1 /*
2 * Copyright (c) 1998, 2026, 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/ciInlineKlass.hpp"
26 #include "ci/ciMethodData.hpp"
27 #include "ci/ciSymbols.hpp"
28 #include "classfile/vmSymbols.hpp"
29 #include "compiler/compileLog.hpp"
30 #include "interpreter/linkResolver.hpp"
31 #include "jvm_io.h"
32 #include "memory/resourceArea.hpp"
33 #include "memory/universe.hpp"
34 #include "oops/oop.inline.hpp"
35 #include "opto/addnode.hpp"
36 #include "opto/castnode.hpp"
37 #include "opto/convertnode.hpp"
38 #include "opto/divnode.hpp"
39 #include "opto/idealGraphPrinter.hpp"
40 #include "opto/idealKit.hpp"
41 #include "opto/inlinetypenode.hpp"
42 #include "opto/matcher.hpp"
43 #include "opto/memnode.hpp"
44 #include "opto/mulnode.hpp"
45 #include "opto/opaquenode.hpp"
46 #include "opto/parse.hpp"
47 #include "opto/runtime.hpp"
48 #include "opto/subtypenode.hpp"
49 #include "runtime/arguments.hpp"
50 #include "runtime/deoptimization.hpp"
51 #include "runtime/globals.hpp"
52 #include "runtime/sharedRuntime.hpp"
53
54 #ifndef PRODUCT
55 extern uint explicit_null_checks_inserted,
56 explicit_null_checks_elided;
57 #endif
58
59 Node* Parse::record_profile_for_speculation_at_array_load(Node* ld) {
60 // Feed unused profile data to type speculation
61 if (UseTypeSpeculation && UseArrayLoadStoreProfile) {
62 ciKlass* array_type = nullptr;
63 ciKlass* element_type = nullptr;
64 ProfilePtrKind element_ptr = ProfileMaybeNull;
65 bool flat_array = true;
66 bool null_free_array = true;
67 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
68 if (element_type != nullptr || element_ptr != ProfileMaybeNull) {
69 ld = record_profile_for_speculation(ld, element_type, element_ptr);
70 }
71 }
72 return ld;
73 }
74
75
76 //---------------------------------array_load----------------------------------
77 void Parse::array_load(BasicType bt) {
78 const Type* elemtype = Type::TOP;
79 Node* prep_array = prepare_array_addressing(bt, 0, elemtype);
80 if (stopped()) return; // guaranteed null or range check
81
82 Node* array_index = pop();
83 Node* array = pop();
84
85 // Handle inline type arrays
86 const TypeOopPtr* element_ptr = elemtype->make_oopptr();
87 const TypeAryPtr* array_type = _gvn.type(array)->is_aryptr();
88
89 if (!array_type->is_not_flat()) {
90 // Cannot statically determine if array is a flat array, emit runtime check
91 assert(UseArrayFlattening && is_reference_type(bt) && element_ptr->can_be_inline_type() &&
92 (!element_ptr->is_inlinetypeptr() || element_ptr->inline_klass()->maybe_flat_in_array()), "array can't be flat");
93 IdealKit ideal(this);
94 IdealVariable res(ideal);
95 ideal.declarations_done();
96 ideal.if_then(flat_array_test(array, /* flat = */ false)); {
97 // Non-flat array
98 sync_kit(ideal);
99 if (!array_type->is_flat()) {
100 assert(array_type->is_flat() || control()->in(0)->as_If()->is_flat_array_check(&_gvn), "Should be found");
101 // Loading from a non-flat array, casting array to not flat.
102 const TypeAryPtr* ary_type = _gvn.type(prep_array)->is_aryptr();
103 ary_type = ary_type->cast_to_not_flat();
104 Node* not_flat_ary = _gvn.transform(new CheckCastPPNode(control(), prep_array, ary_type));
105 Node* adr = get_ptr_to_array_element(not_flat_ary, array_index, bt, ary_type->size(), control());
106 const TypeAryPtr* adr_type = TypeAryPtr::get_array_body_type(bt);
107 DecoratorSet decorator_set = IN_HEAP | IS_ARRAY | C2_CONTROL_DEPENDENT_LOAD;
108 if (needs_range_check(ary_type->size(), array_index)) {
109 // We've emitted a RangeCheck but now insert an additional check between the range check and the actual load.
110 // We cannot pin the load to two separate nodes. Instead, we pin it conservatively here such that it cannot
111 // possibly float above the range check at any point.
112 decorator_set |= C2_UNKNOWN_CONTROL_LOAD;
113 }
114 Node* ld = access_load_at(not_flat_ary, adr, adr_type, element_ptr, bt, decorator_set);
115 if (element_ptr->is_inlinetypeptr()) {
116 ld = InlineTypeNode::make_from_oop(this, ld, element_ptr->inline_klass());
117 }
118 ideal.set(res, ld);
119 }
120 ideal.sync_kit(this);
121 } ideal.else_(); {
122 // Flat array
123 sync_kit(ideal);
124 if (!array_type->is_not_flat()) {
125 if (element_ptr->is_inlinetypeptr()) {
126 ciInlineKlass* vk = element_ptr->inline_klass();
127 Node* flat_array = cast_to_flat_array(array, vk);
128 Node* vt = InlineTypeNode::make_from_flat_array(this, vk, flat_array, array_index);
129 ideal.set(res, vt);
130 } else {
131 // Element type is unknown, and thus we cannot statically determine the exact flat array layout. Emit a
132 // runtime call to correctly load the inline type element from the flat array.
133 Node* inline_type = load_from_unknown_flat_array(array, array_index, element_ptr);
134 bool is_null_free = array_type->is_null_free() ||
135 (!UseNullableAtomicValueFlattening && !UseNullableNonAtomicValueFlattening);
136 if (is_null_free) {
137 inline_type = cast_not_null(inline_type);
138 }
139 ideal.set(res, inline_type);
140 }
141 }
142 ideal.sync_kit(this);
143 } ideal.end_if();
144 sync_kit(ideal);
145 Node* ld = _gvn.transform(ideal.value(res));
146 ld = record_profile_for_speculation_at_array_load(ld);
147 push_node(bt, ld);
148 return;
149 }
150
151 if (elemtype == TypeInt::BOOL) {
152 bt = T_BOOLEAN;
153 }
154 const TypeAryPtr* adr_type = TypeAryPtr::get_array_body_type(bt);
155 Node* adr = get_ptr_to_array_element(prep_array, array_index, bt, array_type->size(), control());
156 Node* ld = access_load_at(array, adr, adr_type, elemtype, bt,
157 IN_HEAP | IS_ARRAY | C2_CONTROL_DEPENDENT_LOAD);
158 ld = record_profile_for_speculation_at_array_load(ld);
159 // Loading an inline type from a non-flat array
160 if (element_ptr != nullptr && element_ptr->is_inlinetypeptr()) {
161 assert(!array_type->is_null_free() || !element_ptr->maybe_null(), "inline type array elements should never be null");
162 ld = InlineTypeNode::make_from_oop(this, ld, element_ptr->inline_klass());
163 }
164 push_node(bt, ld);
165 }
166
167 Node* Parse::load_from_unknown_flat_array(Node* array, Node* array_index, const TypeOopPtr* element_ptr) {
168 // Below membars keep this access to an unknown flat array correctly
169 // ordered with other unknown and known flat array accesses.
170 insert_mem_bar_volatile(Op_MemBarCPUOrder, C->get_alias_index(TypeAryPtr::INLINES));
171
172 Node* call = nullptr;
173 {
174 // Re-execute flat array load if runtime call triggers deoptimization
175 PreserveReexecuteState preexecs(this);
176 jvms()->set_bci(_bci);
177 jvms()->set_should_reexecute(true);
178 inc_sp(2);
179 kill_dead_locals();
180 call = make_runtime_call(RC_NO_LEAF | RC_NO_IO,
181 OptoRuntime::load_unknown_inline_Type(),
182 OptoRuntime::load_unknown_inline_Java(),
183 nullptr, TypeRawPtr::BOTTOM,
184 array, array_index);
185 }
186 make_slow_call_ex(call, env()->Throwable_klass(), false);
187 Node* buffer = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
188
189 insert_mem_bar_volatile(Op_MemBarCPUOrder, C->get_alias_index(TypeAryPtr::INLINES));
190
191 // Keep track of the information that the inline type is in flat arrays
192 const Type* unknown_value = element_ptr->is_instptr()->cast_to_flat_in_array();
193 return _gvn.transform(new CheckCastPPNode(control(), buffer, unknown_value));
194 }
195
196 //--------------------------------array_store----------------------------------
197 void Parse::array_store(BasicType bt) {
198 const Type* elemtype = Type::TOP;
199 Node* prep_array = prepare_array_addressing(bt, type2size[bt], elemtype);
200 if (stopped()) return; // guaranteed null or range check
201
202 Node* adr = get_ptr_to_array_element(prep_array, /* index */peek(0+type2size[bt]), bt,
203 _gvn.type(prep_array)->is_aryptr()->size(), control());
204
205 Node* stored_value_casted = nullptr;
206 if (bt == T_OBJECT) {
207 stored_value_casted = array_store_check(adr, elemtype);
208 if (stopped()) {
209 return;
210 }
211 }
212 Node* const stored_value = pop_node(bt); // Value to store
213 Node* const array_index = pop(); // Index in the array
214 Node* array = pop(); // The array itself
215
216 const TypeAryPtr* array_type = _gvn.type(array)->is_aryptr();
217 const TypeAryPtr* adr_type = TypeAryPtr::get_array_body_type(bt);
218
219 if (elemtype == TypeInt::BOOL) {
220 bt = T_BOOLEAN;
221 } else if (bt == T_OBJECT) {
222 elemtype = elemtype->make_oopptr();
223 const Type* stored_value_casted_type = _gvn.type(stored_value_casted);
224 // Based on the value to be stored, try to determine if the array is not null-free and/or not flat.
225 // This is only legal for non-null stores because the array_store_check always passes for null, even
226 // if the array is null-free. Null stores are handled in GraphKit::inline_array_null_guard().
227 bool not_inline = !stored_value_casted_type->maybe_null() && !stored_value_casted_type->is_oopptr()->can_be_inline_type();
228 bool not_null_free = not_inline;
229 bool not_flat = not_inline || ( stored_value_casted_type->is_inlinetypeptr() &&
230 !stored_value_casted_type->inline_klass()->maybe_flat_in_array());
231 if (!array_type->is_not_null_free() && not_null_free) {
232 // Storing a non-inline type, mark array as not null-free.
233 array_type = array_type->cast_to_not_null_free();
234 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, array_type));
235 replace_in_map(array, cast);
236 array = cast;
237 }
238 if (!array_type->is_not_flat() && not_flat) {
239 // Storing to a non-flat array, mark array as not flat.
240 array_type = array_type->cast_to_not_flat();
241 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, array_type));
242 replace_in_map(array, cast);
243 array = cast;
244 }
245
246 if (array_type->is_null_free() && elemtype->is_inlinetypeptr() && elemtype->inline_klass()->is_empty()) {
247 // Array of null-free empty inline type, there is only 1 state for the elements
248 assert(!stored_value_casted_type->maybe_null(), "should be guaranteed by array store check");
249 return;
250 }
251
252 if (!array_type->is_not_flat()) {
253 // Array might be a flat array, emit runtime checks (for null, a simple inline_array_null_guard is sufficient).
254 assert(UseArrayFlattening && !not_flat && elemtype->is_oopptr()->can_be_inline_type() &&
255 (!array_type->klass_is_exact() || array_type->is_flat()), "array can't be a flat array");
256 // 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.
257 array = inline_array_null_guard(array, stored_value_casted, 3, true);
258 // Reload array type which could have been updated by inline_array_null_guard().
259 array_type = _gvn.type(array)->is_aryptr();
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_not_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 Node* flat_array = cast_to_flat_array(array, vk);
286
287 // Re-execute flat array store if buffering triggers deoptimization
288 PreserveReexecuteState preexecs(this);
289 jvms()->set_should_reexecute(true);
290 inc_sp(3);
291
292 if (!stored_value_casted->is_InlineType()) {
293 assert(_gvn.type(stored_value_casted) == TypePtr::NULL_PTR, "Unexpected value");
294 stored_value_casted = InlineTypeNode::make_null(_gvn, vk);
295 }
296
297 stored_value_casted->as_InlineType()->store_flat_array(this, flat_array, array_index);
298 } else {
299 // Element type is unknown, emit a runtime call since the flat array layout is not statically known.
300 store_to_unknown_flat_array(array, array_index, stored_value_casted);
301 }
302 }
303 ideal.sync_kit(this);
304 }
305 ideal.end_if();
306 sync_kit(ideal);
307 return;
308 } else if (!array_type->is_not_null_free()) {
309 // Array is not flat but may be null free
310 assert(elemtype->is_oopptr()->can_be_inline_type(), "array can't be null-free");
311 array = inline_array_null_guard(array, stored_value_casted, 3, true);
312 }
313 }
314 inc_sp(3);
315 access_store_at(array, adr, adr_type, stored_value, elemtype, bt, MO_UNORDERED | IN_HEAP | IS_ARRAY);
316 dec_sp(3);
317 }
318
319 // Emit a runtime call to store to a flat array whose element type is either unknown (i.e. we do not know the flat
320 // array layout) or not exact (could have different flat array layouts at runtime).
321 void Parse::store_to_unknown_flat_array(Node* array, Node* const idx, Node* non_null_stored_value) {
322 // Below membars keep this access to an unknown flat array correctly
323 // ordered with other unknown and known flat array accesses.
324 insert_mem_bar_volatile(Op_MemBarCPUOrder, C->get_alias_index(TypeAryPtr::INLINES));
325
326 Node* call = nullptr;
327 {
328 // Re-execute flat array store if runtime call triggers deoptimization
329 PreserveReexecuteState preexecs(this);
330 jvms()->set_bci(_bci);
331 jvms()->set_should_reexecute(true);
332 inc_sp(3);
333 kill_dead_locals();
334 call = make_runtime_call(RC_NO_LEAF | RC_NO_IO,
335 OptoRuntime::store_unknown_inline_Type(),
336 OptoRuntime::store_unknown_inline_Java(),
337 nullptr, TypeRawPtr::BOTTOM,
338 non_null_stored_value, array, idx);
339 }
340 make_slow_call_ex(call, env()->Throwable_klass(), false);
341
342 insert_mem_bar_volatile(Op_MemBarCPUOrder, C->get_alias_index(TypeAryPtr::INLINES));
343 }
344
345 //------------------------------array_addressing-------------------------------
346 // Pull array and index from the stack. Compute pointer-to-element.
347 Node* Parse::prepare_array_addressing(BasicType type, int vals, const Type*& elemtype) {
348 Node *idx = peek(0+vals); // Get from stack without popping
349 Node *ary = peek(1+vals); // in case of exception
350
351 // Null check the array base, with correct stack contents
352 ary = null_check(ary, T_ARRAY);
353 // Compile-time detect of null-exception?
354 if (stopped()) return top();
355
356 const TypeAryPtr* arytype = _gvn.type(ary)->is_aryptr();
357 const TypeInt* sizetype = arytype->size();
358 elemtype = arytype->elem();
359
360 if (UseUniqueSubclasses) {
361 const Type* el = elemtype->make_ptr();
362 if (el && el->isa_instptr()) {
363 const TypeInstPtr* toop = el->is_instptr();
364 if (toop->instance_klass()->unique_concrete_subklass()) {
365 // If we load from "AbstractClass[]" we must see "ConcreteSubClass".
366 const Type* subklass = Type::get_const_type(toop->instance_klass());
367 elemtype = subklass->join_speculative(el);
368 }
369 }
370 }
371
372 if (!arytype->is_loaded()) {
373 // Only fails for some -Xcomp runs
374 // The class is unloaded. We have to run this bytecode in the interpreter.
375 ciKlass* klass = arytype->unloaded_klass();
376
377 uncommon_trap(Deoptimization::Reason_unloaded,
378 Deoptimization::Action_reinterpret,
379 klass, "!loaded array");
380 return top();
381 }
382
383 ary = create_speculative_inline_type_array_checks(ary, arytype, elemtype);
384
385 if (needs_range_check(sizetype, idx)) {
386 create_range_check(idx, ary, sizetype);
387 } else if (C->log() != nullptr) {
388 C->log()->elem("observe that='!need_range_check'");
389 }
390
391 // Check for always knowing you are throwing a range-check exception
392 if (stopped()) return top();
393
394 return ary;
395 }
396
397 Node* Parse::get_ptr_to_array_element(Node* array, Node* idx, BasicType elembt, const TypeInt* sizetype, Node* control) {
398 // Make array address computation control dependent to prevent it
399 // from floating above the range check during loop optimizations.
400 Node* ptr = array_element_address(array, idx, elembt, sizetype, control);
401 assert(ptr != top(), "top should go hand-in-hand with stopped");
402
403 return ptr;
404 }
405
406 // 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
407 // be greater or equal the smallest possible array size (i.e. out-of-bounds).
408 bool Parse::needs_range_check(const TypeInt* size_type, const Node* index) const {
409 const TypeInt* index_type = _gvn.type(index)->is_int();
410 return index_type->_hi >= size_type->_lo || index_type->_lo < 0;
411 }
412
413 void Parse::create_range_check(Node* idx, Node* ary, const TypeInt* sizetype) {
414 Node* tst;
415 if (sizetype->_hi <= 0) {
416 // The greatest array bound is negative, so we can conclude that we're
417 // compiling unreachable code, but the unsigned compare trick used below
418 // only works with non-negative lengths. Instead, hack "tst" to be zero so
419 // the uncommon_trap path will always be taken.
420 tst = _gvn.intcon(0);
421 } else {
422 // Range is constant in array-oop, so we can use the original state of mem
423 Node* len = load_array_length(ary);
424
425 // Test length vs index (standard trick using unsigned compare)
426 Node* chk = _gvn.transform(new CmpUNode(idx, len) );
427 BoolTest::mask btest = BoolTest::lt;
428 tst = _gvn.transform(new BoolNode(chk, btest) );
429 }
430 RangeCheckNode* rc = new RangeCheckNode(control(), tst, PROB_MAX, COUNT_UNKNOWN);
431 _gvn.set_type(rc, rc->Value(&_gvn));
432 if (!tst->is_Con()) {
433 record_for_igvn(rc);
434 }
435 set_control(_gvn.transform(new IfTrueNode(rc)));
436 // Branch to failure if out of bounds
437 {
438 PreserveJVMState pjvms(this);
439 set_control(_gvn.transform(new IfFalseNode(rc)));
440 if (C->allow_range_check_smearing()) {
441 // Do not use builtin_throw, since range checks are sometimes
442 // made more stringent by an optimistic transformation.
443 // This creates "tentative" range checks at this point,
444 // which are not guaranteed to throw exceptions.
445 // See IfNode::Ideal, is_range_check, adjust_check.
446 uncommon_trap(Deoptimization::Reason_range_check,
447 Deoptimization::Action_make_not_entrant,
448 nullptr, "range_check");
449 } else {
450 // If we have already recompiled with the range-check-widening
451 // heroic optimization turned off, then we must really be throwing
452 // range check exceptions.
453 builtin_throw(Deoptimization::Reason_range_check);
454 }
455 }
456 }
457
458 // For inline type arrays, we can use the profiling information for array accesses to speculate on the type, flatness,
459 // and null-freeness. We can either prepare the speculative type for later uses or emit explicit speculative checks with
460 // traps now. In the latter case, the speculative type guarantees can avoid additional runtime checks later (e.g.
461 // non-null-free implies non-flat which allows us to remove flatness checks). This makes the graph simpler.
462 Node* Parse::create_speculative_inline_type_array_checks(Node* array, const TypeAryPtr* array_type,
463 const Type*& element_type) {
464 if (!array_type->is_flat() && !array_type->is_not_flat()) {
465 // For arrays that might be flat, speculate that the array has the exact type reported in the profile data such that
466 // we can rely on a fixed memory layout (i.e. either a flat layout or not).
467 array = cast_to_speculative_array_type(array, array_type, element_type);
468 } else if (UseTypeSpeculation && UseArrayLoadStoreProfile) {
469 // Array is known to be either flat or not flat. If possible, update the speculative type by using the profile data
470 // at this bci.
471 array = cast_to_profiled_array_type(array);
472 }
473
474 // Even though the type does not tell us whether we have an inline type array or not, we can still check the profile data
475 // whether we have a non-null-free or non-flat array. Speculating on a non-null-free array doesn't help aaload but could
476 // be profitable for a subsequent aastore.
477 if (!array_type->is_null_free() && !array_type->is_not_null_free()) {
478 array = speculate_non_null_free_array(array, array_type);
479 }
480 if (!array_type->is_flat() && !array_type->is_not_flat()) {
481 array = speculate_non_flat_array(array, array_type);
482 }
483 return array;
484 }
485
486 // Speculate that the array has the exact type reported in the profile data. We emit a trap when this turns out to be
487 // wrong. On the fast path, we add a CheckCastPP to use the exact type.
488 Node* Parse::cast_to_speculative_array_type(Node* const array, const TypeAryPtr*& array_type, const Type*& element_type) {
489 Deoptimization::DeoptReason reason = Deoptimization::Reason_speculate_class_check;
490 ciKlass* speculative_array_type = array_type->speculative_type();
491 if (too_many_traps_or_recompiles(reason) || speculative_array_type == nullptr) {
492 // No speculative type, check profile data at this bci
493 speculative_array_type = nullptr;
494 reason = Deoptimization::Reason_class_check;
495 if (UseArrayLoadStoreProfile && !too_many_traps_or_recompiles(reason)) {
496 ciKlass* profiled_element_type = nullptr;
497 ProfilePtrKind element_ptr = ProfileMaybeNull;
498 bool flat_array = true;
499 bool null_free_array = true;
500 method()->array_access_profiled_type(bci(), speculative_array_type, profiled_element_type, element_ptr, flat_array,
501 null_free_array);
502 }
503 }
504 if (speculative_array_type != nullptr) {
505 // Speculate that this array has the exact type reported by profile data
506 Node* casted_array = nullptr;
507 DEBUG_ONLY(Node* old_control = control();)
508 Node* slow_ctl = type_check_receiver(array, speculative_array_type, 1.0, &casted_array);
509 if (stopped()) {
510 // The check always fails and therefore profile information is incorrect. Don't use it.
511 assert(old_control == slow_ctl, "type check should have been removed");
512 set_control(slow_ctl);
513 } else if (!slow_ctl->is_top()) {
514 { PreserveJVMState pjvms(this);
515 set_control(slow_ctl);
516 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
517 }
518 replace_in_map(array, casted_array);
519 array_type = _gvn.type(casted_array)->is_aryptr();
520 element_type = array_type->elem();
521 return casted_array;
522 }
523 }
524 return array;
525 }
526
527 // Create a CheckCastPP when the speculative type can improve the current type.
528 Node* Parse::cast_to_profiled_array_type(Node* const array) {
529 ciKlass* array_type = nullptr;
530 ciKlass* element_type = nullptr;
531 ProfilePtrKind element_ptr = ProfileMaybeNull;
532 bool flat_array = true;
533 bool null_free_array = true;
534 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
535 if (array_type != nullptr) {
536 return record_profile_for_speculation(array, array_type, ProfileMaybeNull);
537 }
538 return array;
539 }
540
541 // Speculate that the array is non-null-free. We emit a trap when this turns out to be
542 // wrong. On the fast path, we add a CheckCastPP to use the non-null-free type.
543 Node* Parse::speculate_non_null_free_array(Node* const array, const TypeAryPtr*& array_type) {
544 bool null_free_array = true;
545 Deoptimization::DeoptReason reason = Deoptimization::Reason_none;
546 if (array_type->speculative() != nullptr &&
547 array_type->speculative()->is_aryptr()->is_not_null_free() &&
548 !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_class_check)) {
549 null_free_array = false;
550 reason = Deoptimization::Reason_speculate_class_check;
551 } else if (UseArrayLoadStoreProfile && !too_many_traps_or_recompiles(Deoptimization::Reason_class_check)) {
552 ciKlass* profiled_array_type = nullptr;
553 ciKlass* profiled_element_type = nullptr;
554 ProfilePtrKind element_ptr = ProfileMaybeNull;
555 bool flat_array = true;
556 method()->array_access_profiled_type(bci(), profiled_array_type, profiled_element_type, element_ptr, flat_array,
557 null_free_array);
558 reason = Deoptimization::Reason_class_check;
559 }
560 if (!null_free_array) {
561 { // Deoptimize if null-free array
562 BuildCutout unless(this, null_free_array_test(array, /* null_free = */ false), PROB_MAX);
563 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
564 }
565 assert(!stopped(), "null-free array should have been caught earlier");
566 Node* casted_array = _gvn.transform(new CheckCastPPNode(control(), array, array_type->cast_to_not_null_free()));
567 replace_in_map(array, casted_array);
568 array_type = _gvn.type(casted_array)->is_aryptr();
569 return casted_array;
570 }
571 return array;
572 }
573
574 // Speculate that the array is non-flat. We emit a trap when this turns out to be wrong.
575 // On the fast path, we add a CheckCastPP to use the non-flat type.
576 Node* Parse::speculate_non_flat_array(Node* const array, const TypeAryPtr* const array_type) {
577 bool flat_array = true;
578 Deoptimization::DeoptReason reason = Deoptimization::Reason_none;
579 if (array_type->speculative() != nullptr &&
580 array_type->speculative()->is_aryptr()->is_not_flat() &&
581 !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_class_check)) {
582 flat_array = false;
583 reason = Deoptimization::Reason_speculate_class_check;
584 } else if (UseArrayLoadStoreProfile && !too_many_traps_or_recompiles(reason)) {
585 ciKlass* profiled_array_type = nullptr;
586 ciKlass* profiled_element_type = nullptr;
587 ProfilePtrKind element_ptr = ProfileMaybeNull;
588 bool null_free_array = true;
589 method()->array_access_profiled_type(bci(), profiled_array_type, profiled_element_type, element_ptr, flat_array,
590 null_free_array);
591 reason = Deoptimization::Reason_class_check;
592 }
593 if (!flat_array) {
594 { // Deoptimize if flat array
595 BuildCutout unless(this, flat_array_test(array, /* flat = */ false), PROB_MAX);
596 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
597 }
598 assert(!stopped(), "flat array should have been caught earlier");
599 Node* casted_array = _gvn.transform(new CheckCastPPNode(control(), array, array_type->cast_to_not_flat()));
600 replace_in_map(array, casted_array);
601 return casted_array;
602 }
603 return array;
604 }
605
606 // returns IfNode
607 IfNode* Parse::jump_if_fork_int(Node* a, Node* b, BoolTest::mask mask, float prob, float cnt) {
608 Node *cmp = _gvn.transform(new CmpINode(a, b)); // two cases: shiftcount > 32 and shiftcount <= 32
609 Node *tst = _gvn.transform(new BoolNode(cmp, mask));
610 IfNode *iff = create_and_map_if(control(), tst, prob, cnt);
611 return iff;
612 }
613
614
615 // sentinel value for the target bci to mark never taken branches
616 // (according to profiling)
617 static const int never_reached = INT_MAX;
618
619 //------------------------------helper for tableswitch-------------------------
620 void Parse::jump_if_true_fork(IfNode *iff, int dest_bci_if_true, bool unc) {
621 // True branch, use existing map info
622 { PreserveJVMState pjvms(this);
623 Node *iftrue = _gvn.transform( new IfTrueNode (iff) );
624 set_control( iftrue );
625 if (unc) {
626 repush_if_args();
627 uncommon_trap(Deoptimization::Reason_unstable_if,
628 Deoptimization::Action_reinterpret,
629 nullptr,
630 "taken always");
631 } else {
632 assert(dest_bci_if_true != never_reached, "inconsistent dest");
633 merge_new_path(dest_bci_if_true);
634 }
635 }
636
637 // False branch
638 Node *iffalse = _gvn.transform( new IfFalseNode(iff) );
639 set_control( iffalse );
640 }
641
642 void Parse::jump_if_false_fork(IfNode *iff, int dest_bci_if_true, bool unc) {
643 // True branch, use existing map info
644 { PreserveJVMState pjvms(this);
645 Node *iffalse = _gvn.transform( new IfFalseNode (iff) );
646 set_control( iffalse );
647 if (unc) {
648 repush_if_args();
649 uncommon_trap(Deoptimization::Reason_unstable_if,
650 Deoptimization::Action_reinterpret,
651 nullptr,
652 "taken never");
653 } else {
654 assert(dest_bci_if_true != never_reached, "inconsistent dest");
655 merge_new_path(dest_bci_if_true);
656 }
657 }
658
659 // False branch
660 Node *iftrue = _gvn.transform( new IfTrueNode(iff) );
661 set_control( iftrue );
662 }
663
664 void Parse::jump_if_always_fork(int dest_bci, bool unc) {
665 // False branch, use existing map and control()
666 if (unc) {
667 repush_if_args();
668 uncommon_trap(Deoptimization::Reason_unstable_if,
669 Deoptimization::Action_reinterpret,
670 nullptr,
671 "taken never");
672 } else {
673 assert(dest_bci != never_reached, "inconsistent dest");
674 merge_new_path(dest_bci);
675 }
676 }
677
678
679 extern "C" {
680 static int jint_cmp(const void *i, const void *j) {
681 int a = *(jint *)i;
682 int b = *(jint *)j;
683 return a > b ? 1 : a < b ? -1 : 0;
684 }
685 }
686
687
688 class SwitchRange : public StackObj {
689 // a range of integers coupled with a bci destination
690 jint _lo; // inclusive lower limit
691 jint _hi; // inclusive upper limit
692 int _dest;
693 float _cnt; // how many times this range was hit according to profiling
694
695 public:
696 jint lo() const { return _lo; }
697 jint hi() const { return _hi; }
698 int dest() const { return _dest; }
699 bool is_singleton() const { return _lo == _hi; }
700 float cnt() const { return _cnt; }
701
702 void setRange(jint lo, jint hi, int dest, float cnt) {
703 assert(lo <= hi, "must be a non-empty range");
704 _lo = lo, _hi = hi; _dest = dest; _cnt = cnt;
705 assert(_cnt >= 0, "");
706 }
707 bool adjoinRange(jint lo, jint hi, int dest, float cnt, bool trim_ranges) {
708 assert(lo <= hi, "must be a non-empty range");
709 if (lo == _hi+1) {
710 // see merge_ranges() comment below
711 if (trim_ranges) {
712 if (cnt == 0) {
713 if (_cnt != 0) {
714 return false;
715 }
716 if (dest != _dest) {
717 _dest = never_reached;
718 }
719 } else {
720 if (_cnt == 0) {
721 return false;
722 }
723 if (dest != _dest) {
724 return false;
725 }
726 }
727 } else {
728 if (dest != _dest) {
729 return false;
730 }
731 }
732 _hi = hi;
733 _cnt += cnt;
734 return true;
735 }
736 return false;
737 }
738
739 void set (jint value, int dest, float cnt) {
740 setRange(value, value, dest, cnt);
741 }
742 bool adjoin(jint value, int dest, float cnt, bool trim_ranges) {
743 return adjoinRange(value, value, dest, cnt, trim_ranges);
744 }
745 bool adjoin(SwitchRange& other) {
746 return adjoinRange(other._lo, other._hi, other._dest, other._cnt, false);
747 }
748
749 void print() {
750 if (is_singleton())
751 tty->print(" {%d}=>%d (cnt=%f)", lo(), dest(), cnt());
752 else if (lo() == min_jint)
753 tty->print(" {..%d}=>%d (cnt=%f)", hi(), dest(), cnt());
754 else if (hi() == max_jint)
755 tty->print(" {%d..}=>%d (cnt=%f)", lo(), dest(), cnt());
756 else
757 tty->print(" {%d..%d}=>%d (cnt=%f)", lo(), hi(), dest(), cnt());
758 }
759 };
760
761 // We try to minimize the number of ranges and the size of the taken
762 // ones using profiling data. When ranges are created,
763 // SwitchRange::adjoinRange() only allows 2 adjoining ranges to merge
764 // if both were never hit or both were hit to build longer unreached
765 // ranges. Here, we now merge adjoining ranges with the same
766 // destination and finally set destination of unreached ranges to the
767 // special value never_reached because it can help minimize the number
768 // of tests that are necessary.
769 //
770 // For instance:
771 // [0, 1] to target1 sometimes taken
772 // [1, 2] to target1 never taken
773 // [2, 3] to target2 never taken
774 // would lead to:
775 // [0, 1] to target1 sometimes taken
776 // [1, 3] never taken
777 //
778 // (first 2 ranges to target1 are not merged)
779 static void merge_ranges(SwitchRange* ranges, int& rp) {
780 if (rp == 0) {
781 return;
782 }
783 int shift = 0;
784 for (int j = 0; j < rp; j++) {
785 SwitchRange& r1 = ranges[j-shift];
786 SwitchRange& r2 = ranges[j+1];
787 if (r1.adjoin(r2)) {
788 shift++;
789 } else if (shift > 0) {
790 ranges[j+1-shift] = r2;
791 }
792 }
793 rp -= shift;
794 for (int j = 0; j <= rp; j++) {
795 SwitchRange& r = ranges[j];
796 if (r.cnt() == 0 && r.dest() != never_reached) {
797 r.setRange(r.lo(), r.hi(), never_reached, r.cnt());
798 }
799 }
800 }
801
802 //-------------------------------do_tableswitch--------------------------------
803 void Parse::do_tableswitch() {
804 // Get information about tableswitch
805 int default_dest = iter().get_dest_table(0);
806 jint lo_index = iter().get_int_table(1);
807 jint hi_index = iter().get_int_table(2);
808 int len = hi_index - lo_index + 1;
809
810 if (len < 1) {
811 // If this is a backward branch, add safepoint
812 maybe_add_safepoint(default_dest);
813 pop(); // the effect of the instruction execution on the operand stack
814 merge(default_dest);
815 return;
816 }
817
818 ciMethodData* methodData = method()->method_data();
819 ciMultiBranchData* profile = nullptr;
820 if (methodData->is_mature() && UseSwitchProfiling) {
821 ciProfileData* data = methodData->bci_to_data(bci());
822 if (data != nullptr && data->is_MultiBranchData()) {
823 profile = (ciMultiBranchData*)data;
824 }
825 }
826 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if);
827
828 // generate decision tree, using trichotomy when possible
829 int rnum = len+2;
830 bool makes_backward_branch = (default_dest <= bci());
831 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum);
832 int rp = -1;
833 if (lo_index != min_jint) {
834 float cnt = 1.0F;
835 if (profile != nullptr) {
836 cnt = (float)profile->default_count() / (hi_index != max_jint ? 2.0F : 1.0F);
837 }
838 ranges[++rp].setRange(min_jint, lo_index-1, default_dest, cnt);
839 }
840 for (int j = 0; j < len; j++) {
841 jint match_int = lo_index+j;
842 int dest = iter().get_dest_table(j+3);
843 makes_backward_branch |= (dest <= bci());
844 float cnt = 1.0F;
845 if (profile != nullptr) {
846 cnt = (float)profile->count_at(j);
847 }
848 if (rp < 0 || !ranges[rp].adjoin(match_int, dest, cnt, trim_ranges)) {
849 ranges[++rp].set(match_int, dest, cnt);
850 }
851 }
852 jint highest = lo_index+(len-1);
853 assert(ranges[rp].hi() == highest, "");
854 if (highest != max_jint) {
855 float cnt = 1.0F;
856 if (profile != nullptr) {
857 cnt = (float)profile->default_count() / (lo_index != min_jint ? 2.0F : 1.0F);
858 }
859 if (!ranges[rp].adjoinRange(highest+1, max_jint, default_dest, cnt, trim_ranges)) {
860 ranges[++rp].setRange(highest+1, max_jint, default_dest, cnt);
861 }
862 }
863 assert(rp < len+2, "not too many ranges");
864
865 if (trim_ranges) {
866 merge_ranges(ranges, rp);
867 }
868
869 // Safepoint in case if backward branch observed
870 if (makes_backward_branch) {
871 add_safepoint();
872 }
873
874 Node* lookup = pop(); // lookup value
875 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]);
876 }
877
878
879 //------------------------------do_lookupswitch--------------------------------
880 void Parse::do_lookupswitch() {
881 // Get information about lookupswitch
882 int default_dest = iter().get_dest_table(0);
883 jint len = iter().get_int_table(1);
884
885 if (len < 1) { // If this is a backward branch, add safepoint
886 maybe_add_safepoint(default_dest);
887 pop(); // the effect of the instruction execution on the operand stack
888 merge(default_dest);
889 return;
890 }
891
892 ciMethodData* methodData = method()->method_data();
893 ciMultiBranchData* profile = nullptr;
894 if (methodData->is_mature() && UseSwitchProfiling) {
895 ciProfileData* data = methodData->bci_to_data(bci());
896 if (data != nullptr && data->is_MultiBranchData()) {
897 profile = (ciMultiBranchData*)data;
898 }
899 }
900 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if);
901
902 // generate decision tree, using trichotomy when possible
903 jint* table = NEW_RESOURCE_ARRAY(jint, len*3);
904 {
905 for (int j = 0; j < len; j++) {
906 table[3*j+0] = iter().get_int_table(2+2*j);
907 table[3*j+1] = iter().get_dest_table(2+2*j+1);
908 // Handle overflow when converting from uint to jint
909 table[3*j+2] = (profile == nullptr) ? 1 : (jint)MIN2<uint>((uint)max_jint, profile->count_at(j));
910 }
911 qsort(table, len, 3*sizeof(table[0]), jint_cmp);
912 }
913
914 float default_cnt = 1.0F;
915 if (profile != nullptr) {
916 juint defaults = max_juint - len;
917 default_cnt = (float)profile->default_count()/(float)defaults;
918 }
919
920 int rnum = len*2+1;
921 bool makes_backward_branch = (default_dest <= bci());
922 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum);
923 int rp = -1;
924 for (int j = 0; j < len; j++) {
925 jint match_int = table[3*j+0];
926 jint dest = table[3*j+1];
927 jint cnt = table[3*j+2];
928 jint next_lo = rp < 0 ? min_jint : ranges[rp].hi()+1;
929 makes_backward_branch |= (dest <= bci());
930 float c = default_cnt * ((float)match_int - (float)next_lo);
931 if (match_int != next_lo && (rp < 0 || !ranges[rp].adjoinRange(next_lo, match_int-1, default_dest, c, trim_ranges))) {
932 assert(default_dest != never_reached, "sentinel value for dead destinations");
933 ranges[++rp].setRange(next_lo, match_int-1, default_dest, c);
934 }
935 if (rp < 0 || !ranges[rp].adjoin(match_int, dest, (float)cnt, trim_ranges)) {
936 assert(dest != never_reached, "sentinel value for dead destinations");
937 ranges[++rp].set(match_int, dest, (float)cnt);
938 }
939 }
940 jint highest = table[3*(len-1)];
941 assert(ranges[rp].hi() == highest, "");
942 if (highest != max_jint &&
943 !ranges[rp].adjoinRange(highest+1, max_jint, default_dest, default_cnt * ((float)max_jint - (float)highest), trim_ranges)) {
944 ranges[++rp].setRange(highest+1, max_jint, default_dest, default_cnt * ((float)max_jint - (float)highest));
945 }
946 assert(rp < rnum, "not too many ranges");
947
948 if (trim_ranges) {
949 merge_ranges(ranges, rp);
950 }
951
952 // Safepoint in case backward branch observed
953 if (makes_backward_branch) {
954 add_safepoint();
955 }
956
957 Node *lookup = pop(); // lookup value
958 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]);
959 }
960
961 static float if_prob(float taken_cnt, float total_cnt) {
962 assert(taken_cnt <= total_cnt, "");
963 if (total_cnt == 0) {
964 return PROB_FAIR;
965 }
966 float p = taken_cnt / total_cnt;
967 return clamp(p, PROB_MIN, PROB_MAX);
968 }
969
970 static float if_cnt(float cnt) {
971 if (cnt == 0) {
972 return COUNT_UNKNOWN;
973 }
974 return cnt;
975 }
976
977 static float sum_of_cnts(SwitchRange *lo, SwitchRange *hi) {
978 float total_cnt = 0;
979 for (SwitchRange* sr = lo; sr <= hi; sr++) {
980 total_cnt += sr->cnt();
981 }
982 return total_cnt;
983 }
984
985 class SwitchRanges : public ResourceObj {
986 public:
987 SwitchRange* _lo;
988 SwitchRange* _hi;
989 SwitchRange* _mid;
990 float _cost;
991
992 enum {
993 Start,
994 LeftDone,
995 RightDone,
996 Done
997 } _state;
998
999 SwitchRanges(SwitchRange *lo, SwitchRange *hi)
1000 : _lo(lo), _hi(hi), _mid(nullptr),
1001 _cost(0), _state(Start) {
1002 }
1003
1004 SwitchRanges()
1005 : _lo(nullptr), _hi(nullptr), _mid(nullptr),
1006 _cost(0), _state(Start) {}
1007 };
1008
1009 // Estimate cost of performing a binary search on lo..hi
1010 static float compute_tree_cost(SwitchRange *lo, SwitchRange *hi, float total_cnt) {
1011 GrowableArray<SwitchRanges> tree;
1012 SwitchRanges root(lo, hi);
1013 tree.push(root);
1014
1015 float cost = 0;
1016 do {
1017 SwitchRanges& r = *tree.adr_at(tree.length()-1);
1018 if (r._hi != r._lo) {
1019 if (r._mid == nullptr) {
1020 float r_cnt = sum_of_cnts(r._lo, r._hi);
1021
1022 if (r_cnt == 0) {
1023 tree.pop();
1024 cost = 0;
1025 continue;
1026 }
1027
1028 SwitchRange* mid = nullptr;
1029 mid = r._lo;
1030 for (float cnt = 0; ; ) {
1031 assert(mid <= r._hi, "out of bounds");
1032 cnt += mid->cnt();
1033 if (cnt > r_cnt / 2) {
1034 break;
1035 }
1036 mid++;
1037 }
1038 assert(mid <= r._hi, "out of bounds");
1039 r._mid = mid;
1040 r._cost = r_cnt / total_cnt;
1041 }
1042 r._cost += cost;
1043 if (r._state < SwitchRanges::LeftDone && r._mid > r._lo) {
1044 cost = 0;
1045 r._state = SwitchRanges::LeftDone;
1046 tree.push(SwitchRanges(r._lo, r._mid-1));
1047 } else if (r._state < SwitchRanges::RightDone) {
1048 cost = 0;
1049 r._state = SwitchRanges::RightDone;
1050 tree.push(SwitchRanges(r._mid == r._lo ? r._mid+1 : r._mid, r._hi));
1051 } else {
1052 tree.pop();
1053 cost = r._cost;
1054 }
1055 } else {
1056 tree.pop();
1057 cost = r._cost;
1058 }
1059 } while (tree.length() > 0);
1060
1061
1062 return cost;
1063 }
1064
1065 // It sometimes pays off to test most common ranges before the binary search
1066 void Parse::linear_search_switch_ranges(Node* key_val, SwitchRange*& lo, SwitchRange*& hi) {
1067 uint nr = hi - lo + 1;
1068 float total_cnt = sum_of_cnts(lo, hi);
1069
1070 float min = compute_tree_cost(lo, hi, total_cnt);
1071 float extra = 1;
1072 float sub = 0;
1073
1074 SwitchRange* array1 = lo;
1075 SwitchRange* array2 = NEW_RESOURCE_ARRAY(SwitchRange, nr);
1076
1077 SwitchRange* ranges = nullptr;
1078
1079 while (nr >= 2) {
1080 assert(lo == array1 || lo == array2, "one the 2 already allocated arrays");
1081 ranges = (lo == array1) ? array2 : array1;
1082
1083 // Find highest frequency range
1084 SwitchRange* candidate = lo;
1085 for (SwitchRange* sr = lo+1; sr <= hi; sr++) {
1086 if (sr->cnt() > candidate->cnt()) {
1087 candidate = sr;
1088 }
1089 }
1090 SwitchRange most_freq = *candidate;
1091 if (most_freq.cnt() == 0) {
1092 break;
1093 }
1094
1095 // Copy remaining ranges into another array
1096 int shift = 0;
1097 for (uint i = 0; i < nr; i++) {
1098 SwitchRange* sr = &lo[i];
1099 if (sr != candidate) {
1100 ranges[i-shift] = *sr;
1101 } else {
1102 shift++;
1103 if (i > 0 && i < nr-1) {
1104 SwitchRange prev = lo[i-1];
1105 prev.setRange(prev.lo(), sr->hi(), prev.dest(), prev.cnt());
1106 if (prev.adjoin(lo[i+1])) {
1107 shift++;
1108 i++;
1109 }
1110 ranges[i-shift] = prev;
1111 }
1112 }
1113 }
1114 nr -= shift;
1115
1116 // Evaluate cost of testing the most common range and performing a
1117 // binary search on the other ranges
1118 float cost = extra + compute_tree_cost(&ranges[0], &ranges[nr-1], total_cnt);
1119 if (cost >= min) {
1120 break;
1121 }
1122 // swap arrays
1123 lo = &ranges[0];
1124 hi = &ranges[nr-1];
1125
1126 // It pays off: emit the test for the most common range
1127 assert(most_freq.cnt() > 0, "must be taken");
1128 Node* val = _gvn.transform(new SubINode(key_val, _gvn.intcon(most_freq.lo())));
1129 Node* cmp = _gvn.transform(new CmpUNode(val, _gvn.intcon(java_subtract(most_freq.hi(), most_freq.lo()))));
1130 Node* tst = _gvn.transform(new BoolNode(cmp, BoolTest::le));
1131 IfNode* iff = create_and_map_if(control(), tst, if_prob(most_freq.cnt(), total_cnt), if_cnt(most_freq.cnt()));
1132 jump_if_true_fork(iff, most_freq.dest(), false);
1133
1134 sub += most_freq.cnt() / total_cnt;
1135 extra += 1 - sub;
1136 min = cost;
1137 }
1138 }
1139
1140 //----------------------------create_jump_tables-------------------------------
1141 bool Parse::create_jump_tables(Node* key_val, SwitchRange* lo, SwitchRange* hi) {
1142 // Are jumptables enabled
1143 if (!UseJumpTables) return false;
1144
1145 // Are jumptables supported
1146 if (!Matcher::has_match_rule(Op_Jump)) return false;
1147
1148 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if);
1149
1150 // Decide if a guard is needed to lop off big ranges at either (or
1151 // both) end(s) of the input set. We'll call this the default target
1152 // even though we can't be sure that it is the true "default".
1153
1154 bool needs_guard = false;
1155 int default_dest;
1156 int64_t total_outlier_size = 0;
1157 int64_t hi_size = ((int64_t)hi->hi()) - ((int64_t)hi->lo()) + 1;
1158 int64_t lo_size = ((int64_t)lo->hi()) - ((int64_t)lo->lo()) + 1;
1159
1160 if (lo->dest() == hi->dest()) {
1161 total_outlier_size = hi_size + lo_size;
1162 default_dest = lo->dest();
1163 } else if (lo_size > hi_size) {
1164 total_outlier_size = lo_size;
1165 default_dest = lo->dest();
1166 } else {
1167 total_outlier_size = hi_size;
1168 default_dest = hi->dest();
1169 }
1170
1171 float total = sum_of_cnts(lo, hi);
1172 float cost = compute_tree_cost(lo, hi, total);
1173
1174 // If a guard test will eliminate very sparse end ranges, then
1175 // it is worth the cost of an extra jump.
1176 float trimmed_cnt = 0;
1177 if (total_outlier_size > (MaxJumpTableSparseness * 4)) {
1178 needs_guard = true;
1179 if (default_dest == lo->dest()) {
1180 trimmed_cnt += lo->cnt();
1181 lo++;
1182 }
1183 if (default_dest == hi->dest()) {
1184 trimmed_cnt += hi->cnt();
1185 hi--;
1186 }
1187 }
1188
1189 // Find the total number of cases and ranges
1190 int64_t num_cases = ((int64_t)hi->hi()) - ((int64_t)lo->lo()) + 1;
1191 int num_range = hi - lo + 1;
1192
1193 // Don't create table if: too large, too small, or too sparse.
1194 if (num_cases > MaxJumpTableSize)
1195 return false;
1196 if (UseSwitchProfiling) {
1197 // MinJumpTableSize is set so with a well balanced binary tree,
1198 // when the number of ranges is MinJumpTableSize, it's cheaper to
1199 // go through a JumpNode that a tree of IfNodes. Average cost of a
1200 // tree of IfNodes with MinJumpTableSize is
1201 // log2f(MinJumpTableSize) comparisons. So if the cost computed
1202 // from profile data is less than log2f(MinJumpTableSize) then
1203 // going with the binary search is cheaper.
1204 if (cost < log2f(MinJumpTableSize)) {
1205 return false;
1206 }
1207 } else {
1208 if (num_cases < MinJumpTableSize)
1209 return false;
1210 }
1211 if (num_cases > (MaxJumpTableSparseness * num_range))
1212 return false;
1213
1214 // Normalize table lookups to zero
1215 int lowval = lo->lo();
1216 key_val = _gvn.transform( new SubINode(key_val, _gvn.intcon(lowval)) );
1217
1218 // Generate a guard to protect against input keyvals that aren't
1219 // in the switch domain.
1220 if (needs_guard) {
1221 Node* size = _gvn.intcon(num_cases);
1222 Node* cmp = _gvn.transform(new CmpUNode(key_val, size));
1223 Node* tst = _gvn.transform(new BoolNode(cmp, BoolTest::ge));
1224 IfNode* iff = create_and_map_if(control(), tst, if_prob(trimmed_cnt, total), if_cnt(trimmed_cnt));
1225 jump_if_true_fork(iff, default_dest, trim_ranges && trimmed_cnt == 0);
1226
1227 total -= trimmed_cnt;
1228 }
1229
1230 // Create an ideal node JumpTable that has projections
1231 // of all possible ranges for a switch statement
1232 // The key_val input must be converted to a pointer offset and scaled.
1233 // Compare Parse::array_addressing above.
1234
1235 // Clean the 32-bit int into a real 64-bit offset.
1236 // Otherwise, the jint value 0 might turn into an offset of 0x0800000000.
1237 // Make I2L conversion control dependent to prevent it from
1238 // floating above the range check during loop optimizations.
1239 // Do not use a narrow int type here to prevent the data path from dying
1240 // while the control path is not removed. This can happen if the type of key_val
1241 // is later known to be out of bounds of [0, num_cases] and therefore a narrow cast
1242 // would be replaced by TOP while C2 is not able to fold the corresponding range checks.
1243 // Set _carry_dependency for the cast to avoid being removed by IGVN.
1244 #ifdef _LP64
1245 key_val = C->constrained_convI2L(&_gvn, key_val, TypeInt::INT, control(), true /* carry_dependency */);
1246 #endif
1247
1248 // Shift the value by wordsize so we have an index into the table, rather
1249 // than a switch value
1250 Node *shiftWord = _gvn.MakeConX(wordSize);
1251 key_val = _gvn.transform( new MulXNode( key_val, shiftWord));
1252
1253 // Create the JumpNode
1254 Arena* arena = C->comp_arena();
1255 float* probs = (float*)arena->Amalloc(sizeof(float)*num_cases);
1256 int i = 0;
1257 if (total == 0) {
1258 for (SwitchRange* r = lo; r <= hi; r++) {
1259 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) {
1260 probs[i] = 1.0F / num_cases;
1261 }
1262 }
1263 } else {
1264 for (SwitchRange* r = lo; r <= hi; r++) {
1265 float prob = r->cnt()/total;
1266 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) {
1267 probs[i] = prob / (r->hi() - r->lo() + 1);
1268 }
1269 }
1270 }
1271
1272 ciMethodData* methodData = method()->method_data();
1273 ciMultiBranchData* profile = nullptr;
1274 if (methodData->is_mature()) {
1275 ciProfileData* data = methodData->bci_to_data(bci());
1276 if (data != nullptr && data->is_MultiBranchData()) {
1277 profile = (ciMultiBranchData*)data;
1278 }
1279 }
1280
1281 Node* jtn = _gvn.transform(new JumpNode(control(), key_val, num_cases, probs, profile == nullptr ? COUNT_UNKNOWN : total));
1282
1283 // These are the switch destinations hanging off the jumpnode
1284 i = 0;
1285 for (SwitchRange* r = lo; r <= hi; r++) {
1286 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) {
1287 Node* input = _gvn.transform(new JumpProjNode(jtn, i, r->dest(), (int)(j - lowval)));
1288 {
1289 PreserveJVMState pjvms(this);
1290 set_control(input);
1291 jump_if_always_fork(r->dest(), trim_ranges && r->cnt() == 0);
1292 }
1293 }
1294 }
1295 assert(i == num_cases, "miscount of cases");
1296 stop_and_kill_map(); // no more uses for this JVMS
1297 return true;
1298 }
1299
1300 //----------------------------jump_switch_ranges-------------------------------
1301 void Parse::jump_switch_ranges(Node* key_val, SwitchRange *lo, SwitchRange *hi, int switch_depth) {
1302 Block* switch_block = block();
1303 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if);
1304
1305 if (switch_depth == 0) {
1306 // Do special processing for the top-level call.
1307 assert(lo->lo() == min_jint, "initial range must exhaust Type::INT");
1308 assert(hi->hi() == max_jint, "initial range must exhaust Type::INT");
1309
1310 // Decrement pred-numbers for the unique set of nodes.
1311 #ifdef ASSERT
1312 if (!trim_ranges) {
1313 // Ensure that the block's successors are a (duplicate-free) set.
1314 int successors_counted = 0; // block occurrences in [hi..lo]
1315 int unique_successors = switch_block->num_successors();
1316 for (int i = 0; i < unique_successors; i++) {
1317 Block* target = switch_block->successor_at(i);
1318
1319 // Check that the set of successors is the same in both places.
1320 int successors_found = 0;
1321 for (SwitchRange* p = lo; p <= hi; p++) {
1322 if (p->dest() == target->start()) successors_found++;
1323 }
1324 assert(successors_found > 0, "successor must be known");
1325 successors_counted += successors_found;
1326 }
1327 assert(successors_counted == (hi-lo)+1, "no unexpected successors");
1328 }
1329 #endif
1330
1331 // Maybe prune the inputs, based on the type of key_val.
1332 jint min_val = min_jint;
1333 jint max_val = max_jint;
1334 const TypeInt* ti = key_val->bottom_type()->isa_int();
1335 if (ti != nullptr) {
1336 min_val = ti->_lo;
1337 max_val = ti->_hi;
1338 assert(min_val <= max_val, "invalid int type");
1339 }
1340 while (lo->hi() < min_val) {
1341 lo++;
1342 }
1343 if (lo->lo() < min_val) {
1344 lo->setRange(min_val, lo->hi(), lo->dest(), lo->cnt());
1345 }
1346 while (hi->lo() > max_val) {
1347 hi--;
1348 }
1349 if (hi->hi() > max_val) {
1350 hi->setRange(hi->lo(), max_val, hi->dest(), hi->cnt());
1351 }
1352
1353 linear_search_switch_ranges(key_val, lo, hi);
1354 }
1355
1356 #ifndef PRODUCT
1357 if (switch_depth == 0) {
1358 _max_switch_depth = 0;
1359 _est_switch_depth = log2i_graceful((hi - lo + 1) - 1) + 1;
1360 }
1361 SwitchRange* orig_lo = lo;
1362 SwitchRange* orig_hi = hi;
1363 #endif
1364
1365 // The lower-range processing is done iteratively to avoid O(N) stack depth
1366 // when the profiling-based pivot repeatedly selects mid==lo (JDK-8366138).
1367 // The upper-range processing remains recursive but is only reached for
1368 // balanced splits, bounding its depth to O(log N).
1369 // Termination: every iteration either exits or strictly decreases hi-lo:
1370 // lo == mid && mid < hi, increments lo
1371 // lo < mid <= hi, sets hi = mid - 1.
1372 for (int depth = switch_depth;; depth++) {
1373 #ifndef PRODUCT
1374 _max_switch_depth = MAX2(depth, _max_switch_depth);
1375 #endif
1376
1377 assert(lo <= hi, "must be a non-empty set of ranges");
1378 if (lo == hi) {
1379 jump_if_always_fork(lo->dest(), trim_ranges && lo->cnt() == 0);
1380 break;
1381 }
1382
1383 assert(lo->hi() == (lo+1)->lo()-1, "contiguous ranges");
1384 assert(hi->lo() == (hi-1)->hi()+1, "contiguous ranges");
1385
1386 if (create_jump_tables(key_val, lo, hi)) return;
1387
1388 SwitchRange* mid = nullptr;
1389 float total_cnt = sum_of_cnts(lo, hi);
1390
1391 int nr = hi - lo + 1;
1392 // With total_cnt==0 the profiling pivot degenerates to mid==lo
1393 // (0 >= 0/2), producing a linear chain of If nodes instead of a
1394 // balanced tree. A balanced tree is strictly better here: all paths
1395 // are cold, so a balanced split gives fewer comparisons at runtime
1396 // and avoids pathological memory usage in the optimizer.
1397 if (UseSwitchProfiling && total_cnt > 0) {
1398 // Don't keep the binary search tree balanced: pick up mid point
1399 // that split frequencies in half.
1400 float cnt = 0;
1401 for (SwitchRange* sr = lo; sr <= hi; sr++) {
1402 cnt += sr->cnt();
1403 if (cnt >= total_cnt / 2) {
1404 mid = sr;
1405 break;
1406 }
1407 }
1408 } else {
1409 mid = lo + nr/2;
1410
1411 // if there is an easy choice, pivot at a singleton:
1412 if (nr > 3 && !mid->is_singleton() && (mid-1)->is_singleton()) mid--;
1413
1414 assert(lo < mid && mid <= hi, "good pivot choice");
1415 assert(nr != 2 || mid == hi, "should pick higher of 2");
1416 assert(nr != 3 || mid == hi-1, "should pick middle of 3");
1417 }
1418 assert(mid != nullptr, "mid must be set");
1419
1420 Node *test_val = _gvn.intcon(mid == lo ? mid->hi() : mid->lo());
1421
1422 if (mid->is_singleton()) {
1423 IfNode *iff_ne = jump_if_fork_int(key_val, test_val, BoolTest::ne, 1-if_prob(mid->cnt(), total_cnt), if_cnt(mid->cnt()));
1424 jump_if_false_fork(iff_ne, mid->dest(), trim_ranges && mid->cnt() == 0);
1425
1426 // Special Case: If there are exactly three ranges, and the high
1427 // and low range each go to the same place, omit the "gt" test,
1428 // since it will not discriminate anything.
1429 bool eq_test_only = (hi == lo+2 && hi->dest() == lo->dest() && mid == hi-1) || mid == lo;
1430
1431 // if there is a higher range, test for it and process it:
1432 if (mid < hi && !eq_test_only) {
1433 // two comparisons of same values--should enable 1 test for 2 branches
1434 // Use BoolTest::lt instead of BoolTest::gt
1435 float cnt = sum_of_cnts(lo, mid-1);
1436 IfNode *iff_lt = jump_if_fork_int(key_val, test_val, BoolTest::lt, if_prob(cnt, total_cnt), if_cnt(cnt));
1437 Node *iftrue = _gvn.transform( new IfTrueNode(iff_lt) );
1438 Node *iffalse = _gvn.transform( new IfFalseNode(iff_lt) );
1439 { PreserveJVMState pjvms(this);
1440 set_control(iffalse);
1441 jump_switch_ranges(key_val, mid+1, hi, depth+1);
1442 }
1443 set_control(iftrue);
1444 }
1445
1446 } else {
1447 // mid is a range, not a singleton, so treat mid..hi as a unit
1448 float cnt = sum_of_cnts(mid == lo ? mid+1 : mid, hi);
1449 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));
1450
1451 // if there is a higher range, test for it and process it:
1452 if (mid == hi) {
1453 jump_if_true_fork(iff_ge, mid->dest(), trim_ranges && cnt == 0);
1454 } else {
1455 Node *iftrue = _gvn.transform( new IfTrueNode(iff_ge) );
1456 Node *iffalse = _gvn.transform( new IfFalseNode(iff_ge) );
1457 { PreserveJVMState pjvms(this);
1458 set_control(iftrue);
1459 jump_switch_ranges(key_val, mid == lo ? mid+1 : mid, hi, depth+1);
1460 }
1461 set_control(iffalse);
1462 }
1463 }
1464
1465 // Process the lower range: iterate instead of recursing.
1466 if (mid == lo) {
1467 if (mid->is_singleton()) {
1468 lo++;
1469 } else {
1470 jump_if_always_fork(lo->dest(), trim_ranges && lo->cnt() == 0);
1471 break;
1472 }
1473 } else {
1474 hi = mid - 1;
1475 }
1476 }
1477
1478 // Decrease pred_count for each successor after all is done.
1479 if (switch_depth == 0) {
1480 int unique_successors = switch_block->num_successors();
1481 for (int i = 0; i < unique_successors; i++) {
1482 Block* target = switch_block->successor_at(i);
1483 // Throw away the pre-allocated path for each unique successor.
1484 target->next_path_num();
1485 }
1486 }
1487
1488 #ifndef PRODUCT
1489 if (TraceOptoParse && Verbose && WizardMode && switch_depth == 0) {
1490 SwitchRange* r;
1491 int nsing = 0;
1492 for (r = orig_lo; r <= orig_hi; r++) {
1493 if( r->is_singleton() ) nsing++;
1494 }
1495 tty->print(">>> ");
1496 _method->print_short_name();
1497 tty->print_cr(" switch decision tree");
1498 tty->print_cr(" %d ranges (%d singletons), max_depth=%d, est_depth=%d",
1499 (int) (orig_hi-orig_lo+1), nsing, _max_switch_depth, _est_switch_depth);
1500 if (_max_switch_depth > _est_switch_depth) {
1501 tty->print_cr("******** BAD SWITCH DEPTH ********");
1502 }
1503 tty->print(" ");
1504 for (r = orig_lo; r <= orig_hi; r++) {
1505 r->print();
1506 }
1507 tty->cr();
1508 }
1509 #endif
1510 }
1511
1512 Node* Parse::floating_point_mod(Node* a, Node* b, BasicType type) {
1513 assert(type == BasicType::T_FLOAT || type == BasicType::T_DOUBLE, "only float and double are floating points");
1514 CallLeafPureNode* mod = type == BasicType::T_DOUBLE ? static_cast<CallLeafPureNode*>(new ModDNode(C, a, b)) : new ModFNode(C, a, b);
1515
1516 set_predefined_input_for_runtime_call(mod);
1517 mod = _gvn.transform(mod)->as_CallLeafPure();
1518 set_predefined_output_for_runtime_call(mod);
1519 Node* result = _gvn.transform(new ProjNode(mod, TypeFunc::Parms + 0));
1520 record_for_igvn(mod);
1521 return result;
1522 }
1523
1524 void Parse::l2f() {
1525 Node* f2 = pop();
1526 Node* f1 = pop();
1527 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::l2f_Type(),
1528 CAST_FROM_FN_PTR(address, SharedRuntime::l2f),
1529 "l2f", nullptr, //no memory effects
1530 f1, f2);
1531 Node* res = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0));
1532
1533 push(res);
1534 }
1535
1536 // Handle jsr and jsr_w bytecode
1537 void Parse::do_jsr() {
1538 assert(bc() == Bytecodes::_jsr || bc() == Bytecodes::_jsr_w, "wrong bytecode");
1539
1540 // Store information about current state, tagged with new _jsr_bci
1541 int return_bci = iter().next_bci();
1542 int jsr_bci = (bc() == Bytecodes::_jsr) ? iter().get_dest() : iter().get_far_dest();
1543
1544 // The way we do things now, there is only one successor block
1545 // for the jsr, because the target code is cloned by ciTypeFlow.
1546 Block* target = successor_for_bci(jsr_bci);
1547
1548 // What got pushed?
1549 const Type* ret_addr = target->peek();
1550 assert(ret_addr->singleton(), "must be a constant (cloned jsr body)");
1551
1552 // Effect on jsr on stack
1553 push(_gvn.makecon(ret_addr));
1554
1555 // Flow to the jsr.
1556 merge(jsr_bci);
1557 }
1558
1559 // Handle ret bytecode
1560 void Parse::do_ret() {
1561 // Find to whom we return.
1562 assert(block()->num_successors() == 1, "a ret can only go one place now");
1563 Block* target = block()->successor_at(0);
1564 assert(!target->is_ready(), "our arrival must be expected");
1565 int pnum = target->next_path_num();
1566 merge_common(target, pnum);
1567 }
1568
1569 static bool has_injected_profile(BoolTest::mask btest, Node* test, int& taken, int& not_taken) {
1570 if (btest != BoolTest::eq && btest != BoolTest::ne) {
1571 // Only ::eq and ::ne are supported for profile injection.
1572 return false;
1573 }
1574 if (test->is_Cmp() &&
1575 test->in(1)->Opcode() == Op_ProfileBoolean) {
1576 ProfileBooleanNode* profile = (ProfileBooleanNode*)test->in(1);
1577 int false_cnt = profile->false_count();
1578 int true_cnt = profile->true_count();
1579
1580 // Counts matching depends on the actual test operation (::eq or ::ne).
1581 // No need to scale the counts because profile injection was designed
1582 // to feed exact counts into VM.
1583 taken = (btest == BoolTest::eq) ? false_cnt : true_cnt;
1584 not_taken = (btest == BoolTest::eq) ? true_cnt : false_cnt;
1585
1586 profile->consume();
1587 return true;
1588 }
1589 return false;
1590 }
1591
1592 // Give up if too few (or too many, in which case the sum will overflow) counts to be meaningful.
1593 // We also check that individual counters are positive first, otherwise the sum can become positive.
1594 // (check for saturation, integer overflow, and immature counts)
1595 static bool counters_are_meaningful(int counter1, int counter2, int min) {
1596 // check for saturation, including "uint" values too big to fit in "int"
1597 if (counter1 < 0 || counter2 < 0) {
1598 return false;
1599 }
1600 // check for integer overflow of the sum
1601 int64_t sum = (int64_t)counter1 + (int64_t)counter2;
1602 STATIC_ASSERT(sizeof(counter1) < sizeof(sum));
1603 if (sum > INT_MAX) {
1604 return false;
1605 }
1606 // check if mature
1607 return (counter1 + counter2) >= min;
1608 }
1609
1610 //--------------------------dynamic_branch_prediction--------------------------
1611 // Try to gather dynamic branch prediction behavior. Return a probability
1612 // of the branch being taken and set the "cnt" field. Returns a -1.0
1613 // if we need to use static prediction for some reason.
1614 float Parse::dynamic_branch_prediction(float &cnt, BoolTest::mask btest, Node* test) {
1615 ResourceMark rm;
1616
1617 cnt = COUNT_UNKNOWN;
1618
1619 int taken = 0;
1620 int not_taken = 0;
1621
1622 bool use_mdo = !has_injected_profile(btest, test, taken, not_taken);
1623
1624 if (use_mdo) {
1625 // Use MethodData information if it is available
1626 // FIXME: free the ProfileData structure
1627 ciMethodData* methodData = method()->method_data();
1628 if (!methodData->is_mature()) return PROB_UNKNOWN;
1629 ciProfileData* data = methodData->bci_to_data(bci());
1630 if (data == nullptr) {
1631 return PROB_UNKNOWN;
1632 }
1633 if (!data->is_JumpData()) return PROB_UNKNOWN;
1634
1635 // get taken and not taken values
1636 // NOTE: saturated UINT_MAX values become negative,
1637 // as do counts above INT_MAX.
1638 taken = data->as_JumpData()->taken();
1639 not_taken = 0;
1640 if (data->is_BranchData()) {
1641 not_taken = data->as_BranchData()->not_taken();
1642 }
1643
1644 // scale the counts to be commensurate with invocation counts:
1645 // NOTE: overflow for positive values is clamped at INT_MAX
1646 taken = method()->scale_count(taken);
1647 not_taken = method()->scale_count(not_taken);
1648 }
1649 // At this point, saturation or overflow is indicated by INT_MAX
1650 // or a negative value.
1651
1652 // Give up if too few (or too many, in which case the sum will overflow) counts to be meaningful.
1653 // We also check that individual counters are positive first, otherwise the sum can become positive.
1654 if (!counters_are_meaningful(taken, not_taken, 40)) {
1655 if (C->log() != nullptr) {
1656 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d'", iter().get_dest(), taken, not_taken);
1657 }
1658 return PROB_UNKNOWN;
1659 }
1660
1661 // Compute frequency that we arrive here
1662 float sum = taken + not_taken;
1663 // Adjust, if this block is a cloned private block but the
1664 // Jump counts are shared. Taken the private counts for
1665 // just this path instead of the shared counts.
1666 if( block()->count() > 0 )
1667 sum = block()->count();
1668 cnt = sum / FreqCountInvocations;
1669
1670 // Pin probability to sane limits
1671 float prob;
1672 if( !taken )
1673 prob = (0+PROB_MIN) / 2;
1674 else if( !not_taken )
1675 prob = (1+PROB_MAX) / 2;
1676 else { // Compute probability of true path
1677 prob = (float)taken / (float)(taken + not_taken);
1678 if (prob > PROB_MAX) prob = PROB_MAX;
1679 if (prob < PROB_MIN) prob = PROB_MIN;
1680 }
1681
1682 assert((cnt > 0.0f) && (prob > 0.0f),
1683 "Bad frequency assignment in if cnt=%g prob=%g taken=%d not_taken=%d", cnt, prob, taken, not_taken);
1684
1685 if (C->log() != nullptr) {
1686 const char* prob_str = nullptr;
1687 if (prob >= PROB_MAX) prob_str = (prob == PROB_MAX) ? "max" : "always";
1688 if (prob <= PROB_MIN) prob_str = (prob == PROB_MIN) ? "min" : "never";
1689 char prob_str_buf[30];
1690 if (prob_str == nullptr) {
1691 jio_snprintf(prob_str_buf, sizeof(prob_str_buf), "%20.2f", prob);
1692 prob_str = prob_str_buf;
1693 }
1694 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d' cnt='%f' prob='%s'",
1695 iter().get_dest(), taken, not_taken, cnt, prob_str);
1696 }
1697 return prob;
1698 }
1699
1700 //-----------------------------branch_prediction-------------------------------
1701 float Parse::branch_prediction(float& cnt,
1702 BoolTest::mask btest,
1703 int target_bci,
1704 Node* test) {
1705 float prob = dynamic_branch_prediction(cnt, btest, test);
1706 // If prob is unknown, switch to static prediction
1707 if (prob != PROB_UNKNOWN) return prob;
1708
1709 prob = PROB_FAIR; // Set default value
1710 if (btest == BoolTest::eq) // Exactly equal test?
1711 prob = PROB_STATIC_INFREQUENT; // Assume its relatively infrequent
1712 else if (btest == BoolTest::ne)
1713 prob = PROB_STATIC_FREQUENT; // Assume its relatively frequent
1714
1715 // If this is a conditional test guarding a backwards branch,
1716 // assume its a loop-back edge. Make it a likely taken branch.
1717 if (target_bci < bci()) {
1718 if (is_osr_parse()) { // Could be a hot OSR'd loop; force deopt
1719 // Since it's an OSR, we probably have profile data, but since
1720 // branch_prediction returned PROB_UNKNOWN, the counts are too small.
1721 // Let's make a special check here for completely zero counts.
1722 ciMethodData* methodData = method()->method_data();
1723 if (!methodData->is_empty()) {
1724 ciProfileData* data = methodData->bci_to_data(bci());
1725 // Only stop for truly zero counts, which mean an unknown part
1726 // of the OSR-ed method, and we want to deopt to gather more stats.
1727 // If you have ANY counts, then this loop is simply 'cold' relative
1728 // to the OSR loop.
1729 if (data == nullptr ||
1730 (data->as_BranchData()->taken() + data->as_BranchData()->not_taken() == 0)) {
1731 // This is the only way to return PROB_UNKNOWN:
1732 return PROB_UNKNOWN;
1733 }
1734 }
1735 }
1736 prob = PROB_STATIC_FREQUENT; // Likely to take backwards branch
1737 }
1738
1739 assert(prob != PROB_UNKNOWN, "must have some guess at this point");
1740 return prob;
1741 }
1742
1743 // The magic constants are chosen so as to match the output of
1744 // branch_prediction() when the profile reports a zero taken count.
1745 // It is important to distinguish zero counts unambiguously, because
1746 // some branches (e.g., _213_javac.Assembler.eliminate) validly produce
1747 // very small but nonzero probabilities, which if confused with zero
1748 // counts would keep the program recompiling indefinitely.
1749 bool Parse::seems_never_taken(float prob) const {
1750 return prob < PROB_MIN;
1751 }
1752
1753 //-------------------------------repush_if_args--------------------------------
1754 // Push arguments of an "if" bytecode back onto the stack by adjusting _sp.
1755 inline int Parse::repush_if_args() {
1756 if (PrintOpto && WizardMode) {
1757 tty->print("defending against excessive implicit null exceptions on %s @%d in ",
1758 Bytecodes::name(iter().cur_bc()), iter().cur_bci());
1759 method()->print_name(); tty->cr();
1760 }
1761 int bc_depth = - Bytecodes::depth(iter().cur_bc());
1762 assert(bc_depth == 1 || bc_depth == 2, "only two kinds of branches");
1763 DEBUG_ONLY(sync_jvms()); // argument(n) requires a synced jvms
1764 assert(argument(0) != nullptr, "must exist");
1765 assert(bc_depth == 1 || argument(1) != nullptr, "two must exist");
1766 inc_sp(bc_depth);
1767 return bc_depth;
1768 }
1769
1770 // Used by StressUnstableIfTraps
1771 static volatile int _trap_stress_counter = 0;
1772
1773 void Parse::increment_trap_stress_counter(Node*& counter, Node*& incr_store) {
1774 Node* counter_addr = makecon(TypeRawPtr::make((address)&_trap_stress_counter));
1775 counter = make_load(control(), counter_addr, TypeInt::INT, T_INT, MemNode::unordered);
1776 counter = _gvn.transform(new AddINode(counter, intcon(1)));
1777 incr_store = store_to_memory(control(), counter_addr, counter, T_INT, MemNode::unordered);
1778 }
1779
1780 //----------------------------------do_ifnull----------------------------------
1781 void Parse::do_ifnull(BoolTest::mask btest, Node *c) {
1782 int target_bci = iter().get_dest();
1783
1784 Node* counter = nullptr;
1785 Node* incr_store = nullptr;
1786 bool do_stress_trap = StressUnstableIfTraps && ((C->random() % 2) == 0);
1787 if (do_stress_trap) {
1788 increment_trap_stress_counter(counter, incr_store);
1789 }
1790
1791 Block* branch_block = successor_for_bci(target_bci);
1792 Block* next_block = successor_for_bci(iter().next_bci());
1793
1794 float cnt;
1795 float prob = branch_prediction(cnt, btest, target_bci, c);
1796 if (prob == PROB_UNKNOWN) {
1797 // (An earlier version of do_ifnull omitted this trap for OSR methods.)
1798 if (PrintOpto && Verbose) {
1799 tty->print_cr("Never-taken edge stops compilation at bci %d", bci());
1800 }
1801 repush_if_args(); // to gather stats on loop
1802 uncommon_trap(Deoptimization::Reason_unreached,
1803 Deoptimization::Action_reinterpret,
1804 nullptr, "cold");
1805 if (C->eliminate_boxing()) {
1806 // Mark the successor blocks as parsed
1807 branch_block->next_path_num();
1808 next_block->next_path_num();
1809 }
1810 return;
1811 }
1812
1813 NOT_PRODUCT(explicit_null_checks_inserted++);
1814
1815 // Generate real control flow
1816 Node *tst = _gvn.transform( new BoolNode( c, btest ) );
1817
1818 // Sanity check the probability value
1819 assert(prob > 0.0f,"Bad probability in Parser");
1820 // Need xform to put node in hash table
1821 IfNode *iff = create_and_xform_if( control(), tst, prob, cnt );
1822 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser");
1823 // True branch
1824 { PreserveJVMState pjvms(this);
1825 Node* iftrue = _gvn.transform( new IfTrueNode (iff) );
1826 set_control(iftrue);
1827
1828 if (stopped()) { // Path is dead?
1829 NOT_PRODUCT(explicit_null_checks_elided++);
1830 if (C->eliminate_boxing()) {
1831 // Mark the successor block as parsed
1832 branch_block->next_path_num();
1833 }
1834 } else { // Path is live.
1835 adjust_map_after_if(btest, c, prob, branch_block);
1836 if (!stopped()) {
1837 merge(target_bci);
1838 }
1839 }
1840 }
1841
1842 // False branch
1843 Node* iffalse = _gvn.transform( new IfFalseNode(iff) );
1844 set_control(iffalse);
1845
1846 if (stopped()) { // Path is dead?
1847 NOT_PRODUCT(explicit_null_checks_elided++);
1848 if (C->eliminate_boxing()) {
1849 // Mark the successor block as parsed
1850 next_block->next_path_num();
1851 }
1852 } else { // Path is live.
1853 adjust_map_after_if(BoolTest(btest).negate(), c, 1.0-prob, next_block);
1854 }
1855
1856 if (do_stress_trap) {
1857 stress_trap(iff, counter, incr_store);
1858 }
1859 }
1860
1861 //------------------------------------do_if------------------------------------
1862 void Parse::do_if(BoolTest::mask btest, Node* c, bool can_trap, bool new_path, Node** ctrl_taken, Node** mem_taken, Node** io_taken) {
1863 int target_bci = iter().get_dest();
1864
1865 Block* branch_block = successor_for_bci(target_bci);
1866 Block* next_block = successor_for_bci(iter().next_bci());
1867
1868 float cnt;
1869 float prob = branch_prediction(cnt, btest, target_bci, c);
1870 float untaken_prob = 1.0 - prob;
1871
1872 if (prob == PROB_UNKNOWN) {
1873 if (PrintOpto && Verbose) {
1874 tty->print_cr("Never-taken edge stops compilation at bci %d", bci());
1875 }
1876 repush_if_args(); // to gather stats on loop
1877 uncommon_trap(Deoptimization::Reason_unreached,
1878 Deoptimization::Action_reinterpret,
1879 nullptr, "cold");
1880 if (C->eliminate_boxing()) {
1881 // Mark the successor blocks as parsed
1882 branch_block->next_path_num();
1883 next_block->next_path_num();
1884 }
1885 return;
1886 }
1887
1888 Node* counter = nullptr;
1889 Node* incr_store = nullptr;
1890 bool do_stress_trap = StressUnstableIfTraps && ((C->random() % 2) == 0);
1891 if (do_stress_trap) {
1892 increment_trap_stress_counter(counter, incr_store);
1893 }
1894
1895 // Sanity check the probability value
1896 assert(0.0f < prob && prob < 1.0f,"Bad probability in Parser");
1897
1898 bool taken_if_true = true;
1899 // Convert BoolTest to canonical form:
1900 if (!BoolTest(btest).is_canonical()) {
1901 btest = BoolTest(btest).negate();
1902 taken_if_true = false;
1903 // prob is NOT updated here; it remains the probability of the taken
1904 // path (as opposed to the prob of the path guarded by an 'IfTrueNode').
1905 }
1906 assert(btest != BoolTest::eq, "!= is the only canonical exact test");
1907
1908 Node* tst0 = new BoolNode(c, btest);
1909 Node* tst = _gvn.transform(tst0);
1910 BoolTest::mask taken_btest = BoolTest::illegal;
1911 BoolTest::mask untaken_btest = BoolTest::illegal;
1912
1913 if (tst->is_Bool()) {
1914 // Refresh c from the transformed bool node, since it may be
1915 // simpler than the original c. Also re-canonicalize btest.
1916 // This wins when (Bool ne (Conv2B p) 0) => (Bool ne (CmpP p null)).
1917 // That can arise from statements like: if (x instanceof C) ...
1918 if (tst != tst0) {
1919 // Canonicalize one more time since transform can change it.
1920 btest = tst->as_Bool()->_test._test;
1921 if (!BoolTest(btest).is_canonical()) {
1922 // Reverse edges one more time...
1923 tst = _gvn.transform( tst->as_Bool()->negate(&_gvn) );
1924 btest = tst->as_Bool()->_test._test;
1925 assert(BoolTest(btest).is_canonical(), "sanity");
1926 taken_if_true = !taken_if_true;
1927 }
1928 c = tst->in(1);
1929 }
1930 BoolTest::mask neg_btest = BoolTest(btest).negate();
1931 taken_btest = taken_if_true ? btest : neg_btest;
1932 untaken_btest = taken_if_true ? neg_btest : btest;
1933 }
1934
1935 // Generate real control flow
1936 float true_prob = (taken_if_true ? prob : untaken_prob);
1937 IfNode* iff = create_and_map_if(control(), tst, true_prob, cnt);
1938 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser");
1939 Node* taken_branch = new IfTrueNode(iff);
1940 Node* untaken_branch = new IfFalseNode(iff);
1941 if (!taken_if_true) { // Finish conversion to canonical form
1942 Node* tmp = taken_branch;
1943 taken_branch = untaken_branch;
1944 untaken_branch = tmp;
1945 }
1946
1947 // Branch is taken:
1948 { PreserveJVMState pjvms(this);
1949 taken_branch = _gvn.transform(taken_branch);
1950 set_control(taken_branch);
1951
1952 if (stopped()) {
1953 if (C->eliminate_boxing() && !new_path) {
1954 // Mark the successor block as parsed (if we haven't created a new path)
1955 branch_block->next_path_num();
1956 }
1957 } else {
1958 adjust_map_after_if(taken_btest, c, prob, branch_block, can_trap);
1959 if (!stopped()) {
1960 if (new_path) {
1961 // Merge by using a new path
1962 merge_new_path(target_bci);
1963 } else if (ctrl_taken != nullptr) {
1964 // Don't merge but save taken branch to be wired by caller
1965 *ctrl_taken = control();
1966 if (mem_taken != nullptr) {
1967 *mem_taken = reset_memory();
1968 }
1969 if (io_taken != nullptr) {
1970 *io_taken = i_o();
1971 }
1972 } else {
1973 merge(target_bci);
1974 }
1975 }
1976 }
1977 }
1978
1979 untaken_branch = _gvn.transform(untaken_branch);
1980 set_control(untaken_branch);
1981
1982 // Branch not taken.
1983 if (stopped() && ctrl_taken == nullptr) {
1984 if (C->eliminate_boxing()) {
1985 // Mark the successor block as parsed (if caller does not re-wire control flow)
1986 next_block->next_path_num();
1987 }
1988 } else {
1989 adjust_map_after_if(untaken_btest, c, untaken_prob, next_block, can_trap);
1990 }
1991
1992 if (do_stress_trap) {
1993 stress_trap(iff, counter, incr_store);
1994 }
1995 }
1996
1997
1998 static ProfilePtrKind speculative_ptr_kind(const TypeOopPtr* t) {
1999 if (t->speculative() == nullptr) {
2000 return ProfileUnknownNull;
2001 }
2002 if (t->speculative_always_null()) {
2003 return ProfileAlwaysNull;
2004 }
2005 if (t->speculative_maybe_null()) {
2006 return ProfileMaybeNull;
2007 }
2008 return ProfileNeverNull;
2009 }
2010
2011 void Parse::acmp_always_null_input(Node* input, const TypeOopPtr* tinput, BoolTest::mask btest, Node* eq_region) {
2012 if (btest == BoolTest::ne) {
2013 {
2014 PreserveJVMState pjvms(this);
2015 inc_sp(2);
2016 null_check_common(input, T_OBJECT, true, nullptr,
2017 !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_null_check) &&
2018 speculative_ptr_kind(tinput) == ProfileAlwaysNull);
2019 dec_sp(2);
2020 int target_bci = iter().get_dest();
2021 merge(target_bci);
2022 }
2023 record_for_igvn(eq_region);
2024 set_control(_gvn.transform(eq_region));
2025 } else {
2026 inc_sp(2);
2027 null_check_common(input, T_OBJECT, true, nullptr,
2028 !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_null_check) &&
2029 speculative_ptr_kind(tinput) == ProfileAlwaysNull);
2030 dec_sp(2);
2031 }
2032 }
2033
2034 Node* Parse::acmp_null_check(Node* input, const TypeOopPtr* tinput, ProfilePtrKind input_ptr, Node*& null_ctl) {
2035 inc_sp(2);
2036 null_ctl = top();
2037 Node* cast = null_check_oop(input, &null_ctl,
2038 input_ptr == ProfileNeverNull || (input_ptr == ProfileUnknownNull && !too_many_traps_or_recompiles(Deoptimization::Reason_null_check)),
2039 false,
2040 speculative_ptr_kind(tinput) == ProfileNeverNull &&
2041 !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_null_check));
2042 dec_sp(2);
2043 return cast;
2044 }
2045
2046 void Parse::acmp_type_check_or_trap(Node** non_null_input, ciKlass* input_type, Deoptimization::DeoptReason reason) {
2047 Node* slow_ctl = type_check_receiver(*non_null_input, input_type, 1.0, non_null_input);
2048 {
2049 PreserveJVMState pjvms(this);
2050 inc_sp(2);
2051 set_control(slow_ctl);
2052 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
2053 }
2054 }
2055
2056 void Parse::acmp_type_check(Node* input, const TypeOopPtr* tinput, ProfilePtrKind input_ptr, ciKlass* input_type, BoolTest::mask btest, Node* eq_region) {
2057 Node* null_ctl;
2058 Node* cast = acmp_null_check(input, tinput, input_ptr, null_ctl);
2059
2060 if (input_type != nullptr) {
2061 Deoptimization::DeoptReason reason;
2062 if (tinput->speculative_type() != nullptr && !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_class_check)) {
2063 reason = Deoptimization::Reason_speculate_class_check;
2064 } else {
2065 reason = Deoptimization::Reason_class_check;
2066 }
2067 acmp_type_check_or_trap(&cast, input_type, reason);
2068 } else {
2069 // No specific type, check for inline type
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 Node* ne_region = new RegionNode(2);
2076 ne_region->add_req(null_ctl);
2077 ne_region->add_req(control());
2078
2079 record_for_igvn(ne_region);
2080 set_control(_gvn.transform(ne_region));
2081 if (btest == BoolTest::ne) {
2082 {
2083 PreserveJVMState pjvms(this);
2084 if (null_ctl == top()) {
2085 replace_in_map(input, cast);
2086 }
2087 int target_bci = iter().get_dest();
2088 merge(target_bci);
2089 }
2090 record_for_igvn(eq_region);
2091 set_control(_gvn.transform(eq_region));
2092 } else {
2093 if (null_ctl == top()) {
2094 replace_in_map(input, cast);
2095 }
2096 set_control(_gvn.transform(ne_region));
2097 }
2098 }
2099
2100 void Parse::do_acmp(BoolTest::mask btest, Node* left, Node* right) {
2101 ciKlass* left_type = nullptr;
2102 ciKlass* right_type = nullptr;
2103 ProfilePtrKind left_ptr = ProfileUnknownNull;
2104 ProfilePtrKind right_ptr = ProfileUnknownNull;
2105 bool left_inline_type = true;
2106 bool right_inline_type = true;
2107
2108 // Leverage profiling at acmp
2109 if (UseACmpProfile) {
2110 method()->acmp_profiled_type(bci(), left_type, right_type, left_ptr, right_ptr, left_inline_type, right_inline_type);
2111 if (too_many_traps_or_recompiles(Deoptimization::Reason_class_check)) {
2112 left_type = nullptr;
2113 right_type = nullptr;
2114 left_inline_type = true;
2115 right_inline_type = true;
2116 }
2117 if (too_many_traps_or_recompiles(Deoptimization::Reason_null_check)) {
2118 left_ptr = ProfileUnknownNull;
2119 right_ptr = ProfileUnknownNull;
2120 }
2121 }
2122
2123 if (UseTypeSpeculation) {
2124 record_profile_for_speculation(left, left_type, left_ptr);
2125 record_profile_for_speculation(right, right_type, right_ptr);
2126 }
2127
2128 if (!Arguments::is_valhalla_enabled()) {
2129 Node* cmp = CmpP(left, right);
2130 cmp = optimize_cmp_with_klass(cmp);
2131 do_if(btest, cmp);
2132 return;
2133 }
2134
2135 // Check for equality before potentially allocating
2136 if (left == right) {
2137 do_if(btest, makecon(TypeInt::CC_EQ));
2138 return;
2139 }
2140
2141 // Allocate inline type operands and re-execute on deoptimization
2142 if (left->is_InlineType()) {
2143 PreserveReexecuteState preexecs(this);
2144 inc_sp(2);
2145 jvms()->set_should_reexecute(true);
2146 left = left->as_InlineType()->buffer(this);
2147 }
2148 if (right->is_InlineType()) {
2149 PreserveReexecuteState preexecs(this);
2150 inc_sp(2);
2151 jvms()->set_should_reexecute(true);
2152 right = right->as_InlineType()->buffer(this);
2153 }
2154
2155 // First, do a normal pointer comparison
2156 const TypeOopPtr* tleft = _gvn.type(left)->isa_oopptr();
2157 const TypeOopPtr* tright = _gvn.type(right)->isa_oopptr();
2158 Node* cmp = CmpP(left, right);
2159 record_for_igvn(cmp);
2160 cmp = optimize_cmp_with_klass(cmp);
2161 if (tleft == nullptr || !tleft->can_be_inline_type() ||
2162 tright == nullptr || !tright->can_be_inline_type()) {
2163 // This is sufficient, if one of the operands can't be an inline type
2164 do_if(btest, cmp);
2165 return;
2166 }
2167
2168 // Don't add traps to unstable if branches because additional checks are required to
2169 // decide if the operands are equal/substitutable and we therefore shouldn't prune
2170 // branches for one if based on the profiling of the acmp branches.
2171 // Also, OptimizeUnstableIf would set an incorrect re-rexecution state because it
2172 // assumes that there is a 1-1 mapping between the if and the acmp branches and that
2173 // hitting a trap means that we will take the corresponding acmp branch on re-execution.
2174 const bool can_trap = true;
2175
2176 Node* eq_region = nullptr;
2177 if (btest == BoolTest::eq) {
2178 do_if(btest, cmp, !can_trap, true);
2179 if (stopped()) {
2180 // Pointers are equal, operands must be equal
2181 return;
2182 }
2183 } else {
2184 assert(btest == BoolTest::ne, "only eq or ne");
2185 Node* is_not_equal = nullptr;
2186 eq_region = new RegionNode(4);
2187 {
2188 PreserveJVMState pjvms(this);
2189 // Pointers are not equal, but more checks are needed to determine if the operands are (not) substitutable
2190 do_if(btest, cmp, !can_trap, false, &is_not_equal);
2191 if (!stopped()) {
2192 eq_region->init_req(1, control());
2193 }
2194 }
2195 if (is_not_equal == nullptr || is_not_equal->is_top()) {
2196 record_for_igvn(eq_region);
2197 set_control(_gvn.transform(eq_region));
2198 return;
2199 }
2200 set_control(is_not_equal);
2201 }
2202
2203 // Prefer speculative types if available
2204 if (!too_many_traps_or_recompiles(Deoptimization::Reason_speculate_class_check)) {
2205 if (tleft->speculative_type() != nullptr) {
2206 left_type = tleft->speculative_type();
2207 }
2208 if (tright->speculative_type() != nullptr) {
2209 right_type = tright->speculative_type();
2210 }
2211 }
2212
2213 if (speculative_ptr_kind(tleft) != ProfileMaybeNull && speculative_ptr_kind(tleft) != ProfileUnknownNull) {
2214 ProfilePtrKind speculative_left_ptr = speculative_ptr_kind(tleft);
2215 if (speculative_left_ptr == ProfileAlwaysNull && !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_null_assert)) {
2216 left_ptr = speculative_left_ptr;
2217 } else if (speculative_left_ptr == ProfileNeverNull && !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_null_check)) {
2218 left_ptr = speculative_left_ptr;
2219 }
2220 }
2221 if (speculative_ptr_kind(tright) != ProfileMaybeNull && speculative_ptr_kind(tright) != ProfileUnknownNull) {
2222 ProfilePtrKind speculative_right_ptr = speculative_ptr_kind(tright);
2223 if (speculative_right_ptr == ProfileAlwaysNull && !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_null_assert)) {
2224 right_ptr = speculative_right_ptr;
2225 } else if (speculative_right_ptr == ProfileNeverNull && !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_null_check)) {
2226 right_ptr = speculative_right_ptr;
2227 }
2228 }
2229
2230 if (left_ptr == ProfileAlwaysNull) {
2231 // Comparison with null. Assert the input is indeed null and we're done.
2232 acmp_always_null_input(left, tleft, btest, eq_region);
2233 return;
2234 }
2235 if (right_ptr == ProfileAlwaysNull) {
2236 // Comparison with null. Assert the input is indeed null and we're done.
2237 acmp_always_null_input(right, tright, btest, eq_region);
2238 return;
2239 }
2240 if (left_type != nullptr && !left_type->is_inlinetype()) {
2241 // Comparison with an object of known type
2242 acmp_type_check(left, tleft, left_ptr, left_type, btest, eq_region);
2243 return;
2244 }
2245 if (right_type != nullptr && !right_type->is_inlinetype()) {
2246 // Comparison with an object of known type
2247 acmp_type_check(right, tright, right_ptr, right_type, btest, eq_region);
2248 return;
2249 }
2250 if (!left_inline_type) {
2251 // Comparison with an object known not to be an inline type
2252 acmp_type_check(left, tleft, left_ptr, nullptr, btest, eq_region);
2253 return;
2254 }
2255 if (!right_inline_type) {
2256 // Comparison with an object known not to be an inline type
2257 acmp_type_check(right, tright, right_ptr, nullptr, btest, eq_region);
2258 return;
2259 }
2260
2261 // Pointers are not equal, check if first operand is non-null
2262 Node* ne_region = new RegionNode(7);
2263 Node* null_ctl = nullptr;
2264 Node* not_null_left = nullptr;
2265 Node* not_null_right = acmp_null_check(right, tright, right_ptr, null_ctl);
2266 ne_region->init_req(1, null_ctl);
2267
2268 Node* kls_right = nullptr;
2269 if (!stopped()) {
2270 // First operand is non-null, check if it is the speculative inline type if possible
2271 // (which later allows isSubstitutable to be intrinsified), or any inline type if no
2272 // speculation is available.
2273 if (right_type != nullptr && right_type->is_inlinetype()) {
2274 acmp_type_check_or_trap(¬_null_right, right_type, Deoptimization::Reason_speculate_class_check);
2275 } else {
2276 Node* is_value = inline_type_test(not_null_right);
2277 IfNode* is_value_iff = create_and_map_if(control(), is_value, PROB_FAIR, COUNT_UNKNOWN);
2278 Node* not_value = _gvn.transform(new IfFalseNode(is_value_iff));
2279 ne_region->init_req(2, not_value);
2280 set_control(_gvn.transform(new IfTrueNode(is_value_iff)));
2281 }
2282
2283 // The first operand is an inline type, check if the second operand is non-null
2284 not_null_left = acmp_null_check(left, tleft, left_ptr, null_ctl);
2285 ne_region->init_req(3, null_ctl);
2286 if (!stopped()) {
2287 // Check if lhs operand is of a specific speculative inline type (see above).
2288 // If not, we don't need to enforce that the lhs is a value object since we know
2289 // it already for the rhs, and must enforce that they have the same type.
2290 if (left_type != nullptr && left_type->is_inlinetype()) {
2291 acmp_type_check_or_trap(¬_null_left, left_type, Deoptimization::Reason_speculate_class_check);
2292 }
2293 if (!stopped()) {
2294 // Check if both operands are of the same class.
2295 Node* kls_left = load_object_klass(not_null_left);
2296 kls_right = load_object_klass(not_null_right);
2297 Node* kls_cmp = CmpP(kls_left, kls_right);
2298 Node* kls_bol = _gvn.transform(new BoolNode(kls_cmp, BoolTest::ne));
2299 IfNode* kls_iff = create_and_map_if(control(), kls_bol, PROB_FAIR, COUNT_UNKNOWN);
2300 Node* kls_ne = _gvn.transform(new IfTrueNode(kls_iff));
2301 set_control(_gvn.transform(new IfFalseNode(kls_iff)));
2302 ne_region->init_req(4, kls_ne);
2303 }
2304 }
2305 }
2306
2307 if (stopped()) {
2308 record_for_igvn(ne_region);
2309 set_control(_gvn.transform(ne_region));
2310 if (btest == BoolTest::ne) {
2311 {
2312 PreserveJVMState pjvms(this);
2313 int target_bci = iter().get_dest();
2314 merge(target_bci);
2315 }
2316 record_for_igvn(eq_region);
2317 set_control(_gvn.transform(eq_region));
2318 }
2319 return;
2320 }
2321 assert(kls_right != nullptr, "");
2322
2323 IfNode* mask_iff = nullptr;
2324 // If any operand has a precisely known type, isSubstitutable will be intrinsified, so we don't need the fast path
2325 if (UseAcmpFastPath && !_gvn.type(not_null_left)->is_inlinetypeptr() && !_gvn.type(not_null_right)->is_inlinetypeptr()) {
2326 /* Here, we are generating the fast path (the slow path being the call to isSubstitutable)
2327 * See the declarations of _fast_acmp_offset and _fast_acmp_mask in InlineKlass::Members
2328 * for details about the fast path logic, and the meaning of these values.
2329 */
2330 Node* members_addr = off_heap_plus_addr(kls_right, in_bytes(InlineKlass::adr_members_offset()));
2331 Node* members = make_load(control(), members_addr, TypeRawPtr::BOTTOM, T_ADDRESS, MemNode::unordered);
2332 Node* offset_addr = off_heap_plus_addr(members, in_bytes(InlineKlass::fast_acmp_offset_offset()));
2333 Node* offset = make_load(control(), offset_addr, TypeInt::INT, T_INT, MemNode::unordered);
2334
2335 Node* offset_cmp = CmpI(offset, zerocon(T_INT));
2336 Node* offset_bol = _gvn.transform(new BoolNode(offset_cmp, BoolTest::lt));
2337 mask_iff = create_and_map_if(control(), offset_bol, PROB_FAIR, COUNT_UNKNOWN);
2338 Node* slow_path_ctl = _gvn.transform(new IfTrueNode(mask_iff));
2339 Node* fast_path_ctl = _gvn.transform(new IfFalseNode(mask_iff));
2340 set_control(slow_path_ctl);
2341
2342 {
2343 PreserveJVMState jvms(this);
2344 set_control(fast_path_ctl);
2345
2346 Node* offset_l = ConvI2L(offset);
2347 Node* fast_acmp_mask_addr = off_heap_plus_addr(members, in_bytes(InlineKlass::fast_acmp_mask_offset()));
2348 Node* fast_acmp_mask = make_load(control(), fast_acmp_mask_addr, TypeLong::LONG, T_LONG, MemNode::unordered);
2349
2350 // *(left + offset) & mask == *(right + offset) & mask
2351 Node* left_payload_addr = basic_plus_adr(not_null_left, offset_l);
2352 Node* left_payload = make_load(control(), left_payload_addr, TypeLong::LONG, T_LONG, MemNode::unordered, LoadNNode::DependsOnlyOnTest, false, true, true, true);
2353 Node* left_masked = _gvn.transform(new AndLNode(left_payload, fast_acmp_mask));
2354
2355 Node* right_payload_addr = basic_plus_adr(not_null_right, offset_l);
2356 Node* right_payload = make_load(control(), right_payload_addr, TypeLong::LONG, T_LONG, MemNode::unordered, LoadNNode::DependsOnlyOnTest, false, true, true, true);
2357 Node* right_masked = _gvn.transform(new AndLNode(right_payload, fast_acmp_mask));
2358
2359 Node* masked_cmp = CmpL(left_masked, right_masked);
2360
2361 Node* ctl = C->top();
2362 if (btest == BoolTest::eq) {
2363 PreserveJVMState pjvms(this);
2364 do_if(btest, masked_cmp, !can_trap, true, nullptr);
2365 if (!stopped()) {
2366 ctl = control();
2367 }
2368 } else {
2369 assert(btest == BoolTest::ne, "only eq or ne");
2370 PreserveJVMState pjvms(this);
2371 do_if(btest, masked_cmp, !can_trap, false, &ctl);
2372 if (!stopped()) {
2373 eq_region->init_req(3, control());
2374 }
2375 }
2376 ne_region->init_req(6, ctl);
2377 }
2378 }
2379
2380 // Both operands are values types of the same class, we need to perform a
2381 // substitutability test. Delegate to ValueObjectMethods::isSubstitutable().
2382 Node* ne_io_phi = PhiNode::make(ne_region, i_o());
2383 Node* mem = reset_memory();
2384 Node* ne_mem_phi = PhiNode::make(ne_region, mem);
2385
2386 Node* eq_io_phi = nullptr;
2387 Node* eq_mem_phi = nullptr;
2388 if (eq_region != nullptr) {
2389 eq_io_phi = PhiNode::make(eq_region, i_o());
2390 eq_mem_phi = PhiNode::make(eq_region, mem);
2391 }
2392
2393 set_all_memory(mem);
2394
2395 kill_dead_locals();
2396 ciSymbol* subst_method_name = ciSymbols::isSubstitutable_name();
2397 ciMethod* subst_method = ciEnv::current()->ValueObjectMethods_klass()->find_method(subst_method_name, ciSymbols::object_object_boolean_signature());
2398 CallStaticJavaNode* call = new CallStaticJavaNode(C, TypeFunc::make(subst_method), SharedRuntime::get_resolve_static_call_stub(), subst_method);
2399 call->set_override_symbolic_info(true);
2400 call->init_req(TypeFunc::Parms, not_null_left);
2401 call->init_req(TypeFunc::Parms+1, not_null_right);
2402 inc_sp(2);
2403 set_edges_for_java_call(call, false, false);
2404 Node* ret = set_results_for_java_call(call, false, true);
2405 dec_sp(2);
2406
2407 assert(acmp_fast_path_if_from_substitutable_call(&_gvn, call) == mask_iff, "");
2408
2409 // Test the return value of ValueObjectMethods::isSubstitutable()
2410 // This is the last check, do_if can emit traps now.
2411 Node* subst_cmp = _gvn.transform(new CmpINode(ret, intcon(1)));
2412 Node* ctl = C->top();
2413 Node* mem_taken = nullptr;
2414 Node* io_taken = nullptr;
2415 if (btest == BoolTest::eq) {
2416 PreserveJVMState pjvms(this);
2417 // Also merges branch block.
2418 do_if(btest, subst_cmp, can_trap, false, nullptr, &mem_taken, &io_taken);
2419 if (!stopped()) {
2420 ctl = control();
2421 mem_taken = reset_memory();
2422 io_taken = i_o();
2423 }
2424 } else {
2425 assert(btest == BoolTest::ne, "only eq or ne");
2426 PreserveJVMState pjvms(this);
2427 do_if(btest, subst_cmp, can_trap, false, &ctl, &mem_taken, &io_taken);
2428 if (!stopped()) {
2429 eq_region->init_req(2, control());
2430 eq_io_phi->init_req(2, i_o());
2431 eq_mem_phi->init_req(2, reset_memory());
2432 }
2433 }
2434 ne_region->init_req(5, ctl);
2435 ne_io_phi->init_req(5, io_taken);
2436 ne_mem_phi->init_req(5, mem_taken);
2437
2438 // BoolTest::eq: ne_region is fall-through block.
2439 // BoolTest::ne: ne_region is branch block -> merge below.
2440 record_for_igvn(ne_region);
2441 set_control(_gvn.transform(ne_region));
2442 set_i_o(_gvn.transform(ne_io_phi));
2443 set_all_memory(_gvn.transform(ne_mem_phi));
2444
2445 if (btest == BoolTest::ne) {
2446 int target_bci = iter().get_dest();
2447 if (!stopped()) {
2448 PreserveJVMState pjvms(this);
2449 merge(target_bci);
2450 } else if (C->eliminate_boxing()) {
2451 // Mark the branch block as parsed.
2452 Block* branch_block = successor_for_bci(target_bci);
2453 branch_block->next_path_num();
2454 }
2455
2456 // Fall-through block.
2457 record_for_igvn(eq_region);
2458 set_control(_gvn.transform(eq_region));
2459 set_i_o(_gvn.transform(eq_io_phi));
2460 set_all_memory(_gvn.transform(eq_mem_phi));
2461 }
2462 }
2463
2464 /* Detects whether a call to isSubstitutable is under an IfNode guarding the fast path for acmp.
2465 * If so, returns the IfNode branching between the call and the fast path. Returns null otherwise.
2466 *
2467 * The fast path is a LOT easier to generate at parsing time, but can be later proven useless if further
2468 * optimization narrows down the type of operands and allows intrinsification of the substitutability
2469 * check. In this case, the fast path might still apply, but it comes with various downsides, such as
2470 * mismatch access that may hinder optimizations, or buffering requirement. So, when intrinsifying the call,
2471 * we try to remove the fast path.
2472 *
2473 * This test isn't so bad. Loading the fast acmp offset is pretty unique to the fast acmp path.
2474 *
2475 * Clearly, this is only a step before a proper solution for acmp, such as a macro node.
2476 */
2477 IfNode* Parse::acmp_fast_path_if_from_substitutable_call(PhaseGVN* phase, CallStaticJavaNode* call) {
2478 auto is_con_offset = [](Node* node, ByteSize n) -> bool {
2479 if (!node->is_Con()) return false;
2480 TypeNode* con = node->as_Type();
2481 assert(con->type()->is_intptr_t(), "");
2482 return con->type()->is_intptr_t()->is_con(in_bytes(n));
2483 };
2484
2485 assert(call->in(TypeFunc::Control) != nullptr, "");
2486 if (!call->in(TypeFunc::Control)->is_IfProj()) return nullptr;
2487 IfProjNode* if_proj = call->in(TypeFunc::Control)->as_IfProj();
2488 if (if_proj->_con != 1) return nullptr;
2489
2490 assert(if_proj->in(0) != nullptr, "");
2491 assert(if_proj->in(0)->is_If(), "");
2492 IfNode* iff = if_proj->in(0)->as_If();
2493
2494 assert(iff->in(1) != nullptr, "");
2495 if (!iff->in(1)->is_Bool()) return nullptr;
2496 BoolNode* lt = iff->in(1)->as_Bool();
2497 if (lt->_test._test != BoolTest::lt) return nullptr;
2498
2499 assert(lt->in(1) != nullptr, "");
2500 if (lt->in(1)->Opcode() != Op_CmpI) return nullptr;
2501 CmpNode* cmp_i = lt->in(1)->as_Cmp();
2502
2503 assert(cmp_i->in(1) != nullptr, "");
2504 assert(cmp_i->in(2) != nullptr, "");
2505
2506 if (cmp_i->in(1)->Opcode() != Op_LoadI) return nullptr;
2507 LoadNode* load_offset = cmp_i->in(1)->as_Load();
2508 if (!cmp_i->in(2)->is_ConI()) return nullptr;
2509 ConINode* zero_i = cmp_i->in(2)->as_ConI();
2510 assert(zero_i->type()->is_int() != nullptr, "");
2511 if (!zero_i->type()->is_int()->is_con(0)) return nullptr;
2512
2513 assert(load_offset->in(2) != nullptr, "");
2514 if (!load_offset->in(2)->is_AddP()) return nullptr;
2515 AddPNode* offset_addr_add = load_offset->in(2)->as_AddP();
2516
2517 assert(offset_addr_add->in(AddPNode::Base) != nullptr, "");
2518 assert(offset_addr_add->in(AddPNode::Address) != nullptr, "");
2519 assert(offset_addr_add->in(AddPNode::Offset) != nullptr, "");
2520 if (!offset_addr_add->in(AddPNode::Base)->is_top()) return nullptr;
2521 if (offset_addr_add->in(AddPNode::Address)->Opcode() != Op_LoadP) return nullptr;
2522 LoadNode* load_members = offset_addr_add->in(AddPNode::Address)->as_Load();
2523 if (!is_con_offset(offset_addr_add->in(AddPNode::Offset), InlineKlass::fast_acmp_offset_offset())) return nullptr;
2524
2525 assert(load_members->in(2) != nullptr, "");
2526 if (!load_members->in(2)->is_AddP()) return nullptr;
2527 AddPNode* members_addr_add = load_members->in(2)->as_AddP();
2528
2529 assert(members_addr_add->in(AddPNode::Base) != nullptr, "");
2530 assert(members_addr_add->in(AddPNode::Address) != nullptr, "");
2531 assert(members_addr_add->in(AddPNode::Offset) != nullptr, "");
2532 if (!members_addr_add->in(AddPNode::Base)->is_top()) return nullptr;
2533 if (!phase->type(members_addr_add->in(AddPNode::Address))->isa_instklassptr()) return nullptr;
2534 if (!is_con_offset(members_addr_add->in(AddPNode::Offset), InlineKlass::adr_members_offset())) return nullptr;
2535
2536 return iff;
2537 }
2538
2539 // Force unstable if traps to be taken randomly to trigger intermittent bugs such as incorrect debug information.
2540 // Add another if before the unstable if that checks a "random" condition at runtime (a simple shared counter) and
2541 // then either takes the trap or executes the original, unstable if.
2542 void Parse::stress_trap(IfNode* orig_iff, Node* counter, Node* incr_store) {
2543 // Search for an unstable if trap
2544 CallStaticJavaNode* trap = nullptr;
2545 assert(orig_iff->Opcode() == Op_If && orig_iff->outcnt() == 2, "malformed if");
2546 ProjNode* trap_proj = orig_iff->uncommon_trap_proj(trap, Deoptimization::Reason_unstable_if);
2547 if (trap == nullptr || !trap->jvms()->should_reexecute()) {
2548 // No suitable trap found. Remove unused counter load and increment.
2549 C->gvn_replace_by(incr_store, incr_store->in(MemNode::Memory));
2550 return;
2551 }
2552
2553 // Remove trap from optimization list since we add another path to the trap.
2554 bool success = C->remove_unstable_if_trap(trap, true);
2555 assert(success, "Trap already modified");
2556
2557 // Add a check before the original if that will trap with a certain frequency and execute the original if otherwise
2558 int freq_log = (C->random() % 31) + 1; // Random logarithmic frequency in [1, 31]
2559 Node* mask = intcon(right_n_bits(freq_log));
2560 counter = _gvn.transform(new AndINode(counter, mask));
2561 Node* cmp = _gvn.transform(new CmpINode(counter, intcon(0)));
2562 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::mask::eq));
2563 IfNode* iff = _gvn.transform(new IfNode(orig_iff->in(0), bol, orig_iff->_prob, orig_iff->_fcnt))->as_If();
2564 Node* if_true = _gvn.transform(new IfTrueNode(iff));
2565 Node* if_false = _gvn.transform(new IfFalseNode(iff));
2566 assert(!if_true->is_top() && !if_false->is_top(), "trap always / never taken");
2567
2568 // Trap
2569 assert(trap_proj->outcnt() == 1, "some other nodes are dependent on the trap projection");
2570
2571 Node* trap_region = new RegionNode(3);
2572 trap_region->set_req(1, trap_proj);
2573 trap_region->set_req(2, if_true);
2574 trap->set_req(0, _gvn.transform(trap_region));
2575
2576 // Don't trap, execute original if
2577 orig_iff->set_req(0, if_false);
2578 }
2579
2580 bool Parse::path_is_suitable_for_uncommon_trap(float prob) const {
2581 // Randomly skip emitting an uncommon trap
2582 if (StressUnstableIfTraps && ((C->random() % 2) == 0)) {
2583 return false;
2584 }
2585 // Don't want to speculate on uncommon traps when running with -Xcomp
2586 if (!UseInterpreter) {
2587 return false;
2588 }
2589 return seems_never_taken(prob) &&
2590 !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if);
2591 }
2592
2593 void Parse::maybe_add_predicate_after_if(Block* path) {
2594 if (path->is_SEL_head() && path->preds_parsed() == 0) {
2595 // Add predicates at bci of if dominating the loop so traps can be
2596 // recorded on the if's profile data
2597 int bc_depth = repush_if_args();
2598 add_parse_predicates();
2599 dec_sp(bc_depth);
2600 path->set_has_predicates();
2601 }
2602 }
2603
2604
2605 //----------------------------adjust_map_after_if------------------------------
2606 // Adjust the JVM state to reflect the result of taking this path.
2607 // Basically, it means inspecting the CmpNode controlling this
2608 // branch, seeing how it constrains a tested value, and then
2609 // deciding if it's worth our while to encode this constraint
2610 // as graph nodes in the current abstract interpretation map.
2611 void Parse::adjust_map_after_if(BoolTest::mask btest, Node* c, float prob, Block* path, bool can_trap) {
2612 if (!c->is_Cmp()) {
2613 maybe_add_predicate_after_if(path);
2614 return;
2615 }
2616
2617 if (stopped() || btest == BoolTest::illegal) {
2618 return; // nothing to do
2619 }
2620
2621 bool is_fallthrough = (path == successor_for_bci(iter().next_bci()));
2622
2623 if (can_trap && path_is_suitable_for_uncommon_trap(prob)) {
2624 repush_if_args();
2625 Node* call = uncommon_trap(Deoptimization::Reason_unstable_if,
2626 Deoptimization::Action_reinterpret,
2627 nullptr,
2628 (is_fallthrough ? "taken always" : "taken never"));
2629
2630 if (call != nullptr) {
2631 C->record_unstable_if_trap(new UnstableIfTrap(call->as_CallStaticJava(), path));
2632 }
2633 return;
2634 }
2635
2636 if (c->is_FlatArrayCheck()) {
2637 maybe_add_predicate_after_if(path);
2638 return;
2639 }
2640
2641 Node* val = c->in(1);
2642 Node* con = c->in(2);
2643 const Type* tcon = _gvn.type(con);
2644 const Type* tval = _gvn.type(val);
2645 bool have_con = tcon->singleton();
2646 if (tval->singleton()) {
2647 if (!have_con) {
2648 // Swap, so constant is in con.
2649 con = val;
2650 tcon = tval;
2651 val = c->in(2);
2652 tval = _gvn.type(val);
2653 btest = BoolTest(btest).commute();
2654 have_con = true;
2655 } else {
2656 // Do we have two constants? Then leave well enough alone.
2657 have_con = false;
2658 }
2659 }
2660 if (!have_con) { // remaining adjustments need a con
2661 maybe_add_predicate_after_if(path);
2662 return;
2663 }
2664
2665 sharpen_type_after_if(btest, con, tcon, val, tval);
2666 maybe_add_predicate_after_if(path);
2667 }
2668
2669
2670 static Node* extract_obj_from_klass_load(PhaseGVN* gvn, Node* n) {
2671 Node* ldk;
2672 if (n->is_DecodeNKlass()) {
2673 if (n->in(1)->Opcode() != Op_LoadNKlass) {
2674 return nullptr;
2675 } else {
2676 ldk = n->in(1);
2677 }
2678 } else if (n->Opcode() != Op_LoadKlass) {
2679 return nullptr;
2680 } else {
2681 ldk = n;
2682 }
2683 assert(ldk != nullptr && ldk->is_Load(), "should have found a LoadKlass or LoadNKlass node");
2684
2685 Node* adr = ldk->in(MemNode::Address);
2686 intptr_t off = 0;
2687 Node* obj = AddPNode::Ideal_base_and_offset(adr, gvn, off);
2688 if (obj == nullptr || off != oopDesc::klass_offset_in_bytes()) // loading oopDesc::_klass?
2689 return nullptr;
2690 const TypePtr* tp = gvn->type(obj)->is_ptr();
2691 if (tp == nullptr || !(tp->isa_instptr() || tp->isa_aryptr())) // is obj a Java object ptr?
2692 return nullptr;
2693
2694 return obj;
2695 }
2696
2697 // Matches exact and inexact type check IR shapes during parsing.
2698 // On successful match, returns type checked object node and its type after successful check
2699 // as out parameters.
2700 static bool match_type_check(PhaseGVN& gvn,
2701 BoolTest::mask btest,
2702 Node* con, const Type* tcon,
2703 Node* val, const Type* tval,
2704 Node** obj, const TypeOopPtr** cast_type) { // out-parameters
2705 assert(tcon->singleton(), "not a constant: %s", Type::str(tcon));
2706 assert(tcon == gvn.type(con), "mismatch: %s != %s", Type::str(tcon), Type::str(gvn.type(con)));
2707 assert(tval == gvn.type(val), "mismatch: %s != %s", Type::str(tval), Type::str(gvn.type(val)));
2708
2709 // Look for opportunities to sharpen the type of a node whose klass is compared with a constant klass.
2710 // The constant klass being tested against can come from many bytecode instructions (implicitly or explicitly),
2711 // and also from profile data used by speculative casts.
2712 if (btest == BoolTest::eq && tcon->isa_klassptr()) {
2713 // Found:
2714 // Bool(CmpP(LoadKlass(obj._klass), ConP(Foo.klass)), [eq])
2715 // or the narrowOop equivalent.
2716 (*obj) = extract_obj_from_klass_load(&gvn, val);
2717 // Some klass comparisons are not directly in the form
2718 // Bool(CmpP(LoadKlass(obj._klass), ConP(Foo.klass)), [eq]),
2719 // e.g. Bool(CmpP(CastPP(LoadKlass(...)), ConP(klass)), [eq]).
2720 // These patterns with nullable klasses arise from example from
2721 // load_array_klass_from_mirror.
2722 if (*obj == nullptr) { return false; }
2723 (*cast_type) = tcon->isa_klassptr()->as_exact_instance_type();
2724 return true; // found
2725 }
2726
2727 // Match an instanceof check.
2728 // During parsing its IR shape is not canonicalized yet.
2729 //
2730 // obj superklass
2731 // | |
2732 // SubTypeCheck
2733 // |
2734 // Bool [eq] / [ne]
2735 // |
2736 // If
2737 // / \
2738 // T F
2739 // \ /
2740 // Region
2741 // \ ConI ConI
2742 // \ | /
2743 // val -> Phi ConI|CastII <- con
2744 // \ /
2745 // CmpI
2746 // |
2747 // Bool [btest]
2748 // |
2749 //
2750 if (tcon->isa_int() && val->is_Phi() && val->in(0)->as_Region()->is_diamond()) {
2751 RegionNode* diamond = val->in(0)->as_Region();
2752 IfNode* if1 = diamond->in(1)->in(0)->as_If();
2753 BoolNode* b1 = if1->in(1)->isa_Bool();
2754 if (b1 != nullptr && b1->in(1)->isa_SubTypeCheck()) {
2755 assert(b1->_test._test == BoolTest::eq ||
2756 b1->_test._test == BoolTest::ne, "%d", b1->_test._test);
2757
2758 ProjNode* success_proj = if1->proj_out(b1->_test._test == BoolTest::eq ? 1 : 0);
2759 int success_idx = diamond->find_edge(success_proj);
2760 assert(success_idx == 1 || success_idx == 2, "");
2761 assert(val->req() == 3, "not a diamond");
2762
2763 // gen_instanceof() emits 1 on success and 0 on failure.
2764 // Check whether current comparison selects the success value.
2765 const Type* success_tval = gvn.type(val->in(success_idx));
2766 assert(success_tval->isa_int(), "not an int: %s", Type::str(success_tval));
2767 if ((btest == BoolTest::eq && tcon == success_tval) ||
2768 (btest == BoolTest::ne && tcon->join(success_tval)->empty())) {
2769 SubTypeCheckNode* sub = b1->in(1)->as_SubTypeCheck();
2770 Node* obj_or_subklass = sub->in(SubTypeCheckNode::ObjOrSubKlass);
2771 Node* superklass = sub->in(SubTypeCheckNode::SuperKlass);
2772
2773 if (gvn.type(obj_or_subklass)->isa_oopptr()) {
2774 const TypeKlassPtr* klass_ptr_type = gvn.type(superklass)->is_klassptr();
2775 const TypeKlassPtr* improved_klass_ptr_type = klass_ptr_type->try_improve();
2776
2777 (*obj) = obj_or_subklass;
2778 (*cast_type) = improved_klass_ptr_type->as_subtype_instance_type();
2779 return true; // found
2780 }
2781 }
2782 }
2783 }
2784 return false; // not found
2785 }
2786
2787 void Parse::sharpen_type_after_if(BoolTest::mask btest,
2788 Node* con, const Type* tcon,
2789 Node* val, const Type* tval) {
2790 Node* obj = nullptr;
2791 const TypeOopPtr* cast_type = nullptr;
2792 // Insert a cast node with a narrowed type after a successful type check.
2793 if (match_type_check(_gvn, btest, con, tcon, val, tval,
2794 &obj, &cast_type)) {
2795 assert(obj != nullptr && cast_type != nullptr, "missing type check info");
2796 const Type* obj_type = _gvn.type(obj);
2797 const Type* tboth = obj_type->filter_speculative(cast_type);
2798 assert(tboth->higher_equal(obj_type) && tboth->higher_equal(cast_type), "sanity");
2799 if (tboth == Type::TOP && KillPathsReachableByDeadDataNode) {
2800 // Let dead type node cleaning logic prune effectively dead path for us.
2801 // CheckCastPP::Value() == TOP and it will trigger the cleanup during GVN.
2802 // Don't materialize the cast when cleanup is disabled, because
2803 // it kills data and control leaving IR in broken state.
2804 tboth = cast_type;
2805 }
2806 if (tboth != Type::TOP && tboth != obj_type) {
2807 int obj_in_map = map()->find_edge(obj);
2808 if (obj_in_map >= 0 &&
2809 (jvms()->is_loc(obj_in_map) || jvms()->is_stk(obj_in_map))) {
2810 TypeNode* ccast = new CheckCastPPNode(control(), obj, tboth);
2811 // Delay transform() call to allow recovery of pre-cast value at the control merge.
2812 _gvn.set_type_bottom(ccast);
2813 record_for_igvn(ccast);
2814 if (tboth->is_inlinetypeptr()) {
2815 ccast = InlineTypeNode::make_from_oop(this, ccast, tboth->isa_oopptr()->exact_klass(true)->as_inline_klass());
2816 }
2817 // Here's the payoff.
2818 replace_in_map(obj, ccast);
2819 }
2820 }
2821 }
2822
2823 int val_in_map = map()->find_edge(val);
2824 if (val_in_map < 0) return; // replace_in_map would be useless
2825 {
2826 JVMState* jvms = this->jvms();
2827 if (!(jvms->is_loc(val_in_map) ||
2828 jvms->is_stk(val_in_map)))
2829 return; // again, it would be useless
2830 }
2831
2832 // Check for a comparison to a constant, and "know" that the compared
2833 // value is constrained on this path.
2834 assert(tcon->singleton(), "");
2835 ConstraintCastNode* ccast = nullptr;
2836 Node* cast = nullptr;
2837
2838 switch (btest) {
2839 case BoolTest::eq: // Constant test?
2840 {
2841 const Type* tboth = tcon->join_speculative(tval);
2842 if (tboth == tval) break; // Nothing to gain.
2843 if (tcon->isa_int()) {
2844 ccast = new CastIINode(control(), val, tboth);
2845 } else if (tcon == TypePtr::NULL_PTR) {
2846 // Cast to null, but keep the pointer identity temporarily live.
2847 ccast = new CastPPNode(control(), val, tboth);
2848 } else {
2849 const TypeF* tf = tcon->isa_float_constant();
2850 const TypeD* td = tcon->isa_double_constant();
2851 // Exclude tests vs float/double 0 as these could be
2852 // either +0 or -0. Just because you are equal to +0
2853 // doesn't mean you ARE +0!
2854 // Note, following code also replaces Long and Oop values.
2855 if ((!tf || tf->_f != 0.0) &&
2856 (!td || td->_d != 0.0))
2857 cast = con; // Replace non-constant val by con.
2858 }
2859 }
2860 break;
2861
2862 case BoolTest::ne:
2863 if (tcon == TypePtr::NULL_PTR) {
2864 cast = cast_not_null(val, false);
2865 }
2866 break;
2867
2868 default:
2869 // (At this point we could record int range types with CastII.)
2870 break;
2871 }
2872
2873 if (ccast != nullptr) {
2874 const Type* tcc = ccast->as_Type()->type();
2875 assert(tcc != tval && tcc->higher_equal(tval), "must improve");
2876 // Delay transform() call to allow recovery of pre-cast value
2877 // at the control merge.
2878 _gvn.set_type_bottom(ccast);
2879 record_for_igvn(ccast);
2880 cast = ccast;
2881 }
2882
2883 if (cast != nullptr) { // Here's the payoff.
2884 replace_in_map(val, cast);
2885 }
2886 }
2887
2888 /**
2889 * Use speculative type to optimize CmpP node: if comparison is
2890 * against the low level class, cast the object to the speculative
2891 * type if any. CmpP should then go away.
2892 *
2893 * @param c expected CmpP node
2894 * @return result of CmpP on object casted to speculative type
2895 *
2896 */
2897 Node* Parse::optimize_cmp_with_klass(Node* c) {
2898 // If this is transformed by the _gvn to a comparison with the low
2899 // level klass then we may be able to use speculation
2900 if (c->Opcode() == Op_CmpP &&
2901 (c->in(1)->Opcode() == Op_LoadKlass || c->in(1)->Opcode() == Op_DecodeNKlass) &&
2902 c->in(2)->is_Con()) {
2903 Node* load_klass = nullptr;
2904 Node* decode = nullptr;
2905 if (c->in(1)->Opcode() == Op_DecodeNKlass) {
2906 decode = c->in(1);
2907 load_klass = c->in(1)->in(1);
2908 } else {
2909 load_klass = c->in(1);
2910 }
2911 if (load_klass->in(2)->is_AddP()) {
2912 Node* addp = load_klass->in(2);
2913 Node* obj = addp->in(AddPNode::Address);
2914 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
2915 if (obj_type->speculative_type_not_null() != nullptr) {
2916 ciKlass* k = obj_type->speculative_type();
2917 inc_sp(2);
2918 obj = maybe_cast_profiled_obj(obj, k);
2919 dec_sp(2);
2920 if (obj->is_InlineType()) {
2921 assert(obj->as_InlineType()->is_allocated(&_gvn), "must be allocated");
2922 obj = obj->as_InlineType()->get_oop();
2923 }
2924 // Make the CmpP use the casted obj
2925 addp = basic_plus_adr(obj, addp->in(AddPNode::Offset));
2926 load_klass = load_klass->clone();
2927 load_klass->set_req(2, addp);
2928 load_klass = _gvn.transform(load_klass);
2929 if (decode != nullptr) {
2930 decode = decode->clone();
2931 decode->set_req(1, load_klass);
2932 load_klass = _gvn.transform(decode);
2933 }
2934 c = c->clone();
2935 c->set_req(1, load_klass);
2936 c = _gvn.transform(c);
2937 }
2938 }
2939 }
2940 return c;
2941 }
2942
2943 //------------------------------do_one_bytecode--------------------------------
2944 // Parse this bytecode, and alter the Parsers JVM->Node mapping
2945 void Parse::do_one_bytecode() {
2946 Node *a, *b, *c, *d; // Handy temps
2947 BoolTest::mask btest;
2948 int i;
2949
2950 assert(!has_exceptions(), "bytecode entry state must be clear of throws");
2951
2952 if (C->check_node_count(NodeLimitFudgeFactor * 5,
2953 "out of nodes parsing method")) {
2954 return;
2955 }
2956
2957 #ifdef ASSERT
2958 // for setting breakpoints
2959 if (TraceOptoParse) {
2960 tty->print(" @");
2961 dump_bci(bci());
2962 tty->print(" %s", Bytecodes::name(bc()));
2963 tty->cr();
2964 }
2965 #endif
2966
2967 switch (bc()) {
2968 case Bytecodes::_nop:
2969 // do nothing
2970 break;
2971 case Bytecodes::_lconst_0:
2972 push_pair(longcon(0));
2973 break;
2974
2975 case Bytecodes::_lconst_1:
2976 push_pair(longcon(1));
2977 break;
2978
2979 case Bytecodes::_fconst_0:
2980 push(zerocon(T_FLOAT));
2981 break;
2982
2983 case Bytecodes::_fconst_1:
2984 push(makecon(TypeF::ONE));
2985 break;
2986
2987 case Bytecodes::_fconst_2:
2988 push(makecon(TypeF::make(2.0f)));
2989 break;
2990
2991 case Bytecodes::_dconst_0:
2992 push_pair(zerocon(T_DOUBLE));
2993 break;
2994
2995 case Bytecodes::_dconst_1:
2996 push_pair(makecon(TypeD::ONE));
2997 break;
2998
2999 case Bytecodes::_iconst_m1:push(intcon(-1)); break;
3000 case Bytecodes::_iconst_0: push(intcon( 0)); break;
3001 case Bytecodes::_iconst_1: push(intcon( 1)); break;
3002 case Bytecodes::_iconst_2: push(intcon( 2)); break;
3003 case Bytecodes::_iconst_3: push(intcon( 3)); break;
3004 case Bytecodes::_iconst_4: push(intcon( 4)); break;
3005 case Bytecodes::_iconst_5: push(intcon( 5)); break;
3006 case Bytecodes::_bipush: push(intcon(iter().get_constant_u1())); break;
3007 case Bytecodes::_sipush: push(intcon(iter().get_constant_u2())); break;
3008 case Bytecodes::_aconst_null: push(null()); break;
3009
3010 case Bytecodes::_ldc:
3011 case Bytecodes::_ldc_w:
3012 case Bytecodes::_ldc2_w: {
3013 // ciTypeFlow should trap if the ldc is in error state or if the constant is not loaded
3014 assert(!iter().is_in_error(), "ldc is in error state");
3015 ciConstant constant = iter().get_constant();
3016 assert(constant.is_loaded(), "constant is not loaded");
3017 const Type* con_type = Type::make_from_constant(constant);
3018 if (con_type != nullptr) {
3019 push_node(con_type->basic_type(), makecon(con_type));
3020 }
3021 break;
3022 }
3023
3024 case Bytecodes::_aload_0:
3025 push( local(0) );
3026 break;
3027 case Bytecodes::_aload_1:
3028 push( local(1) );
3029 break;
3030 case Bytecodes::_aload_2:
3031 push( local(2) );
3032 break;
3033 case Bytecodes::_aload_3:
3034 push( local(3) );
3035 break;
3036 case Bytecodes::_aload:
3037 push( local(iter().get_index()) );
3038 break;
3039
3040 case Bytecodes::_fload_0:
3041 case Bytecodes::_iload_0:
3042 push( local(0) );
3043 break;
3044 case Bytecodes::_fload_1:
3045 case Bytecodes::_iload_1:
3046 push( local(1) );
3047 break;
3048 case Bytecodes::_fload_2:
3049 case Bytecodes::_iload_2:
3050 push( local(2) );
3051 break;
3052 case Bytecodes::_fload_3:
3053 case Bytecodes::_iload_3:
3054 push( local(3) );
3055 break;
3056 case Bytecodes::_fload:
3057 case Bytecodes::_iload:
3058 push( local(iter().get_index()) );
3059 break;
3060 case Bytecodes::_lload_0:
3061 push_pair_local( 0 );
3062 break;
3063 case Bytecodes::_lload_1:
3064 push_pair_local( 1 );
3065 break;
3066 case Bytecodes::_lload_2:
3067 push_pair_local( 2 );
3068 break;
3069 case Bytecodes::_lload_3:
3070 push_pair_local( 3 );
3071 break;
3072 case Bytecodes::_lload:
3073 push_pair_local( iter().get_index() );
3074 break;
3075
3076 case Bytecodes::_dload_0:
3077 push_pair_local(0);
3078 break;
3079 case Bytecodes::_dload_1:
3080 push_pair_local(1);
3081 break;
3082 case Bytecodes::_dload_2:
3083 push_pair_local(2);
3084 break;
3085 case Bytecodes::_dload_3:
3086 push_pair_local(3);
3087 break;
3088 case Bytecodes::_dload:
3089 push_pair_local(iter().get_index());
3090 break;
3091 case Bytecodes::_fstore_0:
3092 case Bytecodes::_istore_0:
3093 case Bytecodes::_astore_0:
3094 set_local( 0, pop() );
3095 break;
3096 case Bytecodes::_fstore_1:
3097 case Bytecodes::_istore_1:
3098 case Bytecodes::_astore_1:
3099 set_local( 1, pop() );
3100 break;
3101 case Bytecodes::_fstore_2:
3102 case Bytecodes::_istore_2:
3103 case Bytecodes::_astore_2:
3104 set_local( 2, pop() );
3105 break;
3106 case Bytecodes::_fstore_3:
3107 case Bytecodes::_istore_3:
3108 case Bytecodes::_astore_3:
3109 set_local( 3, pop() );
3110 break;
3111 case Bytecodes::_fstore:
3112 case Bytecodes::_istore:
3113 case Bytecodes::_astore:
3114 set_local( iter().get_index(), pop() );
3115 break;
3116 // long stores
3117 case Bytecodes::_lstore_0:
3118 set_pair_local( 0, pop_pair() );
3119 break;
3120 case Bytecodes::_lstore_1:
3121 set_pair_local( 1, pop_pair() );
3122 break;
3123 case Bytecodes::_lstore_2:
3124 set_pair_local( 2, pop_pair() );
3125 break;
3126 case Bytecodes::_lstore_3:
3127 set_pair_local( 3, pop_pair() );
3128 break;
3129 case Bytecodes::_lstore:
3130 set_pair_local( iter().get_index(), pop_pair() );
3131 break;
3132
3133 // double stores
3134 case Bytecodes::_dstore_0:
3135 set_pair_local( 0, pop_pair() );
3136 break;
3137 case Bytecodes::_dstore_1:
3138 set_pair_local( 1, pop_pair() );
3139 break;
3140 case Bytecodes::_dstore_2:
3141 set_pair_local( 2, pop_pair() );
3142 break;
3143 case Bytecodes::_dstore_3:
3144 set_pair_local( 3, pop_pair() );
3145 break;
3146 case Bytecodes::_dstore:
3147 set_pair_local( iter().get_index(), pop_pair() );
3148 break;
3149
3150 case Bytecodes::_pop: dec_sp(1); break;
3151 case Bytecodes::_pop2: dec_sp(2); break;
3152 case Bytecodes::_swap:
3153 a = pop();
3154 b = pop();
3155 push(a);
3156 push(b);
3157 break;
3158 case Bytecodes::_dup:
3159 a = pop();
3160 push(a);
3161 push(a);
3162 break;
3163 case Bytecodes::_dup_x1:
3164 a = pop();
3165 b = pop();
3166 push( a );
3167 push( b );
3168 push( a );
3169 break;
3170 case Bytecodes::_dup_x2:
3171 a = pop();
3172 b = pop();
3173 c = pop();
3174 push( a );
3175 push( c );
3176 push( b );
3177 push( a );
3178 break;
3179 case Bytecodes::_dup2:
3180 a = pop();
3181 b = pop();
3182 push( b );
3183 push( a );
3184 push( b );
3185 push( a );
3186 break;
3187
3188 case Bytecodes::_dup2_x1:
3189 // before: .. c, b, a
3190 // after: .. b, a, c, b, a
3191 // not tested
3192 a = pop();
3193 b = pop();
3194 c = pop();
3195 push( b );
3196 push( a );
3197 push( c );
3198 push( b );
3199 push( a );
3200 break;
3201 case Bytecodes::_dup2_x2:
3202 // before: .. d, c, b, a
3203 // after: .. b, a, d, c, b, a
3204 // not tested
3205 a = pop();
3206 b = pop();
3207 c = pop();
3208 d = pop();
3209 push( b );
3210 push( a );
3211 push( d );
3212 push( c );
3213 push( b );
3214 push( a );
3215 break;
3216
3217 case Bytecodes::_arraylength: {
3218 // Must do null-check with value on expression stack
3219 Node *ary = null_check(peek(), T_ARRAY);
3220 // Compile-time detect of null-exception?
3221 if (stopped()) return;
3222 a = pop();
3223 push(load_array_length(a));
3224 break;
3225 }
3226
3227 case Bytecodes::_baload: array_load(T_BYTE); break;
3228 case Bytecodes::_caload: array_load(T_CHAR); break;
3229 case Bytecodes::_iaload: array_load(T_INT); break;
3230 case Bytecodes::_saload: array_load(T_SHORT); break;
3231 case Bytecodes::_faload: array_load(T_FLOAT); break;
3232 case Bytecodes::_aaload: array_load(T_OBJECT); break;
3233 case Bytecodes::_laload: array_load(T_LONG); break;
3234 case Bytecodes::_daload: array_load(T_DOUBLE); break;
3235 case Bytecodes::_bastore: array_store(T_BYTE); break;
3236 case Bytecodes::_castore: array_store(T_CHAR); break;
3237 case Bytecodes::_iastore: array_store(T_INT); break;
3238 case Bytecodes::_sastore: array_store(T_SHORT); break;
3239 case Bytecodes::_fastore: array_store(T_FLOAT); break;
3240 case Bytecodes::_aastore: array_store(T_OBJECT); break;
3241 case Bytecodes::_lastore: array_store(T_LONG); break;
3242 case Bytecodes::_dastore: array_store(T_DOUBLE); break;
3243
3244 case Bytecodes::_getfield:
3245 do_getfield();
3246 break;
3247
3248 case Bytecodes::_getstatic:
3249 do_getstatic();
3250 break;
3251
3252 case Bytecodes::_putfield:
3253 do_putfield();
3254 break;
3255
3256 case Bytecodes::_putstatic:
3257 do_putstatic();
3258 break;
3259
3260 case Bytecodes::_irem:
3261 // Must keep both values on the expression-stack during null-check
3262 zero_check_int(peek());
3263 // Compile-time detect of null-exception?
3264 if (stopped()) return;
3265 b = pop();
3266 a = pop();
3267 push(_gvn.transform(new ModINode(control(), a, b)));
3268 break;
3269 case Bytecodes::_idiv:
3270 // Must keep both values on the expression-stack during null-check
3271 zero_check_int(peek());
3272 // Compile-time detect of null-exception?
3273 if (stopped()) return;
3274 b = pop();
3275 a = pop();
3276 push( _gvn.transform( new DivINode(control(),a,b) ) );
3277 break;
3278 case Bytecodes::_imul:
3279 b = pop(); a = pop();
3280 push( _gvn.transform( new MulINode(a,b) ) );
3281 break;
3282 case Bytecodes::_iadd:
3283 b = pop(); a = pop();
3284 push( _gvn.transform( new AddINode(a,b) ) );
3285 break;
3286 case Bytecodes::_ineg:
3287 a = pop();
3288 push( _gvn.transform( new SubINode(_gvn.intcon(0),a)) );
3289 break;
3290 case Bytecodes::_isub:
3291 b = pop(); a = pop();
3292 push( _gvn.transform( new SubINode(a,b) ) );
3293 break;
3294 case Bytecodes::_iand:
3295 b = pop(); a = pop();
3296 push( _gvn.transform( new AndINode(a,b) ) );
3297 break;
3298 case Bytecodes::_ior:
3299 b = pop(); a = pop();
3300 push( _gvn.transform( new OrINode(a,b) ) );
3301 break;
3302 case Bytecodes::_ixor:
3303 b = pop(); a = pop();
3304 push( _gvn.transform( new XorINode(a,b) ) );
3305 break;
3306 case Bytecodes::_ishl:
3307 b = pop(); a = pop();
3308 push( _gvn.transform( new LShiftINode(a,b) ) );
3309 break;
3310 case Bytecodes::_ishr:
3311 b = pop(); a = pop();
3312 push( _gvn.transform( new RShiftINode(a,b) ) );
3313 break;
3314 case Bytecodes::_iushr:
3315 b = pop(); a = pop();
3316 push( _gvn.transform( new URShiftINode(a,b) ) );
3317 break;
3318
3319 case Bytecodes::_fneg:
3320 a = pop();
3321 b = _gvn.transform(new NegFNode (a));
3322 push(b);
3323 break;
3324
3325 case Bytecodes::_fsub:
3326 b = pop();
3327 a = pop();
3328 c = _gvn.transform( new SubFNode(a,b) );
3329 push(c);
3330 break;
3331
3332 case Bytecodes::_fadd:
3333 b = pop();
3334 a = pop();
3335 c = _gvn.transform( new AddFNode(a,b) );
3336 push(c);
3337 break;
3338
3339 case Bytecodes::_fmul:
3340 b = pop();
3341 a = pop();
3342 c = _gvn.transform( new MulFNode(a,b) );
3343 push(c);
3344 break;
3345
3346 case Bytecodes::_fdiv:
3347 b = pop();
3348 a = pop();
3349 c = _gvn.transform( new DivFNode(nullptr,a,b) );
3350 push(c);
3351 break;
3352
3353 case Bytecodes::_frem:
3354 // Generate a ModF node.
3355 b = pop();
3356 a = pop();
3357 push(floating_point_mod(a, b, BasicType::T_FLOAT));
3358 break;
3359
3360 case Bytecodes::_fcmpl:
3361 b = pop();
3362 a = pop();
3363 c = _gvn.transform( new CmpF3Node( a, b));
3364 push(c);
3365 break;
3366 case Bytecodes::_fcmpg:
3367 b = pop();
3368 a = pop();
3369
3370 // Same as fcmpl but need to flip the unordered case. Swap the inputs,
3371 // which negates the result sign except for unordered. Flip the unordered
3372 // as well by using CmpF3 which implements unordered-lesser instead of
3373 // unordered-greater semantics. Finally, commute the result bits. Result
3374 // is same as using a CmpF3Greater except we did it with CmpF3 alone.
3375 c = _gvn.transform( new CmpF3Node( b, a));
3376 c = _gvn.transform( new SubINode(_gvn.intcon(0),c) );
3377 push(c);
3378 break;
3379
3380 case Bytecodes::_f2i:
3381 a = pop();
3382 push(_gvn.transform(new ConvF2INode(a)));
3383 break;
3384
3385 case Bytecodes::_d2i:
3386 a = pop_pair();
3387 b = _gvn.transform(new ConvD2INode(a));
3388 push( b );
3389 break;
3390
3391 case Bytecodes::_f2d:
3392 a = pop();
3393 b = _gvn.transform( new ConvF2DNode(a));
3394 push_pair( b );
3395 break;
3396
3397 case Bytecodes::_d2f:
3398 a = pop_pair();
3399 b = _gvn.transform( new ConvD2FNode(a));
3400 push( b );
3401 break;
3402
3403 case Bytecodes::_l2f:
3404 if (Matcher::convL2FSupported()) {
3405 a = pop_pair();
3406 b = _gvn.transform( new ConvL2FNode(a));
3407 push(b);
3408 } else {
3409 l2f();
3410 }
3411 break;
3412
3413 case Bytecodes::_l2d:
3414 a = pop_pair();
3415 b = _gvn.transform( new ConvL2DNode(a));
3416 push_pair(b);
3417 break;
3418
3419 case Bytecodes::_f2l:
3420 a = pop();
3421 b = _gvn.transform( new ConvF2LNode(a));
3422 push_pair(b);
3423 break;
3424
3425 case Bytecodes::_d2l:
3426 a = pop_pair();
3427 b = _gvn.transform( new ConvD2LNode(a));
3428 push_pair(b);
3429 break;
3430
3431 case Bytecodes::_dsub:
3432 b = pop_pair();
3433 a = pop_pair();
3434 c = _gvn.transform( new SubDNode(a,b) );
3435 push_pair(c);
3436 break;
3437
3438 case Bytecodes::_dadd:
3439 b = pop_pair();
3440 a = pop_pair();
3441 c = _gvn.transform( new AddDNode(a,b) );
3442 push_pair(c);
3443 break;
3444
3445 case Bytecodes::_dmul:
3446 b = pop_pair();
3447 a = pop_pair();
3448 c = _gvn.transform( new MulDNode(a,b) );
3449 push_pair(c);
3450 break;
3451
3452 case Bytecodes::_ddiv:
3453 b = pop_pair();
3454 a = pop_pair();
3455 c = _gvn.transform( new DivDNode(nullptr,a,b) );
3456 push_pair(c);
3457 break;
3458
3459 case Bytecodes::_dneg:
3460 a = pop_pair();
3461 b = _gvn.transform(new NegDNode (a));
3462 push_pair(b);
3463 break;
3464
3465 case Bytecodes::_drem:
3466 // Generate a ModD node.
3467 b = pop_pair();
3468 a = pop_pair();
3469 push_pair(floating_point_mod(a, b, BasicType::T_DOUBLE));
3470 break;
3471
3472 case Bytecodes::_dcmpl:
3473 b = pop_pair();
3474 a = pop_pair();
3475 c = _gvn.transform( new CmpD3Node( a, b));
3476 push(c);
3477 break;
3478
3479 case Bytecodes::_dcmpg:
3480 b = pop_pair();
3481 a = pop_pair();
3482 // Same as dcmpl but need to flip the unordered case.
3483 // Commute the inputs, which negates the result sign except for unordered.
3484 // Flip the unordered as well by using CmpD3 which implements
3485 // unordered-lesser instead of unordered-greater semantics.
3486 // Finally, negate the result bits. Result is same as using a
3487 // CmpD3Greater except we did it with CmpD3 alone.
3488 c = _gvn.transform( new CmpD3Node( b, a));
3489 c = _gvn.transform( new SubINode(_gvn.intcon(0),c) );
3490 push(c);
3491 break;
3492
3493
3494 // Note for longs -> lo word is on TOS, hi word is on TOS - 1
3495 case Bytecodes::_land:
3496 b = pop_pair();
3497 a = pop_pair();
3498 c = _gvn.transform( new AndLNode(a,b) );
3499 push_pair(c);
3500 break;
3501 case Bytecodes::_lor:
3502 b = pop_pair();
3503 a = pop_pair();
3504 c = _gvn.transform( new OrLNode(a,b) );
3505 push_pair(c);
3506 break;
3507 case Bytecodes::_lxor:
3508 b = pop_pair();
3509 a = pop_pair();
3510 c = _gvn.transform( new XorLNode(a,b) );
3511 push_pair(c);
3512 break;
3513
3514 case Bytecodes::_lshl:
3515 b = pop(); // the shift count
3516 a = pop_pair(); // value to be shifted
3517 c = _gvn.transform( new LShiftLNode(a,b) );
3518 push_pair(c);
3519 break;
3520 case Bytecodes::_lshr:
3521 b = pop(); // the shift count
3522 a = pop_pair(); // value to be shifted
3523 c = _gvn.transform( new RShiftLNode(a,b) );
3524 push_pair(c);
3525 break;
3526 case Bytecodes::_lushr:
3527 b = pop(); // the shift count
3528 a = pop_pair(); // value to be shifted
3529 c = _gvn.transform( new URShiftLNode(a,b) );
3530 push_pair(c);
3531 break;
3532 case Bytecodes::_lmul:
3533 b = pop_pair();
3534 a = pop_pair();
3535 c = _gvn.transform( new MulLNode(a,b) );
3536 push_pair(c);
3537 break;
3538
3539 case Bytecodes::_lrem:
3540 // Must keep both values on the expression-stack during null-check
3541 assert(peek(0) == top(), "long word order");
3542 zero_check_long(peek(1));
3543 // Compile-time detect of null-exception?
3544 if (stopped()) return;
3545 b = pop_pair();
3546 a = pop_pair();
3547 c = _gvn.transform( new ModLNode(control(),a,b) );
3548 push_pair(c);
3549 break;
3550
3551 case Bytecodes::_ldiv:
3552 // Must keep both values on the expression-stack during null-check
3553 assert(peek(0) == top(), "long word order");
3554 zero_check_long(peek(1));
3555 // Compile-time detect of null-exception?
3556 if (stopped()) return;
3557 b = pop_pair();
3558 a = pop_pair();
3559 c = _gvn.transform( new DivLNode(control(),a,b) );
3560 push_pair(c);
3561 break;
3562
3563 case Bytecodes::_ladd:
3564 b = pop_pair();
3565 a = pop_pair();
3566 c = _gvn.transform( new AddLNode(a,b) );
3567 push_pair(c);
3568 break;
3569 case Bytecodes::_lsub:
3570 b = pop_pair();
3571 a = pop_pair();
3572 c = _gvn.transform( new SubLNode(a,b) );
3573 push_pair(c);
3574 break;
3575 case Bytecodes::_lcmp:
3576 // Safepoints are now inserted _before_ branches. The long-compare
3577 // bytecode painfully produces a 3-way value (-1,0,+1) which requires a
3578 // slew of control flow. These are usually followed by a CmpI vs zero and
3579 // a branch; this pattern then optimizes to the obvious long-compare and
3580 // branch. However, if the branch is backwards there's a Safepoint
3581 // inserted. The inserted Safepoint captures the JVM state at the
3582 // pre-branch point, i.e. it captures the 3-way value. Thus if a
3583 // long-compare is used to control a loop the debug info will force
3584 // computation of the 3-way value, even though the generated code uses a
3585 // long-compare and branch. We try to rectify the situation by inserting
3586 // a SafePoint here and have it dominate and kill the safepoint added at a
3587 // following backwards branch. At this point the JVM state merely holds 2
3588 // longs but not the 3-way value.
3589 switch (iter().next_bc()) {
3590 case Bytecodes::_ifgt:
3591 case Bytecodes::_iflt:
3592 case Bytecodes::_ifge:
3593 case Bytecodes::_ifle:
3594 case Bytecodes::_ifne:
3595 case Bytecodes::_ifeq:
3596 // If this is a backwards branch in the bytecodes, add Safepoint
3597 maybe_add_safepoint(iter().next_get_dest());
3598 default:
3599 break;
3600 }
3601 b = pop_pair();
3602 a = pop_pair();
3603 c = _gvn.transform( new CmpL3Node( a, b ));
3604 push(c);
3605 break;
3606
3607 case Bytecodes::_lneg:
3608 a = pop_pair();
3609 b = _gvn.transform( new SubLNode(longcon(0),a));
3610 push_pair(b);
3611 break;
3612 case Bytecodes::_l2i:
3613 a = pop_pair();
3614 push( _gvn.transform( new ConvL2INode(a)));
3615 break;
3616 case Bytecodes::_i2l:
3617 a = pop();
3618 b = _gvn.transform( new ConvI2LNode(a));
3619 push_pair(b);
3620 break;
3621 case Bytecodes::_i2b:
3622 // Sign extend
3623 a = pop();
3624 a = Compile::narrow_value(T_BYTE, a, TypeInt::BYTE, &_gvn, true);
3625 push(a);
3626 break;
3627 case Bytecodes::_i2s:
3628 a = pop();
3629 a = Compile::narrow_value(T_SHORT, a, TypeInt::SHORT, &_gvn, true);
3630 push(a);
3631 break;
3632 case Bytecodes::_i2c:
3633 a = pop();
3634 a = Compile::narrow_value(T_CHAR, a, TypeInt::CHAR, &_gvn, true);
3635 push(a);
3636 break;
3637
3638 case Bytecodes::_i2f:
3639 a = pop();
3640 b = _gvn.transform( new ConvI2FNode(a) ) ;
3641 push(b);
3642 break;
3643
3644 case Bytecodes::_i2d:
3645 a = pop();
3646 b = _gvn.transform( new ConvI2DNode(a));
3647 push_pair(b);
3648 break;
3649
3650 case Bytecodes::_iinc: // Increment local
3651 i = iter().get_index(); // Get local index
3652 set_local( i, _gvn.transform( new AddINode( _gvn.intcon(iter().get_iinc_con()), local(i) ) ) );
3653 break;
3654
3655 // Exit points of synchronized methods must have an unlock node
3656 case Bytecodes::_return:
3657 return_current(nullptr);
3658 break;
3659
3660 case Bytecodes::_ireturn:
3661 case Bytecodes::_areturn:
3662 case Bytecodes::_freturn:
3663 return_current(pop());
3664 break;
3665 case Bytecodes::_lreturn:
3666 case Bytecodes::_dreturn:
3667 return_current(pop_pair());
3668 break;
3669
3670 case Bytecodes::_athrow:
3671 // null exception oop throws null pointer exception
3672 null_check(peek());
3673 if (stopped()) return;
3674 // Hook the thrown exception directly to subsequent handlers.
3675 if (BailoutToInterpreterForThrows) {
3676 // Keep method interpreted from now on.
3677 uncommon_trap(Deoptimization::Reason_unhandled,
3678 Deoptimization::Action_make_not_compilable);
3679 return;
3680 }
3681 if (env()->jvmti_can_post_on_exceptions()) {
3682 // check if we must post exception events, take uncommon trap if so (with must_throw = false)
3683 uncommon_trap_if_should_post_on_exceptions(Deoptimization::Reason_unhandled, false);
3684 }
3685 // Here if either can_post_on_exceptions or should_post_on_exceptions is false
3686 add_exception_state(make_exception_state(peek()));
3687 break;
3688
3689 case Bytecodes::_goto: // fall through
3690 case Bytecodes::_goto_w: {
3691 int target_bci = (bc() == Bytecodes::_goto) ? iter().get_dest() : iter().get_far_dest();
3692
3693 // If this is a backwards branch in the bytecodes, add Safepoint
3694 maybe_add_safepoint(target_bci);
3695
3696 // Merge the current control into the target basic block
3697 merge(target_bci);
3698
3699 // See if we can get some profile data and hand it off to the next block
3700 Block *target_block = block()->successor_for_bci(target_bci);
3701 if (target_block->pred_count() != 1) break;
3702 ciMethodData* methodData = method()->method_data();
3703 if (!methodData->is_mature()) break;
3704 ciProfileData* data = methodData->bci_to_data(bci());
3705 assert(data != nullptr && data->is_JumpData(), "need JumpData for taken branch");
3706 int taken = ((ciJumpData*)data)->taken();
3707 taken = method()->scale_count(taken);
3708 target_block->set_count(taken);
3709 break;
3710 }
3711
3712 case Bytecodes::_ifnull: btest = BoolTest::eq; goto handle_if_null;
3713 case Bytecodes::_ifnonnull: btest = BoolTest::ne; goto handle_if_null;
3714 handle_if_null:
3715 // If this is a backwards branch in the bytecodes, add Safepoint
3716 maybe_add_safepoint(iter().get_dest());
3717 a = null();
3718 b = pop();
3719 if (b->is_InlineType()) {
3720 // Null checking a scalarized but nullable inline type. Check the null marker
3721 // input instead of the oop input to avoid keeping buffer allocations alive
3722 c = _gvn.transform(new CmpINode(b->as_InlineType()->get_null_marker(), zerocon(T_INT)));
3723 } else {
3724 if (!_gvn.type(b)->speculative_maybe_null() &&
3725 !too_many_traps(Deoptimization::Reason_speculate_null_check)) {
3726 inc_sp(1);
3727 Node* null_ctl = top();
3728 b = null_check_oop(b, &null_ctl, true, true, true);
3729 assert(null_ctl->is_top(), "no null control here");
3730 dec_sp(1);
3731 } else if (_gvn.type(b)->speculative_always_null() &&
3732 !too_many_traps(Deoptimization::Reason_speculate_null_assert)) {
3733 inc_sp(1);
3734 b = null_assert(b);
3735 dec_sp(1);
3736 }
3737 c = _gvn.transform( new CmpPNode(b, a) );
3738 }
3739 do_ifnull(btest, c);
3740 break;
3741
3742 case Bytecodes::_if_acmpeq: btest = BoolTest::eq; goto handle_if_acmp;
3743 case Bytecodes::_if_acmpne: btest = BoolTest::ne; goto handle_if_acmp;
3744 handle_if_acmp:
3745 // If this is a backwards branch in the bytecodes, add Safepoint
3746 maybe_add_safepoint(iter().get_dest());
3747 a = pop();
3748 b = pop();
3749 do_acmp(btest, b, a);
3750 break;
3751
3752 case Bytecodes::_ifeq: btest = BoolTest::eq; goto handle_ifxx;
3753 case Bytecodes::_ifne: btest = BoolTest::ne; goto handle_ifxx;
3754 case Bytecodes::_iflt: btest = BoolTest::lt; goto handle_ifxx;
3755 case Bytecodes::_ifle: btest = BoolTest::le; goto handle_ifxx;
3756 case Bytecodes::_ifgt: btest = BoolTest::gt; goto handle_ifxx;
3757 case Bytecodes::_ifge: btest = BoolTest::ge; goto handle_ifxx;
3758 handle_ifxx:
3759 // If this is a backwards branch in the bytecodes, add Safepoint
3760 maybe_add_safepoint(iter().get_dest());
3761 a = _gvn.intcon(0);
3762 b = pop();
3763 c = _gvn.transform( new CmpINode(b, a) );
3764 do_if(btest, c);
3765 break;
3766
3767 case Bytecodes::_if_icmpeq: btest = BoolTest::eq; goto handle_if_icmp;
3768 case Bytecodes::_if_icmpne: btest = BoolTest::ne; goto handle_if_icmp;
3769 case Bytecodes::_if_icmplt: btest = BoolTest::lt; goto handle_if_icmp;
3770 case Bytecodes::_if_icmple: btest = BoolTest::le; goto handle_if_icmp;
3771 case Bytecodes::_if_icmpgt: btest = BoolTest::gt; goto handle_if_icmp;
3772 case Bytecodes::_if_icmpge: btest = BoolTest::ge; goto handle_if_icmp;
3773 handle_if_icmp:
3774 // If this is a backwards branch in the bytecodes, add Safepoint
3775 maybe_add_safepoint(iter().get_dest());
3776 a = pop();
3777 b = pop();
3778 c = _gvn.transform( new CmpINode( b, a ) );
3779 do_if(btest, c);
3780 break;
3781
3782 case Bytecodes::_tableswitch:
3783 do_tableswitch();
3784 break;
3785
3786 case Bytecodes::_lookupswitch:
3787 do_lookupswitch();
3788 break;
3789
3790 case Bytecodes::_invokestatic:
3791 case Bytecodes::_invokedynamic:
3792 case Bytecodes::_invokespecial:
3793 case Bytecodes::_invokevirtual:
3794 case Bytecodes::_invokeinterface:
3795 do_call();
3796 break;
3797 case Bytecodes::_checkcast:
3798 do_checkcast();
3799 break;
3800 case Bytecodes::_instanceof:
3801 do_instanceof();
3802 break;
3803 case Bytecodes::_anewarray:
3804 do_newarray();
3805 break;
3806 case Bytecodes::_newarray:
3807 do_newarray((BasicType)iter().get_index());
3808 break;
3809 case Bytecodes::_multianewarray:
3810 do_multianewarray();
3811 break;
3812 case Bytecodes::_new:
3813 do_new();
3814 break;
3815
3816 case Bytecodes::_jsr:
3817 case Bytecodes::_jsr_w:
3818 do_jsr();
3819 break;
3820
3821 case Bytecodes::_ret:
3822 do_ret();
3823 break;
3824
3825
3826 case Bytecodes::_monitorenter:
3827 do_monitor_enter();
3828 break;
3829
3830 case Bytecodes::_monitorexit:
3831 do_monitor_exit();
3832 break;
3833
3834 case Bytecodes::_breakpoint:
3835 // Breakpoint set concurrently to compile
3836 // %%% use an uncommon trap?
3837 C->record_failure("breakpoint in method");
3838 return;
3839
3840 default:
3841 #ifndef PRODUCT
3842 map()->dump(99);
3843 #endif
3844 tty->print("\nUnhandled bytecode %s\n", Bytecodes::name(bc()) );
3845 ShouldNotReachHere();
3846 }
3847
3848 #ifndef PRODUCT
3849 if (failing()) { return; }
3850 constexpr int perBytecode = 6;
3851 if (C->should_print_igv(perBytecode)) {
3852 IdealGraphPrinter* printer = C->igv_printer();
3853 char buffer[256];
3854 jio_snprintf(buffer, sizeof(buffer), "Bytecode %d: %s", bci(), Bytecodes::name(bc()));
3855 bool old = printer->traverse_outs();
3856 printer->set_traverse_outs(true);
3857 printer->set_parse(this);
3858 printer->print_graph(buffer);
3859 printer->set_traverse_outs(old);
3860 printer->set_parse(nullptr);
3861 }
3862 #endif
3863 }