Print this page
rev 2237 : [mq]: initial-intrinsification-changes
rev 2238 : [mq]: code-review-comments-vladimir
rev 2239 : [mq]: client_assertion_fauilure
rev 2240 : [mq]: code-review-comments-tom
rev 2242 : [mq]: reflection-unsafe-read-barrier
| Split |
Close |
| Expand all |
| Collapse all |
--- old/src/share/vm/opto/library_call.cpp
+++ new/src/share/vm/opto/library_call.cpp
1 1 /*
2 2 * Copyright (c) 1999, 2011, Oracle and/or its affiliates. All rights reserved.
3 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 4 *
5 5 * This code is free software; you can redistribute it and/or modify it
6 6 * under the terms of the GNU General Public License version 2 only, as
7 7 * published by the Free Software Foundation.
8 8 *
9 9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 12 * version 2 for more details (a copy is included in the LICENSE file that
13 13 * accompanied this code).
14 14 *
15 15 * You should have received a copy of the GNU General Public License version
16 16 * 2 along with this work; if not, write to the Free Software Foundation,
17 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 18 *
19 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 20 * or visit www.oracle.com if you need additional information or have any
21 21 * questions.
22 22 *
23 23 */
24 24
25 25 #include "precompiled.hpp"
26 26 #include "classfile/systemDictionary.hpp"
27 27 #include "classfile/vmSymbols.hpp"
28 28 #include "compiler/compileLog.hpp"
29 29 #include "oops/objArrayKlass.hpp"
30 30 #include "opto/addnode.hpp"
31 31 #include "opto/callGenerator.hpp"
32 32 #include "opto/cfgnode.hpp"
33 33 #include "opto/idealKit.hpp"
34 34 #include "opto/mulnode.hpp"
35 35 #include "opto/parse.hpp"
36 36 #include "opto/runtime.hpp"
37 37 #include "opto/subnode.hpp"
38 38 #include "prims/nativeLookup.hpp"
39 39 #include "runtime/sharedRuntime.hpp"
40 40
41 41 class LibraryIntrinsic : public InlineCallGenerator {
42 42 // Extend the set of intrinsics known to the runtime:
43 43 public:
44 44 private:
45 45 bool _is_virtual;
46 46 vmIntrinsics::ID _intrinsic_id;
47 47
48 48 public:
49 49 LibraryIntrinsic(ciMethod* m, bool is_virtual, vmIntrinsics::ID id)
50 50 : InlineCallGenerator(m),
51 51 _is_virtual(is_virtual),
52 52 _intrinsic_id(id)
53 53 {
54 54 }
55 55 virtual bool is_intrinsic() const { return true; }
56 56 virtual bool is_virtual() const { return _is_virtual; }
57 57 virtual JVMState* generate(JVMState* jvms);
58 58 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; }
59 59 };
60 60
61 61
62 62 // Local helper class for LibraryIntrinsic:
63 63 class LibraryCallKit : public GraphKit {
64 64 private:
65 65 LibraryIntrinsic* _intrinsic; // the library intrinsic being called
66 66
67 67 public:
68 68 LibraryCallKit(JVMState* caller, LibraryIntrinsic* intrinsic)
69 69 : GraphKit(caller),
70 70 _intrinsic(intrinsic)
71 71 {
72 72 }
73 73
74 74 ciMethod* caller() const { return jvms()->method(); }
75 75 int bci() const { return jvms()->bci(); }
76 76 LibraryIntrinsic* intrinsic() const { return _intrinsic; }
77 77 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); }
78 78 ciMethod* callee() const { return _intrinsic->method(); }
79 79 ciSignature* signature() const { return callee()->signature(); }
80 80 int arg_size() const { return callee()->arg_size(); }
81 81
82 82 bool try_to_inline();
83 83
84 84 // Helper functions to inline natives
85 85 void push_result(RegionNode* region, PhiNode* value);
86 86 Node* generate_guard(Node* test, RegionNode* region, float true_prob);
87 87 Node* generate_slow_guard(Node* test, RegionNode* region);
88 88 Node* generate_fair_guard(Node* test, RegionNode* region);
89 89 Node* generate_negative_guard(Node* index, RegionNode* region,
90 90 // resulting CastII of index:
91 91 Node* *pos_index = NULL);
92 92 Node* generate_nonpositive_guard(Node* index, bool never_negative,
93 93 // resulting CastII of index:
94 94 Node* *pos_index = NULL);
95 95 Node* generate_limit_guard(Node* offset, Node* subseq_length,
96 96 Node* array_length,
97 97 RegionNode* region);
98 98 Node* generate_current_thread(Node* &tls_output);
99 99 address basictype2arraycopy(BasicType t, Node *src_offset, Node *dest_offset,
100 100 bool disjoint_bases, const char* &name, bool dest_uninitialized);
101 101 Node* load_mirror_from_klass(Node* klass);
102 102 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null,
103 103 int nargs,
104 104 RegionNode* region, int null_path,
105 105 int offset);
106 106 Node* load_klass_from_mirror(Node* mirror, bool never_see_null, int nargs,
107 107 RegionNode* region, int null_path) {
108 108 int offset = java_lang_Class::klass_offset_in_bytes();
109 109 return load_klass_from_mirror_common(mirror, never_see_null, nargs,
110 110 region, null_path,
111 111 offset);
112 112 }
113 113 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null,
114 114 int nargs,
115 115 RegionNode* region, int null_path) {
116 116 int offset = java_lang_Class::array_klass_offset_in_bytes();
117 117 return load_klass_from_mirror_common(mirror, never_see_null, nargs,
118 118 region, null_path,
119 119 offset);
120 120 }
121 121 Node* generate_access_flags_guard(Node* kls,
122 122 int modifier_mask, int modifier_bits,
123 123 RegionNode* region);
124 124 Node* generate_interface_guard(Node* kls, RegionNode* region);
125 125 Node* generate_array_guard(Node* kls, RegionNode* region) {
126 126 return generate_array_guard_common(kls, region, false, false);
127 127 }
128 128 Node* generate_non_array_guard(Node* kls, RegionNode* region) {
129 129 return generate_array_guard_common(kls, region, false, true);
130 130 }
131 131 Node* generate_objArray_guard(Node* kls, RegionNode* region) {
132 132 return generate_array_guard_common(kls, region, true, false);
133 133 }
134 134 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) {
135 135 return generate_array_guard_common(kls, region, true, true);
136 136 }
137 137 Node* generate_array_guard_common(Node* kls, RegionNode* region,
138 138 bool obj_array, bool not_array);
139 139 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region);
140 140 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id,
141 141 bool is_virtual = false, bool is_static = false);
142 142 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) {
143 143 return generate_method_call(method_id, false, true);
144 144 }
145 145 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) {
146 146 return generate_method_call(method_id, true, false);
147 147 }
148 148
149 149 Node* make_string_method_node(int opcode, Node* str1, Node* cnt1, Node* str2, Node* cnt2);
150 150 bool inline_string_compareTo();
151 151 bool inline_string_indexOf();
152 152 Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i);
153 153 bool inline_string_equals();
154 154 Node* pop_math_arg();
155 155 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
156 156 bool inline_math_native(vmIntrinsics::ID id);
157 157 bool inline_trig(vmIntrinsics::ID id);
|
↓ open down ↓ |
157 lines elided |
↑ open up ↑ |
158 158 bool inline_trans(vmIntrinsics::ID id);
159 159 bool inline_abs(vmIntrinsics::ID id);
160 160 bool inline_sqrt(vmIntrinsics::ID id);
161 161 bool inline_pow(vmIntrinsics::ID id);
162 162 bool inline_exp(vmIntrinsics::ID id);
163 163 bool inline_min_max(vmIntrinsics::ID id);
164 164 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
165 165 // This returns Type::AnyPtr, RawPtr, or OopPtr.
166 166 int classify_unsafe_addr(Node* &base, Node* &offset);
167 167 Node* make_unsafe_address(Node* base, Node* offset);
168 + // Helper for inline_unsafe_access.
169 + // Generates the guards that check whether the result of
170 + // Unsafe.getObject should be recorded in an SATB log buffer.
171 + void insert_g1_pre_barrier(Node* base_oop, Node* offset, Node* pre_val);
168 172 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile);
169 173 bool inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static);
170 174 bool inline_unsafe_allocate();
171 175 bool inline_unsafe_copyMemory();
172 176 bool inline_native_currentThread();
173 177 bool inline_native_time_funcs(bool isNano);
174 178 bool inline_native_isInterrupted();
175 179 bool inline_native_Class_query(vmIntrinsics::ID id);
176 180 bool inline_native_subtype_check();
177 181
178 182 bool inline_native_newArray();
179 183 bool inline_native_getLength();
180 184 bool inline_array_copyOf(bool is_copyOfRange);
181 185 bool inline_array_equals();
182 186 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
183 187 bool inline_native_clone(bool is_virtual);
184 188 bool inline_native_Reflection_getCallerClass();
185 189 bool inline_native_AtomicLong_get();
186 190 bool inline_native_AtomicLong_attemptUpdate();
187 191 bool is_method_invoke_or_aux_frame(JVMState* jvms);
188 192 // Helper function for inlining native object hash method
189 193 bool inline_native_hashcode(bool is_virtual, bool is_static);
190 194 bool inline_native_getClass();
191 195
192 196 // Helper functions for inlining arraycopy
193 197 bool inline_arraycopy();
194 198 void generate_arraycopy(const TypePtr* adr_type,
195 199 BasicType basic_elem_type,
196 200 Node* src, Node* src_offset,
197 201 Node* dest, Node* dest_offset,
198 202 Node* copy_length,
199 203 bool disjoint_bases = false,
200 204 bool length_never_negative = false,
201 205 RegionNode* slow_region = NULL);
202 206 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
203 207 RegionNode* slow_region);
204 208 void generate_clear_array(const TypePtr* adr_type,
205 209 Node* dest,
206 210 BasicType basic_elem_type,
207 211 Node* slice_off,
208 212 Node* slice_len,
209 213 Node* slice_end);
210 214 bool generate_block_arraycopy(const TypePtr* adr_type,
211 215 BasicType basic_elem_type,
212 216 AllocateNode* alloc,
213 217 Node* src, Node* src_offset,
214 218 Node* dest, Node* dest_offset,
215 219 Node* dest_size, bool dest_uninitialized);
216 220 void generate_slow_arraycopy(const TypePtr* adr_type,
217 221 Node* src, Node* src_offset,
218 222 Node* dest, Node* dest_offset,
219 223 Node* copy_length, bool dest_uninitialized);
220 224 Node* generate_checkcast_arraycopy(const TypePtr* adr_type,
221 225 Node* dest_elem_klass,
222 226 Node* src, Node* src_offset,
223 227 Node* dest, Node* dest_offset,
224 228 Node* copy_length, bool dest_uninitialized);
225 229 Node* generate_generic_arraycopy(const TypePtr* adr_type,
226 230 Node* src, Node* src_offset,
227 231 Node* dest, Node* dest_offset,
228 232 Node* copy_length, bool dest_uninitialized);
229 233 void generate_unchecked_arraycopy(const TypePtr* adr_type,
230 234 BasicType basic_elem_type,
231 235 bool disjoint_bases,
|
↓ open down ↓ |
54 lines elided |
↑ open up ↑ |
232 236 Node* src, Node* src_offset,
233 237 Node* dest, Node* dest_offset,
234 238 Node* copy_length, bool dest_uninitialized);
235 239 bool inline_unsafe_CAS(BasicType type);
236 240 bool inline_unsafe_ordered_store(BasicType type);
237 241 bool inline_fp_conversions(vmIntrinsics::ID id);
238 242 bool inline_numberOfLeadingZeros(vmIntrinsics::ID id);
239 243 bool inline_numberOfTrailingZeros(vmIntrinsics::ID id);
240 244 bool inline_bitCount(vmIntrinsics::ID id);
241 245 bool inline_reverseBytes(vmIntrinsics::ID id);
246 +
247 + bool inline_reference_get();
242 248 };
243 249
244 250
245 251 //---------------------------make_vm_intrinsic----------------------------
246 252 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
247 253 vmIntrinsics::ID id = m->intrinsic_id();
248 254 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
249 255
250 256 if (DisableIntrinsic[0] != '\0'
251 257 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) {
252 258 // disabled by a user request on the command line:
253 259 // example: -XX:DisableIntrinsic=_hashCode,_getClass
254 260 return NULL;
255 261 }
256 262
257 263 if (!m->is_loaded()) {
258 264 // do not attempt to inline unloaded methods
259 265 return NULL;
260 266 }
261 267
262 268 // Only a few intrinsics implement a virtual dispatch.
263 269 // They are expensive calls which are also frequently overridden.
264 270 if (is_virtual) {
265 271 switch (id) {
266 272 case vmIntrinsics::_hashCode:
267 273 case vmIntrinsics::_clone:
268 274 // OK, Object.hashCode and Object.clone intrinsics come in both flavors
269 275 break;
270 276 default:
271 277 return NULL;
272 278 }
273 279 }
274 280
275 281 // -XX:-InlineNatives disables nearly all intrinsics:
276 282 if (!InlineNatives) {
277 283 switch (id) {
278 284 case vmIntrinsics::_indexOf:
279 285 case vmIntrinsics::_compareTo:
280 286 case vmIntrinsics::_equals:
281 287 case vmIntrinsics::_equalsC:
282 288 break; // InlineNatives does not control String.compareTo
283 289 default:
284 290 return NULL;
285 291 }
286 292 }
287 293
288 294 switch (id) {
289 295 case vmIntrinsics::_compareTo:
290 296 if (!SpecialStringCompareTo) return NULL;
291 297 break;
292 298 case vmIntrinsics::_indexOf:
293 299 if (!SpecialStringIndexOf) return NULL;
294 300 break;
295 301 case vmIntrinsics::_equals:
296 302 if (!SpecialStringEquals) return NULL;
297 303 break;
298 304 case vmIntrinsics::_equalsC:
299 305 if (!SpecialArraysEquals) return NULL;
300 306 break;
301 307 case vmIntrinsics::_arraycopy:
302 308 if (!InlineArrayCopy) return NULL;
303 309 break;
304 310 case vmIntrinsics::_copyMemory:
305 311 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL;
306 312 if (!InlineArrayCopy) return NULL;
307 313 break;
308 314 case vmIntrinsics::_hashCode:
309 315 if (!InlineObjectHash) return NULL;
310 316 break;
311 317 case vmIntrinsics::_clone:
312 318 case vmIntrinsics::_copyOf:
313 319 case vmIntrinsics::_copyOfRange:
314 320 if (!InlineObjectCopy) return NULL;
315 321 // These also use the arraycopy intrinsic mechanism:
316 322 if (!InlineArrayCopy) return NULL;
317 323 break;
318 324 case vmIntrinsics::_checkIndex:
319 325 // We do not intrinsify this. The optimizer does fine with it.
320 326 return NULL;
321 327
322 328 case vmIntrinsics::_get_AtomicLong:
323 329 case vmIntrinsics::_attemptUpdate:
324 330 if (!InlineAtomicLong) return NULL;
325 331 break;
326 332
327 333 case vmIntrinsics::_getCallerClass:
|
↓ open down ↓ |
76 lines elided |
↑ open up ↑ |
328 334 if (!UseNewReflection) return NULL;
329 335 if (!InlineReflectionGetCallerClass) return NULL;
330 336 if (!JDK_Version::is_gte_jdk14x_version()) return NULL;
331 337 break;
332 338
333 339 case vmIntrinsics::_bitCount_i:
334 340 case vmIntrinsics::_bitCount_l:
335 341 if (!UsePopCountInstruction) return NULL;
336 342 break;
337 343
344 + case vmIntrinsics::_Reference_get:
345 + // It is only when G1 is enabled that we absolutely
346 + // need to use the intrinsic version of Reference.get()
347 + // so that the value in the referent field, if necessary,
348 + // can be registered by the pre-barrier code.
349 + if (!UseG1GC) return NULL;
350 + break;
351 +
338 352 default:
339 353 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
340 354 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
341 355 break;
342 356 }
343 357
344 358 // -XX:-InlineClassNatives disables natives from the Class class.
345 359 // The flag applies to all reflective calls, notably Array.newArray
346 360 // (visible to Java programmers as Array.newInstance).
347 361 if (m->holder()->name() == ciSymbol::java_lang_Class() ||
348 362 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) {
349 363 if (!InlineClassNatives) return NULL;
350 364 }
351 365
352 366 // -XX:-InlineThreadNatives disables natives from the Thread class.
353 367 if (m->holder()->name() == ciSymbol::java_lang_Thread()) {
354 368 if (!InlineThreadNatives) return NULL;
355 369 }
356 370
357 371 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes.
358 372 if (m->holder()->name() == ciSymbol::java_lang_Math() ||
359 373 m->holder()->name() == ciSymbol::java_lang_Float() ||
360 374 m->holder()->name() == ciSymbol::java_lang_Double()) {
361 375 if (!InlineMathNatives) return NULL;
362 376 }
363 377
364 378 // -XX:-InlineUnsafeOps disables natives from the Unsafe class.
365 379 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) {
366 380 if (!InlineUnsafeOps) return NULL;
367 381 }
368 382
369 383 return new LibraryIntrinsic(m, is_virtual, (vmIntrinsics::ID) id);
370 384 }
371 385
372 386 //----------------------register_library_intrinsics-----------------------
373 387 // Initialize this file's data structures, for each Compile instance.
374 388 void Compile::register_library_intrinsics() {
375 389 // Nothing to do here.
376 390 }
377 391
378 392 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
|
↓ open down ↓ |
31 lines elided |
↑ open up ↑ |
379 393 LibraryCallKit kit(jvms, this);
380 394 Compile* C = kit.C;
381 395 int nodes = C->unique();
382 396 #ifndef PRODUCT
383 397 if ((PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) && Verbose) {
384 398 char buf[1000];
385 399 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
386 400 tty->print_cr("Intrinsic %s", str);
387 401 }
388 402 #endif
403 +
389 404 if (kit.try_to_inline()) {
390 405 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) {
391 - tty->print("Inlining intrinsic %s%s at bci:%d in",
392 - vmIntrinsics::name_at(intrinsic_id()),
393 - (is_virtual() ? " (virtual)" : ""), kit.bci());
394 - kit.caller()->print_short_name(tty);
395 - tty->print_cr(" (%d bytes)", kit.caller()->code_size());
406 + if (jvms->has_method()) {
407 + // Not a root compile.
408 + tty->print("Inlining intrinsic %s%s at bci:%d in",
409 + vmIntrinsics::name_at(intrinsic_id()),
410 + (is_virtual() ? " (virtual)" : ""), kit.bci());
411 + kit.caller()->print_short_name(tty);
412 + tty->print_cr(" (%d bytes)", kit.caller()->code_size());
413 + } else {
414 + // Root compile
415 + tty->print_cr("Generating intrinsic %s%s at bci:%d",
416 + vmIntrinsics::name_at(intrinsic_id()),
417 + (is_virtual() ? " (virtual)" : ""), kit.bci());
418 + }
396 419 }
397 420 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
398 421 if (C->log()) {
399 422 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
400 423 vmIntrinsics::name_at(intrinsic_id()),
401 424 (is_virtual() ? " virtual='1'" : ""),
402 425 C->unique() - nodes);
403 426 }
404 427 return kit.transfer_exceptions_into_jvms();
405 428 }
406 429
407 430 if (PrintIntrinsics) {
408 - tty->print("Did not inline intrinsic %s%s at bci:%d in",
431 + if (jvms->has_method()) {
432 + // Not a root compile.
433 + tty->print("Did not inline intrinsic %s%s at bci:%d in",
434 + vmIntrinsics::name_at(intrinsic_id()),
435 + (is_virtual() ? " (virtual)" : ""), kit.bci());
436 + kit.caller()->print_short_name(tty);
437 + tty->print_cr(" (%d bytes)", kit.caller()->code_size());
438 + } else {
439 + // Root compile
440 + tty->print("Did not generate intrinsic %s%s at bci:%d in",
409 441 vmIntrinsics::name_at(intrinsic_id()),
410 442 (is_virtual() ? " (virtual)" : ""), kit.bci());
411 - kit.caller()->print_short_name(tty);
412 - tty->print_cr(" (%d bytes)", kit.caller()->code_size());
443 + }
413 444 }
414 445 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
415 446 return NULL;
416 447 }
417 448
418 449 bool LibraryCallKit::try_to_inline() {
419 450 // Handle symbolic names for otherwise undistinguished boolean switches:
420 451 const bool is_store = true;
421 452 const bool is_native_ptr = true;
422 453 const bool is_static = true;
423 454
455 + if (!jvms()->has_method()) {
456 + // Root JVMState has a null method.
457 + assert(map()->memory()->Opcode() == Op_Parm, "");
458 + // Insert the memory aliasing node
459 + set_all_memory(reset_memory());
460 + }
461 + assert(merged_memory(), "");
462 +
424 463 switch (intrinsic_id()) {
425 464 case vmIntrinsics::_hashCode:
426 465 return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
427 466 case vmIntrinsics::_identityHashCode:
428 467 return inline_native_hashcode(/*!virtual*/ false, is_static);
429 468 case vmIntrinsics::_getClass:
430 469 return inline_native_getClass();
431 470
432 471 case vmIntrinsics::_dsin:
433 472 case vmIntrinsics::_dcos:
434 473 case vmIntrinsics::_dtan:
435 474 case vmIntrinsics::_dabs:
436 475 case vmIntrinsics::_datan2:
437 476 case vmIntrinsics::_dsqrt:
438 477 case vmIntrinsics::_dexp:
439 478 case vmIntrinsics::_dlog:
440 479 case vmIntrinsics::_dlog10:
441 480 case vmIntrinsics::_dpow:
442 481 return inline_math_native(intrinsic_id());
443 482
444 483 case vmIntrinsics::_min:
445 484 case vmIntrinsics::_max:
446 485 return inline_min_max(intrinsic_id());
447 486
448 487 case vmIntrinsics::_arraycopy:
449 488 return inline_arraycopy();
450 489
451 490 case vmIntrinsics::_compareTo:
452 491 return inline_string_compareTo();
453 492 case vmIntrinsics::_indexOf:
454 493 return inline_string_indexOf();
455 494 case vmIntrinsics::_equals:
456 495 return inline_string_equals();
457 496
458 497 case vmIntrinsics::_getObject:
459 498 return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, false);
460 499 case vmIntrinsics::_getBoolean:
461 500 return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, false);
462 501 case vmIntrinsics::_getByte:
463 502 return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, false);
464 503 case vmIntrinsics::_getShort:
465 504 return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, false);
466 505 case vmIntrinsics::_getChar:
467 506 return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, false);
468 507 case vmIntrinsics::_getInt:
469 508 return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, false);
470 509 case vmIntrinsics::_getLong:
471 510 return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, false);
472 511 case vmIntrinsics::_getFloat:
473 512 return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, false);
474 513 case vmIntrinsics::_getDouble:
475 514 return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, false);
476 515
477 516 case vmIntrinsics::_putObject:
478 517 return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, false);
479 518 case vmIntrinsics::_putBoolean:
480 519 return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, false);
481 520 case vmIntrinsics::_putByte:
482 521 return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, false);
483 522 case vmIntrinsics::_putShort:
484 523 return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, false);
485 524 case vmIntrinsics::_putChar:
486 525 return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, false);
487 526 case vmIntrinsics::_putInt:
488 527 return inline_unsafe_access(!is_native_ptr, is_store, T_INT, false);
489 528 case vmIntrinsics::_putLong:
490 529 return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, false);
491 530 case vmIntrinsics::_putFloat:
492 531 return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, false);
493 532 case vmIntrinsics::_putDouble:
494 533 return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, false);
495 534
496 535 case vmIntrinsics::_getByte_raw:
497 536 return inline_unsafe_access(is_native_ptr, !is_store, T_BYTE, false);
498 537 case vmIntrinsics::_getShort_raw:
499 538 return inline_unsafe_access(is_native_ptr, !is_store, T_SHORT, false);
500 539 case vmIntrinsics::_getChar_raw:
501 540 return inline_unsafe_access(is_native_ptr, !is_store, T_CHAR, false);
502 541 case vmIntrinsics::_getInt_raw:
503 542 return inline_unsafe_access(is_native_ptr, !is_store, T_INT, false);
504 543 case vmIntrinsics::_getLong_raw:
505 544 return inline_unsafe_access(is_native_ptr, !is_store, T_LONG, false);
506 545 case vmIntrinsics::_getFloat_raw:
507 546 return inline_unsafe_access(is_native_ptr, !is_store, T_FLOAT, false);
508 547 case vmIntrinsics::_getDouble_raw:
509 548 return inline_unsafe_access(is_native_ptr, !is_store, T_DOUBLE, false);
510 549 case vmIntrinsics::_getAddress_raw:
511 550 return inline_unsafe_access(is_native_ptr, !is_store, T_ADDRESS, false);
512 551
513 552 case vmIntrinsics::_putByte_raw:
514 553 return inline_unsafe_access(is_native_ptr, is_store, T_BYTE, false);
515 554 case vmIntrinsics::_putShort_raw:
516 555 return inline_unsafe_access(is_native_ptr, is_store, T_SHORT, false);
517 556 case vmIntrinsics::_putChar_raw:
518 557 return inline_unsafe_access(is_native_ptr, is_store, T_CHAR, false);
519 558 case vmIntrinsics::_putInt_raw:
520 559 return inline_unsafe_access(is_native_ptr, is_store, T_INT, false);
521 560 case vmIntrinsics::_putLong_raw:
522 561 return inline_unsafe_access(is_native_ptr, is_store, T_LONG, false);
523 562 case vmIntrinsics::_putFloat_raw:
524 563 return inline_unsafe_access(is_native_ptr, is_store, T_FLOAT, false);
525 564 case vmIntrinsics::_putDouble_raw:
526 565 return inline_unsafe_access(is_native_ptr, is_store, T_DOUBLE, false);
527 566 case vmIntrinsics::_putAddress_raw:
528 567 return inline_unsafe_access(is_native_ptr, is_store, T_ADDRESS, false);
529 568
530 569 case vmIntrinsics::_getObjectVolatile:
531 570 return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, true);
532 571 case vmIntrinsics::_getBooleanVolatile:
533 572 return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, true);
534 573 case vmIntrinsics::_getByteVolatile:
535 574 return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, true);
536 575 case vmIntrinsics::_getShortVolatile:
537 576 return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, true);
538 577 case vmIntrinsics::_getCharVolatile:
539 578 return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, true);
540 579 case vmIntrinsics::_getIntVolatile:
541 580 return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, true);
542 581 case vmIntrinsics::_getLongVolatile:
543 582 return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, true);
544 583 case vmIntrinsics::_getFloatVolatile:
545 584 return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, true);
546 585 case vmIntrinsics::_getDoubleVolatile:
547 586 return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, true);
548 587
549 588 case vmIntrinsics::_putObjectVolatile:
550 589 return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, true);
551 590 case vmIntrinsics::_putBooleanVolatile:
552 591 return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, true);
553 592 case vmIntrinsics::_putByteVolatile:
554 593 return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, true);
555 594 case vmIntrinsics::_putShortVolatile:
556 595 return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, true);
557 596 case vmIntrinsics::_putCharVolatile:
558 597 return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, true);
559 598 case vmIntrinsics::_putIntVolatile:
560 599 return inline_unsafe_access(!is_native_ptr, is_store, T_INT, true);
561 600 case vmIntrinsics::_putLongVolatile:
562 601 return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, true);
563 602 case vmIntrinsics::_putFloatVolatile:
564 603 return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, true);
565 604 case vmIntrinsics::_putDoubleVolatile:
566 605 return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, true);
567 606
568 607 case vmIntrinsics::_prefetchRead:
569 608 return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static);
570 609 case vmIntrinsics::_prefetchWrite:
571 610 return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static);
572 611 case vmIntrinsics::_prefetchReadStatic:
573 612 return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static);
574 613 case vmIntrinsics::_prefetchWriteStatic:
575 614 return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static);
576 615
577 616 case vmIntrinsics::_compareAndSwapObject:
578 617 return inline_unsafe_CAS(T_OBJECT);
579 618 case vmIntrinsics::_compareAndSwapInt:
580 619 return inline_unsafe_CAS(T_INT);
581 620 case vmIntrinsics::_compareAndSwapLong:
582 621 return inline_unsafe_CAS(T_LONG);
583 622
584 623 case vmIntrinsics::_putOrderedObject:
585 624 return inline_unsafe_ordered_store(T_OBJECT);
586 625 case vmIntrinsics::_putOrderedInt:
587 626 return inline_unsafe_ordered_store(T_INT);
588 627 case vmIntrinsics::_putOrderedLong:
589 628 return inline_unsafe_ordered_store(T_LONG);
590 629
591 630 case vmIntrinsics::_currentThread:
592 631 return inline_native_currentThread();
593 632 case vmIntrinsics::_isInterrupted:
594 633 return inline_native_isInterrupted();
595 634
596 635 case vmIntrinsics::_currentTimeMillis:
597 636 return inline_native_time_funcs(false);
598 637 case vmIntrinsics::_nanoTime:
599 638 return inline_native_time_funcs(true);
600 639 case vmIntrinsics::_allocateInstance:
601 640 return inline_unsafe_allocate();
602 641 case vmIntrinsics::_copyMemory:
603 642 return inline_unsafe_copyMemory();
604 643 case vmIntrinsics::_newArray:
605 644 return inline_native_newArray();
606 645 case vmIntrinsics::_getLength:
607 646 return inline_native_getLength();
608 647 case vmIntrinsics::_copyOf:
609 648 return inline_array_copyOf(false);
610 649 case vmIntrinsics::_copyOfRange:
611 650 return inline_array_copyOf(true);
612 651 case vmIntrinsics::_equalsC:
613 652 return inline_array_equals();
614 653 case vmIntrinsics::_clone:
615 654 return inline_native_clone(intrinsic()->is_virtual());
616 655
617 656 case vmIntrinsics::_isAssignableFrom:
618 657 return inline_native_subtype_check();
619 658
620 659 case vmIntrinsics::_isInstance:
621 660 case vmIntrinsics::_getModifiers:
622 661 case vmIntrinsics::_isInterface:
623 662 case vmIntrinsics::_isArray:
624 663 case vmIntrinsics::_isPrimitive:
625 664 case vmIntrinsics::_getSuperclass:
626 665 case vmIntrinsics::_getComponentType:
627 666 case vmIntrinsics::_getClassAccessFlags:
628 667 return inline_native_Class_query(intrinsic_id());
629 668
630 669 case vmIntrinsics::_floatToRawIntBits:
631 670 case vmIntrinsics::_floatToIntBits:
632 671 case vmIntrinsics::_intBitsToFloat:
633 672 case vmIntrinsics::_doubleToRawLongBits:
634 673 case vmIntrinsics::_doubleToLongBits:
635 674 case vmIntrinsics::_longBitsToDouble:
636 675 return inline_fp_conversions(intrinsic_id());
637 676
638 677 case vmIntrinsics::_numberOfLeadingZeros_i:
639 678 case vmIntrinsics::_numberOfLeadingZeros_l:
640 679 return inline_numberOfLeadingZeros(intrinsic_id());
641 680
642 681 case vmIntrinsics::_numberOfTrailingZeros_i:
643 682 case vmIntrinsics::_numberOfTrailingZeros_l:
644 683 return inline_numberOfTrailingZeros(intrinsic_id());
645 684
646 685 case vmIntrinsics::_bitCount_i:
647 686 case vmIntrinsics::_bitCount_l:
648 687 return inline_bitCount(intrinsic_id());
649 688
650 689 case vmIntrinsics::_reverseBytes_i:
651 690 case vmIntrinsics::_reverseBytes_l:
652 691 case vmIntrinsics::_reverseBytes_s:
653 692 case vmIntrinsics::_reverseBytes_c:
|
↓ open down ↓ |
220 lines elided |
↑ open up ↑ |
654 693 return inline_reverseBytes((vmIntrinsics::ID) intrinsic_id());
655 694
656 695 case vmIntrinsics::_get_AtomicLong:
657 696 return inline_native_AtomicLong_get();
658 697 case vmIntrinsics::_attemptUpdate:
659 698 return inline_native_AtomicLong_attemptUpdate();
660 699
661 700 case vmIntrinsics::_getCallerClass:
662 701 return inline_native_Reflection_getCallerClass();
663 702
703 + case vmIntrinsics::_Reference_get:
704 + return inline_reference_get();
705 +
664 706 default:
665 707 // If you get here, it may be that someone has added a new intrinsic
666 708 // to the list in vmSymbols.hpp without implementing it here.
667 709 #ifndef PRODUCT
668 710 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
669 711 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
670 712 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
671 713 }
672 714 #endif
673 715 return false;
674 716 }
675 717 }
676 718
677 719 //------------------------------push_result------------------------------
678 720 // Helper function for finishing intrinsics.
679 721 void LibraryCallKit::push_result(RegionNode* region, PhiNode* value) {
680 722 record_for_igvn(region);
681 723 set_control(_gvn.transform(region));
682 724 BasicType value_type = value->type()->basic_type();
683 725 push_node(value_type, _gvn.transform(value));
684 726 }
685 727
686 728 //------------------------------generate_guard---------------------------
687 729 // Helper function for generating guarded fast-slow graph structures.
688 730 // The given 'test', if true, guards a slow path. If the test fails
689 731 // then a fast path can be taken. (We generally hope it fails.)
690 732 // In all cases, GraphKit::control() is updated to the fast path.
691 733 // The returned value represents the control for the slow path.
692 734 // The return value is never 'top'; it is either a valid control
693 735 // or NULL if it is obvious that the slow path can never be taken.
694 736 // Also, if region and the slow control are not NULL, the slow edge
695 737 // is appended to the region.
696 738 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
697 739 if (stopped()) {
698 740 // Already short circuited.
699 741 return NULL;
700 742 }
701 743
702 744 // Build an if node and its projections.
703 745 // If test is true we take the slow path, which we assume is uncommon.
704 746 if (_gvn.type(test) == TypeInt::ZERO) {
705 747 // The slow branch is never taken. No need to build this guard.
706 748 return NULL;
707 749 }
708 750
709 751 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
710 752
711 753 Node* if_slow = _gvn.transform( new (C, 1) IfTrueNode(iff) );
712 754 if (if_slow == top()) {
713 755 // The slow branch is never taken. No need to build this guard.
714 756 return NULL;
715 757 }
716 758
717 759 if (region != NULL)
718 760 region->add_req(if_slow);
719 761
720 762 Node* if_fast = _gvn.transform( new (C, 1) IfFalseNode(iff) );
721 763 set_control(if_fast);
722 764
723 765 return if_slow;
724 766 }
725 767
726 768 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
727 769 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
728 770 }
729 771 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
730 772 return generate_guard(test, region, PROB_FAIR);
731 773 }
732 774
733 775 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
734 776 Node* *pos_index) {
735 777 if (stopped())
736 778 return NULL; // already stopped
737 779 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
738 780 return NULL; // index is already adequately typed
739 781 Node* cmp_lt = _gvn.transform( new (C, 3) CmpINode(index, intcon(0)) );
740 782 Node* bol_lt = _gvn.transform( new (C, 2) BoolNode(cmp_lt, BoolTest::lt) );
741 783 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
742 784 if (is_neg != NULL && pos_index != NULL) {
743 785 // Emulate effect of Parse::adjust_map_after_if.
744 786 Node* ccast = new (C, 2) CastIINode(index, TypeInt::POS);
745 787 ccast->set_req(0, control());
746 788 (*pos_index) = _gvn.transform(ccast);
747 789 }
748 790 return is_neg;
749 791 }
750 792
751 793 inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative,
752 794 Node* *pos_index) {
753 795 if (stopped())
754 796 return NULL; // already stopped
755 797 if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
756 798 return NULL; // index is already adequately typed
757 799 Node* cmp_le = _gvn.transform( new (C, 3) CmpINode(index, intcon(0)) );
758 800 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
759 801 Node* bol_le = _gvn.transform( new (C, 2) BoolNode(cmp_le, le_or_eq) );
760 802 Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN);
761 803 if (is_notp != NULL && pos_index != NULL) {
762 804 // Emulate effect of Parse::adjust_map_after_if.
763 805 Node* ccast = new (C, 2) CastIINode(index, TypeInt::POS1);
764 806 ccast->set_req(0, control());
765 807 (*pos_index) = _gvn.transform(ccast);
766 808 }
767 809 return is_notp;
768 810 }
769 811
770 812 // Make sure that 'position' is a valid limit index, in [0..length].
771 813 // There are two equivalent plans for checking this:
772 814 // A. (offset + copyLength) unsigned<= arrayLength
773 815 // B. offset <= (arrayLength - copyLength)
774 816 // We require that all of the values above, except for the sum and
775 817 // difference, are already known to be non-negative.
776 818 // Plan A is robust in the face of overflow, if offset and copyLength
777 819 // are both hugely positive.
778 820 //
779 821 // Plan B is less direct and intuitive, but it does not overflow at
780 822 // all, since the difference of two non-negatives is always
781 823 // representable. Whenever Java methods must perform the equivalent
782 824 // check they generally use Plan B instead of Plan A.
783 825 // For the moment we use Plan A.
784 826 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
785 827 Node* subseq_length,
786 828 Node* array_length,
787 829 RegionNode* region) {
788 830 if (stopped())
789 831 return NULL; // already stopped
790 832 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
791 833 if (zero_offset && _gvn.eqv_uncast(subseq_length, array_length))
792 834 return NULL; // common case of whole-array copy
793 835 Node* last = subseq_length;
794 836 if (!zero_offset) // last += offset
795 837 last = _gvn.transform( new (C, 3) AddINode(last, offset));
796 838 Node* cmp_lt = _gvn.transform( new (C, 3) CmpUNode(array_length, last) );
797 839 Node* bol_lt = _gvn.transform( new (C, 2) BoolNode(cmp_lt, BoolTest::lt) );
798 840 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
799 841 return is_over;
800 842 }
801 843
802 844
803 845 //--------------------------generate_current_thread--------------------
804 846 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
805 847 ciKlass* thread_klass = env()->Thread_klass();
806 848 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
807 849 Node* thread = _gvn.transform(new (C, 1) ThreadLocalNode());
808 850 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
809 851 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT);
810 852 tls_output = thread;
811 853 return threadObj;
812 854 }
813 855
814 856
815 857 //------------------------------make_string_method_node------------------------
816 858 // Helper method for String intrinsic finctions.
817 859 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* cnt1, Node* str2, Node* cnt2) {
818 860 const int value_offset = java_lang_String::value_offset_in_bytes();
819 861 const int count_offset = java_lang_String::count_offset_in_bytes();
820 862 const int offset_offset = java_lang_String::offset_offset_in_bytes();
821 863
822 864 Node* no_ctrl = NULL;
823 865
824 866 ciInstanceKlass* klass = env()->String_klass();
825 867 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
826 868
827 869 const TypeAryPtr* value_type =
828 870 TypeAryPtr::make(TypePtr::NotNull,
829 871 TypeAry::make(TypeInt::CHAR,TypeInt::POS),
830 872 ciTypeArrayKlass::make(T_CHAR), true, 0);
831 873
832 874 // Get start addr of string and substring
833 875 Node* str1_valuea = basic_plus_adr(str1, str1, value_offset);
834 876 Node* str1_value = make_load(no_ctrl, str1_valuea, value_type, T_OBJECT, string_type->add_offset(value_offset));
835 877 Node* str1_offseta = basic_plus_adr(str1, str1, offset_offset);
836 878 Node* str1_offset = make_load(no_ctrl, str1_offseta, TypeInt::INT, T_INT, string_type->add_offset(offset_offset));
837 879 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR);
838 880
839 881 // Pin loads from String::equals() argument since it could be NULL.
840 882 Node* str2_ctrl = (opcode == Op_StrEquals) ? control() : no_ctrl;
841 883 Node* str2_valuea = basic_plus_adr(str2, str2, value_offset);
842 884 Node* str2_value = make_load(str2_ctrl, str2_valuea, value_type, T_OBJECT, string_type->add_offset(value_offset));
843 885 Node* str2_offseta = basic_plus_adr(str2, str2, offset_offset);
844 886 Node* str2_offset = make_load(str2_ctrl, str2_offseta, TypeInt::INT, T_INT, string_type->add_offset(offset_offset));
845 887 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR);
846 888
847 889 Node* result = NULL;
848 890 switch (opcode) {
849 891 case Op_StrIndexOf:
850 892 result = new (C, 6) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
851 893 str1_start, cnt1, str2_start, cnt2);
852 894 break;
853 895 case Op_StrComp:
854 896 result = new (C, 6) StrCompNode(control(), memory(TypeAryPtr::CHARS),
855 897 str1_start, cnt1, str2_start, cnt2);
856 898 break;
857 899 case Op_StrEquals:
858 900 result = new (C, 5) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
859 901 str1_start, str2_start, cnt1);
860 902 break;
861 903 default:
862 904 ShouldNotReachHere();
863 905 return NULL;
864 906 }
865 907
866 908 // All these intrinsics have checks.
867 909 C->set_has_split_ifs(true); // Has chance for split-if optimization
868 910
869 911 return _gvn.transform(result);
870 912 }
871 913
872 914 //------------------------------inline_string_compareTo------------------------
873 915 bool LibraryCallKit::inline_string_compareTo() {
874 916
875 917 if (!Matcher::has_match_rule(Op_StrComp)) return false;
876 918
877 919 const int value_offset = java_lang_String::value_offset_in_bytes();
878 920 const int count_offset = java_lang_String::count_offset_in_bytes();
879 921 const int offset_offset = java_lang_String::offset_offset_in_bytes();
880 922
881 923 _sp += 2;
882 924 Node *argument = pop(); // pop non-receiver first: it was pushed second
883 925 Node *receiver = pop();
884 926
885 927 // Null check on self without removing any arguments. The argument
886 928 // null check technically happens in the wrong place, which can lead to
887 929 // invalid stack traces when string compare is inlined into a method
888 930 // which handles NullPointerExceptions.
889 931 _sp += 2;
890 932 receiver = do_null_check(receiver, T_OBJECT);
891 933 argument = do_null_check(argument, T_OBJECT);
892 934 _sp -= 2;
893 935 if (stopped()) {
894 936 return true;
895 937 }
896 938
897 939 ciInstanceKlass* klass = env()->String_klass();
898 940 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
899 941 Node* no_ctrl = NULL;
900 942
901 943 // Get counts for string and argument
902 944 Node* receiver_cnta = basic_plus_adr(receiver, receiver, count_offset);
903 945 Node* receiver_cnt = make_load(no_ctrl, receiver_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
904 946
905 947 Node* argument_cnta = basic_plus_adr(argument, argument, count_offset);
906 948 Node* argument_cnt = make_load(no_ctrl, argument_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
907 949
908 950 Node* compare = make_string_method_node(Op_StrComp, receiver, receiver_cnt, argument, argument_cnt);
909 951 push(compare);
910 952 return true;
911 953 }
912 954
913 955 //------------------------------inline_string_equals------------------------
914 956 bool LibraryCallKit::inline_string_equals() {
915 957
916 958 if (!Matcher::has_match_rule(Op_StrEquals)) return false;
917 959
918 960 const int value_offset = java_lang_String::value_offset_in_bytes();
919 961 const int count_offset = java_lang_String::count_offset_in_bytes();
920 962 const int offset_offset = java_lang_String::offset_offset_in_bytes();
921 963
922 964 int nargs = 2;
923 965 _sp += nargs;
924 966 Node* argument = pop(); // pop non-receiver first: it was pushed second
925 967 Node* receiver = pop();
926 968
927 969 // Null check on self without removing any arguments. The argument
928 970 // null check technically happens in the wrong place, which can lead to
929 971 // invalid stack traces when string compare is inlined into a method
930 972 // which handles NullPointerExceptions.
931 973 _sp += nargs;
932 974 receiver = do_null_check(receiver, T_OBJECT);
933 975 //should not do null check for argument for String.equals(), because spec
934 976 //allows to specify NULL as argument.
935 977 _sp -= nargs;
936 978
937 979 if (stopped()) {
938 980 return true;
939 981 }
940 982
941 983 // paths (plus control) merge
942 984 RegionNode* region = new (C, 5) RegionNode(5);
943 985 Node* phi = new (C, 5) PhiNode(region, TypeInt::BOOL);
944 986
945 987 // does source == target string?
946 988 Node* cmp = _gvn.transform(new (C, 3) CmpPNode(receiver, argument));
947 989 Node* bol = _gvn.transform(new (C, 2) BoolNode(cmp, BoolTest::eq));
948 990
949 991 Node* if_eq = generate_slow_guard(bol, NULL);
950 992 if (if_eq != NULL) {
951 993 // receiver == argument
952 994 phi->init_req(2, intcon(1));
953 995 region->init_req(2, if_eq);
954 996 }
955 997
956 998 // get String klass for instanceOf
957 999 ciInstanceKlass* klass = env()->String_klass();
958 1000
959 1001 if (!stopped()) {
960 1002 _sp += nargs; // gen_instanceof might do an uncommon trap
961 1003 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass)));
962 1004 _sp -= nargs;
963 1005 Node* cmp = _gvn.transform(new (C, 3) CmpINode(inst, intcon(1)));
964 1006 Node* bol = _gvn.transform(new (C, 2) BoolNode(cmp, BoolTest::ne));
965 1007
966 1008 Node* inst_false = generate_guard(bol, NULL, PROB_MIN);
967 1009 //instanceOf == true, fallthrough
968 1010
969 1011 if (inst_false != NULL) {
970 1012 phi->init_req(3, intcon(0));
971 1013 region->init_req(3, inst_false);
972 1014 }
973 1015 }
974 1016
975 1017 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
976 1018
977 1019 Node* no_ctrl = NULL;
978 1020 Node* receiver_cnt;
979 1021 Node* argument_cnt;
980 1022
981 1023 if (!stopped()) {
982 1024 // Properly cast the argument to String
983 1025 argument = _gvn.transform(new (C, 2) CheckCastPPNode(control(), argument, string_type));
984 1026
985 1027 // Get counts for string and argument
986 1028 Node* receiver_cnta = basic_plus_adr(receiver, receiver, count_offset);
987 1029 receiver_cnt = make_load(no_ctrl, receiver_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
988 1030
989 1031 // Pin load from argument string since it could be NULL.
990 1032 Node* argument_cnta = basic_plus_adr(argument, argument, count_offset);
991 1033 argument_cnt = make_load(control(), argument_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
992 1034
993 1035 // Check for receiver count != argument count
994 1036 Node* cmp = _gvn.transform( new(C, 3) CmpINode(receiver_cnt, argument_cnt) );
995 1037 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::ne) );
996 1038 Node* if_ne = generate_slow_guard(bol, NULL);
997 1039 if (if_ne != NULL) {
998 1040 phi->init_req(4, intcon(0));
999 1041 region->init_req(4, if_ne);
1000 1042 }
1001 1043 }
1002 1044
1003 1045 // Check for count == 0 is done by mach node StrEquals.
1004 1046
1005 1047 if (!stopped()) {
1006 1048 Node* equals = make_string_method_node(Op_StrEquals, receiver, receiver_cnt, argument, argument_cnt);
1007 1049 phi->init_req(1, equals);
1008 1050 region->init_req(1, control());
1009 1051 }
1010 1052
1011 1053 // post merge
1012 1054 set_control(_gvn.transform(region));
1013 1055 record_for_igvn(region);
1014 1056
1015 1057 push(_gvn.transform(phi));
1016 1058
1017 1059 return true;
1018 1060 }
1019 1061
1020 1062 //------------------------------inline_array_equals----------------------------
1021 1063 bool LibraryCallKit::inline_array_equals() {
1022 1064
1023 1065 if (!Matcher::has_match_rule(Op_AryEq)) return false;
1024 1066
1025 1067 _sp += 2;
1026 1068 Node *argument2 = pop();
1027 1069 Node *argument1 = pop();
1028 1070
1029 1071 Node* equals =
1030 1072 _gvn.transform(new (C, 4) AryEqNode(control(), memory(TypeAryPtr::CHARS),
1031 1073 argument1, argument2) );
1032 1074 push(equals);
1033 1075 return true;
1034 1076 }
1035 1077
1036 1078 // Java version of String.indexOf(constant string)
1037 1079 // class StringDecl {
1038 1080 // StringDecl(char[] ca) {
1039 1081 // offset = 0;
1040 1082 // count = ca.length;
1041 1083 // value = ca;
1042 1084 // }
1043 1085 // int offset;
1044 1086 // int count;
1045 1087 // char[] value;
1046 1088 // }
1047 1089 //
1048 1090 // static int string_indexOf_J(StringDecl string_object, char[] target_object,
1049 1091 // int targetOffset, int cache_i, int md2) {
1050 1092 // int cache = cache_i;
1051 1093 // int sourceOffset = string_object.offset;
1052 1094 // int sourceCount = string_object.count;
1053 1095 // int targetCount = target_object.length;
1054 1096 //
1055 1097 // int targetCountLess1 = targetCount - 1;
1056 1098 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1;
1057 1099 //
1058 1100 // char[] source = string_object.value;
1059 1101 // char[] target = target_object;
1060 1102 // int lastChar = target[targetCountLess1];
1061 1103 //
1062 1104 // outer_loop:
1063 1105 // for (int i = sourceOffset; i < sourceEnd; ) {
1064 1106 // int src = source[i + targetCountLess1];
1065 1107 // if (src == lastChar) {
1066 1108 // // With random strings and a 4-character alphabet,
1067 1109 // // reverse matching at this point sets up 0.8% fewer
1068 1110 // // frames, but (paradoxically) makes 0.3% more probes.
1069 1111 // // Since those probes are nearer the lastChar probe,
1070 1112 // // there is may be a net D$ win with reverse matching.
1071 1113 // // But, reversing loop inhibits unroll of inner loop
1072 1114 // // for unknown reason. So, does running outer loop from
1073 1115 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount)
1074 1116 // for (int j = 0; j < targetCountLess1; j++) {
1075 1117 // if (target[targetOffset + j] != source[i+j]) {
1076 1118 // if ((cache & (1 << source[i+j])) == 0) {
1077 1119 // if (md2 < j+1) {
1078 1120 // i += j+1;
1079 1121 // continue outer_loop;
1080 1122 // }
1081 1123 // }
1082 1124 // i += md2;
1083 1125 // continue outer_loop;
1084 1126 // }
1085 1127 // }
1086 1128 // return i - sourceOffset;
1087 1129 // }
1088 1130 // if ((cache & (1 << src)) == 0) {
1089 1131 // i += targetCountLess1;
1090 1132 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump.
1091 1133 // i++;
1092 1134 // }
1093 1135 // return -1;
1094 1136 // }
1095 1137
1096 1138 //------------------------------string_indexOf------------------------
1097 1139 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i,
1098 1140 jint cache_i, jint md2_i) {
1099 1141
1100 1142 Node* no_ctrl = NULL;
1101 1143 float likely = PROB_LIKELY(0.9);
1102 1144 float unlikely = PROB_UNLIKELY(0.9);
1103 1145
1104 1146 const int nargs = 2; // number of arguments to push back for uncommon trap in predicate
1105 1147
1106 1148 const int value_offset = java_lang_String::value_offset_in_bytes();
1107 1149 const int count_offset = java_lang_String::count_offset_in_bytes();
1108 1150 const int offset_offset = java_lang_String::offset_offset_in_bytes();
1109 1151
1110 1152 ciInstanceKlass* klass = env()->String_klass();
1111 1153 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
1112 1154 const TypeAryPtr* source_type = TypeAryPtr::make(TypePtr::NotNull, TypeAry::make(TypeInt::CHAR,TypeInt::POS), ciTypeArrayKlass::make(T_CHAR), true, 0);
1113 1155
1114 1156 Node* sourceOffseta = basic_plus_adr(string_object, string_object, offset_offset);
1115 1157 Node* sourceOffset = make_load(no_ctrl, sourceOffseta, TypeInt::INT, T_INT, string_type->add_offset(offset_offset));
1116 1158 Node* sourceCounta = basic_plus_adr(string_object, string_object, count_offset);
1117 1159 Node* sourceCount = make_load(no_ctrl, sourceCounta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
1118 1160 Node* sourcea = basic_plus_adr(string_object, string_object, value_offset);
1119 1161 Node* source = make_load(no_ctrl, sourcea, source_type, T_OBJECT, string_type->add_offset(value_offset));
1120 1162
1121 1163 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true)) );
1122 1164 jint target_length = target_array->length();
1123 1165 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
1124 1166 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
1125 1167
1126 1168 IdealKit kit(gvn(), control(), merged_memory(), false, true);
1127 1169 #define __ kit.
1128 1170 Node* zero = __ ConI(0);
1129 1171 Node* one = __ ConI(1);
1130 1172 Node* cache = __ ConI(cache_i);
1131 1173 Node* md2 = __ ConI(md2_i);
1132 1174 Node* lastChar = __ ConI(target_array->char_at(target_length - 1));
1133 1175 Node* targetCount = __ ConI(target_length);
1134 1176 Node* targetCountLess1 = __ ConI(target_length - 1);
1135 1177 Node* targetOffset = __ ConI(targetOffset_i);
1136 1178 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1);
1137 1179
1138 1180 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done();
1139 1181 Node* outer_loop = __ make_label(2 /* goto */);
1140 1182 Node* return_ = __ make_label(1);
1141 1183
1142 1184 __ set(rtn,__ ConI(-1));
1143 1185 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); {
1144 1186 Node* i2 = __ AddI(__ value(i), targetCountLess1);
1145 1187 // pin to prohibit loading of "next iteration" value which may SEGV (rare)
1146 1188 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS);
1147 1189 __ if_then(src, BoolTest::eq, lastChar, unlikely); {
1148 1190 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); {
1149 1191 Node* tpj = __ AddI(targetOffset, __ value(j));
1150 1192 Node* targ = load_array_element(no_ctrl, target, tpj, target_type);
1151 1193 Node* ipj = __ AddI(__ value(i), __ value(j));
1152 1194 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS);
1153 1195 __ if_then(targ, BoolTest::ne, src2); {
1154 1196 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); {
1155 1197 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); {
1156 1198 __ increment(i, __ AddI(__ value(j), one));
1157 1199 __ goto_(outer_loop);
1158 1200 } __ end_if(); __ dead(j);
1159 1201 }__ end_if(); __ dead(j);
1160 1202 __ increment(i, md2);
1161 1203 __ goto_(outer_loop);
1162 1204 }__ end_if();
1163 1205 __ increment(j, one);
1164 1206 }__ end_loop(); __ dead(j);
1165 1207 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i);
1166 1208 __ goto_(return_);
1167 1209 }__ end_if();
1168 1210 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); {
1169 1211 __ increment(i, targetCountLess1);
1170 1212 }__ end_if();
1171 1213 __ increment(i, one);
1172 1214 __ bind(outer_loop);
1173 1215 }__ end_loop(); __ dead(i);
1174 1216 __ bind(return_);
1175 1217
1176 1218 // Final sync IdealKit and GraphKit.
1177 1219 sync_kit(kit);
1178 1220 Node* result = __ value(rtn);
1179 1221 #undef __
1180 1222 C->set_has_loops(true);
1181 1223 return result;
1182 1224 }
1183 1225
1184 1226 //------------------------------inline_string_indexOf------------------------
1185 1227 bool LibraryCallKit::inline_string_indexOf() {
1186 1228
1187 1229 const int value_offset = java_lang_String::value_offset_in_bytes();
1188 1230 const int count_offset = java_lang_String::count_offset_in_bytes();
1189 1231 const int offset_offset = java_lang_String::offset_offset_in_bytes();
1190 1232
1191 1233 _sp += 2;
1192 1234 Node *argument = pop(); // pop non-receiver first: it was pushed second
1193 1235 Node *receiver = pop();
1194 1236
1195 1237 Node* result;
1196 1238 // Disable the use of pcmpestri until it can be guaranteed that
1197 1239 // the load doesn't cross into the uncommited space.
1198 1240 if (Matcher::has_match_rule(Op_StrIndexOf) &&
1199 1241 UseSSE42Intrinsics) {
1200 1242 // Generate SSE4.2 version of indexOf
1201 1243 // We currently only have match rules that use SSE4.2
1202 1244
1203 1245 // Null check on self without removing any arguments. The argument
1204 1246 // null check technically happens in the wrong place, which can lead to
1205 1247 // invalid stack traces when string compare is inlined into a method
1206 1248 // which handles NullPointerExceptions.
1207 1249 _sp += 2;
1208 1250 receiver = do_null_check(receiver, T_OBJECT);
1209 1251 argument = do_null_check(argument, T_OBJECT);
1210 1252 _sp -= 2;
1211 1253
1212 1254 if (stopped()) {
1213 1255 return true;
1214 1256 }
1215 1257
1216 1258 ciInstanceKlass* str_klass = env()->String_klass();
1217 1259 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(str_klass);
1218 1260
1219 1261 // Make the merge point
1220 1262 RegionNode* result_rgn = new (C, 4) RegionNode(4);
1221 1263 Node* result_phi = new (C, 4) PhiNode(result_rgn, TypeInt::INT);
1222 1264 Node* no_ctrl = NULL;
1223 1265
1224 1266 // Get counts for string and substr
1225 1267 Node* source_cnta = basic_plus_adr(receiver, receiver, count_offset);
1226 1268 Node* source_cnt = make_load(no_ctrl, source_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
1227 1269
1228 1270 Node* substr_cnta = basic_plus_adr(argument, argument, count_offset);
1229 1271 Node* substr_cnt = make_load(no_ctrl, substr_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
1230 1272
1231 1273 // Check for substr count > string count
1232 1274 Node* cmp = _gvn.transform( new(C, 3) CmpINode(substr_cnt, source_cnt) );
1233 1275 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::gt) );
1234 1276 Node* if_gt = generate_slow_guard(bol, NULL);
1235 1277 if (if_gt != NULL) {
1236 1278 result_phi->init_req(2, intcon(-1));
1237 1279 result_rgn->init_req(2, if_gt);
1238 1280 }
1239 1281
1240 1282 if (!stopped()) {
1241 1283 // Check for substr count == 0
1242 1284 cmp = _gvn.transform( new(C, 3) CmpINode(substr_cnt, intcon(0)) );
1243 1285 bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) );
1244 1286 Node* if_zero = generate_slow_guard(bol, NULL);
1245 1287 if (if_zero != NULL) {
1246 1288 result_phi->init_req(3, intcon(0));
1247 1289 result_rgn->init_req(3, if_zero);
1248 1290 }
1249 1291 }
1250 1292
1251 1293 if (!stopped()) {
1252 1294 result = make_string_method_node(Op_StrIndexOf, receiver, source_cnt, argument, substr_cnt);
1253 1295 result_phi->init_req(1, result);
1254 1296 result_rgn->init_req(1, control());
1255 1297 }
1256 1298 set_control(_gvn.transform(result_rgn));
1257 1299 record_for_igvn(result_rgn);
1258 1300 result = _gvn.transform(result_phi);
1259 1301
1260 1302 } else { // Use LibraryCallKit::string_indexOf
1261 1303 // don't intrinsify if argument isn't a constant string.
1262 1304 if (!argument->is_Con()) {
1263 1305 return false;
1264 1306 }
1265 1307 const TypeOopPtr* str_type = _gvn.type(argument)->isa_oopptr();
1266 1308 if (str_type == NULL) {
1267 1309 return false;
1268 1310 }
1269 1311 ciInstanceKlass* klass = env()->String_klass();
1270 1312 ciObject* str_const = str_type->const_oop();
1271 1313 if (str_const == NULL || str_const->klass() != klass) {
1272 1314 return false;
1273 1315 }
1274 1316 ciInstance* str = str_const->as_instance();
1275 1317 assert(str != NULL, "must be instance");
1276 1318
1277 1319 ciObject* v = str->field_value_by_offset(value_offset).as_object();
1278 1320 int o = str->field_value_by_offset(offset_offset).as_int();
1279 1321 int c = str->field_value_by_offset(count_offset).as_int();
1280 1322 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array
1281 1323
1282 1324 // constant strings have no offset and count == length which
1283 1325 // simplifies the resulting code somewhat so lets optimize for that.
1284 1326 if (o != 0 || c != pat->length()) {
1285 1327 return false;
1286 1328 }
1287 1329
1288 1330 // Null check on self without removing any arguments. The argument
1289 1331 // null check technically happens in the wrong place, which can lead to
1290 1332 // invalid stack traces when string compare is inlined into a method
1291 1333 // which handles NullPointerExceptions.
1292 1334 _sp += 2;
1293 1335 receiver = do_null_check(receiver, T_OBJECT);
1294 1336 // No null check on the argument is needed since it's a constant String oop.
1295 1337 _sp -= 2;
1296 1338 if (stopped()) {
1297 1339 return true;
1298 1340 }
1299 1341
1300 1342 // The null string as a pattern always returns 0 (match at beginning of string)
1301 1343 if (c == 0) {
1302 1344 push(intcon(0));
1303 1345 return true;
1304 1346 }
1305 1347
1306 1348 // Generate default indexOf
1307 1349 jchar lastChar = pat->char_at(o + (c - 1));
1308 1350 int cache = 0;
1309 1351 int i;
1310 1352 for (i = 0; i < c - 1; i++) {
1311 1353 assert(i < pat->length(), "out of range");
1312 1354 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1)));
1313 1355 }
1314 1356
1315 1357 int md2 = c;
1316 1358 for (i = 0; i < c - 1; i++) {
1317 1359 assert(i < pat->length(), "out of range");
1318 1360 if (pat->char_at(o + i) == lastChar) {
1319 1361 md2 = (c - 1) - i;
1320 1362 }
1321 1363 }
1322 1364
1323 1365 result = string_indexOf(receiver, pat, o, cache, md2);
1324 1366 }
1325 1367
1326 1368 push(result);
1327 1369 return true;
1328 1370 }
1329 1371
1330 1372 //--------------------------pop_math_arg--------------------------------
1331 1373 // Pop a double argument to a math function from the stack
1332 1374 // rounding it if necessary.
1333 1375 Node * LibraryCallKit::pop_math_arg() {
1334 1376 Node *arg = pop_pair();
1335 1377 if( Matcher::strict_fp_requires_explicit_rounding && UseSSE<=1 )
1336 1378 arg = _gvn.transform( new (C, 2) RoundDoubleNode(0, arg) );
1337 1379 return arg;
1338 1380 }
1339 1381
1340 1382 //------------------------------inline_trig----------------------------------
1341 1383 // Inline sin/cos/tan instructions, if possible. If rounding is required, do
1342 1384 // argument reduction which will turn into a fast/slow diamond.
1343 1385 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
1344 1386 _sp += arg_size(); // restore stack pointer
1345 1387 Node* arg = pop_math_arg();
1346 1388 Node* trig = NULL;
1347 1389
1348 1390 switch (id) {
1349 1391 case vmIntrinsics::_dsin:
1350 1392 trig = _gvn.transform((Node*)new (C, 2) SinDNode(arg));
1351 1393 break;
1352 1394 case vmIntrinsics::_dcos:
1353 1395 trig = _gvn.transform((Node*)new (C, 2) CosDNode(arg));
1354 1396 break;
1355 1397 case vmIntrinsics::_dtan:
1356 1398 trig = _gvn.transform((Node*)new (C, 2) TanDNode(arg));
1357 1399 break;
1358 1400 default:
1359 1401 assert(false, "bad intrinsic was passed in");
1360 1402 return false;
1361 1403 }
1362 1404
1363 1405 // Rounding required? Check for argument reduction!
1364 1406 if( Matcher::strict_fp_requires_explicit_rounding ) {
1365 1407
1366 1408 static const double pi_4 = 0.7853981633974483;
1367 1409 static const double neg_pi_4 = -0.7853981633974483;
1368 1410 // pi/2 in 80-bit extended precision
1369 1411 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
1370 1412 // -pi/2 in 80-bit extended precision
1371 1413 // static const unsigned char neg_pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0xbf,0x00,0x00,0x00,0x00,0x00,0x00};
1372 1414 // Cutoff value for using this argument reduction technique
1373 1415 //static const double pi_2_minus_epsilon = 1.564660403643354;
1374 1416 //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
1375 1417
1376 1418 // Pseudocode for sin:
1377 1419 // if (x <= Math.PI / 4.0) {
1378 1420 // if (x >= -Math.PI / 4.0) return fsin(x);
1379 1421 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
1380 1422 // } else {
1381 1423 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0);
1382 1424 // }
1383 1425 // return StrictMath.sin(x);
1384 1426
1385 1427 // Pseudocode for cos:
1386 1428 // if (x <= Math.PI / 4.0) {
1387 1429 // if (x >= -Math.PI / 4.0) return fcos(x);
1388 1430 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0);
1389 1431 // } else {
1390 1432 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
1391 1433 // }
1392 1434 // return StrictMath.cos(x);
1393 1435
1394 1436 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
1395 1437 // requires a special machine instruction to load it. Instead we'll try
1396 1438 // the 'easy' case. If we really need the extra range +/- PI/2 we'll
1397 1439 // probably do the math inside the SIN encoding.
1398 1440
1399 1441 // Make the merge point
1400 1442 RegionNode *r = new (C, 3) RegionNode(3);
1401 1443 Node *phi = new (C, 3) PhiNode(r,Type::DOUBLE);
1402 1444
1403 1445 // Flatten arg so we need only 1 test
1404 1446 Node *abs = _gvn.transform(new (C, 2) AbsDNode(arg));
1405 1447 // Node for PI/4 constant
1406 1448 Node *pi4 = makecon(TypeD::make(pi_4));
1407 1449 // Check PI/4 : abs(arg)
1408 1450 Node *cmp = _gvn.transform(new (C, 3) CmpDNode(pi4,abs));
1409 1451 // Check: If PI/4 < abs(arg) then go slow
1410 1452 Node *bol = _gvn.transform( new (C, 2) BoolNode( cmp, BoolTest::lt ) );
1411 1453 // Branch either way
1412 1454 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1413 1455 set_control(opt_iff(r,iff));
1414 1456
1415 1457 // Set fast path result
1416 1458 phi->init_req(2,trig);
1417 1459
1418 1460 // Slow path - non-blocking leaf call
1419 1461 Node* call = NULL;
1420 1462 switch (id) {
1421 1463 case vmIntrinsics::_dsin:
1422 1464 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1423 1465 CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
1424 1466 "Sin", NULL, arg, top());
1425 1467 break;
1426 1468 case vmIntrinsics::_dcos:
1427 1469 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1428 1470 CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
1429 1471 "Cos", NULL, arg, top());
1430 1472 break;
1431 1473 case vmIntrinsics::_dtan:
1432 1474 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1433 1475 CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
1434 1476 "Tan", NULL, arg, top());
1435 1477 break;
1436 1478 }
1437 1479 assert(control()->in(0) == call, "");
1438 1480 Node* slow_result = _gvn.transform(new (C, 1) ProjNode(call,TypeFunc::Parms));
1439 1481 r->init_req(1,control());
1440 1482 phi->init_req(1,slow_result);
1441 1483
1442 1484 // Post-merge
1443 1485 set_control(_gvn.transform(r));
1444 1486 record_for_igvn(r);
1445 1487 trig = _gvn.transform(phi);
1446 1488
1447 1489 C->set_has_split_ifs(true); // Has chance for split-if optimization
1448 1490 }
1449 1491 // Push result back on JVM stack
1450 1492 push_pair(trig);
1451 1493 return true;
1452 1494 }
1453 1495
1454 1496 //------------------------------inline_sqrt-------------------------------------
1455 1497 // Inline square root instruction, if possible.
1456 1498 bool LibraryCallKit::inline_sqrt(vmIntrinsics::ID id) {
1457 1499 assert(id == vmIntrinsics::_dsqrt, "Not square root");
1458 1500 _sp += arg_size(); // restore stack pointer
1459 1501 push_pair(_gvn.transform(new (C, 2) SqrtDNode(0, pop_math_arg())));
1460 1502 return true;
1461 1503 }
1462 1504
1463 1505 //------------------------------inline_abs-------------------------------------
1464 1506 // Inline absolute value instruction, if possible.
1465 1507 bool LibraryCallKit::inline_abs(vmIntrinsics::ID id) {
1466 1508 assert(id == vmIntrinsics::_dabs, "Not absolute value");
1467 1509 _sp += arg_size(); // restore stack pointer
1468 1510 push_pair(_gvn.transform(new (C, 2) AbsDNode(pop_math_arg())));
1469 1511 return true;
1470 1512 }
1471 1513
1472 1514 //------------------------------inline_exp-------------------------------------
1473 1515 // Inline exp instructions, if possible. The Intel hardware only misses
1474 1516 // really odd corner cases (+/- Infinity). Just uncommon-trap them.
1475 1517 bool LibraryCallKit::inline_exp(vmIntrinsics::ID id) {
1476 1518 assert(id == vmIntrinsics::_dexp, "Not exp");
1477 1519
1478 1520 // If this inlining ever returned NaN in the past, we do not intrinsify it
1479 1521 // every again. NaN results requires StrictMath.exp handling.
1480 1522 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
1481 1523
1482 1524 // Do not intrinsify on older platforms which lack cmove.
1483 1525 if (ConditionalMoveLimit == 0) return false;
1484 1526
1485 1527 _sp += arg_size(); // restore stack pointer
1486 1528 Node *x = pop_math_arg();
1487 1529 Node *result = _gvn.transform(new (C, 2) ExpDNode(0,x));
1488 1530
1489 1531 //-------------------
1490 1532 //result=(result.isNaN())? StrictMath::exp():result;
1491 1533 // Check: If isNaN() by checking result!=result? then go to Strict Math
1492 1534 Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result));
1493 1535 // Build the boolean node
1494 1536 Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) );
1495 1537
1496 1538 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1497 1539 // End the current control-flow path
1498 1540 push_pair(x);
1499 1541 // Math.exp intrinsic returned a NaN, which requires StrictMath.exp
1500 1542 // to handle. Recompile without intrinsifying Math.exp
1501 1543 uncommon_trap(Deoptimization::Reason_intrinsic,
1502 1544 Deoptimization::Action_make_not_entrant);
1503 1545 }
1504 1546
1505 1547 C->set_has_split_ifs(true); // Has chance for split-if optimization
1506 1548
1507 1549 push_pair(result);
1508 1550
1509 1551 return true;
1510 1552 }
1511 1553
1512 1554 //------------------------------inline_pow-------------------------------------
1513 1555 // Inline power instructions, if possible.
1514 1556 bool LibraryCallKit::inline_pow(vmIntrinsics::ID id) {
1515 1557 assert(id == vmIntrinsics::_dpow, "Not pow");
1516 1558
1517 1559 // If this inlining ever returned NaN in the past, we do not intrinsify it
1518 1560 // every again. NaN results requires StrictMath.pow handling.
1519 1561 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
1520 1562
1521 1563 // Do not intrinsify on older platforms which lack cmove.
1522 1564 if (ConditionalMoveLimit == 0) return false;
1523 1565
1524 1566 // Pseudocode for pow
1525 1567 // if (x <= 0.0) {
1526 1568 // if ((double)((int)y)==y) { // if y is int
1527 1569 // result = ((1&(int)y)==0)?-DPow(abs(x), y):DPow(abs(x), y)
1528 1570 // } else {
1529 1571 // result = NaN;
1530 1572 // }
1531 1573 // } else {
1532 1574 // result = DPow(x,y);
1533 1575 // }
1534 1576 // if (result != result)? {
1535 1577 // uncommon_trap();
1536 1578 // }
1537 1579 // return result;
1538 1580
1539 1581 _sp += arg_size(); // restore stack pointer
1540 1582 Node* y = pop_math_arg();
1541 1583 Node* x = pop_math_arg();
1542 1584
1543 1585 Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, x, y) );
1544 1586
1545 1587 // Short form: if not top-level (i.e., Math.pow but inlining Math.pow
1546 1588 // inside of something) then skip the fancy tests and just check for
1547 1589 // NaN result.
1548 1590 Node *result = NULL;
1549 1591 if( jvms()->depth() >= 1 ) {
1550 1592 result = fast_result;
1551 1593 } else {
1552 1594
1553 1595 // Set the merge point for If node with condition of (x <= 0.0)
1554 1596 // There are four possible paths to region node and phi node
1555 1597 RegionNode *r = new (C, 4) RegionNode(4);
1556 1598 Node *phi = new (C, 4) PhiNode(r, Type::DOUBLE);
1557 1599
1558 1600 // Build the first if node: if (x <= 0.0)
1559 1601 // Node for 0 constant
1560 1602 Node *zeronode = makecon(TypeD::ZERO);
1561 1603 // Check x:0
1562 1604 Node *cmp = _gvn.transform(new (C, 3) CmpDNode(x, zeronode));
1563 1605 // Check: If (x<=0) then go complex path
1564 1606 Node *bol1 = _gvn.transform( new (C, 2) BoolNode( cmp, BoolTest::le ) );
1565 1607 // Branch either way
1566 1608 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1567 1609 Node *opt_test = _gvn.transform(if1);
1568 1610 //assert( opt_test->is_If(), "Expect an IfNode");
1569 1611 IfNode *opt_if1 = (IfNode*)opt_test;
1570 1612 // Fast path taken; set region slot 3
1571 1613 Node *fast_taken = _gvn.transform( new (C, 1) IfFalseNode(opt_if1) );
1572 1614 r->init_req(3,fast_taken); // Capture fast-control
1573 1615
1574 1616 // Fast path not-taken, i.e. slow path
1575 1617 Node *complex_path = _gvn.transform( new (C, 1) IfTrueNode(opt_if1) );
1576 1618
1577 1619 // Set fast path result
1578 1620 Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, y, x) );
1579 1621 phi->init_req(3, fast_result);
1580 1622
1581 1623 // Complex path
1582 1624 // Build the second if node (if y is int)
1583 1625 // Node for (int)y
1584 1626 Node *inty = _gvn.transform( new (C, 2) ConvD2INode(y));
1585 1627 // Node for (double)((int) y)
1586 1628 Node *doubleinty= _gvn.transform( new (C, 2) ConvI2DNode(inty));
1587 1629 // Check (double)((int) y) : y
1588 1630 Node *cmpinty= _gvn.transform(new (C, 3) CmpDNode(doubleinty, y));
1589 1631 // Check if (y isn't int) then go to slow path
1590 1632
1591 1633 Node *bol2 = _gvn.transform( new (C, 2) BoolNode( cmpinty, BoolTest::ne ) );
1592 1634 // Branch either way
1593 1635 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1594 1636 Node *slow_path = opt_iff(r,if2); // Set region path 2
1595 1637
1596 1638 // Calculate DPow(abs(x), y)*(1 & (int)y)
1597 1639 // Node for constant 1
1598 1640 Node *conone = intcon(1);
1599 1641 // 1& (int)y
1600 1642 Node *signnode= _gvn.transform( new (C, 3) AndINode(conone, inty) );
1601 1643 // zero node
1602 1644 Node *conzero = intcon(0);
1603 1645 // Check (1&(int)y)==0?
1604 1646 Node *cmpeq1 = _gvn.transform(new (C, 3) CmpINode(signnode, conzero));
1605 1647 // Check if (1&(int)y)!=0?, if so the result is negative
1606 1648 Node *bol3 = _gvn.transform( new (C, 2) BoolNode( cmpeq1, BoolTest::ne ) );
1607 1649 // abs(x)
1608 1650 Node *absx=_gvn.transform( new (C, 2) AbsDNode(x));
1609 1651 // abs(x)^y
1610 1652 Node *absxpowy = _gvn.transform( new (C, 3) PowDNode(0, y, absx) );
1611 1653 // -abs(x)^y
1612 1654 Node *negabsxpowy = _gvn.transform(new (C, 2) NegDNode (absxpowy));
1613 1655 // (1&(int)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
1614 1656 Node *signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
1615 1657 // Set complex path fast result
1616 1658 phi->init_req(2, signresult);
1617 1659
1618 1660 static const jlong nan_bits = CONST64(0x7ff8000000000000);
1619 1661 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
1620 1662 r->init_req(1,slow_path);
1621 1663 phi->init_req(1,slow_result);
1622 1664
1623 1665 // Post merge
1624 1666 set_control(_gvn.transform(r));
1625 1667 record_for_igvn(r);
1626 1668 result=_gvn.transform(phi);
1627 1669 }
1628 1670
1629 1671 //-------------------
1630 1672 //result=(result.isNaN())? uncommon_trap():result;
1631 1673 // Check: If isNaN() by checking result!=result? then go to Strict Math
1632 1674 Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result));
1633 1675 // Build the boolean node
1634 1676 Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) );
1635 1677
1636 1678 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1637 1679 // End the current control-flow path
1638 1680 push_pair(x);
1639 1681 push_pair(y);
1640 1682 // Math.pow intrinsic returned a NaN, which requires StrictMath.pow
1641 1683 // to handle. Recompile without intrinsifying Math.pow.
1642 1684 uncommon_trap(Deoptimization::Reason_intrinsic,
1643 1685 Deoptimization::Action_make_not_entrant);
1644 1686 }
1645 1687
1646 1688 C->set_has_split_ifs(true); // Has chance for split-if optimization
1647 1689
1648 1690 push_pair(result);
1649 1691
1650 1692 return true;
1651 1693 }
1652 1694
1653 1695 //------------------------------inline_trans-------------------------------------
1654 1696 // Inline transcendental instructions, if possible. The Intel hardware gets
1655 1697 // these right, no funny corner cases missed.
1656 1698 bool LibraryCallKit::inline_trans(vmIntrinsics::ID id) {
1657 1699 _sp += arg_size(); // restore stack pointer
1658 1700 Node* arg = pop_math_arg();
1659 1701 Node* trans = NULL;
1660 1702
1661 1703 switch (id) {
1662 1704 case vmIntrinsics::_dlog:
1663 1705 trans = _gvn.transform((Node*)new (C, 2) LogDNode(arg));
1664 1706 break;
1665 1707 case vmIntrinsics::_dlog10:
1666 1708 trans = _gvn.transform((Node*)new (C, 2) Log10DNode(arg));
1667 1709 break;
1668 1710 default:
1669 1711 assert(false, "bad intrinsic was passed in");
1670 1712 return false;
1671 1713 }
1672 1714
1673 1715 // Push result back on JVM stack
1674 1716 push_pair(trans);
1675 1717 return true;
1676 1718 }
1677 1719
1678 1720 //------------------------------runtime_math-----------------------------
1679 1721 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1680 1722 Node* a = NULL;
1681 1723 Node* b = NULL;
1682 1724
1683 1725 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1684 1726 "must be (DD)D or (D)D type");
1685 1727
1686 1728 // Inputs
1687 1729 _sp += arg_size(); // restore stack pointer
1688 1730 if (call_type == OptoRuntime::Math_DD_D_Type()) {
1689 1731 b = pop_math_arg();
1690 1732 }
1691 1733 a = pop_math_arg();
1692 1734
1693 1735 const TypePtr* no_memory_effects = NULL;
1694 1736 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1695 1737 no_memory_effects,
1696 1738 a, top(), b, b ? top() : NULL);
1697 1739 Node* value = _gvn.transform(new (C, 1) ProjNode(trig, TypeFunc::Parms+0));
1698 1740 #ifdef ASSERT
1699 1741 Node* value_top = _gvn.transform(new (C, 1) ProjNode(trig, TypeFunc::Parms+1));
1700 1742 assert(value_top == top(), "second value must be top");
1701 1743 #endif
1702 1744
1703 1745 push_pair(value);
1704 1746 return true;
1705 1747 }
1706 1748
1707 1749 //------------------------------inline_math_native-----------------------------
1708 1750 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1709 1751 switch (id) {
1710 1752 // These intrinsics are not properly supported on all hardware
1711 1753 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) :
1712 1754 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dcos), "COS");
1713 1755 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) :
1714 1756 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dsin), "SIN");
1715 1757 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) :
1716 1758 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dtan), "TAN");
1717 1759
1718 1760 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_trans(id) :
1719 1761 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog), "LOG");
1720 1762 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_trans(id) :
1721 1763 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog10), "LOG10");
1722 1764
1723 1765 // These intrinsics are supported on all hardware
1724 1766 case vmIntrinsics::_dsqrt: return Matcher::has_match_rule(Op_SqrtD) ? inline_sqrt(id) : false;
1725 1767 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_abs(id) : false;
1726 1768
1727 1769 // These intrinsics don't work on X86. The ad implementation doesn't
1728 1770 // handle NaN's properly. Instead of returning infinity, the ad
1729 1771 // implementation returns a NaN on overflow. See bug: 6304089
1730 1772 // Once the ad implementations are fixed, change the code below
1731 1773 // to match the intrinsics above
1732 1774
1733 1775 case vmIntrinsics::_dexp: return
1734 1776 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP");
1735 1777 case vmIntrinsics::_dpow: return
1736 1778 runtime_math(OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
1737 1779
1738 1780 // These intrinsics are not yet correctly implemented
1739 1781 case vmIntrinsics::_datan2:
1740 1782 return false;
1741 1783
1742 1784 default:
1743 1785 ShouldNotReachHere();
1744 1786 return false;
1745 1787 }
1746 1788 }
1747 1789
1748 1790 static bool is_simple_name(Node* n) {
1749 1791 return (n->req() == 1 // constant
1750 1792 || (n->is_Type() && n->as_Type()->type()->singleton())
1751 1793 || n->is_Proj() // parameter or return value
1752 1794 || n->is_Phi() // local of some sort
1753 1795 );
1754 1796 }
1755 1797
1756 1798 //----------------------------inline_min_max-----------------------------------
1757 1799 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
1758 1800 push(generate_min_max(id, argument(0), argument(1)));
1759 1801
1760 1802 return true;
1761 1803 }
1762 1804
1763 1805 Node*
1764 1806 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
1765 1807 // These are the candidate return value:
1766 1808 Node* xvalue = x0;
1767 1809 Node* yvalue = y0;
1768 1810
1769 1811 if (xvalue == yvalue) {
1770 1812 return xvalue;
1771 1813 }
1772 1814
1773 1815 bool want_max = (id == vmIntrinsics::_max);
1774 1816
1775 1817 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
1776 1818 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
1777 1819 if (txvalue == NULL || tyvalue == NULL) return top();
1778 1820 // This is not really necessary, but it is consistent with a
1779 1821 // hypothetical MaxINode::Value method:
1780 1822 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
1781 1823
1782 1824 // %%% This folding logic should (ideally) be in a different place.
1783 1825 // Some should be inside IfNode, and there to be a more reliable
1784 1826 // transformation of ?: style patterns into cmoves. We also want
1785 1827 // more powerful optimizations around cmove and min/max.
1786 1828
1787 1829 // Try to find a dominating comparison of these guys.
1788 1830 // It can simplify the index computation for Arrays.copyOf
1789 1831 // and similar uses of System.arraycopy.
1790 1832 // First, compute the normalized version of CmpI(x, y).
1791 1833 int cmp_op = Op_CmpI;
1792 1834 Node* xkey = xvalue;
1793 1835 Node* ykey = yvalue;
1794 1836 Node* ideal_cmpxy = _gvn.transform( new(C, 3) CmpINode(xkey, ykey) );
1795 1837 if (ideal_cmpxy->is_Cmp()) {
1796 1838 // E.g., if we have CmpI(length - offset, count),
1797 1839 // it might idealize to CmpI(length, count + offset)
1798 1840 cmp_op = ideal_cmpxy->Opcode();
1799 1841 xkey = ideal_cmpxy->in(1);
1800 1842 ykey = ideal_cmpxy->in(2);
1801 1843 }
1802 1844
1803 1845 // Start by locating any relevant comparisons.
1804 1846 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
1805 1847 Node* cmpxy = NULL;
1806 1848 Node* cmpyx = NULL;
1807 1849 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
1808 1850 Node* cmp = start_from->fast_out(k);
1809 1851 if (cmp->outcnt() > 0 && // must have prior uses
1810 1852 cmp->in(0) == NULL && // must be context-independent
1811 1853 cmp->Opcode() == cmp_op) { // right kind of compare
1812 1854 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
1813 1855 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
1814 1856 }
1815 1857 }
1816 1858
1817 1859 const int NCMPS = 2;
1818 1860 Node* cmps[NCMPS] = { cmpxy, cmpyx };
1819 1861 int cmpn;
1820 1862 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1821 1863 if (cmps[cmpn] != NULL) break; // find a result
1822 1864 }
1823 1865 if (cmpn < NCMPS) {
1824 1866 // Look for a dominating test that tells us the min and max.
1825 1867 int depth = 0; // Limit search depth for speed
1826 1868 Node* dom = control();
1827 1869 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
1828 1870 if (++depth >= 100) break;
1829 1871 Node* ifproj = dom;
1830 1872 if (!ifproj->is_Proj()) continue;
1831 1873 Node* iff = ifproj->in(0);
1832 1874 if (!iff->is_If()) continue;
1833 1875 Node* bol = iff->in(1);
1834 1876 if (!bol->is_Bool()) continue;
1835 1877 Node* cmp = bol->in(1);
1836 1878 if (cmp == NULL) continue;
1837 1879 for (cmpn = 0; cmpn < NCMPS; cmpn++)
1838 1880 if (cmps[cmpn] == cmp) break;
1839 1881 if (cmpn == NCMPS) continue;
1840 1882 BoolTest::mask btest = bol->as_Bool()->_test._test;
1841 1883 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
1842 1884 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1843 1885 // At this point, we know that 'x btest y' is true.
1844 1886 switch (btest) {
1845 1887 case BoolTest::eq:
1846 1888 // They are proven equal, so we can collapse the min/max.
1847 1889 // Either value is the answer. Choose the simpler.
1848 1890 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
1849 1891 return yvalue;
1850 1892 return xvalue;
1851 1893 case BoolTest::lt: // x < y
1852 1894 case BoolTest::le: // x <= y
1853 1895 return (want_max ? yvalue : xvalue);
1854 1896 case BoolTest::gt: // x > y
1855 1897 case BoolTest::ge: // x >= y
1856 1898 return (want_max ? xvalue : yvalue);
1857 1899 }
1858 1900 }
1859 1901 }
1860 1902
1861 1903 // We failed to find a dominating test.
1862 1904 // Let's pick a test that might GVN with prior tests.
1863 1905 Node* best_bol = NULL;
1864 1906 BoolTest::mask best_btest = BoolTest::illegal;
1865 1907 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1866 1908 Node* cmp = cmps[cmpn];
1867 1909 if (cmp == NULL) continue;
1868 1910 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
1869 1911 Node* bol = cmp->fast_out(j);
1870 1912 if (!bol->is_Bool()) continue;
1871 1913 BoolTest::mask btest = bol->as_Bool()->_test._test;
1872 1914 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
1873 1915 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1874 1916 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
1875 1917 best_bol = bol->as_Bool();
1876 1918 best_btest = btest;
1877 1919 }
1878 1920 }
1879 1921 }
1880 1922
1881 1923 Node* answer_if_true = NULL;
1882 1924 Node* answer_if_false = NULL;
1883 1925 switch (best_btest) {
1884 1926 default:
1885 1927 if (cmpxy == NULL)
1886 1928 cmpxy = ideal_cmpxy;
1887 1929 best_bol = _gvn.transform( new(C, 2) BoolNode(cmpxy, BoolTest::lt) );
1888 1930 // and fall through:
1889 1931 case BoolTest::lt: // x < y
1890 1932 case BoolTest::le: // x <= y
1891 1933 answer_if_true = (want_max ? yvalue : xvalue);
1892 1934 answer_if_false = (want_max ? xvalue : yvalue);
1893 1935 break;
1894 1936 case BoolTest::gt: // x > y
1895 1937 case BoolTest::ge: // x >= y
1896 1938 answer_if_true = (want_max ? xvalue : yvalue);
1897 1939 answer_if_false = (want_max ? yvalue : xvalue);
1898 1940 break;
1899 1941 }
1900 1942
1901 1943 jint hi, lo;
1902 1944 if (want_max) {
1903 1945 // We can sharpen the minimum.
1904 1946 hi = MAX2(txvalue->_hi, tyvalue->_hi);
1905 1947 lo = MAX2(txvalue->_lo, tyvalue->_lo);
1906 1948 } else {
1907 1949 // We can sharpen the maximum.
1908 1950 hi = MIN2(txvalue->_hi, tyvalue->_hi);
1909 1951 lo = MIN2(txvalue->_lo, tyvalue->_lo);
1910 1952 }
1911 1953
1912 1954 // Use a flow-free graph structure, to avoid creating excess control edges
1913 1955 // which could hinder other optimizations.
1914 1956 // Since Math.min/max is often used with arraycopy, we want
1915 1957 // tightly_coupled_allocation to be able to see beyond min/max expressions.
1916 1958 Node* cmov = CMoveNode::make(C, NULL, best_bol,
1917 1959 answer_if_false, answer_if_true,
1918 1960 TypeInt::make(lo, hi, widen));
1919 1961
1920 1962 return _gvn.transform(cmov);
1921 1963
1922 1964 /*
1923 1965 // This is not as desirable as it may seem, since Min and Max
1924 1966 // nodes do not have a full set of optimizations.
1925 1967 // And they would interfere, anyway, with 'if' optimizations
1926 1968 // and with CMoveI canonical forms.
1927 1969 switch (id) {
1928 1970 case vmIntrinsics::_min:
1929 1971 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
1930 1972 case vmIntrinsics::_max:
1931 1973 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
1932 1974 default:
1933 1975 ShouldNotReachHere();
1934 1976 }
1935 1977 */
1936 1978 }
1937 1979
1938 1980 inline int
1939 1981 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
1940 1982 const TypePtr* base_type = TypePtr::NULL_PTR;
1941 1983 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
1942 1984 if (base_type == NULL) {
1943 1985 // Unknown type.
1944 1986 return Type::AnyPtr;
1945 1987 } else if (base_type == TypePtr::NULL_PTR) {
1946 1988 // Since this is a NULL+long form, we have to switch to a rawptr.
1947 1989 base = _gvn.transform( new (C, 2) CastX2PNode(offset) );
1948 1990 offset = MakeConX(0);
1949 1991 return Type::RawPtr;
1950 1992 } else if (base_type->base() == Type::RawPtr) {
1951 1993 return Type::RawPtr;
1952 1994 } else if (base_type->isa_oopptr()) {
1953 1995 // Base is never null => always a heap address.
1954 1996 if (base_type->ptr() == TypePtr::NotNull) {
1955 1997 return Type::OopPtr;
1956 1998 }
1957 1999 // Offset is small => always a heap address.
1958 2000 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
1959 2001 if (offset_type != NULL &&
1960 2002 base_type->offset() == 0 && // (should always be?)
1961 2003 offset_type->_lo >= 0 &&
1962 2004 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
1963 2005 return Type::OopPtr;
1964 2006 }
1965 2007 // Otherwise, it might either be oop+off or NULL+addr.
1966 2008 return Type::AnyPtr;
1967 2009 } else {
1968 2010 // No information:
1969 2011 return Type::AnyPtr;
1970 2012 }
1971 2013 }
1972 2014
1973 2015 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
1974 2016 int kind = classify_unsafe_addr(base, offset);
1975 2017 if (kind == Type::RawPtr) {
1976 2018 return basic_plus_adr(top(), base, offset);
1977 2019 } else {
1978 2020 return basic_plus_adr(base, offset);
1979 2021 }
1980 2022 }
1981 2023
1982 2024 //-------------------inline_numberOfLeadingZeros_int/long-----------------------
1983 2025 // inline int Integer.numberOfLeadingZeros(int)
1984 2026 // inline int Long.numberOfLeadingZeros(long)
1985 2027 bool LibraryCallKit::inline_numberOfLeadingZeros(vmIntrinsics::ID id) {
1986 2028 assert(id == vmIntrinsics::_numberOfLeadingZeros_i || id == vmIntrinsics::_numberOfLeadingZeros_l, "not numberOfLeadingZeros");
1987 2029 if (id == vmIntrinsics::_numberOfLeadingZeros_i && !Matcher::match_rule_supported(Op_CountLeadingZerosI)) return false;
1988 2030 if (id == vmIntrinsics::_numberOfLeadingZeros_l && !Matcher::match_rule_supported(Op_CountLeadingZerosL)) return false;
1989 2031 _sp += arg_size(); // restore stack pointer
1990 2032 switch (id) {
1991 2033 case vmIntrinsics::_numberOfLeadingZeros_i:
1992 2034 push(_gvn.transform(new (C, 2) CountLeadingZerosINode(pop())));
1993 2035 break;
1994 2036 case vmIntrinsics::_numberOfLeadingZeros_l:
1995 2037 push(_gvn.transform(new (C, 2) CountLeadingZerosLNode(pop_pair())));
1996 2038 break;
1997 2039 default:
1998 2040 ShouldNotReachHere();
1999 2041 }
2000 2042 return true;
2001 2043 }
2002 2044
2003 2045 //-------------------inline_numberOfTrailingZeros_int/long----------------------
2004 2046 // inline int Integer.numberOfTrailingZeros(int)
2005 2047 // inline int Long.numberOfTrailingZeros(long)
2006 2048 bool LibraryCallKit::inline_numberOfTrailingZeros(vmIntrinsics::ID id) {
2007 2049 assert(id == vmIntrinsics::_numberOfTrailingZeros_i || id == vmIntrinsics::_numberOfTrailingZeros_l, "not numberOfTrailingZeros");
2008 2050 if (id == vmIntrinsics::_numberOfTrailingZeros_i && !Matcher::match_rule_supported(Op_CountTrailingZerosI)) return false;
2009 2051 if (id == vmIntrinsics::_numberOfTrailingZeros_l && !Matcher::match_rule_supported(Op_CountTrailingZerosL)) return false;
2010 2052 _sp += arg_size(); // restore stack pointer
2011 2053 switch (id) {
2012 2054 case vmIntrinsics::_numberOfTrailingZeros_i:
2013 2055 push(_gvn.transform(new (C, 2) CountTrailingZerosINode(pop())));
2014 2056 break;
2015 2057 case vmIntrinsics::_numberOfTrailingZeros_l:
2016 2058 push(_gvn.transform(new (C, 2) CountTrailingZerosLNode(pop_pair())));
2017 2059 break;
2018 2060 default:
2019 2061 ShouldNotReachHere();
2020 2062 }
2021 2063 return true;
2022 2064 }
2023 2065
2024 2066 //----------------------------inline_bitCount_int/long-----------------------
2025 2067 // inline int Integer.bitCount(int)
2026 2068 // inline int Long.bitCount(long)
2027 2069 bool LibraryCallKit::inline_bitCount(vmIntrinsics::ID id) {
2028 2070 assert(id == vmIntrinsics::_bitCount_i || id == vmIntrinsics::_bitCount_l, "not bitCount");
2029 2071 if (id == vmIntrinsics::_bitCount_i && !Matcher::has_match_rule(Op_PopCountI)) return false;
2030 2072 if (id == vmIntrinsics::_bitCount_l && !Matcher::has_match_rule(Op_PopCountL)) return false;
2031 2073 _sp += arg_size(); // restore stack pointer
2032 2074 switch (id) {
2033 2075 case vmIntrinsics::_bitCount_i:
2034 2076 push(_gvn.transform(new (C, 2) PopCountINode(pop())));
2035 2077 break;
2036 2078 case vmIntrinsics::_bitCount_l:
2037 2079 push(_gvn.transform(new (C, 2) PopCountLNode(pop_pair())));
2038 2080 break;
2039 2081 default:
2040 2082 ShouldNotReachHere();
2041 2083 }
2042 2084 return true;
2043 2085 }
2044 2086
2045 2087 //----------------------------inline_reverseBytes_int/long/char/short-------------------
2046 2088 // inline Integer.reverseBytes(int)
2047 2089 // inline Long.reverseBytes(long)
2048 2090 // inline Character.reverseBytes(char)
2049 2091 // inline Short.reverseBytes(short)
2050 2092 bool LibraryCallKit::inline_reverseBytes(vmIntrinsics::ID id) {
2051 2093 assert(id == vmIntrinsics::_reverseBytes_i || id == vmIntrinsics::_reverseBytes_l ||
2052 2094 id == vmIntrinsics::_reverseBytes_c || id == vmIntrinsics::_reverseBytes_s,
2053 2095 "not reverse Bytes");
2054 2096 if (id == vmIntrinsics::_reverseBytes_i && !Matcher::has_match_rule(Op_ReverseBytesI)) return false;
2055 2097 if (id == vmIntrinsics::_reverseBytes_l && !Matcher::has_match_rule(Op_ReverseBytesL)) return false;
2056 2098 if (id == vmIntrinsics::_reverseBytes_c && !Matcher::has_match_rule(Op_ReverseBytesUS)) return false;
2057 2099 if (id == vmIntrinsics::_reverseBytes_s && !Matcher::has_match_rule(Op_ReverseBytesS)) return false;
2058 2100 _sp += arg_size(); // restore stack pointer
2059 2101 switch (id) {
2060 2102 case vmIntrinsics::_reverseBytes_i:
2061 2103 push(_gvn.transform(new (C, 2) ReverseBytesINode(0, pop())));
2062 2104 break;
2063 2105 case vmIntrinsics::_reverseBytes_l:
2064 2106 push_pair(_gvn.transform(new (C, 2) ReverseBytesLNode(0, pop_pair())));
2065 2107 break;
2066 2108 case vmIntrinsics::_reverseBytes_c:
2067 2109 push(_gvn.transform(new (C, 2) ReverseBytesUSNode(0, pop())));
2068 2110 break;
2069 2111 case vmIntrinsics::_reverseBytes_s:
2070 2112 push(_gvn.transform(new (C, 2) ReverseBytesSNode(0, pop())));
2071 2113 break;
|
↓ open down ↓ |
1398 lines elided |
↑ open up ↑ |
2072 2114 default:
2073 2115 ;
2074 2116 }
2075 2117 return true;
2076 2118 }
2077 2119
2078 2120 //----------------------------inline_unsafe_access----------------------------
2079 2121
2080 2122 const static BasicType T_ADDRESS_HOLDER = T_LONG;
2081 2123
2124 +// Helper that guards and inserts a G1 pre-barrier.
2125 +void LibraryCallKit::insert_g1_pre_barrier(Node* base_oop, Node* offset, Node* pre_val) {
2126 + assert(UseG1GC, "should not call this otherwise");
2127 +
2128 + // We need to generate the following....
2129 + //
2130 + // if (offset == java_lang_ref_Reference::_reference_offset) {
2131 + // if (base != null) {
2132 + // if (klass(base)->reference_type() != REF_NONE)) {
2133 + // pre_barrier(_, pre_val, ...);
2134 + // }
2135 + // }
2136 + // }
2137 +
2138 + float likely = PROB_LIKELY(0.999);
2139 + float unlikely = PROB_UNLIKELY(0.999);
2140 +
2141 + IdealKit ideal(gvn(), control(), merged_memory());
2142 +#define __ ideal.
2143 +
2144 + const int reference_type_offset = instanceKlass::reference_type_offset_in_bytes() +
2145 + sizeof(oopDesc);
2146 +
2147 + Node* referent_off = __ ConI(java_lang_ref_Reference::referent_offset);
2148 +
2149 + __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2150 + __ if_then(base_oop, BoolTest::ne, null(), likely); {
2151 + Node* k_adr = __ AddP(base_oop, base_oop, __ ConX(oopDesc::klass_offset_in_bytes()));
2152 +
2153 + Node* klass = _gvn.transform( LoadKlassNode::make(gvn(), immutable_memory(), k_adr,
2154 + TypeRawPtr::BOTTOM,
2155 + TypeKlassPtr::OBJECT_OR_NULL) );
2156 +
2157 + Node* ref_typ_adr = __ AddP(klass, klass, __ ConX(reference_type_offset));
2158 + Node* ref_none_val = __ ConI(REF_NONE);
2159 +
2160 + Node* ref_typ = __ load(__ ctrl(), ref_typ_adr, TypeInt::INT, T_INT, Compile::AliasIdxRaw);
2161 +
2162 + __ if_then(ref_typ, BoolTest::ne, ref_none_val, unlikely); {
2163 +
2164 + // Sync IdealKit and graphKit.
2165 + set_all_memory( __ merged_memory());
2166 +
2167 + // Use the pre-barrier to record the value in the referent field
2168 + pre_barrier(false /* do_load */,
2169 + __ ctrl(),
2170 + NULL /* obj */, NULL /* adr */, -1 /* alias_idx */, NULL /* val */, NULL /* val_type */,
2171 + pre_val /* pre_val */,
2172 + T_OBJECT);
2173 + // Update IdealKit memory.
2174 + __ set_all_memory(merged_memory());
2175 + __ set_ctrl(control());
2176 + } __ end_if(); // _ref_type != ref_none
2177 + } __ end_if(); // base != NULL
2178 + } __ end_if(); // offset == referent_offset
2179 +
2180 + // Final sync IdealKit and GraphKit.
2181 + sync_kit(ideal);
2182 +#undef __
2183 +}
2184 +
2185 +
2082 2186 // Interpret Unsafe.fieldOffset cookies correctly:
2083 2187 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
2084 2188
2085 2189 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) {
2086 2190 if (callee()->is_static()) return false; // caller must have the capability!
2087 2191
2088 2192 #ifndef PRODUCT
2089 2193 {
2090 2194 ResourceMark rm;
2091 2195 // Check the signatures.
2092 2196 ciSignature* sig = signature();
2093 2197 #ifdef ASSERT
2094 2198 if (!is_store) {
2095 2199 // Object getObject(Object base, int/long offset), etc.
2096 2200 BasicType rtype = sig->return_type()->basic_type();
2097 2201 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
2098 2202 rtype = T_ADDRESS; // it is really a C void*
2099 2203 assert(rtype == type, "getter must return the expected value");
2100 2204 if (!is_native_ptr) {
2101 2205 assert(sig->count() == 2, "oop getter has 2 arguments");
2102 2206 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2103 2207 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2104 2208 } else {
2105 2209 assert(sig->count() == 1, "native getter has 1 argument");
2106 2210 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
2107 2211 }
2108 2212 } else {
2109 2213 // void putObject(Object base, int/long offset, Object x), etc.
2110 2214 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2111 2215 if (!is_native_ptr) {
2112 2216 assert(sig->count() == 3, "oop putter has 3 arguments");
2113 2217 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2114 2218 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2115 2219 } else {
2116 2220 assert(sig->count() == 2, "native putter has 2 arguments");
2117 2221 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
2118 2222 }
2119 2223 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2120 2224 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
2121 2225 vtype = T_ADDRESS; // it is really a C void*
2122 2226 assert(vtype == type, "putter must accept the expected value");
2123 2227 }
2124 2228 #endif // ASSERT
2125 2229 }
2126 2230 #endif //PRODUCT
2127 2231
2128 2232 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2129 2233
2130 2234 int type_words = type2size[ (type == T_ADDRESS) ? T_LONG : type ];
2131 2235
2132 2236 // Argument words: "this" plus (oop/offset) or (lo/hi) args plus maybe 1 or 2 value words
2133 2237 int nargs = 1 + (is_native_ptr ? 2 : 3) + (is_store ? type_words : 0);
2134 2238
2135 2239 debug_only(int saved_sp = _sp);
2136 2240 _sp += nargs;
2137 2241
2138 2242 Node* val;
2139 2243 debug_only(val = (Node*)(uintptr_t)-1);
2140 2244
2141 2245
2142 2246 if (is_store) {
2143 2247 // Get the value being stored. (Pop it first; it was pushed last.)
2144 2248 switch (type) {
2145 2249 case T_DOUBLE:
2146 2250 case T_LONG:
2147 2251 case T_ADDRESS:
|
↓ open down ↓ |
56 lines elided |
↑ open up ↑ |
2148 2252 val = pop_pair();
2149 2253 break;
2150 2254 default:
2151 2255 val = pop();
2152 2256 }
2153 2257 }
2154 2258
2155 2259 // Build address expression. See the code in inline_unsafe_prefetch.
2156 2260 Node *adr;
2157 2261 Node *heap_base_oop = top();
2262 + Node* offset = top();
2263 +
2158 2264 if (!is_native_ptr) {
2159 2265 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2160 - Node* offset = pop_pair();
2266 + offset = pop_pair();
2161 2267 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2162 2268 Node* base = pop();
2163 2269 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2164 2270 // to be plain byte offsets, which are also the same as those accepted
2165 2271 // by oopDesc::field_base.
2166 2272 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2167 2273 "fieldOffset must be byte-scaled");
2168 2274 // 32-bit machines ignore the high half!
2169 2275 offset = ConvL2X(offset);
2170 2276 adr = make_unsafe_address(base, offset);
2171 2277 heap_base_oop = base;
2172 2278 } else {
2173 2279 Node* ptr = pop_pair();
2174 2280 // Adjust Java long to machine word:
2175 2281 ptr = ConvL2X(ptr);
2176 2282 adr = make_unsafe_address(NULL, ptr);
2177 2283 }
2178 2284
2179 2285 // Pop receiver last: it was pushed first.
2180 2286 Node *receiver = pop();
2181 2287
2182 2288 assert(saved_sp == _sp, "must have correct argument count");
2183 2289
2184 2290 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2185 2291
2186 2292 // First guess at the value type.
2187 2293 const Type *value_type = Type::get_const_basic_type(type);
2188 2294
2189 2295 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM,
2190 2296 // there was not enough information to nail it down.
|
↓ open down ↓ |
20 lines elided |
↑ open up ↑ |
2191 2297 Compile::AliasType* alias_type = C->alias_type(adr_type);
2192 2298 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2193 2299
2194 2300 // We will need memory barriers unless we can determine a unique
2195 2301 // alias category for this reference. (Note: If for some reason
2196 2302 // the barriers get omitted and the unsafe reference begins to "pollute"
2197 2303 // the alias analysis of the rest of the graph, either Compile::can_alias
2198 2304 // or Compile::must_alias will throw a diagnostic assert.)
2199 2305 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
2200 2306
2307 + // If we are reading the value of the referent field of a Reference
2308 + // object (either by using Unsafe directly or through reflection)
2309 + // then, if G1 is enabled, we need to record the referent in an
2310 + // SATB log buffer using the pre-barrier mechanism.
2311 + bool need_read_barrier = UseG1GC && !is_native_ptr && !is_store &&
2312 + offset != top() && heap_base_oop != top();
2313 +
2201 2314 if (!is_store && type == T_OBJECT) {
2202 2315 // Attempt to infer a sharper value type from the offset and base type.
2203 2316 ciKlass* sharpened_klass = NULL;
2204 2317
2205 2318 // See if it is an instance field, with an object type.
2206 2319 if (alias_type->field() != NULL) {
2207 2320 assert(!is_native_ptr, "native pointer op cannot use a java address");
2208 2321 if (alias_type->field()->type()->is_klass()) {
2209 2322 sharpened_klass = alias_type->field()->type()->as_klass();
2210 2323 }
2211 2324 }
2212 2325
2213 2326 // See if it is a narrow oop array.
2214 2327 if (adr_type->isa_aryptr()) {
2215 2328 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2216 2329 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2217 2330 if (elem_type != NULL) {
2218 2331 sharpened_klass = elem_type->klass();
2219 2332 }
2220 2333 }
2221 2334 }
2222 2335
2223 2336 if (sharpened_klass != NULL) {
2224 2337 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2225 2338
2226 2339 // Sharpen the value type.
2227 2340 value_type = tjp;
2228 2341
2229 2342 #ifndef PRODUCT
2230 2343 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) {
2231 2344 tty->print(" from base type: "); adr_type->dump();
2232 2345 tty->print(" sharpened value: "); value_type->dump();
2233 2346 }
2234 2347 #endif
2235 2348 }
2236 2349 }
2237 2350
2238 2351 // Null check on self without removing any arguments. The argument
2239 2352 // null check technically happens in the wrong place, which can lead to
2240 2353 // invalid stack traces when the primitive is inlined into a method
2241 2354 // which handles NullPointerExceptions.
2242 2355 _sp += nargs;
2243 2356 do_null_check(receiver, T_OBJECT);
2244 2357 _sp -= nargs;
2245 2358 if (stopped()) {
2246 2359 return true;
2247 2360 }
2248 2361 // Heap pointers get a null-check from the interpreter,
2249 2362 // as a courtesy. However, this is not guaranteed by Unsafe,
2250 2363 // and it is not possible to fully distinguish unintended nulls
2251 2364 // from intended ones in this API.
2252 2365
2253 2366 if (is_volatile) {
2254 2367 // We need to emit leading and trailing CPU membars (see below) in
2255 2368 // addition to memory membars when is_volatile. This is a little
2256 2369 // too strong, but avoids the need to insert per-alias-type
2257 2370 // volatile membars (for stores; compare Parse::do_put_xxx), which
2258 2371 // we cannot do effectively here because we probably only have a
2259 2372 // rough approximation of type.
2260 2373 need_mem_bar = true;
2261 2374 // For Stores, place a memory ordering barrier now.
2262 2375 if (is_store)
2263 2376 insert_mem_bar(Op_MemBarRelease);
2264 2377 }
2265 2378
2266 2379 // Memory barrier to prevent normal and 'unsafe' accesses from
2267 2380 // bypassing each other. Happens after null checks, so the
2268 2381 // exception paths do not take memory state from the memory barrier,
2269 2382 // so there's no problems making a strong assert about mixing users
2270 2383 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar
2271 2384 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl.
2272 2385 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2273 2386
|
↓ open down ↓ |
63 lines elided |
↑ open up ↑ |
2274 2387 if (!is_store) {
2275 2388 Node* p = make_load(control(), adr, value_type, type, adr_type, is_volatile);
2276 2389 // load value and push onto stack
2277 2390 switch (type) {
2278 2391 case T_BOOLEAN:
2279 2392 case T_CHAR:
2280 2393 case T_BYTE:
2281 2394 case T_SHORT:
2282 2395 case T_INT:
2283 2396 case T_FLOAT:
2397 + push(p);
2398 + break;
2284 2399 case T_OBJECT:
2285 - push( p );
2400 + if (need_read_barrier) {
2401 + insert_g1_pre_barrier(heap_base_oop, offset, p);
2402 + }
2403 + push(p);
2286 2404 break;
2287 2405 case T_ADDRESS:
2288 2406 // Cast to an int type.
2289 2407 p = _gvn.transform( new (C, 2) CastP2XNode(NULL,p) );
2290 2408 p = ConvX2L(p);
2291 2409 push_pair(p);
2292 2410 break;
2293 2411 case T_DOUBLE:
2294 2412 case T_LONG:
2295 2413 push_pair( p );
2296 2414 break;
2297 2415 default: ShouldNotReachHere();
2298 2416 }
2299 2417 } else {
2300 2418 // place effect of store into memory
2301 2419 switch (type) {
2302 2420 case T_DOUBLE:
2303 2421 val = dstore_rounding(val);
2304 2422 break;
2305 2423 case T_ADDRESS:
2306 2424 // Repackage the long as a pointer.
2307 2425 val = ConvL2X(val);
2308 2426 val = _gvn.transform( new (C, 2) CastX2PNode(val) );
2309 2427 break;
2310 2428 }
2311 2429
2312 2430 if (type != T_OBJECT ) {
2313 2431 (void) store_to_memory(control(), adr, val, type, adr_type, is_volatile);
2314 2432 } else {
2315 2433 // Possibly an oop being stored to Java heap or native memory
2316 2434 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
2317 2435 // oop to Java heap.
2318 2436 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type);
2319 2437 } else {
2320 2438 // We can't tell at compile time if we are storing in the Java heap or outside
2321 2439 // of it. So we need to emit code to conditionally do the proper type of
2322 2440 // store.
2323 2441
2324 2442 IdealKit ideal(gvn(), control(), merged_memory());
2325 2443 #define __ ideal.
2326 2444 // QQQ who knows what probability is here??
2327 2445 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
2328 2446 // Sync IdealKit and graphKit.
2329 2447 set_all_memory( __ merged_memory());
2330 2448 set_control(__ ctrl());
2331 2449 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type);
2332 2450 // Update IdealKit memory.
2333 2451 __ set_all_memory(merged_memory());
2334 2452 __ set_ctrl(control());
2335 2453 } __ else_(); {
2336 2454 __ store(__ ctrl(), adr, val, type, alias_type->index(), is_volatile);
2337 2455 } __ end_if();
2338 2456 // Final sync IdealKit and GraphKit.
2339 2457 sync_kit(ideal);
2340 2458 #undef __
2341 2459 }
2342 2460 }
2343 2461 }
2344 2462
2345 2463 if (is_volatile) {
2346 2464 if (!is_store)
2347 2465 insert_mem_bar(Op_MemBarAcquire);
2348 2466 else
2349 2467 insert_mem_bar(Op_MemBarVolatile);
2350 2468 }
2351 2469
2352 2470 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2353 2471
2354 2472 return true;
2355 2473 }
2356 2474
2357 2475 //----------------------------inline_unsafe_prefetch----------------------------
2358 2476
2359 2477 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) {
2360 2478 #ifndef PRODUCT
2361 2479 {
2362 2480 ResourceMark rm;
2363 2481 // Check the signatures.
2364 2482 ciSignature* sig = signature();
2365 2483 #ifdef ASSERT
2366 2484 // Object getObject(Object base, int/long offset), etc.
2367 2485 BasicType rtype = sig->return_type()->basic_type();
2368 2486 if (!is_native_ptr) {
2369 2487 assert(sig->count() == 2, "oop prefetch has 2 arguments");
2370 2488 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object");
2371 2489 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct");
2372 2490 } else {
2373 2491 assert(sig->count() == 1, "native prefetch has 1 argument");
2374 2492 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long");
2375 2493 }
2376 2494 #endif // ASSERT
2377 2495 }
2378 2496 #endif // !PRODUCT
2379 2497
2380 2498 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2381 2499
2382 2500 // Argument words: "this" if not static, plus (oop/offset) or (lo/hi) args
2383 2501 int nargs = (is_static ? 0 : 1) + (is_native_ptr ? 2 : 3);
2384 2502
2385 2503 debug_only(int saved_sp = _sp);
2386 2504 _sp += nargs;
2387 2505
2388 2506 // Build address expression. See the code in inline_unsafe_access.
2389 2507 Node *adr;
2390 2508 if (!is_native_ptr) {
2391 2509 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2392 2510 Node* offset = pop_pair();
2393 2511 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2394 2512 Node* base = pop();
2395 2513 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2396 2514 // to be plain byte offsets, which are also the same as those accepted
2397 2515 // by oopDesc::field_base.
2398 2516 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2399 2517 "fieldOffset must be byte-scaled");
2400 2518 // 32-bit machines ignore the high half!
2401 2519 offset = ConvL2X(offset);
2402 2520 adr = make_unsafe_address(base, offset);
2403 2521 } else {
2404 2522 Node* ptr = pop_pair();
2405 2523 // Adjust Java long to machine word:
2406 2524 ptr = ConvL2X(ptr);
2407 2525 adr = make_unsafe_address(NULL, ptr);
2408 2526 }
2409 2527
2410 2528 if (is_static) {
2411 2529 assert(saved_sp == _sp, "must have correct argument count");
2412 2530 } else {
2413 2531 // Pop receiver last: it was pushed first.
2414 2532 Node *receiver = pop();
2415 2533 assert(saved_sp == _sp, "must have correct argument count");
2416 2534
2417 2535 // Null check on self without removing any arguments. The argument
2418 2536 // null check technically happens in the wrong place, which can lead to
2419 2537 // invalid stack traces when the primitive is inlined into a method
2420 2538 // which handles NullPointerExceptions.
2421 2539 _sp += nargs;
2422 2540 do_null_check(receiver, T_OBJECT);
2423 2541 _sp -= nargs;
2424 2542 if (stopped()) {
2425 2543 return true;
2426 2544 }
2427 2545 }
2428 2546
2429 2547 // Generate the read or write prefetch
2430 2548 Node *prefetch;
2431 2549 if (is_store) {
2432 2550 prefetch = new (C, 3) PrefetchWriteNode(i_o(), adr);
2433 2551 } else {
2434 2552 prefetch = new (C, 3) PrefetchReadNode(i_o(), adr);
2435 2553 }
2436 2554 prefetch->init_req(0, control());
2437 2555 set_i_o(_gvn.transform(prefetch));
2438 2556
2439 2557 return true;
2440 2558 }
2441 2559
2442 2560 //----------------------------inline_unsafe_CAS----------------------------
2443 2561
2444 2562 bool LibraryCallKit::inline_unsafe_CAS(BasicType type) {
2445 2563 // This basic scheme here is the same as inline_unsafe_access, but
2446 2564 // differs in enough details that combining them would make the code
2447 2565 // overly confusing. (This is a true fact! I originally combined
2448 2566 // them, but even I was confused by it!) As much code/comments as
2449 2567 // possible are retained from inline_unsafe_access though to make
2450 2568 // the correspondences clearer. - dl
2451 2569
2452 2570 if (callee()->is_static()) return false; // caller must have the capability!
2453 2571
2454 2572 #ifndef PRODUCT
2455 2573 {
2456 2574 ResourceMark rm;
2457 2575 // Check the signatures.
2458 2576 ciSignature* sig = signature();
2459 2577 #ifdef ASSERT
2460 2578 BasicType rtype = sig->return_type()->basic_type();
2461 2579 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2462 2580 assert(sig->count() == 4, "CAS has 4 arguments");
2463 2581 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2464 2582 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2465 2583 #endif // ASSERT
2466 2584 }
2467 2585 #endif //PRODUCT
2468 2586
2469 2587 // number of stack slots per value argument (1 or 2)
2470 2588 int type_words = type2size[type];
2471 2589
2472 2590 // Cannot inline wide CAS on machines that don't support it natively
2473 2591 if (type2aelembytes(type) > BytesPerInt && !VM_Version::supports_cx8())
2474 2592 return false;
2475 2593
2476 2594 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2477 2595
2478 2596 // Argument words: "this" plus oop plus offset plus oldvalue plus newvalue;
2479 2597 int nargs = 1 + 1 + 2 + type_words + type_words;
2480 2598
2481 2599 // pop arguments: newval, oldval, offset, base, and receiver
2482 2600 debug_only(int saved_sp = _sp);
2483 2601 _sp += nargs;
2484 2602 Node* newval = (type_words == 1) ? pop() : pop_pair();
2485 2603 Node* oldval = (type_words == 1) ? pop() : pop_pair();
2486 2604 Node *offset = pop_pair();
2487 2605 Node *base = pop();
2488 2606 Node *receiver = pop();
2489 2607 assert(saved_sp == _sp, "must have correct argument count");
2490 2608
2491 2609 // Null check receiver.
2492 2610 _sp += nargs;
2493 2611 do_null_check(receiver, T_OBJECT);
2494 2612 _sp -= nargs;
2495 2613 if (stopped()) {
2496 2614 return true;
2497 2615 }
2498 2616
2499 2617 // Build field offset expression.
2500 2618 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2501 2619 // to be plain byte offsets, which are also the same as those accepted
2502 2620 // by oopDesc::field_base.
2503 2621 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2504 2622 // 32-bit machines ignore the high half of long offsets
2505 2623 offset = ConvL2X(offset);
2506 2624 Node* adr = make_unsafe_address(base, offset);
2507 2625 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2508 2626
2509 2627 // (Unlike inline_unsafe_access, there seems no point in trying
2510 2628 // to refine types. Just use the coarse types here.
2511 2629 const Type *value_type = Type::get_const_basic_type(type);
2512 2630 Compile::AliasType* alias_type = C->alias_type(adr_type);
2513 2631 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2514 2632 int alias_idx = C->get_alias_index(adr_type);
2515 2633
2516 2634 // Memory-model-wise, a CAS acts like a little synchronized block,
2517 2635 // so needs barriers on each side. These don't translate into
2518 2636 // actual barriers on most machines, but we still need rest of
2519 2637 // compiler to respect ordering.
2520 2638
2521 2639 insert_mem_bar(Op_MemBarRelease);
2522 2640 insert_mem_bar(Op_MemBarCPUOrder);
2523 2641
2524 2642 // 4984716: MemBars must be inserted before this
2525 2643 // memory node in order to avoid a false
2526 2644 // dependency which will confuse the scheduler.
2527 2645 Node *mem = memory(alias_idx);
2528 2646
2529 2647 // For now, we handle only those cases that actually exist: ints,
2530 2648 // longs, and Object. Adding others should be straightforward.
2531 2649 Node* cas;
|
↓ open down ↓ |
236 lines elided |
↑ open up ↑ |
2532 2650 switch(type) {
2533 2651 case T_INT:
2534 2652 cas = _gvn.transform(new (C, 5) CompareAndSwapINode(control(), mem, adr, newval, oldval));
2535 2653 break;
2536 2654 case T_LONG:
2537 2655 cas = _gvn.transform(new (C, 5) CompareAndSwapLNode(control(), mem, adr, newval, oldval));
2538 2656 break;
2539 2657 case T_OBJECT:
2540 2658 // reference stores need a store barrier.
2541 2659 // (They don't if CAS fails, but it isn't worth checking.)
2542 - pre_barrier(control(), base, adr, alias_idx, newval, value_type->make_oopptr(), T_OBJECT);
2660 + pre_barrier(true /* do_load*/,
2661 + control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
2662 + NULL /* pre_val*/,
2663 + T_OBJECT);
2543 2664 #ifdef _LP64
2544 2665 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2545 2666 Node *newval_enc = _gvn.transform(new (C, 2) EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
2546 2667 Node *oldval_enc = _gvn.transform(new (C, 2) EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
2547 2668 cas = _gvn.transform(new (C, 5) CompareAndSwapNNode(control(), mem, adr,
2548 2669 newval_enc, oldval_enc));
2549 2670 } else
2550 2671 #endif
2551 2672 {
2552 2673 cas = _gvn.transform(new (C, 5) CompareAndSwapPNode(control(), mem, adr, newval, oldval));
2553 2674 }
2554 2675 post_barrier(control(), cas, base, adr, alias_idx, newval, T_OBJECT, true);
2555 2676 break;
2556 2677 default:
2557 2678 ShouldNotReachHere();
2558 2679 break;
2559 2680 }
2560 2681
2561 2682 // SCMemProjNodes represent the memory state of CAS. Their main
2562 2683 // role is to prevent CAS nodes from being optimized away when their
2563 2684 // results aren't used.
2564 2685 Node* proj = _gvn.transform( new (C, 1) SCMemProjNode(cas));
2565 2686 set_memory(proj, alias_idx);
2566 2687
2567 2688 // Add the trailing membar surrounding the access
2568 2689 insert_mem_bar(Op_MemBarCPUOrder);
2569 2690 insert_mem_bar(Op_MemBarAcquire);
2570 2691
2571 2692 push(cas);
2572 2693 return true;
2573 2694 }
2574 2695
2575 2696 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
2576 2697 // This is another variant of inline_unsafe_access, differing in
2577 2698 // that it always issues store-store ("release") barrier and ensures
2578 2699 // store-atomicity (which only matters for "long").
2579 2700
2580 2701 if (callee()->is_static()) return false; // caller must have the capability!
2581 2702
2582 2703 #ifndef PRODUCT
2583 2704 {
2584 2705 ResourceMark rm;
2585 2706 // Check the signatures.
2586 2707 ciSignature* sig = signature();
2587 2708 #ifdef ASSERT
2588 2709 BasicType rtype = sig->return_type()->basic_type();
2589 2710 assert(rtype == T_VOID, "must return void");
2590 2711 assert(sig->count() == 3, "has 3 arguments");
2591 2712 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
2592 2713 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
2593 2714 #endif // ASSERT
2594 2715 }
2595 2716 #endif //PRODUCT
2596 2717
2597 2718 // number of stack slots per value argument (1 or 2)
2598 2719 int type_words = type2size[type];
2599 2720
2600 2721 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2601 2722
2602 2723 // Argument words: "this" plus oop plus offset plus value;
2603 2724 int nargs = 1 + 1 + 2 + type_words;
2604 2725
2605 2726 // pop arguments: val, offset, base, and receiver
2606 2727 debug_only(int saved_sp = _sp);
2607 2728 _sp += nargs;
2608 2729 Node* val = (type_words == 1) ? pop() : pop_pair();
2609 2730 Node *offset = pop_pair();
2610 2731 Node *base = pop();
2611 2732 Node *receiver = pop();
2612 2733 assert(saved_sp == _sp, "must have correct argument count");
2613 2734
2614 2735 // Null check receiver.
2615 2736 _sp += nargs;
2616 2737 do_null_check(receiver, T_OBJECT);
2617 2738 _sp -= nargs;
2618 2739 if (stopped()) {
2619 2740 return true;
2620 2741 }
2621 2742
2622 2743 // Build field offset expression.
2623 2744 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2624 2745 // 32-bit machines ignore the high half of long offsets
2625 2746 offset = ConvL2X(offset);
2626 2747 Node* adr = make_unsafe_address(base, offset);
2627 2748 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2628 2749 const Type *value_type = Type::get_const_basic_type(type);
2629 2750 Compile::AliasType* alias_type = C->alias_type(adr_type);
2630 2751
2631 2752 insert_mem_bar(Op_MemBarRelease);
2632 2753 insert_mem_bar(Op_MemBarCPUOrder);
2633 2754 // Ensure that the store is atomic for longs:
2634 2755 bool require_atomic_access = true;
2635 2756 Node* store;
2636 2757 if (type == T_OBJECT) // reference stores need a store barrier.
2637 2758 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type);
2638 2759 else {
2639 2760 store = store_to_memory(control(), adr, val, type, adr_type, require_atomic_access);
2640 2761 }
2641 2762 insert_mem_bar(Op_MemBarCPUOrder);
2642 2763 return true;
2643 2764 }
2644 2765
2645 2766 bool LibraryCallKit::inline_unsafe_allocate() {
2646 2767 if (callee()->is_static()) return false; // caller must have the capability!
2647 2768 int nargs = 1 + 1;
2648 2769 assert(signature()->size() == nargs-1, "alloc has 1 argument");
2649 2770 null_check_receiver(callee()); // check then ignore argument(0)
2650 2771 _sp += nargs; // set original stack for use by uncommon_trap
2651 2772 Node* cls = do_null_check(argument(1), T_OBJECT);
2652 2773 _sp -= nargs;
2653 2774 if (stopped()) return true;
2654 2775
2655 2776 Node* kls = load_klass_from_mirror(cls, false, nargs, NULL, 0);
2656 2777 _sp += nargs; // set original stack for use by uncommon_trap
2657 2778 kls = do_null_check(kls, T_OBJECT);
2658 2779 _sp -= nargs;
2659 2780 if (stopped()) return true; // argument was like int.class
2660 2781
2661 2782 // Note: The argument might still be an illegal value like
2662 2783 // Serializable.class or Object[].class. The runtime will handle it.
2663 2784 // But we must make an explicit check for initialization.
2664 2785 Node* insp = basic_plus_adr(kls, instanceKlass::init_state_offset_in_bytes() + sizeof(oopDesc));
2665 2786 Node* inst = make_load(NULL, insp, TypeInt::INT, T_INT);
2666 2787 Node* bits = intcon(instanceKlass::fully_initialized);
2667 2788 Node* test = _gvn.transform( new (C, 3) SubINode(inst, bits) );
2668 2789 // The 'test' is non-zero if we need to take a slow path.
2669 2790
2670 2791 Node* obj = new_instance(kls, test);
2671 2792 push(obj);
2672 2793
2673 2794 return true;
2674 2795 }
2675 2796
2676 2797 //------------------------inline_native_time_funcs--------------
2677 2798 // inline code for System.currentTimeMillis() and System.nanoTime()
2678 2799 // these have the same type and signature
2679 2800 bool LibraryCallKit::inline_native_time_funcs(bool isNano) {
2680 2801 address funcAddr = isNano ? CAST_FROM_FN_PTR(address, os::javaTimeNanos) :
2681 2802 CAST_FROM_FN_PTR(address, os::javaTimeMillis);
2682 2803 const char * funcName = isNano ? "nanoTime" : "currentTimeMillis";
2683 2804 const TypeFunc *tf = OptoRuntime::current_time_millis_Type();
2684 2805 const TypePtr* no_memory_effects = NULL;
2685 2806 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
2686 2807 Node* value = _gvn.transform(new (C, 1) ProjNode(time, TypeFunc::Parms+0));
2687 2808 #ifdef ASSERT
2688 2809 Node* value_top = _gvn.transform(new (C, 1) ProjNode(time, TypeFunc::Parms + 1));
2689 2810 assert(value_top == top(), "second value must be top");
2690 2811 #endif
2691 2812 push_pair(value);
2692 2813 return true;
2693 2814 }
2694 2815
2695 2816 //------------------------inline_native_currentThread------------------
2696 2817 bool LibraryCallKit::inline_native_currentThread() {
2697 2818 Node* junk = NULL;
2698 2819 push(generate_current_thread(junk));
2699 2820 return true;
2700 2821 }
2701 2822
2702 2823 //------------------------inline_native_isInterrupted------------------
2703 2824 bool LibraryCallKit::inline_native_isInterrupted() {
2704 2825 const int nargs = 1+1; // receiver + boolean
2705 2826 assert(nargs == arg_size(), "sanity");
2706 2827 // Add a fast path to t.isInterrupted(clear_int):
2707 2828 // (t == Thread.current() && (!TLS._osthread._interrupted || !clear_int))
2708 2829 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
2709 2830 // So, in the common case that the interrupt bit is false,
2710 2831 // we avoid making a call into the VM. Even if the interrupt bit
2711 2832 // is true, if the clear_int argument is false, we avoid the VM call.
2712 2833 // However, if the receiver is not currentThread, we must call the VM,
2713 2834 // because there must be some locking done around the operation.
2714 2835
2715 2836 // We only go to the fast case code if we pass two guards.
2716 2837 // Paths which do not pass are accumulated in the slow_region.
2717 2838 RegionNode* slow_region = new (C, 1) RegionNode(1);
2718 2839 record_for_igvn(slow_region);
2719 2840 RegionNode* result_rgn = new (C, 4) RegionNode(1+3); // fast1, fast2, slow
2720 2841 PhiNode* result_val = new (C, 4) PhiNode(result_rgn, TypeInt::BOOL);
2721 2842 enum { no_int_result_path = 1,
2722 2843 no_clear_result_path = 2,
2723 2844 slow_result_path = 3
2724 2845 };
2725 2846
2726 2847 // (a) Receiving thread must be the current thread.
2727 2848 Node* rec_thr = argument(0);
2728 2849 Node* tls_ptr = NULL;
2729 2850 Node* cur_thr = generate_current_thread(tls_ptr);
2730 2851 Node* cmp_thr = _gvn.transform( new (C, 3) CmpPNode(cur_thr, rec_thr) );
2731 2852 Node* bol_thr = _gvn.transform( new (C, 2) BoolNode(cmp_thr, BoolTest::ne) );
2732 2853
2733 2854 bool known_current_thread = (_gvn.type(bol_thr) == TypeInt::ZERO);
2734 2855 if (!known_current_thread)
2735 2856 generate_slow_guard(bol_thr, slow_region);
2736 2857
2737 2858 // (b) Interrupt bit on TLS must be false.
2738 2859 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
2739 2860 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS);
2740 2861 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
2741 2862 // Set the control input on the field _interrupted read to prevent it floating up.
2742 2863 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT);
2743 2864 Node* cmp_bit = _gvn.transform( new (C, 3) CmpINode(int_bit, intcon(0)) );
2744 2865 Node* bol_bit = _gvn.transform( new (C, 2) BoolNode(cmp_bit, BoolTest::ne) );
2745 2866
2746 2867 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
2747 2868
2748 2869 // First fast path: if (!TLS._interrupted) return false;
2749 2870 Node* false_bit = _gvn.transform( new (C, 1) IfFalseNode(iff_bit) );
2750 2871 result_rgn->init_req(no_int_result_path, false_bit);
2751 2872 result_val->init_req(no_int_result_path, intcon(0));
2752 2873
2753 2874 // drop through to next case
2754 2875 set_control( _gvn.transform(new (C, 1) IfTrueNode(iff_bit)) );
2755 2876
2756 2877 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
2757 2878 Node* clr_arg = argument(1);
2758 2879 Node* cmp_arg = _gvn.transform( new (C, 3) CmpINode(clr_arg, intcon(0)) );
2759 2880 Node* bol_arg = _gvn.transform( new (C, 2) BoolNode(cmp_arg, BoolTest::ne) );
2760 2881 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
2761 2882
2762 2883 // Second fast path: ... else if (!clear_int) return true;
2763 2884 Node* false_arg = _gvn.transform( new (C, 1) IfFalseNode(iff_arg) );
2764 2885 result_rgn->init_req(no_clear_result_path, false_arg);
2765 2886 result_val->init_req(no_clear_result_path, intcon(1));
2766 2887
2767 2888 // drop through to next case
2768 2889 set_control( _gvn.transform(new (C, 1) IfTrueNode(iff_arg)) );
2769 2890
2770 2891 // (d) Otherwise, go to the slow path.
2771 2892 slow_region->add_req(control());
2772 2893 set_control( _gvn.transform(slow_region) );
2773 2894
2774 2895 if (stopped()) {
2775 2896 // There is no slow path.
2776 2897 result_rgn->init_req(slow_result_path, top());
2777 2898 result_val->init_req(slow_result_path, top());
2778 2899 } else {
2779 2900 // non-virtual because it is a private non-static
2780 2901 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
2781 2902
2782 2903 Node* slow_val = set_results_for_java_call(slow_call);
2783 2904 // this->control() comes from set_results_for_java_call
2784 2905
2785 2906 // If we know that the result of the slow call will be true, tell the optimizer!
2786 2907 if (known_current_thread) slow_val = intcon(1);
2787 2908
2788 2909 Node* fast_io = slow_call->in(TypeFunc::I_O);
2789 2910 Node* fast_mem = slow_call->in(TypeFunc::Memory);
2790 2911 // These two phis are pre-filled with copies of of the fast IO and Memory
2791 2912 Node* io_phi = PhiNode::make(result_rgn, fast_io, Type::ABIO);
2792 2913 Node* mem_phi = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
2793 2914
2794 2915 result_rgn->init_req(slow_result_path, control());
2795 2916 io_phi ->init_req(slow_result_path, i_o());
2796 2917 mem_phi ->init_req(slow_result_path, reset_memory());
2797 2918 result_val->init_req(slow_result_path, slow_val);
2798 2919
2799 2920 set_all_memory( _gvn.transform(mem_phi) );
2800 2921 set_i_o( _gvn.transform(io_phi) );
2801 2922 }
2802 2923
2803 2924 push_result(result_rgn, result_val);
2804 2925 C->set_has_split_ifs(true); // Has chance for split-if optimization
2805 2926
2806 2927 return true;
2807 2928 }
2808 2929
2809 2930 //---------------------------load_mirror_from_klass----------------------------
2810 2931 // Given a klass oop, load its java mirror (a java.lang.Class oop).
2811 2932 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
2812 2933 Node* p = basic_plus_adr(klass, Klass::java_mirror_offset_in_bytes() + sizeof(oopDesc));
2813 2934 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT);
2814 2935 }
2815 2936
2816 2937 //-----------------------load_klass_from_mirror_common-------------------------
2817 2938 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
2818 2939 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
2819 2940 // and branch to the given path on the region.
2820 2941 // If never_see_null, take an uncommon trap on null, so we can optimistically
2821 2942 // compile for the non-null case.
2822 2943 // If the region is NULL, force never_see_null = true.
2823 2944 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
2824 2945 bool never_see_null,
2825 2946 int nargs,
2826 2947 RegionNode* region,
2827 2948 int null_path,
2828 2949 int offset) {
2829 2950 if (region == NULL) never_see_null = true;
2830 2951 Node* p = basic_plus_adr(mirror, offset);
2831 2952 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
2832 2953 Node* kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type) );
2833 2954 _sp += nargs; // any deopt will start just before call to enclosing method
2834 2955 Node* null_ctl = top();
2835 2956 kls = null_check_oop(kls, &null_ctl, never_see_null);
2836 2957 if (region != NULL) {
2837 2958 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
2838 2959 region->init_req(null_path, null_ctl);
2839 2960 } else {
2840 2961 assert(null_ctl == top(), "no loose ends");
2841 2962 }
2842 2963 _sp -= nargs;
2843 2964 return kls;
2844 2965 }
2845 2966
2846 2967 //--------------------(inline_native_Class_query helpers)---------------------
2847 2968 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
2848 2969 // Fall through if (mods & mask) == bits, take the guard otherwise.
2849 2970 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
2850 2971 // Branch around if the given klass has the given modifier bit set.
2851 2972 // Like generate_guard, adds a new path onto the region.
2852 2973 Node* modp = basic_plus_adr(kls, Klass::access_flags_offset_in_bytes() + sizeof(oopDesc));
2853 2974 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT);
2854 2975 Node* mask = intcon(modifier_mask);
2855 2976 Node* bits = intcon(modifier_bits);
2856 2977 Node* mbit = _gvn.transform( new (C, 3) AndINode(mods, mask) );
2857 2978 Node* cmp = _gvn.transform( new (C, 3) CmpINode(mbit, bits) );
2858 2979 Node* bol = _gvn.transform( new (C, 2) BoolNode(cmp, BoolTest::ne) );
2859 2980 return generate_fair_guard(bol, region);
2860 2981 }
2861 2982 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
2862 2983 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
2863 2984 }
2864 2985
2865 2986 //-------------------------inline_native_Class_query-------------------
2866 2987 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
2867 2988 int nargs = 1+0; // just the Class mirror, in most cases
2868 2989 const Type* return_type = TypeInt::BOOL;
2869 2990 Node* prim_return_value = top(); // what happens if it's a primitive class?
2870 2991 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
2871 2992 bool expect_prim = false; // most of these guys expect to work on refs
2872 2993
2873 2994 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
2874 2995
2875 2996 switch (id) {
2876 2997 case vmIntrinsics::_isInstance:
2877 2998 nargs = 1+1; // the Class mirror, plus the object getting queried about
2878 2999 // nothing is an instance of a primitive type
2879 3000 prim_return_value = intcon(0);
2880 3001 break;
2881 3002 case vmIntrinsics::_getModifiers:
2882 3003 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
2883 3004 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
2884 3005 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
2885 3006 break;
2886 3007 case vmIntrinsics::_isInterface:
2887 3008 prim_return_value = intcon(0);
2888 3009 break;
2889 3010 case vmIntrinsics::_isArray:
2890 3011 prim_return_value = intcon(0);
2891 3012 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
2892 3013 break;
2893 3014 case vmIntrinsics::_isPrimitive:
2894 3015 prim_return_value = intcon(1);
2895 3016 expect_prim = true; // obviously
2896 3017 break;
2897 3018 case vmIntrinsics::_getSuperclass:
2898 3019 prim_return_value = null();
2899 3020 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
2900 3021 break;
2901 3022 case vmIntrinsics::_getComponentType:
2902 3023 prim_return_value = null();
2903 3024 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
2904 3025 break;
2905 3026 case vmIntrinsics::_getClassAccessFlags:
2906 3027 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
2907 3028 return_type = TypeInt::INT; // not bool! 6297094
2908 3029 break;
2909 3030 default:
2910 3031 ShouldNotReachHere();
2911 3032 }
2912 3033
2913 3034 Node* mirror = argument(0);
2914 3035 Node* obj = (nargs <= 1)? top(): argument(1);
2915 3036
2916 3037 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
2917 3038 if (mirror_con == NULL) return false; // cannot happen?
2918 3039
2919 3040 #ifndef PRODUCT
2920 3041 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) {
2921 3042 ciType* k = mirror_con->java_mirror_type();
2922 3043 if (k) {
2923 3044 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
2924 3045 k->print_name();
2925 3046 tty->cr();
2926 3047 }
2927 3048 }
2928 3049 #endif
2929 3050
2930 3051 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
2931 3052 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT);
2932 3053 record_for_igvn(region);
2933 3054 PhiNode* phi = new (C, PATH_LIMIT) PhiNode(region, return_type);
2934 3055
2935 3056 // The mirror will never be null of Reflection.getClassAccessFlags, however
2936 3057 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
2937 3058 // if it is. See bug 4774291.
2938 3059
2939 3060 // For Reflection.getClassAccessFlags(), the null check occurs in
2940 3061 // the wrong place; see inline_unsafe_access(), above, for a similar
2941 3062 // situation.
2942 3063 _sp += nargs; // set original stack for use by uncommon_trap
2943 3064 mirror = do_null_check(mirror, T_OBJECT);
2944 3065 _sp -= nargs;
2945 3066 // If mirror or obj is dead, only null-path is taken.
2946 3067 if (stopped()) return true;
2947 3068
2948 3069 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
2949 3070
2950 3071 // Now load the mirror's klass metaobject, and null-check it.
2951 3072 // Side-effects region with the control path if the klass is null.
2952 3073 Node* kls = load_klass_from_mirror(mirror, never_see_null, nargs,
2953 3074 region, _prim_path);
2954 3075 // If kls is null, we have a primitive mirror.
2955 3076 phi->init_req(_prim_path, prim_return_value);
2956 3077 if (stopped()) { push_result(region, phi); return true; }
2957 3078
2958 3079 Node* p; // handy temp
2959 3080 Node* null_ctl;
2960 3081
2961 3082 // Now that we have the non-null klass, we can perform the real query.
2962 3083 // For constant classes, the query will constant-fold in LoadNode::Value.
2963 3084 Node* query_value = top();
2964 3085 switch (id) {
2965 3086 case vmIntrinsics::_isInstance:
2966 3087 // nothing is an instance of a primitive type
2967 3088 _sp += nargs; // gen_instanceof might do an uncommon trap
2968 3089 query_value = gen_instanceof(obj, kls);
2969 3090 _sp -= nargs;
2970 3091 break;
2971 3092
2972 3093 case vmIntrinsics::_getModifiers:
2973 3094 p = basic_plus_adr(kls, Klass::modifier_flags_offset_in_bytes() + sizeof(oopDesc));
2974 3095 query_value = make_load(NULL, p, TypeInt::INT, T_INT);
2975 3096 break;
2976 3097
2977 3098 case vmIntrinsics::_isInterface:
2978 3099 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
2979 3100 if (generate_interface_guard(kls, region) != NULL)
2980 3101 // A guard was added. If the guard is taken, it was an interface.
2981 3102 phi->add_req(intcon(1));
2982 3103 // If we fall through, it's a plain class.
2983 3104 query_value = intcon(0);
2984 3105 break;
2985 3106
2986 3107 case vmIntrinsics::_isArray:
2987 3108 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
2988 3109 if (generate_array_guard(kls, region) != NULL)
2989 3110 // A guard was added. If the guard is taken, it was an array.
2990 3111 phi->add_req(intcon(1));
2991 3112 // If we fall through, it's a plain class.
2992 3113 query_value = intcon(0);
2993 3114 break;
2994 3115
2995 3116 case vmIntrinsics::_isPrimitive:
2996 3117 query_value = intcon(0); // "normal" path produces false
2997 3118 break;
2998 3119
2999 3120 case vmIntrinsics::_getSuperclass:
3000 3121 // The rules here are somewhat unfortunate, but we can still do better
3001 3122 // with random logic than with a JNI call.
3002 3123 // Interfaces store null or Object as _super, but must report null.
3003 3124 // Arrays store an intermediate super as _super, but must report Object.
3004 3125 // Other types can report the actual _super.
3005 3126 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3006 3127 if (generate_interface_guard(kls, region) != NULL)
3007 3128 // A guard was added. If the guard is taken, it was an interface.
3008 3129 phi->add_req(null());
3009 3130 if (generate_array_guard(kls, region) != NULL)
3010 3131 // A guard was added. If the guard is taken, it was an array.
3011 3132 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3012 3133 // If we fall through, it's a plain class. Get its _super.
3013 3134 p = basic_plus_adr(kls, Klass::super_offset_in_bytes() + sizeof(oopDesc));
3014 3135 kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL) );
3015 3136 null_ctl = top();
3016 3137 kls = null_check_oop(kls, &null_ctl);
3017 3138 if (null_ctl != top()) {
3018 3139 // If the guard is taken, Object.superClass is null (both klass and mirror).
3019 3140 region->add_req(null_ctl);
3020 3141 phi ->add_req(null());
3021 3142 }
3022 3143 if (!stopped()) {
3023 3144 query_value = load_mirror_from_klass(kls);
3024 3145 }
3025 3146 break;
3026 3147
3027 3148 case vmIntrinsics::_getComponentType:
3028 3149 if (generate_array_guard(kls, region) != NULL) {
3029 3150 // Be sure to pin the oop load to the guard edge just created:
3030 3151 Node* is_array_ctrl = region->in(region->req()-1);
3031 3152 Node* cma = basic_plus_adr(kls, in_bytes(arrayKlass::component_mirror_offset()) + sizeof(oopDesc));
3032 3153 Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT);
3033 3154 phi->add_req(cmo);
3034 3155 }
3035 3156 query_value = null(); // non-array case is null
3036 3157 break;
3037 3158
3038 3159 case vmIntrinsics::_getClassAccessFlags:
3039 3160 p = basic_plus_adr(kls, Klass::access_flags_offset_in_bytes() + sizeof(oopDesc));
3040 3161 query_value = make_load(NULL, p, TypeInt::INT, T_INT);
3041 3162 break;
3042 3163
3043 3164 default:
3044 3165 ShouldNotReachHere();
3045 3166 }
3046 3167
3047 3168 // Fall-through is the normal case of a query to a real class.
3048 3169 phi->init_req(1, query_value);
3049 3170 region->init_req(1, control());
3050 3171
3051 3172 push_result(region, phi);
3052 3173 C->set_has_split_ifs(true); // Has chance for split-if optimization
3053 3174
3054 3175 return true;
3055 3176 }
3056 3177
3057 3178 //--------------------------inline_native_subtype_check------------------------
3058 3179 // This intrinsic takes the JNI calls out of the heart of
3059 3180 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3060 3181 bool LibraryCallKit::inline_native_subtype_check() {
3061 3182 int nargs = 1+1; // the Class mirror, plus the other class getting examined
3062 3183
3063 3184 // Pull both arguments off the stack.
3064 3185 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3065 3186 args[0] = argument(0);
3066 3187 args[1] = argument(1);
3067 3188 Node* klasses[2]; // corresponding Klasses: superk, subk
3068 3189 klasses[0] = klasses[1] = top();
3069 3190
3070 3191 enum {
3071 3192 // A full decision tree on {superc is prim, subc is prim}:
3072 3193 _prim_0_path = 1, // {P,N} => false
3073 3194 // {P,P} & superc!=subc => false
3074 3195 _prim_same_path, // {P,P} & superc==subc => true
3075 3196 _prim_1_path, // {N,P} => false
3076 3197 _ref_subtype_path, // {N,N} & subtype check wins => true
3077 3198 _both_ref_path, // {N,N} & subtype check loses => false
3078 3199 PATH_LIMIT
3079 3200 };
3080 3201
3081 3202 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT);
3082 3203 Node* phi = new (C, PATH_LIMIT) PhiNode(region, TypeInt::BOOL);
3083 3204 record_for_igvn(region);
3084 3205
3085 3206 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3086 3207 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3087 3208 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3088 3209
3089 3210 // First null-check both mirrors and load each mirror's klass metaobject.
3090 3211 int which_arg;
3091 3212 for (which_arg = 0; which_arg <= 1; which_arg++) {
3092 3213 Node* arg = args[which_arg];
3093 3214 _sp += nargs; // set original stack for use by uncommon_trap
3094 3215 arg = do_null_check(arg, T_OBJECT);
3095 3216 _sp -= nargs;
3096 3217 if (stopped()) break;
3097 3218 args[which_arg] = _gvn.transform(arg);
3098 3219
3099 3220 Node* p = basic_plus_adr(arg, class_klass_offset);
3100 3221 Node* kls = LoadKlassNode::make(_gvn, immutable_memory(), p, adr_type, kls_type);
3101 3222 klasses[which_arg] = _gvn.transform(kls);
3102 3223 }
3103 3224
3104 3225 // Having loaded both klasses, test each for null.
3105 3226 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3106 3227 for (which_arg = 0; which_arg <= 1; which_arg++) {
3107 3228 Node* kls = klasses[which_arg];
3108 3229 Node* null_ctl = top();
3109 3230 _sp += nargs; // set original stack for use by uncommon_trap
3110 3231 kls = null_check_oop(kls, &null_ctl, never_see_null);
3111 3232 _sp -= nargs;
3112 3233 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3113 3234 region->init_req(prim_path, null_ctl);
3114 3235 if (stopped()) break;
3115 3236 klasses[which_arg] = kls;
3116 3237 }
3117 3238
3118 3239 if (!stopped()) {
3119 3240 // now we have two reference types, in klasses[0..1]
3120 3241 Node* subk = klasses[1]; // the argument to isAssignableFrom
3121 3242 Node* superk = klasses[0]; // the receiver
3122 3243 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3123 3244 // now we have a successful reference subtype check
3124 3245 region->set_req(_ref_subtype_path, control());
3125 3246 }
3126 3247
3127 3248 // If both operands are primitive (both klasses null), then
3128 3249 // we must return true when they are identical primitives.
3129 3250 // It is convenient to test this after the first null klass check.
3130 3251 set_control(region->in(_prim_0_path)); // go back to first null check
3131 3252 if (!stopped()) {
3132 3253 // Since superc is primitive, make a guard for the superc==subc case.
3133 3254 Node* cmp_eq = _gvn.transform( new (C, 3) CmpPNode(args[0], args[1]) );
3134 3255 Node* bol_eq = _gvn.transform( new (C, 2) BoolNode(cmp_eq, BoolTest::eq) );
3135 3256 generate_guard(bol_eq, region, PROB_FAIR);
3136 3257 if (region->req() == PATH_LIMIT+1) {
3137 3258 // A guard was added. If the added guard is taken, superc==subc.
3138 3259 region->swap_edges(PATH_LIMIT, _prim_same_path);
3139 3260 region->del_req(PATH_LIMIT);
3140 3261 }
3141 3262 region->set_req(_prim_0_path, control()); // Not equal after all.
3142 3263 }
3143 3264
3144 3265 // these are the only paths that produce 'true':
3145 3266 phi->set_req(_prim_same_path, intcon(1));
3146 3267 phi->set_req(_ref_subtype_path, intcon(1));
3147 3268
3148 3269 // pull together the cases:
3149 3270 assert(region->req() == PATH_LIMIT, "sane region");
3150 3271 for (uint i = 1; i < region->req(); i++) {
3151 3272 Node* ctl = region->in(i);
3152 3273 if (ctl == NULL || ctl == top()) {
3153 3274 region->set_req(i, top());
3154 3275 phi ->set_req(i, top());
3155 3276 } else if (phi->in(i) == NULL) {
3156 3277 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3157 3278 }
3158 3279 }
3159 3280
3160 3281 set_control(_gvn.transform(region));
3161 3282 push(_gvn.transform(phi));
3162 3283
3163 3284 return true;
3164 3285 }
3165 3286
3166 3287 //---------------------generate_array_guard_common------------------------
3167 3288 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3168 3289 bool obj_array, bool not_array) {
3169 3290 // If obj_array/non_array==false/false:
3170 3291 // Branch around if the given klass is in fact an array (either obj or prim).
3171 3292 // If obj_array/non_array==false/true:
3172 3293 // Branch around if the given klass is not an array klass of any kind.
3173 3294 // If obj_array/non_array==true/true:
3174 3295 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3175 3296 // If obj_array/non_array==true/false:
3176 3297 // Branch around if the kls is an oop array (Object[] or subtype)
3177 3298 //
3178 3299 // Like generate_guard, adds a new path onto the region.
3179 3300 jint layout_con = 0;
3180 3301 Node* layout_val = get_layout_helper(kls, layout_con);
3181 3302 if (layout_val == NULL) {
3182 3303 bool query = (obj_array
3183 3304 ? Klass::layout_helper_is_objArray(layout_con)
3184 3305 : Klass::layout_helper_is_javaArray(layout_con));
3185 3306 if (query == not_array) {
3186 3307 return NULL; // never a branch
3187 3308 } else { // always a branch
3188 3309 Node* always_branch = control();
3189 3310 if (region != NULL)
3190 3311 region->add_req(always_branch);
3191 3312 set_control(top());
3192 3313 return always_branch;
3193 3314 }
3194 3315 }
3195 3316 // Now test the correct condition.
3196 3317 jint nval = (obj_array
3197 3318 ? ((jint)Klass::_lh_array_tag_type_value
3198 3319 << Klass::_lh_array_tag_shift)
3199 3320 : Klass::_lh_neutral_value);
3200 3321 Node* cmp = _gvn.transform( new(C, 3) CmpINode(layout_val, intcon(nval)) );
3201 3322 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3202 3323 // invert the test if we are looking for a non-array
3203 3324 if (not_array) btest = BoolTest(btest).negate();
3204 3325 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, btest) );
3205 3326 return generate_fair_guard(bol, region);
3206 3327 }
3207 3328
3208 3329
3209 3330 //-----------------------inline_native_newArray--------------------------
3210 3331 bool LibraryCallKit::inline_native_newArray() {
3211 3332 int nargs = 2;
3212 3333 Node* mirror = argument(0);
3213 3334 Node* count_val = argument(1);
3214 3335
3215 3336 _sp += nargs; // set original stack for use by uncommon_trap
3216 3337 mirror = do_null_check(mirror, T_OBJECT);
3217 3338 _sp -= nargs;
3218 3339 // If mirror or obj is dead, only null-path is taken.
3219 3340 if (stopped()) return true;
3220 3341
3221 3342 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3222 3343 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
3223 3344 PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg,
3224 3345 TypeInstPtr::NOTNULL);
3225 3346 PhiNode* result_io = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO);
3226 3347 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY,
3227 3348 TypePtr::BOTTOM);
3228 3349
3229 3350 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3230 3351 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3231 3352 nargs,
3232 3353 result_reg, _slow_path);
3233 3354 Node* normal_ctl = control();
3234 3355 Node* no_array_ctl = result_reg->in(_slow_path);
3235 3356
3236 3357 // Generate code for the slow case. We make a call to newArray().
3237 3358 set_control(no_array_ctl);
3238 3359 if (!stopped()) {
3239 3360 // Either the input type is void.class, or else the
3240 3361 // array klass has not yet been cached. Either the
3241 3362 // ensuing call will throw an exception, or else it
3242 3363 // will cache the array klass for next time.
3243 3364 PreserveJVMState pjvms(this);
3244 3365 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3245 3366 Node* slow_result = set_results_for_java_call(slow_call);
3246 3367 // this->control() comes from set_results_for_java_call
3247 3368 result_reg->set_req(_slow_path, control());
3248 3369 result_val->set_req(_slow_path, slow_result);
3249 3370 result_io ->set_req(_slow_path, i_o());
3250 3371 result_mem->set_req(_slow_path, reset_memory());
3251 3372 }
3252 3373
3253 3374 set_control(normal_ctl);
3254 3375 if (!stopped()) {
3255 3376 // Normal case: The array type has been cached in the java.lang.Class.
3256 3377 // The following call works fine even if the array type is polymorphic.
3257 3378 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3258 3379 Node* obj = new_array(klass_node, count_val, nargs);
3259 3380 result_reg->init_req(_normal_path, control());
3260 3381 result_val->init_req(_normal_path, obj);
3261 3382 result_io ->init_req(_normal_path, i_o());
3262 3383 result_mem->init_req(_normal_path, reset_memory());
3263 3384 }
3264 3385
3265 3386 // Return the combined state.
3266 3387 set_i_o( _gvn.transform(result_io) );
3267 3388 set_all_memory( _gvn.transform(result_mem) );
3268 3389 push_result(result_reg, result_val);
3269 3390 C->set_has_split_ifs(true); // Has chance for split-if optimization
3270 3391
3271 3392 return true;
3272 3393 }
3273 3394
3274 3395 //----------------------inline_native_getLength--------------------------
3275 3396 bool LibraryCallKit::inline_native_getLength() {
3276 3397 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3277 3398
3278 3399 int nargs = 1;
3279 3400 Node* array = argument(0);
3280 3401
3281 3402 _sp += nargs; // set original stack for use by uncommon_trap
3282 3403 array = do_null_check(array, T_OBJECT);
3283 3404 _sp -= nargs;
3284 3405
3285 3406 // If array is dead, only null-path is taken.
3286 3407 if (stopped()) return true;
3287 3408
3288 3409 // Deoptimize if it is a non-array.
3289 3410 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3290 3411
3291 3412 if (non_array != NULL) {
3292 3413 PreserveJVMState pjvms(this);
3293 3414 set_control(non_array);
3294 3415 _sp += nargs; // push the arguments back on the stack
3295 3416 uncommon_trap(Deoptimization::Reason_intrinsic,
3296 3417 Deoptimization::Action_maybe_recompile);
3297 3418 }
3298 3419
3299 3420 // If control is dead, only non-array-path is taken.
3300 3421 if (stopped()) return true;
3301 3422
3302 3423 // The works fine even if the array type is polymorphic.
3303 3424 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3304 3425 push( load_array_length(array) );
3305 3426
3306 3427 C->set_has_split_ifs(true); // Has chance for split-if optimization
3307 3428
3308 3429 return true;
3309 3430 }
3310 3431
3311 3432 //------------------------inline_array_copyOf----------------------------
3312 3433 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3313 3434 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3314 3435
3315 3436 // Restore the stack and pop off the arguments.
3316 3437 int nargs = 3 + (is_copyOfRange? 1: 0);
3317 3438 Node* original = argument(0);
3318 3439 Node* start = is_copyOfRange? argument(1): intcon(0);
3319 3440 Node* end = is_copyOfRange? argument(2): argument(1);
3320 3441 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3321 3442
3322 3443 Node* newcopy;
3323 3444
3324 3445 //set the original stack and the reexecute bit for the interpreter to reexecute
3325 3446 //the bytecode that invokes Arrays.copyOf if deoptimization happens
3326 3447 { PreserveReexecuteState preexecs(this);
3327 3448 _sp += nargs;
3328 3449 jvms()->set_should_reexecute(true);
3329 3450
3330 3451 array_type_mirror = do_null_check(array_type_mirror, T_OBJECT);
3331 3452 original = do_null_check(original, T_OBJECT);
3332 3453
3333 3454 // Check if a null path was taken unconditionally.
3334 3455 if (stopped()) return true;
3335 3456
3336 3457 Node* orig_length = load_array_length(original);
3337 3458
3338 3459 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, 0,
3339 3460 NULL, 0);
3340 3461 klass_node = do_null_check(klass_node, T_OBJECT);
3341 3462
3342 3463 RegionNode* bailout = new (C, 1) RegionNode(1);
3343 3464 record_for_igvn(bailout);
3344 3465
3345 3466 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3346 3467 // Bail out if that is so.
3347 3468 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3348 3469 if (not_objArray != NULL) {
3349 3470 // Improve the klass node's type from the new optimistic assumption:
3350 3471 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3351 3472 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3352 3473 Node* cast = new (C, 2) CastPPNode(klass_node, akls);
3353 3474 cast->init_req(0, control());
3354 3475 klass_node = _gvn.transform(cast);
3355 3476 }
3356 3477
3357 3478 // Bail out if either start or end is negative.
3358 3479 generate_negative_guard(start, bailout, &start);
3359 3480 generate_negative_guard(end, bailout, &end);
3360 3481
3361 3482 Node* length = end;
3362 3483 if (_gvn.type(start) != TypeInt::ZERO) {
3363 3484 length = _gvn.transform( new (C, 3) SubINode(end, start) );
3364 3485 }
3365 3486
3366 3487 // Bail out if length is negative.
3367 3488 // ...Not needed, since the new_array will throw the right exception.
3368 3489 //generate_negative_guard(length, bailout, &length);
3369 3490
3370 3491 if (bailout->req() > 1) {
3371 3492 PreserveJVMState pjvms(this);
3372 3493 set_control( _gvn.transform(bailout) );
3373 3494 uncommon_trap(Deoptimization::Reason_intrinsic,
3374 3495 Deoptimization::Action_maybe_recompile);
3375 3496 }
3376 3497
3377 3498 if (!stopped()) {
3378 3499
3379 3500 // How many elements will we copy from the original?
3380 3501 // The answer is MinI(orig_length - start, length).
3381 3502 Node* orig_tail = _gvn.transform( new(C, 3) SubINode(orig_length, start) );
3382 3503 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3383 3504
3384 3505 const bool raw_mem_only = true;
3385 3506 newcopy = new_array(klass_node, length, 0, raw_mem_only);
3386 3507
3387 3508 // Generate a direct call to the right arraycopy function(s).
3388 3509 // We know the copy is disjoint but we might not know if the
3389 3510 // oop stores need checking.
3390 3511 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
3391 3512 // This will fail a store-check if x contains any non-nulls.
3392 3513 bool disjoint_bases = true;
3393 3514 bool length_never_negative = true;
3394 3515 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
3395 3516 original, start, newcopy, intcon(0), moved,
3396 3517 disjoint_bases, length_never_negative);
3397 3518 }
3398 3519 } //original reexecute and sp are set back here
3399 3520
3400 3521 if(!stopped()) {
3401 3522 push(newcopy);
3402 3523 }
3403 3524
3404 3525 C->set_has_split_ifs(true); // Has chance for split-if optimization
3405 3526
3406 3527 return true;
3407 3528 }
3408 3529
3409 3530
3410 3531 //----------------------generate_virtual_guard---------------------------
3411 3532 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
3412 3533 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
3413 3534 RegionNode* slow_region) {
3414 3535 ciMethod* method = callee();
3415 3536 int vtable_index = method->vtable_index();
3416 3537 // Get the methodOop out of the appropriate vtable entry.
3417 3538 int entry_offset = (instanceKlass::vtable_start_offset() +
3418 3539 vtable_index*vtableEntry::size()) * wordSize +
3419 3540 vtableEntry::method_offset_in_bytes();
3420 3541 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
3421 3542 Node* target_call = make_load(NULL, entry_addr, TypeInstPtr::NOTNULL, T_OBJECT);
3422 3543
3423 3544 // Compare the target method with the expected method (e.g., Object.hashCode).
3424 3545 const TypeInstPtr* native_call_addr = TypeInstPtr::make(method);
3425 3546
3426 3547 Node* native_call = makecon(native_call_addr);
3427 3548 Node* chk_native = _gvn.transform( new(C, 3) CmpPNode(target_call, native_call) );
3428 3549 Node* test_native = _gvn.transform( new(C, 2) BoolNode(chk_native, BoolTest::ne) );
3429 3550
3430 3551 return generate_slow_guard(test_native, slow_region);
3431 3552 }
3432 3553
3433 3554 //-----------------------generate_method_call----------------------------
3434 3555 // Use generate_method_call to make a slow-call to the real
3435 3556 // method if the fast path fails. An alternative would be to
3436 3557 // use a stub like OptoRuntime::slow_arraycopy_Java.
3437 3558 // This only works for expanding the current library call,
3438 3559 // not another intrinsic. (E.g., don't use this for making an
3439 3560 // arraycopy call inside of the copyOf intrinsic.)
3440 3561 CallJavaNode*
3441 3562 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
3442 3563 // When compiling the intrinsic method itself, do not use this technique.
3443 3564 guarantee(callee() != C->method(), "cannot make slow-call to self");
3444 3565
3445 3566 ciMethod* method = callee();
3446 3567 // ensure the JVMS we have will be correct for this call
3447 3568 guarantee(method_id == method->intrinsic_id(), "must match");
3448 3569
3449 3570 const TypeFunc* tf = TypeFunc::make(method);
3450 3571 int tfdc = tf->domain()->cnt();
3451 3572 CallJavaNode* slow_call;
3452 3573 if (is_static) {
3453 3574 assert(!is_virtual, "");
3454 3575 slow_call = new(C, tfdc) CallStaticJavaNode(tf,
3455 3576 SharedRuntime::get_resolve_static_call_stub(),
3456 3577 method, bci());
3457 3578 } else if (is_virtual) {
3458 3579 null_check_receiver(method);
3459 3580 int vtable_index = methodOopDesc::invalid_vtable_index;
3460 3581 if (UseInlineCaches) {
3461 3582 // Suppress the vtable call
3462 3583 } else {
3463 3584 // hashCode and clone are not a miranda methods,
3464 3585 // so the vtable index is fixed.
3465 3586 // No need to use the linkResolver to get it.
3466 3587 vtable_index = method->vtable_index();
3467 3588 }
3468 3589 slow_call = new(C, tfdc) CallDynamicJavaNode(tf,
3469 3590 SharedRuntime::get_resolve_virtual_call_stub(),
3470 3591 method, vtable_index, bci());
3471 3592 } else { // neither virtual nor static: opt_virtual
3472 3593 null_check_receiver(method);
3473 3594 slow_call = new(C, tfdc) CallStaticJavaNode(tf,
3474 3595 SharedRuntime::get_resolve_opt_virtual_call_stub(),
3475 3596 method, bci());
3476 3597 slow_call->set_optimized_virtual(true);
3477 3598 }
3478 3599 set_arguments_for_java_call(slow_call);
3479 3600 set_edges_for_java_call(slow_call);
3480 3601 return slow_call;
3481 3602 }
3482 3603
3483 3604
3484 3605 //------------------------------inline_native_hashcode--------------------
3485 3606 // Build special case code for calls to hashCode on an object.
3486 3607 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
3487 3608 assert(is_static == callee()->is_static(), "correct intrinsic selection");
3488 3609 assert(!(is_virtual && is_static), "either virtual, special, or static");
3489 3610
3490 3611 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
3491 3612
3492 3613 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
3493 3614 PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg,
3494 3615 TypeInt::INT);
3495 3616 PhiNode* result_io = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO);
3496 3617 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY,
3497 3618 TypePtr::BOTTOM);
3498 3619 Node* obj = NULL;
3499 3620 if (!is_static) {
3500 3621 // Check for hashing null object
3501 3622 obj = null_check_receiver(callee());
3502 3623 if (stopped()) return true; // unconditionally null
3503 3624 result_reg->init_req(_null_path, top());
3504 3625 result_val->init_req(_null_path, top());
3505 3626 } else {
3506 3627 // Do a null check, and return zero if null.
3507 3628 // System.identityHashCode(null) == 0
3508 3629 obj = argument(0);
3509 3630 Node* null_ctl = top();
3510 3631 obj = null_check_oop(obj, &null_ctl);
3511 3632 result_reg->init_req(_null_path, null_ctl);
3512 3633 result_val->init_req(_null_path, _gvn.intcon(0));
3513 3634 }
3514 3635
3515 3636 // Unconditionally null? Then return right away.
3516 3637 if (stopped()) {
3517 3638 set_control( result_reg->in(_null_path) );
3518 3639 if (!stopped())
3519 3640 push( result_val ->in(_null_path) );
3520 3641 return true;
3521 3642 }
3522 3643
3523 3644 // After null check, get the object's klass.
3524 3645 Node* obj_klass = load_object_klass(obj);
3525 3646
3526 3647 // This call may be virtual (invokevirtual) or bound (invokespecial).
3527 3648 // For each case we generate slightly different code.
3528 3649
3529 3650 // We only go to the fast case code if we pass a number of guards. The
3530 3651 // paths which do not pass are accumulated in the slow_region.
3531 3652 RegionNode* slow_region = new (C, 1) RegionNode(1);
3532 3653 record_for_igvn(slow_region);
3533 3654
3534 3655 // If this is a virtual call, we generate a funny guard. We pull out
3535 3656 // the vtable entry corresponding to hashCode() from the target object.
3536 3657 // If the target method which we are calling happens to be the native
3537 3658 // Object hashCode() method, we pass the guard. We do not need this
3538 3659 // guard for non-virtual calls -- the caller is known to be the native
3539 3660 // Object hashCode().
3540 3661 if (is_virtual) {
3541 3662 generate_virtual_guard(obj_klass, slow_region);
3542 3663 }
3543 3664
3544 3665 // Get the header out of the object, use LoadMarkNode when available
3545 3666 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
3546 3667 Node* header = make_load(control(), header_addr, TypeX_X, TypeX_X->basic_type());
3547 3668
3548 3669 // Test the header to see if it is unlocked.
3549 3670 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
3550 3671 Node *lmasked_header = _gvn.transform( new (C, 3) AndXNode(header, lock_mask) );
3551 3672 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
3552 3673 Node *chk_unlocked = _gvn.transform( new (C, 3) CmpXNode( lmasked_header, unlocked_val));
3553 3674 Node *test_unlocked = _gvn.transform( new (C, 2) BoolNode( chk_unlocked, BoolTest::ne) );
3554 3675
3555 3676 generate_slow_guard(test_unlocked, slow_region);
3556 3677
3557 3678 // Get the hash value and check to see that it has been properly assigned.
3558 3679 // We depend on hash_mask being at most 32 bits and avoid the use of
3559 3680 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
3560 3681 // vm: see markOop.hpp.
3561 3682 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
3562 3683 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
3563 3684 Node *hshifted_header= _gvn.transform( new (C, 3) URShiftXNode(header, hash_shift) );
3564 3685 // This hack lets the hash bits live anywhere in the mark object now, as long
3565 3686 // as the shift drops the relevant bits into the low 32 bits. Note that
3566 3687 // Java spec says that HashCode is an int so there's no point in capturing
3567 3688 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
3568 3689 hshifted_header = ConvX2I(hshifted_header);
3569 3690 Node *hash_val = _gvn.transform( new (C, 3) AndINode(hshifted_header, hash_mask) );
3570 3691
3571 3692 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
3572 3693 Node *chk_assigned = _gvn.transform( new (C, 3) CmpINode( hash_val, no_hash_val));
3573 3694 Node *test_assigned = _gvn.transform( new (C, 2) BoolNode( chk_assigned, BoolTest::eq) );
3574 3695
3575 3696 generate_slow_guard(test_assigned, slow_region);
3576 3697
3577 3698 Node* init_mem = reset_memory();
3578 3699 // fill in the rest of the null path:
3579 3700 result_io ->init_req(_null_path, i_o());
3580 3701 result_mem->init_req(_null_path, init_mem);
3581 3702
3582 3703 result_val->init_req(_fast_path, hash_val);
3583 3704 result_reg->init_req(_fast_path, control());
3584 3705 result_io ->init_req(_fast_path, i_o());
3585 3706 result_mem->init_req(_fast_path, init_mem);
3586 3707
3587 3708 // Generate code for the slow case. We make a call to hashCode().
3588 3709 set_control(_gvn.transform(slow_region));
3589 3710 if (!stopped()) {
3590 3711 // No need for PreserveJVMState, because we're using up the present state.
3591 3712 set_all_memory(init_mem);
3592 3713 vmIntrinsics::ID hashCode_id = vmIntrinsics::_hashCode;
3593 3714 if (is_static) hashCode_id = vmIntrinsics::_identityHashCode;
3594 3715 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
3595 3716 Node* slow_result = set_results_for_java_call(slow_call);
3596 3717 // this->control() comes from set_results_for_java_call
3597 3718 result_reg->init_req(_slow_path, control());
3598 3719 result_val->init_req(_slow_path, slow_result);
3599 3720 result_io ->set_req(_slow_path, i_o());
3600 3721 result_mem ->set_req(_slow_path, reset_memory());
3601 3722 }
3602 3723
3603 3724 // Return the combined state.
3604 3725 set_i_o( _gvn.transform(result_io) );
3605 3726 set_all_memory( _gvn.transform(result_mem) );
3606 3727 push_result(result_reg, result_val);
3607 3728
3608 3729 return true;
3609 3730 }
3610 3731
3611 3732 //---------------------------inline_native_getClass----------------------------
3612 3733 // Build special case code for calls to getClass on an object.
3613 3734 bool LibraryCallKit::inline_native_getClass() {
3614 3735 Node* obj = null_check_receiver(callee());
3615 3736 if (stopped()) return true;
3616 3737 push( load_mirror_from_klass(load_object_klass(obj)) );
3617 3738 return true;
3618 3739 }
3619 3740
3620 3741 //-----------------inline_native_Reflection_getCallerClass---------------------
3621 3742 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
3622 3743 //
3623 3744 // NOTE that this code must perform the same logic as
3624 3745 // vframeStream::security_get_caller_frame in that it must skip
3625 3746 // Method.invoke() and auxiliary frames.
3626 3747
3627 3748
3628 3749
3629 3750
3630 3751 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
3631 3752 ciMethod* method = callee();
3632 3753
3633 3754 #ifndef PRODUCT
3634 3755 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3635 3756 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
3636 3757 }
3637 3758 #endif
3638 3759
3639 3760 debug_only(int saved_sp = _sp);
3640 3761
3641 3762 // Argument words: (int depth)
3642 3763 int nargs = 1;
3643 3764
3644 3765 _sp += nargs;
3645 3766 Node* caller_depth_node = pop();
3646 3767
3647 3768 assert(saved_sp == _sp, "must have correct argument count");
3648 3769
3649 3770 // The depth value must be a constant in order for the runtime call
3650 3771 // to be eliminated.
3651 3772 const TypeInt* caller_depth_type = _gvn.type(caller_depth_node)->isa_int();
3652 3773 if (caller_depth_type == NULL || !caller_depth_type->is_con()) {
3653 3774 #ifndef PRODUCT
3654 3775 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3655 3776 tty->print_cr(" Bailing out because caller depth was not a constant");
3656 3777 }
3657 3778 #endif
3658 3779 return false;
3659 3780 }
3660 3781 // Note that the JVM state at this point does not include the
3661 3782 // getCallerClass() frame which we are trying to inline. The
3662 3783 // semantics of getCallerClass(), however, are that the "first"
3663 3784 // frame is the getCallerClass() frame, so we subtract one from the
3664 3785 // requested depth before continuing. We don't inline requests of
3665 3786 // getCallerClass(0).
3666 3787 int caller_depth = caller_depth_type->get_con() - 1;
3667 3788 if (caller_depth < 0) {
3668 3789 #ifndef PRODUCT
3669 3790 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3670 3791 tty->print_cr(" Bailing out because caller depth was %d", caller_depth);
3671 3792 }
3672 3793 #endif
3673 3794 return false;
3674 3795 }
3675 3796
3676 3797 if (!jvms()->has_method()) {
3677 3798 #ifndef PRODUCT
3678 3799 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3679 3800 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
3680 3801 }
3681 3802 #endif
3682 3803 return false;
3683 3804 }
3684 3805 int _depth = jvms()->depth(); // cache call chain depth
3685 3806
3686 3807 // Walk back up the JVM state to find the caller at the required
3687 3808 // depth. NOTE that this code must perform the same logic as
3688 3809 // vframeStream::security_get_caller_frame in that it must skip
3689 3810 // Method.invoke() and auxiliary frames. Note also that depth is
3690 3811 // 1-based (1 is the bottom of the inlining).
3691 3812 int inlining_depth = _depth;
3692 3813 JVMState* caller_jvms = NULL;
3693 3814
3694 3815 if (inlining_depth > 0) {
3695 3816 caller_jvms = jvms();
3696 3817 assert(caller_jvms = jvms()->of_depth(inlining_depth), "inlining_depth == our depth");
3697 3818 do {
3698 3819 // The following if-tests should be performed in this order
3699 3820 if (is_method_invoke_or_aux_frame(caller_jvms)) {
3700 3821 // Skip a Method.invoke() or auxiliary frame
3701 3822 } else if (caller_depth > 0) {
3702 3823 // Skip real frame
3703 3824 --caller_depth;
3704 3825 } else {
3705 3826 // We're done: reached desired caller after skipping.
3706 3827 break;
3707 3828 }
3708 3829 caller_jvms = caller_jvms->caller();
3709 3830 --inlining_depth;
3710 3831 } while (inlining_depth > 0);
3711 3832 }
3712 3833
3713 3834 if (inlining_depth == 0) {
3714 3835 #ifndef PRODUCT
3715 3836 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3716 3837 tty->print_cr(" Bailing out because caller depth (%d) exceeded inlining depth (%d)", caller_depth_type->get_con(), _depth);
3717 3838 tty->print_cr(" JVM state at this point:");
3718 3839 for (int i = _depth; i >= 1; i--) {
3719 3840 tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8());
3720 3841 }
3721 3842 }
3722 3843 #endif
3723 3844 return false; // Reached end of inlining
3724 3845 }
3725 3846
3726 3847 // Acquire method holder as java.lang.Class
3727 3848 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
3728 3849 ciInstance* caller_mirror = caller_klass->java_mirror();
3729 3850 // Push this as a constant
3730 3851 push(makecon(TypeInstPtr::make(caller_mirror)));
3731 3852 #ifndef PRODUCT
3732 3853 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3733 3854 tty->print_cr(" Succeeded: caller = %s.%s, caller depth = %d, depth = %d", caller_klass->name()->as_utf8(), caller_jvms->method()->name()->as_utf8(), caller_depth_type->get_con(), _depth);
3734 3855 tty->print_cr(" JVM state at this point:");
3735 3856 for (int i = _depth; i >= 1; i--) {
3736 3857 tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8());
3737 3858 }
3738 3859 }
3739 3860 #endif
3740 3861 return true;
3741 3862 }
3742 3863
3743 3864 // Helper routine for above
3744 3865 bool LibraryCallKit::is_method_invoke_or_aux_frame(JVMState* jvms) {
3745 3866 ciMethod* method = jvms->method();
3746 3867
3747 3868 // Is this the Method.invoke method itself?
3748 3869 if (method->intrinsic_id() == vmIntrinsics::_invoke)
3749 3870 return true;
3750 3871
3751 3872 // Is this a helper, defined somewhere underneath MethodAccessorImpl.
3752 3873 ciKlass* k = method->holder();
3753 3874 if (k->is_instance_klass()) {
3754 3875 ciInstanceKlass* ik = k->as_instance_klass();
3755 3876 for (; ik != NULL; ik = ik->super()) {
3756 3877 if (ik->name() == ciSymbol::sun_reflect_MethodAccessorImpl() &&
3757 3878 ik == env()->find_system_klass(ik->name())) {
3758 3879 return true;
3759 3880 }
3760 3881 }
3761 3882 }
3762 3883 else if (method->is_method_handle_adapter()) {
3763 3884 // This is an internal adapter frame from the MethodHandleCompiler -- skip it
3764 3885 return true;
3765 3886 }
3766 3887
3767 3888 return false;
3768 3889 }
3769 3890
3770 3891 static int value_field_offset = -1; // offset of the "value" field of AtomicLongCSImpl. This is needed by
3771 3892 // inline_native_AtomicLong_attemptUpdate() but it has no way of
3772 3893 // computing it since there is no lookup field by name function in the
3773 3894 // CI interface. This is computed and set by inline_native_AtomicLong_get().
3774 3895 // Using a static variable here is safe even if we have multiple compilation
3775 3896 // threads because the offset is constant. At worst the same offset will be
3776 3897 // computed and stored multiple
3777 3898
3778 3899 bool LibraryCallKit::inline_native_AtomicLong_get() {
3779 3900 // Restore the stack and pop off the argument
3780 3901 _sp+=1;
3781 3902 Node *obj = pop();
3782 3903
3783 3904 // get the offset of the "value" field. Since the CI interfaces
3784 3905 // does not provide a way to look up a field by name, we scan the bytecodes
3785 3906 // to get the field index. We expect the first 2 instructions of the method
3786 3907 // to be:
3787 3908 // 0 aload_0
3788 3909 // 1 getfield "value"
3789 3910 ciMethod* method = callee();
3790 3911 if (value_field_offset == -1)
3791 3912 {
3792 3913 ciField* value_field;
3793 3914 ciBytecodeStream iter(method);
3794 3915 Bytecodes::Code bc = iter.next();
3795 3916
3796 3917 if ((bc != Bytecodes::_aload_0) &&
3797 3918 ((bc != Bytecodes::_aload) || (iter.get_index() != 0)))
3798 3919 return false;
3799 3920 bc = iter.next();
3800 3921 if (bc != Bytecodes::_getfield)
3801 3922 return false;
3802 3923 bool ignore;
3803 3924 value_field = iter.get_field(ignore);
3804 3925 value_field_offset = value_field->offset_in_bytes();
3805 3926 }
3806 3927
3807 3928 // Null check without removing any arguments.
3808 3929 _sp++;
3809 3930 obj = do_null_check(obj, T_OBJECT);
3810 3931 _sp--;
3811 3932 // Check for locking null object
3812 3933 if (stopped()) return true;
3813 3934
3814 3935 Node *adr = basic_plus_adr(obj, obj, value_field_offset);
3815 3936 const TypePtr *adr_type = _gvn.type(adr)->is_ptr();
3816 3937 int alias_idx = C->get_alias_index(adr_type);
3817 3938
3818 3939 Node *result = _gvn.transform(new (C, 3) LoadLLockedNode(control(), memory(alias_idx), adr));
3819 3940
3820 3941 push_pair(result);
3821 3942
3822 3943 return true;
3823 3944 }
3824 3945
3825 3946 bool LibraryCallKit::inline_native_AtomicLong_attemptUpdate() {
3826 3947 // Restore the stack and pop off the arguments
3827 3948 _sp+=5;
3828 3949 Node *newVal = pop_pair();
3829 3950 Node *oldVal = pop_pair();
3830 3951 Node *obj = pop();
3831 3952
3832 3953 // we need the offset of the "value" field which was computed when
3833 3954 // inlining the get() method. Give up if we don't have it.
3834 3955 if (value_field_offset == -1)
3835 3956 return false;
3836 3957
3837 3958 // Null check without removing any arguments.
3838 3959 _sp+=5;
3839 3960 obj = do_null_check(obj, T_OBJECT);
3840 3961 _sp-=5;
3841 3962 // Check for locking null object
3842 3963 if (stopped()) return true;
3843 3964
3844 3965 Node *adr = basic_plus_adr(obj, obj, value_field_offset);
3845 3966 const TypePtr *adr_type = _gvn.type(adr)->is_ptr();
3846 3967 int alias_idx = C->get_alias_index(adr_type);
3847 3968
3848 3969 Node *cas = _gvn.transform(new (C, 5) StoreLConditionalNode(control(), memory(alias_idx), adr, newVal, oldVal));
3849 3970 Node *store_proj = _gvn.transform( new (C, 1) SCMemProjNode(cas));
3850 3971 set_memory(store_proj, alias_idx);
3851 3972 Node *bol = _gvn.transform( new (C, 2) BoolNode( cas, BoolTest::eq ) );
3852 3973
3853 3974 Node *result;
3854 3975 // CMove node is not used to be able fold a possible check code
3855 3976 // after attemptUpdate() call. This code could be transformed
3856 3977 // into CMove node by loop optimizations.
3857 3978 {
3858 3979 RegionNode *r = new (C, 3) RegionNode(3);
3859 3980 result = new (C, 3) PhiNode(r, TypeInt::BOOL);
3860 3981
3861 3982 Node *iff = create_and_xform_if(control(), bol, PROB_FAIR, COUNT_UNKNOWN);
3862 3983 Node *iftrue = opt_iff(r, iff);
3863 3984 r->init_req(1, iftrue);
3864 3985 result->init_req(1, intcon(1));
3865 3986 result->init_req(2, intcon(0));
3866 3987
3867 3988 set_control(_gvn.transform(r));
3868 3989 record_for_igvn(r);
3869 3990
3870 3991 C->set_has_split_ifs(true); // Has chance for split-if optimization
3871 3992 }
3872 3993
3873 3994 push(_gvn.transform(result));
3874 3995 return true;
3875 3996 }
3876 3997
3877 3998 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
3878 3999 // restore the arguments
3879 4000 _sp += arg_size();
3880 4001
3881 4002 switch (id) {
3882 4003 case vmIntrinsics::_floatToRawIntBits:
3883 4004 push(_gvn.transform( new (C, 2) MoveF2INode(pop())));
3884 4005 break;
3885 4006
3886 4007 case vmIntrinsics::_intBitsToFloat:
3887 4008 push(_gvn.transform( new (C, 2) MoveI2FNode(pop())));
3888 4009 break;
3889 4010
3890 4011 case vmIntrinsics::_doubleToRawLongBits:
3891 4012 push_pair(_gvn.transform( new (C, 2) MoveD2LNode(pop_pair())));
3892 4013 break;
3893 4014
3894 4015 case vmIntrinsics::_longBitsToDouble:
3895 4016 push_pair(_gvn.transform( new (C, 2) MoveL2DNode(pop_pair())));
3896 4017 break;
3897 4018
3898 4019 case vmIntrinsics::_doubleToLongBits: {
3899 4020 Node* value = pop_pair();
3900 4021
3901 4022 // two paths (plus control) merge in a wood
3902 4023 RegionNode *r = new (C, 3) RegionNode(3);
3903 4024 Node *phi = new (C, 3) PhiNode(r, TypeLong::LONG);
3904 4025
3905 4026 Node *cmpisnan = _gvn.transform( new (C, 3) CmpDNode(value, value));
3906 4027 // Build the boolean node
3907 4028 Node *bolisnan = _gvn.transform( new (C, 2) BoolNode( cmpisnan, BoolTest::ne ) );
3908 4029
3909 4030 // Branch either way.
3910 4031 // NaN case is less traveled, which makes all the difference.
3911 4032 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
3912 4033 Node *opt_isnan = _gvn.transform(ifisnan);
3913 4034 assert( opt_isnan->is_If(), "Expect an IfNode");
3914 4035 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
3915 4036 Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(opt_ifisnan) );
3916 4037
3917 4038 set_control(iftrue);
3918 4039
3919 4040 static const jlong nan_bits = CONST64(0x7ff8000000000000);
3920 4041 Node *slow_result = longcon(nan_bits); // return NaN
3921 4042 phi->init_req(1, _gvn.transform( slow_result ));
3922 4043 r->init_req(1, iftrue);
3923 4044
3924 4045 // Else fall through
3925 4046 Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(opt_ifisnan) );
3926 4047 set_control(iffalse);
3927 4048
3928 4049 phi->init_req(2, _gvn.transform( new (C, 2) MoveD2LNode(value)));
3929 4050 r->init_req(2, iffalse);
3930 4051
3931 4052 // Post merge
3932 4053 set_control(_gvn.transform(r));
3933 4054 record_for_igvn(r);
3934 4055
3935 4056 Node* result = _gvn.transform(phi);
3936 4057 assert(result->bottom_type()->isa_long(), "must be");
3937 4058 push_pair(result);
3938 4059
3939 4060 C->set_has_split_ifs(true); // Has chance for split-if optimization
3940 4061
3941 4062 break;
3942 4063 }
3943 4064
3944 4065 case vmIntrinsics::_floatToIntBits: {
3945 4066 Node* value = pop();
3946 4067
3947 4068 // two paths (plus control) merge in a wood
3948 4069 RegionNode *r = new (C, 3) RegionNode(3);
3949 4070 Node *phi = new (C, 3) PhiNode(r, TypeInt::INT);
3950 4071
3951 4072 Node *cmpisnan = _gvn.transform( new (C, 3) CmpFNode(value, value));
3952 4073 // Build the boolean node
3953 4074 Node *bolisnan = _gvn.transform( new (C, 2) BoolNode( cmpisnan, BoolTest::ne ) );
3954 4075
3955 4076 // Branch either way.
3956 4077 // NaN case is less traveled, which makes all the difference.
3957 4078 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
3958 4079 Node *opt_isnan = _gvn.transform(ifisnan);
3959 4080 assert( opt_isnan->is_If(), "Expect an IfNode");
3960 4081 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
3961 4082 Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(opt_ifisnan) );
3962 4083
3963 4084 set_control(iftrue);
3964 4085
3965 4086 static const jint nan_bits = 0x7fc00000;
3966 4087 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
3967 4088 phi->init_req(1, _gvn.transform( slow_result ));
3968 4089 r->init_req(1, iftrue);
3969 4090
3970 4091 // Else fall through
3971 4092 Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(opt_ifisnan) );
3972 4093 set_control(iffalse);
3973 4094
3974 4095 phi->init_req(2, _gvn.transform( new (C, 2) MoveF2INode(value)));
3975 4096 r->init_req(2, iffalse);
3976 4097
3977 4098 // Post merge
3978 4099 set_control(_gvn.transform(r));
3979 4100 record_for_igvn(r);
3980 4101
3981 4102 Node* result = _gvn.transform(phi);
3982 4103 assert(result->bottom_type()->isa_int(), "must be");
3983 4104 push(result);
3984 4105
3985 4106 C->set_has_split_ifs(true); // Has chance for split-if optimization
3986 4107
3987 4108 break;
3988 4109 }
3989 4110
3990 4111 default:
3991 4112 ShouldNotReachHere();
3992 4113 }
3993 4114
3994 4115 return true;
3995 4116 }
3996 4117
3997 4118 #ifdef _LP64
3998 4119 #define XTOP ,top() /*additional argument*/
3999 4120 #else //_LP64
4000 4121 #define XTOP /*no additional argument*/
4001 4122 #endif //_LP64
4002 4123
4003 4124 //----------------------inline_unsafe_copyMemory-------------------------
4004 4125 bool LibraryCallKit::inline_unsafe_copyMemory() {
4005 4126 if (callee()->is_static()) return false; // caller must have the capability!
4006 4127 int nargs = 1 + 5 + 3; // 5 args: (src: ptr,off, dst: ptr,off, size)
4007 4128 assert(signature()->size() == nargs-1, "copy has 5 arguments");
4008 4129 null_check_receiver(callee()); // check then ignore argument(0)
4009 4130 if (stopped()) return true;
4010 4131
4011 4132 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4012 4133
4013 4134 Node* src_ptr = argument(1);
4014 4135 Node* src_off = ConvL2X(argument(2));
4015 4136 assert(argument(3)->is_top(), "2nd half of long");
4016 4137 Node* dst_ptr = argument(4);
4017 4138 Node* dst_off = ConvL2X(argument(5));
4018 4139 assert(argument(6)->is_top(), "2nd half of long");
4019 4140 Node* size = ConvL2X(argument(7));
4020 4141 assert(argument(8)->is_top(), "2nd half of long");
4021 4142
4022 4143 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4023 4144 "fieldOffset must be byte-scaled");
4024 4145
4025 4146 Node* src = make_unsafe_address(src_ptr, src_off);
4026 4147 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4027 4148
4028 4149 // Conservatively insert a memory barrier on all memory slices.
4029 4150 // Do not let writes of the copy source or destination float below the copy.
4030 4151 insert_mem_bar(Op_MemBarCPUOrder);
4031 4152
4032 4153 // Call it. Note that the length argument is not scaled.
4033 4154 make_runtime_call(RC_LEAF|RC_NO_FP,
4034 4155 OptoRuntime::fast_arraycopy_Type(),
4035 4156 StubRoutines::unsafe_arraycopy(),
4036 4157 "unsafe_arraycopy",
4037 4158 TypeRawPtr::BOTTOM,
4038 4159 src, dst, size XTOP);
4039 4160
4040 4161 // Do not let reads of the copy destination float above the copy.
4041 4162 insert_mem_bar(Op_MemBarCPUOrder);
4042 4163
4043 4164 return true;
4044 4165 }
4045 4166
4046 4167 //------------------------clone_coping-----------------------------------
4047 4168 // Helper function for inline_native_clone.
4048 4169 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4049 4170 assert(obj_size != NULL, "");
4050 4171 Node* raw_obj = alloc_obj->in(1);
4051 4172 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4052 4173
4053 4174 if (ReduceBulkZeroing) {
4054 4175 // We will be completely responsible for initializing this object -
4055 4176 // mark Initialize node as complete.
4056 4177 AllocateNode* alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4057 4178 // The object was just allocated - there should be no any stores!
4058 4179 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4059 4180 }
4060 4181
4061 4182 // Copy the fastest available way.
4062 4183 // TODO: generate fields copies for small objects instead.
4063 4184 Node* src = obj;
4064 4185 Node* dest = alloc_obj;
4065 4186 Node* size = _gvn.transform(obj_size);
4066 4187
4067 4188 // Exclude the header but include array length to copy by 8 bytes words.
4068 4189 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4069 4190 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4070 4191 instanceOopDesc::base_offset_in_bytes();
4071 4192 // base_off:
4072 4193 // 8 - 32-bit VM
4073 4194 // 12 - 64-bit VM, compressed oops
4074 4195 // 16 - 64-bit VM, normal oops
4075 4196 if (base_off % BytesPerLong != 0) {
4076 4197 assert(UseCompressedOops, "");
4077 4198 if (is_array) {
4078 4199 // Exclude length to copy by 8 bytes words.
4079 4200 base_off += sizeof(int);
4080 4201 } else {
4081 4202 // Include klass to copy by 8 bytes words.
4082 4203 base_off = instanceOopDesc::klass_offset_in_bytes();
4083 4204 }
4084 4205 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4085 4206 }
4086 4207 src = basic_plus_adr(src, base_off);
4087 4208 dest = basic_plus_adr(dest, base_off);
4088 4209
4089 4210 // Compute the length also, if needed:
4090 4211 Node* countx = size;
4091 4212 countx = _gvn.transform( new (C, 3) SubXNode(countx, MakeConX(base_off)) );
4092 4213 countx = _gvn.transform( new (C, 3) URShiftXNode(countx, intcon(LogBytesPerLong) ));
4093 4214
4094 4215 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4095 4216 bool disjoint_bases = true;
4096 4217 generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases,
4097 4218 src, NULL, dest, NULL, countx,
4098 4219 /*dest_uninitialized*/true);
4099 4220
4100 4221 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4101 4222 if (card_mark) {
4102 4223 assert(!is_array, "");
4103 4224 // Put in store barrier for any and all oops we are sticking
4104 4225 // into this object. (We could avoid this if we could prove
4105 4226 // that the object type contains no oop fields at all.)
4106 4227 Node* no_particular_value = NULL;
4107 4228 Node* no_particular_field = NULL;
4108 4229 int raw_adr_idx = Compile::AliasIdxRaw;
4109 4230 post_barrier(control(),
4110 4231 memory(raw_adr_type),
4111 4232 alloc_obj,
4112 4233 no_particular_field,
4113 4234 raw_adr_idx,
4114 4235 no_particular_value,
4115 4236 T_OBJECT,
4116 4237 false);
4117 4238 }
4118 4239
4119 4240 // Do not let reads from the cloned object float above the arraycopy.
4120 4241 insert_mem_bar(Op_MemBarCPUOrder);
4121 4242 }
4122 4243
4123 4244 //------------------------inline_native_clone----------------------------
4124 4245 // Here are the simple edge cases:
4125 4246 // null receiver => normal trap
4126 4247 // virtual and clone was overridden => slow path to out-of-line clone
4127 4248 // not cloneable or finalizer => slow path to out-of-line Object.clone
4128 4249 //
4129 4250 // The general case has two steps, allocation and copying.
4130 4251 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4131 4252 //
4132 4253 // Copying also has two cases, oop arrays and everything else.
4133 4254 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4134 4255 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4135 4256 //
4136 4257 // These steps fold up nicely if and when the cloned object's klass
4137 4258 // can be sharply typed as an object array, a type array, or an instance.
4138 4259 //
4139 4260 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4140 4261 int nargs = 1;
4141 4262 PhiNode* result_val;
4142 4263
4143 4264 //set the original stack and the reexecute bit for the interpreter to reexecute
4144 4265 //the bytecode that invokes Object.clone if deoptimization happens
4145 4266 { PreserveReexecuteState preexecs(this);
4146 4267 jvms()->set_should_reexecute(true);
4147 4268
4148 4269 //null_check_receiver will adjust _sp (push and pop)
4149 4270 Node* obj = null_check_receiver(callee());
4150 4271 if (stopped()) return true;
4151 4272
4152 4273 _sp += nargs;
4153 4274
4154 4275 Node* obj_klass = load_object_klass(obj);
4155 4276 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4156 4277 const TypeOopPtr* toop = ((tklass != NULL)
4157 4278 ? tklass->as_instance_type()
4158 4279 : TypeInstPtr::NOTNULL);
4159 4280
4160 4281 // Conservatively insert a memory barrier on all memory slices.
4161 4282 // Do not let writes into the original float below the clone.
4162 4283 insert_mem_bar(Op_MemBarCPUOrder);
4163 4284
4164 4285 // paths into result_reg:
4165 4286 enum {
4166 4287 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4167 4288 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4168 4289 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4169 4290 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4170 4291 PATH_LIMIT
4171 4292 };
4172 4293 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
4173 4294 result_val = new(C, PATH_LIMIT) PhiNode(result_reg,
4174 4295 TypeInstPtr::NOTNULL);
4175 4296 PhiNode* result_i_o = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO);
4176 4297 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY,
4177 4298 TypePtr::BOTTOM);
4178 4299 record_for_igvn(result_reg);
4179 4300
4180 4301 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4181 4302 int raw_adr_idx = Compile::AliasIdxRaw;
4182 4303 const bool raw_mem_only = true;
4183 4304
4184 4305
4185 4306 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4186 4307 if (array_ctl != NULL) {
4187 4308 // It's an array.
4188 4309 PreserveJVMState pjvms(this);
4189 4310 set_control(array_ctl);
4190 4311 Node* obj_length = load_array_length(obj);
4191 4312 Node* obj_size = NULL;
4192 4313 Node* alloc_obj = new_array(obj_klass, obj_length, 0,
4193 4314 raw_mem_only, &obj_size);
4194 4315
4195 4316 if (!use_ReduceInitialCardMarks()) {
4196 4317 // If it is an oop array, it requires very special treatment,
4197 4318 // because card marking is required on each card of the array.
4198 4319 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4199 4320 if (is_obja != NULL) {
4200 4321 PreserveJVMState pjvms2(this);
4201 4322 set_control(is_obja);
4202 4323 // Generate a direct call to the right arraycopy function(s).
4203 4324 bool disjoint_bases = true;
4204 4325 bool length_never_negative = true;
4205 4326 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
4206 4327 obj, intcon(0), alloc_obj, intcon(0),
4207 4328 obj_length,
4208 4329 disjoint_bases, length_never_negative);
4209 4330 result_reg->init_req(_objArray_path, control());
4210 4331 result_val->init_req(_objArray_path, alloc_obj);
4211 4332 result_i_o ->set_req(_objArray_path, i_o());
4212 4333 result_mem ->set_req(_objArray_path, reset_memory());
4213 4334 }
4214 4335 }
4215 4336 // Otherwise, there are no card marks to worry about.
4216 4337 // (We can dispense with card marks if we know the allocation
4217 4338 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4218 4339 // causes the non-eden paths to take compensating steps to
4219 4340 // simulate a fresh allocation, so that no further
4220 4341 // card marks are required in compiled code to initialize
4221 4342 // the object.)
4222 4343
4223 4344 if (!stopped()) {
4224 4345 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4225 4346
4226 4347 // Present the results of the copy.
4227 4348 result_reg->init_req(_array_path, control());
4228 4349 result_val->init_req(_array_path, alloc_obj);
4229 4350 result_i_o ->set_req(_array_path, i_o());
4230 4351 result_mem ->set_req(_array_path, reset_memory());
4231 4352 }
4232 4353 }
4233 4354
4234 4355 // We only go to the instance fast case code if we pass a number of guards.
4235 4356 // The paths which do not pass are accumulated in the slow_region.
4236 4357 RegionNode* slow_region = new (C, 1) RegionNode(1);
4237 4358 record_for_igvn(slow_region);
4238 4359 if (!stopped()) {
4239 4360 // It's an instance (we did array above). Make the slow-path tests.
4240 4361 // If this is a virtual call, we generate a funny guard. We grab
4241 4362 // the vtable entry corresponding to clone() from the target object.
4242 4363 // If the target method which we are calling happens to be the
4243 4364 // Object clone() method, we pass the guard. We do not need this
4244 4365 // guard for non-virtual calls; the caller is known to be the native
4245 4366 // Object clone().
4246 4367 if (is_virtual) {
4247 4368 generate_virtual_guard(obj_klass, slow_region);
4248 4369 }
4249 4370
4250 4371 // The object must be cloneable and must not have a finalizer.
4251 4372 // Both of these conditions may be checked in a single test.
4252 4373 // We could optimize the cloneable test further, but we don't care.
4253 4374 generate_access_flags_guard(obj_klass,
4254 4375 // Test both conditions:
4255 4376 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
4256 4377 // Must be cloneable but not finalizer:
4257 4378 JVM_ACC_IS_CLONEABLE,
4258 4379 slow_region);
4259 4380 }
4260 4381
4261 4382 if (!stopped()) {
4262 4383 // It's an instance, and it passed the slow-path tests.
4263 4384 PreserveJVMState pjvms(this);
4264 4385 Node* obj_size = NULL;
4265 4386 Node* alloc_obj = new_instance(obj_klass, NULL, raw_mem_only, &obj_size);
4266 4387
4267 4388 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4268 4389
4269 4390 // Present the results of the slow call.
4270 4391 result_reg->init_req(_instance_path, control());
4271 4392 result_val->init_req(_instance_path, alloc_obj);
4272 4393 result_i_o ->set_req(_instance_path, i_o());
4273 4394 result_mem ->set_req(_instance_path, reset_memory());
4274 4395 }
4275 4396
4276 4397 // Generate code for the slow case. We make a call to clone().
4277 4398 set_control(_gvn.transform(slow_region));
4278 4399 if (!stopped()) {
4279 4400 PreserveJVMState pjvms(this);
4280 4401 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4281 4402 Node* slow_result = set_results_for_java_call(slow_call);
4282 4403 // this->control() comes from set_results_for_java_call
4283 4404 result_reg->init_req(_slow_path, control());
4284 4405 result_val->init_req(_slow_path, slow_result);
4285 4406 result_i_o ->set_req(_slow_path, i_o());
4286 4407 result_mem ->set_req(_slow_path, reset_memory());
4287 4408 }
4288 4409
4289 4410 // Return the combined state.
4290 4411 set_control( _gvn.transform(result_reg) );
4291 4412 set_i_o( _gvn.transform(result_i_o) );
4292 4413 set_all_memory( _gvn.transform(result_mem) );
4293 4414 } //original reexecute and sp are set back here
4294 4415
4295 4416 push(_gvn.transform(result_val));
4296 4417
4297 4418 return true;
4298 4419 }
4299 4420
4300 4421
4301 4422 // constants for computing the copy function
4302 4423 enum {
4303 4424 COPYFUNC_UNALIGNED = 0,
4304 4425 COPYFUNC_ALIGNED = 1, // src, dest aligned to HeapWordSize
4305 4426 COPYFUNC_CONJOINT = 0,
4306 4427 COPYFUNC_DISJOINT = 2 // src != dest, or transfer can descend
4307 4428 };
4308 4429
4309 4430 // Note: The condition "disjoint" applies also for overlapping copies
4310 4431 // where an descending copy is permitted (i.e., dest_offset <= src_offset).
4311 4432 static address
4312 4433 select_arraycopy_function(BasicType t, bool aligned, bool disjoint, const char* &name, bool dest_uninitialized) {
4313 4434 int selector =
4314 4435 (aligned ? COPYFUNC_ALIGNED : COPYFUNC_UNALIGNED) +
4315 4436 (disjoint ? COPYFUNC_DISJOINT : COPYFUNC_CONJOINT);
4316 4437
4317 4438 #define RETURN_STUB(xxx_arraycopy) { \
4318 4439 name = #xxx_arraycopy; \
4319 4440 return StubRoutines::xxx_arraycopy(); }
4320 4441
4321 4442 #define RETURN_STUB_PARM(xxx_arraycopy, parm) { \
4322 4443 name = #xxx_arraycopy; \
4323 4444 return StubRoutines::xxx_arraycopy(parm); }
4324 4445
4325 4446 switch (t) {
4326 4447 case T_BYTE:
4327 4448 case T_BOOLEAN:
4328 4449 switch (selector) {
4329 4450 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jbyte_arraycopy);
4330 4451 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jbyte_arraycopy);
4331 4452 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jbyte_disjoint_arraycopy);
4332 4453 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jbyte_disjoint_arraycopy);
4333 4454 }
4334 4455 case T_CHAR:
4335 4456 case T_SHORT:
4336 4457 switch (selector) {
4337 4458 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jshort_arraycopy);
4338 4459 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jshort_arraycopy);
4339 4460 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jshort_disjoint_arraycopy);
4340 4461 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jshort_disjoint_arraycopy);
4341 4462 }
4342 4463 case T_INT:
4343 4464 case T_FLOAT:
4344 4465 switch (selector) {
4345 4466 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jint_arraycopy);
4346 4467 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jint_arraycopy);
4347 4468 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jint_disjoint_arraycopy);
4348 4469 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jint_disjoint_arraycopy);
4349 4470 }
4350 4471 case T_DOUBLE:
4351 4472 case T_LONG:
4352 4473 switch (selector) {
4353 4474 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jlong_arraycopy);
4354 4475 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jlong_arraycopy);
4355 4476 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB(jlong_disjoint_arraycopy);
4356 4477 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB(arrayof_jlong_disjoint_arraycopy);
4357 4478 }
4358 4479 case T_ARRAY:
4359 4480 case T_OBJECT:
4360 4481 switch (selector) {
4361 4482 case COPYFUNC_CONJOINT | COPYFUNC_UNALIGNED: RETURN_STUB_PARM(oop_arraycopy, dest_uninitialized);
4362 4483 case COPYFUNC_CONJOINT | COPYFUNC_ALIGNED: RETURN_STUB_PARM(arrayof_oop_arraycopy, dest_uninitialized);
4363 4484 case COPYFUNC_DISJOINT | COPYFUNC_UNALIGNED: RETURN_STUB_PARM(oop_disjoint_arraycopy, dest_uninitialized);
4364 4485 case COPYFUNC_DISJOINT | COPYFUNC_ALIGNED: RETURN_STUB_PARM(arrayof_oop_disjoint_arraycopy, dest_uninitialized);
4365 4486 }
4366 4487 default:
4367 4488 ShouldNotReachHere();
4368 4489 return NULL;
4369 4490 }
4370 4491
4371 4492 #undef RETURN_STUB
4372 4493 #undef RETURN_STUB_PARM
4373 4494 }
4374 4495
4375 4496 //------------------------------basictype2arraycopy----------------------------
4376 4497 address LibraryCallKit::basictype2arraycopy(BasicType t,
4377 4498 Node* src_offset,
4378 4499 Node* dest_offset,
4379 4500 bool disjoint_bases,
4380 4501 const char* &name,
4381 4502 bool dest_uninitialized) {
4382 4503 const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);;
4383 4504 const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);;
4384 4505
4385 4506 bool aligned = false;
4386 4507 bool disjoint = disjoint_bases;
4387 4508
4388 4509 // if the offsets are the same, we can treat the memory regions as
4389 4510 // disjoint, because either the memory regions are in different arrays,
4390 4511 // or they are identical (which we can treat as disjoint.) We can also
4391 4512 // treat a copy with a destination index less that the source index
4392 4513 // as disjoint since a low->high copy will work correctly in this case.
4393 4514 if (src_offset_inttype != NULL && src_offset_inttype->is_con() &&
4394 4515 dest_offset_inttype != NULL && dest_offset_inttype->is_con()) {
4395 4516 // both indices are constants
4396 4517 int s_offs = src_offset_inttype->get_con();
4397 4518 int d_offs = dest_offset_inttype->get_con();
4398 4519 int element_size = type2aelembytes(t);
4399 4520 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) &&
4400 4521 ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0);
4401 4522 if (s_offs >= d_offs) disjoint = true;
4402 4523 } else if (src_offset == dest_offset && src_offset != NULL) {
4403 4524 // This can occur if the offsets are identical non-constants.
4404 4525 disjoint = true;
4405 4526 }
4406 4527
4407 4528 return select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized);
4408 4529 }
4409 4530
4410 4531
4411 4532 //------------------------------inline_arraycopy-----------------------
4412 4533 bool LibraryCallKit::inline_arraycopy() {
4413 4534 // Restore the stack and pop off the arguments.
4414 4535 int nargs = 5; // 2 oops, 3 ints, no size_t or long
4415 4536 assert(callee()->signature()->size() == nargs, "copy has 5 arguments");
4416 4537
4417 4538 Node *src = argument(0);
4418 4539 Node *src_offset = argument(1);
4419 4540 Node *dest = argument(2);
4420 4541 Node *dest_offset = argument(3);
4421 4542 Node *length = argument(4);
4422 4543
4423 4544 // Compile time checks. If any of these checks cannot be verified at compile time,
4424 4545 // we do not make a fast path for this call. Instead, we let the call remain as it
4425 4546 // is. The checks we choose to mandate at compile time are:
4426 4547 //
4427 4548 // (1) src and dest are arrays.
4428 4549 const Type* src_type = src->Value(&_gvn);
4429 4550 const Type* dest_type = dest->Value(&_gvn);
4430 4551 const TypeAryPtr* top_src = src_type->isa_aryptr();
4431 4552 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4432 4553 if (top_src == NULL || top_src->klass() == NULL ||
4433 4554 top_dest == NULL || top_dest->klass() == NULL) {
4434 4555 // Conservatively insert a memory barrier on all memory slices.
4435 4556 // Do not let writes into the source float below the arraycopy.
4436 4557 insert_mem_bar(Op_MemBarCPUOrder);
4437 4558
4438 4559 // Call StubRoutines::generic_arraycopy stub.
4439 4560 generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT,
4440 4561 src, src_offset, dest, dest_offset, length);
4441 4562
4442 4563 // Do not let reads from the destination float above the arraycopy.
4443 4564 // Since we cannot type the arrays, we don't know which slices
4444 4565 // might be affected. We could restrict this barrier only to those
4445 4566 // memory slices which pertain to array elements--but don't bother.
4446 4567 if (!InsertMemBarAfterArraycopy)
4447 4568 // (If InsertMemBarAfterArraycopy, there is already one in place.)
4448 4569 insert_mem_bar(Op_MemBarCPUOrder);
4449 4570 return true;
4450 4571 }
4451 4572
4452 4573 // (2) src and dest arrays must have elements of the same BasicType
4453 4574 // Figure out the size and type of the elements we will be copying.
4454 4575 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
4455 4576 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4456 4577 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
4457 4578 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
4458 4579
4459 4580 if (src_elem != dest_elem || dest_elem == T_VOID) {
4460 4581 // The component types are not the same or are not recognized. Punt.
4461 4582 // (But, avoid the native method wrapper to JVM_ArrayCopy.)
4462 4583 generate_slow_arraycopy(TypePtr::BOTTOM,
4463 4584 src, src_offset, dest, dest_offset, length,
4464 4585 /*dest_uninitialized*/false);
4465 4586 return true;
4466 4587 }
4467 4588
4468 4589 //---------------------------------------------------------------------------
4469 4590 // We will make a fast path for this call to arraycopy.
4470 4591
4471 4592 // We have the following tests left to perform:
4472 4593 //
4473 4594 // (3) src and dest must not be null.
4474 4595 // (4) src_offset must not be negative.
4475 4596 // (5) dest_offset must not be negative.
4476 4597 // (6) length must not be negative.
4477 4598 // (7) src_offset + length must not exceed length of src.
4478 4599 // (8) dest_offset + length must not exceed length of dest.
4479 4600 // (9) each element of an oop array must be assignable
4480 4601
4481 4602 RegionNode* slow_region = new (C, 1) RegionNode(1);
4482 4603 record_for_igvn(slow_region);
4483 4604
4484 4605 // (3) operands must not be null
4485 4606 // We currently perform our null checks with the do_null_check routine.
4486 4607 // This means that the null exceptions will be reported in the caller
4487 4608 // rather than (correctly) reported inside of the native arraycopy call.
4488 4609 // This should be corrected, given time. We do our null check with the
4489 4610 // stack pointer restored.
4490 4611 _sp += nargs;
4491 4612 src = do_null_check(src, T_ARRAY);
4492 4613 dest = do_null_check(dest, T_ARRAY);
4493 4614 _sp -= nargs;
4494 4615
4495 4616 // (4) src_offset must not be negative.
4496 4617 generate_negative_guard(src_offset, slow_region);
4497 4618
4498 4619 // (5) dest_offset must not be negative.
4499 4620 generate_negative_guard(dest_offset, slow_region);
4500 4621
4501 4622 // (6) length must not be negative (moved to generate_arraycopy()).
4502 4623 // generate_negative_guard(length, slow_region);
4503 4624
4504 4625 // (7) src_offset + length must not exceed length of src.
4505 4626 generate_limit_guard(src_offset, length,
4506 4627 load_array_length(src),
4507 4628 slow_region);
4508 4629
4509 4630 // (8) dest_offset + length must not exceed length of dest.
4510 4631 generate_limit_guard(dest_offset, length,
4511 4632 load_array_length(dest),
4512 4633 slow_region);
4513 4634
4514 4635 // (9) each element of an oop array must be assignable
4515 4636 // The generate_arraycopy subroutine checks this.
4516 4637
4517 4638 // This is where the memory effects are placed:
4518 4639 const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem);
4519 4640 generate_arraycopy(adr_type, dest_elem,
4520 4641 src, src_offset, dest, dest_offset, length,
4521 4642 false, false, slow_region);
4522 4643
4523 4644 return true;
4524 4645 }
4525 4646
4526 4647 //-----------------------------generate_arraycopy----------------------
4527 4648 // Generate an optimized call to arraycopy.
4528 4649 // Caller must guard against non-arrays.
4529 4650 // Caller must determine a common array basic-type for both arrays.
4530 4651 // Caller must validate offsets against array bounds.
4531 4652 // The slow_region has already collected guard failure paths
4532 4653 // (such as out of bounds length or non-conformable array types).
4533 4654 // The generated code has this shape, in general:
4534 4655 //
4535 4656 // if (length == 0) return // via zero_path
4536 4657 // slowval = -1
4537 4658 // if (types unknown) {
4538 4659 // slowval = call generic copy loop
4539 4660 // if (slowval == 0) return // via checked_path
4540 4661 // } else if (indexes in bounds) {
4541 4662 // if ((is object array) && !(array type check)) {
4542 4663 // slowval = call checked copy loop
4543 4664 // if (slowval == 0) return // via checked_path
4544 4665 // } else {
4545 4666 // call bulk copy loop
4546 4667 // return // via fast_path
4547 4668 // }
4548 4669 // }
4549 4670 // // adjust params for remaining work:
4550 4671 // if (slowval != -1) {
4551 4672 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n
4552 4673 // }
4553 4674 // slow_region:
4554 4675 // call slow arraycopy(src, src_offset, dest, dest_offset, length)
4555 4676 // return // via slow_call_path
4556 4677 //
4557 4678 // This routine is used from several intrinsics: System.arraycopy,
4558 4679 // Object.clone (the array subcase), and Arrays.copyOf[Range].
4559 4680 //
4560 4681 void
4561 4682 LibraryCallKit::generate_arraycopy(const TypePtr* adr_type,
4562 4683 BasicType basic_elem_type,
4563 4684 Node* src, Node* src_offset,
4564 4685 Node* dest, Node* dest_offset,
4565 4686 Node* copy_length,
4566 4687 bool disjoint_bases,
4567 4688 bool length_never_negative,
4568 4689 RegionNode* slow_region) {
4569 4690
4570 4691 if (slow_region == NULL) {
4571 4692 slow_region = new(C,1) RegionNode(1);
4572 4693 record_for_igvn(slow_region);
4573 4694 }
4574 4695
4575 4696 Node* original_dest = dest;
4576 4697 AllocateArrayNode* alloc = NULL; // used for zeroing, if needed
4577 4698 bool dest_uninitialized = false;
4578 4699
4579 4700 // See if this is the initialization of a newly-allocated array.
4580 4701 // If so, we will take responsibility here for initializing it to zero.
4581 4702 // (Note: Because tightly_coupled_allocation performs checks on the
4582 4703 // out-edges of the dest, we need to avoid making derived pointers
4583 4704 // from it until we have checked its uses.)
4584 4705 if (ReduceBulkZeroing
4585 4706 && !ZeroTLAB // pointless if already zeroed
4586 4707 && basic_elem_type != T_CONFLICT // avoid corner case
4587 4708 && !_gvn.eqv_uncast(src, dest)
4588 4709 && ((alloc = tightly_coupled_allocation(dest, slow_region))
4589 4710 != NULL)
4590 4711 && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0
4591 4712 && alloc->maybe_set_complete(&_gvn)) {
4592 4713 // "You break it, you buy it."
4593 4714 InitializeNode* init = alloc->initialization();
4594 4715 assert(init->is_complete(), "we just did this");
4595 4716 assert(dest->is_CheckCastPP(), "sanity");
4596 4717 assert(dest->in(0)->in(0) == init, "dest pinned");
4597 4718 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory
4598 4719 // From this point on, every exit path is responsible for
4599 4720 // initializing any non-copied parts of the object to zero.
4600 4721 // Also, if this flag is set we make sure that arraycopy interacts properly
4601 4722 // with G1, eliding pre-barriers. See CR 6627983.
4602 4723 dest_uninitialized = true;
4603 4724 } else {
4604 4725 // No zeroing elimination here.
4605 4726 alloc = NULL;
4606 4727 //original_dest = dest;
4607 4728 //dest_uninitialized = false;
4608 4729 }
4609 4730
4610 4731 // Results are placed here:
4611 4732 enum { fast_path = 1, // normal void-returning assembly stub
4612 4733 checked_path = 2, // special assembly stub with cleanup
4613 4734 slow_call_path = 3, // something went wrong; call the VM
4614 4735 zero_path = 4, // bypass when length of copy is zero
4615 4736 bcopy_path = 5, // copy primitive array by 64-bit blocks
4616 4737 PATH_LIMIT = 6
4617 4738 };
4618 4739 RegionNode* result_region = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
4619 4740 PhiNode* result_i_o = new(C, PATH_LIMIT) PhiNode(result_region, Type::ABIO);
4620 4741 PhiNode* result_memory = new(C, PATH_LIMIT) PhiNode(result_region, Type::MEMORY, adr_type);
4621 4742 record_for_igvn(result_region);
4622 4743 _gvn.set_type_bottom(result_i_o);
4623 4744 _gvn.set_type_bottom(result_memory);
4624 4745 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
4625 4746
4626 4747 // The slow_control path:
4627 4748 Node* slow_control;
4628 4749 Node* slow_i_o = i_o();
4629 4750 Node* slow_mem = memory(adr_type);
4630 4751 debug_only(slow_control = (Node*) badAddress);
4631 4752
4632 4753 // Checked control path:
4633 4754 Node* checked_control = top();
4634 4755 Node* checked_mem = NULL;
4635 4756 Node* checked_i_o = NULL;
4636 4757 Node* checked_value = NULL;
4637 4758
4638 4759 if (basic_elem_type == T_CONFLICT) {
4639 4760 assert(!dest_uninitialized, "");
4640 4761 Node* cv = generate_generic_arraycopy(adr_type,
4641 4762 src, src_offset, dest, dest_offset,
4642 4763 copy_length, dest_uninitialized);
4643 4764 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
4644 4765 checked_control = control();
4645 4766 checked_i_o = i_o();
4646 4767 checked_mem = memory(adr_type);
4647 4768 checked_value = cv;
4648 4769 set_control(top()); // no fast path
4649 4770 }
4650 4771
4651 4772 Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative);
4652 4773 if (not_pos != NULL) {
4653 4774 PreserveJVMState pjvms(this);
4654 4775 set_control(not_pos);
4655 4776
4656 4777 // (6) length must not be negative.
4657 4778 if (!length_never_negative) {
4658 4779 generate_negative_guard(copy_length, slow_region);
4659 4780 }
4660 4781
4661 4782 // copy_length is 0.
4662 4783 if (!stopped() && dest_uninitialized) {
4663 4784 Node* dest_length = alloc->in(AllocateNode::ALength);
4664 4785 if (_gvn.eqv_uncast(copy_length, dest_length)
4665 4786 || _gvn.find_int_con(dest_length, 1) <= 0) {
4666 4787 // There is no zeroing to do. No need for a secondary raw memory barrier.
4667 4788 } else {
4668 4789 // Clear the whole thing since there are no source elements to copy.
4669 4790 generate_clear_array(adr_type, dest, basic_elem_type,
4670 4791 intcon(0), NULL,
4671 4792 alloc->in(AllocateNode::AllocSize));
4672 4793 // Use a secondary InitializeNode as raw memory barrier.
4673 4794 // Currently it is needed only on this path since other
4674 4795 // paths have stub or runtime calls as raw memory barriers.
4675 4796 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize,
4676 4797 Compile::AliasIdxRaw,
4677 4798 top())->as_Initialize();
4678 4799 init->set_complete(&_gvn); // (there is no corresponding AllocateNode)
4679 4800 }
4680 4801 }
4681 4802
4682 4803 // Present the results of the fast call.
4683 4804 result_region->init_req(zero_path, control());
4684 4805 result_i_o ->init_req(zero_path, i_o());
4685 4806 result_memory->init_req(zero_path, memory(adr_type));
4686 4807 }
4687 4808
4688 4809 if (!stopped() && dest_uninitialized) {
4689 4810 // We have to initialize the *uncopied* part of the array to zero.
4690 4811 // The copy destination is the slice dest[off..off+len]. The other slices
4691 4812 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
4692 4813 Node* dest_size = alloc->in(AllocateNode::AllocSize);
4693 4814 Node* dest_length = alloc->in(AllocateNode::ALength);
4694 4815 Node* dest_tail = _gvn.transform( new(C,3) AddINode(dest_offset,
4695 4816 copy_length) );
4696 4817
4697 4818 // If there is a head section that needs zeroing, do it now.
4698 4819 if (find_int_con(dest_offset, -1) != 0) {
4699 4820 generate_clear_array(adr_type, dest, basic_elem_type,
4700 4821 intcon(0), dest_offset,
4701 4822 NULL);
4702 4823 }
4703 4824
4704 4825 // Next, perform a dynamic check on the tail length.
4705 4826 // It is often zero, and we can win big if we prove this.
4706 4827 // There are two wins: Avoid generating the ClearArray
4707 4828 // with its attendant messy index arithmetic, and upgrade
4708 4829 // the copy to a more hardware-friendly word size of 64 bits.
4709 4830 Node* tail_ctl = NULL;
4710 4831 if (!stopped() && !_gvn.eqv_uncast(dest_tail, dest_length)) {
4711 4832 Node* cmp_lt = _gvn.transform( new(C,3) CmpINode(dest_tail, dest_length) );
4712 4833 Node* bol_lt = _gvn.transform( new(C,2) BoolNode(cmp_lt, BoolTest::lt) );
4713 4834 tail_ctl = generate_slow_guard(bol_lt, NULL);
4714 4835 assert(tail_ctl != NULL || !stopped(), "must be an outcome");
4715 4836 }
4716 4837
4717 4838 // At this point, let's assume there is no tail.
4718 4839 if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) {
4719 4840 // There is no tail. Try an upgrade to a 64-bit copy.
4720 4841 bool didit = false;
4721 4842 { PreserveJVMState pjvms(this);
4722 4843 didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc,
4723 4844 src, src_offset, dest, dest_offset,
4724 4845 dest_size, dest_uninitialized);
4725 4846 if (didit) {
4726 4847 // Present the results of the block-copying fast call.
4727 4848 result_region->init_req(bcopy_path, control());
4728 4849 result_i_o ->init_req(bcopy_path, i_o());
4729 4850 result_memory->init_req(bcopy_path, memory(adr_type));
4730 4851 }
4731 4852 }
4732 4853 if (didit)
4733 4854 set_control(top()); // no regular fast path
4734 4855 }
4735 4856
4736 4857 // Clear the tail, if any.
4737 4858 if (tail_ctl != NULL) {
4738 4859 Node* notail_ctl = stopped() ? NULL : control();
4739 4860 set_control(tail_ctl);
4740 4861 if (notail_ctl == NULL) {
4741 4862 generate_clear_array(adr_type, dest, basic_elem_type,
4742 4863 dest_tail, NULL,
4743 4864 dest_size);
4744 4865 } else {
4745 4866 // Make a local merge.
4746 4867 Node* done_ctl = new(C,3) RegionNode(3);
4747 4868 Node* done_mem = new(C,3) PhiNode(done_ctl, Type::MEMORY, adr_type);
4748 4869 done_ctl->init_req(1, notail_ctl);
4749 4870 done_mem->init_req(1, memory(adr_type));
4750 4871 generate_clear_array(adr_type, dest, basic_elem_type,
4751 4872 dest_tail, NULL,
4752 4873 dest_size);
4753 4874 done_ctl->init_req(2, control());
4754 4875 done_mem->init_req(2, memory(adr_type));
4755 4876 set_control( _gvn.transform(done_ctl) );
4756 4877 set_memory( _gvn.transform(done_mem), adr_type );
4757 4878 }
4758 4879 }
4759 4880 }
4760 4881
4761 4882 BasicType copy_type = basic_elem_type;
4762 4883 assert(basic_elem_type != T_ARRAY, "caller must fix this");
4763 4884 if (!stopped() && copy_type == T_OBJECT) {
4764 4885 // If src and dest have compatible element types, we can copy bits.
4765 4886 // Types S[] and D[] are compatible if D is a supertype of S.
4766 4887 //
4767 4888 // If they are not, we will use checked_oop_disjoint_arraycopy,
4768 4889 // which performs a fast optimistic per-oop check, and backs off
4769 4890 // further to JVM_ArrayCopy on the first per-oop check that fails.
4770 4891 // (Actually, we don't move raw bits only; the GC requires card marks.)
4771 4892
4772 4893 // Get the klassOop for both src and dest
4773 4894 Node* src_klass = load_object_klass(src);
4774 4895 Node* dest_klass = load_object_klass(dest);
4775 4896
4776 4897 // Generate the subtype check.
4777 4898 // This might fold up statically, or then again it might not.
4778 4899 //
4779 4900 // Non-static example: Copying List<String>.elements to a new String[].
4780 4901 // The backing store for a List<String> is always an Object[],
4781 4902 // but its elements are always type String, if the generic types
4782 4903 // are correct at the source level.
4783 4904 //
4784 4905 // Test S[] against D[], not S against D, because (probably)
4785 4906 // the secondary supertype cache is less busy for S[] than S.
4786 4907 // This usually only matters when D is an interface.
4787 4908 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
4788 4909 // Plug failing path into checked_oop_disjoint_arraycopy
4789 4910 if (not_subtype_ctrl != top()) {
4790 4911 PreserveJVMState pjvms(this);
4791 4912 set_control(not_subtype_ctrl);
4792 4913 // (At this point we can assume disjoint_bases, since types differ.)
4793 4914 int ek_offset = objArrayKlass::element_klass_offset_in_bytes() + sizeof(oopDesc);
4794 4915 Node* p1 = basic_plus_adr(dest_klass, ek_offset);
4795 4916 Node* n1 = LoadKlassNode::make(_gvn, immutable_memory(), p1, TypeRawPtr::BOTTOM);
4796 4917 Node* dest_elem_klass = _gvn.transform(n1);
4797 4918 Node* cv = generate_checkcast_arraycopy(adr_type,
4798 4919 dest_elem_klass,
4799 4920 src, src_offset, dest, dest_offset,
4800 4921 ConvI2X(copy_length), dest_uninitialized);
4801 4922 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
4802 4923 checked_control = control();
4803 4924 checked_i_o = i_o();
4804 4925 checked_mem = memory(adr_type);
4805 4926 checked_value = cv;
4806 4927 }
4807 4928 // At this point we know we do not need type checks on oop stores.
4808 4929
4809 4930 // Let's see if we need card marks:
4810 4931 if (alloc != NULL && use_ReduceInitialCardMarks()) {
4811 4932 // If we do not need card marks, copy using the jint or jlong stub.
4812 4933 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT);
4813 4934 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type),
4814 4935 "sizes agree");
4815 4936 }
4816 4937 }
4817 4938
4818 4939 if (!stopped()) {
4819 4940 // Generate the fast path, if possible.
4820 4941 PreserveJVMState pjvms(this);
4821 4942 generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases,
4822 4943 src, src_offset, dest, dest_offset,
4823 4944 ConvI2X(copy_length), dest_uninitialized);
4824 4945
4825 4946 // Present the results of the fast call.
4826 4947 result_region->init_req(fast_path, control());
4827 4948 result_i_o ->init_req(fast_path, i_o());
4828 4949 result_memory->init_req(fast_path, memory(adr_type));
4829 4950 }
4830 4951
4831 4952 // Here are all the slow paths up to this point, in one bundle:
4832 4953 slow_control = top();
4833 4954 if (slow_region != NULL)
4834 4955 slow_control = _gvn.transform(slow_region);
4835 4956 debug_only(slow_region = (RegionNode*)badAddress);
4836 4957
4837 4958 set_control(checked_control);
4838 4959 if (!stopped()) {
4839 4960 // Clean up after the checked call.
4840 4961 // The returned value is either 0 or -1^K,
4841 4962 // where K = number of partially transferred array elements.
4842 4963 Node* cmp = _gvn.transform( new(C, 3) CmpINode(checked_value, intcon(0)) );
4843 4964 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) );
4844 4965 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
4845 4966
4846 4967 // If it is 0, we are done, so transfer to the end.
4847 4968 Node* checks_done = _gvn.transform( new(C, 1) IfTrueNode(iff) );
4848 4969 result_region->init_req(checked_path, checks_done);
4849 4970 result_i_o ->init_req(checked_path, checked_i_o);
4850 4971 result_memory->init_req(checked_path, checked_mem);
4851 4972
4852 4973 // If it is not zero, merge into the slow call.
4853 4974 set_control( _gvn.transform( new(C, 1) IfFalseNode(iff) ));
4854 4975 RegionNode* slow_reg2 = new(C, 3) RegionNode(3);
4855 4976 PhiNode* slow_i_o2 = new(C, 3) PhiNode(slow_reg2, Type::ABIO);
4856 4977 PhiNode* slow_mem2 = new(C, 3) PhiNode(slow_reg2, Type::MEMORY, adr_type);
4857 4978 record_for_igvn(slow_reg2);
4858 4979 slow_reg2 ->init_req(1, slow_control);
4859 4980 slow_i_o2 ->init_req(1, slow_i_o);
4860 4981 slow_mem2 ->init_req(1, slow_mem);
4861 4982 slow_reg2 ->init_req(2, control());
4862 4983 slow_i_o2 ->init_req(2, checked_i_o);
4863 4984 slow_mem2 ->init_req(2, checked_mem);
4864 4985
4865 4986 slow_control = _gvn.transform(slow_reg2);
4866 4987 slow_i_o = _gvn.transform(slow_i_o2);
4867 4988 slow_mem = _gvn.transform(slow_mem2);
4868 4989
4869 4990 if (alloc != NULL) {
4870 4991 // We'll restart from the very beginning, after zeroing the whole thing.
4871 4992 // This can cause double writes, but that's OK since dest is brand new.
4872 4993 // So we ignore the low 31 bits of the value returned from the stub.
4873 4994 } else {
4874 4995 // We must continue the copy exactly where it failed, or else
4875 4996 // another thread might see the wrong number of writes to dest.
4876 4997 Node* checked_offset = _gvn.transform( new(C, 3) XorINode(checked_value, intcon(-1)) );
4877 4998 Node* slow_offset = new(C, 3) PhiNode(slow_reg2, TypeInt::INT);
4878 4999 slow_offset->init_req(1, intcon(0));
4879 5000 slow_offset->init_req(2, checked_offset);
4880 5001 slow_offset = _gvn.transform(slow_offset);
4881 5002
4882 5003 // Adjust the arguments by the conditionally incoming offset.
4883 5004 Node* src_off_plus = _gvn.transform( new(C, 3) AddINode(src_offset, slow_offset) );
4884 5005 Node* dest_off_plus = _gvn.transform( new(C, 3) AddINode(dest_offset, slow_offset) );
4885 5006 Node* length_minus = _gvn.transform( new(C, 3) SubINode(copy_length, slow_offset) );
4886 5007
4887 5008 // Tweak the node variables to adjust the code produced below:
4888 5009 src_offset = src_off_plus;
4889 5010 dest_offset = dest_off_plus;
4890 5011 copy_length = length_minus;
4891 5012 }
4892 5013 }
4893 5014
4894 5015 set_control(slow_control);
4895 5016 if (!stopped()) {
4896 5017 // Generate the slow path, if needed.
4897 5018 PreserveJVMState pjvms(this); // replace_in_map may trash the map
4898 5019
4899 5020 set_memory(slow_mem, adr_type);
4900 5021 set_i_o(slow_i_o);
4901 5022
4902 5023 if (dest_uninitialized) {
4903 5024 generate_clear_array(adr_type, dest, basic_elem_type,
4904 5025 intcon(0), NULL,
4905 5026 alloc->in(AllocateNode::AllocSize));
4906 5027 }
4907 5028
4908 5029 generate_slow_arraycopy(adr_type,
4909 5030 src, src_offset, dest, dest_offset,
4910 5031 copy_length, /*dest_uninitialized*/false);
4911 5032
4912 5033 result_region->init_req(slow_call_path, control());
4913 5034 result_i_o ->init_req(slow_call_path, i_o());
4914 5035 result_memory->init_req(slow_call_path, memory(adr_type));
4915 5036 }
4916 5037
4917 5038 // Remove unused edges.
4918 5039 for (uint i = 1; i < result_region->req(); i++) {
4919 5040 if (result_region->in(i) == NULL)
4920 5041 result_region->init_req(i, top());
4921 5042 }
4922 5043
4923 5044 // Finished; return the combined state.
4924 5045 set_control( _gvn.transform(result_region) );
4925 5046 set_i_o( _gvn.transform(result_i_o) );
4926 5047 set_memory( _gvn.transform(result_memory), adr_type );
4927 5048
4928 5049 // The memory edges above are precise in order to model effects around
4929 5050 // array copies accurately to allow value numbering of field loads around
4930 5051 // arraycopy. Such field loads, both before and after, are common in Java
4931 5052 // collections and similar classes involving header/array data structures.
4932 5053 //
4933 5054 // But with low number of register or when some registers are used or killed
4934 5055 // by arraycopy calls it causes registers spilling on stack. See 6544710.
4935 5056 // The next memory barrier is added to avoid it. If the arraycopy can be
4936 5057 // optimized away (which it can, sometimes) then we can manually remove
4937 5058 // the membar also.
4938 5059 //
4939 5060 // Do not let reads from the cloned object float above the arraycopy.
4940 5061 if (InsertMemBarAfterArraycopy || alloc != NULL)
4941 5062 insert_mem_bar(Op_MemBarCPUOrder);
4942 5063 }
4943 5064
4944 5065
4945 5066 // Helper function which determines if an arraycopy immediately follows
4946 5067 // an allocation, with no intervening tests or other escapes for the object.
4947 5068 AllocateArrayNode*
4948 5069 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
4949 5070 RegionNode* slow_region) {
4950 5071 if (stopped()) return NULL; // no fast path
4951 5072 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
4952 5073
4953 5074 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
4954 5075 if (alloc == NULL) return NULL;
4955 5076
4956 5077 Node* rawmem = memory(Compile::AliasIdxRaw);
4957 5078 // Is the allocation's memory state untouched?
4958 5079 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
4959 5080 // Bail out if there have been raw-memory effects since the allocation.
4960 5081 // (Example: There might have been a call or safepoint.)
4961 5082 return NULL;
4962 5083 }
4963 5084 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
4964 5085 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
4965 5086 return NULL;
4966 5087 }
4967 5088
4968 5089 // There must be no unexpected observers of this allocation.
4969 5090 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
4970 5091 Node* obs = ptr->fast_out(i);
4971 5092 if (obs != this->map()) {
4972 5093 return NULL;
4973 5094 }
4974 5095 }
4975 5096
4976 5097 // This arraycopy must unconditionally follow the allocation of the ptr.
4977 5098 Node* alloc_ctl = ptr->in(0);
4978 5099 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
4979 5100
4980 5101 Node* ctl = control();
4981 5102 while (ctl != alloc_ctl) {
4982 5103 // There may be guards which feed into the slow_region.
4983 5104 // Any other control flow means that we might not get a chance
4984 5105 // to finish initializing the allocated object.
4985 5106 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
4986 5107 IfNode* iff = ctl->in(0)->as_If();
4987 5108 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
4988 5109 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
4989 5110 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
4990 5111 ctl = iff->in(0); // This test feeds the known slow_region.
4991 5112 continue;
4992 5113 }
4993 5114 // One more try: Various low-level checks bottom out in
4994 5115 // uncommon traps. If the debug-info of the trap omits
4995 5116 // any reference to the allocation, as we've already
4996 5117 // observed, then there can be no objection to the trap.
4997 5118 bool found_trap = false;
4998 5119 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
4999 5120 Node* obs = not_ctl->fast_out(j);
5000 5121 if (obs->in(0) == not_ctl && obs->is_Call() &&
5001 5122 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
5002 5123 found_trap = true; break;
5003 5124 }
5004 5125 }
5005 5126 if (found_trap) {
5006 5127 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
5007 5128 continue;
5008 5129 }
5009 5130 }
5010 5131 return NULL;
5011 5132 }
5012 5133
5013 5134 // If we get this far, we have an allocation which immediately
5014 5135 // precedes the arraycopy, and we can take over zeroing the new object.
5015 5136 // The arraycopy will finish the initialization, and provide
5016 5137 // a new control state to which we will anchor the destination pointer.
5017 5138
5018 5139 return alloc;
5019 5140 }
5020 5141
5021 5142 // Helper for initialization of arrays, creating a ClearArray.
5022 5143 // It writes zero bits in [start..end), within the body of an array object.
5023 5144 // The memory effects are all chained onto the 'adr_type' alias category.
5024 5145 //
5025 5146 // Since the object is otherwise uninitialized, we are free
5026 5147 // to put a little "slop" around the edges of the cleared area,
5027 5148 // as long as it does not go back into the array's header,
5028 5149 // or beyond the array end within the heap.
5029 5150 //
5030 5151 // The lower edge can be rounded down to the nearest jint and the
5031 5152 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
5032 5153 //
5033 5154 // Arguments:
5034 5155 // adr_type memory slice where writes are generated
5035 5156 // dest oop of the destination array
5036 5157 // basic_elem_type element type of the destination
5037 5158 // slice_idx array index of first element to store
5038 5159 // slice_len number of elements to store (or NULL)
5039 5160 // dest_size total size in bytes of the array object
5040 5161 //
5041 5162 // Exactly one of slice_len or dest_size must be non-NULL.
5042 5163 // If dest_size is non-NULL, zeroing extends to the end of the object.
5043 5164 // If slice_len is non-NULL, the slice_idx value must be a constant.
5044 5165 void
5045 5166 LibraryCallKit::generate_clear_array(const TypePtr* adr_type,
5046 5167 Node* dest,
5047 5168 BasicType basic_elem_type,
5048 5169 Node* slice_idx,
5049 5170 Node* slice_len,
5050 5171 Node* dest_size) {
5051 5172 // one or the other but not both of slice_len and dest_size:
5052 5173 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, "");
5053 5174 if (slice_len == NULL) slice_len = top();
5054 5175 if (dest_size == NULL) dest_size = top();
5055 5176
5056 5177 // operate on this memory slice:
5057 5178 Node* mem = memory(adr_type); // memory slice to operate on
5058 5179
5059 5180 // scaling and rounding of indexes:
5060 5181 int scale = exact_log2(type2aelembytes(basic_elem_type));
5061 5182 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5062 5183 int clear_low = (-1 << scale) & (BytesPerInt - 1);
5063 5184 int bump_bit = (-1 << scale) & BytesPerInt;
5064 5185
5065 5186 // determine constant starts and ends
5066 5187 const intptr_t BIG_NEG = -128;
5067 5188 assert(BIG_NEG + 2*abase < 0, "neg enough");
5068 5189 intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG);
5069 5190 intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG);
5070 5191 if (slice_len_con == 0) {
5071 5192 return; // nothing to do here
5072 5193 }
5073 5194 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
5074 5195 intptr_t end_con = find_intptr_t_con(dest_size, -1);
5075 5196 if (slice_idx_con >= 0 && slice_len_con >= 0) {
5076 5197 assert(end_con < 0, "not two cons");
5077 5198 end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale),
5078 5199 BytesPerLong);
5079 5200 }
5080 5201
5081 5202 if (start_con >= 0 && end_con >= 0) {
5082 5203 // Constant start and end. Simple.
5083 5204 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5084 5205 start_con, end_con, &_gvn);
5085 5206 } else if (start_con >= 0 && dest_size != top()) {
5086 5207 // Constant start, pre-rounded end after the tail of the array.
5087 5208 Node* end = dest_size;
5088 5209 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5089 5210 start_con, end, &_gvn);
5090 5211 } else if (start_con >= 0 && slice_len != top()) {
5091 5212 // Constant start, non-constant end. End needs rounding up.
5092 5213 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
5093 5214 intptr_t end_base = abase + (slice_idx_con << scale);
5094 5215 int end_round = (-1 << scale) & (BytesPerLong - 1);
5095 5216 Node* end = ConvI2X(slice_len);
5096 5217 if (scale != 0)
5097 5218 end = _gvn.transform( new(C,3) LShiftXNode(end, intcon(scale) ));
5098 5219 end_base += end_round;
5099 5220 end = _gvn.transform( new(C,3) AddXNode(end, MakeConX(end_base)) );
5100 5221 end = _gvn.transform( new(C,3) AndXNode(end, MakeConX(~end_round)) );
5101 5222 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5102 5223 start_con, end, &_gvn);
5103 5224 } else if (start_con < 0 && dest_size != top()) {
5104 5225 // Non-constant start, pre-rounded end after the tail of the array.
5105 5226 // This is almost certainly a "round-to-end" operation.
5106 5227 Node* start = slice_idx;
5107 5228 start = ConvI2X(start);
5108 5229 if (scale != 0)
5109 5230 start = _gvn.transform( new(C,3) LShiftXNode( start, intcon(scale) ));
5110 5231 start = _gvn.transform( new(C,3) AddXNode(start, MakeConX(abase)) );
5111 5232 if ((bump_bit | clear_low) != 0) {
5112 5233 int to_clear = (bump_bit | clear_low);
5113 5234 // Align up mod 8, then store a jint zero unconditionally
5114 5235 // just before the mod-8 boundary.
5115 5236 if (((abase + bump_bit) & ~to_clear) - bump_bit
5116 5237 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
5117 5238 bump_bit = 0;
5118 5239 assert((abase & to_clear) == 0, "array base must be long-aligned");
5119 5240 } else {
5120 5241 // Bump 'start' up to (or past) the next jint boundary:
5121 5242 start = _gvn.transform( new(C,3) AddXNode(start, MakeConX(bump_bit)) );
5122 5243 assert((abase & clear_low) == 0, "array base must be int-aligned");
5123 5244 }
5124 5245 // Round bumped 'start' down to jlong boundary in body of array.
5125 5246 start = _gvn.transform( new(C,3) AndXNode(start, MakeConX(~to_clear)) );
5126 5247 if (bump_bit != 0) {
5127 5248 // Store a zero to the immediately preceding jint:
5128 5249 Node* x1 = _gvn.transform( new(C,3) AddXNode(start, MakeConX(-bump_bit)) );
5129 5250 Node* p1 = basic_plus_adr(dest, x1);
5130 5251 mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT);
5131 5252 mem = _gvn.transform(mem);
5132 5253 }
5133 5254 }
5134 5255 Node* end = dest_size; // pre-rounded
5135 5256 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5136 5257 start, end, &_gvn);
5137 5258 } else {
5138 5259 // Non-constant start, unrounded non-constant end.
5139 5260 // (Nobody zeroes a random midsection of an array using this routine.)
5140 5261 ShouldNotReachHere(); // fix caller
5141 5262 }
5142 5263
5143 5264 // Done.
5144 5265 set_memory(mem, adr_type);
5145 5266 }
5146 5267
5147 5268
5148 5269 bool
5149 5270 LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type,
5150 5271 BasicType basic_elem_type,
5151 5272 AllocateNode* alloc,
5152 5273 Node* src, Node* src_offset,
5153 5274 Node* dest, Node* dest_offset,
5154 5275 Node* dest_size, bool dest_uninitialized) {
5155 5276 // See if there is an advantage from block transfer.
5156 5277 int scale = exact_log2(type2aelembytes(basic_elem_type));
5157 5278 if (scale >= LogBytesPerLong)
5158 5279 return false; // it is already a block transfer
5159 5280
5160 5281 // Look at the alignment of the starting offsets.
5161 5282 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5162 5283 const intptr_t BIG_NEG = -128;
5163 5284 assert(BIG_NEG + 2*abase < 0, "neg enough");
5164 5285
5165 5286 intptr_t src_off = abase + ((intptr_t) find_int_con(src_offset, -1) << scale);
5166 5287 intptr_t dest_off = abase + ((intptr_t) find_int_con(dest_offset, -1) << scale);
5167 5288 if (src_off < 0 || dest_off < 0)
5168 5289 // At present, we can only understand constants.
5169 5290 return false;
5170 5291
5171 5292 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
5172 5293 // Non-aligned; too bad.
5173 5294 // One more chance: Pick off an initial 32-bit word.
5174 5295 // This is a common case, since abase can be odd mod 8.
5175 5296 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
5176 5297 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
5177 5298 Node* sptr = basic_plus_adr(src, src_off);
5178 5299 Node* dptr = basic_plus_adr(dest, dest_off);
5179 5300 Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type);
5180 5301 store_to_memory(control(), dptr, sval, T_INT, adr_type);
5181 5302 src_off += BytesPerInt;
5182 5303 dest_off += BytesPerInt;
5183 5304 } else {
5184 5305 return false;
5185 5306 }
5186 5307 }
5187 5308 assert(src_off % BytesPerLong == 0, "");
5188 5309 assert(dest_off % BytesPerLong == 0, "");
5189 5310
5190 5311 // Do this copy by giant steps.
5191 5312 Node* sptr = basic_plus_adr(src, src_off);
5192 5313 Node* dptr = basic_plus_adr(dest, dest_off);
5193 5314 Node* countx = dest_size;
5194 5315 countx = _gvn.transform( new (C, 3) SubXNode(countx, MakeConX(dest_off)) );
5195 5316 countx = _gvn.transform( new (C, 3) URShiftXNode(countx, intcon(LogBytesPerLong)) );
5196 5317
5197 5318 bool disjoint_bases = true; // since alloc != NULL
5198 5319 generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases,
5199 5320 sptr, NULL, dptr, NULL, countx, dest_uninitialized);
5200 5321
5201 5322 return true;
5202 5323 }
5203 5324
5204 5325
5205 5326 // Helper function; generates code for the slow case.
5206 5327 // We make a call to a runtime method which emulates the native method,
5207 5328 // but without the native wrapper overhead.
5208 5329 void
5209 5330 LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type,
5210 5331 Node* src, Node* src_offset,
5211 5332 Node* dest, Node* dest_offset,
5212 5333 Node* copy_length, bool dest_uninitialized) {
5213 5334 assert(!dest_uninitialized, "Invariant");
5214 5335 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON,
5215 5336 OptoRuntime::slow_arraycopy_Type(),
5216 5337 OptoRuntime::slow_arraycopy_Java(),
5217 5338 "slow_arraycopy", adr_type,
5218 5339 src, src_offset, dest, dest_offset,
5219 5340 copy_length);
5220 5341
5221 5342 // Handle exceptions thrown by this fellow:
5222 5343 make_slow_call_ex(call, env()->Throwable_klass(), false);
5223 5344 }
5224 5345
5225 5346 // Helper function; generates code for cases requiring runtime checks.
5226 5347 Node*
5227 5348 LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type,
5228 5349 Node* dest_elem_klass,
5229 5350 Node* src, Node* src_offset,
5230 5351 Node* dest, Node* dest_offset,
5231 5352 Node* copy_length, bool dest_uninitialized) {
5232 5353 if (stopped()) return NULL;
5233 5354
5234 5355 address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized);
5235 5356 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5236 5357 return NULL;
5237 5358 }
5238 5359
5239 5360 // Pick out the parameters required to perform a store-check
5240 5361 // for the target array. This is an optimistic check. It will
5241 5362 // look in each non-null element's class, at the desired klass's
5242 5363 // super_check_offset, for the desired klass.
5243 5364 int sco_offset = Klass::super_check_offset_offset_in_bytes() + sizeof(oopDesc);
5244 5365 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
5245 5366 Node* n3 = new(C, 3) LoadINode(NULL, memory(p3), p3, _gvn.type(p3)->is_ptr());
5246 5367 Node* check_offset = ConvI2X(_gvn.transform(n3));
5247 5368 Node* check_value = dest_elem_klass;
5248 5369
5249 5370 Node* src_start = array_element_address(src, src_offset, T_OBJECT);
5250 5371 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
5251 5372
5252 5373 // (We know the arrays are never conjoint, because their types differ.)
5253 5374 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5254 5375 OptoRuntime::checkcast_arraycopy_Type(),
5255 5376 copyfunc_addr, "checkcast_arraycopy", adr_type,
5256 5377 // five arguments, of which two are
5257 5378 // intptr_t (jlong in LP64)
5258 5379 src_start, dest_start,
5259 5380 copy_length XTOP,
5260 5381 check_offset XTOP,
5261 5382 check_value);
5262 5383
5263 5384 return _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms));
5264 5385 }
5265 5386
5266 5387
5267 5388 // Helper function; generates code for cases requiring runtime checks.
5268 5389 Node*
5269 5390 LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type,
5270 5391 Node* src, Node* src_offset,
5271 5392 Node* dest, Node* dest_offset,
5272 5393 Node* copy_length, bool dest_uninitialized) {
5273 5394 assert(!dest_uninitialized, "Invariant");
5274 5395 if (stopped()) return NULL;
5275 5396 address copyfunc_addr = StubRoutines::generic_arraycopy();
5276 5397 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5277 5398 return NULL;
5278 5399 }
5279 5400
5280 5401 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5281 5402 OptoRuntime::generic_arraycopy_Type(),
5282 5403 copyfunc_addr, "generic_arraycopy", adr_type,
5283 5404 src, src_offset, dest, dest_offset, copy_length);
5284 5405
5285 5406 return _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms));
5286 5407 }
5287 5408
5288 5409 // Helper function; generates the fast out-of-line call to an arraycopy stub.
5289 5410 void
5290 5411 LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type,
5291 5412 BasicType basic_elem_type,
5292 5413 bool disjoint_bases,
5293 5414 Node* src, Node* src_offset,
5294 5415 Node* dest, Node* dest_offset,
5295 5416 Node* copy_length, bool dest_uninitialized) {
5296 5417 if (stopped()) return; // nothing to do
5297 5418
5298 5419 Node* src_start = src;
5299 5420 Node* dest_start = dest;
5300 5421 if (src_offset != NULL || dest_offset != NULL) {
5301 5422 assert(src_offset != NULL && dest_offset != NULL, "");
5302 5423 src_start = array_element_address(src, src_offset, basic_elem_type);
5303 5424 dest_start = array_element_address(dest, dest_offset, basic_elem_type);
5304 5425 }
5305 5426
5306 5427 // Figure out which arraycopy runtime method to call.
5307 5428 const char* copyfunc_name = "arraycopy";
|
↓ open down ↓ |
2755 lines elided |
↑ open up ↑ |
5308 5429 address copyfunc_addr =
5309 5430 basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
5310 5431 disjoint_bases, copyfunc_name, dest_uninitialized);
5311 5432
5312 5433 // Call it. Note that the count_ix value is not scaled to a byte-size.
5313 5434 make_runtime_call(RC_LEAF|RC_NO_FP,
5314 5435 OptoRuntime::fast_arraycopy_Type(),
5315 5436 copyfunc_addr, copyfunc_name, adr_type,
5316 5437 src_start, dest_start, copy_length XTOP);
5317 5438 }
5439 +
5440 +//----------------------------inline_reference_get----------------------------
5441 +
5442 +bool LibraryCallKit::inline_reference_get() {
5443 + const int nargs = 1; // self
5444 +
5445 + guarantee(java_lang_ref_Reference::referent_offset > 0,
5446 + "should have already been set");
5447 +
5448 + int referent_offset = java_lang_ref_Reference::referent_offset;
5449 +
5450 + // Restore the stack and pop off the argument
5451 + _sp += nargs;
5452 + Node *reference_obj = pop();
5453 +
5454 + // Null check on self without removing any arguments.
5455 + _sp += nargs;
5456 + reference_obj = do_null_check(reference_obj, T_OBJECT);
5457 + _sp -= nargs;;
5458 +
5459 + if (stopped()) return true;
5460 +
5461 + Node *adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
5462 +
5463 + ciInstanceKlass* klass = env()->Object_klass();
5464 + const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
5465 +
5466 + Node* no_ctrl = NULL;
5467 + Node *result = make_load(no_ctrl, adr, object_type, T_OBJECT);
5468 +
5469 + // Use the pre-barrier to record the value in the referent field
5470 + pre_barrier(false /* do_load */,
5471 + control(),
5472 + NULL /* obj */, NULL /* adr */, -1 /* alias_idx */, NULL /* val */, NULL /* val_type */,
5473 + result /* pre_val */,
5474 + T_OBJECT);
5475 +
5476 + push(result);
5477 + return true;
5478 +}
5479 +
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX