duke@435: /* fyang@9772: * Copyright (c) 2000, 2019, 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: stefank@2314: #include "precompiled.hpp" stefank@2314: #include "compiler/compileLog.hpp" stefank@2314: #include "memory/allocation.inline.hpp" stefank@2314: #include "opto/addnode.hpp" stefank@2314: #include "opto/callnode.hpp" stefank@2314: #include "opto/connode.hpp" stefank@2314: #include "opto/divnode.hpp" stefank@2314: #include "opto/loopnode.hpp" stefank@2314: #include "opto/mulnode.hpp" stefank@2314: #include "opto/rootnode.hpp" stefank@2314: #include "opto/runtime.hpp" stefank@2314: #include "opto/subnode.hpp" 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: kvn@2747: //------------------------------compute_exact_trip_count----------------------- kvn@2747: // Compute loop exact trip count if possible. Do not recalculate trip count for kvn@2747: // split loops (pre-main-post) which have their limits and inits behind Opaque node. kvn@2747: void IdealLoopTree::compute_exact_trip_count( PhaseIdealLoop *phase ) { kvn@2747: if (!_head->as_Loop()->is_valid_counted_loop()) { kvn@2747: return; kvn@2747: } kvn@2747: CountedLoopNode* cl = _head->as_CountedLoop(); kvn@2747: // Trip count may become nonexact for iteration split loops since kvn@2747: // RCE modifies limits. Note, _trip_count value is not reset since kvn@2747: // it is used to limit unrolling of main loop. kvn@2747: cl->set_nonexact_trip_count(); kvn@2747: kvn@2747: // Loop's test should be part of loop. kvn@2747: if (!phase->is_member(this, phase->get_ctrl(cl->loopexit()->in(CountedLoopEndNode::TestValue)))) kvn@2747: return; // Infinite loop kvn@2747: kvn@2747: #ifdef ASSERT kvn@2747: BoolTest::mask bt = cl->loopexit()->test_trip(); kvn@2747: assert(bt == BoolTest::lt || bt == BoolTest::gt || kvn@2979: bt == BoolTest::ne, "canonical test is expected"); kvn@2747: #endif kvn@2747: kvn@2747: Node* init_n = cl->init_trip(); kvn@2747: Node* limit_n = cl->limit(); kvn@2747: if (init_n != NULL && init_n->is_Con() && kvn@2747: limit_n != NULL && limit_n->is_Con()) { kvn@2747: // Use longs to avoid integer overflow. kvn@2747: int stride_con = cl->stride_con(); vlivanov@4157: jlong init_con = cl->init_trip()->get_int(); vlivanov@4157: jlong limit_con = cl->limit()->get_int(); kvn@2747: int stride_m = stride_con - (stride_con > 0 ? 1 : -1); vlivanov@4157: jlong trip_count = (limit_con - init_con + stride_m)/stride_con; kvn@2747: if (trip_count > 0 && (julong)trip_count < (julong)max_juint) { kvn@2747: // Set exact trip count. kvn@2747: cl->set_exact_trip_count((uint)trip_count); kvn@2747: } kvn@2747: } kvn@2747: } kvn@2747: 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()); kvn@4115: n_inv1 = new (phase->C) 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) { kvn@4115: inv = new (phase->C) SubINode(n_inv1, inv2); duke@435: } else { kvn@4115: inv = new (phase->C) 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) { kvn@4115: addx = new (phase->C) SubINode(inv, x); duke@435: } else { kvn@4115: addx = new (phase->C) AddINode(x, inv); duke@435: } duke@435: phase->register_new_node(addx, phase->get_ctrl(x)); kvn@1976: phase->_igvn.replace_node(n1, addx); kvn@2665: assert(phase->get_loop(phase->get_ctrl(n1)) == this, ""); kvn@2665: _body.yank(n1); 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: // 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 */ vlivanov@7385: || (body_size * body_size + phase->C->live_nodes()) > phase->C->max_node_limit() ) { 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. kvn@2727: // kvn@2727: // orig kvn@2727: // kvn@2727: // stmt1 kvn@2727: // | kvn@2727: // v kvn@2727: // loop predicate kvn@2727: // | kvn@2727: // v kvn@2727: // loop<----+ kvn@2727: // | | kvn@2727: // stmt2 | kvn@2727: // | | kvn@2727: // v | kvn@2727: // if ^ kvn@2727: // / \ | kvn@2727: // / \ | kvn@2727: // v v | kvn@2727: // false true | kvn@2727: // / \ | kvn@2727: // / ----+ kvn@2727: // | kvn@2727: // v kvn@2727: // exit kvn@2727: // kvn@2727: // kvn@2727: // after clone loop kvn@2727: // kvn@2727: // stmt1 kvn@2727: // | kvn@2727: // v kvn@2727: // loop predicate kvn@2727: // / \ kvn@2727: // clone / \ orig kvn@2727: // / \ kvn@2727: // / \ kvn@2727: // v v kvn@2727: // +---->loop clone loop<----+ kvn@2727: // | | | | kvn@2727: // | stmt2 clone stmt2 | kvn@2727: // | | | | kvn@2727: // | v v | kvn@2727: // ^ if clone If ^ kvn@2727: // | / \ / \ | kvn@2727: // | / \ / \ | kvn@2727: // | v v v v | kvn@2727: // | true false false true | kvn@2727: // | / \ / \ | kvn@2727: // +---- \ / ----+ kvn@2727: // \ / kvn@2727: // 1v v2 kvn@2727: // region kvn@2727: // | kvn@2727: // v kvn@2727: // exit kvn@2727: // kvn@2727: // kvn@2727: // after peel and predicate move kvn@2727: // kvn@2727: // stmt1 kvn@2727: // / kvn@2727: // / kvn@2727: // clone / orig kvn@2727: // / kvn@2727: // / +----------+ kvn@2727: // / | | kvn@2727: // / loop predicate | kvn@2727: // / | | kvn@2727: // v v | kvn@2727: // TOP-->loop clone loop<----+ | kvn@2727: // | | | | kvn@2727: // stmt2 clone stmt2 | | kvn@2727: // | | | ^ kvn@2727: // v v | | kvn@2727: // if clone If ^ | kvn@2727: // / \ / \ | | kvn@2727: // / \ / \ | | kvn@2727: // v v v v | | kvn@2727: // true false false true | | kvn@2727: // | \ / \ | | kvn@2727: // | \ / ----+ ^ kvn@2727: // | \ / | kvn@2727: // | 1v v2 | kvn@2727: // v region | kvn@2727: // | | | kvn@2727: // | v | kvn@2727: // | exit | kvn@2727: // | | kvn@2727: // +--------------->-----------------+ kvn@2727: // kvn@2727: // kvn@2727: // final graph kvn@2727: // kvn@2727: // stmt1 kvn@2727: // | kvn@2727: // v kvn@2727: // stmt2 clone kvn@2727: // | kvn@2727: // v kvn@2727: // if clone kvn@2727: // / | kvn@2727: // / | kvn@2727: // v v kvn@2727: // false true kvn@2727: // | | kvn@2727: // | v kvn@2727: // | loop predicate kvn@2727: // | | kvn@2727: // | v kvn@2727: // | loop<----+ kvn@2727: // | | | kvn@2727: // | stmt2 | kvn@2727: // | | | kvn@2727: // | v | kvn@2727: // v if ^ kvn@2727: // | / \ | kvn@2727: // | / \ | kvn@2727: // | v v | kvn@2727: // | false true | kvn@2727: // | | \ | kvn@2727: // v v --+ kvn@2727: // region kvn@2727: // | kvn@2727: // v kvn@2727: // exit kvn@2727: // 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. kvn@2665: #ifndef PRODUCT kvn@2665: if (TraceLoopOpts) { kvn@2665: tty->print("Peel "); kvn@2665: loop->dump_head(); kvn@2665: } kvn@2665: #endif kvn@2727: Node* head = loop->_head; kvn@2727: bool counted_loop = head->is_CountedLoop(); kvn@2727: if (counted_loop) { kvn@2727: CountedLoopNode *cl = head->as_CountedLoop(); duke@435: assert(cl->trip_count() > 0, "peeling a fully unrolled loop"); duke@435: cl->set_trip_count(cl->trip_count() - 1); kvn@2665: if (cl->is_main_loop()) { duke@435: cl->set_normal_loop(); duke@435: #ifndef PRODUCT kvn@2665: 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: } kvn@2727: Node* entry = head->in(LoopNode::EntryControl); 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). kvn@2727: clone_loop( loop, old_new, dom_depth(head) ); 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. kvn@3043: Node* new_entry = old_new[head->in(LoopNode::LoopBackControl)->_idx]; kvn@2727: _igvn.hash_delete(head); kvn@3043: head->set_req(LoopNode::EntryControl, new_entry); kvn@2727: for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) { kvn@2727: Node* old = head->fast_out(j); kvn@2727: if (old->in(0) == loop->_head && old->req() == 3 && old->is_Phi()) { kvn@3043: Node* new_exit_value = old_new[old->in(LoopNode::LoopBackControl)->_idx]; kvn@2727: 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. kvn@2727: Node* new_head = old_new[head->_idx]; kvn@2727: _igvn.hash_delete(new_head); kvn@2727: new_head->set_req(LoopNode::LoopBackControl, C->top()); kvn@2727: for (DUIterator_Fast j2max, j2 = new_head->fast_outs(j2max); j2 < j2max; j2++) { kvn@2727: Node* use = new_head->fast_out(j2); kvn@2727: if (use->in(0) == new_head && 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. kvn@2727: int dd = dom_depth(head); kvn@2727: set_idom(head, 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: } 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: kvn@2735: #define EMPTY_LOOP_SIZE 7 // number of nodes in an empty loop kvn@2735: duke@435: //------------------------------policy_maximally_unroll------------------------ kvn@2735: // Calculate exact loop trip count and return true if loop can be maximally kvn@2735: // unrolled. duke@435: bool IdealLoopTree::policy_maximally_unroll( PhaseIdealLoop *phase ) const { duke@435: CountedLoopNode *cl = _head->as_CountedLoop(); kvn@2694: assert(cl->is_normal_loop(), ""); kvn@2735: if (!cl->is_valid_counted_loop()) kvn@2735: return false; // Malformed counted loop duke@435: kvn@2747: if (!cl->has_exact_trip_count()) { kvn@2747: // Trip count is not exact. duke@435: return false; duke@435: } duke@435: kvn@2747: uint trip_count = cl->trip_count(); kvn@2747: // Note, max_juint is used to indicate unknown trip count. kvn@2747: assert(trip_count > 1, "one iteration loop should be optimized out already"); kvn@2747: assert(trip_count < max_juint, "exact trip_count should be less than max_uint."); 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"); kvn@2694: if (trip_count > unroll_limit || body_size > unroll_limit) { kvn@2694: return false; kvn@2694: } kvn@2694: kvn@2877: // Fully unroll a loop with few iterations regardless next kvn@2877: // conditions since following loop optimizations will split kvn@2877: // such loop anyway (pre-main-post). kvn@2877: if (trip_count <= 3) kvn@2877: return true; kvn@2877: kvn@2735: // Take into account that after unroll conjoined heads and tails will fold, kvn@2735: // otherwise policy_unroll() may allow more unrolling than max unrolling. kvn@2735: uint new_body_size = EMPTY_LOOP_SIZE + (body_size - EMPTY_LOOP_SIZE) * trip_count; kvn@2735: uint tst_body_size = (new_body_size - EMPTY_LOOP_SIZE) / trip_count + EMPTY_LOOP_SIZE; kvn@2735: if (body_size != tst_body_size) // Check for int overflow kvn@2735: return false; kvn@2735: if (new_body_size > unroll_limit || kvn@2735: // Unrolling can result in a large amount of node construction vlivanov@7385: new_body_size >= phase->C->max_node_limit() - phase->C->live_nodes()) { kvn@2735: return false; kvn@2735: } kvn@2735: kvn@2694: // Do not unroll a loop with String intrinsics code. kvn@2694: // String intrinsics are large and have loops. kvn@2694: for (uint k = 0; k < _body.size(); k++) { kvn@2694: Node* n = _body.at(k); kvn@2694: switch (n->Opcode()) { kvn@2694: case Op_StrComp: kvn@2694: case Op_StrEquals: kvn@2694: case Op_StrIndexOf: kvn@4479: case Op_EncodeISOArray: kvn@2694: case Op_AryEq: { kvn@2694: return false; kvn@2694: } kvn@6429: #if INCLUDE_RTM_OPT kvn@6429: case Op_FastLock: kvn@6429: case Op_FastUnlock: { kvn@6429: // Don't unroll RTM locking code because it is large. kvn@6429: if (UseRTMLocking) { kvn@6429: return false; kvn@6429: } kvn@6429: } kvn@6429: #endif kvn@2694: } // switch kvn@2694: } kvn@2694: kvn@2735: return true; // Do 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(); kvn@2694: assert(cl->is_normal_loop() || cl->is_main_loop(), ""); duke@435: kvn@2735: if (!cl->is_valid_counted_loop()) kvn@2735: return false; // Malformed counted loop duke@435: kvn@2877: // Protect against over-unrolling. kvn@2877: // After split at least one iteration will be executed in pre-loop. kvn@2877: if (cl->trip_count() <= (uint)(cl->is_normal_loop() ? 2 : 1)) return false; duke@435: kvn@2865: int future_unroll_ct = cl->unrolled_count() * 2; kvn@5513: if (future_unroll_ct > LoopMaxUnroll) return false; kvn@2735: kvn@2865: // Check for initial stride being a small enough constant kvn@2865: if (abs(cl->stride_con()) > (1<<2)*future_unroll_ct) return false; 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(); kvn@2877: int stride_con = cl->stride_con(); 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. kvn@2694: if (init_n == NULL || !init_n->is_Con() || kvn@2694: limit_n == NULL || !limit_n->is_Con()) { duke@435: Node* phi = cl->phi(); kvn@2694: 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(); kvn@2877: int next_stride = stride_con * 2; // stride after this unroll kvn@2694: if (next_stride > 0) { kvn@2694: if (iv_type->_lo + next_stride <= iv_type->_lo || // overflow kvn@2694: iv_type->_lo + next_stride > iv_type->_hi) { duke@435: return false; // over-unrolling duke@435: } kvn@2694: } else if (next_stride < 0) { kvn@2694: if (iv_type->_hi + next_stride >= iv_type->_hi || // overflow kvn@2694: iv_type->_hi + next_stride < iv_type->_lo) { duke@435: return false; // over-unrolling duke@435: } duke@435: } duke@435: } duke@435: } duke@435: kvn@2877: // After unroll limit will be adjusted: new_limit = limit-stride. kvn@2877: // Bailout if adjustment overflow. kvn@2877: const TypeInt* limit_type = phase->_igvn.type(limit_n)->is_int(); kvn@2877: if (stride_con > 0 && ((limit_type->_hi - stride_con) >= limit_type->_hi) || kvn@2877: stride_con < 0 && ((limit_type->_lo - stride_con) <= limit_type->_lo)) kvn@2877: return false; // overflow kvn@2877: duke@435: // Adjust body_size to determine if we unroll or not duke@435: uint body_size = _body.size(); kvn@3129: // Key test to unroll loop in CRC32 java code kvn@3129: int xors_in_loop = 0; duke@435: // Also count ModL, DivL and MulL which expand mightly kvn@2694: for (uint k = 0; k < _body.size(); k++) { kvn@2694: Node* n = _body.at(k); kvn@2694: switch (n->Opcode()) { kvn@3129: case Op_XorI: xors_in_loop++; break; // CRC32 java code kvn@2694: case Op_ModL: body_size += 30; break; kvn@2694: case Op_DivL: body_size += 30; break; kvn@2694: case Op_MulL: body_size += 10; break; kvn@2694: case Op_StrComp: kvn@2694: case Op_StrEquals: kvn@2694: case Op_StrIndexOf: kvn@4479: case Op_EncodeISOArray: kvn@2694: case Op_AryEq: { kvn@2694: // Do not unroll a loop with String intrinsics code. kvn@2694: // String intrinsics are large and have loops. kvn@2694: return false; kvn@2694: } kvn@6429: #if INCLUDE_RTM_OPT kvn@6429: case Op_FastLock: kvn@6429: case Op_FastUnlock: { kvn@6429: // Don't unroll RTM locking code because it is large. kvn@6429: if (UseRTMLocking) { kvn@6429: return false; kvn@6429: } kvn@6429: } kvn@6429: #endif kvn@2694: } // switch duke@435: } duke@435: duke@435: // Check for being too big kvn@2694: if (body_size > (uint)LoopUnrollLimit) { kvn@3129: if (xors_in_loop >= 4 && body_size < (uint)LoopUnrollLimit*4) return true; kvn@3129: // Normal case: loop too big duke@435: return false; duke@435: } 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 { kvn@2877: 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. kvn@2877: 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. kvn@2877: for (uint i = 0; i < _body.size(); i++) { duke@435: Node *iff = _body[i]; kvn@2877: if (iff->Opcode() == Op_If) { // Test? duke@435: duke@435: // Comparing trip+off vs limit duke@435: Node *bol = iff->in(1); kvn@2877: 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: } kvn@2877: if (bol->as_Bool()->_test._test == BoolTest::ne) kvn@2877: continue; // not RC kvn@2877: duke@435: Node *cmp = bol->in(1); 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. kvn@2985: Node *PhaseIdealLoop::clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n, VectorSet &visited, Node_Stack &clones ) { duke@435: if( get_ctrl(n) != back_ctrl ) return n; duke@435: kvn@2985: // Only visit once kvn@2985: if (visited.test_set(n->_idx)) { kvn@2985: Node *x = clones.find(n->_idx); kvn@2985: if (x != NULL) kvn@2985: return x; kvn@2985: return n; kvn@2985: } kvn@2985: 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 ) { kvn@2985: assert(clones.find(n->_idx) == NULL, "dead loop"); duke@435: x = n->clone(); // Clone a copy of 'n' to preheader kvn@2985: clones.push(x, n->_idx); 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++ ) { kvn@2985: Node *g = clone_up_backedge_goo( back_ctrl, preheader_ctrl, n->in(i), visited, clones ); duke@435: if( g != n->in(i) ) { kvn@2985: if( !x ) { kvn@2985: assert(clones.find(n->_idx) == NULL, "dead loop"); duke@435: x = n->clone(); kvn@2985: clones.push(x, n->_idx); kvn@2985: } 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: roland@7394: bool PhaseIdealLoop::cast_incr_before_loop(Node* incr, Node* ctrl, Node* loop) { roland@7394: Node* castii = new (C) CastIINode(incr, TypeInt::INT, true); roland@7394: castii->set_req(0, ctrl); roland@7394: register_new_node(castii, ctrl); roland@7394: for (DUIterator_Fast imax, i = incr->fast_outs(imax); i < imax; i++) { roland@7394: Node* n = incr->fast_out(i); roland@7394: if (n->is_Phi() && n->in(0) == loop) { roland@7394: int nrep = n->replace_edge(incr, castii); roland@7394: return true; roland@7394: } roland@7394: } roland@7394: return false; roland@7394: } roland@7394: 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: kvn@2665: #ifndef PRODUCT kvn@2665: if (TraceLoopOpts) { kvn@2665: if (peel_only) kvn@2665: tty->print("PeelMainPost "); kvn@2665: else kvn@2665: tty->print("PreMainPost "); kvn@2665: loop->dump_head(); kvn@2665: } kvn@2665: #endif 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(); morris@4779: guarantee(main_end != NULL, "no loop exit node"); 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. kvn@4115: IfFalseNode *new_main_exit = new (C) 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@4115: Node *zer_opaq = new (C) Opaque1Node(C, incr); kvn@4115: Node *zer_cmp = new (C) CmpINode( zer_opaq, limit ); kvn@4115: Node *zer_bol = new (C) 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 kvn@4115: IfNode *zer_iff = new (C) 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 kvn@3847: _igvn.replace_input_of(main_exit, 0, zer_iff); 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 kvn@4115: Node *zer_taken = new (C) 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: kvn@2985: Arena *a = Thread::current()->resource_area(); kvn@2985: VectorSet visited(a); kvn@2985: Node_Stack clones(a, main_head->back_control()->outcnt()); 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(), kvn@2985: main_phi->in(LoopNode::LoopBackControl), kvn@2985: visited, clones); 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, "" ); kvn@4115: IfFalseNode *new_pre_exit = new (C) 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@4115: Node *min_opaq = new (C) Opaque1Node(C, limit); kvn@4115: Node *min_cmp = new (C) CmpINode( pre_incr, min_opaq ); kvn@4115: Node *min_bol = new (C) 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@4115: IfNode *min_iff = new (C) 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 kvn@4115: Node *min_taken = new (C) 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: kvn@2985: visited.Clear(); kvn@2985: clones.clear(); 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(), kvn@2985: pre_phi->in(LoopNode::LoopBackControl), kvn@2985: visited, clones); duke@435: _igvn.hash_delete(main_phi); duke@435: main_phi->set_req( LoopNode::EntryControl, fallpre ); duke@435: } duke@435: } duke@435: roland@7394: // Nodes inside the loop may be control dependent on a predicate roland@7394: // that was moved before the preloop. If the back branch of the main roland@7394: // or post loops becomes dead, those nodes won't be dependent on the roland@7394: // test that guards that loop nest anymore which could lead to an roland@7394: // incorrect array access because it executes independently of the roland@7394: // test that was guarding the loop nest. We add a special CastII on roland@7394: // the if branch that enters the loop, between the input induction roland@7394: // variable value and the induction variable Phi to preserve correct roland@7394: // dependencies. roland@7394: roland@7394: // CastII for the post loop: roland@7394: bool inserted = cast_incr_before_loop(zer_opaq->in(1), zer_taken, post_head); roland@7394: assert(inserted, "no castII inserted"); roland@7394: roland@7394: // CastII for the main loop: roland@7394: inserted = cast_incr_before_loop(pre_incr, min_taken, main_head); roland@7394: assert(inserted, "no castII inserted"); roland@7394: 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, "" ); kvn@4115: Node *pre_limit = new (C) 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@4115: Node *pre_opaq = new (C) 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 > kvn@2979: // kvn@2979: // not-equal test is kept for post loop to handle case kvn@2979: // when init > limit when stride > 0 (and reverse). 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(); kvn@4115: BoolNode* new_bol0 = new (C) 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"); kvn@4115: BoolNode* new_bol1 = new (C) 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(); kvn@4115: BoolNode* new_bol2 = new (C) 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: kvn@2877: // Subtract a trip count for the pre-loop. kvn@2877: main_head->set_trip_count(main_head->trip_count() - 1); kvn@2877: 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); kvn@6630: loop->record_for_igvn(); 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 ) { kvn@2665: assert(LoopUnrollLimit, ""); kvn@2665: CountedLoopNode *loop_head = loop->_head->as_CountedLoop(); kvn@2665: CountedLoopEndNode *loop_end = loop_head->loopexit(); kvn@2665: assert(loop_end, ""); duke@435: #ifndef PRODUCT kvn@2665: if (PrintOpto && VerifyLoopOptimizations) { duke@435: tty->print("Unrolling "); duke@435: loop->dump_head(); kvn@2665: } else if (TraceLoopOpts) { kvn@2747: if (loop_head->trip_count() < (uint)LoopUnrollLimit) { kvn@2877: tty->print("Unroll %d(%2d) ", loop_head->unrolled_count()*2, loop_head->trip_count()); kvn@2735: } else { kvn@2877: tty->print("Unroll %d ", loop_head->unrolled_count()*2); kvn@2735: } kvn@2665: loop->dump_head(); duke@435: } duke@435: #endif 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(); kvn@2665: Node *stride = loop_head->stride(); duke@435: duke@435: Node *opaq = NULL; kvn@2877: if (adjust_min_trip) { // If not maximally unrolling, need adjustment kvn@2877: // Search for zero-trip guard. zmajo@9977: zmajo@9977: // Check the shape of the graph at the loop entry. If an inappropriate zmajo@9977: // graph shape is encountered, the compiler bails out loop unrolling; zmajo@9977: // compilation of the method will still succeed. zmajo@9977: if (!is_canonical_main_loop_entry(loop_head)) { zmajo@9977: return; zmajo@9977: } zmajo@9977: opaq = ctrl->in(0)->in(1)->in(1)->in(2); kvn@2877: // Zero-trip test uses an 'opaque' node which is not shared. kvn@2877: assert(opaq->outcnt() == 1 && opaq->in(1) == limit, ""); duke@435: } duke@435: duke@435: C->set_major_progress(); duke@435: kvn@2877: Node* new_limit = NULL; kvn@2877: if (UnrollLimitCheck) { kvn@2877: int stride_con = stride->get_int(); kvn@2877: int stride_p = (stride_con > 0) ? stride_con : -stride_con; kvn@2877: uint old_trip_count = loop_head->trip_count(); kvn@2877: // Verify that unroll policy result is still valid. kvn@2877: assert(old_trip_count > 1 && kvn@2877: (!adjust_min_trip || stride_p <= (1<<3)*loop_head->unrolled_count()), "sanity"); duke@435: kvn@2877: // Adjust loop limit to keep valid iterations number after unroll. kvn@2877: // Use (limit - stride) instead of (((limit - init)/stride) & (-2))*stride kvn@2877: // which may overflow. kvn@2877: if (!adjust_min_trip) { kvn@2877: assert(old_trip_count > 1 && (old_trip_count & 1) == 0, kvn@2877: "odd trip count for maximally unroll"); kvn@2877: // Don't need to adjust limit for maximally unroll since trip count is even. kvn@2877: } else if (loop_head->has_exact_trip_count() && init->is_Con()) { kvn@2877: // Loop's limit is constant. Loop's init could be constant when pre-loop kvn@2877: // become peeled iteration. vlivanov@4157: jlong init_con = init->get_int(); kvn@2877: // We can keep old loop limit if iterations count stays the same: kvn@2877: // old_trip_count == new_trip_count * 2 kvn@2877: // Note: since old_trip_count >= 2 then new_trip_count >= 1 kvn@2877: // so we also don't need to adjust zero trip test. vlivanov@4157: jlong limit_con = limit->get_int(); kvn@2877: // (stride_con*2) not overflow since stride_con <= 8. kvn@2877: int new_stride_con = stride_con * 2; kvn@2877: int stride_m = new_stride_con - (stride_con > 0 ? 1 : -1); vlivanov@4157: jlong trip_count = (limit_con - init_con + stride_m)/new_stride_con; kvn@2877: // New trip count should satisfy next conditions. kvn@2877: assert(trip_count > 0 && (julong)trip_count < (julong)max_juint/2, "sanity"); kvn@2877: uint new_trip_count = (uint)trip_count; kvn@2877: adjust_min_trip = (old_trip_count != new_trip_count*2); kvn@2877: } duke@435: kvn@2877: if (adjust_min_trip) { kvn@2877: // Step 2: Adjust the trip limit if it is called for. kvn@2877: // The adjustment amount is -stride. Need to make sure if the kvn@2877: // adjustment underflows or overflows, then the main loop is skipped. kvn@2877: Node* cmp = loop_end->cmp_node(); kvn@2877: assert(cmp->in(2) == limit, "sanity"); kvn@2877: assert(opaq != NULL && opaq->in(1) == limit, "sanity"); duke@435: kvn@2877: // Verify that policy_unroll result is still valid. kvn@2877: const TypeInt* limit_type = _igvn.type(limit)->is_int(); kvn@2877: assert(stride_con > 0 && ((limit_type->_hi - stride_con) < limit_type->_hi) || kvn@2877: stride_con < 0 && ((limit_type->_lo - stride_con) > limit_type->_lo), "sanity"); duke@435: kvn@2877: if (limit->is_Con()) { kvn@2877: // The check in policy_unroll and the assert above guarantee kvn@2877: // no underflow if limit is constant. kvn@2877: new_limit = _igvn.intcon(limit->get_int() - stride_con); kvn@2877: set_ctrl(new_limit, C->root()); kvn@2877: } else { kvn@2880: // Limit is not constant. kvn@2899: if (loop_head->unrolled_count() == 1) { // only for first unroll kvn@2880: // Separate limit by Opaque node in case it is an incremented kvn@2880: // variable from previous loop to avoid using pre-incremented kvn@2880: // value which could increase register pressure. kvn@2880: // Otherwise reorg_offsets() optimization will create a separate kvn@2880: // Opaque node for each use of trip-counter and as result kvn@2880: // zero trip guard limit will be different from loop limit. kvn@2880: assert(has_ctrl(opaq), "should have it"); kvn@2880: Node* opaq_ctrl = get_ctrl(opaq); kvn@4115: limit = new (C) Opaque2Node( C, limit ); kvn@2880: register_new_node( limit, opaq_ctrl ); kvn@2880: } aph@9610: if (stride_con > 0 && (java_subtract(limit_type->_lo, stride_con) < limit_type->_lo) || aph@9610: stride_con < 0 && (java_subtract(limit_type->_hi, stride_con) > limit_type->_hi)) { kvn@2877: // No underflow. kvn@4115: new_limit = new (C) SubINode(limit, stride); kvn@2877: } else { kvn@2877: // (limit - stride) may underflow. kvn@2877: // Clamp the adjustment value with MININT or MAXINT: kvn@2877: // kvn@2877: // new_limit = limit-stride kvn@2877: // if (stride > 0) kvn@2877: // new_limit = (limit < new_limit) ? MININT : new_limit; kvn@2877: // else kvn@2877: // new_limit = (limit > new_limit) ? MAXINT : new_limit; kvn@2877: // kvn@2877: BoolTest::mask bt = loop_end->test_trip(); kvn@2877: assert(bt == BoolTest::lt || bt == BoolTest::gt, "canonical test is expected"); kvn@2877: Node* adj_max = _igvn.intcon((stride_con > 0) ? min_jint : max_jint); kvn@2877: set_ctrl(adj_max, C->root()); kvn@2877: Node* old_limit = NULL; kvn@2877: Node* adj_limit = NULL; kvn@2877: Node* bol = limit->is_CMove() ? limit->in(CMoveNode::Condition) : NULL; kvn@2877: if (loop_head->unrolled_count() > 1 && kvn@2877: limit->is_CMove() && limit->Opcode() == Op_CMoveI && kvn@2877: limit->in(CMoveNode::IfTrue) == adj_max && kvn@2877: bol->as_Bool()->_test._test == bt && kvn@2877: bol->in(1)->Opcode() == Op_CmpI && kvn@2877: bol->in(1)->in(2) == limit->in(CMoveNode::IfFalse)) { kvn@2877: // Loop was unrolled before. kvn@2877: // Optimize the limit to avoid nested CMove: kvn@2877: // use original limit as old limit. kvn@2877: old_limit = bol->in(1)->in(1); kvn@2877: // Adjust previous adjusted limit. kvn@2877: adj_limit = limit->in(CMoveNode::IfFalse); kvn@4115: adj_limit = new (C) SubINode(adj_limit, stride); kvn@2877: } else { kvn@2877: old_limit = limit; kvn@4115: adj_limit = new (C) SubINode(limit, stride); kvn@2877: } kvn@2877: assert(old_limit != NULL && adj_limit != NULL, ""); kvn@2877: register_new_node( adj_limit, ctrl ); // adjust amount kvn@4115: Node* adj_cmp = new (C) CmpINode(old_limit, adj_limit); kvn@2877: register_new_node( adj_cmp, ctrl ); kvn@4115: Node* adj_bool = new (C) BoolNode(adj_cmp, bt); kvn@2877: register_new_node( adj_bool, ctrl ); kvn@4115: new_limit = new (C) CMoveINode(adj_bool, adj_limit, adj_max, TypeInt::INT); kvn@2877: } kvn@2877: register_new_node(new_limit, ctrl); kvn@2877: } kvn@2877: assert(new_limit != NULL, ""); kvn@2880: // Replace in loop test. kvn@2929: assert(loop_end->in(1)->in(1) == cmp, "sanity"); kvn@2929: if (cmp->outcnt() == 1 && loop_end->in(1)->outcnt() == 1) { kvn@2929: // Don't need to create new test since only one user. kvn@2929: _igvn.hash_delete(cmp); kvn@2929: cmp->set_req(2, new_limit); kvn@2929: } else { kvn@2929: // Create new test since it is shared. kvn@2929: Node* ctrl2 = loop_end->in(0); kvn@2929: Node* cmp2 = cmp->clone(); kvn@2929: cmp2->set_req(2, new_limit); kvn@2929: register_new_node(cmp2, ctrl2); kvn@2929: Node* bol2 = loop_end->in(1)->clone(); kvn@2929: bol2->set_req(1, cmp2); kvn@2929: register_new_node(bol2, ctrl2); kvn@2929: _igvn.hash_delete(loop_end); kvn@2929: loop_end->set_req(1, bol2); kvn@2929: } kvn@2880: // Step 3: Find the min-trip test guaranteed before a 'main' loop. kvn@2880: // Make it a 1-trip test (means at least 2 trips). kvn@2877: kvn@2880: // Guard test uses an 'opaque' node which is not shared. Hence I kvn@2880: // can edit it's inputs directly. Hammer in the new limit for the kvn@2880: // minimum-trip guard. kvn@2880: assert(opaq->outcnt() == 1, ""); kvn@2880: _igvn.hash_delete(opaq); kvn@2880: opaq->set_req(1, new_limit); kvn@2877: } kvn@2877: kvn@2877: // Adjust max trip count. The trip count is intentionally rounded kvn@2877: // down here (e.g. 15-> 7-> 3-> 1) because if we unwittingly over-unroll, kvn@2877: // the main, unrolled, part of the loop will never execute as it is protected kvn@2877: // by the min-trip test. See bug 4834191 for a case where we over-unrolled kvn@2877: // and later determined that part of the unrolled loop was dead. kvn@2877: loop_head->set_trip_count(old_trip_count / 2); kvn@2877: kvn@2877: // Double the count of original iterations in the unrolled loop body. kvn@2877: loop_head->double_unrolled_count(); kvn@2877: kvn@2877: } else { // LoopLimitCheck kvn@2877: kvn@2877: // Adjust max trip count. The trip count is intentionally rounded kvn@2877: // down here (e.g. 15-> 7-> 3-> 1) because if we unwittingly over-unroll, kvn@2877: // the main, unrolled, part of the loop will never execute as it is protected kvn@2877: // by the min-trip test. See bug 4834191 for a case where we over-unrolled kvn@2877: // and later determined that part of the unrolled loop was dead. kvn@2877: loop_head->set_trip_count(loop_head->trip_count() / 2); kvn@2877: kvn@2877: // Double the count of original iterations in the unrolled loop body. kvn@2877: loop_head->double_unrolled_count(); kvn@2877: kvn@2877: // ----------- kvn@2877: // Step 2: Cut back the trip counter for an unroll amount of 2. kvn@2877: // Loop will normally trip (limit - init)/stride_con. Since it's a kvn@2877: // CountedLoop this is exact (stride divides limit-init exactly). kvn@2877: // We are going to double the loop body, so we want to knock off any kvn@2877: // odd iteration: (trip_cnt & ~1). Then back compute a new limit. kvn@4115: Node *span = new (C) SubINode( limit, init ); kvn@2877: register_new_node( span, ctrl ); kvn@4115: Node *trip = new (C) DivINode( 0, span, stride ); kvn@2877: register_new_node( trip, ctrl ); kvn@2877: Node *mtwo = _igvn.intcon(-2); kvn@2877: set_ctrl(mtwo, C->root()); kvn@4115: Node *rond = new (C) AndINode( trip, mtwo ); kvn@2877: register_new_node( rond, ctrl ); kvn@4115: Node *spn2 = new (C) MulINode( rond, stride ); kvn@2877: register_new_node( spn2, ctrl ); kvn@4115: new_limit = new (C) AddINode( spn2, init ); kvn@2877: register_new_node( new_limit, ctrl ); kvn@2877: kvn@2877: // Hammer in the new limit kvn@2877: Node *ctrl2 = loop_end->in(0); kvn@4115: Node *cmp2 = new (C) CmpINode( loop_head->incr(), new_limit ); kvn@2877: register_new_node( cmp2, ctrl2 ); kvn@4115: Node *bol2 = new (C) BoolNode( cmp2, loop_end->test_trip() ); kvn@2877: register_new_node( bol2, ctrl2 ); kvn@2877: _igvn.hash_delete(loop_end); kvn@2877: loop_end->set_req(CountedLoopEndNode::TestValue, bol2); kvn@2877: kvn@2877: // Step 3: Find the min-trip test guaranteed before a 'main' loop. kvn@2877: // Make it a 1-trip test (means at least 2 trips). kvn@2877: if( adjust_min_trip ) { kvn@2877: assert( new_limit != NULL, "" ); kvn@2877: // Guard test uses an 'opaque' node which is not shared. Hence I kvn@2877: // can edit it's inputs directly. Hammer in the new limit for the kvn@2877: // minimum-trip guard. kvn@2877: assert( opaq->outcnt() == 1, "" ); kvn@2877: _igvn.hash_delete(opaq); kvn@2877: opaq->set_req(1, new_limit); kvn@2877: } kvn@2877: } // LoopLimitCheck 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(); kvn@2877: assert(cl->has_exact_trip_count(), "trip count is not exact"); kvn@2665: assert(cl->trip_count() > 0, ""); kvn@2665: #ifndef PRODUCT kvn@2665: if (TraceLoopOpts) { kvn@2665: tty->print("MaxUnroll %d ", cl->trip_count()); kvn@2665: loop->dump_head(); kvn@2665: } kvn@2665: #endif duke@435: duke@435: // If loop is tripping an odd number of times, peel odd iteration kvn@2665: if ((cl->trip_count() & 1) == 1) { kvn@2665: 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. kvn@2665: if (cl->trip_count() > 0) { kvn@2877: assert((cl->trip_count() & 1) == 0, "missed peeling"); kvn@2665: 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: kvn@2915: //------------------------------adjust_limit----------------------------------- kvn@2915: // Helper function for add_constraint(). roland@9739: Node* PhaseIdealLoop::adjust_limit(int stride_con, Node * scale, Node *offset, Node *rc_limit, Node *loop_limit, Node *pre_ctrl, bool round_up) { kvn@2915: // Compute "I :: (limit-offset)/scale" kvn@4115: Node *con = new (C) SubINode(rc_limit, offset); kvn@2915: register_new_node(con, pre_ctrl); kvn@4115: Node *X = new (C) DivINode(0, con, scale); kvn@2915: register_new_node(X, pre_ctrl); kvn@2915: roland@9739: // When the absolute value of scale is greater than one, the integer roland@9739: // division may round limit down so add one to the limit. roland@9739: if (round_up) { roland@9739: X = new (C) AddINode(X, _igvn.intcon(1)); roland@9739: register_new_node(X, pre_ctrl); roland@9739: } roland@9739: kvn@2915: // Adjust loop limit kvn@2915: loop_limit = (stride_con > 0) kvn@4115: ? (Node*)(new (C) MinINode(loop_limit, X)) kvn@4115: : (Node*)(new (C) MaxINode(loop_limit, X)); kvn@2915: register_new_node(loop_limit, pre_ctrl); kvn@2915: return loop_limit; kvn@2915: } kvn@2915: duke@435: //------------------------------add_constraint--------------------------------- kvn@2877: // Constrain the main loop iterations so the conditions: kvn@2877: // low_limit <= scale_con * I + offset < upper_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). kvn@2877: void PhaseIdealLoop::add_constraint( int stride_con, int scale_con, Node *offset, Node *low_limit, Node *upper_limit, Node *pre_ctrl, Node **pre_limit, Node **main_limit ) { 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. kvn@2915: kvn@2915: Node *scale = _igvn.intcon(scale_con); kvn@2915: set_ctrl(scale, C->root()); kvn@2915: kvn@2915: if ((stride_con^scale_con) >= 0) { // Use XOR to avoid overflow kvn@2877: // The overflow limit: scale*I+offset < upper_limit kvn@2877: // For main-loop compute kvn@2877: // ( if (scale > 0) /* and stride > 0 */ kvn@2877: // I < (upper_limit-offset)/scale kvn@2877: // else /* scale < 0 and stride < 0 */ kvn@2877: // I > (upper_limit-offset)/scale kvn@2877: // ) kvn@2877: // kvn@2915: // (upper_limit-offset) may overflow or underflow. kvn@2877: // But it is fine since main loop will either have kvn@2877: // less iterations or will be skipped in such case. roland@9739: *main_limit = adjust_limit(stride_con, scale, offset, upper_limit, *main_limit, pre_ctrl, false); duke@435: kvn@2877: // The underflow limit: low_limit <= scale*I+offset. kvn@2877: // For pre-loop compute kvn@2877: // NOT(scale*I+offset >= low_limit) kvn@2877: // scale*I+offset < low_limit kvn@2877: // ( if (scale > 0) /* and stride > 0 */ kvn@2877: // I < (low_limit-offset)/scale kvn@2877: // else /* scale < 0 and stride < 0 */ kvn@2877: // I > (low_limit-offset)/scale kvn@2877: // ) kvn@2877: kvn@2877: if (low_limit->get_int() == -max_jint) { kvn@2877: if (!RangeLimitCheck) return; kvn@2877: // We need this guard when scale*pre_limit+offset >= limit kvn@2915: // due to underflow. So we need execute pre-loop until kvn@2915: // scale*I+offset >= min_int. But (min_int-offset) will kvn@2915: // underflow when offset > 0 and X will be > original_limit kvn@2915: // when stride > 0. To avoid it we replace positive offset with 0. kvn@2915: // kvn@2915: // Also (min_int+1 == -max_int) is used instead of min_int here kvn@2915: // to avoid problem with scale == -1 (min_int/(-1) == min_int). kvn@2877: Node* shift = _igvn.intcon(31); kvn@2877: set_ctrl(shift, C->root()); kvn@4115: Node* sign = new (C) RShiftINode(offset, shift); kvn@2915: register_new_node(sign, pre_ctrl); kvn@4115: offset = new (C) AndINode(offset, sign); kvn@2877: register_new_node(offset, pre_ctrl); kvn@2877: } else { kvn@2877: assert(low_limit->get_int() == 0, "wrong low limit for range check"); kvn@2877: // The only problem we have here when offset == min_int kvn@2915: // since (0-min_int) == min_int. It may be fine for stride > 0 kvn@2915: // but for stride < 0 X will be < original_limit. To avoid it kvn@2915: // max(pre_limit, original_limit) is used in do_range_check(). kvn@2877: } kvn@2915: // Pass (-stride) to indicate pre_loop_cond = NOT(main_loop_cond); roland@9739: *pre_limit = adjust_limit((-stride_con), scale, offset, low_limit, *pre_limit, pre_ctrl, roland@9739: scale_con > 1 && stride_con > 0); kvn@2877: kvn@2877: } else { // stride_con*scale_con < 0 kvn@2877: // For negative stride*scale pre-loop checks for overflow and kvn@2877: // post-loop for underflow. kvn@2877: // kvn@2877: // The overflow limit: scale*I+offset < upper_limit kvn@2877: // For pre-loop compute kvn@2877: // NOT(scale*I+offset < upper_limit) kvn@2877: // scale*I+offset >= upper_limit kvn@2877: // scale*I+offset+1 > upper_limit kvn@2877: // ( if (scale < 0) /* and stride > 0 */ kvn@2877: // I < (upper_limit-(offset+1))/scale kvn@2915: // else /* scale > 0 and stride < 0 */ kvn@2877: // I > (upper_limit-(offset+1))/scale kvn@2877: // ) kvn@2915: // kvn@2915: // (upper_limit-offset-1) may underflow or overflow. kvn@2915: // To avoid it min(pre_limit, original_limit) is used kvn@2915: // in do_range_check() for stride > 0 and max() for < 0. kvn@2915: Node *one = _igvn.intcon(1); kvn@2915: set_ctrl(one, C->root()); kvn@2915: kvn@4115: Node *plus_one = new (C) AddINode(offset, one); kvn@2877: register_new_node( plus_one, pre_ctrl ); kvn@2915: // Pass (-stride) to indicate pre_loop_cond = NOT(main_loop_cond); roland@9739: *pre_limit = adjust_limit((-stride_con), scale, plus_one, upper_limit, *pre_limit, pre_ctrl, roland@9739: scale_con < -1 && stride_con > 0); kvn@2877: kvn@2915: if (low_limit->get_int() == -max_jint) { kvn@2915: if (!RangeLimitCheck) return; kvn@2915: // We need this guard when scale*main_limit+offset >= limit kvn@2915: // due to underflow. So we need execute main-loop while kvn@2915: // scale*I+offset+1 > min_int. But (min_int-offset-1) will kvn@2915: // underflow when (offset+1) > 0 and X will be < main_limit kvn@2915: // when scale < 0 (and stride > 0). To avoid it we replace kvn@2915: // positive (offset+1) with 0. kvn@2915: // kvn@2915: // Also (min_int+1 == -max_int) is used instead of min_int here kvn@2915: // to avoid problem with scale == -1 (min_int/(-1) == min_int). kvn@2915: Node* shift = _igvn.intcon(31); kvn@2915: set_ctrl(shift, C->root()); kvn@4115: Node* sign = new (C) RShiftINode(plus_one, shift); kvn@2915: register_new_node(sign, pre_ctrl); kvn@4115: plus_one = new (C) AndINode(plus_one, sign); kvn@2915: register_new_node(plus_one, pre_ctrl); kvn@2915: } else { kvn@2915: assert(low_limit->get_int() == 0, "wrong low limit for range check"); kvn@2915: // The only problem we have here when offset == max_int kvn@2915: // since (max_int+1) == min_int and (0-min_int) == min_int. kvn@2915: // But it is fine since main loop will either have kvn@2915: // less iterations or will be skipped in such case. kvn@2915: } kvn@2915: // The underflow limit: low_limit <= scale*I+offset. kvn@2915: // For main-loop compute kvn@2915: // scale*I+offset+1 > low_limit kvn@2915: // ( if (scale < 0) /* and stride > 0 */ kvn@2915: // I < (low_limit-(offset+1))/scale kvn@2915: // else /* scale > 0 and stride < 0 */ kvn@2915: // I > (low_limit-(offset+1))/scale kvn@2915: // ) kvn@2877: roland@9739: *main_limit = adjust_limit(stride_con, scale, plus_one, low_limit, *main_limit, pre_ctrl, roland@9739: false); 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: } thartmann@8207: if (is_scaled_iv(exp->in(2), iv, p_scale)) { thartmann@8207: if (p_offset != NULL) { thartmann@8207: *p_offset = exp->in(1); thartmann@8207: } thartmann@8207: return true; thartmann@8207: } 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); kvn@4115: Node* offset = new (C) 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)); kvn@4115: Node* offset = new (C) 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 kvn@2665: if (PrintOpto && VerifyLoopOptimizations) { duke@435: tty->print("Range Check Elimination "); duke@435: loop->dump_head(); kvn@2665: } else if (TraceLoopOpts) { kvn@2665: tty->print("RangeCheck "); kvn@2665: loop->dump_head(); duke@435: } duke@435: #endif kvn@2665: assert(RangeCheckElimination, ""); duke@435: CountedLoopNode *cl = loop->_head->as_CountedLoop(); kvn@2665: kvn@2665: // protect against stride not being a constant kvn@2665: if (!cl->stride_is_con()) kvn@2665: return; 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(); kvn@2665: zmajo@9977: // Check graph shape. Cannot optimize a loop if zero-trip zmajo@9977: // Opaque1 node is optimized away and then another round zmajo@9977: // of loop opts attempted. zmajo@9977: if (!is_canonical_main_loop_entry(cl)) { zmajo@9977: return; zmajo@9977: } zmajo@9977: kvn@2665: // Need to find the main-loop zero-trip guard kvn@2665: Node *ctrl = cl->in(LoopNode::EntryControl); kvn@2665: Node *iffm = ctrl->in(0); zmajo@9977: Node *opqzm = iffm->in(1)->in(1)->in(2); kvn@2665: assert(opqzm->in(1) == main_limit, "do not understand situation"); kvn@2665: duke@435: // Find the pre-loop limit; we will expand it's iterations to duke@435: // not ever trip low tests. duke@435: Node *p_f = iffm->in(0); roland@7869: // pre loop may have been optimized out roland@7869: if (p_f->Opcode() != Op_IfFalse) { roland@7869: return; roland@7869: } duke@435: CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd(); kvn@2665: 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. kvn@2665: 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(); kvn@2665: if (orig_limit == NULL || _igvn.type(orig_limit) == Type::TOP) duke@435: return; duke@435: duke@435: // Must know if its a count-up or count-down loop duke@435: duke@435: int stride_con = cl->stride_con(); duke@435: Node *zero = _igvn.intcon(0); duke@435: Node *one = _igvn.intcon(1); kvn@2877: // Use symmetrical int range [-max_jint,max_jint] kvn@2877: Node *mini = _igvn.intcon(-max_jint); duke@435: set_ctrl(zero, C->root()); duke@435: set_ctrl(one, C->root()); kvn@2877: set_ctrl(mini, 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: } kvn@2877: #ifdef ASSERT kvn@2877: if (TraceRangeLimitCheck) { kvn@2877: tty->print_cr("RC bool node%s", flip ? " flipped:" : ":"); kvn@2877: bol->dump(2); kvn@2877: } kvn@2877: #endif 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 kvn@2877: // The underflow and overflow limits: 0 <= scale*I+offset < limit kvn@2877: add_constraint( stride_con, scale_con, offset, zero, limit, pre_ctrl, &pre_limit, &main_limit ); duke@435: if (!conditional_rc) { kvn@2915: // (0-offset)/scale could be outside of loop iterations range. kvn@2915: conditional_rc = !loop->dominates_backedge(iff) || RangeLimitCheck; 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 ) { kvn@2877: case BoolTest::gt: kvn@2877: // Fall into GE case kvn@2877: case BoolTest::ge: kvn@2877: // Convert (I*scale+offset) >= Limit to (I*(-scale)+(-offset)) <= -Limit duke@435: scale_con = -scale_con; kvn@4115: offset = new (C) SubINode( zero, offset ); duke@435: register_new_node( offset, pre_ctrl ); kvn@4115: limit = new (C) SubINode( zero, limit ); duke@435: register_new_node( limit, pre_ctrl ); duke@435: // Fall into LE case kvn@2877: case BoolTest::le: kvn@2877: if (b_test._test != BoolTest::gt) { kvn@2877: // Convert X <= Y to X < Y+1 kvn@4115: limit = new (C) AddINode( limit, one ); kvn@2877: register_new_node( limit, pre_ctrl ); kvn@2877: } duke@435: // Fall into LT case duke@435: case BoolTest::lt: kvn@2877: // The underflow and overflow limits: MIN_INT <= scale*I+offset < limit kvn@2915: // Note: (MIN_INT+1 == -MAX_INT) is used instead of MIN_INT here kvn@2915: // to avoid problem with scale == -1: MIN_INT/(-1) == MIN_INT. kvn@2877: add_constraint( stride_con, scale_con, offset, mini, limit, pre_ctrl, &pre_limit, &main_limit ); duke@435: if (!conditional_rc) { kvn@2915: // ((MIN_INT+1)-offset)/scale could be outside of loop iterations range. kvn@2915: // Note: negative offset is replaced with 0 but (MIN_INT+1)/scale could kvn@2915: // still be outside of loop range. kvn@2915: conditional_rc = !loop->dominates_backedge(iff) || RangeLimitCheck; 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()); kvn@3847: _igvn.replace_input_of(iff, 1, kill_con); 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 iveresov@6070: if (cd->is_Load() && cd->depends_only_on_test()) { // Loads can now float around in the loop duke@435: // Allow the load to float around in the loop, or before it duke@435: // but NOT before the pre-loop. kvn@3847: _igvn.replace_input_of(cd, 0, ctrl); // ctrl, not NULL 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) { kvn@4115: pre_limit = (stride_con > 0) ? (Node*)new (C) MinINode(pre_limit, orig_limit) kvn@4115: : (Node*)new (C) 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. kvn@2877: cl->set_nonexact_trip_count(); kvn@2877: if (!LoopLimitCheck && 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(); kvn@4115: Node *span = new (C) SubINode(main_limit,init); duke@435: register_new_node(span,ctrl); duke@435: Node *rndup = _igvn.intcon(stride_con + ((stride_con>0)?-1:1)); kvn@4115: Node *add = new (C) AddINode(span,rndup); duke@435: register_new_node(add,ctrl); kvn@4115: Node *div = new (C) DivINode(0,add,stride); duke@435: register_new_node(div,ctrl); kvn@4115: Node *mul = new (C) MulINode(div,stride); duke@435: register_new_node(mul,ctrl); kvn@4115: Node *newlim = new (C) 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 kvn@3847: _igvn.replace_input_of(main_cmp, 2, main_limit); duke@435: // The OpaqueNode is unshared by design duke@435: assert( opqzm->outcnt() == 1, "cannot hack shared node" ); kvn@3847: _igvn.replace_input_of(opqzm, 1, main_limit); 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 kvn@2735: if (_body.size() > EMPTY_LOOP_SIZE) kvn@2665: return false; duke@435: kvn@2665: if (!_head->is_CountedLoop()) kvn@2665: return false; // Dead loop duke@435: CountedLoopNode *cl = _head->as_CountedLoop(); kvn@3048: if (!cl->is_valid_counted_loop()) kvn@2665: return false; // Malformed loop kvn@2665: if (!phase->is_member(this, phase->get_ctrl(cl->loopexit()->in(CountedLoopEndNode::TestValue)))) duke@435: return false; // Infinite loop never@2685: 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 never@2685: never@2685: // main and post loops have explicitly created zero trip guard never@2685: bool needs_guard = !cl->is_main_loop() && !cl->is_post_loop(); never@2685: if (needs_guard) { kvn@2747: // Skip guard if values not overlap. kvn@2747: const TypeInt* init_t = phase->_igvn.type(cl->init_trip())->is_int(); kvn@2747: const TypeInt* limit_t = phase->_igvn.type(cl->limit())->is_int(); kvn@2747: int stride_con = cl->stride_con(); kvn@2747: if (stride_con > 0) { kvn@2747: needs_guard = (init_t->_hi >= limit_t->_lo); kvn@2747: } else { kvn@2747: needs_guard = (init_t->_lo <= limit_t->_hi); kvn@2747: } kvn@2747: } kvn@2747: if (needs_guard) { never@2685: // Check for an obvious zero trip guard. kvn@2727: Node* inctrl = PhaseIdealLoop::skip_loop_predicates(cl->in(LoopNode::EntryControl)); never@2685: if (inctrl->Opcode() == Op_IfTrue) { never@2685: // The test should look like just the backedge of a CountedLoop never@2685: Node* iff = inctrl->in(0); never@2685: if (iff->is_If()) { never@2685: Node* bol = iff->in(1); never@2685: if (bol->is_Bool() && bol->as_Bool()->_test._test == cl->loopexit()->test_trip()) { never@2685: Node* cmp = bol->in(1); never@2685: if (cmp->is_Cmp() && cmp->in(1) == cl->init_trip() && cmp->in(2) == cl->limit()) { never@2685: needs_guard = false; never@2685: } never@2685: } never@2685: } never@2685: } never@2685: } never@2685: never@2685: #ifndef PRODUCT never@2685: if (PrintOpto) { never@2685: tty->print("Removing empty loop with%s zero trip guard", needs_guard ? "out" : ""); never@2685: this->dump_head(); never@2685: } else if (TraceLoopOpts) { never@2685: tty->print("Empty with%s zero trip guard ", needs_guard ? "out" : ""); never@2685: this->dump_head(); never@2685: } never@2685: #endif never@2685: never@2685: if (needs_guard) { never@2685: // Peel the loop to ensure there's a zero trip guard never@2685: Node_List old_new; never@2685: phase->do_peeling(this, old_new); never@2685: } never@2685: 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(); kvn@2877: Node *exact_limit = phase->exact_limit(this); kvn@2877: if (exact_limit != cl->limit()) { kvn@2877: // We also need to replace the original limit to collapse loop exit. kvn@2877: Node* cmp = cl->loopexit()->cmp_node(); kvn@2877: assert(cl->limit() == cmp->in(2), "sanity"); fyang@9772: // Duplicate cmp node if it has other users fyang@9772: if (cmp->outcnt() > 1) { fyang@9772: cmp = cmp->clone(); fyang@9772: cmp = phase->_igvn.register_new_node_with_optimizer(cmp); fyang@9772: BoolNode *bol = cl->loopexit()->in(CountedLoopEndNode::TestValue)->as_Bool(); fyang@9772: phase->_igvn.replace_input_of(bol, 1, cmp); // put bol on worklist fyang@9772: } kvn@2877: phase->_igvn._worklist.push(cmp->in(2)); // put limit on worklist kvn@3847: phase->_igvn.replace_input_of(cmp, 2, exact_limit); // put cmp on worklist kvn@2877: } kvn@2877: // Note: the final value after increment should not overflow since kvn@2877: // counted loop has limit check predicate. kvn@4115: Node *final = new (phase->C) SubINode( exact_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: kvn@2747: //------------------------------policy_do_one_iteration_loop------------------- kvn@2747: // Convert one iteration loop into normal code. kvn@2747: bool IdealLoopTree::policy_do_one_iteration_loop( PhaseIdealLoop *phase ) { kvn@2747: if (!_head->as_Loop()->is_valid_counted_loop()) kvn@2747: return false; // Only for counted loop kvn@2747: kvn@2747: CountedLoopNode *cl = _head->as_CountedLoop(); kvn@2747: if (!cl->has_exact_trip_count() || cl->trip_count() != 1) { kvn@2747: return false; kvn@2747: } kvn@2747: kvn@2747: #ifndef PRODUCT kvn@2747: if(TraceLoopOpts) { kvn@2747: tty->print("OneIteration "); kvn@2747: this->dump_head(); kvn@2747: } kvn@2747: #endif kvn@2747: kvn@2747: Node *init_n = cl->init_trip(); kvn@2747: #ifdef ASSERT kvn@2747: // Loop boundaries should be constant since trip count is exact. kvn@2747: assert(init_n->get_int() + cl->stride_con() >= cl->limit()->get_int(), "should be one iteration"); kvn@2747: #endif kvn@2747: // Replace the phi at loop head with the value of the init_trip. kvn@2747: // Then the CountedLoopEnd will collapse (backedge will not be taken) kvn@2747: // and all loop-invariant uses of the exit values will be correct. kvn@2747: phase->_igvn.replace_node(cl->phi(), cl->init_trip()); kvn@2747: phase->C->set_major_progress(); kvn@2747: return true; kvn@2747: } duke@435: duke@435: //============================================================================= duke@435: //------------------------------iteration_split_impl--------------------------- never@836: bool IdealLoopTree::iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new ) { kvn@2747: // Compute exact loop trip count if possible. kvn@2747: compute_exact_trip_count(phase); kvn@2747: kvn@2747: // Convert one iteration loop into normal code. kvn@2747: if (policy_do_one_iteration_loop(phase)) kvn@2747: return true; kvn@2747: duke@435: // Check and remove empty loops (spam micro-benchmarks) kvn@2747: 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). kvn@2747: 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: } kvn@2747: 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); kvn@2747: } 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: kvn@3048: if (!cl->is_valid_counted_loop()) return true; // Ignore various kinds of broken loops duke@435: duke@435: // Do nothing special to pre- and post- loops kvn@2747: 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. kvn@2747: 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); kvn@2747: 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: kvn@2735: // Skip next optimizations if running low on nodes. Note that kvn@2735: // policy_unswitching and policy_maximally_unroll have this check. vlivanov@7385: int nodes_left = phase->C->max_node_limit() - phase->C->live_nodes(); vlivanov@7385: if ((int)(2 * _body.size()) > nodes_left) { kvn@2735: return true; kvn@2735: } 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. kvn@2747: if (should_rce || should_align || should_unroll) { kvn@2747: 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. kvn@2747: 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. kvn@2747: if (should_unroll && !should_peel) kvn@2747: phase->do_unroll(this,old_new, true); duke@435: duke@435: // Adjust the pre-loop limits to align the main body duke@435: // iterations. kvn@2747: if (should_align) duke@435: Unimplemented(); duke@435: duke@435: } else { // Else we have an unchanged counted loop kvn@2747: 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 kvn@2665: 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. kvn@2665: if (_parent /*not the root loop*/ && duke@435: !_irreducible && duke@435: // Also ignore the occasional dead backedge kvn@2665: !tail()->is_top()) { duke@435: adjust_loop_exit_prob(phase); duke@435: } duke@435: duke@435: // Gate unrolling, RCE and peeling efforts. kvn@2665: if (!_child && // If not an inner loop, do not split duke@435: !_irreducible && kvn@474: _allow_optimizations && kvn@2665: !tail()->is_top()) { // Also ignore the occasional dead backedge duke@435: if (!_has_call) { kvn@2665: 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 kvn@2665: // trip counter when there was no major reshaping. kvn@2665: phase->reorg_offsets(this); kvn@2665: kvn@2665: if (_next && !_next->iteration_split(phase, old_new)) never@836: return false; never@836: return true; duke@435: } cfang@1607: cfang@1607: kvn@2727: //============================================================================= never@2118: // Process all the loops in the loop tree and replace any fill thartmann@8476: // patterns with an intrinsic 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(); roland@4159: if (opc == Op_StoreP || opc == Op_StoreN || opc == Op_StoreNKlass || 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 thartmann@8476: // head->phi * elsize + con. head->phi might have a ConvI2L(CastII()). never@2118: Node* elements[4]; thartmann@8476: Node* cast = NULL; 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: } thartmann@8476: if (value->Opcode() == Op_CastII && thartmann@8476: value->as_CastII()->has_range_check()) { thartmann@8476: // Skip range check dependent CastII nodes thartmann@8476: cast = value; thartmann@8476: value = value->in(1); thartmann@8476: } never@2118: #endif never@2118: if (value != head->phi()) { never@2118: msg = "unhandled shift in address"; never@2118: } else { never@2730: if (type2aelembytes(store->as_Mem()->memory_type(), true) != (1 << n->in(2)->get_int())) { never@2730: msg = "scale doesn't match"; never@2730: } else { never@2730: found_index = true; never@2730: shift = n; never@2730: } never@2118: } never@2118: } else if (n->Opcode() == Op_ConvI2L && conv == NULL) { thartmann@8476: conv = n; thartmann@8476: n = n->in(1); thartmann@8476: if (n->Opcode() == Op_CastII && thartmann@8476: n->as_CastII()->has_range_check()) { thartmann@8476: // Skip range check dependent CastII nodes thartmann@8476: cast = n; thartmann@8476: n = n->in(1); thartmann@8476: } thartmann@8476: if (n == head->phi()) { never@2140: found_index = true; 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: morris@4779: CountedLoopEndNode* loop_exit = head->loopexit(); morris@4779: guarantee(loop_exit != NULL, "no loop exit node"); morris@4779: never@2118: // Loop structure is ok never@2118: ok.set(head->_idx); morris@4779: ok.set(loop_exit->_idx); never@2118: ok.set(head->phi()->_idx); never@2118: ok.set(head->incr()->_idx); morris@4779: ok.set(loop_exit->cmp_node()->_idx); morris@4779: ok.set(loop_exit->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); thartmann@8476: if (cast) ok.set(cast->_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 morris@4779: if (n->is_IfTrue() && n->in(0) == loop_exit) 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. morris@4779: if (n == store || n == loop_exit || 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(); kvn@3048: if (!head->is_valid_counted_loop() || !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: rraghavan@8777: Node* exit = head->loopexit()->proj_out(0); rraghavan@8777: if (exit == NULL) { rraghavan@8777: return false; rraghavan@8777: } rraghavan@8777: kvn@2727: #ifndef PRODUCT kvn@2727: if (TraceLoopOpts) { kvn@2727: tty->print("ArrayFill "); kvn@2727: lpt->dump_head(); kvn@2727: } kvn@2727: #endif kvn@2727: 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 kvn@4115: index = new (C) 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. kvn@4115: index = new (C) LShiftXNode(index, shift->in(2)); never@2118: _igvn.register_new_node_with_optimizer(index); never@2118: } kvn@4115: index = new (C) AddPNode(base, base, index); never@2118: _igvn.register_new_node_with_optimizer(index); kvn@4115: Node* from = new (C) AddPNode(base, index, offset); never@2118: _igvn.register_new_node_with_optimizer(from); never@2118: // Compute the number of elements to copy kvn@4115: Node* len = new (C) 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) { kvn@4115: store_value = new (C) MoveF2INode(store_value); never@2118: _igvn.register_new_node_with_optimizer(store_value); never@2118: } else if (t == T_DOUBLE) { kvn@4115: store_value = new (C) MoveD2LNode(store_value); never@2118: _igvn.register_new_node_with_optimizer(store_value); never@2118: } never@2118: goetz@6468: if (CCallingConventionRequiresIntsAsLongs && goetz@6468: // See StubRoutines::select_fill_function for types. FLOAT has been converted to INT. goetz@6468: (t == T_FLOAT || t == T_INT || is_subword_type(t))) { goetz@6468: store_value = new (C) ConvI2LNode(store_value); goetz@6468: _igvn.register_new_node_with_optimizer(store_value); goetz@6468: } goetz@6468: 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(); kvn@4115: CallLeafNode *call = new (C) CallLeafNoFPNode(call_type, fill, kvn@4115: fill_name, TypeAryPtr::get_array_body_type(t)); goetz@6468: uint cnt = 0; goetz@6468: call->init_req(TypeFunc::Parms + cnt++, from); goetz@6468: call->init_req(TypeFunc::Parms + cnt++, store_value); goetz@6468: if (CCallingConventionRequiresIntsAsLongs) { goetz@6468: call->init_req(TypeFunc::Parms + cnt++, C->top()); goetz@6468: } never@2199: #ifdef _LP64 kvn@4115: len = new (C) ConvI2LNode(len); never@2199: _igvn.register_new_node_with_optimizer(len); never@2199: #endif goetz@6468: call->init_req(TypeFunc::Parms + cnt++, len); never@2199: #ifdef _LP64 goetz@6468: call->init_req(TypeFunc::Parms + cnt++, C->top()); never@2199: #endif goetz@6468: call->init_req(TypeFunc::Control, head->init_control()); goetz@6468: call->init_req(TypeFunc::I_O, C->top()); // Does no I/O. goetz@6468: call->init_req(TypeFunc::Memory, mem_phi->in(LoopNode::EntryControl)); goetz@6468: call->init_req(TypeFunc::ReturnAdr, C->start()->proj_out(TypeFunc::ReturnAdr)); goetz@6468: call->init_req(TypeFunc::FramePtr, C->start()->proj_out(TypeFunc::FramePtr)); never@2118: _igvn.register_new_node_with_optimizer(call); kvn@4115: result_ctrl = new (C) ProjNode(call,TypeFunc::Control); never@2118: _igvn.register_new_node_with_optimizer(result_ctrl); kvn@4115: result_mem = new (C) ProjNode(call,TypeFunc::Memory); never@2118: _igvn.register_new_node_with_optimizer(result_mem); never@2118: kvn@4217: /* Disable following optimization until proper fix (add missing checks). kvn@4217: 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: } kvn@4217: */ never@2118: never@2118: // Redirect the old control and memory edges that are outside the loop. 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); roland@8311: lazy_replace(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: }