1 /*
2 * Copyright (c) 1998, 2022, 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 "jvm_io.h"
27 #include "ci/ciMethodData.hpp"
28 #include "ci/ciSymbols.hpp"
29 #include "classfile/vmSymbols.hpp"
30 #include "compiler/compileLog.hpp"
31 #include "interpreter/linkResolver.hpp"
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 int 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 = NULL;
60 ciKlass* element_type = NULL;
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 != NULL || 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 flattened inline type array
87 Node* vt = InlineTypeNode::make_from_flattened(this, elemtype->inline_klass(), ary, adr);
88 push(vt);
89 return;
90 } else if (ary_t->is_null_free()) {
91 // Load from non-flattened inline type array (elements can never be null)
92 bt = T_PRIMITIVE_OBJECT;
93 } else if (!ary_t->is_not_flat()) {
94 // Cannot statically determine if array is flattened, emit runtime check
95 assert(UseFlatArray && is_reference_type(bt) && elemptr->can_be_inline_type() && !ary_t->klass_is_exact() && !ary_t->is_not_null_free() &&
96 (!elemptr->is_inlinetypeptr() || elemptr->inline_klass()->flatten_array()), "array can't be flattened");
97 IdealKit ideal(this);
98 IdealVariable res(ideal);
99 ideal.declarations_done();
100 ideal.if_then(flat_array_test(ary, /* flat = */ false)); {
101 // non-flattened
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 Node* ld = access_load_at(ary, adr, adr_type, elemptr, bt,
106 IN_HEAP | IS_ARRAY | C2_CONTROL_DEPENDENT_LOAD);
107 if (elemptr->is_inlinetypeptr()) {
108 assert(elemptr->maybe_null(), "null free array should be handled above");
109 ld = InlineTypeNode::make_from_oop(this, ld, elemptr->inline_klass(), false);
110 }
111 ideal.sync_kit(this);
112 ideal.set(res, ld);
113 } ideal.else_(); {
114 // flattened
115 sync_kit(ideal);
116 if (elemptr->is_inlinetypeptr()) {
117 // Element type is known, cast and load from flattened representation
118 ciInlineKlass* vk = elemptr->inline_klass();
119 assert(vk->flatten_array() && elemptr->maybe_null(), "never/always flat - should be optimized");
120 ciArrayKlass* array_klass = ciArrayKlass::make(vk, /* null_free */ true);
121 const TypeAryPtr* arytype = TypeOopPtr::make_from_klass(array_klass)->isa_aryptr();
122 Node* cast = _gvn.transform(new CheckCastPPNode(control(), ary, arytype));
123 Node* casted_adr = array_element_address(cast, idx, T_PRIMITIVE_OBJECT, ary_t->size(), control());
124 // Re-execute flattened array load if buffering triggers deoptimization
125 PreserveReexecuteState preexecs(this);
126 jvms()->set_should_reexecute(true);
127 inc_sp(2);
128 Node* vt = InlineTypeNode::make_from_flattened(this, vk, cast, casted_adr)->buffer(this, false);
129 ideal.set(res, vt);
130 ideal.sync_kit(this);
131 } else {
132 // Element type is unknown, emit runtime call
133
134 // Below membars keep this access to an unknown flattened array correctly
135 // ordered with other unknown and known flattened array accesses.
136 insert_mem_bar_volatile(Op_MemBarCPUOrder, C->get_alias_index(TypeAryPtr::INLINES));
137
138 Node* call = NULL;
139 {
140 // Re-execute flattened array load if runtime call triggers deoptimization
141 PreserveReexecuteState preexecs(this);
142 jvms()->set_bci(_bci);
143 jvms()->set_should_reexecute(true);
144 inc_sp(2);
145 kill_dead_locals();
146 call = make_runtime_call(RC_NO_LEAF | RC_NO_IO,
147 OptoRuntime::load_unknown_inline_type(),
148 OptoRuntime::load_unknown_inline_Java(),
149 NULL, TypeRawPtr::BOTTOM,
150 ary, idx);
151 }
152 make_slow_call_ex(call, env()->Throwable_klass(), false);
153 Node* buffer = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
154
155 insert_mem_bar_volatile(Op_MemBarCPUOrder, C->get_alias_index(TypeAryPtr::INLINES));
156
157 // Keep track of the information that the inline type is flattened in arrays
158 const Type* unknown_value = elemptr->is_instptr()->cast_to_flatten_array();
159 buffer = _gvn.transform(new CheckCastPPNode(control(), buffer, unknown_value));
160
161 ideal.sync_kit(this);
162 ideal.set(res, buffer);
163 }
164 } ideal.end_if();
165 sync_kit(ideal);
166 Node* ld = _gvn.transform(ideal.value(res));
167 ld = record_profile_for_speculation_at_array_load(ld);
168 push_node(bt, ld);
169 return;
170 }
171
172 if (elemtype == TypeInt::BOOL) {
173 bt = T_BOOLEAN;
174 }
175 const TypeAryPtr* adr_type = TypeAryPtr::get_array_body_type(bt);
176 Node* ld = access_load_at(ary, adr, adr_type, elemtype, bt,
177 IN_HEAP | IS_ARRAY | C2_CONTROL_DEPENDENT_LOAD);
178 ld = record_profile_for_speculation_at_array_load(ld);
179 // Loading a non-flattened inline type
180 if (elemptr != NULL && elemptr->is_inlinetypeptr()) {
181 assert(!ary_t->is_null_free() || !elemptr->maybe_null(), "inline type array elements should never be null");
182 ld = InlineTypeNode::make_from_oop(this, ld, elemptr->inline_klass(), !elemptr->maybe_null());
183 }
184 push_node(bt, ld);
185 }
186
187
188 //--------------------------------array_store----------------------------------
189 void Parse::array_store(BasicType bt) {
190 const Type* elemtype = Type::TOP;
191 Node* adr = array_addressing(bt, type2size[bt], elemtype);
192 if (stopped()) return; // guaranteed null or range check
193 Node* cast_val = NULL;
194 if (bt == T_OBJECT) {
195 cast_val = array_store_check(adr, elemtype);
196 if (stopped()) return;
197 }
198 Node* val = pop_node(bt); // Value to store
199 Node* idx = pop(); // Index in the array
200 Node* ary = pop(); // The array itself
201
202 const TypeAryPtr* ary_t = _gvn.type(ary)->is_aryptr();
203 const TypeAryPtr* adr_type = TypeAryPtr::get_array_body_type(bt);
204 assert(adr->as_AddP()->in(AddPNode::Base) == ary, "inconsistent address base");
205
206 if (elemtype == TypeInt::BOOL) {
207 bt = T_BOOLEAN;
208 } else if (bt == T_OBJECT) {
209 elemtype = elemtype->make_oopptr();
210 const Type* tval = _gvn.type(cast_val);
211 // We may have lost type information for 'val' here due to the casts
212 // emitted by the array_store_check code (see JDK-6312651)
213 // TODO Remove this code once JDK-6312651 is in.
214 const Type* tval_init = _gvn.type(val);
215 // Based on the value to be stored, try to determine if the array is not null-free and/or not flat.
216 // This is only legal for non-null stores because the array_store_check always passes for null, even
217 // if the array is null-free. Null stores are handled in GraphKit::gen_inline_array_null_guard().
218 bool not_inline = !tval->isa_inlinetype() &&
219 ((!tval_init->maybe_null() && !tval_init->is_oopptr()->can_be_inline_type()) ||
220 (!tval->maybe_null() && !tval->is_oopptr()->can_be_inline_type()));
221 bool not_flattened = not_inline || ((tval_init->is_inlinetypeptr() || tval_init->isa_inlinetype()) && !tval_init->inline_klass()->flatten_array());
222 if (!ary_t->is_not_null_free() && not_inline) {
223 // Storing a non-inline type, mark array as not null-free (-> not flat).
224 ary_t = ary_t->cast_to_not_null_free();
225 Node* cast = _gvn.transform(new CheckCastPPNode(control(), ary, ary_t));
226 replace_in_map(ary, cast);
227 ary = cast;
228 } else if (!ary_t->is_not_flat() && not_flattened) {
229 // Storing a non-flattened value, mark array as not flat.
230 ary_t = ary_t->cast_to_not_flat();
231 Node* cast = _gvn.transform(new CheckCastPPNode(control(), ary, ary_t));
232 replace_in_map(ary, cast);
233 ary = cast;
234 }
235
236 if (ary_t->is_flat()) {
237 // Store to flattened inline type array
238 assert(!tval->maybe_null(), "should be guaranteed by array store check");
239 // Re-execute flattened array store if buffering triggers deoptimization
240 PreserveReexecuteState preexecs(this);
241 inc_sp(3);
242 jvms()->set_should_reexecute(true);
243 cast_val->as_InlineTypeBase()->store_flattened(this, ary, adr, NULL, 0, MO_UNORDERED | IN_HEAP | IS_ARRAY);
244 return;
245 } else if (ary_t->is_null_free()) {
246 // Store to non-flattened inline type array (elements can never be null)
247 assert(!tval->maybe_null(), "should be guaranteed by array store check");
248 if (elemtype->inline_klass()->is_empty()) {
249 // Ignore empty inline stores, array is already initialized.
250 return;
251 }
252 } else if (!ary_t->is_not_flat() && (tval != TypePtr::NULL_PTR || StressReflectiveCode)) {
253 // Array might be flattened, emit runtime checks (for NULL, a simple inline_array_null_guard is sufficient).
254 assert(UseFlatArray && !not_flattened && elemtype->is_oopptr()->can_be_inline_type() &&
255 !ary_t->klass_is_exact() && !ary_t->is_not_null_free(), "array can't be flattened");
256 IdealKit ideal(this);
257 ideal.if_then(flat_array_test(ary, /* flat = */ false)); {
258 // non-flattened
259 assert(ideal.ctrl()->in(0)->as_If()->is_flat_array_check(&_gvn), "Should be found");
260 sync_kit(ideal);
261 Node* cast_ary = inline_array_null_guard(ary, cast_val, 3);
262 inc_sp(3);
263 access_store_at(cast_ary, adr, adr_type, cast_val, elemtype, bt, MO_UNORDERED | IN_HEAP | IS_ARRAY, false);
264 dec_sp(3);
265 ideal.sync_kit(this);
266 } ideal.else_(); {
267 sync_kit(ideal);
268 // flattened
269 Node* null_ctl = top();
270 Node* val = null_check_oop(cast_val, &null_ctl);
271 if (null_ctl != top()) {
272 PreserveJVMState pjvms(this);
273 inc_sp(3);
274 set_control(null_ctl);
275 uncommon_trap(Deoptimization::Reason_null_check, Deoptimization::Action_none);
276 dec_sp(3);
277 }
278 // Try to determine the inline klass
279 ciInlineKlass* vk = NULL;
280 if (tval->isa_inlinetype() || tval->is_inlinetypeptr()) {
281 vk = tval->inline_klass();
282 } else if (tval_init->isa_inlinetype() || tval_init->is_inlinetypeptr()) {
283 vk = tval_init->inline_klass();
284 } else if (elemtype->is_inlinetypeptr()) {
285 vk = elemtype->inline_klass();
286 }
287 Node* casted_ary = ary;
288 if (vk != NULL && !stopped()) {
289 // Element type is known, cast and store to flattened representation
290 assert(vk->flatten_array() && elemtype->maybe_null(), "never/always flat - should be optimized");
291 ciArrayKlass* array_klass = ciArrayKlass::make(vk, /* null_free */ true);
292 const TypeAryPtr* arytype = TypeOopPtr::make_from_klass(array_klass)->isa_aryptr();
293 casted_ary = _gvn.transform(new CheckCastPPNode(control(), casted_ary, arytype));
294 Node* casted_adr = array_element_address(casted_ary, idx, T_OBJECT, arytype->size(), control());
295 if (!val->is_InlineType()) {
296 assert(!gvn().type(val)->maybe_null(), "inline type array elements should never be null");
297 val = InlineTypeNode::make_from_oop(this, val, vk);
298 }
299 // Re-execute flattened array store if buffering triggers deoptimization
300 PreserveReexecuteState preexecs(this);
301 inc_sp(3);
302 jvms()->set_should_reexecute(true);
303 val->as_InlineTypeBase()->store_flattened(this, casted_ary, casted_adr, NULL, 0, MO_UNORDERED | IN_HEAP | IS_ARRAY);
304 } else if (!stopped()) {
305 // Element type is unknown, emit runtime call
306
307 // Below membars keep this access to an unknown flattened array correctly
308 // ordered with other unknown and known flattened array accesses.
309 insert_mem_bar_volatile(Op_MemBarCPUOrder, C->get_alias_index(TypeAryPtr::INLINES));
310
311 make_runtime_call(RC_LEAF,
312 OptoRuntime::store_unknown_inline_type(),
313 CAST_FROM_FN_PTR(address, OptoRuntime::store_unknown_inline),
314 "store_unknown_inline", TypeRawPtr::BOTTOM,
315 val, casted_ary, idx);
316
317 insert_mem_bar_volatile(Op_MemBarCPUOrder, C->get_alias_index(TypeAryPtr::INLINES));
318 }
319 ideal.sync_kit(this);
320 }
321 ideal.end_if();
322 sync_kit(ideal);
323 return;
324 } else if (!ary_t->is_not_null_free()) {
325 // Array is not flattened but may be null free
326 assert(elemtype->is_oopptr()->can_be_inline_type() && !ary_t->klass_is_exact(), "array can't be null-free");
327 ary = inline_array_null_guard(ary, cast_val, 3, true);
328 }
329 }
330 inc_sp(3);
331 access_store_at(ary, adr, adr_type, val, elemtype, bt, MO_UNORDERED | IN_HEAP | IS_ARRAY);
332 dec_sp(3);
333 }
334
335
336 //------------------------------array_addressing-------------------------------
337 // Pull array and index from the stack. Compute pointer-to-element.
338 Node* Parse::array_addressing(BasicType type, int vals, const Type*& elemtype) {
339 Node *idx = peek(0+vals); // Get from stack without popping
340 Node *ary = peek(1+vals); // in case of exception
341
342 // Null check the array base, with correct stack contents
343 ary = null_check(ary, T_ARRAY);
344 // Compile-time detect of null-exception?
345 if (stopped()) return top();
346
347 const TypeAryPtr* arytype = _gvn.type(ary)->is_aryptr();
348 const TypeInt* sizetype = arytype->size();
349 elemtype = arytype->elem();
350
351 if (UseUniqueSubclasses) {
352 const Type* el = elemtype->make_ptr();
353 if (el && el->isa_instptr()) {
354 const TypeInstPtr* toop = el->is_instptr();
355 if (toop->instance_klass()->unique_concrete_subklass()) {
356 // If we load from "AbstractClass[]" we must see "ConcreteSubClass".
357 const Type* subklass = Type::get_const_type(toop->instance_klass());
358 elemtype = subklass->join_speculative(el);
359 }
360 }
361 }
362
363 // Check for big class initializers with all constant offsets
364 // feeding into a known-size array.
365 const TypeInt* idxtype = _gvn.type(idx)->is_int();
366 // See if the highest idx value is less than the lowest array bound,
367 // and if the idx value cannot be negative:
368 bool need_range_check = true;
369 if (idxtype->_hi < sizetype->_lo && idxtype->_lo >= 0) {
370 need_range_check = false;
371 if (C->log() != NULL) C->log()->elem("observe that='!need_range_check'");
372 }
373
374 if (!arytype->is_loaded()) {
375 // Only fails for some -Xcomp runs
376 // The class is unloaded. We have to run this bytecode in the interpreter.
377 ciKlass* klass = arytype->unloaded_klass();
378
379 uncommon_trap(Deoptimization::Reason_unloaded,
380 Deoptimization::Action_reinterpret,
381 klass, "!loaded array");
382 return top();
383 }
384
385 // Do the range check
386 if (GenerateRangeChecks && need_range_check) {
387 Node* tst;
388 if (sizetype->_hi <= 0) {
389 // The greatest array bound is negative, so we can conclude that we're
390 // compiling unreachable code, but the unsigned compare trick used below
391 // only works with non-negative lengths. Instead, hack "tst" to be zero so
392 // the uncommon_trap path will always be taken.
393 tst = _gvn.intcon(0);
394 } else {
395 // Range is constant in array-oop, so we can use the original state of mem
396 Node* len = load_array_length(ary);
397
398 // Test length vs index (standard trick using unsigned compare)
399 Node* chk = _gvn.transform( new CmpUNode(idx, len) );
400 BoolTest::mask btest = BoolTest::lt;
401 tst = _gvn.transform( new BoolNode(chk, btest) );
402 }
403 RangeCheckNode* rc = new RangeCheckNode(control(), tst, PROB_MAX, COUNT_UNKNOWN);
404 _gvn.set_type(rc, rc->Value(&_gvn));
405 if (!tst->is_Con()) {
406 record_for_igvn(rc);
407 }
408 set_control(_gvn.transform(new IfTrueNode(rc)));
409 // Branch to failure if out of bounds
410 {
411 PreserveJVMState pjvms(this);
412 set_control(_gvn.transform(new IfFalseNode(rc)));
413 if (C->allow_range_check_smearing()) {
414 // Do not use builtin_throw, since range checks are sometimes
415 // made more stringent by an optimistic transformation.
416 // This creates "tentative" range checks at this point,
417 // which are not guaranteed to throw exceptions.
418 // See IfNode::Ideal, is_range_check, adjust_check.
419 uncommon_trap(Deoptimization::Reason_range_check,
420 Deoptimization::Action_make_not_entrant,
421 NULL, "range_check");
422 } else {
423 // If we have already recompiled with the range-check-widening
424 // heroic optimization turned off, then we must really be throwing
425 // range check exceptions.
426 builtin_throw(Deoptimization::Reason_range_check);
427 }
428 }
429 }
430 // Check for always knowing you are throwing a range-check exception
431 if (stopped()) return top();
432
433 // This could be an access to an inline type array. We can't tell if it's
434 // flat or not. Knowing the exact type avoids runtime checks and leads to
435 // a much simpler graph shape. Check profile information.
436 if (!arytype->is_flat() && !arytype->is_not_flat()) {
437 // First check the speculative type
438 Deoptimization::DeoptReason reason = Deoptimization::Reason_speculate_class_check;
439 ciKlass* array_type = arytype->speculative_type();
440 if (too_many_traps_or_recompiles(reason) || array_type == NULL) {
441 // No speculative type, check profile data at this bci
442 array_type = NULL;
443 reason = Deoptimization::Reason_class_check;
444 if (UseArrayLoadStoreProfile && !too_many_traps_or_recompiles(reason)) {
445 ciKlass* element_type = NULL;
446 ProfilePtrKind element_ptr = ProfileMaybeNull;
447 bool flat_array = true;
448 bool null_free_array = true;
449 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
450 }
451 }
452 if (array_type != NULL) {
453 // Speculate that this array has the exact type reported by profile data
454 Node* better_ary = NULL;
455 DEBUG_ONLY(Node* old_control = control();)
456 Node* slow_ctl = type_check_receiver(ary, array_type, 1.0, &better_ary);
457 if (stopped()) {
458 // The check always fails and therefore profile information is incorrect. Don't use it.
459 assert(old_control == slow_ctl, "type check should have been removed");
460 set_control(slow_ctl);
461 } else if (!slow_ctl->is_top()) {
462 { PreserveJVMState pjvms(this);
463 set_control(slow_ctl);
464 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
465 }
466 replace_in_map(ary, better_ary);
467 ary = better_ary;
468 arytype = _gvn.type(ary)->is_aryptr();
469 elemtype = arytype->elem();
470 }
471 }
472 } else if (UseTypeSpeculation && UseArrayLoadStoreProfile) {
473 // No need to speculate: feed profile data at this bci for the
474 // array to type speculation
475 ciKlass* array_type = NULL;
476 ciKlass* element_type = NULL;
477 ProfilePtrKind element_ptr = ProfileMaybeNull;
478 bool flat_array = true;
479 bool null_free_array = true;
480 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
481 if (array_type != NULL) {
482 ary = record_profile_for_speculation(ary, array_type, ProfileMaybeNull);
483 }
484 }
485
486 // We have no exact array type from profile data. Check profile data
487 // for a non null-free or non flat array. Non null-free implies non
488 // flat so check this one first. Speculating on a non null-free
489 // array doesn't help aaload but could be profitable for a
490 // subsequent aastore.
491 if (!arytype->is_null_free() && !arytype->is_not_null_free()) {
492 bool null_free_array = true;
493 Deoptimization::DeoptReason reason = Deoptimization::Reason_none;
494 if (arytype->speculative() != NULL &&
495 arytype->speculative()->is_aryptr()->is_not_null_free() &&
496 !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_class_check)) {
497 null_free_array = false;
498 reason = Deoptimization::Reason_speculate_class_check;
499 } else if (UseArrayLoadStoreProfile && !too_many_traps_or_recompiles(Deoptimization::Reason_class_check)) {
500 ciKlass* array_type = NULL;
501 ciKlass* element_type = NULL;
502 ProfilePtrKind element_ptr = ProfileMaybeNull;
503 bool flat_array = true;
504 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
505 reason = Deoptimization::Reason_class_check;
506 }
507 if (!null_free_array) {
508 { // Deoptimize if null-free array
509 BuildCutout unless(this, null_free_array_test(load_object_klass(ary), /* null_free = */ false), PROB_MAX);
510 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
511 }
512 assert(!stopped(), "null-free array should have been caught earlier");
513 Node* better_ary = _gvn.transform(new CheckCastPPNode(control(), ary, arytype->cast_to_not_null_free()));
514 replace_in_map(ary, better_ary);
515 ary = better_ary;
516 arytype = _gvn.type(ary)->is_aryptr();
517 }
518 }
519
520 if (!arytype->is_flat() && !arytype->is_not_flat()) {
521 bool flat_array = true;
522 Deoptimization::DeoptReason reason = Deoptimization::Reason_none;
523 if (arytype->speculative() != NULL &&
524 arytype->speculative()->is_aryptr()->is_not_flat() &&
525 !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_class_check)) {
526 flat_array = false;
527 reason = Deoptimization::Reason_speculate_class_check;
528 } else if (UseArrayLoadStoreProfile && !too_many_traps_or_recompiles(reason)) {
529 ciKlass* array_type = NULL;
530 ciKlass* element_type = NULL;
531 ProfilePtrKind element_ptr = ProfileMaybeNull;
532 bool null_free_array = true;
533 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
534 reason = Deoptimization::Reason_class_check;
535 }
536 if (!flat_array) {
537 { // Deoptimize if flat array
538 BuildCutout unless(this, flat_array_test(ary, /* flat = */ false), PROB_MAX);
539 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
540 }
541 assert(!stopped(), "flat array should have been caught earlier");
542 Node* better_ary = _gvn.transform(new CheckCastPPNode(control(), ary, arytype->cast_to_not_flat()));
543 replace_in_map(ary, better_ary);
544 ary = better_ary;
545 arytype = _gvn.type(ary)->is_aryptr();
546 }
547 }
548
549 // Make array address computation control dependent to prevent it
550 // from floating above the range check during loop optimizations.
551 Node* ptr = array_element_address(ary, idx, type, sizetype, control());
552 assert(ptr != top(), "top should go hand-in-hand with stopped");
553
554 return ptr;
555 }
556
557
558 // returns IfNode
559 IfNode* Parse::jump_if_fork_int(Node* a, Node* b, BoolTest::mask mask, float prob, float cnt) {
560 Node *cmp = _gvn.transform(new CmpINode(a, b)); // two cases: shiftcount > 32 and shiftcount <= 32
561 Node *tst = _gvn.transform(new BoolNode(cmp, mask));
562 IfNode *iff = create_and_map_if(control(), tst, prob, cnt);
563 return iff;
564 }
565
566
567 // sentinel value for the target bci to mark never taken branches
568 // (according to profiling)
569 static const int never_reached = INT_MAX;
570
571 //------------------------------helper for tableswitch-------------------------
572 void Parse::jump_if_true_fork(IfNode *iff, int dest_bci_if_true, bool unc) {
573 // True branch, use existing map info
574 { PreserveJVMState pjvms(this);
575 Node *iftrue = _gvn.transform( new IfTrueNode (iff) );
576 set_control( iftrue );
577 if (unc) {
578 repush_if_args();
579 uncommon_trap(Deoptimization::Reason_unstable_if,
580 Deoptimization::Action_reinterpret,
581 NULL,
582 "taken always");
583 } else {
584 assert(dest_bci_if_true != never_reached, "inconsistent dest");
585 merge_new_path(dest_bci_if_true);
586 }
587 }
588
589 // False branch
590 Node *iffalse = _gvn.transform( new IfFalseNode(iff) );
591 set_control( iffalse );
592 }
593
594 void Parse::jump_if_false_fork(IfNode *iff, int dest_bci_if_true, bool unc) {
595 // True branch, use existing map info
596 { PreserveJVMState pjvms(this);
597 Node *iffalse = _gvn.transform( new IfFalseNode (iff) );
598 set_control( iffalse );
599 if (unc) {
600 repush_if_args();
601 uncommon_trap(Deoptimization::Reason_unstable_if,
602 Deoptimization::Action_reinterpret,
603 NULL,
604 "taken never");
605 } else {
606 assert(dest_bci_if_true != never_reached, "inconsistent dest");
607 merge_new_path(dest_bci_if_true);
608 }
609 }
610
611 // False branch
612 Node *iftrue = _gvn.transform( new IfTrueNode(iff) );
613 set_control( iftrue );
614 }
615
616 void Parse::jump_if_always_fork(int dest_bci, bool unc) {
617 // False branch, use existing map and control()
618 if (unc) {
619 repush_if_args();
620 uncommon_trap(Deoptimization::Reason_unstable_if,
621 Deoptimization::Action_reinterpret,
622 NULL,
623 "taken never");
624 } else {
625 assert(dest_bci != never_reached, "inconsistent dest");
626 merge_new_path(dest_bci);
627 }
628 }
629
630
631 extern "C" {
632 static int jint_cmp(const void *i, const void *j) {
633 int a = *(jint *)i;
634 int b = *(jint *)j;
635 return a > b ? 1 : a < b ? -1 : 0;
636 }
637 }
638
639
640 class SwitchRange : public StackObj {
641 // a range of integers coupled with a bci destination
642 jint _lo; // inclusive lower limit
643 jint _hi; // inclusive upper limit
644 int _dest;
645 float _cnt; // how many times this range was hit according to profiling
646
647 public:
648 jint lo() const { return _lo; }
649 jint hi() const { return _hi; }
650 int dest() const { return _dest; }
651 bool is_singleton() const { return _lo == _hi; }
652 float cnt() const { return _cnt; }
653
654 void setRange(jint lo, jint hi, int dest, float cnt) {
655 assert(lo <= hi, "must be a non-empty range");
656 _lo = lo, _hi = hi; _dest = dest; _cnt = cnt;
657 assert(_cnt >= 0, "");
658 }
659 bool adjoinRange(jint lo, jint hi, int dest, float cnt, bool trim_ranges) {
660 assert(lo <= hi, "must be a non-empty range");
661 if (lo == _hi+1) {
662 // see merge_ranges() comment below
663 if (trim_ranges) {
664 if (cnt == 0) {
665 if (_cnt != 0) {
666 return false;
667 }
668 if (dest != _dest) {
669 _dest = never_reached;
670 }
671 } else {
672 if (_cnt == 0) {
673 return false;
674 }
675 if (dest != _dest) {
676 return false;
677 }
678 }
679 } else {
680 if (dest != _dest) {
681 return false;
682 }
683 }
684 _hi = hi;
685 _cnt += cnt;
686 return true;
687 }
688 return false;
689 }
690
691 void set (jint value, int dest, float cnt) {
692 setRange(value, value, dest, cnt);
693 }
694 bool adjoin(jint value, int dest, float cnt, bool trim_ranges) {
695 return adjoinRange(value, value, dest, cnt, trim_ranges);
696 }
697 bool adjoin(SwitchRange& other) {
698 return adjoinRange(other._lo, other._hi, other._dest, other._cnt, false);
699 }
700
701 void print() {
702 if (is_singleton())
703 tty->print(" {%d}=>%d (cnt=%f)", lo(), dest(), cnt());
704 else if (lo() == min_jint)
705 tty->print(" {..%d}=>%d (cnt=%f)", hi(), dest(), cnt());
706 else if (hi() == max_jint)
707 tty->print(" {%d..}=>%d (cnt=%f)", lo(), dest(), cnt());
708 else
709 tty->print(" {%d..%d}=>%d (cnt=%f)", lo(), hi(), dest(), cnt());
710 }
711 };
712
713 // We try to minimize the number of ranges and the size of the taken
714 // ones using profiling data. When ranges are created,
715 // SwitchRange::adjoinRange() only allows 2 adjoining ranges to merge
716 // if both were never hit or both were hit to build longer unreached
717 // ranges. Here, we now merge adjoining ranges with the same
718 // destination and finally set destination of unreached ranges to the
719 // special value never_reached because it can help minimize the number
720 // of tests that are necessary.
721 //
722 // For instance:
723 // [0, 1] to target1 sometimes taken
724 // [1, 2] to target1 never taken
725 // [2, 3] to target2 never taken
726 // would lead to:
727 // [0, 1] to target1 sometimes taken
728 // [1, 3] never taken
729 //
730 // (first 2 ranges to target1 are not merged)
731 static void merge_ranges(SwitchRange* ranges, int& rp) {
732 if (rp == 0) {
733 return;
734 }
735 int shift = 0;
736 for (int j = 0; j < rp; j++) {
737 SwitchRange& r1 = ranges[j-shift];
738 SwitchRange& r2 = ranges[j+1];
739 if (r1.adjoin(r2)) {
740 shift++;
741 } else if (shift > 0) {
742 ranges[j+1-shift] = r2;
743 }
744 }
745 rp -= shift;
746 for (int j = 0; j <= rp; j++) {
747 SwitchRange& r = ranges[j];
748 if (r.cnt() == 0 && r.dest() != never_reached) {
749 r.setRange(r.lo(), r.hi(), never_reached, r.cnt());
750 }
751 }
752 }
753
754 //-------------------------------do_tableswitch--------------------------------
755 void Parse::do_tableswitch() {
756 // Get information about tableswitch
757 int default_dest = iter().get_dest_table(0);
758 jint lo_index = iter().get_int_table(1);
759 jint hi_index = iter().get_int_table(2);
760 int len = hi_index - lo_index + 1;
761
762 if (len < 1) {
763 // If this is a backward branch, add safepoint
764 maybe_add_safepoint(default_dest);
765 pop(); // the effect of the instruction execution on the operand stack
766 merge(default_dest);
767 return;
768 }
769
770 ciMethodData* methodData = method()->method_data();
771 ciMultiBranchData* profile = NULL;
772 if (methodData->is_mature() && UseSwitchProfiling) {
773 ciProfileData* data = methodData->bci_to_data(bci());
774 if (data != NULL && data->is_MultiBranchData()) {
775 profile = (ciMultiBranchData*)data;
776 }
777 }
778 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if);
779
780 // generate decision tree, using trichotomy when possible
781 int rnum = len+2;
782 bool makes_backward_branch = false;
783 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum);
784 int rp = -1;
785 if (lo_index != min_jint) {
786 float cnt = 1.0F;
787 if (profile != NULL) {
788 cnt = (float)profile->default_count() / (hi_index != max_jint ? 2.0F : 1.0F);
789 }
790 ranges[++rp].setRange(min_jint, lo_index-1, default_dest, cnt);
791 }
792 for (int j = 0; j < len; j++) {
793 jint match_int = lo_index+j;
794 int dest = iter().get_dest_table(j+3);
795 makes_backward_branch |= (dest <= bci());
796 float cnt = 1.0F;
797 if (profile != NULL) {
798 cnt = (float)profile->count_at(j);
799 }
800 if (rp < 0 || !ranges[rp].adjoin(match_int, dest, cnt, trim_ranges)) {
801 ranges[++rp].set(match_int, dest, cnt);
802 }
803 }
804 jint highest = lo_index+(len-1);
805 assert(ranges[rp].hi() == highest, "");
806 if (highest != max_jint) {
807 float cnt = 1.0F;
808 if (profile != NULL) {
809 cnt = (float)profile->default_count() / (lo_index != min_jint ? 2.0F : 1.0F);
810 }
811 if (!ranges[rp].adjoinRange(highest+1, max_jint, default_dest, cnt, trim_ranges)) {
812 ranges[++rp].setRange(highest+1, max_jint, default_dest, cnt);
813 }
814 }
815 assert(rp < len+2, "not too many ranges");
816
817 if (trim_ranges) {
818 merge_ranges(ranges, rp);
819 }
820
821 // Safepoint in case if backward branch observed
822 if (makes_backward_branch) {
823 add_safepoint();
824 }
825
826 Node* lookup = pop(); // lookup value
827 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]);
828 }
829
830
831 //------------------------------do_lookupswitch--------------------------------
832 void Parse::do_lookupswitch() {
833 // Get information about lookupswitch
834 int default_dest = iter().get_dest_table(0);
835 jint len = iter().get_int_table(1);
836
837 if (len < 1) { // If this is a backward branch, add safepoint
838 maybe_add_safepoint(default_dest);
839 pop(); // the effect of the instruction execution on the operand stack
840 merge(default_dest);
841 return;
842 }
843
844 ciMethodData* methodData = method()->method_data();
845 ciMultiBranchData* profile = NULL;
846 if (methodData->is_mature() && UseSwitchProfiling) {
847 ciProfileData* data = methodData->bci_to_data(bci());
848 if (data != NULL && data->is_MultiBranchData()) {
849 profile = (ciMultiBranchData*)data;
850 }
851 }
852 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if);
853
854 // generate decision tree, using trichotomy when possible
855 jint* table = NEW_RESOURCE_ARRAY(jint, len*3);
856 {
857 for (int j = 0; j < len; j++) {
858 table[3*j+0] = iter().get_int_table(2+2*j);
859 table[3*j+1] = iter().get_dest_table(2+2*j+1);
860 // Handle overflow when converting from uint to jint
861 table[3*j+2] = (profile == NULL) ? 1 : (jint)MIN2<uint>((uint)max_jint, profile->count_at(j));
862 }
863 qsort(table, len, 3*sizeof(table[0]), jint_cmp);
864 }
865
866 float default_cnt = 1.0F;
867 if (profile != NULL) {
868 juint defaults = max_juint - len;
869 default_cnt = (float)profile->default_count()/(float)defaults;
870 }
871
872 int rnum = len*2+1;
873 bool makes_backward_branch = false;
874 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum);
875 int rp = -1;
876 for (int j = 0; j < len; j++) {
877 jint match_int = table[3*j+0];
878 jint dest = table[3*j+1];
879 jint cnt = table[3*j+2];
880 jint next_lo = rp < 0 ? min_jint : ranges[rp].hi()+1;
881 makes_backward_branch |= (dest <= bci());
882 float c = default_cnt * ((float)match_int - (float)next_lo);
883 if (match_int != next_lo && (rp < 0 || !ranges[rp].adjoinRange(next_lo, match_int-1, default_dest, c, trim_ranges))) {
884 assert(default_dest != never_reached, "sentinel value for dead destinations");
885 ranges[++rp].setRange(next_lo, match_int-1, default_dest, c);
886 }
887 if (rp < 0 || !ranges[rp].adjoin(match_int, dest, (float)cnt, trim_ranges)) {
888 assert(dest != never_reached, "sentinel value for dead destinations");
889 ranges[++rp].set(match_int, dest, (float)cnt);
890 }
891 }
892 jint highest = table[3*(len-1)];
893 assert(ranges[rp].hi() == highest, "");
894 if (highest != max_jint &&
895 !ranges[rp].adjoinRange(highest+1, max_jint, default_dest, default_cnt * ((float)max_jint - (float)highest), trim_ranges)) {
896 ranges[++rp].setRange(highest+1, max_jint, default_dest, default_cnt * ((float)max_jint - (float)highest));
897 }
898 assert(rp < rnum, "not too many ranges");
899
900 if (trim_ranges) {
901 merge_ranges(ranges, rp);
902 }
903
904 // Safepoint in case backward branch observed
905 if (makes_backward_branch) {
906 add_safepoint();
907 }
908
909 Node *lookup = pop(); // lookup value
910 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]);
911 }
912
913 static float if_prob(float taken_cnt, float total_cnt) {
914 assert(taken_cnt <= total_cnt, "");
915 if (total_cnt == 0) {
916 return PROB_FAIR;
917 }
918 float p = taken_cnt / total_cnt;
919 return clamp(p, PROB_MIN, PROB_MAX);
920 }
921
922 static float if_cnt(float cnt) {
923 if (cnt == 0) {
924 return COUNT_UNKNOWN;
925 }
926 return cnt;
927 }
928
929 static float sum_of_cnts(SwitchRange *lo, SwitchRange *hi) {
930 float total_cnt = 0;
931 for (SwitchRange* sr = lo; sr <= hi; sr++) {
932 total_cnt += sr->cnt();
933 }
934 return total_cnt;
935 }
936
937 class SwitchRanges : public ResourceObj {
938 public:
939 SwitchRange* _lo;
940 SwitchRange* _hi;
941 SwitchRange* _mid;
942 float _cost;
943
944 enum {
945 Start,
946 LeftDone,
947 RightDone,
948 Done
949 } _state;
950
951 SwitchRanges(SwitchRange *lo, SwitchRange *hi)
952 : _lo(lo), _hi(hi), _mid(NULL),
953 _cost(0), _state(Start) {
954 }
955
956 SwitchRanges()
957 : _lo(NULL), _hi(NULL), _mid(NULL),
958 _cost(0), _state(Start) {}
959 };
960
961 // Estimate cost of performing a binary search on lo..hi
962 static float compute_tree_cost(SwitchRange *lo, SwitchRange *hi, float total_cnt) {
963 GrowableArray<SwitchRanges> tree;
964 SwitchRanges root(lo, hi);
965 tree.push(root);
966
967 float cost = 0;
968 do {
969 SwitchRanges& r = *tree.adr_at(tree.length()-1);
970 if (r._hi != r._lo) {
971 if (r._mid == NULL) {
972 float r_cnt = sum_of_cnts(r._lo, r._hi);
973
974 if (r_cnt == 0) {
975 tree.pop();
976 cost = 0;
977 continue;
978 }
979
980 SwitchRange* mid = NULL;
981 mid = r._lo;
982 for (float cnt = 0; ; ) {
983 assert(mid <= r._hi, "out of bounds");
984 cnt += mid->cnt();
985 if (cnt > r_cnt / 2) {
986 break;
987 }
988 mid++;
989 }
990 assert(mid <= r._hi, "out of bounds");
991 r._mid = mid;
992 r._cost = r_cnt / total_cnt;
993 }
994 r._cost += cost;
995 if (r._state < SwitchRanges::LeftDone && r._mid > r._lo) {
996 cost = 0;
997 r._state = SwitchRanges::LeftDone;
998 tree.push(SwitchRanges(r._lo, r._mid-1));
999 } else if (r._state < SwitchRanges::RightDone) {
1000 cost = 0;
1001 r._state = SwitchRanges::RightDone;
1002 tree.push(SwitchRanges(r._mid == r._lo ? r._mid+1 : r._mid, r._hi));
1003 } else {
1004 tree.pop();
1005 cost = r._cost;
1006 }
1007 } else {
1008 tree.pop();
1009 cost = r._cost;
1010 }
1011 } while (tree.length() > 0);
1012
1013
1014 return cost;
1015 }
1016
1017 // It sometimes pays off to test most common ranges before the binary search
1018 void Parse::linear_search_switch_ranges(Node* key_val, SwitchRange*& lo, SwitchRange*& hi) {
1019 uint nr = hi - lo + 1;
1020 float total_cnt = sum_of_cnts(lo, hi);
1021
1022 float min = compute_tree_cost(lo, hi, total_cnt);
1023 float extra = 1;
1024 float sub = 0;
1025
1026 SwitchRange* array1 = lo;
1027 SwitchRange* array2 = NEW_RESOURCE_ARRAY(SwitchRange, nr);
1028
1029 SwitchRange* ranges = NULL;
1030
1031 while (nr >= 2) {
1032 assert(lo == array1 || lo == array2, "one the 2 already allocated arrays");
1033 ranges = (lo == array1) ? array2 : array1;
1034
1035 // Find highest frequency range
1036 SwitchRange* candidate = lo;
1037 for (SwitchRange* sr = lo+1; sr <= hi; sr++) {
1038 if (sr->cnt() > candidate->cnt()) {
1039 candidate = sr;
1040 }
1041 }
1042 SwitchRange most_freq = *candidate;
1043 if (most_freq.cnt() == 0) {
1044 break;
1045 }
1046
1047 // Copy remaining ranges into another array
1048 int shift = 0;
1049 for (uint i = 0; i < nr; i++) {
1050 SwitchRange* sr = &lo[i];
1051 if (sr != candidate) {
1052 ranges[i-shift] = *sr;
1053 } else {
1054 shift++;
1055 if (i > 0 && i < nr-1) {
1056 SwitchRange prev = lo[i-1];
1057 prev.setRange(prev.lo(), sr->hi(), prev.dest(), prev.cnt());
1058 if (prev.adjoin(lo[i+1])) {
1059 shift++;
1060 i++;
1061 }
1062 ranges[i-shift] = prev;
1063 }
1064 }
1065 }
1066 nr -= shift;
1067
1068 // Evaluate cost of testing the most common range and performing a
1069 // binary search on the other ranges
1070 float cost = extra + compute_tree_cost(&ranges[0], &ranges[nr-1], total_cnt);
1071 if (cost >= min) {
1072 break;
1073 }
1074 // swap arrays
1075 lo = &ranges[0];
1076 hi = &ranges[nr-1];
1077
1078 // It pays off: emit the test for the most common range
1079 assert(most_freq.cnt() > 0, "must be taken");
1080 Node* val = _gvn.transform(new SubINode(key_val, _gvn.intcon(most_freq.lo())));
1081 Node* cmp = _gvn.transform(new CmpUNode(val, _gvn.intcon(most_freq.hi() - most_freq.lo())));
1082 Node* tst = _gvn.transform(new BoolNode(cmp, BoolTest::le));
1083 IfNode* iff = create_and_map_if(control(), tst, if_prob(most_freq.cnt(), total_cnt), if_cnt(most_freq.cnt()));
1084 jump_if_true_fork(iff, most_freq.dest(), false);
1085
1086 sub += most_freq.cnt() / total_cnt;
1087 extra += 1 - sub;
1088 min = cost;
1089 }
1090 }
1091
1092 //----------------------------create_jump_tables-------------------------------
1093 bool Parse::create_jump_tables(Node* key_val, SwitchRange* lo, SwitchRange* hi) {
1094 // Are jumptables enabled
1095 if (!UseJumpTables) return false;
1096
1097 // Are jumptables supported
1098 if (!Matcher::has_match_rule(Op_Jump)) return false;
1099
1100 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if);
1101
1102 // Decide if a guard is needed to lop off big ranges at either (or
1103 // both) end(s) of the input set. We'll call this the default target
1104 // even though we can't be sure that it is the true "default".
1105
1106 bool needs_guard = false;
1107 int default_dest;
1108 int64_t total_outlier_size = 0;
1109 int64_t hi_size = ((int64_t)hi->hi()) - ((int64_t)hi->lo()) + 1;
1110 int64_t lo_size = ((int64_t)lo->hi()) - ((int64_t)lo->lo()) + 1;
1111
1112 if (lo->dest() == hi->dest()) {
1113 total_outlier_size = hi_size + lo_size;
1114 default_dest = lo->dest();
1115 } else if (lo_size > hi_size) {
1116 total_outlier_size = lo_size;
1117 default_dest = lo->dest();
1118 } else {
1119 total_outlier_size = hi_size;
1120 default_dest = hi->dest();
1121 }
1122
1123 float total = sum_of_cnts(lo, hi);
1124 float cost = compute_tree_cost(lo, hi, total);
1125
1126 // If a guard test will eliminate very sparse end ranges, then
1127 // it is worth the cost of an extra jump.
1128 float trimmed_cnt = 0;
1129 if (total_outlier_size > (MaxJumpTableSparseness * 4)) {
1130 needs_guard = true;
1131 if (default_dest == lo->dest()) {
1132 trimmed_cnt += lo->cnt();
1133 lo++;
1134 }
1135 if (default_dest == hi->dest()) {
1136 trimmed_cnt += hi->cnt();
1137 hi--;
1138 }
1139 }
1140
1141 // Find the total number of cases and ranges
1142 int64_t num_cases = ((int64_t)hi->hi()) - ((int64_t)lo->lo()) + 1;
1143 int num_range = hi - lo + 1;
1144
1145 // Don't create table if: too large, too small, or too sparse.
1146 if (num_cases > MaxJumpTableSize)
1147 return false;
1148 if (UseSwitchProfiling) {
1149 // MinJumpTableSize is set so with a well balanced binary tree,
1150 // when the number of ranges is MinJumpTableSize, it's cheaper to
1151 // go through a JumpNode that a tree of IfNodes. Average cost of a
1152 // tree of IfNodes with MinJumpTableSize is
1153 // log2f(MinJumpTableSize) comparisons. So if the cost computed
1154 // from profile data is less than log2f(MinJumpTableSize) then
1155 // going with the binary search is cheaper.
1156 if (cost < log2f(MinJumpTableSize)) {
1157 return false;
1158 }
1159 } else {
1160 if (num_cases < MinJumpTableSize)
1161 return false;
1162 }
1163 if (num_cases > (MaxJumpTableSparseness * num_range))
1164 return false;
1165
1166 // Normalize table lookups to zero
1167 int lowval = lo->lo();
1168 key_val = _gvn.transform( new SubINode(key_val, _gvn.intcon(lowval)) );
1169
1170 // Generate a guard to protect against input keyvals that aren't
1171 // in the switch domain.
1172 if (needs_guard) {
1173 Node* size = _gvn.intcon(num_cases);
1174 Node* cmp = _gvn.transform(new CmpUNode(key_val, size));
1175 Node* tst = _gvn.transform(new BoolNode(cmp, BoolTest::ge));
1176 IfNode* iff = create_and_map_if(control(), tst, if_prob(trimmed_cnt, total), if_cnt(trimmed_cnt));
1177 jump_if_true_fork(iff, default_dest, trim_ranges && trimmed_cnt == 0);
1178
1179 total -= trimmed_cnt;
1180 }
1181
1182 // Create an ideal node JumpTable that has projections
1183 // of all possible ranges for a switch statement
1184 // The key_val input must be converted to a pointer offset and scaled.
1185 // Compare Parse::array_addressing above.
1186
1187 // Clean the 32-bit int into a real 64-bit offset.
1188 // Otherwise, the jint value 0 might turn into an offset of 0x0800000000.
1189 // Make I2L conversion control dependent to prevent it from
1190 // floating above the range check during loop optimizations.
1191 // Do not use a narrow int type here to prevent the data path from dying
1192 // while the control path is not removed. This can happen if the type of key_val
1193 // is later known to be out of bounds of [0, num_cases] and therefore a narrow cast
1194 // would be replaced by TOP while C2 is not able to fold the corresponding range checks.
1195 // Set _carry_dependency for the cast to avoid being removed by IGVN.
1196 #ifdef _LP64
1197 key_val = C->constrained_convI2L(&_gvn, key_val, TypeInt::INT, control(), true /* carry_dependency */);
1198 #endif
1199
1200 // Shift the value by wordsize so we have an index into the table, rather
1201 // than a switch value
1202 Node *shiftWord = _gvn.MakeConX(wordSize);
1203 key_val = _gvn.transform( new MulXNode( key_val, shiftWord));
1204
1205 // Create the JumpNode
1206 Arena* arena = C->comp_arena();
1207 float* probs = (float*)arena->Amalloc(sizeof(float)*num_cases);
1208 int i = 0;
1209 if (total == 0) {
1210 for (SwitchRange* r = lo; r <= hi; r++) {
1211 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) {
1212 probs[i] = 1.0F / num_cases;
1213 }
1214 }
1215 } else {
1216 for (SwitchRange* r = lo; r <= hi; r++) {
1217 float prob = r->cnt()/total;
1218 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) {
1219 probs[i] = prob / (r->hi() - r->lo() + 1);
1220 }
1221 }
1222 }
1223
1224 ciMethodData* methodData = method()->method_data();
1225 ciMultiBranchData* profile = NULL;
1226 if (methodData->is_mature()) {
1227 ciProfileData* data = methodData->bci_to_data(bci());
1228 if (data != NULL && data->is_MultiBranchData()) {
1229 profile = (ciMultiBranchData*)data;
1230 }
1231 }
1232
1233 Node* jtn = _gvn.transform(new JumpNode(control(), key_val, num_cases, probs, profile == NULL ? COUNT_UNKNOWN : total));
1234
1235 // These are the switch destinations hanging off the jumpnode
1236 i = 0;
1237 for (SwitchRange* r = lo; r <= hi; r++) {
1238 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) {
1239 Node* input = _gvn.transform(new JumpProjNode(jtn, i, r->dest(), (int)(j - lowval)));
1240 {
1241 PreserveJVMState pjvms(this);
1242 set_control(input);
1243 jump_if_always_fork(r->dest(), trim_ranges && r->cnt() == 0);
1244 }
1245 }
1246 }
1247 assert(i == num_cases, "miscount of cases");
1248 stop_and_kill_map(); // no more uses for this JVMS
1249 return true;
1250 }
1251
1252 //----------------------------jump_switch_ranges-------------------------------
1253 void Parse::jump_switch_ranges(Node* key_val, SwitchRange *lo, SwitchRange *hi, int switch_depth) {
1254 Block* switch_block = block();
1255 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if);
1256
1257 if (switch_depth == 0) {
1258 // Do special processing for the top-level call.
1259 assert(lo->lo() == min_jint, "initial range must exhaust Type::INT");
1260 assert(hi->hi() == max_jint, "initial range must exhaust Type::INT");
1261
1262 // Decrement pred-numbers for the unique set of nodes.
1263 #ifdef ASSERT
1264 if (!trim_ranges) {
1265 // Ensure that the block's successors are a (duplicate-free) set.
1266 int successors_counted = 0; // block occurrences in [hi..lo]
1267 int unique_successors = switch_block->num_successors();
1268 for (int i = 0; i < unique_successors; i++) {
1269 Block* target = switch_block->successor_at(i);
1270
1271 // Check that the set of successors is the same in both places.
1272 int successors_found = 0;
1273 for (SwitchRange* p = lo; p <= hi; p++) {
1274 if (p->dest() == target->start()) successors_found++;
1275 }
1276 assert(successors_found > 0, "successor must be known");
1277 successors_counted += successors_found;
1278 }
1279 assert(successors_counted == (hi-lo)+1, "no unexpected successors");
1280 }
1281 #endif
1282
1283 // Maybe prune the inputs, based on the type of key_val.
1284 jint min_val = min_jint;
1285 jint max_val = max_jint;
1286 const TypeInt* ti = key_val->bottom_type()->isa_int();
1287 if (ti != NULL) {
1288 min_val = ti->_lo;
1289 max_val = ti->_hi;
1290 assert(min_val <= max_val, "invalid int type");
1291 }
1292 while (lo->hi() < min_val) {
1293 lo++;
1294 }
1295 if (lo->lo() < min_val) {
1296 lo->setRange(min_val, lo->hi(), lo->dest(), lo->cnt());
1297 }
1298 while (hi->lo() > max_val) {
1299 hi--;
1300 }
1301 if (hi->hi() > max_val) {
1302 hi->setRange(hi->lo(), max_val, hi->dest(), hi->cnt());
1303 }
1304
1305 linear_search_switch_ranges(key_val, lo, hi);
1306 }
1307
1308 #ifndef PRODUCT
1309 if (switch_depth == 0) {
1310 _max_switch_depth = 0;
1311 _est_switch_depth = log2i_graceful((hi - lo + 1) - 1) + 1;
1312 }
1313 #endif
1314
1315 assert(lo <= hi, "must be a non-empty set of ranges");
1316 if (lo == hi) {
1317 jump_if_always_fork(lo->dest(), trim_ranges && lo->cnt() == 0);
1318 } else {
1319 assert(lo->hi() == (lo+1)->lo()-1, "contiguous ranges");
1320 assert(hi->lo() == (hi-1)->hi()+1, "contiguous ranges");
1321
1322 if (create_jump_tables(key_val, lo, hi)) return;
1323
1324 SwitchRange* mid = NULL;
1325 float total_cnt = sum_of_cnts(lo, hi);
1326
1327 int nr = hi - lo + 1;
1328 if (UseSwitchProfiling) {
1329 // Don't keep the binary search tree balanced: pick up mid point
1330 // that split frequencies in half.
1331 float cnt = 0;
1332 for (SwitchRange* sr = lo; sr <= hi; sr++) {
1333 cnt += sr->cnt();
1334 if (cnt >= total_cnt / 2) {
1335 mid = sr;
1336 break;
1337 }
1338 }
1339 } else {
1340 mid = lo + nr/2;
1341
1342 // if there is an easy choice, pivot at a singleton:
1343 if (nr > 3 && !mid->is_singleton() && (mid-1)->is_singleton()) mid--;
1344
1345 assert(lo < mid && mid <= hi, "good pivot choice");
1346 assert(nr != 2 || mid == hi, "should pick higher of 2");
1347 assert(nr != 3 || mid == hi-1, "should pick middle of 3");
1348 }
1349
1350
1351 Node *test_val = _gvn.intcon(mid == lo ? mid->hi() : mid->lo());
1352
1353 if (mid->is_singleton()) {
1354 IfNode *iff_ne = jump_if_fork_int(key_val, test_val, BoolTest::ne, 1-if_prob(mid->cnt(), total_cnt), if_cnt(mid->cnt()));
1355 jump_if_false_fork(iff_ne, mid->dest(), trim_ranges && mid->cnt() == 0);
1356
1357 // Special Case: If there are exactly three ranges, and the high
1358 // and low range each go to the same place, omit the "gt" test,
1359 // since it will not discriminate anything.
1360 bool eq_test_only = (hi == lo+2 && hi->dest() == lo->dest() && mid == hi-1) || mid == lo;
1361
1362 // if there is a higher range, test for it and process it:
1363 if (mid < hi && !eq_test_only) {
1364 // two comparisons of same values--should enable 1 test for 2 branches
1365 // Use BoolTest::lt instead of BoolTest::gt
1366 float cnt = sum_of_cnts(lo, mid-1);
1367 IfNode *iff_lt = jump_if_fork_int(key_val, test_val, BoolTest::lt, if_prob(cnt, total_cnt), if_cnt(cnt));
1368 Node *iftrue = _gvn.transform( new IfTrueNode(iff_lt) );
1369 Node *iffalse = _gvn.transform( new IfFalseNode(iff_lt) );
1370 { PreserveJVMState pjvms(this);
1371 set_control(iffalse);
1372 jump_switch_ranges(key_val, mid+1, hi, switch_depth+1);
1373 }
1374 set_control(iftrue);
1375 }
1376
1377 } else {
1378 // mid is a range, not a singleton, so treat mid..hi as a unit
1379 float cnt = sum_of_cnts(mid == lo ? mid+1 : mid, hi);
1380 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));
1381
1382 // if there is a higher range, test for it and process it:
1383 if (mid == hi) {
1384 jump_if_true_fork(iff_ge, mid->dest(), trim_ranges && cnt == 0);
1385 } else {
1386 Node *iftrue = _gvn.transform( new IfTrueNode(iff_ge) );
1387 Node *iffalse = _gvn.transform( new IfFalseNode(iff_ge) );
1388 { PreserveJVMState pjvms(this);
1389 set_control(iftrue);
1390 jump_switch_ranges(key_val, mid == lo ? mid+1 : mid, hi, switch_depth+1);
1391 }
1392 set_control(iffalse);
1393 }
1394 }
1395
1396 // in any case, process the lower range
1397 if (mid == lo) {
1398 if (mid->is_singleton()) {
1399 jump_switch_ranges(key_val, lo+1, hi, switch_depth+1);
1400 } else {
1401 jump_if_always_fork(lo->dest(), trim_ranges && lo->cnt() == 0);
1402 }
1403 } else {
1404 jump_switch_ranges(key_val, lo, mid-1, switch_depth+1);
1405 }
1406 }
1407
1408 // Decrease pred_count for each successor after all is done.
1409 if (switch_depth == 0) {
1410 int unique_successors = switch_block->num_successors();
1411 for (int i = 0; i < unique_successors; i++) {
1412 Block* target = switch_block->successor_at(i);
1413 // Throw away the pre-allocated path for each unique successor.
1414 target->next_path_num();
1415 }
1416 }
1417
1418 #ifndef PRODUCT
1419 _max_switch_depth = MAX2(switch_depth, _max_switch_depth);
1420 if (TraceOptoParse && Verbose && WizardMode && switch_depth == 0) {
1421 SwitchRange* r;
1422 int nsing = 0;
1423 for( r = lo; r <= hi; r++ ) {
1424 if( r->is_singleton() ) nsing++;
1425 }
1426 tty->print(">>> ");
1427 _method->print_short_name();
1428 tty->print_cr(" switch decision tree");
1429 tty->print_cr(" %d ranges (%d singletons), max_depth=%d, est_depth=%d",
1430 (int) (hi-lo+1), nsing, _max_switch_depth, _est_switch_depth);
1431 if (_max_switch_depth > _est_switch_depth) {
1432 tty->print_cr("******** BAD SWITCH DEPTH ********");
1433 }
1434 tty->print(" ");
1435 for( r = lo; r <= hi; r++ ) {
1436 r->print();
1437 }
1438 tty->cr();
1439 }
1440 #endif
1441 }
1442
1443 void Parse::modf() {
1444 Node *f2 = pop();
1445 Node *f1 = pop();
1446 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::modf_Type(),
1447 CAST_FROM_FN_PTR(address, SharedRuntime::frem),
1448 "frem", NULL, //no memory effects
1449 f1, f2);
1450 Node* res = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0));
1451
1452 push(res);
1453 }
1454
1455 void Parse::modd() {
1456 Node *d2 = pop_pair();
1457 Node *d1 = pop_pair();
1458 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::Math_DD_D_Type(),
1459 CAST_FROM_FN_PTR(address, SharedRuntime::drem),
1460 "drem", NULL, //no memory effects
1461 d1, top(), d2, top());
1462 Node* res_d = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0));
1463
1464 #ifdef ASSERT
1465 Node* res_top = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 1));
1466 assert(res_top == top(), "second value must be top");
1467 #endif
1468
1469 push_pair(res_d);
1470 }
1471
1472 void Parse::l2f() {
1473 Node* f2 = pop();
1474 Node* f1 = pop();
1475 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::l2f_Type(),
1476 CAST_FROM_FN_PTR(address, SharedRuntime::l2f),
1477 "l2f", NULL, //no memory effects
1478 f1, f2);
1479 Node* res = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0));
1480
1481 push(res);
1482 }
1483
1484 // Handle jsr and jsr_w bytecode
1485 void Parse::do_jsr() {
1486 assert(bc() == Bytecodes::_jsr || bc() == Bytecodes::_jsr_w, "wrong bytecode");
1487
1488 // Store information about current state, tagged with new _jsr_bci
1489 int return_bci = iter().next_bci();
1490 int jsr_bci = (bc() == Bytecodes::_jsr) ? iter().get_dest() : iter().get_far_dest();
1491
1492 // The way we do things now, there is only one successor block
1493 // for the jsr, because the target code is cloned by ciTypeFlow.
1494 Block* target = successor_for_bci(jsr_bci);
1495
1496 // What got pushed?
1497 const Type* ret_addr = target->peek();
1498 assert(ret_addr->singleton(), "must be a constant (cloned jsr body)");
1499
1500 // Effect on jsr on stack
1501 push(_gvn.makecon(ret_addr));
1502
1503 // Flow to the jsr.
1504 merge(jsr_bci);
1505 }
1506
1507 // Handle ret bytecode
1508 void Parse::do_ret() {
1509 // Find to whom we return.
1510 assert(block()->num_successors() == 1, "a ret can only go one place now");
1511 Block* target = block()->successor_at(0);
1512 assert(!target->is_ready(), "our arrival must be expected");
1513 int pnum = target->next_path_num();
1514 merge_common(target, pnum);
1515 }
1516
1517 static bool has_injected_profile(BoolTest::mask btest, Node* test, int& taken, int& not_taken) {
1518 if (btest != BoolTest::eq && btest != BoolTest::ne) {
1519 // Only ::eq and ::ne are supported for profile injection.
1520 return false;
1521 }
1522 if (test->is_Cmp() &&
1523 test->in(1)->Opcode() == Op_ProfileBoolean) {
1524 ProfileBooleanNode* profile = (ProfileBooleanNode*)test->in(1);
1525 int false_cnt = profile->false_count();
1526 int true_cnt = profile->true_count();
1527
1528 // Counts matching depends on the actual test operation (::eq or ::ne).
1529 // No need to scale the counts because profile injection was designed
1530 // to feed exact counts into VM.
1531 taken = (btest == BoolTest::eq) ? false_cnt : true_cnt;
1532 not_taken = (btest == BoolTest::eq) ? true_cnt : false_cnt;
1533
1534 profile->consume();
1535 return true;
1536 }
1537 return false;
1538 }
1539 //--------------------------dynamic_branch_prediction--------------------------
1540 // Try to gather dynamic branch prediction behavior. Return a probability
1541 // of the branch being taken and set the "cnt" field. Returns a -1.0
1542 // if we need to use static prediction for some reason.
1543 float Parse::dynamic_branch_prediction(float &cnt, BoolTest::mask btest, Node* test) {
1544 ResourceMark rm;
1545
1546 cnt = COUNT_UNKNOWN;
1547
1548 int taken = 0;
1549 int not_taken = 0;
1550
1551 bool use_mdo = !has_injected_profile(btest, test, taken, not_taken);
1552
1553 if (use_mdo) {
1554 // Use MethodData information if it is available
1555 // FIXME: free the ProfileData structure
1556 ciMethodData* methodData = method()->method_data();
1557 if (!methodData->is_mature()) return PROB_UNKNOWN;
1558 ciProfileData* data = methodData->bci_to_data(bci());
1559 if (data == NULL) {
1560 return PROB_UNKNOWN;
1561 }
1562 if (!data->is_JumpData()) return PROB_UNKNOWN;
1563
1564 // get taken and not taken values
1565 taken = data->as_JumpData()->taken();
1566 not_taken = 0;
1567 if (data->is_BranchData()) {
1568 not_taken = data->as_BranchData()->not_taken();
1569 }
1570
1571 // scale the counts to be commensurate with invocation counts:
1572 taken = method()->scale_count(taken);
1573 not_taken = method()->scale_count(not_taken);
1574 }
1575
1576 // Give up if too few (or too many, in which case the sum will overflow) counts to be meaningful.
1577 // We also check that individual counters are positive first, otherwise the sum can become positive.
1578 if (taken < 0 || not_taken < 0 || taken + not_taken < 40) {
1579 if (C->log() != NULL) {
1580 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d'", iter().get_dest(), taken, not_taken);
1581 }
1582 return PROB_UNKNOWN;
1583 }
1584
1585 // Compute frequency that we arrive here
1586 float sum = taken + not_taken;
1587 // Adjust, if this block is a cloned private block but the
1588 // Jump counts are shared. Taken the private counts for
1589 // just this path instead of the shared counts.
1590 if( block()->count() > 0 )
1591 sum = block()->count();
1592 cnt = sum / FreqCountInvocations;
1593
1594 // Pin probability to sane limits
1595 float prob;
1596 if( !taken )
1597 prob = (0+PROB_MIN) / 2;
1598 else if( !not_taken )
1599 prob = (1+PROB_MAX) / 2;
1600 else { // Compute probability of true path
1601 prob = (float)taken / (float)(taken + not_taken);
1602 if (prob > PROB_MAX) prob = PROB_MAX;
1603 if (prob < PROB_MIN) prob = PROB_MIN;
1604 }
1605
1606 assert((cnt > 0.0f) && (prob > 0.0f),
1607 "Bad frequency assignment in if");
1608
1609 if (C->log() != NULL) {
1610 const char* prob_str = NULL;
1611 if (prob >= PROB_MAX) prob_str = (prob == PROB_MAX) ? "max" : "always";
1612 if (prob <= PROB_MIN) prob_str = (prob == PROB_MIN) ? "min" : "never";
1613 char prob_str_buf[30];
1614 if (prob_str == NULL) {
1615 jio_snprintf(prob_str_buf, sizeof(prob_str_buf), "%20.2f", prob);
1616 prob_str = prob_str_buf;
1617 }
1618 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d' cnt='%f' prob='%s'",
1619 iter().get_dest(), taken, not_taken, cnt, prob_str);
1620 }
1621 return prob;
1622 }
1623
1624 //-----------------------------branch_prediction-------------------------------
1625 float Parse::branch_prediction(float& cnt,
1626 BoolTest::mask btest,
1627 int target_bci,
1628 Node* test) {
1629 float prob = dynamic_branch_prediction(cnt, btest, test);
1630 // If prob is unknown, switch to static prediction
1631 if (prob != PROB_UNKNOWN) return prob;
1632
1633 prob = PROB_FAIR; // Set default value
1634 if (btest == BoolTest::eq) // Exactly equal test?
1635 prob = PROB_STATIC_INFREQUENT; // Assume its relatively infrequent
1636 else if (btest == BoolTest::ne)
1637 prob = PROB_STATIC_FREQUENT; // Assume its relatively frequent
1638
1639 // If this is a conditional test guarding a backwards branch,
1640 // assume its a loop-back edge. Make it a likely taken branch.
1641 if (target_bci < bci()) {
1642 if (is_osr_parse()) { // Could be a hot OSR'd loop; force deopt
1643 // Since it's an OSR, we probably have profile data, but since
1644 // branch_prediction returned PROB_UNKNOWN, the counts are too small.
1645 // Let's make a special check here for completely zero counts.
1646 ciMethodData* methodData = method()->method_data();
1647 if (!methodData->is_empty()) {
1648 ciProfileData* data = methodData->bci_to_data(bci());
1649 // Only stop for truly zero counts, which mean an unknown part
1650 // of the OSR-ed method, and we want to deopt to gather more stats.
1651 // If you have ANY counts, then this loop is simply 'cold' relative
1652 // to the OSR loop.
1653 if (data == NULL ||
1654 (data->as_BranchData()->taken() + data->as_BranchData()->not_taken() == 0)) {
1655 // This is the only way to return PROB_UNKNOWN:
1656 return PROB_UNKNOWN;
1657 }
1658 }
1659 }
1660 prob = PROB_STATIC_FREQUENT; // Likely to take backwards branch
1661 }
1662
1663 assert(prob != PROB_UNKNOWN, "must have some guess at this point");
1664 return prob;
1665 }
1666
1667 // The magic constants are chosen so as to match the output of
1668 // branch_prediction() when the profile reports a zero taken count.
1669 // It is important to distinguish zero counts unambiguously, because
1670 // some branches (e.g., _213_javac.Assembler.eliminate) validly produce
1671 // very small but nonzero probabilities, which if confused with zero
1672 // counts would keep the program recompiling indefinitely.
1673 bool Parse::seems_never_taken(float prob) const {
1674 return prob < PROB_MIN;
1675 }
1676
1677 // True if the comparison seems to be the kind that will not change its
1678 // statistics from true to false. See comments in adjust_map_after_if.
1679 // This question is only asked along paths which are already
1680 // classified as untaken (by seems_never_taken), so really,
1681 // if a path is never taken, its controlling comparison is
1682 // already acting in a stable fashion. If the comparison
1683 // seems stable, we will put an expensive uncommon trap
1684 // on the untaken path.
1685 bool Parse::seems_stable_comparison() const {
1686 if (C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if)) {
1687 return false;
1688 }
1689 return true;
1690 }
1691
1692 //-------------------------------repush_if_args--------------------------------
1693 // Push arguments of an "if" bytecode back onto the stack by adjusting _sp.
1694 inline int Parse::repush_if_args() {
1695 if (PrintOpto && WizardMode) {
1696 tty->print("defending against excessive implicit null exceptions on %s @%d in ",
1697 Bytecodes::name(iter().cur_bc()), iter().cur_bci());
1698 method()->print_name(); tty->cr();
1699 }
1700 int bc_depth = - Bytecodes::depth(iter().cur_bc());
1701 assert(bc_depth == 1 || bc_depth == 2, "only two kinds of branches");
1702 DEBUG_ONLY(sync_jvms()); // argument(n) requires a synced jvms
1703 assert(argument(0) != NULL, "must exist");
1704 assert(bc_depth == 1 || argument(1) != NULL, "two must exist");
1705 inc_sp(bc_depth);
1706 return bc_depth;
1707 }
1708
1709 //----------------------------------do_ifnull----------------------------------
1710 void Parse::do_ifnull(BoolTest::mask btest, Node *c) {
1711 int target_bci = iter().get_dest();
1712
1713 Block* branch_block = successor_for_bci(target_bci);
1714 Block* next_block = successor_for_bci(iter().next_bci());
1715
1716 float cnt;
1717 float prob = branch_prediction(cnt, btest, target_bci, c);
1718 if (prob == PROB_UNKNOWN) {
1719 // (An earlier version of do_ifnull omitted this trap for OSR methods.)
1720 if (PrintOpto && Verbose) {
1721 tty->print_cr("Never-taken edge stops compilation at bci %d", bci());
1722 }
1723 repush_if_args(); // to gather stats on loop
1724 uncommon_trap(Deoptimization::Reason_unreached,
1725 Deoptimization::Action_reinterpret,
1726 NULL, "cold");
1727 if (C->eliminate_boxing()) {
1728 // Mark the successor blocks as parsed
1729 branch_block->next_path_num();
1730 next_block->next_path_num();
1731 }
1732 return;
1733 }
1734
1735 NOT_PRODUCT(explicit_null_checks_inserted++);
1736
1737 // Generate real control flow
1738 Node *tst = _gvn.transform( new BoolNode( c, btest ) );
1739
1740 // Sanity check the probability value
1741 assert(prob > 0.0f,"Bad probability in Parser");
1742 // Need xform to put node in hash table
1743 IfNode *iff = create_and_xform_if( control(), tst, prob, cnt );
1744 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser");
1745 // True branch
1746 { PreserveJVMState pjvms(this);
1747 Node* iftrue = _gvn.transform( new IfTrueNode (iff) );
1748 set_control(iftrue);
1749
1750 if (stopped()) { // Path is dead?
1751 NOT_PRODUCT(explicit_null_checks_elided++);
1752 if (C->eliminate_boxing()) {
1753 // Mark the successor block as parsed
1754 branch_block->next_path_num();
1755 }
1756 } else { // Path is live.
1757 adjust_map_after_if(btest, c, prob, branch_block);
1758 if (!stopped()) {
1759 merge(target_bci);
1760 }
1761 }
1762 }
1763
1764 // False branch
1765 Node* iffalse = _gvn.transform( new IfFalseNode(iff) );
1766 set_control(iffalse);
1767
1768 if (stopped()) { // Path is dead?
1769 NOT_PRODUCT(explicit_null_checks_elided++);
1770 if (C->eliminate_boxing()) {
1771 // Mark the successor block as parsed
1772 next_block->next_path_num();
1773 }
1774 } else { // Path is live.
1775 adjust_map_after_if(BoolTest(btest).negate(), c, 1.0-prob, next_block);
1776 }
1777 }
1778
1779 //------------------------------------do_if------------------------------------
1780 void Parse::do_if(BoolTest::mask btest, Node* c, bool new_path, Node** ctrl_taken) {
1781 int target_bci = iter().get_dest();
1782
1783 Block* branch_block = successor_for_bci(target_bci);
1784 Block* next_block = successor_for_bci(iter().next_bci());
1785
1786 float cnt;
1787 float prob = branch_prediction(cnt, btest, target_bci, c);
1788 float untaken_prob = 1.0 - prob;
1789
1790 if (prob == PROB_UNKNOWN) {
1791 if (PrintOpto && Verbose) {
1792 tty->print_cr("Never-taken edge stops compilation at bci %d", bci());
1793 }
1794 repush_if_args(); // to gather stats on loop
1795 uncommon_trap(Deoptimization::Reason_unreached,
1796 Deoptimization::Action_reinterpret,
1797 NULL, "cold");
1798 if (C->eliminate_boxing()) {
1799 // Mark the successor blocks as parsed
1800 branch_block->next_path_num();
1801 next_block->next_path_num();
1802 }
1803 return;
1804 }
1805
1806 // Sanity check the probability value
1807 assert(0.0f < prob && prob < 1.0f,"Bad probability in Parser");
1808
1809 bool taken_if_true = true;
1810 // Convert BoolTest to canonical form:
1811 if (!BoolTest(btest).is_canonical()) {
1812 btest = BoolTest(btest).negate();
1813 taken_if_true = false;
1814 // prob is NOT updated here; it remains the probability of the taken
1815 // path (as opposed to the prob of the path guarded by an 'IfTrueNode').
1816 }
1817 assert(btest != BoolTest::eq, "!= is the only canonical exact test");
1818
1819 Node* tst0 = new BoolNode(c, btest);
1820 Node* tst = _gvn.transform(tst0);
1821 BoolTest::mask taken_btest = BoolTest::illegal;
1822 BoolTest::mask untaken_btest = BoolTest::illegal;
1823
1824 if (tst->is_Bool()) {
1825 // Refresh c from the transformed bool node, since it may be
1826 // simpler than the original c. Also re-canonicalize btest.
1827 // This wins when (Bool ne (Conv2B p) 0) => (Bool ne (CmpP p NULL)).
1828 // That can arise from statements like: if (x instanceof C) ...
1829 if (tst != tst0) {
1830 // Canonicalize one more time since transform can change it.
1831 btest = tst->as_Bool()->_test._test;
1832 if (!BoolTest(btest).is_canonical()) {
1833 // Reverse edges one more time...
1834 tst = _gvn.transform( tst->as_Bool()->negate(&_gvn) );
1835 btest = tst->as_Bool()->_test._test;
1836 assert(BoolTest(btest).is_canonical(), "sanity");
1837 taken_if_true = !taken_if_true;
1838 }
1839 c = tst->in(1);
1840 }
1841 BoolTest::mask neg_btest = BoolTest(btest).negate();
1842 taken_btest = taken_if_true ? btest : neg_btest;
1843 untaken_btest = taken_if_true ? neg_btest : btest;
1844 }
1845
1846 // Generate real control flow
1847 float true_prob = (taken_if_true ? prob : untaken_prob);
1848 IfNode* iff = create_and_map_if(control(), tst, true_prob, cnt);
1849 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser");
1850 Node* taken_branch = new IfTrueNode(iff);
1851 Node* untaken_branch = new IfFalseNode(iff);
1852 if (!taken_if_true) { // Finish conversion to canonical form
1853 Node* tmp = taken_branch;
1854 taken_branch = untaken_branch;
1855 untaken_branch = tmp;
1856 }
1857
1858 // Branch is taken:
1859 { PreserveJVMState pjvms(this);
1860 taken_branch = _gvn.transform(taken_branch);
1861 set_control(taken_branch);
1862
1863 if (stopped()) {
1864 if (C->eliminate_boxing() && !new_path) {
1865 // Mark the successor block as parsed (if we haven't created a new path)
1866 branch_block->next_path_num();
1867 }
1868 } else {
1869 adjust_map_after_if(taken_btest, c, prob, branch_block);
1870 if (!stopped()) {
1871 if (new_path) {
1872 // Merge by using a new path
1873 merge_new_path(target_bci);
1874 } else if (ctrl_taken != NULL) {
1875 // Don't merge but save taken branch to be wired by caller
1876 *ctrl_taken = control();
1877 } else {
1878 merge(target_bci);
1879 }
1880 }
1881 }
1882 }
1883
1884 untaken_branch = _gvn.transform(untaken_branch);
1885 set_control(untaken_branch);
1886
1887 // Branch not taken.
1888 if (stopped() && ctrl_taken == NULL) {
1889 if (C->eliminate_boxing()) {
1890 // Mark the successor block as parsed (if caller does not re-wire control flow)
1891 next_block->next_path_num();
1892 }
1893 } else {
1894 adjust_map_after_if(untaken_btest, c, untaken_prob, next_block);
1895 }
1896 }
1897
1898
1899 static ProfilePtrKind speculative_ptr_kind(const TypeOopPtr* t) {
1900 if (t->speculative() == NULL) {
1901 return ProfileUnknownNull;
1902 }
1903 if (t->speculative_always_null()) {
1904 return ProfileAlwaysNull;
1905 }
1906 if (t->speculative_maybe_null()) {
1907 return ProfileMaybeNull;
1908 }
1909 return ProfileNeverNull;
1910 }
1911
1912 void Parse::acmp_always_null_input(Node* input, const TypeOopPtr* tinput, BoolTest::mask btest, Node* eq_region) {
1913 inc_sp(2);
1914 Node* cast = null_check_common(input, T_OBJECT, true, NULL,
1915 !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_null_check) &&
1916 speculative_ptr_kind(tinput) == ProfileAlwaysNull);
1917 dec_sp(2);
1918 if (btest == BoolTest::ne) {
1919 {
1920 PreserveJVMState pjvms(this);
1921 replace_in_map(input, cast);
1922 int target_bci = iter().get_dest();
1923 merge(target_bci);
1924 }
1925 record_for_igvn(eq_region);
1926 set_control(_gvn.transform(eq_region));
1927 } else {
1928 replace_in_map(input, cast);
1929 }
1930 }
1931
1932 Node* Parse::acmp_null_check(Node* input, const TypeOopPtr* tinput, ProfilePtrKind input_ptr, Node*& null_ctl) {
1933 inc_sp(2);
1934 null_ctl = top();
1935 Node* cast = null_check_oop(input, &null_ctl,
1936 input_ptr == ProfileNeverNull || (input_ptr == ProfileUnknownNull && !too_many_traps_or_recompiles(Deoptimization::Reason_null_check)),
1937 false,
1938 speculative_ptr_kind(tinput) == ProfileNeverNull &&
1939 !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_null_check));
1940 dec_sp(2);
1941 assert(!stopped(), "null input should have been caught earlier");
1942 if (cast->is_InlineType()) {
1943 cast = cast->as_InlineType()->get_oop();
1944 }
1945 return cast;
1946 }
1947
1948 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) {
1949 Node* ne_region = new RegionNode(1);
1950 Node* null_ctl;
1951 Node* cast = acmp_null_check(input, tinput, input_ptr, null_ctl);
1952 ne_region->add_req(null_ctl);
1953
1954 Node* slow_ctl = type_check_receiver(cast, input_type, 1.0, &cast);
1955 {
1956 PreserveJVMState pjvms(this);
1957 inc_sp(2);
1958 set_control(slow_ctl);
1959 Deoptimization::DeoptReason reason;
1960 if (tinput->speculative_type() != NULL && !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_class_check)) {
1961 reason = Deoptimization::Reason_speculate_class_check;
1962 } else {
1963 reason = Deoptimization::Reason_class_check;
1964 }
1965 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
1966 }
1967 ne_region->add_req(control());
1968
1969 record_for_igvn(ne_region);
1970 set_control(_gvn.transform(ne_region));
1971 if (btest == BoolTest::ne) {
1972 {
1973 PreserveJVMState pjvms(this);
1974 if (null_ctl == top()) {
1975 replace_in_map(input, cast);
1976 }
1977 int target_bci = iter().get_dest();
1978 merge(target_bci);
1979 }
1980 record_for_igvn(eq_region);
1981 set_control(_gvn.transform(eq_region));
1982 } else {
1983 if (null_ctl == top()) {
1984 replace_in_map(input, cast);
1985 }
1986 set_control(_gvn.transform(ne_region));
1987 }
1988 }
1989
1990 void Parse::acmp_unknown_non_inline_type_input(Node* input, const TypeOopPtr* tinput, ProfilePtrKind input_ptr, BoolTest::mask btest, Node* eq_region) {
1991 Node* ne_region = new RegionNode(1);
1992 Node* null_ctl;
1993 Node* cast = acmp_null_check(input, tinput, input_ptr, null_ctl);
1994 ne_region->add_req(null_ctl);
1995
1996 {
1997 BuildCutout unless(this, inline_type_test(cast, /* is_inline = */ false), PROB_MAX);
1998 inc_sp(2);
1999 uncommon_trap_exact(Deoptimization::Reason_class_check, Deoptimization::Action_maybe_recompile);
2000 }
2001
2002 ne_region->add_req(control());
2003
2004 record_for_igvn(ne_region);
2005 set_control(_gvn.transform(ne_region));
2006 if (btest == BoolTest::ne) {
2007 {
2008 PreserveJVMState pjvms(this);
2009 if (null_ctl == top()) {
2010 replace_in_map(input, cast);
2011 }
2012 int target_bci = iter().get_dest();
2013 merge(target_bci);
2014 }
2015 record_for_igvn(eq_region);
2016 set_control(_gvn.transform(eq_region));
2017 } else {
2018 if (null_ctl == top()) {
2019 replace_in_map(input, cast);
2020 }
2021 set_control(_gvn.transform(ne_region));
2022 }
2023 }
2024
2025 void Parse::do_acmp(BoolTest::mask btest, Node* left, Node* right) {
2026 ciKlass* left_type = NULL;
2027 ciKlass* right_type = NULL;
2028 ProfilePtrKind left_ptr = ProfileUnknownNull;
2029 ProfilePtrKind right_ptr = ProfileUnknownNull;
2030 bool left_inline_type = true;
2031 bool right_inline_type = true;
2032
2033 // Leverage profiling at acmp
2034 if (UseACmpProfile) {
2035 method()->acmp_profiled_type(bci(), left_type, right_type, left_ptr, right_ptr, left_inline_type, right_inline_type);
2036 if (too_many_traps_or_recompiles(Deoptimization::Reason_class_check)) {
2037 left_type = NULL;
2038 right_type = NULL;
2039 left_inline_type = true;
2040 right_inline_type = true;
2041 }
2042 if (too_many_traps_or_recompiles(Deoptimization::Reason_null_check)) {
2043 left_ptr = ProfileUnknownNull;
2044 right_ptr = ProfileUnknownNull;
2045 }
2046 }
2047
2048 if (UseTypeSpeculation) {
2049 record_profile_for_speculation(left, left_type, left_ptr);
2050 record_profile_for_speculation(right, right_type, right_ptr);
2051 }
2052
2053 if (!EnableValhalla) {
2054 Node* cmp = CmpP(left, right);
2055 cmp = optimize_cmp_with_klass(cmp);
2056 do_if(btest, cmp);
2057 return;
2058 }
2059
2060 // Check for equality before potentially allocating
2061 if (left == right) {
2062 do_if(btest, makecon(TypeInt::CC_EQ));
2063 return;
2064 }
2065
2066 // Allocate inline type operands and re-execute on deoptimization
2067 if (left->is_InlineTypeBase()) {
2068 if (_gvn.type(right)->is_zero_type() ||
2069 (right->is_InlineTypeBase() && _gvn.type(right->as_InlineTypeBase()->get_is_init())->is_zero_type())) {
2070 // Null checking a scalarized but nullable inline type. Check the IsInit
2071 // input instead of the oop input to avoid keeping buffer allocations alive.
2072 Node* cmp = CmpI(left->as_InlineTypeBase()->get_is_init(), intcon(0));
2073 do_if(btest, cmp);
2074 return;
2075 } else if (left->is_InlineType()){
2076 PreserveReexecuteState preexecs(this);
2077 inc_sp(2);
2078 jvms()->set_should_reexecute(true);
2079 left = left->as_InlineType()->buffer(this)->get_oop();
2080 }
2081 }
2082 if (right->is_InlineType()) {
2083 PreserveReexecuteState preexecs(this);
2084 inc_sp(2);
2085 jvms()->set_should_reexecute(true);
2086 right = right->as_InlineType()->buffer(this)->get_oop();
2087 }
2088
2089 // First, do a normal pointer comparison
2090 const TypeOopPtr* tleft = _gvn.type(left)->isa_oopptr();
2091 const TypeOopPtr* tright = _gvn.type(right)->isa_oopptr();
2092 Node* cmp = CmpP(left, right);
2093 cmp = optimize_cmp_with_klass(cmp);
2094 if (tleft == NULL || !tleft->can_be_inline_type() ||
2095 tright == NULL || !tright->can_be_inline_type()) {
2096 // This is sufficient, if one of the operands can't be an inline type
2097 do_if(btest, cmp);
2098 return;
2099 }
2100 Node* eq_region = NULL;
2101 if (btest == BoolTest::eq) {
2102 do_if(btest, cmp, true);
2103 if (stopped()) {
2104 return;
2105 }
2106 } else {
2107 assert(btest == BoolTest::ne, "only eq or ne");
2108 Node* is_not_equal = NULL;
2109 eq_region = new RegionNode(3);
2110 {
2111 PreserveJVMState pjvms(this);
2112 do_if(btest, cmp, false, &is_not_equal);
2113 if (!stopped()) {
2114 eq_region->init_req(1, control());
2115 }
2116 }
2117 if (is_not_equal == NULL || is_not_equal->is_top()) {
2118 record_for_igvn(eq_region);
2119 set_control(_gvn.transform(eq_region));
2120 return;
2121 }
2122 set_control(is_not_equal);
2123 }
2124
2125 // Prefer speculative types if available
2126 if (!too_many_traps_or_recompiles(Deoptimization::Reason_speculate_class_check)) {
2127 if (tleft->speculative_type() != NULL) {
2128 left_type = tleft->speculative_type();
2129 }
2130 if (tright->speculative_type() != NULL) {
2131 right_type = tright->speculative_type();
2132 }
2133 }
2134
2135 if (speculative_ptr_kind(tleft) != ProfileMaybeNull && speculative_ptr_kind(tleft) != ProfileUnknownNull) {
2136 ProfilePtrKind speculative_left_ptr = speculative_ptr_kind(tleft);
2137 if (speculative_left_ptr == ProfileAlwaysNull && !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_null_assert)) {
2138 left_ptr = speculative_left_ptr;
2139 } else if (speculative_left_ptr == ProfileNeverNull && !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_null_check)) {
2140 left_ptr = speculative_left_ptr;
2141 }
2142 }
2143 if (speculative_ptr_kind(tright) != ProfileMaybeNull && speculative_ptr_kind(tright) != ProfileUnknownNull) {
2144 ProfilePtrKind speculative_right_ptr = speculative_ptr_kind(tright);
2145 if (speculative_right_ptr == ProfileAlwaysNull && !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_null_assert)) {
2146 right_ptr = speculative_right_ptr;
2147 } else if (speculative_right_ptr == ProfileNeverNull && !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_null_check)) {
2148 right_ptr = speculative_right_ptr;
2149 }
2150 }
2151
2152 if (left_ptr == ProfileAlwaysNull) {
2153 // Comparison with null. Assert the input is indeed null and we're done.
2154 acmp_always_null_input(left, tleft, btest, eq_region);
2155 return;
2156 }
2157 if (right_ptr == ProfileAlwaysNull) {
2158 // Comparison with null. Assert the input is indeed null and we're done.
2159 acmp_always_null_input(right, tright, btest, eq_region);
2160 return;
2161 }
2162 if (left_type != NULL && !left_type->is_inlinetype()) {
2163 // Comparison with an object of known type
2164 acmp_known_non_inline_type_input(left, tleft, left_ptr, left_type, btest, eq_region);
2165 return;
2166 }
2167 if (right_type != NULL && !right_type->is_inlinetype()) {
2168 // Comparison with an object of known type
2169 acmp_known_non_inline_type_input(right, tright, right_ptr, right_type, btest, eq_region);
2170 return;
2171 }
2172 if (!left_inline_type) {
2173 // Comparison with an object known not to be an inline type
2174 acmp_unknown_non_inline_type_input(left, tleft, left_ptr, btest, eq_region);
2175 return;
2176 }
2177 if (!right_inline_type) {
2178 // Comparison with an object known not to be an inline type
2179 acmp_unknown_non_inline_type_input(right, tright, right_ptr, btest, eq_region);
2180 return;
2181 }
2182
2183 // Pointers are not equal, check if first operand is non-null
2184 Node* ne_region = new RegionNode(6);
2185 Node* null_ctl;
2186 Node* not_null_right = acmp_null_check(right, tright, right_ptr, null_ctl);
2187 ne_region->init_req(1, null_ctl);
2188
2189 // First operand is non-null, check if it is an inline type
2190 Node* is_value = inline_type_test(not_null_right);
2191 IfNode* is_value_iff = create_and_map_if(control(), is_value, PROB_FAIR, COUNT_UNKNOWN);
2192 Node* not_value = _gvn.transform(new IfFalseNode(is_value_iff));
2193 ne_region->init_req(2, not_value);
2194 set_control(_gvn.transform(new IfTrueNode(is_value_iff)));
2195
2196 // The first operand is an inline type, check if the second operand is non-null
2197 Node* not_null_left = acmp_null_check(left, tleft, left_ptr, null_ctl);
2198 ne_region->init_req(3, null_ctl);
2199
2200 // Check if both operands are of the same class.
2201 Node* kls_left = load_object_klass(not_null_left);
2202 Node* kls_right = load_object_klass(not_null_right);
2203 Node* kls_cmp = CmpP(kls_left, kls_right);
2204 Node* kls_bol = _gvn.transform(new BoolNode(kls_cmp, BoolTest::ne));
2205 IfNode* kls_iff = create_and_map_if(control(), kls_bol, PROB_FAIR, COUNT_UNKNOWN);
2206 Node* kls_ne = _gvn.transform(new IfTrueNode(kls_iff));
2207 set_control(_gvn.transform(new IfFalseNode(kls_iff)));
2208 ne_region->init_req(4, kls_ne);
2209
2210 if (stopped()) {
2211 record_for_igvn(ne_region);
2212 set_control(_gvn.transform(ne_region));
2213 if (btest == BoolTest::ne) {
2214 {
2215 PreserveJVMState pjvms(this);
2216 int target_bci = iter().get_dest();
2217 merge(target_bci);
2218 }
2219 record_for_igvn(eq_region);
2220 set_control(_gvn.transform(eq_region));
2221 }
2222 return;
2223 }
2224
2225 // Both operands are values types of the same class, we need to perform a
2226 // substitutability test. Delegate to PrimitiveObjectMethods::isSubstitutable().
2227 Node* ne_io_phi = PhiNode::make(ne_region, i_o());
2228 Node* mem = reset_memory();
2229 Node* ne_mem_phi = PhiNode::make(ne_region, mem);
2230
2231 Node* eq_io_phi = NULL;
2232 Node* eq_mem_phi = NULL;
2233 if (eq_region != NULL) {
2234 eq_io_phi = PhiNode::make(eq_region, i_o());
2235 eq_mem_phi = PhiNode::make(eq_region, mem);
2236 }
2237
2238 set_all_memory(mem);
2239
2240 kill_dead_locals();
2241 ciMethod* subst_method = ciEnv::current()->PrimitiveObjectMethods_klass()->find_method(ciSymbols::isSubstitutable_name(), ciSymbols::object_object_boolean_signature());
2242 CallStaticJavaNode *call = new CallStaticJavaNode(C, TypeFunc::make(subst_method), SharedRuntime::get_resolve_static_call_stub(), subst_method);
2243 call->set_override_symbolic_info(true);
2244 call->init_req(TypeFunc::Parms, not_null_left);
2245 call->init_req(TypeFunc::Parms+1, not_null_right);
2246 inc_sp(2);
2247 set_edges_for_java_call(call, false, false);
2248 Node* ret = set_results_for_java_call(call, false, true);
2249 dec_sp(2);
2250
2251 // Test the return value of PrimitiveObjectMethods::isSubstitutable()
2252 Node* subst_cmp = _gvn.transform(new CmpINode(ret, intcon(1)));
2253 Node* ctl = C->top();
2254 if (btest == BoolTest::eq) {
2255 PreserveJVMState pjvms(this);
2256 do_if(btest, subst_cmp);
2257 if (!stopped()) {
2258 ctl = control();
2259 }
2260 } else {
2261 assert(btest == BoolTest::ne, "only eq or ne");
2262 PreserveJVMState pjvms(this);
2263 do_if(btest, subst_cmp, false, &ctl);
2264 if (!stopped()) {
2265 eq_region->init_req(2, control());
2266 eq_io_phi->init_req(2, i_o());
2267 eq_mem_phi->init_req(2, reset_memory());
2268 }
2269 }
2270 ne_region->init_req(5, ctl);
2271 ne_io_phi->init_req(5, i_o());
2272 ne_mem_phi->init_req(5, reset_memory());
2273
2274 record_for_igvn(ne_region);
2275 set_control(_gvn.transform(ne_region));
2276 set_i_o(_gvn.transform(ne_io_phi));
2277 set_all_memory(_gvn.transform(ne_mem_phi));
2278
2279 if (btest == BoolTest::ne) {
2280 {
2281 PreserveJVMState pjvms(this);
2282 int target_bci = iter().get_dest();
2283 merge(target_bci);
2284 }
2285
2286 record_for_igvn(eq_region);
2287 set_control(_gvn.transform(eq_region));
2288 set_i_o(_gvn.transform(eq_io_phi));
2289 set_all_memory(_gvn.transform(eq_mem_phi));
2290 }
2291 }
2292
2293 bool Parse::path_is_suitable_for_uncommon_trap(float prob) const {
2294 // Don't want to speculate on uncommon traps when running with -Xcomp
2295 if (!UseInterpreter) {
2296 return false;
2297 }
2298 return (seems_never_taken(prob) && seems_stable_comparison());
2299 }
2300
2301 void Parse::maybe_add_predicate_after_if(Block* path) {
2302 if (path->is_SEL_head() && path->preds_parsed() == 0) {
2303 // Add predicates at bci of if dominating the loop so traps can be
2304 // recorded on the if's profile data
2305 int bc_depth = repush_if_args();
2306 add_empty_predicates();
2307 dec_sp(bc_depth);
2308 path->set_has_predicates();
2309 }
2310 }
2311
2312
2313 //----------------------------adjust_map_after_if------------------------------
2314 // Adjust the JVM state to reflect the result of taking this path.
2315 // Basically, it means inspecting the CmpNode controlling this
2316 // branch, seeing how it constrains a tested value, and then
2317 // deciding if it's worth our while to encode this constraint
2318 // as graph nodes in the current abstract interpretation map.
2319 void Parse::adjust_map_after_if(BoolTest::mask btest, Node* c, float prob, Block* path) {
2320 if (!c->is_Cmp()) {
2321 maybe_add_predicate_after_if(path);
2322 return;
2323 }
2324
2325 if (stopped() || btest == BoolTest::illegal) {
2326 return; // nothing to do
2327 }
2328
2329 bool is_fallthrough = (path == successor_for_bci(iter().next_bci()));
2330
2331 if (path_is_suitable_for_uncommon_trap(prob)) {
2332 repush_if_args();
2333 Node* call = uncommon_trap(Deoptimization::Reason_unstable_if,
2334 Deoptimization::Action_reinterpret,
2335 NULL,
2336 (is_fallthrough ? "taken always" : "taken never"));
2337
2338 if (call != nullptr) {
2339 C->record_unstable_if_trap(new UnstableIfTrap(call->as_CallStaticJava(), path));
2340 }
2341 return;
2342 }
2343
2344 Node* val = c->in(1);
2345 Node* con = c->in(2);
2346 const Type* tcon = _gvn.type(con);
2347 const Type* tval = _gvn.type(val);
2348 bool have_con = tcon->singleton();
2349 if (tval->singleton()) {
2350 if (!have_con) {
2351 // Swap, so constant is in con.
2352 con = val;
2353 tcon = tval;
2354 val = c->in(2);
2355 tval = _gvn.type(val);
2356 btest = BoolTest(btest).commute();
2357 have_con = true;
2358 } else {
2359 // Do we have two constants? Then leave well enough alone.
2360 have_con = false;
2361 }
2362 }
2363 if (!have_con) { // remaining adjustments need a con
2364 maybe_add_predicate_after_if(path);
2365 return;
2366 }
2367
2368 sharpen_type_after_if(btest, con, tcon, val, tval);
2369 maybe_add_predicate_after_if(path);
2370 }
2371
2372
2373 static Node* extract_obj_from_klass_load(PhaseGVN* gvn, Node* n) {
2374 Node* ldk;
2375 if (n->is_DecodeNKlass()) {
2376 if (n->in(1)->Opcode() != Op_LoadNKlass) {
2377 return NULL;
2378 } else {
2379 ldk = n->in(1);
2380 }
2381 } else if (n->Opcode() != Op_LoadKlass) {
2382 return NULL;
2383 } else {
2384 ldk = n;
2385 }
2386 assert(ldk != NULL && ldk->is_Load(), "should have found a LoadKlass or LoadNKlass node");
2387
2388 Node* adr = ldk->in(MemNode::Address);
2389 intptr_t off = 0;
2390 Node* obj = AddPNode::Ideal_base_and_offset(adr, gvn, off);
2391 if (obj == NULL || off != oopDesc::klass_offset_in_bytes()) // loading oopDesc::_klass?
2392 return NULL;
2393 const TypePtr* tp = gvn->type(obj)->is_ptr();
2394 if (tp == NULL || !(tp->isa_instptr() || tp->isa_aryptr())) // is obj a Java object ptr?
2395 return NULL;
2396
2397 return obj;
2398 }
2399
2400 void Parse::sharpen_type_after_if(BoolTest::mask btest,
2401 Node* con, const Type* tcon,
2402 Node* val, const Type* tval) {
2403 // Look for opportunities to sharpen the type of a node
2404 // whose klass is compared with a constant klass.
2405 if (btest == BoolTest::eq && tcon->isa_klassptr()) {
2406 Node* obj = extract_obj_from_klass_load(&_gvn, val);
2407 const TypeOopPtr* con_type = tcon->isa_klassptr()->as_instance_type();
2408 if (obj != NULL && (con_type->isa_instptr() || con_type->isa_aryptr())) {
2409 // Found:
2410 // Bool(CmpP(LoadKlass(obj._klass), ConP(Foo.klass)), [eq])
2411 // or the narrowOop equivalent.
2412 const Type* obj_type = _gvn.type(obj);
2413 const TypeOopPtr* tboth = obj_type->join_speculative(con_type)->isa_oopptr();
2414 if (tboth != NULL && tboth->klass_is_exact() && tboth != obj_type &&
2415 tboth->higher_equal(obj_type)) {
2416 // obj has to be of the exact type Foo if the CmpP succeeds.
2417 int obj_in_map = map()->find_edge(obj);
2418 JVMState* jvms = this->jvms();
2419 if (obj_in_map >= 0 &&
2420 (jvms->is_loc(obj_in_map) || jvms->is_stk(obj_in_map))) {
2421 TypeNode* ccast = new CheckCastPPNode(control(), obj, tboth);
2422 const Type* tcc = ccast->as_Type()->type();
2423 assert(tcc != obj_type && tcc->higher_equal(obj_type), "must improve");
2424 // Delay transform() call to allow recovery of pre-cast value
2425 // at the control merge.
2426 _gvn.set_type_bottom(ccast);
2427 record_for_igvn(ccast);
2428 // Here's the payoff.
2429 replace_in_map(obj, ccast);
2430 }
2431 }
2432 }
2433 }
2434
2435 int val_in_map = map()->find_edge(val);
2436 if (val_in_map < 0) return; // replace_in_map would be useless
2437 {
2438 JVMState* jvms = this->jvms();
2439 if (!(jvms->is_loc(val_in_map) ||
2440 jvms->is_stk(val_in_map)))
2441 return; // again, it would be useless
2442 }
2443
2444 // Check for a comparison to a constant, and "know" that the compared
2445 // value is constrained on this path.
2446 assert(tcon->singleton(), "");
2447 ConstraintCastNode* ccast = NULL;
2448 Node* cast = NULL;
2449
2450 switch (btest) {
2451 case BoolTest::eq: // Constant test?
2452 {
2453 const Type* tboth = tcon->join_speculative(tval);
2454 if (tboth == tval) break; // Nothing to gain.
2455 if (tcon->isa_int()) {
2456 ccast = new CastIINode(val, tboth);
2457 } else if (tcon == TypePtr::NULL_PTR) {
2458 // Cast to null, but keep the pointer identity temporarily live.
2459 ccast = new CastPPNode(val, tboth);
2460 } else {
2461 const TypeF* tf = tcon->isa_float_constant();
2462 const TypeD* td = tcon->isa_double_constant();
2463 // Exclude tests vs float/double 0 as these could be
2464 // either +0 or -0. Just because you are equal to +0
2465 // doesn't mean you ARE +0!
2466 // Note, following code also replaces Long and Oop values.
2467 if ((!tf || tf->_f != 0.0) &&
2468 (!td || td->_d != 0.0))
2469 cast = con; // Replace non-constant val by con.
2470 }
2471 }
2472 break;
2473
2474 case BoolTest::ne:
2475 if (tcon == TypePtr::NULL_PTR) {
2476 cast = cast_not_null(val, false);
2477 }
2478 break;
2479
2480 default:
2481 // (At this point we could record int range types with CastII.)
2482 break;
2483 }
2484
2485 if (ccast != NULL) {
2486 const Type* tcc = ccast->as_Type()->type();
2487 assert(tcc != tval && tcc->higher_equal(tval), "must improve");
2488 // Delay transform() call to allow recovery of pre-cast value
2489 // at the control merge.
2490 ccast->set_req(0, control());
2491 _gvn.set_type_bottom(ccast);
2492 record_for_igvn(ccast);
2493 cast = ccast;
2494 }
2495
2496 if (cast != NULL) { // Here's the payoff.
2497 replace_in_map(val, cast);
2498 }
2499 }
2500
2501 /**
2502 * Use speculative type to optimize CmpP node: if comparison is
2503 * against the low level class, cast the object to the speculative
2504 * type if any. CmpP should then go away.
2505 *
2506 * @param c expected CmpP node
2507 * @return result of CmpP on object casted to speculative type
2508 *
2509 */
2510 Node* Parse::optimize_cmp_with_klass(Node* c) {
2511 // If this is transformed by the _gvn to a comparison with the low
2512 // level klass then we may be able to use speculation
2513 if (c->Opcode() == Op_CmpP &&
2514 (c->in(1)->Opcode() == Op_LoadKlass || c->in(1)->Opcode() == Op_DecodeNKlass) &&
2515 c->in(2)->is_Con()) {
2516 Node* load_klass = NULL;
2517 Node* decode = NULL;
2518 if (c->in(1)->Opcode() == Op_DecodeNKlass) {
2519 decode = c->in(1);
2520 load_klass = c->in(1)->in(1);
2521 } else {
2522 load_klass = c->in(1);
2523 }
2524 if (load_klass->in(2)->is_AddP()) {
2525 Node* addp = load_klass->in(2);
2526 Node* obj = addp->in(AddPNode::Address);
2527 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
2528 if (obj_type->speculative_type_not_null() != NULL) {
2529 ciKlass* k = obj_type->speculative_type();
2530 inc_sp(2);
2531 obj = maybe_cast_profiled_obj(obj, k);
2532 dec_sp(2);
2533 if (obj->is_InlineType()) {
2534 assert(obj->as_InlineType()->is_allocated(&_gvn), "must be allocated");
2535 obj = obj->as_InlineType()->get_oop();
2536 }
2537 // Make the CmpP use the casted obj
2538 addp = basic_plus_adr(obj, addp->in(AddPNode::Offset));
2539 load_klass = load_klass->clone();
2540 load_klass->set_req(2, addp);
2541 load_klass = _gvn.transform(load_klass);
2542 if (decode != NULL) {
2543 decode = decode->clone();
2544 decode->set_req(1, load_klass);
2545 load_klass = _gvn.transform(decode);
2546 }
2547 c = c->clone();
2548 c->set_req(1, load_klass);
2549 c = _gvn.transform(c);
2550 }
2551 }
2552 }
2553 return c;
2554 }
2555
2556 //------------------------------do_one_bytecode--------------------------------
2557 // Parse this bytecode, and alter the Parsers JVM->Node mapping
2558 void Parse::do_one_bytecode() {
2559 Node *a, *b, *c, *d; // Handy temps
2560 BoolTest::mask btest;
2561 int i;
2562
2563 assert(!has_exceptions(), "bytecode entry state must be clear of throws");
2564
2565 if (C->check_node_count(NodeLimitFudgeFactor * 5,
2566 "out of nodes parsing method")) {
2567 return;
2568 }
2569
2570 #ifdef ASSERT
2571 // for setting breakpoints
2572 if (TraceOptoParse) {
2573 tty->print(" @");
2574 dump_bci(bci());
2575 tty->print(" %s", Bytecodes::name(bc()));
2576 tty->cr();
2577 }
2578 #endif
2579
2580 switch (bc()) {
2581 case Bytecodes::_nop:
2582 // do nothing
2583 break;
2584 case Bytecodes::_lconst_0:
2585 push_pair(longcon(0));
2586 break;
2587
2588 case Bytecodes::_lconst_1:
2589 push_pair(longcon(1));
2590 break;
2591
2592 case Bytecodes::_fconst_0:
2593 push(zerocon(T_FLOAT));
2594 break;
2595
2596 case Bytecodes::_fconst_1:
2597 push(makecon(TypeF::ONE));
2598 break;
2599
2600 case Bytecodes::_fconst_2:
2601 push(makecon(TypeF::make(2.0f)));
2602 break;
2603
2604 case Bytecodes::_dconst_0:
2605 push_pair(zerocon(T_DOUBLE));
2606 break;
2607
2608 case Bytecodes::_dconst_1:
2609 push_pair(makecon(TypeD::ONE));
2610 break;
2611
2612 case Bytecodes::_iconst_m1:push(intcon(-1)); break;
2613 case Bytecodes::_iconst_0: push(intcon( 0)); break;
2614 case Bytecodes::_iconst_1: push(intcon( 1)); break;
2615 case Bytecodes::_iconst_2: push(intcon( 2)); break;
2616 case Bytecodes::_iconst_3: push(intcon( 3)); break;
2617 case Bytecodes::_iconst_4: push(intcon( 4)); break;
2618 case Bytecodes::_iconst_5: push(intcon( 5)); break;
2619 case Bytecodes::_bipush: push(intcon(iter().get_constant_u1())); break;
2620 case Bytecodes::_sipush: push(intcon(iter().get_constant_u2())); break;
2621 case Bytecodes::_aconst_null: push(null()); break;
2622
2623 case Bytecodes::_ldc:
2624 case Bytecodes::_ldc_w:
2625 case Bytecodes::_ldc2_w: {
2626 ciConstant constant = iter().get_constant();
2627 if (constant.is_loaded()) {
2628 const Type* con_type = Type::make_from_constant(constant);
2629 if (con_type != NULL) {
2630 push_node(con_type->basic_type(), makecon(con_type));
2631 }
2632 } else {
2633 // If the constant is unresolved or in error state, run this BC in the interpreter.
2634 if (iter().is_in_error()) {
2635 uncommon_trap(Deoptimization::make_trap_request(Deoptimization::Reason_unhandled,
2636 Deoptimization::Action_none),
2637 NULL, "constant in error state", true /* must_throw */);
2638
2639 } else {
2640 int index = iter().get_constant_pool_index();
2641 uncommon_trap(Deoptimization::make_trap_request(Deoptimization::Reason_unloaded,
2642 Deoptimization::Action_reinterpret,
2643 index),
2644 NULL, "unresolved constant", false /* must_throw */);
2645 }
2646 }
2647 break;
2648 }
2649
2650 case Bytecodes::_aload_0:
2651 push( local(0) );
2652 break;
2653 case Bytecodes::_aload_1:
2654 push( local(1) );
2655 break;
2656 case Bytecodes::_aload_2:
2657 push( local(2) );
2658 break;
2659 case Bytecodes::_aload_3:
2660 push( local(3) );
2661 break;
2662 case Bytecodes::_aload:
2663 push( local(iter().get_index()) );
2664 break;
2665
2666 case Bytecodes::_fload_0:
2667 case Bytecodes::_iload_0:
2668 push( local(0) );
2669 break;
2670 case Bytecodes::_fload_1:
2671 case Bytecodes::_iload_1:
2672 push( local(1) );
2673 break;
2674 case Bytecodes::_fload_2:
2675 case Bytecodes::_iload_2:
2676 push( local(2) );
2677 break;
2678 case Bytecodes::_fload_3:
2679 case Bytecodes::_iload_3:
2680 push( local(3) );
2681 break;
2682 case Bytecodes::_fload:
2683 case Bytecodes::_iload:
2684 push( local(iter().get_index()) );
2685 break;
2686 case Bytecodes::_lload_0:
2687 push_pair_local( 0 );
2688 break;
2689 case Bytecodes::_lload_1:
2690 push_pair_local( 1 );
2691 break;
2692 case Bytecodes::_lload_2:
2693 push_pair_local( 2 );
2694 break;
2695 case Bytecodes::_lload_3:
2696 push_pair_local( 3 );
2697 break;
2698 case Bytecodes::_lload:
2699 push_pair_local( iter().get_index() );
2700 break;
2701
2702 case Bytecodes::_dload_0:
2703 push_pair_local(0);
2704 break;
2705 case Bytecodes::_dload_1:
2706 push_pair_local(1);
2707 break;
2708 case Bytecodes::_dload_2:
2709 push_pair_local(2);
2710 break;
2711 case Bytecodes::_dload_3:
2712 push_pair_local(3);
2713 break;
2714 case Bytecodes::_dload:
2715 push_pair_local(iter().get_index());
2716 break;
2717 case Bytecodes::_fstore_0:
2718 case Bytecodes::_istore_0:
2719 case Bytecodes::_astore_0:
2720 set_local( 0, pop() );
2721 break;
2722 case Bytecodes::_fstore_1:
2723 case Bytecodes::_istore_1:
2724 case Bytecodes::_astore_1:
2725 set_local( 1, pop() );
2726 break;
2727 case Bytecodes::_fstore_2:
2728 case Bytecodes::_istore_2:
2729 case Bytecodes::_astore_2:
2730 set_local( 2, pop() );
2731 break;
2732 case Bytecodes::_fstore_3:
2733 case Bytecodes::_istore_3:
2734 case Bytecodes::_astore_3:
2735 set_local( 3, pop() );
2736 break;
2737 case Bytecodes::_fstore:
2738 case Bytecodes::_istore:
2739 case Bytecodes::_astore:
2740 set_local( iter().get_index(), pop() );
2741 break;
2742 // long stores
2743 case Bytecodes::_lstore_0:
2744 set_pair_local( 0, pop_pair() );
2745 break;
2746 case Bytecodes::_lstore_1:
2747 set_pair_local( 1, pop_pair() );
2748 break;
2749 case Bytecodes::_lstore_2:
2750 set_pair_local( 2, pop_pair() );
2751 break;
2752 case Bytecodes::_lstore_3:
2753 set_pair_local( 3, pop_pair() );
2754 break;
2755 case Bytecodes::_lstore:
2756 set_pair_local( iter().get_index(), pop_pair() );
2757 break;
2758
2759 // double stores
2760 case Bytecodes::_dstore_0:
2761 set_pair_local( 0, dprecision_rounding(pop_pair()) );
2762 break;
2763 case Bytecodes::_dstore_1:
2764 set_pair_local( 1, dprecision_rounding(pop_pair()) );
2765 break;
2766 case Bytecodes::_dstore_2:
2767 set_pair_local( 2, dprecision_rounding(pop_pair()) );
2768 break;
2769 case Bytecodes::_dstore_3:
2770 set_pair_local( 3, dprecision_rounding(pop_pair()) );
2771 break;
2772 case Bytecodes::_dstore:
2773 set_pair_local( iter().get_index(), dprecision_rounding(pop_pair()) );
2774 break;
2775
2776 case Bytecodes::_pop: dec_sp(1); break;
2777 case Bytecodes::_pop2: dec_sp(2); break;
2778 case Bytecodes::_swap:
2779 a = pop();
2780 b = pop();
2781 push(a);
2782 push(b);
2783 break;
2784 case Bytecodes::_dup:
2785 a = pop();
2786 push(a);
2787 push(a);
2788 break;
2789 case Bytecodes::_dup_x1:
2790 a = pop();
2791 b = pop();
2792 push( a );
2793 push( b );
2794 push( a );
2795 break;
2796 case Bytecodes::_dup_x2:
2797 a = pop();
2798 b = pop();
2799 c = pop();
2800 push( a );
2801 push( c );
2802 push( b );
2803 push( a );
2804 break;
2805 case Bytecodes::_dup2:
2806 a = pop();
2807 b = pop();
2808 push( b );
2809 push( a );
2810 push( b );
2811 push( a );
2812 break;
2813
2814 case Bytecodes::_dup2_x1:
2815 // before: .. c, b, a
2816 // after: .. b, a, c, b, a
2817 // not tested
2818 a = pop();
2819 b = pop();
2820 c = pop();
2821 push( b );
2822 push( a );
2823 push( c );
2824 push( b );
2825 push( a );
2826 break;
2827 case Bytecodes::_dup2_x2:
2828 // before: .. d, c, b, a
2829 // after: .. b, a, d, c, b, a
2830 // not tested
2831 a = pop();
2832 b = pop();
2833 c = pop();
2834 d = pop();
2835 push( b );
2836 push( a );
2837 push( d );
2838 push( c );
2839 push( b );
2840 push( a );
2841 break;
2842
2843 case Bytecodes::_arraylength: {
2844 // Must do null-check with value on expression stack
2845 Node *ary = null_check(peek(), T_ARRAY);
2846 // Compile-time detect of null-exception?
2847 if (stopped()) return;
2848 a = pop();
2849 push(load_array_length(a));
2850 break;
2851 }
2852
2853 case Bytecodes::_baload: array_load(T_BYTE); break;
2854 case Bytecodes::_caload: array_load(T_CHAR); break;
2855 case Bytecodes::_iaload: array_load(T_INT); break;
2856 case Bytecodes::_saload: array_load(T_SHORT); break;
2857 case Bytecodes::_faload: array_load(T_FLOAT); break;
2858 case Bytecodes::_aaload: array_load(T_OBJECT); break;
2859 case Bytecodes::_laload: array_load(T_LONG); break;
2860 case Bytecodes::_daload: array_load(T_DOUBLE); break;
2861 case Bytecodes::_bastore: array_store(T_BYTE); break;
2862 case Bytecodes::_castore: array_store(T_CHAR); break;
2863 case Bytecodes::_iastore: array_store(T_INT); break;
2864 case Bytecodes::_sastore: array_store(T_SHORT); break;
2865 case Bytecodes::_fastore: array_store(T_FLOAT); break;
2866 case Bytecodes::_aastore: array_store(T_OBJECT); break;
2867 case Bytecodes::_lastore: array_store(T_LONG); break;
2868 case Bytecodes::_dastore: array_store(T_DOUBLE); break;
2869
2870 case Bytecodes::_getfield:
2871 do_getfield();
2872 break;
2873
2874 case Bytecodes::_getstatic:
2875 do_getstatic();
2876 break;
2877
2878 case Bytecodes::_putfield:
2879 do_putfield();
2880 break;
2881
2882 case Bytecodes::_putstatic:
2883 do_putstatic();
2884 break;
2885
2886 case Bytecodes::_irem:
2887 // Must keep both values on the expression-stack during null-check
2888 zero_check_int(peek());
2889 // Compile-time detect of null-exception?
2890 if (stopped()) return;
2891 b = pop();
2892 a = pop();
2893 push(_gvn.transform(new ModINode(control(), a, b)));
2894 break;
2895 case Bytecodes::_idiv:
2896 // Must keep both values on the expression-stack during null-check
2897 zero_check_int(peek());
2898 // Compile-time detect of null-exception?
2899 if (stopped()) return;
2900 b = pop();
2901 a = pop();
2902 push( _gvn.transform( new DivINode(control(),a,b) ) );
2903 break;
2904 case Bytecodes::_imul:
2905 b = pop(); a = pop();
2906 push( _gvn.transform( new MulINode(a,b) ) );
2907 break;
2908 case Bytecodes::_iadd:
2909 b = pop(); a = pop();
2910 push( _gvn.transform( new AddINode(a,b) ) );
2911 break;
2912 case Bytecodes::_ineg:
2913 a = pop();
2914 push( _gvn.transform( new SubINode(_gvn.intcon(0),a)) );
2915 break;
2916 case Bytecodes::_isub:
2917 b = pop(); a = pop();
2918 push( _gvn.transform( new SubINode(a,b) ) );
2919 break;
2920 case Bytecodes::_iand:
2921 b = pop(); a = pop();
2922 push( _gvn.transform( new AndINode(a,b) ) );
2923 break;
2924 case Bytecodes::_ior:
2925 b = pop(); a = pop();
2926 push( _gvn.transform( new OrINode(a,b) ) );
2927 break;
2928 case Bytecodes::_ixor:
2929 b = pop(); a = pop();
2930 push( _gvn.transform( new XorINode(a,b) ) );
2931 break;
2932 case Bytecodes::_ishl:
2933 b = pop(); a = pop();
2934 push( _gvn.transform( new LShiftINode(a,b) ) );
2935 break;
2936 case Bytecodes::_ishr:
2937 b = pop(); a = pop();
2938 push( _gvn.transform( new RShiftINode(a,b) ) );
2939 break;
2940 case Bytecodes::_iushr:
2941 b = pop(); a = pop();
2942 push( _gvn.transform( new URShiftINode(a,b) ) );
2943 break;
2944
2945 case Bytecodes::_fneg:
2946 a = pop();
2947 b = _gvn.transform(new NegFNode (a));
2948 push(b);
2949 break;
2950
2951 case Bytecodes::_fsub:
2952 b = pop();
2953 a = pop();
2954 c = _gvn.transform( new SubFNode(a,b) );
2955 d = precision_rounding(c);
2956 push( d );
2957 break;
2958
2959 case Bytecodes::_fadd:
2960 b = pop();
2961 a = pop();
2962 c = _gvn.transform( new AddFNode(a,b) );
2963 d = precision_rounding(c);
2964 push( d );
2965 break;
2966
2967 case Bytecodes::_fmul:
2968 b = pop();
2969 a = pop();
2970 c = _gvn.transform( new MulFNode(a,b) );
2971 d = precision_rounding(c);
2972 push( d );
2973 break;
2974
2975 case Bytecodes::_fdiv:
2976 b = pop();
2977 a = pop();
2978 c = _gvn.transform( new DivFNode(0,a,b) );
2979 d = precision_rounding(c);
2980 push( d );
2981 break;
2982
2983 case Bytecodes::_frem:
2984 if (Matcher::has_match_rule(Op_ModF)) {
2985 // Generate a ModF node.
2986 b = pop();
2987 a = pop();
2988 c = _gvn.transform( new ModFNode(0,a,b) );
2989 d = precision_rounding(c);
2990 push( d );
2991 }
2992 else {
2993 // Generate a call.
2994 modf();
2995 }
2996 break;
2997
2998 case Bytecodes::_fcmpl:
2999 b = pop();
3000 a = pop();
3001 c = _gvn.transform( new CmpF3Node( a, b));
3002 push(c);
3003 break;
3004 case Bytecodes::_fcmpg:
3005 b = pop();
3006 a = pop();
3007
3008 // Same as fcmpl but need to flip the unordered case. Swap the inputs,
3009 // which negates the result sign except for unordered. Flip the unordered
3010 // as well by using CmpF3 which implements unordered-lesser instead of
3011 // unordered-greater semantics. Finally, commute the result bits. Result
3012 // is same as using a CmpF3Greater except we did it with CmpF3 alone.
3013 c = _gvn.transform( new CmpF3Node( b, a));
3014 c = _gvn.transform( new SubINode(_gvn.intcon(0),c) );
3015 push(c);
3016 break;
3017
3018 case Bytecodes::_f2i:
3019 a = pop();
3020 push(_gvn.transform(new ConvF2INode(a)));
3021 break;
3022
3023 case Bytecodes::_d2i:
3024 a = pop_pair();
3025 b = _gvn.transform(new ConvD2INode(a));
3026 push( b );
3027 break;
3028
3029 case Bytecodes::_f2d:
3030 a = pop();
3031 b = _gvn.transform( new ConvF2DNode(a));
3032 push_pair( b );
3033 break;
3034
3035 case Bytecodes::_d2f:
3036 a = pop_pair();
3037 b = _gvn.transform( new ConvD2FNode(a));
3038 // This breaks _227_mtrt (speed & correctness) and _222_mpegaudio (speed)
3039 //b = _gvn.transform(new RoundFloatNode(0, b) );
3040 push( b );
3041 break;
3042
3043 case Bytecodes::_l2f:
3044 if (Matcher::convL2FSupported()) {
3045 a = pop_pair();
3046 b = _gvn.transform( new ConvL2FNode(a));
3047 // For x86_32.ad, FILD doesn't restrict precision to 24 or 53 bits.
3048 // Rather than storing the result into an FP register then pushing
3049 // out to memory to round, the machine instruction that implements
3050 // ConvL2D is responsible for rounding.
3051 // c = precision_rounding(b);
3052 push(b);
3053 } else {
3054 l2f();
3055 }
3056 break;
3057
3058 case Bytecodes::_l2d:
3059 a = pop_pair();
3060 b = _gvn.transform( new ConvL2DNode(a));
3061 // For x86_32.ad, rounding is always necessary (see _l2f above).
3062 // c = dprecision_rounding(b);
3063 push_pair(b);
3064 break;
3065
3066 case Bytecodes::_f2l:
3067 a = pop();
3068 b = _gvn.transform( new ConvF2LNode(a));
3069 push_pair(b);
3070 break;
3071
3072 case Bytecodes::_d2l:
3073 a = pop_pair();
3074 b = _gvn.transform( new ConvD2LNode(a));
3075 push_pair(b);
3076 break;
3077
3078 case Bytecodes::_dsub:
3079 b = pop_pair();
3080 a = pop_pair();
3081 c = _gvn.transform( new SubDNode(a,b) );
3082 d = dprecision_rounding(c);
3083 push_pair( d );
3084 break;
3085
3086 case Bytecodes::_dadd:
3087 b = pop_pair();
3088 a = pop_pair();
3089 c = _gvn.transform( new AddDNode(a,b) );
3090 d = dprecision_rounding(c);
3091 push_pair( d );
3092 break;
3093
3094 case Bytecodes::_dmul:
3095 b = pop_pair();
3096 a = pop_pair();
3097 c = _gvn.transform( new MulDNode(a,b) );
3098 d = dprecision_rounding(c);
3099 push_pair( d );
3100 break;
3101
3102 case Bytecodes::_ddiv:
3103 b = pop_pair();
3104 a = pop_pair();
3105 c = _gvn.transform( new DivDNode(0,a,b) );
3106 d = dprecision_rounding(c);
3107 push_pair( d );
3108 break;
3109
3110 case Bytecodes::_dneg:
3111 a = pop_pair();
3112 b = _gvn.transform(new NegDNode (a));
3113 push_pair(b);
3114 break;
3115
3116 case Bytecodes::_drem:
3117 if (Matcher::has_match_rule(Op_ModD)) {
3118 // Generate a ModD node.
3119 b = pop_pair();
3120 a = pop_pair();
3121 // a % b
3122
3123 c = _gvn.transform( new ModDNode(0,a,b) );
3124 d = dprecision_rounding(c);
3125 push_pair( d );
3126 }
3127 else {
3128 // Generate a call.
3129 modd();
3130 }
3131 break;
3132
3133 case Bytecodes::_dcmpl:
3134 b = pop_pair();
3135 a = pop_pair();
3136 c = _gvn.transform( new CmpD3Node( a, b));
3137 push(c);
3138 break;
3139
3140 case Bytecodes::_dcmpg:
3141 b = pop_pair();
3142 a = pop_pair();
3143 // Same as dcmpl but need to flip the unordered case.
3144 // Commute the inputs, which negates the result sign except for unordered.
3145 // Flip the unordered as well by using CmpD3 which implements
3146 // unordered-lesser instead of unordered-greater semantics.
3147 // Finally, negate the result bits. Result is same as using a
3148 // CmpD3Greater except we did it with CmpD3 alone.
3149 c = _gvn.transform( new CmpD3Node( b, a));
3150 c = _gvn.transform( new SubINode(_gvn.intcon(0),c) );
3151 push(c);
3152 break;
3153
3154
3155 // Note for longs -> lo word is on TOS, hi word is on TOS - 1
3156 case Bytecodes::_land:
3157 b = pop_pair();
3158 a = pop_pair();
3159 c = _gvn.transform( new AndLNode(a,b) );
3160 push_pair(c);
3161 break;
3162 case Bytecodes::_lor:
3163 b = pop_pair();
3164 a = pop_pair();
3165 c = _gvn.transform( new OrLNode(a,b) );
3166 push_pair(c);
3167 break;
3168 case Bytecodes::_lxor:
3169 b = pop_pair();
3170 a = pop_pair();
3171 c = _gvn.transform( new XorLNode(a,b) );
3172 push_pair(c);
3173 break;
3174
3175 case Bytecodes::_lshl:
3176 b = pop(); // the shift count
3177 a = pop_pair(); // value to be shifted
3178 c = _gvn.transform( new LShiftLNode(a,b) );
3179 push_pair(c);
3180 break;
3181 case Bytecodes::_lshr:
3182 b = pop(); // the shift count
3183 a = pop_pair(); // value to be shifted
3184 c = _gvn.transform( new RShiftLNode(a,b) );
3185 push_pair(c);
3186 break;
3187 case Bytecodes::_lushr:
3188 b = pop(); // the shift count
3189 a = pop_pair(); // value to be shifted
3190 c = _gvn.transform( new URShiftLNode(a,b) );
3191 push_pair(c);
3192 break;
3193 case Bytecodes::_lmul:
3194 b = pop_pair();
3195 a = pop_pair();
3196 c = _gvn.transform( new MulLNode(a,b) );
3197 push_pair(c);
3198 break;
3199
3200 case Bytecodes::_lrem:
3201 // Must keep both values on the expression-stack during null-check
3202 assert(peek(0) == top(), "long word order");
3203 zero_check_long(peek(1));
3204 // Compile-time detect of null-exception?
3205 if (stopped()) return;
3206 b = pop_pair();
3207 a = pop_pair();
3208 c = _gvn.transform( new ModLNode(control(),a,b) );
3209 push_pair(c);
3210 break;
3211
3212 case Bytecodes::_ldiv:
3213 // Must keep both values on the expression-stack during null-check
3214 assert(peek(0) == top(), "long word order");
3215 zero_check_long(peek(1));
3216 // Compile-time detect of null-exception?
3217 if (stopped()) return;
3218 b = pop_pair();
3219 a = pop_pair();
3220 c = _gvn.transform( new DivLNode(control(),a,b) );
3221 push_pair(c);
3222 break;
3223
3224 case Bytecodes::_ladd:
3225 b = pop_pair();
3226 a = pop_pair();
3227 c = _gvn.transform( new AddLNode(a,b) );
3228 push_pair(c);
3229 break;
3230 case Bytecodes::_lsub:
3231 b = pop_pair();
3232 a = pop_pair();
3233 c = _gvn.transform( new SubLNode(a,b) );
3234 push_pair(c);
3235 break;
3236 case Bytecodes::_lcmp:
3237 // Safepoints are now inserted _before_ branches. The long-compare
3238 // bytecode painfully produces a 3-way value (-1,0,+1) which requires a
3239 // slew of control flow. These are usually followed by a CmpI vs zero and
3240 // a branch; this pattern then optimizes to the obvious long-compare and
3241 // branch. However, if the branch is backwards there's a Safepoint
3242 // inserted. The inserted Safepoint captures the JVM state at the
3243 // pre-branch point, i.e. it captures the 3-way value. Thus if a
3244 // long-compare is used to control a loop the debug info will force
3245 // computation of the 3-way value, even though the generated code uses a
3246 // long-compare and branch. We try to rectify the situation by inserting
3247 // a SafePoint here and have it dominate and kill the safepoint added at a
3248 // following backwards branch. At this point the JVM state merely holds 2
3249 // longs but not the 3-way value.
3250 switch (iter().next_bc()) {
3251 case Bytecodes::_ifgt:
3252 case Bytecodes::_iflt:
3253 case Bytecodes::_ifge:
3254 case Bytecodes::_ifle:
3255 case Bytecodes::_ifne:
3256 case Bytecodes::_ifeq:
3257 // If this is a backwards branch in the bytecodes, add Safepoint
3258 maybe_add_safepoint(iter().next_get_dest());
3259 default:
3260 break;
3261 }
3262 b = pop_pair();
3263 a = pop_pair();
3264 c = _gvn.transform( new CmpL3Node( a, b ));
3265 push(c);
3266 break;
3267
3268 case Bytecodes::_lneg:
3269 a = pop_pair();
3270 b = _gvn.transform( new SubLNode(longcon(0),a));
3271 push_pair(b);
3272 break;
3273 case Bytecodes::_l2i:
3274 a = pop_pair();
3275 push( _gvn.transform( new ConvL2INode(a)));
3276 break;
3277 case Bytecodes::_i2l:
3278 a = pop();
3279 b = _gvn.transform( new ConvI2LNode(a));
3280 push_pair(b);
3281 break;
3282 case Bytecodes::_i2b:
3283 // Sign extend
3284 a = pop();
3285 a = Compile::narrow_value(T_BYTE, a, NULL, &_gvn, true);
3286 push(a);
3287 break;
3288 case Bytecodes::_i2s:
3289 a = pop();
3290 a = Compile::narrow_value(T_SHORT, a, NULL, &_gvn, true);
3291 push(a);
3292 break;
3293 case Bytecodes::_i2c:
3294 a = pop();
3295 a = Compile::narrow_value(T_CHAR, a, NULL, &_gvn, true);
3296 push(a);
3297 break;
3298
3299 case Bytecodes::_i2f:
3300 a = pop();
3301 b = _gvn.transform( new ConvI2FNode(a) ) ;
3302 c = precision_rounding(b);
3303 push (b);
3304 break;
3305
3306 case Bytecodes::_i2d:
3307 a = pop();
3308 b = _gvn.transform( new ConvI2DNode(a));
3309 push_pair(b);
3310 break;
3311
3312 case Bytecodes::_iinc: // Increment local
3313 i = iter().get_index(); // Get local index
3314 set_local( i, _gvn.transform( new AddINode( _gvn.intcon(iter().get_iinc_con()), local(i) ) ) );
3315 break;
3316
3317 // Exit points of synchronized methods must have an unlock node
3318 case Bytecodes::_return:
3319 return_current(NULL);
3320 break;
3321
3322 case Bytecodes::_ireturn:
3323 case Bytecodes::_areturn:
3324 case Bytecodes::_freturn:
3325 return_current(pop());
3326 break;
3327 case Bytecodes::_lreturn:
3328 return_current(pop_pair());
3329 break;
3330 case Bytecodes::_dreturn:
3331 return_current(pop_pair());
3332 break;
3333
3334 case Bytecodes::_athrow:
3335 // null exception oop throws NULL pointer exception
3336 null_check(peek());
3337 if (stopped()) return;
3338 // Hook the thrown exception directly to subsequent handlers.
3339 if (BailoutToInterpreterForThrows) {
3340 // Keep method interpreted from now on.
3341 uncommon_trap(Deoptimization::Reason_unhandled,
3342 Deoptimization::Action_make_not_compilable);
3343 return;
3344 }
3345 if (env()->jvmti_can_post_on_exceptions()) {
3346 // check if we must post exception events, take uncommon trap if so (with must_throw = false)
3347 uncommon_trap_if_should_post_on_exceptions(Deoptimization::Reason_unhandled, false);
3348 }
3349 // Here if either can_post_on_exceptions or should_post_on_exceptions is false
3350 add_exception_state(make_exception_state(peek()));
3351 break;
3352
3353 case Bytecodes::_goto: // fall through
3354 case Bytecodes::_goto_w: {
3355 int target_bci = (bc() == Bytecodes::_goto) ? iter().get_dest() : iter().get_far_dest();
3356
3357 // If this is a backwards branch in the bytecodes, add Safepoint
3358 maybe_add_safepoint(target_bci);
3359
3360 // Merge the current control into the target basic block
3361 merge(target_bci);
3362
3363 // See if we can get some profile data and hand it off to the next block
3364 Block *target_block = block()->successor_for_bci(target_bci);
3365 if (target_block->pred_count() != 1) break;
3366 ciMethodData* methodData = method()->method_data();
3367 if (!methodData->is_mature()) break;
3368 ciProfileData* data = methodData->bci_to_data(bci());
3369 assert(data != NULL && data->is_JumpData(), "need JumpData for taken branch");
3370 int taken = ((ciJumpData*)data)->taken();
3371 taken = method()->scale_count(taken);
3372 target_block->set_count(taken);
3373 break;
3374 }
3375
3376 case Bytecodes::_ifnull: btest = BoolTest::eq; goto handle_if_null;
3377 case Bytecodes::_ifnonnull: btest = BoolTest::ne; goto handle_if_null;
3378 handle_if_null:
3379 // If this is a backwards branch in the bytecodes, add Safepoint
3380 maybe_add_safepoint(iter().get_dest());
3381 a = null();
3382 b = pop();
3383 if (b->is_InlineType()) {
3384 // Null checking a scalarized but nullable inline type. Check the IsInit
3385 // input instead of the oop input to avoid keeping buffer allocations alive
3386 c = _gvn.transform(new CmpINode(b->as_InlineType()->get_is_init(), zerocon(T_INT)));
3387 } else {
3388 if (!_gvn.type(b)->speculative_maybe_null() &&
3389 !too_many_traps(Deoptimization::Reason_speculate_null_check)) {
3390 inc_sp(1);
3391 Node* null_ctl = top();
3392 b = null_check_oop(b, &null_ctl, true, true, true);
3393 assert(null_ctl->is_top(), "no null control here");
3394 dec_sp(1);
3395 } else if (_gvn.type(b)->speculative_always_null() &&
3396 !too_many_traps(Deoptimization::Reason_speculate_null_assert)) {
3397 inc_sp(1);
3398 b = null_assert(b);
3399 dec_sp(1);
3400 }
3401 c = _gvn.transform( new CmpPNode(b, a) );
3402 }
3403 do_ifnull(btest, c);
3404 break;
3405
3406 case Bytecodes::_if_acmpeq: btest = BoolTest::eq; goto handle_if_acmp;
3407 case Bytecodes::_if_acmpne: btest = BoolTest::ne; goto handle_if_acmp;
3408 handle_if_acmp:
3409 // If this is a backwards branch in the bytecodes, add Safepoint
3410 maybe_add_safepoint(iter().get_dest());
3411 a = pop();
3412 b = pop();
3413 do_acmp(btest, b, a);
3414 break;
3415
3416 case Bytecodes::_ifeq: btest = BoolTest::eq; goto handle_ifxx;
3417 case Bytecodes::_ifne: btest = BoolTest::ne; goto handle_ifxx;
3418 case Bytecodes::_iflt: btest = BoolTest::lt; goto handle_ifxx;
3419 case Bytecodes::_ifle: btest = BoolTest::le; goto handle_ifxx;
3420 case Bytecodes::_ifgt: btest = BoolTest::gt; goto handle_ifxx;
3421 case Bytecodes::_ifge: btest = BoolTest::ge; goto handle_ifxx;
3422 handle_ifxx:
3423 // If this is a backwards branch in the bytecodes, add Safepoint
3424 maybe_add_safepoint(iter().get_dest());
3425 a = _gvn.intcon(0);
3426 b = pop();
3427 c = _gvn.transform( new CmpINode(b, a) );
3428 do_if(btest, c);
3429 break;
3430
3431 case Bytecodes::_if_icmpeq: btest = BoolTest::eq; goto handle_if_icmp;
3432 case Bytecodes::_if_icmpne: btest = BoolTest::ne; goto handle_if_icmp;
3433 case Bytecodes::_if_icmplt: btest = BoolTest::lt; goto handle_if_icmp;
3434 case Bytecodes::_if_icmple: btest = BoolTest::le; goto handle_if_icmp;
3435 case Bytecodes::_if_icmpgt: btest = BoolTest::gt; goto handle_if_icmp;
3436 case Bytecodes::_if_icmpge: btest = BoolTest::ge; goto handle_if_icmp;
3437 handle_if_icmp:
3438 // If this is a backwards branch in the bytecodes, add Safepoint
3439 maybe_add_safepoint(iter().get_dest());
3440 a = pop();
3441 b = pop();
3442 c = _gvn.transform( new CmpINode( b, a ) );
3443 do_if(btest, c);
3444 break;
3445
3446 case Bytecodes::_tableswitch:
3447 do_tableswitch();
3448 break;
3449
3450 case Bytecodes::_lookupswitch:
3451 do_lookupswitch();
3452 break;
3453
3454 case Bytecodes::_invokestatic:
3455 case Bytecodes::_invokedynamic:
3456 case Bytecodes::_invokespecial:
3457 case Bytecodes::_invokevirtual:
3458 case Bytecodes::_invokeinterface:
3459 do_call();
3460 break;
3461 case Bytecodes::_checkcast:
3462 do_checkcast();
3463 break;
3464 case Bytecodes::_instanceof:
3465 do_instanceof();
3466 break;
3467 case Bytecodes::_anewarray:
3468 do_newarray();
3469 break;
3470 case Bytecodes::_newarray:
3471 do_newarray((BasicType)iter().get_index());
3472 break;
3473 case Bytecodes::_multianewarray:
3474 do_multianewarray();
3475 break;
3476 case Bytecodes::_new:
3477 do_new();
3478 break;
3479 case Bytecodes::_aconst_init:
3480 do_aconst_init();
3481 break;
3482 case Bytecodes::_withfield:
3483 do_withfield();
3484 break;
3485
3486 case Bytecodes::_jsr:
3487 case Bytecodes::_jsr_w:
3488 do_jsr();
3489 break;
3490
3491 case Bytecodes::_ret:
3492 do_ret();
3493 break;
3494
3495
3496 case Bytecodes::_monitorenter:
3497 do_monitor_enter();
3498 break;
3499
3500 case Bytecodes::_monitorexit:
3501 do_monitor_exit();
3502 break;
3503
3504 case Bytecodes::_breakpoint:
3505 // Breakpoint set concurrently to compile
3506 // %%% use an uncommon trap?
3507 C->record_failure("breakpoint in method");
3508 return;
3509
3510 default:
3511 #ifndef PRODUCT
3512 map()->dump(99);
3513 #endif
3514 tty->print("\nUnhandled bytecode %s\n", Bytecodes::name(bc()) );
3515 ShouldNotReachHere();
3516 }
3517
3518 #ifndef PRODUCT
3519 if (C->should_print_igv(1)) {
3520 IdealGraphPrinter* printer = C->igv_printer();
3521 char buffer[256];
3522 jio_snprintf(buffer, sizeof(buffer), "Bytecode %d: %s", bci(), Bytecodes::name(bc()));
3523 bool old = printer->traverse_outs();
3524 printer->set_traverse_outs(true);
3525 printer->print_method(buffer, 4);
3526 printer->set_traverse_outs(old);
3527 }
3528 #endif
3529 }