1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/src/share/vm/opto/memnode.cpp Wed Apr 27 01:25:04 2016 +0800
1.3 @@ -0,0 +1,4576 @@
1.4 +/*
1.5 + * Copyright (c) 1997, 2014, Oracle and/or its affiliates. All rights reserved.
1.6 + * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
1.7 + *
1.8 + * This code is free software; you can redistribute it and/or modify it
1.9 + * under the terms of the GNU General Public License version 2 only, as
1.10 + * published by the Free Software Foundation.
1.11 + *
1.12 + * This code is distributed in the hope that it will be useful, but WITHOUT
1.13 + * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
1.14 + * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
1.15 + * version 2 for more details (a copy is included in the LICENSE file that
1.16 + * accompanied this code).
1.17 + *
1.18 + * You should have received a copy of the GNU General Public License version
1.19 + * 2 along with this work; if not, write to the Free Software Foundation,
1.20 + * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
1.21 + *
1.22 + * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
1.23 + * or visit www.oracle.com if you need additional information or have any
1.24 + * questions.
1.25 + *
1.26 + */
1.27 +
1.28 +#include "precompiled.hpp"
1.29 +#include "classfile/systemDictionary.hpp"
1.30 +#include "compiler/compileLog.hpp"
1.31 +#include "memory/allocation.inline.hpp"
1.32 +#include "oops/objArrayKlass.hpp"
1.33 +#include "opto/addnode.hpp"
1.34 +#include "opto/cfgnode.hpp"
1.35 +#include "opto/compile.hpp"
1.36 +#include "opto/connode.hpp"
1.37 +#include "opto/loopnode.hpp"
1.38 +#include "opto/machnode.hpp"
1.39 +#include "opto/matcher.hpp"
1.40 +#include "opto/memnode.hpp"
1.41 +#include "opto/mulnode.hpp"
1.42 +#include "opto/phaseX.hpp"
1.43 +#include "opto/regmask.hpp"
1.44 +
1.45 +// Portions of code courtesy of Clifford Click
1.46 +
1.47 +// Optimization - Graph Style
1.48 +
1.49 +static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st);
1.50 +
1.51 +//=============================================================================
1.52 +uint MemNode::size_of() const { return sizeof(*this); }
1.53 +
1.54 +const TypePtr *MemNode::adr_type() const {
1.55 + Node* adr = in(Address);
1.56 + const TypePtr* cross_check = NULL;
1.57 + DEBUG_ONLY(cross_check = _adr_type);
1.58 + return calculate_adr_type(adr->bottom_type(), cross_check);
1.59 +}
1.60 +
1.61 +#ifndef PRODUCT
1.62 +void MemNode::dump_spec(outputStream *st) const {
1.63 + if (in(Address) == NULL) return; // node is dead
1.64 +#ifndef ASSERT
1.65 + // fake the missing field
1.66 + const TypePtr* _adr_type = NULL;
1.67 + if (in(Address) != NULL)
1.68 + _adr_type = in(Address)->bottom_type()->isa_ptr();
1.69 +#endif
1.70 + dump_adr_type(this, _adr_type, st);
1.71 +
1.72 + Compile* C = Compile::current();
1.73 + if( C->alias_type(_adr_type)->is_volatile() )
1.74 + st->print(" Volatile!");
1.75 +}
1.76 +
1.77 +void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) {
1.78 + st->print(" @");
1.79 + if (adr_type == NULL) {
1.80 + st->print("NULL");
1.81 + } else {
1.82 + adr_type->dump_on(st);
1.83 + Compile* C = Compile::current();
1.84 + Compile::AliasType* atp = NULL;
1.85 + if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type);
1.86 + if (atp == NULL)
1.87 + st->print(", idx=?\?;");
1.88 + else if (atp->index() == Compile::AliasIdxBot)
1.89 + st->print(", idx=Bot;");
1.90 + else if (atp->index() == Compile::AliasIdxTop)
1.91 + st->print(", idx=Top;");
1.92 + else if (atp->index() == Compile::AliasIdxRaw)
1.93 + st->print(", idx=Raw;");
1.94 + else {
1.95 + ciField* field = atp->field();
1.96 + if (field) {
1.97 + st->print(", name=");
1.98 + field->print_name_on(st);
1.99 + }
1.100 + st->print(", idx=%d;", atp->index());
1.101 + }
1.102 + }
1.103 +}
1.104 +
1.105 +extern void print_alias_types();
1.106 +
1.107 +#endif
1.108 +
1.109 +Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypeOopPtr *t_oop, Node *load, PhaseGVN *phase) {
1.110 + assert((t_oop != NULL), "sanity");
1.111 + bool is_instance = t_oop->is_known_instance_field();
1.112 + bool is_boxed_value_load = t_oop->is_ptr_to_boxed_value() &&
1.113 + (load != NULL) && load->is_Load() &&
1.114 + (phase->is_IterGVN() != NULL);
1.115 + if (!(is_instance || is_boxed_value_load))
1.116 + return mchain; // don't try to optimize non-instance types
1.117 + uint instance_id = t_oop->instance_id();
1.118 + Node *start_mem = phase->C->start()->proj_out(TypeFunc::Memory);
1.119 + Node *prev = NULL;
1.120 + Node *result = mchain;
1.121 + while (prev != result) {
1.122 + prev = result;
1.123 + if (result == start_mem)
1.124 + break; // hit one of our sentinels
1.125 + // skip over a call which does not affect this memory slice
1.126 + if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
1.127 + Node *proj_in = result->in(0);
1.128 + if (proj_in->is_Allocate() && proj_in->_idx == instance_id) {
1.129 + break; // hit one of our sentinels
1.130 + } else if (proj_in->is_Call()) {
1.131 + CallNode *call = proj_in->as_Call();
1.132 + if (!call->may_modify(t_oop, phase)) { // returns false for instances
1.133 + result = call->in(TypeFunc::Memory);
1.134 + }
1.135 + } else if (proj_in->is_Initialize()) {
1.136 + AllocateNode* alloc = proj_in->as_Initialize()->allocation();
1.137 + // Stop if this is the initialization for the object instance which
1.138 + // which contains this memory slice, otherwise skip over it.
1.139 + if ((alloc == NULL) || (alloc->_idx == instance_id)) {
1.140 + break;
1.141 + }
1.142 + if (is_instance) {
1.143 + result = proj_in->in(TypeFunc::Memory);
1.144 + } else if (is_boxed_value_load) {
1.145 + Node* klass = alloc->in(AllocateNode::KlassNode);
1.146 + const TypeKlassPtr* tklass = phase->type(klass)->is_klassptr();
1.147 + if (tklass->klass_is_exact() && !tklass->klass()->equals(t_oop->klass())) {
1.148 + result = proj_in->in(TypeFunc::Memory); // not related allocation
1.149 + }
1.150 + }
1.151 + } else if (proj_in->is_MemBar()) {
1.152 + result = proj_in->in(TypeFunc::Memory);
1.153 + } else {
1.154 + assert(false, "unexpected projection");
1.155 + }
1.156 + } else if (result->is_ClearArray()) {
1.157 + if (!is_instance || !ClearArrayNode::step_through(&result, instance_id, phase)) {
1.158 + // Can not bypass initialization of the instance
1.159 + // we are looking for.
1.160 + break;
1.161 + }
1.162 + // Otherwise skip it (the call updated 'result' value).
1.163 + } else if (result->is_MergeMem()) {
1.164 + result = step_through_mergemem(phase, result->as_MergeMem(), t_oop, NULL, tty);
1.165 + }
1.166 + }
1.167 + return result;
1.168 +}
1.169 +
1.170 +Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, Node *load, PhaseGVN *phase) {
1.171 + const TypeOopPtr* t_oop = t_adr->isa_oopptr();
1.172 + if (t_oop == NULL)
1.173 + return mchain; // don't try to optimize non-oop types
1.174 + Node* result = optimize_simple_memory_chain(mchain, t_oop, load, phase);
1.175 + bool is_instance = t_oop->is_known_instance_field();
1.176 + PhaseIterGVN *igvn = phase->is_IterGVN();
1.177 + if (is_instance && igvn != NULL && result->is_Phi()) {
1.178 + PhiNode *mphi = result->as_Phi();
1.179 + assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
1.180 + const TypePtr *t = mphi->adr_type();
1.181 + if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM ||
1.182 + t->isa_oopptr() && !t->is_oopptr()->is_known_instance() &&
1.183 + t->is_oopptr()->cast_to_exactness(true)
1.184 + ->is_oopptr()->cast_to_ptr_type(t_oop->ptr())
1.185 + ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop) {
1.186 + // clone the Phi with our address type
1.187 + result = mphi->split_out_instance(t_adr, igvn);
1.188 + } else {
1.189 + assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain");
1.190 + }
1.191 + }
1.192 + return result;
1.193 +}
1.194 +
1.195 +static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) {
1.196 + uint alias_idx = phase->C->get_alias_index(tp);
1.197 + Node *mem = mmem;
1.198 +#ifdef ASSERT
1.199 + {
1.200 + // Check that current type is consistent with the alias index used during graph construction
1.201 + assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
1.202 + bool consistent = adr_check == NULL || adr_check->empty() ||
1.203 + phase->C->must_alias(adr_check, alias_idx );
1.204 + // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
1.205 + if( !consistent && adr_check != NULL && !adr_check->empty() &&
1.206 + tp->isa_aryptr() && tp->offset() == Type::OffsetBot &&
1.207 + adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot &&
1.208 + ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() ||
1.209 + adr_check->offset() == oopDesc::klass_offset_in_bytes() ||
1.210 + adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) {
1.211 + // don't assert if it is dead code.
1.212 + consistent = true;
1.213 + }
1.214 + if( !consistent ) {
1.215 + st->print("alias_idx==%d, adr_check==", alias_idx);
1.216 + if( adr_check == NULL ) {
1.217 + st->print("NULL");
1.218 + } else {
1.219 + adr_check->dump();
1.220 + }
1.221 + st->cr();
1.222 + print_alias_types();
1.223 + assert(consistent, "adr_check must match alias idx");
1.224 + }
1.225 + }
1.226 +#endif
1.227 + // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
1.228 + // means an array I have not precisely typed yet. Do not do any
1.229 + // alias stuff with it any time soon.
1.230 + const TypeOopPtr *toop = tp->isa_oopptr();
1.231 + if( tp->base() != Type::AnyPtr &&
1.232 + !(toop &&
1.233 + toop->klass() != NULL &&
1.234 + toop->klass()->is_java_lang_Object() &&
1.235 + toop->offset() == Type::OffsetBot) ) {
1.236 + // compress paths and change unreachable cycles to TOP
1.237 + // If not, we can update the input infinitely along a MergeMem cycle
1.238 + // Equivalent code in PhiNode::Ideal
1.239 + Node* m = phase->transform(mmem);
1.240 + // If transformed to a MergeMem, get the desired slice
1.241 + // Otherwise the returned node represents memory for every slice
1.242 + mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m;
1.243 + // Update input if it is progress over what we have now
1.244 + }
1.245 + return mem;
1.246 +}
1.247 +
1.248 +//--------------------------Ideal_common---------------------------------------
1.249 +// Look for degenerate control and memory inputs. Bypass MergeMem inputs.
1.250 +// Unhook non-raw memories from complete (macro-expanded) initializations.
1.251 +Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
1.252 + // If our control input is a dead region, kill all below the region
1.253 + Node *ctl = in(MemNode::Control);
1.254 + if (ctl && remove_dead_region(phase, can_reshape))
1.255 + return this;
1.256 + ctl = in(MemNode::Control);
1.257 + // Don't bother trying to transform a dead node
1.258 + if (ctl && ctl->is_top()) return NodeSentinel;
1.259 +
1.260 + PhaseIterGVN *igvn = phase->is_IterGVN();
1.261 + // Wait if control on the worklist.
1.262 + if (ctl && can_reshape && igvn != NULL) {
1.263 + Node* bol = NULL;
1.264 + Node* cmp = NULL;
1.265 + if (ctl->in(0)->is_If()) {
1.266 + assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity");
1.267 + bol = ctl->in(0)->in(1);
1.268 + if (bol->is_Bool())
1.269 + cmp = ctl->in(0)->in(1)->in(1);
1.270 + }
1.271 + if (igvn->_worklist.member(ctl) ||
1.272 + (bol != NULL && igvn->_worklist.member(bol)) ||
1.273 + (cmp != NULL && igvn->_worklist.member(cmp)) ) {
1.274 + // This control path may be dead.
1.275 + // Delay this memory node transformation until the control is processed.
1.276 + phase->is_IterGVN()->_worklist.push(this);
1.277 + return NodeSentinel; // caller will return NULL
1.278 + }
1.279 + }
1.280 + // Ignore if memory is dead, or self-loop
1.281 + Node *mem = in(MemNode::Memory);
1.282 + if (phase->type( mem ) == Type::TOP) return NodeSentinel; // caller will return NULL
1.283 + assert(mem != this, "dead loop in MemNode::Ideal");
1.284 +
1.285 + if (can_reshape && igvn != NULL && igvn->_worklist.member(mem)) {
1.286 + // This memory slice may be dead.
1.287 + // Delay this mem node transformation until the memory is processed.
1.288 + phase->is_IterGVN()->_worklist.push(this);
1.289 + return NodeSentinel; // caller will return NULL
1.290 + }
1.291 +
1.292 + Node *address = in(MemNode::Address);
1.293 + const Type *t_adr = phase->type(address);
1.294 + if (t_adr == Type::TOP) return NodeSentinel; // caller will return NULL
1.295 +
1.296 + if (can_reshape && igvn != NULL &&
1.297 + (igvn->_worklist.member(address) ||
1.298 + igvn->_worklist.size() > 0 && (t_adr != adr_type())) ) {
1.299 + // The address's base and type may change when the address is processed.
1.300 + // Delay this mem node transformation until the address is processed.
1.301 + phase->is_IterGVN()->_worklist.push(this);
1.302 + return NodeSentinel; // caller will return NULL
1.303 + }
1.304 +
1.305 + // Do NOT remove or optimize the next lines: ensure a new alias index
1.306 + // is allocated for an oop pointer type before Escape Analysis.
1.307 + // Note: C++ will not remove it since the call has side effect.
1.308 + if (t_adr->isa_oopptr()) {
1.309 + int alias_idx = phase->C->get_alias_index(t_adr->is_ptr());
1.310 + }
1.311 +
1.312 + Node* base = NULL;
1.313 + if (address->is_AddP()) {
1.314 + base = address->in(AddPNode::Base);
1.315 + }
1.316 + if (base != NULL && phase->type(base)->higher_equal(TypePtr::NULL_PTR) &&
1.317 + !t_adr->isa_rawptr()) {
1.318 + // Note: raw address has TOP base and top->higher_equal(TypePtr::NULL_PTR) is true.
1.319 + // Skip this node optimization if its address has TOP base.
1.320 + return NodeSentinel; // caller will return NULL
1.321 + }
1.322 +
1.323 + // Avoid independent memory operations
1.324 + Node* old_mem = mem;
1.325 +
1.326 + // The code which unhooks non-raw memories from complete (macro-expanded)
1.327 + // initializations was removed. After macro-expansion all stores catched
1.328 + // by Initialize node became raw stores and there is no information
1.329 + // which memory slices they modify. So it is unsafe to move any memory
1.330 + // operation above these stores. Also in most cases hooked non-raw memories
1.331 + // were already unhooked by using information from detect_ptr_independence()
1.332 + // and find_previous_store().
1.333 +
1.334 + if (mem->is_MergeMem()) {
1.335 + MergeMemNode* mmem = mem->as_MergeMem();
1.336 + const TypePtr *tp = t_adr->is_ptr();
1.337 +
1.338 + mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty);
1.339 + }
1.340 +
1.341 + if (mem != old_mem) {
1.342 + set_req(MemNode::Memory, mem);
1.343 + if (can_reshape && old_mem->outcnt() == 0) {
1.344 + igvn->_worklist.push(old_mem);
1.345 + }
1.346 + if (phase->type( mem ) == Type::TOP) return NodeSentinel;
1.347 + return this;
1.348 + }
1.349 +
1.350 + // let the subclass continue analyzing...
1.351 + return NULL;
1.352 +}
1.353 +
1.354 +// Helper function for proving some simple control dominations.
1.355 +// Attempt to prove that all control inputs of 'dom' dominate 'sub'.
1.356 +// Already assumes that 'dom' is available at 'sub', and that 'sub'
1.357 +// is not a constant (dominated by the method's StartNode).
1.358 +// Used by MemNode::find_previous_store to prove that the
1.359 +// control input of a memory operation predates (dominates)
1.360 +// an allocation it wants to look past.
1.361 +bool MemNode::all_controls_dominate(Node* dom, Node* sub) {
1.362 + if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top())
1.363 + return false; // Conservative answer for dead code
1.364 +
1.365 + // Check 'dom'. Skip Proj and CatchProj nodes.
1.366 + dom = dom->find_exact_control(dom);
1.367 + if (dom == NULL || dom->is_top())
1.368 + return false; // Conservative answer for dead code
1.369 +
1.370 + if (dom == sub) {
1.371 + // For the case when, for example, 'sub' is Initialize and the original
1.372 + // 'dom' is Proj node of the 'sub'.
1.373 + return false;
1.374 + }
1.375 +
1.376 + if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub)
1.377 + return true;
1.378 +
1.379 + // 'dom' dominates 'sub' if its control edge and control edges
1.380 + // of all its inputs dominate or equal to sub's control edge.
1.381 +
1.382 + // Currently 'sub' is either Allocate, Initialize or Start nodes.
1.383 + // Or Region for the check in LoadNode::Ideal();
1.384 + // 'sub' should have sub->in(0) != NULL.
1.385 + assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() ||
1.386 + sub->is_Region() || sub->is_Call(), "expecting only these nodes");
1.387 +
1.388 + // Get control edge of 'sub'.
1.389 + Node* orig_sub = sub;
1.390 + sub = sub->find_exact_control(sub->in(0));
1.391 + if (sub == NULL || sub->is_top())
1.392 + return false; // Conservative answer for dead code
1.393 +
1.394 + assert(sub->is_CFG(), "expecting control");
1.395 +
1.396 + if (sub == dom)
1.397 + return true;
1.398 +
1.399 + if (sub->is_Start() || sub->is_Root())
1.400 + return false;
1.401 +
1.402 + {
1.403 + // Check all control edges of 'dom'.
1.404 +
1.405 + ResourceMark rm;
1.406 + Arena* arena = Thread::current()->resource_area();
1.407 + Node_List nlist(arena);
1.408 + Unique_Node_List dom_list(arena);
1.409 +
1.410 + dom_list.push(dom);
1.411 + bool only_dominating_controls = false;
1.412 +
1.413 + for (uint next = 0; next < dom_list.size(); next++) {
1.414 + Node* n = dom_list.at(next);
1.415 + if (n == orig_sub)
1.416 + return false; // One of dom's inputs dominated by sub.
1.417 + if (!n->is_CFG() && n->pinned()) {
1.418 + // Check only own control edge for pinned non-control nodes.
1.419 + n = n->find_exact_control(n->in(0));
1.420 + if (n == NULL || n->is_top())
1.421 + return false; // Conservative answer for dead code
1.422 + assert(n->is_CFG(), "expecting control");
1.423 + dom_list.push(n);
1.424 + } else if (n->is_Con() || n->is_Start() || n->is_Root()) {
1.425 + only_dominating_controls = true;
1.426 + } else if (n->is_CFG()) {
1.427 + if (n->dominates(sub, nlist))
1.428 + only_dominating_controls = true;
1.429 + else
1.430 + return false;
1.431 + } else {
1.432 + // First, own control edge.
1.433 + Node* m = n->find_exact_control(n->in(0));
1.434 + if (m != NULL) {
1.435 + if (m->is_top())
1.436 + return false; // Conservative answer for dead code
1.437 + dom_list.push(m);
1.438 + }
1.439 + // Now, the rest of edges.
1.440 + uint cnt = n->req();
1.441 + for (uint i = 1; i < cnt; i++) {
1.442 + m = n->find_exact_control(n->in(i));
1.443 + if (m == NULL || m->is_top())
1.444 + continue;
1.445 + dom_list.push(m);
1.446 + }
1.447 + }
1.448 + }
1.449 + return only_dominating_controls;
1.450 + }
1.451 +}
1.452 +
1.453 +//---------------------detect_ptr_independence---------------------------------
1.454 +// Used by MemNode::find_previous_store to prove that two base
1.455 +// pointers are never equal.
1.456 +// The pointers are accompanied by their associated allocations,
1.457 +// if any, which have been previously discovered by the caller.
1.458 +bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
1.459 + Node* p2, AllocateNode* a2,
1.460 + PhaseTransform* phase) {
1.461 + // Attempt to prove that these two pointers cannot be aliased.
1.462 + // They may both manifestly be allocations, and they should differ.
1.463 + // Or, if they are not both allocations, they can be distinct constants.
1.464 + // Otherwise, one is an allocation and the other a pre-existing value.
1.465 + if (a1 == NULL && a2 == NULL) { // neither an allocation
1.466 + return (p1 != p2) && p1->is_Con() && p2->is_Con();
1.467 + } else if (a1 != NULL && a2 != NULL) { // both allocations
1.468 + return (a1 != a2);
1.469 + } else if (a1 != NULL) { // one allocation a1
1.470 + // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.)
1.471 + return all_controls_dominate(p2, a1);
1.472 + } else { //(a2 != NULL) // one allocation a2
1.473 + return all_controls_dominate(p1, a2);
1.474 + }
1.475 + return false;
1.476 +}
1.477 +
1.478 +
1.479 +// The logic for reordering loads and stores uses four steps:
1.480 +// (a) Walk carefully past stores and initializations which we
1.481 +// can prove are independent of this load.
1.482 +// (b) Observe that the next memory state makes an exact match
1.483 +// with self (load or store), and locate the relevant store.
1.484 +// (c) Ensure that, if we were to wire self directly to the store,
1.485 +// the optimizer would fold it up somehow.
1.486 +// (d) Do the rewiring, and return, depending on some other part of
1.487 +// the optimizer to fold up the load.
1.488 +// This routine handles steps (a) and (b). Steps (c) and (d) are
1.489 +// specific to loads and stores, so they are handled by the callers.
1.490 +// (Currently, only LoadNode::Ideal has steps (c), (d). More later.)
1.491 +//
1.492 +Node* MemNode::find_previous_store(PhaseTransform* phase) {
1.493 + Node* ctrl = in(MemNode::Control);
1.494 + Node* adr = in(MemNode::Address);
1.495 + intptr_t offset = 0;
1.496 + Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
1.497 + AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
1.498 +
1.499 + if (offset == Type::OffsetBot)
1.500 + return NULL; // cannot unalias unless there are precise offsets
1.501 +
1.502 + const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr();
1.503 +
1.504 + intptr_t size_in_bytes = memory_size();
1.505 +
1.506 + Node* mem = in(MemNode::Memory); // start searching here...
1.507 +
1.508 + int cnt = 50; // Cycle limiter
1.509 + for (;;) { // While we can dance past unrelated stores...
1.510 + if (--cnt < 0) break; // Caught in cycle or a complicated dance?
1.511 +
1.512 + if (mem->is_Store()) {
1.513 + Node* st_adr = mem->in(MemNode::Address);
1.514 + intptr_t st_offset = 0;
1.515 + Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
1.516 + if (st_base == NULL)
1.517 + break; // inscrutable pointer
1.518 + if (st_offset != offset && st_offset != Type::OffsetBot) {
1.519 + const int MAX_STORE = BytesPerLong;
1.520 + if (st_offset >= offset + size_in_bytes ||
1.521 + st_offset <= offset - MAX_STORE ||
1.522 + st_offset <= offset - mem->as_Store()->memory_size()) {
1.523 + // Success: The offsets are provably independent.
1.524 + // (You may ask, why not just test st_offset != offset and be done?
1.525 + // The answer is that stores of different sizes can co-exist
1.526 + // in the same sequence of RawMem effects. We sometimes initialize
1.527 + // a whole 'tile' of array elements with a single jint or jlong.)
1.528 + mem = mem->in(MemNode::Memory);
1.529 + continue; // (a) advance through independent store memory
1.530 + }
1.531 + }
1.532 + if (st_base != base &&
1.533 + detect_ptr_independence(base, alloc,
1.534 + st_base,
1.535 + AllocateNode::Ideal_allocation(st_base, phase),
1.536 + phase)) {
1.537 + // Success: The bases are provably independent.
1.538 + mem = mem->in(MemNode::Memory);
1.539 + continue; // (a) advance through independent store memory
1.540 + }
1.541 +
1.542 + // (b) At this point, if the bases or offsets do not agree, we lose,
1.543 + // since we have not managed to prove 'this' and 'mem' independent.
1.544 + if (st_base == base && st_offset == offset) {
1.545 + return mem; // let caller handle steps (c), (d)
1.546 + }
1.547 +
1.548 + } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
1.549 + InitializeNode* st_init = mem->in(0)->as_Initialize();
1.550 + AllocateNode* st_alloc = st_init->allocation();
1.551 + if (st_alloc == NULL)
1.552 + break; // something degenerated
1.553 + bool known_identical = false;
1.554 + bool known_independent = false;
1.555 + if (alloc == st_alloc)
1.556 + known_identical = true;
1.557 + else if (alloc != NULL)
1.558 + known_independent = true;
1.559 + else if (all_controls_dominate(this, st_alloc))
1.560 + known_independent = true;
1.561 +
1.562 + if (known_independent) {
1.563 + // The bases are provably independent: Either they are
1.564 + // manifestly distinct allocations, or else the control
1.565 + // of this load dominates the store's allocation.
1.566 + int alias_idx = phase->C->get_alias_index(adr_type());
1.567 + if (alias_idx == Compile::AliasIdxRaw) {
1.568 + mem = st_alloc->in(TypeFunc::Memory);
1.569 + } else {
1.570 + mem = st_init->memory(alias_idx);
1.571 + }
1.572 + continue; // (a) advance through independent store memory
1.573 + }
1.574 +
1.575 + // (b) at this point, if we are not looking at a store initializing
1.576 + // the same allocation we are loading from, we lose.
1.577 + if (known_identical) {
1.578 + // From caller, can_see_stored_value will consult find_captured_store.
1.579 + return mem; // let caller handle steps (c), (d)
1.580 + }
1.581 +
1.582 + } else if (addr_t != NULL && addr_t->is_known_instance_field()) {
1.583 + // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
1.584 + if (mem->is_Proj() && mem->in(0)->is_Call()) {
1.585 + CallNode *call = mem->in(0)->as_Call();
1.586 + if (!call->may_modify(addr_t, phase)) {
1.587 + mem = call->in(TypeFunc::Memory);
1.588 + continue; // (a) advance through independent call memory
1.589 + }
1.590 + } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) {
1.591 + mem = mem->in(0)->in(TypeFunc::Memory);
1.592 + continue; // (a) advance through independent MemBar memory
1.593 + } else if (mem->is_ClearArray()) {
1.594 + if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) {
1.595 + // (the call updated 'mem' value)
1.596 + continue; // (a) advance through independent allocation memory
1.597 + } else {
1.598 + // Can not bypass initialization of the instance
1.599 + // we are looking for.
1.600 + return mem;
1.601 + }
1.602 + } else if (mem->is_MergeMem()) {
1.603 + int alias_idx = phase->C->get_alias_index(adr_type());
1.604 + mem = mem->as_MergeMem()->memory_at(alias_idx);
1.605 + continue; // (a) advance through independent MergeMem memory
1.606 + }
1.607 + }
1.608 +
1.609 + // Unless there is an explicit 'continue', we must bail out here,
1.610 + // because 'mem' is an inscrutable memory state (e.g., a call).
1.611 + break;
1.612 + }
1.613 +
1.614 + return NULL; // bail out
1.615 +}
1.616 +
1.617 +//----------------------calculate_adr_type-------------------------------------
1.618 +// Helper function. Notices when the given type of address hits top or bottom.
1.619 +// Also, asserts a cross-check of the type against the expected address type.
1.620 +const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
1.621 + if (t == Type::TOP) return NULL; // does not touch memory any more?
1.622 + #ifdef PRODUCT
1.623 + cross_check = NULL;
1.624 + #else
1.625 + if (!VerifyAliases || is_error_reported() || Node::in_dump()) cross_check = NULL;
1.626 + #endif
1.627 + const TypePtr* tp = t->isa_ptr();
1.628 + if (tp == NULL) {
1.629 + assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
1.630 + return TypePtr::BOTTOM; // touches lots of memory
1.631 + } else {
1.632 + #ifdef ASSERT
1.633 + // %%%% [phh] We don't check the alias index if cross_check is
1.634 + // TypeRawPtr::BOTTOM. Needs to be investigated.
1.635 + if (cross_check != NULL &&
1.636 + cross_check != TypePtr::BOTTOM &&
1.637 + cross_check != TypeRawPtr::BOTTOM) {
1.638 + // Recheck the alias index, to see if it has changed (due to a bug).
1.639 + Compile* C = Compile::current();
1.640 + assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
1.641 + "must stay in the original alias category");
1.642 + // The type of the address must be contained in the adr_type,
1.643 + // disregarding "null"-ness.
1.644 + // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
1.645 + const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
1.646 + assert(cross_check->meet(tp_notnull) == cross_check->remove_speculative(),
1.647 + "real address must not escape from expected memory type");
1.648 + }
1.649 + #endif
1.650 + return tp;
1.651 + }
1.652 +}
1.653 +
1.654 +//------------------------adr_phi_is_loop_invariant----------------------------
1.655 +// A helper function for Ideal_DU_postCCP to check if a Phi in a counted
1.656 +// loop is loop invariant. Make a quick traversal of Phi and associated
1.657 +// CastPP nodes, looking to see if they are a closed group within the loop.
1.658 +bool MemNode::adr_phi_is_loop_invariant(Node* adr_phi, Node* cast) {
1.659 + // The idea is that the phi-nest must boil down to only CastPP nodes
1.660 + // with the same data. This implies that any path into the loop already
1.661 + // includes such a CastPP, and so the original cast, whatever its input,
1.662 + // must be covered by an equivalent cast, with an earlier control input.
1.663 + ResourceMark rm;
1.664 +
1.665 + // The loop entry input of the phi should be the unique dominating
1.666 + // node for every Phi/CastPP in the loop.
1.667 + Unique_Node_List closure;
1.668 + closure.push(adr_phi->in(LoopNode::EntryControl));
1.669 +
1.670 + // Add the phi node and the cast to the worklist.
1.671 + Unique_Node_List worklist;
1.672 + worklist.push(adr_phi);
1.673 + if( cast != NULL ){
1.674 + if( !cast->is_ConstraintCast() ) return false;
1.675 + worklist.push(cast);
1.676 + }
1.677 +
1.678 + // Begin recursive walk of phi nodes.
1.679 + while( worklist.size() ){
1.680 + // Take a node off the worklist
1.681 + Node *n = worklist.pop();
1.682 + if( !closure.member(n) ){
1.683 + // Add it to the closure.
1.684 + closure.push(n);
1.685 + // Make a sanity check to ensure we don't waste too much time here.
1.686 + if( closure.size() > 20) return false;
1.687 + // This node is OK if:
1.688 + // - it is a cast of an identical value
1.689 + // - or it is a phi node (then we add its inputs to the worklist)
1.690 + // Otherwise, the node is not OK, and we presume the cast is not invariant
1.691 + if( n->is_ConstraintCast() ){
1.692 + worklist.push(n->in(1));
1.693 + } else if( n->is_Phi() ) {
1.694 + for( uint i = 1; i < n->req(); i++ ) {
1.695 + worklist.push(n->in(i));
1.696 + }
1.697 + } else {
1.698 + return false;
1.699 + }
1.700 + }
1.701 + }
1.702 +
1.703 + // Quit when the worklist is empty, and we've found no offending nodes.
1.704 + return true;
1.705 +}
1.706 +
1.707 +//------------------------------Ideal_DU_postCCP-------------------------------
1.708 +// Find any cast-away of null-ness and keep its control. Null cast-aways are
1.709 +// going away in this pass and we need to make this memory op depend on the
1.710 +// gating null check.
1.711 +Node *MemNode::Ideal_DU_postCCP( PhaseCCP *ccp ) {
1.712 + return Ideal_common_DU_postCCP(ccp, this, in(MemNode::Address));
1.713 +}
1.714 +
1.715 +// I tried to leave the CastPP's in. This makes the graph more accurate in
1.716 +// some sense; we get to keep around the knowledge that an oop is not-null
1.717 +// after some test. Alas, the CastPP's interfere with GVN (some values are
1.718 +// the regular oop, some are the CastPP of the oop, all merge at Phi's which
1.719 +// cannot collapse, etc). This cost us 10% on SpecJVM, even when I removed
1.720 +// some of the more trivial cases in the optimizer. Removing more useless
1.721 +// Phi's started allowing Loads to illegally float above null checks. I gave
1.722 +// up on this approach. CNC 10/20/2000
1.723 +// This static method may be called not from MemNode (EncodePNode calls it).
1.724 +// Only the control edge of the node 'n' might be updated.
1.725 +Node *MemNode::Ideal_common_DU_postCCP( PhaseCCP *ccp, Node* n, Node* adr ) {
1.726 + Node *skipped_cast = NULL;
1.727 + // Need a null check? Regular static accesses do not because they are
1.728 + // from constant addresses. Array ops are gated by the range check (which
1.729 + // always includes a NULL check). Just check field ops.
1.730 + if( n->in(MemNode::Control) == NULL ) {
1.731 + // Scan upwards for the highest location we can place this memory op.
1.732 + while( true ) {
1.733 + switch( adr->Opcode() ) {
1.734 +
1.735 + case Op_AddP: // No change to NULL-ness, so peek thru AddP's
1.736 + adr = adr->in(AddPNode::Base);
1.737 + continue;
1.738 +
1.739 + case Op_DecodeN: // No change to NULL-ness, so peek thru
1.740 + case Op_DecodeNKlass:
1.741 + adr = adr->in(1);
1.742 + continue;
1.743 +
1.744 + case Op_EncodeP:
1.745 + case Op_EncodePKlass:
1.746 + // EncodeP node's control edge could be set by this method
1.747 + // when EncodeP node depends on CastPP node.
1.748 + //
1.749 + // Use its control edge for memory op because EncodeP may go away
1.750 + // later when it is folded with following or preceding DecodeN node.
1.751 + if (adr->in(0) == NULL) {
1.752 + // Keep looking for cast nodes.
1.753 + adr = adr->in(1);
1.754 + continue;
1.755 + }
1.756 + ccp->hash_delete(n);
1.757 + n->set_req(MemNode::Control, adr->in(0));
1.758 + ccp->hash_insert(n);
1.759 + return n;
1.760 +
1.761 + case Op_CastPP:
1.762 + // If the CastPP is useless, just peek on through it.
1.763 + if( ccp->type(adr) == ccp->type(adr->in(1)) ) {
1.764 + // Remember the cast that we've peeked though. If we peek
1.765 + // through more than one, then we end up remembering the highest
1.766 + // one, that is, if in a loop, the one closest to the top.
1.767 + skipped_cast = adr;
1.768 + adr = adr->in(1);
1.769 + continue;
1.770 + }
1.771 + // CastPP is going away in this pass! We need this memory op to be
1.772 + // control-dependent on the test that is guarding the CastPP.
1.773 + ccp->hash_delete(n);
1.774 + n->set_req(MemNode::Control, adr->in(0));
1.775 + ccp->hash_insert(n);
1.776 + return n;
1.777 +
1.778 + case Op_Phi:
1.779 + // Attempt to float above a Phi to some dominating point.
1.780 + if (adr->in(0) != NULL && adr->in(0)->is_CountedLoop()) {
1.781 + // If we've already peeked through a Cast (which could have set the
1.782 + // control), we can't float above a Phi, because the skipped Cast
1.783 + // may not be loop invariant.
1.784 + if (adr_phi_is_loop_invariant(adr, skipped_cast)) {
1.785 + adr = adr->in(1);
1.786 + continue;
1.787 + }
1.788 + }
1.789 +
1.790 + // Intentional fallthrough!
1.791 +
1.792 + // No obvious dominating point. The mem op is pinned below the Phi
1.793 + // by the Phi itself. If the Phi goes away (no true value is merged)
1.794 + // then the mem op can float, but not indefinitely. It must be pinned
1.795 + // behind the controls leading to the Phi.
1.796 + case Op_CheckCastPP:
1.797 + // These usually stick around to change address type, however a
1.798 + // useless one can be elided and we still need to pick up a control edge
1.799 + if (adr->in(0) == NULL) {
1.800 + // This CheckCastPP node has NO control and is likely useless. But we
1.801 + // need check further up the ancestor chain for a control input to keep
1.802 + // the node in place. 4959717.
1.803 + skipped_cast = adr;
1.804 + adr = adr->in(1);
1.805 + continue;
1.806 + }
1.807 + ccp->hash_delete(n);
1.808 + n->set_req(MemNode::Control, adr->in(0));
1.809 + ccp->hash_insert(n);
1.810 + return n;
1.811 +
1.812 + // List of "safe" opcodes; those that implicitly block the memory
1.813 + // op below any null check.
1.814 + case Op_CastX2P: // no null checks on native pointers
1.815 + case Op_Parm: // 'this' pointer is not null
1.816 + case Op_LoadP: // Loading from within a klass
1.817 + case Op_LoadN: // Loading from within a klass
1.818 + case Op_LoadKlass: // Loading from within a klass
1.819 + case Op_LoadNKlass: // Loading from within a klass
1.820 + case Op_ConP: // Loading from a klass
1.821 + case Op_ConN: // Loading from a klass
1.822 + case Op_ConNKlass: // Loading from a klass
1.823 + case Op_CreateEx: // Sucking up the guts of an exception oop
1.824 + case Op_Con: // Reading from TLS
1.825 + case Op_CMoveP: // CMoveP is pinned
1.826 + case Op_CMoveN: // CMoveN is pinned
1.827 + break; // No progress
1.828 +
1.829 + case Op_Proj: // Direct call to an allocation routine
1.830 + case Op_SCMemProj: // Memory state from store conditional ops
1.831 +#ifdef ASSERT
1.832 + {
1.833 + assert(adr->as_Proj()->_con == TypeFunc::Parms, "must be return value");
1.834 + const Node* call = adr->in(0);
1.835 + if (call->is_CallJava()) {
1.836 + const CallJavaNode* call_java = call->as_CallJava();
1.837 + const TypeTuple *r = call_java->tf()->range();
1.838 + assert(r->cnt() > TypeFunc::Parms, "must return value");
1.839 + const Type* ret_type = r->field_at(TypeFunc::Parms);
1.840 + assert(ret_type && ret_type->isa_ptr(), "must return pointer");
1.841 + // We further presume that this is one of
1.842 + // new_instance_Java, new_array_Java, or
1.843 + // the like, but do not assert for this.
1.844 + } else if (call->is_Allocate()) {
1.845 + // similar case to new_instance_Java, etc.
1.846 + } else if (!call->is_CallLeaf()) {
1.847 + // Projections from fetch_oop (OSR) are allowed as well.
1.848 + ShouldNotReachHere();
1.849 + }
1.850 + }
1.851 +#endif
1.852 + break;
1.853 + default:
1.854 + ShouldNotReachHere();
1.855 + }
1.856 + break;
1.857 + }
1.858 + }
1.859 +
1.860 + return NULL; // No progress
1.861 +}
1.862 +
1.863 +
1.864 +//=============================================================================
1.865 +uint LoadNode::size_of() const { return sizeof(*this); }
1.866 +uint LoadNode::cmp( const Node &n ) const
1.867 +{ return !Type::cmp( _type, ((LoadNode&)n)._type ); }
1.868 +const Type *LoadNode::bottom_type() const { return _type; }
1.869 +uint LoadNode::ideal_reg() const {
1.870 + return _type->ideal_reg();
1.871 +}
1.872 +
1.873 +#ifndef PRODUCT
1.874 +void LoadNode::dump_spec(outputStream *st) const {
1.875 + MemNode::dump_spec(st);
1.876 + if( !Verbose && !WizardMode ) {
1.877 + // standard dump does this in Verbose and WizardMode
1.878 + st->print(" #"); _type->dump_on(st);
1.879 + }
1.880 +}
1.881 +#endif
1.882 +
1.883 +#ifdef ASSERT
1.884 +//----------------------------is_immutable_value-------------------------------
1.885 +// Helper function to allow a raw load without control edge for some cases
1.886 +bool LoadNode::is_immutable_value(Node* adr) {
1.887 + return (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() &&
1.888 + adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
1.889 + (adr->in(AddPNode::Offset)->find_intptr_t_con(-1) ==
1.890 + in_bytes(JavaThread::osthread_offset())));
1.891 +}
1.892 +#endif
1.893 +
1.894 +//----------------------------LoadNode::make-----------------------------------
1.895 +// Polymorphic factory method:
1.896 +Node *LoadNode::make(PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt, MemOrd mo) {
1.897 + Compile* C = gvn.C;
1.898 +
1.899 + // sanity check the alias category against the created node type
1.900 + assert(!(adr_type->isa_oopptr() &&
1.901 + adr_type->offset() == oopDesc::klass_offset_in_bytes()),
1.902 + "use LoadKlassNode instead");
1.903 + assert(!(adr_type->isa_aryptr() &&
1.904 + adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
1.905 + "use LoadRangeNode instead");
1.906 + // Check control edge of raw loads
1.907 + assert( ctl != NULL || C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
1.908 + // oop will be recorded in oop map if load crosses safepoint
1.909 + rt->isa_oopptr() || is_immutable_value(adr),
1.910 + "raw memory operations should have control edge");
1.911 + switch (bt) {
1.912 + case T_BOOLEAN: return new (C) LoadUBNode(ctl, mem, adr, adr_type, rt->is_int(), mo);
1.913 + case T_BYTE: return new (C) LoadBNode (ctl, mem, adr, adr_type, rt->is_int(), mo);
1.914 + case T_INT: return new (C) LoadINode (ctl, mem, adr, adr_type, rt->is_int(), mo);
1.915 + case T_CHAR: return new (C) LoadUSNode(ctl, mem, adr, adr_type, rt->is_int(), mo);
1.916 + case T_SHORT: return new (C) LoadSNode (ctl, mem, adr, adr_type, rt->is_int(), mo);
1.917 + case T_LONG: return new (C) LoadLNode (ctl, mem, adr, adr_type, rt->is_long(), mo);
1.918 + case T_FLOAT: return new (C) LoadFNode (ctl, mem, adr, adr_type, rt, mo);
1.919 + case T_DOUBLE: return new (C) LoadDNode (ctl, mem, adr, adr_type, rt, mo);
1.920 + case T_ADDRESS: return new (C) LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr(), mo);
1.921 + case T_OBJECT:
1.922 +#ifdef _LP64
1.923 + if (adr->bottom_type()->is_ptr_to_narrowoop()) {
1.924 + Node* load = gvn.transform(new (C) LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop(), mo));
1.925 + return new (C) DecodeNNode(load, load->bottom_type()->make_ptr());
1.926 + } else
1.927 +#endif
1.928 + {
1.929 + assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop");
1.930 + return new (C) LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr(), mo);
1.931 + }
1.932 + }
1.933 + ShouldNotReachHere();
1.934 + return (LoadNode*)NULL;
1.935 +}
1.936 +
1.937 +LoadLNode* LoadLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo) {
1.938 + bool require_atomic = true;
1.939 + return new (C) LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), mo, require_atomic);
1.940 +}
1.941 +
1.942 +
1.943 +
1.944 +
1.945 +//------------------------------hash-------------------------------------------
1.946 +uint LoadNode::hash() const {
1.947 + // unroll addition of interesting fields
1.948 + return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
1.949 +}
1.950 +
1.951 +static bool skip_through_membars(Compile::AliasType* atp, const TypeInstPtr* tp, bool eliminate_boxing) {
1.952 + if ((atp != NULL) && (atp->index() >= Compile::AliasIdxRaw)) {
1.953 + bool non_volatile = (atp->field() != NULL) && !atp->field()->is_volatile();
1.954 + bool is_stable_ary = FoldStableValues &&
1.955 + (tp != NULL) && (tp->isa_aryptr() != NULL) &&
1.956 + tp->isa_aryptr()->is_stable();
1.957 +
1.958 + return (eliminate_boxing && non_volatile) || is_stable_ary;
1.959 + }
1.960 +
1.961 + return false;
1.962 +}
1.963 +
1.964 +//---------------------------can_see_stored_value------------------------------
1.965 +// This routine exists to make sure this set of tests is done the same
1.966 +// everywhere. We need to make a coordinated change: first LoadNode::Ideal
1.967 +// will change the graph shape in a way which makes memory alive twice at the
1.968 +// same time (uses the Oracle model of aliasing), then some
1.969 +// LoadXNode::Identity will fold things back to the equivalence-class model
1.970 +// of aliasing.
1.971 +Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const {
1.972 + Node* ld_adr = in(MemNode::Address);
1.973 + intptr_t ld_off = 0;
1.974 + AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
1.975 + const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
1.976 + Compile::AliasType* atp = (tp != NULL) ? phase->C->alias_type(tp) : NULL;
1.977 + // This is more general than load from boxing objects.
1.978 + if (skip_through_membars(atp, tp, phase->C->eliminate_boxing())) {
1.979 + uint alias_idx = atp->index();
1.980 + bool final = !atp->is_rewritable();
1.981 + Node* result = NULL;
1.982 + Node* current = st;
1.983 + // Skip through chains of MemBarNodes checking the MergeMems for
1.984 + // new states for the slice of this load. Stop once any other
1.985 + // kind of node is encountered. Loads from final memory can skip
1.986 + // through any kind of MemBar but normal loads shouldn't skip
1.987 + // through MemBarAcquire since the could allow them to move out of
1.988 + // a synchronized region.
1.989 + while (current->is_Proj()) {
1.990 + int opc = current->in(0)->Opcode();
1.991 + if ((final && (opc == Op_MemBarAcquire ||
1.992 + opc == Op_MemBarAcquireLock ||
1.993 + opc == Op_LoadFence)) ||
1.994 + opc == Op_MemBarRelease ||
1.995 + opc == Op_StoreFence ||
1.996 + opc == Op_MemBarReleaseLock ||
1.997 + opc == Op_MemBarCPUOrder) {
1.998 + Node* mem = current->in(0)->in(TypeFunc::Memory);
1.999 + if (mem->is_MergeMem()) {
1.1000 + MergeMemNode* merge = mem->as_MergeMem();
1.1001 + Node* new_st = merge->memory_at(alias_idx);
1.1002 + if (new_st == merge->base_memory()) {
1.1003 + // Keep searching
1.1004 + current = new_st;
1.1005 + continue;
1.1006 + }
1.1007 + // Save the new memory state for the slice and fall through
1.1008 + // to exit.
1.1009 + result = new_st;
1.1010 + }
1.1011 + }
1.1012 + break;
1.1013 + }
1.1014 + if (result != NULL) {
1.1015 + st = result;
1.1016 + }
1.1017 + }
1.1018 +
1.1019 + // Loop around twice in the case Load -> Initialize -> Store.
1.1020 + // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
1.1021 + for (int trip = 0; trip <= 1; trip++) {
1.1022 +
1.1023 + if (st->is_Store()) {
1.1024 + Node* st_adr = st->in(MemNode::Address);
1.1025 + if (!phase->eqv(st_adr, ld_adr)) {
1.1026 + // Try harder before giving up... Match raw and non-raw pointers.
1.1027 + intptr_t st_off = 0;
1.1028 + AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off);
1.1029 + if (alloc == NULL) return NULL;
1.1030 + if (alloc != ld_alloc) return NULL;
1.1031 + if (ld_off != st_off) return NULL;
1.1032 + // At this point we have proven something like this setup:
1.1033 + // A = Allocate(...)
1.1034 + // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off))
1.1035 + // S = StoreQ(, AddP(, A.Parm , #Off), V)
1.1036 + // (Actually, we haven't yet proven the Q's are the same.)
1.1037 + // In other words, we are loading from a casted version of
1.1038 + // the same pointer-and-offset that we stored to.
1.1039 + // Thus, we are able to replace L by V.
1.1040 + }
1.1041 + // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
1.1042 + if (store_Opcode() != st->Opcode())
1.1043 + return NULL;
1.1044 + return st->in(MemNode::ValueIn);
1.1045 + }
1.1046 +
1.1047 + // A load from a freshly-created object always returns zero.
1.1048 + // (This can happen after LoadNode::Ideal resets the load's memory input
1.1049 + // to find_captured_store, which returned InitializeNode::zero_memory.)
1.1050 + if (st->is_Proj() && st->in(0)->is_Allocate() &&
1.1051 + (st->in(0) == ld_alloc) &&
1.1052 + (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) {
1.1053 + // return a zero value for the load's basic type
1.1054 + // (This is one of the few places where a generic PhaseTransform
1.1055 + // can create new nodes. Think of it as lazily manifesting
1.1056 + // virtually pre-existing constants.)
1.1057 + return phase->zerocon(memory_type());
1.1058 + }
1.1059 +
1.1060 + // A load from an initialization barrier can match a captured store.
1.1061 + if (st->is_Proj() && st->in(0)->is_Initialize()) {
1.1062 + InitializeNode* init = st->in(0)->as_Initialize();
1.1063 + AllocateNode* alloc = init->allocation();
1.1064 + if ((alloc != NULL) && (alloc == ld_alloc)) {
1.1065 + // examine a captured store value
1.1066 + st = init->find_captured_store(ld_off, memory_size(), phase);
1.1067 + if (st != NULL)
1.1068 + continue; // take one more trip around
1.1069 + }
1.1070 + }
1.1071 +
1.1072 + // Load boxed value from result of valueOf() call is input parameter.
1.1073 + if (this->is_Load() && ld_adr->is_AddP() &&
1.1074 + (tp != NULL) && tp->is_ptr_to_boxed_value()) {
1.1075 + intptr_t ignore = 0;
1.1076 + Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore);
1.1077 + if (base != NULL && base->is_Proj() &&
1.1078 + base->as_Proj()->_con == TypeFunc::Parms &&
1.1079 + base->in(0)->is_CallStaticJava() &&
1.1080 + base->in(0)->as_CallStaticJava()->is_boxing_method()) {
1.1081 + return base->in(0)->in(TypeFunc::Parms);
1.1082 + }
1.1083 + }
1.1084 +
1.1085 + break;
1.1086 + }
1.1087 +
1.1088 + return NULL;
1.1089 +}
1.1090 +
1.1091 +//----------------------is_instance_field_load_with_local_phi------------------
1.1092 +bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
1.1093 + if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl &&
1.1094 + in(Address)->is_AddP() ) {
1.1095 + const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr();
1.1096 + // Only instances and boxed values.
1.1097 + if( t_oop != NULL &&
1.1098 + (t_oop->is_ptr_to_boxed_value() ||
1.1099 + t_oop->is_known_instance_field()) &&
1.1100 + t_oop->offset() != Type::OffsetBot &&
1.1101 + t_oop->offset() != Type::OffsetTop) {
1.1102 + return true;
1.1103 + }
1.1104 + }
1.1105 + return false;
1.1106 +}
1.1107 +
1.1108 +//------------------------------Identity---------------------------------------
1.1109 +// Loads are identity if previous store is to same address
1.1110 +Node *LoadNode::Identity( PhaseTransform *phase ) {
1.1111 + // If the previous store-maker is the right kind of Store, and the store is
1.1112 + // to the same address, then we are equal to the value stored.
1.1113 + Node* mem = in(Memory);
1.1114 + Node* value = can_see_stored_value(mem, phase);
1.1115 + if( value ) {
1.1116 + // byte, short & char stores truncate naturally.
1.1117 + // A load has to load the truncated value which requires
1.1118 + // some sort of masking operation and that requires an
1.1119 + // Ideal call instead of an Identity call.
1.1120 + if (memory_size() < BytesPerInt) {
1.1121 + // If the input to the store does not fit with the load's result type,
1.1122 + // it must be truncated via an Ideal call.
1.1123 + if (!phase->type(value)->higher_equal(phase->type(this)))
1.1124 + return this;
1.1125 + }
1.1126 + // (This works even when value is a Con, but LoadNode::Value
1.1127 + // usually runs first, producing the singleton type of the Con.)
1.1128 + return value;
1.1129 + }
1.1130 +
1.1131 + // Search for an existing data phi which was generated before for the same
1.1132 + // instance's field to avoid infinite generation of phis in a loop.
1.1133 + Node *region = mem->in(0);
1.1134 + if (is_instance_field_load_with_local_phi(region)) {
1.1135 + const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr();
1.1136 + int this_index = phase->C->get_alias_index(addr_t);
1.1137 + int this_offset = addr_t->offset();
1.1138 + int this_iid = addr_t->instance_id();
1.1139 + if (!addr_t->is_known_instance() &&
1.1140 + addr_t->is_ptr_to_boxed_value()) {
1.1141 + // Use _idx of address base (could be Phi node) for boxed values.
1.1142 + intptr_t ignore = 0;
1.1143 + Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1.1144 + this_iid = base->_idx;
1.1145 + }
1.1146 + const Type* this_type = bottom_type();
1.1147 + for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1.1148 + Node* phi = region->fast_out(i);
1.1149 + if (phi->is_Phi() && phi != mem &&
1.1150 + phi->as_Phi()->is_same_inst_field(this_type, this_iid, this_index, this_offset)) {
1.1151 + return phi;
1.1152 + }
1.1153 + }
1.1154 + }
1.1155 +
1.1156 + return this;
1.1157 +}
1.1158 +
1.1159 +// We're loading from an object which has autobox behaviour.
1.1160 +// If this object is result of a valueOf call we'll have a phi
1.1161 +// merging a newly allocated object and a load from the cache.
1.1162 +// We want to replace this load with the original incoming
1.1163 +// argument to the valueOf call.
1.1164 +Node* LoadNode::eliminate_autobox(PhaseGVN* phase) {
1.1165 + assert(phase->C->eliminate_boxing(), "sanity");
1.1166 + intptr_t ignore = 0;
1.1167 + Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1.1168 + if ((base == NULL) || base->is_Phi()) {
1.1169 + // Push the loads from the phi that comes from valueOf up
1.1170 + // through it to allow elimination of the loads and the recovery
1.1171 + // of the original value. It is done in split_through_phi().
1.1172 + return NULL;
1.1173 + } else if (base->is_Load() ||
1.1174 + base->is_DecodeN() && base->in(1)->is_Load()) {
1.1175 + // Eliminate the load of boxed value for integer types from the cache
1.1176 + // array by deriving the value from the index into the array.
1.1177 + // Capture the offset of the load and then reverse the computation.
1.1178 +
1.1179 + // Get LoadN node which loads a boxing object from 'cache' array.
1.1180 + if (base->is_DecodeN()) {
1.1181 + base = base->in(1);
1.1182 + }
1.1183 + if (!base->in(Address)->is_AddP()) {
1.1184 + return NULL; // Complex address
1.1185 + }
1.1186 + AddPNode* address = base->in(Address)->as_AddP();
1.1187 + Node* cache_base = address->in(AddPNode::Base);
1.1188 + if ((cache_base != NULL) && cache_base->is_DecodeN()) {
1.1189 + // Get ConP node which is static 'cache' field.
1.1190 + cache_base = cache_base->in(1);
1.1191 + }
1.1192 + if ((cache_base != NULL) && cache_base->is_Con()) {
1.1193 + const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr();
1.1194 + if ((base_type != NULL) && base_type->is_autobox_cache()) {
1.1195 + Node* elements[4];
1.1196 + int shift = exact_log2(type2aelembytes(T_OBJECT));
1.1197 + int count = address->unpack_offsets(elements, ARRAY_SIZE(elements));
1.1198 + if ((count > 0) && elements[0]->is_Con() &&
1.1199 + ((count == 1) ||
1.1200 + (count == 2) && elements[1]->Opcode() == Op_LShiftX &&
1.1201 + elements[1]->in(2) == phase->intcon(shift))) {
1.1202 + ciObjArray* array = base_type->const_oop()->as_obj_array();
1.1203 + // Fetch the box object cache[0] at the base of the array and get its value
1.1204 + ciInstance* box = array->obj_at(0)->as_instance();
1.1205 + ciInstanceKlass* ik = box->klass()->as_instance_klass();
1.1206 + assert(ik->is_box_klass(), "sanity");
1.1207 + assert(ik->nof_nonstatic_fields() == 1, "change following code");
1.1208 + if (ik->nof_nonstatic_fields() == 1) {
1.1209 + // This should be true nonstatic_field_at requires calling
1.1210 + // nof_nonstatic_fields so check it anyway
1.1211 + ciConstant c = box->field_value(ik->nonstatic_field_at(0));
1.1212 + BasicType bt = c.basic_type();
1.1213 + // Only integer types have boxing cache.
1.1214 + assert(bt == T_BOOLEAN || bt == T_CHAR ||
1.1215 + bt == T_BYTE || bt == T_SHORT ||
1.1216 + bt == T_INT || bt == T_LONG, err_msg_res("wrong type = %s", type2name(bt)));
1.1217 + jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int();
1.1218 + if (cache_low != (int)cache_low) {
1.1219 + return NULL; // should not happen since cache is array indexed by value
1.1220 + }
1.1221 + jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift);
1.1222 + if (offset != (int)offset) {
1.1223 + return NULL; // should not happen since cache is array indexed by value
1.1224 + }
1.1225 + // Add up all the offsets making of the address of the load
1.1226 + Node* result = elements[0];
1.1227 + for (int i = 1; i < count; i++) {
1.1228 + result = phase->transform(new (phase->C) AddXNode(result, elements[i]));
1.1229 + }
1.1230 + // Remove the constant offset from the address and then
1.1231 + result = phase->transform(new (phase->C) AddXNode(result, phase->MakeConX(-(int)offset)));
1.1232 + // remove the scaling of the offset to recover the original index.
1.1233 + if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) {
1.1234 + // Peel the shift off directly but wrap it in a dummy node
1.1235 + // since Ideal can't return existing nodes
1.1236 + result = new (phase->C) RShiftXNode(result->in(1), phase->intcon(0));
1.1237 + } else if (result->is_Add() && result->in(2)->is_Con() &&
1.1238 + result->in(1)->Opcode() == Op_LShiftX &&
1.1239 + result->in(1)->in(2) == phase->intcon(shift)) {
1.1240 + // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z)
1.1241 + // but for boxing cache access we know that X<<Z will not overflow
1.1242 + // (there is range check) so we do this optimizatrion by hand here.
1.1243 + Node* add_con = new (phase->C) RShiftXNode(result->in(2), phase->intcon(shift));
1.1244 + result = new (phase->C) AddXNode(result->in(1)->in(1), phase->transform(add_con));
1.1245 + } else {
1.1246 + result = new (phase->C) RShiftXNode(result, phase->intcon(shift));
1.1247 + }
1.1248 +#ifdef _LP64
1.1249 + if (bt != T_LONG) {
1.1250 + result = new (phase->C) ConvL2INode(phase->transform(result));
1.1251 + }
1.1252 +#else
1.1253 + if (bt == T_LONG) {
1.1254 + result = new (phase->C) ConvI2LNode(phase->transform(result));
1.1255 + }
1.1256 +#endif
1.1257 + return result;
1.1258 + }
1.1259 + }
1.1260 + }
1.1261 + }
1.1262 + }
1.1263 + return NULL;
1.1264 +}
1.1265 +
1.1266 +static bool stable_phi(PhiNode* phi, PhaseGVN *phase) {
1.1267 + Node* region = phi->in(0);
1.1268 + if (region == NULL) {
1.1269 + return false; // Wait stable graph
1.1270 + }
1.1271 + uint cnt = phi->req();
1.1272 + for (uint i = 1; i < cnt; i++) {
1.1273 + Node* rc = region->in(i);
1.1274 + if (rc == NULL || phase->type(rc) == Type::TOP)
1.1275 + return false; // Wait stable graph
1.1276 + Node* in = phi->in(i);
1.1277 + if (in == NULL || phase->type(in) == Type::TOP)
1.1278 + return false; // Wait stable graph
1.1279 + }
1.1280 + return true;
1.1281 +}
1.1282 +//------------------------------split_through_phi------------------------------
1.1283 +// Split instance or boxed field load through Phi.
1.1284 +Node *LoadNode::split_through_phi(PhaseGVN *phase) {
1.1285 + Node* mem = in(Memory);
1.1286 + Node* address = in(Address);
1.1287 + const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr();
1.1288 +
1.1289 + assert((t_oop != NULL) &&
1.1290 + (t_oop->is_known_instance_field() ||
1.1291 + t_oop->is_ptr_to_boxed_value()), "invalide conditions");
1.1292 +
1.1293 + Compile* C = phase->C;
1.1294 + intptr_t ignore = 0;
1.1295 + Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1.1296 + bool base_is_phi = (base != NULL) && base->is_Phi();
1.1297 + bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() &&
1.1298 + (base != NULL) && (base == address->in(AddPNode::Base)) &&
1.1299 + phase->type(base)->higher_equal(TypePtr::NOTNULL);
1.1300 +
1.1301 + if (!((mem->is_Phi() || base_is_phi) &&
1.1302 + (load_boxed_values || t_oop->is_known_instance_field()))) {
1.1303 + return NULL; // memory is not Phi
1.1304 + }
1.1305 +
1.1306 + if (mem->is_Phi()) {
1.1307 + if (!stable_phi(mem->as_Phi(), phase)) {
1.1308 + return NULL; // Wait stable graph
1.1309 + }
1.1310 + uint cnt = mem->req();
1.1311 + // Check for loop invariant memory.
1.1312 + if (cnt == 3) {
1.1313 + for (uint i = 1; i < cnt; i++) {
1.1314 + Node* in = mem->in(i);
1.1315 + Node* m = optimize_memory_chain(in, t_oop, this, phase);
1.1316 + if (m == mem) {
1.1317 + set_req(Memory, mem->in(cnt - i));
1.1318 + return this; // made change
1.1319 + }
1.1320 + }
1.1321 + }
1.1322 + }
1.1323 + if (base_is_phi) {
1.1324 + if (!stable_phi(base->as_Phi(), phase)) {
1.1325 + return NULL; // Wait stable graph
1.1326 + }
1.1327 + uint cnt = base->req();
1.1328 + // Check for loop invariant memory.
1.1329 + if (cnt == 3) {
1.1330 + for (uint i = 1; i < cnt; i++) {
1.1331 + if (base->in(i) == base) {
1.1332 + return NULL; // Wait stable graph
1.1333 + }
1.1334 + }
1.1335 + }
1.1336 + }
1.1337 +
1.1338 + bool load_boxed_phi = load_boxed_values && base_is_phi && (base->in(0) == mem->in(0));
1.1339 +
1.1340 + // Split through Phi (see original code in loopopts.cpp).
1.1341 + assert(C->have_alias_type(t_oop), "instance should have alias type");
1.1342 +
1.1343 + // Do nothing here if Identity will find a value
1.1344 + // (to avoid infinite chain of value phis generation).
1.1345 + if (!phase->eqv(this, this->Identity(phase)))
1.1346 + return NULL;
1.1347 +
1.1348 + // Select Region to split through.
1.1349 + Node* region;
1.1350 + if (!base_is_phi) {
1.1351 + assert(mem->is_Phi(), "sanity");
1.1352 + region = mem->in(0);
1.1353 + // Skip if the region dominates some control edge of the address.
1.1354 + if (!MemNode::all_controls_dominate(address, region))
1.1355 + return NULL;
1.1356 + } else if (!mem->is_Phi()) {
1.1357 + assert(base_is_phi, "sanity");
1.1358 + region = base->in(0);
1.1359 + // Skip if the region dominates some control edge of the memory.
1.1360 + if (!MemNode::all_controls_dominate(mem, region))
1.1361 + return NULL;
1.1362 + } else if (base->in(0) != mem->in(0)) {
1.1363 + assert(base_is_phi && mem->is_Phi(), "sanity");
1.1364 + if (MemNode::all_controls_dominate(mem, base->in(0))) {
1.1365 + region = base->in(0);
1.1366 + } else if (MemNode::all_controls_dominate(address, mem->in(0))) {
1.1367 + region = mem->in(0);
1.1368 + } else {
1.1369 + return NULL; // complex graph
1.1370 + }
1.1371 + } else {
1.1372 + assert(base->in(0) == mem->in(0), "sanity");
1.1373 + region = mem->in(0);
1.1374 + }
1.1375 +
1.1376 + const Type* this_type = this->bottom_type();
1.1377 + int this_index = C->get_alias_index(t_oop);
1.1378 + int this_offset = t_oop->offset();
1.1379 + int this_iid = t_oop->instance_id();
1.1380 + if (!t_oop->is_known_instance() && load_boxed_values) {
1.1381 + // Use _idx of address base for boxed values.
1.1382 + this_iid = base->_idx;
1.1383 + }
1.1384 + PhaseIterGVN* igvn = phase->is_IterGVN();
1.1385 + Node* phi = new (C) PhiNode(region, this_type, NULL, this_iid, this_index, this_offset);
1.1386 + for (uint i = 1; i < region->req(); i++) {
1.1387 + Node* x;
1.1388 + Node* the_clone = NULL;
1.1389 + if (region->in(i) == C->top()) {
1.1390 + x = C->top(); // Dead path? Use a dead data op
1.1391 + } else {
1.1392 + x = this->clone(); // Else clone up the data op
1.1393 + the_clone = x; // Remember for possible deletion.
1.1394 + // Alter data node to use pre-phi inputs
1.1395 + if (this->in(0) == region) {
1.1396 + x->set_req(0, region->in(i));
1.1397 + } else {
1.1398 + x->set_req(0, NULL);
1.1399 + }
1.1400 + if (mem->is_Phi() && (mem->in(0) == region)) {
1.1401 + x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone.
1.1402 + }
1.1403 + if (address->is_Phi() && address->in(0) == region) {
1.1404 + x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone
1.1405 + }
1.1406 + if (base_is_phi && (base->in(0) == region)) {
1.1407 + Node* base_x = base->in(i); // Clone address for loads from boxed objects.
1.1408 + Node* adr_x = phase->transform(new (C) AddPNode(base_x,base_x,address->in(AddPNode::Offset)));
1.1409 + x->set_req(Address, adr_x);
1.1410 + }
1.1411 + }
1.1412 + // Check for a 'win' on some paths
1.1413 + const Type *t = x->Value(igvn);
1.1414 +
1.1415 + bool singleton = t->singleton();
1.1416 +
1.1417 + // See comments in PhaseIdealLoop::split_thru_phi().
1.1418 + if (singleton && t == Type::TOP) {
1.1419 + singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
1.1420 + }
1.1421 +
1.1422 + if (singleton) {
1.1423 + x = igvn->makecon(t);
1.1424 + } else {
1.1425 + // We now call Identity to try to simplify the cloned node.
1.1426 + // Note that some Identity methods call phase->type(this).
1.1427 + // Make sure that the type array is big enough for
1.1428 + // our new node, even though we may throw the node away.
1.1429 + // (This tweaking with igvn only works because x is a new node.)
1.1430 + igvn->set_type(x, t);
1.1431 + // If x is a TypeNode, capture any more-precise type permanently into Node
1.1432 + // otherwise it will be not updated during igvn->transform since
1.1433 + // igvn->type(x) is set to x->Value() already.
1.1434 + x->raise_bottom_type(t);
1.1435 + Node *y = x->Identity(igvn);
1.1436 + if (y != x) {
1.1437 + x = y;
1.1438 + } else {
1.1439 + y = igvn->hash_find_insert(x);
1.1440 + if (y) {
1.1441 + x = y;
1.1442 + } else {
1.1443 + // Else x is a new node we are keeping
1.1444 + // We do not need register_new_node_with_optimizer
1.1445 + // because set_type has already been called.
1.1446 + igvn->_worklist.push(x);
1.1447 + }
1.1448 + }
1.1449 + }
1.1450 + if (x != the_clone && the_clone != NULL) {
1.1451 + igvn->remove_dead_node(the_clone);
1.1452 + }
1.1453 + phi->set_req(i, x);
1.1454 + }
1.1455 + // Record Phi
1.1456 + igvn->register_new_node_with_optimizer(phi);
1.1457 + return phi;
1.1458 +}
1.1459 +
1.1460 +//------------------------------Ideal------------------------------------------
1.1461 +// If the load is from Field memory and the pointer is non-null, we can
1.1462 +// zero out the control input.
1.1463 +// If the offset is constant and the base is an object allocation,
1.1464 +// try to hook me up to the exact initializing store.
1.1465 +Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1.1466 + Node* p = MemNode::Ideal_common(phase, can_reshape);
1.1467 + if (p) return (p == NodeSentinel) ? NULL : p;
1.1468 +
1.1469 + Node* ctrl = in(MemNode::Control);
1.1470 + Node* address = in(MemNode::Address);
1.1471 +
1.1472 + // Skip up past a SafePoint control. Cannot do this for Stores because
1.1473 + // pointer stores & cardmarks must stay on the same side of a SafePoint.
1.1474 + if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint &&
1.1475 + phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) {
1.1476 + ctrl = ctrl->in(0);
1.1477 + set_req(MemNode::Control,ctrl);
1.1478 + }
1.1479 +
1.1480 + intptr_t ignore = 0;
1.1481 + Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1.1482 + if (base != NULL
1.1483 + && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) {
1.1484 + // Check for useless control edge in some common special cases
1.1485 + if (in(MemNode::Control) != NULL
1.1486 + && phase->type(base)->higher_equal(TypePtr::NOTNULL)
1.1487 + && all_controls_dominate(base, phase->C->start())) {
1.1488 + // A method-invariant, non-null address (constant or 'this' argument).
1.1489 + set_req(MemNode::Control, NULL);
1.1490 + }
1.1491 + }
1.1492 +
1.1493 + Node* mem = in(MemNode::Memory);
1.1494 + const TypePtr *addr_t = phase->type(address)->isa_ptr();
1.1495 +
1.1496 + if (can_reshape && (addr_t != NULL)) {
1.1497 + // try to optimize our memory input
1.1498 + Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase);
1.1499 + if (opt_mem != mem) {
1.1500 + set_req(MemNode::Memory, opt_mem);
1.1501 + if (phase->type( opt_mem ) == Type::TOP) return NULL;
1.1502 + return this;
1.1503 + }
1.1504 + const TypeOopPtr *t_oop = addr_t->isa_oopptr();
1.1505 + if ((t_oop != NULL) &&
1.1506 + (t_oop->is_known_instance_field() ||
1.1507 + t_oop->is_ptr_to_boxed_value())) {
1.1508 + PhaseIterGVN *igvn = phase->is_IterGVN();
1.1509 + if (igvn != NULL && igvn->_worklist.member(opt_mem)) {
1.1510 + // Delay this transformation until memory Phi is processed.
1.1511 + phase->is_IterGVN()->_worklist.push(this);
1.1512 + return NULL;
1.1513 + }
1.1514 + // Split instance field load through Phi.
1.1515 + Node* result = split_through_phi(phase);
1.1516 + if (result != NULL) return result;
1.1517 +
1.1518 + if (t_oop->is_ptr_to_boxed_value()) {
1.1519 + Node* result = eliminate_autobox(phase);
1.1520 + if (result != NULL) return result;
1.1521 + }
1.1522 + }
1.1523 + }
1.1524 +
1.1525 + // Check for prior store with a different base or offset; make Load
1.1526 + // independent. Skip through any number of them. Bail out if the stores
1.1527 + // are in an endless dead cycle and report no progress. This is a key
1.1528 + // transform for Reflection. However, if after skipping through the Stores
1.1529 + // we can't then fold up against a prior store do NOT do the transform as
1.1530 + // this amounts to using the 'Oracle' model of aliasing. It leaves the same
1.1531 + // array memory alive twice: once for the hoisted Load and again after the
1.1532 + // bypassed Store. This situation only works if EVERYBODY who does
1.1533 + // anti-dependence work knows how to bypass. I.e. we need all
1.1534 + // anti-dependence checks to ask the same Oracle. Right now, that Oracle is
1.1535 + // the alias index stuff. So instead, peek through Stores and IFF we can
1.1536 + // fold up, do so.
1.1537 + Node* prev_mem = find_previous_store(phase);
1.1538 + // Steps (a), (b): Walk past independent stores to find an exact match.
1.1539 + if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) {
1.1540 + // (c) See if we can fold up on the spot, but don't fold up here.
1.1541 + // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or
1.1542 + // just return a prior value, which is done by Identity calls.
1.1543 + if (can_see_stored_value(prev_mem, phase)) {
1.1544 + // Make ready for step (d):
1.1545 + set_req(MemNode::Memory, prev_mem);
1.1546 + return this;
1.1547 + }
1.1548 + }
1.1549 +
1.1550 + return NULL; // No further progress
1.1551 +}
1.1552 +
1.1553 +// Helper to recognize certain Klass fields which are invariant across
1.1554 +// some group of array types (e.g., int[] or all T[] where T < Object).
1.1555 +const Type*
1.1556 +LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
1.1557 + ciKlass* klass) const {
1.1558 + if (tkls->offset() == in_bytes(Klass::modifier_flags_offset())) {
1.1559 + // The field is Klass::_modifier_flags. Return its (constant) value.
1.1560 + // (Folds up the 2nd indirection in aClassConstant.getModifiers().)
1.1561 + assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags");
1.1562 + return TypeInt::make(klass->modifier_flags());
1.1563 + }
1.1564 + if (tkls->offset() == in_bytes(Klass::access_flags_offset())) {
1.1565 + // The field is Klass::_access_flags. Return its (constant) value.
1.1566 + // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
1.1567 + assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags");
1.1568 + return TypeInt::make(klass->access_flags());
1.1569 + }
1.1570 + if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) {
1.1571 + // The field is Klass::_layout_helper. Return its constant value if known.
1.1572 + assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper");
1.1573 + return TypeInt::make(klass->layout_helper());
1.1574 + }
1.1575 +
1.1576 + // No match.
1.1577 + return NULL;
1.1578 +}
1.1579 +
1.1580 +// Try to constant-fold a stable array element.
1.1581 +static const Type* fold_stable_ary_elem(const TypeAryPtr* ary, int off, BasicType loadbt) {
1.1582 + assert(ary->const_oop(), "array should be constant");
1.1583 + assert(ary->is_stable(), "array should be stable");
1.1584 +
1.1585 + // Decode the results of GraphKit::array_element_address.
1.1586 + ciArray* aobj = ary->const_oop()->as_array();
1.1587 + ciConstant con = aobj->element_value_by_offset(off);
1.1588 +
1.1589 + if (con.basic_type() != T_ILLEGAL && !con.is_null_or_zero()) {
1.1590 + const Type* con_type = Type::make_from_constant(con);
1.1591 + if (con_type != NULL) {
1.1592 + if (con_type->isa_aryptr()) {
1.1593 + // Join with the array element type, in case it is also stable.
1.1594 + int dim = ary->stable_dimension();
1.1595 + con_type = con_type->is_aryptr()->cast_to_stable(true, dim-1);
1.1596 + }
1.1597 + if (loadbt == T_NARROWOOP && con_type->isa_oopptr()) {
1.1598 + con_type = con_type->make_narrowoop();
1.1599 + }
1.1600 +#ifndef PRODUCT
1.1601 + if (TraceIterativeGVN) {
1.1602 + tty->print("FoldStableValues: array element [off=%d]: con_type=", off);
1.1603 + con_type->dump(); tty->cr();
1.1604 + }
1.1605 +#endif //PRODUCT
1.1606 + return con_type;
1.1607 + }
1.1608 + }
1.1609 + return NULL;
1.1610 +}
1.1611 +
1.1612 +//------------------------------Value-----------------------------------------
1.1613 +const Type *LoadNode::Value( PhaseTransform *phase ) const {
1.1614 + // Either input is TOP ==> the result is TOP
1.1615 + Node* mem = in(MemNode::Memory);
1.1616 + const Type *t1 = phase->type(mem);
1.1617 + if (t1 == Type::TOP) return Type::TOP;
1.1618 + Node* adr = in(MemNode::Address);
1.1619 + const TypePtr* tp = phase->type(adr)->isa_ptr();
1.1620 + if (tp == NULL || tp->empty()) return Type::TOP;
1.1621 + int off = tp->offset();
1.1622 + assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
1.1623 + Compile* C = phase->C;
1.1624 +
1.1625 + // Try to guess loaded type from pointer type
1.1626 + if (tp->isa_aryptr()) {
1.1627 + const TypeAryPtr* ary = tp->is_aryptr();
1.1628 + const Type* t = ary->elem();
1.1629 +
1.1630 + // Determine whether the reference is beyond the header or not, by comparing
1.1631 + // the offset against the offset of the start of the array's data.
1.1632 + // Different array types begin at slightly different offsets (12 vs. 16).
1.1633 + // We choose T_BYTE as an example base type that is least restrictive
1.1634 + // as to alignment, which will therefore produce the smallest
1.1635 + // possible base offset.
1.1636 + const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1.1637 + const bool off_beyond_header = ((uint)off >= (uint)min_base_off);
1.1638 +
1.1639 + // Try to constant-fold a stable array element.
1.1640 + if (FoldStableValues && ary->is_stable() && ary->const_oop() != NULL) {
1.1641 + // Make sure the reference is not into the header and the offset is constant
1.1642 + if (off_beyond_header && adr->is_AddP() && off != Type::OffsetBot) {
1.1643 + const Type* con_type = fold_stable_ary_elem(ary, off, memory_type());
1.1644 + if (con_type != NULL) {
1.1645 + return con_type;
1.1646 + }
1.1647 + }
1.1648 + }
1.1649 +
1.1650 + // Don't do this for integer types. There is only potential profit if
1.1651 + // the element type t is lower than _type; that is, for int types, if _type is
1.1652 + // more restrictive than t. This only happens here if one is short and the other
1.1653 + // char (both 16 bits), and in those cases we've made an intentional decision
1.1654 + // to use one kind of load over the other. See AndINode::Ideal and 4965907.
1.1655 + // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
1.1656 + //
1.1657 + // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
1.1658 + // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier
1.1659 + // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
1.1660 + // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed,
1.1661 + // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
1.1662 + // In fact, that could have been the original type of p1, and p1 could have
1.1663 + // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
1.1664 + // expression (LShiftL quux 3) independently optimized to the constant 8.
1.1665 + if ((t->isa_int() == NULL) && (t->isa_long() == NULL)
1.1666 + && (_type->isa_vect() == NULL)
1.1667 + && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) {
1.1668 + // t might actually be lower than _type, if _type is a unique
1.1669 + // concrete subclass of abstract class t.
1.1670 + if (off_beyond_header) { // is the offset beyond the header?
1.1671 + const Type* jt = t->join_speculative(_type);
1.1672 + // In any case, do not allow the join, per se, to empty out the type.
1.1673 + if (jt->empty() && !t->empty()) {
1.1674 + // This can happen if a interface-typed array narrows to a class type.
1.1675 + jt = _type;
1.1676 + }
1.1677 +#ifdef ASSERT
1.1678 + if (phase->C->eliminate_boxing() && adr->is_AddP()) {
1.1679 + // The pointers in the autobox arrays are always non-null
1.1680 + Node* base = adr->in(AddPNode::Base);
1.1681 + if ((base != NULL) && base->is_DecodeN()) {
1.1682 + // Get LoadN node which loads IntegerCache.cache field
1.1683 + base = base->in(1);
1.1684 + }
1.1685 + if ((base != NULL) && base->is_Con()) {
1.1686 + const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr();
1.1687 + if ((base_type != NULL) && base_type->is_autobox_cache()) {
1.1688 + // It could be narrow oop
1.1689 + assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity");
1.1690 + }
1.1691 + }
1.1692 + }
1.1693 +#endif
1.1694 + return jt;
1.1695 + }
1.1696 + }
1.1697 + } else if (tp->base() == Type::InstPtr) {
1.1698 + ciEnv* env = C->env();
1.1699 + const TypeInstPtr* tinst = tp->is_instptr();
1.1700 + ciKlass* klass = tinst->klass();
1.1701 + assert( off != Type::OffsetBot ||
1.1702 + // arrays can be cast to Objects
1.1703 + tp->is_oopptr()->klass()->is_java_lang_Object() ||
1.1704 + // unsafe field access may not have a constant offset
1.1705 + C->has_unsafe_access(),
1.1706 + "Field accesses must be precise" );
1.1707 + // For oop loads, we expect the _type to be precise
1.1708 + if (klass == env->String_klass() &&
1.1709 + adr->is_AddP() && off != Type::OffsetBot) {
1.1710 + // For constant Strings treat the final fields as compile time constants.
1.1711 + Node* base = adr->in(AddPNode::Base);
1.1712 + const TypeOopPtr* t = phase->type(base)->isa_oopptr();
1.1713 + if (t != NULL && t->singleton()) {
1.1714 + ciField* field = env->String_klass()->get_field_by_offset(off, false);
1.1715 + if (field != NULL && field->is_final()) {
1.1716 + ciObject* string = t->const_oop();
1.1717 + ciConstant constant = string->as_instance()->field_value(field);
1.1718 + if (constant.basic_type() == T_INT) {
1.1719 + return TypeInt::make(constant.as_int());
1.1720 + } else if (constant.basic_type() == T_ARRAY) {
1.1721 + if (adr->bottom_type()->is_ptr_to_narrowoop()) {
1.1722 + return TypeNarrowOop::make_from_constant(constant.as_object(), true);
1.1723 + } else {
1.1724 + return TypeOopPtr::make_from_constant(constant.as_object(), true);
1.1725 + }
1.1726 + }
1.1727 + }
1.1728 + }
1.1729 + }
1.1730 + // Optimizations for constant objects
1.1731 + ciObject* const_oop = tinst->const_oop();
1.1732 + if (const_oop != NULL) {
1.1733 + // For constant Boxed value treat the target field as a compile time constant.
1.1734 + if (tinst->is_ptr_to_boxed_value()) {
1.1735 + return tinst->get_const_boxed_value();
1.1736 + } else
1.1737 + // For constant CallSites treat the target field as a compile time constant.
1.1738 + if (const_oop->is_call_site()) {
1.1739 + ciCallSite* call_site = const_oop->as_call_site();
1.1740 + ciField* field = call_site->klass()->as_instance_klass()->get_field_by_offset(off, /*is_static=*/ false);
1.1741 + if (field != NULL && field->is_call_site_target()) {
1.1742 + ciMethodHandle* target = call_site->get_target();
1.1743 + if (target != NULL) { // just in case
1.1744 + ciConstant constant(T_OBJECT, target);
1.1745 + const Type* t;
1.1746 + if (adr->bottom_type()->is_ptr_to_narrowoop()) {
1.1747 + t = TypeNarrowOop::make_from_constant(constant.as_object(), true);
1.1748 + } else {
1.1749 + t = TypeOopPtr::make_from_constant(constant.as_object(), true);
1.1750 + }
1.1751 + // Add a dependence for invalidation of the optimization.
1.1752 + if (!call_site->is_constant_call_site()) {
1.1753 + C->dependencies()->assert_call_site_target_value(call_site, target);
1.1754 + }
1.1755 + return t;
1.1756 + }
1.1757 + }
1.1758 + }
1.1759 + }
1.1760 + } else if (tp->base() == Type::KlassPtr) {
1.1761 + assert( off != Type::OffsetBot ||
1.1762 + // arrays can be cast to Objects
1.1763 + tp->is_klassptr()->klass()->is_java_lang_Object() ||
1.1764 + // also allow array-loading from the primary supertype
1.1765 + // array during subtype checks
1.1766 + Opcode() == Op_LoadKlass,
1.1767 + "Field accesses must be precise" );
1.1768 + // For klass/static loads, we expect the _type to be precise
1.1769 + }
1.1770 +
1.1771 + const TypeKlassPtr *tkls = tp->isa_klassptr();
1.1772 + if (tkls != NULL && !StressReflectiveCode) {
1.1773 + ciKlass* klass = tkls->klass();
1.1774 + if (klass->is_loaded() && tkls->klass_is_exact()) {
1.1775 + // We are loading a field from a Klass metaobject whose identity
1.1776 + // is known at compile time (the type is "exact" or "precise").
1.1777 + // Check for fields we know are maintained as constants by the VM.
1.1778 + if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) {
1.1779 + // The field is Klass::_super_check_offset. Return its (constant) value.
1.1780 + // (Folds up type checking code.)
1.1781 + assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
1.1782 + return TypeInt::make(klass->super_check_offset());
1.1783 + }
1.1784 + // Compute index into primary_supers array
1.1785 + juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
1.1786 + // Check for overflowing; use unsigned compare to handle the negative case.
1.1787 + if( depth < ciKlass::primary_super_limit() ) {
1.1788 + // The field is an element of Klass::_primary_supers. Return its (constant) value.
1.1789 + // (Folds up type checking code.)
1.1790 + assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1.1791 + ciKlass *ss = klass->super_of_depth(depth);
1.1792 + return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1.1793 + }
1.1794 + const Type* aift = load_array_final_field(tkls, klass);
1.1795 + if (aift != NULL) return aift;
1.1796 + if (tkls->offset() == in_bytes(ArrayKlass::component_mirror_offset())
1.1797 + && klass->is_array_klass()) {
1.1798 + // The field is ArrayKlass::_component_mirror. Return its (constant) value.
1.1799 + // (Folds up aClassConstant.getComponentType, common in Arrays.copyOf.)
1.1800 + assert(Opcode() == Op_LoadP, "must load an oop from _component_mirror");
1.1801 + return TypeInstPtr::make(klass->as_array_klass()->component_mirror());
1.1802 + }
1.1803 + if (tkls->offset() == in_bytes(Klass::java_mirror_offset())) {
1.1804 + // The field is Klass::_java_mirror. Return its (constant) value.
1.1805 + // (Folds up the 2nd indirection in anObjConstant.getClass().)
1.1806 + assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
1.1807 + return TypeInstPtr::make(klass->java_mirror());
1.1808 + }
1.1809 + }
1.1810 +
1.1811 + // We can still check if we are loading from the primary_supers array at a
1.1812 + // shallow enough depth. Even though the klass is not exact, entries less
1.1813 + // than or equal to its super depth are correct.
1.1814 + if (klass->is_loaded() ) {
1.1815 + ciType *inner = klass;
1.1816 + while( inner->is_obj_array_klass() )
1.1817 + inner = inner->as_obj_array_klass()->base_element_type();
1.1818 + if( inner->is_instance_klass() &&
1.1819 + !inner->as_instance_klass()->flags().is_interface() ) {
1.1820 + // Compute index into primary_supers array
1.1821 + juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
1.1822 + // Check for overflowing; use unsigned compare to handle the negative case.
1.1823 + if( depth < ciKlass::primary_super_limit() &&
1.1824 + depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case
1.1825 + // The field is an element of Klass::_primary_supers. Return its (constant) value.
1.1826 + // (Folds up type checking code.)
1.1827 + assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1.1828 + ciKlass *ss = klass->super_of_depth(depth);
1.1829 + return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1.1830 + }
1.1831 + }
1.1832 + }
1.1833 +
1.1834 + // If the type is enough to determine that the thing is not an array,
1.1835 + // we can give the layout_helper a positive interval type.
1.1836 + // This will help short-circuit some reflective code.
1.1837 + if (tkls->offset() == in_bytes(Klass::layout_helper_offset())
1.1838 + && !klass->is_array_klass() // not directly typed as an array
1.1839 + && !klass->is_interface() // specifically not Serializable & Cloneable
1.1840 + && !klass->is_java_lang_Object() // not the supertype of all T[]
1.1841 + ) {
1.1842 + // Note: When interfaces are reliable, we can narrow the interface
1.1843 + // test to (klass != Serializable && klass != Cloneable).
1.1844 + assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
1.1845 + jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
1.1846 + // The key property of this type is that it folds up tests
1.1847 + // for array-ness, since it proves that the layout_helper is positive.
1.1848 + // Thus, a generic value like the basic object layout helper works fine.
1.1849 + return TypeInt::make(min_size, max_jint, Type::WidenMin);
1.1850 + }
1.1851 + }
1.1852 +
1.1853 + // If we are loading from a freshly-allocated object, produce a zero,
1.1854 + // if the load is provably beyond the header of the object.
1.1855 + // (Also allow a variable load from a fresh array to produce zero.)
1.1856 + const TypeOopPtr *tinst = tp->isa_oopptr();
1.1857 + bool is_instance = (tinst != NULL) && tinst->is_known_instance_field();
1.1858 + bool is_boxed_value = (tinst != NULL) && tinst->is_ptr_to_boxed_value();
1.1859 + if (ReduceFieldZeroing || is_instance || is_boxed_value) {
1.1860 + Node* value = can_see_stored_value(mem,phase);
1.1861 + if (value != NULL && value->is_Con()) {
1.1862 + assert(value->bottom_type()->higher_equal(_type),"sanity");
1.1863 + return value->bottom_type();
1.1864 + }
1.1865 + }
1.1866 +
1.1867 + if (is_instance) {
1.1868 + // If we have an instance type and our memory input is the
1.1869 + // programs's initial memory state, there is no matching store,
1.1870 + // so just return a zero of the appropriate type
1.1871 + Node *mem = in(MemNode::Memory);
1.1872 + if (mem->is_Parm() && mem->in(0)->is_Start()) {
1.1873 + assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
1.1874 + return Type::get_zero_type(_type->basic_type());
1.1875 + }
1.1876 + }
1.1877 + return _type;
1.1878 +}
1.1879 +
1.1880 +//------------------------------match_edge-------------------------------------
1.1881 +// Do we Match on this edge index or not? Match only the address.
1.1882 +uint LoadNode::match_edge(uint idx) const {
1.1883 + return idx == MemNode::Address;
1.1884 +}
1.1885 +
1.1886 +//--------------------------LoadBNode::Ideal--------------------------------------
1.1887 +//
1.1888 +// If the previous store is to the same address as this load,
1.1889 +// and the value stored was larger than a byte, replace this load
1.1890 +// with the value stored truncated to a byte. If no truncation is
1.1891 +// needed, the replacement is done in LoadNode::Identity().
1.1892 +//
1.1893 +Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1.1894 + Node* mem = in(MemNode::Memory);
1.1895 + Node* value = can_see_stored_value(mem,phase);
1.1896 + if( value && !phase->type(value)->higher_equal( _type ) ) {
1.1897 + Node *result = phase->transform( new (phase->C) LShiftINode(value, phase->intcon(24)) );
1.1898 + return new (phase->C) RShiftINode(result, phase->intcon(24));
1.1899 + }
1.1900 + // Identity call will handle the case where truncation is not needed.
1.1901 + return LoadNode::Ideal(phase, can_reshape);
1.1902 +}
1.1903 +
1.1904 +const Type* LoadBNode::Value(PhaseTransform *phase) const {
1.1905 + Node* mem = in(MemNode::Memory);
1.1906 + Node* value = can_see_stored_value(mem,phase);
1.1907 + if (value != NULL && value->is_Con() &&
1.1908 + !value->bottom_type()->higher_equal(_type)) {
1.1909 + // If the input to the store does not fit with the load's result type,
1.1910 + // it must be truncated. We can't delay until Ideal call since
1.1911 + // a singleton Value is needed for split_thru_phi optimization.
1.1912 + int con = value->get_int();
1.1913 + return TypeInt::make((con << 24) >> 24);
1.1914 + }
1.1915 + return LoadNode::Value(phase);
1.1916 +}
1.1917 +
1.1918 +//--------------------------LoadUBNode::Ideal-------------------------------------
1.1919 +//
1.1920 +// If the previous store is to the same address as this load,
1.1921 +// and the value stored was larger than a byte, replace this load
1.1922 +// with the value stored truncated to a byte. If no truncation is
1.1923 +// needed, the replacement is done in LoadNode::Identity().
1.1924 +//
1.1925 +Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
1.1926 + Node* mem = in(MemNode::Memory);
1.1927 + Node* value = can_see_stored_value(mem, phase);
1.1928 + if (value && !phase->type(value)->higher_equal(_type))
1.1929 + return new (phase->C) AndINode(value, phase->intcon(0xFF));
1.1930 + // Identity call will handle the case where truncation is not needed.
1.1931 + return LoadNode::Ideal(phase, can_reshape);
1.1932 +}
1.1933 +
1.1934 +const Type* LoadUBNode::Value(PhaseTransform *phase) const {
1.1935 + Node* mem = in(MemNode::Memory);
1.1936 + Node* value = can_see_stored_value(mem,phase);
1.1937 + if (value != NULL && value->is_Con() &&
1.1938 + !value->bottom_type()->higher_equal(_type)) {
1.1939 + // If the input to the store does not fit with the load's result type,
1.1940 + // it must be truncated. We can't delay until Ideal call since
1.1941 + // a singleton Value is needed for split_thru_phi optimization.
1.1942 + int con = value->get_int();
1.1943 + return TypeInt::make(con & 0xFF);
1.1944 + }
1.1945 + return LoadNode::Value(phase);
1.1946 +}
1.1947 +
1.1948 +//--------------------------LoadUSNode::Ideal-------------------------------------
1.1949 +//
1.1950 +// If the previous store is to the same address as this load,
1.1951 +// and the value stored was larger than a char, replace this load
1.1952 +// with the value stored truncated to a char. If no truncation is
1.1953 +// needed, the replacement is done in LoadNode::Identity().
1.1954 +//
1.1955 +Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1.1956 + Node* mem = in(MemNode::Memory);
1.1957 + Node* value = can_see_stored_value(mem,phase);
1.1958 + if( value && !phase->type(value)->higher_equal( _type ) )
1.1959 + return new (phase->C) AndINode(value,phase->intcon(0xFFFF));
1.1960 + // Identity call will handle the case where truncation is not needed.
1.1961 + return LoadNode::Ideal(phase, can_reshape);
1.1962 +}
1.1963 +
1.1964 +const Type* LoadUSNode::Value(PhaseTransform *phase) const {
1.1965 + Node* mem = in(MemNode::Memory);
1.1966 + Node* value = can_see_stored_value(mem,phase);
1.1967 + if (value != NULL && value->is_Con() &&
1.1968 + !value->bottom_type()->higher_equal(_type)) {
1.1969 + // If the input to the store does not fit with the load's result type,
1.1970 + // it must be truncated. We can't delay until Ideal call since
1.1971 + // a singleton Value is needed for split_thru_phi optimization.
1.1972 + int con = value->get_int();
1.1973 + return TypeInt::make(con & 0xFFFF);
1.1974 + }
1.1975 + return LoadNode::Value(phase);
1.1976 +}
1.1977 +
1.1978 +//--------------------------LoadSNode::Ideal--------------------------------------
1.1979 +//
1.1980 +// If the previous store is to the same address as this load,
1.1981 +// and the value stored was larger than a short, replace this load
1.1982 +// with the value stored truncated to a short. If no truncation is
1.1983 +// needed, the replacement is done in LoadNode::Identity().
1.1984 +//
1.1985 +Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1.1986 + Node* mem = in(MemNode::Memory);
1.1987 + Node* value = can_see_stored_value(mem,phase);
1.1988 + if( value && !phase->type(value)->higher_equal( _type ) ) {
1.1989 + Node *result = phase->transform( new (phase->C) LShiftINode(value, phase->intcon(16)) );
1.1990 + return new (phase->C) RShiftINode(result, phase->intcon(16));
1.1991 + }
1.1992 + // Identity call will handle the case where truncation is not needed.
1.1993 + return LoadNode::Ideal(phase, can_reshape);
1.1994 +}
1.1995 +
1.1996 +const Type* LoadSNode::Value(PhaseTransform *phase) const {
1.1997 + Node* mem = in(MemNode::Memory);
1.1998 + Node* value = can_see_stored_value(mem,phase);
1.1999 + if (value != NULL && value->is_Con() &&
1.2000 + !value->bottom_type()->higher_equal(_type)) {
1.2001 + // If the input to the store does not fit with the load's result type,
1.2002 + // it must be truncated. We can't delay until Ideal call since
1.2003 + // a singleton Value is needed for split_thru_phi optimization.
1.2004 + int con = value->get_int();
1.2005 + return TypeInt::make((con << 16) >> 16);
1.2006 + }
1.2007 + return LoadNode::Value(phase);
1.2008 +}
1.2009 +
1.2010 +//=============================================================================
1.2011 +//----------------------------LoadKlassNode::make------------------------------
1.2012 +// Polymorphic factory method:
1.2013 +Node *LoadKlassNode::make( PhaseGVN& gvn, Node *mem, Node *adr, const TypePtr* at, const TypeKlassPtr *tk ) {
1.2014 + Compile* C = gvn.C;
1.2015 + Node *ctl = NULL;
1.2016 + // sanity check the alias category against the created node type
1.2017 + const TypePtr *adr_type = adr->bottom_type()->isa_ptr();
1.2018 + assert(adr_type != NULL, "expecting TypeKlassPtr");
1.2019 +#ifdef _LP64
1.2020 + if (adr_type->is_ptr_to_narrowklass()) {
1.2021 + assert(UseCompressedClassPointers, "no compressed klasses");
1.2022 + Node* load_klass = gvn.transform(new (C) LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowklass(), MemNode::unordered));
1.2023 + return new (C) DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr());
1.2024 + }
1.2025 +#endif
1.2026 + assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");
1.2027 + return new (C) LoadKlassNode(ctl, mem, adr, at, tk, MemNode::unordered);
1.2028 +}
1.2029 +
1.2030 +//------------------------------Value------------------------------------------
1.2031 +const Type *LoadKlassNode::Value( PhaseTransform *phase ) const {
1.2032 + return klass_value_common(phase);
1.2033 +}
1.2034 +
1.2035 +const Type *LoadNode::klass_value_common( PhaseTransform *phase ) const {
1.2036 + // Either input is TOP ==> the result is TOP
1.2037 + const Type *t1 = phase->type( in(MemNode::Memory) );
1.2038 + if (t1 == Type::TOP) return Type::TOP;
1.2039 + Node *adr = in(MemNode::Address);
1.2040 + const Type *t2 = phase->type( adr );
1.2041 + if (t2 == Type::TOP) return Type::TOP;
1.2042 + const TypePtr *tp = t2->is_ptr();
1.2043 + if (TypePtr::above_centerline(tp->ptr()) ||
1.2044 + tp->ptr() == TypePtr::Null) return Type::TOP;
1.2045 +
1.2046 + // Return a more precise klass, if possible
1.2047 + const TypeInstPtr *tinst = tp->isa_instptr();
1.2048 + if (tinst != NULL) {
1.2049 + ciInstanceKlass* ik = tinst->klass()->as_instance_klass();
1.2050 + int offset = tinst->offset();
1.2051 + if (ik == phase->C->env()->Class_klass()
1.2052 + && (offset == java_lang_Class::klass_offset_in_bytes() ||
1.2053 + offset == java_lang_Class::array_klass_offset_in_bytes())) {
1.2054 + // We are loading a special hidden field from a Class mirror object,
1.2055 + // the field which points to the VM's Klass metaobject.
1.2056 + ciType* t = tinst->java_mirror_type();
1.2057 + // java_mirror_type returns non-null for compile-time Class constants.
1.2058 + if (t != NULL) {
1.2059 + // constant oop => constant klass
1.2060 + if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
1.2061 + if (t->is_void()) {
1.2062 + // We cannot create a void array. Since void is a primitive type return null
1.2063 + // klass. Users of this result need to do a null check on the returned klass.
1.2064 + return TypePtr::NULL_PTR;
1.2065 + }
1.2066 + return TypeKlassPtr::make(ciArrayKlass::make(t));
1.2067 + }
1.2068 + if (!t->is_klass()) {
1.2069 + // a primitive Class (e.g., int.class) has NULL for a klass field
1.2070 + return TypePtr::NULL_PTR;
1.2071 + }
1.2072 + // (Folds up the 1st indirection in aClassConstant.getModifiers().)
1.2073 + return TypeKlassPtr::make(t->as_klass());
1.2074 + }
1.2075 + // non-constant mirror, so we can't tell what's going on
1.2076 + }
1.2077 + if( !ik->is_loaded() )
1.2078 + return _type; // Bail out if not loaded
1.2079 + if (offset == oopDesc::klass_offset_in_bytes()) {
1.2080 + if (tinst->klass_is_exact()) {
1.2081 + return TypeKlassPtr::make(ik);
1.2082 + }
1.2083 + // See if we can become precise: no subklasses and no interface
1.2084 + // (Note: We need to support verified interfaces.)
1.2085 + if (!ik->is_interface() && !ik->has_subklass()) {
1.2086 + //assert(!UseExactTypes, "this code should be useless with exact types");
1.2087 + // Add a dependence; if any subclass added we need to recompile
1.2088 + if (!ik->is_final()) {
1.2089 + // %%% should use stronger assert_unique_concrete_subtype instead
1.2090 + phase->C->dependencies()->assert_leaf_type(ik);
1.2091 + }
1.2092 + // Return precise klass
1.2093 + return TypeKlassPtr::make(ik);
1.2094 + }
1.2095 +
1.2096 + // Return root of possible klass
1.2097 + return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/);
1.2098 + }
1.2099 + }
1.2100 +
1.2101 + // Check for loading klass from an array
1.2102 + const TypeAryPtr *tary = tp->isa_aryptr();
1.2103 + if( tary != NULL ) {
1.2104 + ciKlass *tary_klass = tary->klass();
1.2105 + if (tary_klass != NULL // can be NULL when at BOTTOM or TOP
1.2106 + && tary->offset() == oopDesc::klass_offset_in_bytes()) {
1.2107 + if (tary->klass_is_exact()) {
1.2108 + return TypeKlassPtr::make(tary_klass);
1.2109 + }
1.2110 + ciArrayKlass *ak = tary->klass()->as_array_klass();
1.2111 + // If the klass is an object array, we defer the question to the
1.2112 + // array component klass.
1.2113 + if( ak->is_obj_array_klass() ) {
1.2114 + assert( ak->is_loaded(), "" );
1.2115 + ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass();
1.2116 + if( base_k->is_loaded() && base_k->is_instance_klass() ) {
1.2117 + ciInstanceKlass* ik = base_k->as_instance_klass();
1.2118 + // See if we can become precise: no subklasses and no interface
1.2119 + if (!ik->is_interface() && !ik->has_subklass()) {
1.2120 + //assert(!UseExactTypes, "this code should be useless with exact types");
1.2121 + // Add a dependence; if any subclass added we need to recompile
1.2122 + if (!ik->is_final()) {
1.2123 + phase->C->dependencies()->assert_leaf_type(ik);
1.2124 + }
1.2125 + // Return precise array klass
1.2126 + return TypeKlassPtr::make(ak);
1.2127 + }
1.2128 + }
1.2129 + return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
1.2130 + } else { // Found a type-array?
1.2131 + //assert(!UseExactTypes, "this code should be useless with exact types");
1.2132 + assert( ak->is_type_array_klass(), "" );
1.2133 + return TypeKlassPtr::make(ak); // These are always precise
1.2134 + }
1.2135 + }
1.2136 + }
1.2137 +
1.2138 + // Check for loading klass from an array klass
1.2139 + const TypeKlassPtr *tkls = tp->isa_klassptr();
1.2140 + if (tkls != NULL && !StressReflectiveCode) {
1.2141 + ciKlass* klass = tkls->klass();
1.2142 + if( !klass->is_loaded() )
1.2143 + return _type; // Bail out if not loaded
1.2144 + if( klass->is_obj_array_klass() &&
1.2145 + tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) {
1.2146 + ciKlass* elem = klass->as_obj_array_klass()->element_klass();
1.2147 + // // Always returning precise element type is incorrect,
1.2148 + // // e.g., element type could be object and array may contain strings
1.2149 + // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
1.2150 +
1.2151 + // The array's TypeKlassPtr was declared 'precise' or 'not precise'
1.2152 + // according to the element type's subclassing.
1.2153 + return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/);
1.2154 + }
1.2155 + if( klass->is_instance_klass() && tkls->klass_is_exact() &&
1.2156 + tkls->offset() == in_bytes(Klass::super_offset())) {
1.2157 + ciKlass* sup = klass->as_instance_klass()->super();
1.2158 + // The field is Klass::_super. Return its (constant) value.
1.2159 + // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
1.2160 + return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR;
1.2161 + }
1.2162 + }
1.2163 +
1.2164 + // Bailout case
1.2165 + return LoadNode::Value(phase);
1.2166 +}
1.2167 +
1.2168 +//------------------------------Identity---------------------------------------
1.2169 +// To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
1.2170 +// Also feed through the klass in Allocate(...klass...)._klass.
1.2171 +Node* LoadKlassNode::Identity( PhaseTransform *phase ) {
1.2172 + return klass_identity_common(phase);
1.2173 +}
1.2174 +
1.2175 +Node* LoadNode::klass_identity_common(PhaseTransform *phase ) {
1.2176 + Node* x = LoadNode::Identity(phase);
1.2177 + if (x != this) return x;
1.2178 +
1.2179 + // Take apart the address into an oop and and offset.
1.2180 + // Return 'this' if we cannot.
1.2181 + Node* adr = in(MemNode::Address);
1.2182 + intptr_t offset = 0;
1.2183 + Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
1.2184 + if (base == NULL) return this;
1.2185 + const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
1.2186 + if (toop == NULL) return this;
1.2187 +
1.2188 + // We can fetch the klass directly through an AllocateNode.
1.2189 + // This works even if the klass is not constant (clone or newArray).
1.2190 + if (offset == oopDesc::klass_offset_in_bytes()) {
1.2191 + Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
1.2192 + if (allocated_klass != NULL) {
1.2193 + return allocated_klass;
1.2194 + }
1.2195 + }
1.2196 +
1.2197 + // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*.
1.2198 + // Simplify ak.component_mirror.array_klass to plain ak, ak an ArrayKlass.
1.2199 + // See inline_native_Class_query for occurrences of these patterns.
1.2200 + // Java Example: x.getClass().isAssignableFrom(y)
1.2201 + // Java Example: Array.newInstance(x.getClass().getComponentType(), n)
1.2202 + //
1.2203 + // This improves reflective code, often making the Class
1.2204 + // mirror go completely dead. (Current exception: Class
1.2205 + // mirrors may appear in debug info, but we could clean them out by
1.2206 + // introducing a new debug info operator for Klass*.java_mirror).
1.2207 + if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass()
1.2208 + && (offset == java_lang_Class::klass_offset_in_bytes() ||
1.2209 + offset == java_lang_Class::array_klass_offset_in_bytes())) {
1.2210 + // We are loading a special hidden field from a Class mirror,
1.2211 + // the field which points to its Klass or ArrayKlass metaobject.
1.2212 + if (base->is_Load()) {
1.2213 + Node* adr2 = base->in(MemNode::Address);
1.2214 + const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
1.2215 + if (tkls != NULL && !tkls->empty()
1.2216 + && (tkls->klass()->is_instance_klass() ||
1.2217 + tkls->klass()->is_array_klass())
1.2218 + && adr2->is_AddP()
1.2219 + ) {
1.2220 + int mirror_field = in_bytes(Klass::java_mirror_offset());
1.2221 + if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
1.2222 + mirror_field = in_bytes(ArrayKlass::component_mirror_offset());
1.2223 + }
1.2224 + if (tkls->offset() == mirror_field) {
1.2225 + return adr2->in(AddPNode::Base);
1.2226 + }
1.2227 + }
1.2228 + }
1.2229 + }
1.2230 +
1.2231 + return this;
1.2232 +}
1.2233 +
1.2234 +
1.2235 +//------------------------------Value------------------------------------------
1.2236 +const Type *LoadNKlassNode::Value( PhaseTransform *phase ) const {
1.2237 + const Type *t = klass_value_common(phase);
1.2238 + if (t == Type::TOP)
1.2239 + return t;
1.2240 +
1.2241 + return t->make_narrowklass();
1.2242 +}
1.2243 +
1.2244 +//------------------------------Identity---------------------------------------
1.2245 +// To clean up reflective code, simplify k.java_mirror.as_klass to narrow k.
1.2246 +// Also feed through the klass in Allocate(...klass...)._klass.
1.2247 +Node* LoadNKlassNode::Identity( PhaseTransform *phase ) {
1.2248 + Node *x = klass_identity_common(phase);
1.2249 +
1.2250 + const Type *t = phase->type( x );
1.2251 + if( t == Type::TOP ) return x;
1.2252 + if( t->isa_narrowklass()) return x;
1.2253 + assert (!t->isa_narrowoop(), "no narrow oop here");
1.2254 +
1.2255 + return phase->transform(new (phase->C) EncodePKlassNode(x, t->make_narrowklass()));
1.2256 +}
1.2257 +
1.2258 +//------------------------------Value-----------------------------------------
1.2259 +const Type *LoadRangeNode::Value( PhaseTransform *phase ) const {
1.2260 + // Either input is TOP ==> the result is TOP
1.2261 + const Type *t1 = phase->type( in(MemNode::Memory) );
1.2262 + if( t1 == Type::TOP ) return Type::TOP;
1.2263 + Node *adr = in(MemNode::Address);
1.2264 + const Type *t2 = phase->type( adr );
1.2265 + if( t2 == Type::TOP ) return Type::TOP;
1.2266 + const TypePtr *tp = t2->is_ptr();
1.2267 + if (TypePtr::above_centerline(tp->ptr())) return Type::TOP;
1.2268 + const TypeAryPtr *tap = tp->isa_aryptr();
1.2269 + if( !tap ) return _type;
1.2270 + return tap->size();
1.2271 +}
1.2272 +
1.2273 +//-------------------------------Ideal---------------------------------------
1.2274 +// Feed through the length in AllocateArray(...length...)._length.
1.2275 +Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1.2276 + Node* p = MemNode::Ideal_common(phase, can_reshape);
1.2277 + if (p) return (p == NodeSentinel) ? NULL : p;
1.2278 +
1.2279 + // Take apart the address into an oop and and offset.
1.2280 + // Return 'this' if we cannot.
1.2281 + Node* adr = in(MemNode::Address);
1.2282 + intptr_t offset = 0;
1.2283 + Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
1.2284 + if (base == NULL) return NULL;
1.2285 + const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
1.2286 + if (tary == NULL) return NULL;
1.2287 +
1.2288 + // We can fetch the length directly through an AllocateArrayNode.
1.2289 + // This works even if the length is not constant (clone or newArray).
1.2290 + if (offset == arrayOopDesc::length_offset_in_bytes()) {
1.2291 + AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
1.2292 + if (alloc != NULL) {
1.2293 + Node* allocated_length = alloc->Ideal_length();
1.2294 + Node* len = alloc->make_ideal_length(tary, phase);
1.2295 + if (allocated_length != len) {
1.2296 + // New CastII improves on this.
1.2297 + return len;
1.2298 + }
1.2299 + }
1.2300 + }
1.2301 +
1.2302 + return NULL;
1.2303 +}
1.2304 +
1.2305 +//------------------------------Identity---------------------------------------
1.2306 +// Feed through the length in AllocateArray(...length...)._length.
1.2307 +Node* LoadRangeNode::Identity( PhaseTransform *phase ) {
1.2308 + Node* x = LoadINode::Identity(phase);
1.2309 + if (x != this) return x;
1.2310 +
1.2311 + // Take apart the address into an oop and and offset.
1.2312 + // Return 'this' if we cannot.
1.2313 + Node* adr = in(MemNode::Address);
1.2314 + intptr_t offset = 0;
1.2315 + Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
1.2316 + if (base == NULL) return this;
1.2317 + const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
1.2318 + if (tary == NULL) return this;
1.2319 +
1.2320 + // We can fetch the length directly through an AllocateArrayNode.
1.2321 + // This works even if the length is not constant (clone or newArray).
1.2322 + if (offset == arrayOopDesc::length_offset_in_bytes()) {
1.2323 + AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
1.2324 + if (alloc != NULL) {
1.2325 + Node* allocated_length = alloc->Ideal_length();
1.2326 + // Do not allow make_ideal_length to allocate a CastII node.
1.2327 + Node* len = alloc->make_ideal_length(tary, phase, false);
1.2328 + if (allocated_length == len) {
1.2329 + // Return allocated_length only if it would not be improved by a CastII.
1.2330 + return allocated_length;
1.2331 + }
1.2332 + }
1.2333 + }
1.2334 +
1.2335 + return this;
1.2336 +
1.2337 +}
1.2338 +
1.2339 +//=============================================================================
1.2340 +//---------------------------StoreNode::make-----------------------------------
1.2341 +// Polymorphic factory method:
1.2342 +StoreNode* StoreNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt, MemOrd mo) {
1.2343 + assert((mo == unordered || mo == release), "unexpected");
1.2344 + Compile* C = gvn.C;
1.2345 + assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
1.2346 + ctl != NULL, "raw memory operations should have control edge");
1.2347 +
1.2348 + switch (bt) {
1.2349 + case T_BOOLEAN:
1.2350 + case T_BYTE: return new (C) StoreBNode(ctl, mem, adr, adr_type, val, mo);
1.2351 + case T_INT: return new (C) StoreINode(ctl, mem, adr, adr_type, val, mo);
1.2352 + case T_CHAR:
1.2353 + case T_SHORT: return new (C) StoreCNode(ctl, mem, adr, adr_type, val, mo);
1.2354 + case T_LONG: return new (C) StoreLNode(ctl, mem, adr, adr_type, val, mo);
1.2355 + case T_FLOAT: return new (C) StoreFNode(ctl, mem, adr, adr_type, val, mo);
1.2356 + case T_DOUBLE: return new (C) StoreDNode(ctl, mem, adr, adr_type, val, mo);
1.2357 + case T_METADATA:
1.2358 + case T_ADDRESS:
1.2359 + case T_OBJECT:
1.2360 +#ifdef _LP64
1.2361 + if (adr->bottom_type()->is_ptr_to_narrowoop()) {
1.2362 + val = gvn.transform(new (C) EncodePNode(val, val->bottom_type()->make_narrowoop()));
1.2363 + return new (C) StoreNNode(ctl, mem, adr, adr_type, val, mo);
1.2364 + } else if (adr->bottom_type()->is_ptr_to_narrowklass() ||
1.2365 + (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() &&
1.2366 + adr->bottom_type()->isa_rawptr())) {
1.2367 + val = gvn.transform(new (C) EncodePKlassNode(val, val->bottom_type()->make_narrowklass()));
1.2368 + return new (C) StoreNKlassNode(ctl, mem, adr, adr_type, val, mo);
1.2369 + }
1.2370 +#endif
1.2371 + {
1.2372 + return new (C) StorePNode(ctl, mem, adr, adr_type, val, mo);
1.2373 + }
1.2374 + }
1.2375 + ShouldNotReachHere();
1.2376 + return (StoreNode*)NULL;
1.2377 +}
1.2378 +
1.2379 +StoreLNode* StoreLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) {
1.2380 + bool require_atomic = true;
1.2381 + return new (C) StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic);
1.2382 +}
1.2383 +
1.2384 +
1.2385 +//--------------------------bottom_type----------------------------------------
1.2386 +const Type *StoreNode::bottom_type() const {
1.2387 + return Type::MEMORY;
1.2388 +}
1.2389 +
1.2390 +//------------------------------hash-------------------------------------------
1.2391 +uint StoreNode::hash() const {
1.2392 + // unroll addition of interesting fields
1.2393 + //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
1.2394 +
1.2395 + // Since they are not commoned, do not hash them:
1.2396 + return NO_HASH;
1.2397 +}
1.2398 +
1.2399 +//------------------------------Ideal------------------------------------------
1.2400 +// Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
1.2401 +// When a store immediately follows a relevant allocation/initialization,
1.2402 +// try to capture it into the initialization, or hoist it above.
1.2403 +Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1.2404 + Node* p = MemNode::Ideal_common(phase, can_reshape);
1.2405 + if (p) return (p == NodeSentinel) ? NULL : p;
1.2406 +
1.2407 + Node* mem = in(MemNode::Memory);
1.2408 + Node* address = in(MemNode::Address);
1.2409 +
1.2410 + // Back-to-back stores to same address? Fold em up. Generally
1.2411 + // unsafe if I have intervening uses... Also disallowed for StoreCM
1.2412 + // since they must follow each StoreP operation. Redundant StoreCMs
1.2413 + // are eliminated just before matching in final_graph_reshape.
1.2414 + if (mem->is_Store() && mem->in(MemNode::Address)->eqv_uncast(address) &&
1.2415 + mem->Opcode() != Op_StoreCM) {
1.2416 + // Looking at a dead closed cycle of memory?
1.2417 + assert(mem != mem->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
1.2418 +
1.2419 + assert(Opcode() == mem->Opcode() ||
1.2420 + phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw,
1.2421 + "no mismatched stores, except on raw memory");
1.2422 +
1.2423 + if (mem->outcnt() == 1 && // check for intervening uses
1.2424 + mem->as_Store()->memory_size() <= this->memory_size()) {
1.2425 + // If anybody other than 'this' uses 'mem', we cannot fold 'mem' away.
1.2426 + // For example, 'mem' might be the final state at a conditional return.
1.2427 + // Or, 'mem' might be used by some node which is live at the same time
1.2428 + // 'this' is live, which might be unschedulable. So, require exactly
1.2429 + // ONE user, the 'this' store, until such time as we clone 'mem' for
1.2430 + // each of 'mem's uses (thus making the exactly-1-user-rule hold true).
1.2431 + if (can_reshape) { // (%%% is this an anachronism?)
1.2432 + set_req_X(MemNode::Memory, mem->in(MemNode::Memory),
1.2433 + phase->is_IterGVN());
1.2434 + } else {
1.2435 + // It's OK to do this in the parser, since DU info is always accurate,
1.2436 + // and the parser always refers to nodes via SafePointNode maps.
1.2437 + set_req(MemNode::Memory, mem->in(MemNode::Memory));
1.2438 + }
1.2439 + return this;
1.2440 + }
1.2441 + }
1.2442 +
1.2443 + // Capture an unaliased, unconditional, simple store into an initializer.
1.2444 + // Or, if it is independent of the allocation, hoist it above the allocation.
1.2445 + if (ReduceFieldZeroing && /*can_reshape &&*/
1.2446 + mem->is_Proj() && mem->in(0)->is_Initialize()) {
1.2447 + InitializeNode* init = mem->in(0)->as_Initialize();
1.2448 + intptr_t offset = init->can_capture_store(this, phase, can_reshape);
1.2449 + if (offset > 0) {
1.2450 + Node* moved = init->capture_store(this, offset, phase, can_reshape);
1.2451 + // If the InitializeNode captured me, it made a raw copy of me,
1.2452 + // and I need to disappear.
1.2453 + if (moved != NULL) {
1.2454 + // %%% hack to ensure that Ideal returns a new node:
1.2455 + mem = MergeMemNode::make(phase->C, mem);
1.2456 + return mem; // fold me away
1.2457 + }
1.2458 + }
1.2459 + }
1.2460 +
1.2461 + return NULL; // No further progress
1.2462 +}
1.2463 +
1.2464 +//------------------------------Value-----------------------------------------
1.2465 +const Type *StoreNode::Value( PhaseTransform *phase ) const {
1.2466 + // Either input is TOP ==> the result is TOP
1.2467 + const Type *t1 = phase->type( in(MemNode::Memory) );
1.2468 + if( t1 == Type::TOP ) return Type::TOP;
1.2469 + const Type *t2 = phase->type( in(MemNode::Address) );
1.2470 + if( t2 == Type::TOP ) return Type::TOP;
1.2471 + const Type *t3 = phase->type( in(MemNode::ValueIn) );
1.2472 + if( t3 == Type::TOP ) return Type::TOP;
1.2473 + return Type::MEMORY;
1.2474 +}
1.2475 +
1.2476 +//------------------------------Identity---------------------------------------
1.2477 +// Remove redundant stores:
1.2478 +// Store(m, p, Load(m, p)) changes to m.
1.2479 +// Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
1.2480 +Node *StoreNode::Identity( PhaseTransform *phase ) {
1.2481 + Node* mem = in(MemNode::Memory);
1.2482 + Node* adr = in(MemNode::Address);
1.2483 + Node* val = in(MemNode::ValueIn);
1.2484 +
1.2485 + // Load then Store? Then the Store is useless
1.2486 + if (val->is_Load() &&
1.2487 + val->in(MemNode::Address)->eqv_uncast(adr) &&
1.2488 + val->in(MemNode::Memory )->eqv_uncast(mem) &&
1.2489 + val->as_Load()->store_Opcode() == Opcode()) {
1.2490 + return mem;
1.2491 + }
1.2492 +
1.2493 + // Two stores in a row of the same value?
1.2494 + if (mem->is_Store() &&
1.2495 + mem->in(MemNode::Address)->eqv_uncast(adr) &&
1.2496 + mem->in(MemNode::ValueIn)->eqv_uncast(val) &&
1.2497 + mem->Opcode() == Opcode()) {
1.2498 + return mem;
1.2499 + }
1.2500 +
1.2501 + // Store of zero anywhere into a freshly-allocated object?
1.2502 + // Then the store is useless.
1.2503 + // (It must already have been captured by the InitializeNode.)
1.2504 + if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) {
1.2505 + // a newly allocated object is already all-zeroes everywhere
1.2506 + if (mem->is_Proj() && mem->in(0)->is_Allocate()) {
1.2507 + return mem;
1.2508 + }
1.2509 +
1.2510 + // the store may also apply to zero-bits in an earlier object
1.2511 + Node* prev_mem = find_previous_store(phase);
1.2512 + // Steps (a), (b): Walk past independent stores to find an exact match.
1.2513 + if (prev_mem != NULL) {
1.2514 + Node* prev_val = can_see_stored_value(prev_mem, phase);
1.2515 + if (prev_val != NULL && phase->eqv(prev_val, val)) {
1.2516 + // prev_val and val might differ by a cast; it would be good
1.2517 + // to keep the more informative of the two.
1.2518 + return mem;
1.2519 + }
1.2520 + }
1.2521 + }
1.2522 +
1.2523 + return this;
1.2524 +}
1.2525 +
1.2526 +//------------------------------match_edge-------------------------------------
1.2527 +// Do we Match on this edge index or not? Match only memory & value
1.2528 +uint StoreNode::match_edge(uint idx) const {
1.2529 + return idx == MemNode::Address || idx == MemNode::ValueIn;
1.2530 +}
1.2531 +
1.2532 +//------------------------------cmp--------------------------------------------
1.2533 +// Do not common stores up together. They generally have to be split
1.2534 +// back up anyways, so do not bother.
1.2535 +uint StoreNode::cmp( const Node &n ) const {
1.2536 + return (&n == this); // Always fail except on self
1.2537 +}
1.2538 +
1.2539 +//------------------------------Ideal_masked_input-----------------------------
1.2540 +// Check for a useless mask before a partial-word store
1.2541 +// (StoreB ... (AndI valIn conIa) )
1.2542 +// If (conIa & mask == mask) this simplifies to
1.2543 +// (StoreB ... (valIn) )
1.2544 +Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
1.2545 + Node *val = in(MemNode::ValueIn);
1.2546 + if( val->Opcode() == Op_AndI ) {
1.2547 + const TypeInt *t = phase->type( val->in(2) )->isa_int();
1.2548 + if( t && t->is_con() && (t->get_con() & mask) == mask ) {
1.2549 + set_req(MemNode::ValueIn, val->in(1));
1.2550 + return this;
1.2551 + }
1.2552 + }
1.2553 + return NULL;
1.2554 +}
1.2555 +
1.2556 +
1.2557 +//------------------------------Ideal_sign_extended_input----------------------
1.2558 +// Check for useless sign-extension before a partial-word store
1.2559 +// (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) )
1.2560 +// If (conIL == conIR && conIR <= num_bits) this simplifies to
1.2561 +// (StoreB ... (valIn) )
1.2562 +Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) {
1.2563 + Node *val = in(MemNode::ValueIn);
1.2564 + if( val->Opcode() == Op_RShiftI ) {
1.2565 + const TypeInt *t = phase->type( val->in(2) )->isa_int();
1.2566 + if( t && t->is_con() && (t->get_con() <= num_bits) ) {
1.2567 + Node *shl = val->in(1);
1.2568 + if( shl->Opcode() == Op_LShiftI ) {
1.2569 + const TypeInt *t2 = phase->type( shl->in(2) )->isa_int();
1.2570 + if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) {
1.2571 + set_req(MemNode::ValueIn, shl->in(1));
1.2572 + return this;
1.2573 + }
1.2574 + }
1.2575 + }
1.2576 + }
1.2577 + return NULL;
1.2578 +}
1.2579 +
1.2580 +//------------------------------value_never_loaded-----------------------------------
1.2581 +// Determine whether there are any possible loads of the value stored.
1.2582 +// For simplicity, we actually check if there are any loads from the
1.2583 +// address stored to, not just for loads of the value stored by this node.
1.2584 +//
1.2585 +bool StoreNode::value_never_loaded( PhaseTransform *phase) const {
1.2586 + Node *adr = in(Address);
1.2587 + const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
1.2588 + if (adr_oop == NULL)
1.2589 + return false;
1.2590 + if (!adr_oop->is_known_instance_field())
1.2591 + return false; // if not a distinct instance, there may be aliases of the address
1.2592 + for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
1.2593 + Node *use = adr->fast_out(i);
1.2594 + int opc = use->Opcode();
1.2595 + if (use->is_Load() || use->is_LoadStore()) {
1.2596 + return false;
1.2597 + }
1.2598 + }
1.2599 + return true;
1.2600 +}
1.2601 +
1.2602 +//=============================================================================
1.2603 +//------------------------------Ideal------------------------------------------
1.2604 +// If the store is from an AND mask that leaves the low bits untouched, then
1.2605 +// we can skip the AND operation. If the store is from a sign-extension
1.2606 +// (a left shift, then right shift) we can skip both.
1.2607 +Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
1.2608 + Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
1.2609 + if( progress != NULL ) return progress;
1.2610 +
1.2611 + progress = StoreNode::Ideal_sign_extended_input(phase, 24);
1.2612 + if( progress != NULL ) return progress;
1.2613 +
1.2614 + // Finally check the default case
1.2615 + return StoreNode::Ideal(phase, can_reshape);
1.2616 +}
1.2617 +
1.2618 +//=============================================================================
1.2619 +//------------------------------Ideal------------------------------------------
1.2620 +// If the store is from an AND mask that leaves the low bits untouched, then
1.2621 +// we can skip the AND operation
1.2622 +Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
1.2623 + Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
1.2624 + if( progress != NULL ) return progress;
1.2625 +
1.2626 + progress = StoreNode::Ideal_sign_extended_input(phase, 16);
1.2627 + if( progress != NULL ) return progress;
1.2628 +
1.2629 + // Finally check the default case
1.2630 + return StoreNode::Ideal(phase, can_reshape);
1.2631 +}
1.2632 +
1.2633 +//=============================================================================
1.2634 +//------------------------------Identity---------------------------------------
1.2635 +Node *StoreCMNode::Identity( PhaseTransform *phase ) {
1.2636 + // No need to card mark when storing a null ptr
1.2637 + Node* my_store = in(MemNode::OopStore);
1.2638 + if (my_store->is_Store()) {
1.2639 + const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) );
1.2640 + if( t1 == TypePtr::NULL_PTR ) {
1.2641 + return in(MemNode::Memory);
1.2642 + }
1.2643 + }
1.2644 + return this;
1.2645 +}
1.2646 +
1.2647 +//=============================================================================
1.2648 +//------------------------------Ideal---------------------------------------
1.2649 +Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){
1.2650 + Node* progress = StoreNode::Ideal(phase, can_reshape);
1.2651 + if (progress != NULL) return progress;
1.2652 +
1.2653 + Node* my_store = in(MemNode::OopStore);
1.2654 + if (my_store->is_MergeMem()) {
1.2655 + Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx());
1.2656 + set_req(MemNode::OopStore, mem);
1.2657 + return this;
1.2658 + }
1.2659 +
1.2660 + return NULL;
1.2661 +}
1.2662 +
1.2663 +//------------------------------Value-----------------------------------------
1.2664 +const Type *StoreCMNode::Value( PhaseTransform *phase ) const {
1.2665 + // Either input is TOP ==> the result is TOP
1.2666 + const Type *t = phase->type( in(MemNode::Memory) );
1.2667 + if( t == Type::TOP ) return Type::TOP;
1.2668 + t = phase->type( in(MemNode::Address) );
1.2669 + if( t == Type::TOP ) return Type::TOP;
1.2670 + t = phase->type( in(MemNode::ValueIn) );
1.2671 + if( t == Type::TOP ) return Type::TOP;
1.2672 + // If extra input is TOP ==> the result is TOP
1.2673 + t = phase->type( in(MemNode::OopStore) );
1.2674 + if( t == Type::TOP ) return Type::TOP;
1.2675 +
1.2676 + return StoreNode::Value( phase );
1.2677 +}
1.2678 +
1.2679 +
1.2680 +//=============================================================================
1.2681 +//----------------------------------SCMemProjNode------------------------------
1.2682 +const Type * SCMemProjNode::Value( PhaseTransform *phase ) const
1.2683 +{
1.2684 + return bottom_type();
1.2685 +}
1.2686 +
1.2687 +//=============================================================================
1.2688 +//----------------------------------LoadStoreNode------------------------------
1.2689 +LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required )
1.2690 + : Node(required),
1.2691 + _type(rt),
1.2692 + _adr_type(at)
1.2693 +{
1.2694 + init_req(MemNode::Control, c );
1.2695 + init_req(MemNode::Memory , mem);
1.2696 + init_req(MemNode::Address, adr);
1.2697 + init_req(MemNode::ValueIn, val);
1.2698 + init_class_id(Class_LoadStore);
1.2699 +}
1.2700 +
1.2701 +uint LoadStoreNode::ideal_reg() const {
1.2702 + return _type->ideal_reg();
1.2703 +}
1.2704 +
1.2705 +bool LoadStoreNode::result_not_used() const {
1.2706 + for( DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++ ) {
1.2707 + Node *x = fast_out(i);
1.2708 + if (x->Opcode() == Op_SCMemProj) continue;
1.2709 + return false;
1.2710 + }
1.2711 + return true;
1.2712 +}
1.2713 +
1.2714 +uint LoadStoreNode::size_of() const { return sizeof(*this); }
1.2715 +
1.2716 +//=============================================================================
1.2717 +//----------------------------------LoadStoreConditionalNode--------------------
1.2718 +LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, NULL, TypeInt::BOOL, 5) {
1.2719 + init_req(ExpectedIn, ex );
1.2720 +}
1.2721 +
1.2722 +//=============================================================================
1.2723 +//-------------------------------adr_type--------------------------------------
1.2724 +// Do we Match on this edge index or not? Do not match memory
1.2725 +const TypePtr* ClearArrayNode::adr_type() const {
1.2726 + Node *adr = in(3);
1.2727 + return MemNode::calculate_adr_type(adr->bottom_type());
1.2728 +}
1.2729 +
1.2730 +//------------------------------match_edge-------------------------------------
1.2731 +// Do we Match on this edge index or not? Do not match memory
1.2732 +uint ClearArrayNode::match_edge(uint idx) const {
1.2733 + return idx > 1;
1.2734 +}
1.2735 +
1.2736 +//------------------------------Identity---------------------------------------
1.2737 +// Clearing a zero length array does nothing
1.2738 +Node *ClearArrayNode::Identity( PhaseTransform *phase ) {
1.2739 + return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this;
1.2740 +}
1.2741 +
1.2742 +//------------------------------Idealize---------------------------------------
1.2743 +// Clearing a short array is faster with stores
1.2744 +Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){
1.2745 + const int unit = BytesPerLong;
1.2746 + const TypeX* t = phase->type(in(2))->isa_intptr_t();
1.2747 + if (!t) return NULL;
1.2748 + if (!t->is_con()) return NULL;
1.2749 + intptr_t raw_count = t->get_con();
1.2750 + intptr_t size = raw_count;
1.2751 + if (!Matcher::init_array_count_is_in_bytes) size *= unit;
1.2752 + // Clearing nothing uses the Identity call.
1.2753 + // Negative clears are possible on dead ClearArrays
1.2754 + // (see jck test stmt114.stmt11402.val).
1.2755 + if (size <= 0 || size % unit != 0) return NULL;
1.2756 + intptr_t count = size / unit;
1.2757 + // Length too long; use fast hardware clear
1.2758 + if (size > Matcher::init_array_short_size) return NULL;
1.2759 + Node *mem = in(1);
1.2760 + if( phase->type(mem)==Type::TOP ) return NULL;
1.2761 + Node *adr = in(3);
1.2762 + const Type* at = phase->type(adr);
1.2763 + if( at==Type::TOP ) return NULL;
1.2764 + const TypePtr* atp = at->isa_ptr();
1.2765 + // adjust atp to be the correct array element address type
1.2766 + if (atp == NULL) atp = TypePtr::BOTTOM;
1.2767 + else atp = atp->add_offset(Type::OffsetBot);
1.2768 + // Get base for derived pointer purposes
1.2769 + if( adr->Opcode() != Op_AddP ) Unimplemented();
1.2770 + Node *base = adr->in(1);
1.2771 +
1.2772 + Node *zero = phase->makecon(TypeLong::ZERO);
1.2773 + Node *off = phase->MakeConX(BytesPerLong);
1.2774 + mem = new (phase->C) StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
1.2775 + count--;
1.2776 + while( count-- ) {
1.2777 + mem = phase->transform(mem);
1.2778 + adr = phase->transform(new (phase->C) AddPNode(base,adr,off));
1.2779 + mem = new (phase->C) StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
1.2780 + }
1.2781 + return mem;
1.2782 +}
1.2783 +
1.2784 +//----------------------------step_through----------------------------------
1.2785 +// Return allocation input memory edge if it is different instance
1.2786 +// or itself if it is the one we are looking for.
1.2787 +bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) {
1.2788 + Node* n = *np;
1.2789 + assert(n->is_ClearArray(), "sanity");
1.2790 + intptr_t offset;
1.2791 + AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset);
1.2792 + // This method is called only before Allocate nodes are expanded during
1.2793 + // macro nodes expansion. Before that ClearArray nodes are only generated
1.2794 + // in LibraryCallKit::generate_arraycopy() which follows allocations.
1.2795 + assert(alloc != NULL, "should have allocation");
1.2796 + if (alloc->_idx == instance_id) {
1.2797 + // Can not bypass initialization of the instance we are looking for.
1.2798 + return false;
1.2799 + }
1.2800 + // Otherwise skip it.
1.2801 + InitializeNode* init = alloc->initialization();
1.2802 + if (init != NULL)
1.2803 + *np = init->in(TypeFunc::Memory);
1.2804 + else
1.2805 + *np = alloc->in(TypeFunc::Memory);
1.2806 + return true;
1.2807 +}
1.2808 +
1.2809 +//----------------------------clear_memory-------------------------------------
1.2810 +// Generate code to initialize object storage to zero.
1.2811 +Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
1.2812 + intptr_t start_offset,
1.2813 + Node* end_offset,
1.2814 + PhaseGVN* phase) {
1.2815 + Compile* C = phase->C;
1.2816 + intptr_t offset = start_offset;
1.2817 +
1.2818 + int unit = BytesPerLong;
1.2819 + if ((offset % unit) != 0) {
1.2820 + Node* adr = new (C) AddPNode(dest, dest, phase->MakeConX(offset));
1.2821 + adr = phase->transform(adr);
1.2822 + const TypePtr* atp = TypeRawPtr::BOTTOM;
1.2823 + mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
1.2824 + mem = phase->transform(mem);
1.2825 + offset += BytesPerInt;
1.2826 + }
1.2827 + assert((offset % unit) == 0, "");
1.2828 +
1.2829 + // Initialize the remaining stuff, if any, with a ClearArray.
1.2830 + return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase);
1.2831 +}
1.2832 +
1.2833 +Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
1.2834 + Node* start_offset,
1.2835 + Node* end_offset,
1.2836 + PhaseGVN* phase) {
1.2837 + if (start_offset == end_offset) {
1.2838 + // nothing to do
1.2839 + return mem;
1.2840 + }
1.2841 +
1.2842 + Compile* C = phase->C;
1.2843 + int unit = BytesPerLong;
1.2844 + Node* zbase = start_offset;
1.2845 + Node* zend = end_offset;
1.2846 +
1.2847 + // Scale to the unit required by the CPU:
1.2848 + if (!Matcher::init_array_count_is_in_bytes) {
1.2849 + Node* shift = phase->intcon(exact_log2(unit));
1.2850 + zbase = phase->transform( new(C) URShiftXNode(zbase, shift) );
1.2851 + zend = phase->transform( new(C) URShiftXNode(zend, shift) );
1.2852 + }
1.2853 +
1.2854 + // Bulk clear double-words
1.2855 + Node* zsize = phase->transform( new(C) SubXNode(zend, zbase) );
1.2856 + Node* adr = phase->transform( new(C) AddPNode(dest, dest, start_offset) );
1.2857 + mem = new (C) ClearArrayNode(ctl, mem, zsize, adr);
1.2858 + return phase->transform(mem);
1.2859 +}
1.2860 +
1.2861 +Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
1.2862 + intptr_t start_offset,
1.2863 + intptr_t end_offset,
1.2864 + PhaseGVN* phase) {
1.2865 + if (start_offset == end_offset) {
1.2866 + // nothing to do
1.2867 + return mem;
1.2868 + }
1.2869 +
1.2870 + Compile* C = phase->C;
1.2871 + assert((end_offset % BytesPerInt) == 0, "odd end offset");
1.2872 + intptr_t done_offset = end_offset;
1.2873 + if ((done_offset % BytesPerLong) != 0) {
1.2874 + done_offset -= BytesPerInt;
1.2875 + }
1.2876 + if (done_offset > start_offset) {
1.2877 + mem = clear_memory(ctl, mem, dest,
1.2878 + start_offset, phase->MakeConX(done_offset), phase);
1.2879 + }
1.2880 + if (done_offset < end_offset) { // emit the final 32-bit store
1.2881 + Node* adr = new (C) AddPNode(dest, dest, phase->MakeConX(done_offset));
1.2882 + adr = phase->transform(adr);
1.2883 + const TypePtr* atp = TypeRawPtr::BOTTOM;
1.2884 + mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
1.2885 + mem = phase->transform(mem);
1.2886 + done_offset += BytesPerInt;
1.2887 + }
1.2888 + assert(done_offset == end_offset, "");
1.2889 + return mem;
1.2890 +}
1.2891 +
1.2892 +//=============================================================================
1.2893 +// Do not match memory edge.
1.2894 +uint StrIntrinsicNode::match_edge(uint idx) const {
1.2895 + return idx == 2 || idx == 3;
1.2896 +}
1.2897 +
1.2898 +//------------------------------Ideal------------------------------------------
1.2899 +// Return a node which is more "ideal" than the current node. Strip out
1.2900 +// control copies
1.2901 +Node *StrIntrinsicNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1.2902 + if (remove_dead_region(phase, can_reshape)) return this;
1.2903 + // Don't bother trying to transform a dead node
1.2904 + if (in(0) && in(0)->is_top()) return NULL;
1.2905 +
1.2906 + if (can_reshape) {
1.2907 + Node* mem = phase->transform(in(MemNode::Memory));
1.2908 + // If transformed to a MergeMem, get the desired slice
1.2909 + uint alias_idx = phase->C->get_alias_index(adr_type());
1.2910 + mem = mem->is_MergeMem() ? mem->as_MergeMem()->memory_at(alias_idx) : mem;
1.2911 + if (mem != in(MemNode::Memory)) {
1.2912 + set_req(MemNode::Memory, mem);
1.2913 + return this;
1.2914 + }
1.2915 + }
1.2916 + return NULL;
1.2917 +}
1.2918 +
1.2919 +//------------------------------Value------------------------------------------
1.2920 +const Type *StrIntrinsicNode::Value( PhaseTransform *phase ) const {
1.2921 + if (in(0) && phase->type(in(0)) == Type::TOP) return Type::TOP;
1.2922 + return bottom_type();
1.2923 +}
1.2924 +
1.2925 +//=============================================================================
1.2926 +//------------------------------match_edge-------------------------------------
1.2927 +// Do not match memory edge
1.2928 +uint EncodeISOArrayNode::match_edge(uint idx) const {
1.2929 + return idx == 2 || idx == 3; // EncodeISOArray src (Binary dst len)
1.2930 +}
1.2931 +
1.2932 +//------------------------------Ideal------------------------------------------
1.2933 +// Return a node which is more "ideal" than the current node. Strip out
1.2934 +// control copies
1.2935 +Node *EncodeISOArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1.2936 + return remove_dead_region(phase, can_reshape) ? this : NULL;
1.2937 +}
1.2938 +
1.2939 +//------------------------------Value------------------------------------------
1.2940 +const Type *EncodeISOArrayNode::Value(PhaseTransform *phase) const {
1.2941 + if (in(0) && phase->type(in(0)) == Type::TOP) return Type::TOP;
1.2942 + return bottom_type();
1.2943 +}
1.2944 +
1.2945 +//=============================================================================
1.2946 +MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
1.2947 + : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)),
1.2948 + _adr_type(C->get_adr_type(alias_idx))
1.2949 +{
1.2950 + init_class_id(Class_MemBar);
1.2951 + Node* top = C->top();
1.2952 + init_req(TypeFunc::I_O,top);
1.2953 + init_req(TypeFunc::FramePtr,top);
1.2954 + init_req(TypeFunc::ReturnAdr,top);
1.2955 + if (precedent != NULL)
1.2956 + init_req(TypeFunc::Parms, precedent);
1.2957 +}
1.2958 +
1.2959 +//------------------------------cmp--------------------------------------------
1.2960 +uint MemBarNode::hash() const { return NO_HASH; }
1.2961 +uint MemBarNode::cmp( const Node &n ) const {
1.2962 + return (&n == this); // Always fail except on self
1.2963 +}
1.2964 +
1.2965 +//------------------------------make-------------------------------------------
1.2966 +MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
1.2967 + switch (opcode) {
1.2968 + case Op_MemBarAcquire: return new(C) MemBarAcquireNode(C, atp, pn);
1.2969 + case Op_LoadFence: return new(C) LoadFenceNode(C, atp, pn);
1.2970 + case Op_MemBarRelease: return new(C) MemBarReleaseNode(C, atp, pn);
1.2971 + case Op_StoreFence: return new(C) StoreFenceNode(C, atp, pn);
1.2972 + case Op_MemBarAcquireLock: return new(C) MemBarAcquireLockNode(C, atp, pn);
1.2973 + case Op_MemBarReleaseLock: return new(C) MemBarReleaseLockNode(C, atp, pn);
1.2974 + case Op_MemBarVolatile: return new(C) MemBarVolatileNode(C, atp, pn);
1.2975 + case Op_MemBarCPUOrder: return new(C) MemBarCPUOrderNode(C, atp, pn);
1.2976 + case Op_Initialize: return new(C) InitializeNode(C, atp, pn);
1.2977 + case Op_MemBarStoreStore: return new(C) MemBarStoreStoreNode(C, atp, pn);
1.2978 + default: ShouldNotReachHere(); return NULL;
1.2979 + }
1.2980 +}
1.2981 +
1.2982 +//------------------------------Ideal------------------------------------------
1.2983 +// Return a node which is more "ideal" than the current node. Strip out
1.2984 +// control copies
1.2985 +Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1.2986 + if (remove_dead_region(phase, can_reshape)) return this;
1.2987 + // Don't bother trying to transform a dead node
1.2988 + if (in(0) && in(0)->is_top()) {
1.2989 + return NULL;
1.2990 + }
1.2991 +
1.2992 + // Eliminate volatile MemBars for scalar replaced objects.
1.2993 + if (can_reshape && req() == (Precedent+1)) {
1.2994 + bool eliminate = false;
1.2995 + int opc = Opcode();
1.2996 + if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) {
1.2997 + // Volatile field loads and stores.
1.2998 + Node* my_mem = in(MemBarNode::Precedent);
1.2999 + // The MembarAquire may keep an unused LoadNode alive through the Precedent edge
1.3000 + if ((my_mem != NULL) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) {
1.3001 + // if the Precedent is a decodeN and its input (a Load) is used at more than one place,
1.3002 + // replace this Precedent (decodeN) with the Load instead.
1.3003 + if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1)) {
1.3004 + Node* load_node = my_mem->in(1);
1.3005 + set_req(MemBarNode::Precedent, load_node);
1.3006 + phase->is_IterGVN()->_worklist.push(my_mem);
1.3007 + my_mem = load_node;
1.3008 + } else {
1.3009 + assert(my_mem->unique_out() == this, "sanity");
1.3010 + del_req(Precedent);
1.3011 + phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later
1.3012 + my_mem = NULL;
1.3013 + }
1.3014 + }
1.3015 + if (my_mem != NULL && my_mem->is_Mem()) {
1.3016 + const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr();
1.3017 + // Check for scalar replaced object reference.
1.3018 + if( t_oop != NULL && t_oop->is_known_instance_field() &&
1.3019 + t_oop->offset() != Type::OffsetBot &&
1.3020 + t_oop->offset() != Type::OffsetTop) {
1.3021 + eliminate = true;
1.3022 + }
1.3023 + }
1.3024 + } else if (opc == Op_MemBarRelease) {
1.3025 + // Final field stores.
1.3026 + Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent), phase);
1.3027 + if ((alloc != NULL) && alloc->is_Allocate() &&
1.3028 + alloc->as_Allocate()->_is_non_escaping) {
1.3029 + // The allocated object does not escape.
1.3030 + eliminate = true;
1.3031 + }
1.3032 + }
1.3033 + if (eliminate) {
1.3034 + // Replace MemBar projections by its inputs.
1.3035 + PhaseIterGVN* igvn = phase->is_IterGVN();
1.3036 + igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory));
1.3037 + igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control));
1.3038 + // Must return either the original node (now dead) or a new node
1.3039 + // (Do not return a top here, since that would break the uniqueness of top.)
1.3040 + return new (phase->C) ConINode(TypeInt::ZERO);
1.3041 + }
1.3042 + }
1.3043 + return NULL;
1.3044 +}
1.3045 +
1.3046 +//------------------------------Value------------------------------------------
1.3047 +const Type *MemBarNode::Value( PhaseTransform *phase ) const {
1.3048 + if( !in(0) ) return Type::TOP;
1.3049 + if( phase->type(in(0)) == Type::TOP )
1.3050 + return Type::TOP;
1.3051 + return TypeTuple::MEMBAR;
1.3052 +}
1.3053 +
1.3054 +//------------------------------match------------------------------------------
1.3055 +// Construct projections for memory.
1.3056 +Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) {
1.3057 + switch (proj->_con) {
1.3058 + case TypeFunc::Control:
1.3059 + case TypeFunc::Memory:
1.3060 + return new (m->C) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
1.3061 + }
1.3062 + ShouldNotReachHere();
1.3063 + return NULL;
1.3064 +}
1.3065 +
1.3066 +//===========================InitializeNode====================================
1.3067 +// SUMMARY:
1.3068 +// This node acts as a memory barrier on raw memory, after some raw stores.
1.3069 +// The 'cooked' oop value feeds from the Initialize, not the Allocation.
1.3070 +// The Initialize can 'capture' suitably constrained stores as raw inits.
1.3071 +// It can coalesce related raw stores into larger units (called 'tiles').
1.3072 +// It can avoid zeroing new storage for memory units which have raw inits.
1.3073 +// At macro-expansion, it is marked 'complete', and does not optimize further.
1.3074 +//
1.3075 +// EXAMPLE:
1.3076 +// The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
1.3077 +// ctl = incoming control; mem* = incoming memory
1.3078 +// (Note: A star * on a memory edge denotes I/O and other standard edges.)
1.3079 +// First allocate uninitialized memory and fill in the header:
1.3080 +// alloc = (Allocate ctl mem* 16 #short[].klass ...)
1.3081 +// ctl := alloc.Control; mem* := alloc.Memory*
1.3082 +// rawmem = alloc.Memory; rawoop = alloc.RawAddress
1.3083 +// Then initialize to zero the non-header parts of the raw memory block:
1.3084 +// init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
1.3085 +// ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
1.3086 +// After the initialize node executes, the object is ready for service:
1.3087 +// oop := (CheckCastPP init.Control alloc.RawAddress #short[])
1.3088 +// Suppose its body is immediately initialized as {1,2}:
1.3089 +// store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
1.3090 +// store2 = (StoreC init.Control store1 (+ oop 14) 2)
1.3091 +// mem.SLICE(#short[*]) := store2
1.3092 +//
1.3093 +// DETAILS:
1.3094 +// An InitializeNode collects and isolates object initialization after
1.3095 +// an AllocateNode and before the next possible safepoint. As a
1.3096 +// memory barrier (MemBarNode), it keeps critical stores from drifting
1.3097 +// down past any safepoint or any publication of the allocation.
1.3098 +// Before this barrier, a newly-allocated object may have uninitialized bits.
1.3099 +// After this barrier, it may be treated as a real oop, and GC is allowed.
1.3100 +//
1.3101 +// The semantics of the InitializeNode include an implicit zeroing of
1.3102 +// the new object from object header to the end of the object.
1.3103 +// (The object header and end are determined by the AllocateNode.)
1.3104 +//
1.3105 +// Certain stores may be added as direct inputs to the InitializeNode.
1.3106 +// These stores must update raw memory, and they must be to addresses
1.3107 +// derived from the raw address produced by AllocateNode, and with
1.3108 +// a constant offset. They must be ordered by increasing offset.
1.3109 +// The first one is at in(RawStores), the last at in(req()-1).
1.3110 +// Unlike most memory operations, they are not linked in a chain,
1.3111 +// but are displayed in parallel as users of the rawmem output of
1.3112 +// the allocation.
1.3113 +//
1.3114 +// (See comments in InitializeNode::capture_store, which continue
1.3115 +// the example given above.)
1.3116 +//
1.3117 +// When the associated Allocate is macro-expanded, the InitializeNode
1.3118 +// may be rewritten to optimize collected stores. A ClearArrayNode
1.3119 +// may also be created at that point to represent any required zeroing.
1.3120 +// The InitializeNode is then marked 'complete', prohibiting further
1.3121 +// capturing of nearby memory operations.
1.3122 +//
1.3123 +// During macro-expansion, all captured initializations which store
1.3124 +// constant values of 32 bits or smaller are coalesced (if advantageous)
1.3125 +// into larger 'tiles' 32 or 64 bits. This allows an object to be
1.3126 +// initialized in fewer memory operations. Memory words which are
1.3127 +// covered by neither tiles nor non-constant stores are pre-zeroed
1.3128 +// by explicit stores of zero. (The code shape happens to do all
1.3129 +// zeroing first, then all other stores, with both sequences occurring
1.3130 +// in order of ascending offsets.)
1.3131 +//
1.3132 +// Alternatively, code may be inserted between an AllocateNode and its
1.3133 +// InitializeNode, to perform arbitrary initialization of the new object.
1.3134 +// E.g., the object copying intrinsics insert complex data transfers here.
1.3135 +// The initialization must then be marked as 'complete' disable the
1.3136 +// built-in zeroing semantics and the collection of initializing stores.
1.3137 +//
1.3138 +// While an InitializeNode is incomplete, reads from the memory state
1.3139 +// produced by it are optimizable if they match the control edge and
1.3140 +// new oop address associated with the allocation/initialization.
1.3141 +// They return a stored value (if the offset matches) or else zero.
1.3142 +// A write to the memory state, if it matches control and address,
1.3143 +// and if it is to a constant offset, may be 'captured' by the
1.3144 +// InitializeNode. It is cloned as a raw memory operation and rewired
1.3145 +// inside the initialization, to the raw oop produced by the allocation.
1.3146 +// Operations on addresses which are provably distinct (e.g., to
1.3147 +// other AllocateNodes) are allowed to bypass the initialization.
1.3148 +//
1.3149 +// The effect of all this is to consolidate object initialization
1.3150 +// (both arrays and non-arrays, both piecewise and bulk) into a
1.3151 +// single location, where it can be optimized as a unit.
1.3152 +//
1.3153 +// Only stores with an offset less than TrackedInitializationLimit words
1.3154 +// will be considered for capture by an InitializeNode. This puts a
1.3155 +// reasonable limit on the complexity of optimized initializations.
1.3156 +
1.3157 +//---------------------------InitializeNode------------------------------------
1.3158 +InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
1.3159 + : _is_complete(Incomplete), _does_not_escape(false),
1.3160 + MemBarNode(C, adr_type, rawoop)
1.3161 +{
1.3162 + init_class_id(Class_Initialize);
1.3163 +
1.3164 + assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
1.3165 + assert(in(RawAddress) == rawoop, "proper init");
1.3166 + // Note: allocation() can be NULL, for secondary initialization barriers
1.3167 +}
1.3168 +
1.3169 +// Since this node is not matched, it will be processed by the
1.3170 +// register allocator. Declare that there are no constraints
1.3171 +// on the allocation of the RawAddress edge.
1.3172 +const RegMask &InitializeNode::in_RegMask(uint idx) const {
1.3173 + // This edge should be set to top, by the set_complete. But be conservative.
1.3174 + if (idx == InitializeNode::RawAddress)
1.3175 + return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
1.3176 + return RegMask::Empty;
1.3177 +}
1.3178 +
1.3179 +Node* InitializeNode::memory(uint alias_idx) {
1.3180 + Node* mem = in(Memory);
1.3181 + if (mem->is_MergeMem()) {
1.3182 + return mem->as_MergeMem()->memory_at(alias_idx);
1.3183 + } else {
1.3184 + // incoming raw memory is not split
1.3185 + return mem;
1.3186 + }
1.3187 +}
1.3188 +
1.3189 +bool InitializeNode::is_non_zero() {
1.3190 + if (is_complete()) return false;
1.3191 + remove_extra_zeroes();
1.3192 + return (req() > RawStores);
1.3193 +}
1.3194 +
1.3195 +void InitializeNode::set_complete(PhaseGVN* phase) {
1.3196 + assert(!is_complete(), "caller responsibility");
1.3197 + _is_complete = Complete;
1.3198 +
1.3199 + // After this node is complete, it contains a bunch of
1.3200 + // raw-memory initializations. There is no need for
1.3201 + // it to have anything to do with non-raw memory effects.
1.3202 + // Therefore, tell all non-raw users to re-optimize themselves,
1.3203 + // after skipping the memory effects of this initialization.
1.3204 + PhaseIterGVN* igvn = phase->is_IterGVN();
1.3205 + if (igvn) igvn->add_users_to_worklist(this);
1.3206 +}
1.3207 +
1.3208 +// convenience function
1.3209 +// return false if the init contains any stores already
1.3210 +bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
1.3211 + InitializeNode* init = initialization();
1.3212 + if (init == NULL || init->is_complete()) return false;
1.3213 + init->remove_extra_zeroes();
1.3214 + // for now, if this allocation has already collected any inits, bail:
1.3215 + if (init->is_non_zero()) return false;
1.3216 + init->set_complete(phase);
1.3217 + return true;
1.3218 +}
1.3219 +
1.3220 +void InitializeNode::remove_extra_zeroes() {
1.3221 + if (req() == RawStores) return;
1.3222 + Node* zmem = zero_memory();
1.3223 + uint fill = RawStores;
1.3224 + for (uint i = fill; i < req(); i++) {
1.3225 + Node* n = in(i);
1.3226 + if (n->is_top() || n == zmem) continue; // skip
1.3227 + if (fill < i) set_req(fill, n); // compact
1.3228 + ++fill;
1.3229 + }
1.3230 + // delete any empty spaces created:
1.3231 + while (fill < req()) {
1.3232 + del_req(fill);
1.3233 + }
1.3234 +}
1.3235 +
1.3236 +// Helper for remembering which stores go with which offsets.
1.3237 +intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) {
1.3238 + if (!st->is_Store()) return -1; // can happen to dead code via subsume_node
1.3239 + intptr_t offset = -1;
1.3240 + Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
1.3241 + phase, offset);
1.3242 + if (base == NULL) return -1; // something is dead,
1.3243 + if (offset < 0) return -1; // dead, dead
1.3244 + return offset;
1.3245 +}
1.3246 +
1.3247 +// Helper for proving that an initialization expression is
1.3248 +// "simple enough" to be folded into an object initialization.
1.3249 +// Attempts to prove that a store's initial value 'n' can be captured
1.3250 +// within the initialization without creating a vicious cycle, such as:
1.3251 +// { Foo p = new Foo(); p.next = p; }
1.3252 +// True for constants and parameters and small combinations thereof.
1.3253 +bool InitializeNode::detect_init_independence(Node* n, int& count) {
1.3254 + if (n == NULL) return true; // (can this really happen?)
1.3255 + if (n->is_Proj()) n = n->in(0);
1.3256 + if (n == this) return false; // found a cycle
1.3257 + if (n->is_Con()) return true;
1.3258 + if (n->is_Start()) return true; // params, etc., are OK
1.3259 + if (n->is_Root()) return true; // even better
1.3260 +
1.3261 + Node* ctl = n->in(0);
1.3262 + if (ctl != NULL && !ctl->is_top()) {
1.3263 + if (ctl->is_Proj()) ctl = ctl->in(0);
1.3264 + if (ctl == this) return false;
1.3265 +
1.3266 + // If we already know that the enclosing memory op is pinned right after
1.3267 + // the init, then any control flow that the store has picked up
1.3268 + // must have preceded the init, or else be equal to the init.
1.3269 + // Even after loop optimizations (which might change control edges)
1.3270 + // a store is never pinned *before* the availability of its inputs.
1.3271 + if (!MemNode::all_controls_dominate(n, this))
1.3272 + return false; // failed to prove a good control
1.3273 + }
1.3274 +
1.3275 + // Check data edges for possible dependencies on 'this'.
1.3276 + if ((count += 1) > 20) return false; // complexity limit
1.3277 + for (uint i = 1; i < n->req(); i++) {
1.3278 + Node* m = n->in(i);
1.3279 + if (m == NULL || m == n || m->is_top()) continue;
1.3280 + uint first_i = n->find_edge(m);
1.3281 + if (i != first_i) continue; // process duplicate edge just once
1.3282 + if (!detect_init_independence(m, count)) {
1.3283 + return false;
1.3284 + }
1.3285 + }
1.3286 +
1.3287 + return true;
1.3288 +}
1.3289 +
1.3290 +// Here are all the checks a Store must pass before it can be moved into
1.3291 +// an initialization. Returns zero if a check fails.
1.3292 +// On success, returns the (constant) offset to which the store applies,
1.3293 +// within the initialized memory.
1.3294 +intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase, bool can_reshape) {
1.3295 + const int FAIL = 0;
1.3296 + if (st->req() != MemNode::ValueIn + 1)
1.3297 + return FAIL; // an inscrutable StoreNode (card mark?)
1.3298 + Node* ctl = st->in(MemNode::Control);
1.3299 + if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this))
1.3300 + return FAIL; // must be unconditional after the initialization
1.3301 + Node* mem = st->in(MemNode::Memory);
1.3302 + if (!(mem->is_Proj() && mem->in(0) == this))
1.3303 + return FAIL; // must not be preceded by other stores
1.3304 + Node* adr = st->in(MemNode::Address);
1.3305 + intptr_t offset;
1.3306 + AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
1.3307 + if (alloc == NULL)
1.3308 + return FAIL; // inscrutable address
1.3309 + if (alloc != allocation())
1.3310 + return FAIL; // wrong allocation! (store needs to float up)
1.3311 + Node* val = st->in(MemNode::ValueIn);
1.3312 + int complexity_count = 0;
1.3313 + if (!detect_init_independence(val, complexity_count))
1.3314 + return FAIL; // stored value must be 'simple enough'
1.3315 +
1.3316 + // The Store can be captured only if nothing after the allocation
1.3317 + // and before the Store is using the memory location that the store
1.3318 + // overwrites.
1.3319 + bool failed = false;
1.3320 + // If is_complete_with_arraycopy() is true the shape of the graph is
1.3321 + // well defined and is safe so no need for extra checks.
1.3322 + if (!is_complete_with_arraycopy()) {
1.3323 + // We are going to look at each use of the memory state following
1.3324 + // the allocation to make sure nothing reads the memory that the
1.3325 + // Store writes.
1.3326 + const TypePtr* t_adr = phase->type(adr)->isa_ptr();
1.3327 + int alias_idx = phase->C->get_alias_index(t_adr);
1.3328 + ResourceMark rm;
1.3329 + Unique_Node_List mems;
1.3330 + mems.push(mem);
1.3331 + Node* unique_merge = NULL;
1.3332 + for (uint next = 0; next < mems.size(); ++next) {
1.3333 + Node *m = mems.at(next);
1.3334 + for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
1.3335 + Node *n = m->fast_out(j);
1.3336 + if (n->outcnt() == 0) {
1.3337 + continue;
1.3338 + }
1.3339 + if (n == st) {
1.3340 + continue;
1.3341 + } else if (n->in(0) != NULL && n->in(0) != ctl) {
1.3342 + // If the control of this use is different from the control
1.3343 + // of the Store which is right after the InitializeNode then
1.3344 + // this node cannot be between the InitializeNode and the
1.3345 + // Store.
1.3346 + continue;
1.3347 + } else if (n->is_MergeMem()) {
1.3348 + if (n->as_MergeMem()->memory_at(alias_idx) == m) {
1.3349 + // We can hit a MergeMemNode (that will likely go away
1.3350 + // later) that is a direct use of the memory state
1.3351 + // following the InitializeNode on the same slice as the
1.3352 + // store node that we'd like to capture. We need to check
1.3353 + // the uses of the MergeMemNode.
1.3354 + mems.push(n);
1.3355 + }
1.3356 + } else if (n->is_Mem()) {
1.3357 + Node* other_adr = n->in(MemNode::Address);
1.3358 + if (other_adr == adr) {
1.3359 + failed = true;
1.3360 + break;
1.3361 + } else {
1.3362 + const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr();
1.3363 + if (other_t_adr != NULL) {
1.3364 + int other_alias_idx = phase->C->get_alias_index(other_t_adr);
1.3365 + if (other_alias_idx == alias_idx) {
1.3366 + // A load from the same memory slice as the store right
1.3367 + // after the InitializeNode. We check the control of the
1.3368 + // object/array that is loaded from. If it's the same as
1.3369 + // the store control then we cannot capture the store.
1.3370 + assert(!n->is_Store(), "2 stores to same slice on same control?");
1.3371 + Node* base = other_adr;
1.3372 + assert(base->is_AddP(), err_msg_res("should be addp but is %s", base->Name()));
1.3373 + base = base->in(AddPNode::Base);
1.3374 + if (base != NULL) {
1.3375 + base = base->uncast();
1.3376 + if (base->is_Proj() && base->in(0) == alloc) {
1.3377 + failed = true;
1.3378 + break;
1.3379 + }
1.3380 + }
1.3381 + }
1.3382 + }
1.3383 + }
1.3384 + } else {
1.3385 + failed = true;
1.3386 + break;
1.3387 + }
1.3388 + }
1.3389 + }
1.3390 + }
1.3391 + if (failed) {
1.3392 + if (!can_reshape) {
1.3393 + // We decided we couldn't capture the store during parsing. We
1.3394 + // should try again during the next IGVN once the graph is
1.3395 + // cleaner.
1.3396 + phase->C->record_for_igvn(st);
1.3397 + }
1.3398 + return FAIL;
1.3399 + }
1.3400 +
1.3401 + return offset; // success
1.3402 +}
1.3403 +
1.3404 +// Find the captured store in(i) which corresponds to the range
1.3405 +// [start..start+size) in the initialized object.
1.3406 +// If there is one, return its index i. If there isn't, return the
1.3407 +// negative of the index where it should be inserted.
1.3408 +// Return 0 if the queried range overlaps an initialization boundary
1.3409 +// or if dead code is encountered.
1.3410 +// If size_in_bytes is zero, do not bother with overlap checks.
1.3411 +int InitializeNode::captured_store_insertion_point(intptr_t start,
1.3412 + int size_in_bytes,
1.3413 + PhaseTransform* phase) {
1.3414 + const int FAIL = 0, MAX_STORE = BytesPerLong;
1.3415 +
1.3416 + if (is_complete())
1.3417 + return FAIL; // arraycopy got here first; punt
1.3418 +
1.3419 + assert(allocation() != NULL, "must be present");
1.3420 +
1.3421 + // no negatives, no header fields:
1.3422 + if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL;
1.3423 +
1.3424 + // after a certain size, we bail out on tracking all the stores:
1.3425 + intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
1.3426 + if (start >= ti_limit) return FAIL;
1.3427 +
1.3428 + for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
1.3429 + if (i >= limit) return -(int)i; // not found; here is where to put it
1.3430 +
1.3431 + Node* st = in(i);
1.3432 + intptr_t st_off = get_store_offset(st, phase);
1.3433 + if (st_off < 0) {
1.3434 + if (st != zero_memory()) {
1.3435 + return FAIL; // bail out if there is dead garbage
1.3436 + }
1.3437 + } else if (st_off > start) {
1.3438 + // ...we are done, since stores are ordered
1.3439 + if (st_off < start + size_in_bytes) {
1.3440 + return FAIL; // the next store overlaps
1.3441 + }
1.3442 + return -(int)i; // not found; here is where to put it
1.3443 + } else if (st_off < start) {
1.3444 + if (size_in_bytes != 0 &&
1.3445 + start < st_off + MAX_STORE &&
1.3446 + start < st_off + st->as_Store()->memory_size()) {
1.3447 + return FAIL; // the previous store overlaps
1.3448 + }
1.3449 + } else {
1.3450 + if (size_in_bytes != 0 &&
1.3451 + st->as_Store()->memory_size() != size_in_bytes) {
1.3452 + return FAIL; // mismatched store size
1.3453 + }
1.3454 + return i;
1.3455 + }
1.3456 +
1.3457 + ++i;
1.3458 + }
1.3459 +}
1.3460 +
1.3461 +// Look for a captured store which initializes at the offset 'start'
1.3462 +// with the given size. If there is no such store, and no other
1.3463 +// initialization interferes, then return zero_memory (the memory
1.3464 +// projection of the AllocateNode).
1.3465 +Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
1.3466 + PhaseTransform* phase) {
1.3467 + assert(stores_are_sane(phase), "");
1.3468 + int i = captured_store_insertion_point(start, size_in_bytes, phase);
1.3469 + if (i == 0) {
1.3470 + return NULL; // something is dead
1.3471 + } else if (i < 0) {
1.3472 + return zero_memory(); // just primordial zero bits here
1.3473 + } else {
1.3474 + Node* st = in(i); // here is the store at this position
1.3475 + assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
1.3476 + return st;
1.3477 + }
1.3478 +}
1.3479 +
1.3480 +// Create, as a raw pointer, an address within my new object at 'offset'.
1.3481 +Node* InitializeNode::make_raw_address(intptr_t offset,
1.3482 + PhaseTransform* phase) {
1.3483 + Node* addr = in(RawAddress);
1.3484 + if (offset != 0) {
1.3485 + Compile* C = phase->C;
1.3486 + addr = phase->transform( new (C) AddPNode(C->top(), addr,
1.3487 + phase->MakeConX(offset)) );
1.3488 + }
1.3489 + return addr;
1.3490 +}
1.3491 +
1.3492 +// Clone the given store, converting it into a raw store
1.3493 +// initializing a field or element of my new object.
1.3494 +// Caller is responsible for retiring the original store,
1.3495 +// with subsume_node or the like.
1.3496 +//
1.3497 +// From the example above InitializeNode::InitializeNode,
1.3498 +// here are the old stores to be captured:
1.3499 +// store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
1.3500 +// store2 = (StoreC init.Control store1 (+ oop 14) 2)
1.3501 +//
1.3502 +// Here is the changed code; note the extra edges on init:
1.3503 +// alloc = (Allocate ...)
1.3504 +// rawoop = alloc.RawAddress
1.3505 +// rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
1.3506 +// rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
1.3507 +// init = (Initialize alloc.Control alloc.Memory rawoop
1.3508 +// rawstore1 rawstore2)
1.3509 +//
1.3510 +Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
1.3511 + PhaseTransform* phase, bool can_reshape) {
1.3512 + assert(stores_are_sane(phase), "");
1.3513 +
1.3514 + if (start < 0) return NULL;
1.3515 + assert(can_capture_store(st, phase, can_reshape) == start, "sanity");
1.3516 +
1.3517 + Compile* C = phase->C;
1.3518 + int size_in_bytes = st->memory_size();
1.3519 + int i = captured_store_insertion_point(start, size_in_bytes, phase);
1.3520 + if (i == 0) return NULL; // bail out
1.3521 + Node* prev_mem = NULL; // raw memory for the captured store
1.3522 + if (i > 0) {
1.3523 + prev_mem = in(i); // there is a pre-existing store under this one
1.3524 + set_req(i, C->top()); // temporarily disconnect it
1.3525 + // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
1.3526 + } else {
1.3527 + i = -i; // no pre-existing store
1.3528 + prev_mem = zero_memory(); // a slice of the newly allocated object
1.3529 + if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
1.3530 + set_req(--i, C->top()); // reuse this edge; it has been folded away
1.3531 + else
1.3532 + ins_req(i, C->top()); // build a new edge
1.3533 + }
1.3534 + Node* new_st = st->clone();
1.3535 + new_st->set_req(MemNode::Control, in(Control));
1.3536 + new_st->set_req(MemNode::Memory, prev_mem);
1.3537 + new_st->set_req(MemNode::Address, make_raw_address(start, phase));
1.3538 + new_st = phase->transform(new_st);
1.3539 +
1.3540 + // At this point, new_st might have swallowed a pre-existing store
1.3541 + // at the same offset, or perhaps new_st might have disappeared,
1.3542 + // if it redundantly stored the same value (or zero to fresh memory).
1.3543 +
1.3544 + // In any case, wire it in:
1.3545 + set_req(i, new_st);
1.3546 +
1.3547 + // The caller may now kill the old guy.
1.3548 + DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
1.3549 + assert(check_st == new_st || check_st == NULL, "must be findable");
1.3550 + assert(!is_complete(), "");
1.3551 + return new_st;
1.3552 +}
1.3553 +
1.3554 +static bool store_constant(jlong* tiles, int num_tiles,
1.3555 + intptr_t st_off, int st_size,
1.3556 + jlong con) {
1.3557 + if ((st_off & (st_size-1)) != 0)
1.3558 + return false; // strange store offset (assume size==2**N)
1.3559 + address addr = (address)tiles + st_off;
1.3560 + assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
1.3561 + switch (st_size) {
1.3562 + case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break;
1.3563 + case sizeof(jchar): *(jchar*) addr = (jchar) con; break;
1.3564 + case sizeof(jint): *(jint*) addr = (jint) con; break;
1.3565 + case sizeof(jlong): *(jlong*) addr = (jlong) con; break;
1.3566 + default: return false; // strange store size (detect size!=2**N here)
1.3567 + }
1.3568 + return true; // return success to caller
1.3569 +}
1.3570 +
1.3571 +// Coalesce subword constants into int constants and possibly
1.3572 +// into long constants. The goal, if the CPU permits,
1.3573 +// is to initialize the object with a small number of 64-bit tiles.
1.3574 +// Also, convert floating-point constants to bit patterns.
1.3575 +// Non-constants are not relevant to this pass.
1.3576 +//
1.3577 +// In terms of the running example on InitializeNode::InitializeNode
1.3578 +// and InitializeNode::capture_store, here is the transformation
1.3579 +// of rawstore1 and rawstore2 into rawstore12:
1.3580 +// alloc = (Allocate ...)
1.3581 +// rawoop = alloc.RawAddress
1.3582 +// tile12 = 0x00010002
1.3583 +// rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
1.3584 +// init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
1.3585 +//
1.3586 +void
1.3587 +InitializeNode::coalesce_subword_stores(intptr_t header_size,
1.3588 + Node* size_in_bytes,
1.3589 + PhaseGVN* phase) {
1.3590 + Compile* C = phase->C;
1.3591 +
1.3592 + assert(stores_are_sane(phase), "");
1.3593 + // Note: After this pass, they are not completely sane,
1.3594 + // since there may be some overlaps.
1.3595 +
1.3596 + int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
1.3597 +
1.3598 + intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
1.3599 + intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
1.3600 + size_limit = MIN2(size_limit, ti_limit);
1.3601 + size_limit = align_size_up(size_limit, BytesPerLong);
1.3602 + int num_tiles = size_limit / BytesPerLong;
1.3603 +
1.3604 + // allocate space for the tile map:
1.3605 + const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
1.3606 + jlong tiles_buf[small_len];
1.3607 + Node* nodes_buf[small_len];
1.3608 + jlong inits_buf[small_len];
1.3609 + jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
1.3610 + : NEW_RESOURCE_ARRAY(jlong, num_tiles));
1.3611 + Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
1.3612 + : NEW_RESOURCE_ARRAY(Node*, num_tiles));
1.3613 + jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
1.3614 + : NEW_RESOURCE_ARRAY(jlong, num_tiles));
1.3615 + // tiles: exact bitwise model of all primitive constants
1.3616 + // nodes: last constant-storing node subsumed into the tiles model
1.3617 + // inits: which bytes (in each tile) are touched by any initializations
1.3618 +
1.3619 + //// Pass A: Fill in the tile model with any relevant stores.
1.3620 +
1.3621 + Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
1.3622 + Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
1.3623 + Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
1.3624 + Node* zmem = zero_memory(); // initially zero memory state
1.3625 + for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
1.3626 + Node* st = in(i);
1.3627 + intptr_t st_off = get_store_offset(st, phase);
1.3628 +
1.3629 + // Figure out the store's offset and constant value:
1.3630 + if (st_off < header_size) continue; //skip (ignore header)
1.3631 + if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain)
1.3632 + int st_size = st->as_Store()->memory_size();
1.3633 + if (st_off + st_size > size_limit) break;
1.3634 +
1.3635 + // Record which bytes are touched, whether by constant or not.
1.3636 + if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
1.3637 + continue; // skip (strange store size)
1.3638 +
1.3639 + const Type* val = phase->type(st->in(MemNode::ValueIn));
1.3640 + if (!val->singleton()) continue; //skip (non-con store)
1.3641 + BasicType type = val->basic_type();
1.3642 +
1.3643 + jlong con = 0;
1.3644 + switch (type) {
1.3645 + case T_INT: con = val->is_int()->get_con(); break;
1.3646 + case T_LONG: con = val->is_long()->get_con(); break;
1.3647 + case T_FLOAT: con = jint_cast(val->getf()); break;
1.3648 + case T_DOUBLE: con = jlong_cast(val->getd()); break;
1.3649 + default: continue; //skip (odd store type)
1.3650 + }
1.3651 +
1.3652 + if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
1.3653 + st->Opcode() == Op_StoreL) {
1.3654 + continue; // This StoreL is already optimal.
1.3655 + }
1.3656 +
1.3657 + // Store down the constant.
1.3658 + store_constant(tiles, num_tiles, st_off, st_size, con);
1.3659 +
1.3660 + intptr_t j = st_off >> LogBytesPerLong;
1.3661 +
1.3662 + if (type == T_INT && st_size == BytesPerInt
1.3663 + && (st_off & BytesPerInt) == BytesPerInt) {
1.3664 + jlong lcon = tiles[j];
1.3665 + if (!Matcher::isSimpleConstant64(lcon) &&
1.3666 + st->Opcode() == Op_StoreI) {
1.3667 + // This StoreI is already optimal by itself.
1.3668 + jint* intcon = (jint*) &tiles[j];
1.3669 + intcon[1] = 0; // undo the store_constant()
1.3670 +
1.3671 + // If the previous store is also optimal by itself, back up and
1.3672 + // undo the action of the previous loop iteration... if we can.
1.3673 + // But if we can't, just let the previous half take care of itself.
1.3674 + st = nodes[j];
1.3675 + st_off -= BytesPerInt;
1.3676 + con = intcon[0];
1.3677 + if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) {
1.3678 + assert(st_off >= header_size, "still ignoring header");
1.3679 + assert(get_store_offset(st, phase) == st_off, "must be");
1.3680 + assert(in(i-1) == zmem, "must be");
1.3681 + DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
1.3682 + assert(con == tcon->is_int()->get_con(), "must be");
1.3683 + // Undo the effects of the previous loop trip, which swallowed st:
1.3684 + intcon[0] = 0; // undo store_constant()
1.3685 + set_req(i-1, st); // undo set_req(i, zmem)
1.3686 + nodes[j] = NULL; // undo nodes[j] = st
1.3687 + --old_subword; // undo ++old_subword
1.3688 + }
1.3689 + continue; // This StoreI is already optimal.
1.3690 + }
1.3691 + }
1.3692 +
1.3693 + // This store is not needed.
1.3694 + set_req(i, zmem);
1.3695 + nodes[j] = st; // record for the moment
1.3696 + if (st_size < BytesPerLong) // something has changed
1.3697 + ++old_subword; // includes int/float, but who's counting...
1.3698 + else ++old_long;
1.3699 + }
1.3700 +
1.3701 + if ((old_subword + old_long) == 0)
1.3702 + return; // nothing more to do
1.3703 +
1.3704 + //// Pass B: Convert any non-zero tiles into optimal constant stores.
1.3705 + // Be sure to insert them before overlapping non-constant stores.
1.3706 + // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.)
1.3707 + for (int j = 0; j < num_tiles; j++) {
1.3708 + jlong con = tiles[j];
1.3709 + jlong init = inits[j];
1.3710 + if (con == 0) continue;
1.3711 + jint con0, con1; // split the constant, address-wise
1.3712 + jint init0, init1; // split the init map, address-wise
1.3713 + { union { jlong con; jint intcon[2]; } u;
1.3714 + u.con = con;
1.3715 + con0 = u.intcon[0];
1.3716 + con1 = u.intcon[1];
1.3717 + u.con = init;
1.3718 + init0 = u.intcon[0];
1.3719 + init1 = u.intcon[1];
1.3720 + }
1.3721 +
1.3722 + Node* old = nodes[j];
1.3723 + assert(old != NULL, "need the prior store");
1.3724 + intptr_t offset = (j * BytesPerLong);
1.3725 +
1.3726 + bool split = !Matcher::isSimpleConstant64(con);
1.3727 +
1.3728 + if (offset < header_size) {
1.3729 + assert(offset + BytesPerInt >= header_size, "second int counts");
1.3730 + assert(*(jint*)&tiles[j] == 0, "junk in header");
1.3731 + split = true; // only the second word counts
1.3732 + // Example: int a[] = { 42 ... }
1.3733 + } else if (con0 == 0 && init0 == -1) {
1.3734 + split = true; // first word is covered by full inits
1.3735 + // Example: int a[] = { ... foo(), 42 ... }
1.3736 + } else if (con1 == 0 && init1 == -1) {
1.3737 + split = true; // second word is covered by full inits
1.3738 + // Example: int a[] = { ... 42, foo() ... }
1.3739 + }
1.3740 +
1.3741 + // Here's a case where init0 is neither 0 nor -1:
1.3742 + // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... }
1.3743 + // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
1.3744 + // In this case the tile is not split; it is (jlong)42.
1.3745 + // The big tile is stored down, and then the foo() value is inserted.
1.3746 + // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
1.3747 +
1.3748 + Node* ctl = old->in(MemNode::Control);
1.3749 + Node* adr = make_raw_address(offset, phase);
1.3750 + const TypePtr* atp = TypeRawPtr::BOTTOM;
1.3751 +
1.3752 + // One or two coalesced stores to plop down.
1.3753 + Node* st[2];
1.3754 + intptr_t off[2];
1.3755 + int nst = 0;
1.3756 + if (!split) {
1.3757 + ++new_long;
1.3758 + off[nst] = offset;
1.3759 + st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
1.3760 + phase->longcon(con), T_LONG, MemNode::unordered);
1.3761 + } else {
1.3762 + // Omit either if it is a zero.
1.3763 + if (con0 != 0) {
1.3764 + ++new_int;
1.3765 + off[nst] = offset;
1.3766 + st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
1.3767 + phase->intcon(con0), T_INT, MemNode::unordered);
1.3768 + }
1.3769 + if (con1 != 0) {
1.3770 + ++new_int;
1.3771 + offset += BytesPerInt;
1.3772 + adr = make_raw_address(offset, phase);
1.3773 + off[nst] = offset;
1.3774 + st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
1.3775 + phase->intcon(con1), T_INT, MemNode::unordered);
1.3776 + }
1.3777 + }
1.3778 +
1.3779 + // Insert second store first, then the first before the second.
1.3780 + // Insert each one just before any overlapping non-constant stores.
1.3781 + while (nst > 0) {
1.3782 + Node* st1 = st[--nst];
1.3783 + C->copy_node_notes_to(st1, old);
1.3784 + st1 = phase->transform(st1);
1.3785 + offset = off[nst];
1.3786 + assert(offset >= header_size, "do not smash header");
1.3787 + int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
1.3788 + guarantee(ins_idx != 0, "must re-insert constant store");
1.3789 + if (ins_idx < 0) ins_idx = -ins_idx; // never overlap
1.3790 + if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
1.3791 + set_req(--ins_idx, st1);
1.3792 + else
1.3793 + ins_req(ins_idx, st1);
1.3794 + }
1.3795 + }
1.3796 +
1.3797 + if (PrintCompilation && WizardMode)
1.3798 + tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
1.3799 + old_subword, old_long, new_int, new_long);
1.3800 + if (C->log() != NULL)
1.3801 + C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
1.3802 + old_subword, old_long, new_int, new_long);
1.3803 +
1.3804 + // Clean up any remaining occurrences of zmem:
1.3805 + remove_extra_zeroes();
1.3806 +}
1.3807 +
1.3808 +// Explore forward from in(start) to find the first fully initialized
1.3809 +// word, and return its offset. Skip groups of subword stores which
1.3810 +// together initialize full words. If in(start) is itself part of a
1.3811 +// fully initialized word, return the offset of in(start). If there
1.3812 +// are no following full-word stores, or if something is fishy, return
1.3813 +// a negative value.
1.3814 +intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
1.3815 + int int_map = 0;
1.3816 + intptr_t int_map_off = 0;
1.3817 + const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for
1.3818 +
1.3819 + for (uint i = start, limit = req(); i < limit; i++) {
1.3820 + Node* st = in(i);
1.3821 +
1.3822 + intptr_t st_off = get_store_offset(st, phase);
1.3823 + if (st_off < 0) break; // return conservative answer
1.3824 +
1.3825 + int st_size = st->as_Store()->memory_size();
1.3826 + if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
1.3827 + return st_off; // we found a complete word init
1.3828 + }
1.3829 +
1.3830 + // update the map:
1.3831 +
1.3832 + intptr_t this_int_off = align_size_down(st_off, BytesPerInt);
1.3833 + if (this_int_off != int_map_off) {
1.3834 + // reset the map:
1.3835 + int_map = 0;
1.3836 + int_map_off = this_int_off;
1.3837 + }
1.3838 +
1.3839 + int subword_off = st_off - this_int_off;
1.3840 + int_map |= right_n_bits(st_size) << subword_off;
1.3841 + if ((int_map & FULL_MAP) == FULL_MAP) {
1.3842 + return this_int_off; // we found a complete word init
1.3843 + }
1.3844 +
1.3845 + // Did this store hit or cross the word boundary?
1.3846 + intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt);
1.3847 + if (next_int_off == this_int_off + BytesPerInt) {
1.3848 + // We passed the current int, without fully initializing it.
1.3849 + int_map_off = next_int_off;
1.3850 + int_map >>= BytesPerInt;
1.3851 + } else if (next_int_off > this_int_off + BytesPerInt) {
1.3852 + // We passed the current and next int.
1.3853 + return this_int_off + BytesPerInt;
1.3854 + }
1.3855 + }
1.3856 +
1.3857 + return -1;
1.3858 +}
1.3859 +
1.3860 +
1.3861 +// Called when the associated AllocateNode is expanded into CFG.
1.3862 +// At this point, we may perform additional optimizations.
1.3863 +// Linearize the stores by ascending offset, to make memory
1.3864 +// activity as coherent as possible.
1.3865 +Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
1.3866 + intptr_t header_size,
1.3867 + Node* size_in_bytes,
1.3868 + PhaseGVN* phase) {
1.3869 + assert(!is_complete(), "not already complete");
1.3870 + assert(stores_are_sane(phase), "");
1.3871 + assert(allocation() != NULL, "must be present");
1.3872 +
1.3873 + remove_extra_zeroes();
1.3874 +
1.3875 + if (ReduceFieldZeroing || ReduceBulkZeroing)
1.3876 + // reduce instruction count for common initialization patterns
1.3877 + coalesce_subword_stores(header_size, size_in_bytes, phase);
1.3878 +
1.3879 + Node* zmem = zero_memory(); // initially zero memory state
1.3880 + Node* inits = zmem; // accumulating a linearized chain of inits
1.3881 + #ifdef ASSERT
1.3882 + intptr_t first_offset = allocation()->minimum_header_size();
1.3883 + intptr_t last_init_off = first_offset; // previous init offset
1.3884 + intptr_t last_init_end = first_offset; // previous init offset+size
1.3885 + intptr_t last_tile_end = first_offset; // previous tile offset+size
1.3886 + #endif
1.3887 + intptr_t zeroes_done = header_size;
1.3888 +
1.3889 + bool do_zeroing = true; // we might give up if inits are very sparse
1.3890 + int big_init_gaps = 0; // how many large gaps have we seen?
1.3891 +
1.3892 + if (ZeroTLAB) do_zeroing = false;
1.3893 + if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false;
1.3894 +
1.3895 + for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
1.3896 + Node* st = in(i);
1.3897 + intptr_t st_off = get_store_offset(st, phase);
1.3898 + if (st_off < 0)
1.3899 + break; // unknown junk in the inits
1.3900 + if (st->in(MemNode::Memory) != zmem)
1.3901 + break; // complicated store chains somehow in list
1.3902 +
1.3903 + int st_size = st->as_Store()->memory_size();
1.3904 + intptr_t next_init_off = st_off + st_size;
1.3905 +
1.3906 + if (do_zeroing && zeroes_done < next_init_off) {
1.3907 + // See if this store needs a zero before it or under it.
1.3908 + intptr_t zeroes_needed = st_off;
1.3909 +
1.3910 + if (st_size < BytesPerInt) {
1.3911 + // Look for subword stores which only partially initialize words.
1.3912 + // If we find some, we must lay down some word-level zeroes first,
1.3913 + // underneath the subword stores.
1.3914 + //
1.3915 + // Examples:
1.3916 + // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s
1.3917 + // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y
1.3918 + // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z
1.3919 + //
1.3920 + // Note: coalesce_subword_stores may have already done this,
1.3921 + // if it was prompted by constant non-zero subword initializers.
1.3922 + // But this case can still arise with non-constant stores.
1.3923 +
1.3924 + intptr_t next_full_store = find_next_fullword_store(i, phase);
1.3925 +
1.3926 + // In the examples above:
1.3927 + // in(i) p q r s x y z
1.3928 + // st_off 12 13 14 15 12 13 14
1.3929 + // st_size 1 1 1 1 1 1 1
1.3930 + // next_full_s. 12 16 16 16 16 16 16
1.3931 + // z's_done 12 16 16 16 12 16 12
1.3932 + // z's_needed 12 16 16 16 16 16 16
1.3933 + // zsize 0 0 0 0 4 0 4
1.3934 + if (next_full_store < 0) {
1.3935 + // Conservative tack: Zero to end of current word.
1.3936 + zeroes_needed = align_size_up(zeroes_needed, BytesPerInt);
1.3937 + } else {
1.3938 + // Zero to beginning of next fully initialized word.
1.3939 + // Or, don't zero at all, if we are already in that word.
1.3940 + assert(next_full_store >= zeroes_needed, "must go forward");
1.3941 + assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
1.3942 + zeroes_needed = next_full_store;
1.3943 + }
1.3944 + }
1.3945 +
1.3946 + if (zeroes_needed > zeroes_done) {
1.3947 + intptr_t zsize = zeroes_needed - zeroes_done;
1.3948 + // Do some incremental zeroing on rawmem, in parallel with inits.
1.3949 + zeroes_done = align_size_down(zeroes_done, BytesPerInt);
1.3950 + rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
1.3951 + zeroes_done, zeroes_needed,
1.3952 + phase);
1.3953 + zeroes_done = zeroes_needed;
1.3954 + if (zsize > Matcher::init_array_short_size && ++big_init_gaps > 2)
1.3955 + do_zeroing = false; // leave the hole, next time
1.3956 + }
1.3957 + }
1.3958 +
1.3959 + // Collect the store and move on:
1.3960 + st->set_req(MemNode::Memory, inits);
1.3961 + inits = st; // put it on the linearized chain
1.3962 + set_req(i, zmem); // unhook from previous position
1.3963 +
1.3964 + if (zeroes_done == st_off)
1.3965 + zeroes_done = next_init_off;
1.3966 +
1.3967 + assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
1.3968 +
1.3969 + #ifdef ASSERT
1.3970 + // Various order invariants. Weaker than stores_are_sane because
1.3971 + // a large constant tile can be filled in by smaller non-constant stores.
1.3972 + assert(st_off >= last_init_off, "inits do not reverse");
1.3973 + last_init_off = st_off;
1.3974 + const Type* val = NULL;
1.3975 + if (st_size >= BytesPerInt &&
1.3976 + (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
1.3977 + (int)val->basic_type() < (int)T_OBJECT) {
1.3978 + assert(st_off >= last_tile_end, "tiles do not overlap");
1.3979 + assert(st_off >= last_init_end, "tiles do not overwrite inits");
1.3980 + last_tile_end = MAX2(last_tile_end, next_init_off);
1.3981 + } else {
1.3982 + intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong);
1.3983 + assert(st_tile_end >= last_tile_end, "inits stay with tiles");
1.3984 + assert(st_off >= last_init_end, "inits do not overlap");
1.3985 + last_init_end = next_init_off; // it's a non-tile
1.3986 + }
1.3987 + #endif //ASSERT
1.3988 + }
1.3989 +
1.3990 + remove_extra_zeroes(); // clear out all the zmems left over
1.3991 + add_req(inits);
1.3992 +
1.3993 + if (!ZeroTLAB) {
1.3994 + // If anything remains to be zeroed, zero it all now.
1.3995 + zeroes_done = align_size_down(zeroes_done, BytesPerInt);
1.3996 + // if it is the last unused 4 bytes of an instance, forget about it
1.3997 + intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
1.3998 + if (zeroes_done + BytesPerLong >= size_limit) {
1.3999 + assert(allocation() != NULL, "");
1.4000 + if (allocation()->Opcode() == Op_Allocate) {
1.4001 + Node* klass_node = allocation()->in(AllocateNode::KlassNode);
1.4002 + ciKlass* k = phase->type(klass_node)->is_klassptr()->klass();
1.4003 + if (zeroes_done == k->layout_helper())
1.4004 + zeroes_done = size_limit;
1.4005 + }
1.4006 + }
1.4007 + if (zeroes_done < size_limit) {
1.4008 + rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
1.4009 + zeroes_done, size_in_bytes, phase);
1.4010 + }
1.4011 + }
1.4012 +
1.4013 + set_complete(phase);
1.4014 + return rawmem;
1.4015 +}
1.4016 +
1.4017 +
1.4018 +#ifdef ASSERT
1.4019 +bool InitializeNode::stores_are_sane(PhaseTransform* phase) {
1.4020 + if (is_complete())
1.4021 + return true; // stores could be anything at this point
1.4022 + assert(allocation() != NULL, "must be present");
1.4023 + intptr_t last_off = allocation()->minimum_header_size();
1.4024 + for (uint i = InitializeNode::RawStores; i < req(); i++) {
1.4025 + Node* st = in(i);
1.4026 + intptr_t st_off = get_store_offset(st, phase);
1.4027 + if (st_off < 0) continue; // ignore dead garbage
1.4028 + if (last_off > st_off) {
1.4029 + tty->print_cr("*** bad store offset at %d: " INTX_FORMAT " > " INTX_FORMAT, i, last_off, st_off);
1.4030 + this->dump(2);
1.4031 + assert(false, "ascending store offsets");
1.4032 + return false;
1.4033 + }
1.4034 + last_off = st_off + st->as_Store()->memory_size();
1.4035 + }
1.4036 + return true;
1.4037 +}
1.4038 +#endif //ASSERT
1.4039 +
1.4040 +
1.4041 +
1.4042 +
1.4043 +//============================MergeMemNode=====================================
1.4044 +//
1.4045 +// SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several
1.4046 +// contributing store or call operations. Each contributor provides the memory
1.4047 +// state for a particular "alias type" (see Compile::alias_type). For example,
1.4048 +// if a MergeMem has an input X for alias category #6, then any memory reference
1.4049 +// to alias category #6 may use X as its memory state input, as an exact equivalent
1.4050 +// to using the MergeMem as a whole.
1.4051 +// Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
1.4052 +//
1.4053 +// (Here, the <N> notation gives the index of the relevant adr_type.)
1.4054 +//
1.4055 +// In one special case (and more cases in the future), alias categories overlap.
1.4056 +// The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
1.4057 +// states. Therefore, if a MergeMem has only one contributing input W for Bot,
1.4058 +// it is exactly equivalent to that state W:
1.4059 +// MergeMem(<Bot>: W) <==> W
1.4060 +//
1.4061 +// Usually, the merge has more than one input. In that case, where inputs
1.4062 +// overlap (i.e., one is Bot), the narrower alias type determines the memory
1.4063 +// state for that type, and the wider alias type (Bot) fills in everywhere else:
1.4064 +// Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
1.4065 +// Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
1.4066 +//
1.4067 +// A merge can take a "wide" memory state as one of its narrow inputs.
1.4068 +// This simply means that the merge observes out only the relevant parts of
1.4069 +// the wide input. That is, wide memory states arriving at narrow merge inputs
1.4070 +// are implicitly "filtered" or "sliced" as necessary. (This is rare.)
1.4071 +//
1.4072 +// These rules imply that MergeMem nodes may cascade (via their <Bot> links),
1.4073 +// and that memory slices "leak through":
1.4074 +// MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
1.4075 +//
1.4076 +// But, in such a cascade, repeated memory slices can "block the leak":
1.4077 +// MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
1.4078 +//
1.4079 +// In the last example, Y is not part of the combined memory state of the
1.4080 +// outermost MergeMem. The system must, of course, prevent unschedulable
1.4081 +// memory states from arising, so you can be sure that the state Y is somehow
1.4082 +// a precursor to state Y'.
1.4083 +//
1.4084 +//
1.4085 +// REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
1.4086 +// of each MergeMemNode array are exactly the numerical alias indexes, including
1.4087 +// but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions
1.4088 +// Compile::alias_type (and kin) produce and manage these indexes.
1.4089 +//
1.4090 +// By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
1.4091 +// (Note that this provides quick access to the top node inside MergeMem methods,
1.4092 +// without the need to reach out via TLS to Compile::current.)
1.4093 +//
1.4094 +// As a consequence of what was just described, a MergeMem that represents a full
1.4095 +// memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
1.4096 +// containing all alias categories.
1.4097 +//
1.4098 +// MergeMem nodes never (?) have control inputs, so in(0) is NULL.
1.4099 +//
1.4100 +// All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
1.4101 +// a memory state for the alias type <N>, or else the top node, meaning that
1.4102 +// there is no particular input for that alias type. Note that the length of
1.4103 +// a MergeMem is variable, and may be extended at any time to accommodate new
1.4104 +// memory states at larger alias indexes. When merges grow, they are of course
1.4105 +// filled with "top" in the unused in() positions.
1.4106 +//
1.4107 +// This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
1.4108 +// (Top was chosen because it works smoothly with passes like GCM.)
1.4109 +//
1.4110 +// For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is
1.4111 +// the type of random VM bits like TLS references.) Since it is always the
1.4112 +// first non-Bot memory slice, some low-level loops use it to initialize an
1.4113 +// index variable: for (i = AliasIdxRaw; i < req(); i++).
1.4114 +//
1.4115 +//
1.4116 +// ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns
1.4117 +// the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns
1.4118 +// the memory state for alias type <N>, or (if there is no particular slice at <N>,
1.4119 +// it returns the base memory. To prevent bugs, memory_at does not accept <Top>
1.4120 +// or <Bot> indexes. The iterator MergeMemStream provides robust iteration over
1.4121 +// MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
1.4122 +//
1.4123 +// %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
1.4124 +// really that different from the other memory inputs. An abbreviation called
1.4125 +// "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
1.4126 +//
1.4127 +//
1.4128 +// PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent
1.4129 +// partial memory states. When a Phi splits through a MergeMem, the copy of the Phi
1.4130 +// that "emerges though" the base memory will be marked as excluding the alias types
1.4131 +// of the other (narrow-memory) copies which "emerged through" the narrow edges:
1.4132 +//
1.4133 +// Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
1.4134 +// ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
1.4135 +//
1.4136 +// This strange "subtraction" effect is necessary to ensure IGVN convergence.
1.4137 +// (It is currently unimplemented.) As you can see, the resulting merge is
1.4138 +// actually a disjoint union of memory states, rather than an overlay.
1.4139 +//
1.4140 +
1.4141 +//------------------------------MergeMemNode-----------------------------------
1.4142 +Node* MergeMemNode::make_empty_memory() {
1.4143 + Node* empty_memory = (Node*) Compile::current()->top();
1.4144 + assert(empty_memory->is_top(), "correct sentinel identity");
1.4145 + return empty_memory;
1.4146 +}
1.4147 +
1.4148 +MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
1.4149 + init_class_id(Class_MergeMem);
1.4150 + // all inputs are nullified in Node::Node(int)
1.4151 + // set_input(0, NULL); // no control input
1.4152 +
1.4153 + // Initialize the edges uniformly to top, for starters.
1.4154 + Node* empty_mem = make_empty_memory();
1.4155 + for (uint i = Compile::AliasIdxTop; i < req(); i++) {
1.4156 + init_req(i,empty_mem);
1.4157 + }
1.4158 + assert(empty_memory() == empty_mem, "");
1.4159 +
1.4160 + if( new_base != NULL && new_base->is_MergeMem() ) {
1.4161 + MergeMemNode* mdef = new_base->as_MergeMem();
1.4162 + assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
1.4163 + for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
1.4164 + mms.set_memory(mms.memory2());
1.4165 + }
1.4166 + assert(base_memory() == mdef->base_memory(), "");
1.4167 + } else {
1.4168 + set_base_memory(new_base);
1.4169 + }
1.4170 +}
1.4171 +
1.4172 +// Make a new, untransformed MergeMem with the same base as 'mem'.
1.4173 +// If mem is itself a MergeMem, populate the result with the same edges.
1.4174 +MergeMemNode* MergeMemNode::make(Compile* C, Node* mem) {
1.4175 + return new(C) MergeMemNode(mem);
1.4176 +}
1.4177 +
1.4178 +//------------------------------cmp--------------------------------------------
1.4179 +uint MergeMemNode::hash() const { return NO_HASH; }
1.4180 +uint MergeMemNode::cmp( const Node &n ) const {
1.4181 + return (&n == this); // Always fail except on self
1.4182 +}
1.4183 +
1.4184 +//------------------------------Identity---------------------------------------
1.4185 +Node* MergeMemNode::Identity(PhaseTransform *phase) {
1.4186 + // Identity if this merge point does not record any interesting memory
1.4187 + // disambiguations.
1.4188 + Node* base_mem = base_memory();
1.4189 + Node* empty_mem = empty_memory();
1.4190 + if (base_mem != empty_mem) { // Memory path is not dead?
1.4191 + for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
1.4192 + Node* mem = in(i);
1.4193 + if (mem != empty_mem && mem != base_mem) {
1.4194 + return this; // Many memory splits; no change
1.4195 + }
1.4196 + }
1.4197 + }
1.4198 + return base_mem; // No memory splits; ID on the one true input
1.4199 +}
1.4200 +
1.4201 +//------------------------------Ideal------------------------------------------
1.4202 +// This method is invoked recursively on chains of MergeMem nodes
1.4203 +Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1.4204 + // Remove chain'd MergeMems
1.4205 + //
1.4206 + // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
1.4207 + // relative to the "in(Bot)". Since we are patching both at the same time,
1.4208 + // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
1.4209 + // but rewrite each "in(i)" relative to the new "in(Bot)".
1.4210 + Node *progress = NULL;
1.4211 +
1.4212 +
1.4213 + Node* old_base = base_memory();
1.4214 + Node* empty_mem = empty_memory();
1.4215 + if (old_base == empty_mem)
1.4216 + return NULL; // Dead memory path.
1.4217 +
1.4218 + MergeMemNode* old_mbase;
1.4219 + if (old_base != NULL && old_base->is_MergeMem())
1.4220 + old_mbase = old_base->as_MergeMem();
1.4221 + else
1.4222 + old_mbase = NULL;
1.4223 + Node* new_base = old_base;
1.4224 +
1.4225 + // simplify stacked MergeMems in base memory
1.4226 + if (old_mbase) new_base = old_mbase->base_memory();
1.4227 +
1.4228 + // the base memory might contribute new slices beyond my req()
1.4229 + if (old_mbase) grow_to_match(old_mbase);
1.4230 +
1.4231 + // Look carefully at the base node if it is a phi.
1.4232 + PhiNode* phi_base;
1.4233 + if (new_base != NULL && new_base->is_Phi())
1.4234 + phi_base = new_base->as_Phi();
1.4235 + else
1.4236 + phi_base = NULL;
1.4237 +
1.4238 + Node* phi_reg = NULL;
1.4239 + uint phi_len = (uint)-1;
1.4240 + if (phi_base != NULL && !phi_base->is_copy()) {
1.4241 + // do not examine phi if degraded to a copy
1.4242 + phi_reg = phi_base->region();
1.4243 + phi_len = phi_base->req();
1.4244 + // see if the phi is unfinished
1.4245 + for (uint i = 1; i < phi_len; i++) {
1.4246 + if (phi_base->in(i) == NULL) {
1.4247 + // incomplete phi; do not look at it yet!
1.4248 + phi_reg = NULL;
1.4249 + phi_len = (uint)-1;
1.4250 + break;
1.4251 + }
1.4252 + }
1.4253 + }
1.4254 +
1.4255 + // Note: We do not call verify_sparse on entry, because inputs
1.4256 + // can normalize to the base_memory via subsume_node or similar
1.4257 + // mechanisms. This method repairs that damage.
1.4258 +
1.4259 + assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
1.4260 +
1.4261 + // Look at each slice.
1.4262 + for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
1.4263 + Node* old_in = in(i);
1.4264 + // calculate the old memory value
1.4265 + Node* old_mem = old_in;
1.4266 + if (old_mem == empty_mem) old_mem = old_base;
1.4267 + assert(old_mem == memory_at(i), "");
1.4268 +
1.4269 + // maybe update (reslice) the old memory value
1.4270 +
1.4271 + // simplify stacked MergeMems
1.4272 + Node* new_mem = old_mem;
1.4273 + MergeMemNode* old_mmem;
1.4274 + if (old_mem != NULL && old_mem->is_MergeMem())
1.4275 + old_mmem = old_mem->as_MergeMem();
1.4276 + else
1.4277 + old_mmem = NULL;
1.4278 + if (old_mmem == this) {
1.4279 + // This can happen if loops break up and safepoints disappear.
1.4280 + // A merge of BotPtr (default) with a RawPtr memory derived from a
1.4281 + // safepoint can be rewritten to a merge of the same BotPtr with
1.4282 + // the BotPtr phi coming into the loop. If that phi disappears
1.4283 + // also, we can end up with a self-loop of the mergemem.
1.4284 + // In general, if loops degenerate and memory effects disappear,
1.4285 + // a mergemem can be left looking at itself. This simply means
1.4286 + // that the mergemem's default should be used, since there is
1.4287 + // no longer any apparent effect on this slice.
1.4288 + // Note: If a memory slice is a MergeMem cycle, it is unreachable
1.4289 + // from start. Update the input to TOP.
1.4290 + new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
1.4291 + }
1.4292 + else if (old_mmem != NULL) {
1.4293 + new_mem = old_mmem->memory_at(i);
1.4294 + }
1.4295 + // else preceding memory was not a MergeMem
1.4296 +
1.4297 + // replace equivalent phis (unfortunately, they do not GVN together)
1.4298 + if (new_mem != NULL && new_mem != new_base &&
1.4299 + new_mem->req() == phi_len && new_mem->in(0) == phi_reg) {
1.4300 + if (new_mem->is_Phi()) {
1.4301 + PhiNode* phi_mem = new_mem->as_Phi();
1.4302 + for (uint i = 1; i < phi_len; i++) {
1.4303 + if (phi_base->in(i) != phi_mem->in(i)) {
1.4304 + phi_mem = NULL;
1.4305 + break;
1.4306 + }
1.4307 + }
1.4308 + if (phi_mem != NULL) {
1.4309 + // equivalent phi nodes; revert to the def
1.4310 + new_mem = new_base;
1.4311 + }
1.4312 + }
1.4313 + }
1.4314 +
1.4315 + // maybe store down a new value
1.4316 + Node* new_in = new_mem;
1.4317 + if (new_in == new_base) new_in = empty_mem;
1.4318 +
1.4319 + if (new_in != old_in) {
1.4320 + // Warning: Do not combine this "if" with the previous "if"
1.4321 + // A memory slice might have be be rewritten even if it is semantically
1.4322 + // unchanged, if the base_memory value has changed.
1.4323 + set_req(i, new_in);
1.4324 + progress = this; // Report progress
1.4325 + }
1.4326 + }
1.4327 +
1.4328 + if (new_base != old_base) {
1.4329 + set_req(Compile::AliasIdxBot, new_base);
1.4330 + // Don't use set_base_memory(new_base), because we need to update du.
1.4331 + assert(base_memory() == new_base, "");
1.4332 + progress = this;
1.4333 + }
1.4334 +
1.4335 + if( base_memory() == this ) {
1.4336 + // a self cycle indicates this memory path is dead
1.4337 + set_req(Compile::AliasIdxBot, empty_mem);
1.4338 + }
1.4339 +
1.4340 + // Resolve external cycles by calling Ideal on a MergeMem base_memory
1.4341 + // Recursion must occur after the self cycle check above
1.4342 + if( base_memory()->is_MergeMem() ) {
1.4343 + MergeMemNode *new_mbase = base_memory()->as_MergeMem();
1.4344 + Node *m = phase->transform(new_mbase); // Rollup any cycles
1.4345 + if( m != NULL && (m->is_top() ||
1.4346 + m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) {
1.4347 + // propagate rollup of dead cycle to self
1.4348 + set_req(Compile::AliasIdxBot, empty_mem);
1.4349 + }
1.4350 + }
1.4351 +
1.4352 + if( base_memory() == empty_mem ) {
1.4353 + progress = this;
1.4354 + // Cut inputs during Parse phase only.
1.4355 + // During Optimize phase a dead MergeMem node will be subsumed by Top.
1.4356 + if( !can_reshape ) {
1.4357 + for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
1.4358 + if( in(i) != empty_mem ) { set_req(i, empty_mem); }
1.4359 + }
1.4360 + }
1.4361 + }
1.4362 +
1.4363 + if( !progress && base_memory()->is_Phi() && can_reshape ) {
1.4364 + // Check if PhiNode::Ideal's "Split phis through memory merges"
1.4365 + // transform should be attempted. Look for this->phi->this cycle.
1.4366 + uint merge_width = req();
1.4367 + if (merge_width > Compile::AliasIdxRaw) {
1.4368 + PhiNode* phi = base_memory()->as_Phi();
1.4369 + for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
1.4370 + if (phi->in(i) == this) {
1.4371 + phase->is_IterGVN()->_worklist.push(phi);
1.4372 + break;
1.4373 + }
1.4374 + }
1.4375 + }
1.4376 + }
1.4377 +
1.4378 + assert(progress || verify_sparse(), "please, no dups of base");
1.4379 + return progress;
1.4380 +}
1.4381 +
1.4382 +//-------------------------set_base_memory-------------------------------------
1.4383 +void MergeMemNode::set_base_memory(Node *new_base) {
1.4384 + Node* empty_mem = empty_memory();
1.4385 + set_req(Compile::AliasIdxBot, new_base);
1.4386 + assert(memory_at(req()) == new_base, "must set default memory");
1.4387 + // Clear out other occurrences of new_base:
1.4388 + if (new_base != empty_mem) {
1.4389 + for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
1.4390 + if (in(i) == new_base) set_req(i, empty_mem);
1.4391 + }
1.4392 + }
1.4393 +}
1.4394 +
1.4395 +//------------------------------out_RegMask------------------------------------
1.4396 +const RegMask &MergeMemNode::out_RegMask() const {
1.4397 + return RegMask::Empty;
1.4398 +}
1.4399 +
1.4400 +//------------------------------dump_spec--------------------------------------
1.4401 +#ifndef PRODUCT
1.4402 +void MergeMemNode::dump_spec(outputStream *st) const {
1.4403 + st->print(" {");
1.4404 + Node* base_mem = base_memory();
1.4405 + for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
1.4406 + Node* mem = memory_at(i);
1.4407 + if (mem == base_mem) { st->print(" -"); continue; }
1.4408 + st->print( " N%d:", mem->_idx );
1.4409 + Compile::current()->get_adr_type(i)->dump_on(st);
1.4410 + }
1.4411 + st->print(" }");
1.4412 +}
1.4413 +#endif // !PRODUCT
1.4414 +
1.4415 +
1.4416 +#ifdef ASSERT
1.4417 +static bool might_be_same(Node* a, Node* b) {
1.4418 + if (a == b) return true;
1.4419 + if (!(a->is_Phi() || b->is_Phi())) return false;
1.4420 + // phis shift around during optimization
1.4421 + return true; // pretty stupid...
1.4422 +}
1.4423 +
1.4424 +// verify a narrow slice (either incoming or outgoing)
1.4425 +static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
1.4426 + if (!VerifyAliases) return; // don't bother to verify unless requested
1.4427 + if (is_error_reported()) return; // muzzle asserts when debugging an error
1.4428 + if (Node::in_dump()) return; // muzzle asserts when printing
1.4429 + assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
1.4430 + assert(n != NULL, "");
1.4431 + // Elide intervening MergeMem's
1.4432 + while (n->is_MergeMem()) {
1.4433 + n = n->as_MergeMem()->memory_at(alias_idx);
1.4434 + }
1.4435 + Compile* C = Compile::current();
1.4436 + const TypePtr* n_adr_type = n->adr_type();
1.4437 + if (n == m->empty_memory()) {
1.4438 + // Implicit copy of base_memory()
1.4439 + } else if (n_adr_type != TypePtr::BOTTOM) {
1.4440 + assert(n_adr_type != NULL, "new memory must have a well-defined adr_type");
1.4441 + assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
1.4442 + } else {
1.4443 + // A few places like make_runtime_call "know" that VM calls are narrow,
1.4444 + // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
1.4445 + bool expected_wide_mem = false;
1.4446 + if (n == m->base_memory()) {
1.4447 + expected_wide_mem = true;
1.4448 + } else if (alias_idx == Compile::AliasIdxRaw ||
1.4449 + n == m->memory_at(Compile::AliasIdxRaw)) {
1.4450 + expected_wide_mem = true;
1.4451 + } else if (!C->alias_type(alias_idx)->is_rewritable()) {
1.4452 + // memory can "leak through" calls on channels that
1.4453 + // are write-once. Allow this also.
1.4454 + expected_wide_mem = true;
1.4455 + }
1.4456 + assert(expected_wide_mem, "expected narrow slice replacement");
1.4457 + }
1.4458 +}
1.4459 +#else // !ASSERT
1.4460 +#define verify_memory_slice(m,i,n) (void)(0) // PRODUCT version is no-op
1.4461 +#endif
1.4462 +
1.4463 +
1.4464 +//-----------------------------memory_at---------------------------------------
1.4465 +Node* MergeMemNode::memory_at(uint alias_idx) const {
1.4466 + assert(alias_idx >= Compile::AliasIdxRaw ||
1.4467 + alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0,
1.4468 + "must avoid base_memory and AliasIdxTop");
1.4469 +
1.4470 + // Otherwise, it is a narrow slice.
1.4471 + Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
1.4472 + Compile *C = Compile::current();
1.4473 + if (is_empty_memory(n)) {
1.4474 + // the array is sparse; empty slots are the "top" node
1.4475 + n = base_memory();
1.4476 + assert(Node::in_dump()
1.4477 + || n == NULL || n->bottom_type() == Type::TOP
1.4478 + || n->adr_type() == NULL // address is TOP
1.4479 + || n->adr_type() == TypePtr::BOTTOM
1.4480 + || n->adr_type() == TypeRawPtr::BOTTOM
1.4481 + || Compile::current()->AliasLevel() == 0,
1.4482 + "must be a wide memory");
1.4483 + // AliasLevel == 0 if we are organizing the memory states manually.
1.4484 + // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
1.4485 + } else {
1.4486 + // make sure the stored slice is sane
1.4487 + #ifdef ASSERT
1.4488 + if (is_error_reported() || Node::in_dump()) {
1.4489 + } else if (might_be_same(n, base_memory())) {
1.4490 + // Give it a pass: It is a mostly harmless repetition of the base.
1.4491 + // This can arise normally from node subsumption during optimization.
1.4492 + } else {
1.4493 + verify_memory_slice(this, alias_idx, n);
1.4494 + }
1.4495 + #endif
1.4496 + }
1.4497 + return n;
1.4498 +}
1.4499 +
1.4500 +//---------------------------set_memory_at-------------------------------------
1.4501 +void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
1.4502 + verify_memory_slice(this, alias_idx, n);
1.4503 + Node* empty_mem = empty_memory();
1.4504 + if (n == base_memory()) n = empty_mem; // collapse default
1.4505 + uint need_req = alias_idx+1;
1.4506 + if (req() < need_req) {
1.4507 + if (n == empty_mem) return; // already the default, so do not grow me
1.4508 + // grow the sparse array
1.4509 + do {
1.4510 + add_req(empty_mem);
1.4511 + } while (req() < need_req);
1.4512 + }
1.4513 + set_req( alias_idx, n );
1.4514 +}
1.4515 +
1.4516 +
1.4517 +
1.4518 +//--------------------------iteration_setup------------------------------------
1.4519 +void MergeMemNode::iteration_setup(const MergeMemNode* other) {
1.4520 + if (other != NULL) {
1.4521 + grow_to_match(other);
1.4522 + // invariant: the finite support of mm2 is within mm->req()
1.4523 + #ifdef ASSERT
1.4524 + for (uint i = req(); i < other->req(); i++) {
1.4525 + assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
1.4526 + }
1.4527 + #endif
1.4528 + }
1.4529 + // Replace spurious copies of base_memory by top.
1.4530 + Node* base_mem = base_memory();
1.4531 + if (base_mem != NULL && !base_mem->is_top()) {
1.4532 + for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
1.4533 + if (in(i) == base_mem)
1.4534 + set_req(i, empty_memory());
1.4535 + }
1.4536 + }
1.4537 +}
1.4538 +
1.4539 +//---------------------------grow_to_match-------------------------------------
1.4540 +void MergeMemNode::grow_to_match(const MergeMemNode* other) {
1.4541 + Node* empty_mem = empty_memory();
1.4542 + assert(other->is_empty_memory(empty_mem), "consistent sentinels");
1.4543 + // look for the finite support of the other memory
1.4544 + for (uint i = other->req(); --i >= req(); ) {
1.4545 + if (other->in(i) != empty_mem) {
1.4546 + uint new_len = i+1;
1.4547 + while (req() < new_len) add_req(empty_mem);
1.4548 + break;
1.4549 + }
1.4550 + }
1.4551 +}
1.4552 +
1.4553 +//---------------------------verify_sparse-------------------------------------
1.4554 +#ifndef PRODUCT
1.4555 +bool MergeMemNode::verify_sparse() const {
1.4556 + assert(is_empty_memory(make_empty_memory()), "sane sentinel");
1.4557 + Node* base_mem = base_memory();
1.4558 + // The following can happen in degenerate cases, since empty==top.
1.4559 + if (is_empty_memory(base_mem)) return true;
1.4560 + for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
1.4561 + assert(in(i) != NULL, "sane slice");
1.4562 + if (in(i) == base_mem) return false; // should have been the sentinel value!
1.4563 + }
1.4564 + return true;
1.4565 +}
1.4566 +
1.4567 +bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
1.4568 + Node* n;
1.4569 + n = mm->in(idx);
1.4570 + if (mem == n) return true; // might be empty_memory()
1.4571 + n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
1.4572 + if (mem == n) return true;
1.4573 + while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) {
1.4574 + if (mem == n) return true;
1.4575 + if (n == NULL) break;
1.4576 + }
1.4577 + return false;
1.4578 +}
1.4579 +#endif // !PRODUCT
