Fri, 04 Mar 2016 16:15:48 +0300
8130150: Implement BigInteger.montgomeryMultiply intrinsic
Reviewed-by: kvn, mdoerr
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 // Make array address computation control dependent to prevent it
158 // from floating above the range check during loop optimizations.
159 Node* ptr = array_element_address(ary, idx, type, sizetype, control());
161 if (result2 != NULL) *result2 = elemtype;
163 assert(ptr != top(), "top should go hand-in-hand with stopped");
165 return ptr;
166 }
169 // returns IfNode
170 IfNode* Parse::jump_if_fork_int(Node* a, Node* b, BoolTest::mask mask) {
171 Node *cmp = _gvn.transform( new (C) CmpINode( a, b)); // two cases: shiftcount > 32 and shiftcount <= 32
172 Node *tst = _gvn.transform( new (C) BoolNode( cmp, mask));
173 IfNode *iff = create_and_map_if( control(), tst, ((mask == BoolTest::eq) ? PROB_STATIC_INFREQUENT : PROB_FAIR), COUNT_UNKNOWN );
174 return iff;
175 }
177 // return Region node
178 Node* Parse::jump_if_join(Node* iffalse, Node* iftrue) {
179 Node *region = new (C) RegionNode(3); // 2 results
180 record_for_igvn(region);
181 region->init_req(1, iffalse);
182 region->init_req(2, iftrue );
183 _gvn.set_type(region, Type::CONTROL);
184 region = _gvn.transform(region);
185 set_control (region);
186 return region;
187 }
190 //------------------------------helper for tableswitch-------------------------
191 void Parse::jump_if_true_fork(IfNode *iff, int dest_bci_if_true, int prof_table_index) {
192 // True branch, use existing map info
193 { PreserveJVMState pjvms(this);
194 Node *iftrue = _gvn.transform( new (C) IfTrueNode (iff) );
195 set_control( iftrue );
196 profile_switch_case(prof_table_index);
197 merge_new_path(dest_bci_if_true);
198 }
200 // False branch
201 Node *iffalse = _gvn.transform( new (C) IfFalseNode(iff) );
202 set_control( iffalse );
203 }
205 void Parse::jump_if_false_fork(IfNode *iff, int dest_bci_if_true, int prof_table_index) {
206 // True branch, use existing map info
207 { PreserveJVMState pjvms(this);
208 Node *iffalse = _gvn.transform( new (C) IfFalseNode (iff) );
209 set_control( iffalse );
210 profile_switch_case(prof_table_index);
211 merge_new_path(dest_bci_if_true);
212 }
214 // False branch
215 Node *iftrue = _gvn.transform( new (C) IfTrueNode(iff) );
216 set_control( iftrue );
217 }
219 void Parse::jump_if_always_fork(int dest_bci, int prof_table_index) {
220 // False branch, use existing map and control()
221 profile_switch_case(prof_table_index);
222 merge_new_path(dest_bci);
223 }
226 extern "C" {
227 static int jint_cmp(const void *i, const void *j) {
228 int a = *(jint *)i;
229 int b = *(jint *)j;
230 return a > b ? 1 : a < b ? -1 : 0;
231 }
232 }
235 // Default value for methodData switch indexing. Must be a negative value to avoid
236 // conflict with any legal switch index.
237 #define NullTableIndex -1
239 class SwitchRange : public StackObj {
240 // a range of integers coupled with a bci destination
241 jint _lo; // inclusive lower limit
242 jint _hi; // inclusive upper limit
243 int _dest;
244 int _table_index; // index into method data table
246 public:
247 jint lo() const { return _lo; }
248 jint hi() const { return _hi; }
249 int dest() const { return _dest; }
250 int table_index() const { return _table_index; }
251 bool is_singleton() const { return _lo == _hi; }
253 void setRange(jint lo, jint hi, int dest, int table_index) {
254 assert(lo <= hi, "must be a non-empty range");
255 _lo = lo, _hi = hi; _dest = dest; _table_index = table_index;
256 }
257 bool adjoinRange(jint lo, jint hi, int dest, int table_index) {
258 assert(lo <= hi, "must be a non-empty range");
259 if (lo == _hi+1 && dest == _dest && table_index == _table_index) {
260 _hi = hi;
261 return true;
262 }
263 return false;
264 }
266 void set (jint value, int dest, int table_index) {
267 setRange(value, value, dest, table_index);
268 }
269 bool adjoin(jint value, int dest, int table_index) {
270 return adjoinRange(value, value, dest, table_index);
271 }
273 void print() {
274 if (is_singleton())
275 tty->print(" {%d}=>%d", lo(), dest());
276 else if (lo() == min_jint)
277 tty->print(" {..%d}=>%d", hi(), dest());
278 else if (hi() == max_jint)
279 tty->print(" {%d..}=>%d", lo(), dest());
280 else
281 tty->print(" {%d..%d}=>%d", lo(), hi(), dest());
282 }
283 };
286 //-------------------------------do_tableswitch--------------------------------
287 void Parse::do_tableswitch() {
288 Node* lookup = pop();
290 // Get information about tableswitch
291 int default_dest = iter().get_dest_table(0);
292 int lo_index = iter().get_int_table(1);
293 int hi_index = iter().get_int_table(2);
294 int len = hi_index - lo_index + 1;
296 if (len < 1) {
297 // If this is a backward branch, add safepoint
298 maybe_add_safepoint(default_dest);
299 merge(default_dest);
300 return;
301 }
303 // generate decision tree, using trichotomy when possible
304 int rnum = len+2;
305 bool makes_backward_branch = false;
306 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum);
307 int rp = -1;
308 if (lo_index != min_jint) {
309 ranges[++rp].setRange(min_jint, lo_index-1, default_dest, NullTableIndex);
310 }
311 for (int j = 0; j < len; j++) {
312 jint match_int = lo_index+j;
313 int dest = iter().get_dest_table(j+3);
314 makes_backward_branch |= (dest <= bci());
315 int table_index = method_data_update() ? j : NullTableIndex;
316 if (rp < 0 || !ranges[rp].adjoin(match_int, dest, table_index)) {
317 ranges[++rp].set(match_int, dest, table_index);
318 }
319 }
320 jint highest = lo_index+(len-1);
321 assert(ranges[rp].hi() == highest, "");
322 if (highest != max_jint
323 && !ranges[rp].adjoinRange(highest+1, max_jint, default_dest, NullTableIndex)) {
324 ranges[++rp].setRange(highest+1, max_jint, default_dest, NullTableIndex);
325 }
326 assert(rp < len+2, "not too many ranges");
328 // Safepoint in case if backward branch observed
329 if( makes_backward_branch && UseLoopSafepoints )
330 add_safepoint();
332 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]);
333 }
336 //------------------------------do_lookupswitch--------------------------------
337 void Parse::do_lookupswitch() {
338 Node *lookup = pop(); // lookup value
339 // Get information about lookupswitch
340 int default_dest = iter().get_dest_table(0);
341 int len = iter().get_int_table(1);
343 if (len < 1) { // If this is a backward branch, add safepoint
344 maybe_add_safepoint(default_dest);
345 merge(default_dest);
346 return;
347 }
349 // generate decision tree, using trichotomy when possible
350 jint* table = NEW_RESOURCE_ARRAY(jint, len*2);
351 {
352 for( int j = 0; j < len; j++ ) {
353 table[j+j+0] = iter().get_int_table(2+j+j);
354 table[j+j+1] = iter().get_dest_table(2+j+j+1);
355 }
356 qsort( table, len, 2*sizeof(table[0]), jint_cmp );
357 }
359 int rnum = len*2+1;
360 bool makes_backward_branch = false;
361 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum);
362 int rp = -1;
363 for( int j = 0; j < len; j++ ) {
364 jint match_int = table[j+j+0];
365 int dest = table[j+j+1];
366 int next_lo = rp < 0 ? min_jint : ranges[rp].hi()+1;
367 int table_index = method_data_update() ? j : NullTableIndex;
368 makes_backward_branch |= (dest <= bci());
369 if( match_int != next_lo ) {
370 ranges[++rp].setRange(next_lo, match_int-1, default_dest, NullTableIndex);
371 }
372 if( rp < 0 || !ranges[rp].adjoin(match_int, dest, table_index) ) {
373 ranges[++rp].set(match_int, dest, table_index);
374 }
375 }
376 jint highest = table[2*(len-1)];
377 assert(ranges[rp].hi() == highest, "");
378 if( highest != max_jint
379 && !ranges[rp].adjoinRange(highest+1, max_jint, default_dest, NullTableIndex) ) {
380 ranges[++rp].setRange(highest+1, max_jint, default_dest, NullTableIndex);
381 }
382 assert(rp < rnum, "not too many ranges");
384 // Safepoint in case backward branch observed
385 if( makes_backward_branch && UseLoopSafepoints )
386 add_safepoint();
388 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]);
389 }
391 //----------------------------create_jump_tables-------------------------------
392 bool Parse::create_jump_tables(Node* key_val, SwitchRange* lo, SwitchRange* hi) {
393 // Are jumptables enabled
394 if (!UseJumpTables) return false;
396 // Are jumptables supported
397 if (!Matcher::has_match_rule(Op_Jump)) return false;
399 // Don't make jump table if profiling
400 if (method_data_update()) return false;
402 // Decide if a guard is needed to lop off big ranges at either (or
403 // both) end(s) of the input set. We'll call this the default target
404 // even though we can't be sure that it is the true "default".
406 bool needs_guard = false;
407 int default_dest;
408 int64 total_outlier_size = 0;
409 int64 hi_size = ((int64)hi->hi()) - ((int64)hi->lo()) + 1;
410 int64 lo_size = ((int64)lo->hi()) - ((int64)lo->lo()) + 1;
412 if (lo->dest() == hi->dest()) {
413 total_outlier_size = hi_size + lo_size;
414 default_dest = lo->dest();
415 } else if (lo_size > hi_size) {
416 total_outlier_size = lo_size;
417 default_dest = lo->dest();
418 } else {
419 total_outlier_size = hi_size;
420 default_dest = hi->dest();
421 }
423 // If a guard test will eliminate very sparse end ranges, then
424 // it is worth the cost of an extra jump.
425 if (total_outlier_size > (MaxJumpTableSparseness * 4)) {
426 needs_guard = true;
427 if (default_dest == lo->dest()) lo++;
428 if (default_dest == hi->dest()) hi--;
429 }
431 // Find the total number of cases and ranges
432 int64 num_cases = ((int64)hi->hi()) - ((int64)lo->lo()) + 1;
433 int num_range = hi - lo + 1;
435 // Don't create table if: too large, too small, or too sparse.
436 if (num_cases < MinJumpTableSize || num_cases > MaxJumpTableSize)
437 return false;
438 if (num_cases > (MaxJumpTableSparseness * num_range))
439 return false;
441 // Normalize table lookups to zero
442 int lowval = lo->lo();
443 key_val = _gvn.transform( new (C) SubINode(key_val, _gvn.intcon(lowval)) );
445 // Generate a guard to protect against input keyvals that aren't
446 // in the switch domain.
447 if (needs_guard) {
448 Node* size = _gvn.intcon(num_cases);
449 Node* cmp = _gvn.transform( new (C) CmpUNode(key_val, size) );
450 Node* tst = _gvn.transform( new (C) BoolNode(cmp, BoolTest::ge) );
451 IfNode* iff = create_and_map_if( control(), tst, PROB_FAIR, COUNT_UNKNOWN);
452 jump_if_true_fork(iff, default_dest, NullTableIndex);
453 }
455 // Create an ideal node JumpTable that has projections
456 // of all possible ranges for a switch statement
457 // The key_val input must be converted to a pointer offset and scaled.
458 // Compare Parse::array_addressing above.
459 #ifdef _LP64
460 // Clean the 32-bit int into a real 64-bit offset.
461 // Otherwise, the jint value 0 might turn into an offset of 0x0800000000.
462 const TypeInt* ikeytype = TypeInt::make(0, num_cases-1, Type::WidenMin);
463 // Make I2L conversion control dependent to prevent it from
464 // floating above the range check during loop optimizations.
465 key_val = C->constrained_convI2L(&_gvn, key_val, ikeytype, control());
466 #endif
468 // Shift the value by wordsize so we have an index into the table, rather
469 // than a switch value
470 Node *shiftWord = _gvn.MakeConX(wordSize);
471 key_val = _gvn.transform( new (C) MulXNode( key_val, shiftWord));
473 // Create the JumpNode
474 Node* jtn = _gvn.transform( new (C) JumpNode(control(), key_val, num_cases) );
476 // These are the switch destinations hanging off the jumpnode
477 int i = 0;
478 for (SwitchRange* r = lo; r <= hi; r++) {
479 for (int64 j = r->lo(); j <= r->hi(); j++, i++) {
480 Node* input = _gvn.transform(new (C) JumpProjNode(jtn, i, r->dest(), (int)(j - lowval)));
481 {
482 PreserveJVMState pjvms(this);
483 set_control(input);
484 jump_if_always_fork(r->dest(), r->table_index());
485 }
486 }
487 }
488 assert(i == num_cases, "miscount of cases");
489 stop_and_kill_map(); // no more uses for this JVMS
490 return true;
491 }
493 //----------------------------jump_switch_ranges-------------------------------
494 void Parse::jump_switch_ranges(Node* key_val, SwitchRange *lo, SwitchRange *hi, int switch_depth) {
495 Block* switch_block = block();
497 if (switch_depth == 0) {
498 // Do special processing for the top-level call.
499 assert(lo->lo() == min_jint, "initial range must exhaust Type::INT");
500 assert(hi->hi() == max_jint, "initial range must exhaust Type::INT");
502 // Decrement pred-numbers for the unique set of nodes.
503 #ifdef ASSERT
504 // Ensure that the block's successors are a (duplicate-free) set.
505 int successors_counted = 0; // block occurrences in [hi..lo]
506 int unique_successors = switch_block->num_successors();
507 for (int i = 0; i < unique_successors; i++) {
508 Block* target = switch_block->successor_at(i);
510 // Check that the set of successors is the same in both places.
511 int successors_found = 0;
512 for (SwitchRange* p = lo; p <= hi; p++) {
513 if (p->dest() == target->start()) successors_found++;
514 }
515 assert(successors_found > 0, "successor must be known");
516 successors_counted += successors_found;
517 }
518 assert(successors_counted == (hi-lo)+1, "no unexpected successors");
519 #endif
521 // Maybe prune the inputs, based on the type of key_val.
522 jint min_val = min_jint;
523 jint max_val = max_jint;
524 const TypeInt* ti = key_val->bottom_type()->isa_int();
525 if (ti != NULL) {
526 min_val = ti->_lo;
527 max_val = ti->_hi;
528 assert(min_val <= max_val, "invalid int type");
529 }
530 while (lo->hi() < min_val) lo++;
531 if (lo->lo() < min_val) lo->setRange(min_val, lo->hi(), lo->dest(), lo->table_index());
532 while (hi->lo() > max_val) hi--;
533 if (hi->hi() > max_val) hi->setRange(hi->lo(), max_val, hi->dest(), hi->table_index());
534 }
536 #ifndef PRODUCT
537 if (switch_depth == 0) {
538 _max_switch_depth = 0;
539 _est_switch_depth = log2_intptr((hi-lo+1)-1)+1;
540 }
541 #endif
543 assert(lo <= hi, "must be a non-empty set of ranges");
544 if (lo == hi) {
545 jump_if_always_fork(lo->dest(), lo->table_index());
546 } else {
547 assert(lo->hi() == (lo+1)->lo()-1, "contiguous ranges");
548 assert(hi->lo() == (hi-1)->hi()+1, "contiguous ranges");
550 if (create_jump_tables(key_val, lo, hi)) return;
552 int nr = hi - lo + 1;
554 SwitchRange* mid = lo + nr/2;
555 // if there is an easy choice, pivot at a singleton:
556 if (nr > 3 && !mid->is_singleton() && (mid-1)->is_singleton()) mid--;
558 assert(lo < mid && mid <= hi, "good pivot choice");
559 assert(nr != 2 || mid == hi, "should pick higher of 2");
560 assert(nr != 3 || mid == hi-1, "should pick middle of 3");
562 Node *test_val = _gvn.intcon(mid->lo());
564 if (mid->is_singleton()) {
565 IfNode *iff_ne = jump_if_fork_int(key_val, test_val, BoolTest::ne);
566 jump_if_false_fork(iff_ne, mid->dest(), mid->table_index());
568 // Special Case: If there are exactly three ranges, and the high
569 // and low range each go to the same place, omit the "gt" test,
570 // since it will not discriminate anything.
571 bool eq_test_only = (hi == lo+2 && hi->dest() == lo->dest());
572 if (eq_test_only) {
573 assert(mid == hi-1, "");
574 }
576 // if there is a higher range, test for it and process it:
577 if (mid < hi && !eq_test_only) {
578 // two comparisons of same values--should enable 1 test for 2 branches
579 // Use BoolTest::le instead of BoolTest::gt
580 IfNode *iff_le = jump_if_fork_int(key_val, test_val, BoolTest::le);
581 Node *iftrue = _gvn.transform( new (C) IfTrueNode(iff_le) );
582 Node *iffalse = _gvn.transform( new (C) IfFalseNode(iff_le) );
583 { PreserveJVMState pjvms(this);
584 set_control(iffalse);
585 jump_switch_ranges(key_val, mid+1, hi, switch_depth+1);
586 }
587 set_control(iftrue);
588 }
590 } else {
591 // mid is a range, not a singleton, so treat mid..hi as a unit
592 IfNode *iff_ge = jump_if_fork_int(key_val, test_val, BoolTest::ge);
594 // if there is a higher range, test for it and process it:
595 if (mid == hi) {
596 jump_if_true_fork(iff_ge, mid->dest(), mid->table_index());
597 } else {
598 Node *iftrue = _gvn.transform( new (C) IfTrueNode(iff_ge) );
599 Node *iffalse = _gvn.transform( new (C) IfFalseNode(iff_ge) );
600 { PreserveJVMState pjvms(this);
601 set_control(iftrue);
602 jump_switch_ranges(key_val, mid, hi, switch_depth+1);
603 }
604 set_control(iffalse);
605 }
606 }
608 // in any case, process the lower range
609 jump_switch_ranges(key_val, lo, mid-1, switch_depth+1);
610 }
612 // Decrease pred_count for each successor after all is done.
613 if (switch_depth == 0) {
614 int unique_successors = switch_block->num_successors();
615 for (int i = 0; i < unique_successors; i++) {
616 Block* target = switch_block->successor_at(i);
617 // Throw away the pre-allocated path for each unique successor.
618 target->next_path_num();
619 }
620 }
622 #ifndef PRODUCT
623 _max_switch_depth = MAX2(switch_depth, _max_switch_depth);
624 if (TraceOptoParse && Verbose && WizardMode && switch_depth == 0) {
625 SwitchRange* r;
626 int nsing = 0;
627 for( r = lo; r <= hi; r++ ) {
628 if( r->is_singleton() ) nsing++;
629 }
630 tty->print(">>> ");
631 _method->print_short_name();
632 tty->print_cr(" switch decision tree");
633 tty->print_cr(" %d ranges (%d singletons), max_depth=%d, est_depth=%d",
634 (int) (hi-lo+1), nsing, _max_switch_depth, _est_switch_depth);
635 if (_max_switch_depth > _est_switch_depth) {
636 tty->print_cr("******** BAD SWITCH DEPTH ********");
637 }
638 tty->print(" ");
639 for( r = lo; r <= hi; r++ ) {
640 r->print();
641 }
642 tty->cr();
643 }
644 #endif
645 }
647 void Parse::modf() {
648 Node *f2 = pop();
649 Node *f1 = pop();
650 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::modf_Type(),
651 CAST_FROM_FN_PTR(address, SharedRuntime::frem),
652 "frem", NULL, //no memory effects
653 f1, f2);
654 Node* res = _gvn.transform(new (C) ProjNode(c, TypeFunc::Parms + 0));
656 push(res);
657 }
659 void Parse::modd() {
660 Node *d2 = pop_pair();
661 Node *d1 = pop_pair();
662 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::Math_DD_D_Type(),
663 CAST_FROM_FN_PTR(address, SharedRuntime::drem),
664 "drem", NULL, //no memory effects
665 d1, top(), d2, top());
666 Node* res_d = _gvn.transform(new (C) ProjNode(c, TypeFunc::Parms + 0));
668 #ifdef ASSERT
669 Node* res_top = _gvn.transform(new (C) ProjNode(c, TypeFunc::Parms + 1));
670 assert(res_top == top(), "second value must be top");
671 #endif
673 push_pair(res_d);
674 }
676 void Parse::l2f() {
677 Node* f2 = pop();
678 Node* f1 = pop();
679 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::l2f_Type(),
680 CAST_FROM_FN_PTR(address, SharedRuntime::l2f),
681 "l2f", NULL, //no memory effects
682 f1, f2);
683 Node* res = _gvn.transform(new (C) ProjNode(c, TypeFunc::Parms + 0));
685 push(res);
686 }
688 void Parse::do_irem() {
689 // Must keep both values on the expression-stack during null-check
690 zero_check_int(peek());
691 // Compile-time detect of null-exception?
692 if (stopped()) return;
694 Node* b = pop();
695 Node* a = pop();
697 const Type *t = _gvn.type(b);
698 if (t != Type::TOP) {
699 const TypeInt *ti = t->is_int();
700 if (ti->is_con()) {
701 int divisor = ti->get_con();
702 // check for positive power of 2
703 if (divisor > 0 &&
704 (divisor & ~(divisor-1)) == divisor) {
705 // yes !
706 Node *mask = _gvn.intcon((divisor - 1));
707 // Sigh, must handle negative dividends
708 Node *zero = _gvn.intcon(0);
709 IfNode *ifff = jump_if_fork_int(a, zero, BoolTest::lt);
710 Node *iff = _gvn.transform( new (C) IfFalseNode(ifff) );
711 Node *ift = _gvn.transform( new (C) IfTrueNode (ifff) );
712 Node *reg = jump_if_join(ift, iff);
713 Node *phi = PhiNode::make(reg, NULL, TypeInt::INT);
714 // Negative path; negate/and/negate
715 Node *neg = _gvn.transform( new (C) SubINode(zero, a) );
716 Node *andn= _gvn.transform( new (C) AndINode(neg, mask) );
717 Node *negn= _gvn.transform( new (C) SubINode(zero, andn) );
718 phi->init_req(1, negn);
719 // Fast positive case
720 Node *andx = _gvn.transform( new (C) AndINode(a, mask) );
721 phi->init_req(2, andx);
722 // Push the merge
723 push( _gvn.transform(phi) );
724 return;
725 }
726 }
727 }
728 // Default case
729 push( _gvn.transform( new (C) ModINode(control(),a,b) ) );
730 }
732 // Handle jsr and jsr_w bytecode
733 void Parse::do_jsr() {
734 assert(bc() == Bytecodes::_jsr || bc() == Bytecodes::_jsr_w, "wrong bytecode");
736 // Store information about current state, tagged with new _jsr_bci
737 int return_bci = iter().next_bci();
738 int jsr_bci = (bc() == Bytecodes::_jsr) ? iter().get_dest() : iter().get_far_dest();
740 // Update method data
741 profile_taken_branch(jsr_bci);
743 // The way we do things now, there is only one successor block
744 // for the jsr, because the target code is cloned by ciTypeFlow.
745 Block* target = successor_for_bci(jsr_bci);
747 // What got pushed?
748 const Type* ret_addr = target->peek();
749 assert(ret_addr->singleton(), "must be a constant (cloned jsr body)");
751 // Effect on jsr on stack
752 push(_gvn.makecon(ret_addr));
754 // Flow to the jsr.
755 merge(jsr_bci);
756 }
758 // Handle ret bytecode
759 void Parse::do_ret() {
760 // Find to whom we return.
761 assert(block()->num_successors() == 1, "a ret can only go one place now");
762 Block* target = block()->successor_at(0);
763 assert(!target->is_ready(), "our arrival must be expected");
764 profile_ret(target->flow()->start());
765 int pnum = target->next_path_num();
766 merge_common(target, pnum);
767 }
769 static bool has_injected_profile(BoolTest::mask btest, Node* test, int& taken, int& not_taken) {
770 if (btest != BoolTest::eq && btest != BoolTest::ne) {
771 // Only ::eq and ::ne are supported for profile injection.
772 return false;
773 }
774 if (test->is_Cmp() &&
775 test->in(1)->Opcode() == Op_ProfileBoolean) {
776 ProfileBooleanNode* profile = (ProfileBooleanNode*)test->in(1);
777 int false_cnt = profile->false_count();
778 int true_cnt = profile->true_count();
780 // Counts matching depends on the actual test operation (::eq or ::ne).
781 // No need to scale the counts because profile injection was designed
782 // to feed exact counts into VM.
783 taken = (btest == BoolTest::eq) ? false_cnt : true_cnt;
784 not_taken = (btest == BoolTest::eq) ? true_cnt : false_cnt;
786 profile->consume();
787 return true;
788 }
789 return false;
790 }
791 //--------------------------dynamic_branch_prediction--------------------------
792 // Try to gather dynamic branch prediction behavior. Return a probability
793 // of the branch being taken and set the "cnt" field. Returns a -1.0
794 // if we need to use static prediction for some reason.
795 float Parse::dynamic_branch_prediction(float &cnt, BoolTest::mask btest, Node* test) {
796 ResourceMark rm;
798 cnt = COUNT_UNKNOWN;
800 int taken = 0;
801 int not_taken = 0;
803 bool use_mdo = !has_injected_profile(btest, test, taken, not_taken);
805 if (use_mdo) {
806 // Use MethodData information if it is available
807 // FIXME: free the ProfileData structure
808 ciMethodData* methodData = method()->method_data();
809 if (!methodData->is_mature()) return PROB_UNKNOWN;
810 ciProfileData* data = methodData->bci_to_data(bci());
811 if (!data->is_JumpData()) return PROB_UNKNOWN;
813 // get taken and not taken values
814 taken = data->as_JumpData()->taken();
815 not_taken = 0;
816 if (data->is_BranchData()) {
817 not_taken = data->as_BranchData()->not_taken();
818 }
820 // scale the counts to be commensurate with invocation counts:
821 taken = method()->scale_count(taken);
822 not_taken = method()->scale_count(not_taken);
823 }
825 // Give up if too few (or too many, in which case the sum will overflow) counts to be meaningful.
826 // We also check that individual counters are positive first, otherwise the sum can become positive.
827 if (taken < 0 || not_taken < 0 || taken + not_taken < 40) {
828 if (C->log() != NULL) {
829 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d'", iter().get_dest(), taken, not_taken);
830 }
831 return PROB_UNKNOWN;
832 }
834 // Compute frequency that we arrive here
835 float sum = taken + not_taken;
836 // Adjust, if this block is a cloned private block but the
837 // Jump counts are shared. Taken the private counts for
838 // just this path instead of the shared counts.
839 if( block()->count() > 0 )
840 sum = block()->count();
841 cnt = sum / FreqCountInvocations;
843 // Pin probability to sane limits
844 float prob;
845 if( !taken )
846 prob = (0+PROB_MIN) / 2;
847 else if( !not_taken )
848 prob = (1+PROB_MAX) / 2;
849 else { // Compute probability of true path
850 prob = (float)taken / (float)(taken + not_taken);
851 if (prob > PROB_MAX) prob = PROB_MAX;
852 if (prob < PROB_MIN) prob = PROB_MIN;
853 }
855 assert((cnt > 0.0f) && (prob > 0.0f),
856 "Bad frequency assignment in if");
858 if (C->log() != NULL) {
859 const char* prob_str = NULL;
860 if (prob >= PROB_MAX) prob_str = (prob == PROB_MAX) ? "max" : "always";
861 if (prob <= PROB_MIN) prob_str = (prob == PROB_MIN) ? "min" : "never";
862 char prob_str_buf[30];
863 if (prob_str == NULL) {
864 sprintf(prob_str_buf, "%g", prob);
865 prob_str = prob_str_buf;
866 }
867 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d' cnt='%g' prob='%s'",
868 iter().get_dest(), taken, not_taken, cnt, prob_str);
869 }
870 return prob;
871 }
873 //-----------------------------branch_prediction-------------------------------
874 float Parse::branch_prediction(float& cnt,
875 BoolTest::mask btest,
876 int target_bci,
877 Node* test) {
878 float prob = dynamic_branch_prediction(cnt, btest, test);
879 // If prob is unknown, switch to static prediction
880 if (prob != PROB_UNKNOWN) return prob;
882 prob = PROB_FAIR; // Set default value
883 if (btest == BoolTest::eq) // Exactly equal test?
884 prob = PROB_STATIC_INFREQUENT; // Assume its relatively infrequent
885 else if (btest == BoolTest::ne)
886 prob = PROB_STATIC_FREQUENT; // Assume its relatively frequent
888 // If this is a conditional test guarding a backwards branch,
889 // assume its a loop-back edge. Make it a likely taken branch.
890 if (target_bci < bci()) {
891 if (is_osr_parse()) { // Could be a hot OSR'd loop; force deopt
892 // Since it's an OSR, we probably have profile data, but since
893 // branch_prediction returned PROB_UNKNOWN, the counts are too small.
894 // Let's make a special check here for completely zero counts.
895 ciMethodData* methodData = method()->method_data();
896 if (!methodData->is_empty()) {
897 ciProfileData* data = methodData->bci_to_data(bci());
898 // Only stop for truly zero counts, which mean an unknown part
899 // of the OSR-ed method, and we want to deopt to gather more stats.
900 // If you have ANY counts, then this loop is simply 'cold' relative
901 // to the OSR loop.
902 if (data->as_BranchData()->taken() +
903 data->as_BranchData()->not_taken() == 0 ) {
904 // This is the only way to return PROB_UNKNOWN:
905 return PROB_UNKNOWN;
906 }
907 }
908 }
909 prob = PROB_STATIC_FREQUENT; // Likely to take backwards branch
910 }
912 assert(prob != PROB_UNKNOWN, "must have some guess at this point");
913 return prob;
914 }
916 // The magic constants are chosen so as to match the output of
917 // branch_prediction() when the profile reports a zero taken count.
918 // It is important to distinguish zero counts unambiguously, because
919 // some branches (e.g., _213_javac.Assembler.eliminate) validly produce
920 // very small but nonzero probabilities, which if confused with zero
921 // counts would keep the program recompiling indefinitely.
922 bool Parse::seems_never_taken(float prob) const {
923 return prob < PROB_MIN;
924 }
926 // True if the comparison seems to be the kind that will not change its
927 // statistics from true to false. See comments in adjust_map_after_if.
928 // This question is only asked along paths which are already
929 // classifed as untaken (by seems_never_taken), so really,
930 // if a path is never taken, its controlling comparison is
931 // already acting in a stable fashion. If the comparison
932 // seems stable, we will put an expensive uncommon trap
933 // on the untaken path.
934 bool Parse::seems_stable_comparison() const {
935 if (C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if)) {
936 return false;
937 }
938 return true;
939 }
941 //-------------------------------repush_if_args--------------------------------
942 // Push arguments of an "if" bytecode back onto the stack by adjusting _sp.
943 inline int Parse::repush_if_args() {
944 #ifndef PRODUCT
945 if (PrintOpto && WizardMode) {
946 tty->print("defending against excessive implicit null exceptions on %s @%d in ",
947 Bytecodes::name(iter().cur_bc()), iter().cur_bci());
948 method()->print_name(); tty->cr();
949 }
950 #endif
951 int bc_depth = - Bytecodes::depth(iter().cur_bc());
952 assert(bc_depth == 1 || bc_depth == 2, "only two kinds of branches");
953 DEBUG_ONLY(sync_jvms()); // argument(n) requires a synced jvms
954 assert(argument(0) != NULL, "must exist");
955 assert(bc_depth == 1 || argument(1) != NULL, "two must exist");
956 inc_sp(bc_depth);
957 return bc_depth;
958 }
960 //----------------------------------do_ifnull----------------------------------
961 void Parse::do_ifnull(BoolTest::mask btest, Node *c) {
962 int target_bci = iter().get_dest();
964 Block* branch_block = successor_for_bci(target_bci);
965 Block* next_block = successor_for_bci(iter().next_bci());
967 float cnt;
968 float prob = branch_prediction(cnt, btest, target_bci, c);
969 if (prob == PROB_UNKNOWN) {
970 // (An earlier version of do_ifnull omitted this trap for OSR methods.)
971 #ifndef PRODUCT
972 if (PrintOpto && Verbose)
973 tty->print_cr("Never-taken edge stops compilation at bci %d",bci());
974 #endif
975 repush_if_args(); // to gather stats on loop
976 // We need to mark this branch as taken so that if we recompile we will
977 // see that it is possible. In the tiered system the interpreter doesn't
978 // do profiling and by the time we get to the lower tier from the interpreter
979 // the path may be cold again. Make sure it doesn't look untaken
980 profile_taken_branch(target_bci, !ProfileInterpreter);
981 uncommon_trap(Deoptimization::Reason_unreached,
982 Deoptimization::Action_reinterpret,
983 NULL, "cold");
984 if (C->eliminate_boxing()) {
985 // Mark the successor blocks as parsed
986 branch_block->next_path_num();
987 next_block->next_path_num();
988 }
989 return;
990 }
992 explicit_null_checks_inserted++;
994 // Generate real control flow
995 Node *tst = _gvn.transform( new (C) BoolNode( c, btest ) );
997 // Sanity check the probability value
998 assert(prob > 0.0f,"Bad probability in Parser");
999 // Need xform to put node in hash table
1000 IfNode *iff = create_and_xform_if( control(), tst, prob, cnt );
1001 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser");
1002 // True branch
1003 { PreserveJVMState pjvms(this);
1004 Node* iftrue = _gvn.transform( new (C) IfTrueNode (iff) );
1005 set_control(iftrue);
1007 if (stopped()) { // Path is dead?
1008 explicit_null_checks_elided++;
1009 if (C->eliminate_boxing()) {
1010 // Mark the successor block as parsed
1011 branch_block->next_path_num();
1012 }
1013 } else { // Path is live.
1014 // Update method data
1015 profile_taken_branch(target_bci);
1016 adjust_map_after_if(btest, c, prob, branch_block, next_block);
1017 if (!stopped()) {
1018 merge(target_bci);
1019 }
1020 }
1021 }
1023 // False branch
1024 Node* iffalse = _gvn.transform( new (C) IfFalseNode(iff) );
1025 set_control(iffalse);
1027 if (stopped()) { // Path is dead?
1028 explicit_null_checks_elided++;
1029 if (C->eliminate_boxing()) {
1030 // Mark the successor block as parsed
1031 next_block->next_path_num();
1032 }
1033 } else { // Path is live.
1034 // Update method data
1035 profile_not_taken_branch();
1036 adjust_map_after_if(BoolTest(btest).negate(), c, 1.0-prob,
1037 next_block, branch_block);
1038 }
1039 }
1041 //------------------------------------do_if------------------------------------
1042 void Parse::do_if(BoolTest::mask btest, Node* c) {
1043 int target_bci = iter().get_dest();
1045 Block* branch_block = successor_for_bci(target_bci);
1046 Block* next_block = successor_for_bci(iter().next_bci());
1048 float cnt;
1049 float prob = branch_prediction(cnt, btest, target_bci, c);
1050 float untaken_prob = 1.0 - prob;
1052 if (prob == PROB_UNKNOWN) {
1053 #ifndef PRODUCT
1054 if (PrintOpto && Verbose)
1055 tty->print_cr("Never-taken edge stops compilation at bci %d",bci());
1056 #endif
1057 repush_if_args(); // to gather stats on loop
1058 // We need to mark this branch as taken so that if we recompile we will
1059 // see that it is possible. In the tiered system the interpreter doesn't
1060 // do profiling and by the time we get to the lower tier from the interpreter
1061 // the path may be cold again. Make sure it doesn't look untaken
1062 profile_taken_branch(target_bci, !ProfileInterpreter);
1063 uncommon_trap(Deoptimization::Reason_unreached,
1064 Deoptimization::Action_reinterpret,
1065 NULL, "cold");
1066 if (C->eliminate_boxing()) {
1067 // Mark the successor blocks as parsed
1068 branch_block->next_path_num();
1069 next_block->next_path_num();
1070 }
1071 return;
1072 }
1074 // Sanity check the probability value
1075 assert(0.0f < prob && prob < 1.0f,"Bad probability in Parser");
1077 bool taken_if_true = true;
1078 // Convert BoolTest to canonical form:
1079 if (!BoolTest(btest).is_canonical()) {
1080 btest = BoolTest(btest).negate();
1081 taken_if_true = false;
1082 // prob is NOT updated here; it remains the probability of the taken
1083 // path (as opposed to the prob of the path guarded by an 'IfTrueNode').
1084 }
1085 assert(btest != BoolTest::eq, "!= is the only canonical exact test");
1087 Node* tst0 = new (C) BoolNode(c, btest);
1088 Node* tst = _gvn.transform(tst0);
1089 BoolTest::mask taken_btest = BoolTest::illegal;
1090 BoolTest::mask untaken_btest = BoolTest::illegal;
1092 if (tst->is_Bool()) {
1093 // Refresh c from the transformed bool node, since it may be
1094 // simpler than the original c. Also re-canonicalize btest.
1095 // This wins when (Bool ne (Conv2B p) 0) => (Bool ne (CmpP p NULL)).
1096 // That can arise from statements like: if (x instanceof C) ...
1097 if (tst != tst0) {
1098 // Canonicalize one more time since transform can change it.
1099 btest = tst->as_Bool()->_test._test;
1100 if (!BoolTest(btest).is_canonical()) {
1101 // Reverse edges one more time...
1102 tst = _gvn.transform( tst->as_Bool()->negate(&_gvn) );
1103 btest = tst->as_Bool()->_test._test;
1104 assert(BoolTest(btest).is_canonical(), "sanity");
1105 taken_if_true = !taken_if_true;
1106 }
1107 c = tst->in(1);
1108 }
1109 BoolTest::mask neg_btest = BoolTest(btest).negate();
1110 taken_btest = taken_if_true ? btest : neg_btest;
1111 untaken_btest = taken_if_true ? neg_btest : btest;
1112 }
1114 // Generate real control flow
1115 float true_prob = (taken_if_true ? prob : untaken_prob);
1116 IfNode* iff = create_and_map_if(control(), tst, true_prob, cnt);
1117 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser");
1118 Node* taken_branch = new (C) IfTrueNode(iff);
1119 Node* untaken_branch = new (C) IfFalseNode(iff);
1120 if (!taken_if_true) { // Finish conversion to canonical form
1121 Node* tmp = taken_branch;
1122 taken_branch = untaken_branch;
1123 untaken_branch = tmp;
1124 }
1126 // Branch is taken:
1127 { PreserveJVMState pjvms(this);
1128 taken_branch = _gvn.transform(taken_branch);
1129 set_control(taken_branch);
1131 if (stopped()) {
1132 if (C->eliminate_boxing()) {
1133 // Mark the successor block as parsed
1134 branch_block->next_path_num();
1135 }
1136 } else {
1137 // Update method data
1138 profile_taken_branch(target_bci);
1139 adjust_map_after_if(taken_btest, c, prob, branch_block, next_block);
1140 if (!stopped()) {
1141 merge(target_bci);
1142 }
1143 }
1144 }
1146 untaken_branch = _gvn.transform(untaken_branch);
1147 set_control(untaken_branch);
1149 // Branch not taken.
1150 if (stopped()) {
1151 if (C->eliminate_boxing()) {
1152 // Mark the successor block as parsed
1153 next_block->next_path_num();
1154 }
1155 } else {
1156 // Update method data
1157 profile_not_taken_branch();
1158 adjust_map_after_if(untaken_btest, c, untaken_prob,
1159 next_block, branch_block);
1160 }
1161 }
1163 bool Parse::path_is_suitable_for_uncommon_trap(float prob) const {
1164 // Don't want to speculate on uncommon traps when running with -Xcomp
1165 if (!UseInterpreter) {
1166 return false;
1167 }
1168 return (seems_never_taken(prob) && seems_stable_comparison());
1169 }
1171 //----------------------------adjust_map_after_if------------------------------
1172 // Adjust the JVM state to reflect the result of taking this path.
1173 // Basically, it means inspecting the CmpNode controlling this
1174 // branch, seeing how it constrains a tested value, and then
1175 // deciding if it's worth our while to encode this constraint
1176 // as graph nodes in the current abstract interpretation map.
1177 void Parse::adjust_map_after_if(BoolTest::mask btest, Node* c, float prob,
1178 Block* path, Block* other_path) {
1179 if (stopped() || !c->is_Cmp() || btest == BoolTest::illegal)
1180 return; // nothing to do
1182 bool is_fallthrough = (path == successor_for_bci(iter().next_bci()));
1184 if (path_is_suitable_for_uncommon_trap(prob)) {
1185 repush_if_args();
1186 uncommon_trap(Deoptimization::Reason_unstable_if,
1187 Deoptimization::Action_reinterpret,
1188 NULL,
1189 (is_fallthrough ? "taken always" : "taken never"));
1190 return;
1191 }
1193 Node* val = c->in(1);
1194 Node* con = c->in(2);
1195 const Type* tcon = _gvn.type(con);
1196 const Type* tval = _gvn.type(val);
1197 bool have_con = tcon->singleton();
1198 if (tval->singleton()) {
1199 if (!have_con) {
1200 // Swap, so constant is in con.
1201 con = val;
1202 tcon = tval;
1203 val = c->in(2);
1204 tval = _gvn.type(val);
1205 btest = BoolTest(btest).commute();
1206 have_con = true;
1207 } else {
1208 // Do we have two constants? Then leave well enough alone.
1209 have_con = false;
1210 }
1211 }
1212 if (!have_con) // remaining adjustments need a con
1213 return;
1215 sharpen_type_after_if(btest, con, tcon, val, tval);
1216 }
1219 static Node* extract_obj_from_klass_load(PhaseGVN* gvn, Node* n) {
1220 Node* ldk;
1221 if (n->is_DecodeNKlass()) {
1222 if (n->in(1)->Opcode() != Op_LoadNKlass) {
1223 return NULL;
1224 } else {
1225 ldk = n->in(1);
1226 }
1227 } else if (n->Opcode() != Op_LoadKlass) {
1228 return NULL;
1229 } else {
1230 ldk = n;
1231 }
1232 assert(ldk != NULL && ldk->is_Load(), "should have found a LoadKlass or LoadNKlass node");
1234 Node* adr = ldk->in(MemNode::Address);
1235 intptr_t off = 0;
1236 Node* obj = AddPNode::Ideal_base_and_offset(adr, gvn, off);
1237 if (obj == NULL || off != oopDesc::klass_offset_in_bytes()) // loading oopDesc::_klass?
1238 return NULL;
1239 const TypePtr* tp = gvn->type(obj)->is_ptr();
1240 if (tp == NULL || !(tp->isa_instptr() || tp->isa_aryptr())) // is obj a Java object ptr?
1241 return NULL;
1243 return obj;
1244 }
1246 void Parse::sharpen_type_after_if(BoolTest::mask btest,
1247 Node* con, const Type* tcon,
1248 Node* val, const Type* tval) {
1249 // Look for opportunities to sharpen the type of a node
1250 // whose klass is compared with a constant klass.
1251 if (btest == BoolTest::eq && tcon->isa_klassptr()) {
1252 Node* obj = extract_obj_from_klass_load(&_gvn, val);
1253 const TypeOopPtr* con_type = tcon->isa_klassptr()->as_instance_type();
1254 if (obj != NULL && (con_type->isa_instptr() || con_type->isa_aryptr())) {
1255 // Found:
1256 // Bool(CmpP(LoadKlass(obj._klass), ConP(Foo.klass)), [eq])
1257 // or the narrowOop equivalent.
1258 const Type* obj_type = _gvn.type(obj);
1259 const TypeOopPtr* tboth = obj_type->join_speculative(con_type)->isa_oopptr();
1260 if (tboth != NULL && tboth->klass_is_exact() && tboth != obj_type &&
1261 tboth->higher_equal(obj_type)) {
1262 // obj has to be of the exact type Foo if the CmpP succeeds.
1263 int obj_in_map = map()->find_edge(obj);
1264 JVMState* jvms = this->jvms();
1265 if (obj_in_map >= 0 &&
1266 (jvms->is_loc(obj_in_map) || jvms->is_stk(obj_in_map))) {
1267 TypeNode* ccast = new (C) CheckCastPPNode(control(), obj, tboth);
1268 const Type* tcc = ccast->as_Type()->type();
1269 assert(tcc != obj_type && tcc->higher_equal_speculative(obj_type), "must improve");
1270 // Delay transform() call to allow recovery of pre-cast value
1271 // at the control merge.
1272 _gvn.set_type_bottom(ccast);
1273 record_for_igvn(ccast);
1274 // Here's the payoff.
1275 replace_in_map(obj, ccast);
1276 }
1277 }
1278 }
1279 }
1281 int val_in_map = map()->find_edge(val);
1282 if (val_in_map < 0) return; // replace_in_map would be useless
1283 {
1284 JVMState* jvms = this->jvms();
1285 if (!(jvms->is_loc(val_in_map) ||
1286 jvms->is_stk(val_in_map)))
1287 return; // again, it would be useless
1288 }
1290 // Check for a comparison to a constant, and "know" that the compared
1291 // value is constrained on this path.
1292 assert(tcon->singleton(), "");
1293 ConstraintCastNode* ccast = NULL;
1294 Node* cast = NULL;
1296 switch (btest) {
1297 case BoolTest::eq: // Constant test?
1298 {
1299 const Type* tboth = tcon->join_speculative(tval);
1300 if (tboth == tval) break; // Nothing to gain.
1301 if (tcon->isa_int()) {
1302 ccast = new (C) CastIINode(val, tboth);
1303 } else if (tcon == TypePtr::NULL_PTR) {
1304 // Cast to null, but keep the pointer identity temporarily live.
1305 ccast = new (C) CastPPNode(val, tboth);
1306 } else {
1307 const TypeF* tf = tcon->isa_float_constant();
1308 const TypeD* td = tcon->isa_double_constant();
1309 // Exclude tests vs float/double 0 as these could be
1310 // either +0 or -0. Just because you are equal to +0
1311 // doesn't mean you ARE +0!
1312 // Note, following code also replaces Long and Oop values.
1313 if ((!tf || tf->_f != 0.0) &&
1314 (!td || td->_d != 0.0))
1315 cast = con; // Replace non-constant val by con.
1316 }
1317 }
1318 break;
1320 case BoolTest::ne:
1321 if (tcon == TypePtr::NULL_PTR) {
1322 cast = cast_not_null(val, false);
1323 }
1324 break;
1326 default:
1327 // (At this point we could record int range types with CastII.)
1328 break;
1329 }
1331 if (ccast != NULL) {
1332 const Type* tcc = ccast->as_Type()->type();
1333 assert(tcc != tval && tcc->higher_equal_speculative(tval), "must improve");
1334 // Delay transform() call to allow recovery of pre-cast value
1335 // at the control merge.
1336 ccast->set_req(0, control());
1337 _gvn.set_type_bottom(ccast);
1338 record_for_igvn(ccast);
1339 cast = ccast;
1340 }
1342 if (cast != NULL) { // Here's the payoff.
1343 replace_in_map(val, cast);
1344 }
1345 }
1347 /**
1348 * Use speculative type to optimize CmpP node: if comparison is
1349 * against the low level class, cast the object to the speculative
1350 * type if any. CmpP should then go away.
1351 *
1352 * @param c expected CmpP node
1353 * @return result of CmpP on object casted to speculative type
1354 *
1355 */
1356 Node* Parse::optimize_cmp_with_klass(Node* c) {
1357 // If this is transformed by the _gvn to a comparison with the low
1358 // level klass then we may be able to use speculation
1359 if (c->Opcode() == Op_CmpP &&
1360 (c->in(1)->Opcode() == Op_LoadKlass || c->in(1)->Opcode() == Op_DecodeNKlass) &&
1361 c->in(2)->is_Con()) {
1362 Node* load_klass = NULL;
1363 Node* decode = NULL;
1364 if (c->in(1)->Opcode() == Op_DecodeNKlass) {
1365 decode = c->in(1);
1366 load_klass = c->in(1)->in(1);
1367 } else {
1368 load_klass = c->in(1);
1369 }
1370 if (load_klass->in(2)->is_AddP()) {
1371 Node* addp = load_klass->in(2);
1372 Node* obj = addp->in(AddPNode::Address);
1373 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
1374 if (obj_type->speculative_type() != NULL) {
1375 ciKlass* k = obj_type->speculative_type();
1376 inc_sp(2);
1377 obj = maybe_cast_profiled_obj(obj, k);
1378 dec_sp(2);
1379 // Make the CmpP use the casted obj
1380 addp = basic_plus_adr(obj, addp->in(AddPNode::Offset));
1381 load_klass = load_klass->clone();
1382 load_klass->set_req(2, addp);
1383 load_klass = _gvn.transform(load_klass);
1384 if (decode != NULL) {
1385 decode = decode->clone();
1386 decode->set_req(1, load_klass);
1387 load_klass = _gvn.transform(decode);
1388 }
1389 c = c->clone();
1390 c->set_req(1, load_klass);
1391 c = _gvn.transform(c);
1392 }
1393 }
1394 }
1395 return c;
1396 }
1398 //------------------------------do_one_bytecode--------------------------------
1399 // Parse this bytecode, and alter the Parsers JVM->Node mapping
1400 void Parse::do_one_bytecode() {
1401 Node *a, *b, *c, *d; // Handy temps
1402 BoolTest::mask btest;
1403 int i;
1405 assert(!has_exceptions(), "bytecode entry state must be clear of throws");
1407 if (C->check_node_count(NodeLimitFudgeFactor * 5,
1408 "out of nodes parsing method")) {
1409 return;
1410 }
1412 #ifdef ASSERT
1413 // for setting breakpoints
1414 if (TraceOptoParse) {
1415 tty->print(" @");
1416 dump_bci(bci());
1417 tty->cr();
1418 }
1419 #endif
1421 switch (bc()) {
1422 case Bytecodes::_nop:
1423 // do nothing
1424 break;
1425 case Bytecodes::_lconst_0:
1426 push_pair(longcon(0));
1427 break;
1429 case Bytecodes::_lconst_1:
1430 push_pair(longcon(1));
1431 break;
1433 case Bytecodes::_fconst_0:
1434 push(zerocon(T_FLOAT));
1435 break;
1437 case Bytecodes::_fconst_1:
1438 push(makecon(TypeF::ONE));
1439 break;
1441 case Bytecodes::_fconst_2:
1442 push(makecon(TypeF::make(2.0f)));
1443 break;
1445 case Bytecodes::_dconst_0:
1446 push_pair(zerocon(T_DOUBLE));
1447 break;
1449 case Bytecodes::_dconst_1:
1450 push_pair(makecon(TypeD::ONE));
1451 break;
1453 case Bytecodes::_iconst_m1:push(intcon(-1)); break;
1454 case Bytecodes::_iconst_0: push(intcon( 0)); break;
1455 case Bytecodes::_iconst_1: push(intcon( 1)); break;
1456 case Bytecodes::_iconst_2: push(intcon( 2)); break;
1457 case Bytecodes::_iconst_3: push(intcon( 3)); break;
1458 case Bytecodes::_iconst_4: push(intcon( 4)); break;
1459 case Bytecodes::_iconst_5: push(intcon( 5)); break;
1460 case Bytecodes::_bipush: push(intcon(iter().get_constant_u1())); break;
1461 case Bytecodes::_sipush: push(intcon(iter().get_constant_u2())); break;
1462 case Bytecodes::_aconst_null: push(null()); break;
1463 case Bytecodes::_ldc:
1464 case Bytecodes::_ldc_w:
1465 case Bytecodes::_ldc2_w:
1466 // If the constant is unresolved, run this BC once in the interpreter.
1467 {
1468 ciConstant constant = iter().get_constant();
1469 if (constant.basic_type() == T_OBJECT &&
1470 !constant.as_object()->is_loaded()) {
1471 int index = iter().get_constant_pool_index();
1472 constantTag tag = iter().get_constant_pool_tag(index);
1473 uncommon_trap(Deoptimization::make_trap_request
1474 (Deoptimization::Reason_unloaded,
1475 Deoptimization::Action_reinterpret,
1476 index),
1477 NULL, tag.internal_name());
1478 break;
1479 }
1480 assert(constant.basic_type() != T_OBJECT || constant.as_object()->is_instance(),
1481 "must be java_mirror of klass");
1482 bool pushed = push_constant(constant, true);
1483 guarantee(pushed, "must be possible to push this constant");
1484 }
1486 break;
1488 case Bytecodes::_aload_0:
1489 push( local(0) );
1490 break;
1491 case Bytecodes::_aload_1:
1492 push( local(1) );
1493 break;
1494 case Bytecodes::_aload_2:
1495 push( local(2) );
1496 break;
1497 case Bytecodes::_aload_3:
1498 push( local(3) );
1499 break;
1500 case Bytecodes::_aload:
1501 push( local(iter().get_index()) );
1502 break;
1504 case Bytecodes::_fload_0:
1505 case Bytecodes::_iload_0:
1506 push( local(0) );
1507 break;
1508 case Bytecodes::_fload_1:
1509 case Bytecodes::_iload_1:
1510 push( local(1) );
1511 break;
1512 case Bytecodes::_fload_2:
1513 case Bytecodes::_iload_2:
1514 push( local(2) );
1515 break;
1516 case Bytecodes::_fload_3:
1517 case Bytecodes::_iload_3:
1518 push( local(3) );
1519 break;
1520 case Bytecodes::_fload:
1521 case Bytecodes::_iload:
1522 push( local(iter().get_index()) );
1523 break;
1524 case Bytecodes::_lload_0:
1525 push_pair_local( 0 );
1526 break;
1527 case Bytecodes::_lload_1:
1528 push_pair_local( 1 );
1529 break;
1530 case Bytecodes::_lload_2:
1531 push_pair_local( 2 );
1532 break;
1533 case Bytecodes::_lload_3:
1534 push_pair_local( 3 );
1535 break;
1536 case Bytecodes::_lload:
1537 push_pair_local( iter().get_index() );
1538 break;
1540 case Bytecodes::_dload_0:
1541 push_pair_local(0);
1542 break;
1543 case Bytecodes::_dload_1:
1544 push_pair_local(1);
1545 break;
1546 case Bytecodes::_dload_2:
1547 push_pair_local(2);
1548 break;
1549 case Bytecodes::_dload_3:
1550 push_pair_local(3);
1551 break;
1552 case Bytecodes::_dload:
1553 push_pair_local(iter().get_index());
1554 break;
1555 case Bytecodes::_fstore_0:
1556 case Bytecodes::_istore_0:
1557 case Bytecodes::_astore_0:
1558 set_local( 0, pop() );
1559 break;
1560 case Bytecodes::_fstore_1:
1561 case Bytecodes::_istore_1:
1562 case Bytecodes::_astore_1:
1563 set_local( 1, pop() );
1564 break;
1565 case Bytecodes::_fstore_2:
1566 case Bytecodes::_istore_2:
1567 case Bytecodes::_astore_2:
1568 set_local( 2, pop() );
1569 break;
1570 case Bytecodes::_fstore_3:
1571 case Bytecodes::_istore_3:
1572 case Bytecodes::_astore_3:
1573 set_local( 3, pop() );
1574 break;
1575 case Bytecodes::_fstore:
1576 case Bytecodes::_istore:
1577 case Bytecodes::_astore:
1578 set_local( iter().get_index(), pop() );
1579 break;
1580 // long stores
1581 case Bytecodes::_lstore_0:
1582 set_pair_local( 0, pop_pair() );
1583 break;
1584 case Bytecodes::_lstore_1:
1585 set_pair_local( 1, pop_pair() );
1586 break;
1587 case Bytecodes::_lstore_2:
1588 set_pair_local( 2, pop_pair() );
1589 break;
1590 case Bytecodes::_lstore_3:
1591 set_pair_local( 3, pop_pair() );
1592 break;
1593 case Bytecodes::_lstore:
1594 set_pair_local( iter().get_index(), pop_pair() );
1595 break;
1597 // double stores
1598 case Bytecodes::_dstore_0:
1599 set_pair_local( 0, dstore_rounding(pop_pair()) );
1600 break;
1601 case Bytecodes::_dstore_1:
1602 set_pair_local( 1, dstore_rounding(pop_pair()) );
1603 break;
1604 case Bytecodes::_dstore_2:
1605 set_pair_local( 2, dstore_rounding(pop_pair()) );
1606 break;
1607 case Bytecodes::_dstore_3:
1608 set_pair_local( 3, dstore_rounding(pop_pair()) );
1609 break;
1610 case Bytecodes::_dstore:
1611 set_pair_local( iter().get_index(), dstore_rounding(pop_pair()) );
1612 break;
1614 case Bytecodes::_pop: dec_sp(1); break;
1615 case Bytecodes::_pop2: dec_sp(2); break;
1616 case Bytecodes::_swap:
1617 a = pop();
1618 b = pop();
1619 push(a);
1620 push(b);
1621 break;
1622 case Bytecodes::_dup:
1623 a = pop();
1624 push(a);
1625 push(a);
1626 break;
1627 case Bytecodes::_dup_x1:
1628 a = pop();
1629 b = pop();
1630 push( a );
1631 push( b );
1632 push( a );
1633 break;
1634 case Bytecodes::_dup_x2:
1635 a = pop();
1636 b = pop();
1637 c = pop();
1638 push( a );
1639 push( c );
1640 push( b );
1641 push( a );
1642 break;
1643 case Bytecodes::_dup2:
1644 a = pop();
1645 b = pop();
1646 push( b );
1647 push( a );
1648 push( b );
1649 push( a );
1650 break;
1652 case Bytecodes::_dup2_x1:
1653 // before: .. c, b, a
1654 // after: .. b, a, c, b, a
1655 // not tested
1656 a = pop();
1657 b = pop();
1658 c = pop();
1659 push( b );
1660 push( a );
1661 push( c );
1662 push( b );
1663 push( a );
1664 break;
1665 case Bytecodes::_dup2_x2:
1666 // before: .. d, c, b, a
1667 // after: .. b, a, d, c, b, a
1668 // not tested
1669 a = pop();
1670 b = pop();
1671 c = pop();
1672 d = pop();
1673 push( b );
1674 push( a );
1675 push( d );
1676 push( c );
1677 push( b );
1678 push( a );
1679 break;
1681 case Bytecodes::_arraylength: {
1682 // Must do null-check with value on expression stack
1683 Node *ary = null_check(peek(), T_ARRAY);
1684 // Compile-time detect of null-exception?
1685 if (stopped()) return;
1686 a = pop();
1687 push(load_array_length(a));
1688 break;
1689 }
1691 case Bytecodes::_baload: array_load(T_BYTE); break;
1692 case Bytecodes::_caload: array_load(T_CHAR); break;
1693 case Bytecodes::_iaload: array_load(T_INT); break;
1694 case Bytecodes::_saload: array_load(T_SHORT); break;
1695 case Bytecodes::_faload: array_load(T_FLOAT); break;
1696 case Bytecodes::_aaload: array_load(T_OBJECT); break;
1697 case Bytecodes::_laload: {
1698 a = array_addressing(T_LONG, 0);
1699 if (stopped()) return; // guaranteed null or range check
1700 dec_sp(2); // Pop array and index
1701 push_pair(make_load(control(), a, TypeLong::LONG, T_LONG, TypeAryPtr::LONGS, MemNode::unordered));
1702 break;
1703 }
1704 case Bytecodes::_daload: {
1705 a = array_addressing(T_DOUBLE, 0);
1706 if (stopped()) return; // guaranteed null or range check
1707 dec_sp(2); // Pop array and index
1708 push_pair(make_load(control(), a, Type::DOUBLE, T_DOUBLE, TypeAryPtr::DOUBLES, MemNode::unordered));
1709 break;
1710 }
1711 case Bytecodes::_bastore: array_store(T_BYTE); break;
1712 case Bytecodes::_castore: array_store(T_CHAR); break;
1713 case Bytecodes::_iastore: array_store(T_INT); break;
1714 case Bytecodes::_sastore: array_store(T_SHORT); break;
1715 case Bytecodes::_fastore: array_store(T_FLOAT); break;
1716 case Bytecodes::_aastore: {
1717 d = array_addressing(T_OBJECT, 1);
1718 if (stopped()) return; // guaranteed null or range check
1719 array_store_check();
1720 c = pop(); // Oop to store
1721 b = pop(); // index (already used)
1722 a = pop(); // the array itself
1723 const TypeOopPtr* elemtype = _gvn.type(a)->is_aryptr()->elem()->make_oopptr();
1724 const TypeAryPtr* adr_type = TypeAryPtr::OOPS;
1725 Node* store = store_oop_to_array(control(), a, d, adr_type, c, elemtype, T_OBJECT, MemNode::release);
1726 break;
1727 }
1728 case Bytecodes::_lastore: {
1729 a = array_addressing(T_LONG, 2);
1730 if (stopped()) return; // guaranteed null or range check
1731 c = pop_pair();
1732 dec_sp(2); // Pop array and index
1733 store_to_memory(control(), a, c, T_LONG, TypeAryPtr::LONGS, MemNode::unordered);
1734 break;
1735 }
1736 case Bytecodes::_dastore: {
1737 a = array_addressing(T_DOUBLE, 2);
1738 if (stopped()) return; // guaranteed null or range check
1739 c = pop_pair();
1740 dec_sp(2); // Pop array and index
1741 c = dstore_rounding(c);
1742 store_to_memory(control(), a, c, T_DOUBLE, TypeAryPtr::DOUBLES, MemNode::unordered);
1743 break;
1744 }
1745 case Bytecodes::_getfield:
1746 do_getfield();
1747 break;
1749 case Bytecodes::_getstatic:
1750 do_getstatic();
1751 break;
1753 case Bytecodes::_putfield:
1754 do_putfield();
1755 break;
1757 case Bytecodes::_putstatic:
1758 do_putstatic();
1759 break;
1761 case Bytecodes::_irem:
1762 do_irem();
1763 break;
1764 case Bytecodes::_idiv:
1765 // Must keep both values on the expression-stack during null-check
1766 zero_check_int(peek());
1767 // Compile-time detect of null-exception?
1768 if (stopped()) return;
1769 b = pop();
1770 a = pop();
1771 push( _gvn.transform( new (C) DivINode(control(),a,b) ) );
1772 break;
1773 case Bytecodes::_imul:
1774 b = pop(); a = pop();
1775 push( _gvn.transform( new (C) MulINode(a,b) ) );
1776 break;
1777 case Bytecodes::_iadd:
1778 b = pop(); a = pop();
1779 push( _gvn.transform( new (C) AddINode(a,b) ) );
1780 break;
1781 case Bytecodes::_ineg:
1782 a = pop();
1783 push( _gvn.transform( new (C) SubINode(_gvn.intcon(0),a)) );
1784 break;
1785 case Bytecodes::_isub:
1786 b = pop(); a = pop();
1787 push( _gvn.transform( new (C) SubINode(a,b) ) );
1788 break;
1789 case Bytecodes::_iand:
1790 b = pop(); a = pop();
1791 push( _gvn.transform( new (C) AndINode(a,b) ) );
1792 break;
1793 case Bytecodes::_ior:
1794 b = pop(); a = pop();
1795 push( _gvn.transform( new (C) OrINode(a,b) ) );
1796 break;
1797 case Bytecodes::_ixor:
1798 b = pop(); a = pop();
1799 push( _gvn.transform( new (C) XorINode(a,b) ) );
1800 break;
1801 case Bytecodes::_ishl:
1802 b = pop(); a = pop();
1803 push( _gvn.transform( new (C) LShiftINode(a,b) ) );
1804 break;
1805 case Bytecodes::_ishr:
1806 b = pop(); a = pop();
1807 push( _gvn.transform( new (C) RShiftINode(a,b) ) );
1808 break;
1809 case Bytecodes::_iushr:
1810 b = pop(); a = pop();
1811 push( _gvn.transform( new (C) URShiftINode(a,b) ) );
1812 break;
1814 case Bytecodes::_fneg:
1815 a = pop();
1816 b = _gvn.transform(new (C) NegFNode (a));
1817 push(b);
1818 break;
1820 case Bytecodes::_fsub:
1821 b = pop();
1822 a = pop();
1823 c = _gvn.transform( new (C) SubFNode(a,b) );
1824 d = precision_rounding(c);
1825 push( d );
1826 break;
1828 case Bytecodes::_fadd:
1829 b = pop();
1830 a = pop();
1831 c = _gvn.transform( new (C) AddFNode(a,b) );
1832 d = precision_rounding(c);
1833 push( d );
1834 break;
1836 case Bytecodes::_fmul:
1837 b = pop();
1838 a = pop();
1839 c = _gvn.transform( new (C) MulFNode(a,b) );
1840 d = precision_rounding(c);
1841 push( d );
1842 break;
1844 case Bytecodes::_fdiv:
1845 b = pop();
1846 a = pop();
1847 c = _gvn.transform( new (C) DivFNode(0,a,b) );
1848 d = precision_rounding(c);
1849 push( d );
1850 break;
1852 case Bytecodes::_frem:
1853 if (Matcher::has_match_rule(Op_ModF)) {
1854 // Generate a ModF node.
1855 b = pop();
1856 a = pop();
1857 c = _gvn.transform( new (C) ModFNode(0,a,b) );
1858 d = precision_rounding(c);
1859 push( d );
1860 }
1861 else {
1862 // Generate a call.
1863 modf();
1864 }
1865 break;
1867 case Bytecodes::_fcmpl:
1868 b = pop();
1869 a = pop();
1870 c = _gvn.transform( new (C) CmpF3Node( a, b));
1871 push(c);
1872 break;
1873 case Bytecodes::_fcmpg:
1874 b = pop();
1875 a = pop();
1877 // Same as fcmpl but need to flip the unordered case. Swap the inputs,
1878 // which negates the result sign except for unordered. Flip the unordered
1879 // as well by using CmpF3 which implements unordered-lesser instead of
1880 // unordered-greater semantics. Finally, commute the result bits. Result
1881 // is same as using a CmpF3Greater except we did it with CmpF3 alone.
1882 c = _gvn.transform( new (C) CmpF3Node( b, a));
1883 c = _gvn.transform( new (C) SubINode(_gvn.intcon(0),c) );
1884 push(c);
1885 break;
1887 case Bytecodes::_f2i:
1888 a = pop();
1889 push(_gvn.transform(new (C) ConvF2INode(a)));
1890 break;
1892 case Bytecodes::_d2i:
1893 a = pop_pair();
1894 b = _gvn.transform(new (C) ConvD2INode(a));
1895 push( b );
1896 break;
1898 case Bytecodes::_f2d:
1899 a = pop();
1900 b = _gvn.transform( new (C) ConvF2DNode(a));
1901 push_pair( b );
1902 break;
1904 case Bytecodes::_d2f:
1905 a = pop_pair();
1906 b = _gvn.transform( new (C) ConvD2FNode(a));
1907 // This breaks _227_mtrt (speed & correctness) and _222_mpegaudio (speed)
1908 //b = _gvn.transform(new (C) RoundFloatNode(0, b) );
1909 push( b );
1910 break;
1912 case Bytecodes::_l2f:
1913 if (Matcher::convL2FSupported()) {
1914 a = pop_pair();
1915 b = _gvn.transform( new (C) ConvL2FNode(a));
1916 // For i486.ad, FILD doesn't restrict precision to 24 or 53 bits.
1917 // Rather than storing the result into an FP register then pushing
1918 // out to memory to round, the machine instruction that implements
1919 // ConvL2D is responsible for rounding.
1920 // c = precision_rounding(b);
1921 c = _gvn.transform(b);
1922 push(c);
1923 } else {
1924 l2f();
1925 }
1926 break;
1928 case Bytecodes::_l2d:
1929 a = pop_pair();
1930 b = _gvn.transform( new (C) ConvL2DNode(a));
1931 // For i486.ad, rounding is always necessary (see _l2f above).
1932 // c = dprecision_rounding(b);
1933 c = _gvn.transform(b);
1934 push_pair(c);
1935 break;
1937 case Bytecodes::_f2l:
1938 a = pop();
1939 b = _gvn.transform( new (C) ConvF2LNode(a));
1940 push_pair(b);
1941 break;
1943 case Bytecodes::_d2l:
1944 a = pop_pair();
1945 b = _gvn.transform( new (C) ConvD2LNode(a));
1946 push_pair(b);
1947 break;
1949 case Bytecodes::_dsub:
1950 b = pop_pair();
1951 a = pop_pair();
1952 c = _gvn.transform( new (C) SubDNode(a,b) );
1953 d = dprecision_rounding(c);
1954 push_pair( d );
1955 break;
1957 case Bytecodes::_dadd:
1958 b = pop_pair();
1959 a = pop_pair();
1960 c = _gvn.transform( new (C) AddDNode(a,b) );
1961 d = dprecision_rounding(c);
1962 push_pair( d );
1963 break;
1965 case Bytecodes::_dmul:
1966 b = pop_pair();
1967 a = pop_pair();
1968 c = _gvn.transform( new (C) MulDNode(a,b) );
1969 d = dprecision_rounding(c);
1970 push_pair( d );
1971 break;
1973 case Bytecodes::_ddiv:
1974 b = pop_pair();
1975 a = pop_pair();
1976 c = _gvn.transform( new (C) DivDNode(0,a,b) );
1977 d = dprecision_rounding(c);
1978 push_pair( d );
1979 break;
1981 case Bytecodes::_dneg:
1982 a = pop_pair();
1983 b = _gvn.transform(new (C) NegDNode (a));
1984 push_pair(b);
1985 break;
1987 case Bytecodes::_drem:
1988 if (Matcher::has_match_rule(Op_ModD)) {
1989 // Generate a ModD node.
1990 b = pop_pair();
1991 a = pop_pair();
1992 // a % b
1994 c = _gvn.transform( new (C) ModDNode(0,a,b) );
1995 d = dprecision_rounding(c);
1996 push_pair( d );
1997 }
1998 else {
1999 // Generate a call.
2000 modd();
2001 }
2002 break;
2004 case Bytecodes::_dcmpl:
2005 b = pop_pair();
2006 a = pop_pair();
2007 c = _gvn.transform( new (C) CmpD3Node( a, b));
2008 push(c);
2009 break;
2011 case Bytecodes::_dcmpg:
2012 b = pop_pair();
2013 a = pop_pair();
2014 // Same as dcmpl but need to flip the unordered case.
2015 // Commute the inputs, which negates the result sign except for unordered.
2016 // Flip the unordered as well by using CmpD3 which implements
2017 // unordered-lesser instead of unordered-greater semantics.
2018 // Finally, negate the result bits. Result is same as using a
2019 // CmpD3Greater except we did it with CmpD3 alone.
2020 c = _gvn.transform( new (C) CmpD3Node( b, a));
2021 c = _gvn.transform( new (C) SubINode(_gvn.intcon(0),c) );
2022 push(c);
2023 break;
2026 // Note for longs -> lo word is on TOS, hi word is on TOS - 1
2027 case Bytecodes::_land:
2028 b = pop_pair();
2029 a = pop_pair();
2030 c = _gvn.transform( new (C) AndLNode(a,b) );
2031 push_pair(c);
2032 break;
2033 case Bytecodes::_lor:
2034 b = pop_pair();
2035 a = pop_pair();
2036 c = _gvn.transform( new (C) OrLNode(a,b) );
2037 push_pair(c);
2038 break;
2039 case Bytecodes::_lxor:
2040 b = pop_pair();
2041 a = pop_pair();
2042 c = _gvn.transform( new (C) XorLNode(a,b) );
2043 push_pair(c);
2044 break;
2046 case Bytecodes::_lshl:
2047 b = pop(); // the shift count
2048 a = pop_pair(); // value to be shifted
2049 c = _gvn.transform( new (C) LShiftLNode(a,b) );
2050 push_pair(c);
2051 break;
2052 case Bytecodes::_lshr:
2053 b = pop(); // the shift count
2054 a = pop_pair(); // value to be shifted
2055 c = _gvn.transform( new (C) RShiftLNode(a,b) );
2056 push_pair(c);
2057 break;
2058 case Bytecodes::_lushr:
2059 b = pop(); // the shift count
2060 a = pop_pair(); // value to be shifted
2061 c = _gvn.transform( new (C) URShiftLNode(a,b) );
2062 push_pair(c);
2063 break;
2064 case Bytecodes::_lmul:
2065 b = pop_pair();
2066 a = pop_pair();
2067 c = _gvn.transform( new (C) MulLNode(a,b) );
2068 push_pair(c);
2069 break;
2071 case Bytecodes::_lrem:
2072 // Must keep both values on the expression-stack during null-check
2073 assert(peek(0) == top(), "long word order");
2074 zero_check_long(peek(1));
2075 // Compile-time detect of null-exception?
2076 if (stopped()) return;
2077 b = pop_pair();
2078 a = pop_pair();
2079 c = _gvn.transform( new (C) ModLNode(control(),a,b) );
2080 push_pair(c);
2081 break;
2083 case Bytecodes::_ldiv:
2084 // Must keep both values on the expression-stack during null-check
2085 assert(peek(0) == top(), "long word order");
2086 zero_check_long(peek(1));
2087 // Compile-time detect of null-exception?
2088 if (stopped()) return;
2089 b = pop_pair();
2090 a = pop_pair();
2091 c = _gvn.transform( new (C) DivLNode(control(),a,b) );
2092 push_pair(c);
2093 break;
2095 case Bytecodes::_ladd:
2096 b = pop_pair();
2097 a = pop_pair();
2098 c = _gvn.transform( new (C) AddLNode(a,b) );
2099 push_pair(c);
2100 break;
2101 case Bytecodes::_lsub:
2102 b = pop_pair();
2103 a = pop_pair();
2104 c = _gvn.transform( new (C) SubLNode(a,b) );
2105 push_pair(c);
2106 break;
2107 case Bytecodes::_lcmp:
2108 // Safepoints are now inserted _before_ branches. The long-compare
2109 // bytecode painfully produces a 3-way value (-1,0,+1) which requires a
2110 // slew of control flow. These are usually followed by a CmpI vs zero and
2111 // a branch; this pattern then optimizes to the obvious long-compare and
2112 // branch. However, if the branch is backwards there's a Safepoint
2113 // inserted. The inserted Safepoint captures the JVM state at the
2114 // pre-branch point, i.e. it captures the 3-way value. Thus if a
2115 // long-compare is used to control a loop the debug info will force
2116 // computation of the 3-way value, even though the generated code uses a
2117 // long-compare and branch. We try to rectify the situation by inserting
2118 // a SafePoint here and have it dominate and kill the safepoint added at a
2119 // following backwards branch. At this point the JVM state merely holds 2
2120 // longs but not the 3-way value.
2121 if( UseLoopSafepoints ) {
2122 switch( iter().next_bc() ) {
2123 case Bytecodes::_ifgt:
2124 case Bytecodes::_iflt:
2125 case Bytecodes::_ifge:
2126 case Bytecodes::_ifle:
2127 case Bytecodes::_ifne:
2128 case Bytecodes::_ifeq:
2129 // If this is a backwards branch in the bytecodes, add Safepoint
2130 maybe_add_safepoint(iter().next_get_dest());
2131 }
2132 }
2133 b = pop_pair();
2134 a = pop_pair();
2135 c = _gvn.transform( new (C) CmpL3Node( a, b ));
2136 push(c);
2137 break;
2139 case Bytecodes::_lneg:
2140 a = pop_pair();
2141 b = _gvn.transform( new (C) SubLNode(longcon(0),a));
2142 push_pair(b);
2143 break;
2144 case Bytecodes::_l2i:
2145 a = pop_pair();
2146 push( _gvn.transform( new (C) ConvL2INode(a)));
2147 break;
2148 case Bytecodes::_i2l:
2149 a = pop();
2150 b = _gvn.transform( new (C) ConvI2LNode(a));
2151 push_pair(b);
2152 break;
2153 case Bytecodes::_i2b:
2154 // Sign extend
2155 a = pop();
2156 a = _gvn.transform( new (C) LShiftINode(a,_gvn.intcon(24)) );
2157 a = _gvn.transform( new (C) RShiftINode(a,_gvn.intcon(24)) );
2158 push( a );
2159 break;
2160 case Bytecodes::_i2s:
2161 a = pop();
2162 a = _gvn.transform( new (C) LShiftINode(a,_gvn.intcon(16)) );
2163 a = _gvn.transform( new (C) RShiftINode(a,_gvn.intcon(16)) );
2164 push( a );
2165 break;
2166 case Bytecodes::_i2c:
2167 a = pop();
2168 push( _gvn.transform( new (C) AndINode(a,_gvn.intcon(0xFFFF)) ) );
2169 break;
2171 case Bytecodes::_i2f:
2172 a = pop();
2173 b = _gvn.transform( new (C) ConvI2FNode(a) ) ;
2174 c = precision_rounding(b);
2175 push (b);
2176 break;
2178 case Bytecodes::_i2d:
2179 a = pop();
2180 b = _gvn.transform( new (C) ConvI2DNode(a));
2181 push_pair(b);
2182 break;
2184 case Bytecodes::_iinc: // Increment local
2185 i = iter().get_index(); // Get local index
2186 set_local( i, _gvn.transform( new (C) AddINode( _gvn.intcon(iter().get_iinc_con()), local(i) ) ) );
2187 break;
2189 // Exit points of synchronized methods must have an unlock node
2190 case Bytecodes::_return:
2191 return_current(NULL);
2192 break;
2194 case Bytecodes::_ireturn:
2195 case Bytecodes::_areturn:
2196 case Bytecodes::_freturn:
2197 return_current(pop());
2198 break;
2199 case Bytecodes::_lreturn:
2200 return_current(pop_pair());
2201 break;
2202 case Bytecodes::_dreturn:
2203 return_current(pop_pair());
2204 break;
2206 case Bytecodes::_athrow:
2207 // null exception oop throws NULL pointer exception
2208 null_check(peek());
2209 if (stopped()) return;
2210 // Hook the thrown exception directly to subsequent handlers.
2211 if (BailoutToInterpreterForThrows) {
2212 // Keep method interpreted from now on.
2213 uncommon_trap(Deoptimization::Reason_unhandled,
2214 Deoptimization::Action_make_not_compilable);
2215 return;
2216 }
2217 if (env()->jvmti_can_post_on_exceptions()) {
2218 // check if we must post exception events, take uncommon trap if so (with must_throw = false)
2219 uncommon_trap_if_should_post_on_exceptions(Deoptimization::Reason_unhandled, false);
2220 }
2221 // Here if either can_post_on_exceptions or should_post_on_exceptions is false
2222 add_exception_state(make_exception_state(peek()));
2223 break;
2225 case Bytecodes::_goto: // fall through
2226 case Bytecodes::_goto_w: {
2227 int target_bci = (bc() == Bytecodes::_goto) ? iter().get_dest() : iter().get_far_dest();
2229 // If this is a backwards branch in the bytecodes, add Safepoint
2230 maybe_add_safepoint(target_bci);
2232 // Update method data
2233 profile_taken_branch(target_bci);
2235 // Merge the current control into the target basic block
2236 merge(target_bci);
2238 // See if we can get some profile data and hand it off to the next block
2239 Block *target_block = block()->successor_for_bci(target_bci);
2240 if (target_block->pred_count() != 1) break;
2241 ciMethodData* methodData = method()->method_data();
2242 if (!methodData->is_mature()) break;
2243 ciProfileData* data = methodData->bci_to_data(bci());
2244 assert( data->is_JumpData(), "" );
2245 int taken = ((ciJumpData*)data)->taken();
2246 taken = method()->scale_count(taken);
2247 target_block->set_count(taken);
2248 break;
2249 }
2251 case Bytecodes::_ifnull: btest = BoolTest::eq; goto handle_if_null;
2252 case Bytecodes::_ifnonnull: btest = BoolTest::ne; goto handle_if_null;
2253 handle_if_null:
2254 // If this is a backwards branch in the bytecodes, add Safepoint
2255 maybe_add_safepoint(iter().get_dest());
2256 a = null();
2257 b = pop();
2258 c = _gvn.transform( new (C) CmpPNode(b, a) );
2259 do_ifnull(btest, c);
2260 break;
2262 case Bytecodes::_if_acmpeq: btest = BoolTest::eq; goto handle_if_acmp;
2263 case Bytecodes::_if_acmpne: btest = BoolTest::ne; goto handle_if_acmp;
2264 handle_if_acmp:
2265 // If this is a backwards branch in the bytecodes, add Safepoint
2266 maybe_add_safepoint(iter().get_dest());
2267 a = pop();
2268 b = pop();
2269 c = _gvn.transform( new (C) CmpPNode(b, a) );
2270 c = optimize_cmp_with_klass(c);
2271 do_if(btest, c);
2272 break;
2274 case Bytecodes::_ifeq: btest = BoolTest::eq; goto handle_ifxx;
2275 case Bytecodes::_ifne: btest = BoolTest::ne; goto handle_ifxx;
2276 case Bytecodes::_iflt: btest = BoolTest::lt; goto handle_ifxx;
2277 case Bytecodes::_ifle: btest = BoolTest::le; goto handle_ifxx;
2278 case Bytecodes::_ifgt: btest = BoolTest::gt; goto handle_ifxx;
2279 case Bytecodes::_ifge: btest = BoolTest::ge; goto handle_ifxx;
2280 handle_ifxx:
2281 // If this is a backwards branch in the bytecodes, add Safepoint
2282 maybe_add_safepoint(iter().get_dest());
2283 a = _gvn.intcon(0);
2284 b = pop();
2285 c = _gvn.transform( new (C) CmpINode(b, a) );
2286 do_if(btest, c);
2287 break;
2289 case Bytecodes::_if_icmpeq: btest = BoolTest::eq; goto handle_if_icmp;
2290 case Bytecodes::_if_icmpne: btest = BoolTest::ne; goto handle_if_icmp;
2291 case Bytecodes::_if_icmplt: btest = BoolTest::lt; goto handle_if_icmp;
2292 case Bytecodes::_if_icmple: btest = BoolTest::le; goto handle_if_icmp;
2293 case Bytecodes::_if_icmpgt: btest = BoolTest::gt; goto handle_if_icmp;
2294 case Bytecodes::_if_icmpge: btest = BoolTest::ge; goto handle_if_icmp;
2295 handle_if_icmp:
2296 // If this is a backwards branch in the bytecodes, add Safepoint
2297 maybe_add_safepoint(iter().get_dest());
2298 a = pop();
2299 b = pop();
2300 c = _gvn.transform( new (C) CmpINode( b, a ) );
2301 do_if(btest, c);
2302 break;
2304 case Bytecodes::_tableswitch:
2305 do_tableswitch();
2306 break;
2308 case Bytecodes::_lookupswitch:
2309 do_lookupswitch();
2310 break;
2312 case Bytecodes::_invokestatic:
2313 case Bytecodes::_invokedynamic:
2314 case Bytecodes::_invokespecial:
2315 case Bytecodes::_invokevirtual:
2316 case Bytecodes::_invokeinterface:
2317 do_call();
2318 break;
2319 case Bytecodes::_checkcast:
2320 do_checkcast();
2321 break;
2322 case Bytecodes::_instanceof:
2323 do_instanceof();
2324 break;
2325 case Bytecodes::_anewarray:
2326 do_anewarray();
2327 break;
2328 case Bytecodes::_newarray:
2329 do_newarray((BasicType)iter().get_index());
2330 break;
2331 case Bytecodes::_multianewarray:
2332 do_multianewarray();
2333 break;
2334 case Bytecodes::_new:
2335 do_new();
2336 break;
2338 case Bytecodes::_jsr:
2339 case Bytecodes::_jsr_w:
2340 do_jsr();
2341 break;
2343 case Bytecodes::_ret:
2344 do_ret();
2345 break;
2348 case Bytecodes::_monitorenter:
2349 do_monitor_enter();
2350 break;
2352 case Bytecodes::_monitorexit:
2353 do_monitor_exit();
2354 break;
2356 case Bytecodes::_breakpoint:
2357 // Breakpoint set concurrently to compile
2358 // %%% use an uncommon trap?
2359 C->record_failure("breakpoint in method");
2360 return;
2362 default:
2363 #ifndef PRODUCT
2364 map()->dump(99);
2365 #endif
2366 tty->print("\nUnhandled bytecode %s\n", Bytecodes::name(bc()) );
2367 ShouldNotReachHere();
2368 }
2370 #ifndef PRODUCT
2371 IdealGraphPrinter *printer = IdealGraphPrinter::printer();
2372 if(printer) {
2373 char buffer[256];
2374 sprintf(buffer, "Bytecode %d: %s", bci(), Bytecodes::name(bc()));
2375 bool old = printer->traverse_outs();
2376 printer->set_traverse_outs(true);
2377 printer->print_method(C, buffer, 4);
2378 printer->set_traverse_outs(old);
2379 }
2380 #endif
2381 }