/* * Copyright (c) 1997, 2009, Oracle and/or its affiliates. All rights reserved. * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. * * This code is free software; you can redistribute it and/or modify it * under the terms of the GNU General Public License version 2 only, as * published by the Free Software Foundation. * * This code is distributed in the hope that it will be useful, but WITHOUT * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License * version 2 for more details (a copy is included in the LICENSE file that * accompanied this code). * * You should have received a copy of the GNU General Public License version * 2 along with this work; if not, write to the Free Software Foundation, * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. * * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA * or visit www.oracle.com if you need additional information or have any * questions. * */ // Portions of code courtesy of Clifford Click // Optimization - Graph Style #include "incls/_precompiled.incl" #include "incls/_cfgnode.cpp.incl" //============================================================================= //------------------------------Value------------------------------------------ // Compute the type of the RegionNode. const Type *RegionNode::Value( PhaseTransform *phase ) const { for( uint i=1; itype(n) == Type::CONTROL ) return Type::CONTROL; } return Type::TOP; // All paths dead? Then so are we } //------------------------------Identity--------------------------------------- // Check for Region being Identity. Node *RegionNode::Identity( PhaseTransform *phase ) { // Cannot have Region be an identity, even if it has only 1 input. // Phi users cannot have their Region input folded away for them, // since they need to select the proper data input return this; } //------------------------------merge_region----------------------------------- // If a Region flows into a Region, merge into one big happy merge. This is // hard to do if there is stuff that has to happen static Node *merge_region(RegionNode *region, PhaseGVN *phase) { if( region->Opcode() != Op_Region ) // Do not do to LoopNodes return NULL; Node *progress = NULL; // Progress flag PhaseIterGVN *igvn = phase->is_IterGVN(); uint rreq = region->req(); for( uint i = 1; i < rreq; i++ ) { Node *r = region->in(i); if( r && r->Opcode() == Op_Region && // Found a region? r->in(0) == r && // Not already collapsed? r != region && // Avoid stupid situations r->outcnt() == 2 ) { // Self user and 'region' user only? assert(!r->as_Region()->has_phi(), "no phi users"); if( !progress ) { // No progress if (region->has_phi()) { return NULL; // Only flatten if no Phi users // igvn->hash_delete( phi ); } igvn->hash_delete( region ); progress = region; // Making progress } igvn->hash_delete( r ); // Append inputs to 'r' onto 'region' for( uint j = 1; j < r->req(); j++ ) { // Move an input from 'r' to 'region' region->add_req(r->in(j)); r->set_req(j, phase->C->top()); // Update phis of 'region' //for( uint k = 0; k < max; k++ ) { // Node *phi = region->out(k); // if( phi->is_Phi() ) { // phi->add_req(phi->in(i)); // } //} rreq++; // One more input to Region } // Found a region to merge into Region // Clobber pointer to the now dead 'r' region->set_req(i, phase->C->top()); } } return progress; } //--------------------------------has_phi-------------------------------------- // Helper function: Return any PhiNode that uses this region or NULL PhiNode* RegionNode::has_phi() const { for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) { Node* phi = fast_out(i); if (phi->is_Phi()) { // Check for Phi users assert(phi->in(0) == (Node*)this, "phi uses region only via in(0)"); return phi->as_Phi(); // this one is good enough } } return NULL; } //-----------------------------has_unique_phi---------------------------------- // Helper function: Return the only PhiNode that uses this region or NULL PhiNode* RegionNode::has_unique_phi() const { // Check that only one use is a Phi PhiNode* only_phi = NULL; for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) { Node* phi = fast_out(i); if (phi->is_Phi()) { // Check for Phi users assert(phi->in(0) == (Node*)this, "phi uses region only via in(0)"); if (only_phi == NULL) { only_phi = phi->as_Phi(); } else { return NULL; // multiple phis } } } return only_phi; } //------------------------------check_phi_clipping----------------------------- // Helper function for RegionNode's identification of FP clipping // Check inputs to the Phi static bool check_phi_clipping( PhiNode *phi, ConNode * &min, uint &min_idx, ConNode * &max, uint &max_idx, Node * &val, uint &val_idx ) { min = NULL; max = NULL; val = NULL; min_idx = 0; max_idx = 0; val_idx = 0; uint phi_max = phi->req(); if( phi_max == 4 ) { for( uint j = 1; j < phi_max; ++j ) { Node *n = phi->in(j); int opcode = n->Opcode(); switch( opcode ) { case Op_ConI: { if( min == NULL ) { min = n->Opcode() == Op_ConI ? (ConNode*)n : NULL; min_idx = j; } else { max = n->Opcode() == Op_ConI ? (ConNode*)n : NULL; max_idx = j; if( min->get_int() > max->get_int() ) { // Swap min and max ConNode *temp; uint temp_idx; temp = min; min = max; max = temp; temp_idx = min_idx; min_idx = max_idx; max_idx = temp_idx; } } } break; default: { val = n; val_idx = j; } break; } } } return ( min && max && val && (min->get_int() <= 0) && (max->get_int() >=0) ); } //------------------------------check_if_clipping------------------------------ // Helper function for RegionNode's identification of FP clipping // Check that inputs to Region come from two IfNodes, // // If // False True // If | // False True | // | | | // RegionNode_inputs // static bool check_if_clipping( const RegionNode *region, IfNode * &bot_if, IfNode * &top_if ) { top_if = NULL; bot_if = NULL; // Check control structure above RegionNode for (if ( if ) ) Node *in1 = region->in(1); Node *in2 = region->in(2); Node *in3 = region->in(3); // Check that all inputs are projections if( in1->is_Proj() && in2->is_Proj() && in3->is_Proj() ) { Node *in10 = in1->in(0); Node *in20 = in2->in(0); Node *in30 = in3->in(0); // Check that #1 and #2 are ifTrue and ifFalse from same If if( in10 != NULL && in10->is_If() && in20 != NULL && in20->is_If() && in30 != NULL && in30->is_If() && in10 == in20 && (in1->Opcode() != in2->Opcode()) ) { Node *in100 = in10->in(0); Node *in1000 = (in100 != NULL && in100->is_Proj()) ? in100->in(0) : NULL; // Check that control for in10 comes from other branch of IF from in3 if( in1000 != NULL && in1000->is_If() && in30 == in1000 && (in3->Opcode() != in100->Opcode()) ) { // Control pattern checks top_if = (IfNode*)in1000; bot_if = (IfNode*)in10; } } } return (top_if != NULL); } //------------------------------check_convf2i_clipping------------------------- // Helper function for RegionNode's identification of FP clipping // Verify that the value input to the phi comes from "ConvF2I; LShift; RShift" static bool check_convf2i_clipping( PhiNode *phi, uint idx, ConvF2INode * &convf2i, Node *min, Node *max) { convf2i = NULL; // Check for the RShiftNode Node *rshift = phi->in(idx); assert( rshift, "Previous checks ensure phi input is present"); if( rshift->Opcode() != Op_RShiftI ) { return false; } // Check for the LShiftNode Node *lshift = rshift->in(1); assert( lshift, "Previous checks ensure phi input is present"); if( lshift->Opcode() != Op_LShiftI ) { return false; } // Check for the ConvF2INode Node *conv = lshift->in(1); if( conv->Opcode() != Op_ConvF2I ) { return false; } // Check that shift amounts are only to get sign bits set after F2I jint max_cutoff = max->get_int(); jint min_cutoff = min->get_int(); jint left_shift = lshift->in(2)->get_int(); jint right_shift = rshift->in(2)->get_int(); jint max_post_shift = nth_bit(BitsPerJavaInteger - left_shift - 1); if( left_shift != right_shift || 0 > left_shift || left_shift >= BitsPerJavaInteger || max_post_shift < max_cutoff || max_post_shift < -min_cutoff ) { // Shifts are necessary but current transformation eliminates them return false; } // OK to return the result of ConvF2I without shifting convf2i = (ConvF2INode*)conv; return true; } //------------------------------check_compare_clipping------------------------- // Helper function for RegionNode's identification of FP clipping static bool check_compare_clipping( bool less_than, IfNode *iff, ConNode *limit, Node * & input ) { Node *i1 = iff->in(1); if ( !i1->is_Bool() ) { return false; } BoolNode *bool1 = i1->as_Bool(); if( less_than && bool1->_test._test != BoolTest::le ) { return false; } else if( !less_than && bool1->_test._test != BoolTest::lt ) { return false; } const Node *cmpF = bool1->in(1); if( cmpF->Opcode() != Op_CmpF ) { return false; } // Test that the float value being compared against // is equivalent to the int value used as a limit Node *nodef = cmpF->in(2); if( nodef->Opcode() != Op_ConF ) { return false; } jfloat conf = nodef->getf(); jint coni = limit->get_int(); if( ((int)conf) != coni ) { return false; } input = cmpF->in(1); return true; } //------------------------------is_unreachable_region-------------------------- // Find if the Region node is reachable from the root. bool RegionNode::is_unreachable_region(PhaseGVN *phase) const { assert(req() == 2, ""); // First, cut the simple case of fallthrough region when NONE of // region's phis references itself directly or through a data node. uint max = outcnt(); uint i; for (i = 0; i < max; i++) { Node* phi = raw_out(i); if (phi != NULL && phi->is_Phi()) { assert(phase->eqv(phi->in(0), this) && phi->req() == 2, ""); if (phi->outcnt() == 0) continue; // Safe case - no loops if (phi->outcnt() == 1) { Node* u = phi->raw_out(0); // Skip if only one use is an other Phi or Call or Uncommon trap. // It is safe to consider this case as fallthrough. if (u != NULL && (u->is_Phi() || u->is_CFG())) continue; } // Check when phi references itself directly or through an other node. if (phi->as_Phi()->simple_data_loop_check(phi->in(1)) >= PhiNode::Unsafe) break; // Found possible unsafe data loop. } } if (i >= max) return false; // An unsafe case was NOT found - don't need graph walk. // Unsafe case - check if the Region node is reachable from root. ResourceMark rm; Arena *a = Thread::current()->resource_area(); Node_List nstack(a); VectorSet visited(a); // Mark all control nodes reachable from root outputs Node *n = (Node*)phase->C->root(); nstack.push(n); visited.set(n->_idx); while (nstack.size() != 0) { n = nstack.pop(); uint max = n->outcnt(); for (uint i = 0; i < max; i++) { Node* m = n->raw_out(i); if (m != NULL && m->is_CFG()) { if (phase->eqv(m, this)) { return false; // We reached the Region node - it is not dead. } if (!visited.test_set(m->_idx)) nstack.push(m); } } } return true; // The Region node is unreachable - it is dead. } //------------------------------Ideal------------------------------------------ // Return a node which is more "ideal" than the current node. Must preserve // the CFG, but we can still strip out dead paths. Node *RegionNode::Ideal(PhaseGVN *phase, bool can_reshape) { if( !can_reshape && !in(0) ) return NULL; // Already degraded to a Copy assert(!in(0) || !in(0)->is_Root(), "not a specially hidden merge"); // Check for RegionNode with no Phi users and both inputs come from either // arm of the same IF. If found, then the control-flow split is useless. bool has_phis = false; if (can_reshape) { // Need DU info to check for Phi users has_phis = (has_phi() != NULL); // Cache result if (!has_phis) { // No Phi users? Nothing merging? for (uint i = 1; i < req()-1; i++) { Node *if1 = in(i); if( !if1 ) continue; Node *iff = if1->in(0); if( !iff || !iff->is_If() ) continue; for( uint j=i+1; jin(0) == iff && if1->Opcode() != in(j)->Opcode() ) { // Add the IF Projections to the worklist. They (and the IF itself) // will be eliminated if dead. phase->is_IterGVN()->add_users_to_worklist(iff); set_req(i, iff->in(0));// Skip around the useless IF diamond set_req(j, NULL); return this; // Record progress } } } } } // Remove TOP or NULL input paths. If only 1 input path remains, this Region // degrades to a copy. bool add_to_worklist = false; int cnt = 0; // Count of values merging DEBUG_ONLY( int cnt_orig = req(); ) // Save original inputs count int del_it = 0; // The last input path we delete // For all inputs... for( uint i=1; iis_Region() && n->as_Region()->is_copy() ) { set_req(i, n->nonnull_req()); i--; continue; } if( n->is_Proj() ) { // Remove useless rethrows Node *call = n->in(0); if (call->is_Call() && call->as_Call()->entry_point() == OptoRuntime::rethrow_stub()) { set_req(i, call->in(0)); i--; continue; } } if( phase->type(n) == Type::TOP ) { set_req(i, NULL); // Ignore TOP inputs i--; continue; } cnt++; // One more value merging } else if (can_reshape) { // Else found dead path with DU info PhaseIterGVN *igvn = phase->is_IterGVN(); del_req(i); // Yank path from self del_it = i; uint max = outcnt(); DUIterator j; bool progress = true; while(progress) { // Need to establish property over all users progress = false; for (j = outs(); has_out(j); j++) { Node *n = out(j); if( n->req() != req() && n->is_Phi() ) { assert( n->in(0) == this, "" ); igvn->hash_delete(n); // Yank from hash before hacking edges n->set_req_X(i,NULL,igvn);// Correct DU info n->del_req(i); // Yank path from Phis if( max != outcnt() ) { progress = true; j = refresh_out_pos(j); max = outcnt(); } } } } add_to_worklist = true; i--; } } if (can_reshape && cnt == 1) { // Is it dead loop? // If it is LoopNopde it had 2 (+1 itself) inputs and // one of them was cut. The loop is dead if it was EntryContol. assert(!this->is_Loop() || cnt_orig == 3, "Loop node should have 3 inputs"); if (this->is_Loop() && del_it == LoopNode::EntryControl || !this->is_Loop() && has_phis && is_unreachable_region(phase)) { // Yes, the region will be removed during the next step below. // Cut the backedge input and remove phis since no data paths left. // We don't cut outputs to other nodes here since we need to put them // on the worklist. del_req(1); cnt = 0; assert( req() == 1, "no more inputs expected" ); uint max = outcnt(); bool progress = true; Node *top = phase->C->top(); PhaseIterGVN *igvn = phase->is_IterGVN(); DUIterator j; while(progress) { progress = false; for (j = outs(); has_out(j); j++) { Node *n = out(j); if( n->is_Phi() ) { assert( igvn->eqv(n->in(0), this), "" ); assert( n->req() == 2 && n->in(1) != NULL, "Only one data input expected" ); // Break dead loop data path. // Eagerly replace phis with top to avoid phis copies generation. igvn->replace_node(n, top); if( max != outcnt() ) { progress = true; j = refresh_out_pos(j); max = outcnt(); } } } } add_to_worklist = true; } } if (add_to_worklist) { phase->is_IterGVN()->add_users_to_worklist(this); // Revisit collapsed Phis } if( cnt <= 1 ) { // Only 1 path in? set_req(0, NULL); // Null control input for region copy if( cnt == 0 && !can_reshape) { // Parse phase - leave the node as it is. // No inputs or all inputs are NULL. return NULL; } else if (can_reshape) { // Optimization phase - remove the node PhaseIterGVN *igvn = phase->is_IterGVN(); Node *parent_ctrl; if( cnt == 0 ) { assert( req() == 1, "no inputs expected" ); // During IGVN phase such region will be subsumed by TOP node // so region's phis will have TOP as control node. // Kill phis here to avoid it. PhiNode::is_copy() will be always false. // Also set other user's input to top. parent_ctrl = phase->C->top(); } else { // The fallthrough case since we already checked dead loops above. parent_ctrl = in(1); assert(parent_ctrl != NULL, "Region is a copy of some non-null control"); assert(!igvn->eqv(parent_ctrl, this), "Close dead loop"); } if (!add_to_worklist) igvn->add_users_to_worklist(this); // Check for further allowed opts for (DUIterator_Last imin, i = last_outs(imin); i >= imin; --i) { Node* n = last_out(i); igvn->hash_delete(n); // Remove from worklist before modifying edges if( n->is_Phi() ) { // Collapse all Phis // Eagerly replace phis to avoid copies generation. Node* in; if( cnt == 0 ) { assert( n->req() == 1, "No data inputs expected" ); in = parent_ctrl; // replaced by top } else { assert( n->req() == 2 && n->in(1) != NULL, "Only one data input expected" ); in = n->in(1); // replaced by unique input if( n->as_Phi()->is_unsafe_data_reference(in) ) in = phase->C->top(); // replaced by top } igvn->replace_node(n, in); } else if( n->is_Region() ) { // Update all incoming edges assert( !igvn->eqv(n, this), "Must be removed from DefUse edges"); uint uses_found = 0; for( uint k=1; k < n->req(); k++ ) { if( n->in(k) == this ) { n->set_req(k, parent_ctrl); uses_found++; } } if( uses_found > 1 ) { // (--i) done at the end of the loop. i -= (uses_found - 1); } } else { assert( igvn->eqv(n->in(0), this), "Expect RegionNode to be control parent"); n->set_req(0, parent_ctrl); } #ifdef ASSERT for( uint k=0; k < n->req(); k++ ) { assert( !igvn->eqv(n->in(k), this), "All uses of RegionNode should be gone"); } #endif } // Remove the RegionNode itself from DefUse info igvn->remove_dead_node(this); return NULL; } return this; // Record progress } // If a Region flows into a Region, merge into one big happy merge. if (can_reshape) { Node *m = merge_region(this, phase); if (m != NULL) return m; } // Check if this region is the root of a clipping idiom on floats if( ConvertFloat2IntClipping && can_reshape && req() == 4 ) { // Check that only one use is a Phi and that it simplifies to two constants + PhiNode* phi = has_unique_phi(); if (phi != NULL) { // One Phi user // Check inputs to the Phi ConNode *min; ConNode *max; Node *val; uint min_idx; uint max_idx; uint val_idx; if( check_phi_clipping( phi, min, min_idx, max, max_idx, val, val_idx ) ) { IfNode *top_if; IfNode *bot_if; if( check_if_clipping( this, bot_if, top_if ) ) { // Control pattern checks, now verify compares Node *top_in = NULL; // value being compared against Node *bot_in = NULL; if( check_compare_clipping( true, bot_if, min, bot_in ) && check_compare_clipping( false, top_if, max, top_in ) ) { if( bot_in == top_in ) { PhaseIterGVN *gvn = phase->is_IterGVN(); assert( gvn != NULL, "Only had DefUse info in IterGVN"); // Only remaining check is that bot_in == top_in == (Phi's val + mods) // Check for the ConvF2INode ConvF2INode *convf2i; if( check_convf2i_clipping( phi, val_idx, convf2i, min, max ) && convf2i->in(1) == bot_in ) { // Matched pattern, including LShiftI; RShiftI, replace with integer compares // max test Node *cmp = gvn->register_new_node_with_optimizer(new (phase->C, 3) CmpINode( convf2i, min )); Node *boo = gvn->register_new_node_with_optimizer(new (phase->C, 2) BoolNode( cmp, BoolTest::lt )); IfNode *iff = (IfNode*)gvn->register_new_node_with_optimizer(new (phase->C, 2) IfNode( top_if->in(0), boo, PROB_UNLIKELY_MAG(5), top_if->_fcnt )); Node *if_min= gvn->register_new_node_with_optimizer(new (phase->C, 1) IfTrueNode (iff)); Node *ifF = gvn->register_new_node_with_optimizer(new (phase->C, 1) IfFalseNode(iff)); // min test cmp = gvn->register_new_node_with_optimizer(new (phase->C, 3) CmpINode( convf2i, max )); boo = gvn->register_new_node_with_optimizer(new (phase->C, 2) BoolNode( cmp, BoolTest::gt )); iff = (IfNode*)gvn->register_new_node_with_optimizer(new (phase->C, 2) IfNode( ifF, boo, PROB_UNLIKELY_MAG(5), bot_if->_fcnt )); Node *if_max= gvn->register_new_node_with_optimizer(new (phase->C, 1) IfTrueNode (iff)); ifF = gvn->register_new_node_with_optimizer(new (phase->C, 1) IfFalseNode(iff)); // update input edges to region node set_req_X( min_idx, if_min, gvn ); set_req_X( max_idx, if_max, gvn ); set_req_X( val_idx, ifF, gvn ); // remove unnecessary 'LShiftI; RShiftI' idiom gvn->hash_delete(phi); phi->set_req_X( val_idx, convf2i, gvn ); gvn->hash_find_insert(phi); // Return transformed region node return this; } } } } } } } return NULL; } const RegMask &RegionNode::out_RegMask() const { return RegMask::Empty; } // Find the one non-null required input. RegionNode only Node *Node::nonnull_req() const { assert( is_Region(), "" ); for( uint i = 1; i < _cnt; i++ ) if( in(i) ) return in(i); ShouldNotReachHere(); return NULL; } //============================================================================= // note that these functions assume that the _adr_type field is flattened uint PhiNode::hash() const { const Type* at = _adr_type; return TypeNode::hash() + (at ? at->hash() : 0); } uint PhiNode::cmp( const Node &n ) const { return TypeNode::cmp(n) && _adr_type == ((PhiNode&)n)._adr_type; } static inline const TypePtr* flatten_phi_adr_type(const TypePtr* at) { if (at == NULL || at == TypePtr::BOTTOM) return at; return Compile::current()->alias_type(at)->adr_type(); } //----------------------------make--------------------------------------------- // create a new phi with edges matching r and set (initially) to x PhiNode* PhiNode::make(Node* r, Node* x, const Type *t, const TypePtr* at) { uint preds = r->req(); // Number of predecessor paths assert(t != Type::MEMORY || at == flatten_phi_adr_type(at), "flatten at"); PhiNode* p = new (Compile::current(), preds) PhiNode(r, t, at); for (uint j = 1; j < preds; j++) { // Fill in all inputs, except those which the region does not yet have if (r->in(j) != NULL) p->init_req(j, x); } return p; } PhiNode* PhiNode::make(Node* r, Node* x) { const Type* t = x->bottom_type(); const TypePtr* at = NULL; if (t == Type::MEMORY) at = flatten_phi_adr_type(x->adr_type()); return make(r, x, t, at); } PhiNode* PhiNode::make_blank(Node* r, Node* x) { const Type* t = x->bottom_type(); const TypePtr* at = NULL; if (t == Type::MEMORY) at = flatten_phi_adr_type(x->adr_type()); return new (Compile::current(), r->req()) PhiNode(r, t, at); } //------------------------slice_memory----------------------------------------- // create a new phi with narrowed memory type PhiNode* PhiNode::slice_memory(const TypePtr* adr_type) const { PhiNode* mem = (PhiNode*) clone(); *(const TypePtr**)&mem->_adr_type = adr_type; // convert self-loops, or else we get a bad graph for (uint i = 1; i < req(); i++) { if ((const Node*)in(i) == this) mem->set_req(i, mem); } mem->verify_adr_type(); return mem; } //------------------------split_out_instance----------------------------------- // Split out an instance type from a bottom phi. PhiNode* PhiNode::split_out_instance(const TypePtr* at, PhaseIterGVN *igvn) const { const TypeOopPtr *t_oop = at->isa_oopptr(); assert(t_oop != NULL && t_oop->is_known_instance(), "expecting instance oopptr"); const TypePtr *t = adr_type(); assert(type() == Type::MEMORY && (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM || t->isa_oopptr() && !t->is_oopptr()->is_known_instance() && t->is_oopptr()->cast_to_exactness(true) ->is_oopptr()->cast_to_ptr_type(t_oop->ptr()) ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop), "bottom or raw memory required"); // Check if an appropriate node already exists. Node *region = in(0); for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) { Node* use = region->fast_out(k); if( use->is_Phi()) { PhiNode *phi2 = use->as_Phi(); if (phi2->type() == Type::MEMORY && phi2->adr_type() == at) { return phi2; } } } Compile *C = igvn->C; Arena *a = Thread::current()->resource_area(); Node_Array node_map = new Node_Array(a); Node_Stack stack(a, C->unique() >> 4); PhiNode *nphi = slice_memory(at); igvn->register_new_node_with_optimizer( nphi ); node_map.map(_idx, nphi); stack.push((Node *)this, 1); while(!stack.is_empty()) { PhiNode *ophi = stack.node()->as_Phi(); uint i = stack.index(); assert(i >= 1, "not control edge"); stack.pop(); nphi = node_map[ophi->_idx]->as_Phi(); for (; i < ophi->req(); i++) { Node *in = ophi->in(i); if (in == NULL || igvn->type(in) == Type::TOP) continue; Node *opt = MemNode::optimize_simple_memory_chain(in, at, igvn); PhiNode *optphi = opt->is_Phi() ? opt->as_Phi() : NULL; if (optphi != NULL && optphi->adr_type() == TypePtr::BOTTOM) { opt = node_map[optphi->_idx]; if (opt == NULL) { stack.push(ophi, i); nphi = optphi->slice_memory(at); igvn->register_new_node_with_optimizer( nphi ); node_map.map(optphi->_idx, nphi); ophi = optphi; i = 0; // will get incremented at top of loop continue; } } nphi->set_req(i, opt); } } return nphi; } //------------------------verify_adr_type-------------------------------------- #ifdef ASSERT void PhiNode::verify_adr_type(VectorSet& visited, const TypePtr* at) const { if (visited.test_set(_idx)) return; //already visited // recheck constructor invariants: verify_adr_type(false); // recheck local phi/phi consistency: assert(_adr_type == at || _adr_type == TypePtr::BOTTOM, "adr_type must be consistent across phi nest"); // walk around for (uint i = 1; i < req(); i++) { Node* n = in(i); if (n == NULL) continue; const Node* np = in(i); if (np->is_Phi()) { np->as_Phi()->verify_adr_type(visited, at); } else if (n->bottom_type() == Type::TOP || (n->is_Mem() && n->in(MemNode::Address)->bottom_type() == Type::TOP)) { // ignore top inputs } else { const TypePtr* nat = flatten_phi_adr_type(n->adr_type()); // recheck phi/non-phi consistency at leaves: assert((nat != NULL) == (at != NULL), ""); assert(nat == at || nat == TypePtr::BOTTOM, "adr_type must be consistent at leaves of phi nest"); } } } // Verify a whole nest of phis rooted at this one. void PhiNode::verify_adr_type(bool recursive) const { if (is_error_reported()) return; // muzzle asserts when debugging an error if (Node::in_dump()) return; // muzzle asserts when printing assert((_type == Type::MEMORY) == (_adr_type != NULL), "adr_type for memory phis only"); if (!VerifyAliases) return; // verify thoroughly only if requested assert(_adr_type == flatten_phi_adr_type(_adr_type), "Phi::adr_type must be pre-normalized"); if (recursive) { VectorSet visited(Thread::current()->resource_area()); verify_adr_type(visited, _adr_type); } } #endif //------------------------------Value------------------------------------------ // Compute the type of the PhiNode const Type *PhiNode::Value( PhaseTransform *phase ) const { Node *r = in(0); // RegionNode if( !r ) // Copy or dead return in(1) ? phase->type(in(1)) : Type::TOP; // Note: During parsing, phis are often transformed before their regions. // This means we have to use type_or_null to defend against untyped regions. if( phase->type_or_null(r) == Type::TOP ) // Dead code? return Type::TOP; // Check for trip-counted loop. If so, be smarter. CountedLoopNode *l = r->is_CountedLoop() ? r->as_CountedLoop() : NULL; if( l && l->can_be_counted_loop(phase) && ((const Node*)l->phi() == this) ) { // Trip counted loop! // protect against init_trip() or limit() returning NULL const Node *init = l->init_trip(); const Node *limit = l->limit(); if( init != NULL && limit != NULL && l->stride_is_con() ) { const TypeInt *lo = init ->bottom_type()->isa_int(); const TypeInt *hi = limit->bottom_type()->isa_int(); if( lo && hi ) { // Dying loops might have TOP here int stride = l->stride_con(); if( stride < 0 ) { // Down-counter loop const TypeInt *tmp = lo; lo = hi; hi = tmp; stride = -stride; } if( lo->_hi < hi->_lo ) // Reversed endpoints are well defined :-( return TypeInt::make(lo->_lo,hi->_hi,3); } } } // Until we have harmony between classes and interfaces in the type // lattice, we must tread carefully around phis which implicitly // convert the one to the other. const TypePtr* ttp = _type->make_ptr(); const TypeInstPtr* ttip = (ttp != NULL) ? ttp->isa_instptr() : NULL; const TypeKlassPtr* ttkp = (ttp != NULL) ? ttp->isa_klassptr() : NULL; bool is_intf = false; if (ttip != NULL) { ciKlass* k = ttip->klass(); if (k->is_loaded() && k->is_interface()) is_intf = true; } if (ttkp != NULL) { ciKlass* k = ttkp->klass(); if (k->is_loaded() && k->is_interface()) is_intf = true; } // Default case: merge all inputs const Type *t = Type::TOP; // Merged type starting value for (uint i = 1; i < req(); ++i) {// For all paths in // Reachable control path? if (r->in(i) && phase->type(r->in(i)) == Type::CONTROL) { const Type* ti = phase->type(in(i)); // We assume that each input of an interface-valued Phi is a true // subtype of that interface. This might not be true of the meet // of all the input types. The lattice is not distributive in // such cases. Ward off asserts in type.cpp by refusing to do // meets between interfaces and proper classes. const TypePtr* tip = ti->make_ptr(); const TypeInstPtr* tiip = (tip != NULL) ? tip->isa_instptr() : NULL; if (tiip) { bool ti_is_intf = false; ciKlass* k = tiip->klass(); if (k->is_loaded() && k->is_interface()) ti_is_intf = true; if (is_intf != ti_is_intf) { t = _type; break; } } t = t->meet(ti); } } // The worst-case type (from ciTypeFlow) should be consistent with "t". // That is, we expect that "t->higher_equal(_type)" holds true. // There are various exceptions: // - Inputs which are phis might in fact be widened unnecessarily. // For example, an input might be a widened int while the phi is a short. // - Inputs might be BotPtrs but this phi is dependent on a null check, // and postCCP has removed the cast which encodes the result of the check. // - The type of this phi is an interface, and the inputs are classes. // - Value calls on inputs might produce fuzzy results. // (Occurrences of this case suggest improvements to Value methods.) // // It is not possible to see Type::BOTTOM values as phi inputs, // because the ciTypeFlow pre-pass produces verifier-quality types. const Type* ft = t->filter(_type); // Worst case type #ifdef ASSERT // The following logic has been moved into TypeOopPtr::filter. const Type* jt = t->join(_type); if( jt->empty() ) { // Emptied out??? // Check for evil case of 't' being a class and '_type' expecting an // interface. This can happen because the bytecodes do not contain // enough type info to distinguish a Java-level interface variable // from a Java-level object variable. If we meet 2 classes which // both implement interface I, but their meet is at 'j/l/O' which // doesn't implement I, we have no way to tell if the result should // be 'I' or 'j/l/O'. Thus we'll pick 'j/l/O'. If this then flows // into a Phi which "knows" it's an Interface type we'll have to // uplift the type. if( !t->empty() && ttip && ttip->is_loaded() && ttip->klass()->is_interface() ) { assert(ft == _type, ""); } // Uplift to interface else if( !t->empty() && ttkp && ttkp->is_loaded() && ttkp->klass()->is_interface() ) { assert(ft == _type, ""); } // Uplift to interface // Otherwise it's something stupid like non-overlapping int ranges // found on dying counted loops. else { assert(ft == Type::TOP, ""); } // Canonical empty value } else { // If we have an interface-typed Phi and we narrow to a class type, the join // should report back the class. However, if we have a J/L/Object // class-typed Phi and an interface flows in, it's possible that the meet & // join report an interface back out. This isn't possible but happens // because the type system doesn't interact well with interfaces. const TypePtr *jtp = jt->make_ptr(); const TypeInstPtr *jtip = (jtp != NULL) ? jtp->isa_instptr() : NULL; const TypeKlassPtr *jtkp = (jtp != NULL) ? jtp->isa_klassptr() : NULL; if( jtip && ttip ) { if( jtip->is_loaded() && jtip->klass()->is_interface() && ttip->is_loaded() && !ttip->klass()->is_interface() ) { // Happens in a CTW of rt.jar, 320-341, no extra flags assert(ft == ttip->cast_to_ptr_type(jtip->ptr()) || ft->isa_narrowoop() && ft->make_ptr() == ttip->cast_to_ptr_type(jtip->ptr()), ""); jt = ft; } } if( jtkp && ttkp ) { if( jtkp->is_loaded() && jtkp->klass()->is_interface() && !jtkp->klass_is_exact() && // Keep exact interface klass (6894807) ttkp->is_loaded() && !ttkp->klass()->is_interface() ) { assert(ft == ttkp->cast_to_ptr_type(jtkp->ptr()) || ft->isa_narrowoop() && ft->make_ptr() == ttkp->cast_to_ptr_type(jtkp->ptr()), ""); jt = ft; } } if (jt != ft && jt->base() == ft->base()) { if (jt->isa_int() && jt->is_int()->_lo == ft->is_int()->_lo && jt->is_int()->_hi == ft->is_int()->_hi) jt = ft; if (jt->isa_long() && jt->is_long()->_lo == ft->is_long()->_lo && jt->is_long()->_hi == ft->is_long()->_hi) jt = ft; } if (jt != ft) { tty->print("merge type: "); t->dump(); tty->cr(); tty->print("kill type: "); _type->dump(); tty->cr(); tty->print("join type: "); jt->dump(); tty->cr(); tty->print("filter type: "); ft->dump(); tty->cr(); } assert(jt == ft, ""); } #endif //ASSERT // Deal with conversion problems found in data loops. ft = phase->saturate(ft, phase->type_or_null(this), _type); return ft; } //------------------------------is_diamond_phi--------------------------------- // Does this Phi represent a simple well-shaped diamond merge? Return the // index of the true path or 0 otherwise. int PhiNode::is_diamond_phi() const { // Check for a 2-path merge Node *region = in(0); if( !region ) return 0; if( region->req() != 3 ) return 0; if( req() != 3 ) return 0; // Check that both paths come from the same If Node *ifp1 = region->in(1); Node *ifp2 = region->in(2); if( !ifp1 || !ifp2 ) return 0; Node *iff = ifp1->in(0); if( !iff || !iff->is_If() ) return 0; if( iff != ifp2->in(0) ) return 0; // Check for a proper bool/cmp const Node *b = iff->in(1); if( !b->is_Bool() ) return 0; const Node *cmp = b->in(1); if( !cmp->is_Cmp() ) return 0; // Check for branching opposite expected if( ifp2->Opcode() == Op_IfTrue ) { assert( ifp1->Opcode() == Op_IfFalse, "" ); return 2; } else { assert( ifp1->Opcode() == Op_IfTrue, "" ); return 1; } } //----------------------------check_cmove_id----------------------------------- // Check for CMove'ing a constant after comparing against the constant. // Happens all the time now, since if we compare equality vs a constant in // the parser, we "know" the variable is constant on one path and we force // it. Thus code like "if( x==0 ) {/*EMPTY*/}" ends up inserting a // conditional move: "x = (x==0)?0:x;". Yucko. This fix is slightly more // general in that we don't need constants. Since CMove's are only inserted // in very special circumstances, we do it here on generic Phi's. Node* PhiNode::is_cmove_id(PhaseTransform* phase, int true_path) { assert(true_path !=0, "only diamond shape graph expected"); // is_diamond_phi() has guaranteed the correctness of the nodes sequence: // phi->region->if_proj->ifnode->bool->cmp Node* region = in(0); Node* iff = region->in(1)->in(0); BoolNode* b = iff->in(1)->as_Bool(); Node* cmp = b->in(1); Node* tval = in(true_path); Node* fval = in(3-true_path); Node* id = CMoveNode::is_cmove_id(phase, cmp, tval, fval, b); if (id == NULL) return NULL; // Either value might be a cast that depends on a branch of 'iff'. // Since the 'id' value will float free of the diamond, either // decast or return failure. Node* ctl = id->in(0); if (ctl != NULL && ctl->in(0) == iff) { if (id->is_ConstraintCast()) { return id->in(1); } else { // Don't know how to disentangle this value. return NULL; } } return id; } //------------------------------Identity--------------------------------------- // Check for Region being Identity. Node *PhiNode::Identity( PhaseTransform *phase ) { // Check for no merging going on // (There used to be special-case code here when this->region->is_Loop. // It would check for a tributary phi on the backedge that the main phi // trivially, perhaps with a single cast. The unique_input method // does all this and more, by reducing such tributaries to 'this'.) Node* uin = unique_input(phase); if (uin != NULL) { return uin; } int true_path = is_diamond_phi(); if (true_path != 0) { Node* id = is_cmove_id(phase, true_path); if (id != NULL) return id; } return this; // No identity } //-----------------------------unique_input------------------------------------ // Find the unique value, discounting top, self-loops, and casts. // Return top if there are no inputs, and self if there are multiple. Node* PhiNode::unique_input(PhaseTransform* phase) { // 1) One unique direct input, or // 2) some of the inputs have an intervening ConstraintCast and // the type of input is the same or sharper (more specific) // than the phi's type. // 3) an input is a self loop // // 1) input or 2) input or 3) input __ // / \ / \ \ / \ // \ / | cast phi cast // phi \ / / \ / // phi / -- Node* r = in(0); // RegionNode if (r == NULL) return in(1); // Already degraded to a Copy Node* uncasted_input = NULL; // The unique uncasted input (ConstraintCasts removed) Node* direct_input = NULL; // The unique direct input for (uint i = 1, cnt = req(); i < cnt; ++i) { Node* rc = r->in(i); if (rc == NULL || phase->type(rc) == Type::TOP) continue; // ignore unreachable control path Node* n = in(i); if (n == NULL) continue; Node* un = n->uncast(); if (un == NULL || un == this || phase->type(un) == Type::TOP) { continue; // ignore if top, or in(i) and "this" are in a data cycle } // Check for a unique uncasted input if (uncasted_input == NULL) { uncasted_input = un; } else if (uncasted_input != un) { uncasted_input = NodeSentinel; // no unique uncasted input } // Check for a unique direct input if (direct_input == NULL) { direct_input = n; } else if (direct_input != n) { direct_input = NodeSentinel; // no unique direct input } } if (direct_input == NULL) { return phase->C->top(); // no inputs } assert(uncasted_input != NULL,""); if (direct_input != NodeSentinel) { return direct_input; // one unique direct input } if (uncasted_input != NodeSentinel && phase->type(uncasted_input)->higher_equal(type())) { return uncasted_input; // one unique uncasted input } // Nothing. return NULL; } //------------------------------is_x2logic------------------------------------- // Check for simple convert-to-boolean pattern // If:(C Bool) Region:(IfF IfT) Phi:(Region 0 1) // Convert Phi to an ConvIB. static Node *is_x2logic( PhaseGVN *phase, PhiNode *phi, int true_path ) { assert(true_path !=0, "only diamond shape graph expected"); // Convert the true/false index into an expected 0/1 return. // Map 2->0 and 1->1. int flipped = 2-true_path; // is_diamond_phi() has guaranteed the correctness of the nodes sequence: // phi->region->if_proj->ifnode->bool->cmp Node *region = phi->in(0); Node *iff = region->in(1)->in(0); BoolNode *b = (BoolNode*)iff->in(1); const CmpNode *cmp = (CmpNode*)b->in(1); Node *zero = phi->in(1); Node *one = phi->in(2); const Type *tzero = phase->type( zero ); const Type *tone = phase->type( one ); // Check for compare vs 0 const Type *tcmp = phase->type(cmp->in(2)); if( tcmp != TypeInt::ZERO && tcmp != TypePtr::NULL_PTR ) { // Allow cmp-vs-1 if the other input is bounded by 0-1 if( !(tcmp == TypeInt::ONE && phase->type(cmp->in(1)) == TypeInt::BOOL) ) return NULL; flipped = 1-flipped; // Test is vs 1 instead of 0! } // Check for setting zero/one opposite expected if( tzero == TypeInt::ZERO ) { if( tone == TypeInt::ONE ) { } else return NULL; } else if( tzero == TypeInt::ONE ) { if( tone == TypeInt::ZERO ) { flipped = 1-flipped; } else return NULL; } else return NULL; // Check for boolean test backwards if( b->_test._test == BoolTest::ne ) { } else if( b->_test._test == BoolTest::eq ) { flipped = 1-flipped; } else return NULL; // Build int->bool conversion Node *n = new (phase->C, 2) Conv2BNode( cmp->in(1) ); if( flipped ) n = new (phase->C, 3) XorINode( phase->transform(n), phase->intcon(1) ); return n; } //------------------------------is_cond_add------------------------------------ // Check for simple conditional add pattern: "(P < Q) ? X+Y : X;" // To be profitable the control flow has to disappear; there can be no other // values merging here. We replace the test-and-branch with: // "(sgn(P-Q))&Y) + X". Basically, convert "(P < Q)" into 0 or -1 by // moving the carry bit from (P-Q) into a register with 'sbb EAX,EAX'. // Then convert Y to 0-or-Y and finally add. // This is a key transform for SpecJava _201_compress. static Node* is_cond_add(PhaseGVN *phase, PhiNode *phi, int true_path) { assert(true_path !=0, "only diamond shape graph expected"); // is_diamond_phi() has guaranteed the correctness of the nodes sequence: // phi->region->if_proj->ifnode->bool->cmp RegionNode *region = (RegionNode*)phi->in(0); Node *iff = region->in(1)->in(0); BoolNode* b = iff->in(1)->as_Bool(); const CmpNode *cmp = (CmpNode*)b->in(1); // Make sure only merging this one phi here if (region->has_unique_phi() != phi) return NULL; // Make sure each arm of the diamond has exactly one output, which we assume // is the region. Otherwise, the control flow won't disappear. if (region->in(1)->outcnt() != 1) return NULL; if (region->in(2)->outcnt() != 1) return NULL; // Check for "(P < Q)" of type signed int if (b->_test._test != BoolTest::lt) return NULL; if (cmp->Opcode() != Op_CmpI) return NULL; Node *p = cmp->in(1); Node *q = cmp->in(2); Node *n1 = phi->in( true_path); Node *n2 = phi->in(3-true_path); int op = n1->Opcode(); if( op != Op_AddI // Need zero as additive identity /*&&op != Op_SubI && op != Op_AddP && op != Op_XorI && op != Op_OrI*/ ) return NULL; Node *x = n2; Node *y = n1->in(1); if( n2 == n1->in(1) ) { y = n1->in(2); } else if( n2 == n1->in(1) ) { } else return NULL; // Not so profitable if compare and add are constants if( q->is_Con() && phase->type(q) != TypeInt::ZERO && y->is_Con() ) return NULL; Node *cmplt = phase->transform( new (phase->C, 3) CmpLTMaskNode(p,q) ); Node *j_and = phase->transform( new (phase->C, 3) AndINode(cmplt,y) ); return new (phase->C, 3) AddINode(j_and,x); } //------------------------------is_absolute------------------------------------ // Check for absolute value. static Node* is_absolute( PhaseGVN *phase, PhiNode *phi_root, int true_path) { assert(true_path !=0, "only diamond shape graph expected"); int cmp_zero_idx = 0; // Index of compare input where to look for zero int phi_x_idx = 0; // Index of phi input where to find naked x // ABS ends with the merge of 2 control flow paths. // Find the false path from the true path. With only 2 inputs, 3 - x works nicely. int false_path = 3 - true_path; // is_diamond_phi() has guaranteed the correctness of the nodes sequence: // phi->region->if_proj->ifnode->bool->cmp BoolNode *bol = phi_root->in(0)->in(1)->in(0)->in(1)->as_Bool(); // Check bool sense switch( bol->_test._test ) { case BoolTest::lt: cmp_zero_idx = 1; phi_x_idx = true_path; break; case BoolTest::le: cmp_zero_idx = 2; phi_x_idx = false_path; break; case BoolTest::gt: cmp_zero_idx = 2; phi_x_idx = true_path; break; case BoolTest::ge: cmp_zero_idx = 1; phi_x_idx = false_path; break; default: return NULL; break; } // Test is next Node *cmp = bol->in(1); const Type *tzero = NULL; switch( cmp->Opcode() ) { case Op_CmpF: tzero = TypeF::ZERO; break; // Float ABS case Op_CmpD: tzero = TypeD::ZERO; break; // Double ABS default: return NULL; } // Find zero input of compare; the other input is being abs'd Node *x = NULL; bool flip = false; if( phase->type(cmp->in(cmp_zero_idx)) == tzero ) { x = cmp->in(3 - cmp_zero_idx); } else if( phase->type(cmp->in(3 - cmp_zero_idx)) == tzero ) { // The test is inverted, we should invert the result... x = cmp->in(cmp_zero_idx); flip = true; } else { return NULL; } // Next get the 2 pieces being selected, one is the original value // and the other is the negated value. if( phi_root->in(phi_x_idx) != x ) return NULL; // Check other phi input for subtract node Node *sub = phi_root->in(3 - phi_x_idx); // Allow only Sub(0,X) and fail out for all others; Neg is not OK if( tzero == TypeF::ZERO ) { if( sub->Opcode() != Op_SubF || sub->in(2) != x || phase->type(sub->in(1)) != tzero ) return NULL; x = new (phase->C, 2) AbsFNode(x); if (flip) { x = new (phase->C, 3) SubFNode(sub->in(1), phase->transform(x)); } } else { if( sub->Opcode() != Op_SubD || sub->in(2) != x || phase->type(sub->in(1)) != tzero ) return NULL; x = new (phase->C, 2) AbsDNode(x); if (flip) { x = new (phase->C, 3) SubDNode(sub->in(1), phase->transform(x)); } } return x; } //------------------------------split_once------------------------------------- // Helper for split_flow_path static void split_once(PhaseIterGVN *igvn, Node *phi, Node *val, Node *n, Node *newn) { igvn->hash_delete(n); // Remove from hash before hacking edges uint j = 1; for( uint i = phi->req()-1; i > 0; i-- ) { if( phi->in(i) == val ) { // Found a path with val? // Add to NEW Region/Phi, no DU info newn->set_req( j++, n->in(i) ); // Remove from OLD Region/Phi n->del_req(i); } } // Register the new node but do not transform it. Cannot transform until the // entire Region/Phi conglomerate has been hacked as a single huge transform. igvn->register_new_node_with_optimizer( newn ); // Now I can point to the new node. n->add_req(newn); igvn->_worklist.push(n); } //------------------------------split_flow_path-------------------------------- // Check for merging identical values and split flow paths static Node* split_flow_path(PhaseGVN *phase, PhiNode *phi) { BasicType bt = phi->type()->basic_type(); if( bt == T_ILLEGAL || type2size[bt] <= 0 ) return NULL; // Bail out on funny non-value stuff if( phi->req() <= 3 ) // Need at least 2 matched inputs and a return NULL; // third unequal input to be worth doing // Scan for a constant uint i; for( i = 1; i < phi->req()-1; i++ ) { Node *n = phi->in(i); if( !n ) return NULL; if( phase->type(n) == Type::TOP ) return NULL; if( n->Opcode() == Op_ConP || n->Opcode() == Op_ConN ) break; } if( i >= phi->req() ) // Only split for constants return NULL; Node *val = phi->in(i); // Constant to split for uint hit = 0; // Number of times it occurs for( ; i < phi->req(); i++ ){ // Count occurrences of constant Node *n = phi->in(i); if( !n ) return NULL; if( phase->type(n) == Type::TOP ) return NULL; if( phi->in(i) == val ) hit++; } if( hit <= 1 || // Make sure we find 2 or more hit == phi->req()-1 ) // and not ALL the same value return NULL; // Now start splitting out the flow paths that merge the same value. // Split first the RegionNode. PhaseIterGVN *igvn = phase->is_IterGVN(); Node *r = phi->region(); RegionNode *newr = new (phase->C, hit+1) RegionNode(hit+1); split_once(igvn, phi, val, r, newr); // Now split all other Phis than this one for (DUIterator_Fast kmax, k = r->fast_outs(kmax); k < kmax; k++) { Node* phi2 = r->fast_out(k); if( phi2->is_Phi() && phi2->as_Phi() != phi ) { PhiNode *newphi = PhiNode::make_blank(newr, phi2); split_once(igvn, phi, val, phi2, newphi); } } // Clean up this guy igvn->hash_delete(phi); for( i = phi->req()-1; i > 0; i-- ) { if( phi->in(i) == val ) { phi->del_req(i); } } phi->add_req(val); return phi; } //============================================================================= //------------------------------simple_data_loop_check------------------------- // Try to determining if the phi node in a simple safe/unsafe data loop. // Returns: // enum LoopSafety { Safe = 0, Unsafe, UnsafeLoop }; // Safe - safe case when the phi and it's inputs reference only safe data // nodes; // Unsafe - the phi and it's inputs reference unsafe data nodes but there // is no reference back to the phi - need a graph walk // to determine if it is in a loop; // UnsafeLoop - unsafe case when the phi references itself directly or through // unsafe data node. // Note: a safe data node is a node which could/never reference itself during // GVN transformations. For now it is Con, Proj, Phi, CastPP, CheckCastPP. // I mark Phi nodes as safe node not only because they can reference itself // but also to prevent mistaking the fallthrough case inside an outer loop // as dead loop when the phi references itselfs through an other phi. PhiNode::LoopSafety PhiNode::simple_data_loop_check(Node *in) const { // It is unsafe loop if the phi node references itself directly. if (in == (Node*)this) return UnsafeLoop; // Unsafe loop // Unsafe loop if the phi node references itself through an unsafe data node. // Exclude cases with null inputs or data nodes which could reference // itself (safe for dead loops). if (in != NULL && !in->is_dead_loop_safe()) { // Check inputs of phi's inputs also. // It is much less expensive then full graph walk. uint cnt = in->req(); uint i = (in->is_Proj() && !in->is_CFG()) ? 0 : 1; for (; i < cnt; ++i) { Node* m = in->in(i); if (m == (Node*)this) return UnsafeLoop; // Unsafe loop if (m != NULL && !m->is_dead_loop_safe()) { // Check the most common case (about 30% of all cases): // phi->Load/Store->AddP->(ConP ConP Con)/(Parm Parm Con). Node *m1 = (m->is_AddP() && m->req() > 3) ? m->in(1) : NULL; if (m1 == (Node*)this) return UnsafeLoop; // Unsafe loop if (m1 != NULL && m1 == m->in(2) && m1->is_dead_loop_safe() && m->in(3)->is_Con()) { continue; // Safe case } // The phi references an unsafe node - need full analysis. return Unsafe; } } } return Safe; // Safe case - we can optimize the phi node. } //------------------------------is_unsafe_data_reference----------------------- // If phi can be reached through the data input - it is data loop. bool PhiNode::is_unsafe_data_reference(Node *in) const { assert(req() > 1, ""); // First, check simple cases when phi references itself directly or // through an other node. LoopSafety safety = simple_data_loop_check(in); if (safety == UnsafeLoop) return true; // phi references itself - unsafe loop else if (safety == Safe) return false; // Safe case - phi could be replaced with the unique input. // Unsafe case when we should go through data graph to determine // if the phi references itself. ResourceMark rm; Arena *a = Thread::current()->resource_area(); Node_List nstack(a); VectorSet visited(a); nstack.push(in); // Start with unique input. visited.set(in->_idx); while (nstack.size() != 0) { Node* n = nstack.pop(); uint cnt = n->req(); uint i = (n->is_Proj() && !n->is_CFG()) ? 0 : 1; for (; i < cnt; i++) { Node* m = n->in(i); if (m == (Node*)this) { return true; // Data loop } if (m != NULL && !m->is_dead_loop_safe()) { // Only look for unsafe cases. if (!visited.test_set(m->_idx)) nstack.push(m); } } } return false; // The phi is not reachable from its inputs } //------------------------------Ideal------------------------------------------ // Return a node which is more "ideal" than the current node. Must preserve // the CFG, but we can still strip out dead paths. Node *PhiNode::Ideal(PhaseGVN *phase, bool can_reshape) { // The next should never happen after 6297035 fix. if( is_copy() ) // Already degraded to a Copy ? return NULL; // No change Node *r = in(0); // RegionNode assert(r->in(0) == NULL || !r->in(0)->is_Root(), "not a specially hidden merge"); // Note: During parsing, phis are often transformed before their regions. // This means we have to use type_or_null to defend against untyped regions. if( phase->type_or_null(r) == Type::TOP ) // Dead code? return NULL; // No change Node *top = phase->C->top(); bool new_phi = (outcnt() == 0); // transforming new Phi assert(!can_reshape || !new_phi, "for igvn new phi should be hooked"); // The are 2 situations when only one valid phi's input is left // (in addition to Region input). // One: region is not loop - replace phi with this input. // Two: region is loop - replace phi with top since this data path is dead // and we need to break the dead data loop. Node* progress = NULL; // Record if any progress made for( uint j = 1; j < req(); ++j ){ // For all paths in // Check unreachable control paths Node* rc = r->in(j); Node* n = in(j); // Get the input if (rc == NULL || phase->type(rc) == Type::TOP) { if (n != top) { // Not already top? set_req(j, top); // Nuke it down progress = this; // Record progress } } } if (can_reshape && outcnt() == 0) { // set_req() above may kill outputs if Phi is referenced // only by itself on the dead (top) control path. return top; } Node* uin = unique_input(phase); if (uin == top) { // Simplest case: no alive inputs. if (can_reshape) // IGVN transformation return top; else return NULL; // Identity will return TOP } else if (uin != NULL) { // Only one not-NULL unique input path is left. // Determine if this input is backedge of a loop. // (Skip new phis which have no uses and dead regions). if( outcnt() > 0 && r->in(0) != NULL ) { // First, take the short cut when we know it is a loop and // the EntryControl data path is dead. assert(!r->is_Loop() || r->req() == 3, "Loop node should have 3 inputs"); // Then, check if there is a data loop when phi references itself directly // or through other data nodes. if( r->is_Loop() && !phase->eqv_uncast(uin, in(LoopNode::EntryControl)) || !r->is_Loop() && is_unsafe_data_reference(uin) ) { // Break this data loop to avoid creation of a dead loop. if (can_reshape) { return top; } else { // We can't return top if we are in Parse phase - cut inputs only // let Identity to handle the case. replace_edge(uin, top); return NULL; } } } // One unique input. debug_only(Node* ident = Identity(phase)); // The unique input must eventually be detected by the Identity call. #ifdef ASSERT if (ident != uin && !ident->is_top()) { // print this output before failing assert r->dump(3); this->dump(3); ident->dump(); uin->dump(); } #endif assert(ident == uin || ident->is_top(), "Identity must clean this up"); return NULL; } Node* opt = NULL; int true_path = is_diamond_phi(); if( true_path != 0 ) { // Check for CMove'ing identity. If it would be unsafe, // handle it here. In the safe case, let Identity handle it. Node* unsafe_id = is_cmove_id(phase, true_path); if( unsafe_id != NULL && is_unsafe_data_reference(unsafe_id) ) opt = unsafe_id; // Check for simple convert-to-boolean pattern if( opt == NULL ) opt = is_x2logic(phase, this, true_path); // Check for absolute value if( opt == NULL ) opt = is_absolute(phase, this, true_path); // Check for conditional add if( opt == NULL && can_reshape ) opt = is_cond_add(phase, this, true_path); // These 4 optimizations could subsume the phi: // have to check for a dead data loop creation. if( opt != NULL ) { if( opt == unsafe_id || is_unsafe_data_reference(opt) ) { // Found dead loop. if( can_reshape ) return top; // We can't return top if we are in Parse phase - cut inputs only // to stop further optimizations for this phi. Identity will return TOP. assert(req() == 3, "only diamond merge phi here"); set_req(1, top); set_req(2, top); return NULL; } else { return opt; } } } // Check for merging identical values and split flow paths if (can_reshape) { opt = split_flow_path(phase, this); // This optimization only modifies phi - don't need to check for dead loop. assert(opt == NULL || phase->eqv(opt, this), "do not elide phi"); if (opt != NULL) return opt; } if (in(1) != NULL && in(1)->Opcode() == Op_AddP && can_reshape) { // Try to undo Phi of AddP: // (Phi (AddP base base y) (AddP base2 base2 y)) // becomes: // newbase := (Phi base base2) // (AddP newbase newbase y) // // This occurs as a result of unsuccessful split_thru_phi and // interferes with taking advantage of addressing modes. See the // clone_shift_expressions code in matcher.cpp Node* addp = in(1); const Type* type = addp->in(AddPNode::Base)->bottom_type(); Node* y = addp->in(AddPNode::Offset); if (y != NULL && addp->in(AddPNode::Base) == addp->in(AddPNode::Address)) { // make sure that all the inputs are similar to the first one, // i.e. AddP with base == address and same offset as first AddP bool doit = true; for (uint i = 2; i < req(); i++) { if (in(i) == NULL || in(i)->Opcode() != Op_AddP || in(i)->in(AddPNode::Base) != in(i)->in(AddPNode::Address) || in(i)->in(AddPNode::Offset) != y) { doit = false; break; } // Accumulate type for resulting Phi type = type->meet(in(i)->in(AddPNode::Base)->bottom_type()); } Node* base = NULL; if (doit) { // Check for neighboring AddP nodes in a tree. // If they have a base, use that it. for (DUIterator_Fast kmax, k = this->fast_outs(kmax); k < kmax; k++) { Node* u = this->fast_out(k); if (u->is_AddP()) { Node* base2 = u->in(AddPNode::Base); if (base2 != NULL && !base2->is_top()) { if (base == NULL) base = base2; else if (base != base2) { doit = false; break; } } } } } if (doit) { if (base == NULL) { base = new (phase->C, in(0)->req()) PhiNode(in(0), type, NULL); for (uint i = 1; i < req(); i++) { base->init_req(i, in(i)->in(AddPNode::Base)); } phase->is_IterGVN()->register_new_node_with_optimizer(base); } return new (phase->C, 4) AddPNode(base, base, y); } } } // Split phis through memory merges, so that the memory merges will go away. // Piggy-back this transformation on the search for a unique input.... // It will be as if the merged memory is the unique value of the phi. // (Do not attempt this optimization unless parsing is complete. // It would make the parser's memory-merge logic sick.) // (MergeMemNode is not dead_loop_safe - need to check for dead loop.) if (progress == NULL && can_reshape && type() == Type::MEMORY) { // see if this phi should be sliced uint merge_width = 0; bool saw_self = false; for( uint i=1; iis_MergeMem()) { MergeMemNode* n = ii->as_MergeMem(); merge_width = MAX2(merge_width, n->req()); saw_self = saw_self || phase->eqv(n->base_memory(), this); } } // This restriction is temporarily necessary to ensure termination: if (!saw_self && adr_type() == TypePtr::BOTTOM) merge_width = 0; if (merge_width > Compile::AliasIdxRaw) { // found at least one non-empty MergeMem const TypePtr* at = adr_type(); if (at != TypePtr::BOTTOM) { // Patch the existing phi to select an input from the merge: // Phi:AT1(...MergeMem(m0, m1, m2)...) into // Phi:AT1(...m1...) int alias_idx = phase->C->get_alias_index(at); for (uint i=1; iis_MergeMem()) { MergeMemNode* n = ii->as_MergeMem(); // compress paths and change unreachable cycles to TOP // If not, we can update the input infinitely along a MergeMem cycle // Equivalent code is in MemNode::Ideal_common Node *m = phase->transform(n); if (outcnt() == 0) { // Above transform() may kill us! return top; } // If transformed to a MergeMem, get the desired slice // Otherwise the returned node represents memory for every slice Node *new_mem = (m->is_MergeMem()) ? m->as_MergeMem()->memory_at(alias_idx) : m; // Update input if it is progress over what we have now if (new_mem != ii) { set_req(i, new_mem); progress = this; } } } } else { // We know that at least one MergeMem->base_memory() == this // (saw_self == true). If all other inputs also references this phi // (directly or through data nodes) - it is dead loop. bool saw_safe_input = false; for (uint j = 1; j < req(); ++j) { Node *n = in(j); if (n->is_MergeMem() && n->as_MergeMem()->base_memory() == this) continue; // skip known cases if (!is_unsafe_data_reference(n)) { saw_safe_input = true; // found safe input break; } } if (!saw_safe_input) return top; // all inputs reference back to this phi - dead loop // Phi(...MergeMem(m0, m1:AT1, m2:AT2)...) into // MergeMem(Phi(...m0...), Phi:AT1(...m1...), Phi:AT2(...m2...)) PhaseIterGVN *igvn = phase->is_IterGVN(); Node* hook = new (phase->C, 1) Node(1); PhiNode* new_base = (PhiNode*) clone(); // Must eagerly register phis, since they participate in loops. if (igvn) { igvn->register_new_node_with_optimizer(new_base); hook->add_req(new_base); } MergeMemNode* result = MergeMemNode::make(phase->C, new_base); for (uint i = 1; i < req(); ++i) { Node *ii = in(i); if (ii->is_MergeMem()) { MergeMemNode* n = ii->as_MergeMem(); for (MergeMemStream mms(result, n); mms.next_non_empty2(); ) { // If we have not seen this slice yet, make a phi for it. bool made_new_phi = false; if (mms.is_empty()) { Node* new_phi = new_base->slice_memory(mms.adr_type(phase->C)); made_new_phi = true; if (igvn) { igvn->register_new_node_with_optimizer(new_phi); hook->add_req(new_phi); } mms.set_memory(new_phi); } Node* phi = mms.memory(); assert(made_new_phi || phi->in(i) == n, "replace the i-th merge by a slice"); phi->set_req(i, mms.memory2()); } } } // Distribute all self-loops. { // (Extra braces to hide mms.) for (MergeMemStream mms(result); mms.next_non_empty(); ) { Node* phi = mms.memory(); for (uint i = 1; i < req(); ++i) { if (phi->in(i) == this) phi->set_req(i, phi); } } } // now transform the new nodes, and return the mergemem for (MergeMemStream mms(result); mms.next_non_empty(); ) { Node* phi = mms.memory(); mms.set_memory(phase->transform(phi)); } if (igvn) { // Unhook. igvn->hash_delete(hook); for (uint i = 1; i < hook->req(); i++) { hook->set_req(i, NULL); } } // Replace self with the result. return result; } } // // Other optimizations on the memory chain // const TypePtr* at = adr_type(); for( uint i=1; iis_DecodeN() && ii->bottom_type() == bottom_type()) { // Do optimization if a non dead path exist. if (ii->in(1)->bottom_type() != Type::TOP) { has_decodeN = true; } } else if (!ii->is_Phi()) { may_push = false; } } if (has_decodeN && may_push) { PhaseIterGVN *igvn = phase->is_IterGVN(); // Make narrow type for new phi. const Type* narrow_t = TypeNarrowOop::make(this->bottom_type()->is_ptr()); PhiNode* new_phi = new (phase->C, r->req()) PhiNode(r, narrow_t); uint orig_cnt = req(); for (uint i=1; iis_DecodeN()) { assert(ii->bottom_type() == bottom_type(), "sanity"); new_ii = ii->in(1); } else { assert(ii->is_Phi(), "sanity"); if (ii->as_Phi() == this) { new_ii = new_phi; } else { new_ii = new (phase->C, 2) EncodePNode(ii, narrow_t); igvn->register_new_node_with_optimizer(new_ii); } } new_phi->set_req(i, new_ii); } igvn->register_new_node_with_optimizer(new_phi, this); progress = new (phase->C, 2) DecodeNNode(new_phi, bottom_type()); } } #endif return progress; // Return any progress } //------------------------------is_tripcount----------------------------------- bool PhiNode::is_tripcount() const { return (in(0) != NULL && in(0)->is_CountedLoop() && in(0)->as_CountedLoop()->phi() == this); } //------------------------------out_RegMask------------------------------------ const RegMask &PhiNode::in_RegMask(uint i) const { return i ? out_RegMask() : RegMask::Empty; } const RegMask &PhiNode::out_RegMask() const { uint ideal_reg = Matcher::base2reg[_type->base()]; assert( ideal_reg != Node::NotAMachineReg, "invalid type at Phi" ); if( ideal_reg == 0 ) return RegMask::Empty; return *(Compile::current()->matcher()->idealreg2spillmask[ideal_reg]); } #ifndef PRODUCT void PhiNode::dump_spec(outputStream *st) const { TypeNode::dump_spec(st); if (is_tripcount()) { st->print(" #tripcount"); } } #endif //============================================================================= const Type *GotoNode::Value( PhaseTransform *phase ) const { // If the input is reachable, then we are executed. // If the input is not reachable, then we are not executed. return phase->type(in(0)); } Node *GotoNode::Identity( PhaseTransform *phase ) { return in(0); // Simple copy of incoming control } const RegMask &GotoNode::out_RegMask() const { return RegMask::Empty; } //============================================================================= const RegMask &JumpNode::out_RegMask() const { return RegMask::Empty; } //============================================================================= const RegMask &JProjNode::out_RegMask() const { return RegMask::Empty; } //============================================================================= const RegMask &CProjNode::out_RegMask() const { return RegMask::Empty; } //============================================================================= uint PCTableNode::hash() const { return Node::hash() + _size; } uint PCTableNode::cmp( const Node &n ) const { return _size == ((PCTableNode&)n)._size; } const Type *PCTableNode::bottom_type() const { const Type** f = TypeTuple::fields(_size); for( uint i = 0; i < _size; i++ ) f[i] = Type::CONTROL; return TypeTuple::make(_size, f); } //------------------------------Value------------------------------------------ // Compute the type of the PCTableNode. If reachable it is a tuple of // Control, otherwise the table targets are not reachable const Type *PCTableNode::Value( PhaseTransform *phase ) const { if( phase->type(in(0)) == Type::CONTROL ) return bottom_type(); return Type::TOP; // All paths dead? Then so are we } //------------------------------Ideal------------------------------------------ // Return a node which is more "ideal" than the current node. Strip out // control copies Node *PCTableNode::Ideal(PhaseGVN *phase, bool can_reshape) { return remove_dead_region(phase, can_reshape) ? this : NULL; } //============================================================================= uint JumpProjNode::hash() const { return Node::hash() + _dest_bci; } uint JumpProjNode::cmp( const Node &n ) const { return ProjNode::cmp(n) && _dest_bci == ((JumpProjNode&)n)._dest_bci; } #ifndef PRODUCT void JumpProjNode::dump_spec(outputStream *st) const { ProjNode::dump_spec(st); st->print("@bci %d ",_dest_bci); } #endif //============================================================================= //------------------------------Value------------------------------------------ // Check for being unreachable, or for coming from a Rethrow. Rethrow's cannot // have the default "fall_through_index" path. const Type *CatchNode::Value( PhaseTransform *phase ) const { // Unreachable? Then so are all paths from here. if( phase->type(in(0)) == Type::TOP ) return Type::TOP; // First assume all paths are reachable const Type** f = TypeTuple::fields(_size); for( uint i = 0; i < _size; i++ ) f[i] = Type::CONTROL; // Identify cases that will always throw an exception // () rethrow call // () virtual or interface call with NULL receiver // () call is a check cast with incompatible arguments if( in(1)->is_Proj() ) { Node *i10 = in(1)->in(0); if( i10->is_Call() ) { CallNode *call = i10->as_Call(); // Rethrows always throw exceptions, never return if (call->entry_point() == OptoRuntime::rethrow_stub()) { f[CatchProjNode::fall_through_index] = Type::TOP; } else if( call->req() > TypeFunc::Parms ) { const Type *arg0 = phase->type( call->in(TypeFunc::Parms) ); // Check for null receiver to virtual or interface calls if( call->is_CallDynamicJava() && arg0->higher_equal(TypePtr::NULL_PTR) ) { f[CatchProjNode::fall_through_index] = Type::TOP; } } // End of if not a runtime stub } // End of if have call above me } // End of slot 1 is not a projection return TypeTuple::make(_size, f); } //============================================================================= uint CatchProjNode::hash() const { return Node::hash() + _handler_bci; } uint CatchProjNode::cmp( const Node &n ) const { return ProjNode::cmp(n) && _handler_bci == ((CatchProjNode&)n)._handler_bci; } //------------------------------Identity--------------------------------------- // If only 1 target is possible, choose it if it is the main control Node *CatchProjNode::Identity( PhaseTransform *phase ) { // If my value is control and no other value is, then treat as ID const TypeTuple *t = phase->type(in(0))->is_tuple(); if (t->field_at(_con) != Type::CONTROL) return this; // If we remove the last CatchProj and elide the Catch/CatchProj, then we // also remove any exception table entry. Thus we must know the call // feeding the Catch will not really throw an exception. This is ok for // the main fall-thru control (happens when we know a call can never throw // an exception) or for "rethrow", because a further optimization will // yank the rethrow (happens when we inline a function that can throw an // exception and the caller has no handler). Not legal, e.g., for passing // a NULL receiver to a v-call, or passing bad types to a slow-check-cast. // These cases MUST throw an exception via the runtime system, so the VM // will be looking for a table entry. Node *proj = in(0)->in(1); // Expect a proj feeding CatchNode CallNode *call; if (_con != TypeFunc::Control && // Bail out if not the main control. !(proj->is_Proj() && // AND NOT a rethrow proj->in(0)->is_Call() && (call = proj->in(0)->as_Call()) && call->entry_point() == OptoRuntime::rethrow_stub())) return this; // Search for any other path being control for (uint i = 0; i < t->cnt(); i++) { if (i != _con && t->field_at(i) == Type::CONTROL) return this; } // Only my path is possible; I am identity on control to the jump return in(0)->in(0); } #ifndef PRODUCT void CatchProjNode::dump_spec(outputStream *st) const { ProjNode::dump_spec(st); st->print("@bci %d ",_handler_bci); } #endif //============================================================================= //------------------------------Identity--------------------------------------- // Check for CreateEx being Identity. Node *CreateExNode::Identity( PhaseTransform *phase ) { if( phase->type(in(1)) == Type::TOP ) return in(1); if( phase->type(in(0)) == Type::TOP ) return in(0); // We only come from CatchProj, unless the CatchProj goes away. // If the CatchProj is optimized away, then we just carry the // exception oop through. CallNode *call = in(1)->in(0)->as_Call(); return ( in(0)->is_CatchProj() && in(0)->in(0)->in(1) == in(1) ) ? this : call->in(TypeFunc::Parms); } //============================================================================= //------------------------------Value------------------------------------------ // Check for being unreachable. const Type *NeverBranchNode::Value( PhaseTransform *phase ) const { if (!in(0) || in(0)->is_top()) return Type::TOP; return bottom_type(); } //------------------------------Ideal------------------------------------------ // Check for no longer being part of a loop Node *NeverBranchNode::Ideal(PhaseGVN *phase, bool can_reshape) { if (can_reshape && !in(0)->is_Loop()) { // Dead code elimination can sometimes delete this projection so // if it's not there, there's nothing to do. Node* fallthru = proj_out(0); if (fallthru != NULL) { phase->is_IterGVN()->replace_node(fallthru, in(0)); } return phase->C->top(); } return NULL; } #ifndef PRODUCT void NeverBranchNode::format( PhaseRegAlloc *ra_, outputStream *st) const { st->print("%s", Name()); } #endif