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.
size_of() const59 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
is_inner_loop() const92 bool is_inner_loop() const { return _loop_flags & InnerLoop; }
set_inner_loop()93 void set_inner_loop() { _loop_flags |= InnerLoop; }
94
range_checks_present() const95 bool range_checks_present() const { return _loop_flags & HasRangeChecks; }
is_multiversioned() const96 bool is_multiversioned() const { return _loop_flags & IsMultiversioned; }
is_vectorized_loop() const97 bool is_vectorized_loop() const { return _loop_flags & VectorizedLoop; }
is_partial_peel_loop() const98 bool is_partial_peel_loop() const { return _loop_flags & PartialPeelLoop; }
set_partial_peel_loop()99 void set_partial_peel_loop() { _loop_flags |= PartialPeelLoop; }
partial_peel_has_failed() const100 bool partial_peel_has_failed() const { return _loop_flags & PartialPeelFailed; }
is_strip_mined() const101 bool is_strip_mined() const { return _loop_flags & StripMined; }
is_profile_trip_failed() const102 bool is_profile_trip_failed() const { return _loop_flags & ProfileTripFailed; }
is_subword_loop() const103 bool is_subword_loop() const { return _loop_flags & SubwordLoop; }
104
mark_partial_peel_failed()105 void mark_partial_peel_failed() { _loop_flags |= PartialPeelFailed; }
mark_has_reductions()106 void mark_has_reductions() { _loop_flags |= HasReductions; }
mark_was_slp()107 void mark_was_slp() { _loop_flags |= WasSlpAnalyzed; }
mark_passed_slp()108 void mark_passed_slp() { _loop_flags |= PassedSlpAnalysis; }
mark_do_unroll_only()109 void mark_do_unroll_only() { _loop_flags |= DoUnrollOnly; }
mark_loop_vectorized()110 void mark_loop_vectorized() { _loop_flags |= VectorizedLoop; }
mark_has_atomic_post_loop()111 void mark_has_atomic_post_loop() { _loop_flags |= HasAtomicPostLoop; }
mark_has_range_checks()112 void mark_has_range_checks() { _loop_flags |= HasRangeChecks; }
mark_is_multiversioned()113 void mark_is_multiversioned() { _loop_flags |= IsMultiversioned; }
mark_strip_mined()114 void mark_strip_mined() { _loop_flags |= StripMined; }
clear_strip_mined()115 void clear_strip_mined() { _loop_flags &= ~StripMined; }
mark_profile_trip_failed()116 void mark_profile_trip_failed() { _loop_flags |= ProfileTripFailed; }
mark_subword_loop()117 void mark_subword_loop() { _loop_flags |= SubwordLoop; }
118
unswitch_max()119 int unswitch_max() { return _unswitch_max; }
unswitch_count()120 int unswitch_count() { return _unswitch_count; }
121
has_been_range_checked() const122 int has_been_range_checked() const { return _postloop_flags & LoopRCEChecked; }
set_has_been_range_checked()123 void set_has_been_range_checked() { _postloop_flags |= LoopRCEChecked; }
is_rce_post_loop() const124 int is_rce_post_loop() const { return _postloop_flags & RCEPostLoop; }
set_is_rce_post_loop()125 void set_is_rce_post_loop() { _postloop_flags |= RCEPostLoop; }
126
set_unswitch_count(int val)127 void set_unswitch_count(int val) {
128 assert (val <= unswitch_max(), "too many unswitches");
129 _unswitch_count = val;
130 }
131
set_profile_trip_cnt(float ptc)132 void set_profile_trip_cnt(float ptc) { _profile_trip_cnt = ptc; }
profile_trip_cnt()133 float profile_trip_cnt() { return _profile_trip_cnt; }
134
LoopNode(Node * entry,Node * backedge)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;
can_be_counted_loop(PhaseTransform * phase) const145 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 NOT_DEBUG_RETURN;
skip_strip_mined(int expect_skeleton=1)156 virtual LoopNode* skip_strip_mined(int expect_skeleton = 1) { return this; }
outer_loop_tail() const157 virtual IfTrueNode* outer_loop_tail() const { ShouldNotReachHere(); return NULL; }
outer_loop_end() const158 virtual OuterStripMinedLoopEndNode* outer_loop_end() const { ShouldNotReachHere(); return NULL; }
outer_loop_exit() const159 virtual IfFalseNode* outer_loop_exit() const { ShouldNotReachHere(); return NULL; }
outer_safepoint() const160 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.
size_of() const195 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:
CountedLoopNode(Node * entry,Node * backedge)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
init_control() const228 Node *init_control() const { return in(EntryControl); }
back_control() const229 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.
is_normal_loop() const260 bool is_normal_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Normal; }
is_pre_loop() const261 bool is_pre_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Pre; }
is_main_loop() const262 bool is_main_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Main; }
is_post_loop() const263 bool is_post_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Post; }
is_reduction_loop() const264 bool is_reduction_loop() const { return (_loop_flags&HasReductions) == HasReductions; }
was_slp_analyzed() const265 bool was_slp_analyzed () const { return (_loop_flags&WasSlpAnalyzed) == WasSlpAnalyzed; }
has_passed_slp() const266 bool has_passed_slp () const { return (_loop_flags&PassedSlpAnalysis) == PassedSlpAnalysis; }
is_unroll_only() const267 bool is_unroll_only () const { return (_loop_flags&DoUnrollOnly) == DoUnrollOnly; }
is_main_no_pre_loop() const268 bool is_main_no_pre_loop() const { return _loop_flags & MainHasNoPreLoop; }
has_atomic_post_loop() const269 bool has_atomic_post_loop () const { return (_loop_flags & HasAtomicPostLoop) == HasAtomicPostLoop; }
set_main_no_pre_loop()270 void set_main_no_pre_loop() { _loop_flags |= MainHasNoPreLoop; }
271
main_idx() const272 int main_idx() const { return _main_idx; }
273
274
set_pre_loop(CountedLoopNode * main)275 void set_pre_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Pre ; _main_idx = main->_idx; }
set_main_loop()276 void set_main_loop ( ) { assert(is_normal_loop(),""); _loop_flags |= Main; }
set_post_loop(CountedLoopNode * main)277 void set_post_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Post; _main_idx = main->_idx; }
set_normal_loop()278 void set_normal_loop( ) { _loop_flags &= ~PreMainPostFlagsMask; }
279
set_trip_count(uint tc)280 void set_trip_count(uint tc) { _trip_count = tc; }
trip_count()281 uint trip_count() { return _trip_count; }
282
has_exact_trip_count() const283 bool has_exact_trip_count() const { return (_loop_flags & HasExactTripCount) != 0; }
set_exact_trip_count(uint tc)284 void set_exact_trip_count(uint tc) {
285 _trip_count = tc;
286 _loop_flags |= HasExactTripCount;
287 }
set_nonexact_trip_count()288 void set_nonexact_trip_count() {
289 _loop_flags &= ~HasExactTripCount;
290 }
set_notpassed_slp()291 void set_notpassed_slp() {
292 _loop_flags &= ~PassedSlpAnalysis;
293 }
294
double_unrolled_count()295 void double_unrolled_count() { _unrolled_count_log2++; }
unrolled_count()296 int unrolled_count() { return 1 << MIN2(_unrolled_count_log2, BitsPerInt-3); }
297
set_node_count_before_unroll(int ct)298 void set_node_count_before_unroll(int ct) { _node_count_before_unroll = ct; }
node_count_before_unroll()299 int node_count_before_unroll() { return _node_count_before_unroll; }
set_slp_max_unroll(int unroll_factor)300 void set_slp_max_unroll(int unroll_factor) { _slp_maximum_unroll_factor = unroll_factor; }
slp_max_unroll() const301 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
CountedLoopEndNode(Node * control,Node * test,float prob,float cnt)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
cmp_node() const334 Node *cmp_node() const { return (in(TestValue)->req() >=2) ? in(TestValue)->in(1) : NULL; }
incr() const335 Node *incr() const { Node *tmp = cmp_node(); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; }
limit() const336 Node *limit() const { Node *tmp = cmp_node(); return (tmp && tmp->req()==3) ? tmp->in(2) : NULL; }
stride() const337 Node *stride() const { Node *tmp = incr (); return (tmp && tmp->req()==3) ? tmp->in(2) : NULL; }
init_trip() const338 Node *init_trip() const { Node *tmp = phi (); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; }
339 int stride_con() const;
stride_is_con() const340 bool stride_is_con() const { Node *tmp = stride (); return (tmp != NULL && tmp->is_Con()); }
test_trip() const341 BoolTest::mask test_trip() const { return in(TestValue)->as_Bool()->_test._test; }
phi() const342 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 }
loopnode() const352 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
loopexit_or_null() const373 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
loopexit() const382 inline CountedLoopEndNode* CountedLoopNode::loopexit() const {
383 CountedLoopEndNode* cle = loopexit_or_null();
384 assert(cle != NULL, "loopexit is NULL");
385 return cle;
386 }
387
init_trip() const388 inline Node* CountedLoopNode::init_trip() const {
389 CountedLoopEndNode* cle = loopexit_or_null();
390 return cle != NULL ? cle->init_trip() : NULL;
391 }
stride() const392 inline Node* CountedLoopNode::stride() const {
393 CountedLoopEndNode* cle = loopexit_or_null();
394 return cle != NULL ? cle->stride() : NULL;
395 }
stride_con() const396 inline int CountedLoopNode::stride_con() const {
397 CountedLoopEndNode* cle = loopexit_or_null();
398 return cle != NULL ? cle->stride_con() : 0;
399 }
stride_is_con() const400 inline bool CountedLoopNode::stride_is_con() const {
401 CountedLoopEndNode* cle = loopexit_or_null();
402 return cle != NULL && cle->stride_is_con();
403 }
limit() const404 inline Node* CountedLoopNode::limit() const {
405 CountedLoopEndNode* cle = loopexit_or_null();
406 return cle != NULL ? cle->limit() : NULL;
407 }
incr() const408 inline Node* CountedLoopNode::incr() const {
409 CountedLoopEndNode* cle = loopexit_or_null();
410 return cle != NULL ? cle->incr() : NULL;
411 }
phi() const412 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:
LoopLimitNode(Compile * C,Node * init,Node * limit,Node * stride)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;
bottom_type() const433 virtual const Type *bottom_type() const { return TypeInt::INT; }
ideal_reg() const434 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:
OuterStripMinedLoopNode(Compile * C,Node * entry,Node * backedge)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:
OuterStripMinedLoopEndNode(Node * control,Node * test,float prob,float cnt)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 bool is_expanded(PhaseGVN *phase) const;
474 };
475
476 // -----------------------------IdealLoopTree----------------------------------
477 class IdealLoopTree : public ResourceObj {
478 public:
479 IdealLoopTree *_parent; // Parent in loop tree
480 IdealLoopTree *_next; // Next sibling in loop tree
481 IdealLoopTree *_child; // First child in loop tree
482
483 // The head-tail backedge defines the loop.
484 // If a loop has multiple backedges, this is addressed during cleanup where
485 // we peel off the multiple backedges, merging all edges at the bottom and
486 // ensuring that one proper backedge flow into the loop.
487 Node *_head; // Head of loop
488 Node *_tail; // Tail of loop
489 inline Node *tail(); // Handle lazy update of _tail field
490 inline Node *head(); // Handle lazy update of _head field
491 PhaseIdealLoop* _phase;
492 int _local_loop_unroll_limit;
493 int _local_loop_unroll_factor;
494
495 Node_List _body; // Loop body for inner loops
496
497 uint8_t _nest; // Nesting depth
498 uint8_t _irreducible:1, // True if irreducible
499 _has_call:1, // True if has call safepoint
500 _has_sfpt:1, // True if has non-call safepoint
501 _rce_candidate:1; // True if candidate for range check elimination
502
503 Node_List* _safepts; // List of safepoints in this loop
504 Node_List* _required_safept; // A inner loop cannot delete these safepts;
505 bool _allow_optimizations; // Allow loop optimizations
506
IdealLoopTree(PhaseIdealLoop * phase,Node * head,Node * tail)507 IdealLoopTree( PhaseIdealLoop* phase, Node *head, Node *tail )
508 : _parent(0), _next(0), _child(0),
509 _head(head), _tail(tail),
510 _phase(phase),
511 _local_loop_unroll_limit(0), _local_loop_unroll_factor(0),
512 _nest(0), _irreducible(0), _has_call(0), _has_sfpt(0), _rce_candidate(0),
513 _safepts(NULL),
514 _required_safept(NULL),
515 _allow_optimizations(true)
516 {
517 precond(_head != NULL);
518 precond(_tail != NULL);
519 }
520
521 // Is 'l' a member of 'this'?
522 bool is_member(const IdealLoopTree *l) const; // Test for nested membership
523
524 // Set loop nesting depth. Accumulate has_call bits.
525 int set_nest( uint depth );
526
527 // Split out multiple fall-in edges from the loop header. Move them to a
528 // private RegionNode before the loop. This becomes the loop landing pad.
529 void split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt );
530
531 // Split out the outermost loop from this shared header.
532 void split_outer_loop( PhaseIdealLoop *phase );
533
534 // Merge all the backedges from the shared header into a private Region.
535 // Feed that region as the one backedge to this loop.
536 void merge_many_backedges( PhaseIdealLoop *phase );
537
538 // Split shared headers and insert loop landing pads.
539 // Insert a LoopNode to replace the RegionNode.
540 // Returns TRUE if loop tree is structurally changed.
541 bool beautify_loops( PhaseIdealLoop *phase );
542
543 // Perform optimization to use the loop predicates for null checks and range checks.
544 // Applies to any loop level (not just the innermost one)
545 bool loop_predication( PhaseIdealLoop *phase);
546
547 // Perform iteration-splitting on inner loops. Split iterations to
548 // avoid range checks or one-shot null checks. Returns false if the
549 // current round of loop opts should stop.
550 bool iteration_split( PhaseIdealLoop *phase, Node_List &old_new );
551
552 // Driver for various flavors of iteration splitting. Returns false
553 // if the current round of loop opts should stop.
554 bool iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new );
555
556 // Given dominators, try to find loops with calls that must always be
557 // executed (call dominates loop tail). These loops do not need non-call
558 // safepoints (ncsfpt).
559 void check_safepts(VectorSet &visited, Node_List &stack);
560
561 // Allpaths backwards scan from loop tail, terminating each path at first safepoint
562 // encountered.
563 void allpaths_check_safepts(VectorSet &visited, Node_List &stack);
564
565 // Remove safepoints from loop. Optionally keeping one.
566 void remove_safepoints(PhaseIdealLoop* phase, bool keep_one);
567
568 // Convert to counted loops where possible
569 void counted_loop( PhaseIdealLoop *phase );
570
571 // Check for Node being a loop-breaking test
572 Node *is_loop_exit(Node *iff) const;
573
574 // Remove simplistic dead code from loop body
575 void DCE_loop_body();
576
577 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage.
578 // Replace with a 1-in-10 exit guess.
579 void adjust_loop_exit_prob( PhaseIdealLoop *phase );
580
581 // Return TRUE or FALSE if the loop should never be RCE'd or aligned.
582 // Useful for unrolling loops with NO array accesses.
583 bool policy_peel_only( PhaseIdealLoop *phase ) const;
584
585 // Return TRUE or FALSE if the loop should be unswitched -- clone
586 // loop with an invariant test
587 bool policy_unswitching( PhaseIdealLoop *phase ) const;
588
589 // Micro-benchmark spamming. Remove empty loops.
590 bool do_remove_empty_loop( PhaseIdealLoop *phase );
591
592 // Convert one iteration loop into normal code.
593 bool do_one_iteration_loop( PhaseIdealLoop *phase );
594
595 // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can
596 // move some loop-invariant test (usually a null-check) before the loop.
597 bool policy_peeling(PhaseIdealLoop *phase);
598
599 uint estimate_peeling(PhaseIdealLoop *phase);
600
601 // Return TRUE or FALSE if the loop should be maximally unrolled. Stash any
602 // known trip count in the counted loop node.
603 bool policy_maximally_unroll(PhaseIdealLoop *phase) const;
604
605 // Return TRUE or FALSE if the loop should be unrolled or not. Apply unroll
606 // if the loop is a counted loop and the loop body is small enough.
607 bool policy_unroll(PhaseIdealLoop *phase);
608
609 // Loop analyses to map to a maximal superword unrolling for vectorization.
610 void policy_unroll_slp_analysis(CountedLoopNode *cl, PhaseIdealLoop *phase, int future_unroll_ct);
611
612 // Return TRUE or FALSE if the loop should be range-check-eliminated.
613 // Gather a list of IF tests that are dominated by iteration splitting;
614 // also gather the end of the first split and the start of the 2nd split.
615 bool policy_range_check( PhaseIdealLoop *phase ) const;
616
617 // Return TRUE or FALSE if the loop should be cache-line aligned.
618 // Gather the expression that does the alignment. Note that only
619 // one array base can be aligned in a loop (unless the VM guarantees
620 // mutual alignment). Note that if we vectorize short memory ops
621 // into longer memory ops, we may want to increase alignment.
622 bool policy_align( PhaseIdealLoop *phase ) const;
623
624 // Return TRUE if "iff" is a range check.
625 bool is_range_check_if(IfNode *iff, PhaseIdealLoop *phase, Invariance& invar) const;
626
627 // Estimate the number of nodes required when cloning a loop (body).
628 uint est_loop_clone_sz(uint factor) const;
629 // Estimate the number of nodes required when unrolling a loop (body).
630 uint est_loop_unroll_sz(uint factor) const;
631
632 // Compute loop trip count if possible
633 void compute_trip_count(PhaseIdealLoop* phase);
634
635 // Compute loop trip count from profile data
636 float compute_profile_trip_cnt_helper(Node* n);
637 void compute_profile_trip_cnt( PhaseIdealLoop *phase );
638
639 // Reassociate invariant expressions.
640 void reassociate_invariants(PhaseIdealLoop *phase);
641 // Reassociate invariant add and subtract expressions.
642 Node* reassociate_add_sub(Node* n1, PhaseIdealLoop *phase);
643 // Return nonzero index of invariant operand if invariant and variant
644 // are combined with an Add or Sub. Helper for reassociate_invariants.
645 int is_invariant_addition(Node* n, PhaseIdealLoop *phase);
646
647 // Return true if n is invariant
648 bool is_invariant(Node* n) const;
649
650 // Put loop body on igvn work list
651 void record_for_igvn();
652
is_root()653 bool is_root() { return _parent == NULL; }
654 // A proper/reducible loop w/o any (occasional) dead back-edge.
is_loop()655 bool is_loop() { return !_irreducible && !tail()->is_top(); }
is_counted()656 bool is_counted() { return is_loop() && _head->is_CountedLoop(); }
is_innermost()657 bool is_innermost() { return is_loop() && _child == NULL; }
658
659 void remove_main_post_loops(CountedLoopNode *cl, PhaseIdealLoop *phase);
660
661 #ifndef PRODUCT
662 void dump_head() const; // Dump loop head only
663 void dump() const; // Dump this loop recursively
664 void verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const;
665 #endif
666
667 private:
668 enum { EMPTY_LOOP_SIZE = 7 }; // Number of nodes in an empty loop.
669
670 // Estimate the number of nodes resulting from control and data flow merge.
671 uint est_loop_flow_merge_sz() const;
672 };
673
674 // -----------------------------PhaseIdealLoop---------------------------------
675 // Computes the mapping from Nodes to IdealLoopTrees. Organizes IdealLoopTrees
676 // into a loop tree. Drives the loop-based transformations on the ideal graph.
677 class PhaseIdealLoop : public PhaseTransform {
678 friend class IdealLoopTree;
679 friend class SuperWord;
680 friend class CountedLoopReserveKit;
681 friend class ShenandoahBarrierC2Support;
682 friend class AutoNodeBudget;
683
684 // Pre-computed def-use info
685 PhaseIterGVN &_igvn;
686
687 // Head of loop tree
688 IdealLoopTree* _ltree_root;
689
690 // Array of pre-order numbers, plus post-visited bit.
691 // ZERO for not pre-visited. EVEN for pre-visited but not post-visited.
692 // ODD for post-visited. Other bits are the pre-order number.
693 uint *_preorders;
694 uint _max_preorder;
695
696 const PhaseIdealLoop* _verify_me;
697 bool _verify_only;
698
699 // Allocate _preorders[] array
allocate_preorders()700 void allocate_preorders() {
701 _max_preorder = C->unique()+8;
702 _preorders = NEW_RESOURCE_ARRAY(uint, _max_preorder);
703 memset(_preorders, 0, sizeof(uint) * _max_preorder);
704 }
705
706 // Allocate _preorders[] array
reallocate_preorders()707 void reallocate_preorders() {
708 if ( _max_preorder < C->unique() ) {
709 _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, C->unique());
710 _max_preorder = C->unique();
711 }
712 memset(_preorders, 0, sizeof(uint) * _max_preorder);
713 }
714
715 // Check to grow _preorders[] array for the case when build_loop_tree_impl()
716 // adds new nodes.
check_grow_preorders()717 void check_grow_preorders( ) {
718 if ( _max_preorder < C->unique() ) {
719 uint newsize = _max_preorder<<1; // double size of array
720 _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, newsize);
721 memset(&_preorders[_max_preorder],0,sizeof(uint)*(newsize-_max_preorder));
722 _max_preorder = newsize;
723 }
724 }
725 // Check for pre-visited. Zero for NOT visited; non-zero for visited.
is_visited(Node * n) const726 int is_visited( Node *n ) const { return _preorders[n->_idx]; }
727 // Pre-order numbers are written to the Nodes array as low-bit-set values.
set_preorder_visited(Node * n,int pre_order)728 void set_preorder_visited( Node *n, int pre_order ) {
729 assert( !is_visited( n ), "already set" );
730 _preorders[n->_idx] = (pre_order<<1);
731 };
732 // Return pre-order number.
get_preorder(Node * n) const733 int get_preorder( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]>>1; }
734
735 // Check for being post-visited.
736 // Should be previsited already (checked with assert(is_visited(n))).
is_postvisited(Node * n) const737 int is_postvisited( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]&1; }
738
739 // Mark as post visited
set_postvisited(Node * n)740 void set_postvisited( Node *n ) { assert( !is_postvisited( n ), "" ); _preorders[n->_idx] |= 1; }
741
742 public:
743 // Set/get control node out. Set lower bit to distinguish from IdealLoopTree
744 // Returns true if "n" is a data node, false if it's a control node.
has_ctrl(Node * n) const745 bool has_ctrl( Node *n ) const { return ((intptr_t)_nodes[n->_idx]) & 1; }
746
747 private:
748 // clear out dead code after build_loop_late
749 Node_List _deadlist;
750
751 // Support for faster execution of get_late_ctrl()/dom_lca()
752 // when a node has many uses and dominator depth is deep.
753 Node_Array _dom_lca_tags;
754 void init_dom_lca_tags();
755 void clear_dom_lca_tags();
756
757 // Helper for debugging bad dominance relationships
758 bool verify_dominance(Node* n, Node* use, Node* LCA, Node* early);
759
760 Node* compute_lca_of_uses(Node* n, Node* early, bool verify = false);
761
762 // Inline wrapper for frequent cases:
763 // 1) only one use
764 // 2) a use is the same as the current LCA passed as 'n1'
dom_lca_for_get_late_ctrl(Node * lca,Node * n,Node * tag)765 Node *dom_lca_for_get_late_ctrl( Node *lca, Node *n, Node *tag ) {
766 assert( n->is_CFG(), "" );
767 // Fast-path NULL lca
768 if( lca != NULL && lca != n ) {
769 assert( lca->is_CFG(), "" );
770 // find LCA of all uses
771 n = dom_lca_for_get_late_ctrl_internal( lca, n, tag );
772 }
773 return find_non_split_ctrl(n);
774 }
775 Node *dom_lca_for_get_late_ctrl_internal( Node *lca, Node *n, Node *tag );
776
777 // Helper function for directing control inputs away from CFG split points.
find_non_split_ctrl(Node * ctrl) const778 Node *find_non_split_ctrl( Node *ctrl ) const {
779 if (ctrl != NULL) {
780 if (ctrl->is_MultiBranch()) {
781 ctrl = ctrl->in(0);
782 }
783 assert(ctrl->is_CFG(), "CFG");
784 }
785 return ctrl;
786 }
787
788 Node* cast_incr_before_loop(Node* incr, Node* ctrl, Node* loop);
789
790 #ifdef ASSERT
791 void ensure_zero_trip_guard_proj(Node* node, bool is_main_loop);
792 #endif
793 void copy_skeleton_predicates_to_main_loop_helper(Node* predicate, Node* init, Node* stride, IdealLoopTree* outer_loop, LoopNode* outer_main_head,
794 uint dd_main_head, const uint idx_before_pre_post, const uint idx_after_post_before_pre,
795 Node* zero_trip_guard_proj_main, Node* zero_trip_guard_proj_post, const Node_List &old_new);
796 void copy_skeleton_predicates_to_main_loop(CountedLoopNode* pre_head, Node* init, Node* stride, IdealLoopTree* outer_loop, LoopNode* outer_main_head,
797 uint dd_main_head, const uint idx_before_pre_post, const uint idx_after_post_before_pre,
798 Node* zero_trip_guard_proj_main, Node* zero_trip_guard_proj_post, const Node_List &old_new);
799 Node* clone_skeleton_predicate(Node* iff, Node* new_init, Node* new_stride, Node* predicate, Node* uncommon_proj,
800 Node* current_proj, IdealLoopTree* outer_loop, Node* prev_proj);
801 bool skeleton_predicate_has_opaque(IfNode* iff);
802 void update_main_loop_skeleton_predicates(Node* ctrl, CountedLoopNode* loop_head, Node* init, int stride_con);
803 void insert_loop_limit_check(ProjNode* limit_check_proj, Node* cmp_limit, Node* bol);
804
805 public:
806
igvn() const807 PhaseIterGVN &igvn() const { return _igvn; }
808
809 static bool is_canonical_loop_entry(CountedLoopNode* cl);
810
has_node(Node * n) const811 bool has_node( Node* n ) const {
812 guarantee(n != NULL, "No Node.");
813 return _nodes[n->_idx] != NULL;
814 }
815 // check if transform created new nodes that need _ctrl recorded
816 Node *get_late_ctrl( Node *n, Node *early );
817 Node *get_early_ctrl( Node *n );
818 Node *get_early_ctrl_for_expensive(Node *n, Node* earliest);
819 void set_early_ctrl( Node *n );
820 void set_subtree_ctrl( Node *root );
set_ctrl(Node * n,Node * ctrl)821 void set_ctrl( Node *n, Node *ctrl ) {
822 assert( !has_node(n) || has_ctrl(n), "" );
823 assert( ctrl->in(0), "cannot set dead control node" );
824 assert( ctrl == find_non_split_ctrl(ctrl), "must set legal crtl" );
825 _nodes.map( n->_idx, (Node*)((intptr_t)ctrl + 1) );
826 }
827 // Set control and update loop membership
set_ctrl_and_loop(Node * n,Node * ctrl)828 void set_ctrl_and_loop(Node* n, Node* ctrl) {
829 IdealLoopTree* old_loop = get_loop(get_ctrl(n));
830 IdealLoopTree* new_loop = get_loop(ctrl);
831 if (old_loop != new_loop) {
832 if (old_loop->_child == NULL) old_loop->_body.yank(n);
833 if (new_loop->_child == NULL) new_loop->_body.push(n);
834 }
835 set_ctrl(n, ctrl);
836 }
837 // Control nodes can be replaced or subsumed. During this pass they
838 // get their replacement Node in slot 1. Instead of updating the block
839 // location of all Nodes in the subsumed block, we lazily do it. As we
840 // pull such a subsumed block out of the array, we write back the final
841 // correct block.
get_ctrl(Node * i)842 Node *get_ctrl( Node *i ) {
843
844 assert(has_node(i), "");
845 Node *n = get_ctrl_no_update(i);
846 _nodes.map( i->_idx, (Node*)((intptr_t)n + 1) );
847 assert(has_node(i) && has_ctrl(i), "");
848 assert(n == find_non_split_ctrl(n), "must return legal ctrl" );
849 return n;
850 }
851 // true if CFG node d dominates CFG node n
852 bool is_dominator(Node *d, Node *n);
853 // return get_ctrl for a data node and self(n) for a CFG node
ctrl_or_self(Node * n)854 Node* ctrl_or_self(Node* n) {
855 if (has_ctrl(n))
856 return get_ctrl(n);
857 else {
858 assert (n->is_CFG(), "must be a CFG node");
859 return n;
860 }
861 }
862
get_ctrl_no_update_helper(Node * i) const863 Node *get_ctrl_no_update_helper(Node *i) const {
864 assert(has_ctrl(i), "should be control, not loop");
865 return (Node*)(((intptr_t)_nodes[i->_idx]) & ~1);
866 }
867
get_ctrl_no_update(Node * i) const868 Node *get_ctrl_no_update(Node *i) const {
869 assert( has_ctrl(i), "" );
870 Node *n = get_ctrl_no_update_helper(i);
871 if (!n->in(0)) {
872 // Skip dead CFG nodes
873 do {
874 n = get_ctrl_no_update_helper(n);
875 } while (!n->in(0));
876 n = find_non_split_ctrl(n);
877 }
878 return n;
879 }
880
881 // Check for loop being set
882 // "n" must be a control node. Returns true if "n" is known to be in a loop.
has_loop(Node * n) const883 bool has_loop( Node *n ) const {
884 assert(!has_node(n) || !has_ctrl(n), "");
885 return has_node(n);
886 }
887 // Set loop
set_loop(Node * n,IdealLoopTree * loop)888 void set_loop( Node *n, IdealLoopTree *loop ) {
889 _nodes.map(n->_idx, (Node*)loop);
890 }
891 // Lazy-dazy update of 'get_ctrl' and 'idom_at' mechanisms. Replace
892 // the 'old_node' with 'new_node'. Kill old-node. Add a reference
893 // from old_node to new_node to support the lazy update. Reference
894 // replaces loop reference, since that is not needed for dead node.
lazy_update(Node * old_node,Node * new_node)895 void lazy_update(Node *old_node, Node *new_node) {
896 assert(old_node != new_node, "no cycles please");
897 // Re-use the side array slot for this node to provide the
898 // forwarding pointer.
899 _nodes.map(old_node->_idx, (Node*)((intptr_t)new_node + 1));
900 }
lazy_replace(Node * old_node,Node * new_node)901 void lazy_replace(Node *old_node, Node *new_node) {
902 _igvn.replace_node(old_node, new_node);
903 lazy_update(old_node, new_node);
904 }
905
906 private:
907
908 // Place 'n' in some loop nest, where 'n' is a CFG node
909 void build_loop_tree();
910 int build_loop_tree_impl( Node *n, int pre_order );
911 // Insert loop into the existing loop tree. 'innermost' is a leaf of the
912 // loop tree, not the root.
913 IdealLoopTree *sort( IdealLoopTree *loop, IdealLoopTree *innermost );
914
915 // Place Data nodes in some loop nest
916 void build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack );
917 void build_loop_late ( VectorSet &visited, Node_List &worklist, Node_Stack &nstack );
918 void build_loop_late_post_work(Node* n, bool pinned);
919 void build_loop_late_post(Node* n);
920 void verify_strip_mined_scheduling(Node *n, Node* least);
921
922 // Array of immediate dominance info for each CFG node indexed by node idx
923 private:
924 uint _idom_size;
925 Node **_idom; // Array of immediate dominators
926 uint *_dom_depth; // Used for fast LCA test
927 GrowableArray<uint>* _dom_stk; // For recomputation of dom depth
928
929 // Perform verification that the graph is valid.
PhaseIdealLoop(PhaseIterGVN & igvn)930 PhaseIdealLoop( PhaseIterGVN &igvn) :
931 PhaseTransform(Ideal_Loop),
932 _igvn(igvn),
933 _verify_me(NULL),
934 _verify_only(true),
935 _dom_lca_tags(arena()), // Thread::resource_area
936 _nodes_required(UINT_MAX) {
937 build_and_optimize(LoopOptsVerify);
938 }
939
940 // build the loop tree and perform any requested optimizations
941 void build_and_optimize(LoopOptsMode mode);
942
943 // Dominators for the sea of nodes
944 void Dominators();
945
946 // Compute the Ideal Node to Loop mapping
PhaseIdealLoop(PhaseIterGVN & igvn,LoopOptsMode mode)947 PhaseIdealLoop(PhaseIterGVN &igvn, LoopOptsMode mode) :
948 PhaseTransform(Ideal_Loop),
949 _igvn(igvn),
950 _verify_me(NULL),
951 _verify_only(false),
952 _dom_lca_tags(arena()), // Thread::resource_area
953 _nodes_required(UINT_MAX) {
954 build_and_optimize(mode);
955 }
956
957 // Verify that verify_me made the same decisions as a fresh run.
PhaseIdealLoop(PhaseIterGVN & igvn,const PhaseIdealLoop * verify_me)958 PhaseIdealLoop(PhaseIterGVN &igvn, const PhaseIdealLoop *verify_me) :
959 PhaseTransform(Ideal_Loop),
960 _igvn(igvn),
961 _verify_me(verify_me),
962 _verify_only(false),
963 _dom_lca_tags(arena()), // Thread::resource_area
964 _nodes_required(UINT_MAX) {
965 build_and_optimize(LoopOptsVerify);
966 }
967
968 public:
idom_no_update(Node * d) const969 Node* idom_no_update(Node* d) const {
970 return idom_no_update(d->_idx);
971 }
972
idom_no_update(uint didx) const973 Node* idom_no_update(uint didx) const {
974 assert(didx < _idom_size, "oob");
975 Node* n = _idom[didx];
976 assert(n != NULL,"Bad immediate dominator info.");
977 while (n->in(0) == NULL) { // Skip dead CFG nodes
978 n = (Node*)(((intptr_t)_nodes[n->_idx]) & ~1);
979 assert(n != NULL,"Bad immediate dominator info.");
980 }
981 return n;
982 }
983
idom(Node * d) const984 Node *idom(Node* d) const {
985 return idom(d->_idx);
986 }
987
idom(uint didx) const988 Node *idom(uint didx) const {
989 Node *n = idom_no_update(didx);
990 _idom[didx] = n; // Lazily remove dead CFG nodes from table.
991 return n;
992 }
993
dom_depth(Node * d) const994 uint dom_depth(Node* d) const {
995 guarantee(d != NULL, "Null dominator info.");
996 guarantee(d->_idx < _idom_size, "");
997 return _dom_depth[d->_idx];
998 }
999 void set_idom(Node* d, Node* n, uint dom_depth);
1000 // Locally compute IDOM using dom_lca call
1001 Node *compute_idom( Node *region ) const;
1002 // Recompute dom_depth
1003 void recompute_dom_depth();
1004
1005 // Is safept not required by an outer loop?
1006 bool is_deleteable_safept(Node* sfpt);
1007
1008 // Replace parallel induction variable (parallel to trip counter)
1009 void replace_parallel_iv(IdealLoopTree *loop);
1010
dom_lca(Node * n1,Node * n2) const1011 Node *dom_lca( Node *n1, Node *n2 ) const {
1012 return find_non_split_ctrl(dom_lca_internal(n1, n2));
1013 }
1014 Node *dom_lca_internal( Node *n1, Node *n2 ) const;
1015
1016 // Build and verify the loop tree without modifying the graph. This
1017 // is useful to verify that all inputs properly dominate their uses.
verify(PhaseIterGVN & igvn)1018 static void verify(PhaseIterGVN& igvn) {
1019 #ifdef ASSERT
1020 ResourceMark rm;
1021 PhaseIdealLoop v(igvn);
1022 #endif
1023 }
1024
1025 // Recommended way to use PhaseIdealLoop.
1026 // Run PhaseIdealLoop in some mode and allocates a local scope for memory allocations.
optimize(PhaseIterGVN & igvn,LoopOptsMode mode)1027 static void optimize(PhaseIterGVN &igvn, LoopOptsMode mode) {
1028 ResourceMark rm;
1029 PhaseIdealLoop v(igvn, mode);
1030 }
1031
1032 // True if the method has at least 1 irreducible loop
1033 bool _has_irreducible_loops;
1034
1035 // Per-Node transform
transform(Node * n)1036 virtual Node* transform(Node* n) { return NULL; }
1037
1038 Node* loop_exit_control(Node* x, IdealLoopTree* loop);
1039 Node* loop_exit_test(Node* back_control, IdealLoopTree* loop, Node*& incr, Node*& limit, BoolTest::mask& bt, float& cl_prob);
1040 Node* loop_iv_incr(Node* incr, Node* x, IdealLoopTree* loop, Node*& phi_incr);
1041 Node* loop_iv_stride(Node* incr, IdealLoopTree* loop, Node*& xphi);
1042 PhiNode* loop_iv_phi(Node* xphi, Node* phi_incr, Node* x, IdealLoopTree* loop);
1043
1044 bool is_counted_loop(Node* n, IdealLoopTree* &loop);
1045 IdealLoopTree* insert_outer_loop(IdealLoopTree* loop, LoopNode* outer_l, Node* outer_ift);
1046 IdealLoopTree* create_outer_strip_mined_loop(BoolNode *test, Node *cmp, Node *init_control,
1047 IdealLoopTree* loop, float cl_prob, float le_fcnt,
1048 Node*& entry_control, Node*& iffalse);
1049
1050 Node* exact_limit( IdealLoopTree *loop );
1051
1052 // Return a post-walked LoopNode
get_loop(Node * n) const1053 IdealLoopTree *get_loop( Node *n ) const {
1054 // Dead nodes have no loop, so return the top level loop instead
1055 if (!has_node(n)) return _ltree_root;
1056 assert(!has_ctrl(n), "");
1057 return (IdealLoopTree*)_nodes[n->_idx];
1058 }
1059
ltree_root() const1060 IdealLoopTree* ltree_root() const { return _ltree_root; }
1061
1062 // Is 'n' a (nested) member of 'loop'?
is_member(const IdealLoopTree * loop,Node * n) const1063 int is_member( const IdealLoopTree *loop, Node *n ) const {
1064 return loop->is_member(get_loop(n)); }
1065
1066 // This is the basic building block of the loop optimizations. It clones an
1067 // entire loop body. It makes an old_new loop body mapping; with this
1068 // mapping you can find the new-loop equivalent to an old-loop node. All
1069 // new-loop nodes are exactly equal to their old-loop counterparts, all
1070 // edges are the same. All exits from the old-loop now have a RegionNode
1071 // that merges the equivalent new-loop path. This is true even for the
1072 // normal "loop-exit" condition. All uses of loop-invariant old-loop values
1073 // now come from (one or more) Phis that merge their new-loop equivalents.
1074 // Parameter side_by_side_idom:
1075 // When side_by_size_idom is NULL, the dominator tree is constructed for
1076 // the clone loop to dominate the original. Used in construction of
1077 // pre-main-post loop sequence.
1078 // When nonnull, the clone and original are side-by-side, both are
1079 // dominated by the passed in side_by_side_idom node. Used in
1080 // construction of unswitched loops.
1081 enum CloneLoopMode {
1082 IgnoreStripMined = 0, // Only clone inner strip mined loop
1083 CloneIncludesStripMined = 1, // clone both inner and outer strip mined loops
1084 ControlAroundStripMined = 2 // Only clone inner strip mined loop,
1085 // result control flow branches
1086 // either to inner clone or outer
1087 // strip mined loop.
1088 };
1089 void clone_loop( IdealLoopTree *loop, Node_List &old_new, int dom_depth,
1090 CloneLoopMode mode, Node* side_by_side_idom = NULL);
1091 void clone_loop_handle_data_uses(Node* old, Node_List &old_new,
1092 IdealLoopTree* loop, IdealLoopTree* companion_loop,
1093 Node_List*& split_if_set, Node_List*& split_bool_set,
1094 Node_List*& split_cex_set, Node_List& worklist,
1095 uint new_counter, CloneLoopMode mode);
1096 void clone_outer_loop(LoopNode* head, CloneLoopMode mode, IdealLoopTree *loop,
1097 IdealLoopTree* outer_loop, int dd, Node_List &old_new,
1098 Node_List& extra_data_nodes);
1099
1100 // If we got the effect of peeling, either by actually peeling or by
1101 // making a pre-loop which must execute at least once, we can remove
1102 // all loop-invariant dominated tests in the main body.
1103 void peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new );
1104
1105 // Generate code to do a loop peel for the given loop (and body).
1106 // old_new is a temp array.
1107 void do_peeling( IdealLoopTree *loop, Node_List &old_new );
1108
1109 // Add pre and post loops around the given loop. These loops are used
1110 // during RCE, unrolling and aligning loops.
1111 void insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only );
1112
1113 // Add post loop after the given loop.
1114 Node *insert_post_loop(IdealLoopTree *loop, Node_List &old_new,
1115 CountedLoopNode *main_head, CountedLoopEndNode *main_end,
1116 Node *incr, Node *limit, CountedLoopNode *&post_head);
1117
1118 // Add an RCE'd post loop which we will multi-version adapt for run time test path usage
1119 void insert_scalar_rced_post_loop( IdealLoopTree *loop, Node_List &old_new );
1120
1121 // Add a vector post loop between a vector main loop and the current post loop
1122 void insert_vector_post_loop(IdealLoopTree *loop, Node_List &old_new);
1123 // If Node n lives in the back_ctrl block, we clone a private version of n
1124 // in preheader_ctrl block and return that, otherwise return n.
1125 Node *clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n, VectorSet &visited, Node_Stack &clones );
1126
1127 // Take steps to maximally unroll the loop. Peel any odd iterations, then
1128 // unroll to do double iterations. The next round of major loop transforms
1129 // will repeat till the doubled loop body does all remaining iterations in 1
1130 // pass.
1131 void do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new );
1132
1133 // Unroll the loop body one step - make each trip do 2 iterations.
1134 void do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip );
1135
1136 // Mark vector reduction candidates before loop unrolling
1137 void mark_reductions( IdealLoopTree *loop );
1138
1139 // Return true if exp is a constant times an induction var
1140 bool is_scaled_iv(Node* exp, Node* iv, int* p_scale);
1141
1142 // Return true if exp is a scaled induction var plus (or minus) constant
1143 bool is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth = 0);
1144
1145 // Create a new if above the uncommon_trap_if_pattern for the predicate to be promoted
1146 ProjNode* create_new_if_for_predicate(ProjNode* cont_proj, Node* new_entry, Deoptimization::DeoptReason reason,
1147 int opcode, bool if_cont_is_true_proj = true);
1148
1149 void register_control(Node* n, IdealLoopTree *loop, Node* pred, bool update_body = true);
1150
1151 static Node* skip_all_loop_predicates(Node* entry);
1152 static Node* skip_loop_predicates(Node* entry);
1153
1154 // Find a good location to insert a predicate
1155 static ProjNode* find_predicate_insertion_point(Node* start_c, Deoptimization::DeoptReason reason);
1156 // Find a predicate
1157 static Node* find_predicate(Node* entry);
1158 // Construct a range check for a predicate if
1159 BoolNode* rc_predicate(IdealLoopTree *loop, Node* ctrl,
1160 int scale, Node* offset,
1161 Node* init, Node* limit, jint stride,
1162 Node* range, bool upper, bool &overflow);
1163
1164 // Implementation of the loop predication to promote checks outside the loop
1165 bool loop_predication_impl(IdealLoopTree *loop);
1166 bool loop_predication_impl_helper(IdealLoopTree *loop, ProjNode* proj, ProjNode *predicate_proj,
1167 CountedLoopNode *cl, ConNode* zero, Invariance& invar,
1168 Deoptimization::DeoptReason reason);
1169 bool loop_predication_should_follow_branches(IdealLoopTree *loop, ProjNode *predicate_proj, float& loop_trip_cnt);
1170 void loop_predication_follow_branches(Node *c, IdealLoopTree *loop, float loop_trip_cnt,
1171 PathFrequency& pf, Node_Stack& stack, VectorSet& seen,
1172 Node_List& if_proj_list);
1173 ProjNode* insert_initial_skeleton_predicate(IfNode* iff, IdealLoopTree *loop,
1174 ProjNode* proj, ProjNode *predicate_proj,
1175 ProjNode* upper_bound_proj,
1176 int scale, Node* offset,
1177 Node* init, Node* limit, jint stride,
1178 Node* rng, bool& overflow,
1179 Deoptimization::DeoptReason reason);
1180 Node* add_range_check_predicate(IdealLoopTree* loop, CountedLoopNode* cl,
1181 Node* predicate_proj, int scale_con, Node* offset,
1182 Node* limit, jint stride_con, Node* value);
1183
1184 // Helper function to collect predicate for eliminating the useless ones
1185 void collect_potentially_useful_predicates(IdealLoopTree *loop, Unique_Node_List &predicate_opaque1);
1186 void eliminate_useless_predicates();
1187
1188 // Change the control input of expensive nodes to allow commoning by
1189 // IGVN when it is guaranteed to not result in a more frequent
1190 // execution of the expensive node. Return true if progress.
1191 bool process_expensive_nodes();
1192
1193 // Check whether node has become unreachable
is_node_unreachable(Node * n) const1194 bool is_node_unreachable(Node *n) const {
1195 return !has_node(n) || n->is_unreachable(_igvn);
1196 }
1197
1198 // Eliminate range-checks and other trip-counter vs loop-invariant tests.
1199 int do_range_check( IdealLoopTree *loop, Node_List &old_new );
1200
1201 // Check to see if do_range_check(...) cleaned the main loop of range-checks
1202 void has_range_checks(IdealLoopTree *loop);
1203
1204 // Process post loops which have range checks and try to build a multi-version
1205 // guard to safely determine if we can execute the post loop which was RCE'd.
1206 bool multi_version_post_loops(IdealLoopTree *rce_loop, IdealLoopTree *legacy_loop);
1207
1208 // Cause the rce'd post loop to optimized away, this happens if we cannot complete multiverioning
1209 void poison_rce_post_loop(IdealLoopTree *rce_loop);
1210
1211 // Create a slow version of the loop by cloning the loop
1212 // and inserting an if to select fast-slow versions.
1213 ProjNode* create_slow_version_of_loop(IdealLoopTree *loop,
1214 Node_List &old_new,
1215 int opcode,
1216 CloneLoopMode mode);
1217
1218 // Clone a loop and return the clone head (clone_loop_head).
1219 // Added nodes include int(1), int(0) - disconnected, If, IfTrue, IfFalse,
1220 // This routine was created for usage in CountedLoopReserveKit.
1221 //
1222 // int(1) -> If -> IfTrue -> original_loop_head
1223 // |
1224 // V
1225 // IfFalse -> clone_loop_head (returned by function pointer)
1226 //
1227 LoopNode* create_reserve_version_of_loop(IdealLoopTree *loop, CountedLoopReserveKit* lk);
1228 // Clone loop with an invariant test (that does not exit) and
1229 // insert a clone of the test that selects which version to
1230 // execute.
1231 void do_unswitching (IdealLoopTree *loop, Node_List &old_new);
1232
1233 // Find candidate "if" for unswitching
1234 IfNode* find_unswitching_candidate(const IdealLoopTree *loop) const;
1235
1236 // Range Check Elimination uses this function!
1237 // Constrain the main loop iterations so the affine function:
1238 // low_limit <= scale_con * I + offset < upper_limit
1239 // always holds true. That is, either increase the number of iterations in
1240 // the pre-loop or the post-loop until the condition holds true in the main
1241 // loop. Scale_con, offset and limit are all loop invariant.
1242 void add_constraint(jlong stride_con, jlong scale_con, Node* offset, Node* low_limit, Node* upper_limit, Node* pre_ctrl, Node** pre_limit, Node** main_limit);
1243 // Helper function for add_constraint().
1244 Node* adjust_limit(bool reduce, Node* scale, Node* offset, Node* rc_limit, Node* old_limit, Node* pre_ctrl, bool round);
1245
1246 // Partially peel loop up through last_peel node.
1247 bool partial_peel( IdealLoopTree *loop, Node_List &old_new );
1248
1249 // Create a scheduled list of nodes control dependent on ctrl set.
1250 void scheduled_nodelist( IdealLoopTree *loop, VectorSet& ctrl, Node_List &sched );
1251 // Has a use in the vector set
1252 bool has_use_in_set( Node* n, VectorSet& vset );
1253 // Has use internal to the vector set (ie. not in a phi at the loop head)
1254 bool has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop );
1255 // clone "n" for uses that are outside of loop
1256 int clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist );
1257 // clone "n" for special uses that are in the not_peeled region
1258 void clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
1259 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist );
1260 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
1261 void insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp );
1262 #ifdef ASSERT
1263 // Validate the loop partition sets: peel and not_peel
1264 bool is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, VectorSet& not_peel );
1265 // Ensure that uses outside of loop are of the right form
1266 bool is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
1267 uint orig_exit_idx, uint clone_exit_idx);
1268 bool is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx);
1269 #endif
1270
1271 // Returns nonzero constant stride if-node is a possible iv test (otherwise returns zero.)
1272 int stride_of_possible_iv( Node* iff );
is_possible_iv_test(Node * iff)1273 bool is_possible_iv_test( Node* iff ) { return stride_of_possible_iv(iff) != 0; }
1274 // Return the (unique) control output node that's in the loop (if it exists.)
1275 Node* stay_in_loop( Node* n, IdealLoopTree *loop);
1276 // Insert a signed compare loop exit cloned from an unsigned compare.
1277 IfNode* insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop);
1278 void remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop);
1279 // Utility to register node "n" with PhaseIdealLoop
1280 void register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth);
1281 // Utility to create an if-projection
1282 ProjNode* proj_clone(ProjNode* p, IfNode* iff);
1283 // Force the iff control output to be the live_proj
1284 Node* short_circuit_if(IfNode* iff, ProjNode* live_proj);
1285 // Insert a region before an if projection
1286 RegionNode* insert_region_before_proj(ProjNode* proj);
1287 // Insert a new if before an if projection
1288 ProjNode* insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj);
1289
1290 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1291 // "Nearly" because all Nodes have been cloned from the original in the loop,
1292 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1293 // through the Phi recursively, and return a Bool.
1294 Node *clone_iff( PhiNode *phi, IdealLoopTree *loop );
1295 CmpNode *clone_bool( PhiNode *phi, IdealLoopTree *loop );
1296
1297
1298 // Rework addressing expressions to get the most loop-invariant stuff
1299 // moved out. We'd like to do all associative operators, but it's especially
1300 // important (common) to do address expressions.
1301 Node *remix_address_expressions( Node *n );
1302
1303 // Convert add to muladd to generate MuladdS2I under certain criteria
1304 Node * convert_add_to_muladd(Node * n);
1305
1306 // Attempt to use a conditional move instead of a phi/branch
1307 Node *conditional_move( Node *n );
1308
1309 // Reorganize offset computations to lower register pressure.
1310 // Mostly prevent loop-fallout uses of the pre-incremented trip counter
1311 // (which are then alive with the post-incremented trip counter
1312 // forcing an extra register move)
1313 void reorg_offsets( IdealLoopTree *loop );
1314
1315 // Check for aggressive application of 'split-if' optimization,
1316 // using basic block level info.
1317 void split_if_with_blocks ( VectorSet &visited, Node_Stack &nstack);
1318 Node *split_if_with_blocks_pre ( Node *n );
1319 void split_if_with_blocks_post( Node *n );
1320 Node *has_local_phi_input( Node *n );
1321 // Mark an IfNode as being dominated by a prior test,
1322 // without actually altering the CFG (and hence IDOM info).
1323 void dominated_by( Node *prevdom, Node *iff, bool flip = false, bool exclude_loop_predicate = false );
1324
1325 // Split Node 'n' through merge point
1326 Node *split_thru_region( Node *n, Node *region );
1327 // Split Node 'n' through merge point if there is enough win.
1328 Node *split_thru_phi( Node *n, Node *region, int policy );
1329 // Found an If getting its condition-code input from a Phi in the
1330 // same block. Split thru the Region.
1331 void do_split_if( Node *iff );
1332
1333 // Conversion of fill/copy patterns into intrinsic versions
1334 bool do_intrinsify_fill();
1335 bool intrinsify_fill(IdealLoopTree* lpt);
1336 bool match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value,
1337 Node*& shift, Node*& offset);
1338
1339 private:
1340 // Return a type based on condition control flow
1341 const TypeInt* filtered_type( Node *n, Node* n_ctrl);
filtered_type(Node * n)1342 const TypeInt* filtered_type( Node *n ) { return filtered_type(n, NULL); }
1343 // Helpers for filtered type
1344 const TypeInt* filtered_type_from_dominators( Node* val, Node *val_ctrl);
1345
1346 // Helper functions
1347 Node *spinup( Node *iff, Node *new_false, Node *new_true, Node *region, Node *phi, small_cache *cache );
1348 Node *find_use_block( Node *use, Node *def, Node *old_false, Node *new_false, Node *old_true, Node *new_true );
1349 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 );
1350 bool split_up( Node *n, Node *blk1, Node *blk2 );
1351 void sink_use( Node *use, Node *post_loop );
1352 Node *place_near_use( Node *useblock ) const;
1353 Node* try_move_store_before_loop(Node* n, Node *n_ctrl);
1354 void try_move_store_after_loop(Node* n);
1355 bool identical_backtoback_ifs(Node *n);
1356 bool can_split_if(Node *n_ctrl);
1357
1358 // Determine if a method is too big for a/another round of split-if, based on
1359 // a magic (approximate) ratio derived from the equally magic constant 35000,
1360 // previously used for this purpose (but without relating to the node limit).
must_throttle_split_if()1361 bool must_throttle_split_if() {
1362 uint threshold = C->max_node_limit() * 2 / 5;
1363 return C->live_nodes() > threshold;
1364 }
1365
1366 // A simplistic node request tracking mechanism, where
1367 // = UINT_MAX Request not valid or made final.
1368 // < UINT_MAX Nodes currently requested (estimate).
1369 uint _nodes_required;
1370
1371 enum { REQUIRE_MIN = 70 };
1372
nodes_required() const1373 uint nodes_required() const { return _nodes_required; }
1374
1375 // Given the _currently_ available number of nodes, check whether there is
1376 // "room" for an additional request or not, considering the already required
1377 // number of nodes. Return TRUE if the new request is exceeding the node
1378 // budget limit, otherwise return FALSE. Note that this interpretation will
1379 // act pessimistic on additional requests when new nodes have already been
1380 // generated since the 'begin'. This behaviour fits with the intention that
1381 // node estimates/requests should be made upfront.
exceeding_node_budget(uint required=0)1382 bool exceeding_node_budget(uint required = 0) {
1383 assert(C->live_nodes() < C->max_node_limit(), "sanity");
1384 uint available = C->max_node_limit() - C->live_nodes();
1385 return available < required + _nodes_required + REQUIRE_MIN;
1386 }
1387
require_nodes(uint require,uint minreq=REQUIRE_MIN)1388 uint require_nodes(uint require, uint minreq = REQUIRE_MIN) {
1389 precond(require > 0);
1390 _nodes_required += MAX2(require, minreq);
1391 return _nodes_required;
1392 }
1393
may_require_nodes(uint require,uint minreq=REQUIRE_MIN)1394 bool may_require_nodes(uint require, uint minreq = REQUIRE_MIN) {
1395 return !exceeding_node_budget(require) && require_nodes(require, minreq) > 0;
1396 }
1397
require_nodes_begin()1398 uint require_nodes_begin() {
1399 assert(_nodes_required == UINT_MAX, "Bad state (begin).");
1400 _nodes_required = 0;
1401 return C->live_nodes();
1402 }
1403
1404 // When a node request is final, optionally check that the requested number
1405 // of nodes was reasonably correct with respect to the number of new nodes
1406 // introduced since the last 'begin'. Always check that we have not exceeded
1407 // the maximum node limit.
require_nodes_final(uint live_at_begin,bool check_estimate)1408 void require_nodes_final(uint live_at_begin, bool check_estimate) {
1409 assert(_nodes_required < UINT_MAX, "Bad state (final).");
1410
1411 #ifdef ASSERT
1412 if (check_estimate) {
1413 // Check that the node budget request was not off by too much (x2).
1414 // Should this be the case we _surely_ need to improve the estimates
1415 // used in our budget calculations.
1416 if (C->live_nodes() - live_at_begin > 2 * _nodes_required) {
1417 log_info(compilation)("Bad node estimate: actual = %d >> request = %d",
1418 C->live_nodes() - live_at_begin, _nodes_required);
1419 }
1420 }
1421 #endif
1422 // Assert that we have stayed within the node budget limit.
1423 assert(C->live_nodes() < C->max_node_limit(),
1424 "Exceeding node budget limit: %d + %d > %d (request = %d)",
1425 C->live_nodes() - live_at_begin, live_at_begin,
1426 C->max_node_limit(), _nodes_required);
1427
1428 _nodes_required = UINT_MAX;
1429 }
1430
1431 // Clone loop predicates to slow and fast loop when unswitching a loop
1432 Node* clone_loop_predicates(Node* old_entry, Node* new_entry, bool clone_limit_check, bool is_slow_loop,
1433 uint idx_before_clone, Node_List &old_new);
1434 ProjNode* clone_loop_predicate(ProjNode* predicate_proj, Node* new_entry, Deoptimization::DeoptReason reason,
1435 bool is_slow_loop, uint idx_before_clone, Node_List &old_new);
1436 void clone_concrete_loop_predicates(Deoptimization::DeoptReason reason, ProjNode* old_predicate_proj,
1437 ProjNode* new_predicate_proj, bool is_slow_loop,
1438 uint idx_before_clone, Node_List &old_new);
1439
1440 bool _created_loop_node;
1441
1442 public:
set_created_loop_node()1443 void set_created_loop_node() { _created_loop_node = true; }
created_loop_node()1444 bool created_loop_node() { return _created_loop_node; }
1445 void register_new_node(Node* n, Node* blk);
1446
1447 #ifdef ASSERT
1448 void dump_bad_graph(const char* msg, Node* n, Node* early, Node* LCA);
1449 #endif
1450
1451 #ifndef PRODUCT
1452 void dump() const;
1453 void dump(IdealLoopTree* loop, uint rpo_idx, Node_List &rpo_list) const;
1454 void verify() const; // Major slow :-)
1455 void verify_compare(Node* n, const PhaseIdealLoop* loop_verify, VectorSet &visited) const;
get_loop_idx(Node * n) const1456 IdealLoopTree* get_loop_idx(Node* n) const {
1457 // Dead nodes have no loop, so return the top level loop instead
1458 return _nodes[n->_idx] ? (IdealLoopTree*)_nodes[n->_idx] : _ltree_root;
1459 }
1460 // Print some stats
1461 static void print_statistics();
1462 static int _loop_invokes; // Count of PhaseIdealLoop invokes
1463 static int _loop_work; // Sum of PhaseIdealLoop x _unique
1464 #endif
1465
1466 void rpo(Node* start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list) const;
1467 };
1468
1469
1470 class AutoNodeBudget : public StackObj
1471 {
1472 public:
1473 enum budget_check_t { BUDGET_CHECK, NO_BUDGET_CHECK };
1474
AutoNodeBudget(PhaseIdealLoop * phase,budget_check_t chk=BUDGET_CHECK)1475 AutoNodeBudget(PhaseIdealLoop* phase, budget_check_t chk = BUDGET_CHECK)
1476 : _phase(phase),
1477 _check_at_final(chk == BUDGET_CHECK),
1478 _nodes_at_begin(0)
1479 {
1480 precond(_phase != NULL);
1481
1482 _nodes_at_begin = _phase->require_nodes_begin();
1483 }
1484
~AutoNodeBudget()1485 ~AutoNodeBudget() {
1486 #ifndef PRODUCT
1487 if (TraceLoopOpts) {
1488 uint request = _phase->nodes_required();
1489 uint delta = _phase->C->live_nodes() - _nodes_at_begin;
1490
1491 if (request < delta) {
1492 tty->print_cr("Exceeding node budget: %d < %d", request, delta);
1493 } else {
1494 uint const REQUIRE_MIN = PhaseIdealLoop::REQUIRE_MIN;
1495 // Identify the worst estimates as "poor" ones.
1496 if (request > REQUIRE_MIN && delta > 0) {
1497 if ((delta > REQUIRE_MIN && request > 3 * delta) ||
1498 (delta <= REQUIRE_MIN && request > 10 * delta)) {
1499 tty->print_cr("Poor node estimate: %d >> %d", request, delta);
1500 }
1501 }
1502 }
1503 }
1504 #endif // PRODUCT
1505 _phase->require_nodes_final(_nodes_at_begin, _check_at_final);
1506 }
1507
1508 private:
1509 PhaseIdealLoop* _phase;
1510 bool _check_at_final;
1511 uint _nodes_at_begin;
1512 };
1513
1514
1515 // This kit may be used for making of a reserved copy of a loop before this loop
1516 // goes under non-reversible changes.
1517 //
1518 // Function create_reserve() creates a reserved copy (clone) of the loop.
1519 // The reserved copy is created by calling
1520 // PhaseIdealLoop::create_reserve_version_of_loop - see there how
1521 // the original and reserved loops are connected in the outer graph.
1522 // If create_reserve succeeded, it returns 'true' and _has_reserved is set to 'true'.
1523 //
1524 // By default the reserved copy (clone) of the loop is created as dead code - it is
1525 // dominated in the outer loop by this node chain:
1526 // intcon(1)->If->IfFalse->reserved_copy.
1527 // The original loop is dominated by the the same node chain but IfTrue projection:
1528 // intcon(0)->If->IfTrue->original_loop.
1529 //
1530 // In this implementation of CountedLoopReserveKit the ctor includes create_reserve()
1531 // and the dtor, checks _use_new value.
1532 // If _use_new == false, it "switches" control to reserved copy of the loop
1533 // by simple replacing of node intcon(1) with node intcon(0).
1534 //
1535 // Here is a proposed example of usage (see also SuperWord::output in superword.cpp).
1536 //
1537 // void CountedLoopReserveKit_example()
1538 // {
1539 // CountedLoopReserveKit lrk((phase, lpt, DoReserveCopy = true); // create local object
1540 // if (DoReserveCopy && !lrk.has_reserved()) {
1541 // return; //failed to create reserved loop copy
1542 // }
1543 // ...
1544 // //something is wrong, switch to original loop
1545 /// if(something_is_wrong) return; // ~CountedLoopReserveKit makes the switch
1546 // ...
1547 // //everything worked ok, return with the newly modified loop
1548 // lrk.use_new();
1549 // return; // ~CountedLoopReserveKit does nothing once use_new() was called
1550 // }
1551 //
1552 // Keep in mind, that by default if create_reserve() is not followed by use_new()
1553 // the dtor will "switch to the original" loop.
1554 // NOTE. You you modify outside of the original loop this class is no help.
1555 //
1556 class CountedLoopReserveKit {
1557 private:
1558 PhaseIdealLoop* _phase;
1559 IdealLoopTree* _lpt;
1560 LoopNode* _lp;
1561 IfNode* _iff;
1562 LoopNode* _lp_reserved;
1563 bool _has_reserved;
1564 bool _use_new;
1565 const bool _active; //may be set to false in ctor, then the object is dummy
1566
1567 public:
1568 CountedLoopReserveKit(PhaseIdealLoop* phase, IdealLoopTree *loop, bool active);
1569 ~CountedLoopReserveKit();
use_new()1570 void use_new() {_use_new = true;}
set_iff(IfNode * x)1571 void set_iff(IfNode* x) {_iff = x;}
has_reserved() const1572 bool has_reserved() const { return _active && _has_reserved;}
1573 private:
1574 bool create_reserve();
1575 };// class CountedLoopReserveKit
1576
tail()1577 inline Node* IdealLoopTree::tail() {
1578 // Handle lazy update of _tail field.
1579 if (_tail->in(0) == NULL) {
1580 _tail = _phase->get_ctrl(_tail);
1581 }
1582 return _tail;
1583 }
1584
head()1585 inline Node* IdealLoopTree::head() {
1586 // Handle lazy update of _head field.
1587 if (_head->in(0) == NULL) {
1588 _head = _phase->get_ctrl(_head);
1589 }
1590 return _head;
1591 }
1592
1593 // Iterate over the loop tree using a preorder, left-to-right traversal.
1594 //
1595 // Example that visits all counted loops from within PhaseIdealLoop
1596 //
1597 // for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
1598 // IdealLoopTree* lpt = iter.current();
1599 // if (!lpt->is_counted()) continue;
1600 // ...
1601 class LoopTreeIterator : public StackObj {
1602 private:
1603 IdealLoopTree* _root;
1604 IdealLoopTree* _curnt;
1605
1606 public:
LoopTreeIterator(IdealLoopTree * root)1607 LoopTreeIterator(IdealLoopTree* root) : _root(root), _curnt(root) {}
1608
done()1609 bool done() { return _curnt == NULL; } // Finished iterating?
1610
1611 void next(); // Advance to next loop tree
1612
current()1613 IdealLoopTree* current() { return _curnt; } // Return current value of iterator.
1614 };
1615
1616 #endif // SHARE_OPTO_LOOPNODE_HPP
1617