annotate src/share/vm/opto/macro.cpp @ 423:a1980da045cc

6462850: generate biased locking code in C2 ideal graph Summary: Inline biased locking code in C2 ideal graph during macro nodes expansion Reviewed-by: never
author kvn
date Fri, 07 Nov 2008 09:29:38 -0800
parents f8199438385b
children 87559db65269
rev   line source
duke@0 1 /*
xdono@196 2 * Copyright 2005-2008 Sun Microsystems, Inc. All Rights Reserved.
duke@0 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
duke@0 4 *
duke@0 5 * This code is free software; you can redistribute it and/or modify it
duke@0 6 * under the terms of the GNU General Public License version 2 only, as
duke@0 7 * published by the Free Software Foundation.
duke@0 8 *
duke@0 9 * This code is distributed in the hope that it will be useful, but WITHOUT
duke@0 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
duke@0 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
duke@0 12 * version 2 for more details (a copy is included in the LICENSE file that
duke@0 13 * accompanied this code).
duke@0 14 *
duke@0 15 * You should have received a copy of the GNU General Public License version
duke@0 16 * 2 along with this work; if not, write to the Free Software Foundation,
duke@0 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
duke@0 18 *
duke@0 19 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
duke@0 20 * CA 95054 USA or visit www.sun.com if you need additional information or
duke@0 21 * have any questions.
duke@0 22 *
duke@0 23 */
duke@0 24
duke@0 25 #include "incls/_precompiled.incl"
duke@0 26 #include "incls/_macro.cpp.incl"
duke@0 27
duke@0 28
duke@0 29 //
duke@0 30 // Replace any references to "oldref" in inputs to "use" with "newref".
duke@0 31 // Returns the number of replacements made.
duke@0 32 //
duke@0 33 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) {
duke@0 34 int nreplacements = 0;
duke@0 35 uint req = use->req();
duke@0 36 for (uint j = 0; j < use->len(); j++) {
duke@0 37 Node *uin = use->in(j);
duke@0 38 if (uin == oldref) {
duke@0 39 if (j < req)
duke@0 40 use->set_req(j, newref);
duke@0 41 else
duke@0 42 use->set_prec(j, newref);
duke@0 43 nreplacements++;
duke@0 44 } else if (j >= req && uin == NULL) {
duke@0 45 break;
duke@0 46 }
duke@0 47 }
duke@0 48 return nreplacements;
duke@0 49 }
duke@0 50
duke@0 51 void PhaseMacroExpand::copy_call_debug_info(CallNode *oldcall, CallNode * newcall) {
duke@0 52 // Copy debug information and adjust JVMState information
duke@0 53 uint old_dbg_start = oldcall->tf()->domain()->cnt();
duke@0 54 uint new_dbg_start = newcall->tf()->domain()->cnt();
duke@0 55 int jvms_adj = new_dbg_start - old_dbg_start;
duke@0 56 assert (new_dbg_start == newcall->req(), "argument count mismatch");
kvn@63 57
kvn@63 58 Dict* sosn_map = new Dict(cmpkey,hashkey);
duke@0 59 for (uint i = old_dbg_start; i < oldcall->req(); i++) {
kvn@63 60 Node* old_in = oldcall->in(i);
kvn@63 61 // Clone old SafePointScalarObjectNodes, adjusting their field contents.
kvn@63 62 if (old_in->is_SafePointScalarObject()) {
kvn@63 63 SafePointScalarObjectNode* old_sosn = old_in->as_SafePointScalarObject();
kvn@63 64 uint old_unique = C->unique();
kvn@63 65 Node* new_in = old_sosn->clone(jvms_adj, sosn_map);
kvn@63 66 if (old_unique != C->unique()) {
kvn@63 67 new_in = transform_later(new_in); // Register new node.
kvn@63 68 }
kvn@63 69 old_in = new_in;
kvn@63 70 }
kvn@63 71 newcall->add_req(old_in);
duke@0 72 }
kvn@63 73
duke@0 74 newcall->set_jvms(oldcall->jvms());
duke@0 75 for (JVMState *jvms = newcall->jvms(); jvms != NULL; jvms = jvms->caller()) {
duke@0 76 jvms->set_map(newcall);
duke@0 77 jvms->set_locoff(jvms->locoff()+jvms_adj);
duke@0 78 jvms->set_stkoff(jvms->stkoff()+jvms_adj);
duke@0 79 jvms->set_monoff(jvms->monoff()+jvms_adj);
kvn@63 80 jvms->set_scloff(jvms->scloff()+jvms_adj);
duke@0 81 jvms->set_endoff(jvms->endoff()+jvms_adj);
duke@0 82 }
duke@0 83 }
duke@0 84
kvn@423 85 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word, int mask, int bits, bool return_fast_path) {
kvn@423 86 Node* cmp;
kvn@423 87 if (mask != 0) {
kvn@423 88 Node* and_node = transform_later(new (C, 3) AndXNode(word, MakeConX(mask)));
kvn@423 89 cmp = transform_later(new (C, 3) CmpXNode(and_node, MakeConX(bits)));
kvn@423 90 } else {
kvn@423 91 cmp = word;
kvn@423 92 }
kvn@423 93 Node* bol = transform_later(new (C, 2) BoolNode(cmp, BoolTest::ne));
kvn@423 94 IfNode* iff = new (C, 2) IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN );
kvn@423 95 transform_later(iff);
duke@0 96
kvn@423 97 // Fast path taken.
kvn@423 98 Node *fast_taken = transform_later( new (C, 1) IfFalseNode(iff) );
duke@0 99
duke@0 100 // Fast path not-taken, i.e. slow path
kvn@423 101 Node *slow_taken = transform_later( new (C, 1) IfTrueNode(iff) );
kvn@423 102
kvn@423 103 if (return_fast_path) {
kvn@423 104 region->init_req(edge, slow_taken); // Capture slow-control
kvn@423 105 return fast_taken;
kvn@423 106 } else {
kvn@423 107 region->init_req(edge, fast_taken); // Capture fast-control
kvn@423 108 return slow_taken;
kvn@423 109 }
duke@0 110 }
duke@0 111
duke@0 112 //--------------------copy_predefined_input_for_runtime_call--------------------
duke@0 113 void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) {
duke@0 114 // Set fixed predefined input arguments
duke@0 115 call->init_req( TypeFunc::Control, ctrl );
duke@0 116 call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) );
duke@0 117 call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ?????
duke@0 118 call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) );
duke@0 119 call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) );
duke@0 120 }
duke@0 121
duke@0 122 //------------------------------make_slow_call---------------------------------
duke@0 123 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) {
duke@0 124
duke@0 125 // Slow-path call
duke@0 126 int size = slow_call_type->domain()->cnt();
duke@0 127 CallNode *call = leaf_name
duke@0 128 ? (CallNode*)new (C, size) CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM )
duke@0 129 : (CallNode*)new (C, size) CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), oldcall->jvms()->bci(), TypeRawPtr::BOTTOM );
duke@0 130
duke@0 131 // Slow path call has no side-effects, uses few values
duke@0 132 copy_predefined_input_for_runtime_call(slow_path, oldcall, call );
duke@0 133 if (parm0 != NULL) call->init_req(TypeFunc::Parms+0, parm0);
duke@0 134 if (parm1 != NULL) call->init_req(TypeFunc::Parms+1, parm1);
duke@0 135 copy_call_debug_info(oldcall, call);
duke@0 136 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
duke@0 137 _igvn.hash_delete(oldcall);
duke@0 138 _igvn.subsume_node(oldcall, call);
duke@0 139 transform_later(call);
duke@0 140
duke@0 141 return call;
duke@0 142 }
duke@0 143
duke@0 144 void PhaseMacroExpand::extract_call_projections(CallNode *call) {
duke@0 145 _fallthroughproj = NULL;
duke@0 146 _fallthroughcatchproj = NULL;
duke@0 147 _ioproj_fallthrough = NULL;
duke@0 148 _ioproj_catchall = NULL;
duke@0 149 _catchallcatchproj = NULL;
duke@0 150 _memproj_fallthrough = NULL;
duke@0 151 _memproj_catchall = NULL;
duke@0 152 _resproj = NULL;
duke@0 153 for (DUIterator_Fast imax, i = call->fast_outs(imax); i < imax; i++) {
duke@0 154 ProjNode *pn = call->fast_out(i)->as_Proj();
duke@0 155 switch (pn->_con) {
duke@0 156 case TypeFunc::Control:
duke@0 157 {
duke@0 158 // For Control (fallthrough) and I_O (catch_all_index) we have CatchProj -> Catch -> Proj
duke@0 159 _fallthroughproj = pn;
duke@0 160 DUIterator_Fast jmax, j = pn->fast_outs(jmax);
duke@0 161 const Node *cn = pn->fast_out(j);
duke@0 162 if (cn->is_Catch()) {
duke@0 163 ProjNode *cpn = NULL;
duke@0 164 for (DUIterator_Fast kmax, k = cn->fast_outs(kmax); k < kmax; k++) {
duke@0 165 cpn = cn->fast_out(k)->as_Proj();
duke@0 166 assert(cpn->is_CatchProj(), "must be a CatchProjNode");
duke@0 167 if (cpn->_con == CatchProjNode::fall_through_index)
duke@0 168 _fallthroughcatchproj = cpn;
duke@0 169 else {
duke@0 170 assert(cpn->_con == CatchProjNode::catch_all_index, "must be correct index.");
duke@0 171 _catchallcatchproj = cpn;
duke@0 172 }
duke@0 173 }
duke@0 174 }
duke@0 175 break;
duke@0 176 }
duke@0 177 case TypeFunc::I_O:
duke@0 178 if (pn->_is_io_use)
duke@0 179 _ioproj_catchall = pn;
duke@0 180 else
duke@0 181 _ioproj_fallthrough = pn;
duke@0 182 break;
duke@0 183 case TypeFunc::Memory:
duke@0 184 if (pn->_is_io_use)
duke@0 185 _memproj_catchall = pn;
duke@0 186 else
duke@0 187 _memproj_fallthrough = pn;
duke@0 188 break;
duke@0 189 case TypeFunc::Parms:
duke@0 190 _resproj = pn;
duke@0 191 break;
duke@0 192 default:
duke@0 193 assert(false, "unexpected projection from allocation node.");
duke@0 194 }
duke@0 195 }
duke@0 196
duke@0 197 }
duke@0 198
kvn@73 199 // Eliminate a card mark sequence. p2x is a ConvP2XNode
kvn@73 200 void PhaseMacroExpand::eliminate_card_mark(Node *p2x) {
kvn@73 201 assert(p2x->Opcode() == Op_CastP2X, "ConvP2XNode required");
kvn@73 202 Node *shift = p2x->unique_out();
kvn@73 203 Node *addp = shift->unique_out();
kvn@73 204 for (DUIterator_Last jmin, j = addp->last_outs(jmin); j >= jmin; --j) {
kvn@73 205 Node *st = addp->last_out(j);
kvn@73 206 assert(st->is_Store(), "store required");
kvn@73 207 _igvn.replace_node(st, st->in(MemNode::Memory));
kvn@73 208 }
kvn@73 209 }
kvn@73 210
kvn@73 211 // Search for a memory operation for the specified memory slice.
kvn@256 212 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) {
kvn@73 213 Node *orig_mem = mem;
kvn@73 214 Node *alloc_mem = alloc->in(TypeFunc::Memory);
kvn@256 215 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr();
kvn@73 216 while (true) {
kvn@73 217 if (mem == alloc_mem || mem == start_mem ) {
kvn@73 218 return mem; // hit one of our sentinals
kvn@73 219 } else if (mem->is_MergeMem()) {
kvn@73 220 mem = mem->as_MergeMem()->memory_at(alias_idx);
kvn@73 221 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) {
kvn@73 222 Node *in = mem->in(0);
kvn@73 223 // we can safely skip over safepoints, calls, locks and membars because we
kvn@73 224 // already know that the object is safe to eliminate.
kvn@73 225 if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) {
kvn@73 226 return in;
kvn@256 227 } else if (in->is_Call()) {
kvn@256 228 CallNode *call = in->as_Call();
kvn@256 229 if (!call->may_modify(tinst, phase)) {
kvn@256 230 mem = call->in(TypeFunc::Memory);
kvn@256 231 }
kvn@256 232 mem = in->in(TypeFunc::Memory);
kvn@256 233 } else if (in->is_MemBar()) {
kvn@73 234 mem = in->in(TypeFunc::Memory);
kvn@73 235 } else {
kvn@73 236 assert(false, "unexpected projection");
kvn@73 237 }
kvn@73 238 } else if (mem->is_Store()) {
kvn@73 239 const TypePtr* atype = mem->as_Store()->adr_type();
kvn@73 240 int adr_idx = Compile::current()->get_alias_index(atype);
kvn@73 241 if (adr_idx == alias_idx) {
kvn@73 242 assert(atype->isa_oopptr(), "address type must be oopptr");
kvn@73 243 int adr_offset = atype->offset();
kvn@73 244 uint adr_iid = atype->is_oopptr()->instance_id();
kvn@73 245 // Array elements references have the same alias_idx
kvn@73 246 // but different offset and different instance_id.
kvn@73 247 if (adr_offset == offset && adr_iid == alloc->_idx)
kvn@73 248 return mem;
kvn@73 249 } else {
kvn@73 250 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw");
kvn@73 251 }
kvn@73 252 mem = mem->in(MemNode::Memory);
kvn@73 253 } else {
kvn@73 254 return mem;
kvn@73 255 }
kvn@250 256 assert(mem != orig_mem, "dead memory loop");
kvn@73 257 }
kvn@73 258 }
kvn@73 259
kvn@73 260 //
kvn@73 261 // Given a Memory Phi, compute a value Phi containing the values from stores
kvn@73 262 // on the input paths.
kvn@73 263 // Note: this function is recursive, its depth is limied by the "level" argument
kvn@73 264 // Returns the computed Phi, or NULL if it cannot compute it.
kvn@250 265 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) {
kvn@250 266 assert(mem->is_Phi(), "sanity");
kvn@250 267 int alias_idx = C->get_alias_index(adr_t);
kvn@250 268 int offset = adr_t->offset();
kvn@250 269 int instance_id = adr_t->instance_id();
kvn@250 270
kvn@250 271 // Check if an appropriate value phi already exists.
kvn@250 272 Node* region = mem->in(0);
kvn@250 273 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) {
kvn@250 274 Node* phi = region->fast_out(k);
kvn@250 275 if (phi->is_Phi() && phi != mem &&
kvn@250 276 phi->as_Phi()->is_same_inst_field(phi_type, instance_id, alias_idx, offset)) {
kvn@250 277 return phi;
kvn@250 278 }
kvn@250 279 }
kvn@250 280 // Check if an appropriate new value phi already exists.
kvn@250 281 Node* new_phi = NULL;
kvn@250 282 uint size = value_phis->size();
kvn@250 283 for (uint i=0; i < size; i++) {
kvn@250 284 if ( mem->_idx == value_phis->index_at(i) ) {
kvn@250 285 return value_phis->node_at(i);
kvn@250 286 }
kvn@250 287 }
kvn@73 288
kvn@73 289 if (level <= 0) {
kvn@256 290 return NULL; // Give up: phi tree too deep
kvn@73 291 }
kvn@73 292 Node *start_mem = C->start()->proj_out(TypeFunc::Memory);
kvn@73 293 Node *alloc_mem = alloc->in(TypeFunc::Memory);
kvn@73 294
kvn@73 295 uint length = mem->req();
kvn@73 296 GrowableArray <Node *> values(length, length, NULL);
kvn@73 297
kvn@250 298 // create a new Phi for the value
kvn@250 299 PhiNode *phi = new (C, length) PhiNode(mem->in(0), phi_type, NULL, instance_id, alias_idx, offset);
kvn@250 300 transform_later(phi);
kvn@250 301 value_phis->push(phi, mem->_idx);
kvn@250 302
kvn@73 303 for (uint j = 1; j < length; j++) {
kvn@73 304 Node *in = mem->in(j);
kvn@73 305 if (in == NULL || in->is_top()) {
kvn@73 306 values.at_put(j, in);
kvn@73 307 } else {
kvn@256 308 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn);
kvn@73 309 if (val == start_mem || val == alloc_mem) {
kvn@73 310 // hit a sentinel, return appropriate 0 value
kvn@73 311 values.at_put(j, _igvn.zerocon(ft));
kvn@73 312 continue;
kvn@73 313 }
kvn@73 314 if (val->is_Initialize()) {
kvn@73 315 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
kvn@73 316 }
kvn@73 317 if (val == NULL) {
kvn@73 318 return NULL; // can't find a value on this path
kvn@73 319 }
kvn@73 320 if (val == mem) {
kvn@73 321 values.at_put(j, mem);
kvn@73 322 } else if (val->is_Store()) {
kvn@73 323 values.at_put(j, val->in(MemNode::ValueIn));
kvn@73 324 } else if(val->is_Proj() && val->in(0) == alloc) {
kvn@73 325 values.at_put(j, _igvn.zerocon(ft));
kvn@73 326 } else if (val->is_Phi()) {
kvn@250 327 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1);
kvn@250 328 if (val == NULL) {
kvn@250 329 return NULL;
kvn@73 330 }
kvn@250 331 values.at_put(j, val);
kvn@73 332 } else {
kvn@256 333 assert(false, "unknown node on this path");
kvn@256 334 return NULL; // unknown node on this path
kvn@73 335 }
kvn@73 336 }
kvn@73 337 }
kvn@250 338 // Set Phi's inputs
kvn@73 339 for (uint j = 1; j < length; j++) {
kvn@73 340 if (values.at(j) == mem) {
kvn@73 341 phi->init_req(j, phi);
kvn@73 342 } else {
kvn@73 343 phi->init_req(j, values.at(j));
kvn@73 344 }
kvn@73 345 }
kvn@73 346 return phi;
kvn@73 347 }
kvn@73 348
kvn@73 349 // Search the last value stored into the object's field.
kvn@73 350 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, Node *alloc) {
kvn@223 351 assert(adr_t->is_known_instance_field(), "instance required");
kvn@223 352 int instance_id = adr_t->instance_id();
kvn@223 353 assert((uint)instance_id == alloc->_idx, "wrong allocation");
kvn@73 354
kvn@73 355 int alias_idx = C->get_alias_index(adr_t);
kvn@73 356 int offset = adr_t->offset();
kvn@73 357 Node *start_mem = C->start()->proj_out(TypeFunc::Memory);
kvn@73 358 Node *alloc_ctrl = alloc->in(TypeFunc::Control);
kvn@73 359 Node *alloc_mem = alloc->in(TypeFunc::Memory);
kvn@250 360 Arena *a = Thread::current()->resource_area();
kvn@250 361 VectorSet visited(a);
kvn@73 362
kvn@73 363
kvn@73 364 bool done = sfpt_mem == alloc_mem;
kvn@73 365 Node *mem = sfpt_mem;
kvn@73 366 while (!done) {
kvn@73 367 if (visited.test_set(mem->_idx)) {
kvn@73 368 return NULL; // found a loop, give up
kvn@73 369 }
kvn@256 370 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn);
kvn@73 371 if (mem == start_mem || mem == alloc_mem) {
kvn@73 372 done = true; // hit a sentinel, return appropriate 0 value
kvn@73 373 } else if (mem->is_Initialize()) {
kvn@73 374 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
kvn@73 375 if (mem == NULL) {
kvn@73 376 done = true; // Something go wrong.
kvn@73 377 } else if (mem->is_Store()) {
kvn@73 378 const TypePtr* atype = mem->as_Store()->adr_type();
kvn@73 379 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice");
kvn@73 380 done = true;
kvn@73 381 }
kvn@73 382 } else if (mem->is_Store()) {
kvn@73 383 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr();
kvn@73 384 assert(atype != NULL, "address type must be oopptr");
kvn@73 385 assert(C->get_alias_index(atype) == alias_idx &&
kvn@223 386 atype->is_known_instance_field() && atype->offset() == offset &&
kvn@73 387 atype->instance_id() == instance_id, "store is correct memory slice");
kvn@73 388 done = true;
kvn@73 389 } else if (mem->is_Phi()) {
kvn@73 390 // try to find a phi's unique input
kvn@73 391 Node *unique_input = NULL;
kvn@73 392 Node *top = C->top();
kvn@73 393 for (uint i = 1; i < mem->req(); i++) {
kvn@256 394 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn);
kvn@73 395 if (n == NULL || n == top || n == mem) {
kvn@73 396 continue;
kvn@73 397 } else if (unique_input == NULL) {
kvn@73 398 unique_input = n;
kvn@73 399 } else if (unique_input != n) {
kvn@73 400 unique_input = top;
kvn@73 401 break;
kvn@73 402 }
kvn@73 403 }
kvn@73 404 if (unique_input != NULL && unique_input != top) {
kvn@73 405 mem = unique_input;
kvn@73 406 } else {
kvn@73 407 done = true;
kvn@73 408 }
kvn@73 409 } else {
kvn@73 410 assert(false, "unexpected node");
kvn@73 411 }
kvn@73 412 }
kvn@73 413 if (mem != NULL) {
kvn@73 414 if (mem == start_mem || mem == alloc_mem) {
kvn@73 415 // hit a sentinel, return appropriate 0 value
kvn@73 416 return _igvn.zerocon(ft);
kvn@73 417 } else if (mem->is_Store()) {
kvn@73 418 return mem->in(MemNode::ValueIn);
kvn@73 419 } else if (mem->is_Phi()) {
kvn@73 420 // attempt to produce a Phi reflecting the values on the input paths of the Phi
kvn@250 421 Node_Stack value_phis(a, 8);
kvn@256 422 Node * phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit);
kvn@73 423 if (phi != NULL) {
kvn@73 424 return phi;
kvn@250 425 } else {
kvn@250 426 // Kill all new Phis
kvn@250 427 while(value_phis.is_nonempty()) {
kvn@250 428 Node* n = value_phis.node();
kvn@250 429 _igvn.hash_delete(n);
kvn@250 430 _igvn.subsume_node(n, C->top());
kvn@250 431 value_phis.pop();
kvn@250 432 }
kvn@73 433 }
kvn@73 434 }
kvn@73 435 }
kvn@73 436 // Something go wrong.
kvn@73 437 return NULL;
kvn@73 438 }
kvn@73 439
kvn@73 440 // Check the possibility of scalar replacement.
kvn@73 441 bool PhaseMacroExpand::can_eliminate_allocation(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
kvn@73 442 // Scan the uses of the allocation to check for anything that would
kvn@73 443 // prevent us from eliminating it.
kvn@73 444 NOT_PRODUCT( const char* fail_eliminate = NULL; )
kvn@73 445 DEBUG_ONLY( Node* disq_node = NULL; )
kvn@73 446 bool can_eliminate = true;
kvn@73 447
kvn@73 448 Node* res = alloc->result_cast();
kvn@73 449 const TypeOopPtr* res_type = NULL;
kvn@73 450 if (res == NULL) {
kvn@73 451 // All users were eliminated.
kvn@73 452 } else if (!res->is_CheckCastPP()) {
kvn@73 453 alloc->_is_scalar_replaceable = false; // don't try again
kvn@73 454 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";)
kvn@73 455 can_eliminate = false;
kvn@73 456 } else {
kvn@73 457 res_type = _igvn.type(res)->isa_oopptr();
kvn@73 458 if (res_type == NULL) {
kvn@73 459 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";)
kvn@73 460 can_eliminate = false;
kvn@73 461 } else if (res_type->isa_aryptr()) {
kvn@73 462 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1);
kvn@73 463 if (length < 0) {
kvn@73 464 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";)
kvn@73 465 can_eliminate = false;
kvn@73 466 }
kvn@73 467 }
kvn@73 468 }
kvn@73 469
kvn@73 470 if (can_eliminate && res != NULL) {
kvn@73 471 for (DUIterator_Fast jmax, j = res->fast_outs(jmax);
kvn@73 472 j < jmax && can_eliminate; j++) {
kvn@73 473 Node* use = res->fast_out(j);
kvn@73 474
kvn@73 475 if (use->is_AddP()) {
kvn@73 476 const TypePtr* addp_type = _igvn.type(use)->is_ptr();
kvn@73 477 int offset = addp_type->offset();
kvn@73 478
kvn@73 479 if (offset == Type::OffsetTop || offset == Type::OffsetBot) {
kvn@73 480 NOT_PRODUCT(fail_eliminate = "Undefined field referrence";)
kvn@73 481 can_eliminate = false;
kvn@73 482 break;
kvn@73 483 }
kvn@73 484 for (DUIterator_Fast kmax, k = use->fast_outs(kmax);
kvn@73 485 k < kmax && can_eliminate; k++) {
kvn@73 486 Node* n = use->fast_out(k);
kvn@73 487 if (!n->is_Store() && n->Opcode() != Op_CastP2X) {
kvn@73 488 DEBUG_ONLY(disq_node = n;)
kvn@256 489 if (n->is_Load() || n->is_LoadStore()) {
kvn@73 490 NOT_PRODUCT(fail_eliminate = "Field load";)
kvn@73 491 } else {
kvn@73 492 NOT_PRODUCT(fail_eliminate = "Not store field referrence";)
kvn@73 493 }
kvn@73 494 can_eliminate = false;
kvn@73 495 }
kvn@73 496 }
kvn@73 497 } else if (use->is_SafePoint()) {
kvn@73 498 SafePointNode* sfpt = use->as_SafePoint();
kvn@168 499 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) {
kvn@73 500 // Object is passed as argument.
kvn@73 501 DEBUG_ONLY(disq_node = use;)
kvn@73 502 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";)
kvn@73 503 can_eliminate = false;
kvn@73 504 }
kvn@73 505 Node* sfptMem = sfpt->memory();
kvn@73 506 if (sfptMem == NULL || sfptMem->is_top()) {
kvn@73 507 DEBUG_ONLY(disq_node = use;)
kvn@73 508 NOT_PRODUCT(fail_eliminate = "NULL or TOP memory";)
kvn@73 509 can_eliminate = false;
kvn@73 510 } else {
kvn@73 511 safepoints.append_if_missing(sfpt);
kvn@73 512 }
kvn@73 513 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
kvn@73 514 if (use->is_Phi()) {
kvn@73 515 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
kvn@73 516 NOT_PRODUCT(fail_eliminate = "Object is return value";)
kvn@73 517 } else {
kvn@73 518 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
kvn@73 519 }
kvn@73 520 DEBUG_ONLY(disq_node = use;)
kvn@73 521 } else {
kvn@73 522 if (use->Opcode() == Op_Return) {
kvn@73 523 NOT_PRODUCT(fail_eliminate = "Object is return value";)
kvn@73 524 }else {
kvn@73 525 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
kvn@73 526 }
kvn@73 527 DEBUG_ONLY(disq_node = use;)
kvn@73 528 }
kvn@73 529 can_eliminate = false;
kvn@73 530 }
kvn@73 531 }
kvn@73 532 }
kvn@73 533
kvn@73 534 #ifndef PRODUCT
kvn@73 535 if (PrintEliminateAllocations) {
kvn@73 536 if (can_eliminate) {
kvn@73 537 tty->print("Scalar ");
kvn@73 538 if (res == NULL)
kvn@73 539 alloc->dump();
kvn@73 540 else
kvn@73 541 res->dump();
kvn@73 542 } else {
kvn@73 543 tty->print("NotScalar (%s)", fail_eliminate);
kvn@73 544 if (res == NULL)
kvn@73 545 alloc->dump();
kvn@73 546 else
kvn@73 547 res->dump();
kvn@73 548 #ifdef ASSERT
kvn@73 549 if (disq_node != NULL) {
kvn@73 550 tty->print(" >>>> ");
kvn@73 551 disq_node->dump();
kvn@73 552 }
kvn@73 553 #endif /*ASSERT*/
kvn@73 554 }
kvn@73 555 }
kvn@73 556 #endif
kvn@73 557 return can_eliminate;
kvn@73 558 }
kvn@73 559
kvn@73 560 // Do scalar replacement.
kvn@73 561 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
kvn@73 562 GrowableArray <SafePointNode *> safepoints_done;
kvn@73 563
kvn@73 564 ciKlass* klass = NULL;
kvn@73 565 ciInstanceKlass* iklass = NULL;
kvn@73 566 int nfields = 0;
kvn@73 567 int array_base;
kvn@73 568 int element_size;
kvn@73 569 BasicType basic_elem_type;
kvn@73 570 ciType* elem_type;
kvn@73 571
kvn@73 572 Node* res = alloc->result_cast();
kvn@73 573 const TypeOopPtr* res_type = NULL;
kvn@73 574 if (res != NULL) { // Could be NULL when there are no users
kvn@73 575 res_type = _igvn.type(res)->isa_oopptr();
kvn@73 576 }
kvn@73 577
kvn@73 578 if (res != NULL) {
kvn@73 579 klass = res_type->klass();
kvn@73 580 if (res_type->isa_instptr()) {
kvn@73 581 // find the fields of the class which will be needed for safepoint debug information
kvn@73 582 assert(klass->is_instance_klass(), "must be an instance klass.");
kvn@73 583 iklass = klass->as_instance_klass();
kvn@73 584 nfields = iklass->nof_nonstatic_fields();
kvn@73 585 } else {
kvn@73 586 // find the array's elements which will be needed for safepoint debug information
kvn@73 587 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
kvn@73 588 assert(klass->is_array_klass() && nfields >= 0, "must be an array klass.");
kvn@73 589 elem_type = klass->as_array_klass()->element_type();
kvn@73 590 basic_elem_type = elem_type->basic_type();
kvn@73 591 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
kvn@73 592 element_size = type2aelembytes(basic_elem_type);
kvn@73 593 }
kvn@73 594 }
kvn@73 595 //
kvn@73 596 // Process the safepoint uses
kvn@73 597 //
kvn@73 598 while (safepoints.length() > 0) {
kvn@73 599 SafePointNode* sfpt = safepoints.pop();
kvn@73 600 Node* mem = sfpt->memory();
kvn@73 601 uint first_ind = sfpt->req();
kvn@73 602 SafePointScalarObjectNode* sobj = new (C, 1) SafePointScalarObjectNode(res_type,
kvn@73 603 #ifdef ASSERT
kvn@73 604 alloc,
kvn@73 605 #endif
kvn@73 606 first_ind, nfields);
kvn@73 607 sobj->init_req(0, sfpt->in(TypeFunc::Control));
kvn@73 608 transform_later(sobj);
kvn@73 609
kvn@73 610 // Scan object's fields adding an input to the safepoint for each field.
kvn@73 611 for (int j = 0; j < nfields; j++) {
kvn@309 612 intptr_t offset;
kvn@73 613 ciField* field = NULL;
kvn@73 614 if (iklass != NULL) {
kvn@73 615 field = iklass->nonstatic_field_at(j);
kvn@73 616 offset = field->offset();
kvn@73 617 elem_type = field->type();
kvn@73 618 basic_elem_type = field->layout_type();
kvn@73 619 } else {
kvn@309 620 offset = array_base + j * (intptr_t)element_size;
kvn@73 621 }
kvn@73 622
kvn@73 623 const Type *field_type;
kvn@73 624 // The next code is taken from Parse::do_get_xxx().
kvn@124 625 if (basic_elem_type == T_OBJECT || basic_elem_type == T_ARRAY) {
kvn@73 626 if (!elem_type->is_loaded()) {
kvn@73 627 field_type = TypeInstPtr::BOTTOM;
kvn@73 628 } else if (field != NULL && field->is_constant()) {
kvn@73 629 // This can happen if the constant oop is non-perm.
kvn@73 630 ciObject* con = field->constant_value().as_object();
kvn@73 631 // Do not "join" in the previous type; it doesn't add value,
kvn@73 632 // and may yield a vacuous result if the field is of interface type.
kvn@73 633 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr();
kvn@73 634 assert(field_type != NULL, "field singleton type must be consistent");
kvn@73 635 } else {
kvn@73 636 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
kvn@73 637 }
kvn@124 638 if (UseCompressedOops) {
kvn@221 639 field_type = field_type->make_narrowoop();
kvn@124 640 basic_elem_type = T_NARROWOOP;
kvn@124 641 }
kvn@73 642 } else {
kvn@73 643 field_type = Type::get_const_basic_type(basic_elem_type);
kvn@73 644 }
kvn@73 645
kvn@73 646 const TypeOopPtr *field_addr_type = res_type->add_offset(offset)->isa_oopptr();
kvn@73 647
kvn@73 648 Node *field_val = value_from_mem(mem, basic_elem_type, field_type, field_addr_type, alloc);
kvn@73 649 if (field_val == NULL) {
kvn@73 650 // we weren't able to find a value for this field,
kvn@73 651 // give up on eliminating this allocation
kvn@73 652 alloc->_is_scalar_replaceable = false; // don't try again
kvn@73 653 // remove any extra entries we added to the safepoint
kvn@73 654 uint last = sfpt->req() - 1;
kvn@73 655 for (int k = 0; k < j; k++) {
kvn@73 656 sfpt->del_req(last--);
kvn@73 657 }
kvn@73 658 // rollback processed safepoints
kvn@73 659 while (safepoints_done.length() > 0) {
kvn@73 660 SafePointNode* sfpt_done = safepoints_done.pop();
kvn@73 661 // remove any extra entries we added to the safepoint
kvn@73 662 last = sfpt_done->req() - 1;
kvn@73 663 for (int k = 0; k < nfields; k++) {
kvn@73 664 sfpt_done->del_req(last--);
kvn@73 665 }
kvn@73 666 JVMState *jvms = sfpt_done->jvms();
kvn@73 667 jvms->set_endoff(sfpt_done->req());
kvn@73 668 // Now make a pass over the debug information replacing any references
kvn@73 669 // to SafePointScalarObjectNode with the allocated object.
kvn@73 670 int start = jvms->debug_start();
kvn@73 671 int end = jvms->debug_end();
kvn@73 672 for (int i = start; i < end; i++) {
kvn@73 673 if (sfpt_done->in(i)->is_SafePointScalarObject()) {
kvn@73 674 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject();
kvn@73 675 if (scobj->first_index() == sfpt_done->req() &&
kvn@73 676 scobj->n_fields() == (uint)nfields) {
kvn@73 677 assert(scobj->alloc() == alloc, "sanity");
kvn@73 678 sfpt_done->set_req(i, res);
kvn@73 679 }
kvn@73 680 }
kvn@73 681 }
kvn@73 682 }
kvn@73 683 #ifndef PRODUCT
kvn@73 684 if (PrintEliminateAllocations) {
kvn@73 685 if (field != NULL) {
kvn@73 686 tty->print("=== At SafePoint node %d can't find value of Field: ",
kvn@73 687 sfpt->_idx);
kvn@73 688 field->print();
kvn@73 689 int field_idx = C->get_alias_index(field_addr_type);
kvn@73 690 tty->print(" (alias_idx=%d)", field_idx);
kvn@73 691 } else { // Array's element
kvn@73 692 tty->print("=== At SafePoint node %d can't find value of array element [%d]",
kvn@73 693 sfpt->_idx, j);
kvn@73 694 }
kvn@73 695 tty->print(", which prevents elimination of: ");
kvn@73 696 if (res == NULL)
kvn@73 697 alloc->dump();
kvn@73 698 else
kvn@73 699 res->dump();
kvn@73 700 }
kvn@73 701 #endif
kvn@73 702 return false;
kvn@73 703 }
kvn@124 704 if (UseCompressedOops && field_type->isa_narrowoop()) {
kvn@124 705 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
kvn@124 706 // to be able scalar replace the allocation.
kvn@221 707 if (field_val->is_EncodeP()) {
kvn@221 708 field_val = field_val->in(1);
kvn@221 709 } else {
kvn@221 710 field_val = transform_later(new (C, 2) DecodeNNode(field_val, field_val->bottom_type()->make_ptr()));
kvn@221 711 }
kvn@124 712 }
kvn@73 713 sfpt->add_req(field_val);
kvn@73 714 }
kvn@73 715 JVMState *jvms = sfpt->jvms();
kvn@73 716 jvms->set_endoff(sfpt->req());
kvn@73 717 // Now make a pass over the debug information replacing any references
kvn@73 718 // to the allocated object with "sobj"
kvn@73 719 int start = jvms->debug_start();
kvn@73 720 int end = jvms->debug_end();
kvn@73 721 for (int i = start; i < end; i++) {
kvn@73 722 if (sfpt->in(i) == res) {
kvn@73 723 sfpt->set_req(i, sobj);
kvn@73 724 }
kvn@73 725 }
kvn@73 726 safepoints_done.append_if_missing(sfpt); // keep it for rollback
kvn@73 727 }
kvn@73 728 return true;
kvn@73 729 }
kvn@73 730
kvn@73 731 // Process users of eliminated allocation.
kvn@73 732 void PhaseMacroExpand::process_users_of_allocation(AllocateNode *alloc) {
kvn@73 733 Node* res = alloc->result_cast();
kvn@73 734 if (res != NULL) {
kvn@73 735 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
kvn@73 736 Node *use = res->last_out(j);
kvn@73 737 uint oc1 = res->outcnt();
kvn@73 738
kvn@73 739 if (use->is_AddP()) {
kvn@73 740 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) {
kvn@73 741 Node *n = use->last_out(k);
kvn@73 742 uint oc2 = use->outcnt();
kvn@73 743 if (n->is_Store()) {
kvn@73 744 _igvn.replace_node(n, n->in(MemNode::Memory));
kvn@73 745 } else {
kvn@73 746 assert( n->Opcode() == Op_CastP2X, "CastP2X required");
kvn@73 747 eliminate_card_mark(n);
kvn@73 748 }
kvn@73 749 k -= (oc2 - use->outcnt());
kvn@73 750 }
kvn@73 751 } else {
kvn@73 752 assert( !use->is_SafePoint(), "safepoint uses must have been already elimiated");
kvn@73 753 assert( use->Opcode() == Op_CastP2X, "CastP2X required");
kvn@73 754 eliminate_card_mark(use);
kvn@73 755 }
kvn@73 756 j -= (oc1 - res->outcnt());
kvn@73 757 }
kvn@73 758 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
kvn@73 759 _igvn.remove_dead_node(res);
kvn@73 760 }
kvn@73 761
kvn@73 762 //
kvn@73 763 // Process other users of allocation's projections
kvn@73 764 //
kvn@73 765 if (_resproj != NULL && _resproj->outcnt() != 0) {
kvn@73 766 for (DUIterator_Last jmin, j = _resproj->last_outs(jmin); j >= jmin; ) {
kvn@73 767 Node *use = _resproj->last_out(j);
kvn@73 768 uint oc1 = _resproj->outcnt();
kvn@73 769 if (use->is_Initialize()) {
kvn@73 770 // Eliminate Initialize node.
kvn@73 771 InitializeNode *init = use->as_Initialize();
kvn@73 772 assert(init->outcnt() <= 2, "only a control and memory projection expected");
kvn@73 773 Node *ctrl_proj = init->proj_out(TypeFunc::Control);
kvn@73 774 if (ctrl_proj != NULL) {
kvn@73 775 assert(init->in(TypeFunc::Control) == _fallthroughcatchproj, "allocation control projection");
kvn@73 776 _igvn.replace_node(ctrl_proj, _fallthroughcatchproj);
kvn@73 777 }
kvn@73 778 Node *mem_proj = init->proj_out(TypeFunc::Memory);
kvn@73 779 if (mem_proj != NULL) {
kvn@73 780 Node *mem = init->in(TypeFunc::Memory);
kvn@73 781 #ifdef ASSERT
kvn@73 782 if (mem->is_MergeMem()) {
kvn@73 783 assert(mem->in(TypeFunc::Memory) == _memproj_fallthrough, "allocation memory projection");
kvn@73 784 } else {
kvn@73 785 assert(mem == _memproj_fallthrough, "allocation memory projection");
kvn@73 786 }
kvn@73 787 #endif
kvn@73 788 _igvn.replace_node(mem_proj, mem);
kvn@73 789 }
kvn@73 790 } else if (use->is_AddP()) {
kvn@73 791 // raw memory addresses used only by the initialization
kvn@73 792 _igvn.hash_delete(use);
kvn@73 793 _igvn.subsume_node(use, C->top());
kvn@73 794 } else {
kvn@73 795 assert(false, "only Initialize or AddP expected");
kvn@73 796 }
kvn@73 797 j -= (oc1 - _resproj->outcnt());
kvn@73 798 }
kvn@73 799 }
kvn@73 800 if (_fallthroughcatchproj != NULL) {
kvn@73 801 _igvn.replace_node(_fallthroughcatchproj, alloc->in(TypeFunc::Control));
kvn@73 802 }
kvn@73 803 if (_memproj_fallthrough != NULL) {
kvn@73 804 _igvn.replace_node(_memproj_fallthrough, alloc->in(TypeFunc::Memory));
kvn@73 805 }
kvn@73 806 if (_memproj_catchall != NULL) {
kvn@73 807 _igvn.replace_node(_memproj_catchall, C->top());
kvn@73 808 }
kvn@73 809 if (_ioproj_fallthrough != NULL) {
kvn@73 810 _igvn.replace_node(_ioproj_fallthrough, alloc->in(TypeFunc::I_O));
kvn@73 811 }
kvn@73 812 if (_ioproj_catchall != NULL) {
kvn@73 813 _igvn.replace_node(_ioproj_catchall, C->top());
kvn@73 814 }
kvn@73 815 if (_catchallcatchproj != NULL) {
kvn@73 816 _igvn.replace_node(_catchallcatchproj, C->top());
kvn@73 817 }
kvn@73 818 }
kvn@73 819
kvn@73 820 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) {
kvn@73 821
kvn@73 822 if (!EliminateAllocations || !alloc->_is_scalar_replaceable) {
kvn@73 823 return false;
kvn@73 824 }
kvn@73 825
kvn@73 826 extract_call_projections(alloc);
kvn@73 827
kvn@73 828 GrowableArray <SafePointNode *> safepoints;
kvn@73 829 if (!can_eliminate_allocation(alloc, safepoints)) {
kvn@73 830 return false;
kvn@73 831 }
kvn@73 832
kvn@73 833 if (!scalar_replacement(alloc, safepoints)) {
kvn@73 834 return false;
kvn@73 835 }
kvn@73 836
kvn@73 837 process_users_of_allocation(alloc);
kvn@73 838
kvn@73 839 #ifndef PRODUCT
kvn@73 840 if (PrintEliminateAllocations) {
kvn@73 841 if (alloc->is_AllocateArray())
kvn@73 842 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
kvn@73 843 else
kvn@73 844 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
kvn@73 845 }
kvn@73 846 #endif
kvn@73 847
kvn@73 848 return true;
kvn@73 849 }
kvn@73 850
duke@0 851
duke@0 852 //---------------------------set_eden_pointers-------------------------
duke@0 853 void PhaseMacroExpand::set_eden_pointers(Node* &eden_top_adr, Node* &eden_end_adr) {
duke@0 854 if (UseTLAB) { // Private allocation: load from TLS
duke@0 855 Node* thread = transform_later(new (C, 1) ThreadLocalNode());
duke@0 856 int tlab_top_offset = in_bytes(JavaThread::tlab_top_offset());
duke@0 857 int tlab_end_offset = in_bytes(JavaThread::tlab_end_offset());
duke@0 858 eden_top_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_top_offset);
duke@0 859 eden_end_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_end_offset);
duke@0 860 } else { // Shared allocation: load from globals
duke@0 861 CollectedHeap* ch = Universe::heap();
duke@0 862 address top_adr = (address)ch->top_addr();
duke@0 863 address end_adr = (address)ch->end_addr();
duke@0 864 eden_top_adr = makecon(TypeRawPtr::make(top_adr));
duke@0 865 eden_end_adr = basic_plus_adr(eden_top_adr, end_adr - top_adr);
duke@0 866 }
duke@0 867 }
duke@0 868
duke@0 869
duke@0 870 Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) {
duke@0 871 Node* adr = basic_plus_adr(base, offset);
kvn@423 872 const TypePtr* adr_type = adr->bottom_type()->is_ptr();
coleenp@113 873 Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt);
duke@0 874 transform_later(value);
duke@0 875 return value;
duke@0 876 }
duke@0 877
duke@0 878
duke@0 879 Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) {
duke@0 880 Node* adr = basic_plus_adr(base, offset);
coleenp@113 881 mem = StoreNode::make(_igvn, ctl, mem, adr, NULL, value, bt);
duke@0 882 transform_later(mem);
duke@0 883 return mem;
duke@0 884 }
duke@0 885
duke@0 886 //=============================================================================
duke@0 887 //
duke@0 888 // A L L O C A T I O N
duke@0 889 //
duke@0 890 // Allocation attempts to be fast in the case of frequent small objects.
duke@0 891 // It breaks down like this:
duke@0 892 //
duke@0 893 // 1) Size in doublewords is computed. This is a constant for objects and
duke@0 894 // variable for most arrays. Doubleword units are used to avoid size
duke@0 895 // overflow of huge doubleword arrays. We need doublewords in the end for
duke@0 896 // rounding.
duke@0 897 //
duke@0 898 // 2) Size is checked for being 'too large'. Too-large allocations will go
duke@0 899 // the slow path into the VM. The slow path can throw any required
duke@0 900 // exceptions, and does all the special checks for very large arrays. The
duke@0 901 // size test can constant-fold away for objects. For objects with
duke@0 902 // finalizers it constant-folds the otherway: you always go slow with
duke@0 903 // finalizers.
duke@0 904 //
duke@0 905 // 3) If NOT using TLABs, this is the contended loop-back point.
duke@0 906 // Load-Locked the heap top. If using TLABs normal-load the heap top.
duke@0 907 //
duke@0 908 // 4) Check that heap top + size*8 < max. If we fail go the slow ` route.
duke@0 909 // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish
duke@0 910 // "size*8" we always enter the VM, where "largish" is a constant picked small
duke@0 911 // enough that there's always space between the eden max and 4Gig (old space is
duke@0 912 // there so it's quite large) and large enough that the cost of entering the VM
duke@0 913 // is dwarfed by the cost to initialize the space.
duke@0 914 //
duke@0 915 // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back
duke@0 916 // down. If contended, repeat at step 3. If using TLABs normal-store
duke@0 917 // adjusted heap top back down; there is no contention.
duke@0 918 //
duke@0 919 // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark
duke@0 920 // fields.
duke@0 921 //
duke@0 922 // 7) Merge with the slow-path; cast the raw memory pointer to the correct
duke@0 923 // oop flavor.
duke@0 924 //
duke@0 925 //=============================================================================
duke@0 926 // FastAllocateSizeLimit value is in DOUBLEWORDS.
duke@0 927 // Allocations bigger than this always go the slow route.
duke@0 928 // This value must be small enough that allocation attempts that need to
duke@0 929 // trigger exceptions go the slow route. Also, it must be small enough so
duke@0 930 // that heap_top + size_in_bytes does not wrap around the 4Gig limit.
duke@0 931 //=============================================================================j//
duke@0 932 // %%% Here is an old comment from parseHelper.cpp; is it outdated?
duke@0 933 // The allocator will coalesce int->oop copies away. See comment in
duke@0 934 // coalesce.cpp about how this works. It depends critically on the exact
duke@0 935 // code shape produced here, so if you are changing this code shape
duke@0 936 // make sure the GC info for the heap-top is correct in and around the
duke@0 937 // slow-path call.
duke@0 938 //
duke@0 939
duke@0 940 void PhaseMacroExpand::expand_allocate_common(
duke@0 941 AllocateNode* alloc, // allocation node to be expanded
duke@0 942 Node* length, // array length for an array allocation
duke@0 943 const TypeFunc* slow_call_type, // Type of slow call
duke@0 944 address slow_call_address // Address of slow call
duke@0 945 )
duke@0 946 {
duke@0 947
duke@0 948 Node* ctrl = alloc->in(TypeFunc::Control);
duke@0 949 Node* mem = alloc->in(TypeFunc::Memory);
duke@0 950 Node* i_o = alloc->in(TypeFunc::I_O);
duke@0 951 Node* size_in_bytes = alloc->in(AllocateNode::AllocSize);
duke@0 952 Node* klass_node = alloc->in(AllocateNode::KlassNode);
duke@0 953 Node* initial_slow_test = alloc->in(AllocateNode::InitialTest);
duke@0 954
kvn@73 955 // With escape analysis, the entire memory state was needed to be able to
kvn@73 956 // eliminate the allocation. Since the allocations cannot be eliminated,
kvn@73 957 // optimize it to the raw slice.
kvn@73 958 if (mem->is_MergeMem()) {
kvn@73 959 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
kvn@73 960 }
kvn@73 961
duke@0 962 assert(ctrl != NULL, "must have control");
duke@0 963 // We need a Region and corresponding Phi's to merge the slow-path and fast-path results.
duke@0 964 // they will not be used if "always_slow" is set
duke@0 965 enum { slow_result_path = 1, fast_result_path = 2 };
duke@0 966 Node *result_region;
duke@0 967 Node *result_phi_rawmem;
duke@0 968 Node *result_phi_rawoop;
duke@0 969 Node *result_phi_i_o;
duke@0 970
duke@0 971 // The initial slow comparison is a size check, the comparison
duke@0 972 // we want to do is a BoolTest::gt
duke@0 973 bool always_slow = false;
duke@0 974 int tv = _igvn.find_int_con(initial_slow_test, -1);
duke@0 975 if (tv >= 0) {
duke@0 976 always_slow = (tv == 1);
duke@0 977 initial_slow_test = NULL;
duke@0 978 } else {
duke@0 979 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn);
duke@0 980 }
duke@0 981
ysr@345 982 if (DTraceAllocProbes ||
ysr@345 983 !UseTLAB && (!Universe::heap()->supports_inline_contig_alloc() ||
ysr@345 984 (UseConcMarkSweepGC && CMSIncrementalMode))) {
duke@0 985 // Force slow-path allocation
duke@0 986 always_slow = true;
duke@0 987 initial_slow_test = NULL;
duke@0 988 }
duke@0 989
ysr@345 990
duke@0 991 enum { too_big_or_final_path = 1, need_gc_path = 2 };
duke@0 992 Node *slow_region = NULL;
duke@0 993 Node *toobig_false = ctrl;
duke@0 994
duke@0 995 assert (initial_slow_test == NULL || !always_slow, "arguments must be consistent");
duke@0 996 // generate the initial test if necessary
duke@0 997 if (initial_slow_test != NULL ) {
duke@0 998 slow_region = new (C, 3) RegionNode(3);
duke@0 999
duke@0 1000 // Now make the initial failure test. Usually a too-big test but
duke@0 1001 // might be a TRUE for finalizers or a fancy class check for
duke@0 1002 // newInstance0.
duke@0 1003 IfNode *toobig_iff = new (C, 2) IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
duke@0 1004 transform_later(toobig_iff);
duke@0 1005 // Plug the failing-too-big test into the slow-path region
duke@0 1006 Node *toobig_true = new (C, 1) IfTrueNode( toobig_iff );
duke@0 1007 transform_later(toobig_true);
duke@0 1008 slow_region ->init_req( too_big_or_final_path, toobig_true );
duke@0 1009 toobig_false = new (C, 1) IfFalseNode( toobig_iff );
duke@0 1010 transform_later(toobig_false);
duke@0 1011 } else { // No initial test, just fall into next case
duke@0 1012 toobig_false = ctrl;
duke@0 1013 debug_only(slow_region = NodeSentinel);
duke@0 1014 }
duke@0 1015
duke@0 1016 Node *slow_mem = mem; // save the current memory state for slow path
duke@0 1017 // generate the fast allocation code unless we know that the initial test will always go slow
duke@0 1018 if (!always_slow) {
ysr@345 1019 Node* eden_top_adr;
ysr@345 1020 Node* eden_end_adr;
ysr@345 1021
ysr@345 1022 set_eden_pointers(eden_top_adr, eden_end_adr);
ysr@345 1023
ysr@345 1024 // Load Eden::end. Loop invariant and hoisted.
ysr@345 1025 //
ysr@345 1026 // Note: We set the control input on "eden_end" and "old_eden_top" when using
ysr@345 1027 // a TLAB to work around a bug where these values were being moved across
ysr@345 1028 // a safepoint. These are not oops, so they cannot be include in the oop
ysr@345 1029 // map, but the can be changed by a GC. The proper way to fix this would
ysr@345 1030 // be to set the raw memory state when generating a SafepointNode. However
ysr@345 1031 // this will require extensive changes to the loop optimization in order to
ysr@345 1032 // prevent a degradation of the optimization.
ysr@345 1033 // See comment in memnode.hpp, around line 227 in class LoadPNode.
ysr@345 1034 Node *eden_end = make_load(ctrl, mem, eden_end_adr, 0, TypeRawPtr::BOTTOM, T_ADDRESS);
ysr@345 1035
duke@0 1036 // allocate the Region and Phi nodes for the result
duke@0 1037 result_region = new (C, 3) RegionNode(3);
duke@0 1038 result_phi_rawmem = new (C, 3) PhiNode( result_region, Type::MEMORY, TypeRawPtr::BOTTOM );
duke@0 1039 result_phi_rawoop = new (C, 3) PhiNode( result_region, TypeRawPtr::BOTTOM );
duke@0 1040 result_phi_i_o = new (C, 3) PhiNode( result_region, Type::ABIO ); // I/O is used for Prefetch
duke@0 1041
duke@0 1042 // We need a Region for the loop-back contended case.
duke@0 1043 enum { fall_in_path = 1, contended_loopback_path = 2 };
duke@0 1044 Node *contended_region;
duke@0 1045 Node *contended_phi_rawmem;
duke@0 1046 if( UseTLAB ) {
duke@0 1047 contended_region = toobig_false;
duke@0 1048 contended_phi_rawmem = mem;
duke@0 1049 } else {
duke@0 1050 contended_region = new (C, 3) RegionNode(3);
duke@0 1051 contended_phi_rawmem = new (C, 3) PhiNode( contended_region, Type::MEMORY, TypeRawPtr::BOTTOM);
duke@0 1052 // Now handle the passing-too-big test. We fall into the contended
duke@0 1053 // loop-back merge point.
duke@0 1054 contended_region ->init_req( fall_in_path, toobig_false );
duke@0 1055 contended_phi_rawmem->init_req( fall_in_path, mem );
duke@0 1056 transform_later(contended_region);
duke@0 1057 transform_later(contended_phi_rawmem);
duke@0 1058 }
duke@0 1059
duke@0 1060 // Load(-locked) the heap top.
duke@0 1061 // See note above concerning the control input when using a TLAB
duke@0 1062 Node *old_eden_top = UseTLAB
duke@0 1063 ? new (C, 3) LoadPNode ( ctrl, contended_phi_rawmem, eden_top_adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM )
duke@0 1064 : new (C, 3) LoadPLockedNode( contended_region, contended_phi_rawmem, eden_top_adr );
duke@0 1065
duke@0 1066 transform_later(old_eden_top);
duke@0 1067 // Add to heap top to get a new heap top
duke@0 1068 Node *new_eden_top = new (C, 4) AddPNode( top(), old_eden_top, size_in_bytes );
duke@0 1069 transform_later(new_eden_top);
duke@0 1070 // Check for needing a GC; compare against heap end
duke@0 1071 Node *needgc_cmp = new (C, 3) CmpPNode( new_eden_top, eden_end );
duke@0 1072 transform_later(needgc_cmp);
duke@0 1073 Node *needgc_bol = new (C, 2) BoolNode( needgc_cmp, BoolTest::ge );
duke@0 1074 transform_later(needgc_bol);
duke@0 1075 IfNode *needgc_iff = new (C, 2) IfNode(contended_region, needgc_bol, PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN );
duke@0 1076 transform_later(needgc_iff);
duke@0 1077
duke@0 1078 // Plug the failing-heap-space-need-gc test into the slow-path region
duke@0 1079 Node *needgc_true = new (C, 1) IfTrueNode( needgc_iff );
duke@0 1080 transform_later(needgc_true);
duke@0 1081 if( initial_slow_test ) {
duke@0 1082 slow_region ->init_req( need_gc_path, needgc_true );
duke@0 1083 // This completes all paths into the slow merge point
duke@0 1084 transform_later(slow_region);
duke@0 1085 } else { // No initial slow path needed!
duke@0 1086 // Just fall from the need-GC path straight into the VM call.
duke@0 1087 slow_region = needgc_true;
duke@0 1088 }
duke@0 1089 // No need for a GC. Setup for the Store-Conditional
duke@0 1090 Node *needgc_false = new (C, 1) IfFalseNode( needgc_iff );
duke@0 1091 transform_later(needgc_false);
duke@0 1092
duke@0 1093 // Grab regular I/O before optional prefetch may change it.
duke@0 1094 // Slow-path does no I/O so just set it to the original I/O.
duke@0 1095 result_phi_i_o->init_req( slow_result_path, i_o );
duke@0 1096
duke@0 1097 i_o = prefetch_allocation(i_o, needgc_false, contended_phi_rawmem,
duke@0 1098 old_eden_top, new_eden_top, length);
duke@0 1099
duke@0 1100 // Store (-conditional) the modified eden top back down.
duke@0 1101 // StorePConditional produces flags for a test PLUS a modified raw
duke@0 1102 // memory state.
duke@0 1103 Node *store_eden_top;
duke@0 1104 Node *fast_oop_ctrl;
duke@0 1105 if( UseTLAB ) {
duke@0 1106 store_eden_top = new (C, 4) StorePNode( needgc_false, contended_phi_rawmem, eden_top_adr, TypeRawPtr::BOTTOM, new_eden_top );
duke@0 1107 transform_later(store_eden_top);
duke@0 1108 fast_oop_ctrl = needgc_false; // No contention, so this is the fast path
duke@0 1109 } else {
duke@0 1110 store_eden_top = new (C, 5) StorePConditionalNode( needgc_false, contended_phi_rawmem, eden_top_adr, new_eden_top, old_eden_top );
duke@0 1111 transform_later(store_eden_top);
duke@0 1112 Node *contention_check = new (C, 2) BoolNode( store_eden_top, BoolTest::ne );
duke@0 1113 transform_later(contention_check);
duke@0 1114 store_eden_top = new (C, 1) SCMemProjNode(store_eden_top);
duke@0 1115 transform_later(store_eden_top);
duke@0 1116
duke@0 1117 // If not using TLABs, check to see if there was contention.
duke@0 1118 IfNode *contention_iff = new (C, 2) IfNode ( needgc_false, contention_check, PROB_MIN, COUNT_UNKNOWN );
duke@0 1119 transform_later(contention_iff);
duke@0 1120 Node *contention_true = new (C, 1) IfTrueNode( contention_iff );
duke@0 1121 transform_later(contention_true);
duke@0 1122 // If contention, loopback and try again.
duke@0 1123 contended_region->init_req( contended_loopback_path, contention_true );
duke@0 1124 contended_phi_rawmem->init_req( contended_loopback_path, store_eden_top );
duke@0 1125
duke@0 1126 // Fast-path succeeded with no contention!
duke@0 1127 Node *contention_false = new (C, 1) IfFalseNode( contention_iff );
duke@0 1128 transform_later(contention_false);
duke@0 1129 fast_oop_ctrl = contention_false;
duke@0 1130 }
duke@0 1131
duke@0 1132 // Rename successful fast-path variables to make meaning more obvious
duke@0 1133 Node* fast_oop = old_eden_top;
duke@0 1134 Node* fast_oop_rawmem = store_eden_top;
duke@0 1135 fast_oop_rawmem = initialize_object(alloc,
duke@0 1136 fast_oop_ctrl, fast_oop_rawmem, fast_oop,
duke@0 1137 klass_node, length, size_in_bytes);
duke@0 1138
duke@0 1139 if (ExtendedDTraceProbes) {
duke@0 1140 // Slow-path call
duke@0 1141 int size = TypeFunc::Parms + 2;
duke@0 1142 CallLeafNode *call = new (C, size) CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(),
duke@0 1143 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_object_alloc_base),
duke@0 1144 "dtrace_object_alloc",
duke@0 1145 TypeRawPtr::BOTTOM);
duke@0 1146
duke@0 1147 // Get base of thread-local storage area
duke@0 1148 Node* thread = new (C, 1) ThreadLocalNode();
duke@0 1149 transform_later(thread);
duke@0 1150
duke@0 1151 call->init_req(TypeFunc::Parms+0, thread);
duke@0 1152 call->init_req(TypeFunc::Parms+1, fast_oop);
duke@0 1153 call->init_req( TypeFunc::Control, fast_oop_ctrl );
duke@0 1154 call->init_req( TypeFunc::I_O , top() ) ; // does no i/o
duke@0 1155 call->init_req( TypeFunc::Memory , fast_oop_rawmem );
duke@0 1156 call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) );
duke@0 1157 call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) );
duke@0 1158 transform_later(call);
duke@0 1159 fast_oop_ctrl = new (C, 1) ProjNode(call,TypeFunc::Control);
duke@0 1160 transform_later(fast_oop_ctrl);
duke@0 1161 fast_oop_rawmem = new (C, 1) ProjNode(call,TypeFunc::Memory);
duke@0 1162 transform_later(fast_oop_rawmem);
duke@0 1163 }
duke@0 1164
duke@0 1165 // Plug in the successful fast-path into the result merge point
duke@0 1166 result_region ->init_req( fast_result_path, fast_oop_ctrl );
duke@0 1167 result_phi_rawoop->init_req( fast_result_path, fast_oop );
duke@0 1168 result_phi_i_o ->init_req( fast_result_path, i_o );
duke@0 1169 result_phi_rawmem->init_req( fast_result_path, fast_oop_rawmem );
duke@0 1170 } else {
duke@0 1171 slow_region = ctrl;
duke@0 1172 }
duke@0 1173
duke@0 1174 // Generate slow-path call
duke@0 1175 CallNode *call = new (C, slow_call_type->domain()->cnt())
duke@0 1176 CallStaticJavaNode(slow_call_type, slow_call_address,
duke@0 1177 OptoRuntime::stub_name(slow_call_address),
duke@0 1178 alloc->jvms()->bci(),
duke@0 1179 TypePtr::BOTTOM);
duke@0 1180 call->init_req( TypeFunc::Control, slow_region );
duke@0 1181 call->init_req( TypeFunc::I_O , top() ) ; // does no i/o
duke@0 1182 call->init_req( TypeFunc::Memory , slow_mem ); // may gc ptrs
duke@0 1183 call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) );
duke@0 1184 call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) );
duke@0 1185
duke@0 1186 call->init_req(TypeFunc::Parms+0, klass_node);
duke@0 1187 if (length != NULL) {
duke@0 1188 call->init_req(TypeFunc::Parms+1, length);
duke@0 1189 }
duke@0 1190
duke@0 1191 // Copy debug information and adjust JVMState information, then replace
duke@0 1192 // allocate node with the call
duke@0 1193 copy_call_debug_info((CallNode *) alloc, call);
duke@0 1194 if (!always_slow) {
duke@0 1195 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
duke@0 1196 }
duke@0 1197 _igvn.hash_delete(alloc);
duke@0 1198 _igvn.subsume_node(alloc, call);
duke@0 1199 transform_later(call);
duke@0 1200
duke@0 1201 // Identify the output projections from the allocate node and
duke@0 1202 // adjust any references to them.
duke@0 1203 // The control and io projections look like:
duke@0 1204 //
duke@0 1205 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl)
duke@0 1206 // Allocate Catch
duke@0 1207 // ^---Proj(io) <-------+ ^---CatchProj(io)
duke@0 1208 //
duke@0 1209 // We are interested in the CatchProj nodes.
duke@0 1210 //
duke@0 1211 extract_call_projections(call);
duke@0 1212
duke@0 1213 // An allocate node has separate memory projections for the uses on the control and i_o paths
duke@0 1214 // Replace uses of the control memory projection with result_phi_rawmem (unless we are only generating a slow call)
duke@0 1215 if (!always_slow && _memproj_fallthrough != NULL) {
duke@0 1216 for (DUIterator_Fast imax, i = _memproj_fallthrough->fast_outs(imax); i < imax; i++) {
duke@0 1217 Node *use = _memproj_fallthrough->fast_out(i);
duke@0 1218 _igvn.hash_delete(use);
duke@0 1219 imax -= replace_input(use, _memproj_fallthrough, result_phi_rawmem);
duke@0 1220 _igvn._worklist.push(use);
duke@0 1221 // back up iterator
duke@0 1222 --i;
duke@0 1223 }
duke@0 1224 }
duke@0 1225 // Now change uses of _memproj_catchall to use _memproj_fallthrough and delete _memproj_catchall so
duke@0 1226 // we end up with a call that has only 1 memory projection
duke@0 1227 if (_memproj_catchall != NULL ) {
duke@0 1228 if (_memproj_fallthrough == NULL) {
duke@0 1229 _memproj_fallthrough = new (C, 1) ProjNode(call, TypeFunc::Memory);
duke@0 1230 transform_later(_memproj_fallthrough);
duke@0 1231 }
duke@0 1232 for (DUIterator_Fast imax, i = _memproj_catchall->fast_outs(imax); i < imax; i++) {
duke@0 1233 Node *use = _memproj_catchall->fast_out(i);
duke@0 1234 _igvn.hash_delete(use);
duke@0 1235 imax -= replace_input(use, _memproj_catchall, _memproj_fallthrough);
duke@0 1236 _igvn._worklist.push(use);
duke@0 1237 // back up iterator
duke@0 1238 --i;
duke@0 1239 }
duke@0 1240 }
duke@0 1241
duke@0 1242 mem = result_phi_rawmem;
duke@0 1243
duke@0 1244 // An allocate node has separate i_o projections for the uses on the control and i_o paths
duke@0 1245 // Replace uses of the control i_o projection with result_phi_i_o (unless we are only generating a slow call)
duke@0 1246 if (_ioproj_fallthrough == NULL) {
duke@0 1247 _ioproj_fallthrough = new (C, 1) ProjNode(call, TypeFunc::I_O);
duke@0 1248 transform_later(_ioproj_fallthrough);
duke@0 1249 } else if (!always_slow) {
duke@0 1250 for (DUIterator_Fast imax, i = _ioproj_fallthrough->fast_outs(imax); i < imax; i++) {
duke@0 1251 Node *use = _ioproj_fallthrough->fast_out(i);
duke@0 1252
duke@0 1253 _igvn.hash_delete(use);
duke@0 1254 imax -= replace_input(use, _ioproj_fallthrough, result_phi_i_o);
duke@0 1255 _igvn._worklist.push(use);
duke@0 1256 // back up iterator
duke@0 1257 --i;
duke@0 1258 }
duke@0 1259 }
duke@0 1260 // Now change uses of _ioproj_catchall to use _ioproj_fallthrough and delete _ioproj_catchall so
duke@0 1261 // we end up with a call that has only 1 control projection
duke@0 1262 if (_ioproj_catchall != NULL ) {
duke@0 1263 for (DUIterator_Fast imax, i = _ioproj_catchall->fast_outs(imax); i < imax; i++) {
duke@0 1264 Node *use = _ioproj_catchall->fast_out(i);
duke@0 1265 _igvn.hash_delete(use);
duke@0 1266 imax -= replace_input(use, _ioproj_catchall, _ioproj_fallthrough);
duke@0 1267 _igvn._worklist.push(use);
duke@0 1268 // back up iterator
duke@0 1269 --i;
duke@0 1270 }
duke@0 1271 }
duke@0 1272
duke@0 1273 // if we generated only a slow call, we are done
duke@0 1274 if (always_slow)
duke@0 1275 return;
duke@0 1276
duke@0 1277
duke@0 1278 if (_fallthroughcatchproj != NULL) {
duke@0 1279 ctrl = _fallthroughcatchproj->clone();
duke@0 1280 transform_later(ctrl);
duke@0 1281 _igvn.hash_delete(_fallthroughcatchproj);
duke@0 1282 _igvn.subsume_node(_fallthroughcatchproj, result_region);
duke@0 1283 } else {
duke@0 1284 ctrl = top();
duke@0 1285 }
duke@0 1286 Node *slow_result;
duke@0 1287 if (_resproj == NULL) {
duke@0 1288 // no uses of the allocation result
duke@0 1289 slow_result = top();
duke@0 1290 } else {
duke@0 1291 slow_result = _resproj->clone();
duke@0 1292 transform_later(slow_result);
duke@0 1293 _igvn.hash_delete(_resproj);
duke@0 1294 _igvn.subsume_node(_resproj, result_phi_rawoop);
duke@0 1295 }
duke@0 1296
duke@0 1297 // Plug slow-path into result merge point
duke@0 1298 result_region ->init_req( slow_result_path, ctrl );
duke@0 1299 result_phi_rawoop->init_req( slow_result_path, slow_result);
duke@0 1300 result_phi_rawmem->init_req( slow_result_path, _memproj_fallthrough );
duke@0 1301 transform_later(result_region);
duke@0 1302 transform_later(result_phi_rawoop);
duke@0 1303 transform_later(result_phi_rawmem);
duke@0 1304 transform_later(result_phi_i_o);
duke@0 1305 // This completes all paths into the result merge point
duke@0 1306 }
duke@0 1307
duke@0 1308
duke@0 1309 // Helper for PhaseMacroExpand::expand_allocate_common.
duke@0 1310 // Initializes the newly-allocated storage.
duke@0 1311 Node*
duke@0 1312 PhaseMacroExpand::initialize_object(AllocateNode* alloc,
duke@0 1313 Node* control, Node* rawmem, Node* object,
duke@0 1314 Node* klass_node, Node* length,
duke@0 1315 Node* size_in_bytes) {
duke@0 1316 InitializeNode* init = alloc->initialization();
duke@0 1317 // Store the klass & mark bits
duke@0 1318 Node* mark_node = NULL;
duke@0 1319 // For now only enable fast locking for non-array types
duke@0 1320 if (UseBiasedLocking && (length == NULL)) {
duke@0 1321 mark_node = make_load(NULL, rawmem, klass_node, Klass::prototype_header_offset_in_bytes() + sizeof(oopDesc), TypeRawPtr::BOTTOM, T_ADDRESS);
duke@0 1322 } else {
duke@0 1323 mark_node = makecon(TypeRawPtr::make((address)markOopDesc::prototype()));
duke@0 1324 }
duke@0 1325 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS);
coleenp@113 1326
duke@0 1327 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_OBJECT);
duke@0 1328 int header_size = alloc->minimum_header_size(); // conservatively small
duke@0 1329
duke@0 1330 // Array length
duke@0 1331 if (length != NULL) { // Arrays need length field
duke@0 1332 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
duke@0 1333 // conservatively small header size:
coleenp@113 1334 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
duke@0 1335 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
duke@0 1336 if (k->is_array_klass()) // we know the exact header size in most cases:
duke@0 1337 header_size = Klass::layout_helper_header_size(k->layout_helper());
duke@0 1338 }
duke@0 1339
duke@0 1340 // Clear the object body, if necessary.
duke@0 1341 if (init == NULL) {
duke@0 1342 // The init has somehow disappeared; be cautious and clear everything.
duke@0 1343 //
duke@0 1344 // This can happen if a node is allocated but an uncommon trap occurs
duke@0 1345 // immediately. In this case, the Initialize gets associated with the
duke@0 1346 // trap, and may be placed in a different (outer) loop, if the Allocate
duke@0 1347 // is in a loop. If (this is rare) the inner loop gets unrolled, then
duke@0 1348 // there can be two Allocates to one Initialize. The answer in all these
duke@0 1349 // edge cases is safety first. It is always safe to clear immediately
duke@0 1350 // within an Allocate, and then (maybe or maybe not) clear some more later.
duke@0 1351 if (!ZeroTLAB)
duke@0 1352 rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
duke@0 1353 header_size, size_in_bytes,
duke@0 1354 &_igvn);
duke@0 1355 } else {
duke@0 1356 if (!init->is_complete()) {
duke@0 1357 // Try to win by zeroing only what the init does not store.
duke@0 1358 // We can also try to do some peephole optimizations,
duke@0 1359 // such as combining some adjacent subword stores.
duke@0 1360 rawmem = init->complete_stores(control, rawmem, object,
duke@0 1361 header_size, size_in_bytes, &_igvn);
duke@0 1362 }
duke@0 1363 // We have no more use for this link, since the AllocateNode goes away:
duke@0 1364 init->set_req(InitializeNode::RawAddress, top());
duke@0 1365 // (If we keep the link, it just confuses the register allocator,
duke@0 1366 // who thinks he sees a real use of the address by the membar.)
duke@0 1367 }
duke@0 1368
duke@0 1369 return rawmem;
duke@0 1370 }
duke@0 1371
duke@0 1372 // Generate prefetch instructions for next allocations.
duke@0 1373 Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false,
duke@0 1374 Node*& contended_phi_rawmem,
duke@0 1375 Node* old_eden_top, Node* new_eden_top,
duke@0 1376 Node* length) {
duke@0 1377 if( UseTLAB && AllocatePrefetchStyle == 2 ) {
duke@0 1378 // Generate prefetch allocation with watermark check.
duke@0 1379 // As an allocation hits the watermark, we will prefetch starting
duke@0 1380 // at a "distance" away from watermark.
duke@0 1381 enum { fall_in_path = 1, pf_path = 2 };
duke@0 1382
duke@0 1383 Node *pf_region = new (C, 3) RegionNode(3);
duke@0 1384 Node *pf_phi_rawmem = new (C, 3) PhiNode( pf_region, Type::MEMORY,
duke@0 1385 TypeRawPtr::BOTTOM );
duke@0 1386 // I/O is used for Prefetch
duke@0 1387 Node *pf_phi_abio = new (C, 3) PhiNode( pf_region, Type::ABIO );
duke@0 1388
duke@0 1389 Node *thread = new (C, 1) ThreadLocalNode();
duke@0 1390 transform_later(thread);
duke@0 1391
duke@0 1392 Node *eden_pf_adr = new (C, 4) AddPNode( top()/*not oop*/, thread,
duke@0 1393 _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) );
duke@0 1394 transform_later(eden_pf_adr);
duke@0 1395
duke@0 1396 Node *old_pf_wm = new (C, 3) LoadPNode( needgc_false,
duke@0 1397 contended_phi_rawmem, eden_pf_adr,
duke@0 1398 TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM );
duke@0 1399 transform_later(old_pf_wm);
duke@0 1400
duke@0 1401 // check against new_eden_top
duke@0 1402 Node *need_pf_cmp = new (C, 3) CmpPNode( new_eden_top, old_pf_wm );
duke@0 1403 transform_later(need_pf_cmp);
duke@0 1404 Node *need_pf_bol = new (C, 2) BoolNode( need_pf_cmp, BoolTest::ge );
duke@0 1405 transform_later(need_pf_bol);
duke@0 1406 IfNode *need_pf_iff = new (C, 2) IfNode( needgc_false, need_pf_bol,
duke@0 1407 PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN );
duke@0 1408 transform_later(need_pf_iff);
duke@0 1409
duke@0 1410 // true node, add prefetchdistance
duke@0 1411 Node *need_pf_true = new (C, 1) IfTrueNode( need_pf_iff );
duke@0 1412 transform_later(need_pf_true);
duke@0 1413
duke@0 1414 Node *need_pf_false = new (C, 1) IfFalseNode( need_pf_iff );
duke@0 1415 transform_later(need_pf_false);
duke@0 1416
duke@0 1417 Node *new_pf_wmt = new (C, 4) AddPNode( top(), old_pf_wm,
duke@0 1418 _igvn.MakeConX(AllocatePrefetchDistance) );
duke@0 1419 transform_later(new_pf_wmt );
duke@0 1420 new_pf_wmt->set_req(0, need_pf_true);
duke@0 1421
duke@0 1422 Node *store_new_wmt = new (C, 4) StorePNode( need_pf_true,
duke@0 1423 contended_phi_rawmem, eden_pf_adr,
duke@0 1424 TypeRawPtr::BOTTOM, new_pf_wmt );
duke@0 1425 transform_later(store_new_wmt);
duke@0 1426
duke@0 1427 // adding prefetches
duke@0 1428 pf_phi_abio->init_req( fall_in_path, i_o );
duke@0 1429
duke@0 1430 Node *prefetch_adr;
duke@0 1431 Node *prefetch;
duke@0 1432 uint lines = AllocatePrefetchDistance / AllocatePrefetchStepSize;
duke@0 1433 uint step_size = AllocatePrefetchStepSize;
duke@0 1434 uint distance = 0;
duke@0 1435
duke@0 1436 for ( uint i = 0; i < lines; i++ ) {
duke@0 1437 prefetch_adr = new (C, 4) AddPNode( old_pf_wm, new_pf_wmt,
duke@0 1438 _igvn.MakeConX(distance) );
duke@0 1439 transform_later(prefetch_adr);
duke@0 1440 prefetch = new (C, 3) PrefetchWriteNode( i_o, prefetch_adr );
duke@0 1441 transform_later(prefetch);
duke@0 1442 distance += step_size;
duke@0 1443 i_o = prefetch;
duke@0 1444 }
duke@0 1445 pf_phi_abio->set_req( pf_path, i_o );
duke@0 1446
duke@0 1447 pf_region->init_req( fall_in_path, need_pf_false );
duke@0 1448 pf_region->init_req( pf_path, need_pf_true );
duke@0 1449
duke@0 1450 pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem );
duke@0 1451 pf_phi_rawmem->init_req( pf_path, store_new_wmt );
duke@0 1452
duke@0 1453 transform_later(pf_region);
duke@0 1454 transform_later(pf_phi_rawmem);
duke@0 1455 transform_later(pf_phi_abio);
duke@0 1456
duke@0 1457 needgc_false = pf_region;
duke@0 1458 contended_phi_rawmem = pf_phi_rawmem;
duke@0 1459 i_o = pf_phi_abio;
duke@0 1460 } else if( AllocatePrefetchStyle > 0 ) {
duke@0 1461 // Insert a prefetch for each allocation only on the fast-path
duke@0 1462 Node *prefetch_adr;
duke@0 1463 Node *prefetch;
duke@0 1464 // Generate several prefetch instructions only for arrays.
duke@0 1465 uint lines = (length != NULL) ? AllocatePrefetchLines : 1;
duke@0 1466 uint step_size = AllocatePrefetchStepSize;
duke@0 1467 uint distance = AllocatePrefetchDistance;
duke@0 1468 for ( uint i = 0; i < lines; i++ ) {
duke@0 1469 prefetch_adr = new (C, 4) AddPNode( old_eden_top, new_eden_top,
duke@0 1470 _igvn.MakeConX(distance) );
duke@0 1471 transform_later(prefetch_adr);
duke@0 1472 prefetch = new (C, 3) PrefetchWriteNode( i_o, prefetch_adr );
duke@0 1473 // Do not let it float too high, since if eden_top == eden_end,
duke@0 1474 // both might be null.
duke@0 1475 if( i == 0 ) { // Set control for first prefetch, next follows it
duke@0 1476 prefetch->init_req(0, needgc_false);
duke@0 1477 }
duke@0 1478 transform_later(prefetch);
duke@0 1479 distance += step_size;
duke@0 1480 i_o = prefetch;
duke@0 1481 }
duke@0 1482 }
duke@0 1483 return i_o;
duke@0 1484 }
duke@0 1485
duke@0 1486
duke@0 1487 void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) {
duke@0 1488 expand_allocate_common(alloc, NULL,
duke@0 1489 OptoRuntime::new_instance_Type(),
duke@0 1490 OptoRuntime::new_instance_Java());
duke@0 1491 }
duke@0 1492
duke@0 1493 void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) {
duke@0 1494 Node* length = alloc->in(AllocateNode::ALength);
duke@0 1495 expand_allocate_common(alloc, length,
duke@0 1496 OptoRuntime::new_array_Type(),
duke@0 1497 OptoRuntime::new_array_Java());
duke@0 1498 }
duke@0 1499
duke@0 1500
duke@0 1501 // we have determined that this lock/unlock can be eliminated, we simply
duke@0 1502 // eliminate the node without expanding it.
duke@0 1503 //
duke@0 1504 // Note: The membar's associated with the lock/unlock are currently not
duke@0 1505 // eliminated. This should be investigated as a future enhancement.
duke@0 1506 //
kvn@66 1507 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) {
kvn@66 1508
kvn@66 1509 if (!alock->is_eliminated()) {
kvn@66 1510 return false;
kvn@66 1511 }
kvn@66 1512 // Mark the box lock as eliminated if all correspondent locks are eliminated
kvn@66 1513 // to construct correct debug info.
kvn@66 1514 BoxLockNode* box = alock->box_node()->as_BoxLock();
kvn@66 1515 if (!box->is_eliminated()) {
kvn@66 1516 bool eliminate = true;
kvn@66 1517 for (DUIterator_Fast imax, i = box->fast_outs(imax); i < imax; i++) {
kvn@66 1518 Node *lck = box->fast_out(i);
kvn@66 1519 if (lck->is_Lock() && !lck->as_AbstractLock()->is_eliminated()) {
kvn@66 1520 eliminate = false;
kvn@66 1521 break;
kvn@66 1522 }
kvn@66 1523 }
kvn@66 1524 if (eliminate)
kvn@66 1525 box->set_eliminated();
kvn@66 1526 }
kvn@66 1527
kvn@66 1528 #ifndef PRODUCT
kvn@66 1529 if (PrintEliminateLocks) {
kvn@66 1530 if (alock->is_Lock()) {
kvn@66 1531 tty->print_cr("++++ Eliminating: %d Lock", alock->_idx);
kvn@66 1532 } else {
kvn@66 1533 tty->print_cr("++++ Eliminating: %d Unlock", alock->_idx);
kvn@66 1534 }
kvn@66 1535 }
kvn@66 1536 #endif
kvn@66 1537
kvn@66 1538 Node* mem = alock->in(TypeFunc::Memory);
kvn@66 1539 Node* ctrl = alock->in(TypeFunc::Control);
kvn@66 1540
kvn@66 1541 extract_call_projections(alock);
kvn@66 1542 // There are 2 projections from the lock. The lock node will
kvn@66 1543 // be deleted when its last use is subsumed below.
kvn@66 1544 assert(alock->outcnt() == 2 &&
kvn@66 1545 _fallthroughproj != NULL &&
kvn@66 1546 _memproj_fallthrough != NULL,
kvn@66 1547 "Unexpected projections from Lock/Unlock");
kvn@66 1548
kvn@66 1549 Node* fallthroughproj = _fallthroughproj;
kvn@66 1550 Node* memproj_fallthrough = _memproj_fallthrough;
duke@0 1551
duke@0 1552 // The memory projection from a lock/unlock is RawMem
duke@0 1553 // The input to a Lock is merged memory, so extract its RawMem input
duke@0 1554 // (unless the MergeMem has been optimized away.)
duke@0 1555 if (alock->is_Lock()) {
kvn@66 1556 // Seach for MemBarAcquire node and delete it also.
kvn@66 1557 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
kvn@66 1558 assert(membar != NULL && membar->Opcode() == Op_MemBarAcquire, "");
kvn@66 1559 Node* ctrlproj = membar->proj_out(TypeFunc::Control);
kvn@66 1560 Node* memproj = membar->proj_out(TypeFunc::Memory);
kvn@66 1561 _igvn.hash_delete(ctrlproj);
kvn@66 1562 _igvn.subsume_node(ctrlproj, fallthroughproj);
kvn@66 1563 _igvn.hash_delete(memproj);
kvn@66 1564 _igvn.subsume_node(memproj, memproj_fallthrough);
duke@0 1565 }
duke@0 1566
kvn@66 1567 // Seach for MemBarRelease node and delete it also.
kvn@66 1568 if (alock->is_Unlock() && ctrl != NULL && ctrl->is_Proj() &&
kvn@66 1569 ctrl->in(0)->is_MemBar()) {
kvn@66 1570 MemBarNode* membar = ctrl->in(0)->as_MemBar();
kvn@66 1571 assert(membar->Opcode() == Op_MemBarRelease &&
kvn@66 1572 mem->is_Proj() && membar == mem->in(0), "");
kvn@66 1573 _igvn.hash_delete(fallthroughproj);
kvn@66 1574 _igvn.subsume_node(fallthroughproj, ctrl);
kvn@66 1575 _igvn.hash_delete(memproj_fallthrough);
kvn@66 1576 _igvn.subsume_node(memproj_fallthrough, mem);
kvn@66 1577 fallthroughproj = ctrl;
kvn@66 1578 memproj_fallthrough = mem;
kvn@66 1579 ctrl = membar->in(TypeFunc::Control);
kvn@66 1580 mem = membar->in(TypeFunc::Memory);
kvn@66 1581 }
kvn@66 1582
kvn@66 1583 _igvn.hash_delete(fallthroughproj);
kvn@66 1584 _igvn.subsume_node(fallthroughproj, ctrl);
kvn@66 1585 _igvn.hash_delete(memproj_fallthrough);
kvn@66 1586 _igvn.subsume_node(memproj_fallthrough, mem);
kvn@66 1587 return true;
duke@0 1588 }
duke@0 1589
duke@0 1590
duke@0 1591 //------------------------------expand_lock_node----------------------
duke@0 1592 void PhaseMacroExpand::expand_lock_node(LockNode *lock) {
duke@0 1593
duke@0 1594 Node* ctrl = lock->in(TypeFunc::Control);
duke@0 1595 Node* mem = lock->in(TypeFunc::Memory);
duke@0 1596 Node* obj = lock->obj_node();
duke@0 1597 Node* box = lock->box_node();
kvn@66 1598 Node* flock = lock->fastlock_node();
duke@0 1599
duke@0 1600 // Make the merge point
kvn@423 1601 Node *region;
kvn@423 1602 Node *mem_phi;
kvn@423 1603 Node *slow_path;
duke@0 1604
kvn@423 1605 if (UseOptoBiasInlining) {
kvn@423 1606 /*
kvn@423 1607 * See the full descrition in MacroAssembler::biased_locking_enter().
kvn@423 1608 *
kvn@423 1609 * if( (mark_word & biased_lock_mask) == biased_lock_pattern ) {
kvn@423 1610 * // The object is biased.
kvn@423 1611 * proto_node = klass->prototype_header;
kvn@423 1612 * o_node = thread | proto_node;
kvn@423 1613 * x_node = o_node ^ mark_word;
kvn@423 1614 * if( (x_node & ~age_mask) == 0 ) { // Biased to the current thread ?
kvn@423 1615 * // Done.
kvn@423 1616 * } else {
kvn@423 1617 * if( (x_node & biased_lock_mask) != 0 ) {
kvn@423 1618 * // The klass's prototype header is no longer biased.
kvn@423 1619 * cas(&mark_word, mark_word, proto_node)
kvn@423 1620 * goto cas_lock;
kvn@423 1621 * } else {
kvn@423 1622 * // The klass's prototype header is still biased.
kvn@423 1623 * if( (x_node & epoch_mask) != 0 ) { // Expired epoch?
kvn@423 1624 * old = mark_word;
kvn@423 1625 * new = o_node;
kvn@423 1626 * } else {
kvn@423 1627 * // Different thread or anonymous biased.
kvn@423 1628 * old = mark_word & (epoch_mask | age_mask | biased_lock_mask);
kvn@423 1629 * new = thread | old;
kvn@423 1630 * }
kvn@423 1631 * // Try to rebias.
kvn@423 1632 * if( cas(&mark_word, old, new) == 0 ) {
kvn@423 1633 * // Done.
kvn@423 1634 * } else {
kvn@423 1635 * goto slow_path; // Failed.
kvn@423 1636 * }
kvn@423 1637 * }
kvn@423 1638 * }
kvn@423 1639 * } else {
kvn@423 1640 * // The object is not biased.
kvn@423 1641 * cas_lock:
kvn@423 1642 * if( FastLock(obj) == 0 ) {
kvn@423 1643 * // Done.
kvn@423 1644 * } else {
kvn@423 1645 * slow_path:
kvn@423 1646 * OptoRuntime::complete_monitor_locking_Java(obj);
kvn@423 1647 * }
kvn@423 1648 * }
kvn@423 1649 */
kvn@423 1650
kvn@423 1651 region = new (C, 5) RegionNode(5);
kvn@423 1652 // create a Phi for the memory state
kvn@423 1653 mem_phi = new (C, 5) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
kvn@423 1654
kvn@423 1655 Node* fast_lock_region = new (C, 3) RegionNode(3);
kvn@423 1656 Node* fast_lock_mem_phi = new (C, 3) PhiNode( fast_lock_region, Type::MEMORY, TypeRawPtr::BOTTOM);
kvn@423 1657
kvn@423 1658 // First, check mark word for the biased lock pattern.
kvn@423 1659 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
kvn@423 1660
kvn@423 1661 // Get fast path - mark word has the biased lock pattern.
kvn@423 1662 ctrl = opt_bits_test(ctrl, fast_lock_region, 1, mark_node,
kvn@423 1663 markOopDesc::biased_lock_mask_in_place,
kvn@423 1664 markOopDesc::biased_lock_pattern, true);
kvn@423 1665 // fast_lock_region->in(1) is set to slow path.
kvn@423 1666 fast_lock_mem_phi->init_req(1, mem);
kvn@423 1667
kvn@423 1668 // Now check that the lock is biased to the current thread and has
kvn@423 1669 // the same epoch and bias as Klass::_prototype_header.
kvn@423 1670
kvn@423 1671 // Special-case a fresh allocation to avoid building nodes:
kvn@423 1672 Node* klass_node = AllocateNode::Ideal_klass(obj, &_igvn);
kvn@423 1673 if (klass_node == NULL) {
kvn@423 1674 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
kvn@423 1675 klass_node = transform_later( LoadKlassNode::make(_igvn, mem, k_adr, _igvn.type(k_adr)->is_ptr()) );
kvn@423 1676 klass_node->init_req(0, ctrl);
kvn@423 1677 }
kvn@423 1678 Node *proto_node = make_load(ctrl, mem, klass_node, Klass::prototype_header_offset_in_bytes() + sizeof(oopDesc), TypeX_X, TypeX_X->basic_type());
kvn@423 1679
kvn@423 1680 Node* thread = transform_later(new (C, 1) ThreadLocalNode());
kvn@423 1681 Node* cast_thread = transform_later(new (C, 2) CastP2XNode(ctrl, thread));
kvn@423 1682 Node* o_node = transform_later(new (C, 3) OrXNode(cast_thread, proto_node));
kvn@423 1683 Node* x_node = transform_later(new (C, 3) XorXNode(o_node, mark_node));
kvn@423 1684
kvn@423 1685 // Get slow path - mark word does NOT match the value.
kvn@423 1686 Node* not_biased_ctrl = opt_bits_test(ctrl, region, 3, x_node,
kvn@423 1687 (~markOopDesc::age_mask_in_place), 0);
kvn@423 1688 // region->in(3) is set to fast path - the object is biased to the current thread.
kvn@423 1689 mem_phi->init_req(3, mem);
kvn@423 1690
kvn@423 1691
kvn@423 1692 // Mark word does NOT match the value (thread | Klass::_prototype_header).
kvn@423 1693
kvn@423 1694
kvn@423 1695 // First, check biased pattern.
kvn@423 1696 // Get fast path - _prototype_header has the same biased lock pattern.
kvn@423 1697 ctrl = opt_bits_test(not_biased_ctrl, fast_lock_region, 2, x_node,
kvn@423 1698 markOopDesc::biased_lock_mask_in_place, 0, true);
kvn@423 1699
kvn@423 1700 not_biased_ctrl = fast_lock_region->in(2); // Slow path
kvn@423 1701 // fast_lock_region->in(2) - the prototype header is no longer biased
kvn@423 1702 // and we have to revoke the bias on this object.
kvn@423 1703 // We are going to try to reset the mark of this object to the prototype
kvn@423 1704 // value and fall through to the CAS-based locking scheme.
kvn@423 1705 Node* adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
kvn@423 1706 Node* cas = new (C, 5) StoreXConditionalNode(not_biased_ctrl, mem, adr,
kvn@423 1707 proto_node, mark_node);
kvn@423 1708 transform_later(cas);
kvn@423 1709 Node* proj = transform_later( new (C, 1) SCMemProjNode(cas));
kvn@423 1710 fast_lock_mem_phi->init_req(2, proj);
kvn@423 1711
kvn@423 1712
kvn@423 1713 // Second, check epoch bits.
kvn@423 1714 Node* rebiased_region = new (C, 3) RegionNode(3);
kvn@423 1715 Node* old_phi = new (C, 3) PhiNode( rebiased_region, TypeX_X);
kvn@423 1716 Node* new_phi = new (C, 3) PhiNode( rebiased_region, TypeX_X);
kvn@423 1717
kvn@423 1718 // Get slow path - mark word does NOT match epoch bits.
kvn@423 1719 Node* epoch_ctrl = opt_bits_test(ctrl, rebiased_region, 1, x_node,
kvn@423 1720 markOopDesc::epoch_mask_in_place, 0);
kvn@423 1721 // The epoch of the current bias is not valid, attempt to rebias the object
kvn@423 1722 // toward the current thread.
kvn@423 1723 rebiased_region->init_req(2, epoch_ctrl);
kvn@423 1724 old_phi->init_req(2, mark_node);
kvn@423 1725 new_phi->init_req(2, o_node);
kvn@423 1726
kvn@423 1727 // rebiased_region->in(1) is set to fast path.
kvn@423 1728 // The epoch of the current bias is still valid but we know
kvn@423 1729 // nothing about the owner; it might be set or it might be clear.
kvn@423 1730 Node* cmask = MakeConX(markOopDesc::biased_lock_mask_in_place |
kvn@423 1731 markOopDesc::age_mask_in_place |
kvn@423 1732 markOopDesc::epoch_mask_in_place);
kvn@423 1733 Node* old = transform_later(new (C, 3) AndXNode(mark_node, cmask));
kvn@423 1734 cast_thread = transform_later(new (C, 2) CastP2XNode(ctrl, thread));
kvn@423 1735 Node* new_mark = transform_later(new (C, 3) OrXNode(cast_thread, old));
kvn@423 1736 old_phi->init_req(1, old);
kvn@423 1737 new_phi->init_req(1, new_mark);
kvn@423 1738
kvn@423 1739 transform_later(rebiased_region);
kvn@423 1740 transform_later(old_phi);
kvn@423 1741 transform_later(new_phi);
kvn@423 1742
kvn@423 1743 // Try to acquire the bias of the object using an atomic operation.
kvn@423 1744 // If this fails we will go in to the runtime to revoke the object's bias.
kvn@423 1745 cas = new (C, 5) StoreXConditionalNode(rebiased_region, mem, adr,
kvn@423 1746 new_phi, old_phi);
kvn@423 1747 transform_later(cas);
kvn@423 1748 proj = transform_later( new (C, 1) SCMemProjNode(cas));
kvn@423 1749
kvn@423 1750 // Get slow path - Failed to CAS.
kvn@423 1751 not_biased_ctrl = opt_bits_test(rebiased_region, region, 4, cas, 0, 0);
kvn@423 1752 mem_phi->init_req(4, proj);
kvn@423 1753 // region->in(4) is set to fast path - the object is rebiased to the current thread.
kvn@423 1754
kvn@423 1755 // Failed to CAS.
kvn@423 1756 slow_path = new (C, 3) RegionNode(3);
kvn@423 1757 Node *slow_mem = new (C, 3) PhiNode( slow_path, Type::MEMORY, TypeRawPtr::BOTTOM);
kvn@423 1758
kvn@423 1759 slow_path->init_req(1, not_biased_ctrl); // Capture slow-control
kvn@423 1760 slow_mem->init_req(1, proj);
kvn@423 1761
kvn@423 1762 // Call CAS-based locking scheme (FastLock node).
kvn@423 1763
kvn@423 1764 transform_later(fast_lock_region);
kvn@423 1765 transform_later(fast_lock_mem_phi);
kvn@423 1766
kvn@423 1767 // Get slow path - FastLock failed to lock the object.
kvn@423 1768 ctrl = opt_bits_test(fast_lock_region, region, 2, flock, 0, 0);
kvn@423 1769 mem_phi->init_req(2, fast_lock_mem_phi);
kvn@423 1770 // region->in(2) is set to fast path - the object is locked to the current thread.
kvn@423 1771
kvn@423 1772 slow_path->init_req(2, ctrl); // Capture slow-control
kvn@423 1773 slow_mem->init_req(2, fast_lock_mem_phi);
kvn@423 1774
kvn@423 1775 transform_later(slow_path);
kvn@423 1776 transform_later(slow_mem);
kvn@423 1777 // Reset lock's memory edge.
kvn@423 1778 lock->set_req(TypeFunc::Memory, slow_mem);
kvn@423 1779
kvn@423 1780 } else {
kvn@423 1781 region = new (C, 3) RegionNode(3);
kvn@423 1782 // create a Phi for the memory state
kvn@423 1783 mem_phi = new (C, 3) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
kvn@423 1784
kvn@423 1785 // Optimize test; set region slot 2
kvn@423 1786 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0);
kvn@423 1787 mem_phi->init_req(2, mem);
kvn@423 1788 }
duke@0 1789
duke@0 1790 // Make slow path call
duke@0 1791 CallNode *call = make_slow_call( (CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(), OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path, obj, box );
duke@0 1792
duke@0 1793 extract_call_projections(call);
duke@0 1794
duke@0 1795 // Slow path can only throw asynchronous exceptions, which are always
duke@0 1796 // de-opted. So the compiler thinks the slow-call can never throw an
duke@0 1797 // exception. If it DOES throw an exception we would need the debug
duke@0 1798 // info removed first (since if it throws there is no monitor).
duke@0 1799 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
duke@0 1800 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
duke@0 1801
duke@0 1802 // Capture slow path
duke@0 1803 // disconnect fall-through projection from call and create a new one
duke@0 1804 // hook up users of fall-through projection to region
duke@0 1805 Node *slow_ctrl = _fallthroughproj->clone();
duke@0 1806 transform_later(slow_ctrl);
duke@0 1807 _igvn.hash_delete(_fallthroughproj);
duke@0 1808 _fallthroughproj->disconnect_inputs(NULL);
duke@0 1809 region->init_req(1, slow_ctrl);
duke@0 1810 // region inputs are now complete
duke@0 1811 transform_later(region);
duke@0 1812 _igvn.subsume_node(_fallthroughproj, region);
duke@0 1813
kvn@423 1814 Node *memproj = transform_later( new(C, 1) ProjNode(call, TypeFunc::Memory) );
duke@0 1815 mem_phi->init_req(1, memproj );
duke@0 1816 transform_later(mem_phi);
kvn@423 1817 _igvn.hash_delete(_memproj_fallthrough);
duke@0 1818 _igvn.subsume_node(_memproj_fallthrough, mem_phi);
duke@0 1819 }
duke@0 1820
duke@0 1821 //------------------------------expand_unlock_node----------------------
duke@0 1822 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
duke@0 1823
kvn@66 1824 Node* ctrl = unlock->in(TypeFunc::Control);
duke@0 1825 Node* mem = unlock->in(TypeFunc::Memory);
duke@0 1826 Node* obj = unlock->obj_node();
duke@0 1827 Node* box = unlock->box_node();
duke@0 1828
duke@0 1829 // No need for a null check on unlock
duke@0 1830
duke@0 1831 // Make the merge point
kvn@423 1832 Node *region;
kvn@423 1833 Node *mem_phi;
kvn@423 1834
kvn@423 1835 if (UseOptoBiasInlining) {
kvn@423 1836 // Check for biased locking unlock case, which is a no-op.
kvn@423 1837 // See the full descrition in MacroAssembler::biased_locking_exit().
kvn@423 1838 region = new (C, 4) RegionNode(4);
kvn@423 1839 // create a Phi for the memory state
kvn@423 1840 mem_phi = new (C, 4) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
kvn@423 1841 mem_phi->init_req(3, mem);
kvn@423 1842
kvn@423 1843 Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
kvn@423 1844 ctrl = opt_bits_test(ctrl, region, 3, mark_node,
kvn@423 1845 markOopDesc::biased_lock_mask_in_place,
kvn@423 1846 markOopDesc::biased_lock_pattern);
kvn@423 1847 } else {
kvn@423 1848 region = new (C, 3) RegionNode(3);
kvn@423 1849 // create a Phi for the memory state
kvn@423 1850 mem_phi = new (C, 3) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
kvn@423 1851 }
duke@0 1852
duke@0 1853 FastUnlockNode *funlock = new (C, 3) FastUnlockNode( ctrl, obj, box );
duke@0 1854 funlock = transform_later( funlock )->as_FastUnlock();
duke@0 1855 // Optimize test; set region slot 2
kvn@423 1856 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0);
duke@0 1857
duke@0 1858 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 );
duke@0 1859
duke@0 1860 extract_call_projections(call);
duke@0 1861
duke@0 1862 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
duke@0 1863 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
duke@0 1864
duke@0 1865 // No exceptions for unlocking
duke@0 1866 // Capture slow path
duke@0 1867 // disconnect fall-through projection from call and create a new one
duke@0 1868 // hook up users of fall-through projection to region
duke@0 1869 Node *slow_ctrl = _fallthroughproj->clone();
duke@0 1870 transform_later(slow_ctrl);
duke@0 1871 _igvn.hash_delete(_fallthroughproj);
duke@0 1872 _fallthroughproj->disconnect_inputs(NULL);
duke@0 1873 region->init_req(1, slow_ctrl);
duke@0 1874 // region inputs are now complete
duke@0 1875 transform_later(region);
duke@0 1876 _igvn.subsume_node(_fallthroughproj, region);
duke@0 1877
duke@0 1878 Node *memproj = transform_later( new(C, 1) ProjNode(call, TypeFunc::Memory) );
duke@0 1879 mem_phi->init_req(1, memproj );
duke@0 1880 mem_phi->init_req(2, mem);
duke@0 1881 transform_later(mem_phi);
kvn@423 1882 _igvn.hash_delete(_memproj_fallthrough);
duke@0 1883 _igvn.subsume_node(_memproj_fallthrough, mem_phi);
duke@0 1884 }
duke@0 1885
duke@0 1886 //------------------------------expand_macro_nodes----------------------
duke@0 1887 // Returns true if a failure occurred.
duke@0 1888 bool PhaseMacroExpand::expand_macro_nodes() {
duke@0 1889 if (C->macro_count() == 0)
duke@0 1890 return false;
kvn@73 1891 // attempt to eliminate allocations
kvn@73 1892 bool progress = true;
kvn@73 1893 while (progress) {
kvn@73 1894 progress = false;
kvn@73 1895 for (int i = C->macro_count(); i > 0; i--) {
kvn@73 1896 Node * n = C->macro_node(i-1);
kvn@73 1897 bool success = false;
kvn@73 1898 debug_only(int old_macro_count = C->macro_count(););
kvn@73 1899 switch (n->class_id()) {
kvn@73 1900 case Node::Class_Allocate:
kvn@73 1901 case Node::Class_AllocateArray:
kvn@73 1902 success = eliminate_allocate_node(n->as_Allocate());
kvn@73 1903 break;
kvn@73 1904 case Node::Class_Lock:
kvn@73 1905 case Node::Class_Unlock:
kvn@73 1906 success = eliminate_locking_node(n->as_AbstractLock());
kvn@73 1907 break;
kvn@73 1908 default:
kvn@216 1909 if (n->Opcode() == Op_Opaque1 || n->Opcode() == Op_Opaque2) {
kvn@216 1910 _igvn.add_users_to_worklist(n);
kvn@216 1911 _igvn.hash_delete(n);
kvn@216 1912 _igvn.subsume_node(n, n->in(1));
kvn@216 1913 success = true;
kvn@216 1914 } else {
kvn@216 1915 assert(false, "unknown node type in macro list");
kvn@216 1916 }
kvn@73 1917 }
kvn@73 1918 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
kvn@73 1919 progress = progress || success;
kvn@73 1920 }
kvn@73 1921 }
kvn@73 1922 // Make sure expansion will not cause node limit to be exceeded.
kvn@73 1923 // Worst case is a macro node gets expanded into about 50 nodes.
kvn@73 1924 // Allow 50% more for optimization.
duke@0 1925 if (C->check_node_count(C->macro_count() * 75, "out of nodes before macro expansion" ) )
duke@0 1926 return true;
kvn@73 1927
duke@0 1928 // expand "macro" nodes
duke@0 1929 // nodes are removed from the macro list as they are processed
duke@0 1930 while (C->macro_count() > 0) {
kvn@73 1931 int macro_count = C->macro_count();
kvn@73 1932 Node * n = C->macro_node(macro_count-1);
duke@0 1933 assert(n->is_macro(), "only macro nodes expected here");
duke@0 1934 if (_igvn.type(n) == Type::TOP || n->in(0)->is_top() ) {
duke@0 1935 // node is unreachable, so don't try to expand it
duke@0 1936 C->remove_macro_node(n);
duke@0 1937 continue;
duke@0 1938 }
duke@0 1939 switch (n->class_id()) {
duke@0 1940 case Node::Class_Allocate:
duke@0 1941 expand_allocate(n->as_Allocate());
duke@0 1942 break;
duke@0 1943 case Node::Class_AllocateArray:
duke@0 1944 expand_allocate_array(n->as_AllocateArray());
duke@0 1945 break;
duke@0 1946 case Node::Class_Lock:
duke@0 1947 expand_lock_node(n->as_Lock());
duke@0 1948 break;
duke@0 1949 case Node::Class_Unlock:
duke@0 1950 expand_unlock_node(n->as_Unlock());
duke@0 1951 break;
duke@0 1952 default:
duke@0 1953 assert(false, "unknown node type in macro list");
duke@0 1954 }
kvn@73 1955 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
duke@0 1956 if (C->failing()) return true;
duke@0 1957 }
coleenp@113 1958
coleenp@113 1959 _igvn.set_delay_transform(false);
duke@0 1960 _igvn.optimize();
duke@0 1961 return false;
duke@0 1962 }