1
|
1 |
/*
|
|
2 |
* Copyright 1998-2007 Sun Microsystems, Inc. All Rights Reserved.
|
|
3 |
* DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
|
|
4 |
*
|
|
5 |
* This code is free software; you can redistribute it and/or modify it
|
|
6 |
* under the terms of the GNU General Public License version 2 only, as
|
|
7 |
* published by the Free Software Foundation.
|
|
8 |
*
|
|
9 |
* This code is distributed in the hope that it will be useful, but WITHOUT
|
|
10 |
* ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
|
|
11 |
* FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
|
|
12 |
* version 2 for more details (a copy is included in the LICENSE file that
|
|
13 |
* accompanied this code).
|
|
14 |
*
|
|
15 |
* You should have received a copy of the GNU General Public License version
|
|
16 |
* 2 along with this work; if not, write to the Free Software Foundation,
|
|
17 |
* Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
|
|
18 |
*
|
|
19 |
* Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
|
|
20 |
* CA 95054 USA or visit www.sun.com if you need additional information or
|
|
21 |
* have any questions.
|
|
22 |
*
|
|
23 |
*/
|
|
24 |
|
|
25 |
class CmpNode;
|
|
26 |
class CountedLoopEndNode;
|
|
27 |
class CountedLoopNode;
|
|
28 |
class IdealLoopTree;
|
|
29 |
class LoopNode;
|
|
30 |
class Node;
|
|
31 |
class PhaseIdealLoop;
|
|
32 |
class VectorSet;
|
|
33 |
struct small_cache;
|
|
34 |
|
|
35 |
//
|
|
36 |
// I D E A L I Z E D L O O P S
|
|
37 |
//
|
|
38 |
// Idealized loops are the set of loops I perform more interesting
|
|
39 |
// transformations on, beyond simple hoisting.
|
|
40 |
|
|
41 |
//------------------------------LoopNode---------------------------------------
|
|
42 |
// Simple loop header. Fall in path on left, loop-back path on right.
|
|
43 |
class LoopNode : public RegionNode {
|
|
44 |
// Size is bigger to hold the flags. However, the flags do not change
|
|
45 |
// the semantics so it does not appear in the hash & cmp functions.
|
|
46 |
virtual uint size_of() const { return sizeof(*this); }
|
|
47 |
protected:
|
|
48 |
short _loop_flags;
|
|
49 |
// Names for flag bitfields
|
|
50 |
enum { pre_post_main=0, inner_loop=8, partial_peel_loop=16, partial_peel_failed=32 };
|
|
51 |
char _unswitch_count;
|
|
52 |
enum { _unswitch_max=3 };
|
|
53 |
|
|
54 |
public:
|
|
55 |
// Names for edge indices
|
|
56 |
enum { Self=0, EntryControl, LoopBackControl };
|
|
57 |
|
|
58 |
int is_inner_loop() const { return _loop_flags & inner_loop; }
|
|
59 |
void set_inner_loop() { _loop_flags |= inner_loop; }
|
|
60 |
|
|
61 |
int is_partial_peel_loop() const { return _loop_flags & partial_peel_loop; }
|
|
62 |
void set_partial_peel_loop() { _loop_flags |= partial_peel_loop; }
|
|
63 |
int partial_peel_has_failed() const { return _loop_flags & partial_peel_failed; }
|
|
64 |
void mark_partial_peel_failed() { _loop_flags |= partial_peel_failed; }
|
|
65 |
|
|
66 |
int unswitch_max() { return _unswitch_max; }
|
|
67 |
int unswitch_count() { return _unswitch_count; }
|
|
68 |
void set_unswitch_count(int val) {
|
|
69 |
assert (val <= unswitch_max(), "too many unswitches");
|
|
70 |
_unswitch_count = val;
|
|
71 |
}
|
|
72 |
|
|
73 |
LoopNode( Node *entry, Node *backedge ) : RegionNode(3), _loop_flags(0), _unswitch_count(0) {
|
|
74 |
init_class_id(Class_Loop);
|
|
75 |
init_req(EntryControl, entry);
|
|
76 |
init_req(LoopBackControl, backedge);
|
|
77 |
}
|
|
78 |
|
|
79 |
virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
|
|
80 |
virtual int Opcode() const;
|
|
81 |
bool can_be_counted_loop(PhaseTransform* phase) const {
|
|
82 |
return req() == 3 && in(0) != NULL &&
|
|
83 |
in(1) != NULL && phase->type(in(1)) != Type::TOP &&
|
|
84 |
in(2) != NULL && phase->type(in(2)) != Type::TOP;
|
|
85 |
}
|
|
86 |
#ifndef PRODUCT
|
|
87 |
virtual void dump_spec(outputStream *st) const;
|
|
88 |
#endif
|
|
89 |
};
|
|
90 |
|
|
91 |
//------------------------------Counted Loops----------------------------------
|
|
92 |
// Counted loops are all trip-counted loops, with exactly 1 trip-counter exit
|
|
93 |
// path (and maybe some other exit paths). The trip-counter exit is always
|
|
94 |
// last in the loop. The trip-counter does not have to stride by a constant,
|
|
95 |
// but it does have to stride by a loop-invariant amount; the exit value is
|
|
96 |
// also loop invariant.
|
|
97 |
|
|
98 |
// CountedLoopNodes and CountedLoopEndNodes come in matched pairs. The
|
|
99 |
// CountedLoopNode has the incoming loop control and the loop-back-control
|
|
100 |
// which is always the IfTrue before the matching CountedLoopEndNode. The
|
|
101 |
// CountedLoopEndNode has an incoming control (possibly not the
|
|
102 |
// CountedLoopNode if there is control flow in the loop), the post-increment
|
|
103 |
// trip-counter value, and the limit. The trip-counter value is always of
|
|
104 |
// the form (Op old-trip-counter stride). The old-trip-counter is produced
|
|
105 |
// by a Phi connected to the CountedLoopNode. The stride is loop invariant.
|
|
106 |
// The Op is any commutable opcode, including Add, Mul, Xor. The
|
|
107 |
// CountedLoopEndNode also takes in the loop-invariant limit value.
|
|
108 |
|
|
109 |
// From a CountedLoopNode I can reach the matching CountedLoopEndNode via the
|
|
110 |
// loop-back control. From CountedLoopEndNodes I can reach CountedLoopNodes
|
|
111 |
// via the old-trip-counter from the Op node.
|
|
112 |
|
|
113 |
//------------------------------CountedLoopNode--------------------------------
|
|
114 |
// CountedLoopNodes head simple counted loops. CountedLoopNodes have as
|
|
115 |
// inputs the incoming loop-start control and the loop-back control, so they
|
|
116 |
// act like RegionNodes. They also take in the initial trip counter, the
|
|
117 |
// loop-invariant stride and the loop-invariant limit value. CountedLoopNodes
|
|
118 |
// produce a loop-body control and the trip counter value. Since
|
|
119 |
// CountedLoopNodes behave like RegionNodes I still have a standard CFG model.
|
|
120 |
|
|
121 |
class CountedLoopNode : public LoopNode {
|
|
122 |
// Size is bigger to hold _main_idx. However, _main_idx does not change
|
|
123 |
// the semantics so it does not appear in the hash & cmp functions.
|
|
124 |
virtual uint size_of() const { return sizeof(*this); }
|
|
125 |
|
|
126 |
// For Pre- and Post-loops during debugging ONLY, this holds the index of
|
|
127 |
// the Main CountedLoop. Used to assert that we understand the graph shape.
|
|
128 |
node_idx_t _main_idx;
|
|
129 |
|
|
130 |
// Known trip count calculated by policy_maximally_unroll
|
|
131 |
int _trip_count;
|
|
132 |
|
|
133 |
// Expected trip count from profile data
|
|
134 |
float _profile_trip_cnt;
|
|
135 |
|
|
136 |
// Log2 of original loop bodies in unrolled loop
|
|
137 |
int _unrolled_count_log2;
|
|
138 |
|
|
139 |
// Node count prior to last unrolling - used to decide if
|
|
140 |
// unroll,optimize,unroll,optimize,... is making progress
|
|
141 |
int _node_count_before_unroll;
|
|
142 |
|
|
143 |
public:
|
|
144 |
CountedLoopNode( Node *entry, Node *backedge )
|
|
145 |
: LoopNode(entry, backedge), _trip_count(max_jint),
|
|
146 |
_profile_trip_cnt(COUNT_UNKNOWN), _unrolled_count_log2(0),
|
|
147 |
_node_count_before_unroll(0) {
|
|
148 |
init_class_id(Class_CountedLoop);
|
|
149 |
// Initialize _trip_count to the largest possible value.
|
|
150 |
// Will be reset (lower) if the loop's trip count is known.
|
|
151 |
}
|
|
152 |
|
|
153 |
virtual int Opcode() const;
|
|
154 |
virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
|
|
155 |
|
|
156 |
Node *init_control() const { return in(EntryControl); }
|
|
157 |
Node *back_control() const { return in(LoopBackControl); }
|
|
158 |
CountedLoopEndNode *loopexit() const;
|
|
159 |
Node *init_trip() const;
|
|
160 |
Node *stride() const;
|
|
161 |
int stride_con() const;
|
|
162 |
bool stride_is_con() const;
|
|
163 |
Node *limit() const;
|
|
164 |
Node *incr() const;
|
|
165 |
Node *phi() const;
|
|
166 |
|
|
167 |
// Match increment with optional truncation
|
|
168 |
static Node* match_incr_with_optional_truncation(Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type);
|
|
169 |
|
|
170 |
// A 'main' loop has a pre-loop and a post-loop. The 'main' loop
|
|
171 |
// can run short a few iterations and may start a few iterations in.
|
|
172 |
// It will be RCE'd and unrolled and aligned.
|
|
173 |
|
|
174 |
// A following 'post' loop will run any remaining iterations. Used
|
|
175 |
// during Range Check Elimination, the 'post' loop will do any final
|
|
176 |
// iterations with full checks. Also used by Loop Unrolling, where
|
|
177 |
// the 'post' loop will do any epilog iterations needed. Basically,
|
|
178 |
// a 'post' loop can not profitably be further unrolled or RCE'd.
|
|
179 |
|
|
180 |
// A preceding 'pre' loop will run at least 1 iteration (to do peeling),
|
|
181 |
// it may do under-flow checks for RCE and may do alignment iterations
|
|
182 |
// so the following main loop 'knows' that it is striding down cache
|
|
183 |
// lines.
|
|
184 |
|
|
185 |
// A 'main' loop that is ONLY unrolled or peeled, never RCE'd or
|
|
186 |
// Aligned, may be missing it's pre-loop.
|
|
187 |
enum { Normal=0, Pre=1, Main=2, Post=3, PrePostFlagsMask=3, Main_Has_No_Pre_Loop=4 };
|
|
188 |
int is_normal_loop() const { return (_loop_flags&PrePostFlagsMask) == Normal; }
|
|
189 |
int is_pre_loop () const { return (_loop_flags&PrePostFlagsMask) == Pre; }
|
|
190 |
int is_main_loop () const { return (_loop_flags&PrePostFlagsMask) == Main; }
|
|
191 |
int is_post_loop () const { return (_loop_flags&PrePostFlagsMask) == Post; }
|
|
192 |
int is_main_no_pre_loop() const { return _loop_flags & Main_Has_No_Pre_Loop; }
|
|
193 |
void set_main_no_pre_loop() { _loop_flags |= Main_Has_No_Pre_Loop; }
|
|
194 |
|
|
195 |
|
|
196 |
void set_pre_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Pre ; _main_idx = main->_idx; }
|
|
197 |
void set_main_loop ( ) { assert(is_normal_loop(),""); _loop_flags |= Main; }
|
|
198 |
void set_post_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Post; _main_idx = main->_idx; }
|
|
199 |
void set_normal_loop( ) { _loop_flags &= ~PrePostFlagsMask; }
|
|
200 |
|
|
201 |
void set_trip_count(int tc) { _trip_count = tc; }
|
|
202 |
int trip_count() { return _trip_count; }
|
|
203 |
|
|
204 |
void set_profile_trip_cnt(float ptc) { _profile_trip_cnt = ptc; }
|
|
205 |
float profile_trip_cnt() { return _profile_trip_cnt; }
|
|
206 |
|
|
207 |
void double_unrolled_count() { _unrolled_count_log2++; }
|
|
208 |
int unrolled_count() { return 1 << MIN2(_unrolled_count_log2, BitsPerInt-3); }
|
|
209 |
|
|
210 |
void set_node_count_before_unroll(int ct) { _node_count_before_unroll = ct; }
|
|
211 |
int node_count_before_unroll() { return _node_count_before_unroll; }
|
|
212 |
|
|
213 |
#ifndef PRODUCT
|
|
214 |
virtual void dump_spec(outputStream *st) const;
|
|
215 |
#endif
|
|
216 |
};
|
|
217 |
|
|
218 |
//------------------------------CountedLoopEndNode-----------------------------
|
|
219 |
// CountedLoopEndNodes end simple trip counted loops. They act much like
|
|
220 |
// IfNodes.
|
|
221 |
class CountedLoopEndNode : public IfNode {
|
|
222 |
public:
|
|
223 |
enum { TestControl, TestValue };
|
|
224 |
|
|
225 |
CountedLoopEndNode( Node *control, Node *test, float prob, float cnt )
|
|
226 |
: IfNode( control, test, prob, cnt) {
|
|
227 |
init_class_id(Class_CountedLoopEnd);
|
|
228 |
}
|
|
229 |
virtual int Opcode() const;
|
|
230 |
|
|
231 |
Node *cmp_node() const { return (in(TestValue)->req() >=2) ? in(TestValue)->in(1) : NULL; }
|
|
232 |
Node *incr() const { Node *tmp = cmp_node(); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; }
|
|
233 |
Node *limit() const { Node *tmp = cmp_node(); return (tmp && tmp->req()==3) ? tmp->in(2) : NULL; }
|
|
234 |
Node *stride() const { Node *tmp = incr (); return (tmp && tmp->req()==3) ? tmp->in(2) : NULL; }
|
|
235 |
Node *phi() const { Node *tmp = incr (); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; }
|
|
236 |
Node *init_trip() const { Node *tmp = phi (); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; }
|
|
237 |
int stride_con() const;
|
|
238 |
bool stride_is_con() const { Node *tmp = stride (); return (tmp != NULL && tmp->is_Con()); }
|
|
239 |
BoolTest::mask test_trip() const { return in(TestValue)->as_Bool()->_test._test; }
|
|
240 |
CountedLoopNode *loopnode() const {
|
|
241 |
Node *ln = phi()->in(0);
|
|
242 |
assert( ln->Opcode() == Op_CountedLoop, "malformed loop" );
|
|
243 |
return (CountedLoopNode*)ln; }
|
|
244 |
|
|
245 |
#ifndef PRODUCT
|
|
246 |
virtual void dump_spec(outputStream *st) const;
|
|
247 |
#endif
|
|
248 |
};
|
|
249 |
|
|
250 |
|
|
251 |
inline CountedLoopEndNode *CountedLoopNode::loopexit() const {
|
|
252 |
Node *bc = back_control();
|
|
253 |
if( bc == NULL ) return NULL;
|
|
254 |
Node *le = bc->in(0);
|
|
255 |
if( le->Opcode() != Op_CountedLoopEnd )
|
|
256 |
return NULL;
|
|
257 |
return (CountedLoopEndNode*)le;
|
|
258 |
}
|
|
259 |
inline Node *CountedLoopNode::init_trip() const { return loopexit() ? loopexit()->init_trip() : NULL; }
|
|
260 |
inline Node *CountedLoopNode::stride() const { return loopexit() ? loopexit()->stride() : NULL; }
|
|
261 |
inline int CountedLoopNode::stride_con() const { return loopexit() ? loopexit()->stride_con() : 0; }
|
|
262 |
inline bool CountedLoopNode::stride_is_con() const { return loopexit() && loopexit()->stride_is_con(); }
|
|
263 |
inline Node *CountedLoopNode::limit() const { return loopexit() ? loopexit()->limit() : NULL; }
|
|
264 |
inline Node *CountedLoopNode::incr() const { return loopexit() ? loopexit()->incr() : NULL; }
|
|
265 |
inline Node *CountedLoopNode::phi() const { return loopexit() ? loopexit()->phi() : NULL; }
|
|
266 |
|
|
267 |
|
|
268 |
// -----------------------------IdealLoopTree----------------------------------
|
|
269 |
class IdealLoopTree : public ResourceObj {
|
|
270 |
public:
|
|
271 |
IdealLoopTree *_parent; // Parent in loop tree
|
|
272 |
IdealLoopTree *_next; // Next sibling in loop tree
|
|
273 |
IdealLoopTree *_child; // First child in loop tree
|
|
274 |
|
|
275 |
// The head-tail backedge defines the loop.
|
|
276 |
// If tail is NULL then this loop has multiple backedges as part of the
|
|
277 |
// same loop. During cleanup I'll peel off the multiple backedges; merge
|
|
278 |
// them at the loop bottom and flow 1 real backedge into the loop.
|
|
279 |
Node *_head; // Head of loop
|
|
280 |
Node *_tail; // Tail of loop
|
|
281 |
inline Node *tail(); // Handle lazy update of _tail field
|
|
282 |
PhaseIdealLoop* _phase;
|
|
283 |
|
|
284 |
Node_List _body; // Loop body for inner loops
|
|
285 |
|
|
286 |
uint8 _nest; // Nesting depth
|
|
287 |
uint8 _irreducible:1, // True if irreducible
|
|
288 |
_has_call:1, // True if has call safepoint
|
|
289 |
_has_sfpt:1, // True if has non-call safepoint
|
|
290 |
_rce_candidate:1; // True if candidate for range check elimination
|
|
291 |
|
|
292 |
Node_List* _required_safept; // A inner loop cannot delete these safepts;
|
|
293 |
|
|
294 |
IdealLoopTree( PhaseIdealLoop* phase, Node *head, Node *tail )
|
|
295 |
: _parent(0), _next(0), _child(0),
|
|
296 |
_head(head), _tail(tail),
|
|
297 |
_phase(phase),
|
|
298 |
_required_safept(NULL),
|
|
299 |
_nest(0), _irreducible(0), _has_call(0), _has_sfpt(0), _rce_candidate(0)
|
|
300 |
{ }
|
|
301 |
|
|
302 |
// Is 'l' a member of 'this'?
|
|
303 |
int is_member( const IdealLoopTree *l ) const; // Test for nested membership
|
|
304 |
|
|
305 |
// Set loop nesting depth. Accumulate has_call bits.
|
|
306 |
int set_nest( uint depth );
|
|
307 |
|
|
308 |
// Split out multiple fall-in edges from the loop header. Move them to a
|
|
309 |
// private RegionNode before the loop. This becomes the loop landing pad.
|
|
310 |
void split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt );
|
|
311 |
|
|
312 |
// Split out the outermost loop from this shared header.
|
|
313 |
void split_outer_loop( PhaseIdealLoop *phase );
|
|
314 |
|
|
315 |
// Merge all the backedges from the shared header into a private Region.
|
|
316 |
// Feed that region as the one backedge to this loop.
|
|
317 |
void merge_many_backedges( PhaseIdealLoop *phase );
|
|
318 |
|
|
319 |
// Split shared headers and insert loop landing pads.
|
|
320 |
// Insert a LoopNode to replace the RegionNode.
|
|
321 |
// Returns TRUE if loop tree is structurally changed.
|
|
322 |
bool beautify_loops( PhaseIdealLoop *phase );
|
|
323 |
|
|
324 |
// Perform iteration-splitting on inner loops. Split iterations to avoid
|
|
325 |
// range checks or one-shot null checks.
|
|
326 |
void iteration_split( PhaseIdealLoop *phase, Node_List &old_new );
|
|
327 |
|
|
328 |
// Driver for various flavors of iteration splitting
|
|
329 |
void iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new );
|
|
330 |
|
|
331 |
// Given dominators, try to find loops with calls that must always be
|
|
332 |
// executed (call dominates loop tail). These loops do not need non-call
|
|
333 |
// safepoints (ncsfpt).
|
|
334 |
void check_safepts(VectorSet &visited, Node_List &stack);
|
|
335 |
|
|
336 |
// Allpaths backwards scan from loop tail, terminating each path at first safepoint
|
|
337 |
// encountered.
|
|
338 |
void allpaths_check_safepts(VectorSet &visited, Node_List &stack);
|
|
339 |
|
|
340 |
// Convert to counted loops where possible
|
|
341 |
void counted_loop( PhaseIdealLoop *phase );
|
|
342 |
|
|
343 |
// Check for Node being a loop-breaking test
|
|
344 |
Node *is_loop_exit(Node *iff) const;
|
|
345 |
|
|
346 |
// Returns true if ctrl is executed on every complete iteration
|
|
347 |
bool dominates_backedge(Node* ctrl);
|
|
348 |
|
|
349 |
// Remove simplistic dead code from loop body
|
|
350 |
void DCE_loop_body();
|
|
351 |
|
|
352 |
// Look for loop-exit tests with my 50/50 guesses from the Parsing stage.
|
|
353 |
// Replace with a 1-in-10 exit guess.
|
|
354 |
void adjust_loop_exit_prob( PhaseIdealLoop *phase );
|
|
355 |
|
|
356 |
// Return TRUE or FALSE if the loop should never be RCE'd or aligned.
|
|
357 |
// Useful for unrolling loops with NO array accesses.
|
|
358 |
bool policy_peel_only( PhaseIdealLoop *phase ) const;
|
|
359 |
|
|
360 |
// Return TRUE or FALSE if the loop should be unswitched -- clone
|
|
361 |
// loop with an invariant test
|
|
362 |
bool policy_unswitching( PhaseIdealLoop *phase ) const;
|
|
363 |
|
|
364 |
// Micro-benchmark spamming. Remove empty loops.
|
|
365 |
bool policy_do_remove_empty_loop( PhaseIdealLoop *phase );
|
|
366 |
|
|
367 |
// Return TRUE or FALSE if the loop should be peeled or not. Peel if we can
|
|
368 |
// make some loop-invariant test (usually a null-check) happen before the
|
|
369 |
// loop.
|
|
370 |
bool policy_peeling( PhaseIdealLoop *phase ) const;
|
|
371 |
|
|
372 |
// Return TRUE or FALSE if the loop should be maximally unrolled. Stash any
|
|
373 |
// known trip count in the counted loop node.
|
|
374 |
bool policy_maximally_unroll( PhaseIdealLoop *phase ) const;
|
|
375 |
|
|
376 |
// Return TRUE or FALSE if the loop should be unrolled or not. Unroll if
|
|
377 |
// the loop is a CountedLoop and the body is small enough.
|
|
378 |
bool policy_unroll( PhaseIdealLoop *phase ) const;
|
|
379 |
|
|
380 |
// Return TRUE or FALSE if the loop should be range-check-eliminated.
|
|
381 |
// Gather a list of IF tests that are dominated by iteration splitting;
|
|
382 |
// also gather the end of the first split and the start of the 2nd split.
|
|
383 |
bool policy_range_check( PhaseIdealLoop *phase ) const;
|
|
384 |
|
|
385 |
// Return TRUE or FALSE if the loop should be cache-line aligned.
|
|
386 |
// Gather the expression that does the alignment. Note that only
|
|
387 |
// one array base can be aligned in a loop (unless the VM guarentees
|
|
388 |
// mutual alignment). Note that if we vectorize short memory ops
|
|
389 |
// into longer memory ops, we may want to increase alignment.
|
|
390 |
bool policy_align( PhaseIdealLoop *phase ) const;
|
|
391 |
|
|
392 |
// Compute loop trip count from profile data
|
|
393 |
void compute_profile_trip_cnt( PhaseIdealLoop *phase );
|
|
394 |
|
|
395 |
// Reassociate invariant expressions.
|
|
396 |
void reassociate_invariants(PhaseIdealLoop *phase);
|
|
397 |
// Reassociate invariant add and subtract expressions.
|
|
398 |
Node* reassociate_add_sub(Node* n1, PhaseIdealLoop *phase);
|
|
399 |
// Return nonzero index of invariant operand if invariant and variant
|
|
400 |
// are combined with an Add or Sub. Helper for reassoicate_invariants.
|
|
401 |
int is_invariant_addition(Node* n, PhaseIdealLoop *phase);
|
|
402 |
|
|
403 |
// Return true if n is invariant
|
|
404 |
bool is_invariant(Node* n) const;
|
|
405 |
|
|
406 |
// Put loop body on igvn work list
|
|
407 |
void record_for_igvn();
|
|
408 |
|
|
409 |
bool is_loop() { return !_irreducible && _tail && !_tail->is_top(); }
|
|
410 |
bool is_inner() { return is_loop() && _child == NULL; }
|
|
411 |
bool is_counted() { return is_loop() && _head != NULL && _head->is_CountedLoop(); }
|
|
412 |
|
|
413 |
#ifndef PRODUCT
|
|
414 |
void dump_head( ) const; // Dump loop head only
|
|
415 |
void dump() const; // Dump this loop recursively
|
|
416 |
void verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const;
|
|
417 |
#endif
|
|
418 |
|
|
419 |
};
|
|
420 |
|
|
421 |
// -----------------------------PhaseIdealLoop---------------------------------
|
|
422 |
// Computes the mapping from Nodes to IdealLoopTrees. Organizes IdealLoopTrees into a
|
|
423 |
// loop tree. Drives the loop-based transformations on the ideal graph.
|
|
424 |
class PhaseIdealLoop : public PhaseTransform {
|
|
425 |
friend class IdealLoopTree;
|
|
426 |
friend class SuperWord;
|
|
427 |
// Pre-computed def-use info
|
|
428 |
PhaseIterGVN &_igvn;
|
|
429 |
|
|
430 |
// Head of loop tree
|
|
431 |
IdealLoopTree *_ltree_root;
|
|
432 |
|
|
433 |
// Array of pre-order numbers, plus post-visited bit.
|
|
434 |
// ZERO for not pre-visited. EVEN for pre-visited but not post-visited.
|
|
435 |
// ODD for post-visited. Other bits are the pre-order number.
|
|
436 |
uint *_preorders;
|
|
437 |
uint _max_preorder;
|
|
438 |
|
|
439 |
// Allocate _preorders[] array
|
|
440 |
void allocate_preorders() {
|
|
441 |
_max_preorder = C->unique()+8;
|
|
442 |
_preorders = NEW_RESOURCE_ARRAY(uint, _max_preorder);
|
|
443 |
memset(_preorders, 0, sizeof(uint) * _max_preorder);
|
|
444 |
}
|
|
445 |
|
|
446 |
// Allocate _preorders[] array
|
|
447 |
void reallocate_preorders() {
|
|
448 |
if ( _max_preorder < C->unique() ) {
|
|
449 |
_preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, C->unique());
|
|
450 |
_max_preorder = C->unique();
|
|
451 |
}
|
|
452 |
memset(_preorders, 0, sizeof(uint) * _max_preorder);
|
|
453 |
}
|
|
454 |
|
|
455 |
// Check to grow _preorders[] array for the case when build_loop_tree_impl()
|
|
456 |
// adds new nodes.
|
|
457 |
void check_grow_preorders( ) {
|
|
458 |
if ( _max_preorder < C->unique() ) {
|
|
459 |
uint newsize = _max_preorder<<1; // double size of array
|
|
460 |
_preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, newsize);
|
|
461 |
memset(&_preorders[_max_preorder],0,sizeof(uint)*(newsize-_max_preorder));
|
|
462 |
_max_preorder = newsize;
|
|
463 |
}
|
|
464 |
}
|
|
465 |
// Check for pre-visited. Zero for NOT visited; non-zero for visited.
|
|
466 |
int is_visited( Node *n ) const { return _preorders[n->_idx]; }
|
|
467 |
// Pre-order numbers are written to the Nodes array as low-bit-set values.
|
|
468 |
void set_preorder_visited( Node *n, int pre_order ) {
|
|
469 |
assert( !is_visited( n ), "already set" );
|
|
470 |
_preorders[n->_idx] = (pre_order<<1);
|
|
471 |
};
|
|
472 |
// Return pre-order number.
|
|
473 |
int get_preorder( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]>>1; }
|
|
474 |
|
|
475 |
// Check for being post-visited.
|
|
476 |
// Should be previsited already (checked with assert(is_visited(n))).
|
|
477 |
int is_postvisited( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]&1; }
|
|
478 |
|
|
479 |
// Mark as post visited
|
|
480 |
void set_postvisited( Node *n ) { assert( !is_postvisited( n ), "" ); _preorders[n->_idx] |= 1; }
|
|
481 |
|
|
482 |
// Set/get control node out. Set lower bit to distinguish from IdealLoopTree
|
|
483 |
// Returns true if "n" is a data node, false if it's a control node.
|
|
484 |
bool has_ctrl( Node *n ) const { return ((intptr_t)_nodes[n->_idx]) & 1; }
|
|
485 |
|
|
486 |
// clear out dead code after build_loop_late
|
|
487 |
Node_List _deadlist;
|
|
488 |
|
|
489 |
// Support for faster execution of get_late_ctrl()/dom_lca()
|
|
490 |
// when a node has many uses and dominator depth is deep.
|
|
491 |
Node_Array _dom_lca_tags;
|
|
492 |
void init_dom_lca_tags();
|
|
493 |
void clear_dom_lca_tags();
|
|
494 |
// Inline wrapper for frequent cases:
|
|
495 |
// 1) only one use
|
|
496 |
// 2) a use is the same as the current LCA passed as 'n1'
|
|
497 |
Node *dom_lca_for_get_late_ctrl( Node *lca, Node *n, Node *tag ) {
|
|
498 |
assert( n->is_CFG(), "" );
|
|
499 |
// Fast-path NULL lca
|
|
500 |
if( lca != NULL && lca != n ) {
|
|
501 |
assert( lca->is_CFG(), "" );
|
|
502 |
// find LCA of all uses
|
|
503 |
n = dom_lca_for_get_late_ctrl_internal( lca, n, tag );
|
|
504 |
}
|
|
505 |
return find_non_split_ctrl(n);
|
|
506 |
}
|
|
507 |
Node *dom_lca_for_get_late_ctrl_internal( Node *lca, Node *n, Node *tag );
|
|
508 |
// true if CFG node d dominates CFG node n
|
|
509 |
bool is_dominator(Node *d, Node *n);
|
|
510 |
|
|
511 |
// Helper function for directing control inputs away from CFG split
|
|
512 |
// points.
|
|
513 |
Node *find_non_split_ctrl( Node *ctrl ) const {
|
|
514 |
if (ctrl != NULL) {
|
|
515 |
if (ctrl->is_MultiBranch()) {
|
|
516 |
ctrl = ctrl->in(0);
|
|
517 |
}
|
|
518 |
assert(ctrl->is_CFG(), "CFG");
|
|
519 |
}
|
|
520 |
return ctrl;
|
|
521 |
}
|
|
522 |
|
|
523 |
public:
|
|
524 |
bool has_node( Node* n ) const { return _nodes[n->_idx] != NULL; }
|
|
525 |
// check if transform created new nodes that need _ctrl recorded
|
|
526 |
Node *get_late_ctrl( Node *n, Node *early );
|
|
527 |
Node *get_early_ctrl( Node *n );
|
|
528 |
void set_early_ctrl( Node *n );
|
|
529 |
void set_subtree_ctrl( Node *root );
|
|
530 |
void set_ctrl( Node *n, Node *ctrl ) {
|
|
531 |
assert( !has_node(n) || has_ctrl(n), "" );
|
|
532 |
assert( ctrl->in(0), "cannot set dead control node" );
|
|
533 |
assert( ctrl == find_non_split_ctrl(ctrl), "must set legal crtl" );
|
|
534 |
_nodes.map( n->_idx, (Node*)((intptr_t)ctrl + 1) );
|
|
535 |
}
|
|
536 |
// Set control and update loop membership
|
|
537 |
void set_ctrl_and_loop(Node* n, Node* ctrl) {
|
|
538 |
IdealLoopTree* old_loop = get_loop(get_ctrl(n));
|
|
539 |
IdealLoopTree* new_loop = get_loop(ctrl);
|
|
540 |
if (old_loop != new_loop) {
|
|
541 |
if (old_loop->_child == NULL) old_loop->_body.yank(n);
|
|
542 |
if (new_loop->_child == NULL) new_loop->_body.push(n);
|
|
543 |
}
|
|
544 |
set_ctrl(n, ctrl);
|
|
545 |
}
|
|
546 |
// Control nodes can be replaced or subsumed. During this pass they
|
|
547 |
// get their replacement Node in slot 1. Instead of updating the block
|
|
548 |
// location of all Nodes in the subsumed block, we lazily do it. As we
|
|
549 |
// pull such a subsumed block out of the array, we write back the final
|
|
550 |
// correct block.
|
|
551 |
Node *get_ctrl( Node *i ) {
|
|
552 |
assert(has_node(i), "");
|
|
553 |
Node *n = get_ctrl_no_update(i);
|
|
554 |
_nodes.map( i->_idx, (Node*)((intptr_t)n + 1) );
|
|
555 |
assert(has_node(i) && has_ctrl(i), "");
|
|
556 |
assert(n == find_non_split_ctrl(n), "must return legal ctrl" );
|
|
557 |
return n;
|
|
558 |
}
|
|
559 |
|
|
560 |
private:
|
|
561 |
Node *get_ctrl_no_update( Node *i ) const {
|
|
562 |
assert( has_ctrl(i), "" );
|
|
563 |
Node *n = (Node*)(((intptr_t)_nodes[i->_idx]) & ~1);
|
|
564 |
if (!n->in(0)) {
|
|
565 |
// Skip dead CFG nodes
|
|
566 |
do {
|
|
567 |
n = (Node*)(((intptr_t)_nodes[n->_idx]) & ~1);
|
|
568 |
} while (!n->in(0));
|
|
569 |
n = find_non_split_ctrl(n);
|
|
570 |
}
|
|
571 |
return n;
|
|
572 |
}
|
|
573 |
|
|
574 |
// Check for loop being set
|
|
575 |
// "n" must be a control node. Returns true if "n" is known to be in a loop.
|
|
576 |
bool has_loop( Node *n ) const {
|
|
577 |
assert(!has_node(n) || !has_ctrl(n), "");
|
|
578 |
return has_node(n);
|
|
579 |
}
|
|
580 |
// Set loop
|
|
581 |
void set_loop( Node *n, IdealLoopTree *loop ) {
|
|
582 |
_nodes.map(n->_idx, (Node*)loop);
|
|
583 |
}
|
|
584 |
// Lazy-dazy update of 'get_ctrl' and 'idom_at' mechanisms. Replace
|
|
585 |
// the 'old_node' with 'new_node'. Kill old-node. Add a reference
|
|
586 |
// from old_node to new_node to support the lazy update. Reference
|
|
587 |
// replaces loop reference, since that is not neede for dead node.
|
|
588 |
public:
|
|
589 |
void lazy_update( Node *old_node, Node *new_node ) {
|
|
590 |
assert( old_node != new_node, "no cycles please" );
|
|
591 |
//old_node->set_req( 1, new_node /*NO DU INFO*/ );
|
|
592 |
// Nodes always have DU info now, so re-use the side array slot
|
|
593 |
// for this node to provide the forwarding pointer.
|
|
594 |
_nodes.map( old_node->_idx, (Node*)((intptr_t)new_node + 1) );
|
|
595 |
}
|
|
596 |
void lazy_replace( Node *old_node, Node *new_node ) {
|
|
597 |
_igvn.hash_delete(old_node);
|
|
598 |
_igvn.subsume_node( old_node, new_node );
|
|
599 |
lazy_update( old_node, new_node );
|
|
600 |
}
|
|
601 |
void lazy_replace_proj( Node *old_node, Node *new_node ) {
|
|
602 |
assert( old_node->req() == 1, "use this for Projs" );
|
|
603 |
_igvn.hash_delete(old_node); // Must hash-delete before hacking edges
|
|
604 |
old_node->add_req( NULL );
|
|
605 |
lazy_replace( old_node, new_node );
|
|
606 |
}
|
|
607 |
|
|
608 |
private:
|
|
609 |
|
|
610 |
// Place 'n' in some loop nest, where 'n' is a CFG node
|
|
611 |
void build_loop_tree();
|
|
612 |
int build_loop_tree_impl( Node *n, int pre_order );
|
|
613 |
// Insert loop into the existing loop tree. 'innermost' is a leaf of the
|
|
614 |
// loop tree, not the root.
|
|
615 |
IdealLoopTree *sort( IdealLoopTree *loop, IdealLoopTree *innermost );
|
|
616 |
|
|
617 |
// Place Data nodes in some loop nest
|
|
618 |
void build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack, const PhaseIdealLoop *verify_me );
|
|
619 |
void build_loop_late ( VectorSet &visited, Node_List &worklist, Node_Stack &nstack, const PhaseIdealLoop *verify_me );
|
|
620 |
void build_loop_late_post ( Node* n, const PhaseIdealLoop *verify_me );
|
|
621 |
|
|
622 |
// Array of immediate dominance info for each CFG node indexed by node idx
|
|
623 |
private:
|
|
624 |
uint _idom_size;
|
|
625 |
Node **_idom; // Array of immediate dominators
|
|
626 |
uint *_dom_depth; // Used for fast LCA test
|
|
627 |
GrowableArray<uint>* _dom_stk; // For recomputation of dom depth
|
|
628 |
|
|
629 |
Node* idom_no_update(Node* d) const {
|
|
630 |
assert(d->_idx < _idom_size, "oob");
|
|
631 |
Node* n = _idom[d->_idx];
|
|
632 |
assert(n != NULL,"Bad immediate dominator info.");
|
|
633 |
while (n->in(0) == NULL) { // Skip dead CFG nodes
|
|
634 |
//n = n->in(1);
|
|
635 |
n = (Node*)(((intptr_t)_nodes[n->_idx]) & ~1);
|
|
636 |
assert(n != NULL,"Bad immediate dominator info.");
|
|
637 |
}
|
|
638 |
return n;
|
|
639 |
}
|
|
640 |
Node *idom(Node* d) const {
|
|
641 |
uint didx = d->_idx;
|
|
642 |
Node *n = idom_no_update(d);
|
|
643 |
_idom[didx] = n; // Lazily remove dead CFG nodes from table.
|
|
644 |
return n;
|
|
645 |
}
|
|
646 |
uint dom_depth(Node* d) const {
|
|
647 |
assert(d->_idx < _idom_size, "");
|
|
648 |
return _dom_depth[d->_idx];
|
|
649 |
}
|
|
650 |
void set_idom(Node* d, Node* n, uint dom_depth);
|
|
651 |
// Locally compute IDOM using dom_lca call
|
|
652 |
Node *compute_idom( Node *region ) const;
|
|
653 |
// Recompute dom_depth
|
|
654 |
void recompute_dom_depth();
|
|
655 |
|
|
656 |
// Is safept not required by an outer loop?
|
|
657 |
bool is_deleteable_safept(Node* sfpt);
|
|
658 |
|
|
659 |
public:
|
|
660 |
// Dominators for the sea of nodes
|
|
661 |
void Dominators();
|
|
662 |
Node *dom_lca( Node *n1, Node *n2 ) const {
|
|
663 |
return find_non_split_ctrl(dom_lca_internal(n1, n2));
|
|
664 |
}
|
|
665 |
Node *dom_lca_internal( Node *n1, Node *n2 ) const;
|
|
666 |
|
|
667 |
// Compute the Ideal Node to Loop mapping
|
|
668 |
PhaseIdealLoop( PhaseIterGVN &igvn, const PhaseIdealLoop *verify_me, bool do_split_ifs );
|
|
669 |
|
|
670 |
// True if the method has at least 1 irreducible loop
|
|
671 |
bool _has_irreducible_loops;
|
|
672 |
|
|
673 |
// Per-Node transform
|
|
674 |
virtual Node *transform( Node *a_node ) { return 0; }
|
|
675 |
|
|
676 |
Node *is_counted_loop( Node *x, IdealLoopTree *loop );
|
|
677 |
|
|
678 |
// Return a post-walked LoopNode
|
|
679 |
IdealLoopTree *get_loop( Node *n ) const {
|
|
680 |
// Dead nodes have no loop, so return the top level loop instead
|
|
681 |
if (!has_node(n)) return _ltree_root;
|
|
682 |
assert(!has_ctrl(n), "");
|
|
683 |
return (IdealLoopTree*)_nodes[n->_idx];
|
|
684 |
}
|
|
685 |
|
|
686 |
// Is 'n' a (nested) member of 'loop'?
|
|
687 |
int is_member( const IdealLoopTree *loop, Node *n ) const {
|
|
688 |
return loop->is_member(get_loop(n)); }
|
|
689 |
|
|
690 |
// This is the basic building block of the loop optimizations. It clones an
|
|
691 |
// entire loop body. It makes an old_new loop body mapping; with this
|
|
692 |
// mapping you can find the new-loop equivalent to an old-loop node. All
|
|
693 |
// new-loop nodes are exactly equal to their old-loop counterparts, all
|
|
694 |
// edges are the same. All exits from the old-loop now have a RegionNode
|
|
695 |
// that merges the equivalent new-loop path. This is true even for the
|
|
696 |
// normal "loop-exit" condition. All uses of loop-invariant old-loop values
|
|
697 |
// now come from (one or more) Phis that merge their new-loop equivalents.
|
|
698 |
// Parameter side_by_side_idom:
|
|
699 |
// When side_by_size_idom is NULL, the dominator tree is constructed for
|
|
700 |
// the clone loop to dominate the original. Used in construction of
|
|
701 |
// pre-main-post loop sequence.
|
|
702 |
// When nonnull, the clone and original are side-by-side, both are
|
|
703 |
// dominated by the passed in side_by_side_idom node. Used in
|
|
704 |
// construction of unswitched loops.
|
|
705 |
void clone_loop( IdealLoopTree *loop, Node_List &old_new, int dom_depth,
|
|
706 |
Node* side_by_side_idom = NULL);
|
|
707 |
|
|
708 |
// If we got the effect of peeling, either by actually peeling or by
|
|
709 |
// making a pre-loop which must execute at least once, we can remove
|
|
710 |
// all loop-invariant dominated tests in the main body.
|
|
711 |
void peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new );
|
|
712 |
|
|
713 |
// Generate code to do a loop peel for the given loop (and body).
|
|
714 |
// old_new is a temp array.
|
|
715 |
void do_peeling( IdealLoopTree *loop, Node_List &old_new );
|
|
716 |
|
|
717 |
// Add pre and post loops around the given loop. These loops are used
|
|
718 |
// during RCE, unrolling and aligning loops.
|
|
719 |
void insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only );
|
|
720 |
// If Node n lives in the back_ctrl block, we clone a private version of n
|
|
721 |
// in preheader_ctrl block and return that, otherwise return n.
|
|
722 |
Node *clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n );
|
|
723 |
|
|
724 |
// Take steps to maximally unroll the loop. Peel any odd iterations, then
|
|
725 |
// unroll to do double iterations. The next round of major loop transforms
|
|
726 |
// will repeat till the doubled loop body does all remaining iterations in 1
|
|
727 |
// pass.
|
|
728 |
void do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new );
|
|
729 |
|
|
730 |
// Unroll the loop body one step - make each trip do 2 iterations.
|
|
731 |
void do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip );
|
|
732 |
|
|
733 |
// Return true if exp is a constant times an induction var
|
|
734 |
bool is_scaled_iv(Node* exp, Node* iv, int* p_scale);
|
|
735 |
|
|
736 |
// Return true if exp is a scaled induction var plus (or minus) constant
|
|
737 |
bool is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth = 0);
|
|
738 |
|
|
739 |
// Eliminate range-checks and other trip-counter vs loop-invariant tests.
|
|
740 |
void do_range_check( IdealLoopTree *loop, Node_List &old_new );
|
|
741 |
|
|
742 |
// Create a slow version of the loop by cloning the loop
|
|
743 |
// and inserting an if to select fast-slow versions.
|
|
744 |
ProjNode* create_slow_version_of_loop(IdealLoopTree *loop,
|
|
745 |
Node_List &old_new);
|
|
746 |
|
|
747 |
// Clone loop with an invariant test (that does not exit) and
|
|
748 |
// insert a clone of the test that selects which version to
|
|
749 |
// execute.
|
|
750 |
void do_unswitching (IdealLoopTree *loop, Node_List &old_new);
|
|
751 |
|
|
752 |
// Find candidate "if" for unswitching
|
|
753 |
IfNode* find_unswitching_candidate(const IdealLoopTree *loop) const;
|
|
754 |
|
|
755 |
// Range Check Elimination uses this function!
|
|
756 |
// Constrain the main loop iterations so the affine function:
|
|
757 |
// scale_con * I + offset < limit
|
|
758 |
// always holds true. That is, either increase the number of iterations in
|
|
759 |
// the pre-loop or the post-loop until the condition holds true in the main
|
|
760 |
// loop. Scale_con, offset and limit are all loop invariant.
|
|
761 |
void add_constraint( int stride_con, int scale_con, Node *offset, Node *limit, Node *pre_ctrl, Node **pre_limit, Node **main_limit );
|
|
762 |
|
|
763 |
// Partially peel loop up through last_peel node.
|
|
764 |
bool partial_peel( IdealLoopTree *loop, Node_List &old_new );
|
|
765 |
|
|
766 |
// Create a scheduled list of nodes control dependent on ctrl set.
|
|
767 |
void scheduled_nodelist( IdealLoopTree *loop, VectorSet& ctrl, Node_List &sched );
|
|
768 |
// Has a use in the vector set
|
|
769 |
bool has_use_in_set( Node* n, VectorSet& vset );
|
|
770 |
// Has use internal to the vector set (ie. not in a phi at the loop head)
|
|
771 |
bool has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop );
|
|
772 |
// clone "n" for uses that are outside of loop
|
|
773 |
void clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist );
|
|
774 |
// clone "n" for special uses that are in the not_peeled region
|
|
775 |
void clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
|
|
776 |
VectorSet& not_peel, Node_List& sink_list, Node_List& worklist );
|
|
777 |
// Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
|
|
778 |
void insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp );
|
|
779 |
#ifdef ASSERT
|
|
780 |
// Validate the loop partition sets: peel and not_peel
|
|
781 |
bool is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, VectorSet& not_peel );
|
|
782 |
// Ensure that uses outside of loop are of the right form
|
|
783 |
bool is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
|
|
784 |
uint orig_exit_idx, uint clone_exit_idx);
|
|
785 |
bool is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx);
|
|
786 |
#endif
|
|
787 |
|
|
788 |
// Returns nonzero constant stride if-node is a possible iv test (otherwise returns zero.)
|
|
789 |
int stride_of_possible_iv( Node* iff );
|
|
790 |
bool is_possible_iv_test( Node* iff ) { return stride_of_possible_iv(iff) != 0; }
|
|
791 |
// Return the (unique) control output node that's in the loop (if it exists.)
|
|
792 |
Node* stay_in_loop( Node* n, IdealLoopTree *loop);
|
|
793 |
// Insert a signed compare loop exit cloned from an unsigned compare.
|
|
794 |
IfNode* insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop);
|
|
795 |
void remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop);
|
|
796 |
// Utility to register node "n" with PhaseIdealLoop
|
|
797 |
void register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth);
|
|
798 |
// Utility to create an if-projection
|
|
799 |
ProjNode* proj_clone(ProjNode* p, IfNode* iff);
|
|
800 |
// Force the iff control output to be the live_proj
|
|
801 |
Node* short_circuit_if(IfNode* iff, ProjNode* live_proj);
|
|
802 |
// Insert a region before an if projection
|
|
803 |
RegionNode* insert_region_before_proj(ProjNode* proj);
|
|
804 |
// Insert a new if before an if projection
|
|
805 |
ProjNode* insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj);
|
|
806 |
|
|
807 |
// Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
|
|
808 |
// "Nearly" because all Nodes have been cloned from the original in the loop,
|
|
809 |
// but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
|
|
810 |
// through the Phi recursively, and return a Bool.
|
|
811 |
BoolNode *clone_iff( PhiNode *phi, IdealLoopTree *loop );
|
|
812 |
CmpNode *clone_bool( PhiNode *phi, IdealLoopTree *loop );
|
|
813 |
|
|
814 |
|
|
815 |
// Rework addressing expressions to get the most loop-invariant stuff
|
|
816 |
// moved out. We'd like to do all associative operators, but it's especially
|
|
817 |
// important (common) to do address expressions.
|
|
818 |
Node *remix_address_expressions( Node *n );
|
|
819 |
|
|
820 |
// Attempt to use a conditional move instead of a phi/branch
|
|
821 |
Node *conditional_move( Node *n );
|
|
822 |
|
|
823 |
// Reorganize offset computations to lower register pressure.
|
|
824 |
// Mostly prevent loop-fallout uses of the pre-incremented trip counter
|
|
825 |
// (which are then alive with the post-incremented trip counter
|
|
826 |
// forcing an extra register move)
|
|
827 |
void reorg_offsets( IdealLoopTree *loop );
|
|
828 |
|
|
829 |
// Check for aggressive application of 'split-if' optimization,
|
|
830 |
// using basic block level info.
|
|
831 |
void split_if_with_blocks ( VectorSet &visited, Node_Stack &nstack );
|
|
832 |
Node *split_if_with_blocks_pre ( Node *n );
|
|
833 |
void split_if_with_blocks_post( Node *n );
|
|
834 |
Node *has_local_phi_input( Node *n );
|
|
835 |
// Mark an IfNode as being dominated by a prior test,
|
|
836 |
// without actually altering the CFG (and hence IDOM info).
|
|
837 |
void dominated_by( Node *prevdom, Node *iff );
|
|
838 |
|
|
839 |
// Split Node 'n' through merge point
|
|
840 |
Node *split_thru_region( Node *n, Node *region );
|
|
841 |
// Split Node 'n' through merge point if there is enough win.
|
|
842 |
Node *split_thru_phi( Node *n, Node *region, int policy );
|
|
843 |
// Found an If getting its condition-code input from a Phi in the
|
|
844 |
// same block. Split thru the Region.
|
|
845 |
void do_split_if( Node *iff );
|
|
846 |
|
|
847 |
private:
|
|
848 |
// Return a type based on condition control flow
|
|
849 |
const TypeInt* filtered_type( Node *n, Node* n_ctrl);
|
|
850 |
const TypeInt* filtered_type( Node *n ) { return filtered_type(n, NULL); }
|
|
851 |
// Helpers for filtered type
|
|
852 |
const TypeInt* filtered_type_from_dominators( Node* val, Node *val_ctrl);
|
|
853 |
|
|
854 |
// Helper functions
|
|
855 |
void register_new_node( Node *n, Node *blk );
|
|
856 |
Node *spinup( Node *iff, Node *new_false, Node *new_true, Node *region, Node *phi, small_cache *cache );
|
|
857 |
Node *find_use_block( Node *use, Node *def, Node *old_false, Node *new_false, Node *old_true, Node *new_true );
|
|
858 |
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 );
|
|
859 |
bool split_up( Node *n, Node *blk1, Node *blk2 );
|
|
860 |
void sink_use( Node *use, Node *post_loop );
|
|
861 |
Node *place_near_use( Node *useblock ) const;
|
|
862 |
|
|
863 |
bool _created_loop_node;
|
|
864 |
public:
|
|
865 |
void set_created_loop_node() { _created_loop_node = true; }
|
|
866 |
bool created_loop_node() { return _created_loop_node; }
|
|
867 |
|
|
868 |
#ifndef PRODUCT
|
|
869 |
void dump( ) const;
|
|
870 |
void dump( IdealLoopTree *loop, uint rpo_idx, Node_List &rpo_list ) const;
|
|
871 |
void rpo( Node *start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list ) const;
|
|
872 |
void verify() const; // Major slow :-)
|
|
873 |
void verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const;
|
|
874 |
IdealLoopTree *get_loop_idx(Node* n) const {
|
|
875 |
// Dead nodes have no loop, so return the top level loop instead
|
|
876 |
return _nodes[n->_idx] ? (IdealLoopTree*)_nodes[n->_idx] : _ltree_root;
|
|
877 |
}
|
|
878 |
// Print some stats
|
|
879 |
static void print_statistics();
|
|
880 |
static int _loop_invokes; // Count of PhaseIdealLoop invokes
|
|
881 |
static int _loop_work; // Sum of PhaseIdealLoop x _unique
|
|
882 |
#endif
|
|
883 |
};
|
|
884 |
|
|
885 |
inline Node* IdealLoopTree::tail() {
|
|
886 |
// Handle lazy update of _tail field
|
|
887 |
Node *n = _tail;
|
|
888 |
//while( !n->in(0) ) // Skip dead CFG nodes
|
|
889 |
//n = n->in(1);
|
|
890 |
if (n->in(0) == NULL)
|
|
891 |
n = _phase->get_ctrl(n);
|
|
892 |
_tail = n;
|
|
893 |
return n;
|
|
894 |
}
|
|
895 |
|
|
896 |
|
|
897 |
// Iterate over the loop tree using a preorder, left-to-right traversal.
|
|
898 |
//
|
|
899 |
// Example that visits all counted loops from within PhaseIdealLoop
|
|
900 |
//
|
|
901 |
// for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
|
|
902 |
// IdealLoopTree* lpt = iter.current();
|
|
903 |
// if (!lpt->is_counted()) continue;
|
|
904 |
// ...
|
|
905 |
class LoopTreeIterator : public StackObj {
|
|
906 |
private:
|
|
907 |
IdealLoopTree* _root;
|
|
908 |
IdealLoopTree* _curnt;
|
|
909 |
|
|
910 |
public:
|
|
911 |
LoopTreeIterator(IdealLoopTree* root) : _root(root), _curnt(root) {}
|
|
912 |
|
|
913 |
bool done() { return _curnt == NULL; } // Finished iterating?
|
|
914 |
|
|
915 |
void next(); // Advance to next loop tree
|
|
916 |
|
|
917 |
IdealLoopTree* current() { return _curnt; } // Return current value of iterator.
|
|
918 |
};
|