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 IdealKit ideal(this);
259 ideal.if_then(flat_array_test(array, /* flat = */ false)); {
260 // Non-flat array
261 if (!array_type->is_flat()) {
262 sync_kit(ideal);
263 assert(array_type->is_flat() || ideal.ctrl()->in(0)->as_If()->is_flat_array_check(&_gvn), "Should be found");
264 inc_sp(3);
265 access_store_at(array, adr, adr_type, stored_value_casted, elemtype, bt, MO_UNORDERED | IN_HEAP | IS_ARRAY, false);
266 dec_sp(3);
267 ideal.sync_kit(this);
268 }
269 } ideal.else_(); {
270 // Flat array
271 sync_kit(ideal);
272 if (!array_type->is_not_flat()) {
273 // Try to determine the inline klass type of the stored value
274 ciInlineKlass* vk = nullptr;
275 if (stored_value_casted_type->is_inlinetypeptr()) {
276 vk = stored_value_casted_type->inline_klass();
277 } else if (elemtype->is_inlinetypeptr()) {
278 vk = elemtype->inline_klass();
279 }
280
281 if (vk != nullptr) {
282 // Element type is known, cast and store to flat array layout.
283 Node* flat_array = cast_to_flat_array(array, vk);
284
285 // Re-execute flat array store if buffering triggers deoptimization
286 PreserveReexecuteState preexecs(this);
287 jvms()->set_should_reexecute(true);
288 inc_sp(3);
289
290 if (!stored_value_casted->is_InlineType()) {
291 assert(_gvn.type(stored_value_casted) == TypePtr::NULL_PTR, "Unexpected value");
292 stored_value_casted = InlineTypeNode::make_null(_gvn, vk);
293 }
294
295 stored_value_casted->as_InlineType()->store_flat_array(this, flat_array, array_index);
296 } else {
297 // Element type is unknown, emit a runtime call since the flat array layout is not statically known.
298 store_to_unknown_flat_array(array, array_index, stored_value_casted);
299 }
300 }
301 ideal.sync_kit(this);
302 }
303 ideal.end_if();
304 sync_kit(ideal);
305 return;
306 } else if (!array_type->is_not_null_free()) {
307 // Array is not flat but may be null free
308 assert(elemtype->is_oopptr()->can_be_inline_type(), "array can't be null-free");
309 array = inline_array_null_guard(array, stored_value_casted, 3, true);
310 }
311 }
312 inc_sp(3);
313 access_store_at(array, adr, adr_type, stored_value, elemtype, bt, MO_UNORDERED | IN_HEAP | IS_ARRAY);
314 dec_sp(3);
315 }
316
317 // Emit a runtime call to store to a flat array whose element type is either unknown (i.e. we do not know the flat
318 // array layout) or not exact (could have different flat array layouts at runtime).
319 void Parse::store_to_unknown_flat_array(Node* array, Node* const idx, Node* non_null_stored_value) {
320 // Below membars keep this access to an unknown flat array correctly
321 // ordered with other unknown and known flat array accesses.
322 insert_mem_bar_volatile(Op_MemBarCPUOrder, C->get_alias_index(TypeAryPtr::INLINES));
323
324 Node* call = nullptr;
325 {
326 // Re-execute flat array store if runtime call triggers deoptimization
327 PreserveReexecuteState preexecs(this);
328 jvms()->set_bci(_bci);
329 jvms()->set_should_reexecute(true);
330 inc_sp(3);
331 kill_dead_locals();
332 call = make_runtime_call(RC_NO_LEAF | RC_NO_IO,
333 OptoRuntime::store_unknown_inline_Type(),
334 OptoRuntime::store_unknown_inline_Java(),
335 nullptr, TypeRawPtr::BOTTOM,
336 non_null_stored_value, array, idx);
337 }
338 make_slow_call_ex(call, env()->Throwable_klass(), false);
339
340 insert_mem_bar_volatile(Op_MemBarCPUOrder, C->get_alias_index(TypeAryPtr::INLINES));
341 }
342
343 //------------------------------array_addressing-------------------------------
344 // Pull array and index from the stack. Compute pointer-to-element.
345 Node* Parse::prepare_array_addressing(BasicType type, int vals, const Type*& elemtype) {
346 Node *idx = peek(0+vals); // Get from stack without popping
347 Node *ary = peek(1+vals); // in case of exception
348
349 // Null check the array base, with correct stack contents
350 ary = null_check(ary, T_ARRAY);
351 // Compile-time detect of null-exception?
352 if (stopped()) return top();
353
354 const TypeAryPtr* arytype = _gvn.type(ary)->is_aryptr();
355 const TypeInt* sizetype = arytype->size();
356 elemtype = arytype->elem();
357
358 if (UseUniqueSubclasses) {
359 const Type* el = elemtype->make_ptr();
360 if (el && el->isa_instptr()) {
361 const TypeInstPtr* toop = el->is_instptr();
362 if (toop->instance_klass()->unique_concrete_subklass()) {
363 // If we load from "AbstractClass[]" we must see "ConcreteSubClass".
364 const Type* subklass = Type::get_const_type(toop->instance_klass());
365 elemtype = subklass->join_speculative(el);
366 }
367 }
368 }
369
370 if (!arytype->is_loaded()) {
371 // Only fails for some -Xcomp runs
372 // The class is unloaded. We have to run this bytecode in the interpreter.
373 ciKlass* klass = arytype->unloaded_klass();
374
375 uncommon_trap(Deoptimization::Reason_unloaded,
376 Deoptimization::Action_reinterpret,
377 klass, "!loaded array");
378 return top();
379 }
380
381 ary = create_speculative_inline_type_array_checks(ary, arytype, elemtype);
382
383 if (needs_range_check(sizetype, idx)) {
384 create_range_check(idx, ary, sizetype);
385 } else if (C->log() != nullptr) {
386 C->log()->elem("observe that='!need_range_check'");
387 }
388
389 // Check for always knowing you are throwing a range-check exception
390 if (stopped()) return top();
391
392 return ary;
393 }
394
395 Node* Parse::get_ptr_to_array_element(Node* array, Node* idx, BasicType elembt, const TypeInt* sizetype, Node* control) {
396 // Make array address computation control dependent to prevent it
397 // from floating above the range check during loop optimizations.
398 Node* ptr = array_element_address(array, idx, elembt, sizetype, control);
399 assert(ptr != top(), "top should go hand-in-hand with stopped");
400
401 return ptr;
402 }
403
404 // 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
405 // be greater or equal the smallest possible array size (i.e. out-of-bounds).
406 bool Parse::needs_range_check(const TypeInt* size_type, const Node* index) const {
407 const TypeInt* index_type = _gvn.type(index)->is_int();
408 return index_type->_hi >= size_type->_lo || index_type->_lo < 0;
409 }
410
411 void Parse::create_range_check(Node* idx, Node* ary, const TypeInt* sizetype) {
412 Node* tst;
413 if (sizetype->_hi <= 0) {
414 // The greatest array bound is negative, so we can conclude that we're
415 // compiling unreachable code, but the unsigned compare trick used below
416 // only works with non-negative lengths. Instead, hack "tst" to be zero so
417 // the uncommon_trap path will always be taken.
418 tst = _gvn.intcon(0);
419 } else {
420 // Range is constant in array-oop, so we can use the original state of mem
421 Node* len = load_array_length(ary);
422
423 // Test length vs index (standard trick using unsigned compare)
424 Node* chk = _gvn.transform(new CmpUNode(idx, len) );
425 BoolTest::mask btest = BoolTest::lt;
426 tst = _gvn.transform(new BoolNode(chk, btest) );
427 }
428 RangeCheckNode* rc = new RangeCheckNode(control(), tst, PROB_MAX, COUNT_UNKNOWN);
429 _gvn.set_type(rc, rc->Value(&_gvn));
430 if (!tst->is_Con()) {
431 record_for_igvn(rc);
432 }
433 set_control(_gvn.transform(new IfTrueNode(rc)));
434 // Branch to failure if out of bounds
435 {
436 PreserveJVMState pjvms(this);
437 set_control(_gvn.transform(new IfFalseNode(rc)));
438 if (C->allow_range_check_smearing()) {
439 // Do not use builtin_throw, since range checks are sometimes
440 // made more stringent by an optimistic transformation.
441 // This creates "tentative" range checks at this point,
442 // which are not guaranteed to throw exceptions.
443 // See IfNode::Ideal, is_range_check, adjust_check.
444 uncommon_trap(Deoptimization::Reason_range_check,
445 Deoptimization::Action_make_not_entrant,
446 nullptr, "range_check");
447 } else {
448 // If we have already recompiled with the range-check-widening
449 // heroic optimization turned off, then we must really be throwing
450 // range check exceptions.
451 builtin_throw(Deoptimization::Reason_range_check);
452 }
453 }
454 }
455
456 // For inline type arrays, we can use the profiling information for array accesses to speculate on the type, flatness,
457 // and null-freeness. We can either prepare the speculative type for later uses or emit explicit speculative checks with
458 // traps now. In the latter case, the speculative type guarantees can avoid additional runtime checks later (e.g.
459 // non-null-free implies non-flat which allows us to remove flatness checks). This makes the graph simpler.
460 Node* Parse::create_speculative_inline_type_array_checks(Node* array, const TypeAryPtr* array_type,
461 const Type*& element_type) {
462 if (!array_type->is_flat() && !array_type->is_not_flat()) {
463 // For arrays that might be flat, speculate that the array has the exact type reported in the profile data such that
464 // we can rely on a fixed memory layout (i.e. either a flat layout or not).
465 array = cast_to_speculative_array_type(array, array_type, element_type);
466 } else if (UseTypeSpeculation && UseArrayLoadStoreProfile) {
467 // Array is known to be either flat or not flat. If possible, update the speculative type by using the profile data
468 // at this bci.
469 array = cast_to_profiled_array_type(array);
470 }
471
472 // Even though the type does not tell us whether we have an inline type array or not, we can still check the profile data
473 // whether we have a non-null-free or non-flat array. Speculating on a non-null-free array doesn't help aaload but could
474 // be profitable for a subsequent aastore.
475 if (!array_type->is_null_free() && !array_type->is_not_null_free()) {
476 array = speculate_non_null_free_array(array, array_type);
477 }
478 if (!array_type->is_flat() && !array_type->is_not_flat()) {
479 array = speculate_non_flat_array(array, array_type);
480 }
481 return array;
482 }
483
484 // Speculate that the array has the exact type reported in the profile data. We emit a trap when this turns out to be
485 // wrong. On the fast path, we add a CheckCastPP to use the exact type.
486 Node* Parse::cast_to_speculative_array_type(Node* const array, const TypeAryPtr*& array_type, const Type*& element_type) {
487 Deoptimization::DeoptReason reason = Deoptimization::Reason_speculate_class_check;
488 ciKlass* speculative_array_type = array_type->speculative_type();
489 if (too_many_traps_or_recompiles(reason) || speculative_array_type == nullptr) {
490 // No speculative type, check profile data at this bci
491 speculative_array_type = nullptr;
492 reason = Deoptimization::Reason_class_check;
493 if (UseArrayLoadStoreProfile && !too_many_traps_or_recompiles(reason)) {
494 ciKlass* profiled_element_type = nullptr;
495 ProfilePtrKind element_ptr = ProfileMaybeNull;
496 bool flat_array = true;
497 bool null_free_array = true;
498 method()->array_access_profiled_type(bci(), speculative_array_type, profiled_element_type, element_ptr, flat_array,
499 null_free_array);
500 }
501 }
502 if (speculative_array_type != nullptr) {
503 // Speculate that this array has the exact type reported by profile data
504 Node* casted_array = nullptr;
505 DEBUG_ONLY(Node* old_control = control();)
506 Node* slow_ctl = type_check_receiver(array, speculative_array_type, 1.0, &casted_array);
507 if (stopped()) {
508 // The check always fails and therefore profile information is incorrect. Don't use it.
509 assert(old_control == slow_ctl, "type check should have been removed");
510 set_control(slow_ctl);
511 } else if (!slow_ctl->is_top()) {
512 { PreserveJVMState pjvms(this);
513 set_control(slow_ctl);
514 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
515 }
516 replace_in_map(array, casted_array);
517 array_type = _gvn.type(casted_array)->is_aryptr();
518 element_type = array_type->elem();
519 return casted_array;
520 }
521 }
522 return array;
523 }
524
525 // Create a CheckCastPP when the speculative type can improve the current type.
526 Node* Parse::cast_to_profiled_array_type(Node* const array) {
527 ciKlass* array_type = nullptr;
528 ciKlass* element_type = nullptr;
529 ProfilePtrKind element_ptr = ProfileMaybeNull;
530 bool flat_array = true;
531 bool null_free_array = true;
532 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
533 if (array_type != nullptr) {
534 return record_profile_for_speculation(array, array_type, ProfileMaybeNull);
535 }
536 return array;
537 }
538
539 // Speculate that the array is non-null-free. We emit a trap when this turns out to be
540 // wrong. On the fast path, we add a CheckCastPP to use the non-null-free type.
541 Node* Parse::speculate_non_null_free_array(Node* const array, const TypeAryPtr*& array_type) {
542 bool null_free_array = true;
543 Deoptimization::DeoptReason reason = Deoptimization::Reason_none;
544 if (array_type->speculative() != nullptr &&
545 array_type->speculative()->is_aryptr()->is_not_null_free() &&
546 !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_class_check)) {
547 null_free_array = false;
548 reason = Deoptimization::Reason_speculate_class_check;
549 } else if (UseArrayLoadStoreProfile && !too_many_traps_or_recompiles(Deoptimization::Reason_class_check)) {
550 ciKlass* profiled_array_type = nullptr;
551 ciKlass* profiled_element_type = nullptr;
552 ProfilePtrKind element_ptr = ProfileMaybeNull;
553 bool flat_array = true;
554 method()->array_access_profiled_type(bci(), profiled_array_type, profiled_element_type, element_ptr, flat_array,
555 null_free_array);
556 reason = Deoptimization::Reason_class_check;
557 }
558 if (!null_free_array) {
559 { // Deoptimize if null-free array
560 BuildCutout unless(this, null_free_array_test(array, /* null_free = */ false), PROB_MAX);
561 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
562 }
563 assert(!stopped(), "null-free array should have been caught earlier");
564 Node* casted_array = _gvn.transform(new CheckCastPPNode(control(), array, array_type->cast_to_not_null_free()));
565 replace_in_map(array, casted_array);
566 array_type = _gvn.type(casted_array)->is_aryptr();
567 return casted_array;
568 }
569 return array;
570 }
571
572 // Speculate that the array is non-flat. We emit a trap when this turns out to be wrong.
573 // On the fast path, we add a CheckCastPP to use the non-flat type.
574 Node* Parse::speculate_non_flat_array(Node* const array, const TypeAryPtr* const array_type) {
575 bool flat_array = true;
576 Deoptimization::DeoptReason reason = Deoptimization::Reason_none;
577 if (array_type->speculative() != nullptr &&
578 array_type->speculative()->is_aryptr()->is_not_flat() &&
579 !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_class_check)) {
580 flat_array = false;
581 reason = Deoptimization::Reason_speculate_class_check;
582 } else if (UseArrayLoadStoreProfile && !too_many_traps_or_recompiles(reason)) {
583 ciKlass* profiled_array_type = nullptr;
584 ciKlass* profiled_element_type = nullptr;
585 ProfilePtrKind element_ptr = ProfileMaybeNull;
586 bool null_free_array = true;
587 method()->array_access_profiled_type(bci(), profiled_array_type, profiled_element_type, element_ptr, flat_array,
588 null_free_array);
589 reason = Deoptimization::Reason_class_check;
590 }
591 if (!flat_array) {
592 { // Deoptimize if flat array
593 BuildCutout unless(this, flat_array_test(array, /* flat = */ false), PROB_MAX);
594 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
595 }
596 assert(!stopped(), "flat array should have been caught earlier");
597 Node* casted_array = _gvn.transform(new CheckCastPPNode(control(), array, array_type->cast_to_not_flat()));
598 replace_in_map(array, casted_array);
599 return casted_array;
600 }
601 return array;
602 }
603
604 // returns IfNode
605 IfNode* Parse::jump_if_fork_int(Node* a, Node* b, BoolTest::mask mask, float prob, float cnt) {
606 Node *cmp = _gvn.transform(new CmpINode(a, b)); // two cases: shiftcount > 32 and shiftcount <= 32
607 Node *tst = _gvn.transform(new BoolNode(cmp, mask));
608 IfNode *iff = create_and_map_if(control(), tst, prob, cnt);
609 return iff;
610 }
611
612
613 // sentinel value for the target bci to mark never taken branches
614 // (according to profiling)
615 static const int never_reached = INT_MAX;
616
617 //------------------------------helper for tableswitch-------------------------
618 void Parse::jump_if_true_fork(IfNode *iff, int dest_bci_if_true, bool unc) {
619 // True branch, use existing map info
620 { PreserveJVMState pjvms(this);
621 Node *iftrue = _gvn.transform( new IfTrueNode (iff) );
622 set_control( iftrue );
623 if (unc) {
624 repush_if_args();
625 uncommon_trap(Deoptimization::Reason_unstable_if,
626 Deoptimization::Action_reinterpret,
627 nullptr,
628 "taken always");
629 } else {
630 assert(dest_bci_if_true != never_reached, "inconsistent dest");
631 merge_new_path(dest_bci_if_true);
632 }
633 }
634
635 // False branch
636 Node *iffalse = _gvn.transform( new IfFalseNode(iff) );
637 set_control( iffalse );
638 }
639
640 void Parse::jump_if_false_fork(IfNode *iff, int dest_bci_if_true, bool unc) {
641 // True branch, use existing map info
642 { PreserveJVMState pjvms(this);
643 Node *iffalse = _gvn.transform( new IfFalseNode (iff) );
644 set_control( iffalse );
645 if (unc) {
646 repush_if_args();
647 uncommon_trap(Deoptimization::Reason_unstable_if,
648 Deoptimization::Action_reinterpret,
649 nullptr,
650 "taken never");
651 } else {
652 assert(dest_bci_if_true != never_reached, "inconsistent dest");
653 merge_new_path(dest_bci_if_true);
654 }
655 }
656
657 // False branch
658 Node *iftrue = _gvn.transform( new IfTrueNode(iff) );
659 set_control( iftrue );
660 }
661
662 void Parse::jump_if_always_fork(int dest_bci, bool unc) {
663 // False branch, use existing map and control()
664 if (unc) {
665 repush_if_args();
666 uncommon_trap(Deoptimization::Reason_unstable_if,
667 Deoptimization::Action_reinterpret,
668 nullptr,
669 "taken never");
670 } else {
671 assert(dest_bci != never_reached, "inconsistent dest");
672 merge_new_path(dest_bci);
673 }
674 }
675
676
677 extern "C" {
678 static int jint_cmp(const void *i, const void *j) {
679 int a = *(jint *)i;
680 int b = *(jint *)j;
681 return a > b ? 1 : a < b ? -1 : 0;
682 }
683 }
684
685
686 class SwitchRange : public StackObj {
687 // a range of integers coupled with a bci destination
688 jint _lo; // inclusive lower limit
689 jint _hi; // inclusive upper limit
690 int _dest;
691 float _cnt; // how many times this range was hit according to profiling
692
693 public:
694 jint lo() const { return _lo; }
695 jint hi() const { return _hi; }
696 int dest() const { return _dest; }
697 bool is_singleton() const { return _lo == _hi; }
698 float cnt() const { return _cnt; }
699
700 void setRange(jint lo, jint hi, int dest, float cnt) {
701 assert(lo <= hi, "must be a non-empty range");
702 _lo = lo, _hi = hi; _dest = dest; _cnt = cnt;
703 assert(_cnt >= 0, "");
704 }
705 bool adjoinRange(jint lo, jint hi, int dest, float cnt, bool trim_ranges) {
706 assert(lo <= hi, "must be a non-empty range");
707 if (lo == _hi+1) {
708 // see merge_ranges() comment below
709 if (trim_ranges) {
710 if (cnt == 0) {
711 if (_cnt != 0) {
712 return false;
713 }
714 if (dest != _dest) {
715 _dest = never_reached;
716 }
717 } else {
718 if (_cnt == 0) {
719 return false;
720 }
721 if (dest != _dest) {
722 return false;
723 }
724 }
725 } else {
726 if (dest != _dest) {
727 return false;
728 }
729 }
730 _hi = hi;
731 _cnt += cnt;
732 return true;
733 }
734 return false;
735 }
736
737 void set (jint value, int dest, float cnt) {
738 setRange(value, value, dest, cnt);
739 }
740 bool adjoin(jint value, int dest, float cnt, bool trim_ranges) {
741 return adjoinRange(value, value, dest, cnt, trim_ranges);
742 }
743 bool adjoin(SwitchRange& other) {
744 return adjoinRange(other._lo, other._hi, other._dest, other._cnt, false);
745 }
746
747 void print() {
748 if (is_singleton())
749 tty->print(" {%d}=>%d (cnt=%f)", lo(), dest(), cnt());
750 else if (lo() == min_jint)
751 tty->print(" {..%d}=>%d (cnt=%f)", hi(), dest(), cnt());
752 else if (hi() == max_jint)
753 tty->print(" {%d..}=>%d (cnt=%f)", lo(), dest(), cnt());
754 else
755 tty->print(" {%d..%d}=>%d (cnt=%f)", lo(), hi(), dest(), cnt());
756 }
757 };
758
759 // We try to minimize the number of ranges and the size of the taken
760 // ones using profiling data. When ranges are created,
761 // SwitchRange::adjoinRange() only allows 2 adjoining ranges to merge
762 // if both were never hit or both were hit to build longer unreached
763 // ranges. Here, we now merge adjoining ranges with the same
764 // destination and finally set destination of unreached ranges to the
765 // special value never_reached because it can help minimize the number
766 // of tests that are necessary.
767 //
768 // For instance:
769 // [0, 1] to target1 sometimes taken
770 // [1, 2] to target1 never taken
771 // [2, 3] to target2 never taken
772 // would lead to:
773 // [0, 1] to target1 sometimes taken
774 // [1, 3] never taken
775 //
776 // (first 2 ranges to target1 are not merged)
777 static void merge_ranges(SwitchRange* ranges, int& rp) {
778 if (rp == 0) {
779 return;
780 }
781 int shift = 0;
782 for (int j = 0; j < rp; j++) {
783 SwitchRange& r1 = ranges[j-shift];
784 SwitchRange& r2 = ranges[j+1];
785 if (r1.adjoin(r2)) {
786 shift++;
787 } else if (shift > 0) {
788 ranges[j+1-shift] = r2;
789 }
790 }
791 rp -= shift;
792 for (int j = 0; j <= rp; j++) {
793 SwitchRange& r = ranges[j];
794 if (r.cnt() == 0 && r.dest() != never_reached) {
795 r.setRange(r.lo(), r.hi(), never_reached, r.cnt());
796 }
797 }
798 }
799
800 //-------------------------------do_tableswitch--------------------------------
801 void Parse::do_tableswitch() {
802 // Get information about tableswitch
803 int default_dest = iter().get_dest_table(0);
804 jint lo_index = iter().get_int_table(1);
805 jint hi_index = iter().get_int_table(2);
806 int len = hi_index - lo_index + 1;
807
808 if (len < 1) {
809 // If this is a backward branch, add safepoint
810 maybe_add_safepoint(default_dest);
811 pop(); // the effect of the instruction execution on the operand stack
812 merge(default_dest);
813 return;
814 }
815
816 ciMethodData* methodData = method()->method_data();
817 ciMultiBranchData* profile = nullptr;
818 if (methodData->is_mature() && UseSwitchProfiling) {
819 ciProfileData* data = methodData->bci_to_data(bci());
820 if (data != nullptr && data->is_MultiBranchData()) {
821 profile = (ciMultiBranchData*)data;
822 }
823 }
824 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if);
825
826 // generate decision tree, using trichotomy when possible
827 int rnum = len+2;
828 bool makes_backward_branch = (default_dest <= bci());
829 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum);
830 int rp = -1;
831 if (lo_index != min_jint) {
832 float cnt = 1.0F;
833 if (profile != nullptr) {
834 cnt = (float)profile->default_count() / (hi_index != max_jint ? 2.0F : 1.0F);
835 }
836 ranges[++rp].setRange(min_jint, lo_index-1, default_dest, cnt);
837 }
838 for (int j = 0; j < len; j++) {
839 jint match_int = lo_index+j;
840 int dest = iter().get_dest_table(j+3);
841 makes_backward_branch |= (dest <= bci());
842 float cnt = 1.0F;
843 if (profile != nullptr) {
844 cnt = (float)profile->count_at(j);
845 }
846 if (rp < 0 || !ranges[rp].adjoin(match_int, dest, cnt, trim_ranges)) {
847 ranges[++rp].set(match_int, dest, cnt);
848 }
849 }
850 jint highest = lo_index+(len-1);
851 assert(ranges[rp].hi() == highest, "");
852 if (highest != max_jint) {
853 float cnt = 1.0F;
854 if (profile != nullptr) {
855 cnt = (float)profile->default_count() / (lo_index != min_jint ? 2.0F : 1.0F);
856 }
857 if (!ranges[rp].adjoinRange(highest+1, max_jint, default_dest, cnt, trim_ranges)) {
858 ranges[++rp].setRange(highest+1, max_jint, default_dest, cnt);
859 }
860 }
861 assert(rp < len+2, "not too many ranges");
862
863 if (trim_ranges) {
864 merge_ranges(ranges, rp);
865 }
866
867 // Safepoint in case if backward branch observed
868 if (makes_backward_branch) {
869 add_safepoint();
870 }
871
872 Node* lookup = pop(); // lookup value
873 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]);
874 }
875
876
877 //------------------------------do_lookupswitch--------------------------------
878 void Parse::do_lookupswitch() {
879 // Get information about lookupswitch
880 int default_dest = iter().get_dest_table(0);
881 jint len = iter().get_int_table(1);
882
883 if (len < 1) { // If this is a backward branch, add safepoint
884 maybe_add_safepoint(default_dest);
885 pop(); // the effect of the instruction execution on the operand stack
886 merge(default_dest);
887 return;
888 }
889
890 ciMethodData* methodData = method()->method_data();
891 ciMultiBranchData* profile = nullptr;
892 if (methodData->is_mature() && UseSwitchProfiling) {
893 ciProfileData* data = methodData->bci_to_data(bci());
894 if (data != nullptr && data->is_MultiBranchData()) {
895 profile = (ciMultiBranchData*)data;
896 }
897 }
898 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if);
899
900 // generate decision tree, using trichotomy when possible
901 jint* table = NEW_RESOURCE_ARRAY(jint, len*3);
902 {
903 for (int j = 0; j < len; j++) {
904 table[3*j+0] = iter().get_int_table(2+2*j);
905 table[3*j+1] = iter().get_dest_table(2+2*j+1);
906 // Handle overflow when converting from uint to jint
907 table[3*j+2] = (profile == nullptr) ? 1 : (jint)MIN2<uint>((uint)max_jint, profile->count_at(j));
908 }
909 qsort(table, len, 3*sizeof(table[0]), jint_cmp);
910 }
911
912 float default_cnt = 1.0F;
913 if (profile != nullptr) {
914 juint defaults = max_juint - len;
915 default_cnt = (float)profile->default_count()/(float)defaults;
916 }
917
918 int rnum = len*2+1;
919 bool makes_backward_branch = (default_dest <= bci());
920 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum);
921 int rp = -1;
922 for (int j = 0; j < len; j++) {
923 jint match_int = table[3*j+0];
924 jint dest = table[3*j+1];
925 jint cnt = table[3*j+2];
926 jint next_lo = rp < 0 ? min_jint : ranges[rp].hi()+1;
927 makes_backward_branch |= (dest <= bci());
928 float c = default_cnt * ((float)match_int - (float)next_lo);
929 if (match_int != next_lo && (rp < 0 || !ranges[rp].adjoinRange(next_lo, match_int-1, default_dest, c, trim_ranges))) {
930 assert(default_dest != never_reached, "sentinel value for dead destinations");
931 ranges[++rp].setRange(next_lo, match_int-1, default_dest, c);
932 }
933 if (rp < 0 || !ranges[rp].adjoin(match_int, dest, (float)cnt, trim_ranges)) {
934 assert(dest != never_reached, "sentinel value for dead destinations");
935 ranges[++rp].set(match_int, dest, (float)cnt);
936 }
937 }
938 jint highest = table[3*(len-1)];
939 assert(ranges[rp].hi() == highest, "");
940 if (highest != max_jint &&
941 !ranges[rp].adjoinRange(highest+1, max_jint, default_dest, default_cnt * ((float)max_jint - (float)highest), trim_ranges)) {
942 ranges[++rp].setRange(highest+1, max_jint, default_dest, default_cnt * ((float)max_jint - (float)highest));
943 }
944 assert(rp < rnum, "not too many ranges");
945
946 if (trim_ranges) {
947 merge_ranges(ranges, rp);
948 }
949
950 // Safepoint in case backward branch observed
951 if (makes_backward_branch) {
952 add_safepoint();
953 }
954
955 Node *lookup = pop(); // lookup value
956 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]);
957 }
958
959 static float if_prob(float taken_cnt, float total_cnt) {
960 assert(taken_cnt <= total_cnt, "");
961 if (total_cnt == 0) {
962 return PROB_FAIR;
963 }
964 float p = taken_cnt / total_cnt;
965 return clamp(p, PROB_MIN, PROB_MAX);
966 }
967
968 static float if_cnt(float cnt) {
969 if (cnt == 0) {
970 return COUNT_UNKNOWN;
971 }
972 return cnt;
973 }
974
975 static float sum_of_cnts(SwitchRange *lo, SwitchRange *hi) {
976 float total_cnt = 0;
977 for (SwitchRange* sr = lo; sr <= hi; sr++) {
978 total_cnt += sr->cnt();
979 }
980 return total_cnt;
981 }
982
983 class SwitchRanges : public ResourceObj {
984 public:
985 SwitchRange* _lo;
986 SwitchRange* _hi;
987 SwitchRange* _mid;
988 float _cost;
989
990 enum {
991 Start,
992 LeftDone,
993 RightDone,
994 Done
995 } _state;
996
997 SwitchRanges(SwitchRange *lo, SwitchRange *hi)
998 : _lo(lo), _hi(hi), _mid(nullptr),
999 _cost(0), _state(Start) {
1000 }
1001
1002 SwitchRanges()
1003 : _lo(nullptr), _hi(nullptr), _mid(nullptr),
1004 _cost(0), _state(Start) {}
1005 };
1006
1007 // Estimate cost of performing a binary search on lo..hi
1008 static float compute_tree_cost(SwitchRange *lo, SwitchRange *hi, float total_cnt) {
1009 GrowableArray<SwitchRanges> tree;
1010 SwitchRanges root(lo, hi);
1011 tree.push(root);
1012
1013 float cost = 0;
1014 do {
1015 SwitchRanges& r = *tree.adr_at(tree.length()-1);
1016 if (r._hi != r._lo) {
1017 if (r._mid == nullptr) {
1018 float r_cnt = sum_of_cnts(r._lo, r._hi);
1019
1020 if (r_cnt == 0) {
1021 tree.pop();
1022 cost = 0;
1023 continue;
1024 }
1025
1026 SwitchRange* mid = nullptr;
1027 mid = r._lo;
1028 for (float cnt = 0; ; ) {
1029 assert(mid <= r._hi, "out of bounds");
1030 cnt += mid->cnt();
1031 if (cnt > r_cnt / 2) {
1032 break;
1033 }
1034 mid++;
1035 }
1036 assert(mid <= r._hi, "out of bounds");
1037 r._mid = mid;
1038 r._cost = r_cnt / total_cnt;
1039 }
1040 r._cost += cost;
1041 if (r._state < SwitchRanges::LeftDone && r._mid > r._lo) {
1042 cost = 0;
1043 r._state = SwitchRanges::LeftDone;
1044 tree.push(SwitchRanges(r._lo, r._mid-1));
1045 } else if (r._state < SwitchRanges::RightDone) {
1046 cost = 0;
1047 r._state = SwitchRanges::RightDone;
1048 tree.push(SwitchRanges(r._mid == r._lo ? r._mid+1 : r._mid, r._hi));
1049 } else {
1050 tree.pop();
1051 cost = r._cost;
1052 }
1053 } else {
1054 tree.pop();
1055 cost = r._cost;
1056 }
1057 } while (tree.length() > 0);
1058
1059
1060 return cost;
1061 }
1062
1063 // It sometimes pays off to test most common ranges before the binary search
1064 void Parse::linear_search_switch_ranges(Node* key_val, SwitchRange*& lo, SwitchRange*& hi) {
1065 uint nr = hi - lo + 1;
1066 float total_cnt = sum_of_cnts(lo, hi);
1067
1068 float min = compute_tree_cost(lo, hi, total_cnt);
1069 float extra = 1;
1070 float sub = 0;
1071
1072 SwitchRange* array1 = lo;
1073 SwitchRange* array2 = NEW_RESOURCE_ARRAY(SwitchRange, nr);
1074
1075 SwitchRange* ranges = nullptr;
1076
1077 while (nr >= 2) {
1078 assert(lo == array1 || lo == array2, "one the 2 already allocated arrays");
1079 ranges = (lo == array1) ? array2 : array1;
1080
1081 // Find highest frequency range
1082 SwitchRange* candidate = lo;
1083 for (SwitchRange* sr = lo+1; sr <= hi; sr++) {
1084 if (sr->cnt() > candidate->cnt()) {
1085 candidate = sr;
1086 }
1087 }
1088 SwitchRange most_freq = *candidate;
1089 if (most_freq.cnt() == 0) {
1090 break;
1091 }
1092
1093 // Copy remaining ranges into another array
1094 int shift = 0;
1095 for (uint i = 0; i < nr; i++) {
1096 SwitchRange* sr = &lo[i];
1097 if (sr != candidate) {
1098 ranges[i-shift] = *sr;
1099 } else {
1100 shift++;
1101 if (i > 0 && i < nr-1) {
1102 SwitchRange prev = lo[i-1];
1103 prev.setRange(prev.lo(), sr->hi(), prev.dest(), prev.cnt());
1104 if (prev.adjoin(lo[i+1])) {
1105 shift++;
1106 i++;
1107 }
1108 ranges[i-shift] = prev;
1109 }
1110 }
1111 }
1112 nr -= shift;
1113
1114 // Evaluate cost of testing the most common range and performing a
1115 // binary search on the other ranges
1116 float cost = extra + compute_tree_cost(&ranges[0], &ranges[nr-1], total_cnt);
1117 if (cost >= min) {
1118 break;
1119 }
1120 // swap arrays
1121 lo = &ranges[0];
1122 hi = &ranges[nr-1];
1123
1124 // It pays off: emit the test for the most common range
1125 assert(most_freq.cnt() > 0, "must be taken");
1126 Node* val = _gvn.transform(new SubINode(key_val, _gvn.intcon(most_freq.lo())));
1127 Node* cmp = _gvn.transform(new CmpUNode(val, _gvn.intcon(java_subtract(most_freq.hi(), most_freq.lo()))));
1128 Node* tst = _gvn.transform(new BoolNode(cmp, BoolTest::le));
1129 IfNode* iff = create_and_map_if(control(), tst, if_prob(most_freq.cnt(), total_cnt), if_cnt(most_freq.cnt()));
1130 jump_if_true_fork(iff, most_freq.dest(), false);
1131
1132 sub += most_freq.cnt() / total_cnt;
1133 extra += 1 - sub;
1134 min = cost;
1135 }
1136 }
1137
1138 //----------------------------create_jump_tables-------------------------------
1139 bool Parse::create_jump_tables(Node* key_val, SwitchRange* lo, SwitchRange* hi) {
1140 // Are jumptables enabled
1141 if (!UseJumpTables) return false;
1142
1143 // Are jumptables supported
1144 if (!Matcher::has_match_rule(Op_Jump)) return false;
1145
1146 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if);
1147
1148 // Decide if a guard is needed to lop off big ranges at either (or
1149 // both) end(s) of the input set. We'll call this the default target
1150 // even though we can't be sure that it is the true "default".
1151
1152 bool needs_guard = false;
1153 int default_dest;
1154 int64_t total_outlier_size = 0;
1155 int64_t hi_size = ((int64_t)hi->hi()) - ((int64_t)hi->lo()) + 1;
1156 int64_t lo_size = ((int64_t)lo->hi()) - ((int64_t)lo->lo()) + 1;
1157
1158 if (lo->dest() == hi->dest()) {
1159 total_outlier_size = hi_size + lo_size;
1160 default_dest = lo->dest();
1161 } else if (lo_size > hi_size) {
1162 total_outlier_size = lo_size;
1163 default_dest = lo->dest();
1164 } else {
1165 total_outlier_size = hi_size;
1166 default_dest = hi->dest();
1167 }
1168
1169 float total = sum_of_cnts(lo, hi);
1170 float cost = compute_tree_cost(lo, hi, total);
1171
1172 // If a guard test will eliminate very sparse end ranges, then
1173 // it is worth the cost of an extra jump.
1174 float trimmed_cnt = 0;
1175 if (total_outlier_size > (MaxJumpTableSparseness * 4)) {
1176 needs_guard = true;
1177 if (default_dest == lo->dest()) {
1178 trimmed_cnt += lo->cnt();
1179 lo++;
1180 }
1181 if (default_dest == hi->dest()) {
1182 trimmed_cnt += hi->cnt();
1183 hi--;
1184 }
1185 }
1186
1187 // Find the total number of cases and ranges
1188 int64_t num_cases = ((int64_t)hi->hi()) - ((int64_t)lo->lo()) + 1;
1189 int num_range = hi - lo + 1;
1190
1191 // Don't create table if: too large, too small, or too sparse.
1192 if (num_cases > MaxJumpTableSize)
1193 return false;
1194 if (UseSwitchProfiling) {
1195 // MinJumpTableSize is set so with a well balanced binary tree,
1196 // when the number of ranges is MinJumpTableSize, it's cheaper to
1197 // go through a JumpNode that a tree of IfNodes. Average cost of a
1198 // tree of IfNodes with MinJumpTableSize is
1199 // log2f(MinJumpTableSize) comparisons. So if the cost computed
1200 // from profile data is less than log2f(MinJumpTableSize) then
1201 // going with the binary search is cheaper.
1202 if (cost < log2f(MinJumpTableSize)) {
1203 return false;
1204 }
1205 } else {
1206 if (num_cases < MinJumpTableSize)
1207 return false;
1208 }
1209 if (num_cases > (MaxJumpTableSparseness * num_range))
1210 return false;
1211
1212 // Normalize table lookups to zero
1213 int lowval = lo->lo();
1214 key_val = _gvn.transform( new SubINode(key_val, _gvn.intcon(lowval)) );
1215
1216 // Generate a guard to protect against input keyvals that aren't
1217 // in the switch domain.
1218 if (needs_guard) {
1219 Node* size = _gvn.intcon(num_cases);
1220 Node* cmp = _gvn.transform(new CmpUNode(key_val, size));
1221 Node* tst = _gvn.transform(new BoolNode(cmp, BoolTest::ge));
1222 IfNode* iff = create_and_map_if(control(), tst, if_prob(trimmed_cnt, total), if_cnt(trimmed_cnt));
1223 jump_if_true_fork(iff, default_dest, trim_ranges && trimmed_cnt == 0);
1224
1225 total -= trimmed_cnt;
1226 }
1227
1228 // Create an ideal node JumpTable that has projections
1229 // of all possible ranges for a switch statement
1230 // The key_val input must be converted to a pointer offset and scaled.
1231 // Compare Parse::array_addressing above.
1232
1233 // Clean the 32-bit int into a real 64-bit offset.
1234 // Otherwise, the jint value 0 might turn into an offset of 0x0800000000.
1235 // Make I2L conversion control dependent to prevent it from
1236 // floating above the range check during loop optimizations.
1237 // Do not use a narrow int type here to prevent the data path from dying
1238 // while the control path is not removed. This can happen if the type of key_val
1239 // is later known to be out of bounds of [0, num_cases] and therefore a narrow cast
1240 // would be replaced by TOP while C2 is not able to fold the corresponding range checks.
1241 // Set _carry_dependency for the cast to avoid being removed by IGVN.
1242 #ifdef _LP64
1243 key_val = C->constrained_convI2L(&_gvn, key_val, TypeInt::INT, control(), true /* carry_dependency */);
1244 #endif
1245
1246 // Shift the value by wordsize so we have an index into the table, rather
1247 // than a switch value
1248 Node *shiftWord = _gvn.MakeConX(wordSize);
1249 key_val = _gvn.transform( new MulXNode( key_val, shiftWord));
1250
1251 // Create the JumpNode
1252 Arena* arena = C->comp_arena();
1253 float* probs = (float*)arena->Amalloc(sizeof(float)*num_cases);
1254 int i = 0;
1255 if (total == 0) {
1256 for (SwitchRange* r = lo; r <= hi; r++) {
1257 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) {
1258 probs[i] = 1.0F / num_cases;
1259 }
1260 }
1261 } else {
1262 for (SwitchRange* r = lo; r <= hi; r++) {
1263 float prob = r->cnt()/total;
1264 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) {
1265 probs[i] = prob / (r->hi() - r->lo() + 1);
1266 }
1267 }
1268 }
1269
1270 ciMethodData* methodData = method()->method_data();
1271 ciMultiBranchData* profile = nullptr;
1272 if (methodData->is_mature()) {
1273 ciProfileData* data = methodData->bci_to_data(bci());
1274 if (data != nullptr && data->is_MultiBranchData()) {
1275 profile = (ciMultiBranchData*)data;
1276 }
1277 }
1278
1279 Node* jtn = _gvn.transform(new JumpNode(control(), key_val, num_cases, probs, profile == nullptr ? COUNT_UNKNOWN : total));
1280
1281 // These are the switch destinations hanging off the jumpnode
1282 i = 0;
1283 for (SwitchRange* r = lo; r <= hi; r++) {
1284 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) {
1285 Node* input = _gvn.transform(new JumpProjNode(jtn, i, r->dest(), (int)(j - lowval)));
1286 {
1287 PreserveJVMState pjvms(this);
1288 set_control(input);
1289 jump_if_always_fork(r->dest(), trim_ranges && r->cnt() == 0);
1290 }
1291 }
1292 }
1293 assert(i == num_cases, "miscount of cases");
1294 stop_and_kill_map(); // no more uses for this JVMS
1295 return true;
1296 }
1297
1298 //----------------------------jump_switch_ranges-------------------------------
1299 void Parse::jump_switch_ranges(Node* key_val, SwitchRange *lo, SwitchRange *hi, int switch_depth) {
1300 Block* switch_block = block();
1301 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if);
1302
1303 if (switch_depth == 0) {
1304 // Do special processing for the top-level call.
1305 assert(lo->lo() == min_jint, "initial range must exhaust Type::INT");
1306 assert(hi->hi() == max_jint, "initial range must exhaust Type::INT");
1307
1308 // Decrement pred-numbers for the unique set of nodes.
1309 #ifdef ASSERT
1310 if (!trim_ranges) {
1311 // Ensure that the block's successors are a (duplicate-free) set.
1312 int successors_counted = 0; // block occurrences in [hi..lo]
1313 int unique_successors = switch_block->num_successors();
1314 for (int i = 0; i < unique_successors; i++) {
1315 Block* target = switch_block->successor_at(i);
1316
1317 // Check that the set of successors is the same in both places.
1318 int successors_found = 0;
1319 for (SwitchRange* p = lo; p <= hi; p++) {
1320 if (p->dest() == target->start()) successors_found++;
1321 }
1322 assert(successors_found > 0, "successor must be known");
1323 successors_counted += successors_found;
1324 }
1325 assert(successors_counted == (hi-lo)+1, "no unexpected successors");
1326 }
1327 #endif
1328
1329 // Maybe prune the inputs, based on the type of key_val.
1330 jint min_val = min_jint;
1331 jint max_val = max_jint;
1332 const TypeInt* ti = key_val->bottom_type()->isa_int();
1333 if (ti != nullptr) {
1334 min_val = ti->_lo;
1335 max_val = ti->_hi;
1336 assert(min_val <= max_val, "invalid int type");
1337 }
1338 while (lo->hi() < min_val) {
1339 lo++;
1340 }
1341 if (lo->lo() < min_val) {
1342 lo->setRange(min_val, lo->hi(), lo->dest(), lo->cnt());
1343 }
1344 while (hi->lo() > max_val) {
1345 hi--;
1346 }
1347 if (hi->hi() > max_val) {
1348 hi->setRange(hi->lo(), max_val, hi->dest(), hi->cnt());
1349 }
1350
1351 linear_search_switch_ranges(key_val, lo, hi);
1352 }
1353
1354 #ifndef PRODUCT
1355 if (switch_depth == 0) {
1356 _max_switch_depth = 0;
1357 _est_switch_depth = log2i_graceful((hi - lo + 1) - 1) + 1;
1358 }
1359 SwitchRange* orig_lo = lo;
1360 SwitchRange* orig_hi = hi;
1361 #endif
1362
1363 // The lower-range processing is done iteratively to avoid O(N) stack depth
1364 // when the profiling-based pivot repeatedly selects mid==lo (JDK-8366138).
1365 // The upper-range processing remains recursive but is only reached for
1366 // balanced splits, bounding its depth to O(log N).
1367 // Termination: every iteration either exits or strictly decreases hi-lo:
1368 // lo == mid && mid < hi, increments lo
1369 // lo < mid <= hi, sets hi = mid - 1.
1370 for (int depth = switch_depth;; depth++) {
1371 #ifndef PRODUCT
1372 _max_switch_depth = MAX2(depth, _max_switch_depth);
1373 #endif
1374
1375 assert(lo <= hi, "must be a non-empty set of ranges");
1376 if (lo == hi) {
1377 jump_if_always_fork(lo->dest(), trim_ranges && lo->cnt() == 0);
1378 break;
1379 }
1380
1381 assert(lo->hi() == (lo+1)->lo()-1, "contiguous ranges");
1382 assert(hi->lo() == (hi-1)->hi()+1, "contiguous ranges");
1383
1384 if (create_jump_tables(key_val, lo, hi)) return;
1385
1386 SwitchRange* mid = nullptr;
1387 float total_cnt = sum_of_cnts(lo, hi);
1388
1389 int nr = hi - lo + 1;
1390 // With total_cnt==0 the profiling pivot degenerates to mid==lo
1391 // (0 >= 0/2), producing a linear chain of If nodes instead of a
1392 // balanced tree. A balanced tree is strictly better here: all paths
1393 // are cold, so a balanced split gives fewer comparisons at runtime
1394 // and avoids pathological memory usage in the optimizer.
1395 if (UseSwitchProfiling && total_cnt > 0) {
1396 // Don't keep the binary search tree balanced: pick up mid point
1397 // that split frequencies in half.
1398 float cnt = 0;
1399 for (SwitchRange* sr = lo; sr <= hi; sr++) {
1400 cnt += sr->cnt();
1401 if (cnt >= total_cnt / 2) {
1402 mid = sr;
1403 break;
1404 }
1405 }
1406 } else {
1407 mid = lo + nr/2;
1408
1409 // if there is an easy choice, pivot at a singleton:
1410 if (nr > 3 && !mid->is_singleton() && (mid-1)->is_singleton()) mid--;
1411
1412 assert(lo < mid && mid <= hi, "good pivot choice");
1413 assert(nr != 2 || mid == hi, "should pick higher of 2");
1414 assert(nr != 3 || mid == hi-1, "should pick middle of 3");
1415 }
1416 assert(mid != nullptr, "mid must be set");
1417
1418 Node *test_val = _gvn.intcon(mid == lo ? mid->hi() : mid->lo());
1419
1420 if (mid->is_singleton()) {
1421 IfNode *iff_ne = jump_if_fork_int(key_val, test_val, BoolTest::ne, 1-if_prob(mid->cnt(), total_cnt), if_cnt(mid->cnt()));
1422 jump_if_false_fork(iff_ne, mid->dest(), trim_ranges && mid->cnt() == 0);
1423
1424 // Special Case: If there are exactly three ranges, and the high
1425 // and low range each go to the same place, omit the "gt" test,
1426 // since it will not discriminate anything.
1427 bool eq_test_only = (hi == lo+2 && hi->dest() == lo->dest() && mid == hi-1) || mid == lo;
1428
1429 // if there is a higher range, test for it and process it:
1430 if (mid < hi && !eq_test_only) {
1431 // two comparisons of same values--should enable 1 test for 2 branches
1432 // Use BoolTest::lt instead of BoolTest::gt
1433 float cnt = sum_of_cnts(lo, mid-1);
1434 IfNode *iff_lt = jump_if_fork_int(key_val, test_val, BoolTest::lt, if_prob(cnt, total_cnt), if_cnt(cnt));
1435 Node *iftrue = _gvn.transform( new IfTrueNode(iff_lt) );
1436 Node *iffalse = _gvn.transform( new IfFalseNode(iff_lt) );
1437 { PreserveJVMState pjvms(this);
1438 set_control(iffalse);
1439 jump_switch_ranges(key_val, mid+1, hi, depth+1);
1440 }
1441 set_control(iftrue);
1442 }
1443
1444 } else {
1445 // mid is a range, not a singleton, so treat mid..hi as a unit
1446 float cnt = sum_of_cnts(mid == lo ? mid+1 : mid, hi);
1447 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));
1448
1449 // if there is a higher range, test for it and process it:
1450 if (mid == hi) {
1451 jump_if_true_fork(iff_ge, mid->dest(), trim_ranges && cnt == 0);
1452 } else {
1453 Node *iftrue = _gvn.transform( new IfTrueNode(iff_ge) );
1454 Node *iffalse = _gvn.transform( new IfFalseNode(iff_ge) );
1455 { PreserveJVMState pjvms(this);
1456 set_control(iftrue);
1457 jump_switch_ranges(key_val, mid == lo ? mid+1 : mid, hi, depth+1);
1458 }
1459 set_control(iffalse);
1460 }
1461 }
1462
1463 // Process the lower range: iterate instead of recursing.
1464 if (mid == lo) {
1465 if (mid->is_singleton()) {
1466 lo++;
1467 } else {
1468 jump_if_always_fork(lo->dest(), trim_ranges && lo->cnt() == 0);
1469 break;
1470 }
1471 } else {
1472 hi = mid - 1;
1473 }
1474 }
1475
1476 // Decrease pred_count for each successor after all is done.
1477 if (switch_depth == 0) {
1478 int unique_successors = switch_block->num_successors();
1479 for (int i = 0; i < unique_successors; i++) {
1480 Block* target = switch_block->successor_at(i);
1481 // Throw away the pre-allocated path for each unique successor.
1482 target->next_path_num();
1483 }
1484 }
1485
1486 #ifndef PRODUCT
1487 if (TraceOptoParse && Verbose && WizardMode && switch_depth == 0) {
1488 SwitchRange* r;
1489 int nsing = 0;
1490 for (r = orig_lo; r <= orig_hi; r++) {
1491 if( r->is_singleton() ) nsing++;
1492 }
1493 tty->print(">>> ");
1494 _method->print_short_name();
1495 tty->print_cr(" switch decision tree");
1496 tty->print_cr(" %d ranges (%d singletons), max_depth=%d, est_depth=%d",
1497 (int) (orig_hi-orig_lo+1), nsing, _max_switch_depth, _est_switch_depth);
1498 if (_max_switch_depth > _est_switch_depth) {
1499 tty->print_cr("******** BAD SWITCH DEPTH ********");
1500 }
1501 tty->print(" ");
1502 for (r = orig_lo; r <= orig_hi; r++) {
1503 r->print();
1504 }
1505 tty->cr();
1506 }
1507 #endif
1508 }
1509
1510 Node* Parse::floating_point_mod(Node* a, Node* b, BasicType type) {
1511 assert(type == BasicType::T_FLOAT || type == BasicType::T_DOUBLE, "only float and double are floating points");
1512 CallLeafPureNode* mod = type == BasicType::T_DOUBLE ? static_cast<CallLeafPureNode*>(new ModDNode(C, a, b)) : new ModFNode(C, a, b);
1513
1514 set_predefined_input_for_runtime_call(mod);
1515 mod = _gvn.transform(mod)->as_CallLeafPure();
1516 set_predefined_output_for_runtime_call(mod);
1517 Node* result = _gvn.transform(new ProjNode(mod, TypeFunc::Parms + 0));
1518 record_for_igvn(mod);
1519 return result;
1520 }
1521
1522 void Parse::l2f() {
1523 Node* f2 = pop();
1524 Node* f1 = pop();
1525 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::l2f_Type(),
1526 CAST_FROM_FN_PTR(address, SharedRuntime::l2f),
1527 "l2f", nullptr, //no memory effects
1528 f1, f2);
1529 Node* res = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0));
1530
1531 push(res);
1532 }
1533
1534 // Handle jsr and jsr_w bytecode
1535 void Parse::do_jsr() {
1536 assert(bc() == Bytecodes::_jsr || bc() == Bytecodes::_jsr_w, "wrong bytecode");
1537
1538 // Store information about current state, tagged with new _jsr_bci
1539 int return_bci = iter().next_bci();
1540 int jsr_bci = (bc() == Bytecodes::_jsr) ? iter().get_dest() : iter().get_far_dest();
1541
1542 // The way we do things now, there is only one successor block
1543 // for the jsr, because the target code is cloned by ciTypeFlow.
1544 Block* target = successor_for_bci(jsr_bci);
1545
1546 // What got pushed?
1547 const Type* ret_addr = target->peek();
1548 assert(ret_addr->singleton(), "must be a constant (cloned jsr body)");
1549
1550 // Effect on jsr on stack
1551 push(_gvn.makecon(ret_addr));
1552
1553 // Flow to the jsr.
1554 merge(jsr_bci);
1555 }
1556
1557 // Handle ret bytecode
1558 void Parse::do_ret() {
1559 // Find to whom we return.
1560 assert(block()->num_successors() == 1, "a ret can only go one place now");
1561 Block* target = block()->successor_at(0);
1562 assert(!target->is_ready(), "our arrival must be expected");
1563 int pnum = target->next_path_num();
1564 merge_common(target, pnum);
1565 }
1566
1567 static bool has_injected_profile(BoolTest::mask btest, Node* test, int& taken, int& not_taken) {
1568 if (btest != BoolTest::eq && btest != BoolTest::ne) {
1569 // Only ::eq and ::ne are supported for profile injection.
1570 return false;
1571 }
1572 if (test->is_Cmp() &&
1573 test->in(1)->Opcode() == Op_ProfileBoolean) {
1574 ProfileBooleanNode* profile = (ProfileBooleanNode*)test->in(1);
1575 int false_cnt = profile->false_count();
1576 int true_cnt = profile->true_count();
1577
1578 // Counts matching depends on the actual test operation (::eq or ::ne).
1579 // No need to scale the counts because profile injection was designed
1580 // to feed exact counts into VM.
1581 taken = (btest == BoolTest::eq) ? false_cnt : true_cnt;
1582 not_taken = (btest == BoolTest::eq) ? true_cnt : false_cnt;
1583
1584 profile->consume();
1585 return true;
1586 }
1587 return false;
1588 }
1589
1590 // Give up if too few (or too many, in which case the sum will overflow) counts to be meaningful.
1591 // We also check that individual counters are positive first, otherwise the sum can become positive.
1592 // (check for saturation, integer overflow, and immature counts)
1593 static bool counters_are_meaningful(int counter1, int counter2, int min) {
1594 // check for saturation, including "uint" values too big to fit in "int"
1595 if (counter1 < 0 || counter2 < 0) {
1596 return false;
1597 }
1598 // check for integer overflow of the sum
1599 int64_t sum = (int64_t)counter1 + (int64_t)counter2;
1600 STATIC_ASSERT(sizeof(counter1) < sizeof(sum));
1601 if (sum > INT_MAX) {
1602 return false;
1603 }
1604 // check if mature
1605 return (counter1 + counter2) >= min;
1606 }
1607
1608 //--------------------------dynamic_branch_prediction--------------------------
1609 // Try to gather dynamic branch prediction behavior. Return a probability
1610 // of the branch being taken and set the "cnt" field. Returns a -1.0
1611 // if we need to use static prediction for some reason.
1612 float Parse::dynamic_branch_prediction(float &cnt, BoolTest::mask btest, Node* test) {
1613 ResourceMark rm;
1614
1615 cnt = COUNT_UNKNOWN;
1616
1617 int taken = 0;
1618 int not_taken = 0;
1619
1620 bool use_mdo = !has_injected_profile(btest, test, taken, not_taken);
1621
1622 if (use_mdo) {
1623 // Use MethodData information if it is available
1624 // FIXME: free the ProfileData structure
1625 ciMethodData* methodData = method()->method_data();
1626 if (!methodData->is_mature()) return PROB_UNKNOWN;
1627 ciProfileData* data = methodData->bci_to_data(bci());
1628 if (data == nullptr) {
1629 return PROB_UNKNOWN;
1630 }
1631 if (!data->is_JumpData()) return PROB_UNKNOWN;
1632
1633 // get taken and not taken values
1634 // NOTE: saturated UINT_MAX values become negative,
1635 // as do counts above INT_MAX.
1636 taken = data->as_JumpData()->taken();
1637 not_taken = 0;
1638 if (data->is_BranchData()) {
1639 not_taken = data->as_BranchData()->not_taken();
1640 }
1641
1642 // scale the counts to be commensurate with invocation counts:
1643 // NOTE: overflow for positive values is clamped at INT_MAX
1644 taken = method()->scale_count(taken);
1645 not_taken = method()->scale_count(not_taken);
1646 }
1647 // At this point, saturation or overflow is indicated by INT_MAX
1648 // or a negative value.
1649
1650 // Give up if too few (or too many, in which case the sum will overflow) counts to be meaningful.
1651 // We also check that individual counters are positive first, otherwise the sum can become positive.
1652 if (!counters_are_meaningful(taken, not_taken, 40)) {
1653 if (C->log() != nullptr) {
1654 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d'", iter().get_dest(), taken, not_taken);
1655 }
1656 return PROB_UNKNOWN;
1657 }
1658
1659 // Compute frequency that we arrive here
1660 float sum = taken + not_taken;
1661 // Adjust, if this block is a cloned private block but the
1662 // Jump counts are shared. Taken the private counts for
1663 // just this path instead of the shared counts.
1664 if( block()->count() > 0 )
1665 sum = block()->count();
1666 cnt = sum / FreqCountInvocations;
1667
1668 // Pin probability to sane limits
1669 float prob;
1670 if( !taken )
1671 prob = (0+PROB_MIN) / 2;
1672 else if( !not_taken )
1673 prob = (1+PROB_MAX) / 2;
1674 else { // Compute probability of true path
1675 prob = (float)taken / (float)(taken + not_taken);
1676 if (prob > PROB_MAX) prob = PROB_MAX;
1677 if (prob < PROB_MIN) prob = PROB_MIN;
1678 }
1679
1680 assert((cnt > 0.0f) && (prob > 0.0f),
1681 "Bad frequency assignment in if cnt=%g prob=%g taken=%d not_taken=%d", cnt, prob, taken, not_taken);
1682
1683 if (C->log() != nullptr) {
1684 const char* prob_str = nullptr;
1685 if (prob >= PROB_MAX) prob_str = (prob == PROB_MAX) ? "max" : "always";
1686 if (prob <= PROB_MIN) prob_str = (prob == PROB_MIN) ? "min" : "never";
1687 char prob_str_buf[30];
1688 if (prob_str == nullptr) {
1689 jio_snprintf(prob_str_buf, sizeof(prob_str_buf), "%20.2f", prob);
1690 prob_str = prob_str_buf;
1691 }
1692 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d' cnt='%f' prob='%s'",
1693 iter().get_dest(), taken, not_taken, cnt, prob_str);
1694 }
1695 return prob;
1696 }
1697
1698 //-----------------------------branch_prediction-------------------------------
1699 float Parse::branch_prediction(float& cnt,
1700 BoolTest::mask btest,
1701 int target_bci,
1702 Node* test) {
1703 float prob = dynamic_branch_prediction(cnt, btest, test);
1704 // If prob is unknown, switch to static prediction
1705 if (prob != PROB_UNKNOWN) return prob;
1706
1707 prob = PROB_FAIR; // Set default value
1708 if (btest == BoolTest::eq) // Exactly equal test?
1709 prob = PROB_STATIC_INFREQUENT; // Assume its relatively infrequent
1710 else if (btest == BoolTest::ne)
1711 prob = PROB_STATIC_FREQUENT; // Assume its relatively frequent
1712
1713 // If this is a conditional test guarding a backwards branch,
1714 // assume its a loop-back edge. Make it a likely taken branch.
1715 if (target_bci < bci()) {
1716 if (is_osr_parse()) { // Could be a hot OSR'd loop; force deopt
1717 // Since it's an OSR, we probably have profile data, but since
1718 // branch_prediction returned PROB_UNKNOWN, the counts are too small.
1719 // Let's make a special check here for completely zero counts.
1720 ciMethodData* methodData = method()->method_data();
1721 if (!methodData->is_empty()) {
1722 ciProfileData* data = methodData->bci_to_data(bci());
1723 // Only stop for truly zero counts, which mean an unknown part
1724 // of the OSR-ed method, and we want to deopt to gather more stats.
1725 // If you have ANY counts, then this loop is simply 'cold' relative
1726 // to the OSR loop.
1727 if (data == nullptr ||
1728 (data->as_BranchData()->taken() + data->as_BranchData()->not_taken() == 0)) {
1729 // This is the only way to return PROB_UNKNOWN:
1730 return PROB_UNKNOWN;
1731 }
1732 }
1733 }
1734 prob = PROB_STATIC_FREQUENT; // Likely to take backwards branch
1735 }
1736
1737 assert(prob != PROB_UNKNOWN, "must have some guess at this point");
1738 return prob;
1739 }
1740
1741 // The magic constants are chosen so as to match the output of
1742 // branch_prediction() when the profile reports a zero taken count.
1743 // It is important to distinguish zero counts unambiguously, because
1744 // some branches (e.g., _213_javac.Assembler.eliminate) validly produce
1745 // very small but nonzero probabilities, which if confused with zero
1746 // counts would keep the program recompiling indefinitely.
1747 bool Parse::seems_never_taken(float prob) const {
1748 return prob < PROB_MIN;
1749 }
1750
1751 //-------------------------------repush_if_args--------------------------------
1752 // Push arguments of an "if" bytecode back onto the stack by adjusting _sp.
1753 inline int Parse::repush_if_args() {
1754 if (PrintOpto && WizardMode) {
1755 tty->print("defending against excessive implicit null exceptions on %s @%d in ",
1756 Bytecodes::name(iter().cur_bc()), iter().cur_bci());
1757 method()->print_name(); tty->cr();
1758 }
1759 int bc_depth = - Bytecodes::depth(iter().cur_bc());
1760 assert(bc_depth == 1 || bc_depth == 2, "only two kinds of branches");
1761 DEBUG_ONLY(sync_jvms()); // argument(n) requires a synced jvms
1762 assert(argument(0) != nullptr, "must exist");
1763 assert(bc_depth == 1 || argument(1) != nullptr, "two must exist");
1764 inc_sp(bc_depth);
1765 return bc_depth;
1766 }
1767
1768 // Used by StressUnstableIfTraps
1769 static volatile int _trap_stress_counter = 0;
1770
1771 void Parse::increment_trap_stress_counter(Node*& counter, Node*& incr_store) {
1772 Node* counter_addr = makecon(TypeRawPtr::make((address)&_trap_stress_counter));
1773 counter = make_load(control(), counter_addr, TypeInt::INT, T_INT, MemNode::unordered);
1774 counter = _gvn.transform(new AddINode(counter, intcon(1)));
1775 incr_store = store_to_memory(control(), counter_addr, counter, T_INT, MemNode::unordered);
1776 }
1777
1778 //----------------------------------do_ifnull----------------------------------
1779 void Parse::do_ifnull(BoolTest::mask btest, Node *c) {
1780 int target_bci = iter().get_dest();
1781
1782 Node* counter = nullptr;
1783 Node* incr_store = nullptr;
1784 bool do_stress_trap = StressUnstableIfTraps && ((C->random() % 2) == 0);
1785 if (do_stress_trap) {
1786 increment_trap_stress_counter(counter, incr_store);
1787 }
1788
1789 Block* branch_block = successor_for_bci(target_bci);
1790 Block* next_block = successor_for_bci(iter().next_bci());
1791
1792 float cnt;
1793 float prob = branch_prediction(cnt, btest, target_bci, c);
1794 if (prob == PROB_UNKNOWN) {
1795 // (An earlier version of do_ifnull omitted this trap for OSR methods.)
1796 if (PrintOpto && Verbose) {
1797 tty->print_cr("Never-taken edge stops compilation at bci %d", bci());
1798 }
1799 repush_if_args(); // to gather stats on loop
1800 uncommon_trap(Deoptimization::Reason_unreached,
1801 Deoptimization::Action_reinterpret,
1802 nullptr, "cold");
1803 if (C->eliminate_boxing()) {
1804 // Mark the successor blocks as parsed
1805 branch_block->next_path_num();
1806 next_block->next_path_num();
1807 }
1808 return;
1809 }
1810
1811 NOT_PRODUCT(explicit_null_checks_inserted++);
1812
1813 // Generate real control flow
1814 Node *tst = _gvn.transform( new BoolNode( c, btest ) );
1815
1816 // Sanity check the probability value
1817 assert(prob > 0.0f,"Bad probability in Parser");
1818 // Need xform to put node in hash table
1819 IfNode *iff = create_and_xform_if( control(), tst, prob, cnt );
1820 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser");
1821 // True branch
1822 { PreserveJVMState pjvms(this);
1823 Node* iftrue = _gvn.transform( new IfTrueNode (iff) );
1824 set_control(iftrue);
1825
1826 if (stopped()) { // Path is dead?
1827 NOT_PRODUCT(explicit_null_checks_elided++);
1828 if (C->eliminate_boxing()) {
1829 // Mark the successor block as parsed
1830 branch_block->next_path_num();
1831 }
1832 } else { // Path is live.
1833 adjust_map_after_if(btest, c, prob, branch_block);
1834 if (!stopped()) {
1835 merge(target_bci);
1836 }
1837 }
1838 }
1839
1840 // False branch
1841 Node* iffalse = _gvn.transform( new IfFalseNode(iff) );
1842 set_control(iffalse);
1843
1844 if (stopped()) { // Path is dead?
1845 NOT_PRODUCT(explicit_null_checks_elided++);
1846 if (C->eliminate_boxing()) {
1847 // Mark the successor block as parsed
1848 next_block->next_path_num();
1849 }
1850 } else { // Path is live.
1851 adjust_map_after_if(BoolTest(btest).negate(), c, 1.0-prob, next_block);
1852 }
1853
1854 if (do_stress_trap) {
1855 stress_trap(iff, counter, incr_store);
1856 }
1857 }
1858
1859 //------------------------------------do_if------------------------------------
1860 void Parse::do_if(BoolTest::mask btest, Node* c, bool can_trap, bool new_path, Node** ctrl_taken, Node** mem_taken, Node** io_taken) {
1861 int target_bci = iter().get_dest();
1862
1863 Block* branch_block = successor_for_bci(target_bci);
1864 Block* next_block = successor_for_bci(iter().next_bci());
1865
1866 float cnt;
1867 float prob = branch_prediction(cnt, btest, target_bci, c);
1868 float untaken_prob = 1.0 - prob;
1869
1870 if (prob == PROB_UNKNOWN) {
1871 if (PrintOpto && Verbose) {
1872 tty->print_cr("Never-taken edge stops compilation at bci %d", bci());
1873 }
1874 repush_if_args(); // to gather stats on loop
1875 uncommon_trap(Deoptimization::Reason_unreached,
1876 Deoptimization::Action_reinterpret,
1877 nullptr, "cold");
1878 if (C->eliminate_boxing()) {
1879 // Mark the successor blocks as parsed
1880 branch_block->next_path_num();
1881 next_block->next_path_num();
1882 }
1883 return;
1884 }
1885
1886 Node* counter = nullptr;
1887 Node* incr_store = nullptr;
1888 bool do_stress_trap = StressUnstableIfTraps && ((C->random() % 2) == 0);
1889 if (do_stress_trap) {
1890 increment_trap_stress_counter(counter, incr_store);
1891 }
1892
1893 // Sanity check the probability value
1894 assert(0.0f < prob && prob < 1.0f,"Bad probability in Parser");
1895
1896 bool taken_if_true = true;
1897 // Convert BoolTest to canonical form:
1898 if (!BoolTest(btest).is_canonical()) {
1899 btest = BoolTest(btest).negate();
1900 taken_if_true = false;
1901 // prob is NOT updated here; it remains the probability of the taken
1902 // path (as opposed to the prob of the path guarded by an 'IfTrueNode').
1903 }
1904 assert(btest != BoolTest::eq, "!= is the only canonical exact test");
1905
1906 Node* tst0 = new BoolNode(c, btest);
1907 Node* tst = _gvn.transform(tst0);
1908 BoolTest::mask taken_btest = BoolTest::illegal;
1909 BoolTest::mask untaken_btest = BoolTest::illegal;
1910
1911 if (tst->is_Bool()) {
1912 // Refresh c from the transformed bool node, since it may be
1913 // simpler than the original c. Also re-canonicalize btest.
1914 // This wins when (Bool ne (Conv2B p) 0) => (Bool ne (CmpP p null)).
1915 // That can arise from statements like: if (x instanceof C) ...
1916 if (tst != tst0) {
1917 // Canonicalize one more time since transform can change it.
1918 btest = tst->as_Bool()->_test._test;
1919 if (!BoolTest(btest).is_canonical()) {
1920 // Reverse edges one more time...
1921 tst = _gvn.transform( tst->as_Bool()->negate(&_gvn) );
1922 btest = tst->as_Bool()->_test._test;
1923 assert(BoolTest(btest).is_canonical(), "sanity");
1924 taken_if_true = !taken_if_true;
1925 }
1926 c = tst->in(1);
1927 }
1928 BoolTest::mask neg_btest = BoolTest(btest).negate();
1929 taken_btest = taken_if_true ? btest : neg_btest;
1930 untaken_btest = taken_if_true ? neg_btest : btest;
1931 }
1932
1933 // Generate real control flow
1934 float true_prob = (taken_if_true ? prob : untaken_prob);
1935 IfNode* iff = create_and_map_if(control(), tst, true_prob, cnt);
1936 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser");
1937 Node* taken_branch = new IfTrueNode(iff);
1938 Node* untaken_branch = new IfFalseNode(iff);
1939 if (!taken_if_true) { // Finish conversion to canonical form
1940 Node* tmp = taken_branch;
1941 taken_branch = untaken_branch;
1942 untaken_branch = tmp;
1943 }
1944
1945 // Branch is taken:
1946 { PreserveJVMState pjvms(this);
1947 taken_branch = _gvn.transform(taken_branch);
1948 set_control(taken_branch);
1949
1950 if (stopped()) {
1951 if (C->eliminate_boxing() && !new_path) {
1952 // Mark the successor block as parsed (if we haven't created a new path)
1953 branch_block->next_path_num();
1954 }
1955 } else {
1956 adjust_map_after_if(taken_btest, c, prob, branch_block, can_trap);
1957 if (!stopped()) {
1958 if (new_path) {
1959 // Merge by using a new path
1960 merge_new_path(target_bci);
1961 } else if (ctrl_taken != nullptr) {
1962 // Don't merge but save taken branch to be wired by caller
1963 *ctrl_taken = control();
1964 if (mem_taken != nullptr) {
1965 *mem_taken = reset_memory();
1966 }
1967 if (io_taken != nullptr) {
1968 *io_taken = i_o();
1969 }
1970 } else {
1971 merge(target_bci);
1972 }
1973 }
1974 }
1975 }
1976
1977 untaken_branch = _gvn.transform(untaken_branch);
1978 set_control(untaken_branch);
1979
1980 // Branch not taken.
1981 if (stopped() && ctrl_taken == nullptr) {
1982 if (C->eliminate_boxing()) {
1983 // Mark the successor block as parsed (if caller does not re-wire control flow)
1984 next_block->next_path_num();
1985 }
1986 } else {
1987 adjust_map_after_if(untaken_btest, c, untaken_prob, next_block, can_trap);
1988 }
1989
1990 if (do_stress_trap) {
1991 stress_trap(iff, counter, incr_store);
1992 }
1993 }
1994
1995
1996 static ProfilePtrKind speculative_ptr_kind(const TypeOopPtr* t) {
1997 if (t->speculative() == nullptr) {
1998 return ProfileUnknownNull;
1999 }
2000 if (t->speculative_always_null()) {
2001 return ProfileAlwaysNull;
2002 }
2003 if (t->speculative_maybe_null()) {
2004 return ProfileMaybeNull;
2005 }
2006 return ProfileNeverNull;
2007 }
2008
2009 void Parse::acmp_always_null_input(Node* input, const TypeOopPtr* tinput, BoolTest::mask btest, Node* eq_region) {
2010 if (btest == BoolTest::ne) {
2011 {
2012 PreserveJVMState pjvms(this);
2013 inc_sp(2);
2014 null_check_common(input, T_OBJECT, true, nullptr,
2015 !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_null_check) &&
2016 speculative_ptr_kind(tinput) == ProfileAlwaysNull);
2017 dec_sp(2);
2018 int target_bci = iter().get_dest();
2019 merge(target_bci);
2020 }
2021 record_for_igvn(eq_region);
2022 set_control(_gvn.transform(eq_region));
2023 } else {
2024 inc_sp(2);
2025 null_check_common(input, T_OBJECT, true, nullptr,
2026 !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_null_check) &&
2027 speculative_ptr_kind(tinput) == ProfileAlwaysNull);
2028 dec_sp(2);
2029 }
2030 }
2031
2032 Node* Parse::acmp_null_check(Node* input, const TypeOopPtr* tinput, ProfilePtrKind input_ptr, Node*& null_ctl) {
2033 inc_sp(2);
2034 null_ctl = top();
2035 Node* cast = null_check_oop(input, &null_ctl,
2036 input_ptr == ProfileNeverNull || (input_ptr == ProfileUnknownNull && !too_many_traps_or_recompiles(Deoptimization::Reason_null_check)),
2037 false,
2038 speculative_ptr_kind(tinput) == ProfileNeverNull &&
2039 !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_null_check));
2040 dec_sp(2);
2041 return cast;
2042 }
2043
2044 void Parse::acmp_type_check_or_trap(Node** non_null_input, ciKlass* input_type, Deoptimization::DeoptReason reason) {
2045 Node* slow_ctl = type_check_receiver(*non_null_input, input_type, 1.0, non_null_input);
2046 {
2047 PreserveJVMState pjvms(this);
2048 inc_sp(2);
2049 set_control(slow_ctl);
2050 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
2051 }
2052 }
2053
2054 void Parse::acmp_type_check(Node* input, const TypeOopPtr* tinput, ProfilePtrKind input_ptr, ciKlass* input_type, BoolTest::mask btest, Node* eq_region) {
2055 Node* null_ctl;
2056 Node* cast = acmp_null_check(input, tinput, input_ptr, null_ctl);
2057
2058 if (input_type != nullptr) {
2059 Deoptimization::DeoptReason reason;
2060 if (tinput->speculative_type() != nullptr && !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_class_check)) {
2061 reason = Deoptimization::Reason_speculate_class_check;
2062 } else {
2063 reason = Deoptimization::Reason_class_check;
2064 }
2065 acmp_type_check_or_trap(&cast, input_type, reason);
2066 } else {
2067 // No specific type, check for inline type
2068 BuildCutout unless(this, inline_type_test(cast, /* is_inline = */ false), PROB_MAX);
2069 inc_sp(2);
2070 uncommon_trap_exact(Deoptimization::Reason_class_check, Deoptimization::Action_maybe_recompile);
2071 }
2072
2073 Node* ne_region = new RegionNode(2);
2074 ne_region->add_req(null_ctl);
2075 ne_region->add_req(control());
2076
2077 record_for_igvn(ne_region);
2078 set_control(_gvn.transform(ne_region));
2079 if (btest == BoolTest::ne) {
2080 {
2081 PreserveJVMState pjvms(this);
2082 if (null_ctl == top()) {
2083 replace_in_map(input, cast);
2084 }
2085 int target_bci = iter().get_dest();
2086 merge(target_bci);
2087 }
2088 record_for_igvn(eq_region);
2089 set_control(_gvn.transform(eq_region));
2090 } else {
2091 if (null_ctl == top()) {
2092 replace_in_map(input, cast);
2093 }
2094 set_control(_gvn.transform(ne_region));
2095 }
2096 }
2097
2098 void Parse::do_acmp(BoolTest::mask btest, Node* left, Node* right) {
2099 ciKlass* left_type = nullptr;
2100 ciKlass* right_type = nullptr;
2101 ProfilePtrKind left_ptr = ProfileUnknownNull;
2102 ProfilePtrKind right_ptr = ProfileUnknownNull;
2103 bool left_inline_type = true;
2104 bool right_inline_type = true;
2105
2106 // Leverage profiling at acmp
2107 if (UseACmpProfile) {
2108 method()->acmp_profiled_type(bci(), left_type, right_type, left_ptr, right_ptr, left_inline_type, right_inline_type);
2109 if (too_many_traps_or_recompiles(Deoptimization::Reason_class_check)) {
2110 left_type = nullptr;
2111 right_type = nullptr;
2112 left_inline_type = true;
2113 right_inline_type = true;
2114 }
2115 if (too_many_traps_or_recompiles(Deoptimization::Reason_null_check)) {
2116 left_ptr = ProfileUnknownNull;
2117 right_ptr = ProfileUnknownNull;
2118 }
2119 }
2120
2121 if (UseTypeSpeculation) {
2122 record_profile_for_speculation(left, left_type, left_ptr);
2123 record_profile_for_speculation(right, right_type, right_ptr);
2124 }
2125
2126 if (!Arguments::is_valhalla_enabled()) {
2127 Node* cmp = CmpP(left, right);
2128 cmp = optimize_cmp_with_klass(cmp);
2129 do_if(btest, cmp);
2130 return;
2131 }
2132
2133 // Check for equality before potentially allocating
2134 if (left == right) {
2135 do_if(btest, makecon(TypeInt::CC_EQ));
2136 return;
2137 }
2138
2139 // Allocate inline type operands and re-execute on deoptimization
2140 if (left->is_InlineType()) {
2141 PreserveReexecuteState preexecs(this);
2142 inc_sp(2);
2143 jvms()->set_should_reexecute(true);
2144 left = left->as_InlineType()->buffer(this);
2145 }
2146 if (right->is_InlineType()) {
2147 PreserveReexecuteState preexecs(this);
2148 inc_sp(2);
2149 jvms()->set_should_reexecute(true);
2150 right = right->as_InlineType()->buffer(this);
2151 }
2152
2153 // First, do a normal pointer comparison
2154 const TypeOopPtr* tleft = _gvn.type(left)->isa_oopptr();
2155 const TypeOopPtr* tright = _gvn.type(right)->isa_oopptr();
2156 Node* cmp = CmpP(left, right);
2157 record_for_igvn(cmp);
2158 cmp = optimize_cmp_with_klass(cmp);
2159 if (tleft == nullptr || !tleft->can_be_inline_type() ||
2160 tright == nullptr || !tright->can_be_inline_type()) {
2161 // This is sufficient, if one of the operands can't be an inline type
2162 do_if(btest, cmp);
2163 return;
2164 }
2165
2166 // Don't add traps to unstable if branches because additional checks are required to
2167 // decide if the operands are equal/substitutable and we therefore shouldn't prune
2168 // branches for one if based on the profiling of the acmp branches.
2169 // Also, OptimizeUnstableIf would set an incorrect re-rexecution state because it
2170 // assumes that there is a 1-1 mapping between the if and the acmp branches and that
2171 // hitting a trap means that we will take the corresponding acmp branch on re-execution.
2172 const bool can_trap = true;
2173
2174 Node* eq_region = nullptr;
2175 if (btest == BoolTest::eq) {
2176 do_if(btest, cmp, !can_trap, true);
2177 if (stopped()) {
2178 // Pointers are equal, operands must be equal
2179 return;
2180 }
2181 } else {
2182 assert(btest == BoolTest::ne, "only eq or ne");
2183 Node* is_not_equal = nullptr;
2184 eq_region = new RegionNode(4);
2185 {
2186 PreserveJVMState pjvms(this);
2187 // Pointers are not equal, but more checks are needed to determine if the operands are (not) substitutable
2188 do_if(btest, cmp, !can_trap, false, &is_not_equal);
2189 if (!stopped()) {
2190 eq_region->init_req(1, control());
2191 }
2192 }
2193 if (is_not_equal == nullptr || is_not_equal->is_top()) {
2194 record_for_igvn(eq_region);
2195 set_control(_gvn.transform(eq_region));
2196 return;
2197 }
2198 set_control(is_not_equal);
2199 }
2200
2201 // Prefer speculative types if available
2202 if (!too_many_traps_or_recompiles(Deoptimization::Reason_speculate_class_check)) {
2203 if (tleft->speculative_type() != nullptr) {
2204 left_type = tleft->speculative_type();
2205 }
2206 if (tright->speculative_type() != nullptr) {
2207 right_type = tright->speculative_type();
2208 }
2209 }
2210
2211 if (speculative_ptr_kind(tleft) != ProfileMaybeNull && speculative_ptr_kind(tleft) != ProfileUnknownNull) {
2212 ProfilePtrKind speculative_left_ptr = speculative_ptr_kind(tleft);
2213 if (speculative_left_ptr == ProfileAlwaysNull && !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_null_assert)) {
2214 left_ptr = speculative_left_ptr;
2215 } else if (speculative_left_ptr == ProfileNeverNull && !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_null_check)) {
2216 left_ptr = speculative_left_ptr;
2217 }
2218 }
2219 if (speculative_ptr_kind(tright) != ProfileMaybeNull && speculative_ptr_kind(tright) != ProfileUnknownNull) {
2220 ProfilePtrKind speculative_right_ptr = speculative_ptr_kind(tright);
2221 if (speculative_right_ptr == ProfileAlwaysNull && !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_null_assert)) {
2222 right_ptr = speculative_right_ptr;
2223 } else if (speculative_right_ptr == ProfileNeverNull && !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_null_check)) {
2224 right_ptr = speculative_right_ptr;
2225 }
2226 }
2227
2228 if (left_ptr == ProfileAlwaysNull) {
2229 // Comparison with null. Assert the input is indeed null and we're done.
2230 acmp_always_null_input(left, tleft, btest, eq_region);
2231 return;
2232 }
2233 if (right_ptr == ProfileAlwaysNull) {
2234 // Comparison with null. Assert the input is indeed null and we're done.
2235 acmp_always_null_input(right, tright, btest, eq_region);
2236 return;
2237 }
2238 if (left_type != nullptr && !left_type->is_inlinetype()) {
2239 // Comparison with an object of known type
2240 acmp_type_check(left, tleft, left_ptr, left_type, btest, eq_region);
2241 return;
2242 }
2243 if (right_type != nullptr && !right_type->is_inlinetype()) {
2244 // Comparison with an object of known type
2245 acmp_type_check(right, tright, right_ptr, right_type, btest, eq_region);
2246 return;
2247 }
2248 if (!left_inline_type) {
2249 // Comparison with an object known not to be an inline type
2250 acmp_type_check(left, tleft, left_ptr, nullptr, btest, eq_region);
2251 return;
2252 }
2253 if (!right_inline_type) {
2254 // Comparison with an object known not to be an inline type
2255 acmp_type_check(right, tright, right_ptr, nullptr, btest, eq_region);
2256 return;
2257 }
2258
2259 // Pointers are not equal, check if first operand is non-null
2260 Node* ne_region = new RegionNode(7);
2261 Node* null_ctl = nullptr;
2262 Node* not_null_left = nullptr;
2263 Node* not_null_right = acmp_null_check(right, tright, right_ptr, null_ctl);
2264 ne_region->init_req(1, null_ctl);
2265
2266 Node* kls_right = nullptr;
2267 if (!stopped()) {
2268 // First operand is non-null, check if it is the speculative inline type if possible
2269 // (which later allows isSubstitutable to be intrinsified), or any inline type if no
2270 // speculation is available.
2271 if (right_type != nullptr && right_type->is_inlinetype()) {
2272 acmp_type_check_or_trap(¬_null_right, right_type, Deoptimization::Reason_speculate_class_check);
2273 } else {
2274 Node* is_value = inline_type_test(not_null_right);
2275 IfNode* is_value_iff = create_and_map_if(control(), is_value, PROB_FAIR, COUNT_UNKNOWN);
2276 Node* not_value = _gvn.transform(new IfFalseNode(is_value_iff));
2277 ne_region->init_req(2, not_value);
2278 set_control(_gvn.transform(new IfTrueNode(is_value_iff)));
2279 }
2280
2281 // The first operand is an inline type, check if the second operand is non-null
2282 not_null_left = acmp_null_check(left, tleft, left_ptr, null_ctl);
2283 ne_region->init_req(3, null_ctl);
2284 if (!stopped()) {
2285 // Check if lhs operand is of a specific speculative inline type (see above).
2286 // If not, we don't need to enforce that the lhs is a value object since we know
2287 // it already for the rhs, and must enforce that they have the same type.
2288 if (left_type != nullptr && left_type->is_inlinetype()) {
2289 acmp_type_check_or_trap(¬_null_left, left_type, Deoptimization::Reason_speculate_class_check);
2290 }
2291 if (!stopped()) {
2292 // Check if both operands are of the same class.
2293 Node* kls_left = load_object_klass(not_null_left);
2294 kls_right = load_object_klass(not_null_right);
2295 Node* kls_cmp = CmpP(kls_left, kls_right);
2296 Node* kls_bol = _gvn.transform(new BoolNode(kls_cmp, BoolTest::ne));
2297 IfNode* kls_iff = create_and_map_if(control(), kls_bol, PROB_FAIR, COUNT_UNKNOWN);
2298 Node* kls_ne = _gvn.transform(new IfTrueNode(kls_iff));
2299 set_control(_gvn.transform(new IfFalseNode(kls_iff)));
2300 ne_region->init_req(4, kls_ne);
2301 }
2302 }
2303 }
2304
2305 if (stopped()) {
2306 record_for_igvn(ne_region);
2307 set_control(_gvn.transform(ne_region));
2308 if (btest == BoolTest::ne) {
2309 {
2310 PreserveJVMState pjvms(this);
2311 int target_bci = iter().get_dest();
2312 merge(target_bci);
2313 }
2314 record_for_igvn(eq_region);
2315 set_control(_gvn.transform(eq_region));
2316 }
2317 return;
2318 }
2319 assert(kls_right != nullptr, "");
2320
2321 IfNode* mask_iff = nullptr;
2322 // If any operand has a precisely known type, isSubstitutable will be intrinsified, so we don't need the fast path
2323 if (UseAcmpFastPath && !_gvn.type(not_null_left)->is_inlinetypeptr() && !_gvn.type(not_null_right)->is_inlinetypeptr()) {
2324 /* Here, we are generating the fast path (the slow path being the call to isSubstitutable)
2325 * See the declarations of _fast_acmp_offset and _fast_acmp_mask in InlineKlass::Members
2326 * for details about the fast path logic, and the meaning of these values.
2327 */
2328 Node* members_addr = off_heap_plus_addr(kls_right, in_bytes(InlineKlass::adr_members_offset()));
2329 Node* members = make_load(control(), members_addr, TypeRawPtr::BOTTOM, T_ADDRESS, MemNode::unordered);
2330 Node* offset_addr = off_heap_plus_addr(members, in_bytes(InlineKlass::fast_acmp_offset_offset()));
2331 Node* offset = make_load(control(), offset_addr, TypeInt::INT, T_INT, MemNode::unordered);
2332
2333 Node* offset_cmp = CmpI(offset, zerocon(T_INT));
2334 Node* offset_bol = _gvn.transform(new BoolNode(offset_cmp, BoolTest::lt));
2335 mask_iff = create_and_map_if(control(), offset_bol, PROB_FAIR, COUNT_UNKNOWN);
2336 Node* slow_path_ctl = _gvn.transform(new IfTrueNode(mask_iff));
2337 Node* fast_path_ctl = _gvn.transform(new IfFalseNode(mask_iff));
2338 set_control(slow_path_ctl);
2339
2340 {
2341 PreserveJVMState jvms(this);
2342 set_control(fast_path_ctl);
2343
2344 Node* offset_l = ConvI2L(offset);
2345 Node* fast_acmp_mask_addr = off_heap_plus_addr(members, in_bytes(InlineKlass::fast_acmp_mask_offset()));
2346 Node* fast_acmp_mask = make_load(control(), fast_acmp_mask_addr, TypeLong::LONG, T_LONG, MemNode::unordered);
2347
2348 // *(left + offset) & mask == *(right + offset) & mask
2349 Node* left_payload_addr = basic_plus_adr(not_null_left, offset_l);
2350 Node* left_payload = make_load(control(), left_payload_addr, TypeLong::LONG, T_LONG, MemNode::unordered, LoadNNode::DependsOnlyOnTest, false, true, true, true);
2351 Node* left_masked = _gvn.transform(new AndLNode(left_payload, fast_acmp_mask));
2352
2353 Node* right_payload_addr = basic_plus_adr(not_null_right, offset_l);
2354 Node* right_payload = make_load(control(), right_payload_addr, TypeLong::LONG, T_LONG, MemNode::unordered, LoadNNode::DependsOnlyOnTest, false, true, true, true);
2355 Node* right_masked = _gvn.transform(new AndLNode(right_payload, fast_acmp_mask));
2356
2357 Node* masked_cmp = CmpL(left_masked, right_masked);
2358
2359 Node* ctl = C->top();
2360 if (btest == BoolTest::eq) {
2361 PreserveJVMState pjvms(this);
2362 do_if(btest, masked_cmp, !can_trap, true, nullptr);
2363 if (!stopped()) {
2364 ctl = control();
2365 }
2366 } else {
2367 assert(btest == BoolTest::ne, "only eq or ne");
2368 PreserveJVMState pjvms(this);
2369 do_if(btest, masked_cmp, !can_trap, false, &ctl);
2370 if (!stopped()) {
2371 eq_region->init_req(3, control());
2372 }
2373 }
2374 ne_region->init_req(6, ctl);
2375 }
2376 }
2377
2378 // Both operands are values types of the same class, we need to perform a
2379 // substitutability test. Delegate to ValueObjectMethods::isSubstitutable().
2380 Node* ne_io_phi = PhiNode::make(ne_region, i_o());
2381 Node* mem = reset_memory();
2382 Node* ne_mem_phi = PhiNode::make(ne_region, mem);
2383
2384 Node* eq_io_phi = nullptr;
2385 Node* eq_mem_phi = nullptr;
2386 if (eq_region != nullptr) {
2387 eq_io_phi = PhiNode::make(eq_region, i_o());
2388 eq_mem_phi = PhiNode::make(eq_region, mem);
2389 }
2390
2391 set_all_memory(mem);
2392
2393 kill_dead_locals();
2394 ciSymbol* subst_method_name = ciSymbols::isSubstitutable_name();
2395 ciMethod* subst_method = ciEnv::current()->ValueObjectMethods_klass()->find_method(subst_method_name, ciSymbols::object_object_boolean_signature());
2396 CallStaticJavaNode* call = new CallStaticJavaNode(C, TypeFunc::make(subst_method), SharedRuntime::get_resolve_static_call_stub(), subst_method);
2397 call->set_override_symbolic_info(true);
2398 call->init_req(TypeFunc::Parms, not_null_left);
2399 call->init_req(TypeFunc::Parms+1, not_null_right);
2400 inc_sp(2);
2401 set_edges_for_java_call(call, false, false);
2402 Node* ret = set_results_for_java_call(call, false, true);
2403 dec_sp(2);
2404
2405 assert(acmp_fast_path_if_from_substitutable_call(&_gvn, call) == mask_iff, "");
2406
2407 // Test the return value of ValueObjectMethods::isSubstitutable()
2408 // This is the last check, do_if can emit traps now.
2409 Node* subst_cmp = _gvn.transform(new CmpINode(ret, intcon(1)));
2410 Node* ctl = C->top();
2411 Node* mem_taken = nullptr;
2412 Node* io_taken = nullptr;
2413 if (btest == BoolTest::eq) {
2414 PreserveJVMState pjvms(this);
2415 do_if(btest, subst_cmp, can_trap, false, nullptr, &mem_taken, &io_taken);
2416 if (!stopped()) {
2417 ctl = control();
2418 mem_taken = reset_memory();
2419 io_taken = i_o();
2420 }
2421 } else {
2422 assert(btest == BoolTest::ne, "only eq or ne");
2423 PreserveJVMState pjvms(this);
2424 do_if(btest, subst_cmp, can_trap, false, &ctl, &mem_taken, &io_taken);
2425 if (!stopped()) {
2426 eq_region->init_req(2, control());
2427 eq_io_phi->init_req(2, i_o());
2428 eq_mem_phi->init_req(2, reset_memory());
2429 }
2430 }
2431 ne_region->init_req(5, ctl);
2432 ne_io_phi->init_req(5, io_taken);
2433 ne_mem_phi->init_req(5, mem_taken);
2434
2435 record_for_igvn(ne_region);
2436 set_control(_gvn.transform(ne_region));
2437 set_i_o(_gvn.transform(ne_io_phi));
2438 set_all_memory(_gvn.transform(ne_mem_phi));
2439
2440 if (btest == BoolTest::ne) {
2441 {
2442 PreserveJVMState pjvms(this);
2443 int target_bci = iter().get_dest();
2444 merge(target_bci);
2445 }
2446
2447 record_for_igvn(eq_region);
2448 set_control(_gvn.transform(eq_region));
2449 set_i_o(_gvn.transform(eq_io_phi));
2450 set_all_memory(_gvn.transform(eq_mem_phi));
2451 }
2452 }
2453
2454 /* Detects whether a call to isSubstitutable is under an IfNode guarding the fast path for acmp.
2455 * If so, returns the IfNode branching between the call and the fast path. Returns null otherwise.
2456 *
2457 * The fast path is a LOT easier to generate at parsing time, but can be later proven useless if further
2458 * optimization narrows down the type of operands and allows intrinsification of the substitutability
2459 * check. In this case, the fast path might still apply, but it comes with various downsides, such as
2460 * mismatch access that may hinder optimizations, or buffering requirement. So, when intrinsifying the call,
2461 * we try to remove the fast path.
2462 *
2463 * This test isn't so bad. Loading the fast acmp offset is pretty unique to the fast acmp path.
2464 *
2465 * Clearly, this is only a step before a proper solution for acmp, such as a macro node.
2466 */
2467 IfNode* Parse::acmp_fast_path_if_from_substitutable_call(PhaseGVN* phase, CallStaticJavaNode* call) {
2468 auto is_con_offset = [](Node* node, ByteSize n) -> bool {
2469 if (!node->is_Con()) return false;
2470 TypeNode* con = node->as_Type();
2471 assert(con->type()->is_intptr_t(), "");
2472 return con->type()->is_intptr_t()->is_con(in_bytes(n));
2473 };
2474
2475 assert(call->in(TypeFunc::Control) != nullptr, "");
2476 if (!call->in(TypeFunc::Control)->is_IfProj()) return nullptr;
2477 IfProjNode* if_proj = call->in(TypeFunc::Control)->as_IfProj();
2478 if (if_proj->_con != 1) return nullptr;
2479
2480 assert(if_proj->in(0) != nullptr, "");
2481 assert(if_proj->in(0)->is_If(), "");
2482 IfNode* iff = if_proj->in(0)->as_If();
2483
2484 assert(iff->in(1) != nullptr, "");
2485 if (!iff->in(1)->is_Bool()) return nullptr;
2486 BoolNode* lt = iff->in(1)->as_Bool();
2487 if (lt->_test._test != BoolTest::lt) return nullptr;
2488
2489 assert(lt->in(1) != nullptr, "");
2490 if (lt->in(1)->Opcode() != Op_CmpI) return nullptr;
2491 CmpNode* cmp_i = lt->in(1)->as_Cmp();
2492
2493 assert(cmp_i->in(1) != nullptr, "");
2494 assert(cmp_i->in(2) != nullptr, "");
2495
2496 if (cmp_i->in(1)->Opcode() != Op_LoadI) return nullptr;
2497 LoadNode* load_offset = cmp_i->in(1)->as_Load();
2498 if (!cmp_i->in(2)->is_ConI()) return nullptr;
2499 ConINode* zero_i = cmp_i->in(2)->as_ConI();
2500 assert(zero_i->type()->is_int() != nullptr, "");
2501 if (!zero_i->type()->is_int()->is_con(0)) return nullptr;
2502
2503 assert(load_offset->in(2) != nullptr, "");
2504 if (!load_offset->in(2)->is_AddP()) return nullptr;
2505 AddPNode* offset_addr_add = load_offset->in(2)->as_AddP();
2506
2507 assert(offset_addr_add->in(AddPNode::Base) != nullptr, "");
2508 assert(offset_addr_add->in(AddPNode::Address) != nullptr, "");
2509 assert(offset_addr_add->in(AddPNode::Offset) != nullptr, "");
2510 if (!offset_addr_add->in(AddPNode::Base)->is_top()) return nullptr;
2511 if (offset_addr_add->in(AddPNode::Address)->Opcode() != Op_LoadP) return nullptr;
2512 LoadNode* load_members = offset_addr_add->in(AddPNode::Address)->as_Load();
2513 if (!is_con_offset(offset_addr_add->in(AddPNode::Offset), InlineKlass::fast_acmp_offset_offset())) return nullptr;
2514
2515 assert(load_members->in(2) != nullptr, "");
2516 if (!load_members->in(2)->is_AddP()) return nullptr;
2517 AddPNode* members_addr_add = load_members->in(2)->as_AddP();
2518
2519 assert(members_addr_add->in(AddPNode::Base) != nullptr, "");
2520 assert(members_addr_add->in(AddPNode::Address) != nullptr, "");
2521 assert(members_addr_add->in(AddPNode::Offset) != nullptr, "");
2522 if (!members_addr_add->in(AddPNode::Base)->is_top()) return nullptr;
2523 if (!phase->type(members_addr_add->in(AddPNode::Address))->isa_instklassptr()) return nullptr;
2524 if (!is_con_offset(members_addr_add->in(AddPNode::Offset), InlineKlass::adr_members_offset())) return nullptr;
2525
2526 return iff;
2527 }
2528
2529 // Force unstable if traps to be taken randomly to trigger intermittent bugs such as incorrect debug information.
2530 // Add another if before the unstable if that checks a "random" condition at runtime (a simple shared counter) and
2531 // then either takes the trap or executes the original, unstable if.
2532 void Parse::stress_trap(IfNode* orig_iff, Node* counter, Node* incr_store) {
2533 // Search for an unstable if trap
2534 CallStaticJavaNode* trap = nullptr;
2535 assert(orig_iff->Opcode() == Op_If && orig_iff->outcnt() == 2, "malformed if");
2536 ProjNode* trap_proj = orig_iff->uncommon_trap_proj(trap, Deoptimization::Reason_unstable_if);
2537 if (trap == nullptr || !trap->jvms()->should_reexecute()) {
2538 // No suitable trap found. Remove unused counter load and increment.
2539 C->gvn_replace_by(incr_store, incr_store->in(MemNode::Memory));
2540 return;
2541 }
2542
2543 // Remove trap from optimization list since we add another path to the trap.
2544 bool success = C->remove_unstable_if_trap(trap, true);
2545 assert(success, "Trap already modified");
2546
2547 // Add a check before the original if that will trap with a certain frequency and execute the original if otherwise
2548 int freq_log = (C->random() % 31) + 1; // Random logarithmic frequency in [1, 31]
2549 Node* mask = intcon(right_n_bits(freq_log));
2550 counter = _gvn.transform(new AndINode(counter, mask));
2551 Node* cmp = _gvn.transform(new CmpINode(counter, intcon(0)));
2552 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::mask::eq));
2553 IfNode* iff = _gvn.transform(new IfNode(orig_iff->in(0), bol, orig_iff->_prob, orig_iff->_fcnt))->as_If();
2554 Node* if_true = _gvn.transform(new IfTrueNode(iff));
2555 Node* if_false = _gvn.transform(new IfFalseNode(iff));
2556 assert(!if_true->is_top() && !if_false->is_top(), "trap always / never taken");
2557
2558 // Trap
2559 assert(trap_proj->outcnt() == 1, "some other nodes are dependent on the trap projection");
2560
2561 Node* trap_region = new RegionNode(3);
2562 trap_region->set_req(1, trap_proj);
2563 trap_region->set_req(2, if_true);
2564 trap->set_req(0, _gvn.transform(trap_region));
2565
2566 // Don't trap, execute original if
2567 orig_iff->set_req(0, if_false);
2568 }
2569
2570 bool Parse::path_is_suitable_for_uncommon_trap(float prob) const {
2571 // Randomly skip emitting an uncommon trap
2572 if (StressUnstableIfTraps && ((C->random() % 2) == 0)) {
2573 return false;
2574 }
2575 // Don't want to speculate on uncommon traps when running with -Xcomp
2576 if (!UseInterpreter) {
2577 return false;
2578 }
2579 return seems_never_taken(prob) &&
2580 !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if);
2581 }
2582
2583 void Parse::maybe_add_predicate_after_if(Block* path) {
2584 if (path->is_SEL_head() && path->preds_parsed() == 0) {
2585 // Add predicates at bci of if dominating the loop so traps can be
2586 // recorded on the if's profile data
2587 int bc_depth = repush_if_args();
2588 add_parse_predicates();
2589 dec_sp(bc_depth);
2590 path->set_has_predicates();
2591 }
2592 }
2593
2594
2595 //----------------------------adjust_map_after_if------------------------------
2596 // Adjust the JVM state to reflect the result of taking this path.
2597 // Basically, it means inspecting the CmpNode controlling this
2598 // branch, seeing how it constrains a tested value, and then
2599 // deciding if it's worth our while to encode this constraint
2600 // as graph nodes in the current abstract interpretation map.
2601 void Parse::adjust_map_after_if(BoolTest::mask btest, Node* c, float prob, Block* path, bool can_trap) {
2602 if (!c->is_Cmp()) {
2603 maybe_add_predicate_after_if(path);
2604 return;
2605 }
2606
2607 if (stopped() || btest == BoolTest::illegal) {
2608 return; // nothing to do
2609 }
2610
2611 bool is_fallthrough = (path == successor_for_bci(iter().next_bci()));
2612
2613 if (can_trap && path_is_suitable_for_uncommon_trap(prob)) {
2614 repush_if_args();
2615 Node* call = uncommon_trap(Deoptimization::Reason_unstable_if,
2616 Deoptimization::Action_reinterpret,
2617 nullptr,
2618 (is_fallthrough ? "taken always" : "taken never"));
2619
2620 if (call != nullptr) {
2621 C->record_unstable_if_trap(new UnstableIfTrap(call->as_CallStaticJava(), path));
2622 }
2623 return;
2624 }
2625
2626 if (c->is_FlatArrayCheck()) {
2627 maybe_add_predicate_after_if(path);
2628 return;
2629 }
2630
2631 Node* val = c->in(1);
2632 Node* con = c->in(2);
2633 const Type* tcon = _gvn.type(con);
2634 const Type* tval = _gvn.type(val);
2635 bool have_con = tcon->singleton();
2636 if (tval->singleton()) {
2637 if (!have_con) {
2638 // Swap, so constant is in con.
2639 con = val;
2640 tcon = tval;
2641 val = c->in(2);
2642 tval = _gvn.type(val);
2643 btest = BoolTest(btest).commute();
2644 have_con = true;
2645 } else {
2646 // Do we have two constants? Then leave well enough alone.
2647 have_con = false;
2648 }
2649 }
2650 if (!have_con) { // remaining adjustments need a con
2651 maybe_add_predicate_after_if(path);
2652 return;
2653 }
2654
2655 sharpen_type_after_if(btest, con, tcon, val, tval);
2656 maybe_add_predicate_after_if(path);
2657 }
2658
2659
2660 static Node* extract_obj_from_klass_load(PhaseGVN* gvn, Node* n) {
2661 Node* ldk;
2662 if (n->is_DecodeNKlass()) {
2663 if (n->in(1)->Opcode() != Op_LoadNKlass) {
2664 return nullptr;
2665 } else {
2666 ldk = n->in(1);
2667 }
2668 } else if (n->Opcode() != Op_LoadKlass) {
2669 return nullptr;
2670 } else {
2671 ldk = n;
2672 }
2673 assert(ldk != nullptr && ldk->is_Load(), "should have found a LoadKlass or LoadNKlass node");
2674
2675 Node* adr = ldk->in(MemNode::Address);
2676 intptr_t off = 0;
2677 Node* obj = AddPNode::Ideal_base_and_offset(adr, gvn, off);
2678 if (obj == nullptr || off != oopDesc::klass_offset_in_bytes()) // loading oopDesc::_klass?
2679 return nullptr;
2680 const TypePtr* tp = gvn->type(obj)->is_ptr();
2681 if (tp == nullptr || !(tp->isa_instptr() || tp->isa_aryptr())) // is obj a Java object ptr?
2682 return nullptr;
2683
2684 return obj;
2685 }
2686
2687 // Matches exact and inexact type check IR shapes during parsing.
2688 // On successful match, returns type checked object node and its type after successful check
2689 // as out parameters.
2690 static bool match_type_check(PhaseGVN& gvn,
2691 BoolTest::mask btest,
2692 Node* con, const Type* tcon,
2693 Node* val, const Type* tval,
2694 Node** obj, const TypeOopPtr** cast_type) { // out-parameters
2695 // Look for opportunities to sharpen the type of a node whose klass is compared with a constant klass.
2696 // The constant klass being tested against can come from many bytecode instructions (implicitly or explicitly),
2697 // and also from profile data used by speculative casts.
2698 if (btest == BoolTest::eq && tcon->isa_klassptr()) {
2699 // Found:
2700 // Bool(CmpP(LoadKlass(obj._klass), ConP(Foo.klass)), [eq])
2701 // or the narrowOop equivalent.
2702 (*obj) = extract_obj_from_klass_load(&gvn, val);
2703 // Some klass comparisons are not directly in the form
2704 // Bool(CmpP(LoadKlass(obj._klass), ConP(Foo.klass)), [eq]),
2705 // e.g. Bool(CmpP(CastPP(LoadKlass(...)), ConP(klass)), [eq]).
2706 // These patterns with nullable klasses arise from example from
2707 // load_array_klass_from_mirror.
2708 if (*obj == nullptr) { return false; }
2709 (*cast_type) = tcon->isa_klassptr()->as_instance_type();
2710 return true; // found
2711 }
2712
2713 // Match an instanceof check.
2714 // During parsing its IR shape is not canonicalized yet.
2715 //
2716 // obj superklass
2717 // | |
2718 // SubTypeCheck
2719 // |
2720 // Bool [eq] / [ne]
2721 // |
2722 // If
2723 // / \
2724 // T F
2725 // \ /
2726 // Region
2727 // \ ConI ConI
2728 // \ | /
2729 // val -> Phi ConI <- con
2730 // \ /
2731 // CmpI
2732 // |
2733 // Bool [btest]
2734 // |
2735 //
2736 if (tval->isa_int() && val->is_Phi() && val->in(0)->as_Region()->is_diamond()) {
2737 RegionNode* diamond = val->in(0)->as_Region();
2738 IfNode* if1 = diamond->in(1)->in(0)->as_If();
2739 BoolNode* b1 = if1->in(1)->isa_Bool();
2740 if (b1 != nullptr && b1->in(1)->isa_SubTypeCheck()) {
2741 assert(b1->_test._test == BoolTest::eq ||
2742 b1->_test._test == BoolTest::ne, "%d", b1->_test._test);
2743
2744 ProjNode* success_proj = if1->proj_out(b1->_test._test == BoolTest::eq ? 1 : 0);
2745 int idx = diamond->find_edge(success_proj);
2746 assert(idx == 1 || idx == 2, "");
2747 Node* vcon = val->in(idx);
2748
2749 if ((btest == BoolTest::eq && vcon == con) || (btest == BoolTest::ne && vcon != con)) {
2750 assert(val->find_edge(con) > 0, "mismatch");
2751 SubTypeCheckNode* sub = b1->in(1)->as_SubTypeCheck();
2752 Node* obj_or_subklass = sub->in(SubTypeCheckNode::ObjOrSubKlass);
2753 Node* superklass = sub->in(SubTypeCheckNode::SuperKlass);
2754
2755 if (gvn.type(obj_or_subklass)->isa_oopptr()) {
2756 const TypeKlassPtr* klass_ptr_type = gvn.type(superklass)->is_klassptr();
2757 const TypeKlassPtr* improved_klass_ptr_type = klass_ptr_type->try_improve();
2758
2759 (*obj) = obj_or_subklass;
2760 (*cast_type) = improved_klass_ptr_type->cast_to_exactness(false)->as_instance_type();
2761 return true; // found
2762 }
2763 }
2764 }
2765 }
2766 return false; // not found
2767 }
2768
2769 void Parse::sharpen_type_after_if(BoolTest::mask btest,
2770 Node* con, const Type* tcon,
2771 Node* val, const Type* tval) {
2772 Node* obj = nullptr;
2773 const TypeOopPtr* cast_type = nullptr;
2774 // Insert a cast node with a narrowed type after a successful type check.
2775 if (match_type_check(_gvn, btest, con, tcon, val, tval,
2776 &obj, &cast_type)) {
2777 assert(obj != nullptr && cast_type != nullptr, "missing type check info");
2778 const Type* obj_type = _gvn.type(obj);
2779 const Type* tboth = obj_type->filter_speculative(cast_type);
2780 assert(tboth->higher_equal(obj_type) && tboth->higher_equal(cast_type), "sanity");
2781 if (tboth == Type::TOP && KillPathsReachableByDeadTypeNode) {
2782 // Let dead type node cleaning logic prune effectively dead path for us.
2783 // CheckCastPP::Value() == TOP and it will trigger the cleanup during GVN.
2784 // Don't materialize the cast when cleanup is disabled, because
2785 // it kills data and control leaving IR in broken state.
2786 tboth = cast_type;
2787 }
2788 if (tboth != Type::TOP && tboth != obj_type) {
2789 int obj_in_map = map()->find_edge(obj);
2790 if (obj_in_map >= 0 &&
2791 (jvms()->is_loc(obj_in_map) || jvms()->is_stk(obj_in_map))) {
2792 TypeNode* ccast = new CheckCastPPNode(control(), obj, tboth);
2793 // Delay transform() call to allow recovery of pre-cast value at the control merge.
2794 _gvn.set_type_bottom(ccast);
2795 record_for_igvn(ccast);
2796 if (tboth->is_inlinetypeptr()) {
2797 ccast = InlineTypeNode::make_from_oop(this, ccast, tboth->isa_oopptr()->exact_klass(true)->as_inline_klass());
2798 }
2799 // Here's the payoff.
2800 replace_in_map(obj, ccast);
2801 }
2802 }
2803 }
2804
2805 int val_in_map = map()->find_edge(val);
2806 if (val_in_map < 0) return; // replace_in_map would be useless
2807 {
2808 JVMState* jvms = this->jvms();
2809 if (!(jvms->is_loc(val_in_map) ||
2810 jvms->is_stk(val_in_map)))
2811 return; // again, it would be useless
2812 }
2813
2814 // Check for a comparison to a constant, and "know" that the compared
2815 // value is constrained on this path.
2816 assert(tcon->singleton(), "");
2817 ConstraintCastNode* ccast = nullptr;
2818 Node* cast = nullptr;
2819
2820 switch (btest) {
2821 case BoolTest::eq: // Constant test?
2822 {
2823 const Type* tboth = tcon->join_speculative(tval);
2824 if (tboth == tval) break; // Nothing to gain.
2825 if (tcon->isa_int()) {
2826 ccast = new CastIINode(control(), val, tboth);
2827 } else if (tcon == TypePtr::NULL_PTR) {
2828 // Cast to null, but keep the pointer identity temporarily live.
2829 ccast = new CastPPNode(control(), val, tboth);
2830 } else {
2831 const TypeF* tf = tcon->isa_float_constant();
2832 const TypeD* td = tcon->isa_double_constant();
2833 // Exclude tests vs float/double 0 as these could be
2834 // either +0 or -0. Just because you are equal to +0
2835 // doesn't mean you ARE +0!
2836 // Note, following code also replaces Long and Oop values.
2837 if ((!tf || tf->_f != 0.0) &&
2838 (!td || td->_d != 0.0))
2839 cast = con; // Replace non-constant val by con.
2840 }
2841 }
2842 break;
2843
2844 case BoolTest::ne:
2845 if (tcon == TypePtr::NULL_PTR) {
2846 cast = cast_not_null(val, false);
2847 }
2848 break;
2849
2850 default:
2851 // (At this point we could record int range types with CastII.)
2852 break;
2853 }
2854
2855 if (ccast != nullptr) {
2856 const Type* tcc = ccast->as_Type()->type();
2857 assert(tcc != tval && tcc->higher_equal(tval), "must improve");
2858 // Delay transform() call to allow recovery of pre-cast value
2859 // at the control merge.
2860 _gvn.set_type_bottom(ccast);
2861 record_for_igvn(ccast);
2862 cast = ccast;
2863 }
2864
2865 if (cast != nullptr) { // Here's the payoff.
2866 replace_in_map(val, cast);
2867 }
2868 }
2869
2870 /**
2871 * Use speculative type to optimize CmpP node: if comparison is
2872 * against the low level class, cast the object to the speculative
2873 * type if any. CmpP should then go away.
2874 *
2875 * @param c expected CmpP node
2876 * @return result of CmpP on object casted to speculative type
2877 *
2878 */
2879 Node* Parse::optimize_cmp_with_klass(Node* c) {
2880 // If this is transformed by the _gvn to a comparison with the low
2881 // level klass then we may be able to use speculation
2882 if (c->Opcode() == Op_CmpP &&
2883 (c->in(1)->Opcode() == Op_LoadKlass || c->in(1)->Opcode() == Op_DecodeNKlass) &&
2884 c->in(2)->is_Con()) {
2885 Node* load_klass = nullptr;
2886 Node* decode = nullptr;
2887 if (c->in(1)->Opcode() == Op_DecodeNKlass) {
2888 decode = c->in(1);
2889 load_klass = c->in(1)->in(1);
2890 } else {
2891 load_klass = c->in(1);
2892 }
2893 if (load_klass->in(2)->is_AddP()) {
2894 Node* addp = load_klass->in(2);
2895 Node* obj = addp->in(AddPNode::Address);
2896 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
2897 if (obj_type->speculative_type_not_null() != nullptr) {
2898 ciKlass* k = obj_type->speculative_type();
2899 inc_sp(2);
2900 obj = maybe_cast_profiled_obj(obj, k);
2901 dec_sp(2);
2902 if (obj->is_InlineType()) {
2903 assert(obj->as_InlineType()->is_allocated(&_gvn), "must be allocated");
2904 obj = obj->as_InlineType()->get_oop();
2905 }
2906 // Make the CmpP use the casted obj
2907 addp = basic_plus_adr(obj, addp->in(AddPNode::Offset));
2908 load_klass = load_klass->clone();
2909 load_klass->set_req(2, addp);
2910 load_klass = _gvn.transform(load_klass);
2911 if (decode != nullptr) {
2912 decode = decode->clone();
2913 decode->set_req(1, load_klass);
2914 load_klass = _gvn.transform(decode);
2915 }
2916 c = c->clone();
2917 c->set_req(1, load_klass);
2918 c = _gvn.transform(c);
2919 }
2920 }
2921 }
2922 return c;
2923 }
2924
2925 //------------------------------do_one_bytecode--------------------------------
2926 // Parse this bytecode, and alter the Parsers JVM->Node mapping
2927 void Parse::do_one_bytecode() {
2928 Node *a, *b, *c, *d; // Handy temps
2929 BoolTest::mask btest;
2930 int i;
2931
2932 assert(!has_exceptions(), "bytecode entry state must be clear of throws");
2933
2934 if (C->check_node_count(NodeLimitFudgeFactor * 5,
2935 "out of nodes parsing method")) {
2936 return;
2937 }
2938
2939 #ifdef ASSERT
2940 // for setting breakpoints
2941 if (TraceOptoParse) {
2942 tty->print(" @");
2943 dump_bci(bci());
2944 tty->print(" %s", Bytecodes::name(bc()));
2945 tty->cr();
2946 }
2947 #endif
2948
2949 switch (bc()) {
2950 case Bytecodes::_nop:
2951 // do nothing
2952 break;
2953 case Bytecodes::_lconst_0:
2954 push_pair(longcon(0));
2955 break;
2956
2957 case Bytecodes::_lconst_1:
2958 push_pair(longcon(1));
2959 break;
2960
2961 case Bytecodes::_fconst_0:
2962 push(zerocon(T_FLOAT));
2963 break;
2964
2965 case Bytecodes::_fconst_1:
2966 push(makecon(TypeF::ONE));
2967 break;
2968
2969 case Bytecodes::_fconst_2:
2970 push(makecon(TypeF::make(2.0f)));
2971 break;
2972
2973 case Bytecodes::_dconst_0:
2974 push_pair(zerocon(T_DOUBLE));
2975 break;
2976
2977 case Bytecodes::_dconst_1:
2978 push_pair(makecon(TypeD::ONE));
2979 break;
2980
2981 case Bytecodes::_iconst_m1:push(intcon(-1)); break;
2982 case Bytecodes::_iconst_0: push(intcon( 0)); break;
2983 case Bytecodes::_iconst_1: push(intcon( 1)); break;
2984 case Bytecodes::_iconst_2: push(intcon( 2)); break;
2985 case Bytecodes::_iconst_3: push(intcon( 3)); break;
2986 case Bytecodes::_iconst_4: push(intcon( 4)); break;
2987 case Bytecodes::_iconst_5: push(intcon( 5)); break;
2988 case Bytecodes::_bipush: push(intcon(iter().get_constant_u1())); break;
2989 case Bytecodes::_sipush: push(intcon(iter().get_constant_u2())); break;
2990 case Bytecodes::_aconst_null: push(null()); break;
2991
2992 case Bytecodes::_ldc:
2993 case Bytecodes::_ldc_w:
2994 case Bytecodes::_ldc2_w: {
2995 // ciTypeFlow should trap if the ldc is in error state or if the constant is not loaded
2996 assert(!iter().is_in_error(), "ldc is in error state");
2997 ciConstant constant = iter().get_constant();
2998 assert(constant.is_loaded(), "constant is not loaded");
2999 const Type* con_type = Type::make_from_constant(constant);
3000 if (con_type != nullptr) {
3001 push_node(con_type->basic_type(), makecon(con_type));
3002 }
3003 break;
3004 }
3005
3006 case Bytecodes::_aload_0:
3007 push( local(0) );
3008 break;
3009 case Bytecodes::_aload_1:
3010 push( local(1) );
3011 break;
3012 case Bytecodes::_aload_2:
3013 push( local(2) );
3014 break;
3015 case Bytecodes::_aload_3:
3016 push( local(3) );
3017 break;
3018 case Bytecodes::_aload:
3019 push( local(iter().get_index()) );
3020 break;
3021
3022 case Bytecodes::_fload_0:
3023 case Bytecodes::_iload_0:
3024 push( local(0) );
3025 break;
3026 case Bytecodes::_fload_1:
3027 case Bytecodes::_iload_1:
3028 push( local(1) );
3029 break;
3030 case Bytecodes::_fload_2:
3031 case Bytecodes::_iload_2:
3032 push( local(2) );
3033 break;
3034 case Bytecodes::_fload_3:
3035 case Bytecodes::_iload_3:
3036 push( local(3) );
3037 break;
3038 case Bytecodes::_fload:
3039 case Bytecodes::_iload:
3040 push( local(iter().get_index()) );
3041 break;
3042 case Bytecodes::_lload_0:
3043 push_pair_local( 0 );
3044 break;
3045 case Bytecodes::_lload_1:
3046 push_pair_local( 1 );
3047 break;
3048 case Bytecodes::_lload_2:
3049 push_pair_local( 2 );
3050 break;
3051 case Bytecodes::_lload_3:
3052 push_pair_local( 3 );
3053 break;
3054 case Bytecodes::_lload:
3055 push_pair_local( iter().get_index() );
3056 break;
3057
3058 case Bytecodes::_dload_0:
3059 push_pair_local(0);
3060 break;
3061 case Bytecodes::_dload_1:
3062 push_pair_local(1);
3063 break;
3064 case Bytecodes::_dload_2:
3065 push_pair_local(2);
3066 break;
3067 case Bytecodes::_dload_3:
3068 push_pair_local(3);
3069 break;
3070 case Bytecodes::_dload:
3071 push_pair_local(iter().get_index());
3072 break;
3073 case Bytecodes::_fstore_0:
3074 case Bytecodes::_istore_0:
3075 case Bytecodes::_astore_0:
3076 set_local( 0, pop() );
3077 break;
3078 case Bytecodes::_fstore_1:
3079 case Bytecodes::_istore_1:
3080 case Bytecodes::_astore_1:
3081 set_local( 1, pop() );
3082 break;
3083 case Bytecodes::_fstore_2:
3084 case Bytecodes::_istore_2:
3085 case Bytecodes::_astore_2:
3086 set_local( 2, pop() );
3087 break;
3088 case Bytecodes::_fstore_3:
3089 case Bytecodes::_istore_3:
3090 case Bytecodes::_astore_3:
3091 set_local( 3, pop() );
3092 break;
3093 case Bytecodes::_fstore:
3094 case Bytecodes::_istore:
3095 case Bytecodes::_astore:
3096 set_local( iter().get_index(), pop() );
3097 break;
3098 // long stores
3099 case Bytecodes::_lstore_0:
3100 set_pair_local( 0, pop_pair() );
3101 break;
3102 case Bytecodes::_lstore_1:
3103 set_pair_local( 1, pop_pair() );
3104 break;
3105 case Bytecodes::_lstore_2:
3106 set_pair_local( 2, pop_pair() );
3107 break;
3108 case Bytecodes::_lstore_3:
3109 set_pair_local( 3, pop_pair() );
3110 break;
3111 case Bytecodes::_lstore:
3112 set_pair_local( iter().get_index(), pop_pair() );
3113 break;
3114
3115 // double stores
3116 case Bytecodes::_dstore_0:
3117 set_pair_local( 0, pop_pair() );
3118 break;
3119 case Bytecodes::_dstore_1:
3120 set_pair_local( 1, pop_pair() );
3121 break;
3122 case Bytecodes::_dstore_2:
3123 set_pair_local( 2, pop_pair() );
3124 break;
3125 case Bytecodes::_dstore_3:
3126 set_pair_local( 3, pop_pair() );
3127 break;
3128 case Bytecodes::_dstore:
3129 set_pair_local( iter().get_index(), pop_pair() );
3130 break;
3131
3132 case Bytecodes::_pop: dec_sp(1); break;
3133 case Bytecodes::_pop2: dec_sp(2); break;
3134 case Bytecodes::_swap:
3135 a = pop();
3136 b = pop();
3137 push(a);
3138 push(b);
3139 break;
3140 case Bytecodes::_dup:
3141 a = pop();
3142 push(a);
3143 push(a);
3144 break;
3145 case Bytecodes::_dup_x1:
3146 a = pop();
3147 b = pop();
3148 push( a );
3149 push( b );
3150 push( a );
3151 break;
3152 case Bytecodes::_dup_x2:
3153 a = pop();
3154 b = pop();
3155 c = pop();
3156 push( a );
3157 push( c );
3158 push( b );
3159 push( a );
3160 break;
3161 case Bytecodes::_dup2:
3162 a = pop();
3163 b = pop();
3164 push( b );
3165 push( a );
3166 push( b );
3167 push( a );
3168 break;
3169
3170 case Bytecodes::_dup2_x1:
3171 // before: .. c, b, a
3172 // after: .. b, a, c, b, a
3173 // not tested
3174 a = pop();
3175 b = pop();
3176 c = pop();
3177 push( b );
3178 push( a );
3179 push( c );
3180 push( b );
3181 push( a );
3182 break;
3183 case Bytecodes::_dup2_x2:
3184 // before: .. d, c, b, a
3185 // after: .. b, a, d, c, b, a
3186 // not tested
3187 a = pop();
3188 b = pop();
3189 c = pop();
3190 d = pop();
3191 push( b );
3192 push( a );
3193 push( d );
3194 push( c );
3195 push( b );
3196 push( a );
3197 break;
3198
3199 case Bytecodes::_arraylength: {
3200 // Must do null-check with value on expression stack
3201 Node *ary = null_check(peek(), T_ARRAY);
3202 // Compile-time detect of null-exception?
3203 if (stopped()) return;
3204 a = pop();
3205 push(load_array_length(a));
3206 break;
3207 }
3208
3209 case Bytecodes::_baload: array_load(T_BYTE); break;
3210 case Bytecodes::_caload: array_load(T_CHAR); break;
3211 case Bytecodes::_iaload: array_load(T_INT); break;
3212 case Bytecodes::_saload: array_load(T_SHORT); break;
3213 case Bytecodes::_faload: array_load(T_FLOAT); break;
3214 case Bytecodes::_aaload: array_load(T_OBJECT); break;
3215 case Bytecodes::_laload: array_load(T_LONG); break;
3216 case Bytecodes::_daload: array_load(T_DOUBLE); break;
3217 case Bytecodes::_bastore: array_store(T_BYTE); break;
3218 case Bytecodes::_castore: array_store(T_CHAR); break;
3219 case Bytecodes::_iastore: array_store(T_INT); break;
3220 case Bytecodes::_sastore: array_store(T_SHORT); break;
3221 case Bytecodes::_fastore: array_store(T_FLOAT); break;
3222 case Bytecodes::_aastore: array_store(T_OBJECT); break;
3223 case Bytecodes::_lastore: array_store(T_LONG); break;
3224 case Bytecodes::_dastore: array_store(T_DOUBLE); break;
3225
3226 case Bytecodes::_getfield:
3227 do_getfield();
3228 break;
3229
3230 case Bytecodes::_getstatic:
3231 do_getstatic();
3232 break;
3233
3234 case Bytecodes::_putfield:
3235 do_putfield();
3236 break;
3237
3238 case Bytecodes::_putstatic:
3239 do_putstatic();
3240 break;
3241
3242 case Bytecodes::_irem:
3243 // Must keep both values on the expression-stack during null-check
3244 zero_check_int(peek());
3245 // Compile-time detect of null-exception?
3246 if (stopped()) return;
3247 b = pop();
3248 a = pop();
3249 push(_gvn.transform(new ModINode(control(), a, b)));
3250 break;
3251 case Bytecodes::_idiv:
3252 // Must keep both values on the expression-stack during null-check
3253 zero_check_int(peek());
3254 // Compile-time detect of null-exception?
3255 if (stopped()) return;
3256 b = pop();
3257 a = pop();
3258 push( _gvn.transform( new DivINode(control(),a,b) ) );
3259 break;
3260 case Bytecodes::_imul:
3261 b = pop(); a = pop();
3262 push( _gvn.transform( new MulINode(a,b) ) );
3263 break;
3264 case Bytecodes::_iadd:
3265 b = pop(); a = pop();
3266 push( _gvn.transform( new AddINode(a,b) ) );
3267 break;
3268 case Bytecodes::_ineg:
3269 a = pop();
3270 push( _gvn.transform( new SubINode(_gvn.intcon(0),a)) );
3271 break;
3272 case Bytecodes::_isub:
3273 b = pop(); a = pop();
3274 push( _gvn.transform( new SubINode(a,b) ) );
3275 break;
3276 case Bytecodes::_iand:
3277 b = pop(); a = pop();
3278 push( _gvn.transform( new AndINode(a,b) ) );
3279 break;
3280 case Bytecodes::_ior:
3281 b = pop(); a = pop();
3282 push( _gvn.transform( new OrINode(a,b) ) );
3283 break;
3284 case Bytecodes::_ixor:
3285 b = pop(); a = pop();
3286 push( _gvn.transform( new XorINode(a,b) ) );
3287 break;
3288 case Bytecodes::_ishl:
3289 b = pop(); a = pop();
3290 push( _gvn.transform( new LShiftINode(a,b) ) );
3291 break;
3292 case Bytecodes::_ishr:
3293 b = pop(); a = pop();
3294 push( _gvn.transform( new RShiftINode(a,b) ) );
3295 break;
3296 case Bytecodes::_iushr:
3297 b = pop(); a = pop();
3298 push( _gvn.transform( new URShiftINode(a,b) ) );
3299 break;
3300
3301 case Bytecodes::_fneg:
3302 a = pop();
3303 b = _gvn.transform(new NegFNode (a));
3304 push(b);
3305 break;
3306
3307 case Bytecodes::_fsub:
3308 b = pop();
3309 a = pop();
3310 c = _gvn.transform( new SubFNode(a,b) );
3311 push(c);
3312 break;
3313
3314 case Bytecodes::_fadd:
3315 b = pop();
3316 a = pop();
3317 c = _gvn.transform( new AddFNode(a,b) );
3318 push(c);
3319 break;
3320
3321 case Bytecodes::_fmul:
3322 b = pop();
3323 a = pop();
3324 c = _gvn.transform( new MulFNode(a,b) );
3325 push(c);
3326 break;
3327
3328 case Bytecodes::_fdiv:
3329 b = pop();
3330 a = pop();
3331 c = _gvn.transform( new DivFNode(nullptr,a,b) );
3332 push(c);
3333 break;
3334
3335 case Bytecodes::_frem:
3336 // Generate a ModF node.
3337 b = pop();
3338 a = pop();
3339 push(floating_point_mod(a, b, BasicType::T_FLOAT));
3340 break;
3341
3342 case Bytecodes::_fcmpl:
3343 b = pop();
3344 a = pop();
3345 c = _gvn.transform( new CmpF3Node( a, b));
3346 push(c);
3347 break;
3348 case Bytecodes::_fcmpg:
3349 b = pop();
3350 a = pop();
3351
3352 // Same as fcmpl but need to flip the unordered case. Swap the inputs,
3353 // which negates the result sign except for unordered. Flip the unordered
3354 // as well by using CmpF3 which implements unordered-lesser instead of
3355 // unordered-greater semantics. Finally, commute the result bits. Result
3356 // is same as using a CmpF3Greater except we did it with CmpF3 alone.
3357 c = _gvn.transform( new CmpF3Node( b, a));
3358 c = _gvn.transform( new SubINode(_gvn.intcon(0),c) );
3359 push(c);
3360 break;
3361
3362 case Bytecodes::_f2i:
3363 a = pop();
3364 push(_gvn.transform(new ConvF2INode(a)));
3365 break;
3366
3367 case Bytecodes::_d2i:
3368 a = pop_pair();
3369 b = _gvn.transform(new ConvD2INode(a));
3370 push( b );
3371 break;
3372
3373 case Bytecodes::_f2d:
3374 a = pop();
3375 b = _gvn.transform( new ConvF2DNode(a));
3376 push_pair( b );
3377 break;
3378
3379 case Bytecodes::_d2f:
3380 a = pop_pair();
3381 b = _gvn.transform( new ConvD2FNode(a));
3382 push( b );
3383 break;
3384
3385 case Bytecodes::_l2f:
3386 if (Matcher::convL2FSupported()) {
3387 a = pop_pair();
3388 b = _gvn.transform( new ConvL2FNode(a));
3389 push(b);
3390 } else {
3391 l2f();
3392 }
3393 break;
3394
3395 case Bytecodes::_l2d:
3396 a = pop_pair();
3397 b = _gvn.transform( new ConvL2DNode(a));
3398 push_pair(b);
3399 break;
3400
3401 case Bytecodes::_f2l:
3402 a = pop();
3403 b = _gvn.transform( new ConvF2LNode(a));
3404 push_pair(b);
3405 break;
3406
3407 case Bytecodes::_d2l:
3408 a = pop_pair();
3409 b = _gvn.transform( new ConvD2LNode(a));
3410 push_pair(b);
3411 break;
3412
3413 case Bytecodes::_dsub:
3414 b = pop_pair();
3415 a = pop_pair();
3416 c = _gvn.transform( new SubDNode(a,b) );
3417 push_pair(c);
3418 break;
3419
3420 case Bytecodes::_dadd:
3421 b = pop_pair();
3422 a = pop_pair();
3423 c = _gvn.transform( new AddDNode(a,b) );
3424 push_pair(c);
3425 break;
3426
3427 case Bytecodes::_dmul:
3428 b = pop_pair();
3429 a = pop_pair();
3430 c = _gvn.transform( new MulDNode(a,b) );
3431 push_pair(c);
3432 break;
3433
3434 case Bytecodes::_ddiv:
3435 b = pop_pair();
3436 a = pop_pair();
3437 c = _gvn.transform( new DivDNode(nullptr,a,b) );
3438 push_pair(c);
3439 break;
3440
3441 case Bytecodes::_dneg:
3442 a = pop_pair();
3443 b = _gvn.transform(new NegDNode (a));
3444 push_pair(b);
3445 break;
3446
3447 case Bytecodes::_drem:
3448 // Generate a ModD node.
3449 b = pop_pair();
3450 a = pop_pair();
3451 push_pair(floating_point_mod(a, b, BasicType::T_DOUBLE));
3452 break;
3453
3454 case Bytecodes::_dcmpl:
3455 b = pop_pair();
3456 a = pop_pair();
3457 c = _gvn.transform( new CmpD3Node( a, b));
3458 push(c);
3459 break;
3460
3461 case Bytecodes::_dcmpg:
3462 b = pop_pair();
3463 a = pop_pair();
3464 // Same as dcmpl but need to flip the unordered case.
3465 // Commute the inputs, which negates the result sign except for unordered.
3466 // Flip the unordered as well by using CmpD3 which implements
3467 // unordered-lesser instead of unordered-greater semantics.
3468 // Finally, negate the result bits. Result is same as using a
3469 // CmpD3Greater except we did it with CmpD3 alone.
3470 c = _gvn.transform( new CmpD3Node( b, a));
3471 c = _gvn.transform( new SubINode(_gvn.intcon(0),c) );
3472 push(c);
3473 break;
3474
3475
3476 // Note for longs -> lo word is on TOS, hi word is on TOS - 1
3477 case Bytecodes::_land:
3478 b = pop_pair();
3479 a = pop_pair();
3480 c = _gvn.transform( new AndLNode(a,b) );
3481 push_pair(c);
3482 break;
3483 case Bytecodes::_lor:
3484 b = pop_pair();
3485 a = pop_pair();
3486 c = _gvn.transform( new OrLNode(a,b) );
3487 push_pair(c);
3488 break;
3489 case Bytecodes::_lxor:
3490 b = pop_pair();
3491 a = pop_pair();
3492 c = _gvn.transform( new XorLNode(a,b) );
3493 push_pair(c);
3494 break;
3495
3496 case Bytecodes::_lshl:
3497 b = pop(); // the shift count
3498 a = pop_pair(); // value to be shifted
3499 c = _gvn.transform( new LShiftLNode(a,b) );
3500 push_pair(c);
3501 break;
3502 case Bytecodes::_lshr:
3503 b = pop(); // the shift count
3504 a = pop_pair(); // value to be shifted
3505 c = _gvn.transform( new RShiftLNode(a,b) );
3506 push_pair(c);
3507 break;
3508 case Bytecodes::_lushr:
3509 b = pop(); // the shift count
3510 a = pop_pair(); // value to be shifted
3511 c = _gvn.transform( new URShiftLNode(a,b) );
3512 push_pair(c);
3513 break;
3514 case Bytecodes::_lmul:
3515 b = pop_pair();
3516 a = pop_pair();
3517 c = _gvn.transform( new MulLNode(a,b) );
3518 push_pair(c);
3519 break;
3520
3521 case Bytecodes::_lrem:
3522 // Must keep both values on the expression-stack during null-check
3523 assert(peek(0) == top(), "long word order");
3524 zero_check_long(peek(1));
3525 // Compile-time detect of null-exception?
3526 if (stopped()) return;
3527 b = pop_pair();
3528 a = pop_pair();
3529 c = _gvn.transform( new ModLNode(control(),a,b) );
3530 push_pair(c);
3531 break;
3532
3533 case Bytecodes::_ldiv:
3534 // Must keep both values on the expression-stack during null-check
3535 assert(peek(0) == top(), "long word order");
3536 zero_check_long(peek(1));
3537 // Compile-time detect of null-exception?
3538 if (stopped()) return;
3539 b = pop_pair();
3540 a = pop_pair();
3541 c = _gvn.transform( new DivLNode(control(),a,b) );
3542 push_pair(c);
3543 break;
3544
3545 case Bytecodes::_ladd:
3546 b = pop_pair();
3547 a = pop_pair();
3548 c = _gvn.transform( new AddLNode(a,b) );
3549 push_pair(c);
3550 break;
3551 case Bytecodes::_lsub:
3552 b = pop_pair();
3553 a = pop_pair();
3554 c = _gvn.transform( new SubLNode(a,b) );
3555 push_pair(c);
3556 break;
3557 case Bytecodes::_lcmp:
3558 // Safepoints are now inserted _before_ branches. The long-compare
3559 // bytecode painfully produces a 3-way value (-1,0,+1) which requires a
3560 // slew of control flow. These are usually followed by a CmpI vs zero and
3561 // a branch; this pattern then optimizes to the obvious long-compare and
3562 // branch. However, if the branch is backwards there's a Safepoint
3563 // inserted. The inserted Safepoint captures the JVM state at the
3564 // pre-branch point, i.e. it captures the 3-way value. Thus if a
3565 // long-compare is used to control a loop the debug info will force
3566 // computation of the 3-way value, even though the generated code uses a
3567 // long-compare and branch. We try to rectify the situation by inserting
3568 // a SafePoint here and have it dominate and kill the safepoint added at a
3569 // following backwards branch. At this point the JVM state merely holds 2
3570 // longs but not the 3-way value.
3571 switch (iter().next_bc()) {
3572 case Bytecodes::_ifgt:
3573 case Bytecodes::_iflt:
3574 case Bytecodes::_ifge:
3575 case Bytecodes::_ifle:
3576 case Bytecodes::_ifne:
3577 case Bytecodes::_ifeq:
3578 // If this is a backwards branch in the bytecodes, add Safepoint
3579 maybe_add_safepoint(iter().next_get_dest());
3580 default:
3581 break;
3582 }
3583 b = pop_pair();
3584 a = pop_pair();
3585 c = _gvn.transform( new CmpL3Node( a, b ));
3586 push(c);
3587 break;
3588
3589 case Bytecodes::_lneg:
3590 a = pop_pair();
3591 b = _gvn.transform( new SubLNode(longcon(0),a));
3592 push_pair(b);
3593 break;
3594 case Bytecodes::_l2i:
3595 a = pop_pair();
3596 push( _gvn.transform( new ConvL2INode(a)));
3597 break;
3598 case Bytecodes::_i2l:
3599 a = pop();
3600 b = _gvn.transform( new ConvI2LNode(a));
3601 push_pair(b);
3602 break;
3603 case Bytecodes::_i2b:
3604 // Sign extend
3605 a = pop();
3606 a = Compile::narrow_value(T_BYTE, a, nullptr, &_gvn, true);
3607 push(a);
3608 break;
3609 case Bytecodes::_i2s:
3610 a = pop();
3611 a = Compile::narrow_value(T_SHORT, a, nullptr, &_gvn, true);
3612 push(a);
3613 break;
3614 case Bytecodes::_i2c:
3615 a = pop();
3616 a = Compile::narrow_value(T_CHAR, a, nullptr, &_gvn, true);
3617 push(a);
3618 break;
3619
3620 case Bytecodes::_i2f:
3621 a = pop();
3622 b = _gvn.transform( new ConvI2FNode(a) ) ;
3623 push(b);
3624 break;
3625
3626 case Bytecodes::_i2d:
3627 a = pop();
3628 b = _gvn.transform( new ConvI2DNode(a));
3629 push_pair(b);
3630 break;
3631
3632 case Bytecodes::_iinc: // Increment local
3633 i = iter().get_index(); // Get local index
3634 set_local( i, _gvn.transform( new AddINode( _gvn.intcon(iter().get_iinc_con()), local(i) ) ) );
3635 break;
3636
3637 // Exit points of synchronized methods must have an unlock node
3638 case Bytecodes::_return:
3639 return_current(nullptr);
3640 break;
3641
3642 case Bytecodes::_ireturn:
3643 case Bytecodes::_areturn:
3644 case Bytecodes::_freturn:
3645 return_current(pop());
3646 break;
3647 case Bytecodes::_lreturn:
3648 case Bytecodes::_dreturn:
3649 return_current(pop_pair());
3650 break;
3651
3652 case Bytecodes::_athrow:
3653 // null exception oop throws null pointer exception
3654 null_check(peek());
3655 if (stopped()) return;
3656 // Hook the thrown exception directly to subsequent handlers.
3657 if (BailoutToInterpreterForThrows) {
3658 // Keep method interpreted from now on.
3659 uncommon_trap(Deoptimization::Reason_unhandled,
3660 Deoptimization::Action_make_not_compilable);
3661 return;
3662 }
3663 if (env()->jvmti_can_post_on_exceptions()) {
3664 // check if we must post exception events, take uncommon trap if so (with must_throw = false)
3665 uncommon_trap_if_should_post_on_exceptions(Deoptimization::Reason_unhandled, false);
3666 }
3667 // Here if either can_post_on_exceptions or should_post_on_exceptions is false
3668 add_exception_state(make_exception_state(peek()));
3669 break;
3670
3671 case Bytecodes::_goto: // fall through
3672 case Bytecodes::_goto_w: {
3673 int target_bci = (bc() == Bytecodes::_goto) ? iter().get_dest() : iter().get_far_dest();
3674
3675 // If this is a backwards branch in the bytecodes, add Safepoint
3676 maybe_add_safepoint(target_bci);
3677
3678 // Merge the current control into the target basic block
3679 merge(target_bci);
3680
3681 // See if we can get some profile data and hand it off to the next block
3682 Block *target_block = block()->successor_for_bci(target_bci);
3683 if (target_block->pred_count() != 1) break;
3684 ciMethodData* methodData = method()->method_data();
3685 if (!methodData->is_mature()) break;
3686 ciProfileData* data = methodData->bci_to_data(bci());
3687 assert(data != nullptr && data->is_JumpData(), "need JumpData for taken branch");
3688 int taken = ((ciJumpData*)data)->taken();
3689 taken = method()->scale_count(taken);
3690 target_block->set_count(taken);
3691 break;
3692 }
3693
3694 case Bytecodes::_ifnull: btest = BoolTest::eq; goto handle_if_null;
3695 case Bytecodes::_ifnonnull: btest = BoolTest::ne; goto handle_if_null;
3696 handle_if_null:
3697 // If this is a backwards branch in the bytecodes, add Safepoint
3698 maybe_add_safepoint(iter().get_dest());
3699 a = null();
3700 b = pop();
3701 if (b->is_InlineType()) {
3702 // Null checking a scalarized but nullable inline type. Check the null marker
3703 // input instead of the oop input to avoid keeping buffer allocations alive
3704 c = _gvn.transform(new CmpINode(b->as_InlineType()->get_null_marker(), zerocon(T_INT)));
3705 } else {
3706 if (!_gvn.type(b)->speculative_maybe_null() &&
3707 !too_many_traps(Deoptimization::Reason_speculate_null_check)) {
3708 inc_sp(1);
3709 Node* null_ctl = top();
3710 b = null_check_oop(b, &null_ctl, true, true, true);
3711 assert(null_ctl->is_top(), "no null control here");
3712 dec_sp(1);
3713 } else if (_gvn.type(b)->speculative_always_null() &&
3714 !too_many_traps(Deoptimization::Reason_speculate_null_assert)) {
3715 inc_sp(1);
3716 b = null_assert(b);
3717 dec_sp(1);
3718 }
3719 c = _gvn.transform( new CmpPNode(b, a) );
3720 }
3721 do_ifnull(btest, c);
3722 break;
3723
3724 case Bytecodes::_if_acmpeq: btest = BoolTest::eq; goto handle_if_acmp;
3725 case Bytecodes::_if_acmpne: btest = BoolTest::ne; goto handle_if_acmp;
3726 handle_if_acmp:
3727 // If this is a backwards branch in the bytecodes, add Safepoint
3728 maybe_add_safepoint(iter().get_dest());
3729 a = pop();
3730 b = pop();
3731 do_acmp(btest, b, a);
3732 break;
3733
3734 case Bytecodes::_ifeq: btest = BoolTest::eq; goto handle_ifxx;
3735 case Bytecodes::_ifne: btest = BoolTest::ne; goto handle_ifxx;
3736 case Bytecodes::_iflt: btest = BoolTest::lt; goto handle_ifxx;
3737 case Bytecodes::_ifle: btest = BoolTest::le; goto handle_ifxx;
3738 case Bytecodes::_ifgt: btest = BoolTest::gt; goto handle_ifxx;
3739 case Bytecodes::_ifge: btest = BoolTest::ge; goto handle_ifxx;
3740 handle_ifxx:
3741 // If this is a backwards branch in the bytecodes, add Safepoint
3742 maybe_add_safepoint(iter().get_dest());
3743 a = _gvn.intcon(0);
3744 b = pop();
3745 c = _gvn.transform( new CmpINode(b, a) );
3746 do_if(btest, c);
3747 break;
3748
3749 case Bytecodes::_if_icmpeq: btest = BoolTest::eq; goto handle_if_icmp;
3750 case Bytecodes::_if_icmpne: btest = BoolTest::ne; goto handle_if_icmp;
3751 case Bytecodes::_if_icmplt: btest = BoolTest::lt; goto handle_if_icmp;
3752 case Bytecodes::_if_icmple: btest = BoolTest::le; goto handle_if_icmp;
3753 case Bytecodes::_if_icmpgt: btest = BoolTest::gt; goto handle_if_icmp;
3754 case Bytecodes::_if_icmpge: btest = BoolTest::ge; goto handle_if_icmp;
3755 handle_if_icmp:
3756 // If this is a backwards branch in the bytecodes, add Safepoint
3757 maybe_add_safepoint(iter().get_dest());
3758 a = pop();
3759 b = pop();
3760 c = _gvn.transform( new CmpINode( b, a ) );
3761 do_if(btest, c);
3762 break;
3763
3764 case Bytecodes::_tableswitch:
3765 do_tableswitch();
3766 break;
3767
3768 case Bytecodes::_lookupswitch:
3769 do_lookupswitch();
3770 break;
3771
3772 case Bytecodes::_invokestatic:
3773 case Bytecodes::_invokedynamic:
3774 case Bytecodes::_invokespecial:
3775 case Bytecodes::_invokevirtual:
3776 case Bytecodes::_invokeinterface:
3777 do_call();
3778 break;
3779 case Bytecodes::_checkcast:
3780 do_checkcast();
3781 break;
3782 case Bytecodes::_instanceof:
3783 do_instanceof();
3784 break;
3785 case Bytecodes::_anewarray:
3786 do_newarray();
3787 break;
3788 case Bytecodes::_newarray:
3789 do_newarray((BasicType)iter().get_index());
3790 break;
3791 case Bytecodes::_multianewarray:
3792 do_multianewarray();
3793 break;
3794 case Bytecodes::_new:
3795 do_new();
3796 break;
3797
3798 case Bytecodes::_jsr:
3799 case Bytecodes::_jsr_w:
3800 do_jsr();
3801 break;
3802
3803 case Bytecodes::_ret:
3804 do_ret();
3805 break;
3806
3807
3808 case Bytecodes::_monitorenter:
3809 do_monitor_enter();
3810 break;
3811
3812 case Bytecodes::_monitorexit:
3813 do_monitor_exit();
3814 break;
3815
3816 case Bytecodes::_breakpoint:
3817 // Breakpoint set concurrently to compile
3818 // %%% use an uncommon trap?
3819 C->record_failure("breakpoint in method");
3820 return;
3821
3822 default:
3823 #ifndef PRODUCT
3824 map()->dump(99);
3825 #endif
3826 tty->print("\nUnhandled bytecode %s\n", Bytecodes::name(bc()) );
3827 ShouldNotReachHere();
3828 }
3829
3830 #ifndef PRODUCT
3831 if (failing()) { return; }
3832 constexpr int perBytecode = 6;
3833 if (C->should_print_igv(perBytecode)) {
3834 IdealGraphPrinter* printer = C->igv_printer();
3835 char buffer[256];
3836 jio_snprintf(buffer, sizeof(buffer), "Bytecode %d: %s", bci(), Bytecodes::name(bc()));
3837 bool old = printer->traverse_outs();
3838 printer->set_traverse_outs(true);
3839 printer->set_parse(this);
3840 printer->print_graph(buffer);
3841 printer->set_traverse_outs(old);
3842 printer->set_parse(nullptr);
3843 }
3844 #endif
3845 }