1 /*
2 * Copyright (c) 1998, 2021, 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/systemDictionary.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----------------------------------
array_load(BasicType bt)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----------------------------------
array_store(BasicType bt)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.
array_addressing(BasicType type,int vals,const Type * & elemtype)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
jump_if_fork_int(Node * a,Node * b,BoolTest::mask mask,float prob,float cnt)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 // return Region node
jump_if_join(Node * iffalse,Node * iftrue)222 Node* Parse::jump_if_join(Node* iffalse, Node* iftrue) {
223 Node *region = new RegionNode(3); // 2 results
224 record_for_igvn(region);
225 region->init_req(1, iffalse);
226 region->init_req(2, iftrue );
227 _gvn.set_type(region, Type::CONTROL);
228 region = _gvn.transform(region);
229 set_control (region);
230 return region;
231 }
232
233 // sentinel value for the target bci to mark never taken branches
234 // (according to profiling)
235 static const int never_reached = INT_MAX;
236
237 //------------------------------helper for tableswitch-------------------------
jump_if_true_fork(IfNode * iff,int dest_bci_if_true,bool unc)238 void Parse::jump_if_true_fork(IfNode *iff, int dest_bci_if_true, bool unc) {
239 // True branch, use existing map info
240 { PreserveJVMState pjvms(this);
241 Node *iftrue = _gvn.transform( new IfTrueNode (iff) );
242 set_control( iftrue );
243 if (unc) {
244 repush_if_args();
245 uncommon_trap(Deoptimization::Reason_unstable_if,
246 Deoptimization::Action_reinterpret,
247 NULL,
248 "taken always");
249 } else {
250 assert(dest_bci_if_true != never_reached, "inconsistent dest");
251 merge_new_path(dest_bci_if_true);
252 }
253 }
254
255 // False branch
256 Node *iffalse = _gvn.transform( new IfFalseNode(iff) );
257 set_control( iffalse );
258 }
259
jump_if_false_fork(IfNode * iff,int dest_bci_if_true,bool unc)260 void Parse::jump_if_false_fork(IfNode *iff, int dest_bci_if_true, bool unc) {
261 // True branch, use existing map info
262 { PreserveJVMState pjvms(this);
263 Node *iffalse = _gvn.transform( new IfFalseNode (iff) );
264 set_control( iffalse );
265 if (unc) {
266 repush_if_args();
267 uncommon_trap(Deoptimization::Reason_unstable_if,
268 Deoptimization::Action_reinterpret,
269 NULL,
270 "taken never");
271 } else {
272 assert(dest_bci_if_true != never_reached, "inconsistent dest");
273 merge_new_path(dest_bci_if_true);
274 }
275 }
276
277 // False branch
278 Node *iftrue = _gvn.transform( new IfTrueNode(iff) );
279 set_control( iftrue );
280 }
281
jump_if_always_fork(int dest_bci,bool unc)282 void Parse::jump_if_always_fork(int dest_bci, bool unc) {
283 // False branch, use existing map and control()
284 if (unc) {
285 repush_if_args();
286 uncommon_trap(Deoptimization::Reason_unstable_if,
287 Deoptimization::Action_reinterpret,
288 NULL,
289 "taken never");
290 } else {
291 assert(dest_bci != never_reached, "inconsistent dest");
292 merge_new_path(dest_bci);
293 }
294 }
295
296
297 extern "C" {
jint_cmp(const void * i,const void * j)298 static int jint_cmp(const void *i, const void *j) {
299 int a = *(jint *)i;
300 int b = *(jint *)j;
301 return a > b ? 1 : a < b ? -1 : 0;
302 }
303 }
304
305
306 class SwitchRange : public StackObj {
307 // a range of integers coupled with a bci destination
308 jint _lo; // inclusive lower limit
309 jint _hi; // inclusive upper limit
310 int _dest;
311 float _cnt; // how many times this range was hit according to profiling
312
313 public:
lo() const314 jint lo() const { return _lo; }
hi() const315 jint hi() const { return _hi; }
dest() const316 int dest() const { return _dest; }
is_singleton() const317 bool is_singleton() const { return _lo == _hi; }
cnt() const318 float cnt() const { return _cnt; }
319
setRange(jint lo,jint hi,int dest,float cnt)320 void setRange(jint lo, jint hi, int dest, float cnt) {
321 assert(lo <= hi, "must be a non-empty range");
322 _lo = lo, _hi = hi; _dest = dest; _cnt = cnt;
323 assert(_cnt >= 0, "");
324 }
adjoinRange(jint lo,jint hi,int dest,float cnt,bool trim_ranges)325 bool adjoinRange(jint lo, jint hi, int dest, float cnt, bool trim_ranges) {
326 assert(lo <= hi, "must be a non-empty range");
327 if (lo == _hi+1) {
328 // see merge_ranges() comment below
329 if (trim_ranges) {
330 if (cnt == 0) {
331 if (_cnt != 0) {
332 return false;
333 }
334 if (dest != _dest) {
335 _dest = never_reached;
336 }
337 } else {
338 if (_cnt == 0) {
339 return false;
340 }
341 if (dest != _dest) {
342 return false;
343 }
344 }
345 } else {
346 if (dest != _dest) {
347 return false;
348 }
349 }
350 _hi = hi;
351 _cnt += cnt;
352 return true;
353 }
354 return false;
355 }
356
set(jint value,int dest,float cnt)357 void set (jint value, int dest, float cnt) {
358 setRange(value, value, dest, cnt);
359 }
adjoin(jint value,int dest,float cnt,bool trim_ranges)360 bool adjoin(jint value, int dest, float cnt, bool trim_ranges) {
361 return adjoinRange(value, value, dest, cnt, trim_ranges);
362 }
adjoin(SwitchRange & other)363 bool adjoin(SwitchRange& other) {
364 return adjoinRange(other._lo, other._hi, other._dest, other._cnt, false);
365 }
366
print()367 void print() {
368 if (is_singleton())
369 tty->print(" {%d}=>%d (cnt=%f)", lo(), dest(), cnt());
370 else if (lo() == min_jint)
371 tty->print(" {..%d}=>%d (cnt=%f)", hi(), dest(), cnt());
372 else if (hi() == max_jint)
373 tty->print(" {%d..}=>%d (cnt=%f)", lo(), dest(), cnt());
374 else
375 tty->print(" {%d..%d}=>%d (cnt=%f)", lo(), hi(), dest(), cnt());
376 }
377 };
378
379 // We try to minimize the number of ranges and the size of the taken
380 // ones using profiling data. When ranges are created,
381 // SwitchRange::adjoinRange() only allows 2 adjoining ranges to merge
382 // if both were never hit or both were hit to build longer unreached
383 // ranges. Here, we now merge adjoining ranges with the same
384 // destination and finally set destination of unreached ranges to the
385 // special value never_reached because it can help minimize the number
386 // of tests that are necessary.
387 //
388 // For instance:
389 // [0, 1] to target1 sometimes taken
390 // [1, 2] to target1 never taken
391 // [2, 3] to target2 never taken
392 // would lead to:
393 // [0, 1] to target1 sometimes taken
394 // [1, 3] never taken
395 //
396 // (first 2 ranges to target1 are not merged)
merge_ranges(SwitchRange * ranges,int & rp)397 static void merge_ranges(SwitchRange* ranges, int& rp) {
398 if (rp == 0) {
399 return;
400 }
401 int shift = 0;
402 for (int j = 0; j < rp; j++) {
403 SwitchRange& r1 = ranges[j-shift];
404 SwitchRange& r2 = ranges[j+1];
405 if (r1.adjoin(r2)) {
406 shift++;
407 } else if (shift > 0) {
408 ranges[j+1-shift] = r2;
409 }
410 }
411 rp -= shift;
412 for (int j = 0; j <= rp; j++) {
413 SwitchRange& r = ranges[j];
414 if (r.cnt() == 0 && r.dest() != never_reached) {
415 r.setRange(r.lo(), r.hi(), never_reached, r.cnt());
416 }
417 }
418 }
419
420 //-------------------------------do_tableswitch--------------------------------
do_tableswitch()421 void Parse::do_tableswitch() {
422 // Get information about tableswitch
423 int default_dest = iter().get_dest_table(0);
424 int lo_index = iter().get_int_table(1);
425 int hi_index = iter().get_int_table(2);
426 int len = hi_index - lo_index + 1;
427
428 if (len < 1) {
429 // If this is a backward branch, add safepoint
430 maybe_add_safepoint(default_dest);
431 pop(); // the effect of the instruction execution on the operand stack
432 merge(default_dest);
433 return;
434 }
435
436 ciMethodData* methodData = method()->method_data();
437 ciMultiBranchData* profile = NULL;
438 if (methodData->is_mature() && UseSwitchProfiling) {
439 ciProfileData* data = methodData->bci_to_data(bci());
440 if (data != NULL && data->is_MultiBranchData()) {
441 profile = (ciMultiBranchData*)data;
442 }
443 }
444 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if);
445
446 // generate decision tree, using trichotomy when possible
447 int rnum = len+2;
448 bool makes_backward_branch = false;
449 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum);
450 int rp = -1;
451 if (lo_index != min_jint) {
452 uint cnt = 1;
453 if (profile != NULL) {
454 cnt = profile->default_count() / (hi_index != max_jint ? 2 : 1);
455 }
456 ranges[++rp].setRange(min_jint, lo_index-1, default_dest, cnt);
457 }
458 for (int j = 0; j < len; j++) {
459 jint match_int = lo_index+j;
460 int dest = iter().get_dest_table(j+3);
461 makes_backward_branch |= (dest <= bci());
462 uint cnt = 1;
463 if (profile != NULL) {
464 cnt = profile->count_at(j);
465 }
466 if (rp < 0 || !ranges[rp].adjoin(match_int, dest, cnt, trim_ranges)) {
467 ranges[++rp].set(match_int, dest, cnt);
468 }
469 }
470 jint highest = lo_index+(len-1);
471 assert(ranges[rp].hi() == highest, "");
472 if (highest != max_jint) {
473 uint cnt = 1;
474 if (profile != NULL) {
475 cnt = profile->default_count() / (lo_index != min_jint ? 2 : 1);
476 }
477 if (!ranges[rp].adjoinRange(highest+1, max_jint, default_dest, cnt, trim_ranges)) {
478 ranges[++rp].setRange(highest+1, max_jint, default_dest, cnt);
479 }
480 }
481 assert(rp < len+2, "not too many ranges");
482
483 if (trim_ranges) {
484 merge_ranges(ranges, rp);
485 }
486
487 // Safepoint in case if backward branch observed
488 if (makes_backward_branch && UseLoopSafepoints) {
489 add_safepoint();
490 }
491
492 Node* lookup = pop(); // lookup value
493 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]);
494 }
495
496
497 //------------------------------do_lookupswitch--------------------------------
do_lookupswitch()498 void Parse::do_lookupswitch() {
499 // Get information about lookupswitch
500 int default_dest = iter().get_dest_table(0);
501 int len = iter().get_int_table(1);
502
503 if (len < 1) { // If this is a backward branch, add safepoint
504 maybe_add_safepoint(default_dest);
505 pop(); // the effect of the instruction execution on the operand stack
506 merge(default_dest);
507 return;
508 }
509
510 ciMethodData* methodData = method()->method_data();
511 ciMultiBranchData* profile = NULL;
512 if (methodData->is_mature() && UseSwitchProfiling) {
513 ciProfileData* data = methodData->bci_to_data(bci());
514 if (data != NULL && data->is_MultiBranchData()) {
515 profile = (ciMultiBranchData*)data;
516 }
517 }
518 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if);
519
520 // generate decision tree, using trichotomy when possible
521 jint* table = NEW_RESOURCE_ARRAY(jint, len*3);
522 {
523 for (int j = 0; j < len; j++) {
524 table[3*j+0] = iter().get_int_table(2+2*j);
525 table[3*j+1] = iter().get_dest_table(2+2*j+1);
526 // Handle overflow when converting from uint to jint
527 table[3*j+2] = (profile == NULL) ? 1 : MIN2<uint>(max_jint, profile->count_at(j));
528 }
529 qsort(table, len, 3*sizeof(table[0]), jint_cmp);
530 }
531
532 float defaults = 0;
533 jint prev = min_jint;
534 for (int j = 0; j < len; j++) {
535 jint match_int = table[3*j+0];
536 if (match_int != prev) {
537 defaults += (float)match_int - prev;
538 }
539 prev = match_int+1;
540 }
541 if (prev != min_jint) {
542 defaults += (float)max_jint - prev + 1;
543 }
544 float default_cnt = 1;
545 if (profile != NULL) {
546 default_cnt = profile->default_count()/defaults;
547 }
548
549 int rnum = len*2+1;
550 bool makes_backward_branch = false;
551 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum);
552 int rp = -1;
553 for (int j = 0; j < len; j++) {
554 jint match_int = table[3*j+0];
555 int dest = table[3*j+1];
556 int cnt = table[3*j+2];
557 int next_lo = rp < 0 ? min_jint : ranges[rp].hi()+1;
558 makes_backward_branch |= (dest <= bci());
559 float c = default_cnt * ((float)match_int - next_lo);
560 if (match_int != next_lo && (rp < 0 || !ranges[rp].adjoinRange(next_lo, match_int-1, default_dest, c, trim_ranges))) {
561 assert(default_dest != never_reached, "sentinel value for dead destinations");
562 ranges[++rp].setRange(next_lo, match_int-1, default_dest, c);
563 }
564 if (rp < 0 || !ranges[rp].adjoin(match_int, dest, cnt, trim_ranges)) {
565 assert(dest != never_reached, "sentinel value for dead destinations");
566 ranges[++rp].set(match_int, dest, cnt);
567 }
568 }
569 jint highest = table[3*(len-1)];
570 assert(ranges[rp].hi() == highest, "");
571 if (highest != max_jint &&
572 !ranges[rp].adjoinRange(highest+1, max_jint, default_dest, default_cnt * ((float)max_jint - highest), trim_ranges)) {
573 ranges[++rp].setRange(highest+1, max_jint, default_dest, default_cnt * ((float)max_jint - highest));
574 }
575 assert(rp < rnum, "not too many ranges");
576
577 if (trim_ranges) {
578 merge_ranges(ranges, rp);
579 }
580
581 // Safepoint in case backward branch observed
582 if (makes_backward_branch && UseLoopSafepoints) {
583 add_safepoint();
584 }
585
586 Node *lookup = pop(); // lookup value
587 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]);
588 }
589
if_prob(float taken_cnt,float total_cnt)590 static float if_prob(float taken_cnt, float total_cnt) {
591 assert(taken_cnt <= total_cnt, "");
592 if (total_cnt == 0) {
593 return PROB_FAIR;
594 }
595 float p = taken_cnt / total_cnt;
596 return clamp(p, PROB_MIN, PROB_MAX);
597 }
598
if_cnt(float cnt)599 static float if_cnt(float cnt) {
600 if (cnt == 0) {
601 return COUNT_UNKNOWN;
602 }
603 return cnt;
604 }
605
sum_of_cnts(SwitchRange * lo,SwitchRange * hi)606 static float sum_of_cnts(SwitchRange *lo, SwitchRange *hi) {
607 float total_cnt = 0;
608 for (SwitchRange* sr = lo; sr <= hi; sr++) {
609 total_cnt += sr->cnt();
610 }
611 return total_cnt;
612 }
613
614 class SwitchRanges : public ResourceObj {
615 public:
616 SwitchRange* _lo;
617 SwitchRange* _hi;
618 SwitchRange* _mid;
619 float _cost;
620
621 enum {
622 Start,
623 LeftDone,
624 RightDone,
625 Done
626 } _state;
627
SwitchRanges(SwitchRange * lo,SwitchRange * hi)628 SwitchRanges(SwitchRange *lo, SwitchRange *hi)
629 : _lo(lo), _hi(hi), _mid(NULL),
630 _cost(0), _state(Start) {
631 }
632
SwitchRanges()633 SwitchRanges()
634 : _lo(NULL), _hi(NULL), _mid(NULL),
635 _cost(0), _state(Start) {}
636 };
637
638 // Estimate cost of performing a binary search on lo..hi
compute_tree_cost(SwitchRange * lo,SwitchRange * hi,float total_cnt)639 static float compute_tree_cost(SwitchRange *lo, SwitchRange *hi, float total_cnt) {
640 GrowableArray<SwitchRanges> tree;
641 SwitchRanges root(lo, hi);
642 tree.push(root);
643
644 float cost = 0;
645 do {
646 SwitchRanges& r = *tree.adr_at(tree.length()-1);
647 if (r._hi != r._lo) {
648 if (r._mid == NULL) {
649 float r_cnt = sum_of_cnts(r._lo, r._hi);
650
651 if (r_cnt == 0) {
652 tree.pop();
653 cost = 0;
654 continue;
655 }
656
657 SwitchRange* mid = NULL;
658 mid = r._lo;
659 for (float cnt = 0; ; ) {
660 assert(mid <= r._hi, "out of bounds");
661 cnt += mid->cnt();
662 if (cnt > r_cnt / 2) {
663 break;
664 }
665 mid++;
666 }
667 assert(mid <= r._hi, "out of bounds");
668 r._mid = mid;
669 r._cost = r_cnt / total_cnt;
670 }
671 r._cost += cost;
672 if (r._state < SwitchRanges::LeftDone && r._mid > r._lo) {
673 cost = 0;
674 r._state = SwitchRanges::LeftDone;
675 tree.push(SwitchRanges(r._lo, r._mid-1));
676 } else if (r._state < SwitchRanges::RightDone) {
677 cost = 0;
678 r._state = SwitchRanges::RightDone;
679 tree.push(SwitchRanges(r._mid == r._lo ? r._mid+1 : r._mid, r._hi));
680 } else {
681 tree.pop();
682 cost = r._cost;
683 }
684 } else {
685 tree.pop();
686 cost = r._cost;
687 }
688 } while (tree.length() > 0);
689
690
691 return cost;
692 }
693
694 // It sometimes pays off to test most common ranges before the binary search
linear_search_switch_ranges(Node * key_val,SwitchRange * & lo,SwitchRange * & hi)695 void Parse::linear_search_switch_ranges(Node* key_val, SwitchRange*& lo, SwitchRange*& hi) {
696 uint nr = hi - lo + 1;
697 float total_cnt = sum_of_cnts(lo, hi);
698
699 float min = compute_tree_cost(lo, hi, total_cnt);
700 float extra = 1;
701 float sub = 0;
702
703 SwitchRange* array1 = lo;
704 SwitchRange* array2 = NEW_RESOURCE_ARRAY(SwitchRange, nr);
705
706 SwitchRange* ranges = NULL;
707
708 while (nr >= 2) {
709 assert(lo == array1 || lo == array2, "one the 2 already allocated arrays");
710 ranges = (lo == array1) ? array2 : array1;
711
712 // Find highest frequency range
713 SwitchRange* candidate = lo;
714 for (SwitchRange* sr = lo+1; sr <= hi; sr++) {
715 if (sr->cnt() > candidate->cnt()) {
716 candidate = sr;
717 }
718 }
719 SwitchRange most_freq = *candidate;
720 if (most_freq.cnt() == 0) {
721 break;
722 }
723
724 // Copy remaining ranges into another array
725 int shift = 0;
726 for (uint i = 0; i < nr; i++) {
727 SwitchRange* sr = &lo[i];
728 if (sr != candidate) {
729 ranges[i-shift] = *sr;
730 } else {
731 shift++;
732 if (i > 0 && i < nr-1) {
733 SwitchRange prev = lo[i-1];
734 prev.setRange(prev.lo(), sr->hi(), prev.dest(), prev.cnt());
735 if (prev.adjoin(lo[i+1])) {
736 shift++;
737 i++;
738 }
739 ranges[i-shift] = prev;
740 }
741 }
742 }
743 nr -= shift;
744
745 // Evaluate cost of testing the most common range and performing a
746 // binary search on the other ranges
747 float cost = extra + compute_tree_cost(&ranges[0], &ranges[nr-1], total_cnt);
748 if (cost >= min) {
749 break;
750 }
751 // swap arrays
752 lo = &ranges[0];
753 hi = &ranges[nr-1];
754
755 // It pays off: emit the test for the most common range
756 assert(most_freq.cnt() > 0, "must be taken");
757 Node* val = _gvn.transform(new SubINode(key_val, _gvn.intcon(most_freq.lo())));
758 Node* cmp = _gvn.transform(new CmpUNode(val, _gvn.intcon(most_freq.hi() - most_freq.lo())));
759 Node* tst = _gvn.transform(new BoolNode(cmp, BoolTest::le));
760 IfNode* iff = create_and_map_if(control(), tst, if_prob(most_freq.cnt(), total_cnt), if_cnt(most_freq.cnt()));
761 jump_if_true_fork(iff, most_freq.dest(), false);
762
763 sub += most_freq.cnt() / total_cnt;
764 extra += 1 - sub;
765 min = cost;
766 }
767 }
768
769 //----------------------------create_jump_tables-------------------------------
create_jump_tables(Node * key_val,SwitchRange * lo,SwitchRange * hi)770 bool Parse::create_jump_tables(Node* key_val, SwitchRange* lo, SwitchRange* hi) {
771 // Are jumptables enabled
772 if (!UseJumpTables) return false;
773
774 // Are jumptables supported
775 if (!Matcher::has_match_rule(Op_Jump)) return false;
776
777 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if);
778
779 // Decide if a guard is needed to lop off big ranges at either (or
780 // both) end(s) of the input set. We'll call this the default target
781 // even though we can't be sure that it is the true "default".
782
783 bool needs_guard = false;
784 int default_dest;
785 int64_t total_outlier_size = 0;
786 int64_t hi_size = ((int64_t)hi->hi()) - ((int64_t)hi->lo()) + 1;
787 int64_t lo_size = ((int64_t)lo->hi()) - ((int64_t)lo->lo()) + 1;
788
789 if (lo->dest() == hi->dest()) {
790 total_outlier_size = hi_size + lo_size;
791 default_dest = lo->dest();
792 } else if (lo_size > hi_size) {
793 total_outlier_size = lo_size;
794 default_dest = lo->dest();
795 } else {
796 total_outlier_size = hi_size;
797 default_dest = hi->dest();
798 }
799
800 float total = sum_of_cnts(lo, hi);
801 float cost = compute_tree_cost(lo, hi, total);
802
803 // If a guard test will eliminate very sparse end ranges, then
804 // it is worth the cost of an extra jump.
805 float trimmed_cnt = 0;
806 if (total_outlier_size > (MaxJumpTableSparseness * 4)) {
807 needs_guard = true;
808 if (default_dest == lo->dest()) {
809 trimmed_cnt += lo->cnt();
810 lo++;
811 }
812 if (default_dest == hi->dest()) {
813 trimmed_cnt += hi->cnt();
814 hi--;
815 }
816 }
817
818 // Find the total number of cases and ranges
819 int64_t num_cases = ((int64_t)hi->hi()) - ((int64_t)lo->lo()) + 1;
820 int num_range = hi - lo + 1;
821
822 // Don't create table if: too large, too small, or too sparse.
823 if (num_cases > MaxJumpTableSize)
824 return false;
825 if (UseSwitchProfiling) {
826 // MinJumpTableSize is set so with a well balanced binary tree,
827 // when the number of ranges is MinJumpTableSize, it's cheaper to
828 // go through a JumpNode that a tree of IfNodes. Average cost of a
829 // tree of IfNodes with MinJumpTableSize is
830 // log2f(MinJumpTableSize) comparisons. So if the cost computed
831 // from profile data is less than log2f(MinJumpTableSize) then
832 // going with the binary search is cheaper.
833 if (cost < log2f(MinJumpTableSize)) {
834 return false;
835 }
836 } else {
837 if (num_cases < MinJumpTableSize)
838 return false;
839 }
840 if (num_cases > (MaxJumpTableSparseness * num_range))
841 return false;
842
843 // Normalize table lookups to zero
844 int lowval = lo->lo();
845 key_val = _gvn.transform( new SubINode(key_val, _gvn.intcon(lowval)) );
846
847 // Generate a guard to protect against input keyvals that aren't
848 // in the switch domain.
849 if (needs_guard) {
850 Node* size = _gvn.intcon(num_cases);
851 Node* cmp = _gvn.transform(new CmpUNode(key_val, size));
852 Node* tst = _gvn.transform(new BoolNode(cmp, BoolTest::ge));
853 IfNode* iff = create_and_map_if(control(), tst, if_prob(trimmed_cnt, total), if_cnt(trimmed_cnt));
854 jump_if_true_fork(iff, default_dest, trim_ranges && trimmed_cnt == 0);
855
856 total -= trimmed_cnt;
857 }
858
859 // Create an ideal node JumpTable that has projections
860 // of all possible ranges for a switch statement
861 // The key_val input must be converted to a pointer offset and scaled.
862 // Compare Parse::array_addressing above.
863
864 // Clean the 32-bit int into a real 64-bit offset.
865 // Otherwise, the jint value 0 might turn into an offset of 0x0800000000.
866 const TypeInt* ikeytype = TypeInt::make(0, num_cases, Type::WidenMin);
867 // Make I2L conversion control dependent to prevent it from
868 // floating above the range check during loop optimizations.
869 key_val = C->conv_I2X_index(&_gvn, key_val, ikeytype, control());
870
871 // Shift the value by wordsize so we have an index into the table, rather
872 // than a switch value
873 Node *shiftWord = _gvn.MakeConX(wordSize);
874 key_val = _gvn.transform( new MulXNode( key_val, shiftWord));
875
876 // Create the JumpNode
877 Arena* arena = C->comp_arena();
878 float* probs = (float*)arena->Amalloc(sizeof(float)*num_cases);
879 int i = 0;
880 if (total == 0) {
881 for (SwitchRange* r = lo; r <= hi; r++) {
882 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) {
883 probs[i] = 1.0F / num_cases;
884 }
885 }
886 } else {
887 for (SwitchRange* r = lo; r <= hi; r++) {
888 float prob = r->cnt()/total;
889 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) {
890 probs[i] = prob / (r->hi() - r->lo() + 1);
891 }
892 }
893 }
894
895 ciMethodData* methodData = method()->method_data();
896 ciMultiBranchData* profile = NULL;
897 if (methodData->is_mature()) {
898 ciProfileData* data = methodData->bci_to_data(bci());
899 if (data != NULL && data->is_MultiBranchData()) {
900 profile = (ciMultiBranchData*)data;
901 }
902 }
903
904 Node* jtn = _gvn.transform(new JumpNode(control(), key_val, num_cases, probs, profile == NULL ? COUNT_UNKNOWN : total));
905
906 // These are the switch destinations hanging off the jumpnode
907 i = 0;
908 for (SwitchRange* r = lo; r <= hi; r++) {
909 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) {
910 Node* input = _gvn.transform(new JumpProjNode(jtn, i, r->dest(), (int)(j - lowval)));
911 {
912 PreserveJVMState pjvms(this);
913 set_control(input);
914 jump_if_always_fork(r->dest(), trim_ranges && r->cnt() == 0);
915 }
916 }
917 }
918 assert(i == num_cases, "miscount of cases");
919 stop_and_kill_map(); // no more uses for this JVMS
920 return true;
921 }
922
923 //----------------------------jump_switch_ranges-------------------------------
jump_switch_ranges(Node * key_val,SwitchRange * lo,SwitchRange * hi,int switch_depth)924 void Parse::jump_switch_ranges(Node* key_val, SwitchRange *lo, SwitchRange *hi, int switch_depth) {
925 Block* switch_block = block();
926 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if);
927
928 if (switch_depth == 0) {
929 // Do special processing for the top-level call.
930 assert(lo->lo() == min_jint, "initial range must exhaust Type::INT");
931 assert(hi->hi() == max_jint, "initial range must exhaust Type::INT");
932
933 // Decrement pred-numbers for the unique set of nodes.
934 #ifdef ASSERT
935 if (!trim_ranges) {
936 // Ensure that the block's successors are a (duplicate-free) set.
937 int successors_counted = 0; // block occurrences in [hi..lo]
938 int unique_successors = switch_block->num_successors();
939 for (int i = 0; i < unique_successors; i++) {
940 Block* target = switch_block->successor_at(i);
941
942 // Check that the set of successors is the same in both places.
943 int successors_found = 0;
944 for (SwitchRange* p = lo; p <= hi; p++) {
945 if (p->dest() == target->start()) successors_found++;
946 }
947 assert(successors_found > 0, "successor must be known");
948 successors_counted += successors_found;
949 }
950 assert(successors_counted == (hi-lo)+1, "no unexpected successors");
951 }
952 #endif
953
954 // Maybe prune the inputs, based on the type of key_val.
955 jint min_val = min_jint;
956 jint max_val = max_jint;
957 const TypeInt* ti = key_val->bottom_type()->isa_int();
958 if (ti != NULL) {
959 min_val = ti->_lo;
960 max_val = ti->_hi;
961 assert(min_val <= max_val, "invalid int type");
962 }
963 while (lo->hi() < min_val) {
964 lo++;
965 }
966 if (lo->lo() < min_val) {
967 lo->setRange(min_val, lo->hi(), lo->dest(), lo->cnt());
968 }
969 while (hi->lo() > max_val) {
970 hi--;
971 }
972 if (hi->hi() > max_val) {
973 hi->setRange(hi->lo(), max_val, hi->dest(), hi->cnt());
974 }
975
976 linear_search_switch_ranges(key_val, lo, hi);
977 }
978
979 #ifndef PRODUCT
980 if (switch_depth == 0) {
981 _max_switch_depth = 0;
982 _est_switch_depth = log2_intptr((hi-lo+1)-1)+1;
983 }
984 #endif
985
986 assert(lo <= hi, "must be a non-empty set of ranges");
987 if (lo == hi) {
988 jump_if_always_fork(lo->dest(), trim_ranges && lo->cnt() == 0);
989 } else {
990 assert(lo->hi() == (lo+1)->lo()-1, "contiguous ranges");
991 assert(hi->lo() == (hi-1)->hi()+1, "contiguous ranges");
992
993 if (create_jump_tables(key_val, lo, hi)) return;
994
995 SwitchRange* mid = NULL;
996 float total_cnt = sum_of_cnts(lo, hi);
997
998 int nr = hi - lo + 1;
999 if (UseSwitchProfiling) {
1000 // Don't keep the binary search tree balanced: pick up mid point
1001 // that split frequencies in half.
1002 float cnt = 0;
1003 for (SwitchRange* sr = lo; sr <= hi; sr++) {
1004 cnt += sr->cnt();
1005 if (cnt >= total_cnt / 2) {
1006 mid = sr;
1007 break;
1008 }
1009 }
1010 } else {
1011 mid = lo + nr/2;
1012
1013 // if there is an easy choice, pivot at a singleton:
1014 if (nr > 3 && !mid->is_singleton() && (mid-1)->is_singleton()) mid--;
1015
1016 assert(lo < mid && mid <= hi, "good pivot choice");
1017 assert(nr != 2 || mid == hi, "should pick higher of 2");
1018 assert(nr != 3 || mid == hi-1, "should pick middle of 3");
1019 }
1020
1021
1022 Node *test_val = _gvn.intcon(mid == lo ? mid->hi() : mid->lo());
1023
1024 if (mid->is_singleton()) {
1025 IfNode *iff_ne = jump_if_fork_int(key_val, test_val, BoolTest::ne, 1-if_prob(mid->cnt(), total_cnt), if_cnt(mid->cnt()));
1026 jump_if_false_fork(iff_ne, mid->dest(), trim_ranges && mid->cnt() == 0);
1027
1028 // Special Case: If there are exactly three ranges, and the high
1029 // and low range each go to the same place, omit the "gt" test,
1030 // since it will not discriminate anything.
1031 bool eq_test_only = (hi == lo+2 && hi->dest() == lo->dest() && mid == hi-1) || mid == lo;
1032
1033 // if there is a higher range, test for it and process it:
1034 if (mid < hi && !eq_test_only) {
1035 // two comparisons of same values--should enable 1 test for 2 branches
1036 // Use BoolTest::lt instead of BoolTest::gt
1037 float cnt = sum_of_cnts(lo, mid-1);
1038 IfNode *iff_lt = jump_if_fork_int(key_val, test_val, BoolTest::lt, if_prob(cnt, total_cnt), if_cnt(cnt));
1039 Node *iftrue = _gvn.transform( new IfTrueNode(iff_lt) );
1040 Node *iffalse = _gvn.transform( new IfFalseNode(iff_lt) );
1041 { PreserveJVMState pjvms(this);
1042 set_control(iffalse);
1043 jump_switch_ranges(key_val, mid+1, hi, switch_depth+1);
1044 }
1045 set_control(iftrue);
1046 }
1047
1048 } else {
1049 // mid is a range, not a singleton, so treat mid..hi as a unit
1050 float cnt = sum_of_cnts(mid == lo ? mid+1 : mid, hi);
1051 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));
1052
1053 // if there is a higher range, test for it and process it:
1054 if (mid == hi) {
1055 jump_if_true_fork(iff_ge, mid->dest(), trim_ranges && cnt == 0);
1056 } else {
1057 Node *iftrue = _gvn.transform( new IfTrueNode(iff_ge) );
1058 Node *iffalse = _gvn.transform( new IfFalseNode(iff_ge) );
1059 { PreserveJVMState pjvms(this);
1060 set_control(iftrue);
1061 jump_switch_ranges(key_val, mid == lo ? mid+1 : mid, hi, switch_depth+1);
1062 }
1063 set_control(iffalse);
1064 }
1065 }
1066
1067 // in any case, process the lower range
1068 if (mid == lo) {
1069 if (mid->is_singleton()) {
1070 jump_switch_ranges(key_val, lo+1, hi, switch_depth+1);
1071 } else {
1072 jump_if_always_fork(lo->dest(), trim_ranges && lo->cnt() == 0);
1073 }
1074 } else {
1075 jump_switch_ranges(key_val, lo, mid-1, switch_depth+1);
1076 }
1077 }
1078
1079 // Decrease pred_count for each successor after all is done.
1080 if (switch_depth == 0) {
1081 int unique_successors = switch_block->num_successors();
1082 for (int i = 0; i < unique_successors; i++) {
1083 Block* target = switch_block->successor_at(i);
1084 // Throw away the pre-allocated path for each unique successor.
1085 target->next_path_num();
1086 }
1087 }
1088
1089 #ifndef PRODUCT
1090 _max_switch_depth = MAX2(switch_depth, _max_switch_depth);
1091 if (TraceOptoParse && Verbose && WizardMode && switch_depth == 0) {
1092 SwitchRange* r;
1093 int nsing = 0;
1094 for( r = lo; r <= hi; r++ ) {
1095 if( r->is_singleton() ) nsing++;
1096 }
1097 tty->print(">>> ");
1098 _method->print_short_name();
1099 tty->print_cr(" switch decision tree");
1100 tty->print_cr(" %d ranges (%d singletons), max_depth=%d, est_depth=%d",
1101 (int) (hi-lo+1), nsing, _max_switch_depth, _est_switch_depth);
1102 if (_max_switch_depth > _est_switch_depth) {
1103 tty->print_cr("******** BAD SWITCH DEPTH ********");
1104 }
1105 tty->print(" ");
1106 for( r = lo; r <= hi; r++ ) {
1107 r->print();
1108 }
1109 tty->cr();
1110 }
1111 #endif
1112 }
1113
modf()1114 void Parse::modf() {
1115 Node *f2 = pop();
1116 Node *f1 = pop();
1117 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::modf_Type(),
1118 CAST_FROM_FN_PTR(address, SharedRuntime::frem),
1119 "frem", NULL, //no memory effects
1120 f1, f2);
1121 Node* res = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0));
1122
1123 push(res);
1124 }
1125
modd()1126 void Parse::modd() {
1127 Node *d2 = pop_pair();
1128 Node *d1 = pop_pair();
1129 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::Math_DD_D_Type(),
1130 CAST_FROM_FN_PTR(address, SharedRuntime::drem),
1131 "drem", NULL, //no memory effects
1132 d1, top(), d2, top());
1133 Node* res_d = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0));
1134
1135 #ifdef ASSERT
1136 Node* res_top = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 1));
1137 assert(res_top == top(), "second value must be top");
1138 #endif
1139
1140 push_pair(res_d);
1141 }
1142
l2f()1143 void Parse::l2f() {
1144 Node* f2 = pop();
1145 Node* f1 = pop();
1146 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::l2f_Type(),
1147 CAST_FROM_FN_PTR(address, SharedRuntime::l2f),
1148 "l2f", NULL, //no memory effects
1149 f1, f2);
1150 Node* res = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0));
1151
1152 push(res);
1153 }
1154
do_irem()1155 void Parse::do_irem() {
1156 // Must keep both values on the expression-stack during null-check
1157 zero_check_int(peek());
1158 // Compile-time detect of null-exception?
1159 if (stopped()) return;
1160
1161 Node* b = pop();
1162 Node* a = pop();
1163
1164 const Type *t = _gvn.type(b);
1165 if (t != Type::TOP) {
1166 const TypeInt *ti = t->is_int();
1167 if (ti->is_con()) {
1168 int divisor = ti->get_con();
1169 // check for positive power of 2
1170 if (divisor > 0 &&
1171 (divisor & ~(divisor-1)) == divisor) {
1172 // yes !
1173 Node *mask = _gvn.intcon((divisor - 1));
1174 // Sigh, must handle negative dividends
1175 Node *zero = _gvn.intcon(0);
1176 IfNode *ifff = jump_if_fork_int(a, zero, BoolTest::lt, PROB_FAIR, COUNT_UNKNOWN);
1177 Node *iff = _gvn.transform( new IfFalseNode(ifff) );
1178 Node *ift = _gvn.transform( new IfTrueNode (ifff) );
1179 Node *reg = jump_if_join(ift, iff);
1180 Node *phi = PhiNode::make(reg, NULL, TypeInt::INT);
1181 // Negative path; negate/and/negate
1182 Node *neg = _gvn.transform( new SubINode(zero, a) );
1183 Node *andn= _gvn.transform( new AndINode(neg, mask) );
1184 Node *negn= _gvn.transform( new SubINode(zero, andn) );
1185 phi->init_req(1, negn);
1186 // Fast positive case
1187 Node *andx = _gvn.transform( new AndINode(a, mask) );
1188 phi->init_req(2, andx);
1189 // Push the merge
1190 push( _gvn.transform(phi) );
1191 return;
1192 }
1193 }
1194 }
1195 // Default case
1196 push( _gvn.transform( new ModINode(control(),a,b) ) );
1197 }
1198
1199 // Handle jsr and jsr_w bytecode
do_jsr()1200 void Parse::do_jsr() {
1201 assert(bc() == Bytecodes::_jsr || bc() == Bytecodes::_jsr_w, "wrong bytecode");
1202
1203 // Store information about current state, tagged with new _jsr_bci
1204 int return_bci = iter().next_bci();
1205 int jsr_bci = (bc() == Bytecodes::_jsr) ? iter().get_dest() : iter().get_far_dest();
1206
1207 // The way we do things now, there is only one successor block
1208 // for the jsr, because the target code is cloned by ciTypeFlow.
1209 Block* target = successor_for_bci(jsr_bci);
1210
1211 // What got pushed?
1212 const Type* ret_addr = target->peek();
1213 assert(ret_addr->singleton(), "must be a constant (cloned jsr body)");
1214
1215 // Effect on jsr on stack
1216 push(_gvn.makecon(ret_addr));
1217
1218 // Flow to the jsr.
1219 merge(jsr_bci);
1220 }
1221
1222 // Handle ret bytecode
do_ret()1223 void Parse::do_ret() {
1224 // Find to whom we return.
1225 assert(block()->num_successors() == 1, "a ret can only go one place now");
1226 Block* target = block()->successor_at(0);
1227 assert(!target->is_ready(), "our arrival must be expected");
1228 int pnum = target->next_path_num();
1229 merge_common(target, pnum);
1230 }
1231
has_injected_profile(BoolTest::mask btest,Node * test,int & taken,int & not_taken)1232 static bool has_injected_profile(BoolTest::mask btest, Node* test, int& taken, int& not_taken) {
1233 if (btest != BoolTest::eq && btest != BoolTest::ne) {
1234 // Only ::eq and ::ne are supported for profile injection.
1235 return false;
1236 }
1237 if (test->is_Cmp() &&
1238 test->in(1)->Opcode() == Op_ProfileBoolean) {
1239 ProfileBooleanNode* profile = (ProfileBooleanNode*)test->in(1);
1240 int false_cnt = profile->false_count();
1241 int true_cnt = profile->true_count();
1242
1243 // Counts matching depends on the actual test operation (::eq or ::ne).
1244 // No need to scale the counts because profile injection was designed
1245 // to feed exact counts into VM.
1246 taken = (btest == BoolTest::eq) ? false_cnt : true_cnt;
1247 not_taken = (btest == BoolTest::eq) ? true_cnt : false_cnt;
1248
1249 profile->consume();
1250 return true;
1251 }
1252 return false;
1253 }
1254 //--------------------------dynamic_branch_prediction--------------------------
1255 // Try to gather dynamic branch prediction behavior. Return a probability
1256 // of the branch being taken and set the "cnt" field. Returns a -1.0
1257 // if we need to use static prediction for some reason.
dynamic_branch_prediction(float & cnt,BoolTest::mask btest,Node * test)1258 float Parse::dynamic_branch_prediction(float &cnt, BoolTest::mask btest, Node* test) {
1259 ResourceMark rm;
1260
1261 cnt = COUNT_UNKNOWN;
1262
1263 int taken = 0;
1264 int not_taken = 0;
1265
1266 bool use_mdo = !has_injected_profile(btest, test, taken, not_taken);
1267
1268 if (use_mdo) {
1269 // Use MethodData information if it is available
1270 // FIXME: free the ProfileData structure
1271 ciMethodData* methodData = method()->method_data();
1272 if (!methodData->is_mature()) return PROB_UNKNOWN;
1273 ciProfileData* data = methodData->bci_to_data(bci());
1274 if (data == NULL) {
1275 return PROB_UNKNOWN;
1276 }
1277 if (!data->is_JumpData()) return PROB_UNKNOWN;
1278
1279 // get taken and not taken values
1280 taken = data->as_JumpData()->taken();
1281 not_taken = 0;
1282 if (data->is_BranchData()) {
1283 not_taken = data->as_BranchData()->not_taken();
1284 }
1285
1286 // scale the counts to be commensurate with invocation counts:
1287 taken = method()->scale_count(taken);
1288 not_taken = method()->scale_count(not_taken);
1289 }
1290
1291 // Give up if too few (or too many, in which case the sum will overflow) counts to be meaningful.
1292 // We also check that individual counters are positive first, otherwise the sum can become positive.
1293 if (taken < 0 || not_taken < 0 || taken + not_taken < 40) {
1294 if (C->log() != NULL) {
1295 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d'", iter().get_dest(), taken, not_taken);
1296 }
1297 return PROB_UNKNOWN;
1298 }
1299
1300 // Compute frequency that we arrive here
1301 float sum = taken + not_taken;
1302 // Adjust, if this block is a cloned private block but the
1303 // Jump counts are shared. Taken the private counts for
1304 // just this path instead of the shared counts.
1305 if( block()->count() > 0 )
1306 sum = block()->count();
1307 cnt = sum / FreqCountInvocations;
1308
1309 // Pin probability to sane limits
1310 float prob;
1311 if( !taken )
1312 prob = (0+PROB_MIN) / 2;
1313 else if( !not_taken )
1314 prob = (1+PROB_MAX) / 2;
1315 else { // Compute probability of true path
1316 prob = (float)taken / (float)(taken + not_taken);
1317 if (prob > PROB_MAX) prob = PROB_MAX;
1318 if (prob < PROB_MIN) prob = PROB_MIN;
1319 }
1320
1321 assert((cnt > 0.0f) && (prob > 0.0f),
1322 "Bad frequency assignment in if");
1323
1324 if (C->log() != NULL) {
1325 const char* prob_str = NULL;
1326 if (prob >= PROB_MAX) prob_str = (prob == PROB_MAX) ? "max" : "always";
1327 if (prob <= PROB_MIN) prob_str = (prob == PROB_MIN) ? "min" : "never";
1328 char prob_str_buf[30];
1329 if (prob_str == NULL) {
1330 jio_snprintf(prob_str_buf, sizeof(prob_str_buf), "%20.2f", prob);
1331 prob_str = prob_str_buf;
1332 }
1333 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d' cnt='%f' prob='%s'",
1334 iter().get_dest(), taken, not_taken, cnt, prob_str);
1335 }
1336 return prob;
1337 }
1338
1339 //-----------------------------branch_prediction-------------------------------
branch_prediction(float & cnt,BoolTest::mask btest,int target_bci,Node * test)1340 float Parse::branch_prediction(float& cnt,
1341 BoolTest::mask btest,
1342 int target_bci,
1343 Node* test) {
1344 float prob = dynamic_branch_prediction(cnt, btest, test);
1345 // If prob is unknown, switch to static prediction
1346 if (prob != PROB_UNKNOWN) return prob;
1347
1348 prob = PROB_FAIR; // Set default value
1349 if (btest == BoolTest::eq) // Exactly equal test?
1350 prob = PROB_STATIC_INFREQUENT; // Assume its relatively infrequent
1351 else if (btest == BoolTest::ne)
1352 prob = PROB_STATIC_FREQUENT; // Assume its relatively frequent
1353
1354 // If this is a conditional test guarding a backwards branch,
1355 // assume its a loop-back edge. Make it a likely taken branch.
1356 if (target_bci < bci()) {
1357 if (is_osr_parse()) { // Could be a hot OSR'd loop; force deopt
1358 // Since it's an OSR, we probably have profile data, but since
1359 // branch_prediction returned PROB_UNKNOWN, the counts are too small.
1360 // Let's make a special check here for completely zero counts.
1361 ciMethodData* methodData = method()->method_data();
1362 if (!methodData->is_empty()) {
1363 ciProfileData* data = methodData->bci_to_data(bci());
1364 // Only stop for truly zero counts, which mean an unknown part
1365 // of the OSR-ed method, and we want to deopt to gather more stats.
1366 // If you have ANY counts, then this loop is simply 'cold' relative
1367 // to the OSR loop.
1368 if (data == NULL ||
1369 (data->as_BranchData()->taken() + data->as_BranchData()->not_taken() == 0)) {
1370 // This is the only way to return PROB_UNKNOWN:
1371 return PROB_UNKNOWN;
1372 }
1373 }
1374 }
1375 prob = PROB_STATIC_FREQUENT; // Likely to take backwards branch
1376 }
1377
1378 assert(prob != PROB_UNKNOWN, "must have some guess at this point");
1379 return prob;
1380 }
1381
1382 // The magic constants are chosen so as to match the output of
1383 // branch_prediction() when the profile reports a zero taken count.
1384 // It is important to distinguish zero counts unambiguously, because
1385 // some branches (e.g., _213_javac.Assembler.eliminate) validly produce
1386 // very small but nonzero probabilities, which if confused with zero
1387 // counts would keep the program recompiling indefinitely.
seems_never_taken(float prob) const1388 bool Parse::seems_never_taken(float prob) const {
1389 return prob < PROB_MIN;
1390 }
1391
1392 // True if the comparison seems to be the kind that will not change its
1393 // statistics from true to false. See comments in adjust_map_after_if.
1394 // This question is only asked along paths which are already
1395 // classifed as untaken (by seems_never_taken), so really,
1396 // if a path is never taken, its controlling comparison is
1397 // already acting in a stable fashion. If the comparison
1398 // seems stable, we will put an expensive uncommon trap
1399 // on the untaken path.
seems_stable_comparison() const1400 bool Parse::seems_stable_comparison() const {
1401 if (C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if)) {
1402 return false;
1403 }
1404 return true;
1405 }
1406
1407 //-------------------------------repush_if_args--------------------------------
1408 // Push arguments of an "if" bytecode back onto the stack by adjusting _sp.
repush_if_args()1409 inline int Parse::repush_if_args() {
1410 if (PrintOpto && WizardMode) {
1411 tty->print("defending against excessive implicit null exceptions on %s @%d in ",
1412 Bytecodes::name(iter().cur_bc()), iter().cur_bci());
1413 method()->print_name(); tty->cr();
1414 }
1415 int bc_depth = - Bytecodes::depth(iter().cur_bc());
1416 assert(bc_depth == 1 || bc_depth == 2, "only two kinds of branches");
1417 DEBUG_ONLY(sync_jvms()); // argument(n) requires a synced jvms
1418 assert(argument(0) != NULL, "must exist");
1419 assert(bc_depth == 1 || argument(1) != NULL, "two must exist");
1420 inc_sp(bc_depth);
1421 return bc_depth;
1422 }
1423
1424 //----------------------------------do_ifnull----------------------------------
do_ifnull(BoolTest::mask btest,Node * c)1425 void Parse::do_ifnull(BoolTest::mask btest, Node *c) {
1426 int target_bci = iter().get_dest();
1427
1428 Block* branch_block = successor_for_bci(target_bci);
1429 Block* next_block = successor_for_bci(iter().next_bci());
1430
1431 float cnt;
1432 float prob = branch_prediction(cnt, btest, target_bci, c);
1433 if (prob == PROB_UNKNOWN) {
1434 // (An earlier version of do_ifnull omitted this trap for OSR methods.)
1435 if (PrintOpto && Verbose) {
1436 tty->print_cr("Never-taken edge stops compilation at bci %d", bci());
1437 }
1438 repush_if_args(); // to gather stats on loop
1439 uncommon_trap(Deoptimization::Reason_unreached,
1440 Deoptimization::Action_reinterpret,
1441 NULL, "cold");
1442 if (C->eliminate_boxing()) {
1443 // Mark the successor blocks as parsed
1444 branch_block->next_path_num();
1445 next_block->next_path_num();
1446 }
1447 return;
1448 }
1449
1450 NOT_PRODUCT(explicit_null_checks_inserted++);
1451
1452 // Generate real control flow
1453 Node *tst = _gvn.transform( new BoolNode( c, btest ) );
1454
1455 // Sanity check the probability value
1456 assert(prob > 0.0f,"Bad probability in Parser");
1457 // Need xform to put node in hash table
1458 IfNode *iff = create_and_xform_if( control(), tst, prob, cnt );
1459 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser");
1460 // True branch
1461 { PreserveJVMState pjvms(this);
1462 Node* iftrue = _gvn.transform( new IfTrueNode (iff) );
1463 set_control(iftrue);
1464
1465 if (stopped()) { // Path is dead?
1466 NOT_PRODUCT(explicit_null_checks_elided++);
1467 if (C->eliminate_boxing()) {
1468 // Mark the successor block as parsed
1469 branch_block->next_path_num();
1470 }
1471 } else { // Path is live.
1472 adjust_map_after_if(btest, c, prob, branch_block, next_block);
1473 if (!stopped()) {
1474 merge(target_bci);
1475 }
1476 }
1477 }
1478
1479 // False branch
1480 Node* iffalse = _gvn.transform( new IfFalseNode(iff) );
1481 set_control(iffalse);
1482
1483 if (stopped()) { // Path is dead?
1484 NOT_PRODUCT(explicit_null_checks_elided++);
1485 if (C->eliminate_boxing()) {
1486 // Mark the successor block as parsed
1487 next_block->next_path_num();
1488 }
1489 } else { // Path is live.
1490 adjust_map_after_if(BoolTest(btest).negate(), c, 1.0-prob,
1491 next_block, branch_block);
1492 }
1493 }
1494
1495 //------------------------------------do_if------------------------------------
do_if(BoolTest::mask btest,Node * c)1496 void Parse::do_if(BoolTest::mask btest, Node* c) {
1497 int target_bci = iter().get_dest();
1498
1499 Block* branch_block = successor_for_bci(target_bci);
1500 Block* next_block = successor_for_bci(iter().next_bci());
1501
1502 float cnt;
1503 float prob = branch_prediction(cnt, btest, target_bci, c);
1504 float untaken_prob = 1.0 - prob;
1505
1506 if (prob == PROB_UNKNOWN) {
1507 if (PrintOpto && Verbose) {
1508 tty->print_cr("Never-taken edge stops compilation at bci %d", bci());
1509 }
1510 repush_if_args(); // to gather stats on loop
1511 uncommon_trap(Deoptimization::Reason_unreached,
1512 Deoptimization::Action_reinterpret,
1513 NULL, "cold");
1514 if (C->eliminate_boxing()) {
1515 // Mark the successor blocks as parsed
1516 branch_block->next_path_num();
1517 next_block->next_path_num();
1518 }
1519 return;
1520 }
1521
1522 // Sanity check the probability value
1523 assert(0.0f < prob && prob < 1.0f,"Bad probability in Parser");
1524
1525 bool taken_if_true = true;
1526 // Convert BoolTest to canonical form:
1527 if (!BoolTest(btest).is_canonical()) {
1528 btest = BoolTest(btest).negate();
1529 taken_if_true = false;
1530 // prob is NOT updated here; it remains the probability of the taken
1531 // path (as opposed to the prob of the path guarded by an 'IfTrueNode').
1532 }
1533 assert(btest != BoolTest::eq, "!= is the only canonical exact test");
1534
1535 Node* tst0 = new BoolNode(c, btest);
1536 Node* tst = _gvn.transform(tst0);
1537 BoolTest::mask taken_btest = BoolTest::illegal;
1538 BoolTest::mask untaken_btest = BoolTest::illegal;
1539
1540 if (tst->is_Bool()) {
1541 // Refresh c from the transformed bool node, since it may be
1542 // simpler than the original c. Also re-canonicalize btest.
1543 // This wins when (Bool ne (Conv2B p) 0) => (Bool ne (CmpP p NULL)).
1544 // That can arise from statements like: if (x instanceof C) ...
1545 if (tst != tst0) {
1546 // Canonicalize one more time since transform can change it.
1547 btest = tst->as_Bool()->_test._test;
1548 if (!BoolTest(btest).is_canonical()) {
1549 // Reverse edges one more time...
1550 tst = _gvn.transform( tst->as_Bool()->negate(&_gvn) );
1551 btest = tst->as_Bool()->_test._test;
1552 assert(BoolTest(btest).is_canonical(), "sanity");
1553 taken_if_true = !taken_if_true;
1554 }
1555 c = tst->in(1);
1556 }
1557 BoolTest::mask neg_btest = BoolTest(btest).negate();
1558 taken_btest = taken_if_true ? btest : neg_btest;
1559 untaken_btest = taken_if_true ? neg_btest : btest;
1560 }
1561
1562 // Generate real control flow
1563 float true_prob = (taken_if_true ? prob : untaken_prob);
1564 IfNode* iff = create_and_map_if(control(), tst, true_prob, cnt);
1565 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser");
1566 Node* taken_branch = new IfTrueNode(iff);
1567 Node* untaken_branch = new IfFalseNode(iff);
1568 if (!taken_if_true) { // Finish conversion to canonical form
1569 Node* tmp = taken_branch;
1570 taken_branch = untaken_branch;
1571 untaken_branch = tmp;
1572 }
1573
1574 // Branch is taken:
1575 { PreserveJVMState pjvms(this);
1576 taken_branch = _gvn.transform(taken_branch);
1577 set_control(taken_branch);
1578
1579 if (stopped()) {
1580 if (C->eliminate_boxing()) {
1581 // Mark the successor block as parsed
1582 branch_block->next_path_num();
1583 }
1584 } else {
1585 adjust_map_after_if(taken_btest, c, prob, branch_block, next_block);
1586 if (!stopped()) {
1587 merge(target_bci);
1588 }
1589 }
1590 }
1591
1592 untaken_branch = _gvn.transform(untaken_branch);
1593 set_control(untaken_branch);
1594
1595 // Branch not taken.
1596 if (stopped()) {
1597 if (C->eliminate_boxing()) {
1598 // Mark the successor block as parsed
1599 next_block->next_path_num();
1600 }
1601 } else {
1602 adjust_map_after_if(untaken_btest, c, untaken_prob,
1603 next_block, branch_block);
1604 }
1605 }
1606
path_is_suitable_for_uncommon_trap(float prob) const1607 bool Parse::path_is_suitable_for_uncommon_trap(float prob) const {
1608 // Don't want to speculate on uncommon traps when running with -Xcomp
1609 if (!UseInterpreter) {
1610 return false;
1611 }
1612 return (seems_never_taken(prob) && seems_stable_comparison());
1613 }
1614
maybe_add_predicate_after_if(Block * path)1615 void Parse::maybe_add_predicate_after_if(Block* path) {
1616 if (path->is_SEL_head() && path->preds_parsed() == 0) {
1617 // Add predicates at bci of if dominating the loop so traps can be
1618 // recorded on the if's profile data
1619 int bc_depth = repush_if_args();
1620 add_empty_predicates();
1621 dec_sp(bc_depth);
1622 path->set_has_predicates();
1623 }
1624 }
1625
1626
1627 //----------------------------adjust_map_after_if------------------------------
1628 // Adjust the JVM state to reflect the result of taking this path.
1629 // Basically, it means inspecting the CmpNode controlling this
1630 // branch, seeing how it constrains a tested value, and then
1631 // deciding if it's worth our while to encode this constraint
1632 // as graph nodes in the current abstract interpretation map.
adjust_map_after_if(BoolTest::mask btest,Node * c,float prob,Block * path,Block * other_path)1633 void Parse::adjust_map_after_if(BoolTest::mask btest, Node* c, float prob,
1634 Block* path, Block* other_path) {
1635 if (!c->is_Cmp()) {
1636 maybe_add_predicate_after_if(path);
1637 return;
1638 }
1639
1640 if (stopped() || btest == BoolTest::illegal) {
1641 return; // nothing to do
1642 }
1643
1644 bool is_fallthrough = (path == successor_for_bci(iter().next_bci()));
1645
1646 if (path_is_suitable_for_uncommon_trap(prob)) {
1647 repush_if_args();
1648 uncommon_trap(Deoptimization::Reason_unstable_if,
1649 Deoptimization::Action_reinterpret,
1650 NULL,
1651 (is_fallthrough ? "taken always" : "taken never"));
1652 return;
1653 }
1654
1655 Node* val = c->in(1);
1656 Node* con = c->in(2);
1657 const Type* tcon = _gvn.type(con);
1658 const Type* tval = _gvn.type(val);
1659 bool have_con = tcon->singleton();
1660 if (tval->singleton()) {
1661 if (!have_con) {
1662 // Swap, so constant is in con.
1663 con = val;
1664 tcon = tval;
1665 val = c->in(2);
1666 tval = _gvn.type(val);
1667 btest = BoolTest(btest).commute();
1668 have_con = true;
1669 } else {
1670 // Do we have two constants? Then leave well enough alone.
1671 have_con = false;
1672 }
1673 }
1674 if (!have_con) { // remaining adjustments need a con
1675 maybe_add_predicate_after_if(path);
1676 return;
1677 }
1678
1679 sharpen_type_after_if(btest, con, tcon, val, tval);
1680 maybe_add_predicate_after_if(path);
1681 }
1682
1683
extract_obj_from_klass_load(PhaseGVN * gvn,Node * n)1684 static Node* extract_obj_from_klass_load(PhaseGVN* gvn, Node* n) {
1685 Node* ldk;
1686 if (n->is_DecodeNKlass()) {
1687 if (n->in(1)->Opcode() != Op_LoadNKlass) {
1688 return NULL;
1689 } else {
1690 ldk = n->in(1);
1691 }
1692 } else if (n->Opcode() != Op_LoadKlass) {
1693 return NULL;
1694 } else {
1695 ldk = n;
1696 }
1697 assert(ldk != NULL && ldk->is_Load(), "should have found a LoadKlass or LoadNKlass node");
1698
1699 Node* adr = ldk->in(MemNode::Address);
1700 intptr_t off = 0;
1701 Node* obj = AddPNode::Ideal_base_and_offset(adr, gvn, off);
1702 if (obj == NULL || off != oopDesc::klass_offset_in_bytes()) // loading oopDesc::_klass?
1703 return NULL;
1704 const TypePtr* tp = gvn->type(obj)->is_ptr();
1705 if (tp == NULL || !(tp->isa_instptr() || tp->isa_aryptr())) // is obj a Java object ptr?
1706 return NULL;
1707
1708 return obj;
1709 }
1710
sharpen_type_after_if(BoolTest::mask btest,Node * con,const Type * tcon,Node * val,const Type * tval)1711 void Parse::sharpen_type_after_if(BoolTest::mask btest,
1712 Node* con, const Type* tcon,
1713 Node* val, const Type* tval) {
1714 // Look for opportunities to sharpen the type of a node
1715 // whose klass is compared with a constant klass.
1716 if (btest == BoolTest::eq && tcon->isa_klassptr()) {
1717 Node* obj = extract_obj_from_klass_load(&_gvn, val);
1718 const TypeOopPtr* con_type = tcon->isa_klassptr()->as_instance_type();
1719 if (obj != NULL && (con_type->isa_instptr() || con_type->isa_aryptr())) {
1720 // Found:
1721 // Bool(CmpP(LoadKlass(obj._klass), ConP(Foo.klass)), [eq])
1722 // or the narrowOop equivalent.
1723 const Type* obj_type = _gvn.type(obj);
1724 const TypeOopPtr* tboth = obj_type->join_speculative(con_type)->isa_oopptr();
1725 if (tboth != NULL && tboth->klass_is_exact() && tboth != obj_type &&
1726 tboth->higher_equal(obj_type)) {
1727 // obj has to be of the exact type Foo if the CmpP succeeds.
1728 int obj_in_map = map()->find_edge(obj);
1729 JVMState* jvms = this->jvms();
1730 if (obj_in_map >= 0 &&
1731 (jvms->is_loc(obj_in_map) || jvms->is_stk(obj_in_map))) {
1732 TypeNode* ccast = new CheckCastPPNode(control(), obj, tboth);
1733 const Type* tcc = ccast->as_Type()->type();
1734 assert(tcc != obj_type && tcc->higher_equal(obj_type), "must improve");
1735 // Delay transform() call to allow recovery of pre-cast value
1736 // at the control merge.
1737 _gvn.set_type_bottom(ccast);
1738 record_for_igvn(ccast);
1739 // Here's the payoff.
1740 replace_in_map(obj, ccast);
1741 }
1742 }
1743 }
1744 }
1745
1746 int val_in_map = map()->find_edge(val);
1747 if (val_in_map < 0) return; // replace_in_map would be useless
1748 {
1749 JVMState* jvms = this->jvms();
1750 if (!(jvms->is_loc(val_in_map) ||
1751 jvms->is_stk(val_in_map)))
1752 return; // again, it would be useless
1753 }
1754
1755 // Check for a comparison to a constant, and "know" that the compared
1756 // value is constrained on this path.
1757 assert(tcon->singleton(), "");
1758 ConstraintCastNode* ccast = NULL;
1759 Node* cast = NULL;
1760
1761 switch (btest) {
1762 case BoolTest::eq: // Constant test?
1763 {
1764 const Type* tboth = tcon->join_speculative(tval);
1765 if (tboth == tval) break; // Nothing to gain.
1766 if (tcon->isa_int()) {
1767 ccast = new CastIINode(val, tboth);
1768 } else if (tcon == TypePtr::NULL_PTR) {
1769 // Cast to null, but keep the pointer identity temporarily live.
1770 ccast = new CastPPNode(val, tboth);
1771 } else {
1772 const TypeF* tf = tcon->isa_float_constant();
1773 const TypeD* td = tcon->isa_double_constant();
1774 // Exclude tests vs float/double 0 as these could be
1775 // either +0 or -0. Just because you are equal to +0
1776 // doesn't mean you ARE +0!
1777 // Note, following code also replaces Long and Oop values.
1778 if ((!tf || tf->_f != 0.0) &&
1779 (!td || td->_d != 0.0))
1780 cast = con; // Replace non-constant val by con.
1781 }
1782 }
1783 break;
1784
1785 case BoolTest::ne:
1786 if (tcon == TypePtr::NULL_PTR) {
1787 cast = cast_not_null(val, false);
1788 }
1789 break;
1790
1791 default:
1792 // (At this point we could record int range types with CastII.)
1793 break;
1794 }
1795
1796 if (ccast != NULL) {
1797 const Type* tcc = ccast->as_Type()->type();
1798 assert(tcc != tval && tcc->higher_equal(tval), "must improve");
1799 // Delay transform() call to allow recovery of pre-cast value
1800 // at the control merge.
1801 ccast->set_req(0, control());
1802 _gvn.set_type_bottom(ccast);
1803 record_for_igvn(ccast);
1804 cast = ccast;
1805 }
1806
1807 if (cast != NULL) { // Here's the payoff.
1808 replace_in_map(val, cast);
1809 }
1810 }
1811
1812 /**
1813 * Use speculative type to optimize CmpP node: if comparison is
1814 * against the low level class, cast the object to the speculative
1815 * type if any. CmpP should then go away.
1816 *
1817 * @param c expected CmpP node
1818 * @return result of CmpP on object casted to speculative type
1819 *
1820 */
optimize_cmp_with_klass(Node * c)1821 Node* Parse::optimize_cmp_with_klass(Node* c) {
1822 // If this is transformed by the _gvn to a comparison with the low
1823 // level klass then we may be able to use speculation
1824 if (c->Opcode() == Op_CmpP &&
1825 (c->in(1)->Opcode() == Op_LoadKlass || c->in(1)->Opcode() == Op_DecodeNKlass) &&
1826 c->in(2)->is_Con()) {
1827 Node* load_klass = NULL;
1828 Node* decode = NULL;
1829 if (c->in(1)->Opcode() == Op_DecodeNKlass) {
1830 decode = c->in(1);
1831 load_klass = c->in(1)->in(1);
1832 } else {
1833 load_klass = c->in(1);
1834 }
1835 if (load_klass->in(2)->is_AddP()) {
1836 Node* addp = load_klass->in(2);
1837 Node* obj = addp->in(AddPNode::Address);
1838 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
1839 if (obj_type->speculative_type_not_null() != NULL) {
1840 ciKlass* k = obj_type->speculative_type();
1841 inc_sp(2);
1842 obj = maybe_cast_profiled_obj(obj, k);
1843 dec_sp(2);
1844 // Make the CmpP use the casted obj
1845 addp = basic_plus_adr(obj, addp->in(AddPNode::Offset));
1846 load_klass = load_klass->clone();
1847 load_klass->set_req(2, addp);
1848 load_klass = _gvn.transform(load_klass);
1849 if (decode != NULL) {
1850 decode = decode->clone();
1851 decode->set_req(1, load_klass);
1852 load_klass = _gvn.transform(decode);
1853 }
1854 c = c->clone();
1855 c->set_req(1, load_klass);
1856 c = _gvn.transform(c);
1857 }
1858 }
1859 }
1860 return c;
1861 }
1862
1863 //------------------------------do_one_bytecode--------------------------------
1864 // Parse this bytecode, and alter the Parsers JVM->Node mapping
do_one_bytecode()1865 void Parse::do_one_bytecode() {
1866 Node *a, *b, *c, *d; // Handy temps
1867 BoolTest::mask btest;
1868 int i;
1869
1870 assert(!has_exceptions(), "bytecode entry state must be clear of throws");
1871
1872 if (C->check_node_count(NodeLimitFudgeFactor * 5,
1873 "out of nodes parsing method")) {
1874 return;
1875 }
1876
1877 #ifdef ASSERT
1878 // for setting breakpoints
1879 if (TraceOptoParse) {
1880 tty->print(" @");
1881 dump_bci(bci());
1882 tty->cr();
1883 }
1884 #endif
1885
1886 switch (bc()) {
1887 case Bytecodes::_nop:
1888 // do nothing
1889 break;
1890 case Bytecodes::_lconst_0:
1891 push_pair(longcon(0));
1892 break;
1893
1894 case Bytecodes::_lconst_1:
1895 push_pair(longcon(1));
1896 break;
1897
1898 case Bytecodes::_fconst_0:
1899 push(zerocon(T_FLOAT));
1900 break;
1901
1902 case Bytecodes::_fconst_1:
1903 push(makecon(TypeF::ONE));
1904 break;
1905
1906 case Bytecodes::_fconst_2:
1907 push(makecon(TypeF::make(2.0f)));
1908 break;
1909
1910 case Bytecodes::_dconst_0:
1911 push_pair(zerocon(T_DOUBLE));
1912 break;
1913
1914 case Bytecodes::_dconst_1:
1915 push_pair(makecon(TypeD::ONE));
1916 break;
1917
1918 case Bytecodes::_iconst_m1:push(intcon(-1)); break;
1919 case Bytecodes::_iconst_0: push(intcon( 0)); break;
1920 case Bytecodes::_iconst_1: push(intcon( 1)); break;
1921 case Bytecodes::_iconst_2: push(intcon( 2)); break;
1922 case Bytecodes::_iconst_3: push(intcon( 3)); break;
1923 case Bytecodes::_iconst_4: push(intcon( 4)); break;
1924 case Bytecodes::_iconst_5: push(intcon( 5)); break;
1925 case Bytecodes::_bipush: push(intcon(iter().get_constant_u1())); break;
1926 case Bytecodes::_sipush: push(intcon(iter().get_constant_u2())); break;
1927 case Bytecodes::_aconst_null: push(null()); break;
1928 case Bytecodes::_ldc:
1929 case Bytecodes::_ldc_w:
1930 case Bytecodes::_ldc2_w:
1931 // If the constant is unresolved, run this BC once in the interpreter.
1932 {
1933 ciConstant constant = iter().get_constant();
1934 if (!constant.is_valid() ||
1935 (constant.basic_type() == T_OBJECT &&
1936 !constant.as_object()->is_loaded())) {
1937 int index = iter().get_constant_pool_index();
1938 constantTag tag = iter().get_constant_pool_tag(index);
1939 uncommon_trap(Deoptimization::make_trap_request
1940 (Deoptimization::Reason_unloaded,
1941 Deoptimization::Action_reinterpret,
1942 index),
1943 NULL, tag.internal_name());
1944 break;
1945 }
1946 assert(constant.basic_type() != T_OBJECT || constant.as_object()->is_instance(),
1947 "must be java_mirror of klass");
1948 const Type* con_type = Type::make_from_constant(constant);
1949 if (con_type != NULL) {
1950 push_node(con_type->basic_type(), makecon(con_type));
1951 }
1952 }
1953
1954 break;
1955
1956 case Bytecodes::_aload_0:
1957 push( local(0) );
1958 break;
1959 case Bytecodes::_aload_1:
1960 push( local(1) );
1961 break;
1962 case Bytecodes::_aload_2:
1963 push( local(2) );
1964 break;
1965 case Bytecodes::_aload_3:
1966 push( local(3) );
1967 break;
1968 case Bytecodes::_aload:
1969 push( local(iter().get_index()) );
1970 break;
1971
1972 case Bytecodes::_fload_0:
1973 case Bytecodes::_iload_0:
1974 push( local(0) );
1975 break;
1976 case Bytecodes::_fload_1:
1977 case Bytecodes::_iload_1:
1978 push( local(1) );
1979 break;
1980 case Bytecodes::_fload_2:
1981 case Bytecodes::_iload_2:
1982 push( local(2) );
1983 break;
1984 case Bytecodes::_fload_3:
1985 case Bytecodes::_iload_3:
1986 push( local(3) );
1987 break;
1988 case Bytecodes::_fload:
1989 case Bytecodes::_iload:
1990 push( local(iter().get_index()) );
1991 break;
1992 case Bytecodes::_lload_0:
1993 push_pair_local( 0 );
1994 break;
1995 case Bytecodes::_lload_1:
1996 push_pair_local( 1 );
1997 break;
1998 case Bytecodes::_lload_2:
1999 push_pair_local( 2 );
2000 break;
2001 case Bytecodes::_lload_3:
2002 push_pair_local( 3 );
2003 break;
2004 case Bytecodes::_lload:
2005 push_pair_local( iter().get_index() );
2006 break;
2007
2008 case Bytecodes::_dload_0:
2009 push_pair_local(0);
2010 break;
2011 case Bytecodes::_dload_1:
2012 push_pair_local(1);
2013 break;
2014 case Bytecodes::_dload_2:
2015 push_pair_local(2);
2016 break;
2017 case Bytecodes::_dload_3:
2018 push_pair_local(3);
2019 break;
2020 case Bytecodes::_dload:
2021 push_pair_local(iter().get_index());
2022 break;
2023 case Bytecodes::_fstore_0:
2024 case Bytecodes::_istore_0:
2025 case Bytecodes::_astore_0:
2026 set_local( 0, pop() );
2027 break;
2028 case Bytecodes::_fstore_1:
2029 case Bytecodes::_istore_1:
2030 case Bytecodes::_astore_1:
2031 set_local( 1, pop() );
2032 break;
2033 case Bytecodes::_fstore_2:
2034 case Bytecodes::_istore_2:
2035 case Bytecodes::_astore_2:
2036 set_local( 2, pop() );
2037 break;
2038 case Bytecodes::_fstore_3:
2039 case Bytecodes::_istore_3:
2040 case Bytecodes::_astore_3:
2041 set_local( 3, pop() );
2042 break;
2043 case Bytecodes::_fstore:
2044 case Bytecodes::_istore:
2045 case Bytecodes::_astore:
2046 set_local( iter().get_index(), pop() );
2047 break;
2048 // long stores
2049 case Bytecodes::_lstore_0:
2050 set_pair_local( 0, pop_pair() );
2051 break;
2052 case Bytecodes::_lstore_1:
2053 set_pair_local( 1, pop_pair() );
2054 break;
2055 case Bytecodes::_lstore_2:
2056 set_pair_local( 2, pop_pair() );
2057 break;
2058 case Bytecodes::_lstore_3:
2059 set_pair_local( 3, pop_pair() );
2060 break;
2061 case Bytecodes::_lstore:
2062 set_pair_local( iter().get_index(), pop_pair() );
2063 break;
2064
2065 // double stores
2066 case Bytecodes::_dstore_0:
2067 set_pair_local( 0, dstore_rounding(pop_pair()) );
2068 break;
2069 case Bytecodes::_dstore_1:
2070 set_pair_local( 1, dstore_rounding(pop_pair()) );
2071 break;
2072 case Bytecodes::_dstore_2:
2073 set_pair_local( 2, dstore_rounding(pop_pair()) );
2074 break;
2075 case Bytecodes::_dstore_3:
2076 set_pair_local( 3, dstore_rounding(pop_pair()) );
2077 break;
2078 case Bytecodes::_dstore:
2079 set_pair_local( iter().get_index(), dstore_rounding(pop_pair()) );
2080 break;
2081
2082 case Bytecodes::_pop: dec_sp(1); break;
2083 case Bytecodes::_pop2: dec_sp(2); break;
2084 case Bytecodes::_swap:
2085 a = pop();
2086 b = pop();
2087 push(a);
2088 push(b);
2089 break;
2090 case Bytecodes::_dup:
2091 a = pop();
2092 push(a);
2093 push(a);
2094 break;
2095 case Bytecodes::_dup_x1:
2096 a = pop();
2097 b = pop();
2098 push( a );
2099 push( b );
2100 push( a );
2101 break;
2102 case Bytecodes::_dup_x2:
2103 a = pop();
2104 b = pop();
2105 c = pop();
2106 push( a );
2107 push( c );
2108 push( b );
2109 push( a );
2110 break;
2111 case Bytecodes::_dup2:
2112 a = pop();
2113 b = pop();
2114 push( b );
2115 push( a );
2116 push( b );
2117 push( a );
2118 break;
2119
2120 case Bytecodes::_dup2_x1:
2121 // before: .. c, b, a
2122 // after: .. b, a, c, b, a
2123 // not tested
2124 a = pop();
2125 b = pop();
2126 c = pop();
2127 push( b );
2128 push( a );
2129 push( c );
2130 push( b );
2131 push( a );
2132 break;
2133 case Bytecodes::_dup2_x2:
2134 // before: .. d, c, b, a
2135 // after: .. b, a, d, c, b, a
2136 // not tested
2137 a = pop();
2138 b = pop();
2139 c = pop();
2140 d = pop();
2141 push( b );
2142 push( a );
2143 push( d );
2144 push( c );
2145 push( b );
2146 push( a );
2147 break;
2148
2149 case Bytecodes::_arraylength: {
2150 // Must do null-check with value on expression stack
2151 Node *ary = null_check(peek(), T_ARRAY);
2152 // Compile-time detect of null-exception?
2153 if (stopped()) return;
2154 a = pop();
2155 push(load_array_length(a));
2156 break;
2157 }
2158
2159 case Bytecodes::_baload: array_load(T_BYTE); break;
2160 case Bytecodes::_caload: array_load(T_CHAR); break;
2161 case Bytecodes::_iaload: array_load(T_INT); break;
2162 case Bytecodes::_saload: array_load(T_SHORT); break;
2163 case Bytecodes::_faload: array_load(T_FLOAT); break;
2164 case Bytecodes::_aaload: array_load(T_OBJECT); break;
2165 case Bytecodes::_laload: array_load(T_LONG); break;
2166 case Bytecodes::_daload: array_load(T_DOUBLE); break;
2167 case Bytecodes::_bastore: array_store(T_BYTE); break;
2168 case Bytecodes::_castore: array_store(T_CHAR); break;
2169 case Bytecodes::_iastore: array_store(T_INT); break;
2170 case Bytecodes::_sastore: array_store(T_SHORT); break;
2171 case Bytecodes::_fastore: array_store(T_FLOAT); break;
2172 case Bytecodes::_aastore: array_store(T_OBJECT); break;
2173 case Bytecodes::_lastore: array_store(T_LONG); break;
2174 case Bytecodes::_dastore: array_store(T_DOUBLE); break;
2175
2176 case Bytecodes::_getfield:
2177 do_getfield();
2178 break;
2179
2180 case Bytecodes::_getstatic:
2181 do_getstatic();
2182 break;
2183
2184 case Bytecodes::_putfield:
2185 do_putfield();
2186 break;
2187
2188 case Bytecodes::_putstatic:
2189 do_putstatic();
2190 break;
2191
2192 case Bytecodes::_irem:
2193 do_irem();
2194 break;
2195 case Bytecodes::_idiv:
2196 // Must keep both values on the expression-stack during null-check
2197 zero_check_int(peek());
2198 // Compile-time detect of null-exception?
2199 if (stopped()) return;
2200 b = pop();
2201 a = pop();
2202 push( _gvn.transform( new DivINode(control(),a,b) ) );
2203 break;
2204 case Bytecodes::_imul:
2205 b = pop(); a = pop();
2206 push( _gvn.transform( new MulINode(a,b) ) );
2207 break;
2208 case Bytecodes::_iadd:
2209 b = pop(); a = pop();
2210 push( _gvn.transform( new AddINode(a,b) ) );
2211 break;
2212 case Bytecodes::_ineg:
2213 a = pop();
2214 push( _gvn.transform( new SubINode(_gvn.intcon(0),a)) );
2215 break;
2216 case Bytecodes::_isub:
2217 b = pop(); a = pop();
2218 push( _gvn.transform( new SubINode(a,b) ) );
2219 break;
2220 case Bytecodes::_iand:
2221 b = pop(); a = pop();
2222 push( _gvn.transform( new AndINode(a,b) ) );
2223 break;
2224 case Bytecodes::_ior:
2225 b = pop(); a = pop();
2226 push( _gvn.transform( new OrINode(a,b) ) );
2227 break;
2228 case Bytecodes::_ixor:
2229 b = pop(); a = pop();
2230 push( _gvn.transform( new XorINode(a,b) ) );
2231 break;
2232 case Bytecodes::_ishl:
2233 b = pop(); a = pop();
2234 push( _gvn.transform( new LShiftINode(a,b) ) );
2235 break;
2236 case Bytecodes::_ishr:
2237 b = pop(); a = pop();
2238 push( _gvn.transform( new RShiftINode(a,b) ) );
2239 break;
2240 case Bytecodes::_iushr:
2241 b = pop(); a = pop();
2242 push( _gvn.transform( new URShiftINode(a,b) ) );
2243 break;
2244
2245 case Bytecodes::_fneg:
2246 a = pop();
2247 b = _gvn.transform(new NegFNode (a));
2248 push(b);
2249 break;
2250
2251 case Bytecodes::_fsub:
2252 b = pop();
2253 a = pop();
2254 c = _gvn.transform( new SubFNode(a,b) );
2255 d = precision_rounding(c);
2256 push( d );
2257 break;
2258
2259 case Bytecodes::_fadd:
2260 b = pop();
2261 a = pop();
2262 c = _gvn.transform( new AddFNode(a,b) );
2263 d = precision_rounding(c);
2264 push( d );
2265 break;
2266
2267 case Bytecodes::_fmul:
2268 b = pop();
2269 a = pop();
2270 c = _gvn.transform( new MulFNode(a,b) );
2271 d = precision_rounding(c);
2272 push( d );
2273 break;
2274
2275 case Bytecodes::_fdiv:
2276 b = pop();
2277 a = pop();
2278 c = _gvn.transform( new DivFNode(0,a,b) );
2279 d = precision_rounding(c);
2280 push( d );
2281 break;
2282
2283 case Bytecodes::_frem:
2284 if (Matcher::has_match_rule(Op_ModF)) {
2285 // Generate a ModF node.
2286 b = pop();
2287 a = pop();
2288 c = _gvn.transform( new ModFNode(0,a,b) );
2289 d = precision_rounding(c);
2290 push( d );
2291 }
2292 else {
2293 // Generate a call.
2294 modf();
2295 }
2296 break;
2297
2298 case Bytecodes::_fcmpl:
2299 b = pop();
2300 a = pop();
2301 c = _gvn.transform( new CmpF3Node( a, b));
2302 push(c);
2303 break;
2304 case Bytecodes::_fcmpg:
2305 b = pop();
2306 a = pop();
2307
2308 // Same as fcmpl but need to flip the unordered case. Swap the inputs,
2309 // which negates the result sign except for unordered. Flip the unordered
2310 // as well by using CmpF3 which implements unordered-lesser instead of
2311 // unordered-greater semantics. Finally, commute the result bits. Result
2312 // is same as using a CmpF3Greater except we did it with CmpF3 alone.
2313 c = _gvn.transform( new CmpF3Node( b, a));
2314 c = _gvn.transform( new SubINode(_gvn.intcon(0),c) );
2315 push(c);
2316 break;
2317
2318 case Bytecodes::_f2i:
2319 a = pop();
2320 push(_gvn.transform(new ConvF2INode(a)));
2321 break;
2322
2323 case Bytecodes::_d2i:
2324 a = pop_pair();
2325 b = _gvn.transform(new ConvD2INode(a));
2326 push( b );
2327 break;
2328
2329 case Bytecodes::_f2d:
2330 a = pop();
2331 b = _gvn.transform( new ConvF2DNode(a));
2332 push_pair( b );
2333 break;
2334
2335 case Bytecodes::_d2f:
2336 a = pop_pair();
2337 b = _gvn.transform( new ConvD2FNode(a));
2338 // This breaks _227_mtrt (speed & correctness) and _222_mpegaudio (speed)
2339 //b = _gvn.transform(new RoundFloatNode(0, b) );
2340 push( b );
2341 break;
2342
2343 case Bytecodes::_l2f:
2344 if (Matcher::convL2FSupported()) {
2345 a = pop_pair();
2346 b = _gvn.transform( new ConvL2FNode(a));
2347 // For x86_32.ad, FILD doesn't restrict precision to 24 or 53 bits.
2348 // Rather than storing the result into an FP register then pushing
2349 // out to memory to round, the machine instruction that implements
2350 // ConvL2D is responsible for rounding.
2351 // c = precision_rounding(b);
2352 c = _gvn.transform(b);
2353 push(c);
2354 } else {
2355 l2f();
2356 }
2357 break;
2358
2359 case Bytecodes::_l2d:
2360 a = pop_pair();
2361 b = _gvn.transform( new ConvL2DNode(a));
2362 // For x86_32.ad, rounding is always necessary (see _l2f above).
2363 // c = dprecision_rounding(b);
2364 c = _gvn.transform(b);
2365 push_pair(c);
2366 break;
2367
2368 case Bytecodes::_f2l:
2369 a = pop();
2370 b = _gvn.transform( new ConvF2LNode(a));
2371 push_pair(b);
2372 break;
2373
2374 case Bytecodes::_d2l:
2375 a = pop_pair();
2376 b = _gvn.transform( new ConvD2LNode(a));
2377 push_pair(b);
2378 break;
2379
2380 case Bytecodes::_dsub:
2381 b = pop_pair();
2382 a = pop_pair();
2383 c = _gvn.transform( new SubDNode(a,b) );
2384 d = dprecision_rounding(c);
2385 push_pair( d );
2386 break;
2387
2388 case Bytecodes::_dadd:
2389 b = pop_pair();
2390 a = pop_pair();
2391 c = _gvn.transform( new AddDNode(a,b) );
2392 d = dprecision_rounding(c);
2393 push_pair( d );
2394 break;
2395
2396 case Bytecodes::_dmul:
2397 b = pop_pair();
2398 a = pop_pair();
2399 c = _gvn.transform( new MulDNode(a,b) );
2400 d = dprecision_rounding(c);
2401 push_pair( d );
2402 break;
2403
2404 case Bytecodes::_ddiv:
2405 b = pop_pair();
2406 a = pop_pair();
2407 c = _gvn.transform( new DivDNode(0,a,b) );
2408 d = dprecision_rounding(c);
2409 push_pair( d );
2410 break;
2411
2412 case Bytecodes::_dneg:
2413 a = pop_pair();
2414 b = _gvn.transform(new NegDNode (a));
2415 push_pair(b);
2416 break;
2417
2418 case Bytecodes::_drem:
2419 if (Matcher::has_match_rule(Op_ModD)) {
2420 // Generate a ModD node.
2421 b = pop_pair();
2422 a = pop_pair();
2423 // a % b
2424
2425 c = _gvn.transform( new ModDNode(0,a,b) );
2426 d = dprecision_rounding(c);
2427 push_pair( d );
2428 }
2429 else {
2430 // Generate a call.
2431 modd();
2432 }
2433 break;
2434
2435 case Bytecodes::_dcmpl:
2436 b = pop_pair();
2437 a = pop_pair();
2438 c = _gvn.transform( new CmpD3Node( a, b));
2439 push(c);
2440 break;
2441
2442 case Bytecodes::_dcmpg:
2443 b = pop_pair();
2444 a = pop_pair();
2445 // Same as dcmpl but need to flip the unordered case.
2446 // Commute the inputs, which negates the result sign except for unordered.
2447 // Flip the unordered as well by using CmpD3 which implements
2448 // unordered-lesser instead of unordered-greater semantics.
2449 // Finally, negate the result bits. Result is same as using a
2450 // CmpD3Greater except we did it with CmpD3 alone.
2451 c = _gvn.transform( new CmpD3Node( b, a));
2452 c = _gvn.transform( new SubINode(_gvn.intcon(0),c) );
2453 push(c);
2454 break;
2455
2456
2457 // Note for longs -> lo word is on TOS, hi word is on TOS - 1
2458 case Bytecodes::_land:
2459 b = pop_pair();
2460 a = pop_pair();
2461 c = _gvn.transform( new AndLNode(a,b) );
2462 push_pair(c);
2463 break;
2464 case Bytecodes::_lor:
2465 b = pop_pair();
2466 a = pop_pair();
2467 c = _gvn.transform( new OrLNode(a,b) );
2468 push_pair(c);
2469 break;
2470 case Bytecodes::_lxor:
2471 b = pop_pair();
2472 a = pop_pair();
2473 c = _gvn.transform( new XorLNode(a,b) );
2474 push_pair(c);
2475 break;
2476
2477 case Bytecodes::_lshl:
2478 b = pop(); // the shift count
2479 a = pop_pair(); // value to be shifted
2480 c = _gvn.transform( new LShiftLNode(a,b) );
2481 push_pair(c);
2482 break;
2483 case Bytecodes::_lshr:
2484 b = pop(); // the shift count
2485 a = pop_pair(); // value to be shifted
2486 c = _gvn.transform( new RShiftLNode(a,b) );
2487 push_pair(c);
2488 break;
2489 case Bytecodes::_lushr:
2490 b = pop(); // the shift count
2491 a = pop_pair(); // value to be shifted
2492 c = _gvn.transform( new URShiftLNode(a,b) );
2493 push_pair(c);
2494 break;
2495 case Bytecodes::_lmul:
2496 b = pop_pair();
2497 a = pop_pair();
2498 c = _gvn.transform( new MulLNode(a,b) );
2499 push_pair(c);
2500 break;
2501
2502 case Bytecodes::_lrem:
2503 // Must keep both values on the expression-stack during null-check
2504 assert(peek(0) == top(), "long word order");
2505 zero_check_long(peek(1));
2506 // Compile-time detect of null-exception?
2507 if (stopped()) return;
2508 b = pop_pair();
2509 a = pop_pair();
2510 c = _gvn.transform( new ModLNode(control(),a,b) );
2511 push_pair(c);
2512 break;
2513
2514 case Bytecodes::_ldiv:
2515 // Must keep both values on the expression-stack during null-check
2516 assert(peek(0) == top(), "long word order");
2517 zero_check_long(peek(1));
2518 // Compile-time detect of null-exception?
2519 if (stopped()) return;
2520 b = pop_pair();
2521 a = pop_pair();
2522 c = _gvn.transform( new DivLNode(control(),a,b) );
2523 push_pair(c);
2524 break;
2525
2526 case Bytecodes::_ladd:
2527 b = pop_pair();
2528 a = pop_pair();
2529 c = _gvn.transform( new AddLNode(a,b) );
2530 push_pair(c);
2531 break;
2532 case Bytecodes::_lsub:
2533 b = pop_pair();
2534 a = pop_pair();
2535 c = _gvn.transform( new SubLNode(a,b) );
2536 push_pair(c);
2537 break;
2538 case Bytecodes::_lcmp:
2539 // Safepoints are now inserted _before_ branches. The long-compare
2540 // bytecode painfully produces a 3-way value (-1,0,+1) which requires a
2541 // slew of control flow. These are usually followed by a CmpI vs zero and
2542 // a branch; this pattern then optimizes to the obvious long-compare and
2543 // branch. However, if the branch is backwards there's a Safepoint
2544 // inserted. The inserted Safepoint captures the JVM state at the
2545 // pre-branch point, i.e. it captures the 3-way value. Thus if a
2546 // long-compare is used to control a loop the debug info will force
2547 // computation of the 3-way value, even though the generated code uses a
2548 // long-compare and branch. We try to rectify the situation by inserting
2549 // a SafePoint here and have it dominate and kill the safepoint added at a
2550 // following backwards branch. At this point the JVM state merely holds 2
2551 // longs but not the 3-way value.
2552 if( UseLoopSafepoints ) {
2553 switch( iter().next_bc() ) {
2554 case Bytecodes::_ifgt:
2555 case Bytecodes::_iflt:
2556 case Bytecodes::_ifge:
2557 case Bytecodes::_ifle:
2558 case Bytecodes::_ifne:
2559 case Bytecodes::_ifeq:
2560 // If this is a backwards branch in the bytecodes, add Safepoint
2561 maybe_add_safepoint(iter().next_get_dest());
2562 default:
2563 break;
2564 }
2565 }
2566 b = pop_pair();
2567 a = pop_pair();
2568 c = _gvn.transform( new CmpL3Node( a, b ));
2569 push(c);
2570 break;
2571
2572 case Bytecodes::_lneg:
2573 a = pop_pair();
2574 b = _gvn.transform( new SubLNode(longcon(0),a));
2575 push_pair(b);
2576 break;
2577 case Bytecodes::_l2i:
2578 a = pop_pair();
2579 push( _gvn.transform( new ConvL2INode(a)));
2580 break;
2581 case Bytecodes::_i2l:
2582 a = pop();
2583 b = _gvn.transform( new ConvI2LNode(a));
2584 push_pair(b);
2585 break;
2586 case Bytecodes::_i2b:
2587 // Sign extend
2588 a = pop();
2589 a = _gvn.transform( new LShiftINode(a,_gvn.intcon(24)) );
2590 a = _gvn.transform( new RShiftINode(a,_gvn.intcon(24)) );
2591 push( a );
2592 break;
2593 case Bytecodes::_i2s:
2594 a = pop();
2595 a = _gvn.transform( new LShiftINode(a,_gvn.intcon(16)) );
2596 a = _gvn.transform( new RShiftINode(a,_gvn.intcon(16)) );
2597 push( a );
2598 break;
2599 case Bytecodes::_i2c:
2600 a = pop();
2601 push( _gvn.transform( new AndINode(a,_gvn.intcon(0xFFFF)) ) );
2602 break;
2603
2604 case Bytecodes::_i2f:
2605 a = pop();
2606 b = _gvn.transform( new ConvI2FNode(a) ) ;
2607 c = precision_rounding(b);
2608 push (b);
2609 break;
2610
2611 case Bytecodes::_i2d:
2612 a = pop();
2613 b = _gvn.transform( new ConvI2DNode(a));
2614 push_pair(b);
2615 break;
2616
2617 case Bytecodes::_iinc: // Increment local
2618 i = iter().get_index(); // Get local index
2619 set_local( i, _gvn.transform( new AddINode( _gvn.intcon(iter().get_iinc_con()), local(i) ) ) );
2620 break;
2621
2622 // Exit points of synchronized methods must have an unlock node
2623 case Bytecodes::_return:
2624 return_current(NULL);
2625 break;
2626
2627 case Bytecodes::_ireturn:
2628 case Bytecodes::_areturn:
2629 case Bytecodes::_freturn:
2630 return_current(pop());
2631 break;
2632 case Bytecodes::_lreturn:
2633 return_current(pop_pair());
2634 break;
2635 case Bytecodes::_dreturn:
2636 return_current(pop_pair());
2637 break;
2638
2639 case Bytecodes::_athrow:
2640 // null exception oop throws NULL pointer exception
2641 null_check(peek());
2642 if (stopped()) return;
2643 // Hook the thrown exception directly to subsequent handlers.
2644 if (BailoutToInterpreterForThrows) {
2645 // Keep method interpreted from now on.
2646 uncommon_trap(Deoptimization::Reason_unhandled,
2647 Deoptimization::Action_make_not_compilable);
2648 return;
2649 }
2650 if (env()->jvmti_can_post_on_exceptions()) {
2651 // check if we must post exception events, take uncommon trap if so (with must_throw = false)
2652 uncommon_trap_if_should_post_on_exceptions(Deoptimization::Reason_unhandled, false);
2653 }
2654 // Here if either can_post_on_exceptions or should_post_on_exceptions is false
2655 add_exception_state(make_exception_state(peek()));
2656 break;
2657
2658 case Bytecodes::_goto: // fall through
2659 case Bytecodes::_goto_w: {
2660 int target_bci = (bc() == Bytecodes::_goto) ? iter().get_dest() : iter().get_far_dest();
2661
2662 // If this is a backwards branch in the bytecodes, add Safepoint
2663 maybe_add_safepoint(target_bci);
2664
2665 // Merge the current control into the target basic block
2666 merge(target_bci);
2667
2668 // See if we can get some profile data and hand it off to the next block
2669 Block *target_block = block()->successor_for_bci(target_bci);
2670 if (target_block->pred_count() != 1) break;
2671 ciMethodData* methodData = method()->method_data();
2672 if (!methodData->is_mature()) break;
2673 ciProfileData* data = methodData->bci_to_data(bci());
2674 assert(data != NULL && data->is_JumpData(), "need JumpData for taken branch");
2675 int taken = ((ciJumpData*)data)->taken();
2676 taken = method()->scale_count(taken);
2677 target_block->set_count(taken);
2678 break;
2679 }
2680
2681 case Bytecodes::_ifnull: btest = BoolTest::eq; goto handle_if_null;
2682 case Bytecodes::_ifnonnull: btest = BoolTest::ne; goto handle_if_null;
2683 handle_if_null:
2684 // If this is a backwards branch in the bytecodes, add Safepoint
2685 maybe_add_safepoint(iter().get_dest());
2686 a = null();
2687 b = pop();
2688 if (!_gvn.type(b)->speculative_maybe_null() &&
2689 !too_many_traps(Deoptimization::Reason_speculate_null_check)) {
2690 inc_sp(1);
2691 Node* null_ctl = top();
2692 b = null_check_oop(b, &null_ctl, true, true, true);
2693 assert(null_ctl->is_top(), "no null control here");
2694 dec_sp(1);
2695 } else if (_gvn.type(b)->speculative_always_null() &&
2696 !too_many_traps(Deoptimization::Reason_speculate_null_assert)) {
2697 inc_sp(1);
2698 b = null_assert(b);
2699 dec_sp(1);
2700 }
2701 c = _gvn.transform( new CmpPNode(b, a) );
2702 do_ifnull(btest, c);
2703 break;
2704
2705 case Bytecodes::_if_acmpeq: btest = BoolTest::eq; goto handle_if_acmp;
2706 case Bytecodes::_if_acmpne: btest = BoolTest::ne; goto handle_if_acmp;
2707 handle_if_acmp:
2708 // If this is a backwards branch in the bytecodes, add Safepoint
2709 maybe_add_safepoint(iter().get_dest());
2710 a = pop();
2711 b = pop();
2712 c = _gvn.transform( new CmpPNode(b, a) );
2713 c = optimize_cmp_with_klass(c);
2714 do_if(btest, c);
2715 break;
2716
2717 case Bytecodes::_ifeq: btest = BoolTest::eq; goto handle_ifxx;
2718 case Bytecodes::_ifne: btest = BoolTest::ne; goto handle_ifxx;
2719 case Bytecodes::_iflt: btest = BoolTest::lt; goto handle_ifxx;
2720 case Bytecodes::_ifle: btest = BoolTest::le; goto handle_ifxx;
2721 case Bytecodes::_ifgt: btest = BoolTest::gt; goto handle_ifxx;
2722 case Bytecodes::_ifge: btest = BoolTest::ge; goto handle_ifxx;
2723 handle_ifxx:
2724 // If this is a backwards branch in the bytecodes, add Safepoint
2725 maybe_add_safepoint(iter().get_dest());
2726 a = _gvn.intcon(0);
2727 b = pop();
2728 c = _gvn.transform( new CmpINode(b, a) );
2729 do_if(btest, c);
2730 break;
2731
2732 case Bytecodes::_if_icmpeq: btest = BoolTest::eq; goto handle_if_icmp;
2733 case Bytecodes::_if_icmpne: btest = BoolTest::ne; goto handle_if_icmp;
2734 case Bytecodes::_if_icmplt: btest = BoolTest::lt; goto handle_if_icmp;
2735 case Bytecodes::_if_icmple: btest = BoolTest::le; goto handle_if_icmp;
2736 case Bytecodes::_if_icmpgt: btest = BoolTest::gt; goto handle_if_icmp;
2737 case Bytecodes::_if_icmpge: btest = BoolTest::ge; goto handle_if_icmp;
2738 handle_if_icmp:
2739 // If this is a backwards branch in the bytecodes, add Safepoint
2740 maybe_add_safepoint(iter().get_dest());
2741 a = pop();
2742 b = pop();
2743 c = _gvn.transform( new CmpINode( b, a ) );
2744 do_if(btest, c);
2745 break;
2746
2747 case Bytecodes::_tableswitch:
2748 do_tableswitch();
2749 break;
2750
2751 case Bytecodes::_lookupswitch:
2752 do_lookupswitch();
2753 break;
2754
2755 case Bytecodes::_invokestatic:
2756 case Bytecodes::_invokedynamic:
2757 case Bytecodes::_invokespecial:
2758 case Bytecodes::_invokevirtual:
2759 case Bytecodes::_invokeinterface:
2760 do_call();
2761 break;
2762 case Bytecodes::_checkcast:
2763 do_checkcast();
2764 break;
2765 case Bytecodes::_instanceof:
2766 do_instanceof();
2767 break;
2768 case Bytecodes::_anewarray:
2769 do_anewarray();
2770 break;
2771 case Bytecodes::_newarray:
2772 do_newarray((BasicType)iter().get_index());
2773 break;
2774 case Bytecodes::_multianewarray:
2775 do_multianewarray();
2776 break;
2777 case Bytecodes::_new:
2778 do_new();
2779 break;
2780
2781 case Bytecodes::_jsr:
2782 case Bytecodes::_jsr_w:
2783 do_jsr();
2784 break;
2785
2786 case Bytecodes::_ret:
2787 do_ret();
2788 break;
2789
2790
2791 case Bytecodes::_monitorenter:
2792 do_monitor_enter();
2793 break;
2794
2795 case Bytecodes::_monitorexit:
2796 do_monitor_exit();
2797 break;
2798
2799 case Bytecodes::_breakpoint:
2800 // Breakpoint set concurrently to compile
2801 // %%% use an uncommon trap?
2802 C->record_failure("breakpoint in method");
2803 return;
2804
2805 default:
2806 #ifndef PRODUCT
2807 map()->dump(99);
2808 #endif
2809 tty->print("\nUnhandled bytecode %s\n", Bytecodes::name(bc()) );
2810 ShouldNotReachHere();
2811 }
2812
2813 #ifndef PRODUCT
2814 if (C->should_print(1)) {
2815 IdealGraphPrinter* printer = C->printer();
2816 char buffer[256];
2817 jio_snprintf(buffer, sizeof(buffer), "Bytecode %d: %s", bci(), Bytecodes::name(bc()));
2818 bool old = printer->traverse_outs();
2819 printer->set_traverse_outs(true);
2820 printer->print_method(buffer, 4);
2821 printer->set_traverse_outs(old);
2822 }
2823 #endif
2824 }
2825