Thu, 24 May 2018 19:26:50 +0800
#7046 C2 supports long branch
Contributed-by: fujie
1 /*
2 * Copyright (c) 1998, 2016, 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 */
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/universe.inline.hpp"
32 #include "opto/addnode.hpp"
33 #include "opto/divnode.hpp"
34 #include "opto/idealGraphPrinter.hpp"
35 #include "opto/matcher.hpp"
36 #include "opto/memnode.hpp"
37 #include "opto/mulnode.hpp"
38 #include "opto/parse.hpp"
39 #include "opto/runtime.hpp"
40 #include "runtime/deoptimization.hpp"
41 #include "runtime/sharedRuntime.hpp"
43 extern int explicit_null_checks_inserted,
44 explicit_null_checks_elided;
46 //---------------------------------array_load----------------------------------
47 void Parse::array_load(BasicType elem_type) {
48 const Type* elem = Type::TOP;
49 Node* adr = array_addressing(elem_type, 0, &elem);
50 if (stopped()) return; // guaranteed null or range check
51 dec_sp(2); // Pop array and index
52 const TypeAryPtr* adr_type = TypeAryPtr::get_array_body_type(elem_type);
53 Node* ld = make_load(control(), adr, elem, elem_type, adr_type, MemNode::unordered);
54 push(ld);
55 }
58 //--------------------------------array_store----------------------------------
59 void Parse::array_store(BasicType elem_type) {
60 const Type* elem = Type::TOP;
61 Node* adr = array_addressing(elem_type, 1, &elem);
62 if (stopped()) return; // guaranteed null or range check
63 Node* val = pop();
64 dec_sp(2); // Pop array and index
65 const TypeAryPtr* adr_type = TypeAryPtr::get_array_body_type(elem_type);
66 if (elem == TypeInt::BOOL) {
67 elem_type = T_BOOLEAN;
68 }
69 store_to_memory(control(), adr, val, elem_type, adr_type, StoreNode::release_if_reference(elem_type));
70 }
73 //------------------------------array_addressing-------------------------------
74 // Pull array and index from the stack. Compute pointer-to-element.
75 Node* Parse::array_addressing(BasicType type, int vals, const Type* *result2) {
76 Node *idx = peek(0+vals); // Get from stack without popping
77 Node *ary = peek(1+vals); // in case of exception
79 // Null check the array base, with correct stack contents
80 ary = null_check(ary, T_ARRAY);
81 // Compile-time detect of null-exception?
82 if (stopped()) return top();
84 const TypeAryPtr* arytype = _gvn.type(ary)->is_aryptr();
85 const TypeInt* sizetype = arytype->size();
86 const Type* elemtype = arytype->elem();
88 if (UseUniqueSubclasses && result2 != NULL) {
89 const Type* el = elemtype->make_ptr();
90 if (el && el->isa_instptr()) {
91 const TypeInstPtr* toop = el->is_instptr();
92 if (toop->klass()->as_instance_klass()->unique_concrete_subklass()) {
93 // If we load from "AbstractClass[]" we must see "ConcreteSubClass".
94 const Type* subklass = Type::get_const_type(toop->klass());
95 elemtype = subklass->join_speculative(el);
96 }
97 }
98 }
100 // Check for big class initializers with all constant offsets
101 // feeding into a known-size array.
102 const TypeInt* idxtype = _gvn.type(idx)->is_int();
103 // See if the highest idx value is less than the lowest array bound,
104 // and if the idx value cannot be negative:
105 bool need_range_check = true;
106 if (idxtype->_hi < sizetype->_lo && idxtype->_lo >= 0) {
107 need_range_check = false;
108 if (C->log() != NULL) C->log()->elem("observe that='!need_range_check'");
109 }
111 ciKlass * arytype_klass = arytype->klass();
112 if ((arytype_klass != NULL) && (!arytype_klass->is_loaded())) {
113 // Only fails for some -Xcomp runs
114 // The class is unloaded. We have to run this bytecode in the interpreter.
115 uncommon_trap(Deoptimization::Reason_unloaded,
116 Deoptimization::Action_reinterpret,
117 arytype->klass(), "!loaded array");
118 return top();
119 }
121 // Do the range check
122 if (GenerateRangeChecks && need_range_check) {
123 Node* tst;
124 if (sizetype->_hi <= 0) {
125 // The greatest array bound is negative, so we can conclude that we're
126 // compiling unreachable code, but the unsigned compare trick used below
127 // only works with non-negative lengths. Instead, hack "tst" to be zero so
128 // the uncommon_trap path will always be taken.
129 tst = _gvn.intcon(0);
130 } else {
131 // Range is constant in array-oop, so we can use the original state of mem
132 Node* len = load_array_length(ary);
134 // Test length vs index (standard trick using unsigned compare)
135 Node* chk = _gvn.transform( new (C) CmpUNode(idx, len) );
136 BoolTest::mask btest = BoolTest::lt;
137 tst = _gvn.transform( new (C) BoolNode(chk, btest) );
138 }
139 // Branch to failure if out of bounds
140 { BuildCutout unless(this, tst, PROB_MAX);
141 if (C->allow_range_check_smearing()) {
142 // Do not use builtin_throw, since range checks are sometimes
143 // made more stringent by an optimistic transformation.
144 // This creates "tentative" range checks at this point,
145 // which are not guaranteed to throw exceptions.
146 // See IfNode::Ideal, is_range_check, adjust_check.
147 uncommon_trap(Deoptimization::Reason_range_check,
148 Deoptimization::Action_make_not_entrant,
149 NULL, "range_check");
150 } else {
151 // If we have already recompiled with the range-check-widening
152 // heroic optimization turned off, then we must really be throwing
153 // range check exceptions.
154 builtin_throw(Deoptimization::Reason_range_check, idx);
155 }
156 }
157 }
158 // Check for always knowing you are throwing a range-check exception
159 if (stopped()) return top();
161 // Make array address computation control dependent to prevent it
162 // from floating above the range check during loop optimizations.
163 Node* ptr = array_element_address(ary, idx, type, sizetype, control());
165 if (result2 != NULL) *result2 = elemtype;
167 assert(ptr != top(), "top should go hand-in-hand with stopped");
169 return ptr;
170 }
173 // returns IfNode
174 IfNode* Parse::jump_if_fork_int(Node* a, Node* b, BoolTest::mask mask) {
175 Node *cmp = _gvn.transform( new (C) CmpINode( a, b)); // two cases: shiftcount > 32 and shiftcount <= 32
176 Node *tst = _gvn.transform( new (C) BoolNode( cmp, mask));
177 IfNode *iff = create_and_map_if( control(), tst, ((mask == BoolTest::eq) ? PROB_STATIC_INFREQUENT : PROB_FAIR), COUNT_UNKNOWN );
178 return iff;
179 }
181 // return Region node
182 Node* Parse::jump_if_join(Node* iffalse, Node* iftrue) {
183 Node *region = new (C) RegionNode(3); // 2 results
184 record_for_igvn(region);
185 region->init_req(1, iffalse);
186 region->init_req(2, iftrue );
187 _gvn.set_type(region, Type::CONTROL);
188 region = _gvn.transform(region);
189 set_control (region);
190 return region;
191 }
194 //------------------------------helper for tableswitch-------------------------
195 void Parse::jump_if_true_fork(IfNode *iff, int dest_bci_if_true, int prof_table_index) {
196 // True branch, use existing map info
197 { PreserveJVMState pjvms(this);
198 Node *iftrue = _gvn.transform( new (C) IfTrueNode (iff) );
199 set_control( iftrue );
200 profile_switch_case(prof_table_index);
201 merge_new_path(dest_bci_if_true);
202 }
204 // False branch
205 Node *iffalse = _gvn.transform( new (C) IfFalseNode(iff) );
206 set_control( iffalse );
207 }
209 void Parse::jump_if_false_fork(IfNode *iff, int dest_bci_if_true, int prof_table_index) {
210 // True branch, use existing map info
211 { PreserveJVMState pjvms(this);
212 Node *iffalse = _gvn.transform( new (C) IfFalseNode (iff) );
213 set_control( iffalse );
214 profile_switch_case(prof_table_index);
215 merge_new_path(dest_bci_if_true);
216 }
218 // False branch
219 Node *iftrue = _gvn.transform( new (C) IfTrueNode(iff) );
220 set_control( iftrue );
221 }
223 void Parse::jump_if_always_fork(int dest_bci, int prof_table_index) {
224 // False branch, use existing map and control()
225 profile_switch_case(prof_table_index);
226 merge_new_path(dest_bci);
227 }
230 extern "C" {
231 static int jint_cmp(const void *i, const void *j) {
232 int a = *(jint *)i;
233 int b = *(jint *)j;
234 return a > b ? 1 : a < b ? -1 : 0;
235 }
236 }
239 // Default value for methodData switch indexing. Must be a negative value to avoid
240 // conflict with any legal switch index.
241 #define NullTableIndex -1
243 class SwitchRange : public StackObj {
244 // a range of integers coupled with a bci destination
245 jint _lo; // inclusive lower limit
246 jint _hi; // inclusive upper limit
247 int _dest;
248 int _table_index; // index into method data table
250 public:
251 jint lo() const { return _lo; }
252 jint hi() const { return _hi; }
253 int dest() const { return _dest; }
254 int table_index() const { return _table_index; }
255 bool is_singleton() const { return _lo == _hi; }
257 void setRange(jint lo, jint hi, int dest, int table_index) {
258 assert(lo <= hi, "must be a non-empty range");
259 _lo = lo, _hi = hi; _dest = dest; _table_index = table_index;
260 }
261 bool adjoinRange(jint lo, jint hi, int dest, int table_index) {
262 assert(lo <= hi, "must be a non-empty range");
263 if (lo == _hi+1 && dest == _dest && table_index == _table_index) {
264 _hi = hi;
265 return true;
266 }
267 return false;
268 }
270 void set (jint value, int dest, int table_index) {
271 setRange(value, value, dest, table_index);
272 }
273 bool adjoin(jint value, int dest, int table_index) {
274 return adjoinRange(value, value, dest, table_index);
275 }
277 void print() {
278 if (is_singleton())
279 tty->print(" {%d}=>%d", lo(), dest());
280 else if (lo() == min_jint)
281 tty->print(" {..%d}=>%d", hi(), dest());
282 else if (hi() == max_jint)
283 tty->print(" {%d..}=>%d", lo(), dest());
284 else
285 tty->print(" {%d..%d}=>%d", lo(), hi(), dest());
286 }
287 };
290 //-------------------------------do_tableswitch--------------------------------
291 void Parse::do_tableswitch() {
292 Node* lookup = pop();
294 // Get information about tableswitch
295 int default_dest = iter().get_dest_table(0);
296 int lo_index = iter().get_int_table(1);
297 int hi_index = iter().get_int_table(2);
298 int len = hi_index - lo_index + 1;
300 if (len < 1) {
301 // If this is a backward branch, add safepoint
302 maybe_add_safepoint(default_dest);
303 merge(default_dest);
304 return;
305 }
307 // generate decision tree, using trichotomy when possible
308 int rnum = len+2;
309 bool makes_backward_branch = false;
310 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum);
311 int rp = -1;
312 if (lo_index != min_jint) {
313 ranges[++rp].setRange(min_jint, lo_index-1, default_dest, NullTableIndex);
314 }
315 for (int j = 0; j < len; j++) {
316 jint match_int = lo_index+j;
317 int dest = iter().get_dest_table(j+3);
318 makes_backward_branch |= (dest <= bci());
319 int table_index = method_data_update() ? j : NullTableIndex;
320 if (rp < 0 || !ranges[rp].adjoin(match_int, dest, table_index)) {
321 ranges[++rp].set(match_int, dest, table_index);
322 }
323 }
324 jint highest = lo_index+(len-1);
325 assert(ranges[rp].hi() == highest, "");
326 if (highest != max_jint
327 && !ranges[rp].adjoinRange(highest+1, max_jint, default_dest, NullTableIndex)) {
328 ranges[++rp].setRange(highest+1, max_jint, default_dest, NullTableIndex);
329 }
330 assert(rp < len+2, "not too many ranges");
332 // Safepoint in case if backward branch observed
333 if( makes_backward_branch && UseLoopSafepoints )
334 add_safepoint();
336 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]);
337 }
340 //------------------------------do_lookupswitch--------------------------------
341 void Parse::do_lookupswitch() {
342 Node *lookup = pop(); // lookup value
343 // Get information about lookupswitch
344 int default_dest = iter().get_dest_table(0);
345 int len = iter().get_int_table(1);
347 if (len < 1) { // If this is a backward branch, add safepoint
348 maybe_add_safepoint(default_dest);
349 merge(default_dest);
350 return;
351 }
353 // generate decision tree, using trichotomy when possible
354 jint* table = NEW_RESOURCE_ARRAY(jint, len*2);
355 {
356 for( int j = 0; j < len; j++ ) {
357 table[j+j+0] = iter().get_int_table(2+j+j);
358 table[j+j+1] = iter().get_dest_table(2+j+j+1);
359 }
360 qsort( table, len, 2*sizeof(table[0]), jint_cmp );
361 }
363 int rnum = len*2+1;
364 bool makes_backward_branch = false;
365 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum);
366 int rp = -1;
367 for( int j = 0; j < len; j++ ) {
368 jint match_int = table[j+j+0];
369 int dest = table[j+j+1];
370 int next_lo = rp < 0 ? min_jint : ranges[rp].hi()+1;
371 int table_index = method_data_update() ? j : NullTableIndex;
372 makes_backward_branch |= (dest <= bci());
373 if( match_int != next_lo ) {
374 ranges[++rp].setRange(next_lo, match_int-1, default_dest, NullTableIndex);
375 }
376 if( rp < 0 || !ranges[rp].adjoin(match_int, dest, table_index) ) {
377 ranges[++rp].set(match_int, dest, table_index);
378 }
379 }
380 jint highest = table[2*(len-1)];
381 assert(ranges[rp].hi() == highest, "");
382 if( highest != max_jint
383 && !ranges[rp].adjoinRange(highest+1, max_jint, default_dest, NullTableIndex) ) {
384 ranges[++rp].setRange(highest+1, max_jint, default_dest, NullTableIndex);
385 }
386 assert(rp < rnum, "not too many ranges");
388 // Safepoint in case backward branch observed
389 if( makes_backward_branch && UseLoopSafepoints )
390 add_safepoint();
392 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]);
393 }
395 //----------------------------create_jump_tables-------------------------------
396 bool Parse::create_jump_tables(Node* key_val, SwitchRange* lo, SwitchRange* hi) {
397 // Are jumptables enabled
398 if (!UseJumpTables) return false;
400 // Are jumptables supported
401 if (!Matcher::has_match_rule(Op_Jump)) return false;
403 // Don't make jump table if profiling
404 if (method_data_update()) return false;
406 // Decide if a guard is needed to lop off big ranges at either (or
407 // both) end(s) of the input set. We'll call this the default target
408 // even though we can't be sure that it is the true "default".
410 bool needs_guard = false;
411 int default_dest;
412 int64 total_outlier_size = 0;
413 int64 hi_size = ((int64)hi->hi()) - ((int64)hi->lo()) + 1;
414 int64 lo_size = ((int64)lo->hi()) - ((int64)lo->lo()) + 1;
416 if (lo->dest() == hi->dest()) {
417 total_outlier_size = hi_size + lo_size;
418 default_dest = lo->dest();
419 } else if (lo_size > hi_size) {
420 total_outlier_size = lo_size;
421 default_dest = lo->dest();
422 } else {
423 total_outlier_size = hi_size;
424 default_dest = hi->dest();
425 }
427 // If a guard test will eliminate very sparse end ranges, then
428 // it is worth the cost of an extra jump.
429 if (total_outlier_size > (MaxJumpTableSparseness * 4)) {
430 needs_guard = true;
431 if (default_dest == lo->dest()) lo++;
432 if (default_dest == hi->dest()) hi--;
433 }
435 // Find the total number of cases and ranges
436 int64 num_cases = ((int64)hi->hi()) - ((int64)lo->lo()) + 1;
437 int num_range = hi - lo + 1;
439 // Don't create table if: too large, too small, or too sparse.
440 if (num_cases < MinJumpTableSize || num_cases > MaxJumpTableSize)
441 return false;
442 if (num_cases > (MaxJumpTableSparseness * num_range))
443 return false;
445 // Normalize table lookups to zero
446 int lowval = lo->lo();
447 key_val = _gvn.transform( new (C) SubINode(key_val, _gvn.intcon(lowval)) );
449 // Generate a guard to protect against input keyvals that aren't
450 // in the switch domain.
451 if (needs_guard) {
452 Node* size = _gvn.intcon(num_cases);
453 Node* cmp = _gvn.transform( new (C) CmpUNode(key_val, size) );
454 Node* tst = _gvn.transform( new (C) BoolNode(cmp, BoolTest::ge) );
455 IfNode* iff = create_and_map_if( control(), tst, PROB_FAIR, COUNT_UNKNOWN);
456 jump_if_true_fork(iff, default_dest, NullTableIndex);
457 }
459 // Create an ideal node JumpTable that has projections
460 // of all possible ranges for a switch statement
461 // The key_val input must be converted to a pointer offset and scaled.
462 // Compare Parse::array_addressing above.
463 #ifdef _LP64
464 // Clean the 32-bit int into a real 64-bit offset.
465 // Otherwise, the jint value 0 might turn into an offset of 0x0800000000.
466 const TypeInt* ikeytype = TypeInt::make(0, num_cases-1, Type::WidenMin);
467 // Make I2L conversion control dependent to prevent it from
468 // floating above the range check during loop optimizations.
469 key_val = C->constrained_convI2L(&_gvn, key_val, ikeytype, control());
470 #endif
472 // Shift the value by wordsize so we have an index into the table, rather
473 // than a switch value
474 Node *shiftWord = _gvn.MakeConX(wordSize);
475 key_val = _gvn.transform( new (C) MulXNode( key_val, shiftWord));
477 // Create the JumpNode
478 Node* jtn = _gvn.transform( new (C) JumpNode(control(), key_val, num_cases) );
480 // These are the switch destinations hanging off the jumpnode
481 int i = 0;
482 for (SwitchRange* r = lo; r <= hi; r++) {
483 for (int64 j = r->lo(); j <= r->hi(); j++, i++) {
484 Node* input = _gvn.transform(new (C) JumpProjNode(jtn, i, r->dest(), (int)(j - lowval)));
485 {
486 PreserveJVMState pjvms(this);
487 set_control(input);
488 jump_if_always_fork(r->dest(), r->table_index());
489 }
490 }
491 }
492 assert(i == num_cases, "miscount of cases");
493 stop_and_kill_map(); // no more uses for this JVMS
494 return true;
495 }
497 //----------------------------jump_switch_ranges-------------------------------
498 void Parse::jump_switch_ranges(Node* key_val, SwitchRange *lo, SwitchRange *hi, int switch_depth) {
499 Block* switch_block = block();
501 if (switch_depth == 0) {
502 // Do special processing for the top-level call.
503 assert(lo->lo() == min_jint, "initial range must exhaust Type::INT");
504 assert(hi->hi() == max_jint, "initial range must exhaust Type::INT");
506 // Decrement pred-numbers for the unique set of nodes.
507 #ifdef ASSERT
508 // Ensure that the block's successors are a (duplicate-free) set.
509 int successors_counted = 0; // block occurrences in [hi..lo]
510 int unique_successors = switch_block->num_successors();
511 for (int i = 0; i < unique_successors; i++) {
512 Block* target = switch_block->successor_at(i);
514 // Check that the set of successors is the same in both places.
515 int successors_found = 0;
516 for (SwitchRange* p = lo; p <= hi; p++) {
517 if (p->dest() == target->start()) successors_found++;
518 }
519 assert(successors_found > 0, "successor must be known");
520 successors_counted += successors_found;
521 }
522 assert(successors_counted == (hi-lo)+1, "no unexpected successors");
523 #endif
525 // Maybe prune the inputs, based on the type of key_val.
526 jint min_val = min_jint;
527 jint max_val = max_jint;
528 const TypeInt* ti = key_val->bottom_type()->isa_int();
529 if (ti != NULL) {
530 min_val = ti->_lo;
531 max_val = ti->_hi;
532 assert(min_val <= max_val, "invalid int type");
533 }
534 while (lo->hi() < min_val) lo++;
535 if (lo->lo() < min_val) lo->setRange(min_val, lo->hi(), lo->dest(), lo->table_index());
536 while (hi->lo() > max_val) hi--;
537 if (hi->hi() > max_val) hi->setRange(hi->lo(), max_val, hi->dest(), hi->table_index());
538 }
540 #ifndef PRODUCT
541 if (switch_depth == 0) {
542 _max_switch_depth = 0;
543 _est_switch_depth = log2_intptr((hi-lo+1)-1)+1;
544 }
545 #endif
547 assert(lo <= hi, "must be a non-empty set of ranges");
548 if (lo == hi) {
549 jump_if_always_fork(lo->dest(), lo->table_index());
550 } else {
551 assert(lo->hi() == (lo+1)->lo()-1, "contiguous ranges");
552 assert(hi->lo() == (hi-1)->hi()+1, "contiguous ranges");
554 if (create_jump_tables(key_val, lo, hi)) return;
556 int nr = hi - lo + 1;
558 SwitchRange* mid = lo + nr/2;
559 // if there is an easy choice, pivot at a singleton:
560 if (nr > 3 && !mid->is_singleton() && (mid-1)->is_singleton()) mid--;
562 assert(lo < mid && mid <= hi, "good pivot choice");
563 assert(nr != 2 || mid == hi, "should pick higher of 2");
564 assert(nr != 3 || mid == hi-1, "should pick middle of 3");
566 Node *test_val = _gvn.intcon(mid->lo());
568 if (mid->is_singleton()) {
569 IfNode *iff_ne = jump_if_fork_int(key_val, test_val, BoolTest::ne);
570 jump_if_false_fork(iff_ne, mid->dest(), mid->table_index());
572 // Special Case: If there are exactly three ranges, and the high
573 // and low range each go to the same place, omit the "gt" test,
574 // since it will not discriminate anything.
575 bool eq_test_only = (hi == lo+2 && hi->dest() == lo->dest());
576 if (eq_test_only) {
577 assert(mid == hi-1, "");
578 }
580 // if there is a higher range, test for it and process it:
581 if (mid < hi && !eq_test_only) {
582 // two comparisons of same values--should enable 1 test for 2 branches
583 // Use BoolTest::le instead of BoolTest::gt
584 IfNode *iff_le = jump_if_fork_int(key_val, test_val, BoolTest::le);
585 Node *iftrue = _gvn.transform( new (C) IfTrueNode(iff_le) );
586 Node *iffalse = _gvn.transform( new (C) IfFalseNode(iff_le) );
587 { PreserveJVMState pjvms(this);
588 set_control(iffalse);
589 jump_switch_ranges(key_val, mid+1, hi, switch_depth+1);
590 }
591 set_control(iftrue);
592 }
594 } else {
595 // mid is a range, not a singleton, so treat mid..hi as a unit
596 IfNode *iff_ge = jump_if_fork_int(key_val, test_val, BoolTest::ge);
598 // if there is a higher range, test for it and process it:
599 if (mid == hi) {
600 jump_if_true_fork(iff_ge, mid->dest(), mid->table_index());
601 } else {
602 Node *iftrue = _gvn.transform( new (C) IfTrueNode(iff_ge) );
603 Node *iffalse = _gvn.transform( new (C) IfFalseNode(iff_ge) );
604 { PreserveJVMState pjvms(this);
605 set_control(iftrue);
606 jump_switch_ranges(key_val, mid, hi, switch_depth+1);
607 }
608 set_control(iffalse);
609 }
610 }
612 // in any case, process the lower range
613 jump_switch_ranges(key_val, lo, mid-1, switch_depth+1);
614 }
616 // Decrease pred_count for each successor after all is done.
617 if (switch_depth == 0) {
618 int unique_successors = switch_block->num_successors();
619 for (int i = 0; i < unique_successors; i++) {
620 Block* target = switch_block->successor_at(i);
621 // Throw away the pre-allocated path for each unique successor.
622 target->next_path_num();
623 }
624 }
626 #ifndef PRODUCT
627 _max_switch_depth = MAX2(switch_depth, _max_switch_depth);
628 if (TraceOptoParse && Verbose && WizardMode && switch_depth == 0) {
629 SwitchRange* r;
630 int nsing = 0;
631 for( r = lo; r <= hi; r++ ) {
632 if( r->is_singleton() ) nsing++;
633 }
634 tty->print(">>> ");
635 _method->print_short_name();
636 tty->print_cr(" switch decision tree");
637 tty->print_cr(" %d ranges (%d singletons), max_depth=%d, est_depth=%d",
638 (int) (hi-lo+1), nsing, _max_switch_depth, _est_switch_depth);
639 if (_max_switch_depth > _est_switch_depth) {
640 tty->print_cr("******** BAD SWITCH DEPTH ********");
641 }
642 tty->print(" ");
643 for( r = lo; r <= hi; r++ ) {
644 r->print();
645 }
646 tty->cr();
647 }
648 #endif
649 }
651 void Parse::modf() {
652 Node *f2 = pop();
653 Node *f1 = pop();
654 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::modf_Type(),
655 CAST_FROM_FN_PTR(address, SharedRuntime::frem),
656 "frem", NULL, //no memory effects
657 f1, f2);
658 Node* res = _gvn.transform(new (C) ProjNode(c, TypeFunc::Parms + 0));
660 push(res);
661 }
663 void Parse::modd() {
664 Node *d2 = pop_pair();
665 Node *d1 = pop_pair();
666 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::Math_DD_D_Type(),
667 CAST_FROM_FN_PTR(address, SharedRuntime::drem),
668 "drem", NULL, //no memory effects
669 d1, top(), d2, top());
670 Node* res_d = _gvn.transform(new (C) ProjNode(c, TypeFunc::Parms + 0));
672 #ifdef ASSERT
673 Node* res_top = _gvn.transform(new (C) ProjNode(c, TypeFunc::Parms + 1));
674 assert(res_top == top(), "second value must be top");
675 #endif
677 push_pair(res_d);
678 }
680 void Parse::l2f() {
681 Node* f2 = pop();
682 Node* f1 = pop();
683 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::l2f_Type(),
684 CAST_FROM_FN_PTR(address, SharedRuntime::l2f),
685 "l2f", NULL, //no memory effects
686 f1, f2);
687 Node* res = _gvn.transform(new (C) ProjNode(c, TypeFunc::Parms + 0));
689 push(res);
690 }
692 void Parse::do_irem() {
693 // Must keep both values on the expression-stack during null-check
694 zero_check_int(peek());
695 // Compile-time detect of null-exception?
696 if (stopped()) return;
698 Node* b = pop();
699 Node* a = pop();
701 const Type *t = _gvn.type(b);
702 if (t != Type::TOP) {
703 const TypeInt *ti = t->is_int();
704 if (ti->is_con()) {
705 int divisor = ti->get_con();
706 // check for positive power of 2
707 if (divisor > 0 &&
708 (divisor & ~(divisor-1)) == divisor) {
709 // yes !
710 Node *mask = _gvn.intcon((divisor - 1));
711 // Sigh, must handle negative dividends
712 Node *zero = _gvn.intcon(0);
713 IfNode *ifff = jump_if_fork_int(a, zero, BoolTest::lt);
714 Node *iff = _gvn.transform( new (C) IfFalseNode(ifff) );
715 Node *ift = _gvn.transform( new (C) IfTrueNode (ifff) );
716 Node *reg = jump_if_join(ift, iff);
717 Node *phi = PhiNode::make(reg, NULL, TypeInt::INT);
718 // Negative path; negate/and/negate
719 Node *neg = _gvn.transform( new (C) SubINode(zero, a) );
720 Node *andn= _gvn.transform( new (C) AndINode(neg, mask) );
721 Node *negn= _gvn.transform( new (C) SubINode(zero, andn) );
722 phi->init_req(1, negn);
723 // Fast positive case
724 Node *andx = _gvn.transform( new (C) AndINode(a, mask) );
725 phi->init_req(2, andx);
726 // Push the merge
727 push( _gvn.transform(phi) );
728 return;
729 }
730 }
731 }
732 // Default case
733 push( _gvn.transform( new (C) ModINode(control(),a,b) ) );
734 }
736 // Handle jsr and jsr_w bytecode
737 void Parse::do_jsr() {
738 assert(bc() == Bytecodes::_jsr || bc() == Bytecodes::_jsr_w, "wrong bytecode");
740 // Store information about current state, tagged with new _jsr_bci
741 int return_bci = iter().next_bci();
742 int jsr_bci = (bc() == Bytecodes::_jsr) ? iter().get_dest() : iter().get_far_dest();
744 // Update method data
745 profile_taken_branch(jsr_bci);
747 // The way we do things now, there is only one successor block
748 // for the jsr, because the target code is cloned by ciTypeFlow.
749 Block* target = successor_for_bci(jsr_bci);
751 // What got pushed?
752 const Type* ret_addr = target->peek();
753 assert(ret_addr->singleton(), "must be a constant (cloned jsr body)");
755 // Effect on jsr on stack
756 push(_gvn.makecon(ret_addr));
758 // Flow to the jsr.
759 merge(jsr_bci);
760 }
762 // Handle ret bytecode
763 void Parse::do_ret() {
764 // Find to whom we return.
765 assert(block()->num_successors() == 1, "a ret can only go one place now");
766 Block* target = block()->successor_at(0);
767 assert(!target->is_ready(), "our arrival must be expected");
768 profile_ret(target->flow()->start());
769 int pnum = target->next_path_num();
770 merge_common(target, pnum);
771 }
773 static bool has_injected_profile(BoolTest::mask btest, Node* test, int& taken, int& not_taken) {
774 if (btest != BoolTest::eq && btest != BoolTest::ne) {
775 // Only ::eq and ::ne are supported for profile injection.
776 return false;
777 }
778 if (test->is_Cmp() &&
779 test->in(1)->Opcode() == Op_ProfileBoolean) {
780 ProfileBooleanNode* profile = (ProfileBooleanNode*)test->in(1);
781 int false_cnt = profile->false_count();
782 int true_cnt = profile->true_count();
784 // Counts matching depends on the actual test operation (::eq or ::ne).
785 // No need to scale the counts because profile injection was designed
786 // to feed exact counts into VM.
787 taken = (btest == BoolTest::eq) ? false_cnt : true_cnt;
788 not_taken = (btest == BoolTest::eq) ? true_cnt : false_cnt;
790 profile->consume();
791 return true;
792 }
793 return false;
794 }
795 //--------------------------dynamic_branch_prediction--------------------------
796 // Try to gather dynamic branch prediction behavior. Return a probability
797 // of the branch being taken and set the "cnt" field. Returns a -1.0
798 // if we need to use static prediction for some reason.
799 float Parse::dynamic_branch_prediction(float &cnt, BoolTest::mask btest, Node* test) {
800 ResourceMark rm;
802 cnt = COUNT_UNKNOWN;
804 int taken = 0;
805 int not_taken = 0;
807 bool use_mdo = !has_injected_profile(btest, test, taken, not_taken);
809 if (use_mdo) {
810 // Use MethodData information if it is available
811 // FIXME: free the ProfileData structure
812 ciMethodData* methodData = method()->method_data();
813 if (!methodData->is_mature()) return PROB_UNKNOWN;
814 ciProfileData* data = methodData->bci_to_data(bci());
815 if (!data->is_JumpData()) return PROB_UNKNOWN;
817 // get taken and not taken values
818 taken = data->as_JumpData()->taken();
819 not_taken = 0;
820 if (data->is_BranchData()) {
821 not_taken = data->as_BranchData()->not_taken();
822 }
824 // scale the counts to be commensurate with invocation counts:
825 taken = method()->scale_count(taken);
826 not_taken = method()->scale_count(not_taken);
827 }
829 // Give up if too few (or too many, in which case the sum will overflow) counts to be meaningful.
830 // We also check that individual counters are positive first, otherwise the sum can become positive.
831 if (taken < 0 || not_taken < 0 || taken + not_taken < 40) {
832 if (C->log() != NULL) {
833 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d'", iter().get_dest(), taken, not_taken);
834 }
835 return PROB_UNKNOWN;
836 }
838 // Compute frequency that we arrive here
839 float sum = taken + not_taken;
840 // Adjust, if this block is a cloned private block but the
841 // Jump counts are shared. Taken the private counts for
842 // just this path instead of the shared counts.
843 if( block()->count() > 0 )
844 sum = block()->count();
845 cnt = sum / FreqCountInvocations;
847 // Pin probability to sane limits
848 float prob;
849 if( !taken )
850 prob = (0+PROB_MIN) / 2;
851 else if( !not_taken )
852 prob = (1+PROB_MAX) / 2;
853 else { // Compute probability of true path
854 prob = (float)taken / (float)(taken + not_taken);
855 if (prob > PROB_MAX) prob = PROB_MAX;
856 if (prob < PROB_MIN) prob = PROB_MIN;
857 }
859 assert((cnt > 0.0f) && (prob > 0.0f),
860 "Bad frequency assignment in if");
862 if (C->log() != NULL) {
863 const char* prob_str = NULL;
864 if (prob >= PROB_MAX) prob_str = (prob == PROB_MAX) ? "max" : "always";
865 if (prob <= PROB_MIN) prob_str = (prob == PROB_MIN) ? "min" : "never";
866 char prob_str_buf[30];
867 if (prob_str == NULL) {
868 sprintf(prob_str_buf, "%g", prob);
869 prob_str = prob_str_buf;
870 }
871 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d' cnt='%g' prob='%s'",
872 iter().get_dest(), taken, not_taken, cnt, prob_str);
873 }
874 return prob;
875 }
877 //-----------------------------branch_prediction-------------------------------
878 float Parse::branch_prediction(float& cnt,
879 BoolTest::mask btest,
880 int target_bci,
881 Node* test) {
882 float prob = dynamic_branch_prediction(cnt, btest, test);
883 // If prob is unknown, switch to static prediction
884 if (prob != PROB_UNKNOWN) return prob;
886 prob = PROB_FAIR; // Set default value
887 if (btest == BoolTest::eq) // Exactly equal test?
888 prob = PROB_STATIC_INFREQUENT; // Assume its relatively infrequent
889 else if (btest == BoolTest::ne)
890 prob = PROB_STATIC_FREQUENT; // Assume its relatively frequent
892 // If this is a conditional test guarding a backwards branch,
893 // assume its a loop-back edge. Make it a likely taken branch.
894 if (target_bci < bci()) {
895 if (is_osr_parse()) { // Could be a hot OSR'd loop; force deopt
896 // Since it's an OSR, we probably have profile data, but since
897 // branch_prediction returned PROB_UNKNOWN, the counts are too small.
898 // Let's make a special check here for completely zero counts.
899 ciMethodData* methodData = method()->method_data();
900 if (!methodData->is_empty()) {
901 ciProfileData* data = methodData->bci_to_data(bci());
902 // Only stop for truly zero counts, which mean an unknown part
903 // of the OSR-ed method, and we want to deopt to gather more stats.
904 // If you have ANY counts, then this loop is simply 'cold' relative
905 // to the OSR loop.
906 if (data->as_BranchData()->taken() +
907 data->as_BranchData()->not_taken() == 0 ) {
908 // This is the only way to return PROB_UNKNOWN:
909 return PROB_UNKNOWN;
910 }
911 }
912 }
913 prob = PROB_STATIC_FREQUENT; // Likely to take backwards branch
914 }
916 assert(prob != PROB_UNKNOWN, "must have some guess at this point");
917 return prob;
918 }
920 // The magic constants are chosen so as to match the output of
921 // branch_prediction() when the profile reports a zero taken count.
922 // It is important to distinguish zero counts unambiguously, because
923 // some branches (e.g., _213_javac.Assembler.eliminate) validly produce
924 // very small but nonzero probabilities, which if confused with zero
925 // counts would keep the program recompiling indefinitely.
926 bool Parse::seems_never_taken(float prob) const {
927 return prob < PROB_MIN;
928 }
930 // True if the comparison seems to be the kind that will not change its
931 // statistics from true to false. See comments in adjust_map_after_if.
932 // This question is only asked along paths which are already
933 // classifed as untaken (by seems_never_taken), so really,
934 // if a path is never taken, its controlling comparison is
935 // already acting in a stable fashion. If the comparison
936 // seems stable, we will put an expensive uncommon trap
937 // on the untaken path.
938 bool Parse::seems_stable_comparison() const {
939 if (C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if)) {
940 return false;
941 }
942 return true;
943 }
945 //-------------------------------repush_if_args--------------------------------
946 // Push arguments of an "if" bytecode back onto the stack by adjusting _sp.
947 inline int Parse::repush_if_args() {
948 #ifndef PRODUCT
949 if (PrintOpto && WizardMode) {
950 tty->print("defending against excessive implicit null exceptions on %s @%d in ",
951 Bytecodes::name(iter().cur_bc()), iter().cur_bci());
952 method()->print_name(); tty->cr();
953 }
954 #endif
955 int bc_depth = - Bytecodes::depth(iter().cur_bc());
956 assert(bc_depth == 1 || bc_depth == 2, "only two kinds of branches");
957 DEBUG_ONLY(sync_jvms()); // argument(n) requires a synced jvms
958 assert(argument(0) != NULL, "must exist");
959 assert(bc_depth == 1 || argument(1) != NULL, "two must exist");
960 inc_sp(bc_depth);
961 return bc_depth;
962 }
964 //----------------------------------do_ifnull----------------------------------
965 void Parse::do_ifnull(BoolTest::mask btest, Node *c) {
966 int target_bci = iter().get_dest();
968 Block* branch_block = successor_for_bci(target_bci);
969 Block* next_block = successor_for_bci(iter().next_bci());
971 float cnt;
972 float prob = branch_prediction(cnt, btest, target_bci, c);
973 if (prob == PROB_UNKNOWN) {
974 // (An earlier version of do_ifnull omitted this trap for OSR methods.)
975 #ifndef PRODUCT
976 if (PrintOpto && Verbose)
977 tty->print_cr("Never-taken edge stops compilation at bci %d",bci());
978 #endif
979 repush_if_args(); // to gather stats on loop
980 // We need to mark this branch as taken so that if we recompile we will
981 // see that it is possible. In the tiered system the interpreter doesn't
982 // do profiling and by the time we get to the lower tier from the interpreter
983 // the path may be cold again. Make sure it doesn't look untaken
984 profile_taken_branch(target_bci, !ProfileInterpreter);
985 uncommon_trap(Deoptimization::Reason_unreached,
986 Deoptimization::Action_reinterpret,
987 NULL, "cold");
988 if (C->eliminate_boxing()) {
989 // Mark the successor blocks as parsed
990 branch_block->next_path_num();
991 next_block->next_path_num();
992 }
993 return;
994 }
996 explicit_null_checks_inserted++;
998 // Generate real control flow
999 Node *tst = _gvn.transform( new (C) BoolNode( c, btest ) );
1001 // Sanity check the probability value
1002 assert(prob > 0.0f,"Bad probability in Parser");
1003 // Need xform to put node in hash table
1004 IfNode *iff = create_and_xform_if( control(), tst, prob, cnt );
1005 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser");
1006 // True branch
1007 { PreserveJVMState pjvms(this);
1008 Node* iftrue = _gvn.transform( new (C) IfTrueNode (iff) );
1009 set_control(iftrue);
1011 if (stopped()) { // Path is dead?
1012 explicit_null_checks_elided++;
1013 if (C->eliminate_boxing()) {
1014 // Mark the successor block as parsed
1015 branch_block->next_path_num();
1016 }
1017 } else { // Path is live.
1018 // Update method data
1019 profile_taken_branch(target_bci);
1020 adjust_map_after_if(btest, c, prob, branch_block, next_block);
1021 if (!stopped()) {
1022 merge(target_bci);
1023 }
1024 }
1025 }
1027 // False branch
1028 Node* iffalse = _gvn.transform( new (C) IfFalseNode(iff) );
1029 set_control(iffalse);
1031 if (stopped()) { // Path is dead?
1032 explicit_null_checks_elided++;
1033 if (C->eliminate_boxing()) {
1034 // Mark the successor block as parsed
1035 next_block->next_path_num();
1036 }
1037 } else { // Path is live.
1038 // Update method data
1039 profile_not_taken_branch();
1040 adjust_map_after_if(BoolTest(btest).negate(), c, 1.0-prob,
1041 next_block, branch_block);
1042 }
1043 }
1045 //------------------------------------do_if------------------------------------
1046 void Parse::do_if(BoolTest::mask btest, Node* c) {
1047 int target_bci = iter().get_dest();
1049 Block* branch_block = successor_for_bci(target_bci);
1050 Block* next_block = successor_for_bci(iter().next_bci());
1052 float cnt;
1053 float prob = branch_prediction(cnt, btest, target_bci, c);
1054 float untaken_prob = 1.0 - prob;
1056 if (prob == PROB_UNKNOWN) {
1057 #ifndef PRODUCT
1058 if (PrintOpto && Verbose)
1059 tty->print_cr("Never-taken edge stops compilation at bci %d",bci());
1060 #endif
1061 repush_if_args(); // to gather stats on loop
1062 // We need to mark this branch as taken so that if we recompile we will
1063 // see that it is possible. In the tiered system the interpreter doesn't
1064 // do profiling and by the time we get to the lower tier from the interpreter
1065 // the path may be cold again. Make sure it doesn't look untaken
1066 profile_taken_branch(target_bci, !ProfileInterpreter);
1067 uncommon_trap(Deoptimization::Reason_unreached,
1068 Deoptimization::Action_reinterpret,
1069 NULL, "cold");
1070 if (C->eliminate_boxing()) {
1071 // Mark the successor blocks as parsed
1072 branch_block->next_path_num();
1073 next_block->next_path_num();
1074 }
1075 return;
1076 }
1078 // Sanity check the probability value
1079 assert(0.0f < prob && prob < 1.0f,"Bad probability in Parser");
1081 bool taken_if_true = true;
1082 // Convert BoolTest to canonical form:
1083 if (!BoolTest(btest).is_canonical()) {
1084 btest = BoolTest(btest).negate();
1085 taken_if_true = false;
1086 // prob is NOT updated here; it remains the probability of the taken
1087 // path (as opposed to the prob of the path guarded by an 'IfTrueNode').
1088 }
1089 assert(btest != BoolTest::eq, "!= is the only canonical exact test");
1091 Node* tst0 = new (C) BoolNode(c, btest);
1092 Node* tst = _gvn.transform(tst0);
1093 BoolTest::mask taken_btest = BoolTest::illegal;
1094 BoolTest::mask untaken_btest = BoolTest::illegal;
1096 if (tst->is_Bool()) {
1097 // Refresh c from the transformed bool node, since it may be
1098 // simpler than the original c. Also re-canonicalize btest.
1099 // This wins when (Bool ne (Conv2B p) 0) => (Bool ne (CmpP p NULL)).
1100 // That can arise from statements like: if (x instanceof C) ...
1101 if (tst != tst0) {
1102 // Canonicalize one more time since transform can change it.
1103 btest = tst->as_Bool()->_test._test;
1104 if (!BoolTest(btest).is_canonical()) {
1105 // Reverse edges one more time...
1106 tst = _gvn.transform( tst->as_Bool()->negate(&_gvn) );
1107 btest = tst->as_Bool()->_test._test;
1108 assert(BoolTest(btest).is_canonical(), "sanity");
1109 taken_if_true = !taken_if_true;
1110 }
1111 c = tst->in(1);
1112 }
1113 BoolTest::mask neg_btest = BoolTest(btest).negate();
1114 taken_btest = taken_if_true ? btest : neg_btest;
1115 untaken_btest = taken_if_true ? neg_btest : btest;
1116 }
1118 // Generate real control flow
1119 float true_prob = (taken_if_true ? prob : untaken_prob);
1120 IfNode* iff = create_and_map_if(control(), tst, true_prob, cnt);
1121 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser");
1122 Node* taken_branch = new (C) IfTrueNode(iff);
1123 Node* untaken_branch = new (C) IfFalseNode(iff);
1124 if (!taken_if_true) { // Finish conversion to canonical form
1125 Node* tmp = taken_branch;
1126 taken_branch = untaken_branch;
1127 untaken_branch = tmp;
1128 }
1130 // Branch is taken:
1131 { PreserveJVMState pjvms(this);
1132 taken_branch = _gvn.transform(taken_branch);
1133 set_control(taken_branch);
1135 if (stopped()) {
1136 if (C->eliminate_boxing()) {
1137 // Mark the successor block as parsed
1138 branch_block->next_path_num();
1139 }
1140 } else {
1141 // Update method data
1142 profile_taken_branch(target_bci);
1143 adjust_map_after_if(taken_btest, c, prob, branch_block, next_block);
1144 if (!stopped()) {
1145 merge(target_bci);
1146 }
1147 }
1148 }
1150 untaken_branch = _gvn.transform(untaken_branch);
1151 set_control(untaken_branch);
1153 // Branch not taken.
1154 if (stopped()) {
1155 if (C->eliminate_boxing()) {
1156 // Mark the successor block as parsed
1157 next_block->next_path_num();
1158 }
1159 } else {
1160 // Update method data
1161 profile_not_taken_branch();
1162 adjust_map_after_if(untaken_btest, c, untaken_prob,
1163 next_block, branch_block);
1164 }
1165 }
1167 bool Parse::path_is_suitable_for_uncommon_trap(float prob) const {
1168 // Don't want to speculate on uncommon traps when running with -Xcomp
1169 if (!UseInterpreter) {
1170 return false;
1171 }
1172 return (seems_never_taken(prob) && seems_stable_comparison());
1173 }
1175 //----------------------------adjust_map_after_if------------------------------
1176 // Adjust the JVM state to reflect the result of taking this path.
1177 // Basically, it means inspecting the CmpNode controlling this
1178 // branch, seeing how it constrains a tested value, and then
1179 // deciding if it's worth our while to encode this constraint
1180 // as graph nodes in the current abstract interpretation map.
1181 void Parse::adjust_map_after_if(BoolTest::mask btest, Node* c, float prob,
1182 Block* path, Block* other_path) {
1183 if (stopped() || !c->is_Cmp() || btest == BoolTest::illegal)
1184 return; // nothing to do
1186 bool is_fallthrough = (path == successor_for_bci(iter().next_bci()));
1188 if (path_is_suitable_for_uncommon_trap(prob)) {
1189 repush_if_args();
1190 uncommon_trap(Deoptimization::Reason_unstable_if,
1191 Deoptimization::Action_reinterpret,
1192 NULL,
1193 (is_fallthrough ? "taken always" : "taken never"));
1194 return;
1195 }
1197 Node* val = c->in(1);
1198 Node* con = c->in(2);
1199 const Type* tcon = _gvn.type(con);
1200 const Type* tval = _gvn.type(val);
1201 bool have_con = tcon->singleton();
1202 if (tval->singleton()) {
1203 if (!have_con) {
1204 // Swap, so constant is in con.
1205 con = val;
1206 tcon = tval;
1207 val = c->in(2);
1208 tval = _gvn.type(val);
1209 btest = BoolTest(btest).commute();
1210 have_con = true;
1211 } else {
1212 // Do we have two constants? Then leave well enough alone.
1213 have_con = false;
1214 }
1215 }
1216 if (!have_con) // remaining adjustments need a con
1217 return;
1219 sharpen_type_after_if(btest, con, tcon, val, tval);
1220 }
1223 static Node* extract_obj_from_klass_load(PhaseGVN* gvn, Node* n) {
1224 Node* ldk;
1225 if (n->is_DecodeNKlass()) {
1226 if (n->in(1)->Opcode() != Op_LoadNKlass) {
1227 return NULL;
1228 } else {
1229 ldk = n->in(1);
1230 }
1231 } else if (n->Opcode() != Op_LoadKlass) {
1232 return NULL;
1233 } else {
1234 ldk = n;
1235 }
1236 assert(ldk != NULL && ldk->is_Load(), "should have found a LoadKlass or LoadNKlass node");
1238 Node* adr = ldk->in(MemNode::Address);
1239 intptr_t off = 0;
1240 Node* obj = AddPNode::Ideal_base_and_offset(adr, gvn, off);
1241 if (obj == NULL || off != oopDesc::klass_offset_in_bytes()) // loading oopDesc::_klass?
1242 return NULL;
1243 const TypePtr* tp = gvn->type(obj)->is_ptr();
1244 if (tp == NULL || !(tp->isa_instptr() || tp->isa_aryptr())) // is obj a Java object ptr?
1245 return NULL;
1247 return obj;
1248 }
1250 void Parse::sharpen_type_after_if(BoolTest::mask btest,
1251 Node* con, const Type* tcon,
1252 Node* val, const Type* tval) {
1253 // Look for opportunities to sharpen the type of a node
1254 // whose klass is compared with a constant klass.
1255 if (btest == BoolTest::eq && tcon->isa_klassptr()) {
1256 Node* obj = extract_obj_from_klass_load(&_gvn, val);
1257 const TypeOopPtr* con_type = tcon->isa_klassptr()->as_instance_type();
1258 if (obj != NULL && (con_type->isa_instptr() || con_type->isa_aryptr())) {
1259 // Found:
1260 // Bool(CmpP(LoadKlass(obj._klass), ConP(Foo.klass)), [eq])
1261 // or the narrowOop equivalent.
1262 const Type* obj_type = _gvn.type(obj);
1263 const TypeOopPtr* tboth = obj_type->join_speculative(con_type)->isa_oopptr();
1264 if (tboth != NULL && tboth->klass_is_exact() && tboth != obj_type &&
1265 tboth->higher_equal(obj_type)) {
1266 // obj has to be of the exact type Foo if the CmpP succeeds.
1267 int obj_in_map = map()->find_edge(obj);
1268 JVMState* jvms = this->jvms();
1269 if (obj_in_map >= 0 &&
1270 (jvms->is_loc(obj_in_map) || jvms->is_stk(obj_in_map))) {
1271 TypeNode* ccast = new (C) CheckCastPPNode(control(), obj, tboth);
1272 const Type* tcc = ccast->as_Type()->type();
1273 assert(tcc != obj_type && tcc->higher_equal_speculative(obj_type), "must improve");
1274 // Delay transform() call to allow recovery of pre-cast value
1275 // at the control merge.
1276 _gvn.set_type_bottom(ccast);
1277 record_for_igvn(ccast);
1278 // Here's the payoff.
1279 replace_in_map(obj, ccast);
1280 }
1281 }
1282 }
1283 }
1285 int val_in_map = map()->find_edge(val);
1286 if (val_in_map < 0) return; // replace_in_map would be useless
1287 {
1288 JVMState* jvms = this->jvms();
1289 if (!(jvms->is_loc(val_in_map) ||
1290 jvms->is_stk(val_in_map)))
1291 return; // again, it would be useless
1292 }
1294 // Check for a comparison to a constant, and "know" that the compared
1295 // value is constrained on this path.
1296 assert(tcon->singleton(), "");
1297 ConstraintCastNode* ccast = NULL;
1298 Node* cast = NULL;
1300 switch (btest) {
1301 case BoolTest::eq: // Constant test?
1302 {
1303 const Type* tboth = tcon->join_speculative(tval);
1304 if (tboth == tval) break; // Nothing to gain.
1305 if (tcon->isa_int()) {
1306 ccast = new (C) CastIINode(val, tboth);
1307 } else if (tcon == TypePtr::NULL_PTR) {
1308 // Cast to null, but keep the pointer identity temporarily live.
1309 ccast = new (C) CastPPNode(val, tboth);
1310 } else {
1311 const TypeF* tf = tcon->isa_float_constant();
1312 const TypeD* td = tcon->isa_double_constant();
1313 // Exclude tests vs float/double 0 as these could be
1314 // either +0 or -0. Just because you are equal to +0
1315 // doesn't mean you ARE +0!
1316 // Note, following code also replaces Long and Oop values.
1317 if ((!tf || tf->_f != 0.0) &&
1318 (!td || td->_d != 0.0))
1319 cast = con; // Replace non-constant val by con.
1320 }
1321 }
1322 break;
1324 case BoolTest::ne:
1325 if (tcon == TypePtr::NULL_PTR) {
1326 cast = cast_not_null(val, false);
1327 }
1328 break;
1330 default:
1331 // (At this point we could record int range types with CastII.)
1332 break;
1333 }
1335 if (ccast != NULL) {
1336 const Type* tcc = ccast->as_Type()->type();
1337 assert(tcc != tval && tcc->higher_equal_speculative(tval), "must improve");
1338 // Delay transform() call to allow recovery of pre-cast value
1339 // at the control merge.
1340 ccast->set_req(0, control());
1341 _gvn.set_type_bottom(ccast);
1342 record_for_igvn(ccast);
1343 cast = ccast;
1344 }
1346 if (cast != NULL) { // Here's the payoff.
1347 replace_in_map(val, cast);
1348 }
1349 }
1351 /**
1352 * Use speculative type to optimize CmpP node: if comparison is
1353 * against the low level class, cast the object to the speculative
1354 * type if any. CmpP should then go away.
1355 *
1356 * @param c expected CmpP node
1357 * @return result of CmpP on object casted to speculative type
1358 *
1359 */
1360 Node* Parse::optimize_cmp_with_klass(Node* c) {
1361 // If this is transformed by the _gvn to a comparison with the low
1362 // level klass then we may be able to use speculation
1363 if (c->Opcode() == Op_CmpP &&
1364 (c->in(1)->Opcode() == Op_LoadKlass || c->in(1)->Opcode() == Op_DecodeNKlass) &&
1365 c->in(2)->is_Con()) {
1366 Node* load_klass = NULL;
1367 Node* decode = NULL;
1368 if (c->in(1)->Opcode() == Op_DecodeNKlass) {
1369 decode = c->in(1);
1370 load_klass = c->in(1)->in(1);
1371 } else {
1372 load_klass = c->in(1);
1373 }
1374 if (load_klass->in(2)->is_AddP()) {
1375 Node* addp = load_klass->in(2);
1376 Node* obj = addp->in(AddPNode::Address);
1377 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
1378 if (obj_type->speculative_type() != NULL) {
1379 ciKlass* k = obj_type->speculative_type();
1380 inc_sp(2);
1381 obj = maybe_cast_profiled_obj(obj, k);
1382 dec_sp(2);
1383 // Make the CmpP use the casted obj
1384 addp = basic_plus_adr(obj, addp->in(AddPNode::Offset));
1385 load_klass = load_klass->clone();
1386 load_klass->set_req(2, addp);
1387 load_klass = _gvn.transform(load_klass);
1388 if (decode != NULL) {
1389 decode = decode->clone();
1390 decode->set_req(1, load_klass);
1391 load_klass = _gvn.transform(decode);
1392 }
1393 c = c->clone();
1394 c->set_req(1, load_klass);
1395 c = _gvn.transform(c);
1396 }
1397 }
1398 }
1399 return c;
1400 }
1402 //------------------------------do_one_bytecode--------------------------------
1403 // Parse this bytecode, and alter the Parsers JVM->Node mapping
1404 void Parse::do_one_bytecode() {
1405 Node *a, *b, *c, *d; // Handy temps
1406 BoolTest::mask btest;
1407 int i;
1409 assert(!has_exceptions(), "bytecode entry state must be clear of throws");
1411 if (C->check_node_count(NodeLimitFudgeFactor * 5,
1412 "out of nodes parsing method")) {
1413 return;
1414 }
1416 #ifdef ASSERT
1417 // for setting breakpoints
1418 if (TraceOptoParse) {
1419 tty->print(" @");
1420 dump_bci(bci());
1421 tty->cr();
1422 }
1423 #endif
1425 switch (bc()) {
1426 case Bytecodes::_nop:
1427 // do nothing
1428 break;
1429 case Bytecodes::_lconst_0:
1430 push_pair(longcon(0));
1431 break;
1433 case Bytecodes::_lconst_1:
1434 push_pair(longcon(1));
1435 break;
1437 case Bytecodes::_fconst_0:
1438 push(zerocon(T_FLOAT));
1439 break;
1441 case Bytecodes::_fconst_1:
1442 push(makecon(TypeF::ONE));
1443 break;
1445 case Bytecodes::_fconst_2:
1446 push(makecon(TypeF::make(2.0f)));
1447 break;
1449 case Bytecodes::_dconst_0:
1450 push_pair(zerocon(T_DOUBLE));
1451 break;
1453 case Bytecodes::_dconst_1:
1454 push_pair(makecon(TypeD::ONE));
1455 break;
1457 case Bytecodes::_iconst_m1:push(intcon(-1)); break;
1458 case Bytecodes::_iconst_0: push(intcon( 0)); break;
1459 case Bytecodes::_iconst_1: push(intcon( 1)); break;
1460 case Bytecodes::_iconst_2: push(intcon( 2)); break;
1461 case Bytecodes::_iconst_3: push(intcon( 3)); break;
1462 case Bytecodes::_iconst_4: push(intcon( 4)); break;
1463 case Bytecodes::_iconst_5: push(intcon( 5)); break;
1464 case Bytecodes::_bipush: push(intcon(iter().get_constant_u1())); break;
1465 case Bytecodes::_sipush: push(intcon(iter().get_constant_u2())); break;
1466 case Bytecodes::_aconst_null: push(null()); break;
1467 case Bytecodes::_ldc:
1468 case Bytecodes::_ldc_w:
1469 case Bytecodes::_ldc2_w:
1470 // If the constant is unresolved, run this BC once in the interpreter.
1471 {
1472 ciConstant constant = iter().get_constant();
1473 if (constant.basic_type() == T_OBJECT &&
1474 !constant.as_object()->is_loaded()) {
1475 int index = iter().get_constant_pool_index();
1476 constantTag tag = iter().get_constant_pool_tag(index);
1477 uncommon_trap(Deoptimization::make_trap_request
1478 (Deoptimization::Reason_unloaded,
1479 Deoptimization::Action_reinterpret,
1480 index),
1481 NULL, tag.internal_name());
1482 break;
1483 }
1484 assert(constant.basic_type() != T_OBJECT || constant.as_object()->is_instance(),
1485 "must be java_mirror of klass");
1486 bool pushed = push_constant(constant, true);
1487 guarantee(pushed, "must be possible to push this constant");
1488 }
1490 break;
1492 case Bytecodes::_aload_0:
1493 push( local(0) );
1494 break;
1495 case Bytecodes::_aload_1:
1496 push( local(1) );
1497 break;
1498 case Bytecodes::_aload_2:
1499 push( local(2) );
1500 break;
1501 case Bytecodes::_aload_3:
1502 push( local(3) );
1503 break;
1504 case Bytecodes::_aload:
1505 push( local(iter().get_index()) );
1506 break;
1508 case Bytecodes::_fload_0:
1509 case Bytecodes::_iload_0:
1510 push( local(0) );
1511 break;
1512 case Bytecodes::_fload_1:
1513 case Bytecodes::_iload_1:
1514 push( local(1) );
1515 break;
1516 case Bytecodes::_fload_2:
1517 case Bytecodes::_iload_2:
1518 push( local(2) );
1519 break;
1520 case Bytecodes::_fload_3:
1521 case Bytecodes::_iload_3:
1522 push( local(3) );
1523 break;
1524 case Bytecodes::_fload:
1525 case Bytecodes::_iload:
1526 push( local(iter().get_index()) );
1527 break;
1528 case Bytecodes::_lload_0:
1529 push_pair_local( 0 );
1530 break;
1531 case Bytecodes::_lload_1:
1532 push_pair_local( 1 );
1533 break;
1534 case Bytecodes::_lload_2:
1535 push_pair_local( 2 );
1536 break;
1537 case Bytecodes::_lload_3:
1538 push_pair_local( 3 );
1539 break;
1540 case Bytecodes::_lload:
1541 push_pair_local( iter().get_index() );
1542 break;
1544 case Bytecodes::_dload_0:
1545 push_pair_local(0);
1546 break;
1547 case Bytecodes::_dload_1:
1548 push_pair_local(1);
1549 break;
1550 case Bytecodes::_dload_2:
1551 push_pair_local(2);
1552 break;
1553 case Bytecodes::_dload_3:
1554 push_pair_local(3);
1555 break;
1556 case Bytecodes::_dload:
1557 push_pair_local(iter().get_index());
1558 break;
1559 case Bytecodes::_fstore_0:
1560 case Bytecodes::_istore_0:
1561 case Bytecodes::_astore_0:
1562 set_local( 0, pop() );
1563 break;
1564 case Bytecodes::_fstore_1:
1565 case Bytecodes::_istore_1:
1566 case Bytecodes::_astore_1:
1567 set_local( 1, pop() );
1568 break;
1569 case Bytecodes::_fstore_2:
1570 case Bytecodes::_istore_2:
1571 case Bytecodes::_astore_2:
1572 set_local( 2, pop() );
1573 break;
1574 case Bytecodes::_fstore_3:
1575 case Bytecodes::_istore_3:
1576 case Bytecodes::_astore_3:
1577 set_local( 3, pop() );
1578 break;
1579 case Bytecodes::_fstore:
1580 case Bytecodes::_istore:
1581 case Bytecodes::_astore:
1582 set_local( iter().get_index(), pop() );
1583 break;
1584 // long stores
1585 case Bytecodes::_lstore_0:
1586 set_pair_local( 0, pop_pair() );
1587 break;
1588 case Bytecodes::_lstore_1:
1589 set_pair_local( 1, pop_pair() );
1590 break;
1591 case Bytecodes::_lstore_2:
1592 set_pair_local( 2, pop_pair() );
1593 break;
1594 case Bytecodes::_lstore_3:
1595 set_pair_local( 3, pop_pair() );
1596 break;
1597 case Bytecodes::_lstore:
1598 set_pair_local( iter().get_index(), pop_pair() );
1599 break;
1601 // double stores
1602 case Bytecodes::_dstore_0:
1603 set_pair_local( 0, dstore_rounding(pop_pair()) );
1604 break;
1605 case Bytecodes::_dstore_1:
1606 set_pair_local( 1, dstore_rounding(pop_pair()) );
1607 break;
1608 case Bytecodes::_dstore_2:
1609 set_pair_local( 2, dstore_rounding(pop_pair()) );
1610 break;
1611 case Bytecodes::_dstore_3:
1612 set_pair_local( 3, dstore_rounding(pop_pair()) );
1613 break;
1614 case Bytecodes::_dstore:
1615 set_pair_local( iter().get_index(), dstore_rounding(pop_pair()) );
1616 break;
1618 case Bytecodes::_pop: dec_sp(1); break;
1619 case Bytecodes::_pop2: dec_sp(2); break;
1620 case Bytecodes::_swap:
1621 a = pop();
1622 b = pop();
1623 push(a);
1624 push(b);
1625 break;
1626 case Bytecodes::_dup:
1627 a = pop();
1628 push(a);
1629 push(a);
1630 break;
1631 case Bytecodes::_dup_x1:
1632 a = pop();
1633 b = pop();
1634 push( a );
1635 push( b );
1636 push( a );
1637 break;
1638 case Bytecodes::_dup_x2:
1639 a = pop();
1640 b = pop();
1641 c = pop();
1642 push( a );
1643 push( c );
1644 push( b );
1645 push( a );
1646 break;
1647 case Bytecodes::_dup2:
1648 a = pop();
1649 b = pop();
1650 push( b );
1651 push( a );
1652 push( b );
1653 push( a );
1654 break;
1656 case Bytecodes::_dup2_x1:
1657 // before: .. c, b, a
1658 // after: .. b, a, c, b, a
1659 // not tested
1660 a = pop();
1661 b = pop();
1662 c = pop();
1663 push( b );
1664 push( a );
1665 push( c );
1666 push( b );
1667 push( a );
1668 break;
1669 case Bytecodes::_dup2_x2:
1670 // before: .. d, c, b, a
1671 // after: .. b, a, d, c, b, a
1672 // not tested
1673 a = pop();
1674 b = pop();
1675 c = pop();
1676 d = pop();
1677 push( b );
1678 push( a );
1679 push( d );
1680 push( c );
1681 push( b );
1682 push( a );
1683 break;
1685 case Bytecodes::_arraylength: {
1686 // Must do null-check with value on expression stack
1687 Node *ary = null_check(peek(), T_ARRAY);
1688 // Compile-time detect of null-exception?
1689 if (stopped()) return;
1690 a = pop();
1691 push(load_array_length(a));
1692 break;
1693 }
1695 case Bytecodes::_baload: array_load(T_BYTE); break;
1696 case Bytecodes::_caload: array_load(T_CHAR); break;
1697 case Bytecodes::_iaload: array_load(T_INT); break;
1698 case Bytecodes::_saload: array_load(T_SHORT); break;
1699 case Bytecodes::_faload: array_load(T_FLOAT); break;
1700 case Bytecodes::_aaload: array_load(T_OBJECT); break;
1701 case Bytecodes::_laload: {
1702 a = array_addressing(T_LONG, 0);
1703 if (stopped()) return; // guaranteed null or range check
1704 dec_sp(2); // Pop array and index
1705 push_pair(make_load(control(), a, TypeLong::LONG, T_LONG, TypeAryPtr::LONGS, MemNode::unordered));
1706 break;
1707 }
1708 case Bytecodes::_daload: {
1709 a = array_addressing(T_DOUBLE, 0);
1710 if (stopped()) return; // guaranteed null or range check
1711 dec_sp(2); // Pop array and index
1712 push_pair(make_load(control(), a, Type::DOUBLE, T_DOUBLE, TypeAryPtr::DOUBLES, MemNode::unordered));
1713 break;
1714 }
1715 case Bytecodes::_bastore: array_store(T_BYTE); break;
1716 case Bytecodes::_castore: array_store(T_CHAR); break;
1717 case Bytecodes::_iastore: array_store(T_INT); break;
1718 case Bytecodes::_sastore: array_store(T_SHORT); break;
1719 case Bytecodes::_fastore: array_store(T_FLOAT); break;
1720 case Bytecodes::_aastore: {
1721 d = array_addressing(T_OBJECT, 1);
1722 if (stopped()) return; // guaranteed null or range check
1723 array_store_check();
1724 c = pop(); // Oop to store
1725 b = pop(); // index (already used)
1726 a = pop(); // the array itself
1727 const TypeOopPtr* elemtype = _gvn.type(a)->is_aryptr()->elem()->make_oopptr();
1728 const TypeAryPtr* adr_type = TypeAryPtr::OOPS;
1729 Node* store = store_oop_to_array(control(), a, d, adr_type, c, elemtype, T_OBJECT, MemNode::release);
1730 break;
1731 }
1732 case Bytecodes::_lastore: {
1733 a = array_addressing(T_LONG, 2);
1734 if (stopped()) return; // guaranteed null or range check
1735 c = pop_pair();
1736 dec_sp(2); // Pop array and index
1737 store_to_memory(control(), a, c, T_LONG, TypeAryPtr::LONGS, MemNode::unordered);
1738 break;
1739 }
1740 case Bytecodes::_dastore: {
1741 a = array_addressing(T_DOUBLE, 2);
1742 if (stopped()) return; // guaranteed null or range check
1743 c = pop_pair();
1744 dec_sp(2); // Pop array and index
1745 c = dstore_rounding(c);
1746 store_to_memory(control(), a, c, T_DOUBLE, TypeAryPtr::DOUBLES, MemNode::unordered);
1747 break;
1748 }
1749 case Bytecodes::_getfield:
1750 do_getfield();
1751 break;
1753 case Bytecodes::_getstatic:
1754 do_getstatic();
1755 break;
1757 case Bytecodes::_putfield:
1758 do_putfield();
1759 break;
1761 case Bytecodes::_putstatic:
1762 do_putstatic();
1763 break;
1765 case Bytecodes::_irem:
1766 do_irem();
1767 break;
1768 case Bytecodes::_idiv:
1769 // Must keep both values on the expression-stack during null-check
1770 zero_check_int(peek());
1771 // Compile-time detect of null-exception?
1772 if (stopped()) return;
1773 b = pop();
1774 a = pop();
1775 push( _gvn.transform( new (C) DivINode(control(),a,b) ) );
1776 break;
1777 case Bytecodes::_imul:
1778 b = pop(); a = pop();
1779 push( _gvn.transform( new (C) MulINode(a,b) ) );
1780 break;
1781 case Bytecodes::_iadd:
1782 b = pop(); a = pop();
1783 push( _gvn.transform( new (C) AddINode(a,b) ) );
1784 break;
1785 case Bytecodes::_ineg:
1786 a = pop();
1787 push( _gvn.transform( new (C) SubINode(_gvn.intcon(0),a)) );
1788 break;
1789 case Bytecodes::_isub:
1790 b = pop(); a = pop();
1791 push( _gvn.transform( new (C) SubINode(a,b) ) );
1792 break;
1793 case Bytecodes::_iand:
1794 b = pop(); a = pop();
1795 push( _gvn.transform( new (C) AndINode(a,b) ) );
1796 break;
1797 case Bytecodes::_ior:
1798 b = pop(); a = pop();
1799 push( _gvn.transform( new (C) OrINode(a,b) ) );
1800 break;
1801 case Bytecodes::_ixor:
1802 b = pop(); a = pop();
1803 push( _gvn.transform( new (C) XorINode(a,b) ) );
1804 break;
1805 case Bytecodes::_ishl:
1806 b = pop(); a = pop();
1807 push( _gvn.transform( new (C) LShiftINode(a,b) ) );
1808 break;
1809 case Bytecodes::_ishr:
1810 b = pop(); a = pop();
1811 push( _gvn.transform( new (C) RShiftINode(a,b) ) );
1812 break;
1813 case Bytecodes::_iushr:
1814 b = pop(); a = pop();
1815 push( _gvn.transform( new (C) URShiftINode(a,b) ) );
1816 break;
1818 case Bytecodes::_fneg:
1819 a = pop();
1820 b = _gvn.transform(new (C) NegFNode (a));
1821 push(b);
1822 break;
1824 case Bytecodes::_fsub:
1825 b = pop();
1826 a = pop();
1827 c = _gvn.transform( new (C) SubFNode(a,b) );
1828 d = precision_rounding(c);
1829 push( d );
1830 break;
1832 case Bytecodes::_fadd:
1833 b = pop();
1834 a = pop();
1835 c = _gvn.transform( new (C) AddFNode(a,b) );
1836 d = precision_rounding(c);
1837 push( d );
1838 break;
1840 case Bytecodes::_fmul:
1841 b = pop();
1842 a = pop();
1843 c = _gvn.transform( new (C) MulFNode(a,b) );
1844 d = precision_rounding(c);
1845 push( d );
1846 break;
1848 case Bytecodes::_fdiv:
1849 b = pop();
1850 a = pop();
1851 c = _gvn.transform( new (C) DivFNode(0,a,b) );
1852 d = precision_rounding(c);
1853 push( d );
1854 break;
1856 case Bytecodes::_frem:
1857 if (Matcher::has_match_rule(Op_ModF)) {
1858 // Generate a ModF node.
1859 b = pop();
1860 a = pop();
1861 c = _gvn.transform( new (C) ModFNode(0,a,b) );
1862 d = precision_rounding(c);
1863 push( d );
1864 }
1865 else {
1866 // Generate a call.
1867 modf();
1868 }
1869 break;
1871 case Bytecodes::_fcmpl:
1872 b = pop();
1873 a = pop();
1874 c = _gvn.transform( new (C) CmpF3Node( a, b));
1875 push(c);
1876 break;
1877 case Bytecodes::_fcmpg:
1878 b = pop();
1879 a = pop();
1881 // Same as fcmpl but need to flip the unordered case. Swap the inputs,
1882 // which negates the result sign except for unordered. Flip the unordered
1883 // as well by using CmpF3 which implements unordered-lesser instead of
1884 // unordered-greater semantics. Finally, commute the result bits. Result
1885 // is same as using a CmpF3Greater except we did it with CmpF3 alone.
1886 c = _gvn.transform( new (C) CmpF3Node( b, a));
1887 c = _gvn.transform( new (C) SubINode(_gvn.intcon(0),c) );
1888 push(c);
1889 break;
1891 case Bytecodes::_f2i:
1892 a = pop();
1893 push(_gvn.transform(new (C) ConvF2INode(a)));
1894 break;
1896 case Bytecodes::_d2i:
1897 a = pop_pair();
1898 b = _gvn.transform(new (C) ConvD2INode(a));
1899 push( b );
1900 break;
1902 case Bytecodes::_f2d:
1903 a = pop();
1904 b = _gvn.transform( new (C) ConvF2DNode(a));
1905 push_pair( b );
1906 break;
1908 case Bytecodes::_d2f:
1909 a = pop_pair();
1910 b = _gvn.transform( new (C) ConvD2FNode(a));
1911 // This breaks _227_mtrt (speed & correctness) and _222_mpegaudio (speed)
1912 //b = _gvn.transform(new (C) RoundFloatNode(0, b) );
1913 push( b );
1914 break;
1916 case Bytecodes::_l2f:
1917 if (Matcher::convL2FSupported()) {
1918 a = pop_pair();
1919 b = _gvn.transform( new (C) ConvL2FNode(a));
1920 // For i486.ad, FILD doesn't restrict precision to 24 or 53 bits.
1921 // Rather than storing the result into an FP register then pushing
1922 // out to memory to round, the machine instruction that implements
1923 // ConvL2D is responsible for rounding.
1924 // c = precision_rounding(b);
1925 c = _gvn.transform(b);
1926 push(c);
1927 } else {
1928 l2f();
1929 }
1930 break;
1932 case Bytecodes::_l2d:
1933 a = pop_pair();
1934 b = _gvn.transform( new (C) ConvL2DNode(a));
1935 // For i486.ad, rounding is always necessary (see _l2f above).
1936 // c = dprecision_rounding(b);
1937 c = _gvn.transform(b);
1938 push_pair(c);
1939 break;
1941 case Bytecodes::_f2l:
1942 a = pop();
1943 b = _gvn.transform( new (C) ConvF2LNode(a));
1944 push_pair(b);
1945 break;
1947 case Bytecodes::_d2l:
1948 a = pop_pair();
1949 b = _gvn.transform( new (C) ConvD2LNode(a));
1950 push_pair(b);
1951 break;
1953 case Bytecodes::_dsub:
1954 b = pop_pair();
1955 a = pop_pair();
1956 c = _gvn.transform( new (C) SubDNode(a,b) );
1957 d = dprecision_rounding(c);
1958 push_pair( d );
1959 break;
1961 case Bytecodes::_dadd:
1962 b = pop_pair();
1963 a = pop_pair();
1964 c = _gvn.transform( new (C) AddDNode(a,b) );
1965 d = dprecision_rounding(c);
1966 push_pair( d );
1967 break;
1969 case Bytecodes::_dmul:
1970 b = pop_pair();
1971 a = pop_pair();
1972 c = _gvn.transform( new (C) MulDNode(a,b) );
1973 d = dprecision_rounding(c);
1974 push_pair( d );
1975 break;
1977 case Bytecodes::_ddiv:
1978 b = pop_pair();
1979 a = pop_pair();
1980 c = _gvn.transform( new (C) DivDNode(0,a,b) );
1981 d = dprecision_rounding(c);
1982 push_pair( d );
1983 break;
1985 case Bytecodes::_dneg:
1986 a = pop_pair();
1987 b = _gvn.transform(new (C) NegDNode (a));
1988 push_pair(b);
1989 break;
1991 case Bytecodes::_drem:
1992 if (Matcher::has_match_rule(Op_ModD)) {
1993 // Generate a ModD node.
1994 b = pop_pair();
1995 a = pop_pair();
1996 // a % b
1998 c = _gvn.transform( new (C) ModDNode(0,a,b) );
1999 d = dprecision_rounding(c);
2000 push_pair( d );
2001 }
2002 else {
2003 // Generate a call.
2004 modd();
2005 }
2006 break;
2008 case Bytecodes::_dcmpl:
2009 b = pop_pair();
2010 a = pop_pair();
2011 c = _gvn.transform( new (C) CmpD3Node( a, b));
2012 push(c);
2013 break;
2015 case Bytecodes::_dcmpg:
2016 b = pop_pair();
2017 a = pop_pair();
2018 // Same as dcmpl but need to flip the unordered case.
2019 // Commute the inputs, which negates the result sign except for unordered.
2020 // Flip the unordered as well by using CmpD3 which implements
2021 // unordered-lesser instead of unordered-greater semantics.
2022 // Finally, negate the result bits. Result is same as using a
2023 // CmpD3Greater except we did it with CmpD3 alone.
2024 c = _gvn.transform( new (C) CmpD3Node( b, a));
2025 c = _gvn.transform( new (C) SubINode(_gvn.intcon(0),c) );
2026 push(c);
2027 break;
2030 // Note for longs -> lo word is on TOS, hi word is on TOS - 1
2031 case Bytecodes::_land:
2032 b = pop_pair();
2033 a = pop_pair();
2034 c = _gvn.transform( new (C) AndLNode(a,b) );
2035 push_pair(c);
2036 break;
2037 case Bytecodes::_lor:
2038 b = pop_pair();
2039 a = pop_pair();
2040 c = _gvn.transform( new (C) OrLNode(a,b) );
2041 push_pair(c);
2042 break;
2043 case Bytecodes::_lxor:
2044 b = pop_pair();
2045 a = pop_pair();
2046 c = _gvn.transform( new (C) XorLNode(a,b) );
2047 push_pair(c);
2048 break;
2050 case Bytecodes::_lshl:
2051 b = pop(); // the shift count
2052 a = pop_pair(); // value to be shifted
2053 c = _gvn.transform( new (C) LShiftLNode(a,b) );
2054 push_pair(c);
2055 break;
2056 case Bytecodes::_lshr:
2057 b = pop(); // the shift count
2058 a = pop_pair(); // value to be shifted
2059 c = _gvn.transform( new (C) RShiftLNode(a,b) );
2060 push_pair(c);
2061 break;
2062 case Bytecodes::_lushr:
2063 b = pop(); // the shift count
2064 a = pop_pair(); // value to be shifted
2065 c = _gvn.transform( new (C) URShiftLNode(a,b) );
2066 push_pair(c);
2067 break;
2068 case Bytecodes::_lmul:
2069 b = pop_pair();
2070 a = pop_pair();
2071 c = _gvn.transform( new (C) MulLNode(a,b) );
2072 push_pair(c);
2073 break;
2075 case Bytecodes::_lrem:
2076 // Must keep both values on the expression-stack during null-check
2077 assert(peek(0) == top(), "long word order");
2078 zero_check_long(peek(1));
2079 // Compile-time detect of null-exception?
2080 if (stopped()) return;
2081 b = pop_pair();
2082 a = pop_pair();
2083 c = _gvn.transform( new (C) ModLNode(control(),a,b) );
2084 push_pair(c);
2085 break;
2087 case Bytecodes::_ldiv:
2088 // Must keep both values on the expression-stack during null-check
2089 assert(peek(0) == top(), "long word order");
2090 zero_check_long(peek(1));
2091 // Compile-time detect of null-exception?
2092 if (stopped()) return;
2093 b = pop_pair();
2094 a = pop_pair();
2095 c = _gvn.transform( new (C) DivLNode(control(),a,b) );
2096 push_pair(c);
2097 break;
2099 case Bytecodes::_ladd:
2100 b = pop_pair();
2101 a = pop_pair();
2102 c = _gvn.transform( new (C) AddLNode(a,b) );
2103 push_pair(c);
2104 break;
2105 case Bytecodes::_lsub:
2106 b = pop_pair();
2107 a = pop_pair();
2108 c = _gvn.transform( new (C) SubLNode(a,b) );
2109 push_pair(c);
2110 break;
2111 case Bytecodes::_lcmp:
2112 // Safepoints are now inserted _before_ branches. The long-compare
2113 // bytecode painfully produces a 3-way value (-1,0,+1) which requires a
2114 // slew of control flow. These are usually followed by a CmpI vs zero and
2115 // a branch; this pattern then optimizes to the obvious long-compare and
2116 // branch. However, if the branch is backwards there's a Safepoint
2117 // inserted. The inserted Safepoint captures the JVM state at the
2118 // pre-branch point, i.e. it captures the 3-way value. Thus if a
2119 // long-compare is used to control a loop the debug info will force
2120 // computation of the 3-way value, even though the generated code uses a
2121 // long-compare and branch. We try to rectify the situation by inserting
2122 // a SafePoint here and have it dominate and kill the safepoint added at a
2123 // following backwards branch. At this point the JVM state merely holds 2
2124 // longs but not the 3-way value.
2125 if( UseLoopSafepoints ) {
2126 switch( iter().next_bc() ) {
2127 case Bytecodes::_ifgt:
2128 case Bytecodes::_iflt:
2129 case Bytecodes::_ifge:
2130 case Bytecodes::_ifle:
2131 case Bytecodes::_ifne:
2132 case Bytecodes::_ifeq:
2133 // If this is a backwards branch in the bytecodes, add Safepoint
2134 maybe_add_safepoint(iter().next_get_dest());
2135 }
2136 }
2137 b = pop_pair();
2138 a = pop_pair();
2139 c = _gvn.transform( new (C) CmpL3Node( a, b ));
2140 push(c);
2141 break;
2143 case Bytecodes::_lneg:
2144 a = pop_pair();
2145 b = _gvn.transform( new (C) SubLNode(longcon(0),a));
2146 push_pair(b);
2147 break;
2148 case Bytecodes::_l2i:
2149 a = pop_pair();
2150 push( _gvn.transform( new (C) ConvL2INode(a)));
2151 break;
2152 case Bytecodes::_i2l:
2153 a = pop();
2154 b = _gvn.transform( new (C) ConvI2LNode(a));
2155 push_pair(b);
2156 break;
2157 case Bytecodes::_i2b:
2158 // Sign extend
2159 a = pop();
2160 a = _gvn.transform( new (C) LShiftINode(a,_gvn.intcon(24)) );
2161 a = _gvn.transform( new (C) RShiftINode(a,_gvn.intcon(24)) );
2162 push( a );
2163 break;
2164 case Bytecodes::_i2s:
2165 a = pop();
2166 a = _gvn.transform( new (C) LShiftINode(a,_gvn.intcon(16)) );
2167 a = _gvn.transform( new (C) RShiftINode(a,_gvn.intcon(16)) );
2168 push( a );
2169 break;
2170 case Bytecodes::_i2c:
2171 a = pop();
2172 push( _gvn.transform( new (C) AndINode(a,_gvn.intcon(0xFFFF)) ) );
2173 break;
2175 case Bytecodes::_i2f:
2176 a = pop();
2177 b = _gvn.transform( new (C) ConvI2FNode(a) ) ;
2178 c = precision_rounding(b);
2179 push (b);
2180 break;
2182 case Bytecodes::_i2d:
2183 a = pop();
2184 b = _gvn.transform( new (C) ConvI2DNode(a));
2185 push_pair(b);
2186 break;
2188 case Bytecodes::_iinc: // Increment local
2189 i = iter().get_index(); // Get local index
2190 set_local( i, _gvn.transform( new (C) AddINode( _gvn.intcon(iter().get_iinc_con()), local(i) ) ) );
2191 break;
2193 // Exit points of synchronized methods must have an unlock node
2194 case Bytecodes::_return:
2195 return_current(NULL);
2196 break;
2198 case Bytecodes::_ireturn:
2199 case Bytecodes::_areturn:
2200 case Bytecodes::_freturn:
2201 return_current(pop());
2202 break;
2203 case Bytecodes::_lreturn:
2204 return_current(pop_pair());
2205 break;
2206 case Bytecodes::_dreturn:
2207 return_current(pop_pair());
2208 break;
2210 case Bytecodes::_athrow:
2211 // null exception oop throws NULL pointer exception
2212 null_check(peek());
2213 if (stopped()) return;
2214 // Hook the thrown exception directly to subsequent handlers.
2215 if (BailoutToInterpreterForThrows) {
2216 // Keep method interpreted from now on.
2217 uncommon_trap(Deoptimization::Reason_unhandled,
2218 Deoptimization::Action_make_not_compilable);
2219 return;
2220 }
2221 if (env()->jvmti_can_post_on_exceptions()) {
2222 // check if we must post exception events, take uncommon trap if so (with must_throw = false)
2223 uncommon_trap_if_should_post_on_exceptions(Deoptimization::Reason_unhandled, false);
2224 }
2225 // Here if either can_post_on_exceptions or should_post_on_exceptions is false
2226 add_exception_state(make_exception_state(peek()));
2227 break;
2229 case Bytecodes::_goto: // fall through
2230 case Bytecodes::_goto_w: {
2231 int target_bci = (bc() == Bytecodes::_goto) ? iter().get_dest() : iter().get_far_dest();
2233 // If this is a backwards branch in the bytecodes, add Safepoint
2234 maybe_add_safepoint(target_bci);
2236 // Update method data
2237 profile_taken_branch(target_bci);
2239 // Merge the current control into the target basic block
2240 merge(target_bci);
2242 // See if we can get some profile data and hand it off to the next block
2243 Block *target_block = block()->successor_for_bci(target_bci);
2244 if (target_block->pred_count() != 1) break;
2245 ciMethodData* methodData = method()->method_data();
2246 if (!methodData->is_mature()) break;
2247 ciProfileData* data = methodData->bci_to_data(bci());
2248 assert( data->is_JumpData(), "" );
2249 int taken = ((ciJumpData*)data)->taken();
2250 taken = method()->scale_count(taken);
2251 target_block->set_count(taken);
2252 break;
2253 }
2255 case Bytecodes::_ifnull: btest = BoolTest::eq; goto handle_if_null;
2256 case Bytecodes::_ifnonnull: btest = BoolTest::ne; goto handle_if_null;
2257 handle_if_null:
2258 // If this is a backwards branch in the bytecodes, add Safepoint
2259 maybe_add_safepoint(iter().get_dest());
2260 a = null();
2261 b = pop();
2262 c = _gvn.transform( new (C) CmpPNode(b, a) );
2263 do_ifnull(btest, c);
2264 break;
2266 case Bytecodes::_if_acmpeq: btest = BoolTest::eq; goto handle_if_acmp;
2267 case Bytecodes::_if_acmpne: btest = BoolTest::ne; goto handle_if_acmp;
2268 handle_if_acmp:
2269 // If this is a backwards branch in the bytecodes, add Safepoint
2270 maybe_add_safepoint(iter().get_dest());
2271 a = pop();
2272 b = pop();
2273 c = _gvn.transform( new (C) CmpPNode(b, a) );
2274 c = optimize_cmp_with_klass(c);
2275 do_if(btest, c);
2276 break;
2278 case Bytecodes::_ifeq: btest = BoolTest::eq; goto handle_ifxx;
2279 case Bytecodes::_ifne: btest = BoolTest::ne; goto handle_ifxx;
2280 case Bytecodes::_iflt: btest = BoolTest::lt; goto handle_ifxx;
2281 case Bytecodes::_ifle: btest = BoolTest::le; goto handle_ifxx;
2282 case Bytecodes::_ifgt: btest = BoolTest::gt; goto handle_ifxx;
2283 case Bytecodes::_ifge: btest = BoolTest::ge; goto handle_ifxx;
2284 handle_ifxx:
2285 // If this is a backwards branch in the bytecodes, add Safepoint
2286 maybe_add_safepoint(iter().get_dest());
2287 a = _gvn.intcon(0);
2288 b = pop();
2289 c = _gvn.transform( new (C) CmpINode(b, a) );
2290 do_if(btest, c);
2291 break;
2293 case Bytecodes::_if_icmpeq: btest = BoolTest::eq; goto handle_if_icmp;
2294 case Bytecodes::_if_icmpne: btest = BoolTest::ne; goto handle_if_icmp;
2295 case Bytecodes::_if_icmplt: btest = BoolTest::lt; goto handle_if_icmp;
2296 case Bytecodes::_if_icmple: btest = BoolTest::le; goto handle_if_icmp;
2297 case Bytecodes::_if_icmpgt: btest = BoolTest::gt; goto handle_if_icmp;
2298 case Bytecodes::_if_icmpge: btest = BoolTest::ge; goto handle_if_icmp;
2299 handle_if_icmp:
2300 // If this is a backwards branch in the bytecodes, add Safepoint
2301 maybe_add_safepoint(iter().get_dest());
2302 a = pop();
2303 b = pop();
2304 c = _gvn.transform( new (C) CmpINode( b, a ) );
2305 do_if(btest, c);
2306 break;
2308 case Bytecodes::_tableswitch:
2309 do_tableswitch();
2310 break;
2312 case Bytecodes::_lookupswitch:
2313 do_lookupswitch();
2314 break;
2316 case Bytecodes::_invokestatic:
2317 case Bytecodes::_invokedynamic:
2318 case Bytecodes::_invokespecial:
2319 case Bytecodes::_invokevirtual:
2320 case Bytecodes::_invokeinterface:
2321 do_call();
2322 break;
2323 case Bytecodes::_checkcast:
2324 do_checkcast();
2325 break;
2326 case Bytecodes::_instanceof:
2327 do_instanceof();
2328 break;
2329 case Bytecodes::_anewarray:
2330 do_anewarray();
2331 break;
2332 case Bytecodes::_newarray:
2333 do_newarray((BasicType)iter().get_index());
2334 break;
2335 case Bytecodes::_multianewarray:
2336 do_multianewarray();
2337 break;
2338 case Bytecodes::_new:
2339 do_new();
2340 break;
2342 case Bytecodes::_jsr:
2343 case Bytecodes::_jsr_w:
2344 do_jsr();
2345 break;
2347 case Bytecodes::_ret:
2348 do_ret();
2349 break;
2352 case Bytecodes::_monitorenter:
2353 do_monitor_enter();
2354 break;
2356 case Bytecodes::_monitorexit:
2357 do_monitor_exit();
2358 break;
2360 case Bytecodes::_breakpoint:
2361 // Breakpoint set concurrently to compile
2362 // %%% use an uncommon trap?
2363 C->record_failure("breakpoint in method");
2364 return;
2366 default:
2367 #ifndef PRODUCT
2368 map()->dump(99);
2369 #endif
2370 tty->print("\nUnhandled bytecode %s\n", Bytecodes::name(bc()) );
2371 ShouldNotReachHere();
2372 }
2374 #ifndef PRODUCT
2375 IdealGraphPrinter *printer = IdealGraphPrinter::printer();
2376 if(printer) {
2377 char buffer[256];
2378 sprintf(buffer, "Bytecode %d: %s", bci(), Bytecodes::name(bc()));
2379 bool old = printer->traverse_outs();
2380 printer->set_traverse_outs(true);
2381 printer->print_method(C, buffer, 4);
2382 printer->set_traverse_outs(old);
2383 }
2384 #endif
2385 }