Tue, 09 Mar 2010 20:16:19 +0100
6919934: JSR 292 needs to support x86 C1
Summary: This implements JSR 292 support for C1 x86.
Reviewed-by: never, jrose, kvn
1 /*
2 * Copyright 1998-2009 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 */
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 class Invariance;
34 struct small_cache;
36 //
37 // I D E A L I Z E D L O O P S
38 //
39 // Idealized loops are the set of loops I perform more interesting
40 // transformations on, beyond simple hoisting.
42 //------------------------------LoopNode---------------------------------------
43 // Simple loop header. Fall in path on left, loop-back path on right.
44 class LoopNode : public RegionNode {
45 // Size is bigger to hold the flags. However, the flags do not change
46 // the semantics so it does not appear in the hash & cmp functions.
47 virtual uint size_of() const { return sizeof(*this); }
48 protected:
49 short _loop_flags;
50 // Names for flag bitfields
51 enum { pre_post_main=0, inner_loop=8, partial_peel_loop=16, partial_peel_failed=32 };
52 char _unswitch_count;
53 enum { _unswitch_max=3 };
55 public:
56 // Names for edge indices
57 enum { Self=0, EntryControl, LoopBackControl };
59 int is_inner_loop() const { return _loop_flags & inner_loop; }
60 void set_inner_loop() { _loop_flags |= inner_loop; }
62 int is_partial_peel_loop() const { return _loop_flags & partial_peel_loop; }
63 void set_partial_peel_loop() { _loop_flags |= partial_peel_loop; }
64 int partial_peel_has_failed() const { return _loop_flags & partial_peel_failed; }
65 void mark_partial_peel_failed() { _loop_flags |= partial_peel_failed; }
67 int unswitch_max() { return _unswitch_max; }
68 int unswitch_count() { return _unswitch_count; }
69 void set_unswitch_count(int val) {
70 assert (val <= unswitch_max(), "too many unswitches");
71 _unswitch_count = val;
72 }
74 LoopNode( Node *entry, Node *backedge ) : RegionNode(3), _loop_flags(0), _unswitch_count(0) {
75 init_class_id(Class_Loop);
76 init_req(EntryControl, entry);
77 init_req(LoopBackControl, backedge);
78 }
80 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
81 virtual int Opcode() const;
82 bool can_be_counted_loop(PhaseTransform* phase) const {
83 return req() == 3 && in(0) != NULL &&
84 in(1) != NULL && phase->type(in(1)) != Type::TOP &&
85 in(2) != NULL && phase->type(in(2)) != Type::TOP;
86 }
87 #ifndef PRODUCT
88 virtual void dump_spec(outputStream *st) const;
89 #endif
90 };
92 //------------------------------Counted Loops----------------------------------
93 // Counted loops are all trip-counted loops, with exactly 1 trip-counter exit
94 // path (and maybe some other exit paths). The trip-counter exit is always
95 // last in the loop. The trip-counter does not have to stride by a constant,
96 // but it does have to stride by a loop-invariant amount; the exit value is
97 // also loop invariant.
99 // CountedLoopNodes and CountedLoopEndNodes come in matched pairs. The
100 // CountedLoopNode has the incoming loop control and the loop-back-control
101 // which is always the IfTrue before the matching CountedLoopEndNode. The
102 // CountedLoopEndNode has an incoming control (possibly not the
103 // CountedLoopNode if there is control flow in the loop), the post-increment
104 // trip-counter value, and the limit. The trip-counter value is always of
105 // the form (Op old-trip-counter stride). The old-trip-counter is produced
106 // by a Phi connected to the CountedLoopNode. The stride is loop invariant.
107 // The Op is any commutable opcode, including Add, Mul, Xor. The
108 // CountedLoopEndNode also takes in the loop-invariant limit value.
110 // From a CountedLoopNode I can reach the matching CountedLoopEndNode via the
111 // loop-back control. From CountedLoopEndNodes I can reach CountedLoopNodes
112 // via the old-trip-counter from the Op node.
114 //------------------------------CountedLoopNode--------------------------------
115 // CountedLoopNodes head simple counted loops. CountedLoopNodes have as
116 // inputs the incoming loop-start control and the loop-back control, so they
117 // act like RegionNodes. They also take in the initial trip counter, the
118 // loop-invariant stride and the loop-invariant limit value. CountedLoopNodes
119 // produce a loop-body control and the trip counter value. Since
120 // CountedLoopNodes behave like RegionNodes I still have a standard CFG model.
122 class CountedLoopNode : public LoopNode {
123 // Size is bigger to hold _main_idx. However, _main_idx does not change
124 // the semantics so it does not appear in the hash & cmp functions.
125 virtual uint size_of() const { return sizeof(*this); }
127 // For Pre- and Post-loops during debugging ONLY, this holds the index of
128 // the Main CountedLoop. Used to assert that we understand the graph shape.
129 node_idx_t _main_idx;
131 // Known trip count calculated by policy_maximally_unroll
132 int _trip_count;
134 // Expected trip count from profile data
135 float _profile_trip_cnt;
137 // Log2 of original loop bodies in unrolled loop
138 int _unrolled_count_log2;
140 // Node count prior to last unrolling - used to decide if
141 // unroll,optimize,unroll,optimize,... is making progress
142 int _node_count_before_unroll;
144 public:
145 CountedLoopNode( Node *entry, Node *backedge )
146 : LoopNode(entry, backedge), _trip_count(max_jint),
147 _profile_trip_cnt(COUNT_UNKNOWN), _unrolled_count_log2(0),
148 _node_count_before_unroll(0) {
149 init_class_id(Class_CountedLoop);
150 // Initialize _trip_count to the largest possible value.
151 // Will be reset (lower) if the loop's trip count is known.
152 }
154 virtual int Opcode() const;
155 virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
157 Node *init_control() const { return in(EntryControl); }
158 Node *back_control() const { return in(LoopBackControl); }
159 CountedLoopEndNode *loopexit() const;
160 Node *init_trip() const;
161 Node *stride() const;
162 int stride_con() const;
163 bool stride_is_con() const;
164 Node *limit() const;
165 Node *incr() const;
166 Node *phi() const;
168 // Match increment with optional truncation
169 static Node* match_incr_with_optional_truncation(Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type);
171 // A 'main' loop has a pre-loop and a post-loop. The 'main' loop
172 // can run short a few iterations and may start a few iterations in.
173 // It will be RCE'd and unrolled and aligned.
175 // A following 'post' loop will run any remaining iterations. Used
176 // during Range Check Elimination, the 'post' loop will do any final
177 // iterations with full checks. Also used by Loop Unrolling, where
178 // the 'post' loop will do any epilog iterations needed. Basically,
179 // a 'post' loop can not profitably be further unrolled or RCE'd.
181 // A preceding 'pre' loop will run at least 1 iteration (to do peeling),
182 // it may do under-flow checks for RCE and may do alignment iterations
183 // so the following main loop 'knows' that it is striding down cache
184 // lines.
186 // A 'main' loop that is ONLY unrolled or peeled, never RCE'd or
187 // Aligned, may be missing it's pre-loop.
188 enum { Normal=0, Pre=1, Main=2, Post=3, PrePostFlagsMask=3, Main_Has_No_Pre_Loop=4 };
189 int is_normal_loop() const { return (_loop_flags&PrePostFlagsMask) == Normal; }
190 int is_pre_loop () const { return (_loop_flags&PrePostFlagsMask) == Pre; }
191 int is_main_loop () const { return (_loop_flags&PrePostFlagsMask) == Main; }
192 int is_post_loop () const { return (_loop_flags&PrePostFlagsMask) == Post; }
193 int is_main_no_pre_loop() const { return _loop_flags & Main_Has_No_Pre_Loop; }
194 void set_main_no_pre_loop() { _loop_flags |= Main_Has_No_Pre_Loop; }
196 int main_idx() const { return _main_idx; }
199 void set_pre_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Pre ; _main_idx = main->_idx; }
200 void set_main_loop ( ) { assert(is_normal_loop(),""); _loop_flags |= Main; }
201 void set_post_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Post; _main_idx = main->_idx; }
202 void set_normal_loop( ) { _loop_flags &= ~PrePostFlagsMask; }
204 void set_trip_count(int tc) { _trip_count = tc; }
205 int trip_count() { return _trip_count; }
207 void set_profile_trip_cnt(float ptc) { _profile_trip_cnt = ptc; }
208 float profile_trip_cnt() { return _profile_trip_cnt; }
210 void double_unrolled_count() { _unrolled_count_log2++; }
211 int unrolled_count() { return 1 << MIN2(_unrolled_count_log2, BitsPerInt-3); }
213 void set_node_count_before_unroll(int ct) { _node_count_before_unroll = ct; }
214 int node_count_before_unroll() { return _node_count_before_unroll; }
216 #ifndef PRODUCT
217 virtual void dump_spec(outputStream *st) const;
218 #endif
219 };
221 //------------------------------CountedLoopEndNode-----------------------------
222 // CountedLoopEndNodes end simple trip counted loops. They act much like
223 // IfNodes.
224 class CountedLoopEndNode : public IfNode {
225 public:
226 enum { TestControl, TestValue };
228 CountedLoopEndNode( Node *control, Node *test, float prob, float cnt )
229 : IfNode( control, test, prob, cnt) {
230 init_class_id(Class_CountedLoopEnd);
231 }
232 virtual int Opcode() const;
234 Node *cmp_node() const { return (in(TestValue)->req() >=2) ? in(TestValue)->in(1) : NULL; }
235 Node *incr() const { Node *tmp = cmp_node(); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; }
236 Node *limit() const { Node *tmp = cmp_node(); return (tmp && tmp->req()==3) ? tmp->in(2) : NULL; }
237 Node *stride() const { Node *tmp = incr (); return (tmp && tmp->req()==3) ? tmp->in(2) : NULL; }
238 Node *phi() const { Node *tmp = incr (); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; }
239 Node *init_trip() const { Node *tmp = phi (); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; }
240 int stride_con() const;
241 bool stride_is_con() const { Node *tmp = stride (); return (tmp != NULL && tmp->is_Con()); }
242 BoolTest::mask test_trip() const { return in(TestValue)->as_Bool()->_test._test; }
243 CountedLoopNode *loopnode() const {
244 Node *ln = phi()->in(0);
245 assert( ln->Opcode() == Op_CountedLoop, "malformed loop" );
246 return (CountedLoopNode*)ln; }
248 #ifndef PRODUCT
249 virtual void dump_spec(outputStream *st) const;
250 #endif
251 };
254 inline CountedLoopEndNode *CountedLoopNode::loopexit() const {
255 Node *bc = back_control();
256 if( bc == NULL ) return NULL;
257 Node *le = bc->in(0);
258 if( le->Opcode() != Op_CountedLoopEnd )
259 return NULL;
260 return (CountedLoopEndNode*)le;
261 }
262 inline Node *CountedLoopNode::init_trip() const { return loopexit() ? loopexit()->init_trip() : NULL; }
263 inline Node *CountedLoopNode::stride() const { return loopexit() ? loopexit()->stride() : NULL; }
264 inline int CountedLoopNode::stride_con() const { return loopexit() ? loopexit()->stride_con() : 0; }
265 inline bool CountedLoopNode::stride_is_con() const { return loopexit() && loopexit()->stride_is_con(); }
266 inline Node *CountedLoopNode::limit() const { return loopexit() ? loopexit()->limit() : NULL; }
267 inline Node *CountedLoopNode::incr() const { return loopexit() ? loopexit()->incr() : NULL; }
268 inline Node *CountedLoopNode::phi() const { return loopexit() ? loopexit()->phi() : NULL; }
271 // -----------------------------IdealLoopTree----------------------------------
272 class IdealLoopTree : public ResourceObj {
273 public:
274 IdealLoopTree *_parent; // Parent in loop tree
275 IdealLoopTree *_next; // Next sibling in loop tree
276 IdealLoopTree *_child; // First child in loop tree
278 // The head-tail backedge defines the loop.
279 // If tail is NULL then this loop has multiple backedges as part of the
280 // same loop. During cleanup I'll peel off the multiple backedges; merge
281 // them at the loop bottom and flow 1 real backedge into the loop.
282 Node *_head; // Head of loop
283 Node *_tail; // Tail of loop
284 inline Node *tail(); // Handle lazy update of _tail field
285 PhaseIdealLoop* _phase;
287 Node_List _body; // Loop body for inner loops
289 uint8 _nest; // Nesting depth
290 uint8 _irreducible:1, // True if irreducible
291 _has_call:1, // True if has call safepoint
292 _has_sfpt:1, // True if has non-call safepoint
293 _rce_candidate:1; // True if candidate for range check elimination
295 Node_List* _required_safept; // A inner loop cannot delete these safepts;
296 bool _allow_optimizations; // Allow loop optimizations
298 IdealLoopTree( PhaseIdealLoop* phase, Node *head, Node *tail )
299 : _parent(0), _next(0), _child(0),
300 _head(head), _tail(tail),
301 _phase(phase),
302 _required_safept(NULL),
303 _allow_optimizations(true),
304 _nest(0), _irreducible(0), _has_call(0), _has_sfpt(0), _rce_candidate(0)
305 { }
307 // Is 'l' a member of 'this'?
308 int is_member( const IdealLoopTree *l ) const; // Test for nested membership
310 // Set loop nesting depth. Accumulate has_call bits.
311 int set_nest( uint depth );
313 // Split out multiple fall-in edges from the loop header. Move them to a
314 // private RegionNode before the loop. This becomes the loop landing pad.
315 void split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt );
317 // Split out the outermost loop from this shared header.
318 void split_outer_loop( PhaseIdealLoop *phase );
320 // Merge all the backedges from the shared header into a private Region.
321 // Feed that region as the one backedge to this loop.
322 void merge_many_backedges( PhaseIdealLoop *phase );
324 // Split shared headers and insert loop landing pads.
325 // Insert a LoopNode to replace the RegionNode.
326 // Returns TRUE if loop tree is structurally changed.
327 bool beautify_loops( PhaseIdealLoop *phase );
329 // Perform optimization to use the loop predicates for null checks and range checks.
330 // Applies to any loop level (not just the innermost one)
331 bool loop_predication( PhaseIdealLoop *phase);
333 // Perform iteration-splitting on inner loops. Split iterations to
334 // avoid range checks or one-shot null checks. Returns false if the
335 // current round of loop opts should stop.
336 bool iteration_split( PhaseIdealLoop *phase, Node_List &old_new );
338 // Driver for various flavors of iteration splitting. Returns false
339 // if the current round of loop opts should stop.
340 bool iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new );
342 // Given dominators, try to find loops with calls that must always be
343 // executed (call dominates loop tail). These loops do not need non-call
344 // safepoints (ncsfpt).
345 void check_safepts(VectorSet &visited, Node_List &stack);
347 // Allpaths backwards scan from loop tail, terminating each path at first safepoint
348 // encountered.
349 void allpaths_check_safepts(VectorSet &visited, Node_List &stack);
351 // Convert to counted loops where possible
352 void counted_loop( PhaseIdealLoop *phase );
354 // Check for Node being a loop-breaking test
355 Node *is_loop_exit(Node *iff) const;
357 // Returns true if ctrl is executed on every complete iteration
358 bool dominates_backedge(Node* ctrl);
360 // Remove simplistic dead code from loop body
361 void DCE_loop_body();
363 // Look for loop-exit tests with my 50/50 guesses from the Parsing stage.
364 // Replace with a 1-in-10 exit guess.
365 void adjust_loop_exit_prob( PhaseIdealLoop *phase );
367 // Return TRUE or FALSE if the loop should never be RCE'd or aligned.
368 // Useful for unrolling loops with NO array accesses.
369 bool policy_peel_only( PhaseIdealLoop *phase ) const;
371 // Return TRUE or FALSE if the loop should be unswitched -- clone
372 // loop with an invariant test
373 bool policy_unswitching( PhaseIdealLoop *phase ) const;
375 // Micro-benchmark spamming. Remove empty loops.
376 bool policy_do_remove_empty_loop( PhaseIdealLoop *phase );
378 // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can
379 // make some loop-invariant test (usually a null-check) happen before the
380 // loop.
381 bool policy_peeling( PhaseIdealLoop *phase ) const;
383 // Return TRUE or FALSE if the loop should be maximally unrolled. Stash any
384 // known trip count in the counted loop node.
385 bool policy_maximally_unroll( PhaseIdealLoop *phase ) const;
387 // Return TRUE or FALSE if the loop should be unrolled or not. Unroll if
388 // the loop is a CountedLoop and the body is small enough.
389 bool policy_unroll( PhaseIdealLoop *phase ) const;
391 // Return TRUE or FALSE if the loop should be range-check-eliminated.
392 // Gather a list of IF tests that are dominated by iteration splitting;
393 // also gather the end of the first split and the start of the 2nd split.
394 bool policy_range_check( PhaseIdealLoop *phase ) const;
396 // Return TRUE or FALSE if the loop should be cache-line aligned.
397 // Gather the expression that does the alignment. Note that only
398 // one array base can be aligned in a loop (unless the VM guarantees
399 // mutual alignment). Note that if we vectorize short memory ops
400 // into longer memory ops, we may want to increase alignment.
401 bool policy_align( PhaseIdealLoop *phase ) const;
403 // Return TRUE if "iff" is a range check.
404 bool is_range_check_if(IfNode *iff, PhaseIdealLoop *phase, Invariance& invar) const;
406 // Compute loop trip count from profile data
407 void compute_profile_trip_cnt( PhaseIdealLoop *phase );
409 // Reassociate invariant expressions.
410 void reassociate_invariants(PhaseIdealLoop *phase);
411 // Reassociate invariant add and subtract expressions.
412 Node* reassociate_add_sub(Node* n1, PhaseIdealLoop *phase);
413 // Return nonzero index of invariant operand if invariant and variant
414 // are combined with an Add or Sub. Helper for reassociate_invariants.
415 int is_invariant_addition(Node* n, PhaseIdealLoop *phase);
417 // Return true if n is invariant
418 bool is_invariant(Node* n) const;
420 // Put loop body on igvn work list
421 void record_for_igvn();
423 bool is_loop() { return !_irreducible && _tail && !_tail->is_top(); }
424 bool is_inner() { return is_loop() && _child == NULL; }
425 bool is_counted() { return is_loop() && _head != NULL && _head->is_CountedLoop(); }
427 #ifndef PRODUCT
428 void dump_head( ) const; // Dump loop head only
429 void dump() const; // Dump this loop recursively
430 void verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const;
431 #endif
433 };
435 // -----------------------------PhaseIdealLoop---------------------------------
436 // Computes the mapping from Nodes to IdealLoopTrees. Organizes IdealLoopTrees into a
437 // loop tree. Drives the loop-based transformations on the ideal graph.
438 class PhaseIdealLoop : public PhaseTransform {
439 friend class IdealLoopTree;
440 friend class SuperWord;
441 // Pre-computed def-use info
442 PhaseIterGVN &_igvn;
444 // Head of loop tree
445 IdealLoopTree *_ltree_root;
447 // Array of pre-order numbers, plus post-visited bit.
448 // ZERO for not pre-visited. EVEN for pre-visited but not post-visited.
449 // ODD for post-visited. Other bits are the pre-order number.
450 uint *_preorders;
451 uint _max_preorder;
453 const PhaseIdealLoop* _verify_me;
454 bool _verify_only;
456 // Allocate _preorders[] array
457 void allocate_preorders() {
458 _max_preorder = C->unique()+8;
459 _preorders = NEW_RESOURCE_ARRAY(uint, _max_preorder);
460 memset(_preorders, 0, sizeof(uint) * _max_preorder);
461 }
463 // Allocate _preorders[] array
464 void reallocate_preorders() {
465 if ( _max_preorder < C->unique() ) {
466 _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, C->unique());
467 _max_preorder = C->unique();
468 }
469 memset(_preorders, 0, sizeof(uint) * _max_preorder);
470 }
472 // Check to grow _preorders[] array for the case when build_loop_tree_impl()
473 // adds new nodes.
474 void check_grow_preorders( ) {
475 if ( _max_preorder < C->unique() ) {
476 uint newsize = _max_preorder<<1; // double size of array
477 _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, newsize);
478 memset(&_preorders[_max_preorder],0,sizeof(uint)*(newsize-_max_preorder));
479 _max_preorder = newsize;
480 }
481 }
482 // Check for pre-visited. Zero for NOT visited; non-zero for visited.
483 int is_visited( Node *n ) const { return _preorders[n->_idx]; }
484 // Pre-order numbers are written to the Nodes array as low-bit-set values.
485 void set_preorder_visited( Node *n, int pre_order ) {
486 assert( !is_visited( n ), "already set" );
487 _preorders[n->_idx] = (pre_order<<1);
488 };
489 // Return pre-order number.
490 int get_preorder( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]>>1; }
492 // Check for being post-visited.
493 // Should be previsited already (checked with assert(is_visited(n))).
494 int is_postvisited( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]&1; }
496 // Mark as post visited
497 void set_postvisited( Node *n ) { assert( !is_postvisited( n ), "" ); _preorders[n->_idx] |= 1; }
499 // Set/get control node out. Set lower bit to distinguish from IdealLoopTree
500 // Returns true if "n" is a data node, false if it's a control node.
501 bool has_ctrl( Node *n ) const { return ((intptr_t)_nodes[n->_idx]) & 1; }
503 // clear out dead code after build_loop_late
504 Node_List _deadlist;
506 // Support for faster execution of get_late_ctrl()/dom_lca()
507 // when a node has many uses and dominator depth is deep.
508 Node_Array _dom_lca_tags;
509 void init_dom_lca_tags();
510 void clear_dom_lca_tags();
512 // Helper for debugging bad dominance relationships
513 bool verify_dominance(Node* n, Node* use, Node* LCA, Node* early);
515 Node* compute_lca_of_uses(Node* n, Node* early, bool verify = false);
517 // Inline wrapper for frequent cases:
518 // 1) only one use
519 // 2) a use is the same as the current LCA passed as 'n1'
520 Node *dom_lca_for_get_late_ctrl( Node *lca, Node *n, Node *tag ) {
521 assert( n->is_CFG(), "" );
522 // Fast-path NULL lca
523 if( lca != NULL && lca != n ) {
524 assert( lca->is_CFG(), "" );
525 // find LCA of all uses
526 n = dom_lca_for_get_late_ctrl_internal( lca, n, tag );
527 }
528 return find_non_split_ctrl(n);
529 }
530 Node *dom_lca_for_get_late_ctrl_internal( Node *lca, Node *n, Node *tag );
532 // Helper function for directing control inputs away from CFG split
533 // points.
534 Node *find_non_split_ctrl( Node *ctrl ) const {
535 if (ctrl != NULL) {
536 if (ctrl->is_MultiBranch()) {
537 ctrl = ctrl->in(0);
538 }
539 assert(ctrl->is_CFG(), "CFG");
540 }
541 return ctrl;
542 }
544 public:
545 bool has_node( Node* n ) const { return _nodes[n->_idx] != NULL; }
546 // check if transform created new nodes that need _ctrl recorded
547 Node *get_late_ctrl( Node *n, Node *early );
548 Node *get_early_ctrl( Node *n );
549 void set_early_ctrl( Node *n );
550 void set_subtree_ctrl( Node *root );
551 void set_ctrl( Node *n, Node *ctrl ) {
552 assert( !has_node(n) || has_ctrl(n), "" );
553 assert( ctrl->in(0), "cannot set dead control node" );
554 assert( ctrl == find_non_split_ctrl(ctrl), "must set legal crtl" );
555 _nodes.map( n->_idx, (Node*)((intptr_t)ctrl + 1) );
556 }
557 // Set control and update loop membership
558 void set_ctrl_and_loop(Node* n, Node* ctrl) {
559 IdealLoopTree* old_loop = get_loop(get_ctrl(n));
560 IdealLoopTree* new_loop = get_loop(ctrl);
561 if (old_loop != new_loop) {
562 if (old_loop->_child == NULL) old_loop->_body.yank(n);
563 if (new_loop->_child == NULL) new_loop->_body.push(n);
564 }
565 set_ctrl(n, ctrl);
566 }
567 // Control nodes can be replaced or subsumed. During this pass they
568 // get their replacement Node in slot 1. Instead of updating the block
569 // location of all Nodes in the subsumed block, we lazily do it. As we
570 // pull such a subsumed block out of the array, we write back the final
571 // correct block.
572 Node *get_ctrl( Node *i ) {
573 assert(has_node(i), "");
574 Node *n = get_ctrl_no_update(i);
575 _nodes.map( i->_idx, (Node*)((intptr_t)n + 1) );
576 assert(has_node(i) && has_ctrl(i), "");
577 assert(n == find_non_split_ctrl(n), "must return legal ctrl" );
578 return n;
579 }
580 // true if CFG node d dominates CFG node n
581 bool is_dominator(Node *d, Node *n);
582 // return get_ctrl for a data node and self(n) for a CFG node
583 Node* ctrl_or_self(Node* n) {
584 if (has_ctrl(n))
585 return get_ctrl(n);
586 else {
587 assert (n->is_CFG(), "must be a CFG node");
588 return n;
589 }
590 }
592 private:
593 Node *get_ctrl_no_update( Node *i ) const {
594 assert( has_ctrl(i), "" );
595 Node *n = (Node*)(((intptr_t)_nodes[i->_idx]) & ~1);
596 if (!n->in(0)) {
597 // Skip dead CFG nodes
598 do {
599 n = (Node*)(((intptr_t)_nodes[n->_idx]) & ~1);
600 } while (!n->in(0));
601 n = find_non_split_ctrl(n);
602 }
603 return n;
604 }
606 // Check for loop being set
607 // "n" must be a control node. Returns true if "n" is known to be in a loop.
608 bool has_loop( Node *n ) const {
609 assert(!has_node(n) || !has_ctrl(n), "");
610 return has_node(n);
611 }
612 // Set loop
613 void set_loop( Node *n, IdealLoopTree *loop ) {
614 _nodes.map(n->_idx, (Node*)loop);
615 }
616 // Lazy-dazy update of 'get_ctrl' and 'idom_at' mechanisms. Replace
617 // the 'old_node' with 'new_node'. Kill old-node. Add a reference
618 // from old_node to new_node to support the lazy update. Reference
619 // replaces loop reference, since that is not needed for dead node.
620 public:
621 void lazy_update( Node *old_node, Node *new_node ) {
622 assert( old_node != new_node, "no cycles please" );
623 //old_node->set_req( 1, new_node /*NO DU INFO*/ );
624 // Nodes always have DU info now, so re-use the side array slot
625 // for this node to provide the forwarding pointer.
626 _nodes.map( old_node->_idx, (Node*)((intptr_t)new_node + 1) );
627 }
628 void lazy_replace( Node *old_node, Node *new_node ) {
629 _igvn.hash_delete(old_node);
630 _igvn.subsume_node( old_node, new_node );
631 lazy_update( old_node, new_node );
632 }
633 void lazy_replace_proj( Node *old_node, Node *new_node ) {
634 assert( old_node->req() == 1, "use this for Projs" );
635 _igvn.hash_delete(old_node); // Must hash-delete before hacking edges
636 old_node->add_req( NULL );
637 lazy_replace( old_node, new_node );
638 }
640 private:
642 // Place 'n' in some loop nest, where 'n' is a CFG node
643 void build_loop_tree();
644 int build_loop_tree_impl( Node *n, int pre_order );
645 // Insert loop into the existing loop tree. 'innermost' is a leaf of the
646 // loop tree, not the root.
647 IdealLoopTree *sort( IdealLoopTree *loop, IdealLoopTree *innermost );
649 // Place Data nodes in some loop nest
650 void build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack );
651 void build_loop_late ( VectorSet &visited, Node_List &worklist, Node_Stack &nstack );
652 void build_loop_late_post ( Node* n );
654 // Array of immediate dominance info for each CFG node indexed by node idx
655 private:
656 uint _idom_size;
657 Node **_idom; // Array of immediate dominators
658 uint *_dom_depth; // Used for fast LCA test
659 GrowableArray<uint>* _dom_stk; // For recomputation of dom depth
661 Node* idom_no_update(Node* d) const {
662 assert(d->_idx < _idom_size, "oob");
663 Node* n = _idom[d->_idx];
664 assert(n != NULL,"Bad immediate dominator info.");
665 while (n->in(0) == NULL) { // Skip dead CFG nodes
666 //n = n->in(1);
667 n = (Node*)(((intptr_t)_nodes[n->_idx]) & ~1);
668 assert(n != NULL,"Bad immediate dominator info.");
669 }
670 return n;
671 }
672 Node *idom(Node* d) const {
673 uint didx = d->_idx;
674 Node *n = idom_no_update(d);
675 _idom[didx] = n; // Lazily remove dead CFG nodes from table.
676 return n;
677 }
678 uint dom_depth(Node* d) const {
679 assert(d->_idx < _idom_size, "");
680 return _dom_depth[d->_idx];
681 }
682 void set_idom(Node* d, Node* n, uint dom_depth);
683 // Locally compute IDOM using dom_lca call
684 Node *compute_idom( Node *region ) const;
685 // Recompute dom_depth
686 void recompute_dom_depth();
688 // Is safept not required by an outer loop?
689 bool is_deleteable_safept(Node* sfpt);
691 // Perform verification that the graph is valid.
692 PhaseIdealLoop( PhaseIterGVN &igvn) :
693 PhaseTransform(Ideal_Loop),
694 _igvn(igvn),
695 _dom_lca_tags(C->comp_arena()),
696 _verify_me(NULL),
697 _verify_only(true) {
698 build_and_optimize(false, false);
699 }
701 // build the loop tree and perform any requested optimizations
702 void build_and_optimize(bool do_split_if, bool do_loop_pred);
704 public:
705 // Dominators for the sea of nodes
706 void Dominators();
707 Node *dom_lca( Node *n1, Node *n2 ) const {
708 return find_non_split_ctrl(dom_lca_internal(n1, n2));
709 }
710 Node *dom_lca_internal( Node *n1, Node *n2 ) const;
712 // Compute the Ideal Node to Loop mapping
713 PhaseIdealLoop( PhaseIterGVN &igvn, bool do_split_ifs, bool do_loop_pred) :
714 PhaseTransform(Ideal_Loop),
715 _igvn(igvn),
716 _dom_lca_tags(C->comp_arena()),
717 _verify_me(NULL),
718 _verify_only(false) {
719 build_and_optimize(do_split_ifs, do_loop_pred);
720 }
722 // Verify that verify_me made the same decisions as a fresh run.
723 PhaseIdealLoop( PhaseIterGVN &igvn, const PhaseIdealLoop *verify_me) :
724 PhaseTransform(Ideal_Loop),
725 _igvn(igvn),
726 _dom_lca_tags(C->comp_arena()),
727 _verify_me(verify_me),
728 _verify_only(false) {
729 build_and_optimize(false, false);
730 }
732 // Build and verify the loop tree without modifying the graph. This
733 // is useful to verify that all inputs properly dominate their uses.
734 static void verify(PhaseIterGVN& igvn) {
735 #ifdef ASSERT
736 PhaseIdealLoop v(igvn);
737 #endif
738 }
740 // True if the method has at least 1 irreducible loop
741 bool _has_irreducible_loops;
743 // Per-Node transform
744 virtual Node *transform( Node *a_node ) { return 0; }
746 Node *is_counted_loop( Node *x, IdealLoopTree *loop );
748 // Return a post-walked LoopNode
749 IdealLoopTree *get_loop( Node *n ) const {
750 // Dead nodes have no loop, so return the top level loop instead
751 if (!has_node(n)) return _ltree_root;
752 assert(!has_ctrl(n), "");
753 return (IdealLoopTree*)_nodes[n->_idx];
754 }
756 // Is 'n' a (nested) member of 'loop'?
757 int is_member( const IdealLoopTree *loop, Node *n ) const {
758 return loop->is_member(get_loop(n)); }
760 // This is the basic building block of the loop optimizations. It clones an
761 // entire loop body. It makes an old_new loop body mapping; with this
762 // mapping you can find the new-loop equivalent to an old-loop node. All
763 // new-loop nodes are exactly equal to their old-loop counterparts, all
764 // edges are the same. All exits from the old-loop now have a RegionNode
765 // that merges the equivalent new-loop path. This is true even for the
766 // normal "loop-exit" condition. All uses of loop-invariant old-loop values
767 // now come from (one or more) Phis that merge their new-loop equivalents.
768 // Parameter side_by_side_idom:
769 // When side_by_size_idom is NULL, the dominator tree is constructed for
770 // the clone loop to dominate the original. Used in construction of
771 // pre-main-post loop sequence.
772 // When nonnull, the clone and original are side-by-side, both are
773 // dominated by the passed in side_by_side_idom node. Used in
774 // construction of unswitched loops.
775 void clone_loop( IdealLoopTree *loop, Node_List &old_new, int dom_depth,
776 Node* side_by_side_idom = NULL);
778 // If we got the effect of peeling, either by actually peeling or by
779 // making a pre-loop which must execute at least once, we can remove
780 // all loop-invariant dominated tests in the main body.
781 void peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new );
783 // Generate code to do a loop peel for the given loop (and body).
784 // old_new is a temp array.
785 void do_peeling( IdealLoopTree *loop, Node_List &old_new );
787 // Add pre and post loops around the given loop. These loops are used
788 // during RCE, unrolling and aligning loops.
789 void insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only );
790 // If Node n lives in the back_ctrl block, we clone a private version of n
791 // in preheader_ctrl block and return that, otherwise return n.
792 Node *clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n );
794 // Take steps to maximally unroll the loop. Peel any odd iterations, then
795 // unroll to do double iterations. The next round of major loop transforms
796 // will repeat till the doubled loop body does all remaining iterations in 1
797 // pass.
798 void do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new );
800 // Unroll the loop body one step - make each trip do 2 iterations.
801 void do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip );
803 // Return true if exp is a constant times an induction var
804 bool is_scaled_iv(Node* exp, Node* iv, int* p_scale);
806 // Return true if exp is a scaled induction var plus (or minus) constant
807 bool is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth = 0);
809 // Return true if proj is for "proj->[region->..]call_uct"
810 bool is_uncommon_trap_proj(ProjNode* proj, bool must_reason_predicate = false);
811 // Return true for "if(test)-> proj -> ...
812 // |
813 // V
814 // other_proj->[region->..]call_uct"
815 bool is_uncommon_trap_if_pattern(ProjNode* proj, bool must_reason_predicate = false);
816 // Create a new if above the uncommon_trap_if_pattern for the predicate to be promoted
817 ProjNode* create_new_if_for_predicate(ProjNode* cont_proj);
818 // Find a good location to insert a predicate
819 ProjNode* find_predicate_insertion_point(Node* start_c);
820 // Construct a range check for a predicate if
821 BoolNode* rc_predicate(Node* ctrl,
822 int scale, Node* offset,
823 Node* init, Node* limit, Node* stride,
824 Node* range);
826 // Implementation of the loop predication to promote checks outside the loop
827 bool loop_predication_impl(IdealLoopTree *loop);
829 // Helper function to collect predicate for eliminating the useless ones
830 void collect_potentially_useful_predicates(IdealLoopTree *loop, Unique_Node_List &predicate_opaque1);
831 void eliminate_useless_predicates();
833 // Eliminate range-checks and other trip-counter vs loop-invariant tests.
834 void do_range_check( IdealLoopTree *loop, Node_List &old_new );
836 // Create a slow version of the loop by cloning the loop
837 // and inserting an if to select fast-slow versions.
838 ProjNode* create_slow_version_of_loop(IdealLoopTree *loop,
839 Node_List &old_new);
841 // Clone loop with an invariant test (that does not exit) and
842 // insert a clone of the test that selects which version to
843 // execute.
844 void do_unswitching (IdealLoopTree *loop, Node_List &old_new);
846 // Find candidate "if" for unswitching
847 IfNode* find_unswitching_candidate(const IdealLoopTree *loop) const;
849 // Range Check Elimination uses this function!
850 // Constrain the main loop iterations so the affine function:
851 // scale_con * I + offset < limit
852 // always holds true. That is, either increase the number of iterations in
853 // the pre-loop or the post-loop until the condition holds true in the main
854 // loop. Scale_con, offset and limit are all loop invariant.
855 void add_constraint( int stride_con, int scale_con, Node *offset, Node *limit, Node *pre_ctrl, Node **pre_limit, Node **main_limit );
857 // Partially peel loop up through last_peel node.
858 bool partial_peel( IdealLoopTree *loop, Node_List &old_new );
860 // Create a scheduled list of nodes control dependent on ctrl set.
861 void scheduled_nodelist( IdealLoopTree *loop, VectorSet& ctrl, Node_List &sched );
862 // Has a use in the vector set
863 bool has_use_in_set( Node* n, VectorSet& vset );
864 // Has use internal to the vector set (ie. not in a phi at the loop head)
865 bool has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop );
866 // clone "n" for uses that are outside of loop
867 void clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist );
868 // clone "n" for special uses that are in the not_peeled region
869 void clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
870 VectorSet& not_peel, Node_List& sink_list, Node_List& worklist );
871 // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
872 void insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp );
873 #ifdef ASSERT
874 // Validate the loop partition sets: peel and not_peel
875 bool is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, VectorSet& not_peel );
876 // Ensure that uses outside of loop are of the right form
877 bool is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
878 uint orig_exit_idx, uint clone_exit_idx);
879 bool is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx);
880 #endif
882 // Returns nonzero constant stride if-node is a possible iv test (otherwise returns zero.)
883 int stride_of_possible_iv( Node* iff );
884 bool is_possible_iv_test( Node* iff ) { return stride_of_possible_iv(iff) != 0; }
885 // Return the (unique) control output node that's in the loop (if it exists.)
886 Node* stay_in_loop( Node* n, IdealLoopTree *loop);
887 // Insert a signed compare loop exit cloned from an unsigned compare.
888 IfNode* insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop);
889 void remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop);
890 // Utility to register node "n" with PhaseIdealLoop
891 void register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth);
892 // Utility to create an if-projection
893 ProjNode* proj_clone(ProjNode* p, IfNode* iff);
894 // Force the iff control output to be the live_proj
895 Node* short_circuit_if(IfNode* iff, ProjNode* live_proj);
896 // Insert a region before an if projection
897 RegionNode* insert_region_before_proj(ProjNode* proj);
898 // Insert a new if before an if projection
899 ProjNode* insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj);
901 // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
902 // "Nearly" because all Nodes have been cloned from the original in the loop,
903 // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
904 // through the Phi recursively, and return a Bool.
905 BoolNode *clone_iff( PhiNode *phi, IdealLoopTree *loop );
906 CmpNode *clone_bool( PhiNode *phi, IdealLoopTree *loop );
909 // Rework addressing expressions to get the most loop-invariant stuff
910 // moved out. We'd like to do all associative operators, but it's especially
911 // important (common) to do address expressions.
912 Node *remix_address_expressions( Node *n );
914 // Attempt to use a conditional move instead of a phi/branch
915 Node *conditional_move( Node *n );
917 // Reorganize offset computations to lower register pressure.
918 // Mostly prevent loop-fallout uses of the pre-incremented trip counter
919 // (which are then alive with the post-incremented trip counter
920 // forcing an extra register move)
921 void reorg_offsets( IdealLoopTree *loop );
923 // Check for aggressive application of 'split-if' optimization,
924 // using basic block level info.
925 void split_if_with_blocks ( VectorSet &visited, Node_Stack &nstack );
926 Node *split_if_with_blocks_pre ( Node *n );
927 void split_if_with_blocks_post( Node *n );
928 Node *has_local_phi_input( Node *n );
929 // Mark an IfNode as being dominated by a prior test,
930 // without actually altering the CFG (and hence IDOM info).
931 void dominated_by( Node *prevdom, Node *iff );
933 // Split Node 'n' through merge point
934 Node *split_thru_region( Node *n, Node *region );
935 // Split Node 'n' through merge point if there is enough win.
936 Node *split_thru_phi( Node *n, Node *region, int policy );
937 // Found an If getting its condition-code input from a Phi in the
938 // same block. Split thru the Region.
939 void do_split_if( Node *iff );
941 private:
942 // Return a type based on condition control flow
943 const TypeInt* filtered_type( Node *n, Node* n_ctrl);
944 const TypeInt* filtered_type( Node *n ) { return filtered_type(n, NULL); }
945 // Helpers for filtered type
946 const TypeInt* filtered_type_from_dominators( Node* val, Node *val_ctrl);
948 // Helper functions
949 Node *spinup( Node *iff, Node *new_false, Node *new_true, Node *region, Node *phi, small_cache *cache );
950 Node *find_use_block( Node *use, Node *def, Node *old_false, Node *new_false, Node *old_true, Node *new_true );
951 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 );
952 bool split_up( Node *n, Node *blk1, Node *blk2 );
953 void sink_use( Node *use, Node *post_loop );
954 Node *place_near_use( Node *useblock ) const;
956 bool _created_loop_node;
957 public:
958 void set_created_loop_node() { _created_loop_node = true; }
959 bool created_loop_node() { return _created_loop_node; }
960 void register_new_node( Node *n, Node *blk );
962 #ifndef PRODUCT
963 void dump( ) const;
964 void dump( IdealLoopTree *loop, uint rpo_idx, Node_List &rpo_list ) const;
965 void rpo( Node *start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list ) const;
966 void verify() const; // Major slow :-)
967 void verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const;
968 IdealLoopTree *get_loop_idx(Node* n) const {
969 // Dead nodes have no loop, so return the top level loop instead
970 return _nodes[n->_idx] ? (IdealLoopTree*)_nodes[n->_idx] : _ltree_root;
971 }
972 // Print some stats
973 static void print_statistics();
974 static int _loop_invokes; // Count of PhaseIdealLoop invokes
975 static int _loop_work; // Sum of PhaseIdealLoop x _unique
976 #endif
977 };
979 inline Node* IdealLoopTree::tail() {
980 // Handle lazy update of _tail field
981 Node *n = _tail;
982 //while( !n->in(0) ) // Skip dead CFG nodes
983 //n = n->in(1);
984 if (n->in(0) == NULL)
985 n = _phase->get_ctrl(n);
986 _tail = n;
987 return n;
988 }
991 // Iterate over the loop tree using a preorder, left-to-right traversal.
992 //
993 // Example that visits all counted loops from within PhaseIdealLoop
994 //
995 // for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
996 // IdealLoopTree* lpt = iter.current();
997 // if (!lpt->is_counted()) continue;
998 // ...
999 class LoopTreeIterator : public StackObj {
1000 private:
1001 IdealLoopTree* _root;
1002 IdealLoopTree* _curnt;
1004 public:
1005 LoopTreeIterator(IdealLoopTree* root) : _root(root), _curnt(root) {}
1007 bool done() { return _curnt == NULL; } // Finished iterating?
1009 void next(); // Advance to next loop tree
1011 IdealLoopTree* current() { return _curnt; } // Return current value of iterator.
1012 };