before) are connected to this new obj (phi) but they still will not be
1804 // marked for elimination since new obj has no escape information.
1805 // Mark all associated (same box and obj) lock and unlock nodes for
1806 // elimination if some of them marked already.
1807 void PhaseMacroExpand::mark_eliminated_box(Node* oldbox, Node* obj) {
1808 if (oldbox->as_BoxLock()->is_eliminated())
1809 return; // This BoxLock node was processed already.
1811 // New implementation (EliminateNestedLocks) has separate BoxLock
1812 // node for each locked region so mark all associated locks/unlocks as
1813 // eliminated even if different objects are referenced in one locked region
1814 // (for example, OSR compilation of nested loop inside locked scope).
1815 if (EliminateNestedLocks ||
1816 oldbox->as_BoxLock()->is_simple_lock_region(NULL, obj)) {
1817 // Box is used only in one lock region. Mark this box as eliminated.
1818 _igvn.hash_delete(oldbox);
1819 oldbox->as_BoxLock()->set_eliminated(); // This changes box's hash value
1820 _igvn.hash_insert(oldbox);
1822 for (uint i = 0; i < oldbox->outcnt(); i++) {
1823 Node* u = oldbox->raw_out(i);
1824 if (u->is_AbstractLock() && !u->as_AbstractLock()->is_non_esc_obj()) {
1825 AbstractLockNode* alock = u->as_AbstractLock();
1826 // Check lock's box since box could be referenced by Lock's debug info.
1827 if (alock->box_node() == oldbox) {
1828 // Mark eliminated all related locks and unlocks.
1829 alock->set_non_esc_obj();
1830 }
1831 }
1832 }
1833 return;
1834 }
1836 // Create new "eliminated" BoxLock node and use it in monitor debug info
1837 // instead of oldbox for the same object.
1838 BoxLockNode* newbox = oldbox->clone()->as_BoxLock();
1840 // Note: BoxLock node is marked eliminated only here and it is used
1841 // to indicate that all associated lock and unlock nodes are marked
1842 // for elimination.
1843 newbox->set_eliminated();
1844 transform_later(newbox);
1846 // Replace old box node with new box for all users of the same object.
1847 for (uint i = 0; i < oldbox->outcnt();) {
1848 bool next_edge = true;
1850 Node* u = oldbox->raw_out(i);
1851 if (u->is_AbstractLock()) {
1852 AbstractLockNode* alock = u->as_AbstractLock();
1853 if (alock->box_node() == oldbox && alock->obj_node()->eqv_uncast(obj)) {
1854 // Replace Box and mark eliminated all related locks and unlocks.
1855 alock->set_non_esc_obj();
1856 _igvn.rehash_node_delayed(alock);
1857 alock->set_box_node(newbox);
1858 next_edge = false;
1859 }
1860 }
1861 if (u->is_FastLock() && u->as_FastLock()->obj_node()->eqv_uncast(obj)) {
1862 FastLockNode* flock = u->as_FastLock();
1863 assert(flock->box_node() == oldbox, "sanity");
1864 _igvn.rehash_node_delayed(flock);
1865 flock->set_box_node(newbox);
1866 next_edge = false;
1867 }
1869 // Replace old box in monitor debug info.
1870 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) {
1871 SafePointNode* sfn = u->as_SafePoint();
1872 JVMState* youngest_jvms = sfn->jvms();
1873 int max_depth = youngest_jvms->depth();
1874 for (int depth = 1; depth <= max_depth; depth++) {
1875 JVMState* jvms = youngest_jvms->of_depth(depth);
1876 int num_mon = jvms->nof_monitors();
1877 // Loop over monitors
1878 for (int idx = 0; idx < num_mon; idx++) {
1879 Node* obj_node = sfn->monitor_obj(jvms, idx);
1880 Node* box_node = sfn->monitor_box(jvms, idx);
1881 if (box_node == oldbox && obj_node->eqv_uncast(obj)) {
1882 int j = jvms->monitor_box_offset(idx);
1883 _igvn.replace_input_of(u, j, newbox);
1884 next_edge = false;
1885 }
1886 }
1887 }
1888 }
1889 if (next_edge) i++;
1890 }
1891 }
1893 //-----------------------mark_eliminated_locking_nodes-----------------------
1894 void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) {
1895 if (EliminateNestedLocks) {
1896 if (alock->is_nested()) {
1897 assert(alock->box_node()->as_BoxLock()->is_eliminated(), "sanity");
1898 return;
1899 } else if (!alock->is_non_esc_obj()) { // Not eliminated or coarsened
1900 // Only Lock node has JVMState needed here.
1901 if (alock->jvms() != NULL && alock->as_Lock()->is_nested_lock_region()) {
1902 // Mark eliminated related nested locks and unlocks.
1903 Node* obj = alock->obj_node();
1904 BoxLockNode* box_node = alock->box_node()->as_BoxLock();
1905 assert(!box_node->is_eliminated(), "should not be marked yet");
1906 // Note: BoxLock node is marked eliminated only here
1907 // and it is used to indicate that all associated lock
1908 // and unlock nodes are marked for elimination.
1909 box_node->set_eliminated(); // Box's hash is always NO_HASH here
1910 for (uint i = 0; i < box_node->outcnt(); i++) {
1911 Node* u = box_node->raw_out(i);
1912 if (u->is_AbstractLock()) {
1913 alock = u->as_AbstractLock();
1914 if (alock->box_node() == box_node) {
1915 // Verify that this Box is referenced only by related locks.
1916 assert(alock->obj_node()->eqv_uncast(obj), "");
1917 // Mark all related locks and unlocks.
1918 alock->set_nested();
1919 }
1920 }
1921 }
1922 }
1923 return;
1924 }
1925 // Process locks for non escaping object
1926 assert(alock->is_non_esc_obj(), "");
1927 } // EliminateNestedLocks
1929 if (alock->is_non_esc_obj()) { // Lock is used for non escaping object
1930 // Look for all locks of this object and mark them and
1931 // corresponding BoxLock nodes as eliminated.
1932 Node* obj = alock->obj_node();
1933 for (uint j = 0; j < obj->outcnt(); j++) {
1934 Node* o = obj->raw_out(j);
1935 if (o->is_AbstractLock() &&
1936 o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) {
1937 alock = o->as_AbstractLock();
1938 Node* box = alock->box_node();
1939 // Replace old box node with new eliminated box for all users
1940 // of the same object and mark related locks as eliminated.
1941 mark_eliminated_box(box, obj);
1942 }
1943 }
1944 }
1945 }
1947 // we have determined that this lock/unlock can be eliminated, we simply
1948 // eliminate the node without expanding it.
1949 //
1950 // Note: The membar's associated with the lock/unlock are currently not
1951 // eliminated. This should be investigated as a future enhancement.
1952 //
1953 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) {
1955 if (!alock->is_eliminated()) {
1956 return false;
1957 }
1958 #ifdef ASSERT
1959 if (!alock->is_coarsened()) {
1960 // Check that new "eliminated" BoxLock node is created.
1961 BoxLockNode* oldbox = alock->box_node()->as_BoxLock();
1962 assert(oldbox->is_eliminated(), "should be done already");
1963 }
1964 #endif
1965 CompileLog* log = C->log();
1966 if (log != NULL) {
1967 log->head("eliminate_lock lock='%d'",
1968 alock->is_Lock());
1969 JVMState* p = alock->jvms();
1970 while (p != NULL) {
1971 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1972 p = p->caller();
1973 }
1974 log->tail("eliminate_lock");
1975 }
1977 #ifndef PRODUCT
1978 if (PrintEliminateLocks) {
1979 if (alock->is_Lock()) {
1980 tty->print_cr("++++ Eliminated: %d Lock", alock->_idx);
1981 } else {
1982 tty->print_cr("++++ Eliminated: %d Unlock", alock->_idx);
1983 }
1984 }
1985 #endif
1987 Node* mem = alock->in(TypeFunc::Memory);
1988 Node* ctrl = alock->in(TypeFunc::Control);
1990 extract_call_projections(alock);
1991 // There are 2 projections from the lock. The lock node will
1992 // be deleted when its last use is subsumed below.
1993 assert(alock->outcnt() == 2 &&
1994 _fallthroughproj != NULL &&
1995 _memproj_fallthrough != NULL,
1996 "Unexpected projections from Lock/Unlock");
1998 Node* fallthroughproj = _fallthroughproj;
1999 Node* memproj_fallthrough = _memproj_fallthrough;
2001 // The memory projection from a lock/unlock is RawMem
2002 // The input to a Lock is merged memory, so extract its RawMem input
2003 // (unless the MergeMem has been optimized away.)
2004 if (alock->is_Lock()) {
2005 // Seach for MemBarAcquireLock node and delete it also.
2006 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
2007 assert(membar != NULL && membar->Opcode() == Op_MemBarAcquireLock, "");
2008 Node* ctrlproj = membar->proj_out(TypeFunc::Control);
2009 Node* memproj = membar->proj_out(TypeFunc::Memory);
2010 _igvn.replace_node(ctrlproj, fallthroughproj);
2011 _igvn.replace_node(memproj, memproj_fallthrough);
2013 // Delete FastLock node also if this Lock node is unique user
2014 // (a loop peeling may clone a Lock node).
2015 Node* flock = alock->as_Lock()->fastlock_node();
2016 if (flock->outcnt() == 1) {
2017 assert(flock->unique_out() == alock, "sanity");
2018 _igvn.replace_node(flock, top());
2019 }
2020 }
2022 // Seach for MemBarReleaseLock node and delete it also.
2023 if (alock->is_Unlock() && ctrl != NULL && ctrl->is_Proj() &&
2024 ctrl->in(0)->is_MemBar()) {
2025 MemBarNode* membar = ctrl->in(0)->as_MemBar();
2026 assert(membar->Opcode() == Op_MemBarReleaseLock &&
2027 mem->is_Proj() && membar == mem->in(0), "");
2028 _igvn.replace_node(fallthroughproj, ctrl);
2029 _igvn.replace_node(memproj_fallthrough, mem);
2030 fallthroughproj = ctrl;
2031 memproj_fallthrough = mem;
2032 ctrl = membar->in(TypeFunc::Control);
2033 mem = membar->in(TypeFunc::Memory);
2034 }
2036 _igvn.replace_node(fallthroughproj, ctrl);
2037 _igvn.replace_node(memproj_fallthrough, mem);
2038 return true;
2039 }
2042 //------------------------------expand_lock_node----------------------
2043 void PhaseMacroExpand::expand_lock_node(LockNode *lock) {
2045 Node* ctrl = lock->in(TypeFunc::Control);
2046 Node* mem = lock->in(TypeFunc::Memory);
2047 Node* obj = lock->obj_node();
2048 Node* box = lock->box_node();
2049 Node* flock = lock->fastlock_node();
2051 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2053 // Make the merge point
2054 Node *region;
2055 Node *mem_phi;
2056 Node *slow_path;
2058 if (UseOptoBiasInlining) {
2059 /*
2060 * See the full description in MacroAssembler::biased_locking_enter().
2061 *
2062 * if( (mark_word & biased_lock_mask) == biased_lock_pattern ) {
2063 * // The object is biased.
2064 * proto_node = klass->prototype_header;
2065 * o_node = thread | proto_node;
2066 * x_node = o_node ^ mark_word;
2067 * if( (x_node & ~age_mask) == 0 ) { // Biased to the current thread ?
2068 * // Done.
2069 * } else {
2070 * if( (x_node & biased_lock_mask) != 0 ) {
2071 * // The klass's prototype header is no longer biased.
2072 * cas(&mark_word, mark_word, proto_node)
2073 * goto cas_lock;
2074 * } else {
2075 * // The klass's prototype header is still biased.
2076 * if( (x_node & epoch_mask) != 0 ) { // Expired epoch?
2077 * old = mark_word;
2078 * new = o_node;
2079 * } else {
2080 * // Different thread or anonymous biased.
2081 * old = mark_word & (epoch_mask | age_mask | biased_lock_mask);
2082 * new = thread | old;
2083 * }
2084 * // Try to rebias.
2085 * if( cas(&mark_word, old, new) == 0 ) {
2086 * // Done.
2087 * } else {
2088 * goto slow_path; // Failed.
2089 * }
2090 * }
2091 * }
2092 * } else {
2093 * // The object is not biased.
2094 * cas_lock:
2095 * if( FastLock(obj) == 0 ) {
2096 * // Done.
2097 * } else {
2098 * slow_path:
2099 * OptoRuntime::complete_monitor_locking_Java(obj);
2100 * }
2101 * }
2102 */
2104 region = new (C, 5) RegionNode(5);
2105 // create a Phi for the memory state
2106 mem_phi = new (C, 5) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2108 Node* fast_lock_region = new (C, 3) RegionNode(3);
2109 Node* fast_lock_mem_phi = new (C, 3) PhiNode( fast_lock_region, Type::MEMORY, TypeRawPtr::BOTTOM);
2111 // First, check mark word for the biased lock pattern.
2112 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2114 // Get fast path - mark word has the biased lock pattern.
2115 ctrl = opt_bits_test(ctrl, fast_lock_region, 1, mark_node,
2116 markOopDesc::biased_lock_mask_in_place,
2117 markOopDesc::biased_lock_pattern, true);
2118 // fast_lock_region->in(1) is set to slow path.
2119 fast_lock_mem_phi->init_req(1, mem);
2121 // Now check that the lock is biased to the current thread and has
2122 // the same epoch and bias as Klass::_prototype_header.
2124 // Special-case a fresh allocation to avoid building nodes:
2125 Node* klass_node = AllocateNode::Ideal_klass(obj, &_igvn);
2126 if (klass_node == NULL) {
2127 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
2128 klass_node = transform_later( LoadKlassNode::make(_igvn, mem, k_adr, _igvn.type(k_adr)->is_ptr()) );
2129 #ifdef _LP64
2130 if (UseCompressedOops && klass_node->is_DecodeN()) {
2131 assert(klass_node->in(1)->Opcode() == Op_LoadNKlass, "sanity");
2132 klass_node->in(1)->init_req(0, ctrl);
2133 } else
2134 #endif
2135 klass_node->init_req(0, ctrl);
2136 }
2137 Node *proto_node = make_load(ctrl, mem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeX_X, TypeX_X->basic_type());
2139 Node* thread = transform_later(new (C, 1) ThreadLocalNode());
2140 Node* cast_thread = transform_later(new (C, 2) CastP2XNode(ctrl, thread));
2141 Node* o_node = transform_later(new (C, 3) OrXNode(cast_thread, proto_node));
2142 Node* x_node = transform_later(new (C, 3) XorXNode(o_node, mark_node));
2144 // Get slow path - mark word does NOT match the value.
2145 Node* not_biased_ctrl = opt_bits_test(ctrl, region, 3, x_node,
2146 (~markOopDesc::age_mask_in_place), 0);
2147 // region->in(3) is set to fast path - the object is biased to the current thread.
2148 mem_phi->init_req(3, mem);
2151 // Mark word does NOT match the value (thread | Klass::_prototype_header).
2154 // First, check biased pattern.
2155 // Get fast path - _prototype_header has the same biased lock pattern.
2156 ctrl = opt_bits_test(not_biased_ctrl, fast_lock_region, 2, x_node,
2157 markOopDesc::biased_lock_mask_in_place, 0, true);
2159 not_biased_ctrl = fast_lock_region->in(2); // Slow path
2160 // fast_lock_region->in(2) - the prototype header is no longer biased
2161 // and we have to revoke the bias on this object.
2162 // We are going to try to reset the mark of this object to the prototype
2163 // value and fall through to the CAS-based locking scheme.
2164 Node* adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
2165 Node* cas = new (C, 5) StoreXConditionalNode(not_biased_ctrl, mem, adr,
2166 proto_node, mark_node);
2167 transform_later(cas);
2168 Node* proj = transform_later( new (C, 1) SCMemProjNode(cas));
2169 fast_lock_mem_phi->init_req(2, proj);
2172 // Second, check epoch bits.
2173 Node* rebiased_region = new (C, 3) RegionNode(3);
2174 Node* old_phi = new (C, 3) PhiNode( rebiased_region, TypeX_X);
2175 Node* new_phi = new (C, 3) PhiNode( rebiased_region, TypeX_X);
2177 // Get slow path - mark word does NOT match epoch bits.
2178 Node* epoch_ctrl = opt_bits_test(ctrl, rebiased_region, 1, x_node,
2179 markOopDesc::epoch_mask_in_place, 0);
2180 // The epoch of the current bias is not valid, attempt to rebias the object
2181 // toward the current thread.
2182 rebiased_region->init_req(2, epoch_ctrl);
2183 old_phi->init_req(2, mark_node);
2184 new_phi->init_req(2, o_node);
2186 // rebiased_region->in(1) is set to fast path.
2187 // The epoch of the current bias is still valid but we know
2188 // nothing about the owner; it might be set or it might be clear.
2189 Node* cmask = MakeConX(markOopDesc::biased_lock_mask_in_place |
2190 markOopDesc::age_mask_in_place |
2191 markOopDesc::epoch_mask_in_place);
2192 Node* old = transform_later(new (C, 3) AndXNode(mark_node, cmask));
2193 cast_thread = transform_later(new (C, 2) CastP2XNode(ctrl, thread));
2194 Node* new_mark = transform_later(new (C, 3) OrXNode(cast_thread, old));
2195 old_phi->init_req(1, old);
2196 new_phi->init_req(1, new_mark);
2198 transform_later(rebiased_region);
2199 transform_later(old_phi);
2200 transform_later(new_phi);
2202 // Try to acquire the bias of the object using an atomic operation.
2203 // If this fails we will go in to the runtime to revoke the object's bias.
2204 cas = new (C, 5) StoreXConditionalNode(rebiased_region, mem, adr,
2205 new_phi, old_phi);
2206 transform_later(cas);
2207 proj = transform_later( new (C, 1) SCMemProjNode(cas));
2209 // Get slow path - Failed to CAS.
2210 not_biased_ctrl = opt_bits_test(rebiased_region, region, 4, cas, 0, 0);
2211 mem_phi->init_req(4, proj);
2212 // region->in(4) is set to fast path - the object is rebiased to the current thread.
2214 // Failed to CAS.
2215 slow_path = new (C, 3) RegionNode(3);
2216 Node *slow_mem = new (C, 3) PhiNode( slow_path, Type::MEMORY, TypeRawPtr::BOTTOM);
2218 slow_path->init_req(1, not_biased_ctrl); // Capture slow-control
2219 slow_mem->init_req(1, proj);
2221 // Call CAS-based locking scheme (FastLock node).
2223 transform_later(fast_lock_region);
2224 transform_later(fast_lock_mem_phi);
2226 // Get slow path - FastLock failed to lock the object.
2227 ctrl = opt_bits_test(fast_lock_region, region, 2, flock, 0, 0);
2228 mem_phi->init_req(2, fast_lock_mem_phi);
2229 // region->in(2) is set to fast path - the object is locked to the current thread.
2231 slow_path->init_req(2, ctrl); // Capture slow-control
2232 slow_mem->init_req(2, fast_lock_mem_phi);
2234 transform_later(slow_path);
2235 transform_later(slow_mem);
2236 // Reset lock's memory edge.
2237 lock->set_req(TypeFunc::Memory, slow_mem);
2239 } else {
2240 region = new (C, 3) RegionNode(3);
2241 // create a Phi for the memory state
2242 mem_phi = new (C, 3) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2244 // Optimize test; set region slot 2
2245 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0);
2246 mem_phi->init_req(2, mem);
2247 }
2249 // Make slow path call
2250 CallNode *call = make_slow_call( (CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(), OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path, obj, box );
2252 extract_call_projections(call);
2254 // Slow path can only throw asynchronous exceptions, which are always
2255 // de-opted. So the compiler thinks the slow-call can never throw an
2256 // exception. If it DOES throw an exception we would need the debug
2257 // info removed first (since if it throws there is no monitor).
2258 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
2259 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
2261 // Capture slow path
2262 // disconnect fall-through projection from call and create a new one
2263 // hook up users of fall-through projection to region
2264 Node *slow_ctrl = _fallthroughproj->clone();
2265 transform_later(slow_ctrl);
2266 _igvn.hash_delete(_fallthroughproj);
2267 _fallthroughproj->disconnect_inputs(NULL);
2268 region->init_req(1, slow_ctrl);
2269 // region inputs are now complete
2270 transform_later(region);
2271 _igvn.replace_node(_fallthroughproj, region);
2273 Node *memproj = transform_later( new(C, 1) ProjNode(call, TypeFunc::Memory) );
2274 mem_phi->init_req(1, memproj );
2275 transform_later(mem_phi);
2276 _igvn.replace_node(_memproj_fallthrough, mem_phi);
2277 }
2279 //------------------------------expand_unlock_node----------------------
2280 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
2282 Node* ctrl = unlock->in(TypeFunc::Control);
2283 Node* mem = unlock->in(TypeFunc::Memory);
2284 Node* obj = unlock->obj_node();
2285 Node* box = unlock->box_node();
2287 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2289 // No need for a null check on unlock
2291 // Make the merge point
2292 Node *region;
2293 Node *mem_phi;
2295 if (UseOptoBiasInlining) {
2296 // Check for biased locking unlock case, which is a no-op.
2297 // See the full description in MacroAssembler::biased_locking_exit().
2298 region = new (C, 4) RegionNode(4);
2299 // create a Phi for the memory state
2300 mem_phi = new (C, 4) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2301 mem_phi->init_req(3, mem);
2303 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2304 ctrl = opt_bits_test(ctrl, region, 3, mark_node,
2305 markOopDesc::biased_lock_mask_in_place,
2306 markOopDesc::biased_lock_pattern);
2307 } else {
2308 region = new (C, 3) RegionNode(3);
2309 // create a Phi for the memory state
2310 mem_phi = new (C, 3) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2311 }
2313 FastUnlockNode *funlock = new (C, 3) FastUnlockNode( ctrl, obj, box );
2314 funlock = transform_later( funlock )->as_FastUnlock();
2315 // Optimize test; set region slot 2
2316 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0);
2318 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 );
2320 extract_call_projections(call);
2322 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
2323 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
2325 // No exceptions for unlocking
2326 // Capture slow path
2327 // disconnect fall-through projection from call and create a new one
2328 // hook up users of fall-through projection to region
2329 Node *slow_ctrl = _fallthroughproj->clone();
2330 transform_later(slow_ctrl);
2331 _igvn.hash_delete(_fallthroughproj);
2332 _fallthroughproj->disconnect_inputs(NULL);
2333 region->init_req(1, slow_ctrl);
2334 // region inputs are now complete
2335 transform_later(region);
2336 _igvn.replace_node(_fallthroughproj, region);
2338 Node *memproj = transform_later( new(C, 1) ProjNode(call, TypeFunc::Memory) );
2339 mem_phi->init_req(1, memproj );
2340 mem_phi->init_req(2, mem);
2341 transform_later(mem_phi);
2342 _igvn.replace_node(_memproj_fallthrough, mem_phi);
2343 }
2345 //---------------------------eliminate_macro_nodes----------------------
2346 // Eliminate scalar replaced allocations and associated locks.
2347 void PhaseMacroExpand::eliminate_macro_nodes() {
2348 if (C->macro_count() == 0)
2349 return;
2351 // First, attempt to eliminate locks
2352 int cnt = C->macro_count();
2353 for (int i=0; i < cnt; i++) {
2354 Node *n = C->macro_node(i);
2355 if (n->is_AbstractLock()) { // Lock and Unlock nodes
2356 // Before elimination mark all associated (same box and obj)
2357 // lock and unlock nodes.
2358 mark_eliminated_locking_nodes(n->as_AbstractLock());
2359 }
2360 }
2361 bool progress = true;
2362 while (progress) {
2363 progress = false;
2364 for (int i = C->macro_count(); i > 0; i--) {
2365 Node * n = C->macro_node(i-1);
2366 bool success = false;
2367 debug_only(int old_macro_count = C->macro_count(););
2368 if (n->is_AbstractLock()) {
2369 success = eliminate_locking_node(n->as_AbstractLock());
2370 }
2371 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2372 progress = progress || success;
2373 }
2374 }
2375 // Next, attempt to eliminate allocations
2376 progress = true;
2377 while (progress) {
2378 progress = false;
2379 for (int i = C->macro_count(); i > 0; i--) {
2380 Node * n = C->macro_node(i-1);
2381 bool success = false;
2382 debug_only(int old_macro_count = C->macro_count(););
2383 switch (n->class_id()) {
2384 case Node::Class_Allocate:
2385 case Node::Class_AllocateArray:
2386 success = eliminate_allocate_node(n->as_Allocate());
2387 break;
2388 case Node::Class_Lock:
2389 case Node::Class_Unlock:
2390 assert(!n->as_AbstractLock()->is_eliminated(), "sanity");
2391 break;
2392 default:
2393 assert(n->Opcode() == Op_LoopLimit ||
2394 n->Opcode() == Op_Opaque1 ||
2395 n->Opcode() == Op_Opaque2, "unknown node type in macro list");
2396 }
2397 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2398 progress = progress || success;
2399 }
2400 }
2401 }
2403 //------------------------------expand_macro_nodes----------------------
2404 // Returns true if a failure occurred.
2405 bool PhaseMacroExpand::expand_macro_nodes() {
2406 // Last attempt to eliminate macro nodes.
2407 eliminate_macro_nodes();
2409 // Make sure expansion will not cause node limit to be exceeded.
2410 // Worst case is a macro node gets expanded into about 50 nodes.
2411 // Allow 50% more for optimization.
2412 if (C->check_node_count(C->macro_count() * 75, "out of nodes before macro expansion" ) )
2413 return true;
2415 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations.
2416 bool progress = true;
2417 while (progress) {
2418 progress = false;
2419 for (int i = C->macro_count(); i > 0; i--) {
2420 Node * n = C->macro_node(i-1);
2421 bool success = false;
2422 debug_only(int old_macro_count = C->macro_count(););
2423 if (n->Opcode() == Op_LoopLimit) {
2424 // Remove it from macro list and put on IGVN worklist to optimize.
2425 C->remove_macro_node(n);
2426 _igvn._worklist.push(n);
2427 success = true;
2428 } else if (n->Opcode() == Op_Opaque1 || n->Opcode() == Op_Opaque2) {
2429 _igvn.replace_node(n, n->in(1));
2430 success = true;
2431 }
2432 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2433 progress = progress || success;
2434 }
2435 }
2437 // expand "macro" nodes
2438 // nodes are removed from the macro list as they are processed
2439 while (C->macro_count() > 0) {
2440 int macro_count = C->macro_count();
2441 Node * n = C->macro_node(macro_count-1);
2442 assert(n->is_macro(), "only macro nodes expected here");
2443 if (_igvn.type(n) == Type::TOP || n->in(0)->is_top() ) {
2444 // node is unreachable, so don't try to expand it
2445 C->remove_macro_node(n);
2446 continue;
2447 }
2448 switch (n->class_id()) {
2449 case Node::Class_Allocate:
2450 expand_allocate(n->as_Allocate());
2451 break;
2452 case Node::Class_AllocateArray:
2453 expand_allocate_array(n->as_AllocateArray());
2454 break;
2455 case Node::Class_Lock:
2456 expand_lock_node(n->as_Lock());
2457 break;
2458 case Node::Class_Unlock:
2459 expand_unlock_node(n->as_Unlock());
2460 break;
2461 default:
2462 assert(false, "unknown node type in macro list");
2463 }
2464 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
2465 if (C->failing()) return true;
2466 }
2468 _igvn.set_delay_transform(false);
2469 _igvn.optimize();
2470 if (C->failing()) return true;
2471 return false;
2472 }