Mon, 26 Sep 2011 10:24:05 -0700
7081933: Use zeroing elimination optimization for large array
Summary: Don't zero new typeArray during runtime call if the allocation is followed by arraycopy into it.
Reviewed-by: twisti
1 /*
2 * Copyright (c) 1998, 2011, 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 _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);
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 _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);
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 = do_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(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 if (!arytype->klass()->is_loaded()) {
108 // Only fails for some -Xcomp runs
109 // The class is unloaded. We have to run this bytecode in the interpreter.
110 uncommon_trap(Deoptimization::Reason_unloaded,
111 Deoptimization::Action_reinterpret,
112 arytype->klass(), "!loaded array");
113 return top();
114 }
116 // Do the range check
117 if (GenerateRangeChecks && need_range_check) {
118 Node* tst;
119 if (sizetype->_hi <= 0) {
120 // The greatest array bound is negative, so we can conclude that we're
121 // compiling unreachable code, but the unsigned compare trick used below
122 // only works with non-negative lengths. Instead, hack "tst" to be zero so
123 // the uncommon_trap path will always be taken.
124 tst = _gvn.intcon(0);
125 } else {
126 // Range is constant in array-oop, so we can use the original state of mem
127 Node* len = load_array_length(ary);
129 // Test length vs index (standard trick using unsigned compare)
130 Node* chk = _gvn.transform( new (C, 3) CmpUNode(idx, len) );
131 BoolTest::mask btest = BoolTest::lt;
132 tst = _gvn.transform( new (C, 2) BoolNode(chk, btest) );
133 }
134 // Branch to failure if out of bounds
135 { BuildCutout unless(this, tst, PROB_MAX);
136 if (C->allow_range_check_smearing()) {
137 // Do not use builtin_throw, since range checks are sometimes
138 // made more stringent by an optimistic transformation.
139 // This creates "tentative" range checks at this point,
140 // which are not guaranteed to throw exceptions.
141 // See IfNode::Ideal, is_range_check, adjust_check.
142 uncommon_trap(Deoptimization::Reason_range_check,
143 Deoptimization::Action_make_not_entrant,
144 NULL, "range_check");
145 } else {
146 // If we have already recompiled with the range-check-widening
147 // heroic optimization turned off, then we must really be throwing
148 // range check exceptions.
149 builtin_throw(Deoptimization::Reason_range_check, idx);
150 }
151 }
152 }
153 // Check for always knowing you are throwing a range-check exception
154 if (stopped()) return top();
156 Node* ptr = array_element_address(ary, idx, type, sizetype);
158 if (result2 != NULL) *result2 = elemtype;
160 assert(ptr != top(), "top should go hand-in-hand with stopped");
162 return ptr;
163 }
166 // returns IfNode
167 IfNode* Parse::jump_if_fork_int(Node* a, Node* b, BoolTest::mask mask) {
168 Node *cmp = _gvn.transform( new (C, 3) CmpINode( a, b)); // two cases: shiftcount > 32 and shiftcount <= 32
169 Node *tst = _gvn.transform( new (C, 2) BoolNode( cmp, mask));
170 IfNode *iff = create_and_map_if( control(), tst, ((mask == BoolTest::eq) ? PROB_STATIC_INFREQUENT : PROB_FAIR), COUNT_UNKNOWN );
171 return iff;
172 }
174 // return Region node
175 Node* Parse::jump_if_join(Node* iffalse, Node* iftrue) {
176 Node *region = new (C, 3) RegionNode(3); // 2 results
177 record_for_igvn(region);
178 region->init_req(1, iffalse);
179 region->init_req(2, iftrue );
180 _gvn.set_type(region, Type::CONTROL);
181 region = _gvn.transform(region);
182 set_control (region);
183 return region;
184 }
187 //------------------------------helper for tableswitch-------------------------
188 void Parse::jump_if_true_fork(IfNode *iff, int dest_bci_if_true, int prof_table_index) {
189 // True branch, use existing map info
190 { PreserveJVMState pjvms(this);
191 Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode (iff) );
192 set_control( iftrue );
193 profile_switch_case(prof_table_index);
194 merge_new_path(dest_bci_if_true);
195 }
197 // False branch
198 Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(iff) );
199 set_control( iffalse );
200 }
202 void Parse::jump_if_false_fork(IfNode *iff, int dest_bci_if_true, int prof_table_index) {
203 // True branch, use existing map info
204 { PreserveJVMState pjvms(this);
205 Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode (iff) );
206 set_control( iffalse );
207 profile_switch_case(prof_table_index);
208 merge_new_path(dest_bci_if_true);
209 }
211 // False branch
212 Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(iff) );
213 set_control( iftrue );
214 }
216 void Parse::jump_if_always_fork(int dest_bci, int prof_table_index) {
217 // False branch, use existing map and control()
218 profile_switch_case(prof_table_index);
219 merge_new_path(dest_bci);
220 }
223 extern "C" {
224 static int jint_cmp(const void *i, const void *j) {
225 int a = *(jint *)i;
226 int b = *(jint *)j;
227 return a > b ? 1 : a < b ? -1 : 0;
228 }
229 }
232 // Default value for methodData switch indexing. Must be a negative value to avoid
233 // conflict with any legal switch index.
234 #define NullTableIndex -1
236 class SwitchRange : public StackObj {
237 // a range of integers coupled with a bci destination
238 jint _lo; // inclusive lower limit
239 jint _hi; // inclusive upper limit
240 int _dest;
241 int _table_index; // index into method data table
243 public:
244 jint lo() const { return _lo; }
245 jint hi() const { return _hi; }
246 int dest() const { return _dest; }
247 int table_index() const { return _table_index; }
248 bool is_singleton() const { return _lo == _hi; }
250 void setRange(jint lo, jint hi, int dest, int table_index) {
251 assert(lo <= hi, "must be a non-empty range");
252 _lo = lo, _hi = hi; _dest = dest; _table_index = table_index;
253 }
254 bool adjoinRange(jint lo, jint hi, int dest, int table_index) {
255 assert(lo <= hi, "must be a non-empty range");
256 if (lo == _hi+1 && dest == _dest && table_index == _table_index) {
257 _hi = hi;
258 return true;
259 }
260 return false;
261 }
263 void set (jint value, int dest, int table_index) {
264 setRange(value, value, dest, table_index);
265 }
266 bool adjoin(jint value, int dest, int table_index) {
267 return adjoinRange(value, value, dest, table_index);
268 }
270 void print(ciEnv* env) {
271 if (is_singleton())
272 tty->print(" {%d}=>%d", lo(), dest());
273 else if (lo() == min_jint)
274 tty->print(" {..%d}=>%d", hi(), dest());
275 else if (hi() == max_jint)
276 tty->print(" {%d..}=>%d", lo(), dest());
277 else
278 tty->print(" {%d..%d}=>%d", lo(), hi(), dest());
279 }
280 };
283 //-------------------------------do_tableswitch--------------------------------
284 void Parse::do_tableswitch() {
285 Node* lookup = pop();
287 // Get information about tableswitch
288 int default_dest = iter().get_dest_table(0);
289 int lo_index = iter().get_int_table(1);
290 int hi_index = iter().get_int_table(2);
291 int len = hi_index - lo_index + 1;
293 if (len < 1) {
294 // If this is a backward branch, add safepoint
295 maybe_add_safepoint(default_dest);
296 merge(default_dest);
297 return;
298 }
300 // generate decision tree, using trichotomy when possible
301 int rnum = len+2;
302 bool makes_backward_branch = false;
303 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum);
304 int rp = -1;
305 if (lo_index != min_jint) {
306 ranges[++rp].setRange(min_jint, lo_index-1, default_dest, NullTableIndex);
307 }
308 for (int j = 0; j < len; j++) {
309 jint match_int = lo_index+j;
310 int dest = iter().get_dest_table(j+3);
311 makes_backward_branch |= (dest <= bci());
312 int table_index = method_data_update() ? j : NullTableIndex;
313 if (rp < 0 || !ranges[rp].adjoin(match_int, dest, table_index)) {
314 ranges[++rp].set(match_int, dest, table_index);
315 }
316 }
317 jint highest = lo_index+(len-1);
318 assert(ranges[rp].hi() == highest, "");
319 if (highest != max_jint
320 && !ranges[rp].adjoinRange(highest+1, max_jint, default_dest, NullTableIndex)) {
321 ranges[++rp].setRange(highest+1, max_jint, default_dest, NullTableIndex);
322 }
323 assert(rp < len+2, "not too many ranges");
325 // Safepoint in case if backward branch observed
326 if( makes_backward_branch && UseLoopSafepoints )
327 add_safepoint();
329 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]);
330 }
333 //------------------------------do_lookupswitch--------------------------------
334 void Parse::do_lookupswitch() {
335 Node *lookup = pop(); // lookup value
336 // Get information about lookupswitch
337 int default_dest = iter().get_dest_table(0);
338 int len = iter().get_int_table(1);
340 if (len < 1) { // If this is a backward branch, add safepoint
341 maybe_add_safepoint(default_dest);
342 merge(default_dest);
343 return;
344 }
346 // generate decision tree, using trichotomy when possible
347 jint* table = NEW_RESOURCE_ARRAY(jint, len*2);
348 {
349 for( int j = 0; j < len; j++ ) {
350 table[j+j+0] = iter().get_int_table(2+j+j);
351 table[j+j+1] = iter().get_dest_table(2+j+j+1);
352 }
353 qsort( table, len, 2*sizeof(table[0]), jint_cmp );
354 }
356 int rnum = len*2+1;
357 bool makes_backward_branch = false;
358 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum);
359 int rp = -1;
360 for( int j = 0; j < len; j++ ) {
361 jint match_int = table[j+j+0];
362 int dest = table[j+j+1];
363 int next_lo = rp < 0 ? min_jint : ranges[rp].hi()+1;
364 int table_index = method_data_update() ? j : NullTableIndex;
365 makes_backward_branch |= (dest <= bci());
366 if( match_int != next_lo ) {
367 ranges[++rp].setRange(next_lo, match_int-1, default_dest, NullTableIndex);
368 }
369 if( rp < 0 || !ranges[rp].adjoin(match_int, dest, table_index) ) {
370 ranges[++rp].set(match_int, dest, table_index);
371 }
372 }
373 jint highest = table[2*(len-1)];
374 assert(ranges[rp].hi() == highest, "");
375 if( highest != max_jint
376 && !ranges[rp].adjoinRange(highest+1, max_jint, default_dest, NullTableIndex) ) {
377 ranges[++rp].setRange(highest+1, max_jint, default_dest, NullTableIndex);
378 }
379 assert(rp < rnum, "not too many ranges");
381 // Safepoint in case backward branch observed
382 if( makes_backward_branch && UseLoopSafepoints )
383 add_safepoint();
385 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]);
386 }
388 //----------------------------create_jump_tables-------------------------------
389 bool Parse::create_jump_tables(Node* key_val, SwitchRange* lo, SwitchRange* hi) {
390 // Are jumptables enabled
391 if (!UseJumpTables) return false;
393 // Are jumptables supported
394 if (!Matcher::has_match_rule(Op_Jump)) return false;
396 // Don't make jump table if profiling
397 if (method_data_update()) return false;
399 // Decide if a guard is needed to lop off big ranges at either (or
400 // both) end(s) of the input set. We'll call this the default target
401 // even though we can't be sure that it is the true "default".
403 bool needs_guard = false;
404 int default_dest;
405 int64 total_outlier_size = 0;
406 int64 hi_size = ((int64)hi->hi()) - ((int64)hi->lo()) + 1;
407 int64 lo_size = ((int64)lo->hi()) - ((int64)lo->lo()) + 1;
409 if (lo->dest() == hi->dest()) {
410 total_outlier_size = hi_size + lo_size;
411 default_dest = lo->dest();
412 } else if (lo_size > hi_size) {
413 total_outlier_size = lo_size;
414 default_dest = lo->dest();
415 } else {
416 total_outlier_size = hi_size;
417 default_dest = hi->dest();
418 }
420 // If a guard test will eliminate very sparse end ranges, then
421 // it is worth the cost of an extra jump.
422 if (total_outlier_size > (MaxJumpTableSparseness * 4)) {
423 needs_guard = true;
424 if (default_dest == lo->dest()) lo++;
425 if (default_dest == hi->dest()) hi--;
426 }
428 // Find the total number of cases and ranges
429 int64 num_cases = ((int64)hi->hi()) - ((int64)lo->lo()) + 1;
430 int num_range = hi - lo + 1;
432 // Don't create table if: too large, too small, or too sparse.
433 if (num_cases < MinJumpTableSize || num_cases > MaxJumpTableSize)
434 return false;
435 if (num_cases > (MaxJumpTableSparseness * num_range))
436 return false;
438 // Normalize table lookups to zero
439 int lowval = lo->lo();
440 key_val = _gvn.transform( new (C, 3) SubINode(key_val, _gvn.intcon(lowval)) );
442 // Generate a guard to protect against input keyvals that aren't
443 // in the switch domain.
444 if (needs_guard) {
445 Node* size = _gvn.intcon(num_cases);
446 Node* cmp = _gvn.transform( new (C, 3) CmpUNode(key_val, size) );
447 Node* tst = _gvn.transform( new (C, 2) BoolNode(cmp, BoolTest::ge) );
448 IfNode* iff = create_and_map_if( control(), tst, PROB_FAIR, COUNT_UNKNOWN);
449 jump_if_true_fork(iff, default_dest, NullTableIndex);
450 }
452 // Create an ideal node JumpTable that has projections
453 // of all possible ranges for a switch statement
454 // The key_val input must be converted to a pointer offset and scaled.
455 // Compare Parse::array_addressing above.
456 #ifdef _LP64
457 // Clean the 32-bit int into a real 64-bit offset.
458 // Otherwise, the jint value 0 might turn into an offset of 0x0800000000.
459 const TypeLong* lkeytype = TypeLong::make(CONST64(0), num_cases-1, Type::WidenMin);
460 key_val = _gvn.transform( new (C, 2) ConvI2LNode(key_val, lkeytype) );
461 #endif
462 // Shift the value by wordsize so we have an index into the table, rather
463 // than a switch value
464 Node *shiftWord = _gvn.MakeConX(wordSize);
465 key_val = _gvn.transform( new (C, 3) MulXNode( key_val, shiftWord));
467 // Create the JumpNode
468 Node* jtn = _gvn.transform( new (C, 2) JumpNode(control(), key_val, num_cases) );
470 // These are the switch destinations hanging off the jumpnode
471 int i = 0;
472 for (SwitchRange* r = lo; r <= hi; r++) {
473 for (int j = r->lo(); j <= r->hi(); j++, i++) {
474 Node* input = _gvn.transform(new (C, 1) JumpProjNode(jtn, i, r->dest(), j - lowval));
475 {
476 PreserveJVMState pjvms(this);
477 set_control(input);
478 jump_if_always_fork(r->dest(), r->table_index());
479 }
480 }
481 }
482 assert(i == num_cases, "miscount of cases");
483 stop_and_kill_map(); // no more uses for this JVMS
484 return true;
485 }
487 //----------------------------jump_switch_ranges-------------------------------
488 void Parse::jump_switch_ranges(Node* key_val, SwitchRange *lo, SwitchRange *hi, int switch_depth) {
489 Block* switch_block = block();
491 if (switch_depth == 0) {
492 // Do special processing for the top-level call.
493 assert(lo->lo() == min_jint, "initial range must exhaust Type::INT");
494 assert(hi->hi() == max_jint, "initial range must exhaust Type::INT");
496 // Decrement pred-numbers for the unique set of nodes.
497 #ifdef ASSERT
498 // Ensure that the block's successors are a (duplicate-free) set.
499 int successors_counted = 0; // block occurrences in [hi..lo]
500 int unique_successors = switch_block->num_successors();
501 for (int i = 0; i < unique_successors; i++) {
502 Block* target = switch_block->successor_at(i);
504 // Check that the set of successors is the same in both places.
505 int successors_found = 0;
506 for (SwitchRange* p = lo; p <= hi; p++) {
507 if (p->dest() == target->start()) successors_found++;
508 }
509 assert(successors_found > 0, "successor must be known");
510 successors_counted += successors_found;
511 }
512 assert(successors_counted == (hi-lo)+1, "no unexpected successors");
513 #endif
515 // Maybe prune the inputs, based on the type of key_val.
516 jint min_val = min_jint;
517 jint max_val = max_jint;
518 const TypeInt* ti = key_val->bottom_type()->isa_int();
519 if (ti != NULL) {
520 min_val = ti->_lo;
521 max_val = ti->_hi;
522 assert(min_val <= max_val, "invalid int type");
523 }
524 while (lo->hi() < min_val) lo++;
525 if (lo->lo() < min_val) lo->setRange(min_val, lo->hi(), lo->dest(), lo->table_index());
526 while (hi->lo() > max_val) hi--;
527 if (hi->hi() > max_val) hi->setRange(hi->lo(), max_val, hi->dest(), hi->table_index());
528 }
530 #ifndef PRODUCT
531 if (switch_depth == 0) {
532 _max_switch_depth = 0;
533 _est_switch_depth = log2_intptr((hi-lo+1)-1)+1;
534 }
535 #endif
537 assert(lo <= hi, "must be a non-empty set of ranges");
538 if (lo == hi) {
539 jump_if_always_fork(lo->dest(), lo->table_index());
540 } else {
541 assert(lo->hi() == (lo+1)->lo()-1, "contiguous ranges");
542 assert(hi->lo() == (hi-1)->hi()+1, "contiguous ranges");
544 if (create_jump_tables(key_val, lo, hi)) return;
546 int nr = hi - lo + 1;
548 SwitchRange* mid = lo + nr/2;
549 // if there is an easy choice, pivot at a singleton:
550 if (nr > 3 && !mid->is_singleton() && (mid-1)->is_singleton()) mid--;
552 assert(lo < mid && mid <= hi, "good pivot choice");
553 assert(nr != 2 || mid == hi, "should pick higher of 2");
554 assert(nr != 3 || mid == hi-1, "should pick middle of 3");
556 Node *test_val = _gvn.intcon(mid->lo());
558 if (mid->is_singleton()) {
559 IfNode *iff_ne = jump_if_fork_int(key_val, test_val, BoolTest::ne);
560 jump_if_false_fork(iff_ne, mid->dest(), mid->table_index());
562 // Special Case: If there are exactly three ranges, and the high
563 // and low range each go to the same place, omit the "gt" test,
564 // since it will not discriminate anything.
565 bool eq_test_only = (hi == lo+2 && hi->dest() == lo->dest());
566 if (eq_test_only) {
567 assert(mid == hi-1, "");
568 }
570 // if there is a higher range, test for it and process it:
571 if (mid < hi && !eq_test_only) {
572 // two comparisons of same values--should enable 1 test for 2 branches
573 // Use BoolTest::le instead of BoolTest::gt
574 IfNode *iff_le = jump_if_fork_int(key_val, test_val, BoolTest::le);
575 Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(iff_le) );
576 Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(iff_le) );
577 { PreserveJVMState pjvms(this);
578 set_control(iffalse);
579 jump_switch_ranges(key_val, mid+1, hi, switch_depth+1);
580 }
581 set_control(iftrue);
582 }
584 } else {
585 // mid is a range, not a singleton, so treat mid..hi as a unit
586 IfNode *iff_ge = jump_if_fork_int(key_val, test_val, BoolTest::ge);
588 // if there is a higher range, test for it and process it:
589 if (mid == hi) {
590 jump_if_true_fork(iff_ge, mid->dest(), mid->table_index());
591 } else {
592 Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(iff_ge) );
593 Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(iff_ge) );
594 { PreserveJVMState pjvms(this);
595 set_control(iftrue);
596 jump_switch_ranges(key_val, mid, hi, switch_depth+1);
597 }
598 set_control(iffalse);
599 }
600 }
602 // in any case, process the lower range
603 jump_switch_ranges(key_val, lo, mid-1, switch_depth+1);
604 }
606 // Decrease pred_count for each successor after all is done.
607 if (switch_depth == 0) {
608 int unique_successors = switch_block->num_successors();
609 for (int i = 0; i < unique_successors; i++) {
610 Block* target = switch_block->successor_at(i);
611 // Throw away the pre-allocated path for each unique successor.
612 target->next_path_num();
613 }
614 }
616 #ifndef PRODUCT
617 _max_switch_depth = MAX2(switch_depth, _max_switch_depth);
618 if (TraceOptoParse && Verbose && WizardMode && switch_depth == 0) {
619 SwitchRange* r;
620 int nsing = 0;
621 for( r = lo; r <= hi; r++ ) {
622 if( r->is_singleton() ) nsing++;
623 }
624 tty->print(">>> ");
625 _method->print_short_name();
626 tty->print_cr(" switch decision tree");
627 tty->print_cr(" %d ranges (%d singletons), max_depth=%d, est_depth=%d",
628 hi-lo+1, nsing, _max_switch_depth, _est_switch_depth);
629 if (_max_switch_depth > _est_switch_depth) {
630 tty->print_cr("******** BAD SWITCH DEPTH ********");
631 }
632 tty->print(" ");
633 for( r = lo; r <= hi; r++ ) {
634 r->print(env());
635 }
636 tty->print_cr("");
637 }
638 #endif
639 }
641 void Parse::modf() {
642 Node *f2 = pop();
643 Node *f1 = pop();
644 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::modf_Type(),
645 CAST_FROM_FN_PTR(address, SharedRuntime::frem),
646 "frem", NULL, //no memory effects
647 f1, f2);
648 Node* res = _gvn.transform(new (C, 1) ProjNode(c, TypeFunc::Parms + 0));
650 push(res);
651 }
653 void Parse::modd() {
654 Node *d2 = pop_pair();
655 Node *d1 = pop_pair();
656 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::Math_DD_D_Type(),
657 CAST_FROM_FN_PTR(address, SharedRuntime::drem),
658 "drem", NULL, //no memory effects
659 d1, top(), d2, top());
660 Node* res_d = _gvn.transform(new (C, 1) ProjNode(c, TypeFunc::Parms + 0));
662 #ifdef ASSERT
663 Node* res_top = _gvn.transform(new (C, 1) ProjNode(c, TypeFunc::Parms + 1));
664 assert(res_top == top(), "second value must be top");
665 #endif
667 push_pair(res_d);
668 }
670 void Parse::l2f() {
671 Node* f2 = pop();
672 Node* f1 = pop();
673 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::l2f_Type(),
674 CAST_FROM_FN_PTR(address, SharedRuntime::l2f),
675 "l2f", NULL, //no memory effects
676 f1, f2);
677 Node* res = _gvn.transform(new (C, 1) ProjNode(c, TypeFunc::Parms + 0));
679 push(res);
680 }
682 void Parse::do_irem() {
683 // Must keep both values on the expression-stack during null-check
684 do_null_check(peek(), T_INT);
685 // Compile-time detect of null-exception?
686 if (stopped()) return;
688 Node* b = pop();
689 Node* a = pop();
691 const Type *t = _gvn.type(b);
692 if (t != Type::TOP) {
693 const TypeInt *ti = t->is_int();
694 if (ti->is_con()) {
695 int divisor = ti->get_con();
696 // check for positive power of 2
697 if (divisor > 0 &&
698 (divisor & ~(divisor-1)) == divisor) {
699 // yes !
700 Node *mask = _gvn.intcon((divisor - 1));
701 // Sigh, must handle negative dividends
702 Node *zero = _gvn.intcon(0);
703 IfNode *ifff = jump_if_fork_int(a, zero, BoolTest::lt);
704 Node *iff = _gvn.transform( new (C, 1) IfFalseNode(ifff) );
705 Node *ift = _gvn.transform( new (C, 1) IfTrueNode (ifff) );
706 Node *reg = jump_if_join(ift, iff);
707 Node *phi = PhiNode::make(reg, NULL, TypeInt::INT);
708 // Negative path; negate/and/negate
709 Node *neg = _gvn.transform( new (C, 3) SubINode(zero, a) );
710 Node *andn= _gvn.transform( new (C, 3) AndINode(neg, mask) );
711 Node *negn= _gvn.transform( new (C, 3) SubINode(zero, andn) );
712 phi->init_req(1, negn);
713 // Fast positive case
714 Node *andx = _gvn.transform( new (C, 3) AndINode(a, mask) );
715 phi->init_req(2, andx);
716 // Push the merge
717 push( _gvn.transform(phi) );
718 return;
719 }
720 }
721 }
722 // Default case
723 push( _gvn.transform( new (C, 3) ModINode(control(),a,b) ) );
724 }
726 // Handle jsr and jsr_w bytecode
727 void Parse::do_jsr() {
728 assert(bc() == Bytecodes::_jsr || bc() == Bytecodes::_jsr_w, "wrong bytecode");
730 // Store information about current state, tagged with new _jsr_bci
731 int return_bci = iter().next_bci();
732 int jsr_bci = (bc() == Bytecodes::_jsr) ? iter().get_dest() : iter().get_far_dest();
734 // Update method data
735 profile_taken_branch(jsr_bci);
737 // The way we do things now, there is only one successor block
738 // for the jsr, because the target code is cloned by ciTypeFlow.
739 Block* target = successor_for_bci(jsr_bci);
741 // What got pushed?
742 const Type* ret_addr = target->peek();
743 assert(ret_addr->singleton(), "must be a constant (cloned jsr body)");
745 // Effect on jsr on stack
746 push(_gvn.makecon(ret_addr));
748 // Flow to the jsr.
749 merge(jsr_bci);
750 }
752 // Handle ret bytecode
753 void Parse::do_ret() {
754 // Find to whom we return.
755 assert(block()->num_successors() == 1, "a ret can only go one place now");
756 Block* target = block()->successor_at(0);
757 assert(!target->is_ready(), "our arrival must be expected");
758 profile_ret(target->flow()->start());
759 int pnum = target->next_path_num();
760 merge_common(target, pnum);
761 }
763 //--------------------------dynamic_branch_prediction--------------------------
764 // Try to gather dynamic branch prediction behavior. Return a probability
765 // of the branch being taken and set the "cnt" field. Returns a -1.0
766 // if we need to use static prediction for some reason.
767 float Parse::dynamic_branch_prediction(float &cnt) {
768 ResourceMark rm;
770 cnt = COUNT_UNKNOWN;
772 // Use MethodData information if it is available
773 // FIXME: free the ProfileData structure
774 ciMethodData* methodData = method()->method_data();
775 if (!methodData->is_mature()) return PROB_UNKNOWN;
776 ciProfileData* data = methodData->bci_to_data(bci());
777 if (!data->is_JumpData()) return PROB_UNKNOWN;
779 // get taken and not taken values
780 int taken = data->as_JumpData()->taken();
781 int not_taken = 0;
782 if (data->is_BranchData()) {
783 not_taken = data->as_BranchData()->not_taken();
784 }
786 // scale the counts to be commensurate with invocation counts:
787 taken = method()->scale_count(taken);
788 not_taken = method()->scale_count(not_taken);
790 // Give up if too few (or too many, in which case the sum will overflow) counts to be meaningful.
791 // We also check that individual counters are positive first, overwise the sum can become positive.
792 if (taken < 0 || not_taken < 0 || taken + not_taken < 40) {
793 if (C->log() != NULL) {
794 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d'", iter().get_dest(), taken, not_taken);
795 }
796 return PROB_UNKNOWN;
797 }
799 // Compute frequency that we arrive here
800 float sum = taken + not_taken;
801 // Adjust, if this block is a cloned private block but the
802 // Jump counts are shared. Taken the private counts for
803 // just this path instead of the shared counts.
804 if( block()->count() > 0 )
805 sum = block()->count();
806 cnt = sum / FreqCountInvocations;
808 // Pin probability to sane limits
809 float prob;
810 if( !taken )
811 prob = (0+PROB_MIN) / 2;
812 else if( !not_taken )
813 prob = (1+PROB_MAX) / 2;
814 else { // Compute probability of true path
815 prob = (float)taken / (float)(taken + not_taken);
816 if (prob > PROB_MAX) prob = PROB_MAX;
817 if (prob < PROB_MIN) prob = PROB_MIN;
818 }
820 assert((cnt > 0.0f) && (prob > 0.0f),
821 "Bad frequency assignment in if");
823 if (C->log() != NULL) {
824 const char* prob_str = NULL;
825 if (prob >= PROB_MAX) prob_str = (prob == PROB_MAX) ? "max" : "always";
826 if (prob <= PROB_MIN) prob_str = (prob == PROB_MIN) ? "min" : "never";
827 char prob_str_buf[30];
828 if (prob_str == NULL) {
829 sprintf(prob_str_buf, "%g", prob);
830 prob_str = prob_str_buf;
831 }
832 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d' cnt='%g' prob='%s'",
833 iter().get_dest(), taken, not_taken, cnt, prob_str);
834 }
835 return prob;
836 }
838 //-----------------------------branch_prediction-------------------------------
839 float Parse::branch_prediction(float& cnt,
840 BoolTest::mask btest,
841 int target_bci) {
842 float prob = dynamic_branch_prediction(cnt);
843 // If prob is unknown, switch to static prediction
844 if (prob != PROB_UNKNOWN) return prob;
846 prob = PROB_FAIR; // Set default value
847 if (btest == BoolTest::eq) // Exactly equal test?
848 prob = PROB_STATIC_INFREQUENT; // Assume its relatively infrequent
849 else if (btest == BoolTest::ne)
850 prob = PROB_STATIC_FREQUENT; // Assume its relatively frequent
852 // If this is a conditional test guarding a backwards branch,
853 // assume its a loop-back edge. Make it a likely taken branch.
854 if (target_bci < bci()) {
855 if (is_osr_parse()) { // Could be a hot OSR'd loop; force deopt
856 // Since it's an OSR, we probably have profile data, but since
857 // branch_prediction returned PROB_UNKNOWN, the counts are too small.
858 // Let's make a special check here for completely zero counts.
859 ciMethodData* methodData = method()->method_data();
860 if (!methodData->is_empty()) {
861 ciProfileData* data = methodData->bci_to_data(bci());
862 // Only stop for truly zero counts, which mean an unknown part
863 // of the OSR-ed method, and we want to deopt to gather more stats.
864 // If you have ANY counts, then this loop is simply 'cold' relative
865 // to the OSR loop.
866 if (data->as_BranchData()->taken() +
867 data->as_BranchData()->not_taken() == 0 ) {
868 // This is the only way to return PROB_UNKNOWN:
869 return PROB_UNKNOWN;
870 }
871 }
872 }
873 prob = PROB_STATIC_FREQUENT; // Likely to take backwards branch
874 }
876 assert(prob != PROB_UNKNOWN, "must have some guess at this point");
877 return prob;
878 }
880 // The magic constants are chosen so as to match the output of
881 // branch_prediction() when the profile reports a zero taken count.
882 // It is important to distinguish zero counts unambiguously, because
883 // some branches (e.g., _213_javac.Assembler.eliminate) validly produce
884 // very small but nonzero probabilities, which if confused with zero
885 // counts would keep the program recompiling indefinitely.
886 bool Parse::seems_never_taken(float prob) {
887 return prob < PROB_MIN;
888 }
890 // True if the comparison seems to be the kind that will not change its
891 // statistics from true to false. See comments in adjust_map_after_if.
892 // This question is only asked along paths which are already
893 // classifed as untaken (by seems_never_taken), so really,
894 // if a path is never taken, its controlling comparison is
895 // already acting in a stable fashion. If the comparison
896 // seems stable, we will put an expensive uncommon trap
897 // on the untaken path. To be conservative, and to allow
898 // partially executed counted loops to be compiled fully,
899 // we will plant uncommon traps only after pointer comparisons.
900 bool Parse::seems_stable_comparison(BoolTest::mask btest, Node* cmp) {
901 for (int depth = 4; depth > 0; depth--) {
902 // The following switch can find CmpP here over half the time for
903 // dynamic language code rich with type tests.
904 // Code using counted loops or array manipulations (typical
905 // of benchmarks) will have many (>80%) CmpI instructions.
906 switch (cmp->Opcode()) {
907 case Op_CmpP:
908 // A never-taken null check looks like CmpP/BoolTest::eq.
909 // These certainly should be closed off as uncommon traps.
910 if (btest == BoolTest::eq)
911 return true;
912 // A never-failed type check looks like CmpP/BoolTest::ne.
913 // Let's put traps on those, too, so that we don't have to compile
914 // unused paths with indeterminate dynamic type information.
915 if (ProfileDynamicTypes)
916 return true;
917 return false;
919 case Op_CmpI:
920 // A small minority (< 10%) of CmpP are masked as CmpI,
921 // as if by boolean conversion ((p == q? 1: 0) != 0).
922 // Detect that here, even if it hasn't optimized away yet.
923 // Specifically, this covers the 'instanceof' operator.
924 if (btest == BoolTest::ne || btest == BoolTest::eq) {
925 if (_gvn.type(cmp->in(2))->singleton() &&
926 cmp->in(1)->is_Phi()) {
927 PhiNode* phi = cmp->in(1)->as_Phi();
928 int true_path = phi->is_diamond_phi();
929 if (true_path > 0 &&
930 _gvn.type(phi->in(1))->singleton() &&
931 _gvn.type(phi->in(2))->singleton()) {
932 // phi->region->if_proj->ifnode->bool->cmp
933 BoolNode* bol = phi->in(0)->in(1)->in(0)->in(1)->as_Bool();
934 btest = bol->_test._test;
935 cmp = bol->in(1);
936 continue;
937 }
938 }
939 }
940 return false;
941 }
942 }
943 return false;
944 }
946 //-------------------------------repush_if_args--------------------------------
947 // Push arguments of an "if" bytecode back onto the stack by adjusting _sp.
948 inline int Parse::repush_if_args() {
949 #ifndef PRODUCT
950 if (PrintOpto && WizardMode) {
951 tty->print("defending against excessive implicit null exceptions on %s @%d in ",
952 Bytecodes::name(iter().cur_bc()), iter().cur_bci());
953 method()->print_name(); tty->cr();
954 }
955 #endif
956 int bc_depth = - Bytecodes::depth(iter().cur_bc());
957 assert(bc_depth == 1 || bc_depth == 2, "only two kinds of branches");
958 DEBUG_ONLY(sync_jvms()); // argument(n) requires a synced jvms
959 assert(argument(0) != NULL, "must exist");
960 assert(bc_depth == 1 || argument(1) != NULL, "two must exist");
961 _sp += bc_depth;
962 return bc_depth;
963 }
965 //----------------------------------do_ifnull----------------------------------
966 void Parse::do_ifnull(BoolTest::mask btest, Node *c) {
967 int target_bci = iter().get_dest();
969 Block* branch_block = successor_for_bci(target_bci);
970 Block* next_block = successor_for_bci(iter().next_bci());
972 float cnt;
973 float prob = branch_prediction(cnt, btest, target_bci);
974 if (prob == PROB_UNKNOWN) {
975 // (An earlier version of do_ifnull omitted this trap for OSR methods.)
976 #ifndef PRODUCT
977 if (PrintOpto && Verbose)
978 tty->print_cr("Never-taken edge stops compilation at bci %d",bci());
979 #endif
980 repush_if_args(); // to gather stats on loop
981 // We need to mark this branch as taken so that if we recompile we will
982 // see that it is possible. In the tiered system the interpreter doesn't
983 // do profiling and by the time we get to the lower tier from the interpreter
984 // the path may be cold again. Make sure it doesn't look untaken
985 profile_taken_branch(target_bci, !ProfileInterpreter);
986 uncommon_trap(Deoptimization::Reason_unreached,
987 Deoptimization::Action_reinterpret,
988 NULL, "cold");
989 if (EliminateAutoBox) {
990 // Mark the successor blocks as parsed
991 branch_block->next_path_num();
992 next_block->next_path_num();
993 }
994 return;
995 }
997 explicit_null_checks_inserted++;
999 // Generate real control flow
1000 Node *tst = _gvn.transform( new (C, 2) BoolNode( c, btest ) );
1002 // Sanity check the probability value
1003 assert(prob > 0.0f,"Bad probability in Parser");
1004 // Need xform to put node in hash table
1005 IfNode *iff = create_and_xform_if( control(), tst, prob, cnt );
1006 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser");
1007 // True branch
1008 { PreserveJVMState pjvms(this);
1009 Node* iftrue = _gvn.transform( new (C, 1) IfTrueNode (iff) );
1010 set_control(iftrue);
1012 if (stopped()) { // Path is dead?
1013 explicit_null_checks_elided++;
1014 if (EliminateAutoBox) {
1015 // Mark the successor block as parsed
1016 branch_block->next_path_num();
1017 }
1018 } else { // Path is live.
1019 // Update method data
1020 profile_taken_branch(target_bci);
1021 adjust_map_after_if(btest, c, prob, branch_block, next_block);
1022 if (!stopped()) {
1023 merge(target_bci);
1024 }
1025 }
1026 }
1028 // False branch
1029 Node* iffalse = _gvn.transform( new (C, 1) IfFalseNode(iff) );
1030 set_control(iffalse);
1032 if (stopped()) { // Path is dead?
1033 explicit_null_checks_elided++;
1034 if (EliminateAutoBox) {
1035 // Mark the successor block as parsed
1036 next_block->next_path_num();
1037 }
1038 } else { // Path is live.
1039 // Update method data
1040 profile_not_taken_branch();
1041 adjust_map_after_if(BoolTest(btest).negate(), c, 1.0-prob,
1042 next_block, branch_block);
1043 }
1044 }
1046 //------------------------------------do_if------------------------------------
1047 void Parse::do_if(BoolTest::mask btest, Node* c) {
1048 int target_bci = iter().get_dest();
1050 Block* branch_block = successor_for_bci(target_bci);
1051 Block* next_block = successor_for_bci(iter().next_bci());
1053 float cnt;
1054 float prob = branch_prediction(cnt, btest, target_bci);
1055 float untaken_prob = 1.0 - prob;
1057 if (prob == PROB_UNKNOWN) {
1058 #ifndef PRODUCT
1059 if (PrintOpto && Verbose)
1060 tty->print_cr("Never-taken edge stops compilation at bci %d",bci());
1061 #endif
1062 repush_if_args(); // to gather stats on loop
1063 // We need to mark this branch as taken so that if we recompile we will
1064 // see that it is possible. In the tiered system the interpreter doesn't
1065 // do profiling and by the time we get to the lower tier from the interpreter
1066 // the path may be cold again. Make sure it doesn't look untaken
1067 profile_taken_branch(target_bci, !ProfileInterpreter);
1068 uncommon_trap(Deoptimization::Reason_unreached,
1069 Deoptimization::Action_reinterpret,
1070 NULL, "cold");
1071 if (EliminateAutoBox) {
1072 // Mark the successor blocks as parsed
1073 branch_block->next_path_num();
1074 next_block->next_path_num();
1075 }
1076 return;
1077 }
1079 // Sanity check the probability value
1080 assert(0.0f < prob && prob < 1.0f,"Bad probability in Parser");
1082 bool taken_if_true = true;
1083 // Convert BoolTest to canonical form:
1084 if (!BoolTest(btest).is_canonical()) {
1085 btest = BoolTest(btest).negate();
1086 taken_if_true = false;
1087 // prob is NOT updated here; it remains the probability of the taken
1088 // path (as opposed to the prob of the path guarded by an 'IfTrueNode').
1089 }
1090 assert(btest != BoolTest::eq, "!= is the only canonical exact test");
1092 Node* tst0 = new (C, 2) BoolNode(c, btest);
1093 Node* tst = _gvn.transform(tst0);
1094 BoolTest::mask taken_btest = BoolTest::illegal;
1095 BoolTest::mask untaken_btest = BoolTest::illegal;
1097 if (tst->is_Bool()) {
1098 // Refresh c from the transformed bool node, since it may be
1099 // simpler than the original c. Also re-canonicalize btest.
1100 // This wins when (Bool ne (Conv2B p) 0) => (Bool ne (CmpP p NULL)).
1101 // That can arise from statements like: if (x instanceof C) ...
1102 if (tst != tst0) {
1103 // Canonicalize one more time since transform can change it.
1104 btest = tst->as_Bool()->_test._test;
1105 if (!BoolTest(btest).is_canonical()) {
1106 // Reverse edges one more time...
1107 tst = _gvn.transform( tst->as_Bool()->negate(&_gvn) );
1108 btest = tst->as_Bool()->_test._test;
1109 assert(BoolTest(btest).is_canonical(), "sanity");
1110 taken_if_true = !taken_if_true;
1111 }
1112 c = tst->in(1);
1113 }
1114 BoolTest::mask neg_btest = BoolTest(btest).negate();
1115 taken_btest = taken_if_true ? btest : neg_btest;
1116 untaken_btest = taken_if_true ? neg_btest : btest;
1117 }
1119 // Generate real control flow
1120 float true_prob = (taken_if_true ? prob : untaken_prob);
1121 IfNode* iff = create_and_map_if(control(), tst, true_prob, cnt);
1122 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser");
1123 Node* taken_branch = new (C, 1) IfTrueNode(iff);
1124 Node* untaken_branch = new (C, 1) IfFalseNode(iff);
1125 if (!taken_if_true) { // Finish conversion to canonical form
1126 Node* tmp = taken_branch;
1127 taken_branch = untaken_branch;
1128 untaken_branch = tmp;
1129 }
1131 // Branch is taken:
1132 { PreserveJVMState pjvms(this);
1133 taken_branch = _gvn.transform(taken_branch);
1134 set_control(taken_branch);
1136 if (stopped()) {
1137 if (EliminateAutoBox) {
1138 // Mark the successor block as parsed
1139 branch_block->next_path_num();
1140 }
1141 } else {
1142 // Update method data
1143 profile_taken_branch(target_bci);
1144 adjust_map_after_if(taken_btest, c, prob, branch_block, next_block);
1145 if (!stopped()) {
1146 merge(target_bci);
1147 }
1148 }
1149 }
1151 untaken_branch = _gvn.transform(untaken_branch);
1152 set_control(untaken_branch);
1154 // Branch not taken.
1155 if (stopped()) {
1156 if (EliminateAutoBox) {
1157 // Mark the successor block as parsed
1158 next_block->next_path_num();
1159 }
1160 } else {
1161 // Update method data
1162 profile_not_taken_branch();
1163 adjust_map_after_if(untaken_btest, c, untaken_prob,
1164 next_block, branch_block);
1165 }
1166 }
1168 //----------------------------adjust_map_after_if------------------------------
1169 // Adjust the JVM state to reflect the result of taking this path.
1170 // Basically, it means inspecting the CmpNode controlling this
1171 // branch, seeing how it constrains a tested value, and then
1172 // deciding if it's worth our while to encode this constraint
1173 // as graph nodes in the current abstract interpretation map.
1174 void Parse::adjust_map_after_if(BoolTest::mask btest, Node* c, float prob,
1175 Block* path, Block* other_path) {
1176 if (stopped() || !c->is_Cmp() || btest == BoolTest::illegal)
1177 return; // nothing to do
1179 bool is_fallthrough = (path == successor_for_bci(iter().next_bci()));
1181 if (seems_never_taken(prob) && seems_stable_comparison(btest, c)) {
1182 // If this might possibly turn into an implicit null check,
1183 // and the null has never yet been seen, we need to generate
1184 // an uncommon trap, so as to recompile instead of suffering
1185 // with very slow branches. (We'll get the slow branches if
1186 // the program ever changes phase and starts seeing nulls here.)
1187 //
1188 // We do not inspect for a null constant, since a node may
1189 // optimize to 'null' later on.
1190 //
1191 // Null checks, and other tests which expect inequality,
1192 // show btest == BoolTest::eq along the non-taken branch.
1193 // On the other hand, type tests, must-be-null tests,
1194 // and other tests which expect pointer equality,
1195 // show btest == BoolTest::ne along the non-taken branch.
1196 // We prune both types of branches if they look unused.
1197 repush_if_args();
1198 // We need to mark this branch as taken so that if we recompile we will
1199 // see that it is possible. In the tiered system the interpreter doesn't
1200 // do profiling and by the time we get to the lower tier from the interpreter
1201 // the path may be cold again. Make sure it doesn't look untaken
1202 if (is_fallthrough) {
1203 profile_not_taken_branch(!ProfileInterpreter);
1204 } else {
1205 profile_taken_branch(iter().get_dest(), !ProfileInterpreter);
1206 }
1207 uncommon_trap(Deoptimization::Reason_unreached,
1208 Deoptimization::Action_reinterpret,
1209 NULL,
1210 (is_fallthrough ? "taken always" : "taken never"));
1211 return;
1212 }
1214 Node* val = c->in(1);
1215 Node* con = c->in(2);
1216 const Type* tcon = _gvn.type(con);
1217 const Type* tval = _gvn.type(val);
1218 bool have_con = tcon->singleton();
1219 if (tval->singleton()) {
1220 if (!have_con) {
1221 // Swap, so constant is in con.
1222 con = val;
1223 tcon = tval;
1224 val = c->in(2);
1225 tval = _gvn.type(val);
1226 btest = BoolTest(btest).commute();
1227 have_con = true;
1228 } else {
1229 // Do we have two constants? Then leave well enough alone.
1230 have_con = false;
1231 }
1232 }
1233 if (!have_con) // remaining adjustments need a con
1234 return;
1237 int val_in_map = map()->find_edge(val);
1238 if (val_in_map < 0) return; // replace_in_map would be useless
1239 {
1240 JVMState* jvms = this->jvms();
1241 if (!(jvms->is_loc(val_in_map) ||
1242 jvms->is_stk(val_in_map)))
1243 return; // again, it would be useless
1244 }
1246 // Check for a comparison to a constant, and "know" that the compared
1247 // value is constrained on this path.
1248 assert(tcon->singleton(), "");
1249 ConstraintCastNode* ccast = NULL;
1250 Node* cast = NULL;
1252 switch (btest) {
1253 case BoolTest::eq: // Constant test?
1254 {
1255 const Type* tboth = tcon->join(tval);
1256 if (tboth == tval) break; // Nothing to gain.
1257 if (tcon->isa_int()) {
1258 ccast = new (C, 2) CastIINode(val, tboth);
1259 } else if (tcon == TypePtr::NULL_PTR) {
1260 // Cast to null, but keep the pointer identity temporarily live.
1261 ccast = new (C, 2) CastPPNode(val, tboth);
1262 } else {
1263 const TypeF* tf = tcon->isa_float_constant();
1264 const TypeD* td = tcon->isa_double_constant();
1265 // Exclude tests vs float/double 0 as these could be
1266 // either +0 or -0. Just because you are equal to +0
1267 // doesn't mean you ARE +0!
1268 if ((!tf || tf->_f != 0.0) &&
1269 (!td || td->_d != 0.0))
1270 cast = con; // Replace non-constant val by con.
1271 }
1272 }
1273 break;
1275 case BoolTest::ne:
1276 if (tcon == TypePtr::NULL_PTR) {
1277 cast = cast_not_null(val, false);
1278 }
1279 break;
1281 default:
1282 // (At this point we could record int range types with CastII.)
1283 break;
1284 }
1286 if (ccast != NULL) {
1287 const Type* tcc = ccast->as_Type()->type();
1288 assert(tcc != tval && tcc->higher_equal(tval), "must improve");
1289 // Delay transform() call to allow recovery of pre-cast value
1290 // at the control merge.
1291 ccast->set_req(0, control());
1292 _gvn.set_type_bottom(ccast);
1293 record_for_igvn(ccast);
1294 cast = ccast;
1295 }
1297 if (cast != NULL) { // Here's the payoff.
1298 replace_in_map(val, cast);
1299 }
1300 }
1303 //------------------------------do_one_bytecode--------------------------------
1304 // Parse this bytecode, and alter the Parsers JVM->Node mapping
1305 void Parse::do_one_bytecode() {
1306 Node *a, *b, *c, *d; // Handy temps
1307 BoolTest::mask btest;
1308 int i;
1310 assert(!has_exceptions(), "bytecode entry state must be clear of throws");
1312 if (C->check_node_count(NodeLimitFudgeFactor * 5,
1313 "out of nodes parsing method")) {
1314 return;
1315 }
1317 #ifdef ASSERT
1318 // for setting breakpoints
1319 if (TraceOptoParse) {
1320 tty->print(" @");
1321 dump_bci(bci());
1322 }
1323 #endif
1325 switch (bc()) {
1326 case Bytecodes::_nop:
1327 // do nothing
1328 break;
1329 case Bytecodes::_lconst_0:
1330 push_pair(longcon(0));
1331 break;
1333 case Bytecodes::_lconst_1:
1334 push_pair(longcon(1));
1335 break;
1337 case Bytecodes::_fconst_0:
1338 push(zerocon(T_FLOAT));
1339 break;
1341 case Bytecodes::_fconst_1:
1342 push(makecon(TypeF::ONE));
1343 break;
1345 case Bytecodes::_fconst_2:
1346 push(makecon(TypeF::make(2.0f)));
1347 break;
1349 case Bytecodes::_dconst_0:
1350 push_pair(zerocon(T_DOUBLE));
1351 break;
1353 case Bytecodes::_dconst_1:
1354 push_pair(makecon(TypeD::ONE));
1355 break;
1357 case Bytecodes::_iconst_m1:push(intcon(-1)); break;
1358 case Bytecodes::_iconst_0: push(intcon( 0)); break;
1359 case Bytecodes::_iconst_1: push(intcon( 1)); break;
1360 case Bytecodes::_iconst_2: push(intcon( 2)); break;
1361 case Bytecodes::_iconst_3: push(intcon( 3)); break;
1362 case Bytecodes::_iconst_4: push(intcon( 4)); break;
1363 case Bytecodes::_iconst_5: push(intcon( 5)); break;
1364 case Bytecodes::_bipush: push(intcon(iter().get_constant_u1())); break;
1365 case Bytecodes::_sipush: push(intcon(iter().get_constant_u2())); break;
1366 case Bytecodes::_aconst_null: push(null()); break;
1367 case Bytecodes::_ldc:
1368 case Bytecodes::_ldc_w:
1369 case Bytecodes::_ldc2_w:
1370 // If the constant is unresolved, run this BC once in the interpreter.
1371 {
1372 ciConstant constant = iter().get_constant();
1373 if (constant.basic_type() == T_OBJECT &&
1374 !constant.as_object()->is_loaded()) {
1375 int index = iter().get_constant_pool_index();
1376 constantTag tag = iter().get_constant_pool_tag(index);
1377 uncommon_trap(Deoptimization::make_trap_request
1378 (Deoptimization::Reason_unloaded,
1379 Deoptimization::Action_reinterpret,
1380 index),
1381 NULL, tag.internal_name());
1382 break;
1383 }
1384 assert(constant.basic_type() != T_OBJECT || !constant.as_object()->is_klass(),
1385 "must be java_mirror of klass");
1386 bool pushed = push_constant(constant, true);
1387 guarantee(pushed, "must be possible to push this constant");
1388 }
1390 break;
1392 case Bytecodes::_aload_0:
1393 push( local(0) );
1394 break;
1395 case Bytecodes::_aload_1:
1396 push( local(1) );
1397 break;
1398 case Bytecodes::_aload_2:
1399 push( local(2) );
1400 break;
1401 case Bytecodes::_aload_3:
1402 push( local(3) );
1403 break;
1404 case Bytecodes::_aload:
1405 push( local(iter().get_index()) );
1406 break;
1408 case Bytecodes::_fload_0:
1409 case Bytecodes::_iload_0:
1410 push( local(0) );
1411 break;
1412 case Bytecodes::_fload_1:
1413 case Bytecodes::_iload_1:
1414 push( local(1) );
1415 break;
1416 case Bytecodes::_fload_2:
1417 case Bytecodes::_iload_2:
1418 push( local(2) );
1419 break;
1420 case Bytecodes::_fload_3:
1421 case Bytecodes::_iload_3:
1422 push( local(3) );
1423 break;
1424 case Bytecodes::_fload:
1425 case Bytecodes::_iload:
1426 push( local(iter().get_index()) );
1427 break;
1428 case Bytecodes::_lload_0:
1429 push_pair_local( 0 );
1430 break;
1431 case Bytecodes::_lload_1:
1432 push_pair_local( 1 );
1433 break;
1434 case Bytecodes::_lload_2:
1435 push_pair_local( 2 );
1436 break;
1437 case Bytecodes::_lload_3:
1438 push_pair_local( 3 );
1439 break;
1440 case Bytecodes::_lload:
1441 push_pair_local( iter().get_index() );
1442 break;
1444 case Bytecodes::_dload_0:
1445 push_pair_local(0);
1446 break;
1447 case Bytecodes::_dload_1:
1448 push_pair_local(1);
1449 break;
1450 case Bytecodes::_dload_2:
1451 push_pair_local(2);
1452 break;
1453 case Bytecodes::_dload_3:
1454 push_pair_local(3);
1455 break;
1456 case Bytecodes::_dload:
1457 push_pair_local(iter().get_index());
1458 break;
1459 case Bytecodes::_fstore_0:
1460 case Bytecodes::_istore_0:
1461 case Bytecodes::_astore_0:
1462 set_local( 0, pop() );
1463 break;
1464 case Bytecodes::_fstore_1:
1465 case Bytecodes::_istore_1:
1466 case Bytecodes::_astore_1:
1467 set_local( 1, pop() );
1468 break;
1469 case Bytecodes::_fstore_2:
1470 case Bytecodes::_istore_2:
1471 case Bytecodes::_astore_2:
1472 set_local( 2, pop() );
1473 break;
1474 case Bytecodes::_fstore_3:
1475 case Bytecodes::_istore_3:
1476 case Bytecodes::_astore_3:
1477 set_local( 3, pop() );
1478 break;
1479 case Bytecodes::_fstore:
1480 case Bytecodes::_istore:
1481 case Bytecodes::_astore:
1482 set_local( iter().get_index(), pop() );
1483 break;
1484 // long stores
1485 case Bytecodes::_lstore_0:
1486 set_pair_local( 0, pop_pair() );
1487 break;
1488 case Bytecodes::_lstore_1:
1489 set_pair_local( 1, pop_pair() );
1490 break;
1491 case Bytecodes::_lstore_2:
1492 set_pair_local( 2, pop_pair() );
1493 break;
1494 case Bytecodes::_lstore_3:
1495 set_pair_local( 3, pop_pair() );
1496 break;
1497 case Bytecodes::_lstore:
1498 set_pair_local( iter().get_index(), pop_pair() );
1499 break;
1501 // double stores
1502 case Bytecodes::_dstore_0:
1503 set_pair_local( 0, dstore_rounding(pop_pair()) );
1504 break;
1505 case Bytecodes::_dstore_1:
1506 set_pair_local( 1, dstore_rounding(pop_pair()) );
1507 break;
1508 case Bytecodes::_dstore_2:
1509 set_pair_local( 2, dstore_rounding(pop_pair()) );
1510 break;
1511 case Bytecodes::_dstore_3:
1512 set_pair_local( 3, dstore_rounding(pop_pair()) );
1513 break;
1514 case Bytecodes::_dstore:
1515 set_pair_local( iter().get_index(), dstore_rounding(pop_pair()) );
1516 break;
1518 case Bytecodes::_pop: _sp -= 1; break;
1519 case Bytecodes::_pop2: _sp -= 2; break;
1520 case Bytecodes::_swap:
1521 a = pop();
1522 b = pop();
1523 push(a);
1524 push(b);
1525 break;
1526 case Bytecodes::_dup:
1527 a = pop();
1528 push(a);
1529 push(a);
1530 break;
1531 case Bytecodes::_dup_x1:
1532 a = pop();
1533 b = pop();
1534 push( a );
1535 push( b );
1536 push( a );
1537 break;
1538 case Bytecodes::_dup_x2:
1539 a = pop();
1540 b = pop();
1541 c = pop();
1542 push( a );
1543 push( c );
1544 push( b );
1545 push( a );
1546 break;
1547 case Bytecodes::_dup2:
1548 a = pop();
1549 b = pop();
1550 push( b );
1551 push( a );
1552 push( b );
1553 push( a );
1554 break;
1556 case Bytecodes::_dup2_x1:
1557 // before: .. c, b, a
1558 // after: .. b, a, c, b, a
1559 // not tested
1560 a = pop();
1561 b = pop();
1562 c = pop();
1563 push( b );
1564 push( a );
1565 push( c );
1566 push( b );
1567 push( a );
1568 break;
1569 case Bytecodes::_dup2_x2:
1570 // before: .. d, c, b, a
1571 // after: .. b, a, d, c, b, a
1572 // not tested
1573 a = pop();
1574 b = pop();
1575 c = pop();
1576 d = pop();
1577 push( b );
1578 push( a );
1579 push( d );
1580 push( c );
1581 push( b );
1582 push( a );
1583 break;
1585 case Bytecodes::_arraylength: {
1586 // Must do null-check with value on expression stack
1587 Node *ary = do_null_check(peek(), T_ARRAY);
1588 // Compile-time detect of null-exception?
1589 if (stopped()) return;
1590 a = pop();
1591 push(load_array_length(a));
1592 break;
1593 }
1595 case Bytecodes::_baload: array_load(T_BYTE); break;
1596 case Bytecodes::_caload: array_load(T_CHAR); break;
1597 case Bytecodes::_iaload: array_load(T_INT); break;
1598 case Bytecodes::_saload: array_load(T_SHORT); break;
1599 case Bytecodes::_faload: array_load(T_FLOAT); break;
1600 case Bytecodes::_aaload: array_load(T_OBJECT); break;
1601 case Bytecodes::_laload: {
1602 a = array_addressing(T_LONG, 0);
1603 if (stopped()) return; // guaranteed null or range check
1604 _sp -= 2; // Pop array and index
1605 push_pair( make_load(control(), a, TypeLong::LONG, T_LONG, TypeAryPtr::LONGS));
1606 break;
1607 }
1608 case Bytecodes::_daload: {
1609 a = array_addressing(T_DOUBLE, 0);
1610 if (stopped()) return; // guaranteed null or range check
1611 _sp -= 2; // Pop array and index
1612 push_pair( make_load(control(), a, Type::DOUBLE, T_DOUBLE, TypeAryPtr::DOUBLES));
1613 break;
1614 }
1615 case Bytecodes::_bastore: array_store(T_BYTE); break;
1616 case Bytecodes::_castore: array_store(T_CHAR); break;
1617 case Bytecodes::_iastore: array_store(T_INT); break;
1618 case Bytecodes::_sastore: array_store(T_SHORT); break;
1619 case Bytecodes::_fastore: array_store(T_FLOAT); break;
1620 case Bytecodes::_aastore: {
1621 d = array_addressing(T_OBJECT, 1);
1622 if (stopped()) return; // guaranteed null or range check
1623 array_store_check();
1624 c = pop(); // Oop to store
1625 b = pop(); // index (already used)
1626 a = pop(); // the array itself
1627 const TypeOopPtr* elemtype = _gvn.type(a)->is_aryptr()->elem()->make_oopptr();
1628 const TypeAryPtr* adr_type = TypeAryPtr::OOPS;
1629 Node* store = store_oop_to_array(control(), a, d, adr_type, c, elemtype, T_OBJECT);
1630 break;
1631 }
1632 case Bytecodes::_lastore: {
1633 a = array_addressing(T_LONG, 2);
1634 if (stopped()) return; // guaranteed null or range check
1635 c = pop_pair();
1636 _sp -= 2; // Pop array and index
1637 store_to_memory(control(), a, c, T_LONG, TypeAryPtr::LONGS);
1638 break;
1639 }
1640 case Bytecodes::_dastore: {
1641 a = array_addressing(T_DOUBLE, 2);
1642 if (stopped()) return; // guaranteed null or range check
1643 c = pop_pair();
1644 _sp -= 2; // Pop array and index
1645 c = dstore_rounding(c);
1646 store_to_memory(control(), a, c, T_DOUBLE, TypeAryPtr::DOUBLES);
1647 break;
1648 }
1649 case Bytecodes::_getfield:
1650 do_getfield();
1651 break;
1653 case Bytecodes::_getstatic:
1654 do_getstatic();
1655 break;
1657 case Bytecodes::_putfield:
1658 do_putfield();
1659 break;
1661 case Bytecodes::_putstatic:
1662 do_putstatic();
1663 break;
1665 case Bytecodes::_irem:
1666 do_irem();
1667 break;
1668 case Bytecodes::_idiv:
1669 // Must keep both values on the expression-stack during null-check
1670 do_null_check(peek(), T_INT);
1671 // Compile-time detect of null-exception?
1672 if (stopped()) return;
1673 b = pop();
1674 a = pop();
1675 push( _gvn.transform( new (C, 3) DivINode(control(),a,b) ) );
1676 break;
1677 case Bytecodes::_imul:
1678 b = pop(); a = pop();
1679 push( _gvn.transform( new (C, 3) MulINode(a,b) ) );
1680 break;
1681 case Bytecodes::_iadd:
1682 b = pop(); a = pop();
1683 push( _gvn.transform( new (C, 3) AddINode(a,b) ) );
1684 break;
1685 case Bytecodes::_ineg:
1686 a = pop();
1687 push( _gvn.transform( new (C, 3) SubINode(_gvn.intcon(0),a)) );
1688 break;
1689 case Bytecodes::_isub:
1690 b = pop(); a = pop();
1691 push( _gvn.transform( new (C, 3) SubINode(a,b) ) );
1692 break;
1693 case Bytecodes::_iand:
1694 b = pop(); a = pop();
1695 push( _gvn.transform( new (C, 3) AndINode(a,b) ) );
1696 break;
1697 case Bytecodes::_ior:
1698 b = pop(); a = pop();
1699 push( _gvn.transform( new (C, 3) OrINode(a,b) ) );
1700 break;
1701 case Bytecodes::_ixor:
1702 b = pop(); a = pop();
1703 push( _gvn.transform( new (C, 3) XorINode(a,b) ) );
1704 break;
1705 case Bytecodes::_ishl:
1706 b = pop(); a = pop();
1707 push( _gvn.transform( new (C, 3) LShiftINode(a,b) ) );
1708 break;
1709 case Bytecodes::_ishr:
1710 b = pop(); a = pop();
1711 push( _gvn.transform( new (C, 3) RShiftINode(a,b) ) );
1712 break;
1713 case Bytecodes::_iushr:
1714 b = pop(); a = pop();
1715 push( _gvn.transform( new (C, 3) URShiftINode(a,b) ) );
1716 break;
1718 case Bytecodes::_fneg:
1719 a = pop();
1720 b = _gvn.transform(new (C, 2) NegFNode (a));
1721 push(b);
1722 break;
1724 case Bytecodes::_fsub:
1725 b = pop();
1726 a = pop();
1727 c = _gvn.transform( new (C, 3) SubFNode(a,b) );
1728 d = precision_rounding(c);
1729 push( d );
1730 break;
1732 case Bytecodes::_fadd:
1733 b = pop();
1734 a = pop();
1735 c = _gvn.transform( new (C, 3) AddFNode(a,b) );
1736 d = precision_rounding(c);
1737 push( d );
1738 break;
1740 case Bytecodes::_fmul:
1741 b = pop();
1742 a = pop();
1743 c = _gvn.transform( new (C, 3) MulFNode(a,b) );
1744 d = precision_rounding(c);
1745 push( d );
1746 break;
1748 case Bytecodes::_fdiv:
1749 b = pop();
1750 a = pop();
1751 c = _gvn.transform( new (C, 3) DivFNode(0,a,b) );
1752 d = precision_rounding(c);
1753 push( d );
1754 break;
1756 case Bytecodes::_frem:
1757 if (Matcher::has_match_rule(Op_ModF)) {
1758 // Generate a ModF node.
1759 b = pop();
1760 a = pop();
1761 c = _gvn.transform( new (C, 3) ModFNode(0,a,b) );
1762 d = precision_rounding(c);
1763 push( d );
1764 }
1765 else {
1766 // Generate a call.
1767 modf();
1768 }
1769 break;
1771 case Bytecodes::_fcmpl:
1772 b = pop();
1773 a = pop();
1774 c = _gvn.transform( new (C, 3) CmpF3Node( a, b));
1775 push(c);
1776 break;
1777 case Bytecodes::_fcmpg:
1778 b = pop();
1779 a = pop();
1781 // Same as fcmpl but need to flip the unordered case. Swap the inputs,
1782 // which negates the result sign except for unordered. Flip the unordered
1783 // as well by using CmpF3 which implements unordered-lesser instead of
1784 // unordered-greater semantics. Finally, commute the result bits. Result
1785 // is same as using a CmpF3Greater except we did it with CmpF3 alone.
1786 c = _gvn.transform( new (C, 3) CmpF3Node( b, a));
1787 c = _gvn.transform( new (C, 3) SubINode(_gvn.intcon(0),c) );
1788 push(c);
1789 break;
1791 case Bytecodes::_f2i:
1792 a = pop();
1793 push(_gvn.transform(new (C, 2) ConvF2INode(a)));
1794 break;
1796 case Bytecodes::_d2i:
1797 a = pop_pair();
1798 b = _gvn.transform(new (C, 2) ConvD2INode(a));
1799 push( b );
1800 break;
1802 case Bytecodes::_f2d:
1803 a = pop();
1804 b = _gvn.transform( new (C, 2) ConvF2DNode(a));
1805 push_pair( b );
1806 break;
1808 case Bytecodes::_d2f:
1809 a = pop_pair();
1810 b = _gvn.transform( new (C, 2) ConvD2FNode(a));
1811 // This breaks _227_mtrt (speed & correctness) and _222_mpegaudio (speed)
1812 //b = _gvn.transform(new (C, 2) RoundFloatNode(0, b) );
1813 push( b );
1814 break;
1816 case Bytecodes::_l2f:
1817 if (Matcher::convL2FSupported()) {
1818 a = pop_pair();
1819 b = _gvn.transform( new (C, 2) ConvL2FNode(a));
1820 // For i486.ad, FILD doesn't restrict precision to 24 or 53 bits.
1821 // Rather than storing the result into an FP register then pushing
1822 // out to memory to round, the machine instruction that implements
1823 // ConvL2D is responsible for rounding.
1824 // c = precision_rounding(b);
1825 c = _gvn.transform(b);
1826 push(c);
1827 } else {
1828 l2f();
1829 }
1830 break;
1832 case Bytecodes::_l2d:
1833 a = pop_pair();
1834 b = _gvn.transform( new (C, 2) ConvL2DNode(a));
1835 // For i486.ad, rounding is always necessary (see _l2f above).
1836 // c = dprecision_rounding(b);
1837 c = _gvn.transform(b);
1838 push_pair(c);
1839 break;
1841 case Bytecodes::_f2l:
1842 a = pop();
1843 b = _gvn.transform( new (C, 2) ConvF2LNode(a));
1844 push_pair(b);
1845 break;
1847 case Bytecodes::_d2l:
1848 a = pop_pair();
1849 b = _gvn.transform( new (C, 2) ConvD2LNode(a));
1850 push_pair(b);
1851 break;
1853 case Bytecodes::_dsub:
1854 b = pop_pair();
1855 a = pop_pair();
1856 c = _gvn.transform( new (C, 3) SubDNode(a,b) );
1857 d = dprecision_rounding(c);
1858 push_pair( d );
1859 break;
1861 case Bytecodes::_dadd:
1862 b = pop_pair();
1863 a = pop_pair();
1864 c = _gvn.transform( new (C, 3) AddDNode(a,b) );
1865 d = dprecision_rounding(c);
1866 push_pair( d );
1867 break;
1869 case Bytecodes::_dmul:
1870 b = pop_pair();
1871 a = pop_pair();
1872 c = _gvn.transform( new (C, 3) MulDNode(a,b) );
1873 d = dprecision_rounding(c);
1874 push_pair( d );
1875 break;
1877 case Bytecodes::_ddiv:
1878 b = pop_pair();
1879 a = pop_pair();
1880 c = _gvn.transform( new (C, 3) DivDNode(0,a,b) );
1881 d = dprecision_rounding(c);
1882 push_pair( d );
1883 break;
1885 case Bytecodes::_dneg:
1886 a = pop_pair();
1887 b = _gvn.transform(new (C, 2) NegDNode (a));
1888 push_pair(b);
1889 break;
1891 case Bytecodes::_drem:
1892 if (Matcher::has_match_rule(Op_ModD)) {
1893 // Generate a ModD node.
1894 b = pop_pair();
1895 a = pop_pair();
1896 // a % b
1898 c = _gvn.transform( new (C, 3) ModDNode(0,a,b) );
1899 d = dprecision_rounding(c);
1900 push_pair( d );
1901 }
1902 else {
1903 // Generate a call.
1904 modd();
1905 }
1906 break;
1908 case Bytecodes::_dcmpl:
1909 b = pop_pair();
1910 a = pop_pair();
1911 c = _gvn.transform( new (C, 3) CmpD3Node( a, b));
1912 push(c);
1913 break;
1915 case Bytecodes::_dcmpg:
1916 b = pop_pair();
1917 a = pop_pair();
1918 // Same as dcmpl but need to flip the unordered case.
1919 // Commute the inputs, which negates the result sign except for unordered.
1920 // Flip the unordered as well by using CmpD3 which implements
1921 // unordered-lesser instead of unordered-greater semantics.
1922 // Finally, negate the result bits. Result is same as using a
1923 // CmpD3Greater except we did it with CmpD3 alone.
1924 c = _gvn.transform( new (C, 3) CmpD3Node( b, a));
1925 c = _gvn.transform( new (C, 3) SubINode(_gvn.intcon(0),c) );
1926 push(c);
1927 break;
1930 // Note for longs -> lo word is on TOS, hi word is on TOS - 1
1931 case Bytecodes::_land:
1932 b = pop_pair();
1933 a = pop_pair();
1934 c = _gvn.transform( new (C, 3) AndLNode(a,b) );
1935 push_pair(c);
1936 break;
1937 case Bytecodes::_lor:
1938 b = pop_pair();
1939 a = pop_pair();
1940 c = _gvn.transform( new (C, 3) OrLNode(a,b) );
1941 push_pair(c);
1942 break;
1943 case Bytecodes::_lxor:
1944 b = pop_pair();
1945 a = pop_pair();
1946 c = _gvn.transform( new (C, 3) XorLNode(a,b) );
1947 push_pair(c);
1948 break;
1950 case Bytecodes::_lshl:
1951 b = pop(); // the shift count
1952 a = pop_pair(); // value to be shifted
1953 c = _gvn.transform( new (C, 3) LShiftLNode(a,b) );
1954 push_pair(c);
1955 break;
1956 case Bytecodes::_lshr:
1957 b = pop(); // the shift count
1958 a = pop_pair(); // value to be shifted
1959 c = _gvn.transform( new (C, 3) RShiftLNode(a,b) );
1960 push_pair(c);
1961 break;
1962 case Bytecodes::_lushr:
1963 b = pop(); // the shift count
1964 a = pop_pair(); // value to be shifted
1965 c = _gvn.transform( new (C, 3) URShiftLNode(a,b) );
1966 push_pair(c);
1967 break;
1968 case Bytecodes::_lmul:
1969 b = pop_pair();
1970 a = pop_pair();
1971 c = _gvn.transform( new (C, 3) MulLNode(a,b) );
1972 push_pair(c);
1973 break;
1975 case Bytecodes::_lrem:
1976 // Must keep both values on the expression-stack during null-check
1977 assert(peek(0) == top(), "long word order");
1978 do_null_check(peek(1), T_LONG);
1979 // Compile-time detect of null-exception?
1980 if (stopped()) return;
1981 b = pop_pair();
1982 a = pop_pair();
1983 c = _gvn.transform( new (C, 3) ModLNode(control(),a,b) );
1984 push_pair(c);
1985 break;
1987 case Bytecodes::_ldiv:
1988 // Must keep both values on the expression-stack during null-check
1989 assert(peek(0) == top(), "long word order");
1990 do_null_check(peek(1), T_LONG);
1991 // Compile-time detect of null-exception?
1992 if (stopped()) return;
1993 b = pop_pair();
1994 a = pop_pair();
1995 c = _gvn.transform( new (C, 3) DivLNode(control(),a,b) );
1996 push_pair(c);
1997 break;
1999 case Bytecodes::_ladd:
2000 b = pop_pair();
2001 a = pop_pair();
2002 c = _gvn.transform( new (C, 3) AddLNode(a,b) );
2003 push_pair(c);
2004 break;
2005 case Bytecodes::_lsub:
2006 b = pop_pair();
2007 a = pop_pair();
2008 c = _gvn.transform( new (C, 3) SubLNode(a,b) );
2009 push_pair(c);
2010 break;
2011 case Bytecodes::_lcmp:
2012 // Safepoints are now inserted _before_ branches. The long-compare
2013 // bytecode painfully produces a 3-way value (-1,0,+1) which requires a
2014 // slew of control flow. These are usually followed by a CmpI vs zero and
2015 // a branch; this pattern then optimizes to the obvious long-compare and
2016 // branch. However, if the branch is backwards there's a Safepoint
2017 // inserted. The inserted Safepoint captures the JVM state at the
2018 // pre-branch point, i.e. it captures the 3-way value. Thus if a
2019 // long-compare is used to control a loop the debug info will force
2020 // computation of the 3-way value, even though the generated code uses a
2021 // long-compare and branch. We try to rectify the situation by inserting
2022 // a SafePoint here and have it dominate and kill the safepoint added at a
2023 // following backwards branch. At this point the JVM state merely holds 2
2024 // longs but not the 3-way value.
2025 if( UseLoopSafepoints ) {
2026 switch( iter().next_bc() ) {
2027 case Bytecodes::_ifgt:
2028 case Bytecodes::_iflt:
2029 case Bytecodes::_ifge:
2030 case Bytecodes::_ifle:
2031 case Bytecodes::_ifne:
2032 case Bytecodes::_ifeq:
2033 // If this is a backwards branch in the bytecodes, add Safepoint
2034 maybe_add_safepoint(iter().next_get_dest());
2035 }
2036 }
2037 b = pop_pair();
2038 a = pop_pair();
2039 c = _gvn.transform( new (C, 3) CmpL3Node( a, b ));
2040 push(c);
2041 break;
2043 case Bytecodes::_lneg:
2044 a = pop_pair();
2045 b = _gvn.transform( new (C, 3) SubLNode(longcon(0),a));
2046 push_pair(b);
2047 break;
2048 case Bytecodes::_l2i:
2049 a = pop_pair();
2050 push( _gvn.transform( new (C, 2) ConvL2INode(a)));
2051 break;
2052 case Bytecodes::_i2l:
2053 a = pop();
2054 b = _gvn.transform( new (C, 2) ConvI2LNode(a));
2055 push_pair(b);
2056 break;
2057 case Bytecodes::_i2b:
2058 // Sign extend
2059 a = pop();
2060 a = _gvn.transform( new (C, 3) LShiftINode(a,_gvn.intcon(24)) );
2061 a = _gvn.transform( new (C, 3) RShiftINode(a,_gvn.intcon(24)) );
2062 push( a );
2063 break;
2064 case Bytecodes::_i2s:
2065 a = pop();
2066 a = _gvn.transform( new (C, 3) LShiftINode(a,_gvn.intcon(16)) );
2067 a = _gvn.transform( new (C, 3) RShiftINode(a,_gvn.intcon(16)) );
2068 push( a );
2069 break;
2070 case Bytecodes::_i2c:
2071 a = pop();
2072 push( _gvn.transform( new (C, 3) AndINode(a,_gvn.intcon(0xFFFF)) ) );
2073 break;
2075 case Bytecodes::_i2f:
2076 a = pop();
2077 b = _gvn.transform( new (C, 2) ConvI2FNode(a) ) ;
2078 c = precision_rounding(b);
2079 push (b);
2080 break;
2082 case Bytecodes::_i2d:
2083 a = pop();
2084 b = _gvn.transform( new (C, 2) ConvI2DNode(a));
2085 push_pair(b);
2086 break;
2088 case Bytecodes::_iinc: // Increment local
2089 i = iter().get_index(); // Get local index
2090 set_local( i, _gvn.transform( new (C, 3) AddINode( _gvn.intcon(iter().get_iinc_con()), local(i) ) ) );
2091 break;
2093 // Exit points of synchronized methods must have an unlock node
2094 case Bytecodes::_return:
2095 return_current(NULL);
2096 break;
2098 case Bytecodes::_ireturn:
2099 case Bytecodes::_areturn:
2100 case Bytecodes::_freturn:
2101 return_current(pop());
2102 break;
2103 case Bytecodes::_lreturn:
2104 return_current(pop_pair());
2105 break;
2106 case Bytecodes::_dreturn:
2107 return_current(pop_pair());
2108 break;
2110 case Bytecodes::_athrow:
2111 // null exception oop throws NULL pointer exception
2112 do_null_check(peek(), T_OBJECT);
2113 if (stopped()) return;
2114 // Hook the thrown exception directly to subsequent handlers.
2115 if (BailoutToInterpreterForThrows) {
2116 // Keep method interpreted from now on.
2117 uncommon_trap(Deoptimization::Reason_unhandled,
2118 Deoptimization::Action_make_not_compilable);
2119 return;
2120 }
2121 if (env()->jvmti_can_post_on_exceptions()) {
2122 // check if we must post exception events, take uncommon trap if so (with must_throw = false)
2123 uncommon_trap_if_should_post_on_exceptions(Deoptimization::Reason_unhandled, false);
2124 }
2125 // Here if either can_post_on_exceptions or should_post_on_exceptions is false
2126 add_exception_state(make_exception_state(peek()));
2127 break;
2129 case Bytecodes::_goto: // fall through
2130 case Bytecodes::_goto_w: {
2131 int target_bci = (bc() == Bytecodes::_goto) ? iter().get_dest() : iter().get_far_dest();
2133 // If this is a backwards branch in the bytecodes, add Safepoint
2134 maybe_add_safepoint(target_bci);
2136 // Update method data
2137 profile_taken_branch(target_bci);
2139 // Merge the current control into the target basic block
2140 merge(target_bci);
2142 // See if we can get some profile data and hand it off to the next block
2143 Block *target_block = block()->successor_for_bci(target_bci);
2144 if (target_block->pred_count() != 1) break;
2145 ciMethodData* methodData = method()->method_data();
2146 if (!methodData->is_mature()) break;
2147 ciProfileData* data = methodData->bci_to_data(bci());
2148 assert( data->is_JumpData(), "" );
2149 int taken = ((ciJumpData*)data)->taken();
2150 taken = method()->scale_count(taken);
2151 target_block->set_count(taken);
2152 break;
2153 }
2155 case Bytecodes::_ifnull: btest = BoolTest::eq; goto handle_if_null;
2156 case Bytecodes::_ifnonnull: btest = BoolTest::ne; goto handle_if_null;
2157 handle_if_null:
2158 // If this is a backwards branch in the bytecodes, add Safepoint
2159 maybe_add_safepoint(iter().get_dest());
2160 a = null();
2161 b = pop();
2162 c = _gvn.transform( new (C, 3) CmpPNode(b, a) );
2163 do_ifnull(btest, c);
2164 break;
2166 case Bytecodes::_if_acmpeq: btest = BoolTest::eq; goto handle_if_acmp;
2167 case Bytecodes::_if_acmpne: btest = BoolTest::ne; goto handle_if_acmp;
2168 handle_if_acmp:
2169 // If this is a backwards branch in the bytecodes, add Safepoint
2170 maybe_add_safepoint(iter().get_dest());
2171 a = pop();
2172 b = pop();
2173 c = _gvn.transform( new (C, 3) CmpPNode(b, a) );
2174 do_if(btest, c);
2175 break;
2177 case Bytecodes::_ifeq: btest = BoolTest::eq; goto handle_ifxx;
2178 case Bytecodes::_ifne: btest = BoolTest::ne; goto handle_ifxx;
2179 case Bytecodes::_iflt: btest = BoolTest::lt; goto handle_ifxx;
2180 case Bytecodes::_ifle: btest = BoolTest::le; goto handle_ifxx;
2181 case Bytecodes::_ifgt: btest = BoolTest::gt; goto handle_ifxx;
2182 case Bytecodes::_ifge: btest = BoolTest::ge; goto handle_ifxx;
2183 handle_ifxx:
2184 // If this is a backwards branch in the bytecodes, add Safepoint
2185 maybe_add_safepoint(iter().get_dest());
2186 a = _gvn.intcon(0);
2187 b = pop();
2188 c = _gvn.transform( new (C, 3) CmpINode(b, a) );
2189 do_if(btest, c);
2190 break;
2192 case Bytecodes::_if_icmpeq: btest = BoolTest::eq; goto handle_if_icmp;
2193 case Bytecodes::_if_icmpne: btest = BoolTest::ne; goto handle_if_icmp;
2194 case Bytecodes::_if_icmplt: btest = BoolTest::lt; goto handle_if_icmp;
2195 case Bytecodes::_if_icmple: btest = BoolTest::le; goto handle_if_icmp;
2196 case Bytecodes::_if_icmpgt: btest = BoolTest::gt; goto handle_if_icmp;
2197 case Bytecodes::_if_icmpge: btest = BoolTest::ge; goto handle_if_icmp;
2198 handle_if_icmp:
2199 // If this is a backwards branch in the bytecodes, add Safepoint
2200 maybe_add_safepoint(iter().get_dest());
2201 a = pop();
2202 b = pop();
2203 c = _gvn.transform( new (C, 3) CmpINode( b, a ) );
2204 do_if(btest, c);
2205 break;
2207 case Bytecodes::_tableswitch:
2208 do_tableswitch();
2209 break;
2211 case Bytecodes::_lookupswitch:
2212 do_lookupswitch();
2213 break;
2215 case Bytecodes::_invokestatic:
2216 case Bytecodes::_invokedynamic:
2217 case Bytecodes::_invokespecial:
2218 case Bytecodes::_invokevirtual:
2219 case Bytecodes::_invokeinterface:
2220 do_call();
2221 break;
2222 case Bytecodes::_checkcast:
2223 do_checkcast();
2224 break;
2225 case Bytecodes::_instanceof:
2226 do_instanceof();
2227 break;
2228 case Bytecodes::_anewarray:
2229 do_anewarray();
2230 break;
2231 case Bytecodes::_newarray:
2232 do_newarray((BasicType)iter().get_index());
2233 break;
2234 case Bytecodes::_multianewarray:
2235 do_multianewarray();
2236 break;
2237 case Bytecodes::_new:
2238 do_new();
2239 break;
2241 case Bytecodes::_jsr:
2242 case Bytecodes::_jsr_w:
2243 do_jsr();
2244 break;
2246 case Bytecodes::_ret:
2247 do_ret();
2248 break;
2251 case Bytecodes::_monitorenter:
2252 do_monitor_enter();
2253 break;
2255 case Bytecodes::_monitorexit:
2256 do_monitor_exit();
2257 break;
2259 case Bytecodes::_breakpoint:
2260 // Breakpoint set concurrently to compile
2261 // %%% use an uncommon trap?
2262 C->record_failure("breakpoint in method");
2263 return;
2265 default:
2266 #ifndef PRODUCT
2267 map()->dump(99);
2268 #endif
2269 tty->print("\nUnhandled bytecode %s\n", Bytecodes::name(bc()) );
2270 ShouldNotReachHere();
2271 }
2273 #ifndef PRODUCT
2274 IdealGraphPrinter *printer = IdealGraphPrinter::printer();
2275 if(printer) {
2276 char buffer[256];
2277 sprintf(buffer, "Bytecode %d: %s", bci(), Bytecodes::name(bc()));
2278 bool old = printer->traverse_outs();
2279 printer->set_traverse_outs(true);
2280 printer->print_method(C, buffer, 4);
2281 printer->set_traverse_outs(old);
2282 }
2283 #endif
2284 }