annotate src/share/vm/opto/loopTransform.cpp @ 5635:94a83e0f9ce1

8017065: C2 allows safepoint checks to leak into G1 pre-barriers Summary: Make all raw loads strictly respect control dependencies, make sure RCE doesn't move raw loads, add verification of G1 pre-barriers. Reviewed-by: kvn, roland
author iveresov
date Tue, 05 Nov 2013 01:57:18 -0800
parents 60a32bb8ff99
children 55fb97c4c58d
rev   line source
duke@0 1 /*
mikael@3718 2 * Copyright (c) 2000, 2012, Oracle and/or its affiliates. All rights reserved.
duke@0 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
duke@0 4 *
duke@0 5 * This code is free software; you can redistribute it and/or modify it
duke@0 6 * under the terms of the GNU General Public License version 2 only, as
duke@0 7 * published by the Free Software Foundation.
duke@0 8 *
duke@0 9 * This code is distributed in the hope that it will be useful, but WITHOUT
duke@0 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
duke@0 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
duke@0 12 * version 2 for more details (a copy is included in the LICENSE file that
duke@0 13 * accompanied this code).
duke@0 14 *
duke@0 15 * You should have received a copy of the GNU General Public License version
duke@0 16 * 2 along with this work; if not, write to the Free Software Foundation,
duke@0 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
duke@0 18 *
trims@1472 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
trims@1472 20 * or visit www.oracle.com if you need additional information or have any
trims@1472 21 * questions.
duke@0 22 *
duke@0 23 */
duke@0 24
stefank@1879 25 #include "precompiled.hpp"
stefank@1879 26 #include "compiler/compileLog.hpp"
stefank@1879 27 #include "memory/allocation.inline.hpp"
stefank@1879 28 #include "opto/addnode.hpp"
stefank@1879 29 #include "opto/callnode.hpp"
stefank@1879 30 #include "opto/connode.hpp"
stefank@1879 31 #include "opto/divnode.hpp"
stefank@1879 32 #include "opto/loopnode.hpp"
stefank@1879 33 #include "opto/mulnode.hpp"
stefank@1879 34 #include "opto/rootnode.hpp"
stefank@1879 35 #include "opto/runtime.hpp"
stefank@1879 36 #include "opto/subnode.hpp"
duke@0 37
duke@0 38 //------------------------------is_loop_exit-----------------------------------
duke@0 39 // Given an IfNode, return the loop-exiting projection or NULL if both
duke@0 40 // arms remain in the loop.
duke@0 41 Node *IdealLoopTree::is_loop_exit(Node *iff) const {
duke@0 42 if( iff->outcnt() != 2 ) return NULL; // Ignore partially dead tests
duke@0 43 PhaseIdealLoop *phase = _phase;
duke@0 44 // Test is an IfNode, has 2 projections. If BOTH are in the loop
duke@0 45 // we need loop unswitching instead of peeling.
duke@0 46 if( !is_member(phase->get_loop( iff->raw_out(0) )) )
duke@0 47 return iff->raw_out(0);
duke@0 48 if( !is_member(phase->get_loop( iff->raw_out(1) )) )
duke@0 49 return iff->raw_out(1);
duke@0 50 return NULL;
duke@0 51 }
duke@0 52
duke@0 53
duke@0 54 //=============================================================================
duke@0 55
duke@0 56
duke@0 57 //------------------------------record_for_igvn----------------------------
duke@0 58 // Put loop body on igvn work list
duke@0 59 void IdealLoopTree::record_for_igvn() {
duke@0 60 for( uint i = 0; i < _body.size(); i++ ) {
duke@0 61 Node *n = _body.at(i);
duke@0 62 _phase->_igvn._worklist.push(n);
duke@0 63 }
duke@0 64 }
duke@0 65
kvn@2312 66 //------------------------------compute_exact_trip_count-----------------------
kvn@2312 67 // Compute loop exact trip count if possible. Do not recalculate trip count for
kvn@2312 68 // split loops (pre-main-post) which have their limits and inits behind Opaque node.
kvn@2312 69 void IdealLoopTree::compute_exact_trip_count( PhaseIdealLoop *phase ) {
kvn@2312 70 if (!_head->as_Loop()->is_valid_counted_loop()) {
kvn@2312 71 return;
kvn@2312 72 }
kvn@2312 73 CountedLoopNode* cl = _head->as_CountedLoop();
kvn@2312 74 // Trip count may become nonexact for iteration split loops since
kvn@2312 75 // RCE modifies limits. Note, _trip_count value is not reset since
kvn@2312 76 // it is used to limit unrolling of main loop.
kvn@2312 77 cl->set_nonexact_trip_count();
kvn@2312 78
kvn@2312 79 // Loop's test should be part of loop.
kvn@2312 80 if (!phase->is_member(this, phase->get_ctrl(cl->loopexit()->in(CountedLoopEndNode::TestValue))))
kvn@2312 81 return; // Infinite loop
kvn@2312 82
kvn@2312 83 #ifdef ASSERT
kvn@2312 84 BoolTest::mask bt = cl->loopexit()->test_trip();
kvn@2312 85 assert(bt == BoolTest::lt || bt == BoolTest::gt ||
kvn@2544 86 bt == BoolTest::ne, "canonical test is expected");
kvn@2312 87 #endif
kvn@2312 88
kvn@2312 89 Node* init_n = cl->init_trip();
kvn@2312 90 Node* limit_n = cl->limit();
kvn@2312 91 if (init_n != NULL && init_n->is_Con() &&
kvn@2312 92 limit_n != NULL && limit_n->is_Con()) {
kvn@2312 93 // Use longs to avoid integer overflow.
kvn@2312 94 int stride_con = cl->stride_con();
vlivanov@3722 95 jlong init_con = cl->init_trip()->get_int();
vlivanov@3722 96 jlong limit_con = cl->limit()->get_int();
kvn@2312 97 int stride_m = stride_con - (stride_con > 0 ? 1 : -1);
vlivanov@3722 98 jlong trip_count = (limit_con - init_con + stride_m)/stride_con;
kvn@2312 99 if (trip_count > 0 && (julong)trip_count < (julong)max_juint) {
kvn@2312 100 // Set exact trip count.
kvn@2312 101 cl->set_exact_trip_count((uint)trip_count);
kvn@2312 102 }
kvn@2312 103 }
kvn@2312 104 }
kvn@2312 105
duke@0 106 //------------------------------compute_profile_trip_cnt----------------------------
duke@0 107 // Compute loop trip count from profile data as
duke@0 108 // (backedge_count + loop_exit_count) / loop_exit_count
duke@0 109 void IdealLoopTree::compute_profile_trip_cnt( PhaseIdealLoop *phase ) {
duke@0 110 if (!_head->is_CountedLoop()) {
duke@0 111 return;
duke@0 112 }
duke@0 113 CountedLoopNode* head = _head->as_CountedLoop();
duke@0 114 if (head->profile_trip_cnt() != COUNT_UNKNOWN) {
duke@0 115 return; // Already computed
duke@0 116 }
duke@0 117 float trip_cnt = (float)max_jint; // default is big
duke@0 118
duke@0 119 Node* back = head->in(LoopNode::LoopBackControl);
duke@0 120 while (back != head) {
duke@0 121 if ((back->Opcode() == Op_IfTrue || back->Opcode() == Op_IfFalse) &&
duke@0 122 back->in(0) &&
duke@0 123 back->in(0)->is_If() &&
duke@0 124 back->in(0)->as_If()->_fcnt != COUNT_UNKNOWN &&
duke@0 125 back->in(0)->as_If()->_prob != PROB_UNKNOWN) {
duke@0 126 break;
duke@0 127 }
duke@0 128 back = phase->idom(back);
duke@0 129 }
duke@0 130 if (back != head) {
duke@0 131 assert((back->Opcode() == Op_IfTrue || back->Opcode() == Op_IfFalse) &&
duke@0 132 back->in(0), "if-projection exists");
duke@0 133 IfNode* back_if = back->in(0)->as_If();
duke@0 134 float loop_back_cnt = back_if->_fcnt * back_if->_prob;
duke@0 135
duke@0 136 // Now compute a loop exit count
duke@0 137 float loop_exit_cnt = 0.0f;
duke@0 138 for( uint i = 0; i < _body.size(); i++ ) {
duke@0 139 Node *n = _body[i];
duke@0 140 if( n->is_If() ) {
duke@0 141 IfNode *iff = n->as_If();
duke@0 142 if( iff->_fcnt != COUNT_UNKNOWN && iff->_prob != PROB_UNKNOWN ) {
duke@0 143 Node *exit = is_loop_exit(iff);
duke@0 144 if( exit ) {
duke@0 145 float exit_prob = iff->_prob;
duke@0 146 if (exit->Opcode() == Op_IfFalse) exit_prob = 1.0 - exit_prob;
duke@0 147 if (exit_prob > PROB_MIN) {
duke@0 148 float exit_cnt = iff->_fcnt * exit_prob;
duke@0 149 loop_exit_cnt += exit_cnt;
duke@0 150 }
duke@0 151 }
duke@0 152 }
duke@0 153 }
duke@0 154 }
duke@0 155 if (loop_exit_cnt > 0.0f) {
duke@0 156 trip_cnt = (loop_back_cnt + loop_exit_cnt) / loop_exit_cnt;
duke@0 157 } else {
duke@0 158 // No exit count so use
duke@0 159 trip_cnt = loop_back_cnt;
duke@0 160 }
duke@0 161 }
duke@0 162 #ifndef PRODUCT
duke@0 163 if (TraceProfileTripCount) {
duke@0 164 tty->print_cr("compute_profile_trip_cnt lp: %d cnt: %f\n", head->_idx, trip_cnt);
duke@0 165 }
duke@0 166 #endif
duke@0 167 head->set_profile_trip_cnt(trip_cnt);
duke@0 168 }
duke@0 169
duke@0 170 //---------------------is_invariant_addition-----------------------------
duke@0 171 // Return nonzero index of invariant operand for an Add or Sub
twisti@605 172 // of (nonconstant) invariant and variant values. Helper for reassociate_invariants.
duke@0 173 int IdealLoopTree::is_invariant_addition(Node* n, PhaseIdealLoop *phase) {
duke@0 174 int op = n->Opcode();
duke@0 175 if (op == Op_AddI || op == Op_SubI) {
duke@0 176 bool in1_invar = this->is_invariant(n->in(1));
duke@0 177 bool in2_invar = this->is_invariant(n->in(2));
duke@0 178 if (in1_invar && !in2_invar) return 1;
duke@0 179 if (!in1_invar && in2_invar) return 2;
duke@0 180 }
duke@0 181 return 0;
duke@0 182 }
duke@0 183
duke@0 184 //---------------------reassociate_add_sub-----------------------------
duke@0 185 // Reassociate invariant add and subtract expressions:
duke@0 186 //
duke@0 187 // inv1 + (x + inv2) => ( inv1 + inv2) + x
duke@0 188 // (x + inv2) + inv1 => ( inv1 + inv2) + x
duke@0 189 // inv1 + (x - inv2) => ( inv1 - inv2) + x
duke@0 190 // inv1 - (inv2 - x) => ( inv1 - inv2) + x
duke@0 191 // (x + inv2) - inv1 => (-inv1 + inv2) + x
duke@0 192 // (x - inv2) + inv1 => ( inv1 - inv2) + x
duke@0 193 // (x - inv2) - inv1 => (-inv1 - inv2) + x
duke@0 194 // inv1 + (inv2 - x) => ( inv1 + inv2) - x
duke@0 195 // inv1 - (x - inv2) => ( inv1 + inv2) - x
duke@0 196 // (inv2 - x) + inv1 => ( inv1 + inv2) - x
duke@0 197 // (inv2 - x) - inv1 => (-inv1 + inv2) - x
duke@0 198 // inv1 - (x + inv2) => ( inv1 - inv2) - x
duke@0 199 //
duke@0 200 Node* IdealLoopTree::reassociate_add_sub(Node* n1, PhaseIdealLoop *phase) {
duke@0 201 if (!n1->is_Add() && !n1->is_Sub() || n1->outcnt() == 0) return NULL;
duke@0 202 if (is_invariant(n1)) return NULL;
duke@0 203 int inv1_idx = is_invariant_addition(n1, phase);
duke@0 204 if (!inv1_idx) return NULL;
duke@0 205 // Don't mess with add of constant (igvn moves them to expression tree root.)
duke@0 206 if (n1->is_Add() && n1->in(2)->is_Con()) return NULL;
duke@0 207 Node* inv1 = n1->in(inv1_idx);
duke@0 208 Node* n2 = n1->in(3 - inv1_idx);
duke@0 209 int inv2_idx = is_invariant_addition(n2, phase);
duke@0 210 if (!inv2_idx) return NULL;
duke@0 211 Node* x = n2->in(3 - inv2_idx);
duke@0 212 Node* inv2 = n2->in(inv2_idx);
duke@0 213
duke@0 214 bool neg_x = n2->is_Sub() && inv2_idx == 1;
duke@0 215 bool neg_inv2 = n2->is_Sub() && inv2_idx == 2;
duke@0 216 bool neg_inv1 = n1->is_Sub() && inv1_idx == 2;
duke@0 217 if (n1->is_Sub() && inv1_idx == 1) {
duke@0 218 neg_x = !neg_x;
duke@0 219 neg_inv2 = !neg_inv2;
duke@0 220 }
duke@0 221 Node* inv1_c = phase->get_ctrl(inv1);
duke@0 222 Node* inv2_c = phase->get_ctrl(inv2);
duke@0 223 Node* n_inv1;
duke@0 224 if (neg_inv1) {
duke@0 225 Node *zero = phase->_igvn.intcon(0);
duke@0 226 phase->set_ctrl(zero, phase->C->root());
kvn@3680 227 n_inv1 = new (phase->C) SubINode(zero, inv1);
duke@0 228 phase->register_new_node(n_inv1, inv1_c);
duke@0 229 } else {
duke@0 230 n_inv1 = inv1;
duke@0 231 }
duke@0 232 Node* inv;
duke@0 233 if (neg_inv2) {
kvn@3680 234 inv = new (phase->C) SubINode(n_inv1, inv2);
duke@0 235 } else {
kvn@3680 236 inv = new (phase->C) AddINode(n_inv1, inv2);
duke@0 237 }
duke@0 238 phase->register_new_node(inv, phase->get_early_ctrl(inv));
duke@0 239
duke@0 240 Node* addx;
duke@0 241 if (neg_x) {
kvn@3680 242 addx = new (phase->C) SubINode(inv, x);
duke@0 243 } else {
kvn@3680 244 addx = new (phase->C) AddINode(x, inv);
duke@0 245 }
duke@0 246 phase->register_new_node(addx, phase->get_ctrl(x));
kvn@1541 247 phase->_igvn.replace_node(n1, addx);
kvn@2230 248 assert(phase->get_loop(phase->get_ctrl(n1)) == this, "");
kvn@2230 249 _body.yank(n1);
duke@0 250 return addx;
duke@0 251 }
duke@0 252
duke@0 253 //---------------------reassociate_invariants-----------------------------
duke@0 254 // Reassociate invariant expressions:
duke@0 255 void IdealLoopTree::reassociate_invariants(PhaseIdealLoop *phase) {
duke@0 256 for (int i = _body.size() - 1; i >= 0; i--) {
duke@0 257 Node *n = _body.at(i);
duke@0 258 for (int j = 0; j < 5; j++) {
duke@0 259 Node* nn = reassociate_add_sub(n, phase);
duke@0 260 if (nn == NULL) break;
duke@0 261 n = nn; // again
duke@0 262 };
duke@0 263 }
duke@0 264 }
duke@0 265
duke@0 266 //------------------------------policy_peeling---------------------------------
duke@0 267 // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can
duke@0 268 // make some loop-invariant test (usually a null-check) happen before the loop.
duke@0 269 bool IdealLoopTree::policy_peeling( PhaseIdealLoop *phase ) const {
duke@0 270 Node *test = ((IdealLoopTree*)this)->tail();
duke@0 271 int body_size = ((IdealLoopTree*)this)->_body.size();
bharadwaj@3880 272 int live_node_count = phase->C->live_nodes();
duke@0 273 // Peeling does loop cloning which can result in O(N^2) node construction
duke@0 274 if( body_size > 255 /* Prevent overflow for large body_size */
bharadwaj@3880 275 || (body_size * body_size + live_node_count > MaxNodeLimit) ) {
duke@0 276 return false; // too large to safely clone
duke@0 277 }
duke@0 278 while( test != _head ) { // Scan till run off top of loop
duke@0 279 if( test->is_If() ) { // Test?
duke@0 280 Node *ctrl = phase->get_ctrl(test->in(1));
duke@0 281 if (ctrl->is_top())
duke@0 282 return false; // Found dead test on live IF? No peeling!
duke@0 283 // Standard IF only has one input value to check for loop invariance
duke@0 284 assert( test->Opcode() == Op_If || test->Opcode() == Op_CountedLoopEnd, "Check this code when new subtype is added");
duke@0 285 // Condition is not a member of this loop?
duke@0 286 if( !is_member(phase->get_loop(ctrl)) &&
duke@0 287 is_loop_exit(test) )
duke@0 288 return true; // Found reason to peel!
duke@0 289 }
duke@0 290 // Walk up dominators to loop _head looking for test which is
duke@0 291 // executed on every path thru loop.
duke@0 292 test = phase->idom(test);
duke@0 293 }
duke@0 294 return false;
duke@0 295 }
duke@0 296
duke@0 297 //------------------------------peeled_dom_test_elim---------------------------
duke@0 298 // If we got the effect of peeling, either by actually peeling or by making
duke@0 299 // a pre-loop which must execute at least once, we can remove all
duke@0 300 // loop-invariant dominated tests in the main body.
duke@0 301 void PhaseIdealLoop::peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new ) {
duke@0 302 bool progress = true;
duke@0 303 while( progress ) {
duke@0 304 progress = false; // Reset for next iteration
duke@0 305 Node *prev = loop->_head->in(LoopNode::LoopBackControl);//loop->tail();
duke@0 306 Node *test = prev->in(0);
duke@0 307 while( test != loop->_head ) { // Scan till run off top of loop
duke@0 308
duke@0 309 int p_op = prev->Opcode();
duke@0 310 if( (p_op == Op_IfFalse || p_op == Op_IfTrue) &&
duke@0 311 test->is_If() && // Test?
duke@0 312 !test->in(1)->is_Con() && // And not already obvious?
duke@0 313 // Condition is not a member of this loop?
duke@0 314 !loop->is_member(get_loop(get_ctrl(test->in(1))))){
duke@0 315 // Walk loop body looking for instances of this test
duke@0 316 for( uint i = 0; i < loop->_body.size(); i++ ) {
duke@0 317 Node *n = loop->_body.at(i);
duke@0 318 if( n->is_If() && n->in(1) == test->in(1) /*&& n != loop->tail()->in(0)*/ ) {
duke@0 319 // IfNode was dominated by version in peeled loop body
duke@0 320 progress = true;
duke@0 321 dominated_by( old_new[prev->_idx], n );
duke@0 322 }
duke@0 323 }
duke@0 324 }
duke@0 325 prev = test;
duke@0 326 test = idom(test);
duke@0 327 } // End of scan tests in loop
duke@0 328
duke@0 329 } // End of while( progress )
duke@0 330 }
duke@0 331
duke@0 332 //------------------------------do_peeling-------------------------------------
duke@0 333 // Peel the first iteration of the given loop.
duke@0 334 // Step 1: Clone the loop body. The clone becomes the peeled iteration.
duke@0 335 // The pre-loop illegally has 2 control users (old & new loops).
duke@0 336 // Step 2: Make the old-loop fall-in edges point to the peeled iteration.
duke@0 337 // Do this by making the old-loop fall-in edges act as if they came
duke@0 338 // around the loopback from the prior iteration (follow the old-loop
duke@0 339 // backedges) and then map to the new peeled iteration. This leaves
duke@0 340 // the pre-loop with only 1 user (the new peeled iteration), but the
duke@0 341 // peeled-loop backedge has 2 users.
duke@0 342 // Step 3: Cut the backedge on the clone (so its not a loop) and remove the
duke@0 343 // extra backedge user.
kvn@2292 344 //
kvn@2292 345 // orig
kvn@2292 346 //
kvn@2292 347 // stmt1
kvn@2292 348 // |
kvn@2292 349 // v
kvn@2292 350 // loop predicate
kvn@2292 351 // |
kvn@2292 352 // v
kvn@2292 353 // loop<----+
kvn@2292 354 // | |
kvn@2292 355 // stmt2 |
kvn@2292 356 // | |
kvn@2292 357 // v |
kvn@2292 358 // if ^
kvn@2292 359 // / \ |
kvn@2292 360 // / \ |
kvn@2292 361 // v v |
kvn@2292 362 // false true |
kvn@2292 363 // / \ |
kvn@2292 364 // / ----+
kvn@2292 365 // |
kvn@2292 366 // v
kvn@2292 367 // exit
kvn@2292 368 //
kvn@2292 369 //
kvn@2292 370 // after clone loop
kvn@2292 371 //
kvn@2292 372 // stmt1
kvn@2292 373 // |
kvn@2292 374 // v
kvn@2292 375 // loop predicate
kvn@2292 376 // / \
kvn@2292 377 // clone / \ orig
kvn@2292 378 // / \
kvn@2292 379 // / \
kvn@2292 380 // v v
kvn@2292 381 // +---->loop clone loop<----+
kvn@2292 382 // | | | |
kvn@2292 383 // | stmt2 clone stmt2 |
kvn@2292 384 // | | | |
kvn@2292 385 // | v v |
kvn@2292 386 // ^ if clone If ^
kvn@2292 387 // | / \ / \ |
kvn@2292 388 // | / \ / \ |
kvn@2292 389 // | v v v v |
kvn@2292 390 // | true false false true |
kvn@2292 391 // | / \ / \ |
kvn@2292 392 // +---- \ / ----+
kvn@2292 393 // \ /
kvn@2292 394 // 1v v2
kvn@2292 395 // region
kvn@2292 396 // |
kvn@2292 397 // v
kvn@2292 398 // exit
kvn@2292 399 //
kvn@2292 400 //
kvn@2292 401 // after peel and predicate move
kvn@2292 402 //
kvn@2292 403 // stmt1
kvn@2292 404 // /
kvn@2292 405 // /
kvn@2292 406 // clone / orig
kvn@2292 407 // /
kvn@2292 408 // / +----------+
kvn@2292 409 // / | |
kvn@2292 410 // / loop predicate |
kvn@2292 411 // / | |
kvn@2292 412 // v v |
kvn@2292 413 // TOP-->loop clone loop<----+ |
kvn@2292 414 // | | | |
kvn@2292 415 // stmt2 clone stmt2 | |
kvn@2292 416 // | | | ^
kvn@2292 417 // v v | |
kvn@2292 418 // if clone If ^ |
kvn@2292 419 // / \ / \ | |
kvn@2292 420 // / \ / \ | |
kvn@2292 421 // v v v v | |
kvn@2292 422 // true false false true | |
kvn@2292 423 // | \ / \ | |
kvn@2292 424 // | \ / ----+ ^
kvn@2292 425 // | \ / |
kvn@2292 426 // | 1v v2 |
kvn@2292 427 // v region |
kvn@2292 428 // | | |
kvn@2292 429 // | v |
kvn@2292 430 // | exit |
kvn@2292 431 // | |
kvn@2292 432 // +--------------->-----------------+
kvn@2292 433 //
kvn@2292 434 //
kvn@2292 435 // final graph
kvn@2292 436 //
kvn@2292 437 // stmt1
kvn@2292 438 // |
kvn@2292 439 // v
kvn@2292 440 // stmt2 clone
kvn@2292 441 // |
kvn@2292 442 // v
kvn@2292 443 // if clone
kvn@2292 444 // / |
kvn@2292 445 // / |
kvn@2292 446 // v v
kvn@2292 447 // false true
kvn@2292 448 // | |
kvn@2292 449 // | v
kvn@2292 450 // | loop predicate
kvn@2292 451 // | |
kvn@2292 452 // | v
kvn@2292 453 // | loop<----+
kvn@2292 454 // | | |
kvn@2292 455 // | stmt2 |
kvn@2292 456 // | | |
kvn@2292 457 // | v |
kvn@2292 458 // v if ^
kvn@2292 459 // | / \ |
kvn@2292 460 // | / \ |
kvn@2292 461 // | v v |
kvn@2292 462 // | false true |
kvn@2292 463 // | | \ |
kvn@2292 464 // v v --+
kvn@2292 465 // region
kvn@2292 466 // |
kvn@2292 467 // v
kvn@2292 468 // exit
kvn@2292 469 //
duke@0 470 void PhaseIdealLoop::do_peeling( IdealLoopTree *loop, Node_List &old_new ) {
duke@0 471
duke@0 472 C->set_major_progress();
duke@0 473 // Peeling a 'main' loop in a pre/main/post situation obfuscates the
duke@0 474 // 'pre' loop from the main and the 'pre' can no longer have it's
duke@0 475 // iterations adjusted. Therefore, we need to declare this loop as
duke@0 476 // no longer a 'main' loop; it will need new pre and post loops before
duke@0 477 // we can do further RCE.
kvn@2230 478 #ifndef PRODUCT
kvn@2230 479 if (TraceLoopOpts) {
kvn@2230 480 tty->print("Peel ");
kvn@2230 481 loop->dump_head();
kvn@2230 482 }
kvn@2230 483 #endif
kvn@2292 484 Node* head = loop->_head;
kvn@2292 485 bool counted_loop = head->is_CountedLoop();
kvn@2292 486 if (counted_loop) {
kvn@2292 487 CountedLoopNode *cl = head->as_CountedLoop();
duke@0 488 assert(cl->trip_count() > 0, "peeling a fully unrolled loop");
duke@0 489 cl->set_trip_count(cl->trip_count() - 1);
kvn@2230 490 if (cl->is_main_loop()) {
duke@0 491 cl->set_normal_loop();
duke@0 492 #ifndef PRODUCT
kvn@2230 493 if (PrintOpto && VerifyLoopOptimizations) {
duke@0 494 tty->print("Peeling a 'main' loop; resetting to 'normal' ");
duke@0 495 loop->dump_head();
duke@0 496 }
duke@0 497 #endif
duke@0 498 }
duke@0 499 }
kvn@2292 500 Node* entry = head->in(LoopNode::EntryControl);
duke@0 501
duke@0 502 // Step 1: Clone the loop body. The clone becomes the peeled iteration.
duke@0 503 // The pre-loop illegally has 2 control users (old & new loops).
kvn@2292 504 clone_loop( loop, old_new, dom_depth(head) );
duke@0 505
duke@0 506 // Step 2: Make the old-loop fall-in edges point to the peeled iteration.
duke@0 507 // Do this by making the old-loop fall-in edges act as if they came
duke@0 508 // around the loopback from the prior iteration (follow the old-loop
duke@0 509 // backedges) and then map to the new peeled iteration. This leaves
duke@0 510 // the pre-loop with only 1 user (the new peeled iteration), but the
duke@0 511 // peeled-loop backedge has 2 users.
kvn@2608 512 Node* new_entry = old_new[head->in(LoopNode::LoopBackControl)->_idx];
kvn@2292 513 _igvn.hash_delete(head);
kvn@2608 514 head->set_req(LoopNode::EntryControl, new_entry);
kvn@2292 515 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
kvn@2292 516 Node* old = head->fast_out(j);
kvn@2292 517 if (old->in(0) == loop->_head && old->req() == 3 && old->is_Phi()) {
kvn@2608 518 Node* new_exit_value = old_new[old->in(LoopNode::LoopBackControl)->_idx];
kvn@2292 519 if (!new_exit_value ) // Backedge value is ALSO loop invariant?
duke@0 520 // Then loop body backedge value remains the same.
duke@0 521 new_exit_value = old->in(LoopNode::LoopBackControl);
duke@0 522 _igvn.hash_delete(old);
duke@0 523 old->set_req(LoopNode::EntryControl, new_exit_value);
duke@0 524 }
duke@0 525 }
duke@0 526
duke@0 527
duke@0 528 // Step 3: Cut the backedge on the clone (so its not a loop) and remove the
duke@0 529 // extra backedge user.
kvn@2292 530 Node* new_head = old_new[head->_idx];
kvn@2292 531 _igvn.hash_delete(new_head);
kvn@2292 532 new_head->set_req(LoopNode::LoopBackControl, C->top());
kvn@2292 533 for (DUIterator_Fast j2max, j2 = new_head->fast_outs(j2max); j2 < j2max; j2++) {
kvn@2292 534 Node* use = new_head->fast_out(j2);
kvn@2292 535 if (use->in(0) == new_head && use->req() == 3 && use->is_Phi()) {
duke@0 536 _igvn.hash_delete(use);
duke@0 537 use->set_req(LoopNode::LoopBackControl, C->top());
duke@0 538 }
duke@0 539 }
duke@0 540
duke@0 541
duke@0 542 // Step 4: Correct dom-depth info. Set to loop-head depth.
kvn@2292 543 int dd = dom_depth(head);
kvn@2292 544 set_idom(head, head->in(1), dd);
duke@0 545 for (uint j3 = 0; j3 < loop->_body.size(); j3++) {
duke@0 546 Node *old = loop->_body.at(j3);
duke@0 547 Node *nnn = old_new[old->_idx];
duke@0 548 if (!has_ctrl(nnn))
duke@0 549 set_idom(nnn, idom(nnn), dd-1);
duke@0 550 }
duke@0 551
duke@0 552 // Now force out all loop-invariant dominating tests. The optimizer
duke@0 553 // finds some, but we _know_ they are all useless.
duke@0 554 peeled_dom_test_elim(loop,old_new);
duke@0 555
duke@0 556 loop->record_for_igvn();
duke@0 557 }
duke@0 558
kvn@2300 559 #define EMPTY_LOOP_SIZE 7 // number of nodes in an empty loop
kvn@2300 560
duke@0 561 //------------------------------policy_maximally_unroll------------------------
kvn@2300 562 // Calculate exact loop trip count and return true if loop can be maximally
kvn@2300 563 // unrolled.
duke@0 564 bool IdealLoopTree::policy_maximally_unroll( PhaseIdealLoop *phase ) const {
duke@0 565 CountedLoopNode *cl = _head->as_CountedLoop();
kvn@2259 566 assert(cl->is_normal_loop(), "");
kvn@2300 567 if (!cl->is_valid_counted_loop())
kvn@2300 568 return false; // Malformed counted loop
duke@0 569
kvn@2312 570 if (!cl->has_exact_trip_count()) {
kvn@2312 571 // Trip count is not exact.
duke@0 572 return false;
duke@0 573 }
duke@0 574
kvn@2312 575 uint trip_count = cl->trip_count();
kvn@2312 576 // Note, max_juint is used to indicate unknown trip count.
kvn@2312 577 assert(trip_count > 1, "one iteration loop should be optimized out already");
kvn@2312 578 assert(trip_count < max_juint, "exact trip_count should be less than max_uint.");
duke@0 579
duke@0 580 // Real policy: if we maximally unroll, does it get too big?
duke@0 581 // Allow the unrolled mess to get larger than standard loop
duke@0 582 // size. After all, it will no longer be a loop.
duke@0 583 uint body_size = _body.size();
duke@0 584 uint unroll_limit = (uint)LoopUnrollLimit * 4;
duke@0 585 assert( (intx)unroll_limit == LoopUnrollLimit * 4, "LoopUnrollLimit must fit in 32bits");
kvn@2259 586 if (trip_count > unroll_limit || body_size > unroll_limit) {
kvn@2259 587 return false;
kvn@2259 588 }
kvn@2259 589
kvn@2442 590 // Fully unroll a loop with few iterations regardless next
kvn@2442 591 // conditions since following loop optimizations will split
kvn@2442 592 // such loop anyway (pre-main-post).
kvn@2442 593 if (trip_count <= 3)
kvn@2442 594 return true;
kvn@2442 595
kvn@2300 596 // Take into account that after unroll conjoined heads and tails will fold,
kvn@2300 597 // otherwise policy_unroll() may allow more unrolling than max unrolling.
kvn@2300 598 uint new_body_size = EMPTY_LOOP_SIZE + (body_size - EMPTY_LOOP_SIZE) * trip_count;
kvn@2300 599 uint tst_body_size = (new_body_size - EMPTY_LOOP_SIZE) / trip_count + EMPTY_LOOP_SIZE;
kvn@2300 600 if (body_size != tst_body_size) // Check for int overflow
kvn@2300 601 return false;
kvn@2300 602 if (new_body_size > unroll_limit ||
kvn@2300 603 // Unrolling can result in a large amount of node construction
bharadwaj@3880 604 new_body_size >= MaxNodeLimit - (uint) phase->C->live_nodes()) {
kvn@2300 605 return false;
kvn@2300 606 }
kvn@2300 607
kvn@2259 608 // Do not unroll a loop with String intrinsics code.
kvn@2259 609 // String intrinsics are large and have loops.
kvn@2259 610 for (uint k = 0; k < _body.size(); k++) {
kvn@2259 611 Node* n = _body.at(k);
kvn@2259 612 switch (n->Opcode()) {
kvn@2259 613 case Op_StrComp:
kvn@2259 614 case Op_StrEquals:
kvn@2259 615 case Op_StrIndexOf:
kvn@4044 616 case Op_EncodeISOArray:
kvn@2259 617 case Op_AryEq: {
kvn@2259 618 return false;
kvn@2259 619 }
kvn@2259 620 } // switch
kvn@2259 621 }
kvn@2259 622
kvn@2300 623 return true; // Do maximally unroll
duke@0 624 }
duke@0 625
duke@0 626
duke@0 627 //------------------------------policy_unroll----------------------------------
duke@0 628 // Return TRUE or FALSE if the loop should be unrolled or not. Unroll if
duke@0 629 // the loop is a CountedLoop and the body is small enough.
duke@0 630 bool IdealLoopTree::policy_unroll( PhaseIdealLoop *phase ) const {
duke@0 631
duke@0 632 CountedLoopNode *cl = _head->as_CountedLoop();
kvn@2259 633 assert(cl->is_normal_loop() || cl->is_main_loop(), "");
duke@0 634
kvn@2300 635 if (!cl->is_valid_counted_loop())
kvn@2300 636 return false; // Malformed counted loop
duke@0 637
kvn@2442 638 // Protect against over-unrolling.
kvn@2442 639 // After split at least one iteration will be executed in pre-loop.
kvn@2442 640 if (cl->trip_count() <= (uint)(cl->is_normal_loop() ? 2 : 1)) return false;
duke@0 641
kvn@2430 642 int future_unroll_ct = cl->unrolled_count() * 2;
kvn@5078 643 if (future_unroll_ct > LoopMaxUnroll) return false;
kvn@2300 644
kvn@2430 645 // Check for initial stride being a small enough constant
kvn@2430 646 if (abs(cl->stride_con()) > (1<<2)*future_unroll_ct) return false;
duke@0 647
duke@0 648 // Don't unroll if the next round of unrolling would push us
duke@0 649 // over the expected trip count of the loop. One is subtracted
duke@0 650 // from the expected trip count because the pre-loop normally
duke@0 651 // executes 1 iteration.
duke@0 652 if (UnrollLimitForProfileCheck > 0 &&
duke@0 653 cl->profile_trip_cnt() != COUNT_UNKNOWN &&
duke@0 654 future_unroll_ct > UnrollLimitForProfileCheck &&
duke@0 655 (float)future_unroll_ct > cl->profile_trip_cnt() - 1.0) {
duke@0 656 return false;
duke@0 657 }
duke@0 658
duke@0 659 // When unroll count is greater than LoopUnrollMin, don't unroll if:
duke@0 660 // the residual iterations are more than 10% of the trip count
duke@0 661 // and rounds of "unroll,optimize" are not making significant progress
duke@0 662 // Progress defined as current size less than 20% larger than previous size.
duke@0 663 if (UseSuperWord && cl->node_count_before_unroll() > 0 &&
duke@0 664 future_unroll_ct > LoopUnrollMin &&
duke@0 665 (future_unroll_ct - 1) * 10.0 > cl->profile_trip_cnt() &&
duke@0 666 1.2 * cl->node_count_before_unroll() < (double)_body.size()) {
duke@0 667 return false;
duke@0 668 }
duke@0 669
duke@0 670 Node *init_n = cl->init_trip();
duke@0 671 Node *limit_n = cl->limit();
kvn@2442 672 int stride_con = cl->stride_con();
duke@0 673 // Non-constant bounds.
duke@0 674 // Protect against over-unrolling when init or/and limit are not constant
duke@0 675 // (so that trip_count's init value is maxint) but iv range is known.
kvn@2259 676 if (init_n == NULL || !init_n->is_Con() ||
kvn@2259 677 limit_n == NULL || !limit_n->is_Con()) {
duke@0 678 Node* phi = cl->phi();
kvn@2259 679 if (phi != NULL) {
duke@0 680 assert(phi->is_Phi() && phi->in(0) == _head, "Counted loop should have iv phi.");
duke@0 681 const TypeInt* iv_type = phase->_igvn.type(phi)->is_int();
kvn@2442 682 int next_stride = stride_con * 2; // stride after this unroll
kvn@2259 683 if (next_stride > 0) {
kvn@2259 684 if (iv_type->_lo + next_stride <= iv_type->_lo || // overflow
kvn@2259 685 iv_type->_lo + next_stride > iv_type->_hi) {
duke@0 686 return false; // over-unrolling
duke@0 687 }
kvn@2259 688 } else if (next_stride < 0) {
kvn@2259 689 if (iv_type->_hi + next_stride >= iv_type->_hi || // overflow
kvn@2259 690 iv_type->_hi + next_stride < iv_type->_lo) {
duke@0 691 return false; // over-unrolling
duke@0 692 }
duke@0 693 }
duke@0 694 }
duke@0 695 }
duke@0 696
kvn@2442 697 // After unroll limit will be adjusted: new_limit = limit-stride.
kvn@2442 698 // Bailout if adjustment overflow.
kvn@2442 699 const TypeInt* limit_type = phase->_igvn.type(limit_n)->is_int();
kvn@2442 700 if (stride_con > 0 && ((limit_type->_hi - stride_con) >= limit_type->_hi) ||
kvn@2442 701 stride_con < 0 && ((limit_type->_lo - stride_con) <= limit_type->_lo))
kvn@2442 702 return false; // overflow
kvn@2442 703
duke@0 704 // Adjust body_size to determine if we unroll or not
duke@0 705 uint body_size = _body.size();
kvn@2694 706 // Key test to unroll loop in CRC32 java code
kvn@2694 707 int xors_in_loop = 0;
duke@0 708 // Also count ModL, DivL and MulL which expand mightly
kvn@2259 709 for (uint k = 0; k < _body.size(); k++) {
kvn@2259 710 Node* n = _body.at(k);
kvn@2259 711 switch (n->Opcode()) {
kvn@2694 712 case Op_XorI: xors_in_loop++; break; // CRC32 java code
kvn@2259 713 case Op_ModL: body_size += 30; break;
kvn@2259 714 case Op_DivL: body_size += 30; break;
kvn@2259 715 case Op_MulL: body_size += 10; break;
rbackman@5614 716 case Op_FlagsProj:
rbackman@5614 717 // Can't handle unrolling of loops containing
rbackman@5614 718 // nodes that generate a FlagsProj at the moment
rbackman@5614 719 return false;
kvn@2259 720 case Op_StrComp:
kvn@2259 721 case Op_StrEquals:
kvn@2259 722 case Op_StrIndexOf:
kvn@4044 723 case Op_EncodeISOArray:
kvn@2259 724 case Op_AryEq: {
kvn@2259 725 // Do not unroll a loop with String intrinsics code.
kvn@2259 726 // String intrinsics are large and have loops.
kvn@2259 727 return false;
kvn@2259 728 }
kvn@2259 729 } // switch
duke@0 730 }
duke@0 731
duke@0 732 // Check for being too big
kvn@2259 733 if (body_size > (uint)LoopUnrollLimit) {
kvn@2694 734 if (xors_in_loop >= 4 && body_size < (uint)LoopUnrollLimit*4) return true;
kvn@2694 735 // Normal case: loop too big
duke@0 736 return false;
duke@0 737 }
duke@0 738
duke@0 739 // Unroll once! (Each trip will soon do double iterations)
duke@0 740 return true;
duke@0 741 }
duke@0 742
duke@0 743 //------------------------------policy_align-----------------------------------
duke@0 744 // Return TRUE or FALSE if the loop should be cache-line aligned. Gather the
duke@0 745 // expression that does the alignment. Note that only one array base can be
twisti@605 746 // aligned in a loop (unless the VM guarantees mutual alignment). Note that
duke@0 747 // if we vectorize short memory ops into longer memory ops, we may want to
duke@0 748 // increase alignment.
duke@0 749 bool IdealLoopTree::policy_align( PhaseIdealLoop *phase ) const {
duke@0 750 return false;
duke@0 751 }
duke@0 752
duke@0 753 //------------------------------policy_range_check-----------------------------
duke@0 754 // Return TRUE or FALSE if the loop should be range-check-eliminated.
duke@0 755 // Actually we do iteration-splitting, a more powerful form of RCE.
duke@0 756 bool IdealLoopTree::policy_range_check( PhaseIdealLoop *phase ) const {
kvn@2442 757 if (!RangeCheckElimination) return false;
duke@0 758
duke@0 759 CountedLoopNode *cl = _head->as_CountedLoop();
duke@0 760 // If we unrolled with no intention of doing RCE and we later
duke@0 761 // changed our minds, we got no pre-loop. Either we need to
duke@0 762 // make a new pre-loop, or we gotta disallow RCE.
kvn@2442 763 if (cl->is_main_no_pre_loop()) return false; // Disallowed for now.
duke@0 764 Node *trip_counter = cl->phi();
duke@0 765
duke@0 766 // Check loop body for tests of trip-counter plus loop-invariant vs
duke@0 767 // loop-invariant.
kvn@2442 768 for (uint i = 0; i < _body.size(); i++) {
duke@0 769 Node *iff = _body[i];
kvn@2442 770 if (iff->Opcode() == Op_If) { // Test?
duke@0 771
duke@0 772 // Comparing trip+off vs limit
duke@0 773 Node *bol = iff->in(1);
kvn@2442 774 if (bol->req() != 2) continue; // dead constant test
cfang@1172 775 if (!bol->is_Bool()) {
cfang@1172 776 assert(UseLoopPredicate && bol->Opcode() == Op_Conv2B, "predicate check only");
cfang@1172 777 continue;
cfang@1172 778 }
kvn@2442 779 if (bol->as_Bool()->_test._test == BoolTest::ne)
kvn@2442 780 continue; // not RC
kvn@2442 781
duke@0 782 Node *cmp = bol->in(1);
rbackman@5356 783 if (cmp->is_FlagsProj()) {
rbackman@5356 784 continue;
rbackman@5356 785 }
duke@0 786
duke@0 787 Node *rc_exp = cmp->in(1);
duke@0 788 Node *limit = cmp->in(2);
duke@0 789
duke@0 790 Node *limit_c = phase->get_ctrl(limit);
duke@0 791 if( limit_c == phase->C->top() )
duke@0 792 return false; // Found dead test on live IF? No RCE!
duke@0 793 if( is_member(phase->get_loop(limit_c) ) ) {
duke@0 794 // Compare might have operands swapped; commute them
duke@0 795 rc_exp = cmp->in(2);
duke@0 796 limit = cmp->in(1);
duke@0 797 limit_c = phase->get_ctrl(limit);
duke@0 798 if( is_member(phase->get_loop(limit_c) ) )
duke@0 799 continue; // Both inputs are loop varying; cannot RCE
duke@0 800 }
duke@0 801
duke@0 802 if (!phase->is_scaled_iv_plus_offset(rc_exp, trip_counter, NULL, NULL)) {
duke@0 803 continue;
duke@0 804 }
duke@0 805 // Yeah! Found a test like 'trip+off vs limit'
duke@0 806 // Test is an IfNode, has 2 projections. If BOTH are in the loop
duke@0 807 // we need loop unswitching instead of iteration splitting.
duke@0 808 if( is_loop_exit(iff) )
duke@0 809 return true; // Found reason to split iterations
duke@0 810 } // End of is IF
duke@0 811 }
duke@0 812
duke@0 813 return false;
duke@0 814 }
duke@0 815
duke@0 816 //------------------------------policy_peel_only-------------------------------
duke@0 817 // Return TRUE or FALSE if the loop should NEVER be RCE'd or aligned. Useful
duke@0 818 // for unrolling loops with NO array accesses.
duke@0 819 bool IdealLoopTree::policy_peel_only( PhaseIdealLoop *phase ) const {
duke@0 820
duke@0 821 for( uint i = 0; i < _body.size(); i++ )
duke@0 822 if( _body[i]->is_Mem() )
duke@0 823 return false;
duke@0 824
duke@0 825 // No memory accesses at all!
duke@0 826 return true;
duke@0 827 }
duke@0 828
duke@0 829 //------------------------------clone_up_backedge_goo--------------------------
duke@0 830 // If Node n lives in the back_ctrl block and cannot float, we clone a private
duke@0 831 // version of n in preheader_ctrl block and return that, otherwise return n.
kvn@2550 832 Node *PhaseIdealLoop::clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n, VectorSet &visited, Node_Stack &clones ) {
duke@0 833 if( get_ctrl(n) != back_ctrl ) return n;
duke@0 834
kvn@2550 835 // Only visit once
kvn@2550 836 if (visited.test_set(n->_idx)) {
kvn@2550 837 Node *x = clones.find(n->_idx);
kvn@2550 838 if (x != NULL)
kvn@2550 839 return x;
kvn@2550 840 return n;
kvn@2550 841 }
kvn@2550 842
duke@0 843 Node *x = NULL; // If required, a clone of 'n'
duke@0 844 // Check for 'n' being pinned in the backedge.
duke@0 845 if( n->in(0) && n->in(0) == back_ctrl ) {
kvn@2550 846 assert(clones.find(n->_idx) == NULL, "dead loop");
duke@0 847 x = n->clone(); // Clone a copy of 'n' to preheader
kvn@2550 848 clones.push(x, n->_idx);
duke@0 849 x->set_req( 0, preheader_ctrl ); // Fix x's control input to preheader
duke@0 850 }
duke@0 851
duke@0 852 // Recursive fixup any other input edges into x.
duke@0 853 // If there are no changes we can just return 'n', otherwise
duke@0 854 // we need to clone a private copy and change it.
duke@0 855 for( uint i = 1; i < n->req(); i++ ) {
kvn@2550 856 Node *g = clone_up_backedge_goo( back_ctrl, preheader_ctrl, n->in(i), visited, clones );
duke@0 857 if( g != n->in(i) ) {
kvn@2550 858 if( !x ) {
kvn@2550 859 assert(clones.find(n->_idx) == NULL, "dead loop");
duke@0 860 x = n->clone();
kvn@2550 861 clones.push(x, n->_idx);
kvn@2550 862 }
duke@0 863 x->set_req(i, g);
duke@0 864 }
duke@0 865 }
duke@0 866 if( x ) { // x can legally float to pre-header location
duke@0 867 register_new_node( x, preheader_ctrl );
duke@0 868 return x;
duke@0 869 } else { // raise n to cover LCA of uses
duke@0 870 set_ctrl( n, find_non_split_ctrl(back_ctrl->in(0)) );
duke@0 871 }
duke@0 872 return n;
duke@0 873 }
duke@0 874
duke@0 875 //------------------------------insert_pre_post_loops--------------------------
duke@0 876 // Insert pre and post loops. If peel_only is set, the pre-loop can not have
duke@0 877 // more iterations added. It acts as a 'peel' only, no lower-bound RCE, no
duke@0 878 // alignment. Useful to unroll loops that do no array accesses.
duke@0 879 void PhaseIdealLoop::insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only ) {
duke@0 880
kvn@2230 881 #ifndef PRODUCT
kvn@2230 882 if (TraceLoopOpts) {
kvn@2230 883 if (peel_only)
kvn@2230 884 tty->print("PeelMainPost ");
kvn@2230 885 else
kvn@2230 886 tty->print("PreMainPost ");
kvn@2230 887 loop->dump_head();
kvn@2230 888 }
kvn@2230 889 #endif
duke@0 890 C->set_major_progress();
duke@0 891
duke@0 892 // Find common pieces of the loop being guarded with pre & post loops
duke@0 893 CountedLoopNode *main_head = loop->_head->as_CountedLoop();
duke@0 894 assert( main_head->is_normal_loop(), "" );
duke@0 895 CountedLoopEndNode *main_end = main_head->loopexit();
morris@4344 896 guarantee(main_end != NULL, "no loop exit node");
duke@0 897 assert( main_end->outcnt() == 2, "1 true, 1 false path only" );
duke@0 898 uint dd_main_head = dom_depth(main_head);
duke@0 899 uint max = main_head->outcnt();
duke@0 900
duke@0 901 Node *pre_header= main_head->in(LoopNode::EntryControl);
duke@0 902 Node *init = main_head->init_trip();
duke@0 903 Node *incr = main_end ->incr();
duke@0 904 Node *limit = main_end ->limit();
duke@0 905 Node *stride = main_end ->stride();
duke@0 906 Node *cmp = main_end ->cmp_node();
duke@0 907 BoolTest::mask b_test = main_end->test_trip();
duke@0 908
duke@0 909 // Need only 1 user of 'bol' because I will be hacking the loop bounds.
duke@0 910 Node *bol = main_end->in(CountedLoopEndNode::TestValue);
duke@0 911 if( bol->outcnt() != 1 ) {
duke@0 912 bol = bol->clone();
duke@0 913 register_new_node(bol,main_end->in(CountedLoopEndNode::TestControl));
duke@0 914 _igvn.hash_delete(main_end);
duke@0 915 main_end->set_req(CountedLoopEndNode::TestValue, bol);
duke@0 916 }
duke@0 917 // Need only 1 user of 'cmp' because I will be hacking the loop bounds.
duke@0 918 if( cmp->outcnt() != 1 ) {
duke@0 919 cmp = cmp->clone();
duke@0 920 register_new_node(cmp,main_end->in(CountedLoopEndNode::TestControl));
duke@0 921 _igvn.hash_delete(bol);
duke@0 922 bol->set_req(1, cmp);
duke@0 923 }
duke@0 924
duke@0 925 //------------------------------
duke@0 926 // Step A: Create Post-Loop.
duke@0 927 Node* main_exit = main_end->proj_out(false);
duke@0 928 assert( main_exit->Opcode() == Op_IfFalse, "" );
duke@0 929 int dd_main_exit = dom_depth(main_exit);
duke@0 930
duke@0 931 // Step A1: Clone the loop body. The clone becomes the post-loop. The main
duke@0 932 // loop pre-header illegally has 2 control users (old & new loops).
duke@0 933 clone_loop( loop, old_new, dd_main_exit );
duke@0 934 assert( old_new[main_end ->_idx]->Opcode() == Op_CountedLoopEnd, "" );
duke@0 935 CountedLoopNode *post_head = old_new[main_head->_idx]->as_CountedLoop();
duke@0 936 post_head->set_post_loop(main_head);
duke@0 937
kvn@400 938 // Reduce the post-loop trip count.
kvn@400 939 CountedLoopEndNode* post_end = old_new[main_end ->_idx]->as_CountedLoopEnd();
kvn@400 940 post_end->_prob = PROB_FAIR;
kvn@400 941
duke@0 942 // Build the main-loop normal exit.
kvn@3680 943 IfFalseNode *new_main_exit = new (C) IfFalseNode(main_end);
duke@0 944 _igvn.register_new_node_with_optimizer( new_main_exit );
duke@0 945 set_idom(new_main_exit, main_end, dd_main_exit );
duke@0 946 set_loop(new_main_exit, loop->_parent);
duke@0 947
duke@0 948 // Step A2: Build a zero-trip guard for the post-loop. After leaving the
duke@0 949 // main-loop, the post-loop may not execute at all. We 'opaque' the incr
duke@0 950 // (the main-loop trip-counter exit value) because we will be changing
duke@0 951 // the exit value (via unrolling) so we cannot constant-fold away the zero
duke@0 952 // trip guard until all unrolling is done.
kvn@3680 953 Node *zer_opaq = new (C) Opaque1Node(C, incr);
kvn@3680 954 Node *zer_cmp = new (C) CmpINode( zer_opaq, limit );
kvn@3680 955 Node *zer_bol = new (C) BoolNode( zer_cmp, b_test );
duke@0 956 register_new_node( zer_opaq, new_main_exit );
duke@0 957 register_new_node( zer_cmp , new_main_exit );
duke@0 958 register_new_node( zer_bol , new_main_exit );
duke@0 959
duke@0 960 // Build the IfNode
kvn@3680 961 IfNode *zer_iff = new (C) IfNode( new_main_exit, zer_bol, PROB_FAIR, COUNT_UNKNOWN );
duke@0 962 _igvn.register_new_node_with_optimizer( zer_iff );
duke@0 963 set_idom(zer_iff, new_main_exit, dd_main_exit);
duke@0 964 set_loop(zer_iff, loop->_parent);
duke@0 965
duke@0 966 // Plug in the false-path, taken if we need to skip post-loop
kvn@3412 967 _igvn.replace_input_of(main_exit, 0, zer_iff);
duke@0 968 set_idom(main_exit, zer_iff, dd_main_exit);
duke@0 969 set_idom(main_exit->unique_out(), zer_iff, dd_main_exit);
duke@0 970 // Make the true-path, must enter the post loop
kvn@3680 971 Node *zer_taken = new (C) IfTrueNode( zer_iff );
duke@0 972 _igvn.register_new_node_with_optimizer( zer_taken );
duke@0 973 set_idom(zer_taken, zer_iff, dd_main_exit);
duke@0 974 set_loop(zer_taken, loop->_parent);
duke@0 975 // Plug in the true path
duke@0 976 _igvn.hash_delete( post_head );
duke@0 977 post_head->set_req(LoopNode::EntryControl, zer_taken);
duke@0 978 set_idom(post_head, zer_taken, dd_main_exit);
duke@0 979
kvn@2550 980 Arena *a = Thread::current()->resource_area();
kvn@2550 981 VectorSet visited(a);
kvn@2550 982 Node_Stack clones(a, main_head->back_control()->outcnt());
duke@0 983 // Step A3: Make the fall-in values to the post-loop come from the
duke@0 984 // fall-out values of the main-loop.
duke@0 985 for (DUIterator_Fast imax, i = main_head->fast_outs(imax); i < imax; i++) {
duke@0 986 Node* main_phi = main_head->fast_out(i);
duke@0 987 if( main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() >0 ) {
duke@0 988 Node *post_phi = old_new[main_phi->_idx];
duke@0 989 Node *fallmain = clone_up_backedge_goo(main_head->back_control(),
duke@0 990 post_head->init_control(),
kvn@2550 991 main_phi->in(LoopNode::LoopBackControl),
kvn@2550 992 visited, clones);
duke@0 993 _igvn.hash_delete(post_phi);
duke@0 994 post_phi->set_req( LoopNode::EntryControl, fallmain );
duke@0 995 }
duke@0 996 }
duke@0 997
duke@0 998 // Update local caches for next stanza
duke@0 999 main_exit = new_main_exit;
duke@0 1000
duke@0 1001
duke@0 1002 //------------------------------
duke@0 1003 // Step B: Create Pre-Loop.
duke@0 1004
duke@0 1005 // Step B1: Clone the loop body. The clone becomes the pre-loop. The main
duke@0 1006 // loop pre-header illegally has 2 control users (old & new loops).
duke@0 1007 clone_loop( loop, old_new, dd_main_head );
duke@0 1008 CountedLoopNode* pre_head = old_new[main_head->_idx]->as_CountedLoop();
duke@0 1009 CountedLoopEndNode* pre_end = old_new[main_end ->_idx]->as_CountedLoopEnd();
duke@0 1010 pre_head->set_pre_loop(main_head);
duke@0 1011 Node *pre_incr = old_new[incr->_idx];
duke@0 1012
kvn@400 1013 // Reduce the pre-loop trip count.
kvn@400 1014 pre_end->_prob = PROB_FAIR;
kvn@400 1015
duke@0 1016 // Find the pre-loop normal exit.
duke@0 1017 Node* pre_exit = pre_end->proj_out(false);
duke@0 1018 assert( pre_exit->Opcode() == Op_IfFalse, "" );
kvn@3680 1019 IfFalseNode *new_pre_exit = new (C) IfFalseNode(pre_end);
duke@0 1020 _igvn.register_new_node_with_optimizer( new_pre_exit );
duke@0 1021 set_idom(new_pre_exit, pre_end, dd_main_head);
duke@0 1022 set_loop(new_pre_exit, loop->_parent);
duke@0 1023
duke@0 1024 // Step B2: Build a zero-trip guard for the main-loop. After leaving the
duke@0 1025 // pre-loop, the main-loop may not execute at all. Later in life this
duke@0 1026 // zero-trip guard will become the minimum-trip guard when we unroll
duke@0 1027 // the main-loop.
kvn@3680 1028 Node *min_opaq = new (C) Opaque1Node(C, limit);
kvn@3680 1029 Node *min_cmp = new (C) CmpINode( pre_incr, min_opaq );
kvn@3680 1030 Node *min_bol = new (C) BoolNode( min_cmp, b_test );
duke@0 1031 register_new_node( min_opaq, new_pre_exit );
duke@0 1032 register_new_node( min_cmp , new_pre_exit );
duke@0 1033 register_new_node( min_bol , new_pre_exit );
duke@0 1034
kvn@400 1035 // Build the IfNode (assume the main-loop is executed always).
kvn@3680 1036 IfNode *min_iff = new (C) IfNode( new_pre_exit, min_bol, PROB_ALWAYS, COUNT_UNKNOWN );
duke@0 1037 _igvn.register_new_node_with_optimizer( min_iff );
duke@0 1038 set_idom(min_iff, new_pre_exit, dd_main_head);
duke@0 1039 set_loop(min_iff, loop->_parent);
duke@0 1040
duke@0 1041 // Plug in the false-path, taken if we need to skip main-loop
duke@0 1042 _igvn.hash_delete( pre_exit );
duke@0 1043 pre_exit->set_req(0, min_iff);
duke@0 1044 set_idom(pre_exit, min_iff, dd_main_head);
duke@0 1045 set_idom(pre_exit->unique_out(), min_iff, dd_main_head);
duke@0 1046 // Make the true-path, must enter the main loop
kvn@3680 1047 Node *min_taken = new (C) IfTrueNode( min_iff );
duke@0 1048 _igvn.register_new_node_with_optimizer( min_taken );
duke@0 1049 set_idom(min_taken, min_iff, dd_main_head);
duke@0 1050 set_loop(min_taken, loop->_parent);
duke@0 1051 // Plug in the true path
duke@0 1052 _igvn.hash_delete( main_head );
duke@0 1053 main_head->set_req(LoopNode::EntryControl, min_taken);
duke@0 1054 set_idom(main_head, min_taken, dd_main_head);
duke@0 1055
kvn@2550 1056 visited.Clear();
kvn@2550 1057 clones.clear();
duke@0 1058 // Step B3: Make the fall-in values to the main-loop come from the
duke@0 1059 // fall-out values of the pre-loop.
duke@0 1060 for (DUIterator_Fast i2max, i2 = main_head->fast_outs(i2max); i2 < i2max; i2++) {
duke@0 1061 Node* main_phi = main_head->fast_out(i2);
duke@0 1062 if( main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() > 0 ) {
duke@0 1063 Node *pre_phi = old_new[main_phi->_idx];
duke@0 1064 Node *fallpre = clone_up_backedge_goo(pre_head->back_control(),
duke@0 1065 main_head->init_control(),
kvn@2550 1066 pre_phi->in(LoopNode::LoopBackControl),
kvn@2550 1067 visited, clones);
duke@0 1068 _igvn.hash_delete(main_phi);
duke@0 1069 main_phi->set_req( LoopNode::EntryControl, fallpre );
duke@0 1070 }
duke@0 1071 }
duke@0 1072
duke@0 1073 // Step B4: Shorten the pre-loop to run only 1 iteration (for now).
duke@0 1074 // RCE and alignment may change this later.
duke@0 1075 Node *cmp_end = pre_end->cmp_node();
duke@0 1076 assert( cmp_end->in(2) == limit, "" );
kvn@3680 1077 Node *pre_limit = new (C) AddINode( init, stride );
duke@0 1078
duke@0 1079 // Save the original loop limit in this Opaque1 node for
duke@0 1080 // use by range check elimination.
kvn@3680 1081 Node *pre_opaq = new (C) Opaque1Node(C, pre_limit, limit);
duke@0 1082
duke@0 1083 register_new_node( pre_limit, pre_head->in(0) );
duke@0 1084 register_new_node( pre_opaq , pre_head->in(0) );
duke@0 1085
duke@0 1086 // Since no other users of pre-loop compare, I can hack limit directly
duke@0 1087 assert( cmp_end->outcnt() == 1, "no other users" );
duke@0 1088 _igvn.hash_delete(cmp_end);
duke@0 1089 cmp_end->set_req(2, peel_only ? pre_limit : pre_opaq);
duke@0 1090
duke@0 1091 // Special case for not-equal loop bounds:
duke@0 1092 // Change pre loop test, main loop test, and the
duke@0 1093 // main loop guard test to use lt or gt depending on stride
duke@0 1094 // direction:
duke@0 1095 // positive stride use <
duke@0 1096 // negative stride use >
kvn@2544 1097 //
kvn@2544 1098 // not-equal test is kept for post loop to handle case
kvn@2544 1099 // when init > limit when stride > 0 (and reverse).
duke@0 1100
duke@0 1101 if (pre_end->in(CountedLoopEndNode::TestValue)->as_Bool()->_test._test == BoolTest::ne) {
duke@0 1102
duke@0 1103 BoolTest::mask new_test = (main_end->stride_con() > 0) ? BoolTest::lt : BoolTest::gt;
duke@0 1104 // Modify pre loop end condition
duke@0 1105 Node* pre_bol = pre_end->in(CountedLoopEndNode::TestValue)->as_Bool();
kvn@3680 1106 BoolNode* new_bol0 = new (C) BoolNode(pre_bol->in(1), new_test);
duke@0 1107 register_new_node( new_bol0, pre_head->in(0) );
duke@0 1108 _igvn.hash_delete(pre_end);
duke@0 1109 pre_end->set_req(CountedLoopEndNode::TestValue, new_bol0);
duke@0 1110 // Modify main loop guard condition
duke@0 1111 assert(min_iff->in(CountedLoopEndNode::TestValue) == min_bol, "guard okay");
kvn@3680 1112 BoolNode* new_bol1 = new (C) BoolNode(min_bol->in(1), new_test);
duke@0 1113 register_new_node( new_bol1, new_pre_exit );
duke@0 1114 _igvn.hash_delete(min_iff);
duke@0 1115 min_iff->set_req(CountedLoopEndNode::TestValue, new_bol1);
duke@0 1116 // Modify main loop end condition
duke@0 1117 BoolNode* main_bol = main_end->in(CountedLoopEndNode::TestValue)->as_Bool();
kvn@3680 1118 BoolNode* new_bol2 = new (C) BoolNode(main_bol->in(1), new_test);
duke@0 1119 register_new_node( new_bol2, main_end->in(CountedLoopEndNode::TestControl) );
duke@0 1120 _igvn.hash_delete(main_end);
duke@0 1121 main_end->set_req(CountedLoopEndNode::TestValue, new_bol2);
duke@0 1122 }
duke@0 1123
duke@0 1124 // Flag main loop
duke@0 1125 main_head->set_main_loop();
duke@0 1126 if( peel_only ) main_head->set_main_no_pre_loop();
duke@0 1127
kvn@2442 1128 // Subtract a trip count for the pre-loop.
kvn@2442 1129 main_head->set_trip_count(main_head->trip_count() - 1);
kvn@2442 1130
duke@0 1131 // It's difficult to be precise about the trip-counts
duke@0 1132 // for the pre/post loops. They are usually very short,
duke@0 1133 // so guess that 4 trips is a reasonable value.
duke@0 1134 post_head->set_profile_trip_cnt(4.0);
duke@0 1135 pre_head->set_profile_trip_cnt(4.0);
duke@0 1136
duke@0 1137 // Now force out all loop-invariant dominating tests. The optimizer
duke@0 1138 // finds some, but we _know_ they are all useless.
duke@0 1139 peeled_dom_test_elim(loop,old_new);
duke@0 1140 }
duke@0 1141
duke@0 1142 //------------------------------is_invariant-----------------------------
duke@0 1143 // Return true if n is invariant
duke@0 1144 bool IdealLoopTree::is_invariant(Node* n) const {
cfang@1172 1145 Node *n_c = _phase->has_ctrl(n) ? _phase->get_ctrl(n) : n;
duke@0 1146 if (n_c->is_top()) return false;
duke@0 1147 return !is_member(_phase->get_loop(n_c));
duke@0 1148 }
duke@0 1149
duke@0 1150
duke@0 1151 //------------------------------do_unroll--------------------------------------
duke@0 1152 // Unroll the loop body one step - make each trip do 2 iterations.
duke@0 1153 void PhaseIdealLoop::do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip ) {
kvn@2230 1154 assert(LoopUnrollLimit, "");
kvn@2230 1155 CountedLoopNode *loop_head = loop->_head->as_CountedLoop();
kvn@2230 1156 CountedLoopEndNode *loop_end = loop_head->loopexit();
kvn@2230 1157 assert(loop_end, "");
duke@0 1158 #ifndef PRODUCT
kvn@2230 1159 if (PrintOpto && VerifyLoopOptimizations) {
duke@0 1160 tty->print("Unrolling ");
duke@0 1161 loop->dump_head();
kvn@2230 1162 } else if (TraceLoopOpts) {
kvn@2312 1163 if (loop_head->trip_count() < (uint)LoopUnrollLimit) {
kvn@2442 1164 tty->print("Unroll %d(%2d) ", loop_head->unrolled_count()*2, loop_head->trip_count());
kvn@2300 1165 } else {
kvn@2442 1166 tty->print("Unroll %d ", loop_head->unrolled_count()*2);
kvn@2300 1167 }
kvn@2230 1168 loop->dump_head();
duke@0 1169 }
duke@0 1170 #endif
duke@0 1171
duke@0 1172 // Remember loop node count before unrolling to detect
duke@0 1173 // if rounds of unroll,optimize are making progress
duke@0 1174 loop_head->set_node_count_before_unroll(loop->_body.size());
duke@0 1175
duke@0 1176 Node *ctrl = loop_head->in(LoopNode::EntryControl);
duke@0 1177 Node *limit = loop_head->limit();
duke@0 1178 Node *init = loop_head->init_trip();
kvn@2230 1179 Node *stride = loop_head->stride();
duke@0 1180
duke@0 1181 Node *opaq = NULL;
kvn@2442 1182 if (adjust_min_trip) { // If not maximally unrolling, need adjustment
kvn@2442 1183 // Search for zero-trip guard.
duke@0 1184 assert( loop_head->is_main_loop(), "" );
duke@0 1185 assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" );
duke@0 1186 Node *iff = ctrl->in(0);
duke@0 1187 assert( iff->Opcode() == Op_If, "" );
duke@0 1188 Node *bol = iff->in(1);
duke@0 1189 assert( bol->Opcode() == Op_Bool, "" );
duke@0 1190 Node *cmp = bol->in(1);
duke@0 1191 assert( cmp->Opcode() == Op_CmpI, "" );
duke@0 1192 opaq = cmp->in(2);
kvn@2442 1193 // Occasionally it's possible for a zero-trip guard Opaque1 node to be
duke@0 1194 // optimized away and then another round of loop opts attempted.
duke@0 1195 // We can not optimize this particular loop in that case.
kvn@2442 1196 if (opaq->Opcode() != Op_Opaque1)
kvn@2442 1197 return; // Cannot find zero-trip guard! Bail out!
kvn@2442 1198 // Zero-trip test uses an 'opaque' node which is not shared.
kvn@2442 1199 assert(opaq->outcnt() == 1 && opaq->in(1) == limit, "");
duke@0 1200 }
duke@0 1201
duke@0 1202 C->set_major_progress();
duke@0 1203
kvn@2442 1204 Node* new_limit = NULL;
kvn@2442 1205 if (UnrollLimitCheck) {
kvn@2442 1206 int stride_con = stride->get_int();
kvn@2442 1207 int stride_p = (stride_con > 0) ? stride_con : -stride_con;
kvn@2442 1208 uint old_trip_count = loop_head->trip_count();
kvn@2442 1209 // Verify that unroll policy result is still valid.
kvn@2442 1210 assert(old_trip_count > 1 &&
kvn@2442 1211 (!adjust_min_trip || stride_p <= (1<<3)*loop_head->unrolled_count()), "sanity");
duke@0 1212
kvn@2442 1213 // Adjust loop limit to keep valid iterations number after unroll.
kvn@2442 1214 // Use (limit - stride) instead of (((limit - init)/stride) & (-2))*stride
kvn@2442 1215 // which may overflow.
kvn@2442 1216 if (!adjust_min_trip) {
kvn@2442 1217 assert(old_trip_count > 1 && (old_trip_count & 1) == 0,
kvn@2442 1218 "odd trip count for maximally unroll");
kvn@2442 1219 // Don't need to adjust limit for maximally unroll since trip count is even.
kvn@2442 1220 } else if (loop_head->has_exact_trip_count() && init->is_Con()) {
kvn@2442 1221 // Loop's limit is constant. Loop's init could be constant when pre-loop
kvn@2442 1222 // become peeled iteration.
vlivanov@3722 1223 jlong init_con = init->get_int();
kvn@2442 1224 // We can keep old loop limit if iterations count stays the same:
kvn@2442 1225 // old_trip_count == new_trip_count * 2
kvn@2442 1226 // Note: since old_trip_count >= 2 then new_trip_count >= 1
kvn@2442 1227 // so we also don't need to adjust zero trip test.
vlivanov@3722 1228 jlong limit_con = limit->get_int();
kvn@2442 1229 // (stride_con*2) not overflow since stride_con <= 8.
kvn@2442 1230 int new_stride_con = stride_con * 2;
kvn@2442 1231 int stride_m = new_stride_con - (stride_con > 0 ? 1 : -1);
vlivanov@3722 1232 jlong trip_count = (limit_con - init_con + stride_m)/new_stride_con;
kvn@2442 1233 // New trip count should satisfy next conditions.
kvn@2442 1234 assert(trip_count > 0 && (julong)trip_count < (julong)max_juint/2, "sanity");
kvn@2442 1235 uint new_trip_count = (uint)trip_count;
kvn@2442 1236 adjust_min_trip = (old_trip_count != new_trip_count*2);
kvn@2442 1237 }
duke@0 1238
kvn@2442 1239 if (adjust_min_trip) {
kvn@2442 1240 // Step 2: Adjust the trip limit if it is called for.
kvn@2442 1241 // The adjustment amount is -stride. Need to make sure if the
kvn@2442 1242 // adjustment underflows or overflows, then the main loop is skipped.
kvn@2442 1243 Node* cmp = loop_end->cmp_node();
kvn@2442 1244 assert(cmp->in(2) == limit, "sanity");
kvn@2442 1245 assert(opaq != NULL && opaq->in(1) == limit, "sanity");
duke@0 1246
kvn@2442 1247 // Verify that policy_unroll result is still valid.
kvn@2442 1248 const TypeInt* limit_type = _igvn.type(limit)->is_int();
kvn@2442 1249 assert(stride_con > 0 && ((limit_type->_hi - stride_con) < limit_type->_hi) ||
kvn@2442 1250 stride_con < 0 && ((limit_type->_lo - stride_con) > limit_type->_lo), "sanity");
duke@0 1251
kvn@2442 1252 if (limit->is_Con()) {
kvn@2442 1253 // The check in policy_unroll and the assert above guarantee
kvn@2442 1254 // no underflow if limit is constant.
kvn@2442 1255 new_limit = _igvn.intcon(limit->get_int() - stride_con);
kvn@2442 1256 set_ctrl(new_limit, C->root());
kvn@2442 1257 } else {
kvn@2445 1258 // Limit is not constant.
kvn@2464 1259 if (loop_head->unrolled_count() == 1) { // only for first unroll
kvn@2445 1260 // Separate limit by Opaque node in case it is an incremented
kvn@2445 1261 // variable from previous loop to avoid using pre-incremented
kvn@2445 1262 // value which could increase register pressure.
kvn@2445 1263 // Otherwise reorg_offsets() optimization will create a separate
kvn@2445 1264 // Opaque node for each use of trip-counter and as result
kvn@2445 1265 // zero trip guard limit will be different from loop limit.
kvn@2445 1266 assert(has_ctrl(opaq), "should have it");
kvn@2445 1267 Node* opaq_ctrl = get_ctrl(opaq);
kvn@3680 1268 limit = new (C) Opaque2Node( C, limit );
kvn@2445 1269 register_new_node( limit, opaq_ctrl );
kvn@2445 1270 }
kvn@2442 1271 if (stride_con > 0 && ((limit_type->_lo - stride_con) < limit_type->_lo) ||
kvn@2442 1272 stride_con < 0 && ((limit_type->_hi - stride_con) > limit_type->_hi)) {
kvn@2442 1273 // No underflow.
kvn@3680 1274 new_limit = new (C) SubINode(limit, stride);
kvn@2442 1275 } else {
kvn@2442 1276 // (limit - stride) may underflow.
kvn@2442 1277 // Clamp the adjustment value with MININT or MAXINT:
kvn@2442 1278 //
kvn@2442 1279 // new_limit = limit-stride
kvn@2442 1280 // if (stride > 0)
kvn@2442 1281 // new_limit = (limit < new_limit) ? MININT : new_limit;
kvn@2442 1282 // else
kvn@2442 1283 // new_limit = (limit > new_limit) ? MAXINT : new_limit;
kvn@2442 1284 //
kvn@2442 1285 BoolTest::mask bt = loop_end->test_trip();
kvn@2442 1286 assert(bt == BoolTest::lt || bt == BoolTest::gt, "canonical test is expected");
kvn@2442 1287 Node* adj_max = _igvn.intcon((stride_con > 0) ? min_jint : max_jint);
kvn@2442 1288 set_ctrl(adj_max, C->root());
kvn@2442 1289 Node* old_limit = NULL;
kvn@2442 1290 Node* adj_limit = NULL;
kvn@2442 1291 Node* bol = limit->is_CMove() ? limit->in(CMoveNode::Condition) : NULL;
kvn@2442 1292 if (loop_head->unrolled_count() > 1 &&
kvn@2442 1293 limit->is_CMove() && limit->Opcode() == Op_CMoveI &&
kvn@2442 1294 limit->in(CMoveNode::IfTrue) == adj_max &&
kvn@2442 1295 bol->as_Bool()->_test._test == bt &&
kvn@2442 1296 bol->in(1)->Opcode() == Op_CmpI &&
kvn@2442 1297 bol->in(1)->in(2) == limit->in(CMoveNode::IfFalse)) {
kvn@2442 1298 // Loop was unrolled before.
kvn@2442 1299 // Optimize the limit to avoid nested CMove:
kvn@2442 1300 // use original limit as old limit.
kvn@2442 1301 old_limit = bol->in(1)->in(1);
kvn@2442 1302 // Adjust previous adjusted limit.
kvn@2442 1303 adj_limit = limit->in(CMoveNode::IfFalse);
kvn@3680 1304 adj_limit = new (C) SubINode(adj_limit, stride);
kvn@2442 1305 } else {
kvn@2442 1306 old_limit = limit;
kvn@3680 1307 adj_limit = new (C) SubINode(limit, stride);
kvn@2442 1308 }
kvn@2442 1309 assert(old_limit != NULL && adj_limit != NULL, "");
kvn@2442 1310 register_new_node( adj_limit, ctrl ); // adjust amount
kvn@3680 1311 Node* adj_cmp = new (C) CmpINode(old_limit, adj_limit);
kvn@2442 1312 register_new_node( adj_cmp, ctrl );
kvn@3680 1313 Node* adj_bool = new (C) BoolNode(adj_cmp, bt);
kvn@2442 1314 register_new_node( adj_bool, ctrl );
kvn@3680 1315 new_limit = new (C) CMoveINode(adj_bool, adj_limit, adj_max, TypeInt::INT);
kvn@2442 1316 }
kvn@2442 1317 register_new_node(new_limit, ctrl);
kvn@2442 1318 }
kvn@2442 1319 assert(new_limit != NULL, "");
kvn@2445 1320 // Replace in loop test.
kvn@2494 1321 assert(loop_end->in(1)->in(1) == cmp, "sanity");
kvn@2494 1322 if (cmp->outcnt() == 1 && loop_end->in(1)->outcnt() == 1) {
kvn@2494 1323 // Don't need to create new test since only one user.
kvn@2494 1324 _igvn.hash_delete(cmp);
kvn@2494 1325 cmp->set_req(2, new_limit);
kvn@2494 1326 } else {
kvn@2494 1327 // Create new test since it is shared.
kvn@2494 1328 Node* ctrl2 = loop_end->in(0);
kvn@2494 1329 Node* cmp2 = cmp->clone();
kvn@2494 1330 cmp2->set_req(2, new_limit);
kvn@2494 1331 register_new_node(cmp2, ctrl2);
kvn@2494 1332 Node* bol2 = loop_end->in(1)->clone();
kvn@2494 1333 bol2->set_req(1, cmp2);
kvn@2494 1334 register_new_node(bol2, ctrl2);
kvn@2494 1335 _igvn.hash_delete(loop_end);
kvn@2494 1336 loop_end->set_req(1, bol2);
kvn@2494 1337 }
kvn@2445 1338 // Step 3: Find the min-trip test guaranteed before a 'main' loop.
kvn@2445 1339 // Make it a 1-trip test (means at least 2 trips).
kvn@2442 1340
kvn@2445 1341 // Guard test uses an 'opaque' node which is not shared. Hence I
kvn@2445 1342 // can edit it's inputs directly. Hammer in the new limit for the
kvn@2445 1343 // minimum-trip guard.
kvn@2445 1344 assert(opaq->outcnt() == 1, "");
kvn@2445 1345 _igvn.hash_delete(opaq);
kvn@2445 1346 opaq->set_req(1, new_limit);
kvn@2442 1347 }
kvn@2442 1348
kvn@2442 1349 // Adjust max trip count. The trip count is intentionally rounded
kvn@2442 1350 // down here (e.g. 15-> 7-> 3-> 1) because if we unwittingly over-unroll,
kvn@2442 1351 // the main, unrolled, part of the loop will never execute as it is protected
kvn@2442 1352 // by the min-trip test. See bug 4834191 for a case where we over-unrolled
kvn@2442 1353 // and later determined that part of the unrolled loop was dead.
kvn@2442 1354 loop_head->set_trip_count(old_trip_count / 2);
kvn@2442 1355
kvn@2442 1356 // Double the count of original iterations in the unrolled loop body.
kvn@2442 1357 loop_head->double_unrolled_count();
kvn@2442 1358
kvn@2442 1359 } else { // LoopLimitCheck
kvn@2442 1360
kvn@2442 1361 // Adjust max trip count. The trip count is intentionally rounded
kvn@2442 1362 // down here (e.g. 15-> 7-> 3-> 1) because if we unwittingly over-unroll,
kvn@2442 1363 // the main, unrolled, part of the loop will never execute as it is protected
kvn@2442 1364 // by the min-trip test. See bug 4834191 for a case where we over-unrolled
kvn@2442 1365 // and later determined that part of the unrolled loop was dead.
kvn@2442 1366 loop_head->set_trip_count(loop_head->trip_count() / 2);
kvn@2442 1367
kvn@2442 1368 // Double the count of original iterations in the unrolled loop body.
kvn@2442 1369 loop_head->double_unrolled_count();
kvn@2442 1370
kvn@2442 1371 // -----------
kvn@2442 1372 // Step 2: Cut back the trip counter for an unroll amount of 2.
kvn@2442 1373 // Loop will normally trip (limit - init)/stride_con. Since it's a
kvn@2442 1374 // CountedLoop this is exact (stride divides limit-init exactly).
kvn@2442 1375 // We are going to double the loop body, so we want to knock off any
kvn@2442 1376 // odd iteration: (trip_cnt & ~1). Then back compute a new limit.
kvn@3680 1377 Node *span = new (C) SubINode( limit, init );
kvn@2442 1378 register_new_node( span, ctrl );
kvn@3680 1379 Node *trip = new (C) DivINode( 0, span, stride );
kvn@2442 1380 register_new_node( trip, ctrl );
kvn@2442 1381 Node *mtwo = _igvn.intcon(-2);
kvn@2442 1382 set_ctrl(mtwo, C->root());
kvn@3680 1383 Node *rond = new (C) AndINode( trip, mtwo );
kvn@2442 1384 register_new_node( rond, ctrl );
kvn@3680 1385 Node *spn2 = new (C) MulINode( rond, stride );
kvn@2442 1386 register_new_node( spn2, ctrl );
kvn@3680 1387 new_limit = new (C) AddINode( spn2, init );
kvn@2442 1388 register_new_node( new_limit, ctrl );
kvn@2442 1389
kvn@2442 1390 // Hammer in the new limit
kvn@2442 1391 Node *ctrl2 = loop_end->in(0);
kvn@3680 1392 Node *cmp2 = new (C) CmpINode( loop_head->incr(), new_limit );
kvn@2442 1393 register_new_node( cmp2, ctrl2 );
kvn@3680 1394 Node *bol2 = new (C) BoolNode( cmp2, loop_end->test_trip() );
kvn@2442 1395 register_new_node( bol2, ctrl2 );
kvn@2442 1396 _igvn.hash_delete(loop_end);
kvn@2442 1397 loop_end->set_req(CountedLoopEndNode::TestValue, bol2);
kvn@2442 1398
kvn@2442 1399 // Step 3: Find the min-trip test guaranteed before a 'main' loop.
kvn@2442 1400 // Make it a 1-trip test (means at least 2 trips).
kvn@2442 1401 if( adjust_min_trip ) {
kvn@2442 1402 assert( new_limit != NULL, "" );
kvn@2442 1403 // Guard test uses an 'opaque' node which is not shared. Hence I
kvn@2442 1404 // can edit it's inputs directly. Hammer in the new limit for the
kvn@2442 1405 // minimum-trip guard.
kvn@2442 1406 assert( opaq->outcnt() == 1, "" );
kvn@2442 1407 _igvn.hash_delete(opaq);
kvn@2442 1408 opaq->set_req(1, new_limit);
kvn@2442 1409 }
kvn@2442 1410 } // LoopLimitCheck
duke@0 1411
duke@0 1412 // ---------
duke@0 1413 // Step 4: Clone the loop body. Move it inside the loop. This loop body
duke@0 1414 // represents the odd iterations; since the loop trips an even number of
duke@0 1415 // times its backedge is never taken. Kill the backedge.
duke@0 1416 uint dd = dom_depth(loop_head);
duke@0 1417 clone_loop( loop, old_new, dd );
duke@0 1418
duke@0 1419 // Make backedges of the clone equal to backedges of the original.
duke@0 1420 // Make the fall-in from the original come from the fall-out of the clone.
duke@0 1421 for (DUIterator_Fast jmax, j = loop_head->fast_outs(jmax); j < jmax; j++) {
duke@0 1422 Node* phi = loop_head->fast_out(j);
duke@0 1423 if( phi->is_Phi() && phi->in(0) == loop_head && phi->outcnt() > 0 ) {
duke@0 1424 Node *newphi = old_new[phi->_idx];
duke@0 1425 _igvn.hash_delete( phi );
duke@0 1426 _igvn.hash_delete( newphi );
duke@0 1427
duke@0 1428 phi ->set_req(LoopNode:: EntryControl, newphi->in(LoopNode::LoopBackControl));
duke@0 1429 newphi->set_req(LoopNode::LoopBackControl, phi ->in(LoopNode::LoopBackControl));
duke@0 1430 phi ->set_req(LoopNode::LoopBackControl, C->top());
duke@0 1431 }
duke@0 1432 }
duke@0 1433 Node *clone_head = old_new[loop_head->_idx];
duke@0 1434 _igvn.hash_delete( clone_head );
duke@0 1435 loop_head ->set_req(LoopNode:: EntryControl, clone_head->in(LoopNode::LoopBackControl));
duke@0 1436 clone_head->set_req(LoopNode::LoopBackControl, loop_head ->in(LoopNode::LoopBackControl));
duke@0 1437 loop_head ->set_req(LoopNode::LoopBackControl, C->top());
duke@0 1438 loop->_head = clone_head; // New loop header
duke@0 1439
duke@0 1440 set_idom(loop_head, loop_head ->in(LoopNode::EntryControl), dd);
duke@0 1441 set_idom(clone_head, clone_head->in(LoopNode::EntryControl), dd);
duke@0 1442
duke@0 1443 // Kill the clone's backedge
duke@0 1444 Node *newcle = old_new[loop_end->_idx];
duke@0 1445 _igvn.hash_delete( newcle );
duke@0 1446 Node *one = _igvn.intcon(1);
duke@0 1447 set_ctrl(one, C->root());
duke@0 1448 newcle->set_req(1, one);
duke@0 1449 // Force clone into same loop body
duke@0 1450 uint max = loop->_body.size();
duke@0 1451 for( uint k = 0; k < max; k++ ) {
duke@0 1452 Node *old = loop->_body.at(k);
duke@0 1453 Node *nnn = old_new[old->_idx];
duke@0 1454 loop->_body.push(nnn);
duke@0 1455 if (!has_ctrl(old))
duke@0 1456 set_loop(nnn, loop);
duke@0 1457 }
never@367 1458
never@367 1459 loop->record_for_igvn();
duke@0 1460 }
duke@0 1461
duke@0 1462 //------------------------------do_maximally_unroll----------------------------
duke@0 1463
duke@0 1464 void PhaseIdealLoop::do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new ) {
duke@0 1465 CountedLoopNode *cl = loop->_head->as_CountedLoop();
kvn@2442 1466 assert(cl->has_exact_trip_count(), "trip count is not exact");
kvn@2230 1467 assert(cl->trip_count() > 0, "");
kvn@2230 1468 #ifndef PRODUCT
kvn@2230 1469 if (TraceLoopOpts) {
kvn@2230 1470 tty->print("MaxUnroll %d ", cl->trip_count());
kvn@2230 1471 loop->dump_head();
kvn@2230 1472 }
kvn@2230 1473 #endif
duke@0 1474
duke@0 1475 // If loop is tripping an odd number of times, peel odd iteration
kvn@2230 1476 if ((cl->trip_count() & 1) == 1) {
kvn@2230 1477 do_peeling(loop, old_new);
duke@0 1478 }
duke@0 1479
duke@0 1480 // Now its tripping an even number of times remaining. Double loop body.
duke@0 1481 // Do not adjust pre-guards; they are not needed and do not exist.
kvn@2230 1482 if (cl->trip_count() > 0) {
kvn@2442 1483 assert((cl->trip_count() & 1) == 0, "missed peeling");
kvn@2230 1484 do_unroll(loop, old_new, false);
duke@0 1485 }
duke@0 1486 }
duke@0 1487
duke@0 1488 //------------------------------dominates_backedge---------------------------------
duke@0 1489 // Returns true if ctrl is executed on every complete iteration
duke@0 1490 bool IdealLoopTree::dominates_backedge(Node* ctrl) {
duke@0 1491 assert(ctrl->is_CFG(), "must be control");
duke@0 1492 Node* backedge = _head->as_Loop()->in(LoopNode::LoopBackControl);
duke@0 1493 return _phase->dom_lca_internal(ctrl, backedge) == ctrl;
duke@0 1494 }
duke@0 1495
kvn@2480 1496 //------------------------------adjust_limit-----------------------------------
kvn@2480 1497 // Helper function for add_constraint().
kvn@2480 1498 Node* PhaseIdealLoop::adjust_limit(int stride_con, Node * scale, Node *offset, Node *rc_limit, Node *loop_limit, Node *pre_ctrl) {
kvn@2480 1499 // Compute "I :: (limit-offset)/scale"
kvn@3680 1500 Node *con = new (C) SubINode(rc_limit, offset);
kvn@2480 1501 register_new_node(con, pre_ctrl);
kvn@3680 1502 Node *X = new (C) DivINode(0, con, scale);
kvn@2480 1503 register_new_node(X, pre_ctrl);
kvn@2480 1504
kvn@2480 1505 // Adjust loop limit
kvn@2480 1506 loop_limit = (stride_con > 0)
kvn@3680 1507 ? (Node*)(new (C) MinINode(loop_limit, X))
kvn@3680 1508 : (Node*)(new (C) MaxINode(loop_limit, X));
kvn@2480 1509 register_new_node(loop_limit, pre_ctrl);
kvn@2480 1510 return loop_limit;
kvn@2480 1511 }
kvn@2480 1512
duke@0 1513 //------------------------------add_constraint---------------------------------
kvn@2442 1514 // Constrain the main loop iterations so the conditions:
kvn@2442 1515 // low_limit <= scale_con * I + offset < upper_limit
duke@0 1516 // always holds true. That is, either increase the number of iterations in
duke@0 1517 // the pre-loop or the post-loop until the condition holds true in the main
duke@0 1518 // loop. Stride, scale, offset and limit are all loop invariant. Further,
duke@0 1519 // stride and scale are constants (offset and limit often are).
kvn@2442 1520 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@0 1521 // For positive stride, the pre-loop limit always uses a MAX function
duke@0 1522 // and the main loop a MIN function. For negative stride these are
duke@0 1523 // reversed.
duke@0 1524
duke@0 1525 // Also for positive stride*scale the affine function is increasing, so the
duke@0 1526 // pre-loop must check for underflow and the post-loop for overflow.
duke@0 1527 // Negative stride*scale reverses this; pre-loop checks for overflow and
duke@0 1528 // post-loop for underflow.
kvn@2480 1529
kvn@2480 1530 Node *scale = _igvn.intcon(scale_con);
kvn@2480 1531 set_ctrl(scale, C->root());
kvn@2480 1532
kvn@2480 1533 if ((stride_con^scale_con) >= 0) { // Use XOR to avoid overflow
kvn@2442 1534 // The overflow limit: scale*I+offset < upper_limit
kvn@2442 1535 // For main-loop compute
kvn@2442 1536 // ( if (scale > 0) /* and stride > 0 */
kvn@2442 1537 // I < (upper_limit-offset)/scale
kvn@2442 1538 // else /* scale < 0 and stride < 0 */
kvn@2442 1539 // I > (upper_limit-offset)/scale
kvn@2442 1540 // )
kvn@2442 1541 //
kvn@2480 1542 // (upper_limit-offset) may overflow or underflow.
kvn@2442 1543 // But it is fine since main loop will either have
kvn@2442 1544 // less iterations or will be skipped in such case.
kvn@2480 1545 *main_limit = adjust_limit(stride_con, scale, offset, upper_limit, *main_limit, pre_ctrl);
duke@0 1546
kvn@2442 1547 // The underflow limit: low_limit <= scale*I+offset.
kvn@2442 1548 // For pre-loop compute
kvn@2442 1549 // NOT(scale*I+offset >= low_limit)
kvn@2442 1550 // scale*I+offset < low_limit
kvn@2442 1551 // ( if (scale > 0) /* and stride > 0 */
kvn@2442 1552 // I < (low_limit-offset)/scale
kvn@2442 1553 // else /* scale < 0 and stride < 0 */
kvn@2442 1554 // I > (low_limit-offset)/scale
kvn@2442 1555 // )
kvn@2442 1556
kvn@2442 1557 if (low_limit->get_int() == -max_jint) {
kvn@2442 1558 if (!RangeLimitCheck) return;
kvn@2442 1559 // We need this guard when scale*pre_limit+offset >= limit
kvn@2480 1560 // due to underflow. So we need execute pre-loop until
kvn@2480 1561 // scale*I+offset >= min_int. But (min_int-offset) will
kvn@2480 1562 // underflow when offset > 0 and X will be > original_limit
kvn@2480 1563 // when stride > 0. To avoid it we replace positive offset with 0.
kvn@2480 1564 //
kvn@2480 1565 // Also (min_int+1 == -max_int) is used instead of min_int here
kvn@2480 1566 // to avoid problem with scale == -1 (min_int/(-1) == min_int).
kvn@2442 1567 Node* shift = _igvn.intcon(31);
kvn@2442 1568 set_ctrl(shift, C->root());
kvn@3680 1569 Node* sign = new (C) RShiftINode(offset, shift);
kvn@2480 1570 register_new_node(sign, pre_ctrl);
kvn@3680 1571 offset = new (C) AndINode(offset, sign);
kvn@2442 1572 register_new_node(offset, pre_ctrl);
kvn@2442 1573 } else {
kvn@2442 1574 assert(low_limit->get_int() == 0, "wrong low limit for range check");
kvn@2442 1575 // The only problem we have here when offset == min_int
kvn@2480 1576 // since (0-min_int) == min_int. It may be fine for stride > 0
kvn@2480 1577 // but for stride < 0 X will be < original_limit. To avoid it
kvn@2480 1578 // max(pre_limit, original_limit) is used in do_range_check().
kvn@2442 1579 }
kvn@2480 1580 // Pass (-stride) to indicate pre_loop_cond = NOT(main_loop_cond);
kvn@2480 1581 *pre_limit = adjust_limit((-stride_con), scale, offset, low_limit, *pre_limit, pre_ctrl);
kvn@2442 1582
kvn@2442 1583 } else { // stride_con*scale_con < 0
kvn@2442 1584 // For negative stride*scale pre-loop checks for overflow and
kvn@2442 1585 // post-loop for underflow.
kvn@2442 1586 //
kvn@2442 1587 // The overflow limit: scale*I+offset < upper_limit
kvn@2442 1588 // For pre-loop compute
kvn@2442 1589 // NOT(scale*I+offset < upper_limit)
kvn@2442 1590 // scale*I+offset >= upper_limit
kvn@2442 1591 // scale*I+offset+1 > upper_limit
kvn@2442 1592 // ( if (scale < 0) /* and stride > 0 */
kvn@2442 1593 // I < (upper_limit-(offset+1))/scale
kvn@2480 1594 // else /* scale > 0 and stride < 0 */
kvn@2442 1595 // I > (upper_limit-(offset+1))/scale
kvn@2442 1596 // )
kvn@2480 1597 //
kvn@2480 1598 // (upper_limit-offset-1) may underflow or overflow.
kvn@2480 1599 // To avoid it min(pre_limit, original_limit) is used
kvn@2480 1600 // in do_range_check() for stride > 0 and max() for < 0.
kvn@2480 1601 Node *one = _igvn.intcon(1);
kvn@2480 1602 set_ctrl(one, C->root());
kvn@2480 1603
kvn@3680 1604 Node *plus_one = new (C) AddINode(offset, one);
kvn@2442 1605 register_new_node( plus_one, pre_ctrl );
kvn@2480 1606 // Pass (-stride) to indicate pre_loop_cond = NOT(main_loop_cond);
kvn@2480 1607 *pre_limit = adjust_limit((-stride_con), scale, plus_one, upper_limit, *pre_limit, pre_ctrl);
kvn@2442 1608
kvn@2480 1609 if (low_limit->get_int() == -max_jint) {
kvn@2480 1610 if (!RangeLimitCheck) return;
kvn@2480 1611 // We need this guard when scale*main_limit+offset >= limit
kvn@2480 1612 // due to underflow. So we need execute main-loop while
kvn@2480 1613 // scale*I+offset+1 > min_int. But (min_int-offset-1) will
kvn@2480 1614 // underflow when (offset+1) > 0 and X will be < main_limit
kvn@2480 1615 // when scale < 0 (and stride > 0). To avoid it we replace
kvn@2480 1616 // positive (offset+1) with 0.
kvn@2480 1617 //
kvn@2480 1618 // Also (min_int+1 == -max_int) is used instead of min_int here
kvn@2480 1619 // to avoid problem with scale == -1 (min_int/(-1) == min_int).
kvn@2480 1620 Node* shift = _igvn.intcon(31);
kvn@2480 1621 set_ctrl(shift, C->root());
kvn@3680 1622 Node* sign = new (C) RShiftINode(plus_one, shift);
kvn@2480 1623 register_new_node(sign, pre_ctrl);
kvn@3680 1624 plus_one = new (C) AndINode(plus_one, sign);
kvn@2480 1625 register_new_node(plus_one, pre_ctrl);
kvn@2480 1626 } else {
kvn@2480 1627 assert(low_limit->get_int() == 0, "wrong low limit for range check");
kvn@2480 1628 // The only problem we have here when offset == max_int
kvn@2480 1629 // since (max_int+1) == min_int and (0-min_int) == min_int.
kvn@2480 1630 // But it is fine since main loop will either have
kvn@2480 1631 // less iterations or will be skipped in such case.
kvn@2480 1632 }
kvn@2480 1633 // The underflow limit: low_limit <= scale*I+offset.
kvn@2480 1634 // For main-loop compute
kvn@2480 1635 // scale*I+offset+1 > low_limit
kvn@2480 1636 // ( if (scale < 0) /* and stride > 0 */
kvn@2480 1637 // I < (low_limit-(offset+1))/scale
kvn@2480 1638 // else /* scale > 0 and stride < 0 */
kvn@2480 1639 // I > (low_limit-(offset+1))/scale
kvn@2480 1640 // )
kvn@2442 1641
kvn@2480 1642 *main_limit = adjust_limit(stride_con, scale, plus_one, low_limit, *main_limit, pre_ctrl);
duke@0 1643 }
duke@0 1644 }
duke@0 1645
duke@0 1646
duke@0 1647 //------------------------------is_scaled_iv---------------------------------
duke@0 1648 // Return true if exp is a constant times an induction var
duke@0 1649 bool PhaseIdealLoop::is_scaled_iv(Node* exp, Node* iv, int* p_scale) {
duke@0 1650 if (exp == iv) {
duke@0 1651 if (p_scale != NULL) {
duke@0 1652 *p_scale = 1;
duke@0 1653 }
duke@0 1654 return true;
duke@0 1655 }
duke@0 1656 int opc = exp->Opcode();
duke@0 1657 if (opc == Op_MulI) {
duke@0 1658 if (exp->in(1) == iv && exp->in(2)->is_Con()) {
duke@0 1659 if (p_scale != NULL) {
duke@0 1660 *p_scale = exp->in(2)->get_int();
duke@0 1661 }
duke@0 1662 return true;
duke@0 1663 }
duke@0 1664 if (exp->in(2) == iv && exp->in(1)->is_Con()) {
duke@0 1665 if (p_scale != NULL) {
duke@0 1666 *p_scale = exp->in(1)->get_int();
duke@0 1667 }
duke@0 1668 return true;
duke@0 1669 }
duke@0 1670 } else if (opc == Op_LShiftI) {
duke@0 1671 if (exp->in(1) == iv && exp->in(2)->is_Con()) {
duke@0 1672 if (p_scale != NULL) {
duke@0 1673 *p_scale = 1 << exp->in(2)->get_int();
duke@0 1674 }
duke@0 1675 return true;
duke@0 1676 }
duke@0 1677 }
duke@0 1678 return false;
duke@0 1679 }
duke@0 1680
duke@0 1681 //-----------------------------is_scaled_iv_plus_offset------------------------------
duke@0 1682 // Return true if exp is a simple induction variable expression: k1*iv + (invar + k2)
duke@0 1683 bool PhaseIdealLoop::is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth) {
duke@0 1684 if (is_scaled_iv(exp, iv, p_scale)) {
duke@0 1685 if (p_offset != NULL) {
duke@0 1686 Node *zero = _igvn.intcon(0);
duke@0 1687 set_ctrl(zero, C->root());
duke@0 1688 *p_offset = zero;
duke@0 1689 }
duke@0 1690 return true;
duke@0 1691 }
duke@0 1692 int opc = exp->Opcode();
duke@0 1693 if (opc == Op_AddI) {
duke@0 1694 if (is_scaled_iv(exp->in(1), iv, p_scale)) {
duke@0 1695 if (p_offset != NULL) {
duke@0 1696 *p_offset = exp->in(2);
duke@0 1697 }
duke@0 1698 return true;
duke@0 1699 }
duke@0 1700 if (exp->in(2)->is_Con()) {
duke@0 1701 Node* offset2 = NULL;
duke@0 1702 if (depth < 2 &&
duke@0 1703 is_scaled_iv_plus_offset(exp->in(1), iv, p_scale,
duke@0 1704 p_offset != NULL ? &offset2 : NULL, depth+1)) {
duke@0 1705 if (p_offset != NULL) {
duke@0 1706 Node *ctrl_off2 = get_ctrl(offset2);
kvn@3680 1707 Node* offset = new (C) AddINode(offset2, exp->in(2));
duke@0 1708 register_new_node(offset, ctrl_off2);
duke@0 1709 *p_offset = offset;
duke@0 1710 }
duke@0 1711 return true;
duke@0 1712 }
duke@0 1713 }
duke@0 1714 } else if (opc == Op_SubI) {
duke@0 1715 if (is_scaled_iv(exp->in(1), iv, p_scale)) {
duke@0 1716 if (p_offset != NULL) {
duke@0 1717 Node *zero = _igvn.intcon(0);
duke@0 1718 set_ctrl(zero, C->root());
duke@0 1719 Node *ctrl_off = get_ctrl(exp->in(2));
kvn@3680 1720 Node* offset = new (C) SubINode(zero, exp->in(2));
duke@0 1721 register_new_node(offset, ctrl_off);
duke@0 1722 *p_offset = offset;
duke@0 1723 }
duke@0 1724 return true;
duke@0 1725 }
duke@0 1726 if (is_scaled_iv(exp->in(2), iv, p_scale)) {
duke@0 1727 if (p_offset != NULL) {
duke@0 1728 *p_scale *= -1;
duke@0 1729 *p_offset = exp->in(1);
duke@0 1730 }
duke@0 1731 return true;
duke@0 1732 }
duke@0 1733 }
duke@0 1734 return false;
duke@0 1735 }
duke@0 1736
duke@0 1737 //------------------------------do_range_check---------------------------------
duke@0 1738 // Eliminate range-checks and other trip-counter vs loop-invariant tests.
duke@0 1739 void PhaseIdealLoop::do_range_check( IdealLoopTree *loop, Node_List &old_new ) {
duke@0 1740 #ifndef PRODUCT
kvn@2230 1741 if (PrintOpto && VerifyLoopOptimizations) {
duke@0 1742 tty->print("Range Check Elimination ");
duke@0 1743 loop->dump_head();
kvn@2230 1744 } else if (TraceLoopOpts) {
kvn@2230 1745 tty->print("RangeCheck ");
kvn@2230 1746 loop->dump_head();
duke@0 1747 }
duke@0 1748 #endif
kvn@2230 1749 assert(RangeCheckElimination, "");
duke@0 1750 CountedLoopNode *cl = loop->_head->as_CountedLoop();
kvn@2230 1751 assert(cl->is_main_loop(), "");
kvn@2230 1752
kvn@2230 1753 // protect against stride not being a constant
kvn@2230 1754 if (!cl->stride_is_con())
kvn@2230 1755 return;
duke@0 1756
duke@0 1757 // Find the trip counter; we are iteration splitting based on it
duke@0 1758 Node *trip_counter = cl->phi();
duke@0 1759 // Find the main loop limit; we will trim it's iterations
duke@0 1760 // to not ever trip end tests
duke@0 1761 Node *main_limit = cl->limit();
kvn@2230 1762
kvn@2230 1763 // Need to find the main-loop zero-trip guard
kvn@2230 1764 Node *ctrl = cl->in(LoopNode::EntryControl);
kvn@2230 1765 assert(ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "");
kvn@2230 1766 Node *iffm = ctrl->in(0);
kvn@2230 1767 assert(iffm->Opcode() == Op_If, "");
kvn@2230 1768 Node *bolzm = iffm->in(1);
kvn@2230 1769 assert(bolzm->Opcode() == Op_Bool, "");
kvn@2230 1770 Node *cmpzm = bolzm->in(1);
kvn@2230 1771 assert(cmpzm->is_Cmp(), "");
kvn@2230 1772 Node *opqzm = cmpzm->in(2);
kvn@2442 1773 // Can not optimize a loop if zero-trip Opaque1 node is optimized
kvn@2230 1774 // away and then another round of loop opts attempted.
kvn@2230 1775 if (opqzm->Opcode() != Op_Opaque1)
kvn@2230 1776 return;
kvn@2230 1777 assert(opqzm->in(1) == main_limit, "do not understand situation");
kvn@2230 1778
duke@0 1779 // Find the pre-loop limit; we will expand it's iterations to
duke@0 1780 // not ever trip low tests.
duke@0 1781 Node *p_f = iffm->in(0);
kvn@2230 1782 assert(p_f->Opcode() == Op_IfFalse, "");
duke@0 1783 CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd();
kvn@2230 1784 assert(pre_end->loopnode()->is_pre_loop(), "");
duke@0 1785 Node *pre_opaq1 = pre_end->limit();
duke@0 1786 // Occasionally it's possible for a pre-loop Opaque1 node to be
duke@0 1787 // optimized away and then another round of loop opts attempted.
duke@0 1788 // We can not optimize this particular loop in that case.
kvn@2230 1789 if (pre_opaq1->Opcode() != Op_Opaque1)
duke@0 1790 return;
duke@0 1791 Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1;
duke@0 1792 Node *pre_limit = pre_opaq->in(1);
duke@0 1793
duke@0 1794 // Where do we put new limit calculations
duke@0 1795 Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl);
duke@0 1796
duke@0 1797 // Ensure the original loop limit is available from the
duke@0 1798 // pre-loop Opaque1 node.
duke@0 1799 Node *orig_limit = pre_opaq->original_loop_limit();
kvn@2230 1800 if (orig_limit == NULL || _igvn.type(orig_limit) == Type::TOP)
duke@0 1801 return;
duke@0 1802
duke@0 1803 // Must know if its a count-up or count-down loop
duke@0 1804
duke@0 1805 int stride_con = cl->stride_con();
duke@0 1806 Node *zero = _igvn.intcon(0);
duke@0 1807 Node *one = _igvn.intcon(1);
kvn@2442 1808 // Use symmetrical int range [-max_jint,max_jint]
kvn@2442 1809 Node *mini = _igvn.intcon(-max_jint);
duke@0 1810 set_ctrl(zero, C->root());
duke@0 1811 set_ctrl(one, C->root());
kvn@2442 1812 set_ctrl(mini, C->root());
duke@0 1813
duke@0 1814 // Range checks that do not dominate the loop backedge (ie.
duke@0 1815 // conditionally executed) can lengthen the pre loop limit beyond
duke@0 1816 // the original loop limit. To prevent this, the pre limit is
duke@0 1817 // (for stride > 0) MINed with the original loop limit (MAXed
duke@0 1818 // stride < 0) when some range_check (rc) is conditionally
duke@0 1819 // executed.
duke@0 1820 bool conditional_rc = false;
duke@0 1821
duke@0 1822 // Check loop body for tests of trip-counter plus loop-invariant vs
duke@0 1823 // loop-invariant.
duke@0 1824 for( uint i = 0; i < loop->_body.size(); i++ ) {
duke@0 1825 Node *iff = loop->_body[i];
duke@0 1826 if( iff->Opcode() == Op_If ) { // Test?
duke@0 1827
duke@0 1828 // Test is an IfNode, has 2 projections. If BOTH are in the loop
duke@0 1829 // we need loop unswitching instead of iteration splitting.
duke@0 1830 Node *exit = loop->is_loop_exit(iff);
duke@0 1831 if( !exit ) continue;
duke@0 1832 int flip = (exit->Opcode() == Op_IfTrue) ? 1 : 0;
duke@0 1833
duke@0 1834 // Get boolean condition to test
duke@0 1835 Node *i1 = iff->in(1);
duke@0 1836 if( !i1->is_Bool() ) continue;
duke@0 1837 BoolNode *bol = i1->as_Bool();
duke@0 1838 BoolTest b_test = bol->_test;
duke@0 1839 // Flip sense of test if exit condition is flipped
duke@0 1840 if( flip )
duke@0 1841 b_test = b_test.negate();
duke@0 1842
duke@0 1843 // Get compare
duke@0 1844 Node *cmp = bol->in(1);
duke@0 1845
duke@0 1846 // Look for trip_counter + offset vs limit
duke@0 1847 Node *rc_exp = cmp->in(1);
duke@0 1848 Node *limit = cmp->in(2);
duke@0 1849 jint scale_con= 1; // Assume trip counter not scaled
duke@0 1850
duke@0 1851 Node *limit_c = get_ctrl(limit);
duke@0 1852 if( loop->is_member(get_loop(limit_c) ) ) {
duke@0 1853 // Compare might have operands swapped; commute them
duke@0 1854 b_test = b_test.commute();
duke@0 1855 rc_exp = cmp->in(2);
duke@0 1856 limit = cmp->in(1);
duke@0 1857 limit_c = get_ctrl(limit);
duke@0 1858 if( loop->is_member(get_loop(limit_c) ) )
duke@0 1859 continue; // Both inputs are loop varying; cannot RCE
duke@0 1860 }
duke@0 1861 // Here we know 'limit' is loop invariant
duke@0 1862
duke@0 1863 // 'limit' maybe pinned below the zero trip test (probably from a
duke@0 1864 // previous round of rce), in which case, it can't be used in the
duke@0 1865 // zero trip test expression which must occur before the zero test's if.
duke@0 1866 if( limit_c == ctrl ) {
duke@0 1867 continue; // Don't rce this check but continue looking for other candidates.
duke@0 1868 }
duke@0 1869
duke@0 1870 // Check for scaled induction variable plus an offset
duke@0 1871 Node *offset = NULL;
duke@0 1872
duke@0 1873 if (!is_scaled_iv_plus_offset(rc_exp, trip_counter, &scale_con, &offset)) {
duke@0 1874 continue;
duke@0 1875 }
duke@0 1876
duke@0 1877 Node *offset_c = get_ctrl(offset);
duke@0 1878 if( loop->is_member( get_loop(offset_c) ) )
duke@0 1879 continue; // Offset is not really loop invariant
duke@0 1880 // Here we know 'offset' is loop invariant.
duke@0 1881
duke@0 1882 // As above for the 'limit', the 'offset' maybe pinned below the
duke@0 1883 // zero trip test.
duke@0 1884 if( offset_c == ctrl ) {
duke@0 1885 continue; // Don't rce this check but continue looking for other candidates.
duke@0 1886 }
kvn@2442 1887 #ifdef ASSERT
kvn@2442 1888 if (TraceRangeLimitCheck) {
kvn@2442 1889 tty->print_cr("RC bool node%s", flip ? " flipped:" : ":");
kvn@2442 1890 bol->dump(2);
kvn@2442 1891 }
kvn@2442 1892 #endif
duke@0 1893 // At this point we have the expression as:
duke@0 1894 // scale_con * trip_counter + offset :: limit
duke@0 1895 // where scale_con, offset and limit are loop invariant. Trip_counter
duke@0 1896 // monotonically increases by stride_con, a constant. Both (or either)
duke@0 1897 // stride_con and scale_con can be negative which will flip about the
duke@0 1898 // sense of the test.
duke@0 1899
duke@0 1900 // Adjust pre and main loop limits to guard the correct iteration set
duke@0 1901 if( cmp->Opcode() == Op_CmpU ) {// Unsigned compare is really 2 tests
duke@0 1902 if( b_test._test == BoolTest::lt ) { // Range checks always use lt
kvn@2442 1903 // The underflow and overflow limits: 0 <= scale*I+offset < limit
kvn@2442 1904 add_constraint( stride_con, scale_con, offset, zero, limit, pre_ctrl, &pre_limit, &main_limit );
duke@0 1905 if (!conditional_rc) {
kvn@2480 1906 // (0-offset)/scale could be outside of loop iterations range.
kvn@2480 1907 conditional_rc = !loop->dominates_backedge(iff) || RangeLimitCheck;
duke@0 1908 }
duke@0 1909 } else {
duke@0 1910 #ifndef PRODUCT
duke@0 1911 if( PrintOpto )
duke@0 1912 tty->print_cr("missed RCE opportunity");
duke@0 1913 #endif
duke@0 1914 continue; // In release mode, ignore it
duke@0 1915 }
duke@0 1916 } else { // Otherwise work on normal compares
duke@0 1917 switch( b_test._test ) {
kvn@2442 1918 case BoolTest::gt:
kvn@2442 1919 // Fall into GE case
kvn@2442 1920 case BoolTest::ge:
kvn@2442 1921 // Convert (I*scale+offset) >= Limit to (I*(-scale)+(-offset)) <= -Limit
duke@0 1922 scale_con = -scale_con;
kvn@3680 1923 offset = new (C) SubINode( zero, offset );
duke@0 1924 register_new_node( offset, pre_ctrl );
kvn@3680 1925 limit = new (C) SubINode( zero, limit );
duke@0 1926 register_new_node( limit, pre_ctrl );
duke@0 1927 // Fall into LE case
kvn@2442 1928 case BoolTest::le:
kvn@2442 1929 if (b_test._test != BoolTest::gt) {
kvn@2442 1930 // Convert X <= Y to X < Y+1
kvn@3680 1931 limit = new (C) AddINode( limit, one );
kvn@2442 1932 register_new_node( limit, pre_ctrl );
kvn@2442 1933 }
duke@0 1934 // Fall into LT case
duke@0 1935 case BoolTest::lt:
kvn@2442 1936 // The underflow and overflow limits: MIN_INT <= scale*I+offset < limit
kvn@2480 1937 // Note: (MIN_INT+1 == -MAX_INT) is used instead of MIN_INT here
kvn@2480 1938 // to avoid problem with scale == -1: MIN_INT/(-1) == MIN_INT.
kvn@2442 1939 add_constraint( stride_con, scale_con, offset, mini, limit, pre_ctrl, &pre_limit, &main_limit );
duke@0 1940 if (!conditional_rc) {
kvn@2480 1941 // ((MIN_INT+1)-offset)/scale could be outside of loop iterations range.
kvn@2480 1942 // Note: negative offset is replaced with 0 but (MIN_INT+1)/scale could
kvn@2480 1943 // still be outside of loop range.
kvn@2480 1944 conditional_rc = !loop->dominates_backedge(iff) || RangeLimitCheck;
duke@0 1945 }
duke@0 1946 break;
duke@0 1947 default:
duke@0 1948 #ifndef PRODUCT
duke@0 1949 if( PrintOpto )
duke@0 1950 tty->print_cr("missed RCE opportunity");
duke@0 1951 #endif
duke@0 1952 continue; // Unhandled case
duke@0 1953 }
duke@0 1954 }
duke@0 1955
duke@0 1956 // Kill the eliminated test
duke@0 1957 C->set_major_progress();
duke@0 1958 Node *kill_con = _igvn.intcon( 1-flip );
duke@0 1959 set_ctrl(kill_con, C->root());
kvn@3412 1960 _igvn.replace_input_of(iff, 1, kill_con);
duke@0 1961 // Find surviving projection
duke@0 1962 assert(iff->is_If(), "");
duke@0 1963 ProjNode* dp = ((IfNode*)iff)->proj_out(1-flip);
duke@0 1964 // Find loads off the surviving projection; remove their control edge
duke@0 1965 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) {
duke@0 1966 Node* cd = dp->fast_out(i); // Control-dependent node
iveresov@5635 1967 if (cd->is_Load() && cd->depends_only_on_test()) { // Loads can now float around in the loop
duke@0 1968 // Allow the load to float around in the loop, or before it
duke@0 1969 // but NOT before the pre-loop.
kvn@3412 1970 _igvn.replace_input_of(cd, 0, ctrl); // ctrl, not NULL
duke@0 1971 --i;
duke@0 1972 --imax;
duke@0 1973 }
duke@0 1974 }
duke@0 1975
duke@0 1976 } // End of is IF
duke@0 1977
duke@0 1978 }
duke@0 1979
duke@0 1980 // Update loop limits
duke@0 1981 if (conditional_rc) {
kvn@3680 1982 pre_limit = (stride_con > 0) ? (Node*)new (C) MinINode(pre_limit, orig_limit)
kvn@3680 1983 : (Node*)new (C) MaxINode(pre_limit, orig_limit);
duke@0 1984 register_new_node(pre_limit, pre_ctrl);
duke@0 1985 }
duke@0 1986 _igvn.hash_delete(pre_opaq);
duke@0 1987 pre_opaq->set_req(1, pre_limit);
duke@0 1988
duke@0 1989 // Note:: we are making the main loop limit no longer precise;
duke@0 1990 // need to round up based on stride.
kvn@2442 1991 cl->set_nonexact_trip_count();
kvn@2442 1992 if (!LoopLimitCheck && stride_con != 1 && stride_con != -1) { // Cutout for common case
duke@0 1993 // "Standard" round-up logic: ([main_limit-init+(y-1)]/y)*y+init
duke@0 1994 // Hopefully, compiler will optimize for powers of 2.
duke@0 1995 Node *ctrl = get_ctrl(main_limit);
duke@0 1996 Node *stride = cl->stride();
duke@0 1997 Node *init = cl->init_trip();
kvn@3680 1998 Node *span = new (C) SubINode(main_limit,init);
duke@0 1999 register_new_node(span,ctrl);
duke@0 2000 Node *rndup = _igvn.intcon(stride_con + ((stride_con>0)?-1:1));
kvn@3680 2001 Node *add = new (C) AddINode(span,rndup);
duke@0 2002 register_new_node(add,ctrl);
kvn@3680 2003 Node *div = new (C) DivINode(0,add,stride);
duke@0 2004 register_new_node(div,ctrl);
kvn@3680 2005 Node *mul = new (C) MulINode(div,stride);
duke@0 2006 register_new_node(mul,ctrl);
kvn@3680 2007 Node *newlim = new (C) AddINode(mul,init);
duke@0 2008 register_new_node(newlim,ctrl);
duke@0 2009 main_limit = newlim;
duke@0 2010 }
duke@0 2011
duke@0 2012 Node *main_cle = cl->loopexit();
duke@0 2013 Node *main_bol = main_cle->in(1);
duke@0 2014 // Hacking loop bounds; need private copies of exit test
duke@0 2015 if( main_bol->outcnt() > 1 ) {// BoolNode shared?
duke@0 2016 _igvn.hash_delete(main_cle);
duke@0 2017 main_bol = main_bol->clone();// Clone a private BoolNode
duke@0 2018 register_new_node( main_bol, main_cle->in(0) );
duke@0 2019 main_cle->set_req(1,main_bol);
duke@0 2020 }
duke@0 2021 Node *main_cmp = main_bol->in(1);
duke@0 2022 if( main_cmp->outcnt() > 1 ) { // CmpNode shared?
duke@0 2023 _igvn.hash_delete(main_bol);
duke@0 2024 main_cmp = main_cmp->clone();// Clone a private CmpNode
duke@0 2025 register_new_node( main_cmp, main_cle->in(0) );
duke@0 2026 main_bol->set_req(1,main_cmp);
duke@0 2027 }
duke@0 2028 // Hack the now-private loop bounds
kvn@3412 2029 _igvn.replace_input_of(main_cmp, 2, main_limit);
duke@0 2030 // The OpaqueNode is unshared by design
duke@0 2031 assert( opqzm->outcnt() == 1, "cannot hack shared node" );
kvn@3412 2032 _igvn.replace_input_of(opqzm, 1, main_limit);
duke@0 2033 }
duke@0 2034
duke@0 2035 //------------------------------DCE_loop_body----------------------------------
duke@0 2036 // Remove simplistic dead code from loop body
duke@0 2037 void IdealLoopTree::DCE_loop_body() {
duke@0 2038 for( uint i = 0; i < _body.size(); i++ )
duke@0 2039 if( _body.at(i)->outcnt() == 0 )
duke@0 2040 _body.map( i--, _body.pop() );
duke@0 2041 }
duke@0 2042
duke@0 2043
duke@0 2044 //------------------------------adjust_loop_exit_prob--------------------------
duke@0 2045 // Look for loop-exit tests with the 50/50 (or worse) guesses from the parsing stage.
duke@0 2046 // Replace with a 1-in-10 exit guess.
duke@0 2047 void IdealLoopTree::adjust_loop_exit_prob( PhaseIdealLoop *phase ) {
duke@0 2048 Node *test = tail();
duke@0 2049 while( test != _head ) {
duke@0 2050 uint top = test->Opcode();
duke@0 2051 if( top == Op_IfTrue || top == Op_IfFalse ) {
duke@0 2052 int test_con = ((ProjNode*)test)->_con;
duke@0 2053 assert(top == (uint)(test_con? Op_IfTrue: Op_IfFalse), "sanity");
duke@0 2054 IfNode *iff = test->in(0)->as_If();
duke@0 2055 if( iff->outcnt() == 2 ) { // Ignore dead tests
duke@0 2056 Node *bol = iff->in(1);
duke@0 2057 if( bol && bol->req() > 1 && bol->in(1) &&
duke@0 2058 ((bol->in(1)->Opcode() == Op_StorePConditional ) ||
kvn@420 2059 (bol->in(1)->Opcode() == Op_StoreIConditional ) ||
duke@0 2060 (bol->in(1)->Opcode() == Op_StoreLConditional ) ||
duke@0 2061 (bol->in(1)->Opcode() == Op_CompareAndSwapI ) ||
duke@0 2062 (bol->in(1)->Opcode() == Op_CompareAndSwapL ) ||
coleenp@113 2063 (bol->in(1)->Opcode() == Op_CompareAndSwapP ) ||
coleenp@113 2064 (bol->in(1)->Opcode() == Op_CompareAndSwapN )))
duke@0 2065 return; // Allocation loops RARELY take backedge
duke@0 2066 // Find the OTHER exit path from the IF
duke@0 2067 Node* ex = iff->proj_out(1-test_con);
duke@0 2068 float p = iff->_prob;
duke@0 2069 if( !phase->is_member( this, ex ) && iff->_fcnt == COUNT_UNKNOWN ) {
duke@0 2070 if( top == Op_IfTrue ) {
duke@0 2071 if( p < (PROB_FAIR + PROB_UNLIKELY_MAG(3))) {
duke@0 2072 iff->_prob = PROB_STATIC_FREQUENT;
duke@0 2073 }
duke@0 2074 } else {
duke@0 2075 if( p > (PROB_FAIR - PROB_UNLIKELY_MAG(3))) {
duke@0 2076 iff->_prob = PROB_STATIC_INFREQUENT;
duke@0 2077 }
duke@0 2078 }
duke@0 2079 }
duke@0 2080 }
duke@0 2081 }
duke@0 2082 test = phase->idom(test);
duke@0 2083 }
duke@0 2084 }
duke@0 2085
duke@0 2086
duke@0 2087 //------------------------------policy_do_remove_empty_loop--------------------
duke@0 2088 // Micro-benchmark spamming. Policy is to always remove empty loops.
duke@0 2089 // The 'DO' part is to replace the trip counter with the value it will
duke@0 2090 // have on the last iteration. This will break the loop.
duke@0 2091 bool IdealLoopTree::policy_do_remove_empty_loop( PhaseIdealLoop *phase ) {
duke@0 2092 // Minimum size must be empty loop
kvn@2300 2093 if (_body.size() > EMPTY_LOOP_SIZE)
kvn@2230 2094 return false;
duke@0 2095
kvn@2230 2096 if (!_head->is_CountedLoop())
kvn@2230 2097 return false; // Dead loop
duke@0 2098 CountedLoopNode *cl = _head->as_CountedLoop();
kvn@2613 2099 if (!cl->is_valid_counted_loop())
kvn@2230 2100 return false; // Malformed loop
kvn@2230 2101 if (!phase->is_member(this, phase->get_ctrl(cl->loopexit()->in(CountedLoopEndNode::TestValue))))
duke@0 2102 return false; // Infinite loop
never@2250 2103
duke@0 2104 #ifdef ASSERT
duke@0 2105 // Ensure only one phi which is the iv.
duke@0 2106 Node* iv = NULL;
duke@0 2107 for (DUIterator_Fast imax, i = cl->fast_outs(imax); i < imax; i++) {
duke@0 2108 Node* n = cl->fast_out(i);
duke@0 2109 if (n->Opcode() == Op_Phi) {
duke@0 2110 assert(iv == NULL, "Too many phis" );
duke@0 2111 iv = n;
duke@0 2112 }
duke@0 2113 }
duke@0 2114 assert(iv == cl->phi(), "Wrong phi" );
duke@0 2115 #endif
never@2250 2116
never@2250 2117 // main and post loops have explicitly created zero trip guard
never@2250 2118 bool needs_guard = !cl->is_main_loop() && !cl->is_post_loop();
never@2250 2119 if (needs_guard) {
kvn@2312 2120 // Skip guard if values not overlap.
kvn@2312 2121 const TypeInt* init_t = phase->_igvn.type(cl->init_trip())->is_int();
kvn@2312 2122 const TypeInt* limit_t = phase->_igvn.type(cl->limit())->is_int();
kvn@2312 2123 int stride_con = cl->stride_con();
kvn@2312 2124 if (stride_con > 0) {
kvn@2312 2125 needs_guard = (init_t->_hi >= limit_t->_lo);
kvn@2312 2126 } else {
kvn@2312 2127 needs_guard = (init_t->_lo <= limit_t->_hi);
kvn@2312 2128 }
kvn@2312 2129 }
kvn@2312 2130 if (needs_guard) {
never@2250 2131 // Check for an obvious zero trip guard.
kvn@2292 2132 Node* inctrl = PhaseIdealLoop::skip_loop_predicates(cl->in(LoopNode::EntryControl));
never@2250 2133 if (inctrl->Opcode() == Op_IfTrue) {
never@2250 2134 // The test should look like just the backedge of a CountedLoop
never@2250 2135 Node* iff = inctrl->in(0);
never@2250 2136 if (iff->is_If()) {
never@2250 2137 Node* bol = iff->in(1);
never@2250 2138 if (bol->is_Bool() && bol->as_Bool()->_test._test == cl->loopexit()->test_trip()) {
never@2250 2139 Node* cmp = bol->in(1);
never@2250 2140 if (cmp->is_Cmp() && cmp->in(1) == cl->init_trip() && cmp->in(2) == cl->limit()) {
never@2250 2141 needs_guard = false;
never@2250 2142 }
never@2250 2143 }
never@2250 2144 }
never@2250 2145 }
never@2250 2146 }
never@2250 2147
never@2250 2148 #ifndef PRODUCT
never@2250 2149 if (PrintOpto) {
never@2250 2150 tty->print("Removing empty loop with%s zero trip guard", needs_guard ? "out" : "");
never@2250 2151 this->dump_head();
never@2250 2152 } else if (TraceLoopOpts) {
never@2250 2153 tty->print("Empty with%s zero trip guard ", needs_guard ? "out" : "");
never@2250 2154 this->dump_head();
never@2250 2155 }
never@2250 2156 #endif
never@2250 2157
never@2250 2158 if (needs_guard) {
never@2250 2159 // Peel the loop to ensure there's a zero trip guard
never@2250 2160 Node_List old_new;
never@2250 2161 phase->do_peeling(this, old_new);
never@2250 2162 }
never@2250 2163
duke@0 2164 // Replace the phi at loop head with the final value of the last
duke@0 2165 // iteration. Then the CountedLoopEnd will collapse (backedge never
duke@0 2166 // taken) and all loop-invariant uses of the exit values will be correct.
duke@0 2167 Node *phi = cl->phi();
kvn@2442 2168 Node *exact_limit = phase->exact_limit(this);
kvn@2442 2169 if (exact_limit != cl->limit()) {
kvn@2442 2170 // We also need to replace the original limit to collapse loop exit.
kvn@2442 2171 Node* cmp = cl->loopexit()->cmp_node();
kvn@2442 2172 assert(cl->limit() == cmp->in(2), "sanity");
kvn@2442 2173 phase->_igvn._worklist.push(cmp->in(2)); // put limit on worklist
kvn@3412 2174 phase->_igvn.replace_input_of(cmp, 2, exact_limit); // put cmp on worklist
kvn@2442 2175 }
kvn@2442 2176 // Note: the final value after increment should not overflow since
kvn@2442 2177 // counted loop has limit check predicate.
kvn@3680 2178 Node *final = new (phase->C) SubINode( exact_limit, cl->stride() );
duke@0 2179 phase->register_new_node(final,cl->in(LoopNode::EntryControl));
kvn@1541 2180 phase->_igvn.replace_node(phi,final);
duke@0 2181 phase->C->set_major_progress();
duke@0 2182 return true;
duke@0 2183 }
duke@0 2184
kvn@2312 2185 //------------------------------policy_do_one_iteration_loop-------------------
kvn@2312 2186 // Convert one iteration loop into normal code.
kvn@2312 2187 bool IdealLoopTree::policy_do_one_iteration_loop( PhaseIdealLoop *phase ) {
kvn@2312 2188 if (!_head->as_Loop()->is_valid_counted_loop())
kvn@2312 2189 return false; // Only for counted loop
kvn@2312 2190
kvn@2312 2191 CountedLoopNode *cl = _head->as_CountedLoop();
kvn@2312 2192 if (!cl->has_exact_trip_count() || cl->trip_count() != 1) {
kvn@2312 2193 return false;
kvn@2312 2194 }
kvn@2312 2195
kvn@2312 2196 #ifndef PRODUCT
kvn@2312 2197 if(TraceLoopOpts) {
kvn@2312 2198 tty->print("OneIteration ");
kvn@2312 2199 this->dump_head();
kvn@2312 2200 }
kvn@2312 2201 #endif
kvn@2312 2202
kvn@2312 2203 Node *init_n = cl->init_trip();
kvn@2312 2204 #ifdef ASSERT
kvn@2312 2205 // Loop boundaries should be constant since trip count is exact.
kvn@2312 2206 assert(init_n->get_int() + cl->stride_con() >= cl->limit()->get_int(), "should be one iteration");
kvn@2312 2207 #endif
kvn@2312 2208 // Replace the phi at loop head with the value of the init_trip.
kvn@2312 2209 // Then the CountedLoopEnd will collapse (backedge will not be taken)
kvn@2312 2210 // and all loop-invariant uses of the exit values will be correct.
kvn@2312 2211 phase->_igvn.replace_node(cl->phi(), cl->init_trip());
kvn@2312 2212 phase->C->set_major_progress();
kvn@2312 2213 return true;
kvn@2312 2214 }
duke@0 2215
duke@0 2216 //=============================================================================
duke@0 2217 //------------------------------iteration_split_impl---------------------------
never@401 2218 bool IdealLoopTree::iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new ) {
kvn@2312 2219 // Compute exact loop trip count if possible.
kvn@2312 2220 compute_exact_trip_count(phase);
kvn@2312 2221
kvn@2312 2222 // Convert one iteration loop into normal code.
kvn@2312 2223 if (policy_do_one_iteration_loop(phase))
kvn@2312 2224 return true;
kvn@2312 2225
duke@0 2226 // Check and remove empty loops (spam micro-benchmarks)
kvn@2312 2227 if (policy_do_remove_empty_loop(phase))
cfang@1172 2228 return true; // Here we removed an empty loop
duke@0 2229
duke@0 2230 bool should_peel = policy_peeling(phase); // Should we peel?
duke@0 2231
duke@0 2232 bool should_unswitch = policy_unswitching(phase);
duke@0 2233
duke@0 2234 // Non-counted loops may be peeled; exactly 1 iteration is peeled.
duke@0 2235 // This removes loop-invariant tests (usually null checks).
kvn@2312 2236 if (!_head->is_CountedLoop()) { // Non-counted loop
duke@0 2237 if (PartialPeelLoop && phase->partial_peel(this, old_new)) {
never@401 2238 // Partial peel succeeded so terminate this round of loop opts
never@401 2239 return false;
duke@0 2240 }
kvn@2312 2241 if (should_peel) { // Should we peel?
duke@0 2242 #ifndef PRODUCT
duke@0 2243 if (PrintOpto) tty->print_cr("should_peel");
duke@0 2244 #endif
duke@0 2245 phase->do_peeling(this,old_new);
kvn@2312 2246 } else if (should_unswitch) {
duke@0 2247 phase->do_unswitching(this, old_new);
duke@0 2248 }
never@401 2249 return true;
duke@0 2250 }
duke@0 2251 CountedLoopNode *cl = _head->as_CountedLoop();
duke@0 2252
kvn@2613 2253 if (!cl->is_valid_counted_loop()) return true; // Ignore various kinds of broken loops
duke@0 2254
duke@0 2255 // Do nothing special to pre- and post- loops
kvn@2312 2256 if (cl->is_pre_loop() || cl->is_post_loop()) return true;
duke@0 2257
duke@0 2258 // Compute loop trip count from profile data
duke@0 2259 compute_profile_trip_cnt(phase);
duke@0 2260
duke@0 2261 // Before attempting fancy unrolling, RCE or alignment, see if we want
duke@0 2262 // to completely unroll this loop or do loop unswitching.
kvn@2312 2263 if (cl->is_normal_loop()) {
cfang@789 2264 if (should_unswitch) {
cfang@789 2265 phase->do_unswitching(this, old_new);
cfang@789 2266 return true;
cfang@789 2267 }
duke@0 2268 bool should_maximally_unroll = policy_maximally_unroll(phase);
kvn@2312 2269 if (should_maximally_unroll) {
duke@0 2270 // Here we did some unrolling and peeling. Eventually we will
duke@0 2271 // completely unroll this loop and it will no longer be a loop.
duke@0 2272 phase->do_maximally_unroll(this,old_new);
never@401 2273 return true;
duke@0 2274 }
duke@0 2275 }
duke@0 2276
kvn@2300 2277 // Skip next optimizations if running low on nodes. Note that
kvn@2300 2278 // policy_unswitching and policy_maximally_unroll have this check.
bharadwaj@3880 2279 uint nodes_left = MaxNodeLimit - (uint) phase->C->live_nodes();
kvn@2300 2280 if ((2 * _body.size()) > nodes_left) {
kvn@2300 2281 return true;
kvn@2300 2282 }
duke@0 2283
duke@0 2284 // Counted loops may be peeled, may need some iterations run up
duke@0 2285 // front for RCE, and may want to align loop refs to a cache
duke@0 2286 // line. Thus we clone a full loop up front whose trip count is
duke@0 2287 // at least 1 (if peeling), but may be several more.
duke@0 2288
duke@0 2289 // The main loop will start cache-line aligned with at least 1
duke@0 2290 // iteration of the unrolled body (zero-trip test required) and
duke@0 2291 // will have some range checks removed.
duke@0 2292
duke@0 2293 // A post-loop will finish any odd iterations (leftover after
duke@0 2294 // unrolling), plus any needed for RCE purposes.
duke@0 2295
duke@0 2296 bool should_unroll = policy_unroll(phase);
duke@0 2297
duke@0 2298 bool should_rce = policy_range_check(phase);
duke@0 2299
duke@0 2300 bool should_align = policy_align(phase);
duke@0 2301
duke@0 2302 // If not RCE'ing (iteration splitting) or Aligning, then we do not
duke@0 2303 // need a pre-loop. We may still need to peel an initial iteration but
duke@0 2304 // we will not be needing an unknown number of pre-iterations.
duke@0 2305 //
duke@0 2306 // Basically, if may_rce_align reports FALSE first time through,
duke@0 2307 // we will not be able to later do RCE or Aligning on this loop.
duke@0 2308 bool may_rce_align = !policy_peel_only(phase) || should_rce || should_align;
duke@0 2309
duke@0 2310 // If we have any of these conditions (RCE, alignment, unrolling) met, then
duke@0 2311 // we switch to the pre-/main-/post-loop model. This model also covers
duke@0 2312 // peeling.
kvn@2312 2313 if (should_rce || should_align || should_unroll) {
kvn@2312 2314 if (cl->is_normal_loop()) // Convert to 'pre/main/post' loops
duke@0 2315 phase->insert_pre_post_loops(this,old_new, !may_rce_align);
duke@0 2316
duke@0 2317 // Adjust the pre- and main-loop limits to let the pre and post loops run
duke@0 2318 // with full checks, but the main-loop with no checks. Remove said
duke@0 2319 // checks from the main body.
kvn@2312 2320 if (should_rce)
duke@0 2321 phase->do_range_check(this,old_new);
duke@0 2322
duke@0 2323 // Double loop body for unrolling. Adjust the minimum-trip test (will do
duke@0 2324 // twice as many iterations as before) and the main body limit (only do
duke@0 2325 // an even number of trips). If we are peeling, we might enable some RCE
duke@0 2326 // and we'd rather unroll the post-RCE'd loop SO... do not unroll if
duke@0 2327 // peeling.
kvn@2312 2328 if (should_unroll && !should_peel)
kvn@2312 2329 phase->do_unroll(this,old_new, true);
duke@0 2330
duke@0 2331 // Adjust the pre-loop limits to align the main body
duke@0 2332 // iterations.
kvn@2312 2333 if (should_align)
duke@0 2334 Unimplemented();
duke@0 2335
duke@0 2336 } else { // Else we have an unchanged counted loop
kvn@2312 2337 if (should_peel) // Might want to peel but do nothing else
duke@0 2338 phase->do_peeling(this,old_new);
duke@0 2339 }
never@401 2340 return true;
duke@0 2341 }
duke@0 2342
duke@0 2343
duke@0 2344 //=============================================================================
duke@0 2345 //------------------------------iteration_split--------------------------------
never@401 2346 bool IdealLoopTree::iteration_split( PhaseIdealLoop *phase, Node_List &old_new ) {
duke@0 2347 // Recursively iteration split nested loops
kvn@2230 2348 if (_child && !_child->iteration_split(phase, old_new))
never@401 2349 return false;
duke@0 2350
duke@0 2351 // Clean out prior deadwood
duke@0 2352 DCE_loop_body();
duke@0 2353
duke@0 2354
duke@0 2355 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage.
duke@0 2356 // Replace with a 1-in-10 exit guess.
kvn@2230 2357 if (_parent /*not the root loop*/ &&
duke@0 2358 !_irreducible &&
duke@0 2359 // Also ignore the occasional dead backedge
kvn@2230 2360 !tail()->is_top()) {
duke@0 2361 adjust_loop_exit_prob(phase);
duke@0 2362 }
duke@0 2363
duke@0 2364 // Gate unrolling, RCE and peeling efforts.
kvn@2230 2365 if (!_child && // If not an inner loop, do not split
duke@0 2366 !_irreducible &&
kvn@39 2367 _allow_optimizations &&
kvn@2230 2368 !tail()->is_top()) { // Also ignore the occasional dead backedge
duke@0 2369 if (!_has_call) {
kvn@2230 2370 if (!iteration_split_impl(phase, old_new)) {
cfang@1172 2371 return false;
cfang@1172 2372 }
duke@0 2373 } else if (policy_unswitching(phase)) {
duke@0 2374 phase->do_unswitching(this, old_new);
duke@0 2375 }
duke@0 2376 }
duke@0 2377
duke@0 2378 // Minor offset re-organization to remove loop-fallout uses of
kvn@2230 2379 // trip counter when there was no major reshaping.
kvn@2230 2380 phase->reorg_offsets(this);
kvn@2230 2381
kvn@2230 2382 if (_next && !_next->iteration_split(phase, old_new))
never@401 2383 return false;
never@401 2384 return true;
duke@0 2385 }
cfang@1172 2386
cfang@1172 2387
kvn@2292 2388 //=============================================================================
never@1683 2389 // Process all the loops in the loop tree and replace any fill
never@1683 2390 // patterns with an intrisc version.
never@1683 2391 bool PhaseIdealLoop::do_intrinsify_fill() {
never@1683 2392 bool changed = false;
never@1683 2393 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
never@1683 2394 IdealLoopTree* lpt = iter.current();
never@1683 2395 changed |= intrinsify_fill(lpt);
never@1683 2396 }
never@1683 2397 return changed;
never@1683 2398 }
never@1683 2399
never@1683 2400
never@1683 2401 // Examine an inner loop looking for a a single store of an invariant
never@1683 2402 // value in a unit stride loop,
never@1683 2403 bool PhaseIdealLoop::match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value,
never@1683 2404 Node*& shift, Node*& con) {
never@1683 2405 const char* msg = NULL;
never@1683 2406 Node* msg_node = NULL;
never@1683 2407
never@1683 2408 store_value = NULL;
never@1683 2409 con = NULL;
never@1683 2410 shift = NULL;
never@1683 2411
never@1683 2412 // Process the loop looking for stores. If there are multiple
never@1683 2413 // stores or extra control flow give at this point.
never@1683 2414 CountedLoopNode* head = lpt->_head->as_CountedLoop();
never@1683 2415 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) {
never@1683 2416 Node* n = lpt->_body.at(i);
never@1683 2417 if (n->outcnt() == 0) continue; // Ignore dead
never@1683 2418 if (n->is_Store()) {
never@1683 2419 if (store != NULL) {
never@1683 2420 msg = "multiple stores";
never@1683 2421 break;
never@1683 2422 }
never@1683 2423 int opc = n->Opcode();
roland@3724 2424 if (opc == Op_StoreP || opc == Op_StoreN || opc == Op_StoreNKlass || opc == Op_StoreCM) {
never@1683 2425 msg = "oop fills not handled";
never@1683 2426 break;
never@1683 2427 }
never@1683 2428 Node* value = n->in(MemNode::ValueIn);
never@1683 2429 if (!lpt->is_invariant(value)) {
never@1683 2430 msg = "variant store value";
never@1705 2431 } else if (!_igvn.type(n->in(MemNode::Address))->isa_aryptr()) {
never@1705 2432 msg = "not array address";
never@1683 2433 }
never@1683 2434 store = n;
never@1683 2435 store_value = value;
never@1683 2436 } else if (n->is_If() && n != head->loopexit()) {
never@1683 2437 msg = "extra control flow";
never@1683 2438 msg_node = n;
never@1683 2439 }
never@1683 2440 }
never@1683 2441
never@1683 2442 if (store == NULL) {
never@1683 2443 // No store in loop
never@1683 2444 return false;
never@1683 2445 }
never@1683 2446
never@1683 2447 if (msg == NULL && head->stride_con() != 1) {
never@1683 2448 // could handle negative strides too
never@1683 2449 if (head->stride_con() < 0) {
never@1683 2450 msg = "negative stride";
never@1683 2451 } else {
never@1683 2452 msg = "non-unit stride";
never@1683 2453 }
never@1683 2454 }
never@1683 2455
never@1683 2456 if (msg == NULL && !store->in(MemNode::Address)->is_AddP()) {
never@1683 2457 msg = "can't handle store address";
never@1683 2458 msg_node = store->in(MemNode::Address);
never@1683 2459 }
never@1683 2460
never@1733 2461 if (msg == NULL &&
never@1733 2462 (!store->in(MemNode::Memory)->is_Phi() ||
never@1733 2463 store->in(MemNode::Memory)->in(LoopNode::LoopBackControl) != store)) {
never@1733 2464 msg = "store memory isn't proper phi";
never@1733 2465 msg_node = store->in(MemNode::Memory);
never@1733 2466 }
never@1733 2467
never@1683 2468 // Make sure there is an appropriate fill routine
never@1683 2469 BasicType t = store->as_Mem()->memory_type();
never@1683 2470 const char* fill_name;
never@1683 2471 if (msg == NULL &&
never@1683 2472 StubRoutines::select_fill_function(t, false, fill_name) == NULL) {
never@1683 2473 msg = "unsupported store";
never@1683 2474 msg_node = store;
never@1683 2475 }
never@1683 2476
never@1683 2477 if (msg != NULL) {
never@1683 2478 #ifndef PRODUCT
never@1683 2479 if (TraceOptimizeFill) {
never@1683 2480 tty->print_cr("not fill intrinsic candidate: %s", msg);
never@1683 2481 if (msg_node != NULL) msg_node->dump();
never@1683 2482 }
never@1683 2483 #endif
never@1683 2484 return false;
never@1683 2485 }
never@1683 2486
never@1683 2487 // Make sure the address expression can be handled. It should be
never@1683 2488 // head->phi * elsize + con. head->phi might have a ConvI2L.
never@1683 2489 Node* elements[4];
never@1683 2490 Node* conv = NULL;
never@1705 2491 bool found_index = false;
never@1683 2492 int count = store->in(MemNode::Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements));
never@1683 2493 for (int e = 0; e < count; e++) {
never@1683 2494 Node* n = elements[e];
never@1683 2495 if (n->is_Con() && con == NULL) {
never@1683 2496 con = n;
never@1683 2497 } else if (n->Opcode() == Op_LShiftX && shift == NULL) {
never@1683 2498 Node* value = n->in(1);
never@1683 2499 #ifdef _LP64
never@1683 2500 if (value->Opcode() == Op_ConvI2L) {
never@1683 2501 conv = value;
never@1683 2502 value = value->in(1);
never@1683 2503 }
never@1683 2504 #endif
never@1683 2505 if (value != head->phi()) {
never@1683 2506 msg = "unhandled shift in address";
never@1683 2507 } else {
never@2295 2508 if (type2aelembytes(store->as_Mem()->memory_type(), true) != (1 << n->in(2)->get_int())) {
never@2295 2509 msg = "scale doesn't match";
never@2295 2510 } else {
never@2295 2511 found_index = true;
never@2295 2512 shift = n;
never@2295 2513 }
never@1683 2514 }
never@1683 2515 } else if (n->Opcode() == Op_ConvI2L && conv == NULL) {
never@1683 2516 if (n->in(1) == head->phi()) {
never@1705 2517 found_index = true;
never@1683 2518 conv = n;
never@1683 2519 } else {
never@1683 2520 msg = "unhandled input to ConvI2L";
never@1683 2521 }
never@1683 2522 } else if (n == head->phi()) {
never@1683 2523 // no shift, check below for allowed cases
never@1705 2524 found_index = true;
never@1683 2525 } else {
never@1683 2526 msg = "unhandled node in address";
never@1683 2527 msg_node = n;
never@1683 2528 }
never@1683 2529 }
never@1683 2530
never@1683 2531 if (count == -1) {
never@1683 2532 msg = "malformed address expression";
never@1683 2533 msg_node = store;
never@1683 2534 }
never@1683 2535
never@1705 2536 if (!found_index) {
never@1705 2537 msg = "missing use of index";
never@1705 2538 }
never@1705 2539
never@1683 2540 // byte sized items won't have a shift
never@1683 2541 if (msg == NULL && shift == NULL && t != T_BYTE && t != T_BOOLEAN) {
never@1683 2542 msg = "can't find shift";
never@1683 2543 msg_node = store;
never@1683 2544 }
never@1683 2545
never@1683 2546 if (msg != NULL) {
never@1683 2547 #ifndef PRODUCT
never@1683 2548 if (TraceOptimizeFill) {
never@1683 2549 tty->print_cr("not fill intrinsic: %s", msg);
never@1683 2550 if (msg_node != NULL) msg_node->dump();
never@1683 2551 }
never@1683 2552 #endif
never@1683 2553 return false;
never@1683 2554 }
never@1683 2555
never@1683 2556 // No make sure all the other nodes in the loop can be handled
never@1683 2557 VectorSet ok(Thread::current()->resource_area());
never@1683 2558
never@1683 2559 // store related values are ok
never@1683 2560 ok.set(store->_idx);
never@1683 2561 ok.set(store->in(MemNode::Memory)->_idx);
never@1683 2562
morris@4344 2563 CountedLoopEndNode* loop_exit = head->loopexit();
morris@4344 2564 guarantee(loop_exit != NULL, "no loop exit node");
morris@4344 2565
never@1683 2566 // Loop structure is ok
never@1683 2567 ok.set(head->_idx);
morris@4344 2568 ok.set(loop_exit->_idx);
never@1683 2569 ok.set(head->phi()->_idx);
never@1683 2570 ok.set(head->incr()->_idx);
morris@4344 2571 ok.set(loop_exit->cmp_node()->_idx);
morris@4344 2572 ok.set(loop_exit->in(1)->_idx);
never@1683 2573
never@1683 2574 // Address elements are ok
never@1683 2575 if (con) ok.set(con->_idx);
never@1683 2576 if (shift) ok.set(shift->_idx);
never@1683 2577 if (conv) ok.set(conv->_idx);
never@1683 2578
never@1683 2579 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) {
never@1683 2580 Node* n = lpt->_body.at(i);
never@1683 2581 if (n->outcnt() == 0) continue; // Ignore dead
never@1683 2582 if (ok.test(n->_idx)) continue;
never@1683 2583 // Backedge projection is ok
morris@4344 2584 if (n->is_IfTrue() && n->in(0) == loop_exit) continue;
never@1683 2585 if (!n->is_AddP()) {
never@1683 2586 msg = "unhandled node";
never@1683 2587 msg_node = n;
never@1683 2588 break;
never@1683 2589 }
never@1683 2590 }
never@1683 2591
never@1683 2592 // Make sure no unexpected values are used outside the loop
never@1683 2593 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) {
never@1683 2594 Node* n = lpt->_body.at(i);
never@1683 2595 // These values can be replaced with other nodes if they are used
never@1683 2596 // outside the loop.
morris@4344 2597 if (n == store || n == loop_exit || n == head->incr() || n == store->in(MemNode::Memory)) continue;
never@1683 2598 for (SimpleDUIterator iter(n); iter.has_next(); iter.next()) {
never@1683 2599 Node* use = iter.get();
never@1683 2600 if (!lpt->_body.contains(use)) {
never@1683 2601 msg = "node is used outside loop";
never@1683 2602 // lpt->_body.dump();
never@1683 2603 msg_node = n;
never@1683 2604 break;
never@1683 2605 }
never@1683 2606 }
never@1683 2607 }
never@1683 2608
never@1683 2609 #ifdef ASSERT
never@1683 2610 if (TraceOptimizeFill) {
never@1683 2611 if (msg != NULL) {
never@1683 2612 tty->print_cr("no fill intrinsic: %s", msg);
never@1683 2613 if (msg_node != NULL) msg_node->dump();
never@1683 2614 } else {
never@1683 2615 tty->print_cr("fill intrinsic for:");
never@1683 2616 }
never@1683 2617 store->dump();
never@1683 2618 if (Verbose) {
never@1683 2619 lpt->_body.dump();
never@1683 2620 }
never@1683 2621 }
never@1683 2622 #endif
never@1683 2623
never@1683 2624 return msg == NULL;
never@1683 2625 }
never@1683 2626
never@1683 2627
never@1683 2628
never@1683 2629 bool PhaseIdealLoop::intrinsify_fill(IdealLoopTree* lpt) {
never@1683 2630 // Only for counted inner loops
never@1683 2631 if (!lpt->is_counted() || !lpt->is_inner()) {
never@1683 2632 return false;
never@1683 2633 }
never@1683 2634
never@1683 2635 // Must have constant stride
never@1683 2636 CountedLoopNode* head = lpt->_head->as_CountedLoop();
kvn@2613 2637 if (!head->is_valid_counted_loop() || !head->is_normal_loop()) {
never@1683 2638 return false;
never@1683 2639 }
never@1683 2640
never@1683 2641 // Check that the body only contains a store of a loop invariant
never@1683 2642 // value that is indexed by the loop phi.
never@1683 2643 Node* store = NULL;
never@1683 2644 Node* store_value = NULL;
never@1683 2645 Node* shift = NULL;
never@1683 2646 Node* offset = NULL;
never@1683 2647 if (!match_fill_loop(lpt, store, store_value, shift, offset)) {
never@1683 2648 return false;
never@1683 2649 }
never@1683 2650
kvn@2292 2651 #ifndef PRODUCT
kvn@2292 2652 if (TraceLoopOpts) {
kvn@2292 2653 tty->print("ArrayFill ");
kvn@2292 2654 lpt->dump_head();
kvn@2292 2655 }
kvn@2292 2656 #endif
kvn@2292 2657
never@1683 2658 // Now replace the whole loop body by a call to a fill routine that
never@1683 2659 // covers the same region as the loop.
never@1683 2660 Node* base = store->in(MemNode::Address)->as_AddP()->in(AddPNode::Base);
never@1683 2661
never@1683 2662 // Build an expression for the beginning of the copy region
never@1683 2663 Node* index = head->init_trip();
never@1683 2664 #ifdef _LP64
kvn@3680 2665 index = new (C) ConvI2LNode(index);
never@1683 2666 _igvn.register_new_node_with_optimizer(index);
never@1683 2667 #endif
never@1683 2668 if (shift != NULL) {
never@1683 2669 // byte arrays don't require a shift but others do.
kvn@3680 2670 index = new (C) LShiftXNode(index, shift->in(2));
never@1683 2671 _igvn.register_new_node_with_optimizer(index);
never@1683 2672 }
kvn@3680 2673 index = new (C) AddPNode(base, base, index);
never@1683 2674 _igvn.register_new_node_with_optimizer(index);
kvn@3680 2675 Node* from = new (C) AddPNode(base, index, offset);
never@1683 2676 _igvn.register_new_node_with_optimizer(from);
never@1683 2677 // Compute the number of elements to copy
kvn@3680 2678 Node* len = new (C) SubINode(head->limit(), head->init_trip());
never@1683 2679 _igvn.register_new_node_with_optimizer(len);
never@1683 2680
never@1683 2681 BasicType t = store->as_Mem()->memory_type();
never@1683 2682 bool aligned = false;
never@1683 2683 if (offset != NULL && head->init_trip()->is_Con()) {
never@1683 2684 int element_size = type2aelembytes(t);
never@1683 2685 aligned = (offset->find_intptr_t_type()->get_con() + head->init_trip()->get_int() * element_size) % HeapWordSize == 0;
never@1683 2686 }
never@1683 2687
never@1683 2688 // Build a call to the fill routine
never@1683 2689 const char* fill_name;
never@1683 2690 address fill = StubRoutines::select_fill_function(t, aligned, fill_name);
never@1683 2691 assert(fill != NULL, "what?");
never@1683 2692
never@1683 2693 // Convert float/double to int/long for fill routines
never@1683 2694 if (t == T_FLOAT) {
kvn@3680 2695 store_value = new (C) MoveF2INode(store_value);
never@1683 2696 _igvn.register_new_node_with_optimizer(store_value);
never@1683 2697 } else if (t == T_DOUBLE) {
kvn@3680 2698 store_value = new (C) MoveD2LNode(store_value);
never@1683 2699 _igvn.register_new_node_with_optimizer(store_value);
never@1683 2700 }
never@1683 2701
never@1683 2702 Node* mem_phi = store->in(MemNode::Memory);
never@1683 2703 Node* result_ctrl;
never@1683 2704 Node* result_mem;
never@1683 2705 const TypeFunc* call_type = OptoRuntime::array_fill_Type();
kvn@3680 2706 CallLeafNode *call = new (C) CallLeafNoFPNode(call_type, fill,
kvn@3680 2707 fill_name, TypeAryPtr::get_array_body_type(t));
never@1683 2708 call->init_req(TypeFunc::Parms+0, from);
never@1683 2709 call->init_req(TypeFunc::Parms+1, store_value);
never@1764 2710 #ifdef _LP64
kvn@3680 2711 len = new (C) ConvI2LNode(len);
never@1764 2712 _igvn.register_new_node_with_optimizer(len);
never@1764 2713 #endif
never@1683 2714 call->init_req(TypeFunc::Parms+2, len);
never@1764 2715 #ifdef _LP64
never@1764 2716 call->init_req(TypeFunc::Parms+3, C->top());
never@1764 2717 #endif
never@1683 2718 call->init_req( TypeFunc::Control, head->init_control());
never@1683 2719 call->init_req( TypeFunc::I_O , C->top() ) ; // does no i/o
never@1683 2720 call->init_req( TypeFunc::Memory , mem_phi->in(LoopNode::EntryControl) );
never@1683 2721 call->init_req( TypeFunc::ReturnAdr, C->start()->proj_out(TypeFunc::ReturnAdr) );
never@1683 2722 call->init_req( TypeFunc::FramePtr, C->start()->proj_out(TypeFunc::FramePtr) );
never@1683 2723 _igvn.register_new_node_with_optimizer(call);
kvn@3680 2724 result_ctrl = new (C) ProjNode(call,TypeFunc::Control);
never@1683 2725 _igvn.register_new_node_with_optimizer(result_ctrl);
kvn@3680 2726 result_mem = new (C) ProjNode(call,TypeFunc::Memory);
never@1683 2727 _igvn.register_new_node_with_optimizer(result_mem);
never@1683 2728
kvn@3782 2729 /* Disable following optimization until proper fix (add missing checks).
kvn@3782 2730
never@1683 2731 // If this fill is tightly coupled to an allocation and overwrites
never@1683 2732 // the whole body, allow it to take over the zeroing.
never@1683 2733 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, this);
never@1683 2734 if (alloc != NULL && alloc->is_AllocateArray()) {
never@1683 2735 Node* length = alloc->as_AllocateArray()->Ideal_length();
never@1683 2736 if (head->limit() == length &&
never@1683 2737 head->init_trip() == _igvn.intcon(0)) {
never@1683 2738 if (TraceOptimizeFill) {
never@1683 2739 tty->print_cr("Eliminated zeroing in allocation");
never@1683 2740 }
never@1683 2741 alloc->maybe_set_complete(&_igvn);
never@1683 2742 } else {
never@1683 2743 #ifdef ASSERT
never@1683 2744 if (TraceOptimizeFill) {
never@1683 2745 tty->print_cr("filling array but bounds don't match");
never@1683 2746 alloc->dump();
never@1683 2747 head->init_trip()->dump();
never@1683 2748 head->limit()->dump();
never@1683 2749 length->dump();
never@1683 2750 }
never@1683 2751 #endif
never@1683 2752 }
never@1683 2753 }
kvn@3782 2754 */
never@1683 2755
never@1683 2756 // Redirect the old control and memory edges that are outside the loop.
never@1683 2757 Node* exit = head->loopexit()->proj_out(0);
never@1733 2758 // Sometimes the memory phi of the head is used as the outgoing
never@1733 2759 // state of the loop. It's safe in this case to replace it with the
never@1733 2760 // result_mem.
never@1733 2761 _igvn.replace_node(store->in(MemNode::Memory), result_mem);
never@1683 2762 _igvn.replace_node(exit, result_ctrl);
never@1683 2763 _igvn.replace_node(store, result_mem);
never@1683 2764 // Any uses the increment outside of the loop become the loop limit.
never@1683 2765 _igvn.replace_node(head->incr(), head->limit());
never@1683 2766
never@1683 2767 // Disconnect the head from the loop.
never@1683 2768 for (uint i = 0; i < lpt->_body.size(); i++) {
never@1683 2769 Node* n = lpt->_body.at(i);
never@1683 2770 _igvn.replace_node(n, C->top());
never@1683 2771 }
never@1683 2772
never@1683 2773 return true;
never@1683 2774 }