1 /*
2 * Copyright (c) 1998, 2025, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "ci/ciMethodData.hpp"
26 #include "ci/ciSymbols.hpp"
27 #include "classfile/vmSymbols.hpp"
28 #include "compiler/compileLog.hpp"
29 #include "interpreter/linkResolver.hpp"
30 #include "jvm_io.h"
31 #include "memory/resourceArea.hpp"
32 #include "memory/universe.hpp"
33 #include "oops/oop.inline.hpp"
34 #include "opto/addnode.hpp"
35 #include "opto/castnode.hpp"
36 #include "opto/convertnode.hpp"
37 #include "opto/divnode.hpp"
38 #include "opto/idealGraphPrinter.hpp"
39 #include "opto/idealKit.hpp"
40 #include "opto/inlinetypenode.hpp"
41 #include "opto/matcher.hpp"
42 #include "opto/memnode.hpp"
43 #include "opto/mulnode.hpp"
44 #include "opto/opaquenode.hpp"
45 #include "opto/parse.hpp"
46 #include "opto/runtime.hpp"
47 #include "runtime/deoptimization.hpp"
48 #include "runtime/sharedRuntime.hpp"
49
50 #ifndef PRODUCT
51 extern uint explicit_null_checks_inserted,
52 explicit_null_checks_elided;
53 #endif
54
55 Node* Parse::record_profile_for_speculation_at_array_load(Node* ld) {
56 // Feed unused profile data to type speculation
57 if (UseTypeSpeculation && UseArrayLoadStoreProfile) {
58 ciKlass* array_type = nullptr;
59 ciKlass* element_type = nullptr;
60 ProfilePtrKind element_ptr = ProfileMaybeNull;
61 bool flat_array = true;
62 bool null_free_array = true;
63 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
64 if (element_type != nullptr || element_ptr != ProfileMaybeNull) {
65 ld = record_profile_for_speculation(ld, element_type, element_ptr);
66 }
67 }
68 return ld;
69 }
70
71
72 //---------------------------------array_load----------------------------------
73 void Parse::array_load(BasicType bt) {
74 const Type* elemtype = Type::TOP;
75 Node* adr = array_addressing(bt, 0, elemtype);
76 if (stopped()) return; // guaranteed null or range check
77
78 Node* array_index = pop();
79 Node* array = pop();
80
81 // Handle inline type arrays
82 const TypeOopPtr* element_ptr = elemtype->make_oopptr();
83 const TypeAryPtr* array_type = _gvn.type(array)->is_aryptr();
84
85 if (!array_type->is_not_flat()) {
86 // Cannot statically determine if array is a flat array, emit runtime check
87 assert(UseArrayFlattening && is_reference_type(bt) && element_ptr->can_be_inline_type() &&
88 (!element_ptr->is_inlinetypeptr() || element_ptr->inline_klass()->flat_in_array()), "array can't be flat");
89 IdealKit ideal(this);
90 IdealVariable res(ideal);
91 ideal.declarations_done();
92 ideal.if_then(flat_array_test(array, /* flat = */ false)); {
93 // Non-flat array
94 sync_kit(ideal);
95 if (!array_type->is_flat()) {
96 assert(array_type->is_flat() || control()->in(0)->as_If()->is_flat_array_check(&_gvn), "Should be found");
97 const TypeAryPtr* adr_type = TypeAryPtr::get_array_body_type(bt);
98 DecoratorSet decorator_set = IN_HEAP | IS_ARRAY | C2_CONTROL_DEPENDENT_LOAD;
99 if (needs_range_check(array_type->size(), array_index)) {
100 // We've emitted a RangeCheck but now insert an additional check between the range check and the actual load.
101 // We cannot pin the load to two separate nodes. Instead, we pin it conservatively here such that it cannot
102 // possibly float above the range check at any point.
103 decorator_set |= C2_UNKNOWN_CONTROL_LOAD;
104 }
105 Node* ld = access_load_at(array, adr, adr_type, element_ptr, bt, decorator_set);
106 if (element_ptr->is_inlinetypeptr()) {
107 ld = InlineTypeNode::make_from_oop(this, ld, element_ptr->inline_klass());
108 }
109 ideal.set(res, ld);
110 }
111 ideal.sync_kit(this);
112 } ideal.else_(); {
113 // Flat array
114 sync_kit(ideal);
115 if (!array_type->is_not_flat()) {
116 if (element_ptr->is_inlinetypeptr()) {
117 // Element type is known, cast and load from flat array layout.
118 ciInlineKlass* vk = element_ptr->inline_klass();
119 bool is_null_free = array_type->is_null_free() || !vk->has_nullable_atomic_layout();
120 bool is_not_null_free = array_type->is_not_null_free() || (!vk->has_atomic_layout() && !vk->has_non_atomic_layout());
121 if (is_null_free) {
122 // TODO 8350865 Impossible type
123 is_not_null_free = false;
124 }
125 bool is_naturally_atomic = is_null_free && vk->nof_declared_nonstatic_fields() <= 1;
126 bool may_need_atomicity = !is_naturally_atomic && ((!is_not_null_free && vk->has_atomic_layout()) || (!is_null_free && vk->has_nullable_atomic_layout()));
127
128 adr = flat_array_element_address(array, array_index, vk, is_null_free, is_not_null_free, may_need_atomicity);
129 int nm_offset = is_null_free ? -1 : vk->null_marker_offset_in_payload();
130 Node* vt = InlineTypeNode::make_from_flat(this, vk, array, adr, array_index, nullptr, 0, may_need_atomicity, nm_offset);
131 ideal.set(res, vt);
132 } else {
133 // Element type is unknown, and thus we cannot statically determine the exact flat array layout. Emit a
134 // runtime call to correctly load the inline type element from the flat array.
135 Node* inline_type = load_from_unknown_flat_array(array, array_index, element_ptr);
136 ideal.set(res, inline_type);
137 }
138 }
139 ideal.sync_kit(this);
140 } ideal.end_if();
141 sync_kit(ideal);
142 Node* ld = _gvn.transform(ideal.value(res));
143 ld = record_profile_for_speculation_at_array_load(ld);
144 push_node(bt, ld);
145 return;
146 }
147
148 if (elemtype == TypeInt::BOOL) {
149 bt = T_BOOLEAN;
150 }
151 const TypeAryPtr* adr_type = TypeAryPtr::get_array_body_type(bt);
152 Node* ld = access_load_at(array, adr, adr_type, elemtype, bt,
153 IN_HEAP | IS_ARRAY | C2_CONTROL_DEPENDENT_LOAD);
154 ld = record_profile_for_speculation_at_array_load(ld);
155 // Loading an inline type from a non-flat array
156 if (element_ptr != nullptr && element_ptr->is_inlinetypeptr()) {
157 assert(!array_type->is_null_free() || !element_ptr->maybe_null(), "inline type array elements should never be null");
158 ld = InlineTypeNode::make_from_oop(this, ld, element_ptr->inline_klass());
159 }
160 push_node(bt, ld);
161 }
162
163 Node* Parse::load_from_unknown_flat_array(Node* array, Node* array_index, const TypeOopPtr* element_ptr) {
164 // Below membars keep this access to an unknown flat array correctly
165 // ordered with other unknown and known flat array accesses.
166 insert_mem_bar_volatile(Op_MemBarCPUOrder, C->get_alias_index(TypeAryPtr::INLINES));
167
168 Node* call = nullptr;
169 {
170 // Re-execute flat array load if runtime call triggers deoptimization
171 PreserveReexecuteState preexecs(this);
172 jvms()->set_bci(_bci);
173 jvms()->set_should_reexecute(true);
174 inc_sp(2);
175 kill_dead_locals();
176 call = make_runtime_call(RC_NO_LEAF | RC_NO_IO,
177 OptoRuntime::load_unknown_inline_Type(),
178 OptoRuntime::load_unknown_inline_Java(),
179 nullptr, TypeRawPtr::BOTTOM,
180 array, array_index);
181 }
182 make_slow_call_ex(call, env()->Throwable_klass(), false);
183 Node* buffer = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
184
185 insert_mem_bar_volatile(Op_MemBarCPUOrder, C->get_alias_index(TypeAryPtr::INLINES));
186
187 // Keep track of the information that the inline type is in flat arrays
188 const Type* unknown_value = element_ptr->is_instptr()->cast_to_flat_in_array();
189 return _gvn.transform(new CheckCastPPNode(control(), buffer, unknown_value));
190 }
191
192 //--------------------------------array_store----------------------------------
193 void Parse::array_store(BasicType bt) {
194 const Type* elemtype = Type::TOP;
195 Node* adr = array_addressing(bt, type2size[bt], elemtype);
196 if (stopped()) return; // guaranteed null or range check
197 Node* stored_value_casted = nullptr;
198 if (bt == T_OBJECT) {
199 stored_value_casted = array_store_check(adr, elemtype);
200 if (stopped()) {
201 return;
202 }
203 }
204 Node* const stored_value = pop_node(bt); // Value to store
205 Node* const array_index = pop(); // Index in the array
206 Node* array = pop(); // The array itself
207
208 const TypeAryPtr* array_type = _gvn.type(array)->is_aryptr();
209 const TypeAryPtr* adr_type = TypeAryPtr::get_array_body_type(bt);
210
211 if (elemtype == TypeInt::BOOL) {
212 bt = T_BOOLEAN;
213 } else if (bt == T_OBJECT) {
214 elemtype = elemtype->make_oopptr();
215 const Type* stored_value_casted_type = _gvn.type(stored_value_casted);
216 // Based on the value to be stored, try to determine if the array is not null-free and/or not flat.
217 // This is only legal for non-null stores because the array_store_check always passes for null, even
218 // if the array is null-free. Null stores are handled in GraphKit::inline_array_null_guard().
219 bool not_inline = !stored_value_casted_type->maybe_null() && !stored_value_casted_type->is_oopptr()->can_be_inline_type();
220 bool not_null_free = not_inline;
221 bool not_flat = not_inline || ( stored_value_casted_type->is_inlinetypeptr() &&
222 !stored_value_casted_type->inline_klass()->flat_in_array());
223 if (!array_type->is_not_null_free() && not_null_free) {
224 // Storing a non-inline type, mark array as not null-free.
225 array_type = array_type->cast_to_not_null_free();
226 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, array_type));
227 replace_in_map(array, cast);
228 array = cast;
229 }
230 if (!array_type->is_not_flat() && not_flat) {
231 // Storing to a non-flat array, mark array as not flat.
232 array_type = array_type->cast_to_not_flat();
233 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, array_type));
234 replace_in_map(array, cast);
235 array = cast;
236 }
237
238 if (!array_type->is_flat() && array_type->is_null_free()) {
239 // Store to non-flat null-free inline type array (elements can never be null)
240 assert(!stored_value_casted_type->maybe_null(), "should be guaranteed by array store check");
241 if (elemtype->is_inlinetypeptr() && elemtype->inline_klass()->is_empty()) {
242 // Ignore empty inline stores, array is already initialized.
243 return;
244 }
245 } else if (!array_type->is_not_flat()) {
246 // Array might be a flat array, emit runtime checks (for nullptr, a simple inline_array_null_guard is sufficient).
247 assert(UseArrayFlattening && !not_flat && elemtype->is_oopptr()->can_be_inline_type() &&
248 (!array_type->klass_is_exact() || array_type->is_flat()), "array can't be a flat array");
249 // TODO 8350865 Depending on the available layouts, we can avoid this check in below flat/not-flat branches. Also the safe_for_replace arg is now always true.
250 array = inline_array_null_guard(array, stored_value_casted, 3, true);
251 IdealKit ideal(this);
252 ideal.if_then(flat_array_test(array, /* flat = */ false)); {
253 // Non-flat array
254 if (!array_type->is_flat()) {
255 sync_kit(ideal);
256 assert(array_type->is_flat() || ideal.ctrl()->in(0)->as_If()->is_flat_array_check(&_gvn), "Should be found");
257 inc_sp(3);
258 access_store_at(array, adr, adr_type, stored_value_casted, elemtype, bt, MO_UNORDERED | IN_HEAP | IS_ARRAY, false);
259 dec_sp(3);
260 ideal.sync_kit(this);
261 }
262 } ideal.else_(); {
263 // Flat array
264 sync_kit(ideal);
265 if (!array_type->is_not_flat()) {
266 // Try to determine the inline klass type of the stored value
267 ciInlineKlass* vk = nullptr;
268 if (stored_value_casted_type->is_inlinetypeptr()) {
269 vk = stored_value_casted_type->inline_klass();
270 } else if (elemtype->is_inlinetypeptr()) {
271 vk = elemtype->inline_klass();
272 }
273
274 if (vk != nullptr) {
275 // Element type is known, cast and store to flat array layout.
276 bool is_null_free = array_type->is_null_free() || !vk->has_nullable_atomic_layout();
277 bool is_not_null_free = array_type->is_not_null_free() || (!vk->has_atomic_layout() && !vk->has_non_atomic_layout());
278 if (is_null_free) {
279 // TODO 8350865 Impossible type
280 is_not_null_free = false;
281 }
282 bool is_naturally_atomic = is_null_free && vk->nof_declared_nonstatic_fields() <= 1;
283 bool may_need_atomicity = !is_naturally_atomic && ((!is_not_null_free && vk->has_atomic_layout()) || (!is_null_free && vk->has_nullable_atomic_layout()));
284
285 // Re-execute flat array store if buffering triggers deoptimization
286 PreserveReexecuteState preexecs(this);
287 jvms()->set_should_reexecute(true);
288 inc_sp(3);
289
290 if (!stored_value_casted->is_InlineType()) {
291 assert(_gvn.type(stored_value_casted) == TypePtr::NULL_PTR, "Unexpected value");
292 stored_value_casted = InlineTypeNode::make_null(_gvn, vk);
293 }
294 adr = flat_array_element_address(array, array_index, vk, is_null_free, is_not_null_free, may_need_atomicity);
295 int nm_offset = is_null_free ? -1 : vk->null_marker_offset_in_payload();
296 stored_value_casted->as_InlineType()->store_flat(this, array, adr, array_index, nullptr, 0, may_need_atomicity, nm_offset, MO_UNORDERED | IN_HEAP | IS_ARRAY);
297 } else {
298 // Element type is unknown, emit a runtime call since the flat array layout is not statically known.
299 store_to_unknown_flat_array(array, array_index, stored_value_casted);
300 }
301 }
302 ideal.sync_kit(this);
303 }
304 ideal.end_if();
305 sync_kit(ideal);
306 return;
307 } else if (!array_type->is_not_null_free()) {
308 // Array is not flat but may be null free
309 assert(elemtype->is_oopptr()->can_be_inline_type(), "array can't be null-free");
310 array = inline_array_null_guard(array, stored_value_casted, 3, true);
311 }
312 }
313 inc_sp(3);
314 access_store_at(array, adr, adr_type, stored_value, elemtype, bt, MO_UNORDERED | IN_HEAP | IS_ARRAY);
315 dec_sp(3);
316 }
317
318 // Emit a runtime call to store to a flat array whose element type is either unknown (i.e. we do not know the flat
319 // array layout) or not exact (could have different flat array layouts at runtime).
320 void Parse::store_to_unknown_flat_array(Node* array, Node* const idx, Node* non_null_stored_value) {
321 // Below membars keep this access to an unknown flat array correctly
322 // ordered with other unknown and known flat array accesses.
323 insert_mem_bar_volatile(Op_MemBarCPUOrder, C->get_alias_index(TypeAryPtr::INLINES));
324
325 Node* call = nullptr;
326 {
327 // Re-execute flat array store if runtime call triggers deoptimization
328 PreserveReexecuteState preexecs(this);
329 jvms()->set_bci(_bci);
330 jvms()->set_should_reexecute(true);
331 inc_sp(3);
332 kill_dead_locals();
333 call = make_runtime_call(RC_NO_LEAF | RC_NO_IO,
334 OptoRuntime::store_unknown_inline_Type(),
335 OptoRuntime::store_unknown_inline_Java(),
336 nullptr, TypeRawPtr::BOTTOM,
337 non_null_stored_value, array, idx);
338 }
339 make_slow_call_ex(call, env()->Throwable_klass(), false);
340
341 insert_mem_bar_volatile(Op_MemBarCPUOrder, C->get_alias_index(TypeAryPtr::INLINES));
342 }
343
344 //------------------------------array_addressing-------------------------------
345 // Pull array and index from the stack. Compute pointer-to-element.
346 Node* Parse::array_addressing(BasicType type, int vals, const Type*& elemtype) {
347 Node *idx = peek(0+vals); // Get from stack without popping
348 Node *ary = peek(1+vals); // in case of exception
349
350 // Null check the array base, with correct stack contents
351 ary = null_check(ary, T_ARRAY);
352 // Compile-time detect of null-exception?
353 if (stopped()) return top();
354
355 const TypeAryPtr* arytype = _gvn.type(ary)->is_aryptr();
356 const TypeInt* sizetype = arytype->size();
357 elemtype = arytype->elem();
358
359 if (UseUniqueSubclasses) {
360 const Type* el = elemtype->make_ptr();
361 if (el && el->isa_instptr()) {
362 const TypeInstPtr* toop = el->is_instptr();
363 if (toop->instance_klass()->unique_concrete_subklass()) {
364 // If we load from "AbstractClass[]" we must see "ConcreteSubClass".
365 const Type* subklass = Type::get_const_type(toop->instance_klass());
366 elemtype = subklass->join_speculative(el);
367 }
368 }
369 }
370
371 if (!arytype->is_loaded()) {
372 // Only fails for some -Xcomp runs
373 // The class is unloaded. We have to run this bytecode in the interpreter.
374 ciKlass* klass = arytype->unloaded_klass();
375
376 uncommon_trap(Deoptimization::Reason_unloaded,
377 Deoptimization::Action_reinterpret,
378 klass, "!loaded array");
379 return top();
380 }
381
382 ary = create_speculative_inline_type_array_checks(ary, arytype, elemtype);
383
384 if (needs_range_check(sizetype, idx)) {
385 create_range_check(idx, ary, sizetype);
386 } else if (C->log() != nullptr) {
387 C->log()->elem("observe that='!need_range_check'");
388 }
389
390 // Check for always knowing you are throwing a range-check exception
391 if (stopped()) return top();
392
393 // Make array address computation control dependent to prevent it
394 // from floating above the range check during loop optimizations.
395 Node* ptr = array_element_address(ary, idx, type, sizetype, control());
396 assert(ptr != top(), "top should go hand-in-hand with stopped");
397
398 return ptr;
399 }
400
401 // 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
402 // be greater or equal the smallest possible array size (i.e. out-of-bounds).
403 bool Parse::needs_range_check(const TypeInt* size_type, const Node* index) const {
404 const TypeInt* index_type = _gvn.type(index)->is_int();
405 return index_type->_hi >= size_type->_lo || index_type->_lo < 0;
406 }
407
408 void Parse::create_range_check(Node* idx, Node* ary, const TypeInt* sizetype) {
409 Node* tst;
410 if (sizetype->_hi <= 0) {
411 // The greatest array bound is negative, so we can conclude that we're
412 // compiling unreachable code, but the unsigned compare trick used below
413 // only works with non-negative lengths. Instead, hack "tst" to be zero so
414 // the uncommon_trap path will always be taken.
415 tst = _gvn.intcon(0);
416 } else {
417 // Range is constant in array-oop, so we can use the original state of mem
418 Node* len = load_array_length(ary);
419
420 // Test length vs index (standard trick using unsigned compare)
421 Node* chk = _gvn.transform(new CmpUNode(idx, len) );
422 BoolTest::mask btest = BoolTest::lt;
423 tst = _gvn.transform(new BoolNode(chk, btest) );
424 }
425 RangeCheckNode* rc = new RangeCheckNode(control(), tst, PROB_MAX, COUNT_UNKNOWN);
426 _gvn.set_type(rc, rc->Value(&_gvn));
427 if (!tst->is_Con()) {
428 record_for_igvn(rc);
429 }
430 set_control(_gvn.transform(new IfTrueNode(rc)));
431 // Branch to failure if out of bounds
432 {
433 PreserveJVMState pjvms(this);
434 set_control(_gvn.transform(new IfFalseNode(rc)));
435 if (C->allow_range_check_smearing()) {
436 // Do not use builtin_throw, since range checks are sometimes
437 // made more stringent by an optimistic transformation.
438 // This creates "tentative" range checks at this point,
439 // which are not guaranteed to throw exceptions.
440 // See IfNode::Ideal, is_range_check, adjust_check.
441 uncommon_trap(Deoptimization::Reason_range_check,
442 Deoptimization::Action_make_not_entrant,
443 nullptr, "range_check");
444 } else {
445 // If we have already recompiled with the range-check-widening
446 // heroic optimization turned off, then we must really be throwing
447 // range check exceptions.
448 builtin_throw(Deoptimization::Reason_range_check);
449 }
450 }
451 }
452
453 // For inline type arrays, we can use the profiling information for array accesses to speculate on the type, flatness,
454 // and null-freeness. We can either prepare the speculative type for later uses or emit explicit speculative checks with
455 // traps now. In the latter case, the speculative type guarantees can avoid additional runtime checks later (e.g.
456 // non-null-free implies non-flat which allows us to remove flatness checks). This makes the graph simpler.
457 Node* Parse::create_speculative_inline_type_array_checks(Node* array, const TypeAryPtr* array_type,
458 const Type*& element_type) {
459 if (!array_type->is_flat() && !array_type->is_not_flat()) {
460 // For arrays that might be flat, speculate that the array has the exact type reported in the profile data such that
461 // we can rely on a fixed memory layout (i.e. either a flat layout or not).
462 array = cast_to_speculative_array_type(array, array_type, element_type);
463 } else if (UseTypeSpeculation && UseArrayLoadStoreProfile) {
464 // Array is known to be either flat or not flat. If possible, update the speculative type by using the profile data
465 // at this bci.
466 array = cast_to_profiled_array_type(array);
467 }
468
469 // Even though the type does not tell us whether we have an inline type array or not, we can still check the profile data
470 // whether we have a non-null-free or non-flat array. Speculating on a non-null-free array doesn't help aaload but could
471 // be profitable for a subsequent aastore.
472 if (!array_type->is_null_free() && !array_type->is_not_null_free()) {
473 array = speculate_non_null_free_array(array, array_type);
474 }
475 if (!array_type->is_flat() && !array_type->is_not_flat()) {
476 array = speculate_non_flat_array(array, array_type);
477 }
478 return array;
479 }
480
481 // Speculate that the array has the exact type reported in the profile data. We emit a trap when this turns out to be
482 // wrong. On the fast path, we add a CheckCastPP to use the exact type.
483 Node* Parse::cast_to_speculative_array_type(Node* const array, const TypeAryPtr*& array_type, const Type*& element_type) {
484 Deoptimization::DeoptReason reason = Deoptimization::Reason_speculate_class_check;
485 ciKlass* speculative_array_type = array_type->speculative_type();
486 if (too_many_traps_or_recompiles(reason) || speculative_array_type == nullptr) {
487 // No speculative type, check profile data at this bci
488 speculative_array_type = nullptr;
489 reason = Deoptimization::Reason_class_check;
490 if (UseArrayLoadStoreProfile && !too_many_traps_or_recompiles(reason)) {
491 ciKlass* profiled_element_type = nullptr;
492 ProfilePtrKind element_ptr = ProfileMaybeNull;
493 bool flat_array = true;
494 bool null_free_array = true;
495 method()->array_access_profiled_type(bci(), speculative_array_type, profiled_element_type, element_ptr, flat_array,
496 null_free_array);
497 }
498 }
499 if (speculative_array_type != nullptr) {
500 // Speculate that this array has the exact type reported by profile data
501 Node* casted_array = nullptr;
502 DEBUG_ONLY(Node* old_control = control();)
503 Node* slow_ctl = type_check_receiver(array, speculative_array_type, 1.0, &casted_array);
504 if (stopped()) {
505 // The check always fails and therefore profile information is incorrect. Don't use it.
506 assert(old_control == slow_ctl, "type check should have been removed");
507 set_control(slow_ctl);
508 } else if (!slow_ctl->is_top()) {
509 { PreserveJVMState pjvms(this);
510 set_control(slow_ctl);
511 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
512 }
513 replace_in_map(array, casted_array);
514 array_type = _gvn.type(casted_array)->is_aryptr();
515 element_type = array_type->elem();
516 return casted_array;
517 }
518 }
519 return array;
520 }
521
522 // Create a CheckCastPP when the speculative type can improve the current type.
523 Node* Parse::cast_to_profiled_array_type(Node* const array) {
524 ciKlass* array_type = nullptr;
525 ciKlass* element_type = nullptr;
526 ProfilePtrKind element_ptr = ProfileMaybeNull;
527 bool flat_array = true;
528 bool null_free_array = true;
529 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
530 if (array_type != nullptr) {
531 return record_profile_for_speculation(array, array_type, ProfileMaybeNull);
532 }
533 return array;
534 }
535
536 // Speculate that the array is non-null-free. We emit a trap when this turns out to be
537 // wrong. On the fast path, we add a CheckCastPP to use the non-null-free type.
538 Node* Parse::speculate_non_null_free_array(Node* const array, const TypeAryPtr*& array_type) {
539 bool null_free_array = true;
540 Deoptimization::DeoptReason reason = Deoptimization::Reason_none;
541 if (array_type->speculative() != nullptr &&
542 array_type->speculative()->is_aryptr()->is_not_null_free() &&
543 !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_class_check)) {
544 null_free_array = false;
545 reason = Deoptimization::Reason_speculate_class_check;
546 } else if (UseArrayLoadStoreProfile && !too_many_traps_or_recompiles(Deoptimization::Reason_class_check)) {
547 ciKlass* profiled_array_type = nullptr;
548 ciKlass* profiled_element_type = nullptr;
549 ProfilePtrKind element_ptr = ProfileMaybeNull;
550 bool flat_array = true;
551 method()->array_access_profiled_type(bci(), profiled_array_type, profiled_element_type, element_ptr, flat_array,
552 null_free_array);
553 reason = Deoptimization::Reason_class_check;
554 }
555 if (!null_free_array) {
556 { // Deoptimize if null-free array
557 BuildCutout unless(this, null_free_array_test(array, /* null_free = */ false), PROB_MAX);
558 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
559 }
560 assert(!stopped(), "null-free array should have been caught earlier");
561 Node* casted_array = _gvn.transform(new CheckCastPPNode(control(), array, array_type->cast_to_not_null_free()));
562 replace_in_map(array, casted_array);
563 array_type = _gvn.type(casted_array)->is_aryptr();
564 return casted_array;
565 }
566 return array;
567 }
568
569 // Speculate that the array is non-flat. We emit a trap when this turns out to be wrong.
570 // On the fast path, we add a CheckCastPP to use the non-flat type.
571 Node* Parse::speculate_non_flat_array(Node* const array, const TypeAryPtr* const array_type) {
572 bool flat_array = true;
573 Deoptimization::DeoptReason reason = Deoptimization::Reason_none;
574 if (array_type->speculative() != nullptr &&
575 array_type->speculative()->is_aryptr()->is_not_flat() &&
576 !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_class_check)) {
577 flat_array = false;
578 reason = Deoptimization::Reason_speculate_class_check;
579 } else if (UseArrayLoadStoreProfile && !too_many_traps_or_recompiles(reason)) {
580 ciKlass* profiled_array_type = nullptr;
581 ciKlass* profiled_element_type = nullptr;
582 ProfilePtrKind element_ptr = ProfileMaybeNull;
583 bool null_free_array = true;
584 method()->array_access_profiled_type(bci(), profiled_array_type, profiled_element_type, element_ptr, flat_array,
585 null_free_array);
586 reason = Deoptimization::Reason_class_check;
587 }
588 if (!flat_array) {
589 { // Deoptimize if flat array
590 BuildCutout unless(this, flat_array_test(array, /* flat = */ false), PROB_MAX);
591 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
592 }
593 assert(!stopped(), "flat array should have been caught earlier");
594 Node* casted_array = _gvn.transform(new CheckCastPPNode(control(), array, array_type->cast_to_not_flat()));
595 replace_in_map(array, casted_array);
596 return casted_array;
597 }
598 return array;
599 }
600
601 // returns IfNode
602 IfNode* Parse::jump_if_fork_int(Node* a, Node* b, BoolTest::mask mask, float prob, float cnt) {
603 Node *cmp = _gvn.transform(new CmpINode(a, b)); // two cases: shiftcount > 32 and shiftcount <= 32
604 Node *tst = _gvn.transform(new BoolNode(cmp, mask));
605 IfNode *iff = create_and_map_if(control(), tst, prob, cnt);
606 return iff;
607 }
608
609
610 // sentinel value for the target bci to mark never taken branches
611 // (according to profiling)
612 static const int never_reached = INT_MAX;
613
614 //------------------------------helper for tableswitch-------------------------
615 void Parse::jump_if_true_fork(IfNode *iff, int dest_bci_if_true, bool unc) {
616 // True branch, use existing map info
617 { PreserveJVMState pjvms(this);
618 Node *iftrue = _gvn.transform( new IfTrueNode (iff) );
619 set_control( iftrue );
620 if (unc) {
621 repush_if_args();
622 uncommon_trap(Deoptimization::Reason_unstable_if,
623 Deoptimization::Action_reinterpret,
624 nullptr,
625 "taken always");
626 } else {
627 assert(dest_bci_if_true != never_reached, "inconsistent dest");
628 merge_new_path(dest_bci_if_true);
629 }
630 }
631
632 // False branch
633 Node *iffalse = _gvn.transform( new IfFalseNode(iff) );
634 set_control( iffalse );
635 }
636
637 void Parse::jump_if_false_fork(IfNode *iff, int dest_bci_if_true, bool unc) {
638 // True branch, use existing map info
639 { PreserveJVMState pjvms(this);
640 Node *iffalse = _gvn.transform( new IfFalseNode (iff) );
641 set_control( iffalse );
642 if (unc) {
643 repush_if_args();
644 uncommon_trap(Deoptimization::Reason_unstable_if,
645 Deoptimization::Action_reinterpret,
646 nullptr,
647 "taken never");
648 } else {
649 assert(dest_bci_if_true != never_reached, "inconsistent dest");
650 merge_new_path(dest_bci_if_true);
651 }
652 }
653
654 // False branch
655 Node *iftrue = _gvn.transform( new IfTrueNode(iff) );
656 set_control( iftrue );
657 }
658
659 void Parse::jump_if_always_fork(int dest_bci, bool unc) {
660 // False branch, use existing map and control()
661 if (unc) {
662 repush_if_args();
663 uncommon_trap(Deoptimization::Reason_unstable_if,
664 Deoptimization::Action_reinterpret,
665 nullptr,
666 "taken never");
667 } else {
668 assert(dest_bci != never_reached, "inconsistent dest");
669 merge_new_path(dest_bci);
670 }
671 }
672
673
674 extern "C" {
675 static int jint_cmp(const void *i, const void *j) {
676 int a = *(jint *)i;
677 int b = *(jint *)j;
678 return a > b ? 1 : a < b ? -1 : 0;
679 }
680 }
681
682
683 class SwitchRange : public StackObj {
684 // a range of integers coupled with a bci destination
685 jint _lo; // inclusive lower limit
686 jint _hi; // inclusive upper limit
687 int _dest;
688 float _cnt; // how many times this range was hit according to profiling
689
690 public:
691 jint lo() const { return _lo; }
692 jint hi() const { return _hi; }
693 int dest() const { return _dest; }
694 bool is_singleton() const { return _lo == _hi; }
695 float cnt() const { return _cnt; }
696
697 void setRange(jint lo, jint hi, int dest, float cnt) {
698 assert(lo <= hi, "must be a non-empty range");
699 _lo = lo, _hi = hi; _dest = dest; _cnt = cnt;
700 assert(_cnt >= 0, "");
701 }
702 bool adjoinRange(jint lo, jint hi, int dest, float cnt, bool trim_ranges) {
703 assert(lo <= hi, "must be a non-empty range");
704 if (lo == _hi+1) {
705 // see merge_ranges() comment below
706 if (trim_ranges) {
707 if (cnt == 0) {
708 if (_cnt != 0) {
709 return false;
710 }
711 if (dest != _dest) {
712 _dest = never_reached;
713 }
714 } else {
715 if (_cnt == 0) {
716 return false;
717 }
718 if (dest != _dest) {
719 return false;
720 }
721 }
722 } else {
723 if (dest != _dest) {
724 return false;
725 }
726 }
727 _hi = hi;
728 _cnt += cnt;
729 return true;
730 }
731 return false;
732 }
733
734 void set (jint value, int dest, float cnt) {
735 setRange(value, value, dest, cnt);
736 }
737 bool adjoin(jint value, int dest, float cnt, bool trim_ranges) {
738 return adjoinRange(value, value, dest, cnt, trim_ranges);
739 }
740 bool adjoin(SwitchRange& other) {
741 return adjoinRange(other._lo, other._hi, other._dest, other._cnt, false);
742 }
743
744 void print() {
745 if (is_singleton())
746 tty->print(" {%d}=>%d (cnt=%f)", lo(), dest(), cnt());
747 else if (lo() == min_jint)
748 tty->print(" {..%d}=>%d (cnt=%f)", hi(), dest(), cnt());
749 else if (hi() == max_jint)
750 tty->print(" {%d..}=>%d (cnt=%f)", lo(), dest(), cnt());
751 else
752 tty->print(" {%d..%d}=>%d (cnt=%f)", lo(), hi(), dest(), cnt());
753 }
754 };
755
756 // We try to minimize the number of ranges and the size of the taken
757 // ones using profiling data. When ranges are created,
758 // SwitchRange::adjoinRange() only allows 2 adjoining ranges to merge
759 // if both were never hit or both were hit to build longer unreached
760 // ranges. Here, we now merge adjoining ranges with the same
761 // destination and finally set destination of unreached ranges to the
762 // special value never_reached because it can help minimize the number
763 // of tests that are necessary.
764 //
765 // For instance:
766 // [0, 1] to target1 sometimes taken
767 // [1, 2] to target1 never taken
768 // [2, 3] to target2 never taken
769 // would lead to:
770 // [0, 1] to target1 sometimes taken
771 // [1, 3] never taken
772 //
773 // (first 2 ranges to target1 are not merged)
774 static void merge_ranges(SwitchRange* ranges, int& rp) {
775 if (rp == 0) {
776 return;
777 }
778 int shift = 0;
779 for (int j = 0; j < rp; j++) {
780 SwitchRange& r1 = ranges[j-shift];
781 SwitchRange& r2 = ranges[j+1];
782 if (r1.adjoin(r2)) {
783 shift++;
784 } else if (shift > 0) {
785 ranges[j+1-shift] = r2;
786 }
787 }
788 rp -= shift;
789 for (int j = 0; j <= rp; j++) {
790 SwitchRange& r = ranges[j];
791 if (r.cnt() == 0 && r.dest() != never_reached) {
792 r.setRange(r.lo(), r.hi(), never_reached, r.cnt());
793 }
794 }
795 }
796
797 //-------------------------------do_tableswitch--------------------------------
798 void Parse::do_tableswitch() {
799 // Get information about tableswitch
800 int default_dest = iter().get_dest_table(0);
801 jint lo_index = iter().get_int_table(1);
802 jint hi_index = iter().get_int_table(2);
803 int len = hi_index - lo_index + 1;
804
805 if (len < 1) {
806 // If this is a backward branch, add safepoint
807 maybe_add_safepoint(default_dest);
808 pop(); // the effect of the instruction execution on the operand stack
809 merge(default_dest);
810 return;
811 }
812
813 ciMethodData* methodData = method()->method_data();
814 ciMultiBranchData* profile = nullptr;
815 if (methodData->is_mature() && UseSwitchProfiling) {
816 ciProfileData* data = methodData->bci_to_data(bci());
817 if (data != nullptr && data->is_MultiBranchData()) {
818 profile = (ciMultiBranchData*)data;
819 }
820 }
821 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if);
822
823 // generate decision tree, using trichotomy when possible
824 int rnum = len+2;
825 bool makes_backward_branch = (default_dest <= bci());
826 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum);
827 int rp = -1;
828 if (lo_index != min_jint) {
829 float cnt = 1.0F;
830 if (profile != nullptr) {
831 cnt = (float)profile->default_count() / (hi_index != max_jint ? 2.0F : 1.0F);
832 }
833 ranges[++rp].setRange(min_jint, lo_index-1, default_dest, cnt);
834 }
835 for (int j = 0; j < len; j++) {
836 jint match_int = lo_index+j;
837 int dest = iter().get_dest_table(j+3);
838 makes_backward_branch |= (dest <= bci());
839 float cnt = 1.0F;
840 if (profile != nullptr) {
841 cnt = (float)profile->count_at(j);
842 }
843 if (rp < 0 || !ranges[rp].adjoin(match_int, dest, cnt, trim_ranges)) {
844 ranges[++rp].set(match_int, dest, cnt);
845 }
846 }
847 jint highest = lo_index+(len-1);
848 assert(ranges[rp].hi() == highest, "");
849 if (highest != max_jint) {
850 float cnt = 1.0F;
851 if (profile != nullptr) {
852 cnt = (float)profile->default_count() / (lo_index != min_jint ? 2.0F : 1.0F);
853 }
854 if (!ranges[rp].adjoinRange(highest+1, max_jint, default_dest, cnt, trim_ranges)) {
855 ranges[++rp].setRange(highest+1, max_jint, default_dest, cnt);
856 }
857 }
858 assert(rp < len+2, "not too many ranges");
859
860 if (trim_ranges) {
861 merge_ranges(ranges, rp);
862 }
863
864 // Safepoint in case if backward branch observed
865 if (makes_backward_branch) {
866 add_safepoint();
867 }
868
869 Node* lookup = pop(); // lookup value
870 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]);
871 }
872
873
874 //------------------------------do_lookupswitch--------------------------------
875 void Parse::do_lookupswitch() {
876 // Get information about lookupswitch
877 int default_dest = iter().get_dest_table(0);
878 jint len = iter().get_int_table(1);
879
880 if (len < 1) { // If this is a backward branch, add safepoint
881 maybe_add_safepoint(default_dest);
882 pop(); // the effect of the instruction execution on the operand stack
883 merge(default_dest);
884 return;
885 }
886
887 ciMethodData* methodData = method()->method_data();
888 ciMultiBranchData* profile = nullptr;
889 if (methodData->is_mature() && UseSwitchProfiling) {
890 ciProfileData* data = methodData->bci_to_data(bci());
891 if (data != nullptr && data->is_MultiBranchData()) {
892 profile = (ciMultiBranchData*)data;
893 }
894 }
895 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if);
896
897 // generate decision tree, using trichotomy when possible
898 jint* table = NEW_RESOURCE_ARRAY(jint, len*3);
899 {
900 for (int j = 0; j < len; j++) {
901 table[3*j+0] = iter().get_int_table(2+2*j);
902 table[3*j+1] = iter().get_dest_table(2+2*j+1);
903 // Handle overflow when converting from uint to jint
904 table[3*j+2] = (profile == nullptr) ? 1 : (jint)MIN2<uint>((uint)max_jint, profile->count_at(j));
905 }
906 qsort(table, len, 3*sizeof(table[0]), jint_cmp);
907 }
908
909 float default_cnt = 1.0F;
910 if (profile != nullptr) {
911 juint defaults = max_juint - len;
912 default_cnt = (float)profile->default_count()/(float)defaults;
913 }
914
915 int rnum = len*2+1;
916 bool makes_backward_branch = (default_dest <= bci());
917 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum);
918 int rp = -1;
919 for (int j = 0; j < len; j++) {
920 jint match_int = table[3*j+0];
921 jint dest = table[3*j+1];
922 jint cnt = table[3*j+2];
923 jint next_lo = rp < 0 ? min_jint : ranges[rp].hi()+1;
924 makes_backward_branch |= (dest <= bci());
925 float c = default_cnt * ((float)match_int - (float)next_lo);
926 if (match_int != next_lo && (rp < 0 || !ranges[rp].adjoinRange(next_lo, match_int-1, default_dest, c, trim_ranges))) {
927 assert(default_dest != never_reached, "sentinel value for dead destinations");
928 ranges[++rp].setRange(next_lo, match_int-1, default_dest, c);
929 }
930 if (rp < 0 || !ranges[rp].adjoin(match_int, dest, (float)cnt, trim_ranges)) {
931 assert(dest != never_reached, "sentinel value for dead destinations");
932 ranges[++rp].set(match_int, dest, (float)cnt);
933 }
934 }
935 jint highest = table[3*(len-1)];
936 assert(ranges[rp].hi() == highest, "");
937 if (highest != max_jint &&
938 !ranges[rp].adjoinRange(highest+1, max_jint, default_dest, default_cnt * ((float)max_jint - (float)highest), trim_ranges)) {
939 ranges[++rp].setRange(highest+1, max_jint, default_dest, default_cnt * ((float)max_jint - (float)highest));
940 }
941 assert(rp < rnum, "not too many ranges");
942
943 if (trim_ranges) {
944 merge_ranges(ranges, rp);
945 }
946
947 // Safepoint in case backward branch observed
948 if (makes_backward_branch) {
949 add_safepoint();
950 }
951
952 Node *lookup = pop(); // lookup value
953 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]);
954 }
955
956 static float if_prob(float taken_cnt, float total_cnt) {
957 assert(taken_cnt <= total_cnt, "");
958 if (total_cnt == 0) {
959 return PROB_FAIR;
960 }
961 float p = taken_cnt / total_cnt;
962 return clamp(p, PROB_MIN, PROB_MAX);
963 }
964
965 static float if_cnt(float cnt) {
966 if (cnt == 0) {
967 return COUNT_UNKNOWN;
968 }
969 return cnt;
970 }
971
972 static float sum_of_cnts(SwitchRange *lo, SwitchRange *hi) {
973 float total_cnt = 0;
974 for (SwitchRange* sr = lo; sr <= hi; sr++) {
975 total_cnt += sr->cnt();
976 }
977 return total_cnt;
978 }
979
980 class SwitchRanges : public ResourceObj {
981 public:
982 SwitchRange* _lo;
983 SwitchRange* _hi;
984 SwitchRange* _mid;
985 float _cost;
986
987 enum {
988 Start,
989 LeftDone,
990 RightDone,
991 Done
992 } _state;
993
994 SwitchRanges(SwitchRange *lo, SwitchRange *hi)
995 : _lo(lo), _hi(hi), _mid(nullptr),
996 _cost(0), _state(Start) {
997 }
998
999 SwitchRanges()
1000 : _lo(nullptr), _hi(nullptr), _mid(nullptr),
1001 _cost(0), _state(Start) {}
1002 };
1003
1004 // Estimate cost of performing a binary search on lo..hi
1005 static float compute_tree_cost(SwitchRange *lo, SwitchRange *hi, float total_cnt) {
1006 GrowableArray<SwitchRanges> tree;
1007 SwitchRanges root(lo, hi);
1008 tree.push(root);
1009
1010 float cost = 0;
1011 do {
1012 SwitchRanges& r = *tree.adr_at(tree.length()-1);
1013 if (r._hi != r._lo) {
1014 if (r._mid == nullptr) {
1015 float r_cnt = sum_of_cnts(r._lo, r._hi);
1016
1017 if (r_cnt == 0) {
1018 tree.pop();
1019 cost = 0;
1020 continue;
1021 }
1022
1023 SwitchRange* mid = nullptr;
1024 mid = r._lo;
1025 for (float cnt = 0; ; ) {
1026 assert(mid <= r._hi, "out of bounds");
1027 cnt += mid->cnt();
1028 if (cnt > r_cnt / 2) {
1029 break;
1030 }
1031 mid++;
1032 }
1033 assert(mid <= r._hi, "out of bounds");
1034 r._mid = mid;
1035 r._cost = r_cnt / total_cnt;
1036 }
1037 r._cost += cost;
1038 if (r._state < SwitchRanges::LeftDone && r._mid > r._lo) {
1039 cost = 0;
1040 r._state = SwitchRanges::LeftDone;
1041 tree.push(SwitchRanges(r._lo, r._mid-1));
1042 } else if (r._state < SwitchRanges::RightDone) {
1043 cost = 0;
1044 r._state = SwitchRanges::RightDone;
1045 tree.push(SwitchRanges(r._mid == r._lo ? r._mid+1 : r._mid, r._hi));
1046 } else {
1047 tree.pop();
1048 cost = r._cost;
1049 }
1050 } else {
1051 tree.pop();
1052 cost = r._cost;
1053 }
1054 } while (tree.length() > 0);
1055
1056
1057 return cost;
1058 }
1059
1060 // It sometimes pays off to test most common ranges before the binary search
1061 void Parse::linear_search_switch_ranges(Node* key_val, SwitchRange*& lo, SwitchRange*& hi) {
1062 uint nr = hi - lo + 1;
1063 float total_cnt = sum_of_cnts(lo, hi);
1064
1065 float min = compute_tree_cost(lo, hi, total_cnt);
1066 float extra = 1;
1067 float sub = 0;
1068
1069 SwitchRange* array1 = lo;
1070 SwitchRange* array2 = NEW_RESOURCE_ARRAY(SwitchRange, nr);
1071
1072 SwitchRange* ranges = nullptr;
1073
1074 while (nr >= 2) {
1075 assert(lo == array1 || lo == array2, "one the 2 already allocated arrays");
1076 ranges = (lo == array1) ? array2 : array1;
1077
1078 // Find highest frequency range
1079 SwitchRange* candidate = lo;
1080 for (SwitchRange* sr = lo+1; sr <= hi; sr++) {
1081 if (sr->cnt() > candidate->cnt()) {
1082 candidate = sr;
1083 }
1084 }
1085 SwitchRange most_freq = *candidate;
1086 if (most_freq.cnt() == 0) {
1087 break;
1088 }
1089
1090 // Copy remaining ranges into another array
1091 int shift = 0;
1092 for (uint i = 0; i < nr; i++) {
1093 SwitchRange* sr = &lo[i];
1094 if (sr != candidate) {
1095 ranges[i-shift] = *sr;
1096 } else {
1097 shift++;
1098 if (i > 0 && i < nr-1) {
1099 SwitchRange prev = lo[i-1];
1100 prev.setRange(prev.lo(), sr->hi(), prev.dest(), prev.cnt());
1101 if (prev.adjoin(lo[i+1])) {
1102 shift++;
1103 i++;
1104 }
1105 ranges[i-shift] = prev;
1106 }
1107 }
1108 }
1109 nr -= shift;
1110
1111 // Evaluate cost of testing the most common range and performing a
1112 // binary search on the other ranges
1113 float cost = extra + compute_tree_cost(&ranges[0], &ranges[nr-1], total_cnt);
1114 if (cost >= min) {
1115 break;
1116 }
1117 // swap arrays
1118 lo = &ranges[0];
1119 hi = &ranges[nr-1];
1120
1121 // It pays off: emit the test for the most common range
1122 assert(most_freq.cnt() > 0, "must be taken");
1123 Node* val = _gvn.transform(new SubINode(key_val, _gvn.intcon(most_freq.lo())));
1124 Node* cmp = _gvn.transform(new CmpUNode(val, _gvn.intcon(java_subtract(most_freq.hi(), most_freq.lo()))));
1125 Node* tst = _gvn.transform(new BoolNode(cmp, BoolTest::le));
1126 IfNode* iff = create_and_map_if(control(), tst, if_prob(most_freq.cnt(), total_cnt), if_cnt(most_freq.cnt()));
1127 jump_if_true_fork(iff, most_freq.dest(), false);
1128
1129 sub += most_freq.cnt() / total_cnt;
1130 extra += 1 - sub;
1131 min = cost;
1132 }
1133 }
1134
1135 //----------------------------create_jump_tables-------------------------------
1136 bool Parse::create_jump_tables(Node* key_val, SwitchRange* lo, SwitchRange* hi) {
1137 // Are jumptables enabled
1138 if (!UseJumpTables) return false;
1139
1140 // Are jumptables supported
1141 if (!Matcher::has_match_rule(Op_Jump)) return false;
1142
1143 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if);
1144
1145 // Decide if a guard is needed to lop off big ranges at either (or
1146 // both) end(s) of the input set. We'll call this the default target
1147 // even though we can't be sure that it is the true "default".
1148
1149 bool needs_guard = false;
1150 int default_dest;
1151 int64_t total_outlier_size = 0;
1152 int64_t hi_size = ((int64_t)hi->hi()) - ((int64_t)hi->lo()) + 1;
1153 int64_t lo_size = ((int64_t)lo->hi()) - ((int64_t)lo->lo()) + 1;
1154
1155 if (lo->dest() == hi->dest()) {
1156 total_outlier_size = hi_size + lo_size;
1157 default_dest = lo->dest();
1158 } else if (lo_size > hi_size) {
1159 total_outlier_size = lo_size;
1160 default_dest = lo->dest();
1161 } else {
1162 total_outlier_size = hi_size;
1163 default_dest = hi->dest();
1164 }
1165
1166 float total = sum_of_cnts(lo, hi);
1167 float cost = compute_tree_cost(lo, hi, total);
1168
1169 // If a guard test will eliminate very sparse end ranges, then
1170 // it is worth the cost of an extra jump.
1171 float trimmed_cnt = 0;
1172 if (total_outlier_size > (MaxJumpTableSparseness * 4)) {
1173 needs_guard = true;
1174 if (default_dest == lo->dest()) {
1175 trimmed_cnt += lo->cnt();
1176 lo++;
1177 }
1178 if (default_dest == hi->dest()) {
1179 trimmed_cnt += hi->cnt();
1180 hi--;
1181 }
1182 }
1183
1184 // Find the total number of cases and ranges
1185 int64_t num_cases = ((int64_t)hi->hi()) - ((int64_t)lo->lo()) + 1;
1186 int num_range = hi - lo + 1;
1187
1188 // Don't create table if: too large, too small, or too sparse.
1189 if (num_cases > MaxJumpTableSize)
1190 return false;
1191 if (UseSwitchProfiling) {
1192 // MinJumpTableSize is set so with a well balanced binary tree,
1193 // when the number of ranges is MinJumpTableSize, it's cheaper to
1194 // go through a JumpNode that a tree of IfNodes. Average cost of a
1195 // tree of IfNodes with MinJumpTableSize is
1196 // log2f(MinJumpTableSize) comparisons. So if the cost computed
1197 // from profile data is less than log2f(MinJumpTableSize) then
1198 // going with the binary search is cheaper.
1199 if (cost < log2f(MinJumpTableSize)) {
1200 return false;
1201 }
1202 } else {
1203 if (num_cases < MinJumpTableSize)
1204 return false;
1205 }
1206 if (num_cases > (MaxJumpTableSparseness * num_range))
1207 return false;
1208
1209 // Normalize table lookups to zero
1210 int lowval = lo->lo();
1211 key_val = _gvn.transform( new SubINode(key_val, _gvn.intcon(lowval)) );
1212
1213 // Generate a guard to protect against input keyvals that aren't
1214 // in the switch domain.
1215 if (needs_guard) {
1216 Node* size = _gvn.intcon(num_cases);
1217 Node* cmp = _gvn.transform(new CmpUNode(key_val, size));
1218 Node* tst = _gvn.transform(new BoolNode(cmp, BoolTest::ge));
1219 IfNode* iff = create_and_map_if(control(), tst, if_prob(trimmed_cnt, total), if_cnt(trimmed_cnt));
1220 jump_if_true_fork(iff, default_dest, trim_ranges && trimmed_cnt == 0);
1221
1222 total -= trimmed_cnt;
1223 }
1224
1225 // Create an ideal node JumpTable that has projections
1226 // of all possible ranges for a switch statement
1227 // The key_val input must be converted to a pointer offset and scaled.
1228 // Compare Parse::array_addressing above.
1229
1230 // Clean the 32-bit int into a real 64-bit offset.
1231 // Otherwise, the jint value 0 might turn into an offset of 0x0800000000.
1232 // Make I2L conversion control dependent to prevent it from
1233 // floating above the range check during loop optimizations.
1234 // Do not use a narrow int type here to prevent the data path from dying
1235 // while the control path is not removed. This can happen if the type of key_val
1236 // is later known to be out of bounds of [0, num_cases] and therefore a narrow cast
1237 // would be replaced by TOP while C2 is not able to fold the corresponding range checks.
1238 // Set _carry_dependency for the cast to avoid being removed by IGVN.
1239 #ifdef _LP64
1240 key_val = C->constrained_convI2L(&_gvn, key_val, TypeInt::INT, control(), true /* carry_dependency */);
1241 #endif
1242
1243 // Shift the value by wordsize so we have an index into the table, rather
1244 // than a switch value
1245 Node *shiftWord = _gvn.MakeConX(wordSize);
1246 key_val = _gvn.transform( new MulXNode( key_val, shiftWord));
1247
1248 // Create the JumpNode
1249 Arena* arena = C->comp_arena();
1250 float* probs = (float*)arena->Amalloc(sizeof(float)*num_cases);
1251 int i = 0;
1252 if (total == 0) {
1253 for (SwitchRange* r = lo; r <= hi; r++) {
1254 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) {
1255 probs[i] = 1.0F / num_cases;
1256 }
1257 }
1258 } else {
1259 for (SwitchRange* r = lo; r <= hi; r++) {
1260 float prob = r->cnt()/total;
1261 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) {
1262 probs[i] = prob / (r->hi() - r->lo() + 1);
1263 }
1264 }
1265 }
1266
1267 ciMethodData* methodData = method()->method_data();
1268 ciMultiBranchData* profile = nullptr;
1269 if (methodData->is_mature()) {
1270 ciProfileData* data = methodData->bci_to_data(bci());
1271 if (data != nullptr && data->is_MultiBranchData()) {
1272 profile = (ciMultiBranchData*)data;
1273 }
1274 }
1275
1276 Node* jtn = _gvn.transform(new JumpNode(control(), key_val, num_cases, probs, profile == nullptr ? COUNT_UNKNOWN : total));
1277
1278 // These are the switch destinations hanging off the jumpnode
1279 i = 0;
1280 for (SwitchRange* r = lo; r <= hi; r++) {
1281 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) {
1282 Node* input = _gvn.transform(new JumpProjNode(jtn, i, r->dest(), (int)(j - lowval)));
1283 {
1284 PreserveJVMState pjvms(this);
1285 set_control(input);
1286 jump_if_always_fork(r->dest(), trim_ranges && r->cnt() == 0);
1287 }
1288 }
1289 }
1290 assert(i == num_cases, "miscount of cases");
1291 stop_and_kill_map(); // no more uses for this JVMS
1292 return true;
1293 }
1294
1295 //----------------------------jump_switch_ranges-------------------------------
1296 void Parse::jump_switch_ranges(Node* key_val, SwitchRange *lo, SwitchRange *hi, int switch_depth) {
1297 Block* switch_block = block();
1298 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if);
1299
1300 if (switch_depth == 0) {
1301 // Do special processing for the top-level call.
1302 assert(lo->lo() == min_jint, "initial range must exhaust Type::INT");
1303 assert(hi->hi() == max_jint, "initial range must exhaust Type::INT");
1304
1305 // Decrement pred-numbers for the unique set of nodes.
1306 #ifdef ASSERT
1307 if (!trim_ranges) {
1308 // Ensure that the block's successors are a (duplicate-free) set.
1309 int successors_counted = 0; // block occurrences in [hi..lo]
1310 int unique_successors = switch_block->num_successors();
1311 for (int i = 0; i < unique_successors; i++) {
1312 Block* target = switch_block->successor_at(i);
1313
1314 // Check that the set of successors is the same in both places.
1315 int successors_found = 0;
1316 for (SwitchRange* p = lo; p <= hi; p++) {
1317 if (p->dest() == target->start()) successors_found++;
1318 }
1319 assert(successors_found > 0, "successor must be known");
1320 successors_counted += successors_found;
1321 }
1322 assert(successors_counted == (hi-lo)+1, "no unexpected successors");
1323 }
1324 #endif
1325
1326 // Maybe prune the inputs, based on the type of key_val.
1327 jint min_val = min_jint;
1328 jint max_val = max_jint;
1329 const TypeInt* ti = key_val->bottom_type()->isa_int();
1330 if (ti != nullptr) {
1331 min_val = ti->_lo;
1332 max_val = ti->_hi;
1333 assert(min_val <= max_val, "invalid int type");
1334 }
1335 while (lo->hi() < min_val) {
1336 lo++;
1337 }
1338 if (lo->lo() < min_val) {
1339 lo->setRange(min_val, lo->hi(), lo->dest(), lo->cnt());
1340 }
1341 while (hi->lo() > max_val) {
1342 hi--;
1343 }
1344 if (hi->hi() > max_val) {
1345 hi->setRange(hi->lo(), max_val, hi->dest(), hi->cnt());
1346 }
1347
1348 linear_search_switch_ranges(key_val, lo, hi);
1349 }
1350
1351 #ifndef PRODUCT
1352 if (switch_depth == 0) {
1353 _max_switch_depth = 0;
1354 _est_switch_depth = log2i_graceful((hi - lo + 1) - 1) + 1;
1355 }
1356 #endif
1357
1358 assert(lo <= hi, "must be a non-empty set of ranges");
1359 if (lo == hi) {
1360 jump_if_always_fork(lo->dest(), trim_ranges && lo->cnt() == 0);
1361 } else {
1362 assert(lo->hi() == (lo+1)->lo()-1, "contiguous ranges");
1363 assert(hi->lo() == (hi-1)->hi()+1, "contiguous ranges");
1364
1365 if (create_jump_tables(key_val, lo, hi)) return;
1366
1367 SwitchRange* mid = nullptr;
1368 float total_cnt = sum_of_cnts(lo, hi);
1369
1370 int nr = hi - lo + 1;
1371 if (UseSwitchProfiling) {
1372 // Don't keep the binary search tree balanced: pick up mid point
1373 // that split frequencies in half.
1374 float cnt = 0;
1375 for (SwitchRange* sr = lo; sr <= hi; sr++) {
1376 cnt += sr->cnt();
1377 if (cnt >= total_cnt / 2) {
1378 mid = sr;
1379 break;
1380 }
1381 }
1382 } else {
1383 mid = lo + nr/2;
1384
1385 // if there is an easy choice, pivot at a singleton:
1386 if (nr > 3 && !mid->is_singleton() && (mid-1)->is_singleton()) mid--;
1387
1388 assert(lo < mid && mid <= hi, "good pivot choice");
1389 assert(nr != 2 || mid == hi, "should pick higher of 2");
1390 assert(nr != 3 || mid == hi-1, "should pick middle of 3");
1391 }
1392
1393
1394 Node *test_val = _gvn.intcon(mid == lo ? mid->hi() : mid->lo());
1395
1396 if (mid->is_singleton()) {
1397 IfNode *iff_ne = jump_if_fork_int(key_val, test_val, BoolTest::ne, 1-if_prob(mid->cnt(), total_cnt), if_cnt(mid->cnt()));
1398 jump_if_false_fork(iff_ne, mid->dest(), trim_ranges && mid->cnt() == 0);
1399
1400 // Special Case: If there are exactly three ranges, and the high
1401 // and low range each go to the same place, omit the "gt" test,
1402 // since it will not discriminate anything.
1403 bool eq_test_only = (hi == lo+2 && hi->dest() == lo->dest() && mid == hi-1) || mid == lo;
1404
1405 // if there is a higher range, test for it and process it:
1406 if (mid < hi && !eq_test_only) {
1407 // two comparisons of same values--should enable 1 test for 2 branches
1408 // Use BoolTest::lt instead of BoolTest::gt
1409 float cnt = sum_of_cnts(lo, mid-1);
1410 IfNode *iff_lt = jump_if_fork_int(key_val, test_val, BoolTest::lt, if_prob(cnt, total_cnt), if_cnt(cnt));
1411 Node *iftrue = _gvn.transform( new IfTrueNode(iff_lt) );
1412 Node *iffalse = _gvn.transform( new IfFalseNode(iff_lt) );
1413 { PreserveJVMState pjvms(this);
1414 set_control(iffalse);
1415 jump_switch_ranges(key_val, mid+1, hi, switch_depth+1);
1416 }
1417 set_control(iftrue);
1418 }
1419
1420 } else {
1421 // mid is a range, not a singleton, so treat mid..hi as a unit
1422 float cnt = sum_of_cnts(mid == lo ? mid+1 : mid, hi);
1423 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));
1424
1425 // if there is a higher range, test for it and process it:
1426 if (mid == hi) {
1427 jump_if_true_fork(iff_ge, mid->dest(), trim_ranges && cnt == 0);
1428 } else {
1429 Node *iftrue = _gvn.transform( new IfTrueNode(iff_ge) );
1430 Node *iffalse = _gvn.transform( new IfFalseNode(iff_ge) );
1431 { PreserveJVMState pjvms(this);
1432 set_control(iftrue);
1433 jump_switch_ranges(key_val, mid == lo ? mid+1 : mid, hi, switch_depth+1);
1434 }
1435 set_control(iffalse);
1436 }
1437 }
1438
1439 // in any case, process the lower range
1440 if (mid == lo) {
1441 if (mid->is_singleton()) {
1442 jump_switch_ranges(key_val, lo+1, hi, switch_depth+1);
1443 } else {
1444 jump_if_always_fork(lo->dest(), trim_ranges && lo->cnt() == 0);
1445 }
1446 } else {
1447 jump_switch_ranges(key_val, lo, mid-1, switch_depth+1);
1448 }
1449 }
1450
1451 // Decrease pred_count for each successor after all is done.
1452 if (switch_depth == 0) {
1453 int unique_successors = switch_block->num_successors();
1454 for (int i = 0; i < unique_successors; i++) {
1455 Block* target = switch_block->successor_at(i);
1456 // Throw away the pre-allocated path for each unique successor.
1457 target->next_path_num();
1458 }
1459 }
1460
1461 #ifndef PRODUCT
1462 _max_switch_depth = MAX2(switch_depth, _max_switch_depth);
1463 if (TraceOptoParse && Verbose && WizardMode && switch_depth == 0) {
1464 SwitchRange* r;
1465 int nsing = 0;
1466 for( r = lo; r <= hi; r++ ) {
1467 if( r->is_singleton() ) nsing++;
1468 }
1469 tty->print(">>> ");
1470 _method->print_short_name();
1471 tty->print_cr(" switch decision tree");
1472 tty->print_cr(" %d ranges (%d singletons), max_depth=%d, est_depth=%d",
1473 (int) (hi-lo+1), nsing, _max_switch_depth, _est_switch_depth);
1474 if (_max_switch_depth > _est_switch_depth) {
1475 tty->print_cr("******** BAD SWITCH DEPTH ********");
1476 }
1477 tty->print(" ");
1478 for( r = lo; r <= hi; r++ ) {
1479 r->print();
1480 }
1481 tty->cr();
1482 }
1483 #endif
1484 }
1485
1486 Node* Parse::floating_point_mod(Node* a, Node* b, BasicType type) {
1487 assert(type == BasicType::T_FLOAT || type == BasicType::T_DOUBLE, "only float and double are floating points");
1488 CallNode* mod = type == BasicType::T_DOUBLE ? static_cast<CallNode*>(new ModDNode(C, a, b)) : new ModFNode(C, a, b);
1489
1490 Node* prev_mem = set_predefined_input_for_runtime_call(mod);
1491 mod = _gvn.transform(mod)->as_Call();
1492 set_predefined_output_for_runtime_call(mod, prev_mem, TypeRawPtr::BOTTOM);
1493 Node* result = _gvn.transform(new ProjNode(mod, TypeFunc::Parms + 0));
1494 record_for_igvn(mod);
1495 return result;
1496 }
1497
1498 void Parse::l2f() {
1499 Node* f2 = pop();
1500 Node* f1 = pop();
1501 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::l2f_Type(),
1502 CAST_FROM_FN_PTR(address, SharedRuntime::l2f),
1503 "l2f", nullptr, //no memory effects
1504 f1, f2);
1505 Node* res = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0));
1506
1507 push(res);
1508 }
1509
1510 // Handle jsr and jsr_w bytecode
1511 void Parse::do_jsr() {
1512 assert(bc() == Bytecodes::_jsr || bc() == Bytecodes::_jsr_w, "wrong bytecode");
1513
1514 // Store information about current state, tagged with new _jsr_bci
1515 int return_bci = iter().next_bci();
1516 int jsr_bci = (bc() == Bytecodes::_jsr) ? iter().get_dest() : iter().get_far_dest();
1517
1518 // The way we do things now, there is only one successor block
1519 // for the jsr, because the target code is cloned by ciTypeFlow.
1520 Block* target = successor_for_bci(jsr_bci);
1521
1522 // What got pushed?
1523 const Type* ret_addr = target->peek();
1524 assert(ret_addr->singleton(), "must be a constant (cloned jsr body)");
1525
1526 // Effect on jsr on stack
1527 push(_gvn.makecon(ret_addr));
1528
1529 // Flow to the jsr.
1530 merge(jsr_bci);
1531 }
1532
1533 // Handle ret bytecode
1534 void Parse::do_ret() {
1535 // Find to whom we return.
1536 assert(block()->num_successors() == 1, "a ret can only go one place now");
1537 Block* target = block()->successor_at(0);
1538 assert(!target->is_ready(), "our arrival must be expected");
1539 int pnum = target->next_path_num();
1540 merge_common(target, pnum);
1541 }
1542
1543 static bool has_injected_profile(BoolTest::mask btest, Node* test, int& taken, int& not_taken) {
1544 if (btest != BoolTest::eq && btest != BoolTest::ne) {
1545 // Only ::eq and ::ne are supported for profile injection.
1546 return false;
1547 }
1548 if (test->is_Cmp() &&
1549 test->in(1)->Opcode() == Op_ProfileBoolean) {
1550 ProfileBooleanNode* profile = (ProfileBooleanNode*)test->in(1);
1551 int false_cnt = profile->false_count();
1552 int true_cnt = profile->true_count();
1553
1554 // Counts matching depends on the actual test operation (::eq or ::ne).
1555 // No need to scale the counts because profile injection was designed
1556 // to feed exact counts into VM.
1557 taken = (btest == BoolTest::eq) ? false_cnt : true_cnt;
1558 not_taken = (btest == BoolTest::eq) ? true_cnt : false_cnt;
1559
1560 profile->consume();
1561 return true;
1562 }
1563 return false;
1564 }
1565
1566 // Give up if too few (or too many, in which case the sum will overflow) counts to be meaningful.
1567 // We also check that individual counters are positive first, otherwise the sum can become positive.
1568 // (check for saturation, integer overflow, and immature counts)
1569 static bool counters_are_meaningful(int counter1, int counter2, int min) {
1570 // check for saturation, including "uint" values too big to fit in "int"
1571 if (counter1 < 0 || counter2 < 0) {
1572 return false;
1573 }
1574 // check for integer overflow of the sum
1575 int64_t sum = (int64_t)counter1 + (int64_t)counter2;
1576 STATIC_ASSERT(sizeof(counter1) < sizeof(sum));
1577 if (sum > INT_MAX) {
1578 return false;
1579 }
1580 // check if mature
1581 return (counter1 + counter2) >= min;
1582 }
1583
1584 //--------------------------dynamic_branch_prediction--------------------------
1585 // Try to gather dynamic branch prediction behavior. Return a probability
1586 // of the branch being taken and set the "cnt" field. Returns a -1.0
1587 // if we need to use static prediction for some reason.
1588 float Parse::dynamic_branch_prediction(float &cnt, BoolTest::mask btest, Node* test) {
1589 ResourceMark rm;
1590
1591 cnt = COUNT_UNKNOWN;
1592
1593 int taken = 0;
1594 int not_taken = 0;
1595
1596 bool use_mdo = !has_injected_profile(btest, test, taken, not_taken);
1597
1598 if (use_mdo) {
1599 // Use MethodData information if it is available
1600 // FIXME: free the ProfileData structure
1601 ciMethodData* methodData = method()->method_data();
1602 if (!methodData->is_mature()) return PROB_UNKNOWN;
1603 ciProfileData* data = methodData->bci_to_data(bci());
1604 if (data == nullptr) {
1605 return PROB_UNKNOWN;
1606 }
1607 if (!data->is_JumpData()) return PROB_UNKNOWN;
1608
1609 // get taken and not taken values
1610 // NOTE: saturated UINT_MAX values become negative,
1611 // as do counts above INT_MAX.
1612 taken = data->as_JumpData()->taken();
1613 not_taken = 0;
1614 if (data->is_BranchData()) {
1615 not_taken = data->as_BranchData()->not_taken();
1616 }
1617
1618 // scale the counts to be commensurate with invocation counts:
1619 // NOTE: overflow for positive values is clamped at INT_MAX
1620 taken = method()->scale_count(taken);
1621 not_taken = method()->scale_count(not_taken);
1622 }
1623 // At this point, saturation or overflow is indicated by INT_MAX
1624 // or a negative value.
1625
1626 // Give up if too few (or too many, in which case the sum will overflow) counts to be meaningful.
1627 // We also check that individual counters are positive first, otherwise the sum can become positive.
1628 if (!counters_are_meaningful(taken, not_taken, 40)) {
1629 if (C->log() != nullptr) {
1630 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d'", iter().get_dest(), taken, not_taken);
1631 }
1632 return PROB_UNKNOWN;
1633 }
1634
1635 // Compute frequency that we arrive here
1636 float sum = taken + not_taken;
1637 // Adjust, if this block is a cloned private block but the
1638 // Jump counts are shared. Taken the private counts for
1639 // just this path instead of the shared counts.
1640 if( block()->count() > 0 )
1641 sum = block()->count();
1642 cnt = sum / FreqCountInvocations;
1643
1644 // Pin probability to sane limits
1645 float prob;
1646 if( !taken )
1647 prob = (0+PROB_MIN) / 2;
1648 else if( !not_taken )
1649 prob = (1+PROB_MAX) / 2;
1650 else { // Compute probability of true path
1651 prob = (float)taken / (float)(taken + not_taken);
1652 if (prob > PROB_MAX) prob = PROB_MAX;
1653 if (prob < PROB_MIN) prob = PROB_MIN;
1654 }
1655
1656 assert((cnt > 0.0f) && (prob > 0.0f),
1657 "Bad frequency assignment in if cnt=%g prob=%g taken=%d not_taken=%d", cnt, prob, taken, not_taken);
1658
1659 if (C->log() != nullptr) {
1660 const char* prob_str = nullptr;
1661 if (prob >= PROB_MAX) prob_str = (prob == PROB_MAX) ? "max" : "always";
1662 if (prob <= PROB_MIN) prob_str = (prob == PROB_MIN) ? "min" : "never";
1663 char prob_str_buf[30];
1664 if (prob_str == nullptr) {
1665 jio_snprintf(prob_str_buf, sizeof(prob_str_buf), "%20.2f", prob);
1666 prob_str = prob_str_buf;
1667 }
1668 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d' cnt='%f' prob='%s'",
1669 iter().get_dest(), taken, not_taken, cnt, prob_str);
1670 }
1671 return prob;
1672 }
1673
1674 //-----------------------------branch_prediction-------------------------------
1675 float Parse::branch_prediction(float& cnt,
1676 BoolTest::mask btest,
1677 int target_bci,
1678 Node* test) {
1679 float prob = dynamic_branch_prediction(cnt, btest, test);
1680 // If prob is unknown, switch to static prediction
1681 if (prob != PROB_UNKNOWN) return prob;
1682
1683 prob = PROB_FAIR; // Set default value
1684 if (btest == BoolTest::eq) // Exactly equal test?
1685 prob = PROB_STATIC_INFREQUENT; // Assume its relatively infrequent
1686 else if (btest == BoolTest::ne)
1687 prob = PROB_STATIC_FREQUENT; // Assume its relatively frequent
1688
1689 // If this is a conditional test guarding a backwards branch,
1690 // assume its a loop-back edge. Make it a likely taken branch.
1691 if (target_bci < bci()) {
1692 if (is_osr_parse()) { // Could be a hot OSR'd loop; force deopt
1693 // Since it's an OSR, we probably have profile data, but since
1694 // branch_prediction returned PROB_UNKNOWN, the counts are too small.
1695 // Let's make a special check here for completely zero counts.
1696 ciMethodData* methodData = method()->method_data();
1697 if (!methodData->is_empty()) {
1698 ciProfileData* data = methodData->bci_to_data(bci());
1699 // Only stop for truly zero counts, which mean an unknown part
1700 // of the OSR-ed method, and we want to deopt to gather more stats.
1701 // If you have ANY counts, then this loop is simply 'cold' relative
1702 // to the OSR loop.
1703 if (data == nullptr ||
1704 (data->as_BranchData()->taken() + data->as_BranchData()->not_taken() == 0)) {
1705 // This is the only way to return PROB_UNKNOWN:
1706 return PROB_UNKNOWN;
1707 }
1708 }
1709 }
1710 prob = PROB_STATIC_FREQUENT; // Likely to take backwards branch
1711 }
1712
1713 assert(prob != PROB_UNKNOWN, "must have some guess at this point");
1714 return prob;
1715 }
1716
1717 // The magic constants are chosen so as to match the output of
1718 // branch_prediction() when the profile reports a zero taken count.
1719 // It is important to distinguish zero counts unambiguously, because
1720 // some branches (e.g., _213_javac.Assembler.eliminate) validly produce
1721 // very small but nonzero probabilities, which if confused with zero
1722 // counts would keep the program recompiling indefinitely.
1723 bool Parse::seems_never_taken(float prob) const {
1724 return prob < PROB_MIN;
1725 }
1726
1727 //-------------------------------repush_if_args--------------------------------
1728 // Push arguments of an "if" bytecode back onto the stack by adjusting _sp.
1729 inline int Parse::repush_if_args() {
1730 if (PrintOpto && WizardMode) {
1731 tty->print("defending against excessive implicit null exceptions on %s @%d in ",
1732 Bytecodes::name(iter().cur_bc()), iter().cur_bci());
1733 method()->print_name(); tty->cr();
1734 }
1735 int bc_depth = - Bytecodes::depth(iter().cur_bc());
1736 assert(bc_depth == 1 || bc_depth == 2, "only two kinds of branches");
1737 DEBUG_ONLY(sync_jvms()); // argument(n) requires a synced jvms
1738 assert(argument(0) != nullptr, "must exist");
1739 assert(bc_depth == 1 || argument(1) != nullptr, "two must exist");
1740 inc_sp(bc_depth);
1741 return bc_depth;
1742 }
1743
1744 // Used by StressUnstableIfTraps
1745 static volatile int _trap_stress_counter = 0;
1746
1747 void Parse::increment_trap_stress_counter(Node*& counter, Node*& incr_store) {
1748 Node* counter_addr = makecon(TypeRawPtr::make((address)&_trap_stress_counter));
1749 counter = make_load(control(), counter_addr, TypeInt::INT, T_INT, MemNode::unordered);
1750 counter = _gvn.transform(new AddINode(counter, intcon(1)));
1751 incr_store = store_to_memory(control(), counter_addr, counter, T_INT, MemNode::unordered);
1752 }
1753
1754 //----------------------------------do_ifnull----------------------------------
1755 void Parse::do_ifnull(BoolTest::mask btest, Node *c) {
1756 int target_bci = iter().get_dest();
1757
1758 Node* counter = nullptr;
1759 Node* incr_store = nullptr;
1760 bool do_stress_trap = StressUnstableIfTraps && ((C->random() % 2) == 0);
1761 if (do_stress_trap) {
1762 increment_trap_stress_counter(counter, incr_store);
1763 }
1764
1765 Block* branch_block = successor_for_bci(target_bci);
1766 Block* next_block = successor_for_bci(iter().next_bci());
1767
1768 float cnt;
1769 float prob = branch_prediction(cnt, btest, target_bci, c);
1770 if (prob == PROB_UNKNOWN) {
1771 // (An earlier version of do_ifnull omitted this trap for OSR methods.)
1772 if (PrintOpto && Verbose) {
1773 tty->print_cr("Never-taken edge stops compilation at bci %d", bci());
1774 }
1775 repush_if_args(); // to gather stats on loop
1776 uncommon_trap(Deoptimization::Reason_unreached,
1777 Deoptimization::Action_reinterpret,
1778 nullptr, "cold");
1779 if (C->eliminate_boxing()) {
1780 // Mark the successor blocks as parsed
1781 branch_block->next_path_num();
1782 next_block->next_path_num();
1783 }
1784 return;
1785 }
1786
1787 NOT_PRODUCT(explicit_null_checks_inserted++);
1788
1789 // Generate real control flow
1790 Node *tst = _gvn.transform( new BoolNode( c, btest ) );
1791
1792 // Sanity check the probability value
1793 assert(prob > 0.0f,"Bad probability in Parser");
1794 // Need xform to put node in hash table
1795 IfNode *iff = create_and_xform_if( control(), tst, prob, cnt );
1796 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser");
1797 // True branch
1798 { PreserveJVMState pjvms(this);
1799 Node* iftrue = _gvn.transform( new IfTrueNode (iff) );
1800 set_control(iftrue);
1801
1802 if (stopped()) { // Path is dead?
1803 NOT_PRODUCT(explicit_null_checks_elided++);
1804 if (C->eliminate_boxing()) {
1805 // Mark the successor block as parsed
1806 branch_block->next_path_num();
1807 }
1808 } else { // Path is live.
1809 adjust_map_after_if(btest, c, prob, branch_block);
1810 if (!stopped()) {
1811 merge(target_bci);
1812 }
1813 }
1814 }
1815
1816 // False branch
1817 Node* iffalse = _gvn.transform( new IfFalseNode(iff) );
1818 set_control(iffalse);
1819
1820 if (stopped()) { // Path is dead?
1821 NOT_PRODUCT(explicit_null_checks_elided++);
1822 if (C->eliminate_boxing()) {
1823 // Mark the successor block as parsed
1824 next_block->next_path_num();
1825 }
1826 } else { // Path is live.
1827 adjust_map_after_if(BoolTest(btest).negate(), c, 1.0-prob, next_block);
1828 }
1829
1830 if (do_stress_trap) {
1831 stress_trap(iff, counter, incr_store);
1832 }
1833 }
1834
1835 //------------------------------------do_if------------------------------------
1836 void Parse::do_if(BoolTest::mask btest, Node* c, bool can_trap, bool new_path, Node** ctrl_taken) {
1837 int target_bci = iter().get_dest();
1838
1839 Block* branch_block = successor_for_bci(target_bci);
1840 Block* next_block = successor_for_bci(iter().next_bci());
1841
1842 float cnt;
1843 float prob = branch_prediction(cnt, btest, target_bci, c);
1844 float untaken_prob = 1.0 - prob;
1845
1846 if (prob == PROB_UNKNOWN) {
1847 if (PrintOpto && Verbose) {
1848 tty->print_cr("Never-taken edge stops compilation at bci %d", bci());
1849 }
1850 repush_if_args(); // to gather stats on loop
1851 uncommon_trap(Deoptimization::Reason_unreached,
1852 Deoptimization::Action_reinterpret,
1853 nullptr, "cold");
1854 if (C->eliminate_boxing()) {
1855 // Mark the successor blocks as parsed
1856 branch_block->next_path_num();
1857 next_block->next_path_num();
1858 }
1859 return;
1860 }
1861
1862 Node* counter = nullptr;
1863 Node* incr_store = nullptr;
1864 bool do_stress_trap = StressUnstableIfTraps && ((C->random() % 2) == 0);
1865 if (do_stress_trap) {
1866 increment_trap_stress_counter(counter, incr_store);
1867 }
1868
1869 // Sanity check the probability value
1870 assert(0.0f < prob && prob < 1.0f,"Bad probability in Parser");
1871
1872 bool taken_if_true = true;
1873 // Convert BoolTest to canonical form:
1874 if (!BoolTest(btest).is_canonical()) {
1875 btest = BoolTest(btest).negate();
1876 taken_if_true = false;
1877 // prob is NOT updated here; it remains the probability of the taken
1878 // path (as opposed to the prob of the path guarded by an 'IfTrueNode').
1879 }
1880 assert(btest != BoolTest::eq, "!= is the only canonical exact test");
1881
1882 Node* tst0 = new BoolNode(c, btest);
1883 Node* tst = _gvn.transform(tst0);
1884 BoolTest::mask taken_btest = BoolTest::illegal;
1885 BoolTest::mask untaken_btest = BoolTest::illegal;
1886
1887 if (tst->is_Bool()) {
1888 // Refresh c from the transformed bool node, since it may be
1889 // simpler than the original c. Also re-canonicalize btest.
1890 // This wins when (Bool ne (Conv2B p) 0) => (Bool ne (CmpP p null)).
1891 // That can arise from statements like: if (x instanceof C) ...
1892 if (tst != tst0) {
1893 // Canonicalize one more time since transform can change it.
1894 btest = tst->as_Bool()->_test._test;
1895 if (!BoolTest(btest).is_canonical()) {
1896 // Reverse edges one more time...
1897 tst = _gvn.transform( tst->as_Bool()->negate(&_gvn) );
1898 btest = tst->as_Bool()->_test._test;
1899 assert(BoolTest(btest).is_canonical(), "sanity");
1900 taken_if_true = !taken_if_true;
1901 }
1902 c = tst->in(1);
1903 }
1904 BoolTest::mask neg_btest = BoolTest(btest).negate();
1905 taken_btest = taken_if_true ? btest : neg_btest;
1906 untaken_btest = taken_if_true ? neg_btest : btest;
1907 }
1908
1909 // Generate real control flow
1910 float true_prob = (taken_if_true ? prob : untaken_prob);
1911 IfNode* iff = create_and_map_if(control(), tst, true_prob, cnt);
1912 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser");
1913 Node* taken_branch = new IfTrueNode(iff);
1914 Node* untaken_branch = new IfFalseNode(iff);
1915 if (!taken_if_true) { // Finish conversion to canonical form
1916 Node* tmp = taken_branch;
1917 taken_branch = untaken_branch;
1918 untaken_branch = tmp;
1919 }
1920
1921 // Branch is taken:
1922 { PreserveJVMState pjvms(this);
1923 taken_branch = _gvn.transform(taken_branch);
1924 set_control(taken_branch);
1925
1926 if (stopped()) {
1927 if (C->eliminate_boxing() && !new_path) {
1928 // Mark the successor block as parsed (if we haven't created a new path)
1929 branch_block->next_path_num();
1930 }
1931 } else {
1932 adjust_map_after_if(taken_btest, c, prob, branch_block, can_trap);
1933 if (!stopped()) {
1934 if (new_path) {
1935 // Merge by using a new path
1936 merge_new_path(target_bci);
1937 } else if (ctrl_taken != nullptr) {
1938 // Don't merge but save taken branch to be wired by caller
1939 *ctrl_taken = control();
1940 } else {
1941 merge(target_bci);
1942 }
1943 }
1944 }
1945 }
1946
1947 untaken_branch = _gvn.transform(untaken_branch);
1948 set_control(untaken_branch);
1949
1950 // Branch not taken.
1951 if (stopped() && ctrl_taken == nullptr) {
1952 if (C->eliminate_boxing()) {
1953 // Mark the successor block as parsed (if caller does not re-wire control flow)
1954 next_block->next_path_num();
1955 }
1956 } else {
1957 adjust_map_after_if(untaken_btest, c, untaken_prob, next_block, can_trap);
1958 }
1959
1960 if (do_stress_trap) {
1961 stress_trap(iff, counter, incr_store);
1962 }
1963 }
1964
1965
1966 static ProfilePtrKind speculative_ptr_kind(const TypeOopPtr* t) {
1967 if (t->speculative() == nullptr) {
1968 return ProfileUnknownNull;
1969 }
1970 if (t->speculative_always_null()) {
1971 return ProfileAlwaysNull;
1972 }
1973 if (t->speculative_maybe_null()) {
1974 return ProfileMaybeNull;
1975 }
1976 return ProfileNeverNull;
1977 }
1978
1979 void Parse::acmp_always_null_input(Node* input, const TypeOopPtr* tinput, BoolTest::mask btest, Node* eq_region) {
1980 inc_sp(2);
1981 Node* cast = null_check_common(input, T_OBJECT, true, nullptr,
1982 !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_null_check) &&
1983 speculative_ptr_kind(tinput) == ProfileAlwaysNull);
1984 dec_sp(2);
1985 if (btest == BoolTest::ne) {
1986 {
1987 PreserveJVMState pjvms(this);
1988 replace_in_map(input, cast);
1989 int target_bci = iter().get_dest();
1990 merge(target_bci);
1991 }
1992 record_for_igvn(eq_region);
1993 set_control(_gvn.transform(eq_region));
1994 } else {
1995 replace_in_map(input, cast);
1996 }
1997 }
1998
1999 Node* Parse::acmp_null_check(Node* input, const TypeOopPtr* tinput, ProfilePtrKind input_ptr, Node*& null_ctl) {
2000 inc_sp(2);
2001 null_ctl = top();
2002 Node* cast = null_check_oop(input, &null_ctl,
2003 input_ptr == ProfileNeverNull || (input_ptr == ProfileUnknownNull && !too_many_traps_or_recompiles(Deoptimization::Reason_null_check)),
2004 false,
2005 speculative_ptr_kind(tinput) == ProfileNeverNull &&
2006 !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_null_check));
2007 dec_sp(2);
2008 assert(!stopped(), "null input should have been caught earlier");
2009 return cast;
2010 }
2011
2012 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) {
2013 Node* ne_region = new RegionNode(1);
2014 Node* null_ctl;
2015 Node* cast = acmp_null_check(input, tinput, input_ptr, null_ctl);
2016 ne_region->add_req(null_ctl);
2017
2018 Node* slow_ctl = type_check_receiver(cast, input_type, 1.0, &cast);
2019 {
2020 PreserveJVMState pjvms(this);
2021 inc_sp(2);
2022 set_control(slow_ctl);
2023 Deoptimization::DeoptReason reason;
2024 if (tinput->speculative_type() != nullptr && !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_class_check)) {
2025 reason = Deoptimization::Reason_speculate_class_check;
2026 } else {
2027 reason = Deoptimization::Reason_class_check;
2028 }
2029 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
2030 }
2031 ne_region->add_req(control());
2032
2033 record_for_igvn(ne_region);
2034 set_control(_gvn.transform(ne_region));
2035 if (btest == BoolTest::ne) {
2036 {
2037 PreserveJVMState pjvms(this);
2038 if (null_ctl == top()) {
2039 replace_in_map(input, cast);
2040 }
2041 int target_bci = iter().get_dest();
2042 merge(target_bci);
2043 }
2044 record_for_igvn(eq_region);
2045 set_control(_gvn.transform(eq_region));
2046 } else {
2047 if (null_ctl == top()) {
2048 replace_in_map(input, cast);
2049 }
2050 set_control(_gvn.transform(ne_region));
2051 }
2052 }
2053
2054 void Parse::acmp_unknown_non_inline_type_input(Node* input, const TypeOopPtr* tinput, ProfilePtrKind input_ptr, BoolTest::mask btest, Node* eq_region) {
2055 Node* ne_region = new RegionNode(1);
2056 Node* null_ctl;
2057 Node* cast = acmp_null_check(input, tinput, input_ptr, null_ctl);
2058 ne_region->add_req(null_ctl);
2059
2060 {
2061 BuildCutout unless(this, inline_type_test(cast, /* is_inline = */ false), PROB_MAX);
2062 inc_sp(2);
2063 uncommon_trap_exact(Deoptimization::Reason_class_check, Deoptimization::Action_maybe_recompile);
2064 }
2065
2066 ne_region->add_req(control());
2067
2068 record_for_igvn(ne_region);
2069 set_control(_gvn.transform(ne_region));
2070 if (btest == BoolTest::ne) {
2071 {
2072 PreserveJVMState pjvms(this);
2073 if (null_ctl == top()) {
2074 replace_in_map(input, cast);
2075 }
2076 int target_bci = iter().get_dest();
2077 merge(target_bci);
2078 }
2079 record_for_igvn(eq_region);
2080 set_control(_gvn.transform(eq_region));
2081 } else {
2082 if (null_ctl == top()) {
2083 replace_in_map(input, cast);
2084 }
2085 set_control(_gvn.transform(ne_region));
2086 }
2087 }
2088
2089 void Parse::do_acmp(BoolTest::mask btest, Node* left, Node* right) {
2090 ciKlass* left_type = nullptr;
2091 ciKlass* right_type = nullptr;
2092 ProfilePtrKind left_ptr = ProfileUnknownNull;
2093 ProfilePtrKind right_ptr = ProfileUnknownNull;
2094 bool left_inline_type = true;
2095 bool right_inline_type = true;
2096
2097 // Leverage profiling at acmp
2098 if (UseACmpProfile) {
2099 method()->acmp_profiled_type(bci(), left_type, right_type, left_ptr, right_ptr, left_inline_type, right_inline_type);
2100 if (too_many_traps_or_recompiles(Deoptimization::Reason_class_check)) {
2101 left_type = nullptr;
2102 right_type = nullptr;
2103 left_inline_type = true;
2104 right_inline_type = true;
2105 }
2106 if (too_many_traps_or_recompiles(Deoptimization::Reason_null_check)) {
2107 left_ptr = ProfileUnknownNull;
2108 right_ptr = ProfileUnknownNull;
2109 }
2110 }
2111
2112 if (UseTypeSpeculation) {
2113 record_profile_for_speculation(left, left_type, left_ptr);
2114 record_profile_for_speculation(right, right_type, right_ptr);
2115 }
2116
2117 if (!EnableValhalla) {
2118 Node* cmp = CmpP(left, right);
2119 cmp = optimize_cmp_with_klass(cmp);
2120 do_if(btest, cmp);
2121 return;
2122 }
2123
2124 // Check for equality before potentially allocating
2125 if (left == right) {
2126 do_if(btest, makecon(TypeInt::CC_EQ));
2127 return;
2128 }
2129
2130 // Allocate inline type operands and re-execute on deoptimization
2131 if (left->is_InlineType()) {
2132 if (_gvn.type(right)->is_zero_type() ||
2133 (right->is_InlineType() && _gvn.type(right->as_InlineType()->get_is_init())->is_zero_type())) {
2134 // Null checking a scalarized but nullable inline type. Check the IsInit
2135 // input instead of the oop input to avoid keeping buffer allocations alive.
2136 Node* cmp = CmpI(left->as_InlineType()->get_is_init(), intcon(0));
2137 do_if(btest, cmp);
2138 return;
2139 } else {
2140 PreserveReexecuteState preexecs(this);
2141 inc_sp(2);
2142 jvms()->set_should_reexecute(true);
2143 left = left->as_InlineType()->buffer(this)->get_oop();
2144 }
2145 }
2146 if (right->is_InlineType()) {
2147 PreserveReexecuteState preexecs(this);
2148 inc_sp(2);
2149 jvms()->set_should_reexecute(true);
2150 right = right->as_InlineType()->buffer(this)->get_oop();
2151 }
2152
2153 // First, do a normal pointer comparison
2154 const TypeOopPtr* tleft = _gvn.type(left)->isa_oopptr();
2155 const TypeOopPtr* tright = _gvn.type(right)->isa_oopptr();
2156 Node* cmp = CmpP(left, right);
2157 cmp = optimize_cmp_with_klass(cmp);
2158 if (tleft == nullptr || !tleft->can_be_inline_type() ||
2159 tright == nullptr || !tright->can_be_inline_type()) {
2160 // This is sufficient, if one of the operands can't be an inline type
2161 do_if(btest, cmp);
2162 return;
2163 }
2164
2165 // Don't add traps to unstable if branches because additional checks are required to
2166 // decide if the operands are equal/substitutable and we therefore shouldn't prune
2167 // branches for one if based on the profiling of the acmp branches.
2168 // Also, OptimizeUnstableIf would set an incorrect re-rexecution state because it
2169 // assumes that there is a 1-1 mapping between the if and the acmp branches and that
2170 // hitting a trap means that we will take the corresponding acmp branch on re-execution.
2171 const bool can_trap = true;
2172
2173 Node* eq_region = nullptr;
2174 if (btest == BoolTest::eq) {
2175 do_if(btest, cmp, !can_trap, true);
2176 if (stopped()) {
2177 // Pointers are equal, operands must be equal
2178 return;
2179 }
2180 } else {
2181 assert(btest == BoolTest::ne, "only eq or ne");
2182 Node* is_not_equal = nullptr;
2183 eq_region = new RegionNode(3);
2184 {
2185 PreserveJVMState pjvms(this);
2186 // Pointers are not equal, but more checks are needed to determine if the operands are (not) substitutable
2187 do_if(btest, cmp, !can_trap, false, &is_not_equal);
2188 if (!stopped()) {
2189 eq_region->init_req(1, control());
2190 }
2191 }
2192 if (is_not_equal == nullptr || is_not_equal->is_top()) {
2193 record_for_igvn(eq_region);
2194 set_control(_gvn.transform(eq_region));
2195 return;
2196 }
2197 set_control(is_not_equal);
2198 }
2199
2200 // Prefer speculative types if available
2201 if (!too_many_traps_or_recompiles(Deoptimization::Reason_speculate_class_check)) {
2202 if (tleft->speculative_type() != nullptr) {
2203 left_type = tleft->speculative_type();
2204 }
2205 if (tright->speculative_type() != nullptr) {
2206 right_type = tright->speculative_type();
2207 }
2208 }
2209
2210 if (speculative_ptr_kind(tleft) != ProfileMaybeNull && speculative_ptr_kind(tleft) != ProfileUnknownNull) {
2211 ProfilePtrKind speculative_left_ptr = speculative_ptr_kind(tleft);
2212 if (speculative_left_ptr == ProfileAlwaysNull && !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_null_assert)) {
2213 left_ptr = speculative_left_ptr;
2214 } else if (speculative_left_ptr == ProfileNeverNull && !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_null_check)) {
2215 left_ptr = speculative_left_ptr;
2216 }
2217 }
2218 if (speculative_ptr_kind(tright) != ProfileMaybeNull && speculative_ptr_kind(tright) != ProfileUnknownNull) {
2219 ProfilePtrKind speculative_right_ptr = speculative_ptr_kind(tright);
2220 if (speculative_right_ptr == ProfileAlwaysNull && !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_null_assert)) {
2221 right_ptr = speculative_right_ptr;
2222 } else if (speculative_right_ptr == ProfileNeverNull && !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_null_check)) {
2223 right_ptr = speculative_right_ptr;
2224 }
2225 }
2226
2227 if (left_ptr == ProfileAlwaysNull) {
2228 // Comparison with null. Assert the input is indeed null and we're done.
2229 acmp_always_null_input(left, tleft, btest, eq_region);
2230 return;
2231 }
2232 if (right_ptr == ProfileAlwaysNull) {
2233 // Comparison with null. Assert the input is indeed null and we're done.
2234 acmp_always_null_input(right, tright, btest, eq_region);
2235 return;
2236 }
2237 if (left_type != nullptr && !left_type->is_inlinetype()) {
2238 // Comparison with an object of known type
2239 acmp_known_non_inline_type_input(left, tleft, left_ptr, left_type, btest, eq_region);
2240 return;
2241 }
2242 if (right_type != nullptr && !right_type->is_inlinetype()) {
2243 // Comparison with an object of known type
2244 acmp_known_non_inline_type_input(right, tright, right_ptr, right_type, btest, eq_region);
2245 return;
2246 }
2247 if (!left_inline_type) {
2248 // Comparison with an object known not to be an inline type
2249 acmp_unknown_non_inline_type_input(left, tleft, left_ptr, btest, eq_region);
2250 return;
2251 }
2252 if (!right_inline_type) {
2253 // Comparison with an object known not to be an inline type
2254 acmp_unknown_non_inline_type_input(right, tright, right_ptr, btest, eq_region);
2255 return;
2256 }
2257
2258 // Pointers are not equal, check if first operand is non-null
2259 Node* ne_region = new RegionNode(6);
2260 Node* null_ctl;
2261 Node* not_null_right = acmp_null_check(right, tright, right_ptr, null_ctl);
2262 ne_region->init_req(1, null_ctl);
2263
2264 // First operand is non-null, check if it is an inline type
2265 Node* is_value = inline_type_test(not_null_right);
2266 IfNode* is_value_iff = create_and_map_if(control(), is_value, PROB_FAIR, COUNT_UNKNOWN);
2267 Node* not_value = _gvn.transform(new IfFalseNode(is_value_iff));
2268 ne_region->init_req(2, not_value);
2269 set_control(_gvn.transform(new IfTrueNode(is_value_iff)));
2270
2271 // The first operand is an inline type, check if the second operand is non-null
2272 Node* not_null_left = acmp_null_check(left, tleft, left_ptr, null_ctl);
2273 ne_region->init_req(3, null_ctl);
2274
2275 // Check if both operands are of the same class.
2276 Node* kls_left = load_object_klass(not_null_left);
2277 Node* kls_right = load_object_klass(not_null_right);
2278 Node* kls_cmp = CmpP(kls_left, kls_right);
2279 Node* kls_bol = _gvn.transform(new BoolNode(kls_cmp, BoolTest::ne));
2280 IfNode* kls_iff = create_and_map_if(control(), kls_bol, PROB_FAIR, COUNT_UNKNOWN);
2281 Node* kls_ne = _gvn.transform(new IfTrueNode(kls_iff));
2282 set_control(_gvn.transform(new IfFalseNode(kls_iff)));
2283 ne_region->init_req(4, kls_ne);
2284
2285 if (stopped()) {
2286 record_for_igvn(ne_region);
2287 set_control(_gvn.transform(ne_region));
2288 if (btest == BoolTest::ne) {
2289 {
2290 PreserveJVMState pjvms(this);
2291 int target_bci = iter().get_dest();
2292 merge(target_bci);
2293 }
2294 record_for_igvn(eq_region);
2295 set_control(_gvn.transform(eq_region));
2296 }
2297 return;
2298 }
2299
2300 // Both operands are values types of the same class, we need to perform a
2301 // substitutability test. Delegate to ValueObjectMethods::isSubstitutable().
2302 Node* ne_io_phi = PhiNode::make(ne_region, i_o());
2303 Node* mem = reset_memory();
2304 Node* ne_mem_phi = PhiNode::make(ne_region, mem);
2305
2306 Node* eq_io_phi = nullptr;
2307 Node* eq_mem_phi = nullptr;
2308 if (eq_region != nullptr) {
2309 eq_io_phi = PhiNode::make(eq_region, i_o());
2310 eq_mem_phi = PhiNode::make(eq_region, mem);
2311 }
2312
2313 set_all_memory(mem);
2314
2315 kill_dead_locals();
2316 ciMethod* subst_method = ciEnv::current()->ValueObjectMethods_klass()->find_method(ciSymbols::isSubstitutable_name(), ciSymbols::object_object_boolean_signature());
2317 CallStaticJavaNode *call = new CallStaticJavaNode(C, TypeFunc::make(subst_method), SharedRuntime::get_resolve_static_call_stub(), subst_method);
2318 call->set_override_symbolic_info(true);
2319 call->init_req(TypeFunc::Parms, not_null_left);
2320 call->init_req(TypeFunc::Parms+1, not_null_right);
2321 inc_sp(2);
2322 set_edges_for_java_call(call, false, false);
2323 Node* ret = set_results_for_java_call(call, false, true);
2324 dec_sp(2);
2325
2326 // Test the return value of ValueObjectMethods::isSubstitutable()
2327 // This is the last check, do_if can emit traps now.
2328 Node* subst_cmp = _gvn.transform(new CmpINode(ret, intcon(1)));
2329 Node* ctl = C->top();
2330 if (btest == BoolTest::eq) {
2331 PreserveJVMState pjvms(this);
2332 do_if(btest, subst_cmp, can_trap);
2333 if (!stopped()) {
2334 ctl = control();
2335 }
2336 } else {
2337 assert(btest == BoolTest::ne, "only eq or ne");
2338 PreserveJVMState pjvms(this);
2339 do_if(btest, subst_cmp, can_trap, false, &ctl);
2340 if (!stopped()) {
2341 eq_region->init_req(2, control());
2342 eq_io_phi->init_req(2, i_o());
2343 eq_mem_phi->init_req(2, reset_memory());
2344 }
2345 }
2346 ne_region->init_req(5, ctl);
2347 ne_io_phi->init_req(5, i_o());
2348 ne_mem_phi->init_req(5, reset_memory());
2349
2350 record_for_igvn(ne_region);
2351 set_control(_gvn.transform(ne_region));
2352 set_i_o(_gvn.transform(ne_io_phi));
2353 set_all_memory(_gvn.transform(ne_mem_phi));
2354
2355 if (btest == BoolTest::ne) {
2356 {
2357 PreserveJVMState pjvms(this);
2358 int target_bci = iter().get_dest();
2359 merge(target_bci);
2360 }
2361
2362 record_for_igvn(eq_region);
2363 set_control(_gvn.transform(eq_region));
2364 set_i_o(_gvn.transform(eq_io_phi));
2365 set_all_memory(_gvn.transform(eq_mem_phi));
2366 }
2367 }
2368
2369 // Force unstable if traps to be taken randomly to trigger intermittent bugs such as incorrect debug information.
2370 // Add another if before the unstable if that checks a "random" condition at runtime (a simple shared counter) and
2371 // then either takes the trap or executes the original, unstable if.
2372 void Parse::stress_trap(IfNode* orig_iff, Node* counter, Node* incr_store) {
2373 // Search for an unstable if trap
2374 CallStaticJavaNode* trap = nullptr;
2375 assert(orig_iff->Opcode() == Op_If && orig_iff->outcnt() == 2, "malformed if");
2376 ProjNode* trap_proj = orig_iff->uncommon_trap_proj(trap, Deoptimization::Reason_unstable_if);
2377 if (trap == nullptr || !trap->jvms()->should_reexecute()) {
2378 // No suitable trap found. Remove unused counter load and increment.
2379 C->gvn_replace_by(incr_store, incr_store->in(MemNode::Memory));
2380 return;
2381 }
2382
2383 // Remove trap from optimization list since we add another path to the trap.
2384 bool success = C->remove_unstable_if_trap(trap, true);
2385 assert(success, "Trap already modified");
2386
2387 // Add a check before the original if that will trap with a certain frequency and execute the original if otherwise
2388 int freq_log = (C->random() % 31) + 1; // Random logarithmic frequency in [1, 31]
2389 Node* mask = intcon(right_n_bits(freq_log));
2390 counter = _gvn.transform(new AndINode(counter, mask));
2391 Node* cmp = _gvn.transform(new CmpINode(counter, intcon(0)));
2392 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::mask::eq));
2393 IfNode* iff = _gvn.transform(new IfNode(orig_iff->in(0), bol, orig_iff->_prob, orig_iff->_fcnt))->as_If();
2394 Node* if_true = _gvn.transform(new IfTrueNode(iff));
2395 Node* if_false = _gvn.transform(new IfFalseNode(iff));
2396 assert(!if_true->is_top() && !if_false->is_top(), "trap always / never taken");
2397
2398 // Trap
2399 assert(trap_proj->outcnt() == 1, "some other nodes are dependent on the trap projection");
2400
2401 Node* trap_region = new RegionNode(3);
2402 trap_region->set_req(1, trap_proj);
2403 trap_region->set_req(2, if_true);
2404 trap->set_req(0, _gvn.transform(trap_region));
2405
2406 // Don't trap, execute original if
2407 orig_iff->set_req(0, if_false);
2408 }
2409
2410 bool Parse::path_is_suitable_for_uncommon_trap(float prob) const {
2411 // Randomly skip emitting an uncommon trap
2412 if (StressUnstableIfTraps && ((C->random() % 2) == 0)) {
2413 return false;
2414 }
2415 // Don't want to speculate on uncommon traps when running with -Xcomp
2416 if (!UseInterpreter) {
2417 return false;
2418 }
2419 return seems_never_taken(prob) &&
2420 !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if);
2421 }
2422
2423 void Parse::maybe_add_predicate_after_if(Block* path) {
2424 if (path->is_SEL_head() && path->preds_parsed() == 0) {
2425 // Add predicates at bci of if dominating the loop so traps can be
2426 // recorded on the if's profile data
2427 int bc_depth = repush_if_args();
2428 add_parse_predicates();
2429 dec_sp(bc_depth);
2430 path->set_has_predicates();
2431 }
2432 }
2433
2434
2435 //----------------------------adjust_map_after_if------------------------------
2436 // Adjust the JVM state to reflect the result of taking this path.
2437 // Basically, it means inspecting the CmpNode controlling this
2438 // branch, seeing how it constrains a tested value, and then
2439 // deciding if it's worth our while to encode this constraint
2440 // as graph nodes in the current abstract interpretation map.
2441 void Parse::adjust_map_after_if(BoolTest::mask btest, Node* c, float prob, Block* path, bool can_trap) {
2442 if (!c->is_Cmp()) {
2443 maybe_add_predicate_after_if(path);
2444 return;
2445 }
2446
2447 if (stopped() || btest == BoolTest::illegal) {
2448 return; // nothing to do
2449 }
2450
2451 bool is_fallthrough = (path == successor_for_bci(iter().next_bci()));
2452
2453 if (can_trap && path_is_suitable_for_uncommon_trap(prob)) {
2454 repush_if_args();
2455 Node* call = uncommon_trap(Deoptimization::Reason_unstable_if,
2456 Deoptimization::Action_reinterpret,
2457 nullptr,
2458 (is_fallthrough ? "taken always" : "taken never"));
2459
2460 if (call != nullptr) {
2461 C->record_unstable_if_trap(new UnstableIfTrap(call->as_CallStaticJava(), path));
2462 }
2463 return;
2464 }
2465
2466 Node* val = c->in(1);
2467 Node* con = c->in(2);
2468 const Type* tcon = _gvn.type(con);
2469 const Type* tval = _gvn.type(val);
2470 bool have_con = tcon->singleton();
2471 if (tval->singleton()) {
2472 if (!have_con) {
2473 // Swap, so constant is in con.
2474 con = val;
2475 tcon = tval;
2476 val = c->in(2);
2477 tval = _gvn.type(val);
2478 btest = BoolTest(btest).commute();
2479 have_con = true;
2480 } else {
2481 // Do we have two constants? Then leave well enough alone.
2482 have_con = false;
2483 }
2484 }
2485 if (!have_con) { // remaining adjustments need a con
2486 maybe_add_predicate_after_if(path);
2487 return;
2488 }
2489
2490 sharpen_type_after_if(btest, con, tcon, val, tval);
2491 maybe_add_predicate_after_if(path);
2492 }
2493
2494
2495 static Node* extract_obj_from_klass_load(PhaseGVN* gvn, Node* n) {
2496 Node* ldk;
2497 if (n->is_DecodeNKlass()) {
2498 if (n->in(1)->Opcode() != Op_LoadNKlass) {
2499 return nullptr;
2500 } else {
2501 ldk = n->in(1);
2502 }
2503 } else if (n->Opcode() != Op_LoadKlass) {
2504 return nullptr;
2505 } else {
2506 ldk = n;
2507 }
2508 assert(ldk != nullptr && ldk->is_Load(), "should have found a LoadKlass or LoadNKlass node");
2509
2510 Node* adr = ldk->in(MemNode::Address);
2511 intptr_t off = 0;
2512 Node* obj = AddPNode::Ideal_base_and_offset(adr, gvn, off);
2513 if (obj == nullptr || off != oopDesc::klass_offset_in_bytes()) // loading oopDesc::_klass?
2514 return nullptr;
2515 const TypePtr* tp = gvn->type(obj)->is_ptr();
2516 if (tp == nullptr || !(tp->isa_instptr() || tp->isa_aryptr())) // is obj a Java object ptr?
2517 return nullptr;
2518
2519 return obj;
2520 }
2521
2522 void Parse::sharpen_type_after_if(BoolTest::mask btest,
2523 Node* con, const Type* tcon,
2524 Node* val, const Type* tval) {
2525 // Look for opportunities to sharpen the type of a node
2526 // whose klass is compared with a constant klass.
2527 if (btest == BoolTest::eq && tcon->isa_klassptr()) {
2528 Node* obj = extract_obj_from_klass_load(&_gvn, val);
2529 const TypeOopPtr* con_type = tcon->isa_klassptr()->as_instance_type();
2530 if (obj != nullptr && (con_type->isa_instptr() || con_type->isa_aryptr())) {
2531 // Found:
2532 // Bool(CmpP(LoadKlass(obj._klass), ConP(Foo.klass)), [eq])
2533 // or the narrowOop equivalent.
2534 const Type* obj_type = _gvn.type(obj);
2535 const TypeOopPtr* tboth = obj_type->join_speculative(con_type)->isa_oopptr();
2536 if (tboth != nullptr && tboth->klass_is_exact() && tboth != obj_type &&
2537 tboth->higher_equal(obj_type)) {
2538 // obj has to be of the exact type Foo if the CmpP succeeds.
2539 int obj_in_map = map()->find_edge(obj);
2540 JVMState* jvms = this->jvms();
2541 if (obj_in_map >= 0 &&
2542 (jvms->is_loc(obj_in_map) || jvms->is_stk(obj_in_map))) {
2543 TypeNode* ccast = new CheckCastPPNode(control(), obj, tboth);
2544 const Type* tcc = ccast->as_Type()->type();
2545 assert(tcc != obj_type && tcc->higher_equal(obj_type), "must improve");
2546 // Delay transform() call to allow recovery of pre-cast value
2547 // at the control merge.
2548 _gvn.set_type_bottom(ccast);
2549 record_for_igvn(ccast);
2550 if (tboth->is_inlinetypeptr()) {
2551 ccast = InlineTypeNode::make_from_oop(this, ccast, tboth->exact_klass(true)->as_inline_klass());
2552 }
2553 // Here's the payoff.
2554 replace_in_map(obj, ccast);
2555 }
2556 }
2557 }
2558 }
2559
2560 int val_in_map = map()->find_edge(val);
2561 if (val_in_map < 0) return; // replace_in_map would be useless
2562 {
2563 JVMState* jvms = this->jvms();
2564 if (!(jvms->is_loc(val_in_map) ||
2565 jvms->is_stk(val_in_map)))
2566 return; // again, it would be useless
2567 }
2568
2569 // Check for a comparison to a constant, and "know" that the compared
2570 // value is constrained on this path.
2571 assert(tcon->singleton(), "");
2572 ConstraintCastNode* ccast = nullptr;
2573 Node* cast = nullptr;
2574
2575 switch (btest) {
2576 case BoolTest::eq: // Constant test?
2577 {
2578 const Type* tboth = tcon->join_speculative(tval);
2579 if (tboth == tval) break; // Nothing to gain.
2580 if (tcon->isa_int()) {
2581 ccast = new CastIINode(control(), val, tboth);
2582 } else if (tcon == TypePtr::NULL_PTR) {
2583 // Cast to null, but keep the pointer identity temporarily live.
2584 ccast = new CastPPNode(control(), val, tboth);
2585 } else {
2586 const TypeF* tf = tcon->isa_float_constant();
2587 const TypeD* td = tcon->isa_double_constant();
2588 // Exclude tests vs float/double 0 as these could be
2589 // either +0 or -0. Just because you are equal to +0
2590 // doesn't mean you ARE +0!
2591 // Note, following code also replaces Long and Oop values.
2592 if ((!tf || tf->_f != 0.0) &&
2593 (!td || td->_d != 0.0))
2594 cast = con; // Replace non-constant val by con.
2595 }
2596 }
2597 break;
2598
2599 case BoolTest::ne:
2600 if (tcon == TypePtr::NULL_PTR) {
2601 cast = cast_not_null(val, false);
2602 }
2603 break;
2604
2605 default:
2606 // (At this point we could record int range types with CastII.)
2607 break;
2608 }
2609
2610 if (ccast != nullptr) {
2611 const Type* tcc = ccast->as_Type()->type();
2612 assert(tcc != tval && tcc->higher_equal(tval), "must improve");
2613 // Delay transform() call to allow recovery of pre-cast value
2614 // at the control merge.
2615 _gvn.set_type_bottom(ccast);
2616 record_for_igvn(ccast);
2617 cast = ccast;
2618 }
2619
2620 if (cast != nullptr) { // Here's the payoff.
2621 replace_in_map(val, cast);
2622 }
2623 }
2624
2625 /**
2626 * Use speculative type to optimize CmpP node: if comparison is
2627 * against the low level class, cast the object to the speculative
2628 * type if any. CmpP should then go away.
2629 *
2630 * @param c expected CmpP node
2631 * @return result of CmpP on object casted to speculative type
2632 *
2633 */
2634 Node* Parse::optimize_cmp_with_klass(Node* c) {
2635 // If this is transformed by the _gvn to a comparison with the low
2636 // level klass then we may be able to use speculation
2637 if (c->Opcode() == Op_CmpP &&
2638 (c->in(1)->Opcode() == Op_LoadKlass || c->in(1)->Opcode() == Op_DecodeNKlass) &&
2639 c->in(2)->is_Con()) {
2640 Node* load_klass = nullptr;
2641 Node* decode = nullptr;
2642 if (c->in(1)->Opcode() == Op_DecodeNKlass) {
2643 decode = c->in(1);
2644 load_klass = c->in(1)->in(1);
2645 } else {
2646 load_klass = c->in(1);
2647 }
2648 if (load_klass->in(2)->is_AddP()) {
2649 Node* addp = load_klass->in(2);
2650 Node* obj = addp->in(AddPNode::Address);
2651 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
2652 if (obj_type->speculative_type_not_null() != nullptr) {
2653 ciKlass* k = obj_type->speculative_type();
2654 inc_sp(2);
2655 obj = maybe_cast_profiled_obj(obj, k);
2656 dec_sp(2);
2657 if (obj->is_InlineType()) {
2658 assert(obj->as_InlineType()->is_allocated(&_gvn), "must be allocated");
2659 obj = obj->as_InlineType()->get_oop();
2660 }
2661 // Make the CmpP use the casted obj
2662 addp = basic_plus_adr(obj, addp->in(AddPNode::Offset));
2663 load_klass = load_klass->clone();
2664 load_klass->set_req(2, addp);
2665 load_klass = _gvn.transform(load_klass);
2666 if (decode != nullptr) {
2667 decode = decode->clone();
2668 decode->set_req(1, load_klass);
2669 load_klass = _gvn.transform(decode);
2670 }
2671 c = c->clone();
2672 c->set_req(1, load_klass);
2673 c = _gvn.transform(c);
2674 }
2675 }
2676 }
2677 return c;
2678 }
2679
2680 //------------------------------do_one_bytecode--------------------------------
2681 // Parse this bytecode, and alter the Parsers JVM->Node mapping
2682 void Parse::do_one_bytecode() {
2683 Node *a, *b, *c, *d; // Handy temps
2684 BoolTest::mask btest;
2685 int i;
2686
2687 assert(!has_exceptions(), "bytecode entry state must be clear of throws");
2688
2689 if (C->check_node_count(NodeLimitFudgeFactor * 5,
2690 "out of nodes parsing method")) {
2691 return;
2692 }
2693
2694 #ifdef ASSERT
2695 // for setting breakpoints
2696 if (TraceOptoParse) {
2697 tty->print(" @");
2698 dump_bci(bci());
2699 tty->print(" %s", Bytecodes::name(bc()));
2700 tty->cr();
2701 }
2702 #endif
2703
2704 switch (bc()) {
2705 case Bytecodes::_nop:
2706 // do nothing
2707 break;
2708 case Bytecodes::_lconst_0:
2709 push_pair(longcon(0));
2710 break;
2711
2712 case Bytecodes::_lconst_1:
2713 push_pair(longcon(1));
2714 break;
2715
2716 case Bytecodes::_fconst_0:
2717 push(zerocon(T_FLOAT));
2718 break;
2719
2720 case Bytecodes::_fconst_1:
2721 push(makecon(TypeF::ONE));
2722 break;
2723
2724 case Bytecodes::_fconst_2:
2725 push(makecon(TypeF::make(2.0f)));
2726 break;
2727
2728 case Bytecodes::_dconst_0:
2729 push_pair(zerocon(T_DOUBLE));
2730 break;
2731
2732 case Bytecodes::_dconst_1:
2733 push_pair(makecon(TypeD::ONE));
2734 break;
2735
2736 case Bytecodes::_iconst_m1:push(intcon(-1)); break;
2737 case Bytecodes::_iconst_0: push(intcon( 0)); break;
2738 case Bytecodes::_iconst_1: push(intcon( 1)); break;
2739 case Bytecodes::_iconst_2: push(intcon( 2)); break;
2740 case Bytecodes::_iconst_3: push(intcon( 3)); break;
2741 case Bytecodes::_iconst_4: push(intcon( 4)); break;
2742 case Bytecodes::_iconst_5: push(intcon( 5)); break;
2743 case Bytecodes::_bipush: push(intcon(iter().get_constant_u1())); break;
2744 case Bytecodes::_sipush: push(intcon(iter().get_constant_u2())); break;
2745 case Bytecodes::_aconst_null: push(null()); break;
2746
2747 case Bytecodes::_ldc:
2748 case Bytecodes::_ldc_w:
2749 case Bytecodes::_ldc2_w: {
2750 // ciTypeFlow should trap if the ldc is in error state or if the constant is not loaded
2751 assert(!iter().is_in_error(), "ldc is in error state");
2752 ciConstant constant = iter().get_constant();
2753 assert(constant.is_loaded(), "constant is not loaded");
2754 const Type* con_type = Type::make_from_constant(constant);
2755 if (con_type != nullptr) {
2756 push_node(con_type->basic_type(), makecon(con_type));
2757 }
2758 break;
2759 }
2760
2761 case Bytecodes::_aload_0:
2762 push( local(0) );
2763 break;
2764 case Bytecodes::_aload_1:
2765 push( local(1) );
2766 break;
2767 case Bytecodes::_aload_2:
2768 push( local(2) );
2769 break;
2770 case Bytecodes::_aload_3:
2771 push( local(3) );
2772 break;
2773 case Bytecodes::_aload:
2774 push( local(iter().get_index()) );
2775 break;
2776
2777 case Bytecodes::_fload_0:
2778 case Bytecodes::_iload_0:
2779 push( local(0) );
2780 break;
2781 case Bytecodes::_fload_1:
2782 case Bytecodes::_iload_1:
2783 push( local(1) );
2784 break;
2785 case Bytecodes::_fload_2:
2786 case Bytecodes::_iload_2:
2787 push( local(2) );
2788 break;
2789 case Bytecodes::_fload_3:
2790 case Bytecodes::_iload_3:
2791 push( local(3) );
2792 break;
2793 case Bytecodes::_fload:
2794 case Bytecodes::_iload:
2795 push( local(iter().get_index()) );
2796 break;
2797 case Bytecodes::_lload_0:
2798 push_pair_local( 0 );
2799 break;
2800 case Bytecodes::_lload_1:
2801 push_pair_local( 1 );
2802 break;
2803 case Bytecodes::_lload_2:
2804 push_pair_local( 2 );
2805 break;
2806 case Bytecodes::_lload_3:
2807 push_pair_local( 3 );
2808 break;
2809 case Bytecodes::_lload:
2810 push_pair_local( iter().get_index() );
2811 break;
2812
2813 case Bytecodes::_dload_0:
2814 push_pair_local(0);
2815 break;
2816 case Bytecodes::_dload_1:
2817 push_pair_local(1);
2818 break;
2819 case Bytecodes::_dload_2:
2820 push_pair_local(2);
2821 break;
2822 case Bytecodes::_dload_3:
2823 push_pair_local(3);
2824 break;
2825 case Bytecodes::_dload:
2826 push_pair_local(iter().get_index());
2827 break;
2828 case Bytecodes::_fstore_0:
2829 case Bytecodes::_istore_0:
2830 case Bytecodes::_astore_0:
2831 set_local( 0, pop() );
2832 break;
2833 case Bytecodes::_fstore_1:
2834 case Bytecodes::_istore_1:
2835 case Bytecodes::_astore_1:
2836 set_local( 1, pop() );
2837 break;
2838 case Bytecodes::_fstore_2:
2839 case Bytecodes::_istore_2:
2840 case Bytecodes::_astore_2:
2841 set_local( 2, pop() );
2842 break;
2843 case Bytecodes::_fstore_3:
2844 case Bytecodes::_istore_3:
2845 case Bytecodes::_astore_3:
2846 set_local( 3, pop() );
2847 break;
2848 case Bytecodes::_fstore:
2849 case Bytecodes::_istore:
2850 case Bytecodes::_astore:
2851 set_local( iter().get_index(), pop() );
2852 break;
2853 // long stores
2854 case Bytecodes::_lstore_0:
2855 set_pair_local( 0, pop_pair() );
2856 break;
2857 case Bytecodes::_lstore_1:
2858 set_pair_local( 1, pop_pair() );
2859 break;
2860 case Bytecodes::_lstore_2:
2861 set_pair_local( 2, pop_pair() );
2862 break;
2863 case Bytecodes::_lstore_3:
2864 set_pair_local( 3, pop_pair() );
2865 break;
2866 case Bytecodes::_lstore:
2867 set_pair_local( iter().get_index(), pop_pair() );
2868 break;
2869
2870 // double stores
2871 case Bytecodes::_dstore_0:
2872 set_pair_local( 0, pop_pair() );
2873 break;
2874 case Bytecodes::_dstore_1:
2875 set_pair_local( 1, pop_pair() );
2876 break;
2877 case Bytecodes::_dstore_2:
2878 set_pair_local( 2, pop_pair() );
2879 break;
2880 case Bytecodes::_dstore_3:
2881 set_pair_local( 3, pop_pair() );
2882 break;
2883 case Bytecodes::_dstore:
2884 set_pair_local( iter().get_index(), pop_pair() );
2885 break;
2886
2887 case Bytecodes::_pop: dec_sp(1); break;
2888 case Bytecodes::_pop2: dec_sp(2); break;
2889 case Bytecodes::_swap:
2890 a = pop();
2891 b = pop();
2892 push(a);
2893 push(b);
2894 break;
2895 case Bytecodes::_dup:
2896 a = pop();
2897 push(a);
2898 push(a);
2899 break;
2900 case Bytecodes::_dup_x1:
2901 a = pop();
2902 b = pop();
2903 push( a );
2904 push( b );
2905 push( a );
2906 break;
2907 case Bytecodes::_dup_x2:
2908 a = pop();
2909 b = pop();
2910 c = pop();
2911 push( a );
2912 push( c );
2913 push( b );
2914 push( a );
2915 break;
2916 case Bytecodes::_dup2:
2917 a = pop();
2918 b = pop();
2919 push( b );
2920 push( a );
2921 push( b );
2922 push( a );
2923 break;
2924
2925 case Bytecodes::_dup2_x1:
2926 // before: .. c, b, a
2927 // after: .. b, a, c, b, a
2928 // not tested
2929 a = pop();
2930 b = pop();
2931 c = pop();
2932 push( b );
2933 push( a );
2934 push( c );
2935 push( b );
2936 push( a );
2937 break;
2938 case Bytecodes::_dup2_x2:
2939 // before: .. d, c, b, a
2940 // after: .. b, a, d, c, b, a
2941 // not tested
2942 a = pop();
2943 b = pop();
2944 c = pop();
2945 d = pop();
2946 push( b );
2947 push( a );
2948 push( d );
2949 push( c );
2950 push( b );
2951 push( a );
2952 break;
2953
2954 case Bytecodes::_arraylength: {
2955 // Must do null-check with value on expression stack
2956 Node *ary = null_check(peek(), T_ARRAY);
2957 // Compile-time detect of null-exception?
2958 if (stopped()) return;
2959 a = pop();
2960 push(load_array_length(a));
2961 break;
2962 }
2963
2964 case Bytecodes::_baload: array_load(T_BYTE); break;
2965 case Bytecodes::_caload: array_load(T_CHAR); break;
2966 case Bytecodes::_iaload: array_load(T_INT); break;
2967 case Bytecodes::_saload: array_load(T_SHORT); break;
2968 case Bytecodes::_faload: array_load(T_FLOAT); break;
2969 case Bytecodes::_aaload: array_load(T_OBJECT); break;
2970 case Bytecodes::_laload: array_load(T_LONG); break;
2971 case Bytecodes::_daload: array_load(T_DOUBLE); break;
2972 case Bytecodes::_bastore: array_store(T_BYTE); break;
2973 case Bytecodes::_castore: array_store(T_CHAR); break;
2974 case Bytecodes::_iastore: array_store(T_INT); break;
2975 case Bytecodes::_sastore: array_store(T_SHORT); break;
2976 case Bytecodes::_fastore: array_store(T_FLOAT); break;
2977 case Bytecodes::_aastore: array_store(T_OBJECT); break;
2978 case Bytecodes::_lastore: array_store(T_LONG); break;
2979 case Bytecodes::_dastore: array_store(T_DOUBLE); break;
2980
2981 case Bytecodes::_getfield:
2982 do_getfield();
2983 break;
2984
2985 case Bytecodes::_getstatic:
2986 do_getstatic();
2987 break;
2988
2989 case Bytecodes::_putfield:
2990 do_putfield();
2991 break;
2992
2993 case Bytecodes::_putstatic:
2994 do_putstatic();
2995 break;
2996
2997 case Bytecodes::_irem:
2998 // Must keep both values on the expression-stack during null-check
2999 zero_check_int(peek());
3000 // Compile-time detect of null-exception?
3001 if (stopped()) return;
3002 b = pop();
3003 a = pop();
3004 push(_gvn.transform(new ModINode(control(), a, b)));
3005 break;
3006 case Bytecodes::_idiv:
3007 // Must keep both values on the expression-stack during null-check
3008 zero_check_int(peek());
3009 // Compile-time detect of null-exception?
3010 if (stopped()) return;
3011 b = pop();
3012 a = pop();
3013 push( _gvn.transform( new DivINode(control(),a,b) ) );
3014 break;
3015 case Bytecodes::_imul:
3016 b = pop(); a = pop();
3017 push( _gvn.transform( new MulINode(a,b) ) );
3018 break;
3019 case Bytecodes::_iadd:
3020 b = pop(); a = pop();
3021 push( _gvn.transform( new AddINode(a,b) ) );
3022 break;
3023 case Bytecodes::_ineg:
3024 a = pop();
3025 push( _gvn.transform( new SubINode(_gvn.intcon(0),a)) );
3026 break;
3027 case Bytecodes::_isub:
3028 b = pop(); a = pop();
3029 push( _gvn.transform( new SubINode(a,b) ) );
3030 break;
3031 case Bytecodes::_iand:
3032 b = pop(); a = pop();
3033 push( _gvn.transform( new AndINode(a,b) ) );
3034 break;
3035 case Bytecodes::_ior:
3036 b = pop(); a = pop();
3037 push( _gvn.transform( new OrINode(a,b) ) );
3038 break;
3039 case Bytecodes::_ixor:
3040 b = pop(); a = pop();
3041 push( _gvn.transform( new XorINode(a,b) ) );
3042 break;
3043 case Bytecodes::_ishl:
3044 b = pop(); a = pop();
3045 push( _gvn.transform( new LShiftINode(a,b) ) );
3046 break;
3047 case Bytecodes::_ishr:
3048 b = pop(); a = pop();
3049 push( _gvn.transform( new RShiftINode(a,b) ) );
3050 break;
3051 case Bytecodes::_iushr:
3052 b = pop(); a = pop();
3053 push( _gvn.transform( new URShiftINode(a,b) ) );
3054 break;
3055
3056 case Bytecodes::_fneg:
3057 a = pop();
3058 b = _gvn.transform(new NegFNode (a));
3059 push(b);
3060 break;
3061
3062 case Bytecodes::_fsub:
3063 b = pop();
3064 a = pop();
3065 c = _gvn.transform( new SubFNode(a,b) );
3066 push(c);
3067 break;
3068
3069 case Bytecodes::_fadd:
3070 b = pop();
3071 a = pop();
3072 c = _gvn.transform( new AddFNode(a,b) );
3073 push(c);
3074 break;
3075
3076 case Bytecodes::_fmul:
3077 b = pop();
3078 a = pop();
3079 c = _gvn.transform( new MulFNode(a,b) );
3080 push(c);
3081 break;
3082
3083 case Bytecodes::_fdiv:
3084 b = pop();
3085 a = pop();
3086 c = _gvn.transform( new DivFNode(nullptr,a,b) );
3087 push(c);
3088 break;
3089
3090 case Bytecodes::_frem:
3091 // Generate a ModF node.
3092 b = pop();
3093 a = pop();
3094 push(floating_point_mod(a, b, BasicType::T_FLOAT));
3095 break;
3096
3097 case Bytecodes::_fcmpl:
3098 b = pop();
3099 a = pop();
3100 c = _gvn.transform( new CmpF3Node( a, b));
3101 push(c);
3102 break;
3103 case Bytecodes::_fcmpg:
3104 b = pop();
3105 a = pop();
3106
3107 // Same as fcmpl but need to flip the unordered case. Swap the inputs,
3108 // which negates the result sign except for unordered. Flip the unordered
3109 // as well by using CmpF3 which implements unordered-lesser instead of
3110 // unordered-greater semantics. Finally, commute the result bits. Result
3111 // is same as using a CmpF3Greater except we did it with CmpF3 alone.
3112 c = _gvn.transform( new CmpF3Node( b, a));
3113 c = _gvn.transform( new SubINode(_gvn.intcon(0),c) );
3114 push(c);
3115 break;
3116
3117 case Bytecodes::_f2i:
3118 a = pop();
3119 push(_gvn.transform(new ConvF2INode(a)));
3120 break;
3121
3122 case Bytecodes::_d2i:
3123 a = pop_pair();
3124 b = _gvn.transform(new ConvD2INode(a));
3125 push( b );
3126 break;
3127
3128 case Bytecodes::_f2d:
3129 a = pop();
3130 b = _gvn.transform( new ConvF2DNode(a));
3131 push_pair( b );
3132 break;
3133
3134 case Bytecodes::_d2f:
3135 a = pop_pair();
3136 b = _gvn.transform( new ConvD2FNode(a));
3137 push( b );
3138 break;
3139
3140 case Bytecodes::_l2f:
3141 if (Matcher::convL2FSupported()) {
3142 a = pop_pair();
3143 b = _gvn.transform( new ConvL2FNode(a));
3144 push(b);
3145 } else {
3146 l2f();
3147 }
3148 break;
3149
3150 case Bytecodes::_l2d:
3151 a = pop_pair();
3152 b = _gvn.transform( new ConvL2DNode(a));
3153 push_pair(b);
3154 break;
3155
3156 case Bytecodes::_f2l:
3157 a = pop();
3158 b = _gvn.transform( new ConvF2LNode(a));
3159 push_pair(b);
3160 break;
3161
3162 case Bytecodes::_d2l:
3163 a = pop_pair();
3164 b = _gvn.transform( new ConvD2LNode(a));
3165 push_pair(b);
3166 break;
3167
3168 case Bytecodes::_dsub:
3169 b = pop_pair();
3170 a = pop_pair();
3171 c = _gvn.transform( new SubDNode(a,b) );
3172 push_pair(c);
3173 break;
3174
3175 case Bytecodes::_dadd:
3176 b = pop_pair();
3177 a = pop_pair();
3178 c = _gvn.transform( new AddDNode(a,b) );
3179 push_pair(c);
3180 break;
3181
3182 case Bytecodes::_dmul:
3183 b = pop_pair();
3184 a = pop_pair();
3185 c = _gvn.transform( new MulDNode(a,b) );
3186 push_pair(c);
3187 break;
3188
3189 case Bytecodes::_ddiv:
3190 b = pop_pair();
3191 a = pop_pair();
3192 c = _gvn.transform( new DivDNode(nullptr,a,b) );
3193 push_pair(c);
3194 break;
3195
3196 case Bytecodes::_dneg:
3197 a = pop_pair();
3198 b = _gvn.transform(new NegDNode (a));
3199 push_pair(b);
3200 break;
3201
3202 case Bytecodes::_drem:
3203 // Generate a ModD node.
3204 b = pop_pair();
3205 a = pop_pair();
3206 push_pair(floating_point_mod(a, b, BasicType::T_DOUBLE));
3207 break;
3208
3209 case Bytecodes::_dcmpl:
3210 b = pop_pair();
3211 a = pop_pair();
3212 c = _gvn.transform( new CmpD3Node( a, b));
3213 push(c);
3214 break;
3215
3216 case Bytecodes::_dcmpg:
3217 b = pop_pair();
3218 a = pop_pair();
3219 // Same as dcmpl but need to flip the unordered case.
3220 // Commute the inputs, which negates the result sign except for unordered.
3221 // Flip the unordered as well by using CmpD3 which implements
3222 // unordered-lesser instead of unordered-greater semantics.
3223 // Finally, negate the result bits. Result is same as using a
3224 // CmpD3Greater except we did it with CmpD3 alone.
3225 c = _gvn.transform( new CmpD3Node( b, a));
3226 c = _gvn.transform( new SubINode(_gvn.intcon(0),c) );
3227 push(c);
3228 break;
3229
3230
3231 // Note for longs -> lo word is on TOS, hi word is on TOS - 1
3232 case Bytecodes::_land:
3233 b = pop_pair();
3234 a = pop_pair();
3235 c = _gvn.transform( new AndLNode(a,b) );
3236 push_pair(c);
3237 break;
3238 case Bytecodes::_lor:
3239 b = pop_pair();
3240 a = pop_pair();
3241 c = _gvn.transform( new OrLNode(a,b) );
3242 push_pair(c);
3243 break;
3244 case Bytecodes::_lxor:
3245 b = pop_pair();
3246 a = pop_pair();
3247 c = _gvn.transform( new XorLNode(a,b) );
3248 push_pair(c);
3249 break;
3250
3251 case Bytecodes::_lshl:
3252 b = pop(); // the shift count
3253 a = pop_pair(); // value to be shifted
3254 c = _gvn.transform( new LShiftLNode(a,b) );
3255 push_pair(c);
3256 break;
3257 case Bytecodes::_lshr:
3258 b = pop(); // the shift count
3259 a = pop_pair(); // value to be shifted
3260 c = _gvn.transform( new RShiftLNode(a,b) );
3261 push_pair(c);
3262 break;
3263 case Bytecodes::_lushr:
3264 b = pop(); // the shift count
3265 a = pop_pair(); // value to be shifted
3266 c = _gvn.transform( new URShiftLNode(a,b) );
3267 push_pair(c);
3268 break;
3269 case Bytecodes::_lmul:
3270 b = pop_pair();
3271 a = pop_pair();
3272 c = _gvn.transform( new MulLNode(a,b) );
3273 push_pair(c);
3274 break;
3275
3276 case Bytecodes::_lrem:
3277 // Must keep both values on the expression-stack during null-check
3278 assert(peek(0) == top(), "long word order");
3279 zero_check_long(peek(1));
3280 // Compile-time detect of null-exception?
3281 if (stopped()) return;
3282 b = pop_pair();
3283 a = pop_pair();
3284 c = _gvn.transform( new ModLNode(control(),a,b) );
3285 push_pair(c);
3286 break;
3287
3288 case Bytecodes::_ldiv:
3289 // Must keep both values on the expression-stack during null-check
3290 assert(peek(0) == top(), "long word order");
3291 zero_check_long(peek(1));
3292 // Compile-time detect of null-exception?
3293 if (stopped()) return;
3294 b = pop_pair();
3295 a = pop_pair();
3296 c = _gvn.transform( new DivLNode(control(),a,b) );
3297 push_pair(c);
3298 break;
3299
3300 case Bytecodes::_ladd:
3301 b = pop_pair();
3302 a = pop_pair();
3303 c = _gvn.transform( new AddLNode(a,b) );
3304 push_pair(c);
3305 break;
3306 case Bytecodes::_lsub:
3307 b = pop_pair();
3308 a = pop_pair();
3309 c = _gvn.transform( new SubLNode(a,b) );
3310 push_pair(c);
3311 break;
3312 case Bytecodes::_lcmp:
3313 // Safepoints are now inserted _before_ branches. The long-compare
3314 // bytecode painfully produces a 3-way value (-1,0,+1) which requires a
3315 // slew of control flow. These are usually followed by a CmpI vs zero and
3316 // a branch; this pattern then optimizes to the obvious long-compare and
3317 // branch. However, if the branch is backwards there's a Safepoint
3318 // inserted. The inserted Safepoint captures the JVM state at the
3319 // pre-branch point, i.e. it captures the 3-way value. Thus if a
3320 // long-compare is used to control a loop the debug info will force
3321 // computation of the 3-way value, even though the generated code uses a
3322 // long-compare and branch. We try to rectify the situation by inserting
3323 // a SafePoint here and have it dominate and kill the safepoint added at a
3324 // following backwards branch. At this point the JVM state merely holds 2
3325 // longs but not the 3-way value.
3326 switch (iter().next_bc()) {
3327 case Bytecodes::_ifgt:
3328 case Bytecodes::_iflt:
3329 case Bytecodes::_ifge:
3330 case Bytecodes::_ifle:
3331 case Bytecodes::_ifne:
3332 case Bytecodes::_ifeq:
3333 // If this is a backwards branch in the bytecodes, add Safepoint
3334 maybe_add_safepoint(iter().next_get_dest());
3335 default:
3336 break;
3337 }
3338 b = pop_pair();
3339 a = pop_pair();
3340 c = _gvn.transform( new CmpL3Node( a, b ));
3341 push(c);
3342 break;
3343
3344 case Bytecodes::_lneg:
3345 a = pop_pair();
3346 b = _gvn.transform( new SubLNode(longcon(0),a));
3347 push_pair(b);
3348 break;
3349 case Bytecodes::_l2i:
3350 a = pop_pair();
3351 push( _gvn.transform( new ConvL2INode(a)));
3352 break;
3353 case Bytecodes::_i2l:
3354 a = pop();
3355 b = _gvn.transform( new ConvI2LNode(a));
3356 push_pair(b);
3357 break;
3358 case Bytecodes::_i2b:
3359 // Sign extend
3360 a = pop();
3361 a = Compile::narrow_value(T_BYTE, a, nullptr, &_gvn, true);
3362 push(a);
3363 break;
3364 case Bytecodes::_i2s:
3365 a = pop();
3366 a = Compile::narrow_value(T_SHORT, a, nullptr, &_gvn, true);
3367 push(a);
3368 break;
3369 case Bytecodes::_i2c:
3370 a = pop();
3371 a = Compile::narrow_value(T_CHAR, a, nullptr, &_gvn, true);
3372 push(a);
3373 break;
3374
3375 case Bytecodes::_i2f:
3376 a = pop();
3377 b = _gvn.transform( new ConvI2FNode(a) ) ;
3378 push(b);
3379 break;
3380
3381 case Bytecodes::_i2d:
3382 a = pop();
3383 b = _gvn.transform( new ConvI2DNode(a));
3384 push_pair(b);
3385 break;
3386
3387 case Bytecodes::_iinc: // Increment local
3388 i = iter().get_index(); // Get local index
3389 set_local( i, _gvn.transform( new AddINode( _gvn.intcon(iter().get_iinc_con()), local(i) ) ) );
3390 break;
3391
3392 // Exit points of synchronized methods must have an unlock node
3393 case Bytecodes::_return:
3394 return_current(nullptr);
3395 break;
3396
3397 case Bytecodes::_ireturn:
3398 case Bytecodes::_areturn:
3399 case Bytecodes::_freturn:
3400 return_current(pop());
3401 break;
3402 case Bytecodes::_lreturn:
3403 return_current(pop_pair());
3404 break;
3405 case Bytecodes::_dreturn:
3406 return_current(pop_pair());
3407 break;
3408
3409 case Bytecodes::_athrow:
3410 // null exception oop throws null pointer exception
3411 null_check(peek());
3412 if (stopped()) return;
3413 // Hook the thrown exception directly to subsequent handlers.
3414 if (BailoutToInterpreterForThrows) {
3415 // Keep method interpreted from now on.
3416 uncommon_trap(Deoptimization::Reason_unhandled,
3417 Deoptimization::Action_make_not_compilable);
3418 return;
3419 }
3420 if (env()->jvmti_can_post_on_exceptions()) {
3421 // check if we must post exception events, take uncommon trap if so (with must_throw = false)
3422 uncommon_trap_if_should_post_on_exceptions(Deoptimization::Reason_unhandled, false);
3423 }
3424 // Here if either can_post_on_exceptions or should_post_on_exceptions is false
3425 add_exception_state(make_exception_state(peek()));
3426 break;
3427
3428 case Bytecodes::_goto: // fall through
3429 case Bytecodes::_goto_w: {
3430 int target_bci = (bc() == Bytecodes::_goto) ? iter().get_dest() : iter().get_far_dest();
3431
3432 // If this is a backwards branch in the bytecodes, add Safepoint
3433 maybe_add_safepoint(target_bci);
3434
3435 // Merge the current control into the target basic block
3436 merge(target_bci);
3437
3438 // See if we can get some profile data and hand it off to the next block
3439 Block *target_block = block()->successor_for_bci(target_bci);
3440 if (target_block->pred_count() != 1) break;
3441 ciMethodData* methodData = method()->method_data();
3442 if (!methodData->is_mature()) break;
3443 ciProfileData* data = methodData->bci_to_data(bci());
3444 assert(data != nullptr && data->is_JumpData(), "need JumpData for taken branch");
3445 int taken = ((ciJumpData*)data)->taken();
3446 taken = method()->scale_count(taken);
3447 target_block->set_count(taken);
3448 break;
3449 }
3450
3451 case Bytecodes::_ifnull: btest = BoolTest::eq; goto handle_if_null;
3452 case Bytecodes::_ifnonnull: btest = BoolTest::ne; goto handle_if_null;
3453 handle_if_null:
3454 // If this is a backwards branch in the bytecodes, add Safepoint
3455 maybe_add_safepoint(iter().get_dest());
3456 a = null();
3457 b = pop();
3458 if (b->is_InlineType()) {
3459 // Null checking a scalarized but nullable inline type. Check the IsInit
3460 // input instead of the oop input to avoid keeping buffer allocations alive
3461 c = _gvn.transform(new CmpINode(b->as_InlineType()->get_is_init(), zerocon(T_INT)));
3462 } else {
3463 if (!_gvn.type(b)->speculative_maybe_null() &&
3464 !too_many_traps(Deoptimization::Reason_speculate_null_check)) {
3465 inc_sp(1);
3466 Node* null_ctl = top();
3467 b = null_check_oop(b, &null_ctl, true, true, true);
3468 assert(null_ctl->is_top(), "no null control here");
3469 dec_sp(1);
3470 } else if (_gvn.type(b)->speculative_always_null() &&
3471 !too_many_traps(Deoptimization::Reason_speculate_null_assert)) {
3472 inc_sp(1);
3473 b = null_assert(b);
3474 dec_sp(1);
3475 }
3476 c = _gvn.transform( new CmpPNode(b, a) );
3477 }
3478 do_ifnull(btest, c);
3479 break;
3480
3481 case Bytecodes::_if_acmpeq: btest = BoolTest::eq; goto handle_if_acmp;
3482 case Bytecodes::_if_acmpne: btest = BoolTest::ne; goto handle_if_acmp;
3483 handle_if_acmp:
3484 // If this is a backwards branch in the bytecodes, add Safepoint
3485 maybe_add_safepoint(iter().get_dest());
3486 a = pop();
3487 b = pop();
3488 do_acmp(btest, b, a);
3489 break;
3490
3491 case Bytecodes::_ifeq: btest = BoolTest::eq; goto handle_ifxx;
3492 case Bytecodes::_ifne: btest = BoolTest::ne; goto handle_ifxx;
3493 case Bytecodes::_iflt: btest = BoolTest::lt; goto handle_ifxx;
3494 case Bytecodes::_ifle: btest = BoolTest::le; goto handle_ifxx;
3495 case Bytecodes::_ifgt: btest = BoolTest::gt; goto handle_ifxx;
3496 case Bytecodes::_ifge: btest = BoolTest::ge; goto handle_ifxx;
3497 handle_ifxx:
3498 // If this is a backwards branch in the bytecodes, add Safepoint
3499 maybe_add_safepoint(iter().get_dest());
3500 a = _gvn.intcon(0);
3501 b = pop();
3502 c = _gvn.transform( new CmpINode(b, a) );
3503 do_if(btest, c);
3504 break;
3505
3506 case Bytecodes::_if_icmpeq: btest = BoolTest::eq; goto handle_if_icmp;
3507 case Bytecodes::_if_icmpne: btest = BoolTest::ne; goto handle_if_icmp;
3508 case Bytecodes::_if_icmplt: btest = BoolTest::lt; goto handle_if_icmp;
3509 case Bytecodes::_if_icmple: btest = BoolTest::le; goto handle_if_icmp;
3510 case Bytecodes::_if_icmpgt: btest = BoolTest::gt; goto handle_if_icmp;
3511 case Bytecodes::_if_icmpge: btest = BoolTest::ge; goto handle_if_icmp;
3512 handle_if_icmp:
3513 // If this is a backwards branch in the bytecodes, add Safepoint
3514 maybe_add_safepoint(iter().get_dest());
3515 a = pop();
3516 b = pop();
3517 c = _gvn.transform( new CmpINode( b, a ) );
3518 do_if(btest, c);
3519 break;
3520
3521 case Bytecodes::_tableswitch:
3522 do_tableswitch();
3523 break;
3524
3525 case Bytecodes::_lookupswitch:
3526 do_lookupswitch();
3527 break;
3528
3529 case Bytecodes::_invokestatic:
3530 case Bytecodes::_invokedynamic:
3531 case Bytecodes::_invokespecial:
3532 case Bytecodes::_invokevirtual:
3533 case Bytecodes::_invokeinterface:
3534 do_call();
3535 break;
3536 case Bytecodes::_checkcast:
3537 do_checkcast();
3538 break;
3539 case Bytecodes::_instanceof:
3540 do_instanceof();
3541 break;
3542 case Bytecodes::_anewarray:
3543 do_newarray();
3544 break;
3545 case Bytecodes::_newarray:
3546 do_newarray((BasicType)iter().get_index());
3547 break;
3548 case Bytecodes::_multianewarray:
3549 do_multianewarray();
3550 break;
3551 case Bytecodes::_new:
3552 do_new();
3553 break;
3554
3555 case Bytecodes::_jsr:
3556 case Bytecodes::_jsr_w:
3557 do_jsr();
3558 break;
3559
3560 case Bytecodes::_ret:
3561 do_ret();
3562 break;
3563
3564
3565 case Bytecodes::_monitorenter:
3566 do_monitor_enter();
3567 break;
3568
3569 case Bytecodes::_monitorexit:
3570 do_monitor_exit();
3571 break;
3572
3573 case Bytecodes::_breakpoint:
3574 // Breakpoint set concurrently to compile
3575 // %%% use an uncommon trap?
3576 C->record_failure("breakpoint in method");
3577 return;
3578
3579 default:
3580 #ifndef PRODUCT
3581 map()->dump(99);
3582 #endif
3583 tty->print("\nUnhandled bytecode %s\n", Bytecodes::name(bc()) );
3584 ShouldNotReachHere();
3585 }
3586
3587 #ifndef PRODUCT
3588 if (failing()) { return; }
3589 constexpr int perBytecode = 6;
3590 if (C->should_print_igv(perBytecode)) {
3591 IdealGraphPrinter* printer = C->igv_printer();
3592 char buffer[256];
3593 jio_snprintf(buffer, sizeof(buffer), "Bytecode %d: %s", bci(), Bytecodes::name(bc()));
3594 bool old = printer->traverse_outs();
3595 printer->set_traverse_outs(true);
3596 printer->print_graph(buffer);
3597 printer->set_traverse_outs(old);
3598 }
3599 #endif
3600 }