src/share/vm/opto/loopTransform.cpp

Thu, 28 Apr 2011 16:40:23 -0700

author
kvn
date
Thu, 28 Apr 2011 16:40:23 -0700
changeset 2865
ae93231c7a1f
parent 2750
6c97c830fb6f
child 2877
bad7ecd0b6ed
permissions
-rw-r--r--

7039652: Performance regression after 7004547 changes
Summary: Use unrolled_count() to limit unrolling and use the stride check only for initial stride value.
Reviewed-by: never

duke@435 1 /*
trims@2708 2 * Copyright (c) 2000, 2011, Oracle and/or its affiliates. All rights reserved.
duke@435 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
duke@435 4 *
duke@435 5 * This code is free software; you can redistribute it and/or modify it
duke@435 6 * under the terms of the GNU General Public License version 2 only, as
duke@435 7 * published by the Free Software Foundation.
duke@435 8 *
duke@435 9 * This code is distributed in the hope that it will be useful, but WITHOUT
duke@435 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
duke@435 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
duke@435 12 * version 2 for more details (a copy is included in the LICENSE file that
duke@435 13 * accompanied this code).
duke@435 14 *
duke@435 15 * You should have received a copy of the GNU General Public License version
duke@435 16 * 2 along with this work; if not, write to the Free Software Foundation,
duke@435 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
duke@435 18 *
trims@1907 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
trims@1907 20 * or visit www.oracle.com if you need additional information or have any
trims@1907 21 * questions.
duke@435 22 *
duke@435 23 */
duke@435 24
stefank@2314 25 #include "precompiled.hpp"
stefank@2314 26 #include "compiler/compileLog.hpp"
stefank@2314 27 #include "memory/allocation.inline.hpp"
stefank@2314 28 #include "opto/addnode.hpp"
stefank@2314 29 #include "opto/callnode.hpp"
stefank@2314 30 #include "opto/connode.hpp"
stefank@2314 31 #include "opto/divnode.hpp"
stefank@2314 32 #include "opto/loopnode.hpp"
stefank@2314 33 #include "opto/mulnode.hpp"
stefank@2314 34 #include "opto/rootnode.hpp"
stefank@2314 35 #include "opto/runtime.hpp"
stefank@2314 36 #include "opto/subnode.hpp"
duke@435 37
duke@435 38 //------------------------------is_loop_exit-----------------------------------
duke@435 39 // Given an IfNode, return the loop-exiting projection or NULL if both
duke@435 40 // arms remain in the loop.
duke@435 41 Node *IdealLoopTree::is_loop_exit(Node *iff) const {
duke@435 42 if( iff->outcnt() != 2 ) return NULL; // Ignore partially dead tests
duke@435 43 PhaseIdealLoop *phase = _phase;
duke@435 44 // Test is an IfNode, has 2 projections. If BOTH are in the loop
duke@435 45 // we need loop unswitching instead of peeling.
duke@435 46 if( !is_member(phase->get_loop( iff->raw_out(0) )) )
duke@435 47 return iff->raw_out(0);
duke@435 48 if( !is_member(phase->get_loop( iff->raw_out(1) )) )
duke@435 49 return iff->raw_out(1);
duke@435 50 return NULL;
duke@435 51 }
duke@435 52
duke@435 53
duke@435 54 //=============================================================================
duke@435 55
duke@435 56
duke@435 57 //------------------------------record_for_igvn----------------------------
duke@435 58 // Put loop body on igvn work list
duke@435 59 void IdealLoopTree::record_for_igvn() {
duke@435 60 for( uint i = 0; i < _body.size(); i++ ) {
duke@435 61 Node *n = _body.at(i);
duke@435 62 _phase->_igvn._worklist.push(n);
duke@435 63 }
duke@435 64 }
duke@435 65
kvn@2747 66 //------------------------------compute_exact_trip_count-----------------------
kvn@2747 67 // Compute loop exact trip count if possible. Do not recalculate trip count for
kvn@2747 68 // split loops (pre-main-post) which have their limits and inits behind Opaque node.
kvn@2747 69 void IdealLoopTree::compute_exact_trip_count( PhaseIdealLoop *phase ) {
kvn@2747 70 if (!_head->as_Loop()->is_valid_counted_loop()) {
kvn@2747 71 return;
kvn@2747 72 }
kvn@2747 73 CountedLoopNode* cl = _head->as_CountedLoop();
kvn@2747 74 // Trip count may become nonexact for iteration split loops since
kvn@2747 75 // RCE modifies limits. Note, _trip_count value is not reset since
kvn@2747 76 // it is used to limit unrolling of main loop.
kvn@2747 77 cl->set_nonexact_trip_count();
kvn@2747 78
kvn@2747 79 // Loop's test should be part of loop.
kvn@2747 80 if (!phase->is_member(this, phase->get_ctrl(cl->loopexit()->in(CountedLoopEndNode::TestValue))))
kvn@2747 81 return; // Infinite loop
kvn@2747 82
kvn@2747 83 #ifdef ASSERT
kvn@2747 84 BoolTest::mask bt = cl->loopexit()->test_trip();
kvn@2747 85 assert(bt == BoolTest::lt || bt == BoolTest::gt ||
kvn@2747 86 bt == BoolTest::ne, "canonical test is expected");
kvn@2747 87 #endif
kvn@2747 88
kvn@2747 89 Node* init_n = cl->init_trip();
kvn@2747 90 Node* limit_n = cl->limit();
kvn@2747 91 if (init_n != NULL && init_n->is_Con() &&
kvn@2747 92 limit_n != NULL && limit_n->is_Con()) {
kvn@2747 93 // Use longs to avoid integer overflow.
kvn@2747 94 int stride_con = cl->stride_con();
kvn@2747 95 long init_con = cl->init_trip()->get_int();
kvn@2747 96 long limit_con = cl->limit()->get_int();
kvn@2747 97 int stride_m = stride_con - (stride_con > 0 ? 1 : -1);
kvn@2747 98 long trip_count = (limit_con - init_con + stride_m)/stride_con;
kvn@2747 99 if (trip_count > 0 && (julong)trip_count < (julong)max_juint) {
kvn@2747 100 // Set exact trip count.
kvn@2747 101 cl->set_exact_trip_count((uint)trip_count);
kvn@2747 102 }
kvn@2747 103 }
kvn@2747 104 }
kvn@2747 105
duke@435 106 //------------------------------compute_profile_trip_cnt----------------------------
duke@435 107 // Compute loop trip count from profile data as
duke@435 108 // (backedge_count + loop_exit_count) / loop_exit_count
duke@435 109 void IdealLoopTree::compute_profile_trip_cnt( PhaseIdealLoop *phase ) {
duke@435 110 if (!_head->is_CountedLoop()) {
duke@435 111 return;
duke@435 112 }
duke@435 113 CountedLoopNode* head = _head->as_CountedLoop();
duke@435 114 if (head->profile_trip_cnt() != COUNT_UNKNOWN) {
duke@435 115 return; // Already computed
duke@435 116 }
duke@435 117 float trip_cnt = (float)max_jint; // default is big
duke@435 118
duke@435 119 Node* back = head->in(LoopNode::LoopBackControl);
duke@435 120 while (back != head) {
duke@435 121 if ((back->Opcode() == Op_IfTrue || back->Opcode() == Op_IfFalse) &&
duke@435 122 back->in(0) &&
duke@435 123 back->in(0)->is_If() &&
duke@435 124 back->in(0)->as_If()->_fcnt != COUNT_UNKNOWN &&
duke@435 125 back->in(0)->as_If()->_prob != PROB_UNKNOWN) {
duke@435 126 break;
duke@435 127 }
duke@435 128 back = phase->idom(back);
duke@435 129 }
duke@435 130 if (back != head) {
duke@435 131 assert((back->Opcode() == Op_IfTrue || back->Opcode() == Op_IfFalse) &&
duke@435 132 back->in(0), "if-projection exists");
duke@435 133 IfNode* back_if = back->in(0)->as_If();
duke@435 134 float loop_back_cnt = back_if->_fcnt * back_if->_prob;
duke@435 135
duke@435 136 // Now compute a loop exit count
duke@435 137 float loop_exit_cnt = 0.0f;
duke@435 138 for( uint i = 0; i < _body.size(); i++ ) {
duke@435 139 Node *n = _body[i];
duke@435 140 if( n->is_If() ) {
duke@435 141 IfNode *iff = n->as_If();
duke@435 142 if( iff->_fcnt != COUNT_UNKNOWN && iff->_prob != PROB_UNKNOWN ) {
duke@435 143 Node *exit = is_loop_exit(iff);
duke@435 144 if( exit ) {
duke@435 145 float exit_prob = iff->_prob;
duke@435 146 if (exit->Opcode() == Op_IfFalse) exit_prob = 1.0 - exit_prob;
duke@435 147 if (exit_prob > PROB_MIN) {
duke@435 148 float exit_cnt = iff->_fcnt * exit_prob;
duke@435 149 loop_exit_cnt += exit_cnt;
duke@435 150 }
duke@435 151 }
duke@435 152 }
duke@435 153 }
duke@435 154 }
duke@435 155 if (loop_exit_cnt > 0.0f) {
duke@435 156 trip_cnt = (loop_back_cnt + loop_exit_cnt) / loop_exit_cnt;
duke@435 157 } else {
duke@435 158 // No exit count so use
duke@435 159 trip_cnt = loop_back_cnt;
duke@435 160 }
duke@435 161 }
duke@435 162 #ifndef PRODUCT
duke@435 163 if (TraceProfileTripCount) {
duke@435 164 tty->print_cr("compute_profile_trip_cnt lp: %d cnt: %f\n", head->_idx, trip_cnt);
duke@435 165 }
duke@435 166 #endif
duke@435 167 head->set_profile_trip_cnt(trip_cnt);
duke@435 168 }
duke@435 169
duke@435 170 //---------------------is_invariant_addition-----------------------------
duke@435 171 // Return nonzero index of invariant operand for an Add or Sub
twisti@1040 172 // of (nonconstant) invariant and variant values. Helper for reassociate_invariants.
duke@435 173 int IdealLoopTree::is_invariant_addition(Node* n, PhaseIdealLoop *phase) {
duke@435 174 int op = n->Opcode();
duke@435 175 if (op == Op_AddI || op == Op_SubI) {
duke@435 176 bool in1_invar = this->is_invariant(n->in(1));
duke@435 177 bool in2_invar = this->is_invariant(n->in(2));
duke@435 178 if (in1_invar && !in2_invar) return 1;
duke@435 179 if (!in1_invar && in2_invar) return 2;
duke@435 180 }
duke@435 181 return 0;
duke@435 182 }
duke@435 183
duke@435 184 //---------------------reassociate_add_sub-----------------------------
duke@435 185 // Reassociate invariant add and subtract expressions:
duke@435 186 //
duke@435 187 // inv1 + (x + inv2) => ( inv1 + inv2) + x
duke@435 188 // (x + inv2) + inv1 => ( inv1 + inv2) + x
duke@435 189 // inv1 + (x - inv2) => ( inv1 - inv2) + x
duke@435 190 // inv1 - (inv2 - x) => ( inv1 - inv2) + x
duke@435 191 // (x + inv2) - inv1 => (-inv1 + inv2) + x
duke@435 192 // (x - inv2) + inv1 => ( inv1 - inv2) + x
duke@435 193 // (x - inv2) - inv1 => (-inv1 - inv2) + x
duke@435 194 // inv1 + (inv2 - x) => ( inv1 + inv2) - x
duke@435 195 // inv1 - (x - inv2) => ( inv1 + inv2) - x
duke@435 196 // (inv2 - x) + inv1 => ( inv1 + inv2) - x
duke@435 197 // (inv2 - x) - inv1 => (-inv1 + inv2) - x
duke@435 198 // inv1 - (x + inv2) => ( inv1 - inv2) - x
duke@435 199 //
duke@435 200 Node* IdealLoopTree::reassociate_add_sub(Node* n1, PhaseIdealLoop *phase) {
duke@435 201 if (!n1->is_Add() && !n1->is_Sub() || n1->outcnt() == 0) return NULL;
duke@435 202 if (is_invariant(n1)) return NULL;
duke@435 203 int inv1_idx = is_invariant_addition(n1, phase);
duke@435 204 if (!inv1_idx) return NULL;
duke@435 205 // Don't mess with add of constant (igvn moves them to expression tree root.)
duke@435 206 if (n1->is_Add() && n1->in(2)->is_Con()) return NULL;
duke@435 207 Node* inv1 = n1->in(inv1_idx);
duke@435 208 Node* n2 = n1->in(3 - inv1_idx);
duke@435 209 int inv2_idx = is_invariant_addition(n2, phase);
duke@435 210 if (!inv2_idx) return NULL;
duke@435 211 Node* x = n2->in(3 - inv2_idx);
duke@435 212 Node* inv2 = n2->in(inv2_idx);
duke@435 213
duke@435 214 bool neg_x = n2->is_Sub() && inv2_idx == 1;
duke@435 215 bool neg_inv2 = n2->is_Sub() && inv2_idx == 2;
duke@435 216 bool neg_inv1 = n1->is_Sub() && inv1_idx == 2;
duke@435 217 if (n1->is_Sub() && inv1_idx == 1) {
duke@435 218 neg_x = !neg_x;
duke@435 219 neg_inv2 = !neg_inv2;
duke@435 220 }
duke@435 221 Node* inv1_c = phase->get_ctrl(inv1);
duke@435 222 Node* inv2_c = phase->get_ctrl(inv2);
duke@435 223 Node* n_inv1;
duke@435 224 if (neg_inv1) {
duke@435 225 Node *zero = phase->_igvn.intcon(0);
duke@435 226 phase->set_ctrl(zero, phase->C->root());
duke@435 227 n_inv1 = new (phase->C, 3) SubINode(zero, inv1);
duke@435 228 phase->register_new_node(n_inv1, inv1_c);
duke@435 229 } else {
duke@435 230 n_inv1 = inv1;
duke@435 231 }
duke@435 232 Node* inv;
duke@435 233 if (neg_inv2) {
duke@435 234 inv = new (phase->C, 3) SubINode(n_inv1, inv2);
duke@435 235 } else {
duke@435 236 inv = new (phase->C, 3) AddINode(n_inv1, inv2);
duke@435 237 }
duke@435 238 phase->register_new_node(inv, phase->get_early_ctrl(inv));
duke@435 239
duke@435 240 Node* addx;
duke@435 241 if (neg_x) {
duke@435 242 addx = new (phase->C, 3) SubINode(inv, x);
duke@435 243 } else {
duke@435 244 addx = new (phase->C, 3) AddINode(x, inv);
duke@435 245 }
duke@435 246 phase->register_new_node(addx, phase->get_ctrl(x));
kvn@1976 247 phase->_igvn.replace_node(n1, addx);
kvn@2665 248 assert(phase->get_loop(phase->get_ctrl(n1)) == this, "");
kvn@2665 249 _body.yank(n1);
duke@435 250 return addx;
duke@435 251 }
duke@435 252
duke@435 253 //---------------------reassociate_invariants-----------------------------
duke@435 254 // Reassociate invariant expressions:
duke@435 255 void IdealLoopTree::reassociate_invariants(PhaseIdealLoop *phase) {
duke@435 256 for (int i = _body.size() - 1; i >= 0; i--) {
duke@435 257 Node *n = _body.at(i);
duke@435 258 for (int j = 0; j < 5; j++) {
duke@435 259 Node* nn = reassociate_add_sub(n, phase);
duke@435 260 if (nn == NULL) break;
duke@435 261 n = nn; // again
duke@435 262 };
duke@435 263 }
duke@435 264 }
duke@435 265
duke@435 266 //------------------------------policy_peeling---------------------------------
duke@435 267 // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can
duke@435 268 // make some loop-invariant test (usually a null-check) happen before the loop.
duke@435 269 bool IdealLoopTree::policy_peeling( PhaseIdealLoop *phase ) const {
duke@435 270 Node *test = ((IdealLoopTree*)this)->tail();
duke@435 271 int body_size = ((IdealLoopTree*)this)->_body.size();
duke@435 272 int uniq = phase->C->unique();
duke@435 273 // Peeling does loop cloning which can result in O(N^2) node construction
duke@435 274 if( body_size > 255 /* Prevent overflow for large body_size */
duke@435 275 || (body_size * body_size + uniq > MaxNodeLimit) ) {
duke@435 276 return false; // too large to safely clone
duke@435 277 }
duke@435 278 while( test != _head ) { // Scan till run off top of loop
duke@435 279 if( test->is_If() ) { // Test?
duke@435 280 Node *ctrl = phase->get_ctrl(test->in(1));
duke@435 281 if (ctrl->is_top())
duke@435 282 return false; // Found dead test on live IF? No peeling!
duke@435 283 // Standard IF only has one input value to check for loop invariance
duke@435 284 assert( test->Opcode() == Op_If || test->Opcode() == Op_CountedLoopEnd, "Check this code when new subtype is added");
duke@435 285 // Condition is not a member of this loop?
duke@435 286 if( !is_member(phase->get_loop(ctrl)) &&
duke@435 287 is_loop_exit(test) )
duke@435 288 return true; // Found reason to peel!
duke@435 289 }
duke@435 290 // Walk up dominators to loop _head looking for test which is
duke@435 291 // executed on every path thru loop.
duke@435 292 test = phase->idom(test);
duke@435 293 }
duke@435 294 return false;
duke@435 295 }
duke@435 296
duke@435 297 //------------------------------peeled_dom_test_elim---------------------------
duke@435 298 // If we got the effect of peeling, either by actually peeling or by making
duke@435 299 // a pre-loop which must execute at least once, we can remove all
duke@435 300 // loop-invariant dominated tests in the main body.
duke@435 301 void PhaseIdealLoop::peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new ) {
duke@435 302 bool progress = true;
duke@435 303 while( progress ) {
duke@435 304 progress = false; // Reset for next iteration
duke@435 305 Node *prev = loop->_head->in(LoopNode::LoopBackControl);//loop->tail();
duke@435 306 Node *test = prev->in(0);
duke@435 307 while( test != loop->_head ) { // Scan till run off top of loop
duke@435 308
duke@435 309 int p_op = prev->Opcode();
duke@435 310 if( (p_op == Op_IfFalse || p_op == Op_IfTrue) &&
duke@435 311 test->is_If() && // Test?
duke@435 312 !test->in(1)->is_Con() && // And not already obvious?
duke@435 313 // Condition is not a member of this loop?
duke@435 314 !loop->is_member(get_loop(get_ctrl(test->in(1))))){
duke@435 315 // Walk loop body looking for instances of this test
duke@435 316 for( uint i = 0; i < loop->_body.size(); i++ ) {
duke@435 317 Node *n = loop->_body.at(i);
duke@435 318 if( n->is_If() && n->in(1) == test->in(1) /*&& n != loop->tail()->in(0)*/ ) {
duke@435 319 // IfNode was dominated by version in peeled loop body
duke@435 320 progress = true;
duke@435 321 dominated_by( old_new[prev->_idx], n );
duke@435 322 }
duke@435 323 }
duke@435 324 }
duke@435 325 prev = test;
duke@435 326 test = idom(test);
duke@435 327 } // End of scan tests in loop
duke@435 328
duke@435 329 } // End of while( progress )
duke@435 330 }
duke@435 331
duke@435 332 //------------------------------do_peeling-------------------------------------
duke@435 333 // Peel the first iteration of the given loop.
duke@435 334 // Step 1: Clone the loop body. The clone becomes the peeled iteration.
duke@435 335 // The pre-loop illegally has 2 control users (old & new loops).
duke@435 336 // Step 2: Make the old-loop fall-in edges point to the peeled iteration.
duke@435 337 // Do this by making the old-loop fall-in edges act as if they came
duke@435 338 // around the loopback from the prior iteration (follow the old-loop
duke@435 339 // backedges) and then map to the new peeled iteration. This leaves
duke@435 340 // the pre-loop with only 1 user (the new peeled iteration), but the
duke@435 341 // peeled-loop backedge has 2 users.
duke@435 342 // Step 3: Cut the backedge on the clone (so its not a loop) and remove the
duke@435 343 // extra backedge user.
kvn@2727 344 //
kvn@2727 345 // orig
kvn@2727 346 //
kvn@2727 347 // stmt1
kvn@2727 348 // |
kvn@2727 349 // v
kvn@2727 350 // loop predicate
kvn@2727 351 // |
kvn@2727 352 // v
kvn@2727 353 // loop<----+
kvn@2727 354 // | |
kvn@2727 355 // stmt2 |
kvn@2727 356 // | |
kvn@2727 357 // v |
kvn@2727 358 // if ^
kvn@2727 359 // / \ |
kvn@2727 360 // / \ |
kvn@2727 361 // v v |
kvn@2727 362 // false true |
kvn@2727 363 // / \ |
kvn@2727 364 // / ----+
kvn@2727 365 // |
kvn@2727 366 // v
kvn@2727 367 // exit
kvn@2727 368 //
kvn@2727 369 //
kvn@2727 370 // after clone loop
kvn@2727 371 //
kvn@2727 372 // stmt1
kvn@2727 373 // |
kvn@2727 374 // v
kvn@2727 375 // loop predicate
kvn@2727 376 // / \
kvn@2727 377 // clone / \ orig
kvn@2727 378 // / \
kvn@2727 379 // / \
kvn@2727 380 // v v
kvn@2727 381 // +---->loop clone loop<----+
kvn@2727 382 // | | | |
kvn@2727 383 // | stmt2 clone stmt2 |
kvn@2727 384 // | | | |
kvn@2727 385 // | v v |
kvn@2727 386 // ^ if clone If ^
kvn@2727 387 // | / \ / \ |
kvn@2727 388 // | / \ / \ |
kvn@2727 389 // | v v v v |
kvn@2727 390 // | true false false true |
kvn@2727 391 // | / \ / \ |
kvn@2727 392 // +---- \ / ----+
kvn@2727 393 // \ /
kvn@2727 394 // 1v v2
kvn@2727 395 // region
kvn@2727 396 // |
kvn@2727 397 // v
kvn@2727 398 // exit
kvn@2727 399 //
kvn@2727 400 //
kvn@2727 401 // after peel and predicate move
kvn@2727 402 //
kvn@2727 403 // stmt1
kvn@2727 404 // /
kvn@2727 405 // /
kvn@2727 406 // clone / orig
kvn@2727 407 // /
kvn@2727 408 // / +----------+
kvn@2727 409 // / | |
kvn@2727 410 // / loop predicate |
kvn@2727 411 // / | |
kvn@2727 412 // v v |
kvn@2727 413 // TOP-->loop clone loop<----+ |
kvn@2727 414 // | | | |
kvn@2727 415 // stmt2 clone stmt2 | |
kvn@2727 416 // | | | ^
kvn@2727 417 // v v | |
kvn@2727 418 // if clone If ^ |
kvn@2727 419 // / \ / \ | |
kvn@2727 420 // / \ / \ | |
kvn@2727 421 // v v v v | |
kvn@2727 422 // true false false true | |
kvn@2727 423 // | \ / \ | |
kvn@2727 424 // | \ / ----+ ^
kvn@2727 425 // | \ / |
kvn@2727 426 // | 1v v2 |
kvn@2727 427 // v region |
kvn@2727 428 // | | |
kvn@2727 429 // | v |
kvn@2727 430 // | exit |
kvn@2727 431 // | |
kvn@2727 432 // +--------------->-----------------+
kvn@2727 433 //
kvn@2727 434 //
kvn@2727 435 // final graph
kvn@2727 436 //
kvn@2727 437 // stmt1
kvn@2727 438 // |
kvn@2727 439 // v
kvn@2727 440 // stmt2 clone
kvn@2727 441 // |
kvn@2727 442 // v
kvn@2727 443 // if clone
kvn@2727 444 // / |
kvn@2727 445 // / |
kvn@2727 446 // v v
kvn@2727 447 // false true
kvn@2727 448 // | |
kvn@2727 449 // | v
kvn@2727 450 // | loop predicate
kvn@2727 451 // | |
kvn@2727 452 // | v
kvn@2727 453 // | loop<----+
kvn@2727 454 // | | |
kvn@2727 455 // | stmt2 |
kvn@2727 456 // | | |
kvn@2727 457 // | v |
kvn@2727 458 // v if ^
kvn@2727 459 // | / \ |
kvn@2727 460 // | / \ |
kvn@2727 461 // | v v |
kvn@2727 462 // | false true |
kvn@2727 463 // | | \ |
kvn@2727 464 // v v --+
kvn@2727 465 // region
kvn@2727 466 // |
kvn@2727 467 // v
kvn@2727 468 // exit
kvn@2727 469 //
duke@435 470 void PhaseIdealLoop::do_peeling( IdealLoopTree *loop, Node_List &old_new ) {
duke@435 471
duke@435 472 C->set_major_progress();
duke@435 473 // Peeling a 'main' loop in a pre/main/post situation obfuscates the
duke@435 474 // 'pre' loop from the main and the 'pre' can no longer have it's
duke@435 475 // iterations adjusted. Therefore, we need to declare this loop as
duke@435 476 // no longer a 'main' loop; it will need new pre and post loops before
duke@435 477 // we can do further RCE.
kvn@2665 478 #ifndef PRODUCT
kvn@2665 479 if (TraceLoopOpts) {
kvn@2665 480 tty->print("Peel ");
kvn@2665 481 loop->dump_head();
kvn@2665 482 }
kvn@2665 483 #endif
kvn@2727 484 Node* head = loop->_head;
kvn@2727 485 bool counted_loop = head->is_CountedLoop();
kvn@2727 486 if (counted_loop) {
kvn@2727 487 CountedLoopNode *cl = head->as_CountedLoop();
duke@435 488 assert(cl->trip_count() > 0, "peeling a fully unrolled loop");
duke@435 489 cl->set_trip_count(cl->trip_count() - 1);
kvn@2665 490 if (cl->is_main_loop()) {
duke@435 491 cl->set_normal_loop();
duke@435 492 #ifndef PRODUCT
kvn@2665 493 if (PrintOpto && VerifyLoopOptimizations) {
duke@435 494 tty->print("Peeling a 'main' loop; resetting to 'normal' ");
duke@435 495 loop->dump_head();
duke@435 496 }
duke@435 497 #endif
duke@435 498 }
duke@435 499 }
kvn@2727 500 Node* entry = head->in(LoopNode::EntryControl);
duke@435 501
duke@435 502 // Step 1: Clone the loop body. The clone becomes the peeled iteration.
duke@435 503 // The pre-loop illegally has 2 control users (old & new loops).
kvn@2727 504 clone_loop( loop, old_new, dom_depth(head) );
duke@435 505
duke@435 506 // Step 2: Make the old-loop fall-in edges point to the peeled iteration.
duke@435 507 // Do this by making the old-loop fall-in edges act as if they came
duke@435 508 // around the loopback from the prior iteration (follow the old-loop
duke@435 509 // backedges) and then map to the new peeled iteration. This leaves
duke@435 510 // the pre-loop with only 1 user (the new peeled iteration), but the
duke@435 511 // peeled-loop backedge has 2 users.
kvn@2727 512 Node* new_exit_value = old_new[head->in(LoopNode::LoopBackControl)->_idx];
kvn@2727 513 new_exit_value = move_loop_predicates(entry, new_exit_value);
kvn@2727 514 _igvn.hash_delete(head);
kvn@2727 515 head->set_req(LoopNode::EntryControl, new_exit_value);
kvn@2727 516 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
kvn@2727 517 Node* old = head->fast_out(j);
kvn@2727 518 if (old->in(0) == loop->_head && old->req() == 3 && old->is_Phi()) {
kvn@2727 519 new_exit_value = old_new[old->in(LoopNode::LoopBackControl)->_idx];
kvn@2727 520 if (!new_exit_value ) // Backedge value is ALSO loop invariant?
duke@435 521 // Then loop body backedge value remains the same.
duke@435 522 new_exit_value = old->in(LoopNode::LoopBackControl);
duke@435 523 _igvn.hash_delete(old);
duke@435 524 old->set_req(LoopNode::EntryControl, new_exit_value);
duke@435 525 }
duke@435 526 }
duke@435 527
duke@435 528
duke@435 529 // Step 3: Cut the backedge on the clone (so its not a loop) and remove the
duke@435 530 // extra backedge user.
kvn@2727 531 Node* new_head = old_new[head->_idx];
kvn@2727 532 _igvn.hash_delete(new_head);
kvn@2727 533 new_head->set_req(LoopNode::LoopBackControl, C->top());
kvn@2727 534 for (DUIterator_Fast j2max, j2 = new_head->fast_outs(j2max); j2 < j2max; j2++) {
kvn@2727 535 Node* use = new_head->fast_out(j2);
kvn@2727 536 if (use->in(0) == new_head && use->req() == 3 && use->is_Phi()) {
duke@435 537 _igvn.hash_delete(use);
duke@435 538 use->set_req(LoopNode::LoopBackControl, C->top());
duke@435 539 }
duke@435 540 }
duke@435 541
duke@435 542
duke@435 543 // Step 4: Correct dom-depth info. Set to loop-head depth.
kvn@2727 544 int dd = dom_depth(head);
kvn@2727 545 set_idom(head, head->in(1), dd);
duke@435 546 for (uint j3 = 0; j3 < loop->_body.size(); j3++) {
duke@435 547 Node *old = loop->_body.at(j3);
duke@435 548 Node *nnn = old_new[old->_idx];
duke@435 549 if (!has_ctrl(nnn))
duke@435 550 set_idom(nnn, idom(nnn), dd-1);
duke@435 551 // While we're at it, remove any SafePoints from the peeled code
kvn@2727 552 if (old->Opcode() == Op_SafePoint) {
duke@435 553 Node *nnn = old_new[old->_idx];
duke@435 554 lazy_replace(nnn,nnn->in(TypeFunc::Control));
duke@435 555 }
duke@435 556 }
duke@435 557
duke@435 558 // Now force out all loop-invariant dominating tests. The optimizer
duke@435 559 // finds some, but we _know_ they are all useless.
duke@435 560 peeled_dom_test_elim(loop,old_new);
duke@435 561
duke@435 562 loop->record_for_igvn();
duke@435 563 }
duke@435 564
kvn@2735 565 #define EMPTY_LOOP_SIZE 7 // number of nodes in an empty loop
kvn@2735 566
duke@435 567 //------------------------------policy_maximally_unroll------------------------
kvn@2735 568 // Calculate exact loop trip count and return true if loop can be maximally
kvn@2735 569 // unrolled.
duke@435 570 bool IdealLoopTree::policy_maximally_unroll( PhaseIdealLoop *phase ) const {
duke@435 571 CountedLoopNode *cl = _head->as_CountedLoop();
kvn@2694 572 assert(cl->is_normal_loop(), "");
kvn@2735 573 if (!cl->is_valid_counted_loop())
kvn@2735 574 return false; // Malformed counted loop
duke@435 575
kvn@2747 576 if (!cl->has_exact_trip_count()) {
kvn@2747 577 // Trip count is not exact.
duke@435 578 return false;
duke@435 579 }
duke@435 580
kvn@2747 581 uint trip_count = cl->trip_count();
kvn@2747 582 // Note, max_juint is used to indicate unknown trip count.
kvn@2747 583 assert(trip_count > 1, "one iteration loop should be optimized out already");
kvn@2747 584 assert(trip_count < max_juint, "exact trip_count should be less than max_uint.");
duke@435 585
duke@435 586 // Real policy: if we maximally unroll, does it get too big?
duke@435 587 // Allow the unrolled mess to get larger than standard loop
duke@435 588 // size. After all, it will no longer be a loop.
duke@435 589 uint body_size = _body.size();
duke@435 590 uint unroll_limit = (uint)LoopUnrollLimit * 4;
duke@435 591 assert( (intx)unroll_limit == LoopUnrollLimit * 4, "LoopUnrollLimit must fit in 32bits");
kvn@2694 592 if (trip_count > unroll_limit || body_size > unroll_limit) {
kvn@2694 593 return false;
kvn@2694 594 }
kvn@2694 595
kvn@2735 596 // Take into account that after unroll conjoined heads and tails will fold,
kvn@2735 597 // otherwise policy_unroll() may allow more unrolling than max unrolling.
kvn@2735 598 uint new_body_size = EMPTY_LOOP_SIZE + (body_size - EMPTY_LOOP_SIZE) * trip_count;
kvn@2735 599 uint tst_body_size = (new_body_size - EMPTY_LOOP_SIZE) / trip_count + EMPTY_LOOP_SIZE;
kvn@2735 600 if (body_size != tst_body_size) // Check for int overflow
kvn@2735 601 return false;
kvn@2735 602 if (new_body_size > unroll_limit ||
kvn@2735 603 // Unrolling can result in a large amount of node construction
kvn@2735 604 new_body_size >= MaxNodeLimit - phase->C->unique()) {
kvn@2735 605 return false;
kvn@2735 606 }
kvn@2735 607
kvn@2699 608 // Currently we don't have policy to optimize one iteration loops.
kvn@2699 609 // Maximally unrolling transformation is used for that:
kvn@2699 610 // it is peeled and the original loop become non reachable (dead).
kvn@2735 611 // Also fully unroll a loop with few iterations regardless next
kvn@2735 612 // conditions since following loop optimizations will split
kvn@2735 613 // such loop anyway (pre-main-post).
kvn@2735 614 if (trip_count <= 3)
kvn@2699 615 return true;
kvn@2699 616
kvn@2694 617 // Do not unroll a loop with String intrinsics code.
kvn@2694 618 // String intrinsics are large and have loops.
kvn@2694 619 for (uint k = 0; k < _body.size(); k++) {
kvn@2694 620 Node* n = _body.at(k);
kvn@2694 621 switch (n->Opcode()) {
kvn@2694 622 case Op_StrComp:
kvn@2694 623 case Op_StrEquals:
kvn@2694 624 case Op_StrIndexOf:
kvn@2694 625 case Op_AryEq: {
kvn@2694 626 return false;
kvn@2694 627 }
kvn@2694 628 } // switch
kvn@2694 629 }
kvn@2694 630
kvn@2735 631 return true; // Do maximally unroll
duke@435 632 }
duke@435 633
duke@435 634
kvn@2865 635 #define MAX_UNROLL 16 // maximum number of unrolls for main loop
kvn@2865 636
duke@435 637 //------------------------------policy_unroll----------------------------------
duke@435 638 // Return TRUE or FALSE if the loop should be unrolled or not. Unroll if
duke@435 639 // the loop is a CountedLoop and the body is small enough.
duke@435 640 bool IdealLoopTree::policy_unroll( PhaseIdealLoop *phase ) const {
duke@435 641
duke@435 642 CountedLoopNode *cl = _head->as_CountedLoop();
kvn@2694 643 assert(cl->is_normal_loop() || cl->is_main_loop(), "");
duke@435 644
kvn@2735 645 if (!cl->is_valid_counted_loop())
kvn@2735 646 return false; // Malformed counted loop
duke@435 647
duke@435 648 // protect against over-unrolling
kvn@2694 649 if (cl->trip_count() <= 1) return false;
duke@435 650
kvn@2865 651 int future_unroll_ct = cl->unrolled_count() * 2;
kvn@2865 652 if (future_unroll_ct > MAX_UNROLL) return false;
kvn@2735 653
kvn@2865 654 // Check for initial stride being a small enough constant
kvn@2865 655 if (abs(cl->stride_con()) > (1<<2)*future_unroll_ct) return false;
duke@435 656
duke@435 657 // Don't unroll if the next round of unrolling would push us
duke@435 658 // over the expected trip count of the loop. One is subtracted
duke@435 659 // from the expected trip count because the pre-loop normally
duke@435 660 // executes 1 iteration.
duke@435 661 if (UnrollLimitForProfileCheck > 0 &&
duke@435 662 cl->profile_trip_cnt() != COUNT_UNKNOWN &&
duke@435 663 future_unroll_ct > UnrollLimitForProfileCheck &&
duke@435 664 (float)future_unroll_ct > cl->profile_trip_cnt() - 1.0) {
duke@435 665 return false;
duke@435 666 }
duke@435 667
duke@435 668 // When unroll count is greater than LoopUnrollMin, don't unroll if:
duke@435 669 // the residual iterations are more than 10% of the trip count
duke@435 670 // and rounds of "unroll,optimize" are not making significant progress
duke@435 671 // Progress defined as current size less than 20% larger than previous size.
duke@435 672 if (UseSuperWord && cl->node_count_before_unroll() > 0 &&
duke@435 673 future_unroll_ct > LoopUnrollMin &&
duke@435 674 (future_unroll_ct - 1) * 10.0 > cl->profile_trip_cnt() &&
duke@435 675 1.2 * cl->node_count_before_unroll() < (double)_body.size()) {
duke@435 676 return false;
duke@435 677 }
duke@435 678
duke@435 679 Node *init_n = cl->init_trip();
duke@435 680 Node *limit_n = cl->limit();
duke@435 681 // Non-constant bounds.
duke@435 682 // Protect against over-unrolling when init or/and limit are not constant
duke@435 683 // (so that trip_count's init value is maxint) but iv range is known.
kvn@2694 684 if (init_n == NULL || !init_n->is_Con() ||
kvn@2694 685 limit_n == NULL || !limit_n->is_Con()) {
duke@435 686 Node* phi = cl->phi();
kvn@2694 687 if (phi != NULL) {
duke@435 688 assert(phi->is_Phi() && phi->in(0) == _head, "Counted loop should have iv phi.");
duke@435 689 const TypeInt* iv_type = phase->_igvn.type(phi)->is_int();
duke@435 690 int next_stride = cl->stride_con() * 2; // stride after this unroll
kvn@2694 691 if (next_stride > 0) {
kvn@2694 692 if (iv_type->_lo + next_stride <= iv_type->_lo || // overflow
kvn@2694 693 iv_type->_lo + next_stride > iv_type->_hi) {
duke@435 694 return false; // over-unrolling
duke@435 695 }
kvn@2694 696 } else if (next_stride < 0) {
kvn@2694 697 if (iv_type->_hi + next_stride >= iv_type->_hi || // overflow
kvn@2694 698 iv_type->_hi + next_stride < iv_type->_lo) {
duke@435 699 return false; // over-unrolling
duke@435 700 }
duke@435 701 }
duke@435 702 }
duke@435 703 }
duke@435 704
duke@435 705 // Adjust body_size to determine if we unroll or not
duke@435 706 uint body_size = _body.size();
duke@435 707 // Key test to unroll CaffeineMark's Logic test
duke@435 708 int xors_in_loop = 0;
duke@435 709 // Also count ModL, DivL and MulL which expand mightly
kvn@2694 710 for (uint k = 0; k < _body.size(); k++) {
kvn@2694 711 Node* n = _body.at(k);
kvn@2694 712 switch (n->Opcode()) {
kvn@2694 713 case Op_XorI: xors_in_loop++; break; // CaffeineMark's Logic test
kvn@2694 714 case Op_ModL: body_size += 30; break;
kvn@2694 715 case Op_DivL: body_size += 30; break;
kvn@2694 716 case Op_MulL: body_size += 10; break;
kvn@2694 717 case Op_StrComp:
kvn@2694 718 case Op_StrEquals:
kvn@2694 719 case Op_StrIndexOf:
kvn@2694 720 case Op_AryEq: {
kvn@2694 721 // Do not unroll a loop with String intrinsics code.
kvn@2694 722 // String intrinsics are large and have loops.
kvn@2694 723 return false;
kvn@2694 724 }
kvn@2694 725 } // switch
duke@435 726 }
duke@435 727
duke@435 728 // Check for being too big
kvn@2694 729 if (body_size > (uint)LoopUnrollLimit) {
kvn@2694 730 if (xors_in_loop >= 4 && body_size < (uint)LoopUnrollLimit*4) return true;
duke@435 731 // Normal case: loop too big
duke@435 732 return false;
duke@435 733 }
duke@435 734
duke@435 735 // Unroll once! (Each trip will soon do double iterations)
duke@435 736 return true;
duke@435 737 }
duke@435 738
duke@435 739 //------------------------------policy_align-----------------------------------
duke@435 740 // Return TRUE or FALSE if the loop should be cache-line aligned. Gather the
duke@435 741 // expression that does the alignment. Note that only one array base can be
twisti@1040 742 // aligned in a loop (unless the VM guarantees mutual alignment). Note that
duke@435 743 // if we vectorize short memory ops into longer memory ops, we may want to
duke@435 744 // increase alignment.
duke@435 745 bool IdealLoopTree::policy_align( PhaseIdealLoop *phase ) const {
duke@435 746 return false;
duke@435 747 }
duke@435 748
duke@435 749 //------------------------------policy_range_check-----------------------------
duke@435 750 // Return TRUE or FALSE if the loop should be range-check-eliminated.
duke@435 751 // Actually we do iteration-splitting, a more powerful form of RCE.
duke@435 752 bool IdealLoopTree::policy_range_check( PhaseIdealLoop *phase ) const {
duke@435 753 if( !RangeCheckElimination ) return false;
duke@435 754
duke@435 755 CountedLoopNode *cl = _head->as_CountedLoop();
duke@435 756 // If we unrolled with no intention of doing RCE and we later
duke@435 757 // changed our minds, we got no pre-loop. Either we need to
duke@435 758 // make a new pre-loop, or we gotta disallow RCE.
duke@435 759 if( cl->is_main_no_pre_loop() ) return false; // Disallowed for now.
duke@435 760 Node *trip_counter = cl->phi();
duke@435 761
duke@435 762 // Check loop body for tests of trip-counter plus loop-invariant vs
duke@435 763 // loop-invariant.
duke@435 764 for( uint i = 0; i < _body.size(); i++ ) {
duke@435 765 Node *iff = _body[i];
duke@435 766 if( iff->Opcode() == Op_If ) { // Test?
duke@435 767
duke@435 768 // Comparing trip+off vs limit
duke@435 769 Node *bol = iff->in(1);
duke@435 770 if( bol->req() != 2 ) continue; // dead constant test
cfang@1607 771 if (!bol->is_Bool()) {
cfang@1607 772 assert(UseLoopPredicate && bol->Opcode() == Op_Conv2B, "predicate check only");
cfang@1607 773 continue;
cfang@1607 774 }
duke@435 775 Node *cmp = bol->in(1);
duke@435 776
duke@435 777 Node *rc_exp = cmp->in(1);
duke@435 778 Node *limit = cmp->in(2);
duke@435 779
duke@435 780 Node *limit_c = phase->get_ctrl(limit);
duke@435 781 if( limit_c == phase->C->top() )
duke@435 782 return false; // Found dead test on live IF? No RCE!
duke@435 783 if( is_member(phase->get_loop(limit_c) ) ) {
duke@435 784 // Compare might have operands swapped; commute them
duke@435 785 rc_exp = cmp->in(2);
duke@435 786 limit = cmp->in(1);
duke@435 787 limit_c = phase->get_ctrl(limit);
duke@435 788 if( is_member(phase->get_loop(limit_c) ) )
duke@435 789 continue; // Both inputs are loop varying; cannot RCE
duke@435 790 }
duke@435 791
duke@435 792 if (!phase->is_scaled_iv_plus_offset(rc_exp, trip_counter, NULL, NULL)) {
duke@435 793 continue;
duke@435 794 }
duke@435 795 // Yeah! Found a test like 'trip+off vs limit'
duke@435 796 // Test is an IfNode, has 2 projections. If BOTH are in the loop
duke@435 797 // we need loop unswitching instead of iteration splitting.
duke@435 798 if( is_loop_exit(iff) )
duke@435 799 return true; // Found reason to split iterations
duke@435 800 } // End of is IF
duke@435 801 }
duke@435 802
duke@435 803 return false;
duke@435 804 }
duke@435 805
duke@435 806 //------------------------------policy_peel_only-------------------------------
duke@435 807 // Return TRUE or FALSE if the loop should NEVER be RCE'd or aligned. Useful
duke@435 808 // for unrolling loops with NO array accesses.
duke@435 809 bool IdealLoopTree::policy_peel_only( PhaseIdealLoop *phase ) const {
duke@435 810
duke@435 811 for( uint i = 0; i < _body.size(); i++ )
duke@435 812 if( _body[i]->is_Mem() )
duke@435 813 return false;
duke@435 814
duke@435 815 // No memory accesses at all!
duke@435 816 return true;
duke@435 817 }
duke@435 818
duke@435 819 //------------------------------clone_up_backedge_goo--------------------------
duke@435 820 // If Node n lives in the back_ctrl block and cannot float, we clone a private
duke@435 821 // version of n in preheader_ctrl block and return that, otherwise return n.
duke@435 822 Node *PhaseIdealLoop::clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n ) {
duke@435 823 if( get_ctrl(n) != back_ctrl ) return n;
duke@435 824
duke@435 825 Node *x = NULL; // If required, a clone of 'n'
duke@435 826 // Check for 'n' being pinned in the backedge.
duke@435 827 if( n->in(0) && n->in(0) == back_ctrl ) {
duke@435 828 x = n->clone(); // Clone a copy of 'n' to preheader
duke@435 829 x->set_req( 0, preheader_ctrl ); // Fix x's control input to preheader
duke@435 830 }
duke@435 831
duke@435 832 // Recursive fixup any other input edges into x.
duke@435 833 // If there are no changes we can just return 'n', otherwise
duke@435 834 // we need to clone a private copy and change it.
duke@435 835 for( uint i = 1; i < n->req(); i++ ) {
duke@435 836 Node *g = clone_up_backedge_goo( back_ctrl, preheader_ctrl, n->in(i) );
duke@435 837 if( g != n->in(i) ) {
duke@435 838 if( !x )
duke@435 839 x = n->clone();
duke@435 840 x->set_req(i, g);
duke@435 841 }
duke@435 842 }
duke@435 843 if( x ) { // x can legally float to pre-header location
duke@435 844 register_new_node( x, preheader_ctrl );
duke@435 845 return x;
duke@435 846 } else { // raise n to cover LCA of uses
duke@435 847 set_ctrl( n, find_non_split_ctrl(back_ctrl->in(0)) );
duke@435 848 }
duke@435 849 return n;
duke@435 850 }
duke@435 851
duke@435 852 //------------------------------insert_pre_post_loops--------------------------
duke@435 853 // Insert pre and post loops. If peel_only is set, the pre-loop can not have
duke@435 854 // more iterations added. It acts as a 'peel' only, no lower-bound RCE, no
duke@435 855 // alignment. Useful to unroll loops that do no array accesses.
duke@435 856 void PhaseIdealLoop::insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only ) {
duke@435 857
kvn@2665 858 #ifndef PRODUCT
kvn@2665 859 if (TraceLoopOpts) {
kvn@2665 860 if (peel_only)
kvn@2665 861 tty->print("PeelMainPost ");
kvn@2665 862 else
kvn@2665 863 tty->print("PreMainPost ");
kvn@2665 864 loop->dump_head();
kvn@2665 865 }
kvn@2665 866 #endif
duke@435 867 C->set_major_progress();
duke@435 868
duke@435 869 // Find common pieces of the loop being guarded with pre & post loops
duke@435 870 CountedLoopNode *main_head = loop->_head->as_CountedLoop();
duke@435 871 assert( main_head->is_normal_loop(), "" );
duke@435 872 CountedLoopEndNode *main_end = main_head->loopexit();
duke@435 873 assert( main_end->outcnt() == 2, "1 true, 1 false path only" );
duke@435 874 uint dd_main_head = dom_depth(main_head);
duke@435 875 uint max = main_head->outcnt();
duke@435 876
duke@435 877 Node *pre_header= main_head->in(LoopNode::EntryControl);
duke@435 878 Node *init = main_head->init_trip();
duke@435 879 Node *incr = main_end ->incr();
duke@435 880 Node *limit = main_end ->limit();
duke@435 881 Node *stride = main_end ->stride();
duke@435 882 Node *cmp = main_end ->cmp_node();
duke@435 883 BoolTest::mask b_test = main_end->test_trip();
duke@435 884
duke@435 885 // Need only 1 user of 'bol' because I will be hacking the loop bounds.
duke@435 886 Node *bol = main_end->in(CountedLoopEndNode::TestValue);
duke@435 887 if( bol->outcnt() != 1 ) {
duke@435 888 bol = bol->clone();
duke@435 889 register_new_node(bol,main_end->in(CountedLoopEndNode::TestControl));
duke@435 890 _igvn.hash_delete(main_end);
duke@435 891 main_end->set_req(CountedLoopEndNode::TestValue, bol);
duke@435 892 }
duke@435 893 // Need only 1 user of 'cmp' because I will be hacking the loop bounds.
duke@435 894 if( cmp->outcnt() != 1 ) {
duke@435 895 cmp = cmp->clone();
duke@435 896 register_new_node(cmp,main_end->in(CountedLoopEndNode::TestControl));
duke@435 897 _igvn.hash_delete(bol);
duke@435 898 bol->set_req(1, cmp);
duke@435 899 }
duke@435 900
duke@435 901 //------------------------------
duke@435 902 // Step A: Create Post-Loop.
duke@435 903 Node* main_exit = main_end->proj_out(false);
duke@435 904 assert( main_exit->Opcode() == Op_IfFalse, "" );
duke@435 905 int dd_main_exit = dom_depth(main_exit);
duke@435 906
duke@435 907 // Step A1: Clone the loop body. The clone becomes the post-loop. The main
duke@435 908 // loop pre-header illegally has 2 control users (old & new loops).
duke@435 909 clone_loop( loop, old_new, dd_main_exit );
duke@435 910 assert( old_new[main_end ->_idx]->Opcode() == Op_CountedLoopEnd, "" );
duke@435 911 CountedLoopNode *post_head = old_new[main_head->_idx]->as_CountedLoop();
duke@435 912 post_head->set_post_loop(main_head);
duke@435 913
kvn@835 914 // Reduce the post-loop trip count.
kvn@835 915 CountedLoopEndNode* post_end = old_new[main_end ->_idx]->as_CountedLoopEnd();
kvn@835 916 post_end->_prob = PROB_FAIR;
kvn@835 917
duke@435 918 // Build the main-loop normal exit.
duke@435 919 IfFalseNode *new_main_exit = new (C, 1) IfFalseNode(main_end);
duke@435 920 _igvn.register_new_node_with_optimizer( new_main_exit );
duke@435 921 set_idom(new_main_exit, main_end, dd_main_exit );
duke@435 922 set_loop(new_main_exit, loop->_parent);
duke@435 923
duke@435 924 // Step A2: Build a zero-trip guard for the post-loop. After leaving the
duke@435 925 // main-loop, the post-loop may not execute at all. We 'opaque' the incr
duke@435 926 // (the main-loop trip-counter exit value) because we will be changing
duke@435 927 // the exit value (via unrolling) so we cannot constant-fold away the zero
duke@435 928 // trip guard until all unrolling is done.
kvn@651 929 Node *zer_opaq = new (C, 2) Opaque1Node(C, incr);
duke@435 930 Node *zer_cmp = new (C, 3) CmpINode( zer_opaq, limit );
duke@435 931 Node *zer_bol = new (C, 2) BoolNode( zer_cmp, b_test );
duke@435 932 register_new_node( zer_opaq, new_main_exit );
duke@435 933 register_new_node( zer_cmp , new_main_exit );
duke@435 934 register_new_node( zer_bol , new_main_exit );
duke@435 935
duke@435 936 // Build the IfNode
duke@435 937 IfNode *zer_iff = new (C, 2) IfNode( new_main_exit, zer_bol, PROB_FAIR, COUNT_UNKNOWN );
duke@435 938 _igvn.register_new_node_with_optimizer( zer_iff );
duke@435 939 set_idom(zer_iff, new_main_exit, dd_main_exit);
duke@435 940 set_loop(zer_iff, loop->_parent);
duke@435 941
duke@435 942 // Plug in the false-path, taken if we need to skip post-loop
duke@435 943 _igvn.hash_delete( main_exit );
duke@435 944 main_exit->set_req(0, zer_iff);
duke@435 945 _igvn._worklist.push(main_exit);
duke@435 946 set_idom(main_exit, zer_iff, dd_main_exit);
duke@435 947 set_idom(main_exit->unique_out(), zer_iff, dd_main_exit);
duke@435 948 // Make the true-path, must enter the post loop
duke@435 949 Node *zer_taken = new (C, 1) IfTrueNode( zer_iff );
duke@435 950 _igvn.register_new_node_with_optimizer( zer_taken );
duke@435 951 set_idom(zer_taken, zer_iff, dd_main_exit);
duke@435 952 set_loop(zer_taken, loop->_parent);
duke@435 953 // Plug in the true path
duke@435 954 _igvn.hash_delete( post_head );
duke@435 955 post_head->set_req(LoopNode::EntryControl, zer_taken);
duke@435 956 set_idom(post_head, zer_taken, dd_main_exit);
duke@435 957
duke@435 958 // Step A3: Make the fall-in values to the post-loop come from the
duke@435 959 // fall-out values of the main-loop.
duke@435 960 for (DUIterator_Fast imax, i = main_head->fast_outs(imax); i < imax; i++) {
duke@435 961 Node* main_phi = main_head->fast_out(i);
duke@435 962 if( main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() >0 ) {
duke@435 963 Node *post_phi = old_new[main_phi->_idx];
duke@435 964 Node *fallmain = clone_up_backedge_goo(main_head->back_control(),
duke@435 965 post_head->init_control(),
duke@435 966 main_phi->in(LoopNode::LoopBackControl));
duke@435 967 _igvn.hash_delete(post_phi);
duke@435 968 post_phi->set_req( LoopNode::EntryControl, fallmain );
duke@435 969 }
duke@435 970 }
duke@435 971
duke@435 972 // Update local caches for next stanza
duke@435 973 main_exit = new_main_exit;
duke@435 974
duke@435 975
duke@435 976 //------------------------------
duke@435 977 // Step B: Create Pre-Loop.
duke@435 978
duke@435 979 // Step B1: Clone the loop body. The clone becomes the pre-loop. The main
duke@435 980 // loop pre-header illegally has 2 control users (old & new loops).
duke@435 981 clone_loop( loop, old_new, dd_main_head );
duke@435 982 CountedLoopNode* pre_head = old_new[main_head->_idx]->as_CountedLoop();
duke@435 983 CountedLoopEndNode* pre_end = old_new[main_end ->_idx]->as_CountedLoopEnd();
duke@435 984 pre_head->set_pre_loop(main_head);
duke@435 985 Node *pre_incr = old_new[incr->_idx];
duke@435 986
kvn@835 987 // Reduce the pre-loop trip count.
kvn@835 988 pre_end->_prob = PROB_FAIR;
kvn@835 989
duke@435 990 // Find the pre-loop normal exit.
duke@435 991 Node* pre_exit = pre_end->proj_out(false);
duke@435 992 assert( pre_exit->Opcode() == Op_IfFalse, "" );
duke@435 993 IfFalseNode *new_pre_exit = new (C, 1) IfFalseNode(pre_end);
duke@435 994 _igvn.register_new_node_with_optimizer( new_pre_exit );
duke@435 995 set_idom(new_pre_exit, pre_end, dd_main_head);
duke@435 996 set_loop(new_pre_exit, loop->_parent);
duke@435 997
duke@435 998 // Step B2: Build a zero-trip guard for the main-loop. After leaving the
duke@435 999 // pre-loop, the main-loop may not execute at all. Later in life this
duke@435 1000 // zero-trip guard will become the minimum-trip guard when we unroll
duke@435 1001 // the main-loop.
kvn@651 1002 Node *min_opaq = new (C, 2) Opaque1Node(C, limit);
duke@435 1003 Node *min_cmp = new (C, 3) CmpINode( pre_incr, min_opaq );
duke@435 1004 Node *min_bol = new (C, 2) BoolNode( min_cmp, b_test );
duke@435 1005 register_new_node( min_opaq, new_pre_exit );
duke@435 1006 register_new_node( min_cmp , new_pre_exit );
duke@435 1007 register_new_node( min_bol , new_pre_exit );
duke@435 1008
kvn@835 1009 // Build the IfNode (assume the main-loop is executed always).
kvn@835 1010 IfNode *min_iff = new (C, 2) IfNode( new_pre_exit, min_bol, PROB_ALWAYS, COUNT_UNKNOWN );
duke@435 1011 _igvn.register_new_node_with_optimizer( min_iff );
duke@435 1012 set_idom(min_iff, new_pre_exit, dd_main_head);
duke@435 1013 set_loop(min_iff, loop->_parent);
duke@435 1014
duke@435 1015 // Plug in the false-path, taken if we need to skip main-loop
duke@435 1016 _igvn.hash_delete( pre_exit );
duke@435 1017 pre_exit->set_req(0, min_iff);
duke@435 1018 set_idom(pre_exit, min_iff, dd_main_head);
duke@435 1019 set_idom(pre_exit->unique_out(), min_iff, dd_main_head);
duke@435 1020 // Make the true-path, must enter the main loop
duke@435 1021 Node *min_taken = new (C, 1) IfTrueNode( min_iff );
duke@435 1022 _igvn.register_new_node_with_optimizer( min_taken );
duke@435 1023 set_idom(min_taken, min_iff, dd_main_head);
duke@435 1024 set_loop(min_taken, loop->_parent);
duke@435 1025 // Plug in the true path
duke@435 1026 _igvn.hash_delete( main_head );
duke@435 1027 main_head->set_req(LoopNode::EntryControl, min_taken);
duke@435 1028 set_idom(main_head, min_taken, dd_main_head);
duke@435 1029
duke@435 1030 // Step B3: Make the fall-in values to the main-loop come from the
duke@435 1031 // fall-out values of the pre-loop.
duke@435 1032 for (DUIterator_Fast i2max, i2 = main_head->fast_outs(i2max); i2 < i2max; i2++) {
duke@435 1033 Node* main_phi = main_head->fast_out(i2);
duke@435 1034 if( main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() > 0 ) {
duke@435 1035 Node *pre_phi = old_new[main_phi->_idx];
duke@435 1036 Node *fallpre = clone_up_backedge_goo(pre_head->back_control(),
duke@435 1037 main_head->init_control(),
duke@435 1038 pre_phi->in(LoopNode::LoopBackControl));
duke@435 1039 _igvn.hash_delete(main_phi);
duke@435 1040 main_phi->set_req( LoopNode::EntryControl, fallpre );
duke@435 1041 }
duke@435 1042 }
duke@435 1043
duke@435 1044 // Step B4: Shorten the pre-loop to run only 1 iteration (for now).
duke@435 1045 // RCE and alignment may change this later.
duke@435 1046 Node *cmp_end = pre_end->cmp_node();
duke@435 1047 assert( cmp_end->in(2) == limit, "" );
duke@435 1048 Node *pre_limit = new (C, 3) AddINode( init, stride );
duke@435 1049
duke@435 1050 // Save the original loop limit in this Opaque1 node for
duke@435 1051 // use by range check elimination.
kvn@651 1052 Node *pre_opaq = new (C, 3) Opaque1Node(C, pre_limit, limit);
duke@435 1053
duke@435 1054 register_new_node( pre_limit, pre_head->in(0) );
duke@435 1055 register_new_node( pre_opaq , pre_head->in(0) );
duke@435 1056
duke@435 1057 // Since no other users of pre-loop compare, I can hack limit directly
duke@435 1058 assert( cmp_end->outcnt() == 1, "no other users" );
duke@435 1059 _igvn.hash_delete(cmp_end);
duke@435 1060 cmp_end->set_req(2, peel_only ? pre_limit : pre_opaq);
duke@435 1061
duke@435 1062 // Special case for not-equal loop bounds:
duke@435 1063 // Change pre loop test, main loop test, and the
duke@435 1064 // main loop guard test to use lt or gt depending on stride
duke@435 1065 // direction:
duke@435 1066 // positive stride use <
duke@435 1067 // negative stride use >
duke@435 1068
duke@435 1069 if (pre_end->in(CountedLoopEndNode::TestValue)->as_Bool()->_test._test == BoolTest::ne) {
duke@435 1070
duke@435 1071 BoolTest::mask new_test = (main_end->stride_con() > 0) ? BoolTest::lt : BoolTest::gt;
duke@435 1072 // Modify pre loop end condition
duke@435 1073 Node* pre_bol = pre_end->in(CountedLoopEndNode::TestValue)->as_Bool();
duke@435 1074 BoolNode* new_bol0 = new (C, 2) BoolNode(pre_bol->in(1), new_test);
duke@435 1075 register_new_node( new_bol0, pre_head->in(0) );
duke@435 1076 _igvn.hash_delete(pre_end);
duke@435 1077 pre_end->set_req(CountedLoopEndNode::TestValue, new_bol0);
duke@435 1078 // Modify main loop guard condition
duke@435 1079 assert(min_iff->in(CountedLoopEndNode::TestValue) == min_bol, "guard okay");
duke@435 1080 BoolNode* new_bol1 = new (C, 2) BoolNode(min_bol->in(1), new_test);
duke@435 1081 register_new_node( new_bol1, new_pre_exit );
duke@435 1082 _igvn.hash_delete(min_iff);
duke@435 1083 min_iff->set_req(CountedLoopEndNode::TestValue, new_bol1);
duke@435 1084 // Modify main loop end condition
duke@435 1085 BoolNode* main_bol = main_end->in(CountedLoopEndNode::TestValue)->as_Bool();
duke@435 1086 BoolNode* new_bol2 = new (C, 2) BoolNode(main_bol->in(1), new_test);
duke@435 1087 register_new_node( new_bol2, main_end->in(CountedLoopEndNode::TestControl) );
duke@435 1088 _igvn.hash_delete(main_end);
duke@435 1089 main_end->set_req(CountedLoopEndNode::TestValue, new_bol2);
duke@435 1090 }
duke@435 1091
duke@435 1092 // Flag main loop
duke@435 1093 main_head->set_main_loop();
duke@435 1094 if( peel_only ) main_head->set_main_no_pre_loop();
duke@435 1095
duke@435 1096 // It's difficult to be precise about the trip-counts
duke@435 1097 // for the pre/post loops. They are usually very short,
duke@435 1098 // so guess that 4 trips is a reasonable value.
duke@435 1099 post_head->set_profile_trip_cnt(4.0);
duke@435 1100 pre_head->set_profile_trip_cnt(4.0);
duke@435 1101
duke@435 1102 // Now force out all loop-invariant dominating tests. The optimizer
duke@435 1103 // finds some, but we _know_ they are all useless.
duke@435 1104 peeled_dom_test_elim(loop,old_new);
duke@435 1105 }
duke@435 1106
duke@435 1107 //------------------------------is_invariant-----------------------------
duke@435 1108 // Return true if n is invariant
duke@435 1109 bool IdealLoopTree::is_invariant(Node* n) const {
cfang@1607 1110 Node *n_c = _phase->has_ctrl(n) ? _phase->get_ctrl(n) : n;
duke@435 1111 if (n_c->is_top()) return false;
duke@435 1112 return !is_member(_phase->get_loop(n_c));
duke@435 1113 }
duke@435 1114
duke@435 1115
duke@435 1116 //------------------------------do_unroll--------------------------------------
duke@435 1117 // Unroll the loop body one step - make each trip do 2 iterations.
duke@435 1118 void PhaseIdealLoop::do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip ) {
kvn@2665 1119 assert(LoopUnrollLimit, "");
kvn@2665 1120 CountedLoopNode *loop_head = loop->_head->as_CountedLoop();
kvn@2665 1121 CountedLoopEndNode *loop_end = loop_head->loopexit();
kvn@2665 1122 assert(loop_end, "");
duke@435 1123 #ifndef PRODUCT
kvn@2665 1124 if (PrintOpto && VerifyLoopOptimizations) {
duke@435 1125 tty->print("Unrolling ");
duke@435 1126 loop->dump_head();
kvn@2665 1127 } else if (TraceLoopOpts) {
kvn@2747 1128 if (loop_head->trip_count() < (uint)LoopUnrollLimit) {
kvn@2735 1129 tty->print("Unroll %d(%2d) ", loop_head->unrolled_count()*2, loop_head->trip_count());
kvn@2735 1130 } else {
kvn@2735 1131 tty->print("Unroll %d ", loop_head->unrolled_count()*2);
kvn@2735 1132 }
kvn@2665 1133 loop->dump_head();
duke@435 1134 }
duke@435 1135 #endif
duke@435 1136
duke@435 1137 // Remember loop node count before unrolling to detect
duke@435 1138 // if rounds of unroll,optimize are making progress
duke@435 1139 loop_head->set_node_count_before_unroll(loop->_body.size());
duke@435 1140
duke@435 1141 Node *ctrl = loop_head->in(LoopNode::EntryControl);
duke@435 1142 Node *limit = loop_head->limit();
duke@435 1143 Node *init = loop_head->init_trip();
kvn@2665 1144 Node *stride = loop_head->stride();
duke@435 1145
duke@435 1146 Node *opaq = NULL;
duke@435 1147 if( adjust_min_trip ) { // If not maximally unrolling, need adjustment
duke@435 1148 assert( loop_head->is_main_loop(), "" );
duke@435 1149 assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" );
duke@435 1150 Node *iff = ctrl->in(0);
duke@435 1151 assert( iff->Opcode() == Op_If, "" );
duke@435 1152 Node *bol = iff->in(1);
duke@435 1153 assert( bol->Opcode() == Op_Bool, "" );
duke@435 1154 Node *cmp = bol->in(1);
duke@435 1155 assert( cmp->Opcode() == Op_CmpI, "" );
duke@435 1156 opaq = cmp->in(2);
duke@435 1157 // Occasionally it's possible for a pre-loop Opaque1 node to be
duke@435 1158 // optimized away and then another round of loop opts attempted.
duke@435 1159 // We can not optimize this particular loop in that case.
duke@435 1160 if( opaq->Opcode() != Op_Opaque1 )
duke@435 1161 return; // Cannot find pre-loop! Bail out!
duke@435 1162 }
duke@435 1163
duke@435 1164 C->set_major_progress();
duke@435 1165
duke@435 1166 // Adjust max trip count. The trip count is intentionally rounded
duke@435 1167 // down here (e.g. 15-> 7-> 3-> 1) because if we unwittingly over-unroll,
duke@435 1168 // the main, unrolled, part of the loop will never execute as it is protected
duke@435 1169 // by the min-trip test. See bug 4834191 for a case where we over-unrolled
duke@435 1170 // and later determined that part of the unrolled loop was dead.
duke@435 1171 loop_head->set_trip_count(loop_head->trip_count() / 2);
duke@435 1172
duke@435 1173 // Double the count of original iterations in the unrolled loop body.
duke@435 1174 loop_head->double_unrolled_count();
duke@435 1175
duke@435 1176 // -----------
duke@435 1177 // Step 2: Cut back the trip counter for an unroll amount of 2.
duke@435 1178 // Loop will normally trip (limit - init)/stride_con. Since it's a
duke@435 1179 // CountedLoop this is exact (stride divides limit-init exactly).
duke@435 1180 // We are going to double the loop body, so we want to knock off any
duke@435 1181 // odd iteration: (trip_cnt & ~1). Then back compute a new limit.
duke@435 1182 Node *span = new (C, 3) SubINode( limit, init );
duke@435 1183 register_new_node( span, ctrl );
kvn@2665 1184 Node *trip = new (C, 3) DivINode( 0, span, stride );
duke@435 1185 register_new_node( trip, ctrl );
duke@435 1186 Node *mtwo = _igvn.intcon(-2);
duke@435 1187 set_ctrl(mtwo, C->root());
duke@435 1188 Node *rond = new (C, 3) AndINode( trip, mtwo );
duke@435 1189 register_new_node( rond, ctrl );
kvn@2665 1190 Node *spn2 = new (C, 3) MulINode( rond, stride );
duke@435 1191 register_new_node( spn2, ctrl );
duke@435 1192 Node *lim2 = new (C, 3) AddINode( spn2, init );
duke@435 1193 register_new_node( lim2, ctrl );
duke@435 1194
duke@435 1195 // Hammer in the new limit
duke@435 1196 Node *ctrl2 = loop_end->in(0);
duke@435 1197 Node *cmp2 = new (C, 3) CmpINode( loop_head->incr(), lim2 );
duke@435 1198 register_new_node( cmp2, ctrl2 );
duke@435 1199 Node *bol2 = new (C, 2) BoolNode( cmp2, loop_end->test_trip() );
duke@435 1200 register_new_node( bol2, ctrl2 );
duke@435 1201 _igvn.hash_delete(loop_end);
duke@435 1202 loop_end->set_req(CountedLoopEndNode::TestValue, bol2);
duke@435 1203
duke@435 1204 // Step 3: Find the min-trip test guaranteed before a 'main' loop.
duke@435 1205 // Make it a 1-trip test (means at least 2 trips).
duke@435 1206 if( adjust_min_trip ) {
duke@435 1207 // Guard test uses an 'opaque' node which is not shared. Hence I
duke@435 1208 // can edit it's inputs directly. Hammer in the new limit for the
duke@435 1209 // minimum-trip guard.
duke@435 1210 assert( opaq->outcnt() == 1, "" );
duke@435 1211 _igvn.hash_delete(opaq);
duke@435 1212 opaq->set_req(1, lim2);
duke@435 1213 }
duke@435 1214
duke@435 1215 // ---------
duke@435 1216 // Step 4: Clone the loop body. Move it inside the loop. This loop body
duke@435 1217 // represents the odd iterations; since the loop trips an even number of
duke@435 1218 // times its backedge is never taken. Kill the backedge.
duke@435 1219 uint dd = dom_depth(loop_head);
duke@435 1220 clone_loop( loop, old_new, dd );
duke@435 1221
duke@435 1222 // Make backedges of the clone equal to backedges of the original.
duke@435 1223 // Make the fall-in from the original come from the fall-out of the clone.
duke@435 1224 for (DUIterator_Fast jmax, j = loop_head->fast_outs(jmax); j < jmax; j++) {
duke@435 1225 Node* phi = loop_head->fast_out(j);
duke@435 1226 if( phi->is_Phi() && phi->in(0) == loop_head && phi->outcnt() > 0 ) {
duke@435 1227 Node *newphi = old_new[phi->_idx];
duke@435 1228 _igvn.hash_delete( phi );
duke@435 1229 _igvn.hash_delete( newphi );
duke@435 1230
duke@435 1231 phi ->set_req(LoopNode:: EntryControl, newphi->in(LoopNode::LoopBackControl));
duke@435 1232 newphi->set_req(LoopNode::LoopBackControl, phi ->in(LoopNode::LoopBackControl));
duke@435 1233 phi ->set_req(LoopNode::LoopBackControl, C->top());
duke@435 1234 }
duke@435 1235 }
duke@435 1236 Node *clone_head = old_new[loop_head->_idx];
duke@435 1237 _igvn.hash_delete( clone_head );
duke@435 1238 loop_head ->set_req(LoopNode:: EntryControl, clone_head->in(LoopNode::LoopBackControl));
duke@435 1239 clone_head->set_req(LoopNode::LoopBackControl, loop_head ->in(LoopNode::LoopBackControl));
duke@435 1240 loop_head ->set_req(LoopNode::LoopBackControl, C->top());
duke@435 1241 loop->_head = clone_head; // New loop header
duke@435 1242
duke@435 1243 set_idom(loop_head, loop_head ->in(LoopNode::EntryControl), dd);
duke@435 1244 set_idom(clone_head, clone_head->in(LoopNode::EntryControl), dd);
duke@435 1245
duke@435 1246 // Kill the clone's backedge
duke@435 1247 Node *newcle = old_new[loop_end->_idx];
duke@435 1248 _igvn.hash_delete( newcle );
duke@435 1249 Node *one = _igvn.intcon(1);
duke@435 1250 set_ctrl(one, C->root());
duke@435 1251 newcle->set_req(1, one);
duke@435 1252 // Force clone into same loop body
duke@435 1253 uint max = loop->_body.size();
duke@435 1254 for( uint k = 0; k < max; k++ ) {
duke@435 1255 Node *old = loop->_body.at(k);
duke@435 1256 Node *nnn = old_new[old->_idx];
duke@435 1257 loop->_body.push(nnn);
duke@435 1258 if (!has_ctrl(old))
duke@435 1259 set_loop(nnn, loop);
duke@435 1260 }
never@802 1261
never@802 1262 loop->record_for_igvn();
duke@435 1263 }
duke@435 1264
duke@435 1265 //------------------------------do_maximally_unroll----------------------------
duke@435 1266
duke@435 1267 void PhaseIdealLoop::do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new ) {
duke@435 1268 CountedLoopNode *cl = loop->_head->as_CountedLoop();
kvn@2665 1269 assert(cl->trip_count() > 0, "");
kvn@2665 1270 #ifndef PRODUCT
kvn@2665 1271 if (TraceLoopOpts) {
kvn@2665 1272 tty->print("MaxUnroll %d ", cl->trip_count());
kvn@2665 1273 loop->dump_head();
kvn@2665 1274 }
kvn@2665 1275 #endif
duke@435 1276
duke@435 1277 // If loop is tripping an odd number of times, peel odd iteration
kvn@2665 1278 if ((cl->trip_count() & 1) == 1) {
kvn@2665 1279 do_peeling(loop, old_new);
duke@435 1280 }
duke@435 1281
duke@435 1282 // Now its tripping an even number of times remaining. Double loop body.
duke@435 1283 // Do not adjust pre-guards; they are not needed and do not exist.
kvn@2665 1284 if (cl->trip_count() > 0) {
kvn@2665 1285 do_unroll(loop, old_new, false);
duke@435 1286 }
duke@435 1287 }
duke@435 1288
duke@435 1289 //------------------------------dominates_backedge---------------------------------
duke@435 1290 // Returns true if ctrl is executed on every complete iteration
duke@435 1291 bool IdealLoopTree::dominates_backedge(Node* ctrl) {
duke@435 1292 assert(ctrl->is_CFG(), "must be control");
duke@435 1293 Node* backedge = _head->as_Loop()->in(LoopNode::LoopBackControl);
duke@435 1294 return _phase->dom_lca_internal(ctrl, backedge) == ctrl;
duke@435 1295 }
duke@435 1296
duke@435 1297 //------------------------------add_constraint---------------------------------
duke@435 1298 // Constrain the main loop iterations so the condition:
duke@435 1299 // scale_con * I + offset < limit
duke@435 1300 // always holds true. That is, either increase the number of iterations in
duke@435 1301 // the pre-loop or the post-loop until the condition holds true in the main
duke@435 1302 // loop. Stride, scale, offset and limit are all loop invariant. Further,
duke@435 1303 // stride and scale are constants (offset and limit often are).
duke@435 1304 void PhaseIdealLoop::add_constraint( int stride_con, int scale_con, Node *offset, Node *limit, Node *pre_ctrl, Node **pre_limit, Node **main_limit ) {
duke@435 1305
duke@435 1306 // Compute "I :: (limit-offset)/scale_con"
duke@435 1307 Node *con = new (C, 3) SubINode( limit, offset );
duke@435 1308 register_new_node( con, pre_ctrl );
duke@435 1309 Node *scale = _igvn.intcon(scale_con);
duke@435 1310 set_ctrl(scale, C->root());
duke@435 1311 Node *X = new (C, 3) DivINode( 0, con, scale );
duke@435 1312 register_new_node( X, pre_ctrl );
duke@435 1313
duke@435 1314 // For positive stride, the pre-loop limit always uses a MAX function
duke@435 1315 // and the main loop a MIN function. For negative stride these are
duke@435 1316 // reversed.
duke@435 1317
duke@435 1318 // Also for positive stride*scale the affine function is increasing, so the
duke@435 1319 // pre-loop must check for underflow and the post-loop for overflow.
duke@435 1320 // Negative stride*scale reverses this; pre-loop checks for overflow and
duke@435 1321 // post-loop for underflow.
duke@435 1322 if( stride_con*scale_con > 0 ) {
duke@435 1323 // Compute I < (limit-offset)/scale_con
duke@435 1324 // Adjust main-loop last iteration to be MIN/MAX(main_loop,X)
duke@435 1325 *main_limit = (stride_con > 0)
duke@435 1326 ? (Node*)(new (C, 3) MinINode( *main_limit, X ))
duke@435 1327 : (Node*)(new (C, 3) MaxINode( *main_limit, X ));
duke@435 1328 register_new_node( *main_limit, pre_ctrl );
duke@435 1329
duke@435 1330 } else {
duke@435 1331 // Compute (limit-offset)/scale_con + SGN(-scale_con) <= I
duke@435 1332 // Add the negation of the main-loop constraint to the pre-loop.
duke@435 1333 // See footnote [++] below for a derivation of the limit expression.
duke@435 1334 Node *incr = _igvn.intcon(scale_con > 0 ? -1 : 1);
duke@435 1335 set_ctrl(incr, C->root());
duke@435 1336 Node *adj = new (C, 3) AddINode( X, incr );
duke@435 1337 register_new_node( adj, pre_ctrl );
duke@435 1338 *pre_limit = (scale_con > 0)
duke@435 1339 ? (Node*)new (C, 3) MinINode( *pre_limit, adj )
duke@435 1340 : (Node*)new (C, 3) MaxINode( *pre_limit, adj );
duke@435 1341 register_new_node( *pre_limit, pre_ctrl );
duke@435 1342
duke@435 1343 // [++] Here's the algebra that justifies the pre-loop limit expression:
duke@435 1344 //
duke@435 1345 // NOT( scale_con * I + offset < limit )
duke@435 1346 // ==
duke@435 1347 // scale_con * I + offset >= limit
duke@435 1348 // ==
duke@435 1349 // SGN(scale_con) * I >= (limit-offset)/|scale_con|
duke@435 1350 // ==
duke@435 1351 // (limit-offset)/|scale_con| <= I * SGN(scale_con)
duke@435 1352 // ==
duke@435 1353 // (limit-offset)/|scale_con|-1 < I * SGN(scale_con)
duke@435 1354 // ==
duke@435 1355 // ( if (scale_con > 0) /*common case*/
duke@435 1356 // (limit-offset)/scale_con - 1 < I
duke@435 1357 // else
duke@435 1358 // (limit-offset)/scale_con + 1 > I
duke@435 1359 // )
duke@435 1360 // ( if (scale_con > 0) /*common case*/
duke@435 1361 // (limit-offset)/scale_con + SGN(-scale_con) < I
duke@435 1362 // else
duke@435 1363 // (limit-offset)/scale_con + SGN(-scale_con) > I
duke@435 1364 }
duke@435 1365 }
duke@435 1366
duke@435 1367
duke@435 1368 //------------------------------is_scaled_iv---------------------------------
duke@435 1369 // Return true if exp is a constant times an induction var
duke@435 1370 bool PhaseIdealLoop::is_scaled_iv(Node* exp, Node* iv, int* p_scale) {
duke@435 1371 if (exp == iv) {
duke@435 1372 if (p_scale != NULL) {
duke@435 1373 *p_scale = 1;
duke@435 1374 }
duke@435 1375 return true;
duke@435 1376 }
duke@435 1377 int opc = exp->Opcode();
duke@435 1378 if (opc == Op_MulI) {
duke@435 1379 if (exp->in(1) == iv && exp->in(2)->is_Con()) {
duke@435 1380 if (p_scale != NULL) {
duke@435 1381 *p_scale = exp->in(2)->get_int();
duke@435 1382 }
duke@435 1383 return true;
duke@435 1384 }
duke@435 1385 if (exp->in(2) == iv && exp->in(1)->is_Con()) {
duke@435 1386 if (p_scale != NULL) {
duke@435 1387 *p_scale = exp->in(1)->get_int();
duke@435 1388 }
duke@435 1389 return true;
duke@435 1390 }
duke@435 1391 } else if (opc == Op_LShiftI) {
duke@435 1392 if (exp->in(1) == iv && exp->in(2)->is_Con()) {
duke@435 1393 if (p_scale != NULL) {
duke@435 1394 *p_scale = 1 << exp->in(2)->get_int();
duke@435 1395 }
duke@435 1396 return true;
duke@435 1397 }
duke@435 1398 }
duke@435 1399 return false;
duke@435 1400 }
duke@435 1401
duke@435 1402 //-----------------------------is_scaled_iv_plus_offset------------------------------
duke@435 1403 // Return true if exp is a simple induction variable expression: k1*iv + (invar + k2)
duke@435 1404 bool PhaseIdealLoop::is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth) {
duke@435 1405 if (is_scaled_iv(exp, iv, p_scale)) {
duke@435 1406 if (p_offset != NULL) {
duke@435 1407 Node *zero = _igvn.intcon(0);
duke@435 1408 set_ctrl(zero, C->root());
duke@435 1409 *p_offset = zero;
duke@435 1410 }
duke@435 1411 return true;
duke@435 1412 }
duke@435 1413 int opc = exp->Opcode();
duke@435 1414 if (opc == Op_AddI) {
duke@435 1415 if (is_scaled_iv(exp->in(1), iv, p_scale)) {
duke@435 1416 if (p_offset != NULL) {
duke@435 1417 *p_offset = exp->in(2);
duke@435 1418 }
duke@435 1419 return true;
duke@435 1420 }
duke@435 1421 if (exp->in(2)->is_Con()) {
duke@435 1422 Node* offset2 = NULL;
duke@435 1423 if (depth < 2 &&
duke@435 1424 is_scaled_iv_plus_offset(exp->in(1), iv, p_scale,
duke@435 1425 p_offset != NULL ? &offset2 : NULL, depth+1)) {
duke@435 1426 if (p_offset != NULL) {
duke@435 1427 Node *ctrl_off2 = get_ctrl(offset2);
duke@435 1428 Node* offset = new (C, 3) AddINode(offset2, exp->in(2));
duke@435 1429 register_new_node(offset, ctrl_off2);
duke@435 1430 *p_offset = offset;
duke@435 1431 }
duke@435 1432 return true;
duke@435 1433 }
duke@435 1434 }
duke@435 1435 } else if (opc == Op_SubI) {
duke@435 1436 if (is_scaled_iv(exp->in(1), iv, p_scale)) {
duke@435 1437 if (p_offset != NULL) {
duke@435 1438 Node *zero = _igvn.intcon(0);
duke@435 1439 set_ctrl(zero, C->root());
duke@435 1440 Node *ctrl_off = get_ctrl(exp->in(2));
duke@435 1441 Node* offset = new (C, 3) SubINode(zero, exp->in(2));
duke@435 1442 register_new_node(offset, ctrl_off);
duke@435 1443 *p_offset = offset;
duke@435 1444 }
duke@435 1445 return true;
duke@435 1446 }
duke@435 1447 if (is_scaled_iv(exp->in(2), iv, p_scale)) {
duke@435 1448 if (p_offset != NULL) {
duke@435 1449 *p_scale *= -1;
duke@435 1450 *p_offset = exp->in(1);
duke@435 1451 }
duke@435 1452 return true;
duke@435 1453 }
duke@435 1454 }
duke@435 1455 return false;
duke@435 1456 }
duke@435 1457
duke@435 1458 //------------------------------do_range_check---------------------------------
duke@435 1459 // Eliminate range-checks and other trip-counter vs loop-invariant tests.
duke@435 1460 void PhaseIdealLoop::do_range_check( IdealLoopTree *loop, Node_List &old_new ) {
duke@435 1461 #ifndef PRODUCT
kvn@2665 1462 if (PrintOpto && VerifyLoopOptimizations) {
duke@435 1463 tty->print("Range Check Elimination ");
duke@435 1464 loop->dump_head();
kvn@2665 1465 } else if (TraceLoopOpts) {
kvn@2665 1466 tty->print("RangeCheck ");
kvn@2665 1467 loop->dump_head();
duke@435 1468 }
duke@435 1469 #endif
kvn@2665 1470 assert(RangeCheckElimination, "");
duke@435 1471 CountedLoopNode *cl = loop->_head->as_CountedLoop();
kvn@2665 1472 assert(cl->is_main_loop(), "");
kvn@2665 1473
kvn@2665 1474 // protect against stride not being a constant
kvn@2665 1475 if (!cl->stride_is_con())
kvn@2665 1476 return;
duke@435 1477
duke@435 1478 // Find the trip counter; we are iteration splitting based on it
duke@435 1479 Node *trip_counter = cl->phi();
duke@435 1480 // Find the main loop limit; we will trim it's iterations
duke@435 1481 // to not ever trip end tests
duke@435 1482 Node *main_limit = cl->limit();
kvn@2665 1483
kvn@2665 1484 // Need to find the main-loop zero-trip guard
kvn@2665 1485 Node *ctrl = cl->in(LoopNode::EntryControl);
kvn@2665 1486 assert(ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "");
kvn@2665 1487 Node *iffm = ctrl->in(0);
kvn@2665 1488 assert(iffm->Opcode() == Op_If, "");
kvn@2665 1489 Node *bolzm = iffm->in(1);
kvn@2665 1490 assert(bolzm->Opcode() == Op_Bool, "");
kvn@2665 1491 Node *cmpzm = bolzm->in(1);
kvn@2665 1492 assert(cmpzm->is_Cmp(), "");
kvn@2665 1493 Node *opqzm = cmpzm->in(2);
kvn@2665 1494 // Can not optimize a loop if pre-loop Opaque1 node is optimized
kvn@2665 1495 // away and then another round of loop opts attempted.
kvn@2665 1496 if (opqzm->Opcode() != Op_Opaque1)
kvn@2665 1497 return;
kvn@2665 1498 assert(opqzm->in(1) == main_limit, "do not understand situation");
kvn@2665 1499
duke@435 1500 // Find the pre-loop limit; we will expand it's iterations to
duke@435 1501 // not ever trip low tests.
duke@435 1502 Node *p_f = iffm->in(0);
kvn@2665 1503 assert(p_f->Opcode() == Op_IfFalse, "");
duke@435 1504 CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd();
kvn@2665 1505 assert(pre_end->loopnode()->is_pre_loop(), "");
duke@435 1506 Node *pre_opaq1 = pre_end->limit();
duke@435 1507 // Occasionally it's possible for a pre-loop Opaque1 node to be
duke@435 1508 // optimized away and then another round of loop opts attempted.
duke@435 1509 // We can not optimize this particular loop in that case.
kvn@2665 1510 if (pre_opaq1->Opcode() != Op_Opaque1)
duke@435 1511 return;
duke@435 1512 Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1;
duke@435 1513 Node *pre_limit = pre_opaq->in(1);
duke@435 1514
duke@435 1515 // Where do we put new limit calculations
duke@435 1516 Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl);
duke@435 1517
duke@435 1518 // Ensure the original loop limit is available from the
duke@435 1519 // pre-loop Opaque1 node.
duke@435 1520 Node *orig_limit = pre_opaq->original_loop_limit();
kvn@2665 1521 if (orig_limit == NULL || _igvn.type(orig_limit) == Type::TOP)
duke@435 1522 return;
duke@435 1523
duke@435 1524 // Must know if its a count-up or count-down loop
duke@435 1525
duke@435 1526 int stride_con = cl->stride_con();
duke@435 1527 Node *zero = _igvn.intcon(0);
duke@435 1528 Node *one = _igvn.intcon(1);
duke@435 1529 set_ctrl(zero, C->root());
duke@435 1530 set_ctrl(one, C->root());
duke@435 1531
duke@435 1532 // Range checks that do not dominate the loop backedge (ie.
duke@435 1533 // conditionally executed) can lengthen the pre loop limit beyond
duke@435 1534 // the original loop limit. To prevent this, the pre limit is
duke@435 1535 // (for stride > 0) MINed with the original loop limit (MAXed
duke@435 1536 // stride < 0) when some range_check (rc) is conditionally
duke@435 1537 // executed.
duke@435 1538 bool conditional_rc = false;
duke@435 1539
duke@435 1540 // Check loop body for tests of trip-counter plus loop-invariant vs
duke@435 1541 // loop-invariant.
duke@435 1542 for( uint i = 0; i < loop->_body.size(); i++ ) {
duke@435 1543 Node *iff = loop->_body[i];
duke@435 1544 if( iff->Opcode() == Op_If ) { // Test?
duke@435 1545
duke@435 1546 // Test is an IfNode, has 2 projections. If BOTH are in the loop
duke@435 1547 // we need loop unswitching instead of iteration splitting.
duke@435 1548 Node *exit = loop->is_loop_exit(iff);
duke@435 1549 if( !exit ) continue;
duke@435 1550 int flip = (exit->Opcode() == Op_IfTrue) ? 1 : 0;
duke@435 1551
duke@435 1552 // Get boolean condition to test
duke@435 1553 Node *i1 = iff->in(1);
duke@435 1554 if( !i1->is_Bool() ) continue;
duke@435 1555 BoolNode *bol = i1->as_Bool();
duke@435 1556 BoolTest b_test = bol->_test;
duke@435 1557 // Flip sense of test if exit condition is flipped
duke@435 1558 if( flip )
duke@435 1559 b_test = b_test.negate();
duke@435 1560
duke@435 1561 // Get compare
duke@435 1562 Node *cmp = bol->in(1);
duke@435 1563
duke@435 1564 // Look for trip_counter + offset vs limit
duke@435 1565 Node *rc_exp = cmp->in(1);
duke@435 1566 Node *limit = cmp->in(2);
duke@435 1567 jint scale_con= 1; // Assume trip counter not scaled
duke@435 1568
duke@435 1569 Node *limit_c = get_ctrl(limit);
duke@435 1570 if( loop->is_member(get_loop(limit_c) ) ) {
duke@435 1571 // Compare might have operands swapped; commute them
duke@435 1572 b_test = b_test.commute();
duke@435 1573 rc_exp = cmp->in(2);
duke@435 1574 limit = cmp->in(1);
duke@435 1575 limit_c = get_ctrl(limit);
duke@435 1576 if( loop->is_member(get_loop(limit_c) ) )
duke@435 1577 continue; // Both inputs are loop varying; cannot RCE
duke@435 1578 }
duke@435 1579 // Here we know 'limit' is loop invariant
duke@435 1580
duke@435 1581 // 'limit' maybe pinned below the zero trip test (probably from a
duke@435 1582 // previous round of rce), in which case, it can't be used in the
duke@435 1583 // zero trip test expression which must occur before the zero test's if.
duke@435 1584 if( limit_c == ctrl ) {
duke@435 1585 continue; // Don't rce this check but continue looking for other candidates.
duke@435 1586 }
duke@435 1587
duke@435 1588 // Check for scaled induction variable plus an offset
duke@435 1589 Node *offset = NULL;
duke@435 1590
duke@435 1591 if (!is_scaled_iv_plus_offset(rc_exp, trip_counter, &scale_con, &offset)) {
duke@435 1592 continue;
duke@435 1593 }
duke@435 1594
duke@435 1595 Node *offset_c = get_ctrl(offset);
duke@435 1596 if( loop->is_member( get_loop(offset_c) ) )
duke@435 1597 continue; // Offset is not really loop invariant
duke@435 1598 // Here we know 'offset' is loop invariant.
duke@435 1599
duke@435 1600 // As above for the 'limit', the 'offset' maybe pinned below the
duke@435 1601 // zero trip test.
duke@435 1602 if( offset_c == ctrl ) {
duke@435 1603 continue; // Don't rce this check but continue looking for other candidates.
duke@435 1604 }
duke@435 1605
duke@435 1606 // At this point we have the expression as:
duke@435 1607 // scale_con * trip_counter + offset :: limit
duke@435 1608 // where scale_con, offset and limit are loop invariant. Trip_counter
duke@435 1609 // monotonically increases by stride_con, a constant. Both (or either)
duke@435 1610 // stride_con and scale_con can be negative which will flip about the
duke@435 1611 // sense of the test.
duke@435 1612
duke@435 1613 // Adjust pre and main loop limits to guard the correct iteration set
duke@435 1614 if( cmp->Opcode() == Op_CmpU ) {// Unsigned compare is really 2 tests
duke@435 1615 if( b_test._test == BoolTest::lt ) { // Range checks always use lt
duke@435 1616 // The overflow limit: scale*I+offset < limit
duke@435 1617 add_constraint( stride_con, scale_con, offset, limit, pre_ctrl, &pre_limit, &main_limit );
duke@435 1618 // The underflow limit: 0 <= scale*I+offset.
duke@435 1619 // Some math yields: -scale*I-(offset+1) < 0
duke@435 1620 Node *plus_one = new (C, 3) AddINode( offset, one );
duke@435 1621 register_new_node( plus_one, pre_ctrl );
duke@435 1622 Node *neg_offset = new (C, 3) SubINode( zero, plus_one );
duke@435 1623 register_new_node( neg_offset, pre_ctrl );
duke@435 1624 add_constraint( stride_con, -scale_con, neg_offset, zero, pre_ctrl, &pre_limit, &main_limit );
duke@435 1625 if (!conditional_rc) {
duke@435 1626 conditional_rc = !loop->dominates_backedge(iff);
duke@435 1627 }
duke@435 1628 } else {
duke@435 1629 #ifndef PRODUCT
duke@435 1630 if( PrintOpto )
duke@435 1631 tty->print_cr("missed RCE opportunity");
duke@435 1632 #endif
duke@435 1633 continue; // In release mode, ignore it
duke@435 1634 }
duke@435 1635 } else { // Otherwise work on normal compares
duke@435 1636 switch( b_test._test ) {
duke@435 1637 case BoolTest::ge: // Convert X >= Y to -X <= -Y
duke@435 1638 scale_con = -scale_con;
duke@435 1639 offset = new (C, 3) SubINode( zero, offset );
duke@435 1640 register_new_node( offset, pre_ctrl );
duke@435 1641 limit = new (C, 3) SubINode( zero, limit );
duke@435 1642 register_new_node( limit, pre_ctrl );
duke@435 1643 // Fall into LE case
duke@435 1644 case BoolTest::le: // Convert X <= Y to X < Y+1
duke@435 1645 limit = new (C, 3) AddINode( limit, one );
duke@435 1646 register_new_node( limit, pre_ctrl );
duke@435 1647 // Fall into LT case
duke@435 1648 case BoolTest::lt:
duke@435 1649 add_constraint( stride_con, scale_con, offset, limit, pre_ctrl, &pre_limit, &main_limit );
duke@435 1650 if (!conditional_rc) {
duke@435 1651 conditional_rc = !loop->dominates_backedge(iff);
duke@435 1652 }
duke@435 1653 break;
duke@435 1654 default:
duke@435 1655 #ifndef PRODUCT
duke@435 1656 if( PrintOpto )
duke@435 1657 tty->print_cr("missed RCE opportunity");
duke@435 1658 #endif
duke@435 1659 continue; // Unhandled case
duke@435 1660 }
duke@435 1661 }
duke@435 1662
duke@435 1663 // Kill the eliminated test
duke@435 1664 C->set_major_progress();
duke@435 1665 Node *kill_con = _igvn.intcon( 1-flip );
duke@435 1666 set_ctrl(kill_con, C->root());
duke@435 1667 _igvn.hash_delete(iff);
duke@435 1668 iff->set_req(1, kill_con);
duke@435 1669 _igvn._worklist.push(iff);
duke@435 1670 // Find surviving projection
duke@435 1671 assert(iff->is_If(), "");
duke@435 1672 ProjNode* dp = ((IfNode*)iff)->proj_out(1-flip);
duke@435 1673 // Find loads off the surviving projection; remove their control edge
duke@435 1674 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) {
duke@435 1675 Node* cd = dp->fast_out(i); // Control-dependent node
duke@435 1676 if( cd->is_Load() ) { // Loads can now float around in the loop
duke@435 1677 _igvn.hash_delete(cd);
duke@435 1678 // Allow the load to float around in the loop, or before it
duke@435 1679 // but NOT before the pre-loop.
duke@435 1680 cd->set_req(0, ctrl); // ctrl, not NULL
duke@435 1681 _igvn._worklist.push(cd);
duke@435 1682 --i;
duke@435 1683 --imax;
duke@435 1684 }
duke@435 1685 }
duke@435 1686
duke@435 1687 } // End of is IF
duke@435 1688
duke@435 1689 }
duke@435 1690
duke@435 1691 // Update loop limits
duke@435 1692 if (conditional_rc) {
duke@435 1693 pre_limit = (stride_con > 0) ? (Node*)new (C,3) MinINode(pre_limit, orig_limit)
duke@435 1694 : (Node*)new (C,3) MaxINode(pre_limit, orig_limit);
duke@435 1695 register_new_node(pre_limit, pre_ctrl);
duke@435 1696 }
duke@435 1697 _igvn.hash_delete(pre_opaq);
duke@435 1698 pre_opaq->set_req(1, pre_limit);
duke@435 1699
duke@435 1700 // Note:: we are making the main loop limit no longer precise;
duke@435 1701 // need to round up based on stride.
duke@435 1702 if( stride_con != 1 && stride_con != -1 ) { // Cutout for common case
duke@435 1703 // "Standard" round-up logic: ([main_limit-init+(y-1)]/y)*y+init
duke@435 1704 // Hopefully, compiler will optimize for powers of 2.
duke@435 1705 Node *ctrl = get_ctrl(main_limit);
duke@435 1706 Node *stride = cl->stride();
duke@435 1707 Node *init = cl->init_trip();
duke@435 1708 Node *span = new (C, 3) SubINode(main_limit,init);
duke@435 1709 register_new_node(span,ctrl);
duke@435 1710 Node *rndup = _igvn.intcon(stride_con + ((stride_con>0)?-1:1));
duke@435 1711 Node *add = new (C, 3) AddINode(span,rndup);
duke@435 1712 register_new_node(add,ctrl);
duke@435 1713 Node *div = new (C, 3) DivINode(0,add,stride);
duke@435 1714 register_new_node(div,ctrl);
duke@435 1715 Node *mul = new (C, 3) MulINode(div,stride);
duke@435 1716 register_new_node(mul,ctrl);
duke@435 1717 Node *newlim = new (C, 3) AddINode(mul,init);
duke@435 1718 register_new_node(newlim,ctrl);
duke@435 1719 main_limit = newlim;
duke@435 1720 }
duke@435 1721
duke@435 1722 Node *main_cle = cl->loopexit();
duke@435 1723 Node *main_bol = main_cle->in(1);
duke@435 1724 // Hacking loop bounds; need private copies of exit test
duke@435 1725 if( main_bol->outcnt() > 1 ) {// BoolNode shared?
duke@435 1726 _igvn.hash_delete(main_cle);
duke@435 1727 main_bol = main_bol->clone();// Clone a private BoolNode
duke@435 1728 register_new_node( main_bol, main_cle->in(0) );
duke@435 1729 main_cle->set_req(1,main_bol);
duke@435 1730 }
duke@435 1731 Node *main_cmp = main_bol->in(1);
duke@435 1732 if( main_cmp->outcnt() > 1 ) { // CmpNode shared?
duke@435 1733 _igvn.hash_delete(main_bol);
duke@435 1734 main_cmp = main_cmp->clone();// Clone a private CmpNode
duke@435 1735 register_new_node( main_cmp, main_cle->in(0) );
duke@435 1736 main_bol->set_req(1,main_cmp);
duke@435 1737 }
duke@435 1738 // Hack the now-private loop bounds
duke@435 1739 _igvn.hash_delete(main_cmp);
duke@435 1740 main_cmp->set_req(2, main_limit);
duke@435 1741 _igvn._worklist.push(main_cmp);
duke@435 1742 // The OpaqueNode is unshared by design
duke@435 1743 _igvn.hash_delete(opqzm);
duke@435 1744 assert( opqzm->outcnt() == 1, "cannot hack shared node" );
duke@435 1745 opqzm->set_req(1,main_limit);
duke@435 1746 _igvn._worklist.push(opqzm);
duke@435 1747 }
duke@435 1748
duke@435 1749 //------------------------------DCE_loop_body----------------------------------
duke@435 1750 // Remove simplistic dead code from loop body
duke@435 1751 void IdealLoopTree::DCE_loop_body() {
duke@435 1752 for( uint i = 0; i < _body.size(); i++ )
duke@435 1753 if( _body.at(i)->outcnt() == 0 )
duke@435 1754 _body.map( i--, _body.pop() );
duke@435 1755 }
duke@435 1756
duke@435 1757
duke@435 1758 //------------------------------adjust_loop_exit_prob--------------------------
duke@435 1759 // Look for loop-exit tests with the 50/50 (or worse) guesses from the parsing stage.
duke@435 1760 // Replace with a 1-in-10 exit guess.
duke@435 1761 void IdealLoopTree::adjust_loop_exit_prob( PhaseIdealLoop *phase ) {
duke@435 1762 Node *test = tail();
duke@435 1763 while( test != _head ) {
duke@435 1764 uint top = test->Opcode();
duke@435 1765 if( top == Op_IfTrue || top == Op_IfFalse ) {
duke@435 1766 int test_con = ((ProjNode*)test)->_con;
duke@435 1767 assert(top == (uint)(test_con? Op_IfTrue: Op_IfFalse), "sanity");
duke@435 1768 IfNode *iff = test->in(0)->as_If();
duke@435 1769 if( iff->outcnt() == 2 ) { // Ignore dead tests
duke@435 1770 Node *bol = iff->in(1);
duke@435 1771 if( bol && bol->req() > 1 && bol->in(1) &&
duke@435 1772 ((bol->in(1)->Opcode() == Op_StorePConditional ) ||
kvn@855 1773 (bol->in(1)->Opcode() == Op_StoreIConditional ) ||
duke@435 1774 (bol->in(1)->Opcode() == Op_StoreLConditional ) ||
duke@435 1775 (bol->in(1)->Opcode() == Op_CompareAndSwapI ) ||
duke@435 1776 (bol->in(1)->Opcode() == Op_CompareAndSwapL ) ||
coleenp@548 1777 (bol->in(1)->Opcode() == Op_CompareAndSwapP ) ||
coleenp@548 1778 (bol->in(1)->Opcode() == Op_CompareAndSwapN )))
duke@435 1779 return; // Allocation loops RARELY take backedge
duke@435 1780 // Find the OTHER exit path from the IF
duke@435 1781 Node* ex = iff->proj_out(1-test_con);
duke@435 1782 float p = iff->_prob;
duke@435 1783 if( !phase->is_member( this, ex ) && iff->_fcnt == COUNT_UNKNOWN ) {
duke@435 1784 if( top == Op_IfTrue ) {
duke@435 1785 if( p < (PROB_FAIR + PROB_UNLIKELY_MAG(3))) {
duke@435 1786 iff->_prob = PROB_STATIC_FREQUENT;
duke@435 1787 }
duke@435 1788 } else {
duke@435 1789 if( p > (PROB_FAIR - PROB_UNLIKELY_MAG(3))) {
duke@435 1790 iff->_prob = PROB_STATIC_INFREQUENT;
duke@435 1791 }
duke@435 1792 }
duke@435 1793 }
duke@435 1794 }
duke@435 1795 }
duke@435 1796 test = phase->idom(test);
duke@435 1797 }
duke@435 1798 }
duke@435 1799
duke@435 1800
duke@435 1801 //------------------------------policy_do_remove_empty_loop--------------------
duke@435 1802 // Micro-benchmark spamming. Policy is to always remove empty loops.
duke@435 1803 // The 'DO' part is to replace the trip counter with the value it will
duke@435 1804 // have on the last iteration. This will break the loop.
duke@435 1805 bool IdealLoopTree::policy_do_remove_empty_loop( PhaseIdealLoop *phase ) {
duke@435 1806 // Minimum size must be empty loop
kvn@2735 1807 if (_body.size() > EMPTY_LOOP_SIZE)
kvn@2665 1808 return false;
duke@435 1809
kvn@2665 1810 if (!_head->is_CountedLoop())
kvn@2665 1811 return false; // Dead loop
duke@435 1812 CountedLoopNode *cl = _head->as_CountedLoop();
kvn@2665 1813 if (!cl->loopexit())
kvn@2665 1814 return false; // Malformed loop
kvn@2665 1815 if (!phase->is_member(this, phase->get_ctrl(cl->loopexit()->in(CountedLoopEndNode::TestValue))))
duke@435 1816 return false; // Infinite loop
never@2685 1817
duke@435 1818 #ifdef ASSERT
duke@435 1819 // Ensure only one phi which is the iv.
duke@435 1820 Node* iv = NULL;
duke@435 1821 for (DUIterator_Fast imax, i = cl->fast_outs(imax); i < imax; i++) {
duke@435 1822 Node* n = cl->fast_out(i);
duke@435 1823 if (n->Opcode() == Op_Phi) {
duke@435 1824 assert(iv == NULL, "Too many phis" );
duke@435 1825 iv = n;
duke@435 1826 }
duke@435 1827 }
duke@435 1828 assert(iv == cl->phi(), "Wrong phi" );
duke@435 1829 #endif
never@2685 1830
never@2685 1831 // main and post loops have explicitly created zero trip guard
never@2685 1832 bool needs_guard = !cl->is_main_loop() && !cl->is_post_loop();
never@2685 1833 if (needs_guard) {
kvn@2747 1834 // Skip guard if values not overlap.
kvn@2747 1835 const TypeInt* init_t = phase->_igvn.type(cl->init_trip())->is_int();
kvn@2747 1836 const TypeInt* limit_t = phase->_igvn.type(cl->limit())->is_int();
kvn@2747 1837 int stride_con = cl->stride_con();
kvn@2747 1838 if (stride_con > 0) {
kvn@2747 1839 needs_guard = (init_t->_hi >= limit_t->_lo);
kvn@2747 1840 } else {
kvn@2747 1841 needs_guard = (init_t->_lo <= limit_t->_hi);
kvn@2747 1842 }
kvn@2747 1843 }
kvn@2747 1844 if (needs_guard) {
never@2685 1845 // Check for an obvious zero trip guard.
kvn@2727 1846 Node* inctrl = PhaseIdealLoop::skip_loop_predicates(cl->in(LoopNode::EntryControl));
never@2685 1847 if (inctrl->Opcode() == Op_IfTrue) {
never@2685 1848 // The test should look like just the backedge of a CountedLoop
never@2685 1849 Node* iff = inctrl->in(0);
never@2685 1850 if (iff->is_If()) {
never@2685 1851 Node* bol = iff->in(1);
never@2685 1852 if (bol->is_Bool() && bol->as_Bool()->_test._test == cl->loopexit()->test_trip()) {
never@2685 1853 Node* cmp = bol->in(1);
never@2685 1854 if (cmp->is_Cmp() && cmp->in(1) == cl->init_trip() && cmp->in(2) == cl->limit()) {
never@2685 1855 needs_guard = false;
never@2685 1856 }
never@2685 1857 }
never@2685 1858 }
never@2685 1859 }
never@2685 1860 }
never@2685 1861
never@2685 1862 #ifndef PRODUCT
never@2685 1863 if (PrintOpto) {
never@2685 1864 tty->print("Removing empty loop with%s zero trip guard", needs_guard ? "out" : "");
never@2685 1865 this->dump_head();
never@2685 1866 } else if (TraceLoopOpts) {
never@2685 1867 tty->print("Empty with%s zero trip guard ", needs_guard ? "out" : "");
never@2685 1868 this->dump_head();
never@2685 1869 }
never@2685 1870 #endif
never@2685 1871
never@2685 1872 if (needs_guard) {
never@2685 1873 // Peel the loop to ensure there's a zero trip guard
never@2685 1874 Node_List old_new;
never@2685 1875 phase->do_peeling(this, old_new);
never@2685 1876 }
never@2685 1877
duke@435 1878 // Replace the phi at loop head with the final value of the last
duke@435 1879 // iteration. Then the CountedLoopEnd will collapse (backedge never
duke@435 1880 // taken) and all loop-invariant uses of the exit values will be correct.
duke@435 1881 Node *phi = cl->phi();
duke@435 1882 Node *final = new (phase->C, 3) SubINode( cl->limit(), cl->stride() );
duke@435 1883 phase->register_new_node(final,cl->in(LoopNode::EntryControl));
kvn@1976 1884 phase->_igvn.replace_node(phi,final);
duke@435 1885 phase->C->set_major_progress();
duke@435 1886 return true;
duke@435 1887 }
duke@435 1888
kvn@2747 1889 //------------------------------policy_do_one_iteration_loop-------------------
kvn@2747 1890 // Convert one iteration loop into normal code.
kvn@2747 1891 bool IdealLoopTree::policy_do_one_iteration_loop( PhaseIdealLoop *phase ) {
kvn@2747 1892 if (!_head->as_Loop()->is_valid_counted_loop())
kvn@2747 1893 return false; // Only for counted loop
kvn@2747 1894
kvn@2747 1895 CountedLoopNode *cl = _head->as_CountedLoop();
kvn@2747 1896 if (!cl->has_exact_trip_count() || cl->trip_count() != 1) {
kvn@2747 1897 return false;
kvn@2747 1898 }
kvn@2747 1899
kvn@2747 1900 #ifndef PRODUCT
kvn@2747 1901 if(TraceLoopOpts) {
kvn@2747 1902 tty->print("OneIteration ");
kvn@2747 1903 this->dump_head();
kvn@2747 1904 }
kvn@2747 1905 #endif
kvn@2747 1906
kvn@2747 1907 Node *init_n = cl->init_trip();
kvn@2747 1908 #ifdef ASSERT
kvn@2747 1909 // Loop boundaries should be constant since trip count is exact.
kvn@2747 1910 assert(init_n->get_int() + cl->stride_con() >= cl->limit()->get_int(), "should be one iteration");
kvn@2747 1911 #endif
kvn@2747 1912 // Replace the phi at loop head with the value of the init_trip.
kvn@2747 1913 // Then the CountedLoopEnd will collapse (backedge will not be taken)
kvn@2747 1914 // and all loop-invariant uses of the exit values will be correct.
kvn@2747 1915 phase->_igvn.replace_node(cl->phi(), cl->init_trip());
kvn@2747 1916 phase->C->set_major_progress();
kvn@2747 1917 return true;
kvn@2747 1918 }
duke@435 1919
duke@435 1920 //=============================================================================
duke@435 1921 //------------------------------iteration_split_impl---------------------------
never@836 1922 bool IdealLoopTree::iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new ) {
kvn@2747 1923 // Compute exact loop trip count if possible.
kvn@2747 1924 compute_exact_trip_count(phase);
kvn@2747 1925
kvn@2747 1926 // Convert one iteration loop into normal code.
kvn@2747 1927 if (policy_do_one_iteration_loop(phase))
kvn@2747 1928 return true;
kvn@2747 1929
duke@435 1930 // Check and remove empty loops (spam micro-benchmarks)
kvn@2747 1931 if (policy_do_remove_empty_loop(phase))
cfang@1607 1932 return true; // Here we removed an empty loop
duke@435 1933
duke@435 1934 bool should_peel = policy_peeling(phase); // Should we peel?
duke@435 1935
duke@435 1936 bool should_unswitch = policy_unswitching(phase);
duke@435 1937
duke@435 1938 // Non-counted loops may be peeled; exactly 1 iteration is peeled.
duke@435 1939 // This removes loop-invariant tests (usually null checks).
kvn@2747 1940 if (!_head->is_CountedLoop()) { // Non-counted loop
duke@435 1941 if (PartialPeelLoop && phase->partial_peel(this, old_new)) {
never@836 1942 // Partial peel succeeded so terminate this round of loop opts
never@836 1943 return false;
duke@435 1944 }
kvn@2747 1945 if (should_peel) { // Should we peel?
duke@435 1946 #ifndef PRODUCT
duke@435 1947 if (PrintOpto) tty->print_cr("should_peel");
duke@435 1948 #endif
duke@435 1949 phase->do_peeling(this,old_new);
kvn@2747 1950 } else if (should_unswitch) {
duke@435 1951 phase->do_unswitching(this, old_new);
duke@435 1952 }
never@836 1953 return true;
duke@435 1954 }
duke@435 1955 CountedLoopNode *cl = _head->as_CountedLoop();
duke@435 1956
kvn@2747 1957 if (!cl->loopexit()) return true; // Ignore various kinds of broken loops
duke@435 1958
duke@435 1959 // Do nothing special to pre- and post- loops
kvn@2747 1960 if (cl->is_pre_loop() || cl->is_post_loop()) return true;
duke@435 1961
duke@435 1962 // Compute loop trip count from profile data
duke@435 1963 compute_profile_trip_cnt(phase);
duke@435 1964
duke@435 1965 // Before attempting fancy unrolling, RCE or alignment, see if we want
duke@435 1966 // to completely unroll this loop or do loop unswitching.
kvn@2747 1967 if (cl->is_normal_loop()) {
cfang@1224 1968 if (should_unswitch) {
cfang@1224 1969 phase->do_unswitching(this, old_new);
cfang@1224 1970 return true;
cfang@1224 1971 }
duke@435 1972 bool should_maximally_unroll = policy_maximally_unroll(phase);
kvn@2747 1973 if (should_maximally_unroll) {
duke@435 1974 // Here we did some unrolling and peeling. Eventually we will
duke@435 1975 // completely unroll this loop and it will no longer be a loop.
duke@435 1976 phase->do_maximally_unroll(this,old_new);
never@836 1977 return true;
duke@435 1978 }
duke@435 1979 }
duke@435 1980
kvn@2735 1981 // Skip next optimizations if running low on nodes. Note that
kvn@2735 1982 // policy_unswitching and policy_maximally_unroll have this check.
kvn@2735 1983 uint nodes_left = MaxNodeLimit - phase->C->unique();
kvn@2735 1984 if ((2 * _body.size()) > nodes_left) {
kvn@2735 1985 return true;
kvn@2735 1986 }
duke@435 1987
duke@435 1988 // Counted loops may be peeled, may need some iterations run up
duke@435 1989 // front for RCE, and may want to align loop refs to a cache
duke@435 1990 // line. Thus we clone a full loop up front whose trip count is
duke@435 1991 // at least 1 (if peeling), but may be several more.
duke@435 1992
duke@435 1993 // The main loop will start cache-line aligned with at least 1
duke@435 1994 // iteration of the unrolled body (zero-trip test required) and
duke@435 1995 // will have some range checks removed.
duke@435 1996
duke@435 1997 // A post-loop will finish any odd iterations (leftover after
duke@435 1998 // unrolling), plus any needed for RCE purposes.
duke@435 1999
duke@435 2000 bool should_unroll = policy_unroll(phase);
duke@435 2001
duke@435 2002 bool should_rce = policy_range_check(phase);
duke@435 2003
duke@435 2004 bool should_align = policy_align(phase);
duke@435 2005
duke@435 2006 // If not RCE'ing (iteration splitting) or Aligning, then we do not
duke@435 2007 // need a pre-loop. We may still need to peel an initial iteration but
duke@435 2008 // we will not be needing an unknown number of pre-iterations.
duke@435 2009 //
duke@435 2010 // Basically, if may_rce_align reports FALSE first time through,
duke@435 2011 // we will not be able to later do RCE or Aligning on this loop.
duke@435 2012 bool may_rce_align = !policy_peel_only(phase) || should_rce || should_align;
duke@435 2013
duke@435 2014 // If we have any of these conditions (RCE, alignment, unrolling) met, then
duke@435 2015 // we switch to the pre-/main-/post-loop model. This model also covers
duke@435 2016 // peeling.
kvn@2747 2017 if (should_rce || should_align || should_unroll) {
kvn@2747 2018 if (cl->is_normal_loop()) // Convert to 'pre/main/post' loops
duke@435 2019 phase->insert_pre_post_loops(this,old_new, !may_rce_align);
duke@435 2020
duke@435 2021 // Adjust the pre- and main-loop limits to let the pre and post loops run
duke@435 2022 // with full checks, but the main-loop with no checks. Remove said
duke@435 2023 // checks from the main body.
kvn@2747 2024 if (should_rce)
duke@435 2025 phase->do_range_check(this,old_new);
duke@435 2026
duke@435 2027 // Double loop body for unrolling. Adjust the minimum-trip test (will do
duke@435 2028 // twice as many iterations as before) and the main body limit (only do
duke@435 2029 // an even number of trips). If we are peeling, we might enable some RCE
duke@435 2030 // and we'd rather unroll the post-RCE'd loop SO... do not unroll if
duke@435 2031 // peeling.
kvn@2747 2032 if (should_unroll && !should_peel)
kvn@2747 2033 phase->do_unroll(this,old_new, true);
duke@435 2034
duke@435 2035 // Adjust the pre-loop limits to align the main body
duke@435 2036 // iterations.
kvn@2747 2037 if (should_align)
duke@435 2038 Unimplemented();
duke@435 2039
duke@435 2040 } else { // Else we have an unchanged counted loop
kvn@2747 2041 if (should_peel) // Might want to peel but do nothing else
duke@435 2042 phase->do_peeling(this,old_new);
duke@435 2043 }
never@836 2044 return true;
duke@435 2045 }
duke@435 2046
duke@435 2047
duke@435 2048 //=============================================================================
duke@435 2049 //------------------------------iteration_split--------------------------------
never@836 2050 bool IdealLoopTree::iteration_split( PhaseIdealLoop *phase, Node_List &old_new ) {
duke@435 2051 // Recursively iteration split nested loops
kvn@2665 2052 if (_child && !_child->iteration_split(phase, old_new))
never@836 2053 return false;
duke@435 2054
duke@435 2055 // Clean out prior deadwood
duke@435 2056 DCE_loop_body();
duke@435 2057
duke@435 2058
duke@435 2059 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage.
duke@435 2060 // Replace with a 1-in-10 exit guess.
kvn@2665 2061 if (_parent /*not the root loop*/ &&
duke@435 2062 !_irreducible &&
duke@435 2063 // Also ignore the occasional dead backedge
kvn@2665 2064 !tail()->is_top()) {
duke@435 2065 adjust_loop_exit_prob(phase);
duke@435 2066 }
duke@435 2067
duke@435 2068 // Gate unrolling, RCE and peeling efforts.
kvn@2665 2069 if (!_child && // If not an inner loop, do not split
duke@435 2070 !_irreducible &&
kvn@474 2071 _allow_optimizations &&
kvn@2665 2072 !tail()->is_top()) { // Also ignore the occasional dead backedge
duke@435 2073 if (!_has_call) {
kvn@2665 2074 if (!iteration_split_impl(phase, old_new)) {
cfang@1607 2075 return false;
cfang@1607 2076 }
duke@435 2077 } else if (policy_unswitching(phase)) {
duke@435 2078 phase->do_unswitching(this, old_new);
duke@435 2079 }
duke@435 2080 }
duke@435 2081
duke@435 2082 // Minor offset re-organization to remove loop-fallout uses of
kvn@2665 2083 // trip counter when there was no major reshaping.
kvn@2665 2084 phase->reorg_offsets(this);
kvn@2665 2085
kvn@2665 2086 if (_next && !_next->iteration_split(phase, old_new))
never@836 2087 return false;
never@836 2088 return true;
duke@435 2089 }
cfang@1607 2090
cfang@1607 2091
kvn@2727 2092 //=============================================================================
never@2118 2093 // Process all the loops in the loop tree and replace any fill
never@2118 2094 // patterns with an intrisc version.
never@2118 2095 bool PhaseIdealLoop::do_intrinsify_fill() {
never@2118 2096 bool changed = false;
never@2118 2097 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
never@2118 2098 IdealLoopTree* lpt = iter.current();
never@2118 2099 changed |= intrinsify_fill(lpt);
never@2118 2100 }
never@2118 2101 return changed;
never@2118 2102 }
never@2118 2103
never@2118 2104
never@2118 2105 // Examine an inner loop looking for a a single store of an invariant
never@2118 2106 // value in a unit stride loop,
never@2118 2107 bool PhaseIdealLoop::match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value,
never@2118 2108 Node*& shift, Node*& con) {
never@2118 2109 const char* msg = NULL;
never@2118 2110 Node* msg_node = NULL;
never@2118 2111
never@2118 2112 store_value = NULL;
never@2118 2113 con = NULL;
never@2118 2114 shift = NULL;
never@2118 2115
never@2118 2116 // Process the loop looking for stores. If there are multiple
never@2118 2117 // stores or extra control flow give at this point.
never@2118 2118 CountedLoopNode* head = lpt->_head->as_CountedLoop();
never@2118 2119 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) {
never@2118 2120 Node* n = lpt->_body.at(i);
never@2118 2121 if (n->outcnt() == 0) continue; // Ignore dead
never@2118 2122 if (n->is_Store()) {
never@2118 2123 if (store != NULL) {
never@2118 2124 msg = "multiple stores";
never@2118 2125 break;
never@2118 2126 }
never@2118 2127 int opc = n->Opcode();
never@2118 2128 if (opc == Op_StoreP || opc == Op_StoreN || opc == Op_StoreCM) {
never@2118 2129 msg = "oop fills not handled";
never@2118 2130 break;
never@2118 2131 }
never@2118 2132 Node* value = n->in(MemNode::ValueIn);
never@2118 2133 if (!lpt->is_invariant(value)) {
never@2118 2134 msg = "variant store value";
never@2140 2135 } else if (!_igvn.type(n->in(MemNode::Address))->isa_aryptr()) {
never@2140 2136 msg = "not array address";
never@2118 2137 }
never@2118 2138 store = n;
never@2118 2139 store_value = value;
never@2118 2140 } else if (n->is_If() && n != head->loopexit()) {
never@2118 2141 msg = "extra control flow";
never@2118 2142 msg_node = n;
never@2118 2143 }
never@2118 2144 }
never@2118 2145
never@2118 2146 if (store == NULL) {
never@2118 2147 // No store in loop
never@2118 2148 return false;
never@2118 2149 }
never@2118 2150
never@2118 2151 if (msg == NULL && head->stride_con() != 1) {
never@2118 2152 // could handle negative strides too
never@2118 2153 if (head->stride_con() < 0) {
never@2118 2154 msg = "negative stride";
never@2118 2155 } else {
never@2118 2156 msg = "non-unit stride";
never@2118 2157 }
never@2118 2158 }
never@2118 2159
never@2118 2160 if (msg == NULL && !store->in(MemNode::Address)->is_AddP()) {
never@2118 2161 msg = "can't handle store address";
never@2118 2162 msg_node = store->in(MemNode::Address);
never@2118 2163 }
never@2118 2164
never@2168 2165 if (msg == NULL &&
never@2168 2166 (!store->in(MemNode::Memory)->is_Phi() ||
never@2168 2167 store->in(MemNode::Memory)->in(LoopNode::LoopBackControl) != store)) {
never@2168 2168 msg = "store memory isn't proper phi";
never@2168 2169 msg_node = store->in(MemNode::Memory);
never@2168 2170 }
never@2168 2171
never@2118 2172 // Make sure there is an appropriate fill routine
never@2118 2173 BasicType t = store->as_Mem()->memory_type();
never@2118 2174 const char* fill_name;
never@2118 2175 if (msg == NULL &&
never@2118 2176 StubRoutines::select_fill_function(t, false, fill_name) == NULL) {
never@2118 2177 msg = "unsupported store";
never@2118 2178 msg_node = store;
never@2118 2179 }
never@2118 2180
never@2118 2181 if (msg != NULL) {
never@2118 2182 #ifndef PRODUCT
never@2118 2183 if (TraceOptimizeFill) {
never@2118 2184 tty->print_cr("not fill intrinsic candidate: %s", msg);
never@2118 2185 if (msg_node != NULL) msg_node->dump();
never@2118 2186 }
never@2118 2187 #endif
never@2118 2188 return false;
never@2118 2189 }
never@2118 2190
never@2118 2191 // Make sure the address expression can be handled. It should be
never@2118 2192 // head->phi * elsize + con. head->phi might have a ConvI2L.
never@2118 2193 Node* elements[4];
never@2118 2194 Node* conv = NULL;
never@2140 2195 bool found_index = false;
never@2118 2196 int count = store->in(MemNode::Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements));
never@2118 2197 for (int e = 0; e < count; e++) {
never@2118 2198 Node* n = elements[e];
never@2118 2199 if (n->is_Con() && con == NULL) {
never@2118 2200 con = n;
never@2118 2201 } else if (n->Opcode() == Op_LShiftX && shift == NULL) {
never@2118 2202 Node* value = n->in(1);
never@2118 2203 #ifdef _LP64
never@2118 2204 if (value->Opcode() == Op_ConvI2L) {
never@2118 2205 conv = value;
never@2118 2206 value = value->in(1);
never@2118 2207 }
never@2118 2208 #endif
never@2118 2209 if (value != head->phi()) {
never@2118 2210 msg = "unhandled shift in address";
never@2118 2211 } else {
never@2730 2212 if (type2aelembytes(store->as_Mem()->memory_type(), true) != (1 << n->in(2)->get_int())) {
never@2730 2213 msg = "scale doesn't match";
never@2730 2214 } else {
never@2730 2215 found_index = true;
never@2730 2216 shift = n;
never@2730 2217 }
never@2118 2218 }
never@2118 2219 } else if (n->Opcode() == Op_ConvI2L && conv == NULL) {
never@2118 2220 if (n->in(1) == head->phi()) {
never@2140 2221 found_index = true;
never@2118 2222 conv = n;
never@2118 2223 } else {
never@2118 2224 msg = "unhandled input to ConvI2L";
never@2118 2225 }
never@2118 2226 } else if (n == head->phi()) {
never@2118 2227 // no shift, check below for allowed cases
never@2140 2228 found_index = true;
never@2118 2229 } else {
never@2118 2230 msg = "unhandled node in address";
never@2118 2231 msg_node = n;
never@2118 2232 }
never@2118 2233 }
never@2118 2234
never@2118 2235 if (count == -1) {
never@2118 2236 msg = "malformed address expression";
never@2118 2237 msg_node = store;
never@2118 2238 }
never@2118 2239
never@2140 2240 if (!found_index) {
never@2140 2241 msg = "missing use of index";
never@2140 2242 }
never@2140 2243
never@2118 2244 // byte sized items won't have a shift
never@2118 2245 if (msg == NULL && shift == NULL && t != T_BYTE && t != T_BOOLEAN) {
never@2118 2246 msg = "can't find shift";
never@2118 2247 msg_node = store;
never@2118 2248 }
never@2118 2249
never@2118 2250 if (msg != NULL) {
never@2118 2251 #ifndef PRODUCT
never@2118 2252 if (TraceOptimizeFill) {
never@2118 2253 tty->print_cr("not fill intrinsic: %s", msg);
never@2118 2254 if (msg_node != NULL) msg_node->dump();
never@2118 2255 }
never@2118 2256 #endif
never@2118 2257 return false;
never@2118 2258 }
never@2118 2259
never@2118 2260 // No make sure all the other nodes in the loop can be handled
never@2118 2261 VectorSet ok(Thread::current()->resource_area());
never@2118 2262
never@2118 2263 // store related values are ok
never@2118 2264 ok.set(store->_idx);
never@2118 2265 ok.set(store->in(MemNode::Memory)->_idx);
never@2118 2266
never@2118 2267 // Loop structure is ok
never@2118 2268 ok.set(head->_idx);
never@2118 2269 ok.set(head->loopexit()->_idx);
never@2118 2270 ok.set(head->phi()->_idx);
never@2118 2271 ok.set(head->incr()->_idx);
never@2118 2272 ok.set(head->loopexit()->cmp_node()->_idx);
never@2118 2273 ok.set(head->loopexit()->in(1)->_idx);
never@2118 2274
never@2118 2275 // Address elements are ok
never@2118 2276 if (con) ok.set(con->_idx);
never@2118 2277 if (shift) ok.set(shift->_idx);
never@2118 2278 if (conv) ok.set(conv->_idx);
never@2118 2279
never@2118 2280 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) {
never@2118 2281 Node* n = lpt->_body.at(i);
never@2118 2282 if (n->outcnt() == 0) continue; // Ignore dead
never@2118 2283 if (ok.test(n->_idx)) continue;
never@2118 2284 // Backedge projection is ok
never@2118 2285 if (n->is_IfTrue() && n->in(0) == head->loopexit()) continue;
never@2118 2286 if (!n->is_AddP()) {
never@2118 2287 msg = "unhandled node";
never@2118 2288 msg_node = n;
never@2118 2289 break;
never@2118 2290 }
never@2118 2291 }
never@2118 2292
never@2118 2293 // Make sure no unexpected values are used outside the loop
never@2118 2294 for (uint i = 0; msg == NULL && i < lpt->_body.size(); i++) {
never@2118 2295 Node* n = lpt->_body.at(i);
never@2118 2296 // These values can be replaced with other nodes if they are used
never@2118 2297 // outside the loop.
never@2168 2298 if (n == store || n == head->loopexit() || n == head->incr() || n == store->in(MemNode::Memory)) continue;
never@2118 2299 for (SimpleDUIterator iter(n); iter.has_next(); iter.next()) {
never@2118 2300 Node* use = iter.get();
never@2118 2301 if (!lpt->_body.contains(use)) {
never@2118 2302 msg = "node is used outside loop";
never@2118 2303 // lpt->_body.dump();
never@2118 2304 msg_node = n;
never@2118 2305 break;
never@2118 2306 }
never@2118 2307 }
never@2118 2308 }
never@2118 2309
never@2118 2310 #ifdef ASSERT
never@2118 2311 if (TraceOptimizeFill) {
never@2118 2312 if (msg != NULL) {
never@2118 2313 tty->print_cr("no fill intrinsic: %s", msg);
never@2118 2314 if (msg_node != NULL) msg_node->dump();
never@2118 2315 } else {
never@2118 2316 tty->print_cr("fill intrinsic for:");
never@2118 2317 }
never@2118 2318 store->dump();
never@2118 2319 if (Verbose) {
never@2118 2320 lpt->_body.dump();
never@2118 2321 }
never@2118 2322 }
never@2118 2323 #endif
never@2118 2324
never@2118 2325 return msg == NULL;
never@2118 2326 }
never@2118 2327
never@2118 2328
never@2118 2329
never@2118 2330 bool PhaseIdealLoop::intrinsify_fill(IdealLoopTree* lpt) {
never@2118 2331 // Only for counted inner loops
never@2118 2332 if (!lpt->is_counted() || !lpt->is_inner()) {
never@2118 2333 return false;
never@2118 2334 }
never@2118 2335
never@2118 2336 // Must have constant stride
never@2118 2337 CountedLoopNode* head = lpt->_head->as_CountedLoop();
never@2118 2338 if (!head->stride_is_con() || !head->is_normal_loop()) {
never@2118 2339 return false;
never@2118 2340 }
never@2118 2341
never@2118 2342 // Check that the body only contains a store of a loop invariant
never@2118 2343 // value that is indexed by the loop phi.
never@2118 2344 Node* store = NULL;
never@2118 2345 Node* store_value = NULL;
never@2118 2346 Node* shift = NULL;
never@2118 2347 Node* offset = NULL;
never@2118 2348 if (!match_fill_loop(lpt, store, store_value, shift, offset)) {
never@2118 2349 return false;
never@2118 2350 }
never@2118 2351
kvn@2727 2352 #ifndef PRODUCT
kvn@2727 2353 if (TraceLoopOpts) {
kvn@2727 2354 tty->print("ArrayFill ");
kvn@2727 2355 lpt->dump_head();
kvn@2727 2356 }
kvn@2727 2357 #endif
kvn@2727 2358
never@2118 2359 // Now replace the whole loop body by a call to a fill routine that
never@2118 2360 // covers the same region as the loop.
never@2118 2361 Node* base = store->in(MemNode::Address)->as_AddP()->in(AddPNode::Base);
never@2118 2362
never@2118 2363 // Build an expression for the beginning of the copy region
never@2118 2364 Node* index = head->init_trip();
never@2118 2365 #ifdef _LP64
never@2118 2366 index = new (C, 2) ConvI2LNode(index);
never@2118 2367 _igvn.register_new_node_with_optimizer(index);
never@2118 2368 #endif
never@2118 2369 if (shift != NULL) {
never@2118 2370 // byte arrays don't require a shift but others do.
never@2118 2371 index = new (C, 3) LShiftXNode(index, shift->in(2));
never@2118 2372 _igvn.register_new_node_with_optimizer(index);
never@2118 2373 }
never@2118 2374 index = new (C, 4) AddPNode(base, base, index);
never@2118 2375 _igvn.register_new_node_with_optimizer(index);
never@2118 2376 Node* from = new (C, 4) AddPNode(base, index, offset);
never@2118 2377 _igvn.register_new_node_with_optimizer(from);
never@2118 2378 // Compute the number of elements to copy
never@2118 2379 Node* len = new (C, 3) SubINode(head->limit(), head->init_trip());
never@2118 2380 _igvn.register_new_node_with_optimizer(len);
never@2118 2381
never@2118 2382 BasicType t = store->as_Mem()->memory_type();
never@2118 2383 bool aligned = false;
never@2118 2384 if (offset != NULL && head->init_trip()->is_Con()) {
never@2118 2385 int element_size = type2aelembytes(t);
never@2118 2386 aligned = (offset->find_intptr_t_type()->get_con() + head->init_trip()->get_int() * element_size) % HeapWordSize == 0;
never@2118 2387 }
never@2118 2388
never@2118 2389 // Build a call to the fill routine
never@2118 2390 const char* fill_name;
never@2118 2391 address fill = StubRoutines::select_fill_function(t, aligned, fill_name);
never@2118 2392 assert(fill != NULL, "what?");
never@2118 2393
never@2118 2394 // Convert float/double to int/long for fill routines
never@2118 2395 if (t == T_FLOAT) {
never@2118 2396 store_value = new (C, 2) MoveF2INode(store_value);
never@2118 2397 _igvn.register_new_node_with_optimizer(store_value);
never@2118 2398 } else if (t == T_DOUBLE) {
never@2118 2399 store_value = new (C, 2) MoveD2LNode(store_value);
never@2118 2400 _igvn.register_new_node_with_optimizer(store_value);
never@2118 2401 }
never@2118 2402
never@2118 2403 Node* mem_phi = store->in(MemNode::Memory);
never@2118 2404 Node* result_ctrl;
never@2118 2405 Node* result_mem;
never@2118 2406 const TypeFunc* call_type = OptoRuntime::array_fill_Type();
never@2118 2407 int size = call_type->domain()->cnt();
never@2118 2408 CallLeafNode *call = new (C, size) CallLeafNoFPNode(call_type, fill,
never@2118 2409 fill_name, TypeAryPtr::get_array_body_type(t));
never@2118 2410 call->init_req(TypeFunc::Parms+0, from);
never@2118 2411 call->init_req(TypeFunc::Parms+1, store_value);
never@2199 2412 #ifdef _LP64
never@2199 2413 len = new (C, 2) ConvI2LNode(len);
never@2199 2414 _igvn.register_new_node_with_optimizer(len);
never@2199 2415 #endif
never@2118 2416 call->init_req(TypeFunc::Parms+2, len);
never@2199 2417 #ifdef _LP64
never@2199 2418 call->init_req(TypeFunc::Parms+3, C->top());
never@2199 2419 #endif
never@2118 2420 call->init_req( TypeFunc::Control, head->init_control());
never@2118 2421 call->init_req( TypeFunc::I_O , C->top() ) ; // does no i/o
never@2118 2422 call->init_req( TypeFunc::Memory , mem_phi->in(LoopNode::EntryControl) );
never@2118 2423 call->init_req( TypeFunc::ReturnAdr, C->start()->proj_out(TypeFunc::ReturnAdr) );
never@2118 2424 call->init_req( TypeFunc::FramePtr, C->start()->proj_out(TypeFunc::FramePtr) );
never@2118 2425 _igvn.register_new_node_with_optimizer(call);
never@2118 2426 result_ctrl = new (C, 1) ProjNode(call,TypeFunc::Control);
never@2118 2427 _igvn.register_new_node_with_optimizer(result_ctrl);
never@2118 2428 result_mem = new (C, 1) ProjNode(call,TypeFunc::Memory);
never@2118 2429 _igvn.register_new_node_with_optimizer(result_mem);
never@2118 2430
never@2118 2431 // If this fill is tightly coupled to an allocation and overwrites
never@2118 2432 // the whole body, allow it to take over the zeroing.
never@2118 2433 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, this);
never@2118 2434 if (alloc != NULL && alloc->is_AllocateArray()) {
never@2118 2435 Node* length = alloc->as_AllocateArray()->Ideal_length();
never@2118 2436 if (head->limit() == length &&
never@2118 2437 head->init_trip() == _igvn.intcon(0)) {
never@2118 2438 if (TraceOptimizeFill) {
never@2118 2439 tty->print_cr("Eliminated zeroing in allocation");
never@2118 2440 }
never@2118 2441 alloc->maybe_set_complete(&_igvn);
never@2118 2442 } else {
never@2118 2443 #ifdef ASSERT
never@2118 2444 if (TraceOptimizeFill) {
never@2118 2445 tty->print_cr("filling array but bounds don't match");
never@2118 2446 alloc->dump();
never@2118 2447 head->init_trip()->dump();
never@2118 2448 head->limit()->dump();
never@2118 2449 length->dump();
never@2118 2450 }
never@2118 2451 #endif
never@2118 2452 }
never@2118 2453 }
never@2118 2454
never@2118 2455 // Redirect the old control and memory edges that are outside the loop.
never@2118 2456 Node* exit = head->loopexit()->proj_out(0);
never@2168 2457 // Sometimes the memory phi of the head is used as the outgoing
never@2168 2458 // state of the loop. It's safe in this case to replace it with the
never@2168 2459 // result_mem.
never@2168 2460 _igvn.replace_node(store->in(MemNode::Memory), result_mem);
never@2118 2461 _igvn.replace_node(exit, result_ctrl);
never@2118 2462 _igvn.replace_node(store, result_mem);
never@2118 2463 // Any uses the increment outside of the loop become the loop limit.
never@2118 2464 _igvn.replace_node(head->incr(), head->limit());
never@2118 2465
never@2118 2466 // Disconnect the head from the loop.
never@2118 2467 for (uint i = 0; i < lpt->_body.size(); i++) {
never@2118 2468 Node* n = lpt->_body.at(i);
never@2118 2469 _igvn.replace_node(n, C->top());
never@2118 2470 }
never@2118 2471
never@2118 2472 return true;
never@2118 2473 }

mercurial