1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/src/share/vm/opto/loopnode.hpp Wed Apr 27 01:25:04 2016 +0800
1.3 @@ -0,0 +1,1118 @@
1.4 +/*
1.5 + * Copyright (c) 1998, 2013, Oracle and/or its affiliates. All rights reserved.
1.6 + * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
1.7 + *
1.8 + * This code is free software; you can redistribute it and/or modify it
1.9 + * under the terms of the GNU General Public License version 2 only, as
1.10 + * published by the Free Software Foundation.
1.11 + *
1.12 + * This code is distributed in the hope that it will be useful, but WITHOUT
1.13 + * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
1.14 + * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
1.15 + * version 2 for more details (a copy is included in the LICENSE file that
1.16 + * accompanied this code).
1.17 + *
1.18 + * You should have received a copy of the GNU General Public License version
1.19 + * 2 along with this work; if not, write to the Free Software Foundation,
1.20 + * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
1.21 + *
1.22 + * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
1.23 + * or visit www.oracle.com if you need additional information or have any
1.24 + * questions.
1.25 + *
1.26 + */
1.27 +
1.28 +#ifndef SHARE_VM_OPTO_LOOPNODE_HPP
1.29 +#define SHARE_VM_OPTO_LOOPNODE_HPP
1.30 +
1.31 +#include "opto/cfgnode.hpp"
1.32 +#include "opto/multnode.hpp"
1.33 +#include "opto/phaseX.hpp"
1.34 +#include "opto/subnode.hpp"
1.35 +#include "opto/type.hpp"
1.36 +
1.37 +class CmpNode;
1.38 +class CountedLoopEndNode;
1.39 +class CountedLoopNode;
1.40 +class IdealLoopTree;
1.41 +class LoopNode;
1.42 +class Node;
1.43 +class PhaseIdealLoop;
1.44 +class VectorSet;
1.45 +class Invariance;
1.46 +struct small_cache;
1.47 +
1.48 +//
1.49 +// I D E A L I Z E D L O O P S
1.50 +//
1.51 +// Idealized loops are the set of loops I perform more interesting
1.52 +// transformations on, beyond simple hoisting.
1.53 +
1.54 +//------------------------------LoopNode---------------------------------------
1.55 +// Simple loop header. Fall in path on left, loop-back path on right.
1.56 +class LoopNode : public RegionNode {
1.57 + // Size is bigger to hold the flags. However, the flags do not change
1.58 + // the semantics so it does not appear in the hash & cmp functions.
1.59 + virtual uint size_of() const { return sizeof(*this); }
1.60 +protected:
1.61 + short _loop_flags;
1.62 + // Names for flag bitfields
1.63 + enum { Normal=0, Pre=1, Main=2, Post=3, PreMainPostFlagsMask=3,
1.64 + MainHasNoPreLoop=4,
1.65 + HasExactTripCount=8,
1.66 + InnerLoop=16,
1.67 + PartialPeelLoop=32,
1.68 + PartialPeelFailed=64 };
1.69 + char _unswitch_count;
1.70 + enum { _unswitch_max=3 };
1.71 +
1.72 +public:
1.73 + // Names for edge indices
1.74 + enum { Self=0, EntryControl, LoopBackControl };
1.75 +
1.76 + int is_inner_loop() const { return _loop_flags & InnerLoop; }
1.77 + void set_inner_loop() { _loop_flags |= InnerLoop; }
1.78 +
1.79 + int is_partial_peel_loop() const { return _loop_flags & PartialPeelLoop; }
1.80 + void set_partial_peel_loop() { _loop_flags |= PartialPeelLoop; }
1.81 + int partial_peel_has_failed() const { return _loop_flags & PartialPeelFailed; }
1.82 + void mark_partial_peel_failed() { _loop_flags |= PartialPeelFailed; }
1.83 +
1.84 + int unswitch_max() { return _unswitch_max; }
1.85 + int unswitch_count() { return _unswitch_count; }
1.86 + void set_unswitch_count(int val) {
1.87 + assert (val <= unswitch_max(), "too many unswitches");
1.88 + _unswitch_count = val;
1.89 + }
1.90 +
1.91 + LoopNode( Node *entry, Node *backedge ) : RegionNode(3), _loop_flags(0), _unswitch_count(0) {
1.92 + init_class_id(Class_Loop);
1.93 + init_req(EntryControl, entry);
1.94 + init_req(LoopBackControl, backedge);
1.95 + }
1.96 +
1.97 + virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
1.98 + virtual int Opcode() const;
1.99 + bool can_be_counted_loop(PhaseTransform* phase) const {
1.100 + return req() == 3 && in(0) != NULL &&
1.101 + in(1) != NULL && phase->type(in(1)) != Type::TOP &&
1.102 + in(2) != NULL && phase->type(in(2)) != Type::TOP;
1.103 + }
1.104 + bool is_valid_counted_loop() const;
1.105 +#ifndef PRODUCT
1.106 + virtual void dump_spec(outputStream *st) const;
1.107 +#endif
1.108 +};
1.109 +
1.110 +//------------------------------Counted Loops----------------------------------
1.111 +// Counted loops are all trip-counted loops, with exactly 1 trip-counter exit
1.112 +// path (and maybe some other exit paths). The trip-counter exit is always
1.113 +// last in the loop. The trip-counter have to stride by a constant;
1.114 +// the exit value is also loop invariant.
1.115 +
1.116 +// CountedLoopNodes and CountedLoopEndNodes come in matched pairs. The
1.117 +// CountedLoopNode has the incoming loop control and the loop-back-control
1.118 +// which is always the IfTrue before the matching CountedLoopEndNode. The
1.119 +// CountedLoopEndNode has an incoming control (possibly not the
1.120 +// CountedLoopNode if there is control flow in the loop), the post-increment
1.121 +// trip-counter value, and the limit. The trip-counter value is always of
1.122 +// the form (Op old-trip-counter stride). The old-trip-counter is produced
1.123 +// by a Phi connected to the CountedLoopNode. The stride is constant.
1.124 +// The Op is any commutable opcode, including Add, Mul, Xor. The
1.125 +// CountedLoopEndNode also takes in the loop-invariant limit value.
1.126 +
1.127 +// From a CountedLoopNode I can reach the matching CountedLoopEndNode via the
1.128 +// loop-back control. From CountedLoopEndNodes I can reach CountedLoopNodes
1.129 +// via the old-trip-counter from the Op node.
1.130 +
1.131 +//------------------------------CountedLoopNode--------------------------------
1.132 +// CountedLoopNodes head simple counted loops. CountedLoopNodes have as
1.133 +// inputs the incoming loop-start control and the loop-back control, so they
1.134 +// act like RegionNodes. They also take in the initial trip counter, the
1.135 +// loop-invariant stride and the loop-invariant limit value. CountedLoopNodes
1.136 +// produce a loop-body control and the trip counter value. Since
1.137 +// CountedLoopNodes behave like RegionNodes I still have a standard CFG model.
1.138 +
1.139 +class CountedLoopNode : public LoopNode {
1.140 + // Size is bigger to hold _main_idx. However, _main_idx does not change
1.141 + // the semantics so it does not appear in the hash & cmp functions.
1.142 + virtual uint size_of() const { return sizeof(*this); }
1.143 +
1.144 + // For Pre- and Post-loops during debugging ONLY, this holds the index of
1.145 + // the Main CountedLoop. Used to assert that we understand the graph shape.
1.146 + node_idx_t _main_idx;
1.147 +
1.148 + // Known trip count calculated by compute_exact_trip_count()
1.149 + uint _trip_count;
1.150 +
1.151 + // Expected trip count from profile data
1.152 + float _profile_trip_cnt;
1.153 +
1.154 + // Log2 of original loop bodies in unrolled loop
1.155 + int _unrolled_count_log2;
1.156 +
1.157 + // Node count prior to last unrolling - used to decide if
1.158 + // unroll,optimize,unroll,optimize,... is making progress
1.159 + int _node_count_before_unroll;
1.160 +
1.161 +public:
1.162 + CountedLoopNode( Node *entry, Node *backedge )
1.163 + : LoopNode(entry, backedge), _main_idx(0), _trip_count(max_juint),
1.164 + _profile_trip_cnt(COUNT_UNKNOWN), _unrolled_count_log2(0),
1.165 + _node_count_before_unroll(0) {
1.166 + init_class_id(Class_CountedLoop);
1.167 + // Initialize _trip_count to the largest possible value.
1.168 + // Will be reset (lower) if the loop's trip count is known.
1.169 + }
1.170 +
1.171 + virtual int Opcode() const;
1.172 + virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
1.173 +
1.174 + Node *init_control() const { return in(EntryControl); }
1.175 + Node *back_control() const { return in(LoopBackControl); }
1.176 + CountedLoopEndNode *loopexit() const;
1.177 + Node *init_trip() const;
1.178 + Node *stride() const;
1.179 + int stride_con() const;
1.180 + bool stride_is_con() const;
1.181 + Node *limit() const;
1.182 + Node *incr() const;
1.183 + Node *phi() const;
1.184 +
1.185 + // Match increment with optional truncation
1.186 + static Node* match_incr_with_optional_truncation(Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type);
1.187 +
1.188 + // A 'main' loop has a pre-loop and a post-loop. The 'main' loop
1.189 + // can run short a few iterations and may start a few iterations in.
1.190 + // It will be RCE'd and unrolled and aligned.
1.191 +
1.192 + // A following 'post' loop will run any remaining iterations. Used
1.193 + // during Range Check Elimination, the 'post' loop will do any final
1.194 + // iterations with full checks. Also used by Loop Unrolling, where
1.195 + // the 'post' loop will do any epilog iterations needed. Basically,
1.196 + // a 'post' loop can not profitably be further unrolled or RCE'd.
1.197 +
1.198 + // A preceding 'pre' loop will run at least 1 iteration (to do peeling),
1.199 + // it may do under-flow checks for RCE and may do alignment iterations
1.200 + // so the following main loop 'knows' that it is striding down cache
1.201 + // lines.
1.202 +
1.203 + // A 'main' loop that is ONLY unrolled or peeled, never RCE'd or
1.204 + // Aligned, may be missing it's pre-loop.
1.205 + int is_normal_loop() const { return (_loop_flags&PreMainPostFlagsMask) == Normal; }
1.206 + int is_pre_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Pre; }
1.207 + int is_main_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Main; }
1.208 + int is_post_loop () const { return (_loop_flags&PreMainPostFlagsMask) == Post; }
1.209 + int is_main_no_pre_loop() const { return _loop_flags & MainHasNoPreLoop; }
1.210 + void set_main_no_pre_loop() { _loop_flags |= MainHasNoPreLoop; }
1.211 +
1.212 + int main_idx() const { return _main_idx; }
1.213 +
1.214 +
1.215 + void set_pre_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Pre ; _main_idx = main->_idx; }
1.216 + void set_main_loop ( ) { assert(is_normal_loop(),""); _loop_flags |= Main; }
1.217 + void set_post_loop (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Post; _main_idx = main->_idx; }
1.218 + void set_normal_loop( ) { _loop_flags &= ~PreMainPostFlagsMask; }
1.219 +
1.220 + void set_trip_count(uint tc) { _trip_count = tc; }
1.221 + uint trip_count() { return _trip_count; }
1.222 +
1.223 + bool has_exact_trip_count() const { return (_loop_flags & HasExactTripCount) != 0; }
1.224 + void set_exact_trip_count(uint tc) {
1.225 + _trip_count = tc;
1.226 + _loop_flags |= HasExactTripCount;
1.227 + }
1.228 + void set_nonexact_trip_count() {
1.229 + _loop_flags &= ~HasExactTripCount;
1.230 + }
1.231 +
1.232 + void set_profile_trip_cnt(float ptc) { _profile_trip_cnt = ptc; }
1.233 + float profile_trip_cnt() { return _profile_trip_cnt; }
1.234 +
1.235 + void double_unrolled_count() { _unrolled_count_log2++; }
1.236 + int unrolled_count() { return 1 << MIN2(_unrolled_count_log2, BitsPerInt-3); }
1.237 +
1.238 + void set_node_count_before_unroll(int ct) { _node_count_before_unroll = ct; }
1.239 + int node_count_before_unroll() { return _node_count_before_unroll; }
1.240 +
1.241 +#ifndef PRODUCT
1.242 + virtual void dump_spec(outputStream *st) const;
1.243 +#endif
1.244 +};
1.245 +
1.246 +//------------------------------CountedLoopEndNode-----------------------------
1.247 +// CountedLoopEndNodes end simple trip counted loops. They act much like
1.248 +// IfNodes.
1.249 +class CountedLoopEndNode : public IfNode {
1.250 +public:
1.251 + enum { TestControl, TestValue };
1.252 +
1.253 + CountedLoopEndNode( Node *control, Node *test, float prob, float cnt )
1.254 + : IfNode( control, test, prob, cnt) {
1.255 + init_class_id(Class_CountedLoopEnd);
1.256 + }
1.257 + virtual int Opcode() const;
1.258 +
1.259 + Node *cmp_node() const { return (in(TestValue)->req() >=2) ? in(TestValue)->in(1) : NULL; }
1.260 + Node *incr() const { Node *tmp = cmp_node(); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; }
1.261 + Node *limit() const { Node *tmp = cmp_node(); return (tmp && tmp->req()==3) ? tmp->in(2) : NULL; }
1.262 + Node *stride() const { Node *tmp = incr (); return (tmp && tmp->req()==3) ? tmp->in(2) : NULL; }
1.263 + Node *phi() const { Node *tmp = incr (); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; }
1.264 + Node *init_trip() const { Node *tmp = phi (); return (tmp && tmp->req()==3) ? tmp->in(1) : NULL; }
1.265 + int stride_con() const;
1.266 + bool stride_is_con() const { Node *tmp = stride (); return (tmp != NULL && tmp->is_Con()); }
1.267 + BoolTest::mask test_trip() const { return in(TestValue)->as_Bool()->_test._test; }
1.268 + CountedLoopNode *loopnode() const {
1.269 + // The CountedLoopNode that goes with this CountedLoopEndNode may
1.270 + // have been optimized out by the IGVN so be cautious with the
1.271 + // pattern matching on the graph
1.272 + if (phi() == NULL) {
1.273 + return NULL;
1.274 + }
1.275 + Node *ln = phi()->in(0);
1.276 + if (ln->is_CountedLoop() && ln->as_CountedLoop()->loopexit() == this) {
1.277 + return (CountedLoopNode*)ln;
1.278 + }
1.279 + return NULL;
1.280 + }
1.281 +
1.282 +#ifndef PRODUCT
1.283 + virtual void dump_spec(outputStream *st) const;
1.284 +#endif
1.285 +};
1.286 +
1.287 +
1.288 +inline CountedLoopEndNode *CountedLoopNode::loopexit() const {
1.289 + Node *bc = back_control();
1.290 + if( bc == NULL ) return NULL;
1.291 + Node *le = bc->in(0);
1.292 + if( le->Opcode() != Op_CountedLoopEnd )
1.293 + return NULL;
1.294 + return (CountedLoopEndNode*)le;
1.295 +}
1.296 +inline Node *CountedLoopNode::init_trip() const { return loopexit() ? loopexit()->init_trip() : NULL; }
1.297 +inline Node *CountedLoopNode::stride() const { return loopexit() ? loopexit()->stride() : NULL; }
1.298 +inline int CountedLoopNode::stride_con() const { return loopexit() ? loopexit()->stride_con() : 0; }
1.299 +inline bool CountedLoopNode::stride_is_con() const { return loopexit() && loopexit()->stride_is_con(); }
1.300 +inline Node *CountedLoopNode::limit() const { return loopexit() ? loopexit()->limit() : NULL; }
1.301 +inline Node *CountedLoopNode::incr() const { return loopexit() ? loopexit()->incr() : NULL; }
1.302 +inline Node *CountedLoopNode::phi() const { return loopexit() ? loopexit()->phi() : NULL; }
1.303 +
1.304 +//------------------------------LoopLimitNode-----------------------------
1.305 +// Counted Loop limit node which represents exact final iterator value:
1.306 +// trip_count = (limit - init_trip + stride - 1)/stride
1.307 +// final_value= trip_count * stride + init_trip.
1.308 +// Use HW instructions to calculate it when it can overflow in integer.
1.309 +// Note, final_value should fit into integer since counted loop has
1.310 +// limit check: limit <= max_int-stride.
1.311 +class LoopLimitNode : public Node {
1.312 + enum { Init=1, Limit=2, Stride=3 };
1.313 + public:
1.314 + LoopLimitNode( Compile* C, Node *init, Node *limit, Node *stride ) : Node(0,init,limit,stride) {
1.315 + // Put it on the Macro nodes list to optimize during macro nodes expansion.
1.316 + init_flags(Flag_is_macro);
1.317 + C->add_macro_node(this);
1.318 + }
1.319 + virtual int Opcode() const;
1.320 + virtual const Type *bottom_type() const { return TypeInt::INT; }
1.321 + virtual uint ideal_reg() const { return Op_RegI; }
1.322 + virtual const Type *Value( PhaseTransform *phase ) const;
1.323 + virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
1.324 + virtual Node *Identity( PhaseTransform *phase );
1.325 +};
1.326 +
1.327 +// -----------------------------IdealLoopTree----------------------------------
1.328 +class IdealLoopTree : public ResourceObj {
1.329 +public:
1.330 + IdealLoopTree *_parent; // Parent in loop tree
1.331 + IdealLoopTree *_next; // Next sibling in loop tree
1.332 + IdealLoopTree *_child; // First child in loop tree
1.333 +
1.334 + // The head-tail backedge defines the loop.
1.335 + // If tail is NULL then this loop has multiple backedges as part of the
1.336 + // same loop. During cleanup I'll peel off the multiple backedges; merge
1.337 + // them at the loop bottom and flow 1 real backedge into the loop.
1.338 + Node *_head; // Head of loop
1.339 + Node *_tail; // Tail of loop
1.340 + inline Node *tail(); // Handle lazy update of _tail field
1.341 + PhaseIdealLoop* _phase;
1.342 +
1.343 + Node_List _body; // Loop body for inner loops
1.344 +
1.345 + uint8 _nest; // Nesting depth
1.346 + uint8 _irreducible:1, // True if irreducible
1.347 + _has_call:1, // True if has call safepoint
1.348 + _has_sfpt:1, // True if has non-call safepoint
1.349 + _rce_candidate:1; // True if candidate for range check elimination
1.350 +
1.351 + Node_List* _safepts; // List of safepoints in this loop
1.352 + Node_List* _required_safept; // A inner loop cannot delete these safepts;
1.353 + bool _allow_optimizations; // Allow loop optimizations
1.354 +
1.355 + IdealLoopTree( PhaseIdealLoop* phase, Node *head, Node *tail )
1.356 + : _parent(0), _next(0), _child(0),
1.357 + _head(head), _tail(tail),
1.358 + _phase(phase),
1.359 + _safepts(NULL),
1.360 + _required_safept(NULL),
1.361 + _allow_optimizations(true),
1.362 + _nest(0), _irreducible(0), _has_call(0), _has_sfpt(0), _rce_candidate(0)
1.363 + { }
1.364 +
1.365 + // Is 'l' a member of 'this'?
1.366 + int is_member( const IdealLoopTree *l ) const; // Test for nested membership
1.367 +
1.368 + // Set loop nesting depth. Accumulate has_call bits.
1.369 + int set_nest( uint depth );
1.370 +
1.371 + // Split out multiple fall-in edges from the loop header. Move them to a
1.372 + // private RegionNode before the loop. This becomes the loop landing pad.
1.373 + void split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt );
1.374 +
1.375 + // Split out the outermost loop from this shared header.
1.376 + void split_outer_loop( PhaseIdealLoop *phase );
1.377 +
1.378 + // Merge all the backedges from the shared header into a private Region.
1.379 + // Feed that region as the one backedge to this loop.
1.380 + void merge_many_backedges( PhaseIdealLoop *phase );
1.381 +
1.382 + // Split shared headers and insert loop landing pads.
1.383 + // Insert a LoopNode to replace the RegionNode.
1.384 + // Returns TRUE if loop tree is structurally changed.
1.385 + bool beautify_loops( PhaseIdealLoop *phase );
1.386 +
1.387 + // Perform optimization to use the loop predicates for null checks and range checks.
1.388 + // Applies to any loop level (not just the innermost one)
1.389 + bool loop_predication( PhaseIdealLoop *phase);
1.390 +
1.391 + // Perform iteration-splitting on inner loops. Split iterations to
1.392 + // avoid range checks or one-shot null checks. Returns false if the
1.393 + // current round of loop opts should stop.
1.394 + bool iteration_split( PhaseIdealLoop *phase, Node_List &old_new );
1.395 +
1.396 + // Driver for various flavors of iteration splitting. Returns false
1.397 + // if the current round of loop opts should stop.
1.398 + bool iteration_split_impl( PhaseIdealLoop *phase, Node_List &old_new );
1.399 +
1.400 + // Given dominators, try to find loops with calls that must always be
1.401 + // executed (call dominates loop tail). These loops do not need non-call
1.402 + // safepoints (ncsfpt).
1.403 + void check_safepts(VectorSet &visited, Node_List &stack);
1.404 +
1.405 + // Allpaths backwards scan from loop tail, terminating each path at first safepoint
1.406 + // encountered.
1.407 + void allpaths_check_safepts(VectorSet &visited, Node_List &stack);
1.408 +
1.409 + // Convert to counted loops where possible
1.410 + void counted_loop( PhaseIdealLoop *phase );
1.411 +
1.412 + // Check for Node being a loop-breaking test
1.413 + Node *is_loop_exit(Node *iff) const;
1.414 +
1.415 + // Returns true if ctrl is executed on every complete iteration
1.416 + bool dominates_backedge(Node* ctrl);
1.417 +
1.418 + // Remove simplistic dead code from loop body
1.419 + void DCE_loop_body();
1.420 +
1.421 + // Look for loop-exit tests with my 50/50 guesses from the Parsing stage.
1.422 + // Replace with a 1-in-10 exit guess.
1.423 + void adjust_loop_exit_prob( PhaseIdealLoop *phase );
1.424 +
1.425 + // Return TRUE or FALSE if the loop should never be RCE'd or aligned.
1.426 + // Useful for unrolling loops with NO array accesses.
1.427 + bool policy_peel_only( PhaseIdealLoop *phase ) const;
1.428 +
1.429 + // Return TRUE or FALSE if the loop should be unswitched -- clone
1.430 + // loop with an invariant test
1.431 + bool policy_unswitching( PhaseIdealLoop *phase ) const;
1.432 +
1.433 + // Micro-benchmark spamming. Remove empty loops.
1.434 + bool policy_do_remove_empty_loop( PhaseIdealLoop *phase );
1.435 +
1.436 + // Convert one iteration loop into normal code.
1.437 + bool policy_do_one_iteration_loop( PhaseIdealLoop *phase );
1.438 +
1.439 + // Return TRUE or FALSE if the loop should be peeled or not. Peel if we can
1.440 + // make some loop-invariant test (usually a null-check) happen before the
1.441 + // loop.
1.442 + bool policy_peeling( PhaseIdealLoop *phase ) const;
1.443 +
1.444 + // Return TRUE or FALSE if the loop should be maximally unrolled. Stash any
1.445 + // known trip count in the counted loop node.
1.446 + bool policy_maximally_unroll( PhaseIdealLoop *phase ) const;
1.447 +
1.448 + // Return TRUE or FALSE if the loop should be unrolled or not. Unroll if
1.449 + // the loop is a CountedLoop and the body is small enough.
1.450 + bool policy_unroll( PhaseIdealLoop *phase ) const;
1.451 +
1.452 + // Return TRUE or FALSE if the loop should be range-check-eliminated.
1.453 + // Gather a list of IF tests that are dominated by iteration splitting;
1.454 + // also gather the end of the first split and the start of the 2nd split.
1.455 + bool policy_range_check( PhaseIdealLoop *phase ) const;
1.456 +
1.457 + // Return TRUE or FALSE if the loop should be cache-line aligned.
1.458 + // Gather the expression that does the alignment. Note that only
1.459 + // one array base can be aligned in a loop (unless the VM guarantees
1.460 + // mutual alignment). Note that if we vectorize short memory ops
1.461 + // into longer memory ops, we may want to increase alignment.
1.462 + bool policy_align( PhaseIdealLoop *phase ) const;
1.463 +
1.464 + // Return TRUE if "iff" is a range check.
1.465 + bool is_range_check_if(IfNode *iff, PhaseIdealLoop *phase, Invariance& invar) const;
1.466 +
1.467 + // Compute loop exact trip count if possible
1.468 + void compute_exact_trip_count( PhaseIdealLoop *phase );
1.469 +
1.470 + // Compute loop trip count from profile data
1.471 + void compute_profile_trip_cnt( PhaseIdealLoop *phase );
1.472 +
1.473 + // Reassociate invariant expressions.
1.474 + void reassociate_invariants(PhaseIdealLoop *phase);
1.475 + // Reassociate invariant add and subtract expressions.
1.476 + Node* reassociate_add_sub(Node* n1, PhaseIdealLoop *phase);
1.477 + // Return nonzero index of invariant operand if invariant and variant
1.478 + // are combined with an Add or Sub. Helper for reassociate_invariants.
1.479 + int is_invariant_addition(Node* n, PhaseIdealLoop *phase);
1.480 +
1.481 + // Return true if n is invariant
1.482 + bool is_invariant(Node* n) const;
1.483 +
1.484 + // Put loop body on igvn work list
1.485 + void record_for_igvn();
1.486 +
1.487 + bool is_loop() { return !_irreducible && _tail && !_tail->is_top(); }
1.488 + bool is_inner() { return is_loop() && _child == NULL; }
1.489 + bool is_counted() { return is_loop() && _head != NULL && _head->is_CountedLoop(); }
1.490 +
1.491 +#ifndef PRODUCT
1.492 + void dump_head( ) const; // Dump loop head only
1.493 + void dump() const; // Dump this loop recursively
1.494 + void verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const;
1.495 +#endif
1.496 +
1.497 +};
1.498 +
1.499 +// -----------------------------PhaseIdealLoop---------------------------------
1.500 +// Computes the mapping from Nodes to IdealLoopTrees. Organizes IdealLoopTrees into a
1.501 +// loop tree. Drives the loop-based transformations on the ideal graph.
1.502 +class PhaseIdealLoop : public PhaseTransform {
1.503 + friend class IdealLoopTree;
1.504 + friend class SuperWord;
1.505 + // Pre-computed def-use info
1.506 + PhaseIterGVN &_igvn;
1.507 +
1.508 + // Head of loop tree
1.509 + IdealLoopTree *_ltree_root;
1.510 +
1.511 + // Array of pre-order numbers, plus post-visited bit.
1.512 + // ZERO for not pre-visited. EVEN for pre-visited but not post-visited.
1.513 + // ODD for post-visited. Other bits are the pre-order number.
1.514 + uint *_preorders;
1.515 + uint _max_preorder;
1.516 +
1.517 + const PhaseIdealLoop* _verify_me;
1.518 + bool _verify_only;
1.519 +
1.520 + // Allocate _preorders[] array
1.521 + void allocate_preorders() {
1.522 + _max_preorder = C->unique()+8;
1.523 + _preorders = NEW_RESOURCE_ARRAY(uint, _max_preorder);
1.524 + memset(_preorders, 0, sizeof(uint) * _max_preorder);
1.525 + }
1.526 +
1.527 + // Allocate _preorders[] array
1.528 + void reallocate_preorders() {
1.529 + if ( _max_preorder < C->unique() ) {
1.530 + _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, C->unique());
1.531 + _max_preorder = C->unique();
1.532 + }
1.533 + memset(_preorders, 0, sizeof(uint) * _max_preorder);
1.534 + }
1.535 +
1.536 + // Check to grow _preorders[] array for the case when build_loop_tree_impl()
1.537 + // adds new nodes.
1.538 + void check_grow_preorders( ) {
1.539 + if ( _max_preorder < C->unique() ) {
1.540 + uint newsize = _max_preorder<<1; // double size of array
1.541 + _preorders = REALLOC_RESOURCE_ARRAY(uint, _preorders, _max_preorder, newsize);
1.542 + memset(&_preorders[_max_preorder],0,sizeof(uint)*(newsize-_max_preorder));
1.543 + _max_preorder = newsize;
1.544 + }
1.545 + }
1.546 + // Check for pre-visited. Zero for NOT visited; non-zero for visited.
1.547 + int is_visited( Node *n ) const { return _preorders[n->_idx]; }
1.548 + // Pre-order numbers are written to the Nodes array as low-bit-set values.
1.549 + void set_preorder_visited( Node *n, int pre_order ) {
1.550 + assert( !is_visited( n ), "already set" );
1.551 + _preorders[n->_idx] = (pre_order<<1);
1.552 + };
1.553 + // Return pre-order number.
1.554 + int get_preorder( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]>>1; }
1.555 +
1.556 + // Check for being post-visited.
1.557 + // Should be previsited already (checked with assert(is_visited(n))).
1.558 + int is_postvisited( Node *n ) const { assert( is_visited(n), "" ); return _preorders[n->_idx]&1; }
1.559 +
1.560 + // Mark as post visited
1.561 + void set_postvisited( Node *n ) { assert( !is_postvisited( n ), "" ); _preorders[n->_idx] |= 1; }
1.562 +
1.563 + // Set/get control node out. Set lower bit to distinguish from IdealLoopTree
1.564 + // Returns true if "n" is a data node, false if it's a control node.
1.565 + bool has_ctrl( Node *n ) const { return ((intptr_t)_nodes[n->_idx]) & 1; }
1.566 +
1.567 + // clear out dead code after build_loop_late
1.568 + Node_List _deadlist;
1.569 +
1.570 + // Support for faster execution of get_late_ctrl()/dom_lca()
1.571 + // when a node has many uses and dominator depth is deep.
1.572 + Node_Array _dom_lca_tags;
1.573 + void init_dom_lca_tags();
1.574 + void clear_dom_lca_tags();
1.575 +
1.576 + // Helper for debugging bad dominance relationships
1.577 + bool verify_dominance(Node* n, Node* use, Node* LCA, Node* early);
1.578 +
1.579 + Node* compute_lca_of_uses(Node* n, Node* early, bool verify = false);
1.580 +
1.581 + // Inline wrapper for frequent cases:
1.582 + // 1) only one use
1.583 + // 2) a use is the same as the current LCA passed as 'n1'
1.584 + Node *dom_lca_for_get_late_ctrl( Node *lca, Node *n, Node *tag ) {
1.585 + assert( n->is_CFG(), "" );
1.586 + // Fast-path NULL lca
1.587 + if( lca != NULL && lca != n ) {
1.588 + assert( lca->is_CFG(), "" );
1.589 + // find LCA of all uses
1.590 + n = dom_lca_for_get_late_ctrl_internal( lca, n, tag );
1.591 + }
1.592 + return find_non_split_ctrl(n);
1.593 + }
1.594 + Node *dom_lca_for_get_late_ctrl_internal( Node *lca, Node *n, Node *tag );
1.595 +
1.596 + // Helper function for directing control inputs away from CFG split
1.597 + // points.
1.598 + Node *find_non_split_ctrl( Node *ctrl ) const {
1.599 + if (ctrl != NULL) {
1.600 + if (ctrl->is_MultiBranch()) {
1.601 + ctrl = ctrl->in(0);
1.602 + }
1.603 + assert(ctrl->is_CFG(), "CFG");
1.604 + }
1.605 + return ctrl;
1.606 + }
1.607 +
1.608 +public:
1.609 + bool has_node( Node* n ) const {
1.610 + guarantee(n != NULL, "No Node.");
1.611 + return _nodes[n->_idx] != NULL;
1.612 + }
1.613 + // check if transform created new nodes that need _ctrl recorded
1.614 + Node *get_late_ctrl( Node *n, Node *early );
1.615 + Node *get_early_ctrl( Node *n );
1.616 + Node *get_early_ctrl_for_expensive(Node *n, Node* earliest);
1.617 + void set_early_ctrl( Node *n );
1.618 + void set_subtree_ctrl( Node *root );
1.619 + void set_ctrl( Node *n, Node *ctrl ) {
1.620 + assert( !has_node(n) || has_ctrl(n), "" );
1.621 + assert( ctrl->in(0), "cannot set dead control node" );
1.622 + assert( ctrl == find_non_split_ctrl(ctrl), "must set legal crtl" );
1.623 + _nodes.map( n->_idx, (Node*)((intptr_t)ctrl + 1) );
1.624 + }
1.625 + // Set control and update loop membership
1.626 + void set_ctrl_and_loop(Node* n, Node* ctrl) {
1.627 + IdealLoopTree* old_loop = get_loop(get_ctrl(n));
1.628 + IdealLoopTree* new_loop = get_loop(ctrl);
1.629 + if (old_loop != new_loop) {
1.630 + if (old_loop->_child == NULL) old_loop->_body.yank(n);
1.631 + if (new_loop->_child == NULL) new_loop->_body.push(n);
1.632 + }
1.633 + set_ctrl(n, ctrl);
1.634 + }
1.635 + // Control nodes can be replaced or subsumed. During this pass they
1.636 + // get their replacement Node in slot 1. Instead of updating the block
1.637 + // location of all Nodes in the subsumed block, we lazily do it. As we
1.638 + // pull such a subsumed block out of the array, we write back the final
1.639 + // correct block.
1.640 + Node *get_ctrl( Node *i ) {
1.641 + assert(has_node(i), "");
1.642 + Node *n = get_ctrl_no_update(i);
1.643 + _nodes.map( i->_idx, (Node*)((intptr_t)n + 1) );
1.644 + assert(has_node(i) && has_ctrl(i), "");
1.645 + assert(n == find_non_split_ctrl(n), "must return legal ctrl" );
1.646 + return n;
1.647 + }
1.648 + // true if CFG node d dominates CFG node n
1.649 + bool is_dominator(Node *d, Node *n);
1.650 + // return get_ctrl for a data node and self(n) for a CFG node
1.651 + Node* ctrl_or_self(Node* n) {
1.652 + if (has_ctrl(n))
1.653 + return get_ctrl(n);
1.654 + else {
1.655 + assert (n->is_CFG(), "must be a CFG node");
1.656 + return n;
1.657 + }
1.658 + }
1.659 +
1.660 +private:
1.661 + Node *get_ctrl_no_update( Node *i ) const {
1.662 + assert( has_ctrl(i), "" );
1.663 + Node *n = (Node*)(((intptr_t)_nodes[i->_idx]) & ~1);
1.664 + if (!n->in(0)) {
1.665 + // Skip dead CFG nodes
1.666 + do {
1.667 + n = (Node*)(((intptr_t)_nodes[n->_idx]) & ~1);
1.668 + } while (!n->in(0));
1.669 + n = find_non_split_ctrl(n);
1.670 + }
1.671 + return n;
1.672 + }
1.673 +
1.674 + // Check for loop being set
1.675 + // "n" must be a control node. Returns true if "n" is known to be in a loop.
1.676 + bool has_loop( Node *n ) const {
1.677 + assert(!has_node(n) || !has_ctrl(n), "");
1.678 + return has_node(n);
1.679 + }
1.680 + // Set loop
1.681 + void set_loop( Node *n, IdealLoopTree *loop ) {
1.682 + _nodes.map(n->_idx, (Node*)loop);
1.683 + }
1.684 + // Lazy-dazy update of 'get_ctrl' and 'idom_at' mechanisms. Replace
1.685 + // the 'old_node' with 'new_node'. Kill old-node. Add a reference
1.686 + // from old_node to new_node to support the lazy update. Reference
1.687 + // replaces loop reference, since that is not needed for dead node.
1.688 +public:
1.689 + void lazy_update( Node *old_node, Node *new_node ) {
1.690 + assert( old_node != new_node, "no cycles please" );
1.691 + //old_node->set_req( 1, new_node /*NO DU INFO*/ );
1.692 + // Nodes always have DU info now, so re-use the side array slot
1.693 + // for this node to provide the forwarding pointer.
1.694 + _nodes.map( old_node->_idx, (Node*)((intptr_t)new_node + 1) );
1.695 + }
1.696 + void lazy_replace( Node *old_node, Node *new_node ) {
1.697 + _igvn.replace_node( old_node, new_node );
1.698 + lazy_update( old_node, new_node );
1.699 + }
1.700 + void lazy_replace_proj( Node *old_node, Node *new_node ) {
1.701 + assert( old_node->req() == 1, "use this for Projs" );
1.702 + _igvn.hash_delete(old_node); // Must hash-delete before hacking edges
1.703 + old_node->add_req( NULL );
1.704 + lazy_replace( old_node, new_node );
1.705 + }
1.706 +
1.707 +private:
1.708 +
1.709 + // Place 'n' in some loop nest, where 'n' is a CFG node
1.710 + void build_loop_tree();
1.711 + int build_loop_tree_impl( Node *n, int pre_order );
1.712 + // Insert loop into the existing loop tree. 'innermost' is a leaf of the
1.713 + // loop tree, not the root.
1.714 + IdealLoopTree *sort( IdealLoopTree *loop, IdealLoopTree *innermost );
1.715 +
1.716 + // Place Data nodes in some loop nest
1.717 + void build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack );
1.718 + void build_loop_late ( VectorSet &visited, Node_List &worklist, Node_Stack &nstack );
1.719 + void build_loop_late_post ( Node* n );
1.720 +
1.721 + // Array of immediate dominance info for each CFG node indexed by node idx
1.722 +private:
1.723 + uint _idom_size;
1.724 + Node **_idom; // Array of immediate dominators
1.725 + uint *_dom_depth; // Used for fast LCA test
1.726 + GrowableArray<uint>* _dom_stk; // For recomputation of dom depth
1.727 +
1.728 + Node* idom_no_update(Node* d) const {
1.729 + assert(d->_idx < _idom_size, "oob");
1.730 + Node* n = _idom[d->_idx];
1.731 + assert(n != NULL,"Bad immediate dominator info.");
1.732 + while (n->in(0) == NULL) { // Skip dead CFG nodes
1.733 + //n = n->in(1);
1.734 + n = (Node*)(((intptr_t)_nodes[n->_idx]) & ~1);
1.735 + assert(n != NULL,"Bad immediate dominator info.");
1.736 + }
1.737 + return n;
1.738 + }
1.739 + Node *idom(Node* d) const {
1.740 + uint didx = d->_idx;
1.741 + Node *n = idom_no_update(d);
1.742 + _idom[didx] = n; // Lazily remove dead CFG nodes from table.
1.743 + return n;
1.744 + }
1.745 + uint dom_depth(Node* d) const {
1.746 + guarantee(d != NULL, "Null dominator info.");
1.747 + guarantee(d->_idx < _idom_size, "");
1.748 + return _dom_depth[d->_idx];
1.749 + }
1.750 + void set_idom(Node* d, Node* n, uint dom_depth);
1.751 + // Locally compute IDOM using dom_lca call
1.752 + Node *compute_idom( Node *region ) const;
1.753 + // Recompute dom_depth
1.754 + void recompute_dom_depth();
1.755 +
1.756 + // Is safept not required by an outer loop?
1.757 + bool is_deleteable_safept(Node* sfpt);
1.758 +
1.759 + // Replace parallel induction variable (parallel to trip counter)
1.760 + void replace_parallel_iv(IdealLoopTree *loop);
1.761 +
1.762 + // Perform verification that the graph is valid.
1.763 + PhaseIdealLoop( PhaseIterGVN &igvn) :
1.764 + PhaseTransform(Ideal_Loop),
1.765 + _igvn(igvn),
1.766 + _dom_lca_tags(arena()), // Thread::resource_area
1.767 + _verify_me(NULL),
1.768 + _verify_only(true) {
1.769 + build_and_optimize(false, false);
1.770 + }
1.771 +
1.772 + // build the loop tree and perform any requested optimizations
1.773 + void build_and_optimize(bool do_split_if, bool skip_loop_opts);
1.774 +
1.775 +public:
1.776 + // Dominators for the sea of nodes
1.777 + void Dominators();
1.778 + Node *dom_lca( Node *n1, Node *n2 ) const {
1.779 + return find_non_split_ctrl(dom_lca_internal(n1, n2));
1.780 + }
1.781 + Node *dom_lca_internal( Node *n1, Node *n2 ) const;
1.782 +
1.783 + // Compute the Ideal Node to Loop mapping
1.784 + PhaseIdealLoop( PhaseIterGVN &igvn, bool do_split_ifs, bool skip_loop_opts = false) :
1.785 + PhaseTransform(Ideal_Loop),
1.786 + _igvn(igvn),
1.787 + _dom_lca_tags(arena()), // Thread::resource_area
1.788 + _verify_me(NULL),
1.789 + _verify_only(false) {
1.790 + build_and_optimize(do_split_ifs, skip_loop_opts);
1.791 + }
1.792 +
1.793 + // Verify that verify_me made the same decisions as a fresh run.
1.794 + PhaseIdealLoop( PhaseIterGVN &igvn, const PhaseIdealLoop *verify_me) :
1.795 + PhaseTransform(Ideal_Loop),
1.796 + _igvn(igvn),
1.797 + _dom_lca_tags(arena()), // Thread::resource_area
1.798 + _verify_me(verify_me),
1.799 + _verify_only(false) {
1.800 + build_and_optimize(false, false);
1.801 + }
1.802 +
1.803 + // Build and verify the loop tree without modifying the graph. This
1.804 + // is useful to verify that all inputs properly dominate their uses.
1.805 + static void verify(PhaseIterGVN& igvn) {
1.806 +#ifdef ASSERT
1.807 + PhaseIdealLoop v(igvn);
1.808 +#endif
1.809 + }
1.810 +
1.811 + // True if the method has at least 1 irreducible loop
1.812 + bool _has_irreducible_loops;
1.813 +
1.814 + // Per-Node transform
1.815 + virtual Node *transform( Node *a_node ) { return 0; }
1.816 +
1.817 + bool is_counted_loop( Node *x, IdealLoopTree *loop );
1.818 +
1.819 + Node* exact_limit( IdealLoopTree *loop );
1.820 +
1.821 + // Return a post-walked LoopNode
1.822 + IdealLoopTree *get_loop( Node *n ) const {
1.823 + // Dead nodes have no loop, so return the top level loop instead
1.824 + if (!has_node(n)) return _ltree_root;
1.825 + assert(!has_ctrl(n), "");
1.826 + return (IdealLoopTree*)_nodes[n->_idx];
1.827 + }
1.828 +
1.829 + // Is 'n' a (nested) member of 'loop'?
1.830 + int is_member( const IdealLoopTree *loop, Node *n ) const {
1.831 + return loop->is_member(get_loop(n)); }
1.832 +
1.833 + // This is the basic building block of the loop optimizations. It clones an
1.834 + // entire loop body. It makes an old_new loop body mapping; with this
1.835 + // mapping you can find the new-loop equivalent to an old-loop node. All
1.836 + // new-loop nodes are exactly equal to their old-loop counterparts, all
1.837 + // edges are the same. All exits from the old-loop now have a RegionNode
1.838 + // that merges the equivalent new-loop path. This is true even for the
1.839 + // normal "loop-exit" condition. All uses of loop-invariant old-loop values
1.840 + // now come from (one or more) Phis that merge their new-loop equivalents.
1.841 + // Parameter side_by_side_idom:
1.842 + // When side_by_size_idom is NULL, the dominator tree is constructed for
1.843 + // the clone loop to dominate the original. Used in construction of
1.844 + // pre-main-post loop sequence.
1.845 + // When nonnull, the clone and original are side-by-side, both are
1.846 + // dominated by the passed in side_by_side_idom node. Used in
1.847 + // construction of unswitched loops.
1.848 + void clone_loop( IdealLoopTree *loop, Node_List &old_new, int dom_depth,
1.849 + Node* side_by_side_idom = NULL);
1.850 +
1.851 + // If we got the effect of peeling, either by actually peeling or by
1.852 + // making a pre-loop which must execute at least once, we can remove
1.853 + // all loop-invariant dominated tests in the main body.
1.854 + void peeled_dom_test_elim( IdealLoopTree *loop, Node_List &old_new );
1.855 +
1.856 + // Generate code to do a loop peel for the given loop (and body).
1.857 + // old_new is a temp array.
1.858 + void do_peeling( IdealLoopTree *loop, Node_List &old_new );
1.859 +
1.860 + // Add pre and post loops around the given loop. These loops are used
1.861 + // during RCE, unrolling and aligning loops.
1.862 + void insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only );
1.863 + // If Node n lives in the back_ctrl block, we clone a private version of n
1.864 + // in preheader_ctrl block and return that, otherwise return n.
1.865 + Node *clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n, VectorSet &visited, Node_Stack &clones );
1.866 +
1.867 + // Take steps to maximally unroll the loop. Peel any odd iterations, then
1.868 + // unroll to do double iterations. The next round of major loop transforms
1.869 + // will repeat till the doubled loop body does all remaining iterations in 1
1.870 + // pass.
1.871 + void do_maximally_unroll( IdealLoopTree *loop, Node_List &old_new );
1.872 +
1.873 + // Unroll the loop body one step - make each trip do 2 iterations.
1.874 + void do_unroll( IdealLoopTree *loop, Node_List &old_new, bool adjust_min_trip );
1.875 +
1.876 + // Return true if exp is a constant times an induction var
1.877 + bool is_scaled_iv(Node* exp, Node* iv, int* p_scale);
1.878 +
1.879 + // Return true if exp is a scaled induction var plus (or minus) constant
1.880 + bool is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth = 0);
1.881 +
1.882 + // Create a new if above the uncommon_trap_if_pattern for the predicate to be promoted
1.883 + ProjNode* create_new_if_for_predicate(ProjNode* cont_proj, Node* new_entry,
1.884 + Deoptimization::DeoptReason reason);
1.885 + void register_control(Node* n, IdealLoopTree *loop, Node* pred);
1.886 +
1.887 + // Clone loop predicates to cloned loops (peeled, unswitched)
1.888 + static ProjNode* clone_predicate(ProjNode* predicate_proj, Node* new_entry,
1.889 + Deoptimization::DeoptReason reason,
1.890 + PhaseIdealLoop* loop_phase,
1.891 + PhaseIterGVN* igvn);
1.892 +
1.893 + static Node* clone_loop_predicates(Node* old_entry, Node* new_entry,
1.894 + bool clone_limit_check,
1.895 + PhaseIdealLoop* loop_phase,
1.896 + PhaseIterGVN* igvn);
1.897 + Node* clone_loop_predicates(Node* old_entry, Node* new_entry, bool clone_limit_check);
1.898 +
1.899 + static Node* skip_loop_predicates(Node* entry);
1.900 +
1.901 + // Find a good location to insert a predicate
1.902 + static ProjNode* find_predicate_insertion_point(Node* start_c, Deoptimization::DeoptReason reason);
1.903 + // Find a predicate
1.904 + static Node* find_predicate(Node* entry);
1.905 + // Construct a range check for a predicate if
1.906 + BoolNode* rc_predicate(IdealLoopTree *loop, Node* ctrl,
1.907 + int scale, Node* offset,
1.908 + Node* init, Node* limit, Node* stride,
1.909 + Node* range, bool upper);
1.910 +
1.911 + // Implementation of the loop predication to promote checks outside the loop
1.912 + bool loop_predication_impl(IdealLoopTree *loop);
1.913 +
1.914 + // Helper function to collect predicate for eliminating the useless ones
1.915 + void collect_potentially_useful_predicates(IdealLoopTree *loop, Unique_Node_List &predicate_opaque1);
1.916 + void eliminate_useless_predicates();
1.917 +
1.918 + // Change the control input of expensive nodes to allow commoning by
1.919 + // IGVN when it is guaranteed to not result in a more frequent
1.920 + // execution of the expensive node. Return true if progress.
1.921 + bool process_expensive_nodes();
1.922 +
1.923 + // Check whether node has become unreachable
1.924 + bool is_node_unreachable(Node *n) const {
1.925 + return !has_node(n) || n->is_unreachable(_igvn);
1.926 + }
1.927 +
1.928 + // Eliminate range-checks and other trip-counter vs loop-invariant tests.
1.929 + void do_range_check( IdealLoopTree *loop, Node_List &old_new );
1.930 +
1.931 + // Create a slow version of the loop by cloning the loop
1.932 + // and inserting an if to select fast-slow versions.
1.933 + ProjNode* create_slow_version_of_loop(IdealLoopTree *loop,
1.934 + Node_List &old_new);
1.935 +
1.936 + // Clone loop with an invariant test (that does not exit) and
1.937 + // insert a clone of the test that selects which version to
1.938 + // execute.
1.939 + void do_unswitching (IdealLoopTree *loop, Node_List &old_new);
1.940 +
1.941 + // Find candidate "if" for unswitching
1.942 + IfNode* find_unswitching_candidate(const IdealLoopTree *loop) const;
1.943 +
1.944 + // Range Check Elimination uses this function!
1.945 + // Constrain the main loop iterations so the affine function:
1.946 + // low_limit <= scale_con * I + offset < upper_limit
1.947 + // always holds true. That is, either increase the number of iterations in
1.948 + // the pre-loop or the post-loop until the condition holds true in the main
1.949 + // loop. Scale_con, offset and limit are all loop invariant.
1.950 + void add_constraint( int stride_con, int scale_con, Node *offset, Node *low_limit, Node *upper_limit, Node *pre_ctrl, Node **pre_limit, Node **main_limit );
1.951 + // Helper function for add_constraint().
1.952 + Node* adjust_limit( int stride_con, Node * scale, Node *offset, Node *rc_limit, Node *loop_limit, Node *pre_ctrl );
1.953 +
1.954 + // Partially peel loop up through last_peel node.
1.955 + bool partial_peel( IdealLoopTree *loop, Node_List &old_new );
1.956 +
1.957 + // Create a scheduled list of nodes control dependent on ctrl set.
1.958 + void scheduled_nodelist( IdealLoopTree *loop, VectorSet& ctrl, Node_List &sched );
1.959 + // Has a use in the vector set
1.960 + bool has_use_in_set( Node* n, VectorSet& vset );
1.961 + // Has use internal to the vector set (ie. not in a phi at the loop head)
1.962 + bool has_use_internal_to_set( Node* n, VectorSet& vset, IdealLoopTree *loop );
1.963 + // clone "n" for uses that are outside of loop
1.964 + int clone_for_use_outside_loop( IdealLoopTree *loop, Node* n, Node_List& worklist );
1.965 + // clone "n" for special uses that are in the not_peeled region
1.966 + void clone_for_special_use_inside_loop( IdealLoopTree *loop, Node* n,
1.967 + VectorSet& not_peel, Node_List& sink_list, Node_List& worklist );
1.968 + // Insert phi(lp_entry_val, back_edge_val) at use->in(idx) for loop lp if phi does not already exist
1.969 + void insert_phi_for_loop( Node* use, uint idx, Node* lp_entry_val, Node* back_edge_val, LoopNode* lp );
1.970 +#ifdef ASSERT
1.971 + // Validate the loop partition sets: peel and not_peel
1.972 + bool is_valid_loop_partition( IdealLoopTree *loop, VectorSet& peel, Node_List& peel_list, VectorSet& not_peel );
1.973 + // Ensure that uses outside of loop are of the right form
1.974 + bool is_valid_clone_loop_form( IdealLoopTree *loop, Node_List& peel_list,
1.975 + uint orig_exit_idx, uint clone_exit_idx);
1.976 + bool is_valid_clone_loop_exit_use( IdealLoopTree *loop, Node* use, uint exit_idx);
1.977 +#endif
1.978 +
1.979 + // Returns nonzero constant stride if-node is a possible iv test (otherwise returns zero.)
1.980 + int stride_of_possible_iv( Node* iff );
1.981 + bool is_possible_iv_test( Node* iff ) { return stride_of_possible_iv(iff) != 0; }
1.982 + // Return the (unique) control output node that's in the loop (if it exists.)
1.983 + Node* stay_in_loop( Node* n, IdealLoopTree *loop);
1.984 + // Insert a signed compare loop exit cloned from an unsigned compare.
1.985 + IfNode* insert_cmpi_loop_exit(IfNode* if_cmpu, IdealLoopTree *loop);
1.986 + void remove_cmpi_loop_exit(IfNode* if_cmp, IdealLoopTree *loop);
1.987 + // Utility to register node "n" with PhaseIdealLoop
1.988 + void register_node(Node* n, IdealLoopTree *loop, Node* pred, int ddepth);
1.989 + // Utility to create an if-projection
1.990 + ProjNode* proj_clone(ProjNode* p, IfNode* iff);
1.991 + // Force the iff control output to be the live_proj
1.992 + Node* short_circuit_if(IfNode* iff, ProjNode* live_proj);
1.993 + // Insert a region before an if projection
1.994 + RegionNode* insert_region_before_proj(ProjNode* proj);
1.995 + // Insert a new if before an if projection
1.996 + ProjNode* insert_if_before_proj(Node* left, bool Signed, BoolTest::mask relop, Node* right, ProjNode* proj);
1.997 +
1.998 + // Passed in a Phi merging (recursively) some nearly equivalent Bool/Cmps.
1.999 + // "Nearly" because all Nodes have been cloned from the original in the loop,
1.1000 + // but the fall-in edges to the Cmp are different. Clone bool/Cmp pairs
1.1001 + // through the Phi recursively, and return a Bool.
1.1002 + BoolNode *clone_iff( PhiNode *phi, IdealLoopTree *loop );
1.1003 + CmpNode *clone_bool( PhiNode *phi, IdealLoopTree *loop );
1.1004 +
1.1005 +
1.1006 + // Rework addressing expressions to get the most loop-invariant stuff
1.1007 + // moved out. We'd like to do all associative operators, but it's especially
1.1008 + // important (common) to do address expressions.
1.1009 + Node *remix_address_expressions( Node *n );
1.1010 +
1.1011 + // Attempt to use a conditional move instead of a phi/branch
1.1012 + Node *conditional_move( Node *n );
1.1013 +
1.1014 + // Reorganize offset computations to lower register pressure.
1.1015 + // Mostly prevent loop-fallout uses of the pre-incremented trip counter
1.1016 + // (which are then alive with the post-incremented trip counter
1.1017 + // forcing an extra register move)
1.1018 + void reorg_offsets( IdealLoopTree *loop );
1.1019 +
1.1020 + // Check for aggressive application of 'split-if' optimization,
1.1021 + // using basic block level info.
1.1022 + void split_if_with_blocks ( VectorSet &visited, Node_Stack &nstack );
1.1023 + Node *split_if_with_blocks_pre ( Node *n );
1.1024 + void split_if_with_blocks_post( Node *n );
1.1025 + Node *has_local_phi_input( Node *n );
1.1026 + // Mark an IfNode as being dominated by a prior test,
1.1027 + // without actually altering the CFG (and hence IDOM info).
1.1028 + void dominated_by( Node *prevdom, Node *iff, bool flip = false, bool exclude_loop_predicate = false );
1.1029 +
1.1030 + // Split Node 'n' through merge point
1.1031 + Node *split_thru_region( Node *n, Node *region );
1.1032 + // Split Node 'n' through merge point if there is enough win.
1.1033 + Node *split_thru_phi( Node *n, Node *region, int policy );
1.1034 + // Found an If getting its condition-code input from a Phi in the
1.1035 + // same block. Split thru the Region.
1.1036 + void do_split_if( Node *iff );
1.1037 +
1.1038 + // Conversion of fill/copy patterns into intrisic versions
1.1039 + bool do_intrinsify_fill();
1.1040 + bool intrinsify_fill(IdealLoopTree* lpt);
1.1041 + bool match_fill_loop(IdealLoopTree* lpt, Node*& store, Node*& store_value,
1.1042 + Node*& shift, Node*& offset);
1.1043 +
1.1044 +private:
1.1045 + // Return a type based on condition control flow
1.1046 + const TypeInt* filtered_type( Node *n, Node* n_ctrl);
1.1047 + const TypeInt* filtered_type( Node *n ) { return filtered_type(n, NULL); }
1.1048 + // Helpers for filtered type
1.1049 + const TypeInt* filtered_type_from_dominators( Node* val, Node *val_ctrl);
1.1050 +
1.1051 + // Helper functions
1.1052 + Node *spinup( Node *iff, Node *new_false, Node *new_true, Node *region, Node *phi, small_cache *cache );
1.1053 + Node *find_use_block( Node *use, Node *def, Node *old_false, Node *new_false, Node *old_true, Node *new_true );
1.1054 + 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 );
1.1055 + bool split_up( Node *n, Node *blk1, Node *blk2 );
1.1056 + void sink_use( Node *use, Node *post_loop );
1.1057 + Node *place_near_use( Node *useblock ) const;
1.1058 +
1.1059 + bool _created_loop_node;
1.1060 +public:
1.1061 + void set_created_loop_node() { _created_loop_node = true; }
1.1062 + bool created_loop_node() { return _created_loop_node; }
1.1063 + void register_new_node( Node *n, Node *blk );
1.1064 +
1.1065 +#ifdef ASSERT
1.1066 + void dump_bad_graph(const char* msg, Node* n, Node* early, Node* LCA);
1.1067 +#endif
1.1068 +
1.1069 +#ifndef PRODUCT
1.1070 + void dump( ) const;
1.1071 + void dump( IdealLoopTree *loop, uint rpo_idx, Node_List &rpo_list ) const;
1.1072 + void rpo( Node *start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list ) const;
1.1073 + void verify() const; // Major slow :-)
1.1074 + void verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const;
1.1075 + IdealLoopTree *get_loop_idx(Node* n) const {
1.1076 + // Dead nodes have no loop, so return the top level loop instead
1.1077 + return _nodes[n->_idx] ? (IdealLoopTree*)_nodes[n->_idx] : _ltree_root;
1.1078 + }
1.1079 + // Print some stats
1.1080 + static void print_statistics();
1.1081 + static int _loop_invokes; // Count of PhaseIdealLoop invokes
1.1082 + static int _loop_work; // Sum of PhaseIdealLoop x _unique
1.1083 +#endif
1.1084 +};
1.1085 +
1.1086 +inline Node* IdealLoopTree::tail() {
1.1087 +// Handle lazy update of _tail field
1.1088 + Node *n = _tail;
1.1089 + //while( !n->in(0) ) // Skip dead CFG nodes
1.1090 + //n = n->in(1);
1.1091 + if (n->in(0) == NULL)
1.1092 + n = _phase->get_ctrl(n);
1.1093 + _tail = n;
1.1094 + return n;
1.1095 +}
1.1096 +
1.1097 +
1.1098 +// Iterate over the loop tree using a preorder, left-to-right traversal.
1.1099 +//
1.1100 +// Example that visits all counted loops from within PhaseIdealLoop
1.1101 +//
1.1102 +// for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {
1.1103 +// IdealLoopTree* lpt = iter.current();
1.1104 +// if (!lpt->is_counted()) continue;
1.1105 +// ...
1.1106 +class LoopTreeIterator : public StackObj {
1.1107 +private:
1.1108 + IdealLoopTree* _root;
1.1109 + IdealLoopTree* _curnt;
1.1110 +
1.1111 +public:
1.1112 + LoopTreeIterator(IdealLoopTree* root) : _root(root), _curnt(root) {}
1.1113 +
1.1114 + bool done() { return _curnt == NULL; } // Finished iterating?
1.1115 +
1.1116 + void next(); // Advance to next loop tree
1.1117 +
1.1118 + IdealLoopTree* current() { return _curnt; } // Return current value of iterator.
1.1119 +};
1.1120 +
1.1121 +#endif // SHARE_VM_OPTO_LOOPNODE_HPP
