Tue, 22 Jan 2013 15:34:16 -0800
6896617: Optimize sun.nio.cs.ISO_8859_1$Encode.encodeArrayLoop() on x86
Summary: Use SSE4.2 and AVX2 instructions for encodeArray intrinsic.
Reviewed-by: roland
1 /*
2 * Copyright (c) 2005, 2012, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
25 #include "precompiled.hpp"
26 #include "compiler/compileLog.hpp"
27 #include "libadt/vectset.hpp"
28 #include "opto/addnode.hpp"
29 #include "opto/callnode.hpp"
30 #include "opto/cfgnode.hpp"
31 #include "opto/compile.hpp"
32 #include "opto/connode.hpp"
33 #include "opto/locknode.hpp"
34 #include "opto/loopnode.hpp"
35 #include "opto/macro.hpp"
36 #include "opto/memnode.hpp"
37 #include "opto/node.hpp"
38 #include "opto/phaseX.hpp"
39 #include "opto/rootnode.hpp"
40 #include "opto/runtime.hpp"
41 #include "opto/subnode.hpp"
42 #include "opto/type.hpp"
43 #include "runtime/sharedRuntime.hpp"
46 //
47 // Replace any references to "oldref" in inputs to "use" with "newref".
48 // Returns the number of replacements made.
49 //
50 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) {
51 int nreplacements = 0;
52 uint req = use->req();
53 for (uint j = 0; j < use->len(); j++) {
54 Node *uin = use->in(j);
55 if (uin == oldref) {
56 if (j < req)
57 use->set_req(j, newref);
58 else
59 use->set_prec(j, newref);
60 nreplacements++;
61 } else if (j >= req && uin == NULL) {
62 break;
63 }
64 }
65 return nreplacements;
66 }
68 void PhaseMacroExpand::copy_call_debug_info(CallNode *oldcall, CallNode * newcall) {
69 // Copy debug information and adjust JVMState information
70 uint old_dbg_start = oldcall->tf()->domain()->cnt();
71 uint new_dbg_start = newcall->tf()->domain()->cnt();
72 int jvms_adj = new_dbg_start - old_dbg_start;
73 assert (new_dbg_start == newcall->req(), "argument count mismatch");
75 Dict* sosn_map = new Dict(cmpkey,hashkey);
76 for (uint i = old_dbg_start; i < oldcall->req(); i++) {
77 Node* old_in = oldcall->in(i);
78 // Clone old SafePointScalarObjectNodes, adjusting their field contents.
79 if (old_in != NULL && old_in->is_SafePointScalarObject()) {
80 SafePointScalarObjectNode* old_sosn = old_in->as_SafePointScalarObject();
81 uint old_unique = C->unique();
82 Node* new_in = old_sosn->clone(jvms_adj, sosn_map);
83 if (old_unique != C->unique()) {
84 new_in->set_req(0, C->root()); // reset control edge
85 new_in = transform_later(new_in); // Register new node.
86 }
87 old_in = new_in;
88 }
89 newcall->add_req(old_in);
90 }
92 newcall->set_jvms(oldcall->jvms());
93 for (JVMState *jvms = newcall->jvms(); jvms != NULL; jvms = jvms->caller()) {
94 jvms->set_map(newcall);
95 jvms->set_locoff(jvms->locoff()+jvms_adj);
96 jvms->set_stkoff(jvms->stkoff()+jvms_adj);
97 jvms->set_monoff(jvms->monoff()+jvms_adj);
98 jvms->set_scloff(jvms->scloff()+jvms_adj);
99 jvms->set_endoff(jvms->endoff()+jvms_adj);
100 }
101 }
103 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word, int mask, int bits, bool return_fast_path) {
104 Node* cmp;
105 if (mask != 0) {
106 Node* and_node = transform_later(new (C) AndXNode(word, MakeConX(mask)));
107 cmp = transform_later(new (C) CmpXNode(and_node, MakeConX(bits)));
108 } else {
109 cmp = word;
110 }
111 Node* bol = transform_later(new (C) BoolNode(cmp, BoolTest::ne));
112 IfNode* iff = new (C) IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN );
113 transform_later(iff);
115 // Fast path taken.
116 Node *fast_taken = transform_later( new (C) IfFalseNode(iff) );
118 // Fast path not-taken, i.e. slow path
119 Node *slow_taken = transform_later( new (C) IfTrueNode(iff) );
121 if (return_fast_path) {
122 region->init_req(edge, slow_taken); // Capture slow-control
123 return fast_taken;
124 } else {
125 region->init_req(edge, fast_taken); // Capture fast-control
126 return slow_taken;
127 }
128 }
130 //--------------------copy_predefined_input_for_runtime_call--------------------
131 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) {
132 // Set fixed predefined input arguments
133 call->init_req( TypeFunc::Control, ctrl );
134 call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) );
135 call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ?????
136 call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) );
137 call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) );
138 }
140 //------------------------------make_slow_call---------------------------------
141 CallNode* PhaseMacroExpand::make_slow_call(CallNode *oldcall, const TypeFunc* slow_call_type, address slow_call, const char* leaf_name, Node* slow_path, Node* parm0, Node* parm1) {
143 // Slow-path call
144 CallNode *call = leaf_name
145 ? (CallNode*)new (C) CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM )
146 : (CallNode*)new (C) CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), oldcall->jvms()->bci(), TypeRawPtr::BOTTOM );
148 // Slow path call has no side-effects, uses few values
149 copy_predefined_input_for_runtime_call(slow_path, oldcall, call );
150 if (parm0 != NULL) call->init_req(TypeFunc::Parms+0, parm0);
151 if (parm1 != NULL) call->init_req(TypeFunc::Parms+1, parm1);
152 copy_call_debug_info(oldcall, call);
153 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
154 _igvn.replace_node(oldcall, call);
155 transform_later(call);
157 return call;
158 }
160 void PhaseMacroExpand::extract_call_projections(CallNode *call) {
161 _fallthroughproj = NULL;
162 _fallthroughcatchproj = NULL;
163 _ioproj_fallthrough = NULL;
164 _ioproj_catchall = NULL;
165 _catchallcatchproj = NULL;
166 _memproj_fallthrough = NULL;
167 _memproj_catchall = NULL;
168 _resproj = NULL;
169 for (DUIterator_Fast imax, i = call->fast_outs(imax); i < imax; i++) {
170 ProjNode *pn = call->fast_out(i)->as_Proj();
171 switch (pn->_con) {
172 case TypeFunc::Control:
173 {
174 // For Control (fallthrough) and I_O (catch_all_index) we have CatchProj -> Catch -> Proj
175 _fallthroughproj = pn;
176 DUIterator_Fast jmax, j = pn->fast_outs(jmax);
177 const Node *cn = pn->fast_out(j);
178 if (cn->is_Catch()) {
179 ProjNode *cpn = NULL;
180 for (DUIterator_Fast kmax, k = cn->fast_outs(kmax); k < kmax; k++) {
181 cpn = cn->fast_out(k)->as_Proj();
182 assert(cpn->is_CatchProj(), "must be a CatchProjNode");
183 if (cpn->_con == CatchProjNode::fall_through_index)
184 _fallthroughcatchproj = cpn;
185 else {
186 assert(cpn->_con == CatchProjNode::catch_all_index, "must be correct index.");
187 _catchallcatchproj = cpn;
188 }
189 }
190 }
191 break;
192 }
193 case TypeFunc::I_O:
194 if (pn->_is_io_use)
195 _ioproj_catchall = pn;
196 else
197 _ioproj_fallthrough = pn;
198 break;
199 case TypeFunc::Memory:
200 if (pn->_is_io_use)
201 _memproj_catchall = pn;
202 else
203 _memproj_fallthrough = pn;
204 break;
205 case TypeFunc::Parms:
206 _resproj = pn;
207 break;
208 default:
209 assert(false, "unexpected projection from allocation node.");
210 }
211 }
213 }
215 // Eliminate a card mark sequence. p2x is a ConvP2XNode
216 void PhaseMacroExpand::eliminate_card_mark(Node* p2x) {
217 assert(p2x->Opcode() == Op_CastP2X, "ConvP2XNode required");
218 if (!UseG1GC) {
219 // vanilla/CMS post barrier
220 Node *shift = p2x->unique_out();
221 Node *addp = shift->unique_out();
222 for (DUIterator_Last jmin, j = addp->last_outs(jmin); j >= jmin; --j) {
223 Node *mem = addp->last_out(j);
224 if (UseCondCardMark && mem->is_Load()) {
225 assert(mem->Opcode() == Op_LoadB, "unexpected code shape");
226 // The load is checking if the card has been written so
227 // replace it with zero to fold the test.
228 _igvn.replace_node(mem, intcon(0));
229 continue;
230 }
231 assert(mem->is_Store(), "store required");
232 _igvn.replace_node(mem, mem->in(MemNode::Memory));
233 }
234 } else {
235 // G1 pre/post barriers
236 assert(p2x->outcnt() <= 2, "expects 1 or 2 users: Xor and URShift nodes");
237 // It could be only one user, URShift node, in Object.clone() instrinsic
238 // but the new allocation is passed to arraycopy stub and it could not
239 // be scalar replaced. So we don't check the case.
241 // An other case of only one user (Xor) is when the value check for NULL
242 // in G1 post barrier is folded after CCP so the code which used URShift
243 // is removed.
245 // Take Region node before eliminating post barrier since it also
246 // eliminates CastP2X node when it has only one user.
247 Node* this_region = p2x->in(0);
248 assert(this_region != NULL, "");
250 // Remove G1 post barrier.
252 // Search for CastP2X->Xor->URShift->Cmp path which
253 // checks if the store done to a different from the value's region.
254 // And replace Cmp with #0 (false) to collapse G1 post barrier.
255 Node* xorx = NULL;
256 for (DUIterator_Fast imax, i = p2x->fast_outs(imax); i < imax; i++) {
257 Node* u = p2x->fast_out(i);
258 if (u->Opcode() == Op_XorX) {
259 xorx = u;
260 break;
261 }
262 }
263 assert(xorx != NULL, "missing G1 post barrier");
264 Node* shift = xorx->unique_out();
265 Node* cmpx = shift->unique_out();
266 assert(cmpx->is_Cmp() && cmpx->unique_out()->is_Bool() &&
267 cmpx->unique_out()->as_Bool()->_test._test == BoolTest::ne,
268 "missing region check in G1 post barrier");
269 _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ));
271 // Remove G1 pre barrier.
273 // Search "if (marking != 0)" check and set it to "false".
274 // There is no G1 pre barrier if previous stored value is NULL
275 // (for example, after initialization).
276 if (this_region->is_Region() && this_region->req() == 3) {
277 int ind = 1;
278 if (!this_region->in(ind)->is_IfFalse()) {
279 ind = 2;
280 }
281 if (this_region->in(ind)->is_IfFalse()) {
282 Node* bol = this_region->in(ind)->in(0)->in(1);
283 assert(bol->is_Bool(), "");
284 cmpx = bol->in(1);
285 if (bol->as_Bool()->_test._test == BoolTest::ne &&
286 cmpx->is_Cmp() && cmpx->in(2) == intcon(0) &&
287 cmpx->in(1)->is_Load()) {
288 Node* adr = cmpx->in(1)->as_Load()->in(MemNode::Address);
289 const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() +
290 PtrQueue::byte_offset_of_active());
291 if (adr->is_AddP() && adr->in(AddPNode::Base) == top() &&
292 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
293 adr->in(AddPNode::Offset) == MakeConX(marking_offset)) {
294 _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ));
295 }
296 }
297 }
298 }
299 // Now CastP2X can be removed since it is used only on dead path
300 // which currently still alive until igvn optimize it.
301 assert(p2x->outcnt() == 0 || p2x->unique_out()->Opcode() == Op_URShiftX, "");
302 _igvn.replace_node(p2x, top());
303 }
304 }
306 // Search for a memory operation for the specified memory slice.
307 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) {
308 Node *orig_mem = mem;
309 Node *alloc_mem = alloc->in(TypeFunc::Memory);
310 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr();
311 while (true) {
312 if (mem == alloc_mem || mem == start_mem ) {
313 return mem; // hit one of our sentinels
314 } else if (mem->is_MergeMem()) {
315 mem = mem->as_MergeMem()->memory_at(alias_idx);
316 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) {
317 Node *in = mem->in(0);
318 // we can safely skip over safepoints, calls, locks and membars because we
319 // already know that the object is safe to eliminate.
320 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) {
321 return in;
322 } else if (in->is_Call()) {
323 CallNode *call = in->as_Call();
324 if (!call->may_modify(tinst, phase)) {
325 mem = call->in(TypeFunc::Memory);
326 }
327 mem = in->in(TypeFunc::Memory);
328 } else if (in->is_MemBar()) {
329 mem = in->in(TypeFunc::Memory);
330 } else {
331 assert(false, "unexpected projection");
332 }
333 } else if (mem->is_Store()) {
334 const TypePtr* atype = mem->as_Store()->adr_type();
335 int adr_idx = Compile::current()->get_alias_index(atype);
336 if (adr_idx == alias_idx) {
337 assert(atype->isa_oopptr(), "address type must be oopptr");
338 int adr_offset = atype->offset();
339 uint adr_iid = atype->is_oopptr()->instance_id();
340 // Array elements references have the same alias_idx
341 // but different offset and different instance_id.
342 if (adr_offset == offset && adr_iid == alloc->_idx)
343 return mem;
344 } else {
345 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw");
346 }
347 mem = mem->in(MemNode::Memory);
348 } else if (mem->is_ClearArray()) {
349 if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) {
350 // Can not bypass initialization of the instance
351 // we are looking.
352 debug_only(intptr_t offset;)
353 assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity");
354 InitializeNode* init = alloc->as_Allocate()->initialization();
355 // We are looking for stored value, return Initialize node
356 // or memory edge from Allocate node.
357 if (init != NULL)
358 return init;
359 else
360 return alloc->in(TypeFunc::Memory); // It will produce zero value (see callers).
361 }
362 // Otherwise skip it (the call updated 'mem' value).
363 } else if (mem->Opcode() == Op_SCMemProj) {
364 mem = mem->in(0);
365 Node* adr = NULL;
366 if (mem->is_LoadStore()) {
367 adr = mem->in(MemNode::Address);
368 } else {
369 assert(mem->Opcode() == Op_EncodeISOArray, "sanity");
370 adr = mem->in(3); // Destination array
371 }
372 const TypePtr* atype = adr->bottom_type()->is_ptr();
373 int adr_idx = Compile::current()->get_alias_index(atype);
374 if (adr_idx == alias_idx) {
375 assert(false, "Object is not scalar replaceable if a LoadStore node access its field");
376 return NULL;
377 }
378 mem = mem->in(MemNode::Memory);
379 } else {
380 return mem;
381 }
382 assert(mem != orig_mem, "dead memory loop");
383 }
384 }
386 //
387 // Given a Memory Phi, compute a value Phi containing the values from stores
388 // on the input paths.
389 // Note: this function is recursive, its depth is limied by the "level" argument
390 // Returns the computed Phi, or NULL if it cannot compute it.
391 Node *PhaseMacroExpand::value_from_mem_phi(Node *mem, BasicType ft, const Type *phi_type, const TypeOopPtr *adr_t, Node *alloc, Node_Stack *value_phis, int level) {
392 assert(mem->is_Phi(), "sanity");
393 int alias_idx = C->get_alias_index(adr_t);
394 int offset = adr_t->offset();
395 int instance_id = adr_t->instance_id();
397 // Check if an appropriate value phi already exists.
398 Node* region = mem->in(0);
399 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) {
400 Node* phi = region->fast_out(k);
401 if (phi->is_Phi() && phi != mem &&
402 phi->as_Phi()->is_same_inst_field(phi_type, instance_id, alias_idx, offset)) {
403 return phi;
404 }
405 }
406 // Check if an appropriate new value phi already exists.
407 Node* new_phi = value_phis->find(mem->_idx);
408 if (new_phi != NULL)
409 return new_phi;
411 if (level <= 0) {
412 return NULL; // Give up: phi tree too deep
413 }
414 Node *start_mem = C->start()->proj_out(TypeFunc::Memory);
415 Node *alloc_mem = alloc->in(TypeFunc::Memory);
417 uint length = mem->req();
418 GrowableArray <Node *> values(length, length, NULL, false);
420 // create a new Phi for the value
421 PhiNode *phi = new (C) PhiNode(mem->in(0), phi_type, NULL, instance_id, alias_idx, offset);
422 transform_later(phi);
423 value_phis->push(phi, mem->_idx);
425 for (uint j = 1; j < length; j++) {
426 Node *in = mem->in(j);
427 if (in == NULL || in->is_top()) {
428 values.at_put(j, in);
429 } else {
430 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn);
431 if (val == start_mem || val == alloc_mem) {
432 // hit a sentinel, return appropriate 0 value
433 values.at_put(j, _igvn.zerocon(ft));
434 continue;
435 }
436 if (val->is_Initialize()) {
437 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
438 }
439 if (val == NULL) {
440 return NULL; // can't find a value on this path
441 }
442 if (val == mem) {
443 values.at_put(j, mem);
444 } else if (val->is_Store()) {
445 values.at_put(j, val->in(MemNode::ValueIn));
446 } else if(val->is_Proj() && val->in(0) == alloc) {
447 values.at_put(j, _igvn.zerocon(ft));
448 } else if (val->is_Phi()) {
449 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1);
450 if (val == NULL) {
451 return NULL;
452 }
453 values.at_put(j, val);
454 } else if (val->Opcode() == Op_SCMemProj) {
455 assert(val->in(0)->is_LoadStore() || val->in(0)->Opcode() == Op_EncodeISOArray, "sanity");
456 assert(false, "Object is not scalar replaceable if a LoadStore node access its field");
457 return NULL;
458 } else {
459 #ifdef ASSERT
460 val->dump();
461 assert(false, "unknown node on this path");
462 #endif
463 return NULL; // unknown node on this path
464 }
465 }
466 }
467 // Set Phi's inputs
468 for (uint j = 1; j < length; j++) {
469 if (values.at(j) == mem) {
470 phi->init_req(j, phi);
471 } else {
472 phi->init_req(j, values.at(j));
473 }
474 }
475 return phi;
476 }
478 // Search the last value stored into the object's field.
479 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, Node *alloc) {
480 assert(adr_t->is_known_instance_field(), "instance required");
481 int instance_id = adr_t->instance_id();
482 assert((uint)instance_id == alloc->_idx, "wrong allocation");
484 int alias_idx = C->get_alias_index(adr_t);
485 int offset = adr_t->offset();
486 Node *start_mem = C->start()->proj_out(TypeFunc::Memory);
487 Node *alloc_ctrl = alloc->in(TypeFunc::Control);
488 Node *alloc_mem = alloc->in(TypeFunc::Memory);
489 Arena *a = Thread::current()->resource_area();
490 VectorSet visited(a);
493 bool done = sfpt_mem == alloc_mem;
494 Node *mem = sfpt_mem;
495 while (!done) {
496 if (visited.test_set(mem->_idx)) {
497 return NULL; // found a loop, give up
498 }
499 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn);
500 if (mem == start_mem || mem == alloc_mem) {
501 done = true; // hit a sentinel, return appropriate 0 value
502 } else if (mem->is_Initialize()) {
503 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
504 if (mem == NULL) {
505 done = true; // Something go wrong.
506 } else if (mem->is_Store()) {
507 const TypePtr* atype = mem->as_Store()->adr_type();
508 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice");
509 done = true;
510 }
511 } else if (mem->is_Store()) {
512 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr();
513 assert(atype != NULL, "address type must be oopptr");
514 assert(C->get_alias_index(atype) == alias_idx &&
515 atype->is_known_instance_field() && atype->offset() == offset &&
516 atype->instance_id() == instance_id, "store is correct memory slice");
517 done = true;
518 } else if (mem->is_Phi()) {
519 // try to find a phi's unique input
520 Node *unique_input = NULL;
521 Node *top = C->top();
522 for (uint i = 1; i < mem->req(); i++) {
523 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn);
524 if (n == NULL || n == top || n == mem) {
525 continue;
526 } else if (unique_input == NULL) {
527 unique_input = n;
528 } else if (unique_input != n) {
529 unique_input = top;
530 break;
531 }
532 }
533 if (unique_input != NULL && unique_input != top) {
534 mem = unique_input;
535 } else {
536 done = true;
537 }
538 } else {
539 assert(false, "unexpected node");
540 }
541 }
542 if (mem != NULL) {
543 if (mem == start_mem || mem == alloc_mem) {
544 // hit a sentinel, return appropriate 0 value
545 return _igvn.zerocon(ft);
546 } else if (mem->is_Store()) {
547 return mem->in(MemNode::ValueIn);
548 } else if (mem->is_Phi()) {
549 // attempt to produce a Phi reflecting the values on the input paths of the Phi
550 Node_Stack value_phis(a, 8);
551 Node * phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit);
552 if (phi != NULL) {
553 return phi;
554 } else {
555 // Kill all new Phis
556 while(value_phis.is_nonempty()) {
557 Node* n = value_phis.node();
558 _igvn.replace_node(n, C->top());
559 value_phis.pop();
560 }
561 }
562 }
563 }
564 // Something go wrong.
565 return NULL;
566 }
568 // Check the possibility of scalar replacement.
569 bool PhaseMacroExpand::can_eliminate_allocation(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
570 // Scan the uses of the allocation to check for anything that would
571 // prevent us from eliminating it.
572 NOT_PRODUCT( const char* fail_eliminate = NULL; )
573 DEBUG_ONLY( Node* disq_node = NULL; )
574 bool can_eliminate = true;
576 Node* res = alloc->result_cast();
577 const TypeOopPtr* res_type = NULL;
578 if (res == NULL) {
579 // All users were eliminated.
580 } else if (!res->is_CheckCastPP()) {
581 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";)
582 can_eliminate = false;
583 } else {
584 res_type = _igvn.type(res)->isa_oopptr();
585 if (res_type == NULL) {
586 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";)
587 can_eliminate = false;
588 } else if (res_type->isa_aryptr()) {
589 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1);
590 if (length < 0) {
591 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";)
592 can_eliminate = false;
593 }
594 }
595 }
597 if (can_eliminate && res != NULL) {
598 for (DUIterator_Fast jmax, j = res->fast_outs(jmax);
599 j < jmax && can_eliminate; j++) {
600 Node* use = res->fast_out(j);
602 if (use->is_AddP()) {
603 const TypePtr* addp_type = _igvn.type(use)->is_ptr();
604 int offset = addp_type->offset();
606 if (offset == Type::OffsetTop || offset == Type::OffsetBot) {
607 NOT_PRODUCT(fail_eliminate = "Undefined field referrence";)
608 can_eliminate = false;
609 break;
610 }
611 for (DUIterator_Fast kmax, k = use->fast_outs(kmax);
612 k < kmax && can_eliminate; k++) {
613 Node* n = use->fast_out(k);
614 if (!n->is_Store() && n->Opcode() != Op_CastP2X) {
615 DEBUG_ONLY(disq_node = n;)
616 if (n->is_Load() || n->is_LoadStore()) {
617 NOT_PRODUCT(fail_eliminate = "Field load";)
618 } else {
619 NOT_PRODUCT(fail_eliminate = "Not store field referrence";)
620 }
621 can_eliminate = false;
622 }
623 }
624 } else if (use->is_SafePoint()) {
625 SafePointNode* sfpt = use->as_SafePoint();
626 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) {
627 // Object is passed as argument.
628 DEBUG_ONLY(disq_node = use;)
629 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";)
630 can_eliminate = false;
631 }
632 Node* sfptMem = sfpt->memory();
633 if (sfptMem == NULL || sfptMem->is_top()) {
634 DEBUG_ONLY(disq_node = use;)
635 NOT_PRODUCT(fail_eliminate = "NULL or TOP memory";)
636 can_eliminate = false;
637 } else {
638 safepoints.append_if_missing(sfpt);
639 }
640 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
641 if (use->is_Phi()) {
642 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
643 NOT_PRODUCT(fail_eliminate = "Object is return value";)
644 } else {
645 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
646 }
647 DEBUG_ONLY(disq_node = use;)
648 } else {
649 if (use->Opcode() == Op_Return) {
650 NOT_PRODUCT(fail_eliminate = "Object is return value";)
651 }else {
652 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
653 }
654 DEBUG_ONLY(disq_node = use;)
655 }
656 can_eliminate = false;
657 }
658 }
659 }
661 #ifndef PRODUCT
662 if (PrintEliminateAllocations) {
663 if (can_eliminate) {
664 tty->print("Scalar ");
665 if (res == NULL)
666 alloc->dump();
667 else
668 res->dump();
669 } else {
670 tty->print("NotScalar (%s)", fail_eliminate);
671 if (res == NULL)
672 alloc->dump();
673 else
674 res->dump();
675 #ifdef ASSERT
676 if (disq_node != NULL) {
677 tty->print(" >>>> ");
678 disq_node->dump();
679 }
680 #endif /*ASSERT*/
681 }
682 }
683 #endif
684 return can_eliminate;
685 }
687 // Do scalar replacement.
688 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
689 GrowableArray <SafePointNode *> safepoints_done;
691 ciKlass* klass = NULL;
692 ciInstanceKlass* iklass = NULL;
693 int nfields = 0;
694 int array_base;
695 int element_size;
696 BasicType basic_elem_type;
697 ciType* elem_type;
699 Node* res = alloc->result_cast();
700 const TypeOopPtr* res_type = NULL;
701 if (res != NULL) { // Could be NULL when there are no users
702 res_type = _igvn.type(res)->isa_oopptr();
703 }
705 if (res != NULL) {
706 klass = res_type->klass();
707 if (res_type->isa_instptr()) {
708 // find the fields of the class which will be needed for safepoint debug information
709 assert(klass->is_instance_klass(), "must be an instance klass.");
710 iklass = klass->as_instance_klass();
711 nfields = iklass->nof_nonstatic_fields();
712 } else {
713 // find the array's elements which will be needed for safepoint debug information
714 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
715 assert(klass->is_array_klass() && nfields >= 0, "must be an array klass.");
716 elem_type = klass->as_array_klass()->element_type();
717 basic_elem_type = elem_type->basic_type();
718 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
719 element_size = type2aelembytes(basic_elem_type);
720 }
721 }
722 //
723 // Process the safepoint uses
724 //
725 while (safepoints.length() > 0) {
726 SafePointNode* sfpt = safepoints.pop();
727 Node* mem = sfpt->memory();
728 uint first_ind = sfpt->req();
729 SafePointScalarObjectNode* sobj = new (C) SafePointScalarObjectNode(res_type,
730 #ifdef ASSERT
731 alloc,
732 #endif
733 first_ind, nfields);
734 sobj->init_req(0, C->root());
735 transform_later(sobj);
737 // Scan object's fields adding an input to the safepoint for each field.
738 for (int j = 0; j < nfields; j++) {
739 intptr_t offset;
740 ciField* field = NULL;
741 if (iklass != NULL) {
742 field = iklass->nonstatic_field_at(j);
743 offset = field->offset();
744 elem_type = field->type();
745 basic_elem_type = field->layout_type();
746 } else {
747 offset = array_base + j * (intptr_t)element_size;
748 }
750 const Type *field_type;
751 // The next code is taken from Parse::do_get_xxx().
752 if (basic_elem_type == T_OBJECT || basic_elem_type == T_ARRAY) {
753 if (!elem_type->is_loaded()) {
754 field_type = TypeInstPtr::BOTTOM;
755 } else if (field != NULL && field->is_constant() && field->is_static()) {
756 // This can happen if the constant oop is non-perm.
757 ciObject* con = field->constant_value().as_object();
758 // Do not "join" in the previous type; it doesn't add value,
759 // and may yield a vacuous result if the field is of interface type.
760 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr();
761 assert(field_type != NULL, "field singleton type must be consistent");
762 } else {
763 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
764 }
765 if (UseCompressedOops) {
766 field_type = field_type->make_narrowoop();
767 basic_elem_type = T_NARROWOOP;
768 }
769 } else {
770 field_type = Type::get_const_basic_type(basic_elem_type);
771 }
773 const TypeOopPtr *field_addr_type = res_type->add_offset(offset)->isa_oopptr();
775 Node *field_val = value_from_mem(mem, basic_elem_type, field_type, field_addr_type, alloc);
776 if (field_val == NULL) {
777 // We weren't able to find a value for this field,
778 // give up on eliminating this allocation.
780 // Remove any extra entries we added to the safepoint.
781 uint last = sfpt->req() - 1;
782 for (int k = 0; k < j; k++) {
783 sfpt->del_req(last--);
784 }
785 // rollback processed safepoints
786 while (safepoints_done.length() > 0) {
787 SafePointNode* sfpt_done = safepoints_done.pop();
788 // remove any extra entries we added to the safepoint
789 last = sfpt_done->req() - 1;
790 for (int k = 0; k < nfields; k++) {
791 sfpt_done->del_req(last--);
792 }
793 JVMState *jvms = sfpt_done->jvms();
794 jvms->set_endoff(sfpt_done->req());
795 // Now make a pass over the debug information replacing any references
796 // to SafePointScalarObjectNode with the allocated object.
797 int start = jvms->debug_start();
798 int end = jvms->debug_end();
799 for (int i = start; i < end; i++) {
800 if (sfpt_done->in(i)->is_SafePointScalarObject()) {
801 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject();
802 if (scobj->first_index() == sfpt_done->req() &&
803 scobj->n_fields() == (uint)nfields) {
804 assert(scobj->alloc() == alloc, "sanity");
805 sfpt_done->set_req(i, res);
806 }
807 }
808 }
809 }
810 #ifndef PRODUCT
811 if (PrintEliminateAllocations) {
812 if (field != NULL) {
813 tty->print("=== At SafePoint node %d can't find value of Field: ",
814 sfpt->_idx);
815 field->print();
816 int field_idx = C->get_alias_index(field_addr_type);
817 tty->print(" (alias_idx=%d)", field_idx);
818 } else { // Array's element
819 tty->print("=== At SafePoint node %d can't find value of array element [%d]",
820 sfpt->_idx, j);
821 }
822 tty->print(", which prevents elimination of: ");
823 if (res == NULL)
824 alloc->dump();
825 else
826 res->dump();
827 }
828 #endif
829 return false;
830 }
831 if (UseCompressedOops && field_type->isa_narrowoop()) {
832 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
833 // to be able scalar replace the allocation.
834 if (field_val->is_EncodeP()) {
835 field_val = field_val->in(1);
836 } else {
837 field_val = transform_later(new (C) DecodeNNode(field_val, field_val->bottom_type()->make_ptr()));
838 }
839 }
840 sfpt->add_req(field_val);
841 }
842 JVMState *jvms = sfpt->jvms();
843 jvms->set_endoff(sfpt->req());
844 // Now make a pass over the debug information replacing any references
845 // to the allocated object with "sobj"
846 int start = jvms->debug_start();
847 int end = jvms->debug_end();
848 for (int i = start; i < end; i++) {
849 if (sfpt->in(i) == res) {
850 sfpt->set_req(i, sobj);
851 }
852 }
853 safepoints_done.append_if_missing(sfpt); // keep it for rollback
854 }
855 return true;
856 }
858 // Process users of eliminated allocation.
859 void PhaseMacroExpand::process_users_of_allocation(AllocateNode *alloc) {
860 Node* res = alloc->result_cast();
861 if (res != NULL) {
862 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
863 Node *use = res->last_out(j);
864 uint oc1 = res->outcnt();
866 if (use->is_AddP()) {
867 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) {
868 Node *n = use->last_out(k);
869 uint oc2 = use->outcnt();
870 if (n->is_Store()) {
871 #ifdef ASSERT
872 // Verify that there is no dependent MemBarVolatile nodes,
873 // they should be removed during IGVN, see MemBarNode::Ideal().
874 for (DUIterator_Fast pmax, p = n->fast_outs(pmax);
875 p < pmax; p++) {
876 Node* mb = n->fast_out(p);
877 assert(mb->is_Initialize() || !mb->is_MemBar() ||
878 mb->req() <= MemBarNode::Precedent ||
879 mb->in(MemBarNode::Precedent) != n,
880 "MemBarVolatile should be eliminated for non-escaping object");
881 }
882 #endif
883 _igvn.replace_node(n, n->in(MemNode::Memory));
884 } else {
885 eliminate_card_mark(n);
886 }
887 k -= (oc2 - use->outcnt());
888 }
889 } else {
890 eliminate_card_mark(use);
891 }
892 j -= (oc1 - res->outcnt());
893 }
894 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
895 _igvn.remove_dead_node(res);
896 }
898 //
899 // Process other users of allocation's projections
900 //
901 if (_resproj != NULL && _resproj->outcnt() != 0) {
902 for (DUIterator_Last jmin, j = _resproj->last_outs(jmin); j >= jmin; ) {
903 Node *use = _resproj->last_out(j);
904 uint oc1 = _resproj->outcnt();
905 if (use->is_Initialize()) {
906 // Eliminate Initialize node.
907 InitializeNode *init = use->as_Initialize();
908 assert(init->outcnt() <= 2, "only a control and memory projection expected");
909 Node *ctrl_proj = init->proj_out(TypeFunc::Control);
910 if (ctrl_proj != NULL) {
911 assert(init->in(TypeFunc::Control) == _fallthroughcatchproj, "allocation control projection");
912 _igvn.replace_node(ctrl_proj, _fallthroughcatchproj);
913 }
914 Node *mem_proj = init->proj_out(TypeFunc::Memory);
915 if (mem_proj != NULL) {
916 Node *mem = init->in(TypeFunc::Memory);
917 #ifdef ASSERT
918 if (mem->is_MergeMem()) {
919 assert(mem->in(TypeFunc::Memory) == _memproj_fallthrough, "allocation memory projection");
920 } else {
921 assert(mem == _memproj_fallthrough, "allocation memory projection");
922 }
923 #endif
924 _igvn.replace_node(mem_proj, mem);
925 }
926 } else if (use->is_AddP()) {
927 // raw memory addresses used only by the initialization
928 _igvn.replace_node(use, C->top());
929 } else {
930 assert(false, "only Initialize or AddP expected");
931 }
932 j -= (oc1 - _resproj->outcnt());
933 }
934 }
935 if (_fallthroughcatchproj != NULL) {
936 _igvn.replace_node(_fallthroughcatchproj, alloc->in(TypeFunc::Control));
937 }
938 if (_memproj_fallthrough != NULL) {
939 _igvn.replace_node(_memproj_fallthrough, alloc->in(TypeFunc::Memory));
940 }
941 if (_memproj_catchall != NULL) {
942 _igvn.replace_node(_memproj_catchall, C->top());
943 }
944 if (_ioproj_fallthrough != NULL) {
945 _igvn.replace_node(_ioproj_fallthrough, alloc->in(TypeFunc::I_O));
946 }
947 if (_ioproj_catchall != NULL) {
948 _igvn.replace_node(_ioproj_catchall, C->top());
949 }
950 if (_catchallcatchproj != NULL) {
951 _igvn.replace_node(_catchallcatchproj, C->top());
952 }
953 }
955 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) {
957 if (!EliminateAllocations || !alloc->_is_scalar_replaceable) {
958 return false;
959 }
961 extract_call_projections(alloc);
963 GrowableArray <SafePointNode *> safepoints;
964 if (!can_eliminate_allocation(alloc, safepoints)) {
965 return false;
966 }
968 if (!scalar_replacement(alloc, safepoints)) {
969 return false;
970 }
972 CompileLog* log = C->log();
973 if (log != NULL) {
974 Node* klass = alloc->in(AllocateNode::KlassNode);
975 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr();
976 log->head("eliminate_allocation type='%d'",
977 log->identify(tklass->klass()));
978 JVMState* p = alloc->jvms();
979 while (p != NULL) {
980 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
981 p = p->caller();
982 }
983 log->tail("eliminate_allocation");
984 }
986 process_users_of_allocation(alloc);
988 #ifndef PRODUCT
989 if (PrintEliminateAllocations) {
990 if (alloc->is_AllocateArray())
991 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
992 else
993 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
994 }
995 #endif
997 return true;
998 }
1001 //---------------------------set_eden_pointers-------------------------
1002 void PhaseMacroExpand::set_eden_pointers(Node* &eden_top_adr, Node* &eden_end_adr) {
1003 if (UseTLAB) { // Private allocation: load from TLS
1004 Node* thread = transform_later(new (C) ThreadLocalNode());
1005 int tlab_top_offset = in_bytes(JavaThread::tlab_top_offset());
1006 int tlab_end_offset = in_bytes(JavaThread::tlab_end_offset());
1007 eden_top_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_top_offset);
1008 eden_end_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_end_offset);
1009 } else { // Shared allocation: load from globals
1010 CollectedHeap* ch = Universe::heap();
1011 address top_adr = (address)ch->top_addr();
1012 address end_adr = (address)ch->end_addr();
1013 eden_top_adr = makecon(TypeRawPtr::make(top_adr));
1014 eden_end_adr = basic_plus_adr(eden_top_adr, end_adr - top_adr);
1015 }
1016 }
1019 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) {
1020 Node* adr = basic_plus_adr(base, offset);
1021 const TypePtr* adr_type = adr->bottom_type()->is_ptr();
1022 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt);
1023 transform_later(value);
1024 return value;
1025 }
1028 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) {
1029 Node* adr = basic_plus_adr(base, offset);
1030 mem = StoreNode::make(_igvn, ctl, mem, adr, NULL, value, bt);
1031 transform_later(mem);
1032 return mem;
1033 }
1035 //=============================================================================
1036 //
1037 // A L L O C A T I O N
1038 //
1039 // Allocation attempts to be fast in the case of frequent small objects.
1040 // It breaks down like this:
1041 //
1042 // 1) Size in doublewords is computed. This is a constant for objects and
1043 // variable for most arrays. Doubleword units are used to avoid size
1044 // overflow of huge doubleword arrays. We need doublewords in the end for
1045 // rounding.
1046 //
1047 // 2) Size is checked for being 'too large'. Too-large allocations will go
1048 // the slow path into the VM. The slow path can throw any required
1049 // exceptions, and does all the special checks for very large arrays. The
1050 // size test can constant-fold away for objects. For objects with
1051 // finalizers it constant-folds the otherway: you always go slow with
1052 // finalizers.
1053 //
1054 // 3) If NOT using TLABs, this is the contended loop-back point.
1055 // Load-Locked the heap top. If using TLABs normal-load the heap top.
1056 //
1057 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route.
1058 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish
1059 // "size*8" we always enter the VM, where "largish" is a constant picked small
1060 // enough that there's always space between the eden max and 4Gig (old space is
1061 // there so it's quite large) and large enough that the cost of entering the VM
1062 // is dwarfed by the cost to initialize the space.
1063 //
1064 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back
1065 // down. If contended, repeat at step 3. If using TLABs normal-store
1066 // adjusted heap top back down; there is no contention.
1067 //
1068 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark
1069 // fields.
1070 //
1071 // 7) Merge with the slow-path; cast the raw memory pointer to the correct
1072 // oop flavor.
1073 //
1074 //=============================================================================
1075 // FastAllocateSizeLimit value is in DOUBLEWORDS.
1076 // Allocations bigger than this always go the slow route.
1077 // This value must be small enough that allocation attempts that need to
1078 // trigger exceptions go the slow route. Also, it must be small enough so
1079 // that heap_top + size_in_bytes does not wrap around the 4Gig limit.
1080 //=============================================================================j//
1081 // %%% Here is an old comment from parseHelper.cpp; is it outdated?
1082 // The allocator will coalesce int->oop copies away. See comment in
1083 // coalesce.cpp about how this works. It depends critically on the exact
1084 // code shape produced here, so if you are changing this code shape
1085 // make sure the GC info for the heap-top is correct in and around the
1086 // slow-path call.
1087 //
1089 void PhaseMacroExpand::expand_allocate_common(
1090 AllocateNode* alloc, // allocation node to be expanded
1091 Node* length, // array length for an array allocation
1092 const TypeFunc* slow_call_type, // Type of slow call
1093 address slow_call_address // Address of slow call
1094 )
1095 {
1097 Node* ctrl = alloc->in(TypeFunc::Control);
1098 Node* mem = alloc->in(TypeFunc::Memory);
1099 Node* i_o = alloc->in(TypeFunc::I_O);
1100 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize);
1101 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1102 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest);
1104 Node* storestore = alloc->storestore();
1105 if (storestore != NULL) {
1106 // Break this link that is no longer useful and confuses register allocation
1107 storestore->set_req(MemBarNode::Precedent, top());
1108 }
1110 assert(ctrl != NULL, "must have control");
1111 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results.
1112 // they will not be used if "always_slow" is set
1113 enum { slow_result_path = 1, fast_result_path = 2 };
1114 Node *result_region;
1115 Node *result_phi_rawmem;
1116 Node *result_phi_rawoop;
1117 Node *result_phi_i_o;
1119 // The initial slow comparison is a size check, the comparison
1120 // we want to do is a BoolTest::gt
1121 bool always_slow = false;
1122 int tv = _igvn.find_int_con(initial_slow_test, -1);
1123 if (tv >= 0) {
1124 always_slow = (tv == 1);
1125 initial_slow_test = NULL;
1126 } else {
1127 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn);
1128 }
1130 if (C->env()->dtrace_alloc_probes() ||
1131 !UseTLAB && (!Universe::heap()->supports_inline_contig_alloc() ||
1132 (UseConcMarkSweepGC && CMSIncrementalMode))) {
1133 // Force slow-path allocation
1134 always_slow = true;
1135 initial_slow_test = NULL;
1136 }
1139 enum { too_big_or_final_path = 1, need_gc_path = 2 };
1140 Node *slow_region = NULL;
1141 Node *toobig_false = ctrl;
1143 assert (initial_slow_test == NULL || !always_slow, "arguments must be consistent");
1144 // generate the initial test if necessary
1145 if (initial_slow_test != NULL ) {
1146 slow_region = new (C) RegionNode(3);
1148 // Now make the initial failure test. Usually a too-big test but
1149 // might be a TRUE for finalizers or a fancy class check for
1150 // newInstance0.
1151 IfNode *toobig_iff = new (C) IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
1152 transform_later(toobig_iff);
1153 // Plug the failing-too-big test into the slow-path region
1154 Node *toobig_true = new (C) IfTrueNode( toobig_iff );
1155 transform_later(toobig_true);
1156 slow_region ->init_req( too_big_or_final_path, toobig_true );
1157 toobig_false = new (C) IfFalseNode( toobig_iff );
1158 transform_later(toobig_false);
1159 } else { // No initial test, just fall into next case
1160 toobig_false = ctrl;
1161 debug_only(slow_region = NodeSentinel);
1162 }
1164 Node *slow_mem = mem; // save the current memory state for slow path
1165 // generate the fast allocation code unless we know that the initial test will always go slow
1166 if (!always_slow) {
1167 // Fast path modifies only raw memory.
1168 if (mem->is_MergeMem()) {
1169 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1170 }
1172 Node* eden_top_adr;
1173 Node* eden_end_adr;
1175 set_eden_pointers(eden_top_adr, eden_end_adr);
1177 // Load Eden::end. Loop invariant and hoisted.
1178 //
1179 // Note: We set the control input on "eden_end" and "old_eden_top" when using
1180 // a TLAB to work around a bug where these values were being moved across
1181 // a safepoint. These are not oops, so they cannot be include in the oop
1182 // map, but they can be changed by a GC. The proper way to fix this would
1183 // be to set the raw memory state when generating a SafepointNode. However
1184 // this will require extensive changes to the loop optimization in order to
1185 // prevent a degradation of the optimization.
1186 // See comment in memnode.hpp, around line 227 in class LoadPNode.
1187 Node *eden_end = make_load(ctrl, mem, eden_end_adr, 0, TypeRawPtr::BOTTOM, T_ADDRESS);
1189 // allocate the Region and Phi nodes for the result
1190 result_region = new (C) RegionNode(3);
1191 result_phi_rawmem = new (C) PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1192 result_phi_rawoop = new (C) PhiNode(result_region, TypeRawPtr::BOTTOM);
1193 result_phi_i_o = new (C) PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch
1195 // We need a Region for the loop-back contended case.
1196 enum { fall_in_path = 1, contended_loopback_path = 2 };
1197 Node *contended_region;
1198 Node *contended_phi_rawmem;
1199 if (UseTLAB) {
1200 contended_region = toobig_false;
1201 contended_phi_rawmem = mem;
1202 } else {
1203 contended_region = new (C) RegionNode(3);
1204 contended_phi_rawmem = new (C) PhiNode(contended_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1205 // Now handle the passing-too-big test. We fall into the contended
1206 // loop-back merge point.
1207 contended_region ->init_req(fall_in_path, toobig_false);
1208 contended_phi_rawmem->init_req(fall_in_path, mem);
1209 transform_later(contended_region);
1210 transform_later(contended_phi_rawmem);
1211 }
1213 // Load(-locked) the heap top.
1214 // See note above concerning the control input when using a TLAB
1215 Node *old_eden_top = UseTLAB
1216 ? new (C) LoadPNode (ctrl, contended_phi_rawmem, eden_top_adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM)
1217 : new (C) LoadPLockedNode(contended_region, contended_phi_rawmem, eden_top_adr);
1219 transform_later(old_eden_top);
1220 // Add to heap top to get a new heap top
1221 Node *new_eden_top = new (C) AddPNode(top(), old_eden_top, size_in_bytes);
1222 transform_later(new_eden_top);
1223 // Check for needing a GC; compare against heap end
1224 Node *needgc_cmp = new (C) CmpPNode(new_eden_top, eden_end);
1225 transform_later(needgc_cmp);
1226 Node *needgc_bol = new (C) BoolNode(needgc_cmp, BoolTest::ge);
1227 transform_later(needgc_bol);
1228 IfNode *needgc_iff = new (C) IfNode(contended_region, needgc_bol, PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN);
1229 transform_later(needgc_iff);
1231 // Plug the failing-heap-space-need-gc test into the slow-path region
1232 Node *needgc_true = new (C) IfTrueNode(needgc_iff);
1233 transform_later(needgc_true);
1234 if (initial_slow_test) {
1235 slow_region->init_req(need_gc_path, needgc_true);
1236 // This completes all paths into the slow merge point
1237 transform_later(slow_region);
1238 } else { // No initial slow path needed!
1239 // Just fall from the need-GC path straight into the VM call.
1240 slow_region = needgc_true;
1241 }
1242 // No need for a GC. Setup for the Store-Conditional
1243 Node *needgc_false = new (C) IfFalseNode(needgc_iff);
1244 transform_later(needgc_false);
1246 // Grab regular I/O before optional prefetch may change it.
1247 // Slow-path does no I/O so just set it to the original I/O.
1248 result_phi_i_o->init_req(slow_result_path, i_o);
1250 i_o = prefetch_allocation(i_o, needgc_false, contended_phi_rawmem,
1251 old_eden_top, new_eden_top, length);
1253 // Name successful fast-path variables
1254 Node* fast_oop = old_eden_top;
1255 Node* fast_oop_ctrl;
1256 Node* fast_oop_rawmem;
1258 // Store (-conditional) the modified eden top back down.
1259 // StorePConditional produces flags for a test PLUS a modified raw
1260 // memory state.
1261 if (UseTLAB) {
1262 Node* store_eden_top =
1263 new (C) StorePNode(needgc_false, contended_phi_rawmem, eden_top_adr,
1264 TypeRawPtr::BOTTOM, new_eden_top);
1265 transform_later(store_eden_top);
1266 fast_oop_ctrl = needgc_false; // No contention, so this is the fast path
1267 fast_oop_rawmem = store_eden_top;
1268 } else {
1269 Node* store_eden_top =
1270 new (C) StorePConditionalNode(needgc_false, contended_phi_rawmem, eden_top_adr,
1271 new_eden_top, fast_oop/*old_eden_top*/);
1272 transform_later(store_eden_top);
1273 Node *contention_check = new (C) BoolNode(store_eden_top, BoolTest::ne);
1274 transform_later(contention_check);
1275 store_eden_top = new (C) SCMemProjNode(store_eden_top);
1276 transform_later(store_eden_top);
1278 // If not using TLABs, check to see if there was contention.
1279 IfNode *contention_iff = new (C) IfNode (needgc_false, contention_check, PROB_MIN, COUNT_UNKNOWN);
1280 transform_later(contention_iff);
1281 Node *contention_true = new (C) IfTrueNode(contention_iff);
1282 transform_later(contention_true);
1283 // If contention, loopback and try again.
1284 contended_region->init_req(contended_loopback_path, contention_true);
1285 contended_phi_rawmem->init_req(contended_loopback_path, store_eden_top);
1287 // Fast-path succeeded with no contention!
1288 Node *contention_false = new (C) IfFalseNode(contention_iff);
1289 transform_later(contention_false);
1290 fast_oop_ctrl = contention_false;
1292 // Bump total allocated bytes for this thread
1293 Node* thread = new (C) ThreadLocalNode();
1294 transform_later(thread);
1295 Node* alloc_bytes_adr = basic_plus_adr(top()/*not oop*/, thread,
1296 in_bytes(JavaThread::allocated_bytes_offset()));
1297 Node* alloc_bytes = make_load(fast_oop_ctrl, store_eden_top, alloc_bytes_adr,
1298 0, TypeLong::LONG, T_LONG);
1299 #ifdef _LP64
1300 Node* alloc_size = size_in_bytes;
1301 #else
1302 Node* alloc_size = new (C) ConvI2LNode(size_in_bytes);
1303 transform_later(alloc_size);
1304 #endif
1305 Node* new_alloc_bytes = new (C) AddLNode(alloc_bytes, alloc_size);
1306 transform_later(new_alloc_bytes);
1307 fast_oop_rawmem = make_store(fast_oop_ctrl, store_eden_top, alloc_bytes_adr,
1308 0, new_alloc_bytes, T_LONG);
1309 }
1311 InitializeNode* init = alloc->initialization();
1312 fast_oop_rawmem = initialize_object(alloc,
1313 fast_oop_ctrl, fast_oop_rawmem, fast_oop,
1314 klass_node, length, size_in_bytes);
1316 // If initialization is performed by an array copy, any required
1317 // MemBarStoreStore was already added. If the object does not
1318 // escape no need for a MemBarStoreStore. Otherwise we need a
1319 // MemBarStoreStore so that stores that initialize this object
1320 // can't be reordered with a subsequent store that makes this
1321 // object accessible by other threads.
1322 if (init == NULL || (!init->is_complete_with_arraycopy() && !init->does_not_escape())) {
1323 if (init == NULL || init->req() < InitializeNode::RawStores) {
1324 // No InitializeNode or no stores captured by zeroing
1325 // elimination. Simply add the MemBarStoreStore after object
1326 // initialization.
1327 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot, fast_oop_rawmem);
1328 transform_later(mb);
1330 mb->init_req(TypeFunc::Memory, fast_oop_rawmem);
1331 mb->init_req(TypeFunc::Control, fast_oop_ctrl);
1332 fast_oop_ctrl = new (C) ProjNode(mb,TypeFunc::Control);
1333 transform_later(fast_oop_ctrl);
1334 fast_oop_rawmem = new (C) ProjNode(mb,TypeFunc::Memory);
1335 transform_later(fast_oop_rawmem);
1336 } else {
1337 // Add the MemBarStoreStore after the InitializeNode so that
1338 // all stores performing the initialization that were moved
1339 // before the InitializeNode happen before the storestore
1340 // barrier.
1342 Node* init_ctrl = init->proj_out(TypeFunc::Control);
1343 Node* init_mem = init->proj_out(TypeFunc::Memory);
1345 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
1346 transform_later(mb);
1348 Node* ctrl = new (C) ProjNode(init,TypeFunc::Control);
1349 transform_later(ctrl);
1350 Node* mem = new (C) ProjNode(init,TypeFunc::Memory);
1351 transform_later(mem);
1353 // The MemBarStoreStore depends on control and memory coming
1354 // from the InitializeNode
1355 mb->init_req(TypeFunc::Memory, mem);
1356 mb->init_req(TypeFunc::Control, ctrl);
1358 ctrl = new (C) ProjNode(mb,TypeFunc::Control);
1359 transform_later(ctrl);
1360 mem = new (C) ProjNode(mb,TypeFunc::Memory);
1361 transform_later(mem);
1363 // All nodes that depended on the InitializeNode for control
1364 // and memory must now depend on the MemBarNode that itself
1365 // depends on the InitializeNode
1366 _igvn.replace_node(init_ctrl, ctrl);
1367 _igvn.replace_node(init_mem, mem);
1368 }
1369 }
1371 if (C->env()->dtrace_extended_probes()) {
1372 // Slow-path call
1373 int size = TypeFunc::Parms + 2;
1374 CallLeafNode *call = new (C) CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(),
1375 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_object_alloc_base),
1376 "dtrace_object_alloc",
1377 TypeRawPtr::BOTTOM);
1379 // Get base of thread-local storage area
1380 Node* thread = new (C) ThreadLocalNode();
1381 transform_later(thread);
1383 call->init_req(TypeFunc::Parms+0, thread);
1384 call->init_req(TypeFunc::Parms+1, fast_oop);
1385 call->init_req(TypeFunc::Control, fast_oop_ctrl);
1386 call->init_req(TypeFunc::I_O , top()); // does no i/o
1387 call->init_req(TypeFunc::Memory , fast_oop_rawmem);
1388 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1389 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1390 transform_later(call);
1391 fast_oop_ctrl = new (C) ProjNode(call,TypeFunc::Control);
1392 transform_later(fast_oop_ctrl);
1393 fast_oop_rawmem = new (C) ProjNode(call,TypeFunc::Memory);
1394 transform_later(fast_oop_rawmem);
1395 }
1397 // Plug in the successful fast-path into the result merge point
1398 result_region ->init_req(fast_result_path, fast_oop_ctrl);
1399 result_phi_rawoop->init_req(fast_result_path, fast_oop);
1400 result_phi_i_o ->init_req(fast_result_path, i_o);
1401 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem);
1402 } else {
1403 slow_region = ctrl;
1404 result_phi_i_o = i_o; // Rename it to use in the following code.
1405 }
1407 // Generate slow-path call
1408 CallNode *call = new (C) CallStaticJavaNode(slow_call_type, slow_call_address,
1409 OptoRuntime::stub_name(slow_call_address),
1410 alloc->jvms()->bci(),
1411 TypePtr::BOTTOM);
1412 call->init_req( TypeFunc::Control, slow_region );
1413 call->init_req( TypeFunc::I_O , top() ) ; // does no i/o
1414 call->init_req( TypeFunc::Memory , slow_mem ); // may gc ptrs
1415 call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) );
1416 call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) );
1418 call->init_req(TypeFunc::Parms+0, klass_node);
1419 if (length != NULL) {
1420 call->init_req(TypeFunc::Parms+1, length);
1421 }
1423 // Copy debug information and adjust JVMState information, then replace
1424 // allocate node with the call
1425 copy_call_debug_info((CallNode *) alloc, call);
1426 if (!always_slow) {
1427 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1428 } else {
1429 // Hook i_o projection to avoid its elimination during allocation
1430 // replacement (when only a slow call is generated).
1431 call->set_req(TypeFunc::I_O, result_phi_i_o);
1432 }
1433 _igvn.replace_node(alloc, call);
1434 transform_later(call);
1436 // Identify the output projections from the allocate node and
1437 // adjust any references to them.
1438 // The control and io projections look like:
1439 //
1440 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl)
1441 // Allocate Catch
1442 // ^---Proj(io) <-------+ ^---CatchProj(io)
1443 //
1444 // We are interested in the CatchProj nodes.
1445 //
1446 extract_call_projections(call);
1448 // An allocate node has separate memory projections for the uses on
1449 // the control and i_o paths. Replace the control memory projection with
1450 // result_phi_rawmem (unless we are only generating a slow call when
1451 // both memory projections are combined)
1452 if (!always_slow && _memproj_fallthrough != NULL) {
1453 for (DUIterator_Fast imax, i = _memproj_fallthrough->fast_outs(imax); i < imax; i++) {
1454 Node *use = _memproj_fallthrough->fast_out(i);
1455 _igvn.rehash_node_delayed(use);
1456 imax -= replace_input(use, _memproj_fallthrough, result_phi_rawmem);
1457 // back up iterator
1458 --i;
1459 }
1460 }
1461 // Now change uses of _memproj_catchall to use _memproj_fallthrough and delete
1462 // _memproj_catchall so we end up with a call that has only 1 memory projection.
1463 if (_memproj_catchall != NULL ) {
1464 if (_memproj_fallthrough == NULL) {
1465 _memproj_fallthrough = new (C) ProjNode(call, TypeFunc::Memory);
1466 transform_later(_memproj_fallthrough);
1467 }
1468 for (DUIterator_Fast imax, i = _memproj_catchall->fast_outs(imax); i < imax; i++) {
1469 Node *use = _memproj_catchall->fast_out(i);
1470 _igvn.rehash_node_delayed(use);
1471 imax -= replace_input(use, _memproj_catchall, _memproj_fallthrough);
1472 // back up iterator
1473 --i;
1474 }
1475 assert(_memproj_catchall->outcnt() == 0, "all uses must be deleted");
1476 _igvn.remove_dead_node(_memproj_catchall);
1477 }
1479 // An allocate node has separate i_o projections for the uses on the control
1480 // and i_o paths. Always replace the control i_o projection with result i_o
1481 // otherwise incoming i_o become dead when only a slow call is generated
1482 // (it is different from memory projections where both projections are
1483 // combined in such case).
1484 if (_ioproj_fallthrough != NULL) {
1485 for (DUIterator_Fast imax, i = _ioproj_fallthrough->fast_outs(imax); i < imax; i++) {
1486 Node *use = _ioproj_fallthrough->fast_out(i);
1487 _igvn.rehash_node_delayed(use);
1488 imax -= replace_input(use, _ioproj_fallthrough, result_phi_i_o);
1489 // back up iterator
1490 --i;
1491 }
1492 }
1493 // Now change uses of _ioproj_catchall to use _ioproj_fallthrough and delete
1494 // _ioproj_catchall so we end up with a call that has only 1 i_o projection.
1495 if (_ioproj_catchall != NULL ) {
1496 if (_ioproj_fallthrough == NULL) {
1497 _ioproj_fallthrough = new (C) ProjNode(call, TypeFunc::I_O);
1498 transform_later(_ioproj_fallthrough);
1499 }
1500 for (DUIterator_Fast imax, i = _ioproj_catchall->fast_outs(imax); i < imax; i++) {
1501 Node *use = _ioproj_catchall->fast_out(i);
1502 _igvn.rehash_node_delayed(use);
1503 imax -= replace_input(use, _ioproj_catchall, _ioproj_fallthrough);
1504 // back up iterator
1505 --i;
1506 }
1507 assert(_ioproj_catchall->outcnt() == 0, "all uses must be deleted");
1508 _igvn.remove_dead_node(_ioproj_catchall);
1509 }
1511 // if we generated only a slow call, we are done
1512 if (always_slow) {
1513 // Now we can unhook i_o.
1514 if (result_phi_i_o->outcnt() > 1) {
1515 call->set_req(TypeFunc::I_O, top());
1516 } else {
1517 assert(result_phi_i_o->unique_ctrl_out() == call, "");
1518 // Case of new array with negative size known during compilation.
1519 // AllocateArrayNode::Ideal() optimization disconnect unreachable
1520 // following code since call to runtime will throw exception.
1521 // As result there will be no users of i_o after the call.
1522 // Leave i_o attached to this call to avoid problems in preceding graph.
1523 }
1524 return;
1525 }
1528 if (_fallthroughcatchproj != NULL) {
1529 ctrl = _fallthroughcatchproj->clone();
1530 transform_later(ctrl);
1531 _igvn.replace_node(_fallthroughcatchproj, result_region);
1532 } else {
1533 ctrl = top();
1534 }
1535 Node *slow_result;
1536 if (_resproj == NULL) {
1537 // no uses of the allocation result
1538 slow_result = top();
1539 } else {
1540 slow_result = _resproj->clone();
1541 transform_later(slow_result);
1542 _igvn.replace_node(_resproj, result_phi_rawoop);
1543 }
1545 // Plug slow-path into result merge point
1546 result_region ->init_req( slow_result_path, ctrl );
1547 result_phi_rawoop->init_req( slow_result_path, slow_result);
1548 result_phi_rawmem->init_req( slow_result_path, _memproj_fallthrough );
1549 transform_later(result_region);
1550 transform_later(result_phi_rawoop);
1551 transform_later(result_phi_rawmem);
1552 transform_later(result_phi_i_o);
1553 // This completes all paths into the result merge point
1554 }
1557 // Helper for PhaseMacroExpand::expand_allocate_common.
1558 // Initializes the newly-allocated storage.
1559 Node*
1560 PhaseMacroExpand::initialize_object(AllocateNode* alloc,
1561 Node* control, Node* rawmem, Node* object,
1562 Node* klass_node, Node* length,
1563 Node* size_in_bytes) {
1564 InitializeNode* init = alloc->initialization();
1565 // Store the klass & mark bits
1566 Node* mark_node = NULL;
1567 // For now only enable fast locking for non-array types
1568 if (UseBiasedLocking && (length == NULL)) {
1569 mark_node = make_load(control, rawmem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
1570 } else {
1571 mark_node = makecon(TypeRawPtr::make((address)markOopDesc::prototype()));
1572 }
1573 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS);
1575 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
1576 int header_size = alloc->minimum_header_size(); // conservatively small
1578 // Array length
1579 if (length != NULL) { // Arrays need length field
1580 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
1581 // conservatively small header size:
1582 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1583 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
1584 if (k->is_array_klass()) // we know the exact header size in most cases:
1585 header_size = Klass::layout_helper_header_size(k->layout_helper());
1586 }
1588 // Clear the object body, if necessary.
1589 if (init == NULL) {
1590 // The init has somehow disappeared; be cautious and clear everything.
1591 //
1592 // This can happen if a node is allocated but an uncommon trap occurs
1593 // immediately. In this case, the Initialize gets associated with the
1594 // trap, and may be placed in a different (outer) loop, if the Allocate
1595 // is in a loop. If (this is rare) the inner loop gets unrolled, then
1596 // there can be two Allocates to one Initialize. The answer in all these
1597 // edge cases is safety first. It is always safe to clear immediately
1598 // within an Allocate, and then (maybe or maybe not) clear some more later.
1599 if (!ZeroTLAB)
1600 rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
1601 header_size, size_in_bytes,
1602 &_igvn);
1603 } else {
1604 if (!init->is_complete()) {
1605 // Try to win by zeroing only what the init does not store.
1606 // We can also try to do some peephole optimizations,
1607 // such as combining some adjacent subword stores.
1608 rawmem = init->complete_stores(control, rawmem, object,
1609 header_size, size_in_bytes, &_igvn);
1610 }
1611 // We have no more use for this link, since the AllocateNode goes away:
1612 init->set_req(InitializeNode::RawAddress, top());
1613 // (If we keep the link, it just confuses the register allocator,
1614 // who thinks he sees a real use of the address by the membar.)
1615 }
1617 return rawmem;
1618 }
1620 // Generate prefetch instructions for next allocations.
1621 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false,
1622 Node*& contended_phi_rawmem,
1623 Node* old_eden_top, Node* new_eden_top,
1624 Node* length) {
1625 enum { fall_in_path = 1, pf_path = 2 };
1626 if( UseTLAB && AllocatePrefetchStyle == 2 ) {
1627 // Generate prefetch allocation with watermark check.
1628 // As an allocation hits the watermark, we will prefetch starting
1629 // at a "distance" away from watermark.
1631 Node *pf_region = new (C) RegionNode(3);
1632 Node *pf_phi_rawmem = new (C) PhiNode( pf_region, Type::MEMORY,
1633 TypeRawPtr::BOTTOM );
1634 // I/O is used for Prefetch
1635 Node *pf_phi_abio = new (C) PhiNode( pf_region, Type::ABIO );
1637 Node *thread = new (C) ThreadLocalNode();
1638 transform_later(thread);
1640 Node *eden_pf_adr = new (C) AddPNode( top()/*not oop*/, thread,
1641 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) );
1642 transform_later(eden_pf_adr);
1644 Node *old_pf_wm = new (C) LoadPNode( needgc_false,
1645 contended_phi_rawmem, eden_pf_adr,
1646 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM );
1647 transform_later(old_pf_wm);
1649 // check against new_eden_top
1650 Node *need_pf_cmp = new (C) CmpPNode( new_eden_top, old_pf_wm );
1651 transform_later(need_pf_cmp);
1652 Node *need_pf_bol = new (C) BoolNode( need_pf_cmp, BoolTest::ge );
1653 transform_later(need_pf_bol);
1654 IfNode *need_pf_iff = new (C) IfNode( needgc_false, need_pf_bol,
1655 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN );
1656 transform_later(need_pf_iff);
1658 // true node, add prefetchdistance
1659 Node *need_pf_true = new (C) IfTrueNode( need_pf_iff );
1660 transform_later(need_pf_true);
1662 Node *need_pf_false = new (C) IfFalseNode( need_pf_iff );
1663 transform_later(need_pf_false);
1665 Node *new_pf_wmt = new (C) AddPNode( top(), old_pf_wm,
1666 _igvn.MakeConX(AllocatePrefetchDistance) );
1667 transform_later(new_pf_wmt );
1668 new_pf_wmt->set_req(0, need_pf_true);
1670 Node *store_new_wmt = new (C) StorePNode( need_pf_true,
1671 contended_phi_rawmem, eden_pf_adr,
1672 TypeRawPtr::BOTTOM, new_pf_wmt );
1673 transform_later(store_new_wmt);
1675 // adding prefetches
1676 pf_phi_abio->init_req( fall_in_path, i_o );
1678 Node *prefetch_adr;
1679 Node *prefetch;
1680 uint lines = AllocatePrefetchDistance / AllocatePrefetchStepSize;
1681 uint step_size = AllocatePrefetchStepSize;
1682 uint distance = 0;
1684 for ( uint i = 0; i < lines; i++ ) {
1685 prefetch_adr = new (C) AddPNode( old_pf_wm, new_pf_wmt,
1686 _igvn.MakeConX(distance) );
1687 transform_later(prefetch_adr);
1688 prefetch = new (C) PrefetchAllocationNode( i_o, prefetch_adr );
1689 transform_later(prefetch);
1690 distance += step_size;
1691 i_o = prefetch;
1692 }
1693 pf_phi_abio->set_req( pf_path, i_o );
1695 pf_region->init_req( fall_in_path, need_pf_false );
1696 pf_region->init_req( pf_path, need_pf_true );
1698 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem );
1699 pf_phi_rawmem->init_req( pf_path, store_new_wmt );
1701 transform_later(pf_region);
1702 transform_later(pf_phi_rawmem);
1703 transform_later(pf_phi_abio);
1705 needgc_false = pf_region;
1706 contended_phi_rawmem = pf_phi_rawmem;
1707 i_o = pf_phi_abio;
1708 } else if( UseTLAB && AllocatePrefetchStyle == 3 ) {
1709 // Insert a prefetch for each allocation.
1710 // This code is used for Sparc with BIS.
1711 Node *pf_region = new (C) RegionNode(3);
1712 Node *pf_phi_rawmem = new (C) PhiNode( pf_region, Type::MEMORY,
1713 TypeRawPtr::BOTTOM );
1715 // Generate several prefetch instructions.
1716 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1717 uint step_size = AllocatePrefetchStepSize;
1718 uint distance = AllocatePrefetchDistance;
1720 // Next cache address.
1721 Node *cache_adr = new (C) AddPNode(old_eden_top, old_eden_top,
1722 _igvn.MakeConX(distance));
1723 transform_later(cache_adr);
1724 cache_adr = new (C) CastP2XNode(needgc_false, cache_adr);
1725 transform_later(cache_adr);
1726 Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1));
1727 cache_adr = new (C) AndXNode(cache_adr, mask);
1728 transform_later(cache_adr);
1729 cache_adr = new (C) CastX2PNode(cache_adr);
1730 transform_later(cache_adr);
1732 // Prefetch
1733 Node *prefetch = new (C) PrefetchAllocationNode( contended_phi_rawmem, cache_adr );
1734 prefetch->set_req(0, needgc_false);
1735 transform_later(prefetch);
1736 contended_phi_rawmem = prefetch;
1737 Node *prefetch_adr;
1738 distance = step_size;
1739 for ( uint i = 1; i < lines; i++ ) {
1740 prefetch_adr = new (C) AddPNode( cache_adr, cache_adr,
1741 _igvn.MakeConX(distance) );
1742 transform_later(prefetch_adr);
1743 prefetch = new (C) PrefetchAllocationNode( contended_phi_rawmem, prefetch_adr );
1744 transform_later(prefetch);
1745 distance += step_size;
1746 contended_phi_rawmem = prefetch;
1747 }
1748 } else if( AllocatePrefetchStyle > 0 ) {
1749 // Insert a prefetch for each allocation only on the fast-path
1750 Node *prefetch_adr;
1751 Node *prefetch;
1752 // Generate several prefetch instructions.
1753 uint lines = (length != NULL) ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1754 uint step_size = AllocatePrefetchStepSize;
1755 uint distance = AllocatePrefetchDistance;
1756 for ( uint i = 0; i < lines; i++ ) {
1757 prefetch_adr = new (C) AddPNode( old_eden_top, new_eden_top,
1758 _igvn.MakeConX(distance) );
1759 transform_later(prefetch_adr);
1760 prefetch = new (C) PrefetchAllocationNode( i_o, prefetch_adr );
1761 // Do not let it float too high, since if eden_top == eden_end,
1762 // both might be null.
1763 if( i == 0 ) { // Set control for first prefetch, next follows it
1764 prefetch->init_req(0, needgc_false);
1765 }
1766 transform_later(prefetch);
1767 distance += step_size;
1768 i_o = prefetch;
1769 }
1770 }
1771 return i_o;
1772 }
1775 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) {
1776 expand_allocate_common(alloc, NULL,
1777 OptoRuntime::new_instance_Type(),
1778 OptoRuntime::new_instance_Java());
1779 }
1781 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) {
1782 Node* length = alloc->in(AllocateNode::ALength);
1783 InitializeNode* init = alloc->initialization();
1784 Node* klass_node = alloc->in(AllocateNode::KlassNode);
1785 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
1786 address slow_call_address; // Address of slow call
1787 if (init != NULL && init->is_complete_with_arraycopy() &&
1788 k->is_type_array_klass()) {
1789 // Don't zero type array during slow allocation in VM since
1790 // it will be initialized later by arraycopy in compiled code.
1791 slow_call_address = OptoRuntime::new_array_nozero_Java();
1792 } else {
1793 slow_call_address = OptoRuntime::new_array_Java();
1794 }
1795 expand_allocate_common(alloc, length,
1796 OptoRuntime::new_array_Type(),
1797 slow_call_address);
1798 }
1800 //-------------------mark_eliminated_box----------------------------------
1801 //
1802 // During EA obj may point to several objects but after few ideal graph
1803 // transformations (CCP) it may point to only one non escaping object
1804 // (but still using phi), corresponding locks and unlocks will be marked
1805 // for elimination. Later obj could be replaced with a new node (new phi)
1806 // and which does not have escape information. And later after some graph
1807 // reshape other locks and unlocks (which were not marked for elimination
1808 // before) are connected to this new obj (phi) but they still will not be
1809 // marked for elimination since new obj has no escape information.
1810 // Mark all associated (same box and obj) lock and unlock nodes for
1811 // elimination if some of them marked already.
1812 void PhaseMacroExpand::mark_eliminated_box(Node* oldbox, Node* obj) {
1813 if (oldbox->as_BoxLock()->is_eliminated())
1814 return; // This BoxLock node was processed already.
1816 // New implementation (EliminateNestedLocks) has separate BoxLock
1817 // node for each locked region so mark all associated locks/unlocks as
1818 // eliminated even if different objects are referenced in one locked region
1819 // (for example, OSR compilation of nested loop inside locked scope).
1820 if (EliminateNestedLocks ||
1821 oldbox->as_BoxLock()->is_simple_lock_region(NULL, obj)) {
1822 // Box is used only in one lock region. Mark this box as eliminated.
1823 _igvn.hash_delete(oldbox);
1824 oldbox->as_BoxLock()->set_eliminated(); // This changes box's hash value
1825 _igvn.hash_insert(oldbox);
1827 for (uint i = 0; i < oldbox->outcnt(); i++) {
1828 Node* u = oldbox->raw_out(i);
1829 if (u->is_AbstractLock() && !u->as_AbstractLock()->is_non_esc_obj()) {
1830 AbstractLockNode* alock = u->as_AbstractLock();
1831 // Check lock's box since box could be referenced by Lock's debug info.
1832 if (alock->box_node() == oldbox) {
1833 // Mark eliminated all related locks and unlocks.
1834 alock->set_non_esc_obj();
1835 }
1836 }
1837 }
1838 return;
1839 }
1841 // Create new "eliminated" BoxLock node and use it in monitor debug info
1842 // instead of oldbox for the same object.
1843 BoxLockNode* newbox = oldbox->clone()->as_BoxLock();
1845 // Note: BoxLock node is marked eliminated only here and it is used
1846 // to indicate that all associated lock and unlock nodes are marked
1847 // for elimination.
1848 newbox->set_eliminated();
1849 transform_later(newbox);
1851 // Replace old box node with new box for all users of the same object.
1852 for (uint i = 0; i < oldbox->outcnt();) {
1853 bool next_edge = true;
1855 Node* u = oldbox->raw_out(i);
1856 if (u->is_AbstractLock()) {
1857 AbstractLockNode* alock = u->as_AbstractLock();
1858 if (alock->box_node() == oldbox && alock->obj_node()->eqv_uncast(obj)) {
1859 // Replace Box and mark eliminated all related locks and unlocks.
1860 alock->set_non_esc_obj();
1861 _igvn.rehash_node_delayed(alock);
1862 alock->set_box_node(newbox);
1863 next_edge = false;
1864 }
1865 }
1866 if (u->is_FastLock() && u->as_FastLock()->obj_node()->eqv_uncast(obj)) {
1867 FastLockNode* flock = u->as_FastLock();
1868 assert(flock->box_node() == oldbox, "sanity");
1869 _igvn.rehash_node_delayed(flock);
1870 flock->set_box_node(newbox);
1871 next_edge = false;
1872 }
1874 // Replace old box in monitor debug info.
1875 if (u->is_SafePoint() && u->as_SafePoint()->jvms()) {
1876 SafePointNode* sfn = u->as_SafePoint();
1877 JVMState* youngest_jvms = sfn->jvms();
1878 int max_depth = youngest_jvms->depth();
1879 for (int depth = 1; depth <= max_depth; depth++) {
1880 JVMState* jvms = youngest_jvms->of_depth(depth);
1881 int num_mon = jvms->nof_monitors();
1882 // Loop over monitors
1883 for (int idx = 0; idx < num_mon; idx++) {
1884 Node* obj_node = sfn->monitor_obj(jvms, idx);
1885 Node* box_node = sfn->monitor_box(jvms, idx);
1886 if (box_node == oldbox && obj_node->eqv_uncast(obj)) {
1887 int j = jvms->monitor_box_offset(idx);
1888 _igvn.replace_input_of(u, j, newbox);
1889 next_edge = false;
1890 }
1891 }
1892 }
1893 }
1894 if (next_edge) i++;
1895 }
1896 }
1898 //-----------------------mark_eliminated_locking_nodes-----------------------
1899 void PhaseMacroExpand::mark_eliminated_locking_nodes(AbstractLockNode *alock) {
1900 if (EliminateNestedLocks) {
1901 if (alock->is_nested()) {
1902 assert(alock->box_node()->as_BoxLock()->is_eliminated(), "sanity");
1903 return;
1904 } else if (!alock->is_non_esc_obj()) { // Not eliminated or coarsened
1905 // Only Lock node has JVMState needed here.
1906 if (alock->jvms() != NULL && alock->as_Lock()->is_nested_lock_region()) {
1907 // Mark eliminated related nested locks and unlocks.
1908 Node* obj = alock->obj_node();
1909 BoxLockNode* box_node = alock->box_node()->as_BoxLock();
1910 assert(!box_node->is_eliminated(), "should not be marked yet");
1911 // Note: BoxLock node is marked eliminated only here
1912 // and it is used to indicate that all associated lock
1913 // and unlock nodes are marked for elimination.
1914 box_node->set_eliminated(); // Box's hash is always NO_HASH here
1915 for (uint i = 0; i < box_node->outcnt(); i++) {
1916 Node* u = box_node->raw_out(i);
1917 if (u->is_AbstractLock()) {
1918 alock = u->as_AbstractLock();
1919 if (alock->box_node() == box_node) {
1920 // Verify that this Box is referenced only by related locks.
1921 assert(alock->obj_node()->eqv_uncast(obj), "");
1922 // Mark all related locks and unlocks.
1923 alock->set_nested();
1924 }
1925 }
1926 }
1927 }
1928 return;
1929 }
1930 // Process locks for non escaping object
1931 assert(alock->is_non_esc_obj(), "");
1932 } // EliminateNestedLocks
1934 if (alock->is_non_esc_obj()) { // Lock is used for non escaping object
1935 // Look for all locks of this object and mark them and
1936 // corresponding BoxLock nodes as eliminated.
1937 Node* obj = alock->obj_node();
1938 for (uint j = 0; j < obj->outcnt(); j++) {
1939 Node* o = obj->raw_out(j);
1940 if (o->is_AbstractLock() &&
1941 o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) {
1942 alock = o->as_AbstractLock();
1943 Node* box = alock->box_node();
1944 // Replace old box node with new eliminated box for all users
1945 // of the same object and mark related locks as eliminated.
1946 mark_eliminated_box(box, obj);
1947 }
1948 }
1949 }
1950 }
1952 // we have determined that this lock/unlock can be eliminated, we simply
1953 // eliminate the node without expanding it.
1954 //
1955 // Note: The membar's associated with the lock/unlock are currently not
1956 // eliminated. This should be investigated as a future enhancement.
1957 //
1958 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) {
1960 if (!alock->is_eliminated()) {
1961 return false;
1962 }
1963 #ifdef ASSERT
1964 if (!alock->is_coarsened()) {
1965 // Check that new "eliminated" BoxLock node is created.
1966 BoxLockNode* oldbox = alock->box_node()->as_BoxLock();
1967 assert(oldbox->is_eliminated(), "should be done already");
1968 }
1969 #endif
1970 CompileLog* log = C->log();
1971 if (log != NULL) {
1972 log->head("eliminate_lock lock='%d'",
1973 alock->is_Lock());
1974 JVMState* p = alock->jvms();
1975 while (p != NULL) {
1976 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1977 p = p->caller();
1978 }
1979 log->tail("eliminate_lock");
1980 }
1982 #ifndef PRODUCT
1983 if (PrintEliminateLocks) {
1984 if (alock->is_Lock()) {
1985 tty->print_cr("++++ Eliminated: %d Lock", alock->_idx);
1986 } else {
1987 tty->print_cr("++++ Eliminated: %d Unlock", alock->_idx);
1988 }
1989 }
1990 #endif
1992 Node* mem = alock->in(TypeFunc::Memory);
1993 Node* ctrl = alock->in(TypeFunc::Control);
1995 extract_call_projections(alock);
1996 // There are 2 projections from the lock. The lock node will
1997 // be deleted when its last use is subsumed below.
1998 assert(alock->outcnt() == 2 &&
1999 _fallthroughproj != NULL &&
2000 _memproj_fallthrough != NULL,
2001 "Unexpected projections from Lock/Unlock");
2003 Node* fallthroughproj = _fallthroughproj;
2004 Node* memproj_fallthrough = _memproj_fallthrough;
2006 // The memory projection from a lock/unlock is RawMem
2007 // The input to a Lock is merged memory, so extract its RawMem input
2008 // (unless the MergeMem has been optimized away.)
2009 if (alock->is_Lock()) {
2010 // Seach for MemBarAcquireLock node and delete it also.
2011 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
2012 assert(membar != NULL && membar->Opcode() == Op_MemBarAcquireLock, "");
2013 Node* ctrlproj = membar->proj_out(TypeFunc::Control);
2014 Node* memproj = membar->proj_out(TypeFunc::Memory);
2015 _igvn.replace_node(ctrlproj, fallthroughproj);
2016 _igvn.replace_node(memproj, memproj_fallthrough);
2018 // Delete FastLock node also if this Lock node is unique user
2019 // (a loop peeling may clone a Lock node).
2020 Node* flock = alock->as_Lock()->fastlock_node();
2021 if (flock->outcnt() == 1) {
2022 assert(flock->unique_out() == alock, "sanity");
2023 _igvn.replace_node(flock, top());
2024 }
2025 }
2027 // Seach for MemBarReleaseLock node and delete it also.
2028 if (alock->is_Unlock() && ctrl != NULL && ctrl->is_Proj() &&
2029 ctrl->in(0)->is_MemBar()) {
2030 MemBarNode* membar = ctrl->in(0)->as_MemBar();
2031 assert(membar->Opcode() == Op_MemBarReleaseLock &&
2032 mem->is_Proj() && membar == mem->in(0), "");
2033 _igvn.replace_node(fallthroughproj, ctrl);
2034 _igvn.replace_node(memproj_fallthrough, mem);
2035 fallthroughproj = ctrl;
2036 memproj_fallthrough = mem;
2037 ctrl = membar->in(TypeFunc::Control);
2038 mem = membar->in(TypeFunc::Memory);
2039 }
2041 _igvn.replace_node(fallthroughproj, ctrl);
2042 _igvn.replace_node(memproj_fallthrough, mem);
2043 return true;
2044 }
2047 //------------------------------expand_lock_node----------------------
2048 void PhaseMacroExpand::expand_lock_node(LockNode *lock) {
2050 Node* ctrl = lock->in(TypeFunc::Control);
2051 Node* mem = lock->in(TypeFunc::Memory);
2052 Node* obj = lock->obj_node();
2053 Node* box = lock->box_node();
2054 Node* flock = lock->fastlock_node();
2056 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2058 // Make the merge point
2059 Node *region;
2060 Node *mem_phi;
2061 Node *slow_path;
2063 if (UseOptoBiasInlining) {
2064 /*
2065 * See the full description in MacroAssembler::biased_locking_enter().
2066 *
2067 * if( (mark_word & biased_lock_mask) == biased_lock_pattern ) {
2068 * // The object is biased.
2069 * proto_node = klass->prototype_header;
2070 * o_node = thread | proto_node;
2071 * x_node = o_node ^ mark_word;
2072 * if( (x_node & ~age_mask) == 0 ) { // Biased to the current thread ?
2073 * // Done.
2074 * } else {
2075 * if( (x_node & biased_lock_mask) != 0 ) {
2076 * // The klass's prototype header is no longer biased.
2077 * cas(&mark_word, mark_word, proto_node)
2078 * goto cas_lock;
2079 * } else {
2080 * // The klass's prototype header is still biased.
2081 * if( (x_node & epoch_mask) != 0 ) { // Expired epoch?
2082 * old = mark_word;
2083 * new = o_node;
2084 * } else {
2085 * // Different thread or anonymous biased.
2086 * old = mark_word & (epoch_mask | age_mask | biased_lock_mask);
2087 * new = thread | old;
2088 * }
2089 * // Try to rebias.
2090 * if( cas(&mark_word, old, new) == 0 ) {
2091 * // Done.
2092 * } else {
2093 * goto slow_path; // Failed.
2094 * }
2095 * }
2096 * }
2097 * } else {
2098 * // The object is not biased.
2099 * cas_lock:
2100 * if( FastLock(obj) == 0 ) {
2101 * // Done.
2102 * } else {
2103 * slow_path:
2104 * OptoRuntime::complete_monitor_locking_Java(obj);
2105 * }
2106 * }
2107 */
2109 region = new (C) RegionNode(5);
2110 // create a Phi for the memory state
2111 mem_phi = new (C) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2113 Node* fast_lock_region = new (C) RegionNode(3);
2114 Node* fast_lock_mem_phi = new (C) PhiNode( fast_lock_region, Type::MEMORY, TypeRawPtr::BOTTOM);
2116 // First, check mark word for the biased lock pattern.
2117 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2119 // Get fast path - mark word has the biased lock pattern.
2120 ctrl = opt_bits_test(ctrl, fast_lock_region, 1, mark_node,
2121 markOopDesc::biased_lock_mask_in_place,
2122 markOopDesc::biased_lock_pattern, true);
2123 // fast_lock_region->in(1) is set to slow path.
2124 fast_lock_mem_phi->init_req(1, mem);
2126 // Now check that the lock is biased to the current thread and has
2127 // the same epoch and bias as Klass::_prototype_header.
2129 // Special-case a fresh allocation to avoid building nodes:
2130 Node* klass_node = AllocateNode::Ideal_klass(obj, &_igvn);
2131 if (klass_node == NULL) {
2132 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
2133 klass_node = transform_later( LoadKlassNode::make(_igvn, mem, k_adr, _igvn.type(k_adr)->is_ptr()) );
2134 #ifdef _LP64
2135 if (UseCompressedKlassPointers && klass_node->is_DecodeNKlass()) {
2136 assert(klass_node->in(1)->Opcode() == Op_LoadNKlass, "sanity");
2137 klass_node->in(1)->init_req(0, ctrl);
2138 } else
2139 #endif
2140 klass_node->init_req(0, ctrl);
2141 }
2142 Node *proto_node = make_load(ctrl, mem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeX_X, TypeX_X->basic_type());
2144 Node* thread = transform_later(new (C) ThreadLocalNode());
2145 Node* cast_thread = transform_later(new (C) CastP2XNode(ctrl, thread));
2146 Node* o_node = transform_later(new (C) OrXNode(cast_thread, proto_node));
2147 Node* x_node = transform_later(new (C) XorXNode(o_node, mark_node));
2149 // Get slow path - mark word does NOT match the value.
2150 Node* not_biased_ctrl = opt_bits_test(ctrl, region, 3, x_node,
2151 (~markOopDesc::age_mask_in_place), 0);
2152 // region->in(3) is set to fast path - the object is biased to the current thread.
2153 mem_phi->init_req(3, mem);
2156 // Mark word does NOT match the value (thread | Klass::_prototype_header).
2159 // First, check biased pattern.
2160 // Get fast path - _prototype_header has the same biased lock pattern.
2161 ctrl = opt_bits_test(not_biased_ctrl, fast_lock_region, 2, x_node,
2162 markOopDesc::biased_lock_mask_in_place, 0, true);
2164 not_biased_ctrl = fast_lock_region->in(2); // Slow path
2165 // fast_lock_region->in(2) - the prototype header is no longer biased
2166 // and we have to revoke the bias on this object.
2167 // We are going to try to reset the mark of this object to the prototype
2168 // value and fall through to the CAS-based locking scheme.
2169 Node* adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
2170 Node* cas = new (C) StoreXConditionalNode(not_biased_ctrl, mem, adr,
2171 proto_node, mark_node);
2172 transform_later(cas);
2173 Node* proj = transform_later( new (C) SCMemProjNode(cas));
2174 fast_lock_mem_phi->init_req(2, proj);
2177 // Second, check epoch bits.
2178 Node* rebiased_region = new (C) RegionNode(3);
2179 Node* old_phi = new (C) PhiNode( rebiased_region, TypeX_X);
2180 Node* new_phi = new (C) PhiNode( rebiased_region, TypeX_X);
2182 // Get slow path - mark word does NOT match epoch bits.
2183 Node* epoch_ctrl = opt_bits_test(ctrl, rebiased_region, 1, x_node,
2184 markOopDesc::epoch_mask_in_place, 0);
2185 // The epoch of the current bias is not valid, attempt to rebias the object
2186 // toward the current thread.
2187 rebiased_region->init_req(2, epoch_ctrl);
2188 old_phi->init_req(2, mark_node);
2189 new_phi->init_req(2, o_node);
2191 // rebiased_region->in(1) is set to fast path.
2192 // The epoch of the current bias is still valid but we know
2193 // nothing about the owner; it might be set or it might be clear.
2194 Node* cmask = MakeConX(markOopDesc::biased_lock_mask_in_place |
2195 markOopDesc::age_mask_in_place |
2196 markOopDesc::epoch_mask_in_place);
2197 Node* old = transform_later(new (C) AndXNode(mark_node, cmask));
2198 cast_thread = transform_later(new (C) CastP2XNode(ctrl, thread));
2199 Node* new_mark = transform_later(new (C) OrXNode(cast_thread, old));
2200 old_phi->init_req(1, old);
2201 new_phi->init_req(1, new_mark);
2203 transform_later(rebiased_region);
2204 transform_later(old_phi);
2205 transform_later(new_phi);
2207 // Try to acquire the bias of the object using an atomic operation.
2208 // If this fails we will go in to the runtime to revoke the object's bias.
2209 cas = new (C) StoreXConditionalNode(rebiased_region, mem, adr,
2210 new_phi, old_phi);
2211 transform_later(cas);
2212 proj = transform_later( new (C) SCMemProjNode(cas));
2214 // Get slow path - Failed to CAS.
2215 not_biased_ctrl = opt_bits_test(rebiased_region, region, 4, cas, 0, 0);
2216 mem_phi->init_req(4, proj);
2217 // region->in(4) is set to fast path - the object is rebiased to the current thread.
2219 // Failed to CAS.
2220 slow_path = new (C) RegionNode(3);
2221 Node *slow_mem = new (C) PhiNode( slow_path, Type::MEMORY, TypeRawPtr::BOTTOM);
2223 slow_path->init_req(1, not_biased_ctrl); // Capture slow-control
2224 slow_mem->init_req(1, proj);
2226 // Call CAS-based locking scheme (FastLock node).
2228 transform_later(fast_lock_region);
2229 transform_later(fast_lock_mem_phi);
2231 // Get slow path - FastLock failed to lock the object.
2232 ctrl = opt_bits_test(fast_lock_region, region, 2, flock, 0, 0);
2233 mem_phi->init_req(2, fast_lock_mem_phi);
2234 // region->in(2) is set to fast path - the object is locked to the current thread.
2236 slow_path->init_req(2, ctrl); // Capture slow-control
2237 slow_mem->init_req(2, fast_lock_mem_phi);
2239 transform_later(slow_path);
2240 transform_later(slow_mem);
2241 // Reset lock's memory edge.
2242 lock->set_req(TypeFunc::Memory, slow_mem);
2244 } else {
2245 region = new (C) RegionNode(3);
2246 // create a Phi for the memory state
2247 mem_phi = new (C) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2249 // Optimize test; set region slot 2
2250 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0);
2251 mem_phi->init_req(2, mem);
2252 }
2254 // Make slow path call
2255 CallNode *call = make_slow_call( (CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(), OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path, obj, box );
2257 extract_call_projections(call);
2259 // Slow path can only throw asynchronous exceptions, which are always
2260 // de-opted. So the compiler thinks the slow-call can never throw an
2261 // exception. If it DOES throw an exception we would need the debug
2262 // info removed first (since if it throws there is no monitor).
2263 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
2264 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
2266 // Capture slow path
2267 // disconnect fall-through projection from call and create a new one
2268 // hook up users of fall-through projection to region
2269 Node *slow_ctrl = _fallthroughproj->clone();
2270 transform_later(slow_ctrl);
2271 _igvn.hash_delete(_fallthroughproj);
2272 _fallthroughproj->disconnect_inputs(NULL, C);
2273 region->init_req(1, slow_ctrl);
2274 // region inputs are now complete
2275 transform_later(region);
2276 _igvn.replace_node(_fallthroughproj, region);
2278 Node *memproj = transform_later( new(C) ProjNode(call, TypeFunc::Memory) );
2279 mem_phi->init_req(1, memproj );
2280 transform_later(mem_phi);
2281 _igvn.replace_node(_memproj_fallthrough, mem_phi);
2282 }
2284 //------------------------------expand_unlock_node----------------------
2285 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
2287 Node* ctrl = unlock->in(TypeFunc::Control);
2288 Node* mem = unlock->in(TypeFunc::Memory);
2289 Node* obj = unlock->obj_node();
2290 Node* box = unlock->box_node();
2292 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2294 // No need for a null check on unlock
2296 // Make the merge point
2297 Node *region;
2298 Node *mem_phi;
2300 if (UseOptoBiasInlining) {
2301 // Check for biased locking unlock case, which is a no-op.
2302 // See the full description in MacroAssembler::biased_locking_exit().
2303 region = new (C) RegionNode(4);
2304 // create a Phi for the memory state
2305 mem_phi = new (C) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2306 mem_phi->init_req(3, mem);
2308 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2309 ctrl = opt_bits_test(ctrl, region, 3, mark_node,
2310 markOopDesc::biased_lock_mask_in_place,
2311 markOopDesc::biased_lock_pattern);
2312 } else {
2313 region = new (C) RegionNode(3);
2314 // create a Phi for the memory state
2315 mem_phi = new (C) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2316 }
2318 FastUnlockNode *funlock = new (C) FastUnlockNode( ctrl, obj, box );
2319 funlock = transform_later( funlock )->as_FastUnlock();
2320 // Optimize test; set region slot 2
2321 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0);
2323 CallNode *call = make_slow_call( (CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C), "complete_monitor_unlocking_C", slow_path, obj, box );
2325 extract_call_projections(call);
2327 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
2328 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
2330 // No exceptions for unlocking
2331 // Capture slow path
2332 // disconnect fall-through projection from call and create a new one
2333 // hook up users of fall-through projection to region
2334 Node *slow_ctrl = _fallthroughproj->clone();
2335 transform_later(slow_ctrl);
2336 _igvn.hash_delete(_fallthroughproj);
2337 _fallthroughproj->disconnect_inputs(NULL, C);
2338 region->init_req(1, slow_ctrl);
2339 // region inputs are now complete
2340 transform_later(region);
2341 _igvn.replace_node(_fallthroughproj, region);
2343 Node *memproj = transform_later( new(C) ProjNode(call, TypeFunc::Memory) );
2344 mem_phi->init_req(1, memproj );
2345 mem_phi->init_req(2, mem);
2346 transform_later(mem_phi);
2347 _igvn.replace_node(_memproj_fallthrough, mem_phi);
2348 }
2350 //---------------------------eliminate_macro_nodes----------------------
2351 // Eliminate scalar replaced allocations and associated locks.
2352 void PhaseMacroExpand::eliminate_macro_nodes() {
2353 if (C->macro_count() == 0)
2354 return;
2356 // First, attempt to eliminate locks
2357 int cnt = C->macro_count();
2358 for (int i=0; i < cnt; i++) {
2359 Node *n = C->macro_node(i);
2360 if (n->is_AbstractLock()) { // Lock and Unlock nodes
2361 // Before elimination mark all associated (same box and obj)
2362 // lock and unlock nodes.
2363 mark_eliminated_locking_nodes(n->as_AbstractLock());
2364 }
2365 }
2366 bool progress = true;
2367 while (progress) {
2368 progress = false;
2369 for (int i = C->macro_count(); i > 0; i--) {
2370 Node * n = C->macro_node(i-1);
2371 bool success = false;
2372 debug_only(int old_macro_count = C->macro_count(););
2373 if (n->is_AbstractLock()) {
2374 success = eliminate_locking_node(n->as_AbstractLock());
2375 }
2376 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2377 progress = progress || success;
2378 }
2379 }
2380 // Next, attempt to eliminate allocations
2381 progress = true;
2382 while (progress) {
2383 progress = false;
2384 for (int i = C->macro_count(); i > 0; i--) {
2385 Node * n = C->macro_node(i-1);
2386 bool success = false;
2387 debug_only(int old_macro_count = C->macro_count(););
2388 switch (n->class_id()) {
2389 case Node::Class_Allocate:
2390 case Node::Class_AllocateArray:
2391 success = eliminate_allocate_node(n->as_Allocate());
2392 break;
2393 case Node::Class_Lock:
2394 case Node::Class_Unlock:
2395 assert(!n->as_AbstractLock()->is_eliminated(), "sanity");
2396 break;
2397 default:
2398 assert(n->Opcode() == Op_LoopLimit ||
2399 n->Opcode() == Op_Opaque1 ||
2400 n->Opcode() == Op_Opaque2, "unknown node type in macro list");
2401 }
2402 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2403 progress = progress || success;
2404 }
2405 }
2406 }
2408 //------------------------------expand_macro_nodes----------------------
2409 // Returns true if a failure occurred.
2410 bool PhaseMacroExpand::expand_macro_nodes() {
2411 // Last attempt to eliminate macro nodes.
2412 eliminate_macro_nodes();
2414 // Make sure expansion will not cause node limit to be exceeded.
2415 // Worst case is a macro node gets expanded into about 50 nodes.
2416 // Allow 50% more for optimization.
2417 if (C->check_node_count(C->macro_count() * 75, "out of nodes before macro expansion" ) )
2418 return true;
2420 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations.
2421 bool progress = true;
2422 while (progress) {
2423 progress = false;
2424 for (int i = C->macro_count(); i > 0; i--) {
2425 Node * n = C->macro_node(i-1);
2426 bool success = false;
2427 debug_only(int old_macro_count = C->macro_count(););
2428 if (n->Opcode() == Op_LoopLimit) {
2429 // Remove it from macro list and put on IGVN worklist to optimize.
2430 C->remove_macro_node(n);
2431 _igvn._worklist.push(n);
2432 success = true;
2433 } else if (n->Opcode() == Op_Opaque1 || n->Opcode() == Op_Opaque2) {
2434 _igvn.replace_node(n, n->in(1));
2435 success = true;
2436 }
2437 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2438 progress = progress || success;
2439 }
2440 }
2442 // expand "macro" nodes
2443 // nodes are removed from the macro list as they are processed
2444 while (C->macro_count() > 0) {
2445 int macro_count = C->macro_count();
2446 Node * n = C->macro_node(macro_count-1);
2447 assert(n->is_macro(), "only macro nodes expected here");
2448 if (_igvn.type(n) == Type::TOP || n->in(0)->is_top() ) {
2449 // node is unreachable, so don't try to expand it
2450 C->remove_macro_node(n);
2451 continue;
2452 }
2453 switch (n->class_id()) {
2454 case Node::Class_Allocate:
2455 expand_allocate(n->as_Allocate());
2456 break;
2457 case Node::Class_AllocateArray:
2458 expand_allocate_array(n->as_AllocateArray());
2459 break;
2460 case Node::Class_Lock:
2461 expand_lock_node(n->as_Lock());
2462 break;
2463 case Node::Class_Unlock:
2464 expand_unlock_node(n->as_Unlock());
2465 break;
2466 default:
2467 assert(false, "unknown node type in macro list");
2468 }
2469 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
2470 if (C->failing()) return true;
2471 }
2473 _igvn.set_delay_transform(false);
2474 _igvn.optimize();
2475 if (C->failing()) return true;
2476 return false;
2477 }