2
* Copyright (c) 1998, 2024, Oracle and/or its affiliates. All rights reserved.
3
* DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5
* This code is free software; you can redistribute it and/or modify it
6
* under the terms of the GNU General Public License version 2 only, as
7
* published by the Free Software Foundation.
9
* This code is distributed in the hope that it will be useful, but WITHOUT
10
* ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11
* FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12
* version 2 for more details (a copy is included in the LICENSE file that
13
* accompanied this code).
15
* You should have received a copy of the GNU General Public License version
16
* 2 along with this work; if not, write to the Free Software Foundation,
17
* Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
19
* Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20
* or visit www.oracle.com if you need additional information or have any
25
#include "precompiled.hpp"
26
#include "ci/ciMethodData.hpp"
27
#include "compiler/compileLog.hpp"
28
#include "gc/shared/barrierSet.hpp"
29
#include "gc/shared/c2/barrierSetC2.hpp"
30
#include "libadt/vectset.hpp"
31
#include "memory/allocation.inline.hpp"
32
#include "memory/resourceArea.hpp"
33
#include "opto/addnode.hpp"
34
#include "opto/arraycopynode.hpp"
35
#include "opto/callnode.hpp"
36
#include "opto/castnode.hpp"
37
#include "opto/connode.hpp"
38
#include "opto/convertnode.hpp"
39
#include "opto/divnode.hpp"
40
#include "opto/idealGraphPrinter.hpp"
41
#include "opto/loopnode.hpp"
42
#include "opto/movenode.hpp"
43
#include "opto/mulnode.hpp"
44
#include "opto/opaquenode.hpp"
45
#include "opto/predicates.hpp"
46
#include "opto/rootnode.hpp"
47
#include "opto/runtime.hpp"
48
#include "opto/vectorization.hpp"
49
#include "runtime/sharedRuntime.hpp"
50
#include "utilities/checkedCast.hpp"
51
#include "utilities/powerOfTwo.hpp"
53
//=============================================================================
54
//--------------------------is_cloop_ind_var-----------------------------------
55
// Determine if a node is a counted loop induction variable.
56
// NOTE: The method is declared in "node.hpp".
57
bool Node::is_cloop_ind_var() const {59
as_Phi()->region()->is_CountedLoop() &&
60
as_Phi()->region()->as_CountedLoop()->phi() == this);
63
//=============================================================================
64
//------------------------------dump_spec--------------------------------------
65
// Dump special per-node info
67
void LoopNode::dump_spec(outputStream *st) const {68
RegionNode::dump_spec(st);
69
if (is_inner_loop()) st->print( "inner " );
70
if (is_partial_peel_loop()) st->print( "partial_peel " );
71
if (partial_peel_has_failed()) st->print( "partial_peel_failed " );
75
//------------------------------is_valid_counted_loop-------------------------
76
bool LoopNode::is_valid_counted_loop(BasicType bt) const {77
if (is_BaseCountedLoop() && as_BaseCountedLoop()->bt() == bt) {78
BaseCountedLoopNode* l = as_BaseCountedLoop();
79
BaseCountedLoopEndNode* le = l->loopexit_or_null();
81
le->proj_out_or_null(1 /* true */) == l->in(LoopNode::LoopBackControl)) {83
Node* exit = le->proj_out_or_null(0 /* false */);
84
if (exit != nullptr && exit->Opcode() == Op_IfFalse &&
85
phi != nullptr && phi->is_Phi() &&
86
phi->in(LoopNode::LoopBackControl) == l->incr() &&
87
le->loopnode() == l && le->stride_is_con()) {95
//------------------------------get_early_ctrl---------------------------------
96
// Compute earliest legal control
97
Node *PhaseIdealLoop::get_early_ctrl( Node *n ) {98
assert( !n->is_Phi() && !n->is_CFG(), "this code only handles data nodes" );
101
if (n->in(0) && !n->is_expensive()) {103
if (!early->is_CFG()) // Might be a non-CFG multi-def
104
early = get_ctrl(early); // So treat input as a straight data input
107
early = get_ctrl(n->in(1));
110
uint e_d = dom_depth(early);
112
for (; i < n->req(); i++) {113
Node *cin = get_ctrl(n->in(i));
115
// Keep deepest dominator depth
116
uint c_d = dom_depth(cin);
117
if (c_d > e_d) { // Deeper guy?118
early = cin; // Keep deepest found so far
120
} else if (c_d == e_d && // Same depth?
121
early != cin) { // If not equal, must use slower algorithm122
// If same depth but not equal, one _must_ dominate the other
123
// and we want the deeper (i.e., dominated) guy.
127
n1 = idom(n1); // Walk up until break cycle
129
if (n1 == cin || // Walked early up to cin
131
break; // early is deeper; keep him
132
if (n2 == early || // Walked cin up to early
133
dom_depth(n1) < c_d) {134
early = cin; // cin is deeper; keep him
138
e_d = dom_depth(early); // Reset depth register cache
142
// Return earliest legal location
143
assert(early == find_non_split_ctrl(early), "unexpected early control");
145
if (n->is_expensive() && !_verify_only && !_verify_me) {146
assert(n->in(0), "should have control input");
147
early = get_early_ctrl_for_expensive(n, early);
153
//------------------------------get_early_ctrl_for_expensive---------------------------------
154
// Move node up the dominator tree as high as legal while still beneficial
155
Node *PhaseIdealLoop::get_early_ctrl_for_expensive(Node *n, Node* earliest) {156
assert(n->in(0) && n->is_expensive(), "expensive node with control input here");
157
assert(OptimizeExpensiveOps, "optimization off?");
159
Node* ctl = n->in(0);
160
assert(ctl->is_CFG(), "expensive input 0 must be cfg");
161
uint min_dom_depth = dom_depth(earliest);
163
if (!is_dominator(ctl, earliest) && !is_dominator(earliest, ctl)) {164
dump_bad_graph("Bad graph detected in get_early_ctrl_for_expensive", n, earliest, ctl);165
assert(false, "Bad graph detected in get_early_ctrl_for_expensive");
168
if (dom_depth(ctl) < min_dom_depth) {174
// Moving the node out of a loop on the projection of an If
175
// confuses Loop Predication. So, once we hit a loop in an If branch
176
// that doesn't branch to an UNC, we stop. The code that process
177
// expensive nodes will notice the loop and skip over it to try to
178
// move the node further up.
179
if (ctl->is_CountedLoop() && ctl->in(1) != nullptr && ctl->in(1)->in(0) != nullptr && ctl->in(1)->in(0)->is_If()) {180
if (!ctl->in(1)->as_Proj()->is_uncommon_trap_if_pattern()) {183
next = idom(ctl->in(1)->in(0));
184
} else if (ctl->is_Proj()) {185
// We only move it up along a projection if the projection is
186
// the single control projection for its parent: same code path,
187
// if it's a If with UNC or fallthrough of a call.
188
Node* parent_ctl = ctl->in(0);
189
if (parent_ctl == nullptr) {191
} else if (parent_ctl->is_CountedLoopEnd() && parent_ctl->as_CountedLoopEnd()->loopnode() != nullptr) {192
next = parent_ctl->as_CountedLoopEnd()->loopnode()->init_control();
193
} else if (parent_ctl->is_If()) {194
if (!ctl->as_Proj()->is_uncommon_trap_if_pattern()) {197
assert(idom(ctl) == parent_ctl, "strange");
198
next = idom(parent_ctl);
199
} else if (ctl->is_CatchProj()) {200
if (ctl->as_Proj()->_con != CatchProjNode::fall_through_index) {203
assert(parent_ctl->in(0)->in(0)->is_Call(), "strange graph");
204
next = parent_ctl->in(0)->in(0)->in(0);
206
// Check if parent control has a single projection (this
207
// control is the only possible successor of the parent
208
// control). If so, we can try to move the node above the
211
for (DUIterator_Fast imax, i = parent_ctl->fast_outs(imax); i < imax; i++) {212
Node *p = parent_ctl->fast_out(i);
213
if (p->is_Proj() && p->is_CFG()) {215
if (nb_ctl_proj > 1) {221
if (nb_ctl_proj > 1) {224
assert(parent_ctl->is_Start() || parent_ctl->is_MemBar() || parent_ctl->is_Call() ||
225
BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(parent_ctl), "unexpected node");
226
assert(idom(ctl) == parent_ctl, "strange");
227
next = idom(parent_ctl);
232
if (next->is_Root() || next->is_Start() || dom_depth(next) < min_dom_depth) {238
if (ctl != n->in(0)) {239
_igvn.replace_input_of(n, 0, ctl);
240
_igvn.hash_insert(n);
247
//------------------------------set_early_ctrl---------------------------------
248
// Set earliest legal control
249
void PhaseIdealLoop::set_early_ctrl(Node* n, bool update_body) {250
Node *early = get_early_ctrl(n);
252
// Record earliest legal location
254
IdealLoopTree *loop = get_loop(early);
255
if (update_body && loop->_child == nullptr) {260
//------------------------------set_subtree_ctrl-------------------------------
261
// set missing _ctrl entries on new nodes
262
void PhaseIdealLoop::set_subtree_ctrl(Node* n, bool update_body) {263
// Already set? Get out.
264
if (_loop_or_ctrl[n->_idx]) return;
265
// Recursively set _loop_or_ctrl array to indicate where the Node goes
267
for (i = 0; i < n->req(); ++i) {269
if (m && m != C->root()) {270
set_subtree_ctrl(m, update_body);
275
set_early_ctrl(n, update_body);
278
IdealLoopTree* PhaseIdealLoop::insert_outer_loop(IdealLoopTree* loop, LoopNode* outer_l, Node* outer_ift) {279
IdealLoopTree* outer_ilt = new IdealLoopTree(this, outer_l, outer_ift);
280
IdealLoopTree* parent = loop->_parent;
281
IdealLoopTree* sibling = parent->_child;
282
if (sibling == loop) {283
parent->_child = outer_ilt;
285
while (sibling->_next != loop) {286
sibling = sibling->_next;
288
sibling->_next = outer_ilt;
290
outer_ilt->_next = loop->_next;
291
outer_ilt->_parent = parent;
292
outer_ilt->_child = loop;
293
outer_ilt->_nest = loop->_nest;
294
loop->_parent = outer_ilt;
295
loop->_next = nullptr;
297
assert(loop->_nest <= SHRT_MAX, "sanity");
301
// Create a skeleton strip mined outer loop: a Loop head before the
302
// inner strip mined loop, a safepoint and an exit condition guarded
303
// by an opaque node after the inner strip mined loop with a backedge
304
// to the loop head. The inner strip mined loop is left as it is. Only
305
// once loop optimizations are over, do we adjust the inner loop exit
306
// condition to limit its number of iterations, set the outer loop
307
// exit condition and add Phis to the outer loop head. Some loop
308
// optimizations that operate on the inner strip mined loop need to be
309
// aware of the outer strip mined loop: loop unswitching needs to
310
// clone the outer loop as well as the inner, unrolling needs to only
311
// clone the inner loop etc. No optimizations need to change the outer
312
// strip mined loop as it is only a skeleton.
313
IdealLoopTree* PhaseIdealLoop::create_outer_strip_mined_loop(BoolNode *test, Node *cmp, Node *init_control,
314
IdealLoopTree* loop, float cl_prob, float le_fcnt,
315
Node*& entry_control, Node*& iffalse) {316
Node* outer_test = _igvn.intcon(0);
317
set_ctrl(outer_test, C->root());
318
Node *orig = iffalse;
319
iffalse = iffalse->clone();
320
_igvn.register_new_node_with_optimizer(iffalse);
321
set_idom(iffalse, idom(orig), dom_depth(orig));
323
IfNode *outer_le = new OuterStripMinedLoopEndNode(iffalse, outer_test, cl_prob, le_fcnt);
324
Node *outer_ift = new IfTrueNode (outer_le);
325
Node* outer_iff = orig;
326
_igvn.replace_input_of(outer_iff, 0, outer_le);
328
LoopNode *outer_l = new OuterStripMinedLoopNode(C, init_control, outer_ift);
329
entry_control = outer_l;
331
IdealLoopTree* outer_ilt = insert_outer_loop(loop, outer_l, outer_ift);
333
set_loop(iffalse, outer_ilt);
334
// When this code runs, loop bodies have not yet been populated.
335
const bool body_populated = false;
336
register_control(outer_le, outer_ilt, iffalse, body_populated);
337
register_control(outer_ift, outer_ilt, outer_le, body_populated);
338
set_idom(outer_iff, outer_le, dom_depth(outer_le));
339
_igvn.register_new_node_with_optimizer(outer_l);
340
set_loop(outer_l, outer_ilt);
341
set_idom(outer_l, init_control, dom_depth(init_control)+1);
346
void PhaseIdealLoop::insert_loop_limit_check_predicate(ParsePredicateSuccessProj* loop_limit_check_parse_proj,
347
Node* cmp_limit, Node* bol) {348
assert(loop_limit_check_parse_proj->in(0)->is_ParsePredicate(), "must be parse predicate");
349
Node* new_predicate_proj = create_new_if_for_predicate(loop_limit_check_parse_proj, nullptr,
350
Deoptimization::Reason_loop_limit_check,
352
Node* iff = new_predicate_proj->in(0);
353
cmp_limit = _igvn.register_new_node_with_optimizer(cmp_limit);
354
bol = _igvn.register_new_node_with_optimizer(bol);
355
set_subtree_ctrl(bol, false);
356
_igvn.replace_input_of(iff, 1, bol);
359
// report that the loop predication has been actually performed
361
if (TraceLoopLimitCheck) {362
tty->print_cr("Counted Loop Limit Check generated:");363
debug_only( bol->dump(2); )
368
Node* PhaseIdealLoop::loop_exit_control(Node* x, IdealLoopTree* loop) {369
// Counted loop head must be a good RegionNode with only 3 not null
370
// control input edges: Self, Entry, LoopBack.
371
if (x->in(LoopNode::Self) == nullptr || x->req() != 3 || loop->_irreducible) {374
Node *init_control = x->in(LoopNode::EntryControl);
375
Node *back_control = x->in(LoopNode::LoopBackControl);
376
if (init_control == nullptr || back_control == nullptr) { // Partially dead379
// Must also check for TOP when looking for a dead loop
380
if (init_control->is_top() || back_control->is_top()) {384
// Allow funny placement of Safepoint
385
if (back_control->Opcode() == Op_SafePoint) {386
back_control = back_control->in(TypeFunc::Control);
389
// Controlling test for loop
390
Node *iftrue = back_control;
391
uint iftrue_op = iftrue->Opcode();
392
if (iftrue_op != Op_IfTrue &&
393
iftrue_op != Op_IfFalse) {394
// I have a weird back-control. Probably the loop-exit test is in
395
// the middle of the loop and I am looking at some trailing control-flow
396
// merge point. To fix this I would have to partially peel the loop.
397
return nullptr; // Obscure back-control
400
// Get boolean guarding loop-back test
401
Node *iff = iftrue->in(0);
402
if (get_loop(iff) != loop || !iff->in(1)->is_Bool()) {408
Node* PhaseIdealLoop::loop_exit_test(Node* back_control, IdealLoopTree* loop, Node*& incr, Node*& limit, BoolTest::mask& bt, float& cl_prob) {409
Node* iftrue = back_control;
410
uint iftrue_op = iftrue->Opcode();
411
Node* iff = iftrue->in(0);
412
BoolNode* test = iff->in(1)->as_Bool();
413
bt = test->_test._test;
414
cl_prob = iff->as_If()->_prob;
415
if (iftrue_op == Op_IfFalse) {416
bt = BoolTest(bt).negate();
417
cl_prob = 1.0 - cl_prob;
419
// Get backedge compare
420
Node* cmp = test->in(1);
421
if (!cmp->is_Cmp()) {425
// Find the trip-counter increment & limit. Limit must be loop invariant.
430
// need 'loop()' test to tell if limit is loop invariant
433
if (!is_member(loop, get_ctrl(incr))) { // Swapped trip counter and limit?434
Node* tmp = incr; // Then reverse order into the CmpI
437
bt = BoolTest(bt).commute(); // And commute the exit test
439
if (is_member(loop, get_ctrl(limit))) { // Limit must be loop-invariant442
if (!is_member(loop, get_ctrl(incr))) { // Trip counter must be loop-variant448
Node* PhaseIdealLoop::loop_iv_incr(Node* incr, Node* x, IdealLoopTree* loop, Node*& phi_incr) {449
if (incr->is_Phi()) {450
if (incr->as_Phi()->region() != x || incr->req() != 3) {451
return nullptr; // Not simple trip counter expression
454
incr = phi_incr->in(LoopNode::LoopBackControl); // Assume incr is on backedge of Phi
455
if (!is_member(loop, get_ctrl(incr))) { // Trip counter must be loop-variant462
Node* PhaseIdealLoop::loop_iv_stride(Node* incr, IdealLoopTree* loop, Node*& xphi) {463
assert(incr->Opcode() == Op_AddI || incr->Opcode() == Op_AddL, "caller resp.");
466
Node *stride = incr->in(2);
467
if (!stride->is_Con()) { // Oops, swap these468
if (!xphi->is_Con()) { // Is the other guy a constant?469
return nullptr; // Nope, unknown stride, bail out
471
Node *tmp = xphi; // 'incr' is commutative, so ok to swap
478
PhiNode* PhaseIdealLoop::loop_iv_phi(Node* xphi, Node* phi_incr, Node* x, IdealLoopTree* loop) {479
if (!xphi->is_Phi()) {480
return nullptr; // Too much math on the trip counter
482
if (phi_incr != nullptr && phi_incr != xphi) {485
PhiNode *phi = xphi->as_Phi();
487
// Phi must be of loop header; backedge must wrap to increment
488
if (phi->region() != x) {494
static int check_stride_overflow(jlong final_correction, const TypeInteger* limit_t, BasicType bt) {495
if (final_correction > 0) {496
if (limit_t->lo_as_long() > (max_signed_integer(bt) - final_correction)) {499
if (limit_t->hi_as_long() > (max_signed_integer(bt) - final_correction)) {503
if (limit_t->hi_as_long() < (min_signed_integer(bt) - final_correction)) {506
if (limit_t->lo_as_long() < (min_signed_integer(bt) - final_correction)) {513
static bool condition_stride_ok(BoolTest::mask bt, jlong stride_con) {514
// If the condition is inverted and we will be rolling
515
// through MININT to MAXINT, then bail out.
516
if (bt == BoolTest::eq || // Bail out, but this loop trips at most twice!
518
(bt == BoolTest::ne && stride_con != 1 && stride_con != -1) ||
519
// Count down loop rolls through MAXINT
520
((bt == BoolTest::le || bt == BoolTest::lt) && stride_con < 0) ||
521
// Count up loop rolls through MININT
522
((bt == BoolTest::ge || bt == BoolTest::gt) && stride_con > 0)) {523
return false; // Bail out
528
Node* PhaseIdealLoop::loop_nest_replace_iv(Node* iv_to_replace, Node* inner_iv, Node* outer_phi, Node* inner_head,
532
iv_as_long = new ConvI2LNode(inner_iv, TypeLong::INT);
533
register_new_node(iv_as_long, inner_head);
535
iv_as_long = inner_iv;
537
Node* iv_replacement = AddNode::make(outer_phi, iv_as_long, bt);
538
register_new_node(iv_replacement, inner_head);
539
for (DUIterator_Last imin, i = iv_to_replace->last_outs(imin); i >= imin;) {540
Node* u = iv_to_replace->last_out(i);
542
if (!is_dominator(inner_head, ctrl_or_self(u))) {543
assert(u->is_Phi(), "should be a Phi");
544
for (uint j = 1; j < u->req(); j++) {545
if (u->in(j) == iv_to_replace) {546
assert(is_dominator(inner_head, u->in(0)->in(j)), "iv use above loop?");
551
_igvn.rehash_node_delayed(u);
552
int nb = u->replace_edge(iv_to_replace, iv_replacement, &_igvn);
555
return iv_replacement;
558
// Add a Parse Predicate with an uncommon trap on the failing/false path. Normal control will continue on the true path.
559
void PhaseIdealLoop::add_parse_predicate(Deoptimization::DeoptReason reason, Node* inner_head, IdealLoopTree* loop,
560
SafePointNode* sfpt) {561
if (!C->too_many_traps(reason)) {562
ParsePredicateNode* parse_predicate = new ParsePredicateNode(inner_head->in(LoopNode::EntryControl), reason, &_igvn);
563
register_control(parse_predicate, loop, inner_head->in(LoopNode::EntryControl));
564
Node* if_false = new IfFalseNode(parse_predicate);
565
register_control(if_false, _ltree_root, parse_predicate);
566
Node* if_true = new IfTrueNode(parse_predicate);
567
register_control(if_true, loop, parse_predicate);
569
int trap_request = Deoptimization::make_trap_request(reason, Deoptimization::Action_maybe_recompile);
570
address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point();
571
const TypePtr* no_memory_effects = nullptr;
572
JVMState* jvms = sfpt->jvms();
573
CallNode* unc = new CallStaticJavaNode(OptoRuntime::uncommon_trap_Type(), call_addr, "uncommon_trap",
578
if (sfpt->is_Call()) {579
mem = sfpt->proj_out(TypeFunc::Memory);
580
i_o = sfpt->proj_out(TypeFunc::I_O);
582
mem = sfpt->memory();
586
Node *frame = new ParmNode(C->start(), TypeFunc::FramePtr);
587
register_new_node(frame, C->start());
588
Node *ret = new ParmNode(C->start(), TypeFunc::ReturnAdr);
589
register_new_node(ret, C->start());
591
unc->init_req(TypeFunc::Control, if_false);
592
unc->init_req(TypeFunc::I_O, i_o);
593
unc->init_req(TypeFunc::Memory, mem); // may gc ptrs
594
unc->init_req(TypeFunc::FramePtr, frame);
595
unc->init_req(TypeFunc::ReturnAdr, ret);
596
unc->init_req(TypeFunc::Parms+0, _igvn.intcon(trap_request));
597
unc->set_cnt(PROB_UNLIKELY_MAG(4));
598
unc->copy_call_debug_info(&_igvn, sfpt);
600
for (uint i = TypeFunc::Parms; i < unc->req(); i++) {601
set_subtree_ctrl(unc->in(i), false);
603
register_control(unc, _ltree_root, if_false);
605
Node* ctrl = new ProjNode(unc, TypeFunc::Control);
606
register_control(ctrl, _ltree_root, unc);
607
Node* halt = new HaltNode(ctrl, frame, "uncommon trap returned which should never happen" PRODUCT_ONLY(COMMA /*reachable*/false));
608
register_control(halt, _ltree_root, ctrl);
609
_igvn.add_input_to(C->root(), halt);
611
_igvn.replace_input_of(inner_head, LoopNode::EntryControl, if_true);
612
set_idom(inner_head, if_true, dom_depth(inner_head));
616
// Find a safepoint node that dominates the back edge. We need a
617
// SafePointNode so we can use its jvm state to create empty
619
static bool no_side_effect_since_safepoint(Compile* C, Node* x, Node* mem, MergeMemNode* mm, PhaseIdealLoop* phase) {620
SafePointNode* safepoint = nullptr;
621
for (DUIterator_Fast imax, i = x->fast_outs(imax); i < imax; i++) {622
Node* u = x->fast_out(i);
623
if (u->is_memory_phi()) {624
Node* m = u->in(LoopNode::LoopBackControl);
625
if (u->adr_type() == TypePtr::BOTTOM) {626
if (m->is_MergeMem() && mem->is_MergeMem()) {627
if (m != mem DEBUG_ONLY(|| true)) {628
// MergeMemStream can modify m, for example to adjust the length to mem.
629
// This is unfortunate, and probably unnecessary. But as it is, we need
630
// to add m to the igvn worklist, else we may have a modified node that
631
// is not on the igvn worklist.
632
phase->igvn()._worklist.push(m);
633
for (MergeMemStream mms(m->as_MergeMem(), mem->as_MergeMem()); mms.next_non_empty2(); ) {634
if (!mms.is_empty()) {635
if (mms.memory() != mms.memory2()) {639
if (mms.alias_idx() != Compile::AliasIdxBot) {640
mm->set_memory_at(mms.alias_idx(), mem->as_MergeMem()->base_memory());
646
} else if (mem->is_MergeMem()) {647
if (m != mem->as_MergeMem()->base_memory()) {654
if (mem->is_MergeMem()) {655
if (m != mem->as_MergeMem()->memory_at(C->get_alias_index(u->adr_type()))) {659
mm->set_memory_at(C->get_alias_index(u->adr_type()), mem->as_MergeMem()->base_memory());
672
SafePointNode* PhaseIdealLoop::find_safepoint(Node* back_control, Node* x, IdealLoopTree* loop) {673
IfNode* exit_test = back_control->in(0)->as_If();
674
SafePointNode* safepoint = nullptr;
675
if (exit_test->in(0)->is_SafePoint() && exit_test->in(0)->outcnt() == 1) {676
safepoint = exit_test->in(0)->as_SafePoint();
678
Node* c = back_control;
679
while (c != x && c->Opcode() != Op_SafePoint) {683
if (c->Opcode() == Op_SafePoint) {684
safepoint = c->as_SafePoint();
687
if (safepoint == nullptr) {691
Node* mem = safepoint->in(TypeFunc::Memory);
693
// We can only use that safepoint if there's no side effect between the backedge and the safepoint.
695
// mm is used for book keeping
696
MergeMemNode* mm = nullptr;
698
if (mem->is_MergeMem()) {699
mm = mem->clone()->as_MergeMem();
700
_igvn._worklist.push(mm);
701
for (MergeMemStream mms(mem->as_MergeMem()); mms.next_non_empty(); ) {702
if (mms.alias_idx() != Compile::AliasIdxBot && loop != get_loop(ctrl_or_self(mms.memory()))) {703
mm->set_memory_at(mms.alias_idx(), mem->as_MergeMem()->base_memory());
708
if (!no_side_effect_since_safepoint(C, x, mem, mm, this)) {711
assert(mm == nullptr|| _igvn.transform(mm) == mem->as_MergeMem()->base_memory(), "all memory state should have been processed");
715
_igvn.remove_dead_node(mm);
722
// If the loop has the shape of a counted loop but with a long
723
// induction variable, transform the loop in a loop nest: an inner
724
// loop that iterates for at most max int iterations with an integer
725
// induction variable and an outer loop that iterates over the full
726
// range of long values from the initial loop in (at most) max int
729
// x: for (long phi = init; phi < limit; phi += stride) {730
// // phi := Phi(L, init, incr)
731
// // incr := AddL(phi, longcon(stride))
732
// long incr = phi + stride;
733
// ... use phi and incr ...
738
// x: for (long phi = init; (phi += stride) < limit; ) {739
// // phi := Phi(L, AddL(init, stride), incr)
740
// // incr := AddL(phi, longcon(stride))
741
// long incr = phi + stride;
742
// ... use phi and (phi + stride) ...
747
// const ulong inner_iters_limit = INT_MAX - stride - 1; //near 0x7FFFFFF0
748
// assert(stride <= inner_iters_limit); // else abort transform
749
// assert((extralong)limit + stride <= LONG_MAX); // else deopt
750
// outer_head: for (long outer_phi = init;;) {751
// // outer_phi := Phi(outer_head, init, AddL(outer_phi, I2L(inner_phi)))
752
// ulong inner_iters_max = (ulong) MAX(0, ((extralong)limit + stride - outer_phi));
753
// long inner_iters_actual = MIN(inner_iters_limit, inner_iters_max);
754
// assert(inner_iters_actual == (int)inner_iters_actual);
755
// int inner_phi, inner_incr;
756
// x: for (inner_phi = 0;; inner_phi = inner_incr) {757
// // inner_phi := Phi(x, intcon(0), inner_incr)
758
// // inner_incr := AddI(inner_phi, intcon(stride))
759
// inner_incr = inner_phi + stride;
760
// if (inner_incr < inner_iters_actual) {761
// ... use phi=>(outer_phi+inner_phi) ...
766
// if ((outer_phi+inner_phi) < limit) //OR (outer_phi+inner_incr) < limit
771
// The same logic is used to transform an int counted loop that contains long range checks into a loop nest of 2 int
772
// loops with long range checks transformed to int range checks in the inner loop.
773
bool PhaseIdealLoop::create_loop_nest(IdealLoopTree* loop, Node_List &old_new) {774
Node* x = loop->_head;
775
// Only for inner loops
776
if (loop->_child != nullptr || !x->is_BaseCountedLoop() || x->as_Loop()->is_loop_nest_outer_loop()) {780
if (x->is_CountedLoop() && !x->as_CountedLoop()->is_main_loop() && !x->as_CountedLoop()->is_normal_loop()) {784
BaseCountedLoopNode* head = x->as_BaseCountedLoop();
785
BasicType bt = x->as_BaseCountedLoop()->bt();
787
check_counted_loop_shape(loop, x, bt);
791
Atomic::inc(&_long_loop_candidates);
795
jlong stride_con_long = head->stride_con();
796
assert(stride_con_long != 0, "missed some peephole opt");
797
// We can't iterate for more than max int at a time.
798
if (stride_con_long != (jint)stride_con_long || stride_con_long == min_jint) {799
assert(bt == T_LONG, "only for long loops");
802
jint stride_con = checked_cast<jint>(stride_con_long);
803
// The number of iterations for the integer count loop: guarantee no
804
// overflow: max_jint - stride_con max. -1 so there's no need for a
805
// loop limit check if the exit test is <= or >=.
806
int iters_limit = max_jint - ABS(stride_con) - 1;
808
if (bt == T_LONG && StressLongCountedLoop > 0) {809
iters_limit = iters_limit / StressLongCountedLoop;
812
// At least 2 iterations so counted loop construction doesn't fail
813
if (iters_limit/ABS(stride_con) < 2) {817
PhiNode* phi = head->phi()->as_Phi();
818
Node* incr = head->incr();
820
Node* back_control = head->in(LoopNode::LoopBackControl);
822
// data nodes on back branch not supported
823
if (back_control->outcnt() > 1) {827
Node* limit = head->limit();
828
// We'll need to use the loop limit before the inner loop is entered
829
if (!is_dominator(get_ctrl(limit), x)) {833
IfNode* exit_test = head->loopexit();
835
assert(back_control->Opcode() == Op_IfTrue, "wrong projection for back edge");
837
Node_List range_checks;
838
iters_limit = extract_long_range_checks(loop, stride_con, iters_limit, phi, range_checks);
841
// The only purpose of creating a loop nest is to handle long range checks. If there are none, do not proceed further.
842
if (range_checks.size() == 0) {847
// Take what we know about the number of iterations of the long counted loop into account when computing the limit of
849
const Node* init = head->init_trip();
850
const TypeInteger* lo = _igvn.type(init)->is_integer(bt);
851
const TypeInteger* hi = _igvn.type(limit)->is_integer(bt);
852
if (stride_con < 0) {855
if (hi->hi_as_long() <= lo->lo_as_long()) {856
// not a loop after all
860
if (range_checks.size() > 0) {861
// This transformation requires peeling one iteration. Also, if it has range checks and they are eliminated by Loop
862
// Predication, then 2 Hoisted Check Predicates are added for one range check. Finally, transforming a long range
863
// check requires extra logic to be executed before the loop is entered and for the outer loop. As a result, the
864
// transformations can't pay off for a small number of iterations: roughly, if the loop runs for 3 iterations, it's
865
// going to execute as many range checks once transformed with range checks eliminated (1 peeled iteration with
866
// range checks + 2 predicates per range checks) as it would have not transformed. It also has to pay for the extra
867
// logic on loop entry and for the outer loop.
868
loop->compute_trip_count(this);
869
if (head->is_CountedLoop() && head->as_CountedLoop()->has_exact_trip_count()) {870
if (head->as_CountedLoop()->trip_count() <= 3) {874
loop->compute_profile_trip_cnt(this);
875
if (!head->is_profile_trip_failed() && head->profile_trip_cnt() <= 3) {881
julong orig_iters = (julong)hi->hi_as_long() - lo->lo_as_long();
882
iters_limit = checked_cast<int>(MIN2((julong)iters_limit, orig_iters));
884
// We need a safepoint to insert Parse Predicates for the inner loop.
885
SafePointNode* safepoint;
886
if (bt == T_INT && head->as_CountedLoop()->is_strip_mined()) {887
// Loop is strip mined: use the safepoint of the outer strip mined loop
888
OuterStripMinedLoopNode* outer_loop = head->as_CountedLoop()->outer_loop();
889
assert(outer_loop != nullptr, "no outer loop");
890
safepoint = outer_loop->outer_safepoint();
891
outer_loop->transform_to_counted_loop(&_igvn, this);
892
exit_test = head->loopexit();
894
safepoint = find_safepoint(back_control, x, loop);
897
Node* exit_branch = exit_test->proj_out(false);
898
Node* entry_control = head->in(LoopNode::EntryControl);
900
// Clone the control flow of the loop to build an outer loop
901
Node* outer_back_branch = back_control->clone();
902
Node* outer_exit_test = new IfNode(exit_test->in(0), exit_test->in(1), exit_test->_prob, exit_test->_fcnt);
903
Node* inner_exit_branch = exit_branch->clone();
905
LoopNode* outer_head = new LoopNode(entry_control, outer_back_branch);
906
IdealLoopTree* outer_ilt = insert_outer_loop(loop, outer_head, outer_back_branch);
908
const bool body_populated = true;
909
register_control(outer_head, outer_ilt, entry_control, body_populated);
911
_igvn.register_new_node_with_optimizer(inner_exit_branch);
912
set_loop(inner_exit_branch, outer_ilt);
913
set_idom(inner_exit_branch, exit_test, dom_depth(exit_branch));
915
outer_exit_test->set_req(0, inner_exit_branch);
916
register_control(outer_exit_test, outer_ilt, inner_exit_branch, body_populated);
918
_igvn.replace_input_of(exit_branch, 0, outer_exit_test);
919
set_idom(exit_branch, outer_exit_test, dom_depth(exit_branch));
921
outer_back_branch->set_req(0, outer_exit_test);
922
register_control(outer_back_branch, outer_ilt, outer_exit_test, body_populated);
924
_igvn.replace_input_of(x, LoopNode::EntryControl, outer_head);
925
set_idom(x, outer_head, dom_depth(x));
927
// add an iv phi to the outer loop and use it to compute the inner
928
// loop iteration limit
929
Node* outer_phi = phi->clone();
930
outer_phi->set_req(0, outer_head);
931
register_new_node(outer_phi, outer_head);
933
Node* inner_iters_max = nullptr;
934
if (stride_con > 0) {935
inner_iters_max = MaxNode::max_diff_with_zero(limit, outer_phi, TypeInteger::bottom(bt), _igvn);
937
inner_iters_max = MaxNode::max_diff_with_zero(outer_phi, limit, TypeInteger::bottom(bt), _igvn);
940
Node* inner_iters_limit = _igvn.integercon(iters_limit, bt);
941
// inner_iters_max may not fit in a signed integer (iterating from
942
// Long.MIN_VALUE to Long.MAX_VALUE for instance). Use an unsigned
944
const TypeInteger* inner_iters_actual_range = TypeInteger::make(0, iters_limit, Type::WidenMin, bt);
945
Node* inner_iters_actual = MaxNode::unsigned_min(inner_iters_max, inner_iters_limit, inner_iters_actual_range, _igvn);
947
Node* inner_iters_actual_int;
949
inner_iters_actual_int = new ConvL2INode(inner_iters_actual);
950
_igvn.register_new_node_with_optimizer(inner_iters_actual_int);
951
// When the inner loop is transformed to a counted loop, a loop limit check is not expected to be needed because
952
// the loop limit is less or equal to max_jint - stride - 1 (if stride is positive but a similar argument exists for
953
// a negative stride). We add a CastII here to guarantee that, when the counted loop is created in a subsequent loop
954
// opts pass, an accurate range of values for the limits is found.
955
const TypeInt* inner_iters_actual_int_range = TypeInt::make(0, iters_limit, Type::WidenMin);
956
inner_iters_actual_int = new CastIINode(outer_head, inner_iters_actual_int, inner_iters_actual_int_range, ConstraintCastNode::UnconditionalDependency);
957
_igvn.register_new_node_with_optimizer(inner_iters_actual_int);
959
inner_iters_actual_int = inner_iters_actual;
962
Node* int_zero = _igvn.intcon(0);
963
set_ctrl(int_zero, C->root());
964
if (stride_con < 0) {965
inner_iters_actual_int = new SubINode(int_zero, inner_iters_actual_int);
966
_igvn.register_new_node_with_optimizer(inner_iters_actual_int);
969
// Clone the iv data nodes as an integer iv
970
Node* int_stride = _igvn.intcon(stride_con);
971
set_ctrl(int_stride, C->root());
972
Node* inner_phi = new PhiNode(x->in(0), TypeInt::INT);
973
Node* inner_incr = new AddINode(inner_phi, int_stride);
974
Node* inner_cmp = nullptr;
975
inner_cmp = new CmpINode(inner_incr, inner_iters_actual_int);
976
Node* inner_bol = new BoolNode(inner_cmp, exit_test->in(1)->as_Bool()->_test._test);
977
inner_phi->set_req(LoopNode::EntryControl, int_zero);
978
inner_phi->set_req(LoopNode::LoopBackControl, inner_incr);
979
register_new_node(inner_phi, x);
980
register_new_node(inner_incr, x);
981
register_new_node(inner_cmp, x);
982
register_new_node(inner_bol, x);
984
_igvn.replace_input_of(exit_test, 1, inner_bol);
986
// Clone inner loop phis to outer loop
987
for (uint i = 0; i < head->outcnt(); i++) {988
Node* u = head->raw_out(i);
989
if (u->is_Phi() && u != inner_phi && u != phi) {990
assert(u->in(0) == head, "inconsistent");
991
Node* clone = u->clone();
992
clone->set_req(0, outer_head);
993
register_new_node(clone, outer_head);
994
_igvn.replace_input_of(u, LoopNode::EntryControl, clone);
998
// Replace inner loop long iv phi as inner loop int iv phi + outer
1000
Node* iv_add = loop_nest_replace_iv(phi, inner_phi, outer_phi, head, bt);
1002
set_subtree_ctrl(inner_iters_actual_int, body_populated);
1004
LoopNode* inner_head = create_inner_head(loop, head, exit_test);
1006
// Summary of steps from initial loop to loop nest:
1008
// == old IR nodes =>
1010
// entry_control: {...}1012
// for (long phi = init;;) {1013
// // phi := Phi(x, init, incr)
1014
// // incr := AddL(phi, longcon(stride))
1017
// back_control: fallthrough;
1019
// exit_branch: break;
1020
// long incr = phi + stride;
1021
// ... use phi and incr ...
1025
// == new IR nodes (just before final peel) =>
1027
// entry_control: {...}1028
// long adjusted_limit = limit + stride; //because phi_incr != nullptr
1029
// assert(!limit_check_required || (extralong)limit + stride == adjusted_limit); // else deopt
1030
// ulong inner_iters_limit = max_jint - ABS(stride) - 1; //near 0x7FFFFFF0
1032
// for (long outer_phi = init;;) {1033
// // outer_phi := phi->clone(), in(0):=outer_head, => Phi(outer_head, init, incr)
1034
// // REPLACE phi => AddL(outer_phi, I2L(inner_phi))
1035
// // REPLACE incr => AddL(outer_phi, I2L(inner_incr))
1036
// // SO THAT outer_phi := Phi(outer_head, init, AddL(outer_phi, I2L(inner_incr)))
1037
// ulong inner_iters_max = (ulong) MAX(0, ((extralong)adjusted_limit - outer_phi) * SGN(stride));
1038
// int inner_iters_actual_int = (int) MIN(inner_iters_limit, inner_iters_max) * SGN(stride);
1039
// inner_head: x: //in(1) := outer_head
1041
// for (inner_phi = 0;;) {1042
// // inner_phi := Phi(x, intcon(0), inner_phi + stride)
1043
// int inner_incr = inner_phi + stride;
1044
// bool inner_bol = (inner_incr < inner_iters_actual_int);
1045
// exit_test: //exit_test->in(1) := inner_bol;
1046
// if (inner_bol) // WAS (phi < limit)
1047
// back_control: fallthrough;
1049
// inner_exit_branch: break; //exit_branch->clone()
1050
// ... use phi=>(outer_phi+inner_phi) ...
1051
// inner_phi = inner_phi + stride; // inner_incr
1053
// outer_exit_test: //exit_test->clone(), in(0):=inner_exit_branch
1054
// if ((outer_phi+inner_phi) < limit) // WAS (phi < limit)
1055
// outer_back_branch: fallthrough; //back_control->clone(), in(0):=outer_exit_test
1057
// exit_branch: break; //in(0) := outer_exit_test
1061
outer_phi = new ConvI2LNode(outer_phi);
1062
register_new_node(outer_phi, outer_head);
1065
transform_long_range_checks(stride_con, range_checks, outer_phi, inner_iters_actual_int,
1066
inner_phi, iv_add, inner_head);
1067
// Peel one iteration of the loop and use the safepoint at the end
1068
// of the peeled iteration to insert Parse Predicates. If no well
1069
// positioned safepoint peel to guarantee a safepoint in the outer
1071
if (safepoint != nullptr || !loop->_has_call) {1073
do_peeling(loop, old_new);
1075
C->set_major_progress();
1078
if (safepoint != nullptr) {1079
SafePointNode* cloned_sfpt = old_new[safepoint->_idx]->as_SafePoint();
1081
if (UseLoopPredicate) {1082
add_parse_predicate(Deoptimization::Reason_predicate, inner_head, outer_ilt, cloned_sfpt);
1084
if (UseProfiledLoopPredicate) {1085
add_parse_predicate(Deoptimization::Reason_profile_predicate, inner_head, outer_ilt, cloned_sfpt);
1087
add_parse_predicate(Deoptimization::Reason_loop_limit_check, inner_head, outer_ilt, cloned_sfpt);
1092
Atomic::inc(&_long_loop_nests);
1096
inner_head->mark_loop_nest_inner_loop();
1097
outer_head->mark_loop_nest_outer_loop();
1102
int PhaseIdealLoop::extract_long_range_checks(const IdealLoopTree* loop, jint stride_con, int iters_limit, PhiNode* phi,
1103
Node_List& range_checks) {1104
const jlong min_iters = 2;
1105
jlong reduced_iters_limit = iters_limit;
1106
jlong original_iters_limit = iters_limit;
1107
for (uint i = 0; i < loop->_body.size(); i++) {1108
Node* c = loop->_body.at(i);
1109
if (c->is_IfProj() && c->in(0)->is_RangeCheck()) {1110
IfProjNode* if_proj = c->as_IfProj();
1111
CallStaticJavaNode* call = if_proj->is_uncommon_trap_if_pattern();
1112
if (call != nullptr) {1113
Node* range = nullptr;
1114
Node* offset = nullptr;
1116
if (loop->is_range_check_if(if_proj, this, T_LONG, phi, range, offset, scale) &&
1117
loop->is_invariant(range) && loop->is_invariant(offset) &&
1118
scale != min_jlong &&
1119
original_iters_limit / ABS(scale) >= min_iters * ABS(stride_con)) {1120
assert(scale == (jint)scale, "scale should be an int");
1121
reduced_iters_limit = MIN2(reduced_iters_limit, original_iters_limit/ABS(scale));
1122
range_checks.push(c);
1128
return checked_cast<int>(reduced_iters_limit);
1131
// One execution of the inner loop covers a sub-range of the entire iteration range of the loop: [A,Z), aka [A=init,
1132
// Z=limit). If the loop has at least one trip (which is the case here), the iteration variable i always takes A as its
1133
// first value, followed by A+S (S is the stride), next A+2S, etc. The limit is exclusive, so that the final value B of
1134
// i is never Z. It will be B=Z-1 if S=1, or B=Z+1 if S=-1.
1136
// If |S|>1 the formula for the last value B would require a floor operation, specifically B=floor((Z-sgn(S)-A)/S)*S+A,
1137
// which is B=Z-sgn(S)U for some U in [1,|S|]. So when S>0, i ranges as i:[A,Z) or i:[A,B=Z-U], or else (in reverse)
1138
// as i:(Z,A] or i:[B=Z+U,A]. It will become important to reason about this inclusive range [A,B] or [B,A].
1140
// Within the loop there may be many range checks. Each such range check (R.C.) is of the form 0 <= i*K+L < R, where K
1141
// is a scale factor applied to the loop iteration variable i, and L is some offset; K, L, and R are loop-invariant.
1142
// Because R is never negative (see below), this check can always be simplified to an unsigned check i*K+L <u R.
1144
// When a long loop over a 64-bit variable i (outer_iv) is decomposed into a series of shorter sub-loops over a 32-bit
1145
// variable j (inner_iv), j ranges over a shorter interval j:[0,B_2] or [0,Z_2) (assuming S > 0), where the limit is
1146
// chosen to prevent various cases of 32-bit overflow (including multiplications j*K below). In the sub-loop the
1147
// logical value i is offset from j by a 64-bit constant C, so i ranges in i:C+[0,Z_2).
1149
// For S<0, j ranges (in reverse!) through j:[-|B_2|,0] or (-|Z_2|,0]. For either sign of S, we can say i=j+C and j
1150
// ranges through 32-bit ranges [A_2,B_2] or [B_2,A_2] (A_2=0 of course).
1152
// The disjoint union of all the C+[A_2,B_2] ranges from the sub-loops must be identical to the whole range [A,B].
1153
// Assuming S>0, the first C must be A itself, and the next C value is the previous C+B_2, plus S. If |S|=1, the next
1154
// C value is also the previous C+Z_2. In each sub-loop, j counts from j=A_2=0 and i counts from C+0 and exits at
1155
// j=B_2 (i=C+B_2), just before it gets to i=C+Z_2. Both i and j count up (from C and 0) if S>0; otherwise they count
1156
// down (from C and 0 again).
1158
// Returning to range checks, we see that each i*K+L <u R expands to (C+j)*K+L <u R, or j*K+Q <u R, where Q=(C*K+L).
1159
// (Recall that K and L and R are loop-invariant scale, offset and range values for a particular R.C.) This is still a
1160
// 64-bit comparison, so the range check elimination logic will not apply to it. (The R.C.E. transforms operate only on
1161
// 32-bit indexes and comparisons, because they use 64-bit temporary values to avoid overflow; see
1162
// PhaseIdealLoop::add_constraint.)
1164
// We must transform this comparison so that it gets the same answer, but by means of a 32-bit R.C. (using j not i) of
1165
// the form j*K+L_2 <u32 R_2. Note that L_2 and R_2 must be loop-invariant, but only with respect to the sub-loop. Thus, the
1166
// problem reduces to computing values for L_2 and R_2 (for each R.C. in the loop) in the loop header for the sub-loop.
1167
// Then the standard R.C.E. transforms can take those as inputs and further compute the necessary minimum and maximum
1168
// values for the 32-bit counter j within which the range checks can be eliminated.
1170
// So, given j*K+Q <u R, we need to find some j*K+L_2 <u32 R_2, where L_2 and R_2 fit in 32 bits, and the 32-bit operations do
1171
// not overflow. We also need to cover the cases where i*K+L (= j*K+Q) overflows to a 64-bit negative, since that is
1172
// allowed as an input to the R.C., as long as the R.C. as a whole fails.
1174
// If 32-bit multiplication j*K might overflow, we adjust the sub-loop limit Z_2 closer to zero to reduce j's range.
1176
// For each R.C. j*K+Q <u32 R, the range of mathematical values of j*K+Q in the sub-loop is [Q_min, Q_max], where
1177
// Q_min=Q and Q_max=B_2*K+Q (if S>0 and K>0), Q_min=A_2*K+Q and Q_max=Q (if S<0 and K>0),
1178
// Q_min=B_2*K+Q and Q_max=Q if (S>0 and K<0), Q_min=Q and Q_max=A_2*K+Q (if S<0 and K<0)
1180
// Note that the first R.C. value is always Q=(S*K>0 ? Q_min : Q_max). Also Q_{min,max} = Q + {min,max}(A_2*K,B_2*K).1181
// If S*K>0 then, as the loop iterations progress, each R.C. value i*K+L = j*K+Q goes up from Q=Q_min towards Q_max.
1182
// If S*K<0 then j*K+Q starts at Q=Q_max and goes down towards Q_min.
1184
// Case A: Some Negatives (but no overflow).
1186
// |s64_min . . . 0 . . . s64_max|
1187
// | . Q_min..Q_max . 0 . . . . | s64 negative
1188
// | . . . . R=0 R< R< R< R< | (against R values)
1189
// | . . . Q_min..0..Q_max . . . | small mixed
1190
// | . . . . R R R< R< R< | (against R values)
1192
// R values which are out of range (>Q_max+1) are reduced to max(0,Q_max+1). They are marked on the number line as R<.
1194
// So, if Q_min <s64 0, then use this test:
1195
// j*K + s32_trunc(Q_min) <u32 clamp(R, 0, Q_max+1) if S*K>0 (R.C.E. steps upward)
1196
// j*K + s32_trunc(Q_max) <u32 clamp(R, 0, Q_max+1) if S*K<0 (R.C.E. steps downward)
1197
// Both formulas reduce to adding j*K to the 32-bit truncated value of the first R.C. expression value, Q:
1198
// j*K + s32_trunc(Q) <u32 clamp(R, 0, Q_max+1) for all S,K
1200
// If the 32-bit truncation loses information, no harm is done, since certainly the clamp also will return R_2=zero.
1202
// Case B: No Negatives.
1204
// |s64_min . . . 0 . . . s64_max|
1205
// | . . . . 0 Q_min..Q_max . . | small positive
1206
// | . . . . R> R R R< R< | (against R values)
1207
// | . . . . 0 . Q_min..Q_max . | s64 positive
1208
// | . . . . R> R> R R R< | (against R values)
1210
// R values which are out of range (<Q_min or >Q_max+1) are reduced as marked: R> up to Q_min, R< down to Q_max+1.
1211
// Then the whole comparison is shifted left by Q_min, so it can take place at zero, which is a nice 32-bit value.
1213
// So, if both Q_min, Q_max+1 >=s64 0, then use this test:
1214
// j*K + 0 <u32 clamp(R, Q_min, Q_max+1) - Q_min if S*K>0
1216
// j*K + Q - Q_min <u32 clamp(R, Q_min, Q_max+1) - Q_min for all S,K
1218
// Case C: Overflow in the 64-bit domain
1220
// |..Q_max-2^64 . . 0 . . . Q_min..| s64 overflow
1221
// | . . . . R> R> R> R> R | (against R values)
1223
// In this case, Q_min >s64 Q_max+1, even though the mathematical values of Q_min and Q_max+1 are correctly ordered.
1224
// The formulas from the previous case can be used, except that the bad upper bound Q_max is replaced by max_jlong.
1225
// (In fact, we could use any replacement bound from R to max_jlong inclusive, as the input to the clamp function.)
1227
// So if Q_min >=s64 0 but Q_max+1 <s64 0, use this test:
1228
// j*K + 0 <u32 clamp(R, Q_min, max_jlong) - Q_min if S*K>0
1230
// j*K + Q - Q_min <u32 clamp(R, Q_min, max_jlong) - Q_min for all S,K
1232
// Dropping the bad bound means only Q_min is used to reduce the range of R:
1233
// j*K + Q - Q_min <u32 max(Q_min, R) - Q_min for all S,K
1235
// Here the clamp function is a 64-bit min/max that reduces the dynamic range of its R operand to the required [L,H]:
1236
// clamp(X, L, H) := max(L, min(X, H))
1237
// When degenerately L > H, it returns L not H.
1239
// All of the formulas above can be merged into a single one:
1240
// L_clamp = Q_min < 0 ? 0 : Q_min --whether and how far to left-shift
1241
// H_clamp = Q_max+1 < Q_min ? max_jlong : Q_max+1
1242
// = Q_max+1 < 0 && Q_min >= 0 ? max_jlong : Q_max+1
1243
// Q_first = Q = (S*K>0 ? Q_min : Q_max) = (C*K+L)
1244
// R_clamp = clamp(R, L_clamp, H_clamp) --reduced dynamic range
1246
// j*K + Q_first - L_clamp <u32 R_clamp - L_clamp
1248
// j*K + L_2 <u32 R_2
1250
// L_2 = Q_first - L_clamp
1251
// R_2 = R_clamp - L_clamp
1253
// Note on why R is never negative:
1255
// Various details of this transformation would break badly if R could be negative, so this transformation only
1256
// operates after obtaining hard evidence that R<0 is impossible. For example, if R comes from a LoadRange node, we
1257
// know R cannot be negative. For explicit checks (of both int and long) a proof is constructed in
1258
// inline_preconditions_checkIndex, which triggers an uncommon trap if R<0, then wraps R in a ConstraintCastNode with a
1259
// non-negative type. Later on, when IdealLoopTree::is_range_check_if looks for an optimizable R.C., it checks that
1260
// the type of that R node is non-negative. Any "wild" R node that could be negative is not treated as an optimizable
1261
// R.C., but R values from a.length and inside checkIndex are good to go.
1263
void PhaseIdealLoop::transform_long_range_checks(int stride_con, const Node_List &range_checks, Node* outer_phi,
1264
Node* inner_iters_actual_int, Node* inner_phi,
1265
Node* iv_add, LoopNode* inner_head) {1266
Node* long_zero = _igvn.longcon(0);
1267
set_ctrl(long_zero, C->root());
1268
Node* int_zero = _igvn.intcon(0);
1269
set_ctrl(int_zero, this->C->root());
1270
Node* long_one = _igvn.longcon(1);
1271
set_ctrl(long_one, this->C->root());
1272
Node* int_stride = _igvn.intcon(checked_cast<int>(stride_con));
1273
set_ctrl(int_stride, this->C->root());
1275
for (uint i = 0; i < range_checks.size(); i++) {1276
ProjNode* proj = range_checks.at(i)->as_Proj();
1277
ProjNode* unc_proj = proj->other_if_proj();
1278
RangeCheckNode* rc = proj->in(0)->as_RangeCheck();
1280
Node* offset = nullptr;
1281
Node* rc_bol = rc->in(1);
1282
Node* rc_cmp = rc_bol->in(1);
1283
if (rc_cmp->Opcode() == Op_CmpU) {1284
// could be shared and have already been taken care of
1287
bool short_scale = false;
1288
bool ok = is_scaled_iv_plus_offset(rc_cmp->in(1), iv_add, T_LONG, &scale, &offset, &short_scale);
1289
assert(ok, "inconsistent: was tested before");
1290
Node* range = rc_cmp->in(2);
1291
Node* c = rc->in(0);
1292
Node* entry_control = inner_head->in(LoopNode::EntryControl);
1295
Node* K = _igvn.longcon(scale);
1296
set_ctrl(K, this->C->root());
1302
// (int)i*K + L <u64 R
1303
// with K an int into:
1304
// i*(long)K + L <u64 unsigned_min((long)max_jint + L + 1, R)
1305
// to protect against an overflow of (int)i*K
1307
// Because if (int)i*K overflows, there are K,L where:
1308
// (int)i*K + L <u64 R is false because (int)i*K+L overflows to a negative which becomes a huge u64 value.
1309
// But if i*(long)K + L is >u64 (long)max_jint and still is <u64 R, then
1310
// i*(long)K + L <u64 R is true.
1312
// As a consequence simply converting i*K + L <u64 R to i*(long)K + L <u64 R could cause incorrect execution.
1314
// It's always true that:
1315
// (int)i*K <u64 (long)max_jint + 1
1316
// which implies (int)i*K + L <u64 (long)max_jint + 1 + L
1317
// As a consequence:
1318
// i*(long)K + L <u64 unsigned_min((long)max_jint + L + 1, R)
1319
// is always false in case of overflow of i*K
1321
// Note, there are also K,L where i*K overflows and
1322
// i*K + L <u64 R is true, but
1323
// i*(long)K + L <u64 unsigned_min((long)max_jint + L + 1, R) is false
1324
// So this transformation could cause spurious deoptimizations and failed range check elimination
1325
// (but not incorrect execution) for unlikely corner cases with overflow.
1326
// If this causes problems in practice, we could maybe direct execution to a post-loop, instead of deoptimizing.
1327
Node* max_jint_plus_one_long = _igvn.longcon((jlong)max_jint + 1);
1328
set_ctrl(max_jint_plus_one_long, C->root());
1329
Node* max_range = new AddLNode(max_jint_plus_one_long, L);
1330
register_new_node(max_range, entry_control);
1331
R = MaxNode::unsigned_min(R, max_range, TypeLong::POS, _igvn);
1332
set_subtree_ctrl(R, true);
1335
Node* C = outer_phi;
1337
// Start with 64-bit values:
1339
// (C+j)*K + L <u64 R
1340
// j*K + Q <u64 R where Q = Q_first = C*K+L
1341
Node* Q_first = new MulLNode(C, K);
1342
register_new_node(Q_first, entry_control);
1343
Q_first = new AddLNode(Q_first, L);
1344
register_new_node(Q_first, entry_control);
1346
// Compute endpoints of the range of values j*K + Q.
1347
// Q_min = (j=0)*K + Q; Q_max = (j=B_2)*K + Q
1348
Node* Q_min = Q_first;
1350
// Compute the exact ending value B_2 (which is really A_2 if S < 0)
1351
Node* B_2 = new LoopLimitNode(this->C, int_zero, inner_iters_actual_int, int_stride);
1352
register_new_node(B_2, entry_control);
1353
B_2 = new SubINode(B_2, int_stride);
1354
register_new_node(B_2, entry_control);
1355
B_2 = new ConvI2LNode(B_2);
1356
register_new_node(B_2, entry_control);
1358
Node* Q_max = new MulLNode(B_2, K);
1359
register_new_node(Q_max, entry_control);
1360
Q_max = new AddLNode(Q_max, Q_first);
1361
register_new_node(Q_max, entry_control);
1363
if (scale * stride_con < 0) {1366
// Now, mathematically, Q_max > Q_min, and they are close enough so that (Q_max-Q_min) fits in 32 bits.
1368
// L_clamp = Q_min < 0 ? 0 : Q_min
1369
Node* Q_min_cmp = new CmpLNode(Q_min, long_zero);
1370
register_new_node(Q_min_cmp, entry_control);
1371
Node* Q_min_bool = new BoolNode(Q_min_cmp, BoolTest::lt);
1372
register_new_node(Q_min_bool, entry_control);
1373
Node* L_clamp = new CMoveLNode(Q_min_bool, Q_min, long_zero, TypeLong::LONG);
1374
register_new_node(L_clamp, entry_control);
1375
// (This could also be coded bitwise as L_clamp = Q_min & ~(Q_min>>63).)
1377
Node* Q_max_plus_one = new AddLNode(Q_max, long_one);
1378
register_new_node(Q_max_plus_one, entry_control);
1380
// H_clamp = Q_max+1 < Q_min ? max_jlong : Q_max+1
1381
// (Because Q_min and Q_max are close, the overflow check could also be encoded as Q_max+1 < 0 & Q_min >= 0.)
1382
Node* max_jlong_long = _igvn.longcon(max_jlong);
1383
set_ctrl(max_jlong_long, this->C->root());
1384
Node* Q_max_cmp = new CmpLNode(Q_max_plus_one, Q_min);
1385
register_new_node(Q_max_cmp, entry_control);
1386
Node* Q_max_bool = new BoolNode(Q_max_cmp, BoolTest::lt);
1387
register_new_node(Q_max_bool, entry_control);
1388
Node* H_clamp = new CMoveLNode(Q_max_bool, Q_max_plus_one, max_jlong_long, TypeLong::LONG);
1389
register_new_node(H_clamp, entry_control);
1390
// (This could also be coded bitwise as H_clamp = ((Q_max+1)<<1 | M)>>>1 where M = (Q_max+1)>>63 & ~Q_min>>63.)
1392
// R_2 = clamp(R, L_clamp, H_clamp) - L_clamp
1393
// that is: R_2 = clamp(R, L_clamp=0, H_clamp=Q_max) if Q_min < 0
1394
// or else: R_2 = clamp(R, L_clamp, H_clamp) - Q_min if Q_min >= 0
1395
// and also: R_2 = clamp(R, L_clamp, Q_max+1) - L_clamp if Q_min < Q_max+1 (no overflow)
1396
// or else: R_2 = clamp(R, L_clamp, *no limit*)- L_clamp if Q_max+1 < Q_min (overflow)
1397
Node* R_2 = clamp(R, L_clamp, H_clamp);
1398
R_2 = new SubLNode(R_2, L_clamp);
1399
register_new_node(R_2, entry_control);
1400
R_2 = new ConvL2INode(R_2, TypeInt::POS);
1401
register_new_node(R_2, entry_control);
1403
// L_2 = Q_first - L_clamp
1404
// We are subtracting L_clamp from both sides of the <u32 comparison.
1405
// If S*K>0, then Q_first == 0 and the R.C. expression at -L_clamp and steps upward to Q_max-L_clamp.
1406
// If S*K<0, then Q_first != 0 and the R.C. expression starts high and steps downward to Q_min-L_clamp.
1407
Node* L_2 = new SubLNode(Q_first, L_clamp);
1408
register_new_node(L_2, entry_control);
1409
L_2 = new ConvL2INode(L_2, TypeInt::INT);
1410
register_new_node(L_2, entry_control);
1412
// Transform the range check using the computed values L_2/R_2
1413
// from: i*K + L <u64 R
1414
// to: j*K + L_2 <u32 R_2
1416
// (j*K + Q_first) - L_clamp <u32 clamp(R, L_clamp, H_clamp) - L_clamp
1417
K = _igvn.intcon(checked_cast<int>(scale));
1418
set_ctrl(K, this->C->root());
1419
Node* scaled_iv = new MulINode(inner_phi, K);
1420
register_new_node(scaled_iv, c);
1421
Node* scaled_iv_plus_offset = new AddINode(scaled_iv, L_2);
1422
register_new_node(scaled_iv_plus_offset, c);
1424
Node* new_rc_cmp = new CmpUNode(scaled_iv_plus_offset, R_2);
1425
register_new_node(new_rc_cmp, c);
1427
_igvn.replace_input_of(rc_bol, 1, new_rc_cmp);
1431
Node* PhaseIdealLoop::clamp(Node* R, Node* L, Node* H) {1432
Node* min = MaxNode::signed_min(R, H, TypeLong::LONG, _igvn);
1433
set_subtree_ctrl(min, true);
1434
Node* max = MaxNode::signed_max(L, min, TypeLong::LONG, _igvn);
1435
set_subtree_ctrl(max, true);
1439
LoopNode* PhaseIdealLoop::create_inner_head(IdealLoopTree* loop, BaseCountedLoopNode* head,
1440
IfNode* exit_test) {1441
LoopNode* new_inner_head = new LoopNode(head->in(1), head->in(2));
1442
IfNode* new_inner_exit = new IfNode(exit_test->in(0), exit_test->in(1), exit_test->_prob, exit_test->_fcnt);
1443
_igvn.register_new_node_with_optimizer(new_inner_head);
1444
_igvn.register_new_node_with_optimizer(new_inner_exit);
1445
loop->_body.push(new_inner_head);
1446
loop->_body.push(new_inner_exit);
1447
loop->_body.yank(head);
1448
loop->_body.yank(exit_test);
1449
set_loop(new_inner_head, loop);
1450
set_loop(new_inner_exit, loop);
1451
set_idom(new_inner_head, idom(head), dom_depth(head));
1452
set_idom(new_inner_exit, idom(exit_test), dom_depth(exit_test));
1453
lazy_replace(head, new_inner_head);
1454
lazy_replace(exit_test, new_inner_exit);
1455
loop->_head = new_inner_head;
1456
return new_inner_head;
1460
void PhaseIdealLoop::check_counted_loop_shape(IdealLoopTree* loop, Node* x, BasicType bt) {1461
Node* back_control = loop_exit_control(x, loop);
1462
assert(back_control != nullptr, "no back control");
1464
BoolTest::mask mask = BoolTest::illegal;
1466
Node* incr = nullptr;
1467
Node* limit = nullptr;
1469
Node* cmp = loop_exit_test(back_control, loop, incr, limit, mask, cl_prob);
1470
assert(cmp != nullptr && cmp->Opcode() == Op_Cmp(bt), "no exit test");
1472
Node* phi_incr = nullptr;
1473
incr = loop_iv_incr(incr, x, loop, phi_incr);
1474
assert(incr != nullptr && incr->Opcode() == Op_Add(bt), "no incr");
1476
Node* xphi = nullptr;
1477
Node* stride = loop_iv_stride(incr, loop, xphi);
1479
assert(stride != nullptr, "no stride");
1481
PhiNode* phi = loop_iv_phi(xphi, phi_incr, x, loop);
1483
assert(phi != nullptr && phi->in(LoopNode::LoopBackControl) == incr, "No phi");
1485
jlong stride_con = stride->get_integer_as_long(bt);
1487
assert(condition_stride_ok(mask, stride_con), "illegal condition");
1489
assert(mask != BoolTest::ne, "unexpected condition");
1490
assert(phi_incr == nullptr, "bad loop shape");
1491
assert(cmp->in(1) == incr, "bad exit test shape");
1493
// Safepoint on backedge not supported
1494
assert(x->in(LoopNode::LoopBackControl)->Opcode() != Op_SafePoint, "no safepoint on backedge");
1499
// convert an int counted loop to a long counted to stress handling of
1500
// long counted loops
1501
bool PhaseIdealLoop::convert_to_long_loop(Node* cmp, Node* phi, IdealLoopTree* loop) {1502
Unique_Node_List iv_nodes;
1505
bool failed = false;
1507
for (uint i = 0; i < iv_nodes.size() && !failed; i++) {1508
Node* n = iv_nodes.at(i);
1509
switch(n->Opcode()) {1511
Node* clone = new PhiNode(n->in(0), TypeLong::LONG);
1512
old_new.map(n->_idx, clone);
1516
Node* clone = new CmpLNode(nullptr, nullptr);
1517
old_new.map(n->_idx, clone);
1521
Node* clone = new AddLNode(nullptr, nullptr);
1522
old_new.map(n->_idx, clone);
1530
DEBUG_ONLY(n->dump());
1531
fatal("unexpected");1534
for (uint i = 1; i < n->req(); i++) {1535
Node* in = n->in(i);
1536
if (in == nullptr) {1539
if (loop->is_member(get_loop(get_ctrl(in)))) {1546
for (uint i = 0; i < iv_nodes.size(); i++) {1547
Node* n = iv_nodes.at(i);
1548
Node* clone = old_new[n->_idx];
1549
if (clone != nullptr) {1550
_igvn.remove_dead_node(clone);
1556
for (uint i = 0; i < iv_nodes.size(); i++) {1557
Node* n = iv_nodes.at(i);
1558
Node* clone = old_new[n->_idx];
1559
for (uint i = 1; i < n->req(); i++) {1560
Node* in = n->in(i);
1561
if (in == nullptr) {1564
Node* in_clone = old_new[in->_idx];
1565
if (in_clone == nullptr) {1566
assert(_igvn.type(in)->isa_int(), "");
1567
in_clone = new ConvI2LNode(in);
1568
_igvn.register_new_node_with_optimizer(in_clone);
1569
set_subtree_ctrl(in_clone, false);
1571
if (in_clone->in(0) == nullptr) {1572
in_clone->set_req(0, C->top());
1573
clone->set_req(i, in_clone);
1574
in_clone->set_req(0, nullptr);
1576
clone->set_req(i, in_clone);
1579
_igvn.register_new_node_with_optimizer(clone);
1581
set_ctrl(old_new[phi->_idx], phi->in(0));
1583
for (uint i = 0; i < iv_nodes.size(); i++) {1584
Node* n = iv_nodes.at(i);
1585
Node* clone = old_new[n->_idx];
1586
set_subtree_ctrl(clone, false);
1587
Node* m = n->Opcode() == Op_CmpI ? clone : nullptr;
1588
for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {1589
Node* u = n->fast_out(i);
1590
if (iv_nodes.member(u)) {1594
m = new ConvL2INode(clone);
1595
_igvn.register_new_node_with_optimizer(m);
1596
set_subtree_ctrl(m, false);
1598
_igvn.rehash_node_delayed(u);
1599
int nb = u->replace_edge(n, m, &_igvn);
1607
//------------------------------is_counted_loop--------------------------------
1608
bool PhaseIdealLoop::is_counted_loop(Node* x, IdealLoopTree*&loop, BasicType iv_bt) {1609
PhaseGVN *gvn = &_igvn;
1611
Node* back_control = loop_exit_control(x, loop);
1612
if (back_control == nullptr) {1616
BoolTest::mask bt = BoolTest::illegal;
1618
Node* incr = nullptr;
1619
Node* limit = nullptr;
1620
Node* cmp = loop_exit_test(back_control, loop, incr, limit, bt, cl_prob);
1621
if (cmp == nullptr || cmp->Opcode() != Op_Cmp(iv_bt)) {1622
return false; // Avoid pointer & float & 64-bit compares
1625
// Trip-counter increment must be commutative & associative.
1626
if (incr->Opcode() == Op_Cast(iv_bt)) {1630
Node* phi_incr = nullptr;
1631
incr = loop_iv_incr(incr, x, loop, phi_incr);
1632
if (incr == nullptr) {1636
Node* trunc1 = nullptr;
1637
Node* trunc2 = nullptr;
1638
const TypeInteger* iv_trunc_t = nullptr;
1639
Node* orig_incr = incr;
1640
if (!(incr = CountedLoopNode::match_incr_with_optional_truncation(incr, &trunc1, &trunc2, &iv_trunc_t, iv_bt))) {1641
return false; // Funny increment opcode
1643
assert(incr->Opcode() == Op_Add(iv_bt), "wrong increment code");
1645
Node* xphi = nullptr;
1646
Node* stride = loop_iv_stride(incr, loop, xphi);
1648
if (stride == nullptr) {1652
if (xphi->Opcode() == Op_Cast(iv_bt)) {1656
// Stride must be constant
1657
jlong stride_con = stride->get_integer_as_long(iv_bt);
1658
assert(stride_con != 0, "missed some peephole opt");
1660
PhiNode* phi = loop_iv_phi(xphi, phi_incr, x, loop);
1662
if (phi == nullptr ||
1663
(trunc1 == nullptr && phi->in(LoopNode::LoopBackControl) != incr) ||
1664
(trunc1 != nullptr && phi->in(LoopNode::LoopBackControl) != trunc1)) {1668
Node* iftrue = back_control;
1669
uint iftrue_op = iftrue->Opcode();
1670
Node* iff = iftrue->in(0);
1671
BoolNode* test = iff->in(1)->as_Bool();
1673
const TypeInteger* limit_t = gvn->type(limit)->is_integer(iv_bt);
1674
if (trunc1 != nullptr) {1675
// When there is a truncation, we must be sure that after the truncation
1676
// the trip counter will end up higher than the limit, otherwise we are looking
1677
// at an endless loop. Can happen with range checks.
1682
// sum + = array[i];
1687
// If the array is shorter than 0x8000 this exits through a AIOOB
1688
// - Counted loop transformation is ok
1689
// If the array is longer then this is an endless loop
1690
// - No transformation can be done.
1692
const TypeInteger* incr_t = gvn->type(orig_incr)->is_integer(iv_bt);
1693
if (limit_t->hi_as_long() > incr_t->hi_as_long()) {1694
// if the limit can have a higher value than the increment (before the phi)
1699
Node *init_trip = phi->in(LoopNode::EntryControl);
1701
// If iv trunc type is smaller than int, check for possible wrap.
1702
if (!TypeInteger::bottom(iv_bt)->higher_equal(iv_trunc_t)) {1703
assert(trunc1 != nullptr, "must have found some truncation");
1705
// Get a better type for the phi (filtered thru if's)
1706
const TypeInteger* phi_ft = filtered_type(phi);
1708
// Can iv take on a value that will wrap?
1710
// Ensure iv's limit is not within "stride" of the wrap value.
1712
// Example for "short" type
1713
// Truncation ensures value is in the range -32768..32767 (iv_trunc_t)
1714
// If the stride is +10, then the last value of the induction
1715
// variable before the increment (phi_ft->_hi) must be
1716
// <= 32767 - 10 and (phi_ft->_lo) must be >= -32768 to
1717
// ensure no truncation occurs after the increment.
1719
if (stride_con > 0) {1720
if (iv_trunc_t->hi_as_long() - phi_ft->hi_as_long() < stride_con ||
1721
iv_trunc_t->lo_as_long() > phi_ft->lo_as_long()) {1722
return false; // truncation may occur
1724
} else if (stride_con < 0) {1725
if (iv_trunc_t->lo_as_long() - phi_ft->lo_as_long() > stride_con ||
1726
iv_trunc_t->hi_as_long() < phi_ft->hi_as_long()) {1727
return false; // truncation may occur
1730
// No possibility of wrap so truncation can be discarded
1731
// Promote iv type to Int
1733
assert(trunc1 == nullptr && trunc2 == nullptr, "no truncation for int");
1736
if (!condition_stride_ok(bt, stride_con)) {1740
const TypeInteger* init_t = gvn->type(init_trip)->is_integer(iv_bt);
1742
if (stride_con > 0) {1743
if (init_t->lo_as_long() > max_signed_integer(iv_bt) - stride_con) {1744
return false; // cyclic loop
1747
if (init_t->hi_as_long() < min_signed_integer(iv_bt) - stride_con) {1748
return false; // cyclic loop
1752
if (phi_incr != nullptr && bt != BoolTest::ne) {1753
// check if there is a possibility of IV overflowing after the first increment
1754
if (stride_con > 0) {1755
if (init_t->hi_as_long() > max_signed_integer(iv_bt) - stride_con) {1759
if (init_t->lo_as_long() < min_signed_integer(iv_bt) - stride_con) {1765
// =================================================
1766
// ---- SUCCESS! Found A Trip-Counted Loop! -----
1769
if (x->Opcode() == Op_Region) {1770
// x has not yet been transformed to Loop or LongCountedLoop.
1771
// This should only happen if we are inside an infinite loop.
1772
// It happens like this:
1773
// build_loop_tree -> do not attach infinite loop and nested loops
1774
// beautify_loops -> does not transform the infinite and nested loops to LoopNode, because not attached yet
1775
// build_loop_tree -> find and attach infinite and nested loops
1776
// counted_loop -> nested Regions are not yet transformed to LoopNodes, we land here
1777
assert(x->as_Region()->is_in_infinite_subgraph(),
1778
"x can only be a Region and not Loop if inside infinite loop");
1779
// Come back later when Region is transformed to LoopNode
1783
assert(x->Opcode() == Op_Loop || x->Opcode() == Op_LongCountedLoop, "regular loops only");
1784
C->print_method(PHASE_BEFORE_CLOOPS, 3, x);
1786
// ===================================================
1787
// We can only convert this loop to a counted loop if we can guarantee that the iv phi will never overflow at runtime.
1788
// This is an implicit assumption taken by some loop optimizations. We therefore must ensure this property at all cost.
1789
// At this point, we've already excluded some trivial cases where an overflow could have been proven statically.
1790
// But even though we cannot prove that an overflow will *not* happen, we still want to speculatively convert this loop
1791
// to a counted loop. This can be achieved by adding additional iv phi overflow checks before the loop. If they fail,
1792
// we trap and resume execution before the loop without having executed any iteration of the loop, yet.
1794
// These additional iv phi overflow checks can be inserted as Loop Limit Check Predicates above the Loop Limit Check
1795
// Parse Predicate which captures a JVM state just before the entry of the loop. If there is no such Parse Predicate,
1796
// we cannot generate a Loop Limit Check Predicate and thus cannot speculatively convert the loop to a counted loop.
1798
// In the following, we only focus on int loops with stride > 0 to keep things simple. The argumentation and proof
1799
// for stride < 0 is analogously. For long loops, we would replace max_int with max_long.
1802
// The loop to be converted does not always need to have the often used shape:
1807
// // ... equivalent i+=stride
1808
// i+=stride <==> if (i < limit)
1809
// } while (i < limit); goto loop
1813
// where the loop exit check uses the post-incremented iv phi and a '<'-operator.
1815
// We could also have '<='-operator (or '>='-operator for negative strides) or use the pre-incremented iv phi value
1816
// in the loop exit check:
1827
// Let's define the following terms:
1828
// - iv_pre_i: The pre-incremented iv phi before the i-th iteration.
1829
// - iv_post_i: The post-incremented iv phi after the i-th iteration.
1831
// The iv_pre_i and iv_post_i have the following relation:
1832
// iv_pre_i + stride = iv_post_i
1834
// When converting a loop to a counted loop, we want to have a canonicalized loop exit check of the form:
1835
// iv_post_i < adjusted_limit
1837
// If that is not the case, we need to canonicalize the loop exit check by using different values for adjusted_limit:
1838
// (LE1) iv_post_i < limit: Already canonicalized. We can directly use limit as adjusted_limit.
1839
// -> adjusted_limit = limit.
1840
// (LE2) iv_post_i <= limit:
1841
// iv_post_i < limit + 1
1842
// -> adjusted limit = limit + 1
1843
// (LE3) iv_pre_i < limit:
1844
// iv_pre_i + stride < limit + stride
1845
// iv_post_i < limit + stride
1846
// -> adjusted_limit = limit + stride
1847
// (LE4) iv_pre_i <= limit:
1848
// iv_pre_i < limit + 1
1849
// iv_pre_i + stride < limit + stride + 1
1850
// iv_post_i < limit + stride + 1
1851
// -> adjusted_limit = limit + stride + 1
1854
// (AL) limit <= adjusted_limit.
1856
// The following loop invariant has to hold for counted loops with n iterations (i.e. loop exit check true after n-th
1857
// loop iteration) and a canonicalized loop exit check to guarantee that no iv_post_i over- or underflows:
1858
// (INV) For i = 1..n, min_int <= iv_post_i <= max_int
1860
// To prove (INV), we require the following two conditions/assumptions:
1861
// (i): adjusted_limit - 1 + stride <= max_int
1862
// (ii): init < limit
1864
// If we can prove (INV), we know that there can be no over- or underflow of any iv phi value. We prove (INV) by
1865
// induction by assuming (i) and (ii).
1867
// Proof by Induction
1868
// ------------------
1869
// > Base case (i = 1): We show that (INV) holds after the first iteration:
1870
// min_int <= iv_post_1 = init + stride <= max_int
1872
// First, we note that (ii) implies
1873
// (iii) init <= limit - 1
1874
// max_int >= adjusted_limit - 1 + stride [using (i)]
1875
// >= limit - 1 + stride [using (AL)]
1876
// >= init + stride [using (iii)]
1877
// >= min_int [using stride > 0, no underflow]
1878
// Thus, no overflow happens after the first iteration and (INV) holds for i = 1.
1880
// Note that to prove the base case we need (i) and (ii).
1882
// > Induction Hypothesis (i = j, j > 1): Assume that (INV) holds after the j-th iteration:
1883
// min_int <= iv_post_j <= max_int
1884
// > Step case (i = j + 1): We show that (INV) also holds after the j+1-th iteration:
1885
// min_int <= iv_post_{j+1} = iv_post_j + stride <= max_int1887
// If iv_post_j >= adjusted_limit:
1888
// We exit the loop after the j-th iteration, and we don't execute the j+1-th iteration anymore. Thus, there is
1889
// also no iv_{j+1}. Since (INV) holds for iv_j, there is nothing left to prove.1890
// If iv_post_j < adjusted_limit:
1891
// First, we note that:
1892
// (iv) iv_post_j <= adjusted_limit - 1
1893
// max_int >= adjusted_limit - 1 + stride [using (i)]
1894
// >= iv_post_j + stride [using (iv)]
1895
// >= min_int [using stride > 0, no underflow]
1897
// Note that to prove the step case we only need (i).
1899
// Thus, by assuming (i) and (ii), we proved (INV).
1902
// It is therefore enough to add the following two Loop Limit Check Predicates to check assumptions (i) and (ii):
1904
// (1) Loop Limit Check Predicate for (i):
1905
// Using (i): adjusted_limit - 1 + stride <= max_int
1907
// This condition is now restated to use limit instead of adjusted_limit:
1909
// To prevent an overflow of adjusted_limit -1 + stride itself, we rewrite this check to
1910
// max_int - stride + 1 >= adjusted_limit
1911
// We can merge the two constants into
1912
// canonicalized_correction = stride - 1
1914
// max_int - canonicalized_correction >= adjusted_limit
1916
// To directly use limit instead of adjusted_limit in the predicate condition, we split adjusted_limit into:
1917
// adjusted_limit = limit + limit_correction
1918
// Since stride > 0 and limit_correction <= stride + 1, we can restate this with no over- or underflow into:
1919
// max_int - canonicalized_correction - limit_correction >= limit
1920
// Since canonicalized_correction and limit_correction are both constants, we can replace them with a new constant:
1921
// final_correction = canonicalized_correction + limit_correction
1924
// Final predicate condition:
1925
// max_int - final_correction >= limit
1927
// (2) Loop Limit Check Predicate for (ii):
1928
// Using (ii): init < limit
1930
// This Loop Limit Check Predicate is not required if we can prove at compile time that either:
1931
// (2.1) type(init) < type(limit)
1932
// In this case, we know:
1933
// all possible values of init < all possible values of limit
1934
// and we can skip the predicate.
1936
// (2.2) init < limit is already checked before (i.e. found as a dominating check)
1937
// In this case, we do not need to re-check the condition and can skip the predicate.
1938
// This is often found for while- and for-loops which have the following shape:
1940
// if (init < limit) { // Dominating test. Do not need the Loop Limit Check Predicate below.1942
// if (init >= limit) { trap(); } // Here we would insert the Loop Limit Check Predicate1945
// } while (i < limit);
1948
// (2.3) init + stride <= max_int
1949
// In this case, there is no overflow of the iv phi after the first loop iteration.
1950
// In the proof of the base case above we showed that init + stride <= max_int by using assumption (ii):
1952
// In the proof of the step case above, we did not need (ii) anymore. Therefore, if we already know at
1953
// compile time that init + stride <= max_int then we have trivially proven the base case and that
1954
// there is no overflow of the iv phi after the first iteration. In this case, we don't need to check (ii)
1955
// again and can skip the predicate.
1958
// Accounting for (LE3) and (LE4) where we use pre-incremented phis in the loop exit check.
1959
const jlong limit_correction_for_pre_iv_exit_check = (phi_incr != nullptr) ? stride_con : 0;
1961
// Accounting for (LE2) and (LE4) where we use <= or >= in the loop exit check.
1962
const bool includes_limit = (bt == BoolTest::le || bt == BoolTest::ge);
1963
const jlong limit_correction_for_le_ge_exit_check = (includes_limit ? (stride_con > 0 ? 1 : -1) : 0);
1965
const jlong limit_correction = limit_correction_for_pre_iv_exit_check + limit_correction_for_le_ge_exit_check;
1966
const jlong canonicalized_correction = stride_con + (stride_con > 0 ? -1 : 1);
1967
const jlong final_correction = canonicalized_correction + limit_correction;
1969
int sov = check_stride_overflow(final_correction, limit_t, iv_bt);
1970
Node* init_control = x->in(LoopNode::EntryControl);
1972
// If sov==0, limit's type always satisfies the condition, for
1973
// example, when it is an array length.
1976
return false; // Bailout: integer overflow is certain.
1978
// (1) Loop Limit Check Predicate is required because we could not statically prove that
1979
// limit + final_correction = adjusted_limit - 1 + stride <= max_int
1980
assert(!x->as_Loop()->is_loop_nest_inner_loop(), "loop was transformed");
1981
const Predicates predicates(init_control);
1982
const PredicateBlock* loop_limit_check_predicate_block = predicates.loop_limit_check_predicate_block();
1983
if (!loop_limit_check_predicate_block->has_parse_predicate()) {1984
// The Loop Limit Check Parse Predicate is not generated if this method trapped here before.
1986
if (TraceLoopLimitCheck) {1987
tty->print("Missing Loop Limit Check Parse Predicate:");1995
ParsePredicateNode* loop_limit_check_parse_predicate = loop_limit_check_predicate_block->parse_predicate();
1996
if (!is_dominator(get_ctrl(limit), loop_limit_check_parse_predicate->in(0))) {2003
if (stride_con > 0) {2004
cmp_limit = CmpNode::make(limit, _igvn.integercon(max_signed_integer(iv_bt) - final_correction, iv_bt), iv_bt);
2005
bol = new BoolNode(cmp_limit, BoolTest::le);
2007
cmp_limit = CmpNode::make(limit, _igvn.integercon(min_signed_integer(iv_bt) - final_correction, iv_bt), iv_bt);
2008
bol = new BoolNode(cmp_limit, BoolTest::ge);
2011
insert_loop_limit_check_predicate(init_control->as_IfTrue(), cmp_limit, bol);
2015
const bool init_plus_stride_could_overflow =
2016
(stride_con > 0 && init_t->hi_as_long() > max_signed_integer(iv_bt) - stride_con) ||
2017
(stride_con < 0 && init_t->lo_as_long() < min_signed_integer(iv_bt) - stride_con);
2019
const bool init_gte_limit = (stride_con > 0 && init_t->hi_as_long() >= limit_t->lo_as_long()) ||
2020
(stride_con < 0 && init_t->lo_as_long() <= limit_t->hi_as_long());
2022
if (init_gte_limit && // (2.1)
2023
((bt == BoolTest::ne || init_plus_stride_could_overflow) && // (2.3)
2024
!has_dominating_loop_limit_check(init_trip, limit, stride_con, iv_bt, init_control))) { // (2.2)2025
// (2) Iteration Loop Limit Check Predicate is required because neither (2.1), (2.2), nor (2.3) holds.
2026
// We use the following condition:
2027
// - stride > 0: init < limit
2028
// - stride < 0: init > limit
2030
// This predicate is always required if we have a non-equal-operator in the loop exit check (where stride = 1 is
2031
// a requirement). We transform the loop exit check by using a less-than-operator. By doing so, we must always
2032
// check that init < limit. Otherwise, we could have a different number of iterations at runtime.
2034
const Predicates predicates(init_control);
2035
const PredicateBlock* loop_limit_check_predicate_block = predicates.loop_limit_check_predicate_block();
2036
if (!loop_limit_check_predicate_block->has_parse_predicate()) {2037
// The Loop Limit Check Parse Predicate is not generated if this method trapped here before.
2039
if (TraceLoopLimitCheck) {2040
tty->print("Missing Loop Limit Check Parse Predicate:");2048
ParsePredicateNode* loop_limit_check_parse_predicate = loop_limit_check_predicate_block->parse_predicate();
2049
Node* parse_predicate_entry = loop_limit_check_parse_predicate->in(0);
2050
if (!is_dominator(get_ctrl(limit), parse_predicate_entry) ||
2051
!is_dominator(get_ctrl(init_trip), parse_predicate_entry)) {2058
if (stride_con > 0) {2059
cmp_limit = CmpNode::make(init_trip, limit, iv_bt);
2060
bol = new BoolNode(cmp_limit, BoolTest::lt);
2062
cmp_limit = CmpNode::make(init_trip, limit, iv_bt);
2063
bol = new BoolNode(cmp_limit, BoolTest::gt);
2066
insert_loop_limit_check_predicate(init_control->as_IfTrue(), cmp_limit, bol);
2069
if (bt == BoolTest::ne) {2070
// Now we need to canonicalize the loop condition if it is 'ne'.
2071
assert(stride_con == 1 || stride_con == -1, "simple increment only - checked before");
2072
if (stride_con > 0) {2073
// 'ne' can be replaced with 'lt' only when init < limit. This is ensured by the inserted predicate above.
2076
assert(stride_con < 0, "must be");
2077
// 'ne' can be replaced with 'gt' only when init > limit. This is ensured by the inserted predicate above.
2082
Node* sfpt = nullptr;
2083
if (loop->_child == nullptr) {2084
sfpt = find_safepoint(back_control, x, loop);
2087
if (sfpt->Opcode() != Op_SafePoint) {2092
if (x->in(LoopNode::LoopBackControl)->Opcode() == Op_SafePoint) {2093
Node* backedge_sfpt = x->in(LoopNode::LoopBackControl);
2094
if (((iv_bt == T_INT && LoopStripMiningIter != 0) ||
2097
// Leaving the safepoint on the backedge and creating a
2098
// CountedLoop will confuse optimizations. We can't move the
2099
// safepoint around because its jvm state wouldn't match a new
2100
// location. Give up on that loop.
2103
if (is_deleteable_safept(backedge_sfpt)) {2104
lazy_replace(backedge_sfpt, iftrue);
2105
if (loop->_safepts != nullptr) {2106
loop->_safepts->yank(backedge_sfpt);
2108
loop->_tail = iftrue;
2114
if (iv_bt == T_INT &&
2115
!x->as_Loop()->is_loop_nest_inner_loop() &&
2116
StressLongCountedLoop > 0 &&
2117
trunc1 == nullptr &&
2118
convert_to_long_loop(cmp, phi, loop)) {2123
Node* adjusted_limit = limit;
2124
if (phi_incr != nullptr) {2125
// If compare points directly to the phi we need to adjust
2126
// the compare so that it points to the incr. Limit have
2127
// to be adjusted to keep trip count the same and we
2128
// should avoid int overflow.
2130
// i = init; do {} while(i++ < limit);2132
// i = init; do {} while(++i < limit+1);2134
adjusted_limit = gvn->transform(AddNode::make(limit, stride, iv_bt));
2137
if (includes_limit) {2138
// The limit check guaranties that 'limit <= (max_jint - stride)' so
2139
// we can convert 'i <= limit' to 'i < limit+1' since stride != 0.
2141
Node* one = (stride_con > 0) ? gvn->integercon( 1, iv_bt) : gvn->integercon(-1, iv_bt);
2142
adjusted_limit = gvn->transform(AddNode::make(adjusted_limit, one, iv_bt));
2143
if (bt == BoolTest::le)
2145
else if (bt == BoolTest::ge)
2148
ShouldNotReachHere();
2150
set_subtree_ctrl(adjusted_limit, false);
2152
// Build a canonical trip test.
2153
// Clone code, as old values may be in use.
2154
incr = incr->clone();
2155
incr->set_req(1,phi);
2156
incr->set_req(2,stride);
2157
incr = _igvn.register_new_node_with_optimizer(incr);
2158
set_early_ctrl(incr, false);
2159
_igvn.rehash_node_delayed(phi);
2160
phi->set_req_X( LoopNode::LoopBackControl, incr, &_igvn );
2162
// If phi type is more restrictive than Int, raise to
2163
// Int to prevent (almost) infinite recursion in igvn
2164
// which can only handle integer types for constants or minint..maxint.
2165
if (!TypeInteger::bottom(iv_bt)->higher_equal(phi->bottom_type())) {2166
Node* nphi = PhiNode::make(phi->in(0), phi->in(LoopNode::EntryControl), TypeInteger::bottom(iv_bt));
2167
nphi->set_req(LoopNode::LoopBackControl, phi->in(LoopNode::LoopBackControl));
2168
nphi = _igvn.register_new_node_with_optimizer(nphi);
2169
set_ctrl(nphi, get_ctrl(phi));
2170
_igvn.replace_node(phi, nphi);
2171
phi = nphi->as_Phi();
2174
cmp->set_req(1,incr);
2175
cmp->set_req(2, adjusted_limit);
2176
cmp = _igvn.register_new_node_with_optimizer(cmp);
2177
set_ctrl(cmp, iff->in(0));
2179
test = test->clone()->as_Bool();
2180
(*(BoolTest*)&test->_test)._test = bt;
2181
test->set_req(1,cmp);
2182
_igvn.register_new_node_with_optimizer(test);
2183
set_ctrl(test, iff->in(0));
2185
// Replace the old IfNode with a new LoopEndNode
2186
Node *lex = _igvn.register_new_node_with_optimizer(BaseCountedLoopEndNode::make(iff->in(0), test, cl_prob, iff->as_If()->_fcnt, iv_bt));
2187
IfNode *le = lex->as_If();
2188
uint dd = dom_depth(iff);
2189
set_idom(le, le->in(0), dd); // Update dominance for loop exit
2192
// Get the loop-exit control
2193
Node *iffalse = iff->as_If()->proj_out(!(iftrue_op == Op_IfTrue));
2195
// Need to swap loop-exit and loop-back control?
2196
if (iftrue_op == Op_IfFalse) {2197
Node *ift2=_igvn.register_new_node_with_optimizer(new IfTrueNode (le));
2198
Node *iff2=_igvn.register_new_node_with_optimizer(new IfFalseNode(le));
2200
loop->_tail = back_control = ift2;
2201
set_loop(ift2, loop);
2202
set_loop(iff2, get_loop(iffalse));
2204
// Lazy update of 'get_ctrl' mechanism.
2205
lazy_replace(iffalse, iff2);
2206
lazy_replace(iftrue, ift2);
2212
_igvn.rehash_node_delayed(iffalse);
2213
_igvn.rehash_node_delayed(iftrue);
2214
iffalse->set_req_X( 0, le, &_igvn );
2215
iftrue ->set_req_X( 0, le, &_igvn );
2218
set_idom(iftrue, le, dd+1);
2219
set_idom(iffalse, le, dd+1);
2220
assert(iff->outcnt() == 0, "should be dead now");
2221
lazy_replace( iff, le ); // fix 'get_ctrl'
2223
Node* entry_control = init_control;
2224
bool strip_mine_loop = iv_bt == T_INT &&
2225
loop->_child == nullptr &&
2228
is_deleteable_safept(sfpt);
2229
IdealLoopTree* outer_ilt = nullptr;
2230
if (strip_mine_loop) {2231
outer_ilt = create_outer_strip_mined_loop(test, cmp, init_control, loop,
2232
cl_prob, le->_fcnt, entry_control,
2236
// Now setup a new CountedLoopNode to replace the existing LoopNode
2237
BaseCountedLoopNode *l = BaseCountedLoopNode::make(entry_control, back_control, iv_bt);
2238
l->set_unswitch_count(x->as_Loop()->unswitch_count()); // Preserve
2239
// The following assert is approximately true, and defines the intention
2240
// of can_be_counted_loop. It fails, however, because phase->type
2241
// is not yet initialized for this loop and its parts.
2242
//assert(l->can_be_counted_loop(this), "sanity");
2243
_igvn.register_new_node_with_optimizer(l);
2246
// Fix all data nodes placed at the old loop head.
2247
// Uses the lazy-update mechanism of 'get_ctrl'.
2248
lazy_replace( x, l );
2249
set_idom(l, entry_control, dom_depth(entry_control) + 1);
2251
if (iv_bt == T_INT && (LoopStripMiningIter == 0 || strip_mine_loop)) {2252
// Check for immediately preceding SafePoint and remove
2253
if (sfpt != nullptr && (strip_mine_loop || is_deleteable_safept(sfpt))) {2254
if (strip_mine_loop) {2255
Node* outer_le = outer_ilt->_tail->in(0);
2256
Node* sfpt_clone = sfpt->clone();
2257
sfpt_clone->set_req(0, iffalse);
2258
outer_le->set_req(0, sfpt_clone);
2260
Node* polladdr = sfpt_clone->in(TypeFunc::Parms);
2261
if (polladdr != nullptr && polladdr->is_Load()) {2262
// Polling load should be pinned outside inner loop.
2263
Node* new_polladdr = polladdr->clone();
2264
new_polladdr->set_req(0, iffalse);
2265
_igvn.register_new_node_with_optimizer(new_polladdr, polladdr);
2266
set_ctrl(new_polladdr, iffalse);
2267
sfpt_clone->set_req(TypeFunc::Parms, new_polladdr);
2269
// When this code runs, loop bodies have not yet been populated.
2270
const bool body_populated = false;
2271
register_control(sfpt_clone, outer_ilt, iffalse, body_populated);
2272
set_idom(outer_le, sfpt_clone, dom_depth(sfpt_clone));
2274
lazy_replace(sfpt, sfpt->in(TypeFunc::Control));
2275
if (loop->_safepts != nullptr) {2276
loop->_safepts->yank(sfpt);
2282
assert(l->is_valid_counted_loop(iv_bt), "counted loop shape is messed up");
2283
assert(l == loop->_head && l->phi() == phi && l->loopexit_or_null() == lex, "" );
2286
if (TraceLoopOpts) {2287
tty->print("Counted ");2292
C->print_method(PHASE_AFTER_CLOOPS, 3, l);
2294
// Capture bounds of the loop in the induction variable Phi before
2295
// subsequent transformation (iteration splitting) obscures the
2297
l->phi()->as_Phi()->set_type(l->phi()->Value(&_igvn));
2299
if (strip_mine_loop) {2300
l->mark_strip_mined();
2301
l->verify_strip_mined(1);
2302
outer_ilt->_head->as_Loop()->verify_strip_mined(1);
2307
if (x->as_Loop()->is_loop_nest_inner_loop() && iv_bt == T_LONG) {2308
Atomic::inc(&_long_loop_counted_loops);
2311
if (iv_bt == T_LONG && x->as_Loop()->is_loop_nest_outer_loop()) {2312
l->mark_loop_nest_outer_loop();
2318
// Check if there is a dominating loop limit check of the form 'init < limit' starting at the loop entry.
2319
// If there is one, then we do not need to create an additional Loop Limit Check Predicate.
2320
bool PhaseIdealLoop::has_dominating_loop_limit_check(Node* init_trip, Node* limit, const jlong stride_con,
2321
const BasicType iv_bt, Node* loop_entry) {2322
// Eagerly call transform() on the Cmp and Bool node to common them up if possible. This is required in order to
2323
// successfully find a dominated test with the If node below.
2326
if (stride_con > 0) {2327
cmp_limit = _igvn.transform(CmpNode::make(init_trip, limit, iv_bt));
2328
bol = _igvn.transform(new BoolNode(cmp_limit, BoolTest::lt));
2330
cmp_limit = _igvn.transform(CmpNode::make(init_trip, limit, iv_bt));
2331
bol = _igvn.transform(new BoolNode(cmp_limit, BoolTest::gt));
2334
// Check if there is already a dominating init < limit check. If so, we do not need a Loop Limit Check Predicate.
2335
IfNode* iff = new IfNode(loop_entry, bol, PROB_MIN, COUNT_UNKNOWN);
2336
// Also add fake IfProj nodes in order to call transform() on the newly created IfNode.
2337
IfFalseNode* if_false = new IfFalseNode(iff);
2338
IfTrueNode* if_true = new IfTrueNode(iff);
2339
Node* dominated_iff = _igvn.transform(iff);
2340
// ConI node? Found dominating test (IfNode::dominated_by() returns a ConI node).
2341
const bool found_dominating_test = dominated_iff != nullptr && dominated_iff->is_ConI();
2343
// Kill the If with its projections again in the next IGVN round by cutting it off from the graph.
2344
_igvn.replace_input_of(iff, 0, C->top());
2345
_igvn.replace_input_of(iff, 1, C->top());
2346
return found_dominating_test;
2349
//----------------------exact_limit-------------------------------------------
2350
Node* PhaseIdealLoop::exact_limit( IdealLoopTree *loop ) {2351
assert(loop->_head->is_CountedLoop(), "");
2352
CountedLoopNode *cl = loop->_head->as_CountedLoop();
2353
assert(cl->is_valid_counted_loop(T_INT), "");
2355
if (cl->stride_con() == 1 ||
2356
cl->stride_con() == -1 ||
2357
cl->limit()->Opcode() == Op_LoopLimit) {2358
// Old code has exact limit (it could be incorrect in case of int overflow).
2359
// Loop limit is exact with stride == 1. And loop may already have exact limit.
2362
Node *limit = nullptr;
2364
BoolTest::mask bt = cl->loopexit()->test_trip();
2365
assert(bt == BoolTest::lt || bt == BoolTest::gt, "canonical test is expected");
2367
if (cl->has_exact_trip_count()) {2368
// Simple case: loop has constant boundaries.
2369
// Use jlongs to avoid integer overflow.
2370
int stride_con = cl->stride_con();
2371
jlong init_con = cl->init_trip()->get_int();
2372
jlong limit_con = cl->limit()->get_int();
2373
julong trip_cnt = cl->trip_count();
2374
jlong final_con = init_con + trip_cnt*stride_con;
2375
int final_int = (int)final_con;
2376
// The final value should be in integer range since the loop
2377
// is counted and the limit was checked for overflow.
2378
assert(final_con == (jlong)final_int, "final value should be integer");
2379
limit = _igvn.intcon(final_int);
2381
// Create new LoopLimit node to get exact limit (final iv value).
2382
limit = new LoopLimitNode(C, cl->init_trip(), cl->limit(), cl->stride());
2383
register_new_node(limit, cl->in(LoopNode::EntryControl));
2385
assert(limit != nullptr, "sanity");
2389
//------------------------------Ideal------------------------------------------
2390
// Return a node which is more "ideal" than the current node.
2391
// Attempt to convert into a counted-loop.
2392
Node *LoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {2393
if (!can_be_counted_loop(phase) && !is_OuterStripMinedLoop()) {2394
phase->C->set_major_progress();
2396
return RegionNode::Ideal(phase, can_reshape);
2400
void LoopNode::verify_strip_mined(int expect_skeleton) const {2401
const OuterStripMinedLoopNode* outer = nullptr;
2402
const CountedLoopNode* inner = nullptr;
2403
if (is_strip_mined()) {2404
if (!is_valid_counted_loop(T_INT)) {2405
return; // Skip malformed counted loop
2407
assert(is_CountedLoop(), "no Loop should be marked strip mined");
2408
inner = as_CountedLoop();
2409
outer = inner->in(LoopNode::EntryControl)->as_OuterStripMinedLoop();
2410
} else if (is_OuterStripMinedLoop()) {2411
outer = this->as_OuterStripMinedLoop();
2412
inner = outer->unique_ctrl_out()->as_CountedLoop();
2413
assert(inner->is_valid_counted_loop(T_INT) && inner->is_strip_mined(), "OuterStripMinedLoop should have been removed");
2414
assert(!is_strip_mined(), "outer loop shouldn't be marked strip mined");
2416
if (inner != nullptr || outer != nullptr) {2417
assert(inner != nullptr && outer != nullptr, "missing loop in strip mined nest");
2418
Node* outer_tail = outer->in(LoopNode::LoopBackControl);
2419
Node* outer_le = outer_tail->in(0);
2420
assert(outer_le->Opcode() == Op_OuterStripMinedLoopEnd, "tail of outer loop should be an If");
2421
Node* sfpt = outer_le->in(0);
2422
assert(sfpt->Opcode() == Op_SafePoint, "where's the safepoint?");
2423
Node* inner_out = sfpt->in(0);
2424
CountedLoopEndNode* cle = inner_out->in(0)->as_CountedLoopEnd();
2425
assert(cle == inner->loopexit_or_null(), "mismatch");
2426
bool has_skeleton = outer_le->in(1)->bottom_type()->singleton() && outer_le->in(1)->bottom_type()->is_int()->get_con() == 0;
2428
assert(expect_skeleton == 1 || expect_skeleton == -1, "unexpected skeleton node");
2429
assert(outer->outcnt() == 2, "only control nodes");
2431
assert(expect_skeleton == 0 || expect_skeleton == -1, "no skeleton node?");
2434
Node* be = inner->in(LoopNode::LoopBackControl);
2435
for (DUIterator_Fast imax, i = inner->fast_outs(imax); i < imax; i++) {2436
Node* u = inner->fast_out(i);
2439
for (DUIterator_Fast jmax, j = be->fast_outs(jmax); j < jmax; j++) {2440
Node* n = be->fast_out(j);
2442
assert(n->in(0) == be || n->find_prec_edge(be) > 0, "should be on the backedge");
2445
} while (!n->is_Phi());
2454
assert(be_loads <= phis, "wrong number phis that depends on a pinned load");
2455
for (DUIterator_Fast imax, i = outer->fast_outs(imax); i < imax; i++) {2456
Node* u = outer->fast_out(i);
2457
assert(u == outer || u == inner || u->is_Phi(), "nothing between inner and outer loop");
2460
for (DUIterator_Fast imax, i = inner_out->fast_outs(imax); i < imax; i++) {2461
Node* u = inner_out->fast_out(i);
2462
if (u->is_Store()) {2466
// Late optimization of loads on backedge can cause Phi of outer loop to be eliminated but Phi of inner loop is
2467
// not guaranteed to be optimized out.
2468
assert(outer->outcnt() >= phis + 2 - be_loads && outer->outcnt() <= phis + 2 + stores + 1, "only phis");
2470
assert(sfpt->outcnt() == 1, "no data node");
2471
assert(outer_tail->outcnt() == 1 || !has_skeleton, "no data node");
2476
//=============================================================================
2477
//------------------------------Ideal------------------------------------------
2478
// Return a node which is more "ideal" than the current node.
2479
// Attempt to convert into a counted-loop.
2480
Node *CountedLoopNode::Ideal(PhaseGVN *phase, bool can_reshape) {2481
return RegionNode::Ideal(phase, can_reshape);
2484
//------------------------------dump_spec--------------------------------------
2485
// Dump special per-node info
2487
void CountedLoopNode::dump_spec(outputStream *st) const {2488
LoopNode::dump_spec(st);
2489
if (stride_is_con()) {2490
st->print("stride: %d ",stride_con());2492
if (is_pre_loop ()) st->print("pre of N%d" , _main_idx);2493
if (is_main_loop()) st->print("main of N%d", _idx);2494
if (is_post_loop()) st->print("post of N%d", _main_idx);2495
if (is_strip_mined()) st->print(" strip mined");2499
//=============================================================================
2500
jlong BaseCountedLoopEndNode::stride_con() const {2501
return stride()->bottom_type()->is_integer(bt())->get_con_as_long(bt());
2505
BaseCountedLoopEndNode* BaseCountedLoopEndNode::make(Node* control, Node* test, float prob, float cnt, BasicType bt) {2507
return new CountedLoopEndNode(control, test, prob, cnt);
2509
assert(bt == T_LONG, "unsupported");
2510
return new LongCountedLoopEndNode(control, test, prob, cnt);
2513
//=============================================================================
2514
//------------------------------Value-----------------------------------------
2515
const Type* LoopLimitNode::Value(PhaseGVN* phase) const {2516
const Type* init_t = phase->type(in(Init));
2517
const Type* limit_t = phase->type(in(Limit));
2518
const Type* stride_t = phase->type(in(Stride));
2519
// Either input is TOP ==> the result is TOP
2520
if (init_t == Type::TOP) return Type::TOP;
2521
if (limit_t == Type::TOP) return Type::TOP;
2522
if (stride_t == Type::TOP) return Type::TOP;
2524
int stride_con = stride_t->is_int()->get_con();
2525
if (stride_con == 1)
2526
return bottom_type(); // Identity
2528
if (init_t->is_int()->is_con() && limit_t->is_int()->is_con()) {2529
// Use jlongs to avoid integer overflow.
2530
jlong init_con = init_t->is_int()->get_con();
2531
jlong limit_con = limit_t->is_int()->get_con();
2532
int stride_m = stride_con - (stride_con > 0 ? 1 : -1);
2533
jlong trip_count = (limit_con - init_con + stride_m)/stride_con;
2534
jlong final_con = init_con + stride_con*trip_count;
2535
int final_int = (int)final_con;
2536
// The final value should be in integer range since the loop
2537
// is counted and the limit was checked for overflow.
2538
// Assert checks for overflow only if all input nodes are ConINodes, as during CCP
2539
// there might be a temporary overflow from PhiNodes see JDK-8309266
2540
assert((in(Init)->is_ConI() && in(Limit)->is_ConI() && in(Stride)->is_ConI()) ? final_con == (jlong)final_int : true, "final value should be integer");
2541
if (final_con == (jlong)final_int) {2542
return TypeInt::make(final_int);
2544
return bottom_type();
2548
return bottom_type(); // TypeInt::INT
2551
//------------------------------Ideal------------------------------------------
2552
// Return a node which is more "ideal" than the current node.
2553
Node *LoopLimitNode::Ideal(PhaseGVN *phase, bool can_reshape) {2554
if (phase->type(in(Init)) == Type::TOP ||
2555
phase->type(in(Limit)) == Type::TOP ||
2556
phase->type(in(Stride)) == Type::TOP)
2557
return nullptr; // Dead
2559
int stride_con = phase->type(in(Stride))->is_int()->get_con();
2560
if (stride_con == 1)
2561
return nullptr; // Identity
2563
if (in(Init)->is_Con() && in(Limit)->is_Con())
2564
return nullptr; // Value
2566
// Delay following optimizations until all loop optimizations
2567
// done to keep Ideal graph simple.
2568
if (!can_reshape || !phase->C->post_loop_opts_phase()) {2572
const TypeInt* init_t = phase->type(in(Init) )->is_int();
2573
const TypeInt* limit_t = phase->type(in(Limit))->is_int();
2577
if (stride_con > 0) {2578
stride_p = stride_con;
2581
max = (julong)max_jint;
2583
stride_p = -stride_con;
2586
max = (julong)min_jint;
2588
julong range = lim - ini + stride_p;
2590
// Convert to integer expression if it is not overflow.
2591
Node* stride_m = phase->intcon(stride_con - (stride_con > 0 ? 1 : -1));
2592
Node *range = phase->transform(new SubINode(in(Limit), in(Init)));
2593
Node *bias = phase->transform(new AddINode(range, stride_m));
2594
Node *trip = phase->transform(new DivINode(nullptr, bias, in(Stride)));
2595
Node *span = phase->transform(new MulINode(trip, in(Stride)));
2596
return new AddINode(span, in(Init)); // exact limit
2599
if (is_power_of_2(stride_p) || // divisor is 2^n
2600
!Matcher::has_match_rule(Op_LoopLimit)) { // or no specialized Mach node?2601
// Convert to long expression to avoid integer overflow
2602
// and let igvn optimizer convert this division.
2604
Node* init = phase->transform( new ConvI2LNode(in(Init)));
2605
Node* limit = phase->transform( new ConvI2LNode(in(Limit)));
2606
Node* stride = phase->longcon(stride_con);
2607
Node* stride_m = phase->longcon(stride_con - (stride_con > 0 ? 1 : -1));
2609
Node *range = phase->transform(new SubLNode(limit, init));
2610
Node *bias = phase->transform(new AddLNode(range, stride_m));
2612
if (stride_con > 0 && is_power_of_2(stride_p)) {2613
// bias >= 0 if stride >0, so if stride is 2^n we can use &(-stride)
2614
// and avoid generating rounding for division. Zero trip guard should
2615
// guarantee that init < limit but sometimes the guard is missing and
2616
// we can get situation when init > limit. Note, for the empty loop
2617
// optimization zero trip guard is generated explicitly which leaves
2618
// only RCE predicate where exact limit is used and the predicate
2619
// will simply fail forcing recompilation.
2620
Node* neg_stride = phase->longcon(-stride_con);
2621
span = phase->transform(new AndLNode(bias, neg_stride));
2623
Node *trip = phase->transform(new DivLNode(nullptr, bias, stride));
2624
span = phase->transform(new MulLNode(trip, stride));
2626
// Convert back to int
2627
Node *span_int = phase->transform(new ConvL2INode(span));
2628
return new AddINode(span_int, in(Init)); // exact limit
2631
return nullptr; // No progress
2634
//------------------------------Identity---------------------------------------
2635
// If stride == 1 return limit node.
2636
Node* LoopLimitNode::Identity(PhaseGVN* phase) {2637
int stride_con = phase->type(in(Stride))->is_int()->get_con();
2638
if (stride_con == 1 || stride_con == -1)
2643
//=============================================================================
2644
//----------------------match_incr_with_optional_truncation--------------------
2645
// Match increment with optional truncation:
2646
// CHAR: (i+1)&0x7fff, BYTE: ((i+1)<<8)>>8, or SHORT: ((i+1)<<16)>>16
2647
// Return null for failure. Success returns the increment node.
2648
Node* CountedLoopNode::match_incr_with_optional_truncation(Node* expr, Node** trunc1, Node** trunc2,
2649
const TypeInteger** trunc_type,
2652
if (expr == nullptr || expr->req() != 3) return nullptr;
2657
int n1op = n1->Opcode();
2658
const TypeInteger* trunc_t = TypeInteger::bottom(bt);
2661
// Try to strip (n1 & M) or (n1 << N >> N) from n1.
2662
if (n1op == Op_AndI &&
2663
n1->in(2)->is_Con() &&
2664
n1->in(2)->bottom_type()->is_int()->get_con() == 0x7fff) {2665
// %%% This check should match any mask of 2**K-1.
2668
n1op = n1->Opcode();
2669
trunc_t = TypeInt::CHAR;
2670
} else if (n1op == Op_RShiftI &&
2671
n1->in(1) != nullptr &&
2672
n1->in(1)->Opcode() == Op_LShiftI &&
2673
n1->in(2) == n1->in(1)->in(2) &&
2674
n1->in(2)->is_Con()) {2675
jint shift = n1->in(2)->bottom_type()->is_int()->get_con();
2676
// %%% This check should match any shift in [1..31].
2677
if (shift == 16 || shift == 8) {2681
n1op = n1->Opcode();
2683
trunc_t = TypeInt::SHORT;
2684
} else if (shift == 8) {2685
trunc_t = TypeInt::BYTE;
2691
// If (maybe after stripping) it is an AddI, we won:
2692
if (n1op == Op_Add(bt)) {2695
*trunc_type = trunc_t;
2703
LoopNode* CountedLoopNode::skip_strip_mined(int expect_skeleton) {2704
if (is_strip_mined() && in(EntryControl) != nullptr && in(EntryControl)->is_OuterStripMinedLoop()) {2705
verify_strip_mined(expect_skeleton);
2706
return in(EntryControl)->as_Loop();
2711
OuterStripMinedLoopNode* CountedLoopNode::outer_loop() const {2712
assert(is_strip_mined(), "not a strip mined loop");
2713
Node* c = in(EntryControl);
2714
if (c == nullptr || c->is_top() || !c->is_OuterStripMinedLoop()) {2717
return c->as_OuterStripMinedLoop();
2720
IfTrueNode* OuterStripMinedLoopNode::outer_loop_tail() const {2721
Node* c = in(LoopBackControl);
2722
if (c == nullptr || c->is_top()) {2725
return c->as_IfTrue();
2728
IfTrueNode* CountedLoopNode::outer_loop_tail() const {2729
LoopNode* l = outer_loop();
2733
return l->outer_loop_tail();
2736
OuterStripMinedLoopEndNode* OuterStripMinedLoopNode::outer_loop_end() const {2737
IfTrueNode* proj = outer_loop_tail();
2738
if (proj == nullptr) {2741
Node* c = proj->in(0);
2742
if (c == nullptr || c->is_top() || c->outcnt() != 2) {2745
return c->as_OuterStripMinedLoopEnd();
2748
OuterStripMinedLoopEndNode* CountedLoopNode::outer_loop_end() const {2749
LoopNode* l = outer_loop();
2753
return l->outer_loop_end();
2756
IfFalseNode* OuterStripMinedLoopNode::outer_loop_exit() const {2757
IfNode* le = outer_loop_end();
2758
if (le == nullptr) {2761
Node* c = le->proj_out_or_null(false);
2765
return c->as_IfFalse();
2768
IfFalseNode* CountedLoopNode::outer_loop_exit() const {2769
LoopNode* l = outer_loop();
2773
return l->outer_loop_exit();
2776
SafePointNode* OuterStripMinedLoopNode::outer_safepoint() const {2777
IfNode* le = outer_loop_end();
2778
if (le == nullptr) {2781
Node* c = le->in(0);
2782
if (c == nullptr || c->is_top()) {2785
assert(c->Opcode() == Op_SafePoint, "broken outer loop");
2786
return c->as_SafePoint();
2789
SafePointNode* CountedLoopNode::outer_safepoint() const {2790
LoopNode* l = outer_loop();
2794
return l->outer_safepoint();
2797
Node* CountedLoopNode::skip_assertion_predicates_with_halt() {2798
Node* ctrl = in(LoopNode::EntryControl);
2799
if (is_main_loop()) {2800
ctrl = skip_strip_mined()->in(LoopNode::EntryControl);
2802
if (is_main_loop() || is_post_loop()) {2803
AssertionPredicatesWithHalt assertion_predicates(ctrl);
2804
return assertion_predicates.entry();
2810
int CountedLoopNode::stride_con() const {2811
CountedLoopEndNode* cle = loopexit_or_null();
2812
return cle != nullptr ? cle->stride_con() : 0;
2815
BaseCountedLoopNode* BaseCountedLoopNode::make(Node* entry, Node* backedge, BasicType bt) {2817
return new CountedLoopNode(entry, backedge);
2819
assert(bt == T_LONG, "unsupported");
2820
return new LongCountedLoopNode(entry, backedge);
2823
void OuterStripMinedLoopNode::fix_sunk_stores(CountedLoopEndNode* inner_cle, LoopNode* inner_cl, PhaseIterGVN* igvn,
2824
PhaseIdealLoop* iloop) {2825
Node* cle_out = inner_cle->proj_out(false);
2826
Node* cle_tail = inner_cle->proj_out(true);
2827
if (cle_out->outcnt() > 1) {2828
// Look for chains of stores that were sunk
2829
// out of the inner loop and are in the outer loop
2830
for (DUIterator_Fast imax, i = cle_out->fast_outs(imax); i < imax; i++) {2831
Node* u = cle_out->fast_out(i);
2832
if (u->is_Store()) {2833
int alias_idx = igvn->C->get_alias_index(u->adr_type());
2836
Node* next = first->in(MemNode::Memory);
2837
if (!next->is_Store() || next->in(0) != cle_out) {2840
assert(igvn->C->get_alias_index(next->adr_type()) == alias_idx, "");
2845
Node* next = nullptr;
2846
for (DUIterator_Fast jmax, j = last->fast_outs(jmax); j < jmax; j++) {2847
Node* uu = last->fast_out(j);
2848
if (uu->is_Store() && uu->in(0) == cle_out) {2849
assert(next == nullptr, "only one in the outer loop");
2851
assert(igvn->C->get_alias_index(next->adr_type()) == alias_idx, "");
2854
if (next == nullptr) {2859
Node* phi = nullptr;
2860
for (DUIterator_Fast jmax, j = inner_cl->fast_outs(jmax); j < jmax; j++) {2861
Node* uu = inner_cl->fast_out(j);
2863
Node* be = uu->in(LoopNode::LoopBackControl);
2864
if (be->is_Store() && be->in(0) == inner_cl->in(LoopNode::LoopBackControl)) {2865
assert(igvn->C->get_alias_index(uu->adr_type()) != alias_idx && igvn->C->get_alias_index(uu->adr_type()) != Compile::AliasIdxBot, "unexpected store");
2867
if (be == last || be == first->in(MemNode::Memory)) {2868
assert(igvn->C->get_alias_index(uu->adr_type()) == alias_idx || igvn->C->get_alias_index(uu->adr_type()) == Compile::AliasIdxBot, "unexpected alias");
2869
assert(phi == nullptr, "only one phi");
2875
for (DUIterator_Fast jmax, j = inner_cl->fast_outs(jmax); j < jmax; j++) {2876
Node* uu = inner_cl->fast_out(j);
2877
if (uu->is_memory_phi()) {2878
if (uu->adr_type() == igvn->C->get_adr_type(igvn->C->get_alias_index(u->adr_type()))) {2879
assert(phi == uu, "what's that phi?");
2880
} else if (uu->adr_type() == TypePtr::BOTTOM) {2881
Node* n = uu->in(LoopNode::LoopBackControl);
2882
uint limit = igvn->C->live_nodes();
2886
assert(i < limit, "infinite loop");
2889
} else if (n->is_SafePoint() || n->is_MemBar()) {2890
n = n->in(TypeFunc::Memory);
2891
} else if (n->is_Phi()) {2893
} else if (n->is_MergeMem()) {2894
n = n->as_MergeMem()->memory_at(igvn->C->get_alias_index(u->adr_type()));
2895
} else if (n->is_Store() || n->is_LoadStore() || n->is_ClearArray()) {2896
n = n->in(MemNode::Memory);
2899
ShouldNotReachHere();
2906
if (phi == nullptr) {2907
// If an entire chains was sunk, the
2908
// inner loop has no phi for that memory
2909
// slice, create one for the outer loop
2910
phi = PhiNode::make(inner_cl, first->in(MemNode::Memory), Type::MEMORY,
2911
igvn->C->get_adr_type(igvn->C->get_alias_index(u->adr_type())));
2912
phi->set_req(LoopNode::LoopBackControl, last);
2913
phi = register_new_node(phi, inner_cl, igvn, iloop);
2914
igvn->replace_input_of(first, MemNode::Memory, phi);
2916
// Or fix the outer loop fix to include
2917
// that chain of stores.
2918
Node* be = phi->in(LoopNode::LoopBackControl);
2919
assert(!(be->is_Store() && be->in(0) == inner_cl->in(LoopNode::LoopBackControl)), "store on the backedge + sunk stores: unsupported");
2920
if (be == first->in(MemNode::Memory)) {2921
if (be == phi->in(LoopNode::LoopBackControl)) {2922
igvn->replace_input_of(phi, LoopNode::LoopBackControl, last);
2924
igvn->replace_input_of(be, MemNode::Memory, last);
2928
if (be == phi->in(LoopNode::LoopBackControl)) {2929
assert(phi->in(LoopNode::LoopBackControl) == last, "");
2931
assert(be->in(MemNode::Memory) == last, "");
2941
void OuterStripMinedLoopNode::adjust_strip_mined_loop(PhaseIterGVN* igvn) {2942
// Look for the outer & inner strip mined loop, reduce number of
2943
// iterations of the inner loop, set exit condition of outer loop,
2944
// construct required phi nodes for outer loop.
2945
CountedLoopNode* inner_cl = unique_ctrl_out()->as_CountedLoop();
2946
assert(inner_cl->is_strip_mined(), "inner loop should be strip mined");
2947
if (LoopStripMiningIter == 0) {2948
remove_outer_loop_and_safepoint(igvn);
2951
if (LoopStripMiningIter == 1) {2952
transform_to_counted_loop(igvn, nullptr);
2955
Node* inner_iv_phi = inner_cl->phi();
2956
if (inner_iv_phi == nullptr) {2957
IfNode* outer_le = outer_loop_end();
2958
Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt));
2959
igvn->replace_node(outer_le, iff);
2960
inner_cl->clear_strip_mined();
2963
CountedLoopEndNode* inner_cle = inner_cl->loopexit();
2965
int stride = inner_cl->stride_con();
2966
// For a min int stride, LoopStripMiningIter * stride overflows the int range for all values of LoopStripMiningIter
2967
// except 0 or 1. Those values are handled early on in this method and causes the method to return. So for a min int
2968
// stride, the method is guaranteed to return at the next check below.
2969
jlong scaled_iters_long = ((jlong)LoopStripMiningIter) * ABS((jlong)stride);
2970
int scaled_iters = (int)scaled_iters_long;
2971
if ((jlong)scaled_iters != scaled_iters_long) {2972
// Remove outer loop and safepoint (too few iterations)
2973
remove_outer_loop_and_safepoint(igvn);
2976
jlong short_scaled_iters = LoopStripMiningIterShortLoop * ABS(stride);
2977
const TypeInt* inner_iv_t = igvn->type(inner_iv_phi)->is_int();
2978
jlong iter_estimate = (jlong)inner_iv_t->_hi - (jlong)inner_iv_t->_lo;
2979
assert(iter_estimate > 0, "broken");
2980
if (iter_estimate <= short_scaled_iters) {2981
// Remove outer loop and safepoint: loop executes less than LoopStripMiningIterShortLoop
2982
remove_outer_loop_and_safepoint(igvn);
2985
if (iter_estimate <= scaled_iters_long) {2986
// We would only go through one iteration of
2987
// the outer loop: drop the outer loop but
2988
// keep the safepoint so we don't run for
2989
// too long without a safepoint
2990
IfNode* outer_le = outer_loop_end();
2991
Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt));
2992
igvn->replace_node(outer_le, iff);
2993
inner_cl->clear_strip_mined();
2997
Node* cle_tail = inner_cle->proj_out(true);
3000
if (cle_tail->outcnt() > 1) {3001
// Look for nodes on backedge of inner loop and clone them
3002
Unique_Node_List backedge_nodes;
3003
for (DUIterator_Fast imax, i = cle_tail->fast_outs(imax); i < imax; i++) {3004
Node* u = cle_tail->fast_out(i);
3005
if (u != inner_cl) {3006
assert(!u->is_CFG(), "control flow on the backedge?");
3007
backedge_nodes.push(u);
3010
uint last = igvn->C->unique();
3011
for (uint next = 0; next < backedge_nodes.size(); next++) {3012
Node* n = backedge_nodes.at(next);
3013
old_new.map(n->_idx, n->clone());
3014
for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {3015
Node* u = n->fast_out(i);
3016
assert(!u->is_CFG(), "broken");
3017
if (u->_idx >= last) {3021
backedge_nodes.push(u);
3023
assert(u->in(0) == inner_cl, "strange phi on the backedge");
3027
// Put the clones on the outer loop backedge
3028
Node* le_tail = outer_loop_tail();
3029
for (uint next = 0; next < backedge_nodes.size(); next++) {3030
Node *n = old_new[backedge_nodes.at(next)->_idx];
3031
for (uint i = 1; i < n->req(); i++) {3032
if (n->in(i) != nullptr && old_new[n->in(i)->_idx] != nullptr) {3033
n->set_req(i, old_new[n->in(i)->_idx]);
3036
if (n->in(0) != nullptr && n->in(0) == cle_tail) {3037
n->set_req(0, le_tail);
3039
igvn->register_new_node_with_optimizer(n);
3043
Node* iv_phi = nullptr;
3044
// Make a clone of each phi in the inner loop
3045
// for the outer loop
3046
for (uint i = 0; i < inner_cl->outcnt(); i++) {3047
Node* u = inner_cl->raw_out(i);
3049
assert(u->in(0) == inner_cl, "inconsistent");
3050
Node* phi = u->clone();
3051
phi->set_req(0, this);
3052
Node* be = old_new[phi->in(LoopNode::LoopBackControl)->_idx];
3053
if (be != nullptr) {3054
phi->set_req(LoopNode::LoopBackControl, be);
3056
phi = igvn->transform(phi);
3057
igvn->replace_input_of(u, LoopNode::EntryControl, phi);
3058
if (u == inner_iv_phi) {3064
if (iv_phi != nullptr) {3065
// Now adjust the inner loop's exit condition
3066
Node* limit = inner_cl->limit();
3067
// If limit < init for stride > 0 (or limit > init for stride < 0),
3068
// the loop body is run only once. Given limit - init (init - limit resp.)
3069
// would be negative, the unsigned comparison below would cause
3070
// the loop body to be run for LoopStripMiningIter.
3071
Node* max = nullptr;
3073
max = MaxNode::max_diff_with_zero(limit, iv_phi, TypeInt::INT, *igvn);
3075
max = MaxNode::max_diff_with_zero(iv_phi, limit, TypeInt::INT, *igvn);
3077
// sub is positive and can be larger than the max signed int
3078
// value. Use an unsigned min.
3079
Node* const_iters = igvn->intcon(scaled_iters);
3080
Node* min = MaxNode::unsigned_min(max, const_iters, TypeInt::make(0, scaled_iters, Type::WidenMin), *igvn);
3081
// min is the number of iterations for the next inner loop execution:
3082
// unsigned_min(max(limit - iv_phi, 0), scaled_iters) if stride > 0
3083
// unsigned_min(max(iv_phi - limit, 0), scaled_iters) if stride < 0
3085
Node* new_limit = nullptr;
3087
new_limit = igvn->transform(new AddINode(min, iv_phi));
3089
new_limit = igvn->transform(new SubINode(iv_phi, min));
3091
Node* inner_cmp = inner_cle->cmp_node();
3092
Node* inner_bol = inner_cle->in(CountedLoopEndNode::TestValue);
3093
Node* outer_bol = inner_bol;
3094
// cmp node for inner loop may be shared
3095
inner_cmp = inner_cmp->clone();
3096
inner_cmp->set_req(2, new_limit);
3097
inner_bol = inner_bol->clone();
3098
inner_bol->set_req(1, igvn->transform(inner_cmp));
3099
igvn->replace_input_of(inner_cle, CountedLoopEndNode::TestValue, igvn->transform(inner_bol));
3100
// Set the outer loop's exit condition too
3101
igvn->replace_input_of(outer_loop_end(), 1, outer_bol);
3103
assert(false, "should be able to adjust outer loop");
3104
IfNode* outer_le = outer_loop_end();
3105
Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt));
3106
igvn->replace_node(outer_le, iff);
3107
inner_cl->clear_strip_mined();
3111
void OuterStripMinedLoopNode::transform_to_counted_loop(PhaseIterGVN* igvn, PhaseIdealLoop* iloop) {3112
CountedLoopNode* inner_cl = unique_ctrl_out()->as_CountedLoop();
3113
CountedLoopEndNode* cle = inner_cl->loopexit();
3114
Node* inner_test = cle->in(1);
3115
IfNode* outer_le = outer_loop_end();
3116
CountedLoopEndNode* inner_cle = inner_cl->loopexit();
3117
Node* safepoint = outer_safepoint();
3119
fix_sunk_stores(inner_cle, inner_cl, igvn, iloop);
3121
// make counted loop exit test always fail
3122
ConINode* zero = igvn->intcon(0);
3123
if (iloop != nullptr) {3124
iloop->set_ctrl(zero, igvn->C->root());
3126
igvn->replace_input_of(cle, 1, zero);
3127
// replace outer loop end with CountedLoopEndNode with formers' CLE's exit test
3128
Node* new_end = new CountedLoopEndNode(outer_le->in(0), inner_test, cle->_prob, cle->_fcnt);
3129
register_control(new_end, inner_cl, outer_le->in(0), igvn, iloop);
3130
if (iloop == nullptr) {3131
igvn->replace_node(outer_le, new_end);
3133
iloop->lazy_replace(outer_le, new_end);
3135
// the backedge of the inner loop must be rewired to the new loop end
3136
Node* backedge = cle->proj_out(true);
3137
igvn->replace_input_of(backedge, 0, new_end);
3138
if (iloop != nullptr) {3139
iloop->set_idom(backedge, new_end, iloop->dom_depth(new_end) + 1);
3141
// make the outer loop go away
3142
igvn->replace_input_of(in(LoopBackControl), 0, igvn->C->top());
3143
igvn->replace_input_of(this, LoopBackControl, igvn->C->top());
3144
inner_cl->clear_strip_mined();
3145
if (iloop != nullptr) {3146
Unique_Node_List wq;
3149
IdealLoopTree* outer_loop_ilt = iloop->get_loop(this);
3150
IdealLoopTree* loop = iloop->get_loop(inner_cl);
3152
for (uint i = 0; i < wq.size(); i++) {3154
for (uint j = 0; j < n->req(); ++j) {3155
Node* in = n->in(j);
3156
if (in == nullptr || in->is_CFG()) {3159
if (iloop->get_loop(iloop->get_ctrl(in)) != outer_loop_ilt) {3162
assert(!loop->_body.contains(in), "");
3163
loop->_body.push(in);
3167
iloop->set_loop(safepoint, loop);
3168
loop->_body.push(safepoint);
3169
iloop->set_loop(safepoint->in(0), loop);
3170
loop->_body.push(safepoint->in(0));
3171
outer_loop_ilt->_tail = igvn->C->top();
3175
void OuterStripMinedLoopNode::remove_outer_loop_and_safepoint(PhaseIterGVN* igvn) const {3176
CountedLoopNode* inner_cl = unique_ctrl_out()->as_CountedLoop();
3177
Node* outer_sfpt = outer_safepoint();
3178
Node* outer_out = outer_loop_exit();
3179
igvn->replace_node(outer_out, outer_sfpt->in(0));
3180
igvn->replace_input_of(outer_sfpt, 0, igvn->C->top());
3181
inner_cl->clear_strip_mined();
3184
Node* OuterStripMinedLoopNode::register_new_node(Node* node, LoopNode* ctrl, PhaseIterGVN* igvn, PhaseIdealLoop* iloop) {3185
if (iloop == nullptr) {3186
return igvn->transform(node);
3188
iloop->register_new_node(node, ctrl);
3192
Node* OuterStripMinedLoopNode::register_control(Node* node, Node* loop, Node* idom, PhaseIterGVN* igvn,
3193
PhaseIdealLoop* iloop) {3194
if (iloop == nullptr) {3195
return igvn->transform(node);
3197
iloop->register_control(node, iloop->get_loop(loop), idom);
3201
const Type* OuterStripMinedLoopEndNode::Value(PhaseGVN* phase) const {3202
if (!in(0)) return Type::TOP;
3203
if (phase->type(in(0)) == Type::TOP)
3206
// Until expansion, the loop end condition is not set so this should not constant fold.
3207
if (is_expanded(phase)) {3208
return IfNode::Value(phase);
3211
return TypeTuple::IFBOTH;
3214
bool OuterStripMinedLoopEndNode::is_expanded(PhaseGVN *phase) const {3215
// The outer strip mined loop head only has Phi uses after expansion
3216
if (phase->is_IterGVN()) {3217
Node* backedge = proj_out_or_null(true);
3218
if (backedge != nullptr) {3219
Node* head = backedge->unique_ctrl_out_or_null();
3220
if (head != nullptr && head->is_OuterStripMinedLoop()) {3221
if (head->find_out_with(Op_Phi) != nullptr) {3230
Node *OuterStripMinedLoopEndNode::Ideal(PhaseGVN *phase, bool can_reshape) {3231
if (remove_dead_region(phase, can_reshape)) return this;
3236
//------------------------------filtered_type--------------------------------
3237
// Return a type based on condition control flow
3238
// A successful return will be a type that is restricted due
3239
// to a series of dominating if-tests, such as:
3242
// here: "i" type is [1..10)
3245
// or a control flow merge
3248
// phi( , ) -- at top of loop type is [min_int..10)
3252
const TypeInt* PhaseIdealLoop::filtered_type( Node *n, Node* n_ctrl) {3253
assert(n && n->bottom_type()->is_int(), "must be int");
3254
const TypeInt* filtered_t = nullptr;
3256
assert(n_ctrl != nullptr || n_ctrl == C->top(), "valid control");
3257
filtered_t = filtered_type_from_dominators(n, n_ctrl);
3260
Node* phi = n->as_Phi();
3261
Node* region = phi->in(0);
3262
assert(n_ctrl == nullptr || n_ctrl == region, "ctrl parameter must be region");
3263
if (region && region != C->top()) {3264
for (uint i = 1; i < phi->req(); i++) {3265
Node* val = phi->in(i);
3266
Node* use_c = region->in(i);
3267
const TypeInt* val_t = filtered_type_from_dominators(val, use_c);
3268
if (val_t != nullptr) {3269
if (filtered_t == nullptr) {3272
filtered_t = filtered_t->meet(val_t)->is_int();
3278
const TypeInt* n_t = _igvn.type(n)->is_int();
3279
if (filtered_t != nullptr) {3280
n_t = n_t->join(filtered_t)->is_int();
3286
//------------------------------filtered_type_from_dominators--------------------------------
3287
// Return a possibly more restrictive type for val based on condition control flow of dominators
3288
const TypeInt* PhaseIdealLoop::filtered_type_from_dominators( Node* val, Node *use_ctrl) {3289
if (val->is_Con()) {3290
return val->bottom_type()->is_int();
3292
uint if_limit = 10; // Max number of dominating if's visited
3293
const TypeInt* rtn_t = nullptr;
3295
if (use_ctrl && use_ctrl != C->top()) {3296
Node* val_ctrl = get_ctrl(val);
3297
uint val_dom_depth = dom_depth(val_ctrl);
3298
Node* pred = use_ctrl;
3300
while (if_cnt < if_limit) {3301
if ((pred->Opcode() == Op_IfTrue || pred->Opcode() == Op_IfFalse)) {3303
const TypeInt* if_t = IfNode::filtered_int_type(&_igvn, val, pred);
3304
if (if_t != nullptr) {3305
if (rtn_t == nullptr) {3308
rtn_t = rtn_t->join(if_t)->is_int();
3313
if (pred == nullptr || pred == C->top()) {3316
// Stop if going beyond definition block of val
3317
if (dom_depth(pred) < val_dom_depth) {3326
//------------------------------dump_spec--------------------------------------
3327
// Dump special per-node info
3329
void CountedLoopEndNode::dump_spec(outputStream *st) const {3330
if( in(TestValue) != nullptr && in(TestValue)->is_Bool() ) {3331
BoolTest bt( test_trip()); // Added this for g++.
3338
IfNode::dump_spec(st);
3342
//=============================================================================
3343
//------------------------------is_member--------------------------------------
3344
// Is 'l' a member of 'this'?
3345
bool IdealLoopTree::is_member(const IdealLoopTree *l) const {3346
while( l->_nest > _nest ) l = l->_parent;
3350
//------------------------------set_nest---------------------------------------
3351
// Set loop tree nesting depth. Accumulate _has_call bits.
3352
int IdealLoopTree::set_nest( uint depth ) {3353
assert(depth <= SHRT_MAX, "sanity");
3355
int bits = _has_call;
3356
if( _child ) bits |= _child->set_nest(depth+1);
3357
if( bits ) _has_call = 1;
3358
if( _next ) bits |= _next ->set_nest(depth );
3362
//------------------------------split_fall_in----------------------------------
3363
// Split out multiple fall-in edges from the loop header. Move them to a
3364
// private RegionNode before the loop. This becomes the loop landing pad.
3365
void IdealLoopTree::split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt ) {3366
PhaseIterGVN &igvn = phase->_igvn;
3369
// Make a new RegionNode to be the landing pad.
3370
RegionNode* landing_pad = new RegionNode(fall_in_cnt + 1);
3371
phase->set_loop(landing_pad,_parent);
3372
// If _head was irreducible loop entry, landing_pad may now be too
3373
landing_pad->set_loop_status(_head->as_Region()->loop_status());
3374
// Gather all the fall-in control paths into the landing pad
3375
uint icnt = fall_in_cnt;
3376
uint oreq = _head->req();
3377
for( i = oreq-1; i>0; i-- )
3378
if( !phase->is_member( this, _head->in(i) ) )
3379
landing_pad->set_req(icnt--,_head->in(i));
3381
// Peel off PhiNode edges as well
3382
for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {3383
Node *oj = _head->fast_out(j);
3384
if( oj->is_Phi() ) {3385
PhiNode* old_phi = oj->as_Phi();
3386
assert( old_phi->region() == _head, "" );
3387
igvn.hash_delete(old_phi); // Yank from hash before hacking edges
3388
Node *p = PhiNode::make_blank(landing_pad, old_phi);
3389
uint icnt = fall_in_cnt;
3390
for( i = oreq-1; i>0; i-- ) {3391
if( !phase->is_member( this, _head->in(i) ) ) {3392
p->init_req(icnt--, old_phi->in(i));
3393
// Go ahead and clean out old edges from old phi
3394
old_phi->del_req(i);
3397
// Search for CSE's here, because ZKM.jar does a lot of
3398
// loop hackery and we need to be a little incremental
3399
// with the CSE to avoid O(N^2) node blow-up.
3400
Node *p2 = igvn.hash_find_insert(p); // Look for a CSE
3401
if( p2 ) { // Found CSE3402
p->destruct(&igvn); // Recover useless new node
3403
p = p2; // Use old node
3405
igvn.register_new_node_with_optimizer(p, old_phi);
3407
// Make old Phi refer to new Phi.
3408
old_phi->add_req(p);
3409
// Check for the special case of making the old phi useless and
3410
// disappear it. In JavaGrande I have a case where this useless
3411
// Phi is the loop limit and prevents recognizing a CountedLoop
3412
// which in turn prevents removing an empty loop.
3413
Node *id_old_phi = old_phi->Identity(&igvn);
3414
if( id_old_phi != old_phi ) { // Found a simple identity?3415
// Note that I cannot call 'replace_node' here, because
3416
// that will yank the edge from old_phi to the Region and
3417
// I'm mid-iteration over the Region's uses.
3418
for (DUIterator_Last imin, i = old_phi->last_outs(imin); i >= imin; ) {3419
Node* use = old_phi->last_out(i);
3420
igvn.rehash_node_delayed(use);
3421
uint uses_found = 0;
3422
for (uint j = 0; j < use->len(); j++) {3423
if (use->in(j) == old_phi) {3424
if (j < use->req()) use->set_req (j, id_old_phi);
3425
else use->set_prec(j, id_old_phi);
3429
i -= uses_found; // we deleted 1 or more copies of this edge
3432
igvn._worklist.push(old_phi);
3435
// Finally clean out the fall-in edges from the RegionNode
3436
for( i = oreq-1; i>0; i-- ) {3437
if( !phase->is_member( this, _head->in(i) ) ) {3441
igvn.rehash_node_delayed(_head);
3442
// Transform landing pad
3443
igvn.register_new_node_with_optimizer(landing_pad, _head);
3444
// Insert landing pad into the header
3445
_head->add_req(landing_pad);
3448
//------------------------------split_outer_loop-------------------------------
3449
// Split out the outermost loop from this shared header.
3450
void IdealLoopTree::split_outer_loop( PhaseIdealLoop *phase ) {3451
PhaseIterGVN &igvn = phase->_igvn;
3453
// Find index of outermost loop; it should also be my tail.
3455
while( _head->in(outer_idx) != _tail ) outer_idx++;
3457
// Make a LoopNode for the outermost loop.
3458
Node *ctl = _head->in(LoopNode::EntryControl);
3459
Node *outer = new LoopNode( ctl, _head->in(outer_idx) );
3460
outer = igvn.register_new_node_with_optimizer(outer, _head);
3461
phase->set_created_loop_node();
3463
// Outermost loop falls into '_head' loop
3464
_head->set_req(LoopNode::EntryControl, outer);
3465
_head->del_req(outer_idx);
3466
// Split all the Phis up between '_head' loop and 'outer' loop.
3467
for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {3468
Node *out = _head->fast_out(j);
3469
if( out->is_Phi() ) {3470
PhiNode *old_phi = out->as_Phi();
3471
assert( old_phi->region() == _head, "" );
3472
Node *phi = PhiNode::make_blank(outer, old_phi);
3473
phi->init_req(LoopNode::EntryControl, old_phi->in(LoopNode::EntryControl));
3474
phi->init_req(LoopNode::LoopBackControl, old_phi->in(outer_idx));
3475
phi = igvn.register_new_node_with_optimizer(phi, old_phi);
3476
// Make old Phi point to new Phi on the fall-in path
3477
igvn.replace_input_of(old_phi, LoopNode::EntryControl, phi);
3478
old_phi->del_req(outer_idx);
3482
// Use the new loop head instead of the old shared one
3484
phase->set_loop(_head, this);
3487
//------------------------------fix_parent-------------------------------------
3488
static void fix_parent( IdealLoopTree *loop, IdealLoopTree *parent ) {3489
loop->_parent = parent;
3490
if( loop->_child ) fix_parent( loop->_child, loop );
3491
if( loop->_next ) fix_parent( loop->_next , parent );
3494
//------------------------------estimate_path_freq-----------------------------
3495
static float estimate_path_freq( Node *n ) {3496
// Try to extract some path frequency info
3498
for( int i = 0; i < 50; i++ ) { // Skip through a bunch of uncommon tests3499
uint nop = n->Opcode();
3500
if( nop == Op_SafePoint ) { // Skip any safepoint3504
if( nop == Op_CatchProj ) { // Get count from a prior call3505
// Assume call does not always throw exceptions: means the call-site
3506
// count is also the frequency of the fall-through path.
3507
assert( n->is_CatchProj(), "" );
3508
if( ((CatchProjNode*)n)->_con != CatchProjNode::fall_through_index )
3509
return 0.0f; // Assume call exception path is rare
3510
Node *call = n->in(0)->in(0)->in(0);
3511
assert( call->is_Call(), "expect a call here" );
3512
const JVMState *jvms = ((CallNode*)call)->jvms();
3513
ciMethodData* methodData = jvms->method()->method_data();
3514
if (!methodData->is_mature()) return 0.0f; // No call-site data
3515
ciProfileData* data = methodData->bci_to_data(jvms->bci());
3516
if ((data == nullptr) || !data->is_CounterData()) {3517
// no call profile available, try call's control input
3521
return data->as_CounterData()->count()/FreqCountInvocations;
3523
// See if there's a gating IF test
3524
Node *n_c = n->in(0);
3525
if( !n_c->is_If() ) break; // No estimate available
3527
if( iff->_fcnt != COUNT_UNKNOWN ) // Have a valid count?
3528
// Compute how much count comes on this path
3529
return ((nop == Op_IfTrue) ? iff->_prob : 1.0f - iff->_prob) * iff->_fcnt;
3530
// Have no count info. Skip dull uncommon-trap like branches.
3531
if( (nop == Op_IfTrue && iff->_prob < PROB_LIKELY_MAG(5)) ||
3532
(nop == Op_IfFalse && iff->_prob > PROB_UNLIKELY_MAG(5)) )
3534
// Skip through never-taken branch; look for a real loop exit.
3537
return 0.0f; // No estimate available
3540
//------------------------------merge_many_backedges---------------------------
3541
// Merge all the backedges from the shared header into a private Region.
3542
// Feed that region as the one backedge to this loop.
3543
void IdealLoopTree::merge_many_backedges( PhaseIdealLoop *phase ) {3546
// Scan for the top 2 hottest backedges
3547
float hotcnt = 0.0f;
3548
float warmcnt = 0.0f;
3550
// Loop starts at 2 because slot 1 is the fall-in path
3551
for( i = 2; i < _head->req(); i++ ) {3552
float cnt = estimate_path_freq(_head->in(i));
3553
if( cnt > hotcnt ) { // Grab hottest path3557
} else if( cnt > warmcnt ) { // And 2nd hottest path3562
// See if the hottest backedge is worthy of being an inner loop
3563
// by being much hotter than the next hottest backedge.
3564
if( hotcnt <= 0.0001 ||
3565
hotcnt < 2.0*warmcnt ) hot_idx = 0;// No hot backedge
3567
// Peel out the backedges into a private merge point; peel
3568
// them all except optionally hot_idx.
3569
PhaseIterGVN &igvn = phase->_igvn;
3571
Node *hot_tail = nullptr;
3572
// Make a Region for the merge point
3573
Node *r = new RegionNode(1);
3574
for( i = 2; i < _head->req(); i++ ) {3576
r->add_req( _head->in(i) );
3577
else hot_tail = _head->in(i);
3579
igvn.register_new_node_with_optimizer(r, _head);
3580
// Plug region into end of loop _head, followed by hot_tail
3581
while( _head->req() > 3 ) _head->del_req( _head->req()-1 );
3582
igvn.replace_input_of(_head, 2, r);
3583
if( hot_idx ) _head->add_req(hot_tail);
3585
// Split all the Phis up between '_head' loop and the Region 'r'
3586
for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) {3587
Node *out = _head->fast_out(j);
3588
if( out->is_Phi() ) {3589
PhiNode* n = out->as_Phi();
3590
igvn.hash_delete(n); // Delete from hash before hacking edges
3591
Node *hot_phi = nullptr;
3592
Node *phi = new PhiNode(r, n->type(), n->adr_type());
3593
// Check all inputs for the ones to peel out
3595
for( uint i = 2; i < n->req(); i++ ) {3597
phi->set_req( j++, n->in(i) );
3598
else hot_phi = n->in(i);
3600
// Register the phi but do not transform until whole place transforms
3601
igvn.register_new_node_with_optimizer(phi, n);
3602
// Add the merge phi to the old Phi
3603
while( n->req() > 3 ) n->del_req( n->req()-1 );
3604
igvn.replace_input_of(n, 2, phi);
3605
if( hot_idx ) n->add_req(hot_phi);
3610
// Insert a new IdealLoopTree inserted below me. Turn it into a clone
3611
// of self loop tree. Turn self into a loop headed by _head and with
3612
// tail being the new merge point.
3613
IdealLoopTree *ilt = new IdealLoopTree( phase, _head, _tail );
3614
phase->set_loop(_tail,ilt); // Adjust tail
3615
_tail = r; // Self's tail is new merge point
3616
phase->set_loop(r,this);
3617
ilt->_child = _child; // New guy has my children
3618
_child = ilt; // Self has new guy as only child
3619
ilt->_parent = this; // new guy has self for parent
3620
ilt->_nest = _nest; // Same nesting depth (for now)
3622
// Starting with 'ilt', look for child loop trees using the same shared
3623
// header. Flatten these out; they will no longer be loops in the end.
3624
IdealLoopTree **pilt = &_child;
3626
if( ilt->_head == _head ) {3628
for( i = 2; i < _head->req(); i++ )
3629
if( _head->in(i) == ilt->_tail )
3630
break; // Still a loop
3631
if( i == _head->req() ) { // No longer a loop3632
// Flatten ilt. Hang ilt's "_next" list from the end of
3633
// ilt's '_child' list. Move the ilt's _child up to replace ilt.
3634
IdealLoopTree **cp = &ilt->_child;
3635
while( *cp ) cp = &(*cp)->_next; // Find end of child list
3636
*cp = ilt->_next; // Hang next list at end of child list
3637
*pilt = ilt->_child; // Move child up to replace ilt
3638
ilt->_head = nullptr; // Flag as a loop UNIONED into parent
3639
ilt = ilt->_child; // Repeat using new ilt
3640
continue; // do not advance over ilt->_child
3642
assert( ilt->_tail == hot_tail, "expected to only find the hot inner loop here" );
3643
phase->set_loop(_head,ilt);
3645
pilt = &ilt->_child; // Advance to next
3649
if( _child ) fix_parent( _child, this );
3652
//------------------------------beautify_loops---------------------------------
3653
// Split shared headers and insert loop landing pads.
3654
// Insert a LoopNode to replace the RegionNode.
3655
// Return TRUE if loop tree is structurally changed.
3656
bool IdealLoopTree::beautify_loops( PhaseIdealLoop *phase ) {3657
bool result = false;
3658
// Cache parts in locals for easy
3659
PhaseIterGVN &igvn = phase->_igvn;
3661
igvn.hash_delete(_head); // Yank from hash before hacking edges
3663
// Check for multiple fall-in paths. Peel off a landing pad if need be.
3664
int fall_in_cnt = 0;
3665
for( uint i = 1; i < _head->req(); i++ )
3666
if( !phase->is_member( this, _head->in(i) ) )
3668
assert( fall_in_cnt, "at least 1 fall-in path" );
3669
if( fall_in_cnt > 1 ) // Need a loop landing pad to merge fall-ins
3670
split_fall_in( phase, fall_in_cnt );
3672
// Swap inputs to the _head and all Phis to move the fall-in edge to
3675
while( phase->is_member( this, _head->in(fall_in_cnt) ) )
3677
if( fall_in_cnt > 1 ) {3678
// Since I am just swapping inputs I do not need to update def-use info
3679
Node *tmp = _head->in(1);
3680
igvn.rehash_node_delayed(_head);
3681
_head->set_req( 1, _head->in(fall_in_cnt) );
3682
_head->set_req( fall_in_cnt, tmp );
3683
// Swap also all Phis
3684
for (DUIterator_Fast imax, i = _head->fast_outs(imax); i < imax; i++) {3685
Node* phi = _head->fast_out(i);
3686
if( phi->is_Phi() ) {3687
igvn.rehash_node_delayed(phi); // Yank from hash before hacking edges
3689
phi->set_req( 1, phi->in(fall_in_cnt) );
3690
phi->set_req( fall_in_cnt, tmp );
3694
assert( !phase->is_member( this, _head->in(1) ), "left edge is fall-in" );
3695
assert( phase->is_member( this, _head->in(2) ), "right edge is loop" );
3697
// If I am a shared header (multiple backedges), peel off the many
3698
// backedges into a private merge point and use the merge point as
3699
// the one true backedge.
3700
if (_head->req() > 3) {3701
// Merge the many backedges into a single backedge but leave
3702
// the hottest backedge as separate edge for the following peel.
3703
if (!_irreducible) {3704
merge_many_backedges( phase );
3707
// When recursively beautify my children, split_fall_in can change
3708
// loop tree structure when I am an irreducible loop. Then the head
3709
// of my children has a req() not bigger than 3. Here we need to set
3710
// result to true to catch that case in order to tell the caller to
3711
// rebuild loop tree. See issue JDK-8244407 for details.
3715
// If I have one hot backedge, peel off myself loop.
3716
// I better be the outermost loop.
3717
if (_head->req() > 3 && !_irreducible) {3718
split_outer_loop( phase );
3721
} else if (!_head->is_Loop() && !_irreducible) {3722
// Make a new LoopNode to replace the old loop head
3723
Node *l = new LoopNode( _head->in(1), _head->in(2) );
3724
l = igvn.register_new_node_with_optimizer(l, _head);
3725
phase->set_created_loop_node();
3726
// Go ahead and replace _head
3727
phase->_igvn.replace_node( _head, l );
3729
phase->set_loop(_head, this);
3732
// Now recursively beautify nested loops
3733
if( _child ) result |= _child->beautify_loops( phase );
3734
if( _next ) result |= _next ->beautify_loops( phase );
3738
//------------------------------allpaths_check_safepts----------------------------
3739
// Allpaths backwards scan. Starting at the head, traversing all backedges, and the body. Terminating each path at first
3740
// safepoint encountered. Helper for check_safepts.
3741
void IdealLoopTree::allpaths_check_safepts(VectorSet &visited, Node_List &stack) {3742
assert(stack.size() == 0, "empty stack");
3745
visited.set(_head->_idx);
3746
while (stack.size() > 0) {3747
Node* n = stack.pop();
3748
if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {3749
// Terminate this path
3750
} else if (n->Opcode() == Op_SafePoint) {3751
if (_phase->get_loop(n) != this) {3752
if (_required_safept == nullptr) _required_safept = new Node_List();
3753
// save the first we run into on that path: closest to the tail if the head has a single backedge
3754
_required_safept->push(n);
3756
// Terminate this path
3758
uint start = n->is_Region() ? 1 : 0;
3759
uint end = n->is_Region() && (!n->is_Loop() || n == _head) ? n->req() : start + 1;
3760
for (uint i = start; i < end; i++) {3761
Node* in = n->in(i);
3762
assert(in->is_CFG(), "must be");
3763
if (!visited.test_set(in->_idx) && is_member(_phase->get_loop(in))) {3771
//------------------------------check_safepts----------------------------
3772
// Given dominators, try to find loops with calls that must always be
3773
// executed (call dominates loop tail). These loops do not need non-call
3774
// safepoints (ncsfpt).
3776
// A complication is that a safepoint in a inner loop may be needed
3777
// by an outer loop. In the following, the inner loop sees it has a
3778
// call (block 3) on every path from the head (block 2) to the
3779
// backedge (arc 3->2). So it deletes the ncsfpt (non-call safepoint)
3780
// in block 2, _but_ this leaves the outer loop without a safepoint.
3788
// | 2<---+ ncsfpt in 2
3791
// inner 2,3 / 3 | call in 3
3797
// This method creates a list (_required_safept) of ncsfpt nodes that must
3798
// be protected is created for each loop. When a ncsfpt maybe deleted, it
3799
// is first looked for in the lists for the outer loops of the current loop.
3801
// The insights into the problem:
3802
// A) counted loops are okay
3803
// B) innermost loops are okay (only an inner loop can delete
3804
// a ncsfpt needed by an outer loop)
3805
// C) a loop is immune from an inner loop deleting a safepoint
3806
// if the loop has a call on the idom-path
3807
// D) a loop is also immune if it has a ncsfpt (non-call safepoint) on the
3808
// idom-path that is not in a nested loop
3809
// E) otherwise, an ncsfpt on the idom-path that is nested in an inner
3810
// loop needs to be prevented from deletion by an inner loop
3812
// There are two analyses:
3813
// 1) The first, and cheaper one, scans the loop body from
3814
// tail to head following the idom (immediate dominator)
3815
// chain, looking for the cases (C,D,E) above.
3816
// Since inner loops are scanned before outer loops, there is summary
3817
// information about inner loops. Inner loops can be skipped over
3818
// when the tail of an inner loop is encountered.
3820
// 2) The second, invoked if the first fails to find a call or ncsfpt on
3821
// the idom path (which is rare), scans all predecessor control paths
3822
// from the tail to the head, terminating a path when a call or sfpt
3823
// is encountered, to find the ncsfpt's that are closest to the tail.
3825
void IdealLoopTree::check_safepts(VectorSet &visited, Node_List &stack) {3826
// Bottom up traversal
3827
IdealLoopTree* ch = _child;
3828
if (_child) _child->check_safepts(visited, stack);
3829
if (_next) _next ->check_safepts(visited, stack);
3831
if (!_head->is_CountedLoop() && !_has_sfpt && _parent != nullptr) {3832
bool has_call = false; // call on dom-path
3833
bool has_local_ncsfpt = false; // ncsfpt on dom-path at this loop depth
3834
Node* nonlocal_ncsfpt = nullptr; // ncsfpt on dom-path at a deeper depth
3835
if (!_irreducible) {3836
// Scan the dom-path nodes from tail to head
3837
for (Node* n = tail(); n != _head; n = _phase->idom(n)) {3838
if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) {3840
_has_sfpt = 1; // Then no need for a safept!
3842
} else if (n->Opcode() == Op_SafePoint) {3843
if (_phase->get_loop(n) == this) {3844
has_local_ncsfpt = true;
3847
if (nonlocal_ncsfpt == nullptr) {3848
nonlocal_ncsfpt = n; // save the one closest to the tail
3851
IdealLoopTree* nlpt = _phase->get_loop(n);
3853
// If at an inner loop tail, see if the inner loop has already
3854
// recorded seeing a call on the dom-path (and stop.) If not,
3855
// jump to the head of the inner loop.
3856
assert(is_member(nlpt), "nested loop");
3857
Node* tail = nlpt->_tail;
3858
if (tail->in(0)->is_If()) tail = tail->in(0);
3860
// If inner loop has call on dom-path, so does outer loop
3861
if (nlpt->_has_sfpt) {3866
// Skip to head of inner loop
3867
assert(_phase->is_dominator(_head, nlpt->_head), "inner head dominated by outer head");
3870
// this and nlpt (inner loop) have the same loop head. This should not happen because
3871
// during beautify_loops we call merge_many_backedges. However, infinite loops may not
3872
// have been attached to the loop-tree during build_loop_tree before beautify_loops,
3873
// but then attached in the build_loop_tree afterwards, and so still have unmerged
3874
// backedges. Check if we are indeed in an infinite subgraph, and terminate the scan,
3875
// since we have reached the loop head of this.
3876
assert(_head->as_Region()->is_in_infinite_subgraph(),
3877
"only expect unmerged backedges in infinite loops");
3885
// Record safept's that this loop needs preserved when an
3886
// inner loop attempts to delete it's safepoints.
3887
if (_child != nullptr && !has_call && !has_local_ncsfpt) {3888
if (nonlocal_ncsfpt != nullptr) {3889
if (_required_safept == nullptr) _required_safept = new Node_List();
3890
_required_safept->push(nonlocal_ncsfpt);
3892
// Failed to find a suitable safept on the dom-path. Now use
3893
// an all paths walk from tail to head, looking for safepoints to preserve.
3894
allpaths_check_safepts(visited, stack);
3900
//---------------------------is_deleteable_safept----------------------------
3901
// Is safept not required by an outer loop?
3902
bool PhaseIdealLoop::is_deleteable_safept(Node* sfpt) {3903
assert(sfpt->Opcode() == Op_SafePoint, "");
3904
IdealLoopTree* lp = get_loop(sfpt)->_parent;
3905
while (lp != nullptr) {3906
Node_List* sfpts = lp->_required_safept;
3907
if (sfpts != nullptr) {3908
for (uint i = 0; i < sfpts->size(); i++) {3909
if (sfpt == sfpts->at(i))
3918
//---------------------------replace_parallel_iv-------------------------------
3919
// Replace parallel induction variable (parallel to trip counter)
3920
void PhaseIdealLoop::replace_parallel_iv(IdealLoopTree *loop) {3921
assert(loop->_head->is_CountedLoop(), "");
3922
CountedLoopNode *cl = loop->_head->as_CountedLoop();
3923
if (!cl->is_valid_counted_loop(T_INT)) {3924
return; // skip malformed counted loop
3926
Node *incr = cl->incr();
3927
if (incr == nullptr) {3928
return; // Dead loop?
3930
Node *init = cl->init_trip();
3931
Node *phi = cl->phi();
3932
int stride_con = cl->stride_con();
3934
// Visit all children, looking for Phis
3935
for (DUIterator i = cl->outs(); cl->has_out(i); i++) {3936
Node *out = cl->out(i);
3937
// Look for other phis (secondary IVs). Skip dead ones
3938
if (!out->is_Phi() || out == phi || !has_node(out)) {3942
PhiNode* phi2 = out->as_Phi();
3943
Node* incr2 = phi2->in(LoopNode::LoopBackControl);
3944
// Look for induction variables of the form: X += constant
3945
if (phi2->region() != loop->_head ||
3946
incr2->req() != 3 ||
3947
incr2->in(1)->uncast() != phi2 ||
3949
incr2->Opcode() != Op_AddI ||
3950
!incr2->in(2)->is_Con()) {3954
if (incr2->in(1)->is_ConstraintCast() &&
3955
!(incr2->in(1)->in(0)->is_IfProj() && incr2->in(1)->in(0)->in(0)->is_RangeCheck())) {3956
// Skip AddI->CastII->Phi case if CastII is not controlled by local RangeCheck
3959
// Check for parallel induction variable (parallel to trip counter)
3960
// via an affine function. In particular, count-down loops with
3961
// count-up array indices are common. We only RCE references off
3962
// the trip-counter, so we need to convert all these to trip-counter
3964
Node* init2 = phi2->in(LoopNode::EntryControl);
3965
int stride_con2 = incr2->in(2)->get_int();
3967
// The ratio of the two strides cannot be represented as an int
3968
// if stride_con2 is min_int and stride_con is -1.
3969
if (stride_con2 == min_jint && stride_con == -1) {3973
// The general case here gets a little tricky. We want to find the
3974
// GCD of all possible parallel IV's and make a new IV using this
3975
// GCD for the loop. Then all possible IVs are simple multiples of
3976
// the GCD. In practice, this will cover very few extra loops.
3977
// Instead we require 'stride_con2' to be a multiple of 'stride_con',
3978
// where +/-1 is the common case, but other integer multiples are
3979
// also easy to handle.
3980
int ratio_con = stride_con2/stride_con;
3982
if ((ratio_con * stride_con) == stride_con2) { // Check for exact3984
if (TraceLoopOpts) {3985
tty->print("Parallel IV: %d ", phi2->_idx);3989
// Convert to using the trip counter. The parallel induction
3990
// variable differs from the trip counter by a loop-invariant
3991
// amount, the difference between their respective initial values.
3992
// It is scaled by the 'ratio_con'.
3993
Node* ratio = _igvn.intcon(ratio_con);
3994
set_ctrl(ratio, C->root());
3995
Node* ratio_init = new MulINode(init, ratio);
3996
_igvn.register_new_node_with_optimizer(ratio_init, init);
3997
set_early_ctrl(ratio_init, false);
3998
Node* diff = new SubINode(init2, ratio_init);
3999
_igvn.register_new_node_with_optimizer(diff, init2);
4000
set_early_ctrl(diff, false);
4001
Node* ratio_idx = new MulINode(phi, ratio);
4002
_igvn.register_new_node_with_optimizer(ratio_idx, phi);
4003
set_ctrl(ratio_idx, cl);
4004
Node* add = new AddINode(ratio_idx, diff);
4005
_igvn.register_new_node_with_optimizer(add);
4007
_igvn.replace_node( phi2, add );
4008
// Sometimes an induction variable is unused
4009
if (add->outcnt() == 0) {4010
_igvn.remove_dead_node(add);
4012
--i; // deleted this phi; rescan starting with next position
4018
void IdealLoopTree::remove_safepoints(PhaseIdealLoop* phase, bool keep_one) {4019
Node* keep = nullptr;
4021
// Look for a safepoint on the idom-path.
4022
for (Node* i = tail(); i != _head; i = phase->idom(i)) {4023
if (i->Opcode() == Op_SafePoint && phase->get_loop(i) == this) {4030
// Don't remove any safepoints if it is requested to keep a single safepoint and
4031
// no safepoint was found on idom-path. It is not safe to remove any safepoint
4032
// in this case since there's no safepoint dominating all paths in the loop body.
4033
bool prune = !keep_one || keep != nullptr;
4035
// Delete other safepoints in this loop.
4036
Node_List* sfpts = _safepts;
4037
if (prune && sfpts != nullptr) {4038
assert(keep == nullptr || keep->Opcode() == Op_SafePoint, "not safepoint");
4039
for (uint i = 0; i < sfpts->size(); i++) {4040
Node* n = sfpts->at(i);
4041
assert(phase->get_loop(n) == this, "");
4042
if (n != keep && phase->is_deleteable_safept(n)) {4043
phase->lazy_replace(n, n->in(TypeFunc::Control));
4049
//------------------------------counted_loop-----------------------------------
4050
// Convert to counted loops where possible
4051
void IdealLoopTree::counted_loop( PhaseIdealLoop *phase ) {4053
// For grins, set the inner-loop flag here
4055
if (_head->is_Loop()) _head->as_Loop()->set_inner_loop();
4058
IdealLoopTree* loop = this;
4059
if (_head->is_CountedLoop() ||
4060
phase->is_counted_loop(_head, loop, T_INT)) {4062
if (LoopStripMiningIter == 0 || _head->as_CountedLoop()->is_strip_mined()) {4063
// Indicate we do not need a safepoint here
4067
// Remove safepoints
4068
bool keep_one_sfpt = !(_has_call || _has_sfpt);
4069
remove_safepoints(phase, keep_one_sfpt);
4071
// Look for induction variables
4072
phase->replace_parallel_iv(this);
4073
} else if (_head->is_LongCountedLoop() ||
4074
phase->is_counted_loop(_head, loop, T_LONG)) {4075
remove_safepoints(phase, true);
4077
assert(!_head->is_Loop() || !_head->as_Loop()->is_loop_nest_inner_loop(), "transformation to counted loop should not fail");
4078
if (_parent != nullptr && !_irreducible) {4079
// Not a counted loop. Keep one safepoint.
4080
bool keep_one_sfpt = true;
4081
remove_safepoints(phase, keep_one_sfpt);
4086
assert(loop->_child != this || (loop->_head->as_Loop()->is_OuterStripMinedLoop() && _head->as_CountedLoop()->is_strip_mined()), "what kind of loop was added?");
4087
assert(loop->_child != this || (loop->_child->_child == nullptr && loop->_child->_next == nullptr), "would miss some loops");
4088
if (loop->_child && loop->_child != this) loop->_child->counted_loop(phase);
4089
if (loop->_next) loop->_next ->counted_loop(phase);
4093
// The Estimated Loop Clone Size:
4094
// CloneFactor * (~112% * BodySize + BC) + CC + FanOutTerm,
4095
// where BC and CC are totally ad-hoc/magic "body" and "clone" constants,
4096
// respectively, used to ensure that the node usage estimates made are on the
4097
// safe side, for the most part. The FanOutTerm is an attempt to estimate the
4098
// possible additional/excessive nodes generated due to data and control flow
4099
// merging, for edges reaching outside the loop.
4100
uint IdealLoopTree::est_loop_clone_sz(uint factor) const {4102
precond(0 < factor && factor < 16);
4106
uint const sz = _body.size() + (_body.size() + 7) / 2;
4107
uint estimate = factor * (sz + bc) + cc;
4109
assert((estimate - cc) / factor == sz + bc, "overflow");
4111
return estimate + est_loop_flow_merge_sz();
4114
// The Estimated Loop (full-) Unroll Size:
4115
// UnrollFactor * (~106% * BodySize) + CC + FanOutTerm,
4116
// where CC is a (totally) ad-hoc/magic "clone" constant, used to ensure that
4117
// node usage estimates made are on the safe side, for the most part. This is
4118
// a "light" version of the loop clone size calculation (above), based on the
4119
// assumption that most of the loop-construct overhead will be unraveled when
4120
// (fully) unrolled. Defined for unroll factors larger or equal to one (>=1),
4121
// including an overflow check and returning UINT_MAX in case of an overflow.
4122
uint IdealLoopTree::est_loop_unroll_sz(uint factor) const {4124
precond(factor > 0);
4126
// Take into account that after unroll conjoined heads and tails will fold.
4127
uint const b0 = _body.size() - EMPTY_LOOP_SIZE;
4129
uint const sz = b0 + (b0 + 15) / 16;
4130
uint estimate = factor * sz + cc;
4132
if ((estimate - cc) / factor != sz) {4136
return estimate + est_loop_flow_merge_sz();
4139
// Estimate the growth effect (in nodes) of merging control and data flow when
4140
// cloning a loop body, based on the amount of control and data flow reaching
4141
// outside of the (current) loop body.
4142
uint IdealLoopTree::est_loop_flow_merge_sz() const {4144
uint ctrl_edge_out_cnt = 0;
4145
uint data_edge_out_cnt = 0;
4147
for (uint i = 0; i < _body.size(); i++) {4148
Node* node = _body.at(i);
4149
uint outcnt = node->outcnt();
4151
for (uint k = 0; k < outcnt; k++) {4152
Node* out = node->raw_out(k);
4153
if (out == nullptr) continue;
4154
if (out->is_CFG()) {4155
if (!is_member(_phase->get_loop(out))) {4156
ctrl_edge_out_cnt++;
4158
} else if (_phase->has_ctrl(out)) {4159
Node* ctrl = _phase->get_ctrl(out);
4160
assert(ctrl != nullptr, "must be");
4161
assert(ctrl->is_CFG(), "must be");
4162
if (!is_member(_phase->get_loop(ctrl))) {4163
data_edge_out_cnt++;
4168
// Use data and control count (x2.0) in estimate iff both are > 0. This is
4169
// a rather pessimistic estimate for the most part, in particular for some
4170
// complex loops, but still not enough to capture all loops.
4171
if (ctrl_edge_out_cnt > 0 && data_edge_out_cnt > 0) {4172
return 2 * (ctrl_edge_out_cnt + data_edge_out_cnt);
4178
//------------------------------dump_head--------------------------------------
4179
// Dump 1 liner for loop header info
4180
void IdealLoopTree::dump_head() {4182
tty->print("Loop: N%d/N%d ", _head->_idx, _tail->_idx);4183
if (_irreducible) tty->print(" IRREDUCIBLE");4184
Node* entry = _head->is_Loop() ? _head->as_Loop()->skip_strip_mined(-1)->in(LoopNode::EntryControl)
4185
: _head->in(LoopNode::EntryControl);
4186
const Predicates predicates(entry);
4187
if (predicates.loop_limit_check_predicate_block()->is_non_empty()) {4188
tty->print(" limit_check");4190
if (UseProfiledLoopPredicate && predicates.profiled_loop_predicate_block()->is_non_empty()) {4191
tty->print(" profile_predicated");4193
if (UseLoopPredicate && predicates.loop_predicate_block()->is_non_empty()) {4194
tty->print(" predicated");4196
if (_head->is_CountedLoop()) {4197
CountedLoopNode *cl = _head->as_CountedLoop();
4198
tty->print(" counted");4200
Node* init_n = cl->init_trip();
4201
if (init_n != nullptr && init_n->is_Con())
4202
tty->print(" [%d,", cl->init_trip()->get_int());4204
tty->print(" [int,");4205
Node* limit_n = cl->limit();
4206
if (limit_n != nullptr && limit_n->is_Con())
4207
tty->print("%d),", cl->limit()->get_int());4209
tty->print("int),");4210
int stride_con = cl->stride_con();
4211
if (stride_con > 0) tty->print("+");4212
tty->print("%d", stride_con);4214
tty->print(" (%0.f iters) ", cl->profile_trip_cnt());4216
if (cl->is_pre_loop ()) tty->print(" pre" );4217
if (cl->is_main_loop()) tty->print(" main");4218
if (cl->is_post_loop()) tty->print(" post");4219
if (cl->is_vectorized_loop()) tty->print(" vector");4220
if (range_checks_present()) tty->print(" rc ");4222
if (_has_call) tty->print(" has_call");4223
if (_has_sfpt) tty->print(" has_sfpt");4224
if (_rce_candidate) tty->print(" rce");4225
if (_safepts != nullptr && _safepts->size() > 0) {4226
tty->print(" sfpts={"); _safepts->dump_simple(); tty->print(" }");4228
if (_required_safept != nullptr && _required_safept->size() > 0) {4229
tty->print(" req={"); _required_safept->dump_simple(); tty->print(" }");4232
tty->print(" body={"); _body.dump_simple(); tty->print(" }");4234
if (_head->is_Loop() && _head->as_Loop()->is_strip_mined()) {4235
tty->print(" strip_mined");4240
//------------------------------dump-------------------------------------------
4241
// Dump loops by loop tree
4242
void IdealLoopTree::dump() {4244
if (_child) _child->dump();
4245
if (_next) _next ->dump();
4250
static void log_loop_tree_helper(IdealLoopTree* root, IdealLoopTree* loop, CompileLog* log) {4252
if (loop->_child != nullptr) {4253
log->begin_head("loop_tree");4255
log_loop_tree_helper(root, loop->_child, log);
4256
log->tail("loop_tree");4257
assert(loop->_next == nullptr, "what?");
4259
} else if (loop != nullptr) {4260
Node* head = loop->_head;
4261
log->begin_head("loop");4262
log->print(" idx='%d' ", head->_idx);4263
if (loop->_irreducible) log->print("irreducible='1' ");4264
if (head->is_Loop()) {4265
if (head->as_Loop()->is_inner_loop()) log->print("inner_loop='1' ");4266
if (head->as_Loop()->is_partial_peel_loop()) log->print("partial_peel_loop='1' ");4267
} else if (head->is_CountedLoop()) {4268
CountedLoopNode* cl = head->as_CountedLoop();
4269
if (cl->is_pre_loop()) log->print("pre_loop='%d' ", cl->main_idx());4270
if (cl->is_main_loop()) log->print("main_loop='%d' ", cl->_idx);4271
if (cl->is_post_loop()) log->print("post_loop='%d' ", cl->main_idx());4274
log_loop_tree_helper(root, loop->_child, log);
4276
log_loop_tree_helper(root, loop->_next, log);
4280
void PhaseIdealLoop::log_loop_tree() {4281
if (C->log() != nullptr) {4282
log_loop_tree_helper(_ltree_root, _ltree_root, C->log());
4286
// Eliminate all Parse and Template Assertion Predicates that are not associated with a loop anymore. The eliminated
4287
// predicates will be removed during the next round of IGVN.
4288
void PhaseIdealLoop::eliminate_useless_predicates() {4289
if (C->parse_predicate_count() == 0 && C->template_assertion_predicate_count() == 0) {4290
return; // No predicates left.
4293
eliminate_useless_parse_predicates();
4294
eliminate_useless_template_assertion_predicates();
4297
// Eliminate all Parse Predicates that do not belong to a loop anymore by marking them useless. These will be removed
4298
// during the next round of IGVN.
4299
void PhaseIdealLoop::eliminate_useless_parse_predicates() {4300
mark_all_parse_predicates_useless();
4301
if (C->has_loops()) {4302
mark_loop_associated_parse_predicates_useful();
4304
add_useless_parse_predicates_to_igvn_worklist();
4307
void PhaseIdealLoop::mark_all_parse_predicates_useless() const {4308
for (int i = 0; i < C->parse_predicate_count(); i++) {4309
C->parse_predicate(i)->mark_useless();
4313
void PhaseIdealLoop::mark_loop_associated_parse_predicates_useful() {4314
for (LoopTreeIterator iterator(_ltree_root); !iterator.done(); iterator.next()) {4315
IdealLoopTree* loop = iterator.current();
4316
if (loop->can_apply_loop_predication()) {4317
mark_useful_parse_predicates_for_loop(loop);
4322
void PhaseIdealLoop::mark_useful_parse_predicates_for_loop(IdealLoopTree* loop) {4323
Node* entry = loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl);
4324
const Predicates predicates(entry);
4325
ParsePredicateIterator iterator(predicates);
4326
while (iterator.has_next()) {4327
iterator.next()->mark_useful();
4331
void PhaseIdealLoop::add_useless_parse_predicates_to_igvn_worklist() {4332
for (int i = 0; i < C->parse_predicate_count(); i++) {4333
ParsePredicateNode* parse_predicate_node = C->parse_predicate(i);
4334
if (parse_predicate_node->is_useless()) {4335
_igvn._worklist.push(parse_predicate_node);
4341
// Eliminate all Template Assertion Predicates that do not belong to their originally associated loop anymore by
4342
// replacing the Opaque4 node of the If node with true. These nodes will be removed during the next round of IGVN.
4343
void PhaseIdealLoop::eliminate_useless_template_assertion_predicates() {4344
Unique_Node_List useful_predicates;
4345
if (C->has_loops()) {4346
collect_useful_template_assertion_predicates(useful_predicates);
4348
eliminate_useless_template_assertion_predicates(useful_predicates);
4351
void PhaseIdealLoop::collect_useful_template_assertion_predicates(Unique_Node_List& useful_predicates) {4352
for (LoopTreeIterator iterator(_ltree_root); !iterator.done(); iterator.next()) {4353
IdealLoopTree* loop = iterator.current();
4354
if (loop->can_apply_loop_predication()) {4355
collect_useful_template_assertion_predicates_for_loop(loop, useful_predicates);
4360
void PhaseIdealLoop::collect_useful_template_assertion_predicates_for_loop(IdealLoopTree* loop,
4361
Unique_Node_List &useful_predicates) {4362
Node* entry = loop->_head->as_Loop()->skip_strip_mined()->in(LoopNode::EntryControl);
4363
const Predicates predicates(entry);
4364
if (UseProfiledLoopPredicate) {4365
const PredicateBlock* profiled_loop_predicate_block = predicates.profiled_loop_predicate_block();
4366
if (profiled_loop_predicate_block->has_parse_predicate()) {4367
IfProjNode* parse_predicate_proj = profiled_loop_predicate_block->parse_predicate_success_proj();
4368
get_assertion_predicates(parse_predicate_proj, useful_predicates, true);
4372
if (UseLoopPredicate) {4373
const PredicateBlock* loop_predicate_block = predicates.loop_predicate_block();
4374
if (loop_predicate_block->has_parse_predicate()) {4375
IfProjNode* parse_predicate_proj = loop_predicate_block->parse_predicate_success_proj();
4376
get_assertion_predicates(parse_predicate_proj, useful_predicates, true);
4381
void PhaseIdealLoop::eliminate_useless_template_assertion_predicates(Unique_Node_List& useful_predicates) {4382
for (int i = C->template_assertion_predicate_count(); i > 0; i--) {4383
Opaque4Node* opaque4_node = C->template_assertion_predicate_opaq_node(i - 1)->as_Opaque4();
4384
if (!useful_predicates.member(opaque4_node)) { // not in the useful list4385
_igvn.replace_node(opaque4_node, opaque4_node->in(2));
4390
// If a post or main loop is removed due to an assert predicate, the opaque that guards the loop is not needed anymore
4391
void PhaseIdealLoop::eliminate_useless_zero_trip_guard() {4392
if (_zero_trip_guard_opaque_nodes.size() == 0) {4395
Unique_Node_List useful_zero_trip_guard_opaques_nodes;
4396
for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {4397
IdealLoopTree* lpt = iter.current();
4398
if (lpt->_child == nullptr && lpt->is_counted()) {4399
CountedLoopNode* head = lpt->_head->as_CountedLoop();
4400
Node* opaque = head->is_canonical_loop_entry();
4401
if (opaque != nullptr) {4402
useful_zero_trip_guard_opaques_nodes.push(opaque);
4406
for (uint i = 0; i < _zero_trip_guard_opaque_nodes.size(); ++i) {4407
OpaqueZeroTripGuardNode* opaque = ((OpaqueZeroTripGuardNode*)_zero_trip_guard_opaque_nodes.at(i));
4408
DEBUG_ONLY(CountedLoopNode* guarded_loop = opaque->guarded_loop());
4409
if (!useful_zero_trip_guard_opaques_nodes.member(opaque)) {4410
IfNode* iff = opaque->if_node();
4411
IdealLoopTree* loop = get_loop(iff);
4412
while (loop != _ltree_root && loop != nullptr) {4413
loop = loop->_parent;
4415
if (loop == nullptr) {4416
// unreachable from _ltree_root: zero trip guard is in a newly discovered infinite loop.
4417
// We can't tell if the opaque node is useful or not
4418
assert(guarded_loop == nullptr || guarded_loop->is_in_infinite_subgraph(), "");
4420
assert(guarded_loop == nullptr, "");
4421
this->_igvn.replace_node(opaque, opaque->in(1));
4424
assert(guarded_loop != nullptr, "");
4429
//------------------------process_expensive_nodes-----------------------------
4430
// Expensive nodes have their control input set to prevent the GVN
4431
// from commoning them and as a result forcing the resulting node to
4432
// be in a more frequent path. Use CFG information here, to change the
4433
// control inputs so that some expensive nodes can be commoned while
4434
// not executed more frequently.
4435
bool PhaseIdealLoop::process_expensive_nodes() {4436
assert(OptimizeExpensiveOps, "optimization off?");
4438
// Sort nodes to bring similar nodes together
4439
C->sort_expensive_nodes();
4441
bool progress = false;
4443
for (int i = 0; i < C->expensive_count(); ) {4444
Node* n = C->expensive_node(i);
4446
// Find nodes similar to n
4448
for (; i < C->expensive_count() && Compile::cmp_expensive_nodes(n, C->expensive_node(i)) == 0; i++);
4450
// And compare them two by two
4451
for (int j = start; j < end; j++) {4452
Node* n1 = C->expensive_node(j);
4453
if (is_node_unreachable(n1)) {4456
for (int k = j+1; k < end; k++) {4457
Node* n2 = C->expensive_node(k);
4458
if (is_node_unreachable(n2)) {4462
assert(n1 != n2, "should be pair of nodes");
4464
Node* c1 = n1->in(0);
4465
Node* c2 = n2->in(0);
4467
Node* parent_c1 = c1;
4468
Node* parent_c2 = c2;
4470
// The call to get_early_ctrl_for_expensive() moves the
4471
// expensive nodes up but stops at loops that are in a if
4472
// branch. See whether we can exit the loop and move above the
4474
if (c1->is_Loop()) {4475
parent_c1 = c1->in(1);
4477
if (c2->is_Loop()) {4478
parent_c2 = c2->in(1);
4481
if (parent_c1 == parent_c2) {4482
_igvn._worklist.push(n1);
4483
_igvn._worklist.push(n2);
4487
// Look for identical expensive node up the dominator chain.
4488
if (is_dominator(c1, c2)) {4490
} else if (is_dominator(c2, c1)) {4492
} else if (parent_c1->is_Proj() && parent_c1->in(0)->is_If() &&
4493
parent_c2->is_Proj() && parent_c1->in(0) == parent_c2->in(0)) {4494
// Both branches have the same expensive node so move it up
4496
c1 = c2 = idom(parent_c1->in(0));
4498
// Do the actual moves
4499
if (n1->in(0) != c1) {4500
_igvn.replace_input_of(n1, 0, c1);
4503
if (n2->in(0) != c2) {4504
_igvn.replace_input_of(n2, 0, c2);
4515
// Goes over all children of the root of the loop tree. Check if any of them have a path
4516
// down to Root, that does not go via a NeverBranch exit.
4517
bool PhaseIdealLoop::only_has_infinite_loops() {4519
Unique_Node_List worklist;
4520
// start traversal at all loop heads of first-level loops
4521
for (IdealLoopTree* l = _ltree_root->_child; l != nullptr; l = l->_next) {4522
Node* head = l->_head;
4523
assert(head->is_Region(), "");
4524
worklist.push(head);
4526
return RegionNode::are_all_nodes_in_infinite_subgraph(worklist);
4531
//=============================================================================
4532
//----------------------------build_and_optimize-------------------------------
4533
// Create a PhaseLoop. Build the ideal Loop tree. Map each Ideal Node to
4534
// its corresponding LoopNode. If 'optimize' is true, do some loop cleanups.
4535
void PhaseIdealLoop::build_and_optimize() {4536
assert(!C->post_loop_opts_phase(), "no loop opts allowed");
4538
bool do_split_ifs = (_mode == LoopOptsDefault);
4539
bool skip_loop_opts = (_mode == LoopOptsNone);
4540
bool do_max_unroll = (_mode == LoopOptsMaxUnroll);
4543
int old_progress = C->major_progress();
4544
uint orig_worklist_size = _igvn._worklist.size();
4546
// Reset major-progress flag for the driver's heuristics
4547
C->clear_major_progress();
4550
// Capture for later assert
4551
uint unique = C->unique();
4553
_loop_work += unique;
4556
// True if the method has at least 1 irreducible loop
4557
_has_irreducible_loops = false;
4559
_created_loop_node = false;
4562
// Pre-grow the mapping from Nodes to IdealLoopTrees.
4563
_loop_or_ctrl.map(C->unique(), nullptr);
4564
memset(_loop_or_ctrl.adr(), 0, wordSize * C->unique());
4566
// Pre-build the top-level outermost loop tree entry
4567
_ltree_root = new IdealLoopTree( this, C->root(), C->root() );
4568
// Do not need a safepoint at the top level
4569
_ltree_root->_has_sfpt = 1;
4571
// Initialize Dominators.
4572
// Checked in clone_loop_predicate() during beautify_loops().
4575
_dom_depth = nullptr;
4578
// Empty pre-order array
4579
allocate_preorders();
4581
// Build a loop tree on the fly. Build a mapping from CFG nodes to
4582
// IdealLoopTree entries. Data nodes are NOT walked.
4584
// Check for bailout, and return
4589
// Verify that the has_loops() flag set at parse time is consistent
4590
// with the just built loop tree. With infinite loops, it could be
4591
// that one pass of loop opts only finds infinite loops, clears the
4592
// has_loops() flag but adds NeverBranch nodes so the next loop opts
4593
// verification pass finds a non empty loop tree. When the back edge
4594
// is an exception edge, parsing doesn't set has_loops().
4595
assert(_ltree_root->_child == nullptr || C->has_loops() || only_has_infinite_loops() || C->has_exception_backedge(), "parsing found no loops but there are some");
4596
// No loops after all
4597
if( !_ltree_root->_child && !_verify_only ) C->set_has_loops(false);
4599
// There should always be an outer loop containing the Root and Return nodes.
4600
// If not, we have a degenerate empty program. Bail out in this case.
4601
if (!has_node(C->root())) {4602
if (!_verify_only) {4603
C->clear_major_progress();
4604
assert(false, "empty program detected during loop optimization");
4605
C->record_method_not_compilable("empty program detected during loop optimization");4610
BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
4611
// Nothing to do, so get out
4612
bool stop_early = !C->has_loops() && !skip_loop_opts && !do_split_ifs && !do_max_unroll && !_verify_me &&
4613
!_verify_only && !bs->is_gc_specific_loop_opts_pass(_mode);
4614
bool do_expensive_nodes = C->should_optimize_expensive_nodes(_igvn);
4615
bool strip_mined_loops_expanded = bs->strip_mined_loops_expanded(_mode);
4616
if (stop_early && !do_expensive_nodes) {4620
// Set loop nesting depth
4621
_ltree_root->set_nest( 0 );
4623
// Split shared headers and insert loop landing pads.
4624
// Do not bother doing this on the Root loop of course.
4625
if( !_verify_me && !_verify_only && _ltree_root->_child ) {4626
C->print_method(PHASE_BEFORE_BEAUTIFY_LOOPS, 3);
4627
if( _ltree_root->_child->beautify_loops( this ) ) {4628
// Re-build loop tree!
4629
_ltree_root->_child = nullptr;
4630
_loop_or_ctrl.clear();
4631
reallocate_preorders();
4633
// Check for bailout, and return
4637
// Reset loop nesting depth
4638
_ltree_root->set_nest( 0 );
4640
C->print_method(PHASE_AFTER_BEAUTIFY_LOOPS, 3);
4644
// Build Dominators for elision of null checks & loop finding.
4645
// Since nodes do not have a slot for immediate dominator, make
4646
// a persistent side array for that info indexed on node->_idx.
4647
_idom_size = C->unique();
4648
_idom = NEW_RESOURCE_ARRAY( Node*, _idom_size );
4649
_dom_depth = NEW_RESOURCE_ARRAY( uint, _idom_size );
4650
_dom_stk = nullptr; // Allocated on demand in recompute_dom_depth
4651
memset( _dom_depth, 0, _idom_size * sizeof(uint) );
4655
if (!_verify_only) {4656
// As a side effect, Dominators removed any unreachable CFG paths
4657
// into RegionNodes. It doesn't do this test against Root, so
4659
for( uint i = 1; i < C->root()->req(); i++ ) {4660
if (!_loop_or_ctrl[C->root()->in(i)->_idx]) { // Dead path into Root?4661
_igvn.delete_input_of(C->root(), i);
4662
i--; // Rerun same iteration on compressed edges
4666
// Given dominators, try to find inner loops with calls that must
4667
// always be executed (call dominates loop tail). These loops do
4668
// not need a separate safepoint.
4670
_ltree_root->check_safepts(visited, cisstack);
4673
// Walk the DATA nodes and place into loops. Find earliest control
4674
// node. For CFG nodes, the _loop_or_ctrl array starts out and remains
4675
// holding the associated IdealLoopTree pointer. For DATA nodes, the
4676
// _loop_or_ctrl array holds the earliest legal controlling CFG node.
4678
// Allocate stack with enough space to avoid frequent realloc
4679
int stack_size = (C->live_nodes() >> 1) + 16; // (live_nodes>>1)+16 from Java2D stats
4680
Node_Stack nstack(stack_size);
4684
// Don't need C->root() on worklist since
4685
// it will be processed among C->top() inputs
4686
worklist.push(C->top());
4687
visited.set(C->top()->_idx); // Set C->top() as visited now
4688
build_loop_early( visited, worklist, nstack );
4690
// Given early legal placement, try finding counted loops. This placement
4691
// is good enough to discover most loop invariants.
4692
if (!_verify_me && !_verify_only && !strip_mined_loops_expanded) {4693
_ltree_root->counted_loop( this );
4696
// Find latest loop placement. Find ideal loop placement.
4698
init_dom_lca_tags();
4699
// Need C->root() on worklist when processing outs
4700
worklist.push(C->root());
4701
NOT_PRODUCT( C->verify_graph_edges(); )
4702
worklist.push(C->top());
4703
build_loop_late( visited, worklist, nstack );
4704
if (C->failing()) { return; }4707
C->restore_major_progress(old_progress);
4708
assert(C->unique() == unique, "verification _mode made Nodes? ? ?");
4709
assert(_igvn._worklist.size() == orig_worklist_size, "shouldn't push anything");
4713
// clear out the dead code after build_loop_late
4714
while (_deadlist.size()) {4715
_igvn.remove_globally_dead_node(_deadlist.pop());
4718
eliminate_useless_zero_trip_guard();
4721
assert(do_expensive_nodes, "why are we here?");
4722
if (process_expensive_nodes()) {4723
// If we made some progress when processing expensive nodes then
4724
// the IGVN may modify the graph in a way that will allow us to
4725
// make some more progress: we need to try processing expensive
4727
C->set_major_progress();
4732
// Some parser-inserted loop predicates could never be used by loop
4733
// predication or they were moved away from loop during some optimizations.
4734
// For example, peeling. Eliminate them before next loop optimizations.
4735
eliminate_useless_predicates();
4738
C->verify_graph_edges();
4739
if (_verify_me) { // Nested verify pass?4740
// Check to see if the verify _mode is broken
4741
assert(C->unique() == unique, "non-optimize _mode made Nodes? ? ?");
4744
DEBUG_ONLY( if (VerifyLoopOptimizations) { verify(); } );4745
if (TraceLoopOpts && C->has_loops()) {4746
_ltree_root->dump();
4750
if (skip_loop_opts) {4751
C->restore_major_progress(old_progress);
4755
if (do_max_unroll) {4756
for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {4757
IdealLoopTree* lpt = iter.current();
4758
if (lpt->is_innermost() && lpt->_allow_optimizations && !lpt->_has_call && lpt->is_counted()) {4759
lpt->compute_trip_count(this);
4760
if (!lpt->do_one_iteration_loop(this) &&
4761
!lpt->do_remove_empty_loop(this)) {4762
AutoNodeBudget node_budget(this);
4763
if (lpt->_head->as_CountedLoop()->is_normal_loop() &&
4764
lpt->policy_maximally_unroll(this)) {4765
memset( worklist.adr(), 0, worklist.max()*sizeof(Node*) );
4766
do_maximally_unroll(lpt, worklist);
4772
C->restore_major_progress(old_progress);
4776
if (bs->optimize_loops(this, _mode, visited, nstack, worklist)) {4780
if (ReassociateInvariants && !C->major_progress()) {4781
// Reassociate invariants and prep for split_thru_phi
4782
for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {4783
IdealLoopTree* lpt = iter.current();
4784
if (!lpt->is_loop()) {4787
Node* head = lpt->_head;
4788
if (!head->is_BaseCountedLoop() || !lpt->is_innermost()) continue;
4790
// check for vectorized loops, any reassociation of invariants was already done
4791
if (head->is_CountedLoop()) {4792
if (head->as_CountedLoop()->is_unroll_only()) {4795
AutoNodeBudget node_budget(this);
4796
lpt->reassociate_invariants(this);
4799
// Because RCE opportunities can be masked by split_thru_phi,
4800
// look for RCE candidates and inhibit split_thru_phi
4801
// on just their loop-phi's for this pass of loop opts
4802
if (SplitIfBlocks && do_split_ifs &&
4803
head->as_BaseCountedLoop()->is_valid_counted_loop(head->as_BaseCountedLoop()->bt()) &&
4804
(lpt->policy_range_check(this, true, T_LONG) ||
4805
(head->is_CountedLoop() && lpt->policy_range_check(this, true, T_INT)))) {4806
lpt->_rce_candidate = 1; // = true
4811
// Check for aggressive application of split-if and other transforms
4812
// that require basic-block info (like cloning through Phi's)
4813
if (!C->major_progress() && SplitIfBlocks && do_split_ifs) {4815
split_if_with_blocks( visited, nstack);
4816
DEBUG_ONLY( if (VerifyLoopOptimizations) { verify(); } );4819
if (!C->major_progress() && do_expensive_nodes && process_expensive_nodes()) {4820
C->set_major_progress();
4823
// Perform loop predication before iteration splitting
4824
if (UseLoopPredicate && C->has_loops() && !C->major_progress() && (C->parse_predicate_count() > 0)) {4825
_ltree_root->_child->loop_predication(this);
4828
if (OptimizeFill && UseLoopPredicate && C->has_loops() && !C->major_progress()) {4829
if (do_intrinsify_fill()) {4830
C->set_major_progress();
4834
// Perform iteration-splitting on inner loops. Split iterations to avoid
4835
// range checks or one-shot null checks.
4837
// If split-if's didn't hack the graph too bad (no CFG changes)
4838
// then do loop opts.
4839
if (C->has_loops() && !C->major_progress()) {4840
memset( worklist.adr(), 0, worklist.max()*sizeof(Node*) );
4841
_ltree_root->_child->iteration_split( this, worklist );
4842
// No verify after peeling! GCM has hoisted code out of the loop.
4843
// After peeling, the hoisted code could sink inside the peeled area.
4844
// The peeling code does not try to recompute the best location for
4845
// all the code before the peeled area, so the verify pass will always
4846
// complain about it.
4849
// Check for bailout, and return
4854
// Do verify graph edges in any case
4855
NOT_PRODUCT( C->verify_graph_edges(); );
4857
if (!do_split_ifs) {4858
// We saw major progress in Split-If to get here. We forced a
4859
// pass with unrolling and not split-if, however more split-if's
4860
// might make progress. If the unrolling didn't make progress
4861
// then the major-progress flag got cleared and we won't try
4862
// another round of Split-If. In particular the ever-common
4863
// instance-of/check-cast pattern requires at least 2 rounds of
4864
// Split-If to clear out.
4865
C->set_major_progress();
4868
// Repeat loop optimizations if new loops were seen
4869
if (created_loop_node()) {4870
C->set_major_progress();
4873
// Keep loop predicates and perform optimizations with them
4874
// until no more loop optimizations could be done.
4875
// After that switch predicates off and do more loop optimizations.
4876
if (!C->major_progress() && (C->parse_predicate_count() > 0)) {4877
C->mark_parse_predicate_nodes_useless(_igvn);
4878
assert(C->parse_predicate_count() == 0, "should be zero now");
4879
if (TraceLoopOpts) {4880
tty->print_cr("PredicatesOff");4882
C->set_major_progress();
4885
// Auto-vectorize main-loop
4886
if (C->do_superword() && C->has_loops() && !C->major_progress()) {4887
Compile::TracePhase tp("autoVectorize", &timers[_t_autoVectorize]);4889
// Shared data structures for all AutoVectorizations, to reduce allocations
4891
VSharedData vshared;
4892
for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {4893
IdealLoopTree* lpt = iter.current();
4894
AutoVectorizeStatus status = auto_vectorize(lpt, vshared);
4896
if (status == AutoVectorizeStatus::TriedAndFailed) {4897
// We tried vectorization, but failed. From now on only unroll the loop.
4898
CountedLoopNode* cl = lpt->_head->as_CountedLoop();
4899
if (cl->has_passed_slp()) {4900
C->set_major_progress();
4901
cl->set_notpassed_slp();
4902
cl->mark_do_unroll_only();
4908
// Move UnorderedReduction out of counted loop. Can be introduced by AutoVectorization.
4909
if (C->has_loops() && !C->major_progress()) {4910
for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) {4911
IdealLoopTree* lpt = iter.current();
4912
if (lpt->is_counted() && lpt->is_innermost()) {4913
move_unordered_reduction_out_of_loop(lpt);
4920
//------------------------------print_statistics-------------------------------
4921
int PhaseIdealLoop::_loop_invokes=0;// Count of PhaseIdealLoop invokes
4922
int PhaseIdealLoop::_loop_work=0; // Sum of PhaseIdealLoop x unique
4923
volatile int PhaseIdealLoop::_long_loop_candidates=0; // Number of long loops seen
4924
volatile int PhaseIdealLoop::_long_loop_nests=0; // Number of long loops successfully transformed to a nest
4925
volatile int PhaseIdealLoop::_long_loop_counted_loops=0; // Number of long loops successfully transformed to a counted loop
4926
void PhaseIdealLoop::print_statistics() {4927
tty->print_cr("PhaseIdealLoop=%d, sum _unique=%d, long loops=%d/%d/%d", _loop_invokes, _loop_work, _long_loop_counted_loops, _long_loop_nests, _long_loop_candidates);4932
// Build a verify-only PhaseIdealLoop, and see that it agrees with "this".
4933
void PhaseIdealLoop::verify() const {4935
int old_progress = C->major_progress();
4936
bool success = true;
4938
PhaseIdealLoop phase_verify(_igvn, this);
4939
if (C->failing()) return;
4941
// Verify ctrl and idom of every node.
4942
success &= verify_idom_and_nodes(C->root(), &phase_verify);
4944
// Verify loop-tree.
4945
success &= _ltree_root->verify_tree(phase_verify._ltree_root);
4947
assert(success, "VerifyLoopOptimizations failed");
4949
// Major progress was cleared by creating a verify version of PhaseIdealLoop.
4950
C->restore_major_progress(old_progress);
4953
// Perform a BFS starting at n, through all inputs.
4954
// Call verify_idom and verify_node on all nodes of BFS traversal.
4955
bool PhaseIdealLoop::verify_idom_and_nodes(Node* root, const PhaseIdealLoop* phase_verify) const {4956
Unique_Node_List worklist;
4957
worklist.push(root);
4958
bool success = true;
4959
for (uint i = 0; i < worklist.size(); i++) {4960
Node* n = worklist.at(i);
4962
success &= verify_idom(n, phase_verify);
4963
success &= verify_loop_ctrl(n, phase_verify);
4965
for (uint j = 0; j < n->req(); j++) {4966
if (n->in(j) != nullptr) {4967
worklist.push(n->in(j));
4974
// Verify dominator structure (IDOM).
4975
bool PhaseIdealLoop::verify_idom(Node* n, const PhaseIdealLoop* phase_verify) const {4976
// Verify IDOM for all CFG nodes (except root).
4977
if (!n->is_CFG() || n->is_Root()) {4978
return true; // pass
4981
if (n->_idx >= _idom_size) {4982
tty->print("CFG Node with no idom: ");4984
return false; // fail
4987
Node* id = idom_no_update(n);
4988
Node* id_verify = phase_verify->idom_no_update(n);
4989
if (id != id_verify) {4990
tty->print("Mismatching idom for node: ");4992
tty->print(" We have idom: ");4994
tty->print(" Verify has idom: ");4997
return false; // fail
4999
return true; // pass
5002
// Verify "_loop_or_ctrl": control and loop membership.
5003
// (0) _loop_or_ctrl[i] == nullptr -> node not reachable.
5004
// (1) has_ctrl -> check lowest bit. 1 -> data node. 0 -> ctrl node.
5005
// (2) has_ctrl true: get_ctrl_no_update returns ctrl of data node.
5006
// (3) has_ctrl false: get_loop_idx returns IdealLoopTree for ctrl node.
5007
bool PhaseIdealLoop::verify_loop_ctrl(Node* n, const PhaseIdealLoop* phase_verify) const {5008
const uint i = n->_idx;
5009
// The loop-tree was built from def to use (top-down).
5010
// The verification happens from use to def (bottom-up).
5011
// We may thus find nodes during verification that are not in the loop-tree.
5012
if (_loop_or_ctrl[i] == nullptr || phase_verify->_loop_or_ctrl[i] == nullptr) {5013
if (_loop_or_ctrl[i] != nullptr || phase_verify->_loop_or_ctrl[i] != nullptr) {5014
tty->print_cr("Was reachable in only one. this %d, verify %d.",5015
_loop_or_ctrl[i] != nullptr, phase_verify->_loop_or_ctrl[i] != nullptr);
5017
return false; // fail
5019
// Not reachable for both.
5020
return true; // pass
5023
if (n->is_CFG() == has_ctrl(n)) {5024
tty->print_cr("Exactly one should be true: %d for is_CFG, %d for has_ctrl.", n->is_CFG(), has_ctrl(n));5026
return false; // fail
5029
if (has_ctrl(n) != phase_verify->has_ctrl(n)) {5030
tty->print_cr("Mismatch has_ctrl: %d for this, %d for verify.", has_ctrl(n), phase_verify->has_ctrl(n));5032
return false; // fail
5033
} else if (has_ctrl(n)) {5034
assert(phase_verify->has_ctrl(n), "sanity");
5035
// n is a data node.
5036
// Verify that its ctrl is the same.
5038
// Broken part of VerifyLoopOptimizations (A)
5040
// BUG, wrong control set for example in
5041
// PhaseIdealLoop::split_if_with_blocks
5042
// at "set_ctrl(x, new_ctrl);"
5044
if( _loop_or_ctrl[i] != loop_verify->_loop_or_ctrl[i] &&
5045
get_ctrl_no_update(n) != loop_verify->get_ctrl_no_update(n) ) {5046
tty->print("Mismatched control setting for: ");5048
if( fail++ > 10 ) return;
5049
Node *c = get_ctrl_no_update(n);
5050
tty->print("We have it as: ");5051
if( c->in(0) ) c->dump();
5052
else tty->print_cr("N%d",c->_idx);5053
tty->print("Verify thinks: ");5054
if( loop_verify->has_ctrl(n) )
5055
loop_verify->get_ctrl_no_update(n)->dump();
5057
loop_verify->get_loop_idx(n)->dump();
5061
return true; // pass
5063
assert(!phase_verify->has_ctrl(n), "sanity");
5064
// n is a ctrl node.
5065
// Verify that not has_ctrl, and that get_loop_idx is the same.
5067
// Broken part of VerifyLoopOptimizations (B)
5069
// NeverBranch node for example is added to loop outside its scope.
5070
// Once we run build_loop_tree again, it is added to the correct loop.
5072
if (!C->major_progress()) {5073
// Loop selection can be messed up if we did a major progress
5074
// operation, like split-if. Do not verify in that case.
5075
IdealLoopTree *us = get_loop_idx(n);
5076
IdealLoopTree *them = loop_verify->get_loop_idx(n);
5077
if( us->_head != them->_head || us->_tail != them->_tail ) {5078
tty->print("Unequals loops for: ");5080
if( fail++ > 10 ) return;
5081
tty->print("We have it as: ");5083
tty->print("Verify thinks: ");5089
return true; // pass
5093
static int compare_tree(IdealLoopTree* const& a, IdealLoopTree* const& b) {5094
assert(a != nullptr && b != nullptr, "must be");
5095
return a->_head->_idx - b->_head->_idx;
5098
GrowableArray<IdealLoopTree*> IdealLoopTree::collect_sorted_children() const {5099
GrowableArray<IdealLoopTree*> children;
5100
IdealLoopTree* child = _child;
5101
while (child != nullptr) {5102
assert(child->_parent == this, "all must be children of this");
5103
children.insert_sorted<compare_tree>(child);
5104
child = child->_next;
5109
// Verify that tree structures match. Because the CFG can change, siblings
5110
// within the loop tree can be reordered. We attempt to deal with that by
5111
// reordering the verify's loop tree if possible.
5112
bool IdealLoopTree::verify_tree(IdealLoopTree* loop_verify) const {5113
assert(_head == loop_verify->_head, "mismatched loop head");
5114
assert(this->_parent != nullptr || this->_next == nullptr, "is_root_loop implies has_no_sibling");
5116
// Collect the children
5117
GrowableArray<IdealLoopTree*> children = collect_sorted_children();
5118
GrowableArray<IdealLoopTree*> children_verify = loop_verify->collect_sorted_children();
5120
bool success = true;
5122
// Compare the two children lists
5123
for (int i = 0, j = 0; i < children.length() || j < children_verify.length(); ) {5124
IdealLoopTree* child = nullptr;
5125
IdealLoopTree* child_verify = nullptr;
5126
// Read from both lists, if possible.
5127
if (i < children.length()) {5128
child = children.at(i);
5130
if (j < children_verify.length()) {5131
child_verify = children_verify.at(j);
5133
assert(child != nullptr || child_verify != nullptr, "must find at least one");
5134
if (child != nullptr && child_verify != nullptr && child->_head != child_verify->_head) {5135
// We found two non-equal children. Select the smaller one.
5136
if (child->_head->_idx < child_verify->_head->_idx) {5137
child_verify = nullptr;
5142
// Process the two children, or potentially log the failure if we only found one.
5143
if (child_verify == nullptr) {5144
if (child->_irreducible && Compile::current()->major_progress()) {5145
// Irreducible loops can pick a different header (one of its entries).
5147
tty->print_cr("We have a loop that verify does not have");5151
i++; // step for this
5152
} else if (child == nullptr) {5153
if (child_verify->_irreducible && Compile::current()->major_progress()) {5154
// Irreducible loops can pick a different header (one of its entries).
5155
} else if (child_verify->_head->as_Region()->is_in_infinite_subgraph()) {5156
// Infinite loops do not get attached to the loop-tree on their first visit.
5157
// "this" runs before "loop_verify". It is thus possible that we find the
5158
// infinite loop only for "child_verify". Only finding it with "child" would
5159
// mean that we lost it, which is not ok.
5161
tty->print_cr("Verify has a loop that we do not have");5162
child_verify->dump();
5165
j++; // step for verify
5167
assert(child->_head == child_verify->_head, "We have both and they are equal");
5168
success &= child->verify_tree(child_verify); // Recursion
5169
i++; // step for this
5170
j++; // step for verify
5174
// Broken part of VerifyLoopOptimizations (D)
5176
// split_if has to update the _tail, if it is modified. But that is done by
5177
// checking to what loop the iff belongs to. That info can be wrong, and then
5178
// we do not update the _tail correctly.
5180
Node *tail = _tail; // Inline a non-updating version of
5181
while( !tail->in(0) ) // the 'tail()' call.
5183
assert( tail == loop->_tail, "mismatched loop tail" );
5186
if (_head->is_CountedLoop()) {5187
CountedLoopNode *cl = _head->as_CountedLoop();
5189
Node* ctrl = cl->init_control();
5190
Node* back = cl->back_control();
5191
assert(ctrl != nullptr && ctrl->is_CFG(), "sane loop in-ctrl");
5192
assert(back != nullptr && back->is_CFG(), "sane loop backedge");
5193
cl->loopexit(); // assert implied
5196
// Broken part of VerifyLoopOptimizations (E)
5198
// PhaseIdealLoop::split_thru_region creates new nodes for loop that are not added
5199
// to the loop body. Or maybe they are not added to the correct loop.
5200
// at "Node* x = n->clone();"
5202
// Innermost loops need to verify loop bodies,
5203
// but only if no 'major_progress'
5205
if (!Compile::current()->major_progress() && _child == nullptr) {5206
for( uint i = 0; i < _body.size(); i++ ) {5207
Node *n = _body.at(i);
5208
if (n->outcnt() == 0) continue; // Ignore dead
5210
for( j = 0; j < loop->_body.size(); j++ )
5211
if( loop->_body.at(j) == n )
5213
if( j == loop->_body.size() ) { // Not found in loop body5214
// Last ditch effort to avoid assertion: Its possible that we
5215
// have some users (so outcnt not zero) but are still dead.
5216
// Try to find from root.
5217
if (Compile::current()->root()->find(n->_idx)) {5219
tty->print("We have that verify does not: ");5224
for( uint i2 = 0; i2 < loop->_body.size(); i2++ ) {5225
Node *n = loop->_body.at(i2);
5226
if (n->outcnt() == 0) continue; // Ignore dead
5228
for( j = 0; j < _body.size(); j++ )
5229
if( _body.at(j) == n )
5231
if( j == _body.size() ) { // Not found in loop body5232
// Last ditch effort to avoid assertion: Its possible that we
5233
// have some users (so outcnt not zero) but are still dead.
5234
// Try to find from root.
5235
if (Compile::current()->root()->find(n->_idx)) {5237
tty->print("Verify has that we do not: ");5242
assert( !fail, "loop body mismatch" );
5249
//------------------------------set_idom---------------------------------------
5250
void PhaseIdealLoop::set_idom(Node* d, Node* n, uint dom_depth) {5251
_nesting.check(); // Check if a potential reallocation in the resource arena is safe
5253
if (idx >= _idom_size) {5254
uint newsize = next_power_of_2(idx);
5255
_idom = REALLOC_RESOURCE_ARRAY( Node*, _idom,_idom_size,newsize);
5256
_dom_depth = REALLOC_RESOURCE_ARRAY( uint, _dom_depth,_idom_size,newsize);
5257
memset( _dom_depth + _idom_size, 0, (newsize - _idom_size) * sizeof(uint) );
5258
_idom_size = newsize;
5261
_dom_depth[idx] = dom_depth;
5264
//------------------------------recompute_dom_depth---------------------------------------
5265
// The dominator tree is constructed with only parent pointers.
5266
// This recomputes the depth in the tree by first tagging all
5267
// nodes as "no depth yet" marker. The next pass then runs up
5268
// the dom tree from each node marked "no depth yet", and computes
5269
// the depth on the way back down.
5270
void PhaseIdealLoop::recompute_dom_depth() {5271
uint no_depth_marker = C->unique();
5273
// Initialize depth to "no depth yet" and realize all lazy updates
5274
for (i = 0; i < _idom_size; i++) {5275
// Only indices with a _dom_depth has a Node* or null (otherwise uninitialized).
5276
if (_dom_depth[i] > 0 && _idom[i] != nullptr) {5277
_dom_depth[i] = no_depth_marker;
5279
// heal _idom if it has a fwd mapping in _loop_or_ctrl
5280
if (_idom[i]->in(0) == nullptr) {5285
if (_dom_stk == nullptr) {5286
uint init_size = C->live_nodes() / 100; // Guess that 1/100 is a reasonable initial size.
5287
if (init_size < 10) init_size = 10;
5288
_dom_stk = new GrowableArray<uint>(init_size);
5290
// Compute new depth for each node.
5291
for (i = 0; i < _idom_size; i++) {5293
// Run up the dom tree to find a node with a depth
5294
while (_dom_depth[j] == no_depth_marker) {5298
// Compute the depth on the way back down this tree branch
5299
uint dd = _dom_depth[j] + 1;
5300
while (_dom_stk->length() > 0) {5301
uint j = _dom_stk->pop();
5308
//------------------------------sort-------------------------------------------
5309
// Insert 'loop' into the existing loop tree. 'innermost' is a leaf of the
5310
// loop tree, not the root.
5311
IdealLoopTree *PhaseIdealLoop::sort( IdealLoopTree *loop, IdealLoopTree *innermost ) {5312
if( !innermost ) return loop; // New innermost loop
5314
int loop_preorder = get_preorder(loop->_head); // Cache pre-order number
5315
assert( loop_preorder, "not yet post-walked loop" );
5316
IdealLoopTree **pp = &innermost; // Pointer to previous next-pointer
5317
IdealLoopTree *l = *pp; // Do I go before or after 'l'?
5319
// Insert at start of list
5320
while( l ) { // Insertion sort based on pre-order5321
if( l == loop ) return innermost; // Already on list!
5322
int l_preorder = get_preorder(l->_head); // Cache pre-order number
5323
assert( l_preorder, "not yet post-walked l" );
5324
// Check header pre-order number to figure proper nesting
5325
if( loop_preorder > l_preorder )
5326
break; // End of insertion
5327
// If headers tie (e.g., shared headers) check tail pre-order numbers.
5328
// Since I split shared headers, you'd think this could not happen.
5329
// BUT: I must first do the preorder numbering before I can discover I
5330
// have shared headers, so the split headers all get the same preorder
5331
// number as the RegionNode they split from.
5332
if( loop_preorder == l_preorder &&
5333
get_preorder(loop->_tail) < get_preorder(l->_tail) )
5334
break; // Also check for shared headers (same pre#)
5335
pp = &l->_parent; // Chain up list
5339
// Point predecessor to me
5341
// Point me to successor
5342
IdealLoopTree *p = loop->_parent;
5343
loop->_parent = l; // Point me to successor
5344
if( p ) sort( p, innermost ); // Insert my parents into list as well
5348
//------------------------------build_loop_tree--------------------------------
5349
// I use a modified Vick/Tarjan algorithm. I need pre- and a post- visit
5350
// bits. The _loop_or_ctrl[] array is mapped by Node index and holds a null for
5351
// not-yet-pre-walked, pre-order # for pre-but-not-post-walked and holds the
5352
// tightest enclosing IdealLoopTree for post-walked.
5354
// During my forward walk I do a short 1-layer lookahead to see if I can find
5355
// a loop backedge with that doesn't have any work on the backedge. This
5356
// helps me construct nested loops with shared headers better.
5358
// Once I've done the forward recursion, I do the post-work. For each child
5359
// I check to see if there is a backedge. Backedges define a loop! I
5360
// insert an IdealLoopTree at the target of the backedge.
5362
// During the post-work I also check to see if I have several children
5363
// belonging to different loops. If so, then this Node is a decision point
5364
// where control flow can choose to change loop nests. It is at this
5365
// decision point where I can figure out how loops are nested. At this
5366
// time I can properly order the different loop nests from my children.
5367
// Note that there may not be any backedges at the decision point!
5369
// Since the decision point can be far removed from the backedges, I can't
5370
// order my loops at the time I discover them. Thus at the decision point
5371
// I need to inspect loop header pre-order numbers to properly nest my
5372
// loops. This means I need to sort my childrens' loops by pre-order.
5373
// The sort is of size number-of-control-children, which generally limits
5374
// it to size 2 (i.e., I just choose between my 2 target loops).
5375
void PhaseIdealLoop::build_loop_tree() {5376
// Allocate stack of size C->live_nodes()/2 to avoid frequent realloc
5377
GrowableArray <Node *> bltstack(C->live_nodes() >> 1);
5378
Node *n = C->root();
5383
while ( ( stack_size = bltstack.length() ) != 0 ) {5384
n = bltstack.top(); // Leave node on stack
5385
if ( !is_visited(n) ) {5386
// ---- Pre-pass Work ----
5387
// Pre-walked but not post-walked nodes need a pre_order number.
5389
set_preorder_visited( n, pre_order ); // set as visited
5391
// ---- Scan over children ----
5392
// Scan first over control projections that lead to loop headers.
5393
// This helps us find inner-to-outer loops with shared headers better.
5395
// Scan children's children for loop headers.
5396
for ( int i = n->outcnt() - 1; i >= 0; --i ) {5397
Node* m = n->raw_out(i); // Child
5398
if( m->is_CFG() && !is_visited(m) ) { // Only for CFG children5399
// Scan over children's children to find loop
5400
for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {5401
Node* l = m->fast_out(j);
5402
if( is_visited(l) && // Been visited?
5403
!is_postvisited(l) && // But not post-visited
5404
get_preorder(l) < pre_order ) { // And smaller pre-order5405
// Found! Scan the DFS down this path before doing other paths
5414
else if ( !is_postvisited(n) ) {5415
// Note: build_loop_tree_impl() adds out edges on rare occasions,
5416
// such as com.sun.rsasign.am::a.
5417
// For non-recursive version, first, process current children.
5418
// On next iteration, check if additional children were added.
5419
for ( int k = n->outcnt() - 1; k >= 0; --k ) {5420
Node* u = n->raw_out(k);
5421
if ( u->is_CFG() && !is_visited(u) ) {5425
if ( bltstack.length() == stack_size ) {5426
// There were no additional children, post visit node now
5427
(void)bltstack.pop(); // Remove node from stack
5428
pre_order = build_loop_tree_impl( n, pre_order );
5429
// Check for bailout
5433
// Check to grow _preorders[] array for the case when
5434
// build_loop_tree_impl() adds new nodes.
5435
check_grow_preorders();
5439
(void)bltstack.pop(); // Remove post-visited node from stack
5442
DEBUG_ONLY(verify_regions_in_irreducible_loops();)
5445
//------------------------------build_loop_tree_impl---------------------------
5446
int PhaseIdealLoop::build_loop_tree_impl( Node *n, int pre_order ) {5447
// ---- Post-pass Work ----
5448
// Pre-walked but not post-walked nodes need a pre_order number.
5450
// Tightest enclosing loop for this Node
5451
IdealLoopTree *innermost = nullptr;
5453
// For all children, see if any edge is a backedge. If so, make a loop
5454
// for it. Then find the tightest enclosing loop for the self Node.
5455
for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {5456
Node* m = n->fast_out(i); // Child
5457
if( n == m ) continue; // Ignore control self-cycles
5458
if( !m->is_CFG() ) continue;// Ignore non-CFG edges
5460
IdealLoopTree *l; // Child's loop
5461
if( !is_postvisited(m) ) { // Child visited but not post-visited?5463
assert( get_preorder(m) < pre_order, "should be backedge" );
5464
// Check for the RootNode, which is already a LoopNode and is allowed
5465
// to have multiple "backedges".
5466
if( m == C->root()) { // Found the root?5467
l = _ltree_root; // Root is the outermost LoopNode
5468
} else { // Else found a nested loop5469
// Insert a LoopNode to mark this loop.
5470
l = new IdealLoopTree(this, m, n);
5471
} // End of Else found a nested loop
5472
if( !has_loop(m) ) // If 'm' does not already have a loop set
5473
set_loop(m, l); // Set loop header to loop now
5475
} else { // Else not a nested loop5476
if (!_loop_or_ctrl[m->_idx]) continue; // Dead code has no loop
5477
IdealLoopTree* m_loop = get_loop(m);
5478
l = m_loop; // Get previously determined loop
5479
// If successor is header of a loop (nest), move up-loop till it
5480
// is a member of some outer enclosing loop. Since there are no
5481
// shared headers (I've split them already) I only need to go up
5483
while( l && l->_head == m ) // Successor heads loop?
5484
l = l->_parent; // Move up 1 for me
5485
// If this loop is not properly parented, then this loop
5486
// has no exit path out, i.e. its an infinite loop.
5488
// Make loop "reachable" from root so the CFG is reachable. Basically
5489
// insert a bogus loop exit that is never taken. 'm', the loop head,
5490
// points to 'n', one (of possibly many) fall-in paths. There may be
5491
// many backedges as well.
5493
// Here I set the loop to be the root loop. I could have, after
5494
// inserting a bogus loop exit, restarted the recursion and found my
5495
// new loop exit. This would make the infinite loop a first-class
5496
// loop and it would then get properly optimized. What's the use of
5497
// optimizing an infinite loop?
5498
l = _ltree_root; // Oops, found infinite loop
5500
if (!_verify_only) {5501
// Insert the NeverBranch between 'm' and it's control user.
5502
NeverBranchNode *iff = new NeverBranchNode( m );
5503
_igvn.register_new_node_with_optimizer(iff);
5504
set_loop(iff, m_loop);
5505
Node *if_t = new CProjNode( iff, 0 );
5506
_igvn.register_new_node_with_optimizer(if_t);
5507
set_loop(if_t, m_loop);
5509
Node* cfg = nullptr; // Find the One True Control User of m
5510
for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {5511
Node* x = m->fast_out(j);
5512
if (x->is_CFG() && x != m && x != iff)
5515
assert(cfg != nullptr, "must find the control user of m");
5516
uint k = 0; // Probably cfg->in(0)
5517
while( cfg->in(k) != m ) k++; // But check in case cfg is a Region
5518
_igvn.replace_input_of(cfg, k, if_t); // Now point to NeverBranch
5520
// Now create the never-taken loop exit
5521
Node *if_f = new CProjNode( iff, 1 );
5522
_igvn.register_new_node_with_optimizer(if_f);
5524
// Find frame ptr for Halt. Relies on the optimizer
5525
// V-N'ing. Easier and quicker than searching through
5526
// the program structure.
5527
Node *frame = new ParmNode( C->start(), TypeFunc::FramePtr );
5528
_igvn.register_new_node_with_optimizer(frame);
5529
// Halt & Catch Fire
5530
Node* halt = new HaltNode(if_f, frame, "never-taken loop exit reached");
5531
_igvn.register_new_node_with_optimizer(halt);
5533
_igvn.add_input_to(C->root(), halt);
5535
set_loop(C->root(), _ltree_root);
5538
if (is_postvisited(l->_head)) {5539
// We are currently visiting l, but its head has already been post-visited.
5540
// l is irreducible: we just found a second entry m.
5541
_has_irreducible_loops = true;
5542
RegionNode* secondary_entry = m->as_Region();
5543
DEBUG_ONLY(secondary_entry->verify_can_be_irreducible_entry();)
5545
// Walk up the loop-tree, mark all loops that are already post-visited as irreducible
5546
// Since m is a secondary entry to them all.
5547
while( is_postvisited(l->_head) ) {5548
l->_irreducible = 1; // = true
5549
RegionNode* head = l->_head->as_Region();
5550
DEBUG_ONLY(head->verify_can_be_irreducible_entry();)
5552
// Check for bad CFG here to prevent crash, and bailout of compile
5555
if (TraceLoopOpts) {5556
tty->print_cr("bailout: unhandled CFG: infinite irreducible loop");5560
// This is a rare case that we do not want to handle in C2.
5561
C->record_method_not_compilable("unhandled CFG detected during loop optimization");5566
if (!_verify_only) {5567
C->set_has_irreducible_loop(_has_irreducible_loops);
5570
// This Node might be a decision point for loops. It is only if
5571
// it's children belong to several different loops. The sort call
5572
// does a trivial amount of work if there is only 1 child or all
5573
// children belong to the same loop. If however, the children
5574
// belong to different loops, the sort call will properly set the
5575
// _parent pointers to show how the loops nest.
5577
// In any case, it returns the tightest enclosing loop.
5578
innermost = sort( l, innermost );
5581
// Def-use info will have some dead stuff; dead stuff will have no
5584
// Am I a loop header? If so fix up my parent's child and next ptrs.
5585
if( innermost && innermost->_head == n ) {5586
assert( get_loop(n) == innermost, "" );
5587
IdealLoopTree *p = innermost->_parent;
5588
IdealLoopTree *l = innermost;
5589
while( p && l->_head == n ) {5590
l->_next = p->_child; // Put self on parents 'next child'
5591
p->_child = l; // Make self as first child of parent
5592
l = p; // Now walk up the parent chain
5596
// Note that it is possible for a LoopNode to reach here, if the
5597
// backedge has been made unreachable (hence the LoopNode no longer
5598
// denotes a Loop, and will eventually be removed).
5600
// Record tightest enclosing loop for self. Mark as post-visited.
5601
set_loop(n, innermost);
5602
// Also record has_call flag early on
5604
if( n->is_Call() && !n->is_CallLeaf() && !n->is_macro() ) {5605
// Do not count uncommon calls
5606
if( !n->is_CallStaticJava() || !n->as_CallStaticJava()->_name ) {5607
Node *iff = n->in(0)->in(0);
5608
// No any calls for vectorized loops.
5609
if (C->do_superword() ||
5611
(n->in(0)->Opcode() == Op_IfFalse && (1.0 - iff->as_If()->_prob) >= 0.01) ||
5612
iff->as_If()->_prob >= 0.01) {5613
innermost->_has_call = 1;
5616
} else if( n->is_Allocate() && n->as_Allocate()->_is_scalar_replaceable ) {5617
// Disable loop optimizations if the loop has a scalar replaceable
5618
// allocation. This disabling may cause a potential performance lost
5619
// if the allocation is not eliminated for some reason.
5620
innermost->_allow_optimizations = false;
5621
innermost->_has_call = 1; // = true
5622
} else if (n->Opcode() == Op_SafePoint) {5623
// Record all safepoints in this loop.
5624
if (innermost->_safepts == nullptr) innermost->_safepts = new Node_List();
5625
innermost->_safepts->push(n);
5630
// Flag as post-visited now
5636
//--------------------------verify_regions_in_irreducible_loops----------------
5637
// Iterate down from Root through CFG, verify for every region:
5638
// if it is in an irreducible loop it must be marked as such
5639
void PhaseIdealLoop::verify_regions_in_irreducible_loops() {5641
if (!_has_irreducible_loops) {5642
// last build_loop_tree has not found any irreducible loops
5643
// hence no region has to be marked is_in_irreduible_loop
5647
RootNode* root = C->root();
5648
Unique_Node_List worklist; // visit all nodes once
5649
worklist.push(root);
5650
bool failure = false;
5651
for (uint i = 0; i < worklist.size(); i++) {5652
Node* n = worklist.at(i);
5653
if (n->is_Region()) {5654
RegionNode* region = n->as_Region();
5655
if (is_in_irreducible_loop(region) &&
5656
region->loop_status() == RegionNode::LoopStatus::Reducible) {5658
tty->print("irreducible! ");5662
for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {5663
Node* use = n->fast_out(j);
5664
if (use->is_CFG()) {5665
worklist.push(use); // push if was not pushed before
5669
assert(!failure, "region in irreducible loop was marked as reducible");
5672
//---------------------------is_in_irreducible_loop-------------------------
5673
// Analogous to ciTypeFlow::Block::is_in_irreducible_loop
5674
bool PhaseIdealLoop::is_in_irreducible_loop(RegionNode* region) {5675
if (!_has_irreducible_loops) {5676
return false; // no irreducible loop in graph
5678
IdealLoopTree* l = get_loop(region); // l: innermost loop that contains region
5680
if (l->_irreducible) {5681
return true; // found it
5683
if (l == _ltree_root) {5684
return false; // reached root, terimnate
5687
} while (l != nullptr);
5688
assert(region->is_in_infinite_subgraph(), "must be in infinite subgraph");
5689
// We have "l->_parent == nullptr", which happens only for infinite loops,
5690
// where no parent is attached to the loop. We did not find any irreducible
5691
// loop from this block out to lp. Thus lp only has one entry, and no exit
5692
// (it is infinite and reducible). We can always rewrite an infinite loop
5693
// that is nested inside other loops:
5694
// while(condition) { infinite_loop; }5695
// with an equivalent program where the infinite loop is an outermost loop
5696
// that is not nested in any loop:
5697
// while(condition) { break; } infinite_loop;5698
// Thus, we can understand lp as an outermost loop, and can terminate and
5699
// conclude: this block is in no irreducible loop.
5704
//------------------------------build_loop_early-------------------------------
5705
// Put Data nodes into some loop nest, by setting the _loop_or_ctrl[]->loop mapping.
5706
// First pass computes the earliest controlling node possible. This is the
5707
// controlling input with the deepest dominating depth.
5708
void PhaseIdealLoop::build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) {5709
while (worklist.size() != 0) {5710
// Use local variables nstack_top_n & nstack_top_i to cache values
5712
Node *nstack_top_n = worklist.pop();
5713
uint nstack_top_i = 0;
5714
//while_nstack_nonempty:
5716
// Get parent node and next input's index from stack's top.
5717
Node *n = nstack_top_n;
5718
uint i = nstack_top_i;
5719
uint cnt = n->req(); // Count of inputs
5720
if (i == 0) { // Pre-process the node.5721
if( has_node(n) && // Have either loop or control already?
5722
!has_ctrl(n) ) { // Have loop picked out already?5723
// During "merge_many_backedges" we fold up several nested loops
5724
// into a single loop. This makes the members of the original
5725
// loop bodies pointing to dead loops; they need to move up
5726
// to the new UNION'd larger loop. I set the _head field of these
5727
// dead loops to null and the _parent field points to the owning
5728
// loop. Shades of UNION-FIND algorithm.
5730
while( !(ilt = get_loop(n))->_head ) {5731
// Normally I would use a set_loop here. But in this one special
5732
// case, it is legal (and expected) to change what loop a Node
5734
_loop_or_ctrl.map(n->_idx, (Node*)(ilt->_parent));
5736
// Remove safepoints ONLY if I've already seen I don't need one.
5737
// (the old code here would yank a 2nd safepoint after seeing a
5738
// first one, even though the 1st did not dominate in the loop body
5739
// and thus could be avoided indefinitely)
5740
if( !_verify_only && !_verify_me && ilt->_has_sfpt && n->Opcode() == Op_SafePoint &&
5741
is_deleteable_safept(n)) {5742
Node *in = n->in(TypeFunc::Control);
5743
lazy_replace(n,in); // Pull safepoint now
5744
if (ilt->_safepts != nullptr) {5745
ilt->_safepts->yank(n);
5747
// Carry on with the recursion "as if" we are walking
5748
// only the control input
5749
if( !visited.test_set( in->_idx ) ) {5750
worklist.push(in); // Visit this guy later, using worklist
5752
// Get next node from nstack:
5753
// - skip n's inputs processing by setting i > cnt;
5754
// - we also will not call set_early_ctrl(n) since
5755
// has_node(n) == true (see the condition above).
5762
bool done = true; // Assume all n's inputs will be processed
5764
Node *in = n->in(i);
5766
if (in == nullptr) continue;
5767
if (in->pinned() && !in->is_CFG())
5768
set_ctrl(in, in->in(0));
5769
int is_visited = visited.test_set( in->_idx );
5770
if (!has_node(in)) { // No controlling input yet?5771
assert( !in->is_CFG(), "CFG Node with no controlling input?" );
5772
assert( !is_visited, "visit only once" );
5773
nstack.push(n, i); // Save parent node and next input's index.
5774
nstack_top_n = in; // Process current input now.
5776
done = false; // Not all n's inputs processed.
5777
break; // continue while_nstack_nonempty;
5778
} else if (!is_visited) {5779
// This guy has a location picked out for him, but has not yet
5780
// been visited. Happens to all CFG nodes, for instance.
5781
// Visit him using the worklist instead of recursion, to break
5782
// cycles. Since he has a location already we do not need to
5783
// find his location before proceeding with the current Node.
5784
worklist.push(in); // Visit this guy later, using worklist
5788
// All of n's inputs have been processed, complete post-processing.
5790
// Compute earliest point this Node can go.
5791
// CFG, Phi, pinned nodes already know their controlling input.
5793
// Record earliest legal location
5794
set_early_ctrl(n, false);
5796
if (nstack.is_empty()) {5797
// Finished all nodes on stack.
5798
// Process next node on the worklist.
5801
// Get saved parent node and next input's index.
5802
nstack_top_n = nstack.node();
5803
nstack_top_i = nstack.index();
5810
//------------------------------dom_lca_internal--------------------------------
5812
Node *PhaseIdealLoop::dom_lca_internal( Node *n1, Node *n2 ) const {5813
if( !n1 ) return n2; // Handle null original LCA
5814
assert( n1->is_CFG(), "" );
5815
assert( n2->is_CFG(), "" );
5816
// find LCA of all uses
5817
uint d1 = dom_depth(n1);
5818
uint d2 = dom_depth(n2);
5823
} else if (d1 < d2) {5827
// Here d1 == d2. Due to edits of the dominator-tree, sections
5828
// of the tree might have the same depth. These sections have
5829
// to be searched more carefully.
5831
// Scan up all the n1's with equal depth, looking for n2.
5832
Node *t1 = idom(n1);
5833
while (dom_depth(t1) == d1) {5834
if (t1 == n2) return n2;
5837
// Scan up all the n2's with equal depth, looking for n1.
5838
Node *t2 = idom(n2);
5839
while (dom_depth(t2) == d2) {5840
if (t2 == n1) return n1;
5843
// Move up to a new dominator-depth value as well as up the dom-tree.
5853
//------------------------------compute_idom-----------------------------------
5854
// Locally compute IDOM using dom_lca call. Correct only if the incoming
5855
// IDOMs are correct.
5856
Node *PhaseIdealLoop::compute_idom( Node *region ) const {5857
assert( region->is_Region(), "" );
5858
Node *LCA = nullptr;
5859
for( uint i = 1; i < region->req(); i++ ) {5860
if( region->in(i) != C->top() )
5861
LCA = dom_lca( LCA, region->in(i) );
5866
bool PhaseIdealLoop::verify_dominance(Node* n, Node* use, Node* LCA, Node* early) {5867
bool had_error = false;
5869
if (early != C->root()) {5870
// Make sure that there's a dominance path from LCA to early
5872
while (d != early) {5873
if (d == C->root()) {5874
dump_bad_graph("Bad graph detected in compute_lca_of_uses", n, early, LCA);5875
tty->print_cr("*** Use %d isn't dominated by def %d ***", use->_idx, n->_idx);5887
Node* PhaseIdealLoop::compute_lca_of_uses(Node* n, Node* early, bool verify) {5888
// Compute LCA over list of uses
5889
bool had_error = false;
5890
Node *LCA = nullptr;
5891
for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && LCA != early; i++) {5892
Node* c = n->fast_out(i);
5893
if (_loop_or_ctrl[c->_idx] == nullptr)
5894
continue; // Skip the occasional dead node
5895
if( c->is_Phi() ) { // For Phis, we must land above on the path5896
for( uint j=1; j<c->req(); j++ ) {// For all inputs5897
if( c->in(j) == n ) { // Found matching input?5898
Node *use = c->in(0)->in(j);
5899
if (_verify_only && use->is_top()) continue;
5900
LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
5901
if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error;
5905
// For CFG data-users, use is in the block just prior
5906
Node *use = has_ctrl(c) ? get_ctrl(c) : c->in(0);
5907
LCA = dom_lca_for_get_late_ctrl( LCA, use, n );
5908
if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error;
5911
assert(!had_error, "bad dominance");
5915
// Check the shape of the graph at the loop entry. In some cases,
5916
// the shape of the graph does not match the shape outlined below.
5917
// That is caused by the Opaque1 node "protecting" the shape of
5918
// the graph being removed by, for example, the IGVN performed
5919
// in PhaseIdealLoop::build_and_optimize().
5921
// After the Opaque1 node has been removed, optimizations (e.g., split-if,
5922
// loop unswitching, and IGVN, or a combination of them) can freely change
5923
// the graph's shape. As a result, the graph shape outlined below cannot
5924
// be guaranteed anymore.
5925
Node* CountedLoopNode::is_canonical_loop_entry() {5926
if (!is_main_loop() && !is_post_loop()) {5929
Node* ctrl = skip_assertion_predicates_with_halt();
5931
if (ctrl == nullptr || (!ctrl->is_IfTrue() && !ctrl->is_IfFalse())) {5934
Node* iffm = ctrl->in(0);
5935
if (iffm == nullptr || iffm->Opcode() != Op_If) {5938
Node* bolzm = iffm->in(1);
5939
if (bolzm == nullptr || !bolzm->is_Bool()) {5942
Node* cmpzm = bolzm->in(1);
5943
if (cmpzm == nullptr || !cmpzm->is_Cmp()) {5947
uint input = is_main_loop() ? 2 : 1;
5948
if (input >= cmpzm->req() || cmpzm->in(input) == nullptr) {5951
bool res = cmpzm->in(input)->Opcode() == Op_OpaqueZeroTripGuard;
5953
bool found_opaque = false;
5954
for (uint i = 1; i < cmpzm->req(); i++) {5955
Node* opnd = cmpzm->in(i);
5956
if (opnd && opnd->is_Opaque1()) {5957
found_opaque = true;
5961
assert(found_opaque == res, "wrong pattern");
5963
return res ? cmpzm->in(input) : nullptr;
5966
// Find pre loop end from main loop. Returns nullptr if none.
5967
CountedLoopEndNode* CountedLoopNode::find_pre_loop_end() {5968
assert(is_main_loop(), "Can only find pre-loop from main-loop");
5969
// The loop cannot be optimized if the graph shape at the loop entry is
5971
if (is_canonical_loop_entry() == nullptr) {5975
Node* p_f = skip_assertion_predicates_with_halt()->in(0)->in(0);
5976
if (!p_f->is_IfFalse() || !p_f->in(0)->is_CountedLoopEnd()) {5979
CountedLoopEndNode* pre_end = p_f->in(0)->as_CountedLoopEnd();
5980
CountedLoopNode* loop_node = pre_end->loopnode();
5981
if (loop_node == nullptr || !loop_node->is_pre_loop()) {5987
//------------------------------get_late_ctrl----------------------------------
5988
// Compute latest legal control.
5989
Node *PhaseIdealLoop::get_late_ctrl( Node *n, Node *early ) {5990
assert(early != nullptr, "early control should not be null");
5992
Node* LCA = compute_lca_of_uses(n, early);
5994
if (LCA == C->root() && LCA != early) {5995
// def doesn't dominate uses so print some useful debugging output
5996
compute_lca_of_uses(n, early, true);
6000
if (n->is_Load() && LCA != early) {6001
LCA = get_late_ctrl_with_anti_dep(n->as_Load(), early, LCA);
6004
assert(LCA == find_non_split_ctrl(LCA), "unexpected late control");
6008
// if this is a load, check for anti-dependent stores
6009
// We use a conservative algorithm to identify potential interfering
6010
// instructions and for rescheduling the load. The users of the memory
6011
// input of this load are examined. Any use which is not a load and is
6012
// dominated by early is considered a potentially interfering store.
6013
// This can produce false positives.
6014
Node* PhaseIdealLoop::get_late_ctrl_with_anti_dep(LoadNode* n, Node* early, Node* LCA) {6015
int load_alias_idx = C->get_alias_index(n->adr_type());
6016
if (C->alias_type(load_alias_idx)->is_rewritable()) {6017
Unique_Node_List worklist;
6019
Node* mem = n->in(MemNode::Memory);
6020
for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {6021
Node* s = mem->fast_out(i);
6024
for (uint i = 0; i < worklist.size() && LCA != early; i++) {6025
Node* s = worklist.at(i);
6026
if (s->is_Load() || s->Opcode() == Op_SafePoint ||
6027
(s->is_CallStaticJava() && s->as_CallStaticJava()->uncommon_trap_request() != 0) ||
6030
} else if (s->is_MergeMem()) {6031
for (DUIterator_Fast imax, i = s->fast_outs(imax); i < imax; i++) {6032
Node* s1 = s->fast_out(i);
6036
Node* sctrl = has_ctrl(s) ? get_ctrl(s) : s->in(0);
6037
assert(sctrl != nullptr || !s->is_reachable_from_root(), "must have control");
6038
if (sctrl != nullptr && !sctrl->is_top() && is_dominator(early, sctrl)) {6039
const TypePtr* adr_type = s->adr_type();
6040
if (s->is_ArrayCopy()) {6041
// Copy to known instance needs destination type to test for aliasing
6042
const TypePtr* dest_type = s->as_ArrayCopy()->_dest_type;
6043
if (dest_type != TypeOopPtr::BOTTOM) {6044
adr_type = dest_type;
6047
if (C->can_alias(adr_type, load_alias_idx)) {6048
LCA = dom_lca_for_get_late_ctrl(LCA, sctrl, n);
6049
} else if (s->is_CFG() && s->is_Multi()) {6050
// Look for the memory use of s (that is the use of its memory projection)
6051
for (DUIterator_Fast imax, i = s->fast_outs(imax); i < imax; i++) {6052
Node* s1 = s->fast_out(i);
6053
assert(s1->is_Proj(), "projection expected");
6054
if (_igvn.type(s1) == Type::MEMORY) {6055
for (DUIterator_Fast jmax, j = s1->fast_outs(jmax); j < jmax; j++) {6056
Node* s2 = s1->fast_out(j);
6065
// For Phis only consider Region's inputs that were reached by following the memory edges
6067
for (uint i = 0; i < worklist.size(); i++) {6068
Node* s = worklist.at(i);
6069
if (s->is_Phi() && C->can_alias(s->adr_type(), load_alias_idx)) {6071
for (uint j = 1; j < s->req(); j++) {6072
Node* in = s->in(j);
6073
Node* r_in = r->in(j);
6074
// We can't reach any node from a Phi because we don't enqueue Phi's uses above
6075
if (((worklist.member(in) && !in->is_Phi()) || in == mem) && is_dominator(early, r_in)) {6076
LCA = dom_lca_for_get_late_ctrl(LCA, r_in, n);
6086
// Is CFG node 'dominator' dominating node 'n'?
6087
bool PhaseIdealLoop::is_dominator(Node* dominator, Node* n) {6088
if (dominator == n) {6091
assert(dominator->is_CFG() && n->is_CFG(), "must have CFG nodes");
6092
uint dd = dom_depth(dominator);
6093
while (dom_depth(n) >= dd) {6094
if (n == dominator) {6102
// Is CFG node 'dominator' strictly dominating node 'n'?
6103
bool PhaseIdealLoop::is_strict_dominator(Node* dominator, Node* n) {6104
return dominator != n && is_dominator(dominator, n);
6107
//------------------------------dom_lca_for_get_late_ctrl_internal-------------
6108
// Pair-wise LCA with tags.
6109
// Tag each index with the node 'tag' currently being processed
6110
// before advancing up the dominator chain using idom().
6111
// Later calls that find a match to 'tag' know that this path has already
6112
// been considered in the current LCA (which is input 'n1' by convention).
6113
// Since get_late_ctrl() is only called once for each node, the tag array
6114
// does not need to be cleared between calls to get_late_ctrl().
6115
// Algorithm trades a larger constant factor for better asymptotic behavior
6117
Node *PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal(Node *n1, Node *n2, Node *tag_node) {6118
uint d1 = dom_depth(n1);
6119
uint d2 = dom_depth(n2);
6120
jlong tag = tag_node->_idx | (((jlong)_dom_lca_tags_round) << 32);
6124
// current lca is deeper than n2
6125
_dom_lca_tags.at_put_grow(n1->_idx, tag);
6128
} else if (d1 < d2) {6129
// n2 is deeper than current lca
6130
jlong memo = _dom_lca_tags.at_grow(n2->_idx, 0);
6132
return n1; // Return the current LCA
6134
_dom_lca_tags.at_put_grow(n2->_idx, tag);
6138
// Here d1 == d2. Due to edits of the dominator-tree, sections
6139
// of the tree might have the same depth. These sections have
6140
// to be searched more carefully.
6142
// Scan up all the n1's with equal depth, looking for n2.
6143
_dom_lca_tags.at_put_grow(n1->_idx, tag);
6144
Node *t1 = idom(n1);
6145
while (dom_depth(t1) == d1) {6146
if (t1 == n2) return n2;
6147
_dom_lca_tags.at_put_grow(t1->_idx, tag);
6150
// Scan up all the n2's with equal depth, looking for n1.
6151
_dom_lca_tags.at_put_grow(n2->_idx, tag);
6152
Node *t2 = idom(n2);
6153
while (dom_depth(t2) == d2) {6154
if (t2 == n1) return n1;
6155
_dom_lca_tags.at_put_grow(t2->_idx, tag);
6158
// Move up to a new dominator-depth value as well as up the dom-tree.
6168
//------------------------------init_dom_lca_tags------------------------------
6169
// Tag could be a node's integer index, 32bits instead of 64bits in some cases
6170
// Intended use does not involve any growth for the array, so it could
6172
void PhaseIdealLoop::init_dom_lca_tags() {6173
uint limit = C->unique() + 1;
6174
_dom_lca_tags.at_grow(limit, 0);
6175
_dom_lca_tags_round = 0;
6177
for (uint i = 0; i < limit; ++i) {6178
assert(_dom_lca_tags.at(i) == 0, "Must be distinct from each node pointer");
6183
//------------------------------build_loop_late--------------------------------
6184
// Put Data nodes into some loop nest, by setting the _loop_or_ctrl[]->loop mapping.
6185
// Second pass finds latest legal placement, and ideal loop placement.
6186
void PhaseIdealLoop::build_loop_late( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) {6187
while (worklist.size() != 0) {6188
Node *n = worklist.pop();
6190
if (visited.test_set(n->_idx)) continue;
6191
uint cnt = n->outcnt();
6194
assert(_loop_or_ctrl[n->_idx], "no dead nodes");
6195
// Visit all children
6197
Node* use = n->raw_out(i);
6199
// Check for dead uses. Aggressively prune such junk. It might be
6200
// dead in the global sense, but still have local uses so I cannot
6201
// easily call 'remove_dead_node'.
6202
if (_loop_or_ctrl[use->_idx] != nullptr || use->is_top()) { // Not dead?6203
// Due to cycles, we might not hit the same fixed point in the verify
6204
// pass as we do in the regular pass. Instead, visit such phis as
6205
// simple uses of the loop head.
6206
if( use->in(0) && (use->is_CFG() || use->is_Phi()) ) {6207
if( !visited.test(use->_idx) )
6209
} else if( !visited.test_set(use->_idx) ) {6210
nstack.push(n, i); // Save parent and next use's index.
6211
n = use; // Process all children of current use.
6212
cnt = use->outcnt();
6216
// Do not visit around the backedge of loops via data edges.
6217
// push dead code onto a worklist
6218
_deadlist.push(use);
6221
// All of n's children have been processed, complete post-processing.
6222
build_loop_late_post(n);
6223
if (C->failing()) { return; }6224
if (nstack.is_empty()) {6225
// Finished all nodes on stack.
6226
// Process next node on the worklist.
6229
// Get saved parent node and next use's index. Visit the rest of uses.
6239
// Verify that no data node is scheduled in the outer loop of a strip
6241
void PhaseIdealLoop::verify_strip_mined_scheduling(Node *n, Node* least) {6243
if (get_loop(least)->_nest == 0) {6246
IdealLoopTree* loop = get_loop(least);
6247
Node* head = loop->_head;
6248
if (head->is_OuterStripMinedLoop() &&
6249
// Verification can't be applied to fully built strip mined loops
6250
head->as_Loop()->outer_loop_end()->in(1)->find_int_con(-1) == 0) {6251
Node* sfpt = head->as_Loop()->outer_safepoint();
6253
Unique_Node_List wq;
6255
for (uint i = 0; i < wq.size(); i++) {6257
for (uint i = 1; i < m->req(); i++) {6258
Node* nn = m->in(i);
6262
if (nn != nullptr && has_ctrl(nn) && get_loop(get_ctrl(nn)) == loop) {6267
ShouldNotReachHere();
6273
//------------------------------build_loop_late_post---------------------------
6274
// Put Data nodes into some loop nest, by setting the _loop_or_ctrl[]->loop mapping.
6275
// Second pass finds latest legal placement, and ideal loop placement.
6276
void PhaseIdealLoop::build_loop_late_post(Node *n) {6277
build_loop_late_post_work(n, true);
6280
void PhaseIdealLoop::build_loop_late_post_work(Node *n, bool pinned) {6282
if (n->req() == 2 && (n->Opcode() == Op_ConvI2L || n->Opcode() == Op_CastII) && !C->major_progress() && !_verify_only) {6283
_igvn._worklist.push(n); // Maybe we'll normalize it, if no more loops.
6287
if (_verify_only && !n->is_CFG()) {6288
// Check def-use domination.
6289
compute_lca_of_uses(n, get_ctrl(n), true /* verify */);
6293
// CFG and pinned nodes already handled
6295
if( n->in(0)->is_top() ) return; // Dead?
6297
// We'd like +VerifyLoopOptimizations to not believe that Mod's/Loads
6298
// _must_ be pinned (they have to observe their control edge of course).
6299
// Unlike Stores (which modify an unallocable resource, the memory
6300
// state), Mods/Loads can float around. So free them up.
6301
switch( n->Opcode() ) {6308
case Op_LoadB: // Same with Loads; they can sink
6309
case Op_LoadUB: // during loop optimizations.
6321
case Op_LoadD_unaligned:
6322
case Op_LoadL_unaligned:
6323
case Op_StrComp: // Does a bunch of load-like effects
6326
case Op_StrIndexOfChar:
6328
case Op_VectorizedHashCode:
6329
case Op_CountPositives:
6332
if (n->is_CMove() || n->is_ConstraintCast()) {6336
IdealLoopTree *chosen_loop = get_loop(n->is_CFG() ? n : get_ctrl(n));
6337
if( !chosen_loop->_child ) // Inner loop?
6338
chosen_loop->_body.push(n); // Collect inner loops
6341
} else { // No slot zero6342
if( n->is_CFG() ) { // CFG with no slot 0 is dead6343
_loop_or_ctrl.map(n->_idx,nullptr); // No block setting, it's globally dead
6346
assert(!n->is_CFG() || n->outcnt() == 0, "");
6349
// Do I have a "safe range" I can select over?
6350
Node *early = get_ctrl(n);// Early location already computed
6352
// Compute latest point this Node can go
6353
Node *LCA = get_late_ctrl( n, early );
6354
// LCA is null due to uses being dead
6355
if( LCA == nullptr ) {6357
for (DUIterator i1 = n->outs(); n->has_out(i1); i1++) {6358
assert(_loop_or_ctrl[n->out(i1)->_idx] == nullptr, "all uses must also be dead");
6361
_loop_or_ctrl.map(n->_idx, nullptr); // This node is useless
6365
assert(LCA != nullptr && !LCA->is_top(), "no dead nodes");
6367
Node *legal = LCA; // Walk 'legal' up the IDOM chain
6368
Node *least = legal; // Best legal position so far
6369
while( early != legal ) { // While not at earliest legal6370
if (legal->is_Start() && !early->is_Root()) {6372
// Bad graph. Print idom path and fail.
6373
dump_bad_graph("Bad graph detected in build_loop_late", n, early, LCA);6374
assert(false, "Bad graph detected in build_loop_late");
6376
C->record_method_not_compilable("Bad graph detected in build_loop_late");6379
// Find least loop nesting depth
6380
legal = idom(legal); // Bump up the IDOM tree
6381
// Check for lower nesting depth
6382
if( get_loop(legal)->_nest < get_loop(least)->_nest )
6385
assert(early == legal || legal != C->root(), "bad dominance of inputs");
6387
if (least != early) {6388
// Move the node above predicates as far up as possible so a
6389
// following pass of Loop Predication doesn't hoist a predicate
6390
// that depends on it above that node.
6391
PredicateEntryIterator predicate_iterator(least);
6392
while (predicate_iterator.has_next()) {6393
Node* next_predicate_entry = predicate_iterator.next_entry();
6394
if (is_strict_dominator(next_predicate_entry, early)) {6397
least = next_predicate_entry;
6400
// Try not to place code on a loop entry projection
6401
// which can inhibit range check elimination.
6402
if (least != early && !BarrierSet::barrier_set()->barrier_set_c2()->is_gc_specific_loop_opts_pass(_mode)) {6403
Node* ctrl_out = least->unique_ctrl_out_or_null();
6404
if (ctrl_out != nullptr && ctrl_out->is_Loop() &&
6405
least == ctrl_out->in(LoopNode::EntryControl) &&
6406
(ctrl_out->is_CountedLoop() || ctrl_out->is_OuterStripMinedLoop())) {6407
Node* least_dom = idom(least);
6408
if (get_loop(least_dom)->is_member(get_loop(least))) {6413
// Don't extend live ranges of raw oops
6414
if (least != early && n->is_ConstraintCast() && n->in(1)->bottom_type()->isa_rawptr() &&
6415
!n->bottom_type()->isa_rawptr()) {6420
// Broken part of VerifyLoopOptimizations (F)
6422
// _verify_me->get_ctrl_no_update(n) seems to return wrong result
6424
// If verifying, verify that 'verify_me' has a legal location
6425
// and choose it as our location.
6427
Node *v_ctrl = _verify_me->get_ctrl_no_update(n);
6429
while( early != legal ) { // While not at earliest legal6430
if( legal == v_ctrl ) break; // Check for prior good location
6431
legal = idom(legal) ;// Bump up the IDOM tree
6433
// Check for prior good location
6434
if( legal == v_ctrl ) least = legal; // Keep prior if found
6439
// Assign discovered "here or above" point
6440
least = find_non_split_ctrl(least);
6441
verify_strip_mined_scheduling(n, least);
6444
// Collect inner loop bodies
6445
IdealLoopTree *chosen_loop = get_loop(least);
6446
if( !chosen_loop->_child ) // Inner loop?
6447
chosen_loop->_body.push(n);// Collect inner loops
6449
if (!_verify_only && n->Opcode() == Op_OpaqueZeroTripGuard) {6450
_zero_trip_guard_opaque_nodes.push(n);
6456
void PhaseIdealLoop::dump_bad_graph(const char* msg, Node* n, Node* early, Node* LCA) {6457
tty->print_cr("%s", msg);6458
tty->print("n: "); n->dump();6459
tty->print("early(n): "); early->dump();6460
if (n->in(0) != nullptr && !n->in(0)->is_top() &&
6461
n->in(0) != early && !n->in(0)->is_Root()) {6462
tty->print("n->in(0): "); n->in(0)->dump();6464
for (uint i = 1; i < n->req(); i++) {6465
Node* in1 = n->in(i);
6466
if (in1 != nullptr && in1 != n && !in1->is_top()) {6467
tty->print("n->in(%d): ", i); in1->dump();6468
Node* in1_early = get_ctrl(in1);
6469
tty->print("early(n->in(%d)): ", i); in1_early->dump();6470
if (in1->in(0) != nullptr && !in1->in(0)->is_top() &&
6471
in1->in(0) != in1_early && !in1->in(0)->is_Root()) {6472
tty->print("n->in(%d)->in(0): ", i); in1->in(0)->dump();6474
for (uint j = 1; j < in1->req(); j++) {6475
Node* in2 = in1->in(j);
6476
if (in2 != nullptr && in2 != n && in2 != in1 && !in2->is_top()) {6477
tty->print("n->in(%d)->in(%d): ", i, j); in2->dump();6478
Node* in2_early = get_ctrl(in2);
6479
tty->print("early(n->in(%d)->in(%d)): ", i, j); in2_early->dump();6480
if (in2->in(0) != nullptr && !in2->in(0)->is_top() &&
6481
in2->in(0) != in2_early && !in2->in(0)->is_Root()) {6482
tty->print("n->in(%d)->in(%d)->in(0): ", i, j); in2->in(0)->dump();6489
tty->print("LCA(n): "); LCA->dump();6490
for (uint i = 0; i < n->outcnt(); i++) {6491
Node* u1 = n->raw_out(i);
6494
tty->print("n->out(%d): ", i); u1->dump();6496
for (uint j = 0; j < u1->outcnt(); j++) {6497
Node* u2 = u1->raw_out(j);
6498
if (u2 != u1 && u2 != n && u2->is_CFG()) {6499
tty->print("n->out(%d)->out(%d): ", i, j); u2->dump();6503
Node* u1_later = get_ctrl(u1);
6504
tty->print("later(n->out(%d)): ", i); u1_later->dump();6505
if (u1->in(0) != nullptr && !u1->in(0)->is_top() &&
6506
u1->in(0) != u1_later && !u1->in(0)->is_Root()) {6507
tty->print("n->out(%d)->in(0): ", i); u1->in(0)->dump();6509
for (uint j = 0; j < u1->outcnt(); j++) {6510
Node* u2 = u1->raw_out(j);
6511
if (u2 == n || u2 == u1)
6513
tty->print("n->out(%d)->out(%d): ", i, j); u2->dump();6514
if (!u2->is_CFG()) {6515
Node* u2_later = get_ctrl(u2);
6516
tty->print("later(n->out(%d)->out(%d)): ", i, j); u2_later->dump();6517
if (u2->in(0) != nullptr && !u2->in(0)->is_top() &&
6518
u2->in(0) != u2_later && !u2->in(0)->is_Root()) {6519
tty->print("n->out(%d)->in(0): ", i); u2->in(0)->dump();6525
dump_idoms(early, LCA);
6529
// Class to compute the real LCA given an early node and a wrong LCA in a bad graph.
6531
const PhaseIdealLoop* _phase;
6535
int _wrong_lca_index;
6537
// Given idom chains of early and wrong LCA: Walk through idoms starting at StartNode and find the first node which
6538
// is different: Return the previously visited node which must be the real LCA.
6539
// The node lists also contain _early and _wrong_lca, respectively.
6540
Node* find_real_lca(Unique_Node_List& early_with_idoms, Unique_Node_List& wrong_lca_with_idoms) {6541
int early_index = early_with_idoms.size() - 1;
6542
int wrong_lca_index = wrong_lca_with_idoms.size() - 1;
6543
bool found_difference = false;
6545
if (early_with_idoms[early_index] != wrong_lca_with_idoms[wrong_lca_index]) {6546
// First time early and wrong LCA idoms differ. Real LCA must be at the previous index.
6547
found_difference = true;
6552
} while (wrong_lca_index >= 0);
6554
assert(early_index >= 0, "must always find an LCA - cannot be early");
6555
_early_index = early_index;
6556
_wrong_lca_index = wrong_lca_index;
6557
Node* real_lca = early_with_idoms[_early_index + 1]; // Plus one to skip _early.
6558
assert(found_difference || real_lca == _wrong_lca, "wrong LCA dominates early and is therefore the real LCA");
6562
void dump(Node* real_lca) {6564
tty->print_cr("idoms of early \"%d %s\":", _early->_idx, _early->Name());6565
_phase->dump_idom(_early, _early_index + 1);
6568
tty->print_cr("idoms of (wrong) LCA \"%d %s\":", _wrong_lca->_idx, _wrong_lca->Name());6569
_phase->dump_idom(_wrong_lca, _wrong_lca_index + 1);
6572
tty->print("Real LCA of early \"%d %s\" (idom[%d]) and wrong LCA \"%d %s\"",6573
_early->_idx, _early->Name(), _early_index, _wrong_lca->_idx, _wrong_lca->Name());
6574
if (_wrong_lca_index >= 0) {6575
tty->print(" (idom[%d])", _wrong_lca_index);6582
RealLCA(const PhaseIdealLoop* phase, Node* early, Node* wrong_lca)
6583
: _phase(phase), _early(early), _wrong_lca(wrong_lca), _early_index(0), _wrong_lca_index(0) {6584
assert(!wrong_lca->is_Start(), "StartNode is always a common dominator");
6587
void compute_and_dump() {6589
Unique_Node_List early_with_idoms;
6590
Unique_Node_List wrong_lca_with_idoms;
6591
early_with_idoms.push(_early);
6592
wrong_lca_with_idoms.push(_wrong_lca);
6593
_phase->get_idoms(_early, 10000, early_with_idoms);
6594
_phase->get_idoms(_wrong_lca, 10000, wrong_lca_with_idoms);
6595
Node* real_lca = find_real_lca(early_with_idoms, wrong_lca_with_idoms);
6600
// Dump the idom chain of early, of the wrong LCA and dump the real LCA of early and wrong LCA.
6601
void PhaseIdealLoop::dump_idoms(Node* early, Node* wrong_lca) {6602
assert(!is_dominator(early, wrong_lca), "sanity check that early does not dominate wrong lca");
6603
assert(!has_ctrl(early) && !has_ctrl(wrong_lca), "sanity check, no data nodes");
6605
RealLCA real_lca(this, early, wrong_lca);
6606
real_lca.compute_and_dump();
6611
//------------------------------dump-------------------------------------------
6612
void PhaseIdealLoop::dump() const {6614
Node_Stack stack(C->live_nodes() >> 2);
6617
visited.set(C->top()->_idx);
6618
rpo(C->root(), stack, visited, rpo_list);
6619
// Dump root loop indexed by last element in PO order
6620
dump(_ltree_root, rpo_list.size(), rpo_list);
6623
void PhaseIdealLoop::dump(IdealLoopTree* loop, uint idx, Node_List &rpo_list) const {6626
// Now scan for CFG nodes in the same loop
6627
for (uint j = idx; j > 0; j--) {6628
Node* n = rpo_list[j-1];
6629
if (!_loop_or_ctrl[n->_idx]) // Skip dead nodes
6632
if (get_loop(n) != loop) { // Wrong loop nest6633
if (get_loop(n)->_head == n && // Found nested loop?
6634
get_loop(n)->_parent == loop)
6635
dump(get_loop(n), rpo_list.size(), rpo_list); // Print it nested-ly
6639
// Dump controlling node
6640
tty->sp(2 * loop->_nest);
6642
if (n == C->root()) {6645
Node* cached_idom = idom_no_update(n);
6646
Node* computed_idom = n->in(0);
6647
if (n->is_Region()) {6648
computed_idom = compute_idom(n);
6649
// computed_idom() will return n->in(0) when idom(n) is an IfNode (or
6650
// any MultiBranch ctrl node), so apply a similar transform to
6651
// the cached idom returned from idom_no_update.
6652
cached_idom = find_non_split_ctrl(cached_idom);
6654
tty->print(" ID:%d", computed_idom->_idx);6656
if (cached_idom != computed_idom) {6657
tty->print_cr("*** BROKEN IDOM! Computed as: %d, cached as: %d",6658
computed_idom->_idx, cached_idom->_idx);
6661
// Dump nodes it controls
6662
for (uint k = 0; k < _loop_or_ctrl.max(); k++) {6663
// (k < C->unique() && get_ctrl(find(k)) == n)
6664
if (k < C->unique() && _loop_or_ctrl[k] == (Node*)((intptr_t)n + 1)) {6665
Node* m = C->root()->find(k);
6666
if (m && m->outcnt() > 0) {6667
if (!(has_ctrl(m) && get_ctrl_no_update(m) == n)) {6668
tty->print_cr("*** BROKEN CTRL ACCESSOR! _loop_or_ctrl[k] is %p, ctrl is %p",6669
_loop_or_ctrl[k], has_ctrl(m) ? get_ctrl_no_update(m) : nullptr);
6671
tty->sp(2 * loop->_nest + 1);
6679
void PhaseIdealLoop::dump_idom(Node* n, const uint count) const {6681
tty->print_cr("No idom for data nodes");6684
Unique_Node_List idoms;
6685
get_idoms(n, count, idoms);
6686
dump_idoms_in_reverse(n, idoms);
6690
void PhaseIdealLoop::get_idoms(Node* n, const uint count, Unique_Node_List& idoms) const {6692
for (uint i = 0; !next->is_Start() && i < count; i++) {6694
assert(!idoms.member(next), "duplicated idom is not possible");
6699
void PhaseIdealLoop::dump_idoms_in_reverse(const Node* n, const Node_List& idom_list) const {6702
uint node_index_padding_width = static_cast<int>(log10(static_cast<double>(C->unique()))) + 1;
6703
for (int i = idom_list.size() - 1; i >= 0; i--) {6704
if (i == 9 || i == 99) {6707
next = idom_list[i];
6708
tty->print_cr("idom[%d]:%*c%*d %s", i, padding, ' ', node_index_padding_width, next->_idx, next->Name());6710
tty->print_cr("n: %*c%*d %s", padding, ' ', node_index_padding_width, n->_idx, n->Name());6712
#endif // NOT PRODUCT
6714
// Collect a R-P-O for the whole CFG.
6715
// Result list is in post-order (scan backwards for RPO)
6716
void PhaseIdealLoop::rpo(Node* start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list) const {6718
visited.set(start->_idx);
6720
while (stk.is_nonempty()) {6721
Node* m = stk.node();
6722
uint idx = stk.index();
6723
if (idx < m->outcnt()) {6724
stk.set_index(idx + 1);
6725
Node* n = m->raw_out(idx);
6726
if (n->is_CFG() && !visited.test_set(n->_idx)) {6737
//=============================================================================
6738
//------------------------------LoopTreeIterator-------------------------------
6740
// Advance to next loop tree using a preorder, left-to-right traversal.
6741
void LoopTreeIterator::next() {6742
assert(!done(), "must not be done.");
6743
if (_curnt->_child != nullptr) {6744
_curnt = _curnt->_child;
6745
} else if (_curnt->_next != nullptr) {6746
_curnt = _curnt->_next;
6748
while (_curnt != _root && _curnt->_next == nullptr) {6749
_curnt = _curnt->_parent;
6751
if (_curnt == _root) {6753
assert(done(), "must be done.");
6755
assert(_curnt->_next != nullptr, "must be more to do");
6756
_curnt = _curnt->_next;