1 /*
2 * Copyright (c) 1998, 2021, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "jvm_io.h"
27 #include "ci/ciMethodData.hpp"
28 #include "ci/ciSymbols.hpp"
29 #include "classfile/vmSymbols.hpp"
30 #include "compiler/compileLog.hpp"
31 #include "interpreter/linkResolver.hpp"
32 #include "memory/resourceArea.hpp"
33 #include "memory/universe.hpp"
34 #include "oops/oop.inline.hpp"
35 #include "opto/addnode.hpp"
36 #include "opto/castnode.hpp"
37 #include "opto/convertnode.hpp"
38 #include "opto/divnode.hpp"
39 #include "opto/idealGraphPrinter.hpp"
40 #include "opto/idealKit.hpp"
41 #include "opto/inlinetypenode.hpp"
42 #include "opto/matcher.hpp"
43 #include "opto/memnode.hpp"
44 #include "opto/mulnode.hpp"
45 #include "opto/opaquenode.hpp"
46 #include "opto/parse.hpp"
47 #include "opto/runtime.hpp"
48 #include "runtime/deoptimization.hpp"
49 #include "runtime/sharedRuntime.hpp"
50
51 #ifndef PRODUCT
52 extern int explicit_null_checks_inserted,
53 explicit_null_checks_elided;
54 #endif
55
56 Node* Parse::record_profile_for_speculation_at_array_load(Node* ld) {
57 // Feed unused profile data to type speculation
58 if (UseTypeSpeculation && UseArrayLoadStoreProfile) {
59 ciKlass* array_type = NULL;
60 ciKlass* element_type = NULL;
61 ProfilePtrKind element_ptr = ProfileMaybeNull;
62 bool flat_array = true;
63 bool null_free_array = true;
64 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
65 if (element_type != NULL || element_ptr != ProfileMaybeNull) {
66 ld = record_profile_for_speculation(ld, element_type, element_ptr);
67 }
68 }
69 return ld;
70 }
71
72
73 //---------------------------------array_load----------------------------------
74 void Parse::array_load(BasicType bt) {
75 const Type* elemtype = Type::TOP;
76 Node* adr = array_addressing(bt, 0, elemtype);
77 if (stopped()) return; // guaranteed null or range check
78
79 Node* idx = pop();
80 Node* ary = pop();
81
82 // Handle inline type arrays
83 const TypeOopPtr* elemptr = elemtype->make_oopptr();
84 const TypeAryPtr* ary_t = _gvn.type(ary)->is_aryptr();
85 if (ary_t->is_flat()) {
86 // Load from flattened inline type array
87 Node* vt = InlineTypeNode::make_from_flattened(this, elemtype->inline_klass(), ary, adr);
88 push(vt);
89 return;
90 } else if (ary_t->is_null_free()) {
91 // Load from non-flattened inline type array (elements can never be null)
92 bt = T_INLINE_TYPE;
93 } else if (!ary_t->is_not_flat()) {
94 // Cannot statically determine if array is flattened, emit runtime check
95 assert(UseFlatArray && is_reference_type(bt) && elemptr->can_be_inline_type() && !ary_t->klass_is_exact() && !ary_t->is_not_null_free() &&
96 (!elemptr->is_inlinetypeptr() || elemptr->inline_klass()->flatten_array()), "array can't be flattened");
97 IdealKit ideal(this);
98 IdealVariable res(ideal);
99 ideal.declarations_done();
100 ideal.if_then(flat_array_test(ary, /* flat = */ false)); {
101 // non-flattened
102 assert(ideal.ctrl()->in(0)->as_If()->is_flat_array_check(&_gvn), "Should be found");
103 sync_kit(ideal);
104 const TypeAryPtr* adr_type = TypeAryPtr::get_array_body_type(bt);
105 Node* ld = access_load_at(ary, adr, adr_type, elemptr, bt,
106 IN_HEAP | IS_ARRAY | C2_CONTROL_DEPENDENT_LOAD);
107 if (elemptr->is_inlinetypeptr()) {
108 assert(elemptr->maybe_null(), "null free array should be handled above");
109 ld = InlineTypeNode::make_from_oop(this, ld, elemptr->inline_klass(), false);
110 }
111 ideal.sync_kit(this);
112 ideal.set(res, ld);
113 } ideal.else_(); {
114 // flattened
115 sync_kit(ideal);
116 if (elemptr->is_inlinetypeptr()) {
117 // Element type is known, cast and load from flattened representation
118 ciInlineKlass* vk = elemptr->inline_klass();
119 assert(vk->flatten_array() && elemptr->maybe_null(), "never/always flat - should be optimized");
120 ciArrayKlass* array_klass = ciArrayKlass::make(vk, /* null_free */ true);
121 const TypeAryPtr* arytype = TypeOopPtr::make_from_klass(array_klass)->isa_aryptr();
122 Node* cast = _gvn.transform(new CheckCastPPNode(control(), ary, arytype));
123 Node* casted_adr = array_element_address(cast, idx, T_INLINE_TYPE, ary_t->size(), control());
124 // Re-execute flattened array load if buffering triggers deoptimization
125 PreserveReexecuteState preexecs(this);
126 jvms()->set_should_reexecute(true);
127 inc_sp(2);
128 Node* vt = InlineTypeNode::make_from_flattened(this, vk, cast, casted_adr)->buffer(this, false);
129 ideal.set(res, vt);
130 ideal.sync_kit(this);
131 } else {
132 // Element type is unknown, emit runtime call
133
134 // Below membars keep this access to an unknown flattened array correctly
135 // ordered with other unknown and known flattened array accesses.
136 insert_mem_bar_volatile(Op_MemBarCPUOrder, C->get_alias_index(TypeAryPtr::INLINES));
137
138 Node* call = NULL;
139 {
140 // Re-execute flattened array load if runtime call triggers deoptimization
141 PreserveReexecuteState preexecs(this);
142 jvms()->set_bci(_bci);
143 jvms()->set_should_reexecute(true);
144 inc_sp(2);
145 kill_dead_locals();
146 call = make_runtime_call(RC_NO_LEAF | RC_NO_IO,
147 OptoRuntime::load_unknown_inline_type(),
148 OptoRuntime::load_unknown_inline_Java(),
149 NULL, TypeRawPtr::BOTTOM,
150 ary, idx);
151 }
152 make_slow_call_ex(call, env()->Throwable_klass(), false);
153 Node* buffer = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
154
155 insert_mem_bar_volatile(Op_MemBarCPUOrder, C->get_alias_index(TypeAryPtr::INLINES));
156
157 // Keep track of the information that the inline type is flattened in arrays
158 const Type* unknown_value = elemptr->is_instptr()->cast_to_flatten_array();
159 buffer = _gvn.transform(new CheckCastPPNode(control(), buffer, unknown_value));
160
161 ideal.sync_kit(this);
162 ideal.set(res, buffer);
163 }
164 } ideal.end_if();
165 sync_kit(ideal);
166 Node* ld = _gvn.transform(ideal.value(res));
167 ld = record_profile_for_speculation_at_array_load(ld);
168 push_node(bt, ld);
169 return;
170 }
171
172 if (elemtype == TypeInt::BOOL) {
173 bt = T_BOOLEAN;
174 }
175 const TypeAryPtr* adr_type = TypeAryPtr::get_array_body_type(bt);
176 Node* ld = access_load_at(ary, adr, adr_type, elemtype, bt,
177 IN_HEAP | IS_ARRAY | C2_CONTROL_DEPENDENT_LOAD);
178 ld = record_profile_for_speculation_at_array_load(ld);
179 // Loading a non-flattened inline type
180 if (elemptr != NULL && elemptr->is_inlinetypeptr()) {
181 assert(!ary_t->is_null_free() || !elemptr->maybe_null(), "inline type array elements should never be null");
182 ld = InlineTypeNode::make_from_oop(this, ld, elemptr->inline_klass(), !elemptr->maybe_null());
183 }
184 push_node(bt, ld);
185 }
186
187
188 //--------------------------------array_store----------------------------------
189 void Parse::array_store(BasicType bt) {
190 const Type* elemtype = Type::TOP;
191 Node* adr = array_addressing(bt, type2size[bt], elemtype);
192 if (stopped()) return; // guaranteed null or range check
193 Node* cast_val = NULL;
194 if (bt == T_OBJECT) {
195 cast_val = array_store_check(adr, elemtype);
196 if (stopped()) return;
197 }
198 Node* val = pop_node(bt); // Value to store
199 Node* idx = pop(); // Index in the array
200 Node* ary = pop(); // The array itself
201
202 const TypeAryPtr* ary_t = _gvn.type(ary)->is_aryptr();
203 const TypeAryPtr* adr_type = TypeAryPtr::get_array_body_type(bt);
204 assert(adr->as_AddP()->in(AddPNode::Base) == ary, "inconsistent address base");
205
206 if (elemtype == TypeInt::BOOL) {
207 bt = T_BOOLEAN;
208 } else if (bt == T_OBJECT) {
209 elemtype = elemtype->make_oopptr();
210 const Type* tval = _gvn.type(cast_val);
211 // We may have lost type information for 'val' here due to the casts
212 // emitted by the array_store_check code (see JDK-6312651)
213 // TODO Remove this code once JDK-6312651 is in.
214 const Type* tval_init = _gvn.type(val);
215 // Based on the value to be stored, try to determine if the array is not null-free and/or not flat.
216 // This is only legal for non-null stores because the array_store_check always passes for null, even
217 // if the array is null-free. Null stores are handled in GraphKit::gen_inline_array_null_guard().
218 bool not_inline = !tval->isa_inlinetype() &&
219 ((!tval_init->maybe_null() && !tval_init->is_oopptr()->can_be_inline_type()) ||
220 (!tval->maybe_null() && !tval->is_oopptr()->can_be_inline_type()));
221 bool not_flattened = not_inline || ((tval_init->is_inlinetypeptr() || tval_init->isa_inlinetype()) && !tval_init->inline_klass()->flatten_array());
222 if (!ary_t->is_not_null_free() && not_inline) {
223 // Storing a non-inline type, mark array as not null-free (-> not flat).
224 ary_t = ary_t->cast_to_not_null_free();
225 Node* cast = _gvn.transform(new CheckCastPPNode(control(), ary, ary_t));
226 replace_in_map(ary, cast);
227 ary = cast;
228 } else if (!ary_t->is_not_flat() && not_flattened) {
229 // Storing a non-flattened value, mark array as not flat.
230 ary_t = ary_t->cast_to_not_flat();
231 Node* cast = _gvn.transform(new CheckCastPPNode(control(), ary, ary_t));
232 replace_in_map(ary, cast);
233 ary = cast;
234 }
235
236 if (ary_t->is_flat()) {
237 // Store to flattened inline type array
238 assert(!tval->maybe_null(), "should be guaranteed by array store check");
239 // Re-execute flattened array store if buffering triggers deoptimization
240 PreserveReexecuteState preexecs(this);
241 inc_sp(3);
242 jvms()->set_should_reexecute(true);
243 cast_val->as_InlineTypeBase()->store_flattened(this, ary, adr, NULL, 0, MO_UNORDERED | IN_HEAP | IS_ARRAY);
244 return;
245 } else if (ary_t->is_null_free()) {
246 // Store to non-flattened inline type array (elements can never be null)
247 assert(!tval->maybe_null(), "should be guaranteed by array store check");
248 if (elemtype->inline_klass()->is_empty()) {
249 // Ignore empty inline stores, array is already initialized.
250 return;
251 }
252 } else if (!ary_t->is_not_flat() && (tval != TypePtr::NULL_PTR || StressReflectiveCode)) {
253 // Array might be flattened, emit runtime checks (for NULL, a simple inline_array_null_guard is sufficient).
254 assert(UseFlatArray && !not_flattened && elemtype->is_oopptr()->can_be_inline_type() &&
255 !ary_t->klass_is_exact() && !ary_t->is_not_null_free(), "array can't be flattened");
256 IdealKit ideal(this);
257 ideal.if_then(flat_array_test(ary, /* flat = */ false)); {
258 // non-flattened
259 assert(ideal.ctrl()->in(0)->as_If()->is_flat_array_check(&_gvn), "Should be found");
260 sync_kit(ideal);
261 Node* cast_ary = inline_array_null_guard(ary, cast_val, 3);
262 inc_sp(3);
263 access_store_at(cast_ary, adr, adr_type, cast_val, elemtype, bt, MO_UNORDERED | IN_HEAP | IS_ARRAY, false);
264 dec_sp(3);
265 ideal.sync_kit(this);
266 } ideal.else_(); {
267 sync_kit(ideal);
268 Node* val = cast_val;
269 // flattened
270 if (!val->is_InlineType() && tval->maybe_null()) {
271 // Add null check
272 Node* null_ctl = top();
273 val = null_check_oop(val, &null_ctl);
274 if (null_ctl != top()) {
275 PreserveJVMState pjvms(this);
276 inc_sp(3);
277 set_control(null_ctl);
278 uncommon_trap(Deoptimization::Reason_null_check, Deoptimization::Action_none);
279 dec_sp(3);
280 }
281 }
282 // Try to determine the inline klass
283 ciInlineKlass* vk = NULL;
284 if (tval->isa_inlinetype() || tval->is_inlinetypeptr()) {
285 vk = tval->inline_klass();
286 } else if (tval_init->isa_inlinetype() || tval_init->is_inlinetypeptr()) {
287 vk = tval_init->inline_klass();
288 } else if (elemtype->is_inlinetypeptr()) {
289 vk = elemtype->inline_klass();
290 }
291 Node* casted_ary = ary;
292 if (vk != NULL && !stopped()) {
293 // Element type is known, cast and store to flattened representation
294 assert(vk->flatten_array() && elemtype->maybe_null(), "never/always flat - should be optimized");
295 ciArrayKlass* array_klass = ciArrayKlass::make(vk, /* null_free */ true);
296 const TypeAryPtr* arytype = TypeOopPtr::make_from_klass(array_klass)->isa_aryptr();
297 casted_ary = _gvn.transform(new CheckCastPPNode(control(), casted_ary, arytype));
298 Node* casted_adr = array_element_address(casted_ary, idx, T_OBJECT, arytype->size(), control());
299 if (!val->is_InlineType()) {
300 assert(!gvn().type(val)->maybe_null(), "inline type array elements should never be null");
301 val = InlineTypeNode::make_from_oop(this, val, vk);
302 }
303 // Re-execute flattened array store if buffering triggers deoptimization
304 PreserveReexecuteState preexecs(this);
305 inc_sp(3);
306 jvms()->set_should_reexecute(true);
307 val->as_InlineTypeBase()->store_flattened(this, casted_ary, casted_adr, NULL, 0, MO_UNORDERED | IN_HEAP | IS_ARRAY);
308 } else if (!stopped()) {
309 // Element type is unknown, emit runtime call
310
311 // Below membars keep this access to an unknown flattened array correctly
312 // ordered with other unknown and known flattened array accesses.
313 insert_mem_bar_volatile(Op_MemBarCPUOrder, C->get_alias_index(TypeAryPtr::INLINES));
314
315 make_runtime_call(RC_LEAF,
316 OptoRuntime::store_unknown_inline_type(),
317 CAST_FROM_FN_PTR(address, OptoRuntime::store_unknown_inline),
318 "store_unknown_inline", TypeRawPtr::BOTTOM,
319 val, casted_ary, idx);
320
321 insert_mem_bar_volatile(Op_MemBarCPUOrder, C->get_alias_index(TypeAryPtr::INLINES));
322 }
323 ideal.sync_kit(this);
324 }
325 ideal.end_if();
326 sync_kit(ideal);
327 return;
328 } else if (!ary_t->is_not_null_free()) {
329 // Array is not flattened but may be null free
330 assert(elemtype->is_oopptr()->can_be_inline_type() && !ary_t->klass_is_exact(), "array can't be null-free");
331 ary = inline_array_null_guard(ary, cast_val, 3, true);
332 }
333 }
334 inc_sp(3);
335 access_store_at(ary, adr, adr_type, val, elemtype, bt, MO_UNORDERED | IN_HEAP | IS_ARRAY);
336 dec_sp(3);
337 }
338
339
340 //------------------------------array_addressing-------------------------------
341 // Pull array and index from the stack. Compute pointer-to-element.
342 Node* Parse::array_addressing(BasicType type, int vals, const Type*& elemtype) {
343 Node *idx = peek(0+vals); // Get from stack without popping
344 Node *ary = peek(1+vals); // in case of exception
345
346 // Null check the array base, with correct stack contents
347 ary = null_check(ary, T_ARRAY);
348 // Compile-time detect of null-exception?
349 if (stopped()) return top();
350
351 const TypeAryPtr* arytype = _gvn.type(ary)->is_aryptr();
352 const TypeInt* sizetype = arytype->size();
353 elemtype = arytype->elem();
354
355 if (UseUniqueSubclasses) {
356 const Type* el = elemtype->make_ptr();
357 if (el && el->isa_instptr()) {
358 const TypeInstPtr* toop = el->is_instptr();
359 if (toop->klass()->as_instance_klass()->unique_concrete_subklass()) {
360 // If we load from "AbstractClass[]" we must see "ConcreteSubClass".
361 const Type* subklass = Type::get_const_type(toop->klass());
362 elemtype = subklass->join_speculative(el);
363 }
364 }
365 }
366
367 // Check for big class initializers with all constant offsets
368 // feeding into a known-size array.
369 const TypeInt* idxtype = _gvn.type(idx)->is_int();
370 // See if the highest idx value is less than the lowest array bound,
371 // and if the idx value cannot be negative:
372 bool need_range_check = true;
373 if (idxtype->_hi < sizetype->_lo && idxtype->_lo >= 0) {
374 need_range_check = false;
375 if (C->log() != NULL) C->log()->elem("observe that='!need_range_check'");
376 }
377
378 ciKlass * arytype_klass = arytype->klass();
379 if ((arytype_klass != NULL) && (!arytype_klass->is_loaded())) {
380 // Only fails for some -Xcomp runs
381 // The class is unloaded. We have to run this bytecode in the interpreter.
382 uncommon_trap(Deoptimization::Reason_unloaded,
383 Deoptimization::Action_reinterpret,
384 arytype->klass(), "!loaded array");
385 return top();
386 }
387
388 // Do the range check
389 if (GenerateRangeChecks && need_range_check) {
390 Node* tst;
391 if (sizetype->_hi <= 0) {
392 // The greatest array bound is negative, so we can conclude that we're
393 // compiling unreachable code, but the unsigned compare trick used below
394 // only works with non-negative lengths. Instead, hack "tst" to be zero so
395 // the uncommon_trap path will always be taken.
396 tst = _gvn.intcon(0);
397 } else {
398 // Range is constant in array-oop, so we can use the original state of mem
399 Node* len = load_array_length(ary);
400
401 // Test length vs index (standard trick using unsigned compare)
402 Node* chk = _gvn.transform( new CmpUNode(idx, len) );
403 BoolTest::mask btest = BoolTest::lt;
404 tst = _gvn.transform( new BoolNode(chk, btest) );
405 }
406 RangeCheckNode* rc = new RangeCheckNode(control(), tst, PROB_MAX, COUNT_UNKNOWN);
407 _gvn.set_type(rc, rc->Value(&_gvn));
408 if (!tst->is_Con()) {
409 record_for_igvn(rc);
410 }
411 set_control(_gvn.transform(new IfTrueNode(rc)));
412 // Branch to failure if out of bounds
413 {
414 PreserveJVMState pjvms(this);
415 set_control(_gvn.transform(new IfFalseNode(rc)));
416 if (C->allow_range_check_smearing()) {
417 // Do not use builtin_throw, since range checks are sometimes
418 // made more stringent by an optimistic transformation.
419 // This creates "tentative" range checks at this point,
420 // which are not guaranteed to throw exceptions.
421 // See IfNode::Ideal, is_range_check, adjust_check.
422 uncommon_trap(Deoptimization::Reason_range_check,
423 Deoptimization::Action_make_not_entrant,
424 NULL, "range_check");
425 } else {
426 // If we have already recompiled with the range-check-widening
427 // heroic optimization turned off, then we must really be throwing
428 // range check exceptions.
429 builtin_throw(Deoptimization::Reason_range_check, idx);
430 }
431 }
432 }
433 // Check for always knowing you are throwing a range-check exception
434 if (stopped()) return top();
435
436 // This could be an access to an inline type array. We can't tell if it's
437 // flat or not. Knowing the exact type avoids runtime checks and leads to
438 // a much simpler graph shape. Check profile information.
439 if (!arytype->is_flat() && !arytype->is_not_flat()) {
440 // First check the speculative type
441 Deoptimization::DeoptReason reason = Deoptimization::Reason_speculate_class_check;
442 ciKlass* array_type = arytype->speculative_type();
443 if (too_many_traps_or_recompiles(reason) || array_type == NULL) {
444 // No speculative type, check profile data at this bci
445 array_type = NULL;
446 reason = Deoptimization::Reason_class_check;
447 if (UseArrayLoadStoreProfile && !too_many_traps_or_recompiles(reason)) {
448 ciKlass* element_type = NULL;
449 ProfilePtrKind element_ptr = ProfileMaybeNull;
450 bool flat_array = true;
451 bool null_free_array = true;
452 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
453 }
454 }
455 if (array_type != NULL) {
456 // Speculate that this array has the exact type reported by profile data
457 Node* better_ary = NULL;
458 DEBUG_ONLY(Node* old_control = control();)
459 Node* slow_ctl = type_check_receiver(ary, array_type, 1.0, &better_ary);
460 if (stopped()) {
461 // The check always fails and therefore profile information is incorrect. Don't use it.
462 assert(old_control == slow_ctl, "type check should have been removed");
463 set_control(slow_ctl);
464 } else if (!slow_ctl->is_top()) {
465 { PreserveJVMState pjvms(this);
466 set_control(slow_ctl);
467 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
468 }
469 replace_in_map(ary, better_ary);
470 ary = better_ary;
471 arytype = _gvn.type(ary)->is_aryptr();
472 elemtype = arytype->elem();
473 }
474 }
475 } else if (UseTypeSpeculation && UseArrayLoadStoreProfile) {
476 // No need to speculate: feed profile data at this bci for the
477 // array to type speculation
478 ciKlass* array_type = NULL;
479 ciKlass* element_type = NULL;
480 ProfilePtrKind element_ptr = ProfileMaybeNull;
481 bool flat_array = true;
482 bool null_free_array = true;
483 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
484 if (array_type != NULL) {
485 ary = record_profile_for_speculation(ary, array_type, ProfileMaybeNull);
486 }
487 }
488
489 // We have no exact array type from profile data. Check profile data
490 // for a non null-free or non flat array. Non null-free implies non
491 // flat so check this one first. Speculating on a non null-free
492 // array doesn't help aaload but could be profitable for a
493 // subsequent aastore.
494 if (!arytype->is_null_free() && !arytype->is_not_null_free()) {
495 bool null_free_array = true;
496 Deoptimization::DeoptReason reason = Deoptimization::Reason_none;
497 if (arytype->speculative() != NULL &&
498 arytype->speculative()->is_aryptr()->is_not_null_free() &&
499 !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_class_check)) {
500 null_free_array = false;
501 reason = Deoptimization::Reason_speculate_class_check;
502 } else if (UseArrayLoadStoreProfile && !too_many_traps_or_recompiles(Deoptimization::Reason_class_check)) {
503 ciKlass* array_type = NULL;
504 ciKlass* element_type = NULL;
505 ProfilePtrKind element_ptr = ProfileMaybeNull;
506 bool flat_array = true;
507 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
508 reason = Deoptimization::Reason_class_check;
509 }
510 if (!null_free_array) {
511 { // Deoptimize if null-free array
512 BuildCutout unless(this, null_free_array_test(load_object_klass(ary), /* null_free = */ false), PROB_MAX);
513 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
514 }
515 assert(!stopped(), "null-free array should have been caught earlier");
516 Node* better_ary = _gvn.transform(new CheckCastPPNode(control(), ary, arytype->cast_to_not_null_free()));
517 replace_in_map(ary, better_ary);
518 ary = better_ary;
519 arytype = _gvn.type(ary)->is_aryptr();
520 }
521 }
522
523 if (!arytype->is_flat() && !arytype->is_not_flat()) {
524 bool flat_array = true;
525 Deoptimization::DeoptReason reason = Deoptimization::Reason_none;
526 if (arytype->speculative() != NULL &&
527 arytype->speculative()->is_aryptr()->is_not_flat() &&
528 !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_class_check)) {
529 flat_array = false;
530 reason = Deoptimization::Reason_speculate_class_check;
531 } else if (UseArrayLoadStoreProfile && !too_many_traps_or_recompiles(reason)) {
532 ciKlass* array_type = NULL;
533 ciKlass* element_type = NULL;
534 ProfilePtrKind element_ptr = ProfileMaybeNull;
535 bool null_free_array = true;
536 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
537 reason = Deoptimization::Reason_class_check;
538 }
539 if (!flat_array) {
540 { // Deoptimize if flat array
541 BuildCutout unless(this, flat_array_test(ary, /* flat = */ false), PROB_MAX);
542 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
543 }
544 assert(!stopped(), "flat array should have been caught earlier");
545 Node* better_ary = _gvn.transform(new CheckCastPPNode(control(), ary, arytype->cast_to_not_flat()));
546 replace_in_map(ary, better_ary);
547 ary = better_ary;
548 arytype = _gvn.type(ary)->is_aryptr();
549 }
550 }
551
552 // Make array address computation control dependent to prevent it
553 // from floating above the range check during loop optimizations.
554 Node* ptr = array_element_address(ary, idx, type, sizetype, control());
555 assert(ptr != top(), "top should go hand-in-hand with stopped");
556
557 return ptr;
558 }
559
560
561 // returns IfNode
562 IfNode* Parse::jump_if_fork_int(Node* a, Node* b, BoolTest::mask mask, float prob, float cnt) {
563 Node *cmp = _gvn.transform(new CmpINode(a, b)); // two cases: shiftcount > 32 and shiftcount <= 32
564 Node *tst = _gvn.transform(new BoolNode(cmp, mask));
565 IfNode *iff = create_and_map_if(control(), tst, prob, cnt);
566 return iff;
567 }
568
569
570 // sentinel value for the target bci to mark never taken branches
571 // (according to profiling)
572 static const int never_reached = INT_MAX;
573
574 //------------------------------helper for tableswitch-------------------------
575 void Parse::jump_if_true_fork(IfNode *iff, int dest_bci_if_true, bool unc) {
576 // True branch, use existing map info
577 { PreserveJVMState pjvms(this);
578 Node *iftrue = _gvn.transform( new IfTrueNode (iff) );
579 set_control( iftrue );
580 if (unc) {
581 repush_if_args();
582 uncommon_trap(Deoptimization::Reason_unstable_if,
583 Deoptimization::Action_reinterpret,
584 NULL,
585 "taken always");
586 } else {
587 assert(dest_bci_if_true != never_reached, "inconsistent dest");
588 merge_new_path(dest_bci_if_true);
589 }
590 }
591
592 // False branch
593 Node *iffalse = _gvn.transform( new IfFalseNode(iff) );
594 set_control( iffalse );
595 }
596
597 void Parse::jump_if_false_fork(IfNode *iff, int dest_bci_if_true, bool unc) {
598 // True branch, use existing map info
599 { PreserveJVMState pjvms(this);
600 Node *iffalse = _gvn.transform( new IfFalseNode (iff) );
601 set_control( iffalse );
602 if (unc) {
603 repush_if_args();
604 uncommon_trap(Deoptimization::Reason_unstable_if,
605 Deoptimization::Action_reinterpret,
606 NULL,
607 "taken never");
608 } else {
609 assert(dest_bci_if_true != never_reached, "inconsistent dest");
610 merge_new_path(dest_bci_if_true);
611 }
612 }
613
614 // False branch
615 Node *iftrue = _gvn.transform( new IfTrueNode(iff) );
616 set_control( iftrue );
617 }
618
619 void Parse::jump_if_always_fork(int dest_bci, bool unc) {
620 // False branch, use existing map and control()
621 if (unc) {
622 repush_if_args();
623 uncommon_trap(Deoptimization::Reason_unstable_if,
624 Deoptimization::Action_reinterpret,
625 NULL,
626 "taken never");
627 } else {
628 assert(dest_bci != never_reached, "inconsistent dest");
629 merge_new_path(dest_bci);
630 }
631 }
632
633
634 extern "C" {
635 static int jint_cmp(const void *i, const void *j) {
636 int a = *(jint *)i;
637 int b = *(jint *)j;
638 return a > b ? 1 : a < b ? -1 : 0;
639 }
640 }
641
642
643 class SwitchRange : public StackObj {
644 // a range of integers coupled with a bci destination
645 jint _lo; // inclusive lower limit
646 jint _hi; // inclusive upper limit
647 int _dest;
648 float _cnt; // how many times this range was hit according to profiling
649
650 public:
651 jint lo() const { return _lo; }
652 jint hi() const { return _hi; }
653 int dest() const { return _dest; }
654 bool is_singleton() const { return _lo == _hi; }
655 float cnt() const { return _cnt; }
656
657 void setRange(jint lo, jint hi, int dest, float cnt) {
658 assert(lo <= hi, "must be a non-empty range");
659 _lo = lo, _hi = hi; _dest = dest; _cnt = cnt;
660 assert(_cnt >= 0, "");
661 }
662 bool adjoinRange(jint lo, jint hi, int dest, float cnt, bool trim_ranges) {
663 assert(lo <= hi, "must be a non-empty range");
664 if (lo == _hi+1) {
665 // see merge_ranges() comment below
666 if (trim_ranges) {
667 if (cnt == 0) {
668 if (_cnt != 0) {
669 return false;
670 }
671 if (dest != _dest) {
672 _dest = never_reached;
673 }
674 } else {
675 if (_cnt == 0) {
676 return false;
677 }
678 if (dest != _dest) {
679 return false;
680 }
681 }
682 } else {
683 if (dest != _dest) {
684 return false;
685 }
686 }
687 _hi = hi;
688 _cnt += cnt;
689 return true;
690 }
691 return false;
692 }
693
694 void set (jint value, int dest, float cnt) {
695 setRange(value, value, dest, cnt);
696 }
697 bool adjoin(jint value, int dest, float cnt, bool trim_ranges) {
698 return adjoinRange(value, value, dest, cnt, trim_ranges);
699 }
700 bool adjoin(SwitchRange& other) {
701 return adjoinRange(other._lo, other._hi, other._dest, other._cnt, false);
702 }
703
704 void print() {
705 if (is_singleton())
706 tty->print(" {%d}=>%d (cnt=%f)", lo(), dest(), cnt());
707 else if (lo() == min_jint)
708 tty->print(" {..%d}=>%d (cnt=%f)", hi(), dest(), cnt());
709 else if (hi() == max_jint)
710 tty->print(" {%d..}=>%d (cnt=%f)", lo(), dest(), cnt());
711 else
712 tty->print(" {%d..%d}=>%d (cnt=%f)", lo(), hi(), dest(), cnt());
713 }
714 };
715
716 // We try to minimize the number of ranges and the size of the taken
717 // ones using profiling data. When ranges are created,
718 // SwitchRange::adjoinRange() only allows 2 adjoining ranges to merge
719 // if both were never hit or both were hit to build longer unreached
720 // ranges. Here, we now merge adjoining ranges with the same
721 // destination and finally set destination of unreached ranges to the
722 // special value never_reached because it can help minimize the number
723 // of tests that are necessary.
724 //
725 // For instance:
726 // [0, 1] to target1 sometimes taken
727 // [1, 2] to target1 never taken
728 // [2, 3] to target2 never taken
729 // would lead to:
730 // [0, 1] to target1 sometimes taken
731 // [1, 3] never taken
732 //
733 // (first 2 ranges to target1 are not merged)
734 static void merge_ranges(SwitchRange* ranges, int& rp) {
735 if (rp == 0) {
736 return;
737 }
738 int shift = 0;
739 for (int j = 0; j < rp; j++) {
740 SwitchRange& r1 = ranges[j-shift];
741 SwitchRange& r2 = ranges[j+1];
742 if (r1.adjoin(r2)) {
743 shift++;
744 } else if (shift > 0) {
745 ranges[j+1-shift] = r2;
746 }
747 }
748 rp -= shift;
749 for (int j = 0; j <= rp; j++) {
750 SwitchRange& r = ranges[j];
751 if (r.cnt() == 0 && r.dest() != never_reached) {
752 r.setRange(r.lo(), r.hi(), never_reached, r.cnt());
753 }
754 }
755 }
756
757 //-------------------------------do_tableswitch--------------------------------
758 void Parse::do_tableswitch() {
759 // Get information about tableswitch
760 int default_dest = iter().get_dest_table(0);
761 int lo_index = iter().get_int_table(1);
762 int hi_index = iter().get_int_table(2);
763 int len = hi_index - lo_index + 1;
764
765 if (len < 1) {
766 // If this is a backward branch, add safepoint
767 maybe_add_safepoint(default_dest);
768 pop(); // the effect of the instruction execution on the operand stack
769 merge(default_dest);
770 return;
771 }
772
773 ciMethodData* methodData = method()->method_data();
774 ciMultiBranchData* profile = NULL;
775 if (methodData->is_mature() && UseSwitchProfiling) {
776 ciProfileData* data = methodData->bci_to_data(bci());
777 if (data != NULL && data->is_MultiBranchData()) {
778 profile = (ciMultiBranchData*)data;
779 }
780 }
781 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if);
782
783 // generate decision tree, using trichotomy when possible
784 int rnum = len+2;
785 bool makes_backward_branch = false;
786 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum);
787 int rp = -1;
788 if (lo_index != min_jint) {
789 uint cnt = 1;
790 if (profile != NULL) {
791 cnt = profile->default_count() / (hi_index != max_jint ? 2 : 1);
792 }
793 ranges[++rp].setRange(min_jint, lo_index-1, default_dest, cnt);
794 }
795 for (int j = 0; j < len; j++) {
796 jint match_int = lo_index+j;
797 int dest = iter().get_dest_table(j+3);
798 makes_backward_branch |= (dest <= bci());
799 uint cnt = 1;
800 if (profile != NULL) {
801 cnt = profile->count_at(j);
802 }
803 if (rp < 0 || !ranges[rp].adjoin(match_int, dest, cnt, trim_ranges)) {
804 ranges[++rp].set(match_int, dest, cnt);
805 }
806 }
807 jint highest = lo_index+(len-1);
808 assert(ranges[rp].hi() == highest, "");
809 if (highest != max_jint) {
810 uint cnt = 1;
811 if (profile != NULL) {
812 cnt = profile->default_count() / (lo_index != min_jint ? 2 : 1);
813 }
814 if (!ranges[rp].adjoinRange(highest+1, max_jint, default_dest, cnt, trim_ranges)) {
815 ranges[++rp].setRange(highest+1, max_jint, default_dest, cnt);
816 }
817 }
818 assert(rp < len+2, "not too many ranges");
819
820 if (trim_ranges) {
821 merge_ranges(ranges, rp);
822 }
823
824 // Safepoint in case if backward branch observed
825 if (makes_backward_branch) {
826 add_safepoint();
827 }
828
829 Node* lookup = pop(); // lookup value
830 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]);
831 }
832
833
834 //------------------------------do_lookupswitch--------------------------------
835 void Parse::do_lookupswitch() {
836 // Get information about lookupswitch
837 int default_dest = iter().get_dest_table(0);
838 int len = iter().get_int_table(1);
839
840 if (len < 1) { // If this is a backward branch, add safepoint
841 maybe_add_safepoint(default_dest);
842 pop(); // the effect of the instruction execution on the operand stack
843 merge(default_dest);
844 return;
845 }
846
847 ciMethodData* methodData = method()->method_data();
848 ciMultiBranchData* profile = NULL;
849 if (methodData->is_mature() && UseSwitchProfiling) {
850 ciProfileData* data = methodData->bci_to_data(bci());
851 if (data != NULL && data->is_MultiBranchData()) {
852 profile = (ciMultiBranchData*)data;
853 }
854 }
855 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if);
856
857 // generate decision tree, using trichotomy when possible
858 jint* table = NEW_RESOURCE_ARRAY(jint, len*3);
859 {
860 for (int j = 0; j < len; j++) {
861 table[3*j+0] = iter().get_int_table(2+2*j);
862 table[3*j+1] = iter().get_dest_table(2+2*j+1);
863 // Handle overflow when converting from uint to jint
864 table[3*j+2] = (profile == NULL) ? 1 : MIN2<uint>(max_jint, profile->count_at(j));
865 }
866 qsort(table, len, 3*sizeof(table[0]), jint_cmp);
867 }
868
869 float defaults = 0;
870 jint prev = min_jint;
871 for (int j = 0; j < len; j++) {
872 jint match_int = table[3*j+0];
873 if (match_int != prev) {
874 defaults += (float)match_int - prev;
875 }
876 prev = match_int+1;
877 }
878 if (prev != min_jint) {
879 defaults += (float)max_jint - prev + 1;
880 }
881 float default_cnt = 1;
882 if (profile != NULL) {
883 default_cnt = profile->default_count()/defaults;
884 }
885
886 int rnum = len*2+1;
887 bool makes_backward_branch = false;
888 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum);
889 int rp = -1;
890 for (int j = 0; j < len; j++) {
891 jint match_int = table[3*j+0];
892 int dest = table[3*j+1];
893 int cnt = table[3*j+2];
894 int next_lo = rp < 0 ? min_jint : ranges[rp].hi()+1;
895 makes_backward_branch |= (dest <= bci());
896 float c = default_cnt * ((float)match_int - next_lo);
897 if (match_int != next_lo && (rp < 0 || !ranges[rp].adjoinRange(next_lo, match_int-1, default_dest, c, trim_ranges))) {
898 assert(default_dest != never_reached, "sentinel value for dead destinations");
899 ranges[++rp].setRange(next_lo, match_int-1, default_dest, c);
900 }
901 if (rp < 0 || !ranges[rp].adjoin(match_int, dest, cnt, trim_ranges)) {
902 assert(dest != never_reached, "sentinel value for dead destinations");
903 ranges[++rp].set(match_int, dest, cnt);
904 }
905 }
906 jint highest = table[3*(len-1)];
907 assert(ranges[rp].hi() == highest, "");
908 if (highest != max_jint &&
909 !ranges[rp].adjoinRange(highest+1, max_jint, default_dest, default_cnt * ((float)max_jint - highest), trim_ranges)) {
910 ranges[++rp].setRange(highest+1, max_jint, default_dest, default_cnt * ((float)max_jint - highest));
911 }
912 assert(rp < rnum, "not too many ranges");
913
914 if (trim_ranges) {
915 merge_ranges(ranges, rp);
916 }
917
918 // Safepoint in case backward branch observed
919 if (makes_backward_branch) {
920 add_safepoint();
921 }
922
923 Node *lookup = pop(); // lookup value
924 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]);
925 }
926
927 static float if_prob(float taken_cnt, float total_cnt) {
928 assert(taken_cnt <= total_cnt, "");
929 if (total_cnt == 0) {
930 return PROB_FAIR;
931 }
932 float p = taken_cnt / total_cnt;
933 return clamp(p, PROB_MIN, PROB_MAX);
934 }
935
936 static float if_cnt(float cnt) {
937 if (cnt == 0) {
938 return COUNT_UNKNOWN;
939 }
940 return cnt;
941 }
942
943 static float sum_of_cnts(SwitchRange *lo, SwitchRange *hi) {
944 float total_cnt = 0;
945 for (SwitchRange* sr = lo; sr <= hi; sr++) {
946 total_cnt += sr->cnt();
947 }
948 return total_cnt;
949 }
950
951 class SwitchRanges : public ResourceObj {
952 public:
953 SwitchRange* _lo;
954 SwitchRange* _hi;
955 SwitchRange* _mid;
956 float _cost;
957
958 enum {
959 Start,
960 LeftDone,
961 RightDone,
962 Done
963 } _state;
964
965 SwitchRanges(SwitchRange *lo, SwitchRange *hi)
966 : _lo(lo), _hi(hi), _mid(NULL),
967 _cost(0), _state(Start) {
968 }
969
970 SwitchRanges()
971 : _lo(NULL), _hi(NULL), _mid(NULL),
972 _cost(0), _state(Start) {}
973 };
974
975 // Estimate cost of performing a binary search on lo..hi
976 static float compute_tree_cost(SwitchRange *lo, SwitchRange *hi, float total_cnt) {
977 GrowableArray<SwitchRanges> tree;
978 SwitchRanges root(lo, hi);
979 tree.push(root);
980
981 float cost = 0;
982 do {
983 SwitchRanges& r = *tree.adr_at(tree.length()-1);
984 if (r._hi != r._lo) {
985 if (r._mid == NULL) {
986 float r_cnt = sum_of_cnts(r._lo, r._hi);
987
988 if (r_cnt == 0) {
989 tree.pop();
990 cost = 0;
991 continue;
992 }
993
994 SwitchRange* mid = NULL;
995 mid = r._lo;
996 for (float cnt = 0; ; ) {
997 assert(mid <= r._hi, "out of bounds");
998 cnt += mid->cnt();
999 if (cnt > r_cnt / 2) {
1000 break;
1001 }
1002 mid++;
1003 }
1004 assert(mid <= r._hi, "out of bounds");
1005 r._mid = mid;
1006 r._cost = r_cnt / total_cnt;
1007 }
1008 r._cost += cost;
1009 if (r._state < SwitchRanges::LeftDone && r._mid > r._lo) {
1010 cost = 0;
1011 r._state = SwitchRanges::LeftDone;
1012 tree.push(SwitchRanges(r._lo, r._mid-1));
1013 } else if (r._state < SwitchRanges::RightDone) {
1014 cost = 0;
1015 r._state = SwitchRanges::RightDone;
1016 tree.push(SwitchRanges(r._mid == r._lo ? r._mid+1 : r._mid, r._hi));
1017 } else {
1018 tree.pop();
1019 cost = r._cost;
1020 }
1021 } else {
1022 tree.pop();
1023 cost = r._cost;
1024 }
1025 } while (tree.length() > 0);
1026
1027
1028 return cost;
1029 }
1030
1031 // It sometimes pays off to test most common ranges before the binary search
1032 void Parse::linear_search_switch_ranges(Node* key_val, SwitchRange*& lo, SwitchRange*& hi) {
1033 uint nr = hi - lo + 1;
1034 float total_cnt = sum_of_cnts(lo, hi);
1035
1036 float min = compute_tree_cost(lo, hi, total_cnt);
1037 float extra = 1;
1038 float sub = 0;
1039
1040 SwitchRange* array1 = lo;
1041 SwitchRange* array2 = NEW_RESOURCE_ARRAY(SwitchRange, nr);
1042
1043 SwitchRange* ranges = NULL;
1044
1045 while (nr >= 2) {
1046 assert(lo == array1 || lo == array2, "one the 2 already allocated arrays");
1047 ranges = (lo == array1) ? array2 : array1;
1048
1049 // Find highest frequency range
1050 SwitchRange* candidate = lo;
1051 for (SwitchRange* sr = lo+1; sr <= hi; sr++) {
1052 if (sr->cnt() > candidate->cnt()) {
1053 candidate = sr;
1054 }
1055 }
1056 SwitchRange most_freq = *candidate;
1057 if (most_freq.cnt() == 0) {
1058 break;
1059 }
1060
1061 // Copy remaining ranges into another array
1062 int shift = 0;
1063 for (uint i = 0; i < nr; i++) {
1064 SwitchRange* sr = &lo[i];
1065 if (sr != candidate) {
1066 ranges[i-shift] = *sr;
1067 } else {
1068 shift++;
1069 if (i > 0 && i < nr-1) {
1070 SwitchRange prev = lo[i-1];
1071 prev.setRange(prev.lo(), sr->hi(), prev.dest(), prev.cnt());
1072 if (prev.adjoin(lo[i+1])) {
1073 shift++;
1074 i++;
1075 }
1076 ranges[i-shift] = prev;
1077 }
1078 }
1079 }
1080 nr -= shift;
1081
1082 // Evaluate cost of testing the most common range and performing a
1083 // binary search on the other ranges
1084 float cost = extra + compute_tree_cost(&ranges[0], &ranges[nr-1], total_cnt);
1085 if (cost >= min) {
1086 break;
1087 }
1088 // swap arrays
1089 lo = &ranges[0];
1090 hi = &ranges[nr-1];
1091
1092 // It pays off: emit the test for the most common range
1093 assert(most_freq.cnt() > 0, "must be taken");
1094 Node* val = _gvn.transform(new SubINode(key_val, _gvn.intcon(most_freq.lo())));
1095 Node* cmp = _gvn.transform(new CmpUNode(val, _gvn.intcon(most_freq.hi() - most_freq.lo())));
1096 Node* tst = _gvn.transform(new BoolNode(cmp, BoolTest::le));
1097 IfNode* iff = create_and_map_if(control(), tst, if_prob(most_freq.cnt(), total_cnt), if_cnt(most_freq.cnt()));
1098 jump_if_true_fork(iff, most_freq.dest(), false);
1099
1100 sub += most_freq.cnt() / total_cnt;
1101 extra += 1 - sub;
1102 min = cost;
1103 }
1104 }
1105
1106 //----------------------------create_jump_tables-------------------------------
1107 bool Parse::create_jump_tables(Node* key_val, SwitchRange* lo, SwitchRange* hi) {
1108 // Are jumptables enabled
1109 if (!UseJumpTables) return false;
1110
1111 // Are jumptables supported
1112 if (!Matcher::has_match_rule(Op_Jump)) return false;
1113
1114 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if);
1115
1116 // Decide if a guard is needed to lop off big ranges at either (or
1117 // both) end(s) of the input set. We'll call this the default target
1118 // even though we can't be sure that it is the true "default".
1119
1120 bool needs_guard = false;
1121 int default_dest;
1122 int64_t total_outlier_size = 0;
1123 int64_t hi_size = ((int64_t)hi->hi()) - ((int64_t)hi->lo()) + 1;
1124 int64_t lo_size = ((int64_t)lo->hi()) - ((int64_t)lo->lo()) + 1;
1125
1126 if (lo->dest() == hi->dest()) {
1127 total_outlier_size = hi_size + lo_size;
1128 default_dest = lo->dest();
1129 } else if (lo_size > hi_size) {
1130 total_outlier_size = lo_size;
1131 default_dest = lo->dest();
1132 } else {
1133 total_outlier_size = hi_size;
1134 default_dest = hi->dest();
1135 }
1136
1137 float total = sum_of_cnts(lo, hi);
1138 float cost = compute_tree_cost(lo, hi, total);
1139
1140 // If a guard test will eliminate very sparse end ranges, then
1141 // it is worth the cost of an extra jump.
1142 float trimmed_cnt = 0;
1143 if (total_outlier_size > (MaxJumpTableSparseness * 4)) {
1144 needs_guard = true;
1145 if (default_dest == lo->dest()) {
1146 trimmed_cnt += lo->cnt();
1147 lo++;
1148 }
1149 if (default_dest == hi->dest()) {
1150 trimmed_cnt += hi->cnt();
1151 hi--;
1152 }
1153 }
1154
1155 // Find the total number of cases and ranges
1156 int64_t num_cases = ((int64_t)hi->hi()) - ((int64_t)lo->lo()) + 1;
1157 int num_range = hi - lo + 1;
1158
1159 // Don't create table if: too large, too small, or too sparse.
1160 if (num_cases > MaxJumpTableSize)
1161 return false;
1162 if (UseSwitchProfiling) {
1163 // MinJumpTableSize is set so with a well balanced binary tree,
1164 // when the number of ranges is MinJumpTableSize, it's cheaper to
1165 // go through a JumpNode that a tree of IfNodes. Average cost of a
1166 // tree of IfNodes with MinJumpTableSize is
1167 // log2f(MinJumpTableSize) comparisons. So if the cost computed
1168 // from profile data is less than log2f(MinJumpTableSize) then
1169 // going with the binary search is cheaper.
1170 if (cost < log2f(MinJumpTableSize)) {
1171 return false;
1172 }
1173 } else {
1174 if (num_cases < MinJumpTableSize)
1175 return false;
1176 }
1177 if (num_cases > (MaxJumpTableSparseness * num_range))
1178 return false;
1179
1180 // Normalize table lookups to zero
1181 int lowval = lo->lo();
1182 key_val = _gvn.transform( new SubINode(key_val, _gvn.intcon(lowval)) );
1183
1184 // Generate a guard to protect against input keyvals that aren't
1185 // in the switch domain.
1186 if (needs_guard) {
1187 Node* size = _gvn.intcon(num_cases);
1188 Node* cmp = _gvn.transform(new CmpUNode(key_val, size));
1189 Node* tst = _gvn.transform(new BoolNode(cmp, BoolTest::ge));
1190 IfNode* iff = create_and_map_if(control(), tst, if_prob(trimmed_cnt, total), if_cnt(trimmed_cnt));
1191 jump_if_true_fork(iff, default_dest, trim_ranges && trimmed_cnt == 0);
1192
1193 total -= trimmed_cnt;
1194 }
1195
1196 // Create an ideal node JumpTable that has projections
1197 // of all possible ranges for a switch statement
1198 // The key_val input must be converted to a pointer offset and scaled.
1199 // Compare Parse::array_addressing above.
1200
1201 // Clean the 32-bit int into a real 64-bit offset.
1202 // Otherwise, the jint value 0 might turn into an offset of 0x0800000000.
1203 // Make I2L conversion control dependent to prevent it from
1204 // floating above the range check during loop optimizations.
1205 // Do not use a narrow int type here to prevent the data path from dying
1206 // while the control path is not removed. This can happen if the type of key_val
1207 // is later known to be out of bounds of [0, num_cases] and therefore a narrow cast
1208 // would be replaced by TOP while C2 is not able to fold the corresponding range checks.
1209 // Set _carry_dependency for the cast to avoid being removed by IGVN.
1210 #ifdef _LP64
1211 key_val = C->constrained_convI2L(&_gvn, key_val, TypeInt::INT, control(), true /* carry_dependency */);
1212 #endif
1213
1214 // Shift the value by wordsize so we have an index into the table, rather
1215 // than a switch value
1216 Node *shiftWord = _gvn.MakeConX(wordSize);
1217 key_val = _gvn.transform( new MulXNode( key_val, shiftWord));
1218
1219 // Create the JumpNode
1220 Arena* arena = C->comp_arena();
1221 float* probs = (float*)arena->Amalloc(sizeof(float)*num_cases);
1222 int i = 0;
1223 if (total == 0) {
1224 for (SwitchRange* r = lo; r <= hi; r++) {
1225 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) {
1226 probs[i] = 1.0F / num_cases;
1227 }
1228 }
1229 } else {
1230 for (SwitchRange* r = lo; r <= hi; r++) {
1231 float prob = r->cnt()/total;
1232 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) {
1233 probs[i] = prob / (r->hi() - r->lo() + 1);
1234 }
1235 }
1236 }
1237
1238 ciMethodData* methodData = method()->method_data();
1239 ciMultiBranchData* profile = NULL;
1240 if (methodData->is_mature()) {
1241 ciProfileData* data = methodData->bci_to_data(bci());
1242 if (data != NULL && data->is_MultiBranchData()) {
1243 profile = (ciMultiBranchData*)data;
1244 }
1245 }
1246
1247 Node* jtn = _gvn.transform(new JumpNode(control(), key_val, num_cases, probs, profile == NULL ? COUNT_UNKNOWN : total));
1248
1249 // These are the switch destinations hanging off the jumpnode
1250 i = 0;
1251 for (SwitchRange* r = lo; r <= hi; r++) {
1252 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) {
1253 Node* input = _gvn.transform(new JumpProjNode(jtn, i, r->dest(), (int)(j - lowval)));
1254 {
1255 PreserveJVMState pjvms(this);
1256 set_control(input);
1257 jump_if_always_fork(r->dest(), trim_ranges && r->cnt() == 0);
1258 }
1259 }
1260 }
1261 assert(i == num_cases, "miscount of cases");
1262 stop_and_kill_map(); // no more uses for this JVMS
1263 return true;
1264 }
1265
1266 //----------------------------jump_switch_ranges-------------------------------
1267 void Parse::jump_switch_ranges(Node* key_val, SwitchRange *lo, SwitchRange *hi, int switch_depth) {
1268 Block* switch_block = block();
1269 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if);
1270
1271 if (switch_depth == 0) {
1272 // Do special processing for the top-level call.
1273 assert(lo->lo() == min_jint, "initial range must exhaust Type::INT");
1274 assert(hi->hi() == max_jint, "initial range must exhaust Type::INT");
1275
1276 // Decrement pred-numbers for the unique set of nodes.
1277 #ifdef ASSERT
1278 if (!trim_ranges) {
1279 // Ensure that the block's successors are a (duplicate-free) set.
1280 int successors_counted = 0; // block occurrences in [hi..lo]
1281 int unique_successors = switch_block->num_successors();
1282 for (int i = 0; i < unique_successors; i++) {
1283 Block* target = switch_block->successor_at(i);
1284
1285 // Check that the set of successors is the same in both places.
1286 int successors_found = 0;
1287 for (SwitchRange* p = lo; p <= hi; p++) {
1288 if (p->dest() == target->start()) successors_found++;
1289 }
1290 assert(successors_found > 0, "successor must be known");
1291 successors_counted += successors_found;
1292 }
1293 assert(successors_counted == (hi-lo)+1, "no unexpected successors");
1294 }
1295 #endif
1296
1297 // Maybe prune the inputs, based on the type of key_val.
1298 jint min_val = min_jint;
1299 jint max_val = max_jint;
1300 const TypeInt* ti = key_val->bottom_type()->isa_int();
1301 if (ti != NULL) {
1302 min_val = ti->_lo;
1303 max_val = ti->_hi;
1304 assert(min_val <= max_val, "invalid int type");
1305 }
1306 while (lo->hi() < min_val) {
1307 lo++;
1308 }
1309 if (lo->lo() < min_val) {
1310 lo->setRange(min_val, lo->hi(), lo->dest(), lo->cnt());
1311 }
1312 while (hi->lo() > max_val) {
1313 hi--;
1314 }
1315 if (hi->hi() > max_val) {
1316 hi->setRange(hi->lo(), max_val, hi->dest(), hi->cnt());
1317 }
1318
1319 linear_search_switch_ranges(key_val, lo, hi);
1320 }
1321
1322 #ifndef PRODUCT
1323 if (switch_depth == 0) {
1324 _max_switch_depth = 0;
1325 _est_switch_depth = log2i_graceful((hi - lo + 1) - 1) + 1;
1326 }
1327 #endif
1328
1329 assert(lo <= hi, "must be a non-empty set of ranges");
1330 if (lo == hi) {
1331 jump_if_always_fork(lo->dest(), trim_ranges && lo->cnt() == 0);
1332 } else {
1333 assert(lo->hi() == (lo+1)->lo()-1, "contiguous ranges");
1334 assert(hi->lo() == (hi-1)->hi()+1, "contiguous ranges");
1335
1336 if (create_jump_tables(key_val, lo, hi)) return;
1337
1338 SwitchRange* mid = NULL;
1339 float total_cnt = sum_of_cnts(lo, hi);
1340
1341 int nr = hi - lo + 1;
1342 if (UseSwitchProfiling) {
1343 // Don't keep the binary search tree balanced: pick up mid point
1344 // that split frequencies in half.
1345 float cnt = 0;
1346 for (SwitchRange* sr = lo; sr <= hi; sr++) {
1347 cnt += sr->cnt();
1348 if (cnt >= total_cnt / 2) {
1349 mid = sr;
1350 break;
1351 }
1352 }
1353 } else {
1354 mid = lo + nr/2;
1355
1356 // if there is an easy choice, pivot at a singleton:
1357 if (nr > 3 && !mid->is_singleton() && (mid-1)->is_singleton()) mid--;
1358
1359 assert(lo < mid && mid <= hi, "good pivot choice");
1360 assert(nr != 2 || mid == hi, "should pick higher of 2");
1361 assert(nr != 3 || mid == hi-1, "should pick middle of 3");
1362 }
1363
1364
1365 Node *test_val = _gvn.intcon(mid == lo ? mid->hi() : mid->lo());
1366
1367 if (mid->is_singleton()) {
1368 IfNode *iff_ne = jump_if_fork_int(key_val, test_val, BoolTest::ne, 1-if_prob(mid->cnt(), total_cnt), if_cnt(mid->cnt()));
1369 jump_if_false_fork(iff_ne, mid->dest(), trim_ranges && mid->cnt() == 0);
1370
1371 // Special Case: If there are exactly three ranges, and the high
1372 // and low range each go to the same place, omit the "gt" test,
1373 // since it will not discriminate anything.
1374 bool eq_test_only = (hi == lo+2 && hi->dest() == lo->dest() && mid == hi-1) || mid == lo;
1375
1376 // if there is a higher range, test for it and process it:
1377 if (mid < hi && !eq_test_only) {
1378 // two comparisons of same values--should enable 1 test for 2 branches
1379 // Use BoolTest::lt instead of BoolTest::gt
1380 float cnt = sum_of_cnts(lo, mid-1);
1381 IfNode *iff_lt = jump_if_fork_int(key_val, test_val, BoolTest::lt, if_prob(cnt, total_cnt), if_cnt(cnt));
1382 Node *iftrue = _gvn.transform( new IfTrueNode(iff_lt) );
1383 Node *iffalse = _gvn.transform( new IfFalseNode(iff_lt) );
1384 { PreserveJVMState pjvms(this);
1385 set_control(iffalse);
1386 jump_switch_ranges(key_val, mid+1, hi, switch_depth+1);
1387 }
1388 set_control(iftrue);
1389 }
1390
1391 } else {
1392 // mid is a range, not a singleton, so treat mid..hi as a unit
1393 float cnt = sum_of_cnts(mid == lo ? mid+1 : mid, hi);
1394 IfNode *iff_ge = jump_if_fork_int(key_val, test_val, mid == lo ? BoolTest::gt : BoolTest::ge, if_prob(cnt, total_cnt), if_cnt(cnt));
1395
1396 // if there is a higher range, test for it and process it:
1397 if (mid == hi) {
1398 jump_if_true_fork(iff_ge, mid->dest(), trim_ranges && cnt == 0);
1399 } else {
1400 Node *iftrue = _gvn.transform( new IfTrueNode(iff_ge) );
1401 Node *iffalse = _gvn.transform( new IfFalseNode(iff_ge) );
1402 { PreserveJVMState pjvms(this);
1403 set_control(iftrue);
1404 jump_switch_ranges(key_val, mid == lo ? mid+1 : mid, hi, switch_depth+1);
1405 }
1406 set_control(iffalse);
1407 }
1408 }
1409
1410 // in any case, process the lower range
1411 if (mid == lo) {
1412 if (mid->is_singleton()) {
1413 jump_switch_ranges(key_val, lo+1, hi, switch_depth+1);
1414 } else {
1415 jump_if_always_fork(lo->dest(), trim_ranges && lo->cnt() == 0);
1416 }
1417 } else {
1418 jump_switch_ranges(key_val, lo, mid-1, switch_depth+1);
1419 }
1420 }
1421
1422 // Decrease pred_count for each successor after all is done.
1423 if (switch_depth == 0) {
1424 int unique_successors = switch_block->num_successors();
1425 for (int i = 0; i < unique_successors; i++) {
1426 Block* target = switch_block->successor_at(i);
1427 // Throw away the pre-allocated path for each unique successor.
1428 target->next_path_num();
1429 }
1430 }
1431
1432 #ifndef PRODUCT
1433 _max_switch_depth = MAX2(switch_depth, _max_switch_depth);
1434 if (TraceOptoParse && Verbose && WizardMode && switch_depth == 0) {
1435 SwitchRange* r;
1436 int nsing = 0;
1437 for( r = lo; r <= hi; r++ ) {
1438 if( r->is_singleton() ) nsing++;
1439 }
1440 tty->print(">>> ");
1441 _method->print_short_name();
1442 tty->print_cr(" switch decision tree");
1443 tty->print_cr(" %d ranges (%d singletons), max_depth=%d, est_depth=%d",
1444 (int) (hi-lo+1), nsing, _max_switch_depth, _est_switch_depth);
1445 if (_max_switch_depth > _est_switch_depth) {
1446 tty->print_cr("******** BAD SWITCH DEPTH ********");
1447 }
1448 tty->print(" ");
1449 for( r = lo; r <= hi; r++ ) {
1450 r->print();
1451 }
1452 tty->cr();
1453 }
1454 #endif
1455 }
1456
1457 void Parse::modf() {
1458 Node *f2 = pop();
1459 Node *f1 = pop();
1460 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::modf_Type(),
1461 CAST_FROM_FN_PTR(address, SharedRuntime::frem),
1462 "frem", NULL, //no memory effects
1463 f1, f2);
1464 Node* res = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0));
1465
1466 push(res);
1467 }
1468
1469 void Parse::modd() {
1470 Node *d2 = pop_pair();
1471 Node *d1 = pop_pair();
1472 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::Math_DD_D_Type(),
1473 CAST_FROM_FN_PTR(address, SharedRuntime::drem),
1474 "drem", NULL, //no memory effects
1475 d1, top(), d2, top());
1476 Node* res_d = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0));
1477
1478 #ifdef ASSERT
1479 Node* res_top = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 1));
1480 assert(res_top == top(), "second value must be top");
1481 #endif
1482
1483 push_pair(res_d);
1484 }
1485
1486 void Parse::l2f() {
1487 Node* f2 = pop();
1488 Node* f1 = pop();
1489 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::l2f_Type(),
1490 CAST_FROM_FN_PTR(address, SharedRuntime::l2f),
1491 "l2f", NULL, //no memory effects
1492 f1, f2);
1493 Node* res = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0));
1494
1495 push(res);
1496 }
1497
1498 // Handle jsr and jsr_w bytecode
1499 void Parse::do_jsr() {
1500 assert(bc() == Bytecodes::_jsr || bc() == Bytecodes::_jsr_w, "wrong bytecode");
1501
1502 // Store information about current state, tagged with new _jsr_bci
1503 int return_bci = iter().next_bci();
1504 int jsr_bci = (bc() == Bytecodes::_jsr) ? iter().get_dest() : iter().get_far_dest();
1505
1506 // The way we do things now, there is only one successor block
1507 // for the jsr, because the target code is cloned by ciTypeFlow.
1508 Block* target = successor_for_bci(jsr_bci);
1509
1510 // What got pushed?
1511 const Type* ret_addr = target->peek();
1512 assert(ret_addr->singleton(), "must be a constant (cloned jsr body)");
1513
1514 // Effect on jsr on stack
1515 push(_gvn.makecon(ret_addr));
1516
1517 // Flow to the jsr.
1518 merge(jsr_bci);
1519 }
1520
1521 // Handle ret bytecode
1522 void Parse::do_ret() {
1523 // Find to whom we return.
1524 assert(block()->num_successors() == 1, "a ret can only go one place now");
1525 Block* target = block()->successor_at(0);
1526 assert(!target->is_ready(), "our arrival must be expected");
1527 int pnum = target->next_path_num();
1528 merge_common(target, pnum);
1529 }
1530
1531 static bool has_injected_profile(BoolTest::mask btest, Node* test, int& taken, int& not_taken) {
1532 if (btest != BoolTest::eq && btest != BoolTest::ne) {
1533 // Only ::eq and ::ne are supported for profile injection.
1534 return false;
1535 }
1536 if (test->is_Cmp() &&
1537 test->in(1)->Opcode() == Op_ProfileBoolean) {
1538 ProfileBooleanNode* profile = (ProfileBooleanNode*)test->in(1);
1539 int false_cnt = profile->false_count();
1540 int true_cnt = profile->true_count();
1541
1542 // Counts matching depends on the actual test operation (::eq or ::ne).
1543 // No need to scale the counts because profile injection was designed
1544 // to feed exact counts into VM.
1545 taken = (btest == BoolTest::eq) ? false_cnt : true_cnt;
1546 not_taken = (btest == BoolTest::eq) ? true_cnt : false_cnt;
1547
1548 profile->consume();
1549 return true;
1550 }
1551 return false;
1552 }
1553 //--------------------------dynamic_branch_prediction--------------------------
1554 // Try to gather dynamic branch prediction behavior. Return a probability
1555 // of the branch being taken and set the "cnt" field. Returns a -1.0
1556 // if we need to use static prediction for some reason.
1557 float Parse::dynamic_branch_prediction(float &cnt, BoolTest::mask btest, Node* test) {
1558 ResourceMark rm;
1559
1560 cnt = COUNT_UNKNOWN;
1561
1562 int taken = 0;
1563 int not_taken = 0;
1564
1565 bool use_mdo = !has_injected_profile(btest, test, taken, not_taken);
1566
1567 if (use_mdo) {
1568 // Use MethodData information if it is available
1569 // FIXME: free the ProfileData structure
1570 ciMethodData* methodData = method()->method_data();
1571 if (!methodData->is_mature()) return PROB_UNKNOWN;
1572 ciProfileData* data = methodData->bci_to_data(bci());
1573 if (data == NULL) {
1574 return PROB_UNKNOWN;
1575 }
1576 if (!data->is_JumpData()) return PROB_UNKNOWN;
1577
1578 // get taken and not taken values
1579 taken = data->as_JumpData()->taken();
1580 not_taken = 0;
1581 if (data->is_BranchData()) {
1582 not_taken = data->as_BranchData()->not_taken();
1583 }
1584
1585 // scale the counts to be commensurate with invocation counts:
1586 taken = method()->scale_count(taken);
1587 not_taken = method()->scale_count(not_taken);
1588 }
1589
1590 // Give up if too few (or too many, in which case the sum will overflow) counts to be meaningful.
1591 // We also check that individual counters are positive first, otherwise the sum can become positive.
1592 if (taken < 0 || not_taken < 0 || taken + not_taken < 40) {
1593 if (C->log() != NULL) {
1594 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d'", iter().get_dest(), taken, not_taken);
1595 }
1596 return PROB_UNKNOWN;
1597 }
1598
1599 // Compute frequency that we arrive here
1600 float sum = taken + not_taken;
1601 // Adjust, if this block is a cloned private block but the
1602 // Jump counts are shared. Taken the private counts for
1603 // just this path instead of the shared counts.
1604 if( block()->count() > 0 )
1605 sum = block()->count();
1606 cnt = sum / FreqCountInvocations;
1607
1608 // Pin probability to sane limits
1609 float prob;
1610 if( !taken )
1611 prob = (0+PROB_MIN) / 2;
1612 else if( !not_taken )
1613 prob = (1+PROB_MAX) / 2;
1614 else { // Compute probability of true path
1615 prob = (float)taken / (float)(taken + not_taken);
1616 if (prob > PROB_MAX) prob = PROB_MAX;
1617 if (prob < PROB_MIN) prob = PROB_MIN;
1618 }
1619
1620 assert((cnt > 0.0f) && (prob > 0.0f),
1621 "Bad frequency assignment in if");
1622
1623 if (C->log() != NULL) {
1624 const char* prob_str = NULL;
1625 if (prob >= PROB_MAX) prob_str = (prob == PROB_MAX) ? "max" : "always";
1626 if (prob <= PROB_MIN) prob_str = (prob == PROB_MIN) ? "min" : "never";
1627 char prob_str_buf[30];
1628 if (prob_str == NULL) {
1629 jio_snprintf(prob_str_buf, sizeof(prob_str_buf), "%20.2f", prob);
1630 prob_str = prob_str_buf;
1631 }
1632 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d' cnt='%f' prob='%s'",
1633 iter().get_dest(), taken, not_taken, cnt, prob_str);
1634 }
1635 return prob;
1636 }
1637
1638 //-----------------------------branch_prediction-------------------------------
1639 float Parse::branch_prediction(float& cnt,
1640 BoolTest::mask btest,
1641 int target_bci,
1642 Node* test) {
1643 float prob = dynamic_branch_prediction(cnt, btest, test);
1644 // If prob is unknown, switch to static prediction
1645 if (prob != PROB_UNKNOWN) return prob;
1646
1647 prob = PROB_FAIR; // Set default value
1648 if (btest == BoolTest::eq) // Exactly equal test?
1649 prob = PROB_STATIC_INFREQUENT; // Assume its relatively infrequent
1650 else if (btest == BoolTest::ne)
1651 prob = PROB_STATIC_FREQUENT; // Assume its relatively frequent
1652
1653 // If this is a conditional test guarding a backwards branch,
1654 // assume its a loop-back edge. Make it a likely taken branch.
1655 if (target_bci < bci()) {
1656 if (is_osr_parse()) { // Could be a hot OSR'd loop; force deopt
1657 // Since it's an OSR, we probably have profile data, but since
1658 // branch_prediction returned PROB_UNKNOWN, the counts are too small.
1659 // Let's make a special check here for completely zero counts.
1660 ciMethodData* methodData = method()->method_data();
1661 if (!methodData->is_empty()) {
1662 ciProfileData* data = methodData->bci_to_data(bci());
1663 // Only stop for truly zero counts, which mean an unknown part
1664 // of the OSR-ed method, and we want to deopt to gather more stats.
1665 // If you have ANY counts, then this loop is simply 'cold' relative
1666 // to the OSR loop.
1667 if (data == NULL ||
1668 (data->as_BranchData()->taken() + data->as_BranchData()->not_taken() == 0)) {
1669 // This is the only way to return PROB_UNKNOWN:
1670 return PROB_UNKNOWN;
1671 }
1672 }
1673 }
1674 prob = PROB_STATIC_FREQUENT; // Likely to take backwards branch
1675 }
1676
1677 assert(prob != PROB_UNKNOWN, "must have some guess at this point");
1678 return prob;
1679 }
1680
1681 // The magic constants are chosen so as to match the output of
1682 // branch_prediction() when the profile reports a zero taken count.
1683 // It is important to distinguish zero counts unambiguously, because
1684 // some branches (e.g., _213_javac.Assembler.eliminate) validly produce
1685 // very small but nonzero probabilities, which if confused with zero
1686 // counts would keep the program recompiling indefinitely.
1687 bool Parse::seems_never_taken(float prob) const {
1688 return prob < PROB_MIN;
1689 }
1690
1691 // True if the comparison seems to be the kind that will not change its
1692 // statistics from true to false. See comments in adjust_map_after_if.
1693 // This question is only asked along paths which are already
1694 // classifed as untaken (by seems_never_taken), so really,
1695 // if a path is never taken, its controlling comparison is
1696 // already acting in a stable fashion. If the comparison
1697 // seems stable, we will put an expensive uncommon trap
1698 // on the untaken path.
1699 bool Parse::seems_stable_comparison() const {
1700 if (C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if)) {
1701 return false;
1702 }
1703 return true;
1704 }
1705
1706 //-------------------------------repush_if_args--------------------------------
1707 // Push arguments of an "if" bytecode back onto the stack by adjusting _sp.
1708 inline int Parse::repush_if_args() {
1709 if (PrintOpto && WizardMode) {
1710 tty->print("defending against excessive implicit null exceptions on %s @%d in ",
1711 Bytecodes::name(iter().cur_bc()), iter().cur_bci());
1712 method()->print_name(); tty->cr();
1713 }
1714 int bc_depth = - Bytecodes::depth(iter().cur_bc());
1715 assert(bc_depth == 1 || bc_depth == 2, "only two kinds of branches");
1716 DEBUG_ONLY(sync_jvms()); // argument(n) requires a synced jvms
1717 assert(argument(0) != NULL, "must exist");
1718 assert(bc_depth == 1 || argument(1) != NULL, "two must exist");
1719 inc_sp(bc_depth);
1720 return bc_depth;
1721 }
1722
1723 //----------------------------------do_ifnull----------------------------------
1724 void Parse::do_ifnull(BoolTest::mask btest, Node *c) {
1725 int target_bci = iter().get_dest();
1726
1727 Block* branch_block = successor_for_bci(target_bci);
1728 Block* next_block = successor_for_bci(iter().next_bci());
1729
1730 float cnt;
1731 float prob = branch_prediction(cnt, btest, target_bci, c);
1732 if (prob == PROB_UNKNOWN) {
1733 // (An earlier version of do_ifnull omitted this trap for OSR methods.)
1734 if (PrintOpto && Verbose) {
1735 tty->print_cr("Never-taken edge stops compilation at bci %d", bci());
1736 }
1737 repush_if_args(); // to gather stats on loop
1738 uncommon_trap(Deoptimization::Reason_unreached,
1739 Deoptimization::Action_reinterpret,
1740 NULL, "cold");
1741 if (C->eliminate_boxing()) {
1742 // Mark the successor blocks as parsed
1743 branch_block->next_path_num();
1744 next_block->next_path_num();
1745 }
1746 return;
1747 }
1748
1749 NOT_PRODUCT(explicit_null_checks_inserted++);
1750
1751 // Generate real control flow
1752 Node *tst = _gvn.transform( new BoolNode( c, btest ) );
1753
1754 // Sanity check the probability value
1755 assert(prob > 0.0f,"Bad probability in Parser");
1756 // Need xform to put node in hash table
1757 IfNode *iff = create_and_xform_if( control(), tst, prob, cnt );
1758 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser");
1759 // True branch
1760 { PreserveJVMState pjvms(this);
1761 Node* iftrue = _gvn.transform( new IfTrueNode (iff) );
1762 set_control(iftrue);
1763
1764 if (stopped()) { // Path is dead?
1765 NOT_PRODUCT(explicit_null_checks_elided++);
1766 if (C->eliminate_boxing()) {
1767 // Mark the successor block as parsed
1768 branch_block->next_path_num();
1769 }
1770 } else { // Path is live.
1771 adjust_map_after_if(btest, c, prob, branch_block);
1772 if (!stopped()) {
1773 merge(target_bci);
1774 }
1775 }
1776 }
1777
1778 // False branch
1779 Node* iffalse = _gvn.transform( new IfFalseNode(iff) );
1780 set_control(iffalse);
1781
1782 if (stopped()) { // Path is dead?
1783 NOT_PRODUCT(explicit_null_checks_elided++);
1784 if (C->eliminate_boxing()) {
1785 // Mark the successor block as parsed
1786 next_block->next_path_num();
1787 }
1788 } else { // Path is live.
1789 adjust_map_after_if(BoolTest(btest).negate(), c, 1.0-prob, next_block);
1790 }
1791 }
1792
1793 //------------------------------------do_if------------------------------------
1794 void Parse::do_if(BoolTest::mask btest, Node* c, bool new_path, Node** ctrl_taken) {
1795 int target_bci = iter().get_dest();
1796
1797 Block* branch_block = successor_for_bci(target_bci);
1798 Block* next_block = successor_for_bci(iter().next_bci());
1799
1800 float cnt;
1801 float prob = branch_prediction(cnt, btest, target_bci, c);
1802 float untaken_prob = 1.0 - prob;
1803
1804 if (prob == PROB_UNKNOWN) {
1805 if (PrintOpto && Verbose) {
1806 tty->print_cr("Never-taken edge stops compilation at bci %d", bci());
1807 }
1808 repush_if_args(); // to gather stats on loop
1809 uncommon_trap(Deoptimization::Reason_unreached,
1810 Deoptimization::Action_reinterpret,
1811 NULL, "cold");
1812 if (C->eliminate_boxing()) {
1813 // Mark the successor blocks as parsed
1814 branch_block->next_path_num();
1815 next_block->next_path_num();
1816 }
1817 return;
1818 }
1819
1820 // Sanity check the probability value
1821 assert(0.0f < prob && prob < 1.0f,"Bad probability in Parser");
1822
1823 bool taken_if_true = true;
1824 // Convert BoolTest to canonical form:
1825 if (!BoolTest(btest).is_canonical()) {
1826 btest = BoolTest(btest).negate();
1827 taken_if_true = false;
1828 // prob is NOT updated here; it remains the probability of the taken
1829 // path (as opposed to the prob of the path guarded by an 'IfTrueNode').
1830 }
1831 assert(btest != BoolTest::eq, "!= is the only canonical exact test");
1832
1833 Node* tst0 = new BoolNode(c, btest);
1834 Node* tst = _gvn.transform(tst0);
1835 BoolTest::mask taken_btest = BoolTest::illegal;
1836 BoolTest::mask untaken_btest = BoolTest::illegal;
1837
1838 if (tst->is_Bool()) {
1839 // Refresh c from the transformed bool node, since it may be
1840 // simpler than the original c. Also re-canonicalize btest.
1841 // This wins when (Bool ne (Conv2B p) 0) => (Bool ne (CmpP p NULL)).
1842 // That can arise from statements like: if (x instanceof C) ...
1843 if (tst != tst0) {
1844 // Canonicalize one more time since transform can change it.
1845 btest = tst->as_Bool()->_test._test;
1846 if (!BoolTest(btest).is_canonical()) {
1847 // Reverse edges one more time...
1848 tst = _gvn.transform( tst->as_Bool()->negate(&_gvn) );
1849 btest = tst->as_Bool()->_test._test;
1850 assert(BoolTest(btest).is_canonical(), "sanity");
1851 taken_if_true = !taken_if_true;
1852 }
1853 c = tst->in(1);
1854 }
1855 BoolTest::mask neg_btest = BoolTest(btest).negate();
1856 taken_btest = taken_if_true ? btest : neg_btest;
1857 untaken_btest = taken_if_true ? neg_btest : btest;
1858 }
1859
1860 // Generate real control flow
1861 float true_prob = (taken_if_true ? prob : untaken_prob);
1862 IfNode* iff = create_and_map_if(control(), tst, true_prob, cnt);
1863 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser");
1864 Node* taken_branch = new IfTrueNode(iff);
1865 Node* untaken_branch = new IfFalseNode(iff);
1866 if (!taken_if_true) { // Finish conversion to canonical form
1867 Node* tmp = taken_branch;
1868 taken_branch = untaken_branch;
1869 untaken_branch = tmp;
1870 }
1871
1872 // Branch is taken:
1873 { PreserveJVMState pjvms(this);
1874 taken_branch = _gvn.transform(taken_branch);
1875 set_control(taken_branch);
1876
1877 if (stopped()) {
1878 if (C->eliminate_boxing() && !new_path) {
1879 // Mark the successor block as parsed (if we haven't created a new path)
1880 branch_block->next_path_num();
1881 }
1882 } else {
1883 adjust_map_after_if(taken_btest, c, prob, branch_block);
1884 if (!stopped()) {
1885 if (new_path) {
1886 // Merge by using a new path
1887 merge_new_path(target_bci);
1888 } else if (ctrl_taken != NULL) {
1889 // Don't merge but save taken branch to be wired by caller
1890 *ctrl_taken = control();
1891 } else {
1892 merge(target_bci);
1893 }
1894 }
1895 }
1896 }
1897
1898 untaken_branch = _gvn.transform(untaken_branch);
1899 set_control(untaken_branch);
1900
1901 // Branch not taken.
1902 if (stopped() && ctrl_taken == NULL) {
1903 if (C->eliminate_boxing()) {
1904 // Mark the successor block as parsed (if caller does not re-wire control flow)
1905 next_block->next_path_num();
1906 }
1907 } else {
1908 adjust_map_after_if(untaken_btest, c, untaken_prob, next_block);
1909 }
1910 }
1911
1912
1913 static ProfilePtrKind speculative_ptr_kind(const TypeOopPtr* t) {
1914 if (t->speculative() == NULL) {
1915 return ProfileUnknownNull;
1916 }
1917 if (t->speculative_always_null()) {
1918 return ProfileAlwaysNull;
1919 }
1920 if (t->speculative_maybe_null()) {
1921 return ProfileMaybeNull;
1922 }
1923 return ProfileNeverNull;
1924 }
1925
1926 void Parse::acmp_always_null_input(Node* input, const TypeOopPtr* tinput, BoolTest::mask btest, Node* eq_region) {
1927 inc_sp(2);
1928 Node* cast = null_check_common(input, T_OBJECT, true, NULL,
1929 !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_null_check) &&
1930 speculative_ptr_kind(tinput) == ProfileAlwaysNull);
1931 dec_sp(2);
1932 if (btest == BoolTest::ne) {
1933 {
1934 PreserveJVMState pjvms(this);
1935 replace_in_map(input, cast);
1936 int target_bci = iter().get_dest();
1937 merge(target_bci);
1938 }
1939 record_for_igvn(eq_region);
1940 set_control(_gvn.transform(eq_region));
1941 } else {
1942 replace_in_map(input, cast);
1943 }
1944 }
1945
1946 Node* Parse::acmp_null_check(Node* input, const TypeOopPtr* tinput, ProfilePtrKind input_ptr, Node*& null_ctl) {
1947 inc_sp(2);
1948 null_ctl = top();
1949 Node* cast = null_check_oop(input, &null_ctl,
1950 input_ptr == ProfileNeverNull || (input_ptr == ProfileUnknownNull && !too_many_traps_or_recompiles(Deoptimization::Reason_null_check)),
1951 false,
1952 speculative_ptr_kind(tinput) == ProfileNeverNull &&
1953 !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_null_check));
1954 dec_sp(2);
1955 assert(!stopped(), "null input should have been caught earlier");
1956 if (cast->is_InlineType()) {
1957 cast = cast->as_InlineType()->get_oop();
1958 }
1959 return cast;
1960 }
1961
1962 void Parse::acmp_known_non_inline_type_input(Node* input, const TypeOopPtr* tinput, ProfilePtrKind input_ptr, ciKlass* input_type, BoolTest::mask btest, Node* eq_region) {
1963 Node* ne_region = new RegionNode(1);
1964 Node* null_ctl;
1965 Node* cast = acmp_null_check(input, tinput, input_ptr, null_ctl);
1966 ne_region->add_req(null_ctl);
1967
1968 Node* slow_ctl = type_check_receiver(cast, input_type, 1.0, &cast);
1969 {
1970 PreserveJVMState pjvms(this);
1971 inc_sp(2);
1972 set_control(slow_ctl);
1973 Deoptimization::DeoptReason reason;
1974 if (tinput->speculative_type() != NULL && !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_class_check)) {
1975 reason = Deoptimization::Reason_speculate_class_check;
1976 } else {
1977 reason = Deoptimization::Reason_class_check;
1978 }
1979 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
1980 }
1981 ne_region->add_req(control());
1982
1983 record_for_igvn(ne_region);
1984 set_control(_gvn.transform(ne_region));
1985 if (btest == BoolTest::ne) {
1986 {
1987 PreserveJVMState pjvms(this);
1988 if (null_ctl == top()) {
1989 replace_in_map(input, cast);
1990 }
1991 int target_bci = iter().get_dest();
1992 merge(target_bci);
1993 }
1994 record_for_igvn(eq_region);
1995 set_control(_gvn.transform(eq_region));
1996 } else {
1997 if (null_ctl == top()) {
1998 replace_in_map(input, cast);
1999 }
2000 set_control(_gvn.transform(ne_region));
2001 }
2002 }
2003
2004 void Parse::acmp_unknown_non_inline_type_input(Node* input, const TypeOopPtr* tinput, ProfilePtrKind input_ptr, BoolTest::mask btest, Node* eq_region) {
2005 Node* ne_region = new RegionNode(1);
2006 Node* null_ctl;
2007 Node* cast = acmp_null_check(input, tinput, input_ptr, null_ctl);
2008 ne_region->add_req(null_ctl);
2009
2010 {
2011 BuildCutout unless(this, inline_type_test(cast, /* is_inline = */ false), PROB_MAX);
2012 inc_sp(2);
2013 uncommon_trap_exact(Deoptimization::Reason_class_check, Deoptimization::Action_maybe_recompile);
2014 }
2015
2016 ne_region->add_req(control());
2017
2018 record_for_igvn(ne_region);
2019 set_control(_gvn.transform(ne_region));
2020 if (btest == BoolTest::ne) {
2021 {
2022 PreserveJVMState pjvms(this);
2023 if (null_ctl == top()) {
2024 replace_in_map(input, cast);
2025 }
2026 int target_bci = iter().get_dest();
2027 merge(target_bci);
2028 }
2029 record_for_igvn(eq_region);
2030 set_control(_gvn.transform(eq_region));
2031 } else {
2032 if (null_ctl == top()) {
2033 replace_in_map(input, cast);
2034 }
2035 set_control(_gvn.transform(ne_region));
2036 }
2037 }
2038
2039 void Parse::do_acmp(BoolTest::mask btest, Node* left, Node* right) {
2040 ciKlass* left_type = NULL;
2041 ciKlass* right_type = NULL;
2042 ProfilePtrKind left_ptr = ProfileUnknownNull;
2043 ProfilePtrKind right_ptr = ProfileUnknownNull;
2044 bool left_inline_type = true;
2045 bool right_inline_type = true;
2046
2047 // Leverage profiling at acmp
2048 if (UseACmpProfile) {
2049 method()->acmp_profiled_type(bci(), left_type, right_type, left_ptr, right_ptr, left_inline_type, right_inline_type);
2050 if (too_many_traps_or_recompiles(Deoptimization::Reason_class_check)) {
2051 left_type = NULL;
2052 right_type = NULL;
2053 left_inline_type = true;
2054 right_inline_type = true;
2055 }
2056 if (too_many_traps_or_recompiles(Deoptimization::Reason_null_check)) {
2057 left_ptr = ProfileUnknownNull;
2058 right_ptr = ProfileUnknownNull;
2059 }
2060 }
2061
2062 if (UseTypeSpeculation) {
2063 record_profile_for_speculation(left, left_type, left_ptr);
2064 record_profile_for_speculation(right, right_type, right_ptr);
2065 }
2066
2067 if (!EnableValhalla) {
2068 Node* cmp = CmpP(left, right);
2069 cmp = optimize_cmp_with_klass(cmp);
2070 do_if(btest, cmp);
2071 return;
2072 }
2073
2074 // Check for equality before potentially allocating
2075 if (left == right) {
2076 do_if(btest, makecon(TypeInt::CC_EQ));
2077 return;
2078 }
2079
2080 // Allocate inline type operands and re-execute on deoptimization
2081 if (left->is_InlineType()) {
2082 PreserveReexecuteState preexecs(this);
2083 inc_sp(2);
2084 jvms()->set_should_reexecute(true);
2085 left = left->as_InlineType()->buffer(this)->get_oop();
2086 }
2087 if (right->is_InlineType()) {
2088 PreserveReexecuteState preexecs(this);
2089 inc_sp(2);
2090 jvms()->set_should_reexecute(true);
2091 right = right->as_InlineType()->buffer(this)->get_oop();
2092 }
2093
2094 // First, do a normal pointer comparison
2095 const TypeOopPtr* tleft = _gvn.type(left)->isa_oopptr();
2096 const TypeOopPtr* tright = _gvn.type(right)->isa_oopptr();
2097 Node* cmp = CmpP(left, right);
2098 cmp = optimize_cmp_with_klass(cmp);
2099 if (tleft == NULL || !tleft->can_be_inline_type() ||
2100 tright == NULL || !tright->can_be_inline_type()) {
2101 // This is sufficient, if one of the operands can't be an inline type
2102 do_if(btest, cmp);
2103 return;
2104 }
2105 Node* eq_region = NULL;
2106 if (btest == BoolTest::eq) {
2107 do_if(btest, cmp, true);
2108 if (stopped()) {
2109 return;
2110 }
2111 } else {
2112 assert(btest == BoolTest::ne, "only eq or ne");
2113 Node* is_not_equal = NULL;
2114 eq_region = new RegionNode(3);
2115 {
2116 PreserveJVMState pjvms(this);
2117 do_if(btest, cmp, false, &is_not_equal);
2118 if (!stopped()) {
2119 eq_region->init_req(1, control());
2120 }
2121 }
2122 if (is_not_equal == NULL || is_not_equal->is_top()) {
2123 record_for_igvn(eq_region);
2124 set_control(_gvn.transform(eq_region));
2125 return;
2126 }
2127 set_control(is_not_equal);
2128 }
2129
2130 // Prefer speculative types if available
2131 if (!too_many_traps_or_recompiles(Deoptimization::Reason_speculate_class_check)) {
2132 if (tleft->speculative_type() != NULL) {
2133 left_type = tleft->speculative_type();
2134 }
2135 if (tright->speculative_type() != NULL) {
2136 right_type = tright->speculative_type();
2137 }
2138 }
2139
2140 if (speculative_ptr_kind(tleft) != ProfileMaybeNull && speculative_ptr_kind(tleft) != ProfileUnknownNull) {
2141 ProfilePtrKind speculative_left_ptr = speculative_ptr_kind(tleft);
2142 if (speculative_left_ptr == ProfileAlwaysNull && !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_null_assert)) {
2143 left_ptr = speculative_left_ptr;
2144 } else if (speculative_left_ptr == ProfileNeverNull && !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_null_check)) {
2145 left_ptr = speculative_left_ptr;
2146 }
2147 }
2148 if (speculative_ptr_kind(tright) != ProfileMaybeNull && speculative_ptr_kind(tright) != ProfileUnknownNull) {
2149 ProfilePtrKind speculative_right_ptr = speculative_ptr_kind(tright);
2150 if (speculative_right_ptr == ProfileAlwaysNull && !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_null_assert)) {
2151 right_ptr = speculative_right_ptr;
2152 } else if (speculative_right_ptr == ProfileNeverNull && !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_null_check)) {
2153 right_ptr = speculative_right_ptr;
2154 }
2155 }
2156
2157 if (left_ptr == ProfileAlwaysNull) {
2158 // Comparison with null. Assert the input is indeed null and we're done.
2159 acmp_always_null_input(left, tleft, btest, eq_region);
2160 return;
2161 }
2162 if (right_ptr == ProfileAlwaysNull) {
2163 // Comparison with null. Assert the input is indeed null and we're done.
2164 acmp_always_null_input(right, tright, btest, eq_region);
2165 return;
2166 }
2167 if (left_type != NULL && !left_type->is_inlinetype()) {
2168 // Comparison with an object of known type
2169 acmp_known_non_inline_type_input(left, tleft, left_ptr, left_type, btest, eq_region);
2170 return;
2171 }
2172 if (right_type != NULL && !right_type->is_inlinetype()) {
2173 // Comparison with an object of known type
2174 acmp_known_non_inline_type_input(right, tright, right_ptr, right_type, btest, eq_region);
2175 return;
2176 }
2177 if (!left_inline_type) {
2178 // Comparison with an object known not to be an inline type
2179 acmp_unknown_non_inline_type_input(left, tleft, left_ptr, btest, eq_region);
2180 return;
2181 }
2182 if (!right_inline_type) {
2183 // Comparison with an object known not to be an inline type
2184 acmp_unknown_non_inline_type_input(right, tright, right_ptr, btest, eq_region);
2185 return;
2186 }
2187
2188 // Pointers are not equal, check if first operand is non-null
2189 Node* ne_region = new RegionNode(6);
2190 Node* null_ctl;
2191 Node* not_null_right = acmp_null_check(right, tright, right_ptr, null_ctl);
2192 ne_region->init_req(1, null_ctl);
2193
2194 // First operand is non-null, check if it is an inline type
2195 Node* is_value = inline_type_test(not_null_right);
2196 IfNode* is_value_iff = create_and_map_if(control(), is_value, PROB_FAIR, COUNT_UNKNOWN);
2197 Node* not_value = _gvn.transform(new IfFalseNode(is_value_iff));
2198 ne_region->init_req(2, not_value);
2199 set_control(_gvn.transform(new IfTrueNode(is_value_iff)));
2200
2201 // The first operand is an inline type, check if the second operand is non-null
2202 Node* not_null_left = acmp_null_check(left, tleft, left_ptr, null_ctl);
2203 ne_region->init_req(3, null_ctl);
2204
2205 // Check if both operands are of the same class.
2206 Node* kls_left = load_object_klass(not_null_left);
2207 Node* kls_right = load_object_klass(not_null_right);
2208 Node* kls_cmp = CmpP(kls_left, kls_right);
2209 Node* kls_bol = _gvn.transform(new BoolNode(kls_cmp, BoolTest::ne));
2210 IfNode* kls_iff = create_and_map_if(control(), kls_bol, PROB_FAIR, COUNT_UNKNOWN);
2211 Node* kls_ne = _gvn.transform(new IfTrueNode(kls_iff));
2212 set_control(_gvn.transform(new IfFalseNode(kls_iff)));
2213 ne_region->init_req(4, kls_ne);
2214
2215 if (stopped()) {
2216 record_for_igvn(ne_region);
2217 set_control(_gvn.transform(ne_region));
2218 if (btest == BoolTest::ne) {
2219 {
2220 PreserveJVMState pjvms(this);
2221 int target_bci = iter().get_dest();
2222 merge(target_bci);
2223 }
2224 record_for_igvn(eq_region);
2225 set_control(_gvn.transform(eq_region));
2226 }
2227 return;
2228 }
2229
2230 // Both operands are values types of the same class, we need to perform a
2231 // substitutability test. Delegate to PrimitiveObjectMethods::isSubstitutable().
2232 Node* ne_io_phi = PhiNode::make(ne_region, i_o());
2233 Node* mem = reset_memory();
2234 Node* ne_mem_phi = PhiNode::make(ne_region, mem);
2235
2236 Node* eq_io_phi = NULL;
2237 Node* eq_mem_phi = NULL;
2238 if (eq_region != NULL) {
2239 eq_io_phi = PhiNode::make(eq_region, i_o());
2240 eq_mem_phi = PhiNode::make(eq_region, mem);
2241 }
2242
2243 set_all_memory(mem);
2244
2245 kill_dead_locals();
2246 ciMethod* subst_method = ciEnv::current()->PrimitiveObjectMethods_klass()->find_method(ciSymbols::isSubstitutable_name(), ciSymbols::object_object_boolean_signature());
2247 CallStaticJavaNode *call = new CallStaticJavaNode(C, TypeFunc::make(subst_method), SharedRuntime::get_resolve_static_call_stub(), subst_method);
2248 call->set_override_symbolic_info(true);
2249 call->init_req(TypeFunc::Parms, not_null_left);
2250 call->init_req(TypeFunc::Parms+1, not_null_right);
2251 inc_sp(2);
2252 set_edges_for_java_call(call, false, false);
2253 Node* ret = set_results_for_java_call(call, false, true);
2254 dec_sp(2);
2255
2256 // Test the return value of PrimitiveObjectMethods::isSubstitutable()
2257 Node* subst_cmp = _gvn.transform(new CmpINode(ret, intcon(1)));
2258 Node* ctl = C->top();
2259 if (btest == BoolTest::eq) {
2260 PreserveJVMState pjvms(this);
2261 do_if(btest, subst_cmp);
2262 if (!stopped()) {
2263 ctl = control();
2264 }
2265 } else {
2266 assert(btest == BoolTest::ne, "only eq or ne");
2267 PreserveJVMState pjvms(this);
2268 do_if(btest, subst_cmp, false, &ctl);
2269 if (!stopped()) {
2270 eq_region->init_req(2, control());
2271 eq_io_phi->init_req(2, i_o());
2272 eq_mem_phi->init_req(2, reset_memory());
2273 }
2274 }
2275 ne_region->init_req(5, ctl);
2276 ne_io_phi->init_req(5, i_o());
2277 ne_mem_phi->init_req(5, reset_memory());
2278
2279 record_for_igvn(ne_region);
2280 set_control(_gvn.transform(ne_region));
2281 set_i_o(_gvn.transform(ne_io_phi));
2282 set_all_memory(_gvn.transform(ne_mem_phi));
2283
2284 if (btest == BoolTest::ne) {
2285 {
2286 PreserveJVMState pjvms(this);
2287 int target_bci = iter().get_dest();
2288 merge(target_bci);
2289 }
2290
2291 record_for_igvn(eq_region);
2292 set_control(_gvn.transform(eq_region));
2293 set_i_o(_gvn.transform(eq_io_phi));
2294 set_all_memory(_gvn.transform(eq_mem_phi));
2295 }
2296 }
2297
2298 bool Parse::path_is_suitable_for_uncommon_trap(float prob) const {
2299 // Don't want to speculate on uncommon traps when running with -Xcomp
2300 if (!UseInterpreter) {
2301 return false;
2302 }
2303 return (seems_never_taken(prob) && seems_stable_comparison());
2304 }
2305
2306 void Parse::maybe_add_predicate_after_if(Block* path) {
2307 if (path->is_SEL_head() && path->preds_parsed() == 0) {
2308 // Add predicates at bci of if dominating the loop so traps can be
2309 // recorded on the if's profile data
2310 int bc_depth = repush_if_args();
2311 add_empty_predicates();
2312 dec_sp(bc_depth);
2313 path->set_has_predicates();
2314 }
2315 }
2316
2317
2318 //----------------------------adjust_map_after_if------------------------------
2319 // Adjust the JVM state to reflect the result of taking this path.
2320 // Basically, it means inspecting the CmpNode controlling this
2321 // branch, seeing how it constrains a tested value, and then
2322 // deciding if it's worth our while to encode this constraint
2323 // as graph nodes in the current abstract interpretation map.
2324 void Parse::adjust_map_after_if(BoolTest::mask btest, Node* c, float prob, Block* path) {
2325 if (!c->is_Cmp()) {
2326 maybe_add_predicate_after_if(path);
2327 return;
2328 }
2329
2330 if (stopped() || btest == BoolTest::illegal) {
2331 return; // nothing to do
2332 }
2333
2334 bool is_fallthrough = (path == successor_for_bci(iter().next_bci()));
2335
2336 if (path_is_suitable_for_uncommon_trap(prob)) {
2337 repush_if_args();
2338 uncommon_trap(Deoptimization::Reason_unstable_if,
2339 Deoptimization::Action_reinterpret,
2340 NULL,
2341 (is_fallthrough ? "taken always" : "taken never"));
2342 return;
2343 }
2344
2345 Node* val = c->in(1);
2346 Node* con = c->in(2);
2347 const Type* tcon = _gvn.type(con);
2348 const Type* tval = _gvn.type(val);
2349 bool have_con = tcon->singleton();
2350 if (tval->singleton()) {
2351 if (!have_con) {
2352 // Swap, so constant is in con.
2353 con = val;
2354 tcon = tval;
2355 val = c->in(2);
2356 tval = _gvn.type(val);
2357 btest = BoolTest(btest).commute();
2358 have_con = true;
2359 } else {
2360 // Do we have two constants? Then leave well enough alone.
2361 have_con = false;
2362 }
2363 }
2364 if (!have_con) { // remaining adjustments need a con
2365 maybe_add_predicate_after_if(path);
2366 return;
2367 }
2368
2369 sharpen_type_after_if(btest, con, tcon, val, tval);
2370 maybe_add_predicate_after_if(path);
2371 }
2372
2373
2374 static Node* extract_obj_from_klass_load(PhaseGVN* gvn, Node* n) {
2375 Node* ldk;
2376 if (n->is_DecodeNKlass()) {
2377 if (n->in(1)->Opcode() != Op_LoadNKlass) {
2378 return NULL;
2379 } else {
2380 ldk = n->in(1);
2381 }
2382 } else if (n->Opcode() != Op_LoadKlass) {
2383 return NULL;
2384 } else {
2385 ldk = n;
2386 }
2387 assert(ldk != NULL && ldk->is_Load(), "should have found a LoadKlass or LoadNKlass node");
2388
2389 Node* adr = ldk->in(MemNode::Address);
2390 intptr_t off = 0;
2391 Node* obj = AddPNode::Ideal_base_and_offset(adr, gvn, off);
2392 if (obj == NULL || off != oopDesc::klass_offset_in_bytes()) // loading oopDesc::_klass?
2393 return NULL;
2394 const TypePtr* tp = gvn->type(obj)->is_ptr();
2395 if (tp == NULL || !(tp->isa_instptr() || tp->isa_aryptr())) // is obj a Java object ptr?
2396 return NULL;
2397
2398 return obj;
2399 }
2400
2401 void Parse::sharpen_type_after_if(BoolTest::mask btest,
2402 Node* con, const Type* tcon,
2403 Node* val, const Type* tval) {
2404 // Look for opportunities to sharpen the type of a node
2405 // whose klass is compared with a constant klass.
2406 if (btest == BoolTest::eq && tcon->isa_klassptr()) {
2407 Node* obj = extract_obj_from_klass_load(&_gvn, val);
2408 const TypeOopPtr* con_type = tcon->isa_klassptr()->as_instance_type();
2409 if (obj != NULL && (con_type->isa_instptr() || con_type->isa_aryptr())) {
2410 // Found:
2411 // Bool(CmpP(LoadKlass(obj._klass), ConP(Foo.klass)), [eq])
2412 // or the narrowOop equivalent.
2413 const Type* obj_type = _gvn.type(obj);
2414 const TypeOopPtr* tboth = obj_type->join_speculative(con_type)->isa_oopptr();
2415 if (tboth != NULL && tboth->klass_is_exact() && tboth != obj_type &&
2416 tboth->higher_equal(obj_type)) {
2417 // obj has to be of the exact type Foo if the CmpP succeeds.
2418 int obj_in_map = map()->find_edge(obj);
2419 JVMState* jvms = this->jvms();
2420 if (obj_in_map >= 0 &&
2421 (jvms->is_loc(obj_in_map) || jvms->is_stk(obj_in_map))) {
2422 TypeNode* ccast = new CheckCastPPNode(control(), obj, tboth);
2423 const Type* tcc = ccast->as_Type()->type();
2424 assert(tcc != obj_type && tcc->higher_equal(obj_type), "must improve");
2425 // Delay transform() call to allow recovery of pre-cast value
2426 // at the control merge.
2427 _gvn.set_type_bottom(ccast);
2428 record_for_igvn(ccast);
2429 // Here's the payoff.
2430 replace_in_map(obj, ccast);
2431 }
2432 }
2433 }
2434 }
2435
2436 int val_in_map = map()->find_edge(val);
2437 if (val_in_map < 0) return; // replace_in_map would be useless
2438 {
2439 JVMState* jvms = this->jvms();
2440 if (!(jvms->is_loc(val_in_map) ||
2441 jvms->is_stk(val_in_map)))
2442 return; // again, it would be useless
2443 }
2444
2445 // Check for a comparison to a constant, and "know" that the compared
2446 // value is constrained on this path.
2447 assert(tcon->singleton(), "");
2448 ConstraintCastNode* ccast = NULL;
2449 Node* cast = NULL;
2450
2451 switch (btest) {
2452 case BoolTest::eq: // Constant test?
2453 {
2454 const Type* tboth = tcon->join_speculative(tval);
2455 if (tboth == tval) break; // Nothing to gain.
2456 if (tcon->isa_int()) {
2457 ccast = new CastIINode(val, tboth);
2458 } else if (tcon == TypePtr::NULL_PTR) {
2459 // Cast to null, but keep the pointer identity temporarily live.
2460 ccast = new CastPPNode(val, tboth);
2461 } else {
2462 const TypeF* tf = tcon->isa_float_constant();
2463 const TypeD* td = tcon->isa_double_constant();
2464 // Exclude tests vs float/double 0 as these could be
2465 // either +0 or -0. Just because you are equal to +0
2466 // doesn't mean you ARE +0!
2467 // Note, following code also replaces Long and Oop values.
2468 if ((!tf || tf->_f != 0.0) &&
2469 (!td || td->_d != 0.0))
2470 cast = con; // Replace non-constant val by con.
2471 }
2472 }
2473 break;
2474
2475 case BoolTest::ne:
2476 if (tcon == TypePtr::NULL_PTR) {
2477 cast = cast_not_null(val, false);
2478 }
2479 break;
2480
2481 default:
2482 // (At this point we could record int range types with CastII.)
2483 break;
2484 }
2485
2486 if (ccast != NULL) {
2487 const Type* tcc = ccast->as_Type()->type();
2488 assert(tcc != tval && tcc->higher_equal(tval), "must improve");
2489 // Delay transform() call to allow recovery of pre-cast value
2490 // at the control merge.
2491 ccast->set_req(0, control());
2492 _gvn.set_type_bottom(ccast);
2493 record_for_igvn(ccast);
2494 cast = ccast;
2495 }
2496
2497 if (cast != NULL) { // Here's the payoff.
2498 replace_in_map(val, cast);
2499 }
2500 }
2501
2502 /**
2503 * Use speculative type to optimize CmpP node: if comparison is
2504 * against the low level class, cast the object to the speculative
2505 * type if any. CmpP should then go away.
2506 *
2507 * @param c expected CmpP node
2508 * @return result of CmpP on object casted to speculative type
2509 *
2510 */
2511 Node* Parse::optimize_cmp_with_klass(Node* c) {
2512 // If this is transformed by the _gvn to a comparison with the low
2513 // level klass then we may be able to use speculation
2514 if (c->Opcode() == Op_CmpP &&
2515 (c->in(1)->Opcode() == Op_LoadKlass || c->in(1)->Opcode() == Op_DecodeNKlass) &&
2516 c->in(2)->is_Con()) {
2517 Node* load_klass = NULL;
2518 Node* decode = NULL;
2519 if (c->in(1)->Opcode() == Op_DecodeNKlass) {
2520 decode = c->in(1);
2521 load_klass = c->in(1)->in(1);
2522 } else {
2523 load_klass = c->in(1);
2524 }
2525 if (load_klass->in(2)->is_AddP()) {
2526 Node* addp = load_klass->in(2);
2527 Node* obj = addp->in(AddPNode::Address);
2528 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
2529 if (obj_type->speculative_type_not_null() != NULL) {
2530 ciKlass* k = obj_type->speculative_type();
2531 inc_sp(2);
2532 obj = maybe_cast_profiled_obj(obj, k);
2533 dec_sp(2);
2534 if (obj->is_InlineType()) {
2535 assert(obj->as_InlineType()->is_allocated(&_gvn), "must be allocated");
2536 obj = obj->as_InlineType()->get_oop();
2537 }
2538 // Make the CmpP use the casted obj
2539 addp = basic_plus_adr(obj, addp->in(AddPNode::Offset));
2540 load_klass = load_klass->clone();
2541 load_klass->set_req(2, addp);
2542 load_klass = _gvn.transform(load_klass);
2543 if (decode != NULL) {
2544 decode = decode->clone();
2545 decode->set_req(1, load_klass);
2546 load_klass = _gvn.transform(decode);
2547 }
2548 c = c->clone();
2549 c->set_req(1, load_klass);
2550 c = _gvn.transform(c);
2551 }
2552 }
2553 }
2554 return c;
2555 }
2556
2557 //------------------------------do_one_bytecode--------------------------------
2558 // Parse this bytecode, and alter the Parsers JVM->Node mapping
2559 void Parse::do_one_bytecode() {
2560 Node *a, *b, *c, *d; // Handy temps
2561 BoolTest::mask btest;
2562 int i;
2563
2564 assert(!has_exceptions(), "bytecode entry state must be clear of throws");
2565
2566 if (C->check_node_count(NodeLimitFudgeFactor * 5,
2567 "out of nodes parsing method")) {
2568 return;
2569 }
2570
2571 #ifdef ASSERT
2572 // for setting breakpoints
2573 if (TraceOptoParse) {
2574 tty->print(" @");
2575 dump_bci(bci());
2576 tty->cr();
2577 }
2578 #endif
2579
2580 switch (bc()) {
2581 case Bytecodes::_nop:
2582 // do nothing
2583 break;
2584 case Bytecodes::_lconst_0:
2585 push_pair(longcon(0));
2586 break;
2587
2588 case Bytecodes::_lconst_1:
2589 push_pair(longcon(1));
2590 break;
2591
2592 case Bytecodes::_fconst_0:
2593 push(zerocon(T_FLOAT));
2594 break;
2595
2596 case Bytecodes::_fconst_1:
2597 push(makecon(TypeF::ONE));
2598 break;
2599
2600 case Bytecodes::_fconst_2:
2601 push(makecon(TypeF::make(2.0f)));
2602 break;
2603
2604 case Bytecodes::_dconst_0:
2605 push_pair(zerocon(T_DOUBLE));
2606 break;
2607
2608 case Bytecodes::_dconst_1:
2609 push_pair(makecon(TypeD::ONE));
2610 break;
2611
2612 case Bytecodes::_iconst_m1:push(intcon(-1)); break;
2613 case Bytecodes::_iconst_0: push(intcon( 0)); break;
2614 case Bytecodes::_iconst_1: push(intcon( 1)); break;
2615 case Bytecodes::_iconst_2: push(intcon( 2)); break;
2616 case Bytecodes::_iconst_3: push(intcon( 3)); break;
2617 case Bytecodes::_iconst_4: push(intcon( 4)); break;
2618 case Bytecodes::_iconst_5: push(intcon( 5)); break;
2619 case Bytecodes::_bipush: push(intcon(iter().get_constant_u1())); break;
2620 case Bytecodes::_sipush: push(intcon(iter().get_constant_u2())); break;
2621 case Bytecodes::_aconst_null: push(null()); break;
2622 case Bytecodes::_ldc:
2623 case Bytecodes::_ldc_w:
2624 case Bytecodes::_ldc2_w:
2625 // If the constant is unresolved, run this BC once in the interpreter.
2626 {
2627 ciConstant constant = iter().get_constant();
2628 if (!constant.is_valid() ||
2629 (constant.basic_type() == T_OBJECT &&
2630 !constant.as_object()->is_loaded())) {
2631 int index = iter().get_constant_pool_index();
2632 constantTag tag = iter().get_constant_pool_tag(index);
2633 uncommon_trap(Deoptimization::make_trap_request
2634 (Deoptimization::Reason_unloaded,
2635 Deoptimization::Action_reinterpret,
2636 index),
2637 NULL, tag.internal_name());
2638 break;
2639 }
2640 assert(constant.basic_type() != T_OBJECT || constant.as_object()->is_instance(),
2641 "must be java_mirror of klass");
2642 const Type* con_type = Type::make_from_constant(constant);
2643 if (con_type != NULL) {
2644 push_node(con_type->basic_type(), makecon(con_type));
2645 }
2646 }
2647
2648 break;
2649
2650 case Bytecodes::_aload_0:
2651 push( local(0) );
2652 break;
2653 case Bytecodes::_aload_1:
2654 push( local(1) );
2655 break;
2656 case Bytecodes::_aload_2:
2657 push( local(2) );
2658 break;
2659 case Bytecodes::_aload_3:
2660 push( local(3) );
2661 break;
2662 case Bytecodes::_aload:
2663 push( local(iter().get_index()) );
2664 break;
2665
2666 case Bytecodes::_fload_0:
2667 case Bytecodes::_iload_0:
2668 push( local(0) );
2669 break;
2670 case Bytecodes::_fload_1:
2671 case Bytecodes::_iload_1:
2672 push( local(1) );
2673 break;
2674 case Bytecodes::_fload_2:
2675 case Bytecodes::_iload_2:
2676 push( local(2) );
2677 break;
2678 case Bytecodes::_fload_3:
2679 case Bytecodes::_iload_3:
2680 push( local(3) );
2681 break;
2682 case Bytecodes::_fload:
2683 case Bytecodes::_iload:
2684 push( local(iter().get_index()) );
2685 break;
2686 case Bytecodes::_lload_0:
2687 push_pair_local( 0 );
2688 break;
2689 case Bytecodes::_lload_1:
2690 push_pair_local( 1 );
2691 break;
2692 case Bytecodes::_lload_2:
2693 push_pair_local( 2 );
2694 break;
2695 case Bytecodes::_lload_3:
2696 push_pair_local( 3 );
2697 break;
2698 case Bytecodes::_lload:
2699 push_pair_local( iter().get_index() );
2700 break;
2701
2702 case Bytecodes::_dload_0:
2703 push_pair_local(0);
2704 break;
2705 case Bytecodes::_dload_1:
2706 push_pair_local(1);
2707 break;
2708 case Bytecodes::_dload_2:
2709 push_pair_local(2);
2710 break;
2711 case Bytecodes::_dload_3:
2712 push_pair_local(3);
2713 break;
2714 case Bytecodes::_dload:
2715 push_pair_local(iter().get_index());
2716 break;
2717 case Bytecodes::_fstore_0:
2718 case Bytecodes::_istore_0:
2719 case Bytecodes::_astore_0:
2720 set_local( 0, pop() );
2721 break;
2722 case Bytecodes::_fstore_1:
2723 case Bytecodes::_istore_1:
2724 case Bytecodes::_astore_1:
2725 set_local( 1, pop() );
2726 break;
2727 case Bytecodes::_fstore_2:
2728 case Bytecodes::_istore_2:
2729 case Bytecodes::_astore_2:
2730 set_local( 2, pop() );
2731 break;
2732 case Bytecodes::_fstore_3:
2733 case Bytecodes::_istore_3:
2734 case Bytecodes::_astore_3:
2735 set_local( 3, pop() );
2736 break;
2737 case Bytecodes::_fstore:
2738 case Bytecodes::_istore:
2739 case Bytecodes::_astore:
2740 set_local( iter().get_index(), pop() );
2741 break;
2742 // long stores
2743 case Bytecodes::_lstore_0:
2744 set_pair_local( 0, pop_pair() );
2745 break;
2746 case Bytecodes::_lstore_1:
2747 set_pair_local( 1, pop_pair() );
2748 break;
2749 case Bytecodes::_lstore_2:
2750 set_pair_local( 2, pop_pair() );
2751 break;
2752 case Bytecodes::_lstore_3:
2753 set_pair_local( 3, pop_pair() );
2754 break;
2755 case Bytecodes::_lstore:
2756 set_pair_local( iter().get_index(), pop_pair() );
2757 break;
2758
2759 // double stores
2760 case Bytecodes::_dstore_0:
2761 set_pair_local( 0, dstore_rounding(pop_pair()) );
2762 break;
2763 case Bytecodes::_dstore_1:
2764 set_pair_local( 1, dstore_rounding(pop_pair()) );
2765 break;
2766 case Bytecodes::_dstore_2:
2767 set_pair_local( 2, dstore_rounding(pop_pair()) );
2768 break;
2769 case Bytecodes::_dstore_3:
2770 set_pair_local( 3, dstore_rounding(pop_pair()) );
2771 break;
2772 case Bytecodes::_dstore:
2773 set_pair_local( iter().get_index(), dstore_rounding(pop_pair()) );
2774 break;
2775
2776 case Bytecodes::_pop: dec_sp(1); break;
2777 case Bytecodes::_pop2: dec_sp(2); break;
2778 case Bytecodes::_swap:
2779 a = pop();
2780 b = pop();
2781 push(a);
2782 push(b);
2783 break;
2784 case Bytecodes::_dup:
2785 a = pop();
2786 push(a);
2787 push(a);
2788 break;
2789 case Bytecodes::_dup_x1:
2790 a = pop();
2791 b = pop();
2792 push( a );
2793 push( b );
2794 push( a );
2795 break;
2796 case Bytecodes::_dup_x2:
2797 a = pop();
2798 b = pop();
2799 c = pop();
2800 push( a );
2801 push( c );
2802 push( b );
2803 push( a );
2804 break;
2805 case Bytecodes::_dup2:
2806 a = pop();
2807 b = pop();
2808 push( b );
2809 push( a );
2810 push( b );
2811 push( a );
2812 break;
2813
2814 case Bytecodes::_dup2_x1:
2815 // before: .. c, b, a
2816 // after: .. b, a, c, b, a
2817 // not tested
2818 a = pop();
2819 b = pop();
2820 c = pop();
2821 push( b );
2822 push( a );
2823 push( c );
2824 push( b );
2825 push( a );
2826 break;
2827 case Bytecodes::_dup2_x2:
2828 // before: .. d, c, b, a
2829 // after: .. b, a, d, c, b, a
2830 // not tested
2831 a = pop();
2832 b = pop();
2833 c = pop();
2834 d = pop();
2835 push( b );
2836 push( a );
2837 push( d );
2838 push( c );
2839 push( b );
2840 push( a );
2841 break;
2842
2843 case Bytecodes::_arraylength: {
2844 // Must do null-check with value on expression stack
2845 Node *ary = null_check(peek(), T_ARRAY);
2846 // Compile-time detect of null-exception?
2847 if (stopped()) return;
2848 a = pop();
2849 push(load_array_length(a));
2850 break;
2851 }
2852
2853 case Bytecodes::_baload: array_load(T_BYTE); break;
2854 case Bytecodes::_caload: array_load(T_CHAR); break;
2855 case Bytecodes::_iaload: array_load(T_INT); break;
2856 case Bytecodes::_saload: array_load(T_SHORT); break;
2857 case Bytecodes::_faload: array_load(T_FLOAT); break;
2858 case Bytecodes::_aaload: array_load(T_OBJECT); break;
2859 case Bytecodes::_laload: array_load(T_LONG); break;
2860 case Bytecodes::_daload: array_load(T_DOUBLE); break;
2861 case Bytecodes::_bastore: array_store(T_BYTE); break;
2862 case Bytecodes::_castore: array_store(T_CHAR); break;
2863 case Bytecodes::_iastore: array_store(T_INT); break;
2864 case Bytecodes::_sastore: array_store(T_SHORT); break;
2865 case Bytecodes::_fastore: array_store(T_FLOAT); break;
2866 case Bytecodes::_aastore: array_store(T_OBJECT); break;
2867 case Bytecodes::_lastore: array_store(T_LONG); break;
2868 case Bytecodes::_dastore: array_store(T_DOUBLE); break;
2869
2870 case Bytecodes::_getfield:
2871 do_getfield();
2872 break;
2873
2874 case Bytecodes::_getstatic:
2875 do_getstatic();
2876 break;
2877
2878 case Bytecodes::_putfield:
2879 do_putfield();
2880 break;
2881
2882 case Bytecodes::_putstatic:
2883 do_putstatic();
2884 break;
2885
2886 case Bytecodes::_irem:
2887 // Must keep both values on the expression-stack during null-check
2888 zero_check_int(peek());
2889 // Compile-time detect of null-exception?
2890 if (stopped()) return;
2891 b = pop();
2892 a = pop();
2893 push(_gvn.transform(new ModINode(control(), a, b)));
2894 break;
2895 case Bytecodes::_idiv:
2896 // Must keep both values on the expression-stack during null-check
2897 zero_check_int(peek());
2898 // Compile-time detect of null-exception?
2899 if (stopped()) return;
2900 b = pop();
2901 a = pop();
2902 push( _gvn.transform( new DivINode(control(),a,b) ) );
2903 break;
2904 case Bytecodes::_imul:
2905 b = pop(); a = pop();
2906 push( _gvn.transform( new MulINode(a,b) ) );
2907 break;
2908 case Bytecodes::_iadd:
2909 b = pop(); a = pop();
2910 push( _gvn.transform( new AddINode(a,b) ) );
2911 break;
2912 case Bytecodes::_ineg:
2913 a = pop();
2914 push( _gvn.transform( new SubINode(_gvn.intcon(0),a)) );
2915 break;
2916 case Bytecodes::_isub:
2917 b = pop(); a = pop();
2918 push( _gvn.transform( new SubINode(a,b) ) );
2919 break;
2920 case Bytecodes::_iand:
2921 b = pop(); a = pop();
2922 push( _gvn.transform( new AndINode(a,b) ) );
2923 break;
2924 case Bytecodes::_ior:
2925 b = pop(); a = pop();
2926 push( _gvn.transform( new OrINode(a,b) ) );
2927 break;
2928 case Bytecodes::_ixor:
2929 b = pop(); a = pop();
2930 push( _gvn.transform( new XorINode(a,b) ) );
2931 break;
2932 case Bytecodes::_ishl:
2933 b = pop(); a = pop();
2934 push( _gvn.transform( new LShiftINode(a,b) ) );
2935 break;
2936 case Bytecodes::_ishr:
2937 b = pop(); a = pop();
2938 push( _gvn.transform( new RShiftINode(a,b) ) );
2939 break;
2940 case Bytecodes::_iushr:
2941 b = pop(); a = pop();
2942 push( _gvn.transform( new URShiftINode(a,b) ) );
2943 break;
2944
2945 case Bytecodes::_fneg:
2946 a = pop();
2947 b = _gvn.transform(new NegFNode (a));
2948 push(b);
2949 break;
2950
2951 case Bytecodes::_fsub:
2952 b = pop();
2953 a = pop();
2954 c = _gvn.transform( new SubFNode(a,b) );
2955 d = precision_rounding(c);
2956 push( d );
2957 break;
2958
2959 case Bytecodes::_fadd:
2960 b = pop();
2961 a = pop();
2962 c = _gvn.transform( new AddFNode(a,b) );
2963 d = precision_rounding(c);
2964 push( d );
2965 break;
2966
2967 case Bytecodes::_fmul:
2968 b = pop();
2969 a = pop();
2970 c = _gvn.transform( new MulFNode(a,b) );
2971 d = precision_rounding(c);
2972 push( d );
2973 break;
2974
2975 case Bytecodes::_fdiv:
2976 b = pop();
2977 a = pop();
2978 c = _gvn.transform( new DivFNode(0,a,b) );
2979 d = precision_rounding(c);
2980 push( d );
2981 break;
2982
2983 case Bytecodes::_frem:
2984 if (Matcher::has_match_rule(Op_ModF)) {
2985 // Generate a ModF node.
2986 b = pop();
2987 a = pop();
2988 c = _gvn.transform( new ModFNode(0,a,b) );
2989 d = precision_rounding(c);
2990 push( d );
2991 }
2992 else {
2993 // Generate a call.
2994 modf();
2995 }
2996 break;
2997
2998 case Bytecodes::_fcmpl:
2999 b = pop();
3000 a = pop();
3001 c = _gvn.transform( new CmpF3Node( a, b));
3002 push(c);
3003 break;
3004 case Bytecodes::_fcmpg:
3005 b = pop();
3006 a = pop();
3007
3008 // Same as fcmpl but need to flip the unordered case. Swap the inputs,
3009 // which negates the result sign except for unordered. Flip the unordered
3010 // as well by using CmpF3 which implements unordered-lesser instead of
3011 // unordered-greater semantics. Finally, commute the result bits. Result
3012 // is same as using a CmpF3Greater except we did it with CmpF3 alone.
3013 c = _gvn.transform( new CmpF3Node( b, a));
3014 c = _gvn.transform( new SubINode(_gvn.intcon(0),c) );
3015 push(c);
3016 break;
3017
3018 case Bytecodes::_f2i:
3019 a = pop();
3020 push(_gvn.transform(new ConvF2INode(a)));
3021 break;
3022
3023 case Bytecodes::_d2i:
3024 a = pop_pair();
3025 b = _gvn.transform(new ConvD2INode(a));
3026 push( b );
3027 break;
3028
3029 case Bytecodes::_f2d:
3030 a = pop();
3031 b = _gvn.transform( new ConvF2DNode(a));
3032 push_pair( b );
3033 break;
3034
3035 case Bytecodes::_d2f:
3036 a = pop_pair();
3037 b = _gvn.transform( new ConvD2FNode(a));
3038 // This breaks _227_mtrt (speed & correctness) and _222_mpegaudio (speed)
3039 //b = _gvn.transform(new RoundFloatNode(0, b) );
3040 push( b );
3041 break;
3042
3043 case Bytecodes::_l2f:
3044 if (Matcher::convL2FSupported()) {
3045 a = pop_pair();
3046 b = _gvn.transform( new ConvL2FNode(a));
3047 // For x86_32.ad, FILD doesn't restrict precision to 24 or 53 bits.
3048 // Rather than storing the result into an FP register then pushing
3049 // out to memory to round, the machine instruction that implements
3050 // ConvL2D is responsible for rounding.
3051 // c = precision_rounding(b);
3052 c = _gvn.transform(b);
3053 push(c);
3054 } else {
3055 l2f();
3056 }
3057 break;
3058
3059 case Bytecodes::_l2d:
3060 a = pop_pair();
3061 b = _gvn.transform( new ConvL2DNode(a));
3062 // For x86_32.ad, rounding is always necessary (see _l2f above).
3063 // c = dprecision_rounding(b);
3064 c = _gvn.transform(b);
3065 push_pair(c);
3066 break;
3067
3068 case Bytecodes::_f2l:
3069 a = pop();
3070 b = _gvn.transform( new ConvF2LNode(a));
3071 push_pair(b);
3072 break;
3073
3074 case Bytecodes::_d2l:
3075 a = pop_pair();
3076 b = _gvn.transform( new ConvD2LNode(a));
3077 push_pair(b);
3078 break;
3079
3080 case Bytecodes::_dsub:
3081 b = pop_pair();
3082 a = pop_pair();
3083 c = _gvn.transform( new SubDNode(a,b) );
3084 d = dprecision_rounding(c);
3085 push_pair( d );
3086 break;
3087
3088 case Bytecodes::_dadd:
3089 b = pop_pair();
3090 a = pop_pair();
3091 c = _gvn.transform( new AddDNode(a,b) );
3092 d = dprecision_rounding(c);
3093 push_pair( d );
3094 break;
3095
3096 case Bytecodes::_dmul:
3097 b = pop_pair();
3098 a = pop_pair();
3099 c = _gvn.transform( new MulDNode(a,b) );
3100 d = dprecision_rounding(c);
3101 push_pair( d );
3102 break;
3103
3104 case Bytecodes::_ddiv:
3105 b = pop_pair();
3106 a = pop_pair();
3107 c = _gvn.transform( new DivDNode(0,a,b) );
3108 d = dprecision_rounding(c);
3109 push_pair( d );
3110 break;
3111
3112 case Bytecodes::_dneg:
3113 a = pop_pair();
3114 b = _gvn.transform(new NegDNode (a));
3115 push_pair(b);
3116 break;
3117
3118 case Bytecodes::_drem:
3119 if (Matcher::has_match_rule(Op_ModD)) {
3120 // Generate a ModD node.
3121 b = pop_pair();
3122 a = pop_pair();
3123 // a % b
3124
3125 c = _gvn.transform( new ModDNode(0,a,b) );
3126 d = dprecision_rounding(c);
3127 push_pair( d );
3128 }
3129 else {
3130 // Generate a call.
3131 modd();
3132 }
3133 break;
3134
3135 case Bytecodes::_dcmpl:
3136 b = pop_pair();
3137 a = pop_pair();
3138 c = _gvn.transform( new CmpD3Node( a, b));
3139 push(c);
3140 break;
3141
3142 case Bytecodes::_dcmpg:
3143 b = pop_pair();
3144 a = pop_pair();
3145 // Same as dcmpl but need to flip the unordered case.
3146 // Commute the inputs, which negates the result sign except for unordered.
3147 // Flip the unordered as well by using CmpD3 which implements
3148 // unordered-lesser instead of unordered-greater semantics.
3149 // Finally, negate the result bits. Result is same as using a
3150 // CmpD3Greater except we did it with CmpD3 alone.
3151 c = _gvn.transform( new CmpD3Node( b, a));
3152 c = _gvn.transform( new SubINode(_gvn.intcon(0),c) );
3153 push(c);
3154 break;
3155
3156
3157 // Note for longs -> lo word is on TOS, hi word is on TOS - 1
3158 case Bytecodes::_land:
3159 b = pop_pair();
3160 a = pop_pair();
3161 c = _gvn.transform( new AndLNode(a,b) );
3162 push_pair(c);
3163 break;
3164 case Bytecodes::_lor:
3165 b = pop_pair();
3166 a = pop_pair();
3167 c = _gvn.transform( new OrLNode(a,b) );
3168 push_pair(c);
3169 break;
3170 case Bytecodes::_lxor:
3171 b = pop_pair();
3172 a = pop_pair();
3173 c = _gvn.transform( new XorLNode(a,b) );
3174 push_pair(c);
3175 break;
3176
3177 case Bytecodes::_lshl:
3178 b = pop(); // the shift count
3179 a = pop_pair(); // value to be shifted
3180 c = _gvn.transform( new LShiftLNode(a,b) );
3181 push_pair(c);
3182 break;
3183 case Bytecodes::_lshr:
3184 b = pop(); // the shift count
3185 a = pop_pair(); // value to be shifted
3186 c = _gvn.transform( new RShiftLNode(a,b) );
3187 push_pair(c);
3188 break;
3189 case Bytecodes::_lushr:
3190 b = pop(); // the shift count
3191 a = pop_pair(); // value to be shifted
3192 c = _gvn.transform( new URShiftLNode(a,b) );
3193 push_pair(c);
3194 break;
3195 case Bytecodes::_lmul:
3196 b = pop_pair();
3197 a = pop_pair();
3198 c = _gvn.transform( new MulLNode(a,b) );
3199 push_pair(c);
3200 break;
3201
3202 case Bytecodes::_lrem:
3203 // Must keep both values on the expression-stack during null-check
3204 assert(peek(0) == top(), "long word order");
3205 zero_check_long(peek(1));
3206 // Compile-time detect of null-exception?
3207 if (stopped()) return;
3208 b = pop_pair();
3209 a = pop_pair();
3210 c = _gvn.transform( new ModLNode(control(),a,b) );
3211 push_pair(c);
3212 break;
3213
3214 case Bytecodes::_ldiv:
3215 // Must keep both values on the expression-stack during null-check
3216 assert(peek(0) == top(), "long word order");
3217 zero_check_long(peek(1));
3218 // Compile-time detect of null-exception?
3219 if (stopped()) return;
3220 b = pop_pair();
3221 a = pop_pair();
3222 c = _gvn.transform( new DivLNode(control(),a,b) );
3223 push_pair(c);
3224 break;
3225
3226 case Bytecodes::_ladd:
3227 b = pop_pair();
3228 a = pop_pair();
3229 c = _gvn.transform( new AddLNode(a,b) );
3230 push_pair(c);
3231 break;
3232 case Bytecodes::_lsub:
3233 b = pop_pair();
3234 a = pop_pair();
3235 c = _gvn.transform( new SubLNode(a,b) );
3236 push_pair(c);
3237 break;
3238 case Bytecodes::_lcmp:
3239 // Safepoints are now inserted _before_ branches. The long-compare
3240 // bytecode painfully produces a 3-way value (-1,0,+1) which requires a
3241 // slew of control flow. These are usually followed by a CmpI vs zero and
3242 // a branch; this pattern then optimizes to the obvious long-compare and
3243 // branch. However, if the branch is backwards there's a Safepoint
3244 // inserted. The inserted Safepoint captures the JVM state at the
3245 // pre-branch point, i.e. it captures the 3-way value. Thus if a
3246 // long-compare is used to control a loop the debug info will force
3247 // computation of the 3-way value, even though the generated code uses a
3248 // long-compare and branch. We try to rectify the situation by inserting
3249 // a SafePoint here and have it dominate and kill the safepoint added at a
3250 // following backwards branch. At this point the JVM state merely holds 2
3251 // longs but not the 3-way value.
3252 switch (iter().next_bc()) {
3253 case Bytecodes::_ifgt:
3254 case Bytecodes::_iflt:
3255 case Bytecodes::_ifge:
3256 case Bytecodes::_ifle:
3257 case Bytecodes::_ifne:
3258 case Bytecodes::_ifeq:
3259 // If this is a backwards branch in the bytecodes, add Safepoint
3260 maybe_add_safepoint(iter().next_get_dest());
3261 default:
3262 break;
3263 }
3264 b = pop_pair();
3265 a = pop_pair();
3266 c = _gvn.transform( new CmpL3Node( a, b ));
3267 push(c);
3268 break;
3269
3270 case Bytecodes::_lneg:
3271 a = pop_pair();
3272 b = _gvn.transform( new SubLNode(longcon(0),a));
3273 push_pair(b);
3274 break;
3275 case Bytecodes::_l2i:
3276 a = pop_pair();
3277 push( _gvn.transform( new ConvL2INode(a)));
3278 break;
3279 case Bytecodes::_i2l:
3280 a = pop();
3281 b = _gvn.transform( new ConvI2LNode(a));
3282 push_pair(b);
3283 break;
3284 case Bytecodes::_i2b:
3285 // Sign extend
3286 a = pop();
3287 a = Compile::narrow_value(T_BYTE, a, NULL, &_gvn, true);
3288 push(a);
3289 break;
3290 case Bytecodes::_i2s:
3291 a = pop();
3292 a = Compile::narrow_value(T_SHORT, a, NULL, &_gvn, true);
3293 push(a);
3294 break;
3295 case Bytecodes::_i2c:
3296 a = pop();
3297 a = Compile::narrow_value(T_CHAR, a, NULL, &_gvn, true);
3298 push(a);
3299 break;
3300
3301 case Bytecodes::_i2f:
3302 a = pop();
3303 b = _gvn.transform( new ConvI2FNode(a) ) ;
3304 c = precision_rounding(b);
3305 push (b);
3306 break;
3307
3308 case Bytecodes::_i2d:
3309 a = pop();
3310 b = _gvn.transform( new ConvI2DNode(a));
3311 push_pair(b);
3312 break;
3313
3314 case Bytecodes::_iinc: // Increment local
3315 i = iter().get_index(); // Get local index
3316 set_local( i, _gvn.transform( new AddINode( _gvn.intcon(iter().get_iinc_con()), local(i) ) ) );
3317 break;
3318
3319 // Exit points of synchronized methods must have an unlock node
3320 case Bytecodes::_return:
3321 return_current(NULL);
3322 break;
3323
3324 case Bytecodes::_ireturn:
3325 case Bytecodes::_areturn:
3326 case Bytecodes::_freturn:
3327 return_current(pop());
3328 break;
3329 case Bytecodes::_lreturn:
3330 return_current(pop_pair());
3331 break;
3332 case Bytecodes::_dreturn:
3333 return_current(pop_pair());
3334 break;
3335
3336 case Bytecodes::_athrow:
3337 // null exception oop throws NULL pointer exception
3338 null_check(peek());
3339 if (stopped()) return;
3340 // Hook the thrown exception directly to subsequent handlers.
3341 if (BailoutToInterpreterForThrows) {
3342 // Keep method interpreted from now on.
3343 uncommon_trap(Deoptimization::Reason_unhandled,
3344 Deoptimization::Action_make_not_compilable);
3345 return;
3346 }
3347 if (env()->jvmti_can_post_on_exceptions()) {
3348 // check if we must post exception events, take uncommon trap if so (with must_throw = false)
3349 uncommon_trap_if_should_post_on_exceptions(Deoptimization::Reason_unhandled, false);
3350 }
3351 // Here if either can_post_on_exceptions or should_post_on_exceptions is false
3352 add_exception_state(make_exception_state(peek()));
3353 break;
3354
3355 case Bytecodes::_goto: // fall through
3356 case Bytecodes::_goto_w: {
3357 int target_bci = (bc() == Bytecodes::_goto) ? iter().get_dest() : iter().get_far_dest();
3358
3359 // If this is a backwards branch in the bytecodes, add Safepoint
3360 maybe_add_safepoint(target_bci);
3361
3362 // Merge the current control into the target basic block
3363 merge(target_bci);
3364
3365 // See if we can get some profile data and hand it off to the next block
3366 Block *target_block = block()->successor_for_bci(target_bci);
3367 if (target_block->pred_count() != 1) break;
3368 ciMethodData* methodData = method()->method_data();
3369 if (!methodData->is_mature()) break;
3370 ciProfileData* data = methodData->bci_to_data(bci());
3371 assert(data != NULL && data->is_JumpData(), "need JumpData for taken branch");
3372 int taken = ((ciJumpData*)data)->taken();
3373 taken = method()->scale_count(taken);
3374 target_block->set_count(taken);
3375 break;
3376 }
3377
3378 case Bytecodes::_ifnull: btest = BoolTest::eq; goto handle_if_null;
3379 case Bytecodes::_ifnonnull: btest = BoolTest::ne; goto handle_if_null;
3380 handle_if_null:
3381 // If this is a backwards branch in the bytecodes, add Safepoint
3382 maybe_add_safepoint(iter().get_dest());
3383 a = null();
3384 b = pop();
3385 if (b->is_InlineType()) {
3386 // Return constant false because 'b' is always non-null
3387 c = _gvn.makecon(TypeInt::CC_GT);
3388 } else {
3389 if (!_gvn.type(b)->speculative_maybe_null() &&
3390 !too_many_traps(Deoptimization::Reason_speculate_null_check)) {
3391 inc_sp(1);
3392 Node* null_ctl = top();
3393 b = null_check_oop(b, &null_ctl, true, true, true);
3394 assert(null_ctl->is_top(), "no null control here");
3395 dec_sp(1);
3396 } else if (_gvn.type(b)->speculative_always_null() &&
3397 !too_many_traps(Deoptimization::Reason_speculate_null_assert)) {
3398 inc_sp(1);
3399 b = null_assert(b);
3400 dec_sp(1);
3401 }
3402 c = _gvn.transform( new CmpPNode(b, a) );
3403 }
3404 do_ifnull(btest, c);
3405 break;
3406
3407 case Bytecodes::_if_acmpeq: btest = BoolTest::eq; goto handle_if_acmp;
3408 case Bytecodes::_if_acmpne: btest = BoolTest::ne; goto handle_if_acmp;
3409 handle_if_acmp:
3410 // If this is a backwards branch in the bytecodes, add Safepoint
3411 maybe_add_safepoint(iter().get_dest());
3412 a = pop();
3413 b = pop();
3414 do_acmp(btest, b, a);
3415 break;
3416
3417 case Bytecodes::_ifeq: btest = BoolTest::eq; goto handle_ifxx;
3418 case Bytecodes::_ifne: btest = BoolTest::ne; goto handle_ifxx;
3419 case Bytecodes::_iflt: btest = BoolTest::lt; goto handle_ifxx;
3420 case Bytecodes::_ifle: btest = BoolTest::le; goto handle_ifxx;
3421 case Bytecodes::_ifgt: btest = BoolTest::gt; goto handle_ifxx;
3422 case Bytecodes::_ifge: btest = BoolTest::ge; goto handle_ifxx;
3423 handle_ifxx:
3424 // If this is a backwards branch in the bytecodes, add Safepoint
3425 maybe_add_safepoint(iter().get_dest());
3426 a = _gvn.intcon(0);
3427 b = pop();
3428 c = _gvn.transform( new CmpINode(b, a) );
3429 do_if(btest, c);
3430 break;
3431
3432 case Bytecodes::_if_icmpeq: btest = BoolTest::eq; goto handle_if_icmp;
3433 case Bytecodes::_if_icmpne: btest = BoolTest::ne; goto handle_if_icmp;
3434 case Bytecodes::_if_icmplt: btest = BoolTest::lt; goto handle_if_icmp;
3435 case Bytecodes::_if_icmple: btest = BoolTest::le; goto handle_if_icmp;
3436 case Bytecodes::_if_icmpgt: btest = BoolTest::gt; goto handle_if_icmp;
3437 case Bytecodes::_if_icmpge: btest = BoolTest::ge; goto handle_if_icmp;
3438 handle_if_icmp:
3439 // If this is a backwards branch in the bytecodes, add Safepoint
3440 maybe_add_safepoint(iter().get_dest());
3441 a = pop();
3442 b = pop();
3443 c = _gvn.transform( new CmpINode( b, a ) );
3444 do_if(btest, c);
3445 break;
3446
3447 case Bytecodes::_tableswitch:
3448 do_tableswitch();
3449 break;
3450
3451 case Bytecodes::_lookupswitch:
3452 do_lookupswitch();
3453 break;
3454
3455 case Bytecodes::_invokestatic:
3456 case Bytecodes::_invokedynamic:
3457 case Bytecodes::_invokespecial:
3458 case Bytecodes::_invokevirtual:
3459 case Bytecodes::_invokeinterface:
3460 do_call();
3461 break;
3462 case Bytecodes::_checkcast:
3463 do_checkcast();
3464 break;
3465 case Bytecodes::_instanceof:
3466 do_instanceof();
3467 break;
3468 case Bytecodes::_anewarray:
3469 do_newarray();
3470 break;
3471 case Bytecodes::_newarray:
3472 do_newarray((BasicType)iter().get_index());
3473 break;
3474 case Bytecodes::_multianewarray:
3475 do_multianewarray();
3476 break;
3477 case Bytecodes::_new:
3478 do_new();
3479 break;
3480 case Bytecodes::_defaultvalue:
3481 do_defaultvalue();
3482 break;
3483 case Bytecodes::_withfield:
3484 do_withfield();
3485 break;
3486
3487 case Bytecodes::_jsr:
3488 case Bytecodes::_jsr_w:
3489 do_jsr();
3490 break;
3491
3492 case Bytecodes::_ret:
3493 do_ret();
3494 break;
3495
3496
3497 case Bytecodes::_monitorenter:
3498 do_monitor_enter();
3499 break;
3500
3501 case Bytecodes::_monitorexit:
3502 do_monitor_exit();
3503 break;
3504
3505 case Bytecodes::_breakpoint:
3506 // Breakpoint set concurrently to compile
3507 // %%% use an uncommon trap?
3508 C->record_failure("breakpoint in method");
3509 return;
3510
3511 default:
3512 #ifndef PRODUCT
3513 map()->dump(99);
3514 #endif
3515 tty->print("\nUnhandled bytecode %s\n", Bytecodes::name(bc()) );
3516 ShouldNotReachHere();
3517 }
3518
3519 #ifndef PRODUCT
3520 if (C->should_print(1)) {
3521 IdealGraphPrinter* printer = C->printer();
3522 char buffer[256];
3523 jio_snprintf(buffer, sizeof(buffer), "Bytecode %d: %s", bci(), Bytecodes::name(bc()));
3524 bool old = printer->traverse_outs();
3525 printer->set_traverse_outs(true);
3526 printer->print_method(buffer, 4);
3527 printer->set_traverse_outs(old);
3528 }
3529 #endif
3530 }