1 /*
2 * Copyright (c) 2005, 2022, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "compiler/compileLog.hpp"
27 #include "gc/shared/collectedHeap.inline.hpp"
28 #include "gc/shared/tlab_globals.hpp"
29 #include "libadt/vectset.hpp"
30 #include "memory/universe.hpp"
31 #include "opto/addnode.hpp"
32 #include "opto/arraycopynode.hpp"
33 #include "opto/callnode.hpp"
34 #include "opto/castnode.hpp"
35 #include "opto/cfgnode.hpp"
36 #include "opto/compile.hpp"
37 #include "opto/convertnode.hpp"
38 #include "opto/graphKit.hpp"
39 #include "opto/intrinsicnode.hpp"
40 #include "opto/locknode.hpp"
41 #include "opto/loopnode.hpp"
42 #include "opto/macro.hpp"
43 #include "opto/memnode.hpp"
44 #include "opto/narrowptrnode.hpp"
45 #include "opto/node.hpp"
46 #include "opto/opaquenode.hpp"
47 #include "opto/phaseX.hpp"
48 #include "opto/rootnode.hpp"
49 #include "opto/runtime.hpp"
50 #include "opto/subnode.hpp"
51 #include "opto/subtypenode.hpp"
52 #include "opto/type.hpp"
53 #include "prims/jvmtiExport.hpp"
54 #include "runtime/sharedRuntime.hpp"
55 #include "utilities/macros.hpp"
56 #include "utilities/powerOfTwo.hpp"
57 #if INCLUDE_G1GC
58 #include "gc/g1/g1ThreadLocalData.hpp"
59 #endif // INCLUDE_G1GC
60 #if INCLUDE_SHENANDOAHGC
61 #include "gc/shenandoah/c2/shenandoahBarrierSetC2.hpp"
62 #endif
63
64
65 //
66 // Replace any references to "oldref" in inputs to "use" with "newref".
67 // Returns the number of replacements made.
68 //
69 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) {
70 int nreplacements = 0;
71 uint req = use->req();
72 for (uint j = 0; j < use->len(); j++) {
73 Node *uin = use->in(j);
74 if (uin == oldref) {
75 if (j < req)
76 use->set_req(j, newref);
77 else
78 use->set_prec(j, newref);
79 nreplacements++;
80 } else if (j >= req && uin == NULL) {
81 break;
82 }
83 }
84 return nreplacements;
85 }
86
87 void PhaseMacroExpand::migrate_outs(Node *old, Node *target) {
88 assert(old != NULL, "sanity");
89 for (DUIterator_Fast imax, i = old->fast_outs(imax); i < imax; i++) {
90 Node* use = old->fast_out(i);
91 _igvn.rehash_node_delayed(use);
92 imax -= replace_input(use, old, target);
93 // back up iterator
94 --i;
95 }
96 assert(old->outcnt() == 0, "all uses must be deleted");
97 }
98
99 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word, int mask, int bits, bool return_fast_path) {
100 Node* cmp;
101 if (mask != 0) {
102 Node* and_node = transform_later(new AndXNode(word, MakeConX(mask)));
103 cmp = transform_later(new CmpXNode(and_node, MakeConX(bits)));
104 } else {
105 cmp = word;
106 }
107 Node* bol = transform_later(new BoolNode(cmp, BoolTest::ne));
108 IfNode* iff = new IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN );
109 transform_later(iff);
110
111 // Fast path taken.
112 Node *fast_taken = transform_later(new IfFalseNode(iff));
113
114 // Fast path not-taken, i.e. slow path
115 Node *slow_taken = transform_later(new IfTrueNode(iff));
116
117 if (return_fast_path) {
118 region->init_req(edge, slow_taken); // Capture slow-control
119 return fast_taken;
120 } else {
121 region->init_req(edge, fast_taken); // Capture fast-control
122 return slow_taken;
123 }
124 }
125
126 //--------------------copy_predefined_input_for_runtime_call--------------------
127 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) {
128 // Set fixed predefined input arguments
129 call->init_req( TypeFunc::Control, ctrl );
130 call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) );
131 call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ?????
132 call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) );
133 call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) );
134 }
135
136 //------------------------------make_slow_call---------------------------------
137 CallNode* PhaseMacroExpand::make_slow_call(CallNode *oldcall, const TypeFunc* slow_call_type,
138 address slow_call, const char* leaf_name, Node* slow_path,
139 Node* parm0, Node* parm1, Node* parm2) {
140
141 // Slow-path call
142 CallNode *call = leaf_name
143 ? (CallNode*)new CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM )
144 : (CallNode*)new CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), TypeRawPtr::BOTTOM );
145
146 // Slow path call has no side-effects, uses few values
147 copy_predefined_input_for_runtime_call(slow_path, oldcall, call );
148 if (parm0 != NULL) call->init_req(TypeFunc::Parms+0, parm0);
149 if (parm1 != NULL) call->init_req(TypeFunc::Parms+1, parm1);
150 if (parm2 != NULL) call->init_req(TypeFunc::Parms+2, parm2);
151 call->copy_call_debug_info(&_igvn, oldcall);
152 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
153 _igvn.replace_node(oldcall, call);
154 transform_later(call);
155
156 return call;
157 }
158
159 void PhaseMacroExpand::eliminate_gc_barrier(Node* p2x) {
160 BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2();
161 bs->eliminate_gc_barrier(this, p2x);
162 }
163
164 // Search for a memory operation for the specified memory slice.
165 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) {
166 Node *orig_mem = mem;
167 Node *alloc_mem = alloc->in(TypeFunc::Memory);
168 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr();
169 while (true) {
170 if (mem == alloc_mem || mem == start_mem ) {
171 return mem; // hit one of our sentinels
172 } else if (mem->is_MergeMem()) {
173 mem = mem->as_MergeMem()->memory_at(alias_idx);
174 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) {
175 Node *in = mem->in(0);
176 // we can safely skip over safepoints, calls, locks and membars because we
177 // already know that the object is safe to eliminate.
178 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) {
179 return in;
180 } else if (in->is_Call()) {
181 CallNode *call = in->as_Call();
182 if (call->may_modify(tinst, phase)) {
183 assert(call->is_ArrayCopy(), "ArrayCopy is the only call node that doesn't make allocation escape");
184 if (call->as_ArrayCopy()->modifies(offset, offset, phase, false)) {
185 return in;
186 }
187 }
188 mem = in->in(TypeFunc::Memory);
189 } else if (in->is_MemBar()) {
190 ArrayCopyNode* ac = NULL;
191 if (ArrayCopyNode::may_modify(tinst, in->as_MemBar(), phase, ac)) {
192 if (ac != NULL) {
193 assert(ac->is_clonebasic(), "Only basic clone is a non escaping clone");
194 return ac;
195 }
196 }
197 mem = in->in(TypeFunc::Memory);
198 } else {
199 #ifdef ASSERT
200 in->dump();
201 mem->dump();
202 assert(false, "unexpected projection");
203 #endif
204 }
205 } else if (mem->is_Store()) {
206 const TypePtr* atype = mem->as_Store()->adr_type();
207 int adr_idx = phase->C->get_alias_index(atype);
208 if (adr_idx == alias_idx) {
209 assert(atype->isa_oopptr(), "address type must be oopptr");
210 int adr_offset = atype->offset();
211 uint adr_iid = atype->is_oopptr()->instance_id();
212 // Array elements references have the same alias_idx
213 // but different offset and different instance_id.
214 if (adr_offset == offset && adr_iid == alloc->_idx) {
215 return mem;
216 }
217 } else {
218 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw");
219 }
220 mem = mem->in(MemNode::Memory);
221 } else if (mem->is_ClearArray()) {
222 if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) {
223 // Can not bypass initialization of the instance
224 // we are looking.
225 debug_only(intptr_t offset;)
226 assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity");
227 InitializeNode* init = alloc->as_Allocate()->initialization();
228 // We are looking for stored value, return Initialize node
229 // or memory edge from Allocate node.
230 if (init != NULL) {
231 return init;
232 } else {
233 return alloc->in(TypeFunc::Memory); // It will produce zero value (see callers).
234 }
235 }
236 // Otherwise skip it (the call updated 'mem' value).
237 } else if (mem->Opcode() == Op_SCMemProj) {
238 mem = mem->in(0);
239 Node* adr = NULL;
240 if (mem->is_LoadStore()) {
241 adr = mem->in(MemNode::Address);
242 } else {
243 assert(mem->Opcode() == Op_EncodeISOArray ||
244 mem->Opcode() == Op_StrCompressedCopy, "sanity");
245 adr = mem->in(3); // Destination array
246 }
247 const TypePtr* atype = adr->bottom_type()->is_ptr();
248 int adr_idx = phase->C->get_alias_index(atype);
249 if (adr_idx == alias_idx) {
250 DEBUG_ONLY(mem->dump();)
251 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
252 return NULL;
253 }
254 mem = mem->in(MemNode::Memory);
255 } else if (mem->Opcode() == Op_StrInflatedCopy) {
256 Node* adr = mem->in(3); // Destination array
257 const TypePtr* atype = adr->bottom_type()->is_ptr();
258 int adr_idx = phase->C->get_alias_index(atype);
259 if (adr_idx == alias_idx) {
260 DEBUG_ONLY(mem->dump();)
261 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field");
262 return NULL;
263 }
264 mem = mem->in(MemNode::Memory);
265 } else {
266 return mem;
267 }
268 assert(mem != orig_mem, "dead memory loop");
269 }
270 }
271
272 // Generate loads from source of the arraycopy for fields of
273 // destination needed at a deoptimization point
274 Node* PhaseMacroExpand::make_arraycopy_load(ArrayCopyNode* ac, intptr_t offset, Node* ctl, Node* mem, BasicType ft, const Type *ftype, AllocateNode *alloc) {
275 BasicType bt = ft;
276 const Type *type = ftype;
277 if (ft == T_NARROWOOP) {
278 bt = T_OBJECT;
279 type = ftype->make_oopptr();
280 }
281 Node* res = NULL;
282 if (ac->is_clonebasic()) {
283 assert(ac->in(ArrayCopyNode::Src) != ac->in(ArrayCopyNode::Dest), "clone source equals destination");
284 Node* base = ac->in(ArrayCopyNode::Src);
285 Node* adr = _igvn.transform(new AddPNode(base, base, MakeConX(offset)));
286 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset);
287 MergeMemNode* mergemen = _igvn.transform(MergeMemNode::make(mem))->as_MergeMem();
288 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
289 res = ArrayCopyNode::load(bs, &_igvn, ctl, mergemen, adr, adr_type, type, bt);
290 } else {
291 if (ac->modifies(offset, offset, &_igvn, true)) {
292 assert(ac->in(ArrayCopyNode::Dest) == alloc->result_cast(), "arraycopy destination should be allocation's result");
293 uint shift = exact_log2(type2aelembytes(bt));
294 Node* src_pos = ac->in(ArrayCopyNode::SrcPos);
295 Node* dest_pos = ac->in(ArrayCopyNode::DestPos);
296 const TypeInt* src_pos_t = _igvn.type(src_pos)->is_int();
297 const TypeInt* dest_pos_t = _igvn.type(dest_pos)->is_int();
298
299 Node* adr = NULL;
300 const TypePtr* adr_type = NULL;
301 if (src_pos_t->is_con() && dest_pos_t->is_con()) {
302 intptr_t off = ((src_pos_t->get_con() - dest_pos_t->get_con()) << shift) + offset;
303 Node* base = ac->in(ArrayCopyNode::Src);
304 adr = _igvn.transform(new AddPNode(base, base, MakeConX(off)));
305 adr_type = _igvn.type(base)->is_ptr()->add_offset(off);
306 if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) {
307 // Don't emit a new load from src if src == dst but try to get the value from memory instead
308 return value_from_mem(ac->in(TypeFunc::Memory), ctl, ft, ftype, adr_type->isa_oopptr(), alloc);
309 }
310 } else {
311 Node* diff = _igvn.transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
312 #ifdef _LP64
313 diff = _igvn.transform(new ConvI2LNode(diff));
314 #endif
315 diff = _igvn.transform(new LShiftXNode(diff, intcon(shift)));
316
317 Node* off = _igvn.transform(new AddXNode(MakeConX(offset), diff));
318 Node* base = ac->in(ArrayCopyNode::Src);
319 adr = _igvn.transform(new AddPNode(base, base, off));
320 adr_type = _igvn.type(base)->is_ptr()->add_offset(Type::OffsetBot);
321 if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) {
322 // Non constant offset in the array: we can't statically
323 // determine the value
324 return NULL;
325 }
326 }
327 MergeMemNode* mergemen = _igvn.transform(MergeMemNode::make(mem))->as_MergeMem();
328 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
329 res = ArrayCopyNode::load(bs, &_igvn, ctl, mergemen, adr, adr_type, type, bt);
330 }
331 }
332 if (res != NULL) {
333 if (ftype->isa_narrowoop()) {
334 // PhaseMacroExpand::scalar_replacement adds DecodeN nodes
335 res = _igvn.transform(new EncodePNode(res, ftype));
336 }
337 return res;
338 }
339 return NULL;
340 }
341
342 //
343 // Given a Memory Phi, compute a value Phi containing the values from stores
344 // on the input paths.
345 // Note: this function is recursive, its depth is limited by the "level" argument
346 // Returns the computed Phi, or NULL if it cannot compute it.
347 Node *PhaseMacroExpand::value_from_mem_phi(Node *mem, BasicType ft, const Type *phi_type, const TypeOopPtr *adr_t, AllocateNode *alloc, Node_Stack *value_phis, int level) {
348 assert(mem->is_Phi(), "sanity");
349 int alias_idx = C->get_alias_index(adr_t);
350 int offset = adr_t->offset();
351 int instance_id = adr_t->instance_id();
352
353 // Check if an appropriate value phi already exists.
354 Node* region = mem->in(0);
355 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) {
356 Node* phi = region->fast_out(k);
357 if (phi->is_Phi() && phi != mem &&
358 phi->as_Phi()->is_same_inst_field(phi_type, (int)mem->_idx, instance_id, alias_idx, offset)) {
359 return phi;
360 }
361 }
362 // Check if an appropriate new value phi already exists.
363 Node* new_phi = value_phis->find(mem->_idx);
364 if (new_phi != NULL)
365 return new_phi;
366
367 if (level <= 0) {
368 return NULL; // Give up: phi tree too deep
369 }
370 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
371 Node *alloc_mem = alloc->in(TypeFunc::Memory);
372
373 uint length = mem->req();
374 GrowableArray <Node *> values(length, length, NULL);
375
376 // create a new Phi for the value
377 PhiNode *phi = new PhiNode(mem->in(0), phi_type, NULL, mem->_idx, instance_id, alias_idx, offset);
378 transform_later(phi);
379 value_phis->push(phi, mem->_idx);
380
381 for (uint j = 1; j < length; j++) {
382 Node *in = mem->in(j);
383 if (in == NULL || in->is_top()) {
384 values.at_put(j, in);
385 } else {
386 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn);
387 if (val == start_mem || val == alloc_mem) {
388 // hit a sentinel, return appropriate 0 value
389 values.at_put(j, _igvn.zerocon(ft));
390 continue;
391 }
392 if (val->is_Initialize()) {
393 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
394 }
395 if (val == NULL) {
396 return NULL; // can't find a value on this path
397 }
398 if (val == mem) {
399 values.at_put(j, mem);
400 } else if (val->is_Store()) {
401 Node* n = val->in(MemNode::ValueIn);
402 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
403 n = bs->step_over_gc_barrier(n);
404 if (is_subword_type(ft)) {
405 n = Compile::narrow_value(ft, n, phi_type, &_igvn, true);
406 }
407 values.at_put(j, n);
408 } else if(val->is_Proj() && val->in(0) == alloc) {
409 values.at_put(j, _igvn.zerocon(ft));
410 } else if (val->is_Phi()) {
411 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1);
412 if (val == NULL) {
413 return NULL;
414 }
415 values.at_put(j, val);
416 } else if (val->Opcode() == Op_SCMemProj) {
417 assert(val->in(0)->is_LoadStore() ||
418 val->in(0)->Opcode() == Op_EncodeISOArray ||
419 val->in(0)->Opcode() == Op_StrCompressedCopy, "sanity");
420 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
421 return NULL;
422 } else if (val->is_ArrayCopy()) {
423 Node* res = make_arraycopy_load(val->as_ArrayCopy(), offset, val->in(0), val->in(TypeFunc::Memory), ft, phi_type, alloc);
424 if (res == NULL) {
425 return NULL;
426 }
427 values.at_put(j, res);
428 } else {
429 DEBUG_ONLY( val->dump(); )
430 assert(false, "unknown node on this path");
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 }
445
446 // Search the last value stored into the object's field.
447 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, Node *sfpt_ctl, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, AllocateNode *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");
451
452 int alias_idx = C->get_alias_index(adr_t);
453 int offset = adr_t->offset();
454 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
455 Node *alloc_ctrl = alloc->in(TypeFunc::Control);
456 Node *alloc_mem = alloc->in(TypeFunc::Memory);
457 VectorSet visited;
458
459 bool done = sfpt_mem == alloc_mem;
460 Node *mem = sfpt_mem;
461 while (!done) {
462 if (visited.test_set(mem->_idx)) {
463 return NULL; // found a loop, give up
464 }
465 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn);
466 if (mem == start_mem || mem == alloc_mem) {
467 done = true; // hit a sentinel, return appropriate 0 value
468 } else if (mem->is_Initialize()) {
469 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
470 if (mem == NULL) {
471 done = true; // Something go wrong.
472 } else if (mem->is_Store()) {
473 const TypePtr* atype = mem->as_Store()->adr_type();
474 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice");
475 done = true;
476 }
477 } else if (mem->is_Store()) {
478 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr();
479 assert(atype != NULL, "address type must be oopptr");
480 assert(C->get_alias_index(atype) == alias_idx &&
481 atype->is_known_instance_field() && atype->offset() == offset &&
482 atype->instance_id() == instance_id, "store is correct memory slice");
483 done = true;
484 } else if (mem->is_Phi()) {
485 // try to find a phi's unique input
486 Node *unique_input = NULL;
487 Node *top = C->top();
488 for (uint i = 1; i < mem->req(); i++) {
489 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn);
490 if (n == NULL || n == top || n == mem) {
491 continue;
492 } else if (unique_input == NULL) {
493 unique_input = n;
494 } else if (unique_input != n) {
495 unique_input = top;
496 break;
497 }
498 }
499 if (unique_input != NULL && unique_input != top) {
500 mem = unique_input;
501 } else {
502 done = true;
503 }
504 } else if (mem->is_ArrayCopy()) {
505 done = true;
506 } else {
507 DEBUG_ONLY( mem->dump(); )
508 assert(false, "unexpected node");
509 }
510 }
511 if (mem != NULL) {
512 if (mem == start_mem || mem == alloc_mem) {
513 // hit a sentinel, return appropriate 0 value
514 return _igvn.zerocon(ft);
515 } else if (mem->is_Store()) {
516 Node* n = mem->in(MemNode::ValueIn);
517 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
518 n = bs->step_over_gc_barrier(n);
519 return n;
520 } else if (mem->is_Phi()) {
521 // attempt to produce a Phi reflecting the values on the input paths of the Phi
522 Node_Stack value_phis(8);
523 Node* phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit);
524 if (phi != NULL) {
525 return phi;
526 } else {
527 // Kill all new Phis
528 while(value_phis.is_nonempty()) {
529 Node* n = value_phis.node();
530 _igvn.replace_node(n, C->top());
531 value_phis.pop();
532 }
533 }
534 } else if (mem->is_ArrayCopy()) {
535 Node* ctl = mem->in(0);
536 Node* m = mem->in(TypeFunc::Memory);
537 if (sfpt_ctl->is_Proj() && sfpt_ctl->as_Proj()->is_uncommon_trap_proj(Deoptimization::Reason_none)) {
538 // pin the loads in the uncommon trap path
539 ctl = sfpt_ctl;
540 m = sfpt_mem;
541 }
542 return make_arraycopy_load(mem->as_ArrayCopy(), offset, ctl, m, ft, ftype, alloc);
543 }
544 }
545 // Something go wrong.
546 return NULL;
547 }
548
549 // Check the possibility of scalar replacement.
550 bool PhaseMacroExpand::can_eliminate_allocation(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
551 // Scan the uses of the allocation to check for anything that would
552 // prevent us from eliminating it.
553 NOT_PRODUCT( const char* fail_eliminate = NULL; )
554 DEBUG_ONLY( Node* disq_node = NULL; )
555 bool can_eliminate = true;
556
557 Node* res = alloc->result_cast();
558 const TypeOopPtr* res_type = NULL;
559 if (res == NULL) {
560 // All users were eliminated.
561 } else if (!res->is_CheckCastPP()) {
562 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";)
563 can_eliminate = false;
564 } else {
565 res_type = _igvn.type(res)->isa_oopptr();
566 if (res_type == NULL) {
567 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";)
568 can_eliminate = false;
569 } else if (res_type->isa_aryptr()) {
570 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1);
571 if (length < 0) {
572 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";)
573 can_eliminate = false;
574 }
575 }
576 }
577
578 if (can_eliminate && res != NULL) {
579 for (DUIterator_Fast jmax, j = res->fast_outs(jmax);
580 j < jmax && can_eliminate; j++) {
581 Node* use = res->fast_out(j);
582
583 if (use->is_AddP()) {
584 const TypePtr* addp_type = _igvn.type(use)->is_ptr();
585 int offset = addp_type->offset();
586
587 if (offset == Type::OffsetTop || offset == Type::OffsetBot) {
588 NOT_PRODUCT(fail_eliminate = "Undefined field reference";)
589 can_eliminate = false;
590 break;
591 }
592 for (DUIterator_Fast kmax, k = use->fast_outs(kmax);
593 k < kmax && can_eliminate; k++) {
594 Node* n = use->fast_out(k);
595 if (!n->is_Store() && n->Opcode() != Op_CastP2X
596 SHENANDOAHGC_ONLY(&& (!UseShenandoahGC || !ShenandoahBarrierSetC2::is_shenandoah_wb_pre_call(n))) ) {
597 DEBUG_ONLY(disq_node = n;)
598 if (n->is_Load() || n->is_LoadStore()) {
599 NOT_PRODUCT(fail_eliminate = "Field load";)
600 } else {
601 NOT_PRODUCT(fail_eliminate = "Not store field reference";)
602 }
603 can_eliminate = false;
604 }
605 }
606 } else if (use->is_ArrayCopy() &&
607 (use->as_ArrayCopy()->is_clonebasic() ||
608 use->as_ArrayCopy()->is_arraycopy_validated() ||
609 use->as_ArrayCopy()->is_copyof_validated() ||
610 use->as_ArrayCopy()->is_copyofrange_validated()) &&
611 use->in(ArrayCopyNode::Dest) == res) {
612 // ok to eliminate
613 } else if (use->is_SafePoint()) {
614 SafePointNode* sfpt = use->as_SafePoint();
615 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) {
616 // Object is passed as argument.
617 DEBUG_ONLY(disq_node = use;)
618 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";)
619 can_eliminate = false;
620 }
621 Node* sfptMem = sfpt->memory();
622 if (sfptMem == NULL || sfptMem->is_top()) {
623 DEBUG_ONLY(disq_node = use;)
624 NOT_PRODUCT(fail_eliminate = "NULL or TOP memory";)
625 can_eliminate = false;
626 } else {
627 safepoints.append_if_missing(sfpt);
628 }
629 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
630 if (use->is_Phi()) {
631 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
632 NOT_PRODUCT(fail_eliminate = "Object is return value";)
633 } else {
634 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
635 }
636 DEBUG_ONLY(disq_node = use;)
637 } else {
638 if (use->Opcode() == Op_Return) {
639 NOT_PRODUCT(fail_eliminate = "Object is return value";)
640 }else {
641 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
642 }
643 DEBUG_ONLY(disq_node = use;)
644 }
645 can_eliminate = false;
646 }
647 }
648 }
649
650 #ifndef PRODUCT
651 if (PrintEliminateAllocations) {
652 if (can_eliminate) {
653 tty->print("Scalar ");
654 if (res == NULL)
655 alloc->dump();
656 else
657 res->dump();
658 } else if (alloc->_is_scalar_replaceable) {
659 tty->print("NotScalar (%s)", fail_eliminate);
660 if (res == NULL)
661 alloc->dump();
662 else
663 res->dump();
664 #ifdef ASSERT
665 if (disq_node != NULL) {
666 tty->print(" >>>> ");
667 disq_node->dump();
668 }
669 #endif /*ASSERT*/
670 }
671 }
672 #endif
673 return can_eliminate;
674 }
675
676 // Do scalar replacement.
677 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
678 GrowableArray <SafePointNode *> safepoints_done;
679
680 ciKlass* klass = NULL;
681 ciInstanceKlass* iklass = NULL;
682 int nfields = 0;
683 int array_base = 0;
684 int element_size = 0;
685 BasicType basic_elem_type = T_ILLEGAL;
686 ciType* elem_type = NULL;
687
688 Node* res = alloc->result_cast();
689 assert(res == NULL || res->is_CheckCastPP(), "unexpected AllocateNode result");
690 const TypeOopPtr* res_type = NULL;
691 if (res != NULL) { // Could be NULL when there are no users
692 res_type = _igvn.type(res)->isa_oopptr();
693 }
694
695 if (res != NULL) {
696 klass = res_type->klass();
697 if (res_type->isa_instptr()) {
698 // find the fields of the class which will be needed for safepoint debug information
699 assert(klass->is_instance_klass(), "must be an instance klass.");
700 iklass = klass->as_instance_klass();
701 nfields = iklass->nof_nonstatic_fields();
702 } else {
703 // find the array's elements which will be needed for safepoint debug information
704 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
705 assert(klass->is_array_klass() && nfields >= 0, "must be an array klass.");
706 elem_type = klass->as_array_klass()->element_type();
707 basic_elem_type = elem_type->basic_type();
708 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
709 element_size = type2aelembytes(basic_elem_type);
710 }
711 }
712 //
713 // Process the safepoint uses
714 //
715 while (safepoints.length() > 0) {
716 SafePointNode* sfpt = safepoints.pop();
717 Node* mem = sfpt->memory();
718 Node* ctl = sfpt->control();
719 assert(sfpt->jvms() != NULL, "missed JVMS");
720 // Fields of scalar objs are referenced only at the end
721 // of regular debuginfo at the last (youngest) JVMS.
722 // Record relative start index.
723 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff());
724 SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type,
725 #ifdef ASSERT
726 alloc,
727 #endif
728 first_ind, nfields);
729 sobj->init_req(0, C->root());
730 transform_later(sobj);
731
732 // Scan object's fields adding an input to the safepoint for each field.
733 for (int j = 0; j < nfields; j++) {
734 intptr_t offset;
735 ciField* field = NULL;
736 if (iklass != NULL) {
737 field = iklass->nonstatic_field_at(j);
738 offset = field->offset();
739 elem_type = field->type();
740 basic_elem_type = field->layout_type();
741 } else {
742 offset = array_base + j * (intptr_t)element_size;
743 }
744
745 const Type *field_type;
746 // The next code is taken from Parse::do_get_xxx().
747 if (is_reference_type(basic_elem_type)) {
748 if (!elem_type->is_loaded()) {
749 field_type = TypeInstPtr::BOTTOM;
750 } else if (field != NULL && field->is_static_constant()) {
751 // This can happen if the constant oop is non-perm.
752 ciObject* con = field->constant_value().as_object();
753 // Do not "join" in the previous type; it doesn't add value,
754 // and may yield a vacuous result if the field is of interface type.
755 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr();
756 assert(field_type != NULL, "field singleton type must be consistent");
757 } else {
758 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
759 }
760 if (UseCompressedOops) {
761 field_type = field_type->make_narrowoop();
762 basic_elem_type = T_NARROWOOP;
763 }
764 } else {
765 field_type = Type::get_const_basic_type(basic_elem_type);
766 }
767
768 const TypeOopPtr *field_addr_type = res_type->add_offset(offset)->isa_oopptr();
769
770 Node *field_val = value_from_mem(mem, ctl, basic_elem_type, field_type, field_addr_type, alloc);
771 if (field_val == NULL) {
772 // We weren't able to find a value for this field,
773 // give up on eliminating this allocation.
774
775 // Remove any extra entries we added to the safepoint.
776 uint last = sfpt->req() - 1;
777 for (int k = 0; k < j; k++) {
778 sfpt->del_req(last--);
779 }
780 _igvn._worklist.push(sfpt);
781 // rollback processed safepoints
782 while (safepoints_done.length() > 0) {
783 SafePointNode* sfpt_done = safepoints_done.pop();
784 // remove any extra entries we added to the safepoint
785 last = sfpt_done->req() - 1;
786 for (int k = 0; k < nfields; k++) {
787 sfpt_done->del_req(last--);
788 }
789 JVMState *jvms = sfpt_done->jvms();
790 jvms->set_endoff(sfpt_done->req());
791 // Now make a pass over the debug information replacing any references
792 // to SafePointScalarObjectNode with the allocated object.
793 int start = jvms->debug_start();
794 int end = jvms->debug_end();
795 for (int i = start; i < end; i++) {
796 if (sfpt_done->in(i)->is_SafePointScalarObject()) {
797 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject();
798 if (scobj->first_index(jvms) == sfpt_done->req() &&
799 scobj->n_fields() == (uint)nfields) {
800 assert(scobj->alloc() == alloc, "sanity");
801 sfpt_done->set_req(i, res);
802 }
803 }
804 }
805 _igvn._worklist.push(sfpt_done);
806 }
807 #ifndef PRODUCT
808 if (PrintEliminateAllocations) {
809 if (field != NULL) {
810 tty->print("=== At SafePoint node %d can't find value of Field: ",
811 sfpt->_idx);
812 field->print();
813 int field_idx = C->get_alias_index(field_addr_type);
814 tty->print(" (alias_idx=%d)", field_idx);
815 } else { // Array's element
816 tty->print("=== At SafePoint node %d can't find value of array element [%d]",
817 sfpt->_idx, j);
818 }
819 tty->print(", which prevents elimination of: ");
820 if (res == NULL)
821 alloc->dump();
822 else
823 res->dump();
824 }
825 #endif
826 return false;
827 }
828 if (UseCompressedOops && field_type->isa_narrowoop()) {
829 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
830 // to be able scalar replace the allocation.
831 if (field_val->is_EncodeP()) {
832 field_val = field_val->in(1);
833 } else {
834 field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type()));
835 }
836 }
837 sfpt->add_req(field_val);
838 }
839 JVMState *jvms = sfpt->jvms();
840 jvms->set_endoff(sfpt->req());
841 // Now make a pass over the debug information replacing any references
842 // to the allocated object with "sobj"
843 int start = jvms->debug_start();
844 int end = jvms->debug_end();
845 sfpt->replace_edges_in_range(res, sobj, start, end, &_igvn);
846 _igvn._worklist.push(sfpt);
847 safepoints_done.append_if_missing(sfpt); // keep it for rollback
848 }
849 return true;
850 }
851
852 static void disconnect_projections(MultiNode* n, PhaseIterGVN& igvn) {
853 Node* ctl_proj = n->proj_out_or_null(TypeFunc::Control);
854 Node* mem_proj = n->proj_out_or_null(TypeFunc::Memory);
855 if (ctl_proj != NULL) {
856 igvn.replace_node(ctl_proj, n->in(0));
857 }
858 if (mem_proj != NULL) {
859 igvn.replace_node(mem_proj, n->in(TypeFunc::Memory));
860 }
861 }
862
863 // Process users of eliminated allocation.
864 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc) {
865 Node* res = alloc->result_cast();
866 if (res != NULL) {
867 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
868 Node *use = res->last_out(j);
869 uint oc1 = res->outcnt();
870
871 if (use->is_AddP()) {
872 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) {
873 Node *n = use->last_out(k);
874 uint oc2 = use->outcnt();
875 if (n->is_Store()) {
876 #ifdef ASSERT
877 // Verify that there is no dependent MemBarVolatile nodes,
878 // they should be removed during IGVN, see MemBarNode::Ideal().
879 for (DUIterator_Fast pmax, p = n->fast_outs(pmax);
880 p < pmax; p++) {
881 Node* mb = n->fast_out(p);
882 assert(mb->is_Initialize() || !mb->is_MemBar() ||
883 mb->req() <= MemBarNode::Precedent ||
884 mb->in(MemBarNode::Precedent) != n,
885 "MemBarVolatile should be eliminated for non-escaping object");
886 }
887 #endif
888 _igvn.replace_node(n, n->in(MemNode::Memory));
889 } else {
890 eliminate_gc_barrier(n);
891 }
892 k -= (oc2 - use->outcnt());
893 }
894 _igvn.remove_dead_node(use);
895 } else if (use->is_ArrayCopy()) {
896 // Disconnect ArrayCopy node
897 ArrayCopyNode* ac = use->as_ArrayCopy();
898 if (ac->is_clonebasic()) {
899 Node* membar_after = ac->proj_out(TypeFunc::Control)->unique_ctrl_out();
900 disconnect_projections(ac, _igvn);
901 assert(alloc->in(TypeFunc::Memory)->is_Proj() && alloc->in(TypeFunc::Memory)->in(0)->Opcode() == Op_MemBarCPUOrder, "mem barrier expected before allocation");
902 Node* membar_before = alloc->in(TypeFunc::Memory)->in(0);
903 disconnect_projections(membar_before->as_MemBar(), _igvn);
904 if (membar_after->is_MemBar()) {
905 disconnect_projections(membar_after->as_MemBar(), _igvn);
906 }
907 } else {
908 assert(ac->is_arraycopy_validated() ||
909 ac->is_copyof_validated() ||
910 ac->is_copyofrange_validated(), "unsupported");
911 CallProjections callprojs;
912 ac->extract_projections(&callprojs, true);
913
914 _igvn.replace_node(callprojs.fallthrough_ioproj, ac->in(TypeFunc::I_O));
915 _igvn.replace_node(callprojs.fallthrough_memproj, ac->in(TypeFunc::Memory));
916 _igvn.replace_node(callprojs.fallthrough_catchproj, ac->in(TypeFunc::Control));
917
918 // Set control to top. IGVN will remove the remaining projections
919 ac->set_req(0, top());
920 ac->replace_edge(res, top(), &_igvn);
921
922 // Disconnect src right away: it can help find new
923 // opportunities for allocation elimination
924 Node* src = ac->in(ArrayCopyNode::Src);
925 ac->replace_edge(src, top(), &_igvn);
926 // src can be top at this point if src and dest of the
927 // arraycopy were the same
928 if (src->outcnt() == 0 && !src->is_top()) {
929 _igvn.remove_dead_node(src);
930 }
931 }
932 _igvn._worklist.push(ac);
933 } else {
934 eliminate_gc_barrier(use);
935 }
936 j -= (oc1 - res->outcnt());
937 }
938 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
939 _igvn.remove_dead_node(res);
940 }
941
942 //
943 // Process other users of allocation's projections
944 //
945 if (_callprojs.resproj != NULL && _callprojs.resproj->outcnt() != 0) {
946 // First disconnect stores captured by Initialize node.
947 // If Initialize node is eliminated first in the following code,
948 // it will kill such stores and DUIterator_Last will assert.
949 for (DUIterator_Fast jmax, j = _callprojs.resproj->fast_outs(jmax); j < jmax; j++) {
950 Node* use = _callprojs.resproj->fast_out(j);
951 if (use->is_AddP()) {
952 // raw memory addresses used only by the initialization
953 _igvn.replace_node(use, C->top());
954 --j; --jmax;
955 }
956 }
957 for (DUIterator_Last jmin, j = _callprojs.resproj->last_outs(jmin); j >= jmin; ) {
958 Node* use = _callprojs.resproj->last_out(j);
959 uint oc1 = _callprojs.resproj->outcnt();
960 if (use->is_Initialize()) {
961 // Eliminate Initialize node.
962 InitializeNode *init = use->as_Initialize();
963 assert(init->outcnt() <= 2, "only a control and memory projection expected");
964 Node *ctrl_proj = init->proj_out_or_null(TypeFunc::Control);
965 if (ctrl_proj != NULL) {
966 _igvn.replace_node(ctrl_proj, init->in(TypeFunc::Control));
967 #ifdef ASSERT
968 // If the InitializeNode has no memory out, it will die, and tmp will become NULL
969 Node* tmp = init->in(TypeFunc::Control);
970 assert(tmp == NULL || tmp == _callprojs.fallthrough_catchproj, "allocation control projection");
971 #endif
972 }
973 Node *mem_proj = init->proj_out_or_null(TypeFunc::Memory);
974 if (mem_proj != NULL) {
975 Node *mem = init->in(TypeFunc::Memory);
976 #ifdef ASSERT
977 if (mem->is_MergeMem()) {
978 assert(mem->in(TypeFunc::Memory) == _callprojs.fallthrough_memproj, "allocation memory projection");
979 } else {
980 assert(mem == _callprojs.fallthrough_memproj, "allocation memory projection");
981 }
982 #endif
983 _igvn.replace_node(mem_proj, mem);
984 }
985 } else {
986 assert(false, "only Initialize or AddP expected");
987 }
988 j -= (oc1 - _callprojs.resproj->outcnt());
989 }
990 }
991 if (_callprojs.fallthrough_catchproj != NULL) {
992 _igvn.replace_node(_callprojs.fallthrough_catchproj, alloc->in(TypeFunc::Control));
993 }
994 if (_callprojs.fallthrough_memproj != NULL) {
995 _igvn.replace_node(_callprojs.fallthrough_memproj, alloc->in(TypeFunc::Memory));
996 }
997 if (_callprojs.catchall_memproj != NULL) {
998 _igvn.replace_node(_callprojs.catchall_memproj, C->top());
999 }
1000 if (_callprojs.fallthrough_ioproj != NULL) {
1001 _igvn.replace_node(_callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O));
1002 }
1003 if (_callprojs.catchall_ioproj != NULL) {
1004 _igvn.replace_node(_callprojs.catchall_ioproj, C->top());
1005 }
1006 if (_callprojs.catchall_catchproj != NULL) {
1007 _igvn.replace_node(_callprojs.catchall_catchproj, C->top());
1008 }
1009 }
1010
1011 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) {
1012 // If reallocation fails during deoptimization we'll pop all
1013 // interpreter frames for this compiled frame and that won't play
1014 // nice with JVMTI popframe.
1015 // We avoid this issue by eager reallocation when the popframe request
1016 // is received.
1017 if (!EliminateAllocations || !alloc->_is_non_escaping) {
1018 return false;
1019 }
1020 Node* klass = alloc->in(AllocateNode::KlassNode);
1021 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr();
1022 Node* res = alloc->result_cast();
1023 // Eliminate boxing allocations which are not used
1024 // regardless scalar replaceable status.
1025 bool boxing_alloc = C->eliminate_boxing() &&
1026 tklass->klass()->is_instance_klass() &&
1027 tklass->klass()->as_instance_klass()->is_box_klass();
1028 if (!alloc->_is_scalar_replaceable && (!boxing_alloc || (res != NULL))) {
1029 return false;
1030 }
1031
1032 alloc->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
1033
1034 GrowableArray <SafePointNode *> safepoints;
1035 if (!can_eliminate_allocation(alloc, safepoints)) {
1036 return false;
1037 }
1038
1039 if (!alloc->_is_scalar_replaceable) {
1040 assert(res == NULL, "sanity");
1041 // We can only eliminate allocation if all debug info references
1042 // are already replaced with SafePointScalarObject because
1043 // we can't search for a fields value without instance_id.
1044 if (safepoints.length() > 0) {
1045 return false;
1046 }
1047 }
1048
1049 if (!scalar_replacement(alloc, safepoints)) {
1050 return false;
1051 }
1052
1053 CompileLog* log = C->log();
1054 if (log != NULL) {
1055 log->head("eliminate_allocation type='%d'",
1056 log->identify(tklass->klass()));
1057 JVMState* p = alloc->jvms();
1058 while (p != NULL) {
1059 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1060 p = p->caller();
1061 }
1062 log->tail("eliminate_allocation");
1063 }
1064
1065 process_users_of_allocation(alloc);
1066
1067 #ifndef PRODUCT
1068 if (PrintEliminateAllocations) {
1069 if (alloc->is_AllocateArray())
1070 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1071 else
1072 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1073 }
1074 #endif
1075
1076 return true;
1077 }
1078
1079 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) {
1080 // EA should remove all uses of non-escaping boxing node.
1081 if (!C->eliminate_boxing() || boxing->proj_out_or_null(TypeFunc::Parms) != NULL) {
1082 return false;
1083 }
1084
1085 assert(boxing->result_cast() == NULL, "unexpected boxing node result");
1086
1087 boxing->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
1088
1089 const TypeTuple* r = boxing->tf()->range();
1090 assert(r->cnt() > TypeFunc::Parms, "sanity");
1091 const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr();
1092 assert(t != NULL, "sanity");
1093
1094 CompileLog* log = C->log();
1095 if (log != NULL) {
1096 log->head("eliminate_boxing type='%d'",
1097 log->identify(t->klass()));
1098 JVMState* p = boxing->jvms();
1099 while (p != NULL) {
1100 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1101 p = p->caller();
1102 }
1103 log->tail("eliminate_boxing");
1104 }
1105
1106 process_users_of_allocation(boxing);
1107
1108 #ifndef PRODUCT
1109 if (PrintEliminateAllocations) {
1110 tty->print("++++ Eliminated: %d ", boxing->_idx);
1111 boxing->method()->print_short_name(tty);
1112 tty->cr();
1113 }
1114 #endif
1115
1116 return true;
1117 }
1118
1119 //---------------------------set_eden_pointers-------------------------
1120 void PhaseMacroExpand::set_eden_pointers(Node* &eden_top_adr, Node* &eden_end_adr) {
1121 if (UseTLAB) { // Private allocation: load from TLS
1122 Node* thread = transform_later(new ThreadLocalNode());
1123 int tlab_top_offset = in_bytes(JavaThread::tlab_top_offset());
1124 int tlab_end_offset = in_bytes(JavaThread::tlab_end_offset());
1125 eden_top_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_top_offset);
1126 eden_end_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_end_offset);
1127 } else { // Shared allocation: load from globals
1128 CollectedHeap* ch = Universe::heap();
1129 address top_adr = (address)ch->top_addr();
1130 address end_adr = (address)ch->end_addr();
1131 eden_top_adr = makecon(TypeRawPtr::make(top_adr));
1132 eden_end_adr = basic_plus_adr(eden_top_adr, end_adr - top_adr);
1133 }
1134 }
1135
1136
1137 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) {
1138 Node* adr = basic_plus_adr(base, offset);
1139 const TypePtr* adr_type = adr->bottom_type()->is_ptr();
1140 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt, MemNode::unordered);
1141 transform_later(value);
1142 return value;
1143 }
1144
1145
1146 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) {
1147 Node* adr = basic_plus_adr(base, offset);
1148 mem = StoreNode::make(_igvn, ctl, mem, adr, NULL, value, bt, MemNode::unordered);
1149 transform_later(mem);
1150 return mem;
1151 }
1152
1153 //=============================================================================
1154 //
1155 // A L L O C A T I O N
1156 //
1157 // Allocation attempts to be fast in the case of frequent small objects.
1158 // It breaks down like this:
1159 //
1160 // 1) Size in doublewords is computed. This is a constant for objects and
1161 // variable for most arrays. Doubleword units are used to avoid size
1162 // overflow of huge doubleword arrays. We need doublewords in the end for
1163 // rounding.
1164 //
1165 // 2) Size is checked for being 'too large'. Too-large allocations will go
1166 // the slow path into the VM. The slow path can throw any required
1167 // exceptions, and does all the special checks for very large arrays. The
1168 // size test can constant-fold away for objects. For objects with
1169 // finalizers it constant-folds the otherway: you always go slow with
1170 // finalizers.
1171 //
1172 // 3) If NOT using TLABs, this is the contended loop-back point.
1173 // Load-Locked the heap top. If using TLABs normal-load the heap top.
1174 //
1175 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route.
1176 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish
1177 // "size*8" we always enter the VM, where "largish" is a constant picked small
1178 // enough that there's always space between the eden max and 4Gig (old space is
1179 // there so it's quite large) and large enough that the cost of entering the VM
1180 // is dwarfed by the cost to initialize the space.
1181 //
1182 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back
1183 // down. If contended, repeat at step 3. If using TLABs normal-store
1184 // adjusted heap top back down; there is no contention.
1185 //
1186 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark
1187 // fields.
1188 //
1189 // 7) Merge with the slow-path; cast the raw memory pointer to the correct
1190 // oop flavor.
1191 //
1192 //=============================================================================
1193 // FastAllocateSizeLimit value is in DOUBLEWORDS.
1194 // Allocations bigger than this always go the slow route.
1195 // This value must be small enough that allocation attempts that need to
1196 // trigger exceptions go the slow route. Also, it must be small enough so
1197 // that heap_top + size_in_bytes does not wrap around the 4Gig limit.
1198 //=============================================================================j//
1199 // %%% Here is an old comment from parseHelper.cpp; is it outdated?
1200 // The allocator will coalesce int->oop copies away. See comment in
1201 // coalesce.cpp about how this works. It depends critically on the exact
1202 // code shape produced here, so if you are changing this code shape
1203 // make sure the GC info for the heap-top is correct in and around the
1204 // slow-path call.
1205 //
1206
1207 void PhaseMacroExpand::expand_allocate_common(
1208 AllocateNode* alloc, // allocation node to be expanded
1209 Node* length, // array length for an array allocation
1210 const TypeFunc* slow_call_type, // Type of slow call
1211 address slow_call_address, // Address of slow call
1212 Node* valid_length_test // whether length is valid or not
1213 )
1214 {
1215 Node* ctrl = alloc->in(TypeFunc::Control);
1216 Node* mem = alloc->in(TypeFunc::Memory);
1217 Node* i_o = alloc->in(TypeFunc::I_O);
1218 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize);
1219 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1220 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest);
1221 assert(ctrl != NULL, "must have control");
1222
1223 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results.
1224 // they will not be used if "always_slow" is set
1225 enum { slow_result_path = 1, fast_result_path = 2 };
1226 Node *result_region = NULL;
1227 Node *result_phi_rawmem = NULL;
1228 Node *result_phi_rawoop = NULL;
1229 Node *result_phi_i_o = NULL;
1230
1231 // The initial slow comparison is a size check, the comparison
1232 // we want to do is a BoolTest::gt
1233 bool expand_fast_path = true;
1234 int tv = _igvn.find_int_con(initial_slow_test, -1);
1235 if (tv >= 0) {
1236 // InitialTest has constant result
1237 // 0 - can fit in TLAB
1238 // 1 - always too big or negative
1239 assert(tv <= 1, "0 or 1 if a constant");
1240 expand_fast_path = (tv == 0);
1241 initial_slow_test = NULL;
1242 } else {
1243 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn);
1244 }
1245
1246 if (!UseTLAB && !Universe::heap()->supports_inline_contig_alloc()) {
1247 // Force slow-path allocation
1248 expand_fast_path = false;
1249 initial_slow_test = NULL;
1250 }
1251
1252 bool allocation_has_use = (alloc->result_cast() != NULL);
1253 if (!allocation_has_use) {
1254 InitializeNode* init = alloc->initialization();
1255 if (init != NULL) {
1256 init->remove(&_igvn);
1257 }
1258 if (expand_fast_path && (initial_slow_test == NULL)) {
1259 // Remove allocation node and return.
1260 // Size is a non-negative constant -> no initial check needed -> directly to fast path.
1261 // Also, no usages -> empty fast path -> no fall out to slow path -> nothing left.
1262 #ifndef PRODUCT
1263 if (PrintEliminateAllocations) {
1264 tty->print("NotUsed ");
1265 Node* res = alloc->proj_out_or_null(TypeFunc::Parms);
1266 if (res != NULL) {
1267 res->dump();
1268 } else {
1269 alloc->dump();
1270 }
1271 }
1272 #endif
1273 yank_alloc_node(alloc);
1274 return;
1275 }
1276 }
1277
1278 enum { too_big_or_final_path = 1, need_gc_path = 2 };
1279 Node *slow_region = NULL;
1280 Node *toobig_false = ctrl;
1281
1282 // generate the initial test if necessary
1283 if (initial_slow_test != NULL ) {
1284 assert (expand_fast_path, "Only need test if there is a fast path");
1285 slow_region = new RegionNode(3);
1286
1287 // Now make the initial failure test. Usually a too-big test but
1288 // might be a TRUE for finalizers or a fancy class check for
1289 // newInstance0.
1290 IfNode *toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
1291 transform_later(toobig_iff);
1292 // Plug the failing-too-big test into the slow-path region
1293 Node *toobig_true = new IfTrueNode( toobig_iff );
1294 transform_later(toobig_true);
1295 slow_region ->init_req( too_big_or_final_path, toobig_true );
1296 toobig_false = new IfFalseNode( toobig_iff );
1297 transform_later(toobig_false);
1298 } else {
1299 // No initial test, just fall into next case
1300 assert(allocation_has_use || !expand_fast_path, "Should already have been handled");
1301 toobig_false = ctrl;
1302 debug_only(slow_region = NodeSentinel);
1303 }
1304
1305 // If we are here there are several possibilities
1306 // - expand_fast_path is false - then only a slow path is expanded. That's it.
1307 // no_initial_check means a constant allocation.
1308 // - If check always evaluates to false -> expand_fast_path is false (see above)
1309 // - If check always evaluates to true -> directly into fast path (but may bailout to slowpath)
1310 // if !allocation_has_use the fast path is empty
1311 // if !allocation_has_use && no_initial_check
1312 // - Then there are no fastpath that can fall out to slowpath -> no allocation code at all.
1313 // removed by yank_alloc_node above.
1314
1315 Node *slow_mem = mem; // save the current memory state for slow path
1316 // generate the fast allocation code unless we know that the initial test will always go slow
1317 if (expand_fast_path) {
1318 // Fast path modifies only raw memory.
1319 if (mem->is_MergeMem()) {
1320 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1321 }
1322
1323 // allocate the Region and Phi nodes for the result
1324 result_region = new RegionNode(3);
1325 result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1326 result_phi_i_o = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch
1327
1328 // Grab regular I/O before optional prefetch may change it.
1329 // Slow-path does no I/O so just set it to the original I/O.
1330 result_phi_i_o->init_req(slow_result_path, i_o);
1331
1332 // Name successful fast-path variables
1333 Node* fast_oop_ctrl;
1334 Node* fast_oop_rawmem;
1335 if (allocation_has_use) {
1336 Node* needgc_ctrl = NULL;
1337 result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM);
1338
1339 intx prefetch_lines = length != NULL ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1340 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1341 Node* fast_oop = bs->obj_allocate(this, mem, toobig_false, size_in_bytes, i_o, needgc_ctrl,
1342 fast_oop_ctrl, fast_oop_rawmem,
1343 prefetch_lines);
1344
1345 if (initial_slow_test != NULL) {
1346 // This completes all paths into the slow merge point
1347 slow_region->init_req(need_gc_path, needgc_ctrl);
1348 transform_later(slow_region);
1349 } else {
1350 // No initial slow path needed!
1351 // Just fall from the need-GC path straight into the VM call.
1352 slow_region = needgc_ctrl;
1353 }
1354
1355 InitializeNode* init = alloc->initialization();
1356 fast_oop_rawmem = initialize_object(alloc,
1357 fast_oop_ctrl, fast_oop_rawmem, fast_oop,
1358 klass_node, length, size_in_bytes);
1359 expand_initialize_membar(alloc, init, fast_oop_ctrl, fast_oop_rawmem);
1360 expand_dtrace_alloc_probe(alloc, fast_oop, fast_oop_ctrl, fast_oop_rawmem);
1361
1362 result_phi_rawoop->init_req(fast_result_path, fast_oop);
1363 } else {
1364 assert (initial_slow_test != NULL, "sanity");
1365 fast_oop_ctrl = toobig_false;
1366 fast_oop_rawmem = mem;
1367 transform_later(slow_region);
1368 }
1369
1370 // Plug in the successful fast-path into the result merge point
1371 result_region ->init_req(fast_result_path, fast_oop_ctrl);
1372 result_phi_i_o ->init_req(fast_result_path, i_o);
1373 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem);
1374 } else {
1375 slow_region = ctrl;
1376 result_phi_i_o = i_o; // Rename it to use in the following code.
1377 }
1378
1379 // Generate slow-path call
1380 CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address,
1381 OptoRuntime::stub_name(slow_call_address),
1382 TypePtr::BOTTOM);
1383 call->init_req(TypeFunc::Control, slow_region);
1384 call->init_req(TypeFunc::I_O, top()); // does no i/o
1385 call->init_req(TypeFunc::Memory, slow_mem); // may gc ptrs
1386 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1387 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1388
1389 call->init_req(TypeFunc::Parms+0, klass_node);
1390 if (length != NULL) {
1391 call->init_req(TypeFunc::Parms+1, length);
1392 }
1393
1394 // Copy debug information and adjust JVMState information, then replace
1395 // allocate node with the call
1396 call->copy_call_debug_info(&_igvn, alloc);
1397 // For array allocations, copy the valid length check to the call node so Compile::final_graph_reshaping() can verify
1398 // that the call has the expected number of CatchProj nodes (in case the allocation always fails and the fallthrough
1399 // path dies).
1400 if (valid_length_test != NULL) {
1401 call->add_req(valid_length_test);
1402 }
1403 if (expand_fast_path) {
1404 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1405 } else {
1406 // Hook i_o projection to avoid its elimination during allocation
1407 // replacement (when only a slow call is generated).
1408 call->set_req(TypeFunc::I_O, result_phi_i_o);
1409 }
1410 _igvn.replace_node(alloc, call);
1411 transform_later(call);
1412
1413 // Identify the output projections from the allocate node and
1414 // adjust any references to them.
1415 // The control and io projections look like:
1416 //
1417 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl)
1418 // Allocate Catch
1419 // ^---Proj(io) <-------+ ^---CatchProj(io)
1420 //
1421 // We are interested in the CatchProj nodes.
1422 //
1423 call->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
1424
1425 // An allocate node has separate memory projections for the uses on
1426 // the control and i_o paths. Replace the control memory projection with
1427 // result_phi_rawmem (unless we are only generating a slow call when
1428 // both memory projections are combined)
1429 if (expand_fast_path && _callprojs.fallthrough_memproj != NULL) {
1430 migrate_outs(_callprojs.fallthrough_memproj, result_phi_rawmem);
1431 }
1432 // Now change uses of catchall_memproj to use fallthrough_memproj and delete
1433 // catchall_memproj so we end up with a call that has only 1 memory projection.
1434 if (_callprojs.catchall_memproj != NULL ) {
1435 if (_callprojs.fallthrough_memproj == NULL) {
1436 _callprojs.fallthrough_memproj = new ProjNode(call, TypeFunc::Memory);
1437 transform_later(_callprojs.fallthrough_memproj);
1438 }
1439 migrate_outs(_callprojs.catchall_memproj, _callprojs.fallthrough_memproj);
1440 _igvn.remove_dead_node(_callprojs.catchall_memproj);
1441 }
1442
1443 // An allocate node has separate i_o projections for the uses on the control
1444 // and i_o paths. Always replace the control i_o projection with result i_o
1445 // otherwise incoming i_o become dead when only a slow call is generated
1446 // (it is different from memory projections where both projections are
1447 // combined in such case).
1448 if (_callprojs.fallthrough_ioproj != NULL) {
1449 migrate_outs(_callprojs.fallthrough_ioproj, result_phi_i_o);
1450 }
1451 // Now change uses of catchall_ioproj to use fallthrough_ioproj and delete
1452 // catchall_ioproj so we end up with a call that has only 1 i_o projection.
1453 if (_callprojs.catchall_ioproj != NULL ) {
1454 if (_callprojs.fallthrough_ioproj == NULL) {
1455 _callprojs.fallthrough_ioproj = new ProjNode(call, TypeFunc::I_O);
1456 transform_later(_callprojs.fallthrough_ioproj);
1457 }
1458 migrate_outs(_callprojs.catchall_ioproj, _callprojs.fallthrough_ioproj);
1459 _igvn.remove_dead_node(_callprojs.catchall_ioproj);
1460 }
1461
1462 // if we generated only a slow call, we are done
1463 if (!expand_fast_path) {
1464 // Now we can unhook i_o.
1465 if (result_phi_i_o->outcnt() > 1) {
1466 call->set_req(TypeFunc::I_O, top());
1467 } else {
1468 assert(result_phi_i_o->unique_ctrl_out() == call, "sanity");
1469 // Case of new array with negative size known during compilation.
1470 // AllocateArrayNode::Ideal() optimization disconnect unreachable
1471 // following code since call to runtime will throw exception.
1472 // As result there will be no users of i_o after the call.
1473 // Leave i_o attached to this call to avoid problems in preceding graph.
1474 }
1475 return;
1476 }
1477
1478 if (_callprojs.fallthrough_catchproj != NULL) {
1479 ctrl = _callprojs.fallthrough_catchproj->clone();
1480 transform_later(ctrl);
1481 _igvn.replace_node(_callprojs.fallthrough_catchproj, result_region);
1482 } else {
1483 ctrl = top();
1484 }
1485 Node *slow_result;
1486 if (_callprojs.resproj == NULL) {
1487 // no uses of the allocation result
1488 slow_result = top();
1489 } else {
1490 slow_result = _callprojs.resproj->clone();
1491 transform_later(slow_result);
1492 _igvn.replace_node(_callprojs.resproj, result_phi_rawoop);
1493 }
1494
1495 // Plug slow-path into result merge point
1496 result_region->init_req( slow_result_path, ctrl);
1497 transform_later(result_region);
1498 if (allocation_has_use) {
1499 result_phi_rawoop->init_req(slow_result_path, slow_result);
1500 transform_later(result_phi_rawoop);
1501 }
1502 result_phi_rawmem->init_req(slow_result_path, _callprojs.fallthrough_memproj);
1503 transform_later(result_phi_rawmem);
1504 transform_later(result_phi_i_o);
1505 // This completes all paths into the result merge point
1506 }
1507
1508 // Remove alloc node that has no uses.
1509 void PhaseMacroExpand::yank_alloc_node(AllocateNode* alloc) {
1510 Node* ctrl = alloc->in(TypeFunc::Control);
1511 Node* mem = alloc->in(TypeFunc::Memory);
1512 Node* i_o = alloc->in(TypeFunc::I_O);
1513
1514 alloc->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
1515 if (_callprojs.resproj != NULL) {
1516 for (DUIterator_Fast imax, i = _callprojs.resproj->fast_outs(imax); i < imax; i++) {
1517 Node* use = _callprojs.resproj->fast_out(i);
1518 use->isa_MemBar()->remove(&_igvn);
1519 --imax;
1520 --i; // back up iterator
1521 }
1522 assert(_callprojs.resproj->outcnt() == 0, "all uses must be deleted");
1523 _igvn.remove_dead_node(_callprojs.resproj);
1524 }
1525 if (_callprojs.fallthrough_catchproj != NULL) {
1526 migrate_outs(_callprojs.fallthrough_catchproj, ctrl);
1527 _igvn.remove_dead_node(_callprojs.fallthrough_catchproj);
1528 }
1529 if (_callprojs.catchall_catchproj != NULL) {
1530 _igvn.rehash_node_delayed(_callprojs.catchall_catchproj);
1531 _callprojs.catchall_catchproj->set_req(0, top());
1532 }
1533 if (_callprojs.fallthrough_proj != NULL) {
1534 Node* catchnode = _callprojs.fallthrough_proj->unique_ctrl_out();
1535 _igvn.remove_dead_node(catchnode);
1536 _igvn.remove_dead_node(_callprojs.fallthrough_proj);
1537 }
1538 if (_callprojs.fallthrough_memproj != NULL) {
1539 migrate_outs(_callprojs.fallthrough_memproj, mem);
1540 _igvn.remove_dead_node(_callprojs.fallthrough_memproj);
1541 }
1542 if (_callprojs.fallthrough_ioproj != NULL) {
1543 migrate_outs(_callprojs.fallthrough_ioproj, i_o);
1544 _igvn.remove_dead_node(_callprojs.fallthrough_ioproj);
1545 }
1546 if (_callprojs.catchall_memproj != NULL) {
1547 _igvn.rehash_node_delayed(_callprojs.catchall_memproj);
1548 _callprojs.catchall_memproj->set_req(0, top());
1549 }
1550 if (_callprojs.catchall_ioproj != NULL) {
1551 _igvn.rehash_node_delayed(_callprojs.catchall_ioproj);
1552 _callprojs.catchall_ioproj->set_req(0, top());
1553 }
1554 #ifndef PRODUCT
1555 if (PrintEliminateAllocations) {
1556 if (alloc->is_AllocateArray()) {
1557 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1558 } else {
1559 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1560 }
1561 }
1562 #endif
1563 _igvn.remove_dead_node(alloc);
1564 }
1565
1566 void PhaseMacroExpand::expand_initialize_membar(AllocateNode* alloc, InitializeNode* init,
1567 Node*& fast_oop_ctrl, Node*& fast_oop_rawmem) {
1568 // If initialization is performed by an array copy, any required
1569 // MemBarStoreStore was already added. If the object does not
1570 // escape no need for a MemBarStoreStore. If the object does not
1571 // escape in its initializer and memory barrier (MemBarStoreStore or
1572 // stronger) is already added at exit of initializer, also no need
1573 // for a MemBarStoreStore. Otherwise we need a MemBarStoreStore
1574 // so that stores that initialize this object can't be reordered
1575 // with a subsequent store that makes this object accessible by
1576 // other threads.
1577 // Other threads include java threads and JVM internal threads
1578 // (for example concurrent GC threads). Current concurrent GC
1579 // implementation: G1 will not scan newly created object,
1580 // so it's safe to skip storestore barrier when allocation does
1581 // not escape.
1582 if (!alloc->does_not_escape_thread() &&
1583 !alloc->is_allocation_MemBar_redundant() &&
1584 (init == NULL || !init->is_complete_with_arraycopy())) {
1585 if (init == NULL || init->req() < InitializeNode::RawStores) {
1586 // No InitializeNode or no stores captured by zeroing
1587 // elimination. Simply add the MemBarStoreStore after object
1588 // initialization.
1589 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1590 transform_later(mb);
1591
1592 mb->init_req(TypeFunc::Memory, fast_oop_rawmem);
1593 mb->init_req(TypeFunc::Control, fast_oop_ctrl);
1594 fast_oop_ctrl = new ProjNode(mb, TypeFunc::Control);
1595 transform_later(fast_oop_ctrl);
1596 fast_oop_rawmem = new ProjNode(mb, TypeFunc::Memory);
1597 transform_later(fast_oop_rawmem);
1598 } else {
1599 // Add the MemBarStoreStore after the InitializeNode so that
1600 // all stores performing the initialization that were moved
1601 // before the InitializeNode happen before the storestore
1602 // barrier.
1603
1604 Node* init_ctrl = init->proj_out_or_null(TypeFunc::Control);
1605 Node* init_mem = init->proj_out_or_null(TypeFunc::Memory);
1606
1607 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1608 transform_later(mb);
1609
1610 Node* ctrl = new ProjNode(init, TypeFunc::Control);
1611 transform_later(ctrl);
1612 Node* mem = new ProjNode(init, TypeFunc::Memory);
1613 transform_later(mem);
1614
1615 // The MemBarStoreStore depends on control and memory coming
1616 // from the InitializeNode
1617 mb->init_req(TypeFunc::Memory, mem);
1618 mb->init_req(TypeFunc::Control, ctrl);
1619
1620 ctrl = new ProjNode(mb, TypeFunc::Control);
1621 transform_later(ctrl);
1622 mem = new ProjNode(mb, TypeFunc::Memory);
1623 transform_later(mem);
1624
1625 // All nodes that depended on the InitializeNode for control
1626 // and memory must now depend on the MemBarNode that itself
1627 // depends on the InitializeNode
1628 if (init_ctrl != NULL) {
1629 _igvn.replace_node(init_ctrl, ctrl);
1630 }
1631 if (init_mem != NULL) {
1632 _igvn.replace_node(init_mem, mem);
1633 }
1634 }
1635 }
1636 }
1637
1638 void PhaseMacroExpand::expand_dtrace_alloc_probe(AllocateNode* alloc, Node* oop,
1639 Node*& ctrl, Node*& rawmem) {
1640 if (C->env()->dtrace_alloc_probes()) {
1641 // Slow-path call
1642 int size = TypeFunc::Parms + 2;
1643 CallLeafNode *call = new CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(),
1644 CAST_FROM_FN_PTR(address,
1645 static_cast<int (*)(Thread*, oopDesc*)>(SharedRuntime::dtrace_object_alloc)),
1646 "dtrace_object_alloc",
1647 TypeRawPtr::BOTTOM);
1648
1649 // Get base of thread-local storage area
1650 Node* thread = new ThreadLocalNode();
1651 transform_later(thread);
1652
1653 call->init_req(TypeFunc::Parms + 0, thread);
1654 call->init_req(TypeFunc::Parms + 1, oop);
1655 call->init_req(TypeFunc::Control, ctrl);
1656 call->init_req(TypeFunc::I_O , top()); // does no i/o
1657 call->init_req(TypeFunc::Memory , rawmem);
1658 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1659 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1660 transform_later(call);
1661 ctrl = new ProjNode(call, TypeFunc::Control);
1662 transform_later(ctrl);
1663 rawmem = new ProjNode(call, TypeFunc::Memory);
1664 transform_later(rawmem);
1665 }
1666 }
1667
1668 // Helper for PhaseMacroExpand::expand_allocate_common.
1669 // Initializes the newly-allocated storage.
1670 Node*
1671 PhaseMacroExpand::initialize_object(AllocateNode* alloc,
1672 Node* control, Node* rawmem, Node* object,
1673 Node* klass_node, Node* length,
1674 Node* size_in_bytes) {
1675 InitializeNode* init = alloc->initialization();
1676 // Store the klass & mark bits
1677 Node* mark_node = alloc->make_ideal_mark(&_igvn, object, control, rawmem);
1678 if (!mark_node->is_Con()) {
1679 transform_later(mark_node);
1680 }
1681 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, TypeX_X->basic_type());
1682
1683 #ifndef _LP64
1684 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
1685 #endif
1686 int header_size = alloc->minimum_header_size(); // conservatively small
1687
1688 // Array length
1689 if (length != NULL) { // Arrays need length field
1690 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
1691 // conservatively small header size:
1692 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1693 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
1694 if (k->is_array_klass()) // we know the exact header size in most cases:
1695 header_size = Klass::layout_helper_header_size(k->layout_helper());
1696 }
1697
1698 // Clear the object body, if necessary.
1699 if (init == NULL) {
1700 // The init has somehow disappeared; be cautious and clear everything.
1701 //
1702 // This can happen if a node is allocated but an uncommon trap occurs
1703 // immediately. In this case, the Initialize gets associated with the
1704 // trap, and may be placed in a different (outer) loop, if the Allocate
1705 // is in a loop. If (this is rare) the inner loop gets unrolled, then
1706 // there can be two Allocates to one Initialize. The answer in all these
1707 // edge cases is safety first. It is always safe to clear immediately
1708 // within an Allocate, and then (maybe or maybe not) clear some more later.
1709 if (!(UseTLAB && ZeroTLAB)) {
1710 rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
1711 header_size, size_in_bytes,
1712 &_igvn);
1713 }
1714 } else {
1715 if (!init->is_complete()) {
1716 // Try to win by zeroing only what the init does not store.
1717 // We can also try to do some peephole optimizations,
1718 // such as combining some adjacent subword stores.
1719 rawmem = init->complete_stores(control, rawmem, object,
1720 header_size, size_in_bytes, &_igvn);
1721 }
1722 // We have no more use for this link, since the AllocateNode goes away:
1723 init->set_req(InitializeNode::RawAddress, top());
1724 // (If we keep the link, it just confuses the register allocator,
1725 // who thinks he sees a real use of the address by the membar.)
1726 }
1727
1728 return rawmem;
1729 }
1730
1731 // Generate prefetch instructions for next allocations.
1732 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false,
1733 Node*& contended_phi_rawmem,
1734 Node* old_eden_top, Node* new_eden_top,
1735 intx lines) {
1736 enum { fall_in_path = 1, pf_path = 2 };
1737 if( UseTLAB && AllocatePrefetchStyle == 2 ) {
1738 // Generate prefetch allocation with watermark check.
1739 // As an allocation hits the watermark, we will prefetch starting
1740 // at a "distance" away from watermark.
1741
1742 Node *pf_region = new RegionNode(3);
1743 Node *pf_phi_rawmem = new PhiNode( pf_region, Type::MEMORY,
1744 TypeRawPtr::BOTTOM );
1745 // I/O is used for Prefetch
1746 Node *pf_phi_abio = new PhiNode( pf_region, Type::ABIO );
1747
1748 Node *thread = new ThreadLocalNode();
1749 transform_later(thread);
1750
1751 Node *eden_pf_adr = new AddPNode( top()/*not oop*/, thread,
1752 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) );
1753 transform_later(eden_pf_adr);
1754
1755 Node *old_pf_wm = new LoadPNode(needgc_false,
1756 contended_phi_rawmem, eden_pf_adr,
1757 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM,
1758 MemNode::unordered);
1759 transform_later(old_pf_wm);
1760
1761 // check against new_eden_top
1762 Node *need_pf_cmp = new CmpPNode( new_eden_top, old_pf_wm );
1763 transform_later(need_pf_cmp);
1764 Node *need_pf_bol = new BoolNode( need_pf_cmp, BoolTest::ge );
1765 transform_later(need_pf_bol);
1766 IfNode *need_pf_iff = new IfNode( needgc_false, need_pf_bol,
1767 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN );
1768 transform_later(need_pf_iff);
1769
1770 // true node, add prefetchdistance
1771 Node *need_pf_true = new IfTrueNode( need_pf_iff );
1772 transform_later(need_pf_true);
1773
1774 Node *need_pf_false = new IfFalseNode( need_pf_iff );
1775 transform_later(need_pf_false);
1776
1777 Node *new_pf_wmt = new AddPNode( top(), old_pf_wm,
1778 _igvn.MakeConX(AllocatePrefetchDistance) );
1779 transform_later(new_pf_wmt );
1780 new_pf_wmt->set_req(0, need_pf_true);
1781
1782 Node *store_new_wmt = new StorePNode(need_pf_true,
1783 contended_phi_rawmem, eden_pf_adr,
1784 TypeRawPtr::BOTTOM, new_pf_wmt,
1785 MemNode::unordered);
1786 transform_later(store_new_wmt);
1787
1788 // adding prefetches
1789 pf_phi_abio->init_req( fall_in_path, i_o );
1790
1791 Node *prefetch_adr;
1792 Node *prefetch;
1793 uint step_size = AllocatePrefetchStepSize;
1794 uint distance = 0;
1795
1796 for ( intx i = 0; i < lines; i++ ) {
1797 prefetch_adr = new AddPNode( old_pf_wm, new_pf_wmt,
1798 _igvn.MakeConX(distance) );
1799 transform_later(prefetch_adr);
1800 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
1801 transform_later(prefetch);
1802 distance += step_size;
1803 i_o = prefetch;
1804 }
1805 pf_phi_abio->set_req( pf_path, i_o );
1806
1807 pf_region->init_req( fall_in_path, need_pf_false );
1808 pf_region->init_req( pf_path, need_pf_true );
1809
1810 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem );
1811 pf_phi_rawmem->init_req( pf_path, store_new_wmt );
1812
1813 transform_later(pf_region);
1814 transform_later(pf_phi_rawmem);
1815 transform_later(pf_phi_abio);
1816
1817 needgc_false = pf_region;
1818 contended_phi_rawmem = pf_phi_rawmem;
1819 i_o = pf_phi_abio;
1820 } else if( UseTLAB && AllocatePrefetchStyle == 3 ) {
1821 // Insert a prefetch instruction for each allocation.
1822 // This code is used to generate 1 prefetch instruction per cache line.
1823
1824 // Generate several prefetch instructions.
1825 uint step_size = AllocatePrefetchStepSize;
1826 uint distance = AllocatePrefetchDistance;
1827
1828 // Next cache address.
1829 Node *cache_adr = new AddPNode(old_eden_top, old_eden_top,
1830 _igvn.MakeConX(step_size + distance));
1831 transform_later(cache_adr);
1832 cache_adr = new CastP2XNode(needgc_false, cache_adr);
1833 transform_later(cache_adr);
1834 // Address is aligned to execute prefetch to the beginning of cache line size
1835 // (it is important when BIS instruction is used on SPARC as prefetch).
1836 Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1));
1837 cache_adr = new AndXNode(cache_adr, mask);
1838 transform_later(cache_adr);
1839 cache_adr = new CastX2PNode(cache_adr);
1840 transform_later(cache_adr);
1841
1842 // Prefetch
1843 Node *prefetch = new PrefetchAllocationNode( contended_phi_rawmem, cache_adr );
1844 prefetch->set_req(0, needgc_false);
1845 transform_later(prefetch);
1846 contended_phi_rawmem = prefetch;
1847 Node *prefetch_adr;
1848 distance = step_size;
1849 for ( intx i = 1; i < lines; i++ ) {
1850 prefetch_adr = new AddPNode( cache_adr, cache_adr,
1851 _igvn.MakeConX(distance) );
1852 transform_later(prefetch_adr);
1853 prefetch = new PrefetchAllocationNode( contended_phi_rawmem, prefetch_adr );
1854 transform_later(prefetch);
1855 distance += step_size;
1856 contended_phi_rawmem = prefetch;
1857 }
1858 } else if( AllocatePrefetchStyle > 0 ) {
1859 // Insert a prefetch for each allocation only on the fast-path
1860 Node *prefetch_adr;
1861 Node *prefetch;
1862 // Generate several prefetch instructions.
1863 uint step_size = AllocatePrefetchStepSize;
1864 uint distance = AllocatePrefetchDistance;
1865 for ( intx i = 0; i < lines; i++ ) {
1866 prefetch_adr = new AddPNode( old_eden_top, new_eden_top,
1867 _igvn.MakeConX(distance) );
1868 transform_later(prefetch_adr);
1869 prefetch = new PrefetchAllocationNode( i_o, prefetch_adr );
1870 // Do not let it float too high, since if eden_top == eden_end,
1871 // both might be null.
1872 if( i == 0 ) { // Set control for first prefetch, next follows it
1873 prefetch->init_req(0, needgc_false);
1874 }
1875 transform_later(prefetch);
1876 distance += step_size;
1877 i_o = prefetch;
1878 }
1879 }
1880 return i_o;
1881 }
1882
1883
1884 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) {
1885 expand_allocate_common(alloc, NULL,
1886 OptoRuntime::new_instance_Type(),
1887 OptoRuntime::new_instance_Java(), NULL);
1888 }
1889
1890 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) {
1891 Node* length = alloc->in(AllocateNode::ALength);
1892 Node* valid_length_test = alloc->in(AllocateNode::ValidLengthTest);
1893 InitializeNode* init = alloc->initialization();
1894 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1895 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
1896 address slow_call_address; // Address of slow call
1897 if (init != NULL && init->is_complete_with_arraycopy() &&
1898 k->is_type_array_klass()) {
1899 // Don't zero type array during slow allocation in VM since
1900 // it will be initialized later by arraycopy in compiled code.
1901 slow_call_address = OptoRuntime::new_array_nozero_Java();
1902 } else {
1903 slow_call_address = OptoRuntime::new_array_Java();
1904 }
1905 expand_allocate_common(alloc, length,
1906 OptoRuntime::new_array_Type(),
1907 slow_call_address, valid_length_test);
1908 }
1909
1910 //-------------------mark_eliminated_box----------------------------------
1911 //
1912 // During EA obj may point to several objects but after few ideal graph
1913 // transformations (CCP) it may point to only one non escaping object
1914 // (but still using phi), corresponding locks and unlocks will be marked
1915 // for elimination. Later obj could be replaced with a new node (new phi)
1916 // and which does not have escape information. And later after some graph
1917 // reshape other locks and unlocks (which were not marked for elimination
1918 // before) are connected to this new obj (phi) but they still will not be
1919 // marked for elimination since new obj has no escape information.
1920 // Mark all associated (same box and obj) lock and unlock nodes for
1921 // elimination if some of them marked already.
1922 void PhaseMacroExpand::mark_eliminated_box(Node* oldbox, Node* obj) {
1923 if (oldbox->as_BoxLock()->is_eliminated()) {
1924 return; // This BoxLock node was processed already.
1925 }
1926 // New implementation (EliminateNestedLocks) has separate BoxLock
1927 // node for each locked region so mark all associated locks/unlocks as
1928 // eliminated even if different objects are referenced in one locked region
1929 // (for example, OSR compilation of nested loop inside locked scope).
1930 if (EliminateNestedLocks ||
1931 oldbox->as_BoxLock()->is_simple_lock_region(NULL, obj, NULL)) {
1932 // Box is used only in one lock region. Mark this box as eliminated.
1933 _igvn.hash_delete(oldbox);
1934 oldbox->as_BoxLock()->set_eliminated(); // This changes box's hash value
1935 _igvn.hash_insert(oldbox);
1936
1937 for (uint i = 0; i < oldbox->outcnt(); i++) {
1938 Node* u = oldbox->raw_out(i);
1939 if (u->is_AbstractLock() && !u->as_AbstractLock()->is_non_esc_obj()) {
1940 AbstractLockNode* alock = u->as_AbstractLock();
1941 // Check lock's box since box could be referenced by Lock's debug info.
1942 if (alock->box_node() == oldbox) {
1943 // Mark eliminated all related locks and unlocks.
1944 #ifdef ASSERT
1945 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc4");
1946 #endif
1947 alock->set_non_esc_obj();
1948 }
1949 }
1950 }
1951 return;
1952 }
1953
1954 // Create new "eliminated" BoxLock node and use it in monitor debug info
1955 // instead of oldbox for the same object.
1956 BoxLockNode* newbox = oldbox->clone()->as_BoxLock();
1957
1958 // Note: BoxLock node is marked eliminated only here and it is used
1959 // to indicate that all associated lock and unlock nodes are marked
1960 // for elimination.
1961 newbox->set_eliminated();
1962 transform_later(newbox);
1963
1964 // Replace old box node with new box for all users of the same object.
1965 for (uint i = 0; i < oldbox->outcnt();) {
1966 bool next_edge = true;
1967
1968 Node* u = oldbox->raw_out(i);
1969 if (u->is_AbstractLock()) {
1970 AbstractLockNode* alock = u->as_AbstractLock();
1971 if (alock->box_node() == oldbox && alock->obj_node()->eqv_uncast(obj)) {
1972 // Replace Box and mark eliminated all related locks and unlocks.
1973 #ifdef ASSERT
1974 alock->log_lock_optimization(C, "eliminate_lock_set_non_esc5");
1975 #endif
1976 alock->set_non_esc_obj();
1977 _igvn.rehash_node_delayed(alock);
1978 alock->set_box_node(newbox);
1979 next_edge = false;
1980 }
1981 }
1982 if (u->is_FastLock() && u->as_FastLock()->obj_node()->eqv_uncast(obj)) {
1983 FastLockNode* flock = u->as_FastLock();
1984 assert(flock->box_node() == oldbox, "sanity");
1985 _igvn.rehash_node_delayed(flock);
1986 flock->set_box_node(newbox);
1987 next_edge = false;
1988 }
1989
1990 // Replace old box in monitor debug info.
1991 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) {
1992 SafePointNode* sfn = u->as_SafePoint();
1993 JVMState* youngest_jvms = sfn->jvms();
1994 int max_depth = youngest_jvms->depth();
1995 for (int depth = 1; depth <= max_depth; depth++) {
1996 JVMState* jvms = youngest_jvms->of_depth(depth);
1997 int num_mon = jvms->nof_monitors();
1998 // Loop over monitors
1999 for (int idx = 0; idx < num_mon; idx++) {
2000 Node* obj_node = sfn->monitor_obj(jvms, idx);
2001 Node* box_node = sfn->monitor_box(jvms, idx);
2002 if (box_node == oldbox && obj_node->eqv_uncast(obj)) {
2003 int j = jvms->monitor_box_offset(idx);
2004 _igvn.replace_input_of(u, j, newbox);
2005 next_edge = false;
2006 }
2007 }
2008 }
2009 }
2010 if (next_edge) i++;
2011 }
2012 }
2013
2014 //-----------------------mark_eliminated_locking_nodes-----------------------
2015 void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) {
2016 if (EliminateNestedLocks) {
2017 if (alock->is_nested()) {
2018 assert(alock->box_node()->as_BoxLock()->is_eliminated(), "sanity");
2019 return;
2020 } else if (!alock->is_non_esc_obj()) { // Not eliminated or coarsened
2021 // Only Lock node has JVMState needed here.
2022 // Not that preceding claim is documented anywhere else.
2023 if (alock->jvms() != NULL) {
2024 if (alock->as_Lock()->is_nested_lock_region()) {
2025 // Mark eliminated related nested locks and unlocks.
2026 Node* obj = alock->obj_node();
2027 BoxLockNode* box_node = alock->box_node()->as_BoxLock();
2028 assert(!box_node->is_eliminated(), "should not be marked yet");
2029 // Note: BoxLock node is marked eliminated only here
2030 // and it is used to indicate that all associated lock
2031 // and unlock nodes are marked for elimination.
2032 box_node->set_eliminated(); // Box's hash is always NO_HASH here
2033 for (uint i = 0; i < box_node->outcnt(); i++) {
2034 Node* u = box_node->raw_out(i);
2035 if (u->is_AbstractLock()) {
2036 alock = u->as_AbstractLock();
2037 if (alock->box_node() == box_node) {
2038 // Verify that this Box is referenced only by related locks.
2039 assert(alock->obj_node()->eqv_uncast(obj), "");
2040 // Mark all related locks and unlocks.
2041 #ifdef ASSERT
2042 alock->log_lock_optimization(C, "eliminate_lock_set_nested");
2043 #endif
2044 alock->set_nested();
2045 }
2046 }
2047 }
2048 } else {
2049 #ifdef ASSERT
2050 alock->log_lock_optimization(C, "eliminate_lock_NOT_nested_lock_region");
2051 if (C->log() != NULL)
2052 alock->as_Lock()->is_nested_lock_region(C); // rerun for debugging output
2053 #endif
2054 }
2055 }
2056 return;
2057 }
2058 // Process locks for non escaping object
2059 assert(alock->is_non_esc_obj(), "");
2060 } // EliminateNestedLocks
2061
2062 if (alock->is_non_esc_obj()) { // Lock is used for non escaping object
2063 // Look for all locks of this object and mark them and
2064 // corresponding BoxLock nodes as eliminated.
2065 Node* obj = alock->obj_node();
2066 for (uint j = 0; j < obj->outcnt(); j++) {
2067 Node* o = obj->raw_out(j);
2068 if (o->is_AbstractLock() &&
2069 o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) {
2070 alock = o->as_AbstractLock();
2071 Node* box = alock->box_node();
2072 // Replace old box node with new eliminated box for all users
2073 // of the same object and mark related locks as eliminated.
2074 mark_eliminated_box(box, obj);
2075 }
2076 }
2077 }
2078 }
2079
2080 // we have determined that this lock/unlock can be eliminated, we simply
2081 // eliminate the node without expanding it.
2082 //
2083 // Note: The membar's associated with the lock/unlock are currently not
2084 // eliminated. This should be investigated as a future enhancement.
2085 //
2086 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) {
2087
2088 if (!alock->is_eliminated()) {
2089 return false;
2090 }
2091 #ifdef ASSERT
2092 if (!alock->is_coarsened()) {
2093 // Check that new "eliminated" BoxLock node is created.
2094 BoxLockNode* oldbox = alock->box_node()->as_BoxLock();
2095 assert(oldbox->is_eliminated(), "should be done already");
2096 }
2097 #endif
2098
2099 alock->log_lock_optimization(C, "eliminate_lock");
2100
2101 #ifndef PRODUCT
2102 if (PrintEliminateLocks) {
2103 tty->print_cr("++++ Eliminated: %d %s '%s'", alock->_idx, (alock->is_Lock() ? "Lock" : "Unlock"), alock->kind_as_string());
2104 }
2105 #endif
2106
2107 Node* mem = alock->in(TypeFunc::Memory);
2108 Node* ctrl = alock->in(TypeFunc::Control);
2109 guarantee(ctrl != NULL, "missing control projection, cannot replace_node() with NULL");
2110
2111 alock->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
2112 // There are 2 projections from the lock. The lock node will
2113 // be deleted when its last use is subsumed below.
2114 assert(alock->outcnt() == 2 &&
2115 _callprojs.fallthrough_proj != NULL &&
2116 _callprojs.fallthrough_memproj != NULL,
2117 "Unexpected projections from Lock/Unlock");
2118
2119 Node* fallthroughproj = _callprojs.fallthrough_proj;
2120 Node* memproj_fallthrough = _callprojs.fallthrough_memproj;
2121
2122 // The memory projection from a lock/unlock is RawMem
2123 // The input to a Lock is merged memory, so extract its RawMem input
2124 // (unless the MergeMem has been optimized away.)
2125 if (alock->is_Lock()) {
2126 // Search for MemBarAcquireLock node and delete it also.
2127 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
2128 assert(membar != NULL && membar->Opcode() == Op_MemBarAcquireLock, "");
2129 Node* ctrlproj = membar->proj_out(TypeFunc::Control);
2130 Node* memproj = membar->proj_out(TypeFunc::Memory);
2131 _igvn.replace_node(ctrlproj, fallthroughproj);
2132 _igvn.replace_node(memproj, memproj_fallthrough);
2133
2134 // Delete FastLock node also if this Lock node is unique user
2135 // (a loop peeling may clone a Lock node).
2136 Node* flock = alock->as_Lock()->fastlock_node();
2137 if (flock->outcnt() == 1) {
2138 assert(flock->unique_out() == alock, "sanity");
2139 _igvn.replace_node(flock, top());
2140 }
2141 }
2142
2143 // Search for MemBarReleaseLock node and delete it also.
2144 if (alock->is_Unlock() && ctrl->is_Proj() && ctrl->in(0)->is_MemBar()) {
2145 MemBarNode* membar = ctrl->in(0)->as_MemBar();
2146 assert(membar->Opcode() == Op_MemBarReleaseLock &&
2147 mem->is_Proj() && membar == mem->in(0), "");
2148 _igvn.replace_node(fallthroughproj, ctrl);
2149 _igvn.replace_node(memproj_fallthrough, mem);
2150 fallthroughproj = ctrl;
2151 memproj_fallthrough = mem;
2152 ctrl = membar->in(TypeFunc::Control);
2153 mem = membar->in(TypeFunc::Memory);
2154 }
2155
2156 _igvn.replace_node(fallthroughproj, ctrl);
2157 _igvn.replace_node(memproj_fallthrough, mem);
2158 return true;
2159 }
2160
2161
2162 //------------------------------expand_lock_node----------------------
2163 void PhaseMacroExpand::expand_lock_node(LockNode *lock) {
2164
2165 Node* ctrl = lock->in(TypeFunc::Control);
2166 Node* mem = lock->in(TypeFunc::Memory);
2167 Node* obj = lock->obj_node();
2168 Node* box = lock->box_node();
2169 Node* flock = lock->fastlock_node();
2170
2171 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2172
2173 // Make the merge point
2174 Node *region;
2175 Node *mem_phi;
2176 Node *slow_path;
2177
2178 region = new RegionNode(3);
2179 // create a Phi for the memory state
2180 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2181
2182 // Optimize test; set region slot 2
2183 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0);
2184 mem_phi->init_req(2, mem);
2185
2186 // Make slow path call
2187 CallNode *call = make_slow_call((CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(),
2188 OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path,
2189 obj, box, NULL);
2190
2191 call->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
2192
2193 // Slow path can only throw asynchronous exceptions, which are always
2194 // de-opted. So the compiler thinks the slow-call can never throw an
2195 // exception. If it DOES throw an exception we would need the debug
2196 // info removed first (since if it throws there is no monitor).
2197 assert(_callprojs.fallthrough_ioproj == NULL && _callprojs.catchall_ioproj == NULL &&
2198 _callprojs.catchall_memproj == NULL && _callprojs.catchall_catchproj == NULL, "Unexpected projection from Lock");
2199
2200 // Capture slow path
2201 // disconnect fall-through projection from call and create a new one
2202 // hook up users of fall-through projection to region
2203 Node *slow_ctrl = _callprojs.fallthrough_proj->clone();
2204 transform_later(slow_ctrl);
2205 _igvn.hash_delete(_callprojs.fallthrough_proj);
2206 _callprojs.fallthrough_proj->disconnect_inputs(C);
2207 region->init_req(1, slow_ctrl);
2208 // region inputs are now complete
2209 transform_later(region);
2210 _igvn.replace_node(_callprojs.fallthrough_proj, region);
2211
2212 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory));
2213 mem_phi->init_req(1, memproj );
2214 transform_later(mem_phi);
2215 _igvn.replace_node(_callprojs.fallthrough_memproj, mem_phi);
2216 }
2217
2218 //------------------------------expand_unlock_node----------------------
2219 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
2220
2221 Node* ctrl = unlock->in(TypeFunc::Control);
2222 Node* mem = unlock->in(TypeFunc::Memory);
2223 Node* obj = unlock->obj_node();
2224 Node* box = unlock->box_node();
2225
2226 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2227
2228 // No need for a null check on unlock
2229
2230 // Make the merge point
2231 Node *region;
2232 Node *mem_phi;
2233
2234 region = new RegionNode(3);
2235 // create a Phi for the memory state
2236 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2237
2238 FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box );
2239 funlock = transform_later( funlock )->as_FastUnlock();
2240 // Optimize test; set region slot 2
2241 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0);
2242 Node *thread = transform_later(new ThreadLocalNode());
2243
2244 CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(),
2245 CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C),
2246 "complete_monitor_unlocking_C", slow_path, obj, box, thread);
2247
2248 call->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
2249 assert(_callprojs.fallthrough_ioproj == NULL && _callprojs.catchall_ioproj == NULL &&
2250 _callprojs.catchall_memproj == NULL && _callprojs.catchall_catchproj == NULL, "Unexpected projection from Lock");
2251
2252 // No exceptions for unlocking
2253 // Capture slow path
2254 // disconnect fall-through projection from call and create a new one
2255 // hook up users of fall-through projection to region
2256 Node *slow_ctrl = _callprojs.fallthrough_proj->clone();
2257 transform_later(slow_ctrl);
2258 _igvn.hash_delete(_callprojs.fallthrough_proj);
2259 _callprojs.fallthrough_proj->disconnect_inputs(C);
2260 region->init_req(1, slow_ctrl);
2261 // region inputs are now complete
2262 transform_later(region);
2263 _igvn.replace_node(_callprojs.fallthrough_proj, region);
2264
2265 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory) );
2266 mem_phi->init_req(1, memproj );
2267 mem_phi->init_req(2, mem);
2268 transform_later(mem_phi);
2269 _igvn.replace_node(_callprojs.fallthrough_memproj, mem_phi);
2270 }
2271
2272 void PhaseMacroExpand::expand_subtypecheck_node(SubTypeCheckNode *check) {
2273 assert(check->in(SubTypeCheckNode::Control) == NULL, "should be pinned");
2274 Node* bol = check->unique_out();
2275 Node* obj_or_subklass = check->in(SubTypeCheckNode::ObjOrSubKlass);
2276 Node* superklass = check->in(SubTypeCheckNode::SuperKlass);
2277 assert(bol->is_Bool() && bol->as_Bool()->_test._test == BoolTest::ne, "unexpected bool node");
2278
2279 for (DUIterator_Last imin, i = bol->last_outs(imin); i >= imin; --i) {
2280 Node* iff = bol->last_out(i);
2281 assert(iff->is_If(), "where's the if?");
2282
2283 if (iff->in(0)->is_top()) {
2284 _igvn.replace_input_of(iff, 1, C->top());
2285 continue;
2286 }
2287
2288 Node* iftrue = iff->as_If()->proj_out(1);
2289 Node* iffalse = iff->as_If()->proj_out(0);
2290 Node* ctrl = iff->in(0);
2291
2292 Node* subklass = NULL;
2293 if (_igvn.type(obj_or_subklass)->isa_klassptr()) {
2294 subklass = obj_or_subklass;
2295 } else {
2296 Node* k_adr = basic_plus_adr(obj_or_subklass, oopDesc::klass_offset_in_bytes());
2297 subklass = _igvn.transform(LoadKlassNode::make(_igvn, NULL, C->immutable_memory(), k_adr, TypeInstPtr::KLASS));
2298 }
2299
2300 Node* not_subtype_ctrl = Phase::gen_subtype_check(subklass, superklass, &ctrl, NULL, _igvn);
2301
2302 _igvn.replace_input_of(iff, 0, C->top());
2303 _igvn.replace_node(iftrue, not_subtype_ctrl);
2304 _igvn.replace_node(iffalse, ctrl);
2305 }
2306 _igvn.replace_node(check, C->top());
2307 }
2308
2309 //---------------------------eliminate_macro_nodes----------------------
2310 // Eliminate scalar replaced allocations and associated locks.
2311 void PhaseMacroExpand::eliminate_macro_nodes() {
2312 if (C->macro_count() == 0)
2313 return;
2314
2315 // Before elimination may re-mark (change to Nested or NonEscObj)
2316 // all associated (same box and obj) lock and unlock nodes.
2317 int cnt = C->macro_count();
2318 for (int i=0; i < cnt; i++) {
2319 Node *n = C->macro_node(i);
2320 if (n->is_AbstractLock()) { // Lock and Unlock nodes
2321 mark_eliminated_locking_nodes(n->as_AbstractLock());
2322 }
2323 }
2324 // Re-marking may break consistency of Coarsened locks.
2325 if (!C->coarsened_locks_consistent()) {
2326 return; // recompile without Coarsened locks if broken
2327 }
2328
2329 // First, attempt to eliminate locks
2330 bool progress = true;
2331 while (progress) {
2332 progress = false;
2333 for (int i = C->macro_count(); i > 0; i = MIN2(i - 1, C->macro_count())) { // more than 1 element can be eliminated at once
2334 Node* n = C->macro_node(i - 1);
2335 bool success = false;
2336 DEBUG_ONLY(int old_macro_count = C->macro_count();)
2337 if (n->is_AbstractLock()) {
2338 success = eliminate_locking_node(n->as_AbstractLock());
2339 }
2340 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2341 progress = progress || success;
2342 }
2343 }
2344 // Next, attempt to eliminate allocations
2345 _has_locks = false;
2346 progress = true;
2347 while (progress) {
2348 progress = false;
2349 for (int i = C->macro_count(); i > 0; i = MIN2(i - 1, C->macro_count())) { // more than 1 element can be eliminated at once
2350 Node* n = C->macro_node(i - 1);
2351 bool success = false;
2352 DEBUG_ONLY(int old_macro_count = C->macro_count();)
2353 switch (n->class_id()) {
2354 case Node::Class_Allocate:
2355 case Node::Class_AllocateArray:
2356 success = eliminate_allocate_node(n->as_Allocate());
2357 break;
2358 case Node::Class_CallStaticJava:
2359 success = eliminate_boxing_node(n->as_CallStaticJava());
2360 break;
2361 case Node::Class_Lock:
2362 case Node::Class_Unlock:
2363 assert(!n->as_AbstractLock()->is_eliminated(), "sanity");
2364 _has_locks = true;
2365 break;
2366 case Node::Class_ArrayCopy:
2367 break;
2368 case Node::Class_OuterStripMinedLoop:
2369 break;
2370 case Node::Class_SubTypeCheck:
2371 break;
2372 case Node::Class_Opaque1:
2373 break;
2374 default:
2375 assert(n->Opcode() == Op_LoopLimit ||
2376 n->Opcode() == Op_Opaque2 ||
2377 n->Opcode() == Op_Opaque3 ||
2378 n->Opcode() == Op_Opaque4 ||
2379 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(n),
2380 "unknown node type in macro list");
2381 }
2382 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2383 progress = progress || success;
2384 }
2385 }
2386 }
2387
2388 //------------------------------expand_macro_nodes----------------------
2389 // Returns true if a failure occurred.
2390 bool PhaseMacroExpand::expand_macro_nodes() {
2391 // Last attempt to eliminate macro nodes.
2392 eliminate_macro_nodes();
2393 if (C->failing()) return true;
2394
2395 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations.
2396 bool progress = true;
2397 while (progress) {
2398 progress = false;
2399 for (int i = C->macro_count(); i > 0; i--) {
2400 Node* n = C->macro_node(i-1);
2401 bool success = false;
2402 DEBUG_ONLY(int old_macro_count = C->macro_count();)
2403 if (n->Opcode() == Op_LoopLimit) {
2404 // Remove it from macro list and put on IGVN worklist to optimize.
2405 C->remove_macro_node(n);
2406 _igvn._worklist.push(n);
2407 success = true;
2408 } else if (n->Opcode() == Op_CallStaticJava) {
2409 // Remove it from macro list and put on IGVN worklist to optimize.
2410 C->remove_macro_node(n);
2411 _igvn._worklist.push(n);
2412 success = true;
2413 } else if (n->is_Opaque1() || n->Opcode() == Op_Opaque2) {
2414 _igvn.replace_node(n, n->in(1));
2415 success = true;
2416 #if INCLUDE_RTM_OPT
2417 } else if ((n->Opcode() == Op_Opaque3) && ((Opaque3Node*)n)->rtm_opt()) {
2418 assert(C->profile_rtm(), "should be used only in rtm deoptimization code");
2419 assert((n->outcnt() == 1) && n->unique_out()->is_Cmp(), "");
2420 Node* cmp = n->unique_out();
2421 #ifdef ASSERT
2422 // Validate graph.
2423 assert((cmp->outcnt() == 1) && cmp->unique_out()->is_Bool(), "");
2424 BoolNode* bol = cmp->unique_out()->as_Bool();
2425 assert((bol->outcnt() == 1) && bol->unique_out()->is_If() &&
2426 (bol->_test._test == BoolTest::ne), "");
2427 IfNode* ifn = bol->unique_out()->as_If();
2428 assert((ifn->outcnt() == 2) &&
2429 ifn->proj_out(1)->is_uncommon_trap_proj(Deoptimization::Reason_rtm_state_change) != NULL, "");
2430 #endif
2431 Node* repl = n->in(1);
2432 if (!_has_locks) {
2433 // Remove RTM state check if there are no locks in the code.
2434 // Replace input to compare the same value.
2435 repl = (cmp->in(1) == n) ? cmp->in(2) : cmp->in(1);
2436 }
2437 _igvn.replace_node(n, repl);
2438 success = true;
2439 #endif
2440 } else if (n->Opcode() == Op_Opaque4) {
2441 // With Opaque4 nodes, the expectation is that the test of input 1
2442 // is always equal to the constant value of input 2. So we can
2443 // remove the Opaque4 and replace it by input 2. In debug builds,
2444 // leave the non constant test in instead to sanity check that it
2445 // never fails (if it does, that subgraph was constructed so, at
2446 // runtime, a Halt node is executed).
2447 #ifdef ASSERT
2448 _igvn.replace_node(n, n->in(1));
2449 #else
2450 _igvn.replace_node(n, n->in(2));
2451 #endif
2452 success = true;
2453 } else if (n->Opcode() == Op_OuterStripMinedLoop) {
2454 n->as_OuterStripMinedLoop()->adjust_strip_mined_loop(&_igvn);
2455 C->remove_macro_node(n);
2456 success = true;
2457 }
2458 assert(!success || (C->macro_count() == (old_macro_count - 1)), "elimination must have deleted one node from macro list");
2459 progress = progress || success;
2460 }
2461 }
2462
2463 // Clean up the graph so we're less likely to hit the maximum node
2464 // limit
2465 _igvn.set_delay_transform(false);
2466 _igvn.optimize();
2467 if (C->failing()) return true;
2468 _igvn.set_delay_transform(true);
2469
2470
2471 // Because we run IGVN after each expansion, some macro nodes may go
2472 // dead and be removed from the list as we iterate over it. Move
2473 // Allocate nodes (processed in a second pass) at the beginning of
2474 // the list and then iterate from the last element of the list until
2475 // an Allocate node is seen. This is robust to random deletion in
2476 // the list due to nodes going dead.
2477 C->sort_macro_nodes();
2478
2479 // expand arraycopy "macro" nodes first
2480 // For ReduceBulkZeroing, we must first process all arraycopy nodes
2481 // before the allocate nodes are expanded.
2482 while (C->macro_count() > 0) {
2483 int macro_count = C->macro_count();
2484 Node * n = C->macro_node(macro_count-1);
2485 assert(n->is_macro(), "only macro nodes expected here");
2486 if (_igvn.type(n) == Type::TOP || (n->in(0) != NULL && n->in(0)->is_top())) {
2487 // node is unreachable, so don't try to expand it
2488 C->remove_macro_node(n);
2489 continue;
2490 }
2491 if (n->is_Allocate()) {
2492 break;
2493 }
2494 // Make sure expansion will not cause node limit to be exceeded.
2495 // Worst case is a macro node gets expanded into about 200 nodes.
2496 // Allow 50% more for optimization.
2497 if (C->check_node_count(300, "out of nodes before macro expansion")) {
2498 return true;
2499 }
2500
2501 DEBUG_ONLY(int old_macro_count = C->macro_count();)
2502 switch (n->class_id()) {
2503 case Node::Class_Lock:
2504 expand_lock_node(n->as_Lock());
2505 break;
2506 case Node::Class_Unlock:
2507 expand_unlock_node(n->as_Unlock());
2508 break;
2509 case Node::Class_ArrayCopy:
2510 expand_arraycopy_node(n->as_ArrayCopy());
2511 break;
2512 case Node::Class_SubTypeCheck:
2513 expand_subtypecheck_node(n->as_SubTypeCheck());
2514 break;
2515 default:
2516 assert(false, "unknown node type in macro list");
2517 }
2518 assert(C->macro_count() == (old_macro_count - 1), "expansion must have deleted one node from macro list");
2519 if (C->failing()) return true;
2520
2521 // Clean up the graph so we're less likely to hit the maximum node
2522 // limit
2523 _igvn.set_delay_transform(false);
2524 _igvn.optimize();
2525 if (C->failing()) return true;
2526 _igvn.set_delay_transform(true);
2527 }
2528
2529 // All nodes except Allocate nodes are expanded now. There could be
2530 // new optimization opportunities (such as folding newly created
2531 // load from a just allocated object). Run IGVN.
2532
2533 // expand "macro" nodes
2534 // nodes are removed from the macro list as they are processed
2535 while (C->macro_count() > 0) {
2536 int macro_count = C->macro_count();
2537 Node * n = C->macro_node(macro_count-1);
2538 assert(n->is_macro(), "only macro nodes expected here");
2539 if (_igvn.type(n) == Type::TOP || (n->in(0) != NULL && n->in(0)->is_top())) {
2540 // node is unreachable, so don't try to expand it
2541 C->remove_macro_node(n);
2542 continue;
2543 }
2544 // Make sure expansion will not cause node limit to be exceeded.
2545 // Worst case is a macro node gets expanded into about 200 nodes.
2546 // Allow 50% more for optimization.
2547 if (C->check_node_count(300, "out of nodes before macro expansion")) {
2548 return true;
2549 }
2550 switch (n->class_id()) {
2551 case Node::Class_Allocate:
2552 expand_allocate(n->as_Allocate());
2553 break;
2554 case Node::Class_AllocateArray:
2555 expand_allocate_array(n->as_AllocateArray());
2556 break;
2557 default:
2558 assert(false, "unknown node type in macro list");
2559 }
2560 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
2561 if (C->failing()) return true;
2562
2563 // Clean up the graph so we're less likely to hit the maximum node
2564 // limit
2565 _igvn.set_delay_transform(false);
2566 _igvn.optimize();
2567 if (C->failing()) return true;
2568 _igvn.set_delay_transform(true);
2569 }
2570
2571 _igvn.set_delay_transform(false);
2572 return false;
2573 }