Mercurial > hg > truffle
annotate src/share/vm/opto/loopnode.hpp @ 7788:dbbe8ce3bfa5
commands: accept 'server0' as a vm, accept version number which have '-...' suffix (-internal, -ea...)
author | Gilles Duboscq <duboscq@ssw.jku.at> |
---|---|
date | Wed, 13 Feb 2013 18:33:54 +0100 |
parents | b9a9ed0f8eeb |
children | 8b3da8d14c93 |
rev | line source |
---|---|
0 | 1 /* |
6842
b9a9ed0f8eeb
7197424: update copyright year to match last edit in jdk8 hotspot repository
mikael
parents:
6636
diff
changeset
|
2 * Copyright (c) 1998, 2012, Oracle and/or its affiliates. All rights reserved. |
0 | 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 * | |
1552
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
1303
diff
changeset
|
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA |
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
1303
diff
changeset
|
20 * or visit www.oracle.com if you need additional information or have any |
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
1303
diff
changeset
|
21 * questions. |
0 | 22 * |
23 */ | |
24 | |
1972 | 25 #ifndef SHARE_VM_OPTO_LOOPNODE_HPP |
26 #define SHARE_VM_OPTO_LOOPNODE_HPP | |
27 | |
28 #include "opto/cfgnode.hpp" | |
29 #include "opto/multnode.hpp" | |
30 #include "opto/phaseX.hpp" | |
31 #include "opto/subnode.hpp" | |
32 #include "opto/type.hpp" | |
33 | |
0 | 34 class CmpNode; |
35 class CountedLoopEndNode; | |
36 class CountedLoopNode; | |
37 class IdealLoopTree; | |
38 class LoopNode; | |
39 class Node; | |
40 class PhaseIdealLoop; | |
41 class VectorSet; | |
1172 | 42 class Invariance; |
0 | 43 struct small_cache; |
44 | |
45 // | |
46 // I D E A L I Z E D L O O P S | |
47 // | |
48 // Idealized loops are the set of loops I perform more interesting | |
49 // transformations on, beyond simple hoisting. | |
50 | |
51 //------------------------------LoopNode--------------------------------------- | |
52 // Simple loop header. Fall in path on left, loop-back path on right. | |
53 class LoopNode : public RegionNode { | |
54 // Size is bigger to hold the flags. However, the flags do not change | |
55 // the semantics so it does not appear in the hash & cmp functions. | |
56 virtual uint size_of() const { return sizeof(*this); } | |
57 protected: | |
58 short _loop_flags; | |
59 // Names for flag bitfields | |
2465 | 60 enum { Normal=0, Pre=1, Main=2, Post=3, PreMainPostFlagsMask=3, |
61 MainHasNoPreLoop=4, | |
62 HasExactTripCount=8, | |
63 InnerLoop=16, | |
64 PartialPeelLoop=32, | |
65 PartialPeelFailed=64 }; | |
0 | 66 char _unswitch_count; |
67 enum { _unswitch_max=3 }; | |
68 | |
69 public: | |
70 // Names for edge indices | |
71 enum { Self=0, EntryControl, LoopBackControl }; | |
72 | |
2465 | 73 int is_inner_loop() const { return _loop_flags & InnerLoop; } |
74 void set_inner_loop() { _loop_flags |= InnerLoop; } | |
0 | 75 |
2465 | 76 int is_partial_peel_loop() const { return _loop_flags & PartialPeelLoop; } |
77 void set_partial_peel_loop() { _loop_flags |= PartialPeelLoop; } | |
78 int partial_peel_has_failed() const { return _loop_flags & PartialPeelFailed; } | |
79 void mark_partial_peel_failed() { _loop_flags |= PartialPeelFailed; } | |
0 | 80 |
81 int unswitch_max() { return _unswitch_max; } | |
82 int unswitch_count() { return _unswitch_count; } | |
83 void set_unswitch_count(int val) { | |
84 assert (val <= unswitch_max(), "too many unswitches"); | |
85 _unswitch_count = val; | |
86 } | |
87 | |
88 LoopNode( Node *entry, Node *backedge ) : RegionNode(3), _loop_flags(0), _unswitch_count(0) { | |
89 init_class_id(Class_Loop); | |
90 init_req(EntryControl, entry); | |
91 init_req(LoopBackControl, backedge); | |
92 } | |
93 | |
94 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape); | |
95 virtual int Opcode() const; | |
96 bool can_be_counted_loop(PhaseTransform* phase) const { | |
97 return req() == 3 && in(0) != NULL && | |
98 in(1) != NULL && phase->type(in(1)) != Type::TOP && | |
99 in(2) != NULL && phase->type(in(2)) != Type::TOP; | |
100 } | |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
2248
diff
changeset
|
101 bool is_valid_counted_loop() const; |
0 | 102 #ifndef PRODUCT |
103 virtual void dump_spec(outputStream *st) const; | |
104 #endif | |
105 }; | |
106 | |
107 //------------------------------Counted Loops---------------------------------- | |
108 // Counted loops are all trip-counted loops, with exactly 1 trip-counter exit | |
109 // path (and maybe some other exit paths). The trip-counter exit is always | |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
2248
diff
changeset
|
110 // last in the loop. The trip-counter have to stride by a constant; |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
2248
diff
changeset
|
111 // the exit value is also loop invariant. |
0 | 112 |
113 // CountedLoopNodes and CountedLoopEndNodes come in matched pairs. The | |
114 // CountedLoopNode has the incoming loop control and the loop-back-control | |
115 // which is always the IfTrue before the matching CountedLoopEndNode. The | |
116 // CountedLoopEndNode has an incoming control (possibly not the | |
117 // CountedLoopNode if there is control flow in the loop), the post-increment | |
118 // trip-counter value, and the limit. The trip-counter value is always of | |
119 // the form (Op old-trip-counter stride). The old-trip-counter is produced | |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
2248
diff
changeset
|
120 // by a Phi connected to the CountedLoopNode. The stride is constant. |
0 | 121 // The Op is any commutable opcode, including Add, Mul, Xor. The |
122 // CountedLoopEndNode also takes in the loop-invariant limit value. | |
123 | |
124 // From a CountedLoopNode I can reach the matching CountedLoopEndNode via the | |
125 // loop-back control. From CountedLoopEndNodes I can reach CountedLoopNodes | |
126 // via the old-trip-counter from the Op node. | |
127 | |
128 //------------------------------CountedLoopNode-------------------------------- | |
129 // CountedLoopNodes head simple counted loops. CountedLoopNodes have as | |
130 // inputs the incoming loop-start control and the loop-back control, so they | |
131 // act like RegionNodes. They also take in the initial trip counter, the | |
132 // loop-invariant stride and the loop-invariant limit value. CountedLoopNodes | |
133 // produce a loop-body control and the trip counter value. Since | |
134 // CountedLoopNodes behave like RegionNodes I still have a standard CFG model. | |
135 | |
136 class CountedLoopNode : public LoopNode { | |
137 // Size is bigger to hold _main_idx. However, _main_idx does not change | |
138 // the semantics so it does not appear in the hash & cmp functions. | |
139 virtual uint size_of() const { return sizeof(*this); } | |
140 | |
141 // For Pre- and Post-loops during debugging ONLY, this holds the index of | |
142 // the Main CountedLoop. Used to assert that we understand the graph shape. | |
143 node_idx_t _main_idx; | |
144 | |
2465 | 145 // Known trip count calculated by compute_exact_trip_count() |
146 uint _trip_count; | |
0 | 147 |
148 // Expected trip count from profile data | |
149 float _profile_trip_cnt; | |
150 | |
151 // Log2 of original loop bodies in unrolled loop | |
152 int _unrolled_count_log2; | |
153 | |
154 // Node count prior to last unrolling - used to decide if | |
155 // unroll,optimize,unroll,optimize,... is making progress | |
156 int _node_count_before_unroll; | |
157 | |
158 public: | |
159 CountedLoopNode( Node *entry, Node *backedge ) | |
2465 | 160 : LoopNode(entry, backedge), _main_idx(0), _trip_count(max_juint), |
0 | 161 _profile_trip_cnt(COUNT_UNKNOWN), _unrolled_count_log2(0), |
162 _node_count_before_unroll(0) { | |
163 init_class_id(Class_CountedLoop); | |
164 // Initialize _trip_count to the largest possible value. | |
165 // Will be reset (lower) if the loop's trip count is known. | |
166 } | |
167 | |
168 virtual int Opcode() const; | |
169 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape); | |
170 | |
171 Node *init_control() const { return in(EntryControl); } | |
172 Node *back_control() const { return in(LoopBackControl); } | |
173 CountedLoopEndNode *loopexit() const; | |
174 Node *init_trip() const; | |
175 Node *stride() const; | |
176 int stride_con() const; | |
177 bool stride_is_con() const; | |
178 Node *limit() const; | |
179 Node *incr() const; | |
180 Node *phi() const; | |
181 | |
182 // Match increment with optional truncation | |
183 static Node* match_incr_with_optional_truncation(Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type); | |
184 | |
185 // A 'main' loop has a pre-loop and a post-loop. The 'main' loop | |
186 // can run short a few iterations and may start a few iterations in. | |
187 // It will be RCE'd and unrolled and aligned. | |
188 | |
189 // A following 'post' loop will run any remaining iterations. Used | |
190 // during Range Check Elimination, the 'post' loop will do any final | |
191 // iterations with full checks. Also used by Loop Unrolling, where | |
192 // the 'post' loop will do any epilog iterations needed. Basically, | |
193 // a 'post' loop can not profitably be further unrolled or RCE'd. | |
194 | |
195 // A preceding 'pre' loop will run at least 1 iteration (to do peeling), | |
196 // it may do under-flow checks for RCE and may do alignment iterations | |
197 // so the following main loop 'knows' that it is striding down cache | |
198 // lines. | |
199 | |
200 // A 'main' loop that is ONLY unrolled or peeled, never RCE'd or | |
201 // Aligned, may be missing it's pre-loop. | |
2465 | 202 int is_normal_loop() const { return (_loop_flags&PreMainPostFlagsMask) == Normal; } |
203 int is_pre_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Pre; } | |
204 int is_main_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Main; } | |
205 int is_post_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Post; } | |
206 int is_main_no_pre_loop() const { return _loop_flags & MainHasNoPreLoop; } | |
207 void set_main_no_pre_loop() { _loop_flags |= MainHasNoPreLoop; } | |
0 | 208 |
367
194b8e3a2fc4
6384206: Phis which are later unneeded are impairing our ability to inline based on static types
never
parents:
196
diff
changeset
|
209 int main_idx() const { return _main_idx; } |
194b8e3a2fc4
6384206: Phis which are later unneeded are impairing our ability to inline based on static types
never
parents:
196
diff
changeset
|
210 |
0 | 211 |
212 void set_pre_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Pre ; _main_idx = main->_idx; } | |
213 void set_main_loop ( ) { assert(is_normal_loop(),""); _loop_flags |= Main; } | |
214 void set_post_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Post; _main_idx = main->_idx; } | |
2465 | 215 void set_normal_loop( ) { _loop_flags &= ~PreMainPostFlagsMask; } |
216 | |
217 void set_trip_count(uint tc) { _trip_count = tc; } | |
218 uint trip_count() { return _trip_count; } | |
0 | 219 |
2465 | 220 bool has_exact_trip_count() const { return (_loop_flags & HasExactTripCount) != 0; } |
221 void set_exact_trip_count(uint tc) { | |
222 _trip_count = tc; | |
223 _loop_flags |= HasExactTripCount; | |
224 } | |
225 void set_nonexact_trip_count() { | |
226 _loop_flags &= ~HasExactTripCount; | |
227 } | |
0 | 228 |
229 void set_profile_trip_cnt(float ptc) { _profile_trip_cnt = ptc; } | |
230 float profile_trip_cnt() { return _profile_trip_cnt; } | |
231 | |
232 void double_unrolled_count() { _unrolled_count_log2++; } | |
233 int unrolled_count() { return 1 << MIN2(_unrolled_count_log2, BitsPerInt-3); } | |
234 | |
235 void set_node_count_before_unroll(int ct) { _node_count_before_unroll = ct; } | |
236 int node_count_before_unroll() { return _node_count_before_unroll; } | |
237 | |
238 #ifndef PRODUCT | |
239 virtual void dump_spec(outputStream *st) const; | |
240 #endif | |
241 }; | |
242 | |
243 //------------------------------CountedLoopEndNode----------------------------- | |
244 // CountedLoopEndNodes end simple trip counted loops. They act much like | |
245 // IfNodes. | |
246 class CountedLoopEndNode : public IfNode { | |
247 public: | |
248 enum { TestControl, TestValue }; | |
249 | |
250 CountedLoopEndNode( Node *control, Node *test, float prob, float cnt ) | |
251 : IfNode( control, test, prob, cnt) { | |
252 init_class_id(Class_CountedLoopEnd); | |
253 } | |
254 virtual int Opcode() const; | |
255 | |
256 Node *cmp_node() const { return (in(TestValue)->req() >=2) ? in(TestValue)->in(1) : NULL; } | |
257 Node *incr() const { Node *tmp = cmp_node(); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; } | |
258 Node *limit() const { Node *tmp = cmp_node(); return (tmp && tmp->req()==3) ? tmp->in(2) : NULL; } | |
259 Node *stride() const { Node *tmp = incr (); return (tmp && tmp->req()==3) ? tmp->in(2) : NULL; } | |
260 Node *phi() const { Node *tmp = incr (); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; } | |
261 Node *init_trip() const { Node *tmp = phi (); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; } | |
262 int stride_con() const; | |
263 bool stride_is_con() const { Node *tmp = stride (); return (tmp != NULL && tmp->is_Con()); } | |
264 BoolTest::mask test_trip() const { return in(TestValue)->as_Bool()->_test._test; } | |
265 CountedLoopNode *loopnode() const { | |
266 Node *ln = phi()->in(0); | |
267 assert( ln->Opcode() == Op_CountedLoop, "malformed loop" ); | |
268 return (CountedLoopNode*)ln; } | |
269 | |
270 #ifndef PRODUCT | |
271 virtual void dump_spec(outputStream *st) const; | |
272 #endif | |
273 }; | |
274 | |
275 | |
276 inline CountedLoopEndNode *CountedLoopNode::loopexit() const { | |
277 Node *bc = back_control(); | |
278 if( bc == NULL ) return NULL; | |
279 Node *le = bc->in(0); | |
280 if( le->Opcode() != Op_CountedLoopEnd ) | |
281 return NULL; | |
282 return (CountedLoopEndNode*)le; | |
283 } | |
284 inline Node *CountedLoopNode::init_trip() const { return loopexit() ? loopexit()->init_trip() : NULL; } | |
285 inline Node *CountedLoopNode::stride() const { return loopexit() ? loopexit()->stride() : NULL; } | |
286 inline int CountedLoopNode::stride_con() const { return loopexit() ? loopexit()->stride_con() : 0; } | |
287 inline bool CountedLoopNode::stride_is_con() const { return loopexit() && loopexit()->stride_is_con(); } | |
288 inline Node *CountedLoopNode::limit() const { return loopexit() ? loopexit()->limit() : NULL; } | |
289 inline Node *CountedLoopNode::incr() const { return loopexit() ? loopexit()->incr() : NULL; } | |
290 inline Node *CountedLoopNode::phi() const { return loopexit() ? loopexit()->phi() : NULL; } | |
291 | |
3345 | 292 //------------------------------LoopLimitNode----------------------------- |
293 // Counted Loop limit node which represents exact final iterator value: | |
294 // trip_count = (limit - init_trip + stride - 1)/stride | |
295 // final_value= trip_count * stride + init_trip. | |
296 // Use HW instructions to calculate it when it can overflow in integer. | |
297 // Note, final_value should fit into integer since counted loop has | |
298 // limit check: limit <= max_int-stride. | |
299 class LoopLimitNode : public Node { | |
300 enum { Init=1, Limit=2, Stride=3 }; | |
301 public: | |
302 LoopLimitNode( Compile* C, Node *init, Node *limit, Node *stride ) : Node(0,init,limit,stride) { | |
303 // Put it on the Macro nodes list to optimize during macro nodes expansion. | |
304 init_flags(Flag_is_macro); | |
305 C->add_macro_node(this); | |
306 } | |
307 virtual int Opcode() const; | |
308 virtual const Type *bottom_type() const { return TypeInt::INT; } | |
309 virtual uint ideal_reg() const { return Op_RegI; } | |
310 virtual const Type *Value( PhaseTransform *phase ) const; | |
311 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape); | |
312 virtual Node *Identity( PhaseTransform *phase ); | |
313 }; | |
0 | 314 |
315 // -----------------------------IdealLoopTree---------------------------------- | |
316 class IdealLoopTree : public ResourceObj { | |
317 public: | |
318 IdealLoopTree *_parent; // Parent in loop tree | |
319 IdealLoopTree *_next; // Next sibling in loop tree | |
320 IdealLoopTree *_child; // First child in loop tree | |
321 | |
322 // The head-tail backedge defines the loop. | |
323 // If tail is NULL then this loop has multiple backedges as part of the | |
324 // same loop. During cleanup I'll peel off the multiple backedges; merge | |
325 // them at the loop bottom and flow 1 real backedge into the loop. | |
326 Node *_head; // Head of loop | |
327 Node *_tail; // Tail of loop | |
328 inline Node *tail(); // Handle lazy update of _tail field | |
329 PhaseIdealLoop* _phase; | |
330 | |
331 Node_List _body; // Loop body for inner loops | |
332 | |
333 uint8 _nest; // Nesting depth | |
334 uint8 _irreducible:1, // True if irreducible | |
335 _has_call:1, // True if has call safepoint | |
336 _has_sfpt:1, // True if has non-call safepoint | |
337 _rce_candidate:1; // True if candidate for range check elimination | |
338 | |
6636 | 339 Node_List* _safepts; // List of safepoints in this loop |
39
76256d272075
6667612: (Escape Analysis) disable loop cloning if it has a scalar replaceable allocation
kvn
parents:
17
diff
changeset
|
340 Node_List* _required_safept; // A inner loop cannot delete these safepts; |
76256d272075
6667612: (Escape Analysis) disable loop cloning if it has a scalar replaceable allocation
kvn
parents:
17
diff
changeset
|
341 bool _allow_optimizations; // Allow loop optimizations |
0 | 342 |
343 IdealLoopTree( PhaseIdealLoop* phase, Node *head, Node *tail ) | |
344 : _parent(0), _next(0), _child(0), | |
345 _head(head), _tail(tail), | |
346 _phase(phase), | |
6636 | 347 _safepts(NULL), |
0 | 348 _required_safept(NULL), |
39
76256d272075
6667612: (Escape Analysis) disable loop cloning if it has a scalar replaceable allocation
kvn
parents:
17
diff
changeset
|
349 _allow_optimizations(true), |
0 | 350 _nest(0), _irreducible(0), _has_call(0), _has_sfpt(0), _rce_candidate(0) |
351 { } | |
352 | |
353 // Is 'l' a member of 'this'? | |
354 int is_member( const IdealLoopTree *l ) const; // Test for nested membership | |
355 | |
356 // Set loop nesting depth. Accumulate has_call bits. | |
357 int set_nest( uint depth ); | |
358 | |
359 // Split out multiple fall-in edges from the loop header. Move them to a | |
360 // private RegionNode before the loop. This becomes the loop landing pad. | |
361 void split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt ); | |
362 | |
363 // Split out the outermost loop from this shared header. | |
364 void split_outer_loop( PhaseIdealLoop *phase ); | |
365 | |
366 // Merge all the backedges from the shared header into a private Region. | |
367 // Feed that region as the one backedge to this loop. | |
368 void merge_many_backedges( PhaseIdealLoop *phase ); | |
369 | |
370 // Split shared headers and insert loop landing pads. | |
371 // Insert a LoopNode to replace the RegionNode. | |
372 // Returns TRUE if loop tree is structurally changed. | |
373 bool beautify_loops( PhaseIdealLoop *phase ); | |
374 | |
1172 | 375 // Perform optimization to use the loop predicates for null checks and range checks. |
376 // Applies to any loop level (not just the innermost one) | |
377 bool loop_predication( PhaseIdealLoop *phase); | |
378 | |
401
ee8f06bfb27c
6743188: incomplete fix for 6700047 C2 failed in idom_no_update
never
parents:
367
diff
changeset
|
379 // Perform iteration-splitting on inner loops. Split iterations to |
ee8f06bfb27c
6743188: incomplete fix for 6700047 C2 failed in idom_no_update
never
parents:
367
diff
changeset
|
380 // avoid range checks or one-shot null checks. Returns false if the |
ee8f06bfb27c
6743188: incomplete fix for 6700047 C2 failed in idom_no_update
never
parents:
367
diff
changeset
|
381 // current round of loop opts should stop. |
ee8f06bfb27c
6743188: incomplete fix for 6700047 C2 failed in idom_no_update
never
parents:
367
diff
changeset
|
382 bool iteration_split( PhaseIdealLoop *phase, Node_List &old_new ); |
0 | 383 |
401
ee8f06bfb27c
6743188: incomplete fix for 6700047 C2 failed in idom_no_update
never
parents:
367
diff
changeset
|
384 // Driver for various flavors of iteration splitting. Returns false |
ee8f06bfb27c
6743188: incomplete fix for 6700047 C2 failed in idom_no_update
never
parents:
367
diff
changeset
|
385 // if the current round of loop opts should stop. |
ee8f06bfb27c
6743188: incomplete fix for 6700047 C2 failed in idom_no_update
never
parents:
367
diff
changeset
|
386 bool iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new ); |
0 | 387 |
388 // Given dominators, try to find loops with calls that must always be | |
389 // executed (call dominates loop tail). These loops do not need non-call | |
390 // safepoints (ncsfpt). | |
391 void check_safepts(VectorSet &visited, Node_List &stack); | |
392 | |
393 // Allpaths backwards scan from loop tail, terminating each path at first safepoint | |
394 // encountered. | |
395 void allpaths_check_safepts(VectorSet &visited, Node_List &stack); | |
396 | |
397 // Convert to counted loops where possible | |
398 void counted_loop( PhaseIdealLoop *phase ); | |
399 | |
400 // Check for Node being a loop-breaking test | |
401 Node *is_loop_exit(Node *iff) const; | |
402 | |
403 // Returns true if ctrl is executed on every complete iteration | |
404 bool dominates_backedge(Node* ctrl); | |
405 | |
406 // Remove simplistic dead code from loop body | |
407 void DCE_loop_body(); | |
408 | |
409 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage. | |
410 // Replace with a 1-in-10 exit guess. | |
411 void adjust_loop_exit_prob( PhaseIdealLoop *phase ); | |
412 | |
413 // Return TRUE or FALSE if the loop should never be RCE'd or aligned. | |
414 // Useful for unrolling loops with NO array accesses. | |
415 bool policy_peel_only( PhaseIdealLoop *phase ) const; | |
416 | |
417 // Return TRUE or FALSE if the loop should be unswitched -- clone | |
418 // loop with an invariant test | |
419 bool policy_unswitching( PhaseIdealLoop *phase ) const; | |
420 | |
421 // Micro-benchmark spamming. Remove empty loops. | |
422 bool policy_do_remove_empty_loop( PhaseIdealLoop *phase ); | |
423 | |
2465 | 424 // Convert one iteration loop into normal code. |
425 bool policy_do_one_iteration_loop( PhaseIdealLoop *phase ); | |
426 | |
0 | 427 // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can |
428 // make some loop-invariant test (usually a null-check) happen before the | |
429 // loop. | |
430 bool policy_peeling( PhaseIdealLoop *phase ) const; | |
431 | |
432 // Return TRUE or FALSE if the loop should be maximally unrolled. Stash any | |
433 // known trip count in the counted loop node. | |
434 bool policy_maximally_unroll( PhaseIdealLoop *phase ) const; | |
435 | |
436 // Return TRUE or FALSE if the loop should be unrolled or not. Unroll if | |
437 // the loop is a CountedLoop and the body is small enough. | |
438 bool policy_unroll( PhaseIdealLoop *phase ) const; | |
439 | |
440 // Return TRUE or FALSE if the loop should be range-check-eliminated. | |
441 // Gather a list of IF tests that are dominated by iteration splitting; | |
442 // also gather the end of the first split and the start of the 2nd split. | |
443 bool policy_range_check( PhaseIdealLoop *phase ) const; | |
444 | |
445 // Return TRUE or FALSE if the loop should be cache-line aligned. | |
446 // Gather the expression that does the alignment. Note that only | |
605 | 447 // one array base can be aligned in a loop (unless the VM guarantees |
0 | 448 // mutual alignment). Note that if we vectorize short memory ops |
449 // into longer memory ops, we may want to increase alignment. | |
450 bool policy_align( PhaseIdealLoop *phase ) const; | |
451 | |
1172 | 452 // Return TRUE if "iff" is a range check. |
453 bool is_range_check_if(IfNode *iff, PhaseIdealLoop *phase, Invariance& invar) const; | |
454 | |
2465 | 455 // Compute loop exact trip count if possible |
456 void compute_exact_trip_count( PhaseIdealLoop *phase ); | |
457 | |
0 | 458 // Compute loop trip count from profile data |
459 void compute_profile_trip_cnt( PhaseIdealLoop *phase ); | |
460 | |
461 // Reassociate invariant expressions. | |
462 void reassociate_invariants(PhaseIdealLoop *phase); | |
463 // Reassociate invariant add and subtract expressions. | |
464 Node* reassociate_add_sub(Node* n1, PhaseIdealLoop *phase); | |
465 // Return nonzero index of invariant operand if invariant and variant | |
605 | 466 // are combined with an Add or Sub. Helper for reassociate_invariants. |
0 | 467 int is_invariant_addition(Node* n, PhaseIdealLoop *phase); |
468 | |
469 // Return true if n is invariant | |
470 bool is_invariant(Node* n) const; | |
471 | |
472 // Put loop body on igvn work list | |
473 void record_for_igvn(); | |
474 | |
475 bool is_loop() { return !_irreducible && _tail && !_tail->is_top(); } | |
476 bool is_inner() { return is_loop() && _child == NULL; } | |
477 bool is_counted() { return is_loop() && _head != NULL && _head->is_CountedLoop(); } | |
478 | |
479 #ifndef PRODUCT | |
480 void dump_head( ) const; // Dump loop head only | |
481 void dump() const; // Dump this loop recursively | |
482 void verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const; | |
483 #endif | |
484 | |
485 }; | |
486 | |
487 // -----------------------------PhaseIdealLoop--------------------------------- | |
488 // Computes the mapping from Nodes to IdealLoopTrees. Organizes IdealLoopTrees into a | |
489 // loop tree. Drives the loop-based transformations on the ideal graph. | |
490 class PhaseIdealLoop : public PhaseTransform { | |
491 friend class IdealLoopTree; | |
492 friend class SuperWord; | |
493 // Pre-computed def-use info | |
494 PhaseIterGVN &_igvn; | |
495 | |
496 // Head of loop tree | |
497 IdealLoopTree *_ltree_root; | |
498 | |
499 // Array of pre-order numbers, plus post-visited bit. | |
500 // ZERO for not pre-visited. EVEN for pre-visited but not post-visited. | |
501 // ODD for post-visited. Other bits are the pre-order number. | |
502 uint *_preorders; | |
503 uint _max_preorder; | |
504 | |
921
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
605
diff
changeset
|
505 const PhaseIdealLoop* _verify_me; |
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
605
diff
changeset
|
506 bool _verify_only; |
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
605
diff
changeset
|
507 |
0 | 508 // Allocate _preorders[] array |
509 void allocate_preorders() { | |
510 _max_preorder = C->unique()+8; | |
511 _preorders = NEW_RESOURCE_ARRAY(uint, _max_preorder); | |
512 memset(_preorders, 0, sizeof(uint) * _max_preorder); | |
513 } | |
514 | |
515 // Allocate _preorders[] array | |
516 void reallocate_preorders() { | |
517 if ( _max_preorder < C->unique() ) { | |
518 _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, C->unique()); | |
519 _max_preorder = C->unique(); | |
520 } | |
521 memset(_preorders, 0, sizeof(uint) * _max_preorder); | |
522 } | |
523 | |
524 // Check to grow _preorders[] array for the case when build_loop_tree_impl() | |
525 // adds new nodes. | |
526 void check_grow_preorders( ) { | |
527 if ( _max_preorder < C->unique() ) { | |
528 uint newsize = _max_preorder<<1; // double size of array | |
529 _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, newsize); | |
530 memset(&_preorders[_max_preorder],0,sizeof(uint)*(newsize-_max_preorder)); | |
531 _max_preorder = newsize; | |
532 } | |
533 } | |
534 // Check for pre-visited. Zero for NOT visited; non-zero for visited. | |
535 int is_visited( Node *n ) const { return _preorders[n->_idx]; } | |
536 // Pre-order numbers are written to the Nodes array as low-bit-set values. | |
537 void set_preorder_visited( Node *n, int pre_order ) { | |
538 assert( !is_visited( n ), "already set" ); | |
539 _preorders[n->_idx] = (pre_order<<1); | |
540 }; | |
541 // Return pre-order number. | |
542 int get_preorder( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]>>1; } | |
543 | |
544 // Check for being post-visited. | |
545 // Should be previsited already (checked with assert(is_visited(n))). | |
546 int is_postvisited( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]&1; } | |
547 | |
548 // Mark as post visited | |
549 void set_postvisited( Node *n ) { assert( !is_postvisited( n ), "" ); _preorders[n->_idx] |= 1; } | |
550 | |
551 // Set/get control node out. Set lower bit to distinguish from IdealLoopTree | |
552 // Returns true if "n" is a data node, false if it's a control node. | |
553 bool has_ctrl( Node *n ) const { return ((intptr_t)_nodes[n->_idx]) & 1; } | |
554 | |
555 // clear out dead code after build_loop_late | |
556 Node_List _deadlist; | |
557 | |
558 // Support for faster execution of get_late_ctrl()/dom_lca() | |
559 // when a node has many uses and dominator depth is deep. | |
560 Node_Array _dom_lca_tags; | |
561 void init_dom_lca_tags(); | |
562 void clear_dom_lca_tags(); | |
921
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
605
diff
changeset
|
563 |
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
605
diff
changeset
|
564 // Helper for debugging bad dominance relationships |
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
605
diff
changeset
|
565 bool verify_dominance(Node* n, Node* use, Node* LCA, Node* early); |
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
605
diff
changeset
|
566 |
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
605
diff
changeset
|
567 Node* compute_lca_of_uses(Node* n, Node* early, bool verify = false); |
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
605
diff
changeset
|
568 |
0 | 569 // Inline wrapper for frequent cases: |
570 // 1) only one use | |
571 // 2) a use is the same as the current LCA passed as 'n1' | |
572 Node *dom_lca_for_get_late_ctrl( Node *lca, Node *n, Node *tag ) { | |
573 assert( n->is_CFG(), "" ); | |
574 // Fast-path NULL lca | |
575 if( lca != NULL && lca != n ) { | |
576 assert( lca->is_CFG(), "" ); | |
577 // find LCA of all uses | |
578 n = dom_lca_for_get_late_ctrl_internal( lca, n, tag ); | |
579 } | |
580 return find_non_split_ctrl(n); | |
581 } | |
582 Node *dom_lca_for_get_late_ctrl_internal( Node *lca, Node *n, Node *tag ); | |
921
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
605
diff
changeset
|
583 |
0 | 584 // Helper function for directing control inputs away from CFG split |
585 // points. | |
586 Node *find_non_split_ctrl( Node *ctrl ) const { | |
587 if (ctrl != NULL) { | |
588 if (ctrl->is_MultiBranch()) { | |
589 ctrl = ctrl->in(0); | |
590 } | |
591 assert(ctrl->is_CFG(), "CFG"); | |
592 } | |
593 return ctrl; | |
594 } | |
595 | |
596 public: | |
597 bool has_node( Node* n ) const { return _nodes[n->_idx] != NULL; } | |
598 // check if transform created new nodes that need _ctrl recorded | |
599 Node *get_late_ctrl( Node *n, Node *early ); | |
600 Node *get_early_ctrl( Node *n ); | |
601 void set_early_ctrl( Node *n ); | |
602 void set_subtree_ctrl( Node *root ); | |
603 void set_ctrl( Node *n, Node *ctrl ) { | |
604 assert( !has_node(n) || has_ctrl(n), "" ); | |
605 assert( ctrl->in(0), "cannot set dead control node" ); | |
606 assert( ctrl == find_non_split_ctrl(ctrl), "must set legal crtl" ); | |
607 _nodes.map( n->_idx, (Node*)((intptr_t)ctrl + 1) ); | |
608 } | |
609 // Set control and update loop membership | |
610 void set_ctrl_and_loop(Node* n, Node* ctrl) { | |
611 IdealLoopTree* old_loop = get_loop(get_ctrl(n)); | |
612 IdealLoopTree* new_loop = get_loop(ctrl); | |
613 if (old_loop != new_loop) { | |
614 if (old_loop->_child == NULL) old_loop->_body.yank(n); | |
615 if (new_loop->_child == NULL) new_loop->_body.push(n); | |
616 } | |
617 set_ctrl(n, ctrl); | |
618 } | |
619 // Control nodes can be replaced or subsumed. During this pass they | |
620 // get their replacement Node in slot 1. Instead of updating the block | |
621 // location of all Nodes in the subsumed block, we lazily do it. As we | |
622 // pull such a subsumed block out of the array, we write back the final | |
623 // correct block. | |
624 Node *get_ctrl( Node *i ) { | |
625 assert(has_node(i), ""); | |
626 Node *n = get_ctrl_no_update(i); | |
627 _nodes.map( i->_idx, (Node*)((intptr_t)n + 1) ); | |
628 assert(has_node(i) && has_ctrl(i), ""); | |
629 assert(n == find_non_split_ctrl(n), "must return legal ctrl" ); | |
630 return n; | |
631 } | |
1172 | 632 // true if CFG node d dominates CFG node n |
633 bool is_dominator(Node *d, Node *n); | |
634 // return get_ctrl for a data node and self(n) for a CFG node | |
635 Node* ctrl_or_self(Node* n) { | |
636 if (has_ctrl(n)) | |
637 return get_ctrl(n); | |
638 else { | |
639 assert (n->is_CFG(), "must be a CFG node"); | |
640 return n; | |
641 } | |
642 } | |
0 | 643 |
644 private: | |
645 Node *get_ctrl_no_update( Node *i ) const { | |
646 assert( has_ctrl(i), "" ); | |
647 Node *n = (Node*)(((intptr_t)_nodes[i->_idx]) & ~1); | |
648 if (!n->in(0)) { | |
649 // Skip dead CFG nodes | |
650 do { | |
651 n = (Node*)(((intptr_t)_nodes[n->_idx]) & ~1); | |
652 } while (!n->in(0)); | |
653 n = find_non_split_ctrl(n); | |
654 } | |
655 return n; | |
656 } | |
657 | |
658 // Check for loop being set | |
659 // "n" must be a control node. Returns true if "n" is known to be in a loop. | |
660 bool has_loop( Node *n ) const { | |
661 assert(!has_node(n) || !has_ctrl(n), ""); | |
662 return has_node(n); | |
663 } | |
664 // Set loop | |
665 void set_loop( Node *n, IdealLoopTree *loop ) { | |
666 _nodes.map(n->_idx, (Node*)loop); | |
667 } | |
668 // Lazy-dazy update of 'get_ctrl' and 'idom_at' mechanisms. Replace | |
669 // the 'old_node' with 'new_node'. Kill old-node. Add a reference | |
670 // from old_node to new_node to support the lazy update. Reference | |
1172 | 671 // replaces loop reference, since that is not needed for dead node. |
0 | 672 public: |
673 void lazy_update( Node *old_node, Node *new_node ) { | |
674 assert( old_node != new_node, "no cycles please" ); | |
675 //old_node->set_req( 1, new_node /*NO DU INFO*/ ); | |
676 // Nodes always have DU info now, so re-use the side array slot | |
677 // for this node to provide the forwarding pointer. | |
678 _nodes.map( old_node->_idx, (Node*)((intptr_t)new_node + 1) ); | |
679 } | |
680 void lazy_replace( Node *old_node, Node *new_node ) { | |
1621
6027dddc26c6
6677629: PhaseIterGVN::subsume_node() should call hash_delete() and add_users_to_worklist()
kvn
parents:
1552
diff
changeset
|
681 _igvn.replace_node( old_node, new_node ); |
0 | 682 lazy_update( old_node, new_node ); |
683 } | |
684 void lazy_replace_proj( Node *old_node, Node *new_node ) { | |
685 assert( old_node->req() == 1, "use this for Projs" ); | |
686 _igvn.hash_delete(old_node); // Must hash-delete before hacking edges | |
687 old_node->add_req( NULL ); | |
688 lazy_replace( old_node, new_node ); | |
689 } | |
690 | |
691 private: | |
692 | |
693 // Place 'n' in some loop nest, where 'n' is a CFG node | |
694 void build_loop_tree(); | |
695 int build_loop_tree_impl( Node *n, int pre_order ); | |
696 // Insert loop into the existing loop tree. 'innermost' is a leaf of the | |
697 // loop tree, not the root. | |
698 IdealLoopTree *sort( IdealLoopTree *loop, IdealLoopTree *innermost ); | |
699 | |
700 // Place Data nodes in some loop nest | |
921
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
605
diff
changeset
|
701 void build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ); |
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
605
diff
changeset
|
702 void build_loop_late ( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ); |
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
605
diff
changeset
|
703 void build_loop_late_post ( Node* n ); |
0 | 704 |
705 // Array of immediate dominance info for each CFG node indexed by node idx | |
706 private: | |
707 uint _idom_size; | |
708 Node **_idom; // Array of immediate dominators | |
709 uint *_dom_depth; // Used for fast LCA test | |
710 GrowableArray<uint>* _dom_stk; // For recomputation of dom depth | |
711 | |
712 Node* idom_no_update(Node* d) const { | |
713 assert(d->_idx < _idom_size, "oob"); | |
714 Node* n = _idom[d->_idx]; | |
715 assert(n != NULL,"Bad immediate dominator info."); | |
716 while (n->in(0) == NULL) { // Skip dead CFG nodes | |
717 //n = n->in(1); | |
718 n = (Node*)(((intptr_t)_nodes[n->_idx]) & ~1); | |
719 assert(n != NULL,"Bad immediate dominator info."); | |
720 } | |
721 return n; | |
722 } | |
723 Node *idom(Node* d) const { | |
724 uint didx = d->_idx; | |
725 Node *n = idom_no_update(d); | |
726 _idom[didx] = n; // Lazily remove dead CFG nodes from table. | |
727 return n; | |
728 } | |
729 uint dom_depth(Node* d) const { | |
730 assert(d->_idx < _idom_size, ""); | |
731 return _dom_depth[d->_idx]; | |
732 } | |
733 void set_idom(Node* d, Node* n, uint dom_depth); | |
734 // Locally compute IDOM using dom_lca call | |
735 Node *compute_idom( Node *region ) const; | |
736 // Recompute dom_depth | |
737 void recompute_dom_depth(); | |
738 | |
739 // Is safept not required by an outer loop? | |
740 bool is_deleteable_safept(Node* sfpt); | |
741 | |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
2248
diff
changeset
|
742 // Replace parallel induction variable (parallel to trip counter) |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
2248
diff
changeset
|
743 void replace_parallel_iv(IdealLoopTree *loop); |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
2248
diff
changeset
|
744 |
921
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
605
diff
changeset
|
745 // Perform verification that the graph is valid. |
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
605
diff
changeset
|
746 PhaseIdealLoop( PhaseIterGVN &igvn) : |
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
605
diff
changeset
|
747 PhaseTransform(Ideal_Loop), |
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
605
diff
changeset
|
748 _igvn(igvn), |
2248
194c9fdee631
7017240: C2: native memory leak in nsk/regression/b4675027 on windows-x86 in comp mode with G1
kvn
parents:
1972
diff
changeset
|
749 _dom_lca_tags(arena()), // Thread::resource_area |
921
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
605
diff
changeset
|
750 _verify_me(NULL), |
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
605
diff
changeset
|
751 _verify_only(true) { |
4064
670a74b863fc
7107042: assert(no_dead_loop) failed: dead loop detected
kvn
parents:
3845
diff
changeset
|
752 build_and_optimize(false, false); |
921
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
605
diff
changeset
|
753 } |
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
605
diff
changeset
|
754 |
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
605
diff
changeset
|
755 // build the loop tree and perform any requested optimizations |
4064
670a74b863fc
7107042: assert(no_dead_loop) failed: dead loop detected
kvn
parents:
3845
diff
changeset
|
756 void build_and_optimize(bool do_split_if, bool skip_loop_opts); |
921
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
605
diff
changeset
|
757 |
0 | 758 public: |
759 // Dominators for the sea of nodes | |
760 void Dominators(); | |
761 Node *dom_lca( Node *n1, Node *n2 ) const { | |
762 return find_non_split_ctrl(dom_lca_internal(n1, n2)); | |
763 } | |
764 Node *dom_lca_internal( Node *n1, Node *n2 ) const; | |
765 | |
766 // Compute the Ideal Node to Loop mapping | |
4064
670a74b863fc
7107042: assert(no_dead_loop) failed: dead loop detected
kvn
parents:
3845
diff
changeset
|
767 PhaseIdealLoop( PhaseIterGVN &igvn, bool do_split_ifs, bool skip_loop_opts = false) : |
921
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
605
diff
changeset
|
768 PhaseTransform(Ideal_Loop), |
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
605
diff
changeset
|
769 _igvn(igvn), |
2248
194c9fdee631
7017240: C2: native memory leak in nsk/regression/b4675027 on windows-x86 in comp mode with G1
kvn
parents:
1972
diff
changeset
|
770 _dom_lca_tags(arena()), // Thread::resource_area |
921
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
605
diff
changeset
|
771 _verify_me(NULL), |
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
605
diff
changeset
|
772 _verify_only(false) { |
4064
670a74b863fc
7107042: assert(no_dead_loop) failed: dead loop detected
kvn
parents:
3845
diff
changeset
|
773 build_and_optimize(do_split_ifs, skip_loop_opts); |
921
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
605
diff
changeset
|
774 } |
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
605
diff
changeset
|
775 |
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
605
diff
changeset
|
776 // Verify that verify_me made the same decisions as a fresh run. |
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
605
diff
changeset
|
777 PhaseIdealLoop( PhaseIterGVN &igvn, const PhaseIdealLoop *verify_me) : |
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
605
diff
changeset
|
778 PhaseTransform(Ideal_Loop), |
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
605
diff
changeset
|
779 _igvn(igvn), |
2248
194c9fdee631
7017240: C2: native memory leak in nsk/regression/b4675027 on windows-x86 in comp mode with G1
kvn
parents:
1972
diff
changeset
|
780 _dom_lca_tags(arena()), // Thread::resource_area |
921
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
605
diff
changeset
|
781 _verify_me(verify_me), |
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
605
diff
changeset
|
782 _verify_only(false) { |
4064
670a74b863fc
7107042: assert(no_dead_loop) failed: dead loop detected
kvn
parents:
3845
diff
changeset
|
783 build_and_optimize(false, false); |
921
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
605
diff
changeset
|
784 } |
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
605
diff
changeset
|
785 |
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
605
diff
changeset
|
786 // Build and verify the loop tree without modifying the graph. This |
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
605
diff
changeset
|
787 // is useful to verify that all inputs properly dominate their uses. |
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
605
diff
changeset
|
788 static void verify(PhaseIterGVN& igvn) { |
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
605
diff
changeset
|
789 #ifdef ASSERT |
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
605
diff
changeset
|
790 PhaseIdealLoop v(igvn); |
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
605
diff
changeset
|
791 #endif |
046932b72aa2
6862956: PhaseIdealLoop should have a CFG verification mode
never
parents:
605
diff
changeset
|
792 } |
0 | 793 |
794 // True if the method has at least 1 irreducible loop | |
795 bool _has_irreducible_loops; | |
796 | |
797 // Per-Node transform | |
798 virtual Node *transform( Node *a_node ) { return 0; } | |
799 | |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
2248
diff
changeset
|
800 bool is_counted_loop( Node *x, IdealLoopTree *loop ); |
0 | 801 |
3345 | 802 Node* exact_limit( IdealLoopTree *loop ); |
803 | |
0 | 804 // Return a post-walked LoopNode |
805 IdealLoopTree *get_loop( Node *n ) const { | |
806 // Dead nodes have no loop, so return the top level loop instead | |
807 if (!has_node(n)) return _ltree_root; | |
808 assert(!has_ctrl(n), ""); | |
809 return (IdealLoopTree*)_nodes[n->_idx]; | |
810 } | |
811 | |
812 // Is 'n' a (nested) member of 'loop'? | |
813 int is_member( const IdealLoopTree *loop, Node *n ) const { | |
814 return loop->is_member(get_loop(n)); } | |
815 | |
816 // This is the basic building block of the loop optimizations. It clones an | |
817 // entire loop body. It makes an old_new loop body mapping; with this | |
818 // mapping you can find the new-loop equivalent to an old-loop node. All | |
819 // new-loop nodes are exactly equal to their old-loop counterparts, all | |
820 // edges are the same. All exits from the old-loop now have a RegionNode | |
821 // that merges the equivalent new-loop path. This is true even for the | |
822 // normal "loop-exit" condition. All uses of loop-invariant old-loop values | |
823 // now come from (one or more) Phis that merge their new-loop equivalents. | |
824 // Parameter side_by_side_idom: | |
825 // When side_by_size_idom is NULL, the dominator tree is constructed for | |
826 // the clone loop to dominate the original. Used in construction of | |
827 // pre-main-post loop sequence. | |
828 // When nonnull, the clone and original are side-by-side, both are | |
829 // dominated by the passed in side_by_side_idom node. Used in | |
830 // construction of unswitched loops. | |
831 void clone_loop( IdealLoopTree *loop, Node_List &old_new, int dom_depth, | |
832 Node* side_by_side_idom = NULL); | |
833 | |
834 // If we got the effect of peeling, either by actually peeling or by | |
835 // making a pre-loop which must execute at least once, we can remove | |
836 // all loop-invariant dominated tests in the main body. | |
837 void peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new ); | |
838 | |
839 // Generate code to do a loop peel for the given loop (and body). | |
840 // old_new is a temp array. | |
841 void do_peeling( IdealLoopTree *loop, Node_List &old_new ); | |
842 | |
843 // Add pre and post loops around the given loop. These loops are used | |
844 // during RCE, unrolling and aligning loops. | |
845 void insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only ); | |
846 // If Node n lives in the back_ctrl block, we clone a private version of n | |
847 // in preheader_ctrl block and return that, otherwise return n. | |
3788
e3cbc9ddd434
7044738: Loop unroll optimization causes incorrect result
kvn
parents:
3383
diff
changeset
|
848 Node *clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n, VectorSet &visited, Node_Stack &clones ); |
0 | 849 |
850 // Take steps to maximally unroll the loop. Peel any odd iterations, then | |
851 // unroll to do double iterations. The next round of major loop transforms | |
852 // will repeat till the doubled loop body does all remaining iterations in 1 | |
853 // pass. | |
854 void do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new ); | |
855 | |
856 // Unroll the loop body one step - make each trip do 2 iterations. | |
857 void do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip ); | |
858 | |
859 // Return true if exp is a constant times an induction var | |
860 bool is_scaled_iv(Node* exp, Node* iv, int* p_scale); | |
861 | |
862 // Return true if exp is a scaled induction var plus (or minus) constant | |
863 bool is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth = 0); | |
864 | |
1172 | 865 // Return true if proj is for "proj->[region->..]call_uct" |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
2248
diff
changeset
|
866 static bool is_uncommon_trap_proj(ProjNode* proj, Deoptimization::DeoptReason reason); |
1172 | 867 // Return true for "if(test)-> proj -> ... |
868 // | | |
869 // V | |
870 // other_proj->[region->..]call_uct" | |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
2248
diff
changeset
|
871 static bool is_uncommon_trap_if_pattern(ProjNode* proj, Deoptimization::DeoptReason reason); |
1172 | 872 // Create a new if above the uncommon_trap_if_pattern for the predicate to be promoted |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
2248
diff
changeset
|
873 ProjNode* create_new_if_for_predicate(ProjNode* cont_proj, Node* new_entry, |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
2248
diff
changeset
|
874 Deoptimization::DeoptReason reason); |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
2248
diff
changeset
|
875 void register_control(Node* n, IdealLoopTree *loop, Node* pred); |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
2248
diff
changeset
|
876 |
2445 | 877 // Clone loop predicates to cloned loops (peeled, unswitched) |
878 static ProjNode* clone_predicate(ProjNode* predicate_proj, Node* new_entry, | |
879 Deoptimization::DeoptReason reason, | |
880 PhaseIdealLoop* loop_phase, | |
881 PhaseIterGVN* igvn); | |
3845
c96c3eb1efae
7068051: SIGSEGV in PhaseIdealLoop::build_loop_late_post
kvn
parents:
3840
diff
changeset
|
882 |
2445 | 883 static Node* clone_loop_predicates(Node* old_entry, Node* new_entry, |
3345 | 884 bool clone_limit_check, |
2445 | 885 PhaseIdealLoop* loop_phase, |
886 PhaseIterGVN* igvn); | |
3345 | 887 Node* clone_loop_predicates(Node* old_entry, Node* new_entry, bool clone_limit_check); |
2445 | 888 |
889 static Node* skip_loop_predicates(Node* entry); | |
890 | |
891 // Find a good location to insert a predicate | |
2383
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
2248
diff
changeset
|
892 static ProjNode* find_predicate_insertion_point(Node* start_c, Deoptimization::DeoptReason reason); |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
2248
diff
changeset
|
893 // Find a predicate |
9dc311b8473e
7008866: Missing loop predicate for loop with multiple entries
kvn
parents:
2248
diff
changeset
|
894 static Node* find_predicate(Node* entry); |
1172 | 895 // Construct a range check for a predicate if |
3345 | 896 BoolNode* rc_predicate(IdealLoopTree *loop, Node* ctrl, |
1172 | 897 int scale, Node* offset, |
898 Node* init, Node* limit, Node* stride, | |
1303
c047da02984c
6930043: C2: SIGSEGV in javasoft.sqe.tests.lang.arr017.arr01702.arr01702.loop_forw(II)I
never
parents:
1172
diff
changeset
|
899 Node* range, bool upper); |
1172 | 900 |
901 // Implementation of the loop predication to promote checks outside the loop | |
902 bool loop_predication_impl(IdealLoopTree *loop); | |
903 | |
904 // Helper function to collect predicate for eliminating the useless ones | |
905 void collect_potentially_useful_predicates(IdealLoopTree *loop, Unique_Node_List &predicate_opaque1); | |
906 void eliminate_useless_predicates(); | |
907 | |
0 | 908 // Eliminate range-checks and other trip-counter vs loop-invariant tests. |
909 void do_range_check( IdealLoopTree *loop, Node_List &old_new ); | |
910 | |
911 // Create a slow version of the loop by cloning the loop | |
912 // and inserting an if to select fast-slow versions. | |
913 ProjNode* create_slow_version_of_loop(IdealLoopTree *loop, | |
914 Node_List &old_new); | |
915 | |
916 // Clone loop with an invariant test (that does not exit) and | |
917 // insert a clone of the test that selects which version to | |
918 // execute. | |
919 void do_unswitching (IdealLoopTree *loop, Node_List &old_new); | |
920 | |
921 // Find candidate "if" for unswitching | |
922 IfNode* find_unswitching_candidate(const IdealLoopTree *loop) const; | |
923 | |
924 // Range Check Elimination uses this function! | |
925 // Constrain the main loop iterations so the affine function: | |
3345 | 926 // low_limit <= scale_con * I + offset < upper_limit |
0 | 927 // always holds true. That is, either increase the number of iterations in |
928 // the pre-loop or the post-loop until the condition holds true in the main | |
929 // loop. Scale_con, offset and limit are all loop invariant. | |
3345 | 930 void add_constraint( int stride_con, int scale_con, Node *offset, Node *low_limit, Node *upper_limit, Node *pre_ctrl, Node **pre_limit, Node **main_limit ); |
3383
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3345
diff
changeset
|
931 // Helper function for add_constraint(). |
38569792a45a
7044725: -XX:-UnrollLimitCheck -Xcomp : Exception: String index out of range: 29488
kvn
parents:
3345
diff
changeset
|
932 Node* adjust_limit( int stride_con, Node * scale, Node *offset, Node *rc_limit, Node *loop_limit, Node *pre_ctrl ); |
0 | 933 |
934 // Partially peel loop up through last_peel node. | |
935 bool partial_peel( IdealLoopTree *loop, Node_List &old_new ); | |
936 | |
937 // Create a scheduled list of nodes control dependent on ctrl set. | |
938 void scheduled_nodelist( IdealLoopTree *loop, VectorSet& ctrl, Node_List &sched ); | |
939 // Has a use in the vector set | |
940 bool has_use_in_set( Node* n, VectorSet& vset ); | |
941 // Has use internal to the vector set (ie. not in a phi at the loop head) | |
942 bool has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ); | |
943 // clone "n" for uses that are outside of loop | |
944 void clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ); | |
945 // clone "n" for special uses that are in the not_peeled region | |
946 void clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n, | |
947 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ); | |
948 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist | |
949 void insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ); | |
950 #ifdef ASSERT | |
951 // Validate the loop partition sets: peel and not_peel | |
952 bool is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, VectorSet& not_peel ); | |
953 // Ensure that uses outside of loop are of the right form | |
954 bool is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list, | |
955 uint orig_exit_idx, uint clone_exit_idx); | |
956 bool is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx); | |
957 #endif | |
958 | |
959 // Returns nonzero constant stride if-node is a possible iv test (otherwise returns zero.) | |
960 int stride_of_possible_iv( Node* iff ); | |
961 bool is_possible_iv_test( Node* iff ) { return stride_of_possible_iv(iff) != 0; } | |
962 // Return the (unique) control output node that's in the loop (if it exists.) | |
963 Node* stay_in_loop( Node* n, IdealLoopTree *loop); | |
964 // Insert a signed compare loop exit cloned from an unsigned compare. | |
965 IfNode* insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop); | |
966 void remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop); | |
967 // Utility to register node "n" with PhaseIdealLoop | |
968 void register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth); | |
969 // Utility to create an if-projection | |
970 ProjNode* proj_clone(ProjNode* p, IfNode* iff); | |
971 // Force the iff control output to be the live_proj | |
972 Node* short_circuit_if(IfNode* iff, ProjNode* live_proj); | |
973 // Insert a region before an if projection | |
974 RegionNode* insert_region_before_proj(ProjNode* proj); | |
975 // Insert a new if before an if projection | |
976 ProjNode* insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj); | |
977 | |
978 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. | |
979 // "Nearly" because all Nodes have been cloned from the original in the loop, | |
980 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs | |
981 // through the Phi recursively, and return a Bool. | |
982 BoolNode *clone_iff( PhiNode *phi, IdealLoopTree *loop ); | |
983 CmpNode *clone_bool( PhiNode *phi, IdealLoopTree *loop ); | |
984 | |
985 | |
986 // Rework addressing expressions to get the most loop-invariant stuff | |
987 // moved out. We'd like to do all associative operators, but it's especially | |
988 // important (common) to do address expressions. | |
989 Node *remix_address_expressions( Node *n ); | |
990 | |
991 // Attempt to use a conditional move instead of a phi/branch | |
992 Node *conditional_move( Node *n ); | |
993 | |
994 // Reorganize offset computations to lower register pressure. | |
995 // Mostly prevent loop-fallout uses of the pre-incremented trip counter | |
996 // (which are then alive with the post-incremented trip counter | |
997 // forcing an extra register move) | |
998 void reorg_offsets( IdealLoopTree *loop ); | |
999 | |
1000 // Check for aggressive application of 'split-if' optimization, | |
1001 // using basic block level info. | |
1002 void split_if_with_blocks ( VectorSet &visited, Node_Stack &nstack ); | |
1003 Node *split_if_with_blocks_pre ( Node *n ); | |
1004 void split_if_with_blocks_post( Node *n ); | |
1005 Node *has_local_phi_input( Node *n ); | |
1006 // Mark an IfNode as being dominated by a prior test, | |
1007 // without actually altering the CFG (and hence IDOM info). | |
3840
4e761e7e6e12
7070134: Hotspot crashes with sigsegv from PorterStemmer
kvn
parents:
3788
diff
changeset
|
1008 void dominated_by( Node *prevdom, Node *iff, bool flip = false, bool exclude_loop_predicate = false ); |
0 | 1009 |
1010 // Split Node 'n' through merge point | |
1011 Node *split_thru_region( Node *n, Node *region ); | |
1012 // Split Node 'n' through merge point if there is enough win. | |
1013 Node *split_thru_phi( Node *n, Node *region, int policy ); | |
1014 // Found an If getting its condition-code input from a Phi in the | |
1015 // same block. Split thru the Region. | |
1016 void do_split_if( Node *iff ); | |
1017 | |
1763 | 1018 // Conversion of fill/copy patterns into intrisic versions |
1019 bool do_intrinsify_fill(); | |
1020 bool intrinsify_fill(IdealLoopTree* lpt); | |
1021 bool match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value, | |
1022 Node*& shift, Node*& offset); | |
1023 | |
0 | 1024 private: |
1025 // Return a type based on condition control flow | |
1026 const TypeInt* filtered_type( Node *n, Node* n_ctrl); | |
1027 const TypeInt* filtered_type( Node *n ) { return filtered_type(n, NULL); } | |
1028 // Helpers for filtered type | |
1029 const TypeInt* filtered_type_from_dominators( Node* val, Node *val_ctrl); | |
1030 | |
1031 // Helper functions | |
1032 Node *spinup( Node *iff, Node *new_false, Node *new_true, Node *region, Node *phi, small_cache *cache ); | |
1033 Node *find_use_block( Node *use, Node *def, Node *old_false, Node *new_false, Node *old_true, Node *new_true ); | |
1034 void handle_use( Node *use, Node *def, small_cache *cache, Node *region_dom, Node *new_false, Node *new_true, Node *old_false, Node *old_true ); | |
1035 bool split_up( Node *n, Node *blk1, Node *blk2 ); | |
1036 void sink_use( Node *use, Node *post_loop ); | |
1037 Node *place_near_use( Node *useblock ) const; | |
1038 | |
1039 bool _created_loop_node; | |
1040 public: | |
1041 void set_created_loop_node() { _created_loop_node = true; } | |
1042 bool created_loop_node() { return _created_loop_node; } | |
1172 | 1043 void register_new_node( Node *n, Node *blk ); |
0 | 1044 |
4779
c8d8e124380c
7064302: JDK7 build 147 crashed after testing my java 6-compiled web app
kvn
parents:
4064
diff
changeset
|
1045 #ifdef ASSERT |
c8d8e124380c
7064302: JDK7 build 147 crashed after testing my java 6-compiled web app
kvn
parents:
4064
diff
changeset
|
1046 void dump_bad_graph(Node* n, Node* early, Node* LCA); |
c8d8e124380c
7064302: JDK7 build 147 crashed after testing my java 6-compiled web app
kvn
parents:
4064
diff
changeset
|
1047 #endif |
c8d8e124380c
7064302: JDK7 build 147 crashed after testing my java 6-compiled web app
kvn
parents:
4064
diff
changeset
|
1048 |
0 | 1049 #ifndef PRODUCT |
1050 void dump( ) const; | |
1051 void dump( IdealLoopTree *loop, uint rpo_idx, Node_List &rpo_list ) const; | |
1052 void rpo( Node *start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list ) const; | |
1053 void verify() const; // Major slow :-) | |
1054 void verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const; | |
1055 IdealLoopTree *get_loop_idx(Node* n) const { | |
1056 // Dead nodes have no loop, so return the top level loop instead | |
1057 return _nodes[n->_idx] ? (IdealLoopTree*)_nodes[n->_idx] : _ltree_root; | |
1058 } | |
1059 // Print some stats | |
1060 static void print_statistics(); | |
1061 static int _loop_invokes; // Count of PhaseIdealLoop invokes | |
1062 static int _loop_work; // Sum of PhaseIdealLoop x _unique | |
1063 #endif | |
1064 }; | |
1065 | |
1066 inline Node* IdealLoopTree::tail() { | |
1067 // Handle lazy update of _tail field | |
1068 Node *n = _tail; | |
1069 //while( !n->in(0) ) // Skip dead CFG nodes | |
1070 //n = n->in(1); | |
1071 if (n->in(0) == NULL) | |
1072 n = _phase->get_ctrl(n); | |
1073 _tail = n; | |
1074 return n; | |
1075 } | |
1076 | |
1077 | |
1078 // Iterate over the loop tree using a preorder, left-to-right traversal. | |
1079 // | |
1080 // Example that visits all counted loops from within PhaseIdealLoop | |
1081 // | |
1082 // for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) { | |
1083 // IdealLoopTree* lpt = iter.current(); | |
1084 // if (!lpt->is_counted()) continue; | |
1085 // ... | |
1086 class LoopTreeIterator : public StackObj { | |
1087 private: | |
1088 IdealLoopTree* _root; | |
1089 IdealLoopTree* _curnt; | |
1090 | |
1091 public: | |
1092 LoopTreeIterator(IdealLoopTree* root) : _root(root), _curnt(root) {} | |
1093 | |
1094 bool done() { return _curnt == NULL; } // Finished iterating? | |
1095 | |
1096 void next(); // Advance to next loop tree | |
1097 | |
1098 IdealLoopTree* current() { return _curnt; } // Return current value of iterator. | |
1099 }; | |
1972 | 1100 |
1101 #endif // SHARE_VM_OPTO_LOOPNODE_HPP |