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 "runtime/deoptimization.hpp"
46#include "runtime/sharedRuntime.hpp"
47
48#ifndef PRODUCT
49extern int explicit_null_checks_inserted,
50 explicit_null_checks_elided;
51#endif
52
53//---------------------------------array_load----------------------------------
54void Parse::array_load(BasicType bt) {
55 const Type* elemtype = Type::TOP;
56 bool big_val = bt == T_DOUBLE || bt == T_LONG;
57 Node* adr = array_addressing(bt, 0, &elemtype);
58 if (stopped()) return; // guaranteed null or range check
59
60 pop(); // index (already used)
61 Node* array = pop(); // the array itself
62
63 if (elemtype == TypeInt::BOOL) {
64 bt = T_BOOLEAN;
65 } else if (bt == T_OBJECT) {
66 elemtype = _gvn.type(array)->is_aryptr()->elem()->make_oopptr();
67 }
68
69 const TypeAryPtr* adr_type = TypeAryPtr::get_array_body_type(bt);
70
71 Node* ld = access_load_at(array, adr, adr_type, elemtype, bt,
72 IN_HEAP | IS_ARRAY | C2_CONTROL_DEPENDENT_LOAD);
73 if (big_val) {
74 push_pair(ld);
75 } else {
76 push(ld);
77 }
78}
79
80
81//--------------------------------array_store----------------------------------
82void Parse::array_store(BasicType bt) {
83 const Type* elemtype = Type::TOP;
84 bool big_val = bt == T_DOUBLE || bt == T_LONG;
85 Node* adr = array_addressing(bt, big_val ? 2 : 1, &elemtype);
86 if (stopped()) return; // guaranteed null or range check
87 if (bt == T_OBJECT) {
88 array_store_check();
89 }
90 Node* val; // Oop to store
91 if (big_val) {
92 val = pop_pair();
93 } else {
94 val = pop();
95 }
96 pop(); // index (already used)
97 Node* array = pop(); // the array itself
98
99 if (elemtype == TypeInt::BOOL) {
100 bt = T_BOOLEAN;
101 } else if (bt == T_OBJECT) {
102 elemtype = _gvn.type(array)->is_aryptr()->elem()->make_oopptr();
103 }
104
105 const TypeAryPtr* adr_type = TypeAryPtr::get_array_body_type(bt);
106
107 access_store_at(array, adr, adr_type, val, elemtype, bt, MO_UNORDERED | IN_HEAP | IS_ARRAY);
108}
109
110
111//------------------------------array_addressing-------------------------------
112// Pull array and index from the stack. Compute pointer-to-element.
113Node* Parse::array_addressing(BasicType type, int vals, const Type* *result2) {
114 Node *idx = peek(0+vals); // Get from stack without popping
115 Node *ary = peek(1+vals); // in case of exception
116
117 // Null check the array base, with correct stack contents
118 ary = null_check(ary, T_ARRAY);
119 // Compile-time detect of null-exception?
120 if (stopped()) return top();
121
122 const TypeAryPtr* arytype = _gvn.type(ary)->is_aryptr();
123 const TypeInt* sizetype = arytype->size();
124 const Type* elemtype = arytype->elem();
125
126 if (UseUniqueSubclasses && result2 != NULL) {
127 const Type* el = elemtype->make_ptr();
128 if (el && el->isa_instptr()) {
129 const TypeInstPtr* toop = el->is_instptr();
130 if (toop->klass()->as_instance_klass()->unique_concrete_subklass()) {
131 // If we load from "AbstractClass[]" we must see "ConcreteSubClass".
132 const Type* subklass = Type::get_const_type(toop->klass());
133 elemtype = subklass->join_speculative(el);
134 }
135 }
136 }
137
138 // Check for big class initializers with all constant offsets
139 // feeding into a known-size array.
140 const TypeInt* idxtype = _gvn.type(idx)->is_int();
141 // See if the highest idx value is less than the lowest array bound,
142 // and if the idx value cannot be negative:
143 bool need_range_check = true;
144 if (idxtype->_hi < sizetype->_lo && idxtype->_lo >= 0) {
145 need_range_check = false;
146 if (C->log() != NULL) C->log()->elem("observe that='!need_range_check'");
147 }
148
149 ciKlass * arytype_klass = arytype->klass();
150 if ((arytype_klass != NULL) && (!arytype_klass->is_loaded())) {
151 // Only fails for some -Xcomp runs
152 // The class is unloaded. We have to run this bytecode in the interpreter.
153 uncommon_trap(Deoptimization::Reason_unloaded,
154 Deoptimization::Action_reinterpret,
155 arytype->klass(), "!loaded array");
156 return top();
157 }
158
159 // Do the range check
160 if (GenerateRangeChecks && need_range_check) {
161 Node* tst;
162 if (sizetype->_hi <= 0) {
163 // The greatest array bound is negative, so we can conclude that we're
164 // compiling unreachable code, but the unsigned compare trick used below
165 // only works with non-negative lengths. Instead, hack "tst" to be zero so
166 // the uncommon_trap path will always be taken.
167 tst = _gvn.intcon(0);
168 } else {
169 // Range is constant in array-oop, so we can use the original state of mem
170 Node* len = load_array_length(ary);
171
172 // Test length vs index (standard trick using unsigned compare)
173 Node* chk = _gvn.transform( new CmpUNode(idx, len) );
174 BoolTest::mask btest = BoolTest::lt;
175 tst = _gvn.transform( new BoolNode(chk, btest) );
176 }
177 RangeCheckNode* rc = new RangeCheckNode(control(), tst, PROB_MAX, COUNT_UNKNOWN);
178 _gvn.set_type(rc, rc->Value(&_gvn));
179 if (!tst->is_Con()) {
180 record_for_igvn(rc);
181 }
182 set_control(_gvn.transform(new IfTrueNode(rc)));
183 // Branch to failure if out of bounds
184 {
185 PreserveJVMState pjvms(this);
186 set_control(_gvn.transform(new IfFalseNode(rc)));
187 if (C->allow_range_check_smearing()) {
188 // Do not use builtin_throw, since range checks are sometimes
189 // made more stringent by an optimistic transformation.
190 // This creates "tentative" range checks at this point,
191 // which are not guaranteed to throw exceptions.
192 // See IfNode::Ideal, is_range_check, adjust_check.
193 uncommon_trap(Deoptimization::Reason_range_check,
194 Deoptimization::Action_make_not_entrant,
195 NULL, "range_check");
196 } else {
197 // If we have already recompiled with the range-check-widening
198 // heroic optimization turned off, then we must really be throwing
199 // range check exceptions.
200 builtin_throw(Deoptimization::Reason_range_check, idx);
201 }
202 }
203 }
204 // Check for always knowing you are throwing a range-check exception
205 if (stopped()) return top();
206
207 // Make array address computation control dependent to prevent it
208 // from floating above the range check during loop optimizations.
209 Node* ptr = array_element_address(ary, idx, type, sizetype, control());
210
211 if (result2 != NULL) *result2 = elemtype;
212
213 assert(ptr != top(), "top should go hand-in-hand with stopped");
214
215 return ptr;
216}
217
218
219// returns IfNode
220IfNode* Parse::jump_if_fork_int(Node* a, Node* b, BoolTest::mask mask, float prob, float cnt) {
221 Node *cmp = _gvn.transform(new CmpINode(a, b)); // two cases: shiftcount > 32 and shiftcount <= 32
222 Node *tst = _gvn.transform(new BoolNode(cmp, mask));
223 IfNode *iff = create_and_map_if(control(), tst, prob, cnt);
224 return iff;
225}
226
227// return Region node
228Node* Parse::jump_if_join(Node* iffalse, Node* iftrue) {
229 Node *region = new RegionNode(3); // 2 results
230 record_for_igvn(region);
231 region->init_req(1, iffalse);
232 region->init_req(2, iftrue );
233 _gvn.set_type(region, Type::CONTROL);
234 region = _gvn.transform(region);
235 set_control (region);
236 return region;
237}
238
239// sentinel value for the target bci to mark never taken branches
240// (according to profiling)
241static const int never_reached = INT_MAX;
242
243//------------------------------helper for tableswitch-------------------------
244void Parse::jump_if_true_fork(IfNode *iff, int dest_bci_if_true, int prof_table_index, bool unc) {
245 // True branch, use existing map info
246 { PreserveJVMState pjvms(this);
247 Node *iftrue = _gvn.transform( new IfTrueNode (iff) );
248 set_control( iftrue );
249 if (unc) {
250 repush_if_args();
251 uncommon_trap(Deoptimization::Reason_unstable_if,
252 Deoptimization::Action_reinterpret,
253 NULL,
254 "taken always");
255 } else {
256 assert(dest_bci_if_true != never_reached, "inconsistent dest");
257 profile_switch_case(prof_table_index);
258 merge_new_path(dest_bci_if_true);
259 }
260 }
261
262 // False branch
263 Node *iffalse = _gvn.transform( new IfFalseNode(iff) );
264 set_control( iffalse );
265}
266
267void Parse::jump_if_false_fork(IfNode *iff, int dest_bci_if_true, int prof_table_index, bool unc) {
268 // True branch, use existing map info
269 { PreserveJVMState pjvms(this);
270 Node *iffalse = _gvn.transform( new IfFalseNode (iff) );
271 set_control( iffalse );
272 if (unc) {
273 repush_if_args();
274 uncommon_trap(Deoptimization::Reason_unstable_if,
275 Deoptimization::Action_reinterpret,
276 NULL,
277 "taken never");
278 } else {
279 assert(dest_bci_if_true != never_reached, "inconsistent dest");
280 profile_switch_case(prof_table_index);
281 merge_new_path(dest_bci_if_true);
282 }
283 }
284
285 // False branch
286 Node *iftrue = _gvn.transform( new IfTrueNode(iff) );
287 set_control( iftrue );
288}
289
290void Parse::jump_if_always_fork(int dest_bci, int prof_table_index, bool unc) {
291 // False branch, use existing map and control()
292 if (unc) {
293 repush_if_args();
294 uncommon_trap(Deoptimization::Reason_unstable_if,
295 Deoptimization::Action_reinterpret,
296 NULL,
297 "taken never");
298 } else {
299 assert(dest_bci != never_reached, "inconsistent dest");
300 profile_switch_case(prof_table_index);
301 merge_new_path(dest_bci);
302 }
303}
304
305
306extern "C" {
307 static int jint_cmp(const void *i, const void *j) {
308 int a = *(jint *)i;
309 int b = *(jint *)j;
310 return a > b ? 1 : a < b ? -1 : 0;
311 }
312}
313
314
315// Default value for methodData switch indexing. Must be a negative value to avoid
316// conflict with any legal switch index.
317#define NullTableIndex -1
318
319class SwitchRange : public StackObj {
320 // a range of integers coupled with a bci destination
321 jint _lo; // inclusive lower limit
322 jint _hi; // inclusive upper limit
323 int _dest;
324 int _table_index; // index into method data table
325 float _cnt; // how many times this range was hit according to profiling
326
327public:
328 jint lo() const { return _lo; }
329 jint hi() const { return _hi; }
330 int dest() const { return _dest; }
331 int table_index() const { return _table_index; }
332 bool is_singleton() const { return _lo == _hi; }
333 float cnt() const { return _cnt; }
334
335 void setRange(jint lo, jint hi, int dest, int table_index, float cnt) {
336 assert(lo <= hi, "must be a non-empty range");
337 _lo = lo, _hi = hi; _dest = dest; _table_index = table_index; _cnt = cnt;
338 assert(_cnt >= 0, "");
339 }
340 bool adjoinRange(jint lo, jint hi, int dest, int table_index, float cnt, bool trim_ranges) {
341 assert(lo <= hi, "must be a non-empty range");
342 if (lo == _hi+1 && table_index == _table_index) {
343 // see merge_ranges() comment below
344 if (trim_ranges) {
345 if (cnt == 0) {
346 if (_cnt != 0) {
347 return false;
348 }
349 if (dest != _dest) {
350 _dest = never_reached;
351 }
352 } else {
353 if (_cnt == 0) {
354 return false;
355 }
356 if (dest != _dest) {
357 return false;
358 }
359 }
360 } else {
361 if (dest != _dest) {
362 return false;
363 }
364 }
365 _hi = hi;
366 _cnt += cnt;
367 return true;
368 }
369 return false;
370 }
371
372 void set (jint value, int dest, int table_index, float cnt) {
373 setRange(value, value, dest, table_index, cnt);
374 }
375 bool adjoin(jint value, int dest, int table_index, float cnt, bool trim_ranges) {
376 return adjoinRange(value, value, dest, table_index, cnt, trim_ranges);
377 }
378 bool adjoin(SwitchRange& other) {
379 return adjoinRange(other._lo, other._hi, other._dest, other._table_index, other._cnt, false);
380 }
381
382 void print() {
383 if (is_singleton())
384 tty->print(" {%d}=>%d (cnt=%f)", lo(), dest(), cnt());
385 else if (lo() == min_jint)
386 tty->print(" {..%d}=>%d (cnt=%f)", hi(), dest(), cnt());
387 else if (hi() == max_jint)
388 tty->print(" {%d..}=>%d (cnt=%f)", lo(), dest(), cnt());
389 else
390 tty->print(" {%d..%d}=>%d (cnt=%f)", lo(), hi(), dest(), cnt());
391 }
392};
393
394// We try to minimize the number of ranges and the size of the taken
395// ones using profiling data. When ranges are created,
396// SwitchRange::adjoinRange() only allows 2 adjoining ranges to merge
397// if both were never hit or both were hit to build longer unreached
398// ranges. Here, we now merge adjoining ranges with the same
399// destination and finally set destination of unreached ranges to the
400// special value never_reached because it can help minimize the number
401// of tests that are necessary.
402//
403// For instance:
404// [0, 1] to target1 sometimes taken
405// [1, 2] to target1 never taken
406// [2, 3] to target2 never taken
407// would lead to:
408// [0, 1] to target1 sometimes taken
409// [1, 3] never taken
410//
411// (first 2 ranges to target1 are not merged)
412static void merge_ranges(SwitchRange* ranges, int& rp) {
413 if (rp == 0) {
414 return;
415 }
416 int shift = 0;
417 for (int j = 0; j < rp; j++) {
418 SwitchRange& r1 = ranges[j-shift];
419 SwitchRange& r2 = ranges[j+1];
420 if (r1.adjoin(r2)) {
421 shift++;
422 } else if (shift > 0) {
423 ranges[j+1-shift] = r2;
424 }
425 }
426 rp -= shift;
427 for (int j = 0; j <= rp; j++) {
428 SwitchRange& r = ranges[j];
429 if (r.cnt() == 0 && r.dest() != never_reached) {
430 r.setRange(r.lo(), r.hi(), never_reached, r.table_index(), r.cnt());
431 }
432 }
433}
434
435//-------------------------------do_tableswitch--------------------------------
436void Parse::do_tableswitch() {
437 Node* lookup = pop();
438 // Get information about tableswitch
439 int default_dest = iter().get_dest_table(0);
440 int lo_index = iter().get_int_table(1);
441 int hi_index = iter().get_int_table(2);
442 int len = hi_index - lo_index + 1;
443
444 if (len < 1) {
445 // If this is a backward branch, add safepoint
446 maybe_add_safepoint(default_dest);
447 merge(default_dest);
448 return;
449 }
450
451 ciMethodData* methodData = method()->method_data();
452 ciMultiBranchData* profile = NULL;
453 if (methodData->is_mature() && UseSwitchProfiling) {
454 ciProfileData* data = methodData->bci_to_data(bci());
455 if (data != NULL && data->is_MultiBranchData()) {
456 profile = (ciMultiBranchData*)data;
457 }
458 }
459 bool trim_ranges = !method_data_update() && !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if);
460
461 // generate decision tree, using trichotomy when possible
462 int rnum = len+2;
463 bool makes_backward_branch = false;
464 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum);
465 int rp = -1;
466 if (lo_index != min_jint) {
467 uint cnt = 1;
468 if (profile != NULL) {
469 cnt = profile->default_count() / (hi_index != max_jint ? 2 : 1);
470 }
471 ranges[++rp].setRange(min_jint, lo_index-1, default_dest, NullTableIndex, cnt);
472 }
473 for (int j = 0; j < len; j++) {
474 jint match_int = lo_index+j;
475 int dest = iter().get_dest_table(j+3);
476 makes_backward_branch |= (dest <= bci());
477 int table_index = method_data_update() ? j : NullTableIndex;
478 uint cnt = 1;
479 if (profile != NULL) {
480 cnt = profile->count_at(j);
481 }
482 if (rp < 0 || !ranges[rp].adjoin(match_int, dest, table_index, cnt, trim_ranges)) {
483 ranges[++rp].set(match_int, dest, table_index, cnt);
484 }
485 }
486 jint highest = lo_index+(len-1);
487 assert(ranges[rp].hi() == highest, "");
488 if (highest != max_jint) {
489 uint cnt = 1;
490 if (profile != NULL) {
491 cnt = profile->default_count() / (lo_index != min_jint ? 2 : 1);
492 }
493 if (!ranges[rp].adjoinRange(highest+1, max_jint, default_dest, NullTableIndex, cnt, trim_ranges)) {
494 ranges[++rp].setRange(highest+1, max_jint, default_dest, NullTableIndex, cnt);
495 }
496 }
497 assert(rp < len+2, "not too many ranges");
498
499 if (trim_ranges) {
500 merge_ranges(ranges, rp);
501 }
502
503 // Safepoint in case if backward branch observed
504 if( makes_backward_branch && UseLoopSafepoints )
505 add_safepoint();
506
507 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]);
508}
509
510
511//------------------------------do_lookupswitch--------------------------------
512void Parse::do_lookupswitch() {
513 Node *lookup = pop(); // lookup value
514 // Get information about lookupswitch
515 int default_dest = iter().get_dest_table(0);
516 int len = iter().get_int_table(1);
517
518 if (len < 1) { // If this is a backward branch, add safepoint
519 maybe_add_safepoint(default_dest);
520 merge(default_dest);
521 return;
522 }
523
524 ciMethodData* methodData = method()->method_data();
525 ciMultiBranchData* profile = NULL;
526 if (methodData->is_mature() && UseSwitchProfiling) {
527 ciProfileData* data = methodData->bci_to_data(bci());
528 if (data != NULL && data->is_MultiBranchData()) {
529 profile = (ciMultiBranchData*)data;
530 }
531 }
532 bool trim_ranges = !method_data_update() && !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if);
533
534 // generate decision tree, using trichotomy when possible
535 jint* table = NEW_RESOURCE_ARRAY(jint, len*3);
536 {
537 for (int j = 0; j < len; j++) {
538 table[3*j+0] = iter().get_int_table(2+2*j);
539 table[3*j+1] = iter().get_dest_table(2+2*j+1);
540 table[3*j+2] = profile == NULL ? 1 : profile->count_at(j);
541 }
542 qsort(table, len, 3*sizeof(table[0]), jint_cmp);
543 }
544
545 float defaults = 0;
546 jint prev = min_jint;
547 for (int j = 0; j < len; j++) {
548 jint match_int = table[3*j+0];
549 if (match_int != prev) {
550 defaults += (float)match_int - prev;
551 }
552 prev = match_int+1;
553 }
554 if (prev-1 != max_jint) {
555 defaults += (float)max_jint - prev + 1;
556 }
557 float default_cnt = 1;
558 if (profile != NULL) {
559 default_cnt = profile->default_count()/defaults;
560 }
561
562 int rnum = len*2+1;
563 bool makes_backward_branch = false;
564 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum);
565 int rp = -1;
566 for (int j = 0; j < len; j++) {
567 jint match_int = table[3*j+0];
568 int dest = table[3*j+1];
569 int cnt = table[3*j+2];
570 int next_lo = rp < 0 ? min_jint : ranges[rp].hi()+1;
571 int table_index = method_data_update() ? j : NullTableIndex;
572 makes_backward_branch |= (dest <= bci());
573 float c = default_cnt * ((float)match_int - next_lo);
574 if (match_int != next_lo && (rp < 0 || !ranges[rp].adjoinRange(next_lo, match_int-1, default_dest, NullTableIndex, c, trim_ranges))) {
575 assert(default_dest != never_reached, "sentinel value for dead destinations");
576 ranges[++rp].setRange(next_lo, match_int-1, default_dest, NullTableIndex, c);
577 }
578 if (rp < 0 || !ranges[rp].adjoin(match_int, dest, table_index, cnt, trim_ranges)) {
579 assert(dest != never_reached, "sentinel value for dead destinations");
580 ranges[++rp].set(match_int, dest, table_index, cnt);
581 }
582 }
583 jint highest = table[3*(len-1)];
584 assert(ranges[rp].hi() == highest, "");
585 if (highest != max_jint &&
586 !ranges[rp].adjoinRange(highest+1, max_jint, default_dest, NullTableIndex, default_cnt * ((float)max_jint - highest), trim_ranges)) {
587 ranges[++rp].setRange(highest+1, max_jint, default_dest, NullTableIndex, default_cnt * ((float)max_jint - highest));
588 }
589 assert(rp < rnum, "not too many ranges");
590
591 if (trim_ranges) {
592 merge_ranges(ranges, rp);
593 }
594
595 // Safepoint in case backward branch observed
596 if (makes_backward_branch && UseLoopSafepoints)
597 add_safepoint();
598
599 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]);
600}
601
602static float if_prob(float taken_cnt, float total_cnt) {
603 assert(taken_cnt <= total_cnt, "");
604 if (total_cnt == 0) {
605 return PROB_FAIR;
606 }
607 float p = taken_cnt / total_cnt;
608 return MIN2(MAX2(p, PROB_MIN), PROB_MAX);
609}
610
611static float if_cnt(float cnt) {
612 if (cnt == 0) {
613 return COUNT_UNKNOWN;
614 }
615 return cnt;
616}
617
618static float sum_of_cnts(SwitchRange *lo, SwitchRange *hi) {
619 float total_cnt = 0;
620 for (SwitchRange* sr = lo; sr <= hi; sr++) {
621 total_cnt += sr->cnt();
622 }
623 return total_cnt;
624}
625
626class SwitchRanges : public ResourceObj {
627public:
628 SwitchRange* _lo;
629 SwitchRange* _hi;
630 SwitchRange* _mid;
631 float _cost;
632
633 enum {
634 Start,
635 LeftDone,
636 RightDone,
637 Done
638 } _state;
639
640 SwitchRanges(SwitchRange *lo, SwitchRange *hi)
641 : _lo(lo), _hi(hi), _mid(NULL),
642 _cost(0), _state(Start) {
643 }
644
645 SwitchRanges()
646 : _lo(NULL), _hi(NULL), _mid(NULL),
647 _cost(0), _state(Start) {}
648};
649
650// Estimate cost of performing a binary search on lo..hi
651static float compute_tree_cost(SwitchRange *lo, SwitchRange *hi, float total_cnt) {
652 GrowableArray<SwitchRanges> tree;
653 SwitchRanges root(lo, hi);
654 tree.push(root);
655
656 float cost = 0;
657 do {
658 SwitchRanges& r = *tree.adr_at(tree.length()-1);
659 if (r._hi != r._lo) {
660 if (r._mid == NULL) {
661 float r_cnt = sum_of_cnts(r._lo, r._hi);
662
663 if (r_cnt == 0) {
664 tree.pop();
665 cost = 0;
666 continue;
667 }
668
669 SwitchRange* mid = NULL;
670 mid = r._lo;
671 for (float cnt = 0; ; ) {
672 assert(mid <= r._hi, "out of bounds");
673 cnt += mid->cnt();
674 if (cnt > r_cnt / 2) {
675 break;
676 }
677 mid++;
678 }
679 assert(mid <= r._hi, "out of bounds");
680 r._mid = mid;
681 r._cost = r_cnt / total_cnt;
682 }
683 r._cost += cost;
684 if (r._state < SwitchRanges::LeftDone && r._mid > r._lo) {
685 cost = 0;
686 r._state = SwitchRanges::LeftDone;
687 tree.push(SwitchRanges(r._lo, r._mid-1));
688 } else if (r._state < SwitchRanges::RightDone) {
689 cost = 0;
690 r._state = SwitchRanges::RightDone;
691 tree.push(SwitchRanges(r._mid == r._lo ? r._mid+1 : r._mid, r._hi));
692 } else {
693 tree.pop();
694 cost = r._cost;
695 }
696 } else {
697 tree.pop();
698 cost = r._cost;
699 }
700 } while (tree.length() > 0);
701
702
703 return cost;
704}
705
706// It sometimes pays off to test most common ranges before the binary search
707void Parse::linear_search_switch_ranges(Node* key_val, SwitchRange*& lo, SwitchRange*& hi) {
708 uint nr = hi - lo + 1;
709 float total_cnt = sum_of_cnts(lo, hi);
710
711 float min = compute_tree_cost(lo, hi, total_cnt);
712 float extra = 1;
713 float sub = 0;
714
715 SwitchRange* array1 = lo;
716 SwitchRange* array2 = NEW_RESOURCE_ARRAY(SwitchRange, nr);
717
718 SwitchRange* ranges = NULL;
719
720 while (nr >= 2) {
721 assert(lo == array1 || lo == array2, "one the 2 already allocated arrays");
722 ranges = (lo == array1) ? array2 : array1;
723
724 // Find highest frequency range
725 SwitchRange* candidate = lo;
726 for (SwitchRange* sr = lo+1; sr <= hi; sr++) {
727 if (sr->cnt() > candidate->cnt()) {
728 candidate = sr;
729 }
730 }
731 SwitchRange most_freq = *candidate;
732 if (most_freq.cnt() == 0) {
733 break;
734 }
735
736 // Copy remaining ranges into another array
737 int shift = 0;
738 for (uint i = 0; i < nr; i++) {
739 SwitchRange* sr = &lo[i];
740 if (sr != candidate) {
741 ranges[i-shift] = *sr;
742 } else {
743 shift++;
744 if (i > 0 && i < nr-1) {
745 SwitchRange prev = lo[i-1];
746 prev.setRange(prev.lo(), sr->hi(), prev.dest(), prev.table_index(), prev.cnt());
747 if (prev.adjoin(lo[i+1])) {
748 shift++;
749 i++;
750 }
751 ranges[i-shift] = prev;
752 }
753 }
754 }
755 nr -= shift;
756
757 // Evaluate cost of testing the most common range and performing a
758 // binary search on the other ranges
759 float cost = extra + compute_tree_cost(&ranges[0], &ranges[nr-1], total_cnt);
760 if (cost >= min) {
761 break;
762 }
763 // swap arrays
764 lo = &ranges[0];
765 hi = &ranges[nr-1];
766
767 // It pays off: emit the test for the most common range
768 assert(most_freq.cnt() > 0, "must be taken");
769 Node* val = _gvn.transform(new SubINode(key_val, _gvn.intcon(most_freq.lo())));
770 Node* cmp = _gvn.transform(new CmpUNode(val, _gvn.intcon(most_freq.hi() - most_freq.lo())));
771 Node* tst = _gvn.transform(new BoolNode(cmp, BoolTest::le));
772 IfNode* iff = create_and_map_if(control(), tst, if_prob(most_freq.cnt(), total_cnt), if_cnt(most_freq.cnt()));
773 jump_if_true_fork(iff, most_freq.dest(), most_freq.table_index(), false);
774
775 sub += most_freq.cnt() / total_cnt;
776 extra += 1 - sub;
777 min = cost;
778 }
779}
780
781//----------------------------create_jump_tables-------------------------------
782bool Parse::create_jump_tables(Node* key_val, SwitchRange* lo, SwitchRange* hi) {
783 // Are jumptables enabled
784 if (!UseJumpTables) return false;
785
786 // Are jumptables supported
787 if (!Matcher::has_match_rule(Op_Jump)) return false;
788
789 // Don't make jump table if profiling
790 if (method_data_update()) return false;
791
792 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if);
793
794 // Decide if a guard is needed to lop off big ranges at either (or
795 // both) end(s) of the input set. We'll call this the default target
796 // even though we can't be sure that it is the true "default".
797
798 bool needs_guard = false;
799 int default_dest;
800 int64_t total_outlier_size = 0;
801 int64_t hi_size = ((int64_t)hi->hi()) - ((int64_t)hi->lo()) + 1;
802 int64_t lo_size = ((int64_t)lo->hi()) - ((int64_t)lo->lo()) + 1;
803
804 if (lo->dest() == hi->dest()) {
805 total_outlier_size = hi_size + lo_size;
806 default_dest = lo->dest();
807 } else if (lo_size > hi_size) {
808 total_outlier_size = lo_size;
809 default_dest = lo->dest();
810 } else {
811 total_outlier_size = hi_size;
812 default_dest = hi->dest();
813 }
814
815 float total = sum_of_cnts(lo, hi);
816 float cost = compute_tree_cost(lo, hi, total);
817
818 // If a guard test will eliminate very sparse end ranges, then
819 // it is worth the cost of an extra jump.
820 float trimmed_cnt = 0;
821 if (total_outlier_size > (MaxJumpTableSparseness * 4)) {
822 needs_guard = true;
823 if (default_dest == lo->dest()) {
824 trimmed_cnt += lo->cnt();
825 lo++;
826 }
827 if (default_dest == hi->dest()) {
828 trimmed_cnt += hi->cnt();
829 hi--;
830 }
831 }
832
833 // Find the total number of cases and ranges
834 int64_t num_cases = ((int64_t)hi->hi()) - ((int64_t)lo->lo()) + 1;
835 int num_range = hi - lo + 1;
836
837 // Don't create table if: too large, too small, or too sparse.
838 if (num_cases > MaxJumpTableSize)
839 return false;
840 if (UseSwitchProfiling) {
841 // MinJumpTableSize is set so with a well balanced binary tree,
842 // when the number of ranges is MinJumpTableSize, it's cheaper to
843 // go through a JumpNode that a tree of IfNodes. Average cost of a
844 // tree of IfNodes with MinJumpTableSize is
845 // log2f(MinJumpTableSize) comparisons. So if the cost computed
846 // from profile data is less than log2f(MinJumpTableSize) then
847 // going with the binary search is cheaper.
848 if (cost < log2f(MinJumpTableSize)) {
849 return false;
850 }
851 } else {
852 if (num_cases < MinJumpTableSize)
853 return false;
854 }
855 if (num_cases > (MaxJumpTableSparseness * num_range))
856 return false;
857
858 // Normalize table lookups to zero
859 int lowval = lo->lo();
860 key_val = _gvn.transform( new SubINode(key_val, _gvn.intcon(lowval)) );
861
862 // Generate a guard to protect against input keyvals that aren't
863 // in the switch domain.
864 if (needs_guard) {
865 Node* size = _gvn.intcon(num_cases);
866 Node* cmp = _gvn.transform(new CmpUNode(key_val, size));
867 Node* tst = _gvn.transform(new BoolNode(cmp, BoolTest::ge));
868 IfNode* iff = create_and_map_if(control(), tst, if_prob(trimmed_cnt, total), if_cnt(trimmed_cnt));
869 jump_if_true_fork(iff, default_dest, NullTableIndex, trim_ranges && trimmed_cnt == 0);
870
871 total -= trimmed_cnt;
872 }
873
874 // Create an ideal node JumpTable that has projections
875 // of all possible ranges for a switch statement
876 // The key_val input must be converted to a pointer offset and scaled.
877 // Compare Parse::array_addressing above.
878
879 // Clean the 32-bit int into a real 64-bit offset.
880 // Otherwise, the jint value 0 might turn into an offset of 0x0800000000.
881 const TypeInt* ikeytype = TypeInt::make(0, num_cases, Type::WidenMin);
882 // Make I2L conversion control dependent to prevent it from
883 // floating above the range check during loop optimizations.
884 key_val = C->conv_I2X_index(&_gvn, key_val, ikeytype, control());
885
886 // Shift the value by wordsize so we have an index into the table, rather
887 // than a switch value
888 Node *shiftWord = _gvn.MakeConX(wordSize);
889 key_val = _gvn.transform( new MulXNode( key_val, shiftWord));
890
891 // Create the JumpNode
892 Arena* arena = C->comp_arena();
893 float* probs = (float*)arena->Amalloc(sizeof(float)*num_cases);
894 int i = 0;
895 if (total == 0) {
896 for (SwitchRange* r = lo; r <= hi; r++) {
897 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) {
898 probs[i] = 1.0F / num_cases;
899 }
900 }
901 } else {
902 for (SwitchRange* r = lo; r <= hi; r++) {
903 float prob = r->cnt()/total;
904 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) {
905 probs[i] = prob / (r->hi() - r->lo() + 1);
906 }
907 }
908 }
909
910 ciMethodData* methodData = method()->method_data();
911 ciMultiBranchData* profile = NULL;
912 if (methodData->is_mature()) {
913 ciProfileData* data = methodData->bci_to_data(bci());
914 if (data != NULL && data->is_MultiBranchData()) {
915 profile = (ciMultiBranchData*)data;
916 }
917 }
918
919 Node* jtn = _gvn.transform(new JumpNode(control(), key_val, num_cases, probs, profile == NULL ? COUNT_UNKNOWN : total));
920
921 // These are the switch destinations hanging off the jumpnode
922 i = 0;
923 for (SwitchRange* r = lo; r <= hi; r++) {
924 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) {
925 Node* input = _gvn.transform(new JumpProjNode(jtn, i, r->dest(), (int)(j - lowval)));
926 {
927 PreserveJVMState pjvms(this);
928 set_control(input);
929 jump_if_always_fork(r->dest(), r->table_index(), trim_ranges && r->cnt() == 0);
930 }
931 }
932 }
933 assert(i == num_cases, "miscount of cases");
934 stop_and_kill_map(); // no more uses for this JVMS
935 return true;
936}
937
938//----------------------------jump_switch_ranges-------------------------------
939void Parse::jump_switch_ranges(Node* key_val, SwitchRange *lo, SwitchRange *hi, int switch_depth) {
940 Block* switch_block = block();
941 bool trim_ranges = !method_data_update() && !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if);
942
943 if (switch_depth == 0) {
944 // Do special processing for the top-level call.
945 assert(lo->lo() == min_jint, "initial range must exhaust Type::INT");
946 assert(hi->hi() == max_jint, "initial range must exhaust Type::INT");
947
948 // Decrement pred-numbers for the unique set of nodes.
949#ifdef ASSERT
950 if (!trim_ranges) {
951 // Ensure that the block's successors are a (duplicate-free) set.
952 int successors_counted = 0; // block occurrences in [hi..lo]
953 int unique_successors = switch_block->num_successors();
954 for (int i = 0; i < unique_successors; i++) {
955 Block* target = switch_block->successor_at(i);
956
957 // Check that the set of successors is the same in both places.
958 int successors_found = 0;
959 for (SwitchRange* p = lo; p <= hi; p++) {
960 if (p->dest() == target->start()) successors_found++;
961 }
962 assert(successors_found > 0, "successor must be known");
963 successors_counted += successors_found;
964 }
965 assert(successors_counted == (hi-lo)+1, "no unexpected successors");
966 }
967#endif
968
969 // Maybe prune the inputs, based on the type of key_val.
970 jint min_val = min_jint;
971 jint max_val = max_jint;
972 const TypeInt* ti = key_val->bottom_type()->isa_int();
973 if (ti != NULL) {
974 min_val = ti->_lo;
975 max_val = ti->_hi;
976 assert(min_val <= max_val, "invalid int type");
977 }
978 while (lo->hi() < min_val) {
979 lo++;
980 }
981 if (lo->lo() < min_val) {
982 lo->setRange(min_val, lo->hi(), lo->dest(), lo->table_index(), lo->cnt());
983 }
984 while (hi->lo() > max_val) {
985 hi--;
986 }
987 if (hi->hi() > max_val) {
988 hi->setRange(hi->lo(), max_val, hi->dest(), hi->table_index(), hi->cnt());
989 }
990
991 linear_search_switch_ranges(key_val, lo, hi);
992 }
993
994#ifndef PRODUCT
995 if (switch_depth == 0) {
996 _max_switch_depth = 0;
997 _est_switch_depth = log2_intptr((hi-lo+1)-1)+1;
998 }
999#endif
1000
1001 assert(lo <= hi, "must be a non-empty set of ranges");
1002 if (lo == hi) {
1003 jump_if_always_fork(lo->dest(), lo->table_index(), trim_ranges && lo->cnt() == 0);
1004 } else {
1005 assert(lo->hi() == (lo+1)->lo()-1, "contiguous ranges");
1006 assert(hi->lo() == (hi-1)->hi()+1, "contiguous ranges");
1007
1008 if (create_jump_tables(key_val, lo, hi)) return;
1009
1010 SwitchRange* mid = NULL;
1011 float total_cnt = sum_of_cnts(lo, hi);
1012
1013 int nr = hi - lo + 1;
1014 if (UseSwitchProfiling) {
1015 // Don't keep the binary search tree balanced: pick up mid point
1016 // that split frequencies in half.
1017 float cnt = 0;
1018 for (SwitchRange* sr = lo; sr <= hi; sr++) {
1019 cnt += sr->cnt();
1020 if (cnt >= total_cnt / 2) {
1021 mid = sr;
1022 break;
1023 }
1024 }
1025 } else {
1026 mid = lo + nr/2;
1027
1028 // if there is an easy choice, pivot at a singleton:
1029 if (nr > 3 && !mid->is_singleton() && (mid-1)->is_singleton()) mid--;
1030
1031 assert(lo < mid && mid <= hi, "good pivot choice");
1032 assert(nr != 2 || mid == hi, "should pick higher of 2");
1033 assert(nr != 3 || mid == hi-1, "should pick middle of 3");
1034 }
1035
1036
1037 Node *test_val = _gvn.intcon(mid == lo ? mid->hi() : mid->lo());
1038
1039 if (mid->is_singleton()) {
1040 IfNode *iff_ne = jump_if_fork_int(key_val, test_val, BoolTest::ne, 1-if_prob(mid->cnt(), total_cnt), if_cnt(mid->cnt()));
1041 jump_if_false_fork(iff_ne, mid->dest(), mid->table_index(), trim_ranges && mid->cnt() == 0);
1042
1043 // Special Case: If there are exactly three ranges, and the high
1044 // and low range each go to the same place, omit the "gt" test,
1045 // since it will not discriminate anything.
1046 bool eq_test_only = (hi == lo+2 && hi->dest() == lo->dest() && mid == hi-1) || mid == lo;
1047
1048 // if there is a higher range, test for it and process it:
1049 if (mid < hi && !eq_test_only) {
1050 // two comparisons of same values--should enable 1 test for 2 branches
1051 // Use BoolTest::le instead of BoolTest::gt
1052 float cnt = sum_of_cnts(lo, mid-1);
1053 IfNode *iff_le = jump_if_fork_int(key_val, test_val, BoolTest::le, if_prob(cnt, total_cnt), if_cnt(cnt));
1054 Node *iftrue = _gvn.transform( new IfTrueNode(iff_le) );
1055 Node *iffalse = _gvn.transform( new IfFalseNode(iff_le) );
1056 { PreserveJVMState pjvms(this);
1057 set_control(iffalse);
1058 jump_switch_ranges(key_val, mid+1, hi, switch_depth+1);
1059 }
1060 set_control(iftrue);
1061 }
1062
1063 } else {
1064 // mid is a range, not a singleton, so treat mid..hi as a unit
1065 float cnt = sum_of_cnts(mid == lo ? mid+1 : mid, hi);
1066 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));
1067
1068 // if there is a higher range, test for it and process it:
1069 if (mid == hi) {
1070 jump_if_true_fork(iff_ge, mid->dest(), mid->table_index(), trim_ranges && cnt == 0);
1071 } else {
1072 Node *iftrue = _gvn.transform( new IfTrueNode(iff_ge) );
1073 Node *iffalse = _gvn.transform( new IfFalseNode(iff_ge) );
1074 { PreserveJVMState pjvms(this);
1075 set_control(iftrue);
1076 jump_switch_ranges(key_val, mid == lo ? mid+1 : mid, hi, switch_depth+1);
1077 }
1078 set_control(iffalse);
1079 }
1080 }
1081
1082 // in any case, process the lower range
1083 if (mid == lo) {
1084 if (mid->is_singleton()) {
1085 jump_switch_ranges(key_val, lo+1, hi, switch_depth+1);
1086 } else {
1087 jump_if_always_fork(lo->dest(), lo->table_index(), trim_ranges && lo->cnt() == 0);
1088 }
1089 } else {
1090 jump_switch_ranges(key_val, lo, mid-1, switch_depth+1);
1091 }
1092 }
1093
1094 // Decrease pred_count for each successor after all is done.
1095 if (switch_depth == 0) {
1096 int unique_successors = switch_block->num_successors();
1097 for (int i = 0; i < unique_successors; i++) {
1098 Block* target = switch_block->successor_at(i);
1099 // Throw away the pre-allocated path for each unique successor.
1100 target->next_path_num();
1101 }
1102 }
1103
1104#ifndef PRODUCT
1105 _max_switch_depth = MAX2(switch_depth, _max_switch_depth);
1106 if (TraceOptoParse && Verbose && WizardMode && switch_depth == 0) {
1107 SwitchRange* r;
1108 int nsing = 0;
1109 for( r = lo; r <= hi; r++ ) {
1110 if( r->is_singleton() ) nsing++;
1111 }
1112 tty->print(">>> ");
1113 _method->print_short_name();
1114 tty->print_cr(" switch decision tree");
1115 tty->print_cr(" %d ranges (%d singletons), max_depth=%d, est_depth=%d",
1116 (int) (hi-lo+1), nsing, _max_switch_depth, _est_switch_depth);
1117 if (_max_switch_depth > _est_switch_depth) {
1118 tty->print_cr("******** BAD SWITCH DEPTH ********");
1119 }
1120 tty->print(" ");
1121 for( r = lo; r <= hi; r++ ) {
1122 r->print();
1123 }
1124 tty->cr();
1125 }
1126#endif
1127}
1128
1129void Parse::modf() {
1130 Node *f2 = pop();
1131 Node *f1 = pop();
1132 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::modf_Type(),
1133 CAST_FROM_FN_PTR(address, SharedRuntime::frem),
1134 "frem", NULL, //no memory effects
1135 f1, f2);
1136 Node* res = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0));
1137
1138 push(res);
1139}
1140
1141void Parse::modd() {
1142 Node *d2 = pop_pair();
1143 Node *d1 = pop_pair();
1144 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::Math_DD_D_Type(),
1145 CAST_FROM_FN_PTR(address, SharedRuntime::drem),
1146 "drem", NULL, //no memory effects
1147 d1, top(), d2, top());
1148 Node* res_d = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0));
1149
1150#ifdef ASSERT
1151 Node* res_top = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 1));
1152 assert(res_top == top(), "second value must be top");
1153#endif
1154
1155 push_pair(res_d);
1156}
1157
1158void Parse::l2f() {
1159 Node* f2 = pop();
1160 Node* f1 = pop();
1161 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::l2f_Type(),
1162 CAST_FROM_FN_PTR(address, SharedRuntime::l2f),
1163 "l2f", NULL, //no memory effects
1164 f1, f2);
1165 Node* res = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0));
1166
1167 push(res);
1168}
1169
1170void Parse::do_irem() {
1171 // Must keep both values on the expression-stack during null-check
1172 zero_check_int(peek());
1173 // Compile-time detect of null-exception?
1174 if (stopped()) return;
1175
1176 Node* b = pop();
1177 Node* a = pop();
1178
1179 const Type *t = _gvn.type(b);
1180 if (t != Type::TOP) {
1181 const TypeInt *ti = t->is_int();
1182 if (ti->is_con()) {
1183 int divisor = ti->get_con();
1184 // check for positive power of 2
1185 if (divisor > 0 &&
1186 (divisor & ~(divisor-1)) == divisor) {
1187 // yes !
1188 Node *mask = _gvn.intcon((divisor - 1));
1189 // Sigh, must handle negative dividends
1190 Node *zero = _gvn.intcon(0);
1191 IfNode *ifff = jump_if_fork_int(a, zero, BoolTest::lt, PROB_FAIR, COUNT_UNKNOWN);
1192 Node *iff = _gvn.transform( new IfFalseNode(ifff) );
1193 Node *ift = _gvn.transform( new IfTrueNode (ifff) );
1194 Node *reg = jump_if_join(ift, iff);
1195 Node *phi = PhiNode::make(reg, NULL, TypeInt::INT);
1196 // Negative path; negate/and/negate
1197 Node *neg = _gvn.transform( new SubINode(zero, a) );
1198 Node *andn= _gvn.transform( new AndINode(neg, mask) );
1199 Node *negn= _gvn.transform( new SubINode(zero, andn) );
1200 phi->init_req(1, negn);
1201 // Fast positive case
1202 Node *andx = _gvn.transform( new AndINode(a, mask) );
1203 phi->init_req(2, andx);
1204 // Push the merge
1205 push( _gvn.transform(phi) );
1206 return;
1207 }
1208 }
1209 }
1210 // Default case
1211 push( _gvn.transform( new ModINode(control(),a,b) ) );
1212}
1213
1214// Handle jsr and jsr_w bytecode
1215void Parse::do_jsr() {
1216 assert(bc() == Bytecodes::_jsr || bc() == Bytecodes::_jsr_w, "wrong bytecode");
1217
1218 // Store information about current state, tagged with new _jsr_bci
1219 int return_bci = iter().next_bci();
1220 int jsr_bci = (bc() == Bytecodes::_jsr) ? iter().get_dest() : iter().get_far_dest();
1221
1222 // Update method data
1223 profile_taken_branch(jsr_bci);
1224
1225 // The way we do things now, there is only one successor block
1226 // for the jsr, because the target code is cloned by ciTypeFlow.
1227 Block* target = successor_for_bci(jsr_bci);
1228
1229 // What got pushed?
1230 const Type* ret_addr = target->peek();
1231 assert(ret_addr->singleton(), "must be a constant (cloned jsr body)");
1232
1233 // Effect on jsr on stack
1234 push(_gvn.makecon(ret_addr));
1235
1236 // Flow to the jsr.
1237 merge(jsr_bci);
1238}
1239
1240// Handle ret bytecode
1241void Parse::do_ret() {
1242 // Find to whom we return.
1243 assert(block()->num_successors() == 1, "a ret can only go one place now");
1244 Block* target = block()->successor_at(0);
1245 assert(!target->is_ready(), "our arrival must be expected");
1246 profile_ret(target->flow()->start());
1247 int pnum = target->next_path_num();
1248 merge_common(target, pnum);
1249}
1250
1251static bool has_injected_profile(BoolTest::mask btest, Node* test, int& taken, int& not_taken) {
1252 if (btest != BoolTest::eq && btest != BoolTest::ne) {
1253 // Only ::eq and ::ne are supported for profile injection.
1254 return false;
1255 }
1256 if (test->is_Cmp() &&
1257 test->in(1)->Opcode() == Op_ProfileBoolean) {
1258 ProfileBooleanNode* profile = (ProfileBooleanNode*)test->in(1);
1259 int false_cnt = profile->false_count();
1260 int true_cnt = profile->true_count();
1261
1262 // Counts matching depends on the actual test operation (::eq or ::ne).
1263 // No need to scale the counts because profile injection was designed
1264 // to feed exact counts into VM.
1265 taken = (btest == BoolTest::eq) ? false_cnt : true_cnt;
1266 not_taken = (btest == BoolTest::eq) ? true_cnt : false_cnt;
1267
1268 profile->consume();
1269 return true;
1270 }
1271 return false;
1272}
1273//--------------------------dynamic_branch_prediction--------------------------
1274// Try to gather dynamic branch prediction behavior. Return a probability
1275// of the branch being taken and set the "cnt" field. Returns a -1.0
1276// if we need to use static prediction for some reason.
1277float Parse::dynamic_branch_prediction(float &cnt, BoolTest::mask btest, Node* test) {
1278 ResourceMark rm;
1279
1280 cnt = COUNT_UNKNOWN;
1281
1282 int taken = 0;
1283 int not_taken = 0;
1284
1285 bool use_mdo = !has_injected_profile(btest, test, taken, not_taken);
1286
1287 if (use_mdo) {
1288 // Use MethodData information if it is available
1289 // FIXME: free the ProfileData structure
1290 ciMethodData* methodData = method()->method_data();
1291 if (!methodData->is_mature()) return PROB_UNKNOWN;
1292 ciProfileData* data = methodData->bci_to_data(bci());
1293 if (data == NULL) {
1294 return PROB_UNKNOWN;
1295 }
1296 if (!data->is_JumpData()) return PROB_UNKNOWN;
1297
1298 // get taken and not taken values
1299 taken = data->as_JumpData()->taken();
1300 not_taken = 0;
1301 if (data->is_BranchData()) {
1302 not_taken = data->as_BranchData()->not_taken();
1303 }
1304
1305 // scale the counts to be commensurate with invocation counts:
1306 taken = method()->scale_count(taken);
1307 not_taken = method()->scale_count(not_taken);
1308 }
1309
1310 // Give up if too few (or too many, in which case the sum will overflow) counts to be meaningful.
1311 // We also check that individual counters are positive first, otherwise the sum can become positive.
1312 if (taken < 0 || not_taken < 0 || taken + not_taken < 40) {
1313 if (C->log() != NULL) {
1314 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d'", iter().get_dest(), taken, not_taken);
1315 }
1316 return PROB_UNKNOWN;
1317 }
1318
1319 // Compute frequency that we arrive here
1320 float sum = taken + not_taken;
1321 // Adjust, if this block is a cloned private block but the
1322 // Jump counts are shared. Taken the private counts for
1323 // just this path instead of the shared counts.
1324 if( block()->count() > 0 )
1325 sum = block()->count();
1326 cnt = sum / FreqCountInvocations;
1327
1328 // Pin probability to sane limits
1329 float prob;
1330 if( !taken )
1331 prob = (0+PROB_MIN) / 2;
1332 else if( !not_taken )
1333 prob = (1+PROB_MAX) / 2;
1334 else { // Compute probability of true path
1335 prob = (float)taken / (float)(taken + not_taken);
1336 if (prob > PROB_MAX) prob = PROB_MAX;
1337 if (prob < PROB_MIN) prob = PROB_MIN;
1338 }
1339
1340 assert((cnt > 0.0f) && (prob > 0.0f),
1341 "Bad frequency assignment in if");
1342
1343 if (C->log() != NULL) {
1344 const char* prob_str = NULL;
1345 if (prob >= PROB_MAX) prob_str = (prob == PROB_MAX) ? "max" : "always";
1346 if (prob <= PROB_MIN) prob_str = (prob == PROB_MIN) ? "min" : "never";
1347 char prob_str_buf[30];
1348 if (prob_str == NULL) {
1349 jio_snprintf(prob_str_buf, sizeof(prob_str_buf), "%20.2f", prob);
1350 prob_str = prob_str_buf;
1351 }
1352 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d' cnt='%f' prob='%s'",
1353 iter().get_dest(), taken, not_taken, cnt, prob_str);
1354 }
1355 return prob;
1356}
1357
1358//-----------------------------branch_prediction-------------------------------
1359float Parse::branch_prediction(float& cnt,
1360 BoolTest::mask btest,
1361 int target_bci,
1362 Node* test) {
1363 float prob = dynamic_branch_prediction(cnt, btest, test);
1364 // If prob is unknown, switch to static prediction
1365 if (prob != PROB_UNKNOWN) return prob;
1366
1367 prob = PROB_FAIR; // Set default value
1368 if (btest == BoolTest::eq) // Exactly equal test?
1369 prob = PROB_STATIC_INFREQUENT; // Assume its relatively infrequent
1370 else if (btest == BoolTest::ne)
1371 prob = PROB_STATIC_FREQUENT; // Assume its relatively frequent
1372
1373 // If this is a conditional test guarding a backwards branch,
1374 // assume its a loop-back edge. Make it a likely taken branch.
1375 if (target_bci < bci()) {
1376 if (is_osr_parse()) { // Could be a hot OSR'd loop; force deopt
1377 // Since it's an OSR, we probably have profile data, but since
1378 // branch_prediction returned PROB_UNKNOWN, the counts are too small.
1379 // Let's make a special check here for completely zero counts.
1380 ciMethodData* methodData = method()->method_data();
1381 if (!methodData->is_empty()) {
1382 ciProfileData* data = methodData->bci_to_data(bci());
1383 // Only stop for truly zero counts, which mean an unknown part
1384 // of the OSR-ed method, and we want to deopt to gather more stats.
1385 // If you have ANY counts, then this loop is simply 'cold' relative
1386 // to the OSR loop.
1387 if (data == NULL ||
1388 (data->as_BranchData()->taken() + data->as_BranchData()->not_taken() == 0)) {
1389 // This is the only way to return PROB_UNKNOWN:
1390 return PROB_UNKNOWN;
1391 }
1392 }
1393 }
1394 prob = PROB_STATIC_FREQUENT; // Likely to take backwards branch
1395 }
1396
1397 assert(prob != PROB_UNKNOWN, "must have some guess at this point");
1398 return prob;
1399}
1400
1401// The magic constants are chosen so as to match the output of
1402// branch_prediction() when the profile reports a zero taken count.
1403// It is important to distinguish zero counts unambiguously, because
1404// some branches (e.g., _213_javac.Assembler.eliminate) validly produce
1405// very small but nonzero probabilities, which if confused with zero
1406// counts would keep the program recompiling indefinitely.
1407bool Parse::seems_never_taken(float prob) const {
1408 return prob < PROB_MIN;
1409}
1410
1411// True if the comparison seems to be the kind that will not change its
1412// statistics from true to false. See comments in adjust_map_after_if.
1413// This question is only asked along paths which are already
1414// classifed as untaken (by seems_never_taken), so really,
1415// if a path is never taken, its controlling comparison is
1416// already acting in a stable fashion. If the comparison
1417// seems stable, we will put an expensive uncommon trap
1418// on the untaken path.
1419bool Parse::seems_stable_comparison() const {
1420 if (C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if)) {
1421 return false;
1422 }
1423 return true;
1424}
1425
1426//-------------------------------repush_if_args--------------------------------
1427// Push arguments of an "if" bytecode back onto the stack by adjusting _sp.
1428inline int Parse::repush_if_args() {
1429 if (PrintOpto && WizardMode) {
1430 tty->print("defending against excessive implicit null exceptions on %s @%d in ",
1431 Bytecodes::name(iter().cur_bc()), iter().cur_bci());
1432 method()->print_name(); tty->cr();
1433 }
1434 int bc_depth = - Bytecodes::depth(iter().cur_bc());
1435 assert(bc_depth == 1 || bc_depth == 2, "only two kinds of branches");
1436 DEBUG_ONLY(sync_jvms()); // argument(n) requires a synced jvms
1437 assert(argument(0) != NULL, "must exist");
1438 assert(bc_depth == 1 || argument(1) != NULL, "two must exist");
1439 inc_sp(bc_depth);
1440 return bc_depth;
1441}
1442
1443//----------------------------------do_ifnull----------------------------------
1444void Parse::do_ifnull(BoolTest::mask btest, Node *c) {
1445 int target_bci = iter().get_dest();
1446
1447 Block* branch_block = successor_for_bci(target_bci);
1448 Block* next_block = successor_for_bci(iter().next_bci());
1449
1450 float cnt;
1451 float prob = branch_prediction(cnt, btest, target_bci, c);
1452 if (prob == PROB_UNKNOWN) {
1453 // (An earlier version of do_ifnull omitted this trap for OSR methods.)
1454 if (PrintOpto && Verbose) {
1455 tty->print_cr("Never-taken edge stops compilation at bci %d", bci());
1456 }
1457 repush_if_args(); // to gather stats on loop
1458 // We need to mark this branch as taken so that if we recompile we will
1459 // see that it is possible. In the tiered system the interpreter doesn't
1460 // do profiling and by the time we get to the lower tier from the interpreter
1461 // the path may be cold again. Make sure it doesn't look untaken
1462 profile_taken_branch(target_bci, !ProfileInterpreter);
1463 uncommon_trap(Deoptimization::Reason_unreached,
1464 Deoptimization::Action_reinterpret,
1465 NULL, "cold");
1466 if (C->eliminate_boxing()) {
1467 // Mark the successor blocks as parsed
1468 branch_block->next_path_num();
1469 next_block->next_path_num();
1470 }
1471 return;
1472 }
1473
1474 NOT_PRODUCT(explicit_null_checks_inserted++);
1475
1476 // Generate real control flow
1477 Node *tst = _gvn.transform( new BoolNode( c, btest ) );
1478
1479 // Sanity check the probability value
1480 assert(prob > 0.0f,"Bad probability in Parser");
1481 // Need xform to put node in hash table
1482 IfNode *iff = create_and_xform_if( control(), tst, prob, cnt );
1483 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser");
1484 // True branch
1485 { PreserveJVMState pjvms(this);
1486 Node* iftrue = _gvn.transform( new IfTrueNode (iff) );
1487 set_control(iftrue);
1488
1489 if (stopped()) { // Path is dead?
1490 NOT_PRODUCT(explicit_null_checks_elided++);
1491 if (C->eliminate_boxing()) {
1492 // Mark the successor block as parsed
1493 branch_block->next_path_num();
1494 }
1495 } else { // Path is live.
1496 // Update method data
1497 profile_taken_branch(target_bci);
1498 adjust_map_after_if(btest, c, prob, branch_block, next_block);
1499 if (!stopped()) {
1500 merge(target_bci);
1501 }
1502 }
1503 }
1504
1505 // False branch
1506 Node* iffalse = _gvn.transform( new IfFalseNode(iff) );
1507 set_control(iffalse);
1508
1509 if (stopped()) { // Path is dead?
1510 NOT_PRODUCT(explicit_null_checks_elided++);
1511 if (C->eliminate_boxing()) {
1512 // Mark the successor block as parsed
1513 next_block->next_path_num();
1514 }
1515 } else { // Path is live.
1516 // Update method data
1517 profile_not_taken_branch();
1518 adjust_map_after_if(BoolTest(btest).negate(), c, 1.0-prob,
1519 next_block, branch_block);
1520 }
1521}
1522
1523//------------------------------------do_if------------------------------------
1524void Parse::do_if(BoolTest::mask btest, Node* c) {
1525 int target_bci = iter().get_dest();
1526
1527 Block* branch_block = successor_for_bci(target_bci);
1528 Block* next_block = successor_for_bci(iter().next_bci());
1529
1530 float cnt;
1531 float prob = branch_prediction(cnt, btest, target_bci, c);
1532 float untaken_prob = 1.0 - prob;
1533
1534 if (prob == PROB_UNKNOWN) {
1535 if (PrintOpto && Verbose) {
1536 tty->print_cr("Never-taken edge stops compilation at bci %d", bci());
1537 }
1538 repush_if_args(); // to gather stats on loop
1539 // We need to mark this branch as taken so that if we recompile we will
1540 // see that it is possible. In the tiered system the interpreter doesn't
1541 // do profiling and by the time we get to the lower tier from the interpreter
1542 // the path may be cold again. Make sure it doesn't look untaken
1543 profile_taken_branch(target_bci, !ProfileInterpreter);
1544 uncommon_trap(Deoptimization::Reason_unreached,
1545 Deoptimization::Action_reinterpret,
1546 NULL, "cold");
1547 if (C->eliminate_boxing()) {
1548 // Mark the successor blocks as parsed
1549 branch_block->next_path_num();
1550 next_block->next_path_num();
1551 }
1552 return;
1553 }
1554
1555 // Sanity check the probability value
1556 assert(0.0f < prob && prob < 1.0f,"Bad probability in Parser");
1557
1558 bool taken_if_true = true;
1559 // Convert BoolTest to canonical form:
1560 if (!BoolTest(btest).is_canonical()) {
1561 btest = BoolTest(btest).negate();
1562 taken_if_true = false;
1563 // prob is NOT updated here; it remains the probability of the taken
1564 // path (as opposed to the prob of the path guarded by an 'IfTrueNode').
1565 }
1566 assert(btest != BoolTest::eq, "!= is the only canonical exact test");
1567
1568 Node* tst0 = new BoolNode(c, btest);
1569 Node* tst = _gvn.transform(tst0);
1570 BoolTest::mask taken_btest = BoolTest::illegal;
1571 BoolTest::mask untaken_btest = BoolTest::illegal;
1572
1573 if (tst->is_Bool()) {
1574 // Refresh c from the transformed bool node, since it may be
1575 // simpler than the original c. Also re-canonicalize btest.
1576 // This wins when (Bool ne (Conv2B p) 0) => (Bool ne (CmpP p NULL)).
1577 // That can arise from statements like: if (x instanceof C) ...
1578 if (tst != tst0) {
1579 // Canonicalize one more time since transform can change it.
1580 btest = tst->as_Bool()->_test._test;
1581 if (!BoolTest(btest).is_canonical()) {
1582 // Reverse edges one more time...
1583 tst = _gvn.transform( tst->as_Bool()->negate(&_gvn) );
1584 btest = tst->as_Bool()->_test._test;
1585 assert(BoolTest(btest).is_canonical(), "sanity");
1586 taken_if_true = !taken_if_true;
1587 }
1588 c = tst->in(1);
1589 }
1590 BoolTest::mask neg_btest = BoolTest(btest).negate();
1591 taken_btest = taken_if_true ? btest : neg_btest;
1592 untaken_btest = taken_if_true ? neg_btest : btest;
1593 }
1594
1595 // Generate real control flow
1596 float true_prob = (taken_if_true ? prob : untaken_prob);
1597 IfNode* iff = create_and_map_if(control(), tst, true_prob, cnt);
1598 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser");
1599 Node* taken_branch = new IfTrueNode(iff);
1600 Node* untaken_branch = new IfFalseNode(iff);
1601 if (!taken_if_true) { // Finish conversion to canonical form
1602 Node* tmp = taken_branch;
1603 taken_branch = untaken_branch;
1604 untaken_branch = tmp;
1605 }
1606
1607 // Branch is taken:
1608 { PreserveJVMState pjvms(this);
1609 taken_branch = _gvn.transform(taken_branch);
1610 set_control(taken_branch);
1611
1612 if (stopped()) {
1613 if (C->eliminate_boxing()) {
1614 // Mark the successor block as parsed
1615 branch_block->next_path_num();
1616 }
1617 } else {
1618 // Update method data
1619 profile_taken_branch(target_bci);
1620 adjust_map_after_if(taken_btest, c, prob, branch_block, next_block);
1621 if (!stopped()) {
1622 merge(target_bci);
1623 }
1624 }
1625 }
1626
1627 untaken_branch = _gvn.transform(untaken_branch);
1628 set_control(untaken_branch);
1629
1630 // Branch not taken.
1631 if (stopped()) {
1632 if (C->eliminate_boxing()) {
1633 // Mark the successor block as parsed
1634 next_block->next_path_num();
1635 }
1636 } else {
1637 // Update method data
1638 profile_not_taken_branch();
1639 adjust_map_after_if(untaken_btest, c, untaken_prob,
1640 next_block, branch_block);
1641 }
1642}
1643
1644bool Parse::path_is_suitable_for_uncommon_trap(float prob) const {
1645 // Don't want to speculate on uncommon traps when running with -Xcomp
1646 if (!UseInterpreter) {
1647 return false;
1648 }
1649 return (seems_never_taken(prob) && seems_stable_comparison());
1650}
1651
1652void Parse::maybe_add_predicate_after_if(Block* path) {
1653 if (path->is_SEL_head() && path->preds_parsed() == 0) {
1654 // Add predicates at bci of if dominating the loop so traps can be
1655 // recorded on the if's profile data
1656 int bc_depth = repush_if_args();
1657 add_predicate();
1658 dec_sp(bc_depth);
1659 path->set_has_predicates();
1660 }
1661}
1662
1663
1664//----------------------------adjust_map_after_if------------------------------
1665// Adjust the JVM state to reflect the result of taking this path.
1666// Basically, it means inspecting the CmpNode controlling this
1667// branch, seeing how it constrains a tested value, and then
1668// deciding if it's worth our while to encode this constraint
1669// as graph nodes in the current abstract interpretation map.
1670void Parse::adjust_map_after_if(BoolTest::mask btest, Node* c, float prob,
1671 Block* path, Block* other_path) {
1672 if (!c->is_Cmp()) {
1673 maybe_add_predicate_after_if(path);
1674 return;
1675 }
1676
1677 if (stopped() || btest == BoolTest::illegal) {
1678 return; // nothing to do
1679 }
1680
1681 bool is_fallthrough = (path == successor_for_bci(iter().next_bci()));
1682
1683 if (path_is_suitable_for_uncommon_trap(prob)) {
1684 repush_if_args();
1685 uncommon_trap(Deoptimization::Reason_unstable_if,
1686 Deoptimization::Action_reinterpret,
1687 NULL,
1688 (is_fallthrough ? "taken always" : "taken never"));
1689 return;
1690 }
1691
1692 Node* val = c->in(1);
1693 Node* con = c->in(2);
1694 const Type* tcon = _gvn.type(con);
1695 const Type* tval = _gvn.type(val);
1696 bool have_con = tcon->singleton();
1697 if (tval->singleton()) {
1698 if (!have_con) {
1699 // Swap, so constant is in con.
1700 con = val;
1701 tcon = tval;
1702 val = c->in(2);
1703 tval = _gvn.type(val);
1704 btest = BoolTest(btest).commute();
1705 have_con = true;
1706 } else {
1707 // Do we have two constants? Then leave well enough alone.
1708 have_con = false;
1709 }
1710 }
1711 if (!have_con) { // remaining adjustments need a con
1712 maybe_add_predicate_after_if(path);
1713 return;
1714 }
1715
1716 sharpen_type_after_if(btest, con, tcon, val, tval);
1717 maybe_add_predicate_after_if(path);
1718}
1719
1720
1721static Node* extract_obj_from_klass_load(PhaseGVN* gvn, Node* n) {
1722 Node* ldk;
1723 if (n->is_DecodeNKlass()) {
1724 if (n->in(1)->Opcode() != Op_LoadNKlass) {
1725 return NULL;
1726 } else {
1727 ldk = n->in(1);
1728 }
1729 } else if (n->Opcode() != Op_LoadKlass) {
1730 return NULL;
1731 } else {
1732 ldk = n;
1733 }
1734 assert(ldk != NULL && ldk->is_Load(), "should have found a LoadKlass or LoadNKlass node");
1735
1736 Node* adr = ldk->in(MemNode::Address);
1737 intptr_t off = 0;
1738 Node* obj = AddPNode::Ideal_base_and_offset(adr, gvn, off);
1739 if (obj == NULL || off != oopDesc::klass_offset_in_bytes()) // loading oopDesc::_klass?
1740 return NULL;
1741 const TypePtr* tp = gvn->type(obj)->is_ptr();
1742 if (tp == NULL || !(tp->isa_instptr() || tp->isa_aryptr())) // is obj a Java object ptr?
1743 return NULL;
1744
1745 return obj;
1746}
1747
1748void Parse::sharpen_type_after_if(BoolTest::mask btest,
1749 Node* con, const Type* tcon,
1750 Node* val, const Type* tval) {
1751 // Look for opportunities to sharpen the type of a node
1752 // whose klass is compared with a constant klass.
1753 if (btest == BoolTest::eq && tcon->isa_klassptr()) {
1754 Node* obj = extract_obj_from_klass_load(&_gvn, val);
1755 const TypeOopPtr* con_type = tcon->isa_klassptr()->as_instance_type();
1756 if (obj != NULL && (con_type->isa_instptr() || con_type->isa_aryptr())) {
1757 // Found:
1758 // Bool(CmpP(LoadKlass(obj._klass), ConP(Foo.klass)), [eq])
1759 // or the narrowOop equivalent.
1760 const Type* obj_type = _gvn.type(obj);
1761 const TypeOopPtr* tboth = obj_type->join_speculative(con_type)->isa_oopptr();
1762 if (tboth != NULL && tboth->klass_is_exact() && tboth != obj_type &&
1763 tboth->higher_equal(obj_type)) {
1764 // obj has to be of the exact type Foo if the CmpP succeeds.
1765 int obj_in_map = map()->find_edge(obj);
1766 JVMState* jvms = this->jvms();
1767 if (obj_in_map >= 0 &&
1768 (jvms->is_loc(obj_in_map) || jvms->is_stk(obj_in_map))) {
1769 TypeNode* ccast = new CheckCastPPNode(control(), obj, tboth);
1770 const Type* tcc = ccast->as_Type()->type();
1771 assert(tcc != obj_type && tcc->higher_equal(obj_type), "must improve");
1772 // Delay transform() call to allow recovery of pre-cast value
1773 // at the control merge.
1774 _gvn.set_type_bottom(ccast);
1775 record_for_igvn(ccast);
1776 // Here's the payoff.
1777 replace_in_map(obj, ccast);
1778 }
1779 }
1780 }
1781 }
1782
1783 int val_in_map = map()->find_edge(val);
1784 if (val_in_map < 0) return; // replace_in_map would be useless
1785 {
1786 JVMState* jvms = this->jvms();
1787 if (!(jvms->is_loc(val_in_map) ||
1788 jvms->is_stk(val_in_map)))
1789 return; // again, it would be useless
1790 }
1791
1792 // Check for a comparison to a constant, and "know" that the compared
1793 // value is constrained on this path.
1794 assert(tcon->singleton(), "");
1795 ConstraintCastNode* ccast = NULL;
1796 Node* cast = NULL;
1797
1798 switch (btest) {
1799 case BoolTest::eq: // Constant test?
1800 {
1801 const Type* tboth = tcon->join_speculative(tval);
1802 if (tboth == tval) break; // Nothing to gain.
1803 if (tcon->isa_int()) {
1804 ccast = new CastIINode(val, tboth);
1805 } else if (tcon == TypePtr::NULL_PTR) {
1806 // Cast to null, but keep the pointer identity temporarily live.
1807 ccast = new CastPPNode(val, tboth);
1808 } else {
1809 const TypeF* tf = tcon->isa_float_constant();
1810 const TypeD* td = tcon->isa_double_constant();
1811 // Exclude tests vs float/double 0 as these could be
1812 // either +0 or -0. Just because you are equal to +0
1813 // doesn't mean you ARE +0!
1814 // Note, following code also replaces Long and Oop values.
1815 if ((!tf || tf->_f != 0.0) &&
1816 (!td || td->_d != 0.0))
1817 cast = con; // Replace non-constant val by con.
1818 }
1819 }
1820 break;
1821
1822 case BoolTest::ne:
1823 if (tcon == TypePtr::NULL_PTR) {
1824 cast = cast_not_null(val, false);
1825 }
1826 break;
1827
1828 default:
1829 // (At this point we could record int range types with CastII.)
1830 break;
1831 }
1832
1833 if (ccast != NULL) {
1834 const Type* tcc = ccast->as_Type()->type();
1835 assert(tcc != tval && tcc->higher_equal(tval), "must improve");
1836 // Delay transform() call to allow recovery of pre-cast value
1837 // at the control merge.
1838 ccast->set_req(0, control());
1839 _gvn.set_type_bottom(ccast);
1840 record_for_igvn(ccast);
1841 cast = ccast;
1842 }
1843
1844 if (cast != NULL) { // Here's the payoff.
1845 replace_in_map(val, cast);
1846 }
1847}
1848
1849/**
1850 * Use speculative type to optimize CmpP node: if comparison is
1851 * against the low level class, cast the object to the speculative
1852 * type if any. CmpP should then go away.
1853 *
1854 * @param c expected CmpP node
1855 * @return result of CmpP on object casted to speculative type
1856 *
1857 */
1858Node* Parse::optimize_cmp_with_klass(Node* c) {
1859 // If this is transformed by the _gvn to a comparison with the low
1860 // level klass then we may be able to use speculation
1861 if (c->Opcode() == Op_CmpP &&
1862 (c->in(1)->Opcode() == Op_LoadKlass || c->in(1)->Opcode() == Op_DecodeNKlass) &&
1863 c->in(2)->is_Con()) {
1864 Node* load_klass = NULL;
1865 Node* decode = NULL;
1866 if (c->in(1)->Opcode() == Op_DecodeNKlass) {
1867 decode = c->in(1);
1868 load_klass = c->in(1)->in(1);
1869 } else {
1870 load_klass = c->in(1);
1871 }
1872 if (load_klass->in(2)->is_AddP()) {
1873 Node* addp = load_klass->in(2);
1874 Node* obj = addp->in(AddPNode::Address);
1875 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
1876 if (obj_type->speculative_type_not_null() != NULL) {
1877 ciKlass* k = obj_type->speculative_type();
1878 inc_sp(2);
1879 obj = maybe_cast_profiled_obj(obj, k);
1880 dec_sp(2);
1881 // Make the CmpP use the casted obj
1882 addp = basic_plus_adr(obj, addp->in(AddPNode::Offset));
1883 load_klass = load_klass->clone();
1884 load_klass->set_req(2, addp);
1885 load_klass = _gvn.transform(load_klass);
1886 if (decode != NULL) {
1887 decode = decode->clone();
1888 decode->set_req(1, load_klass);
1889 load_klass = _gvn.transform(decode);
1890 }
1891 c = c->clone();
1892 c->set_req(1, load_klass);
1893 c = _gvn.transform(c);
1894 }
1895 }
1896 }
1897 return c;
1898}
1899
1900//------------------------------do_one_bytecode--------------------------------
1901// Parse this bytecode, and alter the Parsers JVM->Node mapping
1902void Parse::do_one_bytecode() {
1903 Node *a, *b, *c, *d; // Handy temps
1904 BoolTest::mask btest;
1905 int i;
1906
1907 assert(!has_exceptions(), "bytecode entry state must be clear of throws");
1908
1909 if (C->check_node_count(NodeLimitFudgeFactor * 5,
1910 "out of nodes parsing method")) {
1911 return;
1912 }
1913
1914#ifdef ASSERT
1915 // for setting breakpoints
1916 if (TraceOptoParse) {
1917 tty->print(" @");
1918 dump_bci(bci());
1919 tty->cr();
1920 }
1921#endif
1922
1923 switch (bc()) {
1924 case Bytecodes::_nop:
1925 // do nothing
1926 break;
1927 case Bytecodes::_lconst_0:
1928 push_pair(longcon(0));
1929 break;
1930
1931 case Bytecodes::_lconst_1:
1932 push_pair(longcon(1));
1933 break;
1934
1935 case Bytecodes::_fconst_0:
1936 push(zerocon(T_FLOAT));
1937 break;
1938
1939 case Bytecodes::_fconst_1:
1940 push(makecon(TypeF::ONE));
1941 break;
1942
1943 case Bytecodes::_fconst_2:
1944 push(makecon(TypeF::make(2.0f)));
1945 break;
1946
1947 case Bytecodes::_dconst_0:
1948 push_pair(zerocon(T_DOUBLE));
1949 break;
1950
1951 case Bytecodes::_dconst_1:
1952 push_pair(makecon(TypeD::ONE));
1953 break;
1954
1955 case Bytecodes::_iconst_m1:push(intcon(-1)); break;
1956 case Bytecodes::_iconst_0: push(intcon( 0)); break;
1957 case Bytecodes::_iconst_1: push(intcon( 1)); break;
1958 case Bytecodes::_iconst_2: push(intcon( 2)); break;
1959 case Bytecodes::_iconst_3: push(intcon( 3)); break;
1960 case Bytecodes::_iconst_4: push(intcon( 4)); break;
1961 case Bytecodes::_iconst_5: push(intcon( 5)); break;
1962 case Bytecodes::_bipush: push(intcon(iter().get_constant_u1())); break;
1963 case Bytecodes::_sipush: push(intcon(iter().get_constant_u2())); break;
1964 case Bytecodes::_aconst_null: push(null()); break;
1965 case Bytecodes::_ldc:
1966 case Bytecodes::_ldc_w:
1967 case Bytecodes::_ldc2_w:
1968 // If the constant is unresolved, run this BC once in the interpreter.
1969 {
1970 ciConstant constant = iter().get_constant();
1971 if (!constant.is_valid() ||
1972 (constant.basic_type() == T_OBJECT &&
1973 !constant.as_object()->is_loaded())) {
1974 int index = iter().get_constant_pool_index();
1975 constantTag tag = iter().get_constant_pool_tag(index);
1976 uncommon_trap(Deoptimization::make_trap_request
1977 (Deoptimization::Reason_unloaded,
1978 Deoptimization::Action_reinterpret,
1979 index),
1980 NULL, tag.internal_name());
1981 break;
1982 }
1983 assert(constant.basic_type() != T_OBJECT || constant.as_object()->is_instance(),
1984 "must be java_mirror of klass");
1985 const Type* con_type = Type::make_from_constant(constant);
1986 if (con_type != NULL) {
1987 push_node(con_type->basic_type(), makecon(con_type));
1988 }
1989 }
1990
1991 break;
1992
1993 case Bytecodes::_aload_0:
1994 push( local(0) );
1995 break;
1996 case Bytecodes::_aload_1:
1997 push( local(1) );
1998 break;
1999 case Bytecodes::_aload_2:
2000 push( local(2) );
2001 break;
2002 case Bytecodes::_aload_3:
2003 push( local(3) );
2004 break;
2005 case Bytecodes::_aload:
2006 push( local(iter().get_index()) );
2007 break;
2008
2009 case Bytecodes::_fload_0:
2010 case Bytecodes::_iload_0:
2011 push( local(0) );
2012 break;
2013 case Bytecodes::_fload_1:
2014 case Bytecodes::_iload_1:
2015 push( local(1) );
2016 break;
2017 case Bytecodes::_fload_2:
2018 case Bytecodes::_iload_2:
2019 push( local(2) );
2020 break;
2021 case Bytecodes::_fload_3:
2022 case Bytecodes::_iload_3:
2023 push( local(3) );
2024 break;
2025 case Bytecodes::_fload:
2026 case Bytecodes::_iload:
2027 push( local(iter().get_index()) );
2028 break;
2029 case Bytecodes::_lload_0:
2030 push_pair_local( 0 );
2031 break;
2032 case Bytecodes::_lload_1:
2033 push_pair_local( 1 );
2034 break;
2035 case Bytecodes::_lload_2:
2036 push_pair_local( 2 );
2037 break;
2038 case Bytecodes::_lload_3:
2039 push_pair_local( 3 );
2040 break;
2041 case Bytecodes::_lload:
2042 push_pair_local( iter().get_index() );
2043 break;
2044
2045 case Bytecodes::_dload_0:
2046 push_pair_local(0);
2047 break;
2048 case Bytecodes::_dload_1:
2049 push_pair_local(1);
2050 break;
2051 case Bytecodes::_dload_2:
2052 push_pair_local(2);
2053 break;
2054 case Bytecodes::_dload_3:
2055 push_pair_local(3);
2056 break;
2057 case Bytecodes::_dload:
2058 push_pair_local(iter().get_index());
2059 break;
2060 case Bytecodes::_fstore_0:
2061 case Bytecodes::_istore_0:
2062 case Bytecodes::_astore_0:
2063 set_local( 0, pop() );
2064 break;
2065 case Bytecodes::_fstore_1:
2066 case Bytecodes::_istore_1:
2067 case Bytecodes::_astore_1:
2068 set_local( 1, pop() );
2069 break;
2070 case Bytecodes::_fstore_2:
2071 case Bytecodes::_istore_2:
2072 case Bytecodes::_astore_2:
2073 set_local( 2, pop() );
2074 break;
2075 case Bytecodes::_fstore_3:
2076 case Bytecodes::_istore_3:
2077 case Bytecodes::_astore_3:
2078 set_local( 3, pop() );
2079 break;
2080 case Bytecodes::_fstore:
2081 case Bytecodes::_istore:
2082 case Bytecodes::_astore:
2083 set_local( iter().get_index(), pop() );
2084 break;
2085 // long stores
2086 case Bytecodes::_lstore_0:
2087 set_pair_local( 0, pop_pair() );
2088 break;
2089 case Bytecodes::_lstore_1:
2090 set_pair_local( 1, pop_pair() );
2091 break;
2092 case Bytecodes::_lstore_2:
2093 set_pair_local( 2, pop_pair() );
2094 break;
2095 case Bytecodes::_lstore_3:
2096 set_pair_local( 3, pop_pair() );
2097 break;
2098 case Bytecodes::_lstore:
2099 set_pair_local( iter().get_index(), pop_pair() );
2100 break;
2101
2102 // double stores
2103 case Bytecodes::_dstore_0:
2104 set_pair_local( 0, dstore_rounding(pop_pair()) );
2105 break;
2106 case Bytecodes::_dstore_1:
2107 set_pair_local( 1, dstore_rounding(pop_pair()) );
2108 break;
2109 case Bytecodes::_dstore_2:
2110 set_pair_local( 2, dstore_rounding(pop_pair()) );
2111 break;
2112 case Bytecodes::_dstore_3:
2113 set_pair_local( 3, dstore_rounding(pop_pair()) );
2114 break;
2115 case Bytecodes::_dstore:
2116 set_pair_local( iter().get_index(), dstore_rounding(pop_pair()) );
2117 break;
2118
2119 case Bytecodes::_pop: dec_sp(1); break;
2120 case Bytecodes::_pop2: dec_sp(2); break;
2121 case Bytecodes::_swap:
2122 a = pop();
2123 b = pop();
2124 push(a);
2125 push(b);
2126 break;
2127 case Bytecodes::_dup:
2128 a = pop();
2129 push(a);
2130 push(a);
2131 break;
2132 case Bytecodes::_dup_x1:
2133 a = pop();
2134 b = pop();
2135 push( a );
2136 push( b );
2137 push( a );
2138 break;
2139 case Bytecodes::_dup_x2:
2140 a = pop();
2141 b = pop();
2142 c = pop();
2143 push( a );
2144 push( c );
2145 push( b );
2146 push( a );
2147 break;
2148 case Bytecodes::_dup2:
2149 a = pop();
2150 b = pop();
2151 push( b );
2152 push( a );
2153 push( b );
2154 push( a );
2155 break;
2156
2157 case Bytecodes::_dup2_x1:
2158 // before: .. c, b, a
2159 // after: .. b, a, c, b, a
2160 // not tested
2161 a = pop();
2162 b = pop();
2163 c = pop();
2164 push( b );
2165 push( a );
2166 push( c );
2167 push( b );
2168 push( a );
2169 break;
2170 case Bytecodes::_dup2_x2:
2171 // before: .. d, c, b, a
2172 // after: .. b, a, d, c, b, a
2173 // not tested
2174 a = pop();
2175 b = pop();
2176 c = pop();
2177 d = pop();
2178 push( b );
2179 push( a );
2180 push( d );
2181 push( c );
2182 push( b );
2183 push( a );
2184 break;
2185
2186 case Bytecodes::_arraylength: {
2187 // Must do null-check with value on expression stack
2188 Node *ary = null_check(peek(), T_ARRAY);
2189 // Compile-time detect of null-exception?
2190 if (stopped()) return;
2191 a = pop();
2192 push(load_array_length(a));
2193 break;
2194 }
2195
2196 case Bytecodes::_baload: array_load(T_BYTE); break;
2197 case Bytecodes::_caload: array_load(T_CHAR); break;
2198 case Bytecodes::_iaload: array_load(T_INT); break;
2199 case Bytecodes::_saload: array_load(T_SHORT); break;
2200 case Bytecodes::_faload: array_load(T_FLOAT); break;
2201 case Bytecodes::_aaload: array_load(T_OBJECT); break;
2202 case Bytecodes::_laload: array_load(T_LONG); break;
2203 case Bytecodes::_daload: array_load(T_DOUBLE); break;
2204 case Bytecodes::_bastore: array_store(T_BYTE); break;
2205 case Bytecodes::_castore: array_store(T_CHAR); break;
2206 case Bytecodes::_iastore: array_store(T_INT); break;
2207 case Bytecodes::_sastore: array_store(T_SHORT); break;
2208 case Bytecodes::_fastore: array_store(T_FLOAT); break;
2209 case Bytecodes::_aastore: array_store(T_OBJECT); break;
2210 case Bytecodes::_lastore: array_store(T_LONG); break;
2211 case Bytecodes::_dastore: array_store(T_DOUBLE); break;
2212
2213 case Bytecodes::_getfield:
2214 do_getfield();
2215 break;
2216
2217 case Bytecodes::_getstatic:
2218 do_getstatic();
2219 break;
2220
2221 case Bytecodes::_putfield:
2222 do_putfield();
2223 break;
2224
2225 case Bytecodes::_putstatic:
2226 do_putstatic();
2227 break;
2228
2229 case Bytecodes::_irem:
2230 do_irem();
2231 break;
2232 case Bytecodes::_idiv:
2233 // Must keep both values on the expression-stack during null-check
2234 zero_check_int(peek());
2235 // Compile-time detect of null-exception?
2236 if (stopped()) return;
2237 b = pop();
2238 a = pop();
2239 push( _gvn.transform( new DivINode(control(),a,b) ) );
2240 break;
2241 case Bytecodes::_imul:
2242 b = pop(); a = pop();
2243 push( _gvn.transform( new MulINode(a,b) ) );
2244 break;
2245 case Bytecodes::_iadd:
2246 b = pop(); a = pop();
2247 push( _gvn.transform( new AddINode(a,b) ) );
2248 break;
2249 case Bytecodes::_ineg:
2250 a = pop();
2251 push( _gvn.transform( new SubINode(_gvn.intcon(0),a)) );
2252 break;
2253 case Bytecodes::_isub:
2254 b = pop(); a = pop();
2255 push( _gvn.transform( new SubINode(a,b) ) );
2256 break;
2257 case Bytecodes::_iand:
2258 b = pop(); a = pop();
2259 push( _gvn.transform( new AndINode(a,b) ) );
2260 break;
2261 case Bytecodes::_ior:
2262 b = pop(); a = pop();
2263 push( _gvn.transform( new OrINode(a,b) ) );
2264 break;
2265 case Bytecodes::_ixor:
2266 b = pop(); a = pop();
2267 push( _gvn.transform( new XorINode(a,b) ) );
2268 break;
2269 case Bytecodes::_ishl:
2270 b = pop(); a = pop();
2271 push( _gvn.transform( new LShiftINode(a,b) ) );
2272 break;
2273 case Bytecodes::_ishr:
2274 b = pop(); a = pop();
2275 push( _gvn.transform( new RShiftINode(a,b) ) );
2276 break;
2277 case Bytecodes::_iushr:
2278 b = pop(); a = pop();
2279 push( _gvn.transform( new URShiftINode(a,b) ) );
2280 break;
2281
2282 case Bytecodes::_fneg:
2283 a = pop();
2284 b = _gvn.transform(new NegFNode (a));
2285 push(b);
2286 break;
2287
2288 case Bytecodes::_fsub:
2289 b = pop();
2290 a = pop();
2291 c = _gvn.transform( new SubFNode(a,b) );
2292 d = precision_rounding(c);
2293 push( d );
2294 break;
2295
2296 case Bytecodes::_fadd:
2297 b = pop();
2298 a = pop();
2299 c = _gvn.transform( new AddFNode(a,b) );
2300 d = precision_rounding(c);
2301 push( d );
2302 break;
2303
2304 case Bytecodes::_fmul:
2305 b = pop();
2306 a = pop();
2307 c = _gvn.transform( new MulFNode(a,b) );
2308 d = precision_rounding(c);
2309 push( d );
2310 break;
2311
2312 case Bytecodes::_fdiv:
2313 b = pop();
2314 a = pop();
2315 c = _gvn.transform( new DivFNode(0,a,b) );
2316 d = precision_rounding(c);
2317 push( d );
2318 break;
2319
2320 case Bytecodes::_frem:
2321 if (Matcher::has_match_rule(Op_ModF)) {
2322 // Generate a ModF node.
2323 b = pop();
2324 a = pop();
2325 c = _gvn.transform( new ModFNode(0,a,b) );
2326 d = precision_rounding(c);
2327 push( d );
2328 }
2329 else {
2330 // Generate a call.
2331 modf();
2332 }
2333 break;
2334
2335 case Bytecodes::_fcmpl:
2336 b = pop();
2337 a = pop();
2338 c = _gvn.transform( new CmpF3Node( a, b));
2339 push(c);
2340 break;
2341 case Bytecodes::_fcmpg:
2342 b = pop();
2343 a = pop();
2344
2345 // Same as fcmpl but need to flip the unordered case. Swap the inputs,
2346 // which negates the result sign except for unordered. Flip the unordered
2347 // as well by using CmpF3 which implements unordered-lesser instead of
2348 // unordered-greater semantics. Finally, commute the result bits. Result
2349 // is same as using a CmpF3Greater except we did it with CmpF3 alone.
2350 c = _gvn.transform( new CmpF3Node( b, a));
2351 c = _gvn.transform( new SubINode(_gvn.intcon(0),c) );
2352 push(c);
2353 break;
2354
2355 case Bytecodes::_f2i:
2356 a = pop();
2357 push(_gvn.transform(new ConvF2INode(a)));
2358 break;
2359
2360 case Bytecodes::_d2i:
2361 a = pop_pair();
2362 b = _gvn.transform(new ConvD2INode(a));
2363 push( b );
2364 break;
2365
2366 case Bytecodes::_f2d:
2367 a = pop();
2368 b = _gvn.transform( new ConvF2DNode(a));
2369 push_pair( b );
2370 break;
2371
2372 case Bytecodes::_d2f:
2373 a = pop_pair();
2374 b = _gvn.transform( new ConvD2FNode(a));
2375 // This breaks _227_mtrt (speed & correctness) and _222_mpegaudio (speed)
2376 //b = _gvn.transform(new RoundFloatNode(0, b) );
2377 push( b );
2378 break;
2379
2380 case Bytecodes::_l2f:
2381 if (Matcher::convL2FSupported()) {
2382 a = pop_pair();
2383 b = _gvn.transform( new ConvL2FNode(a));
2384 // For i486.ad, FILD doesn't restrict precision to 24 or 53 bits.
2385 // Rather than storing the result into an FP register then pushing
2386 // out to memory to round, the machine instruction that implements
2387 // ConvL2D is responsible for rounding.
2388 // c = precision_rounding(b);
2389 c = _gvn.transform(b);
2390 push(c);
2391 } else {
2392 l2f();
2393 }
2394 break;
2395
2396 case Bytecodes::_l2d:
2397 a = pop_pair();
2398 b = _gvn.transform( new ConvL2DNode(a));
2399 // For i486.ad, rounding is always necessary (see _l2f above).
2400 // c = dprecision_rounding(b);
2401 c = _gvn.transform(b);
2402 push_pair(c);
2403 break;
2404
2405 case Bytecodes::_f2l:
2406 a = pop();
2407 b = _gvn.transform( new ConvF2LNode(a));
2408 push_pair(b);
2409 break;
2410
2411 case Bytecodes::_d2l:
2412 a = pop_pair();
2413 b = _gvn.transform( new ConvD2LNode(a));
2414 push_pair(b);
2415 break;
2416
2417 case Bytecodes::_dsub:
2418 b = pop_pair();
2419 a = pop_pair();
2420 c = _gvn.transform( new SubDNode(a,b) );
2421 d = dprecision_rounding(c);
2422 push_pair( d );
2423 break;
2424
2425 case Bytecodes::_dadd:
2426 b = pop_pair();
2427 a = pop_pair();
2428 c = _gvn.transform( new AddDNode(a,b) );
2429 d = dprecision_rounding(c);
2430 push_pair( d );
2431 break;
2432
2433 case Bytecodes::_dmul:
2434 b = pop_pair();
2435 a = pop_pair();
2436 c = _gvn.transform( new MulDNode(a,b) );
2437 d = dprecision_rounding(c);
2438 push_pair( d );
2439 break;
2440
2441 case Bytecodes::_ddiv:
2442 b = pop_pair();
2443 a = pop_pair();
2444 c = _gvn.transform( new DivDNode(0,a,b) );
2445 d = dprecision_rounding(c);
2446 push_pair( d );
2447 break;
2448
2449 case Bytecodes::_dneg:
2450 a = pop_pair();
2451 b = _gvn.transform(new NegDNode (a));
2452 push_pair(b);
2453 break;
2454
2455 case Bytecodes::_drem:
2456 if (Matcher::has_match_rule(Op_ModD)) {
2457 // Generate a ModD node.
2458 b = pop_pair();
2459 a = pop_pair();
2460 // a % b
2461
2462 c = _gvn.transform( new ModDNode(0,a,b) );
2463 d = dprecision_rounding(c);
2464 push_pair( d );
2465 }
2466 else {
2467 // Generate a call.
2468 modd();
2469 }
2470 break;
2471
2472 case Bytecodes::_dcmpl:
2473 b = pop_pair();
2474 a = pop_pair();
2475 c = _gvn.transform( new CmpD3Node( a, b));
2476 push(c);
2477 break;
2478
2479 case Bytecodes::_dcmpg:
2480 b = pop_pair();
2481 a = pop_pair();
2482 // Same as dcmpl but need to flip the unordered case.
2483 // Commute the inputs, which negates the result sign except for unordered.
2484 // Flip the unordered as well by using CmpD3 which implements
2485 // unordered-lesser instead of unordered-greater semantics.
2486 // Finally, negate the result bits. Result is same as using a
2487 // CmpD3Greater except we did it with CmpD3 alone.
2488 c = _gvn.transform( new CmpD3Node( b, a));
2489 c = _gvn.transform( new SubINode(_gvn.intcon(0),c) );
2490 push(c);
2491 break;
2492
2493
2494 // Note for longs -> lo word is on TOS, hi word is on TOS - 1
2495 case Bytecodes::_land:
2496 b = pop_pair();
2497 a = pop_pair();
2498 c = _gvn.transform( new AndLNode(a,b) );
2499 push_pair(c);
2500 break;
2501 case Bytecodes::_lor:
2502 b = pop_pair();
2503 a = pop_pair();
2504 c = _gvn.transform( new OrLNode(a,b) );
2505 push_pair(c);
2506 break;
2507 case Bytecodes::_lxor:
2508 b = pop_pair();
2509 a = pop_pair();
2510 c = _gvn.transform( new XorLNode(a,b) );
2511 push_pair(c);
2512 break;
2513
2514 case Bytecodes::_lshl:
2515 b = pop(); // the shift count
2516 a = pop_pair(); // value to be shifted
2517 c = _gvn.transform( new LShiftLNode(a,b) );
2518 push_pair(c);
2519 break;
2520 case Bytecodes::_lshr:
2521 b = pop(); // the shift count
2522 a = pop_pair(); // value to be shifted
2523 c = _gvn.transform( new RShiftLNode(a,b) );
2524 push_pair(c);
2525 break;
2526 case Bytecodes::_lushr:
2527 b = pop(); // the shift count
2528 a = pop_pair(); // value to be shifted
2529 c = _gvn.transform( new URShiftLNode(a,b) );
2530 push_pair(c);
2531 break;
2532 case Bytecodes::_lmul:
2533 b = pop_pair();
2534 a = pop_pair();
2535 c = _gvn.transform( new MulLNode(a,b) );
2536 push_pair(c);
2537 break;
2538
2539 case Bytecodes::_lrem:
2540 // Must keep both values on the expression-stack during null-check
2541 assert(peek(0) == top(), "long word order");
2542 zero_check_long(peek(1));
2543 // Compile-time detect of null-exception?
2544 if (stopped()) return;
2545 b = pop_pair();
2546 a = pop_pair();
2547 c = _gvn.transform( new ModLNode(control(),a,b) );
2548 push_pair(c);
2549 break;
2550
2551 case Bytecodes::_ldiv:
2552 // Must keep both values on the expression-stack during null-check
2553 assert(peek(0) == top(), "long word order");
2554 zero_check_long(peek(1));
2555 // Compile-time detect of null-exception?
2556 if (stopped()) return;
2557 b = pop_pair();
2558 a = pop_pair();
2559 c = _gvn.transform( new DivLNode(control(),a,b) );
2560 push_pair(c);
2561 break;
2562
2563 case Bytecodes::_ladd:
2564 b = pop_pair();
2565 a = pop_pair();
2566 c = _gvn.transform( new AddLNode(a,b) );
2567 push_pair(c);
2568 break;
2569 case Bytecodes::_lsub:
2570 b = pop_pair();
2571 a = pop_pair();
2572 c = _gvn.transform( new SubLNode(a,b) );
2573 push_pair(c);
2574 break;
2575 case Bytecodes::_lcmp:
2576 // Safepoints are now inserted _before_ branches. The long-compare
2577 // bytecode painfully produces a 3-way value (-1,0,+1) which requires a
2578 // slew of control flow. These are usually followed by a CmpI vs zero and
2579 // a branch; this pattern then optimizes to the obvious long-compare and
2580 // branch. However, if the branch is backwards there's a Safepoint
2581 // inserted. The inserted Safepoint captures the JVM state at the
2582 // pre-branch point, i.e. it captures the 3-way value. Thus if a
2583 // long-compare is used to control a loop the debug info will force
2584 // computation of the 3-way value, even though the generated code uses a
2585 // long-compare and branch. We try to rectify the situation by inserting
2586 // a SafePoint here and have it dominate and kill the safepoint added at a
2587 // following backwards branch. At this point the JVM state merely holds 2
2588 // longs but not the 3-way value.
2589 if( UseLoopSafepoints ) {
2590 switch( iter().next_bc() ) {
2591 case Bytecodes::_ifgt:
2592 case Bytecodes::_iflt:
2593 case Bytecodes::_ifge:
2594 case Bytecodes::_ifle:
2595 case Bytecodes::_ifne:
2596 case Bytecodes::_ifeq:
2597 // If this is a backwards branch in the bytecodes, add Safepoint
2598 maybe_add_safepoint(iter().next_get_dest());
2599 default:
2600 break;
2601 }
2602 }
2603 b = pop_pair();
2604 a = pop_pair();
2605 c = _gvn.transform( new CmpL3Node( a, b ));
2606 push(c);
2607 break;
2608
2609 case Bytecodes::_lneg:
2610 a = pop_pair();
2611 b = _gvn.transform( new SubLNode(longcon(0),a));
2612 push_pair(b);
2613 break;
2614 case Bytecodes::_l2i:
2615 a = pop_pair();
2616 push( _gvn.transform( new ConvL2INode(a)));
2617 break;
2618 case Bytecodes::_i2l:
2619 a = pop();
2620 b = _gvn.transform( new ConvI2LNode(a));
2621 push_pair(b);
2622 break;
2623 case Bytecodes::_i2b:
2624 // Sign extend
2625 a = pop();
2626 a = _gvn.transform( new LShiftINode(a,_gvn.intcon(24)) );
2627 a = _gvn.transform( new RShiftINode(a,_gvn.intcon(24)) );
2628 push( a );
2629 break;
2630 case Bytecodes::_i2s:
2631 a = pop();
2632 a = _gvn.transform( new LShiftINode(a,_gvn.intcon(16)) );
2633 a = _gvn.transform( new RShiftINode(a,_gvn.intcon(16)) );
2634 push( a );
2635 break;
2636 case Bytecodes::_i2c:
2637 a = pop();
2638 push( _gvn.transform( new AndINode(a,_gvn.intcon(0xFFFF)) ) );
2639 break;
2640
2641 case Bytecodes::_i2f:
2642 a = pop();
2643 b = _gvn.transform( new ConvI2FNode(a) ) ;
2644 c = precision_rounding(b);
2645 push (b);
2646 break;
2647
2648 case Bytecodes::_i2d:
2649 a = pop();
2650 b = _gvn.transform( new ConvI2DNode(a));
2651 push_pair(b);
2652 break;
2653
2654 case Bytecodes::_iinc: // Increment local
2655 i = iter().get_index(); // Get local index
2656 set_local( i, _gvn.transform( new AddINode( _gvn.intcon(iter().get_iinc_con()), local(i) ) ) );
2657 break;
2658
2659 // Exit points of synchronized methods must have an unlock node
2660 case Bytecodes::_return:
2661 return_current(NULL);
2662 break;
2663
2664 case Bytecodes::_ireturn:
2665 case Bytecodes::_areturn:
2666 case Bytecodes::_freturn:
2667 return_current(pop());
2668 break;
2669 case Bytecodes::_lreturn:
2670 return_current(pop_pair());
2671 break;
2672 case Bytecodes::_dreturn:
2673 return_current(pop_pair());
2674 break;
2675
2676 case Bytecodes::_athrow:
2677 // null exception oop throws NULL pointer exception
2678 null_check(peek());
2679 if (stopped()) return;
2680 // Hook the thrown exception directly to subsequent handlers.
2681 if (BailoutToInterpreterForThrows) {
2682 // Keep method interpreted from now on.
2683 uncommon_trap(Deoptimization::Reason_unhandled,
2684 Deoptimization::Action_make_not_compilable);
2685 return;
2686 }
2687 if (env()->jvmti_can_post_on_exceptions()) {
2688 // check if we must post exception events, take uncommon trap if so (with must_throw = false)
2689 uncommon_trap_if_should_post_on_exceptions(Deoptimization::Reason_unhandled, false);
2690 }
2691 // Here if either can_post_on_exceptions or should_post_on_exceptions is false
2692 add_exception_state(make_exception_state(peek()));
2693 break;
2694
2695 case Bytecodes::_goto: // fall through
2696 case Bytecodes::_goto_w: {
2697 int target_bci = (bc() == Bytecodes::_goto) ? iter().get_dest() : iter().get_far_dest();
2698
2699 // If this is a backwards branch in the bytecodes, add Safepoint
2700 maybe_add_safepoint(target_bci);
2701
2702 // Update method data
2703 profile_taken_branch(target_bci);
2704
2705 // Merge the current control into the target basic block
2706 merge(target_bci);
2707
2708 // See if we can get some profile data and hand it off to the next block
2709 Block *target_block = block()->successor_for_bci(target_bci);
2710 if (target_block->pred_count() != 1) break;
2711 ciMethodData* methodData = method()->method_data();
2712 if (!methodData->is_mature()) break;
2713 ciProfileData* data = methodData->bci_to_data(bci());
2714 assert(data != NULL && data->is_JumpData(), "need JumpData for taken branch");
2715 int taken = ((ciJumpData*)data)->taken();
2716 taken = method()->scale_count(taken);
2717 target_block->set_count(taken);
2718 break;
2719 }
2720
2721 case Bytecodes::_ifnull: btest = BoolTest::eq; goto handle_if_null;
2722 case Bytecodes::_ifnonnull: btest = BoolTest::ne; goto handle_if_null;
2723 handle_if_null:
2724 // If this is a backwards branch in the bytecodes, add Safepoint
2725 maybe_add_safepoint(iter().get_dest());
2726 a = null();
2727 b = pop();
2728 if (!_gvn.type(b)->speculative_maybe_null() &&
2729 !too_many_traps(Deoptimization::Reason_speculate_null_check)) {
2730 inc_sp(1);
2731 Node* null_ctl = top();
2732 b = null_check_oop(b, &null_ctl, true, true, true);
2733 assert(null_ctl->is_top(), "no null control here");
2734 dec_sp(1);
2735 } else if (_gvn.type(b)->speculative_always_null() &&
2736 !too_many_traps(Deoptimization::Reason_speculate_null_assert)) {
2737 inc_sp(1);
2738 b = null_assert(b);
2739 dec_sp(1);
2740 }
2741 c = _gvn.transform( new CmpPNode(b, a) );
2742 do_ifnull(btest, c);
2743 break;
2744
2745 case Bytecodes::_if_acmpeq: btest = BoolTest::eq; goto handle_if_acmp;
2746 case Bytecodes::_if_acmpne: btest = BoolTest::ne; goto handle_if_acmp;
2747 handle_if_acmp:
2748 // If this is a backwards branch in the bytecodes, add Safepoint
2749 maybe_add_safepoint(iter().get_dest());
2750 a = access_resolve(pop(), 0);
2751 b = access_resolve(pop(), 0);
2752 c = _gvn.transform( new CmpPNode(b, a) );
2753 c = optimize_cmp_with_klass(c);
2754 do_if(btest, c);
2755 break;
2756
2757 case Bytecodes::_ifeq: btest = BoolTest::eq; goto handle_ifxx;
2758 case Bytecodes::_ifne: btest = BoolTest::ne; goto handle_ifxx;
2759 case Bytecodes::_iflt: btest = BoolTest::lt; goto handle_ifxx;
2760 case Bytecodes::_ifle: btest = BoolTest::le; goto handle_ifxx;
2761 case Bytecodes::_ifgt: btest = BoolTest::gt; goto handle_ifxx;
2762 case Bytecodes::_ifge: btest = BoolTest::ge; goto handle_ifxx;
2763 handle_ifxx:
2764 // If this is a backwards branch in the bytecodes, add Safepoint
2765 maybe_add_safepoint(iter().get_dest());
2766 a = _gvn.intcon(0);
2767 b = pop();
2768 c = _gvn.transform( new CmpINode(b, a) );
2769 do_if(btest, c);
2770 break;
2771
2772 case Bytecodes::_if_icmpeq: btest = BoolTest::eq; goto handle_if_icmp;
2773 case Bytecodes::_if_icmpne: btest = BoolTest::ne; goto handle_if_icmp;
2774 case Bytecodes::_if_icmplt: btest = BoolTest::lt; goto handle_if_icmp;
2775 case Bytecodes::_if_icmple: btest = BoolTest::le; goto handle_if_icmp;
2776 case Bytecodes::_if_icmpgt: btest = BoolTest::gt; goto handle_if_icmp;
2777 case Bytecodes::_if_icmpge: btest = BoolTest::ge; goto handle_if_icmp;
2778 handle_if_icmp:
2779 // If this is a backwards branch in the bytecodes, add Safepoint
2780 maybe_add_safepoint(iter().get_dest());
2781 a = pop();
2782 b = pop();
2783 c = _gvn.transform( new CmpINode( b, a ) );
2784 do_if(btest, c);
2785 break;
2786
2787 case Bytecodes::_tableswitch:
2788 do_tableswitch();
2789 break;
2790
2791 case Bytecodes::_lookupswitch:
2792 do_lookupswitch();
2793 break;
2794
2795 case Bytecodes::_invokestatic:
2796 case Bytecodes::_invokedynamic:
2797 case Bytecodes::_invokespecial:
2798 case Bytecodes::_invokevirtual:
2799 case Bytecodes::_invokeinterface:
2800 do_call();
2801 break;
2802 case Bytecodes::_checkcast:
2803 do_checkcast();
2804 break;
2805 case Bytecodes::_instanceof:
2806 do_instanceof();
2807 break;
2808 case Bytecodes::_anewarray:
2809 do_anewarray();
2810 break;
2811 case Bytecodes::_newarray:
2812 do_newarray((BasicType)iter().get_index());
2813 break;
2814 case Bytecodes::_multianewarray:
2815 do_multianewarray();
2816 break;
2817 case Bytecodes::_new:
2818 do_new();
2819 break;
2820
2821 case Bytecodes::_jsr:
2822 case Bytecodes::_jsr_w:
2823 do_jsr();
2824 break;
2825
2826 case Bytecodes::_ret:
2827 do_ret();
2828 break;
2829
2830
2831 case Bytecodes::_monitorenter:
2832 do_monitor_enter();
2833 break;
2834
2835 case Bytecodes::_monitorexit:
2836 do_monitor_exit();
2837 break;
2838
2839 case Bytecodes::_breakpoint:
2840 // Breakpoint set concurrently to compile
2841 // %%% use an uncommon trap?
2842 C->record_failure("breakpoint in method");
2843 return;
2844
2845 default:
2846#ifndef PRODUCT
2847 map()->dump(99);
2848#endif
2849 tty->print("\nUnhandled bytecode %s\n", Bytecodes::name(bc()) );
2850 ShouldNotReachHere();
2851 }
2852
2853#ifndef PRODUCT
2854 IdealGraphPrinter *printer = C->printer();
2855 if (printer && printer->should_print(1)) {
2856 char buffer[256];
2857 jio_snprintf(buffer, sizeof(buffer), "Bytecode %d: %s", bci(), Bytecodes::name(bc()));
2858 bool old = printer->traverse_outs();
2859 printer->set_traverse_outs(true);
2860 printer->print_method(buffer, 4);
2861 printer->set_traverse_outs(old);
2862 }
2863#endif
2864}
2865