Mon, 24 Nov 2014 07:29:03 -0800
8058148: MaxNodeLimit and LiveNodeCountInliningCutoff
Reviewed-by: kvn, roland
1 /*
2 * Copyright (c) 1998, 2014, 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 Node* adr = array_addressing(elem_type, 1);
61 if (stopped()) return; // guaranteed null or range check
62 Node* val = pop();
63 dec_sp(2); // Pop array and index
64 const TypeAryPtr* adr_type = TypeAryPtr::get_array_body_type(elem_type);
65 store_to_memory(control(), adr, val, elem_type, adr_type, StoreNode::release_if_reference(elem_type));
66 }
69 //------------------------------array_addressing-------------------------------
70 // Pull array and index from the stack. Compute pointer-to-element.
71 Node* Parse::array_addressing(BasicType type, int vals, const Type* *result2) {
72 Node *idx = peek(0+vals); // Get from stack without popping
73 Node *ary = peek(1+vals); // in case of exception
75 // Null check the array base, with correct stack contents
76 ary = null_check(ary, T_ARRAY);
77 // Compile-time detect of null-exception?
78 if (stopped()) return top();
80 const TypeAryPtr* arytype = _gvn.type(ary)->is_aryptr();
81 const TypeInt* sizetype = arytype->size();
82 const Type* elemtype = arytype->elem();
84 if (UseUniqueSubclasses && result2 != NULL) {
85 const Type* el = elemtype->make_ptr();
86 if (el && el->isa_instptr()) {
87 const TypeInstPtr* toop = el->is_instptr();
88 if (toop->klass()->as_instance_klass()->unique_concrete_subklass()) {
89 // If we load from "AbstractClass[]" we must see "ConcreteSubClass".
90 const Type* subklass = Type::get_const_type(toop->klass());
91 elemtype = subklass->join_speculative(el);
92 }
93 }
94 }
96 // Check for big class initializers with all constant offsets
97 // feeding into a known-size array.
98 const TypeInt* idxtype = _gvn.type(idx)->is_int();
99 // See if the highest idx value is less than the lowest array bound,
100 // and if the idx value cannot be negative:
101 bool need_range_check = true;
102 if (idxtype->_hi < sizetype->_lo && idxtype->_lo >= 0) {
103 need_range_check = false;
104 if (C->log() != NULL) C->log()->elem("observe that='!need_range_check'");
105 }
107 ciKlass * arytype_klass = arytype->klass();
108 if ((arytype_klass != NULL) && (!arytype_klass->is_loaded())) {
109 // Only fails for some -Xcomp runs
110 // The class is unloaded. We have to run this bytecode in the interpreter.
111 uncommon_trap(Deoptimization::Reason_unloaded,
112 Deoptimization::Action_reinterpret,
113 arytype->klass(), "!loaded array");
114 return top();
115 }
117 // Do the range check
118 if (GenerateRangeChecks && need_range_check) {
119 Node* tst;
120 if (sizetype->_hi <= 0) {
121 // The greatest array bound is negative, so we can conclude that we're
122 // compiling unreachable code, but the unsigned compare trick used below
123 // only works with non-negative lengths. Instead, hack "tst" to be zero so
124 // the uncommon_trap path will always be taken.
125 tst = _gvn.intcon(0);
126 } else {
127 // Range is constant in array-oop, so we can use the original state of mem
128 Node* len = load_array_length(ary);
130 // Test length vs index (standard trick using unsigned compare)
131 Node* chk = _gvn.transform( new (C) CmpUNode(idx, len) );
132 BoolTest::mask btest = BoolTest::lt;
133 tst = _gvn.transform( new (C) BoolNode(chk, btest) );
134 }
135 // Branch to failure if out of bounds
136 { BuildCutout unless(this, tst, PROB_MAX);
137 if (C->allow_range_check_smearing()) {
138 // Do not use builtin_throw, since range checks are sometimes
139 // made more stringent by an optimistic transformation.
140 // This creates "tentative" range checks at this point,
141 // which are not guaranteed to throw exceptions.
142 // See IfNode::Ideal, is_range_check, adjust_check.
143 uncommon_trap(Deoptimization::Reason_range_check,
144 Deoptimization::Action_make_not_entrant,
145 NULL, "range_check");
146 } else {
147 // If we have already recompiled with the range-check-widening
148 // heroic optimization turned off, then we must really be throwing
149 // range check exceptions.
150 builtin_throw(Deoptimization::Reason_range_check, idx);
151 }
152 }
153 }
154 // Check for always knowing you are throwing a range-check exception
155 if (stopped()) return top();
157 Node* ptr = array_element_address(ary, idx, type, sizetype);
159 if (result2 != NULL) *result2 = elemtype;
161 assert(ptr != top(), "top should go hand-in-hand with stopped");
163 return ptr;
164 }
167 // returns IfNode
168 IfNode* Parse::jump_if_fork_int(Node* a, Node* b, BoolTest::mask mask) {
169 Node *cmp = _gvn.transform( new (C) CmpINode( a, b)); // two cases: shiftcount > 32 and shiftcount <= 32
170 Node *tst = _gvn.transform( new (C) BoolNode( cmp, mask));
171 IfNode *iff = create_and_map_if( control(), tst, ((mask == BoolTest::eq) ? PROB_STATIC_INFREQUENT : PROB_FAIR), COUNT_UNKNOWN );
172 return iff;
173 }
175 // return Region node
176 Node* Parse::jump_if_join(Node* iffalse, Node* iftrue) {
177 Node *region = new (C) RegionNode(3); // 2 results
178 record_for_igvn(region);
179 region->init_req(1, iffalse);
180 region->init_req(2, iftrue );
181 _gvn.set_type(region, Type::CONTROL);
182 region = _gvn.transform(region);
183 set_control (region);
184 return region;
185 }
188 //------------------------------helper for tableswitch-------------------------
189 void Parse::jump_if_true_fork(IfNode *iff, int dest_bci_if_true, int prof_table_index) {
190 // True branch, use existing map info
191 { PreserveJVMState pjvms(this);
192 Node *iftrue = _gvn.transform( new (C) IfTrueNode (iff) );
193 set_control( iftrue );
194 profile_switch_case(prof_table_index);
195 merge_new_path(dest_bci_if_true);
196 }
198 // False branch
199 Node *iffalse = _gvn.transform( new (C) IfFalseNode(iff) );
200 set_control( iffalse );
201 }
203 void Parse::jump_if_false_fork(IfNode *iff, int dest_bci_if_true, int prof_table_index) {
204 // True branch, use existing map info
205 { PreserveJVMState pjvms(this);
206 Node *iffalse = _gvn.transform( new (C) IfFalseNode (iff) );
207 set_control( iffalse );
208 profile_switch_case(prof_table_index);
209 merge_new_path(dest_bci_if_true);
210 }
212 // False branch
213 Node *iftrue = _gvn.transform( new (C) IfTrueNode(iff) );
214 set_control( iftrue );
215 }
217 void Parse::jump_if_always_fork(int dest_bci, int prof_table_index) {
218 // False branch, use existing map and control()
219 profile_switch_case(prof_table_index);
220 merge_new_path(dest_bci);
221 }
224 extern "C" {
225 static int jint_cmp(const void *i, const void *j) {
226 int a = *(jint *)i;
227 int b = *(jint *)j;
228 return a > b ? 1 : a < b ? -1 : 0;
229 }
230 }
233 // Default value for methodData switch indexing. Must be a negative value to avoid
234 // conflict with any legal switch index.
235 #define NullTableIndex -1
237 class SwitchRange : public StackObj {
238 // a range of integers coupled with a bci destination
239 jint _lo; // inclusive lower limit
240 jint _hi; // inclusive upper limit
241 int _dest;
242 int _table_index; // index into method data table
244 public:
245 jint lo() const { return _lo; }
246 jint hi() const { return _hi; }
247 int dest() const { return _dest; }
248 int table_index() const { return _table_index; }
249 bool is_singleton() const { return _lo == _hi; }
251 void setRange(jint lo, jint hi, int dest, int table_index) {
252 assert(lo <= hi, "must be a non-empty range");
253 _lo = lo, _hi = hi; _dest = dest; _table_index = table_index;
254 }
255 bool adjoinRange(jint lo, jint hi, int dest, int table_index) {
256 assert(lo <= hi, "must be a non-empty range");
257 if (lo == _hi+1 && dest == _dest && table_index == _table_index) {
258 _hi = hi;
259 return true;
260 }
261 return false;
262 }
264 void set (jint value, int dest, int table_index) {
265 setRange(value, value, dest, table_index);
266 }
267 bool adjoin(jint value, int dest, int table_index) {
268 return adjoinRange(value, value, dest, table_index);
269 }
271 void print() {
272 if (is_singleton())
273 tty->print(" {%d}=>%d", lo(), dest());
274 else if (lo() == min_jint)
275 tty->print(" {..%d}=>%d", hi(), dest());
276 else if (hi() == max_jint)
277 tty->print(" {%d..}=>%d", lo(), dest());
278 else
279 tty->print(" {%d..%d}=>%d", lo(), hi(), dest());
280 }
281 };
284 //-------------------------------do_tableswitch--------------------------------
285 void Parse::do_tableswitch() {
286 Node* lookup = pop();
288 // Get information about tableswitch
289 int default_dest = iter().get_dest_table(0);
290 int lo_index = iter().get_int_table(1);
291 int hi_index = iter().get_int_table(2);
292 int len = hi_index - lo_index + 1;
294 if (len < 1) {
295 // If this is a backward branch, add safepoint
296 maybe_add_safepoint(default_dest);
297 merge(default_dest);
298 return;
299 }
301 // generate decision tree, using trichotomy when possible
302 int rnum = len+2;
303 bool makes_backward_branch = false;
304 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum);
305 int rp = -1;
306 if (lo_index != min_jint) {
307 ranges[++rp].setRange(min_jint, lo_index-1, default_dest, NullTableIndex);
308 }
309 for (int j = 0; j < len; j++) {
310 jint match_int = lo_index+j;
311 int dest = iter().get_dest_table(j+3);
312 makes_backward_branch |= (dest <= bci());
313 int table_index = method_data_update() ? j : NullTableIndex;
314 if (rp < 0 || !ranges[rp].adjoin(match_int, dest, table_index)) {
315 ranges[++rp].set(match_int, dest, table_index);
316 }
317 }
318 jint highest = lo_index+(len-1);
319 assert(ranges[rp].hi() == highest, "");
320 if (highest != max_jint
321 && !ranges[rp].adjoinRange(highest+1, max_jint, default_dest, NullTableIndex)) {
322 ranges[++rp].setRange(highest+1, max_jint, default_dest, NullTableIndex);
323 }
324 assert(rp < len+2, "not too many ranges");
326 // Safepoint in case if backward branch observed
327 if( makes_backward_branch && UseLoopSafepoints )
328 add_safepoint();
330 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]);
331 }
334 //------------------------------do_lookupswitch--------------------------------
335 void Parse::do_lookupswitch() {
336 Node *lookup = pop(); // lookup value
337 // Get information about lookupswitch
338 int default_dest = iter().get_dest_table(0);
339 int len = iter().get_int_table(1);
341 if (len < 1) { // If this is a backward branch, add safepoint
342 maybe_add_safepoint(default_dest);
343 merge(default_dest);
344 return;
345 }
347 // generate decision tree, using trichotomy when possible
348 jint* table = NEW_RESOURCE_ARRAY(jint, len*2);
349 {
350 for( int j = 0; j < len; j++ ) {
351 table[j+j+0] = iter().get_int_table(2+j+j);
352 table[j+j+1] = iter().get_dest_table(2+j+j+1);
353 }
354 qsort( table, len, 2*sizeof(table[0]), jint_cmp );
355 }
357 int rnum = len*2+1;
358 bool makes_backward_branch = false;
359 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum);
360 int rp = -1;
361 for( int j = 0; j < len; j++ ) {
362 jint match_int = table[j+j+0];
363 int dest = table[j+j+1];
364 int next_lo = rp < 0 ? min_jint : ranges[rp].hi()+1;
365 int table_index = method_data_update() ? j : NullTableIndex;
366 makes_backward_branch |= (dest <= bci());
367 if( match_int != next_lo ) {
368 ranges[++rp].setRange(next_lo, match_int-1, default_dest, NullTableIndex);
369 }
370 if( rp < 0 || !ranges[rp].adjoin(match_int, dest, table_index) ) {
371 ranges[++rp].set(match_int, dest, table_index);
372 }
373 }
374 jint highest = table[2*(len-1)];
375 assert(ranges[rp].hi() == highest, "");
376 if( highest != max_jint
377 && !ranges[rp].adjoinRange(highest+1, max_jint, default_dest, NullTableIndex) ) {
378 ranges[++rp].setRange(highest+1, max_jint, default_dest, NullTableIndex);
379 }
380 assert(rp < rnum, "not too many ranges");
382 // Safepoint in case backward branch observed
383 if( makes_backward_branch && UseLoopSafepoints )
384 add_safepoint();
386 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]);
387 }
389 //----------------------------create_jump_tables-------------------------------
390 bool Parse::create_jump_tables(Node* key_val, SwitchRange* lo, SwitchRange* hi) {
391 // Are jumptables enabled
392 if (!UseJumpTables) return false;
394 // Are jumptables supported
395 if (!Matcher::has_match_rule(Op_Jump)) return false;
397 // Don't make jump table if profiling
398 if (method_data_update()) return false;
400 // Decide if a guard is needed to lop off big ranges at either (or
401 // both) end(s) of the input set. We'll call this the default target
402 // even though we can't be sure that it is the true "default".
404 bool needs_guard = false;
405 int default_dest;
406 int64 total_outlier_size = 0;
407 int64 hi_size = ((int64)hi->hi()) - ((int64)hi->lo()) + 1;
408 int64 lo_size = ((int64)lo->hi()) - ((int64)lo->lo()) + 1;
410 if (lo->dest() == hi->dest()) {
411 total_outlier_size = hi_size + lo_size;
412 default_dest = lo->dest();
413 } else if (lo_size > hi_size) {
414 total_outlier_size = lo_size;
415 default_dest = lo->dest();
416 } else {
417 total_outlier_size = hi_size;
418 default_dest = hi->dest();
419 }
421 // If a guard test will eliminate very sparse end ranges, then
422 // it is worth the cost of an extra jump.
423 if (total_outlier_size > (MaxJumpTableSparseness * 4)) {
424 needs_guard = true;
425 if (default_dest == lo->dest()) lo++;
426 if (default_dest == hi->dest()) hi--;
427 }
429 // Find the total number of cases and ranges
430 int64 num_cases = ((int64)hi->hi()) - ((int64)lo->lo()) + 1;
431 int num_range = hi - lo + 1;
433 // Don't create table if: too large, too small, or too sparse.
434 if (num_cases < MinJumpTableSize || num_cases > MaxJumpTableSize)
435 return false;
436 if (num_cases > (MaxJumpTableSparseness * num_range))
437 return false;
439 // Normalize table lookups to zero
440 int lowval = lo->lo();
441 key_val = _gvn.transform( new (C) SubINode(key_val, _gvn.intcon(lowval)) );
443 // Generate a guard to protect against input keyvals that aren't
444 // in the switch domain.
445 if (needs_guard) {
446 Node* size = _gvn.intcon(num_cases);
447 Node* cmp = _gvn.transform( new (C) CmpUNode(key_val, size) );
448 Node* tst = _gvn.transform( new (C) BoolNode(cmp, BoolTest::ge) );
449 IfNode* iff = create_and_map_if( control(), tst, PROB_FAIR, COUNT_UNKNOWN);
450 jump_if_true_fork(iff, default_dest, NullTableIndex);
451 }
453 // Create an ideal node JumpTable that has projections
454 // of all possible ranges for a switch statement
455 // The key_val input must be converted to a pointer offset and scaled.
456 // Compare Parse::array_addressing above.
457 #ifdef _LP64
458 // Clean the 32-bit int into a real 64-bit offset.
459 // Otherwise, the jint value 0 might turn into an offset of 0x0800000000.
460 const TypeLong* lkeytype = TypeLong::make(CONST64(0), num_cases-1, Type::WidenMin);
461 key_val = _gvn.transform( new (C) ConvI2LNode(key_val, lkeytype) );
462 #endif
463 // Shift the value by wordsize so we have an index into the table, rather
464 // than a switch value
465 Node *shiftWord = _gvn.MakeConX(wordSize);
466 key_val = _gvn.transform( new (C) MulXNode( key_val, shiftWord));
468 // Create the JumpNode
469 Node* jtn = _gvn.transform( new (C) JumpNode(control(), key_val, num_cases) );
471 // These are the switch destinations hanging off the jumpnode
472 int i = 0;
473 for (SwitchRange* r = lo; r <= hi; r++) {
474 for (int64 j = r->lo(); j <= r->hi(); j++, i++) {
475 Node* input = _gvn.transform(new (C) JumpProjNode(jtn, i, r->dest(), (int)(j - lowval)));
476 {
477 PreserveJVMState pjvms(this);
478 set_control(input);
479 jump_if_always_fork(r->dest(), r->table_index());
480 }
481 }
482 }
483 assert(i == num_cases, "miscount of cases");
484 stop_and_kill_map(); // no more uses for this JVMS
485 return true;
486 }
488 //----------------------------jump_switch_ranges-------------------------------
489 void Parse::jump_switch_ranges(Node* key_val, SwitchRange *lo, SwitchRange *hi, int switch_depth) {
490 Block* switch_block = block();
492 if (switch_depth == 0) {
493 // Do special processing for the top-level call.
494 assert(lo->lo() == min_jint, "initial range must exhaust Type::INT");
495 assert(hi->hi() == max_jint, "initial range must exhaust Type::INT");
497 // Decrement pred-numbers for the unique set of nodes.
498 #ifdef ASSERT
499 // Ensure that the block's successors are a (duplicate-free) set.
500 int successors_counted = 0; // block occurrences in [hi..lo]
501 int unique_successors = switch_block->num_successors();
502 for (int i = 0; i < unique_successors; i++) {
503 Block* target = switch_block->successor_at(i);
505 // Check that the set of successors is the same in both places.
506 int successors_found = 0;
507 for (SwitchRange* p = lo; p <= hi; p++) {
508 if (p->dest() == target->start()) successors_found++;
509 }
510 assert(successors_found > 0, "successor must be known");
511 successors_counted += successors_found;
512 }
513 assert(successors_counted == (hi-lo)+1, "no unexpected successors");
514 #endif
516 // Maybe prune the inputs, based on the type of key_val.
517 jint min_val = min_jint;
518 jint max_val = max_jint;
519 const TypeInt* ti = key_val->bottom_type()->isa_int();
520 if (ti != NULL) {
521 min_val = ti->_lo;
522 max_val = ti->_hi;
523 assert(min_val <= max_val, "invalid int type");
524 }
525 while (lo->hi() < min_val) lo++;
526 if (lo->lo() < min_val) lo->setRange(min_val, lo->hi(), lo->dest(), lo->table_index());
527 while (hi->lo() > max_val) hi--;
528 if (hi->hi() > max_val) hi->setRange(hi->lo(), max_val, hi->dest(), hi->table_index());
529 }
531 #ifndef PRODUCT
532 if (switch_depth == 0) {
533 _max_switch_depth = 0;
534 _est_switch_depth = log2_intptr((hi-lo+1)-1)+1;
535 }
536 #endif
538 assert(lo <= hi, "must be a non-empty set of ranges");
539 if (lo == hi) {
540 jump_if_always_fork(lo->dest(), lo->table_index());
541 } else {
542 assert(lo->hi() == (lo+1)->lo()-1, "contiguous ranges");
543 assert(hi->lo() == (hi-1)->hi()+1, "contiguous ranges");
545 if (create_jump_tables(key_val, lo, hi)) return;
547 int nr = hi - lo + 1;
549 SwitchRange* mid = lo + nr/2;
550 // if there is an easy choice, pivot at a singleton:
551 if (nr > 3 && !mid->is_singleton() && (mid-1)->is_singleton()) mid--;
553 assert(lo < mid && mid <= hi, "good pivot choice");
554 assert(nr != 2 || mid == hi, "should pick higher of 2");
555 assert(nr != 3 || mid == hi-1, "should pick middle of 3");
557 Node *test_val = _gvn.intcon(mid->lo());
559 if (mid->is_singleton()) {
560 IfNode *iff_ne = jump_if_fork_int(key_val, test_val, BoolTest::ne);
561 jump_if_false_fork(iff_ne, mid->dest(), mid->table_index());
563 // Special Case: If there are exactly three ranges, and the high
564 // and low range each go to the same place, omit the "gt" test,
565 // since it will not discriminate anything.
566 bool eq_test_only = (hi == lo+2 && hi->dest() == lo->dest());
567 if (eq_test_only) {
568 assert(mid == hi-1, "");
569 }
571 // if there is a higher range, test for it and process it:
572 if (mid < hi && !eq_test_only) {
573 // two comparisons of same values--should enable 1 test for 2 branches
574 // Use BoolTest::le instead of BoolTest::gt
575 IfNode *iff_le = jump_if_fork_int(key_val, test_val, BoolTest::le);
576 Node *iftrue = _gvn.transform( new (C) IfTrueNode(iff_le) );
577 Node *iffalse = _gvn.transform( new (C) IfFalseNode(iff_le) );
578 { PreserveJVMState pjvms(this);
579 set_control(iffalse);
580 jump_switch_ranges(key_val, mid+1, hi, switch_depth+1);
581 }
582 set_control(iftrue);
583 }
585 } else {
586 // mid is a range, not a singleton, so treat mid..hi as a unit
587 IfNode *iff_ge = jump_if_fork_int(key_val, test_val, BoolTest::ge);
589 // if there is a higher range, test for it and process it:
590 if (mid == hi) {
591 jump_if_true_fork(iff_ge, mid->dest(), mid->table_index());
592 } else {
593 Node *iftrue = _gvn.transform( new (C) IfTrueNode(iff_ge) );
594 Node *iffalse = _gvn.transform( new (C) IfFalseNode(iff_ge) );
595 { PreserveJVMState pjvms(this);
596 set_control(iftrue);
597 jump_switch_ranges(key_val, mid, hi, switch_depth+1);
598 }
599 set_control(iffalse);
600 }
601 }
603 // in any case, process the lower range
604 jump_switch_ranges(key_val, lo, mid-1, switch_depth+1);
605 }
607 // Decrease pred_count for each successor after all is done.
608 if (switch_depth == 0) {
609 int unique_successors = switch_block->num_successors();
610 for (int i = 0; i < unique_successors; i++) {
611 Block* target = switch_block->successor_at(i);
612 // Throw away the pre-allocated path for each unique successor.
613 target->next_path_num();
614 }
615 }
617 #ifndef PRODUCT
618 _max_switch_depth = MAX2(switch_depth, _max_switch_depth);
619 if (TraceOptoParse && Verbose && WizardMode && switch_depth == 0) {
620 SwitchRange* r;
621 int nsing = 0;
622 for( r = lo; r <= hi; r++ ) {
623 if( r->is_singleton() ) nsing++;
624 }
625 tty->print(">>> ");
626 _method->print_short_name();
627 tty->print_cr(" switch decision tree");
628 tty->print_cr(" %d ranges (%d singletons), max_depth=%d, est_depth=%d",
629 (int) (hi-lo+1), nsing, _max_switch_depth, _est_switch_depth);
630 if (_max_switch_depth > _est_switch_depth) {
631 tty->print_cr("******** BAD SWITCH DEPTH ********");
632 }
633 tty->print(" ");
634 for( r = lo; r <= hi; r++ ) {
635 r->print();
636 }
637 tty->cr();
638 }
639 #endif
640 }
642 void Parse::modf() {
643 Node *f2 = pop();
644 Node *f1 = pop();
645 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::modf_Type(),
646 CAST_FROM_FN_PTR(address, SharedRuntime::frem),
647 "frem", NULL, //no memory effects
648 f1, f2);
649 Node* res = _gvn.transform(new (C) ProjNode(c, TypeFunc::Parms + 0));
651 push(res);
652 }
654 void Parse::modd() {
655 Node *d2 = pop_pair();
656 Node *d1 = pop_pair();
657 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::Math_DD_D_Type(),
658 CAST_FROM_FN_PTR(address, SharedRuntime::drem),
659 "drem", NULL, //no memory effects
660 d1, top(), d2, top());
661 Node* res_d = _gvn.transform(new (C) ProjNode(c, TypeFunc::Parms + 0));
663 #ifdef ASSERT
664 Node* res_top = _gvn.transform(new (C) ProjNode(c, TypeFunc::Parms + 1));
665 assert(res_top == top(), "second value must be top");
666 #endif
668 push_pair(res_d);
669 }
671 void Parse::l2f() {
672 Node* f2 = pop();
673 Node* f1 = pop();
674 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::l2f_Type(),
675 CAST_FROM_FN_PTR(address, SharedRuntime::l2f),
676 "l2f", NULL, //no memory effects
677 f1, f2);
678 Node* res = _gvn.transform(new (C) ProjNode(c, TypeFunc::Parms + 0));
680 push(res);
681 }
683 void Parse::do_irem() {
684 // Must keep both values on the expression-stack during null-check
685 zero_check_int(peek());
686 // Compile-time detect of null-exception?
687 if (stopped()) return;
689 Node* b = pop();
690 Node* a = pop();
692 const Type *t = _gvn.type(b);
693 if (t != Type::TOP) {
694 const TypeInt *ti = t->is_int();
695 if (ti->is_con()) {
696 int divisor = ti->get_con();
697 // check for positive power of 2
698 if (divisor > 0 &&
699 (divisor & ~(divisor-1)) == divisor) {
700 // yes !
701 Node *mask = _gvn.intcon((divisor - 1));
702 // Sigh, must handle negative dividends
703 Node *zero = _gvn.intcon(0);
704 IfNode *ifff = jump_if_fork_int(a, zero, BoolTest::lt);
705 Node *iff = _gvn.transform( new (C) IfFalseNode(ifff) );
706 Node *ift = _gvn.transform( new (C) IfTrueNode (ifff) );
707 Node *reg = jump_if_join(ift, iff);
708 Node *phi = PhiNode::make(reg, NULL, TypeInt::INT);
709 // Negative path; negate/and/negate
710 Node *neg = _gvn.transform( new (C) SubINode(zero, a) );
711 Node *andn= _gvn.transform( new (C) AndINode(neg, mask) );
712 Node *negn= _gvn.transform( new (C) SubINode(zero, andn) );
713 phi->init_req(1, negn);
714 // Fast positive case
715 Node *andx = _gvn.transform( new (C) AndINode(a, mask) );
716 phi->init_req(2, andx);
717 // Push the merge
718 push( _gvn.transform(phi) );
719 return;
720 }
721 }
722 }
723 // Default case
724 push( _gvn.transform( new (C) ModINode(control(),a,b) ) );
725 }
727 // Handle jsr and jsr_w bytecode
728 void Parse::do_jsr() {
729 assert(bc() == Bytecodes::_jsr || bc() == Bytecodes::_jsr_w, "wrong bytecode");
731 // Store information about current state, tagged with new _jsr_bci
732 int return_bci = iter().next_bci();
733 int jsr_bci = (bc() == Bytecodes::_jsr) ? iter().get_dest() : iter().get_far_dest();
735 // Update method data
736 profile_taken_branch(jsr_bci);
738 // The way we do things now, there is only one successor block
739 // for the jsr, because the target code is cloned by ciTypeFlow.
740 Block* target = successor_for_bci(jsr_bci);
742 // What got pushed?
743 const Type* ret_addr = target->peek();
744 assert(ret_addr->singleton(), "must be a constant (cloned jsr body)");
746 // Effect on jsr on stack
747 push(_gvn.makecon(ret_addr));
749 // Flow to the jsr.
750 merge(jsr_bci);
751 }
753 // Handle ret bytecode
754 void Parse::do_ret() {
755 // Find to whom we return.
756 assert(block()->num_successors() == 1, "a ret can only go one place now");
757 Block* target = block()->successor_at(0);
758 assert(!target->is_ready(), "our arrival must be expected");
759 profile_ret(target->flow()->start());
760 int pnum = target->next_path_num();
761 merge_common(target, pnum);
762 }
764 //--------------------------dynamic_branch_prediction--------------------------
765 // Try to gather dynamic branch prediction behavior. Return a probability
766 // of the branch being taken and set the "cnt" field. Returns a -1.0
767 // if we need to use static prediction for some reason.
768 float Parse::dynamic_branch_prediction(float &cnt) {
769 ResourceMark rm;
771 cnt = COUNT_UNKNOWN;
773 // Use MethodData information if it is available
774 // FIXME: free the ProfileData structure
775 ciMethodData* methodData = method()->method_data();
776 if (!methodData->is_mature()) return PROB_UNKNOWN;
777 ciProfileData* data = methodData->bci_to_data(bci());
778 if (!data->is_JumpData()) return PROB_UNKNOWN;
780 // get taken and not taken values
781 int taken = data->as_JumpData()->taken();
782 int not_taken = 0;
783 if (data->is_BranchData()) {
784 not_taken = data->as_BranchData()->not_taken();
785 }
787 // scale the counts to be commensurate with invocation counts:
788 taken = method()->scale_count(taken);
789 not_taken = method()->scale_count(not_taken);
791 // Give up if too few (or too many, in which case the sum will overflow) counts to be meaningful.
792 // We also check that individual counters are positive first, overwise the sum can become positive.
793 if (taken < 0 || not_taken < 0 || taken + not_taken < 40) {
794 if (C->log() != NULL) {
795 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d'", iter().get_dest(), taken, not_taken);
796 }
797 return PROB_UNKNOWN;
798 }
800 // Compute frequency that we arrive here
801 float sum = taken + not_taken;
802 // Adjust, if this block is a cloned private block but the
803 // Jump counts are shared. Taken the private counts for
804 // just this path instead of the shared counts.
805 if( block()->count() > 0 )
806 sum = block()->count();
807 cnt = sum / FreqCountInvocations;
809 // Pin probability to sane limits
810 float prob;
811 if( !taken )
812 prob = (0+PROB_MIN) / 2;
813 else if( !not_taken )
814 prob = (1+PROB_MAX) / 2;
815 else { // Compute probability of true path
816 prob = (float)taken / (float)(taken + not_taken);
817 if (prob > PROB_MAX) prob = PROB_MAX;
818 if (prob < PROB_MIN) prob = PROB_MIN;
819 }
821 assert((cnt > 0.0f) && (prob > 0.0f),
822 "Bad frequency assignment in if");
824 if (C->log() != NULL) {
825 const char* prob_str = NULL;
826 if (prob >= PROB_MAX) prob_str = (prob == PROB_MAX) ? "max" : "always";
827 if (prob <= PROB_MIN) prob_str = (prob == PROB_MIN) ? "min" : "never";
828 char prob_str_buf[30];
829 if (prob_str == NULL) {
830 sprintf(prob_str_buf, "%g", prob);
831 prob_str = prob_str_buf;
832 }
833 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d' cnt='%g' prob='%s'",
834 iter().get_dest(), taken, not_taken, cnt, prob_str);
835 }
836 return prob;
837 }
839 //-----------------------------branch_prediction-------------------------------
840 float Parse::branch_prediction(float& cnt,
841 BoolTest::mask btest,
842 int target_bci) {
843 float prob = dynamic_branch_prediction(cnt);
844 // If prob is unknown, switch to static prediction
845 if (prob != PROB_UNKNOWN) return prob;
847 prob = PROB_FAIR; // Set default value
848 if (btest == BoolTest::eq) // Exactly equal test?
849 prob = PROB_STATIC_INFREQUENT; // Assume its relatively infrequent
850 else if (btest == BoolTest::ne)
851 prob = PROB_STATIC_FREQUENT; // Assume its relatively frequent
853 // If this is a conditional test guarding a backwards branch,
854 // assume its a loop-back edge. Make it a likely taken branch.
855 if (target_bci < bci()) {
856 if (is_osr_parse()) { // Could be a hot OSR'd loop; force deopt
857 // Since it's an OSR, we probably have profile data, but since
858 // branch_prediction returned PROB_UNKNOWN, the counts are too small.
859 // Let's make a special check here for completely zero counts.
860 ciMethodData* methodData = method()->method_data();
861 if (!methodData->is_empty()) {
862 ciProfileData* data = methodData->bci_to_data(bci());
863 // Only stop for truly zero counts, which mean an unknown part
864 // of the OSR-ed method, and we want to deopt to gather more stats.
865 // If you have ANY counts, then this loop is simply 'cold' relative
866 // to the OSR loop.
867 if (data->as_BranchData()->taken() +
868 data->as_BranchData()->not_taken() == 0 ) {
869 // This is the only way to return PROB_UNKNOWN:
870 return PROB_UNKNOWN;
871 }
872 }
873 }
874 prob = PROB_STATIC_FREQUENT; // Likely to take backwards branch
875 }
877 assert(prob != PROB_UNKNOWN, "must have some guess at this point");
878 return prob;
879 }
881 // The magic constants are chosen so as to match the output of
882 // branch_prediction() when the profile reports a zero taken count.
883 // It is important to distinguish zero counts unambiguously, because
884 // some branches (e.g., _213_javac.Assembler.eliminate) validly produce
885 // very small but nonzero probabilities, which if confused with zero
886 // counts would keep the program recompiling indefinitely.
887 bool Parse::seems_never_taken(float prob) const {
888 return prob < PROB_MIN;
889 }
891 // True if the comparison seems to be the kind that will not change its
892 // statistics from true to false. See comments in adjust_map_after_if.
893 // This question is only asked along paths which are already
894 // classifed as untaken (by seems_never_taken), so really,
895 // if a path is never taken, its controlling comparison is
896 // already acting in a stable fashion. If the comparison
897 // seems stable, we will put an expensive uncommon trap
898 // on the untaken path.
899 bool Parse::seems_stable_comparison() const {
900 if (C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if)) {
901 return false;
902 }
903 return true;
904 }
906 //-------------------------------repush_if_args--------------------------------
907 // Push arguments of an "if" bytecode back onto the stack by adjusting _sp.
908 inline int Parse::repush_if_args() {
909 #ifndef PRODUCT
910 if (PrintOpto && WizardMode) {
911 tty->print("defending against excessive implicit null exceptions on %s @%d in ",
912 Bytecodes::name(iter().cur_bc()), iter().cur_bci());
913 method()->print_name(); tty->cr();
914 }
915 #endif
916 int bc_depth = - Bytecodes::depth(iter().cur_bc());
917 assert(bc_depth == 1 || bc_depth == 2, "only two kinds of branches");
918 DEBUG_ONLY(sync_jvms()); // argument(n) requires a synced jvms
919 assert(argument(0) != NULL, "must exist");
920 assert(bc_depth == 1 || argument(1) != NULL, "two must exist");
921 inc_sp(bc_depth);
922 return bc_depth;
923 }
925 //----------------------------------do_ifnull----------------------------------
926 void Parse::do_ifnull(BoolTest::mask btest, Node *c) {
927 int target_bci = iter().get_dest();
929 Block* branch_block = successor_for_bci(target_bci);
930 Block* next_block = successor_for_bci(iter().next_bci());
932 float cnt;
933 float prob = branch_prediction(cnt, btest, target_bci);
934 if (prob == PROB_UNKNOWN) {
935 // (An earlier version of do_ifnull omitted this trap for OSR methods.)
936 #ifndef PRODUCT
937 if (PrintOpto && Verbose)
938 tty->print_cr("Never-taken edge stops compilation at bci %d",bci());
939 #endif
940 repush_if_args(); // to gather stats on loop
941 // We need to mark this branch as taken so that if we recompile we will
942 // see that it is possible. In the tiered system the interpreter doesn't
943 // do profiling and by the time we get to the lower tier from the interpreter
944 // the path may be cold again. Make sure it doesn't look untaken
945 profile_taken_branch(target_bci, !ProfileInterpreter);
946 uncommon_trap(Deoptimization::Reason_unreached,
947 Deoptimization::Action_reinterpret,
948 NULL, "cold");
949 if (C->eliminate_boxing()) {
950 // Mark the successor blocks as parsed
951 branch_block->next_path_num();
952 next_block->next_path_num();
953 }
954 return;
955 }
957 explicit_null_checks_inserted++;
959 // Generate real control flow
960 Node *tst = _gvn.transform( new (C) BoolNode( c, btest ) );
962 // Sanity check the probability value
963 assert(prob > 0.0f,"Bad probability in Parser");
964 // Need xform to put node in hash table
965 IfNode *iff = create_and_xform_if( control(), tst, prob, cnt );
966 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser");
967 // True branch
968 { PreserveJVMState pjvms(this);
969 Node* iftrue = _gvn.transform( new (C) IfTrueNode (iff) );
970 set_control(iftrue);
972 if (stopped()) { // Path is dead?
973 explicit_null_checks_elided++;
974 if (C->eliminate_boxing()) {
975 // Mark the successor block as parsed
976 branch_block->next_path_num();
977 }
978 } else { // Path is live.
979 // Update method data
980 profile_taken_branch(target_bci);
981 adjust_map_after_if(btest, c, prob, branch_block, next_block);
982 if (!stopped()) {
983 merge(target_bci);
984 }
985 }
986 }
988 // False branch
989 Node* iffalse = _gvn.transform( new (C) IfFalseNode(iff) );
990 set_control(iffalse);
992 if (stopped()) { // Path is dead?
993 explicit_null_checks_elided++;
994 if (C->eliminate_boxing()) {
995 // Mark the successor block as parsed
996 next_block->next_path_num();
997 }
998 } else { // Path is live.
999 // Update method data
1000 profile_not_taken_branch();
1001 adjust_map_after_if(BoolTest(btest).negate(), c, 1.0-prob,
1002 next_block, branch_block);
1003 }
1004 }
1006 //------------------------------------do_if------------------------------------
1007 void Parse::do_if(BoolTest::mask btest, Node* c) {
1008 int target_bci = iter().get_dest();
1010 Block* branch_block = successor_for_bci(target_bci);
1011 Block* next_block = successor_for_bci(iter().next_bci());
1013 float cnt;
1014 float prob = branch_prediction(cnt, btest, target_bci);
1015 float untaken_prob = 1.0 - prob;
1017 if (prob == PROB_UNKNOWN) {
1018 #ifndef PRODUCT
1019 if (PrintOpto && Verbose)
1020 tty->print_cr("Never-taken edge stops compilation at bci %d",bci());
1021 #endif
1022 repush_if_args(); // to gather stats on loop
1023 // We need to mark this branch as taken so that if we recompile we will
1024 // see that it is possible. In the tiered system the interpreter doesn't
1025 // do profiling and by the time we get to the lower tier from the interpreter
1026 // the path may be cold again. Make sure it doesn't look untaken
1027 profile_taken_branch(target_bci, !ProfileInterpreter);
1028 uncommon_trap(Deoptimization::Reason_unreached,
1029 Deoptimization::Action_reinterpret,
1030 NULL, "cold");
1031 if (C->eliminate_boxing()) {
1032 // Mark the successor blocks as parsed
1033 branch_block->next_path_num();
1034 next_block->next_path_num();
1035 }
1036 return;
1037 }
1039 // Sanity check the probability value
1040 assert(0.0f < prob && prob < 1.0f,"Bad probability in Parser");
1042 bool taken_if_true = true;
1043 // Convert BoolTest to canonical form:
1044 if (!BoolTest(btest).is_canonical()) {
1045 btest = BoolTest(btest).negate();
1046 taken_if_true = false;
1047 // prob is NOT updated here; it remains the probability of the taken
1048 // path (as opposed to the prob of the path guarded by an 'IfTrueNode').
1049 }
1050 assert(btest != BoolTest::eq, "!= is the only canonical exact test");
1052 Node* tst0 = new (C) BoolNode(c, btest);
1053 Node* tst = _gvn.transform(tst0);
1054 BoolTest::mask taken_btest = BoolTest::illegal;
1055 BoolTest::mask untaken_btest = BoolTest::illegal;
1057 if (tst->is_Bool()) {
1058 // Refresh c from the transformed bool node, since it may be
1059 // simpler than the original c. Also re-canonicalize btest.
1060 // This wins when (Bool ne (Conv2B p) 0) => (Bool ne (CmpP p NULL)).
1061 // That can arise from statements like: if (x instanceof C) ...
1062 if (tst != tst0) {
1063 // Canonicalize one more time since transform can change it.
1064 btest = tst->as_Bool()->_test._test;
1065 if (!BoolTest(btest).is_canonical()) {
1066 // Reverse edges one more time...
1067 tst = _gvn.transform( tst->as_Bool()->negate(&_gvn) );
1068 btest = tst->as_Bool()->_test._test;
1069 assert(BoolTest(btest).is_canonical(), "sanity");
1070 taken_if_true = !taken_if_true;
1071 }
1072 c = tst->in(1);
1073 }
1074 BoolTest::mask neg_btest = BoolTest(btest).negate();
1075 taken_btest = taken_if_true ? btest : neg_btest;
1076 untaken_btest = taken_if_true ? neg_btest : btest;
1077 }
1079 // Generate real control flow
1080 float true_prob = (taken_if_true ? prob : untaken_prob);
1081 IfNode* iff = create_and_map_if(control(), tst, true_prob, cnt);
1082 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser");
1083 Node* taken_branch = new (C) IfTrueNode(iff);
1084 Node* untaken_branch = new (C) IfFalseNode(iff);
1085 if (!taken_if_true) { // Finish conversion to canonical form
1086 Node* tmp = taken_branch;
1087 taken_branch = untaken_branch;
1088 untaken_branch = tmp;
1089 }
1091 // Branch is taken:
1092 { PreserveJVMState pjvms(this);
1093 taken_branch = _gvn.transform(taken_branch);
1094 set_control(taken_branch);
1096 if (stopped()) {
1097 if (C->eliminate_boxing()) {
1098 // Mark the successor block as parsed
1099 branch_block->next_path_num();
1100 }
1101 } else {
1102 // Update method data
1103 profile_taken_branch(target_bci);
1104 adjust_map_after_if(taken_btest, c, prob, branch_block, next_block);
1105 if (!stopped()) {
1106 merge(target_bci);
1107 }
1108 }
1109 }
1111 untaken_branch = _gvn.transform(untaken_branch);
1112 set_control(untaken_branch);
1114 // Branch not taken.
1115 if (stopped()) {
1116 if (C->eliminate_boxing()) {
1117 // Mark the successor block as parsed
1118 next_block->next_path_num();
1119 }
1120 } else {
1121 // Update method data
1122 profile_not_taken_branch();
1123 adjust_map_after_if(untaken_btest, c, untaken_prob,
1124 next_block, branch_block);
1125 }
1126 }
1128 bool Parse::path_is_suitable_for_uncommon_trap(float prob) const {
1129 // Don't want to speculate on uncommon traps when running with -Xcomp
1130 if (!UseInterpreter) {
1131 return false;
1132 }
1133 return (seems_never_taken(prob) && seems_stable_comparison());
1134 }
1136 //----------------------------adjust_map_after_if------------------------------
1137 // Adjust the JVM state to reflect the result of taking this path.
1138 // Basically, it means inspecting the CmpNode controlling this
1139 // branch, seeing how it constrains a tested value, and then
1140 // deciding if it's worth our while to encode this constraint
1141 // as graph nodes in the current abstract interpretation map.
1142 void Parse::adjust_map_after_if(BoolTest::mask btest, Node* c, float prob,
1143 Block* path, Block* other_path) {
1144 if (stopped() || !c->is_Cmp() || btest == BoolTest::illegal)
1145 return; // nothing to do
1147 bool is_fallthrough = (path == successor_for_bci(iter().next_bci()));
1149 if (path_is_suitable_for_uncommon_trap(prob)) {
1150 repush_if_args();
1151 uncommon_trap(Deoptimization::Reason_unstable_if,
1152 Deoptimization::Action_reinterpret,
1153 NULL,
1154 (is_fallthrough ? "taken always" : "taken never"));
1155 return;
1156 }
1158 Node* val = c->in(1);
1159 Node* con = c->in(2);
1160 const Type* tcon = _gvn.type(con);
1161 const Type* tval = _gvn.type(val);
1162 bool have_con = tcon->singleton();
1163 if (tval->singleton()) {
1164 if (!have_con) {
1165 // Swap, so constant is in con.
1166 con = val;
1167 tcon = tval;
1168 val = c->in(2);
1169 tval = _gvn.type(val);
1170 btest = BoolTest(btest).commute();
1171 have_con = true;
1172 } else {
1173 // Do we have two constants? Then leave well enough alone.
1174 have_con = false;
1175 }
1176 }
1177 if (!have_con) // remaining adjustments need a con
1178 return;
1180 sharpen_type_after_if(btest, con, tcon, val, tval);
1181 }
1184 static Node* extract_obj_from_klass_load(PhaseGVN* gvn, Node* n) {
1185 Node* ldk;
1186 if (n->is_DecodeNKlass()) {
1187 if (n->in(1)->Opcode() != Op_LoadNKlass) {
1188 return NULL;
1189 } else {
1190 ldk = n->in(1);
1191 }
1192 } else if (n->Opcode() != Op_LoadKlass) {
1193 return NULL;
1194 } else {
1195 ldk = n;
1196 }
1197 assert(ldk != NULL && ldk->is_Load(), "should have found a LoadKlass or LoadNKlass node");
1199 Node* adr = ldk->in(MemNode::Address);
1200 intptr_t off = 0;
1201 Node* obj = AddPNode::Ideal_base_and_offset(adr, gvn, off);
1202 if (obj == NULL || off != oopDesc::klass_offset_in_bytes()) // loading oopDesc::_klass?
1203 return NULL;
1204 const TypePtr* tp = gvn->type(obj)->is_ptr();
1205 if (tp == NULL || !(tp->isa_instptr() || tp->isa_aryptr())) // is obj a Java object ptr?
1206 return NULL;
1208 return obj;
1209 }
1211 void Parse::sharpen_type_after_if(BoolTest::mask btest,
1212 Node* con, const Type* tcon,
1213 Node* val, const Type* tval) {
1214 // Look for opportunities to sharpen the type of a node
1215 // whose klass is compared with a constant klass.
1216 if (btest == BoolTest::eq && tcon->isa_klassptr()) {
1217 Node* obj = extract_obj_from_klass_load(&_gvn, val);
1218 const TypeOopPtr* con_type = tcon->isa_klassptr()->as_instance_type();
1219 if (obj != NULL && (con_type->isa_instptr() || con_type->isa_aryptr())) {
1220 // Found:
1221 // Bool(CmpP(LoadKlass(obj._klass), ConP(Foo.klass)), [eq])
1222 // or the narrowOop equivalent.
1223 const Type* obj_type = _gvn.type(obj);
1224 const TypeOopPtr* tboth = obj_type->join_speculative(con_type)->isa_oopptr();
1225 if (tboth != NULL && tboth->klass_is_exact() && tboth != obj_type &&
1226 tboth->higher_equal(obj_type)) {
1227 // obj has to be of the exact type Foo if the CmpP succeeds.
1228 int obj_in_map = map()->find_edge(obj);
1229 JVMState* jvms = this->jvms();
1230 if (obj_in_map >= 0 &&
1231 (jvms->is_loc(obj_in_map) || jvms->is_stk(obj_in_map))) {
1232 TypeNode* ccast = new (C) CheckCastPPNode(control(), obj, tboth);
1233 const Type* tcc = ccast->as_Type()->type();
1234 assert(tcc != obj_type && tcc->higher_equal_speculative(obj_type), "must improve");
1235 // Delay transform() call to allow recovery of pre-cast value
1236 // at the control merge.
1237 _gvn.set_type_bottom(ccast);
1238 record_for_igvn(ccast);
1239 // Here's the payoff.
1240 replace_in_map(obj, ccast);
1241 }
1242 }
1243 }
1244 }
1246 int val_in_map = map()->find_edge(val);
1247 if (val_in_map < 0) return; // replace_in_map would be useless
1248 {
1249 JVMState* jvms = this->jvms();
1250 if (!(jvms->is_loc(val_in_map) ||
1251 jvms->is_stk(val_in_map)))
1252 return; // again, it would be useless
1253 }
1255 // Check for a comparison to a constant, and "know" that the compared
1256 // value is constrained on this path.
1257 assert(tcon->singleton(), "");
1258 ConstraintCastNode* ccast = NULL;
1259 Node* cast = NULL;
1261 switch (btest) {
1262 case BoolTest::eq: // Constant test?
1263 {
1264 const Type* tboth = tcon->join_speculative(tval);
1265 if (tboth == tval) break; // Nothing to gain.
1266 if (tcon->isa_int()) {
1267 ccast = new (C) CastIINode(val, tboth);
1268 } else if (tcon == TypePtr::NULL_PTR) {
1269 // Cast to null, but keep the pointer identity temporarily live.
1270 ccast = new (C) CastPPNode(val, tboth);
1271 } else {
1272 const TypeF* tf = tcon->isa_float_constant();
1273 const TypeD* td = tcon->isa_double_constant();
1274 // Exclude tests vs float/double 0 as these could be
1275 // either +0 or -0. Just because you are equal to +0
1276 // doesn't mean you ARE +0!
1277 // Note, following code also replaces Long and Oop values.
1278 if ((!tf || tf->_f != 0.0) &&
1279 (!td || td->_d != 0.0))
1280 cast = con; // Replace non-constant val by con.
1281 }
1282 }
1283 break;
1285 case BoolTest::ne:
1286 if (tcon == TypePtr::NULL_PTR) {
1287 cast = cast_not_null(val, false);
1288 }
1289 break;
1291 default:
1292 // (At this point we could record int range types with CastII.)
1293 break;
1294 }
1296 if (ccast != NULL) {
1297 const Type* tcc = ccast->as_Type()->type();
1298 assert(tcc != tval && tcc->higher_equal_speculative(tval), "must improve");
1299 // Delay transform() call to allow recovery of pre-cast value
1300 // at the control merge.
1301 ccast->set_req(0, control());
1302 _gvn.set_type_bottom(ccast);
1303 record_for_igvn(ccast);
1304 cast = ccast;
1305 }
1307 if (cast != NULL) { // Here's the payoff.
1308 replace_in_map(val, cast);
1309 }
1310 }
1312 /**
1313 * Use speculative type to optimize CmpP node: if comparison is
1314 * against the low level class, cast the object to the speculative
1315 * type if any. CmpP should then go away.
1316 *
1317 * @param c expected CmpP node
1318 * @return result of CmpP on object casted to speculative type
1319 *
1320 */
1321 Node* Parse::optimize_cmp_with_klass(Node* c) {
1322 // If this is transformed by the _gvn to a comparison with the low
1323 // level klass then we may be able to use speculation
1324 if (c->Opcode() == Op_CmpP &&
1325 (c->in(1)->Opcode() == Op_LoadKlass || c->in(1)->Opcode() == Op_DecodeNKlass) &&
1326 c->in(2)->is_Con()) {
1327 Node* load_klass = NULL;
1328 Node* decode = NULL;
1329 if (c->in(1)->Opcode() == Op_DecodeNKlass) {
1330 decode = c->in(1);
1331 load_klass = c->in(1)->in(1);
1332 } else {
1333 load_klass = c->in(1);
1334 }
1335 if (load_klass->in(2)->is_AddP()) {
1336 Node* addp = load_klass->in(2);
1337 Node* obj = addp->in(AddPNode::Address);
1338 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
1339 if (obj_type->speculative_type() != NULL) {
1340 ciKlass* k = obj_type->speculative_type();
1341 inc_sp(2);
1342 obj = maybe_cast_profiled_obj(obj, k);
1343 dec_sp(2);
1344 // Make the CmpP use the casted obj
1345 addp = basic_plus_adr(obj, addp->in(AddPNode::Offset));
1346 load_klass = load_klass->clone();
1347 load_klass->set_req(2, addp);
1348 load_klass = _gvn.transform(load_klass);
1349 if (decode != NULL) {
1350 decode = decode->clone();
1351 decode->set_req(1, load_klass);
1352 load_klass = _gvn.transform(decode);
1353 }
1354 c = c->clone();
1355 c->set_req(1, load_klass);
1356 c = _gvn.transform(c);
1357 }
1358 }
1359 }
1360 return c;
1361 }
1363 //------------------------------do_one_bytecode--------------------------------
1364 // Parse this bytecode, and alter the Parsers JVM->Node mapping
1365 void Parse::do_one_bytecode() {
1366 Node *a, *b, *c, *d; // Handy temps
1367 BoolTest::mask btest;
1368 int i;
1370 assert(!has_exceptions(), "bytecode entry state must be clear of throws");
1372 if (C->check_node_count(NodeLimitFudgeFactor * 5,
1373 "out of nodes parsing method")) {
1374 return;
1375 }
1377 #ifdef ASSERT
1378 // for setting breakpoints
1379 if (TraceOptoParse) {
1380 tty->print(" @");
1381 dump_bci(bci());
1382 tty->cr();
1383 }
1384 #endif
1386 switch (bc()) {
1387 case Bytecodes::_nop:
1388 // do nothing
1389 break;
1390 case Bytecodes::_lconst_0:
1391 push_pair(longcon(0));
1392 break;
1394 case Bytecodes::_lconst_1:
1395 push_pair(longcon(1));
1396 break;
1398 case Bytecodes::_fconst_0:
1399 push(zerocon(T_FLOAT));
1400 break;
1402 case Bytecodes::_fconst_1:
1403 push(makecon(TypeF::ONE));
1404 break;
1406 case Bytecodes::_fconst_2:
1407 push(makecon(TypeF::make(2.0f)));
1408 break;
1410 case Bytecodes::_dconst_0:
1411 push_pair(zerocon(T_DOUBLE));
1412 break;
1414 case Bytecodes::_dconst_1:
1415 push_pair(makecon(TypeD::ONE));
1416 break;
1418 case Bytecodes::_iconst_m1:push(intcon(-1)); break;
1419 case Bytecodes::_iconst_0: push(intcon( 0)); break;
1420 case Bytecodes::_iconst_1: push(intcon( 1)); break;
1421 case Bytecodes::_iconst_2: push(intcon( 2)); break;
1422 case Bytecodes::_iconst_3: push(intcon( 3)); break;
1423 case Bytecodes::_iconst_4: push(intcon( 4)); break;
1424 case Bytecodes::_iconst_5: push(intcon( 5)); break;
1425 case Bytecodes::_bipush: push(intcon(iter().get_constant_u1())); break;
1426 case Bytecodes::_sipush: push(intcon(iter().get_constant_u2())); break;
1427 case Bytecodes::_aconst_null: push(null()); break;
1428 case Bytecodes::_ldc:
1429 case Bytecodes::_ldc_w:
1430 case Bytecodes::_ldc2_w:
1431 // If the constant is unresolved, run this BC once in the interpreter.
1432 {
1433 ciConstant constant = iter().get_constant();
1434 if (constant.basic_type() == T_OBJECT &&
1435 !constant.as_object()->is_loaded()) {
1436 int index = iter().get_constant_pool_index();
1437 constantTag tag = iter().get_constant_pool_tag(index);
1438 uncommon_trap(Deoptimization::make_trap_request
1439 (Deoptimization::Reason_unloaded,
1440 Deoptimization::Action_reinterpret,
1441 index),
1442 NULL, tag.internal_name());
1443 break;
1444 }
1445 assert(constant.basic_type() != T_OBJECT || constant.as_object()->is_instance(),
1446 "must be java_mirror of klass");
1447 bool pushed = push_constant(constant, true);
1448 guarantee(pushed, "must be possible to push this constant");
1449 }
1451 break;
1453 case Bytecodes::_aload_0:
1454 push( local(0) );
1455 break;
1456 case Bytecodes::_aload_1:
1457 push( local(1) );
1458 break;
1459 case Bytecodes::_aload_2:
1460 push( local(2) );
1461 break;
1462 case Bytecodes::_aload_3:
1463 push( local(3) );
1464 break;
1465 case Bytecodes::_aload:
1466 push( local(iter().get_index()) );
1467 break;
1469 case Bytecodes::_fload_0:
1470 case Bytecodes::_iload_0:
1471 push( local(0) );
1472 break;
1473 case Bytecodes::_fload_1:
1474 case Bytecodes::_iload_1:
1475 push( local(1) );
1476 break;
1477 case Bytecodes::_fload_2:
1478 case Bytecodes::_iload_2:
1479 push( local(2) );
1480 break;
1481 case Bytecodes::_fload_3:
1482 case Bytecodes::_iload_3:
1483 push( local(3) );
1484 break;
1485 case Bytecodes::_fload:
1486 case Bytecodes::_iload:
1487 push( local(iter().get_index()) );
1488 break;
1489 case Bytecodes::_lload_0:
1490 push_pair_local( 0 );
1491 break;
1492 case Bytecodes::_lload_1:
1493 push_pair_local( 1 );
1494 break;
1495 case Bytecodes::_lload_2:
1496 push_pair_local( 2 );
1497 break;
1498 case Bytecodes::_lload_3:
1499 push_pair_local( 3 );
1500 break;
1501 case Bytecodes::_lload:
1502 push_pair_local( iter().get_index() );
1503 break;
1505 case Bytecodes::_dload_0:
1506 push_pair_local(0);
1507 break;
1508 case Bytecodes::_dload_1:
1509 push_pair_local(1);
1510 break;
1511 case Bytecodes::_dload_2:
1512 push_pair_local(2);
1513 break;
1514 case Bytecodes::_dload_3:
1515 push_pair_local(3);
1516 break;
1517 case Bytecodes::_dload:
1518 push_pair_local(iter().get_index());
1519 break;
1520 case Bytecodes::_fstore_0:
1521 case Bytecodes::_istore_0:
1522 case Bytecodes::_astore_0:
1523 set_local( 0, pop() );
1524 break;
1525 case Bytecodes::_fstore_1:
1526 case Bytecodes::_istore_1:
1527 case Bytecodes::_astore_1:
1528 set_local( 1, pop() );
1529 break;
1530 case Bytecodes::_fstore_2:
1531 case Bytecodes::_istore_2:
1532 case Bytecodes::_astore_2:
1533 set_local( 2, pop() );
1534 break;
1535 case Bytecodes::_fstore_3:
1536 case Bytecodes::_istore_3:
1537 case Bytecodes::_astore_3:
1538 set_local( 3, pop() );
1539 break;
1540 case Bytecodes::_fstore:
1541 case Bytecodes::_istore:
1542 case Bytecodes::_astore:
1543 set_local( iter().get_index(), pop() );
1544 break;
1545 // long stores
1546 case Bytecodes::_lstore_0:
1547 set_pair_local( 0, pop_pair() );
1548 break;
1549 case Bytecodes::_lstore_1:
1550 set_pair_local( 1, pop_pair() );
1551 break;
1552 case Bytecodes::_lstore_2:
1553 set_pair_local( 2, pop_pair() );
1554 break;
1555 case Bytecodes::_lstore_3:
1556 set_pair_local( 3, pop_pair() );
1557 break;
1558 case Bytecodes::_lstore:
1559 set_pair_local( iter().get_index(), pop_pair() );
1560 break;
1562 // double stores
1563 case Bytecodes::_dstore_0:
1564 set_pair_local( 0, dstore_rounding(pop_pair()) );
1565 break;
1566 case Bytecodes::_dstore_1:
1567 set_pair_local( 1, dstore_rounding(pop_pair()) );
1568 break;
1569 case Bytecodes::_dstore_2:
1570 set_pair_local( 2, dstore_rounding(pop_pair()) );
1571 break;
1572 case Bytecodes::_dstore_3:
1573 set_pair_local( 3, dstore_rounding(pop_pair()) );
1574 break;
1575 case Bytecodes::_dstore:
1576 set_pair_local( iter().get_index(), dstore_rounding(pop_pair()) );
1577 break;
1579 case Bytecodes::_pop: dec_sp(1); break;
1580 case Bytecodes::_pop2: dec_sp(2); break;
1581 case Bytecodes::_swap:
1582 a = pop();
1583 b = pop();
1584 push(a);
1585 push(b);
1586 break;
1587 case Bytecodes::_dup:
1588 a = pop();
1589 push(a);
1590 push(a);
1591 break;
1592 case Bytecodes::_dup_x1:
1593 a = pop();
1594 b = pop();
1595 push( a );
1596 push( b );
1597 push( a );
1598 break;
1599 case Bytecodes::_dup_x2:
1600 a = pop();
1601 b = pop();
1602 c = pop();
1603 push( a );
1604 push( c );
1605 push( b );
1606 push( a );
1607 break;
1608 case Bytecodes::_dup2:
1609 a = pop();
1610 b = pop();
1611 push( b );
1612 push( a );
1613 push( b );
1614 push( a );
1615 break;
1617 case Bytecodes::_dup2_x1:
1618 // before: .. c, b, a
1619 // after: .. b, a, c, b, a
1620 // not tested
1621 a = pop();
1622 b = pop();
1623 c = pop();
1624 push( b );
1625 push( a );
1626 push( c );
1627 push( b );
1628 push( a );
1629 break;
1630 case Bytecodes::_dup2_x2:
1631 // before: .. d, c, b, a
1632 // after: .. b, a, d, c, b, a
1633 // not tested
1634 a = pop();
1635 b = pop();
1636 c = pop();
1637 d = pop();
1638 push( b );
1639 push( a );
1640 push( d );
1641 push( c );
1642 push( b );
1643 push( a );
1644 break;
1646 case Bytecodes::_arraylength: {
1647 // Must do null-check with value on expression stack
1648 Node *ary = null_check(peek(), T_ARRAY);
1649 // Compile-time detect of null-exception?
1650 if (stopped()) return;
1651 a = pop();
1652 push(load_array_length(a));
1653 break;
1654 }
1656 case Bytecodes::_baload: array_load(T_BYTE); break;
1657 case Bytecodes::_caload: array_load(T_CHAR); break;
1658 case Bytecodes::_iaload: array_load(T_INT); break;
1659 case Bytecodes::_saload: array_load(T_SHORT); break;
1660 case Bytecodes::_faload: array_load(T_FLOAT); break;
1661 case Bytecodes::_aaload: array_load(T_OBJECT); break;
1662 case Bytecodes::_laload: {
1663 a = array_addressing(T_LONG, 0);
1664 if (stopped()) return; // guaranteed null or range check
1665 dec_sp(2); // Pop array and index
1666 push_pair(make_load(control(), a, TypeLong::LONG, T_LONG, TypeAryPtr::LONGS, MemNode::unordered));
1667 break;
1668 }
1669 case Bytecodes::_daload: {
1670 a = array_addressing(T_DOUBLE, 0);
1671 if (stopped()) return; // guaranteed null or range check
1672 dec_sp(2); // Pop array and index
1673 push_pair(make_load(control(), a, Type::DOUBLE, T_DOUBLE, TypeAryPtr::DOUBLES, MemNode::unordered));
1674 break;
1675 }
1676 case Bytecodes::_bastore: array_store(T_BYTE); break;
1677 case Bytecodes::_castore: array_store(T_CHAR); break;
1678 case Bytecodes::_iastore: array_store(T_INT); break;
1679 case Bytecodes::_sastore: array_store(T_SHORT); break;
1680 case Bytecodes::_fastore: array_store(T_FLOAT); break;
1681 case Bytecodes::_aastore: {
1682 d = array_addressing(T_OBJECT, 1);
1683 if (stopped()) return; // guaranteed null or range check
1684 array_store_check();
1685 c = pop(); // Oop to store
1686 b = pop(); // index (already used)
1687 a = pop(); // the array itself
1688 const TypeOopPtr* elemtype = _gvn.type(a)->is_aryptr()->elem()->make_oopptr();
1689 const TypeAryPtr* adr_type = TypeAryPtr::OOPS;
1690 Node* store = store_oop_to_array(control(), a, d, adr_type, c, elemtype, T_OBJECT, MemNode::release);
1691 break;
1692 }
1693 case Bytecodes::_lastore: {
1694 a = array_addressing(T_LONG, 2);
1695 if (stopped()) return; // guaranteed null or range check
1696 c = pop_pair();
1697 dec_sp(2); // Pop array and index
1698 store_to_memory(control(), a, c, T_LONG, TypeAryPtr::LONGS, MemNode::unordered);
1699 break;
1700 }
1701 case Bytecodes::_dastore: {
1702 a = array_addressing(T_DOUBLE, 2);
1703 if (stopped()) return; // guaranteed null or range check
1704 c = pop_pair();
1705 dec_sp(2); // Pop array and index
1706 c = dstore_rounding(c);
1707 store_to_memory(control(), a, c, T_DOUBLE, TypeAryPtr::DOUBLES, MemNode::unordered);
1708 break;
1709 }
1710 case Bytecodes::_getfield:
1711 do_getfield();
1712 break;
1714 case Bytecodes::_getstatic:
1715 do_getstatic();
1716 break;
1718 case Bytecodes::_putfield:
1719 do_putfield();
1720 break;
1722 case Bytecodes::_putstatic:
1723 do_putstatic();
1724 break;
1726 case Bytecodes::_irem:
1727 do_irem();
1728 break;
1729 case Bytecodes::_idiv:
1730 // Must keep both values on the expression-stack during null-check
1731 zero_check_int(peek());
1732 // Compile-time detect of null-exception?
1733 if (stopped()) return;
1734 b = pop();
1735 a = pop();
1736 push( _gvn.transform( new (C) DivINode(control(),a,b) ) );
1737 break;
1738 case Bytecodes::_imul:
1739 b = pop(); a = pop();
1740 push( _gvn.transform( new (C) MulINode(a,b) ) );
1741 break;
1742 case Bytecodes::_iadd:
1743 b = pop(); a = pop();
1744 push( _gvn.transform( new (C) AddINode(a,b) ) );
1745 break;
1746 case Bytecodes::_ineg:
1747 a = pop();
1748 push( _gvn.transform( new (C) SubINode(_gvn.intcon(0),a)) );
1749 break;
1750 case Bytecodes::_isub:
1751 b = pop(); a = pop();
1752 push( _gvn.transform( new (C) SubINode(a,b) ) );
1753 break;
1754 case Bytecodes::_iand:
1755 b = pop(); a = pop();
1756 push( _gvn.transform( new (C) AndINode(a,b) ) );
1757 break;
1758 case Bytecodes::_ior:
1759 b = pop(); a = pop();
1760 push( _gvn.transform( new (C) OrINode(a,b) ) );
1761 break;
1762 case Bytecodes::_ixor:
1763 b = pop(); a = pop();
1764 push( _gvn.transform( new (C) XorINode(a,b) ) );
1765 break;
1766 case Bytecodes::_ishl:
1767 b = pop(); a = pop();
1768 push( _gvn.transform( new (C) LShiftINode(a,b) ) );
1769 break;
1770 case Bytecodes::_ishr:
1771 b = pop(); a = pop();
1772 push( _gvn.transform( new (C) RShiftINode(a,b) ) );
1773 break;
1774 case Bytecodes::_iushr:
1775 b = pop(); a = pop();
1776 push( _gvn.transform( new (C) URShiftINode(a,b) ) );
1777 break;
1779 case Bytecodes::_fneg:
1780 a = pop();
1781 b = _gvn.transform(new (C) NegFNode (a));
1782 push(b);
1783 break;
1785 case Bytecodes::_fsub:
1786 b = pop();
1787 a = pop();
1788 c = _gvn.transform( new (C) SubFNode(a,b) );
1789 d = precision_rounding(c);
1790 push( d );
1791 break;
1793 case Bytecodes::_fadd:
1794 b = pop();
1795 a = pop();
1796 c = _gvn.transform( new (C) AddFNode(a,b) );
1797 d = precision_rounding(c);
1798 push( d );
1799 break;
1801 case Bytecodes::_fmul:
1802 b = pop();
1803 a = pop();
1804 c = _gvn.transform( new (C) MulFNode(a,b) );
1805 d = precision_rounding(c);
1806 push( d );
1807 break;
1809 case Bytecodes::_fdiv:
1810 b = pop();
1811 a = pop();
1812 c = _gvn.transform( new (C) DivFNode(0,a,b) );
1813 d = precision_rounding(c);
1814 push( d );
1815 break;
1817 case Bytecodes::_frem:
1818 if (Matcher::has_match_rule(Op_ModF)) {
1819 // Generate a ModF node.
1820 b = pop();
1821 a = pop();
1822 c = _gvn.transform( new (C) ModFNode(0,a,b) );
1823 d = precision_rounding(c);
1824 push( d );
1825 }
1826 else {
1827 // Generate a call.
1828 modf();
1829 }
1830 break;
1832 case Bytecodes::_fcmpl:
1833 b = pop();
1834 a = pop();
1835 c = _gvn.transform( new (C) CmpF3Node( a, b));
1836 push(c);
1837 break;
1838 case Bytecodes::_fcmpg:
1839 b = pop();
1840 a = pop();
1842 // Same as fcmpl but need to flip the unordered case. Swap the inputs,
1843 // which negates the result sign except for unordered. Flip the unordered
1844 // as well by using CmpF3 which implements unordered-lesser instead of
1845 // unordered-greater semantics. Finally, commute the result bits. Result
1846 // is same as using a CmpF3Greater except we did it with CmpF3 alone.
1847 c = _gvn.transform( new (C) CmpF3Node( b, a));
1848 c = _gvn.transform( new (C) SubINode(_gvn.intcon(0),c) );
1849 push(c);
1850 break;
1852 case Bytecodes::_f2i:
1853 a = pop();
1854 push(_gvn.transform(new (C) ConvF2INode(a)));
1855 break;
1857 case Bytecodes::_d2i:
1858 a = pop_pair();
1859 b = _gvn.transform(new (C) ConvD2INode(a));
1860 push( b );
1861 break;
1863 case Bytecodes::_f2d:
1864 a = pop();
1865 b = _gvn.transform( new (C) ConvF2DNode(a));
1866 push_pair( b );
1867 break;
1869 case Bytecodes::_d2f:
1870 a = pop_pair();
1871 b = _gvn.transform( new (C) ConvD2FNode(a));
1872 // This breaks _227_mtrt (speed & correctness) and _222_mpegaudio (speed)
1873 //b = _gvn.transform(new (C) RoundFloatNode(0, b) );
1874 push( b );
1875 break;
1877 case Bytecodes::_l2f:
1878 if (Matcher::convL2FSupported()) {
1879 a = pop_pair();
1880 b = _gvn.transform( new (C) ConvL2FNode(a));
1881 // For i486.ad, FILD doesn't restrict precision to 24 or 53 bits.
1882 // Rather than storing the result into an FP register then pushing
1883 // out to memory to round, the machine instruction that implements
1884 // ConvL2D is responsible for rounding.
1885 // c = precision_rounding(b);
1886 c = _gvn.transform(b);
1887 push(c);
1888 } else {
1889 l2f();
1890 }
1891 break;
1893 case Bytecodes::_l2d:
1894 a = pop_pair();
1895 b = _gvn.transform( new (C) ConvL2DNode(a));
1896 // For i486.ad, rounding is always necessary (see _l2f above).
1897 // c = dprecision_rounding(b);
1898 c = _gvn.transform(b);
1899 push_pair(c);
1900 break;
1902 case Bytecodes::_f2l:
1903 a = pop();
1904 b = _gvn.transform( new (C) ConvF2LNode(a));
1905 push_pair(b);
1906 break;
1908 case Bytecodes::_d2l:
1909 a = pop_pair();
1910 b = _gvn.transform( new (C) ConvD2LNode(a));
1911 push_pair(b);
1912 break;
1914 case Bytecodes::_dsub:
1915 b = pop_pair();
1916 a = pop_pair();
1917 c = _gvn.transform( new (C) SubDNode(a,b) );
1918 d = dprecision_rounding(c);
1919 push_pair( d );
1920 break;
1922 case Bytecodes::_dadd:
1923 b = pop_pair();
1924 a = pop_pair();
1925 c = _gvn.transform( new (C) AddDNode(a,b) );
1926 d = dprecision_rounding(c);
1927 push_pair( d );
1928 break;
1930 case Bytecodes::_dmul:
1931 b = pop_pair();
1932 a = pop_pair();
1933 c = _gvn.transform( new (C) MulDNode(a,b) );
1934 d = dprecision_rounding(c);
1935 push_pair( d );
1936 break;
1938 case Bytecodes::_ddiv:
1939 b = pop_pair();
1940 a = pop_pair();
1941 c = _gvn.transform( new (C) DivDNode(0,a,b) );
1942 d = dprecision_rounding(c);
1943 push_pair( d );
1944 break;
1946 case Bytecodes::_dneg:
1947 a = pop_pair();
1948 b = _gvn.transform(new (C) NegDNode (a));
1949 push_pair(b);
1950 break;
1952 case Bytecodes::_drem:
1953 if (Matcher::has_match_rule(Op_ModD)) {
1954 // Generate a ModD node.
1955 b = pop_pair();
1956 a = pop_pair();
1957 // a % b
1959 c = _gvn.transform( new (C) ModDNode(0,a,b) );
1960 d = dprecision_rounding(c);
1961 push_pair( d );
1962 }
1963 else {
1964 // Generate a call.
1965 modd();
1966 }
1967 break;
1969 case Bytecodes::_dcmpl:
1970 b = pop_pair();
1971 a = pop_pair();
1972 c = _gvn.transform( new (C) CmpD3Node( a, b));
1973 push(c);
1974 break;
1976 case Bytecodes::_dcmpg:
1977 b = pop_pair();
1978 a = pop_pair();
1979 // Same as dcmpl but need to flip the unordered case.
1980 // Commute the inputs, which negates the result sign except for unordered.
1981 // Flip the unordered as well by using CmpD3 which implements
1982 // unordered-lesser instead of unordered-greater semantics.
1983 // Finally, negate the result bits. Result is same as using a
1984 // CmpD3Greater except we did it with CmpD3 alone.
1985 c = _gvn.transform( new (C) CmpD3Node( b, a));
1986 c = _gvn.transform( new (C) SubINode(_gvn.intcon(0),c) );
1987 push(c);
1988 break;
1991 // Note for longs -> lo word is on TOS, hi word is on TOS - 1
1992 case Bytecodes::_land:
1993 b = pop_pair();
1994 a = pop_pair();
1995 c = _gvn.transform( new (C) AndLNode(a,b) );
1996 push_pair(c);
1997 break;
1998 case Bytecodes::_lor:
1999 b = pop_pair();
2000 a = pop_pair();
2001 c = _gvn.transform( new (C) OrLNode(a,b) );
2002 push_pair(c);
2003 break;
2004 case Bytecodes::_lxor:
2005 b = pop_pair();
2006 a = pop_pair();
2007 c = _gvn.transform( new (C) XorLNode(a,b) );
2008 push_pair(c);
2009 break;
2011 case Bytecodes::_lshl:
2012 b = pop(); // the shift count
2013 a = pop_pair(); // value to be shifted
2014 c = _gvn.transform( new (C) LShiftLNode(a,b) );
2015 push_pair(c);
2016 break;
2017 case Bytecodes::_lshr:
2018 b = pop(); // the shift count
2019 a = pop_pair(); // value to be shifted
2020 c = _gvn.transform( new (C) RShiftLNode(a,b) );
2021 push_pair(c);
2022 break;
2023 case Bytecodes::_lushr:
2024 b = pop(); // the shift count
2025 a = pop_pair(); // value to be shifted
2026 c = _gvn.transform( new (C) URShiftLNode(a,b) );
2027 push_pair(c);
2028 break;
2029 case Bytecodes::_lmul:
2030 b = pop_pair();
2031 a = pop_pair();
2032 c = _gvn.transform( new (C) MulLNode(a,b) );
2033 push_pair(c);
2034 break;
2036 case Bytecodes::_lrem:
2037 // Must keep both values on the expression-stack during null-check
2038 assert(peek(0) == top(), "long word order");
2039 zero_check_long(peek(1));
2040 // Compile-time detect of null-exception?
2041 if (stopped()) return;
2042 b = pop_pair();
2043 a = pop_pair();
2044 c = _gvn.transform( new (C) ModLNode(control(),a,b) );
2045 push_pair(c);
2046 break;
2048 case Bytecodes::_ldiv:
2049 // Must keep both values on the expression-stack during null-check
2050 assert(peek(0) == top(), "long word order");
2051 zero_check_long(peek(1));
2052 // Compile-time detect of null-exception?
2053 if (stopped()) return;
2054 b = pop_pair();
2055 a = pop_pair();
2056 c = _gvn.transform( new (C) DivLNode(control(),a,b) );
2057 push_pair(c);
2058 break;
2060 case Bytecodes::_ladd:
2061 b = pop_pair();
2062 a = pop_pair();
2063 c = _gvn.transform( new (C) AddLNode(a,b) );
2064 push_pair(c);
2065 break;
2066 case Bytecodes::_lsub:
2067 b = pop_pair();
2068 a = pop_pair();
2069 c = _gvn.transform( new (C) SubLNode(a,b) );
2070 push_pair(c);
2071 break;
2072 case Bytecodes::_lcmp:
2073 // Safepoints are now inserted _before_ branches. The long-compare
2074 // bytecode painfully produces a 3-way value (-1,0,+1) which requires a
2075 // slew of control flow. These are usually followed by a CmpI vs zero and
2076 // a branch; this pattern then optimizes to the obvious long-compare and
2077 // branch. However, if the branch is backwards there's a Safepoint
2078 // inserted. The inserted Safepoint captures the JVM state at the
2079 // pre-branch point, i.e. it captures the 3-way value. Thus if a
2080 // long-compare is used to control a loop the debug info will force
2081 // computation of the 3-way value, even though the generated code uses a
2082 // long-compare and branch. We try to rectify the situation by inserting
2083 // a SafePoint here and have it dominate and kill the safepoint added at a
2084 // following backwards branch. At this point the JVM state merely holds 2
2085 // longs but not the 3-way value.
2086 if( UseLoopSafepoints ) {
2087 switch( iter().next_bc() ) {
2088 case Bytecodes::_ifgt:
2089 case Bytecodes::_iflt:
2090 case Bytecodes::_ifge:
2091 case Bytecodes::_ifle:
2092 case Bytecodes::_ifne:
2093 case Bytecodes::_ifeq:
2094 // If this is a backwards branch in the bytecodes, add Safepoint
2095 maybe_add_safepoint(iter().next_get_dest());
2096 }
2097 }
2098 b = pop_pair();
2099 a = pop_pair();
2100 c = _gvn.transform( new (C) CmpL3Node( a, b ));
2101 push(c);
2102 break;
2104 case Bytecodes::_lneg:
2105 a = pop_pair();
2106 b = _gvn.transform( new (C) SubLNode(longcon(0),a));
2107 push_pair(b);
2108 break;
2109 case Bytecodes::_l2i:
2110 a = pop_pair();
2111 push( _gvn.transform( new (C) ConvL2INode(a)));
2112 break;
2113 case Bytecodes::_i2l:
2114 a = pop();
2115 b = _gvn.transform( new (C) ConvI2LNode(a));
2116 push_pair(b);
2117 break;
2118 case Bytecodes::_i2b:
2119 // Sign extend
2120 a = pop();
2121 a = _gvn.transform( new (C) LShiftINode(a,_gvn.intcon(24)) );
2122 a = _gvn.transform( new (C) RShiftINode(a,_gvn.intcon(24)) );
2123 push( a );
2124 break;
2125 case Bytecodes::_i2s:
2126 a = pop();
2127 a = _gvn.transform( new (C) LShiftINode(a,_gvn.intcon(16)) );
2128 a = _gvn.transform( new (C) RShiftINode(a,_gvn.intcon(16)) );
2129 push( a );
2130 break;
2131 case Bytecodes::_i2c:
2132 a = pop();
2133 push( _gvn.transform( new (C) AndINode(a,_gvn.intcon(0xFFFF)) ) );
2134 break;
2136 case Bytecodes::_i2f:
2137 a = pop();
2138 b = _gvn.transform( new (C) ConvI2FNode(a) ) ;
2139 c = precision_rounding(b);
2140 push (b);
2141 break;
2143 case Bytecodes::_i2d:
2144 a = pop();
2145 b = _gvn.transform( new (C) ConvI2DNode(a));
2146 push_pair(b);
2147 break;
2149 case Bytecodes::_iinc: // Increment local
2150 i = iter().get_index(); // Get local index
2151 set_local( i, _gvn.transform( new (C) AddINode( _gvn.intcon(iter().get_iinc_con()), local(i) ) ) );
2152 break;
2154 // Exit points of synchronized methods must have an unlock node
2155 case Bytecodes::_return:
2156 return_current(NULL);
2157 break;
2159 case Bytecodes::_ireturn:
2160 case Bytecodes::_areturn:
2161 case Bytecodes::_freturn:
2162 return_current(pop());
2163 break;
2164 case Bytecodes::_lreturn:
2165 return_current(pop_pair());
2166 break;
2167 case Bytecodes::_dreturn:
2168 return_current(pop_pair());
2169 break;
2171 case Bytecodes::_athrow:
2172 // null exception oop throws NULL pointer exception
2173 null_check(peek());
2174 if (stopped()) return;
2175 // Hook the thrown exception directly to subsequent handlers.
2176 if (BailoutToInterpreterForThrows) {
2177 // Keep method interpreted from now on.
2178 uncommon_trap(Deoptimization::Reason_unhandled,
2179 Deoptimization::Action_make_not_compilable);
2180 return;
2181 }
2182 if (env()->jvmti_can_post_on_exceptions()) {
2183 // check if we must post exception events, take uncommon trap if so (with must_throw = false)
2184 uncommon_trap_if_should_post_on_exceptions(Deoptimization::Reason_unhandled, false);
2185 }
2186 // Here if either can_post_on_exceptions or should_post_on_exceptions is false
2187 add_exception_state(make_exception_state(peek()));
2188 break;
2190 case Bytecodes::_goto: // fall through
2191 case Bytecodes::_goto_w: {
2192 int target_bci = (bc() == Bytecodes::_goto) ? iter().get_dest() : iter().get_far_dest();
2194 // If this is a backwards branch in the bytecodes, add Safepoint
2195 maybe_add_safepoint(target_bci);
2197 // Update method data
2198 profile_taken_branch(target_bci);
2200 // Merge the current control into the target basic block
2201 merge(target_bci);
2203 // See if we can get some profile data and hand it off to the next block
2204 Block *target_block = block()->successor_for_bci(target_bci);
2205 if (target_block->pred_count() != 1) break;
2206 ciMethodData* methodData = method()->method_data();
2207 if (!methodData->is_mature()) break;
2208 ciProfileData* data = methodData->bci_to_data(bci());
2209 assert( data->is_JumpData(), "" );
2210 int taken = ((ciJumpData*)data)->taken();
2211 taken = method()->scale_count(taken);
2212 target_block->set_count(taken);
2213 break;
2214 }
2216 case Bytecodes::_ifnull: btest = BoolTest::eq; goto handle_if_null;
2217 case Bytecodes::_ifnonnull: btest = BoolTest::ne; goto handle_if_null;
2218 handle_if_null:
2219 // If this is a backwards branch in the bytecodes, add Safepoint
2220 maybe_add_safepoint(iter().get_dest());
2221 a = null();
2222 b = pop();
2223 c = _gvn.transform( new (C) CmpPNode(b, a) );
2224 do_ifnull(btest, c);
2225 break;
2227 case Bytecodes::_if_acmpeq: btest = BoolTest::eq; goto handle_if_acmp;
2228 case Bytecodes::_if_acmpne: btest = BoolTest::ne; goto handle_if_acmp;
2229 handle_if_acmp:
2230 // If this is a backwards branch in the bytecodes, add Safepoint
2231 maybe_add_safepoint(iter().get_dest());
2232 a = pop();
2233 b = pop();
2234 c = _gvn.transform( new (C) CmpPNode(b, a) );
2235 c = optimize_cmp_with_klass(c);
2236 do_if(btest, c);
2237 break;
2239 case Bytecodes::_ifeq: btest = BoolTest::eq; goto handle_ifxx;
2240 case Bytecodes::_ifne: btest = BoolTest::ne; goto handle_ifxx;
2241 case Bytecodes::_iflt: btest = BoolTest::lt; goto handle_ifxx;
2242 case Bytecodes::_ifle: btest = BoolTest::le; goto handle_ifxx;
2243 case Bytecodes::_ifgt: btest = BoolTest::gt; goto handle_ifxx;
2244 case Bytecodes::_ifge: btest = BoolTest::ge; goto handle_ifxx;
2245 handle_ifxx:
2246 // If this is a backwards branch in the bytecodes, add Safepoint
2247 maybe_add_safepoint(iter().get_dest());
2248 a = _gvn.intcon(0);
2249 b = pop();
2250 c = _gvn.transform( new (C) CmpINode(b, a) );
2251 do_if(btest, c);
2252 break;
2254 case Bytecodes::_if_icmpeq: btest = BoolTest::eq; goto handle_if_icmp;
2255 case Bytecodes::_if_icmpne: btest = BoolTest::ne; goto handle_if_icmp;
2256 case Bytecodes::_if_icmplt: btest = BoolTest::lt; goto handle_if_icmp;
2257 case Bytecodes::_if_icmple: btest = BoolTest::le; goto handle_if_icmp;
2258 case Bytecodes::_if_icmpgt: btest = BoolTest::gt; goto handle_if_icmp;
2259 case Bytecodes::_if_icmpge: btest = BoolTest::ge; goto handle_if_icmp;
2260 handle_if_icmp:
2261 // If this is a backwards branch in the bytecodes, add Safepoint
2262 maybe_add_safepoint(iter().get_dest());
2263 a = pop();
2264 b = pop();
2265 c = _gvn.transform( new (C) CmpINode( b, a ) );
2266 do_if(btest, c);
2267 break;
2269 case Bytecodes::_tableswitch:
2270 do_tableswitch();
2271 break;
2273 case Bytecodes::_lookupswitch:
2274 do_lookupswitch();
2275 break;
2277 case Bytecodes::_invokestatic:
2278 case Bytecodes::_invokedynamic:
2279 case Bytecodes::_invokespecial:
2280 case Bytecodes::_invokevirtual:
2281 case Bytecodes::_invokeinterface:
2282 do_call();
2283 break;
2284 case Bytecodes::_checkcast:
2285 do_checkcast();
2286 break;
2287 case Bytecodes::_instanceof:
2288 do_instanceof();
2289 break;
2290 case Bytecodes::_anewarray:
2291 do_anewarray();
2292 break;
2293 case Bytecodes::_newarray:
2294 do_newarray((BasicType)iter().get_index());
2295 break;
2296 case Bytecodes::_multianewarray:
2297 do_multianewarray();
2298 break;
2299 case Bytecodes::_new:
2300 do_new();
2301 break;
2303 case Bytecodes::_jsr:
2304 case Bytecodes::_jsr_w:
2305 do_jsr();
2306 break;
2308 case Bytecodes::_ret:
2309 do_ret();
2310 break;
2313 case Bytecodes::_monitorenter:
2314 do_monitor_enter();
2315 break;
2317 case Bytecodes::_monitorexit:
2318 do_monitor_exit();
2319 break;
2321 case Bytecodes::_breakpoint:
2322 // Breakpoint set concurrently to compile
2323 // %%% use an uncommon trap?
2324 C->record_failure("breakpoint in method");
2325 return;
2327 default:
2328 #ifndef PRODUCT
2329 map()->dump(99);
2330 #endif
2331 tty->print("\nUnhandled bytecode %s\n", Bytecodes::name(bc()) );
2332 ShouldNotReachHere();
2333 }
2335 #ifndef PRODUCT
2336 IdealGraphPrinter *printer = IdealGraphPrinter::printer();
2337 if(printer) {
2338 char buffer[256];
2339 sprintf(buffer, "Bytecode %d: %s", bci(), Bytecodes::name(bc()));
2340 bool old = printer->traverse_outs();
2341 printer->set_traverse_outs(true);
2342 printer->print_method(C, buffer, 4);
2343 printer->set_traverse_outs(old);
2344 }
2345 #endif
2346 }