Wed, 13 Apr 2011 14:33:03 -0700
6988308: assert((cnt > 0.0f) && (prob > 0.0f)) failed: Bad frequency assignment in if
Summary: Make sure cnt doesn't become negative and integer overflow doesn't happen.
Reviewed-by: kvn, twisti
duke@435 | 1 | /* |
phh@2423 | 2 | * Copyright (c) 2005, 2011, Oracle and/or its affiliates. All rights reserved. |
duke@435 | 3 | * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. |
duke@435 | 4 | * |
duke@435 | 5 | * This code is free software; you can redistribute it and/or modify it |
duke@435 | 6 | * under the terms of the GNU General Public License version 2 only, as |
duke@435 | 7 | * published by the Free Software Foundation. |
duke@435 | 8 | * |
duke@435 | 9 | * This code is distributed in the hope that it will be useful, but WITHOUT |
duke@435 | 10 | * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or |
duke@435 | 11 | * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
duke@435 | 12 | * version 2 for more details (a copy is included in the LICENSE file that |
duke@435 | 13 | * accompanied this code). |
duke@435 | 14 | * |
duke@435 | 15 | * You should have received a copy of the GNU General Public License version |
duke@435 | 16 | * 2 along with this work; if not, write to the Free Software Foundation, |
duke@435 | 17 | * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. |
duke@435 | 18 | * |
trims@1907 | 19 | * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA |
trims@1907 | 20 | * or visit www.oracle.com if you need additional information or have any |
trims@1907 | 21 | * questions. |
duke@435 | 22 | * |
duke@435 | 23 | */ |
duke@435 | 24 | |
stefank@2314 | 25 | #include "precompiled.hpp" |
stefank@2314 | 26 | #include "compiler/compileLog.hpp" |
stefank@2314 | 27 | #include "libadt/vectset.hpp" |
stefank@2314 | 28 | #include "opto/addnode.hpp" |
stefank@2314 | 29 | #include "opto/callnode.hpp" |
stefank@2314 | 30 | #include "opto/cfgnode.hpp" |
stefank@2314 | 31 | #include "opto/compile.hpp" |
stefank@2314 | 32 | #include "opto/connode.hpp" |
stefank@2314 | 33 | #include "opto/locknode.hpp" |
stefank@2314 | 34 | #include "opto/loopnode.hpp" |
stefank@2314 | 35 | #include "opto/macro.hpp" |
stefank@2314 | 36 | #include "opto/memnode.hpp" |
stefank@2314 | 37 | #include "opto/node.hpp" |
stefank@2314 | 38 | #include "opto/phaseX.hpp" |
stefank@2314 | 39 | #include "opto/rootnode.hpp" |
stefank@2314 | 40 | #include "opto/runtime.hpp" |
stefank@2314 | 41 | #include "opto/subnode.hpp" |
stefank@2314 | 42 | #include "opto/type.hpp" |
stefank@2314 | 43 | #include "runtime/sharedRuntime.hpp" |
duke@435 | 44 | |
duke@435 | 45 | |
duke@435 | 46 | // |
duke@435 | 47 | // Replace any references to "oldref" in inputs to "use" with "newref". |
duke@435 | 48 | // Returns the number of replacements made. |
duke@435 | 49 | // |
duke@435 | 50 | int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) { |
duke@435 | 51 | int nreplacements = 0; |
duke@435 | 52 | uint req = use->req(); |
duke@435 | 53 | for (uint j = 0; j < use->len(); j++) { |
duke@435 | 54 | Node *uin = use->in(j); |
duke@435 | 55 | if (uin == oldref) { |
duke@435 | 56 | if (j < req) |
duke@435 | 57 | use->set_req(j, newref); |
duke@435 | 58 | else |
duke@435 | 59 | use->set_prec(j, newref); |
duke@435 | 60 | nreplacements++; |
duke@435 | 61 | } else if (j >= req && uin == NULL) { |
duke@435 | 62 | break; |
duke@435 | 63 | } |
duke@435 | 64 | } |
duke@435 | 65 | return nreplacements; |
duke@435 | 66 | } |
duke@435 | 67 | |
duke@435 | 68 | void PhaseMacroExpand::copy_call_debug_info(CallNode *oldcall, CallNode * newcall) { |
duke@435 | 69 | // Copy debug information and adjust JVMState information |
duke@435 | 70 | uint old_dbg_start = oldcall->tf()->domain()->cnt(); |
duke@435 | 71 | uint new_dbg_start = newcall->tf()->domain()->cnt(); |
duke@435 | 72 | int jvms_adj = new_dbg_start - old_dbg_start; |
duke@435 | 73 | assert (new_dbg_start == newcall->req(), "argument count mismatch"); |
kvn@498 | 74 | |
kvn@498 | 75 | Dict* sosn_map = new Dict(cmpkey,hashkey); |
duke@435 | 76 | for (uint i = old_dbg_start; i < oldcall->req(); i++) { |
kvn@498 | 77 | Node* old_in = oldcall->in(i); |
kvn@498 | 78 | // Clone old SafePointScalarObjectNodes, adjusting their field contents. |
kvn@895 | 79 | if (old_in != NULL && old_in->is_SafePointScalarObject()) { |
kvn@498 | 80 | SafePointScalarObjectNode* old_sosn = old_in->as_SafePointScalarObject(); |
kvn@498 | 81 | uint old_unique = C->unique(); |
kvn@498 | 82 | Node* new_in = old_sosn->clone(jvms_adj, sosn_map); |
kvn@498 | 83 | if (old_unique != C->unique()) { |
kvn@1036 | 84 | new_in->set_req(0, newcall->in(0)); // reset control edge |
kvn@498 | 85 | new_in = transform_later(new_in); // Register new node. |
kvn@498 | 86 | } |
kvn@498 | 87 | old_in = new_in; |
kvn@498 | 88 | } |
kvn@498 | 89 | newcall->add_req(old_in); |
duke@435 | 90 | } |
kvn@498 | 91 | |
duke@435 | 92 | newcall->set_jvms(oldcall->jvms()); |
duke@435 | 93 | for (JVMState *jvms = newcall->jvms(); jvms != NULL; jvms = jvms->caller()) { |
duke@435 | 94 | jvms->set_map(newcall); |
duke@435 | 95 | jvms->set_locoff(jvms->locoff()+jvms_adj); |
duke@435 | 96 | jvms->set_stkoff(jvms->stkoff()+jvms_adj); |
duke@435 | 97 | jvms->set_monoff(jvms->monoff()+jvms_adj); |
kvn@498 | 98 | jvms->set_scloff(jvms->scloff()+jvms_adj); |
duke@435 | 99 | jvms->set_endoff(jvms->endoff()+jvms_adj); |
duke@435 | 100 | } |
duke@435 | 101 | } |
duke@435 | 102 | |
kvn@855 | 103 | Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word, int mask, int bits, bool return_fast_path) { |
kvn@855 | 104 | Node* cmp; |
kvn@855 | 105 | if (mask != 0) { |
kvn@855 | 106 | Node* and_node = transform_later(new (C, 3) AndXNode(word, MakeConX(mask))); |
kvn@855 | 107 | cmp = transform_later(new (C, 3) CmpXNode(and_node, MakeConX(bits))); |
kvn@855 | 108 | } else { |
kvn@855 | 109 | cmp = word; |
kvn@855 | 110 | } |
kvn@855 | 111 | Node* bol = transform_later(new (C, 2) BoolNode(cmp, BoolTest::ne)); |
kvn@855 | 112 | IfNode* iff = new (C, 2) IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN ); |
kvn@855 | 113 | transform_later(iff); |
duke@435 | 114 | |
kvn@855 | 115 | // Fast path taken. |
kvn@855 | 116 | Node *fast_taken = transform_later( new (C, 1) IfFalseNode(iff) ); |
duke@435 | 117 | |
duke@435 | 118 | // Fast path not-taken, i.e. slow path |
kvn@855 | 119 | Node *slow_taken = transform_later( new (C, 1) IfTrueNode(iff) ); |
kvn@855 | 120 | |
kvn@855 | 121 | if (return_fast_path) { |
kvn@855 | 122 | region->init_req(edge, slow_taken); // Capture slow-control |
kvn@855 | 123 | return fast_taken; |
kvn@855 | 124 | } else { |
kvn@855 | 125 | region->init_req(edge, fast_taken); // Capture fast-control |
kvn@855 | 126 | return slow_taken; |
kvn@855 | 127 | } |
duke@435 | 128 | } |
duke@435 | 129 | |
duke@435 | 130 | //--------------------copy_predefined_input_for_runtime_call-------------------- |
duke@435 | 131 | void PhaseMacroExpand::copy_predefined_input_for_runtime_call(Node * ctrl, CallNode* oldcall, CallNode* call) { |
duke@435 | 132 | // Set fixed predefined input arguments |
duke@435 | 133 | call->init_req( TypeFunc::Control, ctrl ); |
duke@435 | 134 | call->init_req( TypeFunc::I_O , oldcall->in( TypeFunc::I_O) ); |
duke@435 | 135 | call->init_req( TypeFunc::Memory , oldcall->in( TypeFunc::Memory ) ); // ????? |
duke@435 | 136 | call->init_req( TypeFunc::ReturnAdr, oldcall->in( TypeFunc::ReturnAdr ) ); |
duke@435 | 137 | call->init_req( TypeFunc::FramePtr, oldcall->in( TypeFunc::FramePtr ) ); |
duke@435 | 138 | } |
duke@435 | 139 | |
duke@435 | 140 | //------------------------------make_slow_call--------------------------------- |
duke@435 | 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) { |
duke@435 | 142 | |
duke@435 | 143 | // Slow-path call |
duke@435 | 144 | int size = slow_call_type->domain()->cnt(); |
duke@435 | 145 | CallNode *call = leaf_name |
duke@435 | 146 | ? (CallNode*)new (C, size) CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM ) |
duke@435 | 147 | : (CallNode*)new (C, size) CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), oldcall->jvms()->bci(), TypeRawPtr::BOTTOM ); |
duke@435 | 148 | |
duke@435 | 149 | // Slow path call has no side-effects, uses few values |
duke@435 | 150 | copy_predefined_input_for_runtime_call(slow_path, oldcall, call ); |
duke@435 | 151 | if (parm0 != NULL) call->init_req(TypeFunc::Parms+0, parm0); |
duke@435 | 152 | if (parm1 != NULL) call->init_req(TypeFunc::Parms+1, parm1); |
duke@435 | 153 | copy_call_debug_info(oldcall, call); |
duke@435 | 154 | call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON. |
kvn@1976 | 155 | _igvn.replace_node(oldcall, call); |
duke@435 | 156 | transform_later(call); |
duke@435 | 157 | |
duke@435 | 158 | return call; |
duke@435 | 159 | } |
duke@435 | 160 | |
duke@435 | 161 | void PhaseMacroExpand::extract_call_projections(CallNode *call) { |
duke@435 | 162 | _fallthroughproj = NULL; |
duke@435 | 163 | _fallthroughcatchproj = NULL; |
duke@435 | 164 | _ioproj_fallthrough = NULL; |
duke@435 | 165 | _ioproj_catchall = NULL; |
duke@435 | 166 | _catchallcatchproj = NULL; |
duke@435 | 167 | _memproj_fallthrough = NULL; |
duke@435 | 168 | _memproj_catchall = NULL; |
duke@435 | 169 | _resproj = NULL; |
duke@435 | 170 | for (DUIterator_Fast imax, i = call->fast_outs(imax); i < imax; i++) { |
duke@435 | 171 | ProjNode *pn = call->fast_out(i)->as_Proj(); |
duke@435 | 172 | switch (pn->_con) { |
duke@435 | 173 | case TypeFunc::Control: |
duke@435 | 174 | { |
duke@435 | 175 | // For Control (fallthrough) and I_O (catch_all_index) we have CatchProj -> Catch -> Proj |
duke@435 | 176 | _fallthroughproj = pn; |
duke@435 | 177 | DUIterator_Fast jmax, j = pn->fast_outs(jmax); |
duke@435 | 178 | const Node *cn = pn->fast_out(j); |
duke@435 | 179 | if (cn->is_Catch()) { |
duke@435 | 180 | ProjNode *cpn = NULL; |
duke@435 | 181 | for (DUIterator_Fast kmax, k = cn->fast_outs(kmax); k < kmax; k++) { |
duke@435 | 182 | cpn = cn->fast_out(k)->as_Proj(); |
duke@435 | 183 | assert(cpn->is_CatchProj(), "must be a CatchProjNode"); |
duke@435 | 184 | if (cpn->_con == CatchProjNode::fall_through_index) |
duke@435 | 185 | _fallthroughcatchproj = cpn; |
duke@435 | 186 | else { |
duke@435 | 187 | assert(cpn->_con == CatchProjNode::catch_all_index, "must be correct index."); |
duke@435 | 188 | _catchallcatchproj = cpn; |
duke@435 | 189 | } |
duke@435 | 190 | } |
duke@435 | 191 | } |
duke@435 | 192 | break; |
duke@435 | 193 | } |
duke@435 | 194 | case TypeFunc::I_O: |
duke@435 | 195 | if (pn->_is_io_use) |
duke@435 | 196 | _ioproj_catchall = pn; |
duke@435 | 197 | else |
duke@435 | 198 | _ioproj_fallthrough = pn; |
duke@435 | 199 | break; |
duke@435 | 200 | case TypeFunc::Memory: |
duke@435 | 201 | if (pn->_is_io_use) |
duke@435 | 202 | _memproj_catchall = pn; |
duke@435 | 203 | else |
duke@435 | 204 | _memproj_fallthrough = pn; |
duke@435 | 205 | break; |
duke@435 | 206 | case TypeFunc::Parms: |
duke@435 | 207 | _resproj = pn; |
duke@435 | 208 | break; |
duke@435 | 209 | default: |
duke@435 | 210 | assert(false, "unexpected projection from allocation node."); |
duke@435 | 211 | } |
duke@435 | 212 | } |
duke@435 | 213 | |
duke@435 | 214 | } |
duke@435 | 215 | |
kvn@508 | 216 | // Eliminate a card mark sequence. p2x is a ConvP2XNode |
kvn@1286 | 217 | void PhaseMacroExpand::eliminate_card_mark(Node* p2x) { |
kvn@508 | 218 | assert(p2x->Opcode() == Op_CastP2X, "ConvP2XNode required"); |
kvn@1286 | 219 | if (!UseG1GC) { |
kvn@1286 | 220 | // vanilla/CMS post barrier |
kvn@1286 | 221 | Node *shift = p2x->unique_out(); |
kvn@1286 | 222 | Node *addp = shift->unique_out(); |
kvn@1286 | 223 | for (DUIterator_Last jmin, j = addp->last_outs(jmin); j >= jmin; --j) { |
kvn@1286 | 224 | Node *st = addp->last_out(j); |
kvn@1286 | 225 | assert(st->is_Store(), "store required"); |
kvn@1286 | 226 | _igvn.replace_node(st, st->in(MemNode::Memory)); |
kvn@1286 | 227 | } |
kvn@1286 | 228 | } else { |
kvn@1286 | 229 | // G1 pre/post barriers |
kvn@1286 | 230 | assert(p2x->outcnt() == 2, "expects 2 users: Xor and URShift nodes"); |
kvn@1286 | 231 | // It could be only one user, URShift node, in Object.clone() instrinsic |
kvn@1286 | 232 | // but the new allocation is passed to arraycopy stub and it could not |
kvn@1286 | 233 | // be scalar replaced. So we don't check the case. |
kvn@1286 | 234 | |
kvn@1286 | 235 | // Remove G1 post barrier. |
kvn@1286 | 236 | |
kvn@1286 | 237 | // Search for CastP2X->Xor->URShift->Cmp path which |
kvn@1286 | 238 | // checks if the store done to a different from the value's region. |
kvn@1286 | 239 | // And replace Cmp with #0 (false) to collapse G1 post barrier. |
kvn@1286 | 240 | Node* xorx = NULL; |
kvn@1286 | 241 | for (DUIterator_Fast imax, i = p2x->fast_outs(imax); i < imax; i++) { |
kvn@1286 | 242 | Node* u = p2x->fast_out(i); |
kvn@1286 | 243 | if (u->Opcode() == Op_XorX) { |
kvn@1286 | 244 | xorx = u; |
kvn@1286 | 245 | break; |
kvn@1286 | 246 | } |
kvn@1286 | 247 | } |
kvn@1286 | 248 | assert(xorx != NULL, "missing G1 post barrier"); |
kvn@1286 | 249 | Node* shift = xorx->unique_out(); |
kvn@1286 | 250 | Node* cmpx = shift->unique_out(); |
kvn@1286 | 251 | assert(cmpx->is_Cmp() && cmpx->unique_out()->is_Bool() && |
kvn@1286 | 252 | cmpx->unique_out()->as_Bool()->_test._test == BoolTest::ne, |
kvn@1286 | 253 | "missing region check in G1 post barrier"); |
kvn@1286 | 254 | _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ)); |
kvn@1286 | 255 | |
kvn@1286 | 256 | // Remove G1 pre barrier. |
kvn@1286 | 257 | |
kvn@1286 | 258 | // Search "if (marking != 0)" check and set it to "false". |
kvn@1286 | 259 | Node* this_region = p2x->in(0); |
kvn@1286 | 260 | assert(this_region != NULL, ""); |
kvn@1286 | 261 | // There is no G1 pre barrier if previous stored value is NULL |
kvn@1286 | 262 | // (for example, after initialization). |
kvn@1286 | 263 | if (this_region->is_Region() && this_region->req() == 3) { |
kvn@1286 | 264 | int ind = 1; |
kvn@1286 | 265 | if (!this_region->in(ind)->is_IfFalse()) { |
kvn@1286 | 266 | ind = 2; |
kvn@1286 | 267 | } |
kvn@1286 | 268 | if (this_region->in(ind)->is_IfFalse()) { |
kvn@1286 | 269 | Node* bol = this_region->in(ind)->in(0)->in(1); |
kvn@1286 | 270 | assert(bol->is_Bool(), ""); |
kvn@1286 | 271 | cmpx = bol->in(1); |
kvn@1286 | 272 | if (bol->as_Bool()->_test._test == BoolTest::ne && |
kvn@1286 | 273 | cmpx->is_Cmp() && cmpx->in(2) == intcon(0) && |
kvn@1286 | 274 | cmpx->in(1)->is_Load()) { |
kvn@1286 | 275 | Node* adr = cmpx->in(1)->as_Load()->in(MemNode::Address); |
kvn@1286 | 276 | const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + |
kvn@1286 | 277 | PtrQueue::byte_offset_of_active()); |
kvn@1286 | 278 | if (adr->is_AddP() && adr->in(AddPNode::Base) == top() && |
kvn@1286 | 279 | adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal && |
kvn@1286 | 280 | adr->in(AddPNode::Offset) == MakeConX(marking_offset)) { |
kvn@1286 | 281 | _igvn.replace_node(cmpx, makecon(TypeInt::CC_EQ)); |
kvn@1286 | 282 | } |
kvn@1286 | 283 | } |
kvn@1286 | 284 | } |
kvn@1286 | 285 | } |
kvn@1286 | 286 | // Now CastP2X can be removed since it is used only on dead path |
kvn@1286 | 287 | // which currently still alive until igvn optimize it. |
kvn@1286 | 288 | assert(p2x->unique_out()->Opcode() == Op_URShiftX, ""); |
kvn@1286 | 289 | _igvn.replace_node(p2x, top()); |
kvn@508 | 290 | } |
kvn@508 | 291 | } |
kvn@508 | 292 | |
kvn@508 | 293 | // Search for a memory operation for the specified memory slice. |
kvn@688 | 294 | static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) { |
kvn@508 | 295 | Node *orig_mem = mem; |
kvn@508 | 296 | Node *alloc_mem = alloc->in(TypeFunc::Memory); |
kvn@688 | 297 | const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr(); |
kvn@508 | 298 | while (true) { |
kvn@508 | 299 | if (mem == alloc_mem || mem == start_mem ) { |
twisti@1040 | 300 | return mem; // hit one of our sentinels |
kvn@508 | 301 | } else if (mem->is_MergeMem()) { |
kvn@508 | 302 | mem = mem->as_MergeMem()->memory_at(alias_idx); |
kvn@508 | 303 | } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) { |
kvn@508 | 304 | Node *in = mem->in(0); |
kvn@508 | 305 | // we can safely skip over safepoints, calls, locks and membars because we |
kvn@508 | 306 | // already know that the object is safe to eliminate. |
kvn@508 | 307 | if (in->is_Initialize() && in->as_Initialize()->allocation() == alloc) { |
kvn@508 | 308 | return in; |
kvn@688 | 309 | } else if (in->is_Call()) { |
kvn@688 | 310 | CallNode *call = in->as_Call(); |
kvn@688 | 311 | if (!call->may_modify(tinst, phase)) { |
kvn@688 | 312 | mem = call->in(TypeFunc::Memory); |
kvn@688 | 313 | } |
kvn@688 | 314 | mem = in->in(TypeFunc::Memory); |
kvn@688 | 315 | } else if (in->is_MemBar()) { |
kvn@508 | 316 | mem = in->in(TypeFunc::Memory); |
kvn@508 | 317 | } else { |
kvn@508 | 318 | assert(false, "unexpected projection"); |
kvn@508 | 319 | } |
kvn@508 | 320 | } else if (mem->is_Store()) { |
kvn@508 | 321 | const TypePtr* atype = mem->as_Store()->adr_type(); |
kvn@508 | 322 | int adr_idx = Compile::current()->get_alias_index(atype); |
kvn@508 | 323 | if (adr_idx == alias_idx) { |
kvn@508 | 324 | assert(atype->isa_oopptr(), "address type must be oopptr"); |
kvn@508 | 325 | int adr_offset = atype->offset(); |
kvn@508 | 326 | uint adr_iid = atype->is_oopptr()->instance_id(); |
kvn@508 | 327 | // Array elements references have the same alias_idx |
kvn@508 | 328 | // but different offset and different instance_id. |
kvn@508 | 329 | if (adr_offset == offset && adr_iid == alloc->_idx) |
kvn@508 | 330 | return mem; |
kvn@508 | 331 | } else { |
kvn@508 | 332 | assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw"); |
kvn@508 | 333 | } |
kvn@508 | 334 | mem = mem->in(MemNode::Memory); |
kvn@1535 | 335 | } else if (mem->is_ClearArray()) { |
kvn@1535 | 336 | if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) { |
kvn@1535 | 337 | // Can not bypass initialization of the instance |
kvn@1535 | 338 | // we are looking. |
kvn@1535 | 339 | debug_only(intptr_t offset;) |
kvn@1535 | 340 | assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity"); |
kvn@1535 | 341 | InitializeNode* init = alloc->as_Allocate()->initialization(); |
kvn@1535 | 342 | // We are looking for stored value, return Initialize node |
kvn@1535 | 343 | // or memory edge from Allocate node. |
kvn@1535 | 344 | if (init != NULL) |
kvn@1535 | 345 | return init; |
kvn@1535 | 346 | else |
kvn@1535 | 347 | return alloc->in(TypeFunc::Memory); // It will produce zero value (see callers). |
kvn@1535 | 348 | } |
kvn@1535 | 349 | // Otherwise skip it (the call updated 'mem' value). |
kvn@1019 | 350 | } else if (mem->Opcode() == Op_SCMemProj) { |
kvn@1019 | 351 | assert(mem->in(0)->is_LoadStore(), "sanity"); |
kvn@1019 | 352 | const TypePtr* atype = mem->in(0)->in(MemNode::Address)->bottom_type()->is_ptr(); |
kvn@1019 | 353 | int adr_idx = Compile::current()->get_alias_index(atype); |
kvn@1019 | 354 | if (adr_idx == alias_idx) { |
kvn@1019 | 355 | assert(false, "Object is not scalar replaceable if a LoadStore node access its field"); |
kvn@1019 | 356 | return NULL; |
kvn@1019 | 357 | } |
kvn@1019 | 358 | mem = mem->in(0)->in(MemNode::Memory); |
kvn@508 | 359 | } else { |
kvn@508 | 360 | return mem; |
kvn@508 | 361 | } |
kvn@682 | 362 | assert(mem != orig_mem, "dead memory loop"); |
kvn@508 | 363 | } |
kvn@508 | 364 | } |
kvn@508 | 365 | |
kvn@508 | 366 | // |
kvn@508 | 367 | // Given a Memory Phi, compute a value Phi containing the values from stores |
kvn@508 | 368 | // on the input paths. |
kvn@508 | 369 | // Note: this function is recursive, its depth is limied by the "level" argument |
kvn@508 | 370 | // Returns the computed Phi, or NULL if it cannot compute it. |
kvn@682 | 371 | 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@682 | 372 | assert(mem->is_Phi(), "sanity"); |
kvn@682 | 373 | int alias_idx = C->get_alias_index(adr_t); |
kvn@682 | 374 | int offset = adr_t->offset(); |
kvn@682 | 375 | int instance_id = adr_t->instance_id(); |
kvn@682 | 376 | |
kvn@682 | 377 | // Check if an appropriate value phi already exists. |
kvn@682 | 378 | Node* region = mem->in(0); |
kvn@682 | 379 | for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) { |
kvn@682 | 380 | Node* phi = region->fast_out(k); |
kvn@682 | 381 | if (phi->is_Phi() && phi != mem && |
kvn@682 | 382 | phi->as_Phi()->is_same_inst_field(phi_type, instance_id, alias_idx, offset)) { |
kvn@682 | 383 | return phi; |
kvn@682 | 384 | } |
kvn@682 | 385 | } |
kvn@682 | 386 | // Check if an appropriate new value phi already exists. |
kvn@682 | 387 | Node* new_phi = NULL; |
kvn@682 | 388 | uint size = value_phis->size(); |
kvn@682 | 389 | for (uint i=0; i < size; i++) { |
kvn@682 | 390 | if ( mem->_idx == value_phis->index_at(i) ) { |
kvn@682 | 391 | return value_phis->node_at(i); |
kvn@682 | 392 | } |
kvn@682 | 393 | } |
kvn@508 | 394 | |
kvn@508 | 395 | if (level <= 0) { |
kvn@688 | 396 | return NULL; // Give up: phi tree too deep |
kvn@508 | 397 | } |
kvn@508 | 398 | Node *start_mem = C->start()->proj_out(TypeFunc::Memory); |
kvn@508 | 399 | Node *alloc_mem = alloc->in(TypeFunc::Memory); |
kvn@508 | 400 | |
kvn@508 | 401 | uint length = mem->req(); |
kvn@508 | 402 | GrowableArray <Node *> values(length, length, NULL); |
kvn@508 | 403 | |
kvn@682 | 404 | // create a new Phi for the value |
kvn@682 | 405 | PhiNode *phi = new (C, length) PhiNode(mem->in(0), phi_type, NULL, instance_id, alias_idx, offset); |
kvn@682 | 406 | transform_later(phi); |
kvn@682 | 407 | value_phis->push(phi, mem->_idx); |
kvn@682 | 408 | |
kvn@508 | 409 | for (uint j = 1; j < length; j++) { |
kvn@508 | 410 | Node *in = mem->in(j); |
kvn@508 | 411 | if (in == NULL || in->is_top()) { |
kvn@508 | 412 | values.at_put(j, in); |
kvn@508 | 413 | } else { |
kvn@688 | 414 | Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn); |
kvn@508 | 415 | if (val == start_mem || val == alloc_mem) { |
kvn@508 | 416 | // hit a sentinel, return appropriate 0 value |
kvn@508 | 417 | values.at_put(j, _igvn.zerocon(ft)); |
kvn@508 | 418 | continue; |
kvn@508 | 419 | } |
kvn@508 | 420 | if (val->is_Initialize()) { |
kvn@508 | 421 | val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn); |
kvn@508 | 422 | } |
kvn@508 | 423 | if (val == NULL) { |
kvn@508 | 424 | return NULL; // can't find a value on this path |
kvn@508 | 425 | } |
kvn@508 | 426 | if (val == mem) { |
kvn@508 | 427 | values.at_put(j, mem); |
kvn@508 | 428 | } else if (val->is_Store()) { |
kvn@508 | 429 | values.at_put(j, val->in(MemNode::ValueIn)); |
kvn@508 | 430 | } else if(val->is_Proj() && val->in(0) == alloc) { |
kvn@508 | 431 | values.at_put(j, _igvn.zerocon(ft)); |
kvn@508 | 432 | } else if (val->is_Phi()) { |
kvn@682 | 433 | val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1); |
kvn@682 | 434 | if (val == NULL) { |
kvn@682 | 435 | return NULL; |
kvn@508 | 436 | } |
kvn@682 | 437 | values.at_put(j, val); |
kvn@1019 | 438 | } else if (val->Opcode() == Op_SCMemProj) { |
kvn@1019 | 439 | assert(val->in(0)->is_LoadStore(), "sanity"); |
kvn@1019 | 440 | assert(false, "Object is not scalar replaceable if a LoadStore node access its field"); |
kvn@1019 | 441 | return NULL; |
kvn@508 | 442 | } else { |
kvn@1019 | 443 | #ifdef ASSERT |
kvn@1019 | 444 | val->dump(); |
kvn@688 | 445 | assert(false, "unknown node on this path"); |
kvn@1019 | 446 | #endif |
kvn@688 | 447 | return NULL; // unknown node on this path |
kvn@508 | 448 | } |
kvn@508 | 449 | } |
kvn@508 | 450 | } |
kvn@682 | 451 | // Set Phi's inputs |
kvn@508 | 452 | for (uint j = 1; j < length; j++) { |
kvn@508 | 453 | if (values.at(j) == mem) { |
kvn@508 | 454 | phi->init_req(j, phi); |
kvn@508 | 455 | } else { |
kvn@508 | 456 | phi->init_req(j, values.at(j)); |
kvn@508 | 457 | } |
kvn@508 | 458 | } |
kvn@508 | 459 | return phi; |
kvn@508 | 460 | } |
kvn@508 | 461 | |
kvn@508 | 462 | // Search the last value stored into the object's field. |
kvn@508 | 463 | Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, Node *alloc) { |
kvn@658 | 464 | assert(adr_t->is_known_instance_field(), "instance required"); |
kvn@658 | 465 | int instance_id = adr_t->instance_id(); |
kvn@658 | 466 | assert((uint)instance_id == alloc->_idx, "wrong allocation"); |
kvn@508 | 467 | |
kvn@508 | 468 | int alias_idx = C->get_alias_index(adr_t); |
kvn@508 | 469 | int offset = adr_t->offset(); |
kvn@508 | 470 | Node *start_mem = C->start()->proj_out(TypeFunc::Memory); |
kvn@508 | 471 | Node *alloc_ctrl = alloc->in(TypeFunc::Control); |
kvn@508 | 472 | Node *alloc_mem = alloc->in(TypeFunc::Memory); |
kvn@682 | 473 | Arena *a = Thread::current()->resource_area(); |
kvn@682 | 474 | VectorSet visited(a); |
kvn@508 | 475 | |
kvn@508 | 476 | |
kvn@508 | 477 | bool done = sfpt_mem == alloc_mem; |
kvn@508 | 478 | Node *mem = sfpt_mem; |
kvn@508 | 479 | while (!done) { |
kvn@508 | 480 | if (visited.test_set(mem->_idx)) { |
kvn@508 | 481 | return NULL; // found a loop, give up |
kvn@508 | 482 | } |
kvn@688 | 483 | mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn); |
kvn@508 | 484 | if (mem == start_mem || mem == alloc_mem) { |
kvn@508 | 485 | done = true; // hit a sentinel, return appropriate 0 value |
kvn@508 | 486 | } else if (mem->is_Initialize()) { |
kvn@508 | 487 | mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn); |
kvn@508 | 488 | if (mem == NULL) { |
kvn@508 | 489 | done = true; // Something go wrong. |
kvn@508 | 490 | } else if (mem->is_Store()) { |
kvn@508 | 491 | const TypePtr* atype = mem->as_Store()->adr_type(); |
kvn@508 | 492 | assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice"); |
kvn@508 | 493 | done = true; |
kvn@508 | 494 | } |
kvn@508 | 495 | } else if (mem->is_Store()) { |
kvn@508 | 496 | const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr(); |
kvn@508 | 497 | assert(atype != NULL, "address type must be oopptr"); |
kvn@508 | 498 | assert(C->get_alias_index(atype) == alias_idx && |
kvn@658 | 499 | atype->is_known_instance_field() && atype->offset() == offset && |
kvn@508 | 500 | atype->instance_id() == instance_id, "store is correct memory slice"); |
kvn@508 | 501 | done = true; |
kvn@508 | 502 | } else if (mem->is_Phi()) { |
kvn@508 | 503 | // try to find a phi's unique input |
kvn@508 | 504 | Node *unique_input = NULL; |
kvn@508 | 505 | Node *top = C->top(); |
kvn@508 | 506 | for (uint i = 1; i < mem->req(); i++) { |
kvn@688 | 507 | Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn); |
kvn@508 | 508 | if (n == NULL || n == top || n == mem) { |
kvn@508 | 509 | continue; |
kvn@508 | 510 | } else if (unique_input == NULL) { |
kvn@508 | 511 | unique_input = n; |
kvn@508 | 512 | } else if (unique_input != n) { |
kvn@508 | 513 | unique_input = top; |
kvn@508 | 514 | break; |
kvn@508 | 515 | } |
kvn@508 | 516 | } |
kvn@508 | 517 | if (unique_input != NULL && unique_input != top) { |
kvn@508 | 518 | mem = unique_input; |
kvn@508 | 519 | } else { |
kvn@508 | 520 | done = true; |
kvn@508 | 521 | } |
kvn@508 | 522 | } else { |
kvn@508 | 523 | assert(false, "unexpected node"); |
kvn@508 | 524 | } |
kvn@508 | 525 | } |
kvn@508 | 526 | if (mem != NULL) { |
kvn@508 | 527 | if (mem == start_mem || mem == alloc_mem) { |
kvn@508 | 528 | // hit a sentinel, return appropriate 0 value |
kvn@508 | 529 | return _igvn.zerocon(ft); |
kvn@508 | 530 | } else if (mem->is_Store()) { |
kvn@508 | 531 | return mem->in(MemNode::ValueIn); |
kvn@508 | 532 | } else if (mem->is_Phi()) { |
kvn@508 | 533 | // attempt to produce a Phi reflecting the values on the input paths of the Phi |
kvn@682 | 534 | Node_Stack value_phis(a, 8); |
kvn@688 | 535 | Node * phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit); |
kvn@508 | 536 | if (phi != NULL) { |
kvn@508 | 537 | return phi; |
kvn@682 | 538 | } else { |
kvn@682 | 539 | // Kill all new Phis |
kvn@682 | 540 | while(value_phis.is_nonempty()) { |
kvn@682 | 541 | Node* n = value_phis.node(); |
kvn@1976 | 542 | _igvn.replace_node(n, C->top()); |
kvn@682 | 543 | value_phis.pop(); |
kvn@682 | 544 | } |
kvn@508 | 545 | } |
kvn@508 | 546 | } |
kvn@508 | 547 | } |
kvn@508 | 548 | // Something go wrong. |
kvn@508 | 549 | return NULL; |
kvn@508 | 550 | } |
kvn@508 | 551 | |
kvn@508 | 552 | // Check the possibility of scalar replacement. |
kvn@508 | 553 | bool PhaseMacroExpand::can_eliminate_allocation(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) { |
kvn@508 | 554 | // Scan the uses of the allocation to check for anything that would |
kvn@508 | 555 | // prevent us from eliminating it. |
kvn@508 | 556 | NOT_PRODUCT( const char* fail_eliminate = NULL; ) |
kvn@508 | 557 | DEBUG_ONLY( Node* disq_node = NULL; ) |
kvn@508 | 558 | bool can_eliminate = true; |
kvn@508 | 559 | |
kvn@508 | 560 | Node* res = alloc->result_cast(); |
kvn@508 | 561 | const TypeOopPtr* res_type = NULL; |
kvn@508 | 562 | if (res == NULL) { |
kvn@508 | 563 | // All users were eliminated. |
kvn@508 | 564 | } else if (!res->is_CheckCastPP()) { |
kvn@508 | 565 | alloc->_is_scalar_replaceable = false; // don't try again |
kvn@508 | 566 | NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";) |
kvn@508 | 567 | can_eliminate = false; |
kvn@508 | 568 | } else { |
kvn@508 | 569 | res_type = _igvn.type(res)->isa_oopptr(); |
kvn@508 | 570 | if (res_type == NULL) { |
kvn@508 | 571 | NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";) |
kvn@508 | 572 | can_eliminate = false; |
kvn@508 | 573 | } else if (res_type->isa_aryptr()) { |
kvn@508 | 574 | int length = alloc->in(AllocateNode::ALength)->find_int_con(-1); |
kvn@508 | 575 | if (length < 0) { |
kvn@508 | 576 | NOT_PRODUCT(fail_eliminate = "Array's size is not constant";) |
kvn@508 | 577 | can_eliminate = false; |
kvn@508 | 578 | } |
kvn@508 | 579 | } |
kvn@508 | 580 | } |
kvn@508 | 581 | |
kvn@508 | 582 | if (can_eliminate && res != NULL) { |
kvn@508 | 583 | for (DUIterator_Fast jmax, j = res->fast_outs(jmax); |
kvn@508 | 584 | j < jmax && can_eliminate; j++) { |
kvn@508 | 585 | Node* use = res->fast_out(j); |
kvn@508 | 586 | |
kvn@508 | 587 | if (use->is_AddP()) { |
kvn@508 | 588 | const TypePtr* addp_type = _igvn.type(use)->is_ptr(); |
kvn@508 | 589 | int offset = addp_type->offset(); |
kvn@508 | 590 | |
kvn@508 | 591 | if (offset == Type::OffsetTop || offset == Type::OffsetBot) { |
kvn@508 | 592 | NOT_PRODUCT(fail_eliminate = "Undefined field referrence";) |
kvn@508 | 593 | can_eliminate = false; |
kvn@508 | 594 | break; |
kvn@508 | 595 | } |
kvn@508 | 596 | for (DUIterator_Fast kmax, k = use->fast_outs(kmax); |
kvn@508 | 597 | k < kmax && can_eliminate; k++) { |
kvn@508 | 598 | Node* n = use->fast_out(k); |
kvn@508 | 599 | if (!n->is_Store() && n->Opcode() != Op_CastP2X) { |
kvn@508 | 600 | DEBUG_ONLY(disq_node = n;) |
kvn@688 | 601 | if (n->is_Load() || n->is_LoadStore()) { |
kvn@508 | 602 | NOT_PRODUCT(fail_eliminate = "Field load";) |
kvn@508 | 603 | } else { |
kvn@508 | 604 | NOT_PRODUCT(fail_eliminate = "Not store field referrence";) |
kvn@508 | 605 | } |
kvn@508 | 606 | can_eliminate = false; |
kvn@508 | 607 | } |
kvn@508 | 608 | } |
kvn@508 | 609 | } else if (use->is_SafePoint()) { |
kvn@508 | 610 | SafePointNode* sfpt = use->as_SafePoint(); |
kvn@603 | 611 | if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) { |
kvn@508 | 612 | // Object is passed as argument. |
kvn@508 | 613 | DEBUG_ONLY(disq_node = use;) |
kvn@508 | 614 | NOT_PRODUCT(fail_eliminate = "Object is passed as argument";) |
kvn@508 | 615 | can_eliminate = false; |
kvn@508 | 616 | } |
kvn@508 | 617 | Node* sfptMem = sfpt->memory(); |
kvn@508 | 618 | if (sfptMem == NULL || sfptMem->is_top()) { |
kvn@508 | 619 | DEBUG_ONLY(disq_node = use;) |
kvn@508 | 620 | NOT_PRODUCT(fail_eliminate = "NULL or TOP memory";) |
kvn@508 | 621 | can_eliminate = false; |
kvn@508 | 622 | } else { |
kvn@508 | 623 | safepoints.append_if_missing(sfpt); |
kvn@508 | 624 | } |
kvn@508 | 625 | } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark |
kvn@508 | 626 | if (use->is_Phi()) { |
kvn@508 | 627 | if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) { |
kvn@508 | 628 | NOT_PRODUCT(fail_eliminate = "Object is return value";) |
kvn@508 | 629 | } else { |
kvn@508 | 630 | NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";) |
kvn@508 | 631 | } |
kvn@508 | 632 | DEBUG_ONLY(disq_node = use;) |
kvn@508 | 633 | } else { |
kvn@508 | 634 | if (use->Opcode() == Op_Return) { |
kvn@508 | 635 | NOT_PRODUCT(fail_eliminate = "Object is return value";) |
kvn@508 | 636 | }else { |
kvn@508 | 637 | NOT_PRODUCT(fail_eliminate = "Object is referenced by node";) |
kvn@508 | 638 | } |
kvn@508 | 639 | DEBUG_ONLY(disq_node = use;) |
kvn@508 | 640 | } |
kvn@508 | 641 | can_eliminate = false; |
kvn@508 | 642 | } |
kvn@508 | 643 | } |
kvn@508 | 644 | } |
kvn@508 | 645 | |
kvn@508 | 646 | #ifndef PRODUCT |
kvn@508 | 647 | if (PrintEliminateAllocations) { |
kvn@508 | 648 | if (can_eliminate) { |
kvn@508 | 649 | tty->print("Scalar "); |
kvn@508 | 650 | if (res == NULL) |
kvn@508 | 651 | alloc->dump(); |
kvn@508 | 652 | else |
kvn@508 | 653 | res->dump(); |
kvn@508 | 654 | } else { |
kvn@508 | 655 | tty->print("NotScalar (%s)", fail_eliminate); |
kvn@508 | 656 | if (res == NULL) |
kvn@508 | 657 | alloc->dump(); |
kvn@508 | 658 | else |
kvn@508 | 659 | res->dump(); |
kvn@508 | 660 | #ifdef ASSERT |
kvn@508 | 661 | if (disq_node != NULL) { |
kvn@508 | 662 | tty->print(" >>>> "); |
kvn@508 | 663 | disq_node->dump(); |
kvn@508 | 664 | } |
kvn@508 | 665 | #endif /*ASSERT*/ |
kvn@508 | 666 | } |
kvn@508 | 667 | } |
kvn@508 | 668 | #endif |
kvn@508 | 669 | return can_eliminate; |
kvn@508 | 670 | } |
kvn@508 | 671 | |
kvn@508 | 672 | // Do scalar replacement. |
kvn@508 | 673 | bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) { |
kvn@508 | 674 | GrowableArray <SafePointNode *> safepoints_done; |
kvn@508 | 675 | |
kvn@508 | 676 | ciKlass* klass = NULL; |
kvn@508 | 677 | ciInstanceKlass* iklass = NULL; |
kvn@508 | 678 | int nfields = 0; |
kvn@508 | 679 | int array_base; |
kvn@508 | 680 | int element_size; |
kvn@508 | 681 | BasicType basic_elem_type; |
kvn@508 | 682 | ciType* elem_type; |
kvn@508 | 683 | |
kvn@508 | 684 | Node* res = alloc->result_cast(); |
kvn@508 | 685 | const TypeOopPtr* res_type = NULL; |
kvn@508 | 686 | if (res != NULL) { // Could be NULL when there are no users |
kvn@508 | 687 | res_type = _igvn.type(res)->isa_oopptr(); |
kvn@508 | 688 | } |
kvn@508 | 689 | |
kvn@508 | 690 | if (res != NULL) { |
kvn@508 | 691 | klass = res_type->klass(); |
kvn@508 | 692 | if (res_type->isa_instptr()) { |
kvn@508 | 693 | // find the fields of the class which will be needed for safepoint debug information |
kvn@508 | 694 | assert(klass->is_instance_klass(), "must be an instance klass."); |
kvn@508 | 695 | iklass = klass->as_instance_klass(); |
kvn@508 | 696 | nfields = iklass->nof_nonstatic_fields(); |
kvn@508 | 697 | } else { |
kvn@508 | 698 | // find the array's elements which will be needed for safepoint debug information |
kvn@508 | 699 | nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1); |
kvn@508 | 700 | assert(klass->is_array_klass() && nfields >= 0, "must be an array klass."); |
kvn@508 | 701 | elem_type = klass->as_array_klass()->element_type(); |
kvn@508 | 702 | basic_elem_type = elem_type->basic_type(); |
kvn@508 | 703 | array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type); |
kvn@508 | 704 | element_size = type2aelembytes(basic_elem_type); |
kvn@508 | 705 | } |
kvn@508 | 706 | } |
kvn@508 | 707 | // |
kvn@508 | 708 | // Process the safepoint uses |
kvn@508 | 709 | // |
kvn@508 | 710 | while (safepoints.length() > 0) { |
kvn@508 | 711 | SafePointNode* sfpt = safepoints.pop(); |
kvn@508 | 712 | Node* mem = sfpt->memory(); |
kvn@508 | 713 | uint first_ind = sfpt->req(); |
kvn@508 | 714 | SafePointScalarObjectNode* sobj = new (C, 1) SafePointScalarObjectNode(res_type, |
kvn@508 | 715 | #ifdef ASSERT |
kvn@508 | 716 | alloc, |
kvn@508 | 717 | #endif |
kvn@508 | 718 | first_ind, nfields); |
kvn@508 | 719 | sobj->init_req(0, sfpt->in(TypeFunc::Control)); |
kvn@508 | 720 | transform_later(sobj); |
kvn@508 | 721 | |
kvn@508 | 722 | // Scan object's fields adding an input to the safepoint for each field. |
kvn@508 | 723 | for (int j = 0; j < nfields; j++) { |
kvn@741 | 724 | intptr_t offset; |
kvn@508 | 725 | ciField* field = NULL; |
kvn@508 | 726 | if (iklass != NULL) { |
kvn@508 | 727 | field = iklass->nonstatic_field_at(j); |
kvn@508 | 728 | offset = field->offset(); |
kvn@508 | 729 | elem_type = field->type(); |
kvn@508 | 730 | basic_elem_type = field->layout_type(); |
kvn@508 | 731 | } else { |
kvn@741 | 732 | offset = array_base + j * (intptr_t)element_size; |
kvn@508 | 733 | } |
kvn@508 | 734 | |
kvn@508 | 735 | const Type *field_type; |
kvn@508 | 736 | // The next code is taken from Parse::do_get_xxx(). |
kvn@559 | 737 | if (basic_elem_type == T_OBJECT || basic_elem_type == T_ARRAY) { |
kvn@508 | 738 | if (!elem_type->is_loaded()) { |
kvn@508 | 739 | field_type = TypeInstPtr::BOTTOM; |
kvn@2037 | 740 | } else if (field != NULL && field->is_constant() && field->is_static()) { |
kvn@508 | 741 | // This can happen if the constant oop is non-perm. |
kvn@508 | 742 | ciObject* con = field->constant_value().as_object(); |
kvn@508 | 743 | // Do not "join" in the previous type; it doesn't add value, |
kvn@508 | 744 | // and may yield a vacuous result if the field is of interface type. |
kvn@508 | 745 | field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr(); |
kvn@508 | 746 | assert(field_type != NULL, "field singleton type must be consistent"); |
kvn@508 | 747 | } else { |
kvn@508 | 748 | field_type = TypeOopPtr::make_from_klass(elem_type->as_klass()); |
kvn@508 | 749 | } |
kvn@559 | 750 | if (UseCompressedOops) { |
kvn@656 | 751 | field_type = field_type->make_narrowoop(); |
kvn@559 | 752 | basic_elem_type = T_NARROWOOP; |
kvn@559 | 753 | } |
kvn@508 | 754 | } else { |
kvn@508 | 755 | field_type = Type::get_const_basic_type(basic_elem_type); |
kvn@508 | 756 | } |
kvn@508 | 757 | |
kvn@508 | 758 | const TypeOopPtr *field_addr_type = res_type->add_offset(offset)->isa_oopptr(); |
kvn@508 | 759 | |
kvn@508 | 760 | Node *field_val = value_from_mem(mem, basic_elem_type, field_type, field_addr_type, alloc); |
kvn@508 | 761 | if (field_val == NULL) { |
kvn@508 | 762 | // we weren't able to find a value for this field, |
kvn@508 | 763 | // give up on eliminating this allocation |
kvn@508 | 764 | alloc->_is_scalar_replaceable = false; // don't try again |
kvn@508 | 765 | // remove any extra entries we added to the safepoint |
kvn@508 | 766 | uint last = sfpt->req() - 1; |
kvn@508 | 767 | for (int k = 0; k < j; k++) { |
kvn@508 | 768 | sfpt->del_req(last--); |
kvn@508 | 769 | } |
kvn@508 | 770 | // rollback processed safepoints |
kvn@508 | 771 | while (safepoints_done.length() > 0) { |
kvn@508 | 772 | SafePointNode* sfpt_done = safepoints_done.pop(); |
kvn@508 | 773 | // remove any extra entries we added to the safepoint |
kvn@508 | 774 | last = sfpt_done->req() - 1; |
kvn@508 | 775 | for (int k = 0; k < nfields; k++) { |
kvn@508 | 776 | sfpt_done->del_req(last--); |
kvn@508 | 777 | } |
kvn@508 | 778 | JVMState *jvms = sfpt_done->jvms(); |
kvn@508 | 779 | jvms->set_endoff(sfpt_done->req()); |
kvn@508 | 780 | // Now make a pass over the debug information replacing any references |
kvn@508 | 781 | // to SafePointScalarObjectNode with the allocated object. |
kvn@508 | 782 | int start = jvms->debug_start(); |
kvn@508 | 783 | int end = jvms->debug_end(); |
kvn@508 | 784 | for (int i = start; i < end; i++) { |
kvn@508 | 785 | if (sfpt_done->in(i)->is_SafePointScalarObject()) { |
kvn@508 | 786 | SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject(); |
kvn@508 | 787 | if (scobj->first_index() == sfpt_done->req() && |
kvn@508 | 788 | scobj->n_fields() == (uint)nfields) { |
kvn@508 | 789 | assert(scobj->alloc() == alloc, "sanity"); |
kvn@508 | 790 | sfpt_done->set_req(i, res); |
kvn@508 | 791 | } |
kvn@508 | 792 | } |
kvn@508 | 793 | } |
kvn@508 | 794 | } |
kvn@508 | 795 | #ifndef PRODUCT |
kvn@508 | 796 | if (PrintEliminateAllocations) { |
kvn@508 | 797 | if (field != NULL) { |
kvn@508 | 798 | tty->print("=== At SafePoint node %d can't find value of Field: ", |
kvn@508 | 799 | sfpt->_idx); |
kvn@508 | 800 | field->print(); |
kvn@508 | 801 | int field_idx = C->get_alias_index(field_addr_type); |
kvn@508 | 802 | tty->print(" (alias_idx=%d)", field_idx); |
kvn@508 | 803 | } else { // Array's element |
kvn@508 | 804 | tty->print("=== At SafePoint node %d can't find value of array element [%d]", |
kvn@508 | 805 | sfpt->_idx, j); |
kvn@508 | 806 | } |
kvn@508 | 807 | tty->print(", which prevents elimination of: "); |
kvn@508 | 808 | if (res == NULL) |
kvn@508 | 809 | alloc->dump(); |
kvn@508 | 810 | else |
kvn@508 | 811 | res->dump(); |
kvn@508 | 812 | } |
kvn@508 | 813 | #endif |
kvn@508 | 814 | return false; |
kvn@508 | 815 | } |
kvn@559 | 816 | if (UseCompressedOops && field_type->isa_narrowoop()) { |
kvn@559 | 817 | // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation |
kvn@559 | 818 | // to be able scalar replace the allocation. |
kvn@656 | 819 | if (field_val->is_EncodeP()) { |
kvn@656 | 820 | field_val = field_val->in(1); |
kvn@656 | 821 | } else { |
kvn@656 | 822 | field_val = transform_later(new (C, 2) DecodeNNode(field_val, field_val->bottom_type()->make_ptr())); |
kvn@656 | 823 | } |
kvn@559 | 824 | } |
kvn@508 | 825 | sfpt->add_req(field_val); |
kvn@508 | 826 | } |
kvn@508 | 827 | JVMState *jvms = sfpt->jvms(); |
kvn@508 | 828 | jvms->set_endoff(sfpt->req()); |
kvn@508 | 829 | // Now make a pass over the debug information replacing any references |
kvn@508 | 830 | // to the allocated object with "sobj" |
kvn@508 | 831 | int start = jvms->debug_start(); |
kvn@508 | 832 | int end = jvms->debug_end(); |
kvn@508 | 833 | for (int i = start; i < end; i++) { |
kvn@508 | 834 | if (sfpt->in(i) == res) { |
kvn@508 | 835 | sfpt->set_req(i, sobj); |
kvn@508 | 836 | } |
kvn@508 | 837 | } |
kvn@508 | 838 | safepoints_done.append_if_missing(sfpt); // keep it for rollback |
kvn@508 | 839 | } |
kvn@508 | 840 | return true; |
kvn@508 | 841 | } |
kvn@508 | 842 | |
kvn@508 | 843 | // Process users of eliminated allocation. |
kvn@508 | 844 | void PhaseMacroExpand::process_users_of_allocation(AllocateNode *alloc) { |
kvn@508 | 845 | Node* res = alloc->result_cast(); |
kvn@508 | 846 | if (res != NULL) { |
kvn@508 | 847 | for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) { |
kvn@508 | 848 | Node *use = res->last_out(j); |
kvn@508 | 849 | uint oc1 = res->outcnt(); |
kvn@508 | 850 | |
kvn@508 | 851 | if (use->is_AddP()) { |
kvn@508 | 852 | for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) { |
kvn@508 | 853 | Node *n = use->last_out(k); |
kvn@508 | 854 | uint oc2 = use->outcnt(); |
kvn@508 | 855 | if (n->is_Store()) { |
kvn@1535 | 856 | #ifdef ASSERT |
kvn@1535 | 857 | // Verify that there is no dependent MemBarVolatile nodes, |
kvn@1535 | 858 | // they should be removed during IGVN, see MemBarNode::Ideal(). |
kvn@1535 | 859 | for (DUIterator_Fast pmax, p = n->fast_outs(pmax); |
kvn@1535 | 860 | p < pmax; p++) { |
kvn@1535 | 861 | Node* mb = n->fast_out(p); |
kvn@1535 | 862 | assert(mb->is_Initialize() || !mb->is_MemBar() || |
kvn@1535 | 863 | mb->req() <= MemBarNode::Precedent || |
kvn@1535 | 864 | mb->in(MemBarNode::Precedent) != n, |
kvn@1535 | 865 | "MemBarVolatile should be eliminated for non-escaping object"); |
kvn@1535 | 866 | } |
kvn@1535 | 867 | #endif |
kvn@508 | 868 | _igvn.replace_node(n, n->in(MemNode::Memory)); |
kvn@508 | 869 | } else { |
kvn@508 | 870 | eliminate_card_mark(n); |
kvn@508 | 871 | } |
kvn@508 | 872 | k -= (oc2 - use->outcnt()); |
kvn@508 | 873 | } |
kvn@508 | 874 | } else { |
kvn@508 | 875 | eliminate_card_mark(use); |
kvn@508 | 876 | } |
kvn@508 | 877 | j -= (oc1 - res->outcnt()); |
kvn@508 | 878 | } |
kvn@508 | 879 | assert(res->outcnt() == 0, "all uses of allocated objects must be deleted"); |
kvn@508 | 880 | _igvn.remove_dead_node(res); |
kvn@508 | 881 | } |
kvn@508 | 882 | |
kvn@508 | 883 | // |
kvn@508 | 884 | // Process other users of allocation's projections |
kvn@508 | 885 | // |
kvn@508 | 886 | if (_resproj != NULL && _resproj->outcnt() != 0) { |
kvn@508 | 887 | for (DUIterator_Last jmin, j = _resproj->last_outs(jmin); j >= jmin; ) { |
kvn@508 | 888 | Node *use = _resproj->last_out(j); |
kvn@508 | 889 | uint oc1 = _resproj->outcnt(); |
kvn@508 | 890 | if (use->is_Initialize()) { |
kvn@508 | 891 | // Eliminate Initialize node. |
kvn@508 | 892 | InitializeNode *init = use->as_Initialize(); |
kvn@508 | 893 | assert(init->outcnt() <= 2, "only a control and memory projection expected"); |
kvn@508 | 894 | Node *ctrl_proj = init->proj_out(TypeFunc::Control); |
kvn@508 | 895 | if (ctrl_proj != NULL) { |
kvn@508 | 896 | assert(init->in(TypeFunc::Control) == _fallthroughcatchproj, "allocation control projection"); |
kvn@508 | 897 | _igvn.replace_node(ctrl_proj, _fallthroughcatchproj); |
kvn@508 | 898 | } |
kvn@508 | 899 | Node *mem_proj = init->proj_out(TypeFunc::Memory); |
kvn@508 | 900 | if (mem_proj != NULL) { |
kvn@508 | 901 | Node *mem = init->in(TypeFunc::Memory); |
kvn@508 | 902 | #ifdef ASSERT |
kvn@508 | 903 | if (mem->is_MergeMem()) { |
kvn@508 | 904 | assert(mem->in(TypeFunc::Memory) == _memproj_fallthrough, "allocation memory projection"); |
kvn@508 | 905 | } else { |
kvn@508 | 906 | assert(mem == _memproj_fallthrough, "allocation memory projection"); |
kvn@508 | 907 | } |
kvn@508 | 908 | #endif |
kvn@508 | 909 | _igvn.replace_node(mem_proj, mem); |
kvn@508 | 910 | } |
kvn@508 | 911 | } else if (use->is_AddP()) { |
kvn@508 | 912 | // raw memory addresses used only by the initialization |
kvn@1143 | 913 | _igvn.replace_node(use, C->top()); |
kvn@508 | 914 | } else { |
kvn@508 | 915 | assert(false, "only Initialize or AddP expected"); |
kvn@508 | 916 | } |
kvn@508 | 917 | j -= (oc1 - _resproj->outcnt()); |
kvn@508 | 918 | } |
kvn@508 | 919 | } |
kvn@508 | 920 | if (_fallthroughcatchproj != NULL) { |
kvn@508 | 921 | _igvn.replace_node(_fallthroughcatchproj, alloc->in(TypeFunc::Control)); |
kvn@508 | 922 | } |
kvn@508 | 923 | if (_memproj_fallthrough != NULL) { |
kvn@508 | 924 | _igvn.replace_node(_memproj_fallthrough, alloc->in(TypeFunc::Memory)); |
kvn@508 | 925 | } |
kvn@508 | 926 | if (_memproj_catchall != NULL) { |
kvn@508 | 927 | _igvn.replace_node(_memproj_catchall, C->top()); |
kvn@508 | 928 | } |
kvn@508 | 929 | if (_ioproj_fallthrough != NULL) { |
kvn@508 | 930 | _igvn.replace_node(_ioproj_fallthrough, alloc->in(TypeFunc::I_O)); |
kvn@508 | 931 | } |
kvn@508 | 932 | if (_ioproj_catchall != NULL) { |
kvn@508 | 933 | _igvn.replace_node(_ioproj_catchall, C->top()); |
kvn@508 | 934 | } |
kvn@508 | 935 | if (_catchallcatchproj != NULL) { |
kvn@508 | 936 | _igvn.replace_node(_catchallcatchproj, C->top()); |
kvn@508 | 937 | } |
kvn@508 | 938 | } |
kvn@508 | 939 | |
kvn@508 | 940 | bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) { |
kvn@508 | 941 | |
kvn@508 | 942 | if (!EliminateAllocations || !alloc->_is_scalar_replaceable) { |
kvn@508 | 943 | return false; |
kvn@508 | 944 | } |
kvn@508 | 945 | |
kvn@508 | 946 | extract_call_projections(alloc); |
kvn@508 | 947 | |
kvn@508 | 948 | GrowableArray <SafePointNode *> safepoints; |
kvn@508 | 949 | if (!can_eliminate_allocation(alloc, safepoints)) { |
kvn@508 | 950 | return false; |
kvn@508 | 951 | } |
kvn@508 | 952 | |
kvn@508 | 953 | if (!scalar_replacement(alloc, safepoints)) { |
kvn@508 | 954 | return false; |
kvn@508 | 955 | } |
kvn@508 | 956 | |
never@1515 | 957 | CompileLog* log = C->log(); |
never@1515 | 958 | if (log != NULL) { |
never@1515 | 959 | Node* klass = alloc->in(AllocateNode::KlassNode); |
never@1515 | 960 | const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr(); |
never@1515 | 961 | log->head("eliminate_allocation type='%d'", |
never@1515 | 962 | log->identify(tklass->klass())); |
never@1515 | 963 | JVMState* p = alloc->jvms(); |
never@1515 | 964 | while (p != NULL) { |
never@1515 | 965 | log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method())); |
never@1515 | 966 | p = p->caller(); |
never@1515 | 967 | } |
never@1515 | 968 | log->tail("eliminate_allocation"); |
never@1515 | 969 | } |
never@1515 | 970 | |
kvn@508 | 971 | process_users_of_allocation(alloc); |
kvn@508 | 972 | |
kvn@508 | 973 | #ifndef PRODUCT |
never@1515 | 974 | if (PrintEliminateAllocations) { |
never@1515 | 975 | if (alloc->is_AllocateArray()) |
never@1515 | 976 | tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx); |
never@1515 | 977 | else |
never@1515 | 978 | tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx); |
never@1515 | 979 | } |
kvn@508 | 980 | #endif |
kvn@508 | 981 | |
kvn@508 | 982 | return true; |
kvn@508 | 983 | } |
kvn@508 | 984 | |
duke@435 | 985 | |
duke@435 | 986 | //---------------------------set_eden_pointers------------------------- |
duke@435 | 987 | void PhaseMacroExpand::set_eden_pointers(Node* &eden_top_adr, Node* &eden_end_adr) { |
duke@435 | 988 | if (UseTLAB) { // Private allocation: load from TLS |
duke@435 | 989 | Node* thread = transform_later(new (C, 1) ThreadLocalNode()); |
duke@435 | 990 | int tlab_top_offset = in_bytes(JavaThread::tlab_top_offset()); |
duke@435 | 991 | int tlab_end_offset = in_bytes(JavaThread::tlab_end_offset()); |
duke@435 | 992 | eden_top_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_top_offset); |
duke@435 | 993 | eden_end_adr = basic_plus_adr(top()/*not oop*/, thread, tlab_end_offset); |
duke@435 | 994 | } else { // Shared allocation: load from globals |
duke@435 | 995 | CollectedHeap* ch = Universe::heap(); |
duke@435 | 996 | address top_adr = (address)ch->top_addr(); |
duke@435 | 997 | address end_adr = (address)ch->end_addr(); |
duke@435 | 998 | eden_top_adr = makecon(TypeRawPtr::make(top_adr)); |
duke@435 | 999 | eden_end_adr = basic_plus_adr(eden_top_adr, end_adr - top_adr); |
duke@435 | 1000 | } |
duke@435 | 1001 | } |
duke@435 | 1002 | |
duke@435 | 1003 | |
duke@435 | 1004 | Node* PhaseMacroExpand::make_load(Node* ctl, Node* mem, Node* base, int offset, const Type* value_type, BasicType bt) { |
duke@435 | 1005 | Node* adr = basic_plus_adr(base, offset); |
kvn@855 | 1006 | const TypePtr* adr_type = adr->bottom_type()->is_ptr(); |
coleenp@548 | 1007 | Node* value = LoadNode::make(_igvn, ctl, mem, adr, adr_type, value_type, bt); |
duke@435 | 1008 | transform_later(value); |
duke@435 | 1009 | return value; |
duke@435 | 1010 | } |
duke@435 | 1011 | |
duke@435 | 1012 | |
duke@435 | 1013 | Node* PhaseMacroExpand::make_store(Node* ctl, Node* mem, Node* base, int offset, Node* value, BasicType bt) { |
duke@435 | 1014 | Node* adr = basic_plus_adr(base, offset); |
coleenp@548 | 1015 | mem = StoreNode::make(_igvn, ctl, mem, adr, NULL, value, bt); |
duke@435 | 1016 | transform_later(mem); |
duke@435 | 1017 | return mem; |
duke@435 | 1018 | } |
duke@435 | 1019 | |
duke@435 | 1020 | //============================================================================= |
duke@435 | 1021 | // |
duke@435 | 1022 | // A L L O C A T I O N |
duke@435 | 1023 | // |
duke@435 | 1024 | // Allocation attempts to be fast in the case of frequent small objects. |
duke@435 | 1025 | // It breaks down like this: |
duke@435 | 1026 | // |
duke@435 | 1027 | // 1) Size in doublewords is computed. This is a constant for objects and |
duke@435 | 1028 | // variable for most arrays. Doubleword units are used to avoid size |
duke@435 | 1029 | // overflow of huge doubleword arrays. We need doublewords in the end for |
duke@435 | 1030 | // rounding. |
duke@435 | 1031 | // |
duke@435 | 1032 | // 2) Size is checked for being 'too large'. Too-large allocations will go |
duke@435 | 1033 | // the slow path into the VM. The slow path can throw any required |
duke@435 | 1034 | // exceptions, and does all the special checks for very large arrays. The |
duke@435 | 1035 | // size test can constant-fold away for objects. For objects with |
duke@435 | 1036 | // finalizers it constant-folds the otherway: you always go slow with |
duke@435 | 1037 | // finalizers. |
duke@435 | 1038 | // |
duke@435 | 1039 | // 3) If NOT using TLABs, this is the contended loop-back point. |
duke@435 | 1040 | // Load-Locked the heap top. If using TLABs normal-load the heap top. |
duke@435 | 1041 | // |
duke@435 | 1042 | // 4) Check that heap top + size*8 < max. If we fail go the slow ` route. |
duke@435 | 1043 | // NOTE: "top+size*8" cannot wrap the 4Gig line! Here's why: for largish |
duke@435 | 1044 | // "size*8" we always enter the VM, where "largish" is a constant picked small |
duke@435 | 1045 | // enough that there's always space between the eden max and 4Gig (old space is |
duke@435 | 1046 | // there so it's quite large) and large enough that the cost of entering the VM |
duke@435 | 1047 | // is dwarfed by the cost to initialize the space. |
duke@435 | 1048 | // |
duke@435 | 1049 | // 5) If NOT using TLABs, Store-Conditional the adjusted heap top back |
duke@435 | 1050 | // down. If contended, repeat at step 3. If using TLABs normal-store |
duke@435 | 1051 | // adjusted heap top back down; there is no contention. |
duke@435 | 1052 | // |
duke@435 | 1053 | // 6) If !ZeroTLAB then Bulk-clear the object/array. Fill in klass & mark |
duke@435 | 1054 | // fields. |
duke@435 | 1055 | // |
duke@435 | 1056 | // 7) Merge with the slow-path; cast the raw memory pointer to the correct |
duke@435 | 1057 | // oop flavor. |
duke@435 | 1058 | // |
duke@435 | 1059 | //============================================================================= |
duke@435 | 1060 | // FastAllocateSizeLimit value is in DOUBLEWORDS. |
duke@435 | 1061 | // Allocations bigger than this always go the slow route. |
duke@435 | 1062 | // This value must be small enough that allocation attempts that need to |
duke@435 | 1063 | // trigger exceptions go the slow route. Also, it must be small enough so |
duke@435 | 1064 | // that heap_top + size_in_bytes does not wrap around the 4Gig limit. |
duke@435 | 1065 | //=============================================================================j// |
duke@435 | 1066 | // %%% Here is an old comment from parseHelper.cpp; is it outdated? |
duke@435 | 1067 | // The allocator will coalesce int->oop copies away. See comment in |
duke@435 | 1068 | // coalesce.cpp about how this works. It depends critically on the exact |
duke@435 | 1069 | // code shape produced here, so if you are changing this code shape |
duke@435 | 1070 | // make sure the GC info for the heap-top is correct in and around the |
duke@435 | 1071 | // slow-path call. |
duke@435 | 1072 | // |
duke@435 | 1073 | |
duke@435 | 1074 | void PhaseMacroExpand::expand_allocate_common( |
duke@435 | 1075 | AllocateNode* alloc, // allocation node to be expanded |
duke@435 | 1076 | Node* length, // array length for an array allocation |
duke@435 | 1077 | const TypeFunc* slow_call_type, // Type of slow call |
duke@435 | 1078 | address slow_call_address // Address of slow call |
duke@435 | 1079 | ) |
duke@435 | 1080 | { |
duke@435 | 1081 | |
duke@435 | 1082 | Node* ctrl = alloc->in(TypeFunc::Control); |
duke@435 | 1083 | Node* mem = alloc->in(TypeFunc::Memory); |
duke@435 | 1084 | Node* i_o = alloc->in(TypeFunc::I_O); |
duke@435 | 1085 | Node* size_in_bytes = alloc->in(AllocateNode::AllocSize); |
duke@435 | 1086 | Node* klass_node = alloc->in(AllocateNode::KlassNode); |
duke@435 | 1087 | Node* initial_slow_test = alloc->in(AllocateNode::InitialTest); |
duke@435 | 1088 | |
duke@435 | 1089 | assert(ctrl != NULL, "must have control"); |
duke@435 | 1090 | // We need a Region and corresponding Phi's to merge the slow-path and fast-path results. |
duke@435 | 1091 | // they will not be used if "always_slow" is set |
duke@435 | 1092 | enum { slow_result_path = 1, fast_result_path = 2 }; |
duke@435 | 1093 | Node *result_region; |
duke@435 | 1094 | Node *result_phi_rawmem; |
duke@435 | 1095 | Node *result_phi_rawoop; |
duke@435 | 1096 | Node *result_phi_i_o; |
duke@435 | 1097 | |
duke@435 | 1098 | // The initial slow comparison is a size check, the comparison |
duke@435 | 1099 | // we want to do is a BoolTest::gt |
duke@435 | 1100 | bool always_slow = false; |
duke@435 | 1101 | int tv = _igvn.find_int_con(initial_slow_test, -1); |
duke@435 | 1102 | if (tv >= 0) { |
duke@435 | 1103 | always_slow = (tv == 1); |
duke@435 | 1104 | initial_slow_test = NULL; |
duke@435 | 1105 | } else { |
duke@435 | 1106 | initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn); |
duke@435 | 1107 | } |
duke@435 | 1108 | |
kvn@1215 | 1109 | if (C->env()->dtrace_alloc_probes() || |
ysr@777 | 1110 | !UseTLAB && (!Universe::heap()->supports_inline_contig_alloc() || |
ysr@777 | 1111 | (UseConcMarkSweepGC && CMSIncrementalMode))) { |
duke@435 | 1112 | // Force slow-path allocation |
duke@435 | 1113 | always_slow = true; |
duke@435 | 1114 | initial_slow_test = NULL; |
duke@435 | 1115 | } |
duke@435 | 1116 | |
ysr@777 | 1117 | |
duke@435 | 1118 | enum { too_big_or_final_path = 1, need_gc_path = 2 }; |
duke@435 | 1119 | Node *slow_region = NULL; |
duke@435 | 1120 | Node *toobig_false = ctrl; |
duke@435 | 1121 | |
duke@435 | 1122 | assert (initial_slow_test == NULL || !always_slow, "arguments must be consistent"); |
duke@435 | 1123 | // generate the initial test if necessary |
duke@435 | 1124 | if (initial_slow_test != NULL ) { |
duke@435 | 1125 | slow_region = new (C, 3) RegionNode(3); |
duke@435 | 1126 | |
duke@435 | 1127 | // Now make the initial failure test. Usually a too-big test but |
duke@435 | 1128 | // might be a TRUE for finalizers or a fancy class check for |
duke@435 | 1129 | // newInstance0. |
duke@435 | 1130 | IfNode *toobig_iff = new (C, 2) IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN); |
duke@435 | 1131 | transform_later(toobig_iff); |
duke@435 | 1132 | // Plug the failing-too-big test into the slow-path region |
duke@435 | 1133 | Node *toobig_true = new (C, 1) IfTrueNode( toobig_iff ); |
duke@435 | 1134 | transform_later(toobig_true); |
duke@435 | 1135 | slow_region ->init_req( too_big_or_final_path, toobig_true ); |
duke@435 | 1136 | toobig_false = new (C, 1) IfFalseNode( toobig_iff ); |
duke@435 | 1137 | transform_later(toobig_false); |
duke@435 | 1138 | } else { // No initial test, just fall into next case |
duke@435 | 1139 | toobig_false = ctrl; |
duke@435 | 1140 | debug_only(slow_region = NodeSentinel); |
duke@435 | 1141 | } |
duke@435 | 1142 | |
duke@435 | 1143 | Node *slow_mem = mem; // save the current memory state for slow path |
duke@435 | 1144 | // generate the fast allocation code unless we know that the initial test will always go slow |
duke@435 | 1145 | if (!always_slow) { |
kvn@1000 | 1146 | // Fast path modifies only raw memory. |
kvn@1000 | 1147 | if (mem->is_MergeMem()) { |
kvn@1000 | 1148 | mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw); |
kvn@1000 | 1149 | } |
kvn@1000 | 1150 | |
ysr@777 | 1151 | Node* eden_top_adr; |
ysr@777 | 1152 | Node* eden_end_adr; |
ysr@777 | 1153 | |
ysr@777 | 1154 | set_eden_pointers(eden_top_adr, eden_end_adr); |
ysr@777 | 1155 | |
ysr@777 | 1156 | // Load Eden::end. Loop invariant and hoisted. |
ysr@777 | 1157 | // |
ysr@777 | 1158 | // Note: We set the control input on "eden_end" and "old_eden_top" when using |
ysr@777 | 1159 | // a TLAB to work around a bug where these values were being moved across |
ysr@777 | 1160 | // a safepoint. These are not oops, so they cannot be include in the oop |
phh@2423 | 1161 | // map, but they can be changed by a GC. The proper way to fix this would |
ysr@777 | 1162 | // be to set the raw memory state when generating a SafepointNode. However |
ysr@777 | 1163 | // this will require extensive changes to the loop optimization in order to |
ysr@777 | 1164 | // prevent a degradation of the optimization. |
ysr@777 | 1165 | // See comment in memnode.hpp, around line 227 in class LoadPNode. |
ysr@777 | 1166 | Node *eden_end = make_load(ctrl, mem, eden_end_adr, 0, TypeRawPtr::BOTTOM, T_ADDRESS); |
ysr@777 | 1167 | |
duke@435 | 1168 | // allocate the Region and Phi nodes for the result |
duke@435 | 1169 | result_region = new (C, 3) RegionNode(3); |
phh@2423 | 1170 | result_phi_rawmem = new (C, 3) PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM); |
phh@2423 | 1171 | result_phi_rawoop = new (C, 3) PhiNode(result_region, TypeRawPtr::BOTTOM); |
phh@2423 | 1172 | result_phi_i_o = new (C, 3) PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch |
duke@435 | 1173 | |
duke@435 | 1174 | // We need a Region for the loop-back contended case. |
duke@435 | 1175 | enum { fall_in_path = 1, contended_loopback_path = 2 }; |
duke@435 | 1176 | Node *contended_region; |
duke@435 | 1177 | Node *contended_phi_rawmem; |
phh@2423 | 1178 | if (UseTLAB) { |
duke@435 | 1179 | contended_region = toobig_false; |
duke@435 | 1180 | contended_phi_rawmem = mem; |
duke@435 | 1181 | } else { |
duke@435 | 1182 | contended_region = new (C, 3) RegionNode(3); |
phh@2423 | 1183 | contended_phi_rawmem = new (C, 3) PhiNode(contended_region, Type::MEMORY, TypeRawPtr::BOTTOM); |
duke@435 | 1184 | // Now handle the passing-too-big test. We fall into the contended |
duke@435 | 1185 | // loop-back merge point. |
phh@2423 | 1186 | contended_region ->init_req(fall_in_path, toobig_false); |
phh@2423 | 1187 | contended_phi_rawmem->init_req(fall_in_path, mem); |
duke@435 | 1188 | transform_later(contended_region); |
duke@435 | 1189 | transform_later(contended_phi_rawmem); |
duke@435 | 1190 | } |
duke@435 | 1191 | |
duke@435 | 1192 | // Load(-locked) the heap top. |
duke@435 | 1193 | // See note above concerning the control input when using a TLAB |
duke@435 | 1194 | Node *old_eden_top = UseTLAB |
phh@2423 | 1195 | ? new (C, 3) LoadPNode (ctrl, contended_phi_rawmem, eden_top_adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM) |
phh@2423 | 1196 | : new (C, 3) LoadPLockedNode(contended_region, contended_phi_rawmem, eden_top_adr); |
duke@435 | 1197 | |
duke@435 | 1198 | transform_later(old_eden_top); |
duke@435 | 1199 | // Add to heap top to get a new heap top |
phh@2423 | 1200 | Node *new_eden_top = new (C, 4) AddPNode(top(), old_eden_top, size_in_bytes); |
duke@435 | 1201 | transform_later(new_eden_top); |
duke@435 | 1202 | // Check for needing a GC; compare against heap end |
phh@2423 | 1203 | Node *needgc_cmp = new (C, 3) CmpPNode(new_eden_top, eden_end); |
duke@435 | 1204 | transform_later(needgc_cmp); |
phh@2423 | 1205 | Node *needgc_bol = new (C, 2) BoolNode(needgc_cmp, BoolTest::ge); |
duke@435 | 1206 | transform_later(needgc_bol); |
phh@2423 | 1207 | IfNode *needgc_iff = new (C, 2) IfNode(contended_region, needgc_bol, PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN); |
duke@435 | 1208 | transform_later(needgc_iff); |
duke@435 | 1209 | |
duke@435 | 1210 | // Plug the failing-heap-space-need-gc test into the slow-path region |
phh@2423 | 1211 | Node *needgc_true = new (C, 1) IfTrueNode(needgc_iff); |
duke@435 | 1212 | transform_later(needgc_true); |
phh@2423 | 1213 | if (initial_slow_test) { |
phh@2423 | 1214 | slow_region->init_req(need_gc_path, needgc_true); |
duke@435 | 1215 | // This completes all paths into the slow merge point |
duke@435 | 1216 | transform_later(slow_region); |
duke@435 | 1217 | } else { // No initial slow path needed! |
duke@435 | 1218 | // Just fall from the need-GC path straight into the VM call. |
phh@2423 | 1219 | slow_region = needgc_true; |
duke@435 | 1220 | } |
duke@435 | 1221 | // No need for a GC. Setup for the Store-Conditional |
phh@2423 | 1222 | Node *needgc_false = new (C, 1) IfFalseNode(needgc_iff); |
duke@435 | 1223 | transform_later(needgc_false); |
duke@435 | 1224 | |
duke@435 | 1225 | // Grab regular I/O before optional prefetch may change it. |
duke@435 | 1226 | // Slow-path does no I/O so just set it to the original I/O. |
phh@2423 | 1227 | result_phi_i_o->init_req(slow_result_path, i_o); |
duke@435 | 1228 | |
duke@435 | 1229 | i_o = prefetch_allocation(i_o, needgc_false, contended_phi_rawmem, |
duke@435 | 1230 | old_eden_top, new_eden_top, length); |
duke@435 | 1231 | |
phh@2423 | 1232 | // Name successful fast-path variables |
phh@2423 | 1233 | Node* fast_oop = old_eden_top; |
phh@2423 | 1234 | Node* fast_oop_ctrl; |
phh@2423 | 1235 | Node* fast_oop_rawmem; |
phh@2423 | 1236 | |
duke@435 | 1237 | // Store (-conditional) the modified eden top back down. |
duke@435 | 1238 | // StorePConditional produces flags for a test PLUS a modified raw |
duke@435 | 1239 | // memory state. |
phh@2423 | 1240 | if (UseTLAB) { |
phh@2423 | 1241 | Node* store_eden_top = |
phh@2423 | 1242 | new (C, 4) StorePNode(needgc_false, contended_phi_rawmem, eden_top_adr, |
phh@2423 | 1243 | TypeRawPtr::BOTTOM, new_eden_top); |
duke@435 | 1244 | transform_later(store_eden_top); |
duke@435 | 1245 | fast_oop_ctrl = needgc_false; // No contention, so this is the fast path |
phh@2423 | 1246 | fast_oop_rawmem = store_eden_top; |
duke@435 | 1247 | } else { |
phh@2423 | 1248 | Node* store_eden_top = |
phh@2423 | 1249 | new (C, 5) StorePConditionalNode(needgc_false, contended_phi_rawmem, eden_top_adr, |
phh@2423 | 1250 | new_eden_top, fast_oop/*old_eden_top*/); |
duke@435 | 1251 | transform_later(store_eden_top); |
phh@2423 | 1252 | Node *contention_check = new (C, 2) BoolNode(store_eden_top, BoolTest::ne); |
duke@435 | 1253 | transform_later(contention_check); |
duke@435 | 1254 | store_eden_top = new (C, 1) SCMemProjNode(store_eden_top); |
duke@435 | 1255 | transform_later(store_eden_top); |
duke@435 | 1256 | |
duke@435 | 1257 | // If not using TLABs, check to see if there was contention. |
phh@2423 | 1258 | IfNode *contention_iff = new (C, 2) IfNode (needgc_false, contention_check, PROB_MIN, COUNT_UNKNOWN); |
duke@435 | 1259 | transform_later(contention_iff); |
phh@2423 | 1260 | Node *contention_true = new (C, 1) IfTrueNode(contention_iff); |
duke@435 | 1261 | transform_later(contention_true); |
duke@435 | 1262 | // If contention, loopback and try again. |
phh@2423 | 1263 | contended_region->init_req(contended_loopback_path, contention_true); |
phh@2423 | 1264 | contended_phi_rawmem->init_req(contended_loopback_path, store_eden_top); |
duke@435 | 1265 | |
duke@435 | 1266 | // Fast-path succeeded with no contention! |
phh@2423 | 1267 | Node *contention_false = new (C, 1) IfFalseNode(contention_iff); |
duke@435 | 1268 | transform_later(contention_false); |
duke@435 | 1269 | fast_oop_ctrl = contention_false; |
phh@2423 | 1270 | |
phh@2423 | 1271 | // Bump total allocated bytes for this thread |
phh@2423 | 1272 | Node* thread = new (C, 1) ThreadLocalNode(); |
phh@2423 | 1273 | transform_later(thread); |
phh@2423 | 1274 | Node* alloc_bytes_adr = basic_plus_adr(top()/*not oop*/, thread, |
phh@2423 | 1275 | in_bytes(JavaThread::allocated_bytes_offset())); |
phh@2423 | 1276 | Node* alloc_bytes = make_load(fast_oop_ctrl, store_eden_top, alloc_bytes_adr, |
phh@2423 | 1277 | 0, TypeLong::LONG, T_LONG); |
phh@2423 | 1278 | #ifdef _LP64 |
phh@2423 | 1279 | Node* alloc_size = size_in_bytes; |
phh@2423 | 1280 | #else |
phh@2423 | 1281 | Node* alloc_size = new (C, 2) ConvI2LNode(size_in_bytes); |
phh@2423 | 1282 | transform_later(alloc_size); |
phh@2423 | 1283 | #endif |
phh@2423 | 1284 | Node* new_alloc_bytes = new (C, 3) AddLNode(alloc_bytes, alloc_size); |
phh@2423 | 1285 | transform_later(new_alloc_bytes); |
phh@2423 | 1286 | fast_oop_rawmem = make_store(fast_oop_ctrl, store_eden_top, alloc_bytes_adr, |
phh@2423 | 1287 | 0, new_alloc_bytes, T_LONG); |
duke@435 | 1288 | } |
duke@435 | 1289 | |
duke@435 | 1290 | fast_oop_rawmem = initialize_object(alloc, |
duke@435 | 1291 | fast_oop_ctrl, fast_oop_rawmem, fast_oop, |
duke@435 | 1292 | klass_node, length, size_in_bytes); |
duke@435 | 1293 | |
kvn@1215 | 1294 | if (C->env()->dtrace_extended_probes()) { |
duke@435 | 1295 | // Slow-path call |
duke@435 | 1296 | int size = TypeFunc::Parms + 2; |
duke@435 | 1297 | CallLeafNode *call = new (C, size) CallLeafNode(OptoRuntime::dtrace_object_alloc_Type(), |
duke@435 | 1298 | CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_object_alloc_base), |
duke@435 | 1299 | "dtrace_object_alloc", |
duke@435 | 1300 | TypeRawPtr::BOTTOM); |
duke@435 | 1301 | |
duke@435 | 1302 | // Get base of thread-local storage area |
duke@435 | 1303 | Node* thread = new (C, 1) ThreadLocalNode(); |
duke@435 | 1304 | transform_later(thread); |
duke@435 | 1305 | |
duke@435 | 1306 | call->init_req(TypeFunc::Parms+0, thread); |
duke@435 | 1307 | call->init_req(TypeFunc::Parms+1, fast_oop); |
phh@2423 | 1308 | call->init_req(TypeFunc::Control, fast_oop_ctrl); |
phh@2423 | 1309 | call->init_req(TypeFunc::I_O , top()); // does no i/o |
phh@2423 | 1310 | call->init_req(TypeFunc::Memory , fast_oop_rawmem); |
phh@2423 | 1311 | call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr)); |
phh@2423 | 1312 | call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr)); |
duke@435 | 1313 | transform_later(call); |
duke@435 | 1314 | fast_oop_ctrl = new (C, 1) ProjNode(call,TypeFunc::Control); |
duke@435 | 1315 | transform_later(fast_oop_ctrl); |
duke@435 | 1316 | fast_oop_rawmem = new (C, 1) ProjNode(call,TypeFunc::Memory); |
duke@435 | 1317 | transform_later(fast_oop_rawmem); |
duke@435 | 1318 | } |
duke@435 | 1319 | |
duke@435 | 1320 | // Plug in the successful fast-path into the result merge point |
phh@2423 | 1321 | result_region ->init_req(fast_result_path, fast_oop_ctrl); |
phh@2423 | 1322 | result_phi_rawoop->init_req(fast_result_path, fast_oop); |
phh@2423 | 1323 | result_phi_i_o ->init_req(fast_result_path, i_o); |
phh@2423 | 1324 | result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem); |
duke@435 | 1325 | } else { |
duke@435 | 1326 | slow_region = ctrl; |
duke@435 | 1327 | } |
duke@435 | 1328 | |
duke@435 | 1329 | // Generate slow-path call |
duke@435 | 1330 | CallNode *call = new (C, slow_call_type->domain()->cnt()) |
duke@435 | 1331 | CallStaticJavaNode(slow_call_type, slow_call_address, |
duke@435 | 1332 | OptoRuntime::stub_name(slow_call_address), |
duke@435 | 1333 | alloc->jvms()->bci(), |
duke@435 | 1334 | TypePtr::BOTTOM); |
duke@435 | 1335 | call->init_req( TypeFunc::Control, slow_region ); |
duke@435 | 1336 | call->init_req( TypeFunc::I_O , top() ) ; // does no i/o |
duke@435 | 1337 | call->init_req( TypeFunc::Memory , slow_mem ); // may gc ptrs |
duke@435 | 1338 | call->init_req( TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr) ); |
duke@435 | 1339 | call->init_req( TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr) ); |
duke@435 | 1340 | |
duke@435 | 1341 | call->init_req(TypeFunc::Parms+0, klass_node); |
duke@435 | 1342 | if (length != NULL) { |
duke@435 | 1343 | call->init_req(TypeFunc::Parms+1, length); |
duke@435 | 1344 | } |
duke@435 | 1345 | |
duke@435 | 1346 | // Copy debug information and adjust JVMState information, then replace |
duke@435 | 1347 | // allocate node with the call |
duke@435 | 1348 | copy_call_debug_info((CallNode *) alloc, call); |
duke@435 | 1349 | if (!always_slow) { |
duke@435 | 1350 | call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON. |
duke@435 | 1351 | } |
kvn@1976 | 1352 | _igvn.replace_node(alloc, call); |
duke@435 | 1353 | transform_later(call); |
duke@435 | 1354 | |
duke@435 | 1355 | // Identify the output projections from the allocate node and |
duke@435 | 1356 | // adjust any references to them. |
duke@435 | 1357 | // The control and io projections look like: |
duke@435 | 1358 | // |
duke@435 | 1359 | // v---Proj(ctrl) <-----+ v---CatchProj(ctrl) |
duke@435 | 1360 | // Allocate Catch |
duke@435 | 1361 | // ^---Proj(io) <-------+ ^---CatchProj(io) |
duke@435 | 1362 | // |
duke@435 | 1363 | // We are interested in the CatchProj nodes. |
duke@435 | 1364 | // |
duke@435 | 1365 | extract_call_projections(call); |
duke@435 | 1366 | |
duke@435 | 1367 | // An allocate node has separate memory projections for the uses on the control and i_o paths |
duke@435 | 1368 | // Replace uses of the control memory projection with result_phi_rawmem (unless we are only generating a slow call) |
duke@435 | 1369 | if (!always_slow && _memproj_fallthrough != NULL) { |
duke@435 | 1370 | for (DUIterator_Fast imax, i = _memproj_fallthrough->fast_outs(imax); i < imax; i++) { |
duke@435 | 1371 | Node *use = _memproj_fallthrough->fast_out(i); |
duke@435 | 1372 | _igvn.hash_delete(use); |
duke@435 | 1373 | imax -= replace_input(use, _memproj_fallthrough, result_phi_rawmem); |
duke@435 | 1374 | _igvn._worklist.push(use); |
duke@435 | 1375 | // back up iterator |
duke@435 | 1376 | --i; |
duke@435 | 1377 | } |
duke@435 | 1378 | } |
duke@435 | 1379 | // Now change uses of _memproj_catchall to use _memproj_fallthrough and delete _memproj_catchall so |
duke@435 | 1380 | // we end up with a call that has only 1 memory projection |
duke@435 | 1381 | if (_memproj_catchall != NULL ) { |
duke@435 | 1382 | if (_memproj_fallthrough == NULL) { |
duke@435 | 1383 | _memproj_fallthrough = new (C, 1) ProjNode(call, TypeFunc::Memory); |
duke@435 | 1384 | transform_later(_memproj_fallthrough); |
duke@435 | 1385 | } |
duke@435 | 1386 | for (DUIterator_Fast imax, i = _memproj_catchall->fast_outs(imax); i < imax; i++) { |
duke@435 | 1387 | Node *use = _memproj_catchall->fast_out(i); |
duke@435 | 1388 | _igvn.hash_delete(use); |
duke@435 | 1389 | imax -= replace_input(use, _memproj_catchall, _memproj_fallthrough); |
duke@435 | 1390 | _igvn._worklist.push(use); |
duke@435 | 1391 | // back up iterator |
duke@435 | 1392 | --i; |
duke@435 | 1393 | } |
duke@435 | 1394 | } |
duke@435 | 1395 | |
duke@435 | 1396 | // An allocate node has separate i_o projections for the uses on the control and i_o paths |
duke@435 | 1397 | // Replace uses of the control i_o projection with result_phi_i_o (unless we are only generating a slow call) |
duke@435 | 1398 | if (_ioproj_fallthrough == NULL) { |
duke@435 | 1399 | _ioproj_fallthrough = new (C, 1) ProjNode(call, TypeFunc::I_O); |
duke@435 | 1400 | transform_later(_ioproj_fallthrough); |
duke@435 | 1401 | } else if (!always_slow) { |
duke@435 | 1402 | for (DUIterator_Fast imax, i = _ioproj_fallthrough->fast_outs(imax); i < imax; i++) { |
duke@435 | 1403 | Node *use = _ioproj_fallthrough->fast_out(i); |
duke@435 | 1404 | |
duke@435 | 1405 | _igvn.hash_delete(use); |
duke@435 | 1406 | imax -= replace_input(use, _ioproj_fallthrough, result_phi_i_o); |
duke@435 | 1407 | _igvn._worklist.push(use); |
duke@435 | 1408 | // back up iterator |
duke@435 | 1409 | --i; |
duke@435 | 1410 | } |
duke@435 | 1411 | } |
duke@435 | 1412 | // Now change uses of _ioproj_catchall to use _ioproj_fallthrough and delete _ioproj_catchall so |
duke@435 | 1413 | // we end up with a call that has only 1 control projection |
duke@435 | 1414 | if (_ioproj_catchall != NULL ) { |
duke@435 | 1415 | for (DUIterator_Fast imax, i = _ioproj_catchall->fast_outs(imax); i < imax; i++) { |
duke@435 | 1416 | Node *use = _ioproj_catchall->fast_out(i); |
duke@435 | 1417 | _igvn.hash_delete(use); |
duke@435 | 1418 | imax -= replace_input(use, _ioproj_catchall, _ioproj_fallthrough); |
duke@435 | 1419 | _igvn._worklist.push(use); |
duke@435 | 1420 | // back up iterator |
duke@435 | 1421 | --i; |
duke@435 | 1422 | } |
duke@435 | 1423 | } |
duke@435 | 1424 | |
duke@435 | 1425 | // if we generated only a slow call, we are done |
duke@435 | 1426 | if (always_slow) |
duke@435 | 1427 | return; |
duke@435 | 1428 | |
duke@435 | 1429 | |
duke@435 | 1430 | if (_fallthroughcatchproj != NULL) { |
duke@435 | 1431 | ctrl = _fallthroughcatchproj->clone(); |
duke@435 | 1432 | transform_later(ctrl); |
kvn@1143 | 1433 | _igvn.replace_node(_fallthroughcatchproj, result_region); |
duke@435 | 1434 | } else { |
duke@435 | 1435 | ctrl = top(); |
duke@435 | 1436 | } |
duke@435 | 1437 | Node *slow_result; |
duke@435 | 1438 | if (_resproj == NULL) { |
duke@435 | 1439 | // no uses of the allocation result |
duke@435 | 1440 | slow_result = top(); |
duke@435 | 1441 | } else { |
duke@435 | 1442 | slow_result = _resproj->clone(); |
duke@435 | 1443 | transform_later(slow_result); |
kvn@1143 | 1444 | _igvn.replace_node(_resproj, result_phi_rawoop); |
duke@435 | 1445 | } |
duke@435 | 1446 | |
duke@435 | 1447 | // Plug slow-path into result merge point |
duke@435 | 1448 | result_region ->init_req( slow_result_path, ctrl ); |
duke@435 | 1449 | result_phi_rawoop->init_req( slow_result_path, slow_result); |
duke@435 | 1450 | result_phi_rawmem->init_req( slow_result_path, _memproj_fallthrough ); |
duke@435 | 1451 | transform_later(result_region); |
duke@435 | 1452 | transform_later(result_phi_rawoop); |
duke@435 | 1453 | transform_later(result_phi_rawmem); |
duke@435 | 1454 | transform_later(result_phi_i_o); |
duke@435 | 1455 | // This completes all paths into the result merge point |
duke@435 | 1456 | } |
duke@435 | 1457 | |
duke@435 | 1458 | |
duke@435 | 1459 | // Helper for PhaseMacroExpand::expand_allocate_common. |
duke@435 | 1460 | // Initializes the newly-allocated storage. |
duke@435 | 1461 | Node* |
duke@435 | 1462 | PhaseMacroExpand::initialize_object(AllocateNode* alloc, |
duke@435 | 1463 | Node* control, Node* rawmem, Node* object, |
duke@435 | 1464 | Node* klass_node, Node* length, |
duke@435 | 1465 | Node* size_in_bytes) { |
duke@435 | 1466 | InitializeNode* init = alloc->initialization(); |
duke@435 | 1467 | // Store the klass & mark bits |
duke@435 | 1468 | Node* mark_node = NULL; |
duke@435 | 1469 | // For now only enable fast locking for non-array types |
duke@435 | 1470 | if (UseBiasedLocking && (length == NULL)) { |
kvn@1964 | 1471 | mark_node = make_load(control, rawmem, klass_node, Klass::prototype_header_offset_in_bytes() + sizeof(oopDesc), TypeRawPtr::BOTTOM, T_ADDRESS); |
duke@435 | 1472 | } else { |
duke@435 | 1473 | mark_node = makecon(TypeRawPtr::make((address)markOopDesc::prototype())); |
duke@435 | 1474 | } |
duke@435 | 1475 | rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS); |
coleenp@548 | 1476 | |
duke@435 | 1477 | rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_OBJECT); |
duke@435 | 1478 | int header_size = alloc->minimum_header_size(); // conservatively small |
duke@435 | 1479 | |
duke@435 | 1480 | // Array length |
duke@435 | 1481 | if (length != NULL) { // Arrays need length field |
duke@435 | 1482 | rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT); |
duke@435 | 1483 | // conservatively small header size: |
coleenp@548 | 1484 | header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE); |
duke@435 | 1485 | ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass(); |
duke@435 | 1486 | if (k->is_array_klass()) // we know the exact header size in most cases: |
duke@435 | 1487 | header_size = Klass::layout_helper_header_size(k->layout_helper()); |
duke@435 | 1488 | } |
duke@435 | 1489 | |
duke@435 | 1490 | // Clear the object body, if necessary. |
duke@435 | 1491 | if (init == NULL) { |
duke@435 | 1492 | // The init has somehow disappeared; be cautious and clear everything. |
duke@435 | 1493 | // |
duke@435 | 1494 | // This can happen if a node is allocated but an uncommon trap occurs |
duke@435 | 1495 | // immediately. In this case, the Initialize gets associated with the |
duke@435 | 1496 | // trap, and may be placed in a different (outer) loop, if the Allocate |
duke@435 | 1497 | // is in a loop. If (this is rare) the inner loop gets unrolled, then |
duke@435 | 1498 | // there can be two Allocates to one Initialize. The answer in all these |
duke@435 | 1499 | // edge cases is safety first. It is always safe to clear immediately |
duke@435 | 1500 | // within an Allocate, and then (maybe or maybe not) clear some more later. |
duke@435 | 1501 | if (!ZeroTLAB) |
duke@435 | 1502 | rawmem = ClearArrayNode::clear_memory(control, rawmem, object, |
duke@435 | 1503 | header_size, size_in_bytes, |
duke@435 | 1504 | &_igvn); |
duke@435 | 1505 | } else { |
duke@435 | 1506 | if (!init->is_complete()) { |
duke@435 | 1507 | // Try to win by zeroing only what the init does not store. |
duke@435 | 1508 | // We can also try to do some peephole optimizations, |
duke@435 | 1509 | // such as combining some adjacent subword stores. |
duke@435 | 1510 | rawmem = init->complete_stores(control, rawmem, object, |
duke@435 | 1511 | header_size, size_in_bytes, &_igvn); |
duke@435 | 1512 | } |
duke@435 | 1513 | // We have no more use for this link, since the AllocateNode goes away: |
duke@435 | 1514 | init->set_req(InitializeNode::RawAddress, top()); |
duke@435 | 1515 | // (If we keep the link, it just confuses the register allocator, |
duke@435 | 1516 | // who thinks he sees a real use of the address by the membar.) |
duke@435 | 1517 | } |
duke@435 | 1518 | |
duke@435 | 1519 | return rawmem; |
duke@435 | 1520 | } |
duke@435 | 1521 | |
duke@435 | 1522 | // Generate prefetch instructions for next allocations. |
duke@435 | 1523 | Node* PhaseMacroExpand::prefetch_allocation(Node* i_o, Node*& needgc_false, |
duke@435 | 1524 | Node*& contended_phi_rawmem, |
duke@435 | 1525 | Node* old_eden_top, Node* new_eden_top, |
duke@435 | 1526 | Node* length) { |
kvn@1802 | 1527 | enum { fall_in_path = 1, pf_path = 2 }; |
duke@435 | 1528 | if( UseTLAB && AllocatePrefetchStyle == 2 ) { |
duke@435 | 1529 | // Generate prefetch allocation with watermark check. |
duke@435 | 1530 | // As an allocation hits the watermark, we will prefetch starting |
duke@435 | 1531 | // at a "distance" away from watermark. |
duke@435 | 1532 | |
duke@435 | 1533 | Node *pf_region = new (C, 3) RegionNode(3); |
duke@435 | 1534 | Node *pf_phi_rawmem = new (C, 3) PhiNode( pf_region, Type::MEMORY, |
duke@435 | 1535 | TypeRawPtr::BOTTOM ); |
duke@435 | 1536 | // I/O is used for Prefetch |
duke@435 | 1537 | Node *pf_phi_abio = new (C, 3) PhiNode( pf_region, Type::ABIO ); |
duke@435 | 1538 | |
duke@435 | 1539 | Node *thread = new (C, 1) ThreadLocalNode(); |
duke@435 | 1540 | transform_later(thread); |
duke@435 | 1541 | |
duke@435 | 1542 | Node *eden_pf_adr = new (C, 4) AddPNode( top()/*not oop*/, thread, |
duke@435 | 1543 | _igvn.MakeConX(in_bytes(JavaThread::tlab_pf_top_offset())) ); |
duke@435 | 1544 | transform_later(eden_pf_adr); |
duke@435 | 1545 | |
duke@435 | 1546 | Node *old_pf_wm = new (C, 3) LoadPNode( needgc_false, |
duke@435 | 1547 | contended_phi_rawmem, eden_pf_adr, |
duke@435 | 1548 | TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM ); |
duke@435 | 1549 | transform_later(old_pf_wm); |
duke@435 | 1550 | |
duke@435 | 1551 | // check against new_eden_top |
duke@435 | 1552 | Node *need_pf_cmp = new (C, 3) CmpPNode( new_eden_top, old_pf_wm ); |
duke@435 | 1553 | transform_later(need_pf_cmp); |
duke@435 | 1554 | Node *need_pf_bol = new (C, 2) BoolNode( need_pf_cmp, BoolTest::ge ); |
duke@435 | 1555 | transform_later(need_pf_bol); |
duke@435 | 1556 | IfNode *need_pf_iff = new (C, 2) IfNode( needgc_false, need_pf_bol, |
duke@435 | 1557 | PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN ); |
duke@435 | 1558 | transform_later(need_pf_iff); |
duke@435 | 1559 | |
duke@435 | 1560 | // true node, add prefetchdistance |
duke@435 | 1561 | Node *need_pf_true = new (C, 1) IfTrueNode( need_pf_iff ); |
duke@435 | 1562 | transform_later(need_pf_true); |
duke@435 | 1563 | |
duke@435 | 1564 | Node *need_pf_false = new (C, 1) IfFalseNode( need_pf_iff ); |
duke@435 | 1565 | transform_later(need_pf_false); |
duke@435 | 1566 | |
duke@435 | 1567 | Node *new_pf_wmt = new (C, 4) AddPNode( top(), old_pf_wm, |
duke@435 | 1568 | _igvn.MakeConX(AllocatePrefetchDistance) ); |
duke@435 | 1569 | transform_later(new_pf_wmt ); |
duke@435 | 1570 | new_pf_wmt->set_req(0, need_pf_true); |
duke@435 | 1571 | |
duke@435 | 1572 | Node *store_new_wmt = new (C, 4) StorePNode( need_pf_true, |
duke@435 | 1573 | contended_phi_rawmem, eden_pf_adr, |
duke@435 | 1574 | TypeRawPtr::BOTTOM, new_pf_wmt ); |
duke@435 | 1575 | transform_later(store_new_wmt); |
duke@435 | 1576 | |
duke@435 | 1577 | // adding prefetches |
duke@435 | 1578 | pf_phi_abio->init_req( fall_in_path, i_o ); |
duke@435 | 1579 | |
duke@435 | 1580 | Node *prefetch_adr; |
duke@435 | 1581 | Node *prefetch; |
duke@435 | 1582 | uint lines = AllocatePrefetchDistance / AllocatePrefetchStepSize; |
duke@435 | 1583 | uint step_size = AllocatePrefetchStepSize; |
duke@435 | 1584 | uint distance = 0; |
duke@435 | 1585 | |
duke@435 | 1586 | for ( uint i = 0; i < lines; i++ ) { |
duke@435 | 1587 | prefetch_adr = new (C, 4) AddPNode( old_pf_wm, new_pf_wmt, |
duke@435 | 1588 | _igvn.MakeConX(distance) ); |
duke@435 | 1589 | transform_later(prefetch_adr); |
duke@435 | 1590 | prefetch = new (C, 3) PrefetchWriteNode( i_o, prefetch_adr ); |
duke@435 | 1591 | transform_later(prefetch); |
duke@435 | 1592 | distance += step_size; |
duke@435 | 1593 | i_o = prefetch; |
duke@435 | 1594 | } |
duke@435 | 1595 | pf_phi_abio->set_req( pf_path, i_o ); |
duke@435 | 1596 | |
duke@435 | 1597 | pf_region->init_req( fall_in_path, need_pf_false ); |
duke@435 | 1598 | pf_region->init_req( pf_path, need_pf_true ); |
duke@435 | 1599 | |
duke@435 | 1600 | pf_phi_rawmem->init_req( fall_in_path, contended_phi_rawmem ); |
duke@435 | 1601 | pf_phi_rawmem->init_req( pf_path, store_new_wmt ); |
duke@435 | 1602 | |
duke@435 | 1603 | transform_later(pf_region); |
duke@435 | 1604 | transform_later(pf_phi_rawmem); |
duke@435 | 1605 | transform_later(pf_phi_abio); |
duke@435 | 1606 | |
duke@435 | 1607 | needgc_false = pf_region; |
duke@435 | 1608 | contended_phi_rawmem = pf_phi_rawmem; |
duke@435 | 1609 | i_o = pf_phi_abio; |
kvn@1802 | 1610 | } else if( UseTLAB && AllocatePrefetchStyle == 3 ) { |
kvn@1802 | 1611 | // Insert a prefetch for each allocation only on the fast-path |
kvn@1802 | 1612 | Node *pf_region = new (C, 3) RegionNode(3); |
kvn@1802 | 1613 | Node *pf_phi_rawmem = new (C, 3) PhiNode( pf_region, Type::MEMORY, |
kvn@1802 | 1614 | TypeRawPtr::BOTTOM ); |
kvn@1802 | 1615 | |
kvn@1802 | 1616 | // Generate several prefetch instructions only for arrays. |
kvn@1802 | 1617 | uint lines = (length != NULL) ? AllocatePrefetchLines : 1; |
kvn@1802 | 1618 | uint step_size = AllocatePrefetchStepSize; |
kvn@1802 | 1619 | uint distance = AllocatePrefetchDistance; |
kvn@1802 | 1620 | |
kvn@1802 | 1621 | // Next cache address. |
kvn@1802 | 1622 | Node *cache_adr = new (C, 4) AddPNode(old_eden_top, old_eden_top, |
kvn@1802 | 1623 | _igvn.MakeConX(distance)); |
kvn@1802 | 1624 | transform_later(cache_adr); |
kvn@1802 | 1625 | cache_adr = new (C, 2) CastP2XNode(needgc_false, cache_adr); |
kvn@1802 | 1626 | transform_later(cache_adr); |
kvn@1802 | 1627 | Node* mask = _igvn.MakeConX(~(intptr_t)(step_size-1)); |
kvn@1802 | 1628 | cache_adr = new (C, 3) AndXNode(cache_adr, mask); |
kvn@1802 | 1629 | transform_later(cache_adr); |
kvn@1802 | 1630 | cache_adr = new (C, 2) CastX2PNode(cache_adr); |
kvn@1802 | 1631 | transform_later(cache_adr); |
kvn@1802 | 1632 | |
kvn@1802 | 1633 | // Prefetch |
kvn@1802 | 1634 | Node *prefetch = new (C, 3) PrefetchWriteNode( contended_phi_rawmem, cache_adr ); |
kvn@1802 | 1635 | prefetch->set_req(0, needgc_false); |
kvn@1802 | 1636 | transform_later(prefetch); |
kvn@1802 | 1637 | contended_phi_rawmem = prefetch; |
kvn@1802 | 1638 | Node *prefetch_adr; |
kvn@1802 | 1639 | distance = step_size; |
kvn@1802 | 1640 | for ( uint i = 1; i < lines; i++ ) { |
kvn@1802 | 1641 | prefetch_adr = new (C, 4) AddPNode( cache_adr, cache_adr, |
kvn@1802 | 1642 | _igvn.MakeConX(distance) ); |
kvn@1802 | 1643 | transform_later(prefetch_adr); |
kvn@1802 | 1644 | prefetch = new (C, 3) PrefetchWriteNode( contended_phi_rawmem, prefetch_adr ); |
kvn@1802 | 1645 | transform_later(prefetch); |
kvn@1802 | 1646 | distance += step_size; |
kvn@1802 | 1647 | contended_phi_rawmem = prefetch; |
kvn@1802 | 1648 | } |
duke@435 | 1649 | } else if( AllocatePrefetchStyle > 0 ) { |
duke@435 | 1650 | // Insert a prefetch for each allocation only on the fast-path |
duke@435 | 1651 | Node *prefetch_adr; |
duke@435 | 1652 | Node *prefetch; |
duke@435 | 1653 | // Generate several prefetch instructions only for arrays. |
duke@435 | 1654 | uint lines = (length != NULL) ? AllocatePrefetchLines : 1; |
duke@435 | 1655 | uint step_size = AllocatePrefetchStepSize; |
duke@435 | 1656 | uint distance = AllocatePrefetchDistance; |
duke@435 | 1657 | for ( uint i = 0; i < lines; i++ ) { |
duke@435 | 1658 | prefetch_adr = new (C, 4) AddPNode( old_eden_top, new_eden_top, |
duke@435 | 1659 | _igvn.MakeConX(distance) ); |
duke@435 | 1660 | transform_later(prefetch_adr); |
duke@435 | 1661 | prefetch = new (C, 3) PrefetchWriteNode( i_o, prefetch_adr ); |
duke@435 | 1662 | // Do not let it float too high, since if eden_top == eden_end, |
duke@435 | 1663 | // both might be null. |
duke@435 | 1664 | if( i == 0 ) { // Set control for first prefetch, next follows it |
duke@435 | 1665 | prefetch->init_req(0, needgc_false); |
duke@435 | 1666 | } |
duke@435 | 1667 | transform_later(prefetch); |
duke@435 | 1668 | distance += step_size; |
duke@435 | 1669 | i_o = prefetch; |
duke@435 | 1670 | } |
duke@435 | 1671 | } |
duke@435 | 1672 | return i_o; |
duke@435 | 1673 | } |
duke@435 | 1674 | |
duke@435 | 1675 | |
duke@435 | 1676 | void PhaseMacroExpand::expand_allocate(AllocateNode *alloc) { |
duke@435 | 1677 | expand_allocate_common(alloc, NULL, |
duke@435 | 1678 | OptoRuntime::new_instance_Type(), |
duke@435 | 1679 | OptoRuntime::new_instance_Java()); |
duke@435 | 1680 | } |
duke@435 | 1681 | |
duke@435 | 1682 | void PhaseMacroExpand::expand_allocate_array(AllocateArrayNode *alloc) { |
duke@435 | 1683 | Node* length = alloc->in(AllocateNode::ALength); |
duke@435 | 1684 | expand_allocate_common(alloc, length, |
duke@435 | 1685 | OptoRuntime::new_array_Type(), |
duke@435 | 1686 | OptoRuntime::new_array_Java()); |
duke@435 | 1687 | } |
duke@435 | 1688 | |
duke@435 | 1689 | |
duke@435 | 1690 | // we have determined that this lock/unlock can be eliminated, we simply |
duke@435 | 1691 | // eliminate the node without expanding it. |
duke@435 | 1692 | // |
duke@435 | 1693 | // Note: The membar's associated with the lock/unlock are currently not |
duke@435 | 1694 | // eliminated. This should be investigated as a future enhancement. |
duke@435 | 1695 | // |
kvn@501 | 1696 | bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) { |
kvn@501 | 1697 | |
kvn@501 | 1698 | if (!alock->is_eliminated()) { |
kvn@501 | 1699 | return false; |
kvn@501 | 1700 | } |
kvn@895 | 1701 | if (alock->is_Lock() && !alock->is_coarsened()) { |
kvn@895 | 1702 | // Create new "eliminated" BoxLock node and use it |
kvn@895 | 1703 | // in monitor debug info for the same object. |
kvn@895 | 1704 | BoxLockNode* oldbox = alock->box_node()->as_BoxLock(); |
kvn@895 | 1705 | Node* obj = alock->obj_node(); |
kvn@895 | 1706 | if (!oldbox->is_eliminated()) { |
kvn@895 | 1707 | BoxLockNode* newbox = oldbox->clone()->as_BoxLock(); |
kvn@895 | 1708 | newbox->set_eliminated(); |
kvn@895 | 1709 | transform_later(newbox); |
kvn@895 | 1710 | // Replace old box node with new box for all users |
kvn@895 | 1711 | // of the same object. |
kvn@895 | 1712 | for (uint i = 0; i < oldbox->outcnt();) { |
kvn@895 | 1713 | |
kvn@895 | 1714 | bool next_edge = true; |
kvn@895 | 1715 | Node* u = oldbox->raw_out(i); |
kvn@895 | 1716 | if (u == alock) { |
kvn@895 | 1717 | i++; |
kvn@895 | 1718 | continue; // It will be removed below |
kvn@895 | 1719 | } |
kvn@895 | 1720 | if (u->is_Lock() && |
kvn@895 | 1721 | u->as_Lock()->obj_node() == obj && |
kvn@895 | 1722 | // oldbox could be referenced in debug info also |
kvn@895 | 1723 | u->as_Lock()->box_node() == oldbox) { |
kvn@895 | 1724 | assert(u->as_Lock()->is_eliminated(), "sanity"); |
kvn@895 | 1725 | _igvn.hash_delete(u); |
kvn@895 | 1726 | u->set_req(TypeFunc::Parms + 1, newbox); |
kvn@895 | 1727 | next_edge = false; |
kvn@895 | 1728 | #ifdef ASSERT |
kvn@895 | 1729 | } else if (u->is_Unlock() && u->as_Unlock()->obj_node() == obj) { |
kvn@895 | 1730 | assert(u->as_Unlock()->is_eliminated(), "sanity"); |
kvn@895 | 1731 | #endif |
kvn@895 | 1732 | } |
kvn@895 | 1733 | // Replace old box in monitor debug info. |
kvn@895 | 1734 | if (u->is_SafePoint() && u->as_SafePoint()->jvms()) { |
kvn@895 | 1735 | SafePointNode* sfn = u->as_SafePoint(); |
kvn@895 | 1736 | JVMState* youngest_jvms = sfn->jvms(); |
kvn@895 | 1737 | int max_depth = youngest_jvms->depth(); |
kvn@895 | 1738 | for (int depth = 1; depth <= max_depth; depth++) { |
kvn@895 | 1739 | JVMState* jvms = youngest_jvms->of_depth(depth); |
kvn@895 | 1740 | int num_mon = jvms->nof_monitors(); |
kvn@895 | 1741 | // Loop over monitors |
kvn@895 | 1742 | for (int idx = 0; idx < num_mon; idx++) { |
kvn@895 | 1743 | Node* obj_node = sfn->monitor_obj(jvms, idx); |
kvn@895 | 1744 | Node* box_node = sfn->monitor_box(jvms, idx); |
kvn@895 | 1745 | if (box_node == oldbox && obj_node == obj) { |
kvn@895 | 1746 | int j = jvms->monitor_box_offset(idx); |
kvn@895 | 1747 | _igvn.hash_delete(u); |
kvn@895 | 1748 | u->set_req(j, newbox); |
kvn@895 | 1749 | next_edge = false; |
kvn@895 | 1750 | } |
kvn@895 | 1751 | } // for (int idx = 0; |
kvn@895 | 1752 | } // for (int depth = 1; |
kvn@895 | 1753 | } // if (u->is_SafePoint() |
kvn@895 | 1754 | if (next_edge) i++; |
kvn@895 | 1755 | } // for (uint i = 0; i < oldbox->outcnt();) |
kvn@895 | 1756 | } // if (!oldbox->is_eliminated()) |
kvn@895 | 1757 | } // if (alock->is_Lock() && !lock->is_coarsened()) |
kvn@501 | 1758 | |
never@1515 | 1759 | CompileLog* log = C->log(); |
never@1515 | 1760 | if (log != NULL) { |
never@1515 | 1761 | log->head("eliminate_lock lock='%d'", |
never@1515 | 1762 | alock->is_Lock()); |
never@1515 | 1763 | JVMState* p = alock->jvms(); |
never@1515 | 1764 | while (p != NULL) { |
never@1515 | 1765 | log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method())); |
never@1515 | 1766 | p = p->caller(); |
never@1515 | 1767 | } |
never@1515 | 1768 | log->tail("eliminate_lock"); |
never@1515 | 1769 | } |
never@1515 | 1770 | |
kvn@501 | 1771 | #ifndef PRODUCT |
kvn@501 | 1772 | if (PrintEliminateLocks) { |
kvn@501 | 1773 | if (alock->is_Lock()) { |
kvn@501 | 1774 | tty->print_cr("++++ Eliminating: %d Lock", alock->_idx); |
kvn@501 | 1775 | } else { |
kvn@501 | 1776 | tty->print_cr("++++ Eliminating: %d Unlock", alock->_idx); |
kvn@501 | 1777 | } |
kvn@501 | 1778 | } |
kvn@501 | 1779 | #endif |
kvn@501 | 1780 | |
kvn@501 | 1781 | Node* mem = alock->in(TypeFunc::Memory); |
kvn@501 | 1782 | Node* ctrl = alock->in(TypeFunc::Control); |
kvn@501 | 1783 | |
kvn@501 | 1784 | extract_call_projections(alock); |
kvn@501 | 1785 | // There are 2 projections from the lock. The lock node will |
kvn@501 | 1786 | // be deleted when its last use is subsumed below. |
kvn@501 | 1787 | assert(alock->outcnt() == 2 && |
kvn@501 | 1788 | _fallthroughproj != NULL && |
kvn@501 | 1789 | _memproj_fallthrough != NULL, |
kvn@501 | 1790 | "Unexpected projections from Lock/Unlock"); |
kvn@501 | 1791 | |
kvn@501 | 1792 | Node* fallthroughproj = _fallthroughproj; |
kvn@501 | 1793 | Node* memproj_fallthrough = _memproj_fallthrough; |
duke@435 | 1794 | |
duke@435 | 1795 | // The memory projection from a lock/unlock is RawMem |
duke@435 | 1796 | // The input to a Lock is merged memory, so extract its RawMem input |
duke@435 | 1797 | // (unless the MergeMem has been optimized away.) |
duke@435 | 1798 | if (alock->is_Lock()) { |
kvn@501 | 1799 | // Seach for MemBarAcquire node and delete it also. |
kvn@501 | 1800 | MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar(); |
kvn@501 | 1801 | assert(membar != NULL && membar->Opcode() == Op_MemBarAcquire, ""); |
kvn@501 | 1802 | Node* ctrlproj = membar->proj_out(TypeFunc::Control); |
kvn@501 | 1803 | Node* memproj = membar->proj_out(TypeFunc::Memory); |
kvn@1143 | 1804 | _igvn.replace_node(ctrlproj, fallthroughproj); |
kvn@1143 | 1805 | _igvn.replace_node(memproj, memproj_fallthrough); |
kvn@895 | 1806 | |
kvn@895 | 1807 | // Delete FastLock node also if this Lock node is unique user |
kvn@895 | 1808 | // (a loop peeling may clone a Lock node). |
kvn@895 | 1809 | Node* flock = alock->as_Lock()->fastlock_node(); |
kvn@895 | 1810 | if (flock->outcnt() == 1) { |
kvn@895 | 1811 | assert(flock->unique_out() == alock, "sanity"); |
kvn@1143 | 1812 | _igvn.replace_node(flock, top()); |
kvn@895 | 1813 | } |
duke@435 | 1814 | } |
duke@435 | 1815 | |
kvn@501 | 1816 | // Seach for MemBarRelease node and delete it also. |
kvn@501 | 1817 | if (alock->is_Unlock() && ctrl != NULL && ctrl->is_Proj() && |
kvn@501 | 1818 | ctrl->in(0)->is_MemBar()) { |
kvn@501 | 1819 | MemBarNode* membar = ctrl->in(0)->as_MemBar(); |
kvn@501 | 1820 | assert(membar->Opcode() == Op_MemBarRelease && |
kvn@501 | 1821 | mem->is_Proj() && membar == mem->in(0), ""); |
kvn@1143 | 1822 | _igvn.replace_node(fallthroughproj, ctrl); |
kvn@1143 | 1823 | _igvn.replace_node(memproj_fallthrough, mem); |
kvn@501 | 1824 | fallthroughproj = ctrl; |
kvn@501 | 1825 | memproj_fallthrough = mem; |
kvn@501 | 1826 | ctrl = membar->in(TypeFunc::Control); |
kvn@501 | 1827 | mem = membar->in(TypeFunc::Memory); |
kvn@501 | 1828 | } |
kvn@501 | 1829 | |
kvn@1143 | 1830 | _igvn.replace_node(fallthroughproj, ctrl); |
kvn@1143 | 1831 | _igvn.replace_node(memproj_fallthrough, mem); |
kvn@501 | 1832 | return true; |
duke@435 | 1833 | } |
duke@435 | 1834 | |
duke@435 | 1835 | |
duke@435 | 1836 | //------------------------------expand_lock_node---------------------- |
duke@435 | 1837 | void PhaseMacroExpand::expand_lock_node(LockNode *lock) { |
duke@435 | 1838 | |
duke@435 | 1839 | Node* ctrl = lock->in(TypeFunc::Control); |
duke@435 | 1840 | Node* mem = lock->in(TypeFunc::Memory); |
duke@435 | 1841 | Node* obj = lock->obj_node(); |
duke@435 | 1842 | Node* box = lock->box_node(); |
kvn@501 | 1843 | Node* flock = lock->fastlock_node(); |
duke@435 | 1844 | |
duke@435 | 1845 | // Make the merge point |
kvn@855 | 1846 | Node *region; |
kvn@855 | 1847 | Node *mem_phi; |
kvn@855 | 1848 | Node *slow_path; |
duke@435 | 1849 | |
kvn@855 | 1850 | if (UseOptoBiasInlining) { |
kvn@855 | 1851 | /* |
twisti@1040 | 1852 | * See the full description in MacroAssembler::biased_locking_enter(). |
kvn@855 | 1853 | * |
kvn@855 | 1854 | * if( (mark_word & biased_lock_mask) == biased_lock_pattern ) { |
kvn@855 | 1855 | * // The object is biased. |
kvn@855 | 1856 | * proto_node = klass->prototype_header; |
kvn@855 | 1857 | * o_node = thread | proto_node; |
kvn@855 | 1858 | * x_node = o_node ^ mark_word; |
kvn@855 | 1859 | * if( (x_node & ~age_mask) == 0 ) { // Biased to the current thread ? |
kvn@855 | 1860 | * // Done. |
kvn@855 | 1861 | * } else { |
kvn@855 | 1862 | * if( (x_node & biased_lock_mask) != 0 ) { |
kvn@855 | 1863 | * // The klass's prototype header is no longer biased. |
kvn@855 | 1864 | * cas(&mark_word, mark_word, proto_node) |
kvn@855 | 1865 | * goto cas_lock; |
kvn@855 | 1866 | * } else { |
kvn@855 | 1867 | * // The klass's prototype header is still biased. |
kvn@855 | 1868 | * if( (x_node & epoch_mask) != 0 ) { // Expired epoch? |
kvn@855 | 1869 | * old = mark_word; |
kvn@855 | 1870 | * new = o_node; |
kvn@855 | 1871 | * } else { |
kvn@855 | 1872 | * // Different thread or anonymous biased. |
kvn@855 | 1873 | * old = mark_word & (epoch_mask | age_mask | biased_lock_mask); |
kvn@855 | 1874 | * new = thread | old; |
kvn@855 | 1875 | * } |
kvn@855 | 1876 | * // Try to rebias. |
kvn@855 | 1877 | * if( cas(&mark_word, old, new) == 0 ) { |
kvn@855 | 1878 | * // Done. |
kvn@855 | 1879 | * } else { |
kvn@855 | 1880 | * goto slow_path; // Failed. |
kvn@855 | 1881 | * } |
kvn@855 | 1882 | * } |
kvn@855 | 1883 | * } |
kvn@855 | 1884 | * } else { |
kvn@855 | 1885 | * // The object is not biased. |
kvn@855 | 1886 | * cas_lock: |
kvn@855 | 1887 | * if( FastLock(obj) == 0 ) { |
kvn@855 | 1888 | * // Done. |
kvn@855 | 1889 | * } else { |
kvn@855 | 1890 | * slow_path: |
kvn@855 | 1891 | * OptoRuntime::complete_monitor_locking_Java(obj); |
kvn@855 | 1892 | * } |
kvn@855 | 1893 | * } |
kvn@855 | 1894 | */ |
kvn@855 | 1895 | |
kvn@855 | 1896 | region = new (C, 5) RegionNode(5); |
kvn@855 | 1897 | // create a Phi for the memory state |
kvn@855 | 1898 | mem_phi = new (C, 5) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); |
kvn@855 | 1899 | |
kvn@855 | 1900 | Node* fast_lock_region = new (C, 3) RegionNode(3); |
kvn@855 | 1901 | Node* fast_lock_mem_phi = new (C, 3) PhiNode( fast_lock_region, Type::MEMORY, TypeRawPtr::BOTTOM); |
kvn@855 | 1902 | |
kvn@855 | 1903 | // First, check mark word for the biased lock pattern. |
kvn@855 | 1904 | Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type()); |
kvn@855 | 1905 | |
kvn@855 | 1906 | // Get fast path - mark word has the biased lock pattern. |
kvn@855 | 1907 | ctrl = opt_bits_test(ctrl, fast_lock_region, 1, mark_node, |
kvn@855 | 1908 | markOopDesc::biased_lock_mask_in_place, |
kvn@855 | 1909 | markOopDesc::biased_lock_pattern, true); |
kvn@855 | 1910 | // fast_lock_region->in(1) is set to slow path. |
kvn@855 | 1911 | fast_lock_mem_phi->init_req(1, mem); |
kvn@855 | 1912 | |
kvn@855 | 1913 | // Now check that the lock is biased to the current thread and has |
kvn@855 | 1914 | // the same epoch and bias as Klass::_prototype_header. |
kvn@855 | 1915 | |
kvn@855 | 1916 | // Special-case a fresh allocation to avoid building nodes: |
kvn@855 | 1917 | Node* klass_node = AllocateNode::Ideal_klass(obj, &_igvn); |
kvn@855 | 1918 | if (klass_node == NULL) { |
kvn@855 | 1919 | Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes()); |
kvn@855 | 1920 | klass_node = transform_later( LoadKlassNode::make(_igvn, mem, k_adr, _igvn.type(k_adr)->is_ptr()) ); |
kvn@925 | 1921 | #ifdef _LP64 |
kvn@925 | 1922 | if (UseCompressedOops && klass_node->is_DecodeN()) { |
kvn@925 | 1923 | assert(klass_node->in(1)->Opcode() == Op_LoadNKlass, "sanity"); |
kvn@925 | 1924 | klass_node->in(1)->init_req(0, ctrl); |
kvn@925 | 1925 | } else |
kvn@925 | 1926 | #endif |
kvn@925 | 1927 | klass_node->init_req(0, ctrl); |
kvn@855 | 1928 | } |
kvn@855 | 1929 | Node *proto_node = make_load(ctrl, mem, klass_node, Klass::prototype_header_offset_in_bytes() + sizeof(oopDesc), TypeX_X, TypeX_X->basic_type()); |
kvn@855 | 1930 | |
kvn@855 | 1931 | Node* thread = transform_later(new (C, 1) ThreadLocalNode()); |
kvn@855 | 1932 | Node* cast_thread = transform_later(new (C, 2) CastP2XNode(ctrl, thread)); |
kvn@855 | 1933 | Node* o_node = transform_later(new (C, 3) OrXNode(cast_thread, proto_node)); |
kvn@855 | 1934 | Node* x_node = transform_later(new (C, 3) XorXNode(o_node, mark_node)); |
kvn@855 | 1935 | |
kvn@855 | 1936 | // Get slow path - mark word does NOT match the value. |
kvn@855 | 1937 | Node* not_biased_ctrl = opt_bits_test(ctrl, region, 3, x_node, |
kvn@855 | 1938 | (~markOopDesc::age_mask_in_place), 0); |
kvn@855 | 1939 | // region->in(3) is set to fast path - the object is biased to the current thread. |
kvn@855 | 1940 | mem_phi->init_req(3, mem); |
kvn@855 | 1941 | |
kvn@855 | 1942 | |
kvn@855 | 1943 | // Mark word does NOT match the value (thread | Klass::_prototype_header). |
kvn@855 | 1944 | |
kvn@855 | 1945 | |
kvn@855 | 1946 | // First, check biased pattern. |
kvn@855 | 1947 | // Get fast path - _prototype_header has the same biased lock pattern. |
kvn@855 | 1948 | ctrl = opt_bits_test(not_biased_ctrl, fast_lock_region, 2, x_node, |
kvn@855 | 1949 | markOopDesc::biased_lock_mask_in_place, 0, true); |
kvn@855 | 1950 | |
kvn@855 | 1951 | not_biased_ctrl = fast_lock_region->in(2); // Slow path |
kvn@855 | 1952 | // fast_lock_region->in(2) - the prototype header is no longer biased |
kvn@855 | 1953 | // and we have to revoke the bias on this object. |
kvn@855 | 1954 | // We are going to try to reset the mark of this object to the prototype |
kvn@855 | 1955 | // value and fall through to the CAS-based locking scheme. |
kvn@855 | 1956 | Node* adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes()); |
kvn@855 | 1957 | Node* cas = new (C, 5) StoreXConditionalNode(not_biased_ctrl, mem, adr, |
kvn@855 | 1958 | proto_node, mark_node); |
kvn@855 | 1959 | transform_later(cas); |
kvn@855 | 1960 | Node* proj = transform_later( new (C, 1) SCMemProjNode(cas)); |
kvn@855 | 1961 | fast_lock_mem_phi->init_req(2, proj); |
kvn@855 | 1962 | |
kvn@855 | 1963 | |
kvn@855 | 1964 | // Second, check epoch bits. |
kvn@855 | 1965 | Node* rebiased_region = new (C, 3) RegionNode(3); |
kvn@855 | 1966 | Node* old_phi = new (C, 3) PhiNode( rebiased_region, TypeX_X); |
kvn@855 | 1967 | Node* new_phi = new (C, 3) PhiNode( rebiased_region, TypeX_X); |
kvn@855 | 1968 | |
kvn@855 | 1969 | // Get slow path - mark word does NOT match epoch bits. |
kvn@855 | 1970 | Node* epoch_ctrl = opt_bits_test(ctrl, rebiased_region, 1, x_node, |
kvn@855 | 1971 | markOopDesc::epoch_mask_in_place, 0); |
kvn@855 | 1972 | // The epoch of the current bias is not valid, attempt to rebias the object |
kvn@855 | 1973 | // toward the current thread. |
kvn@855 | 1974 | rebiased_region->init_req(2, epoch_ctrl); |
kvn@855 | 1975 | old_phi->init_req(2, mark_node); |
kvn@855 | 1976 | new_phi->init_req(2, o_node); |
kvn@855 | 1977 | |
kvn@855 | 1978 | // rebiased_region->in(1) is set to fast path. |
kvn@855 | 1979 | // The epoch of the current bias is still valid but we know |
kvn@855 | 1980 | // nothing about the owner; it might be set or it might be clear. |
kvn@855 | 1981 | Node* cmask = MakeConX(markOopDesc::biased_lock_mask_in_place | |
kvn@855 | 1982 | markOopDesc::age_mask_in_place | |
kvn@855 | 1983 | markOopDesc::epoch_mask_in_place); |
kvn@855 | 1984 | Node* old = transform_later(new (C, 3) AndXNode(mark_node, cmask)); |
kvn@855 | 1985 | cast_thread = transform_later(new (C, 2) CastP2XNode(ctrl, thread)); |
kvn@855 | 1986 | Node* new_mark = transform_later(new (C, 3) OrXNode(cast_thread, old)); |
kvn@855 | 1987 | old_phi->init_req(1, old); |
kvn@855 | 1988 | new_phi->init_req(1, new_mark); |
kvn@855 | 1989 | |
kvn@855 | 1990 | transform_later(rebiased_region); |
kvn@855 | 1991 | transform_later(old_phi); |
kvn@855 | 1992 | transform_later(new_phi); |
kvn@855 | 1993 | |
kvn@855 | 1994 | // Try to acquire the bias of the object using an atomic operation. |
kvn@855 | 1995 | // If this fails we will go in to the runtime to revoke the object's bias. |
kvn@855 | 1996 | cas = new (C, 5) StoreXConditionalNode(rebiased_region, mem, adr, |
kvn@855 | 1997 | new_phi, old_phi); |
kvn@855 | 1998 | transform_later(cas); |
kvn@855 | 1999 | proj = transform_later( new (C, 1) SCMemProjNode(cas)); |
kvn@855 | 2000 | |
kvn@855 | 2001 | // Get slow path - Failed to CAS. |
kvn@855 | 2002 | not_biased_ctrl = opt_bits_test(rebiased_region, region, 4, cas, 0, 0); |
kvn@855 | 2003 | mem_phi->init_req(4, proj); |
kvn@855 | 2004 | // region->in(4) is set to fast path - the object is rebiased to the current thread. |
kvn@855 | 2005 | |
kvn@855 | 2006 | // Failed to CAS. |
kvn@855 | 2007 | slow_path = new (C, 3) RegionNode(3); |
kvn@855 | 2008 | Node *slow_mem = new (C, 3) PhiNode( slow_path, Type::MEMORY, TypeRawPtr::BOTTOM); |
kvn@855 | 2009 | |
kvn@855 | 2010 | slow_path->init_req(1, not_biased_ctrl); // Capture slow-control |
kvn@855 | 2011 | slow_mem->init_req(1, proj); |
kvn@855 | 2012 | |
kvn@855 | 2013 | // Call CAS-based locking scheme (FastLock node). |
kvn@855 | 2014 | |
kvn@855 | 2015 | transform_later(fast_lock_region); |
kvn@855 | 2016 | transform_later(fast_lock_mem_phi); |
kvn@855 | 2017 | |
kvn@855 | 2018 | // Get slow path - FastLock failed to lock the object. |
kvn@855 | 2019 | ctrl = opt_bits_test(fast_lock_region, region, 2, flock, 0, 0); |
kvn@855 | 2020 | mem_phi->init_req(2, fast_lock_mem_phi); |
kvn@855 | 2021 | // region->in(2) is set to fast path - the object is locked to the current thread. |
kvn@855 | 2022 | |
kvn@855 | 2023 | slow_path->init_req(2, ctrl); // Capture slow-control |
kvn@855 | 2024 | slow_mem->init_req(2, fast_lock_mem_phi); |
kvn@855 | 2025 | |
kvn@855 | 2026 | transform_later(slow_path); |
kvn@855 | 2027 | transform_later(slow_mem); |
kvn@855 | 2028 | // Reset lock's memory edge. |
kvn@855 | 2029 | lock->set_req(TypeFunc::Memory, slow_mem); |
kvn@855 | 2030 | |
kvn@855 | 2031 | } else { |
kvn@855 | 2032 | region = new (C, 3) RegionNode(3); |
kvn@855 | 2033 | // create a Phi for the memory state |
kvn@855 | 2034 | mem_phi = new (C, 3) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); |
kvn@855 | 2035 | |
kvn@855 | 2036 | // Optimize test; set region slot 2 |
kvn@855 | 2037 | slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0); |
kvn@855 | 2038 | mem_phi->init_req(2, mem); |
kvn@855 | 2039 | } |
duke@435 | 2040 | |
duke@435 | 2041 | // Make slow path call |
duke@435 | 2042 | CallNode *call = make_slow_call( (CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(), OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path, obj, box ); |
duke@435 | 2043 | |
duke@435 | 2044 | extract_call_projections(call); |
duke@435 | 2045 | |
duke@435 | 2046 | // Slow path can only throw asynchronous exceptions, which are always |
duke@435 | 2047 | // de-opted. So the compiler thinks the slow-call can never throw an |
duke@435 | 2048 | // exception. If it DOES throw an exception we would need the debug |
duke@435 | 2049 | // info removed first (since if it throws there is no monitor). |
duke@435 | 2050 | assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL && |
duke@435 | 2051 | _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock"); |
duke@435 | 2052 | |
duke@435 | 2053 | // Capture slow path |
duke@435 | 2054 | // disconnect fall-through projection from call and create a new one |
duke@435 | 2055 | // hook up users of fall-through projection to region |
duke@435 | 2056 | Node *slow_ctrl = _fallthroughproj->clone(); |
duke@435 | 2057 | transform_later(slow_ctrl); |
duke@435 | 2058 | _igvn.hash_delete(_fallthroughproj); |
duke@435 | 2059 | _fallthroughproj->disconnect_inputs(NULL); |
duke@435 | 2060 | region->init_req(1, slow_ctrl); |
duke@435 | 2061 | // region inputs are now complete |
duke@435 | 2062 | transform_later(region); |
kvn@1143 | 2063 | _igvn.replace_node(_fallthroughproj, region); |
duke@435 | 2064 | |
kvn@855 | 2065 | Node *memproj = transform_later( new(C, 1) ProjNode(call, TypeFunc::Memory) ); |
duke@435 | 2066 | mem_phi->init_req(1, memproj ); |
duke@435 | 2067 | transform_later(mem_phi); |
kvn@1143 | 2068 | _igvn.replace_node(_memproj_fallthrough, mem_phi); |
duke@435 | 2069 | } |
duke@435 | 2070 | |
duke@435 | 2071 | //------------------------------expand_unlock_node---------------------- |
duke@435 | 2072 | void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) { |
duke@435 | 2073 | |
kvn@501 | 2074 | Node* ctrl = unlock->in(TypeFunc::Control); |
duke@435 | 2075 | Node* mem = unlock->in(TypeFunc::Memory); |
duke@435 | 2076 | Node* obj = unlock->obj_node(); |
duke@435 | 2077 | Node* box = unlock->box_node(); |
duke@435 | 2078 | |
duke@435 | 2079 | // No need for a null check on unlock |
duke@435 | 2080 | |
duke@435 | 2081 | // Make the merge point |
kvn@855 | 2082 | Node *region; |
kvn@855 | 2083 | Node *mem_phi; |
kvn@855 | 2084 | |
kvn@855 | 2085 | if (UseOptoBiasInlining) { |
kvn@855 | 2086 | // Check for biased locking unlock case, which is a no-op. |
twisti@1040 | 2087 | // See the full description in MacroAssembler::biased_locking_exit(). |
kvn@855 | 2088 | region = new (C, 4) RegionNode(4); |
kvn@855 | 2089 | // create a Phi for the memory state |
kvn@855 | 2090 | mem_phi = new (C, 4) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); |
kvn@855 | 2091 | mem_phi->init_req(3, mem); |
kvn@855 | 2092 | |
kvn@855 | 2093 | Node* mark_node = make_load(ctrl, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type()); |
kvn@855 | 2094 | ctrl = opt_bits_test(ctrl, region, 3, mark_node, |
kvn@855 | 2095 | markOopDesc::biased_lock_mask_in_place, |
kvn@855 | 2096 | markOopDesc::biased_lock_pattern); |
kvn@855 | 2097 | } else { |
kvn@855 | 2098 | region = new (C, 3) RegionNode(3); |
kvn@855 | 2099 | // create a Phi for the memory state |
kvn@855 | 2100 | mem_phi = new (C, 3) PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM); |
kvn@855 | 2101 | } |
duke@435 | 2102 | |
duke@435 | 2103 | FastUnlockNode *funlock = new (C, 3) FastUnlockNode( ctrl, obj, box ); |
duke@435 | 2104 | funlock = transform_later( funlock )->as_FastUnlock(); |
duke@435 | 2105 | // Optimize test; set region slot 2 |
kvn@855 | 2106 | Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0); |
duke@435 | 2107 | |
duke@435 | 2108 | 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@435 | 2109 | |
duke@435 | 2110 | extract_call_projections(call); |
duke@435 | 2111 | |
duke@435 | 2112 | assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL && |
duke@435 | 2113 | _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock"); |
duke@435 | 2114 | |
duke@435 | 2115 | // No exceptions for unlocking |
duke@435 | 2116 | // Capture slow path |
duke@435 | 2117 | // disconnect fall-through projection from call and create a new one |
duke@435 | 2118 | // hook up users of fall-through projection to region |
duke@435 | 2119 | Node *slow_ctrl = _fallthroughproj->clone(); |
duke@435 | 2120 | transform_later(slow_ctrl); |
duke@435 | 2121 | _igvn.hash_delete(_fallthroughproj); |
duke@435 | 2122 | _fallthroughproj->disconnect_inputs(NULL); |
duke@435 | 2123 | region->init_req(1, slow_ctrl); |
duke@435 | 2124 | // region inputs are now complete |
duke@435 | 2125 | transform_later(region); |
kvn@1143 | 2126 | _igvn.replace_node(_fallthroughproj, region); |
duke@435 | 2127 | |
duke@435 | 2128 | Node *memproj = transform_later( new(C, 1) ProjNode(call, TypeFunc::Memory) ); |
duke@435 | 2129 | mem_phi->init_req(1, memproj ); |
duke@435 | 2130 | mem_phi->init_req(2, mem); |
duke@435 | 2131 | transform_later(mem_phi); |
kvn@1143 | 2132 | _igvn.replace_node(_memproj_fallthrough, mem_phi); |
duke@435 | 2133 | } |
duke@435 | 2134 | |
duke@435 | 2135 | //------------------------------expand_macro_nodes---------------------- |
duke@435 | 2136 | // Returns true if a failure occurred. |
duke@435 | 2137 | bool PhaseMacroExpand::expand_macro_nodes() { |
duke@435 | 2138 | if (C->macro_count() == 0) |
duke@435 | 2139 | return false; |
kvn@895 | 2140 | // First, attempt to eliminate locks |
kvn@508 | 2141 | bool progress = true; |
kvn@508 | 2142 | while (progress) { |
kvn@508 | 2143 | progress = false; |
kvn@508 | 2144 | for (int i = C->macro_count(); i > 0; i--) { |
kvn@508 | 2145 | Node * n = C->macro_node(i-1); |
kvn@508 | 2146 | bool success = false; |
kvn@508 | 2147 | debug_only(int old_macro_count = C->macro_count();); |
kvn@895 | 2148 | if (n->is_AbstractLock()) { |
kvn@895 | 2149 | success = eliminate_locking_node(n->as_AbstractLock()); |
kvn@895 | 2150 | } else if (n->Opcode() == Op_Opaque1 || n->Opcode() == Op_Opaque2) { |
kvn@1143 | 2151 | _igvn.replace_node(n, n->in(1)); |
kvn@895 | 2152 | success = true; |
kvn@895 | 2153 | } |
kvn@895 | 2154 | assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count"); |
kvn@895 | 2155 | progress = progress || success; |
kvn@895 | 2156 | } |
kvn@895 | 2157 | } |
kvn@895 | 2158 | // Next, attempt to eliminate allocations |
kvn@895 | 2159 | progress = true; |
kvn@895 | 2160 | while (progress) { |
kvn@895 | 2161 | progress = false; |
kvn@895 | 2162 | for (int i = C->macro_count(); i > 0; i--) { |
kvn@895 | 2163 | Node * n = C->macro_node(i-1); |
kvn@895 | 2164 | bool success = false; |
kvn@895 | 2165 | debug_only(int old_macro_count = C->macro_count();); |
kvn@508 | 2166 | switch (n->class_id()) { |
kvn@508 | 2167 | case Node::Class_Allocate: |
kvn@508 | 2168 | case Node::Class_AllocateArray: |
kvn@508 | 2169 | success = eliminate_allocate_node(n->as_Allocate()); |
kvn@508 | 2170 | break; |
kvn@508 | 2171 | case Node::Class_Lock: |
kvn@508 | 2172 | case Node::Class_Unlock: |
kvn@895 | 2173 | assert(!n->as_AbstractLock()->is_eliminated(), "sanity"); |
kvn@508 | 2174 | break; |
kvn@508 | 2175 | default: |
kvn@895 | 2176 | assert(false, "unknown node type in macro list"); |
kvn@508 | 2177 | } |
kvn@508 | 2178 | assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count"); |
kvn@508 | 2179 | progress = progress || success; |
kvn@508 | 2180 | } |
kvn@508 | 2181 | } |
kvn@508 | 2182 | // Make sure expansion will not cause node limit to be exceeded. |
kvn@508 | 2183 | // Worst case is a macro node gets expanded into about 50 nodes. |
kvn@508 | 2184 | // Allow 50% more for optimization. |
duke@435 | 2185 | if (C->check_node_count(C->macro_count() * 75, "out of nodes before macro expansion" ) ) |
duke@435 | 2186 | return true; |
kvn@508 | 2187 | |
duke@435 | 2188 | // expand "macro" nodes |
duke@435 | 2189 | // nodes are removed from the macro list as they are processed |
duke@435 | 2190 | while (C->macro_count() > 0) { |
kvn@508 | 2191 | int macro_count = C->macro_count(); |
kvn@508 | 2192 | Node * n = C->macro_node(macro_count-1); |
duke@435 | 2193 | assert(n->is_macro(), "only macro nodes expected here"); |
duke@435 | 2194 | if (_igvn.type(n) == Type::TOP || n->in(0)->is_top() ) { |
duke@435 | 2195 | // node is unreachable, so don't try to expand it |
duke@435 | 2196 | C->remove_macro_node(n); |
duke@435 | 2197 | continue; |
duke@435 | 2198 | } |
duke@435 | 2199 | switch (n->class_id()) { |
duke@435 | 2200 | case Node::Class_Allocate: |
duke@435 | 2201 | expand_allocate(n->as_Allocate()); |
duke@435 | 2202 | break; |
duke@435 | 2203 | case Node::Class_AllocateArray: |
duke@435 | 2204 | expand_allocate_array(n->as_AllocateArray()); |
duke@435 | 2205 | break; |
duke@435 | 2206 | case Node::Class_Lock: |
duke@435 | 2207 | expand_lock_node(n->as_Lock()); |
duke@435 | 2208 | break; |
duke@435 | 2209 | case Node::Class_Unlock: |
duke@435 | 2210 | expand_unlock_node(n->as_Unlock()); |
duke@435 | 2211 | break; |
duke@435 | 2212 | default: |
duke@435 | 2213 | assert(false, "unknown node type in macro list"); |
duke@435 | 2214 | } |
kvn@508 | 2215 | assert(C->macro_count() < macro_count, "must have deleted a node from macro list"); |
duke@435 | 2216 | if (C->failing()) return true; |
duke@435 | 2217 | } |
coleenp@548 | 2218 | |
coleenp@548 | 2219 | _igvn.set_delay_transform(false); |
duke@435 | 2220 | _igvn.optimize(); |
duke@435 | 2221 | return false; |
duke@435 | 2222 | } |