1 /*
2 * Copyright (c) 1998, 2025, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "ci/ciMethodData.hpp"
26 #include "classfile/vmSymbols.hpp"
27 #include "compiler/compileLog.hpp"
28 #include "interpreter/linkResolver.hpp"
29 #include "jvm_io.h"
30 #include "memory/resourceArea.hpp"
31 #include "memory/universe.hpp"
32 #include "oops/oop.inline.hpp"
33 #include "opto/addnode.hpp"
34 #include "opto/castnode.hpp"
35 #include "opto/convertnode.hpp"
36 #include "opto/divnode.hpp"
37 #include "opto/idealGraphPrinter.hpp"
38 #include "opto/matcher.hpp"
39 #include "opto/memnode.hpp"
40 #include "opto/mulnode.hpp"
41 #include "opto/opaquenode.hpp"
42 #include "opto/parse.hpp"
43 #include "opto/runtime.hpp"
44 #include "opto/subtypenode.hpp"
45 #include "runtime/deoptimization.hpp"
46 #include "runtime/sharedRuntime.hpp"
47
48 #ifndef PRODUCT
49 extern uint explicit_null_checks_inserted,
50 explicit_null_checks_elided;
51 #endif
52
53 //---------------------------------array_load----------------------------------
54 void Parse::array_load(BasicType bt) {
55 const Type* elemtype = Type::TOP;
56 bool big_val = bt == T_DOUBLE || bt == T_LONG;
57 Node* adr = array_addressing(bt, 0, elemtype);
58 if (stopped()) return; // guaranteed null or range check
59
60 pop(); // index (already used)
61 Node* array = pop(); // the array itself
62
63 if (elemtype == TypeInt::BOOL) {
64 bt = T_BOOLEAN;
65 }
66 const TypeAryPtr* adr_type = TypeAryPtr::get_array_body_type(bt);
67
68 Node* ld = access_load_at(array, adr, adr_type, elemtype, bt,
69 IN_HEAP | IS_ARRAY | C2_CONTROL_DEPENDENT_LOAD);
70 if (big_val) {
71 push_pair(ld);
72 } else {
73 push(ld);
74 }
75 }
76
77
78 //--------------------------------array_store----------------------------------
79 void Parse::array_store(BasicType bt) {
80 const Type* elemtype = Type::TOP;
81 bool big_val = bt == T_DOUBLE || bt == T_LONG;
82 Node* adr = array_addressing(bt, big_val ? 2 : 1, elemtype);
83 if (stopped()) return; // guaranteed null or range check
84 if (bt == T_OBJECT) {
85 array_store_check();
86 if (stopped()) {
87 return;
88 }
89 }
90 Node* val; // Oop to store
91 if (big_val) {
92 val = pop_pair();
93 } else {
94 val = pop();
95 }
96 pop(); // index (already used)
97 Node* array = pop(); // the array itself
98
99 if (elemtype == TypeInt::BOOL) {
100 bt = T_BOOLEAN;
101 }
102 const TypeAryPtr* adr_type = TypeAryPtr::get_array_body_type(bt);
103
104 access_store_at(array, adr, adr_type, val, elemtype, bt, MO_UNORDERED | IN_HEAP | IS_ARRAY);
105 }
106
107
108 //------------------------------array_addressing-------------------------------
109 // Pull array and index from the stack. Compute pointer-to-element.
110 Node* Parse::array_addressing(BasicType type, int vals, const Type*& elemtype) {
111 Node *idx = peek(0+vals); // Get from stack without popping
112 Node *ary = peek(1+vals); // in case of exception
113
114 // Null check the array base, with correct stack contents
115 ary = null_check(ary, T_ARRAY);
116 // Compile-time detect of null-exception?
117 if (stopped()) return top();
118
119 const TypeAryPtr* arytype = _gvn.type(ary)->is_aryptr();
120 const TypeInt* sizetype = arytype->size();
121 elemtype = arytype->elem();
122
123 if (UseUniqueSubclasses) {
124 const Type* el = elemtype->make_ptr();
125 if (el && el->isa_instptr()) {
126 const TypeInstPtr* toop = el->is_instptr();
127 if (toop->instance_klass()->unique_concrete_subklass()) {
128 // If we load from "AbstractClass[]" we must see "ConcreteSubClass".
129 const Type* subklass = Type::get_const_type(toop->instance_klass());
130 elemtype = subklass->join_speculative(el);
131 }
132 }
133 }
134
135 // Check for big class initializers with all constant offsets
136 // feeding into a known-size array.
137 const TypeInt* idxtype = _gvn.type(idx)->is_int();
138 // See if the highest idx value is less than the lowest array bound,
139 // and if the idx value cannot be negative:
140 bool need_range_check = true;
141 if (idxtype->_hi < sizetype->_lo && idxtype->_lo >= 0) {
142 need_range_check = false;
143 if (C->log() != nullptr) C->log()->elem("observe that='!need_range_check'");
144 }
145
146 if (!arytype->is_loaded()) {
147 // Only fails for some -Xcomp runs
148 // The class is unloaded. We have to run this bytecode in the interpreter.
149 ciKlass* klass = arytype->unloaded_klass();
150
151 uncommon_trap(Deoptimization::Reason_unloaded,
152 Deoptimization::Action_reinterpret,
153 klass, "!loaded array");
154 return top();
155 }
156
157 // Do the range check
158 if (need_range_check) {
159 Node* tst;
160 if (sizetype->_hi <= 0) {
161 // The greatest array bound is negative, so we can conclude that we're
162 // compiling unreachable code, but the unsigned compare trick used below
163 // only works with non-negative lengths. Instead, hack "tst" to be zero so
164 // the uncommon_trap path will always be taken.
165 tst = _gvn.intcon(0);
166 } else {
167 // Range is constant in array-oop, so we can use the original state of mem
168 Node* len = load_array_length(ary);
169
170 // Test length vs index (standard trick using unsigned compare)
171 Node* chk = _gvn.transform( new CmpUNode(idx, len) );
172 BoolTest::mask btest = BoolTest::lt;
173 tst = _gvn.transform( new BoolNode(chk, btest) );
174 }
175 RangeCheckNode* rc = new RangeCheckNode(control(), tst, PROB_MAX, COUNT_UNKNOWN);
176 _gvn.set_type(rc, rc->Value(&_gvn));
177 if (!tst->is_Con()) {
178 record_for_igvn(rc);
179 }
180 set_control(_gvn.transform(new IfTrueNode(rc)));
181 // Branch to failure if out of bounds
182 {
183 PreserveJVMState pjvms(this);
184 set_control(_gvn.transform(new IfFalseNode(rc)));
185 if (C->allow_range_check_smearing()) {
186 // Do not use builtin_throw, since range checks are sometimes
187 // made more stringent by an optimistic transformation.
188 // This creates "tentative" range checks at this point,
189 // which are not guaranteed to throw exceptions.
190 // See IfNode::Ideal, is_range_check, adjust_check.
191 uncommon_trap(Deoptimization::Reason_range_check,
192 Deoptimization::Action_make_not_entrant,
193 nullptr, "range_check");
194 } else {
195 // If we have already recompiled with the range-check-widening
196 // heroic optimization turned off, then we must really be throwing
197 // range check exceptions.
198 builtin_throw(Deoptimization::Reason_range_check);
199 }
200 }
201 }
202 // Check for always knowing you are throwing a range-check exception
203 if (stopped()) return top();
204
205 // Make array address computation control dependent to prevent it
206 // from floating above the range check during loop optimizations.
207 Node* ptr = array_element_address(ary, idx, type, sizetype, control());
208 assert(ptr != top(), "top should go hand-in-hand with stopped");
209
210 return ptr;
211 }
212
213
214 // returns IfNode
215 IfNode* Parse::jump_if_fork_int(Node* a, Node* b, BoolTest::mask mask, float prob, float cnt) {
216 Node *cmp = _gvn.transform(new CmpINode(a, b)); // two cases: shiftcount > 32 and shiftcount <= 32
217 Node *tst = _gvn.transform(new BoolNode(cmp, mask));
218 IfNode *iff = create_and_map_if(control(), tst, prob, cnt);
219 return iff;
220 }
221
222
223 // sentinel value for the target bci to mark never taken branches
224 // (according to profiling)
225 static const int never_reached = INT_MAX;
226
227 //------------------------------helper for tableswitch-------------------------
228 void Parse::jump_if_true_fork(IfNode *iff, int dest_bci_if_true, bool unc) {
229 // True branch, use existing map info
230 { PreserveJVMState pjvms(this);
231 Node *iftrue = _gvn.transform( new IfTrueNode (iff) );
232 set_control( iftrue );
233 if (unc) {
234 repush_if_args();
235 uncommon_trap(Deoptimization::Reason_unstable_if,
236 Deoptimization::Action_reinterpret,
237 nullptr,
238 "taken always");
239 } else {
240 assert(dest_bci_if_true != never_reached, "inconsistent dest");
241 merge_new_path(dest_bci_if_true);
242 }
243 }
244
245 // False branch
246 Node *iffalse = _gvn.transform( new IfFalseNode(iff) );
247 set_control( iffalse );
248 }
249
250 void Parse::jump_if_false_fork(IfNode *iff, int dest_bci_if_true, bool unc) {
251 // True branch, use existing map info
252 { PreserveJVMState pjvms(this);
253 Node *iffalse = _gvn.transform( new IfFalseNode (iff) );
254 set_control( iffalse );
255 if (unc) {
256 repush_if_args();
257 uncommon_trap(Deoptimization::Reason_unstable_if,
258 Deoptimization::Action_reinterpret,
259 nullptr,
260 "taken never");
261 } else {
262 assert(dest_bci_if_true != never_reached, "inconsistent dest");
263 merge_new_path(dest_bci_if_true);
264 }
265 }
266
267 // False branch
268 Node *iftrue = _gvn.transform( new IfTrueNode(iff) );
269 set_control( iftrue );
270 }
271
272 void Parse::jump_if_always_fork(int dest_bci, bool unc) {
273 // False branch, use existing map and control()
274 if (unc) {
275 repush_if_args();
276 uncommon_trap(Deoptimization::Reason_unstable_if,
277 Deoptimization::Action_reinterpret,
278 nullptr,
279 "taken never");
280 } else {
281 assert(dest_bci != never_reached, "inconsistent dest");
282 merge_new_path(dest_bci);
283 }
284 }
285
286
287 extern "C" {
288 static int jint_cmp(const void *i, const void *j) {
289 int a = *(jint *)i;
290 int b = *(jint *)j;
291 return a > b ? 1 : a < b ? -1 : 0;
292 }
293 }
294
295
296 class SwitchRange : public StackObj {
297 // a range of integers coupled with a bci destination
298 jint _lo; // inclusive lower limit
299 jint _hi; // inclusive upper limit
300 int _dest;
301 float _cnt; // how many times this range was hit according to profiling
302
303 public:
304 jint lo() const { return _lo; }
305 jint hi() const { return _hi; }
306 int dest() const { return _dest; }
307 bool is_singleton() const { return _lo == _hi; }
308 float cnt() const { return _cnt; }
309
310 void setRange(jint lo, jint hi, int dest, float cnt) {
311 assert(lo <= hi, "must be a non-empty range");
312 _lo = lo, _hi = hi; _dest = dest; _cnt = cnt;
313 assert(_cnt >= 0, "");
314 }
315 bool adjoinRange(jint lo, jint hi, int dest, float cnt, bool trim_ranges) {
316 assert(lo <= hi, "must be a non-empty range");
317 if (lo == _hi+1) {
318 // see merge_ranges() comment below
319 if (trim_ranges) {
320 if (cnt == 0) {
321 if (_cnt != 0) {
322 return false;
323 }
324 if (dest != _dest) {
325 _dest = never_reached;
326 }
327 } else {
328 if (_cnt == 0) {
329 return false;
330 }
331 if (dest != _dest) {
332 return false;
333 }
334 }
335 } else {
336 if (dest != _dest) {
337 return false;
338 }
339 }
340 _hi = hi;
341 _cnt += cnt;
342 return true;
343 }
344 return false;
345 }
346
347 void set (jint value, int dest, float cnt) {
348 setRange(value, value, dest, cnt);
349 }
350 bool adjoin(jint value, int dest, float cnt, bool trim_ranges) {
351 return adjoinRange(value, value, dest, cnt, trim_ranges);
352 }
353 bool adjoin(SwitchRange& other) {
354 return adjoinRange(other._lo, other._hi, other._dest, other._cnt, false);
355 }
356
357 void print() {
358 if (is_singleton())
359 tty->print(" {%d}=>%d (cnt=%f)", lo(), dest(), cnt());
360 else if (lo() == min_jint)
361 tty->print(" {..%d}=>%d (cnt=%f)", hi(), dest(), cnt());
362 else if (hi() == max_jint)
363 tty->print(" {%d..}=>%d (cnt=%f)", lo(), dest(), cnt());
364 else
365 tty->print(" {%d..%d}=>%d (cnt=%f)", lo(), hi(), dest(), cnt());
366 }
367 };
368
369 // We try to minimize the number of ranges and the size of the taken
370 // ones using profiling data. When ranges are created,
371 // SwitchRange::adjoinRange() only allows 2 adjoining ranges to merge
372 // if both were never hit or both were hit to build longer unreached
373 // ranges. Here, we now merge adjoining ranges with the same
374 // destination and finally set destination of unreached ranges to the
375 // special value never_reached because it can help minimize the number
376 // of tests that are necessary.
377 //
378 // For instance:
379 // [0, 1] to target1 sometimes taken
380 // [1, 2] to target1 never taken
381 // [2, 3] to target2 never taken
382 // would lead to:
383 // [0, 1] to target1 sometimes taken
384 // [1, 3] never taken
385 //
386 // (first 2 ranges to target1 are not merged)
387 static void merge_ranges(SwitchRange* ranges, int& rp) {
388 if (rp == 0) {
389 return;
390 }
391 int shift = 0;
392 for (int j = 0; j < rp; j++) {
393 SwitchRange& r1 = ranges[j-shift];
394 SwitchRange& r2 = ranges[j+1];
395 if (r1.adjoin(r2)) {
396 shift++;
397 } else if (shift > 0) {
398 ranges[j+1-shift] = r2;
399 }
400 }
401 rp -= shift;
402 for (int j = 0; j <= rp; j++) {
403 SwitchRange& r = ranges[j];
404 if (r.cnt() == 0 && r.dest() != never_reached) {
405 r.setRange(r.lo(), r.hi(), never_reached, r.cnt());
406 }
407 }
408 }
409
410 //-------------------------------do_tableswitch--------------------------------
411 void Parse::do_tableswitch() {
412 // Get information about tableswitch
413 int default_dest = iter().get_dest_table(0);
414 jint lo_index = iter().get_int_table(1);
415 jint hi_index = iter().get_int_table(2);
416 int len = hi_index - lo_index + 1;
417
418 if (len < 1) {
419 // If this is a backward branch, add safepoint
420 maybe_add_safepoint(default_dest);
421 pop(); // the effect of the instruction execution on the operand stack
422 merge(default_dest);
423 return;
424 }
425
426 ciMethodData* methodData = method()->method_data();
427 ciMultiBranchData* profile = nullptr;
428 if (methodData->is_mature() && UseSwitchProfiling) {
429 ciProfileData* data = methodData->bci_to_data(bci());
430 if (data != nullptr && data->is_MultiBranchData()) {
431 profile = (ciMultiBranchData*)data;
432 }
433 }
434 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if);
435
436 // generate decision tree, using trichotomy when possible
437 int rnum = len+2;
438 bool makes_backward_branch = (default_dest <= bci());
439 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum);
440 int rp = -1;
441 if (lo_index != min_jint) {
442 float cnt = 1.0F;
443 if (profile != nullptr) {
444 cnt = (float)profile->default_count() / (hi_index != max_jint ? 2.0F : 1.0F);
445 }
446 ranges[++rp].setRange(min_jint, lo_index-1, default_dest, cnt);
447 }
448 for (int j = 0; j < len; j++) {
449 jint match_int = lo_index+j;
450 int dest = iter().get_dest_table(j+3);
451 makes_backward_branch |= (dest <= bci());
452 float cnt = 1.0F;
453 if (profile != nullptr) {
454 cnt = (float)profile->count_at(j);
455 }
456 if (rp < 0 || !ranges[rp].adjoin(match_int, dest, cnt, trim_ranges)) {
457 ranges[++rp].set(match_int, dest, cnt);
458 }
459 }
460 jint highest = lo_index+(len-1);
461 assert(ranges[rp].hi() == highest, "");
462 if (highest != max_jint) {
463 float cnt = 1.0F;
464 if (profile != nullptr) {
465 cnt = (float)profile->default_count() / (lo_index != min_jint ? 2.0F : 1.0F);
466 }
467 if (!ranges[rp].adjoinRange(highest+1, max_jint, default_dest, cnt, trim_ranges)) {
468 ranges[++rp].setRange(highest+1, max_jint, default_dest, cnt);
469 }
470 }
471 assert(rp < len+2, "not too many ranges");
472
473 if (trim_ranges) {
474 merge_ranges(ranges, rp);
475 }
476
477 // Safepoint in case if backward branch observed
478 if (makes_backward_branch) {
479 add_safepoint();
480 }
481
482 Node* lookup = pop(); // lookup value
483 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]);
484 }
485
486
487 //------------------------------do_lookupswitch--------------------------------
488 void Parse::do_lookupswitch() {
489 // Get information about lookupswitch
490 int default_dest = iter().get_dest_table(0);
491 jint len = iter().get_int_table(1);
492
493 if (len < 1) { // If this is a backward branch, add safepoint
494 maybe_add_safepoint(default_dest);
495 pop(); // the effect of the instruction execution on the operand stack
496 merge(default_dest);
497 return;
498 }
499
500 ciMethodData* methodData = method()->method_data();
501 ciMultiBranchData* profile = nullptr;
502 if (methodData->is_mature() && UseSwitchProfiling) {
503 ciProfileData* data = methodData->bci_to_data(bci());
504 if (data != nullptr && data->is_MultiBranchData()) {
505 profile = (ciMultiBranchData*)data;
506 }
507 }
508 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if);
509
510 // generate decision tree, using trichotomy when possible
511 jint* table = NEW_RESOURCE_ARRAY(jint, len*3);
512 {
513 for (int j = 0; j < len; j++) {
514 table[3*j+0] = iter().get_int_table(2+2*j);
515 table[3*j+1] = iter().get_dest_table(2+2*j+1);
516 // Handle overflow when converting from uint to jint
517 table[3*j+2] = (profile == nullptr) ? 1 : (jint)MIN2<uint>((uint)max_jint, profile->count_at(j));
518 }
519 qsort(table, len, 3*sizeof(table[0]), jint_cmp);
520 }
521
522 float default_cnt = 1.0F;
523 if (profile != nullptr) {
524 juint defaults = max_juint - len;
525 default_cnt = (float)profile->default_count()/(float)defaults;
526 }
527
528 int rnum = len*2+1;
529 bool makes_backward_branch = (default_dest <= bci());
530 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum);
531 int rp = -1;
532 for (int j = 0; j < len; j++) {
533 jint match_int = table[3*j+0];
534 jint dest = table[3*j+1];
535 jint cnt = table[3*j+2];
536 jint next_lo = rp < 0 ? min_jint : ranges[rp].hi()+1;
537 makes_backward_branch |= (dest <= bci());
538 float c = default_cnt * ((float)match_int - (float)next_lo);
539 if (match_int != next_lo && (rp < 0 || !ranges[rp].adjoinRange(next_lo, match_int-1, default_dest, c, trim_ranges))) {
540 assert(default_dest != never_reached, "sentinel value for dead destinations");
541 ranges[++rp].setRange(next_lo, match_int-1, default_dest, c);
542 }
543 if (rp < 0 || !ranges[rp].adjoin(match_int, dest, (float)cnt, trim_ranges)) {
544 assert(dest != never_reached, "sentinel value for dead destinations");
545 ranges[++rp].set(match_int, dest, (float)cnt);
546 }
547 }
548 jint highest = table[3*(len-1)];
549 assert(ranges[rp].hi() == highest, "");
550 if (highest != max_jint &&
551 !ranges[rp].adjoinRange(highest+1, max_jint, default_dest, default_cnt * ((float)max_jint - (float)highest), trim_ranges)) {
552 ranges[++rp].setRange(highest+1, max_jint, default_dest, default_cnt * ((float)max_jint - (float)highest));
553 }
554 assert(rp < rnum, "not too many ranges");
555
556 if (trim_ranges) {
557 merge_ranges(ranges, rp);
558 }
559
560 // Safepoint in case backward branch observed
561 if (makes_backward_branch) {
562 add_safepoint();
563 }
564
565 Node *lookup = pop(); // lookup value
566 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]);
567 }
568
569 static float if_prob(float taken_cnt, float total_cnt) {
570 assert(taken_cnt <= total_cnt, "");
571 if (total_cnt == 0) {
572 return PROB_FAIR;
573 }
574 float p = taken_cnt / total_cnt;
575 return clamp(p, PROB_MIN, PROB_MAX);
576 }
577
578 static float if_cnt(float cnt) {
579 if (cnt == 0) {
580 return COUNT_UNKNOWN;
581 }
582 return cnt;
583 }
584
585 static float sum_of_cnts(SwitchRange *lo, SwitchRange *hi) {
586 float total_cnt = 0;
587 for (SwitchRange* sr = lo; sr <= hi; sr++) {
588 total_cnt += sr->cnt();
589 }
590 return total_cnt;
591 }
592
593 class SwitchRanges : public ResourceObj {
594 public:
595 SwitchRange* _lo;
596 SwitchRange* _hi;
597 SwitchRange* _mid;
598 float _cost;
599
600 enum {
601 Start,
602 LeftDone,
603 RightDone,
604 Done
605 } _state;
606
607 SwitchRanges(SwitchRange *lo, SwitchRange *hi)
608 : _lo(lo), _hi(hi), _mid(nullptr),
609 _cost(0), _state(Start) {
610 }
611
612 SwitchRanges()
613 : _lo(nullptr), _hi(nullptr), _mid(nullptr),
614 _cost(0), _state(Start) {}
615 };
616
617 // Estimate cost of performing a binary search on lo..hi
618 static float compute_tree_cost(SwitchRange *lo, SwitchRange *hi, float total_cnt) {
619 GrowableArray<SwitchRanges> tree;
620 SwitchRanges root(lo, hi);
621 tree.push(root);
622
623 float cost = 0;
624 do {
625 SwitchRanges& r = *tree.adr_at(tree.length()-1);
626 if (r._hi != r._lo) {
627 if (r._mid == nullptr) {
628 float r_cnt = sum_of_cnts(r._lo, r._hi);
629
630 if (r_cnt == 0) {
631 tree.pop();
632 cost = 0;
633 continue;
634 }
635
636 SwitchRange* mid = nullptr;
637 mid = r._lo;
638 for (float cnt = 0; ; ) {
639 assert(mid <= r._hi, "out of bounds");
640 cnt += mid->cnt();
641 if (cnt > r_cnt / 2) {
642 break;
643 }
644 mid++;
645 }
646 assert(mid <= r._hi, "out of bounds");
647 r._mid = mid;
648 r._cost = r_cnt / total_cnt;
649 }
650 r._cost += cost;
651 if (r._state < SwitchRanges::LeftDone && r._mid > r._lo) {
652 cost = 0;
653 r._state = SwitchRanges::LeftDone;
654 tree.push(SwitchRanges(r._lo, r._mid-1));
655 } else if (r._state < SwitchRanges::RightDone) {
656 cost = 0;
657 r._state = SwitchRanges::RightDone;
658 tree.push(SwitchRanges(r._mid == r._lo ? r._mid+1 : r._mid, r._hi));
659 } else {
660 tree.pop();
661 cost = r._cost;
662 }
663 } else {
664 tree.pop();
665 cost = r._cost;
666 }
667 } while (tree.length() > 0);
668
669
670 return cost;
671 }
672
673 // It sometimes pays off to test most common ranges before the binary search
674 void Parse::linear_search_switch_ranges(Node* key_val, SwitchRange*& lo, SwitchRange*& hi) {
675 uint nr = hi - lo + 1;
676 float total_cnt = sum_of_cnts(lo, hi);
677
678 float min = compute_tree_cost(lo, hi, total_cnt);
679 float extra = 1;
680 float sub = 0;
681
682 SwitchRange* array1 = lo;
683 SwitchRange* array2 = NEW_RESOURCE_ARRAY(SwitchRange, nr);
684
685 SwitchRange* ranges = nullptr;
686
687 while (nr >= 2) {
688 assert(lo == array1 || lo == array2, "one the 2 already allocated arrays");
689 ranges = (lo == array1) ? array2 : array1;
690
691 // Find highest frequency range
692 SwitchRange* candidate = lo;
693 for (SwitchRange* sr = lo+1; sr <= hi; sr++) {
694 if (sr->cnt() > candidate->cnt()) {
695 candidate = sr;
696 }
697 }
698 SwitchRange most_freq = *candidate;
699 if (most_freq.cnt() == 0) {
700 break;
701 }
702
703 // Copy remaining ranges into another array
704 int shift = 0;
705 for (uint i = 0; i < nr; i++) {
706 SwitchRange* sr = &lo[i];
707 if (sr != candidate) {
708 ranges[i-shift] = *sr;
709 } else {
710 shift++;
711 if (i > 0 && i < nr-1) {
712 SwitchRange prev = lo[i-1];
713 prev.setRange(prev.lo(), sr->hi(), prev.dest(), prev.cnt());
714 if (prev.adjoin(lo[i+1])) {
715 shift++;
716 i++;
717 }
718 ranges[i-shift] = prev;
719 }
720 }
721 }
722 nr -= shift;
723
724 // Evaluate cost of testing the most common range and performing a
725 // binary search on the other ranges
726 float cost = extra + compute_tree_cost(&ranges[0], &ranges[nr-1], total_cnt);
727 if (cost >= min) {
728 break;
729 }
730 // swap arrays
731 lo = &ranges[0];
732 hi = &ranges[nr-1];
733
734 // It pays off: emit the test for the most common range
735 assert(most_freq.cnt() > 0, "must be taken");
736 Node* val = _gvn.transform(new SubINode(key_val, _gvn.intcon(most_freq.lo())));
737 Node* cmp = _gvn.transform(new CmpUNode(val, _gvn.intcon(java_subtract(most_freq.hi(), most_freq.lo()))));
738 Node* tst = _gvn.transform(new BoolNode(cmp, BoolTest::le));
739 IfNode* iff = create_and_map_if(control(), tst, if_prob(most_freq.cnt(), total_cnt), if_cnt(most_freq.cnt()));
740 jump_if_true_fork(iff, most_freq.dest(), false);
741
742 sub += most_freq.cnt() / total_cnt;
743 extra += 1 - sub;
744 min = cost;
745 }
746 }
747
748 //----------------------------create_jump_tables-------------------------------
749 bool Parse::create_jump_tables(Node* key_val, SwitchRange* lo, SwitchRange* hi) {
750 // Are jumptables enabled
751 if (!UseJumpTables) return false;
752
753 // Are jumptables supported
754 if (!Matcher::has_match_rule(Op_Jump)) return false;
755
756 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if);
757
758 // Decide if a guard is needed to lop off big ranges at either (or
759 // both) end(s) of the input set. We'll call this the default target
760 // even though we can't be sure that it is the true "default".
761
762 bool needs_guard = false;
763 int default_dest;
764 int64_t total_outlier_size = 0;
765 int64_t hi_size = ((int64_t)hi->hi()) - ((int64_t)hi->lo()) + 1;
766 int64_t lo_size = ((int64_t)lo->hi()) - ((int64_t)lo->lo()) + 1;
767
768 if (lo->dest() == hi->dest()) {
769 total_outlier_size = hi_size + lo_size;
770 default_dest = lo->dest();
771 } else if (lo_size > hi_size) {
772 total_outlier_size = lo_size;
773 default_dest = lo->dest();
774 } else {
775 total_outlier_size = hi_size;
776 default_dest = hi->dest();
777 }
778
779 float total = sum_of_cnts(lo, hi);
780 float cost = compute_tree_cost(lo, hi, total);
781
782 // If a guard test will eliminate very sparse end ranges, then
783 // it is worth the cost of an extra jump.
784 float trimmed_cnt = 0;
785 if (total_outlier_size > (MaxJumpTableSparseness * 4)) {
786 needs_guard = true;
787 if (default_dest == lo->dest()) {
788 trimmed_cnt += lo->cnt();
789 lo++;
790 }
791 if (default_dest == hi->dest()) {
792 trimmed_cnt += hi->cnt();
793 hi--;
794 }
795 }
796
797 // Find the total number of cases and ranges
798 int64_t num_cases = ((int64_t)hi->hi()) - ((int64_t)lo->lo()) + 1;
799 int num_range = hi - lo + 1;
800
801 // Don't create table if: too large, too small, or too sparse.
802 if (num_cases > MaxJumpTableSize)
803 return false;
804 if (UseSwitchProfiling) {
805 // MinJumpTableSize is set so with a well balanced binary tree,
806 // when the number of ranges is MinJumpTableSize, it's cheaper to
807 // go through a JumpNode that a tree of IfNodes. Average cost of a
808 // tree of IfNodes with MinJumpTableSize is
809 // log2f(MinJumpTableSize) comparisons. So if the cost computed
810 // from profile data is less than log2f(MinJumpTableSize) then
811 // going with the binary search is cheaper.
812 if (cost < log2f(MinJumpTableSize)) {
813 return false;
814 }
815 } else {
816 if (num_cases < MinJumpTableSize)
817 return false;
818 }
819 if (num_cases > (MaxJumpTableSparseness * num_range))
820 return false;
821
822 // Normalize table lookups to zero
823 int lowval = lo->lo();
824 key_val = _gvn.transform( new SubINode(key_val, _gvn.intcon(lowval)) );
825
826 // Generate a guard to protect against input keyvals that aren't
827 // in the switch domain.
828 if (needs_guard) {
829 Node* size = _gvn.intcon(num_cases);
830 Node* cmp = _gvn.transform(new CmpUNode(key_val, size));
831 Node* tst = _gvn.transform(new BoolNode(cmp, BoolTest::ge));
832 IfNode* iff = create_and_map_if(control(), tst, if_prob(trimmed_cnt, total), if_cnt(trimmed_cnt));
833 jump_if_true_fork(iff, default_dest, trim_ranges && trimmed_cnt == 0);
834
835 total -= trimmed_cnt;
836 }
837
838 // Create an ideal node JumpTable that has projections
839 // of all possible ranges for a switch statement
840 // The key_val input must be converted to a pointer offset and scaled.
841 // Compare Parse::array_addressing above.
842
843 // Clean the 32-bit int into a real 64-bit offset.
844 // Otherwise, the jint value 0 might turn into an offset of 0x0800000000.
845 // Make I2L conversion control dependent to prevent it from
846 // floating above the range check during loop optimizations.
847 // Do not use a narrow int type here to prevent the data path from dying
848 // while the control path is not removed. This can happen if the type of key_val
849 // is later known to be out of bounds of [0, num_cases] and therefore a narrow cast
850 // would be replaced by TOP while C2 is not able to fold the corresponding range checks.
851 // Set _carry_dependency for the cast to avoid being removed by IGVN.
852 #ifdef _LP64
853 key_val = C->constrained_convI2L(&_gvn, key_val, TypeInt::INT, control(), true /* carry_dependency */);
854 #endif
855
856 // Shift the value by wordsize so we have an index into the table, rather
857 // than a switch value
858 Node *shiftWord = _gvn.MakeConX(wordSize);
859 key_val = _gvn.transform( new MulXNode( key_val, shiftWord));
860
861 // Create the JumpNode
862 Arena* arena = C->comp_arena();
863 float* probs = (float*)arena->Amalloc(sizeof(float)*num_cases);
864 int i = 0;
865 if (total == 0) {
866 for (SwitchRange* r = lo; r <= hi; r++) {
867 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) {
868 probs[i] = 1.0F / num_cases;
869 }
870 }
871 } else {
872 for (SwitchRange* r = lo; r <= hi; r++) {
873 float prob = r->cnt()/total;
874 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) {
875 probs[i] = prob / (r->hi() - r->lo() + 1);
876 }
877 }
878 }
879
880 ciMethodData* methodData = method()->method_data();
881 ciMultiBranchData* profile = nullptr;
882 if (methodData->is_mature()) {
883 ciProfileData* data = methodData->bci_to_data(bci());
884 if (data != nullptr && data->is_MultiBranchData()) {
885 profile = (ciMultiBranchData*)data;
886 }
887 }
888
889 Node* jtn = _gvn.transform(new JumpNode(control(), key_val, num_cases, probs, profile == nullptr ? COUNT_UNKNOWN : total));
890
891 // These are the switch destinations hanging off the jumpnode
892 i = 0;
893 for (SwitchRange* r = lo; r <= hi; r++) {
894 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) {
895 Node* input = _gvn.transform(new JumpProjNode(jtn, i, r->dest(), (int)(j - lowval)));
896 {
897 PreserveJVMState pjvms(this);
898 set_control(input);
899 jump_if_always_fork(r->dest(), trim_ranges && r->cnt() == 0);
900 }
901 }
902 }
903 assert(i == num_cases, "miscount of cases");
904 stop_and_kill_map(); // no more uses for this JVMS
905 return true;
906 }
907
908 //----------------------------jump_switch_ranges-------------------------------
909 void Parse::jump_switch_ranges(Node* key_val, SwitchRange *lo, SwitchRange *hi, int switch_depth) {
910 Block* switch_block = block();
911 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if);
912
913 if (switch_depth == 0) {
914 // Do special processing for the top-level call.
915 assert(lo->lo() == min_jint, "initial range must exhaust Type::INT");
916 assert(hi->hi() == max_jint, "initial range must exhaust Type::INT");
917
918 // Decrement pred-numbers for the unique set of nodes.
919 #ifdef ASSERT
920 if (!trim_ranges) {
921 // Ensure that the block's successors are a (duplicate-free) set.
922 int successors_counted = 0; // block occurrences in [hi..lo]
923 int unique_successors = switch_block->num_successors();
924 for (int i = 0; i < unique_successors; i++) {
925 Block* target = switch_block->successor_at(i);
926
927 // Check that the set of successors is the same in both places.
928 int successors_found = 0;
929 for (SwitchRange* p = lo; p <= hi; p++) {
930 if (p->dest() == target->start()) successors_found++;
931 }
932 assert(successors_found > 0, "successor must be known");
933 successors_counted += successors_found;
934 }
935 assert(successors_counted == (hi-lo)+1, "no unexpected successors");
936 }
937 #endif
938
939 // Maybe prune the inputs, based on the type of key_val.
940 jint min_val = min_jint;
941 jint max_val = max_jint;
942 const TypeInt* ti = key_val->bottom_type()->isa_int();
943 if (ti != nullptr) {
944 min_val = ti->_lo;
945 max_val = ti->_hi;
946 assert(min_val <= max_val, "invalid int type");
947 }
948 while (lo->hi() < min_val) {
949 lo++;
950 }
951 if (lo->lo() < min_val) {
952 lo->setRange(min_val, lo->hi(), lo->dest(), lo->cnt());
953 }
954 while (hi->lo() > max_val) {
955 hi--;
956 }
957 if (hi->hi() > max_val) {
958 hi->setRange(hi->lo(), max_val, hi->dest(), hi->cnt());
959 }
960
961 linear_search_switch_ranges(key_val, lo, hi);
962 }
963
964 #ifndef PRODUCT
965 if (switch_depth == 0) {
966 _max_switch_depth = 0;
967 _est_switch_depth = log2i_graceful((hi - lo + 1) - 1) + 1;
968 }
969 #endif
970
971 assert(lo <= hi, "must be a non-empty set of ranges");
972 if (lo == hi) {
973 jump_if_always_fork(lo->dest(), trim_ranges && lo->cnt() == 0);
974 } else {
975 assert(lo->hi() == (lo+1)->lo()-1, "contiguous ranges");
976 assert(hi->lo() == (hi-1)->hi()+1, "contiguous ranges");
977
978 if (create_jump_tables(key_val, lo, hi)) return;
979
980 SwitchRange* mid = nullptr;
981 float total_cnt = sum_of_cnts(lo, hi);
982
983 int nr = hi - lo + 1;
984 if (UseSwitchProfiling) {
985 // Don't keep the binary search tree balanced: pick up mid point
986 // that split frequencies in half.
987 float cnt = 0;
988 for (SwitchRange* sr = lo; sr <= hi; sr++) {
989 cnt += sr->cnt();
990 if (cnt >= total_cnt / 2) {
991 mid = sr;
992 break;
993 }
994 }
995 } else {
996 mid = lo + nr/2;
997
998 // if there is an easy choice, pivot at a singleton:
999 if (nr > 3 && !mid->is_singleton() && (mid-1)->is_singleton()) mid--;
1000
1001 assert(lo < mid && mid <= hi, "good pivot choice");
1002 assert(nr != 2 || mid == hi, "should pick higher of 2");
1003 assert(nr != 3 || mid == hi-1, "should pick middle of 3");
1004 }
1005
1006
1007 Node *test_val = _gvn.intcon(mid == lo ? mid->hi() : mid->lo());
1008
1009 if (mid->is_singleton()) {
1010 IfNode *iff_ne = jump_if_fork_int(key_val, test_val, BoolTest::ne, 1-if_prob(mid->cnt(), total_cnt), if_cnt(mid->cnt()));
1011 jump_if_false_fork(iff_ne, mid->dest(), trim_ranges && mid->cnt() == 0);
1012
1013 // Special Case: If there are exactly three ranges, and the high
1014 // and low range each go to the same place, omit the "gt" test,
1015 // since it will not discriminate anything.
1016 bool eq_test_only = (hi == lo+2 && hi->dest() == lo->dest() && mid == hi-1) || mid == lo;
1017
1018 // if there is a higher range, test for it and process it:
1019 if (mid < hi && !eq_test_only) {
1020 // two comparisons of same values--should enable 1 test for 2 branches
1021 // Use BoolTest::lt instead of BoolTest::gt
1022 float cnt = sum_of_cnts(lo, mid-1);
1023 IfNode *iff_lt = jump_if_fork_int(key_val, test_val, BoolTest::lt, if_prob(cnt, total_cnt), if_cnt(cnt));
1024 Node *iftrue = _gvn.transform( new IfTrueNode(iff_lt) );
1025 Node *iffalse = _gvn.transform( new IfFalseNode(iff_lt) );
1026 { PreserveJVMState pjvms(this);
1027 set_control(iffalse);
1028 jump_switch_ranges(key_val, mid+1, hi, switch_depth+1);
1029 }
1030 set_control(iftrue);
1031 }
1032
1033 } else {
1034 // mid is a range, not a singleton, so treat mid..hi as a unit
1035 float cnt = sum_of_cnts(mid == lo ? mid+1 : mid, hi);
1036 IfNode *iff_ge = jump_if_fork_int(key_val, test_val, mid == lo ? BoolTest::gt : BoolTest::ge, if_prob(cnt, total_cnt), if_cnt(cnt));
1037
1038 // if there is a higher range, test for it and process it:
1039 if (mid == hi) {
1040 jump_if_true_fork(iff_ge, mid->dest(), trim_ranges && cnt == 0);
1041 } else {
1042 Node *iftrue = _gvn.transform( new IfTrueNode(iff_ge) );
1043 Node *iffalse = _gvn.transform( new IfFalseNode(iff_ge) );
1044 { PreserveJVMState pjvms(this);
1045 set_control(iftrue);
1046 jump_switch_ranges(key_val, mid == lo ? mid+1 : mid, hi, switch_depth+1);
1047 }
1048 set_control(iffalse);
1049 }
1050 }
1051
1052 // in any case, process the lower range
1053 if (mid == lo) {
1054 if (mid->is_singleton()) {
1055 jump_switch_ranges(key_val, lo+1, hi, switch_depth+1);
1056 } else {
1057 jump_if_always_fork(lo->dest(), trim_ranges && lo->cnt() == 0);
1058 }
1059 } else {
1060 jump_switch_ranges(key_val, lo, mid-1, switch_depth+1);
1061 }
1062 }
1063
1064 // Decrease pred_count for each successor after all is done.
1065 if (switch_depth == 0) {
1066 int unique_successors = switch_block->num_successors();
1067 for (int i = 0; i < unique_successors; i++) {
1068 Block* target = switch_block->successor_at(i);
1069 // Throw away the pre-allocated path for each unique successor.
1070 target->next_path_num();
1071 }
1072 }
1073
1074 #ifndef PRODUCT
1075 _max_switch_depth = MAX2(switch_depth, _max_switch_depth);
1076 if (TraceOptoParse && Verbose && WizardMode && switch_depth == 0) {
1077 SwitchRange* r;
1078 int nsing = 0;
1079 for( r = lo; r <= hi; r++ ) {
1080 if( r->is_singleton() ) nsing++;
1081 }
1082 tty->print(">>> ");
1083 _method->print_short_name();
1084 tty->print_cr(" switch decision tree");
1085 tty->print_cr(" %d ranges (%d singletons), max_depth=%d, est_depth=%d",
1086 (int) (hi-lo+1), nsing, _max_switch_depth, _est_switch_depth);
1087 if (_max_switch_depth > _est_switch_depth) {
1088 tty->print_cr("******** BAD SWITCH DEPTH ********");
1089 }
1090 tty->print(" ");
1091 for( r = lo; r <= hi; r++ ) {
1092 r->print();
1093 }
1094 tty->cr();
1095 }
1096 #endif
1097 }
1098
1099 Node* Parse::floating_point_mod(Node* a, Node* b, BasicType type) {
1100 assert(type == BasicType::T_FLOAT || type == BasicType::T_DOUBLE, "only float and double are floating points");
1101 CallLeafPureNode* mod = type == BasicType::T_DOUBLE ? static_cast<CallLeafPureNode*>(new ModDNode(C, a, b)) : new ModFNode(C, a, b);
1102
1103 set_predefined_input_for_runtime_call(mod);
1104 mod = _gvn.transform(mod)->as_CallLeafPure();
1105 set_predefined_output_for_runtime_call(mod);
1106 Node* result = _gvn.transform(new ProjNode(mod, TypeFunc::Parms + 0));
1107 record_for_igvn(mod);
1108 return result;
1109 }
1110
1111 void Parse::l2f() {
1112 Node* f2 = pop();
1113 Node* f1 = pop();
1114 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::l2f_Type(),
1115 CAST_FROM_FN_PTR(address, SharedRuntime::l2f),
1116 "l2f", nullptr, //no memory effects
1117 f1, f2);
1118 Node* res = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0));
1119
1120 push(res);
1121 }
1122
1123 // Handle jsr and jsr_w bytecode
1124 void Parse::do_jsr() {
1125 assert(bc() == Bytecodes::_jsr || bc() == Bytecodes::_jsr_w, "wrong bytecode");
1126
1127 // Store information about current state, tagged with new _jsr_bci
1128 int return_bci = iter().next_bci();
1129 int jsr_bci = (bc() == Bytecodes::_jsr) ? iter().get_dest() : iter().get_far_dest();
1130
1131 // The way we do things now, there is only one successor block
1132 // for the jsr, because the target code is cloned by ciTypeFlow.
1133 Block* target = successor_for_bci(jsr_bci);
1134
1135 // What got pushed?
1136 const Type* ret_addr = target->peek();
1137 assert(ret_addr->singleton(), "must be a constant (cloned jsr body)");
1138
1139 // Effect on jsr on stack
1140 push(_gvn.makecon(ret_addr));
1141
1142 // Flow to the jsr.
1143 merge(jsr_bci);
1144 }
1145
1146 // Handle ret bytecode
1147 void Parse::do_ret() {
1148 // Find to whom we return.
1149 assert(block()->num_successors() == 1, "a ret can only go one place now");
1150 Block* target = block()->successor_at(0);
1151 assert(!target->is_ready(), "our arrival must be expected");
1152 int pnum = target->next_path_num();
1153 merge_common(target, pnum);
1154 }
1155
1156 static bool has_injected_profile(BoolTest::mask btest, Node* test, int& taken, int& not_taken) {
1157 if (btest != BoolTest::eq && btest != BoolTest::ne) {
1158 // Only ::eq and ::ne are supported for profile injection.
1159 return false;
1160 }
1161 if (test->is_Cmp() &&
1162 test->in(1)->Opcode() == Op_ProfileBoolean) {
1163 ProfileBooleanNode* profile = (ProfileBooleanNode*)test->in(1);
1164 int false_cnt = profile->false_count();
1165 int true_cnt = profile->true_count();
1166
1167 // Counts matching depends on the actual test operation (::eq or ::ne).
1168 // No need to scale the counts because profile injection was designed
1169 // to feed exact counts into VM.
1170 taken = (btest == BoolTest::eq) ? false_cnt : true_cnt;
1171 not_taken = (btest == BoolTest::eq) ? true_cnt : false_cnt;
1172
1173 profile->consume();
1174 return true;
1175 }
1176 return false;
1177 }
1178
1179 // Give up if too few (or too many, in which case the sum will overflow) counts to be meaningful.
1180 // We also check that individual counters are positive first, otherwise the sum can become positive.
1181 // (check for saturation, integer overflow, and immature counts)
1182 static bool counters_are_meaningful(int counter1, int counter2, int min) {
1183 // check for saturation, including "uint" values too big to fit in "int"
1184 if (counter1 < 0 || counter2 < 0) {
1185 return false;
1186 }
1187 // check for integer overflow of the sum
1188 int64_t sum = (int64_t)counter1 + (int64_t)counter2;
1189 STATIC_ASSERT(sizeof(counter1) < sizeof(sum));
1190 if (sum > INT_MAX) {
1191 return false;
1192 }
1193 // check if mature
1194 return (counter1 + counter2) >= min;
1195 }
1196
1197 //--------------------------dynamic_branch_prediction--------------------------
1198 // Try to gather dynamic branch prediction behavior. Return a probability
1199 // of the branch being taken and set the "cnt" field. Returns a -1.0
1200 // if we need to use static prediction for some reason.
1201 float Parse::dynamic_branch_prediction(float &cnt, BoolTest::mask btest, Node* test) {
1202 ResourceMark rm;
1203
1204 cnt = COUNT_UNKNOWN;
1205
1206 int taken = 0;
1207 int not_taken = 0;
1208
1209 bool use_mdo = !has_injected_profile(btest, test, taken, not_taken);
1210
1211 if (use_mdo) {
1212 // Use MethodData information if it is available
1213 // FIXME: free the ProfileData structure
1214 ciMethodData* methodData = method()->method_data();
1215 if (!methodData->is_mature()) return PROB_UNKNOWN;
1216 ciProfileData* data = methodData->bci_to_data(bci());
1217 if (data == nullptr) {
1218 return PROB_UNKNOWN;
1219 }
1220 if (!data->is_JumpData()) return PROB_UNKNOWN;
1221
1222 // get taken and not taken values
1223 // NOTE: saturated UINT_MAX values become negative,
1224 // as do counts above INT_MAX.
1225 taken = data->as_JumpData()->taken();
1226 not_taken = 0;
1227 if (data->is_BranchData()) {
1228 not_taken = data->as_BranchData()->not_taken();
1229 }
1230
1231 // scale the counts to be commensurate with invocation counts:
1232 // NOTE: overflow for positive values is clamped at INT_MAX
1233 taken = method()->scale_count(taken);
1234 not_taken = method()->scale_count(not_taken);
1235 }
1236 // At this point, saturation or overflow is indicated by INT_MAX
1237 // or a negative value.
1238
1239 // Give up if too few (or too many, in which case the sum will overflow) counts to be meaningful.
1240 // We also check that individual counters are positive first, otherwise the sum can become positive.
1241 if (!counters_are_meaningful(taken, not_taken, 40)) {
1242 if (C->log() != nullptr) {
1243 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d'", iter().get_dest(), taken, not_taken);
1244 }
1245 return PROB_UNKNOWN;
1246 }
1247
1248 // Compute frequency that we arrive here
1249 float sum = taken + not_taken;
1250 // Adjust, if this block is a cloned private block but the
1251 // Jump counts are shared. Taken the private counts for
1252 // just this path instead of the shared counts.
1253 if( block()->count() > 0 )
1254 sum = block()->count();
1255 cnt = sum / FreqCountInvocations;
1256
1257 // Pin probability to sane limits
1258 float prob;
1259 if( !taken )
1260 prob = (0+PROB_MIN) / 2;
1261 else if( !not_taken )
1262 prob = (1+PROB_MAX) / 2;
1263 else { // Compute probability of true path
1264 prob = (float)taken / (float)(taken + not_taken);
1265 if (prob > PROB_MAX) prob = PROB_MAX;
1266 if (prob < PROB_MIN) prob = PROB_MIN;
1267 }
1268
1269 assert((cnt > 0.0f) && (prob > 0.0f),
1270 "Bad frequency assignment in if cnt=%g prob=%g taken=%d not_taken=%d", cnt, prob, taken, not_taken);
1271
1272 if (C->log() != nullptr) {
1273 const char* prob_str = nullptr;
1274 if (prob >= PROB_MAX) prob_str = (prob == PROB_MAX) ? "max" : "always";
1275 if (prob <= PROB_MIN) prob_str = (prob == PROB_MIN) ? "min" : "never";
1276 char prob_str_buf[30];
1277 if (prob_str == nullptr) {
1278 jio_snprintf(prob_str_buf, sizeof(prob_str_buf), "%20.2f", prob);
1279 prob_str = prob_str_buf;
1280 // The %20.2f adds many spaces to the string, to avoid some
1281 // picky overflow warning as noted in 8211929. But, 20 is the
1282 // *minimum* width, not *maximum*, so it's not clear how this
1283 // helps prevent overflow. Looks like we were forced to work
1284 // around a bug in gcc. In any case, strip the blanks.
1285 while (*prob_str == ' ') ++prob_str;
1286 }
1287 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d' cnt='%f' prob='%s'",
1288 iter().get_dest(), taken, not_taken, cnt, prob_str);
1289 }
1290 return prob;
1291 }
1292
1293 //-----------------------------branch_prediction-------------------------------
1294 float Parse::branch_prediction(float& cnt,
1295 BoolTest::mask btest,
1296 int target_bci,
1297 Node* test) {
1298 float prob = dynamic_branch_prediction(cnt, btest, test);
1299 // If prob is unknown, switch to static prediction
1300 if (prob != PROB_UNKNOWN) return prob;
1301
1302 prob = PROB_FAIR; // Set default value
1303 if (btest == BoolTest::eq) // Exactly equal test?
1304 prob = PROB_STATIC_INFREQUENT; // Assume its relatively infrequent
1305 else if (btest == BoolTest::ne)
1306 prob = PROB_STATIC_FREQUENT; // Assume its relatively frequent
1307
1308 // If this is a conditional test guarding a backwards branch,
1309 // assume its a loop-back edge. Make it a likely taken branch.
1310 if (target_bci < bci()) {
1311 if (is_osr_parse()) { // Could be a hot OSR'd loop; force deopt
1312 // Since it's an OSR, we probably have profile data, but since
1313 // branch_prediction returned PROB_UNKNOWN, the counts are too small.
1314 // Let's make a special check here for completely zero counts.
1315 ciMethodData* methodData = method()->method_data();
1316 if (!methodData->is_empty()) {
1317 ciProfileData* data = methodData->bci_to_data(bci());
1318 // Only stop for truly zero counts, which mean an unknown part
1319 // of the OSR-ed method, and we want to deopt to gather more stats.
1320 // If you have ANY counts, then this loop is simply 'cold' relative
1321 // to the OSR loop.
1322 if (data == nullptr ||
1323 (data->as_BranchData()->taken() + data->as_BranchData()->not_taken() == 0)) {
1324 // This is the only way to return PROB_UNKNOWN:
1325 return PROB_UNKNOWN;
1326 }
1327 }
1328 }
1329 prob = PROB_STATIC_FREQUENT; // Likely to take backwards branch
1330 }
1331
1332 assert(prob != PROB_UNKNOWN, "must have some guess at this point");
1333 return prob;
1334 }
1335
1336 // The magic constants are chosen so as to match the output of
1337 // branch_prediction() when the profile reports a zero taken count.
1338 // It is important to distinguish zero counts unambiguously, because
1339 // some branches (e.g., _213_javac.Assembler.eliminate) validly produce
1340 // very small but nonzero probabilities, which if confused with zero
1341 // counts would keep the program recompiling indefinitely.
1342 bool Parse::seems_never_taken(float prob) const {
1343 return prob < PROB_MIN;
1344 }
1345
1346 //-------------------------------repush_if_args--------------------------------
1347 // Push arguments of an "if" bytecode back onto the stack by adjusting _sp.
1348 inline int Parse::repush_if_args() {
1349 if (PrintOpto && WizardMode) {
1350 tty->print("defending against excessive implicit null exceptions on %s @%d in ",
1351 Bytecodes::name(iter().cur_bc()), iter().cur_bci());
1352 method()->print_name(); tty->cr();
1353 }
1354 int bc_depth = - Bytecodes::depth(iter().cur_bc());
1355 assert(bc_depth == 1 || bc_depth == 2, "only two kinds of branches");
1356 DEBUG_ONLY(sync_jvms()); // argument(n) requires a synced jvms
1357 assert(argument(0) != nullptr, "must exist");
1358 assert(bc_depth == 1 || argument(1) != nullptr, "two must exist");
1359 inc_sp(bc_depth);
1360 return bc_depth;
1361 }
1362
1363 // Used by StressUnstableIfTraps
1364 static volatile int _trap_stress_counter = 0;
1365
1366 void Parse::increment_trap_stress_counter(Node*& counter, Node*& incr_store) {
1367 Node* counter_addr = makecon(TypeRawPtr::make((address)&_trap_stress_counter));
1368 counter = make_load(control(), counter_addr, TypeInt::INT, T_INT, MemNode::unordered);
1369 counter = _gvn.transform(new AddINode(counter, intcon(1)));
1370 incr_store = store_to_memory(control(), counter_addr, counter, T_INT, MemNode::unordered);
1371 }
1372
1373 //----------------------------------do_ifnull----------------------------------
1374 void Parse::do_ifnull(BoolTest::mask btest, Node *c) {
1375 int target_bci = iter().get_dest();
1376
1377 Node* counter = nullptr;
1378 Node* incr_store = nullptr;
1379 bool do_stress_trap = StressUnstableIfTraps && ((C->random() % 2) == 0);
1380 if (do_stress_trap) {
1381 increment_trap_stress_counter(counter, incr_store);
1382 }
1383
1384 Block* branch_block = successor_for_bci(target_bci);
1385 Block* next_block = successor_for_bci(iter().next_bci());
1386
1387 float cnt;
1388 float prob = branch_prediction(cnt, btest, target_bci, c);
1389 if (prob == PROB_UNKNOWN) {
1390 // (An earlier version of do_ifnull omitted this trap for OSR methods.)
1391 if (PrintOpto && Verbose) {
1392 tty->print_cr("Never-taken edge stops compilation at bci %d", bci());
1393 }
1394 repush_if_args(); // to gather stats on loop
1395 uncommon_trap(Deoptimization::Reason_unreached,
1396 Deoptimization::Action_reinterpret,
1397 nullptr, "cold");
1398 if (C->eliminate_boxing()) {
1399 // Mark the successor blocks as parsed
1400 branch_block->next_path_num();
1401 next_block->next_path_num();
1402 }
1403 return;
1404 }
1405
1406 NOT_PRODUCT(explicit_null_checks_inserted++);
1407
1408 // Generate real control flow
1409 Node *tst = _gvn.transform( new BoolNode( c, btest ) );
1410
1411 // Sanity check the probability value
1412 assert(prob > 0.0f,"Bad probability in Parser");
1413 // Need xform to put node in hash table
1414 IfNode *iff = create_and_xform_if( control(), tst, prob, cnt );
1415 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser");
1416 // True branch
1417 { PreserveJVMState pjvms(this);
1418 Node* iftrue = _gvn.transform( new IfTrueNode (iff) );
1419 set_control(iftrue);
1420
1421 if (stopped()) { // Path is dead?
1422 NOT_PRODUCT(explicit_null_checks_elided++);
1423 if (C->eliminate_boxing()) {
1424 // Mark the successor block as parsed
1425 branch_block->next_path_num();
1426 }
1427 } else { // Path is live.
1428 adjust_map_after_if(btest, c, prob, branch_block);
1429 if (!stopped()) {
1430 merge(target_bci);
1431 }
1432 }
1433 }
1434
1435 // False branch
1436 Node* iffalse = _gvn.transform( new IfFalseNode(iff) );
1437 set_control(iffalse);
1438
1439 if (stopped()) { // Path is dead?
1440 NOT_PRODUCT(explicit_null_checks_elided++);
1441 if (C->eliminate_boxing()) {
1442 // Mark the successor block as parsed
1443 next_block->next_path_num();
1444 }
1445 } else { // Path is live.
1446 adjust_map_after_if(BoolTest(btest).negate(), c, 1.0-prob, next_block);
1447 }
1448
1449 if (do_stress_trap) {
1450 stress_trap(iff, counter, incr_store);
1451 }
1452 }
1453
1454 //------------------------------------do_if------------------------------------
1455 void Parse::do_if(BoolTest::mask btest, Node* c) {
1456 int target_bci = iter().get_dest();
1457
1458 Block* branch_block = successor_for_bci(target_bci);
1459 Block* next_block = successor_for_bci(iter().next_bci());
1460
1461 float cnt;
1462 float prob = branch_prediction(cnt, btest, target_bci, c);
1463 float untaken_prob = 1.0 - prob;
1464
1465 if (prob == PROB_UNKNOWN) {
1466 if (PrintOpto && Verbose) {
1467 tty->print_cr("Never-taken edge stops compilation at bci %d", bci());
1468 }
1469 repush_if_args(); // to gather stats on loop
1470 uncommon_trap(Deoptimization::Reason_unreached,
1471 Deoptimization::Action_reinterpret,
1472 nullptr, "cold");
1473 if (C->eliminate_boxing()) {
1474 // Mark the successor blocks as parsed
1475 branch_block->next_path_num();
1476 next_block->next_path_num();
1477 }
1478 return;
1479 }
1480
1481 Node* counter = nullptr;
1482 Node* incr_store = nullptr;
1483 bool do_stress_trap = StressUnstableIfTraps && ((C->random() % 2) == 0);
1484 if (do_stress_trap) {
1485 increment_trap_stress_counter(counter, incr_store);
1486 }
1487
1488 // Sanity check the probability value
1489 assert(0.0f < prob && prob < 1.0f,"Bad probability in Parser");
1490
1491 bool taken_if_true = true;
1492 // Convert BoolTest to canonical form:
1493 if (!BoolTest(btest).is_canonical()) {
1494 btest = BoolTest(btest).negate();
1495 taken_if_true = false;
1496 // prob is NOT updated here; it remains the probability of the taken
1497 // path (as opposed to the prob of the path guarded by an 'IfTrueNode').
1498 }
1499 assert(btest != BoolTest::eq, "!= is the only canonical exact test");
1500
1501 Node* tst0 = new BoolNode(c, btest);
1502 Node* tst = _gvn.transform(tst0);
1503 BoolTest::mask taken_btest = BoolTest::illegal;
1504 BoolTest::mask untaken_btest = BoolTest::illegal;
1505
1506 if (tst->is_Bool()) {
1507 // Refresh c from the transformed bool node, since it may be
1508 // simpler than the original c. Also re-canonicalize btest.
1509 // This wins when (Bool ne (Conv2B p) 0) => (Bool ne (CmpP p null)).
1510 // That can arise from statements like: if (x instanceof C) ...
1511 if (tst != tst0) {
1512 // Canonicalize one more time since transform can change it.
1513 btest = tst->as_Bool()->_test._test;
1514 if (!BoolTest(btest).is_canonical()) {
1515 // Reverse edges one more time...
1516 tst = _gvn.transform( tst->as_Bool()->negate(&_gvn) );
1517 btest = tst->as_Bool()->_test._test;
1518 assert(BoolTest(btest).is_canonical(), "sanity");
1519 taken_if_true = !taken_if_true;
1520 }
1521 c = tst->in(1);
1522 }
1523 BoolTest::mask neg_btest = BoolTest(btest).negate();
1524 taken_btest = taken_if_true ? btest : neg_btest;
1525 untaken_btest = taken_if_true ? neg_btest : btest;
1526 }
1527
1528 // Generate real control flow
1529 float true_prob = (taken_if_true ? prob : untaken_prob);
1530 IfNode* iff = create_and_map_if(control(), tst, true_prob, cnt);
1531 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser");
1532 Node* taken_branch = new IfTrueNode(iff);
1533 Node* untaken_branch = new IfFalseNode(iff);
1534 if (!taken_if_true) { // Finish conversion to canonical form
1535 Node* tmp = taken_branch;
1536 taken_branch = untaken_branch;
1537 untaken_branch = tmp;
1538 }
1539
1540 // Branch is taken:
1541 { PreserveJVMState pjvms(this);
1542 taken_branch = _gvn.transform(taken_branch);
1543 set_control(taken_branch);
1544
1545 if (stopped()) {
1546 if (C->eliminate_boxing()) {
1547 // Mark the successor block as parsed
1548 branch_block->next_path_num();
1549 }
1550 } else {
1551 adjust_map_after_if(taken_btest, c, prob, branch_block);
1552 if (!stopped()) {
1553 merge(target_bci);
1554 }
1555 }
1556 }
1557
1558 untaken_branch = _gvn.transform(untaken_branch);
1559 set_control(untaken_branch);
1560
1561 // Branch not taken.
1562 if (stopped()) {
1563 if (C->eliminate_boxing()) {
1564 // Mark the successor block as parsed
1565 next_block->next_path_num();
1566 }
1567 } else {
1568 adjust_map_after_if(untaken_btest, c, untaken_prob, next_block);
1569 }
1570
1571 if (do_stress_trap) {
1572 stress_trap(iff, counter, incr_store);
1573 }
1574 }
1575
1576 // Force unstable if traps to be taken randomly to trigger intermittent bugs such as incorrect debug information.
1577 // Add another if before the unstable if that checks a "random" condition at runtime (a simple shared counter) and
1578 // then either takes the trap or executes the original, unstable if.
1579 void Parse::stress_trap(IfNode* orig_iff, Node* counter, Node* incr_store) {
1580 // Search for an unstable if trap
1581 CallStaticJavaNode* trap = nullptr;
1582 assert(orig_iff->Opcode() == Op_If && orig_iff->outcnt() == 2, "malformed if");
1583 ProjNode* trap_proj = orig_iff->uncommon_trap_proj(trap, Deoptimization::Reason_unstable_if);
1584 if (trap == nullptr || !trap->jvms()->should_reexecute()) {
1585 // No suitable trap found. Remove unused counter load and increment.
1586 C->gvn_replace_by(incr_store, incr_store->in(MemNode::Memory));
1587 return;
1588 }
1589
1590 // Remove trap from optimization list since we add another path to the trap.
1591 bool success = C->remove_unstable_if_trap(trap, true);
1592 assert(success, "Trap already modified");
1593
1594 // Add a check before the original if that will trap with a certain frequency and execute the original if otherwise
1595 int freq_log = (C->random() % 31) + 1; // Random logarithmic frequency in [1, 31]
1596 Node* mask = intcon(right_n_bits(freq_log));
1597 counter = _gvn.transform(new AndINode(counter, mask));
1598 Node* cmp = _gvn.transform(new CmpINode(counter, intcon(0)));
1599 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::mask::eq));
1600 IfNode* iff = _gvn.transform(new IfNode(orig_iff->in(0), bol, orig_iff->_prob, orig_iff->_fcnt))->as_If();
1601 Node* if_true = _gvn.transform(new IfTrueNode(iff));
1602 Node* if_false = _gvn.transform(new IfFalseNode(iff));
1603 assert(!if_true->is_top() && !if_false->is_top(), "trap always / never taken");
1604
1605 // Trap
1606 assert(trap_proj->outcnt() == 1, "some other nodes are dependent on the trap projection");
1607
1608 Node* trap_region = new RegionNode(3);
1609 trap_region->set_req(1, trap_proj);
1610 trap_region->set_req(2, if_true);
1611 trap->set_req(0, _gvn.transform(trap_region));
1612
1613 // Don't trap, execute original if
1614 orig_iff->set_req(0, if_false);
1615 }
1616
1617 bool Parse::path_is_suitable_for_uncommon_trap(float prob) const {
1618 // Randomly skip emitting an uncommon trap
1619 if (StressUnstableIfTraps && ((C->random() % 2) == 0)) {
1620 return false;
1621 }
1622 // Don't want to speculate on uncommon traps when running with -Xcomp
1623 if (!UseInterpreter) {
1624 return false;
1625 }
1626 return seems_never_taken(prob) &&
1627 !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if);
1628 }
1629
1630 void Parse::maybe_add_predicate_after_if(Block* path) {
1631 if (path->is_SEL_head() && path->preds_parsed() == 0) {
1632 // Add predicates at bci of if dominating the loop so traps can be
1633 // recorded on the if's profile data
1634 int bc_depth = repush_if_args();
1635 add_parse_predicates();
1636 dec_sp(bc_depth);
1637 path->set_has_predicates();
1638 }
1639 }
1640
1641
1642 //----------------------------adjust_map_after_if------------------------------
1643 // Adjust the JVM state to reflect the result of taking this path.
1644 // Basically, it means inspecting the CmpNode controlling this
1645 // branch, seeing how it constrains a tested value, and then
1646 // deciding if it's worth our while to encode this constraint
1647 // as graph nodes in the current abstract interpretation map.
1648 void Parse::adjust_map_after_if(BoolTest::mask btest, Node* c, float prob, Block* path) {
1649 if (!c->is_Cmp()) {
1650 maybe_add_predicate_after_if(path);
1651 return;
1652 }
1653
1654 if (stopped() || btest == BoolTest::illegal) {
1655 return; // nothing to do
1656 }
1657
1658 bool is_fallthrough = (path == successor_for_bci(iter().next_bci()));
1659
1660 if (path_is_suitable_for_uncommon_trap(prob)) {
1661 repush_if_args();
1662 Node* call = uncommon_trap(Deoptimization::Reason_unstable_if,
1663 Deoptimization::Action_reinterpret,
1664 nullptr,
1665 (is_fallthrough ? "taken always" : "taken never"));
1666
1667 if (call != nullptr) {
1668 C->record_unstable_if_trap(new UnstableIfTrap(call->as_CallStaticJava(), path));
1669 }
1670 return;
1671 }
1672
1673 Node* val = c->in(1);
1674 Node* con = c->in(2);
1675 const Type* tcon = _gvn.type(con);
1676 const Type* tval = _gvn.type(val);
1677 bool have_con = tcon->singleton();
1678 if (tval->singleton()) {
1679 if (!have_con) {
1680 // Swap, so constant is in con.
1681 con = val;
1682 tcon = tval;
1683 val = c->in(2);
1684 tval = _gvn.type(val);
1685 btest = BoolTest(btest).commute();
1686 have_con = true;
1687 } else {
1688 // Do we have two constants? Then leave well enough alone.
1689 have_con = false;
1690 }
1691 }
1692 if (!have_con) { // remaining adjustments need a con
1693 maybe_add_predicate_after_if(path);
1694 return;
1695 }
1696
1697 sharpen_type_after_if(btest, con, tcon, val, tval);
1698 maybe_add_predicate_after_if(path);
1699 }
1700
1701
1702 static Node* extract_obj_from_klass_load(PhaseGVN* gvn, Node* n) {
1703 Node* ldk;
1704 if (n->is_DecodeNKlass()) {
1705 if (n->in(1)->Opcode() != Op_LoadNKlass) {
1706 return nullptr;
1707 } else {
1708 ldk = n->in(1);
1709 }
1710 } else if (n->Opcode() != Op_LoadKlass) {
1711 return nullptr;
1712 } else {
1713 ldk = n;
1714 }
1715 assert(ldk != nullptr && ldk->is_Load(), "should have found a LoadKlass or LoadNKlass node");
1716
1717 Node* adr = ldk->in(MemNode::Address);
1718 intptr_t off = 0;
1719 Node* obj = AddPNode::Ideal_base_and_offset(adr, gvn, off);
1720 if (obj == nullptr || off != oopDesc::klass_offset_in_bytes()) // loading oopDesc::_klass?
1721 return nullptr;
1722 const TypePtr* tp = gvn->type(obj)->is_ptr();
1723 if (tp == nullptr || !(tp->isa_instptr() || tp->isa_aryptr())) // is obj a Java object ptr?
1724 return nullptr;
1725
1726 return obj;
1727 }
1728
1729 // Matches exact and inexact type check IR shapes during parsing.
1730 // On successful match, returns type checked object node and its type after successful check
1731 // as out parameters.
1732 static bool match_type_check(PhaseGVN& gvn,
1733 BoolTest::mask btest,
1734 Node* con, const Type* tcon,
1735 Node* val, const Type* tval,
1736 Node** obj, const TypeOopPtr** cast_type) { // out-parameters
1737 // Look for opportunities to sharpen the type of a node whose klass is compared with a constant klass.
1738 // The constant klass being tested against can come from many bytecode instructions (implicitly or explicitly),
1739 // and also from profile data used by speculative casts.
1740 if (btest == BoolTest::eq && tcon->isa_klassptr()) {
1741 // Found:
1742 // Bool(CmpP(LoadKlass(obj._klass), ConP(Foo.klass)), [eq])
1743 // or the narrowOop equivalent.
1744 (*obj) = extract_obj_from_klass_load(&gvn, val);
1745 (*cast_type) = tcon->isa_klassptr()->as_instance_type();
1746 return true; // found
1747 }
1748
1749 // Match an instanceof check.
1750 // During parsing its IR shape is not canonicalized yet.
1751 //
1752 // obj superklass
1753 // | |
1754 // SubTypeCheck
1755 // |
1756 // Bool [eq] / [ne]
1757 // |
1758 // If
1759 // / \
1760 // T F
1761 // \ /
1762 // Region
1763 // \ ConI ConI
1764 // \ | /
1765 // val -> Phi ConI <- con
1766 // \ /
1767 // CmpI
1768 // |
1769 // Bool [btest]
1770 // |
1771 //
1772 if (tval->isa_int() && val->is_Phi() && val->in(0)->as_Region()->is_diamond()) {
1773 RegionNode* diamond = val->in(0)->as_Region();
1774 IfNode* if1 = diamond->in(1)->in(0)->as_If();
1775 BoolNode* b1 = if1->in(1)->isa_Bool();
1776 if (b1 != nullptr && b1->in(1)->isa_SubTypeCheck()) {
1777 assert(b1->_test._test == BoolTest::eq ||
1778 b1->_test._test == BoolTest::ne, "%d", b1->_test._test);
1779
1780 ProjNode* success_proj = if1->proj_out(b1->_test._test == BoolTest::eq ? 1 : 0);
1781 int idx = diamond->find_edge(success_proj);
1782 assert(idx == 1 || idx == 2, "");
1783 Node* vcon = val->in(idx);
1784
1785 assert(val->find_edge(con) > 0, "");
1786 if ((btest == BoolTest::eq && vcon == con) || (btest == BoolTest::ne && vcon != con)) {
1787 SubTypeCheckNode* sub = b1->in(1)->as_SubTypeCheck();
1788 Node* obj_or_subklass = sub->in(SubTypeCheckNode::ObjOrSubKlass);
1789 Node* superklass = sub->in(SubTypeCheckNode::SuperKlass);
1790
1791 if (gvn.type(obj_or_subklass)->isa_oopptr()) {
1792 const TypeKlassPtr* klass_ptr_type = gvn.type(superklass)->is_klassptr();
1793 const TypeKlassPtr* improved_klass_ptr_type = klass_ptr_type->try_improve();
1794
1795 (*obj) = obj_or_subklass;
1796 (*cast_type) = improved_klass_ptr_type->cast_to_exactness(false)->as_instance_type();
1797 return true; // found
1798 }
1799 }
1800 }
1801 }
1802 return false; // not found
1803 }
1804
1805 void Parse::sharpen_type_after_if(BoolTest::mask btest,
1806 Node* con, const Type* tcon,
1807 Node* val, const Type* tval) {
1808 Node* obj = nullptr;
1809 const TypeOopPtr* cast_type = nullptr;
1810 // Insert a cast node with a narrowed type after a successful type check.
1811 if (match_type_check(_gvn, btest, con, tcon, val, tval,
1812 &obj, &cast_type)) {
1813 assert(obj != nullptr && cast_type != nullptr, "missing type check info");
1814 const Type* obj_type = _gvn.type(obj);
1815 const TypeOopPtr* tboth = obj_type->join_speculative(cast_type)->isa_oopptr();
1816 if (tboth != nullptr && tboth != obj_type && tboth->higher_equal(obj_type)) {
1817 int obj_in_map = map()->find_edge(obj);
1818 JVMState* jvms = this->jvms();
1819 if (obj_in_map >= 0 &&
1820 (jvms->is_loc(obj_in_map) || jvms->is_stk(obj_in_map))) {
1821 TypeNode* ccast = new CheckCastPPNode(control(), obj, tboth);
1822 const Type* tcc = ccast->as_Type()->type();
1823 assert(tcc != obj_type && tcc->higher_equal(obj_type), "must improve");
1824 // Delay transform() call to allow recovery of pre-cast value
1825 // at the control merge.
1826 _gvn.set_type_bottom(ccast);
1827 record_for_igvn(ccast);
1828 // Here's the payoff.
1829 replace_in_map(obj, ccast);
1830 }
1831 }
1832 }
1833
1834 int val_in_map = map()->find_edge(val);
1835 if (val_in_map < 0) return; // replace_in_map would be useless
1836 {
1837 JVMState* jvms = this->jvms();
1838 if (!(jvms->is_loc(val_in_map) ||
1839 jvms->is_stk(val_in_map)))
1840 return; // again, it would be useless
1841 }
1842
1843 // Check for a comparison to a constant, and "know" that the compared
1844 // value is constrained on this path.
1845 assert(tcon->singleton(), "");
1846 ConstraintCastNode* ccast = nullptr;
1847 Node* cast = nullptr;
1848
1849 switch (btest) {
1850 case BoolTest::eq: // Constant test?
1851 {
1852 const Type* tboth = tcon->join_speculative(tval);
1853 if (tboth == tval) break; // Nothing to gain.
1854 if (tcon->isa_int()) {
1855 ccast = new CastIINode(control(), val, tboth);
1856 } else if (tcon == TypePtr::NULL_PTR) {
1857 // Cast to null, but keep the pointer identity temporarily live.
1858 ccast = new CastPPNode(control(), val, tboth);
1859 } else {
1860 const TypeF* tf = tcon->isa_float_constant();
1861 const TypeD* td = tcon->isa_double_constant();
1862 // Exclude tests vs float/double 0 as these could be
1863 // either +0 or -0. Just because you are equal to +0
1864 // doesn't mean you ARE +0!
1865 // Note, following code also replaces Long and Oop values.
1866 if ((!tf || tf->_f != 0.0) &&
1867 (!td || td->_d != 0.0))
1868 cast = con; // Replace non-constant val by con.
1869 }
1870 }
1871 break;
1872
1873 case BoolTest::ne:
1874 if (tcon == TypePtr::NULL_PTR) {
1875 cast = cast_not_null(val, false);
1876 }
1877 break;
1878
1879 default:
1880 // (At this point we could record int range types with CastII.)
1881 break;
1882 }
1883
1884 if (ccast != nullptr) {
1885 const Type* tcc = ccast->as_Type()->type();
1886 assert(tcc != tval && tcc->higher_equal(tval), "must improve");
1887 // Delay transform() call to allow recovery of pre-cast value
1888 // at the control merge.
1889 _gvn.set_type_bottom(ccast);
1890 record_for_igvn(ccast);
1891 cast = ccast;
1892 }
1893
1894 if (cast != nullptr) { // Here's the payoff.
1895 replace_in_map(val, cast);
1896 }
1897 }
1898
1899 /**
1900 * Use speculative type to optimize CmpP node: if comparison is
1901 * against the low level class, cast the object to the speculative
1902 * type if any. CmpP should then go away.
1903 *
1904 * @param c expected CmpP node
1905 * @return result of CmpP on object casted to speculative type
1906 *
1907 */
1908 Node* Parse::optimize_cmp_with_klass(Node* c) {
1909 // If this is transformed by the _gvn to a comparison with the low
1910 // level klass then we may be able to use speculation
1911 if (c->Opcode() == Op_CmpP &&
1912 (c->in(1)->Opcode() == Op_LoadKlass || c->in(1)->Opcode() == Op_DecodeNKlass) &&
1913 c->in(2)->is_Con()) {
1914 Node* load_klass = nullptr;
1915 Node* decode = nullptr;
1916 if (c->in(1)->Opcode() == Op_DecodeNKlass) {
1917 decode = c->in(1);
1918 load_klass = c->in(1)->in(1);
1919 } else {
1920 load_klass = c->in(1);
1921 }
1922 if (load_klass->in(2)->is_AddP()) {
1923 Node* addp = load_klass->in(2);
1924 Node* obj = addp->in(AddPNode::Address);
1925 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
1926 if (obj_type->speculative_type_not_null() != nullptr) {
1927 ciKlass* k = obj_type->speculative_type();
1928 inc_sp(2);
1929 obj = maybe_cast_profiled_obj(obj, k);
1930 dec_sp(2);
1931 // Make the CmpP use the casted obj
1932 addp = basic_plus_adr(obj, addp->in(AddPNode::Offset));
1933 load_klass = load_klass->clone();
1934 load_klass->set_req(2, addp);
1935 load_klass = _gvn.transform(load_klass);
1936 if (decode != nullptr) {
1937 decode = decode->clone();
1938 decode->set_req(1, load_klass);
1939 load_klass = _gvn.transform(decode);
1940 }
1941 c = c->clone();
1942 c->set_req(1, load_klass);
1943 c = _gvn.transform(c);
1944 }
1945 }
1946 }
1947 return c;
1948 }
1949
1950 //------------------------------do_one_bytecode--------------------------------
1951 // Parse this bytecode, and alter the Parsers JVM->Node mapping
1952 void Parse::do_one_bytecode() {
1953 Node *a, *b, *c, *d; // Handy temps
1954 BoolTest::mask btest;
1955 int i;
1956
1957 assert(!has_exceptions(), "bytecode entry state must be clear of throws");
1958
1959 if (C->check_node_count(NodeLimitFudgeFactor * 5,
1960 "out of nodes parsing method")) {
1961 return;
1962 }
1963
1964 #ifdef ASSERT
1965 // for setting breakpoints
1966 if (TraceOptoParse) {
1967 tty->print(" @");
1968 dump_bci(bci());
1969 tty->print(" %s", Bytecodes::name(bc()));
1970 tty->cr();
1971 }
1972 #endif
1973
1974 switch (bc()) {
1975 case Bytecodes::_nop:
1976 // do nothing
1977 break;
1978 case Bytecodes::_lconst_0:
1979 push_pair(longcon(0));
1980 break;
1981
1982 case Bytecodes::_lconst_1:
1983 push_pair(longcon(1));
1984 break;
1985
1986 case Bytecodes::_fconst_0:
1987 push(zerocon(T_FLOAT));
1988 break;
1989
1990 case Bytecodes::_fconst_1:
1991 push(makecon(TypeF::ONE));
1992 break;
1993
1994 case Bytecodes::_fconst_2:
1995 push(makecon(TypeF::make(2.0f)));
1996 break;
1997
1998 case Bytecodes::_dconst_0:
1999 push_pair(zerocon(T_DOUBLE));
2000 break;
2001
2002 case Bytecodes::_dconst_1:
2003 push_pair(makecon(TypeD::ONE));
2004 break;
2005
2006 case Bytecodes::_iconst_m1:push(intcon(-1)); break;
2007 case Bytecodes::_iconst_0: push(intcon( 0)); break;
2008 case Bytecodes::_iconst_1: push(intcon( 1)); break;
2009 case Bytecodes::_iconst_2: push(intcon( 2)); break;
2010 case Bytecodes::_iconst_3: push(intcon( 3)); break;
2011 case Bytecodes::_iconst_4: push(intcon( 4)); break;
2012 case Bytecodes::_iconst_5: push(intcon( 5)); break;
2013 case Bytecodes::_bipush: push(intcon(iter().get_constant_u1())); break;
2014 case Bytecodes::_sipush: push(intcon(iter().get_constant_u2())); break;
2015 case Bytecodes::_aconst_null: push(null()); break;
2016
2017 case Bytecodes::_ldc:
2018 case Bytecodes::_ldc_w:
2019 case Bytecodes::_ldc2_w: {
2020 // ciTypeFlow should trap if the ldc is in error state or if the constant is not loaded
2021 assert(!iter().is_in_error(), "ldc is in error state");
2022 ciConstant constant = iter().get_constant();
2023 assert(constant.is_loaded(), "constant is not loaded");
2024 const Type* con_type = Type::make_from_constant(constant);
2025 if (con_type != nullptr) {
2026 push_node(con_type->basic_type(), makecon(con_type));
2027 }
2028 break;
2029 }
2030
2031 case Bytecodes::_aload_0:
2032 push( local(0) );
2033 break;
2034 case Bytecodes::_aload_1:
2035 push( local(1) );
2036 break;
2037 case Bytecodes::_aload_2:
2038 push( local(2) );
2039 break;
2040 case Bytecodes::_aload_3:
2041 push( local(3) );
2042 break;
2043 case Bytecodes::_aload:
2044 push( local(iter().get_index()) );
2045 break;
2046
2047 case Bytecodes::_fload_0:
2048 case Bytecodes::_iload_0:
2049 push( local(0) );
2050 break;
2051 case Bytecodes::_fload_1:
2052 case Bytecodes::_iload_1:
2053 push( local(1) );
2054 break;
2055 case Bytecodes::_fload_2:
2056 case Bytecodes::_iload_2:
2057 push( local(2) );
2058 break;
2059 case Bytecodes::_fload_3:
2060 case Bytecodes::_iload_3:
2061 push( local(3) );
2062 break;
2063 case Bytecodes::_fload:
2064 case Bytecodes::_iload:
2065 push( local(iter().get_index()) );
2066 break;
2067 case Bytecodes::_lload_0:
2068 push_pair_local( 0 );
2069 break;
2070 case Bytecodes::_lload_1:
2071 push_pair_local( 1 );
2072 break;
2073 case Bytecodes::_lload_2:
2074 push_pair_local( 2 );
2075 break;
2076 case Bytecodes::_lload_3:
2077 push_pair_local( 3 );
2078 break;
2079 case Bytecodes::_lload:
2080 push_pair_local( iter().get_index() );
2081 break;
2082
2083 case Bytecodes::_dload_0:
2084 push_pair_local(0);
2085 break;
2086 case Bytecodes::_dload_1:
2087 push_pair_local(1);
2088 break;
2089 case Bytecodes::_dload_2:
2090 push_pair_local(2);
2091 break;
2092 case Bytecodes::_dload_3:
2093 push_pair_local(3);
2094 break;
2095 case Bytecodes::_dload:
2096 push_pair_local(iter().get_index());
2097 break;
2098 case Bytecodes::_fstore_0:
2099 case Bytecodes::_istore_0:
2100 case Bytecodes::_astore_0:
2101 set_local( 0, pop() );
2102 break;
2103 case Bytecodes::_fstore_1:
2104 case Bytecodes::_istore_1:
2105 case Bytecodes::_astore_1:
2106 set_local( 1, pop() );
2107 break;
2108 case Bytecodes::_fstore_2:
2109 case Bytecodes::_istore_2:
2110 case Bytecodes::_astore_2:
2111 set_local( 2, pop() );
2112 break;
2113 case Bytecodes::_fstore_3:
2114 case Bytecodes::_istore_3:
2115 case Bytecodes::_astore_3:
2116 set_local( 3, pop() );
2117 break;
2118 case Bytecodes::_fstore:
2119 case Bytecodes::_istore:
2120 case Bytecodes::_astore:
2121 set_local( iter().get_index(), pop() );
2122 break;
2123 // long stores
2124 case Bytecodes::_lstore_0:
2125 set_pair_local( 0, pop_pair() );
2126 break;
2127 case Bytecodes::_lstore_1:
2128 set_pair_local( 1, pop_pair() );
2129 break;
2130 case Bytecodes::_lstore_2:
2131 set_pair_local( 2, pop_pair() );
2132 break;
2133 case Bytecodes::_lstore_3:
2134 set_pair_local( 3, pop_pair() );
2135 break;
2136 case Bytecodes::_lstore:
2137 set_pair_local( iter().get_index(), pop_pair() );
2138 break;
2139
2140 // double stores
2141 case Bytecodes::_dstore_0:
2142 set_pair_local( 0, pop_pair() );
2143 break;
2144 case Bytecodes::_dstore_1:
2145 set_pair_local( 1, pop_pair() );
2146 break;
2147 case Bytecodes::_dstore_2:
2148 set_pair_local( 2, pop_pair() );
2149 break;
2150 case Bytecodes::_dstore_3:
2151 set_pair_local( 3, pop_pair() );
2152 break;
2153 case Bytecodes::_dstore:
2154 set_pair_local( iter().get_index(), pop_pair() );
2155 break;
2156
2157 case Bytecodes::_pop: dec_sp(1); break;
2158 case Bytecodes::_pop2: dec_sp(2); break;
2159 case Bytecodes::_swap:
2160 a = pop();
2161 b = pop();
2162 push(a);
2163 push(b);
2164 break;
2165 case Bytecodes::_dup:
2166 a = pop();
2167 push(a);
2168 push(a);
2169 break;
2170 case Bytecodes::_dup_x1:
2171 a = pop();
2172 b = pop();
2173 push( a );
2174 push( b );
2175 push( a );
2176 break;
2177 case Bytecodes::_dup_x2:
2178 a = pop();
2179 b = pop();
2180 c = pop();
2181 push( a );
2182 push( c );
2183 push( b );
2184 push( a );
2185 break;
2186 case Bytecodes::_dup2:
2187 a = pop();
2188 b = pop();
2189 push( b );
2190 push( a );
2191 push( b );
2192 push( a );
2193 break;
2194
2195 case Bytecodes::_dup2_x1:
2196 // before: .. c, b, a
2197 // after: .. b, a, c, b, a
2198 // not tested
2199 a = pop();
2200 b = pop();
2201 c = pop();
2202 push( b );
2203 push( a );
2204 push( c );
2205 push( b );
2206 push( a );
2207 break;
2208 case Bytecodes::_dup2_x2:
2209 // before: .. d, c, b, a
2210 // after: .. b, a, d, c, b, a
2211 // not tested
2212 a = pop();
2213 b = pop();
2214 c = pop();
2215 d = pop();
2216 push( b );
2217 push( a );
2218 push( d );
2219 push( c );
2220 push( b );
2221 push( a );
2222 break;
2223
2224 case Bytecodes::_arraylength: {
2225 // Must do null-check with value on expression stack
2226 Node *ary = null_check(peek(), T_ARRAY);
2227 // Compile-time detect of null-exception?
2228 if (stopped()) return;
2229 a = pop();
2230 push(load_array_length(a));
2231 break;
2232 }
2233
2234 case Bytecodes::_baload: array_load(T_BYTE); break;
2235 case Bytecodes::_caload: array_load(T_CHAR); break;
2236 case Bytecodes::_iaload: array_load(T_INT); break;
2237 case Bytecodes::_saload: array_load(T_SHORT); break;
2238 case Bytecodes::_faload: array_load(T_FLOAT); break;
2239 case Bytecodes::_aaload: array_load(T_OBJECT); break;
2240 case Bytecodes::_laload: array_load(T_LONG); break;
2241 case Bytecodes::_daload: array_load(T_DOUBLE); break;
2242 case Bytecodes::_bastore: array_store(T_BYTE); break;
2243 case Bytecodes::_castore: array_store(T_CHAR); break;
2244 case Bytecodes::_iastore: array_store(T_INT); break;
2245 case Bytecodes::_sastore: array_store(T_SHORT); break;
2246 case Bytecodes::_fastore: array_store(T_FLOAT); break;
2247 case Bytecodes::_aastore: array_store(T_OBJECT); break;
2248 case Bytecodes::_lastore: array_store(T_LONG); break;
2249 case Bytecodes::_dastore: array_store(T_DOUBLE); break;
2250
2251 case Bytecodes::_getfield:
2252 do_getfield();
2253 break;
2254
2255 case Bytecodes::_getstatic:
2256 do_getstatic();
2257 break;
2258
2259 case Bytecodes::_putfield:
2260 do_putfield();
2261 break;
2262
2263 case Bytecodes::_putstatic:
2264 do_putstatic();
2265 break;
2266
2267 case Bytecodes::_irem:
2268 // Must keep both values on the expression-stack during null-check
2269 zero_check_int(peek());
2270 // Compile-time detect of null-exception?
2271 if (stopped()) return;
2272 b = pop();
2273 a = pop();
2274 push(_gvn.transform(new ModINode(control(), a, b)));
2275 break;
2276 case Bytecodes::_idiv:
2277 // Must keep both values on the expression-stack during null-check
2278 zero_check_int(peek());
2279 // Compile-time detect of null-exception?
2280 if (stopped()) return;
2281 b = pop();
2282 a = pop();
2283 push( _gvn.transform( new DivINode(control(),a,b) ) );
2284 break;
2285 case Bytecodes::_imul:
2286 b = pop(); a = pop();
2287 push( _gvn.transform( new MulINode(a,b) ) );
2288 break;
2289 case Bytecodes::_iadd:
2290 b = pop(); a = pop();
2291 push( _gvn.transform( new AddINode(a,b) ) );
2292 break;
2293 case Bytecodes::_ineg:
2294 a = pop();
2295 push( _gvn.transform( new SubINode(_gvn.intcon(0),a)) );
2296 break;
2297 case Bytecodes::_isub:
2298 b = pop(); a = pop();
2299 push( _gvn.transform( new SubINode(a,b) ) );
2300 break;
2301 case Bytecodes::_iand:
2302 b = pop(); a = pop();
2303 push( _gvn.transform( new AndINode(a,b) ) );
2304 break;
2305 case Bytecodes::_ior:
2306 b = pop(); a = pop();
2307 push( _gvn.transform( new OrINode(a,b) ) );
2308 break;
2309 case Bytecodes::_ixor:
2310 b = pop(); a = pop();
2311 push( _gvn.transform( new XorINode(a,b) ) );
2312 break;
2313 case Bytecodes::_ishl:
2314 b = pop(); a = pop();
2315 push( _gvn.transform( new LShiftINode(a,b) ) );
2316 break;
2317 case Bytecodes::_ishr:
2318 b = pop(); a = pop();
2319 push( _gvn.transform( new RShiftINode(a,b) ) );
2320 break;
2321 case Bytecodes::_iushr:
2322 b = pop(); a = pop();
2323 push( _gvn.transform( new URShiftINode(a,b) ) );
2324 break;
2325
2326 case Bytecodes::_fneg:
2327 a = pop();
2328 b = _gvn.transform(new NegFNode (a));
2329 push(b);
2330 break;
2331
2332 case Bytecodes::_fsub:
2333 b = pop();
2334 a = pop();
2335 c = _gvn.transform( new SubFNode(a,b) );
2336 push(c);
2337 break;
2338
2339 case Bytecodes::_fadd:
2340 b = pop();
2341 a = pop();
2342 c = _gvn.transform( new AddFNode(a,b) );
2343 push(c);
2344 break;
2345
2346 case Bytecodes::_fmul:
2347 b = pop();
2348 a = pop();
2349 c = _gvn.transform( new MulFNode(a,b) );
2350 push(c);
2351 break;
2352
2353 case Bytecodes::_fdiv:
2354 b = pop();
2355 a = pop();
2356 c = _gvn.transform( new DivFNode(nullptr,a,b) );
2357 push(c);
2358 break;
2359
2360 case Bytecodes::_frem:
2361 // Generate a ModF node.
2362 b = pop();
2363 a = pop();
2364 push(floating_point_mod(a, b, BasicType::T_FLOAT));
2365 break;
2366
2367 case Bytecodes::_fcmpl:
2368 b = pop();
2369 a = pop();
2370 c = _gvn.transform( new CmpF3Node( a, b));
2371 push(c);
2372 break;
2373 case Bytecodes::_fcmpg:
2374 b = pop();
2375 a = pop();
2376
2377 // Same as fcmpl but need to flip the unordered case. Swap the inputs,
2378 // which negates the result sign except for unordered. Flip the unordered
2379 // as well by using CmpF3 which implements unordered-lesser instead of
2380 // unordered-greater semantics. Finally, commute the result bits. Result
2381 // is same as using a CmpF3Greater except we did it with CmpF3 alone.
2382 c = _gvn.transform( new CmpF3Node( b, a));
2383 c = _gvn.transform( new SubINode(_gvn.intcon(0),c) );
2384 push(c);
2385 break;
2386
2387 case Bytecodes::_f2i:
2388 a = pop();
2389 push(_gvn.transform(new ConvF2INode(a)));
2390 break;
2391
2392 case Bytecodes::_d2i:
2393 a = pop_pair();
2394 b = _gvn.transform(new ConvD2INode(a));
2395 push( b );
2396 break;
2397
2398 case Bytecodes::_f2d:
2399 a = pop();
2400 b = _gvn.transform( new ConvF2DNode(a));
2401 push_pair( b );
2402 break;
2403
2404 case Bytecodes::_d2f:
2405 a = pop_pair();
2406 b = _gvn.transform( new ConvD2FNode(a));
2407 push( b );
2408 break;
2409
2410 case Bytecodes::_l2f:
2411 if (Matcher::convL2FSupported()) {
2412 a = pop_pair();
2413 b = _gvn.transform( new ConvL2FNode(a));
2414 push(b);
2415 } else {
2416 l2f();
2417 }
2418 break;
2419
2420 case Bytecodes::_l2d:
2421 a = pop_pair();
2422 b = _gvn.transform( new ConvL2DNode(a));
2423 push_pair(b);
2424 break;
2425
2426 case Bytecodes::_f2l:
2427 a = pop();
2428 b = _gvn.transform( new ConvF2LNode(a));
2429 push_pair(b);
2430 break;
2431
2432 case Bytecodes::_d2l:
2433 a = pop_pair();
2434 b = _gvn.transform( new ConvD2LNode(a));
2435 push_pair(b);
2436 break;
2437
2438 case Bytecodes::_dsub:
2439 b = pop_pair();
2440 a = pop_pair();
2441 c = _gvn.transform( new SubDNode(a,b) );
2442 push_pair(c);
2443 break;
2444
2445 case Bytecodes::_dadd:
2446 b = pop_pair();
2447 a = pop_pair();
2448 c = _gvn.transform( new AddDNode(a,b) );
2449 push_pair(c);
2450 break;
2451
2452 case Bytecodes::_dmul:
2453 b = pop_pair();
2454 a = pop_pair();
2455 c = _gvn.transform( new MulDNode(a,b) );
2456 push_pair(c);
2457 break;
2458
2459 case Bytecodes::_ddiv:
2460 b = pop_pair();
2461 a = pop_pair();
2462 c = _gvn.transform( new DivDNode(nullptr,a,b) );
2463 push_pair(c);
2464 break;
2465
2466 case Bytecodes::_dneg:
2467 a = pop_pair();
2468 b = _gvn.transform(new NegDNode (a));
2469 push_pair(b);
2470 break;
2471
2472 case Bytecodes::_drem:
2473 // Generate a ModD node.
2474 b = pop_pair();
2475 a = pop_pair();
2476 push_pair(floating_point_mod(a, b, BasicType::T_DOUBLE));
2477 break;
2478
2479 case Bytecodes::_dcmpl:
2480 b = pop_pair();
2481 a = pop_pair();
2482 c = _gvn.transform( new CmpD3Node( a, b));
2483 push(c);
2484 break;
2485
2486 case Bytecodes::_dcmpg:
2487 b = pop_pair();
2488 a = pop_pair();
2489 // Same as dcmpl but need to flip the unordered case.
2490 // Commute the inputs, which negates the result sign except for unordered.
2491 // Flip the unordered as well by using CmpD3 which implements
2492 // unordered-lesser instead of unordered-greater semantics.
2493 // Finally, negate the result bits. Result is same as using a
2494 // CmpD3Greater except we did it with CmpD3 alone.
2495 c = _gvn.transform( new CmpD3Node( b, a));
2496 c = _gvn.transform( new SubINode(_gvn.intcon(0),c) );
2497 push(c);
2498 break;
2499
2500
2501 // Note for longs -> lo word is on TOS, hi word is on TOS - 1
2502 case Bytecodes::_land:
2503 b = pop_pair();
2504 a = pop_pair();
2505 c = _gvn.transform( new AndLNode(a,b) );
2506 push_pair(c);
2507 break;
2508 case Bytecodes::_lor:
2509 b = pop_pair();
2510 a = pop_pair();
2511 c = _gvn.transform( new OrLNode(a,b) );
2512 push_pair(c);
2513 break;
2514 case Bytecodes::_lxor:
2515 b = pop_pair();
2516 a = pop_pair();
2517 c = _gvn.transform( new XorLNode(a,b) );
2518 push_pair(c);
2519 break;
2520
2521 case Bytecodes::_lshl:
2522 b = pop(); // the shift count
2523 a = pop_pair(); // value to be shifted
2524 c = _gvn.transform( new LShiftLNode(a,b) );
2525 push_pair(c);
2526 break;
2527 case Bytecodes::_lshr:
2528 b = pop(); // the shift count
2529 a = pop_pair(); // value to be shifted
2530 c = _gvn.transform( new RShiftLNode(a,b) );
2531 push_pair(c);
2532 break;
2533 case Bytecodes::_lushr:
2534 b = pop(); // the shift count
2535 a = pop_pair(); // value to be shifted
2536 c = _gvn.transform( new URShiftLNode(a,b) );
2537 push_pair(c);
2538 break;
2539 case Bytecodes::_lmul:
2540 b = pop_pair();
2541 a = pop_pair();
2542 c = _gvn.transform( new MulLNode(a,b) );
2543 push_pair(c);
2544 break;
2545
2546 case Bytecodes::_lrem:
2547 // Must keep both values on the expression-stack during null-check
2548 assert(peek(0) == top(), "long word order");
2549 zero_check_long(peek(1));
2550 // Compile-time detect of null-exception?
2551 if (stopped()) return;
2552 b = pop_pair();
2553 a = pop_pair();
2554 c = _gvn.transform( new ModLNode(control(),a,b) );
2555 push_pair(c);
2556 break;
2557
2558 case Bytecodes::_ldiv:
2559 // Must keep both values on the expression-stack during null-check
2560 assert(peek(0) == top(), "long word order");
2561 zero_check_long(peek(1));
2562 // Compile-time detect of null-exception?
2563 if (stopped()) return;
2564 b = pop_pair();
2565 a = pop_pair();
2566 c = _gvn.transform( new DivLNode(control(),a,b) );
2567 push_pair(c);
2568 break;
2569
2570 case Bytecodes::_ladd:
2571 b = pop_pair();
2572 a = pop_pair();
2573 c = _gvn.transform( new AddLNode(a,b) );
2574 push_pair(c);
2575 break;
2576 case Bytecodes::_lsub:
2577 b = pop_pair();
2578 a = pop_pair();
2579 c = _gvn.transform( new SubLNode(a,b) );
2580 push_pair(c);
2581 break;
2582 case Bytecodes::_lcmp:
2583 // Safepoints are now inserted _before_ branches. The long-compare
2584 // bytecode painfully produces a 3-way value (-1,0,+1) which requires a
2585 // slew of control flow. These are usually followed by a CmpI vs zero and
2586 // a branch; this pattern then optimizes to the obvious long-compare and
2587 // branch. However, if the branch is backwards there's a Safepoint
2588 // inserted. The inserted Safepoint captures the JVM state at the
2589 // pre-branch point, i.e. it captures the 3-way value. Thus if a
2590 // long-compare is used to control a loop the debug info will force
2591 // computation of the 3-way value, even though the generated code uses a
2592 // long-compare and branch. We try to rectify the situation by inserting
2593 // a SafePoint here and have it dominate and kill the safepoint added at a
2594 // following backwards branch. At this point the JVM state merely holds 2
2595 // longs but not the 3-way value.
2596 switch (iter().next_bc()) {
2597 case Bytecodes::_ifgt:
2598 case Bytecodes::_iflt:
2599 case Bytecodes::_ifge:
2600 case Bytecodes::_ifle:
2601 case Bytecodes::_ifne:
2602 case Bytecodes::_ifeq:
2603 // If this is a backwards branch in the bytecodes, add Safepoint
2604 maybe_add_safepoint(iter().next_get_dest());
2605 default:
2606 break;
2607 }
2608 b = pop_pair();
2609 a = pop_pair();
2610 c = _gvn.transform( new CmpL3Node( a, b ));
2611 push(c);
2612 break;
2613
2614 case Bytecodes::_lneg:
2615 a = pop_pair();
2616 b = _gvn.transform( new SubLNode(longcon(0),a));
2617 push_pair(b);
2618 break;
2619 case Bytecodes::_l2i:
2620 a = pop_pair();
2621 push( _gvn.transform( new ConvL2INode(a)));
2622 break;
2623 case Bytecodes::_i2l:
2624 a = pop();
2625 b = _gvn.transform( new ConvI2LNode(a));
2626 push_pair(b);
2627 break;
2628 case Bytecodes::_i2b:
2629 // Sign extend
2630 a = pop();
2631 a = Compile::narrow_value(T_BYTE, a, nullptr, &_gvn, true);
2632 push(a);
2633 break;
2634 case Bytecodes::_i2s:
2635 a = pop();
2636 a = Compile::narrow_value(T_SHORT, a, nullptr, &_gvn, true);
2637 push(a);
2638 break;
2639 case Bytecodes::_i2c:
2640 a = pop();
2641 a = Compile::narrow_value(T_CHAR, a, nullptr, &_gvn, true);
2642 push(a);
2643 break;
2644
2645 case Bytecodes::_i2f:
2646 a = pop();
2647 b = _gvn.transform( new ConvI2FNode(a) ) ;
2648 push(b);
2649 break;
2650
2651 case Bytecodes::_i2d:
2652 a = pop();
2653 b = _gvn.transform( new ConvI2DNode(a));
2654 push_pair(b);
2655 break;
2656
2657 case Bytecodes::_iinc: // Increment local
2658 i = iter().get_index(); // Get local index
2659 set_local( i, _gvn.transform( new AddINode( _gvn.intcon(iter().get_iinc_con()), local(i) ) ) );
2660 break;
2661
2662 // Exit points of synchronized methods must have an unlock node
2663 case Bytecodes::_return:
2664 return_current(nullptr);
2665 break;
2666
2667 case Bytecodes::_ireturn:
2668 case Bytecodes::_areturn:
2669 case Bytecodes::_freturn:
2670 return_current(pop());
2671 break;
2672 case Bytecodes::_lreturn:
2673 return_current(pop_pair());
2674 break;
2675 case Bytecodes::_dreturn:
2676 return_current(pop_pair());
2677 break;
2678
2679 case Bytecodes::_athrow:
2680 // null exception oop throws null pointer exception
2681 null_check(peek());
2682 if (stopped()) return;
2683 // Hook the thrown exception directly to subsequent handlers.
2684 if (BailoutToInterpreterForThrows) {
2685 // Keep method interpreted from now on.
2686 uncommon_trap(Deoptimization::Reason_unhandled,
2687 Deoptimization::Action_make_not_compilable);
2688 return;
2689 }
2690 if (env()->jvmti_can_post_on_exceptions()) {
2691 // check if we must post exception events, take uncommon trap if so (with must_throw = false)
2692 uncommon_trap_if_should_post_on_exceptions(Deoptimization::Reason_unhandled, false);
2693 }
2694 // Here if either can_post_on_exceptions or should_post_on_exceptions is false
2695 add_exception_state(make_exception_state(peek()));
2696 break;
2697
2698 case Bytecodes::_goto: // fall through
2699 case Bytecodes::_goto_w: {
2700 int target_bci = (bc() == Bytecodes::_goto) ? iter().get_dest() : iter().get_far_dest();
2701
2702 // If this is a backwards branch in the bytecodes, add Safepoint
2703 maybe_add_safepoint(target_bci);
2704
2705 // Merge the current control into the target basic block
2706 merge(target_bci);
2707
2708 // See if we can get some profile data and hand it off to the next block
2709 Block *target_block = block()->successor_for_bci(target_bci);
2710 if (target_block->pred_count() != 1) break;
2711 ciMethodData* methodData = method()->method_data();
2712 if (!methodData->is_mature()) break;
2713 ciProfileData* data = methodData->bci_to_data(bci());
2714 assert(data != nullptr && data->is_JumpData(), "need JumpData for taken branch");
2715 int taken = ((ciJumpData*)data)->taken();
2716 taken = method()->scale_count(taken);
2717 target_block->set_count(taken);
2718 break;
2719 }
2720
2721 case Bytecodes::_ifnull: btest = BoolTest::eq; goto handle_if_null;
2722 case Bytecodes::_ifnonnull: btest = BoolTest::ne; goto handle_if_null;
2723 handle_if_null:
2724 // If this is a backwards branch in the bytecodes, add Safepoint
2725 maybe_add_safepoint(iter().get_dest());
2726 a = null();
2727 b = pop();
2728 if (!_gvn.type(b)->speculative_maybe_null() &&
2729 !too_many_traps(Deoptimization::Reason_speculate_null_check)) {
2730 inc_sp(1);
2731 Node* null_ctl = top();
2732 b = null_check_oop(b, &null_ctl, true, true, true);
2733 assert(null_ctl->is_top(), "no null control here");
2734 dec_sp(1);
2735 } else if (_gvn.type(b)->speculative_always_null() &&
2736 !too_many_traps(Deoptimization::Reason_speculate_null_assert)) {
2737 inc_sp(1);
2738 b = null_assert(b);
2739 dec_sp(1);
2740 }
2741 c = _gvn.transform( new CmpPNode(b, a) );
2742 do_ifnull(btest, c);
2743 break;
2744
2745 case Bytecodes::_if_acmpeq: btest = BoolTest::eq; goto handle_if_acmp;
2746 case Bytecodes::_if_acmpne: btest = BoolTest::ne; goto handle_if_acmp;
2747 handle_if_acmp:
2748 // If this is a backwards branch in the bytecodes, add Safepoint
2749 maybe_add_safepoint(iter().get_dest());
2750 a = pop();
2751 b = pop();
2752 c = _gvn.transform( new CmpPNode(b, a) );
2753 c = optimize_cmp_with_klass(c);
2754 do_if(btest, c);
2755 break;
2756
2757 case Bytecodes::_ifeq: btest = BoolTest::eq; goto handle_ifxx;
2758 case Bytecodes::_ifne: btest = BoolTest::ne; goto handle_ifxx;
2759 case Bytecodes::_iflt: btest = BoolTest::lt; goto handle_ifxx;
2760 case Bytecodes::_ifle: btest = BoolTest::le; goto handle_ifxx;
2761 case Bytecodes::_ifgt: btest = BoolTest::gt; goto handle_ifxx;
2762 case Bytecodes::_ifge: btest = BoolTest::ge; goto handle_ifxx;
2763 handle_ifxx:
2764 // If this is a backwards branch in the bytecodes, add Safepoint
2765 maybe_add_safepoint(iter().get_dest());
2766 a = _gvn.intcon(0);
2767 b = pop();
2768 c = _gvn.transform( new CmpINode(b, a) );
2769 do_if(btest, c);
2770 break;
2771
2772 case Bytecodes::_if_icmpeq: btest = BoolTest::eq; goto handle_if_icmp;
2773 case Bytecodes::_if_icmpne: btest = BoolTest::ne; goto handle_if_icmp;
2774 case Bytecodes::_if_icmplt: btest = BoolTest::lt; goto handle_if_icmp;
2775 case Bytecodes::_if_icmple: btest = BoolTest::le; goto handle_if_icmp;
2776 case Bytecodes::_if_icmpgt: btest = BoolTest::gt; goto handle_if_icmp;
2777 case Bytecodes::_if_icmpge: btest = BoolTest::ge; goto handle_if_icmp;
2778 handle_if_icmp:
2779 // If this is a backwards branch in the bytecodes, add Safepoint
2780 maybe_add_safepoint(iter().get_dest());
2781 a = pop();
2782 b = pop();
2783 c = _gvn.transform( new CmpINode( b, a ) );
2784 do_if(btest, c);
2785 break;
2786
2787 case Bytecodes::_tableswitch:
2788 do_tableswitch();
2789 break;
2790
2791 case Bytecodes::_lookupswitch:
2792 do_lookupswitch();
2793 break;
2794
2795 case Bytecodes::_invokestatic:
2796 case Bytecodes::_invokedynamic:
2797 case Bytecodes::_invokespecial:
2798 case Bytecodes::_invokevirtual:
2799 case Bytecodes::_invokeinterface:
2800 do_call();
2801 break;
2802 case Bytecodes::_checkcast:
2803 do_checkcast();
2804 break;
2805 case Bytecodes::_instanceof:
2806 do_instanceof();
2807 break;
2808 case Bytecodes::_anewarray:
2809 do_anewarray();
2810 break;
2811 case Bytecodes::_newarray:
2812 do_newarray((BasicType)iter().get_index());
2813 break;
2814 case Bytecodes::_multianewarray:
2815 do_multianewarray();
2816 break;
2817 case Bytecodes::_new:
2818 do_new();
2819 break;
2820
2821 case Bytecodes::_jsr:
2822 case Bytecodes::_jsr_w:
2823 do_jsr();
2824 break;
2825
2826 case Bytecodes::_ret:
2827 do_ret();
2828 break;
2829
2830
2831 case Bytecodes::_monitorenter:
2832 do_monitor_enter();
2833 break;
2834
2835 case Bytecodes::_monitorexit:
2836 do_monitor_exit();
2837 break;
2838
2839 case Bytecodes::_breakpoint:
2840 // Breakpoint set concurrently to compile
2841 // %%% use an uncommon trap?
2842 C->record_failure("breakpoint in method");
2843 return;
2844
2845 default:
2846 #ifndef PRODUCT
2847 map()->dump(99);
2848 #endif
2849 tty->print("\nUnhandled bytecode %s\n", Bytecodes::name(bc()) );
2850 ShouldNotReachHere();
2851 }
2852
2853 #ifndef PRODUCT
2854 if (failing()) { return; }
2855 constexpr int perBytecode = 6;
2856 if (C->should_print_igv(perBytecode)) {
2857 IdealGraphPrinter* printer = C->igv_printer();
2858 char buffer[256];
2859 jio_snprintf(buffer, sizeof(buffer), "Bytecode %d: %s", bci(), Bytecodes::name(bc()));
2860 bool old = printer->traverse_outs();
2861 printer->set_traverse_outs(true);
2862 printer->set_parse(this);
2863 printer->print_graph(buffer);
2864 printer->set_traverse_outs(old);
2865 printer->set_parse(nullptr);
2866 }
2867 #endif
2868 }