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