duke@435: /* xdono@631: * Copyright 1998-2008 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: #include "incls/_precompiled.incl" duke@435: #include "incls/_loopnode.cpp.incl" duke@435: duke@435: //============================================================================= duke@435: //------------------------------is_loop_iv------------------------------------- duke@435: // Determine if a node is Counted loop induction variable. duke@435: // The method is declared in node.hpp. duke@435: const Node* Node::is_loop_iv() const { duke@435: if (this->is_Phi() && !this->as_Phi()->is_copy() && duke@435: this->as_Phi()->region()->is_CountedLoop() && duke@435: this->as_Phi()->region()->as_CountedLoop()->phi() == this) { duke@435: return this; duke@435: } else { duke@435: return NULL; duke@435: } duke@435: } duke@435: duke@435: //============================================================================= duke@435: //------------------------------dump_spec-------------------------------------- duke@435: // Dump special per-node info duke@435: #ifndef PRODUCT duke@435: void LoopNode::dump_spec(outputStream *st) const { duke@435: if( is_inner_loop () ) st->print( "inner " ); duke@435: if( is_partial_peel_loop () ) st->print( "partial_peel " ); duke@435: if( partial_peel_has_failed () ) st->print( "partial_peel_failed " ); duke@435: } duke@435: #endif duke@435: duke@435: //------------------------------get_early_ctrl--------------------------------- duke@435: // Compute earliest legal control duke@435: Node *PhaseIdealLoop::get_early_ctrl( Node *n ) { duke@435: assert( !n->is_Phi() && !n->is_CFG(), "this code only handles data nodes" ); duke@435: uint i; duke@435: Node *early; duke@435: if( n->in(0) ) { duke@435: early = n->in(0); duke@435: if( !early->is_CFG() ) // Might be a non-CFG multi-def duke@435: early = get_ctrl(early); // So treat input as a straight data input duke@435: i = 1; duke@435: } else { duke@435: early = get_ctrl(n->in(1)); duke@435: i = 2; duke@435: } duke@435: uint e_d = dom_depth(early); duke@435: assert( early, "" ); duke@435: for( ; i < n->req(); i++ ) { duke@435: Node *cin = get_ctrl(n->in(i)); duke@435: assert( cin, "" ); duke@435: // Keep deepest dominator depth duke@435: uint c_d = dom_depth(cin); duke@435: if( c_d > e_d ) { // Deeper guy? duke@435: early = cin; // Keep deepest found so far duke@435: e_d = c_d; duke@435: } else if( c_d == e_d && // Same depth? duke@435: early != cin ) { // If not equal, must use slower algorithm duke@435: // If same depth but not equal, one _must_ dominate the other duke@435: // and we want the deeper (i.e., dominated) guy. duke@435: Node *n1 = early; duke@435: Node *n2 = cin; duke@435: while( 1 ) { duke@435: n1 = idom(n1); // Walk up until break cycle duke@435: n2 = idom(n2); duke@435: if( n1 == cin || // Walked early up to cin duke@435: dom_depth(n2) < c_d ) duke@435: break; // early is deeper; keep him duke@435: if( n2 == early || // Walked cin up to early duke@435: dom_depth(n1) < c_d ) { duke@435: early = cin; // cin is deeper; keep him duke@435: break; duke@435: } duke@435: } duke@435: e_d = dom_depth(early); // Reset depth register cache duke@435: } duke@435: } duke@435: duke@435: // Return earliest legal location duke@435: assert(early == find_non_split_ctrl(early), "unexpected early control"); duke@435: duke@435: return early; duke@435: } duke@435: duke@435: //------------------------------set_early_ctrl--------------------------------- duke@435: // Set earliest legal control duke@435: void PhaseIdealLoop::set_early_ctrl( Node *n ) { duke@435: Node *early = get_early_ctrl(n); duke@435: duke@435: // Record earliest legal location duke@435: set_ctrl(n, early); duke@435: } duke@435: duke@435: //------------------------------set_subtree_ctrl------------------------------- duke@435: // set missing _ctrl entries on new nodes duke@435: void PhaseIdealLoop::set_subtree_ctrl( Node *n ) { duke@435: // Already set? Get out. duke@435: if( _nodes[n->_idx] ) return; duke@435: // Recursively set _nodes array to indicate where the Node goes duke@435: uint i; duke@435: for( i = 0; i < n->req(); ++i ) { duke@435: Node *m = n->in(i); duke@435: if( m && m != C->root() ) duke@435: set_subtree_ctrl( m ); duke@435: } duke@435: duke@435: // Fixup self duke@435: set_early_ctrl( n ); duke@435: } duke@435: duke@435: //------------------------------is_counted_loop-------------------------------- duke@435: Node *PhaseIdealLoop::is_counted_loop( Node *x, IdealLoopTree *loop ) { duke@435: PhaseGVN *gvn = &_igvn; duke@435: duke@435: // Counted loop head must be a good RegionNode with only 3 not NULL duke@435: // control input edges: Self, Entry, LoopBack. duke@435: if ( x->in(LoopNode::Self) == NULL || x->req() != 3 ) duke@435: return NULL; duke@435: duke@435: Node *init_control = x->in(LoopNode::EntryControl); duke@435: Node *back_control = x->in(LoopNode::LoopBackControl); duke@435: if( init_control == NULL || back_control == NULL ) // Partially dead duke@435: return NULL; duke@435: // Must also check for TOP when looking for a dead loop duke@435: if( init_control->is_top() || back_control->is_top() ) duke@435: return NULL; duke@435: duke@435: // Allow funny placement of Safepoint duke@435: if( back_control->Opcode() == Op_SafePoint ) duke@435: back_control = back_control->in(TypeFunc::Control); duke@435: duke@435: // Controlling test for loop duke@435: Node *iftrue = back_control; duke@435: uint iftrue_op = iftrue->Opcode(); duke@435: if( iftrue_op != Op_IfTrue && duke@435: iftrue_op != Op_IfFalse ) duke@435: // I have a weird back-control. Probably the loop-exit test is in duke@435: // the middle of the loop and I am looking at some trailing control-flow duke@435: // merge point. To fix this I would have to partially peel the loop. duke@435: return NULL; // Obscure back-control duke@435: duke@435: // Get boolean guarding loop-back test duke@435: Node *iff = iftrue->in(0); duke@435: if( get_loop(iff) != loop || !iff->in(1)->is_Bool() ) return NULL; duke@435: BoolNode *test = iff->in(1)->as_Bool(); duke@435: BoolTest::mask bt = test->_test._test; duke@435: float cl_prob = iff->as_If()->_prob; duke@435: if( iftrue_op == Op_IfFalse ) { duke@435: bt = BoolTest(bt).negate(); duke@435: cl_prob = 1.0 - cl_prob; duke@435: } duke@435: // Get backedge compare duke@435: Node *cmp = test->in(1); duke@435: int cmp_op = cmp->Opcode(); duke@435: if( cmp_op != Op_CmpI ) duke@435: return NULL; // Avoid pointer & float compares duke@435: duke@435: // Find the trip-counter increment & limit. Limit must be loop invariant. duke@435: Node *incr = cmp->in(1); duke@435: Node *limit = cmp->in(2); duke@435: duke@435: // --------- duke@435: // need 'loop()' test to tell if limit is loop invariant duke@435: // --------- duke@435: duke@435: if( !is_member( loop, get_ctrl(incr) ) ) { // Swapped trip counter and limit? duke@435: Node *tmp = incr; // Then reverse order into the CmpI duke@435: incr = limit; duke@435: limit = tmp; duke@435: bt = BoolTest(bt).commute(); // And commute the exit test duke@435: } duke@435: if( is_member( loop, get_ctrl(limit) ) ) // Limit must loop-invariant duke@435: return NULL; duke@435: duke@435: // Trip-counter increment must be commutative & associative. duke@435: uint incr_op = incr->Opcode(); duke@435: if( incr_op == Op_Phi && incr->req() == 3 ) { duke@435: incr = incr->in(2); // Assume incr is on backedge of Phi duke@435: incr_op = incr->Opcode(); duke@435: } duke@435: Node* trunc1 = NULL; duke@435: Node* trunc2 = NULL; duke@435: const TypeInt* iv_trunc_t = NULL; duke@435: if (!(incr = CountedLoopNode::match_incr_with_optional_truncation(incr, &trunc1, &trunc2, &iv_trunc_t))) { duke@435: return NULL; // Funny increment opcode duke@435: } duke@435: duke@435: // Get merge point duke@435: Node *xphi = incr->in(1); duke@435: Node *stride = incr->in(2); duke@435: if( !stride->is_Con() ) { // Oops, swap these duke@435: if( !xphi->is_Con() ) // Is the other guy a constant? duke@435: return NULL; // Nope, unknown stride, bail out duke@435: Node *tmp = xphi; // 'incr' is commutative, so ok to swap duke@435: xphi = stride; duke@435: stride = tmp; duke@435: } duke@435: //if( loop(xphi) != l) return NULL;// Merge point is in inner loop?? duke@435: if( !xphi->is_Phi() ) return NULL; // Too much math on the trip counter duke@435: PhiNode *phi = xphi->as_Phi(); duke@435: duke@435: // Stride must be constant duke@435: const Type *stride_t = stride->bottom_type(); duke@435: int stride_con = stride_t->is_int()->get_con(); duke@435: assert( stride_con, "missed some peephole opt" ); duke@435: duke@435: // Phi must be of loop header; backedge must wrap to increment duke@435: if( phi->region() != x ) return NULL; duke@435: if( trunc1 == NULL && phi->in(LoopNode::LoopBackControl) != incr || duke@435: trunc1 != NULL && phi->in(LoopNode::LoopBackControl) != trunc1 ) { duke@435: return NULL; duke@435: } duke@435: Node *init_trip = phi->in(LoopNode::EntryControl); duke@435: //if (!init_trip->is_Con()) return NULL; // avoid rolling over MAXINT/MININT duke@435: duke@435: // If iv trunc type is smaller than int, check for possible wrap. duke@435: if (!TypeInt::INT->higher_equal(iv_trunc_t)) { duke@435: assert(trunc1 != NULL, "must have found some truncation"); duke@435: duke@435: // Get a better type for the phi (filtered thru if's) duke@435: const TypeInt* phi_ft = filtered_type(phi); duke@435: duke@435: // Can iv take on a value that will wrap? duke@435: // duke@435: // Ensure iv's limit is not within "stride" of the wrap value. duke@435: // duke@435: // Example for "short" type duke@435: // Truncation ensures value is in the range -32768..32767 (iv_trunc_t) duke@435: // If the stride is +10, then the last value of the induction duke@435: // variable before the increment (phi_ft->_hi) must be duke@435: // <= 32767 - 10 and (phi_ft->_lo) must be >= -32768 to duke@435: // ensure no truncation occurs after the increment. duke@435: duke@435: if (stride_con > 0) { duke@435: if (iv_trunc_t->_hi - phi_ft->_hi < stride_con || duke@435: iv_trunc_t->_lo > phi_ft->_lo) { duke@435: return NULL; // truncation may occur duke@435: } duke@435: } else if (stride_con < 0) { duke@435: if (iv_trunc_t->_lo - phi_ft->_lo > stride_con || duke@435: iv_trunc_t->_hi < phi_ft->_hi) { duke@435: return NULL; // truncation may occur duke@435: } duke@435: } duke@435: // No possibility of wrap so truncation can be discarded duke@435: // Promote iv type to Int duke@435: } else { duke@435: assert(trunc1 == NULL && trunc2 == NULL, "no truncation for int"); duke@435: } duke@435: duke@435: // ================================================= duke@435: // ---- SUCCESS! Found A Trip-Counted Loop! ----- duke@435: // duke@435: // Canonicalize the condition on the test. If we can exactly determine duke@435: // the trip-counter exit value, then set limit to that value and use twisti@1040: // a '!=' test. Otherwise use condition '<' for count-up loops and duke@435: // '>' for count-down loops. If the condition is inverted and we will duke@435: // be rolling through MININT to MAXINT, then bail out. duke@435: duke@435: C->print_method("Before CountedLoop", 3); duke@435: duke@435: // Check for SafePoint on backedge and remove duke@435: Node *sfpt = x->in(LoopNode::LoopBackControl); duke@435: if( sfpt->Opcode() == Op_SafePoint && is_deleteable_safept(sfpt)) { duke@435: lazy_replace( sfpt, iftrue ); duke@435: loop->_tail = iftrue; duke@435: } duke@435: duke@435: duke@435: // If compare points to incr, we are ok. Otherwise the compare duke@435: // can directly point to the phi; in this case adjust the compare so that twisti@1040: // it points to the incr by adjusting the limit. duke@435: if( cmp->in(1) == phi || cmp->in(2) == phi ) duke@435: limit = gvn->transform(new (C, 3) AddINode(limit,stride)); duke@435: duke@435: // trip-count for +-tive stride should be: (limit - init_trip + stride - 1)/stride. duke@435: // Final value for iterator should be: trip_count * stride + init_trip. duke@435: const Type *limit_t = limit->bottom_type(); duke@435: const Type *init_t = init_trip->bottom_type(); duke@435: Node *one_p = gvn->intcon( 1); duke@435: Node *one_m = gvn->intcon(-1); duke@435: duke@435: Node *trip_count = NULL; duke@435: Node *hook = new (C, 6) Node(6); duke@435: switch( bt ) { duke@435: case BoolTest::eq: duke@435: return NULL; // Bail out, but this loop trips at most twice! duke@435: case BoolTest::ne: // Ahh, the case we desire duke@435: if( stride_con == 1 ) duke@435: trip_count = gvn->transform(new (C, 3) SubINode(limit,init_trip)); duke@435: else if( stride_con == -1 ) duke@435: trip_count = gvn->transform(new (C, 3) SubINode(init_trip,limit)); duke@435: else duke@435: return NULL; // Odd stride; must prove we hit limit exactly duke@435: set_subtree_ctrl( trip_count ); duke@435: //_loop.map(trip_count->_idx,loop(limit)); duke@435: break; duke@435: case BoolTest::le: // Maybe convert to '<' case duke@435: limit = gvn->transform(new (C, 3) AddINode(limit,one_p)); duke@435: set_subtree_ctrl( limit ); duke@435: hook->init_req(4, limit); duke@435: duke@435: bt = BoolTest::lt; duke@435: // Make the new limit be in the same loop nest as the old limit duke@435: //_loop.map(limit->_idx,limit_loop); duke@435: // Fall into next case duke@435: case BoolTest::lt: { // Maybe convert to '!=' case duke@435: if( stride_con < 0 ) return NULL; // Count down loop rolls through MAXINT duke@435: Node *range = gvn->transform(new (C, 3) SubINode(limit,init_trip)); duke@435: set_subtree_ctrl( range ); duke@435: hook->init_req(0, range); duke@435: duke@435: Node *bias = gvn->transform(new (C, 3) AddINode(range,stride)); duke@435: set_subtree_ctrl( bias ); duke@435: hook->init_req(1, bias); duke@435: duke@435: Node *bias1 = gvn->transform(new (C, 3) AddINode(bias,one_m)); duke@435: set_subtree_ctrl( bias1 ); duke@435: hook->init_req(2, bias1); duke@435: duke@435: trip_count = gvn->transform(new (C, 3) DivINode(0,bias1,stride)); duke@435: set_subtree_ctrl( trip_count ); duke@435: hook->init_req(3, trip_count); duke@435: break; duke@435: } duke@435: duke@435: case BoolTest::ge: // Maybe convert to '>' case duke@435: limit = gvn->transform(new (C, 3) AddINode(limit,one_m)); duke@435: set_subtree_ctrl( limit ); duke@435: hook->init_req(4 ,limit); duke@435: duke@435: bt = BoolTest::gt; duke@435: // Make the new limit be in the same loop nest as the old limit duke@435: //_loop.map(limit->_idx,limit_loop); duke@435: // Fall into next case duke@435: case BoolTest::gt: { // Maybe convert to '!=' case duke@435: if( stride_con > 0 ) return NULL; // count up loop rolls through MININT duke@435: Node *range = gvn->transform(new (C, 3) SubINode(limit,init_trip)); duke@435: set_subtree_ctrl( range ); duke@435: hook->init_req(0, range); duke@435: duke@435: Node *bias = gvn->transform(new (C, 3) AddINode(range,stride)); duke@435: set_subtree_ctrl( bias ); duke@435: hook->init_req(1, bias); duke@435: duke@435: Node *bias1 = gvn->transform(new (C, 3) AddINode(bias,one_p)); duke@435: set_subtree_ctrl( bias1 ); duke@435: hook->init_req(2, bias1); duke@435: duke@435: trip_count = gvn->transform(new (C, 3) DivINode(0,bias1,stride)); duke@435: set_subtree_ctrl( trip_count ); duke@435: hook->init_req(3, trip_count); duke@435: break; duke@435: } duke@435: } duke@435: duke@435: Node *span = gvn->transform(new (C, 3) MulINode(trip_count,stride)); duke@435: set_subtree_ctrl( span ); duke@435: hook->init_req(5, span); duke@435: duke@435: limit = gvn->transform(new (C, 3) AddINode(span,init_trip)); duke@435: set_subtree_ctrl( limit ); duke@435: duke@435: // Build a canonical trip test. duke@435: // Clone code, as old values may be in use. duke@435: incr = incr->clone(); duke@435: incr->set_req(1,phi); duke@435: incr->set_req(2,stride); duke@435: incr = _igvn.register_new_node_with_optimizer(incr); duke@435: set_early_ctrl( incr ); duke@435: _igvn.hash_delete(phi); duke@435: phi->set_req_X( LoopNode::LoopBackControl, incr, &_igvn ); duke@435: duke@435: // If phi type is more restrictive than Int, raise to duke@435: // Int to prevent (almost) infinite recursion in igvn duke@435: // which can only handle integer types for constants or minint..maxint. duke@435: if (!TypeInt::INT->higher_equal(phi->bottom_type())) { duke@435: Node* nphi = PhiNode::make(phi->in(0), phi->in(LoopNode::EntryControl), TypeInt::INT); duke@435: nphi->set_req(LoopNode::LoopBackControl, phi->in(LoopNode::LoopBackControl)); duke@435: nphi = _igvn.register_new_node_with_optimizer(nphi); duke@435: set_ctrl(nphi, get_ctrl(phi)); duke@435: _igvn.subsume_node(phi, nphi); duke@435: phi = nphi->as_Phi(); duke@435: } duke@435: cmp = cmp->clone(); duke@435: cmp->set_req(1,incr); duke@435: cmp->set_req(2,limit); duke@435: cmp = _igvn.register_new_node_with_optimizer(cmp); duke@435: set_ctrl(cmp, iff->in(0)); duke@435: duke@435: Node *tmp = test->clone(); duke@435: assert( tmp->is_Bool(), "" ); duke@435: test = (BoolNode*)tmp; duke@435: (*(BoolTest*)&test->_test)._test = bt; //BoolTest::ne; duke@435: test->set_req(1,cmp); duke@435: _igvn.register_new_node_with_optimizer(test); duke@435: set_ctrl(test, iff->in(0)); duke@435: // If the exit test is dead, STOP! duke@435: if( test == NULL ) return NULL; duke@435: _igvn.hash_delete(iff); duke@435: iff->set_req_X( 1, test, &_igvn ); duke@435: duke@435: // Replace the old IfNode with a new LoopEndNode duke@435: Node *lex = _igvn.register_new_node_with_optimizer(new (C, 2) CountedLoopEndNode( iff->in(0), iff->in(1), cl_prob, iff->as_If()->_fcnt )); duke@435: IfNode *le = lex->as_If(); duke@435: uint dd = dom_depth(iff); duke@435: set_idom(le, le->in(0), dd); // Update dominance for loop exit duke@435: set_loop(le, loop); duke@435: duke@435: // Get the loop-exit control duke@435: Node *if_f = iff->as_If()->proj_out(!(iftrue_op == Op_IfTrue)); duke@435: duke@435: // Need to swap loop-exit and loop-back control? duke@435: if( iftrue_op == Op_IfFalse ) { duke@435: Node *ift2=_igvn.register_new_node_with_optimizer(new (C, 1) IfTrueNode (le)); duke@435: Node *iff2=_igvn.register_new_node_with_optimizer(new (C, 1) IfFalseNode(le)); duke@435: duke@435: loop->_tail = back_control = ift2; duke@435: set_loop(ift2, loop); duke@435: set_loop(iff2, get_loop(if_f)); duke@435: duke@435: // Lazy update of 'get_ctrl' mechanism. duke@435: lazy_replace_proj( if_f , iff2 ); duke@435: lazy_replace_proj( iftrue, ift2 ); duke@435: duke@435: // Swap names duke@435: if_f = iff2; duke@435: iftrue = ift2; duke@435: } else { duke@435: _igvn.hash_delete(if_f ); duke@435: _igvn.hash_delete(iftrue); duke@435: if_f ->set_req_X( 0, le, &_igvn ); duke@435: iftrue->set_req_X( 0, le, &_igvn ); duke@435: } duke@435: duke@435: set_idom(iftrue, le, dd+1); duke@435: set_idom(if_f, le, dd+1); duke@435: duke@435: // Now setup a new CountedLoopNode to replace the existing LoopNode duke@435: CountedLoopNode *l = new (C, 3) CountedLoopNode(init_control, back_control); duke@435: // The following assert is approximately true, and defines the intention duke@435: // of can_be_counted_loop. It fails, however, because phase->type duke@435: // is not yet initialized for this loop and its parts. duke@435: //assert(l->can_be_counted_loop(this), "sanity"); duke@435: _igvn.register_new_node_with_optimizer(l); duke@435: set_loop(l, loop); duke@435: loop->_head = l; duke@435: // Fix all data nodes placed at the old loop head. duke@435: // Uses the lazy-update mechanism of 'get_ctrl'. duke@435: lazy_replace( x, l ); duke@435: set_idom(l, init_control, dom_depth(x)); duke@435: twisti@1040: // Check for immediately preceding SafePoint and remove duke@435: Node *sfpt2 = le->in(0); duke@435: if( sfpt2->Opcode() == Op_SafePoint && is_deleteable_safept(sfpt2)) duke@435: lazy_replace( sfpt2, sfpt2->in(TypeFunc::Control)); duke@435: duke@435: // Free up intermediate goo duke@435: _igvn.remove_dead_node(hook); duke@435: duke@435: C->print_method("After CountedLoop", 3); duke@435: duke@435: // Return trip counter duke@435: return trip_count; duke@435: } duke@435: duke@435: duke@435: //------------------------------Ideal------------------------------------------ duke@435: // Return a node which is more "ideal" than the current node. duke@435: // Attempt to convert into a counted-loop. duke@435: Node *LoopNode::Ideal(PhaseGVN *phase, bool can_reshape) { duke@435: if (!can_be_counted_loop(phase)) { duke@435: phase->C->set_major_progress(); duke@435: } duke@435: return RegionNode::Ideal(phase, can_reshape); duke@435: } duke@435: duke@435: duke@435: //============================================================================= duke@435: //------------------------------Ideal------------------------------------------ duke@435: // Return a node which is more "ideal" than the current node. duke@435: // Attempt to convert into a counted-loop. duke@435: Node *CountedLoopNode::Ideal(PhaseGVN *phase, bool can_reshape) { duke@435: return RegionNode::Ideal(phase, can_reshape); duke@435: } duke@435: duke@435: //------------------------------dump_spec-------------------------------------- duke@435: // Dump special per-node info duke@435: #ifndef PRODUCT duke@435: void CountedLoopNode::dump_spec(outputStream *st) const { duke@435: LoopNode::dump_spec(st); duke@435: if( stride_is_con() ) { duke@435: st->print("stride: %d ",stride_con()); duke@435: } else { duke@435: st->print("stride: not constant "); duke@435: } duke@435: if( is_pre_loop () ) st->print("pre of N%d" , _main_idx ); duke@435: if( is_main_loop() ) st->print("main of N%d", _idx ); duke@435: if( is_post_loop() ) st->print("post of N%d", _main_idx ); duke@435: } duke@435: #endif duke@435: duke@435: //============================================================================= duke@435: int CountedLoopEndNode::stride_con() const { duke@435: return stride()->bottom_type()->is_int()->get_con(); duke@435: } duke@435: duke@435: duke@435: //----------------------match_incr_with_optional_truncation-------------------- duke@435: // Match increment with optional truncation: duke@435: // CHAR: (i+1)&0x7fff, BYTE: ((i+1)<<8)>>8, or SHORT: ((i+1)<<16)>>16 duke@435: // Return NULL for failure. Success returns the increment node. duke@435: Node* CountedLoopNode::match_incr_with_optional_truncation( duke@435: Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type) { duke@435: // Quick cutouts: duke@435: if (expr == NULL || expr->req() != 3) return false; duke@435: duke@435: Node *t1 = NULL; duke@435: Node *t2 = NULL; duke@435: const TypeInt* trunc_t = TypeInt::INT; duke@435: Node* n1 = expr; duke@435: int n1op = n1->Opcode(); duke@435: duke@435: // Try to strip (n1 & M) or (n1 << N >> N) from n1. duke@435: if (n1op == Op_AndI && duke@435: n1->in(2)->is_Con() && duke@435: n1->in(2)->bottom_type()->is_int()->get_con() == 0x7fff) { duke@435: // %%% This check should match any mask of 2**K-1. duke@435: t1 = n1; duke@435: n1 = t1->in(1); duke@435: n1op = n1->Opcode(); duke@435: trunc_t = TypeInt::CHAR; duke@435: } else if (n1op == Op_RShiftI && duke@435: n1->in(1) != NULL && duke@435: n1->in(1)->Opcode() == Op_LShiftI && duke@435: n1->in(2) == n1->in(1)->in(2) && duke@435: n1->in(2)->is_Con()) { duke@435: jint shift = n1->in(2)->bottom_type()->is_int()->get_con(); duke@435: // %%% This check should match any shift in [1..31]. duke@435: if (shift == 16 || shift == 8) { duke@435: t1 = n1; duke@435: t2 = t1->in(1); duke@435: n1 = t2->in(1); duke@435: n1op = n1->Opcode(); duke@435: if (shift == 16) { duke@435: trunc_t = TypeInt::SHORT; duke@435: } else if (shift == 8) { duke@435: trunc_t = TypeInt::BYTE; duke@435: } duke@435: } duke@435: } duke@435: duke@435: // If (maybe after stripping) it is an AddI, we won: duke@435: if (n1op == Op_AddI) { duke@435: *trunc1 = t1; duke@435: *trunc2 = t2; duke@435: *trunc_type = trunc_t; duke@435: return n1; duke@435: } duke@435: duke@435: // failed duke@435: return NULL; duke@435: } duke@435: duke@435: duke@435: //------------------------------filtered_type-------------------------------- duke@435: // Return a type based on condition control flow duke@435: // A successful return will be a type that is restricted due duke@435: // to a series of dominating if-tests, such as: duke@435: // if (i < 10) { duke@435: // if (i > 0) { duke@435: // here: "i" type is [1..10) duke@435: // } duke@435: // } duke@435: // or a control flow merge duke@435: // if (i < 10) { duke@435: // do { duke@435: // phi( , ) -- at top of loop type is [min_int..10) duke@435: // i = ? duke@435: // } while ( i < 10) duke@435: // duke@435: const TypeInt* PhaseIdealLoop::filtered_type( Node *n, Node* n_ctrl) { duke@435: assert(n && n->bottom_type()->is_int(), "must be int"); duke@435: const TypeInt* filtered_t = NULL; duke@435: if (!n->is_Phi()) { duke@435: assert(n_ctrl != NULL || n_ctrl == C->top(), "valid control"); duke@435: filtered_t = filtered_type_from_dominators(n, n_ctrl); duke@435: duke@435: } else { duke@435: Node* phi = n->as_Phi(); duke@435: Node* region = phi->in(0); duke@435: assert(n_ctrl == NULL || n_ctrl == region, "ctrl parameter must be region"); duke@435: if (region && region != C->top()) { duke@435: for (uint i = 1; i < phi->req(); i++) { duke@435: Node* val = phi->in(i); duke@435: Node* use_c = region->in(i); duke@435: const TypeInt* val_t = filtered_type_from_dominators(val, use_c); duke@435: if (val_t != NULL) { duke@435: if (filtered_t == NULL) { duke@435: filtered_t = val_t; duke@435: } else { duke@435: filtered_t = filtered_t->meet(val_t)->is_int(); duke@435: } duke@435: } duke@435: } duke@435: } duke@435: } duke@435: const TypeInt* n_t = _igvn.type(n)->is_int(); duke@435: if (filtered_t != NULL) { duke@435: n_t = n_t->join(filtered_t)->is_int(); duke@435: } duke@435: return n_t; duke@435: } duke@435: duke@435: duke@435: //------------------------------filtered_type_from_dominators-------------------------------- duke@435: // Return a possibly more restrictive type for val based on condition control flow of dominators duke@435: const TypeInt* PhaseIdealLoop::filtered_type_from_dominators( Node* val, Node *use_ctrl) { duke@435: if (val->is_Con()) { duke@435: return val->bottom_type()->is_int(); duke@435: } duke@435: uint if_limit = 10; // Max number of dominating if's visited duke@435: const TypeInt* rtn_t = NULL; duke@435: duke@435: if (use_ctrl && use_ctrl != C->top()) { duke@435: Node* val_ctrl = get_ctrl(val); duke@435: uint val_dom_depth = dom_depth(val_ctrl); duke@435: Node* pred = use_ctrl; duke@435: uint if_cnt = 0; duke@435: while (if_cnt < if_limit) { duke@435: if ((pred->Opcode() == Op_IfTrue || pred->Opcode() == Op_IfFalse)) { duke@435: if_cnt++; never@452: const TypeInt* if_t = IfNode::filtered_int_type(&_igvn, val, pred); duke@435: if (if_t != NULL) { duke@435: if (rtn_t == NULL) { duke@435: rtn_t = if_t; duke@435: } else { duke@435: rtn_t = rtn_t->join(if_t)->is_int(); duke@435: } duke@435: } duke@435: } duke@435: pred = idom(pred); duke@435: if (pred == NULL || pred == C->top()) { duke@435: break; duke@435: } duke@435: // Stop if going beyond definition block of val duke@435: if (dom_depth(pred) < val_dom_depth) { duke@435: break; duke@435: } duke@435: } duke@435: } duke@435: return rtn_t; duke@435: } duke@435: duke@435: duke@435: //------------------------------dump_spec-------------------------------------- duke@435: // Dump special per-node info duke@435: #ifndef PRODUCT duke@435: void CountedLoopEndNode::dump_spec(outputStream *st) const { duke@435: if( in(TestValue)->is_Bool() ) { duke@435: BoolTest bt( test_trip()); // Added this for g++. duke@435: duke@435: st->print("["); duke@435: bt.dump_on(st); duke@435: st->print("]"); duke@435: } duke@435: st->print(" "); duke@435: IfNode::dump_spec(st); duke@435: } duke@435: #endif duke@435: duke@435: //============================================================================= duke@435: //------------------------------is_member-------------------------------------- duke@435: // Is 'l' a member of 'this'? duke@435: int IdealLoopTree::is_member( const IdealLoopTree *l ) const { duke@435: while( l->_nest > _nest ) l = l->_parent; duke@435: return l == this; duke@435: } duke@435: duke@435: //------------------------------set_nest--------------------------------------- duke@435: // Set loop tree nesting depth. Accumulate _has_call bits. duke@435: int IdealLoopTree::set_nest( uint depth ) { duke@435: _nest = depth; duke@435: int bits = _has_call; duke@435: if( _child ) bits |= _child->set_nest(depth+1); duke@435: if( bits ) _has_call = 1; duke@435: if( _next ) bits |= _next ->set_nest(depth ); duke@435: return bits; duke@435: } duke@435: duke@435: //------------------------------split_fall_in---------------------------------- duke@435: // Split out multiple fall-in edges from the loop header. Move them to a duke@435: // private RegionNode before the loop. This becomes the loop landing pad. duke@435: void IdealLoopTree::split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt ) { duke@435: PhaseIterGVN &igvn = phase->_igvn; duke@435: uint i; duke@435: duke@435: // Make a new RegionNode to be the landing pad. duke@435: Node *landing_pad = new (phase->C, fall_in_cnt+1) RegionNode( fall_in_cnt+1 ); duke@435: phase->set_loop(landing_pad,_parent); duke@435: // Gather all the fall-in control paths into the landing pad duke@435: uint icnt = fall_in_cnt; duke@435: uint oreq = _head->req(); duke@435: for( i = oreq-1; i>0; i-- ) duke@435: if( !phase->is_member( this, _head->in(i) ) ) duke@435: landing_pad->set_req(icnt--,_head->in(i)); duke@435: duke@435: // Peel off PhiNode edges as well duke@435: for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) { duke@435: Node *oj = _head->fast_out(j); duke@435: if( oj->is_Phi() ) { duke@435: PhiNode* old_phi = oj->as_Phi(); duke@435: assert( old_phi->region() == _head, "" ); duke@435: igvn.hash_delete(old_phi); // Yank from hash before hacking edges duke@435: Node *p = PhiNode::make_blank(landing_pad, old_phi); duke@435: uint icnt = fall_in_cnt; duke@435: for( i = oreq-1; i>0; i-- ) { duke@435: if( !phase->is_member( this, _head->in(i) ) ) { duke@435: p->init_req(icnt--, old_phi->in(i)); duke@435: // Go ahead and clean out old edges from old phi duke@435: old_phi->del_req(i); duke@435: } duke@435: } duke@435: // Search for CSE's here, because ZKM.jar does a lot of duke@435: // loop hackery and we need to be a little incremental duke@435: // with the CSE to avoid O(N^2) node blow-up. duke@435: Node *p2 = igvn.hash_find_insert(p); // Look for a CSE duke@435: if( p2 ) { // Found CSE duke@435: p->destruct(); // Recover useless new node duke@435: p = p2; // Use old node duke@435: } else { duke@435: igvn.register_new_node_with_optimizer(p, old_phi); duke@435: } duke@435: // Make old Phi refer to new Phi. duke@435: old_phi->add_req(p); duke@435: // Check for the special case of making the old phi useless and duke@435: // disappear it. In JavaGrande I have a case where this useless duke@435: // Phi is the loop limit and prevents recognizing a CountedLoop duke@435: // which in turn prevents removing an empty loop. duke@435: Node *id_old_phi = old_phi->Identity( &igvn ); duke@435: if( id_old_phi != old_phi ) { // Found a simple identity? duke@435: // Note that I cannot call 'subsume_node' here, because duke@435: // that will yank the edge from old_phi to the Region and duke@435: // I'm mid-iteration over the Region's uses. duke@435: for (DUIterator_Last imin, i = old_phi->last_outs(imin); i >= imin; ) { duke@435: Node* use = old_phi->last_out(i); duke@435: igvn.hash_delete(use); duke@435: igvn._worklist.push(use); duke@435: uint uses_found = 0; duke@435: for (uint j = 0; j < use->len(); j++) { duke@435: if (use->in(j) == old_phi) { duke@435: if (j < use->req()) use->set_req (j, id_old_phi); duke@435: else use->set_prec(j, id_old_phi); duke@435: uses_found++; duke@435: } duke@435: } duke@435: i -= uses_found; // we deleted 1 or more copies of this edge duke@435: } duke@435: } duke@435: igvn._worklist.push(old_phi); duke@435: } duke@435: } duke@435: // Finally clean out the fall-in edges from the RegionNode duke@435: for( i = oreq-1; i>0; i-- ) { duke@435: if( !phase->is_member( this, _head->in(i) ) ) { duke@435: _head->del_req(i); duke@435: } duke@435: } duke@435: // Transform landing pad duke@435: igvn.register_new_node_with_optimizer(landing_pad, _head); duke@435: // Insert landing pad into the header duke@435: _head->add_req(landing_pad); duke@435: } duke@435: duke@435: //------------------------------split_outer_loop------------------------------- duke@435: // Split out the outermost loop from this shared header. duke@435: void IdealLoopTree::split_outer_loop( PhaseIdealLoop *phase ) { duke@435: PhaseIterGVN &igvn = phase->_igvn; duke@435: duke@435: // Find index of outermost loop; it should also be my tail. duke@435: uint outer_idx = 1; duke@435: while( _head->in(outer_idx) != _tail ) outer_idx++; duke@435: duke@435: // Make a LoopNode for the outermost loop. duke@435: Node *ctl = _head->in(LoopNode::EntryControl); duke@435: Node *outer = new (phase->C, 3) LoopNode( ctl, _head->in(outer_idx) ); duke@435: outer = igvn.register_new_node_with_optimizer(outer, _head); duke@435: phase->set_created_loop_node(); duke@435: // Outermost loop falls into '_head' loop duke@435: _head->set_req(LoopNode::EntryControl, outer); duke@435: _head->del_req(outer_idx); duke@435: // Split all the Phis up between '_head' loop and 'outer' loop. duke@435: for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) { duke@435: Node *out = _head->fast_out(j); duke@435: if( out->is_Phi() ) { duke@435: PhiNode *old_phi = out->as_Phi(); duke@435: assert( old_phi->region() == _head, "" ); duke@435: Node *phi = PhiNode::make_blank(outer, old_phi); duke@435: phi->init_req(LoopNode::EntryControl, old_phi->in(LoopNode::EntryControl)); duke@435: phi->init_req(LoopNode::LoopBackControl, old_phi->in(outer_idx)); duke@435: phi = igvn.register_new_node_with_optimizer(phi, old_phi); duke@435: // Make old Phi point to new Phi on the fall-in path duke@435: igvn.hash_delete(old_phi); duke@435: old_phi->set_req(LoopNode::EntryControl, phi); duke@435: old_phi->del_req(outer_idx); duke@435: igvn._worklist.push(old_phi); duke@435: } duke@435: } duke@435: duke@435: // Use the new loop head instead of the old shared one duke@435: _head = outer; duke@435: phase->set_loop(_head, this); duke@435: } duke@435: duke@435: //------------------------------fix_parent------------------------------------- duke@435: static void fix_parent( IdealLoopTree *loop, IdealLoopTree *parent ) { duke@435: loop->_parent = parent; duke@435: if( loop->_child ) fix_parent( loop->_child, loop ); duke@435: if( loop->_next ) fix_parent( loop->_next , parent ); duke@435: } duke@435: duke@435: //------------------------------estimate_path_freq----------------------------- duke@435: static float estimate_path_freq( Node *n ) { duke@435: // Try to extract some path frequency info duke@435: IfNode *iff; duke@435: for( int i = 0; i < 50; i++ ) { // Skip through a bunch of uncommon tests duke@435: uint nop = n->Opcode(); duke@435: if( nop == Op_SafePoint ) { // Skip any safepoint duke@435: n = n->in(0); duke@435: continue; duke@435: } duke@435: if( nop == Op_CatchProj ) { // Get count from a prior call duke@435: // Assume call does not always throw exceptions: means the call-site duke@435: // count is also the frequency of the fall-through path. duke@435: assert( n->is_CatchProj(), "" ); duke@435: if( ((CatchProjNode*)n)->_con != CatchProjNode::fall_through_index ) duke@435: return 0.0f; // Assume call exception path is rare duke@435: Node *call = n->in(0)->in(0)->in(0); duke@435: assert( call->is_Call(), "expect a call here" ); duke@435: const JVMState *jvms = ((CallNode*)call)->jvms(); duke@435: ciMethodData* methodData = jvms->method()->method_data(); duke@435: if (!methodData->is_mature()) return 0.0f; // No call-site data duke@435: ciProfileData* data = methodData->bci_to_data(jvms->bci()); duke@435: if ((data == NULL) || !data->is_CounterData()) { duke@435: // no call profile available, try call's control input duke@435: n = n->in(0); duke@435: continue; duke@435: } duke@435: return data->as_CounterData()->count()/FreqCountInvocations; duke@435: } duke@435: // See if there's a gating IF test duke@435: Node *n_c = n->in(0); duke@435: if( !n_c->is_If() ) break; // No estimate available duke@435: iff = n_c->as_If(); duke@435: if( iff->_fcnt != COUNT_UNKNOWN ) // Have a valid count? duke@435: // Compute how much count comes on this path duke@435: return ((nop == Op_IfTrue) ? iff->_prob : 1.0f - iff->_prob) * iff->_fcnt; duke@435: // Have no count info. Skip dull uncommon-trap like branches. duke@435: if( (nop == Op_IfTrue && iff->_prob < PROB_LIKELY_MAG(5)) || duke@435: (nop == Op_IfFalse && iff->_prob > PROB_UNLIKELY_MAG(5)) ) duke@435: break; duke@435: // Skip through never-taken branch; look for a real loop exit. duke@435: n = iff->in(0); duke@435: } duke@435: return 0.0f; // No estimate available duke@435: } duke@435: duke@435: //------------------------------merge_many_backedges--------------------------- duke@435: // Merge all the backedges from the shared header into a private Region. duke@435: // Feed that region as the one backedge to this loop. duke@435: void IdealLoopTree::merge_many_backedges( PhaseIdealLoop *phase ) { duke@435: uint i; duke@435: duke@435: // Scan for the top 2 hottest backedges duke@435: float hotcnt = 0.0f; duke@435: float warmcnt = 0.0f; duke@435: uint hot_idx = 0; duke@435: // Loop starts at 2 because slot 1 is the fall-in path duke@435: for( i = 2; i < _head->req(); i++ ) { duke@435: float cnt = estimate_path_freq(_head->in(i)); duke@435: if( cnt > hotcnt ) { // Grab hottest path duke@435: warmcnt = hotcnt; duke@435: hotcnt = cnt; duke@435: hot_idx = i; duke@435: } else if( cnt > warmcnt ) { // And 2nd hottest path duke@435: warmcnt = cnt; duke@435: } duke@435: } duke@435: duke@435: // See if the hottest backedge is worthy of being an inner loop duke@435: // by being much hotter than the next hottest backedge. duke@435: if( hotcnt <= 0.0001 || duke@435: hotcnt < 2.0*warmcnt ) hot_idx = 0;// No hot backedge duke@435: duke@435: // Peel out the backedges into a private merge point; peel duke@435: // them all except optionally hot_idx. duke@435: PhaseIterGVN &igvn = phase->_igvn; duke@435: duke@435: Node *hot_tail = NULL; duke@435: // Make a Region for the merge point duke@435: Node *r = new (phase->C, 1) RegionNode(1); duke@435: for( i = 2; i < _head->req(); i++ ) { duke@435: if( i != hot_idx ) duke@435: r->add_req( _head->in(i) ); duke@435: else hot_tail = _head->in(i); duke@435: } duke@435: igvn.register_new_node_with_optimizer(r, _head); duke@435: // Plug region into end of loop _head, followed by hot_tail duke@435: while( _head->req() > 3 ) _head->del_req( _head->req()-1 ); duke@435: _head->set_req(2, r); duke@435: if( hot_idx ) _head->add_req(hot_tail); duke@435: duke@435: // Split all the Phis up between '_head' loop and the Region 'r' duke@435: for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) { duke@435: Node *out = _head->fast_out(j); duke@435: if( out->is_Phi() ) { duke@435: PhiNode* n = out->as_Phi(); duke@435: igvn.hash_delete(n); // Delete from hash before hacking edges duke@435: Node *hot_phi = NULL; duke@435: Node *phi = new (phase->C, r->req()) PhiNode(r, n->type(), n->adr_type()); duke@435: // Check all inputs for the ones to peel out duke@435: uint j = 1; duke@435: for( uint i = 2; i < n->req(); i++ ) { duke@435: if( i != hot_idx ) duke@435: phi->set_req( j++, n->in(i) ); duke@435: else hot_phi = n->in(i); duke@435: } duke@435: // Register the phi but do not transform until whole place transforms duke@435: igvn.register_new_node_with_optimizer(phi, n); duke@435: // Add the merge phi to the old Phi duke@435: while( n->req() > 3 ) n->del_req( n->req()-1 ); duke@435: n->set_req(2, phi); duke@435: if( hot_idx ) n->add_req(hot_phi); duke@435: } duke@435: } duke@435: duke@435: duke@435: // Insert a new IdealLoopTree inserted below me. Turn it into a clone duke@435: // of self loop tree. Turn self into a loop headed by _head and with duke@435: // tail being the new merge point. duke@435: IdealLoopTree *ilt = new IdealLoopTree( phase, _head, _tail ); duke@435: phase->set_loop(_tail,ilt); // Adjust tail duke@435: _tail = r; // Self's tail is new merge point duke@435: phase->set_loop(r,this); duke@435: ilt->_child = _child; // New guy has my children duke@435: _child = ilt; // Self has new guy as only child duke@435: ilt->_parent = this; // new guy has self for parent duke@435: ilt->_nest = _nest; // Same nesting depth (for now) duke@435: duke@435: // Starting with 'ilt', look for child loop trees using the same shared duke@435: // header. Flatten these out; they will no longer be loops in the end. duke@435: IdealLoopTree **pilt = &_child; duke@435: while( ilt ) { duke@435: if( ilt->_head == _head ) { duke@435: uint i; duke@435: for( i = 2; i < _head->req(); i++ ) duke@435: if( _head->in(i) == ilt->_tail ) duke@435: break; // Still a loop duke@435: if( i == _head->req() ) { // No longer a loop duke@435: // Flatten ilt. Hang ilt's "_next" list from the end of duke@435: // ilt's '_child' list. Move the ilt's _child up to replace ilt. duke@435: IdealLoopTree **cp = &ilt->_child; duke@435: while( *cp ) cp = &(*cp)->_next; // Find end of child list duke@435: *cp = ilt->_next; // Hang next list at end of child list duke@435: *pilt = ilt->_child; // Move child up to replace ilt duke@435: ilt->_head = NULL; // Flag as a loop UNIONED into parent duke@435: ilt = ilt->_child; // Repeat using new ilt duke@435: continue; // do not advance over ilt->_child duke@435: } duke@435: assert( ilt->_tail == hot_tail, "expected to only find the hot inner loop here" ); duke@435: phase->set_loop(_head,ilt); duke@435: } duke@435: pilt = &ilt->_child; // Advance to next duke@435: ilt = *pilt; duke@435: } duke@435: duke@435: if( _child ) fix_parent( _child, this ); duke@435: } duke@435: duke@435: //------------------------------beautify_loops--------------------------------- duke@435: // Split shared headers and insert loop landing pads. duke@435: // Insert a LoopNode to replace the RegionNode. duke@435: // Return TRUE if loop tree is structurally changed. duke@435: bool IdealLoopTree::beautify_loops( PhaseIdealLoop *phase ) { duke@435: bool result = false; duke@435: // Cache parts in locals for easy duke@435: PhaseIterGVN &igvn = phase->_igvn; duke@435: duke@435: phase->C->print_method("Before beautify loops", 3); duke@435: duke@435: igvn.hash_delete(_head); // Yank from hash before hacking edges duke@435: duke@435: // Check for multiple fall-in paths. Peel off a landing pad if need be. duke@435: int fall_in_cnt = 0; duke@435: for( uint i = 1; i < _head->req(); i++ ) duke@435: if( !phase->is_member( this, _head->in(i) ) ) duke@435: fall_in_cnt++; duke@435: assert( fall_in_cnt, "at least 1 fall-in path" ); duke@435: if( fall_in_cnt > 1 ) // Need a loop landing pad to merge fall-ins duke@435: split_fall_in( phase, fall_in_cnt ); duke@435: duke@435: // Swap inputs to the _head and all Phis to move the fall-in edge to duke@435: // the left. duke@435: fall_in_cnt = 1; duke@435: while( phase->is_member( this, _head->in(fall_in_cnt) ) ) duke@435: fall_in_cnt++; duke@435: if( fall_in_cnt > 1 ) { duke@435: // Since I am just swapping inputs I do not need to update def-use info duke@435: Node *tmp = _head->in(1); duke@435: _head->set_req( 1, _head->in(fall_in_cnt) ); duke@435: _head->set_req( fall_in_cnt, tmp ); duke@435: // Swap also all Phis duke@435: for (DUIterator_Fast imax, i = _head->fast_outs(imax); i < imax; i++) { duke@435: Node* phi = _head->fast_out(i); duke@435: if( phi->is_Phi() ) { duke@435: igvn.hash_delete(phi); // Yank from hash before hacking edges duke@435: tmp = phi->in(1); duke@435: phi->set_req( 1, phi->in(fall_in_cnt) ); duke@435: phi->set_req( fall_in_cnt, tmp ); duke@435: } duke@435: } duke@435: } duke@435: assert( !phase->is_member( this, _head->in(1) ), "left edge is fall-in" ); duke@435: assert( phase->is_member( this, _head->in(2) ), "right edge is loop" ); duke@435: duke@435: // If I am a shared header (multiple backedges), peel off the many duke@435: // backedges into a private merge point and use the merge point as duke@435: // the one true backedge. duke@435: if( _head->req() > 3 ) { duke@435: // Merge the many backedges into a single backedge. duke@435: merge_many_backedges( phase ); duke@435: result = true; duke@435: } duke@435: duke@435: // If I am a shared header (multiple backedges), peel off myself loop. duke@435: // I better be the outermost loop. duke@435: if( _head->req() > 3 ) { duke@435: split_outer_loop( phase ); duke@435: result = true; duke@435: duke@435: } else if( !_head->is_Loop() && !_irreducible ) { duke@435: // Make a new LoopNode to replace the old loop head duke@435: Node *l = new (phase->C, 3) LoopNode( _head->in(1), _head->in(2) ); duke@435: l = igvn.register_new_node_with_optimizer(l, _head); duke@435: phase->set_created_loop_node(); duke@435: // Go ahead and replace _head duke@435: phase->_igvn.subsume_node( _head, l ); duke@435: _head = l; duke@435: phase->set_loop(_head, this); duke@435: for (DUIterator_Fast imax, i = l->fast_outs(imax); i < imax; i++) duke@435: phase->_igvn.add_users_to_worklist(l->fast_out(i)); duke@435: } duke@435: duke@435: // Now recursively beautify nested loops duke@435: if( _child ) result |= _child->beautify_loops( phase ); duke@435: if( _next ) result |= _next ->beautify_loops( phase ); duke@435: return result; duke@435: } duke@435: duke@435: //------------------------------allpaths_check_safepts---------------------------- duke@435: // Allpaths backwards scan from loop tail, terminating each path at first safepoint duke@435: // encountered. Helper for check_safepts. duke@435: void IdealLoopTree::allpaths_check_safepts(VectorSet &visited, Node_List &stack) { duke@435: assert(stack.size() == 0, "empty stack"); duke@435: stack.push(_tail); duke@435: visited.Clear(); duke@435: visited.set(_tail->_idx); duke@435: while (stack.size() > 0) { duke@435: Node* n = stack.pop(); duke@435: if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) { duke@435: // Terminate this path duke@435: } else if (n->Opcode() == Op_SafePoint) { duke@435: if (_phase->get_loop(n) != this) { duke@435: if (_required_safept == NULL) _required_safept = new Node_List(); duke@435: _required_safept->push(n); // save the one closest to the tail duke@435: } duke@435: // Terminate this path duke@435: } else { duke@435: uint start = n->is_Region() ? 1 : 0; duke@435: uint end = n->is_Region() && !n->is_Loop() ? n->req() : start + 1; duke@435: for (uint i = start; i < end; i++) { duke@435: Node* in = n->in(i); duke@435: assert(in->is_CFG(), "must be"); duke@435: if (!visited.test_set(in->_idx) && is_member(_phase->get_loop(in))) { duke@435: stack.push(in); duke@435: } duke@435: } duke@435: } duke@435: } duke@435: } duke@435: duke@435: //------------------------------check_safepts---------------------------- duke@435: // Given dominators, try to find loops with calls that must always be duke@435: // executed (call dominates loop tail). These loops do not need non-call duke@435: // safepoints (ncsfpt). duke@435: // duke@435: // A complication is that a safepoint in a inner loop may be needed duke@435: // by an outer loop. In the following, the inner loop sees it has a duke@435: // call (block 3) on every path from the head (block 2) to the duke@435: // backedge (arc 3->2). So it deletes the ncsfpt (non-call safepoint) duke@435: // in block 2, _but_ this leaves the outer loop without a safepoint. duke@435: // duke@435: // entry 0 duke@435: // | duke@435: // v duke@435: // outer 1,2 +->1 duke@435: // | | duke@435: // | v duke@435: // | 2<---+ ncsfpt in 2 duke@435: // |_/|\ | duke@435: // | v | duke@435: // inner 2,3 / 3 | call in 3 duke@435: // / | | duke@435: // v +--+ duke@435: // exit 4 duke@435: // duke@435: // duke@435: // This method creates a list (_required_safept) of ncsfpt nodes that must duke@435: // be protected is created for each loop. When a ncsfpt maybe deleted, it duke@435: // is first looked for in the lists for the outer loops of the current loop. duke@435: // duke@435: // The insights into the problem: duke@435: // A) counted loops are okay duke@435: // B) innermost loops are okay (only an inner loop can delete duke@435: // a ncsfpt needed by an outer loop) duke@435: // C) a loop is immune from an inner loop deleting a safepoint duke@435: // if the loop has a call on the idom-path duke@435: // D) a loop is also immune if it has a ncsfpt (non-call safepoint) on the duke@435: // idom-path that is not in a nested loop duke@435: // E) otherwise, an ncsfpt on the idom-path that is nested in an inner duke@435: // loop needs to be prevented from deletion by an inner loop duke@435: // duke@435: // There are two analyses: duke@435: // 1) The first, and cheaper one, scans the loop body from duke@435: // tail to head following the idom (immediate dominator) duke@435: // chain, looking for the cases (C,D,E) above. duke@435: // Since inner loops are scanned before outer loops, there is summary duke@435: // information about inner loops. Inner loops can be skipped over duke@435: // when the tail of an inner loop is encountered. duke@435: // duke@435: // 2) The second, invoked if the first fails to find a call or ncsfpt on duke@435: // the idom path (which is rare), scans all predecessor control paths duke@435: // from the tail to the head, terminating a path when a call or sfpt duke@435: // is encountered, to find the ncsfpt's that are closest to the tail. duke@435: // duke@435: void IdealLoopTree::check_safepts(VectorSet &visited, Node_List &stack) { duke@435: // Bottom up traversal duke@435: IdealLoopTree* ch = _child; duke@435: while (ch != NULL) { duke@435: ch->check_safepts(visited, stack); duke@435: ch = ch->_next; duke@435: } duke@435: duke@435: if (!_head->is_CountedLoop() && !_has_sfpt && _parent != NULL && !_irreducible) { duke@435: bool has_call = false; // call on dom-path duke@435: bool has_local_ncsfpt = false; // ncsfpt on dom-path at this loop depth duke@435: Node* nonlocal_ncsfpt = NULL; // ncsfpt on dom-path at a deeper depth duke@435: // Scan the dom-path nodes from tail to head duke@435: for (Node* n = tail(); n != _head; n = _phase->idom(n)) { duke@435: if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) { duke@435: has_call = true; duke@435: _has_sfpt = 1; // Then no need for a safept! duke@435: break; duke@435: } else if (n->Opcode() == Op_SafePoint) { duke@435: if (_phase->get_loop(n) == this) { duke@435: has_local_ncsfpt = true; duke@435: break; duke@435: } duke@435: if (nonlocal_ncsfpt == NULL) { duke@435: nonlocal_ncsfpt = n; // save the one closest to the tail duke@435: } duke@435: } else { duke@435: IdealLoopTree* nlpt = _phase->get_loop(n); duke@435: if (this != nlpt) { duke@435: // If at an inner loop tail, see if the inner loop has already duke@435: // recorded seeing a call on the dom-path (and stop.) If not, duke@435: // jump to the head of the inner loop. duke@435: assert(is_member(nlpt), "nested loop"); duke@435: Node* tail = nlpt->_tail; duke@435: if (tail->in(0)->is_If()) tail = tail->in(0); duke@435: if (n == tail) { duke@435: // If inner loop has call on dom-path, so does outer loop duke@435: if (nlpt->_has_sfpt) { duke@435: has_call = true; duke@435: _has_sfpt = 1; duke@435: break; duke@435: } duke@435: // Skip to head of inner loop duke@435: assert(_phase->is_dominator(_head, nlpt->_head), "inner head dominated by outer head"); duke@435: n = nlpt->_head; duke@435: } duke@435: } duke@435: } duke@435: } duke@435: // Record safept's that this loop needs preserved when an duke@435: // inner loop attempts to delete it's safepoints. duke@435: if (_child != NULL && !has_call && !has_local_ncsfpt) { duke@435: if (nonlocal_ncsfpt != NULL) { duke@435: if (_required_safept == NULL) _required_safept = new Node_List(); duke@435: _required_safept->push(nonlocal_ncsfpt); duke@435: } else { duke@435: // Failed to find a suitable safept on the dom-path. Now use duke@435: // an all paths walk from tail to head, looking for safepoints to preserve. duke@435: allpaths_check_safepts(visited, stack); duke@435: } duke@435: } duke@435: } duke@435: } duke@435: duke@435: //---------------------------is_deleteable_safept---------------------------- duke@435: // Is safept not required by an outer loop? duke@435: bool PhaseIdealLoop::is_deleteable_safept(Node* sfpt) { duke@435: assert(sfpt->Opcode() == Op_SafePoint, ""); duke@435: IdealLoopTree* lp = get_loop(sfpt)->_parent; duke@435: while (lp != NULL) { duke@435: Node_List* sfpts = lp->_required_safept; duke@435: if (sfpts != NULL) { duke@435: for (uint i = 0; i < sfpts->size(); i++) { duke@435: if (sfpt == sfpts->at(i)) duke@435: return false; duke@435: } duke@435: } duke@435: lp = lp->_parent; duke@435: } duke@435: return true; duke@435: } duke@435: duke@435: //------------------------------counted_loop----------------------------------- duke@435: // Convert to counted loops where possible duke@435: void IdealLoopTree::counted_loop( PhaseIdealLoop *phase ) { duke@435: duke@435: // For grins, set the inner-loop flag here duke@435: if( !_child ) { duke@435: if( _head->is_Loop() ) _head->as_Loop()->set_inner_loop(); duke@435: } duke@435: duke@435: if( _head->is_CountedLoop() || duke@435: phase->is_counted_loop( _head, this ) ) { duke@435: _has_sfpt = 1; // Indicate we do not need a safepoint here duke@435: duke@435: // Look for a safepoint to remove duke@435: for (Node* n = tail(); n != _head; n = phase->idom(n)) duke@435: if (n->Opcode() == Op_SafePoint && phase->get_loop(n) == this && duke@435: phase->is_deleteable_safept(n)) duke@435: phase->lazy_replace(n,n->in(TypeFunc::Control)); duke@435: duke@435: CountedLoopNode *cl = _head->as_CountedLoop(); duke@435: Node *incr = cl->incr(); duke@435: if( !incr ) return; // Dead loop? duke@435: Node *init = cl->init_trip(); duke@435: Node *phi = cl->phi(); duke@435: // protect against stride not being a constant duke@435: if( !cl->stride_is_con() ) return; duke@435: int stride_con = cl->stride_con(); duke@435: duke@435: // Look for induction variables duke@435: duke@435: // Visit all children, looking for Phis duke@435: for (DUIterator i = cl->outs(); cl->has_out(i); i++) { duke@435: Node *out = cl->out(i); never@802: if (!out->is_Phi() || out == phi) continue; // Looking for other phis duke@435: PhiNode* phi2 = out->as_Phi(); duke@435: Node *incr2 = phi2->in( LoopNode::LoopBackControl ); duke@435: // Look for induction variables of the form: X += constant duke@435: if( phi2->region() != _head || duke@435: incr2->req() != 3 || duke@435: incr2->in(1) != phi2 || duke@435: incr2 == incr || duke@435: incr2->Opcode() != Op_AddI || duke@435: !incr2->in(2)->is_Con() ) duke@435: continue; duke@435: duke@435: // Check for parallel induction variable (parallel to trip counter) duke@435: // via an affine function. In particular, count-down loops with duke@435: // count-up array indices are common. We only RCE references off duke@435: // the trip-counter, so we need to convert all these to trip-counter duke@435: // expressions. duke@435: Node *init2 = phi2->in( LoopNode::EntryControl ); duke@435: int stride_con2 = incr2->in(2)->get_int(); duke@435: duke@435: // The general case here gets a little tricky. We want to find the duke@435: // GCD of all possible parallel IV's and make a new IV using this duke@435: // GCD for the loop. Then all possible IVs are simple multiples of duke@435: // the GCD. In practice, this will cover very few extra loops. duke@435: // Instead we require 'stride_con2' to be a multiple of 'stride_con', duke@435: // where +/-1 is the common case, but other integer multiples are duke@435: // also easy to handle. duke@435: int ratio_con = stride_con2/stride_con; duke@435: duke@435: if( ratio_con * stride_con == stride_con2 ) { // Check for exact duke@435: // Convert to using the trip counter. The parallel induction duke@435: // variable differs from the trip counter by a loop-invariant duke@435: // amount, the difference between their respective initial values. duke@435: // It is scaled by the 'ratio_con'. duke@435: Compile* C = phase->C; duke@435: Node* ratio = phase->_igvn.intcon(ratio_con); duke@435: phase->set_ctrl(ratio, C->root()); duke@435: Node* ratio_init = new (C, 3) MulINode(init, ratio); duke@435: phase->_igvn.register_new_node_with_optimizer(ratio_init, init); duke@435: phase->set_early_ctrl(ratio_init); duke@435: Node* diff = new (C, 3) SubINode(init2, ratio_init); duke@435: phase->_igvn.register_new_node_with_optimizer(diff, init2); duke@435: phase->set_early_ctrl(diff); duke@435: Node* ratio_idx = new (C, 3) MulINode(phi, ratio); duke@435: phase->_igvn.register_new_node_with_optimizer(ratio_idx, phi); duke@435: phase->set_ctrl(ratio_idx, cl); duke@435: Node* add = new (C, 3) AddINode(ratio_idx, diff); duke@435: phase->_igvn.register_new_node_with_optimizer(add); duke@435: phase->set_ctrl(add, cl); duke@435: phase->_igvn.hash_delete( phi2 ); duke@435: phase->_igvn.subsume_node( phi2, add ); duke@435: // Sometimes an induction variable is unused duke@435: if (add->outcnt() == 0) { duke@435: phase->_igvn.remove_dead_node(add); duke@435: } duke@435: --i; // deleted this phi; rescan starting with next position duke@435: continue; duke@435: } duke@435: } duke@435: } else if (_parent != NULL && !_irreducible) { duke@435: // Not a counted loop. duke@435: // Look for a safepoint on the idom-path to remove, preserving the first one duke@435: bool found = false; duke@435: Node* n = tail(); duke@435: for (; n != _head && !found; n = phase->idom(n)) { duke@435: if (n->Opcode() == Op_SafePoint && phase->get_loop(n) == this) duke@435: found = true; // Found one duke@435: } duke@435: // Skip past it and delete the others duke@435: for (; n != _head; n = phase->idom(n)) { duke@435: if (n->Opcode() == Op_SafePoint && phase->get_loop(n) == this && duke@435: phase->is_deleteable_safept(n)) duke@435: phase->lazy_replace(n,n->in(TypeFunc::Control)); duke@435: } duke@435: } duke@435: duke@435: // Recursively duke@435: if( _child ) _child->counted_loop( phase ); duke@435: if( _next ) _next ->counted_loop( phase ); duke@435: } duke@435: duke@435: #ifndef PRODUCT duke@435: //------------------------------dump_head-------------------------------------- duke@435: // Dump 1 liner for loop header info duke@435: void IdealLoopTree::dump_head( ) const { duke@435: for( uint i=0; i<_nest; i++ ) duke@435: tty->print(" "); duke@435: tty->print("Loop: N%d/N%d ",_head->_idx,_tail->_idx); duke@435: if( _irreducible ) tty->print(" IRREDUCIBLE"); duke@435: if( _head->is_CountedLoop() ) { duke@435: CountedLoopNode *cl = _head->as_CountedLoop(); duke@435: tty->print(" counted"); duke@435: if( cl->is_pre_loop () ) tty->print(" pre" ); duke@435: if( cl->is_main_loop() ) tty->print(" main"); duke@435: if( cl->is_post_loop() ) tty->print(" post"); duke@435: } duke@435: tty->cr(); duke@435: } duke@435: duke@435: //------------------------------dump------------------------------------------- duke@435: // Dump loops by loop tree duke@435: void IdealLoopTree::dump( ) const { duke@435: dump_head(); duke@435: if( _child ) _child->dump(); duke@435: if( _next ) _next ->dump(); duke@435: } duke@435: duke@435: #endif duke@435: never@802: static void log_loop_tree(IdealLoopTree* root, IdealLoopTree* loop, CompileLog* log) { never@802: if (loop == root) { never@802: if (loop->_child != NULL) { never@802: log->begin_head("loop_tree"); never@802: log->end_head(); never@802: if( loop->_child ) log_loop_tree(root, loop->_child, log); never@802: log->tail("loop_tree"); never@802: assert(loop->_next == NULL, "what?"); never@802: } never@802: } else { never@802: Node* head = loop->_head; never@802: log->begin_head("loop"); never@802: log->print(" idx='%d' ", head->_idx); never@802: if (loop->_irreducible) log->print("irreducible='1' "); never@802: if (head->is_Loop()) { never@802: if (head->as_Loop()->is_inner_loop()) log->print("inner_loop='1' "); never@802: if (head->as_Loop()->is_partial_peel_loop()) log->print("partial_peel_loop='1' "); never@802: } never@802: if (head->is_CountedLoop()) { never@802: CountedLoopNode* cl = head->as_CountedLoop(); never@802: if (cl->is_pre_loop()) log->print("pre_loop='%d' ", cl->main_idx()); never@802: if (cl->is_main_loop()) log->print("main_loop='%d' ", cl->_idx); never@802: if (cl->is_post_loop()) log->print("post_loop='%d' ", cl->main_idx()); never@802: } never@802: log->end_head(); never@802: if( loop->_child ) log_loop_tree(root, loop->_child, log); never@802: log->tail("loop"); never@802: if( loop->_next ) log_loop_tree(root, loop->_next, log); never@802: } never@802: } never@802: duke@435: //============================================================================= duke@435: //------------------------------PhaseIdealLoop--------------------------------- duke@435: // Create a PhaseLoop. Build the ideal Loop tree. Map each Ideal Node to duke@435: // its corresponding LoopNode. If 'optimize' is true, do some loop cleanups. duke@435: PhaseIdealLoop::PhaseIdealLoop( PhaseIterGVN &igvn, const PhaseIdealLoop *verify_me, bool do_split_ifs ) duke@435: : PhaseTransform(Ideal_Loop), duke@435: _igvn(igvn), duke@435: _dom_lca_tags(C->comp_arena()) { duke@435: // Reset major-progress flag for the driver's heuristics duke@435: C->clear_major_progress(); duke@435: duke@435: #ifndef PRODUCT duke@435: // Capture for later assert duke@435: uint unique = C->unique(); duke@435: _loop_invokes++; duke@435: _loop_work += unique; duke@435: #endif duke@435: duke@435: // True if the method has at least 1 irreducible loop duke@435: _has_irreducible_loops = false; duke@435: duke@435: _created_loop_node = false; duke@435: duke@435: Arena *a = Thread::current()->resource_area(); duke@435: VectorSet visited(a); duke@435: // Pre-grow the mapping from Nodes to IdealLoopTrees. duke@435: _nodes.map(C->unique(), NULL); duke@435: memset(_nodes.adr(), 0, wordSize * C->unique()); duke@435: duke@435: // Pre-build the top-level outermost loop tree entry duke@435: _ltree_root = new IdealLoopTree( this, C->root(), C->root() ); duke@435: // Do not need a safepoint at the top level duke@435: _ltree_root->_has_sfpt = 1; duke@435: duke@435: // Empty pre-order array duke@435: allocate_preorders(); duke@435: duke@435: // Build a loop tree on the fly. Build a mapping from CFG nodes to duke@435: // IdealLoopTree entries. Data nodes are NOT walked. duke@435: build_loop_tree(); duke@435: // Check for bailout, and return duke@435: if (C->failing()) { duke@435: return; duke@435: } duke@435: duke@435: // No loops after all duke@435: if( !_ltree_root->_child ) C->set_has_loops(false); duke@435: duke@435: // There should always be an outer loop containing the Root and Return nodes. duke@435: // If not, we have a degenerate empty program. Bail out in this case. duke@435: if (!has_node(C->root())) { duke@435: C->clear_major_progress(); duke@435: C->record_method_not_compilable("empty program detected during loop optimization"); duke@435: return; duke@435: } duke@435: duke@435: // Nothing to do, so get out duke@435: if( !C->has_loops() && !do_split_ifs && !verify_me) { duke@435: _igvn.optimize(); // Cleanup NeverBranches duke@435: return; duke@435: } duke@435: duke@435: // Set loop nesting depth duke@435: _ltree_root->set_nest( 0 ); duke@435: duke@435: // Split shared headers and insert loop landing pads. duke@435: // Do not bother doing this on the Root loop of course. duke@435: if( !verify_me && _ltree_root->_child ) { duke@435: if( _ltree_root->_child->beautify_loops( this ) ) { duke@435: // Re-build loop tree! duke@435: _ltree_root->_child = NULL; duke@435: _nodes.clear(); duke@435: reallocate_preorders(); duke@435: build_loop_tree(); duke@435: // Check for bailout, and return duke@435: if (C->failing()) { duke@435: return; duke@435: } duke@435: // Reset loop nesting depth duke@435: _ltree_root->set_nest( 0 ); never@657: never@657: C->print_method("After beautify loops", 3); duke@435: } duke@435: } duke@435: duke@435: // Build Dominators for elision of NULL checks & loop finding. duke@435: // Since nodes do not have a slot for immediate dominator, make twisti@1040: // a persistent side array for that info indexed on node->_idx. duke@435: _idom_size = C->unique(); duke@435: _idom = NEW_RESOURCE_ARRAY( Node*, _idom_size ); duke@435: _dom_depth = NEW_RESOURCE_ARRAY( uint, _idom_size ); duke@435: _dom_stk = NULL; // Allocated on demand in recompute_dom_depth duke@435: memset( _dom_depth, 0, _idom_size * sizeof(uint) ); duke@435: duke@435: Dominators(); duke@435: duke@435: // As a side effect, Dominators removed any unreachable CFG paths duke@435: // into RegionNodes. It doesn't do this test against Root, so duke@435: // we do it here. duke@435: for( uint i = 1; i < C->root()->req(); i++ ) { duke@435: if( !_nodes[C->root()->in(i)->_idx] ) { // Dead path into Root? duke@435: _igvn.hash_delete(C->root()); duke@435: C->root()->del_req(i); duke@435: _igvn._worklist.push(C->root()); duke@435: i--; // Rerun same iteration on compressed edges duke@435: } duke@435: } duke@435: duke@435: // Given dominators, try to find inner loops with calls that must duke@435: // always be executed (call dominates loop tail). These loops do twisti@1040: // not need a separate safepoint. duke@435: Node_List cisstack(a); duke@435: _ltree_root->check_safepts(visited, cisstack); duke@435: duke@435: // Walk the DATA nodes and place into loops. Find earliest control duke@435: // node. For CFG nodes, the _nodes array starts out and remains duke@435: // holding the associated IdealLoopTree pointer. For DATA nodes, the duke@435: // _nodes array holds the earliest legal controlling CFG node. duke@435: duke@435: // Allocate stack with enough space to avoid frequent realloc duke@435: int stack_size = (C->unique() >> 1) + 16; // (unique>>1)+16 from Java2D stats duke@435: Node_Stack nstack( a, stack_size ); duke@435: duke@435: visited.Clear(); duke@435: Node_List worklist(a); duke@435: // Don't need C->root() on worklist since duke@435: // it will be processed among C->top() inputs duke@435: worklist.push( C->top() ); duke@435: visited.set( C->top()->_idx ); // Set C->top() as visited now duke@435: build_loop_early( visited, worklist, nstack, verify_me ); duke@435: duke@435: // Given early legal placement, try finding counted loops. This placement duke@435: // is good enough to discover most loop invariants. duke@435: if( !verify_me ) duke@435: _ltree_root->counted_loop( this ); duke@435: duke@435: // Find latest loop placement. Find ideal loop placement. duke@435: visited.Clear(); duke@435: init_dom_lca_tags(); duke@435: // Need C->root() on worklist when processing outs duke@435: worklist.push( C->root() ); duke@435: NOT_PRODUCT( C->verify_graph_edges(); ) duke@435: worklist.push( C->top() ); duke@435: build_loop_late( visited, worklist, nstack, verify_me ); duke@435: duke@435: // clear out the dead code duke@435: while(_deadlist.size()) { duke@435: igvn.remove_globally_dead_node(_deadlist.pop()); duke@435: } duke@435: duke@435: #ifndef PRODUCT duke@435: C->verify_graph_edges(); duke@435: if( verify_me ) { // Nested verify pass? duke@435: // Check to see if the verify mode is broken duke@435: assert(C->unique() == unique, "non-optimize mode made Nodes? ? ?"); duke@435: return; duke@435: } duke@435: if( VerifyLoopOptimizations ) verify(); duke@435: #endif duke@435: duke@435: if (ReassociateInvariants) { duke@435: // Reassociate invariants and prep for split_thru_phi duke@435: for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) { duke@435: IdealLoopTree* lpt = iter.current(); duke@435: if (!lpt->is_counted() || !lpt->is_inner()) continue; duke@435: duke@435: lpt->reassociate_invariants(this); duke@435: duke@435: // Because RCE opportunities can be masked by split_thru_phi, duke@435: // look for RCE candidates and inhibit split_thru_phi duke@435: // on just their loop-phi's for this pass of loop opts duke@435: if( SplitIfBlocks && do_split_ifs ) { duke@435: if (lpt->policy_range_check(this)) { kvn@474: lpt->_rce_candidate = 1; // = true duke@435: } duke@435: } duke@435: } duke@435: } duke@435: duke@435: // Check for aggressive application of split-if and other transforms duke@435: // that require basic-block info (like cloning through Phi's) duke@435: if( SplitIfBlocks && do_split_ifs ) { duke@435: visited.Clear(); duke@435: split_if_with_blocks( visited, nstack ); duke@435: NOT_PRODUCT( if( VerifyLoopOptimizations ) verify(); ); duke@435: } duke@435: duke@435: // Perform iteration-splitting on inner loops. Split iterations to avoid duke@435: // range checks or one-shot null checks. duke@435: duke@435: // If split-if's didn't hack the graph too bad (no CFG changes) duke@435: // then do loop opts. duke@435: if( C->has_loops() && !C->major_progress() ) { duke@435: memset( worklist.adr(), 0, worklist.Size()*sizeof(Node*) ); duke@435: _ltree_root->_child->iteration_split( this, worklist ); duke@435: // No verify after peeling! GCM has hoisted code out of the loop. duke@435: // After peeling, the hoisted code could sink inside the peeled area. duke@435: // The peeling code does not try to recompute the best location for duke@435: // all the code before the peeled area, so the verify pass will always duke@435: // complain about it. duke@435: } duke@435: // Do verify graph edges in any case duke@435: NOT_PRODUCT( C->verify_graph_edges(); ); duke@435: duke@435: if( !do_split_ifs ) { duke@435: // We saw major progress in Split-If to get here. We forced a duke@435: // pass with unrolling and not split-if, however more split-if's duke@435: // might make progress. If the unrolling didn't make progress duke@435: // then the major-progress flag got cleared and we won't try duke@435: // another round of Split-If. In particular the ever-common duke@435: // instance-of/check-cast pattern requires at least 2 rounds of duke@435: // Split-If to clear out. duke@435: C->set_major_progress(); duke@435: } duke@435: duke@435: // Repeat loop optimizations if new loops were seen duke@435: if (created_loop_node()) { duke@435: C->set_major_progress(); duke@435: } duke@435: duke@435: // Convert scalar to superword operations duke@435: duke@435: if (UseSuperWord && C->has_loops() && !C->major_progress()) { duke@435: // SuperWord transform duke@435: SuperWord sw(this); duke@435: for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) { duke@435: IdealLoopTree* lpt = iter.current(); duke@435: if (lpt->is_counted()) { duke@435: sw.transform_loop(lpt); duke@435: } duke@435: } duke@435: } duke@435: duke@435: // Cleanup any modified bits duke@435: _igvn.optimize(); duke@435: never@802: // disable assert until issue with split_flow_path is resolved (6742111) never@802: // assert(!_has_irreducible_loops || C->parsed_irreducible_loop() || C->is_osr_compilation(), never@802: // "shouldn't introduce irreducible loops"); never@802: never@802: if (C->log() != NULL) { never@802: log_loop_tree(_ltree_root, _ltree_root, C->log()); never@802: } duke@435: } duke@435: duke@435: #ifndef PRODUCT duke@435: //------------------------------print_statistics------------------------------- duke@435: int PhaseIdealLoop::_loop_invokes=0;// Count of PhaseIdealLoop invokes duke@435: int PhaseIdealLoop::_loop_work=0; // Sum of PhaseIdealLoop x unique duke@435: void PhaseIdealLoop::print_statistics() { duke@435: tty->print_cr("PhaseIdealLoop=%d, sum _unique=%d", _loop_invokes, _loop_work); duke@435: } duke@435: duke@435: //------------------------------verify----------------------------------------- duke@435: // Build a verify-only PhaseIdealLoop, and see that it agrees with me. duke@435: static int fail; // debug only, so its multi-thread dont care duke@435: void PhaseIdealLoop::verify() const { duke@435: int old_progress = C->major_progress(); duke@435: ResourceMark rm; duke@435: PhaseIdealLoop loop_verify( _igvn, this, false ); duke@435: VectorSet visited(Thread::current()->resource_area()); duke@435: duke@435: fail = 0; duke@435: verify_compare( C->root(), &loop_verify, visited ); duke@435: assert( fail == 0, "verify loops failed" ); duke@435: // Verify loop structure is the same duke@435: _ltree_root->verify_tree(loop_verify._ltree_root, NULL); duke@435: // Reset major-progress. It was cleared by creating a verify version of duke@435: // PhaseIdealLoop. duke@435: for( int i=0; iset_major_progress(); duke@435: } duke@435: duke@435: //------------------------------verify_compare--------------------------------- duke@435: // Make sure me and the given PhaseIdealLoop agree on key data structures duke@435: void PhaseIdealLoop::verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const { duke@435: if( !n ) return; duke@435: if( visited.test_set( n->_idx ) ) return; duke@435: if( !_nodes[n->_idx] ) { // Unreachable duke@435: assert( !loop_verify->_nodes[n->_idx], "both should be unreachable" ); duke@435: return; duke@435: } duke@435: duke@435: uint i; duke@435: for( i = 0; i < n->req(); i++ ) duke@435: verify_compare( n->in(i), loop_verify, visited ); duke@435: duke@435: // Check the '_nodes' block/loop structure duke@435: i = n->_idx; duke@435: if( has_ctrl(n) ) { // We have control; verify has loop or ctrl duke@435: if( _nodes[i] != loop_verify->_nodes[i] && duke@435: get_ctrl_no_update(n) != loop_verify->get_ctrl_no_update(n) ) { duke@435: tty->print("Mismatched control setting for: "); duke@435: n->dump(); duke@435: if( fail++ > 10 ) return; duke@435: Node *c = get_ctrl_no_update(n); duke@435: tty->print("We have it as: "); duke@435: if( c->in(0) ) c->dump(); duke@435: else tty->print_cr("N%d",c->_idx); duke@435: tty->print("Verify thinks: "); duke@435: if( loop_verify->has_ctrl(n) ) duke@435: loop_verify->get_ctrl_no_update(n)->dump(); duke@435: else duke@435: loop_verify->get_loop_idx(n)->dump(); duke@435: tty->cr(); duke@435: } duke@435: } else { // We have a loop duke@435: IdealLoopTree *us = get_loop_idx(n); duke@435: if( loop_verify->has_ctrl(n) ) { duke@435: tty->print("Mismatched loop setting for: "); duke@435: n->dump(); duke@435: if( fail++ > 10 ) return; duke@435: tty->print("We have it as: "); duke@435: us->dump(); duke@435: tty->print("Verify thinks: "); duke@435: loop_verify->get_ctrl_no_update(n)->dump(); duke@435: tty->cr(); duke@435: } else if (!C->major_progress()) { duke@435: // Loop selection can be messed up if we did a major progress duke@435: // operation, like split-if. Do not verify in that case. duke@435: IdealLoopTree *them = loop_verify->get_loop_idx(n); duke@435: if( us->_head != them->_head || us->_tail != them->_tail ) { duke@435: tty->print("Unequals loops for: "); duke@435: n->dump(); duke@435: if( fail++ > 10 ) return; duke@435: tty->print("We have it as: "); duke@435: us->dump(); duke@435: tty->print("Verify thinks: "); duke@435: them->dump(); duke@435: tty->cr(); duke@435: } duke@435: } duke@435: } duke@435: duke@435: // Check for immediate dominators being equal duke@435: if( i >= _idom_size ) { duke@435: if( !n->is_CFG() ) return; duke@435: tty->print("CFG Node with no idom: "); duke@435: n->dump(); duke@435: return; duke@435: } duke@435: if( !n->is_CFG() ) return; duke@435: if( n == C->root() ) return; // No IDOM here duke@435: duke@435: assert(n->_idx == i, "sanity"); duke@435: Node *id = idom_no_update(n); duke@435: if( id != loop_verify->idom_no_update(n) ) { duke@435: tty->print("Unequals idoms for: "); duke@435: n->dump(); duke@435: if( fail++ > 10 ) return; duke@435: tty->print("We have it as: "); duke@435: id->dump(); duke@435: tty->print("Verify thinks: "); duke@435: loop_verify->idom_no_update(n)->dump(); duke@435: tty->cr(); duke@435: } duke@435: duke@435: } duke@435: duke@435: //------------------------------verify_tree------------------------------------ duke@435: // Verify that tree structures match. Because the CFG can change, siblings duke@435: // within the loop tree can be reordered. We attempt to deal with that by duke@435: // reordering the verify's loop tree if possible. duke@435: void IdealLoopTree::verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const { duke@435: assert( _parent == parent, "Badly formed loop tree" ); duke@435: duke@435: // Siblings not in same order? Attempt to re-order. duke@435: if( _head != loop->_head ) { duke@435: // Find _next pointer to update duke@435: IdealLoopTree **pp = &loop->_parent->_child; duke@435: while( *pp != loop ) duke@435: pp = &((*pp)->_next); duke@435: // Find proper sibling to be next duke@435: IdealLoopTree **nn = &loop->_next; duke@435: while( (*nn) && (*nn)->_head != _head ) duke@435: nn = &((*nn)->_next); duke@435: duke@435: // Check for no match. duke@435: if( !(*nn) ) { duke@435: // Annoyingly, irreducible loops can pick different headers duke@435: // after a major_progress operation, so the rest of the loop duke@435: // tree cannot be matched. duke@435: if (_irreducible && Compile::current()->major_progress()) return; duke@435: assert( 0, "failed to match loop tree" ); duke@435: } duke@435: duke@435: // Move (*nn) to (*pp) duke@435: IdealLoopTree *hit = *nn; duke@435: *nn = hit->_next; duke@435: hit->_next = loop; duke@435: *pp = loop; duke@435: loop = hit; duke@435: // Now try again to verify duke@435: } duke@435: duke@435: assert( _head == loop->_head , "mismatched loop head" ); duke@435: Node *tail = _tail; // Inline a non-updating version of duke@435: while( !tail->in(0) ) // the 'tail()' call. duke@435: tail = tail->in(1); duke@435: assert( tail == loop->_tail, "mismatched loop tail" ); duke@435: duke@435: // Counted loops that are guarded should be able to find their guards duke@435: if( _head->is_CountedLoop() && _head->as_CountedLoop()->is_main_loop() ) { duke@435: CountedLoopNode *cl = _head->as_CountedLoop(); duke@435: Node *init = cl->init_trip(); duke@435: Node *ctrl = cl->in(LoopNode::EntryControl); duke@435: assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" ); duke@435: Node *iff = ctrl->in(0); duke@435: assert( iff->Opcode() == Op_If, "" ); duke@435: Node *bol = iff->in(1); duke@435: assert( bol->Opcode() == Op_Bool, "" ); duke@435: Node *cmp = bol->in(1); duke@435: assert( cmp->Opcode() == Op_CmpI, "" ); duke@435: Node *add = cmp->in(1); duke@435: Node *opaq; duke@435: if( add->Opcode() == Op_Opaque1 ) { duke@435: opaq = add; duke@435: } else { duke@435: assert( add->Opcode() == Op_AddI || add->Opcode() == Op_ConI , "" ); duke@435: assert( add == init, "" ); duke@435: opaq = cmp->in(2); duke@435: } duke@435: assert( opaq->Opcode() == Op_Opaque1, "" ); duke@435: duke@435: } duke@435: duke@435: if (_child != NULL) _child->verify_tree(loop->_child, this); duke@435: if (_next != NULL) _next ->verify_tree(loop->_next, parent); duke@435: // Innermost loops need to verify loop bodies, duke@435: // but only if no 'major_progress' duke@435: int fail = 0; duke@435: if (!Compile::current()->major_progress() && _child == NULL) { duke@435: for( uint i = 0; i < _body.size(); i++ ) { duke@435: Node *n = _body.at(i); duke@435: if (n->outcnt() == 0) continue; // Ignore dead duke@435: uint j; duke@435: for( j = 0; j < loop->_body.size(); j++ ) duke@435: if( loop->_body.at(j) == n ) duke@435: break; duke@435: if( j == loop->_body.size() ) { // Not found in loop body duke@435: // Last ditch effort to avoid assertion: Its possible that we duke@435: // have some users (so outcnt not zero) but are still dead. duke@435: // Try to find from root. duke@435: if (Compile::current()->root()->find(n->_idx)) { duke@435: fail++; duke@435: tty->print("We have that verify does not: "); duke@435: n->dump(); duke@435: } duke@435: } duke@435: } duke@435: for( uint i2 = 0; i2 < loop->_body.size(); i2++ ) { duke@435: Node *n = loop->_body.at(i2); duke@435: if (n->outcnt() == 0) continue; // Ignore dead duke@435: uint j; duke@435: for( j = 0; j < _body.size(); j++ ) duke@435: if( _body.at(j) == n ) duke@435: break; duke@435: if( j == _body.size() ) { // Not found in loop body duke@435: // Last ditch effort to avoid assertion: Its possible that we duke@435: // have some users (so outcnt not zero) but are still dead. duke@435: // Try to find from root. duke@435: if (Compile::current()->root()->find(n->_idx)) { duke@435: fail++; duke@435: tty->print("Verify has that we do not: "); duke@435: n->dump(); duke@435: } duke@435: } duke@435: } duke@435: assert( !fail, "loop body mismatch" ); duke@435: } duke@435: } duke@435: duke@435: #endif duke@435: duke@435: //------------------------------set_idom--------------------------------------- duke@435: void PhaseIdealLoop::set_idom(Node* d, Node* n, uint dom_depth) { duke@435: uint idx = d->_idx; duke@435: if (idx >= _idom_size) { duke@435: uint newsize = _idom_size<<1; duke@435: while( idx >= newsize ) { duke@435: newsize <<= 1; duke@435: } duke@435: _idom = REALLOC_RESOURCE_ARRAY( Node*, _idom,_idom_size,newsize); duke@435: _dom_depth = REALLOC_RESOURCE_ARRAY( uint, _dom_depth,_idom_size,newsize); duke@435: memset( _dom_depth + _idom_size, 0, (newsize - _idom_size) * sizeof(uint) ); duke@435: _idom_size = newsize; duke@435: } duke@435: _idom[idx] = n; duke@435: _dom_depth[idx] = dom_depth; duke@435: } duke@435: duke@435: //------------------------------recompute_dom_depth--------------------------------------- duke@435: // The dominator tree is constructed with only parent pointers. duke@435: // This recomputes the depth in the tree by first tagging all duke@435: // nodes as "no depth yet" marker. The next pass then runs up duke@435: // the dom tree from each node marked "no depth yet", and computes duke@435: // the depth on the way back down. duke@435: void PhaseIdealLoop::recompute_dom_depth() { duke@435: uint no_depth_marker = C->unique(); duke@435: uint i; duke@435: // Initialize depth to "no depth yet" duke@435: for (i = 0; i < _idom_size; i++) { duke@435: if (_dom_depth[i] > 0 && _idom[i] != NULL) { duke@435: _dom_depth[i] = no_depth_marker; duke@435: } duke@435: } duke@435: if (_dom_stk == NULL) { duke@435: uint init_size = C->unique() / 100; // Guess that 1/100 is a reasonable initial size. duke@435: if (init_size < 10) init_size = 10; duke@435: _dom_stk = new (C->node_arena()) GrowableArray(C->node_arena(), init_size, 0, 0); duke@435: } duke@435: // Compute new depth for each node. duke@435: for (i = 0; i < _idom_size; i++) { duke@435: uint j = i; duke@435: // Run up the dom tree to find a node with a depth duke@435: while (_dom_depth[j] == no_depth_marker) { duke@435: _dom_stk->push(j); duke@435: j = _idom[j]->_idx; duke@435: } duke@435: // Compute the depth on the way back down this tree branch duke@435: uint dd = _dom_depth[j] + 1; duke@435: while (_dom_stk->length() > 0) { duke@435: uint j = _dom_stk->pop(); duke@435: _dom_depth[j] = dd; duke@435: dd++; duke@435: } duke@435: } duke@435: } duke@435: duke@435: //------------------------------sort------------------------------------------- duke@435: // Insert 'loop' into the existing loop tree. 'innermost' is a leaf of the duke@435: // loop tree, not the root. duke@435: IdealLoopTree *PhaseIdealLoop::sort( IdealLoopTree *loop, IdealLoopTree *innermost ) { duke@435: if( !innermost ) return loop; // New innermost loop duke@435: duke@435: int loop_preorder = get_preorder(loop->_head); // Cache pre-order number duke@435: assert( loop_preorder, "not yet post-walked loop" ); duke@435: IdealLoopTree **pp = &innermost; // Pointer to previous next-pointer duke@435: IdealLoopTree *l = *pp; // Do I go before or after 'l'? duke@435: duke@435: // Insert at start of list duke@435: while( l ) { // Insertion sort based on pre-order duke@435: if( l == loop ) return innermost; // Already on list! duke@435: int l_preorder = get_preorder(l->_head); // Cache pre-order number duke@435: assert( l_preorder, "not yet post-walked l" ); duke@435: // Check header pre-order number to figure proper nesting duke@435: if( loop_preorder > l_preorder ) duke@435: break; // End of insertion duke@435: // If headers tie (e.g., shared headers) check tail pre-order numbers. duke@435: // Since I split shared headers, you'd think this could not happen. duke@435: // BUT: I must first do the preorder numbering before I can discover I duke@435: // have shared headers, so the split headers all get the same preorder duke@435: // number as the RegionNode they split from. duke@435: if( loop_preorder == l_preorder && duke@435: get_preorder(loop->_tail) < get_preorder(l->_tail) ) duke@435: break; // Also check for shared headers (same pre#) duke@435: pp = &l->_parent; // Chain up list duke@435: l = *pp; duke@435: } duke@435: // Link into list duke@435: // Point predecessor to me duke@435: *pp = loop; duke@435: // Point me to successor duke@435: IdealLoopTree *p = loop->_parent; duke@435: loop->_parent = l; // Point me to successor duke@435: if( p ) sort( p, innermost ); // Insert my parents into list as well duke@435: return innermost; duke@435: } duke@435: duke@435: //------------------------------build_loop_tree-------------------------------- duke@435: // I use a modified Vick/Tarjan algorithm. I need pre- and a post- visit duke@435: // bits. The _nodes[] array is mapped by Node index and holds a NULL for duke@435: // not-yet-pre-walked, pre-order # for pre-but-not-post-walked and holds the duke@435: // tightest enclosing IdealLoopTree for post-walked. duke@435: // duke@435: // During my forward walk I do a short 1-layer lookahead to see if I can find duke@435: // a loop backedge with that doesn't have any work on the backedge. This duke@435: // helps me construct nested loops with shared headers better. duke@435: // duke@435: // Once I've done the forward recursion, I do the post-work. For each child duke@435: // I check to see if there is a backedge. Backedges define a loop! I duke@435: // insert an IdealLoopTree at the target of the backedge. duke@435: // duke@435: // During the post-work I also check to see if I have several children duke@435: // belonging to different loops. If so, then this Node is a decision point duke@435: // where control flow can choose to change loop nests. It is at this duke@435: // decision point where I can figure out how loops are nested. At this duke@435: // time I can properly order the different loop nests from my children. duke@435: // Note that there may not be any backedges at the decision point! duke@435: // duke@435: // Since the decision point can be far removed from the backedges, I can't duke@435: // order my loops at the time I discover them. Thus at the decision point duke@435: // I need to inspect loop header pre-order numbers to properly nest my duke@435: // loops. This means I need to sort my childrens' loops by pre-order. duke@435: // The sort is of size number-of-control-children, which generally limits duke@435: // it to size 2 (i.e., I just choose between my 2 target loops). duke@435: void PhaseIdealLoop::build_loop_tree() { duke@435: // Allocate stack of size C->unique()/2 to avoid frequent realloc duke@435: GrowableArray bltstack(C->unique() >> 1); duke@435: Node *n = C->root(); duke@435: bltstack.push(n); duke@435: int pre_order = 1; duke@435: int stack_size; duke@435: duke@435: while ( ( stack_size = bltstack.length() ) != 0 ) { duke@435: n = bltstack.top(); // Leave node on stack duke@435: if ( !is_visited(n) ) { duke@435: // ---- Pre-pass Work ---- duke@435: // Pre-walked but not post-walked nodes need a pre_order number. duke@435: duke@435: set_preorder_visited( n, pre_order ); // set as visited duke@435: duke@435: // ---- Scan over children ---- duke@435: // Scan first over control projections that lead to loop headers. duke@435: // This helps us find inner-to-outer loops with shared headers better. duke@435: duke@435: // Scan children's children for loop headers. duke@435: for ( int i = n->outcnt() - 1; i >= 0; --i ) { duke@435: Node* m = n->raw_out(i); // Child duke@435: if( m->is_CFG() && !is_visited(m) ) { // Only for CFG children duke@435: // Scan over children's children to find loop duke@435: for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) { duke@435: Node* l = m->fast_out(j); duke@435: if( is_visited(l) && // Been visited? duke@435: !is_postvisited(l) && // But not post-visited duke@435: get_preorder(l) < pre_order ) { // And smaller pre-order duke@435: // Found! Scan the DFS down this path before doing other paths duke@435: bltstack.push(m); duke@435: break; duke@435: } duke@435: } duke@435: } duke@435: } duke@435: pre_order++; duke@435: } duke@435: else if ( !is_postvisited(n) ) { duke@435: // Note: build_loop_tree_impl() adds out edges on rare occasions, duke@435: // such as com.sun.rsasign.am::a. duke@435: // For non-recursive version, first, process current children. duke@435: // On next iteration, check if additional children were added. duke@435: for ( int k = n->outcnt() - 1; k >= 0; --k ) { duke@435: Node* u = n->raw_out(k); duke@435: if ( u->is_CFG() && !is_visited(u) ) { duke@435: bltstack.push(u); duke@435: } duke@435: } duke@435: if ( bltstack.length() == stack_size ) { duke@435: // There were no additional children, post visit node now duke@435: (void)bltstack.pop(); // Remove node from stack duke@435: pre_order = build_loop_tree_impl( n, pre_order ); duke@435: // Check for bailout duke@435: if (C->failing()) { duke@435: return; duke@435: } duke@435: // Check to grow _preorders[] array for the case when duke@435: // build_loop_tree_impl() adds new nodes. duke@435: check_grow_preorders(); duke@435: } duke@435: } duke@435: else { duke@435: (void)bltstack.pop(); // Remove post-visited node from stack duke@435: } duke@435: } duke@435: } duke@435: duke@435: //------------------------------build_loop_tree_impl--------------------------- duke@435: int PhaseIdealLoop::build_loop_tree_impl( Node *n, int pre_order ) { duke@435: // ---- Post-pass Work ---- duke@435: // Pre-walked but not post-walked nodes need a pre_order number. duke@435: duke@435: // Tightest enclosing loop for this Node duke@435: IdealLoopTree *innermost = NULL; duke@435: duke@435: // For all children, see if any edge is a backedge. If so, make a loop duke@435: // for it. Then find the tightest enclosing loop for the self Node. duke@435: for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { duke@435: Node* m = n->fast_out(i); // Child duke@435: if( n == m ) continue; // Ignore control self-cycles duke@435: if( !m->is_CFG() ) continue;// Ignore non-CFG edges duke@435: duke@435: IdealLoopTree *l; // Child's loop duke@435: if( !is_postvisited(m) ) { // Child visited but not post-visited? duke@435: // Found a backedge duke@435: assert( get_preorder(m) < pre_order, "should be backedge" ); duke@435: // Check for the RootNode, which is already a LoopNode and is allowed duke@435: // to have multiple "backedges". duke@435: if( m == C->root()) { // Found the root? duke@435: l = _ltree_root; // Root is the outermost LoopNode duke@435: } else { // Else found a nested loop duke@435: // Insert a LoopNode to mark this loop. duke@435: l = new IdealLoopTree(this, m, n); duke@435: } // End of Else found a nested loop duke@435: if( !has_loop(m) ) // If 'm' does not already have a loop set duke@435: set_loop(m, l); // Set loop header to loop now duke@435: duke@435: } else { // Else not a nested loop duke@435: if( !_nodes[m->_idx] ) continue; // Dead code has no loop duke@435: l = get_loop(m); // Get previously determined loop duke@435: // If successor is header of a loop (nest), move up-loop till it duke@435: // is a member of some outer enclosing loop. Since there are no duke@435: // shared headers (I've split them already) I only need to go up duke@435: // at most 1 level. duke@435: while( l && l->_head == m ) // Successor heads loop? duke@435: l = l->_parent; // Move up 1 for me duke@435: // If this loop is not properly parented, then this loop duke@435: // has no exit path out, i.e. its an infinite loop. duke@435: if( !l ) { duke@435: // Make loop "reachable" from root so the CFG is reachable. Basically duke@435: // insert a bogus loop exit that is never taken. 'm', the loop head, duke@435: // points to 'n', one (of possibly many) fall-in paths. There may be duke@435: // many backedges as well. duke@435: duke@435: // Here I set the loop to be the root loop. I could have, after duke@435: // inserting a bogus loop exit, restarted the recursion and found my duke@435: // new loop exit. This would make the infinite loop a first-class duke@435: // loop and it would then get properly optimized. What's the use of duke@435: // optimizing an infinite loop? duke@435: l = _ltree_root; // Oops, found infinite loop duke@435: duke@435: // Insert the NeverBranch between 'm' and it's control user. duke@435: NeverBranchNode *iff = new (C, 1) NeverBranchNode( m ); duke@435: _igvn.register_new_node_with_optimizer(iff); duke@435: set_loop(iff, l); duke@435: Node *if_t = new (C, 1) CProjNode( iff, 0 ); duke@435: _igvn.register_new_node_with_optimizer(if_t); duke@435: set_loop(if_t, l); duke@435: duke@435: Node* cfg = NULL; // Find the One True Control User of m duke@435: for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) { duke@435: Node* x = m->fast_out(j); duke@435: if (x->is_CFG() && x != m && x != iff) duke@435: { cfg = x; break; } duke@435: } duke@435: assert(cfg != NULL, "must find the control user of m"); duke@435: uint k = 0; // Probably cfg->in(0) duke@435: while( cfg->in(k) != m ) k++; // But check incase cfg is a Region duke@435: cfg->set_req( k, if_t ); // Now point to NeverBranch duke@435: duke@435: // Now create the never-taken loop exit duke@435: Node *if_f = new (C, 1) CProjNode( iff, 1 ); duke@435: _igvn.register_new_node_with_optimizer(if_f); duke@435: set_loop(if_f, l); duke@435: // Find frame ptr for Halt. Relies on the optimizer duke@435: // V-N'ing. Easier and quicker than searching through duke@435: // the program structure. duke@435: Node *frame = new (C, 1) ParmNode( C->start(), TypeFunc::FramePtr ); duke@435: _igvn.register_new_node_with_optimizer(frame); duke@435: // Halt & Catch Fire duke@435: Node *halt = new (C, TypeFunc::Parms) HaltNode( if_f, frame ); duke@435: _igvn.register_new_node_with_optimizer(halt); duke@435: set_loop(halt, l); duke@435: C->root()->add_req(halt); duke@435: set_loop(C->root(), _ltree_root); duke@435: } duke@435: } duke@435: // Weeny check for irreducible. This child was already visited (this duke@435: // IS the post-work phase). Is this child's loop header post-visited duke@435: // as well? If so, then I found another entry into the loop. duke@435: while( is_postvisited(l->_head) ) { duke@435: // found irreducible kvn@474: l->_irreducible = 1; // = true duke@435: l = l->_parent; duke@435: _has_irreducible_loops = true; duke@435: // Check for bad CFG here to prevent crash, and bailout of compile duke@435: if (l == NULL) { duke@435: C->record_method_not_compilable("unhandled CFG detected during loop optimization"); duke@435: return pre_order; duke@435: } duke@435: } duke@435: duke@435: // This Node might be a decision point for loops. It is only if duke@435: // it's children belong to several different loops. The sort call duke@435: // does a trivial amount of work if there is only 1 child or all duke@435: // children belong to the same loop. If however, the children duke@435: // belong to different loops, the sort call will properly set the duke@435: // _parent pointers to show how the loops nest. duke@435: // duke@435: // In any case, it returns the tightest enclosing loop. duke@435: innermost = sort( l, innermost ); duke@435: } duke@435: duke@435: // Def-use info will have some dead stuff; dead stuff will have no duke@435: // loop decided on. duke@435: duke@435: // Am I a loop header? If so fix up my parent's child and next ptrs. duke@435: if( innermost && innermost->_head == n ) { duke@435: assert( get_loop(n) == innermost, "" ); duke@435: IdealLoopTree *p = innermost->_parent; duke@435: IdealLoopTree *l = innermost; duke@435: while( p && l->_head == n ) { duke@435: l->_next = p->_child; // Put self on parents 'next child' duke@435: p->_child = l; // Make self as first child of parent duke@435: l = p; // Now walk up the parent chain duke@435: p = l->_parent; duke@435: } duke@435: } else { duke@435: // Note that it is possible for a LoopNode to reach here, if the duke@435: // backedge has been made unreachable (hence the LoopNode no longer duke@435: // denotes a Loop, and will eventually be removed). duke@435: duke@435: // Record tightest enclosing loop for self. Mark as post-visited. duke@435: set_loop(n, innermost); duke@435: // Also record has_call flag early on duke@435: if( innermost ) { duke@435: if( n->is_Call() && !n->is_CallLeaf() && !n->is_macro() ) { duke@435: // Do not count uncommon calls duke@435: if( !n->is_CallStaticJava() || !n->as_CallStaticJava()->_name ) { duke@435: Node *iff = n->in(0)->in(0); duke@435: if( !iff->is_If() || duke@435: (n->in(0)->Opcode() == Op_IfFalse && duke@435: (1.0 - iff->as_If()->_prob) >= 0.01) || duke@435: (iff->as_If()->_prob >= 0.01) ) duke@435: innermost->_has_call = 1; duke@435: } kvn@474: } else if( n->is_Allocate() && n->as_Allocate()->_is_scalar_replaceable ) { kvn@474: // Disable loop optimizations if the loop has a scalar replaceable kvn@474: // allocation. This disabling may cause a potential performance lost kvn@474: // if the allocation is not eliminated for some reason. kvn@474: innermost->_allow_optimizations = false; kvn@474: innermost->_has_call = 1; // = true duke@435: } duke@435: } duke@435: } duke@435: duke@435: // Flag as post-visited now duke@435: set_postvisited(n); duke@435: return pre_order; duke@435: } duke@435: duke@435: duke@435: //------------------------------build_loop_early------------------------------- duke@435: // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping. duke@435: // First pass computes the earliest controlling node possible. This is the duke@435: // controlling input with the deepest dominating depth. duke@435: void PhaseIdealLoop::build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack, const PhaseIdealLoop *verify_me ) { duke@435: while (worklist.size() != 0) { duke@435: // Use local variables nstack_top_n & nstack_top_i to cache values duke@435: // on nstack's top. duke@435: Node *nstack_top_n = worklist.pop(); duke@435: uint nstack_top_i = 0; duke@435: //while_nstack_nonempty: duke@435: while (true) { duke@435: // Get parent node and next input's index from stack's top. duke@435: Node *n = nstack_top_n; duke@435: uint i = nstack_top_i; duke@435: uint cnt = n->req(); // Count of inputs duke@435: if (i == 0) { // Pre-process the node. duke@435: if( has_node(n) && // Have either loop or control already? duke@435: !has_ctrl(n) ) { // Have loop picked out already? duke@435: // During "merge_many_backedges" we fold up several nested loops duke@435: // into a single loop. This makes the members of the original duke@435: // loop bodies pointing to dead loops; they need to move up duke@435: // to the new UNION'd larger loop. I set the _head field of these duke@435: // dead loops to NULL and the _parent field points to the owning duke@435: // loop. Shades of UNION-FIND algorithm. duke@435: IdealLoopTree *ilt; duke@435: while( !(ilt = get_loop(n))->_head ) { duke@435: // Normally I would use a set_loop here. But in this one special duke@435: // case, it is legal (and expected) to change what loop a Node duke@435: // belongs to. duke@435: _nodes.map(n->_idx, (Node*)(ilt->_parent) ); duke@435: } duke@435: // Remove safepoints ONLY if I've already seen I don't need one. duke@435: // (the old code here would yank a 2nd safepoint after seeing a duke@435: // first one, even though the 1st did not dominate in the loop body duke@435: // and thus could be avoided indefinitely) duke@435: if( !verify_me && ilt->_has_sfpt && n->Opcode() == Op_SafePoint && duke@435: is_deleteable_safept(n)) { duke@435: Node *in = n->in(TypeFunc::Control); duke@435: lazy_replace(n,in); // Pull safepoint now duke@435: // Carry on with the recursion "as if" we are walking duke@435: // only the control input duke@435: if( !visited.test_set( in->_idx ) ) { duke@435: worklist.push(in); // Visit this guy later, using worklist duke@435: } duke@435: // Get next node from nstack: duke@435: // - skip n's inputs processing by setting i > cnt; duke@435: // - we also will not call set_early_ctrl(n) since duke@435: // has_node(n) == true (see the condition above). duke@435: i = cnt + 1; duke@435: } duke@435: } duke@435: } // if (i == 0) duke@435: duke@435: // Visit all inputs duke@435: bool done = true; // Assume all n's inputs will be processed duke@435: while (i < cnt) { duke@435: Node *in = n->in(i); duke@435: ++i; duke@435: if (in == NULL) continue; duke@435: if (in->pinned() && !in->is_CFG()) duke@435: set_ctrl(in, in->in(0)); duke@435: int is_visited = visited.test_set( in->_idx ); duke@435: if (!has_node(in)) { // No controlling input yet? duke@435: assert( !in->is_CFG(), "CFG Node with no controlling input?" ); duke@435: assert( !is_visited, "visit only once" ); duke@435: nstack.push(n, i); // Save parent node and next input's index. duke@435: nstack_top_n = in; // Process current input now. duke@435: nstack_top_i = 0; duke@435: done = false; // Not all n's inputs processed. duke@435: break; // continue while_nstack_nonempty; duke@435: } else if (!is_visited) { duke@435: // This guy has a location picked out for him, but has not yet duke@435: // been visited. Happens to all CFG nodes, for instance. duke@435: // Visit him using the worklist instead of recursion, to break duke@435: // cycles. Since he has a location already we do not need to duke@435: // find his location before proceeding with the current Node. duke@435: worklist.push(in); // Visit this guy later, using worklist duke@435: } duke@435: } duke@435: if (done) { duke@435: // All of n's inputs have been processed, complete post-processing. duke@435: twisti@1040: // Compute earliest point this Node can go. duke@435: // CFG, Phi, pinned nodes already know their controlling input. duke@435: if (!has_node(n)) { duke@435: // Record earliest legal location duke@435: set_early_ctrl( n ); duke@435: } duke@435: if (nstack.is_empty()) { duke@435: // Finished all nodes on stack. duke@435: // Process next node on the worklist. duke@435: break; duke@435: } duke@435: // Get saved parent node and next input's index. duke@435: nstack_top_n = nstack.node(); duke@435: nstack_top_i = nstack.index(); duke@435: nstack.pop(); duke@435: } duke@435: } // while (true) duke@435: } duke@435: } duke@435: duke@435: //------------------------------dom_lca_internal-------------------------------- duke@435: // Pair-wise LCA duke@435: Node *PhaseIdealLoop::dom_lca_internal( Node *n1, Node *n2 ) const { duke@435: if( !n1 ) return n2; // Handle NULL original LCA duke@435: assert( n1->is_CFG(), "" ); duke@435: assert( n2->is_CFG(), "" ); duke@435: // find LCA of all uses duke@435: uint d1 = dom_depth(n1); duke@435: uint d2 = dom_depth(n2); duke@435: while (n1 != n2) { duke@435: if (d1 > d2) { duke@435: n1 = idom(n1); duke@435: d1 = dom_depth(n1); duke@435: } else if (d1 < d2) { duke@435: n2 = idom(n2); duke@435: d2 = dom_depth(n2); duke@435: } else { duke@435: // Here d1 == d2. Due to edits of the dominator-tree, sections duke@435: // of the tree might have the same depth. These sections have duke@435: // to be searched more carefully. duke@435: duke@435: // Scan up all the n1's with equal depth, looking for n2. duke@435: Node *t1 = idom(n1); duke@435: while (dom_depth(t1) == d1) { duke@435: if (t1 == n2) return n2; duke@435: t1 = idom(t1); duke@435: } duke@435: // Scan up all the n2's with equal depth, looking for n1. duke@435: Node *t2 = idom(n2); duke@435: while (dom_depth(t2) == d2) { duke@435: if (t2 == n1) return n1; duke@435: t2 = idom(t2); duke@435: } duke@435: // Move up to a new dominator-depth value as well as up the dom-tree. duke@435: n1 = t1; duke@435: n2 = t2; duke@435: d1 = dom_depth(n1); duke@435: d2 = dom_depth(n2); duke@435: } duke@435: } duke@435: return n1; duke@435: } duke@435: duke@435: //------------------------------compute_idom----------------------------------- duke@435: // Locally compute IDOM using dom_lca call. Correct only if the incoming duke@435: // IDOMs are correct. duke@435: Node *PhaseIdealLoop::compute_idom( Node *region ) const { duke@435: assert( region->is_Region(), "" ); duke@435: Node *LCA = NULL; duke@435: for( uint i = 1; i < region->req(); i++ ) { duke@435: if( region->in(i) != C->top() ) duke@435: LCA = dom_lca( LCA, region->in(i) ); duke@435: } duke@435: return LCA; duke@435: } duke@435: duke@435: //------------------------------get_late_ctrl---------------------------------- duke@435: // Compute latest legal control. duke@435: Node *PhaseIdealLoop::get_late_ctrl( Node *n, Node *early ) { duke@435: assert(early != NULL, "early control should not be NULL"); duke@435: duke@435: // Compute LCA over list of uses duke@435: Node *LCA = NULL; duke@435: for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && LCA != early; i++) { duke@435: Node* c = n->fast_out(i); duke@435: if (_nodes[c->_idx] == NULL) duke@435: continue; // Skip the occasional dead node duke@435: if( c->is_Phi() ) { // For Phis, we must land above on the path duke@435: for( uint j=1; jreq(); j++ ) {// For all inputs duke@435: if( c->in(j) == n ) { // Found matching input? duke@435: Node *use = c->in(0)->in(j); duke@435: LCA = dom_lca_for_get_late_ctrl( LCA, use, n ); duke@435: } duke@435: } duke@435: } else { duke@435: // For CFG data-users, use is in the block just prior duke@435: Node *use = has_ctrl(c) ? get_ctrl(c) : c->in(0); duke@435: LCA = dom_lca_for_get_late_ctrl( LCA, use, n ); duke@435: } duke@435: } duke@435: duke@435: // if this is a load, check for anti-dependent stores duke@435: // We use a conservative algorithm to identify potential interfering duke@435: // instructions and for rescheduling the load. The users of the memory duke@435: // input of this load are examined. Any use which is not a load and is duke@435: // dominated by early is considered a potentially interfering store. duke@435: // This can produce false positives. duke@435: if (n->is_Load() && LCA != early) { duke@435: Node_List worklist; duke@435: duke@435: Node *mem = n->in(MemNode::Memory); duke@435: for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) { duke@435: Node* s = mem->fast_out(i); duke@435: worklist.push(s); duke@435: } duke@435: while(worklist.size() != 0 && LCA != early) { duke@435: Node* s = worklist.pop(); duke@435: if (s->is_Load()) { duke@435: continue; duke@435: } else if (s->is_MergeMem()) { duke@435: for (DUIterator_Fast imax, i = s->fast_outs(imax); i < imax; i++) { duke@435: Node* s1 = s->fast_out(i); duke@435: worklist.push(s1); duke@435: } duke@435: } else { duke@435: Node *sctrl = has_ctrl(s) ? get_ctrl(s) : s->in(0); duke@435: assert(sctrl != NULL || s->outcnt() == 0, "must have control"); duke@435: if (sctrl != NULL && !sctrl->is_top() && is_dominator(early, sctrl)) { duke@435: LCA = dom_lca_for_get_late_ctrl(LCA, sctrl, n); duke@435: } duke@435: } duke@435: } duke@435: } duke@435: duke@435: assert(LCA == find_non_split_ctrl(LCA), "unexpected late control"); duke@435: return LCA; duke@435: } duke@435: duke@435: // true if CFG node d dominates CFG node n duke@435: bool PhaseIdealLoop::is_dominator(Node *d, Node *n) { duke@435: if (d == n) duke@435: return true; duke@435: assert(d->is_CFG() && n->is_CFG(), "must have CFG nodes"); duke@435: uint dd = dom_depth(d); duke@435: while (dom_depth(n) >= dd) { duke@435: if (n == d) duke@435: return true; duke@435: n = idom(n); duke@435: } duke@435: return false; duke@435: } duke@435: duke@435: //------------------------------dom_lca_for_get_late_ctrl_internal------------- duke@435: // Pair-wise LCA with tags. duke@435: // Tag each index with the node 'tag' currently being processed duke@435: // before advancing up the dominator chain using idom(). duke@435: // Later calls that find a match to 'tag' know that this path has already duke@435: // been considered in the current LCA (which is input 'n1' by convention). duke@435: // Since get_late_ctrl() is only called once for each node, the tag array duke@435: // does not need to be cleared between calls to get_late_ctrl(). duke@435: // Algorithm trades a larger constant factor for better asymptotic behavior duke@435: // duke@435: Node *PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal( Node *n1, Node *n2, Node *tag ) { duke@435: uint d1 = dom_depth(n1); duke@435: uint d2 = dom_depth(n2); duke@435: duke@435: do { duke@435: if (d1 > d2) { duke@435: // current lca is deeper than n2 duke@435: _dom_lca_tags.map(n1->_idx, tag); duke@435: n1 = idom(n1); duke@435: d1 = dom_depth(n1); duke@435: } else if (d1 < d2) { duke@435: // n2 is deeper than current lca duke@435: Node *memo = _dom_lca_tags[n2->_idx]; duke@435: if( memo == tag ) { duke@435: return n1; // Return the current LCA duke@435: } duke@435: _dom_lca_tags.map(n2->_idx, tag); duke@435: n2 = idom(n2); duke@435: d2 = dom_depth(n2); duke@435: } else { duke@435: // Here d1 == d2. Due to edits of the dominator-tree, sections duke@435: // of the tree might have the same depth. These sections have duke@435: // to be searched more carefully. duke@435: duke@435: // Scan up all the n1's with equal depth, looking for n2. duke@435: _dom_lca_tags.map(n1->_idx, tag); duke@435: Node *t1 = idom(n1); duke@435: while (dom_depth(t1) == d1) { duke@435: if (t1 == n2) return n2; duke@435: _dom_lca_tags.map(t1->_idx, tag); duke@435: t1 = idom(t1); duke@435: } duke@435: // Scan up all the n2's with equal depth, looking for n1. duke@435: _dom_lca_tags.map(n2->_idx, tag); duke@435: Node *t2 = idom(n2); duke@435: while (dom_depth(t2) == d2) { duke@435: if (t2 == n1) return n1; duke@435: _dom_lca_tags.map(t2->_idx, tag); duke@435: t2 = idom(t2); duke@435: } duke@435: // Move up to a new dominator-depth value as well as up the dom-tree. duke@435: n1 = t1; duke@435: n2 = t2; duke@435: d1 = dom_depth(n1); duke@435: d2 = dom_depth(n2); duke@435: } duke@435: } while (n1 != n2); duke@435: return n1; duke@435: } duke@435: duke@435: //------------------------------init_dom_lca_tags------------------------------ duke@435: // Tag could be a node's integer index, 32bits instead of 64bits in some cases duke@435: // Intended use does not involve any growth for the array, so it could duke@435: // be of fixed size. duke@435: void PhaseIdealLoop::init_dom_lca_tags() { duke@435: uint limit = C->unique() + 1; duke@435: _dom_lca_tags.map( limit, NULL ); duke@435: #ifdef ASSERT duke@435: for( uint i = 0; i < limit; ++i ) { duke@435: assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer"); duke@435: } duke@435: #endif // ASSERT duke@435: } duke@435: duke@435: //------------------------------clear_dom_lca_tags------------------------------ duke@435: // Tag could be a node's integer index, 32bits instead of 64bits in some cases duke@435: // Intended use does not involve any growth for the array, so it could duke@435: // be of fixed size. duke@435: void PhaseIdealLoop::clear_dom_lca_tags() { duke@435: uint limit = C->unique() + 1; duke@435: _dom_lca_tags.map( limit, NULL ); duke@435: _dom_lca_tags.clear(); duke@435: #ifdef ASSERT duke@435: for( uint i = 0; i < limit; ++i ) { duke@435: assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer"); duke@435: } duke@435: #endif // ASSERT duke@435: } duke@435: duke@435: //------------------------------build_loop_late-------------------------------- duke@435: // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping. duke@435: // Second pass finds latest legal placement, and ideal loop placement. duke@435: void PhaseIdealLoop::build_loop_late( VectorSet &visited, Node_List &worklist, Node_Stack &nstack, const PhaseIdealLoop *verify_me ) { duke@435: while (worklist.size() != 0) { duke@435: Node *n = worklist.pop(); duke@435: // Only visit once duke@435: if (visited.test_set(n->_idx)) continue; duke@435: uint cnt = n->outcnt(); duke@435: uint i = 0; duke@435: while (true) { duke@435: assert( _nodes[n->_idx], "no dead nodes" ); duke@435: // Visit all children duke@435: if (i < cnt) { duke@435: Node* use = n->raw_out(i); duke@435: ++i; duke@435: // Check for dead uses. Aggressively prune such junk. It might be duke@435: // dead in the global sense, but still have local uses so I cannot duke@435: // easily call 'remove_dead_node'. duke@435: if( _nodes[use->_idx] != NULL || use->is_top() ) { // Not dead? duke@435: // Due to cycles, we might not hit the same fixed point in the verify duke@435: // pass as we do in the regular pass. Instead, visit such phis as duke@435: // simple uses of the loop head. duke@435: if( use->in(0) && (use->is_CFG() || use->is_Phi()) ) { duke@435: if( !visited.test(use->_idx) ) duke@435: worklist.push(use); duke@435: } else if( !visited.test_set(use->_idx) ) { duke@435: nstack.push(n, i); // Save parent and next use's index. duke@435: n = use; // Process all children of current use. duke@435: cnt = use->outcnt(); duke@435: i = 0; duke@435: } duke@435: } else { duke@435: // Do not visit around the backedge of loops via data edges. duke@435: // push dead code onto a worklist duke@435: _deadlist.push(use); duke@435: } duke@435: } else { duke@435: // All of n's children have been processed, complete post-processing. duke@435: build_loop_late_post(n, verify_me); duke@435: if (nstack.is_empty()) { duke@435: // Finished all nodes on stack. duke@435: // Process next node on the worklist. duke@435: break; duke@435: } duke@435: // Get saved parent node and next use's index. Visit the rest of uses. duke@435: n = nstack.node(); duke@435: cnt = n->outcnt(); duke@435: i = nstack.index(); duke@435: nstack.pop(); duke@435: } duke@435: } duke@435: } duke@435: } duke@435: duke@435: //------------------------------build_loop_late_post--------------------------- duke@435: // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping. duke@435: // Second pass finds latest legal placement, and ideal loop placement. duke@435: void PhaseIdealLoop::build_loop_late_post( Node *n, const PhaseIdealLoop *verify_me ) { duke@435: duke@435: if (n->req() == 2 && n->Opcode() == Op_ConvI2L && !C->major_progress()) { duke@435: _igvn._worklist.push(n); // Maybe we'll normalize it, if no more loops. duke@435: } duke@435: duke@435: // CFG and pinned nodes already handled duke@435: if( n->in(0) ) { duke@435: if( n->in(0)->is_top() ) return; // Dead? duke@435: duke@435: // We'd like +VerifyLoopOptimizations to not believe that Mod's/Loads duke@435: // _must_ be pinned (they have to observe their control edge of course). duke@435: // Unlike Stores (which modify an unallocable resource, the memory duke@435: // state), Mods/Loads can float around. So free them up. duke@435: bool pinned = true; duke@435: switch( n->Opcode() ) { duke@435: case Op_DivI: duke@435: case Op_DivF: duke@435: case Op_DivD: duke@435: case Op_ModI: duke@435: case Op_ModF: duke@435: case Op_ModD: duke@435: case Op_LoadB: // Same with Loads; they can sink twisti@993: case Op_LoadUS: // during loop optimizations. duke@435: case Op_LoadD: duke@435: case Op_LoadF: duke@435: case Op_LoadI: duke@435: case Op_LoadKlass: kvn@728: case Op_LoadNKlass: duke@435: case Op_LoadL: duke@435: case Op_LoadS: duke@435: case Op_LoadP: kvn@728: case Op_LoadN: duke@435: case Op_LoadRange: duke@435: case Op_LoadD_unaligned: duke@435: case Op_LoadL_unaligned: duke@435: case Op_StrComp: // Does a bunch of load-like effects rasbold@604: case Op_AryEq: duke@435: pinned = false; duke@435: } duke@435: if( pinned ) { twisti@1040: IdealLoopTree *chosen_loop = get_loop(n->is_CFG() ? n : get_ctrl(n)); twisti@1040: if( !chosen_loop->_child ) // Inner loop? twisti@1040: chosen_loop->_body.push(n); // Collect inner loops duke@435: return; duke@435: } duke@435: } else { // No slot zero duke@435: if( n->is_CFG() ) { // CFG with no slot 0 is dead duke@435: _nodes.map(n->_idx,0); // No block setting, it's globally dead duke@435: return; duke@435: } duke@435: assert(!n->is_CFG() || n->outcnt() == 0, ""); duke@435: } duke@435: duke@435: // Do I have a "safe range" I can select over? duke@435: Node *early = get_ctrl(n);// Early location already computed duke@435: duke@435: // Compute latest point this Node can go duke@435: Node *LCA = get_late_ctrl( n, early ); duke@435: // LCA is NULL due to uses being dead duke@435: if( LCA == NULL ) { duke@435: #ifdef ASSERT duke@435: for (DUIterator i1 = n->outs(); n->has_out(i1); i1++) { duke@435: assert( _nodes[n->out(i1)->_idx] == NULL, "all uses must also be dead"); duke@435: } duke@435: #endif duke@435: _nodes.map(n->_idx, 0); // This node is useless duke@435: _deadlist.push(n); duke@435: return; duke@435: } duke@435: assert(LCA != NULL && !LCA->is_top(), "no dead nodes"); duke@435: duke@435: Node *legal = LCA; // Walk 'legal' up the IDOM chain duke@435: Node *least = legal; // Best legal position so far duke@435: while( early != legal ) { // While not at earliest legal duke@435: // Find least loop nesting depth duke@435: legal = idom(legal); // Bump up the IDOM tree duke@435: // Check for lower nesting depth duke@435: if( get_loop(legal)->_nest < get_loop(least)->_nest ) duke@435: least = legal; duke@435: } duke@435: duke@435: // Try not to place code on a loop entry projection duke@435: // which can inhibit range check elimination. duke@435: if (least != early) { duke@435: Node* ctrl_out = least->unique_ctrl_out(); duke@435: if (ctrl_out && ctrl_out->is_CountedLoop() && duke@435: least == ctrl_out->in(LoopNode::EntryControl)) { duke@435: Node* least_dom = idom(least); duke@435: if (get_loop(least_dom)->is_member(get_loop(least))) { duke@435: least = least_dom; duke@435: } duke@435: } duke@435: } duke@435: duke@435: #ifdef ASSERT duke@435: // If verifying, verify that 'verify_me' has a legal location duke@435: // and choose it as our location. duke@435: if( verify_me ) { duke@435: Node *v_ctrl = verify_me->get_ctrl_no_update(n); duke@435: Node *legal = LCA; duke@435: while( early != legal ) { // While not at earliest legal duke@435: if( legal == v_ctrl ) break; // Check for prior good location duke@435: legal = idom(legal) ;// Bump up the IDOM tree duke@435: } duke@435: // Check for prior good location duke@435: if( legal == v_ctrl ) least = legal; // Keep prior if found duke@435: } duke@435: #endif duke@435: duke@435: // Assign discovered "here or above" point duke@435: least = find_non_split_ctrl(least); duke@435: set_ctrl(n, least); duke@435: duke@435: // Collect inner loop bodies twisti@1040: IdealLoopTree *chosen_loop = get_loop(least); twisti@1040: if( !chosen_loop->_child ) // Inner loop? twisti@1040: chosen_loop->_body.push(n);// Collect inner loops duke@435: } duke@435: duke@435: #ifndef PRODUCT duke@435: //------------------------------dump------------------------------------------- duke@435: void PhaseIdealLoop::dump( ) const { duke@435: ResourceMark rm; duke@435: Arena* arena = Thread::current()->resource_area(); duke@435: Node_Stack stack(arena, C->unique() >> 2); duke@435: Node_List rpo_list; duke@435: VectorSet visited(arena); duke@435: visited.set(C->top()->_idx); duke@435: rpo( C->root(), stack, visited, rpo_list ); duke@435: // Dump root loop indexed by last element in PO order duke@435: dump( _ltree_root, rpo_list.size(), rpo_list ); duke@435: } duke@435: duke@435: void PhaseIdealLoop::dump( IdealLoopTree *loop, uint idx, Node_List &rpo_list ) const { never@802: loop->dump_head(); duke@435: duke@435: // Now scan for CFG nodes in the same loop duke@435: for( uint j=idx; j > 0; j-- ) { duke@435: Node *n = rpo_list[j-1]; duke@435: if( !_nodes[n->_idx] ) // Skip dead nodes duke@435: continue; duke@435: if( get_loop(n) != loop ) { // Wrong loop nest duke@435: if( get_loop(n)->_head == n && // Found nested loop? duke@435: get_loop(n)->_parent == loop ) duke@435: dump(get_loop(n),rpo_list.size(),rpo_list); // Print it nested-ly duke@435: continue; duke@435: } duke@435: duke@435: // Dump controlling node duke@435: for( uint x = 0; x < loop->_nest; x++ ) duke@435: tty->print(" "); duke@435: tty->print("C"); duke@435: if( n == C->root() ) { duke@435: n->dump(); duke@435: } else { duke@435: Node* cached_idom = idom_no_update(n); duke@435: Node *computed_idom = n->in(0); duke@435: if( n->is_Region() ) { duke@435: computed_idom = compute_idom(n); duke@435: // computed_idom() will return n->in(0) when idom(n) is an IfNode (or duke@435: // any MultiBranch ctrl node), so apply a similar transform to duke@435: // the cached idom returned from idom_no_update. duke@435: cached_idom = find_non_split_ctrl(cached_idom); duke@435: } duke@435: tty->print(" ID:%d",computed_idom->_idx); duke@435: n->dump(); duke@435: if( cached_idom != computed_idom ) { duke@435: tty->print_cr("*** BROKEN IDOM! Computed as: %d, cached as: %d", duke@435: computed_idom->_idx, cached_idom->_idx); duke@435: } duke@435: } duke@435: // Dump nodes it controls duke@435: for( uint k = 0; k < _nodes.Size(); k++ ) { duke@435: // (k < C->unique() && get_ctrl(find(k)) == n) duke@435: if (k < C->unique() && _nodes[k] == (Node*)((intptr_t)n + 1)) { duke@435: Node *m = C->root()->find(k); duke@435: if( m && m->outcnt() > 0 ) { duke@435: if (!(has_ctrl(m) && get_ctrl_no_update(m) == n)) { duke@435: tty->print_cr("*** BROKEN CTRL ACCESSOR! _nodes[k] is %p, ctrl is %p", duke@435: _nodes[k], has_ctrl(m) ? get_ctrl_no_update(m) : NULL); duke@435: } duke@435: for( uint j = 0; j < loop->_nest; j++ ) duke@435: tty->print(" "); duke@435: tty->print(" "); duke@435: m->dump(); duke@435: } duke@435: } duke@435: } duke@435: } duke@435: } duke@435: duke@435: // Collect a R-P-O for the whole CFG. duke@435: // Result list is in post-order (scan backwards for RPO) duke@435: void PhaseIdealLoop::rpo( Node *start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list ) const { duke@435: stk.push(start, 0); duke@435: visited.set(start->_idx); duke@435: duke@435: while (stk.is_nonempty()) { duke@435: Node* m = stk.node(); duke@435: uint idx = stk.index(); duke@435: if (idx < m->outcnt()) { duke@435: stk.set_index(idx + 1); duke@435: Node* n = m->raw_out(idx); duke@435: if (n->is_CFG() && !visited.test_set(n->_idx)) { duke@435: stk.push(n, 0); duke@435: } duke@435: } else { duke@435: rpo_list.push(m); duke@435: stk.pop(); duke@435: } duke@435: } duke@435: } duke@435: #endif duke@435: duke@435: duke@435: //============================================================================= duke@435: //------------------------------LoopTreeIterator----------------------------------- duke@435: duke@435: // Advance to next loop tree using a preorder, left-to-right traversal. duke@435: void LoopTreeIterator::next() { duke@435: assert(!done(), "must not be done."); duke@435: if (_curnt->_child != NULL) { duke@435: _curnt = _curnt->_child; duke@435: } else if (_curnt->_next != NULL) { duke@435: _curnt = _curnt->_next; duke@435: } else { duke@435: while (_curnt != _root && _curnt->_next == NULL) { duke@435: _curnt = _curnt->_parent; duke@435: } duke@435: if (_curnt == _root) { duke@435: _curnt = NULL; duke@435: assert(done(), "must be done."); duke@435: } else { duke@435: assert(_curnt->_next != NULL, "must be more to do"); duke@435: _curnt = _curnt->_next; duke@435: } duke@435: } duke@435: }