duke@435: /* xdono@631: * Copyright 1997-2008 Sun Microsystems, Inc. All Rights Reserved. duke@435: * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. duke@435: * duke@435: * This code is free software; you can redistribute it and/or modify it duke@435: * under the terms of the GNU General Public License version 2 only, as duke@435: * published by the Free Software Foundation. duke@435: * duke@435: * This code is distributed in the hope that it will be useful, but WITHOUT duke@435: * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or duke@435: * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License duke@435: * version 2 for more details (a copy is included in the LICENSE file that duke@435: * accompanied this code). duke@435: * duke@435: * You should have received a copy of the GNU General Public License version duke@435: * 2 along with this work; if not, write to the Free Software Foundation, duke@435: * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. duke@435: * duke@435: * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, duke@435: * CA 95054 USA or visit www.sun.com if you need additional information or duke@435: * have any questions. duke@435: * duke@435: */ duke@435: duke@435: // Portions of code courtesy of Clifford Click duke@435: duke@435: // Optimization - Graph Style duke@435: duke@435: #include "incls/_precompiled.incl" duke@435: #include "incls/_memnode.cpp.incl" duke@435: kvn@509: static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st); kvn@509: duke@435: //============================================================================= duke@435: uint MemNode::size_of() const { return sizeof(*this); } duke@435: duke@435: const TypePtr *MemNode::adr_type() const { duke@435: Node* adr = in(Address); duke@435: const TypePtr* cross_check = NULL; duke@435: DEBUG_ONLY(cross_check = _adr_type); duke@435: return calculate_adr_type(adr->bottom_type(), cross_check); duke@435: } duke@435: duke@435: #ifndef PRODUCT duke@435: void MemNode::dump_spec(outputStream *st) const { duke@435: if (in(Address) == NULL) return; // node is dead duke@435: #ifndef ASSERT duke@435: // fake the missing field duke@435: const TypePtr* _adr_type = NULL; duke@435: if (in(Address) != NULL) duke@435: _adr_type = in(Address)->bottom_type()->isa_ptr(); duke@435: #endif duke@435: dump_adr_type(this, _adr_type, st); duke@435: duke@435: Compile* C = Compile::current(); duke@435: if( C->alias_type(_adr_type)->is_volatile() ) duke@435: st->print(" Volatile!"); duke@435: } duke@435: duke@435: void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) { duke@435: st->print(" @"); duke@435: if (adr_type == NULL) { duke@435: st->print("NULL"); duke@435: } else { duke@435: adr_type->dump_on(st); duke@435: Compile* C = Compile::current(); duke@435: Compile::AliasType* atp = NULL; duke@435: if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type); duke@435: if (atp == NULL) duke@435: st->print(", idx=?\?;"); duke@435: else if (atp->index() == Compile::AliasIdxBot) duke@435: st->print(", idx=Bot;"); duke@435: else if (atp->index() == Compile::AliasIdxTop) duke@435: st->print(", idx=Top;"); duke@435: else if (atp->index() == Compile::AliasIdxRaw) duke@435: st->print(", idx=Raw;"); duke@435: else { duke@435: ciField* field = atp->field(); duke@435: if (field) { duke@435: st->print(", name="); duke@435: field->print_name_on(st); duke@435: } duke@435: st->print(", idx=%d;", atp->index()); duke@435: } duke@435: } duke@435: } duke@435: duke@435: extern void print_alias_types(); duke@435: duke@435: #endif duke@435: kvn@509: Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypePtr *t_adr, PhaseGVN *phase) { kvn@509: const TypeOopPtr *tinst = t_adr->isa_oopptr(); kvn@658: if (tinst == NULL || !tinst->is_known_instance_field()) kvn@509: return mchain; // don't try to optimize non-instance types kvn@509: uint instance_id = tinst->instance_id(); kvn@688: Node *start_mem = phase->C->start()->proj_out(TypeFunc::Memory); kvn@509: Node *prev = NULL; kvn@509: Node *result = mchain; kvn@509: while (prev != result) { kvn@509: prev = result; kvn@688: if (result == start_mem) twisti@1040: break; // hit one of our sentinels kvn@509: // skip over a call which does not affect this memory slice kvn@509: if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) { kvn@509: Node *proj_in = result->in(0); kvn@688: if (proj_in->is_Allocate() && proj_in->_idx == instance_id) { twisti@1040: break; // hit one of our sentinels kvn@688: } else if (proj_in->is_Call()) { kvn@509: CallNode *call = proj_in->as_Call(); kvn@509: if (!call->may_modify(t_adr, phase)) { kvn@509: result = call->in(TypeFunc::Memory); kvn@509: } kvn@509: } else if (proj_in->is_Initialize()) { kvn@509: AllocateNode* alloc = proj_in->as_Initialize()->allocation(); kvn@509: // Stop if this is the initialization for the object instance which kvn@509: // which contains this memory slice, otherwise skip over it. kvn@509: if (alloc != NULL && alloc->_idx != instance_id) { kvn@509: result = proj_in->in(TypeFunc::Memory); kvn@509: } kvn@509: } else if (proj_in->is_MemBar()) { kvn@509: result = proj_in->in(TypeFunc::Memory); kvn@688: } else { kvn@688: assert(false, "unexpected projection"); kvn@509: } kvn@509: } else if (result->is_MergeMem()) { kvn@509: result = step_through_mergemem(phase, result->as_MergeMem(), t_adr, NULL, tty); kvn@509: } kvn@509: } kvn@509: return result; kvn@509: } kvn@509: kvn@509: Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, PhaseGVN *phase) { kvn@509: const TypeOopPtr *t_oop = t_adr->isa_oopptr(); kvn@658: bool is_instance = (t_oop != NULL) && t_oop->is_known_instance_field(); kvn@509: PhaseIterGVN *igvn = phase->is_IterGVN(); kvn@509: Node *result = mchain; kvn@509: result = optimize_simple_memory_chain(result, t_adr, phase); kvn@509: if (is_instance && igvn != NULL && result->is_Phi()) { kvn@509: PhiNode *mphi = result->as_Phi(); kvn@509: assert(mphi->bottom_type() == Type::MEMORY, "memory phi required"); kvn@509: const TypePtr *t = mphi->adr_type(); kvn@598: if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM || kvn@658: t->isa_oopptr() && !t->is_oopptr()->is_known_instance() && kvn@682: t->is_oopptr()->cast_to_exactness(true) kvn@682: ->is_oopptr()->cast_to_ptr_type(t_oop->ptr()) kvn@682: ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop) { kvn@509: // clone the Phi with our address type kvn@509: result = mphi->split_out_instance(t_adr, igvn); kvn@509: } else { kvn@509: assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain"); kvn@509: } kvn@509: } kvn@509: return result; kvn@509: } kvn@509: kvn@499: static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) { kvn@499: uint alias_idx = phase->C->get_alias_index(tp); kvn@499: Node *mem = mmem; kvn@499: #ifdef ASSERT kvn@499: { kvn@499: // Check that current type is consistent with the alias index used during graph construction kvn@499: assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx"); kvn@499: bool consistent = adr_check == NULL || adr_check->empty() || kvn@499: phase->C->must_alias(adr_check, alias_idx ); kvn@499: // Sometimes dead array references collapse to a[-1], a[-2], or a[-3] kvn@499: if( !consistent && adr_check != NULL && !adr_check->empty() && rasbold@604: tp->isa_aryptr() && tp->offset() == Type::OffsetBot && kvn@499: adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot && kvn@499: ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() || kvn@499: adr_check->offset() == oopDesc::klass_offset_in_bytes() || kvn@499: adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) { kvn@499: // don't assert if it is dead code. kvn@499: consistent = true; kvn@499: } kvn@499: if( !consistent ) { kvn@499: st->print("alias_idx==%d, adr_check==", alias_idx); kvn@499: if( adr_check == NULL ) { kvn@499: st->print("NULL"); kvn@499: } else { kvn@499: adr_check->dump(); kvn@499: } kvn@499: st->cr(); kvn@499: print_alias_types(); kvn@499: assert(consistent, "adr_check must match alias idx"); kvn@499: } kvn@499: } kvn@499: #endif kvn@499: // TypeInstPtr::NOTNULL+any is an OOP with unknown offset - generally kvn@499: // means an array I have not precisely typed yet. Do not do any kvn@499: // alias stuff with it any time soon. kvn@499: const TypeOopPtr *tinst = tp->isa_oopptr(); kvn@499: if( tp->base() != Type::AnyPtr && kvn@499: !(tinst && kvn@499: tinst->klass()->is_java_lang_Object() && kvn@499: tinst->offset() == Type::OffsetBot) ) { kvn@499: // compress paths and change unreachable cycles to TOP kvn@499: // If not, we can update the input infinitely along a MergeMem cycle kvn@499: // Equivalent code in PhiNode::Ideal kvn@499: Node* m = phase->transform(mmem); twisti@1040: // If transformed to a MergeMem, get the desired slice kvn@499: // Otherwise the returned node represents memory for every slice kvn@499: mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m; kvn@499: // Update input if it is progress over what we have now kvn@499: } kvn@499: return mem; kvn@499: } kvn@499: duke@435: //--------------------------Ideal_common--------------------------------------- duke@435: // Look for degenerate control and memory inputs. Bypass MergeMem inputs. duke@435: // Unhook non-raw memories from complete (macro-expanded) initializations. duke@435: Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) { duke@435: // If our control input is a dead region, kill all below the region duke@435: Node *ctl = in(MemNode::Control); duke@435: if (ctl && remove_dead_region(phase, can_reshape)) duke@435: return this; kvn@740: ctl = in(MemNode::Control); kvn@740: // Don't bother trying to transform a dead node kvn@740: if( ctl && ctl->is_top() ) return NodeSentinel; duke@435: duke@435: // Ignore if memory is dead, or self-loop duke@435: Node *mem = in(MemNode::Memory); duke@435: if( phase->type( mem ) == Type::TOP ) return NodeSentinel; // caller will return NULL duke@435: assert( mem != this, "dead loop in MemNode::Ideal" ); duke@435: duke@435: Node *address = in(MemNode::Address); duke@435: const Type *t_adr = phase->type( address ); duke@435: if( t_adr == Type::TOP ) return NodeSentinel; // caller will return NULL duke@435: kvn@855: PhaseIterGVN *igvn = phase->is_IterGVN(); kvn@855: if( can_reshape && igvn != NULL && igvn->_worklist.member(address) ) { kvn@855: // The address's base and type may change when the address is processed. kvn@855: // Delay this mem node transformation until the address is processed. kvn@855: phase->is_IterGVN()->_worklist.push(this); kvn@855: return NodeSentinel; // caller will return NULL kvn@855: } kvn@855: duke@435: // Avoid independent memory operations duke@435: Node* old_mem = mem; duke@435: kvn@471: // The code which unhooks non-raw memories from complete (macro-expanded) kvn@471: // initializations was removed. After macro-expansion all stores catched kvn@471: // by Initialize node became raw stores and there is no information kvn@471: // which memory slices they modify. So it is unsafe to move any memory kvn@471: // operation above these stores. Also in most cases hooked non-raw memories kvn@471: // were already unhooked by using information from detect_ptr_independence() kvn@471: // and find_previous_store(). duke@435: duke@435: if (mem->is_MergeMem()) { duke@435: MergeMemNode* mmem = mem->as_MergeMem(); duke@435: const TypePtr *tp = t_adr->is_ptr(); kvn@499: kvn@499: mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty); duke@435: } duke@435: duke@435: if (mem != old_mem) { duke@435: set_req(MemNode::Memory, mem); kvn@740: if (phase->type( mem ) == Type::TOP) return NodeSentinel; duke@435: return this; duke@435: } duke@435: duke@435: // let the subclass continue analyzing... duke@435: return NULL; duke@435: } duke@435: duke@435: // Helper function for proving some simple control dominations. kvn@554: // Attempt to prove that all control inputs of 'dom' dominate 'sub'. duke@435: // Already assumes that 'dom' is available at 'sub', and that 'sub' duke@435: // is not a constant (dominated by the method's StartNode). duke@435: // Used by MemNode::find_previous_store to prove that the duke@435: // control input of a memory operation predates (dominates) duke@435: // an allocation it wants to look past. kvn@554: bool MemNode::all_controls_dominate(Node* dom, Node* sub) { kvn@554: if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top()) kvn@554: return false; // Conservative answer for dead code kvn@554: kvn@628: // Check 'dom'. Skip Proj and CatchProj nodes. kvn@554: dom = dom->find_exact_control(dom); kvn@554: if (dom == NULL || dom->is_top()) kvn@554: return false; // Conservative answer for dead code kvn@554: kvn@628: if (dom == sub) { kvn@628: // For the case when, for example, 'sub' is Initialize and the original kvn@628: // 'dom' is Proj node of the 'sub'. kvn@628: return false; kvn@628: } kvn@628: kvn@590: if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub) kvn@554: return true; kvn@554: kvn@554: // 'dom' dominates 'sub' if its control edge and control edges kvn@554: // of all its inputs dominate or equal to sub's control edge. kvn@554: kvn@554: // Currently 'sub' is either Allocate, Initialize or Start nodes. kvn@598: // Or Region for the check in LoadNode::Ideal(); kvn@598: // 'sub' should have sub->in(0) != NULL. kvn@598: assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() || kvn@598: sub->is_Region(), "expecting only these nodes"); kvn@554: kvn@554: // Get control edge of 'sub'. kvn@628: Node* orig_sub = sub; kvn@554: sub = sub->find_exact_control(sub->in(0)); kvn@554: if (sub == NULL || sub->is_top()) kvn@554: return false; // Conservative answer for dead code kvn@554: kvn@554: assert(sub->is_CFG(), "expecting control"); kvn@554: kvn@554: if (sub == dom) kvn@554: return true; kvn@554: kvn@554: if (sub->is_Start() || sub->is_Root()) kvn@554: return false; kvn@554: kvn@554: { kvn@554: // Check all control edges of 'dom'. kvn@554: kvn@554: ResourceMark rm; kvn@554: Arena* arena = Thread::current()->resource_area(); kvn@554: Node_List nlist(arena); kvn@554: Unique_Node_List dom_list(arena); kvn@554: kvn@554: dom_list.push(dom); kvn@554: bool only_dominating_controls = false; kvn@554: kvn@554: for (uint next = 0; next < dom_list.size(); next++) { kvn@554: Node* n = dom_list.at(next); kvn@628: if (n == orig_sub) kvn@628: return false; // One of dom's inputs dominated by sub. kvn@554: if (!n->is_CFG() && n->pinned()) { kvn@554: // Check only own control edge for pinned non-control nodes. kvn@554: n = n->find_exact_control(n->in(0)); kvn@554: if (n == NULL || n->is_top()) kvn@554: return false; // Conservative answer for dead code kvn@554: assert(n->is_CFG(), "expecting control"); kvn@628: dom_list.push(n); kvn@628: } else if (n->is_Con() || n->is_Start() || n->is_Root()) { kvn@554: only_dominating_controls = true; kvn@554: } else if (n->is_CFG()) { kvn@554: if (n->dominates(sub, nlist)) kvn@554: only_dominating_controls = true; kvn@554: else kvn@554: return false; kvn@554: } else { kvn@554: // First, own control edge. kvn@554: Node* m = n->find_exact_control(n->in(0)); kvn@590: if (m != NULL) { kvn@590: if (m->is_top()) kvn@590: return false; // Conservative answer for dead code kvn@590: dom_list.push(m); kvn@590: } kvn@554: // Now, the rest of edges. kvn@554: uint cnt = n->req(); kvn@554: for (uint i = 1; i < cnt; i++) { kvn@554: m = n->find_exact_control(n->in(i)); kvn@554: if (m == NULL || m->is_top()) kvn@554: continue; kvn@554: dom_list.push(m); duke@435: } duke@435: } duke@435: } kvn@554: return only_dominating_controls; duke@435: } duke@435: } duke@435: duke@435: //---------------------detect_ptr_independence--------------------------------- duke@435: // Used by MemNode::find_previous_store to prove that two base duke@435: // pointers are never equal. duke@435: // The pointers are accompanied by their associated allocations, duke@435: // if any, which have been previously discovered by the caller. duke@435: bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1, duke@435: Node* p2, AllocateNode* a2, duke@435: PhaseTransform* phase) { duke@435: // Attempt to prove that these two pointers cannot be aliased. duke@435: // They may both manifestly be allocations, and they should differ. duke@435: // Or, if they are not both allocations, they can be distinct constants. duke@435: // Otherwise, one is an allocation and the other a pre-existing value. duke@435: if (a1 == NULL && a2 == NULL) { // neither an allocation duke@435: return (p1 != p2) && p1->is_Con() && p2->is_Con(); duke@435: } else if (a1 != NULL && a2 != NULL) { // both allocations duke@435: return (a1 != a2); duke@435: } else if (a1 != NULL) { // one allocation a1 duke@435: // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.) kvn@554: return all_controls_dominate(p2, a1); duke@435: } else { //(a2 != NULL) // one allocation a2 kvn@554: return all_controls_dominate(p1, a2); duke@435: } duke@435: return false; duke@435: } duke@435: duke@435: duke@435: // The logic for reordering loads and stores uses four steps: duke@435: // (a) Walk carefully past stores and initializations which we duke@435: // can prove are independent of this load. duke@435: // (b) Observe that the next memory state makes an exact match duke@435: // with self (load or store), and locate the relevant store. duke@435: // (c) Ensure that, if we were to wire self directly to the store, duke@435: // the optimizer would fold it up somehow. duke@435: // (d) Do the rewiring, and return, depending on some other part of duke@435: // the optimizer to fold up the load. duke@435: // This routine handles steps (a) and (b). Steps (c) and (d) are duke@435: // specific to loads and stores, so they are handled by the callers. duke@435: // (Currently, only LoadNode::Ideal has steps (c), (d). More later.) duke@435: // duke@435: Node* MemNode::find_previous_store(PhaseTransform* phase) { duke@435: Node* ctrl = in(MemNode::Control); duke@435: Node* adr = in(MemNode::Address); duke@435: intptr_t offset = 0; duke@435: Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); duke@435: AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase); duke@435: duke@435: if (offset == Type::OffsetBot) duke@435: return NULL; // cannot unalias unless there are precise offsets duke@435: kvn@509: const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr(); kvn@509: duke@435: intptr_t size_in_bytes = memory_size(); duke@435: duke@435: Node* mem = in(MemNode::Memory); // start searching here... duke@435: duke@435: int cnt = 50; // Cycle limiter duke@435: for (;;) { // While we can dance past unrelated stores... duke@435: if (--cnt < 0) break; // Caught in cycle or a complicated dance? duke@435: duke@435: if (mem->is_Store()) { duke@435: Node* st_adr = mem->in(MemNode::Address); duke@435: intptr_t st_offset = 0; duke@435: Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset); duke@435: if (st_base == NULL) duke@435: break; // inscrutable pointer duke@435: if (st_offset != offset && st_offset != Type::OffsetBot) { duke@435: const int MAX_STORE = BytesPerLong; duke@435: if (st_offset >= offset + size_in_bytes || duke@435: st_offset <= offset - MAX_STORE || duke@435: st_offset <= offset - mem->as_Store()->memory_size()) { duke@435: // Success: The offsets are provably independent. duke@435: // (You may ask, why not just test st_offset != offset and be done? duke@435: // The answer is that stores of different sizes can co-exist duke@435: // in the same sequence of RawMem effects. We sometimes initialize duke@435: // a whole 'tile' of array elements with a single jint or jlong.) duke@435: mem = mem->in(MemNode::Memory); duke@435: continue; // (a) advance through independent store memory duke@435: } duke@435: } duke@435: if (st_base != base && duke@435: detect_ptr_independence(base, alloc, duke@435: st_base, duke@435: AllocateNode::Ideal_allocation(st_base, phase), duke@435: phase)) { duke@435: // Success: The bases are provably independent. duke@435: mem = mem->in(MemNode::Memory); duke@435: continue; // (a) advance through independent store memory duke@435: } duke@435: duke@435: // (b) At this point, if the bases or offsets do not agree, we lose, duke@435: // since we have not managed to prove 'this' and 'mem' independent. duke@435: if (st_base == base && st_offset == offset) { duke@435: return mem; // let caller handle steps (c), (d) duke@435: } duke@435: duke@435: } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) { duke@435: InitializeNode* st_init = mem->in(0)->as_Initialize(); duke@435: AllocateNode* st_alloc = st_init->allocation(); duke@435: if (st_alloc == NULL) duke@435: break; // something degenerated duke@435: bool known_identical = false; duke@435: bool known_independent = false; duke@435: if (alloc == st_alloc) duke@435: known_identical = true; duke@435: else if (alloc != NULL) duke@435: known_independent = true; kvn@554: else if (all_controls_dominate(this, st_alloc)) duke@435: known_independent = true; duke@435: duke@435: if (known_independent) { duke@435: // The bases are provably independent: Either they are duke@435: // manifestly distinct allocations, or else the control duke@435: // of this load dominates the store's allocation. duke@435: int alias_idx = phase->C->get_alias_index(adr_type()); duke@435: if (alias_idx == Compile::AliasIdxRaw) { duke@435: mem = st_alloc->in(TypeFunc::Memory); duke@435: } else { duke@435: mem = st_init->memory(alias_idx); duke@435: } duke@435: continue; // (a) advance through independent store memory duke@435: } duke@435: duke@435: // (b) at this point, if we are not looking at a store initializing duke@435: // the same allocation we are loading from, we lose. duke@435: if (known_identical) { duke@435: // From caller, can_see_stored_value will consult find_captured_store. duke@435: return mem; // let caller handle steps (c), (d) duke@435: } duke@435: kvn@658: } else if (addr_t != NULL && addr_t->is_known_instance_field()) { kvn@509: // Can't use optimize_simple_memory_chain() since it needs PhaseGVN. kvn@509: if (mem->is_Proj() && mem->in(0)->is_Call()) { kvn@509: CallNode *call = mem->in(0)->as_Call(); kvn@509: if (!call->may_modify(addr_t, phase)) { kvn@509: mem = call->in(TypeFunc::Memory); kvn@509: continue; // (a) advance through independent call memory kvn@509: } kvn@509: } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) { kvn@509: mem = mem->in(0)->in(TypeFunc::Memory); kvn@509: continue; // (a) advance through independent MemBar memory kvn@509: } else if (mem->is_MergeMem()) { kvn@509: int alias_idx = phase->C->get_alias_index(adr_type()); kvn@509: mem = mem->as_MergeMem()->memory_at(alias_idx); kvn@509: continue; // (a) advance through independent MergeMem memory kvn@509: } duke@435: } duke@435: duke@435: // Unless there is an explicit 'continue', we must bail out here, duke@435: // because 'mem' is an inscrutable memory state (e.g., a call). duke@435: break; duke@435: } duke@435: duke@435: return NULL; // bail out duke@435: } duke@435: duke@435: //----------------------calculate_adr_type------------------------------------- duke@435: // Helper function. Notices when the given type of address hits top or bottom. duke@435: // Also, asserts a cross-check of the type against the expected address type. duke@435: const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) { duke@435: if (t == Type::TOP) return NULL; // does not touch memory any more? duke@435: #ifdef PRODUCT duke@435: cross_check = NULL; duke@435: #else duke@435: if (!VerifyAliases || is_error_reported() || Node::in_dump()) cross_check = NULL; duke@435: #endif duke@435: const TypePtr* tp = t->isa_ptr(); duke@435: if (tp == NULL) { duke@435: assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide"); duke@435: return TypePtr::BOTTOM; // touches lots of memory duke@435: } else { duke@435: #ifdef ASSERT duke@435: // %%%% [phh] We don't check the alias index if cross_check is duke@435: // TypeRawPtr::BOTTOM. Needs to be investigated. duke@435: if (cross_check != NULL && duke@435: cross_check != TypePtr::BOTTOM && duke@435: cross_check != TypeRawPtr::BOTTOM) { duke@435: // Recheck the alias index, to see if it has changed (due to a bug). duke@435: Compile* C = Compile::current(); duke@435: assert(C->get_alias_index(cross_check) == C->get_alias_index(tp), duke@435: "must stay in the original alias category"); duke@435: // The type of the address must be contained in the adr_type, duke@435: // disregarding "null"-ness. duke@435: // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.) duke@435: const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr(); duke@435: assert(cross_check->meet(tp_notnull) == cross_check, duke@435: "real address must not escape from expected memory type"); duke@435: } duke@435: #endif duke@435: return tp; duke@435: } duke@435: } duke@435: duke@435: //------------------------adr_phi_is_loop_invariant---------------------------- duke@435: // A helper function for Ideal_DU_postCCP to check if a Phi in a counted duke@435: // loop is loop invariant. Make a quick traversal of Phi and associated duke@435: // CastPP nodes, looking to see if they are a closed group within the loop. duke@435: bool MemNode::adr_phi_is_loop_invariant(Node* adr_phi, Node* cast) { duke@435: // The idea is that the phi-nest must boil down to only CastPP nodes duke@435: // with the same data. This implies that any path into the loop already duke@435: // includes such a CastPP, and so the original cast, whatever its input, duke@435: // must be covered by an equivalent cast, with an earlier control input. duke@435: ResourceMark rm; duke@435: duke@435: // The loop entry input of the phi should be the unique dominating duke@435: // node for every Phi/CastPP in the loop. duke@435: Unique_Node_List closure; duke@435: closure.push(adr_phi->in(LoopNode::EntryControl)); duke@435: duke@435: // Add the phi node and the cast to the worklist. duke@435: Unique_Node_List worklist; duke@435: worklist.push(adr_phi); duke@435: if( cast != NULL ){ duke@435: if( !cast->is_ConstraintCast() ) return false; duke@435: worklist.push(cast); duke@435: } duke@435: duke@435: // Begin recursive walk of phi nodes. duke@435: while( worklist.size() ){ duke@435: // Take a node off the worklist duke@435: Node *n = worklist.pop(); duke@435: if( !closure.member(n) ){ duke@435: // Add it to the closure. duke@435: closure.push(n); duke@435: // Make a sanity check to ensure we don't waste too much time here. duke@435: if( closure.size() > 20) return false; duke@435: // This node is OK if: duke@435: // - it is a cast of an identical value duke@435: // - or it is a phi node (then we add its inputs to the worklist) duke@435: // Otherwise, the node is not OK, and we presume the cast is not invariant duke@435: if( n->is_ConstraintCast() ){ duke@435: worklist.push(n->in(1)); duke@435: } else if( n->is_Phi() ) { duke@435: for( uint i = 1; i < n->req(); i++ ) { duke@435: worklist.push(n->in(i)); duke@435: } duke@435: } else { duke@435: return false; duke@435: } duke@435: } duke@435: } duke@435: duke@435: // Quit when the worklist is empty, and we've found no offending nodes. duke@435: return true; duke@435: } duke@435: duke@435: //------------------------------Ideal_DU_postCCP------------------------------- duke@435: // Find any cast-away of null-ness and keep its control. Null cast-aways are duke@435: // going away in this pass and we need to make this memory op depend on the duke@435: // gating null check. kvn@598: Node *MemNode::Ideal_DU_postCCP( PhaseCCP *ccp ) { kvn@598: return Ideal_common_DU_postCCP(ccp, this, in(MemNode::Address)); kvn@598: } duke@435: duke@435: // I tried to leave the CastPP's in. This makes the graph more accurate in duke@435: // some sense; we get to keep around the knowledge that an oop is not-null duke@435: // after some test. Alas, the CastPP's interfere with GVN (some values are duke@435: // the regular oop, some are the CastPP of the oop, all merge at Phi's which duke@435: // cannot collapse, etc). This cost us 10% on SpecJVM, even when I removed duke@435: // some of the more trivial cases in the optimizer. Removing more useless duke@435: // Phi's started allowing Loads to illegally float above null checks. I gave duke@435: // up on this approach. CNC 10/20/2000 kvn@598: // This static method may be called not from MemNode (EncodePNode calls it). kvn@598: // Only the control edge of the node 'n' might be updated. kvn@598: Node *MemNode::Ideal_common_DU_postCCP( PhaseCCP *ccp, Node* n, Node* adr ) { duke@435: Node *skipped_cast = NULL; duke@435: // Need a null check? Regular static accesses do not because they are duke@435: // from constant addresses. Array ops are gated by the range check (which duke@435: // always includes a NULL check). Just check field ops. kvn@598: if( n->in(MemNode::Control) == NULL ) { duke@435: // Scan upwards for the highest location we can place this memory op. duke@435: while( true ) { duke@435: switch( adr->Opcode() ) { duke@435: duke@435: case Op_AddP: // No change to NULL-ness, so peek thru AddP's duke@435: adr = adr->in(AddPNode::Base); duke@435: continue; duke@435: coleenp@548: case Op_DecodeN: // No change to NULL-ness, so peek thru coleenp@548: adr = adr->in(1); coleenp@548: continue; coleenp@548: duke@435: case Op_CastPP: duke@435: // If the CastPP is useless, just peek on through it. duke@435: if( ccp->type(adr) == ccp->type(adr->in(1)) ) { duke@435: // Remember the cast that we've peeked though. If we peek duke@435: // through more than one, then we end up remembering the highest duke@435: // one, that is, if in a loop, the one closest to the top. duke@435: skipped_cast = adr; duke@435: adr = adr->in(1); duke@435: continue; duke@435: } duke@435: // CastPP is going away in this pass! We need this memory op to be duke@435: // control-dependent on the test that is guarding the CastPP. kvn@598: ccp->hash_delete(n); kvn@598: n->set_req(MemNode::Control, adr->in(0)); kvn@598: ccp->hash_insert(n); kvn@598: return n; duke@435: duke@435: case Op_Phi: duke@435: // Attempt to float above a Phi to some dominating point. duke@435: if (adr->in(0) != NULL && adr->in(0)->is_CountedLoop()) { duke@435: // If we've already peeked through a Cast (which could have set the duke@435: // control), we can't float above a Phi, because the skipped Cast duke@435: // may not be loop invariant. duke@435: if (adr_phi_is_loop_invariant(adr, skipped_cast)) { duke@435: adr = adr->in(1); duke@435: continue; duke@435: } duke@435: } duke@435: duke@435: // Intentional fallthrough! duke@435: duke@435: // No obvious dominating point. The mem op is pinned below the Phi duke@435: // by the Phi itself. If the Phi goes away (no true value is merged) duke@435: // then the mem op can float, but not indefinitely. It must be pinned duke@435: // behind the controls leading to the Phi. duke@435: case Op_CheckCastPP: duke@435: // These usually stick around to change address type, however a duke@435: // useless one can be elided and we still need to pick up a control edge duke@435: if (adr->in(0) == NULL) { duke@435: // This CheckCastPP node has NO control and is likely useless. But we duke@435: // need check further up the ancestor chain for a control input to keep duke@435: // the node in place. 4959717. duke@435: skipped_cast = adr; duke@435: adr = adr->in(1); duke@435: continue; duke@435: } kvn@598: ccp->hash_delete(n); kvn@598: n->set_req(MemNode::Control, adr->in(0)); kvn@598: ccp->hash_insert(n); kvn@598: return n; duke@435: duke@435: // List of "safe" opcodes; those that implicitly block the memory duke@435: // op below any null check. duke@435: case Op_CastX2P: // no null checks on native pointers duke@435: case Op_Parm: // 'this' pointer is not null duke@435: case Op_LoadP: // Loading from within a klass coleenp@548: case Op_LoadN: // Loading from within a klass duke@435: case Op_LoadKlass: // Loading from within a klass kvn@599: case Op_LoadNKlass: // Loading from within a klass duke@435: case Op_ConP: // Loading from a klass kvn@598: case Op_ConN: // Loading from a klass duke@435: case Op_CreateEx: // Sucking up the guts of an exception oop duke@435: case Op_Con: // Reading from TLS duke@435: case Op_CMoveP: // CMoveP is pinned kvn@599: case Op_CMoveN: // CMoveN is pinned duke@435: break; // No progress duke@435: duke@435: case Op_Proj: // Direct call to an allocation routine duke@435: case Op_SCMemProj: // Memory state from store conditional ops duke@435: #ifdef ASSERT duke@435: { duke@435: assert(adr->as_Proj()->_con == TypeFunc::Parms, "must be return value"); duke@435: const Node* call = adr->in(0); kvn@598: if (call->is_CallJava()) { kvn@598: const CallJavaNode* call_java = call->as_CallJava(); kvn@499: const TypeTuple *r = call_java->tf()->range(); kvn@499: assert(r->cnt() > TypeFunc::Parms, "must return value"); kvn@499: const Type* ret_type = r->field_at(TypeFunc::Parms); kvn@499: assert(ret_type && ret_type->isa_ptr(), "must return pointer"); duke@435: // We further presume that this is one of duke@435: // new_instance_Java, new_array_Java, or duke@435: // the like, but do not assert for this. duke@435: } else if (call->is_Allocate()) { duke@435: // similar case to new_instance_Java, etc. duke@435: } else if (!call->is_CallLeaf()) { duke@435: // Projections from fetch_oop (OSR) are allowed as well. duke@435: ShouldNotReachHere(); duke@435: } duke@435: } duke@435: #endif duke@435: break; duke@435: default: duke@435: ShouldNotReachHere(); duke@435: } duke@435: break; duke@435: } duke@435: } duke@435: duke@435: return NULL; // No progress duke@435: } duke@435: duke@435: duke@435: //============================================================================= duke@435: uint LoadNode::size_of() const { return sizeof(*this); } duke@435: uint LoadNode::cmp( const Node &n ) const duke@435: { return !Type::cmp( _type, ((LoadNode&)n)._type ); } duke@435: const Type *LoadNode::bottom_type() const { return _type; } duke@435: uint LoadNode::ideal_reg() const { duke@435: return Matcher::base2reg[_type->base()]; duke@435: } duke@435: duke@435: #ifndef PRODUCT duke@435: void LoadNode::dump_spec(outputStream *st) const { duke@435: MemNode::dump_spec(st); duke@435: if( !Verbose && !WizardMode ) { duke@435: // standard dump does this in Verbose and WizardMode duke@435: st->print(" #"); _type->dump_on(st); duke@435: } duke@435: } duke@435: #endif duke@435: duke@435: duke@435: //----------------------------LoadNode::make----------------------------------- duke@435: // Polymorphic factory method: coleenp@548: Node *LoadNode::make( PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt ) { coleenp@548: Compile* C = gvn.C; coleenp@548: duke@435: // sanity check the alias category against the created node type duke@435: assert(!(adr_type->isa_oopptr() && duke@435: adr_type->offset() == oopDesc::klass_offset_in_bytes()), duke@435: "use LoadKlassNode instead"); duke@435: assert(!(adr_type->isa_aryptr() && duke@435: adr_type->offset() == arrayOopDesc::length_offset_in_bytes()), duke@435: "use LoadRangeNode instead"); duke@435: switch (bt) { duke@435: case T_BOOLEAN: twisti@993: case T_BYTE: return new (C, 3) LoadBNode (ctl, mem, adr, adr_type, rt->is_int() ); twisti@993: case T_INT: return new (C, 3) LoadINode (ctl, mem, adr, adr_type, rt->is_int() ); twisti@993: case T_CHAR: return new (C, 3) LoadUSNode(ctl, mem, adr, adr_type, rt->is_int() ); twisti@993: case T_SHORT: return new (C, 3) LoadSNode (ctl, mem, adr, adr_type, rt->is_int() ); twisti@993: case T_LONG: return new (C, 3) LoadLNode (ctl, mem, adr, adr_type, rt->is_long() ); twisti@993: case T_FLOAT: return new (C, 3) LoadFNode (ctl, mem, adr, adr_type, rt ); twisti@993: case T_DOUBLE: return new (C, 3) LoadDNode (ctl, mem, adr, adr_type, rt ); twisti@993: case T_ADDRESS: return new (C, 3) LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr() ); coleenp@548: case T_OBJECT: coleenp@548: #ifdef _LP64 kvn@598: if (adr->bottom_type()->is_ptr_to_narrowoop()) { kvn@656: Node* load = gvn.transform(new (C, 3) LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop())); kvn@656: return new (C, 2) DecodeNNode(load, load->bottom_type()->make_ptr()); coleenp@548: } else coleenp@548: #endif kvn@598: { kvn@598: assert(!adr->bottom_type()->is_ptr_to_narrowoop(), "should have got back a narrow oop"); kvn@598: return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr()); kvn@598: } duke@435: } duke@435: ShouldNotReachHere(); duke@435: return (LoadNode*)NULL; duke@435: } duke@435: duke@435: LoadLNode* LoadLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt) { duke@435: bool require_atomic = true; duke@435: return new (C, 3) LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), require_atomic); duke@435: } duke@435: duke@435: duke@435: duke@435: duke@435: //------------------------------hash------------------------------------------- duke@435: uint LoadNode::hash() const { duke@435: // unroll addition of interesting fields duke@435: return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address); duke@435: } duke@435: duke@435: //---------------------------can_see_stored_value------------------------------ duke@435: // This routine exists to make sure this set of tests is done the same duke@435: // everywhere. We need to make a coordinated change: first LoadNode::Ideal duke@435: // will change the graph shape in a way which makes memory alive twice at the duke@435: // same time (uses the Oracle model of aliasing), then some duke@435: // LoadXNode::Identity will fold things back to the equivalence-class model duke@435: // of aliasing. duke@435: Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const { duke@435: Node* ld_adr = in(MemNode::Address); duke@435: never@452: const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr(); never@452: Compile::AliasType* atp = tp != NULL ? phase->C->alias_type(tp) : NULL; never@452: if (EliminateAutoBox && atp != NULL && atp->index() >= Compile::AliasIdxRaw && never@452: atp->field() != NULL && !atp->field()->is_volatile()) { never@452: uint alias_idx = atp->index(); never@452: bool final = atp->field()->is_final(); never@452: Node* result = NULL; never@452: Node* current = st; never@452: // Skip through chains of MemBarNodes checking the MergeMems for never@452: // new states for the slice of this load. Stop once any other never@452: // kind of node is encountered. Loads from final memory can skip never@452: // through any kind of MemBar but normal loads shouldn't skip never@452: // through MemBarAcquire since the could allow them to move out of never@452: // a synchronized region. never@452: while (current->is_Proj()) { never@452: int opc = current->in(0)->Opcode(); never@452: if ((final && opc == Op_MemBarAcquire) || never@452: opc == Op_MemBarRelease || opc == Op_MemBarCPUOrder) { never@452: Node* mem = current->in(0)->in(TypeFunc::Memory); never@452: if (mem->is_MergeMem()) { never@452: MergeMemNode* merge = mem->as_MergeMem(); never@452: Node* new_st = merge->memory_at(alias_idx); never@452: if (new_st == merge->base_memory()) { never@452: // Keep searching never@452: current = merge->base_memory(); never@452: continue; never@452: } never@452: // Save the new memory state for the slice and fall through never@452: // to exit. never@452: result = new_st; never@452: } never@452: } never@452: break; never@452: } never@452: if (result != NULL) { never@452: st = result; never@452: } never@452: } never@452: never@452: duke@435: // Loop around twice in the case Load -> Initialize -> Store. duke@435: // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.) duke@435: for (int trip = 0; trip <= 1; trip++) { duke@435: duke@435: if (st->is_Store()) { duke@435: Node* st_adr = st->in(MemNode::Address); duke@435: if (!phase->eqv(st_adr, ld_adr)) { duke@435: // Try harder before giving up... Match raw and non-raw pointers. duke@435: intptr_t st_off = 0; duke@435: AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off); duke@435: if (alloc == NULL) return NULL; duke@435: intptr_t ld_off = 0; duke@435: AllocateNode* allo2 = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off); duke@435: if (alloc != allo2) return NULL; duke@435: if (ld_off != st_off) return NULL; duke@435: // At this point we have proven something like this setup: duke@435: // A = Allocate(...) duke@435: // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off)) duke@435: // S = StoreQ(, AddP(, A.Parm , #Off), V) duke@435: // (Actually, we haven't yet proven the Q's are the same.) duke@435: // In other words, we are loading from a casted version of duke@435: // the same pointer-and-offset that we stored to. duke@435: // Thus, we are able to replace L by V. duke@435: } duke@435: // Now prove that we have a LoadQ matched to a StoreQ, for some Q. duke@435: if (store_Opcode() != st->Opcode()) duke@435: return NULL; duke@435: return st->in(MemNode::ValueIn); duke@435: } duke@435: duke@435: intptr_t offset = 0; // scratch duke@435: duke@435: // A load from a freshly-created object always returns zero. duke@435: // (This can happen after LoadNode::Ideal resets the load's memory input duke@435: // to find_captured_store, which returned InitializeNode::zero_memory.) duke@435: if (st->is_Proj() && st->in(0)->is_Allocate() && duke@435: st->in(0) == AllocateNode::Ideal_allocation(ld_adr, phase, offset) && duke@435: offset >= st->in(0)->as_Allocate()->minimum_header_size()) { duke@435: // return a zero value for the load's basic type duke@435: // (This is one of the few places where a generic PhaseTransform duke@435: // can create new nodes. Think of it as lazily manifesting duke@435: // virtually pre-existing constants.) duke@435: return phase->zerocon(memory_type()); duke@435: } duke@435: duke@435: // A load from an initialization barrier can match a captured store. duke@435: if (st->is_Proj() && st->in(0)->is_Initialize()) { duke@435: InitializeNode* init = st->in(0)->as_Initialize(); duke@435: AllocateNode* alloc = init->allocation(); duke@435: if (alloc != NULL && duke@435: alloc == AllocateNode::Ideal_allocation(ld_adr, phase, offset)) { duke@435: // examine a captured store value duke@435: st = init->find_captured_store(offset, memory_size(), phase); duke@435: if (st != NULL) duke@435: continue; // take one more trip around duke@435: } duke@435: } duke@435: duke@435: break; duke@435: } duke@435: duke@435: return NULL; duke@435: } duke@435: kvn@499: //----------------------is_instance_field_load_with_local_phi------------------ kvn@499: bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) { kvn@499: if( in(MemNode::Memory)->is_Phi() && in(MemNode::Memory)->in(0) == ctrl && kvn@499: in(MemNode::Address)->is_AddP() ) { kvn@499: const TypeOopPtr* t_oop = in(MemNode::Address)->bottom_type()->isa_oopptr(); kvn@499: // Only instances. kvn@658: if( t_oop != NULL && t_oop->is_known_instance_field() && kvn@499: t_oop->offset() != Type::OffsetBot && kvn@499: t_oop->offset() != Type::OffsetTop) { kvn@499: return true; kvn@499: } kvn@499: } kvn@499: return false; kvn@499: } kvn@499: duke@435: //------------------------------Identity--------------------------------------- duke@435: // Loads are identity if previous store is to same address duke@435: Node *LoadNode::Identity( PhaseTransform *phase ) { duke@435: // If the previous store-maker is the right kind of Store, and the store is duke@435: // to the same address, then we are equal to the value stored. duke@435: Node* mem = in(MemNode::Memory); duke@435: Node* value = can_see_stored_value(mem, phase); duke@435: if( value ) { duke@435: // byte, short & char stores truncate naturally. duke@435: // A load has to load the truncated value which requires duke@435: // some sort of masking operation and that requires an duke@435: // Ideal call instead of an Identity call. duke@435: if (memory_size() < BytesPerInt) { duke@435: // If the input to the store does not fit with the load's result type, duke@435: // it must be truncated via an Ideal call. duke@435: if (!phase->type(value)->higher_equal(phase->type(this))) duke@435: return this; duke@435: } duke@435: // (This works even when value is a Con, but LoadNode::Value duke@435: // usually runs first, producing the singleton type of the Con.) duke@435: return value; duke@435: } kvn@499: kvn@499: // Search for an existing data phi which was generated before for the same twisti@1040: // instance's field to avoid infinite generation of phis in a loop. kvn@499: Node *region = mem->in(0); kvn@499: if (is_instance_field_load_with_local_phi(region)) { kvn@499: const TypePtr *addr_t = in(MemNode::Address)->bottom_type()->isa_ptr(); kvn@499: int this_index = phase->C->get_alias_index(addr_t); kvn@499: int this_offset = addr_t->offset(); kvn@499: int this_id = addr_t->is_oopptr()->instance_id(); kvn@499: const Type* this_type = bottom_type(); kvn@499: for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { kvn@499: Node* phi = region->fast_out(i); kvn@499: if (phi->is_Phi() && phi != mem && kvn@499: phi->as_Phi()->is_same_inst_field(this_type, this_id, this_index, this_offset)) { kvn@499: return phi; kvn@499: } kvn@499: } kvn@499: } kvn@499: duke@435: return this; duke@435: } duke@435: never@452: never@452: // Returns true if the AliasType refers to the field that holds the never@452: // cached box array. Currently only handles the IntegerCache case. never@452: static bool is_autobox_cache(Compile::AliasType* atp) { never@452: if (atp != NULL && atp->field() != NULL) { never@452: ciField* field = atp->field(); never@452: ciSymbol* klass = field->holder()->name(); never@452: if (field->name() == ciSymbol::cache_field_name() && never@452: field->holder()->uses_default_loader() && never@452: klass == ciSymbol::java_lang_Integer_IntegerCache()) { never@452: return true; never@452: } never@452: } never@452: return false; never@452: } never@452: never@452: // Fetch the base value in the autobox array never@452: static bool fetch_autobox_base(Compile::AliasType* atp, int& cache_offset) { never@452: if (atp != NULL && atp->field() != NULL) { never@452: ciField* field = atp->field(); never@452: ciSymbol* klass = field->holder()->name(); never@452: if (field->name() == ciSymbol::cache_field_name() && never@452: field->holder()->uses_default_loader() && never@452: klass == ciSymbol::java_lang_Integer_IntegerCache()) { never@452: assert(field->is_constant(), "what?"); never@452: ciObjArray* array = field->constant_value().as_object()->as_obj_array(); never@452: // Fetch the box object at the base of the array and get its value never@452: ciInstance* box = array->obj_at(0)->as_instance(); never@452: ciInstanceKlass* ik = box->klass()->as_instance_klass(); never@452: if (ik->nof_nonstatic_fields() == 1) { never@452: // This should be true nonstatic_field_at requires calling never@452: // nof_nonstatic_fields so check it anyway never@452: ciConstant c = box->field_value(ik->nonstatic_field_at(0)); never@452: cache_offset = c.as_int(); never@452: } never@452: return true; never@452: } never@452: } never@452: return false; never@452: } never@452: never@452: // Returns true if the AliasType refers to the value field of an never@452: // autobox object. Currently only handles Integer. never@452: static bool is_autobox_object(Compile::AliasType* atp) { never@452: if (atp != NULL && atp->field() != NULL) { never@452: ciField* field = atp->field(); never@452: ciSymbol* klass = field->holder()->name(); never@452: if (field->name() == ciSymbol::value_name() && never@452: field->holder()->uses_default_loader() && never@452: klass == ciSymbol::java_lang_Integer()) { never@452: return true; never@452: } never@452: } never@452: return false; never@452: } never@452: never@452: never@452: // We're loading from an object which has autobox behaviour. never@452: // If this object is result of a valueOf call we'll have a phi never@452: // merging a newly allocated object and a load from the cache. never@452: // We want to replace this load with the original incoming never@452: // argument to the valueOf call. never@452: Node* LoadNode::eliminate_autobox(PhaseGVN* phase) { never@452: Node* base = in(Address)->in(AddPNode::Base); never@452: if (base->is_Phi() && base->req() == 3) { never@452: AllocateNode* allocation = NULL; never@452: int allocation_index = -1; never@452: int load_index = -1; never@452: for (uint i = 1; i < base->req(); i++) { never@452: allocation = AllocateNode::Ideal_allocation(base->in(i), phase); never@452: if (allocation != NULL) { never@452: allocation_index = i; never@452: load_index = 3 - allocation_index; never@452: break; never@452: } never@452: } kvn@1018: bool has_load = ( allocation != NULL && kvn@1018: (base->in(load_index)->is_Load() || kvn@1018: base->in(load_index)->is_DecodeN() && kvn@1018: base->in(load_index)->in(1)->is_Load()) ); kvn@1018: if (has_load && in(Memory)->is_Phi() && in(Memory)->in(0) == base->in(0)) { never@452: // Push the loads from the phi that comes from valueOf up never@452: // through it to allow elimination of the loads and the recovery never@452: // of the original value. never@452: Node* mem_phi = in(Memory); never@452: Node* offset = in(Address)->in(AddPNode::Offset); never@988: Node* region = base->in(0); never@452: never@452: Node* in1 = clone(); never@452: Node* in1_addr = in1->in(Address)->clone(); never@452: in1_addr->set_req(AddPNode::Base, base->in(allocation_index)); never@452: in1_addr->set_req(AddPNode::Address, base->in(allocation_index)); never@452: in1_addr->set_req(AddPNode::Offset, offset); never@988: in1->set_req(0, region->in(allocation_index)); never@452: in1->set_req(Address, in1_addr); never@452: in1->set_req(Memory, mem_phi->in(allocation_index)); never@452: never@452: Node* in2 = clone(); never@452: Node* in2_addr = in2->in(Address)->clone(); never@452: in2_addr->set_req(AddPNode::Base, base->in(load_index)); never@452: in2_addr->set_req(AddPNode::Address, base->in(load_index)); never@452: in2_addr->set_req(AddPNode::Offset, offset); never@988: in2->set_req(0, region->in(load_index)); never@452: in2->set_req(Address, in2_addr); never@452: in2->set_req(Memory, mem_phi->in(load_index)); never@452: never@452: in1_addr = phase->transform(in1_addr); never@452: in1 = phase->transform(in1); never@452: in2_addr = phase->transform(in2_addr); never@452: in2 = phase->transform(in2); never@452: never@988: PhiNode* result = PhiNode::make_blank(region, this); never@452: result->set_req(allocation_index, in1); never@452: result->set_req(load_index, in2); never@452: return result; never@452: } kvn@1018: } else if (base->is_Load() || kvn@1018: base->is_DecodeN() && base->in(1)->is_Load()) { kvn@1018: if (base->is_DecodeN()) { kvn@1018: // Get LoadN node which loads cached Integer object kvn@1018: base = base->in(1); kvn@1018: } never@452: // Eliminate the load of Integer.value for integers from the cache never@452: // array by deriving the value from the index into the array. never@452: // Capture the offset of the load and then reverse the computation. never@452: Node* load_base = base->in(Address)->in(AddPNode::Base); kvn@1018: if (load_base->is_DecodeN()) { kvn@1018: // Get LoadN node which loads IntegerCache.cache field kvn@1018: load_base = load_base->in(1); kvn@1018: } never@452: if (load_base != NULL) { never@452: Compile::AliasType* atp = phase->C->alias_type(load_base->adr_type()); never@452: intptr_t cache_offset; never@452: int shift = -1; never@452: Node* cache = NULL; never@452: if (is_autobox_cache(atp)) { kvn@464: shift = exact_log2(type2aelembytes(T_OBJECT)); never@452: cache = AddPNode::Ideal_base_and_offset(load_base->in(Address), phase, cache_offset); never@452: } never@452: if (cache != NULL && base->in(Address)->is_AddP()) { never@452: Node* elements[4]; never@452: int count = base->in(Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements)); never@452: int cache_low; never@452: if (count > 0 && fetch_autobox_base(atp, cache_low)) { never@452: int offset = arrayOopDesc::base_offset_in_bytes(memory_type()) - (cache_low << shift); never@452: // Add up all the offsets making of the address of the load never@452: Node* result = elements[0]; never@452: for (int i = 1; i < count; i++) { never@452: result = phase->transform(new (phase->C, 3) AddXNode(result, elements[i])); never@452: } never@452: // Remove the constant offset from the address and then never@452: // remove the scaling of the offset to recover the original index. never@452: result = phase->transform(new (phase->C, 3) AddXNode(result, phase->MakeConX(-offset))); never@452: if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) { never@452: // Peel the shift off directly but wrap it in a dummy node never@452: // since Ideal can't return existing nodes never@452: result = new (phase->C, 3) RShiftXNode(result->in(1), phase->intcon(0)); never@452: } else { never@452: result = new (phase->C, 3) RShiftXNode(result, phase->intcon(shift)); never@452: } never@452: #ifdef _LP64 never@452: result = new (phase->C, 2) ConvL2INode(phase->transform(result)); never@452: #endif never@452: return result; never@452: } never@452: } never@452: } never@452: } never@452: return NULL; never@452: } never@452: kvn@598: //------------------------------split_through_phi------------------------------ kvn@598: // Split instance field load through Phi. kvn@598: Node *LoadNode::split_through_phi(PhaseGVN *phase) { kvn@598: Node* mem = in(MemNode::Memory); kvn@598: Node* address = in(MemNode::Address); kvn@598: const TypePtr *addr_t = phase->type(address)->isa_ptr(); kvn@598: const TypeOopPtr *t_oop = addr_t->isa_oopptr(); kvn@598: kvn@598: assert(mem->is_Phi() && (t_oop != NULL) && kvn@658: t_oop->is_known_instance_field(), "invalide conditions"); kvn@598: kvn@598: Node *region = mem->in(0); kvn@598: if (region == NULL) { kvn@598: return NULL; // Wait stable graph kvn@598: } kvn@598: uint cnt = mem->req(); kvn@598: for( uint i = 1; i < cnt; i++ ) { kvn@598: Node *in = mem->in(i); kvn@598: if( in == NULL ) { kvn@598: return NULL; // Wait stable graph kvn@598: } kvn@598: } kvn@598: // Check for loop invariant. kvn@598: if (cnt == 3) { kvn@598: for( uint i = 1; i < cnt; i++ ) { kvn@598: Node *in = mem->in(i); kvn@598: Node* m = MemNode::optimize_memory_chain(in, addr_t, phase); kvn@598: if (m == mem) { kvn@598: set_req(MemNode::Memory, mem->in(cnt - i)); // Skip this phi. kvn@598: return this; kvn@598: } kvn@598: } kvn@598: } kvn@598: // Split through Phi (see original code in loopopts.cpp). kvn@598: assert(phase->C->have_alias_type(addr_t), "instance should have alias type"); kvn@598: kvn@598: // Do nothing here if Identity will find a value kvn@598: // (to avoid infinite chain of value phis generation). kvn@598: if ( !phase->eqv(this, this->Identity(phase)) ) kvn@598: return NULL; kvn@598: kvn@598: // Skip the split if the region dominates some control edge of the address. kvn@598: if (cnt == 3 && !MemNode::all_controls_dominate(address, region)) kvn@598: return NULL; kvn@598: kvn@598: const Type* this_type = this->bottom_type(); kvn@598: int this_index = phase->C->get_alias_index(addr_t); kvn@598: int this_offset = addr_t->offset(); kvn@598: int this_iid = addr_t->is_oopptr()->instance_id(); kvn@598: int wins = 0; kvn@598: PhaseIterGVN *igvn = phase->is_IterGVN(); kvn@598: Node *phi = new (igvn->C, region->req()) PhiNode(region, this_type, NULL, this_iid, this_index, this_offset); kvn@598: for( uint i = 1; i < region->req(); i++ ) { kvn@598: Node *x; kvn@598: Node* the_clone = NULL; kvn@598: if( region->in(i) == phase->C->top() ) { kvn@598: x = phase->C->top(); // Dead path? Use a dead data op kvn@598: } else { kvn@598: x = this->clone(); // Else clone up the data op kvn@598: the_clone = x; // Remember for possible deletion. kvn@598: // Alter data node to use pre-phi inputs kvn@598: if( this->in(0) == region ) { kvn@598: x->set_req( 0, region->in(i) ); kvn@598: } else { kvn@598: x->set_req( 0, NULL ); kvn@598: } kvn@598: for( uint j = 1; j < this->req(); j++ ) { kvn@598: Node *in = this->in(j); kvn@598: if( in->is_Phi() && in->in(0) == region ) kvn@598: x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone kvn@598: } kvn@598: } kvn@598: // Check for a 'win' on some paths kvn@598: const Type *t = x->Value(igvn); kvn@598: kvn@598: bool singleton = t->singleton(); kvn@598: kvn@598: // See comments in PhaseIdealLoop::split_thru_phi(). kvn@598: if( singleton && t == Type::TOP ) { kvn@598: singleton &= region->is_Loop() && (i != LoopNode::EntryControl); kvn@598: } kvn@598: kvn@598: if( singleton ) { kvn@598: wins++; kvn@598: x = igvn->makecon(t); kvn@598: } else { kvn@598: // We now call Identity to try to simplify the cloned node. kvn@598: // Note that some Identity methods call phase->type(this). kvn@598: // Make sure that the type array is big enough for kvn@598: // our new node, even though we may throw the node away. kvn@598: // (This tweaking with igvn only works because x is a new node.) kvn@598: igvn->set_type(x, t); kvn@728: // If x is a TypeNode, capture any more-precise type permanently into Node twisti@1040: // otherwise it will be not updated during igvn->transform since kvn@728: // igvn->type(x) is set to x->Value() already. kvn@728: x->raise_bottom_type(t); kvn@598: Node *y = x->Identity(igvn); kvn@598: if( y != x ) { kvn@598: wins++; kvn@598: x = y; kvn@598: } else { kvn@598: y = igvn->hash_find(x); kvn@598: if( y ) { kvn@598: wins++; kvn@598: x = y; kvn@598: } else { kvn@598: // Else x is a new node we are keeping kvn@598: // We do not need register_new_node_with_optimizer kvn@598: // because set_type has already been called. kvn@598: igvn->_worklist.push(x); kvn@598: } kvn@598: } kvn@598: } kvn@598: if (x != the_clone && the_clone != NULL) kvn@598: igvn->remove_dead_node(the_clone); kvn@598: phi->set_req(i, x); kvn@598: } kvn@598: if( wins > 0 ) { kvn@598: // Record Phi kvn@598: igvn->register_new_node_with_optimizer(phi); kvn@598: return phi; kvn@598: } kvn@598: igvn->remove_dead_node(phi); kvn@598: return NULL; kvn@598: } never@452: duke@435: //------------------------------Ideal------------------------------------------ duke@435: // If the load is from Field memory and the pointer is non-null, we can duke@435: // zero out the control input. duke@435: // If the offset is constant and the base is an object allocation, duke@435: // try to hook me up to the exact initializing store. duke@435: Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) { duke@435: Node* p = MemNode::Ideal_common(phase, can_reshape); duke@435: if (p) return (p == NodeSentinel) ? NULL : p; duke@435: duke@435: Node* ctrl = in(MemNode::Control); duke@435: Node* address = in(MemNode::Address); duke@435: duke@435: // Skip up past a SafePoint control. Cannot do this for Stores because duke@435: // pointer stores & cardmarks must stay on the same side of a SafePoint. duke@435: if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint && duke@435: phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) { duke@435: ctrl = ctrl->in(0); duke@435: set_req(MemNode::Control,ctrl); duke@435: } duke@435: duke@435: // Check for useless control edge in some common special cases duke@435: if (in(MemNode::Control) != NULL) { duke@435: intptr_t ignore = 0; duke@435: Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore); duke@435: if (base != NULL duke@435: && phase->type(base)->higher_equal(TypePtr::NOTNULL) never@979: && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw kvn@554: && all_controls_dominate(base, phase->C->start())) { duke@435: // A method-invariant, non-null address (constant or 'this' argument). duke@435: set_req(MemNode::Control, NULL); duke@435: } duke@435: } duke@435: never@452: if (EliminateAutoBox && can_reshape && in(Address)->is_AddP()) { never@452: Node* base = in(Address)->in(AddPNode::Base); never@452: if (base != NULL) { never@452: Compile::AliasType* atp = phase->C->alias_type(adr_type()); never@452: if (is_autobox_object(atp)) { never@452: Node* result = eliminate_autobox(phase); never@452: if (result != NULL) return result; never@452: } never@452: } never@452: } never@452: kvn@509: Node* mem = in(MemNode::Memory); kvn@509: const TypePtr *addr_t = phase->type(address)->isa_ptr(); kvn@509: kvn@509: if (addr_t != NULL) { kvn@509: // try to optimize our memory input kvn@509: Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, phase); kvn@509: if (opt_mem != mem) { kvn@509: set_req(MemNode::Memory, opt_mem); kvn@740: if (phase->type( opt_mem ) == Type::TOP) return NULL; kvn@509: return this; kvn@509: } kvn@509: const TypeOopPtr *t_oop = addr_t->isa_oopptr(); kvn@509: if (can_reshape && opt_mem->is_Phi() && kvn@658: (t_oop != NULL) && t_oop->is_known_instance_field()) { kvn@598: // Split instance field load through Phi. kvn@598: Node* result = split_through_phi(phase); kvn@598: if (result != NULL) return result; kvn@509: } kvn@509: } kvn@509: duke@435: // Check for prior store with a different base or offset; make Load duke@435: // independent. Skip through any number of them. Bail out if the stores duke@435: // are in an endless dead cycle and report no progress. This is a key duke@435: // transform for Reflection. However, if after skipping through the Stores duke@435: // we can't then fold up against a prior store do NOT do the transform as duke@435: // this amounts to using the 'Oracle' model of aliasing. It leaves the same duke@435: // array memory alive twice: once for the hoisted Load and again after the duke@435: // bypassed Store. This situation only works if EVERYBODY who does duke@435: // anti-dependence work knows how to bypass. I.e. we need all duke@435: // anti-dependence checks to ask the same Oracle. Right now, that Oracle is duke@435: // the alias index stuff. So instead, peek through Stores and IFF we can duke@435: // fold up, do so. duke@435: Node* prev_mem = find_previous_store(phase); duke@435: // Steps (a), (b): Walk past independent stores to find an exact match. duke@435: if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) { duke@435: // (c) See if we can fold up on the spot, but don't fold up here. twisti@993: // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or duke@435: // just return a prior value, which is done by Identity calls. duke@435: if (can_see_stored_value(prev_mem, phase)) { duke@435: // Make ready for step (d): duke@435: set_req(MemNode::Memory, prev_mem); duke@435: return this; duke@435: } duke@435: } duke@435: duke@435: return NULL; // No further progress duke@435: } duke@435: duke@435: // Helper to recognize certain Klass fields which are invariant across duke@435: // some group of array types (e.g., int[] or all T[] where T < Object). duke@435: const Type* duke@435: LoadNode::load_array_final_field(const TypeKlassPtr *tkls, duke@435: ciKlass* klass) const { duke@435: if (tkls->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc)) { duke@435: // The field is Klass::_modifier_flags. Return its (constant) value. duke@435: // (Folds up the 2nd indirection in aClassConstant.getModifiers().) duke@435: assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags"); duke@435: return TypeInt::make(klass->modifier_flags()); duke@435: } duke@435: if (tkls->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc)) { duke@435: // The field is Klass::_access_flags. Return its (constant) value. duke@435: // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).) duke@435: assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags"); duke@435: return TypeInt::make(klass->access_flags()); duke@435: } duke@435: if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc)) { duke@435: // The field is Klass::_layout_helper. Return its constant value if known. duke@435: assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper"); duke@435: return TypeInt::make(klass->layout_helper()); duke@435: } duke@435: duke@435: // No match. duke@435: return NULL; duke@435: } duke@435: duke@435: //------------------------------Value----------------------------------------- duke@435: const Type *LoadNode::Value( PhaseTransform *phase ) const { duke@435: // Either input is TOP ==> the result is TOP duke@435: Node* mem = in(MemNode::Memory); duke@435: const Type *t1 = phase->type(mem); duke@435: if (t1 == Type::TOP) return Type::TOP; duke@435: Node* adr = in(MemNode::Address); duke@435: const TypePtr* tp = phase->type(adr)->isa_ptr(); duke@435: if (tp == NULL || tp->empty()) return Type::TOP; duke@435: int off = tp->offset(); duke@435: assert(off != Type::OffsetTop, "case covered by TypePtr::empty"); duke@435: duke@435: // Try to guess loaded type from pointer type duke@435: if (tp->base() == Type::AryPtr) { duke@435: const Type *t = tp->is_aryptr()->elem(); duke@435: // Don't do this for integer types. There is only potential profit if duke@435: // the element type t is lower than _type; that is, for int types, if _type is duke@435: // more restrictive than t. This only happens here if one is short and the other duke@435: // char (both 16 bits), and in those cases we've made an intentional decision duke@435: // to use one kind of load over the other. See AndINode::Ideal and 4965907. duke@435: // Also, do not try to narrow the type for a LoadKlass, regardless of offset. duke@435: // duke@435: // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8)) duke@435: // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier duke@435: // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been duke@435: // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed, duke@435: // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any. duke@435: // In fact, that could have been the original type of p1, and p1 could have duke@435: // had an original form like p1:(AddP x x (LShiftL quux 3)), where the duke@435: // expression (LShiftL quux 3) independently optimized to the constant 8. duke@435: if ((t->isa_int() == NULL) && (t->isa_long() == NULL) kvn@728: && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) { duke@435: // t might actually be lower than _type, if _type is a unique duke@435: // concrete subclass of abstract class t. duke@435: // Make sure the reference is not into the header, by comparing duke@435: // the offset against the offset of the start of the array's data. duke@435: // Different array types begin at slightly different offsets (12 vs. 16). duke@435: // We choose T_BYTE as an example base type that is least restrictive duke@435: // as to alignment, which will therefore produce the smallest duke@435: // possible base offset. duke@435: const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE); duke@435: if ((uint)off >= (uint)min_base_off) { // is the offset beyond the header? duke@435: const Type* jt = t->join(_type); duke@435: // In any case, do not allow the join, per se, to empty out the type. duke@435: if (jt->empty() && !t->empty()) { duke@435: // This can happen if a interface-typed array narrows to a class type. duke@435: jt = _type; duke@435: } never@452: never@452: if (EliminateAutoBox) { never@452: // The pointers in the autobox arrays are always non-null never@452: Node* base = in(Address)->in(AddPNode::Base); never@452: if (base != NULL) { never@452: Compile::AliasType* atp = phase->C->alias_type(base->adr_type()); never@452: if (is_autobox_cache(atp)) { never@452: return jt->join(TypePtr::NOTNULL)->is_ptr(); never@452: } never@452: } never@452: } duke@435: return jt; duke@435: } duke@435: } duke@435: } else if (tp->base() == Type::InstPtr) { duke@435: assert( off != Type::OffsetBot || duke@435: // arrays can be cast to Objects duke@435: tp->is_oopptr()->klass()->is_java_lang_Object() || duke@435: // unsafe field access may not have a constant offset duke@435: phase->C->has_unsafe_access(), duke@435: "Field accesses must be precise" ); duke@435: // For oop loads, we expect the _type to be precise duke@435: } else if (tp->base() == Type::KlassPtr) { duke@435: assert( off != Type::OffsetBot || duke@435: // arrays can be cast to Objects duke@435: tp->is_klassptr()->klass()->is_java_lang_Object() || duke@435: // also allow array-loading from the primary supertype duke@435: // array during subtype checks duke@435: Opcode() == Op_LoadKlass, duke@435: "Field accesses must be precise" ); duke@435: // For klass/static loads, we expect the _type to be precise duke@435: } duke@435: duke@435: const TypeKlassPtr *tkls = tp->isa_klassptr(); duke@435: if (tkls != NULL && !StressReflectiveCode) { duke@435: ciKlass* klass = tkls->klass(); duke@435: if (klass->is_loaded() && tkls->klass_is_exact()) { duke@435: // We are loading a field from a Klass metaobject whose identity duke@435: // is known at compile time (the type is "exact" or "precise"). duke@435: // Check for fields we know are maintained as constants by the VM. duke@435: if (tkls->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc)) { duke@435: // The field is Klass::_super_check_offset. Return its (constant) value. duke@435: // (Folds up type checking code.) duke@435: assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset"); duke@435: return TypeInt::make(klass->super_check_offset()); duke@435: } duke@435: // Compute index into primary_supers array duke@435: juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop); duke@435: // Check for overflowing; use unsigned compare to handle the negative case. duke@435: if( depth < ciKlass::primary_super_limit() ) { duke@435: // The field is an element of Klass::_primary_supers. Return its (constant) value. duke@435: // (Folds up type checking code.) duke@435: assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); duke@435: ciKlass *ss = klass->super_of_depth(depth); duke@435: return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; duke@435: } duke@435: const Type* aift = load_array_final_field(tkls, klass); duke@435: if (aift != NULL) return aift; duke@435: if (tkls->offset() == in_bytes(arrayKlass::component_mirror_offset()) + (int)sizeof(oopDesc) duke@435: && klass->is_array_klass()) { duke@435: // The field is arrayKlass::_component_mirror. Return its (constant) value. duke@435: // (Folds up aClassConstant.getComponentType, common in Arrays.copyOf.) duke@435: assert(Opcode() == Op_LoadP, "must load an oop from _component_mirror"); duke@435: return TypeInstPtr::make(klass->as_array_klass()->component_mirror()); duke@435: } duke@435: if (tkls->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc)) { duke@435: // The field is Klass::_java_mirror. Return its (constant) value. duke@435: // (Folds up the 2nd indirection in anObjConstant.getClass().) duke@435: assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror"); duke@435: return TypeInstPtr::make(klass->java_mirror()); duke@435: } duke@435: } duke@435: duke@435: // We can still check if we are loading from the primary_supers array at a duke@435: // shallow enough depth. Even though the klass is not exact, entries less duke@435: // than or equal to its super depth are correct. duke@435: if (klass->is_loaded() ) { duke@435: ciType *inner = klass->klass(); duke@435: while( inner->is_obj_array_klass() ) duke@435: inner = inner->as_obj_array_klass()->base_element_type(); duke@435: if( inner->is_instance_klass() && duke@435: !inner->as_instance_klass()->flags().is_interface() ) { duke@435: // Compute index into primary_supers array duke@435: juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop); duke@435: // Check for overflowing; use unsigned compare to handle the negative case. duke@435: if( depth < ciKlass::primary_super_limit() && duke@435: depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case duke@435: // The field is an element of Klass::_primary_supers. Return its (constant) value. duke@435: // (Folds up type checking code.) duke@435: assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers"); duke@435: ciKlass *ss = klass->super_of_depth(depth); duke@435: return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR; duke@435: } duke@435: } duke@435: } duke@435: duke@435: // If the type is enough to determine that the thing is not an array, duke@435: // we can give the layout_helper a positive interval type. duke@435: // This will help short-circuit some reflective code. duke@435: if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc) duke@435: && !klass->is_array_klass() // not directly typed as an array duke@435: && !klass->is_interface() // specifically not Serializable & Cloneable duke@435: && !klass->is_java_lang_Object() // not the supertype of all T[] duke@435: ) { duke@435: // Note: When interfaces are reliable, we can narrow the interface duke@435: // test to (klass != Serializable && klass != Cloneable). duke@435: assert(Opcode() == Op_LoadI, "must load an int from _layout_helper"); duke@435: jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false); duke@435: // The key property of this type is that it folds up tests duke@435: // for array-ness, since it proves that the layout_helper is positive. duke@435: // Thus, a generic value like the basic object layout helper works fine. duke@435: return TypeInt::make(min_size, max_jint, Type::WidenMin); duke@435: } duke@435: } duke@435: duke@435: // If we are loading from a freshly-allocated object, produce a zero, duke@435: // if the load is provably beyond the header of the object. duke@435: // (Also allow a variable load from a fresh array to produce zero.) duke@435: if (ReduceFieldZeroing) { duke@435: Node* value = can_see_stored_value(mem,phase); duke@435: if (value != NULL && value->is_Con()) duke@435: return value->bottom_type(); duke@435: } duke@435: kvn@499: const TypeOopPtr *tinst = tp->isa_oopptr(); kvn@658: if (tinst != NULL && tinst->is_known_instance_field()) { kvn@499: // If we have an instance type and our memory input is the kvn@499: // programs's initial memory state, there is no matching store, kvn@499: // so just return a zero of the appropriate type kvn@499: Node *mem = in(MemNode::Memory); kvn@499: if (mem->is_Parm() && mem->in(0)->is_Start()) { kvn@499: assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm"); kvn@499: return Type::get_zero_type(_type->basic_type()); kvn@499: } kvn@499: } duke@435: return _type; duke@435: } duke@435: duke@435: //------------------------------match_edge------------------------------------- duke@435: // Do we Match on this edge index or not? Match only the address. duke@435: uint LoadNode::match_edge(uint idx) const { duke@435: return idx == MemNode::Address; duke@435: } duke@435: duke@435: //--------------------------LoadBNode::Ideal-------------------------------------- duke@435: // duke@435: // If the previous store is to the same address as this load, duke@435: // and the value stored was larger than a byte, replace this load duke@435: // with the value stored truncated to a byte. If no truncation is duke@435: // needed, the replacement is done in LoadNode::Identity(). duke@435: // duke@435: Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) { duke@435: Node* mem = in(MemNode::Memory); duke@435: Node* value = can_see_stored_value(mem,phase); duke@435: if( value && !phase->type(value)->higher_equal( _type ) ) { duke@435: Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(24)) ); duke@435: return new (phase->C, 3) RShiftINode(result, phase->intcon(24)); duke@435: } duke@435: // Identity call will handle the case where truncation is not needed. duke@435: return LoadNode::Ideal(phase, can_reshape); duke@435: } duke@435: twisti@993: //--------------------------LoadUSNode::Ideal------------------------------------- duke@435: // duke@435: // If the previous store is to the same address as this load, duke@435: // and the value stored was larger than a char, replace this load duke@435: // with the value stored truncated to a char. If no truncation is duke@435: // needed, the replacement is done in LoadNode::Identity(). duke@435: // twisti@993: Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) { duke@435: Node* mem = in(MemNode::Memory); duke@435: Node* value = can_see_stored_value(mem,phase); duke@435: if( value && !phase->type(value)->higher_equal( _type ) ) duke@435: return new (phase->C, 3) AndINode(value,phase->intcon(0xFFFF)); duke@435: // Identity call will handle the case where truncation is not needed. duke@435: return LoadNode::Ideal(phase, can_reshape); duke@435: } duke@435: duke@435: //--------------------------LoadSNode::Ideal-------------------------------------- duke@435: // duke@435: // If the previous store is to the same address as this load, duke@435: // and the value stored was larger than a short, replace this load duke@435: // with the value stored truncated to a short. If no truncation is duke@435: // needed, the replacement is done in LoadNode::Identity(). duke@435: // duke@435: Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) { duke@435: Node* mem = in(MemNode::Memory); duke@435: Node* value = can_see_stored_value(mem,phase); duke@435: if( value && !phase->type(value)->higher_equal( _type ) ) { duke@435: Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(16)) ); duke@435: return new (phase->C, 3) RShiftINode(result, phase->intcon(16)); duke@435: } duke@435: // Identity call will handle the case where truncation is not needed. duke@435: return LoadNode::Ideal(phase, can_reshape); duke@435: } duke@435: duke@435: //============================================================================= kvn@599: //----------------------------LoadKlassNode::make------------------------------ kvn@599: // Polymorphic factory method: kvn@599: Node *LoadKlassNode::make( PhaseGVN& gvn, Node *mem, Node *adr, const TypePtr* at, const TypeKlassPtr *tk ) { kvn@599: Compile* C = gvn.C; kvn@599: Node *ctl = NULL; kvn@599: // sanity check the alias category against the created node type kvn@599: const TypeOopPtr *adr_type = adr->bottom_type()->isa_oopptr(); kvn@599: assert(adr_type != NULL, "expecting TypeOopPtr"); kvn@599: #ifdef _LP64 kvn@599: if (adr_type->is_ptr_to_narrowoop()) { kvn@656: Node* load_klass = gvn.transform(new (C, 3) LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowoop())); kvn@656: return new (C, 2) DecodeNNode(load_klass, load_klass->bottom_type()->make_ptr()); kvn@603: } kvn@599: #endif kvn@603: assert(!adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop"); kvn@603: return new (C, 3) LoadKlassNode(ctl, mem, adr, at, tk); kvn@599: } kvn@599: duke@435: //------------------------------Value------------------------------------------ duke@435: const Type *LoadKlassNode::Value( PhaseTransform *phase ) const { kvn@599: return klass_value_common(phase); kvn@599: } kvn@599: kvn@599: const Type *LoadNode::klass_value_common( PhaseTransform *phase ) const { duke@435: // Either input is TOP ==> the result is TOP duke@435: const Type *t1 = phase->type( in(MemNode::Memory) ); duke@435: if (t1 == Type::TOP) return Type::TOP; duke@435: Node *adr = in(MemNode::Address); duke@435: const Type *t2 = phase->type( adr ); duke@435: if (t2 == Type::TOP) return Type::TOP; duke@435: const TypePtr *tp = t2->is_ptr(); duke@435: if (TypePtr::above_centerline(tp->ptr()) || duke@435: tp->ptr() == TypePtr::Null) return Type::TOP; duke@435: duke@435: // Return a more precise klass, if possible duke@435: const TypeInstPtr *tinst = tp->isa_instptr(); duke@435: if (tinst != NULL) { duke@435: ciInstanceKlass* ik = tinst->klass()->as_instance_klass(); duke@435: int offset = tinst->offset(); duke@435: if (ik == phase->C->env()->Class_klass() duke@435: && (offset == java_lang_Class::klass_offset_in_bytes() || duke@435: offset == java_lang_Class::array_klass_offset_in_bytes())) { duke@435: // We are loading a special hidden field from a Class mirror object, duke@435: // the field which points to the VM's Klass metaobject. duke@435: ciType* t = tinst->java_mirror_type(); duke@435: // java_mirror_type returns non-null for compile-time Class constants. duke@435: if (t != NULL) { duke@435: // constant oop => constant klass duke@435: if (offset == java_lang_Class::array_klass_offset_in_bytes()) { duke@435: return TypeKlassPtr::make(ciArrayKlass::make(t)); duke@435: } duke@435: if (!t->is_klass()) { duke@435: // a primitive Class (e.g., int.class) has NULL for a klass field duke@435: return TypePtr::NULL_PTR; duke@435: } duke@435: // (Folds up the 1st indirection in aClassConstant.getModifiers().) duke@435: return TypeKlassPtr::make(t->as_klass()); duke@435: } duke@435: // non-constant mirror, so we can't tell what's going on duke@435: } duke@435: if( !ik->is_loaded() ) duke@435: return _type; // Bail out if not loaded duke@435: if (offset == oopDesc::klass_offset_in_bytes()) { duke@435: if (tinst->klass_is_exact()) { duke@435: return TypeKlassPtr::make(ik); duke@435: } duke@435: // See if we can become precise: no subklasses and no interface duke@435: // (Note: We need to support verified interfaces.) duke@435: if (!ik->is_interface() && !ik->has_subklass()) { duke@435: //assert(!UseExactTypes, "this code should be useless with exact types"); duke@435: // Add a dependence; if any subclass added we need to recompile duke@435: if (!ik->is_final()) { duke@435: // %%% should use stronger assert_unique_concrete_subtype instead duke@435: phase->C->dependencies()->assert_leaf_type(ik); duke@435: } duke@435: // Return precise klass duke@435: return TypeKlassPtr::make(ik); duke@435: } duke@435: duke@435: // Return root of possible klass duke@435: return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/); duke@435: } duke@435: } duke@435: duke@435: // Check for loading klass from an array duke@435: const TypeAryPtr *tary = tp->isa_aryptr(); duke@435: if( tary != NULL ) { duke@435: ciKlass *tary_klass = tary->klass(); duke@435: if (tary_klass != NULL // can be NULL when at BOTTOM or TOP duke@435: && tary->offset() == oopDesc::klass_offset_in_bytes()) { duke@435: if (tary->klass_is_exact()) { duke@435: return TypeKlassPtr::make(tary_klass); duke@435: } duke@435: ciArrayKlass *ak = tary->klass()->as_array_klass(); duke@435: // If the klass is an object array, we defer the question to the duke@435: // array component klass. duke@435: if( ak->is_obj_array_klass() ) { duke@435: assert( ak->is_loaded(), "" ); duke@435: ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass(); duke@435: if( base_k->is_loaded() && base_k->is_instance_klass() ) { duke@435: ciInstanceKlass* ik = base_k->as_instance_klass(); duke@435: // See if we can become precise: no subklasses and no interface duke@435: if (!ik->is_interface() && !ik->has_subklass()) { duke@435: //assert(!UseExactTypes, "this code should be useless with exact types"); duke@435: // Add a dependence; if any subclass added we need to recompile duke@435: if (!ik->is_final()) { duke@435: phase->C->dependencies()->assert_leaf_type(ik); duke@435: } duke@435: // Return precise array klass duke@435: return TypeKlassPtr::make(ak); duke@435: } duke@435: } duke@435: return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/); duke@435: } else { // Found a type-array? duke@435: //assert(!UseExactTypes, "this code should be useless with exact types"); duke@435: assert( ak->is_type_array_klass(), "" ); duke@435: return TypeKlassPtr::make(ak); // These are always precise duke@435: } duke@435: } duke@435: } duke@435: duke@435: // Check for loading klass from an array klass duke@435: const TypeKlassPtr *tkls = tp->isa_klassptr(); duke@435: if (tkls != NULL && !StressReflectiveCode) { duke@435: ciKlass* klass = tkls->klass(); duke@435: if( !klass->is_loaded() ) duke@435: return _type; // Bail out if not loaded duke@435: if( klass->is_obj_array_klass() && duke@435: (uint)tkls->offset() == objArrayKlass::element_klass_offset_in_bytes() + sizeof(oopDesc)) { duke@435: ciKlass* elem = klass->as_obj_array_klass()->element_klass(); duke@435: // // Always returning precise element type is incorrect, duke@435: // // e.g., element type could be object and array may contain strings duke@435: // return TypeKlassPtr::make(TypePtr::Constant, elem, 0); duke@435: duke@435: // The array's TypeKlassPtr was declared 'precise' or 'not precise' duke@435: // according to the element type's subclassing. duke@435: return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/); duke@435: } duke@435: if( klass->is_instance_klass() && tkls->klass_is_exact() && duke@435: (uint)tkls->offset() == Klass::super_offset_in_bytes() + sizeof(oopDesc)) { duke@435: ciKlass* sup = klass->as_instance_klass()->super(); duke@435: // The field is Klass::_super. Return its (constant) value. duke@435: // (Folds up the 2nd indirection in aClassConstant.getSuperClass().) duke@435: return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR; duke@435: } duke@435: } duke@435: duke@435: // Bailout case duke@435: return LoadNode::Value(phase); duke@435: } duke@435: duke@435: //------------------------------Identity--------------------------------------- duke@435: // To clean up reflective code, simplify k.java_mirror.as_klass to plain k. duke@435: // Also feed through the klass in Allocate(...klass...)._klass. duke@435: Node* LoadKlassNode::Identity( PhaseTransform *phase ) { kvn@599: return klass_identity_common(phase); kvn@599: } kvn@599: kvn@599: Node* LoadNode::klass_identity_common(PhaseTransform *phase ) { duke@435: Node* x = LoadNode::Identity(phase); duke@435: if (x != this) return x; duke@435: duke@435: // Take apart the address into an oop and and offset. duke@435: // Return 'this' if we cannot. duke@435: Node* adr = in(MemNode::Address); duke@435: intptr_t offset = 0; duke@435: Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); duke@435: if (base == NULL) return this; duke@435: const TypeOopPtr* toop = phase->type(adr)->isa_oopptr(); duke@435: if (toop == NULL) return this; duke@435: duke@435: // We can fetch the klass directly through an AllocateNode. duke@435: // This works even if the klass is not constant (clone or newArray). duke@435: if (offset == oopDesc::klass_offset_in_bytes()) { duke@435: Node* allocated_klass = AllocateNode::Ideal_klass(base, phase); duke@435: if (allocated_klass != NULL) { duke@435: return allocated_klass; duke@435: } duke@435: } duke@435: duke@435: // Simplify k.java_mirror.as_klass to plain k, where k is a klassOop. duke@435: // Simplify ak.component_mirror.array_klass to plain ak, ak an arrayKlass. duke@435: // See inline_native_Class_query for occurrences of these patterns. duke@435: // Java Example: x.getClass().isAssignableFrom(y) duke@435: // Java Example: Array.newInstance(x.getClass().getComponentType(), n) duke@435: // duke@435: // This improves reflective code, often making the Class duke@435: // mirror go completely dead. (Current exception: Class duke@435: // mirrors may appear in debug info, but we could clean them out by duke@435: // introducing a new debug info operator for klassOop.java_mirror). duke@435: if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass() duke@435: && (offset == java_lang_Class::klass_offset_in_bytes() || duke@435: offset == java_lang_Class::array_klass_offset_in_bytes())) { duke@435: // We are loading a special hidden field from a Class mirror, duke@435: // the field which points to its Klass or arrayKlass metaobject. duke@435: if (base->is_Load()) { duke@435: Node* adr2 = base->in(MemNode::Address); duke@435: const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr(); duke@435: if (tkls != NULL && !tkls->empty() duke@435: && (tkls->klass()->is_instance_klass() || duke@435: tkls->klass()->is_array_klass()) duke@435: && adr2->is_AddP() duke@435: ) { duke@435: int mirror_field = Klass::java_mirror_offset_in_bytes(); duke@435: if (offset == java_lang_Class::array_klass_offset_in_bytes()) { duke@435: mirror_field = in_bytes(arrayKlass::component_mirror_offset()); duke@435: } duke@435: if (tkls->offset() == mirror_field + (int)sizeof(oopDesc)) { duke@435: return adr2->in(AddPNode::Base); duke@435: } duke@435: } duke@435: } duke@435: } duke@435: duke@435: return this; duke@435: } duke@435: kvn@599: kvn@599: //------------------------------Value------------------------------------------ kvn@599: const Type *LoadNKlassNode::Value( PhaseTransform *phase ) const { kvn@599: const Type *t = klass_value_common(phase); kvn@656: if (t == Type::TOP) kvn@656: return t; kvn@656: kvn@656: return t->make_narrowoop(); kvn@599: } kvn@599: kvn@599: //------------------------------Identity--------------------------------------- kvn@599: // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k. kvn@599: // Also feed through the klass in Allocate(...klass...)._klass. kvn@599: Node* LoadNKlassNode::Identity( PhaseTransform *phase ) { kvn@599: Node *x = klass_identity_common(phase); kvn@599: kvn@599: const Type *t = phase->type( x ); kvn@599: if( t == Type::TOP ) return x; kvn@599: if( t->isa_narrowoop()) return x; kvn@599: kvn@656: return phase->transform(new (phase->C, 2) EncodePNode(x, t->make_narrowoop())); kvn@599: } kvn@599: duke@435: //------------------------------Value----------------------------------------- duke@435: const Type *LoadRangeNode::Value( PhaseTransform *phase ) const { duke@435: // Either input is TOP ==> the result is TOP duke@435: const Type *t1 = phase->type( in(MemNode::Memory) ); duke@435: if( t1 == Type::TOP ) return Type::TOP; duke@435: Node *adr = in(MemNode::Address); duke@435: const Type *t2 = phase->type( adr ); duke@435: if( t2 == Type::TOP ) return Type::TOP; duke@435: const TypePtr *tp = t2->is_ptr(); duke@435: if (TypePtr::above_centerline(tp->ptr())) return Type::TOP; duke@435: const TypeAryPtr *tap = tp->isa_aryptr(); duke@435: if( !tap ) return _type; duke@435: return tap->size(); duke@435: } duke@435: rasbold@801: //-------------------------------Ideal--------------------------------------- rasbold@801: // Feed through the length in AllocateArray(...length...)._length. rasbold@801: Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) { rasbold@801: Node* p = MemNode::Ideal_common(phase, can_reshape); rasbold@801: if (p) return (p == NodeSentinel) ? NULL : p; rasbold@801: rasbold@801: // Take apart the address into an oop and and offset. rasbold@801: // Return 'this' if we cannot. rasbold@801: Node* adr = in(MemNode::Address); rasbold@801: intptr_t offset = 0; rasbold@801: Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); rasbold@801: if (base == NULL) return NULL; rasbold@801: const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); rasbold@801: if (tary == NULL) return NULL; rasbold@801: rasbold@801: // We can fetch the length directly through an AllocateArrayNode. rasbold@801: // This works even if the length is not constant (clone or newArray). rasbold@801: if (offset == arrayOopDesc::length_offset_in_bytes()) { rasbold@801: AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase); rasbold@801: if (alloc != NULL) { rasbold@801: Node* allocated_length = alloc->Ideal_length(); rasbold@801: Node* len = alloc->make_ideal_length(tary, phase); rasbold@801: if (allocated_length != len) { rasbold@801: // New CastII improves on this. rasbold@801: return len; rasbold@801: } rasbold@801: } rasbold@801: } rasbold@801: rasbold@801: return NULL; rasbold@801: } rasbold@801: duke@435: //------------------------------Identity--------------------------------------- duke@435: // Feed through the length in AllocateArray(...length...)._length. duke@435: Node* LoadRangeNode::Identity( PhaseTransform *phase ) { duke@435: Node* x = LoadINode::Identity(phase); duke@435: if (x != this) return x; duke@435: duke@435: // Take apart the address into an oop and and offset. duke@435: // Return 'this' if we cannot. duke@435: Node* adr = in(MemNode::Address); duke@435: intptr_t offset = 0; duke@435: Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset); duke@435: if (base == NULL) return this; duke@435: const TypeAryPtr* tary = phase->type(adr)->isa_aryptr(); duke@435: if (tary == NULL) return this; duke@435: duke@435: // We can fetch the length directly through an AllocateArrayNode. duke@435: // This works even if the length is not constant (clone or newArray). duke@435: if (offset == arrayOopDesc::length_offset_in_bytes()) { rasbold@801: AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase); rasbold@801: if (alloc != NULL) { rasbold@801: Node* allocated_length = alloc->Ideal_length(); rasbold@801: // Do not allow make_ideal_length to allocate a CastII node. rasbold@801: Node* len = alloc->make_ideal_length(tary, phase, false); rasbold@801: if (allocated_length == len) { rasbold@801: // Return allocated_length only if it would not be improved by a CastII. rasbold@801: return allocated_length; rasbold@801: } duke@435: } duke@435: } duke@435: duke@435: return this; duke@435: duke@435: } rasbold@801: duke@435: //============================================================================= duke@435: //---------------------------StoreNode::make----------------------------------- duke@435: // Polymorphic factory method: coleenp@548: StoreNode* StoreNode::make( PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt ) { coleenp@548: Compile* C = gvn.C; coleenp@548: duke@435: switch (bt) { duke@435: case T_BOOLEAN: duke@435: case T_BYTE: return new (C, 4) StoreBNode(ctl, mem, adr, adr_type, val); duke@435: case T_INT: return new (C, 4) StoreINode(ctl, mem, adr, adr_type, val); duke@435: case T_CHAR: duke@435: case T_SHORT: return new (C, 4) StoreCNode(ctl, mem, adr, adr_type, val); duke@435: case T_LONG: return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val); duke@435: case T_FLOAT: return new (C, 4) StoreFNode(ctl, mem, adr, adr_type, val); duke@435: case T_DOUBLE: return new (C, 4) StoreDNode(ctl, mem, adr, adr_type, val); duke@435: case T_ADDRESS: coleenp@548: case T_OBJECT: coleenp@548: #ifdef _LP64 kvn@598: if (adr->bottom_type()->is_ptr_to_narrowoop() || coleenp@548: (UseCompressedOops && val->bottom_type()->isa_klassptr() && coleenp@548: adr->bottom_type()->isa_rawptr())) { kvn@656: val = gvn.transform(new (C, 2) EncodePNode(val, val->bottom_type()->make_narrowoop())); kvn@656: return new (C, 4) StoreNNode(ctl, mem, adr, adr_type, val); coleenp@548: } else coleenp@548: #endif kvn@656: { kvn@656: return new (C, 4) StorePNode(ctl, mem, adr, adr_type, val); kvn@656: } duke@435: } duke@435: ShouldNotReachHere(); duke@435: return (StoreNode*)NULL; duke@435: } duke@435: duke@435: StoreLNode* StoreLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val) { duke@435: bool require_atomic = true; duke@435: return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val, require_atomic); duke@435: } duke@435: duke@435: duke@435: //--------------------------bottom_type---------------------------------------- duke@435: const Type *StoreNode::bottom_type() const { duke@435: return Type::MEMORY; duke@435: } duke@435: duke@435: //------------------------------hash------------------------------------------- duke@435: uint StoreNode::hash() const { duke@435: // unroll addition of interesting fields duke@435: //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn); duke@435: duke@435: // Since they are not commoned, do not hash them: duke@435: return NO_HASH; duke@435: } duke@435: duke@435: //------------------------------Ideal------------------------------------------ duke@435: // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x). duke@435: // When a store immediately follows a relevant allocation/initialization, duke@435: // try to capture it into the initialization, or hoist it above. duke@435: Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) { duke@435: Node* p = MemNode::Ideal_common(phase, can_reshape); duke@435: if (p) return (p == NodeSentinel) ? NULL : p; duke@435: duke@435: Node* mem = in(MemNode::Memory); duke@435: Node* address = in(MemNode::Address); duke@435: duke@435: // Back-to-back stores to same address? Fold em up. duke@435: // Generally unsafe if I have intervening uses... duke@435: if (mem->is_Store() && phase->eqv_uncast(mem->in(MemNode::Address), address)) { duke@435: // Looking at a dead closed cycle of memory? duke@435: assert(mem != mem->in(MemNode::Memory), "dead loop in StoreNode::Ideal"); duke@435: duke@435: assert(Opcode() == mem->Opcode() || duke@435: phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw, duke@435: "no mismatched stores, except on raw memory"); duke@435: duke@435: if (mem->outcnt() == 1 && // check for intervening uses duke@435: mem->as_Store()->memory_size() <= this->memory_size()) { duke@435: // If anybody other than 'this' uses 'mem', we cannot fold 'mem' away. duke@435: // For example, 'mem' might be the final state at a conditional return. duke@435: // Or, 'mem' might be used by some node which is live at the same time duke@435: // 'this' is live, which might be unschedulable. So, require exactly duke@435: // ONE user, the 'this' store, until such time as we clone 'mem' for duke@435: // each of 'mem's uses (thus making the exactly-1-user-rule hold true). duke@435: if (can_reshape) { // (%%% is this an anachronism?) duke@435: set_req_X(MemNode::Memory, mem->in(MemNode::Memory), duke@435: phase->is_IterGVN()); duke@435: } else { duke@435: // It's OK to do this in the parser, since DU info is always accurate, duke@435: // and the parser always refers to nodes via SafePointNode maps. duke@435: set_req(MemNode::Memory, mem->in(MemNode::Memory)); duke@435: } duke@435: return this; duke@435: } duke@435: } duke@435: duke@435: // Capture an unaliased, unconditional, simple store into an initializer. duke@435: // Or, if it is independent of the allocation, hoist it above the allocation. duke@435: if (ReduceFieldZeroing && /*can_reshape &&*/ duke@435: mem->is_Proj() && mem->in(0)->is_Initialize()) { duke@435: InitializeNode* init = mem->in(0)->as_Initialize(); duke@435: intptr_t offset = init->can_capture_store(this, phase); duke@435: if (offset > 0) { duke@435: Node* moved = init->capture_store(this, offset, phase); duke@435: // If the InitializeNode captured me, it made a raw copy of me, duke@435: // and I need to disappear. duke@435: if (moved != NULL) { duke@435: // %%% hack to ensure that Ideal returns a new node: duke@435: mem = MergeMemNode::make(phase->C, mem); duke@435: return mem; // fold me away duke@435: } duke@435: } duke@435: } duke@435: duke@435: return NULL; // No further progress duke@435: } duke@435: duke@435: //------------------------------Value----------------------------------------- duke@435: const Type *StoreNode::Value( PhaseTransform *phase ) const { duke@435: // Either input is TOP ==> the result is TOP duke@435: const Type *t1 = phase->type( in(MemNode::Memory) ); duke@435: if( t1 == Type::TOP ) return Type::TOP; duke@435: const Type *t2 = phase->type( in(MemNode::Address) ); duke@435: if( t2 == Type::TOP ) return Type::TOP; duke@435: const Type *t3 = phase->type( in(MemNode::ValueIn) ); duke@435: if( t3 == Type::TOP ) return Type::TOP; duke@435: return Type::MEMORY; duke@435: } duke@435: duke@435: //------------------------------Identity--------------------------------------- duke@435: // Remove redundant stores: duke@435: // Store(m, p, Load(m, p)) changes to m. duke@435: // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x). duke@435: Node *StoreNode::Identity( PhaseTransform *phase ) { duke@435: Node* mem = in(MemNode::Memory); duke@435: Node* adr = in(MemNode::Address); duke@435: Node* val = in(MemNode::ValueIn); duke@435: duke@435: // Load then Store? Then the Store is useless duke@435: if (val->is_Load() && duke@435: phase->eqv_uncast( val->in(MemNode::Address), adr ) && duke@435: phase->eqv_uncast( val->in(MemNode::Memory ), mem ) && duke@435: val->as_Load()->store_Opcode() == Opcode()) { duke@435: return mem; duke@435: } duke@435: duke@435: // Two stores in a row of the same value? duke@435: if (mem->is_Store() && duke@435: phase->eqv_uncast( mem->in(MemNode::Address), adr ) && duke@435: phase->eqv_uncast( mem->in(MemNode::ValueIn), val ) && duke@435: mem->Opcode() == Opcode()) { duke@435: return mem; duke@435: } duke@435: duke@435: // Store of zero anywhere into a freshly-allocated object? duke@435: // Then the store is useless. duke@435: // (It must already have been captured by the InitializeNode.) duke@435: if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) { duke@435: // a newly allocated object is already all-zeroes everywhere duke@435: if (mem->is_Proj() && mem->in(0)->is_Allocate()) { duke@435: return mem; duke@435: } duke@435: duke@435: // the store may also apply to zero-bits in an earlier object duke@435: Node* prev_mem = find_previous_store(phase); duke@435: // Steps (a), (b): Walk past independent stores to find an exact match. duke@435: if (prev_mem != NULL) { duke@435: Node* prev_val = can_see_stored_value(prev_mem, phase); duke@435: if (prev_val != NULL && phase->eqv(prev_val, val)) { duke@435: // prev_val and val might differ by a cast; it would be good duke@435: // to keep the more informative of the two. duke@435: return mem; duke@435: } duke@435: } duke@435: } duke@435: duke@435: return this; duke@435: } duke@435: duke@435: //------------------------------match_edge------------------------------------- duke@435: // Do we Match on this edge index or not? Match only memory & value duke@435: uint StoreNode::match_edge(uint idx) const { duke@435: return idx == MemNode::Address || idx == MemNode::ValueIn; duke@435: } duke@435: duke@435: //------------------------------cmp-------------------------------------------- duke@435: // Do not common stores up together. They generally have to be split duke@435: // back up anyways, so do not bother. duke@435: uint StoreNode::cmp( const Node &n ) const { duke@435: return (&n == this); // Always fail except on self duke@435: } duke@435: duke@435: //------------------------------Ideal_masked_input----------------------------- duke@435: // Check for a useless mask before a partial-word store duke@435: // (StoreB ... (AndI valIn conIa) ) duke@435: // If (conIa & mask == mask) this simplifies to duke@435: // (StoreB ... (valIn) ) duke@435: Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) { duke@435: Node *val = in(MemNode::ValueIn); duke@435: if( val->Opcode() == Op_AndI ) { duke@435: const TypeInt *t = phase->type( val->in(2) )->isa_int(); duke@435: if( t && t->is_con() && (t->get_con() & mask) == mask ) { duke@435: set_req(MemNode::ValueIn, val->in(1)); duke@435: return this; duke@435: } duke@435: } duke@435: return NULL; duke@435: } duke@435: duke@435: duke@435: //------------------------------Ideal_sign_extended_input---------------------- duke@435: // Check for useless sign-extension before a partial-word store duke@435: // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) ) duke@435: // If (conIL == conIR && conIR <= num_bits) this simplifies to duke@435: // (StoreB ... (valIn) ) duke@435: Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) { duke@435: Node *val = in(MemNode::ValueIn); duke@435: if( val->Opcode() == Op_RShiftI ) { duke@435: const TypeInt *t = phase->type( val->in(2) )->isa_int(); duke@435: if( t && t->is_con() && (t->get_con() <= num_bits) ) { duke@435: Node *shl = val->in(1); duke@435: if( shl->Opcode() == Op_LShiftI ) { duke@435: const TypeInt *t2 = phase->type( shl->in(2) )->isa_int(); duke@435: if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) { duke@435: set_req(MemNode::ValueIn, shl->in(1)); duke@435: return this; duke@435: } duke@435: } duke@435: } duke@435: } duke@435: return NULL; duke@435: } duke@435: duke@435: //------------------------------value_never_loaded----------------------------------- duke@435: // Determine whether there are any possible loads of the value stored. duke@435: // For simplicity, we actually check if there are any loads from the duke@435: // address stored to, not just for loads of the value stored by this node. duke@435: // duke@435: bool StoreNode::value_never_loaded( PhaseTransform *phase) const { duke@435: Node *adr = in(Address); duke@435: const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr(); duke@435: if (adr_oop == NULL) duke@435: return false; kvn@658: if (!adr_oop->is_known_instance_field()) duke@435: return false; // if not a distinct instance, there may be aliases of the address duke@435: for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) { duke@435: Node *use = adr->fast_out(i); duke@435: int opc = use->Opcode(); duke@435: if (use->is_Load() || use->is_LoadStore()) { duke@435: return false; duke@435: } duke@435: } duke@435: return true; duke@435: } duke@435: duke@435: //============================================================================= duke@435: //------------------------------Ideal------------------------------------------ duke@435: // If the store is from an AND mask that leaves the low bits untouched, then duke@435: // we can skip the AND operation. If the store is from a sign-extension duke@435: // (a left shift, then right shift) we can skip both. duke@435: Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){ duke@435: Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF); duke@435: if( progress != NULL ) return progress; duke@435: duke@435: progress = StoreNode::Ideal_sign_extended_input(phase, 24); duke@435: if( progress != NULL ) return progress; duke@435: duke@435: // Finally check the default case duke@435: return StoreNode::Ideal(phase, can_reshape); duke@435: } duke@435: duke@435: //============================================================================= duke@435: //------------------------------Ideal------------------------------------------ duke@435: // If the store is from an AND mask that leaves the low bits untouched, then duke@435: // we can skip the AND operation duke@435: Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){ duke@435: Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF); duke@435: if( progress != NULL ) return progress; duke@435: duke@435: progress = StoreNode::Ideal_sign_extended_input(phase, 16); duke@435: if( progress != NULL ) return progress; duke@435: duke@435: // Finally check the default case duke@435: return StoreNode::Ideal(phase, can_reshape); duke@435: } duke@435: duke@435: //============================================================================= duke@435: //------------------------------Identity--------------------------------------- duke@435: Node *StoreCMNode::Identity( PhaseTransform *phase ) { duke@435: // No need to card mark when storing a null ptr duke@435: Node* my_store = in(MemNode::OopStore); duke@435: if (my_store->is_Store()) { duke@435: const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) ); duke@435: if( t1 == TypePtr::NULL_PTR ) { duke@435: return in(MemNode::Memory); duke@435: } duke@435: } duke@435: return this; duke@435: } duke@435: duke@435: //------------------------------Value----------------------------------------- duke@435: const Type *StoreCMNode::Value( PhaseTransform *phase ) const { kvn@478: // Either input is TOP ==> the result is TOP kvn@478: const Type *t = phase->type( in(MemNode::Memory) ); kvn@478: if( t == Type::TOP ) return Type::TOP; kvn@478: t = phase->type( in(MemNode::Address) ); kvn@478: if( t == Type::TOP ) return Type::TOP; kvn@478: t = phase->type( in(MemNode::ValueIn) ); kvn@478: if( t == Type::TOP ) return Type::TOP; duke@435: // If extra input is TOP ==> the result is TOP kvn@478: t = phase->type( in(MemNode::OopStore) ); kvn@478: if( t == Type::TOP ) return Type::TOP; duke@435: duke@435: return StoreNode::Value( phase ); duke@435: } duke@435: duke@435: duke@435: //============================================================================= duke@435: //----------------------------------SCMemProjNode------------------------------ duke@435: const Type * SCMemProjNode::Value( PhaseTransform *phase ) const duke@435: { duke@435: return bottom_type(); duke@435: } duke@435: duke@435: //============================================================================= duke@435: LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : Node(5) { duke@435: init_req(MemNode::Control, c ); duke@435: init_req(MemNode::Memory , mem); duke@435: init_req(MemNode::Address, adr); duke@435: init_req(MemNode::ValueIn, val); duke@435: init_req( ExpectedIn, ex ); duke@435: init_class_id(Class_LoadStore); duke@435: duke@435: } duke@435: duke@435: //============================================================================= duke@435: //-------------------------------adr_type-------------------------------------- duke@435: // Do we Match on this edge index or not? Do not match memory duke@435: const TypePtr* ClearArrayNode::adr_type() const { duke@435: Node *adr = in(3); duke@435: return MemNode::calculate_adr_type(adr->bottom_type()); duke@435: } duke@435: duke@435: //------------------------------match_edge------------------------------------- duke@435: // Do we Match on this edge index or not? Do not match memory duke@435: uint ClearArrayNode::match_edge(uint idx) const { duke@435: return idx > 1; duke@435: } duke@435: duke@435: //------------------------------Identity--------------------------------------- duke@435: // Clearing a zero length array does nothing duke@435: Node *ClearArrayNode::Identity( PhaseTransform *phase ) { never@503: return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this; duke@435: } duke@435: duke@435: //------------------------------Idealize--------------------------------------- duke@435: // Clearing a short array is faster with stores duke@435: Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){ duke@435: const int unit = BytesPerLong; duke@435: const TypeX* t = phase->type(in(2))->isa_intptr_t(); duke@435: if (!t) return NULL; duke@435: if (!t->is_con()) return NULL; duke@435: intptr_t raw_count = t->get_con(); duke@435: intptr_t size = raw_count; duke@435: if (!Matcher::init_array_count_is_in_bytes) size *= unit; duke@435: // Clearing nothing uses the Identity call. duke@435: // Negative clears are possible on dead ClearArrays duke@435: // (see jck test stmt114.stmt11402.val). duke@435: if (size <= 0 || size % unit != 0) return NULL; duke@435: intptr_t count = size / unit; duke@435: // Length too long; use fast hardware clear duke@435: if (size > Matcher::init_array_short_size) return NULL; duke@435: Node *mem = in(1); duke@435: if( phase->type(mem)==Type::TOP ) return NULL; duke@435: Node *adr = in(3); duke@435: const Type* at = phase->type(adr); duke@435: if( at==Type::TOP ) return NULL; duke@435: const TypePtr* atp = at->isa_ptr(); duke@435: // adjust atp to be the correct array element address type duke@435: if (atp == NULL) atp = TypePtr::BOTTOM; duke@435: else atp = atp->add_offset(Type::OffsetBot); duke@435: // Get base for derived pointer purposes duke@435: if( adr->Opcode() != Op_AddP ) Unimplemented(); duke@435: Node *base = adr->in(1); duke@435: duke@435: Node *zero = phase->makecon(TypeLong::ZERO); duke@435: Node *off = phase->MakeConX(BytesPerLong); duke@435: mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero); duke@435: count--; duke@435: while( count-- ) { duke@435: mem = phase->transform(mem); duke@435: adr = phase->transform(new (phase->C, 4) AddPNode(base,adr,off)); duke@435: mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero); duke@435: } duke@435: return mem; duke@435: } duke@435: duke@435: //----------------------------clear_memory------------------------------------- duke@435: // Generate code to initialize object storage to zero. duke@435: Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, duke@435: intptr_t start_offset, duke@435: Node* end_offset, duke@435: PhaseGVN* phase) { duke@435: Compile* C = phase->C; duke@435: intptr_t offset = start_offset; duke@435: duke@435: int unit = BytesPerLong; duke@435: if ((offset % unit) != 0) { duke@435: Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(offset)); duke@435: adr = phase->transform(adr); duke@435: const TypePtr* atp = TypeRawPtr::BOTTOM; coleenp@548: mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT); duke@435: mem = phase->transform(mem); duke@435: offset += BytesPerInt; duke@435: } duke@435: assert((offset % unit) == 0, ""); duke@435: duke@435: // Initialize the remaining stuff, if any, with a ClearArray. duke@435: return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase); duke@435: } duke@435: duke@435: Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, duke@435: Node* start_offset, duke@435: Node* end_offset, duke@435: PhaseGVN* phase) { never@503: if (start_offset == end_offset) { never@503: // nothing to do never@503: return mem; never@503: } never@503: duke@435: Compile* C = phase->C; duke@435: int unit = BytesPerLong; duke@435: Node* zbase = start_offset; duke@435: Node* zend = end_offset; duke@435: duke@435: // Scale to the unit required by the CPU: duke@435: if (!Matcher::init_array_count_is_in_bytes) { duke@435: Node* shift = phase->intcon(exact_log2(unit)); duke@435: zbase = phase->transform( new(C,3) URShiftXNode(zbase, shift) ); duke@435: zend = phase->transform( new(C,3) URShiftXNode(zend, shift) ); duke@435: } duke@435: duke@435: Node* zsize = phase->transform( new(C,3) SubXNode(zend, zbase) ); duke@435: Node* zinit = phase->zerocon((unit == BytesPerLong) ? T_LONG : T_INT); duke@435: duke@435: // Bulk clear double-words duke@435: Node* adr = phase->transform( new(C,4) AddPNode(dest, dest, start_offset) ); duke@435: mem = new (C, 4) ClearArrayNode(ctl, mem, zsize, adr); duke@435: return phase->transform(mem); duke@435: } duke@435: duke@435: Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest, duke@435: intptr_t start_offset, duke@435: intptr_t end_offset, duke@435: PhaseGVN* phase) { never@503: if (start_offset == end_offset) { never@503: // nothing to do never@503: return mem; never@503: } never@503: duke@435: Compile* C = phase->C; duke@435: assert((end_offset % BytesPerInt) == 0, "odd end offset"); duke@435: intptr_t done_offset = end_offset; duke@435: if ((done_offset % BytesPerLong) != 0) { duke@435: done_offset -= BytesPerInt; duke@435: } duke@435: if (done_offset > start_offset) { duke@435: mem = clear_memory(ctl, mem, dest, duke@435: start_offset, phase->MakeConX(done_offset), phase); duke@435: } duke@435: if (done_offset < end_offset) { // emit the final 32-bit store duke@435: Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(done_offset)); duke@435: adr = phase->transform(adr); duke@435: const TypePtr* atp = TypeRawPtr::BOTTOM; coleenp@548: mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT); duke@435: mem = phase->transform(mem); duke@435: done_offset += BytesPerInt; duke@435: } duke@435: assert(done_offset == end_offset, ""); duke@435: return mem; duke@435: } duke@435: duke@435: //============================================================================= duke@435: // Do we match on this edge? No memory edges duke@435: uint StrCompNode::match_edge(uint idx) const { duke@435: return idx == 5 || idx == 6; duke@435: } duke@435: duke@435: //------------------------------Ideal------------------------------------------ duke@435: // Return a node which is more "ideal" than the current node. Strip out duke@435: // control copies duke@435: Node *StrCompNode::Ideal(PhaseGVN *phase, bool can_reshape){ duke@435: return remove_dead_region(phase, can_reshape) ? this : NULL; duke@435: } duke@435: rasbold@604: //------------------------------Ideal------------------------------------------ rasbold@604: // Return a node which is more "ideal" than the current node. Strip out rasbold@604: // control copies rasbold@604: Node *AryEqNode::Ideal(PhaseGVN *phase, bool can_reshape){ rasbold@604: return remove_dead_region(phase, can_reshape) ? this : NULL; rasbold@604: } rasbold@604: duke@435: duke@435: //============================================================================= duke@435: MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent) duke@435: : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)), duke@435: _adr_type(C->get_adr_type(alias_idx)) duke@435: { duke@435: init_class_id(Class_MemBar); duke@435: Node* top = C->top(); duke@435: init_req(TypeFunc::I_O,top); duke@435: init_req(TypeFunc::FramePtr,top); duke@435: init_req(TypeFunc::ReturnAdr,top); duke@435: if (precedent != NULL) duke@435: init_req(TypeFunc::Parms, precedent); duke@435: } duke@435: duke@435: //------------------------------cmp-------------------------------------------- duke@435: uint MemBarNode::hash() const { return NO_HASH; } duke@435: uint MemBarNode::cmp( const Node &n ) const { duke@435: return (&n == this); // Always fail except on self duke@435: } duke@435: duke@435: //------------------------------make------------------------------------------- duke@435: MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) { duke@435: int len = Precedent + (pn == NULL? 0: 1); duke@435: switch (opcode) { duke@435: case Op_MemBarAcquire: return new(C, len) MemBarAcquireNode(C, atp, pn); duke@435: case Op_MemBarRelease: return new(C, len) MemBarReleaseNode(C, atp, pn); duke@435: case Op_MemBarVolatile: return new(C, len) MemBarVolatileNode(C, atp, pn); duke@435: case Op_MemBarCPUOrder: return new(C, len) MemBarCPUOrderNode(C, atp, pn); duke@435: case Op_Initialize: return new(C, len) InitializeNode(C, atp, pn); duke@435: default: ShouldNotReachHere(); return NULL; duke@435: } duke@435: } duke@435: duke@435: //------------------------------Ideal------------------------------------------ duke@435: // Return a node which is more "ideal" than the current node. Strip out duke@435: // control copies duke@435: Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) { kvn@740: return remove_dead_region(phase, can_reshape) ? this : NULL; duke@435: } duke@435: duke@435: //------------------------------Value------------------------------------------ duke@435: const Type *MemBarNode::Value( PhaseTransform *phase ) const { duke@435: if( !in(0) ) return Type::TOP; duke@435: if( phase->type(in(0)) == Type::TOP ) duke@435: return Type::TOP; duke@435: return TypeTuple::MEMBAR; duke@435: } duke@435: duke@435: //------------------------------match------------------------------------------ duke@435: // Construct projections for memory. duke@435: Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) { duke@435: switch (proj->_con) { duke@435: case TypeFunc::Control: duke@435: case TypeFunc::Memory: duke@435: return new (m->C, 1) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj); duke@435: } duke@435: ShouldNotReachHere(); duke@435: return NULL; duke@435: } duke@435: duke@435: //===========================InitializeNode==================================== duke@435: // SUMMARY: duke@435: // This node acts as a memory barrier on raw memory, after some raw stores. duke@435: // The 'cooked' oop value feeds from the Initialize, not the Allocation. duke@435: // The Initialize can 'capture' suitably constrained stores as raw inits. duke@435: // It can coalesce related raw stores into larger units (called 'tiles'). duke@435: // It can avoid zeroing new storage for memory units which have raw inits. duke@435: // At macro-expansion, it is marked 'complete', and does not optimize further. duke@435: // duke@435: // EXAMPLE: duke@435: // The object 'new short[2]' occupies 16 bytes in a 32-bit machine. duke@435: // ctl = incoming control; mem* = incoming memory duke@435: // (Note: A star * on a memory edge denotes I/O and other standard edges.) duke@435: // First allocate uninitialized memory and fill in the header: duke@435: // alloc = (Allocate ctl mem* 16 #short[].klass ...) duke@435: // ctl := alloc.Control; mem* := alloc.Memory* duke@435: // rawmem = alloc.Memory; rawoop = alloc.RawAddress duke@435: // Then initialize to zero the non-header parts of the raw memory block: duke@435: // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress) duke@435: // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory duke@435: // After the initialize node executes, the object is ready for service: duke@435: // oop := (CheckCastPP init.Control alloc.RawAddress #short[]) duke@435: // Suppose its body is immediately initialized as {1,2}: duke@435: // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) duke@435: // store2 = (StoreC init.Control store1 (+ oop 14) 2) duke@435: // mem.SLICE(#short[*]) := store2 duke@435: // duke@435: // DETAILS: duke@435: // An InitializeNode collects and isolates object initialization after duke@435: // an AllocateNode and before the next possible safepoint. As a duke@435: // memory barrier (MemBarNode), it keeps critical stores from drifting duke@435: // down past any safepoint or any publication of the allocation. duke@435: // Before this barrier, a newly-allocated object may have uninitialized bits. duke@435: // After this barrier, it may be treated as a real oop, and GC is allowed. duke@435: // duke@435: // The semantics of the InitializeNode include an implicit zeroing of duke@435: // the new object from object header to the end of the object. duke@435: // (The object header and end are determined by the AllocateNode.) duke@435: // duke@435: // Certain stores may be added as direct inputs to the InitializeNode. duke@435: // These stores must update raw memory, and they must be to addresses duke@435: // derived from the raw address produced by AllocateNode, and with duke@435: // a constant offset. They must be ordered by increasing offset. duke@435: // The first one is at in(RawStores), the last at in(req()-1). duke@435: // Unlike most memory operations, they are not linked in a chain, duke@435: // but are displayed in parallel as users of the rawmem output of duke@435: // the allocation. duke@435: // duke@435: // (See comments in InitializeNode::capture_store, which continue duke@435: // the example given above.) duke@435: // duke@435: // When the associated Allocate is macro-expanded, the InitializeNode duke@435: // may be rewritten to optimize collected stores. A ClearArrayNode duke@435: // may also be created at that point to represent any required zeroing. duke@435: // The InitializeNode is then marked 'complete', prohibiting further duke@435: // capturing of nearby memory operations. duke@435: // duke@435: // During macro-expansion, all captured initializations which store twisti@1040: // constant values of 32 bits or smaller are coalesced (if advantageous) duke@435: // into larger 'tiles' 32 or 64 bits. This allows an object to be duke@435: // initialized in fewer memory operations. Memory words which are duke@435: // covered by neither tiles nor non-constant stores are pre-zeroed duke@435: // by explicit stores of zero. (The code shape happens to do all duke@435: // zeroing first, then all other stores, with both sequences occurring duke@435: // in order of ascending offsets.) duke@435: // duke@435: // Alternatively, code may be inserted between an AllocateNode and its duke@435: // InitializeNode, to perform arbitrary initialization of the new object. duke@435: // E.g., the object copying intrinsics insert complex data transfers here. duke@435: // The initialization must then be marked as 'complete' disable the duke@435: // built-in zeroing semantics and the collection of initializing stores. duke@435: // duke@435: // While an InitializeNode is incomplete, reads from the memory state duke@435: // produced by it are optimizable if they match the control edge and duke@435: // new oop address associated with the allocation/initialization. duke@435: // They return a stored value (if the offset matches) or else zero. duke@435: // A write to the memory state, if it matches control and address, duke@435: // and if it is to a constant offset, may be 'captured' by the duke@435: // InitializeNode. It is cloned as a raw memory operation and rewired duke@435: // inside the initialization, to the raw oop produced by the allocation. duke@435: // Operations on addresses which are provably distinct (e.g., to duke@435: // other AllocateNodes) are allowed to bypass the initialization. duke@435: // duke@435: // The effect of all this is to consolidate object initialization duke@435: // (both arrays and non-arrays, both piecewise and bulk) into a duke@435: // single location, where it can be optimized as a unit. duke@435: // duke@435: // Only stores with an offset less than TrackedInitializationLimit words duke@435: // will be considered for capture by an InitializeNode. This puts a duke@435: // reasonable limit on the complexity of optimized initializations. duke@435: duke@435: //---------------------------InitializeNode------------------------------------ duke@435: InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop) duke@435: : _is_complete(false), duke@435: MemBarNode(C, adr_type, rawoop) duke@435: { duke@435: init_class_id(Class_Initialize); duke@435: duke@435: assert(adr_type == Compile::AliasIdxRaw, "only valid atp"); duke@435: assert(in(RawAddress) == rawoop, "proper init"); duke@435: // Note: allocation() can be NULL, for secondary initialization barriers duke@435: } duke@435: duke@435: // Since this node is not matched, it will be processed by the duke@435: // register allocator. Declare that there are no constraints duke@435: // on the allocation of the RawAddress edge. duke@435: const RegMask &InitializeNode::in_RegMask(uint idx) const { duke@435: // This edge should be set to top, by the set_complete. But be conservative. duke@435: if (idx == InitializeNode::RawAddress) duke@435: return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]); duke@435: return RegMask::Empty; duke@435: } duke@435: duke@435: Node* InitializeNode::memory(uint alias_idx) { duke@435: Node* mem = in(Memory); duke@435: if (mem->is_MergeMem()) { duke@435: return mem->as_MergeMem()->memory_at(alias_idx); duke@435: } else { duke@435: // incoming raw memory is not split duke@435: return mem; duke@435: } duke@435: } duke@435: duke@435: bool InitializeNode::is_non_zero() { duke@435: if (is_complete()) return false; duke@435: remove_extra_zeroes(); duke@435: return (req() > RawStores); duke@435: } duke@435: duke@435: void InitializeNode::set_complete(PhaseGVN* phase) { duke@435: assert(!is_complete(), "caller responsibility"); duke@435: _is_complete = true; duke@435: duke@435: // After this node is complete, it contains a bunch of duke@435: // raw-memory initializations. There is no need for duke@435: // it to have anything to do with non-raw memory effects. duke@435: // Therefore, tell all non-raw users to re-optimize themselves, duke@435: // after skipping the memory effects of this initialization. duke@435: PhaseIterGVN* igvn = phase->is_IterGVN(); duke@435: if (igvn) igvn->add_users_to_worklist(this); duke@435: } duke@435: duke@435: // convenience function duke@435: // return false if the init contains any stores already duke@435: bool AllocateNode::maybe_set_complete(PhaseGVN* phase) { duke@435: InitializeNode* init = initialization(); duke@435: if (init == NULL || init->is_complete()) return false; duke@435: init->remove_extra_zeroes(); duke@435: // for now, if this allocation has already collected any inits, bail: duke@435: if (init->is_non_zero()) return false; duke@435: init->set_complete(phase); duke@435: return true; duke@435: } duke@435: duke@435: void InitializeNode::remove_extra_zeroes() { duke@435: if (req() == RawStores) return; duke@435: Node* zmem = zero_memory(); duke@435: uint fill = RawStores; duke@435: for (uint i = fill; i < req(); i++) { duke@435: Node* n = in(i); duke@435: if (n->is_top() || n == zmem) continue; // skip duke@435: if (fill < i) set_req(fill, n); // compact duke@435: ++fill; duke@435: } duke@435: // delete any empty spaces created: duke@435: while (fill < req()) { duke@435: del_req(fill); duke@435: } duke@435: } duke@435: duke@435: // Helper for remembering which stores go with which offsets. duke@435: intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) { duke@435: if (!st->is_Store()) return -1; // can happen to dead code via subsume_node duke@435: intptr_t offset = -1; duke@435: Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address), duke@435: phase, offset); duke@435: if (base == NULL) return -1; // something is dead, duke@435: if (offset < 0) return -1; // dead, dead duke@435: return offset; duke@435: } duke@435: duke@435: // Helper for proving that an initialization expression is duke@435: // "simple enough" to be folded into an object initialization. duke@435: // Attempts to prove that a store's initial value 'n' can be captured duke@435: // within the initialization without creating a vicious cycle, such as: duke@435: // { Foo p = new Foo(); p.next = p; } duke@435: // True for constants and parameters and small combinations thereof. duke@435: bool InitializeNode::detect_init_independence(Node* n, duke@435: bool st_is_pinned, duke@435: int& count) { duke@435: if (n == NULL) return true; // (can this really happen?) duke@435: if (n->is_Proj()) n = n->in(0); duke@435: if (n == this) return false; // found a cycle duke@435: if (n->is_Con()) return true; duke@435: if (n->is_Start()) return true; // params, etc., are OK duke@435: if (n->is_Root()) return true; // even better duke@435: duke@435: Node* ctl = n->in(0); duke@435: if (ctl != NULL && !ctl->is_top()) { duke@435: if (ctl->is_Proj()) ctl = ctl->in(0); duke@435: if (ctl == this) return false; duke@435: duke@435: // If we already know that the enclosing memory op is pinned right after duke@435: // the init, then any control flow that the store has picked up duke@435: // must have preceded the init, or else be equal to the init. duke@435: // Even after loop optimizations (which might change control edges) duke@435: // a store is never pinned *before* the availability of its inputs. kvn@554: if (!MemNode::all_controls_dominate(n, this)) duke@435: return false; // failed to prove a good control duke@435: duke@435: } duke@435: duke@435: // Check data edges for possible dependencies on 'this'. duke@435: if ((count += 1) > 20) return false; // complexity limit duke@435: for (uint i = 1; i < n->req(); i++) { duke@435: Node* m = n->in(i); duke@435: if (m == NULL || m == n || m->is_top()) continue; duke@435: uint first_i = n->find_edge(m); duke@435: if (i != first_i) continue; // process duplicate edge just once duke@435: if (!detect_init_independence(m, st_is_pinned, count)) { duke@435: return false; duke@435: } duke@435: } duke@435: duke@435: return true; duke@435: } duke@435: duke@435: // Here are all the checks a Store must pass before it can be moved into duke@435: // an initialization. Returns zero if a check fails. duke@435: // On success, returns the (constant) offset to which the store applies, duke@435: // within the initialized memory. duke@435: intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase) { duke@435: const int FAIL = 0; duke@435: if (st->req() != MemNode::ValueIn + 1) duke@435: return FAIL; // an inscrutable StoreNode (card mark?) duke@435: Node* ctl = st->in(MemNode::Control); duke@435: if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this)) duke@435: return FAIL; // must be unconditional after the initialization duke@435: Node* mem = st->in(MemNode::Memory); duke@435: if (!(mem->is_Proj() && mem->in(0) == this)) duke@435: return FAIL; // must not be preceded by other stores duke@435: Node* adr = st->in(MemNode::Address); duke@435: intptr_t offset; duke@435: AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset); duke@435: if (alloc == NULL) duke@435: return FAIL; // inscrutable address duke@435: if (alloc != allocation()) duke@435: return FAIL; // wrong allocation! (store needs to float up) duke@435: Node* val = st->in(MemNode::ValueIn); duke@435: int complexity_count = 0; duke@435: if (!detect_init_independence(val, true, complexity_count)) duke@435: return FAIL; // stored value must be 'simple enough' duke@435: duke@435: return offset; // success duke@435: } duke@435: duke@435: // Find the captured store in(i) which corresponds to the range duke@435: // [start..start+size) in the initialized object. duke@435: // If there is one, return its index i. If there isn't, return the duke@435: // negative of the index where it should be inserted. duke@435: // Return 0 if the queried range overlaps an initialization boundary duke@435: // or if dead code is encountered. duke@435: // If size_in_bytes is zero, do not bother with overlap checks. duke@435: int InitializeNode::captured_store_insertion_point(intptr_t start, duke@435: int size_in_bytes, duke@435: PhaseTransform* phase) { duke@435: const int FAIL = 0, MAX_STORE = BytesPerLong; duke@435: duke@435: if (is_complete()) duke@435: return FAIL; // arraycopy got here first; punt duke@435: duke@435: assert(allocation() != NULL, "must be present"); duke@435: duke@435: // no negatives, no header fields: coleenp@548: if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL; duke@435: duke@435: // after a certain size, we bail out on tracking all the stores: duke@435: intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); duke@435: if (start >= ti_limit) return FAIL; duke@435: duke@435: for (uint i = InitializeNode::RawStores, limit = req(); ; ) { duke@435: if (i >= limit) return -(int)i; // not found; here is where to put it duke@435: duke@435: Node* st = in(i); duke@435: intptr_t st_off = get_store_offset(st, phase); duke@435: if (st_off < 0) { duke@435: if (st != zero_memory()) { duke@435: return FAIL; // bail out if there is dead garbage duke@435: } duke@435: } else if (st_off > start) { duke@435: // ...we are done, since stores are ordered duke@435: if (st_off < start + size_in_bytes) { duke@435: return FAIL; // the next store overlaps duke@435: } duke@435: return -(int)i; // not found; here is where to put it duke@435: } else if (st_off < start) { duke@435: if (size_in_bytes != 0 && duke@435: start < st_off + MAX_STORE && duke@435: start < st_off + st->as_Store()->memory_size()) { duke@435: return FAIL; // the previous store overlaps duke@435: } duke@435: } else { duke@435: if (size_in_bytes != 0 && duke@435: st->as_Store()->memory_size() != size_in_bytes) { duke@435: return FAIL; // mismatched store size duke@435: } duke@435: return i; duke@435: } duke@435: duke@435: ++i; duke@435: } duke@435: } duke@435: duke@435: // Look for a captured store which initializes at the offset 'start' duke@435: // with the given size. If there is no such store, and no other duke@435: // initialization interferes, then return zero_memory (the memory duke@435: // projection of the AllocateNode). duke@435: Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes, duke@435: PhaseTransform* phase) { duke@435: assert(stores_are_sane(phase), ""); duke@435: int i = captured_store_insertion_point(start, size_in_bytes, phase); duke@435: if (i == 0) { duke@435: return NULL; // something is dead duke@435: } else if (i < 0) { duke@435: return zero_memory(); // just primordial zero bits here duke@435: } else { duke@435: Node* st = in(i); // here is the store at this position duke@435: assert(get_store_offset(st->as_Store(), phase) == start, "sanity"); duke@435: return st; duke@435: } duke@435: } duke@435: duke@435: // Create, as a raw pointer, an address within my new object at 'offset'. duke@435: Node* InitializeNode::make_raw_address(intptr_t offset, duke@435: PhaseTransform* phase) { duke@435: Node* addr = in(RawAddress); duke@435: if (offset != 0) { duke@435: Compile* C = phase->C; duke@435: addr = phase->transform( new (C, 4) AddPNode(C->top(), addr, duke@435: phase->MakeConX(offset)) ); duke@435: } duke@435: return addr; duke@435: } duke@435: duke@435: // Clone the given store, converting it into a raw store duke@435: // initializing a field or element of my new object. duke@435: // Caller is responsible for retiring the original store, duke@435: // with subsume_node or the like. duke@435: // duke@435: // From the example above InitializeNode::InitializeNode, duke@435: // here are the old stores to be captured: duke@435: // store1 = (StoreC init.Control init.Memory (+ oop 12) 1) duke@435: // store2 = (StoreC init.Control store1 (+ oop 14) 2) duke@435: // duke@435: // Here is the changed code; note the extra edges on init: duke@435: // alloc = (Allocate ...) duke@435: // rawoop = alloc.RawAddress duke@435: // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1) duke@435: // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2) duke@435: // init = (Initialize alloc.Control alloc.Memory rawoop duke@435: // rawstore1 rawstore2) duke@435: // duke@435: Node* InitializeNode::capture_store(StoreNode* st, intptr_t start, duke@435: PhaseTransform* phase) { duke@435: assert(stores_are_sane(phase), ""); duke@435: duke@435: if (start < 0) return NULL; duke@435: assert(can_capture_store(st, phase) == start, "sanity"); duke@435: duke@435: Compile* C = phase->C; duke@435: int size_in_bytes = st->memory_size(); duke@435: int i = captured_store_insertion_point(start, size_in_bytes, phase); duke@435: if (i == 0) return NULL; // bail out duke@435: Node* prev_mem = NULL; // raw memory for the captured store duke@435: if (i > 0) { duke@435: prev_mem = in(i); // there is a pre-existing store under this one duke@435: set_req(i, C->top()); // temporarily disconnect it duke@435: // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect. duke@435: } else { duke@435: i = -i; // no pre-existing store duke@435: prev_mem = zero_memory(); // a slice of the newly allocated object duke@435: if (i > InitializeNode::RawStores && in(i-1) == prev_mem) duke@435: set_req(--i, C->top()); // reuse this edge; it has been folded away duke@435: else duke@435: ins_req(i, C->top()); // build a new edge duke@435: } duke@435: Node* new_st = st->clone(); duke@435: new_st->set_req(MemNode::Control, in(Control)); duke@435: new_st->set_req(MemNode::Memory, prev_mem); duke@435: new_st->set_req(MemNode::Address, make_raw_address(start, phase)); duke@435: new_st = phase->transform(new_st); duke@435: duke@435: // At this point, new_st might have swallowed a pre-existing store duke@435: // at the same offset, or perhaps new_st might have disappeared, duke@435: // if it redundantly stored the same value (or zero to fresh memory). duke@435: duke@435: // In any case, wire it in: duke@435: set_req(i, new_st); duke@435: duke@435: // The caller may now kill the old guy. duke@435: DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase)); duke@435: assert(check_st == new_st || check_st == NULL, "must be findable"); duke@435: assert(!is_complete(), ""); duke@435: return new_st; duke@435: } duke@435: duke@435: static bool store_constant(jlong* tiles, int num_tiles, duke@435: intptr_t st_off, int st_size, duke@435: jlong con) { duke@435: if ((st_off & (st_size-1)) != 0) duke@435: return false; // strange store offset (assume size==2**N) duke@435: address addr = (address)tiles + st_off; duke@435: assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob"); duke@435: switch (st_size) { duke@435: case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break; duke@435: case sizeof(jchar): *(jchar*) addr = (jchar) con; break; duke@435: case sizeof(jint): *(jint*) addr = (jint) con; break; duke@435: case sizeof(jlong): *(jlong*) addr = (jlong) con; break; duke@435: default: return false; // strange store size (detect size!=2**N here) duke@435: } duke@435: return true; // return success to caller duke@435: } duke@435: duke@435: // Coalesce subword constants into int constants and possibly duke@435: // into long constants. The goal, if the CPU permits, duke@435: // is to initialize the object with a small number of 64-bit tiles. duke@435: // Also, convert floating-point constants to bit patterns. duke@435: // Non-constants are not relevant to this pass. duke@435: // duke@435: // In terms of the running example on InitializeNode::InitializeNode duke@435: // and InitializeNode::capture_store, here is the transformation duke@435: // of rawstore1 and rawstore2 into rawstore12: duke@435: // alloc = (Allocate ...) duke@435: // rawoop = alloc.RawAddress duke@435: // tile12 = 0x00010002 duke@435: // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12) duke@435: // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12) duke@435: // duke@435: void duke@435: InitializeNode::coalesce_subword_stores(intptr_t header_size, duke@435: Node* size_in_bytes, duke@435: PhaseGVN* phase) { duke@435: Compile* C = phase->C; duke@435: duke@435: assert(stores_are_sane(phase), ""); duke@435: // Note: After this pass, they are not completely sane, duke@435: // since there may be some overlaps. duke@435: duke@435: int old_subword = 0, old_long = 0, new_int = 0, new_long = 0; duke@435: duke@435: intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize); duke@435: intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit); duke@435: size_limit = MIN2(size_limit, ti_limit); duke@435: size_limit = align_size_up(size_limit, BytesPerLong); duke@435: int num_tiles = size_limit / BytesPerLong; duke@435: duke@435: // allocate space for the tile map: duke@435: const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small duke@435: jlong tiles_buf[small_len]; duke@435: Node* nodes_buf[small_len]; duke@435: jlong inits_buf[small_len]; duke@435: jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0] duke@435: : NEW_RESOURCE_ARRAY(jlong, num_tiles)); duke@435: Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0] duke@435: : NEW_RESOURCE_ARRAY(Node*, num_tiles)); duke@435: jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0] duke@435: : NEW_RESOURCE_ARRAY(jlong, num_tiles)); duke@435: // tiles: exact bitwise model of all primitive constants duke@435: // nodes: last constant-storing node subsumed into the tiles model duke@435: // inits: which bytes (in each tile) are touched by any initializations duke@435: duke@435: //// Pass A: Fill in the tile model with any relevant stores. duke@435: duke@435: Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles); duke@435: Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles); duke@435: Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles); duke@435: Node* zmem = zero_memory(); // initially zero memory state duke@435: for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { duke@435: Node* st = in(i); duke@435: intptr_t st_off = get_store_offset(st, phase); duke@435: duke@435: // Figure out the store's offset and constant value: duke@435: if (st_off < header_size) continue; //skip (ignore header) duke@435: if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain) duke@435: int st_size = st->as_Store()->memory_size(); duke@435: if (st_off + st_size > size_limit) break; duke@435: duke@435: // Record which bytes are touched, whether by constant or not. duke@435: if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1)) duke@435: continue; // skip (strange store size) duke@435: duke@435: const Type* val = phase->type(st->in(MemNode::ValueIn)); duke@435: if (!val->singleton()) continue; //skip (non-con store) duke@435: BasicType type = val->basic_type(); duke@435: duke@435: jlong con = 0; duke@435: switch (type) { duke@435: case T_INT: con = val->is_int()->get_con(); break; duke@435: case T_LONG: con = val->is_long()->get_con(); break; duke@435: case T_FLOAT: con = jint_cast(val->getf()); break; duke@435: case T_DOUBLE: con = jlong_cast(val->getd()); break; duke@435: default: continue; //skip (odd store type) duke@435: } duke@435: duke@435: if (type == T_LONG && Matcher::isSimpleConstant64(con) && duke@435: st->Opcode() == Op_StoreL) { duke@435: continue; // This StoreL is already optimal. duke@435: } duke@435: duke@435: // Store down the constant. duke@435: store_constant(tiles, num_tiles, st_off, st_size, con); duke@435: duke@435: intptr_t j = st_off >> LogBytesPerLong; duke@435: duke@435: if (type == T_INT && st_size == BytesPerInt duke@435: && (st_off & BytesPerInt) == BytesPerInt) { duke@435: jlong lcon = tiles[j]; duke@435: if (!Matcher::isSimpleConstant64(lcon) && duke@435: st->Opcode() == Op_StoreI) { duke@435: // This StoreI is already optimal by itself. duke@435: jint* intcon = (jint*) &tiles[j]; duke@435: intcon[1] = 0; // undo the store_constant() duke@435: duke@435: // If the previous store is also optimal by itself, back up and duke@435: // undo the action of the previous loop iteration... if we can. duke@435: // But if we can't, just let the previous half take care of itself. duke@435: st = nodes[j]; duke@435: st_off -= BytesPerInt; duke@435: con = intcon[0]; duke@435: if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) { duke@435: assert(st_off >= header_size, "still ignoring header"); duke@435: assert(get_store_offset(st, phase) == st_off, "must be"); duke@435: assert(in(i-1) == zmem, "must be"); duke@435: DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn))); duke@435: assert(con == tcon->is_int()->get_con(), "must be"); duke@435: // Undo the effects of the previous loop trip, which swallowed st: duke@435: intcon[0] = 0; // undo store_constant() duke@435: set_req(i-1, st); // undo set_req(i, zmem) duke@435: nodes[j] = NULL; // undo nodes[j] = st duke@435: --old_subword; // undo ++old_subword duke@435: } duke@435: continue; // This StoreI is already optimal. duke@435: } duke@435: } duke@435: duke@435: // This store is not needed. duke@435: set_req(i, zmem); duke@435: nodes[j] = st; // record for the moment duke@435: if (st_size < BytesPerLong) // something has changed duke@435: ++old_subword; // includes int/float, but who's counting... duke@435: else ++old_long; duke@435: } duke@435: duke@435: if ((old_subword + old_long) == 0) duke@435: return; // nothing more to do duke@435: duke@435: //// Pass B: Convert any non-zero tiles into optimal constant stores. duke@435: // Be sure to insert them before overlapping non-constant stores. duke@435: // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.) duke@435: for (int j = 0; j < num_tiles; j++) { duke@435: jlong con = tiles[j]; duke@435: jlong init = inits[j]; duke@435: if (con == 0) continue; duke@435: jint con0, con1; // split the constant, address-wise duke@435: jint init0, init1; // split the init map, address-wise duke@435: { union { jlong con; jint intcon[2]; } u; duke@435: u.con = con; duke@435: con0 = u.intcon[0]; duke@435: con1 = u.intcon[1]; duke@435: u.con = init; duke@435: init0 = u.intcon[0]; duke@435: init1 = u.intcon[1]; duke@435: } duke@435: duke@435: Node* old = nodes[j]; duke@435: assert(old != NULL, "need the prior store"); duke@435: intptr_t offset = (j * BytesPerLong); duke@435: duke@435: bool split = !Matcher::isSimpleConstant64(con); duke@435: duke@435: if (offset < header_size) { duke@435: assert(offset + BytesPerInt >= header_size, "second int counts"); duke@435: assert(*(jint*)&tiles[j] == 0, "junk in header"); duke@435: split = true; // only the second word counts duke@435: // Example: int a[] = { 42 ... } duke@435: } else if (con0 == 0 && init0 == -1) { duke@435: split = true; // first word is covered by full inits duke@435: // Example: int a[] = { ... foo(), 42 ... } duke@435: } else if (con1 == 0 && init1 == -1) { duke@435: split = true; // second word is covered by full inits duke@435: // Example: int a[] = { ... 42, foo() ... } duke@435: } duke@435: duke@435: // Here's a case where init0 is neither 0 nor -1: duke@435: // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... } duke@435: // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF. duke@435: // In this case the tile is not split; it is (jlong)42. duke@435: // The big tile is stored down, and then the foo() value is inserted. duke@435: // (If there were foo(),foo() instead of foo(),0, init0 would be -1.) duke@435: duke@435: Node* ctl = old->in(MemNode::Control); duke@435: Node* adr = make_raw_address(offset, phase); duke@435: const TypePtr* atp = TypeRawPtr::BOTTOM; duke@435: duke@435: // One or two coalesced stores to plop down. duke@435: Node* st[2]; duke@435: intptr_t off[2]; duke@435: int nst = 0; duke@435: if (!split) { duke@435: ++new_long; duke@435: off[nst] = offset; coleenp@548: st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, duke@435: phase->longcon(con), T_LONG); duke@435: } else { duke@435: // Omit either if it is a zero. duke@435: if (con0 != 0) { duke@435: ++new_int; duke@435: off[nst] = offset; coleenp@548: st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, duke@435: phase->intcon(con0), T_INT); duke@435: } duke@435: if (con1 != 0) { duke@435: ++new_int; duke@435: offset += BytesPerInt; duke@435: adr = make_raw_address(offset, phase); duke@435: off[nst] = offset; coleenp@548: st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp, duke@435: phase->intcon(con1), T_INT); duke@435: } duke@435: } duke@435: duke@435: // Insert second store first, then the first before the second. duke@435: // Insert each one just before any overlapping non-constant stores. duke@435: while (nst > 0) { duke@435: Node* st1 = st[--nst]; duke@435: C->copy_node_notes_to(st1, old); duke@435: st1 = phase->transform(st1); duke@435: offset = off[nst]; duke@435: assert(offset >= header_size, "do not smash header"); duke@435: int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase); duke@435: guarantee(ins_idx != 0, "must re-insert constant store"); duke@435: if (ins_idx < 0) ins_idx = -ins_idx; // never overlap duke@435: if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem) duke@435: set_req(--ins_idx, st1); duke@435: else duke@435: ins_req(ins_idx, st1); duke@435: } duke@435: } duke@435: duke@435: if (PrintCompilation && WizardMode) duke@435: tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long", duke@435: old_subword, old_long, new_int, new_long); duke@435: if (C->log() != NULL) duke@435: C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'", duke@435: old_subword, old_long, new_int, new_long); duke@435: duke@435: // Clean up any remaining occurrences of zmem: duke@435: remove_extra_zeroes(); duke@435: } duke@435: duke@435: // Explore forward from in(start) to find the first fully initialized duke@435: // word, and return its offset. Skip groups of subword stores which duke@435: // together initialize full words. If in(start) is itself part of a duke@435: // fully initialized word, return the offset of in(start). If there duke@435: // are no following full-word stores, or if something is fishy, return duke@435: // a negative value. duke@435: intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) { duke@435: int int_map = 0; duke@435: intptr_t int_map_off = 0; duke@435: const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for duke@435: duke@435: for (uint i = start, limit = req(); i < limit; i++) { duke@435: Node* st = in(i); duke@435: duke@435: intptr_t st_off = get_store_offset(st, phase); duke@435: if (st_off < 0) break; // return conservative answer duke@435: duke@435: int st_size = st->as_Store()->memory_size(); duke@435: if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) { duke@435: return st_off; // we found a complete word init duke@435: } duke@435: duke@435: // update the map: duke@435: duke@435: intptr_t this_int_off = align_size_down(st_off, BytesPerInt); duke@435: if (this_int_off != int_map_off) { duke@435: // reset the map: duke@435: int_map = 0; duke@435: int_map_off = this_int_off; duke@435: } duke@435: duke@435: int subword_off = st_off - this_int_off; duke@435: int_map |= right_n_bits(st_size) << subword_off; duke@435: if ((int_map & FULL_MAP) == FULL_MAP) { duke@435: return this_int_off; // we found a complete word init duke@435: } duke@435: duke@435: // Did this store hit or cross the word boundary? duke@435: intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt); duke@435: if (next_int_off == this_int_off + BytesPerInt) { duke@435: // We passed the current int, without fully initializing it. duke@435: int_map_off = next_int_off; duke@435: int_map >>= BytesPerInt; duke@435: } else if (next_int_off > this_int_off + BytesPerInt) { duke@435: // We passed the current and next int. duke@435: return this_int_off + BytesPerInt; duke@435: } duke@435: } duke@435: duke@435: return -1; duke@435: } duke@435: duke@435: duke@435: // Called when the associated AllocateNode is expanded into CFG. duke@435: // At this point, we may perform additional optimizations. duke@435: // Linearize the stores by ascending offset, to make memory duke@435: // activity as coherent as possible. duke@435: Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr, duke@435: intptr_t header_size, duke@435: Node* size_in_bytes, duke@435: PhaseGVN* phase) { duke@435: assert(!is_complete(), "not already complete"); duke@435: assert(stores_are_sane(phase), ""); duke@435: assert(allocation() != NULL, "must be present"); duke@435: duke@435: remove_extra_zeroes(); duke@435: duke@435: if (ReduceFieldZeroing || ReduceBulkZeroing) duke@435: // reduce instruction count for common initialization patterns duke@435: coalesce_subword_stores(header_size, size_in_bytes, phase); duke@435: duke@435: Node* zmem = zero_memory(); // initially zero memory state duke@435: Node* inits = zmem; // accumulating a linearized chain of inits duke@435: #ifdef ASSERT coleenp@548: intptr_t first_offset = allocation()->minimum_header_size(); coleenp@548: intptr_t last_init_off = first_offset; // previous init offset coleenp@548: intptr_t last_init_end = first_offset; // previous init offset+size coleenp@548: intptr_t last_tile_end = first_offset; // previous tile offset+size duke@435: #endif duke@435: intptr_t zeroes_done = header_size; duke@435: duke@435: bool do_zeroing = true; // we might give up if inits are very sparse duke@435: int big_init_gaps = 0; // how many large gaps have we seen? duke@435: duke@435: if (ZeroTLAB) do_zeroing = false; duke@435: if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false; duke@435: duke@435: for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) { duke@435: Node* st = in(i); duke@435: intptr_t st_off = get_store_offset(st, phase); duke@435: if (st_off < 0) duke@435: break; // unknown junk in the inits duke@435: if (st->in(MemNode::Memory) != zmem) duke@435: break; // complicated store chains somehow in list duke@435: duke@435: int st_size = st->as_Store()->memory_size(); duke@435: intptr_t next_init_off = st_off + st_size; duke@435: duke@435: if (do_zeroing && zeroes_done < next_init_off) { duke@435: // See if this store needs a zero before it or under it. duke@435: intptr_t zeroes_needed = st_off; duke@435: duke@435: if (st_size < BytesPerInt) { duke@435: // Look for subword stores which only partially initialize words. duke@435: // If we find some, we must lay down some word-level zeroes first, duke@435: // underneath the subword stores. duke@435: // duke@435: // Examples: duke@435: // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s duke@435: // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y duke@435: // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z duke@435: // duke@435: // Note: coalesce_subword_stores may have already done this, duke@435: // if it was prompted by constant non-zero subword initializers. duke@435: // But this case can still arise with non-constant stores. duke@435: duke@435: intptr_t next_full_store = find_next_fullword_store(i, phase); duke@435: duke@435: // In the examples above: duke@435: // in(i) p q r s x y z duke@435: // st_off 12 13 14 15 12 13 14 duke@435: // st_size 1 1 1 1 1 1 1 duke@435: // next_full_s. 12 16 16 16 16 16 16 duke@435: // z's_done 12 16 16 16 12 16 12 duke@435: // z's_needed 12 16 16 16 16 16 16 duke@435: // zsize 0 0 0 0 4 0 4 duke@435: if (next_full_store < 0) { duke@435: // Conservative tack: Zero to end of current word. duke@435: zeroes_needed = align_size_up(zeroes_needed, BytesPerInt); duke@435: } else { duke@435: // Zero to beginning of next fully initialized word. duke@435: // Or, don't zero at all, if we are already in that word. duke@435: assert(next_full_store >= zeroes_needed, "must go forward"); duke@435: assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary"); duke@435: zeroes_needed = next_full_store; duke@435: } duke@435: } duke@435: duke@435: if (zeroes_needed > zeroes_done) { duke@435: intptr_t zsize = zeroes_needed - zeroes_done; duke@435: // Do some incremental zeroing on rawmem, in parallel with inits. duke@435: zeroes_done = align_size_down(zeroes_done, BytesPerInt); duke@435: rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, duke@435: zeroes_done, zeroes_needed, duke@435: phase); duke@435: zeroes_done = zeroes_needed; duke@435: if (zsize > Matcher::init_array_short_size && ++big_init_gaps > 2) duke@435: do_zeroing = false; // leave the hole, next time duke@435: } duke@435: } duke@435: duke@435: // Collect the store and move on: duke@435: st->set_req(MemNode::Memory, inits); duke@435: inits = st; // put it on the linearized chain duke@435: set_req(i, zmem); // unhook from previous position duke@435: duke@435: if (zeroes_done == st_off) duke@435: zeroes_done = next_init_off; duke@435: duke@435: assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any"); duke@435: duke@435: #ifdef ASSERT duke@435: // Various order invariants. Weaker than stores_are_sane because duke@435: // a large constant tile can be filled in by smaller non-constant stores. duke@435: assert(st_off >= last_init_off, "inits do not reverse"); duke@435: last_init_off = st_off; duke@435: const Type* val = NULL; duke@435: if (st_size >= BytesPerInt && duke@435: (val = phase->type(st->in(MemNode::ValueIn)))->singleton() && duke@435: (int)val->basic_type() < (int)T_OBJECT) { duke@435: assert(st_off >= last_tile_end, "tiles do not overlap"); duke@435: assert(st_off >= last_init_end, "tiles do not overwrite inits"); duke@435: last_tile_end = MAX2(last_tile_end, next_init_off); duke@435: } else { duke@435: intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong); duke@435: assert(st_tile_end >= last_tile_end, "inits stay with tiles"); duke@435: assert(st_off >= last_init_end, "inits do not overlap"); duke@435: last_init_end = next_init_off; // it's a non-tile duke@435: } duke@435: #endif //ASSERT duke@435: } duke@435: duke@435: remove_extra_zeroes(); // clear out all the zmems left over duke@435: add_req(inits); duke@435: duke@435: if (!ZeroTLAB) { duke@435: // If anything remains to be zeroed, zero it all now. duke@435: zeroes_done = align_size_down(zeroes_done, BytesPerInt); duke@435: // if it is the last unused 4 bytes of an instance, forget about it duke@435: intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint); duke@435: if (zeroes_done + BytesPerLong >= size_limit) { duke@435: assert(allocation() != NULL, ""); duke@435: Node* klass_node = allocation()->in(AllocateNode::KlassNode); duke@435: ciKlass* k = phase->type(klass_node)->is_klassptr()->klass(); duke@435: if (zeroes_done == k->layout_helper()) duke@435: zeroes_done = size_limit; duke@435: } duke@435: if (zeroes_done < size_limit) { duke@435: rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr, duke@435: zeroes_done, size_in_bytes, phase); duke@435: } duke@435: } duke@435: duke@435: set_complete(phase); duke@435: return rawmem; duke@435: } duke@435: duke@435: duke@435: #ifdef ASSERT duke@435: bool InitializeNode::stores_are_sane(PhaseTransform* phase) { duke@435: if (is_complete()) duke@435: return true; // stores could be anything at this point coleenp@548: assert(allocation() != NULL, "must be present"); coleenp@548: intptr_t last_off = allocation()->minimum_header_size(); duke@435: for (uint i = InitializeNode::RawStores; i < req(); i++) { duke@435: Node* st = in(i); duke@435: intptr_t st_off = get_store_offset(st, phase); duke@435: if (st_off < 0) continue; // ignore dead garbage duke@435: if (last_off > st_off) { duke@435: tty->print_cr("*** bad store offset at %d: %d > %d", i, last_off, st_off); duke@435: this->dump(2); duke@435: assert(false, "ascending store offsets"); duke@435: return false; duke@435: } duke@435: last_off = st_off + st->as_Store()->memory_size(); duke@435: } duke@435: return true; duke@435: } duke@435: #endif //ASSERT duke@435: duke@435: duke@435: duke@435: duke@435: //============================MergeMemNode===================================== duke@435: // duke@435: // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several duke@435: // contributing store or call operations. Each contributor provides the memory duke@435: // state for a particular "alias type" (see Compile::alias_type). For example, duke@435: // if a MergeMem has an input X for alias category #6, then any memory reference duke@435: // to alias category #6 may use X as its memory state input, as an exact equivalent duke@435: // to using the MergeMem as a whole. duke@435: // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p) duke@435: // duke@435: // (Here, the notation gives the index of the relevant adr_type.) duke@435: // duke@435: // In one special case (and more cases in the future), alias categories overlap. duke@435: // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory duke@435: // states. Therefore, if a MergeMem has only one contributing input W for Bot, duke@435: // it is exactly equivalent to that state W: duke@435: // MergeMem(: W) <==> W duke@435: // duke@435: // Usually, the merge has more than one input. In that case, where inputs duke@435: // overlap (i.e., one is Bot), the narrower alias type determines the memory duke@435: // state for that type, and the wider alias type (Bot) fills in everywhere else: duke@435: // Load<5>( MergeMem(: W, <6>: X), p ) <==> Load<5>(W,p) duke@435: // Load<6>( MergeMem(: W, <6>: X), p ) <==> Load<6>(X,p) duke@435: // duke@435: // A merge can take a "wide" memory state as one of its narrow inputs. duke@435: // This simply means that the merge observes out only the relevant parts of duke@435: // the wide input. That is, wide memory states arriving at narrow merge inputs duke@435: // are implicitly "filtered" or "sliced" as necessary. (This is rare.) duke@435: // duke@435: // These rules imply that MergeMem nodes may cascade (via their links), duke@435: // and that memory slices "leak through": duke@435: // MergeMem(: MergeMem(: W, <7>: Y)) <==> MergeMem(: W, <7>: Y) duke@435: // duke@435: // But, in such a cascade, repeated memory slices can "block the leak": duke@435: // MergeMem(: MergeMem(: W, <7>: Y), <7>: Y') <==> MergeMem(: W, <7>: Y') duke@435: // duke@435: // In the last example, Y is not part of the combined memory state of the duke@435: // outermost MergeMem. The system must, of course, prevent unschedulable duke@435: // memory states from arising, so you can be sure that the state Y is somehow duke@435: // a precursor to state Y'. duke@435: // duke@435: // duke@435: // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array duke@435: // of each MergeMemNode array are exactly the numerical alias indexes, including duke@435: // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions duke@435: // Compile::alias_type (and kin) produce and manage these indexes. duke@435: // duke@435: // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node. duke@435: // (Note that this provides quick access to the top node inside MergeMem methods, duke@435: // without the need to reach out via TLS to Compile::current.) duke@435: // duke@435: // As a consequence of what was just described, a MergeMem that represents a full duke@435: // memory state has an edge in(AliasIdxBot) which is a "wide" memory state, duke@435: // containing all alias categories. duke@435: // duke@435: // MergeMem nodes never (?) have control inputs, so in(0) is NULL. duke@435: // duke@435: // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either duke@435: // a memory state for the alias type , or else the top node, meaning that duke@435: // there is no particular input for that alias type. Note that the length of duke@435: // a MergeMem is variable, and may be extended at any time to accommodate new duke@435: // memory states at larger alias indexes. When merges grow, they are of course duke@435: // filled with "top" in the unused in() positions. duke@435: // duke@435: // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable. duke@435: // (Top was chosen because it works smoothly with passes like GCM.) duke@435: // duke@435: // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is duke@435: // the type of random VM bits like TLS references.) Since it is always the duke@435: // first non-Bot memory slice, some low-level loops use it to initialize an duke@435: // index variable: for (i = AliasIdxRaw; i < req(); i++). duke@435: // duke@435: // duke@435: // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns duke@435: // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns duke@435: // the memory state for alias type , or (if there is no particular slice at , duke@435: // it returns the base memory. To prevent bugs, memory_at does not accept duke@435: // or indexes. The iterator MergeMemStream provides robust iteration over duke@435: // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited. duke@435: // duke@435: // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't duke@435: // really that different from the other memory inputs. An abbreviation called duke@435: // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy. duke@435: // duke@435: // duke@435: // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent duke@435: // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi duke@435: // that "emerges though" the base memory will be marked as excluding the alias types duke@435: // of the other (narrow-memory) copies which "emerged through" the narrow edges: duke@435: // duke@435: // Phi(U, MergeMem(: W, <8>: Y)) duke@435: // ==Ideal=> MergeMem(: Phi(U, W), Phi<8>(U, Y)) duke@435: // duke@435: // This strange "subtraction" effect is necessary to ensure IGVN convergence. duke@435: // (It is currently unimplemented.) As you can see, the resulting merge is duke@435: // actually a disjoint union of memory states, rather than an overlay. duke@435: // duke@435: duke@435: //------------------------------MergeMemNode----------------------------------- duke@435: Node* MergeMemNode::make_empty_memory() { duke@435: Node* empty_memory = (Node*) Compile::current()->top(); duke@435: assert(empty_memory->is_top(), "correct sentinel identity"); duke@435: return empty_memory; duke@435: } duke@435: duke@435: MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) { duke@435: init_class_id(Class_MergeMem); duke@435: // all inputs are nullified in Node::Node(int) duke@435: // set_input(0, NULL); // no control input duke@435: duke@435: // Initialize the edges uniformly to top, for starters. duke@435: Node* empty_mem = make_empty_memory(); duke@435: for (uint i = Compile::AliasIdxTop; i < req(); i++) { duke@435: init_req(i,empty_mem); duke@435: } duke@435: assert(empty_memory() == empty_mem, ""); duke@435: duke@435: if( new_base != NULL && new_base->is_MergeMem() ) { duke@435: MergeMemNode* mdef = new_base->as_MergeMem(); duke@435: assert(mdef->empty_memory() == empty_mem, "consistent sentinels"); duke@435: for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) { duke@435: mms.set_memory(mms.memory2()); duke@435: } duke@435: assert(base_memory() == mdef->base_memory(), ""); duke@435: } else { duke@435: set_base_memory(new_base); duke@435: } duke@435: } duke@435: duke@435: // Make a new, untransformed MergeMem with the same base as 'mem'. duke@435: // If mem is itself a MergeMem, populate the result with the same edges. duke@435: MergeMemNode* MergeMemNode::make(Compile* C, Node* mem) { duke@435: return new(C, 1+Compile::AliasIdxRaw) MergeMemNode(mem); duke@435: } duke@435: duke@435: //------------------------------cmp-------------------------------------------- duke@435: uint MergeMemNode::hash() const { return NO_HASH; } duke@435: uint MergeMemNode::cmp( const Node &n ) const { duke@435: return (&n == this); // Always fail except on self duke@435: } duke@435: duke@435: //------------------------------Identity--------------------------------------- duke@435: Node* MergeMemNode::Identity(PhaseTransform *phase) { duke@435: // Identity if this merge point does not record any interesting memory duke@435: // disambiguations. duke@435: Node* base_mem = base_memory(); duke@435: Node* empty_mem = empty_memory(); duke@435: if (base_mem != empty_mem) { // Memory path is not dead? duke@435: for (uint i = Compile::AliasIdxRaw; i < req(); i++) { duke@435: Node* mem = in(i); duke@435: if (mem != empty_mem && mem != base_mem) { duke@435: return this; // Many memory splits; no change duke@435: } duke@435: } duke@435: } duke@435: return base_mem; // No memory splits; ID on the one true input duke@435: } duke@435: duke@435: //------------------------------Ideal------------------------------------------ duke@435: // This method is invoked recursively on chains of MergeMem nodes duke@435: Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) { duke@435: // Remove chain'd MergeMems duke@435: // duke@435: // This is delicate, because the each "in(i)" (i >= Raw) is interpreted duke@435: // relative to the "in(Bot)". Since we are patching both at the same time, duke@435: // we have to be careful to read each "in(i)" relative to the old "in(Bot)", duke@435: // but rewrite each "in(i)" relative to the new "in(Bot)". duke@435: Node *progress = NULL; duke@435: duke@435: duke@435: Node* old_base = base_memory(); duke@435: Node* empty_mem = empty_memory(); duke@435: if (old_base == empty_mem) duke@435: return NULL; // Dead memory path. duke@435: duke@435: MergeMemNode* old_mbase; duke@435: if (old_base != NULL && old_base->is_MergeMem()) duke@435: old_mbase = old_base->as_MergeMem(); duke@435: else duke@435: old_mbase = NULL; duke@435: Node* new_base = old_base; duke@435: duke@435: // simplify stacked MergeMems in base memory duke@435: if (old_mbase) new_base = old_mbase->base_memory(); duke@435: duke@435: // the base memory might contribute new slices beyond my req() duke@435: if (old_mbase) grow_to_match(old_mbase); duke@435: duke@435: // Look carefully at the base node if it is a phi. duke@435: PhiNode* phi_base; duke@435: if (new_base != NULL && new_base->is_Phi()) duke@435: phi_base = new_base->as_Phi(); duke@435: else duke@435: phi_base = NULL; duke@435: duke@435: Node* phi_reg = NULL; duke@435: uint phi_len = (uint)-1; duke@435: if (phi_base != NULL && !phi_base->is_copy()) { duke@435: // do not examine phi if degraded to a copy duke@435: phi_reg = phi_base->region(); duke@435: phi_len = phi_base->req(); duke@435: // see if the phi is unfinished duke@435: for (uint i = 1; i < phi_len; i++) { duke@435: if (phi_base->in(i) == NULL) { duke@435: // incomplete phi; do not look at it yet! duke@435: phi_reg = NULL; duke@435: phi_len = (uint)-1; duke@435: break; duke@435: } duke@435: } duke@435: } duke@435: duke@435: // Note: We do not call verify_sparse on entry, because inputs duke@435: // can normalize to the base_memory via subsume_node or similar duke@435: // mechanisms. This method repairs that damage. duke@435: duke@435: assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels"); duke@435: duke@435: // Look at each slice. duke@435: for (uint i = Compile::AliasIdxRaw; i < req(); i++) { duke@435: Node* old_in = in(i); duke@435: // calculate the old memory value duke@435: Node* old_mem = old_in; duke@435: if (old_mem == empty_mem) old_mem = old_base; duke@435: assert(old_mem == memory_at(i), ""); duke@435: duke@435: // maybe update (reslice) the old memory value duke@435: duke@435: // simplify stacked MergeMems duke@435: Node* new_mem = old_mem; duke@435: MergeMemNode* old_mmem; duke@435: if (old_mem != NULL && old_mem->is_MergeMem()) duke@435: old_mmem = old_mem->as_MergeMem(); duke@435: else duke@435: old_mmem = NULL; duke@435: if (old_mmem == this) { duke@435: // This can happen if loops break up and safepoints disappear. duke@435: // A merge of BotPtr (default) with a RawPtr memory derived from a duke@435: // safepoint can be rewritten to a merge of the same BotPtr with duke@435: // the BotPtr phi coming into the loop. If that phi disappears duke@435: // also, we can end up with a self-loop of the mergemem. duke@435: // In general, if loops degenerate and memory effects disappear, duke@435: // a mergemem can be left looking at itself. This simply means duke@435: // that the mergemem's default should be used, since there is duke@435: // no longer any apparent effect on this slice. duke@435: // Note: If a memory slice is a MergeMem cycle, it is unreachable duke@435: // from start. Update the input to TOP. duke@435: new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base; duke@435: } duke@435: else if (old_mmem != NULL) { duke@435: new_mem = old_mmem->memory_at(i); duke@435: } twisti@1040: // else preceding memory was not a MergeMem duke@435: duke@435: // replace equivalent phis (unfortunately, they do not GVN together) duke@435: if (new_mem != NULL && new_mem != new_base && duke@435: new_mem->req() == phi_len && new_mem->in(0) == phi_reg) { duke@435: if (new_mem->is_Phi()) { duke@435: PhiNode* phi_mem = new_mem->as_Phi(); duke@435: for (uint i = 1; i < phi_len; i++) { duke@435: if (phi_base->in(i) != phi_mem->in(i)) { duke@435: phi_mem = NULL; duke@435: break; duke@435: } duke@435: } duke@435: if (phi_mem != NULL) { duke@435: // equivalent phi nodes; revert to the def duke@435: new_mem = new_base; duke@435: } duke@435: } duke@435: } duke@435: duke@435: // maybe store down a new value duke@435: Node* new_in = new_mem; duke@435: if (new_in == new_base) new_in = empty_mem; duke@435: duke@435: if (new_in != old_in) { duke@435: // Warning: Do not combine this "if" with the previous "if" duke@435: // A memory slice might have be be rewritten even if it is semantically duke@435: // unchanged, if the base_memory value has changed. duke@435: set_req(i, new_in); duke@435: progress = this; // Report progress duke@435: } duke@435: } duke@435: duke@435: if (new_base != old_base) { duke@435: set_req(Compile::AliasIdxBot, new_base); duke@435: // Don't use set_base_memory(new_base), because we need to update du. duke@435: assert(base_memory() == new_base, ""); duke@435: progress = this; duke@435: } duke@435: duke@435: if( base_memory() == this ) { duke@435: // a self cycle indicates this memory path is dead duke@435: set_req(Compile::AliasIdxBot, empty_mem); duke@435: } duke@435: duke@435: // Resolve external cycles by calling Ideal on a MergeMem base_memory duke@435: // Recursion must occur after the self cycle check above duke@435: if( base_memory()->is_MergeMem() ) { duke@435: MergeMemNode *new_mbase = base_memory()->as_MergeMem(); duke@435: Node *m = phase->transform(new_mbase); // Rollup any cycles duke@435: if( m != NULL && (m->is_top() || duke@435: m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) { duke@435: // propagate rollup of dead cycle to self duke@435: set_req(Compile::AliasIdxBot, empty_mem); duke@435: } duke@435: } duke@435: duke@435: if( base_memory() == empty_mem ) { duke@435: progress = this; duke@435: // Cut inputs during Parse phase only. duke@435: // During Optimize phase a dead MergeMem node will be subsumed by Top. duke@435: if( !can_reshape ) { duke@435: for (uint i = Compile::AliasIdxRaw; i < req(); i++) { duke@435: if( in(i) != empty_mem ) { set_req(i, empty_mem); } duke@435: } duke@435: } duke@435: } duke@435: duke@435: if( !progress && base_memory()->is_Phi() && can_reshape ) { duke@435: // Check if PhiNode::Ideal's "Split phis through memory merges" duke@435: // transform should be attempted. Look for this->phi->this cycle. duke@435: uint merge_width = req(); duke@435: if (merge_width > Compile::AliasIdxRaw) { duke@435: PhiNode* phi = base_memory()->as_Phi(); duke@435: for( uint i = 1; i < phi->req(); ++i ) {// For all paths in duke@435: if (phi->in(i) == this) { duke@435: phase->is_IterGVN()->_worklist.push(phi); duke@435: break; duke@435: } duke@435: } duke@435: } duke@435: } duke@435: kvn@499: assert(progress || verify_sparse(), "please, no dups of base"); duke@435: return progress; duke@435: } duke@435: duke@435: //-------------------------set_base_memory------------------------------------- duke@435: void MergeMemNode::set_base_memory(Node *new_base) { duke@435: Node* empty_mem = empty_memory(); duke@435: set_req(Compile::AliasIdxBot, new_base); duke@435: assert(memory_at(req()) == new_base, "must set default memory"); duke@435: // Clear out other occurrences of new_base: duke@435: if (new_base != empty_mem) { duke@435: for (uint i = Compile::AliasIdxRaw; i < req(); i++) { duke@435: if (in(i) == new_base) set_req(i, empty_mem); duke@435: } duke@435: } duke@435: } duke@435: duke@435: //------------------------------out_RegMask------------------------------------ duke@435: const RegMask &MergeMemNode::out_RegMask() const { duke@435: return RegMask::Empty; duke@435: } duke@435: duke@435: //------------------------------dump_spec-------------------------------------- duke@435: #ifndef PRODUCT duke@435: void MergeMemNode::dump_spec(outputStream *st) const { duke@435: st->print(" {"); duke@435: Node* base_mem = base_memory(); duke@435: for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) { duke@435: Node* mem = memory_at(i); duke@435: if (mem == base_mem) { st->print(" -"); continue; } duke@435: st->print( " N%d:", mem->_idx ); duke@435: Compile::current()->get_adr_type(i)->dump_on(st); duke@435: } duke@435: st->print(" }"); duke@435: } duke@435: #endif // !PRODUCT duke@435: duke@435: duke@435: #ifdef ASSERT duke@435: static bool might_be_same(Node* a, Node* b) { duke@435: if (a == b) return true; duke@435: if (!(a->is_Phi() || b->is_Phi())) return false; duke@435: // phis shift around during optimization duke@435: return true; // pretty stupid... duke@435: } duke@435: duke@435: // verify a narrow slice (either incoming or outgoing) duke@435: static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) { duke@435: if (!VerifyAliases) return; // don't bother to verify unless requested duke@435: if (is_error_reported()) return; // muzzle asserts when debugging an error duke@435: if (Node::in_dump()) return; // muzzle asserts when printing duke@435: assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel"); duke@435: assert(n != NULL, ""); duke@435: // Elide intervening MergeMem's duke@435: while (n->is_MergeMem()) { duke@435: n = n->as_MergeMem()->memory_at(alias_idx); duke@435: } duke@435: Compile* C = Compile::current(); duke@435: const TypePtr* n_adr_type = n->adr_type(); duke@435: if (n == m->empty_memory()) { duke@435: // Implicit copy of base_memory() duke@435: } else if (n_adr_type != TypePtr::BOTTOM) { duke@435: assert(n_adr_type != NULL, "new memory must have a well-defined adr_type"); duke@435: assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice"); duke@435: } else { duke@435: // A few places like make_runtime_call "know" that VM calls are narrow, duke@435: // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM. duke@435: bool expected_wide_mem = false; duke@435: if (n == m->base_memory()) { duke@435: expected_wide_mem = true; duke@435: } else if (alias_idx == Compile::AliasIdxRaw || duke@435: n == m->memory_at(Compile::AliasIdxRaw)) { duke@435: expected_wide_mem = true; duke@435: } else if (!C->alias_type(alias_idx)->is_rewritable()) { duke@435: // memory can "leak through" calls on channels that duke@435: // are write-once. Allow this also. duke@435: expected_wide_mem = true; duke@435: } duke@435: assert(expected_wide_mem, "expected narrow slice replacement"); duke@435: } duke@435: } duke@435: #else // !ASSERT duke@435: #define verify_memory_slice(m,i,n) (0) // PRODUCT version is no-op duke@435: #endif duke@435: duke@435: duke@435: //-----------------------------memory_at--------------------------------------- duke@435: Node* MergeMemNode::memory_at(uint alias_idx) const { duke@435: assert(alias_idx >= Compile::AliasIdxRaw || duke@435: alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0, duke@435: "must avoid base_memory and AliasIdxTop"); duke@435: duke@435: // Otherwise, it is a narrow slice. duke@435: Node* n = alias_idx < req() ? in(alias_idx) : empty_memory(); duke@435: Compile *C = Compile::current(); duke@435: if (is_empty_memory(n)) { duke@435: // the array is sparse; empty slots are the "top" node duke@435: n = base_memory(); duke@435: assert(Node::in_dump() duke@435: || n == NULL || n->bottom_type() == Type::TOP duke@435: || n->adr_type() == TypePtr::BOTTOM duke@435: || n->adr_type() == TypeRawPtr::BOTTOM duke@435: || Compile::current()->AliasLevel() == 0, duke@435: "must be a wide memory"); duke@435: // AliasLevel == 0 if we are organizing the memory states manually. duke@435: // See verify_memory_slice for comments on TypeRawPtr::BOTTOM. duke@435: } else { duke@435: // make sure the stored slice is sane duke@435: #ifdef ASSERT duke@435: if (is_error_reported() || Node::in_dump()) { duke@435: } else if (might_be_same(n, base_memory())) { duke@435: // Give it a pass: It is a mostly harmless repetition of the base. duke@435: // This can arise normally from node subsumption during optimization. duke@435: } else { duke@435: verify_memory_slice(this, alias_idx, n); duke@435: } duke@435: #endif duke@435: } duke@435: return n; duke@435: } duke@435: duke@435: //---------------------------set_memory_at------------------------------------- duke@435: void MergeMemNode::set_memory_at(uint alias_idx, Node *n) { duke@435: verify_memory_slice(this, alias_idx, n); duke@435: Node* empty_mem = empty_memory(); duke@435: if (n == base_memory()) n = empty_mem; // collapse default duke@435: uint need_req = alias_idx+1; duke@435: if (req() < need_req) { duke@435: if (n == empty_mem) return; // already the default, so do not grow me duke@435: // grow the sparse array duke@435: do { duke@435: add_req(empty_mem); duke@435: } while (req() < need_req); duke@435: } duke@435: set_req( alias_idx, n ); duke@435: } duke@435: duke@435: duke@435: duke@435: //--------------------------iteration_setup------------------------------------ duke@435: void MergeMemNode::iteration_setup(const MergeMemNode* other) { duke@435: if (other != NULL) { duke@435: grow_to_match(other); duke@435: // invariant: the finite support of mm2 is within mm->req() duke@435: #ifdef ASSERT duke@435: for (uint i = req(); i < other->req(); i++) { duke@435: assert(other->is_empty_memory(other->in(i)), "slice left uncovered"); duke@435: } duke@435: #endif duke@435: } duke@435: // Replace spurious copies of base_memory by top. duke@435: Node* base_mem = base_memory(); duke@435: if (base_mem != NULL && !base_mem->is_top()) { duke@435: for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) { duke@435: if (in(i) == base_mem) duke@435: set_req(i, empty_memory()); duke@435: } duke@435: } duke@435: } duke@435: duke@435: //---------------------------grow_to_match------------------------------------- duke@435: void MergeMemNode::grow_to_match(const MergeMemNode* other) { duke@435: Node* empty_mem = empty_memory(); duke@435: assert(other->is_empty_memory(empty_mem), "consistent sentinels"); duke@435: // look for the finite support of the other memory duke@435: for (uint i = other->req(); --i >= req(); ) { duke@435: if (other->in(i) != empty_mem) { duke@435: uint new_len = i+1; duke@435: while (req() < new_len) add_req(empty_mem); duke@435: break; duke@435: } duke@435: } duke@435: } duke@435: duke@435: //---------------------------verify_sparse------------------------------------- duke@435: #ifndef PRODUCT duke@435: bool MergeMemNode::verify_sparse() const { duke@435: assert(is_empty_memory(make_empty_memory()), "sane sentinel"); duke@435: Node* base_mem = base_memory(); duke@435: // The following can happen in degenerate cases, since empty==top. duke@435: if (is_empty_memory(base_mem)) return true; duke@435: for (uint i = Compile::AliasIdxRaw; i < req(); i++) { duke@435: assert(in(i) != NULL, "sane slice"); duke@435: if (in(i) == base_mem) return false; // should have been the sentinel value! duke@435: } duke@435: return true; duke@435: } duke@435: duke@435: bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) { duke@435: Node* n; duke@435: n = mm->in(idx); duke@435: if (mem == n) return true; // might be empty_memory() duke@435: n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx); duke@435: if (mem == n) return true; duke@435: while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) { duke@435: if (mem == n) return true; duke@435: if (n == NULL) break; duke@435: } duke@435: return false; duke@435: } duke@435: #endif // !PRODUCT