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
duke@435 | 1 | /* |
never@1356 | 2 | * Copyright 1998-2009 Sun Microsystems, Inc. All Rights Reserved. |
duke@435 | 3 | * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. |
duke@435 | 4 | * |
duke@435 | 5 | * This code is free software; you can redistribute it and/or modify it |
duke@435 | 6 | * under the terms of the GNU General Public License version 2 only, as |
duke@435 | 7 | * published by the Free Software Foundation. |
duke@435 | 8 | * |
duke@435 | 9 | * This code is distributed in the hope that it will be useful, but WITHOUT |
duke@435 | 10 | * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or |
duke@435 | 11 | * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
duke@435 | 12 | * version 2 for more details (a copy is included in the LICENSE file that |
duke@435 | 13 | * accompanied this code). |
duke@435 | 14 | * |
duke@435 | 15 | * You should have received a copy of the GNU General Public License version |
duke@435 | 16 | * 2 along with this work; if not, write to the Free Software Foundation, |
duke@435 | 17 | * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. |
duke@435 | 18 | * |
duke@435 | 19 | * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, |
duke@435 | 20 | * CA 95054 USA or visit www.sun.com if you need additional information or |
duke@435 | 21 | * have any questions. |
duke@435 | 22 | * |
duke@435 | 23 | */ |
duke@435 | 24 | |
duke@435 | 25 | class CmpNode; |
duke@435 | 26 | class CountedLoopEndNode; |
duke@435 | 27 | class CountedLoopNode; |
duke@435 | 28 | class IdealLoopTree; |
duke@435 | 29 | class LoopNode; |
duke@435 | 30 | class Node; |
duke@435 | 31 | class PhaseIdealLoop; |
duke@435 | 32 | class VectorSet; |
cfang@1607 | 33 | class Invariance; |
duke@435 | 34 | struct small_cache; |
duke@435 | 35 | |
duke@435 | 36 | // |
duke@435 | 37 | // I D E A L I Z E D L O O P S |
duke@435 | 38 | // |
duke@435 | 39 | // Idealized loops are the set of loops I perform more interesting |
duke@435 | 40 | // transformations on, beyond simple hoisting. |
duke@435 | 41 | |
duke@435 | 42 | //------------------------------LoopNode--------------------------------------- |
duke@435 | 43 | // Simple loop header. Fall in path on left, loop-back path on right. |
duke@435 | 44 | class LoopNode : public RegionNode { |
duke@435 | 45 | // Size is bigger to hold the flags. However, the flags do not change |
duke@435 | 46 | // the semantics so it does not appear in the hash & cmp functions. |
duke@435 | 47 | virtual uint size_of() const { return sizeof(*this); } |
duke@435 | 48 | protected: |
duke@435 | 49 | short _loop_flags; |
duke@435 | 50 | // Names for flag bitfields |
duke@435 | 51 | enum { pre_post_main=0, inner_loop=8, partial_peel_loop=16, partial_peel_failed=32 }; |
duke@435 | 52 | char _unswitch_count; |
duke@435 | 53 | enum { _unswitch_max=3 }; |
duke@435 | 54 | |
duke@435 | 55 | public: |
duke@435 | 56 | // Names for edge indices |
duke@435 | 57 | enum { Self=0, EntryControl, LoopBackControl }; |
duke@435 | 58 | |
duke@435 | 59 | int is_inner_loop() const { return _loop_flags & inner_loop; } |
duke@435 | 60 | void set_inner_loop() { _loop_flags |= inner_loop; } |
duke@435 | 61 | |
duke@435 | 62 | int is_partial_peel_loop() const { return _loop_flags & partial_peel_loop; } |
duke@435 | 63 | void set_partial_peel_loop() { _loop_flags |= partial_peel_loop; } |
duke@435 | 64 | int partial_peel_has_failed() const { return _loop_flags & partial_peel_failed; } |
duke@435 | 65 | void mark_partial_peel_failed() { _loop_flags |= partial_peel_failed; } |
duke@435 | 66 | |
duke@435 | 67 | int unswitch_max() { return _unswitch_max; } |
duke@435 | 68 | int unswitch_count() { return _unswitch_count; } |
duke@435 | 69 | void set_unswitch_count(int val) { |
duke@435 | 70 | assert (val <= unswitch_max(), "too many unswitches"); |
duke@435 | 71 | _unswitch_count = val; |
duke@435 | 72 | } |
duke@435 | 73 | |
duke@435 | 74 | LoopNode( Node *entry, Node *backedge ) : RegionNode(3), _loop_flags(0), _unswitch_count(0) { |
duke@435 | 75 | init_class_id(Class_Loop); |
duke@435 | 76 | init_req(EntryControl, entry); |
duke@435 | 77 | init_req(LoopBackControl, backedge); |
duke@435 | 78 | } |
duke@435 | 79 | |
duke@435 | 80 | virtual Node *Ideal(PhaseGVN *phase, bool can_reshape); |
duke@435 | 81 | virtual int Opcode() const; |
duke@435 | 82 | bool can_be_counted_loop(PhaseTransform* phase) const { |
duke@435 | 83 | return req() == 3 && in(0) != NULL && |
duke@435 | 84 | in(1) != NULL && phase->type(in(1)) != Type::TOP && |
duke@435 | 85 | in(2) != NULL && phase->type(in(2)) != Type::TOP; |
duke@435 | 86 | } |
duke@435 | 87 | #ifndef PRODUCT |
duke@435 | 88 | virtual void dump_spec(outputStream *st) const; |
duke@435 | 89 | #endif |
duke@435 | 90 | }; |
duke@435 | 91 | |
duke@435 | 92 | //------------------------------Counted Loops---------------------------------- |
duke@435 | 93 | // Counted loops are all trip-counted loops, with exactly 1 trip-counter exit |
duke@435 | 94 | // path (and maybe some other exit paths). The trip-counter exit is always |
duke@435 | 95 | // last in the loop. The trip-counter does not have to stride by a constant, |
duke@435 | 96 | // but it does have to stride by a loop-invariant amount; the exit value is |
duke@435 | 97 | // also loop invariant. |
duke@435 | 98 | |
duke@435 | 99 | // CountedLoopNodes and CountedLoopEndNodes come in matched pairs. The |
duke@435 | 100 | // CountedLoopNode has the incoming loop control and the loop-back-control |
duke@435 | 101 | // which is always the IfTrue before the matching CountedLoopEndNode. The |
duke@435 | 102 | // CountedLoopEndNode has an incoming control (possibly not the |
duke@435 | 103 | // CountedLoopNode if there is control flow in the loop), the post-increment |
duke@435 | 104 | // trip-counter value, and the limit. The trip-counter value is always of |
duke@435 | 105 | // the form (Op old-trip-counter stride). The old-trip-counter is produced |
duke@435 | 106 | // by a Phi connected to the CountedLoopNode. The stride is loop invariant. |
duke@435 | 107 | // The Op is any commutable opcode, including Add, Mul, Xor. The |
duke@435 | 108 | // CountedLoopEndNode also takes in the loop-invariant limit value. |
duke@435 | 109 | |
duke@435 | 110 | // From a CountedLoopNode I can reach the matching CountedLoopEndNode via the |
duke@435 | 111 | // loop-back control. From CountedLoopEndNodes I can reach CountedLoopNodes |
duke@435 | 112 | // via the old-trip-counter from the Op node. |
duke@435 | 113 | |
duke@435 | 114 | //------------------------------CountedLoopNode-------------------------------- |
duke@435 | 115 | // CountedLoopNodes head simple counted loops. CountedLoopNodes have as |
duke@435 | 116 | // inputs the incoming loop-start control and the loop-back control, so they |
duke@435 | 117 | // act like RegionNodes. They also take in the initial trip counter, the |
duke@435 | 118 | // loop-invariant stride and the loop-invariant limit value. CountedLoopNodes |
duke@435 | 119 | // produce a loop-body control and the trip counter value. Since |
duke@435 | 120 | // CountedLoopNodes behave like RegionNodes I still have a standard CFG model. |
duke@435 | 121 | |
duke@435 | 122 | class CountedLoopNode : public LoopNode { |
duke@435 | 123 | // Size is bigger to hold _main_idx. However, _main_idx does not change |
duke@435 | 124 | // the semantics so it does not appear in the hash & cmp functions. |
duke@435 | 125 | virtual uint size_of() const { return sizeof(*this); } |
duke@435 | 126 | |
duke@435 | 127 | // For Pre- and Post-loops during debugging ONLY, this holds the index of |
duke@435 | 128 | // the Main CountedLoop. Used to assert that we understand the graph shape. |
duke@435 | 129 | node_idx_t _main_idx; |
duke@435 | 130 | |
duke@435 | 131 | // Known trip count calculated by policy_maximally_unroll |
duke@435 | 132 | int _trip_count; |
duke@435 | 133 | |
duke@435 | 134 | // Expected trip count from profile data |
duke@435 | 135 | float _profile_trip_cnt; |
duke@435 | 136 | |
duke@435 | 137 | // Log2 of original loop bodies in unrolled loop |
duke@435 | 138 | int _unrolled_count_log2; |
duke@435 | 139 | |
duke@435 | 140 | // Node count prior to last unrolling - used to decide if |
duke@435 | 141 | // unroll,optimize,unroll,optimize,... is making progress |
duke@435 | 142 | int _node_count_before_unroll; |
duke@435 | 143 | |
duke@435 | 144 | public: |
duke@435 | 145 | CountedLoopNode( Node *entry, Node *backedge ) |
duke@435 | 146 | : LoopNode(entry, backedge), _trip_count(max_jint), |
duke@435 | 147 | _profile_trip_cnt(COUNT_UNKNOWN), _unrolled_count_log2(0), |
duke@435 | 148 | _node_count_before_unroll(0) { |
duke@435 | 149 | init_class_id(Class_CountedLoop); |
duke@435 | 150 | // Initialize _trip_count to the largest possible value. |
duke@435 | 151 | // Will be reset (lower) if the loop's trip count is known. |
duke@435 | 152 | } |
duke@435 | 153 | |
duke@435 | 154 | virtual int Opcode() const; |
duke@435 | 155 | virtual Node *Ideal(PhaseGVN *phase, bool can_reshape); |
duke@435 | 156 | |
duke@435 | 157 | Node *init_control() const { return in(EntryControl); } |
duke@435 | 158 | Node *back_control() const { return in(LoopBackControl); } |
duke@435 | 159 | CountedLoopEndNode *loopexit() const; |
duke@435 | 160 | Node *init_trip() const; |
duke@435 | 161 | Node *stride() const; |
duke@435 | 162 | int stride_con() const; |
duke@435 | 163 | bool stride_is_con() const; |
duke@435 | 164 | Node *limit() const; |
duke@435 | 165 | Node *incr() const; |
duke@435 | 166 | Node *phi() const; |
duke@435 | 167 | |
duke@435 | 168 | // Match increment with optional truncation |
duke@435 | 169 | static Node* match_incr_with_optional_truncation(Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type); |
duke@435 | 170 | |
duke@435 | 171 | // A 'main' loop has a pre-loop and a post-loop. The 'main' loop |
duke@435 | 172 | // can run short a few iterations and may start a few iterations in. |
duke@435 | 173 | // It will be RCE'd and unrolled and aligned. |
duke@435 | 174 | |
duke@435 | 175 | // A following 'post' loop will run any remaining iterations. Used |
duke@435 | 176 | // during Range Check Elimination, the 'post' loop will do any final |
duke@435 | 177 | // iterations with full checks. Also used by Loop Unrolling, where |
duke@435 | 178 | // the 'post' loop will do any epilog iterations needed. Basically, |
duke@435 | 179 | // a 'post' loop can not profitably be further unrolled or RCE'd. |
duke@435 | 180 | |
duke@435 | 181 | // A preceding 'pre' loop will run at least 1 iteration (to do peeling), |
duke@435 | 182 | // it may do under-flow checks for RCE and may do alignment iterations |
duke@435 | 183 | // so the following main loop 'knows' that it is striding down cache |
duke@435 | 184 | // lines. |
duke@435 | 185 | |
duke@435 | 186 | // A 'main' loop that is ONLY unrolled or peeled, never RCE'd or |
duke@435 | 187 | // Aligned, may be missing it's pre-loop. |
duke@435 | 188 | enum { Normal=0, Pre=1, Main=2, Post=3, PrePostFlagsMask=3, Main_Has_No_Pre_Loop=4 }; |
duke@435 | 189 | int is_normal_loop() const { return (_loop_flags&PrePostFlagsMask) == Normal; } |
duke@435 | 190 | int is_pre_loop () const { return (_loop_flags&PrePostFlagsMask) == Pre; } |
duke@435 | 191 | int is_main_loop () const { return (_loop_flags&PrePostFlagsMask) == Main; } |
duke@435 | 192 | int is_post_loop () const { return (_loop_flags&PrePostFlagsMask) == Post; } |
duke@435 | 193 | int is_main_no_pre_loop() const { return _loop_flags & Main_Has_No_Pre_Loop; } |
duke@435 | 194 | void set_main_no_pre_loop() { _loop_flags |= Main_Has_No_Pre_Loop; } |
duke@435 | 195 | |
never@802 | 196 | int main_idx() const { return _main_idx; } |
never@802 | 197 | |
duke@435 | 198 | |
duke@435 | 199 | void set_pre_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Pre ; _main_idx = main->_idx; } |
duke@435 | 200 | void set_main_loop ( ) { assert(is_normal_loop(),""); _loop_flags |= Main; } |
duke@435 | 201 | void set_post_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Post; _main_idx = main->_idx; } |
duke@435 | 202 | void set_normal_loop( ) { _loop_flags &= ~PrePostFlagsMask; } |
duke@435 | 203 | |
duke@435 | 204 | void set_trip_count(int tc) { _trip_count = tc; } |
duke@435 | 205 | int trip_count() { return _trip_count; } |
duke@435 | 206 | |
duke@435 | 207 | void set_profile_trip_cnt(float ptc) { _profile_trip_cnt = ptc; } |
duke@435 | 208 | float profile_trip_cnt() { return _profile_trip_cnt; } |
duke@435 | 209 | |
duke@435 | 210 | void double_unrolled_count() { _unrolled_count_log2++; } |
duke@435 | 211 | int unrolled_count() { return 1 << MIN2(_unrolled_count_log2, BitsPerInt-3); } |
duke@435 | 212 | |
duke@435 | 213 | void set_node_count_before_unroll(int ct) { _node_count_before_unroll = ct; } |
duke@435 | 214 | int node_count_before_unroll() { return _node_count_before_unroll; } |
duke@435 | 215 | |
duke@435 | 216 | #ifndef PRODUCT |
duke@435 | 217 | virtual void dump_spec(outputStream *st) const; |
duke@435 | 218 | #endif |
duke@435 | 219 | }; |
duke@435 | 220 | |
duke@435 | 221 | //------------------------------CountedLoopEndNode----------------------------- |
duke@435 | 222 | // CountedLoopEndNodes end simple trip counted loops. They act much like |
duke@435 | 223 | // IfNodes. |
duke@435 | 224 | class CountedLoopEndNode : public IfNode { |
duke@435 | 225 | public: |
duke@435 | 226 | enum { TestControl, TestValue }; |
duke@435 | 227 | |
duke@435 | 228 | CountedLoopEndNode( Node *control, Node *test, float prob, float cnt ) |
duke@435 | 229 | : IfNode( control, test, prob, cnt) { |
duke@435 | 230 | init_class_id(Class_CountedLoopEnd); |
duke@435 | 231 | } |
duke@435 | 232 | virtual int Opcode() const; |
duke@435 | 233 | |
duke@435 | 234 | Node *cmp_node() const { return (in(TestValue)->req() >=2) ? in(TestValue)->in(1) : NULL; } |
duke@435 | 235 | Node *incr() const { Node *tmp = cmp_node(); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; } |
duke@435 | 236 | Node *limit() const { Node *tmp = cmp_node(); return (tmp && tmp->req()==3) ? tmp->in(2) : NULL; } |
duke@435 | 237 | Node *stride() const { Node *tmp = incr (); return (tmp && tmp->req()==3) ? tmp->in(2) : NULL; } |
duke@435 | 238 | Node *phi() const { Node *tmp = incr (); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; } |
duke@435 | 239 | Node *init_trip() const { Node *tmp = phi (); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; } |
duke@435 | 240 | int stride_con() const; |
duke@435 | 241 | bool stride_is_con() const { Node *tmp = stride (); return (tmp != NULL && tmp->is_Con()); } |
duke@435 | 242 | BoolTest::mask test_trip() const { return in(TestValue)->as_Bool()->_test._test; } |
duke@435 | 243 | CountedLoopNode *loopnode() const { |
duke@435 | 244 | Node *ln = phi()->in(0); |
duke@435 | 245 | assert( ln->Opcode() == Op_CountedLoop, "malformed loop" ); |
duke@435 | 246 | return (CountedLoopNode*)ln; } |
duke@435 | 247 | |
duke@435 | 248 | #ifndef PRODUCT |
duke@435 | 249 | virtual void dump_spec(outputStream *st) const; |
duke@435 | 250 | #endif |
duke@435 | 251 | }; |
duke@435 | 252 | |
duke@435 | 253 | |
duke@435 | 254 | inline CountedLoopEndNode *CountedLoopNode::loopexit() const { |
duke@435 | 255 | Node *bc = back_control(); |
duke@435 | 256 | if( bc == NULL ) return NULL; |
duke@435 | 257 | Node *le = bc->in(0); |
duke@435 | 258 | if( le->Opcode() != Op_CountedLoopEnd ) |
duke@435 | 259 | return NULL; |
duke@435 | 260 | return (CountedLoopEndNode*)le; |
duke@435 | 261 | } |
duke@435 | 262 | inline Node *CountedLoopNode::init_trip() const { return loopexit() ? loopexit()->init_trip() : NULL; } |
duke@435 | 263 | inline Node *CountedLoopNode::stride() const { return loopexit() ? loopexit()->stride() : NULL; } |
duke@435 | 264 | inline int CountedLoopNode::stride_con() const { return loopexit() ? loopexit()->stride_con() : 0; } |
duke@435 | 265 | inline bool CountedLoopNode::stride_is_con() const { return loopexit() && loopexit()->stride_is_con(); } |
duke@435 | 266 | inline Node *CountedLoopNode::limit() const { return loopexit() ? loopexit()->limit() : NULL; } |
duke@435 | 267 | inline Node *CountedLoopNode::incr() const { return loopexit() ? loopexit()->incr() : NULL; } |
duke@435 | 268 | inline Node *CountedLoopNode::phi() const { return loopexit() ? loopexit()->phi() : NULL; } |
duke@435 | 269 | |
duke@435 | 270 | |
duke@435 | 271 | // -----------------------------IdealLoopTree---------------------------------- |
duke@435 | 272 | class IdealLoopTree : public ResourceObj { |
duke@435 | 273 | public: |
duke@435 | 274 | IdealLoopTree *_parent; // Parent in loop tree |
duke@435 | 275 | IdealLoopTree *_next; // Next sibling in loop tree |
duke@435 | 276 | IdealLoopTree *_child; // First child in loop tree |
duke@435 | 277 | |
duke@435 | 278 | // The head-tail backedge defines the loop. |
duke@435 | 279 | // If tail is NULL then this loop has multiple backedges as part of the |
duke@435 | 280 | // same loop. During cleanup I'll peel off the multiple backedges; merge |
duke@435 | 281 | // them at the loop bottom and flow 1 real backedge into the loop. |
duke@435 | 282 | Node *_head; // Head of loop |
duke@435 | 283 | Node *_tail; // Tail of loop |
duke@435 | 284 | inline Node *tail(); // Handle lazy update of _tail field |
duke@435 | 285 | PhaseIdealLoop* _phase; |
duke@435 | 286 | |
duke@435 | 287 | Node_List _body; // Loop body for inner loops |
duke@435 | 288 | |
duke@435 | 289 | uint8 _nest; // Nesting depth |
duke@435 | 290 | uint8 _irreducible:1, // True if irreducible |
duke@435 | 291 | _has_call:1, // True if has call safepoint |
duke@435 | 292 | _has_sfpt:1, // True if has non-call safepoint |
duke@435 | 293 | _rce_candidate:1; // True if candidate for range check elimination |
duke@435 | 294 | |
kvn@474 | 295 | Node_List* _required_safept; // A inner loop cannot delete these safepts; |
kvn@474 | 296 | bool _allow_optimizations; // Allow loop optimizations |
duke@435 | 297 | |
duke@435 | 298 | IdealLoopTree( PhaseIdealLoop* phase, Node *head, Node *tail ) |
duke@435 | 299 | : _parent(0), _next(0), _child(0), |
duke@435 | 300 | _head(head), _tail(tail), |
duke@435 | 301 | _phase(phase), |
duke@435 | 302 | _required_safept(NULL), |
kvn@474 | 303 | _allow_optimizations(true), |
duke@435 | 304 | _nest(0), _irreducible(0), _has_call(0), _has_sfpt(0), _rce_candidate(0) |
duke@435 | 305 | { } |
duke@435 | 306 | |
duke@435 | 307 | // Is 'l' a member of 'this'? |
duke@435 | 308 | int is_member( const IdealLoopTree *l ) const; // Test for nested membership |
duke@435 | 309 | |
duke@435 | 310 | // Set loop nesting depth. Accumulate has_call bits. |
duke@435 | 311 | int set_nest( uint depth ); |
duke@435 | 312 | |
duke@435 | 313 | // Split out multiple fall-in edges from the loop header. Move them to a |
duke@435 | 314 | // private RegionNode before the loop. This becomes the loop landing pad. |
duke@435 | 315 | void split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt ); |
duke@435 | 316 | |
duke@435 | 317 | // Split out the outermost loop from this shared header. |
duke@435 | 318 | void split_outer_loop( PhaseIdealLoop *phase ); |
duke@435 | 319 | |
duke@435 | 320 | // Merge all the backedges from the shared header into a private Region. |
duke@435 | 321 | // Feed that region as the one backedge to this loop. |
duke@435 | 322 | void merge_many_backedges( PhaseIdealLoop *phase ); |
duke@435 | 323 | |
duke@435 | 324 | // Split shared headers and insert loop landing pads. |
duke@435 | 325 | // Insert a LoopNode to replace the RegionNode. |
duke@435 | 326 | // Returns TRUE if loop tree is structurally changed. |
duke@435 | 327 | bool beautify_loops( PhaseIdealLoop *phase ); |
duke@435 | 328 | |
cfang@1607 | 329 | // Perform optimization to use the loop predicates for null checks and range checks. |
cfang@1607 | 330 | // Applies to any loop level (not just the innermost one) |
cfang@1607 | 331 | bool loop_predication( PhaseIdealLoop *phase); |
cfang@1607 | 332 | |
never@836 | 333 | // Perform iteration-splitting on inner loops. Split iterations to |
never@836 | 334 | // avoid range checks or one-shot null checks. Returns false if the |
never@836 | 335 | // current round of loop opts should stop. |
never@836 | 336 | bool iteration_split( PhaseIdealLoop *phase, Node_List &old_new ); |
duke@435 | 337 | |
never@836 | 338 | // Driver for various flavors of iteration splitting. Returns false |
never@836 | 339 | // if the current round of loop opts should stop. |
never@836 | 340 | bool iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new ); |
duke@435 | 341 | |
duke@435 | 342 | // Given dominators, try to find loops with calls that must always be |
duke@435 | 343 | // executed (call dominates loop tail). These loops do not need non-call |
duke@435 | 344 | // safepoints (ncsfpt). |
duke@435 | 345 | void check_safepts(VectorSet &visited, Node_List &stack); |
duke@435 | 346 | |
duke@435 | 347 | // Allpaths backwards scan from loop tail, terminating each path at first safepoint |
duke@435 | 348 | // encountered. |
duke@435 | 349 | void allpaths_check_safepts(VectorSet &visited, Node_List &stack); |
duke@435 | 350 | |
duke@435 | 351 | // Convert to counted loops where possible |
duke@435 | 352 | void counted_loop( PhaseIdealLoop *phase ); |
duke@435 | 353 | |
duke@435 | 354 | // Check for Node being a loop-breaking test |
duke@435 | 355 | Node *is_loop_exit(Node *iff) const; |
duke@435 | 356 | |
duke@435 | 357 | // Returns true if ctrl is executed on every complete iteration |
duke@435 | 358 | bool dominates_backedge(Node* ctrl); |
duke@435 | 359 | |
duke@435 | 360 | // Remove simplistic dead code from loop body |
duke@435 | 361 | void DCE_loop_body(); |
duke@435 | 362 | |
duke@435 | 363 | // Look for loop-exit tests with my 50/50 guesses from the Parsing stage. |
duke@435 | 364 | // Replace with a 1-in-10 exit guess. |
duke@435 | 365 | void adjust_loop_exit_prob( PhaseIdealLoop *phase ); |
duke@435 | 366 | |
duke@435 | 367 | // Return TRUE or FALSE if the loop should never be RCE'd or aligned. |
duke@435 | 368 | // Useful for unrolling loops with NO array accesses. |
duke@435 | 369 | bool policy_peel_only( PhaseIdealLoop *phase ) const; |
duke@435 | 370 | |
duke@435 | 371 | // Return TRUE or FALSE if the loop should be unswitched -- clone |
duke@435 | 372 | // loop with an invariant test |
duke@435 | 373 | bool policy_unswitching( PhaseIdealLoop *phase ) const; |
duke@435 | 374 | |
duke@435 | 375 | // Micro-benchmark spamming. Remove empty loops. |
duke@435 | 376 | bool policy_do_remove_empty_loop( PhaseIdealLoop *phase ); |
duke@435 | 377 | |
duke@435 | 378 | // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can |
duke@435 | 379 | // make some loop-invariant test (usually a null-check) happen before the |
duke@435 | 380 | // loop. |
duke@435 | 381 | bool policy_peeling( PhaseIdealLoop *phase ) const; |
duke@435 | 382 | |
duke@435 | 383 | // Return TRUE or FALSE if the loop should be maximally unrolled. Stash any |
duke@435 | 384 | // known trip count in the counted loop node. |
duke@435 | 385 | bool policy_maximally_unroll( PhaseIdealLoop *phase ) const; |
duke@435 | 386 | |
duke@435 | 387 | // Return TRUE or FALSE if the loop should be unrolled or not. Unroll if |
duke@435 | 388 | // the loop is a CountedLoop and the body is small enough. |
duke@435 | 389 | bool policy_unroll( PhaseIdealLoop *phase ) const; |
duke@435 | 390 | |
duke@435 | 391 | // Return TRUE or FALSE if the loop should be range-check-eliminated. |
duke@435 | 392 | // Gather a list of IF tests that are dominated by iteration splitting; |
duke@435 | 393 | // also gather the end of the first split and the start of the 2nd split. |
duke@435 | 394 | bool policy_range_check( PhaseIdealLoop *phase ) const; |
duke@435 | 395 | |
duke@435 | 396 | // Return TRUE or FALSE if the loop should be cache-line aligned. |
duke@435 | 397 | // Gather the expression that does the alignment. Note that only |
twisti@1040 | 398 | // one array base can be aligned in a loop (unless the VM guarantees |
duke@435 | 399 | // mutual alignment). Note that if we vectorize short memory ops |
duke@435 | 400 | // into longer memory ops, we may want to increase alignment. |
duke@435 | 401 | bool policy_align( PhaseIdealLoop *phase ) const; |
duke@435 | 402 | |
cfang@1607 | 403 | // Return TRUE if "iff" is a range check. |
cfang@1607 | 404 | bool is_range_check_if(IfNode *iff, PhaseIdealLoop *phase, Invariance& invar) const; |
cfang@1607 | 405 | |
duke@435 | 406 | // Compute loop trip count from profile data |
duke@435 | 407 | void compute_profile_trip_cnt( PhaseIdealLoop *phase ); |
duke@435 | 408 | |
duke@435 | 409 | // Reassociate invariant expressions. |
duke@435 | 410 | void reassociate_invariants(PhaseIdealLoop *phase); |
duke@435 | 411 | // Reassociate invariant add and subtract expressions. |
duke@435 | 412 | Node* reassociate_add_sub(Node* n1, PhaseIdealLoop *phase); |
duke@435 | 413 | // Return nonzero index of invariant operand if invariant and variant |
twisti@1040 | 414 | // are combined with an Add or Sub. Helper for reassociate_invariants. |
duke@435 | 415 | int is_invariant_addition(Node* n, PhaseIdealLoop *phase); |
duke@435 | 416 | |
duke@435 | 417 | // Return true if n is invariant |
duke@435 | 418 | bool is_invariant(Node* n) const; |
duke@435 | 419 | |
duke@435 | 420 | // Put loop body on igvn work list |
duke@435 | 421 | void record_for_igvn(); |
duke@435 | 422 | |
duke@435 | 423 | bool is_loop() { return !_irreducible && _tail && !_tail->is_top(); } |
duke@435 | 424 | bool is_inner() { return is_loop() && _child == NULL; } |
duke@435 | 425 | bool is_counted() { return is_loop() && _head != NULL && _head->is_CountedLoop(); } |
duke@435 | 426 | |
duke@435 | 427 | #ifndef PRODUCT |
duke@435 | 428 | void dump_head( ) const; // Dump loop head only |
duke@435 | 429 | void dump() const; // Dump this loop recursively |
duke@435 | 430 | void verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const; |
duke@435 | 431 | #endif |
duke@435 | 432 | |
duke@435 | 433 | }; |
duke@435 | 434 | |
duke@435 | 435 | // -----------------------------PhaseIdealLoop--------------------------------- |
duke@435 | 436 | // Computes the mapping from Nodes to IdealLoopTrees. Organizes IdealLoopTrees into a |
duke@435 | 437 | // loop tree. Drives the loop-based transformations on the ideal graph. |
duke@435 | 438 | class PhaseIdealLoop : public PhaseTransform { |
duke@435 | 439 | friend class IdealLoopTree; |
duke@435 | 440 | friend class SuperWord; |
duke@435 | 441 | // Pre-computed def-use info |
duke@435 | 442 | PhaseIterGVN &_igvn; |
duke@435 | 443 | |
duke@435 | 444 | // Head of loop tree |
duke@435 | 445 | IdealLoopTree *_ltree_root; |
duke@435 | 446 | |
duke@435 | 447 | // Array of pre-order numbers, plus post-visited bit. |
duke@435 | 448 | // ZERO for not pre-visited. EVEN for pre-visited but not post-visited. |
duke@435 | 449 | // ODD for post-visited. Other bits are the pre-order number. |
duke@435 | 450 | uint *_preorders; |
duke@435 | 451 | uint _max_preorder; |
duke@435 | 452 | |
never@1356 | 453 | const PhaseIdealLoop* _verify_me; |
never@1356 | 454 | bool _verify_only; |
never@1356 | 455 | |
duke@435 | 456 | // Allocate _preorders[] array |
duke@435 | 457 | void allocate_preorders() { |
duke@435 | 458 | _max_preorder = C->unique()+8; |
duke@435 | 459 | _preorders = NEW_RESOURCE_ARRAY(uint, _max_preorder); |
duke@435 | 460 | memset(_preorders, 0, sizeof(uint) * _max_preorder); |
duke@435 | 461 | } |
duke@435 | 462 | |
duke@435 | 463 | // Allocate _preorders[] array |
duke@435 | 464 | void reallocate_preorders() { |
duke@435 | 465 | if ( _max_preorder < C->unique() ) { |
duke@435 | 466 | _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, C->unique()); |
duke@435 | 467 | _max_preorder = C->unique(); |
duke@435 | 468 | } |
duke@435 | 469 | memset(_preorders, 0, sizeof(uint) * _max_preorder); |
duke@435 | 470 | } |
duke@435 | 471 | |
duke@435 | 472 | // Check to grow _preorders[] array for the case when build_loop_tree_impl() |
duke@435 | 473 | // adds new nodes. |
duke@435 | 474 | void check_grow_preorders( ) { |
duke@435 | 475 | if ( _max_preorder < C->unique() ) { |
duke@435 | 476 | uint newsize = _max_preorder<<1; // double size of array |
duke@435 | 477 | _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, newsize); |
duke@435 | 478 | memset(&_preorders[_max_preorder],0,sizeof(uint)*(newsize-_max_preorder)); |
duke@435 | 479 | _max_preorder = newsize; |
duke@435 | 480 | } |
duke@435 | 481 | } |
duke@435 | 482 | // Check for pre-visited. Zero for NOT visited; non-zero for visited. |
duke@435 | 483 | int is_visited( Node *n ) const { return _preorders[n->_idx]; } |
duke@435 | 484 | // Pre-order numbers are written to the Nodes array as low-bit-set values. |
duke@435 | 485 | void set_preorder_visited( Node *n, int pre_order ) { |
duke@435 | 486 | assert( !is_visited( n ), "already set" ); |
duke@435 | 487 | _preorders[n->_idx] = (pre_order<<1); |
duke@435 | 488 | }; |
duke@435 | 489 | // Return pre-order number. |
duke@435 | 490 | int get_preorder( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]>>1; } |
duke@435 | 491 | |
duke@435 | 492 | // Check for being post-visited. |
duke@435 | 493 | // Should be previsited already (checked with assert(is_visited(n))). |
duke@435 | 494 | int is_postvisited( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]&1; } |
duke@435 | 495 | |
duke@435 | 496 | // Mark as post visited |
duke@435 | 497 | void set_postvisited( Node *n ) { assert( !is_postvisited( n ), "" ); _preorders[n->_idx] |= 1; } |
duke@435 | 498 | |
duke@435 | 499 | // Set/get control node out. Set lower bit to distinguish from IdealLoopTree |
duke@435 | 500 | // Returns true if "n" is a data node, false if it's a control node. |
duke@435 | 501 | bool has_ctrl( Node *n ) const { return ((intptr_t)_nodes[n->_idx]) & 1; } |
duke@435 | 502 | |
duke@435 | 503 | // clear out dead code after build_loop_late |
duke@435 | 504 | Node_List _deadlist; |
duke@435 | 505 | |
duke@435 | 506 | // Support for faster execution of get_late_ctrl()/dom_lca() |
duke@435 | 507 | // when a node has many uses and dominator depth is deep. |
duke@435 | 508 | Node_Array _dom_lca_tags; |
duke@435 | 509 | void init_dom_lca_tags(); |
duke@435 | 510 | void clear_dom_lca_tags(); |
never@1356 | 511 | |
never@1356 | 512 | // Helper for debugging bad dominance relationships |
never@1356 | 513 | bool verify_dominance(Node* n, Node* use, Node* LCA, Node* early); |
never@1356 | 514 | |
never@1356 | 515 | Node* compute_lca_of_uses(Node* n, Node* early, bool verify = false); |
never@1356 | 516 | |
duke@435 | 517 | // Inline wrapper for frequent cases: |
duke@435 | 518 | // 1) only one use |
duke@435 | 519 | // 2) a use is the same as the current LCA passed as 'n1' |
duke@435 | 520 | Node *dom_lca_for_get_late_ctrl( Node *lca, Node *n, Node *tag ) { |
duke@435 | 521 | assert( n->is_CFG(), "" ); |
duke@435 | 522 | // Fast-path NULL lca |
duke@435 | 523 | if( lca != NULL && lca != n ) { |
duke@435 | 524 | assert( lca->is_CFG(), "" ); |
duke@435 | 525 | // find LCA of all uses |
duke@435 | 526 | n = dom_lca_for_get_late_ctrl_internal( lca, n, tag ); |
duke@435 | 527 | } |
duke@435 | 528 | return find_non_split_ctrl(n); |
duke@435 | 529 | } |
duke@435 | 530 | Node *dom_lca_for_get_late_ctrl_internal( Node *lca, Node *n, Node *tag ); |
never@1356 | 531 | |
duke@435 | 532 | // Helper function for directing control inputs away from CFG split |
duke@435 | 533 | // points. |
duke@435 | 534 | Node *find_non_split_ctrl( Node *ctrl ) const { |
duke@435 | 535 | if (ctrl != NULL) { |
duke@435 | 536 | if (ctrl->is_MultiBranch()) { |
duke@435 | 537 | ctrl = ctrl->in(0); |
duke@435 | 538 | } |
duke@435 | 539 | assert(ctrl->is_CFG(), "CFG"); |
duke@435 | 540 | } |
duke@435 | 541 | return ctrl; |
duke@435 | 542 | } |
duke@435 | 543 | |
duke@435 | 544 | public: |
duke@435 | 545 | bool has_node( Node* n ) const { return _nodes[n->_idx] != NULL; } |
duke@435 | 546 | // check if transform created new nodes that need _ctrl recorded |
duke@435 | 547 | Node *get_late_ctrl( Node *n, Node *early ); |
duke@435 | 548 | Node *get_early_ctrl( Node *n ); |
duke@435 | 549 | void set_early_ctrl( Node *n ); |
duke@435 | 550 | void set_subtree_ctrl( Node *root ); |
duke@435 | 551 | void set_ctrl( Node *n, Node *ctrl ) { |
duke@435 | 552 | assert( !has_node(n) || has_ctrl(n), "" ); |
duke@435 | 553 | assert( ctrl->in(0), "cannot set dead control node" ); |
duke@435 | 554 | assert( ctrl == find_non_split_ctrl(ctrl), "must set legal crtl" ); |
duke@435 | 555 | _nodes.map( n->_idx, (Node*)((intptr_t)ctrl + 1) ); |
duke@435 | 556 | } |
duke@435 | 557 | // Set control and update loop membership |
duke@435 | 558 | void set_ctrl_and_loop(Node* n, Node* ctrl) { |
duke@435 | 559 | IdealLoopTree* old_loop = get_loop(get_ctrl(n)); |
duke@435 | 560 | IdealLoopTree* new_loop = get_loop(ctrl); |
duke@435 | 561 | if (old_loop != new_loop) { |
duke@435 | 562 | if (old_loop->_child == NULL) old_loop->_body.yank(n); |
duke@435 | 563 | if (new_loop->_child == NULL) new_loop->_body.push(n); |
duke@435 | 564 | } |
duke@435 | 565 | set_ctrl(n, ctrl); |
duke@435 | 566 | } |
duke@435 | 567 | // Control nodes can be replaced or subsumed. During this pass they |
duke@435 | 568 | // get their replacement Node in slot 1. Instead of updating the block |
duke@435 | 569 | // location of all Nodes in the subsumed block, we lazily do it. As we |
duke@435 | 570 | // pull such a subsumed block out of the array, we write back the final |
duke@435 | 571 | // correct block. |
duke@435 | 572 | Node *get_ctrl( Node *i ) { |
duke@435 | 573 | assert(has_node(i), ""); |
duke@435 | 574 | Node *n = get_ctrl_no_update(i); |
duke@435 | 575 | _nodes.map( i->_idx, (Node*)((intptr_t)n + 1) ); |
duke@435 | 576 | assert(has_node(i) && has_ctrl(i), ""); |
duke@435 | 577 | assert(n == find_non_split_ctrl(n), "must return legal ctrl" ); |
duke@435 | 578 | return n; |
duke@435 | 579 | } |
cfang@1607 | 580 | // true if CFG node d dominates CFG node n |
cfang@1607 | 581 | bool is_dominator(Node *d, Node *n); |
cfang@1607 | 582 | // return get_ctrl for a data node and self(n) for a CFG node |
cfang@1607 | 583 | Node* ctrl_or_self(Node* n) { |
cfang@1607 | 584 | if (has_ctrl(n)) |
cfang@1607 | 585 | return get_ctrl(n); |
cfang@1607 | 586 | else { |
cfang@1607 | 587 | assert (n->is_CFG(), "must be a CFG node"); |
cfang@1607 | 588 | return n; |
cfang@1607 | 589 | } |
cfang@1607 | 590 | } |
duke@435 | 591 | |
duke@435 | 592 | private: |
duke@435 | 593 | Node *get_ctrl_no_update( Node *i ) const { |
duke@435 | 594 | assert( has_ctrl(i), "" ); |
duke@435 | 595 | Node *n = (Node*)(((intptr_t)_nodes[i->_idx]) & ~1); |
duke@435 | 596 | if (!n->in(0)) { |
duke@435 | 597 | // Skip dead CFG nodes |
duke@435 | 598 | do { |
duke@435 | 599 | n = (Node*)(((intptr_t)_nodes[n->_idx]) & ~1); |
duke@435 | 600 | } while (!n->in(0)); |
duke@435 | 601 | n = find_non_split_ctrl(n); |
duke@435 | 602 | } |
duke@435 | 603 | return n; |
duke@435 | 604 | } |
duke@435 | 605 | |
duke@435 | 606 | // Check for loop being set |
duke@435 | 607 | // "n" must be a control node. Returns true if "n" is known to be in a loop. |
duke@435 | 608 | bool has_loop( Node *n ) const { |
duke@435 | 609 | assert(!has_node(n) || !has_ctrl(n), ""); |
duke@435 | 610 | return has_node(n); |
duke@435 | 611 | } |
duke@435 | 612 | // Set loop |
duke@435 | 613 | void set_loop( Node *n, IdealLoopTree *loop ) { |
duke@435 | 614 | _nodes.map(n->_idx, (Node*)loop); |
duke@435 | 615 | } |
duke@435 | 616 | // Lazy-dazy update of 'get_ctrl' and 'idom_at' mechanisms. Replace |
duke@435 | 617 | // the 'old_node' with 'new_node'. Kill old-node. Add a reference |
duke@435 | 618 | // from old_node to new_node to support the lazy update. Reference |
cfang@1607 | 619 | // replaces loop reference, since that is not needed for dead node. |
duke@435 | 620 | public: |
duke@435 | 621 | void lazy_update( Node *old_node, Node *new_node ) { |
duke@435 | 622 | assert( old_node != new_node, "no cycles please" ); |
duke@435 | 623 | //old_node->set_req( 1, new_node /*NO DU INFO*/ ); |
duke@435 | 624 | // Nodes always have DU info now, so re-use the side array slot |
duke@435 | 625 | // for this node to provide the forwarding pointer. |
duke@435 | 626 | _nodes.map( old_node->_idx, (Node*)((intptr_t)new_node + 1) ); |
duke@435 | 627 | } |
duke@435 | 628 | void lazy_replace( Node *old_node, Node *new_node ) { |
duke@435 | 629 | _igvn.hash_delete(old_node); |
duke@435 | 630 | _igvn.subsume_node( old_node, new_node ); |
duke@435 | 631 | lazy_update( old_node, new_node ); |
duke@435 | 632 | } |
duke@435 | 633 | void lazy_replace_proj( Node *old_node, Node *new_node ) { |
duke@435 | 634 | assert( old_node->req() == 1, "use this for Projs" ); |
duke@435 | 635 | _igvn.hash_delete(old_node); // Must hash-delete before hacking edges |
duke@435 | 636 | old_node->add_req( NULL ); |
duke@435 | 637 | lazy_replace( old_node, new_node ); |
duke@435 | 638 | } |
duke@435 | 639 | |
duke@435 | 640 | private: |
duke@435 | 641 | |
duke@435 | 642 | // Place 'n' in some loop nest, where 'n' is a CFG node |
duke@435 | 643 | void build_loop_tree(); |
duke@435 | 644 | int build_loop_tree_impl( Node *n, int pre_order ); |
duke@435 | 645 | // Insert loop into the existing loop tree. 'innermost' is a leaf of the |
duke@435 | 646 | // loop tree, not the root. |
duke@435 | 647 | IdealLoopTree *sort( IdealLoopTree *loop, IdealLoopTree *innermost ); |
duke@435 | 648 | |
duke@435 | 649 | // Place Data nodes in some loop nest |
never@1356 | 650 | void build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ); |
never@1356 | 651 | void build_loop_late ( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ); |
never@1356 | 652 | void build_loop_late_post ( Node* n ); |
duke@435 | 653 | |
duke@435 | 654 | // Array of immediate dominance info for each CFG node indexed by node idx |
duke@435 | 655 | private: |
duke@435 | 656 | uint _idom_size; |
duke@435 | 657 | Node **_idom; // Array of immediate dominators |
duke@435 | 658 | uint *_dom_depth; // Used for fast LCA test |
duke@435 | 659 | GrowableArray<uint>* _dom_stk; // For recomputation of dom depth |
duke@435 | 660 | |
duke@435 | 661 | Node* idom_no_update(Node* d) const { |
duke@435 | 662 | assert(d->_idx < _idom_size, "oob"); |
duke@435 | 663 | Node* n = _idom[d->_idx]; |
duke@435 | 664 | assert(n != NULL,"Bad immediate dominator info."); |
duke@435 | 665 | while (n->in(0) == NULL) { // Skip dead CFG nodes |
duke@435 | 666 | //n = n->in(1); |
duke@435 | 667 | n = (Node*)(((intptr_t)_nodes[n->_idx]) & ~1); |
duke@435 | 668 | assert(n != NULL,"Bad immediate dominator info."); |
duke@435 | 669 | } |
duke@435 | 670 | return n; |
duke@435 | 671 | } |
duke@435 | 672 | Node *idom(Node* d) const { |
duke@435 | 673 | uint didx = d->_idx; |
duke@435 | 674 | Node *n = idom_no_update(d); |
duke@435 | 675 | _idom[didx] = n; // Lazily remove dead CFG nodes from table. |
duke@435 | 676 | return n; |
duke@435 | 677 | } |
duke@435 | 678 | uint dom_depth(Node* d) const { |
duke@435 | 679 | assert(d->_idx < _idom_size, ""); |
duke@435 | 680 | return _dom_depth[d->_idx]; |
duke@435 | 681 | } |
duke@435 | 682 | void set_idom(Node* d, Node* n, uint dom_depth); |
duke@435 | 683 | // Locally compute IDOM using dom_lca call |
duke@435 | 684 | Node *compute_idom( Node *region ) const; |
duke@435 | 685 | // Recompute dom_depth |
duke@435 | 686 | void recompute_dom_depth(); |
duke@435 | 687 | |
duke@435 | 688 | // Is safept not required by an outer loop? |
duke@435 | 689 | bool is_deleteable_safept(Node* sfpt); |
duke@435 | 690 | |
never@1356 | 691 | // Perform verification that the graph is valid. |
never@1356 | 692 | PhaseIdealLoop( PhaseIterGVN &igvn) : |
never@1356 | 693 | PhaseTransform(Ideal_Loop), |
never@1356 | 694 | _igvn(igvn), |
never@1356 | 695 | _dom_lca_tags(C->comp_arena()), |
never@1356 | 696 | _verify_me(NULL), |
never@1356 | 697 | _verify_only(true) { |
cfang@1607 | 698 | build_and_optimize(false, false); |
never@1356 | 699 | } |
never@1356 | 700 | |
never@1356 | 701 | // build the loop tree and perform any requested optimizations |
cfang@1607 | 702 | void build_and_optimize(bool do_split_if, bool do_loop_pred); |
never@1356 | 703 | |
duke@435 | 704 | public: |
duke@435 | 705 | // Dominators for the sea of nodes |
duke@435 | 706 | void Dominators(); |
duke@435 | 707 | Node *dom_lca( Node *n1, Node *n2 ) const { |
duke@435 | 708 | return find_non_split_ctrl(dom_lca_internal(n1, n2)); |
duke@435 | 709 | } |
duke@435 | 710 | Node *dom_lca_internal( Node *n1, Node *n2 ) const; |
duke@435 | 711 | |
duke@435 | 712 | // Compute the Ideal Node to Loop mapping |
cfang@1607 | 713 | PhaseIdealLoop( PhaseIterGVN &igvn, bool do_split_ifs, bool do_loop_pred) : |
never@1356 | 714 | PhaseTransform(Ideal_Loop), |
never@1356 | 715 | _igvn(igvn), |
never@1356 | 716 | _dom_lca_tags(C->comp_arena()), |
never@1356 | 717 | _verify_me(NULL), |
never@1356 | 718 | _verify_only(false) { |
cfang@1607 | 719 | build_and_optimize(do_split_ifs, do_loop_pred); |
never@1356 | 720 | } |
never@1356 | 721 | |
never@1356 | 722 | // Verify that verify_me made the same decisions as a fresh run. |
never@1356 | 723 | PhaseIdealLoop( PhaseIterGVN &igvn, const PhaseIdealLoop *verify_me) : |
never@1356 | 724 | PhaseTransform(Ideal_Loop), |
never@1356 | 725 | _igvn(igvn), |
never@1356 | 726 | _dom_lca_tags(C->comp_arena()), |
never@1356 | 727 | _verify_me(verify_me), |
never@1356 | 728 | _verify_only(false) { |
cfang@1607 | 729 | build_and_optimize(false, false); |
never@1356 | 730 | } |
never@1356 | 731 | |
never@1356 | 732 | // Build and verify the loop tree without modifying the graph. This |
never@1356 | 733 | // is useful to verify that all inputs properly dominate their uses. |
never@1356 | 734 | static void verify(PhaseIterGVN& igvn) { |
never@1356 | 735 | #ifdef ASSERT |
never@1356 | 736 | PhaseIdealLoop v(igvn); |
never@1356 | 737 | #endif |
never@1356 | 738 | } |
duke@435 | 739 | |
duke@435 | 740 | // True if the method has at least 1 irreducible loop |
duke@435 | 741 | bool _has_irreducible_loops; |
duke@435 | 742 | |
duke@435 | 743 | // Per-Node transform |
duke@435 | 744 | virtual Node *transform( Node *a_node ) { return 0; } |
duke@435 | 745 | |
duke@435 | 746 | Node *is_counted_loop( Node *x, IdealLoopTree *loop ); |
duke@435 | 747 | |
duke@435 | 748 | // Return a post-walked LoopNode |
duke@435 | 749 | IdealLoopTree *get_loop( Node *n ) const { |
duke@435 | 750 | // Dead nodes have no loop, so return the top level loop instead |
duke@435 | 751 | if (!has_node(n)) return _ltree_root; |
duke@435 | 752 | assert(!has_ctrl(n), ""); |
duke@435 | 753 | return (IdealLoopTree*)_nodes[n->_idx]; |
duke@435 | 754 | } |
duke@435 | 755 | |
duke@435 | 756 | // Is 'n' a (nested) member of 'loop'? |
duke@435 | 757 | int is_member( const IdealLoopTree *loop, Node *n ) const { |
duke@435 | 758 | return loop->is_member(get_loop(n)); } |
duke@435 | 759 | |
duke@435 | 760 | // This is the basic building block of the loop optimizations. It clones an |
duke@435 | 761 | // entire loop body. It makes an old_new loop body mapping; with this |
duke@435 | 762 | // mapping you can find the new-loop equivalent to an old-loop node. All |
duke@435 | 763 | // new-loop nodes are exactly equal to their old-loop counterparts, all |
duke@435 | 764 | // edges are the same. All exits from the old-loop now have a RegionNode |
duke@435 | 765 | // that merges the equivalent new-loop path. This is true even for the |
duke@435 | 766 | // normal "loop-exit" condition. All uses of loop-invariant old-loop values |
duke@435 | 767 | // now come from (one or more) Phis that merge their new-loop equivalents. |
duke@435 | 768 | // Parameter side_by_side_idom: |
duke@435 | 769 | // When side_by_size_idom is NULL, the dominator tree is constructed for |
duke@435 | 770 | // the clone loop to dominate the original. Used in construction of |
duke@435 | 771 | // pre-main-post loop sequence. |
duke@435 | 772 | // When nonnull, the clone and original are side-by-side, both are |
duke@435 | 773 | // dominated by the passed in side_by_side_idom node. Used in |
duke@435 | 774 | // construction of unswitched loops. |
duke@435 | 775 | void clone_loop( IdealLoopTree *loop, Node_List &old_new, int dom_depth, |
duke@435 | 776 | Node* side_by_side_idom = NULL); |
duke@435 | 777 | |
duke@435 | 778 | // If we got the effect of peeling, either by actually peeling or by |
duke@435 | 779 | // making a pre-loop which must execute at least once, we can remove |
duke@435 | 780 | // all loop-invariant dominated tests in the main body. |
duke@435 | 781 | void peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new ); |
duke@435 | 782 | |
duke@435 | 783 | // Generate code to do a loop peel for the given loop (and body). |
duke@435 | 784 | // old_new is a temp array. |
duke@435 | 785 | void do_peeling( IdealLoopTree *loop, Node_List &old_new ); |
duke@435 | 786 | |
duke@435 | 787 | // Add pre and post loops around the given loop. These loops are used |
duke@435 | 788 | // during RCE, unrolling and aligning loops. |
duke@435 | 789 | void insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only ); |
duke@435 | 790 | // If Node n lives in the back_ctrl block, we clone a private version of n |
duke@435 | 791 | // in preheader_ctrl block and return that, otherwise return n. |
duke@435 | 792 | Node *clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n ); |
duke@435 | 793 | |
duke@435 | 794 | // Take steps to maximally unroll the loop. Peel any odd iterations, then |
duke@435 | 795 | // unroll to do double iterations. The next round of major loop transforms |
duke@435 | 796 | // will repeat till the doubled loop body does all remaining iterations in 1 |
duke@435 | 797 | // pass. |
duke@435 | 798 | void do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new ); |
duke@435 | 799 | |
duke@435 | 800 | // Unroll the loop body one step - make each trip do 2 iterations. |
duke@435 | 801 | void do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip ); |
duke@435 | 802 | |
duke@435 | 803 | // Return true if exp is a constant times an induction var |
duke@435 | 804 | bool is_scaled_iv(Node* exp, Node* iv, int* p_scale); |
duke@435 | 805 | |
duke@435 | 806 | // Return true if exp is a scaled induction var plus (or minus) constant |
duke@435 | 807 | bool is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth = 0); |
duke@435 | 808 | |
cfang@1607 | 809 | // Return true if proj is for "proj->[region->..]call_uct" |
cfang@1607 | 810 | bool is_uncommon_trap_proj(ProjNode* proj, bool must_reason_predicate = false); |
cfang@1607 | 811 | // Return true for "if(test)-> proj -> ... |
cfang@1607 | 812 | // | |
cfang@1607 | 813 | // V |
cfang@1607 | 814 | // other_proj->[region->..]call_uct" |
cfang@1607 | 815 | bool is_uncommon_trap_if_pattern(ProjNode* proj, bool must_reason_predicate = false); |
cfang@1607 | 816 | // Create a new if above the uncommon_trap_if_pattern for the predicate to be promoted |
cfang@1607 | 817 | ProjNode* create_new_if_for_predicate(ProjNode* cont_proj); |
cfang@1607 | 818 | // Find a good location to insert a predicate |
cfang@1607 | 819 | ProjNode* find_predicate_insertion_point(Node* start_c); |
cfang@1607 | 820 | // Construct a range check for a predicate if |
cfang@1607 | 821 | BoolNode* rc_predicate(Node* ctrl, |
cfang@1607 | 822 | int scale, Node* offset, |
cfang@1607 | 823 | Node* init, Node* limit, Node* stride, |
cfang@1607 | 824 | Node* range); |
cfang@1607 | 825 | |
cfang@1607 | 826 | // Implementation of the loop predication to promote checks outside the loop |
cfang@1607 | 827 | bool loop_predication_impl(IdealLoopTree *loop); |
cfang@1607 | 828 | |
cfang@1607 | 829 | // Helper function to collect predicate for eliminating the useless ones |
cfang@1607 | 830 | void collect_potentially_useful_predicates(IdealLoopTree *loop, Unique_Node_List &predicate_opaque1); |
cfang@1607 | 831 | void eliminate_useless_predicates(); |
cfang@1607 | 832 | |
duke@435 | 833 | // Eliminate range-checks and other trip-counter vs loop-invariant tests. |
duke@435 | 834 | void do_range_check( IdealLoopTree *loop, Node_List &old_new ); |
duke@435 | 835 | |
duke@435 | 836 | // Create a slow version of the loop by cloning the loop |
duke@435 | 837 | // and inserting an if to select fast-slow versions. |
duke@435 | 838 | ProjNode* create_slow_version_of_loop(IdealLoopTree *loop, |
duke@435 | 839 | Node_List &old_new); |
duke@435 | 840 | |
duke@435 | 841 | // Clone loop with an invariant test (that does not exit) and |
duke@435 | 842 | // insert a clone of the test that selects which version to |
duke@435 | 843 | // execute. |
duke@435 | 844 | void do_unswitching (IdealLoopTree *loop, Node_List &old_new); |
duke@435 | 845 | |
duke@435 | 846 | // Find candidate "if" for unswitching |
duke@435 | 847 | IfNode* find_unswitching_candidate(const IdealLoopTree *loop) const; |
duke@435 | 848 | |
duke@435 | 849 | // Range Check Elimination uses this function! |
duke@435 | 850 | // Constrain the main loop iterations so the affine function: |
duke@435 | 851 | // scale_con * I + offset < limit |
duke@435 | 852 | // always holds true. That is, either increase the number of iterations in |
duke@435 | 853 | // the pre-loop or the post-loop until the condition holds true in the main |
duke@435 | 854 | // loop. Scale_con, offset and limit are all loop invariant. |
duke@435 | 855 | void add_constraint( int stride_con, int scale_con, Node *offset, Node *limit, Node *pre_ctrl, Node **pre_limit, Node **main_limit ); |
duke@435 | 856 | |
duke@435 | 857 | // Partially peel loop up through last_peel node. |
duke@435 | 858 | bool partial_peel( IdealLoopTree *loop, Node_List &old_new ); |
duke@435 | 859 | |
duke@435 | 860 | // Create a scheduled list of nodes control dependent on ctrl set. |
duke@435 | 861 | void scheduled_nodelist( IdealLoopTree *loop, VectorSet& ctrl, Node_List &sched ); |
duke@435 | 862 | // Has a use in the vector set |
duke@435 | 863 | bool has_use_in_set( Node* n, VectorSet& vset ); |
duke@435 | 864 | // Has use internal to the vector set (ie. not in a phi at the loop head) |
duke@435 | 865 | bool has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop ); |
duke@435 | 866 | // clone "n" for uses that are outside of loop |
duke@435 | 867 | void clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist ); |
duke@435 | 868 | // clone "n" for special uses that are in the not_peeled region |
duke@435 | 869 | void clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n, |
duke@435 | 870 | VectorSet& not_peel, Node_List& sink_list, Node_List& worklist ); |
duke@435 | 871 | // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist |
duke@435 | 872 | void insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp ); |
duke@435 | 873 | #ifdef ASSERT |
duke@435 | 874 | // Validate the loop partition sets: peel and not_peel |
duke@435 | 875 | bool is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, VectorSet& not_peel ); |
duke@435 | 876 | // Ensure that uses outside of loop are of the right form |
duke@435 | 877 | bool is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list, |
duke@435 | 878 | uint orig_exit_idx, uint clone_exit_idx); |
duke@435 | 879 | bool is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx); |
duke@435 | 880 | #endif |
duke@435 | 881 | |
duke@435 | 882 | // Returns nonzero constant stride if-node is a possible iv test (otherwise returns zero.) |
duke@435 | 883 | int stride_of_possible_iv( Node* iff ); |
duke@435 | 884 | bool is_possible_iv_test( Node* iff ) { return stride_of_possible_iv(iff) != 0; } |
duke@435 | 885 | // Return the (unique) control output node that's in the loop (if it exists.) |
duke@435 | 886 | Node* stay_in_loop( Node* n, IdealLoopTree *loop); |
duke@435 | 887 | // Insert a signed compare loop exit cloned from an unsigned compare. |
duke@435 | 888 | IfNode* insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop); |
duke@435 | 889 | void remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop); |
duke@435 | 890 | // Utility to register node "n" with PhaseIdealLoop |
duke@435 | 891 | void register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth); |
duke@435 | 892 | // Utility to create an if-projection |
duke@435 | 893 | ProjNode* proj_clone(ProjNode* p, IfNode* iff); |
duke@435 | 894 | // Force the iff control output to be the live_proj |
duke@435 | 895 | Node* short_circuit_if(IfNode* iff, ProjNode* live_proj); |
duke@435 | 896 | // Insert a region before an if projection |
duke@435 | 897 | RegionNode* insert_region_before_proj(ProjNode* proj); |
duke@435 | 898 | // Insert a new if before an if projection |
duke@435 | 899 | ProjNode* insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj); |
duke@435 | 900 | |
duke@435 | 901 | // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps. |
duke@435 | 902 | // "Nearly" because all Nodes have been cloned from the original in the loop, |
duke@435 | 903 | // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs |
duke@435 | 904 | // through the Phi recursively, and return a Bool. |
duke@435 | 905 | BoolNode *clone_iff( PhiNode *phi, IdealLoopTree *loop ); |
duke@435 | 906 | CmpNode *clone_bool( PhiNode *phi, IdealLoopTree *loop ); |
duke@435 | 907 | |
duke@435 | 908 | |
duke@435 | 909 | // Rework addressing expressions to get the most loop-invariant stuff |
duke@435 | 910 | // moved out. We'd like to do all associative operators, but it's especially |
duke@435 | 911 | // important (common) to do address expressions. |
duke@435 | 912 | Node *remix_address_expressions( Node *n ); |
duke@435 | 913 | |
duke@435 | 914 | // Attempt to use a conditional move instead of a phi/branch |
duke@435 | 915 | Node *conditional_move( Node *n ); |
duke@435 | 916 | |
duke@435 | 917 | // Reorganize offset computations to lower register pressure. |
duke@435 | 918 | // Mostly prevent loop-fallout uses of the pre-incremented trip counter |
duke@435 | 919 | // (which are then alive with the post-incremented trip counter |
duke@435 | 920 | // forcing an extra register move) |
duke@435 | 921 | void reorg_offsets( IdealLoopTree *loop ); |
duke@435 | 922 | |
duke@435 | 923 | // Check for aggressive application of 'split-if' optimization, |
duke@435 | 924 | // using basic block level info. |
duke@435 | 925 | void split_if_with_blocks ( VectorSet &visited, Node_Stack &nstack ); |
duke@435 | 926 | Node *split_if_with_blocks_pre ( Node *n ); |
duke@435 | 927 | void split_if_with_blocks_post( Node *n ); |
duke@435 | 928 | Node *has_local_phi_input( Node *n ); |
duke@435 | 929 | // Mark an IfNode as being dominated by a prior test, |
duke@435 | 930 | // without actually altering the CFG (and hence IDOM info). |
duke@435 | 931 | void dominated_by( Node *prevdom, Node *iff ); |
duke@435 | 932 | |
duke@435 | 933 | // Split Node 'n' through merge point |
duke@435 | 934 | Node *split_thru_region( Node *n, Node *region ); |
duke@435 | 935 | // Split Node 'n' through merge point if there is enough win. |
duke@435 | 936 | Node *split_thru_phi( Node *n, Node *region, int policy ); |
duke@435 | 937 | // Found an If getting its condition-code input from a Phi in the |
duke@435 | 938 | // same block. Split thru the Region. |
duke@435 | 939 | void do_split_if( Node *iff ); |
duke@435 | 940 | |
duke@435 | 941 | private: |
duke@435 | 942 | // Return a type based on condition control flow |
duke@435 | 943 | const TypeInt* filtered_type( Node *n, Node* n_ctrl); |
duke@435 | 944 | const TypeInt* filtered_type( Node *n ) { return filtered_type(n, NULL); } |
duke@435 | 945 | // Helpers for filtered type |
duke@435 | 946 | const TypeInt* filtered_type_from_dominators( Node* val, Node *val_ctrl); |
duke@435 | 947 | |
duke@435 | 948 | // Helper functions |
duke@435 | 949 | Node *spinup( Node *iff, Node *new_false, Node *new_true, Node *region, Node *phi, small_cache *cache ); |
duke@435 | 950 | Node *find_use_block( Node *use, Node *def, Node *old_false, Node *new_false, Node *old_true, Node *new_true ); |
duke@435 | 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 ); |
duke@435 | 952 | bool split_up( Node *n, Node *blk1, Node *blk2 ); |
duke@435 | 953 | void sink_use( Node *use, Node *post_loop ); |
duke@435 | 954 | Node *place_near_use( Node *useblock ) const; |
duke@435 | 955 | |
duke@435 | 956 | bool _created_loop_node; |
duke@435 | 957 | public: |
duke@435 | 958 | void set_created_loop_node() { _created_loop_node = true; } |
duke@435 | 959 | bool created_loop_node() { return _created_loop_node; } |
cfang@1607 | 960 | void register_new_node( Node *n, Node *blk ); |
duke@435 | 961 | |
duke@435 | 962 | #ifndef PRODUCT |
duke@435 | 963 | void dump( ) const; |
duke@435 | 964 | void dump( IdealLoopTree *loop, uint rpo_idx, Node_List &rpo_list ) const; |
duke@435 | 965 | void rpo( Node *start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list ) const; |
duke@435 | 966 | void verify() const; // Major slow :-) |
duke@435 | 967 | void verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const; |
duke@435 | 968 | IdealLoopTree *get_loop_idx(Node* n) const { |
duke@435 | 969 | // Dead nodes have no loop, so return the top level loop instead |
duke@435 | 970 | return _nodes[n->_idx] ? (IdealLoopTree*)_nodes[n->_idx] : _ltree_root; |
duke@435 | 971 | } |
duke@435 | 972 | // Print some stats |
duke@435 | 973 | static void print_statistics(); |
duke@435 | 974 | static int _loop_invokes; // Count of PhaseIdealLoop invokes |
duke@435 | 975 | static int _loop_work; // Sum of PhaseIdealLoop x _unique |
duke@435 | 976 | #endif |
duke@435 | 977 | }; |
duke@435 | 978 | |
duke@435 | 979 | inline Node* IdealLoopTree::tail() { |
duke@435 | 980 | // Handle lazy update of _tail field |
duke@435 | 981 | Node *n = _tail; |
duke@435 | 982 | //while( !n->in(0) ) // Skip dead CFG nodes |
duke@435 | 983 | //n = n->in(1); |
duke@435 | 984 | if (n->in(0) == NULL) |
duke@435 | 985 | n = _phase->get_ctrl(n); |
duke@435 | 986 | _tail = n; |
duke@435 | 987 | return n; |
duke@435 | 988 | } |
duke@435 | 989 | |
duke@435 | 990 | |
duke@435 | 991 | // Iterate over the loop tree using a preorder, left-to-right traversal. |
duke@435 | 992 | // |
duke@435 | 993 | // Example that visits all counted loops from within PhaseIdealLoop |
duke@435 | 994 | // |
duke@435 | 995 | // for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) { |
duke@435 | 996 | // IdealLoopTree* lpt = iter.current(); |
duke@435 | 997 | // if (!lpt->is_counted()) continue; |
duke@435 | 998 | // ... |
duke@435 | 999 | class LoopTreeIterator : public StackObj { |
duke@435 | 1000 | private: |
duke@435 | 1001 | IdealLoopTree* _root; |
duke@435 | 1002 | IdealLoopTree* _curnt; |
duke@435 | 1003 | |
duke@435 | 1004 | public: |
duke@435 | 1005 | LoopTreeIterator(IdealLoopTree* root) : _root(root), _curnt(root) {} |
duke@435 | 1006 | |
duke@435 | 1007 | bool done() { return _curnt == NULL; } // Finished iterating? |
duke@435 | 1008 | |
duke@435 | 1009 | void next(); // Advance to next loop tree |
duke@435 | 1010 | |
duke@435 | 1011 | IdealLoopTree* current() { return _curnt; } // Return current value of iterator. |
duke@435 | 1012 | }; |