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 "ci/ciMethodData.hpp"
27 #include "ci/ciSymbols.hpp"
28 #include "classfile/vmSymbols.hpp"
29 #include "compiler/compileLog.hpp"
30 #include "interpreter/linkResolver.hpp"
31 #include "jvm_io.h"
32 #include "memory/resourceArea.hpp"
33 #include "memory/universe.hpp"
34 #include "oops/oop.inline.hpp"
35 #include "opto/addnode.hpp"
36 #include "opto/castnode.hpp"
37 #include "opto/convertnode.hpp"
38 #include "opto/divnode.hpp"
39 #include "opto/idealGraphPrinter.hpp"
40 #include "opto/idealKit.hpp"
41 #include "opto/inlinetypenode.hpp"
42 #include "opto/matcher.hpp"
43 #include "opto/memnode.hpp"
44 #include "opto/mulnode.hpp"
45 #include "opto/opaquenode.hpp"
46 #include "opto/parse.hpp"
47 #include "opto/runtime.hpp"
48 #include "runtime/deoptimization.hpp"
49 #include "runtime/sharedRuntime.hpp"
50
51 #ifndef PRODUCT
52 extern 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
205 if (elemtype == TypeInt::BOOL) {
206 bt = T_BOOLEAN;
207 } else if (bt == T_OBJECT) {
208 elemtype = elemtype->make_oopptr();
209 const Type* tval = _gvn.type(cast_val);
210 // Based on the value to be stored, try to determine if the array is not null-free and/or not flat.
211 // This is only legal for non-null stores because the array_store_check always passes for null, even
212 // if the array is null-free. Null stores are handled in GraphKit::gen_inline_array_null_guard().
213 bool not_null_free = !tval->maybe_null() && !tval->is_oopptr()->can_be_inline_type();
214 bool not_flattened = not_null_free || (tval->is_inlinetypeptr() && !tval->inline_klass()->flatten_array());
215 if (!ary_t->is_not_null_free() && not_null_free) {
216 // Storing a non-inline type, mark array as not null-free (-> not flat).
217 ary_t = ary_t->cast_to_not_null_free();
218 Node* cast = _gvn.transform(new CheckCastPPNode(control(), ary, ary_t));
219 replace_in_map(ary, cast);
220 ary = cast;
221 } else if (!ary_t->is_not_flat() && not_flattened) {
222 // Storing a non-flattened value, mark array as not flat.
223 ary_t = ary_t->cast_to_not_flat();
224 Node* cast = _gvn.transform(new CheckCastPPNode(control(), ary, ary_t));
225 replace_in_map(ary, cast);
226 ary = cast;
227 }
228
229 if (ary_t->is_flat()) {
230 // Store to flattened inline type array
231 assert(!tval->maybe_null(), "should be guaranteed by array store check");
232 // Re-execute flattened array store if buffering triggers deoptimization
233 PreserveReexecuteState preexecs(this);
234 inc_sp(3);
235 jvms()->set_should_reexecute(true);
236 cast_val->as_InlineType()->store_flattened(this, ary, adr, NULL, 0, MO_UNORDERED | IN_HEAP | IS_ARRAY);
237 return;
238 } else if (ary_t->is_null_free()) {
239 // Store to non-flattened inline type array (elements can never be null)
240 assert(!tval->maybe_null(), "should be guaranteed by array store check");
241 if (elemtype->inline_klass()->is_empty()) {
242 // Ignore empty inline stores, array is already initialized.
243 return;
244 }
245 } else if (!ary_t->is_not_flat() && (tval != TypePtr::NULL_PTR || StressReflectiveCode)) {
246 // Array might be flattened, emit runtime checks (for NULL, a simple inline_array_null_guard is sufficient).
247 assert(UseFlatArray && !not_flattened && elemtype->is_oopptr()->can_be_inline_type() &&
248 !ary_t->klass_is_exact() && !ary_t->is_not_null_free(), "array can't be flattened");
249 IdealKit ideal(this);
250 ideal.if_then(flat_array_test(ary, /* flat = */ false)); {
251 // non-flattened
252 assert(ideal.ctrl()->in(0)->as_If()->is_flat_array_check(&_gvn), "Should be found");
253 sync_kit(ideal);
254 Node* cast_ary = inline_array_null_guard(ary, cast_val, 3);
255 inc_sp(3);
256 access_store_at(cast_ary, adr, adr_type, cast_val, elemtype, bt, MO_UNORDERED | IN_HEAP | IS_ARRAY, false);
257 dec_sp(3);
258 ideal.sync_kit(this);
259 } ideal.else_(); {
260 sync_kit(ideal);
261 // flattened
262 Node* null_ctl = top();
263 Node* val = null_check_oop(cast_val, &null_ctl);
264 if (null_ctl != top()) {
265 PreserveJVMState pjvms(this);
266 inc_sp(3);
267 set_control(null_ctl);
268 uncommon_trap(Deoptimization::Reason_null_check, Deoptimization::Action_none);
269 dec_sp(3);
270 }
271 // Try to determine the inline klass
272 ciInlineKlass* vk = NULL;
273 if (tval->is_inlinetypeptr()) {
274 vk = tval->inline_klass();
275 } else if (elemtype->is_inlinetypeptr()) {
276 vk = elemtype->inline_klass();
277 }
278 Node* casted_ary = ary;
279 if (vk != NULL && !stopped()) {
280 // Element type is known, cast and store to flattened representation
281 assert(vk->flatten_array() && elemtype->maybe_null(), "never/always flat - should be optimized");
282 ciArrayKlass* array_klass = ciArrayKlass::make(vk, /* null_free */ true);
283 const TypeAryPtr* arytype = TypeOopPtr::make_from_klass(array_klass)->isa_aryptr();
284 casted_ary = _gvn.transform(new CheckCastPPNode(control(), casted_ary, arytype));
285 Node* casted_adr = array_element_address(casted_ary, idx, T_OBJECT, arytype->size(), control());
286 if (!val->is_InlineType()) {
287 assert(!gvn().type(val)->maybe_null(), "inline type array elements should never be null");
288 val = InlineTypeNode::make_from_oop(this, val, vk);
289 }
290 // Re-execute flattened array store if buffering triggers deoptimization
291 PreserveReexecuteState preexecs(this);
292 inc_sp(3);
293 jvms()->set_should_reexecute(true);
294 val->as_InlineType()->store_flattened(this, casted_ary, casted_adr, NULL, 0, MO_UNORDERED | IN_HEAP | IS_ARRAY);
295 } else if (!stopped()) {
296 // Element type is unknown, emit runtime call
297
298 // Below membars keep this access to an unknown flattened array correctly
299 // ordered with other unknown and known flattened array accesses.
300 insert_mem_bar_volatile(Op_MemBarCPUOrder, C->get_alias_index(TypeAryPtr::INLINES));
301
302 make_runtime_call(RC_LEAF,
303 OptoRuntime::store_unknown_inline_type(),
304 CAST_FROM_FN_PTR(address, OptoRuntime::store_unknown_inline),
305 "store_unknown_inline", TypeRawPtr::BOTTOM,
306 val, casted_ary, idx);
307
308 insert_mem_bar_volatile(Op_MemBarCPUOrder, C->get_alias_index(TypeAryPtr::INLINES));
309 }
310 ideal.sync_kit(this);
311 }
312 ideal.end_if();
313 sync_kit(ideal);
314 return;
315 } else if (!ary_t->is_not_null_free()) {
316 // Array is not flattened but may be null free
317 assert(elemtype->is_oopptr()->can_be_inline_type() && !ary_t->klass_is_exact(), "array can't be null-free");
318 ary = inline_array_null_guard(ary, cast_val, 3, true);
319 }
320 }
321 inc_sp(3);
322 access_store_at(ary, adr, adr_type, val, elemtype, bt, MO_UNORDERED | IN_HEAP | IS_ARRAY);
323 dec_sp(3);
324 }
325
326
327 //------------------------------array_addressing-------------------------------
328 // Pull array and index from the stack. Compute pointer-to-element.
329 Node* Parse::array_addressing(BasicType type, int vals, const Type*& elemtype) {
330 Node *idx = peek(0+vals); // Get from stack without popping
331 Node *ary = peek(1+vals); // in case of exception
332
333 // Null check the array base, with correct stack contents
334 ary = null_check(ary, T_ARRAY);
335 // Compile-time detect of null-exception?
336 if (stopped()) return top();
337
338 const TypeAryPtr* arytype = _gvn.type(ary)->is_aryptr();
339 const TypeInt* sizetype = arytype->size();
340 elemtype = arytype->elem();
341
342 if (UseUniqueSubclasses) {
343 const Type* el = elemtype->make_ptr();
344 if (el && el->isa_instptr()) {
345 const TypeInstPtr* toop = el->is_instptr();
346 if (toop->instance_klass()->unique_concrete_subklass()) {
347 // If we load from "AbstractClass[]" we must see "ConcreteSubClass".
348 const Type* subklass = Type::get_const_type(toop->instance_klass());
349 elemtype = subklass->join_speculative(el);
350 }
351 }
352 }
353
354 // Check for big class initializers with all constant offsets
355 // feeding into a known-size array.
356 const TypeInt* idxtype = _gvn.type(idx)->is_int();
357 // See if the highest idx value is less than the lowest array bound,
358 // and if the idx value cannot be negative:
359 bool need_range_check = true;
360 if (idxtype->_hi < sizetype->_lo && idxtype->_lo >= 0) {
361 need_range_check = false;
362 if (C->log() != NULL) C->log()->elem("observe that='!need_range_check'");
363 }
364
365 if (!arytype->is_loaded()) {
366 // Only fails for some -Xcomp runs
367 // The class is unloaded. We have to run this bytecode in the interpreter.
368 ciKlass* klass = arytype->unloaded_klass();
369
370 uncommon_trap(Deoptimization::Reason_unloaded,
371 Deoptimization::Action_reinterpret,
372 klass, "!loaded array");
373 return top();
374 }
375
376 // Do the range check
377 if (GenerateRangeChecks && need_range_check) {
378 Node* tst;
379 if (sizetype->_hi <= 0) {
380 // The greatest array bound is negative, so we can conclude that we're
381 // compiling unreachable code, but the unsigned compare trick used below
382 // only works with non-negative lengths. Instead, hack "tst" to be zero so
383 // the uncommon_trap path will always be taken.
384 tst = _gvn.intcon(0);
385 } else {
386 // Range is constant in array-oop, so we can use the original state of mem
387 Node* len = load_array_length(ary);
388
389 // Test length vs index (standard trick using unsigned compare)
390 Node* chk = _gvn.transform( new CmpUNode(idx, len) );
391 BoolTest::mask btest = BoolTest::lt;
392 tst = _gvn.transform( new BoolNode(chk, btest) );
393 }
394 RangeCheckNode* rc = new RangeCheckNode(control(), tst, PROB_MAX, COUNT_UNKNOWN);
395 _gvn.set_type(rc, rc->Value(&_gvn));
396 if (!tst->is_Con()) {
397 record_for_igvn(rc);
398 }
399 set_control(_gvn.transform(new IfTrueNode(rc)));
400 // Branch to failure if out of bounds
401 {
402 PreserveJVMState pjvms(this);
403 set_control(_gvn.transform(new IfFalseNode(rc)));
404 if (C->allow_range_check_smearing()) {
405 // Do not use builtin_throw, since range checks are sometimes
406 // made more stringent by an optimistic transformation.
407 // This creates "tentative" range checks at this point,
408 // which are not guaranteed to throw exceptions.
409 // See IfNode::Ideal, is_range_check, adjust_check.
410 uncommon_trap(Deoptimization::Reason_range_check,
411 Deoptimization::Action_make_not_entrant,
412 NULL, "range_check");
413 } else {
414 // If we have already recompiled with the range-check-widening
415 // heroic optimization turned off, then we must really be throwing
416 // range check exceptions.
417 builtin_throw(Deoptimization::Reason_range_check);
418 }
419 }
420 }
421 // Check for always knowing you are throwing a range-check exception
422 if (stopped()) return top();
423
424 // This could be an access to an inline type array. We can't tell if it's
425 // flat or not. Knowing the exact type avoids runtime checks and leads to
426 // a much simpler graph shape. Check profile information.
427 if (!arytype->is_flat() && !arytype->is_not_flat()) {
428 // First check the speculative type
429 Deoptimization::DeoptReason reason = Deoptimization::Reason_speculate_class_check;
430 ciKlass* array_type = arytype->speculative_type();
431 if (too_many_traps_or_recompiles(reason) || array_type == NULL) {
432 // No speculative type, check profile data at this bci
433 array_type = NULL;
434 reason = Deoptimization::Reason_class_check;
435 if (UseArrayLoadStoreProfile && !too_many_traps_or_recompiles(reason)) {
436 ciKlass* element_type = NULL;
437 ProfilePtrKind element_ptr = ProfileMaybeNull;
438 bool flat_array = true;
439 bool null_free_array = true;
440 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
441 }
442 }
443 if (array_type != NULL) {
444 // Speculate that this array has the exact type reported by profile data
445 Node* better_ary = NULL;
446 DEBUG_ONLY(Node* old_control = control();)
447 Node* slow_ctl = type_check_receiver(ary, array_type, 1.0, &better_ary);
448 if (stopped()) {
449 // The check always fails and therefore profile information is incorrect. Don't use it.
450 assert(old_control == slow_ctl, "type check should have been removed");
451 set_control(slow_ctl);
452 } else if (!slow_ctl->is_top()) {
453 { PreserveJVMState pjvms(this);
454 set_control(slow_ctl);
455 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
456 }
457 replace_in_map(ary, better_ary);
458 ary = better_ary;
459 arytype = _gvn.type(ary)->is_aryptr();
460 elemtype = arytype->elem();
461 }
462 }
463 } else if (UseTypeSpeculation && UseArrayLoadStoreProfile) {
464 // No need to speculate: feed profile data at this bci for the
465 // array to type speculation
466 ciKlass* array_type = NULL;
467 ciKlass* element_type = NULL;
468 ProfilePtrKind element_ptr = ProfileMaybeNull;
469 bool flat_array = true;
470 bool null_free_array = true;
471 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
472 if (array_type != NULL) {
473 ary = record_profile_for_speculation(ary, array_type, ProfileMaybeNull);
474 }
475 }
476
477 // We have no exact array type from profile data. Check profile data
478 // for a non null-free or non flat array. Non null-free implies non
479 // flat so check this one first. Speculating on a non null-free
480 // array doesn't help aaload but could be profitable for a
481 // subsequent aastore.
482 if (!arytype->is_null_free() && !arytype->is_not_null_free()) {
483 bool null_free_array = true;
484 Deoptimization::DeoptReason reason = Deoptimization::Reason_none;
485 if (arytype->speculative() != NULL &&
486 arytype->speculative()->is_aryptr()->is_not_null_free() &&
487 !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_class_check)) {
488 null_free_array = false;
489 reason = Deoptimization::Reason_speculate_class_check;
490 } else if (UseArrayLoadStoreProfile && !too_many_traps_or_recompiles(Deoptimization::Reason_class_check)) {
491 ciKlass* array_type = NULL;
492 ciKlass* element_type = NULL;
493 ProfilePtrKind element_ptr = ProfileMaybeNull;
494 bool flat_array = true;
495 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
496 reason = Deoptimization::Reason_class_check;
497 }
498 if (!null_free_array) {
499 { // Deoptimize if null-free array
500 BuildCutout unless(this, null_free_array_test(load_object_klass(ary), /* null_free = */ false), PROB_MAX);
501 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
502 }
503 assert(!stopped(), "null-free array should have been caught earlier");
504 Node* better_ary = _gvn.transform(new CheckCastPPNode(control(), ary, arytype->cast_to_not_null_free()));
505 replace_in_map(ary, better_ary);
506 ary = better_ary;
507 arytype = _gvn.type(ary)->is_aryptr();
508 }
509 }
510
511 if (!arytype->is_flat() && !arytype->is_not_flat()) {
512 bool flat_array = true;
513 Deoptimization::DeoptReason reason = Deoptimization::Reason_none;
514 if (arytype->speculative() != NULL &&
515 arytype->speculative()->is_aryptr()->is_not_flat() &&
516 !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_class_check)) {
517 flat_array = false;
518 reason = Deoptimization::Reason_speculate_class_check;
519 } else if (UseArrayLoadStoreProfile && !too_many_traps_or_recompiles(reason)) {
520 ciKlass* array_type = NULL;
521 ciKlass* element_type = NULL;
522 ProfilePtrKind element_ptr = ProfileMaybeNull;
523 bool null_free_array = true;
524 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
525 reason = Deoptimization::Reason_class_check;
526 }
527 if (!flat_array) {
528 { // Deoptimize if flat array
529 BuildCutout unless(this, flat_array_test(ary, /* flat = */ false), PROB_MAX);
530 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
531 }
532 assert(!stopped(), "flat array should have been caught earlier");
533 Node* better_ary = _gvn.transform(new CheckCastPPNode(control(), ary, arytype->cast_to_not_flat()));
534 replace_in_map(ary, better_ary);
535 ary = better_ary;
536 arytype = _gvn.type(ary)->is_aryptr();
537 }
538 }
539
540 // Make array address computation control dependent to prevent it
541 // from floating above the range check during loop optimizations.
542 Node* ptr = array_element_address(ary, idx, type, sizetype, control());
543 assert(ptr != top(), "top should go hand-in-hand with stopped");
544
545 return ptr;
546 }
547
548
549 // returns IfNode
550 IfNode* Parse::jump_if_fork_int(Node* a, Node* b, BoolTest::mask mask, float prob, float cnt) {
551 Node *cmp = _gvn.transform(new CmpINode(a, b)); // two cases: shiftcount > 32 and shiftcount <= 32
552 Node *tst = _gvn.transform(new BoolNode(cmp, mask));
553 IfNode *iff = create_and_map_if(control(), tst, prob, cnt);
554 return iff;
555 }
556
557
558 // sentinel value for the target bci to mark never taken branches
559 // (according to profiling)
560 static const int never_reached = INT_MAX;
561
562 //------------------------------helper for tableswitch-------------------------
563 void Parse::jump_if_true_fork(IfNode *iff, int dest_bci_if_true, bool unc) {
564 // True branch, use existing map info
565 { PreserveJVMState pjvms(this);
566 Node *iftrue = _gvn.transform( new IfTrueNode (iff) );
567 set_control( iftrue );
568 if (unc) {
569 repush_if_args();
570 uncommon_trap(Deoptimization::Reason_unstable_if,
571 Deoptimization::Action_reinterpret,
572 NULL,
573 "taken always");
574 } else {
575 assert(dest_bci_if_true != never_reached, "inconsistent dest");
576 merge_new_path(dest_bci_if_true);
577 }
578 }
579
580 // False branch
581 Node *iffalse = _gvn.transform( new IfFalseNode(iff) );
582 set_control( iffalse );
583 }
584
585 void Parse::jump_if_false_fork(IfNode *iff, int dest_bci_if_true, bool unc) {
586 // True branch, use existing map info
587 { PreserveJVMState pjvms(this);
588 Node *iffalse = _gvn.transform( new IfFalseNode (iff) );
589 set_control( iffalse );
590 if (unc) {
591 repush_if_args();
592 uncommon_trap(Deoptimization::Reason_unstable_if,
593 Deoptimization::Action_reinterpret,
594 NULL,
595 "taken never");
596 } else {
597 assert(dest_bci_if_true != never_reached, "inconsistent dest");
598 merge_new_path(dest_bci_if_true);
599 }
600 }
601
602 // False branch
603 Node *iftrue = _gvn.transform( new IfTrueNode(iff) );
604 set_control( iftrue );
605 }
606
607 void Parse::jump_if_always_fork(int dest_bci, bool unc) {
608 // False branch, use existing map and control()
609 if (unc) {
610 repush_if_args();
611 uncommon_trap(Deoptimization::Reason_unstable_if,
612 Deoptimization::Action_reinterpret,
613 NULL,
614 "taken never");
615 } else {
616 assert(dest_bci != never_reached, "inconsistent dest");
617 merge_new_path(dest_bci);
618 }
619 }
620
621
622 extern "C" {
623 static int jint_cmp(const void *i, const void *j) {
624 int a = *(jint *)i;
625 int b = *(jint *)j;
626 return a > b ? 1 : a < b ? -1 : 0;
627 }
628 }
629
630
631 class SwitchRange : public StackObj {
632 // a range of integers coupled with a bci destination
633 jint _lo; // inclusive lower limit
634 jint _hi; // inclusive upper limit
635 int _dest;
636 float _cnt; // how many times this range was hit according to profiling
637
638 public:
639 jint lo() const { return _lo; }
640 jint hi() const { return _hi; }
641 int dest() const { return _dest; }
642 bool is_singleton() const { return _lo == _hi; }
643 float cnt() const { return _cnt; }
644
645 void setRange(jint lo, jint hi, int dest, float cnt) {
646 assert(lo <= hi, "must be a non-empty range");
647 _lo = lo, _hi = hi; _dest = dest; _cnt = cnt;
648 assert(_cnt >= 0, "");
649 }
650 bool adjoinRange(jint lo, jint hi, int dest, float cnt, bool trim_ranges) {
651 assert(lo <= hi, "must be a non-empty range");
652 if (lo == _hi+1) {
653 // see merge_ranges() comment below
654 if (trim_ranges) {
655 if (cnt == 0) {
656 if (_cnt != 0) {
657 return false;
658 }
659 if (dest != _dest) {
660 _dest = never_reached;
661 }
662 } else {
663 if (_cnt == 0) {
664 return false;
665 }
666 if (dest != _dest) {
667 return false;
668 }
669 }
670 } else {
671 if (dest != _dest) {
672 return false;
673 }
674 }
675 _hi = hi;
676 _cnt += cnt;
677 return true;
678 }
679 return false;
680 }
681
682 void set (jint value, int dest, float cnt) {
683 setRange(value, value, dest, cnt);
684 }
685 bool adjoin(jint value, int dest, float cnt, bool trim_ranges) {
686 return adjoinRange(value, value, dest, cnt, trim_ranges);
687 }
688 bool adjoin(SwitchRange& other) {
689 return adjoinRange(other._lo, other._hi, other._dest, other._cnt, false);
690 }
691
692 void print() {
693 if (is_singleton())
694 tty->print(" {%d}=>%d (cnt=%f)", lo(), dest(), cnt());
695 else if (lo() == min_jint)
696 tty->print(" {..%d}=>%d (cnt=%f)", hi(), dest(), cnt());
697 else if (hi() == max_jint)
698 tty->print(" {%d..}=>%d (cnt=%f)", lo(), dest(), cnt());
699 else
700 tty->print(" {%d..%d}=>%d (cnt=%f)", lo(), hi(), dest(), cnt());
701 }
702 };
703
704 // We try to minimize the number of ranges and the size of the taken
705 // ones using profiling data. When ranges are created,
706 // SwitchRange::adjoinRange() only allows 2 adjoining ranges to merge
707 // if both were never hit or both were hit to build longer unreached
708 // ranges. Here, we now merge adjoining ranges with the same
709 // destination and finally set destination of unreached ranges to the
710 // special value never_reached because it can help minimize the number
711 // of tests that are necessary.
712 //
713 // For instance:
714 // [0, 1] to target1 sometimes taken
715 // [1, 2] to target1 never taken
716 // [2, 3] to target2 never taken
717 // would lead to:
718 // [0, 1] to target1 sometimes taken
719 // [1, 3] never taken
720 //
721 // (first 2 ranges to target1 are not merged)
722 static void merge_ranges(SwitchRange* ranges, int& rp) {
723 if (rp == 0) {
724 return;
725 }
726 int shift = 0;
727 for (int j = 0; j < rp; j++) {
728 SwitchRange& r1 = ranges[j-shift];
729 SwitchRange& r2 = ranges[j+1];
730 if (r1.adjoin(r2)) {
731 shift++;
732 } else if (shift > 0) {
733 ranges[j+1-shift] = r2;
734 }
735 }
736 rp -= shift;
737 for (int j = 0; j <= rp; j++) {
738 SwitchRange& r = ranges[j];
739 if (r.cnt() == 0 && r.dest() != never_reached) {
740 r.setRange(r.lo(), r.hi(), never_reached, r.cnt());
741 }
742 }
743 }
744
745 //-------------------------------do_tableswitch--------------------------------
746 void Parse::do_tableswitch() {
747 // Get information about tableswitch
748 int default_dest = iter().get_dest_table(0);
749 jint lo_index = iter().get_int_table(1);
750 jint hi_index = iter().get_int_table(2);
751 int len = hi_index - lo_index + 1;
752
753 if (len < 1) {
754 // If this is a backward branch, add safepoint
755 maybe_add_safepoint(default_dest);
756 pop(); // the effect of the instruction execution on the operand stack
757 merge(default_dest);
758 return;
759 }
760
761 ciMethodData* methodData = method()->method_data();
762 ciMultiBranchData* profile = NULL;
763 if (methodData->is_mature() && UseSwitchProfiling) {
764 ciProfileData* data = methodData->bci_to_data(bci());
765 if (data != NULL && data->is_MultiBranchData()) {
766 profile = (ciMultiBranchData*)data;
767 }
768 }
769 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if);
770
771 // generate decision tree, using trichotomy when possible
772 int rnum = len+2;
773 bool makes_backward_branch = false;
774 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum);
775 int rp = -1;
776 if (lo_index != min_jint) {
777 float cnt = 1.0F;
778 if (profile != NULL) {
779 cnt = (float)profile->default_count() / (hi_index != max_jint ? 2.0F : 1.0F);
780 }
781 ranges[++rp].setRange(min_jint, lo_index-1, default_dest, cnt);
782 }
783 for (int j = 0; j < len; j++) {
784 jint match_int = lo_index+j;
785 int dest = iter().get_dest_table(j+3);
786 makes_backward_branch |= (dest <= bci());
787 float cnt = 1.0F;
788 if (profile != NULL) {
789 cnt = (float)profile->count_at(j);
790 }
791 if (rp < 0 || !ranges[rp].adjoin(match_int, dest, cnt, trim_ranges)) {
792 ranges[++rp].set(match_int, dest, cnt);
793 }
794 }
795 jint highest = lo_index+(len-1);
796 assert(ranges[rp].hi() == highest, "");
797 if (highest != max_jint) {
798 float cnt = 1.0F;
799 if (profile != NULL) {
800 cnt = (float)profile->default_count() / (lo_index != min_jint ? 2.0F : 1.0F);
801 }
802 if (!ranges[rp].adjoinRange(highest+1, max_jint, default_dest, cnt, trim_ranges)) {
803 ranges[++rp].setRange(highest+1, max_jint, default_dest, cnt);
804 }
805 }
806 assert(rp < len+2, "not too many ranges");
807
808 if (trim_ranges) {
809 merge_ranges(ranges, rp);
810 }
811
812 // Safepoint in case if backward branch observed
813 if (makes_backward_branch) {
814 add_safepoint();
815 }
816
817 Node* lookup = pop(); // lookup value
818 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]);
819 }
820
821
822 //------------------------------do_lookupswitch--------------------------------
823 void Parse::do_lookupswitch() {
824 // Get information about lookupswitch
825 int default_dest = iter().get_dest_table(0);
826 jint len = iter().get_int_table(1);
827
828 if (len < 1) { // If this is a backward branch, add safepoint
829 maybe_add_safepoint(default_dest);
830 pop(); // the effect of the instruction execution on the operand stack
831 merge(default_dest);
832 return;
833 }
834
835 ciMethodData* methodData = method()->method_data();
836 ciMultiBranchData* profile = NULL;
837 if (methodData->is_mature() && UseSwitchProfiling) {
838 ciProfileData* data = methodData->bci_to_data(bci());
839 if (data != NULL && data->is_MultiBranchData()) {
840 profile = (ciMultiBranchData*)data;
841 }
842 }
843 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if);
844
845 // generate decision tree, using trichotomy when possible
846 jint* table = NEW_RESOURCE_ARRAY(jint, len*3);
847 {
848 for (int j = 0; j < len; j++) {
849 table[3*j+0] = iter().get_int_table(2+2*j);
850 table[3*j+1] = iter().get_dest_table(2+2*j+1);
851 // Handle overflow when converting from uint to jint
852 table[3*j+2] = (profile == NULL) ? 1 : (jint)MIN2<uint>((uint)max_jint, profile->count_at(j));
853 }
854 qsort(table, len, 3*sizeof(table[0]), jint_cmp);
855 }
856
857 float default_cnt = 1.0F;
858 if (profile != NULL) {
859 juint defaults = max_juint - len;
860 default_cnt = (float)profile->default_count()/(float)defaults;
861 }
862
863 int rnum = len*2+1;
864 bool makes_backward_branch = false;
865 SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum);
866 int rp = -1;
867 for (int j = 0; j < len; j++) {
868 jint match_int = table[3*j+0];
869 jint dest = table[3*j+1];
870 jint cnt = table[3*j+2];
871 jint next_lo = rp < 0 ? min_jint : ranges[rp].hi()+1;
872 makes_backward_branch |= (dest <= bci());
873 float c = default_cnt * ((float)match_int - (float)next_lo);
874 if (match_int != next_lo && (rp < 0 || !ranges[rp].adjoinRange(next_lo, match_int-1, default_dest, c, trim_ranges))) {
875 assert(default_dest != never_reached, "sentinel value for dead destinations");
876 ranges[++rp].setRange(next_lo, match_int-1, default_dest, c);
877 }
878 if (rp < 0 || !ranges[rp].adjoin(match_int, dest, (float)cnt, trim_ranges)) {
879 assert(dest != never_reached, "sentinel value for dead destinations");
880 ranges[++rp].set(match_int, dest, (float)cnt);
881 }
882 }
883 jint highest = table[3*(len-1)];
884 assert(ranges[rp].hi() == highest, "");
885 if (highest != max_jint &&
886 !ranges[rp].adjoinRange(highest+1, max_jint, default_dest, default_cnt * ((float)max_jint - (float)highest), trim_ranges)) {
887 ranges[++rp].setRange(highest+1, max_jint, default_dest, default_cnt * ((float)max_jint - (float)highest));
888 }
889 assert(rp < rnum, "not too many ranges");
890
891 if (trim_ranges) {
892 merge_ranges(ranges, rp);
893 }
894
895 // Safepoint in case backward branch observed
896 if (makes_backward_branch) {
897 add_safepoint();
898 }
899
900 Node *lookup = pop(); // lookup value
901 jump_switch_ranges(lookup, &ranges[0], &ranges[rp]);
902 }
903
904 static float if_prob(float taken_cnt, float total_cnt) {
905 assert(taken_cnt <= total_cnt, "");
906 if (total_cnt == 0) {
907 return PROB_FAIR;
908 }
909 float p = taken_cnt / total_cnt;
910 return clamp(p, PROB_MIN, PROB_MAX);
911 }
912
913 static float if_cnt(float cnt) {
914 if (cnt == 0) {
915 return COUNT_UNKNOWN;
916 }
917 return cnt;
918 }
919
920 static float sum_of_cnts(SwitchRange *lo, SwitchRange *hi) {
921 float total_cnt = 0;
922 for (SwitchRange* sr = lo; sr <= hi; sr++) {
923 total_cnt += sr->cnt();
924 }
925 return total_cnt;
926 }
927
928 class SwitchRanges : public ResourceObj {
929 public:
930 SwitchRange* _lo;
931 SwitchRange* _hi;
932 SwitchRange* _mid;
933 float _cost;
934
935 enum {
936 Start,
937 LeftDone,
938 RightDone,
939 Done
940 } _state;
941
942 SwitchRanges(SwitchRange *lo, SwitchRange *hi)
943 : _lo(lo), _hi(hi), _mid(NULL),
944 _cost(0), _state(Start) {
945 }
946
947 SwitchRanges()
948 : _lo(NULL), _hi(NULL), _mid(NULL),
949 _cost(0), _state(Start) {}
950 };
951
952 // Estimate cost of performing a binary search on lo..hi
953 static float compute_tree_cost(SwitchRange *lo, SwitchRange *hi, float total_cnt) {
954 GrowableArray<SwitchRanges> tree;
955 SwitchRanges root(lo, hi);
956 tree.push(root);
957
958 float cost = 0;
959 do {
960 SwitchRanges& r = *tree.adr_at(tree.length()-1);
961 if (r._hi != r._lo) {
962 if (r._mid == NULL) {
963 float r_cnt = sum_of_cnts(r._lo, r._hi);
964
965 if (r_cnt == 0) {
966 tree.pop();
967 cost = 0;
968 continue;
969 }
970
971 SwitchRange* mid = NULL;
972 mid = r._lo;
973 for (float cnt = 0; ; ) {
974 assert(mid <= r._hi, "out of bounds");
975 cnt += mid->cnt();
976 if (cnt > r_cnt / 2) {
977 break;
978 }
979 mid++;
980 }
981 assert(mid <= r._hi, "out of bounds");
982 r._mid = mid;
983 r._cost = r_cnt / total_cnt;
984 }
985 r._cost += cost;
986 if (r._state < SwitchRanges::LeftDone && r._mid > r._lo) {
987 cost = 0;
988 r._state = SwitchRanges::LeftDone;
989 tree.push(SwitchRanges(r._lo, r._mid-1));
990 } else if (r._state < SwitchRanges::RightDone) {
991 cost = 0;
992 r._state = SwitchRanges::RightDone;
993 tree.push(SwitchRanges(r._mid == r._lo ? r._mid+1 : r._mid, r._hi));
994 } else {
995 tree.pop();
996 cost = r._cost;
997 }
998 } else {
999 tree.pop();
1000 cost = r._cost;
1001 }
1002 } while (tree.length() > 0);
1003
1004
1005 return cost;
1006 }
1007
1008 // It sometimes pays off to test most common ranges before the binary search
1009 void Parse::linear_search_switch_ranges(Node* key_val, SwitchRange*& lo, SwitchRange*& hi) {
1010 uint nr = hi - lo + 1;
1011 float total_cnt = sum_of_cnts(lo, hi);
1012
1013 float min = compute_tree_cost(lo, hi, total_cnt);
1014 float extra = 1;
1015 float sub = 0;
1016
1017 SwitchRange* array1 = lo;
1018 SwitchRange* array2 = NEW_RESOURCE_ARRAY(SwitchRange, nr);
1019
1020 SwitchRange* ranges = NULL;
1021
1022 while (nr >= 2) {
1023 assert(lo == array1 || lo == array2, "one the 2 already allocated arrays");
1024 ranges = (lo == array1) ? array2 : array1;
1025
1026 // Find highest frequency range
1027 SwitchRange* candidate = lo;
1028 for (SwitchRange* sr = lo+1; sr <= hi; sr++) {
1029 if (sr->cnt() > candidate->cnt()) {
1030 candidate = sr;
1031 }
1032 }
1033 SwitchRange most_freq = *candidate;
1034 if (most_freq.cnt() == 0) {
1035 break;
1036 }
1037
1038 // Copy remaining ranges into another array
1039 int shift = 0;
1040 for (uint i = 0; i < nr; i++) {
1041 SwitchRange* sr = &lo[i];
1042 if (sr != candidate) {
1043 ranges[i-shift] = *sr;
1044 } else {
1045 shift++;
1046 if (i > 0 && i < nr-1) {
1047 SwitchRange prev = lo[i-1];
1048 prev.setRange(prev.lo(), sr->hi(), prev.dest(), prev.cnt());
1049 if (prev.adjoin(lo[i+1])) {
1050 shift++;
1051 i++;
1052 }
1053 ranges[i-shift] = prev;
1054 }
1055 }
1056 }
1057 nr -= shift;
1058
1059 // Evaluate cost of testing the most common range and performing a
1060 // binary search on the other ranges
1061 float cost = extra + compute_tree_cost(&ranges[0], &ranges[nr-1], total_cnt);
1062 if (cost >= min) {
1063 break;
1064 }
1065 // swap arrays
1066 lo = &ranges[0];
1067 hi = &ranges[nr-1];
1068
1069 // It pays off: emit the test for the most common range
1070 assert(most_freq.cnt() > 0, "must be taken");
1071 Node* val = _gvn.transform(new SubINode(key_val, _gvn.intcon(most_freq.lo())));
1072 Node* cmp = _gvn.transform(new CmpUNode(val, _gvn.intcon(most_freq.hi() - most_freq.lo())));
1073 Node* tst = _gvn.transform(new BoolNode(cmp, BoolTest::le));
1074 IfNode* iff = create_and_map_if(control(), tst, if_prob(most_freq.cnt(), total_cnt), if_cnt(most_freq.cnt()));
1075 jump_if_true_fork(iff, most_freq.dest(), false);
1076
1077 sub += most_freq.cnt() / total_cnt;
1078 extra += 1 - sub;
1079 min = cost;
1080 }
1081 }
1082
1083 //----------------------------create_jump_tables-------------------------------
1084 bool Parse::create_jump_tables(Node* key_val, SwitchRange* lo, SwitchRange* hi) {
1085 // Are jumptables enabled
1086 if (!UseJumpTables) return false;
1087
1088 // Are jumptables supported
1089 if (!Matcher::has_match_rule(Op_Jump)) return false;
1090
1091 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if);
1092
1093 // Decide if a guard is needed to lop off big ranges at either (or
1094 // both) end(s) of the input set. We'll call this the default target
1095 // even though we can't be sure that it is the true "default".
1096
1097 bool needs_guard = false;
1098 int default_dest;
1099 int64_t total_outlier_size = 0;
1100 int64_t hi_size = ((int64_t)hi->hi()) - ((int64_t)hi->lo()) + 1;
1101 int64_t lo_size = ((int64_t)lo->hi()) - ((int64_t)lo->lo()) + 1;
1102
1103 if (lo->dest() == hi->dest()) {
1104 total_outlier_size = hi_size + lo_size;
1105 default_dest = lo->dest();
1106 } else if (lo_size > hi_size) {
1107 total_outlier_size = lo_size;
1108 default_dest = lo->dest();
1109 } else {
1110 total_outlier_size = hi_size;
1111 default_dest = hi->dest();
1112 }
1113
1114 float total = sum_of_cnts(lo, hi);
1115 float cost = compute_tree_cost(lo, hi, total);
1116
1117 // If a guard test will eliminate very sparse end ranges, then
1118 // it is worth the cost of an extra jump.
1119 float trimmed_cnt = 0;
1120 if (total_outlier_size > (MaxJumpTableSparseness * 4)) {
1121 needs_guard = true;
1122 if (default_dest == lo->dest()) {
1123 trimmed_cnt += lo->cnt();
1124 lo++;
1125 }
1126 if (default_dest == hi->dest()) {
1127 trimmed_cnt += hi->cnt();
1128 hi--;
1129 }
1130 }
1131
1132 // Find the total number of cases and ranges
1133 int64_t num_cases = ((int64_t)hi->hi()) - ((int64_t)lo->lo()) + 1;
1134 int num_range = hi - lo + 1;
1135
1136 // Don't create table if: too large, too small, or too sparse.
1137 if (num_cases > MaxJumpTableSize)
1138 return false;
1139 if (UseSwitchProfiling) {
1140 // MinJumpTableSize is set so with a well balanced binary tree,
1141 // when the number of ranges is MinJumpTableSize, it's cheaper to
1142 // go through a JumpNode that a tree of IfNodes. Average cost of a
1143 // tree of IfNodes with MinJumpTableSize is
1144 // log2f(MinJumpTableSize) comparisons. So if the cost computed
1145 // from profile data is less than log2f(MinJumpTableSize) then
1146 // going with the binary search is cheaper.
1147 if (cost < log2f(MinJumpTableSize)) {
1148 return false;
1149 }
1150 } else {
1151 if (num_cases < MinJumpTableSize)
1152 return false;
1153 }
1154 if (num_cases > (MaxJumpTableSparseness * num_range))
1155 return false;
1156
1157 // Normalize table lookups to zero
1158 int lowval = lo->lo();
1159 key_val = _gvn.transform( new SubINode(key_val, _gvn.intcon(lowval)) );
1160
1161 // Generate a guard to protect against input keyvals that aren't
1162 // in the switch domain.
1163 if (needs_guard) {
1164 Node* size = _gvn.intcon(num_cases);
1165 Node* cmp = _gvn.transform(new CmpUNode(key_val, size));
1166 Node* tst = _gvn.transform(new BoolNode(cmp, BoolTest::ge));
1167 IfNode* iff = create_and_map_if(control(), tst, if_prob(trimmed_cnt, total), if_cnt(trimmed_cnt));
1168 jump_if_true_fork(iff, default_dest, trim_ranges && trimmed_cnt == 0);
1169
1170 total -= trimmed_cnt;
1171 }
1172
1173 // Create an ideal node JumpTable that has projections
1174 // of all possible ranges for a switch statement
1175 // The key_val input must be converted to a pointer offset and scaled.
1176 // Compare Parse::array_addressing above.
1177
1178 // Clean the 32-bit int into a real 64-bit offset.
1179 // Otherwise, the jint value 0 might turn into an offset of 0x0800000000.
1180 // Make I2L conversion control dependent to prevent it from
1181 // floating above the range check during loop optimizations.
1182 // Do not use a narrow int type here to prevent the data path from dying
1183 // while the control path is not removed. This can happen if the type of key_val
1184 // is later known to be out of bounds of [0, num_cases] and therefore a narrow cast
1185 // would be replaced by TOP while C2 is not able to fold the corresponding range checks.
1186 // Set _carry_dependency for the cast to avoid being removed by IGVN.
1187 #ifdef _LP64
1188 key_val = C->constrained_convI2L(&_gvn, key_val, TypeInt::INT, control(), true /* carry_dependency */);
1189 #endif
1190
1191 // Shift the value by wordsize so we have an index into the table, rather
1192 // than a switch value
1193 Node *shiftWord = _gvn.MakeConX(wordSize);
1194 key_val = _gvn.transform( new MulXNode( key_val, shiftWord));
1195
1196 // Create the JumpNode
1197 Arena* arena = C->comp_arena();
1198 float* probs = (float*)arena->Amalloc(sizeof(float)*num_cases);
1199 int i = 0;
1200 if (total == 0) {
1201 for (SwitchRange* r = lo; r <= hi; r++) {
1202 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) {
1203 probs[i] = 1.0F / num_cases;
1204 }
1205 }
1206 } else {
1207 for (SwitchRange* r = lo; r <= hi; r++) {
1208 float prob = r->cnt()/total;
1209 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) {
1210 probs[i] = prob / (r->hi() - r->lo() + 1);
1211 }
1212 }
1213 }
1214
1215 ciMethodData* methodData = method()->method_data();
1216 ciMultiBranchData* profile = NULL;
1217 if (methodData->is_mature()) {
1218 ciProfileData* data = methodData->bci_to_data(bci());
1219 if (data != NULL && data->is_MultiBranchData()) {
1220 profile = (ciMultiBranchData*)data;
1221 }
1222 }
1223
1224 Node* jtn = _gvn.transform(new JumpNode(control(), key_val, num_cases, probs, profile == NULL ? COUNT_UNKNOWN : total));
1225
1226 // These are the switch destinations hanging off the jumpnode
1227 i = 0;
1228 for (SwitchRange* r = lo; r <= hi; r++) {
1229 for (int64_t j = r->lo(); j <= r->hi(); j++, i++) {
1230 Node* input = _gvn.transform(new JumpProjNode(jtn, i, r->dest(), (int)(j - lowval)));
1231 {
1232 PreserveJVMState pjvms(this);
1233 set_control(input);
1234 jump_if_always_fork(r->dest(), trim_ranges && r->cnt() == 0);
1235 }
1236 }
1237 }
1238 assert(i == num_cases, "miscount of cases");
1239 stop_and_kill_map(); // no more uses for this JVMS
1240 return true;
1241 }
1242
1243 //----------------------------jump_switch_ranges-------------------------------
1244 void Parse::jump_switch_ranges(Node* key_val, SwitchRange *lo, SwitchRange *hi, int switch_depth) {
1245 Block* switch_block = block();
1246 bool trim_ranges = !C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if);
1247
1248 if (switch_depth == 0) {
1249 // Do special processing for the top-level call.
1250 assert(lo->lo() == min_jint, "initial range must exhaust Type::INT");
1251 assert(hi->hi() == max_jint, "initial range must exhaust Type::INT");
1252
1253 // Decrement pred-numbers for the unique set of nodes.
1254 #ifdef ASSERT
1255 if (!trim_ranges) {
1256 // Ensure that the block's successors are a (duplicate-free) set.
1257 int successors_counted = 0; // block occurrences in [hi..lo]
1258 int unique_successors = switch_block->num_successors();
1259 for (int i = 0; i < unique_successors; i++) {
1260 Block* target = switch_block->successor_at(i);
1261
1262 // Check that the set of successors is the same in both places.
1263 int successors_found = 0;
1264 for (SwitchRange* p = lo; p <= hi; p++) {
1265 if (p->dest() == target->start()) successors_found++;
1266 }
1267 assert(successors_found > 0, "successor must be known");
1268 successors_counted += successors_found;
1269 }
1270 assert(successors_counted == (hi-lo)+1, "no unexpected successors");
1271 }
1272 #endif
1273
1274 // Maybe prune the inputs, based on the type of key_val.
1275 jint min_val = min_jint;
1276 jint max_val = max_jint;
1277 const TypeInt* ti = key_val->bottom_type()->isa_int();
1278 if (ti != NULL) {
1279 min_val = ti->_lo;
1280 max_val = ti->_hi;
1281 assert(min_val <= max_val, "invalid int type");
1282 }
1283 while (lo->hi() < min_val) {
1284 lo++;
1285 }
1286 if (lo->lo() < min_val) {
1287 lo->setRange(min_val, lo->hi(), lo->dest(), lo->cnt());
1288 }
1289 while (hi->lo() > max_val) {
1290 hi--;
1291 }
1292 if (hi->hi() > max_val) {
1293 hi->setRange(hi->lo(), max_val, hi->dest(), hi->cnt());
1294 }
1295
1296 linear_search_switch_ranges(key_val, lo, hi);
1297 }
1298
1299 #ifndef PRODUCT
1300 if (switch_depth == 0) {
1301 _max_switch_depth = 0;
1302 _est_switch_depth = log2i_graceful((hi - lo + 1) - 1) + 1;
1303 }
1304 #endif
1305
1306 assert(lo <= hi, "must be a non-empty set of ranges");
1307 if (lo == hi) {
1308 jump_if_always_fork(lo->dest(), trim_ranges && lo->cnt() == 0);
1309 } else {
1310 assert(lo->hi() == (lo+1)->lo()-1, "contiguous ranges");
1311 assert(hi->lo() == (hi-1)->hi()+1, "contiguous ranges");
1312
1313 if (create_jump_tables(key_val, lo, hi)) return;
1314
1315 SwitchRange* mid = NULL;
1316 float total_cnt = sum_of_cnts(lo, hi);
1317
1318 int nr = hi - lo + 1;
1319 if (UseSwitchProfiling) {
1320 // Don't keep the binary search tree balanced: pick up mid point
1321 // that split frequencies in half.
1322 float cnt = 0;
1323 for (SwitchRange* sr = lo; sr <= hi; sr++) {
1324 cnt += sr->cnt();
1325 if (cnt >= total_cnt / 2) {
1326 mid = sr;
1327 break;
1328 }
1329 }
1330 } else {
1331 mid = lo + nr/2;
1332
1333 // if there is an easy choice, pivot at a singleton:
1334 if (nr > 3 && !mid->is_singleton() && (mid-1)->is_singleton()) mid--;
1335
1336 assert(lo < mid && mid <= hi, "good pivot choice");
1337 assert(nr != 2 || mid == hi, "should pick higher of 2");
1338 assert(nr != 3 || mid == hi-1, "should pick middle of 3");
1339 }
1340
1341
1342 Node *test_val = _gvn.intcon(mid == lo ? mid->hi() : mid->lo());
1343
1344 if (mid->is_singleton()) {
1345 IfNode *iff_ne = jump_if_fork_int(key_val, test_val, BoolTest::ne, 1-if_prob(mid->cnt(), total_cnt), if_cnt(mid->cnt()));
1346 jump_if_false_fork(iff_ne, mid->dest(), trim_ranges && mid->cnt() == 0);
1347
1348 // Special Case: If there are exactly three ranges, and the high
1349 // and low range each go to the same place, omit the "gt" test,
1350 // since it will not discriminate anything.
1351 bool eq_test_only = (hi == lo+2 && hi->dest() == lo->dest() && mid == hi-1) || mid == lo;
1352
1353 // if there is a higher range, test for it and process it:
1354 if (mid < hi && !eq_test_only) {
1355 // two comparisons of same values--should enable 1 test for 2 branches
1356 // Use BoolTest::lt instead of BoolTest::gt
1357 float cnt = sum_of_cnts(lo, mid-1);
1358 IfNode *iff_lt = jump_if_fork_int(key_val, test_val, BoolTest::lt, if_prob(cnt, total_cnt), if_cnt(cnt));
1359 Node *iftrue = _gvn.transform( new IfTrueNode(iff_lt) );
1360 Node *iffalse = _gvn.transform( new IfFalseNode(iff_lt) );
1361 { PreserveJVMState pjvms(this);
1362 set_control(iffalse);
1363 jump_switch_ranges(key_val, mid+1, hi, switch_depth+1);
1364 }
1365 set_control(iftrue);
1366 }
1367
1368 } else {
1369 // mid is a range, not a singleton, so treat mid..hi as a unit
1370 float cnt = sum_of_cnts(mid == lo ? mid+1 : mid, hi);
1371 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));
1372
1373 // if there is a higher range, test for it and process it:
1374 if (mid == hi) {
1375 jump_if_true_fork(iff_ge, mid->dest(), trim_ranges && cnt == 0);
1376 } else {
1377 Node *iftrue = _gvn.transform( new IfTrueNode(iff_ge) );
1378 Node *iffalse = _gvn.transform( new IfFalseNode(iff_ge) );
1379 { PreserveJVMState pjvms(this);
1380 set_control(iftrue);
1381 jump_switch_ranges(key_val, mid == lo ? mid+1 : mid, hi, switch_depth+1);
1382 }
1383 set_control(iffalse);
1384 }
1385 }
1386
1387 // in any case, process the lower range
1388 if (mid == lo) {
1389 if (mid->is_singleton()) {
1390 jump_switch_ranges(key_val, lo+1, hi, switch_depth+1);
1391 } else {
1392 jump_if_always_fork(lo->dest(), trim_ranges && lo->cnt() == 0);
1393 }
1394 } else {
1395 jump_switch_ranges(key_val, lo, mid-1, switch_depth+1);
1396 }
1397 }
1398
1399 // Decrease pred_count for each successor after all is done.
1400 if (switch_depth == 0) {
1401 int unique_successors = switch_block->num_successors();
1402 for (int i = 0; i < unique_successors; i++) {
1403 Block* target = switch_block->successor_at(i);
1404 // Throw away the pre-allocated path for each unique successor.
1405 target->next_path_num();
1406 }
1407 }
1408
1409 #ifndef PRODUCT
1410 _max_switch_depth = MAX2(switch_depth, _max_switch_depth);
1411 if (TraceOptoParse && Verbose && WizardMode && switch_depth == 0) {
1412 SwitchRange* r;
1413 int nsing = 0;
1414 for( r = lo; r <= hi; r++ ) {
1415 if( r->is_singleton() ) nsing++;
1416 }
1417 tty->print(">>> ");
1418 _method->print_short_name();
1419 tty->print_cr(" switch decision tree");
1420 tty->print_cr(" %d ranges (%d singletons), max_depth=%d, est_depth=%d",
1421 (int) (hi-lo+1), nsing, _max_switch_depth, _est_switch_depth);
1422 if (_max_switch_depth > _est_switch_depth) {
1423 tty->print_cr("******** BAD SWITCH DEPTH ********");
1424 }
1425 tty->print(" ");
1426 for( r = lo; r <= hi; r++ ) {
1427 r->print();
1428 }
1429 tty->cr();
1430 }
1431 #endif
1432 }
1433
1434 void Parse::modf() {
1435 Node *f2 = pop();
1436 Node *f1 = pop();
1437 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::modf_Type(),
1438 CAST_FROM_FN_PTR(address, SharedRuntime::frem),
1439 "frem", NULL, //no memory effects
1440 f1, f2);
1441 Node* res = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0));
1442
1443 push(res);
1444 }
1445
1446 void Parse::modd() {
1447 Node *d2 = pop_pair();
1448 Node *d1 = pop_pair();
1449 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::Math_DD_D_Type(),
1450 CAST_FROM_FN_PTR(address, SharedRuntime::drem),
1451 "drem", NULL, //no memory effects
1452 d1, top(), d2, top());
1453 Node* res_d = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0));
1454
1455 #ifdef ASSERT
1456 Node* res_top = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 1));
1457 assert(res_top == top(), "second value must be top");
1458 #endif
1459
1460 push_pair(res_d);
1461 }
1462
1463 void Parse::l2f() {
1464 Node* f2 = pop();
1465 Node* f1 = pop();
1466 Node* c = make_runtime_call(RC_LEAF, OptoRuntime::l2f_Type(),
1467 CAST_FROM_FN_PTR(address, SharedRuntime::l2f),
1468 "l2f", NULL, //no memory effects
1469 f1, f2);
1470 Node* res = _gvn.transform(new ProjNode(c, TypeFunc::Parms + 0));
1471
1472 push(res);
1473 }
1474
1475 // Handle jsr and jsr_w bytecode
1476 void Parse::do_jsr() {
1477 assert(bc() == Bytecodes::_jsr || bc() == Bytecodes::_jsr_w, "wrong bytecode");
1478
1479 // Store information about current state, tagged with new _jsr_bci
1480 int return_bci = iter().next_bci();
1481 int jsr_bci = (bc() == Bytecodes::_jsr) ? iter().get_dest() : iter().get_far_dest();
1482
1483 // The way we do things now, there is only one successor block
1484 // for the jsr, because the target code is cloned by ciTypeFlow.
1485 Block* target = successor_for_bci(jsr_bci);
1486
1487 // What got pushed?
1488 const Type* ret_addr = target->peek();
1489 assert(ret_addr->singleton(), "must be a constant (cloned jsr body)");
1490
1491 // Effect on jsr on stack
1492 push(_gvn.makecon(ret_addr));
1493
1494 // Flow to the jsr.
1495 merge(jsr_bci);
1496 }
1497
1498 // Handle ret bytecode
1499 void Parse::do_ret() {
1500 // Find to whom we return.
1501 assert(block()->num_successors() == 1, "a ret can only go one place now");
1502 Block* target = block()->successor_at(0);
1503 assert(!target->is_ready(), "our arrival must be expected");
1504 int pnum = target->next_path_num();
1505 merge_common(target, pnum);
1506 }
1507
1508 static bool has_injected_profile(BoolTest::mask btest, Node* test, int& taken, int& not_taken) {
1509 if (btest != BoolTest::eq && btest != BoolTest::ne) {
1510 // Only ::eq and ::ne are supported for profile injection.
1511 return false;
1512 }
1513 if (test->is_Cmp() &&
1514 test->in(1)->Opcode() == Op_ProfileBoolean) {
1515 ProfileBooleanNode* profile = (ProfileBooleanNode*)test->in(1);
1516 int false_cnt = profile->false_count();
1517 int true_cnt = profile->true_count();
1518
1519 // Counts matching depends on the actual test operation (::eq or ::ne).
1520 // No need to scale the counts because profile injection was designed
1521 // to feed exact counts into VM.
1522 taken = (btest == BoolTest::eq) ? false_cnt : true_cnt;
1523 not_taken = (btest == BoolTest::eq) ? true_cnt : false_cnt;
1524
1525 profile->consume();
1526 return true;
1527 }
1528 return false;
1529 }
1530 //--------------------------dynamic_branch_prediction--------------------------
1531 // Try to gather dynamic branch prediction behavior. Return a probability
1532 // of the branch being taken and set the "cnt" field. Returns a -1.0
1533 // if we need to use static prediction for some reason.
1534 float Parse::dynamic_branch_prediction(float &cnt, BoolTest::mask btest, Node* test) {
1535 ResourceMark rm;
1536
1537 cnt = COUNT_UNKNOWN;
1538
1539 int taken = 0;
1540 int not_taken = 0;
1541
1542 bool use_mdo = !has_injected_profile(btest, test, taken, not_taken);
1543
1544 if (use_mdo) {
1545 // Use MethodData information if it is available
1546 // FIXME: free the ProfileData structure
1547 ciMethodData* methodData = method()->method_data();
1548 if (!methodData->is_mature()) return PROB_UNKNOWN;
1549 ciProfileData* data = methodData->bci_to_data(bci());
1550 if (data == NULL) {
1551 return PROB_UNKNOWN;
1552 }
1553 if (!data->is_JumpData()) return PROB_UNKNOWN;
1554
1555 // get taken and not taken values
1556 taken = data->as_JumpData()->taken();
1557 not_taken = 0;
1558 if (data->is_BranchData()) {
1559 not_taken = data->as_BranchData()->not_taken();
1560 }
1561
1562 // scale the counts to be commensurate with invocation counts:
1563 taken = method()->scale_count(taken);
1564 not_taken = method()->scale_count(not_taken);
1565 }
1566
1567 // Give up if too few (or too many, in which case the sum will overflow) counts to be meaningful.
1568 // We also check that individual counters are positive first, otherwise the sum can become positive.
1569 if (taken < 0 || not_taken < 0 || taken + not_taken < 40) {
1570 if (C->log() != NULL) {
1571 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d'", iter().get_dest(), taken, not_taken);
1572 }
1573 return PROB_UNKNOWN;
1574 }
1575
1576 // Compute frequency that we arrive here
1577 float sum = taken + not_taken;
1578 // Adjust, if this block is a cloned private block but the
1579 // Jump counts are shared. Taken the private counts for
1580 // just this path instead of the shared counts.
1581 if( block()->count() > 0 )
1582 sum = block()->count();
1583 cnt = sum / FreqCountInvocations;
1584
1585 // Pin probability to sane limits
1586 float prob;
1587 if( !taken )
1588 prob = (0+PROB_MIN) / 2;
1589 else if( !not_taken )
1590 prob = (1+PROB_MAX) / 2;
1591 else { // Compute probability of true path
1592 prob = (float)taken / (float)(taken + not_taken);
1593 if (prob > PROB_MAX) prob = PROB_MAX;
1594 if (prob < PROB_MIN) prob = PROB_MIN;
1595 }
1596
1597 assert((cnt > 0.0f) && (prob > 0.0f),
1598 "Bad frequency assignment in if");
1599
1600 if (C->log() != NULL) {
1601 const char* prob_str = NULL;
1602 if (prob >= PROB_MAX) prob_str = (prob == PROB_MAX) ? "max" : "always";
1603 if (prob <= PROB_MIN) prob_str = (prob == PROB_MIN) ? "min" : "never";
1604 char prob_str_buf[30];
1605 if (prob_str == NULL) {
1606 jio_snprintf(prob_str_buf, sizeof(prob_str_buf), "%20.2f", prob);
1607 prob_str = prob_str_buf;
1608 }
1609 C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d' cnt='%f' prob='%s'",
1610 iter().get_dest(), taken, not_taken, cnt, prob_str);
1611 }
1612 return prob;
1613 }
1614
1615 //-----------------------------branch_prediction-------------------------------
1616 float Parse::branch_prediction(float& cnt,
1617 BoolTest::mask btest,
1618 int target_bci,
1619 Node* test) {
1620 float prob = dynamic_branch_prediction(cnt, btest, test);
1621 // If prob is unknown, switch to static prediction
1622 if (prob != PROB_UNKNOWN) return prob;
1623
1624 prob = PROB_FAIR; // Set default value
1625 if (btest == BoolTest::eq) // Exactly equal test?
1626 prob = PROB_STATIC_INFREQUENT; // Assume its relatively infrequent
1627 else if (btest == BoolTest::ne)
1628 prob = PROB_STATIC_FREQUENT; // Assume its relatively frequent
1629
1630 // If this is a conditional test guarding a backwards branch,
1631 // assume its a loop-back edge. Make it a likely taken branch.
1632 if (target_bci < bci()) {
1633 if (is_osr_parse()) { // Could be a hot OSR'd loop; force deopt
1634 // Since it's an OSR, we probably have profile data, but since
1635 // branch_prediction returned PROB_UNKNOWN, the counts are too small.
1636 // Let's make a special check here for completely zero counts.
1637 ciMethodData* methodData = method()->method_data();
1638 if (!methodData->is_empty()) {
1639 ciProfileData* data = methodData->bci_to_data(bci());
1640 // Only stop for truly zero counts, which mean an unknown part
1641 // of the OSR-ed method, and we want to deopt to gather more stats.
1642 // If you have ANY counts, then this loop is simply 'cold' relative
1643 // to the OSR loop.
1644 if (data == NULL ||
1645 (data->as_BranchData()->taken() + data->as_BranchData()->not_taken() == 0)) {
1646 // This is the only way to return PROB_UNKNOWN:
1647 return PROB_UNKNOWN;
1648 }
1649 }
1650 }
1651 prob = PROB_STATIC_FREQUENT; // Likely to take backwards branch
1652 }
1653
1654 assert(prob != PROB_UNKNOWN, "must have some guess at this point");
1655 return prob;
1656 }
1657
1658 // The magic constants are chosen so as to match the output of
1659 // branch_prediction() when the profile reports a zero taken count.
1660 // It is important to distinguish zero counts unambiguously, because
1661 // some branches (e.g., _213_javac.Assembler.eliminate) validly produce
1662 // very small but nonzero probabilities, which if confused with zero
1663 // counts would keep the program recompiling indefinitely.
1664 bool Parse::seems_never_taken(float prob) const {
1665 return prob < PROB_MIN;
1666 }
1667
1668 // True if the comparison seems to be the kind that will not change its
1669 // statistics from true to false. See comments in adjust_map_after_if.
1670 // This question is only asked along paths which are already
1671 // classified as untaken (by seems_never_taken), so really,
1672 // if a path is never taken, its controlling comparison is
1673 // already acting in a stable fashion. If the comparison
1674 // seems stable, we will put an expensive uncommon trap
1675 // on the untaken path.
1676 bool Parse::seems_stable_comparison() const {
1677 if (C->too_many_traps(method(), bci(), Deoptimization::Reason_unstable_if)) {
1678 return false;
1679 }
1680 return true;
1681 }
1682
1683 //-------------------------------repush_if_args--------------------------------
1684 // Push arguments of an "if" bytecode back onto the stack by adjusting _sp.
1685 inline int Parse::repush_if_args() {
1686 if (PrintOpto && WizardMode) {
1687 tty->print("defending against excessive implicit null exceptions on %s @%d in ",
1688 Bytecodes::name(iter().cur_bc()), iter().cur_bci());
1689 method()->print_name(); tty->cr();
1690 }
1691 int bc_depth = - Bytecodes::depth(iter().cur_bc());
1692 assert(bc_depth == 1 || bc_depth == 2, "only two kinds of branches");
1693 DEBUG_ONLY(sync_jvms()); // argument(n) requires a synced jvms
1694 assert(argument(0) != NULL, "must exist");
1695 assert(bc_depth == 1 || argument(1) != NULL, "two must exist");
1696 inc_sp(bc_depth);
1697 return bc_depth;
1698 }
1699
1700 //----------------------------------do_ifnull----------------------------------
1701 void Parse::do_ifnull(BoolTest::mask btest, Node *c) {
1702 int target_bci = iter().get_dest();
1703
1704 Block* branch_block = successor_for_bci(target_bci);
1705 Block* next_block = successor_for_bci(iter().next_bci());
1706
1707 float cnt;
1708 float prob = branch_prediction(cnt, btest, target_bci, c);
1709 if (prob == PROB_UNKNOWN) {
1710 // (An earlier version of do_ifnull omitted this trap for OSR methods.)
1711 if (PrintOpto && Verbose) {
1712 tty->print_cr("Never-taken edge stops compilation at bci %d", bci());
1713 }
1714 repush_if_args(); // to gather stats on loop
1715 uncommon_trap(Deoptimization::Reason_unreached,
1716 Deoptimization::Action_reinterpret,
1717 NULL, "cold");
1718 if (C->eliminate_boxing()) {
1719 // Mark the successor blocks as parsed
1720 branch_block->next_path_num();
1721 next_block->next_path_num();
1722 }
1723 return;
1724 }
1725
1726 NOT_PRODUCT(explicit_null_checks_inserted++);
1727
1728 // Generate real control flow
1729 Node *tst = _gvn.transform( new BoolNode( c, btest ) );
1730
1731 // Sanity check the probability value
1732 assert(prob > 0.0f,"Bad probability in Parser");
1733 // Need xform to put node in hash table
1734 IfNode *iff = create_and_xform_if( control(), tst, prob, cnt );
1735 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser");
1736 // True branch
1737 { PreserveJVMState pjvms(this);
1738 Node* iftrue = _gvn.transform( new IfTrueNode (iff) );
1739 set_control(iftrue);
1740
1741 if (stopped()) { // Path is dead?
1742 NOT_PRODUCT(explicit_null_checks_elided++);
1743 if (C->eliminate_boxing()) {
1744 // Mark the successor block as parsed
1745 branch_block->next_path_num();
1746 }
1747 } else { // Path is live.
1748 adjust_map_after_if(btest, c, prob, branch_block);
1749 if (!stopped()) {
1750 merge(target_bci);
1751 }
1752 }
1753 }
1754
1755 // False branch
1756 Node* iffalse = _gvn.transform( new IfFalseNode(iff) );
1757 set_control(iffalse);
1758
1759 if (stopped()) { // Path is dead?
1760 NOT_PRODUCT(explicit_null_checks_elided++);
1761 if (C->eliminate_boxing()) {
1762 // Mark the successor block as parsed
1763 next_block->next_path_num();
1764 }
1765 } else { // Path is live.
1766 adjust_map_after_if(BoolTest(btest).negate(), c, 1.0-prob, next_block);
1767 }
1768 }
1769
1770 //------------------------------------do_if------------------------------------
1771 void Parse::do_if(BoolTest::mask btest, Node* c, bool new_path, Node** ctrl_taken) {
1772 int target_bci = iter().get_dest();
1773
1774 Block* branch_block = successor_for_bci(target_bci);
1775 Block* next_block = successor_for_bci(iter().next_bci());
1776
1777 float cnt;
1778 float prob = branch_prediction(cnt, btest, target_bci, c);
1779 float untaken_prob = 1.0 - prob;
1780
1781 if (prob == PROB_UNKNOWN) {
1782 if (PrintOpto && Verbose) {
1783 tty->print_cr("Never-taken edge stops compilation at bci %d", bci());
1784 }
1785 repush_if_args(); // to gather stats on loop
1786 uncommon_trap(Deoptimization::Reason_unreached,
1787 Deoptimization::Action_reinterpret,
1788 NULL, "cold");
1789 if (C->eliminate_boxing()) {
1790 // Mark the successor blocks as parsed
1791 branch_block->next_path_num();
1792 next_block->next_path_num();
1793 }
1794 return;
1795 }
1796
1797 // Sanity check the probability value
1798 assert(0.0f < prob && prob < 1.0f,"Bad probability in Parser");
1799
1800 bool taken_if_true = true;
1801 // Convert BoolTest to canonical form:
1802 if (!BoolTest(btest).is_canonical()) {
1803 btest = BoolTest(btest).negate();
1804 taken_if_true = false;
1805 // prob is NOT updated here; it remains the probability of the taken
1806 // path (as opposed to the prob of the path guarded by an 'IfTrueNode').
1807 }
1808 assert(btest != BoolTest::eq, "!= is the only canonical exact test");
1809
1810 Node* tst0 = new BoolNode(c, btest);
1811 Node* tst = _gvn.transform(tst0);
1812 BoolTest::mask taken_btest = BoolTest::illegal;
1813 BoolTest::mask untaken_btest = BoolTest::illegal;
1814
1815 if (tst->is_Bool()) {
1816 // Refresh c from the transformed bool node, since it may be
1817 // simpler than the original c. Also re-canonicalize btest.
1818 // This wins when (Bool ne (Conv2B p) 0) => (Bool ne (CmpP p NULL)).
1819 // That can arise from statements like: if (x instanceof C) ...
1820 if (tst != tst0) {
1821 // Canonicalize one more time since transform can change it.
1822 btest = tst->as_Bool()->_test._test;
1823 if (!BoolTest(btest).is_canonical()) {
1824 // Reverse edges one more time...
1825 tst = _gvn.transform( tst->as_Bool()->negate(&_gvn) );
1826 btest = tst->as_Bool()->_test._test;
1827 assert(BoolTest(btest).is_canonical(), "sanity");
1828 taken_if_true = !taken_if_true;
1829 }
1830 c = tst->in(1);
1831 }
1832 BoolTest::mask neg_btest = BoolTest(btest).negate();
1833 taken_btest = taken_if_true ? btest : neg_btest;
1834 untaken_btest = taken_if_true ? neg_btest : btest;
1835 }
1836
1837 // Generate real control flow
1838 float true_prob = (taken_if_true ? prob : untaken_prob);
1839 IfNode* iff = create_and_map_if(control(), tst, true_prob, cnt);
1840 assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser");
1841 Node* taken_branch = new IfTrueNode(iff);
1842 Node* untaken_branch = new IfFalseNode(iff);
1843 if (!taken_if_true) { // Finish conversion to canonical form
1844 Node* tmp = taken_branch;
1845 taken_branch = untaken_branch;
1846 untaken_branch = tmp;
1847 }
1848
1849 // Branch is taken:
1850 { PreserveJVMState pjvms(this);
1851 taken_branch = _gvn.transform(taken_branch);
1852 set_control(taken_branch);
1853
1854 if (stopped()) {
1855 if (C->eliminate_boxing() && !new_path) {
1856 // Mark the successor block as parsed (if we haven't created a new path)
1857 branch_block->next_path_num();
1858 }
1859 } else {
1860 adjust_map_after_if(taken_btest, c, prob, branch_block);
1861 if (!stopped()) {
1862 if (new_path) {
1863 // Merge by using a new path
1864 merge_new_path(target_bci);
1865 } else if (ctrl_taken != NULL) {
1866 // Don't merge but save taken branch to be wired by caller
1867 *ctrl_taken = control();
1868 } else {
1869 merge(target_bci);
1870 }
1871 }
1872 }
1873 }
1874
1875 untaken_branch = _gvn.transform(untaken_branch);
1876 set_control(untaken_branch);
1877
1878 // Branch not taken.
1879 if (stopped() && ctrl_taken == NULL) {
1880 if (C->eliminate_boxing()) {
1881 // Mark the successor block as parsed (if caller does not re-wire control flow)
1882 next_block->next_path_num();
1883 }
1884 } else {
1885 adjust_map_after_if(untaken_btest, c, untaken_prob, next_block);
1886 }
1887 }
1888
1889
1890 static ProfilePtrKind speculative_ptr_kind(const TypeOopPtr* t) {
1891 if (t->speculative() == NULL) {
1892 return ProfileUnknownNull;
1893 }
1894 if (t->speculative_always_null()) {
1895 return ProfileAlwaysNull;
1896 }
1897 if (t->speculative_maybe_null()) {
1898 return ProfileMaybeNull;
1899 }
1900 return ProfileNeverNull;
1901 }
1902
1903 void Parse::acmp_always_null_input(Node* input, const TypeOopPtr* tinput, BoolTest::mask btest, Node* eq_region) {
1904 inc_sp(2);
1905 Node* cast = null_check_common(input, T_OBJECT, true, NULL,
1906 !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_null_check) &&
1907 speculative_ptr_kind(tinput) == ProfileAlwaysNull);
1908 dec_sp(2);
1909 if (btest == BoolTest::ne) {
1910 {
1911 PreserveJVMState pjvms(this);
1912 replace_in_map(input, cast);
1913 int target_bci = iter().get_dest();
1914 merge(target_bci);
1915 }
1916 record_for_igvn(eq_region);
1917 set_control(_gvn.transform(eq_region));
1918 } else {
1919 replace_in_map(input, cast);
1920 }
1921 }
1922
1923 Node* Parse::acmp_null_check(Node* input, const TypeOopPtr* tinput, ProfilePtrKind input_ptr, Node*& null_ctl) {
1924 inc_sp(2);
1925 null_ctl = top();
1926 Node* cast = null_check_oop(input, &null_ctl,
1927 input_ptr == ProfileNeverNull || (input_ptr == ProfileUnknownNull && !too_many_traps_or_recompiles(Deoptimization::Reason_null_check)),
1928 false,
1929 speculative_ptr_kind(tinput) == ProfileNeverNull &&
1930 !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_null_check));
1931 dec_sp(2);
1932 assert(!stopped(), "null input should have been caught earlier");
1933 return cast;
1934 }
1935
1936 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) {
1937 Node* ne_region = new RegionNode(1);
1938 Node* null_ctl;
1939 Node* cast = acmp_null_check(input, tinput, input_ptr, null_ctl);
1940 ne_region->add_req(null_ctl);
1941
1942 Node* slow_ctl = type_check_receiver(cast, input_type, 1.0, &cast);
1943 {
1944 PreserveJVMState pjvms(this);
1945 inc_sp(2);
1946 set_control(slow_ctl);
1947 Deoptimization::DeoptReason reason;
1948 if (tinput->speculative_type() != NULL && !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_class_check)) {
1949 reason = Deoptimization::Reason_speculate_class_check;
1950 } else {
1951 reason = Deoptimization::Reason_class_check;
1952 }
1953 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
1954 }
1955 ne_region->add_req(control());
1956
1957 record_for_igvn(ne_region);
1958 set_control(_gvn.transform(ne_region));
1959 if (btest == BoolTest::ne) {
1960 {
1961 PreserveJVMState pjvms(this);
1962 if (null_ctl == top()) {
1963 replace_in_map(input, cast);
1964 }
1965 int target_bci = iter().get_dest();
1966 merge(target_bci);
1967 }
1968 record_for_igvn(eq_region);
1969 set_control(_gvn.transform(eq_region));
1970 } else {
1971 if (null_ctl == top()) {
1972 replace_in_map(input, cast);
1973 }
1974 set_control(_gvn.transform(ne_region));
1975 }
1976 }
1977
1978 void Parse::acmp_unknown_non_inline_type_input(Node* input, const TypeOopPtr* tinput, ProfilePtrKind input_ptr, BoolTest::mask btest, Node* eq_region) {
1979 Node* ne_region = new RegionNode(1);
1980 Node* null_ctl;
1981 Node* cast = acmp_null_check(input, tinput, input_ptr, null_ctl);
1982 ne_region->add_req(null_ctl);
1983
1984 {
1985 BuildCutout unless(this, inline_type_test(cast, /* is_inline = */ false), PROB_MAX);
1986 inc_sp(2);
1987 uncommon_trap_exact(Deoptimization::Reason_class_check, Deoptimization::Action_maybe_recompile);
1988 }
1989
1990 ne_region->add_req(control());
1991
1992 record_for_igvn(ne_region);
1993 set_control(_gvn.transform(ne_region));
1994 if (btest == BoolTest::ne) {
1995 {
1996 PreserveJVMState pjvms(this);
1997 if (null_ctl == top()) {
1998 replace_in_map(input, cast);
1999 }
2000 int target_bci = iter().get_dest();
2001 merge(target_bci);
2002 }
2003 record_for_igvn(eq_region);
2004 set_control(_gvn.transform(eq_region));
2005 } else {
2006 if (null_ctl == top()) {
2007 replace_in_map(input, cast);
2008 }
2009 set_control(_gvn.transform(ne_region));
2010 }
2011 }
2012
2013 void Parse::do_acmp(BoolTest::mask btest, Node* left, Node* right) {
2014 ciKlass* left_type = NULL;
2015 ciKlass* right_type = NULL;
2016 ProfilePtrKind left_ptr = ProfileUnknownNull;
2017 ProfilePtrKind right_ptr = ProfileUnknownNull;
2018 bool left_inline_type = true;
2019 bool right_inline_type = true;
2020
2021 // Leverage profiling at acmp
2022 if (UseACmpProfile) {
2023 method()->acmp_profiled_type(bci(), left_type, right_type, left_ptr, right_ptr, left_inline_type, right_inline_type);
2024 if (too_many_traps_or_recompiles(Deoptimization::Reason_class_check)) {
2025 left_type = NULL;
2026 right_type = NULL;
2027 left_inline_type = true;
2028 right_inline_type = true;
2029 }
2030 if (too_many_traps_or_recompiles(Deoptimization::Reason_null_check)) {
2031 left_ptr = ProfileUnknownNull;
2032 right_ptr = ProfileUnknownNull;
2033 }
2034 }
2035
2036 if (UseTypeSpeculation) {
2037 record_profile_for_speculation(left, left_type, left_ptr);
2038 record_profile_for_speculation(right, right_type, right_ptr);
2039 }
2040
2041 if (!EnableValhalla) {
2042 Node* cmp = CmpP(left, right);
2043 cmp = optimize_cmp_with_klass(cmp);
2044 do_if(btest, cmp);
2045 return;
2046 }
2047
2048 // Check for equality before potentially allocating
2049 if (left == right) {
2050 do_if(btest, makecon(TypeInt::CC_EQ));
2051 return;
2052 }
2053
2054 // Allocate inline type operands and re-execute on deoptimization
2055 if (left->is_InlineType()) {
2056 if (_gvn.type(right)->is_zero_type() ||
2057 (right->is_InlineType() && _gvn.type(right->as_InlineType()->get_is_init())->is_zero_type())) {
2058 // Null checking a scalarized but nullable inline type. Check the IsInit
2059 // input instead of the oop input to avoid keeping buffer allocations alive.
2060 Node* cmp = CmpI(left->as_InlineType()->get_is_init(), intcon(0));
2061 do_if(btest, cmp);
2062 return;
2063 } else {
2064 PreserveReexecuteState preexecs(this);
2065 inc_sp(2);
2066 jvms()->set_should_reexecute(true);
2067 left = left->as_InlineType()->buffer(this)->get_oop();
2068 }
2069 }
2070 if (right->is_InlineType()) {
2071 PreserveReexecuteState preexecs(this);
2072 inc_sp(2);
2073 jvms()->set_should_reexecute(true);
2074 right = right->as_InlineType()->buffer(this)->get_oop();
2075 }
2076
2077 // First, do a normal pointer comparison
2078 const TypeOopPtr* tleft = _gvn.type(left)->isa_oopptr();
2079 const TypeOopPtr* tright = _gvn.type(right)->isa_oopptr();
2080 Node* cmp = CmpP(left, right);
2081 cmp = optimize_cmp_with_klass(cmp);
2082 if (tleft == NULL || !tleft->can_be_inline_type() ||
2083 tright == NULL || !tright->can_be_inline_type()) {
2084 // This is sufficient, if one of the operands can't be an inline type
2085 do_if(btest, cmp);
2086 return;
2087 }
2088 Node* eq_region = NULL;
2089 if (btest == BoolTest::eq) {
2090 do_if(btest, cmp, true);
2091 if (stopped()) {
2092 return;
2093 }
2094 } else {
2095 assert(btest == BoolTest::ne, "only eq or ne");
2096 Node* is_not_equal = NULL;
2097 eq_region = new RegionNode(3);
2098 {
2099 PreserveJVMState pjvms(this);
2100 do_if(btest, cmp, false, &is_not_equal);
2101 if (!stopped()) {
2102 eq_region->init_req(1, control());
2103 }
2104 }
2105 if (is_not_equal == NULL || is_not_equal->is_top()) {
2106 record_for_igvn(eq_region);
2107 set_control(_gvn.transform(eq_region));
2108 return;
2109 }
2110 set_control(is_not_equal);
2111 }
2112
2113 // Prefer speculative types if available
2114 if (!too_many_traps_or_recompiles(Deoptimization::Reason_speculate_class_check)) {
2115 if (tleft->speculative_type() != NULL) {
2116 left_type = tleft->speculative_type();
2117 }
2118 if (tright->speculative_type() != NULL) {
2119 right_type = tright->speculative_type();
2120 }
2121 }
2122
2123 if (speculative_ptr_kind(tleft) != ProfileMaybeNull && speculative_ptr_kind(tleft) != ProfileUnknownNull) {
2124 ProfilePtrKind speculative_left_ptr = speculative_ptr_kind(tleft);
2125 if (speculative_left_ptr == ProfileAlwaysNull && !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_null_assert)) {
2126 left_ptr = speculative_left_ptr;
2127 } else if (speculative_left_ptr == ProfileNeverNull && !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_null_check)) {
2128 left_ptr = speculative_left_ptr;
2129 }
2130 }
2131 if (speculative_ptr_kind(tright) != ProfileMaybeNull && speculative_ptr_kind(tright) != ProfileUnknownNull) {
2132 ProfilePtrKind speculative_right_ptr = speculative_ptr_kind(tright);
2133 if (speculative_right_ptr == ProfileAlwaysNull && !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_null_assert)) {
2134 right_ptr = speculative_right_ptr;
2135 } else if (speculative_right_ptr == ProfileNeverNull && !too_many_traps_or_recompiles(Deoptimization::Reason_speculate_null_check)) {
2136 right_ptr = speculative_right_ptr;
2137 }
2138 }
2139
2140 if (left_ptr == ProfileAlwaysNull) {
2141 // Comparison with null. Assert the input is indeed null and we're done.
2142 acmp_always_null_input(left, tleft, btest, eq_region);
2143 return;
2144 }
2145 if (right_ptr == ProfileAlwaysNull) {
2146 // Comparison with null. Assert the input is indeed null and we're done.
2147 acmp_always_null_input(right, tright, btest, eq_region);
2148 return;
2149 }
2150 if (left_type != NULL && !left_type->is_inlinetype()) {
2151 // Comparison with an object of known type
2152 acmp_known_non_inline_type_input(left, tleft, left_ptr, left_type, btest, eq_region);
2153 return;
2154 }
2155 if (right_type != NULL && !right_type->is_inlinetype()) {
2156 // Comparison with an object of known type
2157 acmp_known_non_inline_type_input(right, tright, right_ptr, right_type, btest, eq_region);
2158 return;
2159 }
2160 if (!left_inline_type) {
2161 // Comparison with an object known not to be an inline type
2162 acmp_unknown_non_inline_type_input(left, tleft, left_ptr, btest, eq_region);
2163 return;
2164 }
2165 if (!right_inline_type) {
2166 // Comparison with an object known not to be an inline type
2167 acmp_unknown_non_inline_type_input(right, tright, right_ptr, btest, eq_region);
2168 return;
2169 }
2170
2171 // Pointers are not equal, check if first operand is non-null
2172 Node* ne_region = new RegionNode(6);
2173 Node* null_ctl;
2174 Node* not_null_right = acmp_null_check(right, tright, right_ptr, null_ctl);
2175 ne_region->init_req(1, null_ctl);
2176
2177 // First operand is non-null, check if it is an inline type
2178 Node* is_value = inline_type_test(not_null_right);
2179 IfNode* is_value_iff = create_and_map_if(control(), is_value, PROB_FAIR, COUNT_UNKNOWN);
2180 Node* not_value = _gvn.transform(new IfFalseNode(is_value_iff));
2181 ne_region->init_req(2, not_value);
2182 set_control(_gvn.transform(new IfTrueNode(is_value_iff)));
2183
2184 // The first operand is an inline type, check if the second operand is non-null
2185 Node* not_null_left = acmp_null_check(left, tleft, left_ptr, null_ctl);
2186 ne_region->init_req(3, null_ctl);
2187
2188 // Check if both operands are of the same class.
2189 Node* kls_left = load_object_klass(not_null_left);
2190 Node* kls_right = load_object_klass(not_null_right);
2191 Node* kls_cmp = CmpP(kls_left, kls_right);
2192 Node* kls_bol = _gvn.transform(new BoolNode(kls_cmp, BoolTest::ne));
2193 IfNode* kls_iff = create_and_map_if(control(), kls_bol, PROB_FAIR, COUNT_UNKNOWN);
2194 Node* kls_ne = _gvn.transform(new IfTrueNode(kls_iff));
2195 set_control(_gvn.transform(new IfFalseNode(kls_iff)));
2196 ne_region->init_req(4, kls_ne);
2197
2198 if (stopped()) {
2199 record_for_igvn(ne_region);
2200 set_control(_gvn.transform(ne_region));
2201 if (btest == BoolTest::ne) {
2202 {
2203 PreserveJVMState pjvms(this);
2204 int target_bci = iter().get_dest();
2205 merge(target_bci);
2206 }
2207 record_for_igvn(eq_region);
2208 set_control(_gvn.transform(eq_region));
2209 }
2210 return;
2211 }
2212
2213 // Both operands are values types of the same class, we need to perform a
2214 // substitutability test. Delegate to ValueObjectMethods::isSubstitutable().
2215 Node* ne_io_phi = PhiNode::make(ne_region, i_o());
2216 Node* mem = reset_memory();
2217 Node* ne_mem_phi = PhiNode::make(ne_region, mem);
2218
2219 Node* eq_io_phi = NULL;
2220 Node* eq_mem_phi = NULL;
2221 if (eq_region != NULL) {
2222 eq_io_phi = PhiNode::make(eq_region, i_o());
2223 eq_mem_phi = PhiNode::make(eq_region, mem);
2224 }
2225
2226 set_all_memory(mem);
2227
2228 kill_dead_locals();
2229 ciMethod* subst_method = ciEnv::current()->ValueObjectMethods_klass()->find_method(ciSymbols::isSubstitutable_name(), ciSymbols::object_object_boolean_signature());
2230 CallStaticJavaNode *call = new CallStaticJavaNode(C, TypeFunc::make(subst_method), SharedRuntime::get_resolve_static_call_stub(), subst_method);
2231 call->set_override_symbolic_info(true);
2232 call->init_req(TypeFunc::Parms, not_null_left);
2233 call->init_req(TypeFunc::Parms+1, not_null_right);
2234 inc_sp(2);
2235 set_edges_for_java_call(call, false, false);
2236 Node* ret = set_results_for_java_call(call, false, true);
2237 dec_sp(2);
2238
2239 // Test the return value of ValueObjectMethods::isSubstitutable()
2240 Node* subst_cmp = _gvn.transform(new CmpINode(ret, intcon(1)));
2241 Node* ctl = C->top();
2242 if (btest == BoolTest::eq) {
2243 PreserveJVMState pjvms(this);
2244 do_if(btest, subst_cmp);
2245 if (!stopped()) {
2246 ctl = control();
2247 }
2248 } else {
2249 assert(btest == BoolTest::ne, "only eq or ne");
2250 PreserveJVMState pjvms(this);
2251 do_if(btest, subst_cmp, false, &ctl);
2252 if (!stopped()) {
2253 eq_region->init_req(2, control());
2254 eq_io_phi->init_req(2, i_o());
2255 eq_mem_phi->init_req(2, reset_memory());
2256 }
2257 }
2258 ne_region->init_req(5, ctl);
2259 ne_io_phi->init_req(5, i_o());
2260 ne_mem_phi->init_req(5, reset_memory());
2261
2262 record_for_igvn(ne_region);
2263 set_control(_gvn.transform(ne_region));
2264 set_i_o(_gvn.transform(ne_io_phi));
2265 set_all_memory(_gvn.transform(ne_mem_phi));
2266
2267 if (btest == BoolTest::ne) {
2268 {
2269 PreserveJVMState pjvms(this);
2270 int target_bci = iter().get_dest();
2271 merge(target_bci);
2272 }
2273
2274 record_for_igvn(eq_region);
2275 set_control(_gvn.transform(eq_region));
2276 set_i_o(_gvn.transform(eq_io_phi));
2277 set_all_memory(_gvn.transform(eq_mem_phi));
2278 }
2279 }
2280
2281 bool Parse::path_is_suitable_for_uncommon_trap(float prob) const {
2282 // Don't want to speculate on uncommon traps when running with -Xcomp
2283 if (!UseInterpreter) {
2284 return false;
2285 }
2286 return (seems_never_taken(prob) && seems_stable_comparison());
2287 }
2288
2289 void Parse::maybe_add_predicate_after_if(Block* path) {
2290 if (path->is_SEL_head() && path->preds_parsed() == 0) {
2291 // Add predicates at bci of if dominating the loop so traps can be
2292 // recorded on the if's profile data
2293 int bc_depth = repush_if_args();
2294 add_empty_predicates();
2295 dec_sp(bc_depth);
2296 path->set_has_predicates();
2297 }
2298 }
2299
2300
2301 //----------------------------adjust_map_after_if------------------------------
2302 // Adjust the JVM state to reflect the result of taking this path.
2303 // Basically, it means inspecting the CmpNode controlling this
2304 // branch, seeing how it constrains a tested value, and then
2305 // deciding if it's worth our while to encode this constraint
2306 // as graph nodes in the current abstract interpretation map.
2307 void Parse::adjust_map_after_if(BoolTest::mask btest, Node* c, float prob, Block* path) {
2308 if (!c->is_Cmp()) {
2309 maybe_add_predicate_after_if(path);
2310 return;
2311 }
2312
2313 if (stopped() || btest == BoolTest::illegal) {
2314 return; // nothing to do
2315 }
2316
2317 bool is_fallthrough = (path == successor_for_bci(iter().next_bci()));
2318
2319 if (path_is_suitable_for_uncommon_trap(prob)) {
2320 repush_if_args();
2321 Node* call = uncommon_trap(Deoptimization::Reason_unstable_if,
2322 Deoptimization::Action_reinterpret,
2323 NULL,
2324 (is_fallthrough ? "taken always" : "taken never"));
2325
2326 if (call != nullptr) {
2327 C->record_unstable_if_trap(new UnstableIfTrap(call->as_CallStaticJava(), path));
2328 }
2329 return;
2330 }
2331
2332 Node* val = c->in(1);
2333 Node* con = c->in(2);
2334 const Type* tcon = _gvn.type(con);
2335 const Type* tval = _gvn.type(val);
2336 bool have_con = tcon->singleton();
2337 if (tval->singleton()) {
2338 if (!have_con) {
2339 // Swap, so constant is in con.
2340 con = val;
2341 tcon = tval;
2342 val = c->in(2);
2343 tval = _gvn.type(val);
2344 btest = BoolTest(btest).commute();
2345 have_con = true;
2346 } else {
2347 // Do we have two constants? Then leave well enough alone.
2348 have_con = false;
2349 }
2350 }
2351 if (!have_con) { // remaining adjustments need a con
2352 maybe_add_predicate_after_if(path);
2353 return;
2354 }
2355
2356 sharpen_type_after_if(btest, con, tcon, val, tval);
2357 maybe_add_predicate_after_if(path);
2358 }
2359
2360
2361 static Node* extract_obj_from_klass_load(PhaseGVN* gvn, Node* n) {
2362 Node* ldk;
2363 if (n->is_DecodeNKlass()) {
2364 if (n->in(1)->Opcode() != Op_LoadNKlass) {
2365 return NULL;
2366 } else {
2367 ldk = n->in(1);
2368 }
2369 } else if (n->Opcode() != Op_LoadKlass) {
2370 return NULL;
2371 } else {
2372 ldk = n;
2373 }
2374 assert(ldk != NULL && ldk->is_Load(), "should have found a LoadKlass or LoadNKlass node");
2375
2376 Node* adr = ldk->in(MemNode::Address);
2377 intptr_t off = 0;
2378 Node* obj = AddPNode::Ideal_base_and_offset(adr, gvn, off);
2379 if (obj == NULL || off != oopDesc::klass_offset_in_bytes()) // loading oopDesc::_klass?
2380 return NULL;
2381 const TypePtr* tp = gvn->type(obj)->is_ptr();
2382 if (tp == NULL || !(tp->isa_instptr() || tp->isa_aryptr())) // is obj a Java object ptr?
2383 return NULL;
2384
2385 return obj;
2386 }
2387
2388 void Parse::sharpen_type_after_if(BoolTest::mask btest,
2389 Node* con, const Type* tcon,
2390 Node* val, const Type* tval) {
2391 // Look for opportunities to sharpen the type of a node
2392 // whose klass is compared with a constant klass.
2393 if (btest == BoolTest::eq && tcon->isa_klassptr()) {
2394 Node* obj = extract_obj_from_klass_load(&_gvn, val);
2395 const TypeOopPtr* con_type = tcon->isa_klassptr()->as_instance_type();
2396 if (obj != NULL && (con_type->isa_instptr() || con_type->isa_aryptr())) {
2397 // Found:
2398 // Bool(CmpP(LoadKlass(obj._klass), ConP(Foo.klass)), [eq])
2399 // or the narrowOop equivalent.
2400 const Type* obj_type = _gvn.type(obj);
2401 const TypeOopPtr* tboth = obj_type->join_speculative(con_type)->isa_oopptr();
2402 if (tboth != NULL && tboth->klass_is_exact() && tboth != obj_type &&
2403 tboth->higher_equal(obj_type)) {
2404 // obj has to be of the exact type Foo if the CmpP succeeds.
2405 int obj_in_map = map()->find_edge(obj);
2406 JVMState* jvms = this->jvms();
2407 if (obj_in_map >= 0 &&
2408 (jvms->is_loc(obj_in_map) || jvms->is_stk(obj_in_map))) {
2409 TypeNode* ccast = new CheckCastPPNode(control(), obj, tboth);
2410 const Type* tcc = ccast->as_Type()->type();
2411 assert(tcc != obj_type && tcc->higher_equal(obj_type), "must improve");
2412 // Delay transform() call to allow recovery of pre-cast value
2413 // at the control merge.
2414 _gvn.set_type_bottom(ccast);
2415 record_for_igvn(ccast);
2416 // Here's the payoff.
2417 replace_in_map(obj, ccast);
2418 }
2419 }
2420 }
2421 }
2422
2423 int val_in_map = map()->find_edge(val);
2424 if (val_in_map < 0) return; // replace_in_map would be useless
2425 {
2426 JVMState* jvms = this->jvms();
2427 if (!(jvms->is_loc(val_in_map) ||
2428 jvms->is_stk(val_in_map)))
2429 return; // again, it would be useless
2430 }
2431
2432 // Check for a comparison to a constant, and "know" that the compared
2433 // value is constrained on this path.
2434 assert(tcon->singleton(), "");
2435 ConstraintCastNode* ccast = NULL;
2436 Node* cast = NULL;
2437
2438 switch (btest) {
2439 case BoolTest::eq: // Constant test?
2440 {
2441 const Type* tboth = tcon->join_speculative(tval);
2442 if (tboth == tval) break; // Nothing to gain.
2443 if (tcon->isa_int()) {
2444 ccast = new CastIINode(val, tboth);
2445 } else if (tcon == TypePtr::NULL_PTR) {
2446 // Cast to null, but keep the pointer identity temporarily live.
2447 ccast = new CastPPNode(val, tboth);
2448 } else {
2449 const TypeF* tf = tcon->isa_float_constant();
2450 const TypeD* td = tcon->isa_double_constant();
2451 // Exclude tests vs float/double 0 as these could be
2452 // either +0 or -0. Just because you are equal to +0
2453 // doesn't mean you ARE +0!
2454 // Note, following code also replaces Long and Oop values.
2455 if ((!tf || tf->_f != 0.0) &&
2456 (!td || td->_d != 0.0))
2457 cast = con; // Replace non-constant val by con.
2458 }
2459 }
2460 break;
2461
2462 case BoolTest::ne:
2463 if (tcon == TypePtr::NULL_PTR) {
2464 cast = cast_not_null(val, false);
2465 }
2466 break;
2467
2468 default:
2469 // (At this point we could record int range types with CastII.)
2470 break;
2471 }
2472
2473 if (ccast != NULL) {
2474 const Type* tcc = ccast->as_Type()->type();
2475 assert(tcc != tval && tcc->higher_equal(tval), "must improve");
2476 // Delay transform() call to allow recovery of pre-cast value
2477 // at the control merge.
2478 ccast->set_req(0, control());
2479 _gvn.set_type_bottom(ccast);
2480 record_for_igvn(ccast);
2481 cast = ccast;
2482 }
2483
2484 if (cast != NULL) { // Here's the payoff.
2485 replace_in_map(val, cast);
2486 }
2487 }
2488
2489 /**
2490 * Use speculative type to optimize CmpP node: if comparison is
2491 * against the low level class, cast the object to the speculative
2492 * type if any. CmpP should then go away.
2493 *
2494 * @param c expected CmpP node
2495 * @return result of CmpP on object casted to speculative type
2496 *
2497 */
2498 Node* Parse::optimize_cmp_with_klass(Node* c) {
2499 // If this is transformed by the _gvn to a comparison with the low
2500 // level klass then we may be able to use speculation
2501 if (c->Opcode() == Op_CmpP &&
2502 (c->in(1)->Opcode() == Op_LoadKlass || c->in(1)->Opcode() == Op_DecodeNKlass) &&
2503 c->in(2)->is_Con()) {
2504 Node* load_klass = NULL;
2505 Node* decode = NULL;
2506 if (c->in(1)->Opcode() == Op_DecodeNKlass) {
2507 decode = c->in(1);
2508 load_klass = c->in(1)->in(1);
2509 } else {
2510 load_klass = c->in(1);
2511 }
2512 if (load_klass->in(2)->is_AddP()) {
2513 Node* addp = load_klass->in(2);
2514 Node* obj = addp->in(AddPNode::Address);
2515 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
2516 if (obj_type->speculative_type_not_null() != NULL) {
2517 ciKlass* k = obj_type->speculative_type();
2518 inc_sp(2);
2519 obj = maybe_cast_profiled_obj(obj, k);
2520 dec_sp(2);
2521 if (obj->is_InlineType()) {
2522 assert(obj->as_InlineType()->is_allocated(&_gvn), "must be allocated");
2523 obj = obj->as_InlineType()->get_oop();
2524 }
2525 // Make the CmpP use the casted obj
2526 addp = basic_plus_adr(obj, addp->in(AddPNode::Offset));
2527 load_klass = load_klass->clone();
2528 load_klass->set_req(2, addp);
2529 load_klass = _gvn.transform(load_klass);
2530 if (decode != NULL) {
2531 decode = decode->clone();
2532 decode->set_req(1, load_klass);
2533 load_klass = _gvn.transform(decode);
2534 }
2535 c = c->clone();
2536 c->set_req(1, load_klass);
2537 c = _gvn.transform(c);
2538 }
2539 }
2540 }
2541 return c;
2542 }
2543
2544 //------------------------------do_one_bytecode--------------------------------
2545 // Parse this bytecode, and alter the Parsers JVM->Node mapping
2546 void Parse::do_one_bytecode() {
2547 Node *a, *b, *c, *d; // Handy temps
2548 BoolTest::mask btest;
2549 int i;
2550
2551 assert(!has_exceptions(), "bytecode entry state must be clear of throws");
2552
2553 if (C->check_node_count(NodeLimitFudgeFactor * 5,
2554 "out of nodes parsing method")) {
2555 return;
2556 }
2557
2558 #ifdef ASSERT
2559 // for setting breakpoints
2560 if (TraceOptoParse) {
2561 tty->print(" @");
2562 dump_bci(bci());
2563 tty->print(" %s", Bytecodes::name(bc()));
2564 tty->cr();
2565 }
2566 #endif
2567
2568 switch (bc()) {
2569 case Bytecodes::_nop:
2570 // do nothing
2571 break;
2572 case Bytecodes::_lconst_0:
2573 push_pair(longcon(0));
2574 break;
2575
2576 case Bytecodes::_lconst_1:
2577 push_pair(longcon(1));
2578 break;
2579
2580 case Bytecodes::_fconst_0:
2581 push(zerocon(T_FLOAT));
2582 break;
2583
2584 case Bytecodes::_fconst_1:
2585 push(makecon(TypeF::ONE));
2586 break;
2587
2588 case Bytecodes::_fconst_2:
2589 push(makecon(TypeF::make(2.0f)));
2590 break;
2591
2592 case Bytecodes::_dconst_0:
2593 push_pair(zerocon(T_DOUBLE));
2594 break;
2595
2596 case Bytecodes::_dconst_1:
2597 push_pair(makecon(TypeD::ONE));
2598 break;
2599
2600 case Bytecodes::_iconst_m1:push(intcon(-1)); break;
2601 case Bytecodes::_iconst_0: push(intcon( 0)); break;
2602 case Bytecodes::_iconst_1: push(intcon( 1)); break;
2603 case Bytecodes::_iconst_2: push(intcon( 2)); break;
2604 case Bytecodes::_iconst_3: push(intcon( 3)); break;
2605 case Bytecodes::_iconst_4: push(intcon( 4)); break;
2606 case Bytecodes::_iconst_5: push(intcon( 5)); break;
2607 case Bytecodes::_bipush: push(intcon(iter().get_constant_u1())); break;
2608 case Bytecodes::_sipush: push(intcon(iter().get_constant_u2())); break;
2609 case Bytecodes::_aconst_null: push(null()); break;
2610
2611 case Bytecodes::_ldc:
2612 case Bytecodes::_ldc_w:
2613 case Bytecodes::_ldc2_w: {
2614 ciConstant constant = iter().get_constant();
2615 if (constant.is_loaded()) {
2616 const Type* con_type = Type::make_from_constant(constant);
2617 if (con_type != NULL) {
2618 push_node(con_type->basic_type(), makecon(con_type));
2619 }
2620 } else {
2621 // If the constant is unresolved or in error state, run this BC in the interpreter.
2622 if (iter().is_in_error()) {
2623 uncommon_trap(Deoptimization::make_trap_request(Deoptimization::Reason_unhandled,
2624 Deoptimization::Action_none),
2625 NULL, "constant in error state", true /* must_throw */);
2626
2627 } else {
2628 int index = iter().get_constant_pool_index();
2629 uncommon_trap(Deoptimization::make_trap_request(Deoptimization::Reason_unloaded,
2630 Deoptimization::Action_reinterpret,
2631 index),
2632 NULL, "unresolved constant", false /* must_throw */);
2633 }
2634 }
2635 break;
2636 }
2637
2638 case Bytecodes::_aload_0:
2639 push( local(0) );
2640 break;
2641 case Bytecodes::_aload_1:
2642 push( local(1) );
2643 break;
2644 case Bytecodes::_aload_2:
2645 push( local(2) );
2646 break;
2647 case Bytecodes::_aload_3:
2648 push( local(3) );
2649 break;
2650 case Bytecodes::_aload:
2651 push( local(iter().get_index()) );
2652 break;
2653
2654 case Bytecodes::_fload_0:
2655 case Bytecodes::_iload_0:
2656 push( local(0) );
2657 break;
2658 case Bytecodes::_fload_1:
2659 case Bytecodes::_iload_1:
2660 push( local(1) );
2661 break;
2662 case Bytecodes::_fload_2:
2663 case Bytecodes::_iload_2:
2664 push( local(2) );
2665 break;
2666 case Bytecodes::_fload_3:
2667 case Bytecodes::_iload_3:
2668 push( local(3) );
2669 break;
2670 case Bytecodes::_fload:
2671 case Bytecodes::_iload:
2672 push( local(iter().get_index()) );
2673 break;
2674 case Bytecodes::_lload_0:
2675 push_pair_local( 0 );
2676 break;
2677 case Bytecodes::_lload_1:
2678 push_pair_local( 1 );
2679 break;
2680 case Bytecodes::_lload_2:
2681 push_pair_local( 2 );
2682 break;
2683 case Bytecodes::_lload_3:
2684 push_pair_local( 3 );
2685 break;
2686 case Bytecodes::_lload:
2687 push_pair_local( iter().get_index() );
2688 break;
2689
2690 case Bytecodes::_dload_0:
2691 push_pair_local(0);
2692 break;
2693 case Bytecodes::_dload_1:
2694 push_pair_local(1);
2695 break;
2696 case Bytecodes::_dload_2:
2697 push_pair_local(2);
2698 break;
2699 case Bytecodes::_dload_3:
2700 push_pair_local(3);
2701 break;
2702 case Bytecodes::_dload:
2703 push_pair_local(iter().get_index());
2704 break;
2705 case Bytecodes::_fstore_0:
2706 case Bytecodes::_istore_0:
2707 case Bytecodes::_astore_0:
2708 set_local( 0, pop() );
2709 break;
2710 case Bytecodes::_fstore_1:
2711 case Bytecodes::_istore_1:
2712 case Bytecodes::_astore_1:
2713 set_local( 1, pop() );
2714 break;
2715 case Bytecodes::_fstore_2:
2716 case Bytecodes::_istore_2:
2717 case Bytecodes::_astore_2:
2718 set_local( 2, pop() );
2719 break;
2720 case Bytecodes::_fstore_3:
2721 case Bytecodes::_istore_3:
2722 case Bytecodes::_astore_3:
2723 set_local( 3, pop() );
2724 break;
2725 case Bytecodes::_fstore:
2726 case Bytecodes::_istore:
2727 case Bytecodes::_astore:
2728 set_local( iter().get_index(), pop() );
2729 break;
2730 // long stores
2731 case Bytecodes::_lstore_0:
2732 set_pair_local( 0, pop_pair() );
2733 break;
2734 case Bytecodes::_lstore_1:
2735 set_pair_local( 1, pop_pair() );
2736 break;
2737 case Bytecodes::_lstore_2:
2738 set_pair_local( 2, pop_pair() );
2739 break;
2740 case Bytecodes::_lstore_3:
2741 set_pair_local( 3, pop_pair() );
2742 break;
2743 case Bytecodes::_lstore:
2744 set_pair_local( iter().get_index(), pop_pair() );
2745 break;
2746
2747 // double stores
2748 case Bytecodes::_dstore_0:
2749 set_pair_local( 0, dprecision_rounding(pop_pair()) );
2750 break;
2751 case Bytecodes::_dstore_1:
2752 set_pair_local( 1, dprecision_rounding(pop_pair()) );
2753 break;
2754 case Bytecodes::_dstore_2:
2755 set_pair_local( 2, dprecision_rounding(pop_pair()) );
2756 break;
2757 case Bytecodes::_dstore_3:
2758 set_pair_local( 3, dprecision_rounding(pop_pair()) );
2759 break;
2760 case Bytecodes::_dstore:
2761 set_pair_local( iter().get_index(), dprecision_rounding(pop_pair()) );
2762 break;
2763
2764 case Bytecodes::_pop: dec_sp(1); break;
2765 case Bytecodes::_pop2: dec_sp(2); break;
2766 case Bytecodes::_swap:
2767 a = pop();
2768 b = pop();
2769 push(a);
2770 push(b);
2771 break;
2772 case Bytecodes::_dup:
2773 a = pop();
2774 push(a);
2775 push(a);
2776 break;
2777 case Bytecodes::_dup_x1:
2778 a = pop();
2779 b = pop();
2780 push( a );
2781 push( b );
2782 push( a );
2783 break;
2784 case Bytecodes::_dup_x2:
2785 a = pop();
2786 b = pop();
2787 c = pop();
2788 push( a );
2789 push( c );
2790 push( b );
2791 push( a );
2792 break;
2793 case Bytecodes::_dup2:
2794 a = pop();
2795 b = pop();
2796 push( b );
2797 push( a );
2798 push( b );
2799 push( a );
2800 break;
2801
2802 case Bytecodes::_dup2_x1:
2803 // before: .. c, b, a
2804 // after: .. b, a, c, b, a
2805 // not tested
2806 a = pop();
2807 b = pop();
2808 c = pop();
2809 push( b );
2810 push( a );
2811 push( c );
2812 push( b );
2813 push( a );
2814 break;
2815 case Bytecodes::_dup2_x2:
2816 // before: .. d, c, b, a
2817 // after: .. b, a, d, c, b, a
2818 // not tested
2819 a = pop();
2820 b = pop();
2821 c = pop();
2822 d = pop();
2823 push( b );
2824 push( a );
2825 push( d );
2826 push( c );
2827 push( b );
2828 push( a );
2829 break;
2830
2831 case Bytecodes::_arraylength: {
2832 // Must do null-check with value on expression stack
2833 Node *ary = null_check(peek(), T_ARRAY);
2834 // Compile-time detect of null-exception?
2835 if (stopped()) return;
2836 a = pop();
2837 push(load_array_length(a));
2838 break;
2839 }
2840
2841 case Bytecodes::_baload: array_load(T_BYTE); break;
2842 case Bytecodes::_caload: array_load(T_CHAR); break;
2843 case Bytecodes::_iaload: array_load(T_INT); break;
2844 case Bytecodes::_saload: array_load(T_SHORT); break;
2845 case Bytecodes::_faload: array_load(T_FLOAT); break;
2846 case Bytecodes::_aaload: array_load(T_OBJECT); break;
2847 case Bytecodes::_laload: array_load(T_LONG); break;
2848 case Bytecodes::_daload: array_load(T_DOUBLE); break;
2849 case Bytecodes::_bastore: array_store(T_BYTE); break;
2850 case Bytecodes::_castore: array_store(T_CHAR); break;
2851 case Bytecodes::_iastore: array_store(T_INT); break;
2852 case Bytecodes::_sastore: array_store(T_SHORT); break;
2853 case Bytecodes::_fastore: array_store(T_FLOAT); break;
2854 case Bytecodes::_aastore: array_store(T_OBJECT); break;
2855 case Bytecodes::_lastore: array_store(T_LONG); break;
2856 case Bytecodes::_dastore: array_store(T_DOUBLE); break;
2857
2858 case Bytecodes::_getfield:
2859 do_getfield();
2860 break;
2861
2862 case Bytecodes::_getstatic:
2863 do_getstatic();
2864 break;
2865
2866 case Bytecodes::_putfield:
2867 do_putfield();
2868 break;
2869
2870 case Bytecodes::_putstatic:
2871 do_putstatic();
2872 break;
2873
2874 case Bytecodes::_irem:
2875 // Must keep both values on the expression-stack during null-check
2876 zero_check_int(peek());
2877 // Compile-time detect of null-exception?
2878 if (stopped()) return;
2879 b = pop();
2880 a = pop();
2881 push(_gvn.transform(new ModINode(control(), a, b)));
2882 break;
2883 case Bytecodes::_idiv:
2884 // Must keep both values on the expression-stack during null-check
2885 zero_check_int(peek());
2886 // Compile-time detect of null-exception?
2887 if (stopped()) return;
2888 b = pop();
2889 a = pop();
2890 push( _gvn.transform( new DivINode(control(),a,b) ) );
2891 break;
2892 case Bytecodes::_imul:
2893 b = pop(); a = pop();
2894 push( _gvn.transform( new MulINode(a,b) ) );
2895 break;
2896 case Bytecodes::_iadd:
2897 b = pop(); a = pop();
2898 push( _gvn.transform( new AddINode(a,b) ) );
2899 break;
2900 case Bytecodes::_ineg:
2901 a = pop();
2902 push( _gvn.transform( new SubINode(_gvn.intcon(0),a)) );
2903 break;
2904 case Bytecodes::_isub:
2905 b = pop(); a = pop();
2906 push( _gvn.transform( new SubINode(a,b) ) );
2907 break;
2908 case Bytecodes::_iand:
2909 b = pop(); a = pop();
2910 push( _gvn.transform( new AndINode(a,b) ) );
2911 break;
2912 case Bytecodes::_ior:
2913 b = pop(); a = pop();
2914 push( _gvn.transform( new OrINode(a,b) ) );
2915 break;
2916 case Bytecodes::_ixor:
2917 b = pop(); a = pop();
2918 push( _gvn.transform( new XorINode(a,b) ) );
2919 break;
2920 case Bytecodes::_ishl:
2921 b = pop(); a = pop();
2922 push( _gvn.transform( new LShiftINode(a,b) ) );
2923 break;
2924 case Bytecodes::_ishr:
2925 b = pop(); a = pop();
2926 push( _gvn.transform( new RShiftINode(a,b) ) );
2927 break;
2928 case Bytecodes::_iushr:
2929 b = pop(); a = pop();
2930 push( _gvn.transform( new URShiftINode(a,b) ) );
2931 break;
2932
2933 case Bytecodes::_fneg:
2934 a = pop();
2935 b = _gvn.transform(new NegFNode (a));
2936 push(b);
2937 break;
2938
2939 case Bytecodes::_fsub:
2940 b = pop();
2941 a = pop();
2942 c = _gvn.transform( new SubFNode(a,b) );
2943 d = precision_rounding(c);
2944 push( d );
2945 break;
2946
2947 case Bytecodes::_fadd:
2948 b = pop();
2949 a = pop();
2950 c = _gvn.transform( new AddFNode(a,b) );
2951 d = precision_rounding(c);
2952 push( d );
2953 break;
2954
2955 case Bytecodes::_fmul:
2956 b = pop();
2957 a = pop();
2958 c = _gvn.transform( new MulFNode(a,b) );
2959 d = precision_rounding(c);
2960 push( d );
2961 break;
2962
2963 case Bytecodes::_fdiv:
2964 b = pop();
2965 a = pop();
2966 c = _gvn.transform( new DivFNode(0,a,b) );
2967 d = precision_rounding(c);
2968 push( d );
2969 break;
2970
2971 case Bytecodes::_frem:
2972 if (Matcher::has_match_rule(Op_ModF)) {
2973 // Generate a ModF node.
2974 b = pop();
2975 a = pop();
2976 c = _gvn.transform( new ModFNode(0,a,b) );
2977 d = precision_rounding(c);
2978 push( d );
2979 }
2980 else {
2981 // Generate a call.
2982 modf();
2983 }
2984 break;
2985
2986 case Bytecodes::_fcmpl:
2987 b = pop();
2988 a = pop();
2989 c = _gvn.transform( new CmpF3Node( a, b));
2990 push(c);
2991 break;
2992 case Bytecodes::_fcmpg:
2993 b = pop();
2994 a = pop();
2995
2996 // Same as fcmpl but need to flip the unordered case. Swap the inputs,
2997 // which negates the result sign except for unordered. Flip the unordered
2998 // as well by using CmpF3 which implements unordered-lesser instead of
2999 // unordered-greater semantics. Finally, commute the result bits. Result
3000 // is same as using a CmpF3Greater except we did it with CmpF3 alone.
3001 c = _gvn.transform( new CmpF3Node( b, a));
3002 c = _gvn.transform( new SubINode(_gvn.intcon(0),c) );
3003 push(c);
3004 break;
3005
3006 case Bytecodes::_f2i:
3007 a = pop();
3008 push(_gvn.transform(new ConvF2INode(a)));
3009 break;
3010
3011 case Bytecodes::_d2i:
3012 a = pop_pair();
3013 b = _gvn.transform(new ConvD2INode(a));
3014 push( b );
3015 break;
3016
3017 case Bytecodes::_f2d:
3018 a = pop();
3019 b = _gvn.transform( new ConvF2DNode(a));
3020 push_pair( b );
3021 break;
3022
3023 case Bytecodes::_d2f:
3024 a = pop_pair();
3025 b = _gvn.transform( new ConvD2FNode(a));
3026 // This breaks _227_mtrt (speed & correctness) and _222_mpegaudio (speed)
3027 //b = _gvn.transform(new RoundFloatNode(0, b) );
3028 push( b );
3029 break;
3030
3031 case Bytecodes::_l2f:
3032 if (Matcher::convL2FSupported()) {
3033 a = pop_pair();
3034 b = _gvn.transform( new ConvL2FNode(a));
3035 // For x86_32.ad, FILD doesn't restrict precision to 24 or 53 bits.
3036 // Rather than storing the result into an FP register then pushing
3037 // out to memory to round, the machine instruction that implements
3038 // ConvL2D is responsible for rounding.
3039 // c = precision_rounding(b);
3040 push(b);
3041 } else {
3042 l2f();
3043 }
3044 break;
3045
3046 case Bytecodes::_l2d:
3047 a = pop_pair();
3048 b = _gvn.transform( new ConvL2DNode(a));
3049 // For x86_32.ad, rounding is always necessary (see _l2f above).
3050 // c = dprecision_rounding(b);
3051 push_pair(b);
3052 break;
3053
3054 case Bytecodes::_f2l:
3055 a = pop();
3056 b = _gvn.transform( new ConvF2LNode(a));
3057 push_pair(b);
3058 break;
3059
3060 case Bytecodes::_d2l:
3061 a = pop_pair();
3062 b = _gvn.transform( new ConvD2LNode(a));
3063 push_pair(b);
3064 break;
3065
3066 case Bytecodes::_dsub:
3067 b = pop_pair();
3068 a = pop_pair();
3069 c = _gvn.transform( new SubDNode(a,b) );
3070 d = dprecision_rounding(c);
3071 push_pair( d );
3072 break;
3073
3074 case Bytecodes::_dadd:
3075 b = pop_pair();
3076 a = pop_pair();
3077 c = _gvn.transform( new AddDNode(a,b) );
3078 d = dprecision_rounding(c);
3079 push_pair( d );
3080 break;
3081
3082 case Bytecodes::_dmul:
3083 b = pop_pair();
3084 a = pop_pair();
3085 c = _gvn.transform( new MulDNode(a,b) );
3086 d = dprecision_rounding(c);
3087 push_pair( d );
3088 break;
3089
3090 case Bytecodes::_ddiv:
3091 b = pop_pair();
3092 a = pop_pair();
3093 c = _gvn.transform( new DivDNode(0,a,b) );
3094 d = dprecision_rounding(c);
3095 push_pair( d );
3096 break;
3097
3098 case Bytecodes::_dneg:
3099 a = pop_pair();
3100 b = _gvn.transform(new NegDNode (a));
3101 push_pair(b);
3102 break;
3103
3104 case Bytecodes::_drem:
3105 if (Matcher::has_match_rule(Op_ModD)) {
3106 // Generate a ModD node.
3107 b = pop_pair();
3108 a = pop_pair();
3109 // a % b
3110
3111 c = _gvn.transform( new ModDNode(0,a,b) );
3112 d = dprecision_rounding(c);
3113 push_pair( d );
3114 }
3115 else {
3116 // Generate a call.
3117 modd();
3118 }
3119 break;
3120
3121 case Bytecodes::_dcmpl:
3122 b = pop_pair();
3123 a = pop_pair();
3124 c = _gvn.transform( new CmpD3Node( a, b));
3125 push(c);
3126 break;
3127
3128 case Bytecodes::_dcmpg:
3129 b = pop_pair();
3130 a = pop_pair();
3131 // Same as dcmpl but need to flip the unordered case.
3132 // Commute the inputs, which negates the result sign except for unordered.
3133 // Flip the unordered as well by using CmpD3 which implements
3134 // unordered-lesser instead of unordered-greater semantics.
3135 // Finally, negate the result bits. Result is same as using a
3136 // CmpD3Greater except we did it with CmpD3 alone.
3137 c = _gvn.transform( new CmpD3Node( b, a));
3138 c = _gvn.transform( new SubINode(_gvn.intcon(0),c) );
3139 push(c);
3140 break;
3141
3142
3143 // Note for longs -> lo word is on TOS, hi word is on TOS - 1
3144 case Bytecodes::_land:
3145 b = pop_pair();
3146 a = pop_pair();
3147 c = _gvn.transform( new AndLNode(a,b) );
3148 push_pair(c);
3149 break;
3150 case Bytecodes::_lor:
3151 b = pop_pair();
3152 a = pop_pair();
3153 c = _gvn.transform( new OrLNode(a,b) );
3154 push_pair(c);
3155 break;
3156 case Bytecodes::_lxor:
3157 b = pop_pair();
3158 a = pop_pair();
3159 c = _gvn.transform( new XorLNode(a,b) );
3160 push_pair(c);
3161 break;
3162
3163 case Bytecodes::_lshl:
3164 b = pop(); // the shift count
3165 a = pop_pair(); // value to be shifted
3166 c = _gvn.transform( new LShiftLNode(a,b) );
3167 push_pair(c);
3168 break;
3169 case Bytecodes::_lshr:
3170 b = pop(); // the shift count
3171 a = pop_pair(); // value to be shifted
3172 c = _gvn.transform( new RShiftLNode(a,b) );
3173 push_pair(c);
3174 break;
3175 case Bytecodes::_lushr:
3176 b = pop(); // the shift count
3177 a = pop_pair(); // value to be shifted
3178 c = _gvn.transform( new URShiftLNode(a,b) );
3179 push_pair(c);
3180 break;
3181 case Bytecodes::_lmul:
3182 b = pop_pair();
3183 a = pop_pair();
3184 c = _gvn.transform( new MulLNode(a,b) );
3185 push_pair(c);
3186 break;
3187
3188 case Bytecodes::_lrem:
3189 // Must keep both values on the expression-stack during null-check
3190 assert(peek(0) == top(), "long word order");
3191 zero_check_long(peek(1));
3192 // Compile-time detect of null-exception?
3193 if (stopped()) return;
3194 b = pop_pair();
3195 a = pop_pair();
3196 c = _gvn.transform( new ModLNode(control(),a,b) );
3197 push_pair(c);
3198 break;
3199
3200 case Bytecodes::_ldiv:
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 DivLNode(control(),a,b) );
3209 push_pair(c);
3210 break;
3211
3212 case Bytecodes::_ladd:
3213 b = pop_pair();
3214 a = pop_pair();
3215 c = _gvn.transform( new AddLNode(a,b) );
3216 push_pair(c);
3217 break;
3218 case Bytecodes::_lsub:
3219 b = pop_pair();
3220 a = pop_pair();
3221 c = _gvn.transform( new SubLNode(a,b) );
3222 push_pair(c);
3223 break;
3224 case Bytecodes::_lcmp:
3225 // Safepoints are now inserted _before_ branches. The long-compare
3226 // bytecode painfully produces a 3-way value (-1,0,+1) which requires a
3227 // slew of control flow. These are usually followed by a CmpI vs zero and
3228 // a branch; this pattern then optimizes to the obvious long-compare and
3229 // branch. However, if the branch is backwards there's a Safepoint
3230 // inserted. The inserted Safepoint captures the JVM state at the
3231 // pre-branch point, i.e. it captures the 3-way value. Thus if a
3232 // long-compare is used to control a loop the debug info will force
3233 // computation of the 3-way value, even though the generated code uses a
3234 // long-compare and branch. We try to rectify the situation by inserting
3235 // a SafePoint here and have it dominate and kill the safepoint added at a
3236 // following backwards branch. At this point the JVM state merely holds 2
3237 // longs but not the 3-way value.
3238 switch (iter().next_bc()) {
3239 case Bytecodes::_ifgt:
3240 case Bytecodes::_iflt:
3241 case Bytecodes::_ifge:
3242 case Bytecodes::_ifle:
3243 case Bytecodes::_ifne:
3244 case Bytecodes::_ifeq:
3245 // If this is a backwards branch in the bytecodes, add Safepoint
3246 maybe_add_safepoint(iter().next_get_dest());
3247 default:
3248 break;
3249 }
3250 b = pop_pair();
3251 a = pop_pair();
3252 c = _gvn.transform( new CmpL3Node( a, b ));
3253 push(c);
3254 break;
3255
3256 case Bytecodes::_lneg:
3257 a = pop_pair();
3258 b = _gvn.transform( new SubLNode(longcon(0),a));
3259 push_pair(b);
3260 break;
3261 case Bytecodes::_l2i:
3262 a = pop_pair();
3263 push( _gvn.transform( new ConvL2INode(a)));
3264 break;
3265 case Bytecodes::_i2l:
3266 a = pop();
3267 b = _gvn.transform( new ConvI2LNode(a));
3268 push_pair(b);
3269 break;
3270 case Bytecodes::_i2b:
3271 // Sign extend
3272 a = pop();
3273 a = Compile::narrow_value(T_BYTE, a, NULL, &_gvn, true);
3274 push(a);
3275 break;
3276 case Bytecodes::_i2s:
3277 a = pop();
3278 a = Compile::narrow_value(T_SHORT, a, NULL, &_gvn, true);
3279 push(a);
3280 break;
3281 case Bytecodes::_i2c:
3282 a = pop();
3283 a = Compile::narrow_value(T_CHAR, a, NULL, &_gvn, true);
3284 push(a);
3285 break;
3286
3287 case Bytecodes::_i2f:
3288 a = pop();
3289 b = _gvn.transform( new ConvI2FNode(a) ) ;
3290 c = precision_rounding(b);
3291 push (b);
3292 break;
3293
3294 case Bytecodes::_i2d:
3295 a = pop();
3296 b = _gvn.transform( new ConvI2DNode(a));
3297 push_pair(b);
3298 break;
3299
3300 case Bytecodes::_iinc: // Increment local
3301 i = iter().get_index(); // Get local index
3302 set_local( i, _gvn.transform( new AddINode( _gvn.intcon(iter().get_iinc_con()), local(i) ) ) );
3303 break;
3304
3305 // Exit points of synchronized methods must have an unlock node
3306 case Bytecodes::_return:
3307 return_current(NULL);
3308 break;
3309
3310 case Bytecodes::_ireturn:
3311 case Bytecodes::_areturn:
3312 case Bytecodes::_freturn:
3313 return_current(pop());
3314 break;
3315 case Bytecodes::_lreturn:
3316 return_current(pop_pair());
3317 break;
3318 case Bytecodes::_dreturn:
3319 return_current(pop_pair());
3320 break;
3321
3322 case Bytecodes::_athrow:
3323 // null exception oop throws NULL pointer exception
3324 null_check(peek());
3325 if (stopped()) return;
3326 // Hook the thrown exception directly to subsequent handlers.
3327 if (BailoutToInterpreterForThrows) {
3328 // Keep method interpreted from now on.
3329 uncommon_trap(Deoptimization::Reason_unhandled,
3330 Deoptimization::Action_make_not_compilable);
3331 return;
3332 }
3333 if (env()->jvmti_can_post_on_exceptions()) {
3334 // check if we must post exception events, take uncommon trap if so (with must_throw = false)
3335 uncommon_trap_if_should_post_on_exceptions(Deoptimization::Reason_unhandled, false);
3336 }
3337 // Here if either can_post_on_exceptions or should_post_on_exceptions is false
3338 add_exception_state(make_exception_state(peek()));
3339 break;
3340
3341 case Bytecodes::_goto: // fall through
3342 case Bytecodes::_goto_w: {
3343 int target_bci = (bc() == Bytecodes::_goto) ? iter().get_dest() : iter().get_far_dest();
3344
3345 // If this is a backwards branch in the bytecodes, add Safepoint
3346 maybe_add_safepoint(target_bci);
3347
3348 // Merge the current control into the target basic block
3349 merge(target_bci);
3350
3351 // See if we can get some profile data and hand it off to the next block
3352 Block *target_block = block()->successor_for_bci(target_bci);
3353 if (target_block->pred_count() != 1) break;
3354 ciMethodData* methodData = method()->method_data();
3355 if (!methodData->is_mature()) break;
3356 ciProfileData* data = methodData->bci_to_data(bci());
3357 assert(data != NULL && data->is_JumpData(), "need JumpData for taken branch");
3358 int taken = ((ciJumpData*)data)->taken();
3359 taken = method()->scale_count(taken);
3360 target_block->set_count(taken);
3361 break;
3362 }
3363
3364 case Bytecodes::_ifnull: btest = BoolTest::eq; goto handle_if_null;
3365 case Bytecodes::_ifnonnull: btest = BoolTest::ne; goto handle_if_null;
3366 handle_if_null:
3367 // If this is a backwards branch in the bytecodes, add Safepoint
3368 maybe_add_safepoint(iter().get_dest());
3369 a = null();
3370 b = pop();
3371 if (b->is_InlineType()) {
3372 // Null checking a scalarized but nullable inline type. Check the IsInit
3373 // input instead of the oop input to avoid keeping buffer allocations alive
3374 c = _gvn.transform(new CmpINode(b->as_InlineType()->get_is_init(), zerocon(T_INT)));
3375 } else {
3376 if (!_gvn.type(b)->speculative_maybe_null() &&
3377 !too_many_traps(Deoptimization::Reason_speculate_null_check)) {
3378 inc_sp(1);
3379 Node* null_ctl = top();
3380 b = null_check_oop(b, &null_ctl, true, true, true);
3381 assert(null_ctl->is_top(), "no null control here");
3382 dec_sp(1);
3383 } else if (_gvn.type(b)->speculative_always_null() &&
3384 !too_many_traps(Deoptimization::Reason_speculate_null_assert)) {
3385 inc_sp(1);
3386 b = null_assert(b);
3387 dec_sp(1);
3388 }
3389 c = _gvn.transform( new CmpPNode(b, a) );
3390 }
3391 do_ifnull(btest, c);
3392 break;
3393
3394 case Bytecodes::_if_acmpeq: btest = BoolTest::eq; goto handle_if_acmp;
3395 case Bytecodes::_if_acmpne: btest = BoolTest::ne; goto handle_if_acmp;
3396 handle_if_acmp:
3397 // If this is a backwards branch in the bytecodes, add Safepoint
3398 maybe_add_safepoint(iter().get_dest());
3399 a = pop();
3400 b = pop();
3401 do_acmp(btest, b, a);
3402 break;
3403
3404 case Bytecodes::_ifeq: btest = BoolTest::eq; goto handle_ifxx;
3405 case Bytecodes::_ifne: btest = BoolTest::ne; goto handle_ifxx;
3406 case Bytecodes::_iflt: btest = BoolTest::lt; goto handle_ifxx;
3407 case Bytecodes::_ifle: btest = BoolTest::le; goto handle_ifxx;
3408 case Bytecodes::_ifgt: btest = BoolTest::gt; goto handle_ifxx;
3409 case Bytecodes::_ifge: btest = BoolTest::ge; goto handle_ifxx;
3410 handle_ifxx:
3411 // If this is a backwards branch in the bytecodes, add Safepoint
3412 maybe_add_safepoint(iter().get_dest());
3413 a = _gvn.intcon(0);
3414 b = pop();
3415 c = _gvn.transform( new CmpINode(b, a) );
3416 do_if(btest, c);
3417 break;
3418
3419 case Bytecodes::_if_icmpeq: btest = BoolTest::eq; goto handle_if_icmp;
3420 case Bytecodes::_if_icmpne: btest = BoolTest::ne; goto handle_if_icmp;
3421 case Bytecodes::_if_icmplt: btest = BoolTest::lt; goto handle_if_icmp;
3422 case Bytecodes::_if_icmple: btest = BoolTest::le; goto handle_if_icmp;
3423 case Bytecodes::_if_icmpgt: btest = BoolTest::gt; goto handle_if_icmp;
3424 case Bytecodes::_if_icmpge: btest = BoolTest::ge; goto handle_if_icmp;
3425 handle_if_icmp:
3426 // If this is a backwards branch in the bytecodes, add Safepoint
3427 maybe_add_safepoint(iter().get_dest());
3428 a = pop();
3429 b = pop();
3430 c = _gvn.transform( new CmpINode( b, a ) );
3431 do_if(btest, c);
3432 break;
3433
3434 case Bytecodes::_tableswitch:
3435 do_tableswitch();
3436 break;
3437
3438 case Bytecodes::_lookupswitch:
3439 do_lookupswitch();
3440 break;
3441
3442 case Bytecodes::_invokestatic:
3443 case Bytecodes::_invokedynamic:
3444 case Bytecodes::_invokespecial:
3445 case Bytecodes::_invokevirtual:
3446 case Bytecodes::_invokeinterface:
3447 do_call();
3448 break;
3449 case Bytecodes::_checkcast:
3450 do_checkcast();
3451 break;
3452 case Bytecodes::_instanceof:
3453 do_instanceof();
3454 break;
3455 case Bytecodes::_anewarray:
3456 do_newarray();
3457 break;
3458 case Bytecodes::_newarray:
3459 do_newarray((BasicType)iter().get_index());
3460 break;
3461 case Bytecodes::_multianewarray:
3462 do_multianewarray();
3463 break;
3464 case Bytecodes::_new:
3465 do_new();
3466 break;
3467 case Bytecodes::_aconst_init:
3468 do_aconst_init();
3469 break;
3470 case Bytecodes::_withfield:
3471 do_withfield();
3472 break;
3473
3474 case Bytecodes::_jsr:
3475 case Bytecodes::_jsr_w:
3476 do_jsr();
3477 break;
3478
3479 case Bytecodes::_ret:
3480 do_ret();
3481 break;
3482
3483
3484 case Bytecodes::_monitorenter:
3485 do_monitor_enter();
3486 break;
3487
3488 case Bytecodes::_monitorexit:
3489 do_monitor_exit();
3490 break;
3491
3492 case Bytecodes::_breakpoint:
3493 // Breakpoint set concurrently to compile
3494 // %%% use an uncommon trap?
3495 C->record_failure("breakpoint in method");
3496 return;
3497
3498 default:
3499 #ifndef PRODUCT
3500 map()->dump(99);
3501 #endif
3502 tty->print("\nUnhandled bytecode %s\n", Bytecodes::name(bc()) );
3503 ShouldNotReachHere();
3504 }
3505
3506 #ifndef PRODUCT
3507 if (C->should_print_igv(1)) {
3508 IdealGraphPrinter* printer = C->igv_printer();
3509 char buffer[256];
3510 jio_snprintf(buffer, sizeof(buffer), "Bytecode %d: %s", bci(), Bytecodes::name(bc()));
3511 bool old = printer->traverse_outs();
3512 printer->set_traverse_outs(true);
3513 printer->print_method(buffer, 4);
3514 printer->set_traverse_outs(old);
3515 }
3516 #endif
3517 }