1 /*
2 * Copyright (c) 1998, 2024, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
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 "precompiled.hpp"
26 #include "ci/ciMethodData.hpp"
27 #include "classfile/vmSymbols.hpp"
28 #include "compiler/compileLog.hpp"
29 #include "interpreter/linkResolver.hpp"
30 #include "jvm_io.h"
31 #include "memory/resourceArea.hpp"
32 #include "memory/universe.hpp"
33 #include "oops/oop.inline.hpp"
34 #include "opto/addnode.hpp"
35 #include "opto/castnode.hpp"
36 #include "opto/convertnode.hpp"
37 #include "opto/divnode.hpp"
38 #include "opto/idealGraphPrinter.hpp"
39 #include "opto/matcher.hpp"
40 #include "opto/memnode.hpp"
41 #include "opto/mulnode.hpp"
42 #include "opto/opaquenode.hpp"
43 #include "opto/parse.hpp"
44 #include "opto/runtime.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 void Parse::modf() {
1100 Node *f2 = pop();
1101 Node *f1 = pop();
1102 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::modf_Type(),
1103 CAST_FROM_FN_PTR(address, SharedRuntime::frem),
1104 "frem", nullptr, //no memory effects
1105 f1, f2);
1106 Node* res = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0));
1107
1108 push(res);
1109 }
1110
1111 void Parse::modd() {
1112 Node *d2 = pop_pair();
1113 Node *d1 = pop_pair();
1114 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::Math_DD_D_Type(),
1115 CAST_FROM_FN_PTR(address, SharedRuntime::drem),
1116 "drem", nullptr, //no memory effects
1117 d1, top(), d2, top());
1118 Node* res_d = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0));
1119
1120 #ifdef ASSERT
1121 Node* res_top = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 1));
1122 assert(res_top == top(), "second value must be top");
1123 #endif
1124
1125 push_pair(res_d);
1126 }
1127
1128 void Parse::l2f() {
1129 Node* f2 = pop();
1130 Node* f1 = pop();
1131 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::l2f_Type(),
1132 CAST_FROM_FN_PTR(address, SharedRuntime::l2f),
1133 "l2f", nullptr, //no memory effects
1134 f1, f2);
1135 Node* res = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0));
1136
1137 push(res);
1138 }
1139
1140 // Handle jsr and jsr_w bytecode
1141 void Parse::do_jsr() {
1142 assert(bc() == Bytecodes::_jsr || bc() == Bytecodes::_jsr_w, "wrong bytecode");
1143
1144 // Store information about current state, tagged with new _jsr_bci
1145 int return_bci = iter().next_bci();
1146 int jsr_bci = (bc() == Bytecodes::_jsr) ? iter().get_dest() : iter().get_far_dest();
1147
1148 // The way we do things now, there is only one successor block
1149 // for the jsr, because the target code is cloned by ciTypeFlow.
1150 Block* target = successor_for_bci(jsr_bci);
1151
1152 // What got pushed?
1153 const Type* ret_addr = target->peek();
1154 assert(ret_addr->singleton(), "must be a constant (cloned jsr body)");
1155
1156 // Effect on jsr on stack
1157 push(_gvn.makecon(ret_addr));
1158
1159 // Flow to the jsr.
1160 merge(jsr_bci);
1161 }
1162
1163 // Handle ret bytecode
1164 void Parse::do_ret() {
1165 // Find to whom we return.
1166 assert(block()->num_successors() == 1, "a ret can only go one place now");
1167 Block* target = block()->successor_at(0);
1168 assert(!target->is_ready(), "our arrival must be expected");
1169 int pnum = target->next_path_num();
1170 merge_common(target, pnum);
1171 }
1172
1173 static bool has_injected_profile(BoolTest::mask btest, Node* test, int& taken, int& not_taken) {
1174 if (btest != BoolTest::eq && btest != BoolTest::ne) {
1175 // Only ::eq and ::ne are supported for profile injection.
1176 return false;
1177 }
1178 if (test->is_Cmp() &&
1179 test->in(1)->Opcode() == Op_ProfileBoolean) {
1180 ProfileBooleanNode* profile = (ProfileBooleanNode*)test->in(1);
1181 int false_cnt = profile->false_count();
1182 int true_cnt = profile->true_count();
1183
1184 // Counts matching depends on the actual test operation (::eq or ::ne).
1185 // No need to scale the counts because profile injection was designed
1186 // to feed exact counts into VM.
1187 taken = (btest == BoolTest::eq) ? false_cnt : true_cnt;
1188 not_taken = (btest == BoolTest::eq) ? true_cnt : false_cnt;
1189
1190 profile->consume();
1191 return true;
1192 }
1193 return false;
1194 }
1195
1196 // Give up if too few (or too many, in which case the sum will overflow) counts to be meaningful.
1197 // We also check that individual counters are positive first, otherwise the sum can become positive.
1198 // (check for saturation, integer overflow, and immature counts)
1199 static bool counters_are_meaningful(int counter1, int counter2, int min) {
1200 // check for saturation, including "uint" values too big to fit in "int"
1201 if (counter1 < 0 || counter2 < 0) {
1202 return false;
1203 }
1204 // check for integer overflow of the sum
1205 int64_t sum = (int64_t)counter1 + (int64_t)counter2;
1206 STATIC_ASSERT(sizeof(counter1) < sizeof(sum));
1207 if (sum > INT_MAX) {
1208 return false;
1209 }
1210 // check if mature
1211 return (counter1 + counter2) >= min;
1212 }
1213
1214 //--------------------------dynamic_branch_prediction--------------------------
1215 // Try to gather dynamic branch prediction behavior. Return a probability
1216 // of the branch being taken and set the "cnt" field. Returns a -1.0
1217 // if we need to use static prediction for some reason.
1218 float Parse::dynamic_branch_prediction(float &cnt, BoolTest::mask btest, Node* test) {
1219 ResourceMark rm;
1220
1221 cnt = COUNT_UNKNOWN;
1222
1223 int taken = 0;
1224 int not_taken = 0;
1225
1226 bool use_mdo = !has_injected_profile(btest, test, taken, not_taken);
1227
1228 if (use_mdo) {
1229 // Use MethodData information if it is available
1230 // FIXME: free the ProfileData structure
1231 ciMethodData* methodData = method()->method_data();
1232 if (!methodData->is_mature()) return PROB_UNKNOWN;
1233 ciProfileData* data = methodData->bci_to_data(bci());
1234 if (data == nullptr) {
1235 return PROB_UNKNOWN;
1236 }
1237 if (!data->is_JumpData()) return PROB_UNKNOWN;
1238
1239 // get taken and not taken values
1240 // NOTE: saturated UINT_MAX values become negative,
1241 // as do counts above INT_MAX.
1242 taken = data->as_JumpData()->taken();
1243 not_taken = 0;
1244 if (data->is_BranchData()) {
1245 not_taken = data->as_BranchData()->not_taken();
1246 }
1247
1248 // scale the counts to be commensurate with invocation counts:
1249 // NOTE: overflow for positive values is clamped at INT_MAX
1250 taken = method()->scale_count(taken);
1251 not_taken = method()->scale_count(not_taken);
1252 }
1253 // At this point, saturation or overflow is indicated by INT_MAX
1254 // or a negative value.
1255
1256 // Give up if too few (or too many, in which case the sum will overflow) counts to be meaningful.
1257 // We also check that individual counters are positive first, otherwise the sum can become positive.
1258 if (!counters_are_meaningful(taken, not_taken, 40)) {
1259 if (C->log() != nullptr) {
1260 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d'", iter().get_dest(), taken, not_taken);
1261 }
1262 return PROB_UNKNOWN;
1263 }
1264
1265 // Compute frequency that we arrive here
1266 float sum = taken + not_taken;
1267 // Adjust, if this block is a cloned private block but the
1268 // Jump counts are shared. Taken the private counts for
1269 // just this path instead of the shared counts.
1270 if( block()->count() > 0 )
1271 sum = block()->count();
1272 cnt = sum / FreqCountInvocations;
1273
1274 // Pin probability to sane limits
1275 float prob;
1276 if( !taken )
1277 prob = (0+PROB_MIN) / 2;
1278 else if( !not_taken )
1279 prob = (1+PROB_MAX) / 2;
1280 else { // Compute probability of true path
1281 prob = (float)taken / (float)(taken + not_taken);
1282 if (prob > PROB_MAX) prob = PROB_MAX;
1283 if (prob < PROB_MIN) prob = PROB_MIN;
1284 }
1285
1286 assert((cnt > 0.0f) && (prob > 0.0f),
1287 "Bad frequency assignment in if cnt=%g prob=%g taken=%d not_taken=%d", cnt, prob, taken, not_taken);
1288
1289 if (C->log() != nullptr) {
1290 const char* prob_str = nullptr;
1291 if (prob >= PROB_MAX) prob_str = (prob == PROB_MAX) ? "max" : "always";
1292 if (prob <= PROB_MIN) prob_str = (prob == PROB_MIN) ? "min" : "never";
1293 char prob_str_buf[30];
1294 if (prob_str == nullptr) {
1295 jio_snprintf(prob_str_buf, sizeof(prob_str_buf), "%20.2f", prob);
1296 prob_str = prob_str_buf;
1297 // The %20.2f adds many spaces to the string, to avoid some
1298 // picky overflow warning as noted in 8211929. But, 20 is the
1299 // *minimum* width, not *maximum*, so it's not clear how this
1300 // helps prevent overflow. Looks like we were forced to work
1301 // around a bug in gcc. In any case, strip the blanks.
1302 while (*prob_str == ' ') ++prob_str;
1303 }
1304 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d' cnt='%f' prob='%s'",
1305 iter().get_dest(), taken, not_taken, cnt, prob_str);
1306 }
1307 return prob;
1308 }
1309
1310 //-----------------------------branch_prediction-------------------------------
1311 float Parse::branch_prediction(float& cnt,
1312 BoolTest::mask btest,
1313 int target_bci,
1314 Node* test) {
1315 float prob = dynamic_branch_prediction(cnt, btest, test);
1316 // If prob is unknown, switch to static prediction
1317 if (prob != PROB_UNKNOWN) return prob;
1318
1319 prob = PROB_FAIR; // Set default value
1320 if (btest == BoolTest::eq) // Exactly equal test?
1321 prob = PROB_STATIC_INFREQUENT; // Assume its relatively infrequent
1322 else if (btest == BoolTest::ne)
1323 prob = PROB_STATIC_FREQUENT; // Assume its relatively frequent
1324
1325 // If this is a conditional test guarding a backwards branch,
1326 // assume its a loop-back edge. Make it a likely taken branch.
1327 if (target_bci < bci()) {
1328 if (is_osr_parse()) { // Could be a hot OSR'd loop; force deopt
1329 // Since it's an OSR, we probably have profile data, but since
1330 // branch_prediction returned PROB_UNKNOWN, the counts are too small.
1331 // Let's make a special check here for completely zero counts.
1332 ciMethodData* methodData = method()->method_data();
1333 if (!methodData->is_empty()) {
1334 ciProfileData* data = methodData->bci_to_data(bci());
1335 // Only stop for truly zero counts, which mean an unknown part
1336 // of the OSR-ed method, and we want to deopt to gather more stats.
1337 // If you have ANY counts, then this loop is simply 'cold' relative
1338 // to the OSR loop.
1339 if (data == nullptr ||
1340 (data->as_BranchData()->taken() + data->as_BranchData()->not_taken() == 0)) {
1341 // This is the only way to return PROB_UNKNOWN:
1342 return PROB_UNKNOWN;
1343 }
1344 }
1345 }
1346 prob = PROB_STATIC_FREQUENT; // Likely to take backwards branch
1347 }
1348
1349 assert(prob != PROB_UNKNOWN, "must have some guess at this point");
1350 return prob;
1351 }
1352
1353 // The magic constants are chosen so as to match the output of
1354 // branch_prediction() when the profile reports a zero taken count.
1355 // It is important to distinguish zero counts unambiguously, because
1356 // some branches (e.g., _213_javac.Assembler.eliminate) validly produce
1357 // very small but nonzero probabilities, which if confused with zero
1358 // counts would keep the program recompiling indefinitely.
1359 bool Parse::seems_never_taken(float prob) const {
1360 return prob < PROB_MIN;
1361 }
1362
1363 //-------------------------------repush_if_args--------------------------------
1364 // Push arguments of an "if" bytecode back onto the stack by adjusting _sp.
1365 inline int Parse::repush_if_args() {
1366 if (PrintOpto && WizardMode) {
1367 tty->print("defending against excessive implicit null exceptions on %s @%d in ",
1368 Bytecodes::name(iter().cur_bc()), iter().cur_bci());
1369 method()->print_name(); tty->cr();
1370 }
1371 int bc_depth = - Bytecodes::depth(iter().cur_bc());
1372 assert(bc_depth == 1 || bc_depth == 2, "only two kinds of branches");
1373 DEBUG_ONLY(sync_jvms()); // argument(n) requires a synced jvms
1374 assert(argument(0) != nullptr, "must exist");
1375 assert(bc_depth == 1 || argument(1) != nullptr, "two must exist");
1376 inc_sp(bc_depth);
1377 return bc_depth;
1378 }
1379
1380 // Used by StressUnstableIfTraps
1381 static volatile int _trap_stress_counter = 0;
1382
1383 void Parse::increment_trap_stress_counter(Node*& counter, Node*& incr_store) {
1384 Node* counter_addr = makecon(TypeRawPtr::make((address)&_trap_stress_counter));
1385 counter = make_load(control(), counter_addr, TypeInt::INT, T_INT, MemNode::unordered);
1386 counter = _gvn.transform(new AddINode(counter, intcon(1)));
1387 incr_store = store_to_memory(control(), counter_addr, counter, T_INT, MemNode::unordered);
1388 }
1389
1390 //----------------------------------do_ifnull----------------------------------
1391 void Parse::do_ifnull(BoolTest::mask btest, Node *c) {
1392 int target_bci = iter().get_dest();
1393
1394 Node* counter = nullptr;
1395 Node* incr_store = nullptr;
1396 bool do_stress_trap = StressUnstableIfTraps && ((C->random() % 2) == 0);
1397 if (do_stress_trap) {
1398 increment_trap_stress_counter(counter, incr_store);
1399 }
1400
1401 Block* branch_block = successor_for_bci(target_bci);
1402 Block* next_block = successor_for_bci(iter().next_bci());
1403
1404 float cnt;
1405 float prob = branch_prediction(cnt, btest, target_bci, c);
1406 if (prob == PROB_UNKNOWN) {
1407 // (An earlier version of do_ifnull omitted this trap for OSR methods.)
1408 if (PrintOpto && Verbose) {
1409 tty->print_cr("Never-taken edge stops compilation at bci %d", bci());
1410 }
1411 repush_if_args(); // to gather stats on loop
1412 uncommon_trap(Deoptimization::Reason_unreached,
1413 Deoptimization::Action_reinterpret,
1414 nullptr, "cold");
1415 if (C->eliminate_boxing()) {
1416 // Mark the successor blocks as parsed
1417 branch_block->next_path_num();
1418 next_block->next_path_num();
1419 }
1420 return;
1421 }
1422
1423 NOT_PRODUCT(explicit_null_checks_inserted++);
1424
1425 // Generate real control flow
1426 Node *tst = _gvn.transform( new BoolNode( c, btest ) );
1427
1428 // Sanity check the probability value
1429 assert(prob > 0.0f,"Bad probability in Parser");
1430 // Need xform to put node in hash table
1431 IfNode *iff = create_and_xform_if( control(), tst, prob, cnt );
1432 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser");
1433 // True branch
1434 { PreserveJVMState pjvms(this);
1435 Node* iftrue = _gvn.transform( new IfTrueNode (iff) );
1436 set_control(iftrue);
1437
1438 if (stopped()) { // Path is dead?
1439 NOT_PRODUCT(explicit_null_checks_elided++);
1440 if (C->eliminate_boxing()) {
1441 // Mark the successor block as parsed
1442 branch_block->next_path_num();
1443 }
1444 } else { // Path is live.
1445 adjust_map_after_if(btest, c, prob, branch_block);
1446 if (!stopped()) {
1447 merge(target_bci);
1448 }
1449 }
1450 }
1451
1452 // False branch
1453 Node* iffalse = _gvn.transform( new IfFalseNode(iff) );
1454 set_control(iffalse);
1455
1456 if (stopped()) { // Path is dead?
1457 NOT_PRODUCT(explicit_null_checks_elided++);
1458 if (C->eliminate_boxing()) {
1459 // Mark the successor block as parsed
1460 next_block->next_path_num();
1461 }
1462 } else { // Path is live.
1463 adjust_map_after_if(BoolTest(btest).negate(), c, 1.0-prob, next_block);
1464 }
1465
1466 if (do_stress_trap) {
1467 stress_trap(iff, counter, incr_store);
1468 }
1469 }
1470
1471 //------------------------------------do_if------------------------------------
1472 void Parse::do_if(BoolTest::mask btest, Node* c) {
1473 int target_bci = iter().get_dest();
1474
1475 Block* branch_block = successor_for_bci(target_bci);
1476 Block* next_block = successor_for_bci(iter().next_bci());
1477
1478 float cnt;
1479 float prob = branch_prediction(cnt, btest, target_bci, c);
1480 float untaken_prob = 1.0 - prob;
1481
1482 if (prob == PROB_UNKNOWN) {
1483 if (PrintOpto && Verbose) {
1484 tty->print_cr("Never-taken edge stops compilation at bci %d", bci());
1485 }
1486 repush_if_args(); // to gather stats on loop
1487 uncommon_trap(Deoptimization::Reason_unreached,
1488 Deoptimization::Action_reinterpret,
1489 nullptr, "cold");
1490 if (C->eliminate_boxing()) {
1491 // Mark the successor blocks as parsed
1492 branch_block->next_path_num();
1493 next_block->next_path_num();
1494 }
1495 return;
1496 }
1497
1498 Node* counter = nullptr;
1499 Node* incr_store = nullptr;
1500 bool do_stress_trap = StressUnstableIfTraps && ((C->random() % 2) == 0);
1501 if (do_stress_trap) {
1502 increment_trap_stress_counter(counter, incr_store);
1503 }
1504
1505 // Sanity check the probability value
1506 assert(0.0f < prob && prob < 1.0f,"Bad probability in Parser");
1507
1508 bool taken_if_true = true;
1509 // Convert BoolTest to canonical form:
1510 if (!BoolTest(btest).is_canonical()) {
1511 btest = BoolTest(btest).negate();
1512 taken_if_true = false;
1513 // prob is NOT updated here; it remains the probability of the taken
1514 // path (as opposed to the prob of the path guarded by an 'IfTrueNode').
1515 }
1516 assert(btest != BoolTest::eq, "!= is the only canonical exact test");
1517
1518 Node* tst0 = new BoolNode(c, btest);
1519 Node* tst = _gvn.transform(tst0);
1520 BoolTest::mask taken_btest = BoolTest::illegal;
1521 BoolTest::mask untaken_btest = BoolTest::illegal;
1522
1523 if (tst->is_Bool()) {
1524 // Refresh c from the transformed bool node, since it may be
1525 // simpler than the original c. Also re-canonicalize btest.
1526 // This wins when (Bool ne (Conv2B p) 0) => (Bool ne (CmpP p null)).
1527 // That can arise from statements like: if (x instanceof C) ...
1528 if (tst != tst0) {
1529 // Canonicalize one more time since transform can change it.
1530 btest = tst->as_Bool()->_test._test;
1531 if (!BoolTest(btest).is_canonical()) {
1532 // Reverse edges one more time...
1533 tst = _gvn.transform( tst->as_Bool()->negate(&_gvn) );
1534 btest = tst->as_Bool()->_test._test;
1535 assert(BoolTest(btest).is_canonical(), "sanity");
1536 taken_if_true = !taken_if_true;
1537 }
1538 c = tst->in(1);
1539 }
1540 BoolTest::mask neg_btest = BoolTest(btest).negate();
1541 taken_btest = taken_if_true ? btest : neg_btest;
1542 untaken_btest = taken_if_true ? neg_btest : btest;
1543 }
1544
1545 // Generate real control flow
1546 float true_prob = (taken_if_true ? prob : untaken_prob);
1547 IfNode* iff = create_and_map_if(control(), tst, true_prob, cnt);
1548 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser");
1549 Node* taken_branch = new IfTrueNode(iff);
1550 Node* untaken_branch = new IfFalseNode(iff);
1551 if (!taken_if_true) { // Finish conversion to canonical form
1552 Node* tmp = taken_branch;
1553 taken_branch = untaken_branch;
1554 untaken_branch = tmp;
1555 }
1556
1557 // Branch is taken:
1558 { PreserveJVMState pjvms(this);
1559 taken_branch = _gvn.transform(taken_branch);
1560 set_control(taken_branch);
1561
1562 if (stopped()) {
1563 if (C->eliminate_boxing()) {
1564 // Mark the successor block as parsed
1565 branch_block->next_path_num();
1566 }
1567 } else {
1568 adjust_map_after_if(taken_btest, c, prob, branch_block);
1569 if (!stopped()) {
1570 merge(target_bci);
1571 }
1572 }
1573 }
1574
1575 untaken_branch = _gvn.transform(untaken_branch);
1576 set_control(untaken_branch);
1577
1578 // Branch not taken.
1579 if (stopped()) {
1580 if (C->eliminate_boxing()) {
1581 // Mark the successor block as parsed
1582 next_block->next_path_num();
1583 }
1584 } else {
1585 adjust_map_after_if(untaken_btest, c, untaken_prob, next_block);
1586 }
1587
1588 if (do_stress_trap) {
1589 stress_trap(iff, counter, incr_store);
1590 }
1591 }
1592
1593 // Force unstable if traps to be taken randomly to trigger intermittent bugs such as incorrect debug information.
1594 // Add another if before the unstable if that checks a "random" condition at runtime (a simple shared counter) and
1595 // then either takes the trap or executes the original, unstable if.
1596 void Parse::stress_trap(IfNode* orig_iff, Node* counter, Node* incr_store) {
1597 // Search for an unstable if trap
1598 CallStaticJavaNode* trap = nullptr;
1599 assert(orig_iff->Opcode() == Op_If && orig_iff->outcnt() == 2, "malformed if");
1600 ProjNode* trap_proj = orig_iff->uncommon_trap_proj(trap, Deoptimization::Reason_unstable_if);
1601 if (trap == nullptr || !trap->jvms()->should_reexecute()) {
1602 // No suitable trap found. Remove unused counter load and increment.
1603 C->gvn_replace_by(incr_store, incr_store->in(MemNode::Memory));
1604 return;
1605 }
1606
1607 // Remove trap from optimization list since we add another path to the trap.
1608 bool success = C->remove_unstable_if_trap(trap, true);
1609 assert(success, "Trap already modified");
1610
1611 // Add a check before the original if that will trap with a certain frequency and execute the original if otherwise
1612 int freq_log = (C->random() % 31) + 1; // Random logarithmic frequency in [1, 31]
1613 Node* mask = intcon(right_n_bits(freq_log));
1614 counter = _gvn.transform(new AndINode(counter, mask));
1615 Node* cmp = _gvn.transform(new CmpINode(counter, intcon(0)));
1616 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::mask::eq));
1617 IfNode* iff = _gvn.transform(new IfNode(orig_iff->in(0), bol, orig_iff->_prob, orig_iff->_fcnt))->as_If();
1618 Node* if_true = _gvn.transform(new IfTrueNode(iff));
1619 Node* if_false = _gvn.transform(new IfFalseNode(iff));
1620 assert(!if_true->is_top() && !if_false->is_top(), "trap always / never taken");
1621
1622 // Trap
1623 assert(trap_proj->outcnt() == 1, "some other nodes are dependent on the trap projection");
1624
1625 Node* trap_region = new RegionNode(3);
1626 trap_region->set_req(1, trap_proj);
1627 trap_region->set_req(2, if_true);
1628 trap->set_req(0, _gvn.transform(trap_region));
1629
1630 // Don't trap, execute original if
1631 orig_iff->set_req(0, if_false);
1632 }
1633
1634 bool Parse::path_is_suitable_for_uncommon_trap(float prob) const {
1635 // Randomly skip emitting an uncommon trap
1636 if (StressUnstableIfTraps && ((C->random() % 2) == 0)) {
1637 return false;
1638 }
1639 // Don't want to speculate on uncommon traps when running with -Xcomp
1640 if (!UseInterpreter) {
1641 return false;
1642 }
1643 return seems_never_taken(prob) &&
1644 !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if);
1645 }
1646
1647 void Parse::maybe_add_predicate_after_if(Block* path) {
1648 if (path->is_SEL_head() && path->preds_parsed() == 0) {
1649 // Add predicates at bci of if dominating the loop so traps can be
1650 // recorded on the if's profile data
1651 int bc_depth = repush_if_args();
1652 add_parse_predicates();
1653 dec_sp(bc_depth);
1654 path->set_has_predicates();
1655 }
1656 }
1657
1658
1659 //----------------------------adjust_map_after_if------------------------------
1660 // Adjust the JVM state to reflect the result of taking this path.
1661 // Basically, it means inspecting the CmpNode controlling this
1662 // branch, seeing how it constrains a tested value, and then
1663 // deciding if it's worth our while to encode this constraint
1664 // as graph nodes in the current abstract interpretation map.
1665 void Parse::adjust_map_after_if(BoolTest::mask btest, Node* c, float prob, Block* path) {
1666 if (!c->is_Cmp()) {
1667 maybe_add_predicate_after_if(path);
1668 return;
1669 }
1670
1671 if (stopped() || btest == BoolTest::illegal) {
1672 return; // nothing to do
1673 }
1674
1675 bool is_fallthrough = (path == successor_for_bci(iter().next_bci()));
1676
1677 if (path_is_suitable_for_uncommon_trap(prob)) {
1678 repush_if_args();
1679 Node* call = uncommon_trap(Deoptimization::Reason_unstable_if,
1680 Deoptimization::Action_reinterpret,
1681 nullptr,
1682 (is_fallthrough ? "taken always" : "taken never"));
1683
1684 if (call != nullptr) {
1685 C->record_unstable_if_trap(new UnstableIfTrap(call->as_CallStaticJava(), path));
1686 }
1687 return;
1688 }
1689
1690 Node* val = c->in(1);
1691 Node* con = c->in(2);
1692 const Type* tcon = _gvn.type(con);
1693 const Type* tval = _gvn.type(val);
1694 bool have_con = tcon->singleton();
1695 if (tval->singleton()) {
1696 if (!have_con) {
1697 // Swap, so constant is in con.
1698 con = val;
1699 tcon = tval;
1700 val = c->in(2);
1701 tval = _gvn.type(val);
1702 btest = BoolTest(btest).commute();
1703 have_con = true;
1704 } else {
1705 // Do we have two constants? Then leave well enough alone.
1706 have_con = false;
1707 }
1708 }
1709 if (!have_con) { // remaining adjustments need a con
1710 maybe_add_predicate_after_if(path);
1711 return;
1712 }
1713
1714 sharpen_type_after_if(btest, con, tcon, val, tval);
1715 maybe_add_predicate_after_if(path);
1716 }
1717
1718
1719 static Node* extract_obj_from_klass_load(PhaseGVN* gvn, Node* n) {
1720 Node* ldk;
1721 if (n->is_DecodeNKlass()) {
1722 if (n->in(1)->Opcode() != Op_LoadNKlass) {
1723 return nullptr;
1724 } else {
1725 ldk = n->in(1);
1726 }
1727 } else if (n->Opcode() != Op_LoadKlass) {
1728 return nullptr;
1729 } else {
1730 ldk = n;
1731 }
1732 assert(ldk != nullptr && ldk->is_Load(), "should have found a LoadKlass or LoadNKlass node");
1733
1734 Node* adr = ldk->in(MemNode::Address);
1735 intptr_t off = 0;
1736 Node* obj = AddPNode::Ideal_base_and_offset(adr, gvn, off);
1737 if (obj == nullptr || off != oopDesc::klass_offset_in_bytes()) // loading oopDesc::_klass?
1738 return nullptr;
1739 const TypePtr* tp = gvn->type(obj)->is_ptr();
1740 if (tp == nullptr || !(tp->isa_instptr() || tp->isa_aryptr())) // is obj a Java object ptr?
1741 return nullptr;
1742
1743 return obj;
1744 }
1745
1746 void Parse::sharpen_type_after_if(BoolTest::mask btest,
1747 Node* con, const Type* tcon,
1748 Node* val, const Type* tval) {
1749 // Look for opportunities to sharpen the type of a node
1750 // whose klass is compared with a constant klass.
1751 if (btest == BoolTest::eq && tcon->isa_klassptr()) {
1752 Node* obj = extract_obj_from_klass_load(&_gvn, val);
1753 const TypeOopPtr* con_type = tcon->isa_klassptr()->as_instance_type();
1754 if (obj != nullptr && (con_type->isa_instptr() || con_type->isa_aryptr())) {
1755 // Found:
1756 // Bool(CmpP(LoadKlass(obj._klass), ConP(Foo.klass)), [eq])
1757 // or the narrowOop equivalent.
1758 const Type* obj_type = _gvn.type(obj);
1759 const TypeOopPtr* tboth = obj_type->join_speculative(con_type)->isa_oopptr();
1760 if (tboth != nullptr && tboth->klass_is_exact() && tboth != obj_type &&
1761 tboth->higher_equal(obj_type)) {
1762 // obj has to be of the exact type Foo if the CmpP succeeds.
1763 int obj_in_map = map()->find_edge(obj);
1764 JVMState* jvms = this->jvms();
1765 if (obj_in_map >= 0 &&
1766 (jvms->is_loc(obj_in_map) || jvms->is_stk(obj_in_map))) {
1767 TypeNode* ccast = new CheckCastPPNode(control(), obj, tboth);
1768 const Type* tcc = ccast->as_Type()->type();
1769 assert(tcc != obj_type && tcc->higher_equal(obj_type), "must improve");
1770 // Delay transform() call to allow recovery of pre-cast value
1771 // at the control merge.
1772 _gvn.set_type_bottom(ccast);
1773 record_for_igvn(ccast);
1774 // Here's the payoff.
1775 replace_in_map(obj, ccast);
1776 }
1777 }
1778 }
1779 }
1780
1781 int val_in_map = map()->find_edge(val);
1782 if (val_in_map < 0) return; // replace_in_map would be useless
1783 {
1784 JVMState* jvms = this->jvms();
1785 if (!(jvms->is_loc(val_in_map) ||
1786 jvms->is_stk(val_in_map)))
1787 return; // again, it would be useless
1788 }
1789
1790 // Check for a comparison to a constant, and "know" that the compared
1791 // value is constrained on this path.
1792 assert(tcon->singleton(), "");
1793 ConstraintCastNode* ccast = nullptr;
1794 Node* cast = nullptr;
1795
1796 switch (btest) {
1797 case BoolTest::eq: // Constant test?
1798 {
1799 const Type* tboth = tcon->join_speculative(tval);
1800 if (tboth == tval) break; // Nothing to gain.
1801 if (tcon->isa_int()) {
1802 ccast = new CastIINode(control(), val, tboth);
1803 } else if (tcon == TypePtr::NULL_PTR) {
1804 // Cast to null, but keep the pointer identity temporarily live.
1805 ccast = new CastPPNode(control(), val, tboth);
1806 } else {
1807 const TypeF* tf = tcon->isa_float_constant();
1808 const TypeD* td = tcon->isa_double_constant();
1809 // Exclude tests vs float/double 0 as these could be
1810 // either +0 or -0. Just because you are equal to +0
1811 // doesn't mean you ARE +0!
1812 // Note, following code also replaces Long and Oop values.
1813 if ((!tf || tf->_f != 0.0) &&
1814 (!td || td->_d != 0.0))
1815 cast = con; // Replace non-constant val by con.
1816 }
1817 }
1818 break;
1819
1820 case BoolTest::ne:
1821 if (tcon == TypePtr::NULL_PTR) {
1822 cast = cast_not_null(val, false);
1823 }
1824 break;
1825
1826 default:
1827 // (At this point we could record int range types with CastII.)
1828 break;
1829 }
1830
1831 if (ccast != nullptr) {
1832 const Type* tcc = ccast->as_Type()->type();
1833 assert(tcc != tval && tcc->higher_equal(tval), "must improve");
1834 // Delay transform() call to allow recovery of pre-cast value
1835 // at the control merge.
1836 _gvn.set_type_bottom(ccast);
1837 record_for_igvn(ccast);
1838 cast = ccast;
1839 }
1840
1841 if (cast != nullptr) { // Here's the payoff.
1842 replace_in_map(val, cast);
1843 }
1844 }
1845
1846 /**
1847 * Use speculative type to optimize CmpP node: if comparison is
1848 * against the low level class, cast the object to the speculative
1849 * type if any. CmpP should then go away.
1850 *
1851 * @param c expected CmpP node
1852 * @return result of CmpP on object casted to speculative type
1853 *
1854 */
1855 Node* Parse::optimize_cmp_with_klass(Node* c) {
1856 // If this is transformed by the _gvn to a comparison with the low
1857 // level klass then we may be able to use speculation
1858 if (c->Opcode() == Op_CmpP &&
1859 (c->in(1)->Opcode() == Op_LoadKlass || c->in(1)->Opcode() == Op_DecodeNKlass) &&
1860 c->in(2)->is_Con()) {
1861 Node* load_klass = nullptr;
1862 Node* decode = nullptr;
1863 if (c->in(1)->Opcode() == Op_DecodeNKlass) {
1864 decode = c->in(1);
1865 load_klass = c->in(1)->in(1);
1866 } else {
1867 load_klass = c->in(1);
1868 }
1869 if (load_klass->in(2)->is_AddP()) {
1870 Node* addp = load_klass->in(2);
1871 Node* obj = addp->in(AddPNode::Address);
1872 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
1873 if (obj_type->speculative_type_not_null() != nullptr) {
1874 ciKlass* k = obj_type->speculative_type();
1875 inc_sp(2);
1876 obj = maybe_cast_profiled_obj(obj, k);
1877 dec_sp(2);
1878 // Make the CmpP use the casted obj
1879 addp = basic_plus_adr(obj, addp->in(AddPNode::Offset));
1880 load_klass = load_klass->clone();
1881 load_klass->set_req(2, addp);
1882 load_klass = _gvn.transform(load_klass);
1883 if (decode != nullptr) {
1884 decode = decode->clone();
1885 decode->set_req(1, load_klass);
1886 load_klass = _gvn.transform(decode);
1887 }
1888 c = c->clone();
1889 c->set_req(1, load_klass);
1890 c = _gvn.transform(c);
1891 }
1892 }
1893 }
1894 return c;
1895 }
1896
1897 //------------------------------do_one_bytecode--------------------------------
1898 // Parse this bytecode, and alter the Parsers JVM->Node mapping
1899 void Parse::do_one_bytecode() {
1900 Node *a, *b, *c, *d; // Handy temps
1901 BoolTest::mask btest;
1902 int i;
1903
1904 assert(!has_exceptions(), "bytecode entry state must be clear of throws");
1905
1906 if (C->check_node_count(NodeLimitFudgeFactor * 5,
1907 "out of nodes parsing method")) {
1908 return;
1909 }
1910
1911 #ifdef ASSERT
1912 // for setting breakpoints
1913 if (TraceOptoParse) {
1914 tty->print(" @");
1915 dump_bci(bci());
1916 tty->print(" %s", Bytecodes::name(bc()));
1917 tty->cr();
1918 }
1919 #endif
1920
1921 switch (bc()) {
1922 case Bytecodes::_nop:
1923 // do nothing
1924 break;
1925 case Bytecodes::_lconst_0:
1926 push_pair(longcon(0));
1927 break;
1928
1929 case Bytecodes::_lconst_1:
1930 push_pair(longcon(1));
1931 break;
1932
1933 case Bytecodes::_fconst_0:
1934 push(zerocon(T_FLOAT));
1935 break;
1936
1937 case Bytecodes::_fconst_1:
1938 push(makecon(TypeF::ONE));
1939 break;
1940
1941 case Bytecodes::_fconst_2:
1942 push(makecon(TypeF::make(2.0f)));
1943 break;
1944
1945 case Bytecodes::_dconst_0:
1946 push_pair(zerocon(T_DOUBLE));
1947 break;
1948
1949 case Bytecodes::_dconst_1:
1950 push_pair(makecon(TypeD::ONE));
1951 break;
1952
1953 case Bytecodes::_iconst_m1:push(intcon(-1)); break;
1954 case Bytecodes::_iconst_0: push(intcon( 0)); break;
1955 case Bytecodes::_iconst_1: push(intcon( 1)); break;
1956 case Bytecodes::_iconst_2: push(intcon( 2)); break;
1957 case Bytecodes::_iconst_3: push(intcon( 3)); break;
1958 case Bytecodes::_iconst_4: push(intcon( 4)); break;
1959 case Bytecodes::_iconst_5: push(intcon( 5)); break;
1960 case Bytecodes::_bipush: push(intcon(iter().get_constant_u1())); break;
1961 case Bytecodes::_sipush: push(intcon(iter().get_constant_u2())); break;
1962 case Bytecodes::_aconst_null: push(null()); break;
1963
1964 case Bytecodes::_ldc:
1965 case Bytecodes::_ldc_w:
1966 case Bytecodes::_ldc2_w: {
1967 // ciTypeFlow should trap if the ldc is in error state or if the constant is not loaded
1968 assert(!iter().is_in_error(), "ldc is in error state");
1969 ciConstant constant = iter().get_constant();
1970 assert(constant.is_loaded(), "constant is not loaded");
1971 const Type* con_type = Type::make_from_constant(constant);
1972 if (con_type != nullptr) {
1973 push_node(con_type->basic_type(), makecon(con_type));
1974 }
1975 break;
1976 }
1977
1978 case Bytecodes::_aload_0:
1979 push( local(0) );
1980 break;
1981 case Bytecodes::_aload_1:
1982 push( local(1) );
1983 break;
1984 case Bytecodes::_aload_2:
1985 push( local(2) );
1986 break;
1987 case Bytecodes::_aload_3:
1988 push( local(3) );
1989 break;
1990 case Bytecodes::_aload:
1991 push( local(iter().get_index()) );
1992 break;
1993
1994 case Bytecodes::_fload_0:
1995 case Bytecodes::_iload_0:
1996 push( local(0) );
1997 break;
1998 case Bytecodes::_fload_1:
1999 case Bytecodes::_iload_1:
2000 push( local(1) );
2001 break;
2002 case Bytecodes::_fload_2:
2003 case Bytecodes::_iload_2:
2004 push( local(2) );
2005 break;
2006 case Bytecodes::_fload_3:
2007 case Bytecodes::_iload_3:
2008 push( local(3) );
2009 break;
2010 case Bytecodes::_fload:
2011 case Bytecodes::_iload:
2012 push( local(iter().get_index()) );
2013 break;
2014 case Bytecodes::_lload_0:
2015 push_pair_local( 0 );
2016 break;
2017 case Bytecodes::_lload_1:
2018 push_pair_local( 1 );
2019 break;
2020 case Bytecodes::_lload_2:
2021 push_pair_local( 2 );
2022 break;
2023 case Bytecodes::_lload_3:
2024 push_pair_local( 3 );
2025 break;
2026 case Bytecodes::_lload:
2027 push_pair_local( iter().get_index() );
2028 break;
2029
2030 case Bytecodes::_dload_0:
2031 push_pair_local(0);
2032 break;
2033 case Bytecodes::_dload_1:
2034 push_pair_local(1);
2035 break;
2036 case Bytecodes::_dload_2:
2037 push_pair_local(2);
2038 break;
2039 case Bytecodes::_dload_3:
2040 push_pair_local(3);
2041 break;
2042 case Bytecodes::_dload:
2043 push_pair_local(iter().get_index());
2044 break;
2045 case Bytecodes::_fstore_0:
2046 case Bytecodes::_istore_0:
2047 case Bytecodes::_astore_0:
2048 set_local( 0, pop() );
2049 break;
2050 case Bytecodes::_fstore_1:
2051 case Bytecodes::_istore_1:
2052 case Bytecodes::_astore_1:
2053 set_local( 1, pop() );
2054 break;
2055 case Bytecodes::_fstore_2:
2056 case Bytecodes::_istore_2:
2057 case Bytecodes::_astore_2:
2058 set_local( 2, pop() );
2059 break;
2060 case Bytecodes::_fstore_3:
2061 case Bytecodes::_istore_3:
2062 case Bytecodes::_astore_3:
2063 set_local( 3, pop() );
2064 break;
2065 case Bytecodes::_fstore:
2066 case Bytecodes::_istore:
2067 case Bytecodes::_astore:
2068 set_local( iter().get_index(), pop() );
2069 break;
2070 // long stores
2071 case Bytecodes::_lstore_0:
2072 set_pair_local( 0, pop_pair() );
2073 break;
2074 case Bytecodes::_lstore_1:
2075 set_pair_local( 1, pop_pair() );
2076 break;
2077 case Bytecodes::_lstore_2:
2078 set_pair_local( 2, pop_pair() );
2079 break;
2080 case Bytecodes::_lstore_3:
2081 set_pair_local( 3, pop_pair() );
2082 break;
2083 case Bytecodes::_lstore:
2084 set_pair_local( iter().get_index(), pop_pair() );
2085 break;
2086
2087 // double stores
2088 case Bytecodes::_dstore_0:
2089 set_pair_local( 0, dprecision_rounding(pop_pair()) );
2090 break;
2091 case Bytecodes::_dstore_1:
2092 set_pair_local( 1, dprecision_rounding(pop_pair()) );
2093 break;
2094 case Bytecodes::_dstore_2:
2095 set_pair_local( 2, dprecision_rounding(pop_pair()) );
2096 break;
2097 case Bytecodes::_dstore_3:
2098 set_pair_local( 3, dprecision_rounding(pop_pair()) );
2099 break;
2100 case Bytecodes::_dstore:
2101 set_pair_local( iter().get_index(), dprecision_rounding(pop_pair()) );
2102 break;
2103
2104 case Bytecodes::_pop: dec_sp(1); break;
2105 case Bytecodes::_pop2: dec_sp(2); break;
2106 case Bytecodes::_swap:
2107 a = pop();
2108 b = pop();
2109 push(a);
2110 push(b);
2111 break;
2112 case Bytecodes::_dup:
2113 a = pop();
2114 push(a);
2115 push(a);
2116 break;
2117 case Bytecodes::_dup_x1:
2118 a = pop();
2119 b = pop();
2120 push( a );
2121 push( b );
2122 push( a );
2123 break;
2124 case Bytecodes::_dup_x2:
2125 a = pop();
2126 b = pop();
2127 c = pop();
2128 push( a );
2129 push( c );
2130 push( b );
2131 push( a );
2132 break;
2133 case Bytecodes::_dup2:
2134 a = pop();
2135 b = pop();
2136 push( b );
2137 push( a );
2138 push( b );
2139 push( a );
2140 break;
2141
2142 case Bytecodes::_dup2_x1:
2143 // before: .. c, b, a
2144 // after: .. b, a, c, b, a
2145 // not tested
2146 a = pop();
2147 b = pop();
2148 c = pop();
2149 push( b );
2150 push( a );
2151 push( c );
2152 push( b );
2153 push( a );
2154 break;
2155 case Bytecodes::_dup2_x2:
2156 // before: .. d, c, b, a
2157 // after: .. b, a, d, c, b, a
2158 // not tested
2159 a = pop();
2160 b = pop();
2161 c = pop();
2162 d = pop();
2163 push( b );
2164 push( a );
2165 push( d );
2166 push( c );
2167 push( b );
2168 push( a );
2169 break;
2170
2171 case Bytecodes::_arraylength: {
2172 // Must do null-check with value on expression stack
2173 Node *ary = null_check(peek(), T_ARRAY);
2174 // Compile-time detect of null-exception?
2175 if (stopped()) return;
2176 a = pop();
2177 push(load_array_length(a));
2178 break;
2179 }
2180
2181 case Bytecodes::_baload: array_load(T_BYTE); break;
2182 case Bytecodes::_caload: array_load(T_CHAR); break;
2183 case Bytecodes::_iaload: array_load(T_INT); break;
2184 case Bytecodes::_saload: array_load(T_SHORT); break;
2185 case Bytecodes::_faload: array_load(T_FLOAT); break;
2186 case Bytecodes::_aaload: array_load(T_OBJECT); break;
2187 case Bytecodes::_laload: array_load(T_LONG); break;
2188 case Bytecodes::_daload: array_load(T_DOUBLE); break;
2189 case Bytecodes::_bastore: array_store(T_BYTE); break;
2190 case Bytecodes::_castore: array_store(T_CHAR); break;
2191 case Bytecodes::_iastore: array_store(T_INT); break;
2192 case Bytecodes::_sastore: array_store(T_SHORT); break;
2193 case Bytecodes::_fastore: array_store(T_FLOAT); break;
2194 case Bytecodes::_aastore: array_store(T_OBJECT); break;
2195 case Bytecodes::_lastore: array_store(T_LONG); break;
2196 case Bytecodes::_dastore: array_store(T_DOUBLE); break;
2197
2198 case Bytecodes::_getfield:
2199 do_getfield();
2200 break;
2201
2202 case Bytecodes::_getstatic:
2203 do_getstatic();
2204 break;
2205
2206 case Bytecodes::_putfield:
2207 do_putfield();
2208 break;
2209
2210 case Bytecodes::_putstatic:
2211 do_putstatic();
2212 break;
2213
2214 case Bytecodes::_irem:
2215 // Must keep both values on the expression-stack during null-check
2216 zero_check_int(peek());
2217 // Compile-time detect of null-exception?
2218 if (stopped()) return;
2219 b = pop();
2220 a = pop();
2221 push(_gvn.transform(new ModINode(control(), a, b)));
2222 break;
2223 case Bytecodes::_idiv:
2224 // Must keep both values on the expression-stack during null-check
2225 zero_check_int(peek());
2226 // Compile-time detect of null-exception?
2227 if (stopped()) return;
2228 b = pop();
2229 a = pop();
2230 push( _gvn.transform( new DivINode(control(),a,b) ) );
2231 break;
2232 case Bytecodes::_imul:
2233 b = pop(); a = pop();
2234 push( _gvn.transform( new MulINode(a,b) ) );
2235 break;
2236 case Bytecodes::_iadd:
2237 b = pop(); a = pop();
2238 push( _gvn.transform( new AddINode(a,b) ) );
2239 break;
2240 case Bytecodes::_ineg:
2241 a = pop();
2242 push( _gvn.transform( new SubINode(_gvn.intcon(0),a)) );
2243 break;
2244 case Bytecodes::_isub:
2245 b = pop(); a = pop();
2246 push( _gvn.transform( new SubINode(a,b) ) );
2247 break;
2248 case Bytecodes::_iand:
2249 b = pop(); a = pop();
2250 push( _gvn.transform( new AndINode(a,b) ) );
2251 break;
2252 case Bytecodes::_ior:
2253 b = pop(); a = pop();
2254 push( _gvn.transform( new OrINode(a,b) ) );
2255 break;
2256 case Bytecodes::_ixor:
2257 b = pop(); a = pop();
2258 push( _gvn.transform( new XorINode(a,b) ) );
2259 break;
2260 case Bytecodes::_ishl:
2261 b = pop(); a = pop();
2262 push( _gvn.transform( new LShiftINode(a,b) ) );
2263 break;
2264 case Bytecodes::_ishr:
2265 b = pop(); a = pop();
2266 push( _gvn.transform( new RShiftINode(a,b) ) );
2267 break;
2268 case Bytecodes::_iushr:
2269 b = pop(); a = pop();
2270 push( _gvn.transform( new URShiftINode(a,b) ) );
2271 break;
2272
2273 case Bytecodes::_fneg:
2274 a = pop();
2275 b = _gvn.transform(new NegFNode (a));
2276 push(b);
2277 break;
2278
2279 case Bytecodes::_fsub:
2280 b = pop();
2281 a = pop();
2282 c = _gvn.transform( new SubFNode(a,b) );
2283 d = precision_rounding(c);
2284 push( d );
2285 break;
2286
2287 case Bytecodes::_fadd:
2288 b = pop();
2289 a = pop();
2290 c = _gvn.transform( new AddFNode(a,b) );
2291 d = precision_rounding(c);
2292 push( d );
2293 break;
2294
2295 case Bytecodes::_fmul:
2296 b = pop();
2297 a = pop();
2298 c = _gvn.transform( new MulFNode(a,b) );
2299 d = precision_rounding(c);
2300 push( d );
2301 break;
2302
2303 case Bytecodes::_fdiv:
2304 b = pop();
2305 a = pop();
2306 c = _gvn.transform( new DivFNode(nullptr,a,b) );
2307 d = precision_rounding(c);
2308 push( d );
2309 break;
2310
2311 case Bytecodes::_frem:
2312 if (Matcher::has_match_rule(Op_ModF)) {
2313 // Generate a ModF node.
2314 b = pop();
2315 a = pop();
2316 c = _gvn.transform( new ModFNode(nullptr,a,b) );
2317 d = precision_rounding(c);
2318 push( d );
2319 }
2320 else {
2321 // Generate a call.
2322 modf();
2323 }
2324 break;
2325
2326 case Bytecodes::_fcmpl:
2327 b = pop();
2328 a = pop();
2329 c = _gvn.transform( new CmpF3Node( a, b));
2330 push(c);
2331 break;
2332 case Bytecodes::_fcmpg:
2333 b = pop();
2334 a = pop();
2335
2336 // Same as fcmpl but need to flip the unordered case. Swap the inputs,
2337 // which negates the result sign except for unordered. Flip the unordered
2338 // as well by using CmpF3 which implements unordered-lesser instead of
2339 // unordered-greater semantics. Finally, commute the result bits. Result
2340 // is same as using a CmpF3Greater except we did it with CmpF3 alone.
2341 c = _gvn.transform( new CmpF3Node( b, a));
2342 c = _gvn.transform( new SubINode(_gvn.intcon(0),c) );
2343 push(c);
2344 break;
2345
2346 case Bytecodes::_f2i:
2347 a = pop();
2348 push(_gvn.transform(new ConvF2INode(a)));
2349 break;
2350
2351 case Bytecodes::_d2i:
2352 a = pop_pair();
2353 b = _gvn.transform(new ConvD2INode(a));
2354 push( b );
2355 break;
2356
2357 case Bytecodes::_f2d:
2358 a = pop();
2359 b = _gvn.transform( new ConvF2DNode(a));
2360 push_pair( b );
2361 break;
2362
2363 case Bytecodes::_d2f:
2364 a = pop_pair();
2365 b = _gvn.transform( new ConvD2FNode(a));
2366 // This breaks _227_mtrt (speed & correctness) and _222_mpegaudio (speed)
2367 //b = _gvn.transform(new RoundFloatNode(nullptr, b) );
2368 push( b );
2369 break;
2370
2371 case Bytecodes::_l2f:
2372 if (Matcher::convL2FSupported()) {
2373 a = pop_pair();
2374 b = _gvn.transform( new ConvL2FNode(a));
2375 // For x86_32.ad, FILD doesn't restrict precision to 24 or 53 bits.
2376 // Rather than storing the result into an FP register then pushing
2377 // out to memory to round, the machine instruction that implements
2378 // ConvL2D is responsible for rounding.
2379 // c = precision_rounding(b);
2380 push(b);
2381 } else {
2382 l2f();
2383 }
2384 break;
2385
2386 case Bytecodes::_l2d:
2387 a = pop_pair();
2388 b = _gvn.transform( new ConvL2DNode(a));
2389 // For x86_32.ad, rounding is always necessary (see _l2f above).
2390 // c = dprecision_rounding(b);
2391 push_pair(b);
2392 break;
2393
2394 case Bytecodes::_f2l:
2395 a = pop();
2396 b = _gvn.transform( new ConvF2LNode(a));
2397 push_pair(b);
2398 break;
2399
2400 case Bytecodes::_d2l:
2401 a = pop_pair();
2402 b = _gvn.transform( new ConvD2LNode(a));
2403 push_pair(b);
2404 break;
2405
2406 case Bytecodes::_dsub:
2407 b = pop_pair();
2408 a = pop_pair();
2409 c = _gvn.transform( new SubDNode(a,b) );
2410 d = dprecision_rounding(c);
2411 push_pair( d );
2412 break;
2413
2414 case Bytecodes::_dadd:
2415 b = pop_pair();
2416 a = pop_pair();
2417 c = _gvn.transform( new AddDNode(a,b) );
2418 d = dprecision_rounding(c);
2419 push_pair( d );
2420 break;
2421
2422 case Bytecodes::_dmul:
2423 b = pop_pair();
2424 a = pop_pair();
2425 c = _gvn.transform( new MulDNode(a,b) );
2426 d = dprecision_rounding(c);
2427 push_pair( d );
2428 break;
2429
2430 case Bytecodes::_ddiv:
2431 b = pop_pair();
2432 a = pop_pair();
2433 c = _gvn.transform( new DivDNode(nullptr,a,b) );
2434 d = dprecision_rounding(c);
2435 push_pair( d );
2436 break;
2437
2438 case Bytecodes::_dneg:
2439 a = pop_pair();
2440 b = _gvn.transform(new NegDNode (a));
2441 push_pair(b);
2442 break;
2443
2444 case Bytecodes::_drem:
2445 if (Matcher::has_match_rule(Op_ModD)) {
2446 // Generate a ModD node.
2447 b = pop_pair();
2448 a = pop_pair();
2449 // a % b
2450
2451 c = _gvn.transform( new ModDNode(nullptr,a,b) );
2452 d = dprecision_rounding(c);
2453 push_pair( d );
2454 }
2455 else {
2456 // Generate a call.
2457 modd();
2458 }
2459 break;
2460
2461 case Bytecodes::_dcmpl:
2462 b = pop_pair();
2463 a = pop_pair();
2464 c = _gvn.transform( new CmpD3Node( a, b));
2465 push(c);
2466 break;
2467
2468 case Bytecodes::_dcmpg:
2469 b = pop_pair();
2470 a = pop_pair();
2471 // Same as dcmpl but need to flip the unordered case.
2472 // Commute the inputs, which negates the result sign except for unordered.
2473 // Flip the unordered as well by using CmpD3 which implements
2474 // unordered-lesser instead of unordered-greater semantics.
2475 // Finally, negate the result bits. Result is same as using a
2476 // CmpD3Greater except we did it with CmpD3 alone.
2477 c = _gvn.transform( new CmpD3Node( b, a));
2478 c = _gvn.transform( new SubINode(_gvn.intcon(0),c) );
2479 push(c);
2480 break;
2481
2482
2483 // Note for longs -> lo word is on TOS, hi word is on TOS - 1
2484 case Bytecodes::_land:
2485 b = pop_pair();
2486 a = pop_pair();
2487 c = _gvn.transform( new AndLNode(a,b) );
2488 push_pair(c);
2489 break;
2490 case Bytecodes::_lor:
2491 b = pop_pair();
2492 a = pop_pair();
2493 c = _gvn.transform( new OrLNode(a,b) );
2494 push_pair(c);
2495 break;
2496 case Bytecodes::_lxor:
2497 b = pop_pair();
2498 a = pop_pair();
2499 c = _gvn.transform( new XorLNode(a,b) );
2500 push_pair(c);
2501 break;
2502
2503 case Bytecodes::_lshl:
2504 b = pop(); // the shift count
2505 a = pop_pair(); // value to be shifted
2506 c = _gvn.transform( new LShiftLNode(a,b) );
2507 push_pair(c);
2508 break;
2509 case Bytecodes::_lshr:
2510 b = pop(); // the shift count
2511 a = pop_pair(); // value to be shifted
2512 c = _gvn.transform( new RShiftLNode(a,b) );
2513 push_pair(c);
2514 break;
2515 case Bytecodes::_lushr:
2516 b = pop(); // the shift count
2517 a = pop_pair(); // value to be shifted
2518 c = _gvn.transform( new URShiftLNode(a,b) );
2519 push_pair(c);
2520 break;
2521 case Bytecodes::_lmul:
2522 b = pop_pair();
2523 a = pop_pair();
2524 c = _gvn.transform( new MulLNode(a,b) );
2525 push_pair(c);
2526 break;
2527
2528 case Bytecodes::_lrem:
2529 // Must keep both values on the expression-stack during null-check
2530 assert(peek(0) == top(), "long word order");
2531 zero_check_long(peek(1));
2532 // Compile-time detect of null-exception?
2533 if (stopped()) return;
2534 b = pop_pair();
2535 a = pop_pair();
2536 c = _gvn.transform( new ModLNode(control(),a,b) );
2537 push_pair(c);
2538 break;
2539
2540 case Bytecodes::_ldiv:
2541 // Must keep both values on the expression-stack during null-check
2542 assert(peek(0) == top(), "long word order");
2543 zero_check_long(peek(1));
2544 // Compile-time detect of null-exception?
2545 if (stopped()) return;
2546 b = pop_pair();
2547 a = pop_pair();
2548 c = _gvn.transform( new DivLNode(control(),a,b) );
2549 push_pair(c);
2550 break;
2551
2552 case Bytecodes::_ladd:
2553 b = pop_pair();
2554 a = pop_pair();
2555 c = _gvn.transform( new AddLNode(a,b) );
2556 push_pair(c);
2557 break;
2558 case Bytecodes::_lsub:
2559 b = pop_pair();
2560 a = pop_pair();
2561 c = _gvn.transform( new SubLNode(a,b) );
2562 push_pair(c);
2563 break;
2564 case Bytecodes::_lcmp:
2565 // Safepoints are now inserted _before_ branches. The long-compare
2566 // bytecode painfully produces a 3-way value (-1,0,+1) which requires a
2567 // slew of control flow. These are usually followed by a CmpI vs zero and
2568 // a branch; this pattern then optimizes to the obvious long-compare and
2569 // branch. However, if the branch is backwards there's a Safepoint
2570 // inserted. The inserted Safepoint captures the JVM state at the
2571 // pre-branch point, i.e. it captures the 3-way value. Thus if a
2572 // long-compare is used to control a loop the debug info will force
2573 // computation of the 3-way value, even though the generated code uses a
2574 // long-compare and branch. We try to rectify the situation by inserting
2575 // a SafePoint here and have it dominate and kill the safepoint added at a
2576 // following backwards branch. At this point the JVM state merely holds 2
2577 // longs but not the 3-way value.
2578 switch (iter().next_bc()) {
2579 case Bytecodes::_ifgt:
2580 case Bytecodes::_iflt:
2581 case Bytecodes::_ifge:
2582 case Bytecodes::_ifle:
2583 case Bytecodes::_ifne:
2584 case Bytecodes::_ifeq:
2585 // If this is a backwards branch in the bytecodes, add Safepoint
2586 maybe_add_safepoint(iter().next_get_dest());
2587 default:
2588 break;
2589 }
2590 b = pop_pair();
2591 a = pop_pair();
2592 c = _gvn.transform( new CmpL3Node( a, b ));
2593 push(c);
2594 break;
2595
2596 case Bytecodes::_lneg:
2597 a = pop_pair();
2598 b = _gvn.transform( new SubLNode(longcon(0),a));
2599 push_pair(b);
2600 break;
2601 case Bytecodes::_l2i:
2602 a = pop_pair();
2603 push( _gvn.transform( new ConvL2INode(a)));
2604 break;
2605 case Bytecodes::_i2l:
2606 a = pop();
2607 b = _gvn.transform( new ConvI2LNode(a));
2608 push_pair(b);
2609 break;
2610 case Bytecodes::_i2b:
2611 // Sign extend
2612 a = pop();
2613 a = Compile::narrow_value(T_BYTE, a, nullptr, &_gvn, true);
2614 push(a);
2615 break;
2616 case Bytecodes::_i2s:
2617 a = pop();
2618 a = Compile::narrow_value(T_SHORT, a, nullptr, &_gvn, true);
2619 push(a);
2620 break;
2621 case Bytecodes::_i2c:
2622 a = pop();
2623 a = Compile::narrow_value(T_CHAR, a, nullptr, &_gvn, true);
2624 push(a);
2625 break;
2626
2627 case Bytecodes::_i2f:
2628 a = pop();
2629 b = _gvn.transform( new ConvI2FNode(a) ) ;
2630 c = precision_rounding(b);
2631 push (b);
2632 break;
2633
2634 case Bytecodes::_i2d:
2635 a = pop();
2636 b = _gvn.transform( new ConvI2DNode(a));
2637 push_pair(b);
2638 break;
2639
2640 case Bytecodes::_iinc: // Increment local
2641 i = iter().get_index(); // Get local index
2642 set_local( i, _gvn.transform( new AddINode( _gvn.intcon(iter().get_iinc_con()), local(i) ) ) );
2643 break;
2644
2645 // Exit points of synchronized methods must have an unlock node
2646 case Bytecodes::_return:
2647 return_current(nullptr);
2648 break;
2649
2650 case Bytecodes::_ireturn:
2651 case Bytecodes::_areturn:
2652 case Bytecodes::_freturn:
2653 return_current(pop());
2654 break;
2655 case Bytecodes::_lreturn:
2656 return_current(pop_pair());
2657 break;
2658 case Bytecodes::_dreturn:
2659 return_current(pop_pair());
2660 break;
2661
2662 case Bytecodes::_athrow:
2663 // null exception oop throws null pointer exception
2664 null_check(peek());
2665 if (stopped()) return;
2666 // Hook the thrown exception directly to subsequent handlers.
2667 if (BailoutToInterpreterForThrows) {
2668 // Keep method interpreted from now on.
2669 uncommon_trap(Deoptimization::Reason_unhandled,
2670 Deoptimization::Action_make_not_compilable);
2671 return;
2672 }
2673 if (env()->jvmti_can_post_on_exceptions()) {
2674 // check if we must post exception events, take uncommon trap if so (with must_throw = false)
2675 uncommon_trap_if_should_post_on_exceptions(Deoptimization::Reason_unhandled, false);
2676 }
2677 // Here if either can_post_on_exceptions or should_post_on_exceptions is false
2678 add_exception_state(make_exception_state(peek()));
2679 break;
2680
2681 case Bytecodes::_goto: // fall through
2682 case Bytecodes::_goto_w: {
2683 int target_bci = (bc() == Bytecodes::_goto) ? iter().get_dest() : iter().get_far_dest();
2684
2685 // If this is a backwards branch in the bytecodes, add Safepoint
2686 maybe_add_safepoint(target_bci);
2687
2688 // Merge the current control into the target basic block
2689 merge(target_bci);
2690
2691 // See if we can get some profile data and hand it off to the next block
2692 Block *target_block = block()->successor_for_bci(target_bci);
2693 if (target_block->pred_count() != 1) break;
2694 ciMethodData* methodData = method()->method_data();
2695 if (!methodData->is_mature()) break;
2696 ciProfileData* data = methodData->bci_to_data(bci());
2697 assert(data != nullptr && data->is_JumpData(), "need JumpData for taken branch");
2698 int taken = ((ciJumpData*)data)->taken();
2699 taken = method()->scale_count(taken);
2700 target_block->set_count(taken);
2701 break;
2702 }
2703
2704 case Bytecodes::_ifnull: btest = BoolTest::eq; goto handle_if_null;
2705 case Bytecodes::_ifnonnull: btest = BoolTest::ne; goto handle_if_null;
2706 handle_if_null:
2707 // If this is a backwards branch in the bytecodes, add Safepoint
2708 maybe_add_safepoint(iter().get_dest());
2709 a = null();
2710 b = pop();
2711 if (!_gvn.type(b)->speculative_maybe_null() &&
2712 !too_many_traps(Deoptimization::Reason_speculate_null_check)) {
2713 inc_sp(1);
2714 Node* null_ctl = top();
2715 b = null_check_oop(b, &null_ctl, true, true, true);
2716 assert(null_ctl->is_top(), "no null control here");
2717 dec_sp(1);
2718 } else if (_gvn.type(b)->speculative_always_null() &&
2719 !too_many_traps(Deoptimization::Reason_speculate_null_assert)) {
2720 inc_sp(1);
2721 b = null_assert(b);
2722 dec_sp(1);
2723 }
2724 c = _gvn.transform( new CmpPNode(b, a) );
2725 do_ifnull(btest, c);
2726 break;
2727
2728 case Bytecodes::_if_acmpeq: btest = BoolTest::eq; goto handle_if_acmp;
2729 case Bytecodes::_if_acmpne: btest = BoolTest::ne; goto handle_if_acmp;
2730 handle_if_acmp:
2731 // If this is a backwards branch in the bytecodes, add Safepoint
2732 maybe_add_safepoint(iter().get_dest());
2733 a = pop();
2734 b = pop();
2735 c = _gvn.transform( new CmpPNode(b, a) );
2736 c = optimize_cmp_with_klass(c);
2737 do_if(btest, c);
2738 break;
2739
2740 case Bytecodes::_ifeq: btest = BoolTest::eq; goto handle_ifxx;
2741 case Bytecodes::_ifne: btest = BoolTest::ne; goto handle_ifxx;
2742 case Bytecodes::_iflt: btest = BoolTest::lt; goto handle_ifxx;
2743 case Bytecodes::_ifle: btest = BoolTest::le; goto handle_ifxx;
2744 case Bytecodes::_ifgt: btest = BoolTest::gt; goto handle_ifxx;
2745 case Bytecodes::_ifge: btest = BoolTest::ge; goto handle_ifxx;
2746 handle_ifxx:
2747 // If this is a backwards branch in the bytecodes, add Safepoint
2748 maybe_add_safepoint(iter().get_dest());
2749 a = _gvn.intcon(0);
2750 b = pop();
2751 c = _gvn.transform( new CmpINode(b, a) );
2752 do_if(btest, c);
2753 break;
2754
2755 case Bytecodes::_if_icmpeq: btest = BoolTest::eq; goto handle_if_icmp;
2756 case Bytecodes::_if_icmpne: btest = BoolTest::ne; goto handle_if_icmp;
2757 case Bytecodes::_if_icmplt: btest = BoolTest::lt; goto handle_if_icmp;
2758 case Bytecodes::_if_icmple: btest = BoolTest::le; goto handle_if_icmp;
2759 case Bytecodes::_if_icmpgt: btest = BoolTest::gt; goto handle_if_icmp;
2760 case Bytecodes::_if_icmpge: btest = BoolTest::ge; goto handle_if_icmp;
2761 handle_if_icmp:
2762 // If this is a backwards branch in the bytecodes, add Safepoint
2763 maybe_add_safepoint(iter().get_dest());
2764 a = pop();
2765 b = pop();
2766 c = _gvn.transform( new CmpINode( b, a ) );
2767 do_if(btest, c);
2768 break;
2769
2770 case Bytecodes::_tableswitch:
2771 do_tableswitch();
2772 break;
2773
2774 case Bytecodes::_lookupswitch:
2775 do_lookupswitch();
2776 break;
2777
2778 case Bytecodes::_invokestatic:
2779 case Bytecodes::_invokedynamic:
2780 case Bytecodes::_invokespecial:
2781 case Bytecodes::_invokevirtual:
2782 case Bytecodes::_invokeinterface:
2783 do_call();
2784 break;
2785 case Bytecodes::_checkcast:
2786 do_checkcast();
2787 break;
2788 case Bytecodes::_instanceof:
2789 do_instanceof();
2790 break;
2791 case Bytecodes::_anewarray:
2792 do_anewarray();
2793 break;
2794 case Bytecodes::_newarray:
2795 do_newarray((BasicType)iter().get_index());
2796 break;
2797 case Bytecodes::_multianewarray:
2798 do_multianewarray();
2799 break;
2800 case Bytecodes::_new:
2801 do_new();
2802 break;
2803
2804 case Bytecodes::_jsr:
2805 case Bytecodes::_jsr_w:
2806 do_jsr();
2807 break;
2808
2809 case Bytecodes::_ret:
2810 do_ret();
2811 break;
2812
2813
2814 case Bytecodes::_monitorenter:
2815 do_monitor_enter();
2816 break;
2817
2818 case Bytecodes::_monitorexit:
2819 do_monitor_exit();
2820 break;
2821
2822 case Bytecodes::_breakpoint:
2823 // Breakpoint set concurrently to compile
2824 // %%% use an uncommon trap?
2825 C->record_failure("breakpoint in method");
2826 return;
2827
2828 default:
2829 #ifndef PRODUCT
2830 map()->dump(99);
2831 #endif
2832 tty->print("\nUnhandled bytecode %s\n", Bytecodes::name(bc()) );
2833 ShouldNotReachHere();
2834 }
2835
2836 #ifndef PRODUCT
2837 if (failing()) { return; }
2838 constexpr int perBytecode = 6;
2839 if (C->should_print_igv(perBytecode)) {
2840 IdealGraphPrinter* printer = C->igv_printer();
2841 char buffer[256];
2842 jio_snprintf(buffer, sizeof(buffer), "Bytecode %d: %s", bci(), Bytecodes::name(bc()));
2843 bool old = printer->traverse_outs();
2844 printer->set_traverse_outs(true);
2845 printer->print_method(buffer, perBytecode);
2846 printer->set_traverse_outs(old);
2847 }
2848 #endif
2849 }