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 // Re-execute flat array load if buffering triggers deoptimization
129 PreserveReexecuteState preexecs(this);
130 jvms()->set_should_reexecute(true);
131 inc_sp(3);
132
133 adr = flat_array_element_address(array, array_index, vk, is_null_free, is_not_null_free, may_need_atomicity);
134 int nm_offset = is_null_free ? -1 : vk->null_marker_offset_in_payload();
135 Node* vt = InlineTypeNode::make_from_flat(this, vk, array, adr, array_index, nullptr, 0, may_need_atomicity, nm_offset);
136 ideal.set(res, vt);
137 } else {
138 // Element type is unknown, and thus we cannot statically determine the exact flat array layout. Emit a
139 // runtime call to correctly load the inline type element from the flat array.
140 Node* inline_type = load_from_unknown_flat_array(array, array_index, element_ptr);
141 bool is_null_free = array_type->is_null_free() || !UseNullableValueFlattening;
142 if (is_null_free) {
143 inline_type = cast_not_null(inline_type);
144 }
145 ideal.set(res, inline_type);
146 }
147 }
148 ideal.sync_kit(this);
149 } ideal.end_if();
150 sync_kit(ideal);
151 Node* ld = _gvn.transform(ideal.value(res));
152 ld = record_profile_for_speculation_at_array_load(ld);
153 push_node(bt, ld);
154 return;
155 }
156
157 if (elemtype == TypeInt::BOOL) {
158 bt = T_BOOLEAN;
159 }
160 const TypeAryPtr* adr_type = TypeAryPtr::get_array_body_type(bt);
161 Node* ld = access_load_at(array, adr, adr_type, elemtype, bt,
162 IN_HEAP | IS_ARRAY | C2_CONTROL_DEPENDENT_LOAD);
163 ld = record_profile_for_speculation_at_array_load(ld);
164 // Loading an inline type from a non-flat array
165 if (element_ptr != nullptr && element_ptr->is_inlinetypeptr()) {
166 assert(!array_type->is_null_free() || !element_ptr->maybe_null(), "inline type array elements should never be null");
167 ld = InlineTypeNode::make_from_oop(this, ld, element_ptr->inline_klass());
168 }
169 push_node(bt, ld);
170 }
171
172 Node* Parse::load_from_unknown_flat_array(Node* array, Node* array_index, const TypeOopPtr* element_ptr) {
173 // Below membars keep this access to an unknown flat array correctly
174 // ordered with other unknown and known flat array accesses.
175 insert_mem_bar_volatile(Op_MemBarCPUOrder, C->get_alias_index(TypeAryPtr::INLINES));
176
177 Node* call = nullptr;
178 {
179 // Re-execute flat array load if runtime call triggers deoptimization
180 PreserveReexecuteState preexecs(this);
181 jvms()->set_bci(_bci);
182 jvms()->set_should_reexecute(true);
183 inc_sp(2);
184 kill_dead_locals();
185 call = make_runtime_call(RC_NO_LEAF | RC_NO_IO,
186 OptoRuntime::load_unknown_inline_Type(),
187 OptoRuntime::load_unknown_inline_Java(),
188 nullptr, TypeRawPtr::BOTTOM,
189 array, array_index);
190 }
191 make_slow_call_ex(call, env()->Throwable_klass(), false);
192 Node* buffer = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
193
194 insert_mem_bar_volatile(Op_MemBarCPUOrder, C->get_alias_index(TypeAryPtr::INLINES));
195
196 // Keep track of the information that the inline type is in flat arrays
197 const Type* unknown_value = element_ptr->is_instptr()->cast_to_flat_in_array();
198 return _gvn.transform(new CheckCastPPNode(control(), buffer, unknown_value));
199 }
200
201 //--------------------------------array_store----------------------------------
202 void Parse::array_store(BasicType bt) {
203 const Type* elemtype = Type::TOP;
204 Node* adr = array_addressing(bt, type2size[bt], elemtype);
205 if (stopped()) return; // guaranteed null or range check
206 Node* stored_value_casted = nullptr;
207 if (bt == T_OBJECT) {
208 stored_value_casted = array_store_check(adr, elemtype);
209 if (stopped()) {
210 return;
211 }
212 }
213 Node* const stored_value = pop_node(bt); // Value to store
214 Node* const array_index = pop(); // Index in the array
215 Node* array = pop(); // The array itself
216
217 const TypeAryPtr* array_type = _gvn.type(array)->is_aryptr();
218 const TypeAryPtr* adr_type = TypeAryPtr::get_array_body_type(bt);
219
220 if (elemtype == TypeInt::BOOL) {
221 bt = T_BOOLEAN;
222 } else if (bt == T_OBJECT) {
223 elemtype = elemtype->make_oopptr();
224 const Type* stored_value_casted_type = _gvn.type(stored_value_casted);
225 // Based on the value to be stored, try to determine if the array is not null-free and/or not flat.
226 // This is only legal for non-null stores because the array_store_check always passes for null, even
227 // if the array is null-free. Null stores are handled in GraphKit::inline_array_null_guard().
228 bool not_inline = !stored_value_casted_type->maybe_null() && !stored_value_casted_type->is_oopptr()->can_be_inline_type();
229 bool not_null_free = not_inline;
230 bool not_flat = not_inline || ( stored_value_casted_type->is_inlinetypeptr() &&
231 !stored_value_casted_type->inline_klass()->flat_in_array());
232 if (!array_type->is_not_null_free() && not_null_free) {
233 // Storing a non-inline type, mark array as not null-free.
234 array_type = array_type->cast_to_not_null_free();
235 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, array_type));
236 replace_in_map(array, cast);
237 array = cast;
238 }
239 if (!array_type->is_not_flat() && not_flat) {
240 // Storing to a non-flat array, mark array as not flat.
241 array_type = array_type->cast_to_not_flat();
242 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, array_type));
243 replace_in_map(array, cast);
244 array = cast;
245 }
246
247 if (!array_type->is_flat() && array_type->is_null_free()) {
248 // Store to non-flat null-free inline type array (elements can never be null)
249 assert(!stored_value_casted_type->maybe_null(), "should be guaranteed by array store check");
250 if (elemtype->is_inlinetypeptr() && elemtype->inline_klass()->is_empty()) {
251 // Ignore empty inline stores, array is already initialized.
252 return;
253 }
254 } else if (!array_type->is_not_flat()) {
255 // Array might be a flat array, emit runtime checks (for nullptr, a simple inline_array_null_guard is sufficient).
256 assert(UseArrayFlattening && !not_flat && elemtype->is_oopptr()->can_be_inline_type() &&
257 (!array_type->klass_is_exact() || array_type->is_flat()), "array can't be a flat array");
258 // 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.
259 array = inline_array_null_guard(array, stored_value_casted, 3, true);
260 IdealKit ideal(this);
261 ideal.if_then(flat_array_test(array, /* flat = */ false)); {
262 // Non-flat array
263 if (!array_type->is_flat()) {
264 sync_kit(ideal);
265 assert(array_type->is_flat() || ideal.ctrl()->in(0)->as_If()->is_flat_array_check(&_gvn), "Should be found");
266 inc_sp(3);
267 access_store_at(array, adr, adr_type, stored_value_casted, elemtype, bt, MO_UNORDERED | IN_HEAP | IS_ARRAY, false);
268 dec_sp(3);
269 ideal.sync_kit(this);
270 }
271 } ideal.else_(); {
272 // Flat array
273 sync_kit(ideal);
274 if (!array_type->is_not_flat()) {
275 // Try to determine the inline klass type of the stored value
276 ciInlineKlass* vk = nullptr;
277 if (stored_value_casted_type->is_inlinetypeptr()) {
278 vk = stored_value_casted_type->inline_klass();
279 } else if (elemtype->is_inlinetypeptr()) {
280 vk = elemtype->inline_klass();
281 }
282
283 if (vk != nullptr) {
284 // Element type is known, cast and store to flat array layout.
285 bool is_null_free = array_type->is_null_free() || !vk->has_nullable_atomic_layout();
286 bool is_not_null_free = array_type->is_not_null_free() || (!vk->has_atomic_layout() && !vk->has_non_atomic_layout());
287 if (is_null_free) {
288 // TODO 8350865 Impossible type
289 is_not_null_free = false;
290 }
291 bool is_naturally_atomic = (is_null_free && vk->nof_declared_nonstatic_fields() <= 1);
292 bool may_need_atomicity = !is_naturally_atomic && ((!is_not_null_free && vk->has_atomic_layout()) || (!is_null_free && vk->has_nullable_atomic_layout()));
293
294 // Re-execute flat array store if buffering triggers deoptimization
295 PreserveReexecuteState preexecs(this);
296 jvms()->set_should_reexecute(true);
297 inc_sp(3);
298
299 if (!stored_value_casted->is_InlineType()) {
300 assert(_gvn.type(stored_value_casted) == TypePtr::NULL_PTR, "Unexpected value");
301 stored_value_casted = InlineTypeNode::make_null(_gvn, vk);
302 }
303 adr = flat_array_element_address(array, array_index, vk, is_null_free, is_not_null_free, may_need_atomicity);
304 int nm_offset = is_null_free ? -1 : vk->null_marker_offset_in_payload();
305 stored_value_casted->as_InlineType()->store_flat(this, array, adr, array_index, vk, 0, may_need_atomicity, nm_offset, MO_UNORDERED | IN_HEAP | IS_ARRAY);
306 } else {
307 // Element type is unknown, emit a runtime call since the flat array layout is not statically known.
308 store_to_unknown_flat_array(array, array_index, stored_value_casted);
309 }
310 }
311 ideal.sync_kit(this);
312 }
313 ideal.end_if();
314 sync_kit(ideal);
315 return;
316 } else if (!array_type->is_not_null_free()) {
317 // Array is not flat but may be null free
318 assert(elemtype->is_oopptr()->can_be_inline_type(), "array can't be null-free");
319 array = inline_array_null_guard(array, stored_value_casted, 3, true);
320 }
321 }
322 inc_sp(3);
323 access_store_at(array, adr, adr_type, stored_value, elemtype, bt, MO_UNORDERED | IN_HEAP | IS_ARRAY);
324 dec_sp(3);
325 }
326
327 // Emit a runtime call to store to a flat array whose element type is either unknown (i.e. we do not know the flat
328 // array layout) or not exact (could have different flat array layouts at runtime).
329 void Parse::store_to_unknown_flat_array(Node* array, Node* const idx, Node* non_null_stored_value) {
330 // Below membars keep this access to an unknown flat array correctly
331 // ordered with other unknown and known flat array accesses.
332 insert_mem_bar_volatile(Op_MemBarCPUOrder, C->get_alias_index(TypeAryPtr::INLINES));
333
334 Node* call = nullptr;
335 {
336 // Re-execute flat array store if runtime call triggers deoptimization
337 PreserveReexecuteState preexecs(this);
338 jvms()->set_bci(_bci);
339 jvms()->set_should_reexecute(true);
340 inc_sp(3);
341 kill_dead_locals();
342 call = make_runtime_call(RC_NO_LEAF | RC_NO_IO,
343 OptoRuntime::store_unknown_inline_Type(),
344 OptoRuntime::store_unknown_inline_Java(),
345 nullptr, TypeRawPtr::BOTTOM,
346 non_null_stored_value, array, idx);
347 }
348 make_slow_call_ex(call, env()->Throwable_klass(), false);
349
350 insert_mem_bar_volatile(Op_MemBarCPUOrder, C->get_alias_index(TypeAryPtr::INLINES));
351 }
352
353 //------------------------------array_addressing-------------------------------
354 // Pull array and index from the stack. Compute pointer-to-element.
355 Node* Parse::array_addressing(BasicType type, int vals, const Type*& elemtype) {
356 Node *idx = peek(0+vals); // Get from stack without popping
357 Node *ary = peek(1+vals); // in case of exception
358
359 // Null check the array base, with correct stack contents
360 ary = null_check(ary, T_ARRAY);
361 // Compile-time detect of null-exception?
362 if (stopped()) return top();
363
364 const TypeAryPtr* arytype = _gvn.type(ary)->is_aryptr();
365 const TypeInt* sizetype = arytype->size();
366 elemtype = arytype->elem();
367
368 if (UseUniqueSubclasses) {
369 const Type* el = elemtype->make_ptr();
370 if (el && el->isa_instptr()) {
371 const TypeInstPtr* toop = el->is_instptr();
372 if (toop->instance_klass()->unique_concrete_subklass()) {
373 // If we load from "AbstractClass[]" we must see "ConcreteSubClass".
374 const Type* subklass = Type::get_const_type(toop->instance_klass());
375 elemtype = subklass->join_speculative(el);
376 }
377 }
378 }
379
380 if (!arytype->is_loaded()) {
381 // Only fails for some -Xcomp runs
382 // The class is unloaded. We have to run this bytecode in the interpreter.
383 ciKlass* klass = arytype->unloaded_klass();
384
385 uncommon_trap(Deoptimization::Reason_unloaded,
386 Deoptimization::Action_reinterpret,
387 klass, "!loaded array");
388 return top();
389 }
390
391 ary = create_speculative_inline_type_array_checks(ary, arytype, elemtype);
392
393 if (needs_range_check(sizetype, idx)) {
394 create_range_check(idx, ary, sizetype);
395 } else if (C->log() != nullptr) {
396 C->log()->elem("observe that='!need_range_check'");
397 }
398
399 // Check for always knowing you are throwing a range-check exception
400 if (stopped()) return top();
401
402 // Make array address computation control dependent to prevent it
403 // from floating above the range check during loop optimizations.
404 Node* ptr = array_element_address(ary, idx, type, sizetype, control());
405 assert(ptr != top(), "top should go hand-in-hand with stopped");
406
407 return ptr;
408 }
409
410 // 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
411 // be greater or equal the smallest possible array size (i.e. out-of-bounds).
412 bool Parse::needs_range_check(const TypeInt* size_type, const Node* index) const {
413 const TypeInt* index_type = _gvn.type(index)->is_int();
414 return index_type->_hi >= size_type->_lo || index_type->_lo < 0;
415 }
416
417 void Parse::create_range_check(Node* idx, Node* ary, const TypeInt* sizetype) {
418 Node* tst;
419 if (sizetype->_hi <= 0) {
420 // The greatest array bound is negative, so we can conclude that we're
421 // compiling unreachable code, but the unsigned compare trick used below
422 // only works with non-negative lengths. Instead, hack "tst" to be zero so
423 // the uncommon_trap path will always be taken.
424 tst = _gvn.intcon(0);
425 } else {
426 // Range is constant in array-oop, so we can use the original state of mem
427 Node* len = load_array_length(ary);
428
429 // Test length vs index (standard trick using unsigned compare)
430 Node* chk = _gvn.transform(new CmpUNode(idx, len) );
431 BoolTest::mask btest = BoolTest::lt;
432 tst = _gvn.transform(new BoolNode(chk, btest) );
433 }
434 RangeCheckNode* rc = new RangeCheckNode(control(), tst, PROB_MAX, COUNT_UNKNOWN);
435 _gvn.set_type(rc, rc->Value(&_gvn));
436 if (!tst->is_Con()) {
437 record_for_igvn(rc);
438 }
439 set_control(_gvn.transform(new IfTrueNode(rc)));
440 // Branch to failure if out of bounds
441 {
442 PreserveJVMState pjvms(this);
443 set_control(_gvn.transform(new IfFalseNode(rc)));
444 if (C->allow_range_check_smearing()) {
445 // Do not use builtin_throw, since range checks are sometimes
446 // made more stringent by an optimistic transformation.
447 // This creates "tentative" range checks at this point,
448 // which are not guaranteed to throw exceptions.
449 // See IfNode::Ideal, is_range_check, adjust_check.
450 uncommon_trap(Deoptimization::Reason_range_check,
451 Deoptimization::Action_make_not_entrant,
452 nullptr, "range_check");
453 } else {
454 // If we have already recompiled with the range-check-widening
455 // heroic optimization turned off, then we must really be throwing
456 // range check exceptions.
457 builtin_throw(Deoptimization::Reason_range_check);
458 }
459 }
460 }
461
462 // For inline type arrays, we can use the profiling information for array accesses to speculate on the type, flatness,
463 // and null-freeness. We can either prepare the speculative type for later uses or emit explicit speculative checks with
464 // traps now. In the latter case, the speculative type guarantees can avoid additional runtime checks later (e.g.
465 // non-null-free implies non-flat which allows us to remove flatness checks). This makes the graph simpler.
466 Node* Parse::create_speculative_inline_type_array_checks(Node* array, const TypeAryPtr* array_type,
467 const Type*& element_type) {
468 if (!array_type->is_flat() && !array_type->is_not_flat()) {
469 // For arrays that might be flat, speculate that the array has the exact type reported in the profile data such that
470 // we can rely on a fixed memory layout (i.e. either a flat layout or not).
471 array = cast_to_speculative_array_type(array, array_type, element_type);
472 } else if (UseTypeSpeculation && UseArrayLoadStoreProfile) {
473 // Array is known to be either flat or not flat. If possible, update the speculative type by using the profile data
474 // at this bci.
475 array = cast_to_profiled_array_type(array);
476 }
477
478 // Even though the type does not tell us whether we have an inline type array or not, we can still check the profile data
479 // whether we have a non-null-free or non-flat array. Speculating on a non-null-free array doesn't help aaload but could
480 // be profitable for a subsequent aastore.
481 if (!array_type->is_null_free() && !array_type->is_not_null_free()) {
482 array = speculate_non_null_free_array(array, array_type);
483 }
484 if (!array_type->is_flat() && !array_type->is_not_flat()) {
485 array = speculate_non_flat_array(array, array_type);
486 }
487 return array;
488 }
489
490 // Speculate that the array has the exact type reported in the profile data. We emit a trap when this turns out to be
491 // wrong. On the fast path, we add a CheckCastPP to use the exact type.
492 Node* Parse::cast_to_speculative_array_type(Node* const array, const TypeAryPtr*& array_type, const Type*& element_type) {
493 Deoptimization::DeoptReason reason = Deoptimization::Reason_speculate_class_check;
494 ciKlass* speculative_array_type = array_type->speculative_type();
495 if (too_many_traps_or_recompiles(reason) || speculative_array_type == nullptr) {
496 // No speculative type, check profile data at this bci
497 speculative_array_type = nullptr;
498 reason = Deoptimization::Reason_class_check;
499 if (UseArrayLoadStoreProfile && !too_many_traps_or_recompiles(reason)) {
500 ciKlass* profiled_element_type = nullptr;
501 ProfilePtrKind element_ptr = ProfileMaybeNull;
502 bool flat_array = true;
503 bool null_free_array = true;
504 method()->array_access_profiled_type(bci(), speculative_array_type, profiled_element_type, element_ptr, flat_array,
505 null_free_array);
506 }
507 }
508 if (speculative_array_type != nullptr) {
509 // Speculate that this array has the exact type reported by profile data
510 Node* casted_array = nullptr;
511 DEBUG_ONLY(Node* old_control = control();)
512 Node* slow_ctl = type_check_receiver(array, speculative_array_type, 1.0, &casted_array);
513 if (stopped()) {
514 // The check always fails and therefore profile information is incorrect. Don't use it.
515 assert(old_control == slow_ctl, "type check should have been removed");
516 set_control(slow_ctl);
517 } else if (!slow_ctl->is_top()) {
518 { PreserveJVMState pjvms(this);
519 set_control(slow_ctl);
520 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
521 }
522 replace_in_map(array, casted_array);
523 array_type = _gvn.type(casted_array)->is_aryptr();
524 element_type = array_type->elem();
525 return casted_array;
526 }
527 }
528 return array;
529 }
530
531 // Create a CheckCastPP when the speculative type can improve the current type.
532 Node* Parse::cast_to_profiled_array_type(Node* const array) {
533 ciKlass* array_type = nullptr;
534 ciKlass* element_type = nullptr;
535 ProfilePtrKind element_ptr = ProfileMaybeNull;
536 bool flat_array = true;
537 bool null_free_array = true;
538 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
539 if (array_type != nullptr) {
540 return record_profile_for_speculation(array, array_type, ProfileMaybeNull);
541 }
542 return array;
543 }
544
545 // Speculate that the array is non-null-free. We emit a trap when this turns out to be
546 // wrong. On the fast path, we add a CheckCastPP to use the non-null-free type.
547 Node* Parse::speculate_non_null_free_array(Node* const array, const TypeAryPtr*& array_type) {
548 bool null_free_array = true;
549 Deoptimization::DeoptReason reason = Deoptimization::Reason_none;
550 if (array_type->speculative() != nullptr &&
551 array_type->speculative()->is_aryptr()->is_not_null_free() &&
552 !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_class_check)) {
553 null_free_array = false;
554 reason = Deoptimization::Reason_speculate_class_check;
555 } else if (UseArrayLoadStoreProfile && !too_many_traps_or_recompiles(Deoptimization::Reason_class_check)) {
556 ciKlass* profiled_array_type = nullptr;
557 ciKlass* profiled_element_type = nullptr;
558 ProfilePtrKind element_ptr = ProfileMaybeNull;
559 bool flat_array = true;
560 method()->array_access_profiled_type(bci(), profiled_array_type, profiled_element_type, element_ptr, flat_array,
561 null_free_array);
562 reason = Deoptimization::Reason_class_check;
563 }
564 if (!null_free_array) {
565 { // Deoptimize if null-free array
566 BuildCutout unless(this, null_free_array_test(array, /* null_free = */ false), PROB_MAX);
567 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
568 }
569 assert(!stopped(), "null-free array should have been caught earlier");
570 Node* casted_array = _gvn.transform(new CheckCastPPNode(control(), array, array_type->cast_to_not_null_free()));
571 replace_in_map(array, casted_array);
572 array_type = _gvn.type(casted_array)->is_aryptr();
573 return casted_array;
574 }
575 return array;
576 }
577
578 // Speculate that the array is non-flat. We emit a trap when this turns out to be wrong.
579 // On the fast path, we add a CheckCastPP to use the non-flat type.
580 Node* Parse::speculate_non_flat_array(Node* const array, const TypeAryPtr* const array_type) {
581 bool flat_array = true;
582 Deoptimization::DeoptReason reason = Deoptimization::Reason_none;
583 if (array_type->speculative() != nullptr &&
584 array_type->speculative()->is_aryptr()->is_not_flat() &&
585 !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_class_check)) {
586 flat_array = false;
587 reason = Deoptimization::Reason_speculate_class_check;
588 } else if (UseArrayLoadStoreProfile && !too_many_traps_or_recompiles(reason)) {
589 ciKlass* profiled_array_type = nullptr;
590 ciKlass* profiled_element_type = nullptr;
591 ProfilePtrKind element_ptr = ProfileMaybeNull;
592 bool null_free_array = true;
593 method()->array_access_profiled_type(bci(), profiled_array_type, profiled_element_type, element_ptr, flat_array,
594 null_free_array);
595 reason = Deoptimization::Reason_class_check;
596 }
597 if (!flat_array) {
598 { // Deoptimize if flat array
599 BuildCutout unless(this, flat_array_test(array, /* flat = */ false), PROB_MAX);
600 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
601 }
602 assert(!stopped(), "flat array should have been caught earlier");
603 Node* casted_array = _gvn.transform(new CheckCastPPNode(control(), array, array_type->cast_to_not_flat()));
604 replace_in_map(array, casted_array);
605 return casted_array;
606 }
607 return array;
608 }
609
610 // returns IfNode
611 IfNode* Parse::jump_if_fork_int(Node* a, Node* b, BoolTest::mask mask, float prob, float cnt) {
612 Node *cmp = _gvn.transform(new CmpINode(a, b)); // two cases: shiftcount > 32 and shiftcount <= 32
613 Node *tst = _gvn.transform(new BoolNode(cmp, mask));
614 IfNode *iff = create_and_map_if(control(), tst, prob, cnt);
615 return iff;
616 }
617
618
619 // sentinel value for the target bci to mark never taken branches
620 // (according to profiling)
621 static const int never_reached = INT_MAX;
622
623 //------------------------------helper for tableswitch-------------------------
624 void Parse::jump_if_true_fork(IfNode *iff, int dest_bci_if_true, bool unc) {
625 // True branch, use existing map info
626 { PreserveJVMState pjvms(this);
627 Node *iftrue = _gvn.transform( new IfTrueNode (iff) );
628 set_control( iftrue );
629 if (unc) {
630 repush_if_args();
631 uncommon_trap(Deoptimization::Reason_unstable_if,
632 Deoptimization::Action_reinterpret,
633 nullptr,
634 "taken always");
635 } else {
636 assert(dest_bci_if_true != never_reached, "inconsistent dest");
637 merge_new_path(dest_bci_if_true);
638 }
639 }
640
641 // False branch
642 Node *iffalse = _gvn.transform( new IfFalseNode(iff) );
643 set_control( iffalse );
644 }
645
646 void Parse::jump_if_false_fork(IfNode *iff, int dest_bci_if_true, bool unc) {
647 // True branch, use existing map info
648 { PreserveJVMState pjvms(this);
649 Node *iffalse = _gvn.transform( new IfFalseNode (iff) );
650 set_control( iffalse );
651 if (unc) {
652 repush_if_args();
653 uncommon_trap(Deoptimization::Reason_unstable_if,
654 Deoptimization::Action_reinterpret,
655 nullptr,
656 "taken never");
657 } else {
658 assert(dest_bci_if_true != never_reached, "inconsistent dest");
659 merge_new_path(dest_bci_if_true);
660 }
661 }
662
663 // False branch
664 Node *iftrue = _gvn.transform( new IfTrueNode(iff) );
665 set_control( iftrue );
666 }
667
668 void Parse::jump_if_always_fork(int dest_bci, bool unc) {
669 // False branch, use existing map and control()
670 if (unc) {
671 repush_if_args();
672 uncommon_trap(Deoptimization::Reason_unstable_if,
673 Deoptimization::Action_reinterpret,
674 nullptr,
675 "taken never");
676 } else {
677 assert(dest_bci != never_reached, "inconsistent dest");
678 merge_new_path(dest_bci);
679 }
680 }
681
682
683 extern "C" {
684 static int jint_cmp(const void *i, const void *j) {
685 int a = *(jint *)i;
686 int b = *(jint *)j;
687 return a > b ? 1 : a < b ? -1 : 0;
688 }
689 }
690
691
692 class SwitchRange : public StackObj {
693 // a range of integers coupled with a bci destination
694 jint _lo; // inclusive lower limit
695 jint _hi; // inclusive upper limit
696 int _dest;
697 float _cnt; // how many times this range was hit according to profiling
698
699 public:
700 jint lo() const { return _lo; }
701 jint hi() const { return _hi; }
702 int dest() const { return _dest; }
703 bool is_singleton() const { return _lo == _hi; }
704 float cnt() const { return _cnt; }
705
706 void setRange(jint lo, jint hi, int dest, float cnt) {
707 assert(lo <= hi, "must be a non-empty range");
708 _lo = lo, _hi = hi; _dest = dest; _cnt = cnt;
709 assert(_cnt >= 0, "");
710 }
711 bool adjoinRange(jint lo, jint hi, int dest, float cnt, bool trim_ranges) {
712 assert(lo <= hi, "must be a non-empty range");
713 if (lo == _hi+1) {
714 // see merge_ranges() comment below
715 if (trim_ranges) {
716 if (cnt == 0) {
717 if (_cnt != 0) {
718 return false;
719 }
720 if (dest != _dest) {
721 _dest = never_reached;
722 }
723 } else {
724 if (_cnt == 0) {
725 return false;
726 }
727 if (dest != _dest) {
728 return false;
729 }
730 }
731 } else {
732 if (dest != _dest) {
733 return false;
734 }
735 }
736 _hi = hi;
737 _cnt += cnt;
738 return true;
739 }
740 return false;
741 }
742
743 void set (jint value, int dest, float cnt) {
744 setRange(value, value, dest, cnt);
745 }
746 bool adjoin(jint value, int dest, float cnt, bool trim_ranges) {
747 return adjoinRange(value, value, dest, cnt, trim_ranges);
748 }
749 bool adjoin(SwitchRange& other) {
750 return adjoinRange(other._lo, other._hi, other._dest, other._cnt, false);
751 }
752
753 void print() {
754 if (is_singleton())
755 tty->print(" {%d}=>%d (cnt=%f)", lo(), dest(), cnt());
756 else if (lo() == min_jint)
757 tty->print(" {..%d}=>%d (cnt=%f)", hi(), dest(), cnt());
758 else if (hi() == max_jint)
759 tty->print(" {%d..}=>%d (cnt=%f)", lo(), dest(), cnt());
760 else
761 tty->print(" {%d..%d}=>%d (cnt=%f)", lo(), hi(), dest(), cnt());
762 }
763 };
764
765 // We try to minimize the number of ranges and the size of the taken
766 // ones using profiling data. When ranges are created,
767 // SwitchRange::adjoinRange() only allows 2 adjoining ranges to merge
768 // if both were never hit or both were hit to build longer unreached
769 // ranges. Here, we now merge adjoining ranges with the same
770 // destination and finally set destination of unreached ranges to the
771 // special value never_reached because it can help minimize the number
772 // of tests that are necessary.
773 //
774 // For instance:
775 // [0, 1] to target1 sometimes taken
776 // [1, 2] to target1 never taken
777 // [2, 3] to target2 never taken
778 // would lead to:
779 // [0, 1] to target1 sometimes taken
780 // [1, 3] never taken
781 //
782 // (first 2 ranges to target1 are not merged)
783 static void merge_ranges(SwitchRange* ranges, int& rp) {
784 if (rp == 0) {
785 return;
786 }
787 int shift = 0;
788 for (int j = 0; j < rp; j++) {
789 SwitchRange& r1 = ranges[j-shift];
790 SwitchRange& r2 = ranges[j+1];
791 if (r1.adjoin(r2)) {
792 shift++;
793 } else if (shift > 0) {
794 ranges[j+1-shift] = r2;
795 }
796 }
797 rp -= shift;
798 for (int j = 0; j <= rp; j++) {
799 SwitchRange& r = ranges[j];
800 if (r.cnt() == 0 && r.dest() != never_reached) {
801 r.setRange(r.lo(), r.hi(), never_reached, r.cnt());
802 }
803 }
804 }
805
806 //-------------------------------do_tableswitch--------------------------------
807 void Parse::do_tableswitch() {
808 // Get information about tableswitch
809 int default_dest = iter().get_dest_table(0);
810 jint lo_index = iter().get_int_table(1);
811 jint hi_index = iter().get_int_table(2);
812 int len = hi_index - lo_index + 1;
813
814 if (len < 1) {
815 // If this is a backward branch, add safepoint
816 maybe_add_safepoint(default_dest);
817 pop(); // the effect of the instruction execution on the operand stack
818 merge(default_dest);
819 return;
820 }
821
822 ciMethodData* methodData = method()->method_data();
823 ciMultiBranchData* profile = nullptr;
824 if (methodData->is_mature() && UseSwitchProfiling) {
825 ciProfileData* data = methodData->bci_to_data(bci());
826 if (data != nullptr && data->is_MultiBranchData()) {
827 profile = (ciMultiBranchData*)data;
828 }
829 }
830 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if);
831
832 // generate decision tree, using trichotomy when possible
833 int rnum = len+2;
834 bool makes_backward_branch = (default_dest <= bci());
835 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum);
836 int rp = -1;
837 if (lo_index != min_jint) {
838 float cnt = 1.0F;
839 if (profile != nullptr) {
840 cnt = (float)profile->default_count() / (hi_index != max_jint ? 2.0F : 1.0F);
841 }
842 ranges[++rp].setRange(min_jint, lo_index-1, default_dest, cnt);
843 }
844 for (int j = 0; j < len; j++) {
845 jint match_int = lo_index+j;
846 int dest = iter().get_dest_table(j+3);
847 makes_backward_branch |= (dest <= bci());
848 float cnt = 1.0F;
849 if (profile != nullptr) {
850 cnt = (float)profile->count_at(j);
851 }
852 if (rp < 0 || !ranges[rp].adjoin(match_int, dest, cnt, trim_ranges)) {
853 ranges[++rp].set(match_int, dest, cnt);
854 }
855 }
856 jint highest = lo_index+(len-1);
857 assert(ranges[rp].hi() == highest, "");
858 if (highest != max_jint) {
859 float cnt = 1.0F;
860 if (profile != nullptr) {
861 cnt = (float)profile->default_count() / (lo_index != min_jint ? 2.0F : 1.0F);
862 }
863 if (!ranges[rp].adjoinRange(highest+1, max_jint, default_dest, cnt, trim_ranges)) {
864 ranges[++rp].setRange(highest+1, max_jint, default_dest, cnt);
865 }
866 }
867 assert(rp < len+2, "not too many ranges");
868
869 if (trim_ranges) {
870 merge_ranges(ranges, rp);
871 }
872
873 // Safepoint in case if backward branch observed
874 if (makes_backward_branch) {
875 add_safepoint();
876 }
877
878 Node* lookup = pop(); // lookup value
879 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]);
880 }
881
882
883 //------------------------------do_lookupswitch--------------------------------
884 void Parse::do_lookupswitch() {
885 // Get information about lookupswitch
886 int default_dest = iter().get_dest_table(0);
887 jint len = iter().get_int_table(1);
888
889 if (len < 1) { // If this is a backward branch, add safepoint
890 maybe_add_safepoint(default_dest);
891 pop(); // the effect of the instruction execution on the operand stack
892 merge(default_dest);
893 return;
894 }
895
896 ciMethodData* methodData = method()->method_data();
897 ciMultiBranchData* profile = nullptr;
898 if (methodData->is_mature() && UseSwitchProfiling) {
899 ciProfileData* data = methodData->bci_to_data(bci());
900 if (data != nullptr && data->is_MultiBranchData()) {
901 profile = (ciMultiBranchData*)data;
902 }
903 }
904 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if);
905
906 // generate decision tree, using trichotomy when possible
907 jint* table = NEW_RESOURCE_ARRAY(jint, len*3);
908 {
909 for (int j = 0; j < len; j++) {
910 table[3*j+0] = iter().get_int_table(2+2*j);
911 table[3*j+1] = iter().get_dest_table(2+2*j+1);
912 // Handle overflow when converting from uint to jint
913 table[3*j+2] = (profile == nullptr) ? 1 : (jint)MIN2<uint>((uint)max_jint, profile->count_at(j));
914 }
915 qsort(table, len, 3*sizeof(table[0]), jint_cmp);
916 }
917
918 float default_cnt = 1.0F;
919 if (profile != nullptr) {
920 juint defaults = max_juint - len;
921 default_cnt = (float)profile->default_count()/(float)defaults;
922 }
923
924 int rnum = len*2+1;
925 bool makes_backward_branch = (default_dest <= bci());
926 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum);
927 int rp = -1;
928 for (int j = 0; j < len; j++) {
929 jint match_int = table[3*j+0];
930 jint dest = table[3*j+1];
931 jint cnt = table[3*j+2];
932 jint next_lo = rp < 0 ? min_jint : ranges[rp].hi()+1;
933 makes_backward_branch |= (dest <= bci());
934 float c = default_cnt * ((float)match_int - (float)next_lo);
935 if (match_int != next_lo && (rp < 0 || !ranges[rp].adjoinRange(next_lo, match_int-1, default_dest, c, trim_ranges))) {
936 assert(default_dest != never_reached, "sentinel value for dead destinations");
937 ranges[++rp].setRange(next_lo, match_int-1, default_dest, c);
938 }
939 if (rp < 0 || !ranges[rp].adjoin(match_int, dest, (float)cnt, trim_ranges)) {
940 assert(dest != never_reached, "sentinel value for dead destinations");
941 ranges[++rp].set(match_int, dest, (float)cnt);
942 }
943 }
944 jint highest = table[3*(len-1)];
945 assert(ranges[rp].hi() == highest, "");
946 if (highest != max_jint &&
947 !ranges[rp].adjoinRange(highest+1, max_jint, default_dest, default_cnt * ((float)max_jint - (float)highest), trim_ranges)) {
948 ranges[++rp].setRange(highest+1, max_jint, default_dest, default_cnt * ((float)max_jint - (float)highest));
949 }
950 assert(rp < rnum, "not too many ranges");
951
952 if (trim_ranges) {
953 merge_ranges(ranges, rp);
954 }
955
956 // Safepoint in case backward branch observed
957 if (makes_backward_branch) {
958 add_safepoint();
959 }
960
961 Node *lookup = pop(); // lookup value
962 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]);
963 }
964
965 static float if_prob(float taken_cnt, float total_cnt) {
966 assert(taken_cnt <= total_cnt, "");
967 if (total_cnt == 0) {
968 return PROB_FAIR;
969 }
970 float p = taken_cnt / total_cnt;
971 return clamp(p, PROB_MIN, PROB_MAX);
972 }
973
974 static float if_cnt(float cnt) {
975 if (cnt == 0) {
976 return COUNT_UNKNOWN;
977 }
978 return cnt;
979 }
980
981 static float sum_of_cnts(SwitchRange *lo, SwitchRange *hi) {
982 float total_cnt = 0;
983 for (SwitchRange* sr = lo; sr <= hi; sr++) {
984 total_cnt += sr->cnt();
985 }
986 return total_cnt;
987 }
988
989 class SwitchRanges : public ResourceObj {
990 public:
991 SwitchRange* _lo;
992 SwitchRange* _hi;
993 SwitchRange* _mid;
994 float _cost;
995
996 enum {
997 Start,
998 LeftDone,
999 RightDone,
1000 Done
1001 } _state;
1002
1003 SwitchRanges(SwitchRange *lo, SwitchRange *hi)
1004 : _lo(lo), _hi(hi), _mid(nullptr),
1005 _cost(0), _state(Start) {
1006 }
1007
1008 SwitchRanges()
1009 : _lo(nullptr), _hi(nullptr), _mid(nullptr),
1010 _cost(0), _state(Start) {}
1011 };
1012
1013 // Estimate cost of performing a binary search on lo..hi
1014 static float compute_tree_cost(SwitchRange *lo, SwitchRange *hi, float total_cnt) {
1015 GrowableArray<SwitchRanges> tree;
1016 SwitchRanges root(lo, hi);
1017 tree.push(root);
1018
1019 float cost = 0;
1020 do {
1021 SwitchRanges& r = *tree.adr_at(tree.length()-1);
1022 if (r._hi != r._lo) {
1023 if (r._mid == nullptr) {
1024 float r_cnt = sum_of_cnts(r._lo, r._hi);
1025
1026 if (r_cnt == 0) {
1027 tree.pop();
1028 cost = 0;
1029 continue;
1030 }
1031
1032 SwitchRange* mid = nullptr;
1033 mid = r._lo;
1034 for (float cnt = 0; ; ) {
1035 assert(mid <= r._hi, "out of bounds");
1036 cnt += mid->cnt();
1037 if (cnt > r_cnt / 2) {
1038 break;
1039 }
1040 mid++;
1041 }
1042 assert(mid <= r._hi, "out of bounds");
1043 r._mid = mid;
1044 r._cost = r_cnt / total_cnt;
1045 }
1046 r._cost += cost;
1047 if (r._state < SwitchRanges::LeftDone && r._mid > r._lo) {
1048 cost = 0;
1049 r._state = SwitchRanges::LeftDone;
1050 tree.push(SwitchRanges(r._lo, r._mid-1));
1051 } else if (r._state < SwitchRanges::RightDone) {
1052 cost = 0;
1053 r._state = SwitchRanges::RightDone;
1054 tree.push(SwitchRanges(r._mid == r._lo ? r._mid+1 : r._mid, r._hi));
1055 } else {
1056 tree.pop();
1057 cost = r._cost;
1058 }
1059 } else {
1060 tree.pop();
1061 cost = r._cost;
1062 }
1063 } while (tree.length() > 0);
1064
1065
1066 return cost;
1067 }
1068
1069 // It sometimes pays off to test most common ranges before the binary search
1070 void Parse::linear_search_switch_ranges(Node* key_val, SwitchRange*& lo, SwitchRange*& hi) {
1071 uint nr = hi - lo + 1;
1072 float total_cnt = sum_of_cnts(lo, hi);
1073
1074 float min = compute_tree_cost(lo, hi, total_cnt);
1075 float extra = 1;
1076 float sub = 0;
1077
1078 SwitchRange* array1 = lo;
1079 SwitchRange* array2 = NEW_RESOURCE_ARRAY(SwitchRange, nr);
1080
1081 SwitchRange* ranges = nullptr;
1082
1083 while (nr >= 2) {
1084 assert(lo == array1 || lo == array2, "one the 2 already allocated arrays");
1085 ranges = (lo == array1) ? array2 : array1;
1086
1087 // Find highest frequency range
1088 SwitchRange* candidate = lo;
1089 for (SwitchRange* sr = lo+1; sr <= hi; sr++) {
1090 if (sr->cnt() > candidate->cnt()) {
1091 candidate = sr;
1092 }
1093 }
1094 SwitchRange most_freq = *candidate;
1095 if (most_freq.cnt() == 0) {
1096 break;
1097 }
1098
1099 // Copy remaining ranges into another array
1100 int shift = 0;
1101 for (uint i = 0; i < nr; i++) {
1102 SwitchRange* sr = &lo[i];
1103 if (sr != candidate) {
1104 ranges[i-shift] = *sr;
1105 } else {
1106 shift++;
1107 if (i > 0 && i < nr-1) {
1108 SwitchRange prev = lo[i-1];
1109 prev.setRange(prev.lo(), sr->hi(), prev.dest(), prev.cnt());
1110 if (prev.adjoin(lo[i+1])) {
1111 shift++;
1112 i++;
1113 }
1114 ranges[i-shift] = prev;
1115 }
1116 }
1117 }
1118 nr -= shift;
1119
1120 // Evaluate cost of testing the most common range and performing a
1121 // binary search on the other ranges
1122 float cost = extra + compute_tree_cost(&ranges[0], &ranges[nr-1], total_cnt);
1123 if (cost >= min) {
1124 break;
1125 }
1126 // swap arrays
1127 lo = &ranges[0];
1128 hi = &ranges[nr-1];
1129
1130 // It pays off: emit the test for the most common range
1131 assert(most_freq.cnt() > 0, "must be taken");
1132 Node* val = _gvn.transform(new SubINode(key_val, _gvn.intcon(most_freq.lo())));
1133 Node* cmp = _gvn.transform(new CmpUNode(val, _gvn.intcon(java_subtract(most_freq.hi(), most_freq.lo()))));
1134 Node* tst = _gvn.transform(new BoolNode(cmp, BoolTest::le));
1135 IfNode* iff = create_and_map_if(control(), tst, if_prob(most_freq.cnt(), total_cnt), if_cnt(most_freq.cnt()));
1136 jump_if_true_fork(iff, most_freq.dest(), false);
1137
1138 sub += most_freq.cnt() / total_cnt;
1139 extra += 1 - sub;
1140 min = cost;
1141 }
1142 }
1143
1144 //----------------------------create_jump_tables-------------------------------
1145 bool Parse::create_jump_tables(Node* key_val, SwitchRange* lo, SwitchRange* hi) {
1146 // Are jumptables enabled
1147 if (!UseJumpTables) return false;
1148
1149 // Are jumptables supported
1150 if (!Matcher::has_match_rule(Op_Jump)) return false;
1151
1152 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if);
1153
1154 // Decide if a guard is needed to lop off big ranges at either (or
1155 // both) end(s) of the input set. We'll call this the default target
1156 // even though we can't be sure that it is the true "default".
1157
1158 bool needs_guard = false;
1159 int default_dest;
1160 int64_t total_outlier_size = 0;
1161 int64_t hi_size = ((int64_t)hi->hi()) - ((int64_t)hi->lo()) + 1;
1162 int64_t lo_size = ((int64_t)lo->hi()) - ((int64_t)lo->lo()) + 1;
1163
1164 if (lo->dest() == hi->dest()) {
1165 total_outlier_size = hi_size + lo_size;
1166 default_dest = lo->dest();
1167 } else if (lo_size > hi_size) {
1168 total_outlier_size = lo_size;
1169 default_dest = lo->dest();
1170 } else {
1171 total_outlier_size = hi_size;
1172 default_dest = hi->dest();
1173 }
1174
1175 float total = sum_of_cnts(lo, hi);
1176 float cost = compute_tree_cost(lo, hi, total);
1177
1178 // If a guard test will eliminate very sparse end ranges, then
1179 // it is worth the cost of an extra jump.
1180 float trimmed_cnt = 0;
1181 if (total_outlier_size > (MaxJumpTableSparseness * 4)) {
1182 needs_guard = true;
1183 if (default_dest == lo->dest()) {
1184 trimmed_cnt += lo->cnt();
1185 lo++;
1186 }
1187 if (default_dest == hi->dest()) {
1188 trimmed_cnt += hi->cnt();
1189 hi--;
1190 }
1191 }
1192
1193 // Find the total number of cases and ranges
1194 int64_t num_cases = ((int64_t)hi->hi()) - ((int64_t)lo->lo()) + 1;
1195 int num_range = hi - lo + 1;
1196
1197 // Don't create table if: too large, too small, or too sparse.
1198 if (num_cases > MaxJumpTableSize)
1199 return false;
1200 if (UseSwitchProfiling) {
1201 // MinJumpTableSize is set so with a well balanced binary tree,
1202 // when the number of ranges is MinJumpTableSize, it's cheaper to
1203 // go through a JumpNode that a tree of IfNodes. Average cost of a
1204 // tree of IfNodes with MinJumpTableSize is
1205 // log2f(MinJumpTableSize) comparisons. So if the cost computed
1206 // from profile data is less than log2f(MinJumpTableSize) then
1207 // going with the binary search is cheaper.
1208 if (cost < log2f(MinJumpTableSize)) {
1209 return false;
1210 }
1211 } else {
1212 if (num_cases < MinJumpTableSize)
1213 return false;
1214 }
1215 if (num_cases > (MaxJumpTableSparseness * num_range))
1216 return false;
1217
1218 // Normalize table lookups to zero
1219 int lowval = lo->lo();
1220 key_val = _gvn.transform( new SubINode(key_val, _gvn.intcon(lowval)) );
1221
1222 // Generate a guard to protect against input keyvals that aren't
1223 // in the switch domain.
1224 if (needs_guard) {
1225 Node* size = _gvn.intcon(num_cases);
1226 Node* cmp = _gvn.transform(new CmpUNode(key_val, size));
1227 Node* tst = _gvn.transform(new BoolNode(cmp, BoolTest::ge));
1228 IfNode* iff = create_and_map_if(control(), tst, if_prob(trimmed_cnt, total), if_cnt(trimmed_cnt));
1229 jump_if_true_fork(iff, default_dest, trim_ranges && trimmed_cnt == 0);
1230
1231 total -= trimmed_cnt;
1232 }
1233
1234 // Create an ideal node JumpTable that has projections
1235 // of all possible ranges for a switch statement
1236 // The key_val input must be converted to a pointer offset and scaled.
1237 // Compare Parse::array_addressing above.
1238
1239 // Clean the 32-bit int into a real 64-bit offset.
1240 // Otherwise, the jint value 0 might turn into an offset of 0x0800000000.
1241 // Make I2L conversion control dependent to prevent it from
1242 // floating above the range check during loop optimizations.
1243 // Do not use a narrow int type here to prevent the data path from dying
1244 // while the control path is not removed. This can happen if the type of key_val
1245 // is later known to be out of bounds of [0, num_cases] and therefore a narrow cast
1246 // would be replaced by TOP while C2 is not able to fold the corresponding range checks.
1247 // Set _carry_dependency for the cast to avoid being removed by IGVN.
1248 #ifdef _LP64
1249 key_val = C->constrained_convI2L(&_gvn, key_val, TypeInt::INT, control(), true /* carry_dependency */);
1250 #endif
1251
1252 // Shift the value by wordsize so we have an index into the table, rather
1253 // than a switch value
1254 Node *shiftWord = _gvn.MakeConX(wordSize);
1255 key_val = _gvn.transform( new MulXNode( key_val, shiftWord));
1256
1257 // Create the JumpNode
1258 Arena* arena = C->comp_arena();
1259 float* probs = (float*)arena->Amalloc(sizeof(float)*num_cases);
1260 int i = 0;
1261 if (total == 0) {
1262 for (SwitchRange* r = lo; r <= hi; r++) {
1263 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) {
1264 probs[i] = 1.0F / num_cases;
1265 }
1266 }
1267 } else {
1268 for (SwitchRange* r = lo; r <= hi; r++) {
1269 float prob = r->cnt()/total;
1270 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) {
1271 probs[i] = prob / (r->hi() - r->lo() + 1);
1272 }
1273 }
1274 }
1275
1276 ciMethodData* methodData = method()->method_data();
1277 ciMultiBranchData* profile = nullptr;
1278 if (methodData->is_mature()) {
1279 ciProfileData* data = methodData->bci_to_data(bci());
1280 if (data != nullptr && data->is_MultiBranchData()) {
1281 profile = (ciMultiBranchData*)data;
1282 }
1283 }
1284
1285 Node* jtn = _gvn.transform(new JumpNode(control(), key_val, num_cases, probs, profile == nullptr ? COUNT_UNKNOWN : total));
1286
1287 // These are the switch destinations hanging off the jumpnode
1288 i = 0;
1289 for (SwitchRange* r = lo; r <= hi; r++) {
1290 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) {
1291 Node* input = _gvn.transform(new JumpProjNode(jtn, i, r->dest(), (int)(j - lowval)));
1292 {
1293 PreserveJVMState pjvms(this);
1294 set_control(input);
1295 jump_if_always_fork(r->dest(), trim_ranges && r->cnt() == 0);
1296 }
1297 }
1298 }
1299 assert(i == num_cases, "miscount of cases");
1300 stop_and_kill_map(); // no more uses for this JVMS
1301 return true;
1302 }
1303
1304 //----------------------------jump_switch_ranges-------------------------------
1305 void Parse::jump_switch_ranges(Node* key_val, SwitchRange *lo, SwitchRange *hi, int switch_depth) {
1306 Block* switch_block = block();
1307 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if);
1308
1309 if (switch_depth == 0) {
1310 // Do special processing for the top-level call.
1311 assert(lo->lo() == min_jint, "initial range must exhaust Type::INT");
1312 assert(hi->hi() == max_jint, "initial range must exhaust Type::INT");
1313
1314 // Decrement pred-numbers for the unique set of nodes.
1315 #ifdef ASSERT
1316 if (!trim_ranges) {
1317 // Ensure that the block's successors are a (duplicate-free) set.
1318 int successors_counted = 0; // block occurrences in [hi..lo]
1319 int unique_successors = switch_block->num_successors();
1320 for (int i = 0; i < unique_successors; i++) {
1321 Block* target = switch_block->successor_at(i);
1322
1323 // Check that the set of successors is the same in both places.
1324 int successors_found = 0;
1325 for (SwitchRange* p = lo; p <= hi; p++) {
1326 if (p->dest() == target->start()) successors_found++;
1327 }
1328 assert(successors_found > 0, "successor must be known");
1329 successors_counted += successors_found;
1330 }
1331 assert(successors_counted == (hi-lo)+1, "no unexpected successors");
1332 }
1333 #endif
1334
1335 // Maybe prune the inputs, based on the type of key_val.
1336 jint min_val = min_jint;
1337 jint max_val = max_jint;
1338 const TypeInt* ti = key_val->bottom_type()->isa_int();
1339 if (ti != nullptr) {
1340 min_val = ti->_lo;
1341 max_val = ti->_hi;
1342 assert(min_val <= max_val, "invalid int type");
1343 }
1344 while (lo->hi() < min_val) {
1345 lo++;
1346 }
1347 if (lo->lo() < min_val) {
1348 lo->setRange(min_val, lo->hi(), lo->dest(), lo->cnt());
1349 }
1350 while (hi->lo() > max_val) {
1351 hi--;
1352 }
1353 if (hi->hi() > max_val) {
1354 hi->setRange(hi->lo(), max_val, hi->dest(), hi->cnt());
1355 }
1356
1357 linear_search_switch_ranges(key_val, lo, hi);
1358 }
1359
1360 #ifndef PRODUCT
1361 if (switch_depth == 0) {
1362 _max_switch_depth = 0;
1363 _est_switch_depth = log2i_graceful((hi - lo + 1) - 1) + 1;
1364 }
1365 #endif
1366
1367 assert(lo <= hi, "must be a non-empty set of ranges");
1368 if (lo == hi) {
1369 jump_if_always_fork(lo->dest(), trim_ranges && lo->cnt() == 0);
1370 } else {
1371 assert(lo->hi() == (lo+1)->lo()-1, "contiguous ranges");
1372 assert(hi->lo() == (hi-1)->hi()+1, "contiguous ranges");
1373
1374 if (create_jump_tables(key_val, lo, hi)) return;
1375
1376 SwitchRange* mid = nullptr;
1377 float total_cnt = sum_of_cnts(lo, hi);
1378
1379 int nr = hi - lo + 1;
1380 if (UseSwitchProfiling) {
1381 // Don't keep the binary search tree balanced: pick up mid point
1382 // that split frequencies in half.
1383 float cnt = 0;
1384 for (SwitchRange* sr = lo; sr <= hi; sr++) {
1385 cnt += sr->cnt();
1386 if (cnt >= total_cnt / 2) {
1387 mid = sr;
1388 break;
1389 }
1390 }
1391 } else {
1392 mid = lo + nr/2;
1393
1394 // if there is an easy choice, pivot at a singleton:
1395 if (nr > 3 && !mid->is_singleton() && (mid-1)->is_singleton()) mid--;
1396
1397 assert(lo < mid && mid <= hi, "good pivot choice");
1398 assert(nr != 2 || mid == hi, "should pick higher of 2");
1399 assert(nr != 3 || mid == hi-1, "should pick middle of 3");
1400 }
1401
1402
1403 Node *test_val = _gvn.intcon(mid == lo ? mid->hi() : mid->lo());
1404
1405 if (mid->is_singleton()) {
1406 IfNode *iff_ne = jump_if_fork_int(key_val, test_val, BoolTest::ne, 1-if_prob(mid->cnt(), total_cnt), if_cnt(mid->cnt()));
1407 jump_if_false_fork(iff_ne, mid->dest(), trim_ranges && mid->cnt() == 0);
1408
1409 // Special Case: If there are exactly three ranges, and the high
1410 // and low range each go to the same place, omit the "gt" test,
1411 // since it will not discriminate anything.
1412 bool eq_test_only = (hi == lo+2 && hi->dest() == lo->dest() && mid == hi-1) || mid == lo;
1413
1414 // if there is a higher range, test for it and process it:
1415 if (mid < hi && !eq_test_only) {
1416 // two comparisons of same values--should enable 1 test for 2 branches
1417 // Use BoolTest::lt instead of BoolTest::gt
1418 float cnt = sum_of_cnts(lo, mid-1);
1419 IfNode *iff_lt = jump_if_fork_int(key_val, test_val, BoolTest::lt, if_prob(cnt, total_cnt), if_cnt(cnt));
1420 Node *iftrue = _gvn.transform( new IfTrueNode(iff_lt) );
1421 Node *iffalse = _gvn.transform( new IfFalseNode(iff_lt) );
1422 { PreserveJVMState pjvms(this);
1423 set_control(iffalse);
1424 jump_switch_ranges(key_val, mid+1, hi, switch_depth+1);
1425 }
1426 set_control(iftrue);
1427 }
1428
1429 } else {
1430 // mid is a range, not a singleton, so treat mid..hi as a unit
1431 float cnt = sum_of_cnts(mid == lo ? mid+1 : mid, hi);
1432 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));
1433
1434 // if there is a higher range, test for it and process it:
1435 if (mid == hi) {
1436 jump_if_true_fork(iff_ge, mid->dest(), trim_ranges && cnt == 0);
1437 } else {
1438 Node *iftrue = _gvn.transform( new IfTrueNode(iff_ge) );
1439 Node *iffalse = _gvn.transform( new IfFalseNode(iff_ge) );
1440 { PreserveJVMState pjvms(this);
1441 set_control(iftrue);
1442 jump_switch_ranges(key_val, mid == lo ? mid+1 : mid, hi, switch_depth+1);
1443 }
1444 set_control(iffalse);
1445 }
1446 }
1447
1448 // in any case, process the lower range
1449 if (mid == lo) {
1450 if (mid->is_singleton()) {
1451 jump_switch_ranges(key_val, lo+1, hi, switch_depth+1);
1452 } else {
1453 jump_if_always_fork(lo->dest(), trim_ranges && lo->cnt() == 0);
1454 }
1455 } else {
1456 jump_switch_ranges(key_val, lo, mid-1, switch_depth+1);
1457 }
1458 }
1459
1460 // Decrease pred_count for each successor after all is done.
1461 if (switch_depth == 0) {
1462 int unique_successors = switch_block->num_successors();
1463 for (int i = 0; i < unique_successors; i++) {
1464 Block* target = switch_block->successor_at(i);
1465 // Throw away the pre-allocated path for each unique successor.
1466 target->next_path_num();
1467 }
1468 }
1469
1470 #ifndef PRODUCT
1471 _max_switch_depth = MAX2(switch_depth, _max_switch_depth);
1472 if (TraceOptoParse && Verbose && WizardMode && switch_depth == 0) {
1473 SwitchRange* r;
1474 int nsing = 0;
1475 for( r = lo; r <= hi; r++ ) {
1476 if( r->is_singleton() ) nsing++;
1477 }
1478 tty->print(">>> ");
1479 _method->print_short_name();
1480 tty->print_cr(" switch decision tree");
1481 tty->print_cr(" %d ranges (%d singletons), max_depth=%d, est_depth=%d",
1482 (int) (hi-lo+1), nsing, _max_switch_depth, _est_switch_depth);
1483 if (_max_switch_depth > _est_switch_depth) {
1484 tty->print_cr("******** BAD SWITCH DEPTH ********");
1485 }
1486 tty->print(" ");
1487 for( r = lo; r <= hi; r++ ) {
1488 r->print();
1489 }
1490 tty->cr();
1491 }
1492 #endif
1493 }
1494
1495 Node* Parse::floating_point_mod(Node* a, Node* b, BasicType type) {
1496 assert(type == BasicType::T_FLOAT || type == BasicType::T_DOUBLE, "only float and double are floating points");
1497 CallNode* mod = type == BasicType::T_DOUBLE ? static_cast<CallNode*>(new ModDNode(C, a, b)) : new ModFNode(C, a, b);
1498
1499 Node* prev_mem = set_predefined_input_for_runtime_call(mod);
1500 mod = _gvn.transform(mod)->as_Call();
1501 set_predefined_output_for_runtime_call(mod, prev_mem, TypeRawPtr::BOTTOM);
1502 Node* result = _gvn.transform(new ProjNode(mod, TypeFunc::Parms + 0));
1503 record_for_igvn(mod);
1504 return result;
1505 }
1506
1507 void Parse::l2f() {
1508 Node* f2 = pop();
1509 Node* f1 = pop();
1510 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::l2f_Type(),
1511 CAST_FROM_FN_PTR(address, SharedRuntime::l2f),
1512 "l2f", nullptr, //no memory effects
1513 f1, f2);
1514 Node* res = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0));
1515
1516 push(res);
1517 }
1518
1519 // Handle jsr and jsr_w bytecode
1520 void Parse::do_jsr() {
1521 assert(bc() == Bytecodes::_jsr || bc() == Bytecodes::_jsr_w, "wrong bytecode");
1522
1523 // Store information about current state, tagged with new _jsr_bci
1524 int return_bci = iter().next_bci();
1525 int jsr_bci = (bc() == Bytecodes::_jsr) ? iter().get_dest() : iter().get_far_dest();
1526
1527 // The way we do things now, there is only one successor block
1528 // for the jsr, because the target code is cloned by ciTypeFlow.
1529 Block* target = successor_for_bci(jsr_bci);
1530
1531 // What got pushed?
1532 const Type* ret_addr = target->peek();
1533 assert(ret_addr->singleton(), "must be a constant (cloned jsr body)");
1534
1535 // Effect on jsr on stack
1536 push(_gvn.makecon(ret_addr));
1537
1538 // Flow to the jsr.
1539 merge(jsr_bci);
1540 }
1541
1542 // Handle ret bytecode
1543 void Parse::do_ret() {
1544 // Find to whom we return.
1545 assert(block()->num_successors() == 1, "a ret can only go one place now");
1546 Block* target = block()->successor_at(0);
1547 assert(!target->is_ready(), "our arrival must be expected");
1548 int pnum = target->next_path_num();
1549 merge_common(target, pnum);
1550 }
1551
1552 static bool has_injected_profile(BoolTest::mask btest, Node* test, int& taken, int& not_taken) {
1553 if (btest != BoolTest::eq && btest != BoolTest::ne) {
1554 // Only ::eq and ::ne are supported for profile injection.
1555 return false;
1556 }
1557 if (test->is_Cmp() &&
1558 test->in(1)->Opcode() == Op_ProfileBoolean) {
1559 ProfileBooleanNode* profile = (ProfileBooleanNode*)test->in(1);
1560 int false_cnt = profile->false_count();
1561 int true_cnt = profile->true_count();
1562
1563 // Counts matching depends on the actual test operation (::eq or ::ne).
1564 // No need to scale the counts because profile injection was designed
1565 // to feed exact counts into VM.
1566 taken = (btest == BoolTest::eq) ? false_cnt : true_cnt;
1567 not_taken = (btest == BoolTest::eq) ? true_cnt : false_cnt;
1568
1569 profile->consume();
1570 return true;
1571 }
1572 return false;
1573 }
1574
1575 // Give up if too few (or too many, in which case the sum will overflow) counts to be meaningful.
1576 // We also check that individual counters are positive first, otherwise the sum can become positive.
1577 // (check for saturation, integer overflow, and immature counts)
1578 static bool counters_are_meaningful(int counter1, int counter2, int min) {
1579 // check for saturation, including "uint" values too big to fit in "int"
1580 if (counter1 < 0 || counter2 < 0) {
1581 return false;
1582 }
1583 // check for integer overflow of the sum
1584 int64_t sum = (int64_t)counter1 + (int64_t)counter2;
1585 STATIC_ASSERT(sizeof(counter1) < sizeof(sum));
1586 if (sum > INT_MAX) {
1587 return false;
1588 }
1589 // check if mature
1590 return (counter1 + counter2) >= min;
1591 }
1592
1593 //--------------------------dynamic_branch_prediction--------------------------
1594 // Try to gather dynamic branch prediction behavior. Return a probability
1595 // of the branch being taken and set the "cnt" field. Returns a -1.0
1596 // if we need to use static prediction for some reason.
1597 float Parse::dynamic_branch_prediction(float &cnt, BoolTest::mask btest, Node* test) {
1598 ResourceMark rm;
1599
1600 cnt = COUNT_UNKNOWN;
1601
1602 int taken = 0;
1603 int not_taken = 0;
1604
1605 bool use_mdo = !has_injected_profile(btest, test, taken, not_taken);
1606
1607 if (use_mdo) {
1608 // Use MethodData information if it is available
1609 // FIXME: free the ProfileData structure
1610 ciMethodData* methodData = method()->method_data();
1611 if (!methodData->is_mature()) return PROB_UNKNOWN;
1612 ciProfileData* data = methodData->bci_to_data(bci());
1613 if (data == nullptr) {
1614 return PROB_UNKNOWN;
1615 }
1616 if (!data->is_JumpData()) return PROB_UNKNOWN;
1617
1618 // get taken and not taken values
1619 // NOTE: saturated UINT_MAX values become negative,
1620 // as do counts above INT_MAX.
1621 taken = data->as_JumpData()->taken();
1622 not_taken = 0;
1623 if (data->is_BranchData()) {
1624 not_taken = data->as_BranchData()->not_taken();
1625 }
1626
1627 // scale the counts to be commensurate with invocation counts:
1628 // NOTE: overflow for positive values is clamped at INT_MAX
1629 taken = method()->scale_count(taken);
1630 not_taken = method()->scale_count(not_taken);
1631 }
1632 // At this point, saturation or overflow is indicated by INT_MAX
1633 // or a negative value.
1634
1635 // Give up if too few (or too many, in which case the sum will overflow) counts to be meaningful.
1636 // We also check that individual counters are positive first, otherwise the sum can become positive.
1637 if (!counters_are_meaningful(taken, not_taken, 40)) {
1638 if (C->log() != nullptr) {
1639 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d'", iter().get_dest(), taken, not_taken);
1640 }
1641 return PROB_UNKNOWN;
1642 }
1643
1644 // Compute frequency that we arrive here
1645 float sum = taken + not_taken;
1646 // Adjust, if this block is a cloned private block but the
1647 // Jump counts are shared. Taken the private counts for
1648 // just this path instead of the shared counts.
1649 if( block()->count() > 0 )
1650 sum = block()->count();
1651 cnt = sum / FreqCountInvocations;
1652
1653 // Pin probability to sane limits
1654 float prob;
1655 if( !taken )
1656 prob = (0+PROB_MIN) / 2;
1657 else if( !not_taken )
1658 prob = (1+PROB_MAX) / 2;
1659 else { // Compute probability of true path
1660 prob = (float)taken / (float)(taken + not_taken);
1661 if (prob > PROB_MAX) prob = PROB_MAX;
1662 if (prob < PROB_MIN) prob = PROB_MIN;
1663 }
1664
1665 assert((cnt > 0.0f) && (prob > 0.0f),
1666 "Bad frequency assignment in if cnt=%g prob=%g taken=%d not_taken=%d", cnt, prob, taken, not_taken);
1667
1668 if (C->log() != nullptr) {
1669 const char* prob_str = nullptr;
1670 if (prob >= PROB_MAX) prob_str = (prob == PROB_MAX) ? "max" : "always";
1671 if (prob <= PROB_MIN) prob_str = (prob == PROB_MIN) ? "min" : "never";
1672 char prob_str_buf[30];
1673 if (prob_str == nullptr) {
1674 jio_snprintf(prob_str_buf, sizeof(prob_str_buf), "%20.2f", prob);
1675 prob_str = prob_str_buf;
1676 }
1677 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d' cnt='%f' prob='%s'",
1678 iter().get_dest(), taken, not_taken, cnt, prob_str);
1679 }
1680 return prob;
1681 }
1682
1683 //-----------------------------branch_prediction-------------------------------
1684 float Parse::branch_prediction(float& cnt,
1685 BoolTest::mask btest,
1686 int target_bci,
1687 Node* test) {
1688 float prob = dynamic_branch_prediction(cnt, btest, test);
1689 // If prob is unknown, switch to static prediction
1690 if (prob != PROB_UNKNOWN) return prob;
1691
1692 prob = PROB_FAIR; // Set default value
1693 if (btest == BoolTest::eq) // Exactly equal test?
1694 prob = PROB_STATIC_INFREQUENT; // Assume its relatively infrequent
1695 else if (btest == BoolTest::ne)
1696 prob = PROB_STATIC_FREQUENT; // Assume its relatively frequent
1697
1698 // If this is a conditional test guarding a backwards branch,
1699 // assume its a loop-back edge. Make it a likely taken branch.
1700 if (target_bci < bci()) {
1701 if (is_osr_parse()) { // Could be a hot OSR'd loop; force deopt
1702 // Since it's an OSR, we probably have profile data, but since
1703 // branch_prediction returned PROB_UNKNOWN, the counts are too small.
1704 // Let's make a special check here for completely zero counts.
1705 ciMethodData* methodData = method()->method_data();
1706 if (!methodData->is_empty()) {
1707 ciProfileData* data = methodData->bci_to_data(bci());
1708 // Only stop for truly zero counts, which mean an unknown part
1709 // of the OSR-ed method, and we want to deopt to gather more stats.
1710 // If you have ANY counts, then this loop is simply 'cold' relative
1711 // to the OSR loop.
1712 if (data == nullptr ||
1713 (data->as_BranchData()->taken() + data->as_BranchData()->not_taken() == 0)) {
1714 // This is the only way to return PROB_UNKNOWN:
1715 return PROB_UNKNOWN;
1716 }
1717 }
1718 }
1719 prob = PROB_STATIC_FREQUENT; // Likely to take backwards branch
1720 }
1721
1722 assert(prob != PROB_UNKNOWN, "must have some guess at this point");
1723 return prob;
1724 }
1725
1726 // The magic constants are chosen so as to match the output of
1727 // branch_prediction() when the profile reports a zero taken count.
1728 // It is important to distinguish zero counts unambiguously, because
1729 // some branches (e.g., _213_javac.Assembler.eliminate) validly produce
1730 // very small but nonzero probabilities, which if confused with zero
1731 // counts would keep the program recompiling indefinitely.
1732 bool Parse::seems_never_taken(float prob) const {
1733 return prob < PROB_MIN;
1734 }
1735
1736 //-------------------------------repush_if_args--------------------------------
1737 // Push arguments of an "if" bytecode back onto the stack by adjusting _sp.
1738 inline int Parse::repush_if_args() {
1739 if (PrintOpto && WizardMode) {
1740 tty->print("defending against excessive implicit null exceptions on %s @%d in ",
1741 Bytecodes::name(iter().cur_bc()), iter().cur_bci());
1742 method()->print_name(); tty->cr();
1743 }
1744 int bc_depth = - Bytecodes::depth(iter().cur_bc());
1745 assert(bc_depth == 1 || bc_depth == 2, "only two kinds of branches");
1746 DEBUG_ONLY(sync_jvms()); // argument(n) requires a synced jvms
1747 assert(argument(0) != nullptr, "must exist");
1748 assert(bc_depth == 1 || argument(1) != nullptr, "two must exist");
1749 inc_sp(bc_depth);
1750 return bc_depth;
1751 }
1752
1753 // Used by StressUnstableIfTraps
1754 static volatile int _trap_stress_counter = 0;
1755
1756 void Parse::increment_trap_stress_counter(Node*& counter, Node*& incr_store) {
1757 Node* counter_addr = makecon(TypeRawPtr::make((address)&_trap_stress_counter));
1758 counter = make_load(control(), counter_addr, TypeInt::INT, T_INT, MemNode::unordered);
1759 counter = _gvn.transform(new AddINode(counter, intcon(1)));
1760 incr_store = store_to_memory(control(), counter_addr, counter, T_INT, MemNode::unordered);
1761 }
1762
1763 //----------------------------------do_ifnull----------------------------------
1764 void Parse::do_ifnull(BoolTest::mask btest, Node *c) {
1765 int target_bci = iter().get_dest();
1766
1767 Node* counter = nullptr;
1768 Node* incr_store = nullptr;
1769 bool do_stress_trap = StressUnstableIfTraps && ((C->random() % 2) == 0);
1770 if (do_stress_trap) {
1771 increment_trap_stress_counter(counter, incr_store);
1772 }
1773
1774 Block* branch_block = successor_for_bci(target_bci);
1775 Block* next_block = successor_for_bci(iter().next_bci());
1776
1777 float cnt;
1778 float prob = branch_prediction(cnt, btest, target_bci, c);
1779 if (prob == PROB_UNKNOWN) {
1780 // (An earlier version of do_ifnull omitted this trap for OSR methods.)
1781 if (PrintOpto && Verbose) {
1782 tty->print_cr("Never-taken edge stops compilation at bci %d", bci());
1783 }
1784 repush_if_args(); // to gather stats on loop
1785 uncommon_trap(Deoptimization::Reason_unreached,
1786 Deoptimization::Action_reinterpret,
1787 nullptr, "cold");
1788 if (C->eliminate_boxing()) {
1789 // Mark the successor blocks as parsed
1790 branch_block->next_path_num();
1791 next_block->next_path_num();
1792 }
1793 return;
1794 }
1795
1796 NOT_PRODUCT(explicit_null_checks_inserted++);
1797
1798 // Generate real control flow
1799 Node *tst = _gvn.transform( new BoolNode( c, btest ) );
1800
1801 // Sanity check the probability value
1802 assert(prob > 0.0f,"Bad probability in Parser");
1803 // Need xform to put node in hash table
1804 IfNode *iff = create_and_xform_if( control(), tst, prob, cnt );
1805 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser");
1806 // True branch
1807 { PreserveJVMState pjvms(this);
1808 Node* iftrue = _gvn.transform( new IfTrueNode (iff) );
1809 set_control(iftrue);
1810
1811 if (stopped()) { // Path is dead?
1812 NOT_PRODUCT(explicit_null_checks_elided++);
1813 if (C->eliminate_boxing()) {
1814 // Mark the successor block as parsed
1815 branch_block->next_path_num();
1816 }
1817 } else { // Path is live.
1818 adjust_map_after_if(btest, c, prob, branch_block);
1819 if (!stopped()) {
1820 merge(target_bci);
1821 }
1822 }
1823 }
1824
1825 // False branch
1826 Node* iffalse = _gvn.transform( new IfFalseNode(iff) );
1827 set_control(iffalse);
1828
1829 if (stopped()) { // Path is dead?
1830 NOT_PRODUCT(explicit_null_checks_elided++);
1831 if (C->eliminate_boxing()) {
1832 // Mark the successor block as parsed
1833 next_block->next_path_num();
1834 }
1835 } else { // Path is live.
1836 adjust_map_after_if(BoolTest(btest).negate(), c, 1.0-prob, next_block);
1837 }
1838
1839 if (do_stress_trap) {
1840 stress_trap(iff, counter, incr_store);
1841 }
1842 }
1843
1844 //------------------------------------do_if------------------------------------
1845 void Parse::do_if(BoolTest::mask btest, Node* c, bool can_trap, bool new_path, Node** ctrl_taken) {
1846 int target_bci = iter().get_dest();
1847
1848 Block* branch_block = successor_for_bci(target_bci);
1849 Block* next_block = successor_for_bci(iter().next_bci());
1850
1851 float cnt;
1852 float prob = branch_prediction(cnt, btest, target_bci, c);
1853 float untaken_prob = 1.0 - prob;
1854
1855 if (prob == PROB_UNKNOWN) {
1856 if (PrintOpto && Verbose) {
1857 tty->print_cr("Never-taken edge stops compilation at bci %d", bci());
1858 }
1859 repush_if_args(); // to gather stats on loop
1860 uncommon_trap(Deoptimization::Reason_unreached,
1861 Deoptimization::Action_reinterpret,
1862 nullptr, "cold");
1863 if (C->eliminate_boxing()) {
1864 // Mark the successor blocks as parsed
1865 branch_block->next_path_num();
1866 next_block->next_path_num();
1867 }
1868 return;
1869 }
1870
1871 Node* counter = nullptr;
1872 Node* incr_store = nullptr;
1873 bool do_stress_trap = StressUnstableIfTraps && ((C->random() % 2) == 0);
1874 if (do_stress_trap) {
1875 increment_trap_stress_counter(counter, incr_store);
1876 }
1877
1878 // Sanity check the probability value
1879 assert(0.0f < prob && prob < 1.0f,"Bad probability in Parser");
1880
1881 bool taken_if_true = true;
1882 // Convert BoolTest to canonical form:
1883 if (!BoolTest(btest).is_canonical()) {
1884 btest = BoolTest(btest).negate();
1885 taken_if_true = false;
1886 // prob is NOT updated here; it remains the probability of the taken
1887 // path (as opposed to the prob of the path guarded by an 'IfTrueNode').
1888 }
1889 assert(btest != BoolTest::eq, "!= is the only canonical exact test");
1890
1891 Node* tst0 = new BoolNode(c, btest);
1892 Node* tst = _gvn.transform(tst0);
1893 BoolTest::mask taken_btest = BoolTest::illegal;
1894 BoolTest::mask untaken_btest = BoolTest::illegal;
1895
1896 if (tst->is_Bool()) {
1897 // Refresh c from the transformed bool node, since it may be
1898 // simpler than the original c. Also re-canonicalize btest.
1899 // This wins when (Bool ne (Conv2B p) 0) => (Bool ne (CmpP p null)).
1900 // That can arise from statements like: if (x instanceof C) ...
1901 if (tst != tst0) {
1902 // Canonicalize one more time since transform can change it.
1903 btest = tst->as_Bool()->_test._test;
1904 if (!BoolTest(btest).is_canonical()) {
1905 // Reverse edges one more time...
1906 tst = _gvn.transform( tst->as_Bool()->negate(&_gvn) );
1907 btest = tst->as_Bool()->_test._test;
1908 assert(BoolTest(btest).is_canonical(), "sanity");
1909 taken_if_true = !taken_if_true;
1910 }
1911 c = tst->in(1);
1912 }
1913 BoolTest::mask neg_btest = BoolTest(btest).negate();
1914 taken_btest = taken_if_true ? btest : neg_btest;
1915 untaken_btest = taken_if_true ? neg_btest : btest;
1916 }
1917
1918 // Generate real control flow
1919 float true_prob = (taken_if_true ? prob : untaken_prob);
1920 IfNode* iff = create_and_map_if(control(), tst, true_prob, cnt);
1921 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser");
1922 Node* taken_branch = new IfTrueNode(iff);
1923 Node* untaken_branch = new IfFalseNode(iff);
1924 if (!taken_if_true) { // Finish conversion to canonical form
1925 Node* tmp = taken_branch;
1926 taken_branch = untaken_branch;
1927 untaken_branch = tmp;
1928 }
1929
1930 // Branch is taken:
1931 { PreserveJVMState pjvms(this);
1932 taken_branch = _gvn.transform(taken_branch);
1933 set_control(taken_branch);
1934
1935 if (stopped()) {
1936 if (C->eliminate_boxing() && !new_path) {
1937 // Mark the successor block as parsed (if we haven't created a new path)
1938 branch_block->next_path_num();
1939 }
1940 } else {
1941 adjust_map_after_if(taken_btest, c, prob, branch_block, can_trap);
1942 if (!stopped()) {
1943 if (new_path) {
1944 // Merge by using a new path
1945 merge_new_path(target_bci);
1946 } else if (ctrl_taken != nullptr) {
1947 // Don't merge but save taken branch to be wired by caller
1948 *ctrl_taken = control();
1949 } else {
1950 merge(target_bci);
1951 }
1952 }
1953 }
1954 }
1955
1956 untaken_branch = _gvn.transform(untaken_branch);
1957 set_control(untaken_branch);
1958
1959 // Branch not taken.
1960 if (stopped() && ctrl_taken == nullptr) {
1961 if (C->eliminate_boxing()) {
1962 // Mark the successor block as parsed (if caller does not re-wire control flow)
1963 next_block->next_path_num();
1964 }
1965 } else {
1966 adjust_map_after_if(untaken_btest, c, untaken_prob, next_block, can_trap);
1967 }
1968
1969 if (do_stress_trap) {
1970 stress_trap(iff, counter, incr_store);
1971 }
1972 }
1973
1974
1975 static ProfilePtrKind speculative_ptr_kind(const TypeOopPtr* t) {
1976 if (t->speculative() == nullptr) {
1977 return ProfileUnknownNull;
1978 }
1979 if (t->speculative_always_null()) {
1980 return ProfileAlwaysNull;
1981 }
1982 if (t->speculative_maybe_null()) {
1983 return ProfileMaybeNull;
1984 }
1985 return ProfileNeverNull;
1986 }
1987
1988 void Parse::acmp_always_null_input(Node* input, const TypeOopPtr* tinput, BoolTest::mask btest, Node* eq_region) {
1989 inc_sp(2);
1990 Node* cast = null_check_common(input, T_OBJECT, true, nullptr,
1991 !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_null_check) &&
1992 speculative_ptr_kind(tinput) == ProfileAlwaysNull);
1993 dec_sp(2);
1994 if (btest == BoolTest::ne) {
1995 {
1996 PreserveJVMState pjvms(this);
1997 replace_in_map(input, cast);
1998 int target_bci = iter().get_dest();
1999 merge(target_bci);
2000 }
2001 record_for_igvn(eq_region);
2002 set_control(_gvn.transform(eq_region));
2003 } else {
2004 replace_in_map(input, cast);
2005 }
2006 }
2007
2008 Node* Parse::acmp_null_check(Node* input, const TypeOopPtr* tinput, ProfilePtrKind input_ptr, Node*& null_ctl) {
2009 inc_sp(2);
2010 null_ctl = top();
2011 Node* cast = null_check_oop(input, &null_ctl,
2012 input_ptr == ProfileNeverNull || (input_ptr == ProfileUnknownNull && !too_many_traps_or_recompiles(Deoptimization::Reason_null_check)),
2013 false,
2014 speculative_ptr_kind(tinput) == ProfileNeverNull &&
2015 !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_null_check));
2016 dec_sp(2);
2017 assert(!stopped(), "null input should have been caught earlier");
2018 return cast;
2019 }
2020
2021 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) {
2022 Node* ne_region = new RegionNode(1);
2023 Node* null_ctl;
2024 Node* cast = acmp_null_check(input, tinput, input_ptr, null_ctl);
2025 ne_region->add_req(null_ctl);
2026
2027 Node* slow_ctl = type_check_receiver(cast, input_type, 1.0, &cast);
2028 {
2029 PreserveJVMState pjvms(this);
2030 inc_sp(2);
2031 set_control(slow_ctl);
2032 Deoptimization::DeoptReason reason;
2033 if (tinput->speculative_type() != nullptr && !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_class_check)) {
2034 reason = Deoptimization::Reason_speculate_class_check;
2035 } else {
2036 reason = Deoptimization::Reason_class_check;
2037 }
2038 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
2039 }
2040 ne_region->add_req(control());
2041
2042 record_for_igvn(ne_region);
2043 set_control(_gvn.transform(ne_region));
2044 if (btest == BoolTest::ne) {
2045 {
2046 PreserveJVMState pjvms(this);
2047 if (null_ctl == top()) {
2048 replace_in_map(input, cast);
2049 }
2050 int target_bci = iter().get_dest();
2051 merge(target_bci);
2052 }
2053 record_for_igvn(eq_region);
2054 set_control(_gvn.transform(eq_region));
2055 } else {
2056 if (null_ctl == top()) {
2057 replace_in_map(input, cast);
2058 }
2059 set_control(_gvn.transform(ne_region));
2060 }
2061 }
2062
2063 void Parse::acmp_unknown_non_inline_type_input(Node* input, const TypeOopPtr* tinput, ProfilePtrKind input_ptr, BoolTest::mask btest, Node* eq_region) {
2064 Node* ne_region = new RegionNode(1);
2065 Node* null_ctl;
2066 Node* cast = acmp_null_check(input, tinput, input_ptr, null_ctl);
2067 ne_region->add_req(null_ctl);
2068
2069 {
2070 BuildCutout unless(this, inline_type_test(cast, /* is_inline = */ false), PROB_MAX);
2071 inc_sp(2);
2072 uncommon_trap_exact(Deoptimization::Reason_class_check, Deoptimization::Action_maybe_recompile);
2073 }
2074
2075 ne_region->add_req(control());
2076
2077 record_for_igvn(ne_region);
2078 set_control(_gvn.transform(ne_region));
2079 if (btest == BoolTest::ne) {
2080 {
2081 PreserveJVMState pjvms(this);
2082 if (null_ctl == top()) {
2083 replace_in_map(input, cast);
2084 }
2085 int target_bci = iter().get_dest();
2086 merge(target_bci);
2087 }
2088 record_for_igvn(eq_region);
2089 set_control(_gvn.transform(eq_region));
2090 } else {
2091 if (null_ctl == top()) {
2092 replace_in_map(input, cast);
2093 }
2094 set_control(_gvn.transform(ne_region));
2095 }
2096 }
2097
2098 void Parse::do_acmp(BoolTest::mask btest, Node* left, Node* right) {
2099 ciKlass* left_type = nullptr;
2100 ciKlass* right_type = nullptr;
2101 ProfilePtrKind left_ptr = ProfileUnknownNull;
2102 ProfilePtrKind right_ptr = ProfileUnknownNull;
2103 bool left_inline_type = true;
2104 bool right_inline_type = true;
2105
2106 // Leverage profiling at acmp
2107 if (UseACmpProfile) {
2108 method()->acmp_profiled_type(bci(), left_type, right_type, left_ptr, right_ptr, left_inline_type, right_inline_type);
2109 if (too_many_traps_or_recompiles(Deoptimization::Reason_class_check)) {
2110 left_type = nullptr;
2111 right_type = nullptr;
2112 left_inline_type = true;
2113 right_inline_type = true;
2114 }
2115 if (too_many_traps_or_recompiles(Deoptimization::Reason_null_check)) {
2116 left_ptr = ProfileUnknownNull;
2117 right_ptr = ProfileUnknownNull;
2118 }
2119 }
2120
2121 if (UseTypeSpeculation) {
2122 record_profile_for_speculation(left, left_type, left_ptr);
2123 record_profile_for_speculation(right, right_type, right_ptr);
2124 }
2125
2126 if (!EnableValhalla) {
2127 Node* cmp = CmpP(left, right);
2128 cmp = optimize_cmp_with_klass(cmp);
2129 do_if(btest, cmp);
2130 return;
2131 }
2132
2133 // Check for equality before potentially allocating
2134 if (left == right) {
2135 do_if(btest, makecon(TypeInt::CC_EQ));
2136 return;
2137 }
2138
2139 // Allocate inline type operands and re-execute on deoptimization
2140 if (left->is_InlineType()) {
2141 if (_gvn.type(right)->is_zero_type() ||
2142 (right->is_InlineType() && _gvn.type(right->as_InlineType()->get_is_init())->is_zero_type())) {
2143 // Null checking a scalarized but nullable inline type. Check the IsInit
2144 // input instead of the oop input to avoid keeping buffer allocations alive.
2145 Node* cmp = CmpI(left->as_InlineType()->get_is_init(), intcon(0));
2146 do_if(btest, cmp);
2147 return;
2148 } else {
2149 PreserveReexecuteState preexecs(this);
2150 inc_sp(2);
2151 jvms()->set_should_reexecute(true);
2152 left = left->as_InlineType()->buffer(this)->get_oop();
2153 }
2154 }
2155 if (right->is_InlineType()) {
2156 PreserveReexecuteState preexecs(this);
2157 inc_sp(2);
2158 jvms()->set_should_reexecute(true);
2159 right = right->as_InlineType()->buffer(this)->get_oop();
2160 }
2161
2162 // First, do a normal pointer comparison
2163 const TypeOopPtr* tleft = _gvn.type(left)->isa_oopptr();
2164 const TypeOopPtr* tright = _gvn.type(right)->isa_oopptr();
2165 Node* cmp = CmpP(left, right);
2166 cmp = optimize_cmp_with_klass(cmp);
2167 if (tleft == nullptr || !tleft->can_be_inline_type() ||
2168 tright == nullptr || !tright->can_be_inline_type()) {
2169 // This is sufficient, if one of the operands can't be an inline type
2170 do_if(btest, cmp);
2171 return;
2172 }
2173
2174 // Don't add traps to unstable if branches because additional checks are required to
2175 // decide if the operands are equal/substitutable and we therefore shouldn't prune
2176 // branches for one if based on the profiling of the acmp branches.
2177 // Also, OptimizeUnstableIf would set an incorrect re-rexecution state because it
2178 // assumes that there is a 1-1 mapping between the if and the acmp branches and that
2179 // hitting a trap means that we will take the corresponding acmp branch on re-execution.
2180 const bool can_trap = true;
2181
2182 Node* eq_region = nullptr;
2183 if (btest == BoolTest::eq) {
2184 do_if(btest, cmp, !can_trap, true);
2185 if (stopped()) {
2186 // Pointers are equal, operands must be equal
2187 return;
2188 }
2189 } else {
2190 assert(btest == BoolTest::ne, "only eq or ne");
2191 Node* is_not_equal = nullptr;
2192 eq_region = new RegionNode(3);
2193 {
2194 PreserveJVMState pjvms(this);
2195 // Pointers are not equal, but more checks are needed to determine if the operands are (not) substitutable
2196 do_if(btest, cmp, !can_trap, false, &is_not_equal);
2197 if (!stopped()) {
2198 eq_region->init_req(1, control());
2199 }
2200 }
2201 if (is_not_equal == nullptr || is_not_equal->is_top()) {
2202 record_for_igvn(eq_region);
2203 set_control(_gvn.transform(eq_region));
2204 return;
2205 }
2206 set_control(is_not_equal);
2207 }
2208
2209 // Prefer speculative types if available
2210 if (!too_many_traps_or_recompiles(Deoptimization::Reason_speculate_class_check)) {
2211 if (tleft->speculative_type() != nullptr) {
2212 left_type = tleft->speculative_type();
2213 }
2214 if (tright->speculative_type() != nullptr) {
2215 right_type = tright->speculative_type();
2216 }
2217 }
2218
2219 if (speculative_ptr_kind(tleft) != ProfileMaybeNull && speculative_ptr_kind(tleft) != ProfileUnknownNull) {
2220 ProfilePtrKind speculative_left_ptr = speculative_ptr_kind(tleft);
2221 if (speculative_left_ptr == ProfileAlwaysNull && !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_null_assert)) {
2222 left_ptr = speculative_left_ptr;
2223 } else if (speculative_left_ptr == ProfileNeverNull && !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_null_check)) {
2224 left_ptr = speculative_left_ptr;
2225 }
2226 }
2227 if (speculative_ptr_kind(tright) != ProfileMaybeNull && speculative_ptr_kind(tright) != ProfileUnknownNull) {
2228 ProfilePtrKind speculative_right_ptr = speculative_ptr_kind(tright);
2229 if (speculative_right_ptr == ProfileAlwaysNull && !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_null_assert)) {
2230 right_ptr = speculative_right_ptr;
2231 } else if (speculative_right_ptr == ProfileNeverNull && !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_null_check)) {
2232 right_ptr = speculative_right_ptr;
2233 }
2234 }
2235
2236 if (left_ptr == ProfileAlwaysNull) {
2237 // Comparison with null. Assert the input is indeed null and we're done.
2238 acmp_always_null_input(left, tleft, btest, eq_region);
2239 return;
2240 }
2241 if (right_ptr == ProfileAlwaysNull) {
2242 // Comparison with null. Assert the input is indeed null and we're done.
2243 acmp_always_null_input(right, tright, btest, eq_region);
2244 return;
2245 }
2246 if (left_type != nullptr && !left_type->is_inlinetype()) {
2247 // Comparison with an object of known type
2248 acmp_known_non_inline_type_input(left, tleft, left_ptr, left_type, btest, eq_region);
2249 return;
2250 }
2251 if (right_type != nullptr && !right_type->is_inlinetype()) {
2252 // Comparison with an object of known type
2253 acmp_known_non_inline_type_input(right, tright, right_ptr, right_type, btest, eq_region);
2254 return;
2255 }
2256 if (!left_inline_type) {
2257 // Comparison with an object known not to be an inline type
2258 acmp_unknown_non_inline_type_input(left, tleft, left_ptr, btest, eq_region);
2259 return;
2260 }
2261 if (!right_inline_type) {
2262 // Comparison with an object known not to be an inline type
2263 acmp_unknown_non_inline_type_input(right, tright, right_ptr, btest, eq_region);
2264 return;
2265 }
2266
2267 // Pointers are not equal, check if first operand is non-null
2268 Node* ne_region = new RegionNode(6);
2269 Node* null_ctl;
2270 Node* not_null_right = acmp_null_check(right, tright, right_ptr, null_ctl);
2271 ne_region->init_req(1, null_ctl);
2272
2273 // First operand is non-null, check if it is an inline type
2274 Node* is_value = inline_type_test(not_null_right);
2275 IfNode* is_value_iff = create_and_map_if(control(), is_value, PROB_FAIR, COUNT_UNKNOWN);
2276 Node* not_value = _gvn.transform(new IfFalseNode(is_value_iff));
2277 ne_region->init_req(2, not_value);
2278 set_control(_gvn.transform(new IfTrueNode(is_value_iff)));
2279
2280 // The first operand is an inline type, check if the second operand is non-null
2281 Node* not_null_left = acmp_null_check(left, tleft, left_ptr, null_ctl);
2282 ne_region->init_req(3, null_ctl);
2283
2284 // Check if both operands are of the same class.
2285 Node* kls_left = load_object_klass(not_null_left);
2286 Node* kls_right = load_object_klass(not_null_right);
2287 Node* kls_cmp = CmpP(kls_left, kls_right);
2288 Node* kls_bol = _gvn.transform(new BoolNode(kls_cmp, BoolTest::ne));
2289 IfNode* kls_iff = create_and_map_if(control(), kls_bol, PROB_FAIR, COUNT_UNKNOWN);
2290 Node* kls_ne = _gvn.transform(new IfTrueNode(kls_iff));
2291 set_control(_gvn.transform(new IfFalseNode(kls_iff)));
2292 ne_region->init_req(4, kls_ne);
2293
2294 if (stopped()) {
2295 record_for_igvn(ne_region);
2296 set_control(_gvn.transform(ne_region));
2297 if (btest == BoolTest::ne) {
2298 {
2299 PreserveJVMState pjvms(this);
2300 int target_bci = iter().get_dest();
2301 merge(target_bci);
2302 }
2303 record_for_igvn(eq_region);
2304 set_control(_gvn.transform(eq_region));
2305 }
2306 return;
2307 }
2308
2309 // Both operands are values types of the same class, we need to perform a
2310 // substitutability test. Delegate to ValueObjectMethods::isSubstitutable().
2311 Node* ne_io_phi = PhiNode::make(ne_region, i_o());
2312 Node* mem = reset_memory();
2313 Node* ne_mem_phi = PhiNode::make(ne_region, mem);
2314
2315 Node* eq_io_phi = nullptr;
2316 Node* eq_mem_phi = nullptr;
2317 if (eq_region != nullptr) {
2318 eq_io_phi = PhiNode::make(eq_region, i_o());
2319 eq_mem_phi = PhiNode::make(eq_region, mem);
2320 }
2321
2322 set_all_memory(mem);
2323
2324 kill_dead_locals();
2325 ciMethod* subst_method = ciEnv::current()->ValueObjectMethods_klass()->find_method(ciSymbols::isSubstitutable_name(), ciSymbols::object_object_boolean_signature());
2326 CallStaticJavaNode *call = new CallStaticJavaNode(C, TypeFunc::make(subst_method), SharedRuntime::get_resolve_static_call_stub(), subst_method);
2327 call->set_override_symbolic_info(true);
2328 call->init_req(TypeFunc::Parms, not_null_left);
2329 call->init_req(TypeFunc::Parms+1, not_null_right);
2330 inc_sp(2);
2331 set_edges_for_java_call(call, false, false);
2332 Node* ret = set_results_for_java_call(call, false, true);
2333 dec_sp(2);
2334
2335 // Test the return value of ValueObjectMethods::isSubstitutable()
2336 // This is the last check, do_if can emit traps now.
2337 Node* subst_cmp = _gvn.transform(new CmpINode(ret, intcon(1)));
2338 Node* ctl = C->top();
2339 if (btest == BoolTest::eq) {
2340 PreserveJVMState pjvms(this);
2341 do_if(btest, subst_cmp, can_trap);
2342 if (!stopped()) {
2343 ctl = control();
2344 }
2345 } else {
2346 assert(btest == BoolTest::ne, "only eq or ne");
2347 PreserveJVMState pjvms(this);
2348 do_if(btest, subst_cmp, can_trap, false, &ctl);
2349 if (!stopped()) {
2350 eq_region->init_req(2, control());
2351 eq_io_phi->init_req(2, i_o());
2352 eq_mem_phi->init_req(2, reset_memory());
2353 }
2354 }
2355 ne_region->init_req(5, ctl);
2356 ne_io_phi->init_req(5, i_o());
2357 ne_mem_phi->init_req(5, reset_memory());
2358
2359 record_for_igvn(ne_region);
2360 set_control(_gvn.transform(ne_region));
2361 set_i_o(_gvn.transform(ne_io_phi));
2362 set_all_memory(_gvn.transform(ne_mem_phi));
2363
2364 if (btest == BoolTest::ne) {
2365 {
2366 PreserveJVMState pjvms(this);
2367 int target_bci = iter().get_dest();
2368 merge(target_bci);
2369 }
2370
2371 record_for_igvn(eq_region);
2372 set_control(_gvn.transform(eq_region));
2373 set_i_o(_gvn.transform(eq_io_phi));
2374 set_all_memory(_gvn.transform(eq_mem_phi));
2375 }
2376 }
2377
2378 // Force unstable if traps to be taken randomly to trigger intermittent bugs such as incorrect debug information.
2379 // Add another if before the unstable if that checks a "random" condition at runtime (a simple shared counter) and
2380 // then either takes the trap or executes the original, unstable if.
2381 void Parse::stress_trap(IfNode* orig_iff, Node* counter, Node* incr_store) {
2382 // Search for an unstable if trap
2383 CallStaticJavaNode* trap = nullptr;
2384 assert(orig_iff->Opcode() == Op_If && orig_iff->outcnt() == 2, "malformed if");
2385 ProjNode* trap_proj = orig_iff->uncommon_trap_proj(trap, Deoptimization::Reason_unstable_if);
2386 if (trap == nullptr || !trap->jvms()->should_reexecute()) {
2387 // No suitable trap found. Remove unused counter load and increment.
2388 C->gvn_replace_by(incr_store, incr_store->in(MemNode::Memory));
2389 return;
2390 }
2391
2392 // Remove trap from optimization list since we add another path to the trap.
2393 bool success = C->remove_unstable_if_trap(trap, true);
2394 assert(success, "Trap already modified");
2395
2396 // Add a check before the original if that will trap with a certain frequency and execute the original if otherwise
2397 int freq_log = (C->random() % 31) + 1; // Random logarithmic frequency in [1, 31]
2398 Node* mask = intcon(right_n_bits(freq_log));
2399 counter = _gvn.transform(new AndINode(counter, mask));
2400 Node* cmp = _gvn.transform(new CmpINode(counter, intcon(0)));
2401 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::mask::eq));
2402 IfNode* iff = _gvn.transform(new IfNode(orig_iff->in(0), bol, orig_iff->_prob, orig_iff->_fcnt))->as_If();
2403 Node* if_true = _gvn.transform(new IfTrueNode(iff));
2404 Node* if_false = _gvn.transform(new IfFalseNode(iff));
2405 assert(!if_true->is_top() && !if_false->is_top(), "trap always / never taken");
2406
2407 // Trap
2408 assert(trap_proj->outcnt() == 1, "some other nodes are dependent on the trap projection");
2409
2410 Node* trap_region = new RegionNode(3);
2411 trap_region->set_req(1, trap_proj);
2412 trap_region->set_req(2, if_true);
2413 trap->set_req(0, _gvn.transform(trap_region));
2414
2415 // Don't trap, execute original if
2416 orig_iff->set_req(0, if_false);
2417 }
2418
2419 bool Parse::path_is_suitable_for_uncommon_trap(float prob) const {
2420 // Randomly skip emitting an uncommon trap
2421 if (StressUnstableIfTraps && ((C->random() % 2) == 0)) {
2422 return false;
2423 }
2424 // Don't want to speculate on uncommon traps when running with -Xcomp
2425 if (!UseInterpreter) {
2426 return false;
2427 }
2428 return seems_never_taken(prob) &&
2429 !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if);
2430 }
2431
2432 void Parse::maybe_add_predicate_after_if(Block* path) {
2433 if (path->is_SEL_head() && path->preds_parsed() == 0) {
2434 // Add predicates at bci of if dominating the loop so traps can be
2435 // recorded on the if's profile data
2436 int bc_depth = repush_if_args();
2437 add_parse_predicates();
2438 dec_sp(bc_depth);
2439 path->set_has_predicates();
2440 }
2441 }
2442
2443
2444 //----------------------------adjust_map_after_if------------------------------
2445 // Adjust the JVM state to reflect the result of taking this path.
2446 // Basically, it means inspecting the CmpNode controlling this
2447 // branch, seeing how it constrains a tested value, and then
2448 // deciding if it's worth our while to encode this constraint
2449 // as graph nodes in the current abstract interpretation map.
2450 void Parse::adjust_map_after_if(BoolTest::mask btest, Node* c, float prob, Block* path, bool can_trap) {
2451 if (!c->is_Cmp()) {
2452 maybe_add_predicate_after_if(path);
2453 return;
2454 }
2455
2456 if (stopped() || btest == BoolTest::illegal) {
2457 return; // nothing to do
2458 }
2459
2460 bool is_fallthrough = (path == successor_for_bci(iter().next_bci()));
2461
2462 if (can_trap && path_is_suitable_for_uncommon_trap(prob)) {
2463 repush_if_args();
2464 Node* call = uncommon_trap(Deoptimization::Reason_unstable_if,
2465 Deoptimization::Action_reinterpret,
2466 nullptr,
2467 (is_fallthrough ? "taken always" : "taken never"));
2468
2469 if (call != nullptr) {
2470 C->record_unstable_if_trap(new UnstableIfTrap(call->as_CallStaticJava(), path));
2471 }
2472 return;
2473 }
2474
2475 Node* val = c->in(1);
2476 Node* con = c->in(2);
2477 const Type* tcon = _gvn.type(con);
2478 const Type* tval = _gvn.type(val);
2479 bool have_con = tcon->singleton();
2480 if (tval->singleton()) {
2481 if (!have_con) {
2482 // Swap, so constant is in con.
2483 con = val;
2484 tcon = tval;
2485 val = c->in(2);
2486 tval = _gvn.type(val);
2487 btest = BoolTest(btest).commute();
2488 have_con = true;
2489 } else {
2490 // Do we have two constants? Then leave well enough alone.
2491 have_con = false;
2492 }
2493 }
2494 if (!have_con) { // remaining adjustments need a con
2495 maybe_add_predicate_after_if(path);
2496 return;
2497 }
2498
2499 sharpen_type_after_if(btest, con, tcon, val, tval);
2500 maybe_add_predicate_after_if(path);
2501 }
2502
2503
2504 static Node* extract_obj_from_klass_load(PhaseGVN* gvn, Node* n) {
2505 Node* ldk;
2506 if (n->is_DecodeNKlass()) {
2507 if (n->in(1)->Opcode() != Op_LoadNKlass) {
2508 return nullptr;
2509 } else {
2510 ldk = n->in(1);
2511 }
2512 } else if (n->Opcode() != Op_LoadKlass) {
2513 return nullptr;
2514 } else {
2515 ldk = n;
2516 }
2517 assert(ldk != nullptr && ldk->is_Load(), "should have found a LoadKlass or LoadNKlass node");
2518
2519 Node* adr = ldk->in(MemNode::Address);
2520 intptr_t off = 0;
2521 Node* obj = AddPNode::Ideal_base_and_offset(adr, gvn, off);
2522 if (obj == nullptr || off != oopDesc::klass_offset_in_bytes()) // loading oopDesc::_klass?
2523 return nullptr;
2524 const TypePtr* tp = gvn->type(obj)->is_ptr();
2525 if (tp == nullptr || !(tp->isa_instptr() || tp->isa_aryptr())) // is obj a Java object ptr?
2526 return nullptr;
2527
2528 return obj;
2529 }
2530
2531 void Parse::sharpen_type_after_if(BoolTest::mask btest,
2532 Node* con, const Type* tcon,
2533 Node* val, const Type* tval) {
2534 // Look for opportunities to sharpen the type of a node
2535 // whose klass is compared with a constant klass.
2536 if (btest == BoolTest::eq && tcon->isa_klassptr()) {
2537 Node* obj = extract_obj_from_klass_load(&_gvn, val);
2538 const TypeOopPtr* con_type = tcon->isa_klassptr()->as_instance_type();
2539 if (obj != nullptr && (con_type->isa_instptr() || con_type->isa_aryptr())) {
2540 // Found:
2541 // Bool(CmpP(LoadKlass(obj._klass), ConP(Foo.klass)), [eq])
2542 // or the narrowOop equivalent.
2543 const Type* obj_type = _gvn.type(obj);
2544 const TypeOopPtr* tboth = obj_type->join_speculative(con_type)->isa_oopptr();
2545 if (tboth != nullptr && tboth->klass_is_exact() && tboth != obj_type &&
2546 tboth->higher_equal(obj_type)) {
2547 // obj has to be of the exact type Foo if the CmpP succeeds.
2548 int obj_in_map = map()->find_edge(obj);
2549 JVMState* jvms = this->jvms();
2550 if (obj_in_map >= 0 &&
2551 (jvms->is_loc(obj_in_map) || jvms->is_stk(obj_in_map))) {
2552 TypeNode* ccast = new CheckCastPPNode(control(), obj, tboth);
2553 const Type* tcc = ccast->as_Type()->type();
2554 assert(tcc != obj_type && tcc->higher_equal(obj_type), "must improve");
2555 // Delay transform() call to allow recovery of pre-cast value
2556 // at the control merge.
2557 _gvn.set_type_bottom(ccast);
2558 record_for_igvn(ccast);
2559 if (tboth->is_inlinetypeptr()) {
2560 ccast = InlineTypeNode::make_from_oop(this, ccast, tboth->exact_klass(true)->as_inline_klass());
2561 }
2562 // Here's the payoff.
2563 replace_in_map(obj, ccast);
2564 }
2565 }
2566 }
2567 }
2568
2569 int val_in_map = map()->find_edge(val);
2570 if (val_in_map < 0) return; // replace_in_map would be useless
2571 {
2572 JVMState* jvms = this->jvms();
2573 if (!(jvms->is_loc(val_in_map) ||
2574 jvms->is_stk(val_in_map)))
2575 return; // again, it would be useless
2576 }
2577
2578 // Check for a comparison to a constant, and "know" that the compared
2579 // value is constrained on this path.
2580 assert(tcon->singleton(), "");
2581 ConstraintCastNode* ccast = nullptr;
2582 Node* cast = nullptr;
2583
2584 switch (btest) {
2585 case BoolTest::eq: // Constant test?
2586 {
2587 const Type* tboth = tcon->join_speculative(tval);
2588 if (tboth == tval) break; // Nothing to gain.
2589 if (tcon->isa_int()) {
2590 ccast = new CastIINode(control(), val, tboth);
2591 } else if (tcon == TypePtr::NULL_PTR) {
2592 // Cast to null, but keep the pointer identity temporarily live.
2593 ccast = new CastPPNode(control(), val, tboth);
2594 } else {
2595 const TypeF* tf = tcon->isa_float_constant();
2596 const TypeD* td = tcon->isa_double_constant();
2597 // Exclude tests vs float/double 0 as these could be
2598 // either +0 or -0. Just because you are equal to +0
2599 // doesn't mean you ARE +0!
2600 // Note, following code also replaces Long and Oop values.
2601 if ((!tf || tf->_f != 0.0) &&
2602 (!td || td->_d != 0.0))
2603 cast = con; // Replace non-constant val by con.
2604 }
2605 }
2606 break;
2607
2608 case BoolTest::ne:
2609 if (tcon == TypePtr::NULL_PTR) {
2610 cast = cast_not_null(val, false);
2611 }
2612 break;
2613
2614 default:
2615 // (At this point we could record int range types with CastII.)
2616 break;
2617 }
2618
2619 if (ccast != nullptr) {
2620 const Type* tcc = ccast->as_Type()->type();
2621 assert(tcc != tval && tcc->higher_equal(tval), "must improve");
2622 // Delay transform() call to allow recovery of pre-cast value
2623 // at the control merge.
2624 _gvn.set_type_bottom(ccast);
2625 record_for_igvn(ccast);
2626 cast = ccast;
2627 }
2628
2629 if (cast != nullptr) { // Here's the payoff.
2630 replace_in_map(val, cast);
2631 }
2632 }
2633
2634 /**
2635 * Use speculative type to optimize CmpP node: if comparison is
2636 * against the low level class, cast the object to the speculative
2637 * type if any. CmpP should then go away.
2638 *
2639 * @param c expected CmpP node
2640 * @return result of CmpP on object casted to speculative type
2641 *
2642 */
2643 Node* Parse::optimize_cmp_with_klass(Node* c) {
2644 // If this is transformed by the _gvn to a comparison with the low
2645 // level klass then we may be able to use speculation
2646 if (c->Opcode() == Op_CmpP &&
2647 (c->in(1)->Opcode() == Op_LoadKlass || c->in(1)->Opcode() == Op_DecodeNKlass) &&
2648 c->in(2)->is_Con()) {
2649 Node* load_klass = nullptr;
2650 Node* decode = nullptr;
2651 if (c->in(1)->Opcode() == Op_DecodeNKlass) {
2652 decode = c->in(1);
2653 load_klass = c->in(1)->in(1);
2654 } else {
2655 load_klass = c->in(1);
2656 }
2657 if (load_klass->in(2)->is_AddP()) {
2658 Node* addp = load_klass->in(2);
2659 Node* obj = addp->in(AddPNode::Address);
2660 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
2661 if (obj_type->speculative_type_not_null() != nullptr) {
2662 ciKlass* k = obj_type->speculative_type();
2663 inc_sp(2);
2664 obj = maybe_cast_profiled_obj(obj, k);
2665 dec_sp(2);
2666 if (obj->is_InlineType()) {
2667 assert(obj->as_InlineType()->is_allocated(&_gvn), "must be allocated");
2668 obj = obj->as_InlineType()->get_oop();
2669 }
2670 // Make the CmpP use the casted obj
2671 addp = basic_plus_adr(obj, addp->in(AddPNode::Offset));
2672 load_klass = load_klass->clone();
2673 load_klass->set_req(2, addp);
2674 load_klass = _gvn.transform(load_klass);
2675 if (decode != nullptr) {
2676 decode = decode->clone();
2677 decode->set_req(1, load_klass);
2678 load_klass = _gvn.transform(decode);
2679 }
2680 c = c->clone();
2681 c->set_req(1, load_klass);
2682 c = _gvn.transform(c);
2683 }
2684 }
2685 }
2686 return c;
2687 }
2688
2689 //------------------------------do_one_bytecode--------------------------------
2690 // Parse this bytecode, and alter the Parsers JVM->Node mapping
2691 void Parse::do_one_bytecode() {
2692 Node *a, *b, *c, *d; // Handy temps
2693 BoolTest::mask btest;
2694 int i;
2695
2696 assert(!has_exceptions(), "bytecode entry state must be clear of throws");
2697
2698 if (C->check_node_count(NodeLimitFudgeFactor * 5,
2699 "out of nodes parsing method")) {
2700 return;
2701 }
2702
2703 #ifdef ASSERT
2704 // for setting breakpoints
2705 if (TraceOptoParse) {
2706 tty->print(" @");
2707 dump_bci(bci());
2708 tty->print(" %s", Bytecodes::name(bc()));
2709 tty->cr();
2710 }
2711 #endif
2712
2713 switch (bc()) {
2714 case Bytecodes::_nop:
2715 // do nothing
2716 break;
2717 case Bytecodes::_lconst_0:
2718 push_pair(longcon(0));
2719 break;
2720
2721 case Bytecodes::_lconst_1:
2722 push_pair(longcon(1));
2723 break;
2724
2725 case Bytecodes::_fconst_0:
2726 push(zerocon(T_FLOAT));
2727 break;
2728
2729 case Bytecodes::_fconst_1:
2730 push(makecon(TypeF::ONE));
2731 break;
2732
2733 case Bytecodes::_fconst_2:
2734 push(makecon(TypeF::make(2.0f)));
2735 break;
2736
2737 case Bytecodes::_dconst_0:
2738 push_pair(zerocon(T_DOUBLE));
2739 break;
2740
2741 case Bytecodes::_dconst_1:
2742 push_pair(makecon(TypeD::ONE));
2743 break;
2744
2745 case Bytecodes::_iconst_m1:push(intcon(-1)); break;
2746 case Bytecodes::_iconst_0: push(intcon( 0)); break;
2747 case Bytecodes::_iconst_1: push(intcon( 1)); break;
2748 case Bytecodes::_iconst_2: push(intcon( 2)); break;
2749 case Bytecodes::_iconst_3: push(intcon( 3)); break;
2750 case Bytecodes::_iconst_4: push(intcon( 4)); break;
2751 case Bytecodes::_iconst_5: push(intcon( 5)); break;
2752 case Bytecodes::_bipush: push(intcon(iter().get_constant_u1())); break;
2753 case Bytecodes::_sipush: push(intcon(iter().get_constant_u2())); break;
2754 case Bytecodes::_aconst_null: push(null()); break;
2755
2756 case Bytecodes::_ldc:
2757 case Bytecodes::_ldc_w:
2758 case Bytecodes::_ldc2_w: {
2759 // ciTypeFlow should trap if the ldc is in error state or if the constant is not loaded
2760 assert(!iter().is_in_error(), "ldc is in error state");
2761 ciConstant constant = iter().get_constant();
2762 assert(constant.is_loaded(), "constant is not loaded");
2763 const Type* con_type = Type::make_from_constant(constant);
2764 if (con_type != nullptr) {
2765 push_node(con_type->basic_type(), makecon(con_type));
2766 }
2767 break;
2768 }
2769
2770 case Bytecodes::_aload_0:
2771 push( local(0) );
2772 break;
2773 case Bytecodes::_aload_1:
2774 push( local(1) );
2775 break;
2776 case Bytecodes::_aload_2:
2777 push( local(2) );
2778 break;
2779 case Bytecodes::_aload_3:
2780 push( local(3) );
2781 break;
2782 case Bytecodes::_aload:
2783 push( local(iter().get_index()) );
2784 break;
2785
2786 case Bytecodes::_fload_0:
2787 case Bytecodes::_iload_0:
2788 push( local(0) );
2789 break;
2790 case Bytecodes::_fload_1:
2791 case Bytecodes::_iload_1:
2792 push( local(1) );
2793 break;
2794 case Bytecodes::_fload_2:
2795 case Bytecodes::_iload_2:
2796 push( local(2) );
2797 break;
2798 case Bytecodes::_fload_3:
2799 case Bytecodes::_iload_3:
2800 push( local(3) );
2801 break;
2802 case Bytecodes::_fload:
2803 case Bytecodes::_iload:
2804 push( local(iter().get_index()) );
2805 break;
2806 case Bytecodes::_lload_0:
2807 push_pair_local( 0 );
2808 break;
2809 case Bytecodes::_lload_1:
2810 push_pair_local( 1 );
2811 break;
2812 case Bytecodes::_lload_2:
2813 push_pair_local( 2 );
2814 break;
2815 case Bytecodes::_lload_3:
2816 push_pair_local( 3 );
2817 break;
2818 case Bytecodes::_lload:
2819 push_pair_local( iter().get_index() );
2820 break;
2821
2822 case Bytecodes::_dload_0:
2823 push_pair_local(0);
2824 break;
2825 case Bytecodes::_dload_1:
2826 push_pair_local(1);
2827 break;
2828 case Bytecodes::_dload_2:
2829 push_pair_local(2);
2830 break;
2831 case Bytecodes::_dload_3:
2832 push_pair_local(3);
2833 break;
2834 case Bytecodes::_dload:
2835 push_pair_local(iter().get_index());
2836 break;
2837 case Bytecodes::_fstore_0:
2838 case Bytecodes::_istore_0:
2839 case Bytecodes::_astore_0:
2840 set_local( 0, pop() );
2841 break;
2842 case Bytecodes::_fstore_1:
2843 case Bytecodes::_istore_1:
2844 case Bytecodes::_astore_1:
2845 set_local( 1, pop() );
2846 break;
2847 case Bytecodes::_fstore_2:
2848 case Bytecodes::_istore_2:
2849 case Bytecodes::_astore_2:
2850 set_local( 2, pop() );
2851 break;
2852 case Bytecodes::_fstore_3:
2853 case Bytecodes::_istore_3:
2854 case Bytecodes::_astore_3:
2855 set_local( 3, pop() );
2856 break;
2857 case Bytecodes::_fstore:
2858 case Bytecodes::_istore:
2859 case Bytecodes::_astore:
2860 set_local( iter().get_index(), pop() );
2861 break;
2862 // long stores
2863 case Bytecodes::_lstore_0:
2864 set_pair_local( 0, pop_pair() );
2865 break;
2866 case Bytecodes::_lstore_1:
2867 set_pair_local( 1, pop_pair() );
2868 break;
2869 case Bytecodes::_lstore_2:
2870 set_pair_local( 2, pop_pair() );
2871 break;
2872 case Bytecodes::_lstore_3:
2873 set_pair_local( 3, pop_pair() );
2874 break;
2875 case Bytecodes::_lstore:
2876 set_pair_local( iter().get_index(), pop_pair() );
2877 break;
2878
2879 // double stores
2880 case Bytecodes::_dstore_0:
2881 set_pair_local( 0, pop_pair() );
2882 break;
2883 case Bytecodes::_dstore_1:
2884 set_pair_local( 1, pop_pair() );
2885 break;
2886 case Bytecodes::_dstore_2:
2887 set_pair_local( 2, pop_pair() );
2888 break;
2889 case Bytecodes::_dstore_3:
2890 set_pair_local( 3, pop_pair() );
2891 break;
2892 case Bytecodes::_dstore:
2893 set_pair_local( iter().get_index(), pop_pair() );
2894 break;
2895
2896 case Bytecodes::_pop: dec_sp(1); break;
2897 case Bytecodes::_pop2: dec_sp(2); break;
2898 case Bytecodes::_swap:
2899 a = pop();
2900 b = pop();
2901 push(a);
2902 push(b);
2903 break;
2904 case Bytecodes::_dup:
2905 a = pop();
2906 push(a);
2907 push(a);
2908 break;
2909 case Bytecodes::_dup_x1:
2910 a = pop();
2911 b = pop();
2912 push( a );
2913 push( b );
2914 push( a );
2915 break;
2916 case Bytecodes::_dup_x2:
2917 a = pop();
2918 b = pop();
2919 c = pop();
2920 push( a );
2921 push( c );
2922 push( b );
2923 push( a );
2924 break;
2925 case Bytecodes::_dup2:
2926 a = pop();
2927 b = pop();
2928 push( b );
2929 push( a );
2930 push( b );
2931 push( a );
2932 break;
2933
2934 case Bytecodes::_dup2_x1:
2935 // before: .. c, b, a
2936 // after: .. b, a, c, b, a
2937 // not tested
2938 a = pop();
2939 b = pop();
2940 c = pop();
2941 push( b );
2942 push( a );
2943 push( c );
2944 push( b );
2945 push( a );
2946 break;
2947 case Bytecodes::_dup2_x2:
2948 // before: .. d, c, b, a
2949 // after: .. b, a, d, c, b, a
2950 // not tested
2951 a = pop();
2952 b = pop();
2953 c = pop();
2954 d = pop();
2955 push( b );
2956 push( a );
2957 push( d );
2958 push( c );
2959 push( b );
2960 push( a );
2961 break;
2962
2963 case Bytecodes::_arraylength: {
2964 // Must do null-check with value on expression stack
2965 Node *ary = null_check(peek(), T_ARRAY);
2966 // Compile-time detect of null-exception?
2967 if (stopped()) return;
2968 a = pop();
2969 push(load_array_length(a));
2970 break;
2971 }
2972
2973 case Bytecodes::_baload: array_load(T_BYTE); break;
2974 case Bytecodes::_caload: array_load(T_CHAR); break;
2975 case Bytecodes::_iaload: array_load(T_INT); break;
2976 case Bytecodes::_saload: array_load(T_SHORT); break;
2977 case Bytecodes::_faload: array_load(T_FLOAT); break;
2978 case Bytecodes::_aaload: array_load(T_OBJECT); break;
2979 case Bytecodes::_laload: array_load(T_LONG); break;
2980 case Bytecodes::_daload: array_load(T_DOUBLE); break;
2981 case Bytecodes::_bastore: array_store(T_BYTE); break;
2982 case Bytecodes::_castore: array_store(T_CHAR); break;
2983 case Bytecodes::_iastore: array_store(T_INT); break;
2984 case Bytecodes::_sastore: array_store(T_SHORT); break;
2985 case Bytecodes::_fastore: array_store(T_FLOAT); break;
2986 case Bytecodes::_aastore: array_store(T_OBJECT); break;
2987 case Bytecodes::_lastore: array_store(T_LONG); break;
2988 case Bytecodes::_dastore: array_store(T_DOUBLE); break;
2989
2990 case Bytecodes::_getfield:
2991 do_getfield();
2992 break;
2993
2994 case Bytecodes::_getstatic:
2995 do_getstatic();
2996 break;
2997
2998 case Bytecodes::_putfield:
2999 do_putfield();
3000 break;
3001
3002 case Bytecodes::_putstatic:
3003 do_putstatic();
3004 break;
3005
3006 case Bytecodes::_irem:
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 ModINode(control(), a, b)));
3014 break;
3015 case Bytecodes::_idiv:
3016 // Must keep both values on the expression-stack during null-check
3017 zero_check_int(peek());
3018 // Compile-time detect of null-exception?
3019 if (stopped()) return;
3020 b = pop();
3021 a = pop();
3022 push( _gvn.transform( new DivINode(control(),a,b) ) );
3023 break;
3024 case Bytecodes::_imul:
3025 b = pop(); a = pop();
3026 push( _gvn.transform( new MulINode(a,b) ) );
3027 break;
3028 case Bytecodes::_iadd:
3029 b = pop(); a = pop();
3030 push( _gvn.transform( new AddINode(a,b) ) );
3031 break;
3032 case Bytecodes::_ineg:
3033 a = pop();
3034 push( _gvn.transform( new SubINode(_gvn.intcon(0),a)) );
3035 break;
3036 case Bytecodes::_isub:
3037 b = pop(); a = pop();
3038 push( _gvn.transform( new SubINode(a,b) ) );
3039 break;
3040 case Bytecodes::_iand:
3041 b = pop(); a = pop();
3042 push( _gvn.transform( new AndINode(a,b) ) );
3043 break;
3044 case Bytecodes::_ior:
3045 b = pop(); a = pop();
3046 push( _gvn.transform( new OrINode(a,b) ) );
3047 break;
3048 case Bytecodes::_ixor:
3049 b = pop(); a = pop();
3050 push( _gvn.transform( new XorINode(a,b) ) );
3051 break;
3052 case Bytecodes::_ishl:
3053 b = pop(); a = pop();
3054 push( _gvn.transform( new LShiftINode(a,b) ) );
3055 break;
3056 case Bytecodes::_ishr:
3057 b = pop(); a = pop();
3058 push( _gvn.transform( new RShiftINode(a,b) ) );
3059 break;
3060 case Bytecodes::_iushr:
3061 b = pop(); a = pop();
3062 push( _gvn.transform( new URShiftINode(a,b) ) );
3063 break;
3064
3065 case Bytecodes::_fneg:
3066 a = pop();
3067 b = _gvn.transform(new NegFNode (a));
3068 push(b);
3069 break;
3070
3071 case Bytecodes::_fsub:
3072 b = pop();
3073 a = pop();
3074 c = _gvn.transform( new SubFNode(a,b) );
3075 push(c);
3076 break;
3077
3078 case Bytecodes::_fadd:
3079 b = pop();
3080 a = pop();
3081 c = _gvn.transform( new AddFNode(a,b) );
3082 push(c);
3083 break;
3084
3085 case Bytecodes::_fmul:
3086 b = pop();
3087 a = pop();
3088 c = _gvn.transform( new MulFNode(a,b) );
3089 push(c);
3090 break;
3091
3092 case Bytecodes::_fdiv:
3093 b = pop();
3094 a = pop();
3095 c = _gvn.transform( new DivFNode(nullptr,a,b) );
3096 push(c);
3097 break;
3098
3099 case Bytecodes::_frem:
3100 // Generate a ModF node.
3101 b = pop();
3102 a = pop();
3103 push(floating_point_mod(a, b, BasicType::T_FLOAT));
3104 break;
3105
3106 case Bytecodes::_fcmpl:
3107 b = pop();
3108 a = pop();
3109 c = _gvn.transform( new CmpF3Node( a, b));
3110 push(c);
3111 break;
3112 case Bytecodes::_fcmpg:
3113 b = pop();
3114 a = pop();
3115
3116 // Same as fcmpl but need to flip the unordered case. Swap the inputs,
3117 // which negates the result sign except for unordered. Flip the unordered
3118 // as well by using CmpF3 which implements unordered-lesser instead of
3119 // unordered-greater semantics. Finally, commute the result bits. Result
3120 // is same as using a CmpF3Greater except we did it with CmpF3 alone.
3121 c = _gvn.transform( new CmpF3Node( b, a));
3122 c = _gvn.transform( new SubINode(_gvn.intcon(0),c) );
3123 push(c);
3124 break;
3125
3126 case Bytecodes::_f2i:
3127 a = pop();
3128 push(_gvn.transform(new ConvF2INode(a)));
3129 break;
3130
3131 case Bytecodes::_d2i:
3132 a = pop_pair();
3133 b = _gvn.transform(new ConvD2INode(a));
3134 push( b );
3135 break;
3136
3137 case Bytecodes::_f2d:
3138 a = pop();
3139 b = _gvn.transform( new ConvF2DNode(a));
3140 push_pair( b );
3141 break;
3142
3143 case Bytecodes::_d2f:
3144 a = pop_pair();
3145 b = _gvn.transform( new ConvD2FNode(a));
3146 push( b );
3147 break;
3148
3149 case Bytecodes::_l2f:
3150 if (Matcher::convL2FSupported()) {
3151 a = pop_pair();
3152 b = _gvn.transform( new ConvL2FNode(a));
3153 push(b);
3154 } else {
3155 l2f();
3156 }
3157 break;
3158
3159 case Bytecodes::_l2d:
3160 a = pop_pair();
3161 b = _gvn.transform( new ConvL2DNode(a));
3162 push_pair(b);
3163 break;
3164
3165 case Bytecodes::_f2l:
3166 a = pop();
3167 b = _gvn.transform( new ConvF2LNode(a));
3168 push_pair(b);
3169 break;
3170
3171 case Bytecodes::_d2l:
3172 a = pop_pair();
3173 b = _gvn.transform( new ConvD2LNode(a));
3174 push_pair(b);
3175 break;
3176
3177 case Bytecodes::_dsub:
3178 b = pop_pair();
3179 a = pop_pair();
3180 c = _gvn.transform( new SubDNode(a,b) );
3181 push_pair(c);
3182 break;
3183
3184 case Bytecodes::_dadd:
3185 b = pop_pair();
3186 a = pop_pair();
3187 c = _gvn.transform( new AddDNode(a,b) );
3188 push_pair(c);
3189 break;
3190
3191 case Bytecodes::_dmul:
3192 b = pop_pair();
3193 a = pop_pair();
3194 c = _gvn.transform( new MulDNode(a,b) );
3195 push_pair(c);
3196 break;
3197
3198 case Bytecodes::_ddiv:
3199 b = pop_pair();
3200 a = pop_pair();
3201 c = _gvn.transform( new DivDNode(nullptr,a,b) );
3202 push_pair(c);
3203 break;
3204
3205 case Bytecodes::_dneg:
3206 a = pop_pair();
3207 b = _gvn.transform(new NegDNode (a));
3208 push_pair(b);
3209 break;
3210
3211 case Bytecodes::_drem:
3212 // Generate a ModD node.
3213 b = pop_pair();
3214 a = pop_pair();
3215 push_pair(floating_point_mod(a, b, BasicType::T_DOUBLE));
3216 break;
3217
3218 case Bytecodes::_dcmpl:
3219 b = pop_pair();
3220 a = pop_pair();
3221 c = _gvn.transform( new CmpD3Node( a, b));
3222 push(c);
3223 break;
3224
3225 case Bytecodes::_dcmpg:
3226 b = pop_pair();
3227 a = pop_pair();
3228 // Same as dcmpl but need to flip the unordered case.
3229 // Commute the inputs, which negates the result sign except for unordered.
3230 // Flip the unordered as well by using CmpD3 which implements
3231 // unordered-lesser instead of unordered-greater semantics.
3232 // Finally, negate the result bits. Result is same as using a
3233 // CmpD3Greater except we did it with CmpD3 alone.
3234 c = _gvn.transform( new CmpD3Node( b, a));
3235 c = _gvn.transform( new SubINode(_gvn.intcon(0),c) );
3236 push(c);
3237 break;
3238
3239
3240 // Note for longs -> lo word is on TOS, hi word is on TOS - 1
3241 case Bytecodes::_land:
3242 b = pop_pair();
3243 a = pop_pair();
3244 c = _gvn.transform( new AndLNode(a,b) );
3245 push_pair(c);
3246 break;
3247 case Bytecodes::_lor:
3248 b = pop_pair();
3249 a = pop_pair();
3250 c = _gvn.transform( new OrLNode(a,b) );
3251 push_pair(c);
3252 break;
3253 case Bytecodes::_lxor:
3254 b = pop_pair();
3255 a = pop_pair();
3256 c = _gvn.transform( new XorLNode(a,b) );
3257 push_pair(c);
3258 break;
3259
3260 case Bytecodes::_lshl:
3261 b = pop(); // the shift count
3262 a = pop_pair(); // value to be shifted
3263 c = _gvn.transform( new LShiftLNode(a,b) );
3264 push_pair(c);
3265 break;
3266 case Bytecodes::_lshr:
3267 b = pop(); // the shift count
3268 a = pop_pair(); // value to be shifted
3269 c = _gvn.transform( new RShiftLNode(a,b) );
3270 push_pair(c);
3271 break;
3272 case Bytecodes::_lushr:
3273 b = pop(); // the shift count
3274 a = pop_pair(); // value to be shifted
3275 c = _gvn.transform( new URShiftLNode(a,b) );
3276 push_pair(c);
3277 break;
3278 case Bytecodes::_lmul:
3279 b = pop_pair();
3280 a = pop_pair();
3281 c = _gvn.transform( new MulLNode(a,b) );
3282 push_pair(c);
3283 break;
3284
3285 case Bytecodes::_lrem:
3286 // Must keep both values on the expression-stack during null-check
3287 assert(peek(0) == top(), "long word order");
3288 zero_check_long(peek(1));
3289 // Compile-time detect of null-exception?
3290 if (stopped()) return;
3291 b = pop_pair();
3292 a = pop_pair();
3293 c = _gvn.transform( new ModLNode(control(),a,b) );
3294 push_pair(c);
3295 break;
3296
3297 case Bytecodes::_ldiv:
3298 // Must keep both values on the expression-stack during null-check
3299 assert(peek(0) == top(), "long word order");
3300 zero_check_long(peek(1));
3301 // Compile-time detect of null-exception?
3302 if (stopped()) return;
3303 b = pop_pair();
3304 a = pop_pair();
3305 c = _gvn.transform( new DivLNode(control(),a,b) );
3306 push_pair(c);
3307 break;
3308
3309 case Bytecodes::_ladd:
3310 b = pop_pair();
3311 a = pop_pair();
3312 c = _gvn.transform( new AddLNode(a,b) );
3313 push_pair(c);
3314 break;
3315 case Bytecodes::_lsub:
3316 b = pop_pair();
3317 a = pop_pair();
3318 c = _gvn.transform( new SubLNode(a,b) );
3319 push_pair(c);
3320 break;
3321 case Bytecodes::_lcmp:
3322 // Safepoints are now inserted _before_ branches. The long-compare
3323 // bytecode painfully produces a 3-way value (-1,0,+1) which requires a
3324 // slew of control flow. These are usually followed by a CmpI vs zero and
3325 // a branch; this pattern then optimizes to the obvious long-compare and
3326 // branch. However, if the branch is backwards there's a Safepoint
3327 // inserted. The inserted Safepoint captures the JVM state at the
3328 // pre-branch point, i.e. it captures the 3-way value. Thus if a
3329 // long-compare is used to control a loop the debug info will force
3330 // computation of the 3-way value, even though the generated code uses a
3331 // long-compare and branch. We try to rectify the situation by inserting
3332 // a SafePoint here and have it dominate and kill the safepoint added at a
3333 // following backwards branch. At this point the JVM state merely holds 2
3334 // longs but not the 3-way value.
3335 switch (iter().next_bc()) {
3336 case Bytecodes::_ifgt:
3337 case Bytecodes::_iflt:
3338 case Bytecodes::_ifge:
3339 case Bytecodes::_ifle:
3340 case Bytecodes::_ifne:
3341 case Bytecodes::_ifeq:
3342 // If this is a backwards branch in the bytecodes, add Safepoint
3343 maybe_add_safepoint(iter().next_get_dest());
3344 default:
3345 break;
3346 }
3347 b = pop_pair();
3348 a = pop_pair();
3349 c = _gvn.transform( new CmpL3Node( a, b ));
3350 push(c);
3351 break;
3352
3353 case Bytecodes::_lneg:
3354 a = pop_pair();
3355 b = _gvn.transform( new SubLNode(longcon(0),a));
3356 push_pair(b);
3357 break;
3358 case Bytecodes::_l2i:
3359 a = pop_pair();
3360 push( _gvn.transform( new ConvL2INode(a)));
3361 break;
3362 case Bytecodes::_i2l:
3363 a = pop();
3364 b = _gvn.transform( new ConvI2LNode(a));
3365 push_pair(b);
3366 break;
3367 case Bytecodes::_i2b:
3368 // Sign extend
3369 a = pop();
3370 a = Compile::narrow_value(T_BYTE, a, nullptr, &_gvn, true);
3371 push(a);
3372 break;
3373 case Bytecodes::_i2s:
3374 a = pop();
3375 a = Compile::narrow_value(T_SHORT, a, nullptr, &_gvn, true);
3376 push(a);
3377 break;
3378 case Bytecodes::_i2c:
3379 a = pop();
3380 a = Compile::narrow_value(T_CHAR, a, nullptr, &_gvn, true);
3381 push(a);
3382 break;
3383
3384 case Bytecodes::_i2f:
3385 a = pop();
3386 b = _gvn.transform( new ConvI2FNode(a) ) ;
3387 push(b);
3388 break;
3389
3390 case Bytecodes::_i2d:
3391 a = pop();
3392 b = _gvn.transform( new ConvI2DNode(a));
3393 push_pair(b);
3394 break;
3395
3396 case Bytecodes::_iinc: // Increment local
3397 i = iter().get_index(); // Get local index
3398 set_local( i, _gvn.transform( new AddINode( _gvn.intcon(iter().get_iinc_con()), local(i) ) ) );
3399 break;
3400
3401 // Exit points of synchronized methods must have an unlock node
3402 case Bytecodes::_return:
3403 return_current(nullptr);
3404 break;
3405
3406 case Bytecodes::_ireturn:
3407 case Bytecodes::_areturn:
3408 case Bytecodes::_freturn:
3409 return_current(pop());
3410 break;
3411 case Bytecodes::_lreturn:
3412 return_current(pop_pair());
3413 break;
3414 case Bytecodes::_dreturn:
3415 return_current(pop_pair());
3416 break;
3417
3418 case Bytecodes::_athrow:
3419 // null exception oop throws null pointer exception
3420 null_check(peek());
3421 if (stopped()) return;
3422 // Hook the thrown exception directly to subsequent handlers.
3423 if (BailoutToInterpreterForThrows) {
3424 // Keep method interpreted from now on.
3425 uncommon_trap(Deoptimization::Reason_unhandled,
3426 Deoptimization::Action_make_not_compilable);
3427 return;
3428 }
3429 if (env()->jvmti_can_post_on_exceptions()) {
3430 // check if we must post exception events, take uncommon trap if so (with must_throw = false)
3431 uncommon_trap_if_should_post_on_exceptions(Deoptimization::Reason_unhandled, false);
3432 }
3433 // Here if either can_post_on_exceptions or should_post_on_exceptions is false
3434 add_exception_state(make_exception_state(peek()));
3435 break;
3436
3437 case Bytecodes::_goto: // fall through
3438 case Bytecodes::_goto_w: {
3439 int target_bci = (bc() == Bytecodes::_goto) ? iter().get_dest() : iter().get_far_dest();
3440
3441 // If this is a backwards branch in the bytecodes, add Safepoint
3442 maybe_add_safepoint(target_bci);
3443
3444 // Merge the current control into the target basic block
3445 merge(target_bci);
3446
3447 // See if we can get some profile data and hand it off to the next block
3448 Block *target_block = block()->successor_for_bci(target_bci);
3449 if (target_block->pred_count() != 1) break;
3450 ciMethodData* methodData = method()->method_data();
3451 if (!methodData->is_mature()) break;
3452 ciProfileData* data = methodData->bci_to_data(bci());
3453 assert(data != nullptr && data->is_JumpData(), "need JumpData for taken branch");
3454 int taken = ((ciJumpData*)data)->taken();
3455 taken = method()->scale_count(taken);
3456 target_block->set_count(taken);
3457 break;
3458 }
3459
3460 case Bytecodes::_ifnull: btest = BoolTest::eq; goto handle_if_null;
3461 case Bytecodes::_ifnonnull: btest = BoolTest::ne; goto handle_if_null;
3462 handle_if_null:
3463 // If this is a backwards branch in the bytecodes, add Safepoint
3464 maybe_add_safepoint(iter().get_dest());
3465 a = null();
3466 b = pop();
3467 if (b->is_InlineType()) {
3468 // Null checking a scalarized but nullable inline type. Check the IsInit
3469 // input instead of the oop input to avoid keeping buffer allocations alive
3470 c = _gvn.transform(new CmpINode(b->as_InlineType()->get_is_init(), zerocon(T_INT)));
3471 } else {
3472 if (!_gvn.type(b)->speculative_maybe_null() &&
3473 !too_many_traps(Deoptimization::Reason_speculate_null_check)) {
3474 inc_sp(1);
3475 Node* null_ctl = top();
3476 b = null_check_oop(b, &null_ctl, true, true, true);
3477 assert(null_ctl->is_top(), "no null control here");
3478 dec_sp(1);
3479 } else if (_gvn.type(b)->speculative_always_null() &&
3480 !too_many_traps(Deoptimization::Reason_speculate_null_assert)) {
3481 inc_sp(1);
3482 b = null_assert(b);
3483 dec_sp(1);
3484 }
3485 c = _gvn.transform( new CmpPNode(b, a) );
3486 }
3487 do_ifnull(btest, c);
3488 break;
3489
3490 case Bytecodes::_if_acmpeq: btest = BoolTest::eq; goto handle_if_acmp;
3491 case Bytecodes::_if_acmpne: btest = BoolTest::ne; goto handle_if_acmp;
3492 handle_if_acmp:
3493 // If this is a backwards branch in the bytecodes, add Safepoint
3494 maybe_add_safepoint(iter().get_dest());
3495 a = pop();
3496 b = pop();
3497 do_acmp(btest, b, a);
3498 break;
3499
3500 case Bytecodes::_ifeq: btest = BoolTest::eq; goto handle_ifxx;
3501 case Bytecodes::_ifne: btest = BoolTest::ne; goto handle_ifxx;
3502 case Bytecodes::_iflt: btest = BoolTest::lt; goto handle_ifxx;
3503 case Bytecodes::_ifle: btest = BoolTest::le; goto handle_ifxx;
3504 case Bytecodes::_ifgt: btest = BoolTest::gt; goto handle_ifxx;
3505 case Bytecodes::_ifge: btest = BoolTest::ge; goto handle_ifxx;
3506 handle_ifxx:
3507 // If this is a backwards branch in the bytecodes, add Safepoint
3508 maybe_add_safepoint(iter().get_dest());
3509 a = _gvn.intcon(0);
3510 b = pop();
3511 c = _gvn.transform( new CmpINode(b, a) );
3512 do_if(btest, c);
3513 break;
3514
3515 case Bytecodes::_if_icmpeq: btest = BoolTest::eq; goto handle_if_icmp;
3516 case Bytecodes::_if_icmpne: btest = BoolTest::ne; goto handle_if_icmp;
3517 case Bytecodes::_if_icmplt: btest = BoolTest::lt; goto handle_if_icmp;
3518 case Bytecodes::_if_icmple: btest = BoolTest::le; goto handle_if_icmp;
3519 case Bytecodes::_if_icmpgt: btest = BoolTest::gt; goto handle_if_icmp;
3520 case Bytecodes::_if_icmpge: btest = BoolTest::ge; goto handle_if_icmp;
3521 handle_if_icmp:
3522 // If this is a backwards branch in the bytecodes, add Safepoint
3523 maybe_add_safepoint(iter().get_dest());
3524 a = pop();
3525 b = pop();
3526 c = _gvn.transform( new CmpINode( b, a ) );
3527 do_if(btest, c);
3528 break;
3529
3530 case Bytecodes::_tableswitch:
3531 do_tableswitch();
3532 break;
3533
3534 case Bytecodes::_lookupswitch:
3535 do_lookupswitch();
3536 break;
3537
3538 case Bytecodes::_invokestatic:
3539 case Bytecodes::_invokedynamic:
3540 case Bytecodes::_invokespecial:
3541 case Bytecodes::_invokevirtual:
3542 case Bytecodes::_invokeinterface:
3543 do_call();
3544 break;
3545 case Bytecodes::_checkcast:
3546 do_checkcast();
3547 break;
3548 case Bytecodes::_instanceof:
3549 do_instanceof();
3550 break;
3551 case Bytecodes::_anewarray:
3552 do_newarray();
3553 break;
3554 case Bytecodes::_newarray:
3555 do_newarray((BasicType)iter().get_index());
3556 break;
3557 case Bytecodes::_multianewarray:
3558 do_multianewarray();
3559 break;
3560 case Bytecodes::_new:
3561 do_new();
3562 break;
3563
3564 case Bytecodes::_jsr:
3565 case Bytecodes::_jsr_w:
3566 do_jsr();
3567 break;
3568
3569 case Bytecodes::_ret:
3570 do_ret();
3571 break;
3572
3573
3574 case Bytecodes::_monitorenter:
3575 do_monitor_enter();
3576 break;
3577
3578 case Bytecodes::_monitorexit:
3579 do_monitor_exit();
3580 break;
3581
3582 case Bytecodes::_breakpoint:
3583 // Breakpoint set concurrently to compile
3584 // %%% use an uncommon trap?
3585 C->record_failure("breakpoint in method");
3586 return;
3587
3588 default:
3589 #ifndef PRODUCT
3590 map()->dump(99);
3591 #endif
3592 tty->print("\nUnhandled bytecode %s\n", Bytecodes::name(bc()) );
3593 ShouldNotReachHere();
3594 }
3595
3596 #ifndef PRODUCT
3597 if (failing()) { return; }
3598 constexpr int perBytecode = 6;
3599 if (C->should_print_igv(perBytecode)) {
3600 IdealGraphPrinter* printer = C->igv_printer();
3601 char buffer[256];
3602 jio_snprintf(buffer, sizeof(buffer), "Bytecode %d: %s", bci(), Bytecodes::name(bc()));
3603 bool old = printer->traverse_outs();
3604 printer->set_traverse_outs(true);
3605 printer->print_graph(buffer);
3606 printer->set_traverse_outs(old);
3607 }
3608 #endif
3609 }