0
|
1 /*
|
|
2 * Copyright 2000-2007 Sun Microsystems, Inc. All Rights Reserved.
|
|
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
|
|
4 *
|
|
5 * This code is free software; you can redistribute it and/or modify it
|
|
6 * under the terms of the GNU General Public License version 2 only, as
|
|
7 * published by the Free Software Foundation.
|
|
8 *
|
|
9 * This code is distributed in the hope that it will be useful, but WITHOUT
|
|
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
|
|
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
|
|
12 * version 2 for more details (a copy is included in the LICENSE file that
|
|
13 * accompanied this code).
|
|
14 *
|
|
15 * You should have received a copy of the GNU General Public License version
|
|
16 * 2 along with this work; if not, write to the Free Software Foundation,
|
|
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
|
|
18 *
|
|
19 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
|
|
20 * CA 95054 USA or visit www.sun.com if you need additional information or
|
|
21 * have any questions.
|
|
22 *
|
|
23 */
|
|
24
|
|
25 #include "incls/_precompiled.incl"
|
|
26 #include "incls/_loopTransform.cpp.incl"
|
|
27
|
|
28 //------------------------------is_loop_exit-----------------------------------
|
|
29 // Given an IfNode, return the loop-exiting projection or NULL if both
|
|
30 // arms remain in the loop.
|
|
31 Node *IdealLoopTree::is_loop_exit(Node *iff) const {
|
|
32 if( iff->outcnt() != 2 ) return NULL; // Ignore partially dead tests
|
|
33 PhaseIdealLoop *phase = _phase;
|
|
34 // Test is an IfNode, has 2 projections. If BOTH are in the loop
|
|
35 // we need loop unswitching instead of peeling.
|
|
36 if( !is_member(phase->get_loop( iff->raw_out(0) )) )
|
|
37 return iff->raw_out(0);
|
|
38 if( !is_member(phase->get_loop( iff->raw_out(1) )) )
|
|
39 return iff->raw_out(1);
|
|
40 return NULL;
|
|
41 }
|
|
42
|
|
43
|
|
44 //=============================================================================
|
|
45
|
|
46
|
|
47 //------------------------------record_for_igvn----------------------------
|
|
48 // Put loop body on igvn work list
|
|
49 void IdealLoopTree::record_for_igvn() {
|
|
50 for( uint i = 0; i < _body.size(); i++ ) {
|
|
51 Node *n = _body.at(i);
|
|
52 _phase->_igvn._worklist.push(n);
|
|
53 }
|
|
54 }
|
|
55
|
|
56 //------------------------------compute_profile_trip_cnt----------------------------
|
|
57 // Compute loop trip count from profile data as
|
|
58 // (backedge_count + loop_exit_count) / loop_exit_count
|
|
59 void IdealLoopTree::compute_profile_trip_cnt( PhaseIdealLoop *phase ) {
|
|
60 if (!_head->is_CountedLoop()) {
|
|
61 return;
|
|
62 }
|
|
63 CountedLoopNode* head = _head->as_CountedLoop();
|
|
64 if (head->profile_trip_cnt() != COUNT_UNKNOWN) {
|
|
65 return; // Already computed
|
|
66 }
|
|
67 float trip_cnt = (float)max_jint; // default is big
|
|
68
|
|
69 Node* back = head->in(LoopNode::LoopBackControl);
|
|
70 while (back != head) {
|
|
71 if ((back->Opcode() == Op_IfTrue || back->Opcode() == Op_IfFalse) &&
|
|
72 back->in(0) &&
|
|
73 back->in(0)->is_If() &&
|
|
74 back->in(0)->as_If()->_fcnt != COUNT_UNKNOWN &&
|
|
75 back->in(0)->as_If()->_prob != PROB_UNKNOWN) {
|
|
76 break;
|
|
77 }
|
|
78 back = phase->idom(back);
|
|
79 }
|
|
80 if (back != head) {
|
|
81 assert((back->Opcode() == Op_IfTrue || back->Opcode() == Op_IfFalse) &&
|
|
82 back->in(0), "if-projection exists");
|
|
83 IfNode* back_if = back->in(0)->as_If();
|
|
84 float loop_back_cnt = back_if->_fcnt * back_if->_prob;
|
|
85
|
|
86 // Now compute a loop exit count
|
|
87 float loop_exit_cnt = 0.0f;
|
|
88 for( uint i = 0; i < _body.size(); i++ ) {
|
|
89 Node *n = _body[i];
|
|
90 if( n->is_If() ) {
|
|
91 IfNode *iff = n->as_If();
|
|
92 if( iff->_fcnt != COUNT_UNKNOWN && iff->_prob != PROB_UNKNOWN ) {
|
|
93 Node *exit = is_loop_exit(iff);
|
|
94 if( exit ) {
|
|
95 float exit_prob = iff->_prob;
|
|
96 if (exit->Opcode() == Op_IfFalse) exit_prob = 1.0 - exit_prob;
|
|
97 if (exit_prob > PROB_MIN) {
|
|
98 float exit_cnt = iff->_fcnt * exit_prob;
|
|
99 loop_exit_cnt += exit_cnt;
|
|
100 }
|
|
101 }
|
|
102 }
|
|
103 }
|
|
104 }
|
|
105 if (loop_exit_cnt > 0.0f) {
|
|
106 trip_cnt = (loop_back_cnt + loop_exit_cnt) / loop_exit_cnt;
|
|
107 } else {
|
|
108 // No exit count so use
|
|
109 trip_cnt = loop_back_cnt;
|
|
110 }
|
|
111 }
|
|
112 #ifndef PRODUCT
|
|
113 if (TraceProfileTripCount) {
|
|
114 tty->print_cr("compute_profile_trip_cnt lp: %d cnt: %f\n", head->_idx, trip_cnt);
|
|
115 }
|
|
116 #endif
|
|
117 head->set_profile_trip_cnt(trip_cnt);
|
|
118 }
|
|
119
|
|
120 //---------------------is_invariant_addition-----------------------------
|
|
121 // Return nonzero index of invariant operand for an Add or Sub
|
|
122 // of (nonconstant) invariant and variant values. Helper for reassoicate_invariants.
|
|
123 int IdealLoopTree::is_invariant_addition(Node* n, PhaseIdealLoop *phase) {
|
|
124 int op = n->Opcode();
|
|
125 if (op == Op_AddI || op == Op_SubI) {
|
|
126 bool in1_invar = this->is_invariant(n->in(1));
|
|
127 bool in2_invar = this->is_invariant(n->in(2));
|
|
128 if (in1_invar && !in2_invar) return 1;
|
|
129 if (!in1_invar && in2_invar) return 2;
|
|
130 }
|
|
131 return 0;
|
|
132 }
|
|
133
|
|
134 //---------------------reassociate_add_sub-----------------------------
|
|
135 // Reassociate invariant add and subtract expressions:
|
|
136 //
|
|
137 // inv1 + (x + inv2) => ( inv1 + inv2) + x
|
|
138 // (x + inv2) + inv1 => ( inv1 + inv2) + x
|
|
139 // inv1 + (x - inv2) => ( inv1 - inv2) + x
|
|
140 // inv1 - (inv2 - x) => ( inv1 - inv2) + x
|
|
141 // (x + inv2) - inv1 => (-inv1 + inv2) + x
|
|
142 // (x - inv2) + inv1 => ( inv1 - inv2) + x
|
|
143 // (x - inv2) - inv1 => (-inv1 - inv2) + x
|
|
144 // inv1 + (inv2 - x) => ( inv1 + inv2) - x
|
|
145 // inv1 - (x - inv2) => ( inv1 + inv2) - x
|
|
146 // (inv2 - x) + inv1 => ( inv1 + inv2) - x
|
|
147 // (inv2 - x) - inv1 => (-inv1 + inv2) - x
|
|
148 // inv1 - (x + inv2) => ( inv1 - inv2) - x
|
|
149 //
|
|
150 Node* IdealLoopTree::reassociate_add_sub(Node* n1, PhaseIdealLoop *phase) {
|
|
151 if (!n1->is_Add() && !n1->is_Sub() || n1->outcnt() == 0) return NULL;
|
|
152 if (is_invariant(n1)) return NULL;
|
|
153 int inv1_idx = is_invariant_addition(n1, phase);
|
|
154 if (!inv1_idx) return NULL;
|
|
155 // Don't mess with add of constant (igvn moves them to expression tree root.)
|
|
156 if (n1->is_Add() && n1->in(2)->is_Con()) return NULL;
|
|
157 Node* inv1 = n1->in(inv1_idx);
|
|
158 Node* n2 = n1->in(3 - inv1_idx);
|
|
159 int inv2_idx = is_invariant_addition(n2, phase);
|
|
160 if (!inv2_idx) return NULL;
|
|
161 Node* x = n2->in(3 - inv2_idx);
|
|
162 Node* inv2 = n2->in(inv2_idx);
|
|
163
|
|
164 bool neg_x = n2->is_Sub() && inv2_idx == 1;
|
|
165 bool neg_inv2 = n2->is_Sub() && inv2_idx == 2;
|
|
166 bool neg_inv1 = n1->is_Sub() && inv1_idx == 2;
|
|
167 if (n1->is_Sub() && inv1_idx == 1) {
|
|
168 neg_x = !neg_x;
|
|
169 neg_inv2 = !neg_inv2;
|
|
170 }
|
|
171 Node* inv1_c = phase->get_ctrl(inv1);
|
|
172 Node* inv2_c = phase->get_ctrl(inv2);
|
|
173 Node* n_inv1;
|
|
174 if (neg_inv1) {
|
|
175 Node *zero = phase->_igvn.intcon(0);
|
|
176 phase->set_ctrl(zero, phase->C->root());
|
|
177 n_inv1 = new (phase->C, 3) SubINode(zero, inv1);
|
|
178 phase->register_new_node(n_inv1, inv1_c);
|
|
179 } else {
|
|
180 n_inv1 = inv1;
|
|
181 }
|
|
182 Node* inv;
|
|
183 if (neg_inv2) {
|
|
184 inv = new (phase->C, 3) SubINode(n_inv1, inv2);
|
|
185 } else {
|
|
186 inv = new (phase->C, 3) AddINode(n_inv1, inv2);
|
|
187 }
|
|
188 phase->register_new_node(inv, phase->get_early_ctrl(inv));
|
|
189
|
|
190 Node* addx;
|
|
191 if (neg_x) {
|
|
192 addx = new (phase->C, 3) SubINode(inv, x);
|
|
193 } else {
|
|
194 addx = new (phase->C, 3) AddINode(x, inv);
|
|
195 }
|
|
196 phase->register_new_node(addx, phase->get_ctrl(x));
|
|
197 phase->_igvn.hash_delete(n1);
|
|
198 phase->_igvn.subsume_node(n1, addx);
|
|
199 return addx;
|
|
200 }
|
|
201
|
|
202 //---------------------reassociate_invariants-----------------------------
|
|
203 // Reassociate invariant expressions:
|
|
204 void IdealLoopTree::reassociate_invariants(PhaseIdealLoop *phase) {
|
|
205 for (int i = _body.size() - 1; i >= 0; i--) {
|
|
206 Node *n = _body.at(i);
|
|
207 for (int j = 0; j < 5; j++) {
|
|
208 Node* nn = reassociate_add_sub(n, phase);
|
|
209 if (nn == NULL) break;
|
|
210 n = nn; // again
|
|
211 };
|
|
212 }
|
|
213 }
|
|
214
|
|
215 //------------------------------policy_peeling---------------------------------
|
|
216 // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can
|
|
217 // make some loop-invariant test (usually a null-check) happen before the loop.
|
|
218 bool IdealLoopTree::policy_peeling( PhaseIdealLoop *phase ) const {
|
|
219 Node *test = ((IdealLoopTree*)this)->tail();
|
|
220 int body_size = ((IdealLoopTree*)this)->_body.size();
|
|
221 int uniq = phase->C->unique();
|
|
222 // Peeling does loop cloning which can result in O(N^2) node construction
|
|
223 if( body_size > 255 /* Prevent overflow for large body_size */
|
|
224 || (body_size * body_size + uniq > MaxNodeLimit) ) {
|
|
225 return false; // too large to safely clone
|
|
226 }
|
|
227 while( test != _head ) { // Scan till run off top of loop
|
|
228 if( test->is_If() ) { // Test?
|
|
229 Node *ctrl = phase->get_ctrl(test->in(1));
|
|
230 if (ctrl->is_top())
|
|
231 return false; // Found dead test on live IF? No peeling!
|
|
232 // Standard IF only has one input value to check for loop invariance
|
|
233 assert( test->Opcode() == Op_If || test->Opcode() == Op_CountedLoopEnd, "Check this code when new subtype is added");
|
|
234 // Condition is not a member of this loop?
|
|
235 if( !is_member(phase->get_loop(ctrl)) &&
|
|
236 is_loop_exit(test) )
|
|
237 return true; // Found reason to peel!
|
|
238 }
|
|
239 // Walk up dominators to loop _head looking for test which is
|
|
240 // executed on every path thru loop.
|
|
241 test = phase->idom(test);
|
|
242 }
|
|
243 return false;
|
|
244 }
|
|
245
|
|
246 //------------------------------peeled_dom_test_elim---------------------------
|
|
247 // If we got the effect of peeling, either by actually peeling or by making
|
|
248 // a pre-loop which must execute at least once, we can remove all
|
|
249 // loop-invariant dominated tests in the main body.
|
|
250 void PhaseIdealLoop::peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new ) {
|
|
251 bool progress = true;
|
|
252 while( progress ) {
|
|
253 progress = false; // Reset for next iteration
|
|
254 Node *prev = loop->_head->in(LoopNode::LoopBackControl);//loop->tail();
|
|
255 Node *test = prev->in(0);
|
|
256 while( test != loop->_head ) { // Scan till run off top of loop
|
|
257
|
|
258 int p_op = prev->Opcode();
|
|
259 if( (p_op == Op_IfFalse || p_op == Op_IfTrue) &&
|
|
260 test->is_If() && // Test?
|
|
261 !test->in(1)->is_Con() && // And not already obvious?
|
|
262 // Condition is not a member of this loop?
|
|
263 !loop->is_member(get_loop(get_ctrl(test->in(1))))){
|
|
264 // Walk loop body looking for instances of this test
|
|
265 for( uint i = 0; i < loop->_body.size(); i++ ) {
|
|
266 Node *n = loop->_body.at(i);
|
|
267 if( n->is_If() && n->in(1) == test->in(1) /*&& n != loop->tail()->in(0)*/ ) {
|
|
268 // IfNode was dominated by version in peeled loop body
|
|
269 progress = true;
|
|
270 dominated_by( old_new[prev->_idx], n );
|
|
271 }
|
|
272 }
|
|
273 }
|
|
274 prev = test;
|
|
275 test = idom(test);
|
|
276 } // End of scan tests in loop
|
|
277
|
|
278 } // End of while( progress )
|
|
279 }
|
|
280
|
|
281 //------------------------------do_peeling-------------------------------------
|
|
282 // Peel the first iteration of the given loop.
|
|
283 // Step 1: Clone the loop body. The clone becomes the peeled iteration.
|
|
284 // The pre-loop illegally has 2 control users (old & new loops).
|
|
285 // Step 2: Make the old-loop fall-in edges point to the peeled iteration.
|
|
286 // Do this by making the old-loop fall-in edges act as if they came
|
|
287 // around the loopback from the prior iteration (follow the old-loop
|
|
288 // backedges) and then map to the new peeled iteration. This leaves
|
|
289 // the pre-loop with only 1 user (the new peeled iteration), but the
|
|
290 // peeled-loop backedge has 2 users.
|
|
291 // Step 3: Cut the backedge on the clone (so its not a loop) and remove the
|
|
292 // extra backedge user.
|
|
293 void PhaseIdealLoop::do_peeling( IdealLoopTree *loop, Node_List &old_new ) {
|
|
294
|
|
295 C->set_major_progress();
|
|
296 // Peeling a 'main' loop in a pre/main/post situation obfuscates the
|
|
297 // 'pre' loop from the main and the 'pre' can no longer have it's
|
|
298 // iterations adjusted. Therefore, we need to declare this loop as
|
|
299 // no longer a 'main' loop; it will need new pre and post loops before
|
|
300 // we can do further RCE.
|
|
301 Node *h = loop->_head;
|
|
302 if( h->is_CountedLoop() ) {
|
|
303 CountedLoopNode *cl = h->as_CountedLoop();
|
|
304 assert(cl->trip_count() > 0, "peeling a fully unrolled loop");
|
|
305 cl->set_trip_count(cl->trip_count() - 1);
|
|
306 if( cl->is_main_loop() ) {
|
|
307 cl->set_normal_loop();
|
|
308 #ifndef PRODUCT
|
|
309 if( PrintOpto && VerifyLoopOptimizations ) {
|
|
310 tty->print("Peeling a 'main' loop; resetting to 'normal' ");
|
|
311 loop->dump_head();
|
|
312 }
|
|
313 #endif
|
|
314 }
|
|
315 }
|
|
316
|
|
317 // Step 1: Clone the loop body. The clone becomes the peeled iteration.
|
|
318 // The pre-loop illegally has 2 control users (old & new loops).
|
|
319 clone_loop( loop, old_new, dom_depth(loop->_head) );
|
|
320
|
|
321
|
|
322 // Step 2: Make the old-loop fall-in edges point to the peeled iteration.
|
|
323 // Do this by making the old-loop fall-in edges act as if they came
|
|
324 // around the loopback from the prior iteration (follow the old-loop
|
|
325 // backedges) and then map to the new peeled iteration. This leaves
|
|
326 // the pre-loop with only 1 user (the new peeled iteration), but the
|
|
327 // peeled-loop backedge has 2 users.
|
|
328 for (DUIterator_Fast jmax, j = loop->_head->fast_outs(jmax); j < jmax; j++) {
|
|
329 Node* old = loop->_head->fast_out(j);
|
|
330 if( old->in(0) == loop->_head && old->req() == 3 &&
|
|
331 (old->is_Loop() || old->is_Phi()) ) {
|
|
332 Node *new_exit_value = old_new[old->in(LoopNode::LoopBackControl)->_idx];
|
|
333 if( !new_exit_value ) // Backedge value is ALSO loop invariant?
|
|
334 // Then loop body backedge value remains the same.
|
|
335 new_exit_value = old->in(LoopNode::LoopBackControl);
|
|
336 _igvn.hash_delete(old);
|
|
337 old->set_req(LoopNode::EntryControl, new_exit_value);
|
|
338 }
|
|
339 }
|
|
340
|
|
341
|
|
342 // Step 3: Cut the backedge on the clone (so its not a loop) and remove the
|
|
343 // extra backedge user.
|
|
344 Node *nnn = old_new[loop->_head->_idx];
|
|
345 _igvn.hash_delete(nnn);
|
|
346 nnn->set_req(LoopNode::LoopBackControl, C->top());
|
|
347 for (DUIterator_Fast j2max, j2 = nnn->fast_outs(j2max); j2 < j2max; j2++) {
|
|
348 Node* use = nnn->fast_out(j2);
|
|
349 if( use->in(0) == nnn && use->req() == 3 && use->is_Phi() ) {
|
|
350 _igvn.hash_delete(use);
|
|
351 use->set_req(LoopNode::LoopBackControl, C->top());
|
|
352 }
|
|
353 }
|
|
354
|
|
355
|
|
356 // Step 4: Correct dom-depth info. Set to loop-head depth.
|
|
357 int dd = dom_depth(loop->_head);
|
|
358 set_idom(loop->_head, loop->_head->in(1), dd);
|
|
359 for (uint j3 = 0; j3 < loop->_body.size(); j3++) {
|
|
360 Node *old = loop->_body.at(j3);
|
|
361 Node *nnn = old_new[old->_idx];
|
|
362 if (!has_ctrl(nnn))
|
|
363 set_idom(nnn, idom(nnn), dd-1);
|
|
364 // While we're at it, remove any SafePoints from the peeled code
|
|
365 if( old->Opcode() == Op_SafePoint ) {
|
|
366 Node *nnn = old_new[old->_idx];
|
|
367 lazy_replace(nnn,nnn->in(TypeFunc::Control));
|
|
368 }
|
|
369 }
|
|
370
|
|
371 // Now force out all loop-invariant dominating tests. The optimizer
|
|
372 // finds some, but we _know_ they are all useless.
|
|
373 peeled_dom_test_elim(loop,old_new);
|
|
374
|
|
375 loop->record_for_igvn();
|
|
376 }
|
|
377
|
|
378 //------------------------------policy_maximally_unroll------------------------
|
|
379 // Return exact loop trip count, or 0 if not maximally unrolling
|
|
380 bool IdealLoopTree::policy_maximally_unroll( PhaseIdealLoop *phase ) const {
|
|
381 CountedLoopNode *cl = _head->as_CountedLoop();
|
|
382 assert( cl->is_normal_loop(), "" );
|
|
383
|
|
384 Node *init_n = cl->init_trip();
|
|
385 Node *limit_n = cl->limit();
|
|
386
|
|
387 // Non-constant bounds
|
|
388 if( init_n == NULL || !init_n->is_Con() ||
|
|
389 limit_n == NULL || !limit_n->is_Con() ||
|
|
390 // protect against stride not being a constant
|
|
391 !cl->stride_is_con() ) {
|
|
392 return false;
|
|
393 }
|
|
394 int init = init_n->get_int();
|
|
395 int limit = limit_n->get_int();
|
|
396 int span = limit - init;
|
|
397 int stride = cl->stride_con();
|
|
398
|
|
399 if (init >= limit || stride > span) {
|
|
400 // return a false (no maximally unroll) and the regular unroll/peel
|
|
401 // route will make a small mess which CCP will fold away.
|
|
402 return false;
|
|
403 }
|
|
404 uint trip_count = span/stride; // trip_count can be greater than 2 Gig.
|
|
405 assert( (int)trip_count*stride == span, "must divide evenly" );
|
|
406
|
|
407 // Real policy: if we maximally unroll, does it get too big?
|
|
408 // Allow the unrolled mess to get larger than standard loop
|
|
409 // size. After all, it will no longer be a loop.
|
|
410 uint body_size = _body.size();
|
|
411 uint unroll_limit = (uint)LoopUnrollLimit * 4;
|
|
412 assert( (intx)unroll_limit == LoopUnrollLimit * 4, "LoopUnrollLimit must fit in 32bits");
|
|
413 cl->set_trip_count(trip_count);
|
|
414 if( trip_count <= unroll_limit && body_size <= unroll_limit ) {
|
|
415 uint new_body_size = body_size * trip_count;
|
|
416 if (new_body_size <= unroll_limit &&
|
|
417 body_size == new_body_size / trip_count &&
|
|
418 // Unrolling can result in a large amount of node construction
|
|
419 new_body_size < MaxNodeLimit - phase->C->unique()) {
|
|
420 return true; // maximally unroll
|
|
421 }
|
|
422 }
|
|
423
|
|
424 return false; // Do not maximally unroll
|
|
425 }
|
|
426
|
|
427
|
|
428 //------------------------------policy_unroll----------------------------------
|
|
429 // Return TRUE or FALSE if the loop should be unrolled or not. Unroll if
|
|
430 // the loop is a CountedLoop and the body is small enough.
|
|
431 bool IdealLoopTree::policy_unroll( PhaseIdealLoop *phase ) const {
|
|
432
|
|
433 CountedLoopNode *cl = _head->as_CountedLoop();
|
|
434 assert( cl->is_normal_loop() || cl->is_main_loop(), "" );
|
|
435
|
|
436 // protect against stride not being a constant
|
|
437 if( !cl->stride_is_con() ) return false;
|
|
438
|
|
439 // protect against over-unrolling
|
|
440 if( cl->trip_count() <= 1 ) return false;
|
|
441
|
|
442 int future_unroll_ct = cl->unrolled_count() * 2;
|
|
443
|
|
444 // Don't unroll if the next round of unrolling would push us
|
|
445 // over the expected trip count of the loop. One is subtracted
|
|
446 // from the expected trip count because the pre-loop normally
|
|
447 // executes 1 iteration.
|
|
448 if (UnrollLimitForProfileCheck > 0 &&
|
|
449 cl->profile_trip_cnt() != COUNT_UNKNOWN &&
|
|
450 future_unroll_ct > UnrollLimitForProfileCheck &&
|
|
451 (float)future_unroll_ct > cl->profile_trip_cnt() - 1.0) {
|
|
452 return false;
|
|
453 }
|
|
454
|
|
455 // When unroll count is greater than LoopUnrollMin, don't unroll if:
|
|
456 // the residual iterations are more than 10% of the trip count
|
|
457 // and rounds of "unroll,optimize" are not making significant progress
|
|
458 // Progress defined as current size less than 20% larger than previous size.
|
|
459 if (UseSuperWord && cl->node_count_before_unroll() > 0 &&
|
|
460 future_unroll_ct > LoopUnrollMin &&
|
|
461 (future_unroll_ct - 1) * 10.0 > cl->profile_trip_cnt() &&
|
|
462 1.2 * cl->node_count_before_unroll() < (double)_body.size()) {
|
|
463 return false;
|
|
464 }
|
|
465
|
|
466 Node *init_n = cl->init_trip();
|
|
467 Node *limit_n = cl->limit();
|
|
468 // Non-constant bounds.
|
|
469 // Protect against over-unrolling when init or/and limit are not constant
|
|
470 // (so that trip_count's init value is maxint) but iv range is known.
|
|
471 if( init_n == NULL || !init_n->is_Con() ||
|
|
472 limit_n == NULL || !limit_n->is_Con() ) {
|
|
473 Node* phi = cl->phi();
|
|
474 if( phi != NULL ) {
|
|
475 assert(phi->is_Phi() && phi->in(0) == _head, "Counted loop should have iv phi.");
|
|
476 const TypeInt* iv_type = phase->_igvn.type(phi)->is_int();
|
|
477 int next_stride = cl->stride_con() * 2; // stride after this unroll
|
|
478 if( next_stride > 0 ) {
|
|
479 if( iv_type->_lo + next_stride <= iv_type->_lo || // overflow
|
|
480 iv_type->_lo + next_stride > iv_type->_hi ) {
|
|
481 return false; // over-unrolling
|
|
482 }
|
|
483 } else if( next_stride < 0 ) {
|
|
484 if( iv_type->_hi + next_stride >= iv_type->_hi || // overflow
|
|
485 iv_type->_hi + next_stride < iv_type->_lo ) {
|
|
486 return false; // over-unrolling
|
|
487 }
|
|
488 }
|
|
489 }
|
|
490 }
|
|
491
|
|
492 // Adjust body_size to determine if we unroll or not
|
|
493 uint body_size = _body.size();
|
|
494 // Key test to unroll CaffeineMark's Logic test
|
|
495 int xors_in_loop = 0;
|
|
496 // Also count ModL, DivL and MulL which expand mightly
|
|
497 for( uint k = 0; k < _body.size(); k++ ) {
|
|
498 switch( _body.at(k)->Opcode() ) {
|
|
499 case Op_XorI: xors_in_loop++; break; // CaffeineMark's Logic test
|
|
500 case Op_ModL: body_size += 30; break;
|
|
501 case Op_DivL: body_size += 30; break;
|
|
502 case Op_MulL: body_size += 10; break;
|
|
503 }
|
|
504 }
|
|
505
|
|
506 // Check for being too big
|
|
507 if( body_size > (uint)LoopUnrollLimit ) {
|
|
508 if( xors_in_loop >= 4 && body_size < (uint)LoopUnrollLimit*4) return true;
|
|
509 // Normal case: loop too big
|
|
510 return false;
|
|
511 }
|
|
512
|
|
513 // Check for stride being a small enough constant
|
|
514 if( abs(cl->stride_con()) > (1<<3) ) return false;
|
|
515
|
|
516 // Unroll once! (Each trip will soon do double iterations)
|
|
517 return true;
|
|
518 }
|
|
519
|
|
520 //------------------------------policy_align-----------------------------------
|
|
521 // Return TRUE or FALSE if the loop should be cache-line aligned. Gather the
|
|
522 // expression that does the alignment. Note that only one array base can be
|
|
523 // aligned in a loop (unless the VM guarentees mutual alignment). Note that
|
|
524 // if we vectorize short memory ops into longer memory ops, we may want to
|
|
525 // increase alignment.
|
|
526 bool IdealLoopTree::policy_align( PhaseIdealLoop *phase ) const {
|
|
527 return false;
|
|
528 }
|
|
529
|
|
530 //------------------------------policy_range_check-----------------------------
|
|
531 // Return TRUE or FALSE if the loop should be range-check-eliminated.
|
|
532 // Actually we do iteration-splitting, a more powerful form of RCE.
|
|
533 bool IdealLoopTree::policy_range_check( PhaseIdealLoop *phase ) const {
|
|
534 if( !RangeCheckElimination ) return false;
|
|
535
|
|
536 CountedLoopNode *cl = _head->as_CountedLoop();
|
|
537 // If we unrolled with no intention of doing RCE and we later
|
|
538 // changed our minds, we got no pre-loop. Either we need to
|
|
539 // make a new pre-loop, or we gotta disallow RCE.
|
|
540 if( cl->is_main_no_pre_loop() ) return false; // Disallowed for now.
|
|
541 Node *trip_counter = cl->phi();
|
|
542
|
|
543 // Check loop body for tests of trip-counter plus loop-invariant vs
|
|
544 // loop-invariant.
|
|
545 for( uint i = 0; i < _body.size(); i++ ) {
|
|
546 Node *iff = _body[i];
|
|
547 if( iff->Opcode() == Op_If ) { // Test?
|
|
548
|
|
549 // Comparing trip+off vs limit
|
|
550 Node *bol = iff->in(1);
|
|
551 if( bol->req() != 2 ) continue; // dead constant test
|
|
552 Node *cmp = bol->in(1);
|
|
553
|
|
554 Node *rc_exp = cmp->in(1);
|
|
555 Node *limit = cmp->in(2);
|
|
556
|
|
557 Node *limit_c = phase->get_ctrl(limit);
|
|
558 if( limit_c == phase->C->top() )
|
|
559 return false; // Found dead test on live IF? No RCE!
|
|
560 if( is_member(phase->get_loop(limit_c) ) ) {
|
|
561 // Compare might have operands swapped; commute them
|
|
562 rc_exp = cmp->in(2);
|
|
563 limit = cmp->in(1);
|
|
564 limit_c = phase->get_ctrl(limit);
|
|
565 if( is_member(phase->get_loop(limit_c) ) )
|
|
566 continue; // Both inputs are loop varying; cannot RCE
|
|
567 }
|
|
568
|
|
569 if (!phase->is_scaled_iv_plus_offset(rc_exp, trip_counter, NULL, NULL)) {
|
|
570 continue;
|
|
571 }
|
|
572 // Yeah! Found a test like 'trip+off vs limit'
|
|
573 // Test is an IfNode, has 2 projections. If BOTH are in the loop
|
|
574 // we need loop unswitching instead of iteration splitting.
|
|
575 if( is_loop_exit(iff) )
|
|
576 return true; // Found reason to split iterations
|
|
577 } // End of is IF
|
|
578 }
|
|
579
|
|
580 return false;
|
|
581 }
|
|
582
|
|
583 //------------------------------policy_peel_only-------------------------------
|
|
584 // Return TRUE or FALSE if the loop should NEVER be RCE'd or aligned. Useful
|
|
585 // for unrolling loops with NO array accesses.
|
|
586 bool IdealLoopTree::policy_peel_only( PhaseIdealLoop *phase ) const {
|
|
587
|
|
588 for( uint i = 0; i < _body.size(); i++ )
|
|
589 if( _body[i]->is_Mem() )
|
|
590 return false;
|
|
591
|
|
592 // No memory accesses at all!
|
|
593 return true;
|
|
594 }
|
|
595
|
|
596 //------------------------------clone_up_backedge_goo--------------------------
|
|
597 // If Node n lives in the back_ctrl block and cannot float, we clone a private
|
|
598 // version of n in preheader_ctrl block and return that, otherwise return n.
|
|
599 Node *PhaseIdealLoop::clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n ) {
|
|
600 if( get_ctrl(n) != back_ctrl ) return n;
|
|
601
|
|
602 Node *x = NULL; // If required, a clone of 'n'
|
|
603 // Check for 'n' being pinned in the backedge.
|
|
604 if( n->in(0) && n->in(0) == back_ctrl ) {
|
|
605 x = n->clone(); // Clone a copy of 'n' to preheader
|
|
606 x->set_req( 0, preheader_ctrl ); // Fix x's control input to preheader
|
|
607 }
|
|
608
|
|
609 // Recursive fixup any other input edges into x.
|
|
610 // If there are no changes we can just return 'n', otherwise
|
|
611 // we need to clone a private copy and change it.
|
|
612 for( uint i = 1; i < n->req(); i++ ) {
|
|
613 Node *g = clone_up_backedge_goo( back_ctrl, preheader_ctrl, n->in(i) );
|
|
614 if( g != n->in(i) ) {
|
|
615 if( !x )
|
|
616 x = n->clone();
|
|
617 x->set_req(i, g);
|
|
618 }
|
|
619 }
|
|
620 if( x ) { // x can legally float to pre-header location
|
|
621 register_new_node( x, preheader_ctrl );
|
|
622 return x;
|
|
623 } else { // raise n to cover LCA of uses
|
|
624 set_ctrl( n, find_non_split_ctrl(back_ctrl->in(0)) );
|
|
625 }
|
|
626 return n;
|
|
627 }
|
|
628
|
|
629 //------------------------------insert_pre_post_loops--------------------------
|
|
630 // Insert pre and post loops. If peel_only is set, the pre-loop can not have
|
|
631 // more iterations added. It acts as a 'peel' only, no lower-bound RCE, no
|
|
632 // alignment. Useful to unroll loops that do no array accesses.
|
|
633 void PhaseIdealLoop::insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only ) {
|
|
634
|
|
635 C->set_major_progress();
|
|
636
|
|
637 // Find common pieces of the loop being guarded with pre & post loops
|
|
638 CountedLoopNode *main_head = loop->_head->as_CountedLoop();
|
|
639 assert( main_head->is_normal_loop(), "" );
|
|
640 CountedLoopEndNode *main_end = main_head->loopexit();
|
|
641 assert( main_end->outcnt() == 2, "1 true, 1 false path only" );
|
|
642 uint dd_main_head = dom_depth(main_head);
|
|
643 uint max = main_head->outcnt();
|
|
644
|
|
645 Node *pre_header= main_head->in(LoopNode::EntryControl);
|
|
646 Node *init = main_head->init_trip();
|
|
647 Node *incr = main_end ->incr();
|
|
648 Node *limit = main_end ->limit();
|
|
649 Node *stride = main_end ->stride();
|
|
650 Node *cmp = main_end ->cmp_node();
|
|
651 BoolTest::mask b_test = main_end->test_trip();
|
|
652
|
|
653 // Need only 1 user of 'bol' because I will be hacking the loop bounds.
|
|
654 Node *bol = main_end->in(CountedLoopEndNode::TestValue);
|
|
655 if( bol->outcnt() != 1 ) {
|
|
656 bol = bol->clone();
|
|
657 register_new_node(bol,main_end->in(CountedLoopEndNode::TestControl));
|
|
658 _igvn.hash_delete(main_end);
|
|
659 main_end->set_req(CountedLoopEndNode::TestValue, bol);
|
|
660 }
|
|
661 // Need only 1 user of 'cmp' because I will be hacking the loop bounds.
|
|
662 if( cmp->outcnt() != 1 ) {
|
|
663 cmp = cmp->clone();
|
|
664 register_new_node(cmp,main_end->in(CountedLoopEndNode::TestControl));
|
|
665 _igvn.hash_delete(bol);
|
|
666 bol->set_req(1, cmp);
|
|
667 }
|
|
668
|
|
669 //------------------------------
|
|
670 // Step A: Create Post-Loop.
|
|
671 Node* main_exit = main_end->proj_out(false);
|
|
672 assert( main_exit->Opcode() == Op_IfFalse, "" );
|
|
673 int dd_main_exit = dom_depth(main_exit);
|
|
674
|
|
675 // Step A1: Clone the loop body. The clone becomes the post-loop. The main
|
|
676 // loop pre-header illegally has 2 control users (old & new loops).
|
|
677 clone_loop( loop, old_new, dd_main_exit );
|
|
678 assert( old_new[main_end ->_idx]->Opcode() == Op_CountedLoopEnd, "" );
|
|
679 CountedLoopNode *post_head = old_new[main_head->_idx]->as_CountedLoop();
|
|
680 post_head->set_post_loop(main_head);
|
|
681
|
|
682 // Build the main-loop normal exit.
|
|
683 IfFalseNode *new_main_exit = new (C, 1) IfFalseNode(main_end);
|
|
684 _igvn.register_new_node_with_optimizer( new_main_exit );
|
|
685 set_idom(new_main_exit, main_end, dd_main_exit );
|
|
686 set_loop(new_main_exit, loop->_parent);
|
|
687
|
|
688 // Step A2: Build a zero-trip guard for the post-loop. After leaving the
|
|
689 // main-loop, the post-loop may not execute at all. We 'opaque' the incr
|
|
690 // (the main-loop trip-counter exit value) because we will be changing
|
|
691 // the exit value (via unrolling) so we cannot constant-fold away the zero
|
|
692 // trip guard until all unrolling is done.
|
|
693 Node *zer_opaq = new (C, 2) Opaque1Node(incr);
|
|
694 Node *zer_cmp = new (C, 3) CmpINode( zer_opaq, limit );
|
|
695 Node *zer_bol = new (C, 2) BoolNode( zer_cmp, b_test );
|
|
696 register_new_node( zer_opaq, new_main_exit );
|
|
697 register_new_node( zer_cmp , new_main_exit );
|
|
698 register_new_node( zer_bol , new_main_exit );
|
|
699
|
|
700 // Build the IfNode
|
|
701 IfNode *zer_iff = new (C, 2) IfNode( new_main_exit, zer_bol, PROB_FAIR, COUNT_UNKNOWN );
|
|
702 _igvn.register_new_node_with_optimizer( zer_iff );
|
|
703 set_idom(zer_iff, new_main_exit, dd_main_exit);
|
|
704 set_loop(zer_iff, loop->_parent);
|
|
705
|
|
706 // Plug in the false-path, taken if we need to skip post-loop
|
|
707 _igvn.hash_delete( main_exit );
|
|
708 main_exit->set_req(0, zer_iff);
|
|
709 _igvn._worklist.push(main_exit);
|
|
710 set_idom(main_exit, zer_iff, dd_main_exit);
|
|
711 set_idom(main_exit->unique_out(), zer_iff, dd_main_exit);
|
|
712 // Make the true-path, must enter the post loop
|
|
713 Node *zer_taken = new (C, 1) IfTrueNode( zer_iff );
|
|
714 _igvn.register_new_node_with_optimizer( zer_taken );
|
|
715 set_idom(zer_taken, zer_iff, dd_main_exit);
|
|
716 set_loop(zer_taken, loop->_parent);
|
|
717 // Plug in the true path
|
|
718 _igvn.hash_delete( post_head );
|
|
719 post_head->set_req(LoopNode::EntryControl, zer_taken);
|
|
720 set_idom(post_head, zer_taken, dd_main_exit);
|
|
721
|
|
722 // Step A3: Make the fall-in values to the post-loop come from the
|
|
723 // fall-out values of the main-loop.
|
|
724 for (DUIterator_Fast imax, i = main_head->fast_outs(imax); i < imax; i++) {
|
|
725 Node* main_phi = main_head->fast_out(i);
|
|
726 if( main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() >0 ) {
|
|
727 Node *post_phi = old_new[main_phi->_idx];
|
|
728 Node *fallmain = clone_up_backedge_goo(main_head->back_control(),
|
|
729 post_head->init_control(),
|
|
730 main_phi->in(LoopNode::LoopBackControl));
|
|
731 _igvn.hash_delete(post_phi);
|
|
732 post_phi->set_req( LoopNode::EntryControl, fallmain );
|
|
733 }
|
|
734 }
|
|
735
|
|
736 // Update local caches for next stanza
|
|
737 main_exit = new_main_exit;
|
|
738
|
|
739
|
|
740 //------------------------------
|
|
741 // Step B: Create Pre-Loop.
|
|
742
|
|
743 // Step B1: Clone the loop body. The clone becomes the pre-loop. The main
|
|
744 // loop pre-header illegally has 2 control users (old & new loops).
|
|
745 clone_loop( loop, old_new, dd_main_head );
|
|
746 CountedLoopNode* pre_head = old_new[main_head->_idx]->as_CountedLoop();
|
|
747 CountedLoopEndNode* pre_end = old_new[main_end ->_idx]->as_CountedLoopEnd();
|
|
748 pre_head->set_pre_loop(main_head);
|
|
749 Node *pre_incr = old_new[incr->_idx];
|
|
750
|
|
751 // Find the pre-loop normal exit.
|
|
752 Node* pre_exit = pre_end->proj_out(false);
|
|
753 assert( pre_exit->Opcode() == Op_IfFalse, "" );
|
|
754 IfFalseNode *new_pre_exit = new (C, 1) IfFalseNode(pre_end);
|
|
755 _igvn.register_new_node_with_optimizer( new_pre_exit );
|
|
756 set_idom(new_pre_exit, pre_end, dd_main_head);
|
|
757 set_loop(new_pre_exit, loop->_parent);
|
|
758
|
|
759 // Step B2: Build a zero-trip guard for the main-loop. After leaving the
|
|
760 // pre-loop, the main-loop may not execute at all. Later in life this
|
|
761 // zero-trip guard will become the minimum-trip guard when we unroll
|
|
762 // the main-loop.
|
|
763 Node *min_opaq = new (C, 2) Opaque1Node(limit);
|
|
764 Node *min_cmp = new (C, 3) CmpINode( pre_incr, min_opaq );
|
|
765 Node *min_bol = new (C, 2) BoolNode( min_cmp, b_test );
|
|
766 register_new_node( min_opaq, new_pre_exit );
|
|
767 register_new_node( min_cmp , new_pre_exit );
|
|
768 register_new_node( min_bol , new_pre_exit );
|
|
769
|
|
770 // Build the IfNode
|
|
771 IfNode *min_iff = new (C, 2) IfNode( new_pre_exit, min_bol, PROB_FAIR, COUNT_UNKNOWN );
|
|
772 _igvn.register_new_node_with_optimizer( min_iff );
|
|
773 set_idom(min_iff, new_pre_exit, dd_main_head);
|
|
774 set_loop(min_iff, loop->_parent);
|
|
775
|
|
776 // Plug in the false-path, taken if we need to skip main-loop
|
|
777 _igvn.hash_delete( pre_exit );
|
|
778 pre_exit->set_req(0, min_iff);
|
|
779 set_idom(pre_exit, min_iff, dd_main_head);
|
|
780 set_idom(pre_exit->unique_out(), min_iff, dd_main_head);
|
|
781 // Make the true-path, must enter the main loop
|
|
782 Node *min_taken = new (C, 1) IfTrueNode( min_iff );
|
|
783 _igvn.register_new_node_with_optimizer( min_taken );
|
|
784 set_idom(min_taken, min_iff, dd_main_head);
|
|
785 set_loop(min_taken, loop->_parent);
|
|
786 // Plug in the true path
|
|
787 _igvn.hash_delete( main_head );
|
|
788 main_head->set_req(LoopNode::EntryControl, min_taken);
|
|
789 set_idom(main_head, min_taken, dd_main_head);
|
|
790
|
|
791 // Step B3: Make the fall-in values to the main-loop come from the
|
|
792 // fall-out values of the pre-loop.
|
|
793 for (DUIterator_Fast i2max, i2 = main_head->fast_outs(i2max); i2 < i2max; i2++) {
|
|
794 Node* main_phi = main_head->fast_out(i2);
|
|
795 if( main_phi->is_Phi() && main_phi->in(0) == main_head && main_phi->outcnt() > 0 ) {
|
|
796 Node *pre_phi = old_new[main_phi->_idx];
|
|
797 Node *fallpre = clone_up_backedge_goo(pre_head->back_control(),
|
|
798 main_head->init_control(),
|
|
799 pre_phi->in(LoopNode::LoopBackControl));
|
|
800 _igvn.hash_delete(main_phi);
|
|
801 main_phi->set_req( LoopNode::EntryControl, fallpre );
|
|
802 }
|
|
803 }
|
|
804
|
|
805 // Step B4: Shorten the pre-loop to run only 1 iteration (for now).
|
|
806 // RCE and alignment may change this later.
|
|
807 Node *cmp_end = pre_end->cmp_node();
|
|
808 assert( cmp_end->in(2) == limit, "" );
|
|
809 Node *pre_limit = new (C, 3) AddINode( init, stride );
|
|
810
|
|
811 // Save the original loop limit in this Opaque1 node for
|
|
812 // use by range check elimination.
|
|
813 Node *pre_opaq = new (C, 3) Opaque1Node(pre_limit, limit);
|
|
814
|
|
815 register_new_node( pre_limit, pre_head->in(0) );
|
|
816 register_new_node( pre_opaq , pre_head->in(0) );
|
|
817
|
|
818 // Since no other users of pre-loop compare, I can hack limit directly
|
|
819 assert( cmp_end->outcnt() == 1, "no other users" );
|
|
820 _igvn.hash_delete(cmp_end);
|
|
821 cmp_end->set_req(2, peel_only ? pre_limit : pre_opaq);
|
|
822
|
|
823 // Special case for not-equal loop bounds:
|
|
824 // Change pre loop test, main loop test, and the
|
|
825 // main loop guard test to use lt or gt depending on stride
|
|
826 // direction:
|
|
827 // positive stride use <
|
|
828 // negative stride use >
|
|
829
|
|
830 if (pre_end->in(CountedLoopEndNode::TestValue)->as_Bool()->_test._test == BoolTest::ne) {
|
|
831
|
|
832 BoolTest::mask new_test = (main_end->stride_con() > 0) ? BoolTest::lt : BoolTest::gt;
|
|
833 // Modify pre loop end condition
|
|
834 Node* pre_bol = pre_end->in(CountedLoopEndNode::TestValue)->as_Bool();
|
|
835 BoolNode* new_bol0 = new (C, 2) BoolNode(pre_bol->in(1), new_test);
|
|
836 register_new_node( new_bol0, pre_head->in(0) );
|
|
837 _igvn.hash_delete(pre_end);
|
|
838 pre_end->set_req(CountedLoopEndNode::TestValue, new_bol0);
|
|
839 // Modify main loop guard condition
|
|
840 assert(min_iff->in(CountedLoopEndNode::TestValue) == min_bol, "guard okay");
|
|
841 BoolNode* new_bol1 = new (C, 2) BoolNode(min_bol->in(1), new_test);
|
|
842 register_new_node( new_bol1, new_pre_exit );
|
|
843 _igvn.hash_delete(min_iff);
|
|
844 min_iff->set_req(CountedLoopEndNode::TestValue, new_bol1);
|
|
845 // Modify main loop end condition
|
|
846 BoolNode* main_bol = main_end->in(CountedLoopEndNode::TestValue)->as_Bool();
|
|
847 BoolNode* new_bol2 = new (C, 2) BoolNode(main_bol->in(1), new_test);
|
|
848 register_new_node( new_bol2, main_end->in(CountedLoopEndNode::TestControl) );
|
|
849 _igvn.hash_delete(main_end);
|
|
850 main_end->set_req(CountedLoopEndNode::TestValue, new_bol2);
|
|
851 }
|
|
852
|
|
853 // Flag main loop
|
|
854 main_head->set_main_loop();
|
|
855 if( peel_only ) main_head->set_main_no_pre_loop();
|
|
856
|
|
857 // It's difficult to be precise about the trip-counts
|
|
858 // for the pre/post loops. They are usually very short,
|
|
859 // so guess that 4 trips is a reasonable value.
|
|
860 post_head->set_profile_trip_cnt(4.0);
|
|
861 pre_head->set_profile_trip_cnt(4.0);
|
|
862
|
|
863 // Now force out all loop-invariant dominating tests. The optimizer
|
|
864 // finds some, but we _know_ they are all useless.
|
|
865 peeled_dom_test_elim(loop,old_new);
|
|
866 }
|
|
867
|
|
868 //------------------------------is_invariant-----------------------------
|
|
869 // Return true if n is invariant
|
|
870 bool IdealLoopTree::is_invariant(Node* n) const {
|
|
871 Node *n_c = _phase->get_ctrl(n);
|
|
872 if (n_c->is_top()) return false;
|
|
873 return !is_member(_phase->get_loop(n_c));
|
|
874 }
|
|
875
|
|
876
|
|
877 //------------------------------do_unroll--------------------------------------
|
|
878 // Unroll the loop body one step - make each trip do 2 iterations.
|
|
879 void PhaseIdealLoop::do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip ) {
|
|
880 assert( LoopUnrollLimit, "" );
|
|
881 #ifndef PRODUCT
|
|
882 if( PrintOpto && VerifyLoopOptimizations ) {
|
|
883 tty->print("Unrolling ");
|
|
884 loop->dump_head();
|
|
885 }
|
|
886 #endif
|
|
887 CountedLoopNode *loop_head = loop->_head->as_CountedLoop();
|
|
888 CountedLoopEndNode *loop_end = loop_head->loopexit();
|
|
889 assert( loop_end, "" );
|
|
890
|
|
891 // Remember loop node count before unrolling to detect
|
|
892 // if rounds of unroll,optimize are making progress
|
|
893 loop_head->set_node_count_before_unroll(loop->_body.size());
|
|
894
|
|
895 Node *ctrl = loop_head->in(LoopNode::EntryControl);
|
|
896 Node *limit = loop_head->limit();
|
|
897 Node *init = loop_head->init_trip();
|
|
898 Node *strid = loop_head->stride();
|
|
899
|
|
900 Node *opaq = NULL;
|
|
901 if( adjust_min_trip ) { // If not maximally unrolling, need adjustment
|
|
902 assert( loop_head->is_main_loop(), "" );
|
|
903 assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" );
|
|
904 Node *iff = ctrl->in(0);
|
|
905 assert( iff->Opcode() == Op_If, "" );
|
|
906 Node *bol = iff->in(1);
|
|
907 assert( bol->Opcode() == Op_Bool, "" );
|
|
908 Node *cmp = bol->in(1);
|
|
909 assert( cmp->Opcode() == Op_CmpI, "" );
|
|
910 opaq = cmp->in(2);
|
|
911 // Occasionally it's possible for a pre-loop Opaque1 node to be
|
|
912 // optimized away and then another round of loop opts attempted.
|
|
913 // We can not optimize this particular loop in that case.
|
|
914 if( opaq->Opcode() != Op_Opaque1 )
|
|
915 return; // Cannot find pre-loop! Bail out!
|
|
916 }
|
|
917
|
|
918 C->set_major_progress();
|
|
919
|
|
920 // Adjust max trip count. The trip count is intentionally rounded
|
|
921 // down here (e.g. 15-> 7-> 3-> 1) because if we unwittingly over-unroll,
|
|
922 // the main, unrolled, part of the loop will never execute as it is protected
|
|
923 // by the min-trip test. See bug 4834191 for a case where we over-unrolled
|
|
924 // and later determined that part of the unrolled loop was dead.
|
|
925 loop_head->set_trip_count(loop_head->trip_count() / 2);
|
|
926
|
|
927 // Double the count of original iterations in the unrolled loop body.
|
|
928 loop_head->double_unrolled_count();
|
|
929
|
|
930 // -----------
|
|
931 // Step 2: Cut back the trip counter for an unroll amount of 2.
|
|
932 // Loop will normally trip (limit - init)/stride_con. Since it's a
|
|
933 // CountedLoop this is exact (stride divides limit-init exactly).
|
|
934 // We are going to double the loop body, so we want to knock off any
|
|
935 // odd iteration: (trip_cnt & ~1). Then back compute a new limit.
|
|
936 Node *span = new (C, 3) SubINode( limit, init );
|
|
937 register_new_node( span, ctrl );
|
|
938 Node *trip = new (C, 3) DivINode( 0, span, strid );
|
|
939 register_new_node( trip, ctrl );
|
|
940 Node *mtwo = _igvn.intcon(-2);
|
|
941 set_ctrl(mtwo, C->root());
|
|
942 Node *rond = new (C, 3) AndINode( trip, mtwo );
|
|
943 register_new_node( rond, ctrl );
|
|
944 Node *spn2 = new (C, 3) MulINode( rond, strid );
|
|
945 register_new_node( spn2, ctrl );
|
|
946 Node *lim2 = new (C, 3) AddINode( spn2, init );
|
|
947 register_new_node( lim2, ctrl );
|
|
948
|
|
949 // Hammer in the new limit
|
|
950 Node *ctrl2 = loop_end->in(0);
|
|
951 Node *cmp2 = new (C, 3) CmpINode( loop_head->incr(), lim2 );
|
|
952 register_new_node( cmp2, ctrl2 );
|
|
953 Node *bol2 = new (C, 2) BoolNode( cmp2, loop_end->test_trip() );
|
|
954 register_new_node( bol2, ctrl2 );
|
|
955 _igvn.hash_delete(loop_end);
|
|
956 loop_end->set_req(CountedLoopEndNode::TestValue, bol2);
|
|
957
|
|
958 // Step 3: Find the min-trip test guaranteed before a 'main' loop.
|
|
959 // Make it a 1-trip test (means at least 2 trips).
|
|
960 if( adjust_min_trip ) {
|
|
961 // Guard test uses an 'opaque' node which is not shared. Hence I
|
|
962 // can edit it's inputs directly. Hammer in the new limit for the
|
|
963 // minimum-trip guard.
|
|
964 assert( opaq->outcnt() == 1, "" );
|
|
965 _igvn.hash_delete(opaq);
|
|
966 opaq->set_req(1, lim2);
|
|
967 }
|
|
968
|
|
969 // ---------
|
|
970 // Step 4: Clone the loop body. Move it inside the loop. This loop body
|
|
971 // represents the odd iterations; since the loop trips an even number of
|
|
972 // times its backedge is never taken. Kill the backedge.
|
|
973 uint dd = dom_depth(loop_head);
|
|
974 clone_loop( loop, old_new, dd );
|
|
975
|
|
976 // Make backedges of the clone equal to backedges of the original.
|
|
977 // Make the fall-in from the original come from the fall-out of the clone.
|
|
978 for (DUIterator_Fast jmax, j = loop_head->fast_outs(jmax); j < jmax; j++) {
|
|
979 Node* phi = loop_head->fast_out(j);
|
|
980 if( phi->is_Phi() && phi->in(0) == loop_head && phi->outcnt() > 0 ) {
|
|
981 Node *newphi = old_new[phi->_idx];
|
|
982 _igvn.hash_delete( phi );
|
|
983 _igvn.hash_delete( newphi );
|
|
984
|
|
985 phi ->set_req(LoopNode:: EntryControl, newphi->in(LoopNode::LoopBackControl));
|
|
986 newphi->set_req(LoopNode::LoopBackControl, phi ->in(LoopNode::LoopBackControl));
|
|
987 phi ->set_req(LoopNode::LoopBackControl, C->top());
|
|
988 }
|
|
989 }
|
|
990 Node *clone_head = old_new[loop_head->_idx];
|
|
991 _igvn.hash_delete( clone_head );
|
|
992 loop_head ->set_req(LoopNode:: EntryControl, clone_head->in(LoopNode::LoopBackControl));
|
|
993 clone_head->set_req(LoopNode::LoopBackControl, loop_head ->in(LoopNode::LoopBackControl));
|
|
994 loop_head ->set_req(LoopNode::LoopBackControl, C->top());
|
|
995 loop->_head = clone_head; // New loop header
|
|
996
|
|
997 set_idom(loop_head, loop_head ->in(LoopNode::EntryControl), dd);
|
|
998 set_idom(clone_head, clone_head->in(LoopNode::EntryControl), dd);
|
|
999
|
|
1000 // Kill the clone's backedge
|
|
1001 Node *newcle = old_new[loop_end->_idx];
|
|
1002 _igvn.hash_delete( newcle );
|
|
1003 Node *one = _igvn.intcon(1);
|
|
1004 set_ctrl(one, C->root());
|
|
1005 newcle->set_req(1, one);
|
|
1006 // Force clone into same loop body
|
|
1007 uint max = loop->_body.size();
|
|
1008 for( uint k = 0; k < max; k++ ) {
|
|
1009 Node *old = loop->_body.at(k);
|
|
1010 Node *nnn = old_new[old->_idx];
|
|
1011 loop->_body.push(nnn);
|
|
1012 if (!has_ctrl(old))
|
|
1013 set_loop(nnn, loop);
|
|
1014 }
|
|
1015 }
|
|
1016
|
|
1017 //------------------------------do_maximally_unroll----------------------------
|
|
1018
|
|
1019 void PhaseIdealLoop::do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new ) {
|
|
1020 CountedLoopNode *cl = loop->_head->as_CountedLoop();
|
|
1021 assert( cl->trip_count() > 0, "");
|
|
1022
|
|
1023 // If loop is tripping an odd number of times, peel odd iteration
|
|
1024 if( (cl->trip_count() & 1) == 1 ) {
|
|
1025 do_peeling( loop, old_new );
|
|
1026 }
|
|
1027
|
|
1028 // Now its tripping an even number of times remaining. Double loop body.
|
|
1029 // Do not adjust pre-guards; they are not needed and do not exist.
|
|
1030 if( cl->trip_count() > 0 ) {
|
|
1031 do_unroll( loop, old_new, false );
|
|
1032 }
|
|
1033 }
|
|
1034
|
|
1035 //------------------------------dominates_backedge---------------------------------
|
|
1036 // Returns true if ctrl is executed on every complete iteration
|
|
1037 bool IdealLoopTree::dominates_backedge(Node* ctrl) {
|
|
1038 assert(ctrl->is_CFG(), "must be control");
|
|
1039 Node* backedge = _head->as_Loop()->in(LoopNode::LoopBackControl);
|
|
1040 return _phase->dom_lca_internal(ctrl, backedge) == ctrl;
|
|
1041 }
|
|
1042
|
|
1043 //------------------------------add_constraint---------------------------------
|
|
1044 // Constrain the main loop iterations so the condition:
|
|
1045 // scale_con * I + offset < limit
|
|
1046 // always holds true. That is, either increase the number of iterations in
|
|
1047 // the pre-loop or the post-loop until the condition holds true in the main
|
|
1048 // loop. Stride, scale, offset and limit are all loop invariant. Further,
|
|
1049 // stride and scale are constants (offset and limit often are).
|
|
1050 void PhaseIdealLoop::add_constraint( int stride_con, int scale_con, Node *offset, Node *limit, Node *pre_ctrl, Node **pre_limit, Node **main_limit ) {
|
|
1051
|
|
1052 // Compute "I :: (limit-offset)/scale_con"
|
|
1053 Node *con = new (C, 3) SubINode( limit, offset );
|
|
1054 register_new_node( con, pre_ctrl );
|
|
1055 Node *scale = _igvn.intcon(scale_con);
|
|
1056 set_ctrl(scale, C->root());
|
|
1057 Node *X = new (C, 3) DivINode( 0, con, scale );
|
|
1058 register_new_node( X, pre_ctrl );
|
|
1059
|
|
1060 // For positive stride, the pre-loop limit always uses a MAX function
|
|
1061 // and the main loop a MIN function. For negative stride these are
|
|
1062 // reversed.
|
|
1063
|
|
1064 // Also for positive stride*scale the affine function is increasing, so the
|
|
1065 // pre-loop must check for underflow and the post-loop for overflow.
|
|
1066 // Negative stride*scale reverses this; pre-loop checks for overflow and
|
|
1067 // post-loop for underflow.
|
|
1068 if( stride_con*scale_con > 0 ) {
|
|
1069 // Compute I < (limit-offset)/scale_con
|
|
1070 // Adjust main-loop last iteration to be MIN/MAX(main_loop,X)
|
|
1071 *main_limit = (stride_con > 0)
|
|
1072 ? (Node*)(new (C, 3) MinINode( *main_limit, X ))
|
|
1073 : (Node*)(new (C, 3) MaxINode( *main_limit, X ));
|
|
1074 register_new_node( *main_limit, pre_ctrl );
|
|
1075
|
|
1076 } else {
|
|
1077 // Compute (limit-offset)/scale_con + SGN(-scale_con) <= I
|
|
1078 // Add the negation of the main-loop constraint to the pre-loop.
|
|
1079 // See footnote [++] below for a derivation of the limit expression.
|
|
1080 Node *incr = _igvn.intcon(scale_con > 0 ? -1 : 1);
|
|
1081 set_ctrl(incr, C->root());
|
|
1082 Node *adj = new (C, 3) AddINode( X, incr );
|
|
1083 register_new_node( adj, pre_ctrl );
|
|
1084 *pre_limit = (scale_con > 0)
|
|
1085 ? (Node*)new (C, 3) MinINode( *pre_limit, adj )
|
|
1086 : (Node*)new (C, 3) MaxINode( *pre_limit, adj );
|
|
1087 register_new_node( *pre_limit, pre_ctrl );
|
|
1088
|
|
1089 // [++] Here's the algebra that justifies the pre-loop limit expression:
|
|
1090 //
|
|
1091 // NOT( scale_con * I + offset < limit )
|
|
1092 // ==
|
|
1093 // scale_con * I + offset >= limit
|
|
1094 // ==
|
|
1095 // SGN(scale_con) * I >= (limit-offset)/|scale_con|
|
|
1096 // ==
|
|
1097 // (limit-offset)/|scale_con| <= I * SGN(scale_con)
|
|
1098 // ==
|
|
1099 // (limit-offset)/|scale_con|-1 < I * SGN(scale_con)
|
|
1100 // ==
|
|
1101 // ( if (scale_con > 0) /*common case*/
|
|
1102 // (limit-offset)/scale_con - 1 < I
|
|
1103 // else
|
|
1104 // (limit-offset)/scale_con + 1 > I
|
|
1105 // )
|
|
1106 // ( if (scale_con > 0) /*common case*/
|
|
1107 // (limit-offset)/scale_con + SGN(-scale_con) < I
|
|
1108 // else
|
|
1109 // (limit-offset)/scale_con + SGN(-scale_con) > I
|
|
1110 }
|
|
1111 }
|
|
1112
|
|
1113
|
|
1114 //------------------------------is_scaled_iv---------------------------------
|
|
1115 // Return true if exp is a constant times an induction var
|
|
1116 bool PhaseIdealLoop::is_scaled_iv(Node* exp, Node* iv, int* p_scale) {
|
|
1117 if (exp == iv) {
|
|
1118 if (p_scale != NULL) {
|
|
1119 *p_scale = 1;
|
|
1120 }
|
|
1121 return true;
|
|
1122 }
|
|
1123 int opc = exp->Opcode();
|
|
1124 if (opc == Op_MulI) {
|
|
1125 if (exp->in(1) == iv && exp->in(2)->is_Con()) {
|
|
1126 if (p_scale != NULL) {
|
|
1127 *p_scale = exp->in(2)->get_int();
|
|
1128 }
|
|
1129 return true;
|
|
1130 }
|
|
1131 if (exp->in(2) == iv && exp->in(1)->is_Con()) {
|
|
1132 if (p_scale != NULL) {
|
|
1133 *p_scale = exp->in(1)->get_int();
|
|
1134 }
|
|
1135 return true;
|
|
1136 }
|
|
1137 } else if (opc == Op_LShiftI) {
|
|
1138 if (exp->in(1) == iv && exp->in(2)->is_Con()) {
|
|
1139 if (p_scale != NULL) {
|
|
1140 *p_scale = 1 << exp->in(2)->get_int();
|
|
1141 }
|
|
1142 return true;
|
|
1143 }
|
|
1144 }
|
|
1145 return false;
|
|
1146 }
|
|
1147
|
|
1148 //-----------------------------is_scaled_iv_plus_offset------------------------------
|
|
1149 // Return true if exp is a simple induction variable expression: k1*iv + (invar + k2)
|
|
1150 bool PhaseIdealLoop::is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth) {
|
|
1151 if (is_scaled_iv(exp, iv, p_scale)) {
|
|
1152 if (p_offset != NULL) {
|
|
1153 Node *zero = _igvn.intcon(0);
|
|
1154 set_ctrl(zero, C->root());
|
|
1155 *p_offset = zero;
|
|
1156 }
|
|
1157 return true;
|
|
1158 }
|
|
1159 int opc = exp->Opcode();
|
|
1160 if (opc == Op_AddI) {
|
|
1161 if (is_scaled_iv(exp->in(1), iv, p_scale)) {
|
|
1162 if (p_offset != NULL) {
|
|
1163 *p_offset = exp->in(2);
|
|
1164 }
|
|
1165 return true;
|
|
1166 }
|
|
1167 if (exp->in(2)->is_Con()) {
|
|
1168 Node* offset2 = NULL;
|
|
1169 if (depth < 2 &&
|
|
1170 is_scaled_iv_plus_offset(exp->in(1), iv, p_scale,
|
|
1171 p_offset != NULL ? &offset2 : NULL, depth+1)) {
|
|
1172 if (p_offset != NULL) {
|
|
1173 Node *ctrl_off2 = get_ctrl(offset2);
|
|
1174 Node* offset = new (C, 3) AddINode(offset2, exp->in(2));
|
|
1175 register_new_node(offset, ctrl_off2);
|
|
1176 *p_offset = offset;
|
|
1177 }
|
|
1178 return true;
|
|
1179 }
|
|
1180 }
|
|
1181 } else if (opc == Op_SubI) {
|
|
1182 if (is_scaled_iv(exp->in(1), iv, p_scale)) {
|
|
1183 if (p_offset != NULL) {
|
|
1184 Node *zero = _igvn.intcon(0);
|
|
1185 set_ctrl(zero, C->root());
|
|
1186 Node *ctrl_off = get_ctrl(exp->in(2));
|
|
1187 Node* offset = new (C, 3) SubINode(zero, exp->in(2));
|
|
1188 register_new_node(offset, ctrl_off);
|
|
1189 *p_offset = offset;
|
|
1190 }
|
|
1191 return true;
|
|
1192 }
|
|
1193 if (is_scaled_iv(exp->in(2), iv, p_scale)) {
|
|
1194 if (p_offset != NULL) {
|
|
1195 *p_scale *= -1;
|
|
1196 *p_offset = exp->in(1);
|
|
1197 }
|
|
1198 return true;
|
|
1199 }
|
|
1200 }
|
|
1201 return false;
|
|
1202 }
|
|
1203
|
|
1204 //------------------------------do_range_check---------------------------------
|
|
1205 // Eliminate range-checks and other trip-counter vs loop-invariant tests.
|
|
1206 void PhaseIdealLoop::do_range_check( IdealLoopTree *loop, Node_List &old_new ) {
|
|
1207 #ifndef PRODUCT
|
|
1208 if( PrintOpto && VerifyLoopOptimizations ) {
|
|
1209 tty->print("Range Check Elimination ");
|
|
1210 loop->dump_head();
|
|
1211 }
|
|
1212 #endif
|
|
1213 assert( RangeCheckElimination, "" );
|
|
1214 CountedLoopNode *cl = loop->_head->as_CountedLoop();
|
|
1215 assert( cl->is_main_loop(), "" );
|
|
1216
|
|
1217 // Find the trip counter; we are iteration splitting based on it
|
|
1218 Node *trip_counter = cl->phi();
|
|
1219 // Find the main loop limit; we will trim it's iterations
|
|
1220 // to not ever trip end tests
|
|
1221 Node *main_limit = cl->limit();
|
|
1222 // Find the pre-loop limit; we will expand it's iterations to
|
|
1223 // not ever trip low tests.
|
|
1224 Node *ctrl = cl->in(LoopNode::EntryControl);
|
|
1225 assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" );
|
|
1226 Node *iffm = ctrl->in(0);
|
|
1227 assert( iffm->Opcode() == Op_If, "" );
|
|
1228 Node *p_f = iffm->in(0);
|
|
1229 assert( p_f->Opcode() == Op_IfFalse, "" );
|
|
1230 CountedLoopEndNode *pre_end = p_f->in(0)->as_CountedLoopEnd();
|
|
1231 assert( pre_end->loopnode()->is_pre_loop(), "" );
|
|
1232 Node *pre_opaq1 = pre_end->limit();
|
|
1233 // Occasionally it's possible for a pre-loop Opaque1 node to be
|
|
1234 // optimized away and then another round of loop opts attempted.
|
|
1235 // We can not optimize this particular loop in that case.
|
|
1236 if( pre_opaq1->Opcode() != Op_Opaque1 )
|
|
1237 return;
|
|
1238 Opaque1Node *pre_opaq = (Opaque1Node*)pre_opaq1;
|
|
1239 Node *pre_limit = pre_opaq->in(1);
|
|
1240
|
|
1241 // Where do we put new limit calculations
|
|
1242 Node *pre_ctrl = pre_end->loopnode()->in(LoopNode::EntryControl);
|
|
1243
|
|
1244 // Ensure the original loop limit is available from the
|
|
1245 // pre-loop Opaque1 node.
|
|
1246 Node *orig_limit = pre_opaq->original_loop_limit();
|
|
1247 if( orig_limit == NULL || _igvn.type(orig_limit) == Type::TOP )
|
|
1248 return;
|
|
1249
|
|
1250 // Need to find the main-loop zero-trip guard
|
|
1251 Node *bolzm = iffm->in(1);
|
|
1252 assert( bolzm->Opcode() == Op_Bool, "" );
|
|
1253 Node *cmpzm = bolzm->in(1);
|
|
1254 assert( cmpzm->is_Cmp(), "" );
|
|
1255 Node *opqzm = cmpzm->in(2);
|
|
1256 if( opqzm->Opcode() != Op_Opaque1 )
|
|
1257 return;
|
|
1258 assert( opqzm->in(1) == main_limit, "do not understand situation" );
|
|
1259
|
|
1260 // Must know if its a count-up or count-down loop
|
|
1261
|
|
1262 // protect against stride not being a constant
|
|
1263 if ( !cl->stride_is_con() ) {
|
|
1264 return;
|
|
1265 }
|
|
1266 int stride_con = cl->stride_con();
|
|
1267 Node *zero = _igvn.intcon(0);
|
|
1268 Node *one = _igvn.intcon(1);
|
|
1269 set_ctrl(zero, C->root());
|
|
1270 set_ctrl(one, C->root());
|
|
1271
|
|
1272 // Range checks that do not dominate the loop backedge (ie.
|
|
1273 // conditionally executed) can lengthen the pre loop limit beyond
|
|
1274 // the original loop limit. To prevent this, the pre limit is
|
|
1275 // (for stride > 0) MINed with the original loop limit (MAXed
|
|
1276 // stride < 0) when some range_check (rc) is conditionally
|
|
1277 // executed.
|
|
1278 bool conditional_rc = false;
|
|
1279
|
|
1280 // Check loop body for tests of trip-counter plus loop-invariant vs
|
|
1281 // loop-invariant.
|
|
1282 for( uint i = 0; i < loop->_body.size(); i++ ) {
|
|
1283 Node *iff = loop->_body[i];
|
|
1284 if( iff->Opcode() == Op_If ) { // Test?
|
|
1285
|
|
1286 // Test is an IfNode, has 2 projections. If BOTH are in the loop
|
|
1287 // we need loop unswitching instead of iteration splitting.
|
|
1288 Node *exit = loop->is_loop_exit(iff);
|
|
1289 if( !exit ) continue;
|
|
1290 int flip = (exit->Opcode() == Op_IfTrue) ? 1 : 0;
|
|
1291
|
|
1292 // Get boolean condition to test
|
|
1293 Node *i1 = iff->in(1);
|
|
1294 if( !i1->is_Bool() ) continue;
|
|
1295 BoolNode *bol = i1->as_Bool();
|
|
1296 BoolTest b_test = bol->_test;
|
|
1297 // Flip sense of test if exit condition is flipped
|
|
1298 if( flip )
|
|
1299 b_test = b_test.negate();
|
|
1300
|
|
1301 // Get compare
|
|
1302 Node *cmp = bol->in(1);
|
|
1303
|
|
1304 // Look for trip_counter + offset vs limit
|
|
1305 Node *rc_exp = cmp->in(1);
|
|
1306 Node *limit = cmp->in(2);
|
|
1307 jint scale_con= 1; // Assume trip counter not scaled
|
|
1308
|
|
1309 Node *limit_c = get_ctrl(limit);
|
|
1310 if( loop->is_member(get_loop(limit_c) ) ) {
|
|
1311 // Compare might have operands swapped; commute them
|
|
1312 b_test = b_test.commute();
|
|
1313 rc_exp = cmp->in(2);
|
|
1314 limit = cmp->in(1);
|
|
1315 limit_c = get_ctrl(limit);
|
|
1316 if( loop->is_member(get_loop(limit_c) ) )
|
|
1317 continue; // Both inputs are loop varying; cannot RCE
|
|
1318 }
|
|
1319 // Here we know 'limit' is loop invariant
|
|
1320
|
|
1321 // 'limit' maybe pinned below the zero trip test (probably from a
|
|
1322 // previous round of rce), in which case, it can't be used in the
|
|
1323 // zero trip test expression which must occur before the zero test's if.
|
|
1324 if( limit_c == ctrl ) {
|
|
1325 continue; // Don't rce this check but continue looking for other candidates.
|
|
1326 }
|
|
1327
|
|
1328 // Check for scaled induction variable plus an offset
|
|
1329 Node *offset = NULL;
|
|
1330
|
|
1331 if (!is_scaled_iv_plus_offset(rc_exp, trip_counter, &scale_con, &offset)) {
|
|
1332 continue;
|
|
1333 }
|
|
1334
|
|
1335 Node *offset_c = get_ctrl(offset);
|
|
1336 if( loop->is_member( get_loop(offset_c) ) )
|
|
1337 continue; // Offset is not really loop invariant
|
|
1338 // Here we know 'offset' is loop invariant.
|
|
1339
|
|
1340 // As above for the 'limit', the 'offset' maybe pinned below the
|
|
1341 // zero trip test.
|
|
1342 if( offset_c == ctrl ) {
|
|
1343 continue; // Don't rce this check but continue looking for other candidates.
|
|
1344 }
|
|
1345
|
|
1346 // At this point we have the expression as:
|
|
1347 // scale_con * trip_counter + offset :: limit
|
|
1348 // where scale_con, offset and limit are loop invariant. Trip_counter
|
|
1349 // monotonically increases by stride_con, a constant. Both (or either)
|
|
1350 // stride_con and scale_con can be negative which will flip about the
|
|
1351 // sense of the test.
|
|
1352
|
|
1353 // Adjust pre and main loop limits to guard the correct iteration set
|
|
1354 if( cmp->Opcode() == Op_CmpU ) {// Unsigned compare is really 2 tests
|
|
1355 if( b_test._test == BoolTest::lt ) { // Range checks always use lt
|
|
1356 // The overflow limit: scale*I+offset < limit
|
|
1357 add_constraint( stride_con, scale_con, offset, limit, pre_ctrl, &pre_limit, &main_limit );
|
|
1358 // The underflow limit: 0 <= scale*I+offset.
|
|
1359 // Some math yields: -scale*I-(offset+1) < 0
|
|
1360 Node *plus_one = new (C, 3) AddINode( offset, one );
|
|
1361 register_new_node( plus_one, pre_ctrl );
|
|
1362 Node *neg_offset = new (C, 3) SubINode( zero, plus_one );
|
|
1363 register_new_node( neg_offset, pre_ctrl );
|
|
1364 add_constraint( stride_con, -scale_con, neg_offset, zero, pre_ctrl, &pre_limit, &main_limit );
|
|
1365 if (!conditional_rc) {
|
|
1366 conditional_rc = !loop->dominates_backedge(iff);
|
|
1367 }
|
|
1368 } else {
|
|
1369 #ifndef PRODUCT
|
|
1370 if( PrintOpto )
|
|
1371 tty->print_cr("missed RCE opportunity");
|
|
1372 #endif
|
|
1373 continue; // In release mode, ignore it
|
|
1374 }
|
|
1375 } else { // Otherwise work on normal compares
|
|
1376 switch( b_test._test ) {
|
|
1377 case BoolTest::ge: // Convert X >= Y to -X <= -Y
|
|
1378 scale_con = -scale_con;
|
|
1379 offset = new (C, 3) SubINode( zero, offset );
|
|
1380 register_new_node( offset, pre_ctrl );
|
|
1381 limit = new (C, 3) SubINode( zero, limit );
|
|
1382 register_new_node( limit, pre_ctrl );
|
|
1383 // Fall into LE case
|
|
1384 case BoolTest::le: // Convert X <= Y to X < Y+1
|
|
1385 limit = new (C, 3) AddINode( limit, one );
|
|
1386 register_new_node( limit, pre_ctrl );
|
|
1387 // Fall into LT case
|
|
1388 case BoolTest::lt:
|
|
1389 add_constraint( stride_con, scale_con, offset, limit, pre_ctrl, &pre_limit, &main_limit );
|
|
1390 if (!conditional_rc) {
|
|
1391 conditional_rc = !loop->dominates_backedge(iff);
|
|
1392 }
|
|
1393 break;
|
|
1394 default:
|
|
1395 #ifndef PRODUCT
|
|
1396 if( PrintOpto )
|
|
1397 tty->print_cr("missed RCE opportunity");
|
|
1398 #endif
|
|
1399 continue; // Unhandled case
|
|
1400 }
|
|
1401 }
|
|
1402
|
|
1403 // Kill the eliminated test
|
|
1404 C->set_major_progress();
|
|
1405 Node *kill_con = _igvn.intcon( 1-flip );
|
|
1406 set_ctrl(kill_con, C->root());
|
|
1407 _igvn.hash_delete(iff);
|
|
1408 iff->set_req(1, kill_con);
|
|
1409 _igvn._worklist.push(iff);
|
|
1410 // Find surviving projection
|
|
1411 assert(iff->is_If(), "");
|
|
1412 ProjNode* dp = ((IfNode*)iff)->proj_out(1-flip);
|
|
1413 // Find loads off the surviving projection; remove their control edge
|
|
1414 for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) {
|
|
1415 Node* cd = dp->fast_out(i); // Control-dependent node
|
|
1416 if( cd->is_Load() ) { // Loads can now float around in the loop
|
|
1417 _igvn.hash_delete(cd);
|
|
1418 // Allow the load to float around in the loop, or before it
|
|
1419 // but NOT before the pre-loop.
|
|
1420 cd->set_req(0, ctrl); // ctrl, not NULL
|
|
1421 _igvn._worklist.push(cd);
|
|
1422 --i;
|
|
1423 --imax;
|
|
1424 }
|
|
1425 }
|
|
1426
|
|
1427 } // End of is IF
|
|
1428
|
|
1429 }
|
|
1430
|
|
1431 // Update loop limits
|
|
1432 if (conditional_rc) {
|
|
1433 pre_limit = (stride_con > 0) ? (Node*)new (C,3) MinINode(pre_limit, orig_limit)
|
|
1434 : (Node*)new (C,3) MaxINode(pre_limit, orig_limit);
|
|
1435 register_new_node(pre_limit, pre_ctrl);
|
|
1436 }
|
|
1437 _igvn.hash_delete(pre_opaq);
|
|
1438 pre_opaq->set_req(1, pre_limit);
|
|
1439
|
|
1440 // Note:: we are making the main loop limit no longer precise;
|
|
1441 // need to round up based on stride.
|
|
1442 if( stride_con != 1 && stride_con != -1 ) { // Cutout for common case
|
|
1443 // "Standard" round-up logic: ([main_limit-init+(y-1)]/y)*y+init
|
|
1444 // Hopefully, compiler will optimize for powers of 2.
|
|
1445 Node *ctrl = get_ctrl(main_limit);
|
|
1446 Node *stride = cl->stride();
|
|
1447 Node *init = cl->init_trip();
|
|
1448 Node *span = new (C, 3) SubINode(main_limit,init);
|
|
1449 register_new_node(span,ctrl);
|
|
1450 Node *rndup = _igvn.intcon(stride_con + ((stride_con>0)?-1:1));
|
|
1451 Node *add = new (C, 3) AddINode(span,rndup);
|
|
1452 register_new_node(add,ctrl);
|
|
1453 Node *div = new (C, 3) DivINode(0,add,stride);
|
|
1454 register_new_node(div,ctrl);
|
|
1455 Node *mul = new (C, 3) MulINode(div,stride);
|
|
1456 register_new_node(mul,ctrl);
|
|
1457 Node *newlim = new (C, 3) AddINode(mul,init);
|
|
1458 register_new_node(newlim,ctrl);
|
|
1459 main_limit = newlim;
|
|
1460 }
|
|
1461
|
|
1462 Node *main_cle = cl->loopexit();
|
|
1463 Node *main_bol = main_cle->in(1);
|
|
1464 // Hacking loop bounds; need private copies of exit test
|
|
1465 if( main_bol->outcnt() > 1 ) {// BoolNode shared?
|
|
1466 _igvn.hash_delete(main_cle);
|
|
1467 main_bol = main_bol->clone();// Clone a private BoolNode
|
|
1468 register_new_node( main_bol, main_cle->in(0) );
|
|
1469 main_cle->set_req(1,main_bol);
|
|
1470 }
|
|
1471 Node *main_cmp = main_bol->in(1);
|
|
1472 if( main_cmp->outcnt() > 1 ) { // CmpNode shared?
|
|
1473 _igvn.hash_delete(main_bol);
|
|
1474 main_cmp = main_cmp->clone();// Clone a private CmpNode
|
|
1475 register_new_node( main_cmp, main_cle->in(0) );
|
|
1476 main_bol->set_req(1,main_cmp);
|
|
1477 }
|
|
1478 // Hack the now-private loop bounds
|
|
1479 _igvn.hash_delete(main_cmp);
|
|
1480 main_cmp->set_req(2, main_limit);
|
|
1481 _igvn._worklist.push(main_cmp);
|
|
1482 // The OpaqueNode is unshared by design
|
|
1483 _igvn.hash_delete(opqzm);
|
|
1484 assert( opqzm->outcnt() == 1, "cannot hack shared node" );
|
|
1485 opqzm->set_req(1,main_limit);
|
|
1486 _igvn._worklist.push(opqzm);
|
|
1487 }
|
|
1488
|
|
1489 //------------------------------DCE_loop_body----------------------------------
|
|
1490 // Remove simplistic dead code from loop body
|
|
1491 void IdealLoopTree::DCE_loop_body() {
|
|
1492 for( uint i = 0; i < _body.size(); i++ )
|
|
1493 if( _body.at(i)->outcnt() == 0 )
|
|
1494 _body.map( i--, _body.pop() );
|
|
1495 }
|
|
1496
|
|
1497
|
|
1498 //------------------------------adjust_loop_exit_prob--------------------------
|
|
1499 // Look for loop-exit tests with the 50/50 (or worse) guesses from the parsing stage.
|
|
1500 // Replace with a 1-in-10 exit guess.
|
|
1501 void IdealLoopTree::adjust_loop_exit_prob( PhaseIdealLoop *phase ) {
|
|
1502 Node *test = tail();
|
|
1503 while( test != _head ) {
|
|
1504 uint top = test->Opcode();
|
|
1505 if( top == Op_IfTrue || top == Op_IfFalse ) {
|
|
1506 int test_con = ((ProjNode*)test)->_con;
|
|
1507 assert(top == (uint)(test_con? Op_IfTrue: Op_IfFalse), "sanity");
|
|
1508 IfNode *iff = test->in(0)->as_If();
|
|
1509 if( iff->outcnt() == 2 ) { // Ignore dead tests
|
|
1510 Node *bol = iff->in(1);
|
|
1511 if( bol && bol->req() > 1 && bol->in(1) &&
|
|
1512 ((bol->in(1)->Opcode() == Op_StorePConditional ) ||
|
|
1513 (bol->in(1)->Opcode() == Op_StoreLConditional ) ||
|
|
1514 (bol->in(1)->Opcode() == Op_CompareAndSwapI ) ||
|
|
1515 (bol->in(1)->Opcode() == Op_CompareAndSwapL ) ||
|
|
1516 (bol->in(1)->Opcode() == Op_CompareAndSwapP )))
|
|
1517 return; // Allocation loops RARELY take backedge
|
|
1518 // Find the OTHER exit path from the IF
|
|
1519 Node* ex = iff->proj_out(1-test_con);
|
|
1520 float p = iff->_prob;
|
|
1521 if( !phase->is_member( this, ex ) && iff->_fcnt == COUNT_UNKNOWN ) {
|
|
1522 if( top == Op_IfTrue ) {
|
|
1523 if( p < (PROB_FAIR + PROB_UNLIKELY_MAG(3))) {
|
|
1524 iff->_prob = PROB_STATIC_FREQUENT;
|
|
1525 }
|
|
1526 } else {
|
|
1527 if( p > (PROB_FAIR - PROB_UNLIKELY_MAG(3))) {
|
|
1528 iff->_prob = PROB_STATIC_INFREQUENT;
|
|
1529 }
|
|
1530 }
|
|
1531 }
|
|
1532 }
|
|
1533 }
|
|
1534 test = phase->idom(test);
|
|
1535 }
|
|
1536 }
|
|
1537
|
|
1538
|
|
1539 //------------------------------policy_do_remove_empty_loop--------------------
|
|
1540 // Micro-benchmark spamming. Policy is to always remove empty loops.
|
|
1541 // The 'DO' part is to replace the trip counter with the value it will
|
|
1542 // have on the last iteration. This will break the loop.
|
|
1543 bool IdealLoopTree::policy_do_remove_empty_loop( PhaseIdealLoop *phase ) {
|
|
1544 // Minimum size must be empty loop
|
|
1545 if( _body.size() > 7/*number of nodes in an empty loop*/ ) return false;
|
|
1546
|
|
1547 if( !_head->is_CountedLoop() ) return false; // Dead loop
|
|
1548 CountedLoopNode *cl = _head->as_CountedLoop();
|
|
1549 if( !cl->loopexit() ) return false; // Malformed loop
|
|
1550 if( !phase->is_member(this,phase->get_ctrl(cl->loopexit()->in(CountedLoopEndNode::TestValue)) ) )
|
|
1551 return false; // Infinite loop
|
|
1552 #ifndef PRODUCT
|
|
1553 if( PrintOpto )
|
|
1554 tty->print_cr("Removing empty loop");
|
|
1555 #endif
|
|
1556 #ifdef ASSERT
|
|
1557 // Ensure only one phi which is the iv.
|
|
1558 Node* iv = NULL;
|
|
1559 for (DUIterator_Fast imax, i = cl->fast_outs(imax); i < imax; i++) {
|
|
1560 Node* n = cl->fast_out(i);
|
|
1561 if (n->Opcode() == Op_Phi) {
|
|
1562 assert(iv == NULL, "Too many phis" );
|
|
1563 iv = n;
|
|
1564 }
|
|
1565 }
|
|
1566 assert(iv == cl->phi(), "Wrong phi" );
|
|
1567 #endif
|
|
1568 // Replace the phi at loop head with the final value of the last
|
|
1569 // iteration. Then the CountedLoopEnd will collapse (backedge never
|
|
1570 // taken) and all loop-invariant uses of the exit values will be correct.
|
|
1571 Node *phi = cl->phi();
|
|
1572 Node *final = new (phase->C, 3) SubINode( cl->limit(), cl->stride() );
|
|
1573 phase->register_new_node(final,cl->in(LoopNode::EntryControl));
|
|
1574 phase->_igvn.hash_delete(phi);
|
|
1575 phase->_igvn.subsume_node(phi,final);
|
|
1576 phase->C->set_major_progress();
|
|
1577 return true;
|
|
1578 }
|
|
1579
|
|
1580
|
|
1581 //=============================================================================
|
|
1582 //------------------------------iteration_split_impl---------------------------
|
|
1583 void IdealLoopTree::iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new ) {
|
|
1584 // Check and remove empty loops (spam micro-benchmarks)
|
|
1585 if( policy_do_remove_empty_loop(phase) )
|
|
1586 return; // Here we removed an empty loop
|
|
1587
|
|
1588 bool should_peel = policy_peeling(phase); // Should we peel?
|
|
1589
|
|
1590 bool should_unswitch = policy_unswitching(phase);
|
|
1591
|
|
1592 // Non-counted loops may be peeled; exactly 1 iteration is peeled.
|
|
1593 // This removes loop-invariant tests (usually null checks).
|
|
1594 if( !_head->is_CountedLoop() ) { // Non-counted loop
|
|
1595 if (PartialPeelLoop && phase->partial_peel(this, old_new)) {
|
|
1596 return;
|
|
1597 }
|
|
1598 if( should_peel ) { // Should we peel?
|
|
1599 #ifndef PRODUCT
|
|
1600 if (PrintOpto) tty->print_cr("should_peel");
|
|
1601 #endif
|
|
1602 phase->do_peeling(this,old_new);
|
|
1603 } else if( should_unswitch ) {
|
|
1604 phase->do_unswitching(this, old_new);
|
|
1605 }
|
|
1606 return;
|
|
1607 }
|
|
1608 CountedLoopNode *cl = _head->as_CountedLoop();
|
|
1609
|
|
1610 if( !cl->loopexit() ) return; // Ignore various kinds of broken loops
|
|
1611
|
|
1612 // Do nothing special to pre- and post- loops
|
|
1613 if( cl->is_pre_loop() || cl->is_post_loop() ) return;
|
|
1614
|
|
1615 // Compute loop trip count from profile data
|
|
1616 compute_profile_trip_cnt(phase);
|
|
1617
|
|
1618 // Before attempting fancy unrolling, RCE or alignment, see if we want
|
|
1619 // to completely unroll this loop or do loop unswitching.
|
|
1620 if( cl->is_normal_loop() ) {
|
|
1621 bool should_maximally_unroll = policy_maximally_unroll(phase);
|
|
1622 if( should_maximally_unroll ) {
|
|
1623 // Here we did some unrolling and peeling. Eventually we will
|
|
1624 // completely unroll this loop and it will no longer be a loop.
|
|
1625 phase->do_maximally_unroll(this,old_new);
|
|
1626 return;
|
|
1627 }
|
|
1628 if (should_unswitch) {
|
|
1629 phase->do_unswitching(this, old_new);
|
|
1630 return;
|
|
1631 }
|
|
1632 }
|
|
1633
|
|
1634
|
|
1635 // Counted loops may be peeled, may need some iterations run up
|
|
1636 // front for RCE, and may want to align loop refs to a cache
|
|
1637 // line. Thus we clone a full loop up front whose trip count is
|
|
1638 // at least 1 (if peeling), but may be several more.
|
|
1639
|
|
1640 // The main loop will start cache-line aligned with at least 1
|
|
1641 // iteration of the unrolled body (zero-trip test required) and
|
|
1642 // will have some range checks removed.
|
|
1643
|
|
1644 // A post-loop will finish any odd iterations (leftover after
|
|
1645 // unrolling), plus any needed for RCE purposes.
|
|
1646
|
|
1647 bool should_unroll = policy_unroll(phase);
|
|
1648
|
|
1649 bool should_rce = policy_range_check(phase);
|
|
1650
|
|
1651 bool should_align = policy_align(phase);
|
|
1652
|
|
1653 // If not RCE'ing (iteration splitting) or Aligning, then we do not
|
|
1654 // need a pre-loop. We may still need to peel an initial iteration but
|
|
1655 // we will not be needing an unknown number of pre-iterations.
|
|
1656 //
|
|
1657 // Basically, if may_rce_align reports FALSE first time through,
|
|
1658 // we will not be able to later do RCE or Aligning on this loop.
|
|
1659 bool may_rce_align = !policy_peel_only(phase) || should_rce || should_align;
|
|
1660
|
|
1661 // If we have any of these conditions (RCE, alignment, unrolling) met, then
|
|
1662 // we switch to the pre-/main-/post-loop model. This model also covers
|
|
1663 // peeling.
|
|
1664 if( should_rce || should_align || should_unroll ) {
|
|
1665 if( cl->is_normal_loop() ) // Convert to 'pre/main/post' loops
|
|
1666 phase->insert_pre_post_loops(this,old_new, !may_rce_align);
|
|
1667
|
|
1668 // Adjust the pre- and main-loop limits to let the pre and post loops run
|
|
1669 // with full checks, but the main-loop with no checks. Remove said
|
|
1670 // checks from the main body.
|
|
1671 if( should_rce )
|
|
1672 phase->do_range_check(this,old_new);
|
|
1673
|
|
1674 // Double loop body for unrolling. Adjust the minimum-trip test (will do
|
|
1675 // twice as many iterations as before) and the main body limit (only do
|
|
1676 // an even number of trips). If we are peeling, we might enable some RCE
|
|
1677 // and we'd rather unroll the post-RCE'd loop SO... do not unroll if
|
|
1678 // peeling.
|
|
1679 if( should_unroll && !should_peel )
|
|
1680 phase->do_unroll(this,old_new, true);
|
|
1681
|
|
1682 // Adjust the pre-loop limits to align the main body
|
|
1683 // iterations.
|
|
1684 if( should_align )
|
|
1685 Unimplemented();
|
|
1686
|
|
1687 } else { // Else we have an unchanged counted loop
|
|
1688 if( should_peel ) // Might want to peel but do nothing else
|
|
1689 phase->do_peeling(this,old_new);
|
|
1690 }
|
|
1691 }
|
|
1692
|
|
1693
|
|
1694 //=============================================================================
|
|
1695 //------------------------------iteration_split--------------------------------
|
|
1696 void IdealLoopTree::iteration_split( PhaseIdealLoop *phase, Node_List &old_new ) {
|
|
1697 // Recursively iteration split nested loops
|
|
1698 if( _child ) _child->iteration_split( phase, old_new );
|
|
1699
|
|
1700 // Clean out prior deadwood
|
|
1701 DCE_loop_body();
|
|
1702
|
|
1703
|
|
1704 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage.
|
|
1705 // Replace with a 1-in-10 exit guess.
|
|
1706 if( _parent /*not the root loop*/ &&
|
|
1707 !_irreducible &&
|
|
1708 // Also ignore the occasional dead backedge
|
|
1709 !tail()->is_top() ) {
|
|
1710 adjust_loop_exit_prob(phase);
|
|
1711 }
|
|
1712
|
|
1713
|
|
1714 // Gate unrolling, RCE and peeling efforts.
|
|
1715 if( !_child && // If not an inner loop, do not split
|
|
1716 !_irreducible &&
|
|
1717 !tail()->is_top() ) { // Also ignore the occasional dead backedge
|
|
1718 if (!_has_call) {
|
|
1719 iteration_split_impl( phase, old_new );
|
|
1720 } else if (policy_unswitching(phase)) {
|
|
1721 phase->do_unswitching(this, old_new);
|
|
1722 }
|
|
1723 }
|
|
1724
|
|
1725 // Minor offset re-organization to remove loop-fallout uses of
|
|
1726 // trip counter.
|
|
1727 if( _head->is_CountedLoop() ) phase->reorg_offsets( this );
|
|
1728 if( _next ) _next->iteration_split( phase, old_new );
|
|
1729 }
|