1 /*
2 * Copyright (c) 1998, 2025, 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 "ci/ciMethodData.hpp"
27 #include "ci/ciSymbols.hpp"
28 #include "classfile/vmSymbols.hpp"
29 #include "compiler/compileLog.hpp"
30 #include "interpreter/linkResolver.hpp"
31 #include "jvm_io.h"
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 uint 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 = nullptr;
60 ciKlass* element_type = nullptr;
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 != nullptr || 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* array_index = pop();
80 Node* array = pop();
81
82 // Handle inline type arrays
83 const TypeOopPtr* element_ptr = elemtype->make_oopptr();
84 const TypeAryPtr* array_type = _gvn.type(array)->is_aryptr();
85 if (array_type->is_flat()) {
86 // Load from flat inline type array
87 Node* inline_type;
88 if (element_ptr->klass_is_exact()) {
89 inline_type = InlineTypeNode::make_from_flat(this, elemtype->inline_klass(), array, adr);
90 } else {
91 // Element type of flat array is not exact. Therefore, we cannot determine the flat array layout statically.
92 // Emit a runtime call to load the element from the flat array.
93 inline_type = load_from_unknown_flat_array(array, array_index, element_ptr);
94 inline_type = record_profile_for_speculation_at_array_load(inline_type);
95 }
96 push(inline_type);
97 return;
98 }
99
100 if (!array_type->is_not_flat()) {
101 // Cannot statically determine if array is a flat array, emit runtime check
102 assert(UseFlatArray && is_reference_type(bt) && element_ptr->can_be_inline_type() && !array_type->is_not_null_free() &&
103 (!element_ptr->is_inlinetypeptr() || element_ptr->inline_klass()->flat_in_array()), "array can't be flat");
104 IdealKit ideal(this);
105 IdealVariable res(ideal);
106 ideal.declarations_done();
107 ideal.if_then(flat_array_test(array, /* flat = */ false)); {
108 // Non-flat array
109 assert(ideal.ctrl()->in(0)->as_If()->is_flat_array_check(&_gvn), "Should be found");
110 sync_kit(ideal);
111 const TypeAryPtr* adr_type = TypeAryPtr::get_array_body_type(bt);
112 DecoratorSet decorator_set = IN_HEAP | IS_ARRAY | C2_CONTROL_DEPENDENT_LOAD;
113 if (needs_range_check(array_type->size(), array_index)) {
114 // We've emitted a RangeCheck but now insert an additional check between the range check and the actual load.
115 // We cannot pin the load to two separate nodes. Instead, we pin it conservatively here such that it cannot
116 // possibly float above the range check at any point.
117 decorator_set |= C2_UNKNOWN_CONTROL_LOAD;
118 }
119 Node* ld = access_load_at(array, adr, adr_type, element_ptr, bt, decorator_set);
120 if (element_ptr->is_inlinetypeptr()) {
121 assert(element_ptr->maybe_null(), "null free array should be handled above");
122 ld = InlineTypeNode::make_from_oop(this, ld, element_ptr->inline_klass(), false);
123 }
124 ideal.sync_kit(this);
125 ideal.set(res, ld);
126 } ideal.else_(); {
127 // Flat array
128 sync_kit(ideal);
129 if (element_ptr->is_inlinetypeptr()) {
130 // Element type is known, cast and load from flat array layout.
131 ciInlineKlass* vk = element_ptr->inline_klass();
132 assert(vk->flat_in_array() && element_ptr->maybe_null(), "never/always flat - should be optimized");
133 ciArrayKlass* array_klass = ciArrayKlass::make(vk, /* null_free */ true);
134 const TypeAryPtr* arytype = TypeOopPtr::make_from_klass(array_klass)->isa_aryptr();
135 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, arytype));
136 Node* casted_adr = array_element_address(cast, array_index, T_OBJECT, array_type->size(), control());
137 // Re-execute flat array load if buffering triggers deoptimization
138 PreserveReexecuteState preexecs(this);
139 jvms()->set_should_reexecute(true);
140 inc_sp(2);
141 Node* vt = InlineTypeNode::make_from_flat(this, vk, cast, casted_adr)->buffer(this, false);
142 ideal.set(res, vt);
143 ideal.sync_kit(this);
144 } else {
145 // Element type is unknown, and thus we cannot statically determine the exact flat array layout. Emit a
146 // runtime call to correctly load the inline type element from the flat array.
147 Node* inline_type = load_from_unknown_flat_array(array, array_index, element_ptr);
148 ideal.sync_kit(this);
149 ideal.set(res, inline_type);
150 }
151 } ideal.end_if();
152 sync_kit(ideal);
153 Node* ld = _gvn.transform(ideal.value(res));
154 ld = record_profile_for_speculation_at_array_load(ld);
155 push_node(bt, ld);
156 return;
157 }
158
159 if (array_type->is_null_free()) {
160 // Load from non-flat inline type array (elements can never be null)
161 bt = T_OBJECT;
162 }
163
164 if (elemtype == TypeInt::BOOL) {
165 bt = T_BOOLEAN;
166 }
167 const TypeAryPtr* adr_type = TypeAryPtr::get_array_body_type(bt);
168 Node* ld = access_load_at(array, adr, adr_type, elemtype, bt,
169 IN_HEAP | IS_ARRAY | C2_CONTROL_DEPENDENT_LOAD);
170 ld = record_profile_for_speculation_at_array_load(ld);
171 // Loading an inline type from a non-flat array
172 if (element_ptr != nullptr && element_ptr->is_inlinetypeptr()) {
173 assert(!array_type->is_null_free() || !element_ptr->maybe_null(), "inline type array elements should never be null");
174 ld = InlineTypeNode::make_from_oop(this, ld, element_ptr->inline_klass(), !element_ptr->maybe_null());
175 }
176 push_node(bt, ld);
177 }
178
179 Node* Parse::load_from_unknown_flat_array(Node* array, Node* array_index, const TypeOopPtr* element_ptr) {
180 // Below membars keep this access to an unknown flat array correctly
181 // ordered with other unknown and known flat array accesses.
182 insert_mem_bar_volatile(Op_MemBarCPUOrder, C->get_alias_index(TypeAryPtr::INLINES));
183
184 Node* call = nullptr;
185 {
186 // Re-execute flat array load if runtime call triggers deoptimization
187 PreserveReexecuteState preexecs(this);
188 jvms()->set_bci(_bci);
189 jvms()->set_should_reexecute(true);
190 inc_sp(2);
191 kill_dead_locals();
192 call = make_runtime_call(RC_NO_LEAF | RC_NO_IO,
193 OptoRuntime::load_unknown_inline_Type(),
194 OptoRuntime::load_unknown_inline_Java(),
195 nullptr, TypeRawPtr::BOTTOM,
196 array, array_index);
197 }
198 make_slow_call_ex(call, env()->Throwable_klass(), false);
199 Node* buffer = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
200
201 insert_mem_bar_volatile(Op_MemBarCPUOrder, C->get_alias_index(TypeAryPtr::INLINES));
202
203 // Keep track of the information that the inline type is in flat arrays
204 const Type* unknown_value = element_ptr->is_instptr()->cast_to_flat_in_array();
205 return _gvn.transform(new CheckCastPPNode(control(), buffer, unknown_value));
206 }
207
208 //--------------------------------array_store----------------------------------
209 void Parse::array_store(BasicType bt) {
210 const Type* elemtype = Type::TOP;
211 Node* adr = array_addressing(bt, type2size[bt], elemtype);
212 if (stopped()) return; // guaranteed null or range check
213 Node* stored_value_casted = nullptr;
214 if (bt == T_OBJECT) {
215 stored_value_casted = array_store_check(adr, elemtype);
216 if (stopped()) {
217 return;
218 }
219 }
220 Node* const stored_value = pop_node(bt); // Value to store
221 Node* const array_index = pop(); // Index in the array
222 Node* array = pop(); // The array itself
223
224 const TypeAryPtr* array_type = _gvn.type(array)->is_aryptr();
225 const TypeAryPtr* adr_type = TypeAryPtr::get_array_body_type(bt);
226
227 if (elemtype == TypeInt::BOOL) {
228 bt = T_BOOLEAN;
229 } else if (bt == T_OBJECT) {
230 elemtype = elemtype->make_oopptr();
231 const Type* stored_value_casted_type = _gvn.type(stored_value_casted);
232 // Based on the value to be stored, try to determine if the array is not null-free and/or not flat.
233 // This is only legal for non-null stores because the array_store_check always passes for null, even
234 // if the array is null-free. Null stores are handled in GraphKit::inline_array_null_guard().
235 bool not_null_free = !stored_value_casted_type->maybe_null() &&
236 !stored_value_casted_type->is_oopptr()->can_be_inline_type();
237 bool not_flat = not_null_free || (stored_value_casted_type->is_inlinetypeptr() &&
238 !stored_value_casted_type->inline_klass()->flat_in_array());
239 if (!array_type->is_not_null_free() && not_null_free) {
240 // Storing a non-inline type, mark array as not null-free (-> not flat).
241 array_type = array_type->cast_to_not_null_free();
242 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, array_type));
243 replace_in_map(array, cast);
244 array = cast;
245 } else if (!array_type->is_not_flat() && not_flat) {
246 // Storing to a non-flat array, mark array as not flat.
247 array_type = array_type->cast_to_not_flat();
248 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, array_type));
249 replace_in_map(array, cast);
250 array = cast;
251 }
252
253 if (array_type->is_flat()) {
254 // Store to flat inline type array
255 assert(!stored_value_casted_type->maybe_null(), "should be guaranteed by array store check");
256 if (array_type->klass_is_exact()) {
257 // Store to exact flat inline type array where we know the flat array layout statically.
258 // Re-execute flat array store if buffering triggers deoptimization
259 PreserveReexecuteState preexecs(this);
260 inc_sp(3);
261 jvms()->set_should_reexecute(true);
262 stored_value_casted->as_InlineType()->store_flat(this, array, adr, nullptr, 0, MO_UNORDERED | IN_HEAP | IS_ARRAY);
263 } else {
264 // Element type of flat array is not exact. Therefore, we cannot determine the flat array layout statically.
265 // Emit a runtime call to store the element to the flat array.
266 store_to_unknown_flat_array(array, array_index, stored_value_casted);
267 }
268 return;
269 }
270 if (array_type->is_null_free()) {
271 // Store to non-flat null-free inline type array (elements can never be null)
272 assert(!stored_value_casted_type->maybe_null(), "should be guaranteed by array store check");
273 if (elemtype->inline_klass()->is_empty()) {
274 // Ignore empty inline stores, array is already initialized.
275 return;
276 }
277 } else if (!array_type->is_not_flat() && (stored_value_casted_type != TypePtr::NULL_PTR || StressReflectiveCode)) {
278 // Array might be a flat array, emit runtime checks (for nullptr, a simple inline_array_null_guard is sufficient).
279 assert(UseFlatArray && !not_flat && elemtype->is_oopptr()->can_be_inline_type() &&
280 !array_type->klass_is_exact() && !array_type->is_not_null_free(), "array can't be a flat array");
281 IdealKit ideal(this);
282 ideal.if_then(flat_array_test(array, /* flat = */ false)); {
283 // Non-flat array
284 assert(ideal.ctrl()->in(0)->as_If()->is_flat_array_check(&_gvn), "Should be found");
285 sync_kit(ideal);
286 Node* cast_array = inline_array_null_guard(array, stored_value_casted, 3);
287 inc_sp(3);
288 access_store_at(cast_array, adr, adr_type, stored_value_casted, elemtype, bt, MO_UNORDERED | IN_HEAP | IS_ARRAY, false);
289 dec_sp(3);
290 ideal.sync_kit(this);
291 } ideal.else_(); {
292 sync_kit(ideal);
293 // flat array
294 Node* null_ctl = top();
295 Node* null_checked_stored_value_casted = null_check_oop(stored_value_casted, &null_ctl);
296 if (null_ctl != top()) {
297 PreserveJVMState pjvms(this);
298 inc_sp(3);
299 set_control(null_ctl);
300 uncommon_trap(Deoptimization::Reason_null_check, Deoptimization::Action_none);
301 dec_sp(3);
302 }
303 // Try to determine the inline klass
304 ciInlineKlass* inline_Klass = nullptr;
305 if (stored_value_casted_type->is_inlinetypeptr()) {
306 inline_Klass = stored_value_casted_type->inline_klass();
307 } else if (elemtype->is_inlinetypeptr()) {
308 inline_Klass = elemtype->inline_klass();
309 }
310 if (!stopped()) {
311 if (inline_Klass != nullptr) {
312 // Element type is known, cast and store to flat array layout.
313 assert(inline_Klass->flat_in_array() && elemtype->maybe_null(), "never/always flat - should be optimized");
314 ciArrayKlass* array_klass = ciArrayKlass::make(inline_Klass, /* null_free */ true);
315 const TypeAryPtr* arytype = TypeOopPtr::make_from_klass(array_klass)->isa_aryptr();
316 Node* casted_array = _gvn.transform(new CheckCastPPNode(control(), array, arytype));
317 Node* casted_adr = array_element_address(casted_array, array_index, T_OBJECT, arytype->size(), control());
318 if (!null_checked_stored_value_casted->is_InlineType()) {
319 assert(!gvn().type(null_checked_stored_value_casted)->maybe_null(),
320 "inline type array elements should never be null");
321 null_checked_stored_value_casted = InlineTypeNode::make_from_oop(this, null_checked_stored_value_casted,
322 inline_Klass);
323 }
324 // Re-execute flat array store if buffering triggers deoptimization
325 PreserveReexecuteState preexecs(this);
326 inc_sp(3);
327 jvms()->set_should_reexecute(true);
328 null_checked_stored_value_casted->as_InlineType()->store_flat(this, casted_array, casted_adr, nullptr, 0, MO_UNORDERED | IN_HEAP | IS_ARRAY);
329 } else {
330 // Element type is unknown, emit a runtime call since the flat array layout is not statically known.
331 store_to_unknown_flat_array(array, array_index, null_checked_stored_value_casted);
332 }
333 }
334 ideal.sync_kit(this);
335 }
336 ideal.end_if();
337 sync_kit(ideal);
338 return;
339 } else if (!array_type->is_not_null_free()) {
340 // Array is not flat but may be null free
341 assert(elemtype->is_oopptr()->can_be_inline_type(), "array can't be null-free");
342 array = inline_array_null_guard(array, stored_value_casted, 3, true);
343 }
344 }
345 inc_sp(3);
346 access_store_at(array, adr, adr_type, stored_value, elemtype, bt, MO_UNORDERED | IN_HEAP | IS_ARRAY);
347 dec_sp(3);
348 }
349
350 // Emit a runtime call to store to a flat array whose element type is either unknown (i.e. we do not know the flat
351 // array layout) or not exact (could have different flat array layouts at runtime).
352 void Parse::store_to_unknown_flat_array(Node* array, Node* const idx, Node* non_null_stored_value) {
353 // Below membars keep this access to an unknown flat array correctly
354 // ordered with other unknown and known flat array accesses.
355 insert_mem_bar_volatile(Op_MemBarCPUOrder, C->get_alias_index(TypeAryPtr::INLINES));
356
357 Node* call = nullptr;
358 {
359 // Re-execute flat array store if runtime call triggers deoptimization
360 PreserveReexecuteState preexecs(this);
361 jvms()->set_bci(_bci);
362 jvms()->set_should_reexecute(true);
363 inc_sp(3);
364 kill_dead_locals();
365 call = make_runtime_call(RC_NO_LEAF | RC_NO_IO,
366 OptoRuntime::store_unknown_inline_Type(),
367 OptoRuntime::store_unknown_inline_Java(),
368 nullptr, TypeRawPtr::BOTTOM,
369 non_null_stored_value, array, idx);
370 }
371 make_slow_call_ex(call, env()->Throwable_klass(), false);
372
373 insert_mem_bar_volatile(Op_MemBarCPUOrder, C->get_alias_index(TypeAryPtr::INLINES));
374 }
375
376 //------------------------------array_addressing-------------------------------
377 // Pull array and index from the stack. Compute pointer-to-element.
378 Node* Parse::array_addressing(BasicType type, int vals, const Type*& elemtype) {
379 Node *idx = peek(0+vals); // Get from stack without popping
380 Node *ary = peek(1+vals); // in case of exception
381
382 // Null check the array base, with correct stack contents
383 ary = null_check(ary, T_ARRAY);
384 // Compile-time detect of null-exception?
385 if (stopped()) return top();
386
387 const TypeAryPtr* arytype = _gvn.type(ary)->is_aryptr();
388 const TypeInt* sizetype = arytype->size();
389 elemtype = arytype->elem();
390
391 if (UseUniqueSubclasses) {
392 const Type* el = elemtype->make_ptr();
393 if (el && el->isa_instptr()) {
394 const TypeInstPtr* toop = el->is_instptr();
395 if (toop->instance_klass()->unique_concrete_subklass()) {
396 // If we load from "AbstractClass[]" we must see "ConcreteSubClass".
397 const Type* subklass = Type::get_const_type(toop->instance_klass());
398 elemtype = subklass->join_speculative(el);
399 }
400 }
401 }
402
403 if (!arytype->is_loaded()) {
404 // Only fails for some -Xcomp runs
405 // The class is unloaded. We have to run this bytecode in the interpreter.
406 ciKlass* klass = arytype->unloaded_klass();
407
408 uncommon_trap(Deoptimization::Reason_unloaded,
409 Deoptimization::Action_reinterpret,
410 klass, "!loaded array");
411 return top();
412 }
413
414 ary = create_speculative_inline_type_array_checks(ary, arytype, elemtype);
415
416 if (needs_range_check(sizetype, idx)) {
417 create_range_check(idx, ary, sizetype);
418 } else if (C->log() != nullptr) {
419 C->log()->elem("observe that='!need_range_check'");
420 }
421
422 // Check for always knowing you are throwing a range-check exception
423 if (stopped()) return top();
424
425 // Make array address computation control dependent to prevent it
426 // from floating above the range check during loop optimizations.
427 Node* ptr = array_element_address(ary, idx, type, sizetype, control());
428 assert(ptr != top(), "top should go hand-in-hand with stopped");
429
430 return ptr;
431 }
432
433 // Check if we need a range check for an array access. This is the case if the index is either negative or if it could
434 // be greater or equal the smallest possible array size (i.e. out-of-bounds).
435 bool Parse::needs_range_check(const TypeInt* size_type, const Node* index) const {
436 const TypeInt* index_type = _gvn.type(index)->is_int();
437 return index_type->_hi >= size_type->_lo || index_type->_lo < 0;
438 }
439
440 void Parse::create_range_check(Node* idx, Node* ary, const TypeInt* sizetype) {
441 Node* tst;
442 if (sizetype->_hi <= 0) {
443 // The greatest array bound is negative, so we can conclude that we're
444 // compiling unreachable code, but the unsigned compare trick used below
445 // only works with non-negative lengths. Instead, hack "tst" to be zero so
446 // the uncommon_trap path will always be taken.
447 tst = _gvn.intcon(0);
448 } else {
449 // Range is constant in array-oop, so we can use the original state of mem
450 Node* len = load_array_length(ary);
451
452 // Test length vs index (standard trick using unsigned compare)
453 Node* chk = _gvn.transform(new CmpUNode(idx, len) );
454 BoolTest::mask btest = BoolTest::lt;
455 tst = _gvn.transform(new BoolNode(chk, btest) );
456 }
457 RangeCheckNode* rc = new RangeCheckNode(control(), tst, PROB_MAX, COUNT_UNKNOWN);
458 _gvn.set_type(rc, rc->Value(&_gvn));
459 if (!tst->is_Con()) {
460 record_for_igvn(rc);
461 }
462 set_control(_gvn.transform(new IfTrueNode(rc)));
463 // Branch to failure if out of bounds
464 {
465 PreserveJVMState pjvms(this);
466 set_control(_gvn.transform(new IfFalseNode(rc)));
467 if (C->allow_range_check_smearing()) {
468 // Do not use builtin_throw, since range checks are sometimes
469 // made more stringent by an optimistic transformation.
470 // This creates "tentative" range checks at this point,
471 // which are not guaranteed to throw exceptions.
472 // See IfNode::Ideal, is_range_check, adjust_check.
473 uncommon_trap(Deoptimization::Reason_range_check,
474 Deoptimization::Action_make_not_entrant,
475 nullptr, "range_check");
476 } else {
477 // If we have already recompiled with the range-check-widening
478 // heroic optimization turned off, then we must really be throwing
479 // range check exceptions.
480 builtin_throw(Deoptimization::Reason_range_check);
481 }
482 }
483 }
484
485 // For inline type arrays, we can use the profiling information for array accesses to speculate on the type, flatness,
486 // and null-freeness. We can either prepare the speculative type for later uses or emit explicit speculative checks with
487 // traps now. In the latter case, the speculative type guarantees can avoid additional runtime checks later (e.g.
488 // non-null-free implies non-flat which allows us to remove flatness checks). This makes the graph simpler.
489 Node* Parse::create_speculative_inline_type_array_checks(Node* array, const TypeAryPtr* array_type,
490 const Type*& element_type) {
491 if (!array_type->is_flat() && !array_type->is_not_flat()) {
492 // For arrays that might be flat, speculate that the array has the exact type reported in the profile data such that
493 // we can rely on a fixed memory layout (i.e. either a flat layout or not).
494 array = cast_to_speculative_array_type(array, array_type, element_type);
495 } else if (UseTypeSpeculation && UseArrayLoadStoreProfile) {
496 // Array is known to be either flat or not flat. If possible, update the speculative type by using the profile data
497 // at this bci.
498 array = cast_to_profiled_array_type(array);
499 }
500
501 // Even though the type does not tell us whether we have an inline type array or not, we can still check the profile data
502 // whether we have a non-null-free or non-flat array. Since non-null-free implies non-flat, we check this first.
503 // Speculating on a non-null-free array doesn't help aaload but could be profitable for a subsequent aastore.
504 if (!array_type->is_null_free() && !array_type->is_not_null_free()) {
505 array = speculate_non_null_free_array(array, array_type);
506 }
507
508 if (!array_type->is_flat() && !array_type->is_not_flat()) {
509 array = speculate_non_flat_array(array, array_type);
510 }
511 return array;
512 }
513
514 // Speculate that the array has the exact type reported in the profile data. We emit a trap when this turns out to be
515 // wrong. On the fast path, we add a CheckCastPP to use the exact type.
516 Node* Parse::cast_to_speculative_array_type(Node* const array, const TypeAryPtr*& array_type, const Type*& element_type) {
517 Deoptimization::DeoptReason reason = Deoptimization::Reason_speculate_class_check;
518 ciKlass* speculative_array_type = array_type->speculative_type();
519 if (too_many_traps_or_recompiles(reason) || speculative_array_type == nullptr) {
520 // No speculative type, check profile data at this bci
521 speculative_array_type = nullptr;
522 reason = Deoptimization::Reason_class_check;
523 if (UseArrayLoadStoreProfile && !too_many_traps_or_recompiles(reason)) {
524 ciKlass* profiled_element_type = nullptr;
525 ProfilePtrKind element_ptr = ProfileMaybeNull;
526 bool flat_array = true;
527 bool null_free_array = true;
528 method()->array_access_profiled_type(bci(), speculative_array_type, profiled_element_type, element_ptr, flat_array,
529 null_free_array);
530 }
531 }
532 if (speculative_array_type != nullptr) {
533 // Speculate that this array has the exact type reported by profile data
534 Node* casted_array = nullptr;
535 DEBUG_ONLY(Node* old_control = control();)
536 Node* slow_ctl = type_check_receiver(array, speculative_array_type, 1.0, &casted_array);
537 if (stopped()) {
538 // The check always fails and therefore profile information is incorrect. Don't use it.
539 assert(old_control == slow_ctl, "type check should have been removed");
540 set_control(slow_ctl);
541 } else if (!slow_ctl->is_top()) {
542 { PreserveJVMState pjvms(this);
543 set_control(slow_ctl);
544 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
545 }
546 replace_in_map(array, casted_array);
547 array_type = _gvn.type(casted_array)->is_aryptr();
548 element_type = array_type->elem();
549 return casted_array;
550 }
551 }
552 return array;
553 }
554
555 // Create a CheckCastPP when the speculative type can improve the current type.
556 Node* Parse::cast_to_profiled_array_type(Node* const array) {
557 ciKlass* array_type = nullptr;
558 ciKlass* element_type = nullptr;
559 ProfilePtrKind element_ptr = ProfileMaybeNull;
560 bool flat_array = true;
561 bool null_free_array = true;
562 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
563 if (array_type != nullptr) {
564 return record_profile_for_speculation(array, array_type, ProfileMaybeNull);
565 }
566 return array;
567 }
568
569 // Speculate that the array is non-null-free. This will imply non-flatness. We emit a trap when this turns out to be
570 // wrong. On the fast path, we add a CheckCastPP to use the non-null-free type.
571 Node* Parse::speculate_non_null_free_array(Node* const array, const TypeAryPtr*& array_type) {
572 bool null_free_array = true;
573 Deoptimization::DeoptReason reason = Deoptimization::Reason_none;
574 if (array_type->speculative() != nullptr &&
575 array_type->speculative()->is_aryptr()->is_not_null_free() &&
576 !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_class_check)) {
577 null_free_array = false;
578 reason = Deoptimization::Reason_speculate_class_check;
579 } else if (UseArrayLoadStoreProfile && !too_many_traps_or_recompiles(Deoptimization::Reason_class_check)) {
580 ciKlass* profiled_array_type = nullptr;
581 ciKlass* profiled_element_type = nullptr;
582 ProfilePtrKind element_ptr = ProfileMaybeNull;
583 bool flat_array = true;
584 method()->array_access_profiled_type(bci(), profiled_array_type, profiled_element_type, element_ptr, flat_array,
585 null_free_array);
586 reason = Deoptimization::Reason_class_check;
587 }
588 if (!null_free_array) {
589 { // Deoptimize if null-free array
590 BuildCutout unless(this, null_free_array_test(array, /* null_free = */ false), PROB_MAX);
591 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
592 }
593 assert(!stopped(), "null-free array should have been caught earlier");
594 Node* casted_array = _gvn.transform(new CheckCastPPNode(control(), array, array_type->cast_to_not_null_free()));
595 replace_in_map(array, casted_array);
596 array_type = _gvn.type(casted_array)->is_aryptr();
597 return casted_array;
598 }
599 return array;
600 }
601
602 // Speculate that the array is non-flat. We emit a trap when this turns out to be wrong. On the fast path, we add a
603 // CheckCastPP to use the non-flat type.
604 Node* Parse::speculate_non_flat_array(Node* const array, const TypeAryPtr* const array_type) {
605 bool flat_array = true;
606 Deoptimization::DeoptReason reason = Deoptimization::Reason_none;
607 if (array_type->speculative() != nullptr &&
608 array_type->speculative()->is_aryptr()->is_not_flat() &&
609 !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_class_check)) {
610 flat_array = false;
611 reason = Deoptimization::Reason_speculate_class_check;
612 } else if (UseArrayLoadStoreProfile && !too_many_traps_or_recompiles(reason)) {
613 ciKlass* profiled_array_type = nullptr;
614 ciKlass* profiled_element_type = nullptr;
615 ProfilePtrKind element_ptr = ProfileMaybeNull;
616 bool null_free_array = true;
617 method()->array_access_profiled_type(bci(), profiled_array_type, profiled_element_type, element_ptr, flat_array,
618 null_free_array);
619 reason = Deoptimization::Reason_class_check;
620 }
621 if (!flat_array) {
622 { // Deoptimize if flat array
623 BuildCutout unless(this, flat_array_test(array, /* flat = */ false), PROB_MAX);
624 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
625 }
626 assert(!stopped(), "flat array should have been caught earlier");
627 Node* casted_array = _gvn.transform(new CheckCastPPNode(control(), array, array_type->cast_to_not_flat()));
628 replace_in_map(array, casted_array);
629 return casted_array;
630 }
631 return array;
632 }
633
634 // returns IfNode
635 IfNode* Parse::jump_if_fork_int(Node* a, Node* b, BoolTest::mask mask, float prob, float cnt) {
636 Node *cmp = _gvn.transform(new CmpINode(a, b)); // two cases: shiftcount > 32 and shiftcount <= 32
637 Node *tst = _gvn.transform(new BoolNode(cmp, mask));
638 IfNode *iff = create_and_map_if(control(), tst, prob, cnt);
639 return iff;
640 }
641
642
643 // sentinel value for the target bci to mark never taken branches
644 // (according to profiling)
645 static const int never_reached = INT_MAX;
646
647 //------------------------------helper for tableswitch-------------------------
648 void Parse::jump_if_true_fork(IfNode *iff, int dest_bci_if_true, bool unc) {
649 // True branch, use existing map info
650 { PreserveJVMState pjvms(this);
651 Node *iftrue = _gvn.transform( new IfTrueNode (iff) );
652 set_control( iftrue );
653 if (unc) {
654 repush_if_args();
655 uncommon_trap(Deoptimization::Reason_unstable_if,
656 Deoptimization::Action_reinterpret,
657 nullptr,
658 "taken always");
659 } else {
660 assert(dest_bci_if_true != never_reached, "inconsistent dest");
661 merge_new_path(dest_bci_if_true);
662 }
663 }
664
665 // False branch
666 Node *iffalse = _gvn.transform( new IfFalseNode(iff) );
667 set_control( iffalse );
668 }
669
670 void Parse::jump_if_false_fork(IfNode *iff, int dest_bci_if_true, bool unc) {
671 // True branch, use existing map info
672 { PreserveJVMState pjvms(this);
673 Node *iffalse = _gvn.transform( new IfFalseNode (iff) );
674 set_control( iffalse );
675 if (unc) {
676 repush_if_args();
677 uncommon_trap(Deoptimization::Reason_unstable_if,
678 Deoptimization::Action_reinterpret,
679 nullptr,
680 "taken never");
681 } else {
682 assert(dest_bci_if_true != never_reached, "inconsistent dest");
683 merge_new_path(dest_bci_if_true);
684 }
685 }
686
687 // False branch
688 Node *iftrue = _gvn.transform( new IfTrueNode(iff) );
689 set_control( iftrue );
690 }
691
692 void Parse::jump_if_always_fork(int dest_bci, bool unc) {
693 // False branch, use existing map and control()
694 if (unc) {
695 repush_if_args();
696 uncommon_trap(Deoptimization::Reason_unstable_if,
697 Deoptimization::Action_reinterpret,
698 nullptr,
699 "taken never");
700 } else {
701 assert(dest_bci != never_reached, "inconsistent dest");
702 merge_new_path(dest_bci);
703 }
704 }
705
706
707 extern "C" {
708 static int jint_cmp(const void *i, const void *j) {
709 int a = *(jint *)i;
710 int b = *(jint *)j;
711 return a > b ? 1 : a < b ? -1 : 0;
712 }
713 }
714
715
716 class SwitchRange : public StackObj {
717 // a range of integers coupled with a bci destination
718 jint _lo; // inclusive lower limit
719 jint _hi; // inclusive upper limit
720 int _dest;
721 float _cnt; // how many times this range was hit according to profiling
722
723 public:
724 jint lo() const { return _lo; }
725 jint hi() const { return _hi; }
726 int dest() const { return _dest; }
727 bool is_singleton() const { return _lo == _hi; }
728 float cnt() const { return _cnt; }
729
730 void setRange(jint lo, jint hi, int dest, float cnt) {
731 assert(lo <= hi, "must be a non-empty range");
732 _lo = lo, _hi = hi; _dest = dest; _cnt = cnt;
733 assert(_cnt >= 0, "");
734 }
735 bool adjoinRange(jint lo, jint hi, int dest, float cnt, bool trim_ranges) {
736 assert(lo <= hi, "must be a non-empty range");
737 if (lo == _hi+1) {
738 // see merge_ranges() comment below
739 if (trim_ranges) {
740 if (cnt == 0) {
741 if (_cnt != 0) {
742 return false;
743 }
744 if (dest != _dest) {
745 _dest = never_reached;
746 }
747 } else {
748 if (_cnt == 0) {
749 return false;
750 }
751 if (dest != _dest) {
752 return false;
753 }
754 }
755 } else {
756 if (dest != _dest) {
757 return false;
758 }
759 }
760 _hi = hi;
761 _cnt += cnt;
762 return true;
763 }
764 return false;
765 }
766
767 void set (jint value, int dest, float cnt) {
768 setRange(value, value, dest, cnt);
769 }
770 bool adjoin(jint value, int dest, float cnt, bool trim_ranges) {
771 return adjoinRange(value, value, dest, cnt, trim_ranges);
772 }
773 bool adjoin(SwitchRange& other) {
774 return adjoinRange(other._lo, other._hi, other._dest, other._cnt, false);
775 }
776
777 void print() {
778 if (is_singleton())
779 tty->print(" {%d}=>%d (cnt=%f)", lo(), dest(), cnt());
780 else if (lo() == min_jint)
781 tty->print(" {..%d}=>%d (cnt=%f)", hi(), dest(), cnt());
782 else if (hi() == max_jint)
783 tty->print(" {%d..}=>%d (cnt=%f)", lo(), dest(), cnt());
784 else
785 tty->print(" {%d..%d}=>%d (cnt=%f)", lo(), hi(), dest(), cnt());
786 }
787 };
788
789 // We try to minimize the number of ranges and the size of the taken
790 // ones using profiling data. When ranges are created,
791 // SwitchRange::adjoinRange() only allows 2 adjoining ranges to merge
792 // if both were never hit or both were hit to build longer unreached
793 // ranges. Here, we now merge adjoining ranges with the same
794 // destination and finally set destination of unreached ranges to the
795 // special value never_reached because it can help minimize the number
796 // of tests that are necessary.
797 //
798 // For instance:
799 // [0, 1] to target1 sometimes taken
800 // [1, 2] to target1 never taken
801 // [2, 3] to target2 never taken
802 // would lead to:
803 // [0, 1] to target1 sometimes taken
804 // [1, 3] never taken
805 //
806 // (first 2 ranges to target1 are not merged)
807 static void merge_ranges(SwitchRange* ranges, int& rp) {
808 if (rp == 0) {
809 return;
810 }
811 int shift = 0;
812 for (int j = 0; j < rp; j++) {
813 SwitchRange& r1 = ranges[j-shift];
814 SwitchRange& r2 = ranges[j+1];
815 if (r1.adjoin(r2)) {
816 shift++;
817 } else if (shift > 0) {
818 ranges[j+1-shift] = r2;
819 }
820 }
821 rp -= shift;
822 for (int j = 0; j <= rp; j++) {
823 SwitchRange& r = ranges[j];
824 if (r.cnt() == 0 && r.dest() != never_reached) {
825 r.setRange(r.lo(), r.hi(), never_reached, r.cnt());
826 }
827 }
828 }
829
830 //-------------------------------do_tableswitch--------------------------------
831 void Parse::do_tableswitch() {
832 // Get information about tableswitch
833 int default_dest = iter().get_dest_table(0);
834 jint lo_index = iter().get_int_table(1);
835 jint hi_index = iter().get_int_table(2);
836 int len = hi_index - lo_index + 1;
837
838 if (len < 1) {
839 // If this is a backward branch, add safepoint
840 maybe_add_safepoint(default_dest);
841 pop(); // the effect of the instruction execution on the operand stack
842 merge(default_dest);
843 return;
844 }
845
846 ciMethodData* methodData = method()->method_data();
847 ciMultiBranchData* profile = nullptr;
848 if (methodData->is_mature() && UseSwitchProfiling) {
849 ciProfileData* data = methodData->bci_to_data(bci());
850 if (data != nullptr && data->is_MultiBranchData()) {
851 profile = (ciMultiBranchData*)data;
852 }
853 }
854 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if);
855
856 // generate decision tree, using trichotomy when possible
857 int rnum = len+2;
858 bool makes_backward_branch = (default_dest <= bci());
859 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum);
860 int rp = -1;
861 if (lo_index != min_jint) {
862 float cnt = 1.0F;
863 if (profile != nullptr) {
864 cnt = (float)profile->default_count() / (hi_index != max_jint ? 2.0F : 1.0F);
865 }
866 ranges[++rp].setRange(min_jint, lo_index-1, default_dest, cnt);
867 }
868 for (int j = 0; j < len; j++) {
869 jint match_int = lo_index+j;
870 int dest = iter().get_dest_table(j+3);
871 makes_backward_branch |= (dest <= bci());
872 float cnt = 1.0F;
873 if (profile != nullptr) {
874 cnt = (float)profile->count_at(j);
875 }
876 if (rp < 0 || !ranges[rp].adjoin(match_int, dest, cnt, trim_ranges)) {
877 ranges[++rp].set(match_int, dest, cnt);
878 }
879 }
880 jint highest = lo_index+(len-1);
881 assert(ranges[rp].hi() == highest, "");
882 if (highest != max_jint) {
883 float cnt = 1.0F;
884 if (profile != nullptr) {
885 cnt = (float)profile->default_count() / (lo_index != min_jint ? 2.0F : 1.0F);
886 }
887 if (!ranges[rp].adjoinRange(highest+1, max_jint, default_dest, cnt, trim_ranges)) {
888 ranges[++rp].setRange(highest+1, max_jint, default_dest, cnt);
889 }
890 }
891 assert(rp < len+2, "not too many ranges");
892
893 if (trim_ranges) {
894 merge_ranges(ranges, rp);
895 }
896
897 // Safepoint in case if backward branch observed
898 if (makes_backward_branch) {
899 add_safepoint();
900 }
901
902 Node* lookup = pop(); // lookup value
903 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]);
904 }
905
906
907 //------------------------------do_lookupswitch--------------------------------
908 void Parse::do_lookupswitch() {
909 // Get information about lookupswitch
910 int default_dest = iter().get_dest_table(0);
911 jint len = iter().get_int_table(1);
912
913 if (len < 1) { // If this is a backward branch, add safepoint
914 maybe_add_safepoint(default_dest);
915 pop(); // the effect of the instruction execution on the operand stack
916 merge(default_dest);
917 return;
918 }
919
920 ciMethodData* methodData = method()->method_data();
921 ciMultiBranchData* profile = nullptr;
922 if (methodData->is_mature() && UseSwitchProfiling) {
923 ciProfileData* data = methodData->bci_to_data(bci());
924 if (data != nullptr && data->is_MultiBranchData()) {
925 profile = (ciMultiBranchData*)data;
926 }
927 }
928 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if);
929
930 // generate decision tree, using trichotomy when possible
931 jint* table = NEW_RESOURCE_ARRAY(jint, len*3);
932 {
933 for (int j = 0; j < len; j++) {
934 table[3*j+0] = iter().get_int_table(2+2*j);
935 table[3*j+1] = iter().get_dest_table(2+2*j+1);
936 // Handle overflow when converting from uint to jint
937 table[3*j+2] = (profile == nullptr) ? 1 : (jint)MIN2<uint>((uint)max_jint, profile->count_at(j));
938 }
939 qsort(table, len, 3*sizeof(table[0]), jint_cmp);
940 }
941
942 float default_cnt = 1.0F;
943 if (profile != nullptr) {
944 juint defaults = max_juint - len;
945 default_cnt = (float)profile->default_count()/(float)defaults;
946 }
947
948 int rnum = len*2+1;
949 bool makes_backward_branch = (default_dest <= bci());
950 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum);
951 int rp = -1;
952 for (int j = 0; j < len; j++) {
953 jint match_int = table[3*j+0];
954 jint dest = table[3*j+1];
955 jint cnt = table[3*j+2];
956 jint next_lo = rp < 0 ? min_jint : ranges[rp].hi()+1;
957 makes_backward_branch |= (dest <= bci());
958 float c = default_cnt * ((float)match_int - (float)next_lo);
959 if (match_int != next_lo && (rp < 0 || !ranges[rp].adjoinRange(next_lo, match_int-1, default_dest, c, trim_ranges))) {
960 assert(default_dest != never_reached, "sentinel value for dead destinations");
961 ranges[++rp].setRange(next_lo, match_int-1, default_dest, c);
962 }
963 if (rp < 0 || !ranges[rp].adjoin(match_int, dest, (float)cnt, trim_ranges)) {
964 assert(dest != never_reached, "sentinel value for dead destinations");
965 ranges[++rp].set(match_int, dest, (float)cnt);
966 }
967 }
968 jint highest = table[3*(len-1)];
969 assert(ranges[rp].hi() == highest, "");
970 if (highest != max_jint &&
971 !ranges[rp].adjoinRange(highest+1, max_jint, default_dest, default_cnt * ((float)max_jint - (float)highest), trim_ranges)) {
972 ranges[++rp].setRange(highest+1, max_jint, default_dest, default_cnt * ((float)max_jint - (float)highest));
973 }
974 assert(rp < rnum, "not too many ranges");
975
976 if (trim_ranges) {
977 merge_ranges(ranges, rp);
978 }
979
980 // Safepoint in case backward branch observed
981 if (makes_backward_branch) {
982 add_safepoint();
983 }
984
985 Node *lookup = pop(); // lookup value
986 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]);
987 }
988
989 static float if_prob(float taken_cnt, float total_cnt) {
990 assert(taken_cnt <= total_cnt, "");
991 if (total_cnt == 0) {
992 return PROB_FAIR;
993 }
994 float p = taken_cnt / total_cnt;
995 return clamp(p, PROB_MIN, PROB_MAX);
996 }
997
998 static float if_cnt(float cnt) {
999 if (cnt == 0) {
1000 return COUNT_UNKNOWN;
1001 }
1002 return cnt;
1003 }
1004
1005 static float sum_of_cnts(SwitchRange *lo, SwitchRange *hi) {
1006 float total_cnt = 0;
1007 for (SwitchRange* sr = lo; sr <= hi; sr++) {
1008 total_cnt += sr->cnt();
1009 }
1010 return total_cnt;
1011 }
1012
1013 class SwitchRanges : public ResourceObj {
1014 public:
1015 SwitchRange* _lo;
1016 SwitchRange* _hi;
1017 SwitchRange* _mid;
1018 float _cost;
1019
1020 enum {
1021 Start,
1022 LeftDone,
1023 RightDone,
1024 Done
1025 } _state;
1026
1027 SwitchRanges(SwitchRange *lo, SwitchRange *hi)
1028 : _lo(lo), _hi(hi), _mid(nullptr),
1029 _cost(0), _state(Start) {
1030 }
1031
1032 SwitchRanges()
1033 : _lo(nullptr), _hi(nullptr), _mid(nullptr),
1034 _cost(0), _state(Start) {}
1035 };
1036
1037 // Estimate cost of performing a binary search on lo..hi
1038 static float compute_tree_cost(SwitchRange *lo, SwitchRange *hi, float total_cnt) {
1039 GrowableArray<SwitchRanges> tree;
1040 SwitchRanges root(lo, hi);
1041 tree.push(root);
1042
1043 float cost = 0;
1044 do {
1045 SwitchRanges& r = *tree.adr_at(tree.length()-1);
1046 if (r._hi != r._lo) {
1047 if (r._mid == nullptr) {
1048 float r_cnt = sum_of_cnts(r._lo, r._hi);
1049
1050 if (r_cnt == 0) {
1051 tree.pop();
1052 cost = 0;
1053 continue;
1054 }
1055
1056 SwitchRange* mid = nullptr;
1057 mid = r._lo;
1058 for (float cnt = 0; ; ) {
1059 assert(mid <= r._hi, "out of bounds");
1060 cnt += mid->cnt();
1061 if (cnt > r_cnt / 2) {
1062 break;
1063 }
1064 mid++;
1065 }
1066 assert(mid <= r._hi, "out of bounds");
1067 r._mid = mid;
1068 r._cost = r_cnt / total_cnt;
1069 }
1070 r._cost += cost;
1071 if (r._state < SwitchRanges::LeftDone && r._mid > r._lo) {
1072 cost = 0;
1073 r._state = SwitchRanges::LeftDone;
1074 tree.push(SwitchRanges(r._lo, r._mid-1));
1075 } else if (r._state < SwitchRanges::RightDone) {
1076 cost = 0;
1077 r._state = SwitchRanges::RightDone;
1078 tree.push(SwitchRanges(r._mid == r._lo ? r._mid+1 : r._mid, r._hi));
1079 } else {
1080 tree.pop();
1081 cost = r._cost;
1082 }
1083 } else {
1084 tree.pop();
1085 cost = r._cost;
1086 }
1087 } while (tree.length() > 0);
1088
1089
1090 return cost;
1091 }
1092
1093 // It sometimes pays off to test most common ranges before the binary search
1094 void Parse::linear_search_switch_ranges(Node* key_val, SwitchRange*& lo, SwitchRange*& hi) {
1095 uint nr = hi - lo + 1;
1096 float total_cnt = sum_of_cnts(lo, hi);
1097
1098 float min = compute_tree_cost(lo, hi, total_cnt);
1099 float extra = 1;
1100 float sub = 0;
1101
1102 SwitchRange* array1 = lo;
1103 SwitchRange* array2 = NEW_RESOURCE_ARRAY(SwitchRange, nr);
1104
1105 SwitchRange* ranges = nullptr;
1106
1107 while (nr >= 2) {
1108 assert(lo == array1 || lo == array2, "one the 2 already allocated arrays");
1109 ranges = (lo == array1) ? array2 : array1;
1110
1111 // Find highest frequency range
1112 SwitchRange* candidate = lo;
1113 for (SwitchRange* sr = lo+1; sr <= hi; sr++) {
1114 if (sr->cnt() > candidate->cnt()) {
1115 candidate = sr;
1116 }
1117 }
1118 SwitchRange most_freq = *candidate;
1119 if (most_freq.cnt() == 0) {
1120 break;
1121 }
1122
1123 // Copy remaining ranges into another array
1124 int shift = 0;
1125 for (uint i = 0; i < nr; i++) {
1126 SwitchRange* sr = &lo[i];
1127 if (sr != candidate) {
1128 ranges[i-shift] = *sr;
1129 } else {
1130 shift++;
1131 if (i > 0 && i < nr-1) {
1132 SwitchRange prev = lo[i-1];
1133 prev.setRange(prev.lo(), sr->hi(), prev.dest(), prev.cnt());
1134 if (prev.adjoin(lo[i+1])) {
1135 shift++;
1136 i++;
1137 }
1138 ranges[i-shift] = prev;
1139 }
1140 }
1141 }
1142 nr -= shift;
1143
1144 // Evaluate cost of testing the most common range and performing a
1145 // binary search on the other ranges
1146 float cost = extra + compute_tree_cost(&ranges[0], &ranges[nr-1], total_cnt);
1147 if (cost >= min) {
1148 break;
1149 }
1150 // swap arrays
1151 lo = &ranges[0];
1152 hi = &ranges[nr-1];
1153
1154 // It pays off: emit the test for the most common range
1155 assert(most_freq.cnt() > 0, "must be taken");
1156 Node* val = _gvn.transform(new SubINode(key_val, _gvn.intcon(most_freq.lo())));
1157 Node* cmp = _gvn.transform(new CmpUNode(val, _gvn.intcon(java_subtract(most_freq.hi(), most_freq.lo()))));
1158 Node* tst = _gvn.transform(new BoolNode(cmp, BoolTest::le));
1159 IfNode* iff = create_and_map_if(control(), tst, if_prob(most_freq.cnt(), total_cnt), if_cnt(most_freq.cnt()));
1160 jump_if_true_fork(iff, most_freq.dest(), false);
1161
1162 sub += most_freq.cnt() / total_cnt;
1163 extra += 1 - sub;
1164 min = cost;
1165 }
1166 }
1167
1168 //----------------------------create_jump_tables-------------------------------
1169 bool Parse::create_jump_tables(Node* key_val, SwitchRange* lo, SwitchRange* hi) {
1170 // Are jumptables enabled
1171 if (!UseJumpTables) return false;
1172
1173 // Are jumptables supported
1174 if (!Matcher::has_match_rule(Op_Jump)) return false;
1175
1176 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if);
1177
1178 // Decide if a guard is needed to lop off big ranges at either (or
1179 // both) end(s) of the input set. We'll call this the default target
1180 // even though we can't be sure that it is the true "default".
1181
1182 bool needs_guard = false;
1183 int default_dest;
1184 int64_t total_outlier_size = 0;
1185 int64_t hi_size = ((int64_t)hi->hi()) - ((int64_t)hi->lo()) + 1;
1186 int64_t lo_size = ((int64_t)lo->hi()) - ((int64_t)lo->lo()) + 1;
1187
1188 if (lo->dest() == hi->dest()) {
1189 total_outlier_size = hi_size + lo_size;
1190 default_dest = lo->dest();
1191 } else if (lo_size > hi_size) {
1192 total_outlier_size = lo_size;
1193 default_dest = lo->dest();
1194 } else {
1195 total_outlier_size = hi_size;
1196 default_dest = hi->dest();
1197 }
1198
1199 float total = sum_of_cnts(lo, hi);
1200 float cost = compute_tree_cost(lo, hi, total);
1201
1202 // If a guard test will eliminate very sparse end ranges, then
1203 // it is worth the cost of an extra jump.
1204 float trimmed_cnt = 0;
1205 if (total_outlier_size > (MaxJumpTableSparseness * 4)) {
1206 needs_guard = true;
1207 if (default_dest == lo->dest()) {
1208 trimmed_cnt += lo->cnt();
1209 lo++;
1210 }
1211 if (default_dest == hi->dest()) {
1212 trimmed_cnt += hi->cnt();
1213 hi--;
1214 }
1215 }
1216
1217 // Find the total number of cases and ranges
1218 int64_t num_cases = ((int64_t)hi->hi()) - ((int64_t)lo->lo()) + 1;
1219 int num_range = hi - lo + 1;
1220
1221 // Don't create table if: too large, too small, or too sparse.
1222 if (num_cases > MaxJumpTableSize)
1223 return false;
1224 if (UseSwitchProfiling) {
1225 // MinJumpTableSize is set so with a well balanced binary tree,
1226 // when the number of ranges is MinJumpTableSize, it's cheaper to
1227 // go through a JumpNode that a tree of IfNodes. Average cost of a
1228 // tree of IfNodes with MinJumpTableSize is
1229 // log2f(MinJumpTableSize) comparisons. So if the cost computed
1230 // from profile data is less than log2f(MinJumpTableSize) then
1231 // going with the binary search is cheaper.
1232 if (cost < log2f(MinJumpTableSize)) {
1233 return false;
1234 }
1235 } else {
1236 if (num_cases < MinJumpTableSize)
1237 return false;
1238 }
1239 if (num_cases > (MaxJumpTableSparseness * num_range))
1240 return false;
1241
1242 // Normalize table lookups to zero
1243 int lowval = lo->lo();
1244 key_val = _gvn.transform( new SubINode(key_val, _gvn.intcon(lowval)) );
1245
1246 // Generate a guard to protect against input keyvals that aren't
1247 // in the switch domain.
1248 if (needs_guard) {
1249 Node* size = _gvn.intcon(num_cases);
1250 Node* cmp = _gvn.transform(new CmpUNode(key_val, size));
1251 Node* tst = _gvn.transform(new BoolNode(cmp, BoolTest::ge));
1252 IfNode* iff = create_and_map_if(control(), tst, if_prob(trimmed_cnt, total), if_cnt(trimmed_cnt));
1253 jump_if_true_fork(iff, default_dest, trim_ranges && trimmed_cnt == 0);
1254
1255 total -= trimmed_cnt;
1256 }
1257
1258 // Create an ideal node JumpTable that has projections
1259 // of all possible ranges for a switch statement
1260 // The key_val input must be converted to a pointer offset and scaled.
1261 // Compare Parse::array_addressing above.
1262
1263 // Clean the 32-bit int into a real 64-bit offset.
1264 // Otherwise, the jint value 0 might turn into an offset of 0x0800000000.
1265 // Make I2L conversion control dependent to prevent it from
1266 // floating above the range check during loop optimizations.
1267 // Do not use a narrow int type here to prevent the data path from dying
1268 // while the control path is not removed. This can happen if the type of key_val
1269 // is later known to be out of bounds of [0, num_cases] and therefore a narrow cast
1270 // would be replaced by TOP while C2 is not able to fold the corresponding range checks.
1271 // Set _carry_dependency for the cast to avoid being removed by IGVN.
1272 #ifdef _LP64
1273 key_val = C->constrained_convI2L(&_gvn, key_val, TypeInt::INT, control(), true /* carry_dependency */);
1274 #endif
1275
1276 // Shift the value by wordsize so we have an index into the table, rather
1277 // than a switch value
1278 Node *shiftWord = _gvn.MakeConX(wordSize);
1279 key_val = _gvn.transform( new MulXNode( key_val, shiftWord));
1280
1281 // Create the JumpNode
1282 Arena* arena = C->comp_arena();
1283 float* probs = (float*)arena->Amalloc(sizeof(float)*num_cases);
1284 int i = 0;
1285 if (total == 0) {
1286 for (SwitchRange* r = lo; r <= hi; r++) {
1287 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) {
1288 probs[i] = 1.0F / num_cases;
1289 }
1290 }
1291 } else {
1292 for (SwitchRange* r = lo; r <= hi; r++) {
1293 float prob = r->cnt()/total;
1294 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) {
1295 probs[i] = prob / (r->hi() - r->lo() + 1);
1296 }
1297 }
1298 }
1299
1300 ciMethodData* methodData = method()->method_data();
1301 ciMultiBranchData* profile = nullptr;
1302 if (methodData->is_mature()) {
1303 ciProfileData* data = methodData->bci_to_data(bci());
1304 if (data != nullptr && data->is_MultiBranchData()) {
1305 profile = (ciMultiBranchData*)data;
1306 }
1307 }
1308
1309 Node* jtn = _gvn.transform(new JumpNode(control(), key_val, num_cases, probs, profile == nullptr ? COUNT_UNKNOWN : total));
1310
1311 // These are the switch destinations hanging off the jumpnode
1312 i = 0;
1313 for (SwitchRange* r = lo; r <= hi; r++) {
1314 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) {
1315 Node* input = _gvn.transform(new JumpProjNode(jtn, i, r->dest(), (int)(j - lowval)));
1316 {
1317 PreserveJVMState pjvms(this);
1318 set_control(input);
1319 jump_if_always_fork(r->dest(), trim_ranges && r->cnt() == 0);
1320 }
1321 }
1322 }
1323 assert(i == num_cases, "miscount of cases");
1324 stop_and_kill_map(); // no more uses for this JVMS
1325 return true;
1326 }
1327
1328 //----------------------------jump_switch_ranges-------------------------------
1329 void Parse::jump_switch_ranges(Node* key_val, SwitchRange *lo, SwitchRange *hi, int switch_depth) {
1330 Block* switch_block = block();
1331 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if);
1332
1333 if (switch_depth == 0) {
1334 // Do special processing for the top-level call.
1335 assert(lo->lo() == min_jint, "initial range must exhaust Type::INT");
1336 assert(hi->hi() == max_jint, "initial range must exhaust Type::INT");
1337
1338 // Decrement pred-numbers for the unique set of nodes.
1339 #ifdef ASSERT
1340 if (!trim_ranges) {
1341 // Ensure that the block's successors are a (duplicate-free) set.
1342 int successors_counted = 0; // block occurrences in [hi..lo]
1343 int unique_successors = switch_block->num_successors();
1344 for (int i = 0; i < unique_successors; i++) {
1345 Block* target = switch_block->successor_at(i);
1346
1347 // Check that the set of successors is the same in both places.
1348 int successors_found = 0;
1349 for (SwitchRange* p = lo; p <= hi; p++) {
1350 if (p->dest() == target->start()) successors_found++;
1351 }
1352 assert(successors_found > 0, "successor must be known");
1353 successors_counted += successors_found;
1354 }
1355 assert(successors_counted == (hi-lo)+1, "no unexpected successors");
1356 }
1357 #endif
1358
1359 // Maybe prune the inputs, based on the type of key_val.
1360 jint min_val = min_jint;
1361 jint max_val = max_jint;
1362 const TypeInt* ti = key_val->bottom_type()->isa_int();
1363 if (ti != nullptr) {
1364 min_val = ti->_lo;
1365 max_val = ti->_hi;
1366 assert(min_val <= max_val, "invalid int type");
1367 }
1368 while (lo->hi() < min_val) {
1369 lo++;
1370 }
1371 if (lo->lo() < min_val) {
1372 lo->setRange(min_val, lo->hi(), lo->dest(), lo->cnt());
1373 }
1374 while (hi->lo() > max_val) {
1375 hi--;
1376 }
1377 if (hi->hi() > max_val) {
1378 hi->setRange(hi->lo(), max_val, hi->dest(), hi->cnt());
1379 }
1380
1381 linear_search_switch_ranges(key_val, lo, hi);
1382 }
1383
1384 #ifndef PRODUCT
1385 if (switch_depth == 0) {
1386 _max_switch_depth = 0;
1387 _est_switch_depth = log2i_graceful((hi - lo + 1) - 1) + 1;
1388 }
1389 #endif
1390
1391 assert(lo <= hi, "must be a non-empty set of ranges");
1392 if (lo == hi) {
1393 jump_if_always_fork(lo->dest(), trim_ranges && lo->cnt() == 0);
1394 } else {
1395 assert(lo->hi() == (lo+1)->lo()-1, "contiguous ranges");
1396 assert(hi->lo() == (hi-1)->hi()+1, "contiguous ranges");
1397
1398 if (create_jump_tables(key_val, lo, hi)) return;
1399
1400 SwitchRange* mid = nullptr;
1401 float total_cnt = sum_of_cnts(lo, hi);
1402
1403 int nr = hi - lo + 1;
1404 if (UseSwitchProfiling) {
1405 // Don't keep the binary search tree balanced: pick up mid point
1406 // that split frequencies in half.
1407 float cnt = 0;
1408 for (SwitchRange* sr = lo; sr <= hi; sr++) {
1409 cnt += sr->cnt();
1410 if (cnt >= total_cnt / 2) {
1411 mid = sr;
1412 break;
1413 }
1414 }
1415 } else {
1416 mid = lo + nr/2;
1417
1418 // if there is an easy choice, pivot at a singleton:
1419 if (nr > 3 && !mid->is_singleton() && (mid-1)->is_singleton()) mid--;
1420
1421 assert(lo < mid && mid <= hi, "good pivot choice");
1422 assert(nr != 2 || mid == hi, "should pick higher of 2");
1423 assert(nr != 3 || mid == hi-1, "should pick middle of 3");
1424 }
1425
1426
1427 Node *test_val = _gvn.intcon(mid == lo ? mid->hi() : mid->lo());
1428
1429 if (mid->is_singleton()) {
1430 IfNode *iff_ne = jump_if_fork_int(key_val, test_val, BoolTest::ne, 1-if_prob(mid->cnt(), total_cnt), if_cnt(mid->cnt()));
1431 jump_if_false_fork(iff_ne, mid->dest(), trim_ranges && mid->cnt() == 0);
1432
1433 // Special Case: If there are exactly three ranges, and the high
1434 // and low range each go to the same place, omit the "gt" test,
1435 // since it will not discriminate anything.
1436 bool eq_test_only = (hi == lo+2 && hi->dest() == lo->dest() && mid == hi-1) || mid == lo;
1437
1438 // if there is a higher range, test for it and process it:
1439 if (mid < hi && !eq_test_only) {
1440 // two comparisons of same values--should enable 1 test for 2 branches
1441 // Use BoolTest::lt instead of BoolTest::gt
1442 float cnt = sum_of_cnts(lo, mid-1);
1443 IfNode *iff_lt = jump_if_fork_int(key_val, test_val, BoolTest::lt, if_prob(cnt, total_cnt), if_cnt(cnt));
1444 Node *iftrue = _gvn.transform( new IfTrueNode(iff_lt) );
1445 Node *iffalse = _gvn.transform( new IfFalseNode(iff_lt) );
1446 { PreserveJVMState pjvms(this);
1447 set_control(iffalse);
1448 jump_switch_ranges(key_val, mid+1, hi, switch_depth+1);
1449 }
1450 set_control(iftrue);
1451 }
1452
1453 } else {
1454 // mid is a range, not a singleton, so treat mid..hi as a unit
1455 float cnt = sum_of_cnts(mid == lo ? mid+1 : mid, hi);
1456 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));
1457
1458 // if there is a higher range, test for it and process it:
1459 if (mid == hi) {
1460 jump_if_true_fork(iff_ge, mid->dest(), trim_ranges && cnt == 0);
1461 } else {
1462 Node *iftrue = _gvn.transform( new IfTrueNode(iff_ge) );
1463 Node *iffalse = _gvn.transform( new IfFalseNode(iff_ge) );
1464 { PreserveJVMState pjvms(this);
1465 set_control(iftrue);
1466 jump_switch_ranges(key_val, mid == lo ? mid+1 : mid, hi, switch_depth+1);
1467 }
1468 set_control(iffalse);
1469 }
1470 }
1471
1472 // in any case, process the lower range
1473 if (mid == lo) {
1474 if (mid->is_singleton()) {
1475 jump_switch_ranges(key_val, lo+1, hi, switch_depth+1);
1476 } else {
1477 jump_if_always_fork(lo->dest(), trim_ranges && lo->cnt() == 0);
1478 }
1479 } else {
1480 jump_switch_ranges(key_val, lo, mid-1, switch_depth+1);
1481 }
1482 }
1483
1484 // Decrease pred_count for each successor after all is done.
1485 if (switch_depth == 0) {
1486 int unique_successors = switch_block->num_successors();
1487 for (int i = 0; i < unique_successors; i++) {
1488 Block* target = switch_block->successor_at(i);
1489 // Throw away the pre-allocated path for each unique successor.
1490 target->next_path_num();
1491 }
1492 }
1493
1494 #ifndef PRODUCT
1495 _max_switch_depth = MAX2(switch_depth, _max_switch_depth);
1496 if (TraceOptoParse && Verbose && WizardMode && switch_depth == 0) {
1497 SwitchRange* r;
1498 int nsing = 0;
1499 for( r = lo; r <= hi; r++ ) {
1500 if( r->is_singleton() ) nsing++;
1501 }
1502 tty->print(">>> ");
1503 _method->print_short_name();
1504 tty->print_cr(" switch decision tree");
1505 tty->print_cr(" %d ranges (%d singletons), max_depth=%d, est_depth=%d",
1506 (int) (hi-lo+1), nsing, _max_switch_depth, _est_switch_depth);
1507 if (_max_switch_depth > _est_switch_depth) {
1508 tty->print_cr("******** BAD SWITCH DEPTH ********");
1509 }
1510 tty->print(" ");
1511 for( r = lo; r <= hi; r++ ) {
1512 r->print();
1513 }
1514 tty->cr();
1515 }
1516 #endif
1517 }
1518
1519 void Parse::modf() {
1520 Node *f2 = pop();
1521 Node *f1 = pop();
1522 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::modf_Type(),
1523 CAST_FROM_FN_PTR(address, SharedRuntime::frem),
1524 "frem", nullptr, //no memory effects
1525 f1, f2);
1526 Node* res = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0));
1527
1528 push(res);
1529 }
1530
1531 void Parse::modd() {
1532 Node *d2 = pop_pair();
1533 Node *d1 = pop_pair();
1534 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::Math_DD_D_Type(),
1535 CAST_FROM_FN_PTR(address, SharedRuntime::drem),
1536 "drem", nullptr, //no memory effects
1537 d1, top(), d2, top());
1538 Node* res_d = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0));
1539
1540 #ifdef ASSERT
1541 Node* res_top = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 1));
1542 assert(res_top == top(), "second value must be top");
1543 #endif
1544
1545 push_pair(res_d);
1546 }
1547
1548 void Parse::l2f() {
1549 Node* f2 = pop();
1550 Node* f1 = pop();
1551 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::l2f_Type(),
1552 CAST_FROM_FN_PTR(address, SharedRuntime::l2f),
1553 "l2f", nullptr, //no memory effects
1554 f1, f2);
1555 Node* res = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0));
1556
1557 push(res);
1558 }
1559
1560 // Handle jsr and jsr_w bytecode
1561 void Parse::do_jsr() {
1562 assert(bc() == Bytecodes::_jsr || bc() == Bytecodes::_jsr_w, "wrong bytecode");
1563
1564 // Store information about current state, tagged with new _jsr_bci
1565 int return_bci = iter().next_bci();
1566 int jsr_bci = (bc() == Bytecodes::_jsr) ? iter().get_dest() : iter().get_far_dest();
1567
1568 // The way we do things now, there is only one successor block
1569 // for the jsr, because the target code is cloned by ciTypeFlow.
1570 Block* target = successor_for_bci(jsr_bci);
1571
1572 // What got pushed?
1573 const Type* ret_addr = target->peek();
1574 assert(ret_addr->singleton(), "must be a constant (cloned jsr body)");
1575
1576 // Effect on jsr on stack
1577 push(_gvn.makecon(ret_addr));
1578
1579 // Flow to the jsr.
1580 merge(jsr_bci);
1581 }
1582
1583 // Handle ret bytecode
1584 void Parse::do_ret() {
1585 // Find to whom we return.
1586 assert(block()->num_successors() == 1, "a ret can only go one place now");
1587 Block* target = block()->successor_at(0);
1588 assert(!target->is_ready(), "our arrival must be expected");
1589 int pnum = target->next_path_num();
1590 merge_common(target, pnum);
1591 }
1592
1593 static bool has_injected_profile(BoolTest::mask btest, Node* test, int& taken, int& not_taken) {
1594 if (btest != BoolTest::eq && btest != BoolTest::ne) {
1595 // Only ::eq and ::ne are supported for profile injection.
1596 return false;
1597 }
1598 if (test->is_Cmp() &&
1599 test->in(1)->Opcode() == Op_ProfileBoolean) {
1600 ProfileBooleanNode* profile = (ProfileBooleanNode*)test->in(1);
1601 int false_cnt = profile->false_count();
1602 int true_cnt = profile->true_count();
1603
1604 // Counts matching depends on the actual test operation (::eq or ::ne).
1605 // No need to scale the counts because profile injection was designed
1606 // to feed exact counts into VM.
1607 taken = (btest == BoolTest::eq) ? false_cnt : true_cnt;
1608 not_taken = (btest == BoolTest::eq) ? true_cnt : false_cnt;
1609
1610 profile->consume();
1611 return true;
1612 }
1613 return false;
1614 }
1615
1616 // Give up if too few (or too many, in which case the sum will overflow) counts to be meaningful.
1617 // We also check that individual counters are positive first, otherwise the sum can become positive.
1618 // (check for saturation, integer overflow, and immature counts)
1619 static bool counters_are_meaningful(int counter1, int counter2, int min) {
1620 // check for saturation, including "uint" values too big to fit in "int"
1621 if (counter1 < 0 || counter2 < 0) {
1622 return false;
1623 }
1624 // check for integer overflow of the sum
1625 int64_t sum = (int64_t)counter1 + (int64_t)counter2;
1626 STATIC_ASSERT(sizeof(counter1) < sizeof(sum));
1627 if (sum > INT_MAX) {
1628 return false;
1629 }
1630 // check if mature
1631 return (counter1 + counter2) >= min;
1632 }
1633
1634 //--------------------------dynamic_branch_prediction--------------------------
1635 // Try to gather dynamic branch prediction behavior. Return a probability
1636 // of the branch being taken and set the "cnt" field. Returns a -1.0
1637 // if we need to use static prediction for some reason.
1638 float Parse::dynamic_branch_prediction(float &cnt, BoolTest::mask btest, Node* test) {
1639 ResourceMark rm;
1640
1641 cnt = COUNT_UNKNOWN;
1642
1643 int taken = 0;
1644 int not_taken = 0;
1645
1646 bool use_mdo = !has_injected_profile(btest, test, taken, not_taken);
1647
1648 if (use_mdo) {
1649 // Use MethodData information if it is available
1650 // FIXME: free the ProfileData structure
1651 ciMethodData* methodData = method()->method_data();
1652 if (!methodData->is_mature()) return PROB_UNKNOWN;
1653 ciProfileData* data = methodData->bci_to_data(bci());
1654 if (data == nullptr) {
1655 return PROB_UNKNOWN;
1656 }
1657 if (!data->is_JumpData()) return PROB_UNKNOWN;
1658
1659 // get taken and not taken values
1660 // NOTE: saturated UINT_MAX values become negative,
1661 // as do counts above INT_MAX.
1662 taken = data->as_JumpData()->taken();
1663 not_taken = 0;
1664 if (data->is_BranchData()) {
1665 not_taken = data->as_BranchData()->not_taken();
1666 }
1667
1668 // scale the counts to be commensurate with invocation counts:
1669 // NOTE: overflow for positive values is clamped at INT_MAX
1670 taken = method()->scale_count(taken);
1671 not_taken = method()->scale_count(not_taken);
1672 }
1673 // At this point, saturation or overflow is indicated by INT_MAX
1674 // or a negative value.
1675
1676 // Give up if too few (or too many, in which case the sum will overflow) counts to be meaningful.
1677 // We also check that individual counters are positive first, otherwise the sum can become positive.
1678 if (!counters_are_meaningful(taken, not_taken, 40)) {
1679 if (C->log() != nullptr) {
1680 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d'", iter().get_dest(), taken, not_taken);
1681 }
1682 return PROB_UNKNOWN;
1683 }
1684
1685 // Compute frequency that we arrive here
1686 float sum = taken + not_taken;
1687 // Adjust, if this block is a cloned private block but the
1688 // Jump counts are shared. Taken the private counts for
1689 // just this path instead of the shared counts.
1690 if( block()->count() > 0 )
1691 sum = block()->count();
1692 cnt = sum / FreqCountInvocations;
1693
1694 // Pin probability to sane limits
1695 float prob;
1696 if( !taken )
1697 prob = (0+PROB_MIN) / 2;
1698 else if( !not_taken )
1699 prob = (1+PROB_MAX) / 2;
1700 else { // Compute probability of true path
1701 prob = (float)taken / (float)(taken + not_taken);
1702 if (prob > PROB_MAX) prob = PROB_MAX;
1703 if (prob < PROB_MIN) prob = PROB_MIN;
1704 }
1705
1706 assert((cnt > 0.0f) && (prob > 0.0f),
1707 "Bad frequency assignment in if cnt=%g prob=%g taken=%d not_taken=%d", cnt, prob, taken, not_taken);
1708
1709 if (C->log() != nullptr) {
1710 const char* prob_str = nullptr;
1711 if (prob >= PROB_MAX) prob_str = (prob == PROB_MAX) ? "max" : "always";
1712 if (prob <= PROB_MIN) prob_str = (prob == PROB_MIN) ? "min" : "never";
1713 char prob_str_buf[30];
1714 if (prob_str == nullptr) {
1715 jio_snprintf(prob_str_buf, sizeof(prob_str_buf), "%20.2f", prob);
1716 prob_str = prob_str_buf;
1717 }
1718 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d' cnt='%f' prob='%s'",
1719 iter().get_dest(), taken, not_taken, cnt, prob_str);
1720 }
1721 return prob;
1722 }
1723
1724 //-----------------------------branch_prediction-------------------------------
1725 float Parse::branch_prediction(float& cnt,
1726 BoolTest::mask btest,
1727 int target_bci,
1728 Node* test) {
1729 float prob = dynamic_branch_prediction(cnt, btest, test);
1730 // If prob is unknown, switch to static prediction
1731 if (prob != PROB_UNKNOWN) return prob;
1732
1733 prob = PROB_FAIR; // Set default value
1734 if (btest == BoolTest::eq) // Exactly equal test?
1735 prob = PROB_STATIC_INFREQUENT; // Assume its relatively infrequent
1736 else if (btest == BoolTest::ne)
1737 prob = PROB_STATIC_FREQUENT; // Assume its relatively frequent
1738
1739 // If this is a conditional test guarding a backwards branch,
1740 // assume its a loop-back edge. Make it a likely taken branch.
1741 if (target_bci < bci()) {
1742 if (is_osr_parse()) { // Could be a hot OSR'd loop; force deopt
1743 // Since it's an OSR, we probably have profile data, but since
1744 // branch_prediction returned PROB_UNKNOWN, the counts are too small.
1745 // Let's make a special check here for completely zero counts.
1746 ciMethodData* methodData = method()->method_data();
1747 if (!methodData->is_empty()) {
1748 ciProfileData* data = methodData->bci_to_data(bci());
1749 // Only stop for truly zero counts, which mean an unknown part
1750 // of the OSR-ed method, and we want to deopt to gather more stats.
1751 // If you have ANY counts, then this loop is simply 'cold' relative
1752 // to the OSR loop.
1753 if (data == nullptr ||
1754 (data->as_BranchData()->taken() + data->as_BranchData()->not_taken() == 0)) {
1755 // This is the only way to return PROB_UNKNOWN:
1756 return PROB_UNKNOWN;
1757 }
1758 }
1759 }
1760 prob = PROB_STATIC_FREQUENT; // Likely to take backwards branch
1761 }
1762
1763 assert(prob != PROB_UNKNOWN, "must have some guess at this point");
1764 return prob;
1765 }
1766
1767 // The magic constants are chosen so as to match the output of
1768 // branch_prediction() when the profile reports a zero taken count.
1769 // It is important to distinguish zero counts unambiguously, because
1770 // some branches (e.g., _213_javac.Assembler.eliminate) validly produce
1771 // very small but nonzero probabilities, which if confused with zero
1772 // counts would keep the program recompiling indefinitely.
1773 bool Parse::seems_never_taken(float prob) const {
1774 return prob < PROB_MIN;
1775 }
1776
1777 //-------------------------------repush_if_args--------------------------------
1778 // Push arguments of an "if" bytecode back onto the stack by adjusting _sp.
1779 inline int Parse::repush_if_args() {
1780 if (PrintOpto && WizardMode) {
1781 tty->print("defending against excessive implicit null exceptions on %s @%d in ",
1782 Bytecodes::name(iter().cur_bc()), iter().cur_bci());
1783 method()->print_name(); tty->cr();
1784 }
1785 int bc_depth = - Bytecodes::depth(iter().cur_bc());
1786 assert(bc_depth == 1 || bc_depth == 2, "only two kinds of branches");
1787 DEBUG_ONLY(sync_jvms()); // argument(n) requires a synced jvms
1788 assert(argument(0) != nullptr, "must exist");
1789 assert(bc_depth == 1 || argument(1) != nullptr, "two must exist");
1790 inc_sp(bc_depth);
1791 return bc_depth;
1792 }
1793
1794 // Used by StressUnstableIfTraps
1795 static volatile int _trap_stress_counter = 0;
1796
1797 void Parse::increment_trap_stress_counter(Node*& counter, Node*& incr_store) {
1798 Node* counter_addr = makecon(TypeRawPtr::make((address)&_trap_stress_counter));
1799 counter = make_load(control(), counter_addr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered);
1800 counter = _gvn.transform(new AddINode(counter, intcon(1)));
1801 incr_store = store_to_memory(control(), counter_addr, counter, T_INT, Compile::AliasIdxRaw, MemNode::unordered);
1802 }
1803
1804 //----------------------------------do_ifnull----------------------------------
1805 void Parse::do_ifnull(BoolTest::mask btest, Node *c) {
1806 int target_bci = iter().get_dest();
1807
1808 Node* counter = nullptr;
1809 Node* incr_store = nullptr;
1810 bool do_stress_trap = StressUnstableIfTraps && ((C->random() % 2) == 0);
1811 if (do_stress_trap) {
1812 increment_trap_stress_counter(counter, incr_store);
1813 }
1814
1815 Block* branch_block = successor_for_bci(target_bci);
1816 Block* next_block = successor_for_bci(iter().next_bci());
1817
1818 float cnt;
1819 float prob = branch_prediction(cnt, btest, target_bci, c);
1820 if (prob == PROB_UNKNOWN) {
1821 // (An earlier version of do_ifnull omitted this trap for OSR methods.)
1822 if (PrintOpto && Verbose) {
1823 tty->print_cr("Never-taken edge stops compilation at bci %d", bci());
1824 }
1825 repush_if_args(); // to gather stats on loop
1826 uncommon_trap(Deoptimization::Reason_unreached,
1827 Deoptimization::Action_reinterpret,
1828 nullptr, "cold");
1829 if (C->eliminate_boxing()) {
1830 // Mark the successor blocks as parsed
1831 branch_block->next_path_num();
1832 next_block->next_path_num();
1833 }
1834 return;
1835 }
1836
1837 NOT_PRODUCT(explicit_null_checks_inserted++);
1838
1839 // Generate real control flow
1840 Node *tst = _gvn.transform( new BoolNode( c, btest ) );
1841
1842 // Sanity check the probability value
1843 assert(prob > 0.0f,"Bad probability in Parser");
1844 // Need xform to put node in hash table
1845 IfNode *iff = create_and_xform_if( control(), tst, prob, cnt );
1846 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser");
1847 // True branch
1848 { PreserveJVMState pjvms(this);
1849 Node* iftrue = _gvn.transform( new IfTrueNode (iff) );
1850 set_control(iftrue);
1851
1852 if (stopped()) { // Path is dead?
1853 NOT_PRODUCT(explicit_null_checks_elided++);
1854 if (C->eliminate_boxing()) {
1855 // Mark the successor block as parsed
1856 branch_block->next_path_num();
1857 }
1858 } else { // Path is live.
1859 adjust_map_after_if(btest, c, prob, branch_block);
1860 if (!stopped()) {
1861 merge(target_bci);
1862 }
1863 }
1864 }
1865
1866 // False branch
1867 Node* iffalse = _gvn.transform( new IfFalseNode(iff) );
1868 set_control(iffalse);
1869
1870 if (stopped()) { // Path is dead?
1871 NOT_PRODUCT(explicit_null_checks_elided++);
1872 if (C->eliminate_boxing()) {
1873 // Mark the successor block as parsed
1874 next_block->next_path_num();
1875 }
1876 } else { // Path is live.
1877 adjust_map_after_if(BoolTest(btest).negate(), c, 1.0-prob, next_block);
1878 }
1879
1880 if (do_stress_trap) {
1881 stress_trap(iff, counter, incr_store);
1882 }
1883 }
1884
1885 //------------------------------------do_if------------------------------------
1886 void Parse::do_if(BoolTest::mask btest, Node* c, bool can_trap, bool new_path, Node** ctrl_taken) {
1887 int target_bci = iter().get_dest();
1888
1889 Block* branch_block = successor_for_bci(target_bci);
1890 Block* next_block = successor_for_bci(iter().next_bci());
1891
1892 float cnt;
1893 float prob = branch_prediction(cnt, btest, target_bci, c);
1894 float untaken_prob = 1.0 - prob;
1895
1896 if (prob == PROB_UNKNOWN) {
1897 if (PrintOpto && Verbose) {
1898 tty->print_cr("Never-taken edge stops compilation at bci %d", bci());
1899 }
1900 repush_if_args(); // to gather stats on loop
1901 uncommon_trap(Deoptimization::Reason_unreached,
1902 Deoptimization::Action_reinterpret,
1903 nullptr, "cold");
1904 if (C->eliminate_boxing()) {
1905 // Mark the successor blocks as parsed
1906 branch_block->next_path_num();
1907 next_block->next_path_num();
1908 }
1909 return;
1910 }
1911
1912 Node* counter = nullptr;
1913 Node* incr_store = nullptr;
1914 bool do_stress_trap = StressUnstableIfTraps && ((C->random() % 2) == 0);
1915 if (do_stress_trap) {
1916 increment_trap_stress_counter(counter, incr_store);
1917 }
1918
1919 // Sanity check the probability value
1920 assert(0.0f < prob && prob < 1.0f,"Bad probability in Parser");
1921
1922 bool taken_if_true = true;
1923 // Convert BoolTest to canonical form:
1924 if (!BoolTest(btest).is_canonical()) {
1925 btest = BoolTest(btest).negate();
1926 taken_if_true = false;
1927 // prob is NOT updated here; it remains the probability of the taken
1928 // path (as opposed to the prob of the path guarded by an 'IfTrueNode').
1929 }
1930 assert(btest != BoolTest::eq, "!= is the only canonical exact test");
1931
1932 Node* tst0 = new BoolNode(c, btest);
1933 Node* tst = _gvn.transform(tst0);
1934 BoolTest::mask taken_btest = BoolTest::illegal;
1935 BoolTest::mask untaken_btest = BoolTest::illegal;
1936
1937 if (tst->is_Bool()) {
1938 // Refresh c from the transformed bool node, since it may be
1939 // simpler than the original c. Also re-canonicalize btest.
1940 // This wins when (Bool ne (Conv2B p) 0) => (Bool ne (CmpP p null)).
1941 // That can arise from statements like: if (x instanceof C) ...
1942 if (tst != tst0) {
1943 // Canonicalize one more time since transform can change it.
1944 btest = tst->as_Bool()->_test._test;
1945 if (!BoolTest(btest).is_canonical()) {
1946 // Reverse edges one more time...
1947 tst = _gvn.transform( tst->as_Bool()->negate(&_gvn) );
1948 btest = tst->as_Bool()->_test._test;
1949 assert(BoolTest(btest).is_canonical(), "sanity");
1950 taken_if_true = !taken_if_true;
1951 }
1952 c = tst->in(1);
1953 }
1954 BoolTest::mask neg_btest = BoolTest(btest).negate();
1955 taken_btest = taken_if_true ? btest : neg_btest;
1956 untaken_btest = taken_if_true ? neg_btest : btest;
1957 }
1958
1959 // Generate real control flow
1960 float true_prob = (taken_if_true ? prob : untaken_prob);
1961 IfNode* iff = create_and_map_if(control(), tst, true_prob, cnt);
1962 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser");
1963 Node* taken_branch = new IfTrueNode(iff);
1964 Node* untaken_branch = new IfFalseNode(iff);
1965 if (!taken_if_true) { // Finish conversion to canonical form
1966 Node* tmp = taken_branch;
1967 taken_branch = untaken_branch;
1968 untaken_branch = tmp;
1969 }
1970
1971 // Branch is taken:
1972 { PreserveJVMState pjvms(this);
1973 taken_branch = _gvn.transform(taken_branch);
1974 set_control(taken_branch);
1975
1976 if (stopped()) {
1977 if (C->eliminate_boxing() && !new_path) {
1978 // Mark the successor block as parsed (if we haven't created a new path)
1979 branch_block->next_path_num();
1980 }
1981 } else {
1982 adjust_map_after_if(taken_btest, c, prob, branch_block, can_trap);
1983 if (!stopped()) {
1984 if (new_path) {
1985 // Merge by using a new path
1986 merge_new_path(target_bci);
1987 } else if (ctrl_taken != nullptr) {
1988 // Don't merge but save taken branch to be wired by caller
1989 *ctrl_taken = control();
1990 } else {
1991 merge(target_bci);
1992 }
1993 }
1994 }
1995 }
1996
1997 untaken_branch = _gvn.transform(untaken_branch);
1998 set_control(untaken_branch);
1999
2000 // Branch not taken.
2001 if (stopped() && ctrl_taken == nullptr) {
2002 if (C->eliminate_boxing()) {
2003 // Mark the successor block as parsed (if caller does not re-wire control flow)
2004 next_block->next_path_num();
2005 }
2006 } else {
2007 adjust_map_after_if(untaken_btest, c, untaken_prob, next_block, can_trap);
2008 }
2009
2010 if (do_stress_trap) {
2011 stress_trap(iff, counter, incr_store);
2012 }
2013 }
2014
2015
2016 static ProfilePtrKind speculative_ptr_kind(const TypeOopPtr* t) {
2017 if (t->speculative() == nullptr) {
2018 return ProfileUnknownNull;
2019 }
2020 if (t->speculative_always_null()) {
2021 return ProfileAlwaysNull;
2022 }
2023 if (t->speculative_maybe_null()) {
2024 return ProfileMaybeNull;
2025 }
2026 return ProfileNeverNull;
2027 }
2028
2029 void Parse::acmp_always_null_input(Node* input, const TypeOopPtr* tinput, BoolTest::mask btest, Node* eq_region) {
2030 inc_sp(2);
2031 Node* cast = null_check_common(input, T_OBJECT, true, nullptr,
2032 !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_null_check) &&
2033 speculative_ptr_kind(tinput) == ProfileAlwaysNull);
2034 dec_sp(2);
2035 if (btest == BoolTest::ne) {
2036 {
2037 PreserveJVMState pjvms(this);
2038 replace_in_map(input, cast);
2039 int target_bci = iter().get_dest();
2040 merge(target_bci);
2041 }
2042 record_for_igvn(eq_region);
2043 set_control(_gvn.transform(eq_region));
2044 } else {
2045 replace_in_map(input, cast);
2046 }
2047 }
2048
2049 Node* Parse::acmp_null_check(Node* input, const TypeOopPtr* tinput, ProfilePtrKind input_ptr, Node*& null_ctl) {
2050 inc_sp(2);
2051 null_ctl = top();
2052 Node* cast = null_check_oop(input, &null_ctl,
2053 input_ptr == ProfileNeverNull || (input_ptr == ProfileUnknownNull && !too_many_traps_or_recompiles(Deoptimization::Reason_null_check)),
2054 false,
2055 speculative_ptr_kind(tinput) == ProfileNeverNull &&
2056 !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_null_check));
2057 dec_sp(2);
2058 assert(!stopped(), "null input should have been caught earlier");
2059 return cast;
2060 }
2061
2062 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) {
2063 Node* ne_region = new RegionNode(1);
2064 Node* null_ctl;
2065 Node* cast = acmp_null_check(input, tinput, input_ptr, null_ctl);
2066 ne_region->add_req(null_ctl);
2067
2068 Node* slow_ctl = type_check_receiver(cast, input_type, 1.0, &cast);
2069 {
2070 PreserveJVMState pjvms(this);
2071 inc_sp(2);
2072 set_control(slow_ctl);
2073 Deoptimization::DeoptReason reason;
2074 if (tinput->speculative_type() != nullptr && !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_class_check)) {
2075 reason = Deoptimization::Reason_speculate_class_check;
2076 } else {
2077 reason = Deoptimization::Reason_class_check;
2078 }
2079 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
2080 }
2081 ne_region->add_req(control());
2082
2083 record_for_igvn(ne_region);
2084 set_control(_gvn.transform(ne_region));
2085 if (btest == BoolTest::ne) {
2086 {
2087 PreserveJVMState pjvms(this);
2088 if (null_ctl == top()) {
2089 replace_in_map(input, cast);
2090 }
2091 int target_bci = iter().get_dest();
2092 merge(target_bci);
2093 }
2094 record_for_igvn(eq_region);
2095 set_control(_gvn.transform(eq_region));
2096 } else {
2097 if (null_ctl == top()) {
2098 replace_in_map(input, cast);
2099 }
2100 set_control(_gvn.transform(ne_region));
2101 }
2102 }
2103
2104 void Parse::acmp_unknown_non_inline_type_input(Node* input, const TypeOopPtr* tinput, ProfilePtrKind input_ptr, BoolTest::mask btest, Node* eq_region) {
2105 Node* ne_region = new RegionNode(1);
2106 Node* null_ctl;
2107 Node* cast = acmp_null_check(input, tinput, input_ptr, null_ctl);
2108 ne_region->add_req(null_ctl);
2109
2110 {
2111 BuildCutout unless(this, inline_type_test(cast, /* is_inline = */ false), PROB_MAX);
2112 inc_sp(2);
2113 uncommon_trap_exact(Deoptimization::Reason_class_check, Deoptimization::Action_maybe_recompile);
2114 }
2115
2116 ne_region->add_req(control());
2117
2118 record_for_igvn(ne_region);
2119 set_control(_gvn.transform(ne_region));
2120 if (btest == BoolTest::ne) {
2121 {
2122 PreserveJVMState pjvms(this);
2123 if (null_ctl == top()) {
2124 replace_in_map(input, cast);
2125 }
2126 int target_bci = iter().get_dest();
2127 merge(target_bci);
2128 }
2129 record_for_igvn(eq_region);
2130 set_control(_gvn.transform(eq_region));
2131 } else {
2132 if (null_ctl == top()) {
2133 replace_in_map(input, cast);
2134 }
2135 set_control(_gvn.transform(ne_region));
2136 }
2137 }
2138
2139 void Parse::do_acmp(BoolTest::mask btest, Node* left, Node* right) {
2140 ciKlass* left_type = nullptr;
2141 ciKlass* right_type = nullptr;
2142 ProfilePtrKind left_ptr = ProfileUnknownNull;
2143 ProfilePtrKind right_ptr = ProfileUnknownNull;
2144 bool left_inline_type = true;
2145 bool right_inline_type = true;
2146
2147 // Leverage profiling at acmp
2148 if (UseACmpProfile) {
2149 method()->acmp_profiled_type(bci(), left_type, right_type, left_ptr, right_ptr, left_inline_type, right_inline_type);
2150 if (too_many_traps_or_recompiles(Deoptimization::Reason_class_check)) {
2151 left_type = nullptr;
2152 right_type = nullptr;
2153 left_inline_type = true;
2154 right_inline_type = true;
2155 }
2156 if (too_many_traps_or_recompiles(Deoptimization::Reason_null_check)) {
2157 left_ptr = ProfileUnknownNull;
2158 right_ptr = ProfileUnknownNull;
2159 }
2160 }
2161
2162 if (UseTypeSpeculation) {
2163 record_profile_for_speculation(left, left_type, left_ptr);
2164 record_profile_for_speculation(right, right_type, right_ptr);
2165 }
2166
2167 if (!EnableValhalla) {
2168 Node* cmp = CmpP(left, right);
2169 cmp = optimize_cmp_with_klass(cmp);
2170 do_if(btest, cmp);
2171 return;
2172 }
2173
2174 // Check for equality before potentially allocating
2175 if (left == right) {
2176 do_if(btest, makecon(TypeInt::CC_EQ));
2177 return;
2178 }
2179
2180 // Allocate inline type operands and re-execute on deoptimization
2181 if (left->is_InlineType()) {
2182 if (_gvn.type(right)->is_zero_type() ||
2183 (right->is_InlineType() && _gvn.type(right->as_InlineType()->get_is_init())->is_zero_type())) {
2184 // Null checking a scalarized but nullable inline type. Check the IsInit
2185 // input instead of the oop input to avoid keeping buffer allocations alive.
2186 Node* cmp = CmpI(left->as_InlineType()->get_is_init(), intcon(0));
2187 do_if(btest, cmp);
2188 return;
2189 } else {
2190 PreserveReexecuteState preexecs(this);
2191 inc_sp(2);
2192 jvms()->set_should_reexecute(true);
2193 left = left->as_InlineType()->buffer(this)->get_oop();
2194 }
2195 }
2196 if (right->is_InlineType()) {
2197 PreserveReexecuteState preexecs(this);
2198 inc_sp(2);
2199 jvms()->set_should_reexecute(true);
2200 right = right->as_InlineType()->buffer(this)->get_oop();
2201 }
2202
2203 // First, do a normal pointer comparison
2204 const TypeOopPtr* tleft = _gvn.type(left)->isa_oopptr();
2205 const TypeOopPtr* tright = _gvn.type(right)->isa_oopptr();
2206 Node* cmp = CmpP(left, right);
2207 cmp = optimize_cmp_with_klass(cmp);
2208 if (tleft == nullptr || !tleft->can_be_inline_type() ||
2209 tright == nullptr || !tright->can_be_inline_type()) {
2210 // This is sufficient, if one of the operands can't be an inline type
2211 do_if(btest, cmp);
2212 return;
2213 }
2214
2215 // Don't add traps to unstable if branches because additional checks are required to
2216 // decide if the operands are equal/substitutable and we therefore shouldn't prune
2217 // branches for one if based on the profiling of the acmp branches.
2218 // Also, OptimizeUnstableIf would set an incorrect re-rexecution state because it
2219 // assumes that there is a 1-1 mapping between the if and the acmp branches and that
2220 // hitting a trap means that we will take the corresponding acmp branch on re-execution.
2221 const bool can_trap = true;
2222
2223 Node* eq_region = nullptr;
2224 if (btest == BoolTest::eq) {
2225 do_if(btest, cmp, !can_trap, true);
2226 if (stopped()) {
2227 // Pointers are equal, operands must be equal
2228 return;
2229 }
2230 } else {
2231 assert(btest == BoolTest::ne, "only eq or ne");
2232 Node* is_not_equal = nullptr;
2233 eq_region = new RegionNode(3);
2234 {
2235 PreserveJVMState pjvms(this);
2236 // Pointers are not equal, but more checks are needed to determine if the operands are (not) substitutable
2237 do_if(btest, cmp, !can_trap, false, &is_not_equal);
2238 if (!stopped()) {
2239 eq_region->init_req(1, control());
2240 }
2241 }
2242 if (is_not_equal == nullptr || is_not_equal->is_top()) {
2243 record_for_igvn(eq_region);
2244 set_control(_gvn.transform(eq_region));
2245 return;
2246 }
2247 set_control(is_not_equal);
2248 }
2249
2250 // Prefer speculative types if available
2251 if (!too_many_traps_or_recompiles(Deoptimization::Reason_speculate_class_check)) {
2252 if (tleft->speculative_type() != nullptr) {
2253 left_type = tleft->speculative_type();
2254 }
2255 if (tright->speculative_type() != nullptr) {
2256 right_type = tright->speculative_type();
2257 }
2258 }
2259
2260 if (speculative_ptr_kind(tleft) != ProfileMaybeNull && speculative_ptr_kind(tleft) != ProfileUnknownNull) {
2261 ProfilePtrKind speculative_left_ptr = speculative_ptr_kind(tleft);
2262 if (speculative_left_ptr == ProfileAlwaysNull && !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_null_assert)) {
2263 left_ptr = speculative_left_ptr;
2264 } else if (speculative_left_ptr == ProfileNeverNull && !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_null_check)) {
2265 left_ptr = speculative_left_ptr;
2266 }
2267 }
2268 if (speculative_ptr_kind(tright) != ProfileMaybeNull && speculative_ptr_kind(tright) != ProfileUnknownNull) {
2269 ProfilePtrKind speculative_right_ptr = speculative_ptr_kind(tright);
2270 if (speculative_right_ptr == ProfileAlwaysNull && !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_null_assert)) {
2271 right_ptr = speculative_right_ptr;
2272 } else if (speculative_right_ptr == ProfileNeverNull && !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_null_check)) {
2273 right_ptr = speculative_right_ptr;
2274 }
2275 }
2276
2277 if (left_ptr == ProfileAlwaysNull) {
2278 // Comparison with null. Assert the input is indeed null and we're done.
2279 acmp_always_null_input(left, tleft, btest, eq_region);
2280 return;
2281 }
2282 if (right_ptr == ProfileAlwaysNull) {
2283 // Comparison with null. Assert the input is indeed null and we're done.
2284 acmp_always_null_input(right, tright, btest, eq_region);
2285 return;
2286 }
2287 if (left_type != nullptr && !left_type->is_inlinetype()) {
2288 // Comparison with an object of known type
2289 acmp_known_non_inline_type_input(left, tleft, left_ptr, left_type, btest, eq_region);
2290 return;
2291 }
2292 if (right_type != nullptr && !right_type->is_inlinetype()) {
2293 // Comparison with an object of known type
2294 acmp_known_non_inline_type_input(right, tright, right_ptr, right_type, btest, eq_region);
2295 return;
2296 }
2297 if (!left_inline_type) {
2298 // Comparison with an object known not to be an inline type
2299 acmp_unknown_non_inline_type_input(left, tleft, left_ptr, btest, eq_region);
2300 return;
2301 }
2302 if (!right_inline_type) {
2303 // Comparison with an object known not to be an inline type
2304 acmp_unknown_non_inline_type_input(right, tright, right_ptr, btest, eq_region);
2305 return;
2306 }
2307
2308 // Pointers are not equal, check if first operand is non-null
2309 Node* ne_region = new RegionNode(6);
2310 Node* null_ctl;
2311 Node* not_null_right = acmp_null_check(right, tright, right_ptr, null_ctl);
2312 ne_region->init_req(1, null_ctl);
2313
2314 // First operand is non-null, check if it is an inline type
2315 Node* is_value = inline_type_test(not_null_right);
2316 IfNode* is_value_iff = create_and_map_if(control(), is_value, PROB_FAIR, COUNT_UNKNOWN);
2317 Node* not_value = _gvn.transform(new IfFalseNode(is_value_iff));
2318 ne_region->init_req(2, not_value);
2319 set_control(_gvn.transform(new IfTrueNode(is_value_iff)));
2320
2321 // The first operand is an inline type, check if the second operand is non-null
2322 Node* not_null_left = acmp_null_check(left, tleft, left_ptr, null_ctl);
2323 ne_region->init_req(3, null_ctl);
2324
2325 // Check if both operands are of the same class.
2326 Node* kls_left = load_object_klass(not_null_left);
2327 Node* kls_right = load_object_klass(not_null_right);
2328 Node* kls_cmp = CmpP(kls_left, kls_right);
2329 Node* kls_bol = _gvn.transform(new BoolNode(kls_cmp, BoolTest::ne));
2330 IfNode* kls_iff = create_and_map_if(control(), kls_bol, PROB_FAIR, COUNT_UNKNOWN);
2331 Node* kls_ne = _gvn.transform(new IfTrueNode(kls_iff));
2332 set_control(_gvn.transform(new IfFalseNode(kls_iff)));
2333 ne_region->init_req(4, kls_ne);
2334
2335 if (stopped()) {
2336 record_for_igvn(ne_region);
2337 set_control(_gvn.transform(ne_region));
2338 if (btest == BoolTest::ne) {
2339 {
2340 PreserveJVMState pjvms(this);
2341 int target_bci = iter().get_dest();
2342 merge(target_bci);
2343 }
2344 record_for_igvn(eq_region);
2345 set_control(_gvn.transform(eq_region));
2346 }
2347 return;
2348 }
2349
2350 // Both operands are values types of the same class, we need to perform a
2351 // substitutability test. Delegate to ValueObjectMethods::isSubstitutable().
2352 Node* ne_io_phi = PhiNode::make(ne_region, i_o());
2353 Node* mem = reset_memory();
2354 Node* ne_mem_phi = PhiNode::make(ne_region, mem);
2355
2356 Node* eq_io_phi = nullptr;
2357 Node* eq_mem_phi = nullptr;
2358 if (eq_region != nullptr) {
2359 eq_io_phi = PhiNode::make(eq_region, i_o());
2360 eq_mem_phi = PhiNode::make(eq_region, mem);
2361 }
2362
2363 set_all_memory(mem);
2364
2365 kill_dead_locals();
2366 ciMethod* subst_method = ciEnv::current()->ValueObjectMethods_klass()->find_method(ciSymbols::isSubstitutable_name(), ciSymbols::object_object_boolean_signature());
2367 CallStaticJavaNode *call = new CallStaticJavaNode(C, TypeFunc::make(subst_method), SharedRuntime::get_resolve_static_call_stub(), subst_method);
2368 call->set_override_symbolic_info(true);
2369 call->init_req(TypeFunc::Parms, not_null_left);
2370 call->init_req(TypeFunc::Parms+1, not_null_right);
2371 inc_sp(2);
2372 set_edges_for_java_call(call, false, false);
2373 Node* ret = set_results_for_java_call(call, false, true);
2374 dec_sp(2);
2375
2376 // Test the return value of ValueObjectMethods::isSubstitutable()
2377 // This is the last check, do_if can emit traps now.
2378 Node* subst_cmp = _gvn.transform(new CmpINode(ret, intcon(1)));
2379 Node* ctl = C->top();
2380 if (btest == BoolTest::eq) {
2381 PreserveJVMState pjvms(this);
2382 do_if(btest, subst_cmp, can_trap);
2383 if (!stopped()) {
2384 ctl = control();
2385 }
2386 } else {
2387 assert(btest == BoolTest::ne, "only eq or ne");
2388 PreserveJVMState pjvms(this);
2389 do_if(btest, subst_cmp, can_trap, false, &ctl);
2390 if (!stopped()) {
2391 eq_region->init_req(2, control());
2392 eq_io_phi->init_req(2, i_o());
2393 eq_mem_phi->init_req(2, reset_memory());
2394 }
2395 }
2396 ne_region->init_req(5, ctl);
2397 ne_io_phi->init_req(5, i_o());
2398 ne_mem_phi->init_req(5, reset_memory());
2399
2400 record_for_igvn(ne_region);
2401 set_control(_gvn.transform(ne_region));
2402 set_i_o(_gvn.transform(ne_io_phi));
2403 set_all_memory(_gvn.transform(ne_mem_phi));
2404
2405 if (btest == BoolTest::ne) {
2406 {
2407 PreserveJVMState pjvms(this);
2408 int target_bci = iter().get_dest();
2409 merge(target_bci);
2410 }
2411
2412 record_for_igvn(eq_region);
2413 set_control(_gvn.transform(eq_region));
2414 set_i_o(_gvn.transform(eq_io_phi));
2415 set_all_memory(_gvn.transform(eq_mem_phi));
2416 }
2417 }
2418
2419 // Force unstable if traps to be taken randomly to trigger intermittent bugs such as incorrect debug information.
2420 // Add another if before the unstable if that checks a "random" condition at runtime (a simple shared counter) and
2421 // then either takes the trap or executes the original, unstable if.
2422 void Parse::stress_trap(IfNode* orig_iff, Node* counter, Node* incr_store) {
2423 // Search for an unstable if trap
2424 CallStaticJavaNode* trap = nullptr;
2425 assert(orig_iff->Opcode() == Op_If && orig_iff->outcnt() == 2, "malformed if");
2426 ProjNode* trap_proj = orig_iff->uncommon_trap_proj(trap, Deoptimization::Reason_unstable_if);
2427 if (trap == nullptr || !trap->jvms()->should_reexecute()) {
2428 // No suitable trap found. Remove unused counter load and increment.
2429 C->gvn_replace_by(incr_store, incr_store->in(MemNode::Memory));
2430 return;
2431 }
2432
2433 // Remove trap from optimization list since we add another path to the trap.
2434 bool success = C->remove_unstable_if_trap(trap, true);
2435 assert(success, "Trap already modified");
2436
2437 // Add a check before the original if that will trap with a certain frequency and execute the original if otherwise
2438 int freq_log = (C->random() % 31) + 1; // Random logarithmic frequency in [1, 31]
2439 Node* mask = intcon(right_n_bits(freq_log));
2440 counter = _gvn.transform(new AndINode(counter, mask));
2441 Node* cmp = _gvn.transform(new CmpINode(counter, intcon(0)));
2442 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::mask::eq));
2443 IfNode* iff = _gvn.transform(new IfNode(orig_iff->in(0), bol, orig_iff->_prob, orig_iff->_fcnt))->as_If();
2444 Node* if_true = _gvn.transform(new IfTrueNode(iff));
2445 Node* if_false = _gvn.transform(new IfFalseNode(iff));
2446 assert(!if_true->is_top() && !if_false->is_top(), "trap always / never taken");
2447
2448 // Trap
2449 assert(trap_proj->outcnt() == 1, "some other nodes are dependent on the trap projection");
2450
2451 Node* trap_region = new RegionNode(3);
2452 trap_region->set_req(1, trap_proj);
2453 trap_region->set_req(2, if_true);
2454 trap->set_req(0, _gvn.transform(trap_region));
2455
2456 // Don't trap, execute original if
2457 orig_iff->set_req(0, if_false);
2458 }
2459
2460 bool Parse::path_is_suitable_for_uncommon_trap(float prob) const {
2461 // Randomly skip emitting an uncommon trap
2462 if (StressUnstableIfTraps && ((C->random() % 2) == 0)) {
2463 return false;
2464 }
2465 // Don't want to speculate on uncommon traps when running with -Xcomp
2466 if (!UseInterpreter) {
2467 return false;
2468 }
2469 return seems_never_taken(prob) &&
2470 !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if);
2471 }
2472
2473 void Parse::maybe_add_predicate_after_if(Block* path) {
2474 if (path->is_SEL_head() && path->preds_parsed() == 0) {
2475 // Add predicates at bci of if dominating the loop so traps can be
2476 // recorded on the if's profile data
2477 int bc_depth = repush_if_args();
2478 add_parse_predicates();
2479 dec_sp(bc_depth);
2480 path->set_has_predicates();
2481 }
2482 }
2483
2484
2485 //----------------------------adjust_map_after_if------------------------------
2486 // Adjust the JVM state to reflect the result of taking this path.
2487 // Basically, it means inspecting the CmpNode controlling this
2488 // branch, seeing how it constrains a tested value, and then
2489 // deciding if it's worth our while to encode this constraint
2490 // as graph nodes in the current abstract interpretation map.
2491 void Parse::adjust_map_after_if(BoolTest::mask btest, Node* c, float prob, Block* path, bool can_trap) {
2492 if (!c->is_Cmp()) {
2493 maybe_add_predicate_after_if(path);
2494 return;
2495 }
2496
2497 if (stopped() || btest == BoolTest::illegal) {
2498 return; // nothing to do
2499 }
2500
2501 bool is_fallthrough = (path == successor_for_bci(iter().next_bci()));
2502
2503 if (can_trap && path_is_suitable_for_uncommon_trap(prob)) {
2504 repush_if_args();
2505 Node* call = uncommon_trap(Deoptimization::Reason_unstable_if,
2506 Deoptimization::Action_reinterpret,
2507 nullptr,
2508 (is_fallthrough ? "taken always" : "taken never"));
2509
2510 if (call != nullptr) {
2511 C->record_unstable_if_trap(new UnstableIfTrap(call->as_CallStaticJava(), path));
2512 }
2513 return;
2514 }
2515
2516 Node* val = c->in(1);
2517 Node* con = c->in(2);
2518 const Type* tcon = _gvn.type(con);
2519 const Type* tval = _gvn.type(val);
2520 bool have_con = tcon->singleton();
2521 if (tval->singleton()) {
2522 if (!have_con) {
2523 // Swap, so constant is in con.
2524 con = val;
2525 tcon = tval;
2526 val = c->in(2);
2527 tval = _gvn.type(val);
2528 btest = BoolTest(btest).commute();
2529 have_con = true;
2530 } else {
2531 // Do we have two constants? Then leave well enough alone.
2532 have_con = false;
2533 }
2534 }
2535 if (!have_con) { // remaining adjustments need a con
2536 maybe_add_predicate_after_if(path);
2537 return;
2538 }
2539
2540 sharpen_type_after_if(btest, con, tcon, val, tval);
2541 maybe_add_predicate_after_if(path);
2542 }
2543
2544
2545 static Node* extract_obj_from_klass_load(PhaseGVN* gvn, Node* n) {
2546 Node* ldk;
2547 if (n->is_DecodeNKlass()) {
2548 if (n->in(1)->Opcode() != Op_LoadNKlass) {
2549 return nullptr;
2550 } else {
2551 ldk = n->in(1);
2552 }
2553 } else if (n->Opcode() != Op_LoadKlass) {
2554 return nullptr;
2555 } else {
2556 ldk = n;
2557 }
2558 assert(ldk != nullptr && ldk->is_Load(), "should have found a LoadKlass or LoadNKlass node");
2559
2560 Node* adr = ldk->in(MemNode::Address);
2561 intptr_t off = 0;
2562 Node* obj = AddPNode::Ideal_base_and_offset(adr, gvn, off);
2563 if (obj == nullptr || off != oopDesc::klass_offset_in_bytes()) // loading oopDesc::_klass?
2564 return nullptr;
2565 const TypePtr* tp = gvn->type(obj)->is_ptr();
2566 if (tp == nullptr || !(tp->isa_instptr() || tp->isa_aryptr())) // is obj a Java object ptr?
2567 return nullptr;
2568
2569 return obj;
2570 }
2571
2572 void Parse::sharpen_type_after_if(BoolTest::mask btest,
2573 Node* con, const Type* tcon,
2574 Node* val, const Type* tval) {
2575 // Look for opportunities to sharpen the type of a node
2576 // whose klass is compared with a constant klass.
2577 if (btest == BoolTest::eq && tcon->isa_klassptr()) {
2578 Node* obj = extract_obj_from_klass_load(&_gvn, val);
2579 const TypeOopPtr* con_type = tcon->isa_klassptr()->as_instance_type();
2580 if (obj != nullptr && (con_type->isa_instptr() || con_type->isa_aryptr())) {
2581 // Found:
2582 // Bool(CmpP(LoadKlass(obj._klass), ConP(Foo.klass)), [eq])
2583 // or the narrowOop equivalent.
2584 const Type* obj_type = _gvn.type(obj);
2585 const TypeOopPtr* tboth = obj_type->join_speculative(con_type)->isa_oopptr();
2586 if (tboth != nullptr && tboth->klass_is_exact() && tboth != obj_type &&
2587 tboth->higher_equal(obj_type)) {
2588 // obj has to be of the exact type Foo if the CmpP succeeds.
2589 int obj_in_map = map()->find_edge(obj);
2590 JVMState* jvms = this->jvms();
2591 if (obj_in_map >= 0 &&
2592 (jvms->is_loc(obj_in_map) || jvms->is_stk(obj_in_map))) {
2593 TypeNode* ccast = new CheckCastPPNode(control(), obj, tboth);
2594 const Type* tcc = ccast->as_Type()->type();
2595 assert(tcc != obj_type && tcc->higher_equal(obj_type), "must improve");
2596 // Delay transform() call to allow recovery of pre-cast value
2597 // at the control merge.
2598 _gvn.set_type_bottom(ccast);
2599 record_for_igvn(ccast);
2600 if (tboth->is_inlinetypeptr()) {
2601 ccast = InlineTypeNode::make_from_oop(this, ccast, tboth->exact_klass(true)->as_inline_klass());
2602 }
2603 // Here's the payoff.
2604 replace_in_map(obj, ccast);
2605 }
2606 }
2607 }
2608 }
2609
2610 int val_in_map = map()->find_edge(val);
2611 if (val_in_map < 0) return; // replace_in_map would be useless
2612 {
2613 JVMState* jvms = this->jvms();
2614 if (!(jvms->is_loc(val_in_map) ||
2615 jvms->is_stk(val_in_map)))
2616 return; // again, it would be useless
2617 }
2618
2619 // Check for a comparison to a constant, and "know" that the compared
2620 // value is constrained on this path.
2621 assert(tcon->singleton(), "");
2622 ConstraintCastNode* ccast = nullptr;
2623 Node* cast = nullptr;
2624
2625 switch (btest) {
2626 case BoolTest::eq: // Constant test?
2627 {
2628 const Type* tboth = tcon->join_speculative(tval);
2629 if (tboth == tval) break; // Nothing to gain.
2630 if (tcon->isa_int()) {
2631 ccast = new CastIINode(control(), val, tboth);
2632 } else if (tcon == TypePtr::NULL_PTR) {
2633 // Cast to null, but keep the pointer identity temporarily live.
2634 ccast = new CastPPNode(control(), val, tboth);
2635 } else {
2636 const TypeF* tf = tcon->isa_float_constant();
2637 const TypeD* td = tcon->isa_double_constant();
2638 // Exclude tests vs float/double 0 as these could be
2639 // either +0 or -0. Just because you are equal to +0
2640 // doesn't mean you ARE +0!
2641 // Note, following code also replaces Long and Oop values.
2642 if ((!tf || tf->_f != 0.0) &&
2643 (!td || td->_d != 0.0))
2644 cast = con; // Replace non-constant val by con.
2645 }
2646 }
2647 break;
2648
2649 case BoolTest::ne:
2650 if (tcon == TypePtr::NULL_PTR) {
2651 cast = cast_not_null(val, false);
2652 }
2653 break;
2654
2655 default:
2656 // (At this point we could record int range types with CastII.)
2657 break;
2658 }
2659
2660 if (ccast != nullptr) {
2661 const Type* tcc = ccast->as_Type()->type();
2662 assert(tcc != tval && tcc->higher_equal(tval), "must improve");
2663 // Delay transform() call to allow recovery of pre-cast value
2664 // at the control merge.
2665 _gvn.set_type_bottom(ccast);
2666 record_for_igvn(ccast);
2667 cast = ccast;
2668 }
2669
2670 if (cast != nullptr) { // Here's the payoff.
2671 replace_in_map(val, cast);
2672 }
2673 }
2674
2675 /**
2676 * Use speculative type to optimize CmpP node: if comparison is
2677 * against the low level class, cast the object to the speculative
2678 * type if any. CmpP should then go away.
2679 *
2680 * @param c expected CmpP node
2681 * @return result of CmpP on object casted to speculative type
2682 *
2683 */
2684 Node* Parse::optimize_cmp_with_klass(Node* c) {
2685 // If this is transformed by the _gvn to a comparison with the low
2686 // level klass then we may be able to use speculation
2687 if (c->Opcode() == Op_CmpP &&
2688 (c->in(1)->Opcode() == Op_LoadKlass || c->in(1)->Opcode() == Op_DecodeNKlass) &&
2689 c->in(2)->is_Con()) {
2690 Node* load_klass = nullptr;
2691 Node* decode = nullptr;
2692 if (c->in(1)->Opcode() == Op_DecodeNKlass) {
2693 decode = c->in(1);
2694 load_klass = c->in(1)->in(1);
2695 } else {
2696 load_klass = c->in(1);
2697 }
2698 if (load_klass->in(2)->is_AddP()) {
2699 Node* addp = load_klass->in(2);
2700 Node* obj = addp->in(AddPNode::Address);
2701 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
2702 if (obj_type->speculative_type_not_null() != nullptr) {
2703 ciKlass* k = obj_type->speculative_type();
2704 inc_sp(2);
2705 obj = maybe_cast_profiled_obj(obj, k);
2706 dec_sp(2);
2707 if (obj->is_InlineType()) {
2708 assert(obj->as_InlineType()->is_allocated(&_gvn), "must be allocated");
2709 obj = obj->as_InlineType()->get_oop();
2710 }
2711 // Make the CmpP use the casted obj
2712 addp = basic_plus_adr(obj, addp->in(AddPNode::Offset));
2713 load_klass = load_klass->clone();
2714 load_klass->set_req(2, addp);
2715 load_klass = _gvn.transform(load_klass);
2716 if (decode != nullptr) {
2717 decode = decode->clone();
2718 decode->set_req(1, load_klass);
2719 load_klass = _gvn.transform(decode);
2720 }
2721 c = c->clone();
2722 c->set_req(1, load_klass);
2723 c = _gvn.transform(c);
2724 }
2725 }
2726 }
2727 return c;
2728 }
2729
2730 //------------------------------do_one_bytecode--------------------------------
2731 // Parse this bytecode, and alter the Parsers JVM->Node mapping
2732 void Parse::do_one_bytecode() {
2733 Node *a, *b, *c, *d; // Handy temps
2734 BoolTest::mask btest;
2735 int i;
2736
2737 assert(!has_exceptions(), "bytecode entry state must be clear of throws");
2738
2739 if (C->check_node_count(NodeLimitFudgeFactor * 5,
2740 "out of nodes parsing method")) {
2741 return;
2742 }
2743
2744 #ifdef ASSERT
2745 // for setting breakpoints
2746 if (TraceOptoParse) {
2747 tty->print(" @");
2748 dump_bci(bci());
2749 tty->print(" %s", Bytecodes::name(bc()));
2750 tty->cr();
2751 }
2752 #endif
2753
2754 switch (bc()) {
2755 case Bytecodes::_nop:
2756 // do nothing
2757 break;
2758 case Bytecodes::_lconst_0:
2759 push_pair(longcon(0));
2760 break;
2761
2762 case Bytecodes::_lconst_1:
2763 push_pair(longcon(1));
2764 break;
2765
2766 case Bytecodes::_fconst_0:
2767 push(zerocon(T_FLOAT));
2768 break;
2769
2770 case Bytecodes::_fconst_1:
2771 push(makecon(TypeF::ONE));
2772 break;
2773
2774 case Bytecodes::_fconst_2:
2775 push(makecon(TypeF::make(2.0f)));
2776 break;
2777
2778 case Bytecodes::_dconst_0:
2779 push_pair(zerocon(T_DOUBLE));
2780 break;
2781
2782 case Bytecodes::_dconst_1:
2783 push_pair(makecon(TypeD::ONE));
2784 break;
2785
2786 case Bytecodes::_iconst_m1:push(intcon(-1)); break;
2787 case Bytecodes::_iconst_0: push(intcon( 0)); break;
2788 case Bytecodes::_iconst_1: push(intcon( 1)); break;
2789 case Bytecodes::_iconst_2: push(intcon( 2)); break;
2790 case Bytecodes::_iconst_3: push(intcon( 3)); break;
2791 case Bytecodes::_iconst_4: push(intcon( 4)); break;
2792 case Bytecodes::_iconst_5: push(intcon( 5)); break;
2793 case Bytecodes::_bipush: push(intcon(iter().get_constant_u1())); break;
2794 case Bytecodes::_sipush: push(intcon(iter().get_constant_u2())); break;
2795 case Bytecodes::_aconst_null: push(null()); break;
2796
2797 case Bytecodes::_ldc:
2798 case Bytecodes::_ldc_w:
2799 case Bytecodes::_ldc2_w: {
2800 // ciTypeFlow should trap if the ldc is in error state or if the constant is not loaded
2801 assert(!iter().is_in_error(), "ldc is in error state");
2802 ciConstant constant = iter().get_constant();
2803 assert(constant.is_loaded(), "constant is not loaded");
2804 const Type* con_type = Type::make_from_constant(constant);
2805 if (con_type != nullptr) {
2806 push_node(con_type->basic_type(), makecon(con_type));
2807 }
2808 break;
2809 }
2810
2811 case Bytecodes::_aload_0:
2812 push( local(0) );
2813 break;
2814 case Bytecodes::_aload_1:
2815 push( local(1) );
2816 break;
2817 case Bytecodes::_aload_2:
2818 push( local(2) );
2819 break;
2820 case Bytecodes::_aload_3:
2821 push( local(3) );
2822 break;
2823 case Bytecodes::_aload:
2824 push( local(iter().get_index()) );
2825 break;
2826
2827 case Bytecodes::_fload_0:
2828 case Bytecodes::_iload_0:
2829 push( local(0) );
2830 break;
2831 case Bytecodes::_fload_1:
2832 case Bytecodes::_iload_1:
2833 push( local(1) );
2834 break;
2835 case Bytecodes::_fload_2:
2836 case Bytecodes::_iload_2:
2837 push( local(2) );
2838 break;
2839 case Bytecodes::_fload_3:
2840 case Bytecodes::_iload_3:
2841 push( local(3) );
2842 break;
2843 case Bytecodes::_fload:
2844 case Bytecodes::_iload:
2845 push( local(iter().get_index()) );
2846 break;
2847 case Bytecodes::_lload_0:
2848 push_pair_local( 0 );
2849 break;
2850 case Bytecodes::_lload_1:
2851 push_pair_local( 1 );
2852 break;
2853 case Bytecodes::_lload_2:
2854 push_pair_local( 2 );
2855 break;
2856 case Bytecodes::_lload_3:
2857 push_pair_local( 3 );
2858 break;
2859 case Bytecodes::_lload:
2860 push_pair_local( iter().get_index() );
2861 break;
2862
2863 case Bytecodes::_dload_0:
2864 push_pair_local(0);
2865 break;
2866 case Bytecodes::_dload_1:
2867 push_pair_local(1);
2868 break;
2869 case Bytecodes::_dload_2:
2870 push_pair_local(2);
2871 break;
2872 case Bytecodes::_dload_3:
2873 push_pair_local(3);
2874 break;
2875 case Bytecodes::_dload:
2876 push_pair_local(iter().get_index());
2877 break;
2878 case Bytecodes::_fstore_0:
2879 case Bytecodes::_istore_0:
2880 case Bytecodes::_astore_0:
2881 set_local( 0, pop() );
2882 break;
2883 case Bytecodes::_fstore_1:
2884 case Bytecodes::_istore_1:
2885 case Bytecodes::_astore_1:
2886 set_local( 1, pop() );
2887 break;
2888 case Bytecodes::_fstore_2:
2889 case Bytecodes::_istore_2:
2890 case Bytecodes::_astore_2:
2891 set_local( 2, pop() );
2892 break;
2893 case Bytecodes::_fstore_3:
2894 case Bytecodes::_istore_3:
2895 case Bytecodes::_astore_3:
2896 set_local( 3, pop() );
2897 break;
2898 case Bytecodes::_fstore:
2899 case Bytecodes::_istore:
2900 case Bytecodes::_astore:
2901 set_local( iter().get_index(), pop() );
2902 break;
2903 // long stores
2904 case Bytecodes::_lstore_0:
2905 set_pair_local( 0, pop_pair() );
2906 break;
2907 case Bytecodes::_lstore_1:
2908 set_pair_local( 1, pop_pair() );
2909 break;
2910 case Bytecodes::_lstore_2:
2911 set_pair_local( 2, pop_pair() );
2912 break;
2913 case Bytecodes::_lstore_3:
2914 set_pair_local( 3, pop_pair() );
2915 break;
2916 case Bytecodes::_lstore:
2917 set_pair_local( iter().get_index(), pop_pair() );
2918 break;
2919
2920 // double stores
2921 case Bytecodes::_dstore_0:
2922 set_pair_local( 0, dprecision_rounding(pop_pair()) );
2923 break;
2924 case Bytecodes::_dstore_1:
2925 set_pair_local( 1, dprecision_rounding(pop_pair()) );
2926 break;
2927 case Bytecodes::_dstore_2:
2928 set_pair_local( 2, dprecision_rounding(pop_pair()) );
2929 break;
2930 case Bytecodes::_dstore_3:
2931 set_pair_local( 3, dprecision_rounding(pop_pair()) );
2932 break;
2933 case Bytecodes::_dstore:
2934 set_pair_local( iter().get_index(), dprecision_rounding(pop_pair()) );
2935 break;
2936
2937 case Bytecodes::_pop: dec_sp(1); break;
2938 case Bytecodes::_pop2: dec_sp(2); break;
2939 case Bytecodes::_swap:
2940 a = pop();
2941 b = pop();
2942 push(a);
2943 push(b);
2944 break;
2945 case Bytecodes::_dup:
2946 a = pop();
2947 push(a);
2948 push(a);
2949 break;
2950 case Bytecodes::_dup_x1:
2951 a = pop();
2952 b = pop();
2953 push( a );
2954 push( b );
2955 push( a );
2956 break;
2957 case Bytecodes::_dup_x2:
2958 a = pop();
2959 b = pop();
2960 c = pop();
2961 push( a );
2962 push( c );
2963 push( b );
2964 push( a );
2965 break;
2966 case Bytecodes::_dup2:
2967 a = pop();
2968 b = pop();
2969 push( b );
2970 push( a );
2971 push( b );
2972 push( a );
2973 break;
2974
2975 case Bytecodes::_dup2_x1:
2976 // before: .. c, b, a
2977 // after: .. b, a, c, b, a
2978 // not tested
2979 a = pop();
2980 b = pop();
2981 c = pop();
2982 push( b );
2983 push( a );
2984 push( c );
2985 push( b );
2986 push( a );
2987 break;
2988 case Bytecodes::_dup2_x2:
2989 // before: .. d, c, b, a
2990 // after: .. b, a, d, c, b, a
2991 // not tested
2992 a = pop();
2993 b = pop();
2994 c = pop();
2995 d = pop();
2996 push( b );
2997 push( a );
2998 push( d );
2999 push( c );
3000 push( b );
3001 push( a );
3002 break;
3003
3004 case Bytecodes::_arraylength: {
3005 // Must do null-check with value on expression stack
3006 Node *ary = null_check(peek(), T_ARRAY);
3007 // Compile-time detect of null-exception?
3008 if (stopped()) return;
3009 a = pop();
3010 push(load_array_length(a));
3011 break;
3012 }
3013
3014 case Bytecodes::_baload: array_load(T_BYTE); break;
3015 case Bytecodes::_caload: array_load(T_CHAR); break;
3016 case Bytecodes::_iaload: array_load(T_INT); break;
3017 case Bytecodes::_saload: array_load(T_SHORT); break;
3018 case Bytecodes::_faload: array_load(T_FLOAT); break;
3019 case Bytecodes::_aaload: array_load(T_OBJECT); break;
3020 case Bytecodes::_laload: array_load(T_LONG); break;
3021 case Bytecodes::_daload: array_load(T_DOUBLE); break;
3022 case Bytecodes::_bastore: array_store(T_BYTE); break;
3023 case Bytecodes::_castore: array_store(T_CHAR); break;
3024 case Bytecodes::_iastore: array_store(T_INT); break;
3025 case Bytecodes::_sastore: array_store(T_SHORT); break;
3026 case Bytecodes::_fastore: array_store(T_FLOAT); break;
3027 case Bytecodes::_aastore: array_store(T_OBJECT); break;
3028 case Bytecodes::_lastore: array_store(T_LONG); break;
3029 case Bytecodes::_dastore: array_store(T_DOUBLE); break;
3030
3031 case Bytecodes::_getfield:
3032 do_getfield();
3033 break;
3034
3035 case Bytecodes::_getstatic:
3036 do_getstatic();
3037 break;
3038
3039 case Bytecodes::_putfield:
3040 do_putfield();
3041 break;
3042
3043 case Bytecodes::_putstatic:
3044 do_putstatic();
3045 break;
3046
3047 case Bytecodes::_irem:
3048 // Must keep both values on the expression-stack during null-check
3049 zero_check_int(peek());
3050 // Compile-time detect of null-exception?
3051 if (stopped()) return;
3052 b = pop();
3053 a = pop();
3054 push(_gvn.transform(new ModINode(control(), a, b)));
3055 break;
3056 case Bytecodes::_idiv:
3057 // Must keep both values on the expression-stack during null-check
3058 zero_check_int(peek());
3059 // Compile-time detect of null-exception?
3060 if (stopped()) return;
3061 b = pop();
3062 a = pop();
3063 push( _gvn.transform( new DivINode(control(),a,b) ) );
3064 break;
3065 case Bytecodes::_imul:
3066 b = pop(); a = pop();
3067 push( _gvn.transform( new MulINode(a,b) ) );
3068 break;
3069 case Bytecodes::_iadd:
3070 b = pop(); a = pop();
3071 push( _gvn.transform( new AddINode(a,b) ) );
3072 break;
3073 case Bytecodes::_ineg:
3074 a = pop();
3075 push( _gvn.transform( new SubINode(_gvn.intcon(0),a)) );
3076 break;
3077 case Bytecodes::_isub:
3078 b = pop(); a = pop();
3079 push( _gvn.transform( new SubINode(a,b) ) );
3080 break;
3081 case Bytecodes::_iand:
3082 b = pop(); a = pop();
3083 push( _gvn.transform( new AndINode(a,b) ) );
3084 break;
3085 case Bytecodes::_ior:
3086 b = pop(); a = pop();
3087 push( _gvn.transform( new OrINode(a,b) ) );
3088 break;
3089 case Bytecodes::_ixor:
3090 b = pop(); a = pop();
3091 push( _gvn.transform( new XorINode(a,b) ) );
3092 break;
3093 case Bytecodes::_ishl:
3094 b = pop(); a = pop();
3095 push( _gvn.transform( new LShiftINode(a,b) ) );
3096 break;
3097 case Bytecodes::_ishr:
3098 b = pop(); a = pop();
3099 push( _gvn.transform( new RShiftINode(a,b) ) );
3100 break;
3101 case Bytecodes::_iushr:
3102 b = pop(); a = pop();
3103 push( _gvn.transform( new URShiftINode(a,b) ) );
3104 break;
3105
3106 case Bytecodes::_fneg:
3107 a = pop();
3108 b = _gvn.transform(new NegFNode (a));
3109 push(b);
3110 break;
3111
3112 case Bytecodes::_fsub:
3113 b = pop();
3114 a = pop();
3115 c = _gvn.transform( new SubFNode(a,b) );
3116 d = precision_rounding(c);
3117 push( d );
3118 break;
3119
3120 case Bytecodes::_fadd:
3121 b = pop();
3122 a = pop();
3123 c = _gvn.transform( new AddFNode(a,b) );
3124 d = precision_rounding(c);
3125 push( d );
3126 break;
3127
3128 case Bytecodes::_fmul:
3129 b = pop();
3130 a = pop();
3131 c = _gvn.transform( new MulFNode(a,b) );
3132 d = precision_rounding(c);
3133 push( d );
3134 break;
3135
3136 case Bytecodes::_fdiv:
3137 b = pop();
3138 a = pop();
3139 c = _gvn.transform( new DivFNode(nullptr,a,b) );
3140 d = precision_rounding(c);
3141 push( d );
3142 break;
3143
3144 case Bytecodes::_frem:
3145 if (Matcher::has_match_rule(Op_ModF)) {
3146 // Generate a ModF node.
3147 b = pop();
3148 a = pop();
3149 c = _gvn.transform( new ModFNode(nullptr,a,b) );
3150 d = precision_rounding(c);
3151 push( d );
3152 }
3153 else {
3154 // Generate a call.
3155 modf();
3156 }
3157 break;
3158
3159 case Bytecodes::_fcmpl:
3160 b = pop();
3161 a = pop();
3162 c = _gvn.transform( new CmpF3Node( a, b));
3163 push(c);
3164 break;
3165 case Bytecodes::_fcmpg:
3166 b = pop();
3167 a = pop();
3168
3169 // Same as fcmpl but need to flip the unordered case. Swap the inputs,
3170 // which negates the result sign except for unordered. Flip the unordered
3171 // as well by using CmpF3 which implements unordered-lesser instead of
3172 // unordered-greater semantics. Finally, commute the result bits. Result
3173 // is same as using a CmpF3Greater except we did it with CmpF3 alone.
3174 c = _gvn.transform( new CmpF3Node( b, a));
3175 c = _gvn.transform( new SubINode(_gvn.intcon(0),c) );
3176 push(c);
3177 break;
3178
3179 case Bytecodes::_f2i:
3180 a = pop();
3181 push(_gvn.transform(new ConvF2INode(a)));
3182 break;
3183
3184 case Bytecodes::_d2i:
3185 a = pop_pair();
3186 b = _gvn.transform(new ConvD2INode(a));
3187 push( b );
3188 break;
3189
3190 case Bytecodes::_f2d:
3191 a = pop();
3192 b = _gvn.transform( new ConvF2DNode(a));
3193 push_pair( b );
3194 break;
3195
3196 case Bytecodes::_d2f:
3197 a = pop_pair();
3198 b = _gvn.transform( new ConvD2FNode(a));
3199 // This breaks _227_mtrt (speed & correctness) and _222_mpegaudio (speed)
3200 //b = _gvn.transform(new RoundFloatNode(nullptr, b) );
3201 push( b );
3202 break;
3203
3204 case Bytecodes::_l2f:
3205 if (Matcher::convL2FSupported()) {
3206 a = pop_pair();
3207 b = _gvn.transform( new ConvL2FNode(a));
3208 // For x86_32.ad, FILD doesn't restrict precision to 24 or 53 bits.
3209 // Rather than storing the result into an FP register then pushing
3210 // out to memory to round, the machine instruction that implements
3211 // ConvL2D is responsible for rounding.
3212 // c = precision_rounding(b);
3213 push(b);
3214 } else {
3215 l2f();
3216 }
3217 break;
3218
3219 case Bytecodes::_l2d:
3220 a = pop_pair();
3221 b = _gvn.transform( new ConvL2DNode(a));
3222 // For x86_32.ad, rounding is always necessary (see _l2f above).
3223 // c = dprecision_rounding(b);
3224 push_pair(b);
3225 break;
3226
3227 case Bytecodes::_f2l:
3228 a = pop();
3229 b = _gvn.transform( new ConvF2LNode(a));
3230 push_pair(b);
3231 break;
3232
3233 case Bytecodes::_d2l:
3234 a = pop_pair();
3235 b = _gvn.transform( new ConvD2LNode(a));
3236 push_pair(b);
3237 break;
3238
3239 case Bytecodes::_dsub:
3240 b = pop_pair();
3241 a = pop_pair();
3242 c = _gvn.transform( new SubDNode(a,b) );
3243 d = dprecision_rounding(c);
3244 push_pair( d );
3245 break;
3246
3247 case Bytecodes::_dadd:
3248 b = pop_pair();
3249 a = pop_pair();
3250 c = _gvn.transform( new AddDNode(a,b) );
3251 d = dprecision_rounding(c);
3252 push_pair( d );
3253 break;
3254
3255 case Bytecodes::_dmul:
3256 b = pop_pair();
3257 a = pop_pair();
3258 c = _gvn.transform( new MulDNode(a,b) );
3259 d = dprecision_rounding(c);
3260 push_pair( d );
3261 break;
3262
3263 case Bytecodes::_ddiv:
3264 b = pop_pair();
3265 a = pop_pair();
3266 c = _gvn.transform( new DivDNode(nullptr,a,b) );
3267 d = dprecision_rounding(c);
3268 push_pair( d );
3269 break;
3270
3271 case Bytecodes::_dneg:
3272 a = pop_pair();
3273 b = _gvn.transform(new NegDNode (a));
3274 push_pair(b);
3275 break;
3276
3277 case Bytecodes::_drem:
3278 if (Matcher::has_match_rule(Op_ModD)) {
3279 // Generate a ModD node.
3280 b = pop_pair();
3281 a = pop_pair();
3282 // a % b
3283
3284 c = _gvn.transform( new ModDNode(nullptr,a,b) );
3285 d = dprecision_rounding(c);
3286 push_pair( d );
3287 }
3288 else {
3289 // Generate a call.
3290 modd();
3291 }
3292 break;
3293
3294 case Bytecodes::_dcmpl:
3295 b = pop_pair();
3296 a = pop_pair();
3297 c = _gvn.transform( new CmpD3Node( a, b));
3298 push(c);
3299 break;
3300
3301 case Bytecodes::_dcmpg:
3302 b = pop_pair();
3303 a = pop_pair();
3304 // Same as dcmpl but need to flip the unordered case.
3305 // Commute the inputs, which negates the result sign except for unordered.
3306 // Flip the unordered as well by using CmpD3 which implements
3307 // unordered-lesser instead of unordered-greater semantics.
3308 // Finally, negate the result bits. Result is same as using a
3309 // CmpD3Greater except we did it with CmpD3 alone.
3310 c = _gvn.transform( new CmpD3Node( b, a));
3311 c = _gvn.transform( new SubINode(_gvn.intcon(0),c) );
3312 push(c);
3313 break;
3314
3315
3316 // Note for longs -> lo word is on TOS, hi word is on TOS - 1
3317 case Bytecodes::_land:
3318 b = pop_pair();
3319 a = pop_pair();
3320 c = _gvn.transform( new AndLNode(a,b) );
3321 push_pair(c);
3322 break;
3323 case Bytecodes::_lor:
3324 b = pop_pair();
3325 a = pop_pair();
3326 c = _gvn.transform( new OrLNode(a,b) );
3327 push_pair(c);
3328 break;
3329 case Bytecodes::_lxor:
3330 b = pop_pair();
3331 a = pop_pair();
3332 c = _gvn.transform( new XorLNode(a,b) );
3333 push_pair(c);
3334 break;
3335
3336 case Bytecodes::_lshl:
3337 b = pop(); // the shift count
3338 a = pop_pair(); // value to be shifted
3339 c = _gvn.transform( new LShiftLNode(a,b) );
3340 push_pair(c);
3341 break;
3342 case Bytecodes::_lshr:
3343 b = pop(); // the shift count
3344 a = pop_pair(); // value to be shifted
3345 c = _gvn.transform( new RShiftLNode(a,b) );
3346 push_pair(c);
3347 break;
3348 case Bytecodes::_lushr:
3349 b = pop(); // the shift count
3350 a = pop_pair(); // value to be shifted
3351 c = _gvn.transform( new URShiftLNode(a,b) );
3352 push_pair(c);
3353 break;
3354 case Bytecodes::_lmul:
3355 b = pop_pair();
3356 a = pop_pair();
3357 c = _gvn.transform( new MulLNode(a,b) );
3358 push_pair(c);
3359 break;
3360
3361 case Bytecodes::_lrem:
3362 // Must keep both values on the expression-stack during null-check
3363 assert(peek(0) == top(), "long word order");
3364 zero_check_long(peek(1));
3365 // Compile-time detect of null-exception?
3366 if (stopped()) return;
3367 b = pop_pair();
3368 a = pop_pair();
3369 c = _gvn.transform( new ModLNode(control(),a,b) );
3370 push_pair(c);
3371 break;
3372
3373 case Bytecodes::_ldiv:
3374 // Must keep both values on the expression-stack during null-check
3375 assert(peek(0) == top(), "long word order");
3376 zero_check_long(peek(1));
3377 // Compile-time detect of null-exception?
3378 if (stopped()) return;
3379 b = pop_pair();
3380 a = pop_pair();
3381 c = _gvn.transform( new DivLNode(control(),a,b) );
3382 push_pair(c);
3383 break;
3384
3385 case Bytecodes::_ladd:
3386 b = pop_pair();
3387 a = pop_pair();
3388 c = _gvn.transform( new AddLNode(a,b) );
3389 push_pair(c);
3390 break;
3391 case Bytecodes::_lsub:
3392 b = pop_pair();
3393 a = pop_pair();
3394 c = _gvn.transform( new SubLNode(a,b) );
3395 push_pair(c);
3396 break;
3397 case Bytecodes::_lcmp:
3398 // Safepoints are now inserted _before_ branches. The long-compare
3399 // bytecode painfully produces a 3-way value (-1,0,+1) which requires a
3400 // slew of control flow. These are usually followed by a CmpI vs zero and
3401 // a branch; this pattern then optimizes to the obvious long-compare and
3402 // branch. However, if the branch is backwards there's a Safepoint
3403 // inserted. The inserted Safepoint captures the JVM state at the
3404 // pre-branch point, i.e. it captures the 3-way value. Thus if a
3405 // long-compare is used to control a loop the debug info will force
3406 // computation of the 3-way value, even though the generated code uses a
3407 // long-compare and branch. We try to rectify the situation by inserting
3408 // a SafePoint here and have it dominate and kill the safepoint added at a
3409 // following backwards branch. At this point the JVM state merely holds 2
3410 // longs but not the 3-way value.
3411 switch (iter().next_bc()) {
3412 case Bytecodes::_ifgt:
3413 case Bytecodes::_iflt:
3414 case Bytecodes::_ifge:
3415 case Bytecodes::_ifle:
3416 case Bytecodes::_ifne:
3417 case Bytecodes::_ifeq:
3418 // If this is a backwards branch in the bytecodes, add Safepoint
3419 maybe_add_safepoint(iter().next_get_dest());
3420 default:
3421 break;
3422 }
3423 b = pop_pair();
3424 a = pop_pair();
3425 c = _gvn.transform( new CmpL3Node( a, b ));
3426 push(c);
3427 break;
3428
3429 case Bytecodes::_lneg:
3430 a = pop_pair();
3431 b = _gvn.transform( new SubLNode(longcon(0),a));
3432 push_pair(b);
3433 break;
3434 case Bytecodes::_l2i:
3435 a = pop_pair();
3436 push( _gvn.transform( new ConvL2INode(a)));
3437 break;
3438 case Bytecodes::_i2l:
3439 a = pop();
3440 b = _gvn.transform( new ConvI2LNode(a));
3441 push_pair(b);
3442 break;
3443 case Bytecodes::_i2b:
3444 // Sign extend
3445 a = pop();
3446 a = Compile::narrow_value(T_BYTE, a, nullptr, &_gvn, true);
3447 push(a);
3448 break;
3449 case Bytecodes::_i2s:
3450 a = pop();
3451 a = Compile::narrow_value(T_SHORT, a, nullptr, &_gvn, true);
3452 push(a);
3453 break;
3454 case Bytecodes::_i2c:
3455 a = pop();
3456 a = Compile::narrow_value(T_CHAR, a, nullptr, &_gvn, true);
3457 push(a);
3458 break;
3459
3460 case Bytecodes::_i2f:
3461 a = pop();
3462 b = _gvn.transform( new ConvI2FNode(a) ) ;
3463 c = precision_rounding(b);
3464 push (b);
3465 break;
3466
3467 case Bytecodes::_i2d:
3468 a = pop();
3469 b = _gvn.transform( new ConvI2DNode(a));
3470 push_pair(b);
3471 break;
3472
3473 case Bytecodes::_iinc: // Increment local
3474 i = iter().get_index(); // Get local index
3475 set_local( i, _gvn.transform( new AddINode( _gvn.intcon(iter().get_iinc_con()), local(i) ) ) );
3476 break;
3477
3478 // Exit points of synchronized methods must have an unlock node
3479 case Bytecodes::_return:
3480 return_current(nullptr);
3481 break;
3482
3483 case Bytecodes::_ireturn:
3484 case Bytecodes::_areturn:
3485 case Bytecodes::_freturn:
3486 return_current(pop());
3487 break;
3488 case Bytecodes::_lreturn:
3489 return_current(pop_pair());
3490 break;
3491 case Bytecodes::_dreturn:
3492 return_current(pop_pair());
3493 break;
3494
3495 case Bytecodes::_athrow:
3496 // null exception oop throws null pointer exception
3497 null_check(peek());
3498 if (stopped()) return;
3499 // Hook the thrown exception directly to subsequent handlers.
3500 if (BailoutToInterpreterForThrows) {
3501 // Keep method interpreted from now on.
3502 uncommon_trap(Deoptimization::Reason_unhandled,
3503 Deoptimization::Action_make_not_compilable);
3504 return;
3505 }
3506 if (env()->jvmti_can_post_on_exceptions()) {
3507 // check if we must post exception events, take uncommon trap if so (with must_throw = false)
3508 uncommon_trap_if_should_post_on_exceptions(Deoptimization::Reason_unhandled, false);
3509 }
3510 // Here if either can_post_on_exceptions or should_post_on_exceptions is false
3511 add_exception_state(make_exception_state(peek()));
3512 break;
3513
3514 case Bytecodes::_goto: // fall through
3515 case Bytecodes::_goto_w: {
3516 int target_bci = (bc() == Bytecodes::_goto) ? iter().get_dest() : iter().get_far_dest();
3517
3518 // If this is a backwards branch in the bytecodes, add Safepoint
3519 maybe_add_safepoint(target_bci);
3520
3521 // Merge the current control into the target basic block
3522 merge(target_bci);
3523
3524 // See if we can get some profile data and hand it off to the next block
3525 Block *target_block = block()->successor_for_bci(target_bci);
3526 if (target_block->pred_count() != 1) break;
3527 ciMethodData* methodData = method()->method_data();
3528 if (!methodData->is_mature()) break;
3529 ciProfileData* data = methodData->bci_to_data(bci());
3530 assert(data != nullptr && data->is_JumpData(), "need JumpData for taken branch");
3531 int taken = ((ciJumpData*)data)->taken();
3532 taken = method()->scale_count(taken);
3533 target_block->set_count(taken);
3534 break;
3535 }
3536
3537 case Bytecodes::_ifnull: btest = BoolTest::eq; goto handle_if_null;
3538 case Bytecodes::_ifnonnull: btest = BoolTest::ne; goto handle_if_null;
3539 handle_if_null:
3540 // If this is a backwards branch in the bytecodes, add Safepoint
3541 maybe_add_safepoint(iter().get_dest());
3542 a = null();
3543 b = pop();
3544 if (b->is_InlineType()) {
3545 // Null checking a scalarized but nullable inline type. Check the IsInit
3546 // input instead of the oop input to avoid keeping buffer allocations alive
3547 c = _gvn.transform(new CmpINode(b->as_InlineType()->get_is_init(), zerocon(T_INT)));
3548 } else {
3549 if (!_gvn.type(b)->speculative_maybe_null() &&
3550 !too_many_traps(Deoptimization::Reason_speculate_null_check)) {
3551 inc_sp(1);
3552 Node* null_ctl = top();
3553 b = null_check_oop(b, &null_ctl, true, true, true);
3554 assert(null_ctl->is_top(), "no null control here");
3555 dec_sp(1);
3556 } else if (_gvn.type(b)->speculative_always_null() &&
3557 !too_many_traps(Deoptimization::Reason_speculate_null_assert)) {
3558 inc_sp(1);
3559 b = null_assert(b);
3560 dec_sp(1);
3561 }
3562 c = _gvn.transform( new CmpPNode(b, a) );
3563 }
3564 do_ifnull(btest, c);
3565 break;
3566
3567 case Bytecodes::_if_acmpeq: btest = BoolTest::eq; goto handle_if_acmp;
3568 case Bytecodes::_if_acmpne: btest = BoolTest::ne; goto handle_if_acmp;
3569 handle_if_acmp:
3570 // If this is a backwards branch in the bytecodes, add Safepoint
3571 maybe_add_safepoint(iter().get_dest());
3572 a = pop();
3573 b = pop();
3574 do_acmp(btest, b, a);
3575 break;
3576
3577 case Bytecodes::_ifeq: btest = BoolTest::eq; goto handle_ifxx;
3578 case Bytecodes::_ifne: btest = BoolTest::ne; goto handle_ifxx;
3579 case Bytecodes::_iflt: btest = BoolTest::lt; goto handle_ifxx;
3580 case Bytecodes::_ifle: btest = BoolTest::le; goto handle_ifxx;
3581 case Bytecodes::_ifgt: btest = BoolTest::gt; goto handle_ifxx;
3582 case Bytecodes::_ifge: btest = BoolTest::ge; goto handle_ifxx;
3583 handle_ifxx:
3584 // If this is a backwards branch in the bytecodes, add Safepoint
3585 maybe_add_safepoint(iter().get_dest());
3586 a = _gvn.intcon(0);
3587 b = pop();
3588 c = _gvn.transform( new CmpINode(b, a) );
3589 do_if(btest, c);
3590 break;
3591
3592 case Bytecodes::_if_icmpeq: btest = BoolTest::eq; goto handle_if_icmp;
3593 case Bytecodes::_if_icmpne: btest = BoolTest::ne; goto handle_if_icmp;
3594 case Bytecodes::_if_icmplt: btest = BoolTest::lt; goto handle_if_icmp;
3595 case Bytecodes::_if_icmple: btest = BoolTest::le; goto handle_if_icmp;
3596 case Bytecodes::_if_icmpgt: btest = BoolTest::gt; goto handle_if_icmp;
3597 case Bytecodes::_if_icmpge: btest = BoolTest::ge; goto handle_if_icmp;
3598 handle_if_icmp:
3599 // If this is a backwards branch in the bytecodes, add Safepoint
3600 maybe_add_safepoint(iter().get_dest());
3601 a = pop();
3602 b = pop();
3603 c = _gvn.transform( new CmpINode( b, a ) );
3604 do_if(btest, c);
3605 break;
3606
3607 case Bytecodes::_tableswitch:
3608 do_tableswitch();
3609 break;
3610
3611 case Bytecodes::_lookupswitch:
3612 do_lookupswitch();
3613 break;
3614
3615 case Bytecodes::_invokestatic:
3616 case Bytecodes::_invokedynamic:
3617 case Bytecodes::_invokespecial:
3618 case Bytecodes::_invokevirtual:
3619 case Bytecodes::_invokeinterface:
3620 do_call();
3621 break;
3622 case Bytecodes::_checkcast:
3623 do_checkcast();
3624 break;
3625 case Bytecodes::_instanceof:
3626 do_instanceof();
3627 break;
3628 case Bytecodes::_anewarray:
3629 do_newarray();
3630 break;
3631 case Bytecodes::_newarray:
3632 do_newarray((BasicType)iter().get_index());
3633 break;
3634 case Bytecodes::_multianewarray:
3635 do_multianewarray();
3636 break;
3637 case Bytecodes::_new:
3638 do_new();
3639 break;
3640
3641 case Bytecodes::_jsr:
3642 case Bytecodes::_jsr_w:
3643 do_jsr();
3644 break;
3645
3646 case Bytecodes::_ret:
3647 do_ret();
3648 break;
3649
3650
3651 case Bytecodes::_monitorenter:
3652 do_monitor_enter();
3653 break;
3654
3655 case Bytecodes::_monitorexit:
3656 do_monitor_exit();
3657 break;
3658
3659 case Bytecodes::_breakpoint:
3660 // Breakpoint set concurrently to compile
3661 // %%% use an uncommon trap?
3662 C->record_failure("breakpoint in method");
3663 return;
3664
3665 default:
3666 #ifndef PRODUCT
3667 map()->dump(99);
3668 #endif
3669 tty->print("\nUnhandled bytecode %s\n", Bytecodes::name(bc()) );
3670 ShouldNotReachHere();
3671 }
3672
3673 #ifndef PRODUCT
3674 if (failing()) { return; }
3675 constexpr int perBytecode = 6;
3676 if (C->should_print_igv(perBytecode)) {
3677 IdealGraphPrinter* printer = C->igv_printer();
3678 char buffer[256];
3679 jio_snprintf(buffer, sizeof(buffer), "Bytecode %d: %s", bci(), Bytecodes::name(bc()));
3680 bool old = printer->traverse_outs();
3681 printer->set_traverse_outs(true);
3682 printer->print_method(buffer, perBytecode);
3683 printer->set_traverse_outs(old);
3684 }
3685 #endif
3686 }