1 | /* |
2 | * Copyright (c) 1998, 2019, Oracle and/or its affiliates. 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 Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA |
20 | * or visit www.oracle.com if you need additional information or have any |
21 | * questions. |
22 | * |
23 | */ |
24 | |
25 | #ifndef SHARE_OPTO_LOOPNODE_HPP |
26 | #define SHARE_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 | |
34 | class CmpNode; |
35 | class CountedLoopEndNode; |
36 | class CountedLoopNode; |
37 | class IdealLoopTree; |
38 | class LoopNode; |
39 | class Node; |
40 | class OuterStripMinedLoopEndNode; |
41 | class PathFrequency; |
42 | class PhaseIdealLoop; |
43 | class CountedLoopReserveKit; |
44 | class VectorSet; |
45 | class Invariance; |
46 | struct small_cache; |
47 | |
48 | // |
49 | // I D E A L I Z E D L O O P S |
50 | // |
51 | // Idealized loops are the set of loops I perform more interesting |
52 | // transformations on, beyond simple hoisting. |
53 | |
54 | //------------------------------LoopNode--------------------------------------- |
55 | // Simple loop header. Fall in path on left, loop-back path on right. |
56 | class LoopNode : public RegionNode { |
57 | // Size is bigger to hold the flags. However, the flags do not change |
58 | // the semantics so it does not appear in the hash & cmp functions. |
59 | virtual uint size_of() const { return sizeof(*this); } |
60 | protected: |
61 | uint _loop_flags; |
62 | // Names for flag bitfields |
63 | enum { Normal=0, Pre=1, Main=2, Post=3, PreMainPostFlagsMask=3, |
64 | MainHasNoPreLoop=4, |
65 | HasExactTripCount=8, |
66 | InnerLoop=16, |
67 | PartialPeelLoop=32, |
68 | PartialPeelFailed=64, |
69 | HasReductions=128, |
70 | WasSlpAnalyzed=256, |
71 | PassedSlpAnalysis=512, |
72 | DoUnrollOnly=1024, |
73 | VectorizedLoop=2048, |
74 | HasAtomicPostLoop=4096, |
75 | HasRangeChecks=8192, |
76 | IsMultiversioned=16384, |
77 | StripMined=32768, |
78 | SubwordLoop=65536, |
79 | ProfileTripFailed=131072}; |
80 | char _unswitch_count; |
81 | enum { _unswitch_max=3 }; |
82 | char _postloop_flags; |
83 | enum { LoopNotRCEChecked = 0, LoopRCEChecked = 1, RCEPostLoop = 2 }; |
84 | |
85 | // Expected trip count from profile data |
86 | float _profile_trip_cnt; |
87 | |
88 | public: |
89 | // Names for edge indices |
90 | enum { Self=0, EntryControl, LoopBackControl }; |
91 | |
92 | bool is_inner_loop() const { return _loop_flags & InnerLoop; } |
93 | void set_inner_loop() { _loop_flags |= InnerLoop; } |
94 | |
95 | bool range_checks_present() const { return _loop_flags & HasRangeChecks; } |
96 | bool is_multiversioned() const { return _loop_flags & IsMultiversioned; } |
97 | bool is_vectorized_loop() const { return _loop_flags & VectorizedLoop; } |
98 | bool is_partial_peel_loop() const { return _loop_flags & PartialPeelLoop; } |
99 | void set_partial_peel_loop() { _loop_flags |= PartialPeelLoop; } |
100 | bool partial_peel_has_failed() const { return _loop_flags & PartialPeelFailed; } |
101 | bool is_strip_mined() const { return _loop_flags & StripMined; } |
102 | bool is_profile_trip_failed() const { return _loop_flags & ProfileTripFailed; } |
103 | bool is_subword_loop() const { return _loop_flags & SubwordLoop; } |
104 | |
105 | void mark_partial_peel_failed() { _loop_flags |= PartialPeelFailed; } |
106 | void mark_has_reductions() { _loop_flags |= HasReductions; } |
107 | void mark_was_slp() { _loop_flags |= WasSlpAnalyzed; } |
108 | void mark_passed_slp() { _loop_flags |= PassedSlpAnalysis; } |
109 | void mark_do_unroll_only() { _loop_flags |= DoUnrollOnly; } |
110 | void mark_loop_vectorized() { _loop_flags |= VectorizedLoop; } |
111 | void mark_has_atomic_post_loop() { _loop_flags |= HasAtomicPostLoop; } |
112 | void mark_has_range_checks() { _loop_flags |= HasRangeChecks; } |
113 | void mark_is_multiversioned() { _loop_flags |= IsMultiversioned; } |
114 | void mark_strip_mined() { _loop_flags |= StripMined; } |
115 | void clear_strip_mined() { _loop_flags &= ~StripMined; } |
116 | void mark_profile_trip_failed() { _loop_flags |= ProfileTripFailed; } |
117 | void mark_subword_loop() { _loop_flags |= SubwordLoop; } |
118 | |
119 | int unswitch_max() { return _unswitch_max; } |
120 | int unswitch_count() { return _unswitch_count; } |
121 | |
122 | int has_been_range_checked() const { return _postloop_flags & LoopRCEChecked; } |
123 | void set_has_been_range_checked() { _postloop_flags |= LoopRCEChecked; } |
124 | int is_rce_post_loop() const { return _postloop_flags & RCEPostLoop; } |
125 | void set_is_rce_post_loop() { _postloop_flags |= RCEPostLoop; } |
126 | |
127 | void set_unswitch_count(int val) { |
128 | assert (val <= unswitch_max(), "too many unswitches" ); |
129 | _unswitch_count = val; |
130 | } |
131 | |
132 | void set_profile_trip_cnt(float ptc) { _profile_trip_cnt = ptc; } |
133 | float profile_trip_cnt() { return _profile_trip_cnt; } |
134 | |
135 | LoopNode(Node *entry, Node *backedge) |
136 | : RegionNode(3), _loop_flags(0), _unswitch_count(0), |
137 | _postloop_flags(0), _profile_trip_cnt(COUNT_UNKNOWN) { |
138 | init_class_id(Class_Loop); |
139 | init_req(EntryControl, entry); |
140 | init_req(LoopBackControl, backedge); |
141 | } |
142 | |
143 | virtual Node *Ideal(PhaseGVN *phase, bool can_reshape); |
144 | virtual int Opcode() const; |
145 | bool can_be_counted_loop(PhaseTransform* phase) const { |
146 | return req() == 3 && in(0) != NULL && |
147 | in(1) != NULL && phase->type(in(1)) != Type::TOP && |
148 | in(2) != NULL && phase->type(in(2)) != Type::TOP; |
149 | } |
150 | bool is_valid_counted_loop() const; |
151 | #ifndef PRODUCT |
152 | virtual void dump_spec(outputStream *st) const; |
153 | #endif |
154 | |
155 | void verify_strip_mined(int expect_skeleton) const; |
156 | virtual LoopNode* skip_strip_mined(int expect_skeleton = 1) { return this; } |
157 | virtual IfTrueNode* outer_loop_tail() const { ShouldNotReachHere(); return NULL; } |
158 | virtual OuterStripMinedLoopEndNode* outer_loop_end() const { ShouldNotReachHere(); return NULL; } |
159 | virtual IfFalseNode* outer_loop_exit() const { ShouldNotReachHere(); return NULL; } |
160 | virtual SafePointNode* outer_safepoint() const { ShouldNotReachHere(); return NULL; } |
161 | }; |
162 | |
163 | //------------------------------Counted Loops---------------------------------- |
164 | // Counted loops are all trip-counted loops, with exactly 1 trip-counter exit |
165 | // path (and maybe some other exit paths). The trip-counter exit is always |
166 | // last in the loop. The trip-counter have to stride by a constant; |
167 | // the exit value is also loop invariant. |
168 | |
169 | // CountedLoopNodes and CountedLoopEndNodes come in matched pairs. The |
170 | // CountedLoopNode has the incoming loop control and the loop-back-control |
171 | // which is always the IfTrue before the matching CountedLoopEndNode. The |
172 | // CountedLoopEndNode has an incoming control (possibly not the |
173 | // CountedLoopNode if there is control flow in the loop), the post-increment |
174 | // trip-counter value, and the limit. The trip-counter value is always of |
175 | // the form (Op old-trip-counter stride). The old-trip-counter is produced |
176 | // by a Phi connected to the CountedLoopNode. The stride is constant. |
177 | // The Op is any commutable opcode, including Add, Mul, Xor. The |
178 | // CountedLoopEndNode also takes in the loop-invariant limit value. |
179 | |
180 | // From a CountedLoopNode I can reach the matching CountedLoopEndNode via the |
181 | // loop-back control. From CountedLoopEndNodes I can reach CountedLoopNodes |
182 | // via the old-trip-counter from the Op node. |
183 | |
184 | //------------------------------CountedLoopNode-------------------------------- |
185 | // CountedLoopNodes head simple counted loops. CountedLoopNodes have as |
186 | // inputs the incoming loop-start control and the loop-back control, so they |
187 | // act like RegionNodes. They also take in the initial trip counter, the |
188 | // loop-invariant stride and the loop-invariant limit value. CountedLoopNodes |
189 | // produce a loop-body control and the trip counter value. Since |
190 | // CountedLoopNodes behave like RegionNodes I still have a standard CFG model. |
191 | |
192 | class CountedLoopNode : public LoopNode { |
193 | // Size is bigger to hold _main_idx. However, _main_idx does not change |
194 | // the semantics so it does not appear in the hash & cmp functions. |
195 | virtual uint size_of() const { return sizeof(*this); } |
196 | |
197 | // For Pre- and Post-loops during debugging ONLY, this holds the index of |
198 | // the Main CountedLoop. Used to assert that we understand the graph shape. |
199 | node_idx_t _main_idx; |
200 | |
201 | // Known trip count calculated by compute_exact_trip_count() |
202 | uint _trip_count; |
203 | |
204 | // Log2 of original loop bodies in unrolled loop |
205 | int _unrolled_count_log2; |
206 | |
207 | // Node count prior to last unrolling - used to decide if |
208 | // unroll,optimize,unroll,optimize,... is making progress |
209 | int _node_count_before_unroll; |
210 | |
211 | // If slp analysis is performed we record the maximum |
212 | // vector mapped unroll factor here |
213 | int _slp_maximum_unroll_factor; |
214 | |
215 | public: |
216 | CountedLoopNode( Node *entry, Node *backedge ) |
217 | : LoopNode(entry, backedge), _main_idx(0), _trip_count(max_juint), |
218 | _unrolled_count_log2(0), _node_count_before_unroll(0), |
219 | _slp_maximum_unroll_factor(0) { |
220 | init_class_id(Class_CountedLoop); |
221 | // Initialize _trip_count to the largest possible value. |
222 | // Will be reset (lower) if the loop's trip count is known. |
223 | } |
224 | |
225 | virtual int Opcode() const; |
226 | virtual Node *Ideal(PhaseGVN *phase, bool can_reshape); |
227 | |
228 | Node *init_control() const { return in(EntryControl); } |
229 | Node *back_control() const { return in(LoopBackControl); } |
230 | CountedLoopEndNode *loopexit_or_null() const; |
231 | CountedLoopEndNode *loopexit() const; |
232 | Node *init_trip() const; |
233 | Node *stride() const; |
234 | int stride_con() const; |
235 | bool stride_is_con() const; |
236 | Node *limit() const; |
237 | Node *incr() const; |
238 | Node *phi() const; |
239 | |
240 | // Match increment with optional truncation |
241 | static Node* match_incr_with_optional_truncation(Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type); |
242 | |
243 | // A 'main' loop has a pre-loop and a post-loop. The 'main' loop |
244 | // can run short a few iterations and may start a few iterations in. |
245 | // It will be RCE'd and unrolled and aligned. |
246 | |
247 | // A following 'post' loop will run any remaining iterations. Used |
248 | // during Range Check Elimination, the 'post' loop will do any final |
249 | // iterations with full checks. Also used by Loop Unrolling, where |
250 | // the 'post' loop will do any epilog iterations needed. Basically, |
251 | // a 'post' loop can not profitably be further unrolled or RCE'd. |
252 | |
253 | // A preceding 'pre' loop will run at least 1 iteration (to do peeling), |
254 | // it may do under-flow checks for RCE and may do alignment iterations |
255 | // so the following main loop 'knows' that it is striding down cache |
256 | // lines. |
257 | |
258 | // A 'main' loop that is ONLY unrolled or peeled, never RCE'd or |
259 | // Aligned, may be missing it's pre-loop. |
260 | bool is_normal_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Normal; } |
261 | bool is_pre_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Pre; } |
262 | bool is_main_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Main; } |
263 | bool is_post_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Post; } |
264 | bool is_reduction_loop() const { return (_loop_flags&HasReductions) == HasReductions; } |
265 | bool was_slp_analyzed () const { return (_loop_flags&WasSlpAnalyzed) == WasSlpAnalyzed; } |
266 | bool has_passed_slp () const { return (_loop_flags&PassedSlpAnalysis) == PassedSlpAnalysis; } |
267 | bool is_unroll_only () const { return (_loop_flags&DoUnrollOnly) == DoUnrollOnly; } |
268 | bool is_main_no_pre_loop() const { return _loop_flags & MainHasNoPreLoop; } |
269 | bool has_atomic_post_loop () const { return (_loop_flags & HasAtomicPostLoop) == HasAtomicPostLoop; } |
270 | void set_main_no_pre_loop() { _loop_flags |= MainHasNoPreLoop; } |
271 | |
272 | int main_idx() const { return _main_idx; } |
273 | |
274 | |
275 | void set_pre_loop (CountedLoopNode *main) { assert(is_normal_loop(),"" ); _loop_flags |= Pre ; _main_idx = main->_idx; } |
276 | void set_main_loop ( ) { assert(is_normal_loop(),"" ); _loop_flags |= Main; } |
277 | void set_post_loop (CountedLoopNode *main) { assert(is_normal_loop(),"" ); _loop_flags |= Post; _main_idx = main->_idx; } |
278 | void set_normal_loop( ) { _loop_flags &= ~PreMainPostFlagsMask; } |
279 | |
280 | void set_trip_count(uint tc) { _trip_count = tc; } |
281 | uint trip_count() { return _trip_count; } |
282 | |
283 | bool has_exact_trip_count() const { return (_loop_flags & HasExactTripCount) != 0; } |
284 | void set_exact_trip_count(uint tc) { |
285 | _trip_count = tc; |
286 | _loop_flags |= HasExactTripCount; |
287 | } |
288 | void set_nonexact_trip_count() { |
289 | _loop_flags &= ~HasExactTripCount; |
290 | } |
291 | void set_notpassed_slp() { |
292 | _loop_flags &= ~PassedSlpAnalysis; |
293 | } |
294 | |
295 | void double_unrolled_count() { _unrolled_count_log2++; } |
296 | int unrolled_count() { return 1 << MIN2(_unrolled_count_log2, BitsPerInt-3); } |
297 | |
298 | void set_node_count_before_unroll(int ct) { _node_count_before_unroll = ct; } |
299 | int node_count_before_unroll() { return _node_count_before_unroll; } |
300 | void set_slp_max_unroll(int unroll_factor) { _slp_maximum_unroll_factor = unroll_factor; } |
301 | int slp_max_unroll() const { return _slp_maximum_unroll_factor; } |
302 | |
303 | virtual LoopNode* skip_strip_mined(int expect_skeleton = 1); |
304 | OuterStripMinedLoopNode* outer_loop() const; |
305 | virtual IfTrueNode* outer_loop_tail() const; |
306 | virtual OuterStripMinedLoopEndNode* outer_loop_end() const; |
307 | virtual IfFalseNode* outer_loop_exit() const; |
308 | virtual SafePointNode* outer_safepoint() const; |
309 | |
310 | // If this is a main loop in a pre/main/post loop nest, walk over |
311 | // the predicates that were inserted by |
312 | // duplicate_predicates()/add_range_check_predicate() |
313 | static Node* skip_predicates_from_entry(Node* ctrl); |
314 | Node* skip_predicates(); |
315 | |
316 | #ifndef PRODUCT |
317 | virtual void dump_spec(outputStream *st) const; |
318 | #endif |
319 | }; |
320 | |
321 | //------------------------------CountedLoopEndNode----------------------------- |
322 | // CountedLoopEndNodes end simple trip counted loops. They act much like |
323 | // IfNodes. |
324 | class CountedLoopEndNode : public IfNode { |
325 | public: |
326 | enum { TestControl, TestValue }; |
327 | |
328 | CountedLoopEndNode( Node *control, Node *test, float prob, float cnt ) |
329 | : IfNode( control, test, prob, cnt) { |
330 | init_class_id(Class_CountedLoopEnd); |
331 | } |
332 | virtual int Opcode() const; |
333 | |
334 | Node *cmp_node() const { return (in(TestValue)->req() >=2) ? in(TestValue)->in(1) : NULL; } |
335 | Node *incr() const { Node *tmp = cmp_node(); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; } |
336 | Node *limit() const { Node *tmp = cmp_node(); return (tmp && tmp->req()==3) ? tmp->in(2) : NULL; } |
337 | Node *stride() const { Node *tmp = incr (); return (tmp && tmp->req()==3) ? tmp->in(2) : NULL; } |
338 | Node *init_trip() const { Node *tmp = phi (); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; } |
339 | int stride_con() const; |
340 | bool stride_is_con() const { Node *tmp = stride (); return (tmp != NULL && tmp->is_Con()); } |
341 | BoolTest::mask test_trip() const { return in(TestValue)->as_Bool()->_test._test; } |
342 | PhiNode *phi() const { |
343 | Node *tmp = incr(); |
344 | if (tmp && tmp->req() == 3) { |
345 | Node* phi = tmp->in(1); |
346 | if (phi->is_Phi()) { |
347 | return phi->as_Phi(); |
348 | } |
349 | } |
350 | return NULL; |
351 | } |
352 | CountedLoopNode *loopnode() const { |
353 | // The CountedLoopNode that goes with this CountedLoopEndNode may |
354 | // have been optimized out by the IGVN so be cautious with the |
355 | // pattern matching on the graph |
356 | PhiNode* iv_phi = phi(); |
357 | if (iv_phi == NULL) { |
358 | return NULL; |
359 | } |
360 | Node *ln = iv_phi->in(0); |
361 | if (ln->is_CountedLoop() && ln->as_CountedLoop()->loopexit_or_null() == this) { |
362 | return (CountedLoopNode*)ln; |
363 | } |
364 | return NULL; |
365 | } |
366 | |
367 | #ifndef PRODUCT |
368 | virtual void dump_spec(outputStream *st) const; |
369 | #endif |
370 | }; |
371 | |
372 | |
373 | inline CountedLoopEndNode* CountedLoopNode::loopexit_or_null() const { |
374 | Node* bctrl = back_control(); |
375 | if (bctrl == NULL) return NULL; |
376 | |
377 | Node* lexit = bctrl->in(0); |
378 | return (CountedLoopEndNode*) |
379 | (lexit->Opcode() == Op_CountedLoopEnd ? lexit : NULL); |
380 | } |
381 | |
382 | inline CountedLoopEndNode* CountedLoopNode::loopexit() const { |
383 | CountedLoopEndNode* cle = loopexit_or_null(); |
384 | assert(cle != NULL, "loopexit is NULL" ); |
385 | return cle; |
386 | } |
387 | |
388 | inline Node* CountedLoopNode::init_trip() const { |
389 | CountedLoopEndNode* cle = loopexit_or_null(); |
390 | return cle != NULL ? cle->init_trip() : NULL; |
391 | } |
392 | inline Node* CountedLoopNode::stride() const { |
393 | CountedLoopEndNode* cle = loopexit_or_null(); |
394 | return cle != NULL ? cle->stride() : NULL; |
395 | } |
396 | inline int CountedLoopNode::stride_con() const { |
397 | CountedLoopEndNode* cle = loopexit_or_null(); |
398 | return cle != NULL ? cle->stride_con() : 0; |
399 | } |
400 | inline bool CountedLoopNode::stride_is_con() const { |
401 | CountedLoopEndNode* cle = loopexit_or_null(); |
402 | return cle != NULL && cle->stride_is_con(); |
403 | } |
404 | inline Node* CountedLoopNode::limit() const { |
405 | CountedLoopEndNode* cle = loopexit_or_null(); |
406 | return cle != NULL ? cle->limit() : NULL; |
407 | } |
408 | inline Node* CountedLoopNode::incr() const { |
409 | CountedLoopEndNode* cle = loopexit_or_null(); |
410 | return cle != NULL ? cle->incr() : NULL; |
411 | } |
412 | inline Node* CountedLoopNode::phi() const { |
413 | CountedLoopEndNode* cle = loopexit_or_null(); |
414 | return cle != NULL ? cle->phi() : NULL; |
415 | } |
416 | |
417 | //------------------------------LoopLimitNode----------------------------- |
418 | // Counted Loop limit node which represents exact final iterator value: |
419 | // trip_count = (limit - init_trip + stride - 1)/stride |
420 | // final_value= trip_count * stride + init_trip. |
421 | // Use HW instructions to calculate it when it can overflow in integer. |
422 | // Note, final_value should fit into integer since counted loop has |
423 | // limit check: limit <= max_int-stride. |
424 | class LoopLimitNode : public Node { |
425 | enum { Init=1, Limit=2, Stride=3 }; |
426 | public: |
427 | LoopLimitNode( Compile* C, Node *init, Node *limit, Node *stride ) : Node(0,init,limit,stride) { |
428 | // Put it on the Macro nodes list to optimize during macro nodes expansion. |
429 | init_flags(Flag_is_macro); |
430 | C->add_macro_node(this); |
431 | } |
432 | virtual int Opcode() const; |
433 | virtual const Type *bottom_type() const { return TypeInt::INT; } |
434 | virtual uint ideal_reg() const { return Op_RegI; } |
435 | virtual const Type* Value(PhaseGVN* phase) const; |
436 | virtual Node *Ideal(PhaseGVN *phase, bool can_reshape); |
437 | virtual Node* Identity(PhaseGVN* phase); |
438 | }; |
439 | |
440 | // Support for strip mining |
441 | class OuterStripMinedLoopNode : public LoopNode { |
442 | private: |
443 | CountedLoopNode* inner_loop() const; |
444 | public: |
445 | OuterStripMinedLoopNode(Compile* C, Node *entry, Node *backedge) |
446 | : LoopNode(entry, backedge) { |
447 | init_class_id(Class_OuterStripMinedLoop); |
448 | init_flags(Flag_is_macro); |
449 | C->add_macro_node(this); |
450 | } |
451 | |
452 | virtual int Opcode() const; |
453 | |
454 | virtual IfTrueNode* outer_loop_tail() const; |
455 | virtual OuterStripMinedLoopEndNode* outer_loop_end() const; |
456 | virtual IfFalseNode* outer_loop_exit() const; |
457 | virtual SafePointNode* outer_safepoint() const; |
458 | void adjust_strip_mined_loop(PhaseIterGVN* igvn); |
459 | }; |
460 | |
461 | class OuterStripMinedLoopEndNode : public IfNode { |
462 | public: |
463 | OuterStripMinedLoopEndNode(Node *control, Node *test, float prob, float cnt) |
464 | : IfNode(control, test, prob, cnt) { |
465 | init_class_id(Class_OuterStripMinedLoopEnd); |
466 | } |
467 | |
468 | virtual int Opcode() const; |
469 | |
470 | virtual const Type* Value(PhaseGVN* phase) const; |
471 | virtual Node *Ideal(PhaseGVN *phase, bool can_reshape); |
472 | }; |
473 | |
474 | // -----------------------------IdealLoopTree---------------------------------- |
475 | class IdealLoopTree : public ResourceObj { |
476 | public: |
477 | IdealLoopTree *_parent; // Parent in loop tree |
478 | IdealLoopTree *_next; // Next sibling in loop tree |
479 | IdealLoopTree *_child; // First child in loop tree |
480 | |
481 | // The head-tail backedge defines the loop. |
482 | // If a loop has multiple backedges, this is addressed during cleanup where |
483 | // we peel off the multiple backedges, merging all edges at the bottom and |
484 | // ensuring that one proper backedge flow into the loop. |
485 | Node *_head; // Head of loop |
486 | Node *_tail; // Tail of loop |
487 | inline Node *tail(); // Handle lazy update of _tail field |
488 | PhaseIdealLoop* _phase; |
489 | int _local_loop_unroll_limit; |
490 | int _local_loop_unroll_factor; |
491 | |
492 | Node_List _body; // Loop body for inner loops |
493 | |
494 | uint8_t _nest; // Nesting depth |
495 | uint8_t _irreducible:1, // True if irreducible |
496 | _has_call:1, // True if has call safepoint |
497 | _has_sfpt:1, // True if has non-call safepoint |
498 | _rce_candidate:1; // True if candidate for range check elimination |
499 | |
500 | Node_List* _safepts; // List of safepoints in this loop |
501 | Node_List* _required_safept; // A inner loop cannot delete these safepts; |
502 | bool _allow_optimizations; // Allow loop optimizations |
503 | |
504 | IdealLoopTree( PhaseIdealLoop* phase, Node *head, Node *tail ) |
505 | : _parent(0), _next(0), _child(0), |
506 | _head(head), _tail(tail), |
507 | _phase(phase), |
508 | _local_loop_unroll_limit(0), _local_loop_unroll_factor(0), |
509 | _nest(0), _irreducible(0), _has_call(0), _has_sfpt(0), _rce_candidate(0), |
510 | _safepts(NULL), |
511 | _required_safept(NULL), |
512 | _allow_optimizations(true) |
513 | { |
514 | precond(_head != NULL); |
515 | precond(_tail != NULL); |
516 | } |
517 | |
518 | // Is 'l' a member of 'this'? |
519 | bool is_member(const IdealLoopTree *l) const; // Test for nested membership |
520 | |
521 | // Set loop nesting depth. Accumulate has_call bits. |
522 | int set_nest( uint depth ); |
523 | |
524 | // Split out multiple fall-in edges from the loop header. Move them to a |
525 | // private RegionNode before the loop. This becomes the loop landing pad. |
526 | void split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt ); |
527 | |
528 | // Split out the outermost loop from this shared header. |
529 | void split_outer_loop( PhaseIdealLoop *phase ); |
530 | |
531 | // Merge all the backedges from the shared header into a private Region. |
532 | // Feed that region as the one backedge to this loop. |
533 | void merge_many_backedges( PhaseIdealLoop *phase ); |
534 | |
535 | // Split shared headers and insert loop landing pads. |
536 | // Insert a LoopNode to replace the RegionNode. |
537 | // Returns TRUE if loop tree is structurally changed. |
538 | bool beautify_loops( PhaseIdealLoop *phase ); |
539 | |
540 | // Perform optimization to use the loop predicates for null checks and range checks. |
541 | // Applies to any loop level (not just the innermost one) |
542 | bool loop_predication( PhaseIdealLoop *phase); |
543 | |
544 | // Perform iteration-splitting on inner loops. Split iterations to |
545 | // avoid range checks or one-shot null checks. Returns false if the |
546 | // current round of loop opts should stop. |
547 | bool iteration_split( PhaseIdealLoop *phase, Node_List &old_new ); |
548 | |
549 | // Driver for various flavors of iteration splitting. Returns false |
550 | // if the current round of loop opts should stop. |
551 | bool iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new ); |
552 | |
553 | // Given dominators, try to find loops with calls that must always be |
554 | // executed (call dominates loop tail). These loops do not need non-call |
555 | // safepoints (ncsfpt). |
556 | void check_safepts(VectorSet &visited, Node_List &stack); |
557 | |
558 | // Allpaths backwards scan from loop tail, terminating each path at first safepoint |
559 | // encountered. |
560 | void allpaths_check_safepts(VectorSet &visited, Node_List &stack); |
561 | |
562 | // Remove safepoints from loop. Optionally keeping one. |
563 | void remove_safepoints(PhaseIdealLoop* phase, bool keep_one); |
564 | |
565 | // Convert to counted loops where possible |
566 | void counted_loop( PhaseIdealLoop *phase ); |
567 | |
568 | // Check for Node being a loop-breaking test |
569 | Node *is_loop_exit(Node *iff) const; |
570 | |
571 | // Remove simplistic dead code from loop body |
572 | void DCE_loop_body(); |
573 | |
574 | // Look for loop-exit tests with my 50/50 guesses from the Parsing stage. |
575 | // Replace with a 1-in-10 exit guess. |
576 | void adjust_loop_exit_prob( PhaseIdealLoop *phase ); |
577 | |
578 | // Return TRUE or FALSE if the loop should never be RCE'd or aligned. |
579 | // Useful for unrolling loops with NO array accesses. |
580 | bool policy_peel_only( PhaseIdealLoop *phase ) const; |
581 | |
582 | // Return TRUE or FALSE if the loop should be unswitched -- clone |
583 | // loop with an invariant test |
584 | bool policy_unswitching( PhaseIdealLoop *phase ) const; |
585 | |
586 | // Micro-benchmark spamming. Remove empty loops. |
587 | bool do_remove_empty_loop( PhaseIdealLoop *phase ); |
588 | |
589 | // Convert one iteration loop into normal code. |
590 | bool do_one_iteration_loop( PhaseIdealLoop *phase ); |
591 | |
592 | // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can |
593 | // move some loop-invariant test (usually a null-check) before the loop. |
594 | bool policy_peeling(PhaseIdealLoop *phase); |
595 | |
596 | uint estimate_peeling(PhaseIdealLoop *phase); |
597 | |
598 | // Return TRUE or FALSE if the loop should be maximally unrolled. Stash any |
599 | // known trip count in the counted loop node. |
600 | bool policy_maximally_unroll(PhaseIdealLoop *phase) const; |
601 | |
602 | // Return TRUE or FALSE if the loop should be unrolled or not. Apply unroll |
603 | // if the loop is a counted loop and the loop body is small enough. |
604 | bool policy_unroll(PhaseIdealLoop *phase); |
605 | |
606 | // Loop analyses to map to a maximal superword unrolling for vectorization. |
607 | void policy_unroll_slp_analysis(CountedLoopNode *cl, PhaseIdealLoop *phase, int future_unroll_ct); |
608 | |
609 | // Return TRUE or FALSE if the loop should be range-check-eliminated. |
610 | // Gather a list of IF tests that are dominated by iteration splitting; |
611 | // also gather the end of the first split and the start of the 2nd split. |
612 | bool policy_range_check( PhaseIdealLoop *phase ) const; |
613 | |
614 | // Return TRUE or FALSE if the loop should be cache-line aligned. |
615 | // Gather the expression that does the alignment. Note that only |
616 | // one array base can be aligned in a loop (unless the VM guarantees |
617 | // mutual alignment). Note that if we vectorize short memory ops |
618 | // into longer memory ops, we may want to increase alignment. |
619 | bool policy_align( PhaseIdealLoop *phase ) const; |
620 | |
621 | // Return TRUE if "iff" is a range check. |
622 | bool is_range_check_if(IfNode *iff, PhaseIdealLoop *phase, Invariance& invar) const; |
623 | |
624 | // Estimate the number of nodes required when cloning a loop (body). |
625 | uint est_loop_clone_sz(uint factor) const; |
626 | |
627 | // Compute loop trip count if possible |
628 | void compute_trip_count(PhaseIdealLoop* phase); |
629 | |
630 | // Compute loop trip count from profile data |
631 | float compute_profile_trip_cnt_helper(Node* n); |
632 | void compute_profile_trip_cnt( PhaseIdealLoop *phase ); |
633 | |
634 | // Reassociate invariant expressions. |
635 | void reassociate_invariants(PhaseIdealLoop *phase); |
636 | // Reassociate invariant add and subtract expressions. |
637 | Node* reassociate_add_sub(Node* n1, PhaseIdealLoop *phase); |
638 | // Return nonzero index of invariant operand if invariant and variant |
639 | // are combined with an Add or Sub. Helper for reassociate_invariants. |
640 | int is_invariant_addition(Node* n, PhaseIdealLoop *phase); |
641 | |
642 | // Return true if n is invariant |
643 | bool is_invariant(Node* n) const; |
644 | |
645 | // Put loop body on igvn work list |
646 | void record_for_igvn(); |
647 | |
648 | bool is_root() { return _parent == NULL; } |
649 | // A proper/reducible loop w/o any (occasional) dead back-edge. |
650 | bool is_loop() { return !_irreducible && !tail()->is_top(); } |
651 | bool is_counted() { return is_loop() && _head->is_CountedLoop(); } |
652 | bool is_innermost() { return is_loop() && _child == NULL; } |
653 | |
654 | void remove_main_post_loops(CountedLoopNode *cl, PhaseIdealLoop *phase); |
655 | |
656 | #ifndef PRODUCT |
657 | void dump_head( ) const; // Dump loop head only |
658 | void dump() const; // Dump this loop recursively |
659 | void verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const; |
660 | #endif |
661 | |
662 | }; |
663 | |
664 | // -----------------------------PhaseIdealLoop--------------------------------- |
665 | // Computes the mapping from Nodes to IdealLoopTrees. Organizes IdealLoopTrees |
666 | // into a loop tree. Drives the loop-based transformations on the ideal graph. |
667 | class PhaseIdealLoop : public PhaseTransform { |
668 | friend class IdealLoopTree; |
669 | friend class SuperWord; |
670 | friend class CountedLoopReserveKit; |
671 | friend class ShenandoahBarrierC2Support; |
672 | friend class AutoNodeBudget; |
673 | |
674 | // Pre-computed def-use info |
675 | PhaseIterGVN &_igvn; |
676 | |
677 | // Head of loop tree |
678 | IdealLoopTree *_ltree_root; |
679 | |
680 | // Array of pre-order numbers, plus post-visited bit. |
681 | // ZERO for not pre-visited. EVEN for pre-visited but not post-visited. |
682 | // ODD for post-visited. Other bits are the pre-order number. |
683 | uint *_preorders; |
684 | uint _max_preorder; |
685 | |
686 | const PhaseIdealLoop* _verify_me; |
687 | bool _verify_only; |
688 | |
689 | // Allocate _preorders[] array |
690 | void allocate_preorders() { |
691 | _max_preorder = C->unique()+8; |
692 | _preorders = NEW_RESOURCE_ARRAY(uint, _max_preorder); |
693 | memset(_preorders, 0, sizeof(uint) * _max_preorder); |
694 | } |
695 | |
696 | // Allocate _preorders[] array |
697 | void reallocate_preorders() { |
698 | if ( _max_preorder < C->unique() ) { |
699 | _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, C->unique()); |
700 | _max_preorder = C->unique(); |
701 | } |
702 | memset(_preorders, 0, sizeof(uint) * _max_preorder); |
703 | } |
704 | |
705 | // Check to grow _preorders[] array for the case when build_loop_tree_impl() |
706 | // adds new nodes. |
707 | void check_grow_preorders( ) { |
708 | if ( _max_preorder < C->unique() ) { |
709 | uint newsize = _max_preorder<<1; // double size of array |
710 | _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, newsize); |
711 | memset(&_preorders[_max_preorder],0,sizeof(uint)*(newsize-_max_preorder)); |
712 | _max_preorder = newsize; |
713 | } |
714 | } |
715 | // Check for pre-visited. Zero for NOT visited; non-zero for visited. |
716 | int is_visited( Node *n ) const { return _preorders[n->_idx]; } |
717 | // Pre-order numbers are written to the Nodes array as low-bit-set values. |
718 | void set_preorder_visited( Node *n, int pre_order ) { |
719 | assert( !is_visited( n ), "already set" ); |
720 | _preorders[n->_idx] = (pre_order<<1); |
721 | }; |
722 | // Return pre-order number. |
723 | int get_preorder( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]>>1; } |
724 | |
725 | // Check for being post-visited. |
726 | // Should be previsited already (checked with assert(is_visited(n))). |
727 | int is_postvisited( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]&1; } |
728 | |
729 | // Mark as post visited |
730 | void set_postvisited( Node *n ) { assert( !is_postvisited( n ), "" ); _preorders[n->_idx] |= 1; } |
731 | |
732 | public: |
733 | // Set/get control node out. Set lower bit to distinguish from IdealLoopTree |
734 | // Returns true if "n" is a data node, false if it's a control node. |
735 | bool has_ctrl( Node *n ) const { return ((intptr_t)_nodes[n->_idx]) & 1; } |
736 | |
737 | private: |
738 | // clear out dead code after build_loop_late |
739 | Node_List _deadlist; |
740 | |
741 | // Support for faster execution of get_late_ctrl()/dom_lca() |
742 | // when a node has many uses and dominator depth is deep. |
743 | Node_Array _dom_lca_tags; |
744 | void init_dom_lca_tags(); |
745 | void clear_dom_lca_tags(); |
746 | |
747 | // Helper for debugging bad dominance relationships |
748 | bool verify_dominance(Node* n, Node* use, Node* LCA, Node* early); |
749 | |
750 | Node* compute_lca_of_uses(Node* n, Node* early, bool verify = false); |
751 | |
752 | // Inline wrapper for frequent cases: |
753 | // 1) only one use |
754 | // 2) a use is the same as the current LCA passed as 'n1' |
755 | Node *dom_lca_for_get_late_ctrl( Node *lca, Node *n, Node *tag ) { |
756 | assert( n->is_CFG(), "" ); |
757 | // Fast-path NULL lca |
758 | if( lca != NULL && lca != n ) { |
759 | assert( lca->is_CFG(), "" ); |
760 | // find LCA of all uses |
761 | n = dom_lca_for_get_late_ctrl_internal( lca, n, tag ); |
762 | } |
763 | return find_non_split_ctrl(n); |
764 | } |
765 | Node *dom_lca_for_get_late_ctrl_internal( Node *lca, Node *n, Node *tag ); |
766 | |
767 | // Helper function for directing control inputs away from CFG split points. |
768 | Node *find_non_split_ctrl( Node *ctrl ) const { |
769 | if (ctrl != NULL) { |
770 | if (ctrl->is_MultiBranch()) { |
771 | ctrl = ctrl->in(0); |
772 | } |
773 | assert(ctrl->is_CFG(), "CFG" ); |
774 | } |
775 | return ctrl; |
776 | } |
777 | |
778 | Node* cast_incr_before_loop(Node* incr, Node* ctrl, Node* loop); |
779 | void duplicate_predicates_helper(Node* predicate, Node* start, Node* end, IdealLoopTree* outer_loop, |
780 | LoopNode* outer_main_head, uint dd_main_head); |
781 | void duplicate_predicates(CountedLoopNode* pre_head, Node* start, Node* end, IdealLoopTree* outer_loop, |
782 | LoopNode* outer_main_head, uint dd_main_head); |
783 | Node* clone_skeleton_predicate(Node* iff, Node* value, Node* predicate, Node* uncommon_proj, |
784 | Node* current_proj, IdealLoopTree* outer_loop, Node* prev_proj); |
785 | bool skeleton_predicate_has_opaque(IfNode* iff); |
786 | void update_skeleton_predicates(Node* ctrl, CountedLoopNode* loop_head, Node* init, int stride_con); |
787 | void insert_loop_limit_check(ProjNode* limit_check_proj, Node* cmp_limit, Node* bol); |
788 | |
789 | public: |
790 | |
791 | PhaseIterGVN &igvn() const { return _igvn; } |
792 | |
793 | static bool is_canonical_loop_entry(CountedLoopNode* cl); |
794 | |
795 | bool has_node( Node* n ) const { |
796 | guarantee(n != NULL, "No Node." ); |
797 | return _nodes[n->_idx] != NULL; |
798 | } |
799 | // check if transform created new nodes that need _ctrl recorded |
800 | Node *get_late_ctrl( Node *n, Node *early ); |
801 | Node *get_early_ctrl( Node *n ); |
802 | Node *get_early_ctrl_for_expensive(Node *n, Node* earliest); |
803 | void set_early_ctrl( Node *n ); |
804 | void set_subtree_ctrl( Node *root ); |
805 | void set_ctrl( Node *n, Node *ctrl ) { |
806 | assert( !has_node(n) || has_ctrl(n), "" ); |
807 | assert( ctrl->in(0), "cannot set dead control node" ); |
808 | assert( ctrl == find_non_split_ctrl(ctrl), "must set legal crtl" ); |
809 | _nodes.map( n->_idx, (Node*)((intptr_t)ctrl + 1) ); |
810 | } |
811 | // Set control and update loop membership |
812 | void set_ctrl_and_loop(Node* n, Node* ctrl) { |
813 | IdealLoopTree* old_loop = get_loop(get_ctrl(n)); |
814 | IdealLoopTree* new_loop = get_loop(ctrl); |
815 | if (old_loop != new_loop) { |
816 | if (old_loop->_child == NULL) old_loop->_body.yank(n); |
817 | if (new_loop->_child == NULL) new_loop->_body.push(n); |
818 | } |
819 | set_ctrl(n, ctrl); |
820 | } |
821 | // Control nodes can be replaced or subsumed. During this pass they |
822 | // get their replacement Node in slot 1. Instead of updating the block |
823 | // location of all Nodes in the subsumed block, we lazily do it. As we |
824 | // pull such a subsumed block out of the array, we write back the final |
825 | // correct block. |
826 | Node *get_ctrl( Node *i ) { |
827 | |
828 | assert(has_node(i), "" ); |
829 | Node *n = get_ctrl_no_update(i); |
830 | _nodes.map( i->_idx, (Node*)((intptr_t)n + 1) ); |
831 | assert(has_node(i) && has_ctrl(i), "" ); |
832 | assert(n == find_non_split_ctrl(n), "must return legal ctrl" ); |
833 | return n; |
834 | } |
835 | // true if CFG node d dominates CFG node n |
836 | bool is_dominator(Node *d, Node *n); |
837 | // return get_ctrl for a data node and self(n) for a CFG node |
838 | Node* ctrl_or_self(Node* n) { |
839 | if (has_ctrl(n)) |
840 | return get_ctrl(n); |
841 | else { |
842 | assert (n->is_CFG(), "must be a CFG node" ); |
843 | return n; |
844 | } |
845 | } |
846 | |
847 | Node *get_ctrl_no_update_helper(Node *i) const { |
848 | assert(has_ctrl(i), "should be control, not loop" ); |
849 | return (Node*)(((intptr_t)_nodes[i->_idx]) & ~1); |
850 | } |
851 | |
852 | Node *get_ctrl_no_update(Node *i) const { |
853 | assert( has_ctrl(i), "" ); |
854 | Node *n = get_ctrl_no_update_helper(i); |
855 | if (!n->in(0)) { |
856 | // Skip dead CFG nodes |
857 | do { |
858 | n = get_ctrl_no_update_helper(n); |
859 | } while (!n->in(0)); |
860 | n = find_non_split_ctrl(n); |
861 | } |
862 | return n; |
863 | } |
864 | |
865 | // Check for loop being set |
866 | // "n" must be a control node. Returns true if "n" is known to be in a loop. |
867 | bool has_loop( Node *n ) const { |
868 | assert(!has_node(n) || !has_ctrl(n), "" ); |
869 | return has_node(n); |
870 | } |
871 | // Set loop |
872 | void set_loop( Node *n, IdealLoopTree *loop ) { |
873 | _nodes.map(n->_idx, (Node*)loop); |
874 | } |
875 | // Lazy-dazy update of 'get_ctrl' and 'idom_at' mechanisms. Replace |
876 | // the 'old_node' with 'new_node'. Kill old-node. Add a reference |
877 | // from old_node to new_node to support the lazy update. Reference |
878 | // replaces loop reference, since that is not needed for dead node. |
879 | void lazy_update(Node *old_node, Node *new_node) { |
880 | assert(old_node != new_node, "no cycles please" ); |
881 | // Re-use the side array slot for this node to provide the |
882 | // forwarding pointer. |
883 | _nodes.map(old_node->_idx, (Node*)((intptr_t)new_node + 1)); |
884 | } |
885 | void lazy_replace(Node *old_node, Node *new_node) { |
886 | _igvn.replace_node(old_node, new_node); |
887 | lazy_update(old_node, new_node); |
888 | } |
889 | |
890 | private: |
891 | |
892 | // Place 'n' in some loop nest, where 'n' is a CFG node |
893 | void build_loop_tree(); |
894 | int build_loop_tree_impl( Node *n, int pre_order ); |
895 | // Insert loop into the existing loop tree. 'innermost' is a leaf of the |
896 | // loop tree, not the root. |
897 | IdealLoopTree *sort( IdealLoopTree *loop, IdealLoopTree *innermost ); |
898 | |
899 | // Place Data nodes in some loop nest |
900 | void build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ); |
901 | void build_loop_late ( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ); |
902 | void build_loop_late_post_work(Node* n, bool pinned); |
903 | void build_loop_late_post(Node* n); |
904 | void verify_strip_mined_scheduling(Node *n, Node* least); |
905 | |
906 | // Array of immediate dominance info for each CFG node indexed by node idx |
907 | private: |
908 | uint _idom_size; |
909 | Node **_idom; // Array of immediate dominators |
910 | uint *_dom_depth; // Used for fast LCA test |
911 | GrowableArray<uint>* _dom_stk; // For recomputation of dom depth |
912 | |
913 | // Perform verification that the graph is valid. |
914 | PhaseIdealLoop( PhaseIterGVN &igvn) : |
915 | PhaseTransform(Ideal_Loop), |
916 | _igvn(igvn), |
917 | _verify_me(NULL), |
918 | _verify_only(true), |
919 | _dom_lca_tags(arena()), // Thread::resource_area |
920 | _nodes_required(UINT_MAX) { |
921 | build_and_optimize(LoopOptsVerify); |
922 | } |
923 | |
924 | // build the loop tree and perform any requested optimizations |
925 | void build_and_optimize(LoopOptsMode mode); |
926 | |
927 | // Dominators for the sea of nodes |
928 | void Dominators(); |
929 | |
930 | // Compute the Ideal Node to Loop mapping |
931 | PhaseIdealLoop(PhaseIterGVN &igvn, LoopOptsMode mode) : |
932 | PhaseTransform(Ideal_Loop), |
933 | _igvn(igvn), |
934 | _verify_me(NULL), |
935 | _verify_only(false), |
936 | _dom_lca_tags(arena()), // Thread::resource_area |
937 | _nodes_required(UINT_MAX) { |
938 | build_and_optimize(mode); |
939 | } |
940 | |
941 | // Verify that verify_me made the same decisions as a fresh run. |
942 | PhaseIdealLoop(PhaseIterGVN &igvn, const PhaseIdealLoop *verify_me) : |
943 | PhaseTransform(Ideal_Loop), |
944 | _igvn(igvn), |
945 | _verify_me(verify_me), |
946 | _verify_only(false), |
947 | _dom_lca_tags(arena()), // Thread::resource_area |
948 | _nodes_required(UINT_MAX) { |
949 | build_and_optimize(LoopOptsVerify); |
950 | } |
951 | |
952 | public: |
953 | Node* idom_no_update(Node* d) const { |
954 | return idom_no_update(d->_idx); |
955 | } |
956 | |
957 | Node* idom_no_update(uint didx) const { |
958 | assert(didx < _idom_size, "oob" ); |
959 | Node* n = _idom[didx]; |
960 | assert(n != NULL,"Bad immediate dominator info." ); |
961 | while (n->in(0) == NULL) { // Skip dead CFG nodes |
962 | n = (Node*)(((intptr_t)_nodes[n->_idx]) & ~1); |
963 | assert(n != NULL,"Bad immediate dominator info." ); |
964 | } |
965 | return n; |
966 | } |
967 | |
968 | Node *idom(Node* d) const { |
969 | return idom(d->_idx); |
970 | } |
971 | |
972 | Node *idom(uint didx) const { |
973 | Node *n = idom_no_update(didx); |
974 | _idom[didx] = n; // Lazily remove dead CFG nodes from table. |
975 | return n; |
976 | } |
977 | |
978 | uint dom_depth(Node* d) const { |
979 | guarantee(d != NULL, "Null dominator info." ); |
980 | guarantee(d->_idx < _idom_size, "" ); |
981 | return _dom_depth[d->_idx]; |
982 | } |
983 | void set_idom(Node* d, Node* n, uint dom_depth); |
984 | // Locally compute IDOM using dom_lca call |
985 | Node *compute_idom( Node *region ) const; |
986 | // Recompute dom_depth |
987 | void recompute_dom_depth(); |
988 | |
989 | // Is safept not required by an outer loop? |
990 | bool is_deleteable_safept(Node* sfpt); |
991 | |
992 | // Replace parallel induction variable (parallel to trip counter) |
993 | void replace_parallel_iv(IdealLoopTree *loop); |
994 | |
995 | Node *dom_lca( Node *n1, Node *n2 ) const { |
996 | return find_non_split_ctrl(dom_lca_internal(n1, n2)); |
997 | } |
998 | Node *dom_lca_internal( Node *n1, Node *n2 ) const; |
999 | |
1000 | // Build and verify the loop tree without modifying the graph. This |
1001 | // is useful to verify that all inputs properly dominate their uses. |
1002 | static void verify(PhaseIterGVN& igvn) { |
1003 | #ifdef ASSERT |
1004 | ResourceMark rm; |
1005 | PhaseIdealLoop v(igvn); |
1006 | #endif |
1007 | } |
1008 | |
1009 | // Recommended way to use PhaseIdealLoop. |
1010 | // Run PhaseIdealLoop in some mode and allocates a local scope for memory allocations. |
1011 | static void optimize(PhaseIterGVN &igvn, LoopOptsMode mode) { |
1012 | ResourceMark rm; |
1013 | PhaseIdealLoop v(igvn, mode); |
1014 | } |
1015 | |
1016 | // True if the method has at least 1 irreducible loop |
1017 | bool _has_irreducible_loops; |
1018 | |
1019 | // Per-Node transform |
1020 | virtual Node *transform( Node *a_node ) { return 0; } |
1021 | |
1022 | bool is_counted_loop(Node* x, IdealLoopTree*& loop); |
1023 | IdealLoopTree* create_outer_strip_mined_loop(BoolNode *test, Node *cmp, Node *init_control, |
1024 | IdealLoopTree* loop, float cl_prob, float le_fcnt, |
1025 | Node*& entry_control, Node*& iffalse); |
1026 | |
1027 | Node* exact_limit( IdealLoopTree *loop ); |
1028 | |
1029 | // Return a post-walked LoopNode |
1030 | IdealLoopTree *get_loop( Node *n ) const { |
1031 | // Dead nodes have no loop, so return the top level loop instead |
1032 | if (!has_node(n)) return _ltree_root; |
1033 | assert(!has_ctrl(n), "" ); |
1034 | return (IdealLoopTree*)_nodes[n->_idx]; |
1035 | } |
1036 | |
1037 | IdealLoopTree *ltree_root() const { return _ltree_root; } |
1038 | |
1039 | // Is 'n' a (nested) member of 'loop'? |
1040 | int is_member( const IdealLoopTree *loop, Node *n ) const { |
1041 | return loop->is_member(get_loop(n)); } |
1042 | |
1043 | // This is the basic building block of the loop optimizations. It clones an |
1044 | // entire loop body. It makes an old_new loop body mapping; with this |
1045 | // mapping you can find the new-loop equivalent to an old-loop node. All |
1046 | // new-loop nodes are exactly equal to their old-loop counterparts, all |
1047 | // edges are the same. All exits from the old-loop now have a RegionNode |
1048 | // that merges the equivalent new-loop path. This is true even for the |
1049 | // normal "loop-exit" condition. All uses of loop-invariant old-loop values |
1050 | // now come from (one or more) Phis that merge their new-loop equivalents. |
1051 | // Parameter side_by_side_idom: |
1052 | // When side_by_size_idom is NULL, the dominator tree is constructed for |
1053 | // the clone loop to dominate the original. Used in construction of |
1054 | // pre-main-post loop sequence. |
1055 | // When nonnull, the clone and original are side-by-side, both are |
1056 | // dominated by the passed in side_by_side_idom node. Used in |
1057 | // construction of unswitched loops. |
1058 | enum CloneLoopMode { |
1059 | IgnoreStripMined = 0, // Only clone inner strip mined loop |
1060 | CloneIncludesStripMined = 1, // clone both inner and outer strip mined loops |
1061 | ControlAroundStripMined = 2 // Only clone inner strip mined loop, |
1062 | // result control flow branches |
1063 | // either to inner clone or outer |
1064 | // strip mined loop. |
1065 | }; |
1066 | void clone_loop( IdealLoopTree *loop, Node_List &old_new, int dom_depth, |
1067 | CloneLoopMode mode, Node* side_by_side_idom = NULL); |
1068 | void clone_loop_handle_data_uses(Node* old, Node_List &old_new, |
1069 | IdealLoopTree* loop, IdealLoopTree* companion_loop, |
1070 | Node_List*& split_if_set, Node_List*& split_bool_set, |
1071 | Node_List*& split_cex_set, Node_List& worklist, |
1072 | uint new_counter, CloneLoopMode mode); |
1073 | void clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop, |
1074 | IdealLoopTree* outer_loop, int dd, Node_List &old_new, |
1075 | Node_List& ); |
1076 | |
1077 | // If we got the effect of peeling, either by actually peeling or by |
1078 | // making a pre-loop which must execute at least once, we can remove |
1079 | // all loop-invariant dominated tests in the main body. |
1080 | void peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new ); |
1081 | |
1082 | // Generate code to do a loop peel for the given loop (and body). |
1083 | // old_new is a temp array. |
1084 | void do_peeling( IdealLoopTree *loop, Node_List &old_new ); |
1085 | |
1086 | // Add pre and post loops around the given loop. These loops are used |
1087 | // during RCE, unrolling and aligning loops. |
1088 | void insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only ); |
1089 | |
1090 | // Add post loop after the given loop. |
1091 | Node *insert_post_loop(IdealLoopTree *loop, Node_List &old_new, |
1092 | CountedLoopNode *main_head, CountedLoopEndNode *main_end, |
1093 | Node *incr, Node *limit, CountedLoopNode *&post_head); |
1094 | |
1095 | // Add an RCE'd post loop which we will multi-version adapt for run time test path usage |
1096 | void insert_scalar_rced_post_loop( IdealLoopTree *loop, Node_List &old_new ); |
1097 | |
1098 | // Add a vector post loop between a vector main loop and the current post loop |
1099 | void insert_vector_post_loop(IdealLoopTree *loop, Node_List &old_new); |
1100 | // If Node n lives in the back_ctrl block, we clone a private version of n |
1101 | // in preheader_ctrl block and return that, otherwise return n. |
1102 | Node *clone_up_backedge_goo( Node *back_ctrl, Node *, Node *n, VectorSet &visited, Node_Stack &clones ); |
1103 | |
1104 | // Take steps to maximally unroll the loop. Peel any odd iterations, then |
1105 | // unroll to do double iterations. The next round of major loop transforms |
1106 | // will repeat till the doubled loop body does all remaining iterations in 1 |
1107 | // pass. |
1108 | void do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new ); |
1109 | |
1110 | // Unroll the loop body one step - make each trip do 2 iterations. |
1111 | void do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip ); |
1112 | |
1113 | // Mark vector reduction candidates before loop unrolling |
1114 | void mark_reductions( IdealLoopTree *loop ); |
1115 | |
1116 | // Return true if exp is a constant times an induction var |
1117 | bool is_scaled_iv(Node* exp, Node* iv, int* p_scale); |
1118 | |
1119 | // Return true if exp is a scaled induction var plus (or minus) constant |
1120 | bool is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth = 0); |
1121 | |
1122 | // Create a new if above the uncommon_trap_if_pattern for the predicate to be promoted |
1123 | ProjNode* create_new_if_for_predicate(ProjNode* cont_proj, Node* new_entry, |
1124 | Deoptimization::DeoptReason reason, |
1125 | int opcode); |
1126 | void register_control(Node* n, IdealLoopTree *loop, Node* pred); |
1127 | |
1128 | // Clone loop predicates to cloned loops (peeled, unswitched) |
1129 | static ProjNode* clone_predicate(ProjNode* predicate_proj, Node* new_entry, |
1130 | Deoptimization::DeoptReason reason, |
1131 | PhaseIdealLoop* loop_phase, |
1132 | PhaseIterGVN* igvn); |
1133 | |
1134 | static void clone_loop_predicates_fix_mem(ProjNode* dom_proj , ProjNode* proj, |
1135 | PhaseIdealLoop* loop_phase, |
1136 | PhaseIterGVN* igvn); |
1137 | |
1138 | static Node* clone_loop_predicates(Node* old_entry, Node* new_entry, |
1139 | bool clone_limit_check, |
1140 | PhaseIdealLoop* loop_phase, |
1141 | PhaseIterGVN* igvn); |
1142 | Node* clone_loop_predicates(Node* old_entry, Node* new_entry, bool clone_limit_check); |
1143 | |
1144 | static Node* skip_all_loop_predicates(Node* entry); |
1145 | static Node* skip_loop_predicates(Node* entry); |
1146 | |
1147 | // Find a good location to insert a predicate |
1148 | static ProjNode* find_predicate_insertion_point(Node* start_c, Deoptimization::DeoptReason reason); |
1149 | // Find a predicate |
1150 | static Node* find_predicate(Node* entry); |
1151 | // Construct a range check for a predicate if |
1152 | BoolNode* rc_predicate(IdealLoopTree *loop, Node* ctrl, |
1153 | int scale, Node* offset, |
1154 | Node* init, Node* limit, jint stride, |
1155 | Node* range, bool upper, bool &overflow); |
1156 | |
1157 | // Implementation of the loop predication to promote checks outside the loop |
1158 | bool loop_predication_impl(IdealLoopTree *loop); |
1159 | bool loop_predication_impl_helper(IdealLoopTree *loop, ProjNode* proj, ProjNode *predicate_proj, |
1160 | CountedLoopNode *cl, ConNode* zero, Invariance& invar, |
1161 | Deoptimization::DeoptReason reason); |
1162 | bool loop_predication_should_follow_branches(IdealLoopTree *loop, ProjNode *predicate_proj, float& loop_trip_cnt); |
1163 | void loop_predication_follow_branches(Node *c, IdealLoopTree *loop, float loop_trip_cnt, |
1164 | PathFrequency& pf, Node_Stack& stack, VectorSet& seen, |
1165 | Node_List& if_proj_list); |
1166 | ProjNode* insert_skeleton_predicate(IfNode* iff, IdealLoopTree *loop, |
1167 | ProjNode* proj, ProjNode *predicate_proj, |
1168 | ProjNode* upper_bound_proj, |
1169 | int scale, Node* offset, |
1170 | Node* init, Node* limit, jint stride, |
1171 | Node* rng, bool& overflow, |
1172 | Deoptimization::DeoptReason reason); |
1173 | Node* add_range_check_predicate(IdealLoopTree* loop, CountedLoopNode* cl, |
1174 | Node* predicate_proj, int scale_con, Node* offset, |
1175 | Node* limit, jint stride_con, Node* value); |
1176 | |
1177 | // Helper function to collect predicate for eliminating the useless ones |
1178 | void collect_potentially_useful_predicates(IdealLoopTree *loop, Unique_Node_List &predicate_opaque1); |
1179 | void eliminate_useless_predicates(); |
1180 | |
1181 | // Change the control input of expensive nodes to allow commoning by |
1182 | // IGVN when it is guaranteed to not result in a more frequent |
1183 | // execution of the expensive node. Return true if progress. |
1184 | bool process_expensive_nodes(); |
1185 | |
1186 | // Check whether node has become unreachable |
1187 | bool is_node_unreachable(Node *n) const { |
1188 | return !has_node(n) || n->is_unreachable(_igvn); |
1189 | } |
1190 | |
1191 | // Eliminate range-checks and other trip-counter vs loop-invariant tests. |
1192 | int do_range_check( IdealLoopTree *loop, Node_List &old_new ); |
1193 | |
1194 | // Check to see if do_range_check(...) cleaned the main loop of range-checks |
1195 | void has_range_checks(IdealLoopTree *loop); |
1196 | |
1197 | // Process post loops which have range checks and try to build a multi-version |
1198 | // guard to safely determine if we can execute the post loop which was RCE'd. |
1199 | bool multi_version_post_loops(IdealLoopTree *rce_loop, IdealLoopTree *legacy_loop); |
1200 | |
1201 | // Cause the rce'd post loop to optimized away, this happens if we cannot complete multiverioning |
1202 | void poison_rce_post_loop(IdealLoopTree *rce_loop); |
1203 | |
1204 | // Create a slow version of the loop by cloning the loop |
1205 | // and inserting an if to select fast-slow versions. |
1206 | ProjNode* create_slow_version_of_loop(IdealLoopTree *loop, |
1207 | Node_List &old_new, |
1208 | int opcode, |
1209 | CloneLoopMode mode); |
1210 | |
1211 | // Clone a loop and return the clone head (clone_loop_head). |
1212 | // Added nodes include int(1), int(0) - disconnected, If, IfTrue, IfFalse, |
1213 | // This routine was created for usage in CountedLoopReserveKit. |
1214 | // |
1215 | // int(1) -> If -> IfTrue -> original_loop_head |
1216 | // | |
1217 | // V |
1218 | // IfFalse -> clone_loop_head (returned by function pointer) |
1219 | // |
1220 | LoopNode* create_reserve_version_of_loop(IdealLoopTree *loop, CountedLoopReserveKit* lk); |
1221 | // Clone loop with an invariant test (that does not exit) and |
1222 | // insert a clone of the test that selects which version to |
1223 | // execute. |
1224 | void do_unswitching (IdealLoopTree *loop, Node_List &old_new); |
1225 | |
1226 | // Find candidate "if" for unswitching |
1227 | IfNode* find_unswitching_candidate(const IdealLoopTree *loop) const; |
1228 | |
1229 | // Range Check Elimination uses this function! |
1230 | // Constrain the main loop iterations so the affine function: |
1231 | // low_limit <= scale_con * I + offset < upper_limit |
1232 | // always holds true. That is, either increase the number of iterations in |
1233 | // the pre-loop or the post-loop until the condition holds true in the main |
1234 | // loop. Scale_con, offset and limit are all loop invariant. |
1235 | 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 ); |
1236 | // Helper function for add_constraint(). |
1237 | Node* adjust_limit(int stride_con, Node * scale, Node *offset, Node *rc_limit, Node *loop_limit, Node *pre_ctrl, bool round_up); |
1238 | |
1239 | // Partially peel loop up through last_peel node. |
1240 | bool partial_peel( IdealLoopTree *loop, Node_List &old_new ); |
1241 | |
1242 | // Create a scheduled list of nodes control dependent on ctrl set. |
1243 | void scheduled_nodelist( IdealLoopTree *loop, VectorSet& ctrl, Node_List &sched ); |
1244 | // Has a use in the vector set |
1245 | bool has_use_in_set( Node* n, VectorSet& vset ); |
1246 | // Has use internal to the vector set (ie. not in a phi at the loop head) |
1247 | bool has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ); |
1248 | // clone "n" for uses that are outside of loop |
1249 | int clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ); |
1250 | // clone "n" for special uses that are in the not_peeled region |
1251 | void clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n, |
1252 | VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ); |
1253 | // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist |
1254 | void insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ); |
1255 | #ifdef ASSERT |
1256 | // Validate the loop partition sets: peel and not_peel |
1257 | bool is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, VectorSet& not_peel ); |
1258 | // Ensure that uses outside of loop are of the right form |
1259 | bool is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list, |
1260 | uint orig_exit_idx, uint clone_exit_idx); |
1261 | bool is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx); |
1262 | #endif |
1263 | |
1264 | // Returns nonzero constant stride if-node is a possible iv test (otherwise returns zero.) |
1265 | int stride_of_possible_iv( Node* iff ); |
1266 | bool is_possible_iv_test( Node* iff ) { return stride_of_possible_iv(iff) != 0; } |
1267 | // Return the (unique) control output node that's in the loop (if it exists.) |
1268 | Node* stay_in_loop( Node* n, IdealLoopTree *loop); |
1269 | // Insert a signed compare loop exit cloned from an unsigned compare. |
1270 | IfNode* insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop); |
1271 | void remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop); |
1272 | // Utility to register node "n" with PhaseIdealLoop |
1273 | void register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth); |
1274 | // Utility to create an if-projection |
1275 | ProjNode* proj_clone(ProjNode* p, IfNode* iff); |
1276 | // Force the iff control output to be the live_proj |
1277 | Node* short_circuit_if(IfNode* iff, ProjNode* live_proj); |
1278 | // Insert a region before an if projection |
1279 | RegionNode* insert_region_before_proj(ProjNode* proj); |
1280 | // Insert a new if before an if projection |
1281 | ProjNode* insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj); |
1282 | |
1283 | // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. |
1284 | // "Nearly" because all Nodes have been cloned from the original in the loop, |
1285 | // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs |
1286 | // through the Phi recursively, and return a Bool. |
1287 | Node *clone_iff( PhiNode *phi, IdealLoopTree *loop ); |
1288 | CmpNode *clone_bool( PhiNode *phi, IdealLoopTree *loop ); |
1289 | |
1290 | |
1291 | // Rework addressing expressions to get the most loop-invariant stuff |
1292 | // moved out. We'd like to do all associative operators, but it's especially |
1293 | // important (common) to do address expressions. |
1294 | Node *remix_address_expressions( Node *n ); |
1295 | |
1296 | // Convert add to muladd to generate MuladdS2I under certain criteria |
1297 | Node * convert_add_to_muladd(Node * n); |
1298 | |
1299 | // Attempt to use a conditional move instead of a phi/branch |
1300 | Node *conditional_move( Node *n ); |
1301 | |
1302 | // Reorganize offset computations to lower register pressure. |
1303 | // Mostly prevent loop-fallout uses of the pre-incremented trip counter |
1304 | // (which are then alive with the post-incremented trip counter |
1305 | // forcing an extra register move) |
1306 | void reorg_offsets( IdealLoopTree *loop ); |
1307 | |
1308 | // Check for aggressive application of 'split-if' optimization, |
1309 | // using basic block level info. |
1310 | void split_if_with_blocks ( VectorSet &visited, Node_Stack &nstack); |
1311 | Node *split_if_with_blocks_pre ( Node *n ); |
1312 | void split_if_with_blocks_post( Node *n ); |
1313 | Node *has_local_phi_input( Node *n ); |
1314 | // Mark an IfNode as being dominated by a prior test, |
1315 | // without actually altering the CFG (and hence IDOM info). |
1316 | void dominated_by( Node *prevdom, Node *iff, bool flip = false, bool exclude_loop_predicate = false ); |
1317 | |
1318 | // Split Node 'n' through merge point |
1319 | Node *split_thru_region( Node *n, Node *region ); |
1320 | // Split Node 'n' through merge point if there is enough win. |
1321 | Node *split_thru_phi( Node *n, Node *region, int policy ); |
1322 | // Found an If getting its condition-code input from a Phi in the |
1323 | // same block. Split thru the Region. |
1324 | void do_split_if( Node *iff ); |
1325 | |
1326 | // Conversion of fill/copy patterns into intrisic versions |
1327 | bool do_intrinsify_fill(); |
1328 | bool intrinsify_fill(IdealLoopTree* lpt); |
1329 | bool match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value, |
1330 | Node*& shift, Node*& offset); |
1331 | |
1332 | private: |
1333 | // Return a type based on condition control flow |
1334 | const TypeInt* filtered_type( Node *n, Node* n_ctrl); |
1335 | const TypeInt* filtered_type( Node *n ) { return filtered_type(n, NULL); } |
1336 | // Helpers for filtered type |
1337 | const TypeInt* filtered_type_from_dominators( Node* val, Node *val_ctrl); |
1338 | |
1339 | // Helper functions |
1340 | Node *spinup( Node *iff, Node *new_false, Node *new_true, Node *region, Node *phi, small_cache *cache ); |
1341 | Node *find_use_block( Node *use, Node *def, Node *old_false, Node *new_false, Node *old_true, Node *new_true ); |
1342 | 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 ); |
1343 | bool split_up( Node *n, Node *blk1, Node *blk2 ); |
1344 | void sink_use( Node *use, Node *post_loop ); |
1345 | Node *place_near_use( Node *useblock ) const; |
1346 | Node* try_move_store_before_loop(Node* n, Node *n_ctrl); |
1347 | void try_move_store_after_loop(Node* n); |
1348 | bool identical_backtoback_ifs(Node *n); |
1349 | bool can_split_if(Node *n_ctrl); |
1350 | |
1351 | // Determine if a method is too big for a/another round of split-if, based on |
1352 | // a magic (approximate) ratio derived from the equally magic constant 35000, |
1353 | // previously used for this purpose (but without relating to the node limit). |
1354 | bool must_throttle_split_if() { |
1355 | uint threshold = C->max_node_limit() * 2 / 5; |
1356 | return C->live_nodes() > threshold; |
1357 | } |
1358 | |
1359 | // A simplistic node request tracking mechanism, where |
1360 | // = UINT_MAX Request not valid or made final. |
1361 | // < UINT_MAX Nodes currently requested (estimate). |
1362 | uint _nodes_required; |
1363 | |
1364 | enum { REQUIRE_MIN = 70 }; |
1365 | |
1366 | uint nodes_required() const { return _nodes_required; } |
1367 | |
1368 | // Given the _currently_ available number of nodes, check whether there is |
1369 | // "room" for an additional request or not, considering the already required |
1370 | // number of nodes. Return TRUE if the new request is exceeding the node |
1371 | // budget limit, otherwise return FALSE. Note that this interpretation will |
1372 | // act pessimistic on additional requests when new nodes have already been |
1373 | // generated since the 'begin'. This behaviour fits with the intention that |
1374 | // node estimates/requests should be made upfront. |
1375 | bool exceeding_node_budget(uint required = 0) { |
1376 | assert(C->live_nodes() < C->max_node_limit(), "sanity" ); |
1377 | uint available = C->max_node_limit() - C->live_nodes(); |
1378 | return available < required + _nodes_required; |
1379 | } |
1380 | |
1381 | uint require_nodes(uint require, uint minreq = REQUIRE_MIN) { |
1382 | precond(require > 0); |
1383 | _nodes_required += MAX2(require, minreq); |
1384 | return _nodes_required; |
1385 | } |
1386 | |
1387 | bool may_require_nodes(uint require, uint minreq = REQUIRE_MIN) { |
1388 | return !exceeding_node_budget(require) && require_nodes(require, minreq) > 0; |
1389 | } |
1390 | |
1391 | uint require_nodes_begin() { |
1392 | assert(_nodes_required == UINT_MAX, "Bad state (begin)." ); |
1393 | _nodes_required = 0; |
1394 | return C->live_nodes(); |
1395 | } |
1396 | |
1397 | // When a node request is final, optionally check that the requested number |
1398 | // of nodes was reasonably correct with respect to the number of new nodes |
1399 | // introduced since the last 'begin'. Always check that we have not exceeded |
1400 | // the maximum node limit. |
1401 | void require_nodes_final(uint live_at_begin, bool check_estimate) { |
1402 | assert(_nodes_required < UINT_MAX, "Bad state (final)." ); |
1403 | |
1404 | if (check_estimate) { |
1405 | // Assert that the node budget request was not off by too much (x2). |
1406 | // Should this be the case we _surely_ need to improve the estimates |
1407 | // used in our budget calculations. |
1408 | assert(C->live_nodes() - live_at_begin <= 2 * _nodes_required, |
1409 | "Bad node estimate: actual = %d >> request = %d" , |
1410 | C->live_nodes() - live_at_begin, _nodes_required); |
1411 | } |
1412 | // Assert that we have stayed within the node budget limit. |
1413 | assert(C->live_nodes() < C->max_node_limit(), |
1414 | "Exceeding node budget limit: %d + %d > %d (request = %d)" , |
1415 | C->live_nodes() - live_at_begin, live_at_begin, |
1416 | C->max_node_limit(), _nodes_required); |
1417 | |
1418 | _nodes_required = UINT_MAX; |
1419 | } |
1420 | |
1421 | bool _created_loop_node; |
1422 | |
1423 | public: |
1424 | void set_created_loop_node() { _created_loop_node = true; } |
1425 | bool created_loop_node() { return _created_loop_node; } |
1426 | void register_new_node( Node *n, Node *blk ); |
1427 | |
1428 | #ifdef ASSERT |
1429 | void dump_bad_graph(const char* msg, Node* n, Node* early, Node* LCA); |
1430 | #endif |
1431 | |
1432 | #ifndef PRODUCT |
1433 | void dump( ) const; |
1434 | void dump( IdealLoopTree *loop, uint rpo_idx, Node_List &rpo_list ) const; |
1435 | void verify() const; // Major slow :-) |
1436 | void verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const; |
1437 | IdealLoopTree *get_loop_idx(Node* n) const { |
1438 | // Dead nodes have no loop, so return the top level loop instead |
1439 | return _nodes[n->_idx] ? (IdealLoopTree*)_nodes[n->_idx] : _ltree_root; |
1440 | } |
1441 | // Print some stats |
1442 | static void print_statistics(); |
1443 | static int _loop_invokes; // Count of PhaseIdealLoop invokes |
1444 | static int _loop_work; // Sum of PhaseIdealLoop x _unique |
1445 | #endif |
1446 | void rpo( Node *start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list ) const; |
1447 | }; |
1448 | |
1449 | |
1450 | class AutoNodeBudget : public StackObj |
1451 | { |
1452 | public: |
1453 | enum budget_check_t { BUDGET_CHECK, NO_BUDGET_CHECK }; |
1454 | |
1455 | AutoNodeBudget(PhaseIdealLoop* phase, budget_check_t chk = BUDGET_CHECK) |
1456 | : _phase(phase), |
1457 | _check_at_final(chk == BUDGET_CHECK), |
1458 | _nodes_at_begin(0) |
1459 | { |
1460 | precond(_phase != NULL); |
1461 | |
1462 | _nodes_at_begin = _phase->require_nodes_begin(); |
1463 | } |
1464 | |
1465 | ~AutoNodeBudget() { |
1466 | #ifndef PRODUCT |
1467 | if (TraceLoopOpts) { |
1468 | uint request = _phase->nodes_required(); |
1469 | uint delta = _phase->C->live_nodes() - _nodes_at_begin; |
1470 | |
1471 | if (request < delta) { |
1472 | tty->print_cr("Exceeding node budget: %d < %d" , request, delta); |
1473 | } else { |
1474 | uint const REQUIRE_MIN = PhaseIdealLoop::REQUIRE_MIN; |
1475 | // Identify the worst estimates as "poor" ones. |
1476 | if (request > REQUIRE_MIN && delta > 0) { |
1477 | if ((delta > REQUIRE_MIN && request > 3 * delta) || |
1478 | (delta <= REQUIRE_MIN && request > 10 * delta)) { |
1479 | tty->print_cr("Poor node estimate: %d >> %d" , request, delta); |
1480 | } |
1481 | } |
1482 | } |
1483 | } |
1484 | #endif // PRODUCT |
1485 | _phase->require_nodes_final(_nodes_at_begin, _check_at_final); |
1486 | } |
1487 | |
1488 | private: |
1489 | PhaseIdealLoop* _phase; |
1490 | bool _check_at_final; |
1491 | uint _nodes_at_begin; |
1492 | }; |
1493 | |
1494 | |
1495 | // This kit may be used for making of a reserved copy of a loop before this loop |
1496 | // goes under non-reversible changes. |
1497 | // |
1498 | // Function create_reserve() creates a reserved copy (clone) of the loop. |
1499 | // The reserved copy is created by calling |
1500 | // PhaseIdealLoop::create_reserve_version_of_loop - see there how |
1501 | // the original and reserved loops are connected in the outer graph. |
1502 | // If create_reserve succeeded, it returns 'true' and _has_reserved is set to 'true'. |
1503 | // |
1504 | // By default the reserved copy (clone) of the loop is created as dead code - it is |
1505 | // dominated in the outer loop by this node chain: |
1506 | // intcon(1)->If->IfFalse->reserved_copy. |
1507 | // The original loop is dominated by the the same node chain but IfTrue projection: |
1508 | // intcon(0)->If->IfTrue->original_loop. |
1509 | // |
1510 | // In this implementation of CountedLoopReserveKit the ctor includes create_reserve() |
1511 | // and the dtor, checks _use_new value. |
1512 | // If _use_new == false, it "switches" control to reserved copy of the loop |
1513 | // by simple replacing of node intcon(1) with node intcon(0). |
1514 | // |
1515 | // Here is a proposed example of usage (see also SuperWord::output in superword.cpp). |
1516 | // |
1517 | // void CountedLoopReserveKit_example() |
1518 | // { |
1519 | // CountedLoopReserveKit lrk((phase, lpt, DoReserveCopy = true); // create local object |
1520 | // if (DoReserveCopy && !lrk.has_reserved()) { |
1521 | // return; //failed to create reserved loop copy |
1522 | // } |
1523 | // ... |
1524 | // //something is wrong, switch to original loop |
1525 | /// if(something_is_wrong) return; // ~CountedLoopReserveKit makes the switch |
1526 | // ... |
1527 | // //everything worked ok, return with the newly modified loop |
1528 | // lrk.use_new(); |
1529 | // return; // ~CountedLoopReserveKit does nothing once use_new() was called |
1530 | // } |
1531 | // |
1532 | // Keep in mind, that by default if create_reserve() is not followed by use_new() |
1533 | // the dtor will "switch to the original" loop. |
1534 | // NOTE. You you modify outside of the original loop this class is no help. |
1535 | // |
1536 | class CountedLoopReserveKit { |
1537 | private: |
1538 | PhaseIdealLoop* _phase; |
1539 | IdealLoopTree* _lpt; |
1540 | LoopNode* _lp; |
1541 | IfNode* _iff; |
1542 | LoopNode* _lp_reserved; |
1543 | bool _has_reserved; |
1544 | bool _use_new; |
1545 | const bool _active; //may be set to false in ctor, then the object is dummy |
1546 | |
1547 | public: |
1548 | CountedLoopReserveKit(PhaseIdealLoop* phase, IdealLoopTree *loop, bool active); |
1549 | ~CountedLoopReserveKit(); |
1550 | void use_new() {_use_new = true;} |
1551 | void set_iff(IfNode* x) {_iff = x;} |
1552 | bool has_reserved() const { return _active && _has_reserved;} |
1553 | private: |
1554 | bool create_reserve(); |
1555 | };// class CountedLoopReserveKit |
1556 | |
1557 | inline Node* IdealLoopTree::tail() { |
1558 | // Handle lazy update of _tail field. |
1559 | if (_tail->in(0) == NULL) { |
1560 | _tail = _phase->get_ctrl(_tail); |
1561 | } |
1562 | return _tail; |
1563 | } |
1564 | |
1565 | |
1566 | // Iterate over the loop tree using a preorder, left-to-right traversal. |
1567 | // |
1568 | // Example that visits all counted loops from within PhaseIdealLoop |
1569 | // |
1570 | // for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) { |
1571 | // IdealLoopTree* lpt = iter.current(); |
1572 | // if (!lpt->is_counted()) continue; |
1573 | // ... |
1574 | class LoopTreeIterator : public StackObj { |
1575 | private: |
1576 | IdealLoopTree* _root; |
1577 | IdealLoopTree* _curnt; |
1578 | |
1579 | public: |
1580 | LoopTreeIterator(IdealLoopTree* root) : _root(root), _curnt(root) {} |
1581 | |
1582 | bool done() { return _curnt == NULL; } // Finished iterating? |
1583 | |
1584 | void next(); // Advance to next loop tree |
1585 | |
1586 | IdealLoopTree* current() { return _curnt; } // Return current value of iterator. |
1587 | }; |
1588 | |
1589 | #endif // SHARE_OPTO_LOOPNODE_HPP |
1590 | |