duke@435: /* trims@1907: * Copyright (c) 2000, 2010, Oracle and/or its affiliates. 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: * trims@1907: * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA trims@1907: * or visit www.oracle.com if you need additional information or have any trims@1907: * questions. duke@435: * duke@435: */ duke@435: duke@435: #include "incls/_precompiled.incl" duke@435: #include "incls/_loopTransform.cpp.incl" duke@435: duke@435: //------------------------------is_loop_exit----------------------------------- duke@435: // Given an IfNode, return the loop-exiting projection or NULL if both duke@435: // arms remain in the loop. duke@435: Node *IdealLoopTree::is_loop_exit(Node *iff) const { duke@435: if( iff->outcnt() != 2 ) return NULL; // Ignore partially dead tests duke@435: PhaseIdealLoop *phase = _phase; duke@435: // Test is an IfNode, has 2 projections. If BOTH are in the loop duke@435: // we need loop unswitching instead of peeling. duke@435: if( !is_member(phase->get_loop( iff->raw_out(0) )) ) duke@435: return iff->raw_out(0); duke@435: if( !is_member(phase->get_loop( iff->raw_out(1) )) ) duke@435: return iff->raw_out(1); duke@435: return NULL; duke@435: } duke@435: duke@435: duke@435: //============================================================================= duke@435: duke@435: duke@435: //------------------------------record_for_igvn---------------------------- duke@435: // Put loop body on igvn work list duke@435: void IdealLoopTree::record_for_igvn() { duke@435: for( uint i = 0; i < _body.size(); i++ ) { duke@435: Node *n = _body.at(i); duke@435: _phase->_igvn._worklist.push(n); duke@435: } duke@435: } duke@435: duke@435: //------------------------------compute_profile_trip_cnt---------------------------- duke@435: // Compute loop trip count from profile data as duke@435: // (backedge_count + loop_exit_count) / loop_exit_count duke@435: void IdealLoopTree::compute_profile_trip_cnt( PhaseIdealLoop *phase ) { duke@435: if (!_head->is_CountedLoop()) { duke@435: return; duke@435: } duke@435: CountedLoopNode* head = _head->as_CountedLoop(); duke@435: if (head->profile_trip_cnt() != COUNT_UNKNOWN) { duke@435: return; // Already computed duke@435: } duke@435: float trip_cnt = (float)max_jint; // default is big duke@435: duke@435: Node* back = head->in(LoopNode::LoopBackControl); duke@435: while (back != head) { duke@435: if ((back->Opcode() == Op_IfTrue || back->Opcode() == Op_IfFalse) && duke@435: back->in(0) && duke@435: back->in(0)->is_If() && duke@435: back->in(0)->as_If()->_fcnt != COUNT_UNKNOWN && duke@435: back->in(0)->as_If()->_prob != PROB_UNKNOWN) { duke@435: break; duke@435: } duke@435: back = phase->idom(back); duke@435: } duke@435: if (back != head) { duke@435: assert((back->Opcode() == Op_IfTrue || back->Opcode() == Op_IfFalse) && duke@435: back->in(0), "if-projection exists"); duke@435: IfNode* back_if = back->in(0)->as_If(); duke@435: float loop_back_cnt = back_if->_fcnt * back_if->_prob; duke@435: duke@435: // Now compute a loop exit count duke@435: float loop_exit_cnt = 0.0f; duke@435: for( uint i = 0; i < _body.size(); i++ ) { duke@435: Node *n = _body[i]; duke@435: if( n->is_If() ) { duke@435: IfNode *iff = n->as_If(); duke@435: if( iff->_fcnt != COUNT_UNKNOWN && iff->_prob != PROB_UNKNOWN ) { duke@435: Node *exit = is_loop_exit(iff); duke@435: if( exit ) { duke@435: float exit_prob = iff->_prob; duke@435: if (exit->Opcode() == Op_IfFalse) exit_prob = 1.0 - exit_prob; duke@435: if (exit_prob > PROB_MIN) { duke@435: float exit_cnt = iff->_fcnt * exit_prob; duke@435: loop_exit_cnt += exit_cnt; duke@435: } duke@435: } duke@435: } duke@435: } duke@435: } duke@435: if (loop_exit_cnt > 0.0f) { duke@435: trip_cnt = (loop_back_cnt + loop_exit_cnt) / loop_exit_cnt; duke@435: } else { duke@435: // No exit count so use duke@435: trip_cnt = loop_back_cnt; duke@435: } duke@435: } duke@435: #ifndef PRODUCT duke@435: if (TraceProfileTripCount) { duke@435: tty->print_cr("compute_profile_trip_cnt lp: %d cnt: %f\n", head->_idx, trip_cnt); duke@435: } duke@435: #endif duke@435: head->set_profile_trip_cnt(trip_cnt); duke@435: } duke@435: duke@435: //---------------------is_invariant_addition----------------------------- duke@435: // Return nonzero index of invariant operand for an Add or Sub twisti@1040: // of (nonconstant) invariant and variant values. Helper for reassociate_invariants. duke@435: int IdealLoopTree::is_invariant_addition(Node* n, PhaseIdealLoop *phase) { duke@435: int op = n->Opcode(); duke@435: if (op == Op_AddI || op == Op_SubI) { duke@435: bool in1_invar = this->is_invariant(n->in(1)); duke@435: bool in2_invar = this->is_invariant(n->in(2)); duke@435: if (in1_invar && !in2_invar) return 1; duke@435: if (!in1_invar && in2_invar) return 2; duke@435: } duke@435: return 0; duke@435: } duke@435: duke@435: //---------------------reassociate_add_sub----------------------------- duke@435: // Reassociate invariant add and subtract expressions: duke@435: // duke@435: // inv1 + (x + inv2) => ( inv1 + inv2) + x duke@435: // (x + inv2) + inv1 => ( inv1 + inv2) + x duke@435: // inv1 + (x - inv2) => ( inv1 - inv2) + x duke@435: // inv1 - (inv2 - x) => ( inv1 - inv2) + x duke@435: // (x + inv2) - inv1 => (-inv1 + inv2) + x duke@435: // (x - inv2) + inv1 => ( inv1 - inv2) + x duke@435: // (x - inv2) - inv1 => (-inv1 - inv2) + x duke@435: // inv1 + (inv2 - x) => ( inv1 + inv2) - x duke@435: // inv1 - (x - inv2) => ( inv1 + inv2) - x duke@435: // (inv2 - x) + inv1 => ( inv1 + inv2) - x duke@435: // (inv2 - x) - inv1 => (-inv1 + inv2) - x duke@435: // inv1 - (x + inv2) => ( inv1 - inv2) - x duke@435: // duke@435: Node* IdealLoopTree::reassociate_add_sub(Node* n1, PhaseIdealLoop *phase) { duke@435: if (!n1->is_Add() && !n1->is_Sub() || n1->outcnt() == 0) return NULL; duke@435: if (is_invariant(n1)) return NULL; duke@435: int inv1_idx = is_invariant_addition(n1, phase); duke@435: if (!inv1_idx) return NULL; duke@435: // Don't mess with add of constant (igvn moves them to expression tree root.) duke@435: if (n1->is_Add() && n1->in(2)->is_Con()) return NULL; duke@435: Node* inv1 = n1->in(inv1_idx); duke@435: Node* n2 = n1->in(3 - inv1_idx); duke@435: int inv2_idx = is_invariant_addition(n2, phase); duke@435: if (!inv2_idx) return NULL; duke@435: Node* x = n2->in(3 - inv2_idx); duke@435: Node* inv2 = n2->in(inv2_idx); duke@435: duke@435: bool neg_x = n2->is_Sub() && inv2_idx == 1; duke@435: bool neg_inv2 = n2->is_Sub() && inv2_idx == 2; duke@435: bool neg_inv1 = n1->is_Sub() && inv1_idx == 2; duke@435: if (n1->is_Sub() && inv1_idx == 1) { duke@435: neg_x = !neg_x; duke@435: neg_inv2 = !neg_inv2; duke@435: } duke@435: Node* inv1_c = phase->get_ctrl(inv1); duke@435: Node* inv2_c = phase->get_ctrl(inv2); duke@435: Node* n_inv1; duke@435: if (neg_inv1) { duke@435: Node *zero = phase->_igvn.intcon(0); duke@435: phase->set_ctrl(zero, phase->C->root()); duke@435: n_inv1 = new (phase->C, 3) SubINode(zero, inv1); duke@435: phase->register_new_node(n_inv1, inv1_c); duke@435: } else { duke@435: n_inv1 = inv1; duke@435: } duke@435: Node* inv; duke@435: if (neg_inv2) { duke@435: inv = new (phase->C, 3) SubINode(n_inv1, inv2); duke@435: } else { duke@435: inv = new (phase->C, 3) AddINode(n_inv1, inv2); duke@435: } duke@435: phase->register_new_node(inv, phase->get_early_ctrl(inv)); duke@435: duke@435: Node* addx; duke@435: if (neg_x) { duke@435: addx = new (phase->C, 3) SubINode(inv, x); duke@435: } else { duke@435: addx = new (phase->C, 3) AddINode(x, inv); duke@435: } duke@435: phase->register_new_node(addx, phase->get_ctrl(x)); kvn@1976: phase->_igvn.replace_node(n1, addx); duke@435: return addx; duke@435: } duke@435: duke@435: //---------------------reassociate_invariants----------------------------- duke@435: // Reassociate invariant expressions: duke@435: void IdealLoopTree::reassociate_invariants(PhaseIdealLoop *phase) { duke@435: for (int i = _body.size() - 1; i >= 0; i--) { duke@435: Node *n = _body.at(i); duke@435: for (int j = 0; j < 5; j++) { duke@435: Node* nn = reassociate_add_sub(n, phase); duke@435: if (nn == NULL) break; duke@435: n = nn; // again duke@435: }; duke@435: } duke@435: } duke@435: duke@435: //------------------------------policy_peeling--------------------------------- duke@435: // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can duke@435: // make some loop-invariant test (usually a null-check) happen before the loop. duke@435: bool IdealLoopTree::policy_peeling( PhaseIdealLoop *phase ) const { duke@435: Node *test = ((IdealLoopTree*)this)->tail(); duke@435: int body_size = ((IdealLoopTree*)this)->_body.size(); duke@435: int uniq = phase->C->unique(); duke@435: // Peeling does loop cloning which can result in O(N^2) node construction duke@435: if( body_size > 255 /* Prevent overflow for large body_size */ duke@435: || (body_size * body_size + uniq > MaxNodeLimit) ) { duke@435: return false; // too large to safely clone duke@435: } duke@435: while( test != _head ) { // Scan till run off top of loop duke@435: if( test->is_If() ) { // Test? duke@435: Node *ctrl = phase->get_ctrl(test->in(1)); duke@435: if (ctrl->is_top()) duke@435: return false; // Found dead test on live IF? No peeling! duke@435: // Standard IF only has one input value to check for loop invariance duke@435: assert( test->Opcode() == Op_If || test->Opcode() == Op_CountedLoopEnd, "Check this code when new subtype is added"); duke@435: // Condition is not a member of this loop? duke@435: if( !is_member(phase->get_loop(ctrl)) && duke@435: is_loop_exit(test) ) duke@435: return true; // Found reason to peel! duke@435: } duke@435: // Walk up dominators to loop _head looking for test which is duke@435: // executed on every path thru loop. duke@435: test = phase->idom(test); duke@435: } duke@435: return false; duke@435: } duke@435: duke@435: //------------------------------peeled_dom_test_elim--------------------------- duke@435: // If we got the effect of peeling, either by actually peeling or by making duke@435: // a pre-loop which must execute at least once, we can remove all duke@435: // loop-invariant dominated tests in the main body. duke@435: void PhaseIdealLoop::peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new ) { duke@435: bool progress = true; duke@435: while( progress ) { duke@435: progress = false; // Reset for next iteration duke@435: Node *prev = loop->_head->in(LoopNode::LoopBackControl);//loop->tail(); duke@435: Node *test = prev->in(0); duke@435: while( test != loop->_head ) { // Scan till run off top of loop duke@435: duke@435: int p_op = prev->Opcode(); duke@435: if( (p_op == Op_IfFalse || p_op == Op_IfTrue) && duke@435: test->is_If() && // Test? duke@435: !test->in(1)->is_Con() && // And not already obvious? duke@435: // Condition is not a member of this loop? duke@435: !loop->is_member(get_loop(get_ctrl(test->in(1))))){ duke@435: // Walk loop body looking for instances of this test duke@435: for( uint i = 0; i < loop->_body.size(); i++ ) { duke@435: Node *n = loop->_body.at(i); duke@435: if( n->is_If() && n->in(1) == test->in(1) /*&& n != loop->tail()->in(0)*/ ) { duke@435: // IfNode was dominated by version in peeled loop body duke@435: progress = true; duke@435: dominated_by( old_new[prev->_idx], n ); duke@435: } duke@435: } duke@435: } duke@435: prev = test; duke@435: test = idom(test); duke@435: } // End of scan tests in loop duke@435: duke@435: } // End of while( progress ) duke@435: } duke@435: duke@435: //------------------------------do_peeling------------------------------------- duke@435: // Peel the first iteration of the given loop. duke@435: // Step 1: Clone the loop body. The clone becomes the peeled iteration. duke@435: // The pre-loop illegally has 2 control users (old & new loops). duke@435: // Step 2: Make the old-loop fall-in edges point to the peeled iteration. duke@435: // Do this by making the old-loop fall-in edges act as if they came duke@435: // around the loopback from the prior iteration (follow the old-loop duke@435: // backedges) and then map to the new peeled iteration. This leaves duke@435: // the pre-loop with only 1 user (the new peeled iteration), but the duke@435: // peeled-loop backedge has 2 users. duke@435: // Step 3: Cut the backedge on the clone (so its not a loop) and remove the duke@435: // extra backedge user. duke@435: void PhaseIdealLoop::do_peeling( IdealLoopTree *loop, Node_List &old_new ) { duke@435: duke@435: C->set_major_progress(); duke@435: // Peeling a 'main' loop in a pre/main/post situation obfuscates the duke@435: // 'pre' loop from the main and the 'pre' can no longer have it's duke@435: // iterations adjusted. Therefore, we need to declare this loop as duke@435: // no longer a 'main' loop; it will need new pre and post loops before duke@435: // we can do further RCE. duke@435: Node *h = loop->_head; duke@435: if( h->is_CountedLoop() ) { duke@435: CountedLoopNode *cl = h->as_CountedLoop(); duke@435: assert(cl->trip_count() > 0, "peeling a fully unrolled loop"); duke@435: cl->set_trip_count(cl->trip_count() - 1); duke@435: if( cl->is_main_loop() ) { duke@435: cl->set_normal_loop(); duke@435: #ifndef PRODUCT duke@435: if( PrintOpto && VerifyLoopOptimizations ) { duke@435: tty->print("Peeling a 'main' loop; resetting to 'normal' "); duke@435: loop->dump_head(); duke@435: } duke@435: #endif duke@435: } duke@435: } duke@435: duke@435: // Step 1: Clone the loop body. The clone becomes the peeled iteration. duke@435: // The pre-loop illegally has 2 control users (old & new loops). duke@435: clone_loop( loop, old_new, dom_depth(loop->_head) ); duke@435: duke@435: duke@435: // Step 2: Make the old-loop fall-in edges point to the peeled iteration. duke@435: // Do this by making the old-loop fall-in edges act as if they came duke@435: // around the loopback from the prior iteration (follow the old-loop duke@435: // backedges) and then map to the new peeled iteration. This leaves duke@435: // the pre-loop with only 1 user (the new peeled iteration), but the duke@435: // peeled-loop backedge has 2 users. duke@435: for (DUIterator_Fast jmax, j = loop->_head->fast_outs(jmax); j < jmax; j++) { duke@435: Node* old = loop->_head->fast_out(j); duke@435: if( old->in(0) == loop->_head && old->req() == 3 && duke@435: (old->is_Loop() || old->is_Phi()) ) { duke@435: Node *new_exit_value = old_new[old->in(LoopNode::LoopBackControl)->_idx]; duke@435: if( !new_exit_value ) // Backedge value is ALSO loop invariant? duke@435: // Then loop body backedge value remains the same. duke@435: new_exit_value = old->in(LoopNode::LoopBackControl); duke@435: _igvn.hash_delete(old); duke@435: old->set_req(LoopNode::EntryControl, new_exit_value); duke@435: } duke@435: } duke@435: duke@435: duke@435: // Step 3: Cut the backedge on the clone (so its not a loop) and remove the duke@435: // extra backedge user. duke@435: Node *nnn = old_new[loop->_head->_idx]; duke@435: _igvn.hash_delete(nnn); duke@435: nnn->set_req(LoopNode::LoopBackControl, C->top()); duke@435: for (DUIterator_Fast j2max, j2 = nnn->fast_outs(j2max); j2 < j2max; j2++) { duke@435: Node* use = nnn->fast_out(j2); duke@435: if( use->in(0) == nnn && use->req() == 3 && use->is_Phi() ) { duke@435: _igvn.hash_delete(use); duke@435: use->set_req(LoopNode::LoopBackControl, C->top()); duke@435: } duke@435: } duke@435: duke@435: duke@435: // Step 4: Correct dom-depth info. Set to loop-head depth. duke@435: int dd = dom_depth(loop->_head); duke@435: set_idom(loop->_head, loop->_head->in(1), dd); duke@435: for (uint j3 = 0; j3 < loop->_body.size(); j3++) { duke@435: Node *old = loop->_body.at(j3); duke@435: Node *nnn = old_new[old->_idx]; duke@435: if (!has_ctrl(nnn)) duke@435: set_idom(nnn, idom(nnn), dd-1); duke@435: // While we're at it, remove any SafePoints from the peeled code duke@435: if( old->Opcode() == Op_SafePoint ) { duke@435: Node *nnn = old_new[old->_idx]; duke@435: lazy_replace(nnn,nnn->in(TypeFunc::Control)); duke@435: } duke@435: } duke@435: duke@435: // Now force out all loop-invariant dominating tests. The optimizer duke@435: // finds some, but we _know_ they are all useless. duke@435: peeled_dom_test_elim(loop,old_new); duke@435: duke@435: loop->record_for_igvn(); duke@435: } duke@435: duke@435: //------------------------------policy_maximally_unroll------------------------ duke@435: // Return exact loop trip count, or 0 if not maximally unrolling duke@435: bool IdealLoopTree::policy_maximally_unroll( PhaseIdealLoop *phase ) const { duke@435: CountedLoopNode *cl = _head->as_CountedLoop(); duke@435: assert( cl->is_normal_loop(), "" ); duke@435: duke@435: Node *init_n = cl->init_trip(); duke@435: Node *limit_n = cl->limit(); duke@435: duke@435: // Non-constant bounds duke@435: if( init_n == NULL || !init_n->is_Con() || duke@435: limit_n == NULL || !limit_n->is_Con() || duke@435: // protect against stride not being a constant duke@435: !cl->stride_is_con() ) { duke@435: return false; duke@435: } duke@435: int init = init_n->get_int(); duke@435: int limit = limit_n->get_int(); duke@435: int span = limit - init; duke@435: int stride = cl->stride_con(); duke@435: duke@435: if (init >= limit || stride > span) { duke@435: // return a false (no maximally unroll) and the regular unroll/peel duke@435: // route will make a small mess which CCP will fold away. duke@435: return false; duke@435: } duke@435: uint trip_count = span/stride; // trip_count can be greater than 2 Gig. duke@435: assert( (int)trip_count*stride == span, "must divide evenly" ); duke@435: duke@435: // Real policy: if we maximally unroll, does it get too big? duke@435: // Allow the unrolled mess to get larger than standard loop duke@435: // size. After all, it will no longer be a loop. duke@435: uint body_size = _body.size(); duke@435: uint unroll_limit = (uint)LoopUnrollLimit * 4; duke@435: assert( (intx)unroll_limit == LoopUnrollLimit * 4, "LoopUnrollLimit must fit in 32bits"); duke@435: cl->set_trip_count(trip_count); duke@435: if( trip_count <= unroll_limit && body_size <= unroll_limit ) { duke@435: uint new_body_size = body_size * trip_count; duke@435: if (new_body_size <= unroll_limit && duke@435: body_size == new_body_size / trip_count && duke@435: // Unrolling can result in a large amount of node construction duke@435: new_body_size < MaxNodeLimit - phase->C->unique()) { duke@435: return true; // maximally unroll duke@435: } duke@435: } duke@435: duke@435: return false; // Do not maximally unroll duke@435: } duke@435: duke@435: duke@435: //------------------------------policy_unroll---------------------------------- duke@435: // Return TRUE or FALSE if the loop should be unrolled or not. Unroll if duke@435: // the loop is a CountedLoop and the body is small enough. duke@435: bool IdealLoopTree::policy_unroll( PhaseIdealLoop *phase ) const { duke@435: duke@435: CountedLoopNode *cl = _head->as_CountedLoop(); duke@435: assert( cl->is_normal_loop() || cl->is_main_loop(), "" ); duke@435: duke@435: // protect against stride not being a constant duke@435: if( !cl->stride_is_con() ) return false; duke@435: duke@435: // protect against over-unrolling duke@435: if( cl->trip_count() <= 1 ) return false; duke@435: duke@435: int future_unroll_ct = cl->unrolled_count() * 2; duke@435: duke@435: // Don't unroll if the next round of unrolling would push us duke@435: // over the expected trip count of the loop. One is subtracted duke@435: // from the expected trip count because the pre-loop normally duke@435: // executes 1 iteration. duke@435: if (UnrollLimitForProfileCheck > 0 && duke@435: cl->profile_trip_cnt() != COUNT_UNKNOWN && duke@435: future_unroll_ct > UnrollLimitForProfileCheck && duke@435: (float)future_unroll_ct > cl->profile_trip_cnt() - 1.0) { duke@435: return false; duke@435: } duke@435: duke@435: // When unroll count is greater than LoopUnrollMin, don't unroll if: duke@435: // the residual iterations are more than 10% of the trip count duke@435: // and rounds of "unroll,optimize" are not making significant progress duke@435: // Progress defined as current size less than 20% larger than previous size. duke@435: if (UseSuperWord && cl->node_count_before_unroll() > 0 && duke@435: future_unroll_ct > LoopUnrollMin && duke@435: (future_unroll_ct - 1) * 10.0 > cl->profile_trip_cnt() && duke@435: 1.2 * cl->node_count_before_unroll() < (double)_body.size()) { duke@435: return false; duke@435: } duke@435: duke@435: Node *init_n = cl->init_trip(); duke@435: Node *limit_n = cl->limit(); duke@435: // Non-constant bounds. duke@435: // Protect against over-unrolling when init or/and limit are not constant duke@435: // (so that trip_count's init value is maxint) but iv range is known. duke@435: if( init_n == NULL || !init_n->is_Con() || duke@435: limit_n == NULL || !limit_n->is_Con() ) { duke@435: Node* phi = cl->phi(); duke@435: if( phi != NULL ) { duke@435: assert(phi->is_Phi() && phi->in(0) == _head, "Counted loop should have iv phi."); duke@435: const TypeInt* iv_type = phase->_igvn.type(phi)->is_int(); duke@435: int next_stride = cl->stride_con() * 2; // stride after this unroll duke@435: if( next_stride > 0 ) { duke@435: if( iv_type->_lo + next_stride <= iv_type->_lo || // overflow duke@435: iv_type->_lo + next_stride > iv_type->_hi ) { duke@435: return false; // over-unrolling duke@435: } duke@435: } else if( next_stride < 0 ) { duke@435: if( iv_type->_hi + next_stride >= iv_type->_hi || // overflow duke@435: iv_type->_hi + next_stride < iv_type->_lo ) { duke@435: return false; // over-unrolling duke@435: } duke@435: } duke@435: } duke@435: } duke@435: duke@435: // Adjust body_size to determine if we unroll or not duke@435: uint body_size = _body.size(); duke@435: // Key test to unroll CaffeineMark's Logic test duke@435: int xors_in_loop = 0; duke@435: // Also count ModL, DivL and MulL which expand mightly duke@435: for( uint k = 0; k < _body.size(); k++ ) { duke@435: switch( _body.at(k)->Opcode() ) { duke@435: case Op_XorI: xors_in_loop++; break; // CaffeineMark's Logic test duke@435: case Op_ModL: body_size += 30; break; duke@435: case Op_DivL: body_size += 30; break; duke@435: case Op_MulL: body_size += 10; break; duke@435: } duke@435: } duke@435: duke@435: // Check for being too big duke@435: if( body_size > (uint)LoopUnrollLimit ) { duke@435: if( xors_in_loop >= 4 && body_size < (uint)LoopUnrollLimit*4) return true; duke@435: // Normal case: loop too big duke@435: return false; duke@435: } duke@435: duke@435: // Check for stride being a small enough constant duke@435: if( abs(cl->stride_con()) > (1<<3) ) return false; duke@435: duke@435: // Unroll once! (Each trip will soon do double iterations) duke@435: return true; duke@435: } duke@435: duke@435: //------------------------------policy_align----------------------------------- duke@435: // Return TRUE or FALSE if the loop should be cache-line aligned. Gather the duke@435: // expression that does the alignment. Note that only one array base can be twisti@1040: // aligned in a loop (unless the VM guarantees mutual alignment). Note that duke@435: // if we vectorize short memory ops into longer memory ops, we may want to duke@435: // increase alignment. duke@435: bool IdealLoopTree::policy_align( PhaseIdealLoop *phase ) const { duke@435: return false; duke@435: } duke@435: duke@435: //------------------------------policy_range_check----------------------------- duke@435: // Return TRUE or FALSE if the loop should be range-check-eliminated. duke@435: // Actually we do iteration-splitting, a more powerful form of RCE. duke@435: bool IdealLoopTree::policy_range_check( PhaseIdealLoop *phase ) const { duke@435: if( !RangeCheckElimination ) return false; duke@435: duke@435: CountedLoopNode *cl = _head->as_CountedLoop(); duke@435: // If we unrolled with no intention of doing RCE and we later duke@435: // changed our minds, we got no pre-loop. Either we need to duke@435: // make a new pre-loop, or we gotta disallow RCE. duke@435: if( cl->is_main_no_pre_loop() ) return false; // Disallowed for now. duke@435: Node *trip_counter = cl->phi(); duke@435: duke@435: // Check loop body for tests of trip-counter plus loop-invariant vs duke@435: // loop-invariant. duke@435: for( uint i = 0; i < _body.size(); i++ ) { duke@435: Node *iff = _body[i]; duke@435: if( iff->Opcode() == Op_If ) { // Test? duke@435: duke@435: // Comparing trip+off vs limit duke@435: Node *bol = iff->in(1); duke@435: if( bol->req() != 2 ) continue; // dead constant test cfang@1607: if (!bol->is_Bool()) { cfang@1607: assert(UseLoopPredicate && bol->Opcode() == Op_Conv2B, "predicate check only"); cfang@1607: continue; cfang@1607: } duke@435: Node *cmp = bol->in(1); duke@435: duke@435: Node *rc_exp = cmp->in(1); duke@435: Node *limit = cmp->in(2); duke@435: duke@435: Node *limit_c = phase->get_ctrl(limit); duke@435: if( limit_c == phase->C->top() ) duke@435: return false; // Found dead test on live IF? No RCE! duke@435: if( is_member(phase->get_loop(limit_c) ) ) { duke@435: // Compare might have operands swapped; commute them duke@435: rc_exp = cmp->in(2); duke@435: limit = cmp->in(1); duke@435: limit_c = phase->get_ctrl(limit); duke@435: if( is_member(phase->get_loop(limit_c) ) ) duke@435: continue; // Both inputs are loop varying; cannot RCE duke@435: } duke@435: duke@435: if (!phase->is_scaled_iv_plus_offset(rc_exp, trip_counter, NULL, NULL)) { duke@435: continue; duke@435: } duke@435: // Yeah! Found a test like 'trip+off vs limit' duke@435: // Test is an IfNode, has 2 projections. If BOTH are in the loop duke@435: // we need loop unswitching instead of iteration splitting. duke@435: if( is_loop_exit(iff) ) duke@435: return true; // Found reason to split iterations duke@435: } // End of is IF duke@435: } duke@435: duke@435: return false; duke@435: } duke@435: duke@435: //------------------------------policy_peel_only------------------------------- duke@435: // Return TRUE or FALSE if the loop should NEVER be RCE'd or aligned. Useful duke@435: // for unrolling loops with NO array accesses. duke@435: bool IdealLoopTree::policy_peel_only( PhaseIdealLoop *phase ) const { duke@435: duke@435: for( uint i = 0; i < _body.size(); i++ ) duke@435: if( _body[i]->is_Mem() ) duke@435: return false; duke@435: duke@435: // No memory accesses at all! duke@435: return true; duke@435: } duke@435: duke@435: //------------------------------clone_up_backedge_goo-------------------------- duke@435: // If Node n lives in the back_ctrl block and cannot float, we clone a private duke@435: // version of n in preheader_ctrl block and return that, otherwise return n. duke@435: Node *PhaseIdealLoop::clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n ) { duke@435: if( get_ctrl(n) != back_ctrl ) return n; duke@435: duke@435: Node *x = NULL; // If required, a clone of 'n' duke@435: // Check for 'n' being pinned in the backedge. duke@435: if( n->in(0) && n->in(0) == back_ctrl ) { duke@435: x = n->clone(); // Clone a copy of 'n' to preheader duke@435: x->set_req( 0, preheader_ctrl ); // Fix x's control input to preheader duke@435: } duke@435: duke@435: // Recursive fixup any other input edges into x. duke@435: // If there are no changes we can just return 'n', otherwise duke@435: // we need to clone a private copy and change it. duke@435: for( uint i = 1; i < n->req(); i++ ) { duke@435: Node *g = clone_up_backedge_goo( back_ctrl, preheader_ctrl, n->in(i) ); duke@435: if( g != n->in(i) ) { duke@435: if( !x ) duke@435: x = n->clone(); duke@435: x->set_req(i, g); duke@435: } duke@435: } duke@435: if( x ) { // x can legally float to pre-header location duke@435: register_new_node( x, preheader_ctrl ); duke@435: return x; duke@435: } else { // raise n to cover LCA of uses duke@435: set_ctrl( n, find_non_split_ctrl(back_ctrl->in(0)) ); duke@435: } duke@435: return n; duke@435: } duke@435: duke@435: //------------------------------insert_pre_post_loops-------------------------- duke@435: // Insert pre and post loops. If peel_only is set, the pre-loop can not have duke@435: // more iterations added. It acts as a 'peel' only, no lower-bound RCE, no duke@435: // alignment. Useful to unroll loops that do no array accesses. duke@435: void PhaseIdealLoop::insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only ) { duke@435: duke@435: C->set_major_progress(); duke@435: duke@435: // Find common pieces of the loop being guarded with pre & post loops duke@435: CountedLoopNode *main_head = loop->_head->as_CountedLoop(); duke@435: assert( main_head->is_normal_loop(), "" ); duke@435: CountedLoopEndNode *main_end = main_head->loopexit(); duke@435: assert( main_end->outcnt() == 2, "1 true, 1 false path only" ); duke@435: uint dd_main_head = dom_depth(main_head); duke@435: uint max = main_head->outcnt(); duke@435: duke@435: Node *pre_header= main_head->in(LoopNode::EntryControl); duke@435: Node *init = main_head->init_trip(); duke@435: Node *incr = main_end ->incr(); duke@435: Node *limit = main_end ->limit(); duke@435: Node *stride = main_end ->stride(); duke@435: Node *cmp = main_end ->cmp_node(); duke@435: BoolTest::mask b_test = main_end->test_trip(); duke@435: duke@435: // Need only 1 user of 'bol' because I will be hacking the loop bounds. duke@435: Node *bol = main_end->in(CountedLoopEndNode::TestValue); duke@435: if( bol->outcnt() != 1 ) { duke@435: bol = bol->clone(); duke@435: register_new_node(bol,main_end->in(CountedLoopEndNode::TestControl)); duke@435: _igvn.hash_delete(main_end); duke@435: main_end->set_req(CountedLoopEndNode::TestValue, bol); duke@435: } duke@435: // Need only 1 user of 'cmp' because I will be hacking the loop bounds. duke@435: if( cmp->outcnt() != 1 ) { duke@435: cmp = cmp->clone(); duke@435: register_new_node(cmp,main_end->in(CountedLoopEndNode::TestControl)); duke@435: _igvn.hash_delete(bol); duke@435: bol->set_req(1, cmp); duke@435: } duke@435: duke@435: //------------------------------ duke@435: // Step A: Create Post-Loop. duke@435: Node* main_exit = main_end->proj_out(false); duke@435: assert( main_exit->Opcode() == Op_IfFalse, "" ); duke@435: int dd_main_exit = dom_depth(main_exit); duke@435: duke@435: // Step A1: Clone the loop body. The clone becomes the post-loop. The main duke@435: // loop pre-header illegally has 2 control users (old & new loops). duke@435: clone_loop( loop, old_new, dd_main_exit ); duke@435: assert( old_new[main_end ->_idx]->Opcode() == Op_CountedLoopEnd, "" ); duke@435: CountedLoopNode *post_head = old_new[main_head->_idx]->as_CountedLoop(); duke@435: post_head->set_post_loop(main_head); duke@435: kvn@835: // Reduce the post-loop trip count. kvn@835: CountedLoopEndNode* post_end = old_new[main_end ->_idx]->as_CountedLoopEnd(); kvn@835: post_end->_prob = PROB_FAIR; kvn@835: duke@435: // Build the main-loop normal exit. duke@435: IfFalseNode *new_main_exit = new (C, 1) IfFalseNode(main_end); duke@435: _igvn.register_new_node_with_optimizer( new_main_exit ); duke@435: set_idom(new_main_exit, main_end, dd_main_exit ); duke@435: set_loop(new_main_exit, loop->_parent); duke@435: duke@435: // Step A2: Build a zero-trip guard for the post-loop. After leaving the duke@435: // main-loop, the post-loop may not execute at all. We 'opaque' the incr duke@435: // (the main-loop trip-counter exit value) because we will be changing duke@435: // the exit value (via unrolling) so we cannot constant-fold away the zero duke@435: // trip guard until all unrolling is done. kvn@651: Node *zer_opaq = new (C, 2) Opaque1Node(C, incr); duke@435: Node *zer_cmp = new (C, 3) CmpINode( zer_opaq, limit ); duke@435: Node *zer_bol = new (C, 2) BoolNode( zer_cmp, b_test ); duke@435: register_new_node( zer_opaq, new_main_exit ); duke@435: register_new_node( zer_cmp , new_main_exit ); duke@435: register_new_node( zer_bol , new_main_exit ); duke@435: duke@435: // Build the IfNode duke@435: IfNode *zer_iff = new (C, 2) IfNode( new_main_exit, zer_bol, PROB_FAIR, COUNT_UNKNOWN ); duke@435: _igvn.register_new_node_with_optimizer( zer_iff ); duke@435: set_idom(zer_iff, new_main_exit, dd_main_exit); duke@435: set_loop(zer_iff, loop->_parent); duke@435: duke@435: // Plug in the false-path, taken if we need to skip post-loop duke@435: _igvn.hash_delete( main_exit ); duke@435: main_exit->set_req(0, zer_iff); duke@435: _igvn._worklist.push(main_exit); duke@435: set_idom(main_exit, zer_iff, dd_main_exit); duke@435: set_idom(main_exit->unique_out(), zer_iff, dd_main_exit); duke@435: // Make the true-path, must enter the post loop duke@435: Node *zer_taken = new (C, 1) IfTrueNode( zer_iff ); duke@435: _igvn.register_new_node_with_optimizer( zer_taken ); duke@435: set_idom(zer_taken, zer_iff, dd_main_exit); duke@435: set_loop(zer_taken, loop->_parent); duke@435: // Plug in the true path duke@435: _igvn.hash_delete( post_head ); duke@435: post_head->set_req(LoopNode::EntryControl, zer_taken); duke@435: set_idom(post_head, zer_taken, dd_main_exit); duke@435: duke@435: // Step A3: Make the fall-in values to the post-loop come from the duke@435: // fall-out values of the main-loop. duke@435: for (DUIterator_Fast imax, i = main_head->fast_outs(imax); i < imax; i++) { duke@435: Node* main_phi = main_head->fast_out(i); duke@435: if( main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() >0 ) { duke@435: Node *post_phi = old_new[main_phi->_idx]; duke@435: Node *fallmain = clone_up_backedge_goo(main_head->back_control(), duke@435: post_head->init_control(), duke@435: main_phi->in(LoopNode::LoopBackControl)); duke@435: _igvn.hash_delete(post_phi); duke@435: post_phi->set_req( LoopNode::EntryControl, fallmain ); duke@435: } duke@435: } duke@435: duke@435: // Update local caches for next stanza duke@435: main_exit = new_main_exit; duke@435: duke@435: duke@435: //------------------------------ duke@435: // Step B: Create Pre-Loop. duke@435: duke@435: // Step B1: Clone the loop body. The clone becomes the pre-loop. The main duke@435: // loop pre-header illegally has 2 control users (old & new loops). duke@435: clone_loop( loop, old_new, dd_main_head ); duke@435: CountedLoopNode* pre_head = old_new[main_head->_idx]->as_CountedLoop(); duke@435: CountedLoopEndNode* pre_end = old_new[main_end ->_idx]->as_CountedLoopEnd(); duke@435: pre_head->set_pre_loop(main_head); duke@435: Node *pre_incr = old_new[incr->_idx]; duke@435: kvn@835: // Reduce the pre-loop trip count. kvn@835: pre_end->_prob = PROB_FAIR; kvn@835: duke@435: // Find the pre-loop normal exit. duke@435: Node* pre_exit = pre_end->proj_out(false); duke@435: assert( pre_exit->Opcode() == Op_IfFalse, "" ); duke@435: IfFalseNode *new_pre_exit = new (C, 1) IfFalseNode(pre_end); duke@435: _igvn.register_new_node_with_optimizer( new_pre_exit ); duke@435: set_idom(new_pre_exit, pre_end, dd_main_head); duke@435: set_loop(new_pre_exit, loop->_parent); duke@435: duke@435: // Step B2: Build a zero-trip guard for the main-loop. After leaving the duke@435: // pre-loop, the main-loop may not execute at all. Later in life this duke@435: // zero-trip guard will become the minimum-trip guard when we unroll duke@435: // the main-loop. kvn@651: Node *min_opaq = new (C, 2) Opaque1Node(C, limit); duke@435: Node *min_cmp = new (C, 3) CmpINode( pre_incr, min_opaq ); duke@435: Node *min_bol = new (C, 2) BoolNode( min_cmp, b_test ); duke@435: register_new_node( min_opaq, new_pre_exit ); duke@435: register_new_node( min_cmp , new_pre_exit ); duke@435: register_new_node( min_bol , new_pre_exit ); duke@435: kvn@835: // Build the IfNode (assume the main-loop is executed always). kvn@835: IfNode *min_iff = new (C, 2) IfNode( new_pre_exit, min_bol, PROB_ALWAYS, COUNT_UNKNOWN ); duke@435: _igvn.register_new_node_with_optimizer( min_iff ); duke@435: set_idom(min_iff, new_pre_exit, dd_main_head); duke@435: set_loop(min_iff, loop->_parent); duke@435: duke@435: // Plug in the false-path, taken if we need to skip main-loop duke@435: _igvn.hash_delete( pre_exit ); duke@435: pre_exit->set_req(0, min_iff); duke@435: set_idom(pre_exit, min_iff, dd_main_head); duke@435: set_idom(pre_exit->unique_out(), min_iff, dd_main_head); duke@435: // Make the true-path, must enter the main loop duke@435: Node *min_taken = new (C, 1) IfTrueNode( min_iff ); duke@435: _igvn.register_new_node_with_optimizer( min_taken ); duke@435: set_idom(min_taken, min_iff, dd_main_head); duke@435: set_loop(min_taken, loop->_parent); duke@435: // Plug in the true path duke@435: _igvn.hash_delete( main_head ); duke@435: main_head->set_req(LoopNode::EntryControl, min_taken); duke@435: set_idom(main_head, min_taken, dd_main_head); duke@435: duke@435: // Step B3: Make the fall-in values to the main-loop come from the duke@435: // fall-out values of the pre-loop. duke@435: for (DUIterator_Fast i2max, i2 = main_head->fast_outs(i2max); i2 < i2max; i2++) { duke@435: Node* main_phi = main_head->fast_out(i2); duke@435: if( main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() > 0 ) { duke@435: Node *pre_phi = old_new[main_phi->_idx]; duke@435: Node *fallpre = clone_up_backedge_goo(pre_head->back_control(), duke@435: main_head->init_control(), duke@435: pre_phi->in(LoopNode::LoopBackControl)); duke@435: _igvn.hash_delete(main_phi); duke@435: main_phi->set_req( LoopNode::EntryControl, fallpre ); duke@435: } duke@435: } duke@435: duke@435: // Step B4: Shorten the pre-loop to run only 1 iteration (for now). duke@435: // RCE and alignment may change this later. duke@435: Node *cmp_end = pre_end->cmp_node(); duke@435: assert( cmp_end->in(2) == limit, "" ); duke@435: Node *pre_limit = new (C, 3) AddINode( init, stride ); duke@435: duke@435: // Save the original loop limit in this Opaque1 node for duke@435: // use by range check elimination. kvn@651: Node *pre_opaq = new (C, 3) Opaque1Node(C, pre_limit, limit); duke@435: duke@435: register_new_node( pre_limit, pre_head->in(0) ); duke@435: register_new_node( pre_opaq , pre_head->in(0) ); duke@435: duke@435: // Since no other users of pre-loop compare, I can hack limit directly duke@435: assert( cmp_end->outcnt() == 1, "no other users" ); duke@435: _igvn.hash_delete(cmp_end); duke@435: cmp_end->set_req(2, peel_only ? pre_limit : pre_opaq); duke@435: duke@435: // Special case for not-equal loop bounds: duke@435: // Change pre loop test, main loop test, and the duke@435: // main loop guard test to use lt or gt depending on stride duke@435: // direction: duke@435: // positive stride use < duke@435: // negative stride use > duke@435: duke@435: if (pre_end->in(CountedLoopEndNode::TestValue)->as_Bool()->_test._test == BoolTest::ne) { duke@435: duke@435: BoolTest::mask new_test = (main_end->stride_con() > 0) ? BoolTest::lt : BoolTest::gt; duke@435: // Modify pre loop end condition duke@435: Node* pre_bol = pre_end->in(CountedLoopEndNode::TestValue)->as_Bool(); duke@435: BoolNode* new_bol0 = new (C, 2) BoolNode(pre_bol->in(1), new_test); duke@435: register_new_node( new_bol0, pre_head->in(0) ); duke@435: _igvn.hash_delete(pre_end); duke@435: pre_end->set_req(CountedLoopEndNode::TestValue, new_bol0); duke@435: // Modify main loop guard condition duke@435: assert(min_iff->in(CountedLoopEndNode::TestValue) == min_bol, "guard okay"); duke@435: BoolNode* new_bol1 = new (C, 2) BoolNode(min_bol->in(1), new_test); duke@435: register_new_node( new_bol1, new_pre_exit ); duke@435: _igvn.hash_delete(min_iff); duke@435: min_iff->set_req(CountedLoopEndNode::TestValue, new_bol1); duke@435: // Modify main loop end condition duke@435: BoolNode* main_bol = main_end->in(CountedLoopEndNode::TestValue)->as_Bool(); duke@435: BoolNode* new_bol2 = new (C, 2) BoolNode(main_bol->in(1), new_test); duke@435: register_new_node( new_bol2, main_end->in(CountedLoopEndNode::TestControl) ); duke@435: _igvn.hash_delete(main_end); duke@435: main_end->set_req(CountedLoopEndNode::TestValue, new_bol2); duke@435: } duke@435: duke@435: // Flag main loop duke@435: main_head->set_main_loop(); duke@435: if( peel_only ) main_head->set_main_no_pre_loop(); duke@435: duke@435: // It's difficult to be precise about the trip-counts duke@435: // for the pre/post loops. They are usually very short, duke@435: // so guess that 4 trips is a reasonable value. duke@435: post_head->set_profile_trip_cnt(4.0); duke@435: pre_head->set_profile_trip_cnt(4.0); duke@435: duke@435: // Now force out all loop-invariant dominating tests. The optimizer duke@435: // finds some, but we _know_ they are all useless. duke@435: peeled_dom_test_elim(loop,old_new); duke@435: } duke@435: duke@435: //------------------------------is_invariant----------------------------- duke@435: // Return true if n is invariant duke@435: bool IdealLoopTree::is_invariant(Node* n) const { cfang@1607: Node *n_c = _phase->has_ctrl(n) ? _phase->get_ctrl(n) : n; duke@435: if (n_c->is_top()) return false; duke@435: return !is_member(_phase->get_loop(n_c)); duke@435: } duke@435: duke@435: duke@435: //------------------------------do_unroll-------------------------------------- duke@435: // Unroll the loop body one step - make each trip do 2 iterations. duke@435: void PhaseIdealLoop::do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip ) { duke@435: assert( LoopUnrollLimit, "" ); duke@435: #ifndef PRODUCT duke@435: if( PrintOpto && VerifyLoopOptimizations ) { duke@435: tty->print("Unrolling "); duke@435: loop->dump_head(); duke@435: } duke@435: #endif duke@435: CountedLoopNode *loop_head = loop->_head->as_CountedLoop(); duke@435: CountedLoopEndNode *loop_end = loop_head->loopexit(); duke@435: assert( loop_end, "" ); duke@435: duke@435: // Remember loop node count before unrolling to detect duke@435: // if rounds of unroll,optimize are making progress duke@435: loop_head->set_node_count_before_unroll(loop->_body.size()); duke@435: duke@435: Node *ctrl = loop_head->in(LoopNode::EntryControl); duke@435: Node *limit = loop_head->limit(); duke@435: Node *init = loop_head->init_trip(); duke@435: Node *strid = loop_head->stride(); duke@435: duke@435: Node *opaq = NULL; duke@435: if( adjust_min_trip ) { // If not maximally unrolling, need adjustment duke@435: assert( loop_head->is_main_loop(), "" ); 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: opaq = cmp->in(2); duke@435: // Occasionally it's possible for a pre-loop Opaque1 node to be duke@435: // optimized away and then another round of loop opts attempted. duke@435: // We can not optimize this particular loop in that case. duke@435: if( opaq->Opcode() != Op_Opaque1 ) duke@435: return; // Cannot find pre-loop! Bail out! duke@435: } duke@435: duke@435: C->set_major_progress(); duke@435: duke@435: // Adjust max trip count. The trip count is intentionally rounded duke@435: // down here (e.g. 15-> 7-> 3-> 1) because if we unwittingly over-unroll, duke@435: // the main, unrolled, part of the loop will never execute as it is protected duke@435: // by the min-trip test. See bug 4834191 for a case where we over-unrolled duke@435: // and later determined that part of the unrolled loop was dead. duke@435: loop_head->set_trip_count(loop_head->trip_count() / 2); duke@435: duke@435: // Double the count of original iterations in the unrolled loop body. duke@435: loop_head->double_unrolled_count(); duke@435: duke@435: // ----------- duke@435: // Step 2: Cut back the trip counter for an unroll amount of 2. duke@435: // Loop will normally trip (limit - init)/stride_con. Since it's a duke@435: // CountedLoop this is exact (stride divides limit-init exactly). duke@435: // We are going to double the loop body, so we want to knock off any duke@435: // odd iteration: (trip_cnt & ~1). Then back compute a new limit. duke@435: Node *span = new (C, 3) SubINode( limit, init ); duke@435: register_new_node( span, ctrl ); duke@435: Node *trip = new (C, 3) DivINode( 0, span, strid ); duke@435: register_new_node( trip, ctrl ); duke@435: Node *mtwo = _igvn.intcon(-2); duke@435: set_ctrl(mtwo, C->root()); duke@435: Node *rond = new (C, 3) AndINode( trip, mtwo ); duke@435: register_new_node( rond, ctrl ); duke@435: Node *spn2 = new (C, 3) MulINode( rond, strid ); duke@435: register_new_node( spn2, ctrl ); duke@435: Node *lim2 = new (C, 3) AddINode( spn2, init ); duke@435: register_new_node( lim2, ctrl ); duke@435: duke@435: // Hammer in the new limit duke@435: Node *ctrl2 = loop_end->in(0); duke@435: Node *cmp2 = new (C, 3) CmpINode( loop_head->incr(), lim2 ); duke@435: register_new_node( cmp2, ctrl2 ); duke@435: Node *bol2 = new (C, 2) BoolNode( cmp2, loop_end->test_trip() ); duke@435: register_new_node( bol2, ctrl2 ); duke@435: _igvn.hash_delete(loop_end); duke@435: loop_end->set_req(CountedLoopEndNode::TestValue, bol2); duke@435: duke@435: // Step 3: Find the min-trip test guaranteed before a 'main' loop. duke@435: // Make it a 1-trip test (means at least 2 trips). duke@435: if( adjust_min_trip ) { duke@435: // Guard test uses an 'opaque' node which is not shared. Hence I duke@435: // can edit it's inputs directly. Hammer in the new limit for the duke@435: // minimum-trip guard. duke@435: assert( opaq->outcnt() == 1, "" ); duke@435: _igvn.hash_delete(opaq); duke@435: opaq->set_req(1, lim2); duke@435: } duke@435: duke@435: // --------- duke@435: // Step 4: Clone the loop body. Move it inside the loop. This loop body duke@435: // represents the odd iterations; since the loop trips an even number of duke@435: // times its backedge is never taken. Kill the backedge. duke@435: uint dd = dom_depth(loop_head); duke@435: clone_loop( loop, old_new, dd ); duke@435: duke@435: // Make backedges of the clone equal to backedges of the original. duke@435: // Make the fall-in from the original come from the fall-out of the clone. duke@435: for (DUIterator_Fast jmax, j = loop_head->fast_outs(jmax); j < jmax; j++) { duke@435: Node* phi = loop_head->fast_out(j); duke@435: if( phi->is_Phi() && phi->in(0) == loop_head && phi->outcnt() > 0 ) { duke@435: Node *newphi = old_new[phi->_idx]; duke@435: _igvn.hash_delete( phi ); duke@435: _igvn.hash_delete( newphi ); duke@435: duke@435: phi ->set_req(LoopNode:: EntryControl, newphi->in(LoopNode::LoopBackControl)); duke@435: newphi->set_req(LoopNode::LoopBackControl, phi ->in(LoopNode::LoopBackControl)); duke@435: phi ->set_req(LoopNode::LoopBackControl, C->top()); duke@435: } duke@435: } duke@435: Node *clone_head = old_new[loop_head->_idx]; duke@435: _igvn.hash_delete( clone_head ); duke@435: loop_head ->set_req(LoopNode:: EntryControl, clone_head->in(LoopNode::LoopBackControl)); duke@435: clone_head->set_req(LoopNode::LoopBackControl, loop_head ->in(LoopNode::LoopBackControl)); duke@435: loop_head ->set_req(LoopNode::LoopBackControl, C->top()); duke@435: loop->_head = clone_head; // New loop header duke@435: duke@435: set_idom(loop_head, loop_head ->in(LoopNode::EntryControl), dd); duke@435: set_idom(clone_head, clone_head->in(LoopNode::EntryControl), dd); duke@435: duke@435: // Kill the clone's backedge duke@435: Node *newcle = old_new[loop_end->_idx]; duke@435: _igvn.hash_delete( newcle ); duke@435: Node *one = _igvn.intcon(1); duke@435: set_ctrl(one, C->root()); duke@435: newcle->set_req(1, one); duke@435: // Force clone into same loop body duke@435: uint max = loop->_body.size(); duke@435: for( uint k = 0; k < max; k++ ) { duke@435: Node *old = loop->_body.at(k); duke@435: Node *nnn = old_new[old->_idx]; duke@435: loop->_body.push(nnn); duke@435: if (!has_ctrl(old)) duke@435: set_loop(nnn, loop); duke@435: } never@802: never@802: loop->record_for_igvn(); duke@435: } duke@435: duke@435: //------------------------------do_maximally_unroll---------------------------- duke@435: duke@435: void PhaseIdealLoop::do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new ) { duke@435: CountedLoopNode *cl = loop->_head->as_CountedLoop(); duke@435: assert( cl->trip_count() > 0, ""); duke@435: duke@435: // If loop is tripping an odd number of times, peel odd iteration duke@435: if( (cl->trip_count() & 1) == 1 ) { duke@435: do_peeling( loop, old_new ); duke@435: } duke@435: duke@435: // Now its tripping an even number of times remaining. Double loop body. duke@435: // Do not adjust pre-guards; they are not needed and do not exist. duke@435: if( cl->trip_count() > 0 ) { duke@435: do_unroll( loop, old_new, false ); duke@435: } duke@435: } duke@435: duke@435: //------------------------------dominates_backedge--------------------------------- duke@435: // Returns true if ctrl is executed on every complete iteration duke@435: bool IdealLoopTree::dominates_backedge(Node* ctrl) { duke@435: assert(ctrl->is_CFG(), "must be control"); duke@435: Node* backedge = _head->as_Loop()->in(LoopNode::LoopBackControl); duke@435: return _phase->dom_lca_internal(ctrl, backedge) == ctrl; duke@435: } duke@435: duke@435: //------------------------------add_constraint--------------------------------- duke@435: // Constrain the main loop iterations so the condition: duke@435: // scale_con * I + offset < limit duke@435: // always holds true. That is, either increase the number of iterations in duke@435: // the pre-loop or the post-loop until the condition holds true in the main duke@435: // loop. Stride, scale, offset and limit are all loop invariant. Further, duke@435: // stride and scale are constants (offset and limit often are). duke@435: void PhaseIdealLoop::add_constraint( int stride_con, int scale_con, Node *offset, Node *limit, Node *pre_ctrl, Node **pre_limit, Node **main_limit ) { duke@435: duke@435: // Compute "I :: (limit-offset)/scale_con" duke@435: Node *con = new (C, 3) SubINode( limit, offset ); duke@435: register_new_node( con, pre_ctrl ); duke@435: Node *scale = _igvn.intcon(scale_con); duke@435: set_ctrl(scale, C->root()); duke@435: Node *X = new (C, 3) DivINode( 0, con, scale ); duke@435: register_new_node( X, pre_ctrl ); duke@435: duke@435: // For positive stride, the pre-loop limit always uses a MAX function duke@435: // and the main loop a MIN function. For negative stride these are duke@435: // reversed. duke@435: duke@435: // Also for positive stride*scale the affine function is increasing, so the duke@435: // pre-loop must check for underflow and the post-loop for overflow. duke@435: // Negative stride*scale reverses this; pre-loop checks for overflow and duke@435: // post-loop for underflow. duke@435: if( stride_con*scale_con > 0 ) { duke@435: // Compute I < (limit-offset)/scale_con duke@435: // Adjust main-loop last iteration to be MIN/MAX(main_loop,X) duke@435: *main_limit = (stride_con > 0) duke@435: ? (Node*)(new (C, 3) MinINode( *main_limit, X )) duke@435: : (Node*)(new (C, 3) MaxINode( *main_limit, X )); duke@435: register_new_node( *main_limit, pre_ctrl ); duke@435: duke@435: } else { duke@435: // Compute (limit-offset)/scale_con + SGN(-scale_con) <= I duke@435: // Add the negation of the main-loop constraint to the pre-loop. duke@435: // See footnote [++] below for a derivation of the limit expression. duke@435: Node *incr = _igvn.intcon(scale_con > 0 ? -1 : 1); duke@435: set_ctrl(incr, C->root()); duke@435: Node *adj = new (C, 3) AddINode( X, incr ); duke@435: register_new_node( adj, pre_ctrl ); duke@435: *pre_limit = (scale_con > 0) duke@435: ? (Node*)new (C, 3) MinINode( *pre_limit, adj ) duke@435: : (Node*)new (C, 3) MaxINode( *pre_limit, adj ); duke@435: register_new_node( *pre_limit, pre_ctrl ); duke@435: duke@435: // [++] Here's the algebra that justifies the pre-loop limit expression: duke@435: // duke@435: // NOT( scale_con * I + offset < limit ) duke@435: // == duke@435: // scale_con * I + offset >= limit duke@435: // == duke@435: // SGN(scale_con) * I >= (limit-offset)/|scale_con| duke@435: // == duke@435: // (limit-offset)/|scale_con| <= I * SGN(scale_con) duke@435: // == duke@435: // (limit-offset)/|scale_con|-1 < I * SGN(scale_con) duke@435: // == duke@435: // ( if (scale_con > 0) /*common case*/ duke@435: // (limit-offset)/scale_con - 1 < I duke@435: // else duke@435: // (limit-offset)/scale_con + 1 > I duke@435: // ) duke@435: // ( if (scale_con > 0) /*common case*/ duke@435: // (limit-offset)/scale_con + SGN(-scale_con) < I duke@435: // else duke@435: // (limit-offset)/scale_con + SGN(-scale_con) > I duke@435: } duke@435: } duke@435: duke@435: duke@435: //------------------------------is_scaled_iv--------------------------------- duke@435: // Return true if exp is a constant times an induction var duke@435: bool PhaseIdealLoop::is_scaled_iv(Node* exp, Node* iv, int* p_scale) { duke@435: if (exp == iv) { duke@435: if (p_scale != NULL) { duke@435: *p_scale = 1; duke@435: } duke@435: return true; duke@435: } duke@435: int opc = exp->Opcode(); duke@435: if (opc == Op_MulI) { duke@435: if (exp->in(1) == iv && exp->in(2)->is_Con()) { duke@435: if (p_scale != NULL) { duke@435: *p_scale = exp->in(2)->get_int(); duke@435: } duke@435: return true; duke@435: } duke@435: if (exp->in(2) == iv && exp->in(1)->is_Con()) { duke@435: if (p_scale != NULL) { duke@435: *p_scale = exp->in(1)->get_int(); duke@435: } duke@435: return true; duke@435: } duke@435: } else if (opc == Op_LShiftI) { duke@435: if (exp->in(1) == iv && exp->in(2)->is_Con()) { duke@435: if (p_scale != NULL) { duke@435: *p_scale = 1 << exp->in(2)->get_int(); duke@435: } duke@435: return true; duke@435: } duke@435: } duke@435: return false; duke@435: } duke@435: duke@435: //-----------------------------is_scaled_iv_plus_offset------------------------------ duke@435: // Return true if exp is a simple induction variable expression: k1*iv + (invar + k2) duke@435: bool PhaseIdealLoop::is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth) { duke@435: if (is_scaled_iv(exp, iv, p_scale)) { duke@435: if (p_offset != NULL) { duke@435: Node *zero = _igvn.intcon(0); duke@435: set_ctrl(zero, C->root()); duke@435: *p_offset = zero; duke@435: } duke@435: return true; duke@435: } duke@435: int opc = exp->Opcode(); duke@435: if (opc == Op_AddI) { duke@435: if (is_scaled_iv(exp->in(1), iv, p_scale)) { duke@435: if (p_offset != NULL) { duke@435: *p_offset = exp->in(2); duke@435: } duke@435: return true; duke@435: } duke@435: if (exp->in(2)->is_Con()) { duke@435: Node* offset2 = NULL; duke@435: if (depth < 2 && duke@435: is_scaled_iv_plus_offset(exp->in(1), iv, p_scale, duke@435: p_offset != NULL ? &offset2 : NULL, depth+1)) { duke@435: if (p_offset != NULL) { duke@435: Node *ctrl_off2 = get_ctrl(offset2); duke@435: Node* offset = new (C, 3) AddINode(offset2, exp->in(2)); duke@435: register_new_node(offset, ctrl_off2); duke@435: *p_offset = offset; duke@435: } duke@435: return true; duke@435: } duke@435: } duke@435: } else if (opc == Op_SubI) { duke@435: if (is_scaled_iv(exp->in(1), iv, p_scale)) { duke@435: if (p_offset != NULL) { duke@435: Node *zero = _igvn.intcon(0); duke@435: set_ctrl(zero, C->root()); duke@435: Node *ctrl_off = get_ctrl(exp->in(2)); duke@435: Node* offset = new (C, 3) SubINode(zero, exp->in(2)); duke@435: register_new_node(offset, ctrl_off); duke@435: *p_offset = offset; duke@435: } duke@435: return true; duke@435: } duke@435: if (is_scaled_iv(exp->in(2), iv, p_scale)) { duke@435: if (p_offset != NULL) { duke@435: *p_scale *= -1; duke@435: *p_offset = exp->in(1); duke@435: } duke@435: return true; duke@435: } duke@435: } duke@435: return false; duke@435: } duke@435: duke@435: //------------------------------do_range_check--------------------------------- duke@435: // Eliminate range-checks and other trip-counter vs loop-invariant tests. duke@435: void PhaseIdealLoop::do_range_check( IdealLoopTree *loop, Node_List &old_new ) { duke@435: #ifndef PRODUCT duke@435: if( PrintOpto && VerifyLoopOptimizations ) { duke@435: tty->print("Range Check Elimination "); duke@435: loop->dump_head(); duke@435: } duke@435: #endif duke@435: assert( RangeCheckElimination, "" ); duke@435: CountedLoopNode *cl = loop->_head->as_CountedLoop(); duke@435: assert( cl->is_main_loop(), "" ); duke@435: duke@435: // Find the trip counter; we are iteration splitting based on it duke@435: Node *trip_counter = cl->phi(); duke@435: // Find the main loop limit; we will trim it's iterations duke@435: // to not ever trip end tests duke@435: Node *main_limit = cl->limit(); duke@435: // Find the pre-loop limit; we will expand it's iterations to duke@435: // not ever trip low tests. duke@435: Node *ctrl = cl->in(LoopNode::EntryControl); duke@435: assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" ); duke@435: Node *iffm = ctrl->in(0); duke@435: assert( iffm->Opcode() == Op_If, "" ); duke@435: Node *p_f = iffm->in(0); duke@435: assert( p_f->Opcode() == Op_IfFalse, "" ); duke@435: CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd(); duke@435: assert( pre_end->loopnode()->is_pre_loop(), "" ); duke@435: Node *pre_opaq1 = pre_end->limit(); duke@435: // Occasionally it's possible for a pre-loop Opaque1 node to be duke@435: // optimized away and then another round of loop opts attempted. duke@435: // We can not optimize this particular loop in that case. duke@435: if( pre_opaq1->Opcode() != Op_Opaque1 ) duke@435: return; duke@435: Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1; duke@435: Node *pre_limit = pre_opaq->in(1); duke@435: duke@435: // Where do we put new limit calculations duke@435: Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl); duke@435: duke@435: // Ensure the original loop limit is available from the duke@435: // pre-loop Opaque1 node. duke@435: Node *orig_limit = pre_opaq->original_loop_limit(); duke@435: if( orig_limit == NULL || _igvn.type(orig_limit) == Type::TOP ) duke@435: return; duke@435: duke@435: // Need to find the main-loop zero-trip guard duke@435: Node *bolzm = iffm->in(1); duke@435: assert( bolzm->Opcode() == Op_Bool, "" ); duke@435: Node *cmpzm = bolzm->in(1); duke@435: assert( cmpzm->is_Cmp(), "" ); duke@435: Node *opqzm = cmpzm->in(2); duke@435: if( opqzm->Opcode() != Op_Opaque1 ) duke@435: return; duke@435: assert( opqzm->in(1) == main_limit, "do not understand situation" ); duke@435: duke@435: // Must know if its a count-up or count-down loop duke@435: duke@435: // protect against stride not being a constant duke@435: if ( !cl->stride_is_con() ) { duke@435: return; duke@435: } duke@435: int stride_con = cl->stride_con(); duke@435: Node *zero = _igvn.intcon(0); duke@435: Node *one = _igvn.intcon(1); duke@435: set_ctrl(zero, C->root()); duke@435: set_ctrl(one, C->root()); duke@435: duke@435: // Range checks that do not dominate the loop backedge (ie. duke@435: // conditionally executed) can lengthen the pre loop limit beyond duke@435: // the original loop limit. To prevent this, the pre limit is duke@435: // (for stride > 0) MINed with the original loop limit (MAXed duke@435: // stride < 0) when some range_check (rc) is conditionally duke@435: // executed. duke@435: bool conditional_rc = false; duke@435: duke@435: // Check loop body for tests of trip-counter plus loop-invariant vs duke@435: // loop-invariant. duke@435: for( uint i = 0; i < loop->_body.size(); i++ ) { duke@435: Node *iff = loop->_body[i]; duke@435: if( iff->Opcode() == Op_If ) { // Test? duke@435: duke@435: // Test is an IfNode, has 2 projections. If BOTH are in the loop duke@435: // we need loop unswitching instead of iteration splitting. duke@435: Node *exit = loop->is_loop_exit(iff); duke@435: if( !exit ) continue; duke@435: int flip = (exit->Opcode() == Op_IfTrue) ? 1 : 0; duke@435: duke@435: // Get boolean condition to test duke@435: Node *i1 = iff->in(1); duke@435: if( !i1->is_Bool() ) continue; duke@435: BoolNode *bol = i1->as_Bool(); duke@435: BoolTest b_test = bol->_test; duke@435: // Flip sense of test if exit condition is flipped duke@435: if( flip ) duke@435: b_test = b_test.negate(); duke@435: duke@435: // Get compare duke@435: Node *cmp = bol->in(1); duke@435: duke@435: // Look for trip_counter + offset vs limit duke@435: Node *rc_exp = cmp->in(1); duke@435: Node *limit = cmp->in(2); duke@435: jint scale_con= 1; // Assume trip counter not scaled duke@435: duke@435: Node *limit_c = get_ctrl(limit); duke@435: if( loop->is_member(get_loop(limit_c) ) ) { duke@435: // Compare might have operands swapped; commute them duke@435: b_test = b_test.commute(); duke@435: rc_exp = cmp->in(2); duke@435: limit = cmp->in(1); duke@435: limit_c = get_ctrl(limit); duke@435: if( loop->is_member(get_loop(limit_c) ) ) duke@435: continue; // Both inputs are loop varying; cannot RCE duke@435: } duke@435: // Here we know 'limit' is loop invariant duke@435: duke@435: // 'limit' maybe pinned below the zero trip test (probably from a duke@435: // previous round of rce), in which case, it can't be used in the duke@435: // zero trip test expression which must occur before the zero test's if. duke@435: if( limit_c == ctrl ) { duke@435: continue; // Don't rce this check but continue looking for other candidates. duke@435: } duke@435: duke@435: // Check for scaled induction variable plus an offset duke@435: Node *offset = NULL; duke@435: duke@435: if (!is_scaled_iv_plus_offset(rc_exp, trip_counter, &scale_con, &offset)) { duke@435: continue; duke@435: } duke@435: duke@435: Node *offset_c = get_ctrl(offset); duke@435: if( loop->is_member( get_loop(offset_c) ) ) duke@435: continue; // Offset is not really loop invariant duke@435: // Here we know 'offset' is loop invariant. duke@435: duke@435: // As above for the 'limit', the 'offset' maybe pinned below the duke@435: // zero trip test. duke@435: if( offset_c == ctrl ) { duke@435: continue; // Don't rce this check but continue looking for other candidates. duke@435: } duke@435: duke@435: // At this point we have the expression as: duke@435: // scale_con * trip_counter + offset :: limit duke@435: // where scale_con, offset and limit are loop invariant. Trip_counter duke@435: // monotonically increases by stride_con, a constant. Both (or either) duke@435: // stride_con and scale_con can be negative which will flip about the duke@435: // sense of the test. duke@435: duke@435: // Adjust pre and main loop limits to guard the correct iteration set duke@435: if( cmp->Opcode() == Op_CmpU ) {// Unsigned compare is really 2 tests duke@435: if( b_test._test == BoolTest::lt ) { // Range checks always use lt duke@435: // The overflow limit: scale*I+offset < limit duke@435: add_constraint( stride_con, scale_con, offset, limit, pre_ctrl, &pre_limit, &main_limit ); duke@435: // The underflow limit: 0 <= scale*I+offset. duke@435: // Some math yields: -scale*I-(offset+1) < 0 duke@435: Node *plus_one = new (C, 3) AddINode( offset, one ); duke@435: register_new_node( plus_one, pre_ctrl ); duke@435: Node *neg_offset = new (C, 3) SubINode( zero, plus_one ); duke@435: register_new_node( neg_offset, pre_ctrl ); duke@435: add_constraint( stride_con, -scale_con, neg_offset, zero, pre_ctrl, &pre_limit, &main_limit ); duke@435: if (!conditional_rc) { duke@435: conditional_rc = !loop->dominates_backedge(iff); duke@435: } duke@435: } else { duke@435: #ifndef PRODUCT duke@435: if( PrintOpto ) duke@435: tty->print_cr("missed RCE opportunity"); duke@435: #endif duke@435: continue; // In release mode, ignore it duke@435: } duke@435: } else { // Otherwise work on normal compares duke@435: switch( b_test._test ) { duke@435: case BoolTest::ge: // Convert X >= Y to -X <= -Y duke@435: scale_con = -scale_con; duke@435: offset = new (C, 3) SubINode( zero, offset ); duke@435: register_new_node( offset, pre_ctrl ); duke@435: limit = new (C, 3) SubINode( zero, limit ); duke@435: register_new_node( limit, pre_ctrl ); duke@435: // Fall into LE case duke@435: case BoolTest::le: // Convert X <= Y to X < Y+1 duke@435: limit = new (C, 3) AddINode( limit, one ); duke@435: register_new_node( limit, pre_ctrl ); duke@435: // Fall into LT case duke@435: case BoolTest::lt: duke@435: add_constraint( stride_con, scale_con, offset, limit, pre_ctrl, &pre_limit, &main_limit ); duke@435: if (!conditional_rc) { duke@435: conditional_rc = !loop->dominates_backedge(iff); duke@435: } duke@435: break; duke@435: default: duke@435: #ifndef PRODUCT duke@435: if( PrintOpto ) duke@435: tty->print_cr("missed RCE opportunity"); duke@435: #endif duke@435: continue; // Unhandled case duke@435: } duke@435: } duke@435: duke@435: // Kill the eliminated test duke@435: C->set_major_progress(); duke@435: Node *kill_con = _igvn.intcon( 1-flip ); duke@435: set_ctrl(kill_con, C->root()); duke@435: _igvn.hash_delete(iff); duke@435: iff->set_req(1, kill_con); duke@435: _igvn._worklist.push(iff); duke@435: // Find surviving projection duke@435: assert(iff->is_If(), ""); duke@435: ProjNode* dp = ((IfNode*)iff)->proj_out(1-flip); duke@435: // Find loads off the surviving projection; remove their control edge duke@435: for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) { duke@435: Node* cd = dp->fast_out(i); // Control-dependent node duke@435: if( cd->is_Load() ) { // Loads can now float around in the loop duke@435: _igvn.hash_delete(cd); duke@435: // Allow the load to float around in the loop, or before it duke@435: // but NOT before the pre-loop. duke@435: cd->set_req(0, ctrl); // ctrl, not NULL duke@435: _igvn._worklist.push(cd); duke@435: --i; duke@435: --imax; duke@435: } duke@435: } duke@435: duke@435: } // End of is IF duke@435: duke@435: } duke@435: duke@435: // Update loop limits duke@435: if (conditional_rc) { duke@435: pre_limit = (stride_con > 0) ? (Node*)new (C,3) MinINode(pre_limit, orig_limit) duke@435: : (Node*)new (C,3) MaxINode(pre_limit, orig_limit); duke@435: register_new_node(pre_limit, pre_ctrl); duke@435: } duke@435: _igvn.hash_delete(pre_opaq); duke@435: pre_opaq->set_req(1, pre_limit); duke@435: duke@435: // Note:: we are making the main loop limit no longer precise; duke@435: // need to round up based on stride. duke@435: if( stride_con != 1 && stride_con != -1 ) { // Cutout for common case duke@435: // "Standard" round-up logic: ([main_limit-init+(y-1)]/y)*y+init duke@435: // Hopefully, compiler will optimize for powers of 2. duke@435: Node *ctrl = get_ctrl(main_limit); duke@435: Node *stride = cl->stride(); duke@435: Node *init = cl->init_trip(); duke@435: Node *span = new (C, 3) SubINode(main_limit,init); duke@435: register_new_node(span,ctrl); duke@435: Node *rndup = _igvn.intcon(stride_con + ((stride_con>0)?-1:1)); duke@435: Node *add = new (C, 3) AddINode(span,rndup); duke@435: register_new_node(add,ctrl); duke@435: Node *div = new (C, 3) DivINode(0,add,stride); duke@435: register_new_node(div,ctrl); duke@435: Node *mul = new (C, 3) MulINode(div,stride); duke@435: register_new_node(mul,ctrl); duke@435: Node *newlim = new (C, 3) AddINode(mul,init); duke@435: register_new_node(newlim,ctrl); duke@435: main_limit = newlim; duke@435: } duke@435: duke@435: Node *main_cle = cl->loopexit(); duke@435: Node *main_bol = main_cle->in(1); duke@435: // Hacking loop bounds; need private copies of exit test duke@435: if( main_bol->outcnt() > 1 ) {// BoolNode shared? duke@435: _igvn.hash_delete(main_cle); duke@435: main_bol = main_bol->clone();// Clone a private BoolNode duke@435: register_new_node( main_bol, main_cle->in(0) ); duke@435: main_cle->set_req(1,main_bol); duke@435: } duke@435: Node *main_cmp = main_bol->in(1); duke@435: if( main_cmp->outcnt() > 1 ) { // CmpNode shared? duke@435: _igvn.hash_delete(main_bol); duke@435: main_cmp = main_cmp->clone();// Clone a private CmpNode duke@435: register_new_node( main_cmp, main_cle->in(0) ); duke@435: main_bol->set_req(1,main_cmp); duke@435: } duke@435: // Hack the now-private loop bounds duke@435: _igvn.hash_delete(main_cmp); duke@435: main_cmp->set_req(2, main_limit); duke@435: _igvn._worklist.push(main_cmp); duke@435: // The OpaqueNode is unshared by design duke@435: _igvn.hash_delete(opqzm); duke@435: assert( opqzm->outcnt() == 1, "cannot hack shared node" ); duke@435: opqzm->set_req(1,main_limit); duke@435: _igvn._worklist.push(opqzm); duke@435: } duke@435: duke@435: //------------------------------DCE_loop_body---------------------------------- duke@435: // Remove simplistic dead code from loop body duke@435: void IdealLoopTree::DCE_loop_body() { duke@435: for( uint i = 0; i < _body.size(); i++ ) duke@435: if( _body.at(i)->outcnt() == 0 ) duke@435: _body.map( i--, _body.pop() ); duke@435: } duke@435: duke@435: duke@435: //------------------------------adjust_loop_exit_prob-------------------------- duke@435: // Look for loop-exit tests with the 50/50 (or worse) guesses from the parsing stage. duke@435: // Replace with a 1-in-10 exit guess. duke@435: void IdealLoopTree::adjust_loop_exit_prob( PhaseIdealLoop *phase ) { duke@435: Node *test = tail(); duke@435: while( test != _head ) { duke@435: uint top = test->Opcode(); duke@435: if( top == Op_IfTrue || top == Op_IfFalse ) { duke@435: int test_con = ((ProjNode*)test)->_con; duke@435: assert(top == (uint)(test_con? Op_IfTrue: Op_IfFalse), "sanity"); duke@435: IfNode *iff = test->in(0)->as_If(); duke@435: if( iff->outcnt() == 2 ) { // Ignore dead tests duke@435: Node *bol = iff->in(1); duke@435: if( bol && bol->req() > 1 && bol->in(1) && duke@435: ((bol->in(1)->Opcode() == Op_StorePConditional ) || kvn@855: (bol->in(1)->Opcode() == Op_StoreIConditional ) || duke@435: (bol->in(1)->Opcode() == Op_StoreLConditional ) || duke@435: (bol->in(1)->Opcode() == Op_CompareAndSwapI ) || duke@435: (bol->in(1)->Opcode() == Op_CompareAndSwapL ) || coleenp@548: (bol->in(1)->Opcode() == Op_CompareAndSwapP ) || coleenp@548: (bol->in(1)->Opcode() == Op_CompareAndSwapN ))) duke@435: return; // Allocation loops RARELY take backedge duke@435: // Find the OTHER exit path from the IF duke@435: Node* ex = iff->proj_out(1-test_con); duke@435: float p = iff->_prob; duke@435: if( !phase->is_member( this, ex ) && iff->_fcnt == COUNT_UNKNOWN ) { duke@435: if( top == Op_IfTrue ) { duke@435: if( p < (PROB_FAIR + PROB_UNLIKELY_MAG(3))) { duke@435: iff->_prob = PROB_STATIC_FREQUENT; duke@435: } duke@435: } else { duke@435: if( p > (PROB_FAIR - PROB_UNLIKELY_MAG(3))) { duke@435: iff->_prob = PROB_STATIC_INFREQUENT; duke@435: } duke@435: } duke@435: } duke@435: } duke@435: } duke@435: test = phase->idom(test); duke@435: } duke@435: } duke@435: duke@435: duke@435: //------------------------------policy_do_remove_empty_loop-------------------- duke@435: // Micro-benchmark spamming. Policy is to always remove empty loops. duke@435: // The 'DO' part is to replace the trip counter with the value it will duke@435: // have on the last iteration. This will break the loop. duke@435: bool IdealLoopTree::policy_do_remove_empty_loop( PhaseIdealLoop *phase ) { duke@435: // Minimum size must be empty loop duke@435: if( _body.size() > 7/*number of nodes in an empty loop*/ ) return false; duke@435: duke@435: if( !_head->is_CountedLoop() ) return false; // Dead loop duke@435: CountedLoopNode *cl = _head->as_CountedLoop(); duke@435: if( !cl->loopexit() ) return false; // Malformed loop duke@435: if( !phase->is_member(this,phase->get_ctrl(cl->loopexit()->in(CountedLoopEndNode::TestValue)) ) ) duke@435: return false; // Infinite loop duke@435: #ifndef PRODUCT duke@435: if( PrintOpto ) duke@435: tty->print_cr("Removing empty loop"); duke@435: #endif duke@435: #ifdef ASSERT duke@435: // Ensure only one phi which is the iv. duke@435: Node* iv = NULL; duke@435: for (DUIterator_Fast imax, i = cl->fast_outs(imax); i < imax; i++) { duke@435: Node* n = cl->fast_out(i); duke@435: if (n->Opcode() == Op_Phi) { duke@435: assert(iv == NULL, "Too many phis" ); duke@435: iv = n; duke@435: } duke@435: } duke@435: assert(iv == cl->phi(), "Wrong phi" ); duke@435: #endif duke@435: // Replace the phi at loop head with the final value of the last duke@435: // iteration. Then the CountedLoopEnd will collapse (backedge never duke@435: // taken) and all loop-invariant uses of the exit values will be correct. duke@435: Node *phi = cl->phi(); duke@435: Node *final = new (phase->C, 3) SubINode( cl->limit(), cl->stride() ); duke@435: phase->register_new_node(final,cl->in(LoopNode::EntryControl)); kvn@1976: phase->_igvn.replace_node(phi,final); duke@435: phase->C->set_major_progress(); duke@435: return true; duke@435: } duke@435: duke@435: duke@435: //============================================================================= duke@435: //------------------------------iteration_split_impl--------------------------- never@836: bool IdealLoopTree::iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new ) { duke@435: // Check and remove empty loops (spam micro-benchmarks) duke@435: if( policy_do_remove_empty_loop(phase) ) cfang@1607: return true; // Here we removed an empty loop duke@435: duke@435: bool should_peel = policy_peeling(phase); // Should we peel? duke@435: duke@435: bool should_unswitch = policy_unswitching(phase); duke@435: duke@435: // Non-counted loops may be peeled; exactly 1 iteration is peeled. duke@435: // This removes loop-invariant tests (usually null checks). duke@435: if( !_head->is_CountedLoop() ) { // Non-counted loop duke@435: if (PartialPeelLoop && phase->partial_peel(this, old_new)) { never@836: // Partial peel succeeded so terminate this round of loop opts never@836: return false; duke@435: } duke@435: if( should_peel ) { // Should we peel? duke@435: #ifndef PRODUCT duke@435: if (PrintOpto) tty->print_cr("should_peel"); duke@435: #endif duke@435: phase->do_peeling(this,old_new); duke@435: } else if( should_unswitch ) { duke@435: phase->do_unswitching(this, old_new); duke@435: } never@836: return true; duke@435: } duke@435: CountedLoopNode *cl = _head->as_CountedLoop(); duke@435: never@836: if( !cl->loopexit() ) return true; // Ignore various kinds of broken loops duke@435: duke@435: // Do nothing special to pre- and post- loops never@836: if( cl->is_pre_loop() || cl->is_post_loop() ) return true; duke@435: duke@435: // Compute loop trip count from profile data duke@435: compute_profile_trip_cnt(phase); duke@435: duke@435: // Before attempting fancy unrolling, RCE or alignment, see if we want duke@435: // to completely unroll this loop or do loop unswitching. duke@435: if( cl->is_normal_loop() ) { cfang@1224: if (should_unswitch) { cfang@1224: phase->do_unswitching(this, old_new); cfang@1224: return true; cfang@1224: } duke@435: bool should_maximally_unroll = policy_maximally_unroll(phase); duke@435: if( should_maximally_unroll ) { duke@435: // Here we did some unrolling and peeling. Eventually we will duke@435: // completely unroll this loop and it will no longer be a loop. duke@435: phase->do_maximally_unroll(this,old_new); never@836: return true; duke@435: } duke@435: } duke@435: duke@435: duke@435: // Counted loops may be peeled, may need some iterations run up duke@435: // front for RCE, and may want to align loop refs to a cache duke@435: // line. Thus we clone a full loop up front whose trip count is duke@435: // at least 1 (if peeling), but may be several more. duke@435: duke@435: // The main loop will start cache-line aligned with at least 1 duke@435: // iteration of the unrolled body (zero-trip test required) and duke@435: // will have some range checks removed. duke@435: duke@435: // A post-loop will finish any odd iterations (leftover after duke@435: // unrolling), plus any needed for RCE purposes. duke@435: duke@435: bool should_unroll = policy_unroll(phase); duke@435: duke@435: bool should_rce = policy_range_check(phase); duke@435: duke@435: bool should_align = policy_align(phase); duke@435: duke@435: // If not RCE'ing (iteration splitting) or Aligning, then we do not duke@435: // need a pre-loop. We may still need to peel an initial iteration but duke@435: // we will not be needing an unknown number of pre-iterations. duke@435: // duke@435: // Basically, if may_rce_align reports FALSE first time through, duke@435: // we will not be able to later do RCE or Aligning on this loop. duke@435: bool may_rce_align = !policy_peel_only(phase) || should_rce || should_align; duke@435: duke@435: // If we have any of these conditions (RCE, alignment, unrolling) met, then duke@435: // we switch to the pre-/main-/post-loop model. This model also covers duke@435: // peeling. duke@435: if( should_rce || should_align || should_unroll ) { duke@435: if( cl->is_normal_loop() ) // Convert to 'pre/main/post' loops duke@435: phase->insert_pre_post_loops(this,old_new, !may_rce_align); duke@435: duke@435: // Adjust the pre- and main-loop limits to let the pre and post loops run duke@435: // with full checks, but the main-loop with no checks. Remove said duke@435: // checks from the main body. duke@435: if( should_rce ) duke@435: phase->do_range_check(this,old_new); duke@435: duke@435: // Double loop body for unrolling. Adjust the minimum-trip test (will do duke@435: // twice as many iterations as before) and the main body limit (only do duke@435: // an even number of trips). If we are peeling, we might enable some RCE duke@435: // and we'd rather unroll the post-RCE'd loop SO... do not unroll if duke@435: // peeling. cfang@1607: if( should_unroll && !should_peel ) cfang@1607: phase->do_unroll(this,old_new, true); duke@435: duke@435: // Adjust the pre-loop limits to align the main body duke@435: // iterations. duke@435: if( should_align ) duke@435: Unimplemented(); duke@435: duke@435: } else { // Else we have an unchanged counted loop duke@435: if( should_peel ) // Might want to peel but do nothing else duke@435: phase->do_peeling(this,old_new); duke@435: } never@836: return true; duke@435: } duke@435: duke@435: duke@435: //============================================================================= duke@435: //------------------------------iteration_split-------------------------------- never@836: bool IdealLoopTree::iteration_split( PhaseIdealLoop *phase, Node_List &old_new ) { duke@435: // Recursively iteration split nested loops never@836: if( _child && !_child->iteration_split( phase, old_new )) never@836: return false; duke@435: duke@435: // Clean out prior deadwood duke@435: DCE_loop_body(); duke@435: duke@435: duke@435: // Look for loop-exit tests with my 50/50 guesses from the Parsing stage. duke@435: // Replace with a 1-in-10 exit guess. duke@435: if( _parent /*not the root loop*/ && duke@435: !_irreducible && duke@435: // Also ignore the occasional dead backedge duke@435: !tail()->is_top() ) { duke@435: adjust_loop_exit_prob(phase); duke@435: } duke@435: duke@435: duke@435: // Gate unrolling, RCE and peeling efforts. duke@435: if( !_child && // If not an inner loop, do not split duke@435: !_irreducible && kvn@474: _allow_optimizations && duke@435: !tail()->is_top() ) { // Also ignore the occasional dead backedge duke@435: if (!_has_call) { cfang@1607: if (!iteration_split_impl( phase, old_new )) { cfang@1607: return false; cfang@1607: } duke@435: } else if (policy_unswitching(phase)) { duke@435: phase->do_unswitching(this, old_new); duke@435: } duke@435: } duke@435: duke@435: // Minor offset re-organization to remove loop-fallout uses of duke@435: // trip counter. duke@435: if( _head->is_CountedLoop() ) phase->reorg_offsets( this ); never@836: if( _next && !_next->iteration_split( phase, old_new )) never@836: return false; never@836: return true; duke@435: } cfang@1607: cfang@1607: //-------------------------------is_uncommon_trap_proj---------------------------- cfang@1607: // Return true if proj is the form of "proj->[region->..]call_uct" cfang@1607: bool PhaseIdealLoop::is_uncommon_trap_proj(ProjNode* proj, bool must_reason_predicate) { cfang@1607: int path_limit = 10; cfang@1607: assert(proj, "invalid argument"); cfang@1607: Node* out = proj; cfang@1607: for (int ct = 0; ct < path_limit; ct++) { cfang@1607: out = out->unique_ctrl_out(); cfang@1607: if (out == NULL || out->is_Root() || out->is_Start()) cfang@1607: return false; cfang@1607: if (out->is_CallStaticJava()) { cfang@1607: int req = out->as_CallStaticJava()->uncommon_trap_request(); cfang@1607: if (req != 0) { cfang@1607: Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(req); cfang@1607: if (!must_reason_predicate || reason == Deoptimization::Reason_predicate){ cfang@1607: return true; cfang@1607: } cfang@1607: } cfang@1607: return false; // don't do further after call cfang@1607: } cfang@1607: } cfang@1607: return false; cfang@1607: } cfang@1607: cfang@1607: //-------------------------------is_uncommon_trap_if_pattern------------------------- cfang@1607: // Return true for "if(test)-> proj -> ... cfang@1607: // | cfang@1607: // V cfang@1607: // other_proj->[region->..]call_uct" cfang@1607: // cfang@1607: // "must_reason_predicate" means the uct reason must be Reason_predicate cfang@1607: bool PhaseIdealLoop::is_uncommon_trap_if_pattern(ProjNode *proj, bool must_reason_predicate) { cfang@1607: Node *in0 = proj->in(0); cfang@1607: if (!in0->is_If()) return false; kvn@1715: // Variation of a dead If node. kvn@1715: if (in0->outcnt() < 2) return false; cfang@1607: IfNode* iff = in0->as_If(); cfang@1607: cfang@1607: // we need "If(Conv2B(Opaque1(...)))" pattern for must_reason_predicate cfang@1607: if (must_reason_predicate) { cfang@1607: if (iff->in(1)->Opcode() != Op_Conv2B || cfang@1607: iff->in(1)->in(1)->Opcode() != Op_Opaque1) { cfang@1607: return false; cfang@1607: } cfang@1607: } cfang@1607: cfang@1607: ProjNode* other_proj = iff->proj_out(1-proj->_con)->as_Proj(); cfang@1607: return is_uncommon_trap_proj(other_proj, must_reason_predicate); cfang@1607: } cfang@1607: cfang@1607: //------------------------------create_new_if_for_predicate------------------------ cfang@1607: // create a new if above the uct_if_pattern for the predicate to be promoted. cfang@1607: // cfang@1607: // before after cfang@1607: // ---------- ---------- cfang@1607: // ctrl ctrl cfang@1607: // | | cfang@1607: // | | cfang@1607: // v v cfang@1607: // iff new_iff cfang@1607: // / \ / \ cfang@1607: // / \ / \ cfang@1607: // v v v v cfang@1607: // uncommon_proj cont_proj if_uct if_cont cfang@1607: // \ | | | | cfang@1607: // \ | | | | cfang@1607: // v v v | v cfang@1607: // rgn loop | iff cfang@1607: // | | / \ cfang@1607: // | | / \ cfang@1607: // v | v v cfang@1607: // uncommon_trap | uncommon_proj cont_proj cfang@1607: // \ \ | | cfang@1607: // \ \ | | cfang@1607: // v v v v cfang@1607: // rgn loop cfang@1607: // | cfang@1607: // | cfang@1607: // v cfang@1607: // uncommon_trap cfang@1607: // cfang@1607: // cfang@1607: // We will create a region to guard the uct call if there is no one there. cfang@1607: // The true projecttion (if_cont) of the new_iff is returned. cfang@1607: ProjNode* PhaseIdealLoop::create_new_if_for_predicate(ProjNode* cont_proj) { cfang@1607: assert(is_uncommon_trap_if_pattern(cont_proj, true), "must be a uct if pattern!"); cfang@1607: IfNode* iff = cont_proj->in(0)->as_If(); cfang@1607: cfang@1607: ProjNode *uncommon_proj = iff->proj_out(1 - cont_proj->_con); cfang@1607: Node *rgn = uncommon_proj->unique_ctrl_out(); cfang@1607: assert(rgn->is_Region() || rgn->is_Call(), "must be a region or call uct"); cfang@1607: cfang@1607: if (!rgn->is_Region()) { // create a region to guard the call cfang@1607: assert(rgn->is_Call(), "must be call uct"); cfang@1607: CallNode* call = rgn->as_Call(); cfang@1607: rgn = new (C, 1) RegionNode(1); cfang@1607: _igvn.set_type(rgn, rgn->bottom_type()); cfang@1607: rgn->add_req(uncommon_proj); cfang@1607: set_idom(rgn, idom(uncommon_proj), dom_depth(uncommon_proj)+1); cfang@1607: _igvn.hash_delete(call); cfang@1607: call->set_req(0, rgn); cfang@1607: } cfang@1607: cfang@1607: // Create new_iff cfang@1607: uint iffdd = dom_depth(iff); cfang@1607: IdealLoopTree* lp = get_loop(iff); cfang@1607: IfNode *new_iff = new (C, 2) IfNode(iff->in(0), NULL, iff->_prob, iff->_fcnt); cfang@1607: register_node(new_iff, lp, idom(iff), iffdd); cfang@1607: Node *if_cont = new (C, 1) IfTrueNode(new_iff); cfang@1607: Node *if_uct = new (C, 1) IfFalseNode(new_iff); cfang@1607: if (cont_proj->is_IfFalse()) { cfang@1607: // Swap cfang@1607: Node* tmp = if_uct; if_uct = if_cont; if_cont = tmp; cfang@1607: } cfang@1607: register_node(if_cont, lp, new_iff, iffdd); cfang@1607: register_node(if_uct, get_loop(rgn), new_iff, iffdd); cfang@1607: cfang@1607: // if_cont to iff cfang@1607: _igvn.hash_delete(iff); cfang@1607: iff->set_req(0, if_cont); cfang@1607: set_idom(iff, if_cont, dom_depth(iff)); cfang@1607: cfang@1607: // if_uct to rgn cfang@1607: _igvn.hash_delete(rgn); cfang@1607: rgn->add_req(if_uct); cfang@1607: Node* ridom = idom(rgn); cfang@1607: Node* nrdom = dom_lca(ridom, new_iff); cfang@1607: set_idom(rgn, nrdom, dom_depth(rgn)); cfang@1607: cfang@1607: // rgn must have no phis cfang@1607: assert(!rgn->as_Region()->has_phi(), "region must have no phis"); cfang@1607: cfang@1607: return if_cont->as_Proj(); cfang@1607: } cfang@1607: cfang@1607: //------------------------------find_predicate_insertion_point-------------------------- cfang@1607: // Find a good location to insert a predicate cfang@1607: ProjNode* PhaseIdealLoop::find_predicate_insertion_point(Node* start_c) { cfang@1607: if (start_c == C->root() || !start_c->is_Proj()) cfang@1607: return NULL; cfang@1607: if (is_uncommon_trap_if_pattern(start_c->as_Proj(), true/*Reason_Predicate*/)) { cfang@1607: return start_c->as_Proj(); cfang@1607: } cfang@1607: return NULL; cfang@1607: } cfang@1607: cfang@1607: //------------------------------Invariance----------------------------------- cfang@1607: // Helper class for loop_predication_impl to compute invariance on the fly and cfang@1607: // clone invariants. cfang@1607: class Invariance : public StackObj { cfang@1607: VectorSet _visited, _invariant; cfang@1607: Node_Stack _stack; cfang@1607: VectorSet _clone_visited; cfang@1607: Node_List _old_new; // map of old to new (clone) cfang@1607: IdealLoopTree* _lpt; cfang@1607: PhaseIdealLoop* _phase; cfang@1607: cfang@1607: // Helper function to set up the invariance for invariance computation cfang@1607: // If n is a known invariant, set up directly. Otherwise, look up the cfang@1607: // the possibility to push n onto the stack for further processing. cfang@1607: void visit(Node* use, Node* n) { cfang@1607: if (_lpt->is_invariant(n)) { // known invariant cfang@1607: _invariant.set(n->_idx); cfang@1607: } else if (!n->is_CFG()) { cfang@1607: Node *n_ctrl = _phase->ctrl_or_self(n); cfang@1607: Node *u_ctrl = _phase->ctrl_or_self(use); // self if use is a CFG cfang@1607: if (_phase->is_dominator(n_ctrl, u_ctrl)) { cfang@1607: _stack.push(n, n->in(0) == NULL ? 1 : 0); cfang@1607: } cfang@1607: } cfang@1607: } cfang@1607: cfang@1607: // Compute invariance for "the_node" and (possibly) all its inputs recursively cfang@1607: // on the fly cfang@1607: void compute_invariance(Node* n) { cfang@1607: assert(_visited.test(n->_idx), "must be"); cfang@1607: visit(n, n); cfang@1607: while (_stack.is_nonempty()) { cfang@1607: Node* n = _stack.node(); cfang@1607: uint idx = _stack.index(); cfang@1607: if (idx == n->req()) { // all inputs are processed cfang@1607: _stack.pop(); cfang@1607: // n is invariant if it's inputs are all invariant cfang@1607: bool all_inputs_invariant = true; cfang@1607: for (uint i = 0; i < n->req(); i++) { cfang@1607: Node* in = n->in(i); cfang@1607: if (in == NULL) continue; cfang@1607: assert(_visited.test(in->_idx), "must have visited input"); cfang@1607: if (!_invariant.test(in->_idx)) { // bad guy cfang@1607: all_inputs_invariant = false; cfang@1607: break; cfang@1607: } cfang@1607: } cfang@1607: if (all_inputs_invariant) { cfang@1607: _invariant.set(n->_idx); // I am a invariant too cfang@1607: } cfang@1607: } else { // process next input cfang@1607: _stack.set_index(idx + 1); cfang@1607: Node* m = n->in(idx); cfang@1607: if (m != NULL && !_visited.test_set(m->_idx)) { cfang@1607: visit(n, m); cfang@1607: } cfang@1607: } cfang@1607: } cfang@1607: } cfang@1607: cfang@1607: // Helper function to set up _old_new map for clone_nodes. cfang@1607: // If n is a known invariant, set up directly ("clone" of n == n). cfang@1607: // Otherwise, push n onto the stack for real cloning. cfang@1607: void clone_visit(Node* n) { cfang@1607: assert(_invariant.test(n->_idx), "must be invariant"); cfang@1607: if (_lpt->is_invariant(n)) { // known invariant cfang@1607: _old_new.map(n->_idx, n); cfang@1607: } else{ // to be cloned cfang@1607: assert (!n->is_CFG(), "should not see CFG here"); cfang@1607: _stack.push(n, n->in(0) == NULL ? 1 : 0); cfang@1607: } cfang@1607: } cfang@1607: cfang@1607: // Clone "n" and (possibly) all its inputs recursively cfang@1607: void clone_nodes(Node* n, Node* ctrl) { cfang@1607: clone_visit(n); cfang@1607: while (_stack.is_nonempty()) { cfang@1607: Node* n = _stack.node(); cfang@1607: uint idx = _stack.index(); cfang@1607: if (idx == n->req()) { // all inputs processed, clone n! cfang@1607: _stack.pop(); cfang@1607: // clone invariant node cfang@1607: Node* n_cl = n->clone(); cfang@1607: _old_new.map(n->_idx, n_cl); cfang@1607: _phase->register_new_node(n_cl, ctrl); cfang@1607: for (uint i = 0; i < n->req(); i++) { cfang@1607: Node* in = n_cl->in(i); cfang@1607: if (in == NULL) continue; cfang@1607: n_cl->set_req(i, _old_new[in->_idx]); cfang@1607: } cfang@1607: } else { // process next input cfang@1607: _stack.set_index(idx + 1); cfang@1607: Node* m = n->in(idx); cfang@1607: if (m != NULL && !_clone_visited.test_set(m->_idx)) { cfang@1607: clone_visit(m); // visit the input cfang@1607: } cfang@1607: } cfang@1607: } cfang@1607: } cfang@1607: cfang@1607: public: cfang@1607: Invariance(Arena* area, IdealLoopTree* lpt) : cfang@1607: _lpt(lpt), _phase(lpt->_phase), cfang@1607: _visited(area), _invariant(area), _stack(area, 10 /* guess */), cfang@1607: _clone_visited(area), _old_new(area) cfang@1607: {} cfang@1607: cfang@1607: // Map old to n for invariance computation and clone cfang@1607: void map_ctrl(Node* old, Node* n) { cfang@1607: assert(old->is_CFG() && n->is_CFG(), "must be"); cfang@1607: _old_new.map(old->_idx, n); // "clone" of old is n cfang@1607: _invariant.set(old->_idx); // old is invariant cfang@1607: _clone_visited.set(old->_idx); cfang@1607: } cfang@1607: cfang@1607: // Driver function to compute invariance cfang@1607: bool is_invariant(Node* n) { cfang@1607: if (!_visited.test_set(n->_idx)) cfang@1607: compute_invariance(n); cfang@1607: return (_invariant.test(n->_idx) != 0); cfang@1607: } cfang@1607: cfang@1607: // Driver function to clone invariant cfang@1607: Node* clone(Node* n, Node* ctrl) { cfang@1607: assert(ctrl->is_CFG(), "must be"); cfang@1607: assert(_invariant.test(n->_idx), "must be an invariant"); cfang@1607: if (!_clone_visited.test(n->_idx)) cfang@1607: clone_nodes(n, ctrl); cfang@1607: return _old_new[n->_idx]; cfang@1607: } cfang@1607: }; cfang@1607: cfang@1607: //------------------------------is_range_check_if ----------------------------------- cfang@1607: // Returns true if the predicate of iff is in "scale*iv + offset u< load_range(ptr)" format cfang@1607: // Note: this function is particularly designed for loop predication. We require load_range cfang@1607: // and offset to be loop invariant computed on the fly by "invar" cfang@1607: bool IdealLoopTree::is_range_check_if(IfNode *iff, PhaseIdealLoop *phase, Invariance& invar) const { cfang@1607: if (!is_loop_exit(iff)) { cfang@1607: return false; cfang@1607: } cfang@1607: if (!iff->in(1)->is_Bool()) { cfang@1607: return false; cfang@1607: } cfang@1607: const BoolNode *bol = iff->in(1)->as_Bool(); cfang@1607: if (bol->_test._test != BoolTest::lt) { cfang@1607: return false; cfang@1607: } cfang@1607: if (!bol->in(1)->is_Cmp()) { cfang@1607: return false; cfang@1607: } cfang@1607: const CmpNode *cmp = bol->in(1)->as_Cmp(); cfang@1607: if (cmp->Opcode() != Op_CmpU ) { cfang@1607: return false; cfang@1607: } never@2118: Node* range = cmp->in(2); never@2118: if (range->Opcode() != Op_LoadRange) { never@2118: const TypeInt* tint = phase->_igvn.type(range)->isa_int(); never@2118: if (!OptimizeFill || tint == NULL || tint->empty() || tint->_lo < 0) { never@2118: // Allow predication on positive values that aren't LoadRanges. never@2118: // This allows optimization of loops where the length of the never@2118: // array is a known value and doesn't need to be loaded back never@2118: // from the array. never@2118: return false; never@2118: } cfang@1607: } never@2118: if (!invar.is_invariant(range)) { cfang@1607: return false; cfang@1607: } cfang@1607: Node *iv = _head->as_CountedLoop()->phi(); cfang@1607: int scale = 0; cfang@1607: Node *offset = NULL; cfang@1607: if (!phase->is_scaled_iv_plus_offset(cmp->in(1), iv, &scale, &offset)) { cfang@1607: return false; cfang@1607: } cfang@1607: if(offset && !invar.is_invariant(offset)) { // offset must be invariant cfang@1607: return false; cfang@1607: } cfang@1607: return true; cfang@1607: } cfang@1607: cfang@1607: //------------------------------rc_predicate----------------------------------- cfang@1607: // Create a range check predicate cfang@1607: // cfang@1607: // for (i = init; i < limit; i += stride) { cfang@1607: // a[scale*i+offset] cfang@1607: // } cfang@1607: // cfang@1607: // Compute max(scale*i + offset) for init <= i < limit and build the predicate cfang@1607: // as "max(scale*i + offset) u< a.length". cfang@1607: // cfang@1607: // There are two cases for max(scale*i + offset): cfang@1607: // (1) stride*scale > 0 cfang@1607: // max(scale*i + offset) = scale*(limit-stride) + offset cfang@1607: // (2) stride*scale < 0 cfang@1607: // max(scale*i + offset) = scale*init + offset cfang@1607: BoolNode* PhaseIdealLoop::rc_predicate(Node* ctrl, cfang@1607: int scale, Node* offset, cfang@1607: Node* init, Node* limit, Node* stride, never@1738: Node* range, bool upper) { never@1738: DEBUG_ONLY(ttyLocker ttyl); never@1738: if (TraceLoopPredicate) tty->print("rc_predicate "); never@1738: cfang@1607: Node* max_idx_expr = init; cfang@1607: int stride_con = stride->get_int(); never@1738: if ((stride_con > 0) == (scale > 0) == upper) { cfang@1607: max_idx_expr = new (C, 3) SubINode(limit, stride); cfang@1607: register_new_node(max_idx_expr, ctrl); never@1738: if (TraceLoopPredicate) tty->print("(limit - stride) "); never@1738: } else { never@1738: if (TraceLoopPredicate) tty->print("init "); cfang@1607: } cfang@1607: cfang@1607: if (scale != 1) { cfang@1607: ConNode* con_scale = _igvn.intcon(scale); cfang@1607: max_idx_expr = new (C, 3) MulINode(max_idx_expr, con_scale); cfang@1607: register_new_node(max_idx_expr, ctrl); never@1738: if (TraceLoopPredicate) tty->print("* %d ", scale); cfang@1607: } cfang@1607: cfang@1607: if (offset && (!offset->is_Con() || offset->get_int() != 0)){ cfang@1607: max_idx_expr = new (C, 3) AddINode(max_idx_expr, offset); cfang@1607: register_new_node(max_idx_expr, ctrl); never@1738: if (TraceLoopPredicate) never@1738: if (offset->is_Con()) tty->print("+ %d ", offset->get_int()); never@1738: else tty->print("+ offset "); cfang@1607: } cfang@1607: cfang@1607: CmpUNode* cmp = new (C, 3) CmpUNode(max_idx_expr, range); cfang@1607: register_new_node(cmp, ctrl); cfang@1607: BoolNode* bol = new (C, 2) BoolNode(cmp, BoolTest::lt); cfang@1607: register_new_node(bol, ctrl); never@1738: never@1738: if (TraceLoopPredicate) tty->print_cr("_head->is_Loop()) { never@1710: // Could be a simple region when irreducible loops are present. never@1710: return false; never@1710: } never@1710: never@1710: CountedLoopNode *cl = NULL; never@1710: if (loop->_head->is_CountedLoop()) { never@1710: cl = loop->_head->as_CountedLoop(); never@1710: // do nothing for iteration-splitted loops never@1710: if (!cl->is_normal_loop()) return false; never@1710: } never@1710: cfang@1607: // Too many traps seen? cfang@1607: bool tmt = C->too_many_traps(C->method(), 0, Deoptimization::Reason_predicate); cfang@1607: int tc = C->trap_count(Deoptimization::Reason_predicate); cfang@1607: if (tmt || tc > 0) { cfang@1607: if (TraceLoopPredicate) { cfang@1607: tty->print_cr("too many predicate traps: %d", tc); cfang@1607: C->method()->print(); // which method has too many predicate traps cfang@1607: tty->print_cr(""); cfang@1607: } cfang@1607: return false; cfang@1607: } cfang@1607: cfang@1607: LoopNode *lpn = loop->_head->as_Loop(); cfang@1607: Node* entry = lpn->in(LoopNode::EntryControl); cfang@1607: cfang@1607: ProjNode *predicate_proj = find_predicate_insertion_point(entry); cfang@1607: if (!predicate_proj){ cfang@1607: #ifndef PRODUCT cfang@1607: if (TraceLoopPredicate) { cfang@1607: tty->print("missing predicate:"); cfang@1607: loop->dump_head(); cfang@1607: } cfang@1607: #endif cfang@1607: return false; cfang@1607: } cfang@1607: cfang@1607: ConNode* zero = _igvn.intcon(0); cfang@1607: set_ctrl(zero, C->root()); cfang@1607: Node *cond_false = new (C, 2) Conv2BNode(zero); cfang@1607: register_new_node(cond_false, C->root()); cfang@1607: ConNode* one = _igvn.intcon(1); cfang@1607: set_ctrl(one, C->root()); cfang@1607: Node *cond_true = new (C, 2) Conv2BNode(one); cfang@1607: register_new_node(cond_true, C->root()); cfang@1607: cfang@1607: ResourceArea *area = Thread::current()->resource_area(); cfang@1607: Invariance invar(area, loop); cfang@1607: cfang@1607: // Create list of if-projs such that a newer proj dominates all older cfang@1607: // projs in the list, and they all dominate loop->tail() cfang@1607: Node_List if_proj_list(area); cfang@1607: LoopNode *head = loop->_head->as_Loop(); cfang@1607: Node *current_proj = loop->tail(); //start from tail cfang@1607: while ( current_proj != head ) { cfang@1607: if (loop == get_loop(current_proj) && // still in the loop ? cfang@1607: current_proj->is_Proj() && // is a projection ? cfang@1607: current_proj->in(0)->Opcode() == Op_If) { // is a if projection ? cfang@1607: if_proj_list.push(current_proj); cfang@1607: } cfang@1607: current_proj = idom(current_proj); cfang@1607: } cfang@1607: cfang@1607: bool hoisted = false; // true if at least one proj is promoted cfang@1607: while (if_proj_list.size() > 0) { cfang@1607: // Following are changed to nonnull when a predicate can be hoisted cfang@1607: ProjNode* new_predicate_proj = NULL; cfang@1607: cfang@1607: ProjNode* proj = if_proj_list.pop()->as_Proj(); cfang@1607: IfNode* iff = proj->in(0)->as_If(); cfang@1607: cfang@1607: if (!is_uncommon_trap_if_pattern(proj)) { cfang@1607: if (loop->is_loop_exit(iff)) { cfang@1607: // stop processing the remaining projs in the list because the execution of them cfang@1607: // depends on the condition of "iff" (iff->in(1)). cfang@1607: break; cfang@1607: } else { cfang@1607: // Both arms are inside the loop. There are two cases: cfang@1607: // (1) there is one backward branch. In this case, any remaining proj cfang@1607: // in the if_proj list post-dominates "iff". So, the condition of "iff" cfang@1607: // does not determine the execution the remining projs directly, and we cfang@1607: // can safely continue. cfang@1607: // (2) both arms are forwarded, i.e. a diamond shape. In this case, "proj" cfang@1607: // does not dominate loop->tail(), so it can not be in the if_proj list. cfang@1607: continue; cfang@1607: } cfang@1607: } cfang@1607: cfang@1607: Node* test = iff->in(1); cfang@1607: if (!test->is_Bool()){ //Conv2B, ... cfang@1607: continue; cfang@1607: } cfang@1607: BoolNode* bol = test->as_Bool(); cfang@1607: if (invar.is_invariant(bol)) { cfang@1607: // Invariant test cfang@1607: new_predicate_proj = create_new_if_for_predicate(predicate_proj); cfang@1607: Node* ctrl = new_predicate_proj->in(0)->as_If()->in(0); never@1738: BoolNode* new_predicate_bol = invar.clone(bol, ctrl)->as_Bool(); never@1738: never@1738: // Negate test if necessary never@1738: bool negated = false; never@1738: if (proj->_con != predicate_proj->_con) { never@1738: new_predicate_bol = new (C, 2) BoolNode(new_predicate_bol->in(1), new_predicate_bol->_test.negate()); never@1738: register_new_node(new_predicate_bol, ctrl); never@1738: negated = true; never@1738: } never@1738: IfNode* new_predicate_iff = new_predicate_proj->in(0)->as_If(); never@1738: _igvn.hash_delete(new_predicate_iff); never@1738: new_predicate_iff->set_req(1, new_predicate_bol); never@1738: if (TraceLoopPredicate) tty->print_cr("invariant if%s: %d", negated ? " negated" : "", new_predicate_iff->_idx); never@1738: cfang@1607: } else if (cl != NULL && loop->is_range_check_if(iff, this, invar)) { never@1738: assert(proj->_con == predicate_proj->_con, "must match"); never@1738: never@1738: // Range check for counted loops cfang@1607: const Node* cmp = bol->in(1)->as_Cmp(); cfang@1607: Node* idx = cmp->in(1); cfang@1607: assert(!invar.is_invariant(idx), "index is variant"); never@2118: assert(cmp->in(2)->Opcode() == Op_LoadRange || OptimizeFill, "must be"); never@2118: Node* rng = cmp->in(2); never@2118: assert(invar.is_invariant(rng), "range must be invariant"); cfang@1607: int scale = 1; cfang@1607: Node* offset = zero; cfang@1607: bool ok = is_scaled_iv_plus_offset(idx, cl->phi(), &scale, &offset); cfang@1607: assert(ok, "must be index expression"); never@1738: never@1738: Node* init = cl->init_trip(); never@1738: Node* limit = cl->limit(); never@1738: Node* stride = cl->stride(); never@1738: never@1738: // Build if's for the upper and lower bound tests. The never@1738: // lower_bound test will dominate the upper bound test and all never@1738: // cloned or created nodes will use the lower bound test as never@1738: // their declared control. never@1738: ProjNode* lower_bound_proj = create_new_if_for_predicate(predicate_proj); never@1738: ProjNode* upper_bound_proj = create_new_if_for_predicate(predicate_proj); never@1738: assert(upper_bound_proj->in(0)->as_If()->in(0) == lower_bound_proj, "should dominate"); never@1738: Node *ctrl = lower_bound_proj->in(0)->as_If()->in(0); never@1738: never@1738: // Perform cloning to keep Invariance state correct since the never@1738: // late schedule will place invariant things in the loop. never@2118: rng = invar.clone(rng, ctrl); cfang@1607: if (offset && offset != zero) { cfang@1607: assert(invar.is_invariant(offset), "offset must be loop invariant"); cfang@1607: offset = invar.clone(offset, ctrl); cfang@1607: } cfang@1607: never@1738: // Test the lower bound never@2118: Node* lower_bound_bol = rc_predicate(ctrl, scale, offset, init, limit, stride, rng, false); never@1738: IfNode* lower_bound_iff = lower_bound_proj->in(0)->as_If(); never@1738: _igvn.hash_delete(lower_bound_iff); never@1738: lower_bound_iff->set_req(1, lower_bound_bol); never@1738: if (TraceLoopPredicate) tty->print_cr("lower bound check if: %d", lower_bound_iff->_idx); never@1738: never@1738: // Test the upper bound never@2118: Node* upper_bound_bol = rc_predicate(ctrl, scale, offset, init, limit, stride, rng, true); never@1738: IfNode* upper_bound_iff = upper_bound_proj->in(0)->as_If(); never@1738: _igvn.hash_delete(upper_bound_iff); never@1738: upper_bound_iff->set_req(1, upper_bound_bol); never@1738: if (TraceLoopPredicate) tty->print_cr("upper bound check if: %d", lower_bound_iff->_idx); never@1738: never@1738: // Fall through into rest of the clean up code which will move never@1738: // any dependent nodes onto the upper bound test. never@1738: new_predicate_proj = upper_bound_proj; never@1738: } else { cfang@1607: // The other proj of the "iff" is a uncommon trap projection, and we can assume cfang@1607: // the other proj will not be executed ("executed" means uct raised). cfang@1607: continue; never@1738: } cfang@1607: never@1738: // Success - attach condition (new_predicate_bol) to predicate if never@1738: invar.map_ctrl(proj, new_predicate_proj); // so that invariance test can be appropriate never@1738: never@1738: // Eliminate the old if in the loop body never@1738: _igvn.hash_delete(iff); never@1738: iff->set_req(1, proj->is_IfFalse() ? cond_false : cond_true); never@1738: never@1738: Node* ctrl = new_predicate_proj; // new control never@1738: ProjNode* dp = proj; // old control never@1738: assert(get_loop(dp) == loop, "guaranteed at the time of collecting proj"); never@1738: // Find nodes (depends only on the test) off the surviving projection; never@1738: // move them outside the loop with the control of proj_clone never@1738: for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) { never@1738: Node* cd = dp->fast_out(i); // Control-dependent node never@1738: if (cd->depends_only_on_test()) { never@1738: assert(cd->in(0) == dp, ""); never@1738: _igvn.hash_delete(cd); never@1738: cd->set_req(0, ctrl); // ctrl, not NULL never@1738: set_early_ctrl(cd); never@1738: _igvn._worklist.push(cd); never@1738: IdealLoopTree *new_loop = get_loop(get_ctrl(cd)); never@1738: if (new_loop != loop) { never@1738: if (!loop->_child) loop->_body.yank(cd); never@1738: if (!new_loop->_child ) new_loop->_body.push(cd); never@1738: } never@1738: --i; never@1738: --imax; cfang@1607: } never@1738: } cfang@1607: never@1738: hoisted = true; never@1738: C->set_major_progress(); cfang@1607: } // end while cfang@1607: cfang@1607: #ifndef PRODUCT never@1738: // report that the loop predication has been actually performed never@1738: // for this loop never@1738: if (TraceLoopPredicate && hoisted) { never@1738: tty->print("Loop Predication Performed:"); never@1738: loop->dump_head(); never@1738: } cfang@1607: #endif cfang@1607: cfang@1607: return hoisted; cfang@1607: } cfang@1607: cfang@1607: //------------------------------loop_predication-------------------------------- cfang@1607: // driver routine for loop predication optimization cfang@1607: bool IdealLoopTree::loop_predication( PhaseIdealLoop *phase) { cfang@1607: bool hoisted = false; cfang@1607: // Recursively promote predicates cfang@1607: if ( _child ) { cfang@1607: hoisted = _child->loop_predication( phase); cfang@1607: } cfang@1607: cfang@1607: // self cfang@1607: if (!_irreducible && !tail()->is_top()) { cfang@1607: hoisted |= phase->loop_predication_impl(this); cfang@1607: } cfang@1607: cfang@1607: if ( _next ) { //sibling cfang@1607: hoisted |= _next->loop_predication( phase); cfang@1607: } cfang@1607: cfang@1607: return hoisted; cfang@1607: } never@2118: never@2118: never@2118: // Process all the loops in the loop tree and replace any fill never@2118: // patterns with an intrisc version. never@2118: bool PhaseIdealLoop::do_intrinsify_fill() { never@2118: bool changed = false; never@2118: for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) { never@2118: IdealLoopTree* lpt = iter.current(); never@2118: changed |= intrinsify_fill(lpt); never@2118: } never@2118: return changed; never@2118: } never@2118: never@2118: never@2118: // Examine an inner loop looking for a a single store of an invariant never@2118: // value in a unit stride loop, never@2118: bool PhaseIdealLoop::match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value, never@2118: Node*& shift, Node*& con) { never@2118: const char* msg = NULL; never@2118: Node* msg_node = NULL; never@2118: never@2118: store_value = NULL; never@2118: con = NULL; never@2118: shift = NULL; never@2118: never@2118: // Process the loop looking for stores. If there are multiple never@2118: // stores or extra control flow give at this point. never@2118: CountedLoopNode* head = lpt->_head->as_CountedLoop(); never@2118: for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) { never@2118: Node* n = lpt->_body.at(i); never@2118: if (n->outcnt() == 0) continue; // Ignore dead never@2118: if (n->is_Store()) { never@2118: if (store != NULL) { never@2118: msg = "multiple stores"; never@2118: break; never@2118: } never@2118: int opc = n->Opcode(); never@2118: if (opc == Op_StoreP || opc == Op_StoreN || opc == Op_StoreCM) { never@2118: msg = "oop fills not handled"; never@2118: break; never@2118: } never@2118: Node* value = n->in(MemNode::ValueIn); never@2118: if (!lpt->is_invariant(value)) { never@2118: msg = "variant store value"; never@2140: } else if (!_igvn.type(n->in(MemNode::Address))->isa_aryptr()) { never@2140: msg = "not array address"; never@2118: } never@2118: store = n; never@2118: store_value = value; never@2118: } else if (n->is_If() && n != head->loopexit()) { never@2118: msg = "extra control flow"; never@2118: msg_node = n; never@2118: } never@2118: } never@2118: never@2118: if (store == NULL) { never@2118: // No store in loop never@2118: return false; never@2118: } never@2118: never@2118: if (msg == NULL && head->stride_con() != 1) { never@2118: // could handle negative strides too never@2118: if (head->stride_con() < 0) { never@2118: msg = "negative stride"; never@2118: } else { never@2118: msg = "non-unit stride"; never@2118: } never@2118: } never@2118: never@2118: if (msg == NULL && !store->in(MemNode::Address)->is_AddP()) { never@2118: msg = "can't handle store address"; never@2118: msg_node = store->in(MemNode::Address); never@2118: } never@2118: never@2168: if (msg == NULL && never@2168: (!store->in(MemNode::Memory)->is_Phi() || never@2168: store->in(MemNode::Memory)->in(LoopNode::LoopBackControl) != store)) { never@2168: msg = "store memory isn't proper phi"; never@2168: msg_node = store->in(MemNode::Memory); never@2168: } never@2168: never@2118: // Make sure there is an appropriate fill routine never@2118: BasicType t = store->as_Mem()->memory_type(); never@2118: const char* fill_name; never@2118: if (msg == NULL && never@2118: StubRoutines::select_fill_function(t, false, fill_name) == NULL) { never@2118: msg = "unsupported store"; never@2118: msg_node = store; never@2118: } never@2118: never@2118: if (msg != NULL) { never@2118: #ifndef PRODUCT never@2118: if (TraceOptimizeFill) { never@2118: tty->print_cr("not fill intrinsic candidate: %s", msg); never@2118: if (msg_node != NULL) msg_node->dump(); never@2118: } never@2118: #endif never@2118: return false; never@2118: } never@2118: never@2118: // Make sure the address expression can be handled. It should be never@2118: // head->phi * elsize + con. head->phi might have a ConvI2L. never@2118: Node* elements[4]; never@2118: Node* conv = NULL; never@2140: bool found_index = false; never@2118: int count = store->in(MemNode::Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements)); never@2118: for (int e = 0; e < count; e++) { never@2118: Node* n = elements[e]; never@2118: if (n->is_Con() && con == NULL) { never@2118: con = n; never@2118: } else if (n->Opcode() == Op_LShiftX && shift == NULL) { never@2118: Node* value = n->in(1); never@2118: #ifdef _LP64 never@2118: if (value->Opcode() == Op_ConvI2L) { never@2118: conv = value; never@2118: value = value->in(1); never@2118: } never@2118: #endif never@2118: if (value != head->phi()) { never@2118: msg = "unhandled shift in address"; never@2118: } else { never@2140: found_index = true; never@2118: shift = n; never@2118: assert(type2aelembytes(store->as_Mem()->memory_type(), true) == 1 << shift->in(2)->get_int(), "scale should match"); never@2118: } never@2118: } else if (n->Opcode() == Op_ConvI2L && conv == NULL) { never@2118: if (n->in(1) == head->phi()) { never@2140: found_index = true; never@2118: conv = n; never@2118: } else { never@2118: msg = "unhandled input to ConvI2L"; never@2118: } never@2118: } else if (n == head->phi()) { never@2118: // no shift, check below for allowed cases never@2140: found_index = true; never@2118: } else { never@2118: msg = "unhandled node in address"; never@2118: msg_node = n; never@2118: } never@2118: } never@2118: never@2118: if (count == -1) { never@2118: msg = "malformed address expression"; never@2118: msg_node = store; never@2118: } never@2118: never@2140: if (!found_index) { never@2140: msg = "missing use of index"; never@2140: } never@2140: never@2118: // byte sized items won't have a shift never@2118: if (msg == NULL && shift == NULL && t != T_BYTE && t != T_BOOLEAN) { never@2118: msg = "can't find shift"; never@2118: msg_node = store; never@2118: } never@2118: never@2118: if (msg != NULL) { never@2118: #ifndef PRODUCT never@2118: if (TraceOptimizeFill) { never@2118: tty->print_cr("not fill intrinsic: %s", msg); never@2118: if (msg_node != NULL) msg_node->dump(); never@2118: } never@2118: #endif never@2118: return false; never@2118: } never@2118: never@2118: // No make sure all the other nodes in the loop can be handled never@2118: VectorSet ok(Thread::current()->resource_area()); never@2118: never@2118: // store related values are ok never@2118: ok.set(store->_idx); never@2118: ok.set(store->in(MemNode::Memory)->_idx); never@2118: never@2118: // Loop structure is ok never@2118: ok.set(head->_idx); never@2118: ok.set(head->loopexit()->_idx); never@2118: ok.set(head->phi()->_idx); never@2118: ok.set(head->incr()->_idx); never@2118: ok.set(head->loopexit()->cmp_node()->_idx); never@2118: ok.set(head->loopexit()->in(1)->_idx); never@2118: never@2118: // Address elements are ok never@2118: if (con) ok.set(con->_idx); never@2118: if (shift) ok.set(shift->_idx); never@2118: if (conv) ok.set(conv->_idx); never@2118: never@2118: for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) { never@2118: Node* n = lpt->_body.at(i); never@2118: if (n->outcnt() == 0) continue; // Ignore dead never@2118: if (ok.test(n->_idx)) continue; never@2118: // Backedge projection is ok never@2118: if (n->is_IfTrue() && n->in(0) == head->loopexit()) continue; never@2118: if (!n->is_AddP()) { never@2118: msg = "unhandled node"; never@2118: msg_node = n; never@2118: break; never@2118: } never@2118: } never@2118: never@2118: // Make sure no unexpected values are used outside the loop never@2118: for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) { never@2118: Node* n = lpt->_body.at(i); never@2118: // These values can be replaced with other nodes if they are used never@2118: // outside the loop. never@2168: if (n == store || n == head->loopexit() || n == head->incr() || n == store->in(MemNode::Memory)) continue; never@2118: for (SimpleDUIterator iter(n); iter.has_next(); iter.next()) { never@2118: Node* use = iter.get(); never@2118: if (!lpt->_body.contains(use)) { never@2118: msg = "node is used outside loop"; never@2118: // lpt->_body.dump(); never@2118: msg_node = n; never@2118: break; never@2118: } never@2118: } never@2118: } never@2118: never@2118: #ifdef ASSERT never@2118: if (TraceOptimizeFill) { never@2118: if (msg != NULL) { never@2118: tty->print_cr("no fill intrinsic: %s", msg); never@2118: if (msg_node != NULL) msg_node->dump(); never@2118: } else { never@2118: tty->print_cr("fill intrinsic for:"); never@2118: } never@2118: store->dump(); never@2118: if (Verbose) { never@2118: lpt->_body.dump(); never@2118: } never@2118: } never@2118: #endif never@2118: never@2118: return msg == NULL; never@2118: } never@2118: never@2118: never@2118: never@2118: bool PhaseIdealLoop::intrinsify_fill(IdealLoopTree* lpt) { never@2118: // Only for counted inner loops never@2118: if (!lpt->is_counted() || !lpt->is_inner()) { never@2118: return false; never@2118: } never@2118: never@2118: // Must have constant stride never@2118: CountedLoopNode* head = lpt->_head->as_CountedLoop(); never@2118: if (!head->stride_is_con() || !head->is_normal_loop()) { never@2118: return false; never@2118: } never@2118: never@2118: // Check that the body only contains a store of a loop invariant never@2118: // value that is indexed by the loop phi. never@2118: Node* store = NULL; never@2118: Node* store_value = NULL; never@2118: Node* shift = NULL; never@2118: Node* offset = NULL; never@2118: if (!match_fill_loop(lpt, store, store_value, shift, offset)) { never@2118: return false; never@2118: } never@2118: never@2118: // Now replace the whole loop body by a call to a fill routine that never@2118: // covers the same region as the loop. never@2118: Node* base = store->in(MemNode::Address)->as_AddP()->in(AddPNode::Base); never@2118: never@2118: // Build an expression for the beginning of the copy region never@2118: Node* index = head->init_trip(); never@2118: #ifdef _LP64 never@2118: index = new (C, 2) ConvI2LNode(index); never@2118: _igvn.register_new_node_with_optimizer(index); never@2118: #endif never@2118: if (shift != NULL) { never@2118: // byte arrays don't require a shift but others do. never@2118: index = new (C, 3) LShiftXNode(index, shift->in(2)); never@2118: _igvn.register_new_node_with_optimizer(index); never@2118: } never@2118: index = new (C, 4) AddPNode(base, base, index); never@2118: _igvn.register_new_node_with_optimizer(index); never@2118: Node* from = new (C, 4) AddPNode(base, index, offset); never@2118: _igvn.register_new_node_with_optimizer(from); never@2118: // Compute the number of elements to copy never@2118: Node* len = new (C, 3) SubINode(head->limit(), head->init_trip()); never@2118: _igvn.register_new_node_with_optimizer(len); never@2118: never@2118: BasicType t = store->as_Mem()->memory_type(); never@2118: bool aligned = false; never@2118: if (offset != NULL && head->init_trip()->is_Con()) { never@2118: int element_size = type2aelembytes(t); never@2118: aligned = (offset->find_intptr_t_type()->get_con() + head->init_trip()->get_int() * element_size) % HeapWordSize == 0; never@2118: } never@2118: never@2118: // Build a call to the fill routine never@2118: const char* fill_name; never@2118: address fill = StubRoutines::select_fill_function(t, aligned, fill_name); never@2118: assert(fill != NULL, "what?"); never@2118: never@2118: // Convert float/double to int/long for fill routines never@2118: if (t == T_FLOAT) { never@2118: store_value = new (C, 2) MoveF2INode(store_value); never@2118: _igvn.register_new_node_with_optimizer(store_value); never@2118: } else if (t == T_DOUBLE) { never@2118: store_value = new (C, 2) MoveD2LNode(store_value); never@2118: _igvn.register_new_node_with_optimizer(store_value); never@2118: } never@2118: never@2118: Node* mem_phi = store->in(MemNode::Memory); never@2118: Node* result_ctrl; never@2118: Node* result_mem; never@2118: const TypeFunc* call_type = OptoRuntime::array_fill_Type(); never@2118: int size = call_type->domain()->cnt(); never@2118: CallLeafNode *call = new (C, size) CallLeafNoFPNode(call_type, fill, never@2118: fill_name, TypeAryPtr::get_array_body_type(t)); never@2118: call->init_req(TypeFunc::Parms+0, from); never@2118: call->init_req(TypeFunc::Parms+1, store_value); never@2199: #ifdef _LP64 never@2199: len = new (C, 2) ConvI2LNode(len); never@2199: _igvn.register_new_node_with_optimizer(len); never@2199: #endif never@2118: call->init_req(TypeFunc::Parms+2, len); never@2199: #ifdef _LP64 never@2199: call->init_req(TypeFunc::Parms+3, C->top()); never@2199: #endif never@2118: call->init_req( TypeFunc::Control, head->init_control()); never@2118: call->init_req( TypeFunc::I_O , C->top() ) ; // does no i/o never@2118: call->init_req( TypeFunc::Memory , mem_phi->in(LoopNode::EntryControl) ); never@2118: call->init_req( TypeFunc::ReturnAdr, C->start()->proj_out(TypeFunc::ReturnAdr) ); never@2118: call->init_req( TypeFunc::FramePtr, C->start()->proj_out(TypeFunc::FramePtr) ); never@2118: _igvn.register_new_node_with_optimizer(call); never@2118: result_ctrl = new (C, 1) ProjNode(call,TypeFunc::Control); never@2118: _igvn.register_new_node_with_optimizer(result_ctrl); never@2118: result_mem = new (C, 1) ProjNode(call,TypeFunc::Memory); never@2118: _igvn.register_new_node_with_optimizer(result_mem); never@2118: never@2118: // If this fill is tightly coupled to an allocation and overwrites never@2118: // the whole body, allow it to take over the zeroing. never@2118: AllocateNode* alloc = AllocateNode::Ideal_allocation(base, this); never@2118: if (alloc != NULL && alloc->is_AllocateArray()) { never@2118: Node* length = alloc->as_AllocateArray()->Ideal_length(); never@2118: if (head->limit() == length && never@2118: head->init_trip() == _igvn.intcon(0)) { never@2118: if (TraceOptimizeFill) { never@2118: tty->print_cr("Eliminated zeroing in allocation"); never@2118: } never@2118: alloc->maybe_set_complete(&_igvn); never@2118: } else { never@2118: #ifdef ASSERT never@2118: if (TraceOptimizeFill) { never@2118: tty->print_cr("filling array but bounds don't match"); never@2118: alloc->dump(); never@2118: head->init_trip()->dump(); never@2118: head->limit()->dump(); never@2118: length->dump(); never@2118: } never@2118: #endif never@2118: } never@2118: } never@2118: never@2118: // Redirect the old control and memory edges that are outside the loop. never@2118: Node* exit = head->loopexit()->proj_out(0); never@2168: // Sometimes the memory phi of the head is used as the outgoing never@2168: // state of the loop. It's safe in this case to replace it with the never@2168: // result_mem. never@2168: _igvn.replace_node(store->in(MemNode::Memory), result_mem); never@2118: _igvn.replace_node(exit, result_ctrl); never@2118: _igvn.replace_node(store, result_mem); never@2118: // Any uses the increment outside of the loop become the loop limit. never@2118: _igvn.replace_node(head->incr(), head->limit()); never@2118: never@2118: // Disconnect the head from the loop. never@2118: for (uint i = 0; i < lpt->_body.size(); i++) { never@2118: Node* n = lpt->_body.at(i); never@2118: _igvn.replace_node(n, C->top()); never@2118: } never@2118: never@2118: return true; never@2118: }