1 /*
2 * Copyright (c) 1998, 2018, 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 "opto/valuetypenode.hpp"
46 #include "runtime/deoptimization.hpp"
47 #include "runtime/sharedRuntime.hpp"
48
49 #ifndef PRODUCT
50 extern int explicit_null_checks_inserted,
51 explicit_null_checks_elided;
52 #endif
53
54 //---------------------------------array_load----------------------------------
55 void Parse::array_load(BasicType bt) {
56 const Type* elemtype = Type::TOP;
57 Node* adr = array_addressing(bt, 0, &elemtype);
58 if (stopped()) return; // guaranteed null or range check
59
60 Node* idx = pop();
61 Node* ary = pop();
62
63 // Handle value type arrays
64 const TypeOopPtr* elemptr = elemtype->make_oopptr();
65 if (elemtype->isa_valuetype() != NULL) {
66 // Load from flattened value type array
67 ciValueKlass* vk = elemtype->is_valuetype()->value_klass();
68 ValueTypeNode* vt = ValueTypeNode::make_from_flattened(this, vk, ary, adr);
69 push(vt);
70 return;
71 } else if (elemptr != NULL && elemptr->is_valuetypeptr()) {
72 // Load from non-flattened value type array (elements can never be null)
73 bt = T_VALUETYPE;
74 assert(elemptr->meet(TypePtr::NULL_PTR) != elemptr, "value type array elements should never be null");
75 } else if (ValueArrayFlatten && elemptr != NULL && elemptr->can_be_value_type()) {
76 // Cannot statically determine if array is flattened, emit runtime check
77 gen_flattened_array_guard(ary, 2);
78 }
79
80 if (elemtype == TypeInt::BOOL) {
81 bt = T_BOOLEAN;
82 } else if (bt == T_OBJECT) {
83 elemtype = _gvn.type(ary)->is_aryptr()->elem()->make_oopptr();
84 }
85
86 const TypeAryPtr* adr_type = TypeAryPtr::get_array_body_type(bt);
87 Node* ld = access_load_at(ary, adr, adr_type, elemtype, bt,
88 IN_HEAP | IN_HEAP_ARRAY | C2_CONTROL_DEPENDENT_LOAD);
89 if (bt == T_VALUETYPE) {
90 // Loading a non-flattened (but flattenable) value type from an array
91 assert(!gvn().type(ld)->is_ptr()->maybe_null(), "value type array elements should never be null");
92 ld = ValueTypeNode::make_from_oop(this, ld, elemptr->value_klass());
93 }
94
95 push_node(bt, ld);
96 }
97
98
99 //--------------------------------array_store----------------------------------
100 void Parse::array_store(BasicType bt) {
101 const Type* elemtype = Type::TOP;
102 Node* adr = array_addressing(bt, type2size[bt], &elemtype);
103 if (stopped()) return; // guaranteed null or range check
104 Node* cast_val = NULL;
105 if (bt == T_OBJECT) {
106 cast_val = array_store_check();
107 if (stopped()) return;
108 }
109 Node* val = pop_node(bt); // Value to store
110 Node* idx = pop(); // Index in the array
111 Node* ary = pop(); // The array itself
112
113 // Handle value type arrays
114 if (bt == T_OBJECT) {
115 const TypeValueType* vt = _gvn.type(cast_val)->isa_valuetype();
116 if (vt != NULL && vt->value_klass()->flatten_array()) {
117 // Store to flattened value type array
118 if (elemtype->isa_valuetype() == NULL) {
119 // Object/interface array must be flattened, cast it
120 assert(elemtype->make_oopptr()->can_be_value_type(), "must be object or interface array");
121 ciArrayKlass* array_klass = ciArrayKlass::make(vt->value_klass());
122 const TypeAryPtr* arytype = TypeOopPtr::make_from_klass(array_klass)->isa_aryptr();
123 ary = _gvn.transform(new CheckCastPPNode(control(), ary, arytype));
124 adr = array_element_address(ary, idx, T_OBJECT, arytype->size(), control());
125 }
126 cast_val->as_ValueType()->store_flattened(this, ary, adr);
127 return;
128 } else if (vt != NULL) {
129 // Store to non-flattened value type array
130 bt = T_VALUETYPE;
131 val = cast_val;
132 }
133 }
134
135 if (elemtype == TypeInt::BOOL) {
136 bt = T_BOOLEAN;
137 } else if (bt == T_OBJECT) {
138 elemtype = _gvn.type(ary)->is_aryptr()->elem()->make_oopptr();
139 }
140
141 const TypeAryPtr* adr_type = TypeAryPtr::get_array_body_type(bt);
142 access_store_at(control(), ary, adr, adr_type, val, elemtype, bt, MO_UNORDERED | IN_HEAP | IN_HEAP_ARRAY);
143 }
144
145
146 //------------------------------array_addressing-------------------------------
147 // Pull array and index from the stack. Compute pointer-to-element.
148 Node* Parse::array_addressing(BasicType type, int vals, const Type* *result2) {
149 Node *idx = peek(0+vals); // Get from stack without popping
150 Node *ary = peek(1+vals); // in case of exception
151
152 // Null check the array base, with correct stack contents
153 ary = null_check(ary, T_ARRAY);
154 // Compile-time detect of null-exception?
155 if (stopped()) return top();
156
157 const TypeAryPtr* arytype = _gvn.type(ary)->is_aryptr();
158 const TypeInt* sizetype = arytype->size();
159 const Type* elemtype = arytype->elem();
160
161 if (UseUniqueSubclasses && result2 != NULL) {
162 const Type* el = elemtype->make_ptr();
163 if (el && el->isa_instptr()) {
164 const TypeInstPtr* toop = el->is_instptr();
165 if (toop->klass()->as_instance_klass()->unique_concrete_subklass()) {
166 // If we load from "AbstractClass[]" we must see "ConcreteSubClass".
167 const Type* subklass = Type::get_const_type(toop->klass());
168 elemtype = subklass->join_speculative(el);
169 }
170 }
171 }
172
173 // Check for big class initializers with all constant offsets
174 // feeding into a known-size array.
175 const TypeInt* idxtype = _gvn.type(idx)->is_int();
176 // See if the highest idx value is less than the lowest array bound,
177 // and if the idx value cannot be negative:
178 bool need_range_check = true;
179 if (idxtype->_hi < sizetype->_lo && idxtype->_lo >= 0) {
180 need_range_check = false;
181 if (C->log() != NULL) C->log()->elem("observe that='!need_range_check'");
182 }
183
184 ciKlass * arytype_klass = arytype->klass();
185 if ((arytype_klass != NULL) && (!arytype_klass->is_loaded())) {
186 // Only fails for some -Xcomp runs
187 // The class is unloaded. We have to run this bytecode in the interpreter.
188 uncommon_trap(Deoptimization::Reason_unloaded,
189 Deoptimization::Action_reinterpret,
190 arytype->klass(), "!loaded array");
191 return top();
192 }
193
194 // Do the range check
195 if (GenerateRangeChecks && need_range_check) {
196 Node* tst;
197 if (sizetype->_hi <= 0) {
198 // The greatest array bound is negative, so we can conclude that we're
199 // compiling unreachable code, but the unsigned compare trick used below
200 // only works with non-negative lengths. Instead, hack "tst" to be zero so
201 // the uncommon_trap path will always be taken.
202 tst = _gvn.intcon(0);
203 } else {
204 // Range is constant in array-oop, so we can use the original state of mem
205 Node* len = load_array_length(ary);
206
207 // Test length vs index (standard trick using unsigned compare)
208 Node* chk = _gvn.transform( new CmpUNode(idx, len) );
209 BoolTest::mask btest = BoolTest::lt;
210 tst = _gvn.transform( new BoolNode(chk, btest) );
211 }
212 RangeCheckNode* rc = new RangeCheckNode(control(), tst, PROB_MAX, COUNT_UNKNOWN);
213 _gvn.set_type(rc, rc->Value(&_gvn));
214 if (!tst->is_Con()) {
215 record_for_igvn(rc);
216 }
217 set_control(_gvn.transform(new IfTrueNode(rc)));
218 // Branch to failure if out of bounds
219 {
220 PreserveJVMState pjvms(this);
221 set_control(_gvn.transform(new IfFalseNode(rc)));
222 if (C->allow_range_check_smearing()) {
223 // Do not use builtin_throw, since range checks are sometimes
224 // made more stringent by an optimistic transformation.
225 // This creates "tentative" range checks at this point,
226 // which are not guaranteed to throw exceptions.
227 // See IfNode::Ideal, is_range_check, adjust_check.
228 uncommon_trap(Deoptimization::Reason_range_check,
229 Deoptimization::Action_make_not_entrant,
230 NULL, "range_check");
231 } else {
232 // If we have already recompiled with the range-check-widening
233 // heroic optimization turned off, then we must really be throwing
234 // range check exceptions.
235 builtin_throw(Deoptimization::Reason_range_check, idx);
236 }
237 }
238 }
239 // Check for always knowing you are throwing a range-check exception
240 if (stopped()) return top();
241
242 // Make array address computation control dependent to prevent it
243 // from floating above the range check during loop optimizations.
244 Node* ptr = array_element_address(ary, idx, type, sizetype, control());
245
246 if (result2 != NULL) *result2 = elemtype;
247
248 assert(ptr != top(), "top should go hand-in-hand with stopped");
249
250 return ptr;
251 }
252
253
254 // returns IfNode
255 IfNode* Parse::jump_if_fork_int(Node* a, Node* b, BoolTest::mask mask, float prob, float cnt) {
256 Node *cmp = _gvn.transform(new CmpINode(a, b)); // two cases: shiftcount > 32 and shiftcount <= 32
257 Node *tst = _gvn.transform(new BoolNode(cmp, mask));
258 IfNode *iff = create_and_map_if(control(), tst, prob, cnt);
259 return iff;
260 }
261
262 // return Region node
263 Node* Parse::jump_if_join(Node* iffalse, Node* iftrue) {
264 Node *region = new RegionNode(3); // 2 results
265 record_for_igvn(region);
266 region->init_req(1, iffalse);
267 region->init_req(2, iftrue );
268 _gvn.set_type(region, Type::CONTROL);
269 region = _gvn.transform(region);
270 set_control (region);
271 return region;
272 }
273
274 // sentinel value for the target bci to mark never taken branches
275 // (according to profiling)
276 static const int never_reached = INT_MAX;
277
278 //------------------------------helper for tableswitch-------------------------
279 void Parse::jump_if_true_fork(IfNode *iff, int dest_bci_if_true, int prof_table_index, bool unc) {
280 // True branch, use existing map info
281 { PreserveJVMState pjvms(this);
282 Node *iftrue = _gvn.transform( new IfTrueNode (iff) );
283 set_control( iftrue );
284 if (unc) {
285 repush_if_args();
286 uncommon_trap(Deoptimization::Reason_unstable_if,
287 Deoptimization::Action_reinterpret,
288 NULL,
289 "taken always");
290 } else {
291 assert(dest_bci_if_true != never_reached, "inconsistent dest");
292 profile_switch_case(prof_table_index);
293 merge_new_path(dest_bci_if_true);
294 }
295 }
296
297 // False branch
298 Node *iffalse = _gvn.transform( new IfFalseNode(iff) );
299 set_control( iffalse );
300 }
301
302 void Parse::jump_if_false_fork(IfNode *iff, int dest_bci_if_true, int prof_table_index, bool unc) {
303 // True branch, use existing map info
304 { PreserveJVMState pjvms(this);
305 Node *iffalse = _gvn.transform( new IfFalseNode (iff) );
306 set_control( iffalse );
307 if (unc) {
308 repush_if_args();
309 uncommon_trap(Deoptimization::Reason_unstable_if,
310 Deoptimization::Action_reinterpret,
311 NULL,
312 "taken never");
313 } else {
314 assert(dest_bci_if_true != never_reached, "inconsistent dest");
315 profile_switch_case(prof_table_index);
316 merge_new_path(dest_bci_if_true);
317 }
318 }
319
320 // False branch
321 Node *iftrue = _gvn.transform( new IfTrueNode(iff) );
322 set_control( iftrue );
323 }
324
325 void Parse::jump_if_always_fork(int dest_bci, int prof_table_index, bool unc) {
326 // False branch, use existing map and control()
327 if (unc) {
328 repush_if_args();
329 uncommon_trap(Deoptimization::Reason_unstable_if,
330 Deoptimization::Action_reinterpret,
331 NULL,
332 "taken never");
333 } else {
334 assert(dest_bci != never_reached, "inconsistent dest");
335 profile_switch_case(prof_table_index);
336 merge_new_path(dest_bci);
337 }
338 }
339
340
341 extern "C" {
342 static int jint_cmp(const void *i, const void *j) {
343 int a = *(jint *)i;
344 int b = *(jint *)j;
345 return a > b ? 1 : a < b ? -1 : 0;
346 }
347 }
348
349
350 // Default value for methodData switch indexing. Must be a negative value to avoid
351 // conflict with any legal switch index.
352 #define NullTableIndex -1
353
354 class SwitchRange : public StackObj {
355 // a range of integers coupled with a bci destination
356 jint _lo; // inclusive lower limit
357 jint _hi; // inclusive upper limit
358 int _dest;
359 int _table_index; // index into method data table
360 float _cnt; // how many times this range was hit according to profiling
361
362 public:
363 jint lo() const { return _lo; }
364 jint hi() const { return _hi; }
365 int dest() const { return _dest; }
366 int table_index() const { return _table_index; }
367 bool is_singleton() const { return _lo == _hi; }
368 float cnt() const { return _cnt; }
369
370 void setRange(jint lo, jint hi, int dest, int table_index, float cnt) {
371 assert(lo <= hi, "must be a non-empty range");
372 _lo = lo, _hi = hi; _dest = dest; _table_index = table_index; _cnt = cnt;
373 assert(_cnt >= 0, "");
374 }
375 bool adjoinRange(jint lo, jint hi, int dest, int table_index, float cnt, bool trim_ranges) {
376 assert(lo <= hi, "must be a non-empty range");
377 if (lo == _hi+1 && table_index == _table_index) {
378 // see merge_ranges() comment below
379 if (trim_ranges) {
380 if (cnt == 0) {
381 if (_cnt != 0) {
382 return false;
383 }
384 if (dest != _dest) {
385 _dest = never_reached;
386 }
387 } else {
388 if (_cnt == 0) {
389 return false;
390 }
391 if (dest != _dest) {
392 return false;
393 }
394 }
395 } else {
396 if (dest != _dest) {
397 return false;
398 }
399 }
400 _hi = hi;
401 _cnt += cnt;
402 return true;
403 }
404 return false;
405 }
406
407 void set (jint value, int dest, int table_index, float cnt) {
408 setRange(value, value, dest, table_index, cnt);
409 }
410 bool adjoin(jint value, int dest, int table_index, float cnt, bool trim_ranges) {
411 return adjoinRange(value, value, dest, table_index, cnt, trim_ranges);
412 }
413 bool adjoin(SwitchRange& other) {
414 return adjoinRange(other._lo, other._hi, other._dest, other._table_index, other._cnt, false);
415 }
416
417 void print() {
418 if (is_singleton())
419 tty->print(" {%d}=>%d (cnt=%f)", lo(), dest(), cnt());
420 else if (lo() == min_jint)
421 tty->print(" {..%d}=>%d (cnt=%f)", hi(), dest(), cnt());
422 else if (hi() == max_jint)
423 tty->print(" {%d..}=>%d (cnt=%f)", lo(), dest(), cnt());
424 else
425 tty->print(" {%d..%d}=>%d (cnt=%f)", lo(), hi(), dest(), cnt());
426 }
427 };
428
429 // We try to minimize the number of ranges and the size of the taken
430 // ones using profiling data. When ranges are created,
431 // SwitchRange::adjoinRange() only allows 2 adjoining ranges to merge
432 // if both were never hit or both were hit to build longer unreached
433 // ranges. Here, we now merge adjoining ranges with the same
434 // destination and finally set destination of unreached ranges to the
435 // special value never_reached because it can help minimize the number
436 // of tests that are necessary.
437 //
438 // For instance:
439 // [0, 1] to target1 sometimes taken
440 // [1, 2] to target1 never taken
441 // [2, 3] to target2 never taken
442 // would lead to:
443 // [0, 1] to target1 sometimes taken
444 // [1, 3] never taken
445 //
446 // (first 2 ranges to target1 are not merged)
447 static void merge_ranges(SwitchRange* ranges, int& rp) {
448 if (rp == 0) {
449 return;
450 }
451 int shift = 0;
452 for (int j = 0; j < rp; j++) {
453 SwitchRange& r1 = ranges[j-shift];
454 SwitchRange& r2 = ranges[j+1];
455 if (r1.adjoin(r2)) {
456 shift++;
457 } else if (shift > 0) {
458 ranges[j+1-shift] = r2;
459 }
460 }
461 rp -= shift;
462 for (int j = 0; j <= rp; j++) {
463 SwitchRange& r = ranges[j];
464 if (r.cnt() == 0 && r.dest() != never_reached) {
465 r.setRange(r.lo(), r.hi(), never_reached, r.table_index(), r.cnt());
466 }
467 }
468 }
469
470 //-------------------------------do_tableswitch--------------------------------
471 void Parse::do_tableswitch() {
472 Node* lookup = pop();
473 // Get information about tableswitch
474 int default_dest = iter().get_dest_table(0);
475 int lo_index = iter().get_int_table(1);
476 int hi_index = iter().get_int_table(2);
477 int len = hi_index - lo_index + 1;
478
479 if (len < 1) {
480 // If this is a backward branch, add safepoint
481 maybe_add_safepoint(default_dest);
482 merge(default_dest);
483 return;
484 }
485
486 ciMethodData* methodData = method()->method_data();
487 ciMultiBranchData* profile = NULL;
488 if (methodData->is_mature() && UseSwitchProfiling) {
489 ciProfileData* data = methodData->bci_to_data(bci());
490 if (data != NULL && data->is_MultiBranchData()) {
491 profile = (ciMultiBranchData*)data;
492 }
493 }
494 bool trim_ranges = !method_data_update() && !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if);
495
496 // generate decision tree, using trichotomy when possible
497 int rnum = len+2;
498 bool makes_backward_branch = false;
499 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum);
500 int rp = -1;
501 if (lo_index != min_jint) {
502 uint cnt = 1;
503 if (profile != NULL) {
504 cnt = profile->default_count() / (hi_index != max_jint ? 2 : 1);
505 }
506 ranges[++rp].setRange(min_jint, lo_index-1, default_dest, NullTableIndex, cnt);
507 }
508 for (int j = 0; j < len; j++) {
509 jint match_int = lo_index+j;
510 int dest = iter().get_dest_table(j+3);
511 makes_backward_branch |= (dest <= bci());
512 int table_index = method_data_update() ? j : NullTableIndex;
513 uint cnt = 1;
514 if (profile != NULL) {
515 cnt = profile->count_at(j);
516 }
517 if (rp < 0 || !ranges[rp].adjoin(match_int, dest, table_index, cnt, trim_ranges)) {
518 ranges[++rp].set(match_int, dest, table_index, cnt);
519 }
520 }
521 jint highest = lo_index+(len-1);
522 assert(ranges[rp].hi() == highest, "");
523 if (highest != max_jint) {
524 uint cnt = 1;
525 if (profile != NULL) {
526 cnt = profile->default_count() / (lo_index != min_jint ? 2 : 1);
527 }
528 if (!ranges[rp].adjoinRange(highest+1, max_jint, default_dest, NullTableIndex, cnt, trim_ranges)) {
529 ranges[++rp].setRange(highest+1, max_jint, default_dest, NullTableIndex, cnt);
530 }
531 }
532 assert(rp < len+2, "not too many ranges");
533
534 if (trim_ranges) {
535 merge_ranges(ranges, rp);
536 }
537
538 // Safepoint in case if backward branch observed
539 if( makes_backward_branch && UseLoopSafepoints )
540 add_safepoint();
541
542 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]);
543 }
544
545
546 //------------------------------do_lookupswitch--------------------------------
547 void Parse::do_lookupswitch() {
548 Node *lookup = pop(); // lookup value
549 // Get information about lookupswitch
550 int default_dest = iter().get_dest_table(0);
551 int len = iter().get_int_table(1);
552
553 if (len < 1) { // If this is a backward branch, add safepoint
554 maybe_add_safepoint(default_dest);
555 merge(default_dest);
556 return;
557 }
558
559 ciMethodData* methodData = method()->method_data();
560 ciMultiBranchData* profile = NULL;
561 if (methodData->is_mature() && UseSwitchProfiling) {
562 ciProfileData* data = methodData->bci_to_data(bci());
563 if (data != NULL && data->is_MultiBranchData()) {
564 profile = (ciMultiBranchData*)data;
565 }
566 }
567 bool trim_ranges = !method_data_update() && !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if);
568
569 // generate decision tree, using trichotomy when possible
570 jint* table = NEW_RESOURCE_ARRAY(jint, len*3);
571 {
572 for (int j = 0; j < len; j++) {
573 table[3*j+0] = iter().get_int_table(2+2*j);
574 table[3*j+1] = iter().get_dest_table(2+2*j+1);
575 table[3*j+2] = profile == NULL ? 1 : profile->count_at(j);
576 }
577 qsort(table, len, 3*sizeof(table[0]), jint_cmp);
578 }
579
580 float defaults = 0;
581 jint prev = min_jint;
582 for (int j = 0; j < len; j++) {
583 jint match_int = table[3*j+0];
584 if (match_int != prev) {
585 defaults += (float)match_int - prev;
586 }
587 prev = match_int+1;
588 }
589 if (prev-1 != max_jint) {
590 defaults += (float)max_jint - prev + 1;
591 }
592 float default_cnt = 1;
593 if (profile != NULL) {
594 default_cnt = profile->default_count()/defaults;
595 }
596
597 int rnum = len*2+1;
598 bool makes_backward_branch = false;
599 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum);
600 int rp = -1;
601 for (int j = 0; j < len; j++) {
602 jint match_int = table[3*j+0];
603 int dest = table[3*j+1];
604 int cnt = table[3*j+2];
605 int next_lo = rp < 0 ? min_jint : ranges[rp].hi()+1;
606 int table_index = method_data_update() ? j : NullTableIndex;
607 makes_backward_branch |= (dest <= bci());
608 float c = default_cnt * ((float)match_int - next_lo);
609 if (match_int != next_lo && (rp < 0 || !ranges[rp].adjoinRange(next_lo, match_int-1, default_dest, NullTableIndex, c, trim_ranges))) {
610 assert(default_dest != never_reached, "sentinel value for dead destinations");
611 ranges[++rp].setRange(next_lo, match_int-1, default_dest, NullTableIndex, c);
612 }
613 if (rp < 0 || !ranges[rp].adjoin(match_int, dest, table_index, cnt, trim_ranges)) {
614 assert(dest != never_reached, "sentinel value for dead destinations");
615 ranges[++rp].set(match_int, dest, table_index, cnt);
616 }
617 }
618 jint highest = table[3*(len-1)];
619 assert(ranges[rp].hi() == highest, "");
620 if (highest != max_jint &&
621 !ranges[rp].adjoinRange(highest+1, max_jint, default_dest, NullTableIndex, default_cnt * ((float)max_jint - highest), trim_ranges)) {
622 ranges[++rp].setRange(highest+1, max_jint, default_dest, NullTableIndex, default_cnt * ((float)max_jint - highest));
623 }
624 assert(rp < rnum, "not too many ranges");
625
626 if (trim_ranges) {
627 merge_ranges(ranges, rp);
628 }
629
630 // Safepoint in case backward branch observed
631 if (makes_backward_branch && UseLoopSafepoints)
632 add_safepoint();
633
634 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]);
635 }
636
637 static float if_prob(float taken_cnt, float total_cnt) {
638 assert(taken_cnt <= total_cnt, "");
639 if (total_cnt == 0) {
640 return PROB_FAIR;
641 }
642 float p = taken_cnt / total_cnt;
643 return MIN2(MAX2(p, PROB_MIN), PROB_MAX);
644 }
645
646 static float if_cnt(float cnt) {
647 if (cnt == 0) {
648 return COUNT_UNKNOWN;
649 }
650 return cnt;
651 }
652
653 static float sum_of_cnts(SwitchRange *lo, SwitchRange *hi) {
654 float total_cnt = 0;
655 for (SwitchRange* sr = lo; sr <= hi; sr++) {
656 total_cnt += sr->cnt();
657 }
658 return total_cnt;
659 }
660
661 class SwitchRanges : public ResourceObj {
662 public:
663 SwitchRange* _lo;
664 SwitchRange* _hi;
665 SwitchRange* _mid;
666 float _cost;
667
668 enum {
669 Start,
670 LeftDone,
671 RightDone,
672 Done
673 } _state;
674
675 SwitchRanges(SwitchRange *lo, SwitchRange *hi)
676 : _lo(lo), _hi(hi), _mid(NULL),
677 _cost(0), _state(Start) {
678 }
679
680 SwitchRanges()
681 : _lo(NULL), _hi(NULL), _mid(NULL),
682 _cost(0), _state(Start) {}
683 };
684
685 // Estimate cost of performing a binary search on lo..hi
686 static float compute_tree_cost(SwitchRange *lo, SwitchRange *hi, float total_cnt) {
687 GrowableArray<SwitchRanges> tree;
688 SwitchRanges root(lo, hi);
689 tree.push(root);
690
691 float cost = 0;
692 do {
693 SwitchRanges& r = *tree.adr_at(tree.length()-1);
694 if (r._hi != r._lo) {
695 if (r._mid == NULL) {
696 float r_cnt = sum_of_cnts(r._lo, r._hi);
697
698 if (r_cnt == 0) {
699 tree.pop();
700 cost = 0;
701 continue;
702 }
703
704 SwitchRange* mid = NULL;
705 mid = r._lo;
706 for (float cnt = 0; ; ) {
707 assert(mid <= r._hi, "out of bounds");
708 cnt += mid->cnt();
709 if (cnt > r_cnt / 2) {
710 break;
711 }
712 mid++;
713 }
714 assert(mid <= r._hi, "out of bounds");
715 r._mid = mid;
716 r._cost = r_cnt / total_cnt;
717 }
718 r._cost += cost;
719 if (r._state < SwitchRanges::LeftDone && r._mid > r._lo) {
720 cost = 0;
721 r._state = SwitchRanges::LeftDone;
722 tree.push(SwitchRanges(r._lo, r._mid-1));
723 } else if (r._state < SwitchRanges::RightDone) {
724 cost = 0;
725 r._state = SwitchRanges::RightDone;
726 tree.push(SwitchRanges(r._mid == r._lo ? r._mid+1 : r._mid, r._hi));
727 } else {
728 tree.pop();
729 cost = r._cost;
730 }
731 } else {
732 tree.pop();
733 cost = r._cost;
734 }
735 } while (tree.length() > 0);
736
737
738 return cost;
739 }
740
741 // It sometimes pays off to test most common ranges before the binary search
742 void Parse::linear_search_switch_ranges(Node* key_val, SwitchRange*& lo, SwitchRange*& hi) {
743 uint nr = hi - lo + 1;
744 float total_cnt = sum_of_cnts(lo, hi);
745
746 float min = compute_tree_cost(lo, hi, total_cnt);
747 float extra = 1;
748 float sub = 0;
749
750 SwitchRange* array1 = lo;
751 SwitchRange* array2 = NEW_RESOURCE_ARRAY(SwitchRange, nr);
752
753 SwitchRange* ranges = NULL;
754
755 while (nr >= 2) {
756 assert(lo == array1 || lo == array2, "one the 2 already allocated arrays");
757 ranges = (lo == array1) ? array2 : array1;
758
759 // Find highest frequency range
760 SwitchRange* candidate = lo;
761 for (SwitchRange* sr = lo+1; sr <= hi; sr++) {
762 if (sr->cnt() > candidate->cnt()) {
763 candidate = sr;
764 }
765 }
766 SwitchRange most_freq = *candidate;
767 if (most_freq.cnt() == 0) {
768 break;
769 }
770
771 // Copy remaining ranges into another array
772 int shift = 0;
773 for (uint i = 0; i < nr; i++) {
774 SwitchRange* sr = &lo[i];
775 if (sr != candidate) {
776 ranges[i-shift] = *sr;
777 } else {
778 shift++;
779 if (i > 0 && i < nr-1) {
780 SwitchRange prev = lo[i-1];
781 prev.setRange(prev.lo(), sr->hi(), prev.dest(), prev.table_index(), prev.cnt());
782 if (prev.adjoin(lo[i+1])) {
783 shift++;
784 i++;
785 }
786 ranges[i-shift] = prev;
787 }
788 }
789 }
790 nr -= shift;
791
792 // Evaluate cost of testing the most common range and performing a
793 // binary search on the other ranges
794 float cost = extra + compute_tree_cost(&ranges[0], &ranges[nr-1], total_cnt);
795 if (cost >= min) {
796 break;
797 }
798 // swap arrays
799 lo = &ranges[0];
800 hi = &ranges[nr-1];
801
802 // It pays off: emit the test for the most common range
803 assert(most_freq.cnt() > 0, "must be taken");
804 Node* val = _gvn.transform(new SubINode(key_val, _gvn.intcon(most_freq.lo())));
805 Node* cmp = _gvn.transform(new CmpUNode(val, _gvn.intcon(most_freq.hi() - most_freq.lo())));
806 Node* tst = _gvn.transform(new BoolNode(cmp, BoolTest::le));
807 IfNode* iff = create_and_map_if(control(), tst, if_prob(most_freq.cnt(), total_cnt), if_cnt(most_freq.cnt()));
808 jump_if_true_fork(iff, most_freq.dest(), most_freq.table_index(), false);
809
810 sub += most_freq.cnt() / total_cnt;
811 extra += 1 - sub;
812 min = cost;
813 }
814 }
815
816 //----------------------------create_jump_tables-------------------------------
817 bool Parse::create_jump_tables(Node* key_val, SwitchRange* lo, SwitchRange* hi) {
818 // Are jumptables enabled
819 if (!UseJumpTables) return false;
820
821 // Are jumptables supported
822 if (!Matcher::has_match_rule(Op_Jump)) return false;
823
824 // Don't make jump table if profiling
825 if (method_data_update()) return false;
826
827 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if);
828
829 // Decide if a guard is needed to lop off big ranges at either (or
830 // both) end(s) of the input set. We'll call this the default target
831 // even though we can't be sure that it is the true "default".
832
833 bool needs_guard = false;
834 int default_dest;
835 int64_t total_outlier_size = 0;
836 int64_t hi_size = ((int64_t)hi->hi()) - ((int64_t)hi->lo()) + 1;
837 int64_t lo_size = ((int64_t)lo->hi()) - ((int64_t)lo->lo()) + 1;
838
839 if (lo->dest() == hi->dest()) {
840 total_outlier_size = hi_size + lo_size;
841 default_dest = lo->dest();
842 } else if (lo_size > hi_size) {
843 total_outlier_size = lo_size;
844 default_dest = lo->dest();
845 } else {
846 total_outlier_size = hi_size;
847 default_dest = hi->dest();
848 }
849
850 float total = sum_of_cnts(lo, hi);
851 float cost = compute_tree_cost(lo, hi, total);
852
853 // If a guard test will eliminate very sparse end ranges, then
854 // it is worth the cost of an extra jump.
855 float trimmed_cnt = 0;
856 if (total_outlier_size > (MaxJumpTableSparseness * 4)) {
857 needs_guard = true;
858 if (default_dest == lo->dest()) {
859 trimmed_cnt += lo->cnt();
860 lo++;
861 }
862 if (default_dest == hi->dest()) {
863 trimmed_cnt += hi->cnt();
864 hi--;
865 }
866 }
867
868 // Find the total number of cases and ranges
869 int64_t num_cases = ((int64_t)hi->hi()) - ((int64_t)lo->lo()) + 1;
870 int num_range = hi - lo + 1;
871
872 // Don't create table if: too large, too small, or too sparse.
873 if (num_cases > MaxJumpTableSize)
874 return false;
875 if (UseSwitchProfiling) {
876 // MinJumpTableSize is set so with a well balanced binary tree,
877 // when the number of ranges is MinJumpTableSize, it's cheaper to
878 // go through a JumpNode that a tree of IfNodes. Average cost of a
879 // tree of IfNodes with MinJumpTableSize is
880 // log2f(MinJumpTableSize) comparisons. So if the cost computed
881 // from profile data is less than log2f(MinJumpTableSize) then
882 // going with the binary search is cheaper.
883 if (cost < log2f(MinJumpTableSize)) {
884 return false;
885 }
886 } else {
887 if (num_cases < MinJumpTableSize)
888 return false;
889 }
890 if (num_cases > (MaxJumpTableSparseness * num_range))
891 return false;
892
893 // Normalize table lookups to zero
894 int lowval = lo->lo();
895 key_val = _gvn.transform( new SubINode(key_val, _gvn.intcon(lowval)) );
896
897 // Generate a guard to protect against input keyvals that aren't
898 // in the switch domain.
899 if (needs_guard) {
900 Node* size = _gvn.intcon(num_cases);
901 Node* cmp = _gvn.transform(new CmpUNode(key_val, size));
902 Node* tst = _gvn.transform(new BoolNode(cmp, BoolTest::ge));
903 IfNode* iff = create_and_map_if(control(), tst, if_prob(trimmed_cnt, total), if_cnt(trimmed_cnt));
904 jump_if_true_fork(iff, default_dest, NullTableIndex, trim_ranges && trimmed_cnt == 0);
905
906 total -= trimmed_cnt;
907 }
908
909 // Create an ideal node JumpTable that has projections
910 // of all possible ranges for a switch statement
911 // The key_val input must be converted to a pointer offset and scaled.
912 // Compare Parse::array_addressing above.
913
914 // Clean the 32-bit int into a real 64-bit offset.
915 // Otherwise, the jint value 0 might turn into an offset of 0x0800000000.
916 const TypeInt* ikeytype = TypeInt::make(0, num_cases, Type::WidenMin);
917 // Make I2L conversion control dependent to prevent it from
918 // floating above the range check during loop optimizations.
919 key_val = C->conv_I2X_index(&_gvn, key_val, ikeytype, control());
920
921 // Shift the value by wordsize so we have an index into the table, rather
922 // than a switch value
923 Node *shiftWord = _gvn.MakeConX(wordSize);
924 key_val = _gvn.transform( new MulXNode( key_val, shiftWord));
925
926 // Create the JumpNode
927 Arena* arena = C->comp_arena();
928 float* probs = (float*)arena->Amalloc(sizeof(float)*num_cases);
929 int i = 0;
930 if (total == 0) {
931 for (SwitchRange* r = lo; r <= hi; r++) {
932 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) {
933 probs[i] = 1.0F / num_cases;
934 }
935 }
936 } else {
937 for (SwitchRange* r = lo; r <= hi; r++) {
938 float prob = r->cnt()/total;
939 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) {
940 probs[i] = prob / (r->hi() - r->lo() + 1);
941 }
942 }
943 }
944
945 ciMethodData* methodData = method()->method_data();
946 ciMultiBranchData* profile = NULL;
947 if (methodData->is_mature()) {
948 ciProfileData* data = methodData->bci_to_data(bci());
949 if (data != NULL && data->is_MultiBranchData()) {
950 profile = (ciMultiBranchData*)data;
951 }
952 }
953
954 Node* jtn = _gvn.transform(new JumpNode(control(), key_val, num_cases, probs, profile == NULL ? COUNT_UNKNOWN : total));
955
956 // These are the switch destinations hanging off the jumpnode
957 i = 0;
958 for (SwitchRange* r = lo; r <= hi; r++) {
959 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) {
960 Node* input = _gvn.transform(new JumpProjNode(jtn, i, r->dest(), (int)(j - lowval)));
961 {
962 PreserveJVMState pjvms(this);
963 set_control(input);
964 jump_if_always_fork(r->dest(), r->table_index(), trim_ranges && r->cnt() == 0);
965 }
966 }
967 }
968 assert(i == num_cases, "miscount of cases");
969 stop_and_kill_map(); // no more uses for this JVMS
970 return true;
971 }
972
973 //----------------------------jump_switch_ranges-------------------------------
974 void Parse::jump_switch_ranges(Node* key_val, SwitchRange *lo, SwitchRange *hi, int switch_depth) {
975 Block* switch_block = block();
976 bool trim_ranges = !method_data_update() && !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if);
977
978 if (switch_depth == 0) {
979 // Do special processing for the top-level call.
980 assert(lo->lo() == min_jint, "initial range must exhaust Type::INT");
981 assert(hi->hi() == max_jint, "initial range must exhaust Type::INT");
982
983 // Decrement pred-numbers for the unique set of nodes.
984 #ifdef ASSERT
985 if (!trim_ranges) {
986 // Ensure that the block's successors are a (duplicate-free) set.
987 int successors_counted = 0; // block occurrences in [hi..lo]
988 int unique_successors = switch_block->num_successors();
989 for (int i = 0; i < unique_successors; i++) {
990 Block* target = switch_block->successor_at(i);
991
992 // Check that the set of successors is the same in both places.
993 int successors_found = 0;
994 for (SwitchRange* p = lo; p <= hi; p++) {
995 if (p->dest() == target->start()) successors_found++;
996 }
997 assert(successors_found > 0, "successor must be known");
998 successors_counted += successors_found;
999 }
1000 assert(successors_counted == (hi-lo)+1, "no unexpected successors");
1001 }
1002 #endif
1003
1004 // Maybe prune the inputs, based on the type of key_val.
1005 jint min_val = min_jint;
1006 jint max_val = max_jint;
1007 const TypeInt* ti = key_val->bottom_type()->isa_int();
1008 if (ti != NULL) {
1009 min_val = ti->_lo;
1010 max_val = ti->_hi;
1011 assert(min_val <= max_val, "invalid int type");
1012 }
1013 while (lo->hi() < min_val) {
1014 lo++;
1015 }
1016 if (lo->lo() < min_val) {
1017 lo->setRange(min_val, lo->hi(), lo->dest(), lo->table_index(), lo->cnt());
1018 }
1019 while (hi->lo() > max_val) {
1020 hi--;
1021 }
1022 if (hi->hi() > max_val) {
1023 hi->setRange(hi->lo(), max_val, hi->dest(), hi->table_index(), hi->cnt());
1024 }
1025
1026 linear_search_switch_ranges(key_val, lo, hi);
1027 }
1028
1029 #ifndef PRODUCT
1030 if (switch_depth == 0) {
1031 _max_switch_depth = 0;
1032 _est_switch_depth = log2_intptr((hi-lo+1)-1)+1;
1033 }
1034 #endif
1035
1036 assert(lo <= hi, "must be a non-empty set of ranges");
1037 if (lo == hi) {
1038 jump_if_always_fork(lo->dest(), lo->table_index(), trim_ranges && lo->cnt() == 0);
1039 } else {
1040 assert(lo->hi() == (lo+1)->lo()-1, "contiguous ranges");
1041 assert(hi->lo() == (hi-1)->hi()+1, "contiguous ranges");
1042
1043 if (create_jump_tables(key_val, lo, hi)) return;
1044
1045 SwitchRange* mid = NULL;
1046 float total_cnt = sum_of_cnts(lo, hi);
1047
1048 int nr = hi - lo + 1;
1049 if (UseSwitchProfiling) {
1050 // Don't keep the binary search tree balanced: pick up mid point
1051 // that split frequencies in half.
1052 float cnt = 0;
1053 for (SwitchRange* sr = lo; sr <= hi; sr++) {
1054 cnt += sr->cnt();
1055 if (cnt >= total_cnt / 2) {
1056 mid = sr;
1057 break;
1058 }
1059 }
1060 } else {
1061 mid = lo + nr/2;
1062
1063 // if there is an easy choice, pivot at a singleton:
1064 if (nr > 3 && !mid->is_singleton() && (mid-1)->is_singleton()) mid--;
1065
1066 assert(lo < mid && mid <= hi, "good pivot choice");
1067 assert(nr != 2 || mid == hi, "should pick higher of 2");
1068 assert(nr != 3 || mid == hi-1, "should pick middle of 3");
1069 }
1070
1071
1072 Node *test_val = _gvn.intcon(mid == lo ? mid->hi() : mid->lo());
1073
1074 if (mid->is_singleton()) {
1075 IfNode *iff_ne = jump_if_fork_int(key_val, test_val, BoolTest::ne, 1-if_prob(mid->cnt(), total_cnt), if_cnt(mid->cnt()));
1076 jump_if_false_fork(iff_ne, mid->dest(), mid->table_index(), trim_ranges && mid->cnt() == 0);
1077
1078 // Special Case: If there are exactly three ranges, and the high
1079 // and low range each go to the same place, omit the "gt" test,
1080 // since it will not discriminate anything.
1081 bool eq_test_only = (hi == lo+2 && hi->dest() == lo->dest() && mid == hi-1) || mid == lo;
1082
1083 // if there is a higher range, test for it and process it:
1084 if (mid < hi && !eq_test_only) {
1085 // two comparisons of same values--should enable 1 test for 2 branches
1086 // Use BoolTest::le instead of BoolTest::gt
1087 float cnt = sum_of_cnts(lo, mid-1);
1088 IfNode *iff_le = jump_if_fork_int(key_val, test_val, BoolTest::le, if_prob(cnt, total_cnt), if_cnt(cnt));
1089 Node *iftrue = _gvn.transform( new IfTrueNode(iff_le) );
1090 Node *iffalse = _gvn.transform( new IfFalseNode(iff_le) );
1091 { PreserveJVMState pjvms(this);
1092 set_control(iffalse);
1093 jump_switch_ranges(key_val, mid+1, hi, switch_depth+1);
1094 }
1095 set_control(iftrue);
1096 }
1097
1098 } else {
1099 // mid is a range, not a singleton, so treat mid..hi as a unit
1100 float cnt = sum_of_cnts(mid == lo ? mid+1 : mid, hi);
1101 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));
1102
1103 // if there is a higher range, test for it and process it:
1104 if (mid == hi) {
1105 jump_if_true_fork(iff_ge, mid->dest(), mid->table_index(), trim_ranges && cnt == 0);
1106 } else {
1107 Node *iftrue = _gvn.transform( new IfTrueNode(iff_ge) );
1108 Node *iffalse = _gvn.transform( new IfFalseNode(iff_ge) );
1109 { PreserveJVMState pjvms(this);
1110 set_control(iftrue);
1111 jump_switch_ranges(key_val, mid == lo ? mid+1 : mid, hi, switch_depth+1);
1112 }
1113 set_control(iffalse);
1114 }
1115 }
1116
1117 // in any case, process the lower range
1118 if (mid == lo) {
1119 if (mid->is_singleton()) {
1120 jump_switch_ranges(key_val, lo+1, hi, switch_depth+1);
1121 } else {
1122 jump_if_always_fork(lo->dest(), lo->table_index(), trim_ranges && lo->cnt() == 0);
1123 }
1124 } else {
1125 jump_switch_ranges(key_val, lo, mid-1, switch_depth+1);
1126 }
1127 }
1128
1129 // Decrease pred_count for each successor after all is done.
1130 if (switch_depth == 0) {
1131 int unique_successors = switch_block->num_successors();
1132 for (int i = 0; i < unique_successors; i++) {
1133 Block* target = switch_block->successor_at(i);
1134 // Throw away the pre-allocated path for each unique successor.
1135 target->next_path_num();
1136 }
1137 }
1138
1139 #ifndef PRODUCT
1140 _max_switch_depth = MAX2(switch_depth, _max_switch_depth);
1141 if (TraceOptoParse && Verbose && WizardMode && switch_depth == 0) {
1142 SwitchRange* r;
1143 int nsing = 0;
1144 for( r = lo; r <= hi; r++ ) {
1145 if( r->is_singleton() ) nsing++;
1146 }
1147 tty->print(">>> ");
1148 _method->print_short_name();
1149 tty->print_cr(" switch decision tree");
1150 tty->print_cr(" %d ranges (%d singletons), max_depth=%d, est_depth=%d",
1151 (int) (hi-lo+1), nsing, _max_switch_depth, _est_switch_depth);
1152 if (_max_switch_depth > _est_switch_depth) {
1153 tty->print_cr("******** BAD SWITCH DEPTH ********");
1154 }
1155 tty->print(" ");
1156 for( r = lo; r <= hi; r++ ) {
1157 r->print();
1158 }
1159 tty->cr();
1160 }
1161 #endif
1162 }
1163
1164 void Parse::modf() {
1165 Node *f2 = pop();
1166 Node *f1 = pop();
1167 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::modf_Type(),
1168 CAST_FROM_FN_PTR(address, SharedRuntime::frem),
1169 "frem", NULL, //no memory effects
1170 f1, f2);
1171 Node* res = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0));
1172
1173 push(res);
1174 }
1175
1176 void Parse::modd() {
1177 Node *d2 = pop_pair();
1178 Node *d1 = pop_pair();
1179 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::Math_DD_D_Type(),
1180 CAST_FROM_FN_PTR(address, SharedRuntime::drem),
1181 "drem", NULL, //no memory effects
1182 d1, top(), d2, top());
1183 Node* res_d = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0));
1184
1185 #ifdef ASSERT
1186 Node* res_top = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 1));
1187 assert(res_top == top(), "second value must be top");
1188 #endif
1189
1190 push_pair(res_d);
1191 }
1192
1193 void Parse::l2f() {
1194 Node* f2 = pop();
1195 Node* f1 = pop();
1196 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::l2f_Type(),
1197 CAST_FROM_FN_PTR(address, SharedRuntime::l2f),
1198 "l2f", NULL, //no memory effects
1199 f1, f2);
1200 Node* res = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0));
1201
1202 push(res);
1203 }
1204
1205 void Parse::do_irem() {
1206 // Must keep both values on the expression-stack during null-check
1207 zero_check_int(peek());
1208 // Compile-time detect of null-exception?
1209 if (stopped()) return;
1210
1211 Node* b = pop();
1212 Node* a = pop();
1213
1214 const Type *t = _gvn.type(b);
1215 if (t != Type::TOP) {
1216 const TypeInt *ti = t->is_int();
1217 if (ti->is_con()) {
1218 int divisor = ti->get_con();
1219 // check for positive power of 2
1220 if (divisor > 0 &&
1221 (divisor & ~(divisor-1)) == divisor) {
1222 // yes !
1223 Node *mask = _gvn.intcon((divisor - 1));
1224 // Sigh, must handle negative dividends
1225 Node *zero = _gvn.intcon(0);
1226 IfNode *ifff = jump_if_fork_int(a, zero, BoolTest::lt, PROB_FAIR, COUNT_UNKNOWN);
1227 Node *iff = _gvn.transform( new IfFalseNode(ifff) );
1228 Node *ift = _gvn.transform( new IfTrueNode (ifff) );
1229 Node *reg = jump_if_join(ift, iff);
1230 Node *phi = PhiNode::make(reg, NULL, TypeInt::INT);
1231 // Negative path; negate/and/negate
1232 Node *neg = _gvn.transform( new SubINode(zero, a) );
1233 Node *andn= _gvn.transform( new AndINode(neg, mask) );
1234 Node *negn= _gvn.transform( new SubINode(zero, andn) );
1235 phi->init_req(1, negn);
1236 // Fast positive case
1237 Node *andx = _gvn.transform( new AndINode(a, mask) );
1238 phi->init_req(2, andx);
1239 // Push the merge
1240 push( _gvn.transform(phi) );
1241 return;
1242 }
1243 }
1244 }
1245 // Default case
1246 push( _gvn.transform( new ModINode(control(),a,b) ) );
1247 }
1248
1249 // Handle jsr and jsr_w bytecode
1250 void Parse::do_jsr() {
1251 assert(bc() == Bytecodes::_jsr || bc() == Bytecodes::_jsr_w, "wrong bytecode");
1252
1253 // Store information about current state, tagged with new _jsr_bci
1254 int return_bci = iter().next_bci();
1255 int jsr_bci = (bc() == Bytecodes::_jsr) ? iter().get_dest() : iter().get_far_dest();
1256
1257 // Update method data
1258 profile_taken_branch(jsr_bci);
1259
1260 // The way we do things now, there is only one successor block
1261 // for the jsr, because the target code is cloned by ciTypeFlow.
1262 Block* target = successor_for_bci(jsr_bci);
1263
1264 // What got pushed?
1265 const Type* ret_addr = target->peek();
1266 assert(ret_addr->singleton(), "must be a constant (cloned jsr body)");
1267
1268 // Effect on jsr on stack
1269 push(_gvn.makecon(ret_addr));
1270
1271 // Flow to the jsr.
1272 merge(jsr_bci);
1273 }
1274
1275 // Handle ret bytecode
1276 void Parse::do_ret() {
1277 // Find to whom we return.
1278 assert(block()->num_successors() == 1, "a ret can only go one place now");
1279 Block* target = block()->successor_at(0);
1280 assert(!target->is_ready(), "our arrival must be expected");
1281 profile_ret(target->flow()->start());
1282 int pnum = target->next_path_num();
1283 merge_common(target, pnum);
1284 }
1285
1286 static bool has_injected_profile(BoolTest::mask btest, Node* test, int& taken, int& not_taken) {
1287 if (btest != BoolTest::eq && btest != BoolTest::ne) {
1288 // Only ::eq and ::ne are supported for profile injection.
1289 return false;
1290 }
1291 if (test->is_Cmp() &&
1292 test->in(1)->Opcode() == Op_ProfileBoolean) {
1293 ProfileBooleanNode* profile = (ProfileBooleanNode*)test->in(1);
1294 int false_cnt = profile->false_count();
1295 int true_cnt = profile->true_count();
1296
1297 // Counts matching depends on the actual test operation (::eq or ::ne).
1298 // No need to scale the counts because profile injection was designed
1299 // to feed exact counts into VM.
1300 taken = (btest == BoolTest::eq) ? false_cnt : true_cnt;
1301 not_taken = (btest == BoolTest::eq) ? true_cnt : false_cnt;
1302
1303 profile->consume();
1304 return true;
1305 }
1306 return false;
1307 }
1308 //--------------------------dynamic_branch_prediction--------------------------
1309 // Try to gather dynamic branch prediction behavior. Return a probability
1310 // of the branch being taken and set the "cnt" field. Returns a -1.0
1311 // if we need to use static prediction for some reason.
1312 float Parse::dynamic_branch_prediction(float &cnt, BoolTest::mask btest, Node* test) {
1313 ResourceMark rm;
1314
1315 cnt = COUNT_UNKNOWN;
1316
1317 int taken = 0;
1318 int not_taken = 0;
1319
1320 bool use_mdo = !has_injected_profile(btest, test, taken, not_taken);
1321
1322 if (use_mdo) {
1323 // Use MethodData information if it is available
1324 // FIXME: free the ProfileData structure
1325 ciMethodData* methodData = method()->method_data();
1326 if (!methodData->is_mature()) return PROB_UNKNOWN;
1327 ciProfileData* data = methodData->bci_to_data(bci());
1328 if (data == NULL) {
1329 return PROB_UNKNOWN;
1330 }
1331 if (!data->is_JumpData()) return PROB_UNKNOWN;
1332
1333 // get taken and not taken values
1334 taken = data->as_JumpData()->taken();
1335 not_taken = 0;
1336 if (data->is_BranchData()) {
1337 not_taken = data->as_BranchData()->not_taken();
1338 }
1339
1340 // scale the counts to be commensurate with invocation counts:
1341 taken = method()->scale_count(taken);
1342 not_taken = method()->scale_count(not_taken);
1343 }
1344
1345 // Give up if too few (or too many, in which case the sum will overflow) counts to be meaningful.
1346 // We also check that individual counters are positive first, otherwise the sum can become positive.
1347 if (taken < 0 || not_taken < 0 || taken + not_taken < 40) {
1348 if (C->log() != NULL) {
1349 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d'", iter().get_dest(), taken, not_taken);
1350 }
1351 return PROB_UNKNOWN;
1352 }
1353
1354 // Compute frequency that we arrive here
1355 float sum = taken + not_taken;
1356 // Adjust, if this block is a cloned private block but the
1357 // Jump counts are shared. Taken the private counts for
1358 // just this path instead of the shared counts.
1359 if( block()->count() > 0 )
1360 sum = block()->count();
1361 cnt = sum / FreqCountInvocations;
1362
1363 // Pin probability to sane limits
1364 float prob;
1365 if( !taken )
1366 prob = (0+PROB_MIN) / 2;
1367 else if( !not_taken )
1368 prob = (1+PROB_MAX) / 2;
1369 else { // Compute probability of true path
1370 prob = (float)taken / (float)(taken + not_taken);
1371 if (prob > PROB_MAX) prob = PROB_MAX;
1372 if (prob < PROB_MIN) prob = PROB_MIN;
1373 }
1374
1375 assert((cnt > 0.0f) && (prob > 0.0f),
1376 "Bad frequency assignment in if");
1377
1378 if (C->log() != NULL) {
1379 const char* prob_str = NULL;
1380 if (prob >= PROB_MAX) prob_str = (prob == PROB_MAX) ? "max" : "always";
1381 if (prob <= PROB_MIN) prob_str = (prob == PROB_MIN) ? "min" : "never";
1382 char prob_str_buf[30];
1383 if (prob_str == NULL) {
1384 sprintf(prob_str_buf, "%g", prob);
1385 prob_str = prob_str_buf;
1386 }
1387 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d' cnt='%f' prob='%s'",
1388 iter().get_dest(), taken, not_taken, cnt, prob_str);
1389 }
1390 return prob;
1391 }
1392
1393 //-----------------------------branch_prediction-------------------------------
1394 float Parse::branch_prediction(float& cnt,
1395 BoolTest::mask btest,
1396 int target_bci,
1397 Node* test) {
1398 float prob = dynamic_branch_prediction(cnt, btest, test);
1399 // If prob is unknown, switch to static prediction
1400 if (prob != PROB_UNKNOWN) return prob;
1401
1402 prob = PROB_FAIR; // Set default value
1403 if (btest == BoolTest::eq) // Exactly equal test?
1404 prob = PROB_STATIC_INFREQUENT; // Assume its relatively infrequent
1405 else if (btest == BoolTest::ne)
1406 prob = PROB_STATIC_FREQUENT; // Assume its relatively frequent
1407
1408 // If this is a conditional test guarding a backwards branch,
1409 // assume its a loop-back edge. Make it a likely taken branch.
1410 if (target_bci < bci()) {
1411 if (is_osr_parse()) { // Could be a hot OSR'd loop; force deopt
1412 // Since it's an OSR, we probably have profile data, but since
1413 // branch_prediction returned PROB_UNKNOWN, the counts are too small.
1414 // Let's make a special check here for completely zero counts.
1415 ciMethodData* methodData = method()->method_data();
1416 if (!methodData->is_empty()) {
1417 ciProfileData* data = methodData->bci_to_data(bci());
1418 // Only stop for truly zero counts, which mean an unknown part
1419 // of the OSR-ed method, and we want to deopt to gather more stats.
1420 // If you have ANY counts, then this loop is simply 'cold' relative
1421 // to the OSR loop.
1422 if (data == NULL ||
1423 (data->as_BranchData()->taken() + data->as_BranchData()->not_taken() == 0)) {
1424 // This is the only way to return PROB_UNKNOWN:
1425 return PROB_UNKNOWN;
1426 }
1427 }
1428 }
1429 prob = PROB_STATIC_FREQUENT; // Likely to take backwards branch
1430 }
1431
1432 assert(prob != PROB_UNKNOWN, "must have some guess at this point");
1433 return prob;
1434 }
1435
1436 // The magic constants are chosen so as to match the output of
1437 // branch_prediction() when the profile reports a zero taken count.
1438 // It is important to distinguish zero counts unambiguously, because
1439 // some branches (e.g., _213_javac.Assembler.eliminate) validly produce
1440 // very small but nonzero probabilities, which if confused with zero
1441 // counts would keep the program recompiling indefinitely.
1442 bool Parse::seems_never_taken(float prob) const {
1443 return prob < PROB_MIN;
1444 }
1445
1446 // True if the comparison seems to be the kind that will not change its
1447 // statistics from true to false. See comments in adjust_map_after_if.
1448 // This question is only asked along paths which are already
1449 // classifed as untaken (by seems_never_taken), so really,
1450 // if a path is never taken, its controlling comparison is
1451 // already acting in a stable fashion. If the comparison
1452 // seems stable, we will put an expensive uncommon trap
1453 // on the untaken path.
1454 bool Parse::seems_stable_comparison() const {
1455 if (C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if)) {
1456 return false;
1457 }
1458 return true;
1459 }
1460
1461 //-------------------------------repush_if_args--------------------------------
1462 // Push arguments of an "if" bytecode back onto the stack by adjusting _sp.
1463 inline int Parse::repush_if_args() {
1464 if (PrintOpto && WizardMode) {
1465 tty->print("defending against excessive implicit null exceptions on %s @%d in ",
1466 Bytecodes::name(iter().cur_bc()), iter().cur_bci());
1467 method()->print_name(); tty->cr();
1468 }
1469 int bc_depth = - Bytecodes::depth(iter().cur_bc());
1470 assert(bc_depth == 1 || bc_depth == 2, "only two kinds of branches");
1471 DEBUG_ONLY(sync_jvms()); // argument(n) requires a synced jvms
1472 assert(argument(0) != NULL, "must exist");
1473 assert(bc_depth == 1 || argument(1) != NULL, "two must exist");
1474 inc_sp(bc_depth);
1475 return bc_depth;
1476 }
1477
1478 //----------------------------------do_ifnull----------------------------------
1479 void Parse::do_ifnull(BoolTest::mask btest, Node *c) {
1480 int target_bci = iter().get_dest();
1481
1482 Block* branch_block = successor_for_bci(target_bci);
1483 Block* next_block = successor_for_bci(iter().next_bci());
1484
1485 float cnt;
1486 float prob = branch_prediction(cnt, btest, target_bci, c);
1487 if (prob == PROB_UNKNOWN) {
1488 // (An earlier version of do_ifnull omitted this trap for OSR methods.)
1489 if (PrintOpto && Verbose) {
1490 tty->print_cr("Never-taken edge stops compilation at bci %d", bci());
1491 }
1492 repush_if_args(); // to gather stats on loop
1493 // We need to mark this branch as taken so that if we recompile we will
1494 // see that it is possible. In the tiered system the interpreter doesn't
1495 // do profiling and by the time we get to the lower tier from the interpreter
1496 // the path may be cold again. Make sure it doesn't look untaken
1497 profile_taken_branch(target_bci, !ProfileInterpreter);
1498 uncommon_trap(Deoptimization::Reason_unreached,
1499 Deoptimization::Action_reinterpret,
1500 NULL, "cold");
1501 if (C->eliminate_boxing()) {
1502 // Mark the successor blocks as parsed
1503 branch_block->next_path_num();
1504 next_block->next_path_num();
1505 }
1506 return;
1507 }
1508
1509 NOT_PRODUCT(explicit_null_checks_inserted++);
1510
1511 // Generate real control flow
1512 Node *tst = _gvn.transform( new BoolNode( c, btest ) );
1513
1514 // Sanity check the probability value
1515 assert(prob > 0.0f,"Bad probability in Parser");
1516 // Need xform to put node in hash table
1517 IfNode *iff = create_and_xform_if( control(), tst, prob, cnt );
1518 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser");
1519 // True branch
1520 { PreserveJVMState pjvms(this);
1521 Node* iftrue = _gvn.transform( new IfTrueNode (iff) );
1522 set_control(iftrue);
1523
1524 if (stopped()) { // Path is dead?
1525 NOT_PRODUCT(explicit_null_checks_elided++);
1526 if (C->eliminate_boxing()) {
1527 // Mark the successor block as parsed
1528 branch_block->next_path_num();
1529 }
1530 } else { // Path is live.
1531 // Update method data
1532 profile_taken_branch(target_bci);
1533 adjust_map_after_if(btest, c, prob, branch_block, next_block);
1534 if (!stopped()) {
1535 merge(target_bci);
1536 }
1537 }
1538 }
1539
1540 // False branch
1541 Node* iffalse = _gvn.transform( new IfFalseNode(iff) );
1542 set_control(iffalse);
1543
1544 if (stopped()) { // Path is dead?
1545 NOT_PRODUCT(explicit_null_checks_elided++);
1546 if (C->eliminate_boxing()) {
1547 // Mark the successor block as parsed
1548 next_block->next_path_num();
1549 }
1550 } else { // Path is live.
1551 // Update method data
1552 profile_not_taken_branch();
1553 adjust_map_after_if(BoolTest(btest).negate(), c, 1.0-prob,
1554 next_block, branch_block);
1555 }
1556 }
1557
1558 //------------------------------------do_if------------------------------------
1559 void Parse::do_if(BoolTest::mask btest, Node* c) {
1560 int target_bci = iter().get_dest();
1561
1562 Block* branch_block = successor_for_bci(target_bci);
1563 Block* next_block = successor_for_bci(iter().next_bci());
1564
1565 float cnt;
1566 float prob = branch_prediction(cnt, btest, target_bci, c);
1567 float untaken_prob = 1.0 - prob;
1568
1569 if (prob == PROB_UNKNOWN) {
1570 if (PrintOpto && Verbose) {
1571 tty->print_cr("Never-taken edge stops compilation at bci %d", bci());
1572 }
1573 repush_if_args(); // to gather stats on loop
1574 // We need to mark this branch as taken so that if we recompile we will
1575 // see that it is possible. In the tiered system the interpreter doesn't
1576 // do profiling and by the time we get to the lower tier from the interpreter
1577 // the path may be cold again. Make sure it doesn't look untaken
1578 profile_taken_branch(target_bci, !ProfileInterpreter);
1579 uncommon_trap(Deoptimization::Reason_unreached,
1580 Deoptimization::Action_reinterpret,
1581 NULL, "cold");
1582 if (C->eliminate_boxing()) {
1583 // Mark the successor blocks as parsed
1584 branch_block->next_path_num();
1585 next_block->next_path_num();
1586 }
1587 return;
1588 }
1589
1590 // Sanity check the probability value
1591 assert(0.0f < prob && prob < 1.0f,"Bad probability in Parser");
1592
1593 bool taken_if_true = true;
1594 // Convert BoolTest to canonical form:
1595 if (!BoolTest(btest).is_canonical()) {
1596 btest = BoolTest(btest).negate();
1597 taken_if_true = false;
1598 // prob is NOT updated here; it remains the probability of the taken
1599 // path (as opposed to the prob of the path guarded by an 'IfTrueNode').
1600 }
1601 assert(btest != BoolTest::eq, "!= is the only canonical exact test");
1602
1603 Node* tst0 = new BoolNode(c, btest);
1604 Node* tst = _gvn.transform(tst0);
1605 BoolTest::mask taken_btest = BoolTest::illegal;
1606 BoolTest::mask untaken_btest = BoolTest::illegal;
1607
1608 if (tst->is_Bool()) {
1609 // Refresh c from the transformed bool node, since it may be
1610 // simpler than the original c. Also re-canonicalize btest.
1611 // This wins when (Bool ne (Conv2B p) 0) => (Bool ne (CmpP p NULL)).
1612 // That can arise from statements like: if (x instanceof C) ...
1613 if (tst != tst0) {
1614 // Canonicalize one more time since transform can change it.
1615 btest = tst->as_Bool()->_test._test;
1616 if (!BoolTest(btest).is_canonical()) {
1617 // Reverse edges one more time...
1618 tst = _gvn.transform( tst->as_Bool()->negate(&_gvn) );
1619 btest = tst->as_Bool()->_test._test;
1620 assert(BoolTest(btest).is_canonical(), "sanity");
1621 taken_if_true = !taken_if_true;
1622 }
1623 c = tst->in(1);
1624 }
1625 BoolTest::mask neg_btest = BoolTest(btest).negate();
1626 taken_btest = taken_if_true ? btest : neg_btest;
1627 untaken_btest = taken_if_true ? neg_btest : btest;
1628 }
1629
1630 // Generate real control flow
1631 float true_prob = (taken_if_true ? prob : untaken_prob);
1632 IfNode* iff = create_and_map_if(control(), tst, true_prob, cnt);
1633 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser");
1634 Node* taken_branch = new IfTrueNode(iff);
1635 Node* untaken_branch = new IfFalseNode(iff);
1636 if (!taken_if_true) { // Finish conversion to canonical form
1637 Node* tmp = taken_branch;
1638 taken_branch = untaken_branch;
1639 untaken_branch = tmp;
1640 }
1641
1642 // Branch is taken:
1643 { PreserveJVMState pjvms(this);
1644 taken_branch = _gvn.transform(taken_branch);
1645 set_control(taken_branch);
1646
1647 if (stopped()) {
1648 if (C->eliminate_boxing()) {
1649 // Mark the successor block as parsed
1650 branch_block->next_path_num();
1651 }
1652 } else {
1653 // Update method data
1654 profile_taken_branch(target_bci);
1655 adjust_map_after_if(taken_btest, c, prob, branch_block, next_block);
1656 if (!stopped()) {
1657 merge(target_bci);
1658 }
1659 }
1660 }
1661
1662 untaken_branch = _gvn.transform(untaken_branch);
1663 set_control(untaken_branch);
1664
1665 // Branch not taken.
1666 if (stopped()) {
1667 if (C->eliminate_boxing()) {
1668 // Mark the successor block as parsed
1669 next_block->next_path_num();
1670 }
1671 } else {
1672 // Update method data
1673 profile_not_taken_branch();
1674 adjust_map_after_if(untaken_btest, c, untaken_prob,
1675 next_block, branch_block);
1676 }
1677 }
1678
1679 bool Parse::path_is_suitable_for_uncommon_trap(float prob) const {
1680 // Don't want to speculate on uncommon traps when running with -Xcomp
1681 if (!UseInterpreter) {
1682 return false;
1683 }
1684 return (seems_never_taken(prob) && seems_stable_comparison());
1685 }
1686
1687 //----------------------------adjust_map_after_if------------------------------
1688 // Adjust the JVM state to reflect the result of taking this path.
1689 // Basically, it means inspecting the CmpNode controlling this
1690 // branch, seeing how it constrains a tested value, and then
1691 // deciding if it's worth our while to encode this constraint
1692 // as graph nodes in the current abstract interpretation map.
1693 void Parse::adjust_map_after_if(BoolTest::mask btest, Node* c, float prob,
1694 Block* path, Block* other_path) {
1695 if (stopped() || !c->is_Cmp() || btest == BoolTest::illegal)
1696 return; // nothing to do
1697
1698 bool is_fallthrough = (path == successor_for_bci(iter().next_bci()));
1699
1700 if (path_is_suitable_for_uncommon_trap(prob)) {
1701 repush_if_args();
1702 uncommon_trap(Deoptimization::Reason_unstable_if,
1703 Deoptimization::Action_reinterpret,
1704 NULL,
1705 (is_fallthrough ? "taken always" : "taken never"));
1706 return;
1707 }
1708
1709 Node* val = c->in(1);
1710 Node* con = c->in(2);
1711 const Type* tcon = _gvn.type(con);
1712 const Type* tval = _gvn.type(val);
1713 bool have_con = tcon->singleton();
1714 if (tval->singleton()) {
1715 if (!have_con) {
1716 // Swap, so constant is in con.
1717 con = val;
1718 tcon = tval;
1719 val = c->in(2);
1720 tval = _gvn.type(val);
1721 btest = BoolTest(btest).commute();
1722 have_con = true;
1723 } else {
1724 // Do we have two constants? Then leave well enough alone.
1725 have_con = false;
1726 }
1727 }
1728 if (!have_con) // remaining adjustments need a con
1729 return;
1730
1731 sharpen_type_after_if(btest, con, tcon, val, tval);
1732 }
1733
1734
1735 static Node* extract_obj_from_klass_load(PhaseGVN* gvn, Node* n) {
1736 Node* ldk;
1737 if (n->is_DecodeNKlass()) {
1738 if (n->in(1)->Opcode() != Op_LoadNKlass) {
1739 return NULL;
1740 } else {
1741 ldk = n->in(1);
1742 }
1743 } else if (n->Opcode() != Op_LoadKlass) {
1744 return NULL;
1745 } else {
1746 ldk = n;
1747 }
1748 assert(ldk != NULL && ldk->is_Load(), "should have found a LoadKlass or LoadNKlass node");
1749
1750 Node* adr = ldk->in(MemNode::Address);
1751 intptr_t off = 0;
1752 Node* obj = AddPNode::Ideal_base_and_offset(adr, gvn, off);
1753 if (obj == NULL || off != oopDesc::klass_offset_in_bytes()) // loading oopDesc::_klass?
1754 return NULL;
1755 const TypePtr* tp = gvn->type(obj)->is_ptr();
1756 if (tp == NULL || !(tp->isa_instptr() || tp->isa_aryptr())) // is obj a Java object ptr?
1757 return NULL;
1758
1759 return obj;
1760 }
1761
1762 void Parse::sharpen_type_after_if(BoolTest::mask btest,
1763 Node* con, const Type* tcon,
1764 Node* val, const Type* tval) {
1765 // Look for opportunities to sharpen the type of a node
1766 // whose klass is compared with a constant klass.
1767 if (btest == BoolTest::eq && tcon->isa_klassptr()) {
1768 Node* obj = extract_obj_from_klass_load(&_gvn, val);
1769 const TypeOopPtr* con_type = tcon->isa_klassptr()->as_instance_type();
1770 if (obj != NULL && (con_type->isa_instptr() || con_type->isa_aryptr())) {
1771 // Found:
1772 // Bool(CmpP(LoadKlass(obj._klass), ConP(Foo.klass)), [eq])
1773 // or the narrowOop equivalent.
1774 const Type* obj_type = _gvn.type(obj);
1775 const TypeOopPtr* tboth = obj_type->join_speculative(con_type)->isa_oopptr();
1776 if (tboth != NULL && tboth->klass_is_exact() && tboth != obj_type &&
1777 tboth->higher_equal(obj_type)) {
1778 // obj has to be of the exact type Foo if the CmpP succeeds.
1779 int obj_in_map = map()->find_edge(obj);
1780 JVMState* jvms = this->jvms();
1781 if (obj_in_map >= 0 &&
1782 (jvms->is_loc(obj_in_map) || jvms->is_stk(obj_in_map))) {
1783 TypeNode* ccast = new CheckCastPPNode(control(), obj, tboth);
1784 const Type* tcc = ccast->as_Type()->type();
1785 assert(tcc != obj_type && tcc->higher_equal(obj_type), "must improve");
1786 // Delay transform() call to allow recovery of pre-cast value
1787 // at the control merge.
1788 _gvn.set_type_bottom(ccast);
1789 record_for_igvn(ccast);
1790 // Here's the payoff.
1791 replace_in_map(obj, ccast);
1792 }
1793 }
1794 }
1795 }
1796
1797 int val_in_map = map()->find_edge(val);
1798 if (val_in_map < 0) return; // replace_in_map would be useless
1799 {
1800 JVMState* jvms = this->jvms();
1801 if (!(jvms->is_loc(val_in_map) ||
1802 jvms->is_stk(val_in_map)))
1803 return; // again, it would be useless
1804 }
1805
1806 // Check for a comparison to a constant, and "know" that the compared
1807 // value is constrained on this path.
1808 assert(tcon->singleton(), "");
1809 ConstraintCastNode* ccast = NULL;
1810 Node* cast = NULL;
1811
1812 switch (btest) {
1813 case BoolTest::eq: // Constant test?
1814 {
1815 const Type* tboth = tcon->join_speculative(tval);
1816 if (tboth == tval) break; // Nothing to gain.
1817 if (tcon->isa_int()) {
1818 ccast = new CastIINode(val, tboth);
1819 } else if (tcon == TypePtr::NULL_PTR) {
1820 // Cast to null, but keep the pointer identity temporarily live.
1821 ccast = new CastPPNode(val, tboth);
1822 } else {
1823 const TypeF* tf = tcon->isa_float_constant();
1824 const TypeD* td = tcon->isa_double_constant();
1825 // Exclude tests vs float/double 0 as these could be
1826 // either +0 or -0. Just because you are equal to +0
1827 // doesn't mean you ARE +0!
1828 // Note, following code also replaces Long and Oop values.
1829 if ((!tf || tf->_f != 0.0) &&
1830 (!td || td->_d != 0.0))
1831 cast = con; // Replace non-constant val by con.
1832 }
1833 }
1834 break;
1835
1836 case BoolTest::ne:
1837 if (tcon == TypePtr::NULL_PTR) {
1838 cast = cast_not_null(val, false);
1839 }
1840 break;
1841
1842 default:
1843 // (At this point we could record int range types with CastII.)
1844 break;
1845 }
1846
1847 if (ccast != NULL) {
1848 const Type* tcc = ccast->as_Type()->type();
1849 assert(tcc != tval && tcc->higher_equal(tval), "must improve");
1850 // Delay transform() call to allow recovery of pre-cast value
1851 // at the control merge.
1852 ccast->set_req(0, control());
1853 _gvn.set_type_bottom(ccast);
1854 record_for_igvn(ccast);
1855 cast = ccast;
1856 }
1857
1858 if (cast != NULL) { // Here's the payoff.
1859 replace_in_map(val, cast);
1860 }
1861 }
1862
1863 /**
1864 * Use speculative type to optimize CmpP node: if comparison is
1865 * against the low level class, cast the object to the speculative
1866 * type if any. CmpP should then go away.
1867 *
1868 * @param c expected CmpP node
1869 * @return result of CmpP on object casted to speculative type
1870 *
1871 */
1872 Node* Parse::optimize_cmp_with_klass(Node* c) {
1873 // If this is transformed by the _gvn to a comparison with the low
1874 // level klass then we may be able to use speculation
1875 if (c->Opcode() == Op_CmpP &&
1876 (c->in(1)->Opcode() == Op_LoadKlass || c->in(1)->Opcode() == Op_DecodeNKlass) &&
1877 c->in(2)->is_Con()) {
1878 Node* load_klass = NULL;
1879 Node* decode = NULL;
1880 if (c->in(1)->Opcode() == Op_DecodeNKlass) {
1881 decode = c->in(1);
1882 load_klass = c->in(1)->in(1);
1883 } else {
1884 load_klass = c->in(1);
1885 }
1886 if (load_klass->in(2)->is_AddP()) {
1887 Node* addp = load_klass->in(2);
1888 Node* obj = addp->in(AddPNode::Address);
1889 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
1890 if (obj_type->speculative_type_not_null() != NULL) {
1891 ciKlass* k = obj_type->speculative_type();
1892 inc_sp(2);
1893 obj = maybe_cast_profiled_obj(obj, k);
1894 dec_sp(2);
1895 // Make the CmpP use the casted obj
1896 addp = basic_plus_adr(obj, addp->in(AddPNode::Offset));
1897 load_klass = load_klass->clone();
1898 load_klass->set_req(2, addp);
1899 load_klass = _gvn.transform(load_klass);
1900 if (decode != NULL) {
1901 decode = decode->clone();
1902 decode->set_req(1, load_klass);
1903 load_klass = _gvn.transform(decode);
1904 }
1905 c = c->clone();
1906 c->set_req(1, load_klass);
1907 c = _gvn.transform(c);
1908 }
1909 }
1910 }
1911 return c;
1912 }
1913
1914 //------------------------------do_one_bytecode--------------------------------
1915 // Parse this bytecode, and alter the Parsers JVM->Node mapping
1916 void Parse::do_one_bytecode() {
1917 Node *a, *b, *c, *d; // Handy temps
1918 BoolTest::mask btest;
1919 int i;
1920
1921 assert(!has_exceptions(), "bytecode entry state must be clear of throws");
1922
1923 if (C->check_node_count(NodeLimitFudgeFactor * 5,
1924 "out of nodes parsing method")) {
1925 return;
1926 }
1927
1928 #ifdef ASSERT
1929 // for setting breakpoints
1930 if (TraceOptoParse) {
1931 tty->print(" @");
1932 dump_bci(bci());
1933 tty->cr();
1934 }
1935 #endif
1936
1937 switch (bc()) {
1938 case Bytecodes::_nop:
1939 // do nothing
1940 break;
1941 case Bytecodes::_lconst_0:
1942 push_pair(longcon(0));
1943 break;
1944
1945 case Bytecodes::_lconst_1:
1946 push_pair(longcon(1));
1947 break;
1948
1949 case Bytecodes::_fconst_0:
1950 push(zerocon(T_FLOAT));
1951 break;
1952
1953 case Bytecodes::_fconst_1:
1954 push(makecon(TypeF::ONE));
1955 break;
1956
1957 case Bytecodes::_fconst_2:
1958 push(makecon(TypeF::make(2.0f)));
1959 break;
1960
1961 case Bytecodes::_dconst_0:
1962 push_pair(zerocon(T_DOUBLE));
1963 break;
1964
1965 case Bytecodes::_dconst_1:
1966 push_pair(makecon(TypeD::ONE));
1967 break;
1968
1969 case Bytecodes::_iconst_m1:push(intcon(-1)); break;
1970 case Bytecodes::_iconst_0: push(intcon( 0)); break;
1971 case Bytecodes::_iconst_1: push(intcon( 1)); break;
1972 case Bytecodes::_iconst_2: push(intcon( 2)); break;
1973 case Bytecodes::_iconst_3: push(intcon( 3)); break;
1974 case Bytecodes::_iconst_4: push(intcon( 4)); break;
1975 case Bytecodes::_iconst_5: push(intcon( 5)); break;
1976 case Bytecodes::_bipush: push(intcon(iter().get_constant_u1())); break;
1977 case Bytecodes::_sipush: push(intcon(iter().get_constant_u2())); break;
1978 case Bytecodes::_aconst_null: push(null()); break;
1979 case Bytecodes::_ldc:
1980 case Bytecodes::_ldc_w:
1981 case Bytecodes::_ldc2_w:
1982 // If the constant is unresolved, run this BC once in the interpreter.
1983 {
1984 ciConstant constant = iter().get_constant();
1985 if (!constant.is_valid() ||
1986 (constant.basic_type() == T_OBJECT &&
1987 !constant.as_object()->is_loaded())) {
1988 int index = iter().get_constant_pool_index();
1989 constantTag tag = iter().get_constant_pool_tag(index);
1990 uncommon_trap(Deoptimization::make_trap_request
1991 (Deoptimization::Reason_unloaded,
1992 Deoptimization::Action_reinterpret,
1993 index),
1994 NULL, tag.internal_name());
1995 break;
1996 }
1997 assert(constant.basic_type() != T_OBJECT || constant.as_object()->is_instance(),
1998 "must be java_mirror of klass");
1999 const Type* con_type = Type::make_from_constant(constant);
2000 if (con_type != NULL) {
2001 push_node(con_type->basic_type(), makecon(con_type));
2002 }
2003 }
2004
2005 break;
2006
2007 case Bytecodes::_aload_0:
2008 push( local(0) );
2009 break;
2010 case Bytecodes::_aload_1:
2011 push( local(1) );
2012 break;
2013 case Bytecodes::_aload_2:
2014 push( local(2) );
2015 break;
2016 case Bytecodes::_aload_3:
2017 push( local(3) );
2018 break;
2019 case Bytecodes::_aload:
2020 push( local(iter().get_index()) );
2021 break;
2022
2023 case Bytecodes::_fload_0:
2024 case Bytecodes::_iload_0:
2025 push( local(0) );
2026 break;
2027 case Bytecodes::_fload_1:
2028 case Bytecodes::_iload_1:
2029 push( local(1) );
2030 break;
2031 case Bytecodes::_fload_2:
2032 case Bytecodes::_iload_2:
2033 push( local(2) );
2034 break;
2035 case Bytecodes::_fload_3:
2036 case Bytecodes::_iload_3:
2037 push( local(3) );
2038 break;
2039 case Bytecodes::_fload:
2040 case Bytecodes::_iload:
2041 push( local(iter().get_index()) );
2042 break;
2043 case Bytecodes::_lload_0:
2044 push_pair_local( 0 );
2045 break;
2046 case Bytecodes::_lload_1:
2047 push_pair_local( 1 );
2048 break;
2049 case Bytecodes::_lload_2:
2050 push_pair_local( 2 );
2051 break;
2052 case Bytecodes::_lload_3:
2053 push_pair_local( 3 );
2054 break;
2055 case Bytecodes::_lload:
2056 push_pair_local( iter().get_index() );
2057 break;
2058
2059 case Bytecodes::_dload_0:
2060 push_pair_local(0);
2061 break;
2062 case Bytecodes::_dload_1:
2063 push_pair_local(1);
2064 break;
2065 case Bytecodes::_dload_2:
2066 push_pair_local(2);
2067 break;
2068 case Bytecodes::_dload_3:
2069 push_pair_local(3);
2070 break;
2071 case Bytecodes::_dload:
2072 push_pair_local(iter().get_index());
2073 break;
2074 case Bytecodes::_fstore_0:
2075 case Bytecodes::_istore_0:
2076 case Bytecodes::_astore_0:
2077 set_local( 0, pop() );
2078 break;
2079 case Bytecodes::_fstore_1:
2080 case Bytecodes::_istore_1:
2081 case Bytecodes::_astore_1:
2082 set_local( 1, pop() );
2083 break;
2084 case Bytecodes::_fstore_2:
2085 case Bytecodes::_istore_2:
2086 case Bytecodes::_astore_2:
2087 set_local( 2, pop() );
2088 break;
2089 case Bytecodes::_fstore_3:
2090 case Bytecodes::_istore_3:
2091 case Bytecodes::_astore_3:
2092 set_local( 3, pop() );
2093 break;
2094 case Bytecodes::_fstore:
2095 case Bytecodes::_istore:
2096 case Bytecodes::_astore:
2097 set_local( iter().get_index(), pop() );
2098 break;
2099 // long stores
2100 case Bytecodes::_lstore_0:
2101 set_pair_local( 0, pop_pair() );
2102 break;
2103 case Bytecodes::_lstore_1:
2104 set_pair_local( 1, pop_pair() );
2105 break;
2106 case Bytecodes::_lstore_2:
2107 set_pair_local( 2, pop_pair() );
2108 break;
2109 case Bytecodes::_lstore_3:
2110 set_pair_local( 3, pop_pair() );
2111 break;
2112 case Bytecodes::_lstore:
2113 set_pair_local( iter().get_index(), pop_pair() );
2114 break;
2115
2116 // double stores
2117 case Bytecodes::_dstore_0:
2118 set_pair_local( 0, dstore_rounding(pop_pair()) );
2119 break;
2120 case Bytecodes::_dstore_1:
2121 set_pair_local( 1, dstore_rounding(pop_pair()) );
2122 break;
2123 case Bytecodes::_dstore_2:
2124 set_pair_local( 2, dstore_rounding(pop_pair()) );
2125 break;
2126 case Bytecodes::_dstore_3:
2127 set_pair_local( 3, dstore_rounding(pop_pair()) );
2128 break;
2129 case Bytecodes::_dstore:
2130 set_pair_local( iter().get_index(), dstore_rounding(pop_pair()) );
2131 break;
2132
2133 case Bytecodes::_pop: dec_sp(1); break;
2134 case Bytecodes::_pop2: dec_sp(2); break;
2135 case Bytecodes::_swap:
2136 a = pop();
2137 b = pop();
2138 push(a);
2139 push(b);
2140 break;
2141 case Bytecodes::_dup:
2142 a = pop();
2143 push(a);
2144 push(a);
2145 break;
2146 case Bytecodes::_dup_x1:
2147 a = pop();
2148 b = pop();
2149 push( a );
2150 push( b );
2151 push( a );
2152 break;
2153 case Bytecodes::_dup_x2:
2154 a = pop();
2155 b = pop();
2156 c = pop();
2157 push( a );
2158 push( c );
2159 push( b );
2160 push( a );
2161 break;
2162 case Bytecodes::_dup2:
2163 a = pop();
2164 b = pop();
2165 push( b );
2166 push( a );
2167 push( b );
2168 push( a );
2169 break;
2170
2171 case Bytecodes::_dup2_x1:
2172 // before: .. c, b, a
2173 // after: .. b, a, c, b, a
2174 // not tested
2175 a = pop();
2176 b = pop();
2177 c = pop();
2178 push( b );
2179 push( a );
2180 push( c );
2181 push( b );
2182 push( a );
2183 break;
2184 case Bytecodes::_dup2_x2:
2185 // before: .. d, c, b, a
2186 // after: .. b, a, d, c, b, a
2187 // not tested
2188 a = pop();
2189 b = pop();
2190 c = pop();
2191 d = pop();
2192 push( b );
2193 push( a );
2194 push( d );
2195 push( c );
2196 push( b );
2197 push( a );
2198 break;
2199
2200 case Bytecodes::_arraylength: {
2201 // Must do null-check with value on expression stack
2202 Node *ary = null_check(peek(), T_ARRAY);
2203 // Compile-time detect of null-exception?
2204 if (stopped()) return;
2205 a = pop();
2206 push(load_array_length(a));
2207 break;
2208 }
2209
2210 case Bytecodes::_baload: array_load(T_BYTE); break;
2211 case Bytecodes::_caload: array_load(T_CHAR); break;
2212 case Bytecodes::_iaload: array_load(T_INT); break;
2213 case Bytecodes::_saload: array_load(T_SHORT); break;
2214 case Bytecodes::_faload: array_load(T_FLOAT); break;
2215 case Bytecodes::_aaload: array_load(T_OBJECT); break;
2216 case Bytecodes::_laload: array_load(T_LONG); break;
2217 case Bytecodes::_daload: array_load(T_DOUBLE); break;
2218 case Bytecodes::_bastore: array_store(T_BYTE); break;
2219 case Bytecodes::_castore: array_store(T_CHAR); break;
2220 case Bytecodes::_iastore: array_store(T_INT); break;
2221 case Bytecodes::_sastore: array_store(T_SHORT); break;
2222 case Bytecodes::_fastore: array_store(T_FLOAT); break;
2223 case Bytecodes::_aastore: array_store(T_OBJECT); break;
2224 case Bytecodes::_lastore: array_store(T_LONG); break;
2225 case Bytecodes::_dastore: array_store(T_DOUBLE); break;
2226
2227 case Bytecodes::_getfield:
2228 do_getfield();
2229 break;
2230
2231 case Bytecodes::_getstatic:
2232 do_getstatic();
2233 break;
2234
2235 case Bytecodes::_putfield:
2236 do_putfield();
2237 break;
2238
2239 case Bytecodes::_putstatic:
2240 do_putstatic();
2241 break;
2242
2243 case Bytecodes::_irem:
2244 do_irem();
2245 break;
2246 case Bytecodes::_idiv:
2247 // Must keep both values on the expression-stack during null-check
2248 zero_check_int(peek());
2249 // Compile-time detect of null-exception?
2250 if (stopped()) return;
2251 b = pop();
2252 a = pop();
2253 push( _gvn.transform( new DivINode(control(),a,b) ) );
2254 break;
2255 case Bytecodes::_imul:
2256 b = pop(); a = pop();
2257 push( _gvn.transform( new MulINode(a,b) ) );
2258 break;
2259 case Bytecodes::_iadd:
2260 b = pop(); a = pop();
2261 push( _gvn.transform( new AddINode(a,b) ) );
2262 break;
2263 case Bytecodes::_ineg:
2264 a = pop();
2265 push( _gvn.transform( new SubINode(_gvn.intcon(0),a)) );
2266 break;
2267 case Bytecodes::_isub:
2268 b = pop(); a = pop();
2269 push( _gvn.transform( new SubINode(a,b) ) );
2270 break;
2271 case Bytecodes::_iand:
2272 b = pop(); a = pop();
2273 push( _gvn.transform( new AndINode(a,b) ) );
2274 break;
2275 case Bytecodes::_ior:
2276 b = pop(); a = pop();
2277 push( _gvn.transform( new OrINode(a,b) ) );
2278 break;
2279 case Bytecodes::_ixor:
2280 b = pop(); a = pop();
2281 push( _gvn.transform( new XorINode(a,b) ) );
2282 break;
2283 case Bytecodes::_ishl:
2284 b = pop(); a = pop();
2285 push( _gvn.transform( new LShiftINode(a,b) ) );
2286 break;
2287 case Bytecodes::_ishr:
2288 b = pop(); a = pop();
2289 push( _gvn.transform( new RShiftINode(a,b) ) );
2290 break;
2291 case Bytecodes::_iushr:
2292 b = pop(); a = pop();
2293 push( _gvn.transform( new URShiftINode(a,b) ) );
2294 break;
2295
2296 case Bytecodes::_fneg:
2297 a = pop();
2298 b = _gvn.transform(new NegFNode (a));
2299 push(b);
2300 break;
2301
2302 case Bytecodes::_fsub:
2303 b = pop();
2304 a = pop();
2305 c = _gvn.transform( new SubFNode(a,b) );
2306 d = precision_rounding(c);
2307 push( d );
2308 break;
2309
2310 case Bytecodes::_fadd:
2311 b = pop();
2312 a = pop();
2313 c = _gvn.transform( new AddFNode(a,b) );
2314 d = precision_rounding(c);
2315 push( d );
2316 break;
2317
2318 case Bytecodes::_fmul:
2319 b = pop();
2320 a = pop();
2321 c = _gvn.transform( new MulFNode(a,b) );
2322 d = precision_rounding(c);
2323 push( d );
2324 break;
2325
2326 case Bytecodes::_fdiv:
2327 b = pop();
2328 a = pop();
2329 c = _gvn.transform( new DivFNode(0,a,b) );
2330 d = precision_rounding(c);
2331 push( d );
2332 break;
2333
2334 case Bytecodes::_frem:
2335 if (Matcher::has_match_rule(Op_ModF)) {
2336 // Generate a ModF node.
2337 b = pop();
2338 a = pop();
2339 c = _gvn.transform( new ModFNode(0,a,b) );
2340 d = precision_rounding(c);
2341 push( d );
2342 }
2343 else {
2344 // Generate a call.
2345 modf();
2346 }
2347 break;
2348
2349 case Bytecodes::_fcmpl:
2350 b = pop();
2351 a = pop();
2352 c = _gvn.transform( new CmpF3Node( a, b));
2353 push(c);
2354 break;
2355 case Bytecodes::_fcmpg:
2356 b = pop();
2357 a = pop();
2358
2359 // Same as fcmpl but need to flip the unordered case. Swap the inputs,
2360 // which negates the result sign except for unordered. Flip the unordered
2361 // as well by using CmpF3 which implements unordered-lesser instead of
2362 // unordered-greater semantics. Finally, commute the result bits. Result
2363 // is same as using a CmpF3Greater except we did it with CmpF3 alone.
2364 c = _gvn.transform( new CmpF3Node( b, a));
2365 c = _gvn.transform( new SubINode(_gvn.intcon(0),c) );
2366 push(c);
2367 break;
2368
2369 case Bytecodes::_f2i:
2370 a = pop();
2371 push(_gvn.transform(new ConvF2INode(a)));
2372 break;
2373
2374 case Bytecodes::_d2i:
2375 a = pop_pair();
2376 b = _gvn.transform(new ConvD2INode(a));
2377 push( b );
2378 break;
2379
2380 case Bytecodes::_f2d:
2381 a = pop();
2382 b = _gvn.transform( new ConvF2DNode(a));
2383 push_pair( b );
2384 break;
2385
2386 case Bytecodes::_d2f:
2387 a = pop_pair();
2388 b = _gvn.transform( new ConvD2FNode(a));
2389 // This breaks _227_mtrt (speed & correctness) and _222_mpegaudio (speed)
2390 //b = _gvn.transform(new RoundFloatNode(0, b) );
2391 push( b );
2392 break;
2393
2394 case Bytecodes::_l2f:
2395 if (Matcher::convL2FSupported()) {
2396 a = pop_pair();
2397 b = _gvn.transform( new ConvL2FNode(a));
2398 // For i486.ad, FILD doesn't restrict precision to 24 or 53 bits.
2399 // Rather than storing the result into an FP register then pushing
2400 // out to memory to round, the machine instruction that implements
2401 // ConvL2D is responsible for rounding.
2402 // c = precision_rounding(b);
2403 c = _gvn.transform(b);
2404 push(c);
2405 } else {
2406 l2f();
2407 }
2408 break;
2409
2410 case Bytecodes::_l2d:
2411 a = pop_pair();
2412 b = _gvn.transform( new ConvL2DNode(a));
2413 // For i486.ad, rounding is always necessary (see _l2f above).
2414 // c = dprecision_rounding(b);
2415 c = _gvn.transform(b);
2416 push_pair(c);
2417 break;
2418
2419 case Bytecodes::_f2l:
2420 a = pop();
2421 b = _gvn.transform( new ConvF2LNode(a));
2422 push_pair(b);
2423 break;
2424
2425 case Bytecodes::_d2l:
2426 a = pop_pair();
2427 b = _gvn.transform( new ConvD2LNode(a));
2428 push_pair(b);
2429 break;
2430
2431 case Bytecodes::_dsub:
2432 b = pop_pair();
2433 a = pop_pair();
2434 c = _gvn.transform( new SubDNode(a,b) );
2435 d = dprecision_rounding(c);
2436 push_pair( d );
2437 break;
2438
2439 case Bytecodes::_dadd:
2440 b = pop_pair();
2441 a = pop_pair();
2442 c = _gvn.transform( new AddDNode(a,b) );
2443 d = dprecision_rounding(c);
2444 push_pair( d );
2445 break;
2446
2447 case Bytecodes::_dmul:
2448 b = pop_pair();
2449 a = pop_pair();
2450 c = _gvn.transform( new MulDNode(a,b) );
2451 d = dprecision_rounding(c);
2452 push_pair( d );
2453 break;
2454
2455 case Bytecodes::_ddiv:
2456 b = pop_pair();
2457 a = pop_pair();
2458 c = _gvn.transform( new DivDNode(0,a,b) );
2459 d = dprecision_rounding(c);
2460 push_pair( d );
2461 break;
2462
2463 case Bytecodes::_dneg:
2464 a = pop_pair();
2465 b = _gvn.transform(new NegDNode (a));
2466 push_pair(b);
2467 break;
2468
2469 case Bytecodes::_drem:
2470 if (Matcher::has_match_rule(Op_ModD)) {
2471 // Generate a ModD node.
2472 b = pop_pair();
2473 a = pop_pair();
2474 // a % b
2475
2476 c = _gvn.transform( new ModDNode(0,a,b) );
2477 d = dprecision_rounding(c);
2478 push_pair( d );
2479 }
2480 else {
2481 // Generate a call.
2482 modd();
2483 }
2484 break;
2485
2486 case Bytecodes::_dcmpl:
2487 b = pop_pair();
2488 a = pop_pair();
2489 c = _gvn.transform( new CmpD3Node( a, b));
2490 push(c);
2491 break;
2492
2493 case Bytecodes::_dcmpg:
2494 b = pop_pair();
2495 a = pop_pair();
2496 // Same as dcmpl but need to flip the unordered case.
2497 // Commute the inputs, which negates the result sign except for unordered.
2498 // Flip the unordered as well by using CmpD3 which implements
2499 // unordered-lesser instead of unordered-greater semantics.
2500 // Finally, negate the result bits. Result is same as using a
2501 // CmpD3Greater except we did it with CmpD3 alone.
2502 c = _gvn.transform( new CmpD3Node( b, a));
2503 c = _gvn.transform( new SubINode(_gvn.intcon(0),c) );
2504 push(c);
2505 break;
2506
2507
2508 // Note for longs -> lo word is on TOS, hi word is on TOS - 1
2509 case Bytecodes::_land:
2510 b = pop_pair();
2511 a = pop_pair();
2512 c = _gvn.transform( new AndLNode(a,b) );
2513 push_pair(c);
2514 break;
2515 case Bytecodes::_lor:
2516 b = pop_pair();
2517 a = pop_pair();
2518 c = _gvn.transform( new OrLNode(a,b) );
2519 push_pair(c);
2520 break;
2521 case Bytecodes::_lxor:
2522 b = pop_pair();
2523 a = pop_pair();
2524 c = _gvn.transform( new XorLNode(a,b) );
2525 push_pair(c);
2526 break;
2527
2528 case Bytecodes::_lshl:
2529 b = pop(); // the shift count
2530 a = pop_pair(); // value to be shifted
2531 c = _gvn.transform( new LShiftLNode(a,b) );
2532 push_pair(c);
2533 break;
2534 case Bytecodes::_lshr:
2535 b = pop(); // the shift count
2536 a = pop_pair(); // value to be shifted
2537 c = _gvn.transform( new RShiftLNode(a,b) );
2538 push_pair(c);
2539 break;
2540 case Bytecodes::_lushr:
2541 b = pop(); // the shift count
2542 a = pop_pair(); // value to be shifted
2543 c = _gvn.transform( new URShiftLNode(a,b) );
2544 push_pair(c);
2545 break;
2546 case Bytecodes::_lmul:
2547 b = pop_pair();
2548 a = pop_pair();
2549 c = _gvn.transform( new MulLNode(a,b) );
2550 push_pair(c);
2551 break;
2552
2553 case Bytecodes::_lrem:
2554 // Must keep both values on the expression-stack during null-check
2555 assert(peek(0) == top(), "long word order");
2556 zero_check_long(peek(1));
2557 // Compile-time detect of null-exception?
2558 if (stopped()) return;
2559 b = pop_pair();
2560 a = pop_pair();
2561 c = _gvn.transform( new ModLNode(control(),a,b) );
2562 push_pair(c);
2563 break;
2564
2565 case Bytecodes::_ldiv:
2566 // Must keep both values on the expression-stack during null-check
2567 assert(peek(0) == top(), "long word order");
2568 zero_check_long(peek(1));
2569 // Compile-time detect of null-exception?
2570 if (stopped()) return;
2571 b = pop_pair();
2572 a = pop_pair();
2573 c = _gvn.transform( new DivLNode(control(),a,b) );
2574 push_pair(c);
2575 break;
2576
2577 case Bytecodes::_ladd:
2578 b = pop_pair();
2579 a = pop_pair();
2580 c = _gvn.transform( new AddLNode(a,b) );
2581 push_pair(c);
2582 break;
2583 case Bytecodes::_lsub:
2584 b = pop_pair();
2585 a = pop_pair();
2586 c = _gvn.transform( new SubLNode(a,b) );
2587 push_pair(c);
2588 break;
2589 case Bytecodes::_lcmp:
2590 // Safepoints are now inserted _before_ branches. The long-compare
2591 // bytecode painfully produces a 3-way value (-1,0,+1) which requires a
2592 // slew of control flow. These are usually followed by a CmpI vs zero and
2593 // a branch; this pattern then optimizes to the obvious long-compare and
2594 // branch. However, if the branch is backwards there's a Safepoint
2595 // inserted. The inserted Safepoint captures the JVM state at the
2596 // pre-branch point, i.e. it captures the 3-way value. Thus if a
2597 // long-compare is used to control a loop the debug info will force
2598 // computation of the 3-way value, even though the generated code uses a
2599 // long-compare and branch. We try to rectify the situation by inserting
2600 // a SafePoint here and have it dominate and kill the safepoint added at a
2601 // following backwards branch. At this point the JVM state merely holds 2
2602 // longs but not the 3-way value.
2603 if( UseLoopSafepoints ) {
2604 switch( iter().next_bc() ) {
2605 case Bytecodes::_ifgt:
2606 case Bytecodes::_iflt:
2607 case Bytecodes::_ifge:
2608 case Bytecodes::_ifle:
2609 case Bytecodes::_ifne:
2610 case Bytecodes::_ifeq:
2611 // If this is a backwards branch in the bytecodes, add Safepoint
2612 maybe_add_safepoint(iter().next_get_dest());
2613 default:
2614 break;
2615 }
2616 }
2617 b = pop_pair();
2618 a = pop_pair();
2619 c = _gvn.transform( new CmpL3Node( a, b ));
2620 push(c);
2621 break;
2622
2623 case Bytecodes::_lneg:
2624 a = pop_pair();
2625 b = _gvn.transform( new SubLNode(longcon(0),a));
2626 push_pair(b);
2627 break;
2628 case Bytecodes::_l2i:
2629 a = pop_pair();
2630 push( _gvn.transform( new ConvL2INode(a)));
2631 break;
2632 case Bytecodes::_i2l:
2633 a = pop();
2634 b = _gvn.transform( new ConvI2LNode(a));
2635 push_pair(b);
2636 break;
2637 case Bytecodes::_i2b:
2638 // Sign extend
2639 a = pop();
2640 a = _gvn.transform( new LShiftINode(a,_gvn.intcon(24)) );
2641 a = _gvn.transform( new RShiftINode(a,_gvn.intcon(24)) );
2642 push( a );
2643 break;
2644 case Bytecodes::_i2s:
2645 a = pop();
2646 a = _gvn.transform( new LShiftINode(a,_gvn.intcon(16)) );
2647 a = _gvn.transform( new RShiftINode(a,_gvn.intcon(16)) );
2648 push( a );
2649 break;
2650 case Bytecodes::_i2c:
2651 a = pop();
2652 push( _gvn.transform( new AndINode(a,_gvn.intcon(0xFFFF)) ) );
2653 break;
2654
2655 case Bytecodes::_i2f:
2656 a = pop();
2657 b = _gvn.transform( new ConvI2FNode(a) ) ;
2658 c = precision_rounding(b);
2659 push (b);
2660 break;
2661
2662 case Bytecodes::_i2d:
2663 a = pop();
2664 b = _gvn.transform( new ConvI2DNode(a));
2665 push_pair(b);
2666 break;
2667
2668 case Bytecodes::_iinc: // Increment local
2669 i = iter().get_index(); // Get local index
2670 set_local( i, _gvn.transform( new AddINode( _gvn.intcon(iter().get_iinc_con()), local(i) ) ) );
2671 break;
2672
2673 // Exit points of synchronized methods must have an unlock node
2674 case Bytecodes::_return:
2675 return_current(NULL);
2676 break;
2677
2678 case Bytecodes::_ireturn:
2679 case Bytecodes::_areturn:
2680 case Bytecodes::_freturn:
2681 return_current(pop());
2682 break;
2683 case Bytecodes::_lreturn:
2684 return_current(pop_pair());
2685 break;
2686 case Bytecodes::_dreturn:
2687 return_current(pop_pair());
2688 break;
2689
2690 case Bytecodes::_athrow:
2691 // null exception oop throws NULL pointer exception
2692 null_check(peek());
2693 if (stopped()) return;
2694 // Hook the thrown exception directly to subsequent handlers.
2695 if (BailoutToInterpreterForThrows) {
2696 // Keep method interpreted from now on.
2697 uncommon_trap(Deoptimization::Reason_unhandled,
2698 Deoptimization::Action_make_not_compilable);
2699 return;
2700 }
2701 if (env()->jvmti_can_post_on_exceptions()) {
2702 // check if we must post exception events, take uncommon trap if so (with must_throw = false)
2703 uncommon_trap_if_should_post_on_exceptions(Deoptimization::Reason_unhandled, false);
2704 }
2705 // Here if either can_post_on_exceptions or should_post_on_exceptions is false
2706 add_exception_state(make_exception_state(peek()));
2707 break;
2708
2709 case Bytecodes::_goto: // fall through
2710 case Bytecodes::_goto_w: {
2711 int target_bci = (bc() == Bytecodes::_goto) ? iter().get_dest() : iter().get_far_dest();
2712
2713 // If this is a backwards branch in the bytecodes, add Safepoint
2714 maybe_add_safepoint(target_bci);
2715
2716 // Update method data
2717 profile_taken_branch(target_bci);
2718
2719 // Merge the current control into the target basic block
2720 merge(target_bci);
2721
2722 // See if we can get some profile data and hand it off to the next block
2723 Block *target_block = block()->successor_for_bci(target_bci);
2724 if (target_block->pred_count() != 1) break;
2725 ciMethodData* methodData = method()->method_data();
2726 if (!methodData->is_mature()) break;
2727 ciProfileData* data = methodData->bci_to_data(bci());
2728 assert(data != NULL && data->is_JumpData(), "need JumpData for taken branch");
2729 int taken = ((ciJumpData*)data)->taken();
2730 taken = method()->scale_count(taken);
2731 target_block->set_count(taken);
2732 break;
2733 }
2734
2735 case Bytecodes::_ifnull: btest = BoolTest::eq; goto handle_if_null;
2736 case Bytecodes::_ifnonnull: btest = BoolTest::ne; goto handle_if_null;
2737 handle_if_null:
2738 // If this is a backwards branch in the bytecodes, add Safepoint
2739 maybe_add_safepoint(iter().get_dest());
2740 a = null();
2741 b = pop();
2742 if (b->is_ValueType()) {
2743 // Return constant false because 'b' is always non-null
2744 c = _gvn.intcon(1);
2745 } else {
2746 if (!_gvn.type(b)->speculative_maybe_null() &&
2747 !too_many_traps(Deoptimization::Reason_speculate_null_check)) {
2748 inc_sp(1);
2749 Node* null_ctl = top();
2750 b = null_check_oop(b, &null_ctl, true, true, true);
2751 assert(null_ctl->is_top(), "no null control here");
2752 dec_sp(1);
2753 } else if (_gvn.type(b)->speculative_always_null() &&
2754 !too_many_traps(Deoptimization::Reason_speculate_null_assert)) {
2755 inc_sp(1);
2756 b = null_assert(b);
2757 dec_sp(1);
2758 }
2759 c = _gvn.transform( new CmpPNode(b, a) );
2760 }
2761 do_ifnull(btest, c);
2762 break;
2763
2764 case Bytecodes::_if_acmpeq: btest = BoolTest::eq; goto handle_if_acmp;
2765 case Bytecodes::_if_acmpne: btest = BoolTest::ne; goto handle_if_acmp;
2766 handle_if_acmp:
2767 // If this is a backwards branch in the bytecodes, add Safepoint
2768 maybe_add_safepoint(iter().get_dest());
2769 a = pop();
2770 b = pop();
2771 c = _gvn.transform(acmp(a, b));
2772 c = optimize_cmp_with_klass(c);
2773 do_if(btest, c);
2774 break;
2775
2776 case Bytecodes::_ifeq: btest = BoolTest::eq; goto handle_ifxx;
2777 case Bytecodes::_ifne: btest = BoolTest::ne; goto handle_ifxx;
2778 case Bytecodes::_iflt: btest = BoolTest::lt; goto handle_ifxx;
2779 case Bytecodes::_ifle: btest = BoolTest::le; goto handle_ifxx;
2780 case Bytecodes::_ifgt: btest = BoolTest::gt; goto handle_ifxx;
2781 case Bytecodes::_ifge: btest = BoolTest::ge; goto handle_ifxx;
2782 handle_ifxx:
2783 // If this is a backwards branch in the bytecodes, add Safepoint
2784 maybe_add_safepoint(iter().get_dest());
2785 a = _gvn.intcon(0);
2786 b = pop();
2787 c = _gvn.transform( new CmpINode(b, a) );
2788 do_if(btest, c);
2789 break;
2790
2791 case Bytecodes::_if_icmpeq: btest = BoolTest::eq; goto handle_if_icmp;
2792 case Bytecodes::_if_icmpne: btest = BoolTest::ne; goto handle_if_icmp;
2793 case Bytecodes::_if_icmplt: btest = BoolTest::lt; goto handle_if_icmp;
2794 case Bytecodes::_if_icmple: btest = BoolTest::le; goto handle_if_icmp;
2795 case Bytecodes::_if_icmpgt: btest = BoolTest::gt; goto handle_if_icmp;
2796 case Bytecodes::_if_icmpge: btest = BoolTest::ge; goto handle_if_icmp;
2797 handle_if_icmp:
2798 // If this is a backwards branch in the bytecodes, add Safepoint
2799 maybe_add_safepoint(iter().get_dest());
2800 a = pop();
2801 b = pop();
2802 c = _gvn.transform( new CmpINode( b, a ) );
2803 do_if(btest, c);
2804 break;
2805
2806 case Bytecodes::_tableswitch:
2807 do_tableswitch();
2808 break;
2809
2810 case Bytecodes::_lookupswitch:
2811 do_lookupswitch();
2812 break;
2813
2814 case Bytecodes::_invokestatic:
2815 case Bytecodes::_invokedynamic:
2816 case Bytecodes::_invokespecial:
2817 case Bytecodes::_invokevirtual:
2818 case Bytecodes::_invokeinterface:
2819 do_call();
2820 break;
2821 case Bytecodes::_checkcast:
2822 do_checkcast();
2823 break;
2824 case Bytecodes::_instanceof:
2825 do_instanceof();
2826 break;
2827 case Bytecodes::_anewarray:
2828 do_newarray();
2829 break;
2830 case Bytecodes::_newarray:
2831 do_newarray((BasicType)iter().get_index());
2832 break;
2833 case Bytecodes::_multianewarray:
2834 do_multianewarray();
2835 break;
2836 case Bytecodes::_new:
2837 do_new();
2838 break;
2839 case Bytecodes::_defaultvalue:
2840 do_defaultvalue();
2841 break;
2842 case Bytecodes::_withfield:
2843 do_withfield();
2844 break;
2845
2846 case Bytecodes::_jsr:
2847 case Bytecodes::_jsr_w:
2848 do_jsr();
2849 break;
2850
2851 case Bytecodes::_ret:
2852 do_ret();
2853 break;
2854
2855
2856 case Bytecodes::_monitorenter:
2857 do_monitor_enter();
2858 break;
2859
2860 case Bytecodes::_monitorexit:
2861 do_monitor_exit();
2862 break;
2863
2864 case Bytecodes::_breakpoint:
2865 // Breakpoint set concurrently to compile
2866 // %%% use an uncommon trap?
2867 C->record_failure("breakpoint in method");
2868 return;
2869
2870 default:
2871 #ifndef PRODUCT
2872 map()->dump(99);
2873 #endif
2874 tty->print("\nUnhandled bytecode %s\n", Bytecodes::name(bc()) );
2875 ShouldNotReachHere();
2876 }
2877
2878 #ifndef PRODUCT
2879 IdealGraphPrinter *printer = C->printer();
2880 if (printer && printer->should_print(1)) {
2881 char buffer[256];
2882 sprintf(buffer, "Bytecode %d: %s", bci(), Bytecodes::name(bc()));
2883 bool old = printer->traverse_outs();
2884 printer->set_traverse_outs(true);
2885 printer->print_method(buffer, 4);
2886 printer->set_traverse_outs(old);
2887 }
2888 #endif
2889 }