1 /*
2 * Copyright (c) 1998, 2022, 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 "jvm_io.h"
27 #include "ci/ciMethodData.hpp"
28 #include "classfile/vmSymbols.hpp"
29 #include "compiler/compileLog.hpp"
30 #include "interpreter/linkResolver.hpp"
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 int 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->klass()->as_instance_klass()->unique_concrete_subklass()) {
128 // If we load from "AbstractClass[]" we must see "ConcreteSubClass".
129 const Type* subklass = Type::get_const_type(toop->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() != NULL) C->log()->elem("observe that='!need_range_check'");
144 }
145
146 ciKlass * arytype_klass = arytype->klass();
147 if ((arytype_klass != NULL) && (!arytype_klass->is_loaded())) {
148 // Only fails for some -Xcomp runs
149 // The class is unloaded. We have to run this bytecode in the interpreter.
150 uncommon_trap(Deoptimization::Reason_unloaded,
151 Deoptimization::Action_reinterpret,
152 arytype->klass(), "!loaded array");
153 return top();
154 }
155
156 // Do the range check
157 if (GenerateRangeChecks && need_range_check) {
158 Node* tst;
159 if (sizetype->_hi <= 0) {
160 // The greatest array bound is negative, so we can conclude that we're
161 // compiling unreachable code, but the unsigned compare trick used below
162 // only works with non-negative lengths. Instead, hack "tst" to be zero so
163 // the uncommon_trap path will always be taken.
164 tst = _gvn.intcon(0);
165 } else {
166 // Range is constant in array-oop, so we can use the original state of mem
167 Node* len = load_array_length(ary);
168
169 // Test length vs index (standard trick using unsigned compare)
170 Node* chk = _gvn.transform( new CmpUNode(idx, len) );
171 BoolTest::mask btest = BoolTest::lt;
172 tst = _gvn.transform( new BoolNode(chk, btest) );
173 }
174 RangeCheckNode* rc = new RangeCheckNode(control(), tst, PROB_MAX, COUNT_UNKNOWN);
175 _gvn.set_type(rc, rc->Value(&_gvn));
176 if (!tst->is_Con()) {
177 record_for_igvn(rc);
178 }
179 set_control(_gvn.transform(new IfTrueNode(rc)));
180 // Branch to failure if out of bounds
181 {
182 PreserveJVMState pjvms(this);
183 set_control(_gvn.transform(new IfFalseNode(rc)));
184 if (C->allow_range_check_smearing()) {
185 // Do not use builtin_throw, since range checks are sometimes
186 // made more stringent by an optimistic transformation.
187 // This creates "tentative" range checks at this point,
188 // which are not guaranteed to throw exceptions.
189 // See IfNode::Ideal, is_range_check, adjust_check.
190 uncommon_trap(Deoptimization::Reason_range_check,
191 Deoptimization::Action_make_not_entrant,
192 NULL, "range_check");
193 } else {
194 // If we have already recompiled with the range-check-widening
195 // heroic optimization turned off, then we must really be throwing
196 // range check exceptions.
197 builtin_throw(Deoptimization::Reason_range_check, idx);
198 }
199 }
200 }
201 // Check for always knowing you are throwing a range-check exception
202 if (stopped()) return top();
203
204 // Make array address computation control dependent to prevent it
205 // from floating above the range check during loop optimizations.
206 Node* ptr = array_element_address(ary, idx, type, sizetype, control());
207 assert(ptr != top(), "top should go hand-in-hand with stopped");
208
209 return ptr;
210 }
211
212
213 // returns IfNode
214 IfNode* Parse::jump_if_fork_int(Node* a, Node* b, BoolTest::mask mask, float prob, float cnt) {
215 Node *cmp = _gvn.transform(new CmpINode(a, b)); // two cases: shiftcount > 32 and shiftcount <= 32
216 Node *tst = _gvn.transform(new BoolNode(cmp, mask));
217 IfNode *iff = create_and_map_if(control(), tst, prob, cnt);
218 return iff;
219 }
220
221
222 // sentinel value for the target bci to mark never taken branches
223 // (according to profiling)
224 static const int never_reached = INT_MAX;
225
226 //------------------------------helper for tableswitch-------------------------
227 void Parse::jump_if_true_fork(IfNode *iff, int dest_bci_if_true, bool unc) {
228 // True branch, use existing map info
229 { PreserveJVMState pjvms(this);
230 Node *iftrue = _gvn.transform( new IfTrueNode (iff) );
231 set_control( iftrue );
232 if (unc) {
233 repush_if_args();
234 uncommon_trap(Deoptimization::Reason_unstable_if,
235 Deoptimization::Action_reinterpret,
236 NULL,
237 "taken always");
238 } else {
239 assert(dest_bci_if_true != never_reached, "inconsistent dest");
240 merge_new_path(dest_bci_if_true);
241 }
242 }
243
244 // False branch
245 Node *iffalse = _gvn.transform( new IfFalseNode(iff) );
246 set_control( iffalse );
247 }
248
249 void Parse::jump_if_false_fork(IfNode *iff, int dest_bci_if_true, bool unc) {
250 // True branch, use existing map info
251 { PreserveJVMState pjvms(this);
252 Node *iffalse = _gvn.transform( new IfFalseNode (iff) );
253 set_control( iffalse );
254 if (unc) {
255 repush_if_args();
256 uncommon_trap(Deoptimization::Reason_unstable_if,
257 Deoptimization::Action_reinterpret,
258 NULL,
259 "taken never");
260 } else {
261 assert(dest_bci_if_true != never_reached, "inconsistent dest");
262 merge_new_path(dest_bci_if_true);
263 }
264 }
265
266 // False branch
267 Node *iftrue = _gvn.transform( new IfTrueNode(iff) );
268 set_control( iftrue );
269 }
270
271 void Parse::jump_if_always_fork(int dest_bci, bool unc) {
272 // False branch, use existing map and control()
273 if (unc) {
274 repush_if_args();
275 uncommon_trap(Deoptimization::Reason_unstable_if,
276 Deoptimization::Action_reinterpret,
277 NULL,
278 "taken never");
279 } else {
280 assert(dest_bci != never_reached, "inconsistent dest");
281 merge_new_path(dest_bci);
282 }
283 }
284
285
286 extern "C" {
287 static int jint_cmp(const void *i, const void *j) {
288 int a = *(jint *)i;
289 int b = *(jint *)j;
290 return a > b ? 1 : a < b ? -1 : 0;
291 }
292 }
293
294
295 class SwitchRange : public StackObj {
296 // a range of integers coupled with a bci destination
297 jint _lo; // inclusive lower limit
298 jint _hi; // inclusive upper limit
299 int _dest;
300 float _cnt; // how many times this range was hit according to profiling
301
302 public:
303 jint lo() const { return _lo; }
304 jint hi() const { return _hi; }
305 int dest() const { return _dest; }
306 bool is_singleton() const { return _lo == _hi; }
307 float cnt() const { return _cnt; }
308
309 void setRange(jint lo, jint hi, int dest, float cnt) {
310 assert(lo <= hi, "must be a non-empty range");
311 _lo = lo, _hi = hi; _dest = dest; _cnt = cnt;
312 assert(_cnt >= 0, "");
313 }
314 bool adjoinRange(jint lo, jint hi, int dest, float cnt, bool trim_ranges) {
315 assert(lo <= hi, "must be a non-empty range");
316 if (lo == _hi+1) {
317 // see merge_ranges() comment below
318 if (trim_ranges) {
319 if (cnt == 0) {
320 if (_cnt != 0) {
321 return false;
322 }
323 if (dest != _dest) {
324 _dest = never_reached;
325 }
326 } else {
327 if (_cnt == 0) {
328 return false;
329 }
330 if (dest != _dest) {
331 return false;
332 }
333 }
334 } else {
335 if (dest != _dest) {
336 return false;
337 }
338 }
339 _hi = hi;
340 _cnt += cnt;
341 return true;
342 }
343 return false;
344 }
345
346 void set (jint value, int dest, float cnt) {
347 setRange(value, value, dest, cnt);
348 }
349 bool adjoin(jint value, int dest, float cnt, bool trim_ranges) {
350 return adjoinRange(value, value, dest, cnt, trim_ranges);
351 }
352 bool adjoin(SwitchRange& other) {
353 return adjoinRange(other._lo, other._hi, other._dest, other._cnt, false);
354 }
355
356 void print() {
357 if (is_singleton())
358 tty->print(" {%d}=>%d (cnt=%f)", lo(), dest(), cnt());
359 else if (lo() == min_jint)
360 tty->print(" {..%d}=>%d (cnt=%f)", hi(), dest(), cnt());
361 else if (hi() == max_jint)
362 tty->print(" {%d..}=>%d (cnt=%f)", lo(), dest(), cnt());
363 else
364 tty->print(" {%d..%d}=>%d (cnt=%f)", lo(), hi(), dest(), cnt());
365 }
366 };
367
368 // We try to minimize the number of ranges and the size of the taken
369 // ones using profiling data. When ranges are created,
370 // SwitchRange::adjoinRange() only allows 2 adjoining ranges to merge
371 // if both were never hit or both were hit to build longer unreached
372 // ranges. Here, we now merge adjoining ranges with the same
373 // destination and finally set destination of unreached ranges to the
374 // special value never_reached because it can help minimize the number
375 // of tests that are necessary.
376 //
377 // For instance:
378 // [0, 1] to target1 sometimes taken
379 // [1, 2] to target1 never taken
380 // [2, 3] to target2 never taken
381 // would lead to:
382 // [0, 1] to target1 sometimes taken
383 // [1, 3] never taken
384 //
385 // (first 2 ranges to target1 are not merged)
386 static void merge_ranges(SwitchRange* ranges, int& rp) {
387 if (rp == 0) {
388 return;
389 }
390 int shift = 0;
391 for (int j = 0; j < rp; j++) {
392 SwitchRange& r1 = ranges[j-shift];
393 SwitchRange& r2 = ranges[j+1];
394 if (r1.adjoin(r2)) {
395 shift++;
396 } else if (shift > 0) {
397 ranges[j+1-shift] = r2;
398 }
399 }
400 rp -= shift;
401 for (int j = 0; j <= rp; j++) {
402 SwitchRange& r = ranges[j];
403 if (r.cnt() == 0 && r.dest() != never_reached) {
404 r.setRange(r.lo(), r.hi(), never_reached, r.cnt());
405 }
406 }
407 }
408
409 //-------------------------------do_tableswitch--------------------------------
410 void Parse::do_tableswitch() {
411 // Get information about tableswitch
412 int default_dest = iter().get_dest_table(0);
413 jint lo_index = iter().get_int_table(1);
414 jint hi_index = iter().get_int_table(2);
415 int len = hi_index - lo_index + 1;
416
417 if (len < 1) {
418 // If this is a backward branch, add safepoint
419 maybe_add_safepoint(default_dest);
420 pop(); // the effect of the instruction execution on the operand stack
421 merge(default_dest);
422 return;
423 }
424
425 ciMethodData* methodData = method()->method_data();
426 ciMultiBranchData* profile = NULL;
427 if (methodData->is_mature() && UseSwitchProfiling) {
428 ciProfileData* data = methodData->bci_to_data(bci());
429 if (data != NULL && data->is_MultiBranchData()) {
430 profile = (ciMultiBranchData*)data;
431 }
432 }
433 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if);
434
435 // generate decision tree, using trichotomy when possible
436 int rnum = len+2;
437 bool makes_backward_branch = false;
438 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum);
439 int rp = -1;
440 if (lo_index != min_jint) {
441 float cnt = 1.0F;
442 if (profile != NULL) {
443 cnt = (float)profile->default_count() / (hi_index != max_jint ? 2.0F : 1.0F);
444 }
445 ranges[++rp].setRange(min_jint, lo_index-1, default_dest, cnt);
446 }
447 for (int j = 0; j < len; j++) {
448 jint match_int = lo_index+j;
449 int dest = iter().get_dest_table(j+3);
450 makes_backward_branch |= (dest <= bci());
451 float cnt = 1.0F;
452 if (profile != NULL) {
453 cnt = (float)profile->count_at(j);
454 }
455 if (rp < 0 || !ranges[rp].adjoin(match_int, dest, cnt, trim_ranges)) {
456 ranges[++rp].set(match_int, dest, cnt);
457 }
458 }
459 jint highest = lo_index+(len-1);
460 assert(ranges[rp].hi() == highest, "");
461 if (highest != max_jint) {
462 float cnt = 1.0F;
463 if (profile != NULL) {
464 cnt = (float)profile->default_count() / (lo_index != min_jint ? 2.0F : 1.0F);
465 }
466 if (!ranges[rp].adjoinRange(highest+1, max_jint, default_dest, cnt, trim_ranges)) {
467 ranges[++rp].setRange(highest+1, max_jint, default_dest, cnt);
468 }
469 }
470 assert(rp < len+2, "not too many ranges");
471
472 if (trim_ranges) {
473 merge_ranges(ranges, rp);
474 }
475
476 // Safepoint in case if backward branch observed
477 if (makes_backward_branch) {
478 add_safepoint();
479 }
480
481 Node* lookup = pop(); // lookup value
482 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]);
483 }
484
485
486 //------------------------------do_lookupswitch--------------------------------
487 void Parse::do_lookupswitch() {
488 // Get information about lookupswitch
489 int default_dest = iter().get_dest_table(0);
490 jint len = iter().get_int_table(1);
491
492 if (len < 1) { // If this is a backward branch, add safepoint
493 maybe_add_safepoint(default_dest);
494 pop(); // the effect of the instruction execution on the operand stack
495 merge(default_dest);
496 return;
497 }
498
499 ciMethodData* methodData = method()->method_data();
500 ciMultiBranchData* profile = NULL;
501 if (methodData->is_mature() && UseSwitchProfiling) {
502 ciProfileData* data = methodData->bci_to_data(bci());
503 if (data != NULL && data->is_MultiBranchData()) {
504 profile = (ciMultiBranchData*)data;
505 }
506 }
507 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if);
508
509 // generate decision tree, using trichotomy when possible
510 jint* table = NEW_RESOURCE_ARRAY(jint, len*3);
511 {
512 for (int j = 0; j < len; j++) {
513 table[3*j+0] = iter().get_int_table(2+2*j);
514 table[3*j+1] = iter().get_dest_table(2+2*j+1);
515 // Handle overflow when converting from uint to jint
516 table[3*j+2] = (profile == NULL) ? 1 : (jint)MIN2<uint>((uint)max_jint, profile->count_at(j));
517 }
518 qsort(table, len, 3*sizeof(table[0]), jint_cmp);
519 }
520
521 float default_cnt = 1.0F;
522 if (profile != NULL) {
523 juint defaults = max_juint - len;
524 default_cnt = (float)profile->default_count()/(float)defaults;
525 }
526
527 int rnum = len*2+1;
528 bool makes_backward_branch = false;
529 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum);
530 int rp = -1;
531 for (int j = 0; j < len; j++) {
532 jint match_int = table[3*j+0];
533 jint dest = table[3*j+1];
534 jint cnt = table[3*j+2];
535 jint next_lo = rp < 0 ? min_jint : ranges[rp].hi()+1;
536 makes_backward_branch |= (dest <= bci());
537 float c = default_cnt * ((float)match_int - (float)next_lo);
538 if (match_int != next_lo && (rp < 0 || !ranges[rp].adjoinRange(next_lo, match_int-1, default_dest, c, trim_ranges))) {
539 assert(default_dest != never_reached, "sentinel value for dead destinations");
540 ranges[++rp].setRange(next_lo, match_int-1, default_dest, c);
541 }
542 if (rp < 0 || !ranges[rp].adjoin(match_int, dest, (float)cnt, trim_ranges)) {
543 assert(dest != never_reached, "sentinel value for dead destinations");
544 ranges[++rp].set(match_int, dest, (float)cnt);
545 }
546 }
547 jint highest = table[3*(len-1)];
548 assert(ranges[rp].hi() == highest, "");
549 if (highest != max_jint &&
550 !ranges[rp].adjoinRange(highest+1, max_jint, default_dest, default_cnt * ((float)max_jint - (float)highest), trim_ranges)) {
551 ranges[++rp].setRange(highest+1, max_jint, default_dest, default_cnt * ((float)max_jint - (float)highest));
552 }
553 assert(rp < rnum, "not too many ranges");
554
555 if (trim_ranges) {
556 merge_ranges(ranges, rp);
557 }
558
559 // Safepoint in case backward branch observed
560 if (makes_backward_branch) {
561 add_safepoint();
562 }
563
564 Node *lookup = pop(); // lookup value
565 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]);
566 }
567
568 static float if_prob(float taken_cnt, float total_cnt) {
569 assert(taken_cnt <= total_cnt, "");
570 if (total_cnt == 0) {
571 return PROB_FAIR;
572 }
573 float p = taken_cnt / total_cnt;
574 return clamp(p, PROB_MIN, PROB_MAX);
575 }
576
577 static float if_cnt(float cnt) {
578 if (cnt == 0) {
579 return COUNT_UNKNOWN;
580 }
581 return cnt;
582 }
583
584 static float sum_of_cnts(SwitchRange *lo, SwitchRange *hi) {
585 float total_cnt = 0;
586 for (SwitchRange* sr = lo; sr <= hi; sr++) {
587 total_cnt += sr->cnt();
588 }
589 return total_cnt;
590 }
591
592 class SwitchRanges : public ResourceObj {
593 public:
594 SwitchRange* _lo;
595 SwitchRange* _hi;
596 SwitchRange* _mid;
597 float _cost;
598
599 enum {
600 Start,
601 LeftDone,
602 RightDone,
603 Done
604 } _state;
605
606 SwitchRanges(SwitchRange *lo, SwitchRange *hi)
607 : _lo(lo), _hi(hi), _mid(NULL),
608 _cost(0), _state(Start) {
609 }
610
611 SwitchRanges()
612 : _lo(NULL), _hi(NULL), _mid(NULL),
613 _cost(0), _state(Start) {}
614 };
615
616 // Estimate cost of performing a binary search on lo..hi
617 static float compute_tree_cost(SwitchRange *lo, SwitchRange *hi, float total_cnt) {
618 GrowableArray<SwitchRanges> tree;
619 SwitchRanges root(lo, hi);
620 tree.push(root);
621
622 float cost = 0;
623 do {
624 SwitchRanges& r = *tree.adr_at(tree.length()-1);
625 if (r._hi != r._lo) {
626 if (r._mid == NULL) {
627 float r_cnt = sum_of_cnts(r._lo, r._hi);
628
629 if (r_cnt == 0) {
630 tree.pop();
631 cost = 0;
632 continue;
633 }
634
635 SwitchRange* mid = NULL;
636 mid = r._lo;
637 for (float cnt = 0; ; ) {
638 assert(mid <= r._hi, "out of bounds");
639 cnt += mid->cnt();
640 if (cnt > r_cnt / 2) {
641 break;
642 }
643 mid++;
644 }
645 assert(mid <= r._hi, "out of bounds");
646 r._mid = mid;
647 r._cost = r_cnt / total_cnt;
648 }
649 r._cost += cost;
650 if (r._state < SwitchRanges::LeftDone && r._mid > r._lo) {
651 cost = 0;
652 r._state = SwitchRanges::LeftDone;
653 tree.push(SwitchRanges(r._lo, r._mid-1));
654 } else if (r._state < SwitchRanges::RightDone) {
655 cost = 0;
656 r._state = SwitchRanges::RightDone;
657 tree.push(SwitchRanges(r._mid == r._lo ? r._mid+1 : r._mid, r._hi));
658 } else {
659 tree.pop();
660 cost = r._cost;
661 }
662 } else {
663 tree.pop();
664 cost = r._cost;
665 }
666 } while (tree.length() > 0);
667
668
669 return cost;
670 }
671
672 // It sometimes pays off to test most common ranges before the binary search
673 void Parse::linear_search_switch_ranges(Node* key_val, SwitchRange*& lo, SwitchRange*& hi) {
674 uint nr = hi - lo + 1;
675 float total_cnt = sum_of_cnts(lo, hi);
676
677 float min = compute_tree_cost(lo, hi, total_cnt);
678 float extra = 1;
679 float sub = 0;
680
681 SwitchRange* array1 = lo;
682 SwitchRange* array2 = NEW_RESOURCE_ARRAY(SwitchRange, nr);
683
684 SwitchRange* ranges = NULL;
685
686 while (nr >= 2) {
687 assert(lo == array1 || lo == array2, "one the 2 already allocated arrays");
688 ranges = (lo == array1) ? array2 : array1;
689
690 // Find highest frequency range
691 SwitchRange* candidate = lo;
692 for (SwitchRange* sr = lo+1; sr <= hi; sr++) {
693 if (sr->cnt() > candidate->cnt()) {
694 candidate = sr;
695 }
696 }
697 SwitchRange most_freq = *candidate;
698 if (most_freq.cnt() == 0) {
699 break;
700 }
701
702 // Copy remaining ranges into another array
703 int shift = 0;
704 for (uint i = 0; i < nr; i++) {
705 SwitchRange* sr = &lo[i];
706 if (sr != candidate) {
707 ranges[i-shift] = *sr;
708 } else {
709 shift++;
710 if (i > 0 && i < nr-1) {
711 SwitchRange prev = lo[i-1];
712 prev.setRange(prev.lo(), sr->hi(), prev.dest(), prev.cnt());
713 if (prev.adjoin(lo[i+1])) {
714 shift++;
715 i++;
716 }
717 ranges[i-shift] = prev;
718 }
719 }
720 }
721 nr -= shift;
722
723 // Evaluate cost of testing the most common range and performing a
724 // binary search on the other ranges
725 float cost = extra + compute_tree_cost(&ranges[0], &ranges[nr-1], total_cnt);
726 if (cost >= min) {
727 break;
728 }
729 // swap arrays
730 lo = &ranges[0];
731 hi = &ranges[nr-1];
732
733 // It pays off: emit the test for the most common range
734 assert(most_freq.cnt() > 0, "must be taken");
735 Node* val = _gvn.transform(new SubINode(key_val, _gvn.intcon(most_freq.lo())));
736 Node* cmp = _gvn.transform(new CmpUNode(val, _gvn.intcon(most_freq.hi() - most_freq.lo())));
737 Node* tst = _gvn.transform(new BoolNode(cmp, BoolTest::le));
738 IfNode* iff = create_and_map_if(control(), tst, if_prob(most_freq.cnt(), total_cnt), if_cnt(most_freq.cnt()));
739 jump_if_true_fork(iff, most_freq.dest(), false);
740
741 sub += most_freq.cnt() / total_cnt;
742 extra += 1 - sub;
743 min = cost;
744 }
745 }
746
747 //----------------------------create_jump_tables-------------------------------
748 bool Parse::create_jump_tables(Node* key_val, SwitchRange* lo, SwitchRange* hi) {
749 // Are jumptables enabled
750 if (!UseJumpTables) return false;
751
752 // Are jumptables supported
753 if (!Matcher::has_match_rule(Op_Jump)) return false;
754
755 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if);
756
757 // Decide if a guard is needed to lop off big ranges at either (or
758 // both) end(s) of the input set. We'll call this the default target
759 // even though we can't be sure that it is the true "default".
760
761 bool needs_guard = false;
762 int default_dest;
763 int64_t total_outlier_size = 0;
764 int64_t hi_size = ((int64_t)hi->hi()) - ((int64_t)hi->lo()) + 1;
765 int64_t lo_size = ((int64_t)lo->hi()) - ((int64_t)lo->lo()) + 1;
766
767 if (lo->dest() == hi->dest()) {
768 total_outlier_size = hi_size + lo_size;
769 default_dest = lo->dest();
770 } else if (lo_size > hi_size) {
771 total_outlier_size = lo_size;
772 default_dest = lo->dest();
773 } else {
774 total_outlier_size = hi_size;
775 default_dest = hi->dest();
776 }
777
778 float total = sum_of_cnts(lo, hi);
779 float cost = compute_tree_cost(lo, hi, total);
780
781 // If a guard test will eliminate very sparse end ranges, then
782 // it is worth the cost of an extra jump.
783 float trimmed_cnt = 0;
784 if (total_outlier_size > (MaxJumpTableSparseness * 4)) {
785 needs_guard = true;
786 if (default_dest == lo->dest()) {
787 trimmed_cnt += lo->cnt();
788 lo++;
789 }
790 if (default_dest == hi->dest()) {
791 trimmed_cnt += hi->cnt();
792 hi--;
793 }
794 }
795
796 // Find the total number of cases and ranges
797 int64_t num_cases = ((int64_t)hi->hi()) - ((int64_t)lo->lo()) + 1;
798 int num_range = hi - lo + 1;
799
800 // Don't create table if: too large, too small, or too sparse.
801 if (num_cases > MaxJumpTableSize)
802 return false;
803 if (UseSwitchProfiling) {
804 // MinJumpTableSize is set so with a well balanced binary tree,
805 // when the number of ranges is MinJumpTableSize, it's cheaper to
806 // go through a JumpNode that a tree of IfNodes. Average cost of a
807 // tree of IfNodes with MinJumpTableSize is
808 // log2f(MinJumpTableSize) comparisons. So if the cost computed
809 // from profile data is less than log2f(MinJumpTableSize) then
810 // going with the binary search is cheaper.
811 if (cost < log2f(MinJumpTableSize)) {
812 return false;
813 }
814 } else {
815 if (num_cases < MinJumpTableSize)
816 return false;
817 }
818 if (num_cases > (MaxJumpTableSparseness * num_range))
819 return false;
820
821 // Normalize table lookups to zero
822 int lowval = lo->lo();
823 key_val = _gvn.transform( new SubINode(key_val, _gvn.intcon(lowval)) );
824
825 // Generate a guard to protect against input keyvals that aren't
826 // in the switch domain.
827 if (needs_guard) {
828 Node* size = _gvn.intcon(num_cases);
829 Node* cmp = _gvn.transform(new CmpUNode(key_val, size));
830 Node* tst = _gvn.transform(new BoolNode(cmp, BoolTest::ge));
831 IfNode* iff = create_and_map_if(control(), tst, if_prob(trimmed_cnt, total), if_cnt(trimmed_cnt));
832 jump_if_true_fork(iff, default_dest, trim_ranges && trimmed_cnt == 0);
833
834 total -= trimmed_cnt;
835 }
836
837 // Create an ideal node JumpTable that has projections
838 // of all possible ranges for a switch statement
839 // The key_val input must be converted to a pointer offset and scaled.
840 // Compare Parse::array_addressing above.
841
842 // Clean the 32-bit int into a real 64-bit offset.
843 // Otherwise, the jint value 0 might turn into an offset of 0x0800000000.
844 // Make I2L conversion control dependent to prevent it from
845 // floating above the range check during loop optimizations.
846 // Do not use a narrow int type here to prevent the data path from dying
847 // while the control path is not removed. This can happen if the type of key_val
848 // is later known to be out of bounds of [0, num_cases] and therefore a narrow cast
849 // would be replaced by TOP while C2 is not able to fold the corresponding range checks.
850 // Set _carry_dependency for the cast to avoid being removed by IGVN.
851 #ifdef _LP64
852 key_val = C->constrained_convI2L(&_gvn, key_val, TypeInt::INT, control(), true /* carry_dependency */);
853 #endif
854
855 // Shift the value by wordsize so we have an index into the table, rather
856 // than a switch value
857 Node *shiftWord = _gvn.MakeConX(wordSize);
858 key_val = _gvn.transform( new MulXNode( key_val, shiftWord));
859
860 // Create the JumpNode
861 Arena* arena = C->comp_arena();
862 float* probs = (float*)arena->Amalloc(sizeof(float)*num_cases);
863 int i = 0;
864 if (total == 0) {
865 for (SwitchRange* r = lo; r <= hi; r++) {
866 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) {
867 probs[i] = 1.0F / num_cases;
868 }
869 }
870 } else {
871 for (SwitchRange* r = lo; r <= hi; r++) {
872 float prob = r->cnt()/total;
873 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) {
874 probs[i] = prob / (r->hi() - r->lo() + 1);
875 }
876 }
877 }
878
879 ciMethodData* methodData = method()->method_data();
880 ciMultiBranchData* profile = NULL;
881 if (methodData->is_mature()) {
882 ciProfileData* data = methodData->bci_to_data(bci());
883 if (data != NULL && data->is_MultiBranchData()) {
884 profile = (ciMultiBranchData*)data;
885 }
886 }
887
888 Node* jtn = _gvn.transform(new JumpNode(control(), key_val, num_cases, probs, profile == NULL ? COUNT_UNKNOWN : total));
889
890 // These are the switch destinations hanging off the jumpnode
891 i = 0;
892 for (SwitchRange* r = lo; r <= hi; r++) {
893 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) {
894 Node* input = _gvn.transform(new JumpProjNode(jtn, i, r->dest(), (int)(j - lowval)));
895 {
896 PreserveJVMState pjvms(this);
897 set_control(input);
898 jump_if_always_fork(r->dest(), trim_ranges && r->cnt() == 0);
899 }
900 }
901 }
902 assert(i == num_cases, "miscount of cases");
903 stop_and_kill_map(); // no more uses for this JVMS
904 return true;
905 }
906
907 //----------------------------jump_switch_ranges-------------------------------
908 void Parse::jump_switch_ranges(Node* key_val, SwitchRange *lo, SwitchRange *hi, int switch_depth) {
909 Block* switch_block = block();
910 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if);
911
912 if (switch_depth == 0) {
913 // Do special processing for the top-level call.
914 assert(lo->lo() == min_jint, "initial range must exhaust Type::INT");
915 assert(hi->hi() == max_jint, "initial range must exhaust Type::INT");
916
917 // Decrement pred-numbers for the unique set of nodes.
918 #ifdef ASSERT
919 if (!trim_ranges) {
920 // Ensure that the block's successors are a (duplicate-free) set.
921 int successors_counted = 0; // block occurrences in [hi..lo]
922 int unique_successors = switch_block->num_successors();
923 for (int i = 0; i < unique_successors; i++) {
924 Block* target = switch_block->successor_at(i);
925
926 // Check that the set of successors is the same in both places.
927 int successors_found = 0;
928 for (SwitchRange* p = lo; p <= hi; p++) {
929 if (p->dest() == target->start()) successors_found++;
930 }
931 assert(successors_found > 0, "successor must be known");
932 successors_counted += successors_found;
933 }
934 assert(successors_counted == (hi-lo)+1, "no unexpected successors");
935 }
936 #endif
937
938 // Maybe prune the inputs, based on the type of key_val.
939 jint min_val = min_jint;
940 jint max_val = max_jint;
941 const TypeInt* ti = key_val->bottom_type()->isa_int();
942 if (ti != NULL) {
943 min_val = ti->_lo;
944 max_val = ti->_hi;
945 assert(min_val <= max_val, "invalid int type");
946 }
947 while (lo->hi() < min_val) {
948 lo++;
949 }
950 if (lo->lo() < min_val) {
951 lo->setRange(min_val, lo->hi(), lo->dest(), lo->cnt());
952 }
953 while (hi->lo() > max_val) {
954 hi--;
955 }
956 if (hi->hi() > max_val) {
957 hi->setRange(hi->lo(), max_val, hi->dest(), hi->cnt());
958 }
959
960 linear_search_switch_ranges(key_val, lo, hi);
961 }
962
963 #ifndef PRODUCT
964 if (switch_depth == 0) {
965 _max_switch_depth = 0;
966 _est_switch_depth = log2i_graceful((hi - lo + 1) - 1) + 1;
967 }
968 #endif
969
970 assert(lo <= hi, "must be a non-empty set of ranges");
971 if (lo == hi) {
972 jump_if_always_fork(lo->dest(), trim_ranges && lo->cnt() == 0);
973 } else {
974 assert(lo->hi() == (lo+1)->lo()-1, "contiguous ranges");
975 assert(hi->lo() == (hi-1)->hi()+1, "contiguous ranges");
976
977 if (create_jump_tables(key_val, lo, hi)) return;
978
979 SwitchRange* mid = NULL;
980 float total_cnt = sum_of_cnts(lo, hi);
981
982 int nr = hi - lo + 1;
983 if (UseSwitchProfiling) {
984 // Don't keep the binary search tree balanced: pick up mid point
985 // that split frequencies in half.
986 float cnt = 0;
987 for (SwitchRange* sr = lo; sr <= hi; sr++) {
988 cnt += sr->cnt();
989 if (cnt >= total_cnt / 2) {
990 mid = sr;
991 break;
992 }
993 }
994 } else {
995 mid = lo + nr/2;
996
997 // if there is an easy choice, pivot at a singleton:
998 if (nr > 3 && !mid->is_singleton() && (mid-1)->is_singleton()) mid--;
999
1000 assert(lo < mid && mid <= hi, "good pivot choice");
1001 assert(nr != 2 || mid == hi, "should pick higher of 2");
1002 assert(nr != 3 || mid == hi-1, "should pick middle of 3");
1003 }
1004
1005
1006 Node *test_val = _gvn.intcon(mid == lo ? mid->hi() : mid->lo());
1007
1008 if (mid->is_singleton()) {
1009 IfNode *iff_ne = jump_if_fork_int(key_val, test_val, BoolTest::ne, 1-if_prob(mid->cnt(), total_cnt), if_cnt(mid->cnt()));
1010 jump_if_false_fork(iff_ne, mid->dest(), trim_ranges && mid->cnt() == 0);
1011
1012 // Special Case: If there are exactly three ranges, and the high
1013 // and low range each go to the same place, omit the "gt" test,
1014 // since it will not discriminate anything.
1015 bool eq_test_only = (hi == lo+2 && hi->dest() == lo->dest() && mid == hi-1) || mid == lo;
1016
1017 // if there is a higher range, test for it and process it:
1018 if (mid < hi && !eq_test_only) {
1019 // two comparisons of same values--should enable 1 test for 2 branches
1020 // Use BoolTest::lt instead of BoolTest::gt
1021 float cnt = sum_of_cnts(lo, mid-1);
1022 IfNode *iff_lt = jump_if_fork_int(key_val, test_val, BoolTest::lt, if_prob(cnt, total_cnt), if_cnt(cnt));
1023 Node *iftrue = _gvn.transform( new IfTrueNode(iff_lt) );
1024 Node *iffalse = _gvn.transform( new IfFalseNode(iff_lt) );
1025 { PreserveJVMState pjvms(this);
1026 set_control(iffalse);
1027 jump_switch_ranges(key_val, mid+1, hi, switch_depth+1);
1028 }
1029 set_control(iftrue);
1030 }
1031
1032 } else {
1033 // mid is a range, not a singleton, so treat mid..hi as a unit
1034 float cnt = sum_of_cnts(mid == lo ? mid+1 : mid, hi);
1035 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));
1036
1037 // if there is a higher range, test for it and process it:
1038 if (mid == hi) {
1039 jump_if_true_fork(iff_ge, mid->dest(), trim_ranges && cnt == 0);
1040 } else {
1041 Node *iftrue = _gvn.transform( new IfTrueNode(iff_ge) );
1042 Node *iffalse = _gvn.transform( new IfFalseNode(iff_ge) );
1043 { PreserveJVMState pjvms(this);
1044 set_control(iftrue);
1045 jump_switch_ranges(key_val, mid == lo ? mid+1 : mid, hi, switch_depth+1);
1046 }
1047 set_control(iffalse);
1048 }
1049 }
1050
1051 // in any case, process the lower range
1052 if (mid == lo) {
1053 if (mid->is_singleton()) {
1054 jump_switch_ranges(key_val, lo+1, hi, switch_depth+1);
1055 } else {
1056 jump_if_always_fork(lo->dest(), trim_ranges && lo->cnt() == 0);
1057 }
1058 } else {
1059 jump_switch_ranges(key_val, lo, mid-1, switch_depth+1);
1060 }
1061 }
1062
1063 // Decrease pred_count for each successor after all is done.
1064 if (switch_depth == 0) {
1065 int unique_successors = switch_block->num_successors();
1066 for (int i = 0; i < unique_successors; i++) {
1067 Block* target = switch_block->successor_at(i);
1068 // Throw away the pre-allocated path for each unique successor.
1069 target->next_path_num();
1070 }
1071 }
1072
1073 #ifndef PRODUCT
1074 _max_switch_depth = MAX2(switch_depth, _max_switch_depth);
1075 if (TraceOptoParse && Verbose && WizardMode && switch_depth == 0) {
1076 SwitchRange* r;
1077 int nsing = 0;
1078 for( r = lo; r <= hi; r++ ) {
1079 if( r->is_singleton() ) nsing++;
1080 }
1081 tty->print(">>> ");
1082 _method->print_short_name();
1083 tty->print_cr(" switch decision tree");
1084 tty->print_cr(" %d ranges (%d singletons), max_depth=%d, est_depth=%d",
1085 (int) (hi-lo+1), nsing, _max_switch_depth, _est_switch_depth);
1086 if (_max_switch_depth > _est_switch_depth) {
1087 tty->print_cr("******** BAD SWITCH DEPTH ********");
1088 }
1089 tty->print(" ");
1090 for( r = lo; r <= hi; r++ ) {
1091 r->print();
1092 }
1093 tty->cr();
1094 }
1095 #endif
1096 }
1097
1098 void Parse::modf() {
1099 Node *f2 = pop();
1100 Node *f1 = pop();
1101 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::modf_Type(),
1102 CAST_FROM_FN_PTR(address, SharedRuntime::frem),
1103 "frem", NULL, //no memory effects
1104 f1, f2);
1105 Node* res = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0));
1106
1107 push(res);
1108 }
1109
1110 void Parse::modd() {
1111 Node *d2 = pop_pair();
1112 Node *d1 = pop_pair();
1113 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::Math_DD_D_Type(),
1114 CAST_FROM_FN_PTR(address, SharedRuntime::drem),
1115 "drem", NULL, //no memory effects
1116 d1, top(), d2, top());
1117 Node* res_d = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0));
1118
1119 #ifdef ASSERT
1120 Node* res_top = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 1));
1121 assert(res_top == top(), "second value must be top");
1122 #endif
1123
1124 push_pair(res_d);
1125 }
1126
1127 void Parse::l2f() {
1128 Node* f2 = pop();
1129 Node* f1 = pop();
1130 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::l2f_Type(),
1131 CAST_FROM_FN_PTR(address, SharedRuntime::l2f),
1132 "l2f", NULL, //no memory effects
1133 f1, f2);
1134 Node* res = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0));
1135
1136 push(res);
1137 }
1138
1139 // Handle jsr and jsr_w bytecode
1140 void Parse::do_jsr() {
1141 assert(bc() == Bytecodes::_jsr || bc() == Bytecodes::_jsr_w, "wrong bytecode");
1142
1143 // Store information about current state, tagged with new _jsr_bci
1144 int return_bci = iter().next_bci();
1145 int jsr_bci = (bc() == Bytecodes::_jsr) ? iter().get_dest() : iter().get_far_dest();
1146
1147 // The way we do things now, there is only one successor block
1148 // for the jsr, because the target code is cloned by ciTypeFlow.
1149 Block* target = successor_for_bci(jsr_bci);
1150
1151 // What got pushed?
1152 const Type* ret_addr = target->peek();
1153 assert(ret_addr->singleton(), "must be a constant (cloned jsr body)");
1154
1155 // Effect on jsr on stack
1156 push(_gvn.makecon(ret_addr));
1157
1158 // Flow to the jsr.
1159 merge(jsr_bci);
1160 }
1161
1162 // Handle ret bytecode
1163 void Parse::do_ret() {
1164 // Find to whom we return.
1165 assert(block()->num_successors() == 1, "a ret can only go one place now");
1166 Block* target = block()->successor_at(0);
1167 assert(!target->is_ready(), "our arrival must be expected");
1168 int pnum = target->next_path_num();
1169 merge_common(target, pnum);
1170 }
1171
1172 static bool has_injected_profile(BoolTest::mask btest, Node* test, int& taken, int& not_taken) {
1173 if (btest != BoolTest::eq && btest != BoolTest::ne) {
1174 // Only ::eq and ::ne are supported for profile injection.
1175 return false;
1176 }
1177 if (test->is_Cmp() &&
1178 test->in(1)->Opcode() == Op_ProfileBoolean) {
1179 ProfileBooleanNode* profile = (ProfileBooleanNode*)test->in(1);
1180 int false_cnt = profile->false_count();
1181 int true_cnt = profile->true_count();
1182
1183 // Counts matching depends on the actual test operation (::eq or ::ne).
1184 // No need to scale the counts because profile injection was designed
1185 // to feed exact counts into VM.
1186 taken = (btest == BoolTest::eq) ? false_cnt : true_cnt;
1187 not_taken = (btest == BoolTest::eq) ? true_cnt : false_cnt;
1188
1189 profile->consume();
1190 return true;
1191 }
1192 return false;
1193 }
1194 //--------------------------dynamic_branch_prediction--------------------------
1195 // Try to gather dynamic branch prediction behavior. Return a probability
1196 // of the branch being taken and set the "cnt" field. Returns a -1.0
1197 // if we need to use static prediction for some reason.
1198 float Parse::dynamic_branch_prediction(float &cnt, BoolTest::mask btest, Node* test) {
1199 ResourceMark rm;
1200
1201 cnt = COUNT_UNKNOWN;
1202
1203 int taken = 0;
1204 int not_taken = 0;
1205
1206 bool use_mdo = !has_injected_profile(btest, test, taken, not_taken);
1207
1208 if (use_mdo) {
1209 // Use MethodData information if it is available
1210 // FIXME: free the ProfileData structure
1211 ciMethodData* methodData = method()->method_data();
1212 if (!methodData->is_mature()) return PROB_UNKNOWN;
1213 ciProfileData* data = methodData->bci_to_data(bci());
1214 if (data == NULL) {
1215 return PROB_UNKNOWN;
1216 }
1217 if (!data->is_JumpData()) return PROB_UNKNOWN;
1218
1219 // get taken and not taken values
1220 taken = data->as_JumpData()->taken();
1221 not_taken = 0;
1222 if (data->is_BranchData()) {
1223 not_taken = data->as_BranchData()->not_taken();
1224 }
1225
1226 // scale the counts to be commensurate with invocation counts:
1227 taken = method()->scale_count(taken);
1228 not_taken = method()->scale_count(not_taken);
1229 }
1230
1231 // Give up if too few (or too many, in which case the sum will overflow) counts to be meaningful.
1232 // We also check that individual counters are positive first, otherwise the sum can become positive.
1233 if (taken < 0 || not_taken < 0 || taken + not_taken < 40) {
1234 if (C->log() != NULL) {
1235 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d'", iter().get_dest(), taken, not_taken);
1236 }
1237 return PROB_UNKNOWN;
1238 }
1239
1240 // Compute frequency that we arrive here
1241 float sum = taken + not_taken;
1242 // Adjust, if this block is a cloned private block but the
1243 // Jump counts are shared. Taken the private counts for
1244 // just this path instead of the shared counts.
1245 if( block()->count() > 0 )
1246 sum = block()->count();
1247 cnt = sum / FreqCountInvocations;
1248
1249 // Pin probability to sane limits
1250 float prob;
1251 if( !taken )
1252 prob = (0+PROB_MIN) / 2;
1253 else if( !not_taken )
1254 prob = (1+PROB_MAX) / 2;
1255 else { // Compute probability of true path
1256 prob = (float)taken / (float)(taken + not_taken);
1257 if (prob > PROB_MAX) prob = PROB_MAX;
1258 if (prob < PROB_MIN) prob = PROB_MIN;
1259 }
1260
1261 assert((cnt > 0.0f) && (prob > 0.0f),
1262 "Bad frequency assignment in if");
1263
1264 if (C->log() != NULL) {
1265 const char* prob_str = NULL;
1266 if (prob >= PROB_MAX) prob_str = (prob == PROB_MAX) ? "max" : "always";
1267 if (prob <= PROB_MIN) prob_str = (prob == PROB_MIN) ? "min" : "never";
1268 char prob_str_buf[30];
1269 if (prob_str == NULL) {
1270 jio_snprintf(prob_str_buf, sizeof(prob_str_buf), "%20.2f", prob);
1271 prob_str = prob_str_buf;
1272 }
1273 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d' cnt='%f' prob='%s'",
1274 iter().get_dest(), taken, not_taken, cnt, prob_str);
1275 }
1276 return prob;
1277 }
1278
1279 //-----------------------------branch_prediction-------------------------------
1280 float Parse::branch_prediction(float& cnt,
1281 BoolTest::mask btest,
1282 int target_bci,
1283 Node* test) {
1284 float prob = dynamic_branch_prediction(cnt, btest, test);
1285 // If prob is unknown, switch to static prediction
1286 if (prob != PROB_UNKNOWN) return prob;
1287
1288 prob = PROB_FAIR; // Set default value
1289 if (btest == BoolTest::eq) // Exactly equal test?
1290 prob = PROB_STATIC_INFREQUENT; // Assume its relatively infrequent
1291 else if (btest == BoolTest::ne)
1292 prob = PROB_STATIC_FREQUENT; // Assume its relatively frequent
1293
1294 // If this is a conditional test guarding a backwards branch,
1295 // assume its a loop-back edge. Make it a likely taken branch.
1296 if (target_bci < bci()) {
1297 if (is_osr_parse()) { // Could be a hot OSR'd loop; force deopt
1298 // Since it's an OSR, we probably have profile data, but since
1299 // branch_prediction returned PROB_UNKNOWN, the counts are too small.
1300 // Let's make a special check here for completely zero counts.
1301 ciMethodData* methodData = method()->method_data();
1302 if (!methodData->is_empty()) {
1303 ciProfileData* data = methodData->bci_to_data(bci());
1304 // Only stop for truly zero counts, which mean an unknown part
1305 // of the OSR-ed method, and we want to deopt to gather more stats.
1306 // If you have ANY counts, then this loop is simply 'cold' relative
1307 // to the OSR loop.
1308 if (data == NULL ||
1309 (data->as_BranchData()->taken() + data->as_BranchData()->not_taken() == 0)) {
1310 // This is the only way to return PROB_UNKNOWN:
1311 return PROB_UNKNOWN;
1312 }
1313 }
1314 }
1315 prob = PROB_STATIC_FREQUENT; // Likely to take backwards branch
1316 }
1317
1318 assert(prob != PROB_UNKNOWN, "must have some guess at this point");
1319 return prob;
1320 }
1321
1322 // The magic constants are chosen so as to match the output of
1323 // branch_prediction() when the profile reports a zero taken count.
1324 // It is important to distinguish zero counts unambiguously, because
1325 // some branches (e.g., _213_javac.Assembler.eliminate) validly produce
1326 // very small but nonzero probabilities, which if confused with zero
1327 // counts would keep the program recompiling indefinitely.
1328 bool Parse::seems_never_taken(float prob) const {
1329 return prob < PROB_MIN;
1330 }
1331
1332 // True if the comparison seems to be the kind that will not change its
1333 // statistics from true to false. See comments in adjust_map_after_if.
1334 // This question is only asked along paths which are already
1335 // classifed as untaken (by seems_never_taken), so really,
1336 // if a path is never taken, its controlling comparison is
1337 // already acting in a stable fashion. If the comparison
1338 // seems stable, we will put an expensive uncommon trap
1339 // on the untaken path.
1340 bool Parse::seems_stable_comparison() const {
1341 if (C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if)) {
1342 return false;
1343 }
1344 return true;
1345 }
1346
1347 //-------------------------------repush_if_args--------------------------------
1348 // Push arguments of an "if" bytecode back onto the stack by adjusting _sp.
1349 inline int Parse::repush_if_args() {
1350 if (PrintOpto && WizardMode) {
1351 tty->print("defending against excessive implicit null exceptions on %s @%d in ",
1352 Bytecodes::name(iter().cur_bc()), iter().cur_bci());
1353 method()->print_name(); tty->cr();
1354 }
1355 int bc_depth = - Bytecodes::depth(iter().cur_bc());
1356 assert(bc_depth == 1 || bc_depth == 2, "only two kinds of branches");
1357 DEBUG_ONLY(sync_jvms()); // argument(n) requires a synced jvms
1358 assert(argument(0) != NULL, "must exist");
1359 assert(bc_depth == 1 || argument(1) != NULL, "two must exist");
1360 inc_sp(bc_depth);
1361 return bc_depth;
1362 }
1363
1364 //----------------------------------do_ifnull----------------------------------
1365 void Parse::do_ifnull(BoolTest::mask btest, Node *c) {
1366 int target_bci = iter().get_dest();
1367
1368 Block* branch_block = successor_for_bci(target_bci);
1369 Block* next_block = successor_for_bci(iter().next_bci());
1370
1371 float cnt;
1372 float prob = branch_prediction(cnt, btest, target_bci, c);
1373 if (prob == PROB_UNKNOWN) {
1374 // (An earlier version of do_ifnull omitted this trap for OSR methods.)
1375 if (PrintOpto && Verbose) {
1376 tty->print_cr("Never-taken edge stops compilation at bci %d", bci());
1377 }
1378 repush_if_args(); // to gather stats on loop
1379 uncommon_trap(Deoptimization::Reason_unreached,
1380 Deoptimization::Action_reinterpret,
1381 NULL, "cold");
1382 if (C->eliminate_boxing()) {
1383 // Mark the successor blocks as parsed
1384 branch_block->next_path_num();
1385 next_block->next_path_num();
1386 }
1387 return;
1388 }
1389
1390 NOT_PRODUCT(explicit_null_checks_inserted++);
1391
1392 // Generate real control flow
1393 Node *tst = _gvn.transform( new BoolNode( c, btest ) );
1394
1395 // Sanity check the probability value
1396 assert(prob > 0.0f,"Bad probability in Parser");
1397 // Need xform to put node in hash table
1398 IfNode *iff = create_and_xform_if( control(), tst, prob, cnt );
1399 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser");
1400 // True branch
1401 { PreserveJVMState pjvms(this);
1402 Node* iftrue = _gvn.transform( new IfTrueNode (iff) );
1403 set_control(iftrue);
1404
1405 if (stopped()) { // Path is dead?
1406 NOT_PRODUCT(explicit_null_checks_elided++);
1407 if (C->eliminate_boxing()) {
1408 // Mark the successor block as parsed
1409 branch_block->next_path_num();
1410 }
1411 } else { // Path is live.
1412 adjust_map_after_if(btest, c, prob, branch_block, next_block);
1413 if (!stopped()) {
1414 merge(target_bci);
1415 }
1416 }
1417 }
1418
1419 // False branch
1420 Node* iffalse = _gvn.transform( new IfFalseNode(iff) );
1421 set_control(iffalse);
1422
1423 if (stopped()) { // Path is dead?
1424 NOT_PRODUCT(explicit_null_checks_elided++);
1425 if (C->eliminate_boxing()) {
1426 // Mark the successor block as parsed
1427 next_block->next_path_num();
1428 }
1429 } else { // Path is live.
1430 adjust_map_after_if(BoolTest(btest).negate(), c, 1.0-prob,
1431 next_block, branch_block);
1432 }
1433 }
1434
1435 //------------------------------------do_if------------------------------------
1436 void Parse::do_if(BoolTest::mask btest, Node* c) {
1437 int target_bci = iter().get_dest();
1438
1439 Block* branch_block = successor_for_bci(target_bci);
1440 Block* next_block = successor_for_bci(iter().next_bci());
1441
1442 float cnt;
1443 float prob = branch_prediction(cnt, btest, target_bci, c);
1444 float untaken_prob = 1.0 - prob;
1445
1446 if (prob == PROB_UNKNOWN) {
1447 if (PrintOpto && Verbose) {
1448 tty->print_cr("Never-taken edge stops compilation at bci %d", bci());
1449 }
1450 repush_if_args(); // to gather stats on loop
1451 uncommon_trap(Deoptimization::Reason_unreached,
1452 Deoptimization::Action_reinterpret,
1453 NULL, "cold");
1454 if (C->eliminate_boxing()) {
1455 // Mark the successor blocks as parsed
1456 branch_block->next_path_num();
1457 next_block->next_path_num();
1458 }
1459 return;
1460 }
1461
1462 // Sanity check the probability value
1463 assert(0.0f < prob && prob < 1.0f,"Bad probability in Parser");
1464
1465 bool taken_if_true = true;
1466 // Convert BoolTest to canonical form:
1467 if (!BoolTest(btest).is_canonical()) {
1468 btest = BoolTest(btest).negate();
1469 taken_if_true = false;
1470 // prob is NOT updated here; it remains the probability of the taken
1471 // path (as opposed to the prob of the path guarded by an 'IfTrueNode').
1472 }
1473 assert(btest != BoolTest::eq, "!= is the only canonical exact test");
1474
1475 Node* tst0 = new BoolNode(c, btest);
1476 Node* tst = _gvn.transform(tst0);
1477 BoolTest::mask taken_btest = BoolTest::illegal;
1478 BoolTest::mask untaken_btest = BoolTest::illegal;
1479
1480 if (tst->is_Bool()) {
1481 // Refresh c from the transformed bool node, since it may be
1482 // simpler than the original c. Also re-canonicalize btest.
1483 // This wins when (Bool ne (Conv2B p) 0) => (Bool ne (CmpP p NULL)).
1484 // That can arise from statements like: if (x instanceof C) ...
1485 if (tst != tst0) {
1486 // Canonicalize one more time since transform can change it.
1487 btest = tst->as_Bool()->_test._test;
1488 if (!BoolTest(btest).is_canonical()) {
1489 // Reverse edges one more time...
1490 tst = _gvn.transform( tst->as_Bool()->negate(&_gvn) );
1491 btest = tst->as_Bool()->_test._test;
1492 assert(BoolTest(btest).is_canonical(), "sanity");
1493 taken_if_true = !taken_if_true;
1494 }
1495 c = tst->in(1);
1496 }
1497 BoolTest::mask neg_btest = BoolTest(btest).negate();
1498 taken_btest = taken_if_true ? btest : neg_btest;
1499 untaken_btest = taken_if_true ? neg_btest : btest;
1500 }
1501
1502 // Generate real control flow
1503 float true_prob = (taken_if_true ? prob : untaken_prob);
1504 IfNode* iff = create_and_map_if(control(), tst, true_prob, cnt);
1505 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser");
1506 Node* taken_branch = new IfTrueNode(iff);
1507 Node* untaken_branch = new IfFalseNode(iff);
1508 if (!taken_if_true) { // Finish conversion to canonical form
1509 Node* tmp = taken_branch;
1510 taken_branch = untaken_branch;
1511 untaken_branch = tmp;
1512 }
1513
1514 // Branch is taken:
1515 { PreserveJVMState pjvms(this);
1516 taken_branch = _gvn.transform(taken_branch);
1517 set_control(taken_branch);
1518
1519 if (stopped()) {
1520 if (C->eliminate_boxing()) {
1521 // Mark the successor block as parsed
1522 branch_block->next_path_num();
1523 }
1524 } else {
1525 adjust_map_after_if(taken_btest, c, prob, branch_block, next_block);
1526 if (!stopped()) {
1527 merge(target_bci);
1528 }
1529 }
1530 }
1531
1532 untaken_branch = _gvn.transform(untaken_branch);
1533 set_control(untaken_branch);
1534
1535 // Branch not taken.
1536 if (stopped()) {
1537 if (C->eliminate_boxing()) {
1538 // Mark the successor block as parsed
1539 next_block->next_path_num();
1540 }
1541 } else {
1542 adjust_map_after_if(untaken_btest, c, untaken_prob,
1543 next_block, branch_block);
1544 }
1545 }
1546
1547 bool Parse::path_is_suitable_for_uncommon_trap(float prob) const {
1548 // Don't want to speculate on uncommon traps when running with -Xcomp
1549 if (!UseInterpreter) {
1550 return false;
1551 }
1552 return (seems_never_taken(prob) && seems_stable_comparison());
1553 }
1554
1555 void Parse::maybe_add_predicate_after_if(Block* path) {
1556 if (path->is_SEL_head() && path->preds_parsed() == 0) {
1557 // Add predicates at bci of if dominating the loop so traps can be
1558 // recorded on the if's profile data
1559 int bc_depth = repush_if_args();
1560 add_empty_predicates();
1561 dec_sp(bc_depth);
1562 path->set_has_predicates();
1563 }
1564 }
1565
1566
1567 //----------------------------adjust_map_after_if------------------------------
1568 // Adjust the JVM state to reflect the result of taking this path.
1569 // Basically, it means inspecting the CmpNode controlling this
1570 // branch, seeing how it constrains a tested value, and then
1571 // deciding if it's worth our while to encode this constraint
1572 // as graph nodes in the current abstract interpretation map.
1573 void Parse::adjust_map_after_if(BoolTest::mask btest, Node* c, float prob,
1574 Block* path, Block* other_path) {
1575 if (!c->is_Cmp()) {
1576 maybe_add_predicate_after_if(path);
1577 return;
1578 }
1579
1580 if (stopped() || btest == BoolTest::illegal) {
1581 return; // nothing to do
1582 }
1583
1584 bool is_fallthrough = (path == successor_for_bci(iter().next_bci()));
1585
1586 if (path_is_suitable_for_uncommon_trap(prob)) {
1587 repush_if_args();
1588 uncommon_trap(Deoptimization::Reason_unstable_if,
1589 Deoptimization::Action_reinterpret,
1590 NULL,
1591 (is_fallthrough ? "taken always" : "taken never"));
1592 return;
1593 }
1594
1595 Node* val = c->in(1);
1596 Node* con = c->in(2);
1597 const Type* tcon = _gvn.type(con);
1598 const Type* tval = _gvn.type(val);
1599 bool have_con = tcon->singleton();
1600 if (tval->singleton()) {
1601 if (!have_con) {
1602 // Swap, so constant is in con.
1603 con = val;
1604 tcon = tval;
1605 val = c->in(2);
1606 tval = _gvn.type(val);
1607 btest = BoolTest(btest).commute();
1608 have_con = true;
1609 } else {
1610 // Do we have two constants? Then leave well enough alone.
1611 have_con = false;
1612 }
1613 }
1614 if (!have_con) { // remaining adjustments need a con
1615 maybe_add_predicate_after_if(path);
1616 return;
1617 }
1618
1619 sharpen_type_after_if(btest, con, tcon, val, tval);
1620 maybe_add_predicate_after_if(path);
1621 }
1622
1623
1624 static Node* extract_obj_from_klass_load(PhaseGVN* gvn, Node* n) {
1625 Node* ldk;
1626 if (n->is_DecodeNKlass()) {
1627 if (n->in(1)->Opcode() != Op_LoadNKlass) {
1628 return NULL;
1629 } else {
1630 ldk = n->in(1);
1631 }
1632 } else if (n->Opcode() != Op_LoadKlass) {
1633 return NULL;
1634 } else {
1635 ldk = n;
1636 }
1637 assert(ldk != NULL && ldk->is_Load(), "should have found a LoadKlass or LoadNKlass node");
1638
1639 Node* adr = ldk->in(MemNode::Address);
1640 intptr_t off = 0;
1641 Node* obj = AddPNode::Ideal_base_and_offset(adr, gvn, off);
1642 if (obj == NULL || off != oopDesc::klass_offset_in_bytes()) // loading oopDesc::_klass?
1643 return NULL;
1644 const TypePtr* tp = gvn->type(obj)->is_ptr();
1645 if (tp == NULL || !(tp->isa_instptr() || tp->isa_aryptr())) // is obj a Java object ptr?
1646 return NULL;
1647
1648 return obj;
1649 }
1650
1651 void Parse::sharpen_type_after_if(BoolTest::mask btest,
1652 Node* con, const Type* tcon,
1653 Node* val, const Type* tval) {
1654 // Look for opportunities to sharpen the type of a node
1655 // whose klass is compared with a constant klass.
1656 if (btest == BoolTest::eq && tcon->isa_klassptr()) {
1657 Node* obj = extract_obj_from_klass_load(&_gvn, val);
1658 const TypeOopPtr* con_type = tcon->isa_klassptr()->as_instance_type();
1659 if (obj != NULL && (con_type->isa_instptr() || con_type->isa_aryptr())) {
1660 // Found:
1661 // Bool(CmpP(LoadKlass(obj._klass), ConP(Foo.klass)), [eq])
1662 // or the narrowOop equivalent.
1663 const Type* obj_type = _gvn.type(obj);
1664 const TypeOopPtr* tboth = obj_type->join_speculative(con_type)->isa_oopptr();
1665 if (tboth != NULL && tboth->klass_is_exact() && tboth != obj_type &&
1666 tboth->higher_equal(obj_type)) {
1667 // obj has to be of the exact type Foo if the CmpP succeeds.
1668 int obj_in_map = map()->find_edge(obj);
1669 JVMState* jvms = this->jvms();
1670 if (obj_in_map >= 0 &&
1671 (jvms->is_loc(obj_in_map) || jvms->is_stk(obj_in_map))) {
1672 TypeNode* ccast = new CheckCastPPNode(control(), obj, tboth);
1673 const Type* tcc = ccast->as_Type()->type();
1674 assert(tcc != obj_type && tcc->higher_equal(obj_type), "must improve");
1675 // Delay transform() call to allow recovery of pre-cast value
1676 // at the control merge.
1677 _gvn.set_type_bottom(ccast);
1678 record_for_igvn(ccast);
1679 // Here's the payoff.
1680 replace_in_map(obj, ccast);
1681 }
1682 }
1683 }
1684 }
1685
1686 int val_in_map = map()->find_edge(val);
1687 if (val_in_map < 0) return; // replace_in_map would be useless
1688 {
1689 JVMState* jvms = this->jvms();
1690 if (!(jvms->is_loc(val_in_map) ||
1691 jvms->is_stk(val_in_map)))
1692 return; // again, it would be useless
1693 }
1694
1695 // Check for a comparison to a constant, and "know" that the compared
1696 // value is constrained on this path.
1697 assert(tcon->singleton(), "");
1698 ConstraintCastNode* ccast = NULL;
1699 Node* cast = NULL;
1700
1701 switch (btest) {
1702 case BoolTest::eq: // Constant test?
1703 {
1704 const Type* tboth = tcon->join_speculative(tval);
1705 if (tboth == tval) break; // Nothing to gain.
1706 if (tcon->isa_int()) {
1707 ccast = new CastIINode(val, tboth);
1708 } else if (tcon == TypePtr::NULL_PTR) {
1709 // Cast to null, but keep the pointer identity temporarily live.
1710 ccast = new CastPPNode(val, tboth);
1711 } else {
1712 const TypeF* tf = tcon->isa_float_constant();
1713 const TypeD* td = tcon->isa_double_constant();
1714 // Exclude tests vs float/double 0 as these could be
1715 // either +0 or -0. Just because you are equal to +0
1716 // doesn't mean you ARE +0!
1717 // Note, following code also replaces Long and Oop values.
1718 if ((!tf || tf->_f != 0.0) &&
1719 (!td || td->_d != 0.0))
1720 cast = con; // Replace non-constant val by con.
1721 }
1722 }
1723 break;
1724
1725 case BoolTest::ne:
1726 if (tcon == TypePtr::NULL_PTR) {
1727 cast = cast_not_null(val, false);
1728 }
1729 break;
1730
1731 default:
1732 // (At this point we could record int range types with CastII.)
1733 break;
1734 }
1735
1736 if (ccast != NULL) {
1737 const Type* tcc = ccast->as_Type()->type();
1738 assert(tcc != tval && tcc->higher_equal(tval), "must improve");
1739 // Delay transform() call to allow recovery of pre-cast value
1740 // at the control merge.
1741 ccast->set_req(0, control());
1742 _gvn.set_type_bottom(ccast);
1743 record_for_igvn(ccast);
1744 cast = ccast;
1745 }
1746
1747 if (cast != NULL) { // Here's the payoff.
1748 replace_in_map(val, cast);
1749 }
1750 }
1751
1752 /**
1753 * Use speculative type to optimize CmpP node: if comparison is
1754 * against the low level class, cast the object to the speculative
1755 * type if any. CmpP should then go away.
1756 *
1757 * @param c expected CmpP node
1758 * @return result of CmpP on object casted to speculative type
1759 *
1760 */
1761 Node* Parse::optimize_cmp_with_klass(Node* c) {
1762 // If this is transformed by the _gvn to a comparison with the low
1763 // level klass then we may be able to use speculation
1764 if (c->Opcode() == Op_CmpP &&
1765 (c->in(1)->Opcode() == Op_LoadKlass || c->in(1)->Opcode() == Op_DecodeNKlass) &&
1766 c->in(2)->is_Con()) {
1767 Node* load_klass = NULL;
1768 Node* decode = NULL;
1769 if (c->in(1)->Opcode() == Op_DecodeNKlass) {
1770 decode = c->in(1);
1771 load_klass = c->in(1)->in(1);
1772 } else {
1773 load_klass = c->in(1);
1774 }
1775 if (load_klass->in(2)->is_AddP()) {
1776 Node* addp = load_klass->in(2);
1777 Node* obj = addp->in(AddPNode::Address);
1778 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
1779 if (obj_type->speculative_type_not_null() != NULL) {
1780 ciKlass* k = obj_type->speculative_type();
1781 inc_sp(2);
1782 obj = maybe_cast_profiled_obj(obj, k);
1783 dec_sp(2);
1784 // Make the CmpP use the casted obj
1785 addp = basic_plus_adr(obj, addp->in(AddPNode::Offset));
1786 load_klass = load_klass->clone();
1787 load_klass->set_req(2, addp);
1788 load_klass = _gvn.transform(load_klass);
1789 if (decode != NULL) {
1790 decode = decode->clone();
1791 decode->set_req(1, load_klass);
1792 load_klass = _gvn.transform(decode);
1793 }
1794 c = c->clone();
1795 c->set_req(1, load_klass);
1796 c = _gvn.transform(c);
1797 }
1798 }
1799 }
1800 return c;
1801 }
1802
1803 //------------------------------do_one_bytecode--------------------------------
1804 // Parse this bytecode, and alter the Parsers JVM->Node mapping
1805 void Parse::do_one_bytecode() {
1806 Node *a, *b, *c, *d; // Handy temps
1807 BoolTest::mask btest;
1808 int i;
1809
1810 assert(!has_exceptions(), "bytecode entry state must be clear of throws");
1811
1812 if (C->check_node_count(NodeLimitFudgeFactor * 5,
1813 "out of nodes parsing method")) {
1814 return;
1815 }
1816
1817 #ifdef ASSERT
1818 // for setting breakpoints
1819 if (TraceOptoParse) {
1820 tty->print(" @");
1821 dump_bci(bci());
1822 tty->cr();
1823 }
1824 #endif
1825
1826 switch (bc()) {
1827 case Bytecodes::_nop:
1828 // do nothing
1829 break;
1830 case Bytecodes::_lconst_0:
1831 push_pair(longcon(0));
1832 break;
1833
1834 case Bytecodes::_lconst_1:
1835 push_pair(longcon(1));
1836 break;
1837
1838 case Bytecodes::_fconst_0:
1839 push(zerocon(T_FLOAT));
1840 break;
1841
1842 case Bytecodes::_fconst_1:
1843 push(makecon(TypeF::ONE));
1844 break;
1845
1846 case Bytecodes::_fconst_2:
1847 push(makecon(TypeF::make(2.0f)));
1848 break;
1849
1850 case Bytecodes::_dconst_0:
1851 push_pair(zerocon(T_DOUBLE));
1852 break;
1853
1854 case Bytecodes::_dconst_1:
1855 push_pair(makecon(TypeD::ONE));
1856 break;
1857
1858 case Bytecodes::_iconst_m1:push(intcon(-1)); break;
1859 case Bytecodes::_iconst_0: push(intcon( 0)); break;
1860 case Bytecodes::_iconst_1: push(intcon( 1)); break;
1861 case Bytecodes::_iconst_2: push(intcon( 2)); break;
1862 case Bytecodes::_iconst_3: push(intcon( 3)); break;
1863 case Bytecodes::_iconst_4: push(intcon( 4)); break;
1864 case Bytecodes::_iconst_5: push(intcon( 5)); break;
1865 case Bytecodes::_bipush: push(intcon(iter().get_constant_u1())); break;
1866 case Bytecodes::_sipush: push(intcon(iter().get_constant_u2())); break;
1867 case Bytecodes::_aconst_null: push(null()); break;
1868
1869 case Bytecodes::_ldc:
1870 case Bytecodes::_ldc_w:
1871 case Bytecodes::_ldc2_w: {
1872 ciConstant constant = iter().get_constant();
1873 if (constant.is_loaded()) {
1874 const Type* con_type = Type::make_from_constant(constant);
1875 if (con_type != NULL) {
1876 push_node(con_type->basic_type(), makecon(con_type));
1877 }
1878 } else {
1879 // If the constant is unresolved or in error state, run this BC in the interpreter.
1880 if (iter().is_in_error()) {
1881 uncommon_trap(Deoptimization::make_trap_request(Deoptimization::Reason_unhandled,
1882 Deoptimization::Action_none),
1883 NULL, "constant in error state", true /* must_throw */);
1884
1885 } else {
1886 int index = iter().get_constant_pool_index();
1887 uncommon_trap(Deoptimization::make_trap_request(Deoptimization::Reason_unloaded,
1888 Deoptimization::Action_reinterpret,
1889 index),
1890 NULL, "unresolved constant", false /* must_throw */);
1891 }
1892 }
1893 break;
1894 }
1895
1896 case Bytecodes::_aload_0:
1897 push( local(0) );
1898 break;
1899 case Bytecodes::_aload_1:
1900 push( local(1) );
1901 break;
1902 case Bytecodes::_aload_2:
1903 push( local(2) );
1904 break;
1905 case Bytecodes::_aload_3:
1906 push( local(3) );
1907 break;
1908 case Bytecodes::_aload:
1909 push( local(iter().get_index()) );
1910 break;
1911
1912 case Bytecodes::_fload_0:
1913 case Bytecodes::_iload_0:
1914 push( local(0) );
1915 break;
1916 case Bytecodes::_fload_1:
1917 case Bytecodes::_iload_1:
1918 push( local(1) );
1919 break;
1920 case Bytecodes::_fload_2:
1921 case Bytecodes::_iload_2:
1922 push( local(2) );
1923 break;
1924 case Bytecodes::_fload_3:
1925 case Bytecodes::_iload_3:
1926 push( local(3) );
1927 break;
1928 case Bytecodes::_fload:
1929 case Bytecodes::_iload:
1930 push( local(iter().get_index()) );
1931 break;
1932 case Bytecodes::_lload_0:
1933 push_pair_local( 0 );
1934 break;
1935 case Bytecodes::_lload_1:
1936 push_pair_local( 1 );
1937 break;
1938 case Bytecodes::_lload_2:
1939 push_pair_local( 2 );
1940 break;
1941 case Bytecodes::_lload_3:
1942 push_pair_local( 3 );
1943 break;
1944 case Bytecodes::_lload:
1945 push_pair_local( iter().get_index() );
1946 break;
1947
1948 case Bytecodes::_dload_0:
1949 push_pair_local(0);
1950 break;
1951 case Bytecodes::_dload_1:
1952 push_pair_local(1);
1953 break;
1954 case Bytecodes::_dload_2:
1955 push_pair_local(2);
1956 break;
1957 case Bytecodes::_dload_3:
1958 push_pair_local(3);
1959 break;
1960 case Bytecodes::_dload:
1961 push_pair_local(iter().get_index());
1962 break;
1963 case Bytecodes::_fstore_0:
1964 case Bytecodes::_istore_0:
1965 case Bytecodes::_astore_0:
1966 set_local( 0, pop() );
1967 break;
1968 case Bytecodes::_fstore_1:
1969 case Bytecodes::_istore_1:
1970 case Bytecodes::_astore_1:
1971 set_local( 1, pop() );
1972 break;
1973 case Bytecodes::_fstore_2:
1974 case Bytecodes::_istore_2:
1975 case Bytecodes::_astore_2:
1976 set_local( 2, pop() );
1977 break;
1978 case Bytecodes::_fstore_3:
1979 case Bytecodes::_istore_3:
1980 case Bytecodes::_astore_3:
1981 set_local( 3, pop() );
1982 break;
1983 case Bytecodes::_fstore:
1984 case Bytecodes::_istore:
1985 case Bytecodes::_astore:
1986 set_local( iter().get_index(), pop() );
1987 break;
1988 // long stores
1989 case Bytecodes::_lstore_0:
1990 set_pair_local( 0, pop_pair() );
1991 break;
1992 case Bytecodes::_lstore_1:
1993 set_pair_local( 1, pop_pair() );
1994 break;
1995 case Bytecodes::_lstore_2:
1996 set_pair_local( 2, pop_pair() );
1997 break;
1998 case Bytecodes::_lstore_3:
1999 set_pair_local( 3, pop_pair() );
2000 break;
2001 case Bytecodes::_lstore:
2002 set_pair_local( iter().get_index(), pop_pair() );
2003 break;
2004
2005 // double stores
2006 case Bytecodes::_dstore_0:
2007 set_pair_local( 0, dprecision_rounding(pop_pair()) );
2008 break;
2009 case Bytecodes::_dstore_1:
2010 set_pair_local( 1, dprecision_rounding(pop_pair()) );
2011 break;
2012 case Bytecodes::_dstore_2:
2013 set_pair_local( 2, dprecision_rounding(pop_pair()) );
2014 break;
2015 case Bytecodes::_dstore_3:
2016 set_pair_local( 3, dprecision_rounding(pop_pair()) );
2017 break;
2018 case Bytecodes::_dstore:
2019 set_pair_local( iter().get_index(), dprecision_rounding(pop_pair()) );
2020 break;
2021
2022 case Bytecodes::_pop: dec_sp(1); break;
2023 case Bytecodes::_pop2: dec_sp(2); break;
2024 case Bytecodes::_swap:
2025 a = pop();
2026 b = pop();
2027 push(a);
2028 push(b);
2029 break;
2030 case Bytecodes::_dup:
2031 a = pop();
2032 push(a);
2033 push(a);
2034 break;
2035 case Bytecodes::_dup_x1:
2036 a = pop();
2037 b = pop();
2038 push( a );
2039 push( b );
2040 push( a );
2041 break;
2042 case Bytecodes::_dup_x2:
2043 a = pop();
2044 b = pop();
2045 c = pop();
2046 push( a );
2047 push( c );
2048 push( b );
2049 push( a );
2050 break;
2051 case Bytecodes::_dup2:
2052 a = pop();
2053 b = pop();
2054 push( b );
2055 push( a );
2056 push( b );
2057 push( a );
2058 break;
2059
2060 case Bytecodes::_dup2_x1:
2061 // before: .. c, b, a
2062 // after: .. b, a, c, b, a
2063 // not tested
2064 a = pop();
2065 b = pop();
2066 c = pop();
2067 push( b );
2068 push( a );
2069 push( c );
2070 push( b );
2071 push( a );
2072 break;
2073 case Bytecodes::_dup2_x2:
2074 // before: .. d, c, b, a
2075 // after: .. b, a, d, c, b, a
2076 // not tested
2077 a = pop();
2078 b = pop();
2079 c = pop();
2080 d = pop();
2081 push( b );
2082 push( a );
2083 push( d );
2084 push( c );
2085 push( b );
2086 push( a );
2087 break;
2088
2089 case Bytecodes::_arraylength: {
2090 // Must do null-check with value on expression stack
2091 Node *ary = null_check(peek(), T_ARRAY);
2092 // Compile-time detect of null-exception?
2093 if (stopped()) return;
2094 a = pop();
2095 push(load_array_length(a));
2096 break;
2097 }
2098
2099 case Bytecodes::_baload: array_load(T_BYTE); break;
2100 case Bytecodes::_caload: array_load(T_CHAR); break;
2101 case Bytecodes::_iaload: array_load(T_INT); break;
2102 case Bytecodes::_saload: array_load(T_SHORT); break;
2103 case Bytecodes::_faload: array_load(T_FLOAT); break;
2104 case Bytecodes::_aaload: array_load(T_OBJECT); break;
2105 case Bytecodes::_laload: array_load(T_LONG); break;
2106 case Bytecodes::_daload: array_load(T_DOUBLE); break;
2107 case Bytecodes::_bastore: array_store(T_BYTE); break;
2108 case Bytecodes::_castore: array_store(T_CHAR); break;
2109 case Bytecodes::_iastore: array_store(T_INT); break;
2110 case Bytecodes::_sastore: array_store(T_SHORT); break;
2111 case Bytecodes::_fastore: array_store(T_FLOAT); break;
2112 case Bytecodes::_aastore: array_store(T_OBJECT); break;
2113 case Bytecodes::_lastore: array_store(T_LONG); break;
2114 case Bytecodes::_dastore: array_store(T_DOUBLE); break;
2115
2116 case Bytecodes::_getfield:
2117 do_getfield();
2118 break;
2119
2120 case Bytecodes::_getstatic:
2121 do_getstatic();
2122 break;
2123
2124 case Bytecodes::_putfield:
2125 do_putfield();
2126 break;
2127
2128 case Bytecodes::_putstatic:
2129 do_putstatic();
2130 break;
2131
2132 case Bytecodes::_irem:
2133 // Must keep both values on the expression-stack during null-check
2134 zero_check_int(peek());
2135 // Compile-time detect of null-exception?
2136 if (stopped()) return;
2137 b = pop();
2138 a = pop();
2139 push(_gvn.transform(new ModINode(control(), a, b)));
2140 break;
2141 case Bytecodes::_idiv:
2142 // Must keep both values on the expression-stack during null-check
2143 zero_check_int(peek());
2144 // Compile-time detect of null-exception?
2145 if (stopped()) return;
2146 b = pop();
2147 a = pop();
2148 push( _gvn.transform( new DivINode(control(),a,b) ) );
2149 break;
2150 case Bytecodes::_imul:
2151 b = pop(); a = pop();
2152 push( _gvn.transform( new MulINode(a,b) ) );
2153 break;
2154 case Bytecodes::_iadd:
2155 b = pop(); a = pop();
2156 push( _gvn.transform( new AddINode(a,b) ) );
2157 break;
2158 case Bytecodes::_ineg:
2159 a = pop();
2160 push( _gvn.transform( new SubINode(_gvn.intcon(0),a)) );
2161 break;
2162 case Bytecodes::_isub:
2163 b = pop(); a = pop();
2164 push( _gvn.transform( new SubINode(a,b) ) );
2165 break;
2166 case Bytecodes::_iand:
2167 b = pop(); a = pop();
2168 push( _gvn.transform( new AndINode(a,b) ) );
2169 break;
2170 case Bytecodes::_ior:
2171 b = pop(); a = pop();
2172 push( _gvn.transform( new OrINode(a,b) ) );
2173 break;
2174 case Bytecodes::_ixor:
2175 b = pop(); a = pop();
2176 push( _gvn.transform( new XorINode(a,b) ) );
2177 break;
2178 case Bytecodes::_ishl:
2179 b = pop(); a = pop();
2180 push( _gvn.transform( new LShiftINode(a,b) ) );
2181 break;
2182 case Bytecodes::_ishr:
2183 b = pop(); a = pop();
2184 push( _gvn.transform( new RShiftINode(a,b) ) );
2185 break;
2186 case Bytecodes::_iushr:
2187 b = pop(); a = pop();
2188 push( _gvn.transform( new URShiftINode(a,b) ) );
2189 break;
2190
2191 case Bytecodes::_fneg:
2192 a = pop();
2193 b = _gvn.transform(new NegFNode (a));
2194 push(b);
2195 break;
2196
2197 case Bytecodes::_fsub:
2198 b = pop();
2199 a = pop();
2200 c = _gvn.transform( new SubFNode(a,b) );
2201 d = precision_rounding(c);
2202 push( d );
2203 break;
2204
2205 case Bytecodes::_fadd:
2206 b = pop();
2207 a = pop();
2208 c = _gvn.transform( new AddFNode(a,b) );
2209 d = precision_rounding(c);
2210 push( d );
2211 break;
2212
2213 case Bytecodes::_fmul:
2214 b = pop();
2215 a = pop();
2216 c = _gvn.transform( new MulFNode(a,b) );
2217 d = precision_rounding(c);
2218 push( d );
2219 break;
2220
2221 case Bytecodes::_fdiv:
2222 b = pop();
2223 a = pop();
2224 c = _gvn.transform( new DivFNode(0,a,b) );
2225 d = precision_rounding(c);
2226 push( d );
2227 break;
2228
2229 case Bytecodes::_frem:
2230 if (Matcher::has_match_rule(Op_ModF)) {
2231 // Generate a ModF node.
2232 b = pop();
2233 a = pop();
2234 c = _gvn.transform( new ModFNode(0,a,b) );
2235 d = precision_rounding(c);
2236 push( d );
2237 }
2238 else {
2239 // Generate a call.
2240 modf();
2241 }
2242 break;
2243
2244 case Bytecodes::_fcmpl:
2245 b = pop();
2246 a = pop();
2247 c = _gvn.transform( new CmpF3Node( a, b));
2248 push(c);
2249 break;
2250 case Bytecodes::_fcmpg:
2251 b = pop();
2252 a = pop();
2253
2254 // Same as fcmpl but need to flip the unordered case. Swap the inputs,
2255 // which negates the result sign except for unordered. Flip the unordered
2256 // as well by using CmpF3 which implements unordered-lesser instead of
2257 // unordered-greater semantics. Finally, commute the result bits. Result
2258 // is same as using a CmpF3Greater except we did it with CmpF3 alone.
2259 c = _gvn.transform( new CmpF3Node( b, a));
2260 c = _gvn.transform( new SubINode(_gvn.intcon(0),c) );
2261 push(c);
2262 break;
2263
2264 case Bytecodes::_f2i:
2265 a = pop();
2266 push(_gvn.transform(new ConvF2INode(a)));
2267 break;
2268
2269 case Bytecodes::_d2i:
2270 a = pop_pair();
2271 b = _gvn.transform(new ConvD2INode(a));
2272 push( b );
2273 break;
2274
2275 case Bytecodes::_f2d:
2276 a = pop();
2277 b = _gvn.transform( new ConvF2DNode(a));
2278 push_pair( b );
2279 break;
2280
2281 case Bytecodes::_d2f:
2282 a = pop_pair();
2283 b = _gvn.transform( new ConvD2FNode(a));
2284 // This breaks _227_mtrt (speed & correctness) and _222_mpegaudio (speed)
2285 //b = _gvn.transform(new RoundFloatNode(0, b) );
2286 push( b );
2287 break;
2288
2289 case Bytecodes::_l2f:
2290 if (Matcher::convL2FSupported()) {
2291 a = pop_pair();
2292 b = _gvn.transform( new ConvL2FNode(a));
2293 // For x86_32.ad, FILD doesn't restrict precision to 24 or 53 bits.
2294 // Rather than storing the result into an FP register then pushing
2295 // out to memory to round, the machine instruction that implements
2296 // ConvL2D is responsible for rounding.
2297 // c = precision_rounding(b);
2298 push(b);
2299 } else {
2300 l2f();
2301 }
2302 break;
2303
2304 case Bytecodes::_l2d:
2305 a = pop_pair();
2306 b = _gvn.transform( new ConvL2DNode(a));
2307 // For x86_32.ad, rounding is always necessary (see _l2f above).
2308 // c = dprecision_rounding(b);
2309 push_pair(b);
2310 break;
2311
2312 case Bytecodes::_f2l:
2313 a = pop();
2314 b = _gvn.transform( new ConvF2LNode(a));
2315 push_pair(b);
2316 break;
2317
2318 case Bytecodes::_d2l:
2319 a = pop_pair();
2320 b = _gvn.transform( new ConvD2LNode(a));
2321 push_pair(b);
2322 break;
2323
2324 case Bytecodes::_dsub:
2325 b = pop_pair();
2326 a = pop_pair();
2327 c = _gvn.transform( new SubDNode(a,b) );
2328 d = dprecision_rounding(c);
2329 push_pair( d );
2330 break;
2331
2332 case Bytecodes::_dadd:
2333 b = pop_pair();
2334 a = pop_pair();
2335 c = _gvn.transform( new AddDNode(a,b) );
2336 d = dprecision_rounding(c);
2337 push_pair( d );
2338 break;
2339
2340 case Bytecodes::_dmul:
2341 b = pop_pair();
2342 a = pop_pair();
2343 c = _gvn.transform( new MulDNode(a,b) );
2344 d = dprecision_rounding(c);
2345 push_pair( d );
2346 break;
2347
2348 case Bytecodes::_ddiv:
2349 b = pop_pair();
2350 a = pop_pair();
2351 c = _gvn.transform( new DivDNode(0,a,b) );
2352 d = dprecision_rounding(c);
2353 push_pair( d );
2354 break;
2355
2356 case Bytecodes::_dneg:
2357 a = pop_pair();
2358 b = _gvn.transform(new NegDNode (a));
2359 push_pair(b);
2360 break;
2361
2362 case Bytecodes::_drem:
2363 if (Matcher::has_match_rule(Op_ModD)) {
2364 // Generate a ModD node.
2365 b = pop_pair();
2366 a = pop_pair();
2367 // a % b
2368
2369 c = _gvn.transform( new ModDNode(0,a,b) );
2370 d = dprecision_rounding(c);
2371 push_pair( d );
2372 }
2373 else {
2374 // Generate a call.
2375 modd();
2376 }
2377 break;
2378
2379 case Bytecodes::_dcmpl:
2380 b = pop_pair();
2381 a = pop_pair();
2382 c = _gvn.transform( new CmpD3Node( a, b));
2383 push(c);
2384 break;
2385
2386 case Bytecodes::_dcmpg:
2387 b = pop_pair();
2388 a = pop_pair();
2389 // Same as dcmpl but need to flip the unordered case.
2390 // Commute the inputs, which negates the result sign except for unordered.
2391 // Flip the unordered as well by using CmpD3 which implements
2392 // unordered-lesser instead of unordered-greater semantics.
2393 // Finally, negate the result bits. Result is same as using a
2394 // CmpD3Greater except we did it with CmpD3 alone.
2395 c = _gvn.transform( new CmpD3Node( b, a));
2396 c = _gvn.transform( new SubINode(_gvn.intcon(0),c) );
2397 push(c);
2398 break;
2399
2400
2401 // Note for longs -> lo word is on TOS, hi word is on TOS - 1
2402 case Bytecodes::_land:
2403 b = pop_pair();
2404 a = pop_pair();
2405 c = _gvn.transform( new AndLNode(a,b) );
2406 push_pair(c);
2407 break;
2408 case Bytecodes::_lor:
2409 b = pop_pair();
2410 a = pop_pair();
2411 c = _gvn.transform( new OrLNode(a,b) );
2412 push_pair(c);
2413 break;
2414 case Bytecodes::_lxor:
2415 b = pop_pair();
2416 a = pop_pair();
2417 c = _gvn.transform( new XorLNode(a,b) );
2418 push_pair(c);
2419 break;
2420
2421 case Bytecodes::_lshl:
2422 b = pop(); // the shift count
2423 a = pop_pair(); // value to be shifted
2424 c = _gvn.transform( new LShiftLNode(a,b) );
2425 push_pair(c);
2426 break;
2427 case Bytecodes::_lshr:
2428 b = pop(); // the shift count
2429 a = pop_pair(); // value to be shifted
2430 c = _gvn.transform( new RShiftLNode(a,b) );
2431 push_pair(c);
2432 break;
2433 case Bytecodes::_lushr:
2434 b = pop(); // the shift count
2435 a = pop_pair(); // value to be shifted
2436 c = _gvn.transform( new URShiftLNode(a,b) );
2437 push_pair(c);
2438 break;
2439 case Bytecodes::_lmul:
2440 b = pop_pair();
2441 a = pop_pair();
2442 c = _gvn.transform( new MulLNode(a,b) );
2443 push_pair(c);
2444 break;
2445
2446 case Bytecodes::_lrem:
2447 // Must keep both values on the expression-stack during null-check
2448 assert(peek(0) == top(), "long word order");
2449 zero_check_long(peek(1));
2450 // Compile-time detect of null-exception?
2451 if (stopped()) return;
2452 b = pop_pair();
2453 a = pop_pair();
2454 c = _gvn.transform( new ModLNode(control(),a,b) );
2455 push_pair(c);
2456 break;
2457
2458 case Bytecodes::_ldiv:
2459 // Must keep both values on the expression-stack during null-check
2460 assert(peek(0) == top(), "long word order");
2461 zero_check_long(peek(1));
2462 // Compile-time detect of null-exception?
2463 if (stopped()) return;
2464 b = pop_pair();
2465 a = pop_pair();
2466 c = _gvn.transform( new DivLNode(control(),a,b) );
2467 push_pair(c);
2468 break;
2469
2470 case Bytecodes::_ladd:
2471 b = pop_pair();
2472 a = pop_pair();
2473 c = _gvn.transform( new AddLNode(a,b) );
2474 push_pair(c);
2475 break;
2476 case Bytecodes::_lsub:
2477 b = pop_pair();
2478 a = pop_pair();
2479 c = _gvn.transform( new SubLNode(a,b) );
2480 push_pair(c);
2481 break;
2482 case Bytecodes::_lcmp:
2483 // Safepoints are now inserted _before_ branches. The long-compare
2484 // bytecode painfully produces a 3-way value (-1,0,+1) which requires a
2485 // slew of control flow. These are usually followed by a CmpI vs zero and
2486 // a branch; this pattern then optimizes to the obvious long-compare and
2487 // branch. However, if the branch is backwards there's a Safepoint
2488 // inserted. The inserted Safepoint captures the JVM state at the
2489 // pre-branch point, i.e. it captures the 3-way value. Thus if a
2490 // long-compare is used to control a loop the debug info will force
2491 // computation of the 3-way value, even though the generated code uses a
2492 // long-compare and branch. We try to rectify the situation by inserting
2493 // a SafePoint here and have it dominate and kill the safepoint added at a
2494 // following backwards branch. At this point the JVM state merely holds 2
2495 // longs but not the 3-way value.
2496 switch (iter().next_bc()) {
2497 case Bytecodes::_ifgt:
2498 case Bytecodes::_iflt:
2499 case Bytecodes::_ifge:
2500 case Bytecodes::_ifle:
2501 case Bytecodes::_ifne:
2502 case Bytecodes::_ifeq:
2503 // If this is a backwards branch in the bytecodes, add Safepoint
2504 maybe_add_safepoint(iter().next_get_dest());
2505 default:
2506 break;
2507 }
2508 b = pop_pair();
2509 a = pop_pair();
2510 c = _gvn.transform( new CmpL3Node( a, b ));
2511 push(c);
2512 break;
2513
2514 case Bytecodes::_lneg:
2515 a = pop_pair();
2516 b = _gvn.transform( new SubLNode(longcon(0),a));
2517 push_pair(b);
2518 break;
2519 case Bytecodes::_l2i:
2520 a = pop_pair();
2521 push( _gvn.transform( new ConvL2INode(a)));
2522 break;
2523 case Bytecodes::_i2l:
2524 a = pop();
2525 b = _gvn.transform( new ConvI2LNode(a));
2526 push_pair(b);
2527 break;
2528 case Bytecodes::_i2b:
2529 // Sign extend
2530 a = pop();
2531 a = Compile::narrow_value(T_BYTE, a, NULL, &_gvn, true);
2532 push(a);
2533 break;
2534 case Bytecodes::_i2s:
2535 a = pop();
2536 a = Compile::narrow_value(T_SHORT, a, NULL, &_gvn, true);
2537 push(a);
2538 break;
2539 case Bytecodes::_i2c:
2540 a = pop();
2541 a = Compile::narrow_value(T_CHAR, a, NULL, &_gvn, true);
2542 push(a);
2543 break;
2544
2545 case Bytecodes::_i2f:
2546 a = pop();
2547 b = _gvn.transform( new ConvI2FNode(a) ) ;
2548 c = precision_rounding(b);
2549 push (b);
2550 break;
2551
2552 case Bytecodes::_i2d:
2553 a = pop();
2554 b = _gvn.transform( new ConvI2DNode(a));
2555 push_pair(b);
2556 break;
2557
2558 case Bytecodes::_iinc: // Increment local
2559 i = iter().get_index(); // Get local index
2560 set_local( i, _gvn.transform( new AddINode( _gvn.intcon(iter().get_iinc_con()), local(i) ) ) );
2561 break;
2562
2563 // Exit points of synchronized methods must have an unlock node
2564 case Bytecodes::_return:
2565 return_current(NULL);
2566 break;
2567
2568 case Bytecodes::_ireturn:
2569 case Bytecodes::_areturn:
2570 case Bytecodes::_freturn:
2571 return_current(pop());
2572 break;
2573 case Bytecodes::_lreturn:
2574 return_current(pop_pair());
2575 break;
2576 case Bytecodes::_dreturn:
2577 return_current(pop_pair());
2578 break;
2579
2580 case Bytecodes::_athrow:
2581 // null exception oop throws NULL pointer exception
2582 null_check(peek());
2583 if (stopped()) return;
2584 // Hook the thrown exception directly to subsequent handlers.
2585 if (BailoutToInterpreterForThrows) {
2586 // Keep method interpreted from now on.
2587 uncommon_trap(Deoptimization::Reason_unhandled,
2588 Deoptimization::Action_make_not_compilable);
2589 return;
2590 }
2591 if (env()->jvmti_can_post_on_exceptions()) {
2592 // check if we must post exception events, take uncommon trap if so (with must_throw = false)
2593 uncommon_trap_if_should_post_on_exceptions(Deoptimization::Reason_unhandled, false);
2594 }
2595 // Here if either can_post_on_exceptions or should_post_on_exceptions is false
2596 add_exception_state(make_exception_state(peek()));
2597 break;
2598
2599 case Bytecodes::_goto: // fall through
2600 case Bytecodes::_goto_w: {
2601 int target_bci = (bc() == Bytecodes::_goto) ? iter().get_dest() : iter().get_far_dest();
2602
2603 // If this is a backwards branch in the bytecodes, add Safepoint
2604 maybe_add_safepoint(target_bci);
2605
2606 // Merge the current control into the target basic block
2607 merge(target_bci);
2608
2609 // See if we can get some profile data and hand it off to the next block
2610 Block *target_block = block()->successor_for_bci(target_bci);
2611 if (target_block->pred_count() != 1) break;
2612 ciMethodData* methodData = method()->method_data();
2613 if (!methodData->is_mature()) break;
2614 ciProfileData* data = methodData->bci_to_data(bci());
2615 assert(data != NULL && data->is_JumpData(), "need JumpData for taken branch");
2616 int taken = ((ciJumpData*)data)->taken();
2617 taken = method()->scale_count(taken);
2618 target_block->set_count(taken);
2619 break;
2620 }
2621
2622 case Bytecodes::_ifnull: btest = BoolTest::eq; goto handle_if_null;
2623 case Bytecodes::_ifnonnull: btest = BoolTest::ne; goto handle_if_null;
2624 handle_if_null:
2625 // If this is a backwards branch in the bytecodes, add Safepoint
2626 maybe_add_safepoint(iter().get_dest());
2627 a = null();
2628 b = pop();
2629 if (!_gvn.type(b)->speculative_maybe_null() &&
2630 !too_many_traps(Deoptimization::Reason_speculate_null_check)) {
2631 inc_sp(1);
2632 Node* null_ctl = top();
2633 b = null_check_oop(b, &null_ctl, true, true, true);
2634 assert(null_ctl->is_top(), "no null control here");
2635 dec_sp(1);
2636 } else if (_gvn.type(b)->speculative_always_null() &&
2637 !too_many_traps(Deoptimization::Reason_speculate_null_assert)) {
2638 inc_sp(1);
2639 b = null_assert(b);
2640 dec_sp(1);
2641 }
2642 c = _gvn.transform( new CmpPNode(b, a) );
2643 do_ifnull(btest, c);
2644 break;
2645
2646 case Bytecodes::_if_acmpeq: btest = BoolTest::eq; goto handle_if_acmp;
2647 case Bytecodes::_if_acmpne: btest = BoolTest::ne; goto handle_if_acmp;
2648 handle_if_acmp:
2649 // If this is a backwards branch in the bytecodes, add Safepoint
2650 maybe_add_safepoint(iter().get_dest());
2651 a = pop();
2652 b = pop();
2653 c = _gvn.transform( new CmpPNode(b, a) );
2654 c = optimize_cmp_with_klass(c);
2655 do_if(btest, c);
2656 break;
2657
2658 case Bytecodes::_ifeq: btest = BoolTest::eq; goto handle_ifxx;
2659 case Bytecodes::_ifne: btest = BoolTest::ne; goto handle_ifxx;
2660 case Bytecodes::_iflt: btest = BoolTest::lt; goto handle_ifxx;
2661 case Bytecodes::_ifle: btest = BoolTest::le; goto handle_ifxx;
2662 case Bytecodes::_ifgt: btest = BoolTest::gt; goto handle_ifxx;
2663 case Bytecodes::_ifge: btest = BoolTest::ge; goto handle_ifxx;
2664 handle_ifxx:
2665 // If this is a backwards branch in the bytecodes, add Safepoint
2666 maybe_add_safepoint(iter().get_dest());
2667 a = _gvn.intcon(0);
2668 b = pop();
2669 c = _gvn.transform( new CmpINode(b, a) );
2670 do_if(btest, c);
2671 break;
2672
2673 case Bytecodes::_if_icmpeq: btest = BoolTest::eq; goto handle_if_icmp;
2674 case Bytecodes::_if_icmpne: btest = BoolTest::ne; goto handle_if_icmp;
2675 case Bytecodes::_if_icmplt: btest = BoolTest::lt; goto handle_if_icmp;
2676 case Bytecodes::_if_icmple: btest = BoolTest::le; goto handle_if_icmp;
2677 case Bytecodes::_if_icmpgt: btest = BoolTest::gt; goto handle_if_icmp;
2678 case Bytecodes::_if_icmpge: btest = BoolTest::ge; goto handle_if_icmp;
2679 handle_if_icmp:
2680 // If this is a backwards branch in the bytecodes, add Safepoint
2681 maybe_add_safepoint(iter().get_dest());
2682 a = pop();
2683 b = pop();
2684 c = _gvn.transform( new CmpINode( b, a ) );
2685 do_if(btest, c);
2686 break;
2687
2688 case Bytecodes::_tableswitch:
2689 do_tableswitch();
2690 break;
2691
2692 case Bytecodes::_lookupswitch:
2693 do_lookupswitch();
2694 break;
2695
2696 case Bytecodes::_invokestatic:
2697 case Bytecodes::_invokedynamic:
2698 case Bytecodes::_invokespecial:
2699 case Bytecodes::_invokevirtual:
2700 case Bytecodes::_invokeinterface:
2701 do_call();
2702 break;
2703 case Bytecodes::_checkcast:
2704 do_checkcast();
2705 break;
2706 case Bytecodes::_instanceof:
2707 do_instanceof();
2708 break;
2709 case Bytecodes::_anewarray:
2710 do_anewarray();
2711 break;
2712 case Bytecodes::_newarray:
2713 do_newarray((BasicType)iter().get_index());
2714 break;
2715 case Bytecodes::_multianewarray:
2716 do_multianewarray();
2717 break;
2718 case Bytecodes::_new:
2719 do_new();
2720 break;
2721
2722 case Bytecodes::_jsr:
2723 case Bytecodes::_jsr_w:
2724 do_jsr();
2725 break;
2726
2727 case Bytecodes::_ret:
2728 do_ret();
2729 break;
2730
2731
2732 case Bytecodes::_monitorenter:
2733 do_monitor_enter();
2734 break;
2735
2736 case Bytecodes::_monitorexit:
2737 do_monitor_exit();
2738 break;
2739
2740 case Bytecodes::_breakpoint:
2741 // Breakpoint set concurrently to compile
2742 // %%% use an uncommon trap?
2743 C->record_failure("breakpoint in method");
2744 return;
2745
2746 default:
2747 #ifndef PRODUCT
2748 map()->dump(99);
2749 #endif
2750 tty->print("\nUnhandled bytecode %s\n", Bytecodes::name(bc()) );
2751 ShouldNotReachHere();
2752 }
2753
2754 #ifndef PRODUCT
2755 if (C->should_print_igv(1)) {
2756 IdealGraphPrinter* printer = C->igv_printer();
2757 char buffer[256];
2758 jio_snprintf(buffer, sizeof(buffer), "Bytecode %d: %s", bci(), Bytecodes::name(bc()));
2759 bool old = printer->traverse_outs();
2760 printer->set_traverse_outs(true);
2761 printer->print_method(buffer, 4);
2762 printer->set_traverse_outs(old);
2763 }
2764 #endif
2765 }