1 /*
2 * Copyright (c) 1998, 2024, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "ci/ciMethodData.hpp"
27 #include "ci/ciSymbols.hpp"
28 #include "classfile/vmSymbols.hpp"
29 #include "compiler/compileLog.hpp"
30 #include "interpreter/linkResolver.hpp"
31 #include "jvm_io.h"
32 #include "memory/resourceArea.hpp"
33 #include "memory/universe.hpp"
34 #include "oops/oop.inline.hpp"
35 #include "opto/addnode.hpp"
36 #include "opto/castnode.hpp"
37 #include "opto/convertnode.hpp"
38 #include "opto/divnode.hpp"
39 #include "opto/idealGraphPrinter.hpp"
40 #include "opto/idealKit.hpp"
41 #include "opto/inlinetypenode.hpp"
42 #include "opto/matcher.hpp"
43 #include "opto/memnode.hpp"
44 #include "opto/mulnode.hpp"
45 #include "opto/opaquenode.hpp"
46 #include "opto/parse.hpp"
47 #include "opto/runtime.hpp"
48 #include "runtime/deoptimization.hpp"
49 #include "runtime/sharedRuntime.hpp"
50
51 #ifndef PRODUCT
52 extern uint explicit_null_checks_inserted,
53 explicit_null_checks_elided;
54 #endif
55
56 Node* Parse::record_profile_for_speculation_at_array_load(Node* ld) {
57 // Feed unused profile data to type speculation
58 if (UseTypeSpeculation && UseArrayLoadStoreProfile) {
59 ciKlass* array_type = nullptr;
60 ciKlass* element_type = nullptr;
61 ProfilePtrKind element_ptr = ProfileMaybeNull;
62 bool flat_array = true;
63 bool null_free_array = true;
64 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
65 if (element_type != nullptr || element_ptr != ProfileMaybeNull) {
66 ld = record_profile_for_speculation(ld, element_type, element_ptr);
67 }
68 }
69 return ld;
70 }
71
72
73 //---------------------------------array_load----------------------------------
74 void Parse::array_load(BasicType bt) {
75 const Type* elemtype = Type::TOP;
76 Node* adr = array_addressing(bt, 0, elemtype);
77 if (stopped()) return; // guaranteed null or range check
78
79 Node* idx = pop();
80 Node* ary = pop();
81
82 // Handle inline type arrays
83 const TypeOopPtr* elemptr = elemtype->make_oopptr();
84 const TypeAryPtr* ary_t = _gvn.type(ary)->is_aryptr();
85 if (ary_t->is_flat()) {
86 // Load from flat inline type array
87 Node* vt = InlineTypeNode::make_from_flat(this, elemtype->inline_klass(), ary, adr);
88 push(vt);
89 return;
90 } else if (ary_t->is_null_free()) {
91 // Load from non-flat inline type array (elements can never be null)
92 bt = T_OBJECT;
93 } else if (!ary_t->is_not_flat()) {
94 // Cannot statically determine if array is a flat array, emit runtime check
95 assert(UseFlatArray && is_reference_type(bt) && elemptr->can_be_inline_type() && !ary_t->is_not_null_free() &&
96 (!elemptr->is_inlinetypeptr() || elemptr->inline_klass()->flat_in_array()), "array can't be flat");
97 IdealKit ideal(this);
98 IdealVariable res(ideal);
99 ideal.declarations_done();
100 ideal.if_then(flat_array_test(ary, /* flat = */ false)); {
101 // non-flat array
102 assert(ideal.ctrl()->in(0)->as_If()->is_flat_array_check(&_gvn), "Should be found");
103 sync_kit(ideal);
104 const TypeAryPtr* adr_type = TypeAryPtr::get_array_body_type(bt);
105 DecoratorSet decorator_set = IN_HEAP | IS_ARRAY | C2_CONTROL_DEPENDENT_LOAD;
106 if (needs_range_check(ary_t->size(), idx)) {
107 // We've emitted a RangeCheck but now insert an additional check between the range check and the actual load.
108 // We cannot pin the load to two separate nodes. Instead, we pin it conservatively here such that it cannot
109 // possibly float above the range check at any point.
110 decorator_set |= C2_UNKNOWN_CONTROL_LOAD;
111 }
112 Node* ld = access_load_at(ary, adr, adr_type, elemptr, bt, decorator_set);
113 if (elemptr->is_inlinetypeptr()) {
114 assert(elemptr->maybe_null(), "null free array should be handled above");
115 ld = InlineTypeNode::make_from_oop(this, ld, elemptr->inline_klass(), false);
116 }
117 ideal.sync_kit(this);
118 ideal.set(res, ld);
119 } ideal.else_(); {
120 // flat array
121 sync_kit(ideal);
122 if (elemptr->is_inlinetypeptr()) {
123 // Element type is known, cast and load from flat representation
124 ciInlineKlass* vk = elemptr->inline_klass();
125 assert(vk->flat_in_array() && elemptr->maybe_null(), "never/always flat - should be optimized");
126 ciArrayKlass* array_klass = ciArrayKlass::make(vk, /* null_free */ true);
127 const TypeAryPtr* arytype = TypeOopPtr::make_from_klass(array_klass)->isa_aryptr();
128 Node* cast = _gvn.transform(new CheckCastPPNode(control(), ary, arytype));
129 Node* casted_adr = array_element_address(cast, idx, T_OBJECT, ary_t->size(), control());
130 // Re-execute flat array load if buffering triggers deoptimization
131 PreserveReexecuteState preexecs(this);
132 jvms()->set_should_reexecute(true);
133 inc_sp(2);
134 Node* vt = InlineTypeNode::make_from_flat(this, vk, cast, casted_adr)->buffer(this, false);
135 ideal.set(res, vt);
136 ideal.sync_kit(this);
137 } else {
138 // Element type is unknown, emit runtime call
139
140 // Below membars keep this access to an unknown flat array correctly
141 // ordered with other unknown and known flat array accesses.
142 insert_mem_bar_volatile(Op_MemBarCPUOrder, C->get_alias_index(TypeAryPtr::INLINES));
143
144 Node* call = nullptr;
145 {
146 // Re-execute flat array load if runtime call triggers deoptimization
147 PreserveReexecuteState preexecs(this);
148 jvms()->set_bci(_bci);
149 jvms()->set_should_reexecute(true);
150 inc_sp(2);
151 kill_dead_locals();
152 call = make_runtime_call(RC_NO_LEAF | RC_NO_IO,
153 OptoRuntime::load_unknown_inline_Type(),
154 OptoRuntime::load_unknown_inline_Java(),
155 nullptr, TypeRawPtr::BOTTOM,
156 ary, idx);
157 }
158 make_slow_call_ex(call, env()->Throwable_klass(), false);
159 Node* buffer = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
160
161 insert_mem_bar_volatile(Op_MemBarCPUOrder, C->get_alias_index(TypeAryPtr::INLINES));
162
163 // Keep track of the information that the inline type is in flat arrays
164 const Type* unknown_value = elemptr->is_instptr()->cast_to_flat_in_array();
165 buffer = _gvn.transform(new CheckCastPPNode(control(), buffer, unknown_value));
166
167 ideal.sync_kit(this);
168 ideal.set(res, buffer);
169 }
170 } ideal.end_if();
171 sync_kit(ideal);
172 Node* ld = _gvn.transform(ideal.value(res));
173 ld = record_profile_for_speculation_at_array_load(ld);
174 push_node(bt, ld);
175 return;
176 }
177
178 if (elemtype == TypeInt::BOOL) {
179 bt = T_BOOLEAN;
180 }
181 const TypeAryPtr* adr_type = TypeAryPtr::get_array_body_type(bt);
182 Node* ld = access_load_at(ary, adr, adr_type, elemtype, bt,
183 IN_HEAP | IS_ARRAY | C2_CONTROL_DEPENDENT_LOAD);
184 ld = record_profile_for_speculation_at_array_load(ld);
185 // Loading an inline type from a non-flat array
186 if (elemptr != nullptr && elemptr->is_inlinetypeptr()) {
187 assert(!ary_t->is_null_free() || !elemptr->maybe_null(), "inline type array elements should never be null");
188 ld = InlineTypeNode::make_from_oop(this, ld, elemptr->inline_klass(), !elemptr->maybe_null());
189 }
190 push_node(bt, ld);
191 }
192
193
194 //--------------------------------array_store----------------------------------
195 void Parse::array_store(BasicType bt) {
196 const Type* elemtype = Type::TOP;
197 Node* adr = array_addressing(bt, type2size[bt], elemtype);
198 if (stopped()) return; // guaranteed null or range check
199 Node* cast_val = nullptr;
200 if (bt == T_OBJECT) {
201 cast_val = array_store_check(adr, elemtype);
202 if (stopped()) return;
203 }
204 Node* val = pop_node(bt); // Value to store
205 Node* idx = pop(); // Index in the array
206 Node* ary = pop(); // The array itself
207
208 const TypeAryPtr* ary_t = _gvn.type(ary)->is_aryptr();
209 const TypeAryPtr* adr_type = TypeAryPtr::get_array_body_type(bt);
210
211 if (elemtype == TypeInt::BOOL) {
212 bt = T_BOOLEAN;
213 } else if (bt == T_OBJECT) {
214 elemtype = elemtype->make_oopptr();
215 const Type* tval = _gvn.type(cast_val);
216 // Based on the value to be stored, try to determine if the array is not null-free and/or not flat.
217 // This is only legal for non-null stores because the array_store_check always passes for null, even
218 // if the array is null-free. Null stores are handled in GraphKit::gen_inline_array_null_guard().
219 bool not_null_free = !tval->maybe_null() && !tval->is_oopptr()->can_be_inline_type();
220 bool not_flat = not_null_free || (tval->is_inlinetypeptr() && !tval->inline_klass()->flat_in_array());
221 if (!ary_t->is_not_null_free() && not_null_free) {
222 // Storing a non-inline type, mark array as not null-free (-> not flat).
223 ary_t = ary_t->cast_to_not_null_free();
224 Node* cast = _gvn.transform(new CheckCastPPNode(control(), ary, ary_t));
225 replace_in_map(ary, cast);
226 ary = cast;
227 } else if (!ary_t->is_not_flat() && not_flat) {
228 // Storing to a non-flat array, mark array as not flat.
229 ary_t = ary_t->cast_to_not_flat();
230 Node* cast = _gvn.transform(new CheckCastPPNode(control(), ary, ary_t));
231 replace_in_map(ary, cast);
232 ary = cast;
233 }
234
235 if (ary_t->is_flat()) {
236 // Store to flat inline type array
237 assert(!tval->maybe_null(), "should be guaranteed by array store check");
238 // Re-execute flat array store if buffering triggers deoptimization
239 PreserveReexecuteState preexecs(this);
240 inc_sp(3);
241 jvms()->set_should_reexecute(true);
242 cast_val->as_InlineType()->store_flat(this, ary, adr, nullptr, 0, MO_UNORDERED | IN_HEAP | IS_ARRAY);
243 return;
244 } else if (ary_t->is_null_free()) {
245 // Store to non-flat inline type array (elements can never be null)
246 assert(!tval->maybe_null(), "should be guaranteed by array store check");
247 if (elemtype->inline_klass()->is_empty()) {
248 // Ignore empty inline stores, array is already initialized.
249 return;
250 }
251 } else if (!ary_t->is_not_flat() && (tval != TypePtr::NULL_PTR || StressReflectiveCode)) {
252 // Array might be a flat array, emit runtime checks (for nullptr, a simple inline_array_null_guard is sufficient).
253 assert(UseFlatArray && !not_flat && elemtype->is_oopptr()->can_be_inline_type() &&
254 !ary_t->klass_is_exact() && !ary_t->is_not_null_free(), "array can't be a flat array");
255 IdealKit ideal(this);
256 ideal.if_then(flat_array_test(ary, /* flat = */ false)); {
257 // non-flat array
258 assert(ideal.ctrl()->in(0)->as_If()->is_flat_array_check(&_gvn), "Should be found");
259 sync_kit(ideal);
260 Node* cast_ary = inline_array_null_guard(ary, cast_val, 3);
261 inc_sp(3);
262 access_store_at(cast_ary, adr, adr_type, cast_val, elemtype, bt, MO_UNORDERED | IN_HEAP | IS_ARRAY, false);
263 dec_sp(3);
264 ideal.sync_kit(this);
265 } ideal.else_(); {
266 sync_kit(ideal);
267 // flat array
268 Node* null_ctl = top();
269 Node* val = null_check_oop(cast_val, &null_ctl);
270 if (null_ctl != top()) {
271 PreserveJVMState pjvms(this);
272 inc_sp(3);
273 set_control(null_ctl);
274 uncommon_trap(Deoptimization::Reason_null_check, Deoptimization::Action_none);
275 dec_sp(3);
276 }
277 // Try to determine the inline klass
278 ciInlineKlass* vk = nullptr;
279 if (tval->is_inlinetypeptr()) {
280 vk = tval->inline_klass();
281 } else if (elemtype->is_inlinetypeptr()) {
282 vk = elemtype->inline_klass();
283 }
284 Node* casted_ary = ary;
285 if (vk != nullptr && !stopped()) {
286 // Element type is known, cast and store to flat representation
287 assert(vk->flat_in_array() && elemtype->maybe_null(), "never/always flat - should be optimized");
288 ciArrayKlass* array_klass = ciArrayKlass::make(vk, /* null_free */ true);
289 const TypeAryPtr* arytype = TypeOopPtr::make_from_klass(array_klass)->isa_aryptr();
290 casted_ary = _gvn.transform(new CheckCastPPNode(control(), casted_ary, arytype));
291 Node* casted_adr = array_element_address(casted_ary, idx, T_OBJECT, arytype->size(), control());
292 if (!val->is_InlineType()) {
293 assert(!gvn().type(val)->maybe_null(), "inline type array elements should never be null");
294 val = InlineTypeNode::make_from_oop(this, val, vk);
295 }
296 // Re-execute flat array store if buffering triggers deoptimization
297 PreserveReexecuteState preexecs(this);
298 inc_sp(3);
299 jvms()->set_should_reexecute(true);
300 val->as_InlineType()->store_flat(this, casted_ary, casted_adr, nullptr, 0, MO_UNORDERED | IN_HEAP | IS_ARRAY);
301 } else if (!stopped()) {
302 // Element type is unknown, emit runtime call
303
304 // Below membars keep this access to an unknown flat array correctly
305 // ordered with other unknown and known flat array accesses.
306 insert_mem_bar_volatile(Op_MemBarCPUOrder, C->get_alias_index(TypeAryPtr::INLINES));
307
308 make_runtime_call(RC_LEAF,
309 OptoRuntime::store_unknown_inline_Type(),
310 CAST_FROM_FN_PTR(address, OptoRuntime::store_unknown_inline_C),
311 "store_unknown_inline", TypeRawPtr::BOTTOM,
312 val, casted_ary, idx);
313
314 insert_mem_bar_volatile(Op_MemBarCPUOrder, C->get_alias_index(TypeAryPtr::INLINES));
315 }
316 ideal.sync_kit(this);
317 }
318 ideal.end_if();
319 sync_kit(ideal);
320 return;
321 } else if (!ary_t->is_not_null_free()) {
322 // Array is not flat but may be null free
323 assert(elemtype->is_oopptr()->can_be_inline_type(), "array can't be null-free");
324 ary = inline_array_null_guard(ary, cast_val, 3, true);
325 }
326 }
327 inc_sp(3);
328 access_store_at(ary, adr, adr_type, val, elemtype, bt, MO_UNORDERED | IN_HEAP | IS_ARRAY);
329 dec_sp(3);
330 }
331
332
333 //------------------------------array_addressing-------------------------------
334 // Pull array and index from the stack. Compute pointer-to-element.
335 Node* Parse::array_addressing(BasicType type, int vals, const Type*& elemtype) {
336 Node *idx = peek(0+vals); // Get from stack without popping
337 Node *ary = peek(1+vals); // in case of exception
338
339 // Null check the array base, with correct stack contents
340 ary = null_check(ary, T_ARRAY);
341 // Compile-time detect of null-exception?
342 if (stopped()) return top();
343
344 const TypeAryPtr* arytype = _gvn.type(ary)->is_aryptr();
345 const TypeInt* sizetype = arytype->size();
346 elemtype = arytype->elem();
347
348 if (UseUniqueSubclasses) {
349 const Type* el = elemtype->make_ptr();
350 if (el && el->isa_instptr()) {
351 const TypeInstPtr* toop = el->is_instptr();
352 if (toop->instance_klass()->unique_concrete_subklass()) {
353 // If we load from "AbstractClass[]" we must see "ConcreteSubClass".
354 const Type* subklass = Type::get_const_type(toop->instance_klass());
355 elemtype = subklass->join_speculative(el);
356 }
357 }
358 }
359
360 if (!arytype->is_loaded()) {
361 // Only fails for some -Xcomp runs
362 // The class is unloaded. We have to run this bytecode in the interpreter.
363 ciKlass* klass = arytype->unloaded_klass();
364
365 uncommon_trap(Deoptimization::Reason_unloaded,
366 Deoptimization::Action_reinterpret,
367 klass, "!loaded array");
368 return top();
369 }
370
371 ary = create_speculative_inline_type_array_checks(ary, arytype, elemtype);
372
373 if (needs_range_check(sizetype, idx)) {
374 create_range_check(idx, ary, sizetype);
375 } else if (C->log() != nullptr) {
376 C->log()->elem("observe that='!need_range_check'");
377 }
378
379 // Check for always knowing you are throwing a range-check exception
380 if (stopped()) return top();
381
382 // Make array address computation control dependent to prevent it
383 // from floating above the range check during loop optimizations.
384 Node* ptr = array_element_address(ary, idx, type, sizetype, control());
385 assert(ptr != top(), "top should go hand-in-hand with stopped");
386
387 return ptr;
388 }
389
390 // 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
391 // be greater or equal the smallest possible array size (i.e. out-of-bounds).
392 bool Parse::needs_range_check(const TypeInt* size_type, const Node* index) const {
393 const TypeInt* index_type = _gvn.type(index)->is_int();
394 return index_type->_hi >= size_type->_lo || index_type->_lo < 0;
395 }
396
397 void Parse::create_range_check(Node* idx, Node* ary, const TypeInt* sizetype) {
398 Node* tst;
399 if (sizetype->_hi <= 0) {
400 // The greatest array bound is negative, so we can conclude that we're
401 // compiling unreachable code, but the unsigned compare trick used below
402 // only works with non-negative lengths. Instead, hack "tst" to be zero so
403 // the uncommon_trap path will always be taken.
404 tst = _gvn.intcon(0);
405 } else {
406 // Range is constant in array-oop, so we can use the original state of mem
407 Node* len = load_array_length(ary);
408
409 // Test length vs index (standard trick using unsigned compare)
410 Node* chk = _gvn.transform(new CmpUNode(idx, len) );
411 BoolTest::mask btest = BoolTest::lt;
412 tst = _gvn.transform(new BoolNode(chk, btest) );
413 }
414 RangeCheckNode* rc = new RangeCheckNode(control(), tst, PROB_MAX, COUNT_UNKNOWN);
415 _gvn.set_type(rc, rc->Value(&_gvn));
416 if (!tst->is_Con()) {
417 record_for_igvn(rc);
418 }
419 set_control(_gvn.transform(new IfTrueNode(rc)));
420 // Branch to failure if out of bounds
421 {
422 PreserveJVMState pjvms(this);
423 set_control(_gvn.transform(new IfFalseNode(rc)));
424 if (C->allow_range_check_smearing()) {
425 // Do not use builtin_throw, since range checks are sometimes
426 // made more stringent by an optimistic transformation.
427 // This creates "tentative" range checks at this point,
428 // which are not guaranteed to throw exceptions.
429 // See IfNode::Ideal, is_range_check, adjust_check.
430 uncommon_trap(Deoptimization::Reason_range_check,
431 Deoptimization::Action_make_not_entrant,
432 nullptr, "range_check");
433 } else {
434 // If we have already recompiled with the range-check-widening
435 // heroic optimization turned off, then we must really be throwing
436 // range check exceptions.
437 builtin_throw(Deoptimization::Reason_range_check);
438 }
439 }
440 }
441
442 // For inline type arrays, we can use the profiling information for array accesses to speculate on the type, flatness,
443 // and null-freeness. We can either prepare the speculative type for later uses or emit explicit speculative checks with
444 // traps now. In the latter case, the speculative type guarantees can avoid additional runtime checks later (e.g.
445 // non-null-free implies non-flat which allows us to remove flatness checks). This makes the graph simpler.
446 Node* Parse::create_speculative_inline_type_array_checks(Node* array, const TypeAryPtr* array_type,
447 const Type*& element_type) {
448 if (!array_type->is_flat() && !array_type->is_not_flat()) {
449 // For arrays that might be flat, speculate that the array has the exact type reported in the profile data such that
450 // we can rely on a fixed memory layout (i.e. either a flat layout or not).
451 array = cast_to_speculative_array_type(array, array_type, element_type);
452 } else if (UseTypeSpeculation && UseArrayLoadStoreProfile) {
453 // Array is known to be either flat or not flat. If possible, update the speculative type by using the profile data
454 // at this bci.
455 array = cast_to_profiled_array_type(array);
456 }
457
458 // Even though the type does not tell us whether we have an inline type array or not, we can still check the profile data
459 // whether we have a non-null-free or non-flat array. Since non-null-free implies non-flat, we check this first.
460 // Speculating on a non-null-free array doesn't help aaload but could be profitable for a subsequent aastore.
461 if (!array_type->is_null_free() && !array_type->is_not_null_free()) {
462 array = speculate_non_null_free_array(array, array_type);
463 }
464
465 if (!array_type->is_flat() && !array_type->is_not_flat()) {
466 array = speculate_non_flat_array(array, array_type);
467 }
468 return array;
469 }
470
471 // Speculate that the array has the exact type reported in the profile data. We emit a trap when this turns out to be
472 // wrong. On the fast path, we add a CheckCastPP to use the exact type.
473 Node* Parse::cast_to_speculative_array_type(Node* const array, const TypeAryPtr*& array_type, const Type*& element_type) {
474 Deoptimization::DeoptReason reason = Deoptimization::Reason_speculate_class_check;
475 ciKlass* speculative_array_type = array_type->speculative_type();
476 if (too_many_traps_or_recompiles(reason) || speculative_array_type == nullptr) {
477 // No speculative type, check profile data at this bci
478 speculative_array_type = nullptr;
479 reason = Deoptimization::Reason_class_check;
480 if (UseArrayLoadStoreProfile && !too_many_traps_or_recompiles(reason)) {
481 ciKlass* profiled_element_type = nullptr;
482 ProfilePtrKind element_ptr = ProfileMaybeNull;
483 bool flat_array = true;
484 bool null_free_array = true;
485 method()->array_access_profiled_type(bci(), speculative_array_type, profiled_element_type, element_ptr, flat_array,
486 null_free_array);
487 }
488 }
489 if (speculative_array_type != nullptr) {
490 // Speculate that this array has the exact type reported by profile data
491 Node* casted_array = nullptr;
492 DEBUG_ONLY(Node* old_control = control();)
493 Node* slow_ctl = type_check_receiver(array, speculative_array_type, 1.0, &casted_array);
494 if (stopped()) {
495 // The check always fails and therefore profile information is incorrect. Don't use it.
496 assert(old_control == slow_ctl, "type check should have been removed");
497 set_control(slow_ctl);
498 } else if (!slow_ctl->is_top()) {
499 { PreserveJVMState pjvms(this);
500 set_control(slow_ctl);
501 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
502 }
503 replace_in_map(array, casted_array);
504 array_type = _gvn.type(casted_array)->is_aryptr();
505 element_type = array_type->elem();
506 return casted_array;
507 }
508 }
509 return array;
510 }
511
512 // Create a CheckCastPP when the speculative type can improve the current type.
513 Node* Parse::cast_to_profiled_array_type(Node* const array) {
514 ciKlass* array_type = nullptr;
515 ciKlass* element_type = nullptr;
516 ProfilePtrKind element_ptr = ProfileMaybeNull;
517 bool flat_array = true;
518 bool null_free_array = true;
519 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
520 if (array_type != nullptr) {
521 return record_profile_for_speculation(array, array_type, ProfileMaybeNull);
522 }
523 return array;
524 }
525
526 // Speculate that the array is non-null-free. This will imply non-flatness. We emit a trap when this turns out to be
527 // wrong. On the fast path, we add a CheckCastPP to use the non-null-free type.
528 Node* Parse::speculate_non_null_free_array(Node* const array, const TypeAryPtr*& array_type) {
529 bool null_free_array = true;
530 Deoptimization::DeoptReason reason = Deoptimization::Reason_none;
531 if (array_type->speculative() != nullptr &&
532 array_type->speculative()->is_aryptr()->is_not_null_free() &&
533 !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_class_check)) {
534 null_free_array = false;
535 reason = Deoptimization::Reason_speculate_class_check;
536 } else if (UseArrayLoadStoreProfile && !too_many_traps_or_recompiles(Deoptimization::Reason_class_check)) {
537 ciKlass* profiled_array_type = nullptr;
538 ciKlass* profiled_element_type = nullptr;
539 ProfilePtrKind element_ptr = ProfileMaybeNull;
540 bool flat_array = true;
541 method()->array_access_profiled_type(bci(), profiled_array_type, profiled_element_type, element_ptr, flat_array,
542 null_free_array);
543 reason = Deoptimization::Reason_class_check;
544 }
545 if (!null_free_array) {
546 { // Deoptimize if null-free array
547 BuildCutout unless(this, null_free_array_test(array, /* null_free = */ false), PROB_MAX);
548 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
549 }
550 assert(!stopped(), "null-free array should have been caught earlier");
551 Node* casted_array = _gvn.transform(new CheckCastPPNode(control(), array, array_type->cast_to_not_null_free()));
552 replace_in_map(array, casted_array);
553 array_type = _gvn.type(casted_array)->is_aryptr();
554 return casted_array;
555 }
556 return array;
557 }
558
559 // Speculate that the array is non-flat. We emit a trap when this turns out to be wrong. On the fast path, we add a
560 // CheckCastPP to use the non-flat type.
561 Node* Parse::speculate_non_flat_array(Node* const array, const TypeAryPtr* const array_type) {
562 bool flat_array = true;
563 Deoptimization::DeoptReason reason = Deoptimization::Reason_none;
564 if (array_type->speculative() != nullptr &&
565 array_type->speculative()->is_aryptr()->is_not_flat() &&
566 !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_class_check)) {
567 flat_array = false;
568 reason = Deoptimization::Reason_speculate_class_check;
569 } else if (UseArrayLoadStoreProfile && !too_many_traps_or_recompiles(reason)) {
570 ciKlass* profiled_array_type = nullptr;
571 ciKlass* profiled_element_type = nullptr;
572 ProfilePtrKind element_ptr = ProfileMaybeNull;
573 bool null_free_array = true;
574 method()->array_access_profiled_type(bci(), profiled_array_type, profiled_element_type, element_ptr, flat_array,
575 null_free_array);
576 reason = Deoptimization::Reason_class_check;
577 }
578 if (!flat_array) {
579 { // Deoptimize if flat array
580 BuildCutout unless(this, flat_array_test(array, /* flat = */ false), PROB_MAX);
581 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
582 }
583 assert(!stopped(), "flat array should have been caught earlier");
584 Node* casted_array = _gvn.transform(new CheckCastPPNode(control(), array, array_type->cast_to_not_flat()));
585 replace_in_map(array, casted_array);
586 return casted_array;
587 }
588 return array;
589 }
590
591 // returns IfNode
592 IfNode* Parse::jump_if_fork_int(Node* a, Node* b, BoolTest::mask mask, float prob, float cnt) {
593 Node *cmp = _gvn.transform(new CmpINode(a, b)); // two cases: shiftcount > 32 and shiftcount <= 32
594 Node *tst = _gvn.transform(new BoolNode(cmp, mask));
595 IfNode *iff = create_and_map_if(control(), tst, prob, cnt);
596 return iff;
597 }
598
599
600 // sentinel value for the target bci to mark never taken branches
601 // (according to profiling)
602 static const int never_reached = INT_MAX;
603
604 //------------------------------helper for tableswitch-------------------------
605 void Parse::jump_if_true_fork(IfNode *iff, int dest_bci_if_true, bool unc) {
606 // True branch, use existing map info
607 { PreserveJVMState pjvms(this);
608 Node *iftrue = _gvn.transform( new IfTrueNode (iff) );
609 set_control( iftrue );
610 if (unc) {
611 repush_if_args();
612 uncommon_trap(Deoptimization::Reason_unstable_if,
613 Deoptimization::Action_reinterpret,
614 nullptr,
615 "taken always");
616 } else {
617 assert(dest_bci_if_true != never_reached, "inconsistent dest");
618 merge_new_path(dest_bci_if_true);
619 }
620 }
621
622 // False branch
623 Node *iffalse = _gvn.transform( new IfFalseNode(iff) );
624 set_control( iffalse );
625 }
626
627 void Parse::jump_if_false_fork(IfNode *iff, int dest_bci_if_true, bool unc) {
628 // True branch, use existing map info
629 { PreserveJVMState pjvms(this);
630 Node *iffalse = _gvn.transform( new IfFalseNode (iff) );
631 set_control( iffalse );
632 if (unc) {
633 repush_if_args();
634 uncommon_trap(Deoptimization::Reason_unstable_if,
635 Deoptimization::Action_reinterpret,
636 nullptr,
637 "taken never");
638 } else {
639 assert(dest_bci_if_true != never_reached, "inconsistent dest");
640 merge_new_path(dest_bci_if_true);
641 }
642 }
643
644 // False branch
645 Node *iftrue = _gvn.transform( new IfTrueNode(iff) );
646 set_control( iftrue );
647 }
648
649 void Parse::jump_if_always_fork(int dest_bci, bool unc) {
650 // False branch, use existing map and control()
651 if (unc) {
652 repush_if_args();
653 uncommon_trap(Deoptimization::Reason_unstable_if,
654 Deoptimization::Action_reinterpret,
655 nullptr,
656 "taken never");
657 } else {
658 assert(dest_bci != never_reached, "inconsistent dest");
659 merge_new_path(dest_bci);
660 }
661 }
662
663
664 extern "C" {
665 static int jint_cmp(const void *i, const void *j) {
666 int a = *(jint *)i;
667 int b = *(jint *)j;
668 return a > b ? 1 : a < b ? -1 : 0;
669 }
670 }
671
672
673 class SwitchRange : public StackObj {
674 // a range of integers coupled with a bci destination
675 jint _lo; // inclusive lower limit
676 jint _hi; // inclusive upper limit
677 int _dest;
678 float _cnt; // how many times this range was hit according to profiling
679
680 public:
681 jint lo() const { return _lo; }
682 jint hi() const { return _hi; }
683 int dest() const { return _dest; }
684 bool is_singleton() const { return _lo == _hi; }
685 float cnt() const { return _cnt; }
686
687 void setRange(jint lo, jint hi, int dest, float cnt) {
688 assert(lo <= hi, "must be a non-empty range");
689 _lo = lo, _hi = hi; _dest = dest; _cnt = cnt;
690 assert(_cnt >= 0, "");
691 }
692 bool adjoinRange(jint lo, jint hi, int dest, float cnt, bool trim_ranges) {
693 assert(lo <= hi, "must be a non-empty range");
694 if (lo == _hi+1) {
695 // see merge_ranges() comment below
696 if (trim_ranges) {
697 if (cnt == 0) {
698 if (_cnt != 0) {
699 return false;
700 }
701 if (dest != _dest) {
702 _dest = never_reached;
703 }
704 } else {
705 if (_cnt == 0) {
706 return false;
707 }
708 if (dest != _dest) {
709 return false;
710 }
711 }
712 } else {
713 if (dest != _dest) {
714 return false;
715 }
716 }
717 _hi = hi;
718 _cnt += cnt;
719 return true;
720 }
721 return false;
722 }
723
724 void set (jint value, int dest, float cnt) {
725 setRange(value, value, dest, cnt);
726 }
727 bool adjoin(jint value, int dest, float cnt, bool trim_ranges) {
728 return adjoinRange(value, value, dest, cnt, trim_ranges);
729 }
730 bool adjoin(SwitchRange& other) {
731 return adjoinRange(other._lo, other._hi, other._dest, other._cnt, false);
732 }
733
734 void print() {
735 if (is_singleton())
736 tty->print(" {%d}=>%d (cnt=%f)", lo(), dest(), cnt());
737 else if (lo() == min_jint)
738 tty->print(" {..%d}=>%d (cnt=%f)", hi(), dest(), cnt());
739 else if (hi() == max_jint)
740 tty->print(" {%d..}=>%d (cnt=%f)", lo(), dest(), cnt());
741 else
742 tty->print(" {%d..%d}=>%d (cnt=%f)", lo(), hi(), dest(), cnt());
743 }
744 };
745
746 // We try to minimize the number of ranges and the size of the taken
747 // ones using profiling data. When ranges are created,
748 // SwitchRange::adjoinRange() only allows 2 adjoining ranges to merge
749 // if both were never hit or both were hit to build longer unreached
750 // ranges. Here, we now merge adjoining ranges with the same
751 // destination and finally set destination of unreached ranges to the
752 // special value never_reached because it can help minimize the number
753 // of tests that are necessary.
754 //
755 // For instance:
756 // [0, 1] to target1 sometimes taken
757 // [1, 2] to target1 never taken
758 // [2, 3] to target2 never taken
759 // would lead to:
760 // [0, 1] to target1 sometimes taken
761 // [1, 3] never taken
762 //
763 // (first 2 ranges to target1 are not merged)
764 static void merge_ranges(SwitchRange* ranges, int& rp) {
765 if (rp == 0) {
766 return;
767 }
768 int shift = 0;
769 for (int j = 0; j < rp; j++) {
770 SwitchRange& r1 = ranges[j-shift];
771 SwitchRange& r2 = ranges[j+1];
772 if (r1.adjoin(r2)) {
773 shift++;
774 } else if (shift > 0) {
775 ranges[j+1-shift] = r2;
776 }
777 }
778 rp -= shift;
779 for (int j = 0; j <= rp; j++) {
780 SwitchRange& r = ranges[j];
781 if (r.cnt() == 0 && r.dest() != never_reached) {
782 r.setRange(r.lo(), r.hi(), never_reached, r.cnt());
783 }
784 }
785 }
786
787 //-------------------------------do_tableswitch--------------------------------
788 void Parse::do_tableswitch() {
789 // Get information about tableswitch
790 int default_dest = iter().get_dest_table(0);
791 jint lo_index = iter().get_int_table(1);
792 jint hi_index = iter().get_int_table(2);
793 int len = hi_index - lo_index + 1;
794
795 if (len < 1) {
796 // If this is a backward branch, add safepoint
797 maybe_add_safepoint(default_dest);
798 pop(); // the effect of the instruction execution on the operand stack
799 merge(default_dest);
800 return;
801 }
802
803 ciMethodData* methodData = method()->method_data();
804 ciMultiBranchData* profile = nullptr;
805 if (methodData->is_mature() && UseSwitchProfiling) {
806 ciProfileData* data = methodData->bci_to_data(bci());
807 if (data != nullptr && data->is_MultiBranchData()) {
808 profile = (ciMultiBranchData*)data;
809 }
810 }
811 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if);
812
813 // generate decision tree, using trichotomy when possible
814 int rnum = len+2;
815 bool makes_backward_branch = (default_dest <= bci());
816 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum);
817 int rp = -1;
818 if (lo_index != min_jint) {
819 float cnt = 1.0F;
820 if (profile != nullptr) {
821 cnt = (float)profile->default_count() / (hi_index != max_jint ? 2.0F : 1.0F);
822 }
823 ranges[++rp].setRange(min_jint, lo_index-1, default_dest, cnt);
824 }
825 for (int j = 0; j < len; j++) {
826 jint match_int = lo_index+j;
827 int dest = iter().get_dest_table(j+3);
828 makes_backward_branch |= (dest <= bci());
829 float cnt = 1.0F;
830 if (profile != nullptr) {
831 cnt = (float)profile->count_at(j);
832 }
833 if (rp < 0 || !ranges[rp].adjoin(match_int, dest, cnt, trim_ranges)) {
834 ranges[++rp].set(match_int, dest, cnt);
835 }
836 }
837 jint highest = lo_index+(len-1);
838 assert(ranges[rp].hi() == highest, "");
839 if (highest != max_jint) {
840 float cnt = 1.0F;
841 if (profile != nullptr) {
842 cnt = (float)profile->default_count() / (lo_index != min_jint ? 2.0F : 1.0F);
843 }
844 if (!ranges[rp].adjoinRange(highest+1, max_jint, default_dest, cnt, trim_ranges)) {
845 ranges[++rp].setRange(highest+1, max_jint, default_dest, cnt);
846 }
847 }
848 assert(rp < len+2, "not too many ranges");
849
850 if (trim_ranges) {
851 merge_ranges(ranges, rp);
852 }
853
854 // Safepoint in case if backward branch observed
855 if (makes_backward_branch) {
856 add_safepoint();
857 }
858
859 Node* lookup = pop(); // lookup value
860 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]);
861 }
862
863
864 //------------------------------do_lookupswitch--------------------------------
865 void Parse::do_lookupswitch() {
866 // Get information about lookupswitch
867 int default_dest = iter().get_dest_table(0);
868 jint len = iter().get_int_table(1);
869
870 if (len < 1) { // If this is a backward branch, add safepoint
871 maybe_add_safepoint(default_dest);
872 pop(); // the effect of the instruction execution on the operand stack
873 merge(default_dest);
874 return;
875 }
876
877 ciMethodData* methodData = method()->method_data();
878 ciMultiBranchData* profile = nullptr;
879 if (methodData->is_mature() && UseSwitchProfiling) {
880 ciProfileData* data = methodData->bci_to_data(bci());
881 if (data != nullptr && data->is_MultiBranchData()) {
882 profile = (ciMultiBranchData*)data;
883 }
884 }
885 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if);
886
887 // generate decision tree, using trichotomy when possible
888 jint* table = NEW_RESOURCE_ARRAY(jint, len*3);
889 {
890 for (int j = 0; j < len; j++) {
891 table[3*j+0] = iter().get_int_table(2+2*j);
892 table[3*j+1] = iter().get_dest_table(2+2*j+1);
893 // Handle overflow when converting from uint to jint
894 table[3*j+2] = (profile == nullptr) ? 1 : (jint)MIN2<uint>((uint)max_jint, profile->count_at(j));
895 }
896 qsort(table, len, 3*sizeof(table[0]), jint_cmp);
897 }
898
899 float default_cnt = 1.0F;
900 if (profile != nullptr) {
901 juint defaults = max_juint - len;
902 default_cnt = (float)profile->default_count()/(float)defaults;
903 }
904
905 int rnum = len*2+1;
906 bool makes_backward_branch = (default_dest <= bci());
907 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum);
908 int rp = -1;
909 for (int j = 0; j < len; j++) {
910 jint match_int = table[3*j+0];
911 jint dest = table[3*j+1];
912 jint cnt = table[3*j+2];
913 jint next_lo = rp < 0 ? min_jint : ranges[rp].hi()+1;
914 makes_backward_branch |= (dest <= bci());
915 float c = default_cnt * ((float)match_int - (float)next_lo);
916 if (match_int != next_lo && (rp < 0 || !ranges[rp].adjoinRange(next_lo, match_int-1, default_dest, c, trim_ranges))) {
917 assert(default_dest != never_reached, "sentinel value for dead destinations");
918 ranges[++rp].setRange(next_lo, match_int-1, default_dest, c);
919 }
920 if (rp < 0 || !ranges[rp].adjoin(match_int, dest, (float)cnt, trim_ranges)) {
921 assert(dest != never_reached, "sentinel value for dead destinations");
922 ranges[++rp].set(match_int, dest, (float)cnt);
923 }
924 }
925 jint highest = table[3*(len-1)];
926 assert(ranges[rp].hi() == highest, "");
927 if (highest != max_jint &&
928 !ranges[rp].adjoinRange(highest+1, max_jint, default_dest, default_cnt * ((float)max_jint - (float)highest), trim_ranges)) {
929 ranges[++rp].setRange(highest+1, max_jint, default_dest, default_cnt * ((float)max_jint - (float)highest));
930 }
931 assert(rp < rnum, "not too many ranges");
932
933 if (trim_ranges) {
934 merge_ranges(ranges, rp);
935 }
936
937 // Safepoint in case backward branch observed
938 if (makes_backward_branch) {
939 add_safepoint();
940 }
941
942 Node *lookup = pop(); // lookup value
943 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]);
944 }
945
946 static float if_prob(float taken_cnt, float total_cnt) {
947 assert(taken_cnt <= total_cnt, "");
948 if (total_cnt == 0) {
949 return PROB_FAIR;
950 }
951 float p = taken_cnt / total_cnt;
952 return clamp(p, PROB_MIN, PROB_MAX);
953 }
954
955 static float if_cnt(float cnt) {
956 if (cnt == 0) {
957 return COUNT_UNKNOWN;
958 }
959 return cnt;
960 }
961
962 static float sum_of_cnts(SwitchRange *lo, SwitchRange *hi) {
963 float total_cnt = 0;
964 for (SwitchRange* sr = lo; sr <= hi; sr++) {
965 total_cnt += sr->cnt();
966 }
967 return total_cnt;
968 }
969
970 class SwitchRanges : public ResourceObj {
971 public:
972 SwitchRange* _lo;
973 SwitchRange* _hi;
974 SwitchRange* _mid;
975 float _cost;
976
977 enum {
978 Start,
979 LeftDone,
980 RightDone,
981 Done
982 } _state;
983
984 SwitchRanges(SwitchRange *lo, SwitchRange *hi)
985 : _lo(lo), _hi(hi), _mid(nullptr),
986 _cost(0), _state(Start) {
987 }
988
989 SwitchRanges()
990 : _lo(nullptr), _hi(nullptr), _mid(nullptr),
991 _cost(0), _state(Start) {}
992 };
993
994 // Estimate cost of performing a binary search on lo..hi
995 static float compute_tree_cost(SwitchRange *lo, SwitchRange *hi, float total_cnt) {
996 GrowableArray<SwitchRanges> tree;
997 SwitchRanges root(lo, hi);
998 tree.push(root);
999
1000 float cost = 0;
1001 do {
1002 SwitchRanges& r = *tree.adr_at(tree.length()-1);
1003 if (r._hi != r._lo) {
1004 if (r._mid == nullptr) {
1005 float r_cnt = sum_of_cnts(r._lo, r._hi);
1006
1007 if (r_cnt == 0) {
1008 tree.pop();
1009 cost = 0;
1010 continue;
1011 }
1012
1013 SwitchRange* mid = nullptr;
1014 mid = r._lo;
1015 for (float cnt = 0; ; ) {
1016 assert(mid <= r._hi, "out of bounds");
1017 cnt += mid->cnt();
1018 if (cnt > r_cnt / 2) {
1019 break;
1020 }
1021 mid++;
1022 }
1023 assert(mid <= r._hi, "out of bounds");
1024 r._mid = mid;
1025 r._cost = r_cnt / total_cnt;
1026 }
1027 r._cost += cost;
1028 if (r._state < SwitchRanges::LeftDone && r._mid > r._lo) {
1029 cost = 0;
1030 r._state = SwitchRanges::LeftDone;
1031 tree.push(SwitchRanges(r._lo, r._mid-1));
1032 } else if (r._state < SwitchRanges::RightDone) {
1033 cost = 0;
1034 r._state = SwitchRanges::RightDone;
1035 tree.push(SwitchRanges(r._mid == r._lo ? r._mid+1 : r._mid, r._hi));
1036 } else {
1037 tree.pop();
1038 cost = r._cost;
1039 }
1040 } else {
1041 tree.pop();
1042 cost = r._cost;
1043 }
1044 } while (tree.length() > 0);
1045
1046
1047 return cost;
1048 }
1049
1050 // It sometimes pays off to test most common ranges before the binary search
1051 void Parse::linear_search_switch_ranges(Node* key_val, SwitchRange*& lo, SwitchRange*& hi) {
1052 uint nr = hi - lo + 1;
1053 float total_cnt = sum_of_cnts(lo, hi);
1054
1055 float min = compute_tree_cost(lo, hi, total_cnt);
1056 float extra = 1;
1057 float sub = 0;
1058
1059 SwitchRange* array1 = lo;
1060 SwitchRange* array2 = NEW_RESOURCE_ARRAY(SwitchRange, nr);
1061
1062 SwitchRange* ranges = nullptr;
1063
1064 while (nr >= 2) {
1065 assert(lo == array1 || lo == array2, "one the 2 already allocated arrays");
1066 ranges = (lo == array1) ? array2 : array1;
1067
1068 // Find highest frequency range
1069 SwitchRange* candidate = lo;
1070 for (SwitchRange* sr = lo+1; sr <= hi; sr++) {
1071 if (sr->cnt() > candidate->cnt()) {
1072 candidate = sr;
1073 }
1074 }
1075 SwitchRange most_freq = *candidate;
1076 if (most_freq.cnt() == 0) {
1077 break;
1078 }
1079
1080 // Copy remaining ranges into another array
1081 int shift = 0;
1082 for (uint i = 0; i < nr; i++) {
1083 SwitchRange* sr = &lo[i];
1084 if (sr != candidate) {
1085 ranges[i-shift] = *sr;
1086 } else {
1087 shift++;
1088 if (i > 0 && i < nr-1) {
1089 SwitchRange prev = lo[i-1];
1090 prev.setRange(prev.lo(), sr->hi(), prev.dest(), prev.cnt());
1091 if (prev.adjoin(lo[i+1])) {
1092 shift++;
1093 i++;
1094 }
1095 ranges[i-shift] = prev;
1096 }
1097 }
1098 }
1099 nr -= shift;
1100
1101 // Evaluate cost of testing the most common range and performing a
1102 // binary search on the other ranges
1103 float cost = extra + compute_tree_cost(&ranges[0], &ranges[nr-1], total_cnt);
1104 if (cost >= min) {
1105 break;
1106 }
1107 // swap arrays
1108 lo = &ranges[0];
1109 hi = &ranges[nr-1];
1110
1111 // It pays off: emit the test for the most common range
1112 assert(most_freq.cnt() > 0, "must be taken");
1113 Node* val = _gvn.transform(new SubINode(key_val, _gvn.intcon(most_freq.lo())));
1114 Node* cmp = _gvn.transform(new CmpUNode(val, _gvn.intcon(java_subtract(most_freq.hi(), most_freq.lo()))));
1115 Node* tst = _gvn.transform(new BoolNode(cmp, BoolTest::le));
1116 IfNode* iff = create_and_map_if(control(), tst, if_prob(most_freq.cnt(), total_cnt), if_cnt(most_freq.cnt()));
1117 jump_if_true_fork(iff, most_freq.dest(), false);
1118
1119 sub += most_freq.cnt() / total_cnt;
1120 extra += 1 - sub;
1121 min = cost;
1122 }
1123 }
1124
1125 //----------------------------create_jump_tables-------------------------------
1126 bool Parse::create_jump_tables(Node* key_val, SwitchRange* lo, SwitchRange* hi) {
1127 // Are jumptables enabled
1128 if (!UseJumpTables) return false;
1129
1130 // Are jumptables supported
1131 if (!Matcher::has_match_rule(Op_Jump)) return false;
1132
1133 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if);
1134
1135 // Decide if a guard is needed to lop off big ranges at either (or
1136 // both) end(s) of the input set. We'll call this the default target
1137 // even though we can't be sure that it is the true "default".
1138
1139 bool needs_guard = false;
1140 int default_dest;
1141 int64_t total_outlier_size = 0;
1142 int64_t hi_size = ((int64_t)hi->hi()) - ((int64_t)hi->lo()) + 1;
1143 int64_t lo_size = ((int64_t)lo->hi()) - ((int64_t)lo->lo()) + 1;
1144
1145 if (lo->dest() == hi->dest()) {
1146 total_outlier_size = hi_size + lo_size;
1147 default_dest = lo->dest();
1148 } else if (lo_size > hi_size) {
1149 total_outlier_size = lo_size;
1150 default_dest = lo->dest();
1151 } else {
1152 total_outlier_size = hi_size;
1153 default_dest = hi->dest();
1154 }
1155
1156 float total = sum_of_cnts(lo, hi);
1157 float cost = compute_tree_cost(lo, hi, total);
1158
1159 // If a guard test will eliminate very sparse end ranges, then
1160 // it is worth the cost of an extra jump.
1161 float trimmed_cnt = 0;
1162 if (total_outlier_size > (MaxJumpTableSparseness * 4)) {
1163 needs_guard = true;
1164 if (default_dest == lo->dest()) {
1165 trimmed_cnt += lo->cnt();
1166 lo++;
1167 }
1168 if (default_dest == hi->dest()) {
1169 trimmed_cnt += hi->cnt();
1170 hi--;
1171 }
1172 }
1173
1174 // Find the total number of cases and ranges
1175 int64_t num_cases = ((int64_t)hi->hi()) - ((int64_t)lo->lo()) + 1;
1176 int num_range = hi - lo + 1;
1177
1178 // Don't create table if: too large, too small, or too sparse.
1179 if (num_cases > MaxJumpTableSize)
1180 return false;
1181 if (UseSwitchProfiling) {
1182 // MinJumpTableSize is set so with a well balanced binary tree,
1183 // when the number of ranges is MinJumpTableSize, it's cheaper to
1184 // go through a JumpNode that a tree of IfNodes. Average cost of a
1185 // tree of IfNodes with MinJumpTableSize is
1186 // log2f(MinJumpTableSize) comparisons. So if the cost computed
1187 // from profile data is less than log2f(MinJumpTableSize) then
1188 // going with the binary search is cheaper.
1189 if (cost < log2f(MinJumpTableSize)) {
1190 return false;
1191 }
1192 } else {
1193 if (num_cases < MinJumpTableSize)
1194 return false;
1195 }
1196 if (num_cases > (MaxJumpTableSparseness * num_range))
1197 return false;
1198
1199 // Normalize table lookups to zero
1200 int lowval = lo->lo();
1201 key_val = _gvn.transform( new SubINode(key_val, _gvn.intcon(lowval)) );
1202
1203 // Generate a guard to protect against input keyvals that aren't
1204 // in the switch domain.
1205 if (needs_guard) {
1206 Node* size = _gvn.intcon(num_cases);
1207 Node* cmp = _gvn.transform(new CmpUNode(key_val, size));
1208 Node* tst = _gvn.transform(new BoolNode(cmp, BoolTest::ge));
1209 IfNode* iff = create_and_map_if(control(), tst, if_prob(trimmed_cnt, total), if_cnt(trimmed_cnt));
1210 jump_if_true_fork(iff, default_dest, trim_ranges && trimmed_cnt == 0);
1211
1212 total -= trimmed_cnt;
1213 }
1214
1215 // Create an ideal node JumpTable that has projections
1216 // of all possible ranges for a switch statement
1217 // The key_val input must be converted to a pointer offset and scaled.
1218 // Compare Parse::array_addressing above.
1219
1220 // Clean the 32-bit int into a real 64-bit offset.
1221 // Otherwise, the jint value 0 might turn into an offset of 0x0800000000.
1222 // Make I2L conversion control dependent to prevent it from
1223 // floating above the range check during loop optimizations.
1224 // Do not use a narrow int type here to prevent the data path from dying
1225 // while the control path is not removed. This can happen if the type of key_val
1226 // is later known to be out of bounds of [0, num_cases] and therefore a narrow cast
1227 // would be replaced by TOP while C2 is not able to fold the corresponding range checks.
1228 // Set _carry_dependency for the cast to avoid being removed by IGVN.
1229 #ifdef _LP64
1230 key_val = C->constrained_convI2L(&_gvn, key_val, TypeInt::INT, control(), true /* carry_dependency */);
1231 #endif
1232
1233 // Shift the value by wordsize so we have an index into the table, rather
1234 // than a switch value
1235 Node *shiftWord = _gvn.MakeConX(wordSize);
1236 key_val = _gvn.transform( new MulXNode( key_val, shiftWord));
1237
1238 // Create the JumpNode
1239 Arena* arena = C->comp_arena();
1240 float* probs = (float*)arena->Amalloc(sizeof(float)*num_cases);
1241 int i = 0;
1242 if (total == 0) {
1243 for (SwitchRange* r = lo; r <= hi; r++) {
1244 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) {
1245 probs[i] = 1.0F / num_cases;
1246 }
1247 }
1248 } else {
1249 for (SwitchRange* r = lo; r <= hi; r++) {
1250 float prob = r->cnt()/total;
1251 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) {
1252 probs[i] = prob / (r->hi() - r->lo() + 1);
1253 }
1254 }
1255 }
1256
1257 ciMethodData* methodData = method()->method_data();
1258 ciMultiBranchData* profile = nullptr;
1259 if (methodData->is_mature()) {
1260 ciProfileData* data = methodData->bci_to_data(bci());
1261 if (data != nullptr && data->is_MultiBranchData()) {
1262 profile = (ciMultiBranchData*)data;
1263 }
1264 }
1265
1266 Node* jtn = _gvn.transform(new JumpNode(control(), key_val, num_cases, probs, profile == nullptr ? COUNT_UNKNOWN : total));
1267
1268 // These are the switch destinations hanging off the jumpnode
1269 i = 0;
1270 for (SwitchRange* r = lo; r <= hi; r++) {
1271 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) {
1272 Node* input = _gvn.transform(new JumpProjNode(jtn, i, r->dest(), (int)(j - lowval)));
1273 {
1274 PreserveJVMState pjvms(this);
1275 set_control(input);
1276 jump_if_always_fork(r->dest(), trim_ranges && r->cnt() == 0);
1277 }
1278 }
1279 }
1280 assert(i == num_cases, "miscount of cases");
1281 stop_and_kill_map(); // no more uses for this JVMS
1282 return true;
1283 }
1284
1285 //----------------------------jump_switch_ranges-------------------------------
1286 void Parse::jump_switch_ranges(Node* key_val, SwitchRange *lo, SwitchRange *hi, int switch_depth) {
1287 Block* switch_block = block();
1288 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if);
1289
1290 if (switch_depth == 0) {
1291 // Do special processing for the top-level call.
1292 assert(lo->lo() == min_jint, "initial range must exhaust Type::INT");
1293 assert(hi->hi() == max_jint, "initial range must exhaust Type::INT");
1294
1295 // Decrement pred-numbers for the unique set of nodes.
1296 #ifdef ASSERT
1297 if (!trim_ranges) {
1298 // Ensure that the block's successors are a (duplicate-free) set.
1299 int successors_counted = 0; // block occurrences in [hi..lo]
1300 int unique_successors = switch_block->num_successors();
1301 for (int i = 0; i < unique_successors; i++) {
1302 Block* target = switch_block->successor_at(i);
1303
1304 // Check that the set of successors is the same in both places.
1305 int successors_found = 0;
1306 for (SwitchRange* p = lo; p <= hi; p++) {
1307 if (p->dest() == target->start()) successors_found++;
1308 }
1309 assert(successors_found > 0, "successor must be known");
1310 successors_counted += successors_found;
1311 }
1312 assert(successors_counted == (hi-lo)+1, "no unexpected successors");
1313 }
1314 #endif
1315
1316 // Maybe prune the inputs, based on the type of key_val.
1317 jint min_val = min_jint;
1318 jint max_val = max_jint;
1319 const TypeInt* ti = key_val->bottom_type()->isa_int();
1320 if (ti != nullptr) {
1321 min_val = ti->_lo;
1322 max_val = ti->_hi;
1323 assert(min_val <= max_val, "invalid int type");
1324 }
1325 while (lo->hi() < min_val) {
1326 lo++;
1327 }
1328 if (lo->lo() < min_val) {
1329 lo->setRange(min_val, lo->hi(), lo->dest(), lo->cnt());
1330 }
1331 while (hi->lo() > max_val) {
1332 hi--;
1333 }
1334 if (hi->hi() > max_val) {
1335 hi->setRange(hi->lo(), max_val, hi->dest(), hi->cnt());
1336 }
1337
1338 linear_search_switch_ranges(key_val, lo, hi);
1339 }
1340
1341 #ifndef PRODUCT
1342 if (switch_depth == 0) {
1343 _max_switch_depth = 0;
1344 _est_switch_depth = log2i_graceful((hi - lo + 1) - 1) + 1;
1345 }
1346 #endif
1347
1348 assert(lo <= hi, "must be a non-empty set of ranges");
1349 if (lo == hi) {
1350 jump_if_always_fork(lo->dest(), trim_ranges && lo->cnt() == 0);
1351 } else {
1352 assert(lo->hi() == (lo+1)->lo()-1, "contiguous ranges");
1353 assert(hi->lo() == (hi-1)->hi()+1, "contiguous ranges");
1354
1355 if (create_jump_tables(key_val, lo, hi)) return;
1356
1357 SwitchRange* mid = nullptr;
1358 float total_cnt = sum_of_cnts(lo, hi);
1359
1360 int nr = hi - lo + 1;
1361 if (UseSwitchProfiling) {
1362 // Don't keep the binary search tree balanced: pick up mid point
1363 // that split frequencies in half.
1364 float cnt = 0;
1365 for (SwitchRange* sr = lo; sr <= hi; sr++) {
1366 cnt += sr->cnt();
1367 if (cnt >= total_cnt / 2) {
1368 mid = sr;
1369 break;
1370 }
1371 }
1372 } else {
1373 mid = lo + nr/2;
1374
1375 // if there is an easy choice, pivot at a singleton:
1376 if (nr > 3 && !mid->is_singleton() && (mid-1)->is_singleton()) mid--;
1377
1378 assert(lo < mid && mid <= hi, "good pivot choice");
1379 assert(nr != 2 || mid == hi, "should pick higher of 2");
1380 assert(nr != 3 || mid == hi-1, "should pick middle of 3");
1381 }
1382
1383
1384 Node *test_val = _gvn.intcon(mid == lo ? mid->hi() : mid->lo());
1385
1386 if (mid->is_singleton()) {
1387 IfNode *iff_ne = jump_if_fork_int(key_val, test_val, BoolTest::ne, 1-if_prob(mid->cnt(), total_cnt), if_cnt(mid->cnt()));
1388 jump_if_false_fork(iff_ne, mid->dest(), trim_ranges && mid->cnt() == 0);
1389
1390 // Special Case: If there are exactly three ranges, and the high
1391 // and low range each go to the same place, omit the "gt" test,
1392 // since it will not discriminate anything.
1393 bool eq_test_only = (hi == lo+2 && hi->dest() == lo->dest() && mid == hi-1) || mid == lo;
1394
1395 // if there is a higher range, test for it and process it:
1396 if (mid < hi && !eq_test_only) {
1397 // two comparisons of same values--should enable 1 test for 2 branches
1398 // Use BoolTest::lt instead of BoolTest::gt
1399 float cnt = sum_of_cnts(lo, mid-1);
1400 IfNode *iff_lt = jump_if_fork_int(key_val, test_val, BoolTest::lt, if_prob(cnt, total_cnt), if_cnt(cnt));
1401 Node *iftrue = _gvn.transform( new IfTrueNode(iff_lt) );
1402 Node *iffalse = _gvn.transform( new IfFalseNode(iff_lt) );
1403 { PreserveJVMState pjvms(this);
1404 set_control(iffalse);
1405 jump_switch_ranges(key_val, mid+1, hi, switch_depth+1);
1406 }
1407 set_control(iftrue);
1408 }
1409
1410 } else {
1411 // mid is a range, not a singleton, so treat mid..hi as a unit
1412 float cnt = sum_of_cnts(mid == lo ? mid+1 : mid, hi);
1413 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));
1414
1415 // if there is a higher range, test for it and process it:
1416 if (mid == hi) {
1417 jump_if_true_fork(iff_ge, mid->dest(), trim_ranges && cnt == 0);
1418 } else {
1419 Node *iftrue = _gvn.transform( new IfTrueNode(iff_ge) );
1420 Node *iffalse = _gvn.transform( new IfFalseNode(iff_ge) );
1421 { PreserveJVMState pjvms(this);
1422 set_control(iftrue);
1423 jump_switch_ranges(key_val, mid == lo ? mid+1 : mid, hi, switch_depth+1);
1424 }
1425 set_control(iffalse);
1426 }
1427 }
1428
1429 // in any case, process the lower range
1430 if (mid == lo) {
1431 if (mid->is_singleton()) {
1432 jump_switch_ranges(key_val, lo+1, hi, switch_depth+1);
1433 } else {
1434 jump_if_always_fork(lo->dest(), trim_ranges && lo->cnt() == 0);
1435 }
1436 } else {
1437 jump_switch_ranges(key_val, lo, mid-1, switch_depth+1);
1438 }
1439 }
1440
1441 // Decrease pred_count for each successor after all is done.
1442 if (switch_depth == 0) {
1443 int unique_successors = switch_block->num_successors();
1444 for (int i = 0; i < unique_successors; i++) {
1445 Block* target = switch_block->successor_at(i);
1446 // Throw away the pre-allocated path for each unique successor.
1447 target->next_path_num();
1448 }
1449 }
1450
1451 #ifndef PRODUCT
1452 _max_switch_depth = MAX2(switch_depth, _max_switch_depth);
1453 if (TraceOptoParse && Verbose && WizardMode && switch_depth == 0) {
1454 SwitchRange* r;
1455 int nsing = 0;
1456 for( r = lo; r <= hi; r++ ) {
1457 if( r->is_singleton() ) nsing++;
1458 }
1459 tty->print(">>> ");
1460 _method->print_short_name();
1461 tty->print_cr(" switch decision tree");
1462 tty->print_cr(" %d ranges (%d singletons), max_depth=%d, est_depth=%d",
1463 (int) (hi-lo+1), nsing, _max_switch_depth, _est_switch_depth);
1464 if (_max_switch_depth > _est_switch_depth) {
1465 tty->print_cr("******** BAD SWITCH DEPTH ********");
1466 }
1467 tty->print(" ");
1468 for( r = lo; r <= hi; r++ ) {
1469 r->print();
1470 }
1471 tty->cr();
1472 }
1473 #endif
1474 }
1475
1476 void Parse::modf() {
1477 Node *f2 = pop();
1478 Node *f1 = pop();
1479 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::modf_Type(),
1480 CAST_FROM_FN_PTR(address, SharedRuntime::frem),
1481 "frem", nullptr, //no memory effects
1482 f1, f2);
1483 Node* res = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0));
1484
1485 push(res);
1486 }
1487
1488 void Parse::modd() {
1489 Node *d2 = pop_pair();
1490 Node *d1 = pop_pair();
1491 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::Math_DD_D_Type(),
1492 CAST_FROM_FN_PTR(address, SharedRuntime::drem),
1493 "drem", nullptr, //no memory effects
1494 d1, top(), d2, top());
1495 Node* res_d = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0));
1496
1497 #ifdef ASSERT
1498 Node* res_top = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 1));
1499 assert(res_top == top(), "second value must be top");
1500 #endif
1501
1502 push_pair(res_d);
1503 }
1504
1505 void Parse::l2f() {
1506 Node* f2 = pop();
1507 Node* f1 = pop();
1508 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::l2f_Type(),
1509 CAST_FROM_FN_PTR(address, SharedRuntime::l2f),
1510 "l2f", nullptr, //no memory effects
1511 f1, f2);
1512 Node* res = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0));
1513
1514 push(res);
1515 }
1516
1517 // Handle jsr and jsr_w bytecode
1518 void Parse::do_jsr() {
1519 assert(bc() == Bytecodes::_jsr || bc() == Bytecodes::_jsr_w, "wrong bytecode");
1520
1521 // Store information about current state, tagged with new _jsr_bci
1522 int return_bci = iter().next_bci();
1523 int jsr_bci = (bc() == Bytecodes::_jsr) ? iter().get_dest() : iter().get_far_dest();
1524
1525 // The way we do things now, there is only one successor block
1526 // for the jsr, because the target code is cloned by ciTypeFlow.
1527 Block* target = successor_for_bci(jsr_bci);
1528
1529 // What got pushed?
1530 const Type* ret_addr = target->peek();
1531 assert(ret_addr->singleton(), "must be a constant (cloned jsr body)");
1532
1533 // Effect on jsr on stack
1534 push(_gvn.makecon(ret_addr));
1535
1536 // Flow to the jsr.
1537 merge(jsr_bci);
1538 }
1539
1540 // Handle ret bytecode
1541 void Parse::do_ret() {
1542 // Find to whom we return.
1543 assert(block()->num_successors() == 1, "a ret can only go one place now");
1544 Block* target = block()->successor_at(0);
1545 assert(!target->is_ready(), "our arrival must be expected");
1546 int pnum = target->next_path_num();
1547 merge_common(target, pnum);
1548 }
1549
1550 static bool has_injected_profile(BoolTest::mask btest, Node* test, int& taken, int& not_taken) {
1551 if (btest != BoolTest::eq && btest != BoolTest::ne) {
1552 // Only ::eq and ::ne are supported for profile injection.
1553 return false;
1554 }
1555 if (test->is_Cmp() &&
1556 test->in(1)->Opcode() == Op_ProfileBoolean) {
1557 ProfileBooleanNode* profile = (ProfileBooleanNode*)test->in(1);
1558 int false_cnt = profile->false_count();
1559 int true_cnt = profile->true_count();
1560
1561 // Counts matching depends on the actual test operation (::eq or ::ne).
1562 // No need to scale the counts because profile injection was designed
1563 // to feed exact counts into VM.
1564 taken = (btest == BoolTest::eq) ? false_cnt : true_cnt;
1565 not_taken = (btest == BoolTest::eq) ? true_cnt : false_cnt;
1566
1567 profile->consume();
1568 return true;
1569 }
1570 return false;
1571 }
1572
1573 // Give up if too few (or too many, in which case the sum will overflow) counts to be meaningful.
1574 // We also check that individual counters are positive first, otherwise the sum can become positive.
1575 // (check for saturation, integer overflow, and immature counts)
1576 static bool counters_are_meaningful(int counter1, int counter2, int min) {
1577 // check for saturation, including "uint" values too big to fit in "int"
1578 if (counter1 < 0 || counter2 < 0) {
1579 return false;
1580 }
1581 // check for integer overflow of the sum
1582 int64_t sum = (int64_t)counter1 + (int64_t)counter2;
1583 STATIC_ASSERT(sizeof(counter1) < sizeof(sum));
1584 if (sum > INT_MAX) {
1585 return false;
1586 }
1587 // check if mature
1588 return (counter1 + counter2) >= min;
1589 }
1590
1591 //--------------------------dynamic_branch_prediction--------------------------
1592 // Try to gather dynamic branch prediction behavior. Return a probability
1593 // of the branch being taken and set the "cnt" field. Returns a -1.0
1594 // if we need to use static prediction for some reason.
1595 float Parse::dynamic_branch_prediction(float &cnt, BoolTest::mask btest, Node* test) {
1596 ResourceMark rm;
1597
1598 cnt = COUNT_UNKNOWN;
1599
1600 int taken = 0;
1601 int not_taken = 0;
1602
1603 bool use_mdo = !has_injected_profile(btest, test, taken, not_taken);
1604
1605 if (use_mdo) {
1606 // Use MethodData information if it is available
1607 // FIXME: free the ProfileData structure
1608 ciMethodData* methodData = method()->method_data();
1609 if (!methodData->is_mature()) return PROB_UNKNOWN;
1610 ciProfileData* data = methodData->bci_to_data(bci());
1611 if (data == nullptr) {
1612 return PROB_UNKNOWN;
1613 }
1614 if (!data->is_JumpData()) return PROB_UNKNOWN;
1615
1616 // get taken and not taken values
1617 // NOTE: saturated UINT_MAX values become negative,
1618 // as do counts above INT_MAX.
1619 taken = data->as_JumpData()->taken();
1620 not_taken = 0;
1621 if (data->is_BranchData()) {
1622 not_taken = data->as_BranchData()->not_taken();
1623 }
1624
1625 // scale the counts to be commensurate with invocation counts:
1626 // NOTE: overflow for positive values is clamped at INT_MAX
1627 taken = method()->scale_count(taken);
1628 not_taken = method()->scale_count(not_taken);
1629 }
1630 // At this point, saturation or overflow is indicated by INT_MAX
1631 // or a negative value.
1632
1633 // Give up if too few (or too many, in which case the sum will overflow) counts to be meaningful.
1634 // We also check that individual counters are positive first, otherwise the sum can become positive.
1635 if (!counters_are_meaningful(taken, not_taken, 40)) {
1636 if (C->log() != nullptr) {
1637 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d'", iter().get_dest(), taken, not_taken);
1638 }
1639 return PROB_UNKNOWN;
1640 }
1641
1642 // Compute frequency that we arrive here
1643 float sum = taken + not_taken;
1644 // Adjust, if this block is a cloned private block but the
1645 // Jump counts are shared. Taken the private counts for
1646 // just this path instead of the shared counts.
1647 if( block()->count() > 0 )
1648 sum = block()->count();
1649 cnt = sum / FreqCountInvocations;
1650
1651 // Pin probability to sane limits
1652 float prob;
1653 if( !taken )
1654 prob = (0+PROB_MIN) / 2;
1655 else if( !not_taken )
1656 prob = (1+PROB_MAX) / 2;
1657 else { // Compute probability of true path
1658 prob = (float)taken / (float)(taken + not_taken);
1659 if (prob > PROB_MAX) prob = PROB_MAX;
1660 if (prob < PROB_MIN) prob = PROB_MIN;
1661 }
1662
1663 assert((cnt > 0.0f) && (prob > 0.0f),
1664 "Bad frequency assignment in if cnt=%g prob=%g taken=%d not_taken=%d", cnt, prob, taken, not_taken);
1665
1666 if (C->log() != nullptr) {
1667 const char* prob_str = nullptr;
1668 if (prob >= PROB_MAX) prob_str = (prob == PROB_MAX) ? "max" : "always";
1669 if (prob <= PROB_MIN) prob_str = (prob == PROB_MIN) ? "min" : "never";
1670 char prob_str_buf[30];
1671 if (prob_str == nullptr) {
1672 jio_snprintf(prob_str_buf, sizeof(prob_str_buf), "%20.2f", prob);
1673 prob_str = prob_str_buf;
1674 }
1675 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d' cnt='%f' prob='%s'",
1676 iter().get_dest(), taken, not_taken, cnt, prob_str);
1677 }
1678 return prob;
1679 }
1680
1681 //-----------------------------branch_prediction-------------------------------
1682 float Parse::branch_prediction(float& cnt,
1683 BoolTest::mask btest,
1684 int target_bci,
1685 Node* test) {
1686 float prob = dynamic_branch_prediction(cnt, btest, test);
1687 // If prob is unknown, switch to static prediction
1688 if (prob != PROB_UNKNOWN) return prob;
1689
1690 prob = PROB_FAIR; // Set default value
1691 if (btest == BoolTest::eq) // Exactly equal test?
1692 prob = PROB_STATIC_INFREQUENT; // Assume its relatively infrequent
1693 else if (btest == BoolTest::ne)
1694 prob = PROB_STATIC_FREQUENT; // Assume its relatively frequent
1695
1696 // If this is a conditional test guarding a backwards branch,
1697 // assume its a loop-back edge. Make it a likely taken branch.
1698 if (target_bci < bci()) {
1699 if (is_osr_parse()) { // Could be a hot OSR'd loop; force deopt
1700 // Since it's an OSR, we probably have profile data, but since
1701 // branch_prediction returned PROB_UNKNOWN, the counts are too small.
1702 // Let's make a special check here for completely zero counts.
1703 ciMethodData* methodData = method()->method_data();
1704 if (!methodData->is_empty()) {
1705 ciProfileData* data = methodData->bci_to_data(bci());
1706 // Only stop for truly zero counts, which mean an unknown part
1707 // of the OSR-ed method, and we want to deopt to gather more stats.
1708 // If you have ANY counts, then this loop is simply 'cold' relative
1709 // to the OSR loop.
1710 if (data == nullptr ||
1711 (data->as_BranchData()->taken() + data->as_BranchData()->not_taken() == 0)) {
1712 // This is the only way to return PROB_UNKNOWN:
1713 return PROB_UNKNOWN;
1714 }
1715 }
1716 }
1717 prob = PROB_STATIC_FREQUENT; // Likely to take backwards branch
1718 }
1719
1720 assert(prob != PROB_UNKNOWN, "must have some guess at this point");
1721 return prob;
1722 }
1723
1724 // The magic constants are chosen so as to match the output of
1725 // branch_prediction() when the profile reports a zero taken count.
1726 // It is important to distinguish zero counts unambiguously, because
1727 // some branches (e.g., _213_javac.Assembler.eliminate) validly produce
1728 // very small but nonzero probabilities, which if confused with zero
1729 // counts would keep the program recompiling indefinitely.
1730 bool Parse::seems_never_taken(float prob) const {
1731 return prob < PROB_MIN;
1732 }
1733
1734 //-------------------------------repush_if_args--------------------------------
1735 // Push arguments of an "if" bytecode back onto the stack by adjusting _sp.
1736 inline int Parse::repush_if_args() {
1737 if (PrintOpto && WizardMode) {
1738 tty->print("defending against excessive implicit null exceptions on %s @%d in ",
1739 Bytecodes::name(iter().cur_bc()), iter().cur_bci());
1740 method()->print_name(); tty->cr();
1741 }
1742 int bc_depth = - Bytecodes::depth(iter().cur_bc());
1743 assert(bc_depth == 1 || bc_depth == 2, "only two kinds of branches");
1744 DEBUG_ONLY(sync_jvms()); // argument(n) requires a synced jvms
1745 assert(argument(0) != nullptr, "must exist");
1746 assert(bc_depth == 1 || argument(1) != nullptr, "two must exist");
1747 inc_sp(bc_depth);
1748 return bc_depth;
1749 }
1750
1751 // Used by StressUnstableIfTraps
1752 static volatile int _trap_stress_counter = 0;
1753
1754 void Parse::increment_trap_stress_counter(Node*& counter, Node*& incr_store) {
1755 Node* counter_addr = makecon(TypeRawPtr::make((address)&_trap_stress_counter));
1756 counter = make_load(control(), counter_addr, TypeInt::INT, T_INT, Compile::AliasIdxRaw, MemNode::unordered);
1757 counter = _gvn.transform(new AddINode(counter, intcon(1)));
1758 incr_store = store_to_memory(control(), counter_addr, counter, T_INT, Compile::AliasIdxRaw, MemNode::unordered);
1759 }
1760
1761 //----------------------------------do_ifnull----------------------------------
1762 void Parse::do_ifnull(BoolTest::mask btest, Node *c) {
1763 int target_bci = iter().get_dest();
1764
1765 Node* counter = nullptr;
1766 Node* incr_store = nullptr;
1767 bool do_stress_trap = StressUnstableIfTraps && ((C->random() % 2) == 0);
1768 if (do_stress_trap) {
1769 increment_trap_stress_counter(counter, incr_store);
1770 }
1771
1772 Block* branch_block = successor_for_bci(target_bci);
1773 Block* next_block = successor_for_bci(iter().next_bci());
1774
1775 float cnt;
1776 float prob = branch_prediction(cnt, btest, target_bci, c);
1777 if (prob == PROB_UNKNOWN) {
1778 // (An earlier version of do_ifnull omitted this trap for OSR methods.)
1779 if (PrintOpto && Verbose) {
1780 tty->print_cr("Never-taken edge stops compilation at bci %d", bci());
1781 }
1782 repush_if_args(); // to gather stats on loop
1783 uncommon_trap(Deoptimization::Reason_unreached,
1784 Deoptimization::Action_reinterpret,
1785 nullptr, "cold");
1786 if (C->eliminate_boxing()) {
1787 // Mark the successor blocks as parsed
1788 branch_block->next_path_num();
1789 next_block->next_path_num();
1790 }
1791 return;
1792 }
1793
1794 NOT_PRODUCT(explicit_null_checks_inserted++);
1795
1796 // Generate real control flow
1797 Node *tst = _gvn.transform( new BoolNode( c, btest ) );
1798
1799 // Sanity check the probability value
1800 assert(prob > 0.0f,"Bad probability in Parser");
1801 // Need xform to put node in hash table
1802 IfNode *iff = create_and_xform_if( control(), tst, prob, cnt );
1803 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser");
1804 // True branch
1805 { PreserveJVMState pjvms(this);
1806 Node* iftrue = _gvn.transform( new IfTrueNode (iff) );
1807 set_control(iftrue);
1808
1809 if (stopped()) { // Path is dead?
1810 NOT_PRODUCT(explicit_null_checks_elided++);
1811 if (C->eliminate_boxing()) {
1812 // Mark the successor block as parsed
1813 branch_block->next_path_num();
1814 }
1815 } else { // Path is live.
1816 adjust_map_after_if(btest, c, prob, branch_block);
1817 if (!stopped()) {
1818 merge(target_bci);
1819 }
1820 }
1821 }
1822
1823 // False branch
1824 Node* iffalse = _gvn.transform( new IfFalseNode(iff) );
1825 set_control(iffalse);
1826
1827 if (stopped()) { // Path is dead?
1828 NOT_PRODUCT(explicit_null_checks_elided++);
1829 if (C->eliminate_boxing()) {
1830 // Mark the successor block as parsed
1831 next_block->next_path_num();
1832 }
1833 } else { // Path is live.
1834 adjust_map_after_if(BoolTest(btest).negate(), c, 1.0-prob, next_block);
1835 }
1836
1837 if (do_stress_trap) {
1838 stress_trap(iff, counter, incr_store);
1839 }
1840 }
1841
1842 //------------------------------------do_if------------------------------------
1843 void Parse::do_if(BoolTest::mask btest, Node* c, bool can_trap, bool new_path, Node** ctrl_taken) {
1844 int target_bci = iter().get_dest();
1845
1846 Block* branch_block = successor_for_bci(target_bci);
1847 Block* next_block = successor_for_bci(iter().next_bci());
1848
1849 float cnt;
1850 float prob = branch_prediction(cnt, btest, target_bci, c);
1851 float untaken_prob = 1.0 - prob;
1852
1853 if (prob == PROB_UNKNOWN) {
1854 if (PrintOpto && Verbose) {
1855 tty->print_cr("Never-taken edge stops compilation at bci %d", bci());
1856 }
1857 repush_if_args(); // to gather stats on loop
1858 uncommon_trap(Deoptimization::Reason_unreached,
1859 Deoptimization::Action_reinterpret,
1860 nullptr, "cold");
1861 if (C->eliminate_boxing()) {
1862 // Mark the successor blocks as parsed
1863 branch_block->next_path_num();
1864 next_block->next_path_num();
1865 }
1866 return;
1867 }
1868
1869 Node* counter = nullptr;
1870 Node* incr_store = nullptr;
1871 bool do_stress_trap = StressUnstableIfTraps && ((C->random() % 2) == 0);
1872 if (do_stress_trap) {
1873 increment_trap_stress_counter(counter, incr_store);
1874 }
1875
1876 // Sanity check the probability value
1877 assert(0.0f < prob && prob < 1.0f,"Bad probability in Parser");
1878
1879 bool taken_if_true = true;
1880 // Convert BoolTest to canonical form:
1881 if (!BoolTest(btest).is_canonical()) {
1882 btest = BoolTest(btest).negate();
1883 taken_if_true = false;
1884 // prob is NOT updated here; it remains the probability of the taken
1885 // path (as opposed to the prob of the path guarded by an 'IfTrueNode').
1886 }
1887 assert(btest != BoolTest::eq, "!= is the only canonical exact test");
1888
1889 Node* tst0 = new BoolNode(c, btest);
1890 Node* tst = _gvn.transform(tst0);
1891 BoolTest::mask taken_btest = BoolTest::illegal;
1892 BoolTest::mask untaken_btest = BoolTest::illegal;
1893
1894 if (tst->is_Bool()) {
1895 // Refresh c from the transformed bool node, since it may be
1896 // simpler than the original c. Also re-canonicalize btest.
1897 // This wins when (Bool ne (Conv2B p) 0) => (Bool ne (CmpP p null)).
1898 // That can arise from statements like: if (x instanceof C) ...
1899 if (tst != tst0) {
1900 // Canonicalize one more time since transform can change it.
1901 btest = tst->as_Bool()->_test._test;
1902 if (!BoolTest(btest).is_canonical()) {
1903 // Reverse edges one more time...
1904 tst = _gvn.transform( tst->as_Bool()->negate(&_gvn) );
1905 btest = tst->as_Bool()->_test._test;
1906 assert(BoolTest(btest).is_canonical(), "sanity");
1907 taken_if_true = !taken_if_true;
1908 }
1909 c = tst->in(1);
1910 }
1911 BoolTest::mask neg_btest = BoolTest(btest).negate();
1912 taken_btest = taken_if_true ? btest : neg_btest;
1913 untaken_btest = taken_if_true ? neg_btest : btest;
1914 }
1915
1916 // Generate real control flow
1917 float true_prob = (taken_if_true ? prob : untaken_prob);
1918 IfNode* iff = create_and_map_if(control(), tst, true_prob, cnt);
1919 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser");
1920 Node* taken_branch = new IfTrueNode(iff);
1921 Node* untaken_branch = new IfFalseNode(iff);
1922 if (!taken_if_true) { // Finish conversion to canonical form
1923 Node* tmp = taken_branch;
1924 taken_branch = untaken_branch;
1925 untaken_branch = tmp;
1926 }
1927
1928 // Branch is taken:
1929 { PreserveJVMState pjvms(this);
1930 taken_branch = _gvn.transform(taken_branch);
1931 set_control(taken_branch);
1932
1933 if (stopped()) {
1934 if (C->eliminate_boxing() && !new_path) {
1935 // Mark the successor block as parsed (if we haven't created a new path)
1936 branch_block->next_path_num();
1937 }
1938 } else {
1939 adjust_map_after_if(taken_btest, c, prob, branch_block, can_trap);
1940 if (!stopped()) {
1941 if (new_path) {
1942 // Merge by using a new path
1943 merge_new_path(target_bci);
1944 } else if (ctrl_taken != nullptr) {
1945 // Don't merge but save taken branch to be wired by caller
1946 *ctrl_taken = control();
1947 } else {
1948 merge(target_bci);
1949 }
1950 }
1951 }
1952 }
1953
1954 untaken_branch = _gvn.transform(untaken_branch);
1955 set_control(untaken_branch);
1956
1957 // Branch not taken.
1958 if (stopped() && ctrl_taken == nullptr) {
1959 if (C->eliminate_boxing()) {
1960 // Mark the successor block as parsed (if caller does not re-wire control flow)
1961 next_block->next_path_num();
1962 }
1963 } else {
1964 adjust_map_after_if(untaken_btest, c, untaken_prob, next_block, can_trap);
1965 }
1966
1967 if (do_stress_trap) {
1968 stress_trap(iff, counter, incr_store);
1969 }
1970 }
1971
1972
1973 static ProfilePtrKind speculative_ptr_kind(const TypeOopPtr* t) {
1974 if (t->speculative() == nullptr) {
1975 return ProfileUnknownNull;
1976 }
1977 if (t->speculative_always_null()) {
1978 return ProfileAlwaysNull;
1979 }
1980 if (t->speculative_maybe_null()) {
1981 return ProfileMaybeNull;
1982 }
1983 return ProfileNeverNull;
1984 }
1985
1986 void Parse::acmp_always_null_input(Node* input, const TypeOopPtr* tinput, BoolTest::mask btest, Node* eq_region) {
1987 inc_sp(2);
1988 Node* cast = null_check_common(input, T_OBJECT, true, nullptr,
1989 !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_null_check) &&
1990 speculative_ptr_kind(tinput) == ProfileAlwaysNull);
1991 dec_sp(2);
1992 if (btest == BoolTest::ne) {
1993 {
1994 PreserveJVMState pjvms(this);
1995 replace_in_map(input, cast);
1996 int target_bci = iter().get_dest();
1997 merge(target_bci);
1998 }
1999 record_for_igvn(eq_region);
2000 set_control(_gvn.transform(eq_region));
2001 } else {
2002 replace_in_map(input, cast);
2003 }
2004 }
2005
2006 Node* Parse::acmp_null_check(Node* input, const TypeOopPtr* tinput, ProfilePtrKind input_ptr, Node*& null_ctl) {
2007 inc_sp(2);
2008 null_ctl = top();
2009 Node* cast = null_check_oop(input, &null_ctl,
2010 input_ptr == ProfileNeverNull || (input_ptr == ProfileUnknownNull && !too_many_traps_or_recompiles(Deoptimization::Reason_null_check)),
2011 false,
2012 speculative_ptr_kind(tinput) == ProfileNeverNull &&
2013 !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_null_check));
2014 dec_sp(2);
2015 assert(!stopped(), "null input should have been caught earlier");
2016 return cast;
2017 }
2018
2019 void Parse::acmp_known_non_inline_type_input(Node* input, const TypeOopPtr* tinput, ProfilePtrKind input_ptr, ciKlass* input_type, BoolTest::mask btest, Node* eq_region) {
2020 Node* ne_region = new RegionNode(1);
2021 Node* null_ctl;
2022 Node* cast = acmp_null_check(input, tinput, input_ptr, null_ctl);
2023 ne_region->add_req(null_ctl);
2024
2025 Node* slow_ctl = type_check_receiver(cast, input_type, 1.0, &cast);
2026 {
2027 PreserveJVMState pjvms(this);
2028 inc_sp(2);
2029 set_control(slow_ctl);
2030 Deoptimization::DeoptReason reason;
2031 if (tinput->speculative_type() != nullptr && !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_class_check)) {
2032 reason = Deoptimization::Reason_speculate_class_check;
2033 } else {
2034 reason = Deoptimization::Reason_class_check;
2035 }
2036 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
2037 }
2038 ne_region->add_req(control());
2039
2040 record_for_igvn(ne_region);
2041 set_control(_gvn.transform(ne_region));
2042 if (btest == BoolTest::ne) {
2043 {
2044 PreserveJVMState pjvms(this);
2045 if (null_ctl == top()) {
2046 replace_in_map(input, cast);
2047 }
2048 int target_bci = iter().get_dest();
2049 merge(target_bci);
2050 }
2051 record_for_igvn(eq_region);
2052 set_control(_gvn.transform(eq_region));
2053 } else {
2054 if (null_ctl == top()) {
2055 replace_in_map(input, cast);
2056 }
2057 set_control(_gvn.transform(ne_region));
2058 }
2059 }
2060
2061 void Parse::acmp_unknown_non_inline_type_input(Node* input, const TypeOopPtr* tinput, ProfilePtrKind input_ptr, BoolTest::mask btest, Node* eq_region) {
2062 Node* ne_region = new RegionNode(1);
2063 Node* null_ctl;
2064 Node* cast = acmp_null_check(input, tinput, input_ptr, null_ctl);
2065 ne_region->add_req(null_ctl);
2066
2067 {
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 ne_region->add_req(control());
2074
2075 record_for_igvn(ne_region);
2076 set_control(_gvn.transform(ne_region));
2077 if (btest == BoolTest::ne) {
2078 {
2079 PreserveJVMState pjvms(this);
2080 if (null_ctl == top()) {
2081 replace_in_map(input, cast);
2082 }
2083 int target_bci = iter().get_dest();
2084 merge(target_bci);
2085 }
2086 record_for_igvn(eq_region);
2087 set_control(_gvn.transform(eq_region));
2088 } else {
2089 if (null_ctl == top()) {
2090 replace_in_map(input, cast);
2091 }
2092 set_control(_gvn.transform(ne_region));
2093 }
2094 }
2095
2096 void Parse::do_acmp(BoolTest::mask btest, Node* left, Node* right) {
2097 ciKlass* left_type = nullptr;
2098 ciKlass* right_type = nullptr;
2099 ProfilePtrKind left_ptr = ProfileUnknownNull;
2100 ProfilePtrKind right_ptr = ProfileUnknownNull;
2101 bool left_inline_type = true;
2102 bool right_inline_type = true;
2103
2104 // Leverage profiling at acmp
2105 if (UseACmpProfile) {
2106 method()->acmp_profiled_type(bci(), left_type, right_type, left_ptr, right_ptr, left_inline_type, right_inline_type);
2107 if (too_many_traps_or_recompiles(Deoptimization::Reason_class_check)) {
2108 left_type = nullptr;
2109 right_type = nullptr;
2110 left_inline_type = true;
2111 right_inline_type = true;
2112 }
2113 if (too_many_traps_or_recompiles(Deoptimization::Reason_null_check)) {
2114 left_ptr = ProfileUnknownNull;
2115 right_ptr = ProfileUnknownNull;
2116 }
2117 }
2118
2119 if (UseTypeSpeculation) {
2120 record_profile_for_speculation(left, left_type, left_ptr);
2121 record_profile_for_speculation(right, right_type, right_ptr);
2122 }
2123
2124 if (!EnableValhalla) {
2125 Node* cmp = CmpP(left, right);
2126 cmp = optimize_cmp_with_klass(cmp);
2127 do_if(btest, cmp);
2128 return;
2129 }
2130
2131 // Check for equality before potentially allocating
2132 if (left == right) {
2133 do_if(btest, makecon(TypeInt::CC_EQ));
2134 return;
2135 }
2136
2137 // Allocate inline type operands and re-execute on deoptimization
2138 if (left->is_InlineType()) {
2139 if (_gvn.type(right)->is_zero_type() ||
2140 (right->is_InlineType() && _gvn.type(right->as_InlineType()->get_is_init())->is_zero_type())) {
2141 // Null checking a scalarized but nullable inline type. Check the IsInit
2142 // input instead of the oop input to avoid keeping buffer allocations alive.
2143 Node* cmp = CmpI(left->as_InlineType()->get_is_init(), intcon(0));
2144 do_if(btest, cmp);
2145 return;
2146 } else {
2147 PreserveReexecuteState preexecs(this);
2148 inc_sp(2);
2149 jvms()->set_should_reexecute(true);
2150 left = left->as_InlineType()->buffer(this)->get_oop();
2151 }
2152 }
2153 if (right->is_InlineType()) {
2154 PreserveReexecuteState preexecs(this);
2155 inc_sp(2);
2156 jvms()->set_should_reexecute(true);
2157 right = right->as_InlineType()->buffer(this)->get_oop();
2158 }
2159
2160 // First, do a normal pointer comparison
2161 const TypeOopPtr* tleft = _gvn.type(left)->isa_oopptr();
2162 const TypeOopPtr* tright = _gvn.type(right)->isa_oopptr();
2163 Node* cmp = CmpP(left, right);
2164 cmp = optimize_cmp_with_klass(cmp);
2165 if (tleft == nullptr || !tleft->can_be_inline_type() ||
2166 tright == nullptr || !tright->can_be_inline_type()) {
2167 // This is sufficient, if one of the operands can't be an inline type
2168 do_if(btest, cmp);
2169 return;
2170 }
2171
2172 // Don't add traps to unstable if branches because additional checks are required to
2173 // decide if the operands are equal/substitutable and we therefore shouldn't prune
2174 // branches for one if based on the profiling of the acmp branches.
2175 // Also, OptimizeUnstableIf would set an incorrect re-rexecution state because it
2176 // assumes that there is a 1-1 mapping between the if and the acmp branches and that
2177 // hitting a trap means that we will take the corresponding acmp branch on re-execution.
2178 const bool can_trap = true;
2179
2180 Node* eq_region = nullptr;
2181 if (btest == BoolTest::eq) {
2182 do_if(btest, cmp, !can_trap, true);
2183 if (stopped()) {
2184 // Pointers are equal, operands must be equal
2185 return;
2186 }
2187 } else {
2188 assert(btest == BoolTest::ne, "only eq or ne");
2189 Node* is_not_equal = nullptr;
2190 eq_region = new RegionNode(3);
2191 {
2192 PreserveJVMState pjvms(this);
2193 // Pointers are not equal, but more checks are needed to determine if the operands are (not) substitutable
2194 do_if(btest, cmp, !can_trap, false, &is_not_equal);
2195 if (!stopped()) {
2196 eq_region->init_req(1, control());
2197 }
2198 }
2199 if (is_not_equal == nullptr || is_not_equal->is_top()) {
2200 record_for_igvn(eq_region);
2201 set_control(_gvn.transform(eq_region));
2202 return;
2203 }
2204 set_control(is_not_equal);
2205 }
2206
2207 // Prefer speculative types if available
2208 if (!too_many_traps_or_recompiles(Deoptimization::Reason_speculate_class_check)) {
2209 if (tleft->speculative_type() != nullptr) {
2210 left_type = tleft->speculative_type();
2211 }
2212 if (tright->speculative_type() != nullptr) {
2213 right_type = tright->speculative_type();
2214 }
2215 }
2216
2217 if (speculative_ptr_kind(tleft) != ProfileMaybeNull && speculative_ptr_kind(tleft) != ProfileUnknownNull) {
2218 ProfilePtrKind speculative_left_ptr = speculative_ptr_kind(tleft);
2219 if (speculative_left_ptr == ProfileAlwaysNull && !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_null_assert)) {
2220 left_ptr = speculative_left_ptr;
2221 } else if (speculative_left_ptr == ProfileNeverNull && !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_null_check)) {
2222 left_ptr = speculative_left_ptr;
2223 }
2224 }
2225 if (speculative_ptr_kind(tright) != ProfileMaybeNull && speculative_ptr_kind(tright) != ProfileUnknownNull) {
2226 ProfilePtrKind speculative_right_ptr = speculative_ptr_kind(tright);
2227 if (speculative_right_ptr == ProfileAlwaysNull && !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_null_assert)) {
2228 right_ptr = speculative_right_ptr;
2229 } else if (speculative_right_ptr == ProfileNeverNull && !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_null_check)) {
2230 right_ptr = speculative_right_ptr;
2231 }
2232 }
2233
2234 if (left_ptr == ProfileAlwaysNull) {
2235 // Comparison with null. Assert the input is indeed null and we're done.
2236 acmp_always_null_input(left, tleft, btest, eq_region);
2237 return;
2238 }
2239 if (right_ptr == ProfileAlwaysNull) {
2240 // Comparison with null. Assert the input is indeed null and we're done.
2241 acmp_always_null_input(right, tright, btest, eq_region);
2242 return;
2243 }
2244 if (left_type != nullptr && !left_type->is_inlinetype()) {
2245 // Comparison with an object of known type
2246 acmp_known_non_inline_type_input(left, tleft, left_ptr, left_type, btest, eq_region);
2247 return;
2248 }
2249 if (right_type != nullptr && !right_type->is_inlinetype()) {
2250 // Comparison with an object of known type
2251 acmp_known_non_inline_type_input(right, tright, right_ptr, right_type, btest, eq_region);
2252 return;
2253 }
2254 if (!left_inline_type) {
2255 // Comparison with an object known not to be an inline type
2256 acmp_unknown_non_inline_type_input(left, tleft, left_ptr, btest, eq_region);
2257 return;
2258 }
2259 if (!right_inline_type) {
2260 // Comparison with an object known not to be an inline type
2261 acmp_unknown_non_inline_type_input(right, tright, right_ptr, btest, eq_region);
2262 return;
2263 }
2264
2265 // Pointers are not equal, check if first operand is non-null
2266 Node* ne_region = new RegionNode(6);
2267 Node* null_ctl;
2268 Node* not_null_right = acmp_null_check(right, tright, right_ptr, null_ctl);
2269 ne_region->init_req(1, null_ctl);
2270
2271 // First operand is non-null, check if it is an inline type
2272 Node* is_value = inline_type_test(not_null_right);
2273 IfNode* is_value_iff = create_and_map_if(control(), is_value, PROB_FAIR, COUNT_UNKNOWN);
2274 Node* not_value = _gvn.transform(new IfFalseNode(is_value_iff));
2275 ne_region->init_req(2, not_value);
2276 set_control(_gvn.transform(new IfTrueNode(is_value_iff)));
2277
2278 // The first operand is an inline type, check if the second operand is non-null
2279 Node* not_null_left = acmp_null_check(left, tleft, left_ptr, null_ctl);
2280 ne_region->init_req(3, null_ctl);
2281
2282 // Check if both operands are of the same class.
2283 Node* kls_left = load_object_klass(not_null_left);
2284 Node* kls_right = load_object_klass(not_null_right);
2285 Node* kls_cmp = CmpP(kls_left, kls_right);
2286 Node* kls_bol = _gvn.transform(new BoolNode(kls_cmp, BoolTest::ne));
2287 IfNode* kls_iff = create_and_map_if(control(), kls_bol, PROB_FAIR, COUNT_UNKNOWN);
2288 Node* kls_ne = _gvn.transform(new IfTrueNode(kls_iff));
2289 set_control(_gvn.transform(new IfFalseNode(kls_iff)));
2290 ne_region->init_req(4, kls_ne);
2291
2292 if (stopped()) {
2293 record_for_igvn(ne_region);
2294 set_control(_gvn.transform(ne_region));
2295 if (btest == BoolTest::ne) {
2296 {
2297 PreserveJVMState pjvms(this);
2298 int target_bci = iter().get_dest();
2299 merge(target_bci);
2300 }
2301 record_for_igvn(eq_region);
2302 set_control(_gvn.transform(eq_region));
2303 }
2304 return;
2305 }
2306
2307 // Both operands are values types of the same class, we need to perform a
2308 // substitutability test. Delegate to ValueObjectMethods::isSubstitutable().
2309 Node* ne_io_phi = PhiNode::make(ne_region, i_o());
2310 Node* mem = reset_memory();
2311 Node* ne_mem_phi = PhiNode::make(ne_region, mem);
2312
2313 Node* eq_io_phi = nullptr;
2314 Node* eq_mem_phi = nullptr;
2315 if (eq_region != nullptr) {
2316 eq_io_phi = PhiNode::make(eq_region, i_o());
2317 eq_mem_phi = PhiNode::make(eq_region, mem);
2318 }
2319
2320 set_all_memory(mem);
2321
2322 kill_dead_locals();
2323 ciMethod* subst_method = ciEnv::current()->ValueObjectMethods_klass()->find_method(ciSymbols::isSubstitutable_name(), ciSymbols::object_object_boolean_signature());
2324 CallStaticJavaNode *call = new CallStaticJavaNode(C, TypeFunc::make(subst_method), SharedRuntime::get_resolve_static_call_stub(), subst_method);
2325 call->set_override_symbolic_info(true);
2326 call->init_req(TypeFunc::Parms, not_null_left);
2327 call->init_req(TypeFunc::Parms+1, not_null_right);
2328 inc_sp(2);
2329 set_edges_for_java_call(call, false, false);
2330 Node* ret = set_results_for_java_call(call, false, true);
2331 dec_sp(2);
2332
2333 // Test the return value of ValueObjectMethods::isSubstitutable()
2334 // This is the last check, do_if can emit traps now.
2335 Node* subst_cmp = _gvn.transform(new CmpINode(ret, intcon(1)));
2336 Node* ctl = C->top();
2337 if (btest == BoolTest::eq) {
2338 PreserveJVMState pjvms(this);
2339 do_if(btest, subst_cmp, can_trap);
2340 if (!stopped()) {
2341 ctl = control();
2342 }
2343 } else {
2344 assert(btest == BoolTest::ne, "only eq or ne");
2345 PreserveJVMState pjvms(this);
2346 do_if(btest, subst_cmp, can_trap, false, &ctl);
2347 if (!stopped()) {
2348 eq_region->init_req(2, control());
2349 eq_io_phi->init_req(2, i_o());
2350 eq_mem_phi->init_req(2, reset_memory());
2351 }
2352 }
2353 ne_region->init_req(5, ctl);
2354 ne_io_phi->init_req(5, i_o());
2355 ne_mem_phi->init_req(5, reset_memory());
2356
2357 record_for_igvn(ne_region);
2358 set_control(_gvn.transform(ne_region));
2359 set_i_o(_gvn.transform(ne_io_phi));
2360 set_all_memory(_gvn.transform(ne_mem_phi));
2361
2362 if (btest == BoolTest::ne) {
2363 {
2364 PreserveJVMState pjvms(this);
2365 int target_bci = iter().get_dest();
2366 merge(target_bci);
2367 }
2368
2369 record_for_igvn(eq_region);
2370 set_control(_gvn.transform(eq_region));
2371 set_i_o(_gvn.transform(eq_io_phi));
2372 set_all_memory(_gvn.transform(eq_mem_phi));
2373 }
2374 }
2375
2376 // Force unstable if traps to be taken randomly to trigger intermittent bugs such as incorrect debug information.
2377 // Add another if before the unstable if that checks a "random" condition at runtime (a simple shared counter) and
2378 // then either takes the trap or executes the original, unstable if.
2379 void Parse::stress_trap(IfNode* orig_iff, Node* counter, Node* incr_store) {
2380 // Search for an unstable if trap
2381 CallStaticJavaNode* trap = nullptr;
2382 assert(orig_iff->Opcode() == Op_If && orig_iff->outcnt() == 2, "malformed if");
2383 ProjNode* trap_proj = orig_iff->uncommon_trap_proj(trap, Deoptimization::Reason_unstable_if);
2384 if (trap == nullptr || !trap->jvms()->should_reexecute()) {
2385 // No suitable trap found. Remove unused counter load and increment.
2386 C->gvn_replace_by(incr_store, incr_store->in(MemNode::Memory));
2387 return;
2388 }
2389
2390 // Remove trap from optimization list since we add another path to the trap.
2391 bool success = C->remove_unstable_if_trap(trap, true);
2392 assert(success, "Trap already modified");
2393
2394 // Add a check before the original if that will trap with a certain frequency and execute the original if otherwise
2395 int freq_log = (C->random() % 31) + 1; // Random logarithmic frequency in [1, 31]
2396 Node* mask = intcon(right_n_bits(freq_log));
2397 counter = _gvn.transform(new AndINode(counter, mask));
2398 Node* cmp = _gvn.transform(new CmpINode(counter, intcon(0)));
2399 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::mask::eq));
2400 IfNode* iff = _gvn.transform(new IfNode(orig_iff->in(0), bol, orig_iff->_prob, orig_iff->_fcnt))->as_If();
2401 Node* if_true = _gvn.transform(new IfTrueNode(iff));
2402 Node* if_false = _gvn.transform(new IfFalseNode(iff));
2403 assert(!if_true->is_top() && !if_false->is_top(), "trap always / never taken");
2404
2405 // Trap
2406 assert(trap_proj->outcnt() == 1, "some other nodes are dependent on the trap projection");
2407
2408 Node* trap_region = new RegionNode(3);
2409 trap_region->set_req(1, trap_proj);
2410 trap_region->set_req(2, if_true);
2411 trap->set_req(0, _gvn.transform(trap_region));
2412
2413 // Don't trap, execute original if
2414 orig_iff->set_req(0, if_false);
2415 }
2416
2417 bool Parse::path_is_suitable_for_uncommon_trap(float prob) const {
2418 // Randomly skip emitting an uncommon trap
2419 if (StressUnstableIfTraps && ((C->random() % 2) == 0)) {
2420 return false;
2421 }
2422 // Don't want to speculate on uncommon traps when running with -Xcomp
2423 if (!UseInterpreter) {
2424 return false;
2425 }
2426 return seems_never_taken(prob) &&
2427 !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if);
2428 }
2429
2430 void Parse::maybe_add_predicate_after_if(Block* path) {
2431 if (path->is_SEL_head() && path->preds_parsed() == 0) {
2432 // Add predicates at bci of if dominating the loop so traps can be
2433 // recorded on the if's profile data
2434 int bc_depth = repush_if_args();
2435 add_parse_predicates();
2436 dec_sp(bc_depth);
2437 path->set_has_predicates();
2438 }
2439 }
2440
2441
2442 //----------------------------adjust_map_after_if------------------------------
2443 // Adjust the JVM state to reflect the result of taking this path.
2444 // Basically, it means inspecting the CmpNode controlling this
2445 // branch, seeing how it constrains a tested value, and then
2446 // deciding if it's worth our while to encode this constraint
2447 // as graph nodes in the current abstract interpretation map.
2448 void Parse::adjust_map_after_if(BoolTest::mask btest, Node* c, float prob, Block* path, bool can_trap) {
2449 if (!c->is_Cmp()) {
2450 maybe_add_predicate_after_if(path);
2451 return;
2452 }
2453
2454 if (stopped() || btest == BoolTest::illegal) {
2455 return; // nothing to do
2456 }
2457
2458 bool is_fallthrough = (path == successor_for_bci(iter().next_bci()));
2459
2460 if (can_trap && path_is_suitable_for_uncommon_trap(prob)) {
2461 repush_if_args();
2462 Node* call = uncommon_trap(Deoptimization::Reason_unstable_if,
2463 Deoptimization::Action_reinterpret,
2464 nullptr,
2465 (is_fallthrough ? "taken always" : "taken never"));
2466
2467 if (call != nullptr) {
2468 C->record_unstable_if_trap(new UnstableIfTrap(call->as_CallStaticJava(), path));
2469 }
2470 return;
2471 }
2472
2473 Node* val = c->in(1);
2474 Node* con = c->in(2);
2475 const Type* tcon = _gvn.type(con);
2476 const Type* tval = _gvn.type(val);
2477 bool have_con = tcon->singleton();
2478 if (tval->singleton()) {
2479 if (!have_con) {
2480 // Swap, so constant is in con.
2481 con = val;
2482 tcon = tval;
2483 val = c->in(2);
2484 tval = _gvn.type(val);
2485 btest = BoolTest(btest).commute();
2486 have_con = true;
2487 } else {
2488 // Do we have two constants? Then leave well enough alone.
2489 have_con = false;
2490 }
2491 }
2492 if (!have_con) { // remaining adjustments need a con
2493 maybe_add_predicate_after_if(path);
2494 return;
2495 }
2496
2497 sharpen_type_after_if(btest, con, tcon, val, tval);
2498 maybe_add_predicate_after_if(path);
2499 }
2500
2501
2502 static Node* extract_obj_from_klass_load(PhaseGVN* gvn, Node* n) {
2503 Node* ldk;
2504 if (n->is_DecodeNKlass()) {
2505 if (n->in(1)->Opcode() != Op_LoadNKlass) {
2506 return nullptr;
2507 } else {
2508 ldk = n->in(1);
2509 }
2510 } else if (n->Opcode() != Op_LoadKlass) {
2511 return nullptr;
2512 } else {
2513 ldk = n;
2514 }
2515 assert(ldk != nullptr && ldk->is_Load(), "should have found a LoadKlass or LoadNKlass node");
2516
2517 Node* adr = ldk->in(MemNode::Address);
2518 intptr_t off = 0;
2519 Node* obj = AddPNode::Ideal_base_and_offset(adr, gvn, off);
2520 if (obj == nullptr || off != oopDesc::klass_offset_in_bytes()) // loading oopDesc::_klass?
2521 return nullptr;
2522 const TypePtr* tp = gvn->type(obj)->is_ptr();
2523 if (tp == nullptr || !(tp->isa_instptr() || tp->isa_aryptr())) // is obj a Java object ptr?
2524 return nullptr;
2525
2526 return obj;
2527 }
2528
2529 void Parse::sharpen_type_after_if(BoolTest::mask btest,
2530 Node* con, const Type* tcon,
2531 Node* val, const Type* tval) {
2532 // Look for opportunities to sharpen the type of a node
2533 // whose klass is compared with a constant klass.
2534 if (btest == BoolTest::eq && tcon->isa_klassptr()) {
2535 Node* obj = extract_obj_from_klass_load(&_gvn, val);
2536 const TypeOopPtr* con_type = tcon->isa_klassptr()->as_instance_type();
2537 if (obj != nullptr && (con_type->isa_instptr() || con_type->isa_aryptr())) {
2538 // Found:
2539 // Bool(CmpP(LoadKlass(obj._klass), ConP(Foo.klass)), [eq])
2540 // or the narrowOop equivalent.
2541 const Type* obj_type = _gvn.type(obj);
2542 const TypeOopPtr* tboth = obj_type->join_speculative(con_type)->isa_oopptr();
2543 if (tboth != nullptr && tboth->klass_is_exact() && tboth != obj_type &&
2544 tboth->higher_equal(obj_type)) {
2545 // obj has to be of the exact type Foo if the CmpP succeeds.
2546 int obj_in_map = map()->find_edge(obj);
2547 JVMState* jvms = this->jvms();
2548 if (obj_in_map >= 0 &&
2549 (jvms->is_loc(obj_in_map) || jvms->is_stk(obj_in_map))) {
2550 TypeNode* ccast = new CheckCastPPNode(control(), obj, tboth);
2551 const Type* tcc = ccast->as_Type()->type();
2552 assert(tcc != obj_type && tcc->higher_equal(obj_type), "must improve");
2553 // Delay transform() call to allow recovery of pre-cast value
2554 // at the control merge.
2555 _gvn.set_type_bottom(ccast);
2556 record_for_igvn(ccast);
2557 if (tboth->is_inlinetypeptr()) {
2558 ccast = InlineTypeNode::make_from_oop(this, ccast, tboth->exact_klass(true)->as_inline_klass());
2559 }
2560 // Here's the payoff.
2561 replace_in_map(obj, ccast);
2562 }
2563 }
2564 }
2565 }
2566
2567 int val_in_map = map()->find_edge(val);
2568 if (val_in_map < 0) return; // replace_in_map would be useless
2569 {
2570 JVMState* jvms = this->jvms();
2571 if (!(jvms->is_loc(val_in_map) ||
2572 jvms->is_stk(val_in_map)))
2573 return; // again, it would be useless
2574 }
2575
2576 // Check for a comparison to a constant, and "know" that the compared
2577 // value is constrained on this path.
2578 assert(tcon->singleton(), "");
2579 ConstraintCastNode* ccast = nullptr;
2580 Node* cast = nullptr;
2581
2582 switch (btest) {
2583 case BoolTest::eq: // Constant test?
2584 {
2585 const Type* tboth = tcon->join_speculative(tval);
2586 if (tboth == tval) break; // Nothing to gain.
2587 if (tcon->isa_int()) {
2588 ccast = new CastIINode(control(), val, tboth);
2589 } else if (tcon == TypePtr::NULL_PTR) {
2590 // Cast to null, but keep the pointer identity temporarily live.
2591 ccast = new CastPPNode(control(), val, tboth);
2592 } else {
2593 const TypeF* tf = tcon->isa_float_constant();
2594 const TypeD* td = tcon->isa_double_constant();
2595 // Exclude tests vs float/double 0 as these could be
2596 // either +0 or -0. Just because you are equal to +0
2597 // doesn't mean you ARE +0!
2598 // Note, following code also replaces Long and Oop values.
2599 if ((!tf || tf->_f != 0.0) &&
2600 (!td || td->_d != 0.0))
2601 cast = con; // Replace non-constant val by con.
2602 }
2603 }
2604 break;
2605
2606 case BoolTest::ne:
2607 if (tcon == TypePtr::NULL_PTR) {
2608 cast = cast_not_null(val, false);
2609 }
2610 break;
2611
2612 default:
2613 // (At this point we could record int range types with CastII.)
2614 break;
2615 }
2616
2617 if (ccast != nullptr) {
2618 const Type* tcc = ccast->as_Type()->type();
2619 assert(tcc != tval && tcc->higher_equal(tval), "must improve");
2620 // Delay transform() call to allow recovery of pre-cast value
2621 // at the control merge.
2622 _gvn.set_type_bottom(ccast);
2623 record_for_igvn(ccast);
2624 cast = ccast;
2625 }
2626
2627 if (cast != nullptr) { // Here's the payoff.
2628 replace_in_map(val, cast);
2629 }
2630 }
2631
2632 /**
2633 * Use speculative type to optimize CmpP node: if comparison is
2634 * against the low level class, cast the object to the speculative
2635 * type if any. CmpP should then go away.
2636 *
2637 * @param c expected CmpP node
2638 * @return result of CmpP on object casted to speculative type
2639 *
2640 */
2641 Node* Parse::optimize_cmp_with_klass(Node* c) {
2642 // If this is transformed by the _gvn to a comparison with the low
2643 // level klass then we may be able to use speculation
2644 if (c->Opcode() == Op_CmpP &&
2645 (c->in(1)->Opcode() == Op_LoadKlass || c->in(1)->Opcode() == Op_DecodeNKlass) &&
2646 c->in(2)->is_Con()) {
2647 Node* load_klass = nullptr;
2648 Node* decode = nullptr;
2649 if (c->in(1)->Opcode() == Op_DecodeNKlass) {
2650 decode = c->in(1);
2651 load_klass = c->in(1)->in(1);
2652 } else {
2653 load_klass = c->in(1);
2654 }
2655 if (load_klass->in(2)->is_AddP()) {
2656 Node* addp = load_klass->in(2);
2657 Node* obj = addp->in(AddPNode::Address);
2658 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
2659 if (obj_type->speculative_type_not_null() != nullptr) {
2660 ciKlass* k = obj_type->speculative_type();
2661 inc_sp(2);
2662 obj = maybe_cast_profiled_obj(obj, k);
2663 dec_sp(2);
2664 if (obj->is_InlineType()) {
2665 assert(obj->as_InlineType()->is_allocated(&_gvn), "must be allocated");
2666 obj = obj->as_InlineType()->get_oop();
2667 }
2668 // Make the CmpP use the casted obj
2669 addp = basic_plus_adr(obj, addp->in(AddPNode::Offset));
2670 load_klass = load_klass->clone();
2671 load_klass->set_req(2, addp);
2672 load_klass = _gvn.transform(load_klass);
2673 if (decode != nullptr) {
2674 decode = decode->clone();
2675 decode->set_req(1, load_klass);
2676 load_klass = _gvn.transform(decode);
2677 }
2678 c = c->clone();
2679 c->set_req(1, load_klass);
2680 c = _gvn.transform(c);
2681 }
2682 }
2683 }
2684 return c;
2685 }
2686
2687 //------------------------------do_one_bytecode--------------------------------
2688 // Parse this bytecode, and alter the Parsers JVM->Node mapping
2689 void Parse::do_one_bytecode() {
2690 Node *a, *b, *c, *d; // Handy temps
2691 BoolTest::mask btest;
2692 int i;
2693
2694 assert(!has_exceptions(), "bytecode entry state must be clear of throws");
2695
2696 if (C->check_node_count(NodeLimitFudgeFactor * 5,
2697 "out of nodes parsing method")) {
2698 return;
2699 }
2700
2701 #ifdef ASSERT
2702 // for setting breakpoints
2703 if (TraceOptoParse) {
2704 tty->print(" @");
2705 dump_bci(bci());
2706 tty->print(" %s", Bytecodes::name(bc()));
2707 tty->cr();
2708 }
2709 #endif
2710
2711 switch (bc()) {
2712 case Bytecodes::_nop:
2713 // do nothing
2714 break;
2715 case Bytecodes::_lconst_0:
2716 push_pair(longcon(0));
2717 break;
2718
2719 case Bytecodes::_lconst_1:
2720 push_pair(longcon(1));
2721 break;
2722
2723 case Bytecodes::_fconst_0:
2724 push(zerocon(T_FLOAT));
2725 break;
2726
2727 case Bytecodes::_fconst_1:
2728 push(makecon(TypeF::ONE));
2729 break;
2730
2731 case Bytecodes::_fconst_2:
2732 push(makecon(TypeF::make(2.0f)));
2733 break;
2734
2735 case Bytecodes::_dconst_0:
2736 push_pair(zerocon(T_DOUBLE));
2737 break;
2738
2739 case Bytecodes::_dconst_1:
2740 push_pair(makecon(TypeD::ONE));
2741 break;
2742
2743 case Bytecodes::_iconst_m1:push(intcon(-1)); break;
2744 case Bytecodes::_iconst_0: push(intcon( 0)); break;
2745 case Bytecodes::_iconst_1: push(intcon( 1)); break;
2746 case Bytecodes::_iconst_2: push(intcon( 2)); break;
2747 case Bytecodes::_iconst_3: push(intcon( 3)); break;
2748 case Bytecodes::_iconst_4: push(intcon( 4)); break;
2749 case Bytecodes::_iconst_5: push(intcon( 5)); break;
2750 case Bytecodes::_bipush: push(intcon(iter().get_constant_u1())); break;
2751 case Bytecodes::_sipush: push(intcon(iter().get_constant_u2())); break;
2752 case Bytecodes::_aconst_null: push(null()); break;
2753
2754 case Bytecodes::_ldc:
2755 case Bytecodes::_ldc_w:
2756 case Bytecodes::_ldc2_w: {
2757 ciConstant constant = iter().get_constant();
2758 if (constant.is_loaded()) {
2759 const Type* con_type = Type::make_from_constant(constant);
2760 if (con_type != nullptr) {
2761 push_node(con_type->basic_type(), makecon(con_type));
2762 }
2763 } else {
2764 // If the constant is unresolved or in error state, run this BC in the interpreter.
2765 if (iter().is_in_error()) {
2766 uncommon_trap(Deoptimization::make_trap_request(Deoptimization::Reason_unhandled,
2767 Deoptimization::Action_none),
2768 nullptr, "constant in error state", true /* must_throw */);
2769
2770 } else {
2771 int index = iter().get_constant_pool_index();
2772 uncommon_trap(Deoptimization::make_trap_request(Deoptimization::Reason_unloaded,
2773 Deoptimization::Action_reinterpret,
2774 index),
2775 nullptr, "unresolved constant", false /* must_throw */);
2776 }
2777 }
2778 break;
2779 }
2780
2781 case Bytecodes::_aload_0:
2782 push( local(0) );
2783 break;
2784 case Bytecodes::_aload_1:
2785 push( local(1) );
2786 break;
2787 case Bytecodes::_aload_2:
2788 push( local(2) );
2789 break;
2790 case Bytecodes::_aload_3:
2791 push( local(3) );
2792 break;
2793 case Bytecodes::_aload:
2794 push( local(iter().get_index()) );
2795 break;
2796
2797 case Bytecodes::_fload_0:
2798 case Bytecodes::_iload_0:
2799 push( local(0) );
2800 break;
2801 case Bytecodes::_fload_1:
2802 case Bytecodes::_iload_1:
2803 push( local(1) );
2804 break;
2805 case Bytecodes::_fload_2:
2806 case Bytecodes::_iload_2:
2807 push( local(2) );
2808 break;
2809 case Bytecodes::_fload_3:
2810 case Bytecodes::_iload_3:
2811 push( local(3) );
2812 break;
2813 case Bytecodes::_fload:
2814 case Bytecodes::_iload:
2815 push( local(iter().get_index()) );
2816 break;
2817 case Bytecodes::_lload_0:
2818 push_pair_local( 0 );
2819 break;
2820 case Bytecodes::_lload_1:
2821 push_pair_local( 1 );
2822 break;
2823 case Bytecodes::_lload_2:
2824 push_pair_local( 2 );
2825 break;
2826 case Bytecodes::_lload_3:
2827 push_pair_local( 3 );
2828 break;
2829 case Bytecodes::_lload:
2830 push_pair_local( iter().get_index() );
2831 break;
2832
2833 case Bytecodes::_dload_0:
2834 push_pair_local(0);
2835 break;
2836 case Bytecodes::_dload_1:
2837 push_pair_local(1);
2838 break;
2839 case Bytecodes::_dload_2:
2840 push_pair_local(2);
2841 break;
2842 case Bytecodes::_dload_3:
2843 push_pair_local(3);
2844 break;
2845 case Bytecodes::_dload:
2846 push_pair_local(iter().get_index());
2847 break;
2848 case Bytecodes::_fstore_0:
2849 case Bytecodes::_istore_0:
2850 case Bytecodes::_astore_0:
2851 set_local( 0, pop() );
2852 break;
2853 case Bytecodes::_fstore_1:
2854 case Bytecodes::_istore_1:
2855 case Bytecodes::_astore_1:
2856 set_local( 1, pop() );
2857 break;
2858 case Bytecodes::_fstore_2:
2859 case Bytecodes::_istore_2:
2860 case Bytecodes::_astore_2:
2861 set_local( 2, pop() );
2862 break;
2863 case Bytecodes::_fstore_3:
2864 case Bytecodes::_istore_3:
2865 case Bytecodes::_astore_3:
2866 set_local( 3, pop() );
2867 break;
2868 case Bytecodes::_fstore:
2869 case Bytecodes::_istore:
2870 case Bytecodes::_astore:
2871 set_local( iter().get_index(), pop() );
2872 break;
2873 // long stores
2874 case Bytecodes::_lstore_0:
2875 set_pair_local( 0, pop_pair() );
2876 break;
2877 case Bytecodes::_lstore_1:
2878 set_pair_local( 1, pop_pair() );
2879 break;
2880 case Bytecodes::_lstore_2:
2881 set_pair_local( 2, pop_pair() );
2882 break;
2883 case Bytecodes::_lstore_3:
2884 set_pair_local( 3, pop_pair() );
2885 break;
2886 case Bytecodes::_lstore:
2887 set_pair_local( iter().get_index(), pop_pair() );
2888 break;
2889
2890 // double stores
2891 case Bytecodes::_dstore_0:
2892 set_pair_local( 0, dprecision_rounding(pop_pair()) );
2893 break;
2894 case Bytecodes::_dstore_1:
2895 set_pair_local( 1, dprecision_rounding(pop_pair()) );
2896 break;
2897 case Bytecodes::_dstore_2:
2898 set_pair_local( 2, dprecision_rounding(pop_pair()) );
2899 break;
2900 case Bytecodes::_dstore_3:
2901 set_pair_local( 3, dprecision_rounding(pop_pair()) );
2902 break;
2903 case Bytecodes::_dstore:
2904 set_pair_local( iter().get_index(), dprecision_rounding(pop_pair()) );
2905 break;
2906
2907 case Bytecodes::_pop: dec_sp(1); break;
2908 case Bytecodes::_pop2: dec_sp(2); break;
2909 case Bytecodes::_swap:
2910 a = pop();
2911 b = pop();
2912 push(a);
2913 push(b);
2914 break;
2915 case Bytecodes::_dup:
2916 a = pop();
2917 push(a);
2918 push(a);
2919 break;
2920 case Bytecodes::_dup_x1:
2921 a = pop();
2922 b = pop();
2923 push( a );
2924 push( b );
2925 push( a );
2926 break;
2927 case Bytecodes::_dup_x2:
2928 a = pop();
2929 b = pop();
2930 c = pop();
2931 push( a );
2932 push( c );
2933 push( b );
2934 push( a );
2935 break;
2936 case Bytecodes::_dup2:
2937 a = pop();
2938 b = pop();
2939 push( b );
2940 push( a );
2941 push( b );
2942 push( a );
2943 break;
2944
2945 case Bytecodes::_dup2_x1:
2946 // before: .. c, b, a
2947 // after: .. b, a, c, b, a
2948 // not tested
2949 a = pop();
2950 b = pop();
2951 c = pop();
2952 push( b );
2953 push( a );
2954 push( c );
2955 push( b );
2956 push( a );
2957 break;
2958 case Bytecodes::_dup2_x2:
2959 // before: .. d, c, b, a
2960 // after: .. b, a, d, c, b, a
2961 // not tested
2962 a = pop();
2963 b = pop();
2964 c = pop();
2965 d = pop();
2966 push( b );
2967 push( a );
2968 push( d );
2969 push( c );
2970 push( b );
2971 push( a );
2972 break;
2973
2974 case Bytecodes::_arraylength: {
2975 // Must do null-check with value on expression stack
2976 Node *ary = null_check(peek(), T_ARRAY);
2977 // Compile-time detect of null-exception?
2978 if (stopped()) return;
2979 a = pop();
2980 push(load_array_length(a));
2981 break;
2982 }
2983
2984 case Bytecodes::_baload: array_load(T_BYTE); break;
2985 case Bytecodes::_caload: array_load(T_CHAR); break;
2986 case Bytecodes::_iaload: array_load(T_INT); break;
2987 case Bytecodes::_saload: array_load(T_SHORT); break;
2988 case Bytecodes::_faload: array_load(T_FLOAT); break;
2989 case Bytecodes::_aaload: array_load(T_OBJECT); break;
2990 case Bytecodes::_laload: array_load(T_LONG); break;
2991 case Bytecodes::_daload: array_load(T_DOUBLE); break;
2992 case Bytecodes::_bastore: array_store(T_BYTE); break;
2993 case Bytecodes::_castore: array_store(T_CHAR); break;
2994 case Bytecodes::_iastore: array_store(T_INT); break;
2995 case Bytecodes::_sastore: array_store(T_SHORT); break;
2996 case Bytecodes::_fastore: array_store(T_FLOAT); break;
2997 case Bytecodes::_aastore: array_store(T_OBJECT); break;
2998 case Bytecodes::_lastore: array_store(T_LONG); break;
2999 case Bytecodes::_dastore: array_store(T_DOUBLE); break;
3000
3001 case Bytecodes::_getfield:
3002 do_getfield();
3003 break;
3004
3005 case Bytecodes::_getstatic:
3006 do_getstatic();
3007 break;
3008
3009 case Bytecodes::_putfield:
3010 do_putfield();
3011 break;
3012
3013 case Bytecodes::_putstatic:
3014 do_putstatic();
3015 break;
3016
3017 case Bytecodes::_irem:
3018 // Must keep both values on the expression-stack during null-check
3019 zero_check_int(peek());
3020 // Compile-time detect of null-exception?
3021 if (stopped()) return;
3022 b = pop();
3023 a = pop();
3024 push(_gvn.transform(new ModINode(control(), a, b)));
3025 break;
3026 case Bytecodes::_idiv:
3027 // Must keep both values on the expression-stack during null-check
3028 zero_check_int(peek());
3029 // Compile-time detect of null-exception?
3030 if (stopped()) return;
3031 b = pop();
3032 a = pop();
3033 push( _gvn.transform( new DivINode(control(),a,b) ) );
3034 break;
3035 case Bytecodes::_imul:
3036 b = pop(); a = pop();
3037 push( _gvn.transform( new MulINode(a,b) ) );
3038 break;
3039 case Bytecodes::_iadd:
3040 b = pop(); a = pop();
3041 push( _gvn.transform( new AddINode(a,b) ) );
3042 break;
3043 case Bytecodes::_ineg:
3044 a = pop();
3045 push( _gvn.transform( new SubINode(_gvn.intcon(0),a)) );
3046 break;
3047 case Bytecodes::_isub:
3048 b = pop(); a = pop();
3049 push( _gvn.transform( new SubINode(a,b) ) );
3050 break;
3051 case Bytecodes::_iand:
3052 b = pop(); a = pop();
3053 push( _gvn.transform( new AndINode(a,b) ) );
3054 break;
3055 case Bytecodes::_ior:
3056 b = pop(); a = pop();
3057 push( _gvn.transform( new OrINode(a,b) ) );
3058 break;
3059 case Bytecodes::_ixor:
3060 b = pop(); a = pop();
3061 push( _gvn.transform( new XorINode(a,b) ) );
3062 break;
3063 case Bytecodes::_ishl:
3064 b = pop(); a = pop();
3065 push( _gvn.transform( new LShiftINode(a,b) ) );
3066 break;
3067 case Bytecodes::_ishr:
3068 b = pop(); a = pop();
3069 push( _gvn.transform( new RShiftINode(a,b) ) );
3070 break;
3071 case Bytecodes::_iushr:
3072 b = pop(); a = pop();
3073 push( _gvn.transform( new URShiftINode(a,b) ) );
3074 break;
3075
3076 case Bytecodes::_fneg:
3077 a = pop();
3078 b = _gvn.transform(new NegFNode (a));
3079 push(b);
3080 break;
3081
3082 case Bytecodes::_fsub:
3083 b = pop();
3084 a = pop();
3085 c = _gvn.transform( new SubFNode(a,b) );
3086 d = precision_rounding(c);
3087 push( d );
3088 break;
3089
3090 case Bytecodes::_fadd:
3091 b = pop();
3092 a = pop();
3093 c = _gvn.transform( new AddFNode(a,b) );
3094 d = precision_rounding(c);
3095 push( d );
3096 break;
3097
3098 case Bytecodes::_fmul:
3099 b = pop();
3100 a = pop();
3101 c = _gvn.transform( new MulFNode(a,b) );
3102 d = precision_rounding(c);
3103 push( d );
3104 break;
3105
3106 case Bytecodes::_fdiv:
3107 b = pop();
3108 a = pop();
3109 c = _gvn.transform( new DivFNode(nullptr,a,b) );
3110 d = precision_rounding(c);
3111 push( d );
3112 break;
3113
3114 case Bytecodes::_frem:
3115 if (Matcher::has_match_rule(Op_ModF)) {
3116 // Generate a ModF node.
3117 b = pop();
3118 a = pop();
3119 c = _gvn.transform( new ModFNode(nullptr,a,b) );
3120 d = precision_rounding(c);
3121 push( d );
3122 }
3123 else {
3124 // Generate a call.
3125 modf();
3126 }
3127 break;
3128
3129 case Bytecodes::_fcmpl:
3130 b = pop();
3131 a = pop();
3132 c = _gvn.transform( new CmpF3Node( a, b));
3133 push(c);
3134 break;
3135 case Bytecodes::_fcmpg:
3136 b = pop();
3137 a = pop();
3138
3139 // Same as fcmpl but need to flip the unordered case. Swap the inputs,
3140 // which negates the result sign except for unordered. Flip the unordered
3141 // as well by using CmpF3 which implements unordered-lesser instead of
3142 // unordered-greater semantics. Finally, commute the result bits. Result
3143 // is same as using a CmpF3Greater except we did it with CmpF3 alone.
3144 c = _gvn.transform( new CmpF3Node( b, a));
3145 c = _gvn.transform( new SubINode(_gvn.intcon(0),c) );
3146 push(c);
3147 break;
3148
3149 case Bytecodes::_f2i:
3150 a = pop();
3151 push(_gvn.transform(new ConvF2INode(a)));
3152 break;
3153
3154 case Bytecodes::_d2i:
3155 a = pop_pair();
3156 b = _gvn.transform(new ConvD2INode(a));
3157 push( b );
3158 break;
3159
3160 case Bytecodes::_f2d:
3161 a = pop();
3162 b = _gvn.transform( new ConvF2DNode(a));
3163 push_pair( b );
3164 break;
3165
3166 case Bytecodes::_d2f:
3167 a = pop_pair();
3168 b = _gvn.transform( new ConvD2FNode(a));
3169 // This breaks _227_mtrt (speed & correctness) and _222_mpegaudio (speed)
3170 //b = _gvn.transform(new RoundFloatNode(nullptr, b) );
3171 push( b );
3172 break;
3173
3174 case Bytecodes::_l2f:
3175 if (Matcher::convL2FSupported()) {
3176 a = pop_pair();
3177 b = _gvn.transform( new ConvL2FNode(a));
3178 // For x86_32.ad, FILD doesn't restrict precision to 24 or 53 bits.
3179 // Rather than storing the result into an FP register then pushing
3180 // out to memory to round, the machine instruction that implements
3181 // ConvL2D is responsible for rounding.
3182 // c = precision_rounding(b);
3183 push(b);
3184 } else {
3185 l2f();
3186 }
3187 break;
3188
3189 case Bytecodes::_l2d:
3190 a = pop_pair();
3191 b = _gvn.transform( new ConvL2DNode(a));
3192 // For x86_32.ad, rounding is always necessary (see _l2f above).
3193 // c = dprecision_rounding(b);
3194 push_pair(b);
3195 break;
3196
3197 case Bytecodes::_f2l:
3198 a = pop();
3199 b = _gvn.transform( new ConvF2LNode(a));
3200 push_pair(b);
3201 break;
3202
3203 case Bytecodes::_d2l:
3204 a = pop_pair();
3205 b = _gvn.transform( new ConvD2LNode(a));
3206 push_pair(b);
3207 break;
3208
3209 case Bytecodes::_dsub:
3210 b = pop_pair();
3211 a = pop_pair();
3212 c = _gvn.transform( new SubDNode(a,b) );
3213 d = dprecision_rounding(c);
3214 push_pair( d );
3215 break;
3216
3217 case Bytecodes::_dadd:
3218 b = pop_pair();
3219 a = pop_pair();
3220 c = _gvn.transform( new AddDNode(a,b) );
3221 d = dprecision_rounding(c);
3222 push_pair( d );
3223 break;
3224
3225 case Bytecodes::_dmul:
3226 b = pop_pair();
3227 a = pop_pair();
3228 c = _gvn.transform( new MulDNode(a,b) );
3229 d = dprecision_rounding(c);
3230 push_pair( d );
3231 break;
3232
3233 case Bytecodes::_ddiv:
3234 b = pop_pair();
3235 a = pop_pair();
3236 c = _gvn.transform( new DivDNode(nullptr,a,b) );
3237 d = dprecision_rounding(c);
3238 push_pair( d );
3239 break;
3240
3241 case Bytecodes::_dneg:
3242 a = pop_pair();
3243 b = _gvn.transform(new NegDNode (a));
3244 push_pair(b);
3245 break;
3246
3247 case Bytecodes::_drem:
3248 if (Matcher::has_match_rule(Op_ModD)) {
3249 // Generate a ModD node.
3250 b = pop_pair();
3251 a = pop_pair();
3252 // a % b
3253
3254 c = _gvn.transform( new ModDNode(nullptr,a,b) );
3255 d = dprecision_rounding(c);
3256 push_pair( d );
3257 }
3258 else {
3259 // Generate a call.
3260 modd();
3261 }
3262 break;
3263
3264 case Bytecodes::_dcmpl:
3265 b = pop_pair();
3266 a = pop_pair();
3267 c = _gvn.transform( new CmpD3Node( a, b));
3268 push(c);
3269 break;
3270
3271 case Bytecodes::_dcmpg:
3272 b = pop_pair();
3273 a = pop_pair();
3274 // Same as dcmpl but need to flip the unordered case.
3275 // Commute the inputs, which negates the result sign except for unordered.
3276 // Flip the unordered as well by using CmpD3 which implements
3277 // unordered-lesser instead of unordered-greater semantics.
3278 // Finally, negate the result bits. Result is same as using a
3279 // CmpD3Greater except we did it with CmpD3 alone.
3280 c = _gvn.transform( new CmpD3Node( b, a));
3281 c = _gvn.transform( new SubINode(_gvn.intcon(0),c) );
3282 push(c);
3283 break;
3284
3285
3286 // Note for longs -> lo word is on TOS, hi word is on TOS - 1
3287 case Bytecodes::_land:
3288 b = pop_pair();
3289 a = pop_pair();
3290 c = _gvn.transform( new AndLNode(a,b) );
3291 push_pair(c);
3292 break;
3293 case Bytecodes::_lor:
3294 b = pop_pair();
3295 a = pop_pair();
3296 c = _gvn.transform( new OrLNode(a,b) );
3297 push_pair(c);
3298 break;
3299 case Bytecodes::_lxor:
3300 b = pop_pair();
3301 a = pop_pair();
3302 c = _gvn.transform( new XorLNode(a,b) );
3303 push_pair(c);
3304 break;
3305
3306 case Bytecodes::_lshl:
3307 b = pop(); // the shift count
3308 a = pop_pair(); // value to be shifted
3309 c = _gvn.transform( new LShiftLNode(a,b) );
3310 push_pair(c);
3311 break;
3312 case Bytecodes::_lshr:
3313 b = pop(); // the shift count
3314 a = pop_pair(); // value to be shifted
3315 c = _gvn.transform( new RShiftLNode(a,b) );
3316 push_pair(c);
3317 break;
3318 case Bytecodes::_lushr:
3319 b = pop(); // the shift count
3320 a = pop_pair(); // value to be shifted
3321 c = _gvn.transform( new URShiftLNode(a,b) );
3322 push_pair(c);
3323 break;
3324 case Bytecodes::_lmul:
3325 b = pop_pair();
3326 a = pop_pair();
3327 c = _gvn.transform( new MulLNode(a,b) );
3328 push_pair(c);
3329 break;
3330
3331 case Bytecodes::_lrem:
3332 // Must keep both values on the expression-stack during null-check
3333 assert(peek(0) == top(), "long word order");
3334 zero_check_long(peek(1));
3335 // Compile-time detect of null-exception?
3336 if (stopped()) return;
3337 b = pop_pair();
3338 a = pop_pair();
3339 c = _gvn.transform( new ModLNode(control(),a,b) );
3340 push_pair(c);
3341 break;
3342
3343 case Bytecodes::_ldiv:
3344 // Must keep both values on the expression-stack during null-check
3345 assert(peek(0) == top(), "long word order");
3346 zero_check_long(peek(1));
3347 // Compile-time detect of null-exception?
3348 if (stopped()) return;
3349 b = pop_pair();
3350 a = pop_pair();
3351 c = _gvn.transform( new DivLNode(control(),a,b) );
3352 push_pair(c);
3353 break;
3354
3355 case Bytecodes::_ladd:
3356 b = pop_pair();
3357 a = pop_pair();
3358 c = _gvn.transform( new AddLNode(a,b) );
3359 push_pair(c);
3360 break;
3361 case Bytecodes::_lsub:
3362 b = pop_pair();
3363 a = pop_pair();
3364 c = _gvn.transform( new SubLNode(a,b) );
3365 push_pair(c);
3366 break;
3367 case Bytecodes::_lcmp:
3368 // Safepoints are now inserted _before_ branches. The long-compare
3369 // bytecode painfully produces a 3-way value (-1,0,+1) which requires a
3370 // slew of control flow. These are usually followed by a CmpI vs zero and
3371 // a branch; this pattern then optimizes to the obvious long-compare and
3372 // branch. However, if the branch is backwards there's a Safepoint
3373 // inserted. The inserted Safepoint captures the JVM state at the
3374 // pre-branch point, i.e. it captures the 3-way value. Thus if a
3375 // long-compare is used to control a loop the debug info will force
3376 // computation of the 3-way value, even though the generated code uses a
3377 // long-compare and branch. We try to rectify the situation by inserting
3378 // a SafePoint here and have it dominate and kill the safepoint added at a
3379 // following backwards branch. At this point the JVM state merely holds 2
3380 // longs but not the 3-way value.
3381 switch (iter().next_bc()) {
3382 case Bytecodes::_ifgt:
3383 case Bytecodes::_iflt:
3384 case Bytecodes::_ifge:
3385 case Bytecodes::_ifle:
3386 case Bytecodes::_ifne:
3387 case Bytecodes::_ifeq:
3388 // If this is a backwards branch in the bytecodes, add Safepoint
3389 maybe_add_safepoint(iter().next_get_dest());
3390 default:
3391 break;
3392 }
3393 b = pop_pair();
3394 a = pop_pair();
3395 c = _gvn.transform( new CmpL3Node( a, b ));
3396 push(c);
3397 break;
3398
3399 case Bytecodes::_lneg:
3400 a = pop_pair();
3401 b = _gvn.transform( new SubLNode(longcon(0),a));
3402 push_pair(b);
3403 break;
3404 case Bytecodes::_l2i:
3405 a = pop_pair();
3406 push( _gvn.transform( new ConvL2INode(a)));
3407 break;
3408 case Bytecodes::_i2l:
3409 a = pop();
3410 b = _gvn.transform( new ConvI2LNode(a));
3411 push_pair(b);
3412 break;
3413 case Bytecodes::_i2b:
3414 // Sign extend
3415 a = pop();
3416 a = Compile::narrow_value(T_BYTE, a, nullptr, &_gvn, true);
3417 push(a);
3418 break;
3419 case Bytecodes::_i2s:
3420 a = pop();
3421 a = Compile::narrow_value(T_SHORT, a, nullptr, &_gvn, true);
3422 push(a);
3423 break;
3424 case Bytecodes::_i2c:
3425 a = pop();
3426 a = Compile::narrow_value(T_CHAR, a, nullptr, &_gvn, true);
3427 push(a);
3428 break;
3429
3430 case Bytecodes::_i2f:
3431 a = pop();
3432 b = _gvn.transform( new ConvI2FNode(a) ) ;
3433 c = precision_rounding(b);
3434 push (b);
3435 break;
3436
3437 case Bytecodes::_i2d:
3438 a = pop();
3439 b = _gvn.transform( new ConvI2DNode(a));
3440 push_pair(b);
3441 break;
3442
3443 case Bytecodes::_iinc: // Increment local
3444 i = iter().get_index(); // Get local index
3445 set_local( i, _gvn.transform( new AddINode( _gvn.intcon(iter().get_iinc_con()), local(i) ) ) );
3446 break;
3447
3448 // Exit points of synchronized methods must have an unlock node
3449 case Bytecodes::_return:
3450 return_current(nullptr);
3451 break;
3452
3453 case Bytecodes::_ireturn:
3454 case Bytecodes::_areturn:
3455 case Bytecodes::_freturn:
3456 return_current(pop());
3457 break;
3458 case Bytecodes::_lreturn:
3459 return_current(pop_pair());
3460 break;
3461 case Bytecodes::_dreturn:
3462 return_current(pop_pair());
3463 break;
3464
3465 case Bytecodes::_athrow:
3466 // null exception oop throws null pointer exception
3467 null_check(peek());
3468 if (stopped()) return;
3469 // Hook the thrown exception directly to subsequent handlers.
3470 if (BailoutToInterpreterForThrows) {
3471 // Keep method interpreted from now on.
3472 uncommon_trap(Deoptimization::Reason_unhandled,
3473 Deoptimization::Action_make_not_compilable);
3474 return;
3475 }
3476 if (env()->jvmti_can_post_on_exceptions()) {
3477 // check if we must post exception events, take uncommon trap if so (with must_throw = false)
3478 uncommon_trap_if_should_post_on_exceptions(Deoptimization::Reason_unhandled, false);
3479 }
3480 // Here if either can_post_on_exceptions or should_post_on_exceptions is false
3481 add_exception_state(make_exception_state(peek()));
3482 break;
3483
3484 case Bytecodes::_goto: // fall through
3485 case Bytecodes::_goto_w: {
3486 int target_bci = (bc() == Bytecodes::_goto) ? iter().get_dest() : iter().get_far_dest();
3487
3488 // If this is a backwards branch in the bytecodes, add Safepoint
3489 maybe_add_safepoint(target_bci);
3490
3491 // Merge the current control into the target basic block
3492 merge(target_bci);
3493
3494 // See if we can get some profile data and hand it off to the next block
3495 Block *target_block = block()->successor_for_bci(target_bci);
3496 if (target_block->pred_count() != 1) break;
3497 ciMethodData* methodData = method()->method_data();
3498 if (!methodData->is_mature()) break;
3499 ciProfileData* data = methodData->bci_to_data(bci());
3500 assert(data != nullptr && data->is_JumpData(), "need JumpData for taken branch");
3501 int taken = ((ciJumpData*)data)->taken();
3502 taken = method()->scale_count(taken);
3503 target_block->set_count(taken);
3504 break;
3505 }
3506
3507 case Bytecodes::_ifnull: btest = BoolTest::eq; goto handle_if_null;
3508 case Bytecodes::_ifnonnull: btest = BoolTest::ne; goto handle_if_null;
3509 handle_if_null:
3510 // If this is a backwards branch in the bytecodes, add Safepoint
3511 maybe_add_safepoint(iter().get_dest());
3512 a = null();
3513 b = pop();
3514 if (b->is_InlineType()) {
3515 // Null checking a scalarized but nullable inline type. Check the IsInit
3516 // input instead of the oop input to avoid keeping buffer allocations alive
3517 c = _gvn.transform(new CmpINode(b->as_InlineType()->get_is_init(), zerocon(T_INT)));
3518 } else {
3519 if (!_gvn.type(b)->speculative_maybe_null() &&
3520 !too_many_traps(Deoptimization::Reason_speculate_null_check)) {
3521 inc_sp(1);
3522 Node* null_ctl = top();
3523 b = null_check_oop(b, &null_ctl, true, true, true);
3524 assert(null_ctl->is_top(), "no null control here");
3525 dec_sp(1);
3526 } else if (_gvn.type(b)->speculative_always_null() &&
3527 !too_many_traps(Deoptimization::Reason_speculate_null_assert)) {
3528 inc_sp(1);
3529 b = null_assert(b);
3530 dec_sp(1);
3531 }
3532 c = _gvn.transform( new CmpPNode(b, a) );
3533 }
3534 do_ifnull(btest, c);
3535 break;
3536
3537 case Bytecodes::_if_acmpeq: btest = BoolTest::eq; goto handle_if_acmp;
3538 case Bytecodes::_if_acmpne: btest = BoolTest::ne; goto handle_if_acmp;
3539 handle_if_acmp:
3540 // If this is a backwards branch in the bytecodes, add Safepoint
3541 maybe_add_safepoint(iter().get_dest());
3542 a = pop();
3543 b = pop();
3544 do_acmp(btest, b, a);
3545 break;
3546
3547 case Bytecodes::_ifeq: btest = BoolTest::eq; goto handle_ifxx;
3548 case Bytecodes::_ifne: btest = BoolTest::ne; goto handle_ifxx;
3549 case Bytecodes::_iflt: btest = BoolTest::lt; goto handle_ifxx;
3550 case Bytecodes::_ifle: btest = BoolTest::le; goto handle_ifxx;
3551 case Bytecodes::_ifgt: btest = BoolTest::gt; goto handle_ifxx;
3552 case Bytecodes::_ifge: btest = BoolTest::ge; goto handle_ifxx;
3553 handle_ifxx:
3554 // If this is a backwards branch in the bytecodes, add Safepoint
3555 maybe_add_safepoint(iter().get_dest());
3556 a = _gvn.intcon(0);
3557 b = pop();
3558 c = _gvn.transform( new CmpINode(b, a) );
3559 do_if(btest, c);
3560 break;
3561
3562 case Bytecodes::_if_icmpeq: btest = BoolTest::eq; goto handle_if_icmp;
3563 case Bytecodes::_if_icmpne: btest = BoolTest::ne; goto handle_if_icmp;
3564 case Bytecodes::_if_icmplt: btest = BoolTest::lt; goto handle_if_icmp;
3565 case Bytecodes::_if_icmple: btest = BoolTest::le; goto handle_if_icmp;
3566 case Bytecodes::_if_icmpgt: btest = BoolTest::gt; goto handle_if_icmp;
3567 case Bytecodes::_if_icmpge: btest = BoolTest::ge; goto handle_if_icmp;
3568 handle_if_icmp:
3569 // If this is a backwards branch in the bytecodes, add Safepoint
3570 maybe_add_safepoint(iter().get_dest());
3571 a = pop();
3572 b = pop();
3573 c = _gvn.transform( new CmpINode( b, a ) );
3574 do_if(btest, c);
3575 break;
3576
3577 case Bytecodes::_tableswitch:
3578 do_tableswitch();
3579 break;
3580
3581 case Bytecodes::_lookupswitch:
3582 do_lookupswitch();
3583 break;
3584
3585 case Bytecodes::_invokestatic:
3586 case Bytecodes::_invokedynamic:
3587 case Bytecodes::_invokespecial:
3588 case Bytecodes::_invokevirtual:
3589 case Bytecodes::_invokeinterface:
3590 do_call();
3591 break;
3592 case Bytecodes::_checkcast:
3593 do_checkcast();
3594 break;
3595 case Bytecodes::_instanceof:
3596 do_instanceof();
3597 break;
3598 case Bytecodes::_anewarray:
3599 do_newarray();
3600 break;
3601 case Bytecodes::_newarray:
3602 do_newarray((BasicType)iter().get_index());
3603 break;
3604 case Bytecodes::_multianewarray:
3605 do_multianewarray();
3606 break;
3607 case Bytecodes::_new:
3608 do_new();
3609 break;
3610
3611 case Bytecodes::_jsr:
3612 case Bytecodes::_jsr_w:
3613 do_jsr();
3614 break;
3615
3616 case Bytecodes::_ret:
3617 do_ret();
3618 break;
3619
3620
3621 case Bytecodes::_monitorenter:
3622 do_monitor_enter();
3623 break;
3624
3625 case Bytecodes::_monitorexit:
3626 do_monitor_exit();
3627 break;
3628
3629 case Bytecodes::_breakpoint:
3630 // Breakpoint set concurrently to compile
3631 // %%% use an uncommon trap?
3632 C->record_failure("breakpoint in method");
3633 return;
3634
3635 default:
3636 #ifndef PRODUCT
3637 map()->dump(99);
3638 #endif
3639 tty->print("\nUnhandled bytecode %s\n", Bytecodes::name(bc()) );
3640 ShouldNotReachHere();
3641 }
3642
3643 #ifndef PRODUCT
3644 if (failing()) { return; }
3645 constexpr int perBytecode = 6;
3646 if (C->should_print_igv(perBytecode)) {
3647 IdealGraphPrinter* printer = C->igv_printer();
3648 char buffer[256];
3649 jio_snprintf(buffer, sizeof(buffer), "Bytecode %d: %s", bci(), Bytecodes::name(bc()));
3650 bool old = printer->traverse_outs();
3651 printer->set_traverse_outs(true);
3652 printer->print_method(buffer, perBytecode);
3653 printer->set_traverse_outs(old);
3654 }
3655 #endif
3656 }