Tue, 09 Mar 2010 20:16:19 +0100
6919934: JSR 292 needs to support x86 C1
Summary: This implements JSR 292 support for C1 x86.
Reviewed-by: never, jrose, kvn
1 /*
2 * Copyright 2005-2009 Sun Microsystems, Inc. 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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
20 * CA 95054 USA or visit www.sun.com if you need additional information or
21 * have any questions.
22 *
23 */
25 #include "incls/_precompiled.incl"
26 #include "incls/_macro.cpp.incl"
29 //
30 // Replace any references to "oldref" in inputs to "use" with "newref".
31 // Returns the number of replacements made.
32 //
33 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) {
34 int nreplacements = 0;
35 uint req = use->req();
36 for (uint j = 0; j < use->len(); j++) {
37 Node *uin = use->in(j);
38 if (uin == oldref) {
39 if (j < req)
40 use->set_req(j, newref);
41 else
42 use->set_prec(j, newref);
43 nreplacements++;
44 } else if (j >= req && uin == NULL) {
45 break;
46 }
47 }
48 return nreplacements;
49 }
51 void PhaseMacroExpand::copy_call_debug_info(CallNode *oldcall, CallNode * newcall) {
52 // Copy debug information and adjust JVMState information
53 uint old_dbg_start = oldcall->tf()->domain()->cnt();
54 uint new_dbg_start = newcall->tf()->domain()->cnt();
55 int jvms_adj = new_dbg_start - old_dbg_start;
56 assert (new_dbg_start == newcall->req(), "argument count mismatch");
58 Dict* sosn_map = new Dict(cmpkey,hashkey);
59 for (uint i = old_dbg_start; i < oldcall->req(); i++) {
60 Node* old_in = oldcall->in(i);
61 // Clone old SafePointScalarObjectNodes, adjusting their field contents.
62 if (old_in != NULL && old_in->is_SafePointScalarObject()) {
63 SafePointScalarObjectNode* old_sosn = old_in->as_SafePointScalarObject();
64 uint old_unique = C->unique();
65 Node* new_in = old_sosn->clone(jvms_adj, sosn_map);
66 if (old_unique != C->unique()) {
67 new_in->set_req(0, newcall->in(0)); // reset control edge
68 new_in = transform_later(new_in); // Register new node.
69 }
70 old_in = new_in;
71 }
72 newcall->add_req(old_in);
73 }
75 newcall->set_jvms(oldcall->jvms());
76 for (JVMState *jvms = newcall->jvms(); jvms != NULL; jvms = jvms->caller()) {
77 jvms->set_map(newcall);
78 jvms->set_locoff(jvms->locoff()+jvms_adj);
79 jvms->set_stkoff(jvms->stkoff()+jvms_adj);
80 jvms->set_monoff(jvms->monoff()+jvms_adj);
81 jvms->set_scloff(jvms->scloff()+jvms_adj);
82 jvms->set_endoff(jvms->endoff()+jvms_adj);
83 }
84 }
86 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word, int mask, int bits, bool return_fast_path) {
87 Node* cmp;
88 if (mask != 0) {
89 Node* and_node = transform_later(new (C, 3) AndXNode(word, MakeConX(mask)));
90 cmp = transform_later(new (C, 3) CmpXNode(and_node, MakeConX(bits)));
91 } else {
92 cmp = word;
93 }
94 Node* bol = transform_later(new (C, 2) BoolNode(cmp, BoolTest::ne));
95 IfNode* iff = new (C, 2) IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN );
96 transform_later(iff);
98 // Fast path taken.
99 Node *fast_taken = transform_later( new (C, 1) IfFalseNode(iff) );
101 // Fast path not-taken, i.e. slow path
102 Node *slow_taken = transform_later( new (C, 1) IfTrueNode(iff) );
104 if (return_fast_path) {
105 region->init_req(edge, slow_taken); // Capture slow-control
106 return fast_taken;
107 } else {
108 region->init_req(edge, fast_taken); // Capture fast-control
109 return slow_taken;
110 }
111 }
113 //--------------------copy_predefined_input_for_runtime_call--------------------
114 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) {
115 // Set fixed predefined input arguments
116 call->init_req( TypeFunc::Control, ctrl );
117 call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) );
118 call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ?????
119 call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) );
120 call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) );
121 }
123 //------------------------------make_slow_call---------------------------------
124 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) {
126 // Slow-path call
127 int size = slow_call_type->domain()->cnt();
128 CallNode *call = leaf_name
129 ? (CallNode*)new (C, size) CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM )
130 : (CallNode*)new (C, size) CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), oldcall->jvms()->bci(), TypeRawPtr::BOTTOM );
132 // Slow path call has no side-effects, uses few values
133 copy_predefined_input_for_runtime_call(slow_path, oldcall, call );
134 if (parm0 != NULL) call->init_req(TypeFunc::Parms+0, parm0);
135 if (parm1 != NULL) call->init_req(TypeFunc::Parms+1, parm1);
136 copy_call_debug_info(oldcall, call);
137 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
138 _igvn.hash_delete(oldcall);
139 _igvn.subsume_node(oldcall, call);
140 transform_later(call);
142 return call;
143 }
145 void PhaseMacroExpand::extract_call_projections(CallNode *call) {
146 _fallthroughproj = NULL;
147 _fallthroughcatchproj = NULL;
148 _ioproj_fallthrough = NULL;
149 _ioproj_catchall = NULL;
150 _catchallcatchproj = NULL;
151 _memproj_fallthrough = NULL;
152 _memproj_catchall = NULL;
153 _resproj = NULL;
154 for (DUIterator_Fast imax, i = call->fast_outs(imax); i < imax; i++) {
155 ProjNode *pn = call->fast_out(i)->as_Proj();
156 switch (pn->_con) {
157 case TypeFunc::Control:
158 {
159 // For Control (fallthrough) and I_O (catch_all_index) we have CatchProj -> Catch -> Proj
160 _fallthroughproj = pn;
161 DUIterator_Fast jmax, j = pn->fast_outs(jmax);
162 const Node *cn = pn->fast_out(j);
163 if (cn->is_Catch()) {
164 ProjNode *cpn = NULL;
165 for (DUIterator_Fast kmax, k = cn->fast_outs(kmax); k < kmax; k++) {
166 cpn = cn->fast_out(k)->as_Proj();
167 assert(cpn->is_CatchProj(), "must be a CatchProjNode");
168 if (cpn->_con == CatchProjNode::fall_through_index)
169 _fallthroughcatchproj = cpn;
170 else {
171 assert(cpn->_con == CatchProjNode::catch_all_index, "must be correct index.");
172 _catchallcatchproj = cpn;
173 }
174 }
175 }
176 break;
177 }
178 case TypeFunc::I_O:
179 if (pn->_is_io_use)
180 _ioproj_catchall = pn;
181 else
182 _ioproj_fallthrough = pn;
183 break;
184 case TypeFunc::Memory:
185 if (pn->_is_io_use)
186 _memproj_catchall = pn;
187 else
188 _memproj_fallthrough = pn;
189 break;
190 case TypeFunc::Parms:
191 _resproj = pn;
192 break;
193 default:
194 assert(false, "unexpected projection from allocation node.");
195 }
196 }
198 }
200 // Eliminate a card mark sequence. p2x is a ConvP2XNode
201 void PhaseMacroExpand::eliminate_card_mark(Node* p2x) {
202 assert(p2x->Opcode() == Op_CastP2X, "ConvP2XNode required");
203 if (!UseG1GC) {
204 // vanilla/CMS post barrier
205 Node *shift = p2x->unique_out();
206 Node *addp = shift->unique_out();
207 for (DUIterator_Last jmin, j = addp->last_outs(jmin); j >= jmin; --j) {
208 Node *st = addp->last_out(j);
209 assert(st->is_Store(), "store required");
210 _igvn.replace_node(st, st->in(MemNode::Memory));
211 }
212 } else {
213 // G1 pre/post barriers
214 assert(p2x->outcnt() == 2, "expects 2 users: Xor and URShift nodes");
215 // It could be only one user, URShift node, in Object.clone() instrinsic
216 // but the new allocation is passed to arraycopy stub and it could not
217 // be scalar replaced. So we don't check the case.
219 // Remove G1 post barrier.
221 // Search for CastP2X->Xor->URShift->Cmp path which
222 // checks if the store done to a different from the value's region.
223 // And replace Cmp with #0 (false) to collapse G1 post barrier.
224 Node* xorx = NULL;
225 for (DUIterator_Fast imax, i = p2x->fast_outs(imax); i < imax; i++) {
226 Node* u = p2x->fast_out(i);
227 if (u->Opcode() == Op_XorX) {
228 xorx = u;
229 break;
230 }
231 }
232 assert(xorx != NULL, "missing G1 post barrier");
233 Node* shift = xorx->unique_out();
234 Node* cmpx = shift->unique_out();
235 assert(cmpx->is_Cmp() && cmpx->unique_out()->is_Bool() &&
236 cmpx->unique_out()->as_Bool()->_test._test == BoolTest::ne,
237 "missing region check in G1 post barrier");
238 _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ));
240 // Remove G1 pre barrier.
242 // Search "if (marking != 0)" check and set it to "false".
243 Node* this_region = p2x->in(0);
244 assert(this_region != NULL, "");
245 // There is no G1 pre barrier if previous stored value is NULL
246 // (for example, after initialization).
247 if (this_region->is_Region() && this_region->req() == 3) {
248 int ind = 1;
249 if (!this_region->in(ind)->is_IfFalse()) {
250 ind = 2;
251 }
252 if (this_region->in(ind)->is_IfFalse()) {
253 Node* bol = this_region->in(ind)->in(0)->in(1);
254 assert(bol->is_Bool(), "");
255 cmpx = bol->in(1);
256 if (bol->as_Bool()->_test._test == BoolTest::ne &&
257 cmpx->is_Cmp() && cmpx->in(2) == intcon(0) &&
258 cmpx->in(1)->is_Load()) {
259 Node* adr = cmpx->in(1)->as_Load()->in(MemNode::Address);
260 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() +
261 PtrQueue::byte_offset_of_active());
262 if (adr->is_AddP() && adr->in(AddPNode::Base) == top() &&
263 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
264 adr->in(AddPNode::Offset) == MakeConX(marking_offset)) {
265 _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ));
266 }
267 }
268 }
269 }
270 // Now CastP2X can be removed since it is used only on dead path
271 // which currently still alive until igvn optimize it.
272 assert(p2x->unique_out()->Opcode() == Op_URShiftX, "");
273 _igvn.replace_node(p2x, top());
274 }
275 }
277 // Search for a memory operation for the specified memory slice.
278 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) {
279 Node *orig_mem = mem;
280 Node *alloc_mem = alloc->in(TypeFunc::Memory);
281 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr();
282 while (true) {
283 if (mem == alloc_mem || mem == start_mem ) {
284 return mem; // hit one of our sentinels
285 } else if (mem->is_MergeMem()) {
286 mem = mem->as_MergeMem()->memory_at(alias_idx);
287 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) {
288 Node *in = mem->in(0);
289 // we can safely skip over safepoints, calls, locks and membars because we
290 // already know that the object is safe to eliminate.
291 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) {
292 return in;
293 } else if (in->is_Call()) {
294 CallNode *call = in->as_Call();
295 if (!call->may_modify(tinst, phase)) {
296 mem = call->in(TypeFunc::Memory);
297 }
298 mem = in->in(TypeFunc::Memory);
299 } else if (in->is_MemBar()) {
300 mem = in->in(TypeFunc::Memory);
301 } else {
302 assert(false, "unexpected projection");
303 }
304 } else if (mem->is_Store()) {
305 const TypePtr* atype = mem->as_Store()->adr_type();
306 int adr_idx = Compile::current()->get_alias_index(atype);
307 if (adr_idx == alias_idx) {
308 assert(atype->isa_oopptr(), "address type must be oopptr");
309 int adr_offset = atype->offset();
310 uint adr_iid = atype->is_oopptr()->instance_id();
311 // Array elements references have the same alias_idx
312 // but different offset and different instance_id.
313 if (adr_offset == offset && adr_iid == alloc->_idx)
314 return mem;
315 } else {
316 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw");
317 }
318 mem = mem->in(MemNode::Memory);
319 } else if (mem->is_ClearArray()) {
320 if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) {
321 // Can not bypass initialization of the instance
322 // we are looking.
323 debug_only(intptr_t offset;)
324 assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity");
325 InitializeNode* init = alloc->as_Allocate()->initialization();
326 // We are looking for stored value, return Initialize node
327 // or memory edge from Allocate node.
328 if (init != NULL)
329 return init;
330 else
331 return alloc->in(TypeFunc::Memory); // It will produce zero value (see callers).
332 }
333 // Otherwise skip it (the call updated 'mem' value).
334 } else if (mem->Opcode() == Op_SCMemProj) {
335 assert(mem->in(0)->is_LoadStore(), "sanity");
336 const TypePtr* atype = mem->in(0)->in(MemNode::Address)->bottom_type()->is_ptr();
337 int adr_idx = Compile::current()->get_alias_index(atype);
338 if (adr_idx == alias_idx) {
339 assert(false, "Object is not scalar replaceable if a LoadStore node access its field");
340 return NULL;
341 }
342 mem = mem->in(0)->in(MemNode::Memory);
343 } else {
344 return mem;
345 }
346 assert(mem != orig_mem, "dead memory loop");
347 }
348 }
350 //
351 // Given a Memory Phi, compute a value Phi containing the values from stores
352 // on the input paths.
353 // Note: this function is recursive, its depth is limied by the "level" argument
354 // Returns the computed Phi, or NULL if it cannot compute it.
355 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) {
356 assert(mem->is_Phi(), "sanity");
357 int alias_idx = C->get_alias_index(adr_t);
358 int offset = adr_t->offset();
359 int instance_id = adr_t->instance_id();
361 // Check if an appropriate value phi already exists.
362 Node* region = mem->in(0);
363 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) {
364 Node* phi = region->fast_out(k);
365 if (phi->is_Phi() && phi != mem &&
366 phi->as_Phi()->is_same_inst_field(phi_type, instance_id, alias_idx, offset)) {
367 return phi;
368 }
369 }
370 // Check if an appropriate new value phi already exists.
371 Node* new_phi = NULL;
372 uint size = value_phis->size();
373 for (uint i=0; i < size; i++) {
374 if ( mem->_idx == value_phis->index_at(i) ) {
375 return value_phis->node_at(i);
376 }
377 }
379 if (level <= 0) {
380 return NULL; // Give up: phi tree too deep
381 }
382 Node *start_mem = C->start()->proj_out(TypeFunc::Memory);
383 Node *alloc_mem = alloc->in(TypeFunc::Memory);
385 uint length = mem->req();
386 GrowableArray <Node *> values(length, length, NULL);
388 // create a new Phi for the value
389 PhiNode *phi = new (C, length) PhiNode(mem->in(0), phi_type, NULL, instance_id, alias_idx, offset);
390 transform_later(phi);
391 value_phis->push(phi, mem->_idx);
393 for (uint j = 1; j < length; j++) {
394 Node *in = mem->in(j);
395 if (in == NULL || in->is_top()) {
396 values.at_put(j, in);
397 } else {
398 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn);
399 if (val == start_mem || val == alloc_mem) {
400 // hit a sentinel, return appropriate 0 value
401 values.at_put(j, _igvn.zerocon(ft));
402 continue;
403 }
404 if (val->is_Initialize()) {
405 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
406 }
407 if (val == NULL) {
408 return NULL; // can't find a value on this path
409 }
410 if (val == mem) {
411 values.at_put(j, mem);
412 } else if (val->is_Store()) {
413 values.at_put(j, val->in(MemNode::ValueIn));
414 } else if(val->is_Proj() && val->in(0) == alloc) {
415 values.at_put(j, _igvn.zerocon(ft));
416 } else if (val->is_Phi()) {
417 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1);
418 if (val == NULL) {
419 return NULL;
420 }
421 values.at_put(j, val);
422 } else if (val->Opcode() == Op_SCMemProj) {
423 assert(val->in(0)->is_LoadStore(), "sanity");
424 assert(false, "Object is not scalar replaceable if a LoadStore node access its field");
425 return NULL;
426 } else {
427 #ifdef ASSERT
428 val->dump();
429 assert(false, "unknown node on this path");
430 #endif
431 return NULL; // unknown node on this path
432 }
433 }
434 }
435 // Set Phi's inputs
436 for (uint j = 1; j < length; j++) {
437 if (values.at(j) == mem) {
438 phi->init_req(j, phi);
439 } else {
440 phi->init_req(j, values.at(j));
441 }
442 }
443 return phi;
444 }
446 // Search the last value stored into the object's field.
447 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, Node *alloc) {
448 assert(adr_t->is_known_instance_field(), "instance required");
449 int instance_id = adr_t->instance_id();
450 assert((uint)instance_id == alloc->_idx, "wrong allocation");
452 int alias_idx = C->get_alias_index(adr_t);
453 int offset = adr_t->offset();
454 Node *start_mem = C->start()->proj_out(TypeFunc::Memory);
455 Node *alloc_ctrl = alloc->in(TypeFunc::Control);
456 Node *alloc_mem = alloc->in(TypeFunc::Memory);
457 Arena *a = Thread::current()->resource_area();
458 VectorSet visited(a);
461 bool done = sfpt_mem == alloc_mem;
462 Node *mem = sfpt_mem;
463 while (!done) {
464 if (visited.test_set(mem->_idx)) {
465 return NULL; // found a loop, give up
466 }
467 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn);
468 if (mem == start_mem || mem == alloc_mem) {
469 done = true; // hit a sentinel, return appropriate 0 value
470 } else if (mem->is_Initialize()) {
471 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
472 if (mem == NULL) {
473 done = true; // Something go wrong.
474 } else if (mem->is_Store()) {
475 const TypePtr* atype = mem->as_Store()->adr_type();
476 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice");
477 done = true;
478 }
479 } else if (mem->is_Store()) {
480 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr();
481 assert(atype != NULL, "address type must be oopptr");
482 assert(C->get_alias_index(atype) == alias_idx &&
483 atype->is_known_instance_field() && atype->offset() == offset &&
484 atype->instance_id() == instance_id, "store is correct memory slice");
485 done = true;
486 } else if (mem->is_Phi()) {
487 // try to find a phi's unique input
488 Node *unique_input = NULL;
489 Node *top = C->top();
490 for (uint i = 1; i < mem->req(); i++) {
491 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn);
492 if (n == NULL || n == top || n == mem) {
493 continue;
494 } else if (unique_input == NULL) {
495 unique_input = n;
496 } else if (unique_input != n) {
497 unique_input = top;
498 break;
499 }
500 }
501 if (unique_input != NULL && unique_input != top) {
502 mem = unique_input;
503 } else {
504 done = true;
505 }
506 } else {
507 assert(false, "unexpected node");
508 }
509 }
510 if (mem != NULL) {
511 if (mem == start_mem || mem == alloc_mem) {
512 // hit a sentinel, return appropriate 0 value
513 return _igvn.zerocon(ft);
514 } else if (mem->is_Store()) {
515 return mem->in(MemNode::ValueIn);
516 } else if (mem->is_Phi()) {
517 // attempt to produce a Phi reflecting the values on the input paths of the Phi
518 Node_Stack value_phis(a, 8);
519 Node * phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit);
520 if (phi != NULL) {
521 return phi;
522 } else {
523 // Kill all new Phis
524 while(value_phis.is_nonempty()) {
525 Node* n = value_phis.node();
526 _igvn.hash_delete(n);
527 _igvn.subsume_node(n, C->top());
528 value_phis.pop();
529 }
530 }
531 }
532 }
533 // Something go wrong.
534 return NULL;
535 }
537 // Check the possibility of scalar replacement.
538 bool PhaseMacroExpand::can_eliminate_allocation(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
539 // Scan the uses of the allocation to check for anything that would
540 // prevent us from eliminating it.
541 NOT_PRODUCT( const char* fail_eliminate = NULL; )
542 DEBUG_ONLY( Node* disq_node = NULL; )
543 bool can_eliminate = true;
545 Node* res = alloc->result_cast();
546 const TypeOopPtr* res_type = NULL;
547 if (res == NULL) {
548 // All users were eliminated.
549 } else if (!res->is_CheckCastPP()) {
550 alloc->_is_scalar_replaceable = false; // don't try again
551 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";)
552 can_eliminate = false;
553 } else {
554 res_type = _igvn.type(res)->isa_oopptr();
555 if (res_type == NULL) {
556 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";)
557 can_eliminate = false;
558 } else if (res_type->isa_aryptr()) {
559 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1);
560 if (length < 0) {
561 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";)
562 can_eliminate = false;
563 }
564 }
565 }
567 if (can_eliminate && res != NULL) {
568 for (DUIterator_Fast jmax, j = res->fast_outs(jmax);
569 j < jmax && can_eliminate; j++) {
570 Node* use = res->fast_out(j);
572 if (use->is_AddP()) {
573 const TypePtr* addp_type = _igvn.type(use)->is_ptr();
574 int offset = addp_type->offset();
576 if (offset == Type::OffsetTop || offset == Type::OffsetBot) {
577 NOT_PRODUCT(fail_eliminate = "Undefined field referrence";)
578 can_eliminate = false;
579 break;
580 }
581 for (DUIterator_Fast kmax, k = use->fast_outs(kmax);
582 k < kmax && can_eliminate; k++) {
583 Node* n = use->fast_out(k);
584 if (!n->is_Store() && n->Opcode() != Op_CastP2X) {
585 DEBUG_ONLY(disq_node = n;)
586 if (n->is_Load() || n->is_LoadStore()) {
587 NOT_PRODUCT(fail_eliminate = "Field load";)
588 } else {
589 NOT_PRODUCT(fail_eliminate = "Not store field referrence";)
590 }
591 can_eliminate = false;
592 }
593 }
594 } else if (use->is_SafePoint()) {
595 SafePointNode* sfpt = use->as_SafePoint();
596 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) {
597 // Object is passed as argument.
598 DEBUG_ONLY(disq_node = use;)
599 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";)
600 can_eliminate = false;
601 }
602 Node* sfptMem = sfpt->memory();
603 if (sfptMem == NULL || sfptMem->is_top()) {
604 DEBUG_ONLY(disq_node = use;)
605 NOT_PRODUCT(fail_eliminate = "NULL or TOP memory";)
606 can_eliminate = false;
607 } else {
608 safepoints.append_if_missing(sfpt);
609 }
610 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
611 if (use->is_Phi()) {
612 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
613 NOT_PRODUCT(fail_eliminate = "Object is return value";)
614 } else {
615 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
616 }
617 DEBUG_ONLY(disq_node = use;)
618 } else {
619 if (use->Opcode() == Op_Return) {
620 NOT_PRODUCT(fail_eliminate = "Object is return value";)
621 }else {
622 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
623 }
624 DEBUG_ONLY(disq_node = use;)
625 }
626 can_eliminate = false;
627 }
628 }
629 }
631 #ifndef PRODUCT
632 if (PrintEliminateAllocations) {
633 if (can_eliminate) {
634 tty->print("Scalar ");
635 if (res == NULL)
636 alloc->dump();
637 else
638 res->dump();
639 } else {
640 tty->print("NotScalar (%s)", fail_eliminate);
641 if (res == NULL)
642 alloc->dump();
643 else
644 res->dump();
645 #ifdef ASSERT
646 if (disq_node != NULL) {
647 tty->print(" >>>> ");
648 disq_node->dump();
649 }
650 #endif /*ASSERT*/
651 }
652 }
653 #endif
654 return can_eliminate;
655 }
657 // Do scalar replacement.
658 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
659 GrowableArray <SafePointNode *> safepoints_done;
661 ciKlass* klass = NULL;
662 ciInstanceKlass* iklass = NULL;
663 int nfields = 0;
664 int array_base;
665 int element_size;
666 BasicType basic_elem_type;
667 ciType* elem_type;
669 Node* res = alloc->result_cast();
670 const TypeOopPtr* res_type = NULL;
671 if (res != NULL) { // Could be NULL when there are no users
672 res_type = _igvn.type(res)->isa_oopptr();
673 }
675 if (res != NULL) {
676 klass = res_type->klass();
677 if (res_type->isa_instptr()) {
678 // find the fields of the class which will be needed for safepoint debug information
679 assert(klass->is_instance_klass(), "must be an instance klass.");
680 iklass = klass->as_instance_klass();
681 nfields = iklass->nof_nonstatic_fields();
682 } else {
683 // find the array's elements which will be needed for safepoint debug information
684 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
685 assert(klass->is_array_klass() && nfields >= 0, "must be an array klass.");
686 elem_type = klass->as_array_klass()->element_type();
687 basic_elem_type = elem_type->basic_type();
688 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
689 element_size = type2aelembytes(basic_elem_type);
690 }
691 }
692 //
693 // Process the safepoint uses
694 //
695 while (safepoints.length() > 0) {
696 SafePointNode* sfpt = safepoints.pop();
697 Node* mem = sfpt->memory();
698 uint first_ind = sfpt->req();
699 SafePointScalarObjectNode* sobj = new (C, 1) SafePointScalarObjectNode(res_type,
700 #ifdef ASSERT
701 alloc,
702 #endif
703 first_ind, nfields);
704 sobj->init_req(0, sfpt->in(TypeFunc::Control));
705 transform_later(sobj);
707 // Scan object's fields adding an input to the safepoint for each field.
708 for (int j = 0; j < nfields; j++) {
709 intptr_t offset;
710 ciField* field = NULL;
711 if (iklass != NULL) {
712 field = iklass->nonstatic_field_at(j);
713 offset = field->offset();
714 elem_type = field->type();
715 basic_elem_type = field->layout_type();
716 } else {
717 offset = array_base + j * (intptr_t)element_size;
718 }
720 const Type *field_type;
721 // The next code is taken from Parse::do_get_xxx().
722 if (basic_elem_type == T_OBJECT || basic_elem_type == T_ARRAY) {
723 if (!elem_type->is_loaded()) {
724 field_type = TypeInstPtr::BOTTOM;
725 } else if (field != NULL && field->is_constant()) {
726 // This can happen if the constant oop is non-perm.
727 ciObject* con = field->constant_value().as_object();
728 // Do not "join" in the previous type; it doesn't add value,
729 // and may yield a vacuous result if the field is of interface type.
730 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr();
731 assert(field_type != NULL, "field singleton type must be consistent");
732 } else {
733 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
734 }
735 if (UseCompressedOops) {
736 field_type = field_type->make_narrowoop();
737 basic_elem_type = T_NARROWOOP;
738 }
739 } else {
740 field_type = Type::get_const_basic_type(basic_elem_type);
741 }
743 const TypeOopPtr *field_addr_type = res_type->add_offset(offset)->isa_oopptr();
745 Node *field_val = value_from_mem(mem, basic_elem_type, field_type, field_addr_type, alloc);
746 if (field_val == NULL) {
747 // we weren't able to find a value for this field,
748 // give up on eliminating this allocation
749 alloc->_is_scalar_replaceable = false; // don't try again
750 // remove any extra entries we added to the safepoint
751 uint last = sfpt->req() - 1;
752 for (int k = 0; k < j; k++) {
753 sfpt->del_req(last--);
754 }
755 // rollback processed safepoints
756 while (safepoints_done.length() > 0) {
757 SafePointNode* sfpt_done = safepoints_done.pop();
758 // remove any extra entries we added to the safepoint
759 last = sfpt_done->req() - 1;
760 for (int k = 0; k < nfields; k++) {
761 sfpt_done->del_req(last--);
762 }
763 JVMState *jvms = sfpt_done->jvms();
764 jvms->set_endoff(sfpt_done->req());
765 // Now make a pass over the debug information replacing any references
766 // to SafePointScalarObjectNode with the allocated object.
767 int start = jvms->debug_start();
768 int end = jvms->debug_end();
769 for (int i = start; i < end; i++) {
770 if (sfpt_done->in(i)->is_SafePointScalarObject()) {
771 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject();
772 if (scobj->first_index() == sfpt_done->req() &&
773 scobj->n_fields() == (uint)nfields) {
774 assert(scobj->alloc() == alloc, "sanity");
775 sfpt_done->set_req(i, res);
776 }
777 }
778 }
779 }
780 #ifndef PRODUCT
781 if (PrintEliminateAllocations) {
782 if (field != NULL) {
783 tty->print("=== At SafePoint node %d can't find value of Field: ",
784 sfpt->_idx);
785 field->print();
786 int field_idx = C->get_alias_index(field_addr_type);
787 tty->print(" (alias_idx=%d)", field_idx);
788 } else { // Array's element
789 tty->print("=== At SafePoint node %d can't find value of array element [%d]",
790 sfpt->_idx, j);
791 }
792 tty->print(", which prevents elimination of: ");
793 if (res == NULL)
794 alloc->dump();
795 else
796 res->dump();
797 }
798 #endif
799 return false;
800 }
801 if (UseCompressedOops && field_type->isa_narrowoop()) {
802 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
803 // to be able scalar replace the allocation.
804 if (field_val->is_EncodeP()) {
805 field_val = field_val->in(1);
806 } else {
807 field_val = transform_later(new (C, 2) DecodeNNode(field_val, field_val->bottom_type()->make_ptr()));
808 }
809 }
810 sfpt->add_req(field_val);
811 }
812 JVMState *jvms = sfpt->jvms();
813 jvms->set_endoff(sfpt->req());
814 // Now make a pass over the debug information replacing any references
815 // to the allocated object with "sobj"
816 int start = jvms->debug_start();
817 int end = jvms->debug_end();
818 for (int i = start; i < end; i++) {
819 if (sfpt->in(i) == res) {
820 sfpt->set_req(i, sobj);
821 }
822 }
823 safepoints_done.append_if_missing(sfpt); // keep it for rollback
824 }
825 return true;
826 }
828 // Process users of eliminated allocation.
829 void PhaseMacroExpand::process_users_of_allocation(AllocateNode *alloc) {
830 Node* res = alloc->result_cast();
831 if (res != NULL) {
832 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
833 Node *use = res->last_out(j);
834 uint oc1 = res->outcnt();
836 if (use->is_AddP()) {
837 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) {
838 Node *n = use->last_out(k);
839 uint oc2 = use->outcnt();
840 if (n->is_Store()) {
841 #ifdef ASSERT
842 // Verify that there is no dependent MemBarVolatile nodes,
843 // they should be removed during IGVN, see MemBarNode::Ideal().
844 for (DUIterator_Fast pmax, p = n->fast_outs(pmax);
845 p < pmax; p++) {
846 Node* mb = n->fast_out(p);
847 assert(mb->is_Initialize() || !mb->is_MemBar() ||
848 mb->req() <= MemBarNode::Precedent ||
849 mb->in(MemBarNode::Precedent) != n,
850 "MemBarVolatile should be eliminated for non-escaping object");
851 }
852 #endif
853 _igvn.replace_node(n, n->in(MemNode::Memory));
854 } else {
855 eliminate_card_mark(n);
856 }
857 k -= (oc2 - use->outcnt());
858 }
859 } else {
860 eliminate_card_mark(use);
861 }
862 j -= (oc1 - res->outcnt());
863 }
864 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
865 _igvn.remove_dead_node(res);
866 }
868 //
869 // Process other users of allocation's projections
870 //
871 if (_resproj != NULL && _resproj->outcnt() != 0) {
872 for (DUIterator_Last jmin, j = _resproj->last_outs(jmin); j >= jmin; ) {
873 Node *use = _resproj->last_out(j);
874 uint oc1 = _resproj->outcnt();
875 if (use->is_Initialize()) {
876 // Eliminate Initialize node.
877 InitializeNode *init = use->as_Initialize();
878 assert(init->outcnt() <= 2, "only a control and memory projection expected");
879 Node *ctrl_proj = init->proj_out(TypeFunc::Control);
880 if (ctrl_proj != NULL) {
881 assert(init->in(TypeFunc::Control) == _fallthroughcatchproj, "allocation control projection");
882 _igvn.replace_node(ctrl_proj, _fallthroughcatchproj);
883 }
884 Node *mem_proj = init->proj_out(TypeFunc::Memory);
885 if (mem_proj != NULL) {
886 Node *mem = init->in(TypeFunc::Memory);
887 #ifdef ASSERT
888 if (mem->is_MergeMem()) {
889 assert(mem->in(TypeFunc::Memory) == _memproj_fallthrough, "allocation memory projection");
890 } else {
891 assert(mem == _memproj_fallthrough, "allocation memory projection");
892 }
893 #endif
894 _igvn.replace_node(mem_proj, mem);
895 }
896 } else if (use->is_AddP()) {
897 // raw memory addresses used only by the initialization
898 _igvn.replace_node(use, C->top());
899 } else {
900 assert(false, "only Initialize or AddP expected");
901 }
902 j -= (oc1 - _resproj->outcnt());
903 }
904 }
905 if (_fallthroughcatchproj != NULL) {
906 _igvn.replace_node(_fallthroughcatchproj, alloc->in(TypeFunc::Control));
907 }
908 if (_memproj_fallthrough != NULL) {
909 _igvn.replace_node(_memproj_fallthrough, alloc->in(TypeFunc::Memory));
910 }
911 if (_memproj_catchall != NULL) {
912 _igvn.replace_node(_memproj_catchall, C->top());
913 }
914 if (_ioproj_fallthrough != NULL) {
915 _igvn.replace_node(_ioproj_fallthrough, alloc->in(TypeFunc::I_O));
916 }
917 if (_ioproj_catchall != NULL) {
918 _igvn.replace_node(_ioproj_catchall, C->top());
919 }
920 if (_catchallcatchproj != NULL) {
921 _igvn.replace_node(_catchallcatchproj, C->top());
922 }
923 }
925 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) {
927 if (!EliminateAllocations || !alloc->_is_scalar_replaceable) {
928 return false;
929 }
931 extract_call_projections(alloc);
933 GrowableArray <SafePointNode *> safepoints;
934 if (!can_eliminate_allocation(alloc, safepoints)) {
935 return false;
936 }
938 if (!scalar_replacement(alloc, safepoints)) {
939 return false;
940 }
942 CompileLog* log = C->log();
943 if (log != NULL) {
944 Node* klass = alloc->in(AllocateNode::KlassNode);
945 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr();
946 log->head("eliminate_allocation type='%d'",
947 log->identify(tklass->klass()));
948 JVMState* p = alloc->jvms();
949 while (p != NULL) {
950 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
951 p = p->caller();
952 }
953 log->tail("eliminate_allocation");
954 }
956 process_users_of_allocation(alloc);
958 #ifndef PRODUCT
959 if (PrintEliminateAllocations) {
960 if (alloc->is_AllocateArray())
961 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
962 else
963 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
964 }
965 #endif
967 return true;
968 }
971 //---------------------------set_eden_pointers-------------------------
972 void PhaseMacroExpand::set_eden_pointers(Node* &eden_top_adr, Node* &eden_end_adr) {
973 if (UseTLAB) { // Private allocation: load from TLS
974 Node* thread = transform_later(new (C, 1) ThreadLocalNode());
975 int tlab_top_offset = in_bytes(JavaThread::tlab_top_offset());
976 int tlab_end_offset = in_bytes(JavaThread::tlab_end_offset());
977 eden_top_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_top_offset);
978 eden_end_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_end_offset);
979 } else { // Shared allocation: load from globals
980 CollectedHeap* ch = Universe::heap();
981 address top_adr = (address)ch->top_addr();
982 address end_adr = (address)ch->end_addr();
983 eden_top_adr = makecon(TypeRawPtr::make(top_adr));
984 eden_end_adr = basic_plus_adr(eden_top_adr, end_adr - top_adr);
985 }
986 }
989 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) {
990 Node* adr = basic_plus_adr(base, offset);
991 const TypePtr* adr_type = adr->bottom_type()->is_ptr();
992 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt);
993 transform_later(value);
994 return value;
995 }
998 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) {
999 Node* adr = basic_plus_adr(base, offset);
1000 mem = StoreNode::make(_igvn, ctl, mem, adr, NULL, value, bt);
1001 transform_later(mem);
1002 return mem;
1003 }
1005 //=============================================================================
1006 //
1007 // A L L O C A T I O N
1008 //
1009 // Allocation attempts to be fast in the case of frequent small objects.
1010 // It breaks down like this:
1011 //
1012 // 1) Size in doublewords is computed. This is a constant for objects and
1013 // variable for most arrays. Doubleword units are used to avoid size
1014 // overflow of huge doubleword arrays. We need doublewords in the end for
1015 // rounding.
1016 //
1017 // 2) Size is checked for being 'too large'. Too-large allocations will go
1018 // the slow path into the VM. The slow path can throw any required
1019 // exceptions, and does all the special checks for very large arrays. The
1020 // size test can constant-fold away for objects. For objects with
1021 // finalizers it constant-folds the otherway: you always go slow with
1022 // finalizers.
1023 //
1024 // 3) If NOT using TLABs, this is the contended loop-back point.
1025 // Load-Locked the heap top. If using TLABs normal-load the heap top.
1026 //
1027 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route.
1028 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish
1029 // "size*8" we always enter the VM, where "largish" is a constant picked small
1030 // enough that there's always space between the eden max and 4Gig (old space is
1031 // there so it's quite large) and large enough that the cost of entering the VM
1032 // is dwarfed by the cost to initialize the space.
1033 //
1034 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back
1035 // down. If contended, repeat at step 3. If using TLABs normal-store
1036 // adjusted heap top back down; there is no contention.
1037 //
1038 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark
1039 // fields.
1040 //
1041 // 7) Merge with the slow-path; cast the raw memory pointer to the correct
1042 // oop flavor.
1043 //
1044 //=============================================================================
1045 // FastAllocateSizeLimit value is in DOUBLEWORDS.
1046 // Allocations bigger than this always go the slow route.
1047 // This value must be small enough that allocation attempts that need to
1048 // trigger exceptions go the slow route. Also, it must be small enough so
1049 // that heap_top + size_in_bytes does not wrap around the 4Gig limit.
1050 //=============================================================================j//
1051 // %%% Here is an old comment from parseHelper.cpp; is it outdated?
1052 // The allocator will coalesce int->oop copies away. See comment in
1053 // coalesce.cpp about how this works. It depends critically on the exact
1054 // code shape produced here, so if you are changing this code shape
1055 // make sure the GC info for the heap-top is correct in and around the
1056 // slow-path call.
1057 //
1059 void PhaseMacroExpand::expand_allocate_common(
1060 AllocateNode* alloc, // allocation node to be expanded
1061 Node* length, // array length for an array allocation
1062 const TypeFunc* slow_call_type, // Type of slow call
1063 address slow_call_address // Address of slow call
1064 )
1065 {
1067 Node* ctrl = alloc->in(TypeFunc::Control);
1068 Node* mem = alloc->in(TypeFunc::Memory);
1069 Node* i_o = alloc->in(TypeFunc::I_O);
1070 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize);
1071 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1072 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest);
1074 assert(ctrl != NULL, "must have control");
1075 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results.
1076 // they will not be used if "always_slow" is set
1077 enum { slow_result_path = 1, fast_result_path = 2 };
1078 Node *result_region;
1079 Node *result_phi_rawmem;
1080 Node *result_phi_rawoop;
1081 Node *result_phi_i_o;
1083 // The initial slow comparison is a size check, the comparison
1084 // we want to do is a BoolTest::gt
1085 bool always_slow = false;
1086 int tv = _igvn.find_int_con(initial_slow_test, -1);
1087 if (tv >= 0) {
1088 always_slow = (tv == 1);
1089 initial_slow_test = NULL;
1090 } else {
1091 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn);
1092 }
1094 if (C->env()->dtrace_alloc_probes() ||
1095 !UseTLAB && (!Universe::heap()->supports_inline_contig_alloc() ||
1096 (UseConcMarkSweepGC && CMSIncrementalMode))) {
1097 // Force slow-path allocation
1098 always_slow = true;
1099 initial_slow_test = NULL;
1100 }
1103 enum { too_big_or_final_path = 1, need_gc_path = 2 };
1104 Node *slow_region = NULL;
1105 Node *toobig_false = ctrl;
1107 assert (initial_slow_test == NULL || !always_slow, "arguments must be consistent");
1108 // generate the initial test if necessary
1109 if (initial_slow_test != NULL ) {
1110 slow_region = new (C, 3) RegionNode(3);
1112 // Now make the initial failure test. Usually a too-big test but
1113 // might be a TRUE for finalizers or a fancy class check for
1114 // newInstance0.
1115 IfNode *toobig_iff = new (C, 2) IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
1116 transform_later(toobig_iff);
1117 // Plug the failing-too-big test into the slow-path region
1118 Node *toobig_true = new (C, 1) IfTrueNode( toobig_iff );
1119 transform_later(toobig_true);
1120 slow_region ->init_req( too_big_or_final_path, toobig_true );
1121 toobig_false = new (C, 1) IfFalseNode( toobig_iff );
1122 transform_later(toobig_false);
1123 } else { // No initial test, just fall into next case
1124 toobig_false = ctrl;
1125 debug_only(slow_region = NodeSentinel);
1126 }
1128 Node *slow_mem = mem; // save the current memory state for slow path
1129 // generate the fast allocation code unless we know that the initial test will always go slow
1130 if (!always_slow) {
1131 // Fast path modifies only raw memory.
1132 if (mem->is_MergeMem()) {
1133 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1134 }
1136 Node* eden_top_adr;
1137 Node* eden_end_adr;
1139 set_eden_pointers(eden_top_adr, eden_end_adr);
1141 // Load Eden::end. Loop invariant and hoisted.
1142 //
1143 // Note: We set the control input on "eden_end" and "old_eden_top" when using
1144 // a TLAB to work around a bug where these values were being moved across
1145 // a safepoint. These are not oops, so they cannot be include in the oop
1146 // map, but the can be changed by a GC. The proper way to fix this would
1147 // be to set the raw memory state when generating a SafepointNode. However
1148 // this will require extensive changes to the loop optimization in order to
1149 // prevent a degradation of the optimization.
1150 // See comment in memnode.hpp, around line 227 in class LoadPNode.
1151 Node *eden_end = make_load(ctrl, mem, eden_end_adr, 0, TypeRawPtr::BOTTOM, T_ADDRESS);
1153 // allocate the Region and Phi nodes for the result
1154 result_region = new (C, 3) RegionNode(3);
1155 result_phi_rawmem = new (C, 3) PhiNode( result_region, Type::MEMORY, TypeRawPtr::BOTTOM );
1156 result_phi_rawoop = new (C, 3) PhiNode( result_region, TypeRawPtr::BOTTOM );
1157 result_phi_i_o = new (C, 3) PhiNode( result_region, Type::ABIO ); // I/O is used for Prefetch
1159 // We need a Region for the loop-back contended case.
1160 enum { fall_in_path = 1, contended_loopback_path = 2 };
1161 Node *contended_region;
1162 Node *contended_phi_rawmem;
1163 if( UseTLAB ) {
1164 contended_region = toobig_false;
1165 contended_phi_rawmem = mem;
1166 } else {
1167 contended_region = new (C, 3) RegionNode(3);
1168 contended_phi_rawmem = new (C, 3) PhiNode( contended_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1169 // Now handle the passing-too-big test. We fall into the contended
1170 // loop-back merge point.
1171 contended_region ->init_req( fall_in_path, toobig_false );
1172 contended_phi_rawmem->init_req( fall_in_path, mem );
1173 transform_later(contended_region);
1174 transform_later(contended_phi_rawmem);
1175 }
1177 // Load(-locked) the heap top.
1178 // See note above concerning the control input when using a TLAB
1179 Node *old_eden_top = UseTLAB
1180 ? new (C, 3) LoadPNode ( ctrl, contended_phi_rawmem, eden_top_adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM )
1181 : new (C, 3) LoadPLockedNode( contended_region, contended_phi_rawmem, eden_top_adr );
1183 transform_later(old_eden_top);
1184 // Add to heap top to get a new heap top
1185 Node *new_eden_top = new (C, 4) AddPNode( top(), old_eden_top, size_in_bytes );
1186 transform_later(new_eden_top);
1187 // Check for needing a GC; compare against heap end
1188 Node *needgc_cmp = new (C, 3) CmpPNode( new_eden_top, eden_end );
1189 transform_later(needgc_cmp);
1190 Node *needgc_bol = new (C, 2) BoolNode( needgc_cmp, BoolTest::ge );
1191 transform_later(needgc_bol);
1192 IfNode *needgc_iff = new (C, 2) IfNode(contended_region, needgc_bol, PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN );
1193 transform_later(needgc_iff);
1195 // Plug the failing-heap-space-need-gc test into the slow-path region
1196 Node *needgc_true = new (C, 1) IfTrueNode( needgc_iff );
1197 transform_later(needgc_true);
1198 if( initial_slow_test ) {
1199 slow_region ->init_req( need_gc_path, needgc_true );
1200 // This completes all paths into the slow merge point
1201 transform_later(slow_region);
1202 } else { // No initial slow path needed!
1203 // Just fall from the need-GC path straight into the VM call.
1204 slow_region = needgc_true;
1205 }
1206 // No need for a GC. Setup for the Store-Conditional
1207 Node *needgc_false = new (C, 1) IfFalseNode( needgc_iff );
1208 transform_later(needgc_false);
1210 // Grab regular I/O before optional prefetch may change it.
1211 // Slow-path does no I/O so just set it to the original I/O.
1212 result_phi_i_o->init_req( slow_result_path, i_o );
1214 i_o = prefetch_allocation(i_o, needgc_false, contended_phi_rawmem,
1215 old_eden_top, new_eden_top, length);
1217 // Store (-conditional) the modified eden top back down.
1218 // StorePConditional produces flags for a test PLUS a modified raw
1219 // memory state.
1220 Node *store_eden_top;
1221 Node *fast_oop_ctrl;
1222 if( UseTLAB ) {
1223 store_eden_top = new (C, 4) StorePNode( needgc_false, contended_phi_rawmem, eden_top_adr, TypeRawPtr::BOTTOM, new_eden_top );
1224 transform_later(store_eden_top);
1225 fast_oop_ctrl = needgc_false; // No contention, so this is the fast path
1226 } else {
1227 store_eden_top = new (C, 5) StorePConditionalNode( needgc_false, contended_phi_rawmem, eden_top_adr, new_eden_top, old_eden_top );
1228 transform_later(store_eden_top);
1229 Node *contention_check = new (C, 2) BoolNode( store_eden_top, BoolTest::ne );
1230 transform_later(contention_check);
1231 store_eden_top = new (C, 1) SCMemProjNode(store_eden_top);
1232 transform_later(store_eden_top);
1234 // If not using TLABs, check to see if there was contention.
1235 IfNode *contention_iff = new (C, 2) IfNode ( needgc_false, contention_check, PROB_MIN, COUNT_UNKNOWN );
1236 transform_later(contention_iff);
1237 Node *contention_true = new (C, 1) IfTrueNode( contention_iff );
1238 transform_later(contention_true);
1239 // If contention, loopback and try again.
1240 contended_region->init_req( contended_loopback_path, contention_true );
1241 contended_phi_rawmem->init_req( contended_loopback_path, store_eden_top );
1243 // Fast-path succeeded with no contention!
1244 Node *contention_false = new (C, 1) IfFalseNode( contention_iff );
1245 transform_later(contention_false);
1246 fast_oop_ctrl = contention_false;
1247 }
1249 // Rename successful fast-path variables to make meaning more obvious
1250 Node* fast_oop = old_eden_top;
1251 Node* fast_oop_rawmem = store_eden_top;
1252 fast_oop_rawmem = initialize_object(alloc,
1253 fast_oop_ctrl, fast_oop_rawmem, fast_oop,
1254 klass_node, length, size_in_bytes);
1256 if (C->env()->dtrace_extended_probes()) {
1257 // Slow-path call
1258 int size = TypeFunc::Parms + 2;
1259 CallLeafNode *call = new (C, size) CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(),
1260 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_object_alloc_base),
1261 "dtrace_object_alloc",
1262 TypeRawPtr::BOTTOM);
1264 // Get base of thread-local storage area
1265 Node* thread = new (C, 1) ThreadLocalNode();
1266 transform_later(thread);
1268 call->init_req(TypeFunc::Parms+0, thread);
1269 call->init_req(TypeFunc::Parms+1, fast_oop);
1270 call->init_req( TypeFunc::Control, fast_oop_ctrl );
1271 call->init_req( TypeFunc::I_O , top() ) ; // does no i/o
1272 call->init_req( TypeFunc::Memory , fast_oop_rawmem );
1273 call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) );
1274 call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) );
1275 transform_later(call);
1276 fast_oop_ctrl = new (C, 1) ProjNode(call,TypeFunc::Control);
1277 transform_later(fast_oop_ctrl);
1278 fast_oop_rawmem = new (C, 1) ProjNode(call,TypeFunc::Memory);
1279 transform_later(fast_oop_rawmem);
1280 }
1282 // Plug in the successful fast-path into the result merge point
1283 result_region ->init_req( fast_result_path, fast_oop_ctrl );
1284 result_phi_rawoop->init_req( fast_result_path, fast_oop );
1285 result_phi_i_o ->init_req( fast_result_path, i_o );
1286 result_phi_rawmem->init_req( fast_result_path, fast_oop_rawmem );
1287 } else {
1288 slow_region = ctrl;
1289 }
1291 // Generate slow-path call
1292 CallNode *call = new (C, slow_call_type->domain()->cnt())
1293 CallStaticJavaNode(slow_call_type, slow_call_address,
1294 OptoRuntime::stub_name(slow_call_address),
1295 alloc->jvms()->bci(),
1296 TypePtr::BOTTOM);
1297 call->init_req( TypeFunc::Control, slow_region );
1298 call->init_req( TypeFunc::I_O , top() ) ; // does no i/o
1299 call->init_req( TypeFunc::Memory , slow_mem ); // may gc ptrs
1300 call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) );
1301 call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) );
1303 call->init_req(TypeFunc::Parms+0, klass_node);
1304 if (length != NULL) {
1305 call->init_req(TypeFunc::Parms+1, length);
1306 }
1308 // Copy debug information and adjust JVMState information, then replace
1309 // allocate node with the call
1310 copy_call_debug_info((CallNode *) alloc, call);
1311 if (!always_slow) {
1312 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1313 }
1314 _igvn.hash_delete(alloc);
1315 _igvn.subsume_node(alloc, call);
1316 transform_later(call);
1318 // Identify the output projections from the allocate node and
1319 // adjust any references to them.
1320 // The control and io projections look like:
1321 //
1322 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl)
1323 // Allocate Catch
1324 // ^---Proj(io) <-------+ ^---CatchProj(io)
1325 //
1326 // We are interested in the CatchProj nodes.
1327 //
1328 extract_call_projections(call);
1330 // An allocate node has separate memory projections for the uses on the control and i_o paths
1331 // Replace uses of the control memory projection with result_phi_rawmem (unless we are only generating a slow call)
1332 if (!always_slow && _memproj_fallthrough != NULL) {
1333 for (DUIterator_Fast imax, i = _memproj_fallthrough->fast_outs(imax); i < imax; i++) {
1334 Node *use = _memproj_fallthrough->fast_out(i);
1335 _igvn.hash_delete(use);
1336 imax -= replace_input(use, _memproj_fallthrough, result_phi_rawmem);
1337 _igvn._worklist.push(use);
1338 // back up iterator
1339 --i;
1340 }
1341 }
1342 // Now change uses of _memproj_catchall to use _memproj_fallthrough and delete _memproj_catchall so
1343 // we end up with a call that has only 1 memory projection
1344 if (_memproj_catchall != NULL ) {
1345 if (_memproj_fallthrough == NULL) {
1346 _memproj_fallthrough = new (C, 1) ProjNode(call, TypeFunc::Memory);
1347 transform_later(_memproj_fallthrough);
1348 }
1349 for (DUIterator_Fast imax, i = _memproj_catchall->fast_outs(imax); i < imax; i++) {
1350 Node *use = _memproj_catchall->fast_out(i);
1351 _igvn.hash_delete(use);
1352 imax -= replace_input(use, _memproj_catchall, _memproj_fallthrough);
1353 _igvn._worklist.push(use);
1354 // back up iterator
1355 --i;
1356 }
1357 }
1359 // An allocate node has separate i_o projections for the uses on the control and i_o paths
1360 // Replace uses of the control i_o projection with result_phi_i_o (unless we are only generating a slow call)
1361 if (_ioproj_fallthrough == NULL) {
1362 _ioproj_fallthrough = new (C, 1) ProjNode(call, TypeFunc::I_O);
1363 transform_later(_ioproj_fallthrough);
1364 } else if (!always_slow) {
1365 for (DUIterator_Fast imax, i = _ioproj_fallthrough->fast_outs(imax); i < imax; i++) {
1366 Node *use = _ioproj_fallthrough->fast_out(i);
1368 _igvn.hash_delete(use);
1369 imax -= replace_input(use, _ioproj_fallthrough, result_phi_i_o);
1370 _igvn._worklist.push(use);
1371 // back up iterator
1372 --i;
1373 }
1374 }
1375 // Now change uses of _ioproj_catchall to use _ioproj_fallthrough and delete _ioproj_catchall so
1376 // we end up with a call that has only 1 control projection
1377 if (_ioproj_catchall != NULL ) {
1378 for (DUIterator_Fast imax, i = _ioproj_catchall->fast_outs(imax); i < imax; i++) {
1379 Node *use = _ioproj_catchall->fast_out(i);
1380 _igvn.hash_delete(use);
1381 imax -= replace_input(use, _ioproj_catchall, _ioproj_fallthrough);
1382 _igvn._worklist.push(use);
1383 // back up iterator
1384 --i;
1385 }
1386 }
1388 // if we generated only a slow call, we are done
1389 if (always_slow)
1390 return;
1393 if (_fallthroughcatchproj != NULL) {
1394 ctrl = _fallthroughcatchproj->clone();
1395 transform_later(ctrl);
1396 _igvn.replace_node(_fallthroughcatchproj, result_region);
1397 } else {
1398 ctrl = top();
1399 }
1400 Node *slow_result;
1401 if (_resproj == NULL) {
1402 // no uses of the allocation result
1403 slow_result = top();
1404 } else {
1405 slow_result = _resproj->clone();
1406 transform_later(slow_result);
1407 _igvn.replace_node(_resproj, result_phi_rawoop);
1408 }
1410 // Plug slow-path into result merge point
1411 result_region ->init_req( slow_result_path, ctrl );
1412 result_phi_rawoop->init_req( slow_result_path, slow_result);
1413 result_phi_rawmem->init_req( slow_result_path, _memproj_fallthrough );
1414 transform_later(result_region);
1415 transform_later(result_phi_rawoop);
1416 transform_later(result_phi_rawmem);
1417 transform_later(result_phi_i_o);
1418 // This completes all paths into the result merge point
1419 }
1422 // Helper for PhaseMacroExpand::expand_allocate_common.
1423 // Initializes the newly-allocated storage.
1424 Node*
1425 PhaseMacroExpand::initialize_object(AllocateNode* alloc,
1426 Node* control, Node* rawmem, Node* object,
1427 Node* klass_node, Node* length,
1428 Node* size_in_bytes) {
1429 InitializeNode* init = alloc->initialization();
1430 // Store the klass & mark bits
1431 Node* mark_node = NULL;
1432 // For now only enable fast locking for non-array types
1433 if (UseBiasedLocking && (length == NULL)) {
1434 mark_node = make_load(NULL, rawmem, klass_node, Klass::prototype_header_offset_in_bytes() + sizeof(oopDesc), TypeRawPtr::BOTTOM, T_ADDRESS);
1435 } else {
1436 mark_node = makecon(TypeRawPtr::make((address)markOopDesc::prototype()));
1437 }
1438 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS);
1440 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_OBJECT);
1441 int header_size = alloc->minimum_header_size(); // conservatively small
1443 // Array length
1444 if (length != NULL) { // Arrays need length field
1445 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
1446 // conservatively small header size:
1447 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1448 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
1449 if (k->is_array_klass()) // we know the exact header size in most cases:
1450 header_size = Klass::layout_helper_header_size(k->layout_helper());
1451 }
1453 // Clear the object body, if necessary.
1454 if (init == NULL) {
1455 // The init has somehow disappeared; be cautious and clear everything.
1456 //
1457 // This can happen if a node is allocated but an uncommon trap occurs
1458 // immediately. In this case, the Initialize gets associated with the
1459 // trap, and may be placed in a different (outer) loop, if the Allocate
1460 // is in a loop. If (this is rare) the inner loop gets unrolled, then
1461 // there can be two Allocates to one Initialize. The answer in all these
1462 // edge cases is safety first. It is always safe to clear immediately
1463 // within an Allocate, and then (maybe or maybe not) clear some more later.
1464 if (!ZeroTLAB)
1465 rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
1466 header_size, size_in_bytes,
1467 &_igvn);
1468 } else {
1469 if (!init->is_complete()) {
1470 // Try to win by zeroing only what the init does not store.
1471 // We can also try to do some peephole optimizations,
1472 // such as combining some adjacent subword stores.
1473 rawmem = init->complete_stores(control, rawmem, object,
1474 header_size, size_in_bytes, &_igvn);
1475 }
1476 // We have no more use for this link, since the AllocateNode goes away:
1477 init->set_req(InitializeNode::RawAddress, top());
1478 // (If we keep the link, it just confuses the register allocator,
1479 // who thinks he sees a real use of the address by the membar.)
1480 }
1482 return rawmem;
1483 }
1485 // Generate prefetch instructions for next allocations.
1486 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false,
1487 Node*& contended_phi_rawmem,
1488 Node* old_eden_top, Node* new_eden_top,
1489 Node* length) {
1490 if( UseTLAB && AllocatePrefetchStyle == 2 ) {
1491 // Generate prefetch allocation with watermark check.
1492 // As an allocation hits the watermark, we will prefetch starting
1493 // at a "distance" away from watermark.
1494 enum { fall_in_path = 1, pf_path = 2 };
1496 Node *pf_region = new (C, 3) RegionNode(3);
1497 Node *pf_phi_rawmem = new (C, 3) PhiNode( pf_region, Type::MEMORY,
1498 TypeRawPtr::BOTTOM );
1499 // I/O is used for Prefetch
1500 Node *pf_phi_abio = new (C, 3) PhiNode( pf_region, Type::ABIO );
1502 Node *thread = new (C, 1) ThreadLocalNode();
1503 transform_later(thread);
1505 Node *eden_pf_adr = new (C, 4) AddPNode( top()/*not oop*/, thread,
1506 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) );
1507 transform_later(eden_pf_adr);
1509 Node *old_pf_wm = new (C, 3) LoadPNode( needgc_false,
1510 contended_phi_rawmem, eden_pf_adr,
1511 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM );
1512 transform_later(old_pf_wm);
1514 // check against new_eden_top
1515 Node *need_pf_cmp = new (C, 3) CmpPNode( new_eden_top, old_pf_wm );
1516 transform_later(need_pf_cmp);
1517 Node *need_pf_bol = new (C, 2) BoolNode( need_pf_cmp, BoolTest::ge );
1518 transform_later(need_pf_bol);
1519 IfNode *need_pf_iff = new (C, 2) IfNode( needgc_false, need_pf_bol,
1520 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN );
1521 transform_later(need_pf_iff);
1523 // true node, add prefetchdistance
1524 Node *need_pf_true = new (C, 1) IfTrueNode( need_pf_iff );
1525 transform_later(need_pf_true);
1527 Node *need_pf_false = new (C, 1) IfFalseNode( need_pf_iff );
1528 transform_later(need_pf_false);
1530 Node *new_pf_wmt = new (C, 4) AddPNode( top(), old_pf_wm,
1531 _igvn.MakeConX(AllocatePrefetchDistance) );
1532 transform_later(new_pf_wmt );
1533 new_pf_wmt->set_req(0, need_pf_true);
1535 Node *store_new_wmt = new (C, 4) StorePNode( need_pf_true,
1536 contended_phi_rawmem, eden_pf_adr,
1537 TypeRawPtr::BOTTOM, new_pf_wmt );
1538 transform_later(store_new_wmt);
1540 // adding prefetches
1541 pf_phi_abio->init_req( fall_in_path, i_o );
1543 Node *prefetch_adr;
1544 Node *prefetch;
1545 uint lines = AllocatePrefetchDistance / AllocatePrefetchStepSize;
1546 uint step_size = AllocatePrefetchStepSize;
1547 uint distance = 0;
1549 for ( uint i = 0; i < lines; i++ ) {
1550 prefetch_adr = new (C, 4) AddPNode( old_pf_wm, new_pf_wmt,
1551 _igvn.MakeConX(distance) );
1552 transform_later(prefetch_adr);
1553 prefetch = new (C, 3) PrefetchWriteNode( i_o, prefetch_adr );
1554 transform_later(prefetch);
1555 distance += step_size;
1556 i_o = prefetch;
1557 }
1558 pf_phi_abio->set_req( pf_path, i_o );
1560 pf_region->init_req( fall_in_path, need_pf_false );
1561 pf_region->init_req( pf_path, need_pf_true );
1563 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem );
1564 pf_phi_rawmem->init_req( pf_path, store_new_wmt );
1566 transform_later(pf_region);
1567 transform_later(pf_phi_rawmem);
1568 transform_later(pf_phi_abio);
1570 needgc_false = pf_region;
1571 contended_phi_rawmem = pf_phi_rawmem;
1572 i_o = pf_phi_abio;
1573 } else if( AllocatePrefetchStyle > 0 ) {
1574 // Insert a prefetch for each allocation only on the fast-path
1575 Node *prefetch_adr;
1576 Node *prefetch;
1577 // Generate several prefetch instructions only for arrays.
1578 uint lines = (length != NULL) ? AllocatePrefetchLines : 1;
1579 uint step_size = AllocatePrefetchStepSize;
1580 uint distance = AllocatePrefetchDistance;
1581 for ( uint i = 0; i < lines; i++ ) {
1582 prefetch_adr = new (C, 4) AddPNode( old_eden_top, new_eden_top,
1583 _igvn.MakeConX(distance) );
1584 transform_later(prefetch_adr);
1585 prefetch = new (C, 3) PrefetchWriteNode( i_o, prefetch_adr );
1586 // Do not let it float too high, since if eden_top == eden_end,
1587 // both might be null.
1588 if( i == 0 ) { // Set control for first prefetch, next follows it
1589 prefetch->init_req(0, needgc_false);
1590 }
1591 transform_later(prefetch);
1592 distance += step_size;
1593 i_o = prefetch;
1594 }
1595 }
1596 return i_o;
1597 }
1600 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) {
1601 expand_allocate_common(alloc, NULL,
1602 OptoRuntime::new_instance_Type(),
1603 OptoRuntime::new_instance_Java());
1604 }
1606 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) {
1607 Node* length = alloc->in(AllocateNode::ALength);
1608 expand_allocate_common(alloc, length,
1609 OptoRuntime::new_array_Type(),
1610 OptoRuntime::new_array_Java());
1611 }
1614 // we have determined that this lock/unlock can be eliminated, we simply
1615 // eliminate the node without expanding it.
1616 //
1617 // Note: The membar's associated with the lock/unlock are currently not
1618 // eliminated. This should be investigated as a future enhancement.
1619 //
1620 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) {
1622 if (!alock->is_eliminated()) {
1623 return false;
1624 }
1625 if (alock->is_Lock() && !alock->is_coarsened()) {
1626 // Create new "eliminated" BoxLock node and use it
1627 // in monitor debug info for the same object.
1628 BoxLockNode* oldbox = alock->box_node()->as_BoxLock();
1629 Node* obj = alock->obj_node();
1630 if (!oldbox->is_eliminated()) {
1631 BoxLockNode* newbox = oldbox->clone()->as_BoxLock();
1632 newbox->set_eliminated();
1633 transform_later(newbox);
1634 // Replace old box node with new box for all users
1635 // of the same object.
1636 for (uint i = 0; i < oldbox->outcnt();) {
1638 bool next_edge = true;
1639 Node* u = oldbox->raw_out(i);
1640 if (u == alock) {
1641 i++;
1642 continue; // It will be removed below
1643 }
1644 if (u->is_Lock() &&
1645 u->as_Lock()->obj_node() == obj &&
1646 // oldbox could be referenced in debug info also
1647 u->as_Lock()->box_node() == oldbox) {
1648 assert(u->as_Lock()->is_eliminated(), "sanity");
1649 _igvn.hash_delete(u);
1650 u->set_req(TypeFunc::Parms + 1, newbox);
1651 next_edge = false;
1652 #ifdef ASSERT
1653 } else if (u->is_Unlock() && u->as_Unlock()->obj_node() == obj) {
1654 assert(u->as_Unlock()->is_eliminated(), "sanity");
1655 #endif
1656 }
1657 // Replace old box in monitor debug info.
1658 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) {
1659 SafePointNode* sfn = u->as_SafePoint();
1660 JVMState* youngest_jvms = sfn->jvms();
1661 int max_depth = youngest_jvms->depth();
1662 for (int depth = 1; depth <= max_depth; depth++) {
1663 JVMState* jvms = youngest_jvms->of_depth(depth);
1664 int num_mon = jvms->nof_monitors();
1665 // Loop over monitors
1666 for (int idx = 0; idx < num_mon; idx++) {
1667 Node* obj_node = sfn->monitor_obj(jvms, idx);
1668 Node* box_node = sfn->monitor_box(jvms, idx);
1669 if (box_node == oldbox && obj_node == obj) {
1670 int j = jvms->monitor_box_offset(idx);
1671 _igvn.hash_delete(u);
1672 u->set_req(j, newbox);
1673 next_edge = false;
1674 }
1675 } // for (int idx = 0;
1676 } // for (int depth = 1;
1677 } // if (u->is_SafePoint()
1678 if (next_edge) i++;
1679 } // for (uint i = 0; i < oldbox->outcnt();)
1680 } // if (!oldbox->is_eliminated())
1681 } // if (alock->is_Lock() && !lock->is_coarsened())
1683 CompileLog* log = C->log();
1684 if (log != NULL) {
1685 log->head("eliminate_lock lock='%d'",
1686 alock->is_Lock());
1687 JVMState* p = alock->jvms();
1688 while (p != NULL) {
1689 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1690 p = p->caller();
1691 }
1692 log->tail("eliminate_lock");
1693 }
1695 #ifndef PRODUCT
1696 if (PrintEliminateLocks) {
1697 if (alock->is_Lock()) {
1698 tty->print_cr("++++ Eliminating: %d Lock", alock->_idx);
1699 } else {
1700 tty->print_cr("++++ Eliminating: %d Unlock", alock->_idx);
1701 }
1702 }
1703 #endif
1705 Node* mem = alock->in(TypeFunc::Memory);
1706 Node* ctrl = alock->in(TypeFunc::Control);
1708 extract_call_projections(alock);
1709 // There are 2 projections from the lock. The lock node will
1710 // be deleted when its last use is subsumed below.
1711 assert(alock->outcnt() == 2 &&
1712 _fallthroughproj != NULL &&
1713 _memproj_fallthrough != NULL,
1714 "Unexpected projections from Lock/Unlock");
1716 Node* fallthroughproj = _fallthroughproj;
1717 Node* memproj_fallthrough = _memproj_fallthrough;
1719 // The memory projection from a lock/unlock is RawMem
1720 // The input to a Lock is merged memory, so extract its RawMem input
1721 // (unless the MergeMem has been optimized away.)
1722 if (alock->is_Lock()) {
1723 // Seach for MemBarAcquire node and delete it also.
1724 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
1725 assert(membar != NULL && membar->Opcode() == Op_MemBarAcquire, "");
1726 Node* ctrlproj = membar->proj_out(TypeFunc::Control);
1727 Node* memproj = membar->proj_out(TypeFunc::Memory);
1728 _igvn.replace_node(ctrlproj, fallthroughproj);
1729 _igvn.replace_node(memproj, memproj_fallthrough);
1731 // Delete FastLock node also if this Lock node is unique user
1732 // (a loop peeling may clone a Lock node).
1733 Node* flock = alock->as_Lock()->fastlock_node();
1734 if (flock->outcnt() == 1) {
1735 assert(flock->unique_out() == alock, "sanity");
1736 _igvn.replace_node(flock, top());
1737 }
1738 }
1740 // Seach for MemBarRelease node and delete it also.
1741 if (alock->is_Unlock() && ctrl != NULL && ctrl->is_Proj() &&
1742 ctrl->in(0)->is_MemBar()) {
1743 MemBarNode* membar = ctrl->in(0)->as_MemBar();
1744 assert(membar->Opcode() == Op_MemBarRelease &&
1745 mem->is_Proj() && membar == mem->in(0), "");
1746 _igvn.replace_node(fallthroughproj, ctrl);
1747 _igvn.replace_node(memproj_fallthrough, mem);
1748 fallthroughproj = ctrl;
1749 memproj_fallthrough = mem;
1750 ctrl = membar->in(TypeFunc::Control);
1751 mem = membar->in(TypeFunc::Memory);
1752 }
1754 _igvn.replace_node(fallthroughproj, ctrl);
1755 _igvn.replace_node(memproj_fallthrough, mem);
1756 return true;
1757 }
1760 //------------------------------expand_lock_node----------------------
1761 void PhaseMacroExpand::expand_lock_node(LockNode *lock) {
1763 Node* ctrl = lock->in(TypeFunc::Control);
1764 Node* mem = lock->in(TypeFunc::Memory);
1765 Node* obj = lock->obj_node();
1766 Node* box = lock->box_node();
1767 Node* flock = lock->fastlock_node();
1769 // Make the merge point
1770 Node *region;
1771 Node *mem_phi;
1772 Node *slow_path;
1774 if (UseOptoBiasInlining) {
1775 /*
1776 * See the full description in MacroAssembler::biased_locking_enter().
1777 *
1778 * if( (mark_word & biased_lock_mask) == biased_lock_pattern ) {
1779 * // The object is biased.
1780 * proto_node = klass->prototype_header;
1781 * o_node = thread | proto_node;
1782 * x_node = o_node ^ mark_word;
1783 * if( (x_node & ~age_mask) == 0 ) { // Biased to the current thread ?
1784 * // Done.
1785 * } else {
1786 * if( (x_node & biased_lock_mask) != 0 ) {
1787 * // The klass's prototype header is no longer biased.
1788 * cas(&mark_word, mark_word, proto_node)
1789 * goto cas_lock;
1790 * } else {
1791 * // The klass's prototype header is still biased.
1792 * if( (x_node & epoch_mask) != 0 ) { // Expired epoch?
1793 * old = mark_word;
1794 * new = o_node;
1795 * } else {
1796 * // Different thread or anonymous biased.
1797 * old = mark_word & (epoch_mask | age_mask | biased_lock_mask);
1798 * new = thread | old;
1799 * }
1800 * // Try to rebias.
1801 * if( cas(&mark_word, old, new) == 0 ) {
1802 * // Done.
1803 * } else {
1804 * goto slow_path; // Failed.
1805 * }
1806 * }
1807 * }
1808 * } else {
1809 * // The object is not biased.
1810 * cas_lock:
1811 * if( FastLock(obj) == 0 ) {
1812 * // Done.
1813 * } else {
1814 * slow_path:
1815 * OptoRuntime::complete_monitor_locking_Java(obj);
1816 * }
1817 * }
1818 */
1820 region = new (C, 5) RegionNode(5);
1821 // create a Phi for the memory state
1822 mem_phi = new (C, 5) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
1824 Node* fast_lock_region = new (C, 3) RegionNode(3);
1825 Node* fast_lock_mem_phi = new (C, 3) PhiNode( fast_lock_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1827 // First, check mark word for the biased lock pattern.
1828 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
1830 // Get fast path - mark word has the biased lock pattern.
1831 ctrl = opt_bits_test(ctrl, fast_lock_region, 1, mark_node,
1832 markOopDesc::biased_lock_mask_in_place,
1833 markOopDesc::biased_lock_pattern, true);
1834 // fast_lock_region->in(1) is set to slow path.
1835 fast_lock_mem_phi->init_req(1, mem);
1837 // Now check that the lock is biased to the current thread and has
1838 // the same epoch and bias as Klass::_prototype_header.
1840 // Special-case a fresh allocation to avoid building nodes:
1841 Node* klass_node = AllocateNode::Ideal_klass(obj, &_igvn);
1842 if (klass_node == NULL) {
1843 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1844 klass_node = transform_later( LoadKlassNode::make(_igvn, mem, k_adr, _igvn.type(k_adr)->is_ptr()) );
1845 #ifdef _LP64
1846 if (UseCompressedOops && klass_node->is_DecodeN()) {
1847 assert(klass_node->in(1)->Opcode() == Op_LoadNKlass, "sanity");
1848 klass_node->in(1)->init_req(0, ctrl);
1849 } else
1850 #endif
1851 klass_node->init_req(0, ctrl);
1852 }
1853 Node *proto_node = make_load(ctrl, mem, klass_node, Klass::prototype_header_offset_in_bytes() + sizeof(oopDesc), TypeX_X, TypeX_X->basic_type());
1855 Node* thread = transform_later(new (C, 1) ThreadLocalNode());
1856 Node* cast_thread = transform_later(new (C, 2) CastP2XNode(ctrl, thread));
1857 Node* o_node = transform_later(new (C, 3) OrXNode(cast_thread, proto_node));
1858 Node* x_node = transform_later(new (C, 3) XorXNode(o_node, mark_node));
1860 // Get slow path - mark word does NOT match the value.
1861 Node* not_biased_ctrl = opt_bits_test(ctrl, region, 3, x_node,
1862 (~markOopDesc::age_mask_in_place), 0);
1863 // region->in(3) is set to fast path - the object is biased to the current thread.
1864 mem_phi->init_req(3, mem);
1867 // Mark word does NOT match the value (thread | Klass::_prototype_header).
1870 // First, check biased pattern.
1871 // Get fast path - _prototype_header has the same biased lock pattern.
1872 ctrl = opt_bits_test(not_biased_ctrl, fast_lock_region, 2, x_node,
1873 markOopDesc::biased_lock_mask_in_place, 0, true);
1875 not_biased_ctrl = fast_lock_region->in(2); // Slow path
1876 // fast_lock_region->in(2) - the prototype header is no longer biased
1877 // and we have to revoke the bias on this object.
1878 // We are going to try to reset the mark of this object to the prototype
1879 // value and fall through to the CAS-based locking scheme.
1880 Node* adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
1881 Node* cas = new (C, 5) StoreXConditionalNode(not_biased_ctrl, mem, adr,
1882 proto_node, mark_node);
1883 transform_later(cas);
1884 Node* proj = transform_later( new (C, 1) SCMemProjNode(cas));
1885 fast_lock_mem_phi->init_req(2, proj);
1888 // Second, check epoch bits.
1889 Node* rebiased_region = new (C, 3) RegionNode(3);
1890 Node* old_phi = new (C, 3) PhiNode( rebiased_region, TypeX_X);
1891 Node* new_phi = new (C, 3) PhiNode( rebiased_region, TypeX_X);
1893 // Get slow path - mark word does NOT match epoch bits.
1894 Node* epoch_ctrl = opt_bits_test(ctrl, rebiased_region, 1, x_node,
1895 markOopDesc::epoch_mask_in_place, 0);
1896 // The epoch of the current bias is not valid, attempt to rebias the object
1897 // toward the current thread.
1898 rebiased_region->init_req(2, epoch_ctrl);
1899 old_phi->init_req(2, mark_node);
1900 new_phi->init_req(2, o_node);
1902 // rebiased_region->in(1) is set to fast path.
1903 // The epoch of the current bias is still valid but we know
1904 // nothing about the owner; it might be set or it might be clear.
1905 Node* cmask = MakeConX(markOopDesc::biased_lock_mask_in_place |
1906 markOopDesc::age_mask_in_place |
1907 markOopDesc::epoch_mask_in_place);
1908 Node* old = transform_later(new (C, 3) AndXNode(mark_node, cmask));
1909 cast_thread = transform_later(new (C, 2) CastP2XNode(ctrl, thread));
1910 Node* new_mark = transform_later(new (C, 3) OrXNode(cast_thread, old));
1911 old_phi->init_req(1, old);
1912 new_phi->init_req(1, new_mark);
1914 transform_later(rebiased_region);
1915 transform_later(old_phi);
1916 transform_later(new_phi);
1918 // Try to acquire the bias of the object using an atomic operation.
1919 // If this fails we will go in to the runtime to revoke the object's bias.
1920 cas = new (C, 5) StoreXConditionalNode(rebiased_region, mem, adr,
1921 new_phi, old_phi);
1922 transform_later(cas);
1923 proj = transform_later( new (C, 1) SCMemProjNode(cas));
1925 // Get slow path - Failed to CAS.
1926 not_biased_ctrl = opt_bits_test(rebiased_region, region, 4, cas, 0, 0);
1927 mem_phi->init_req(4, proj);
1928 // region->in(4) is set to fast path - the object is rebiased to the current thread.
1930 // Failed to CAS.
1931 slow_path = new (C, 3) RegionNode(3);
1932 Node *slow_mem = new (C, 3) PhiNode( slow_path, Type::MEMORY, TypeRawPtr::BOTTOM);
1934 slow_path->init_req(1, not_biased_ctrl); // Capture slow-control
1935 slow_mem->init_req(1, proj);
1937 // Call CAS-based locking scheme (FastLock node).
1939 transform_later(fast_lock_region);
1940 transform_later(fast_lock_mem_phi);
1942 // Get slow path - FastLock failed to lock the object.
1943 ctrl = opt_bits_test(fast_lock_region, region, 2, flock, 0, 0);
1944 mem_phi->init_req(2, fast_lock_mem_phi);
1945 // region->in(2) is set to fast path - the object is locked to the current thread.
1947 slow_path->init_req(2, ctrl); // Capture slow-control
1948 slow_mem->init_req(2, fast_lock_mem_phi);
1950 transform_later(slow_path);
1951 transform_later(slow_mem);
1952 // Reset lock's memory edge.
1953 lock->set_req(TypeFunc::Memory, slow_mem);
1955 } else {
1956 region = new (C, 3) RegionNode(3);
1957 // create a Phi for the memory state
1958 mem_phi = new (C, 3) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
1960 // Optimize test; set region slot 2
1961 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0);
1962 mem_phi->init_req(2, mem);
1963 }
1965 // Make slow path call
1966 CallNode *call = make_slow_call( (CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(), OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path, obj, box );
1968 extract_call_projections(call);
1970 // Slow path can only throw asynchronous exceptions, which are always
1971 // de-opted. So the compiler thinks the slow-call can never throw an
1972 // exception. If it DOES throw an exception we would need the debug
1973 // info removed first (since if it throws there is no monitor).
1974 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
1975 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
1977 // Capture slow path
1978 // disconnect fall-through projection from call and create a new one
1979 // hook up users of fall-through projection to region
1980 Node *slow_ctrl = _fallthroughproj->clone();
1981 transform_later(slow_ctrl);
1982 _igvn.hash_delete(_fallthroughproj);
1983 _fallthroughproj->disconnect_inputs(NULL);
1984 region->init_req(1, slow_ctrl);
1985 // region inputs are now complete
1986 transform_later(region);
1987 _igvn.replace_node(_fallthroughproj, region);
1989 Node *memproj = transform_later( new(C, 1) ProjNode(call, TypeFunc::Memory) );
1990 mem_phi->init_req(1, memproj );
1991 transform_later(mem_phi);
1992 _igvn.replace_node(_memproj_fallthrough, mem_phi);
1993 }
1995 //------------------------------expand_unlock_node----------------------
1996 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
1998 Node* ctrl = unlock->in(TypeFunc::Control);
1999 Node* mem = unlock->in(TypeFunc::Memory);
2000 Node* obj = unlock->obj_node();
2001 Node* box = unlock->box_node();
2003 // No need for a null check on unlock
2005 // Make the merge point
2006 Node *region;
2007 Node *mem_phi;
2009 if (UseOptoBiasInlining) {
2010 // Check for biased locking unlock case, which is a no-op.
2011 // See the full description in MacroAssembler::biased_locking_exit().
2012 region = new (C, 4) RegionNode(4);
2013 // create a Phi for the memory state
2014 mem_phi = new (C, 4) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2015 mem_phi->init_req(3, mem);
2017 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2018 ctrl = opt_bits_test(ctrl, region, 3, mark_node,
2019 markOopDesc::biased_lock_mask_in_place,
2020 markOopDesc::biased_lock_pattern);
2021 } else {
2022 region = new (C, 3) RegionNode(3);
2023 // create a Phi for the memory state
2024 mem_phi = new (C, 3) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2025 }
2027 FastUnlockNode *funlock = new (C, 3) FastUnlockNode( ctrl, obj, box );
2028 funlock = transform_later( funlock )->as_FastUnlock();
2029 // Optimize test; set region slot 2
2030 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0);
2032 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 );
2034 extract_call_projections(call);
2036 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
2037 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
2039 // No exceptions for unlocking
2040 // Capture slow path
2041 // disconnect fall-through projection from call and create a new one
2042 // hook up users of fall-through projection to region
2043 Node *slow_ctrl = _fallthroughproj->clone();
2044 transform_later(slow_ctrl);
2045 _igvn.hash_delete(_fallthroughproj);
2046 _fallthroughproj->disconnect_inputs(NULL);
2047 region->init_req(1, slow_ctrl);
2048 // region inputs are now complete
2049 transform_later(region);
2050 _igvn.replace_node(_fallthroughproj, region);
2052 Node *memproj = transform_later( new(C, 1) ProjNode(call, TypeFunc::Memory) );
2053 mem_phi->init_req(1, memproj );
2054 mem_phi->init_req(2, mem);
2055 transform_later(mem_phi);
2056 _igvn.replace_node(_memproj_fallthrough, mem_phi);
2057 }
2059 //------------------------------expand_macro_nodes----------------------
2060 // Returns true if a failure occurred.
2061 bool PhaseMacroExpand::expand_macro_nodes() {
2062 if (C->macro_count() == 0)
2063 return false;
2064 // First, attempt to eliminate locks
2065 bool progress = true;
2066 while (progress) {
2067 progress = false;
2068 for (int i = C->macro_count(); i > 0; i--) {
2069 Node * n = C->macro_node(i-1);
2070 bool success = false;
2071 debug_only(int old_macro_count = C->macro_count(););
2072 if (n->is_AbstractLock()) {
2073 success = eliminate_locking_node(n->as_AbstractLock());
2074 } else if (n->Opcode() == Op_Opaque1 || n->Opcode() == Op_Opaque2) {
2075 _igvn.replace_node(n, n->in(1));
2076 success = true;
2077 }
2078 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2079 progress = progress || success;
2080 }
2081 }
2082 // Next, attempt to eliminate allocations
2083 progress = true;
2084 while (progress) {
2085 progress = false;
2086 for (int i = C->macro_count(); i > 0; i--) {
2087 Node * n = C->macro_node(i-1);
2088 bool success = false;
2089 debug_only(int old_macro_count = C->macro_count(););
2090 switch (n->class_id()) {
2091 case Node::Class_Allocate:
2092 case Node::Class_AllocateArray:
2093 success = eliminate_allocate_node(n->as_Allocate());
2094 break;
2095 case Node::Class_Lock:
2096 case Node::Class_Unlock:
2097 assert(!n->as_AbstractLock()->is_eliminated(), "sanity");
2098 break;
2099 default:
2100 assert(false, "unknown node type in macro list");
2101 }
2102 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2103 progress = progress || success;
2104 }
2105 }
2106 // Make sure expansion will not cause node limit to be exceeded.
2107 // Worst case is a macro node gets expanded into about 50 nodes.
2108 // Allow 50% more for optimization.
2109 if (C->check_node_count(C->macro_count() * 75, "out of nodes before macro expansion" ) )
2110 return true;
2112 // expand "macro" nodes
2113 // nodes are removed from the macro list as they are processed
2114 while (C->macro_count() > 0) {
2115 int macro_count = C->macro_count();
2116 Node * n = C->macro_node(macro_count-1);
2117 assert(n->is_macro(), "only macro nodes expected here");
2118 if (_igvn.type(n) == Type::TOP || n->in(0)->is_top() ) {
2119 // node is unreachable, so don't try to expand it
2120 C->remove_macro_node(n);
2121 continue;
2122 }
2123 switch (n->class_id()) {
2124 case Node::Class_Allocate:
2125 expand_allocate(n->as_Allocate());
2126 break;
2127 case Node::Class_AllocateArray:
2128 expand_allocate_array(n->as_AllocateArray());
2129 break;
2130 case Node::Class_Lock:
2131 expand_lock_node(n->as_Lock());
2132 break;
2133 case Node::Class_Unlock:
2134 expand_unlock_node(n->as_Unlock());
2135 break;
2136 default:
2137 assert(false, "unknown node type in macro list");
2138 }
2139 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
2140 if (C->failing()) return true;
2141 }
2143 _igvn.set_delay_transform(false);
2144 _igvn.optimize();
2145 return false;
2146 }