rev 4136 : 7153771: array bound check elimination for c1
Summary: when possible optimize out array bound checks, inserting predicates when needed.
Reviewed-by:

   1 /*
   2  * Copyright (c) 1999, 2012, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "precompiled.hpp"
  26 #include "c1/c1_CodeStubs.hpp"
  27 #include "c1/c1_FrameMap.hpp"
  28 #include "c1/c1_LIRAssembler.hpp"
  29 #include "c1/c1_MacroAssembler.hpp"
  30 #include "c1/c1_Runtime1.hpp"
  31 #include "nativeInst_x86.hpp"
  32 #include "runtime/sharedRuntime.hpp"
  33 #include "utilities/macros.hpp"
  34 #include "vmreg_x86.inline.hpp"
  35 #if INCLUDE_ALL_GCS
  36 #include "gc_implementation/g1/g1SATBCardTableModRefBS.hpp"
  37 #endif // INCLUDE_ALL_GCS
  38 
  39 
  40 #define __ ce->masm()->
  41 
  42 float ConversionStub::float_zero = 0.0;
  43 double ConversionStub::double_zero = 0.0;
  44 
  45 void ConversionStub::emit_code(LIR_Assembler* ce) {
  46   __ bind(_entry);
  47   assert(bytecode() == Bytecodes::_f2i || bytecode() == Bytecodes::_d2i, "other conversions do not require stub");
  48 
  49 
  50   if (input()->is_single_xmm()) {
  51     __ comiss(input()->as_xmm_float_reg(),
  52               ExternalAddress((address)&float_zero));
  53   } else if (input()->is_double_xmm()) {
  54     __ comisd(input()->as_xmm_double_reg(),
  55               ExternalAddress((address)&double_zero));
  56   } else {
  57     LP64_ONLY(ShouldNotReachHere());
  58     __ push(rax);
  59     __ ftst();
  60     __ fnstsw_ax();
  61     __ sahf();
  62     __ pop(rax);
  63   }
  64 
  65   Label NaN, do_return;
  66   __ jccb(Assembler::parity, NaN);
  67   __ jccb(Assembler::below, do_return);
  68 
  69   // input is > 0 -> return maxInt
  70   // result register already contains 0x80000000, so subtracting 1 gives 0x7fffffff
  71   __ decrement(result()->as_register());
  72   __ jmpb(do_return);
  73 
  74   // input is NaN -> return 0
  75   __ bind(NaN);
  76   __ xorptr(result()->as_register(), result()->as_register());
  77 
  78   __ bind(do_return);
  79   __ jmp(_continuation);
  80 }
  81 
  82 void CounterOverflowStub::emit_code(LIR_Assembler* ce) {
  83   __ bind(_entry);
  84   ce->store_parameter(_method->as_register(), 1);
  85   ce->store_parameter(_bci, 0);
  86   __ call(RuntimeAddress(Runtime1::entry_for(Runtime1::counter_overflow_id)));
  87   ce->add_call_info_here(_info);
  88   ce->verify_oop_map(_info);
  89   __ jmp(_continuation);
  90 }
  91 
  92 RangeCheckStub::RangeCheckStub(CodeEmitInfo* info, LIR_Opr index,
  93                                bool throw_index_out_of_bounds_exception)
  94   : _throw_index_out_of_bounds_exception(throw_index_out_of_bounds_exception)
  95   , _index(index)
  96 {
  97   assert(info != NULL, "must have info");
  98   _info = new CodeEmitInfo(info);
  99 }
 100 
 101 
 102 void RangeCheckStub::emit_code(LIR_Assembler* ce) {
 103   __ bind(_entry);








 104   // pass the array index on stack because all registers must be preserved
 105   if (_index->is_cpu_register()) {
 106     ce->store_parameter(_index->as_register(), 0);
 107   } else {
 108     ce->store_parameter(_index->as_jint(), 0);
 109   }
 110   Runtime1::StubID stub_id;
 111   if (_throw_index_out_of_bounds_exception) {
 112     stub_id = Runtime1::throw_index_exception_id;
 113   } else {
 114     stub_id = Runtime1::throw_range_check_failed_id;
 115   }
 116   __ call(RuntimeAddress(Runtime1::entry_for(stub_id)));
 117   ce->add_call_info_here(_info);
 118   debug_only(__ should_not_reach_here());
 119 }
 120 












 121 
 122 void DivByZeroStub::emit_code(LIR_Assembler* ce) {
 123   if (_offset != -1) {
 124     ce->compilation()->implicit_exception_table()->append(_offset, __ offset());
 125   }
 126   __ bind(_entry);
 127   __ call(RuntimeAddress(Runtime1::entry_for(Runtime1::throw_div0_exception_id)));
 128   ce->add_call_info_here(_info);
 129   debug_only(__ should_not_reach_here());
 130 }
 131 
 132 
 133 // Implementation of NewInstanceStub
 134 
 135 NewInstanceStub::NewInstanceStub(LIR_Opr klass_reg, LIR_Opr result, ciInstanceKlass* klass, CodeEmitInfo* info, Runtime1::StubID stub_id) {
 136   _result = result;
 137   _klass = klass;
 138   _klass_reg = klass_reg;
 139   _info = new CodeEmitInfo(info);
 140   assert(stub_id == Runtime1::new_instance_id                 ||
 141          stub_id == Runtime1::fast_new_instance_id            ||
 142          stub_id == Runtime1::fast_new_instance_init_check_id,
 143          "need new_instance id");
 144   _stub_id   = stub_id;
 145 }
 146 
 147 
 148 void NewInstanceStub::emit_code(LIR_Assembler* ce) {
 149   assert(__ rsp_offset() == 0, "frame size should be fixed");
 150   __ bind(_entry);
 151   __ movptr(rdx, _klass_reg->as_register());
 152   __ call(RuntimeAddress(Runtime1::entry_for(_stub_id)));
 153   ce->add_call_info_here(_info);
 154   ce->verify_oop_map(_info);
 155   assert(_result->as_register() == rax, "result must in rax,");
 156   __ jmp(_continuation);
 157 }
 158 
 159 
 160 // Implementation of NewTypeArrayStub
 161 
 162 NewTypeArrayStub::NewTypeArrayStub(LIR_Opr klass_reg, LIR_Opr length, LIR_Opr result, CodeEmitInfo* info) {
 163   _klass_reg = klass_reg;
 164   _length = length;
 165   _result = result;
 166   _info = new CodeEmitInfo(info);
 167 }
 168 
 169 
 170 void NewTypeArrayStub::emit_code(LIR_Assembler* ce) {
 171   assert(__ rsp_offset() == 0, "frame size should be fixed");
 172   __ bind(_entry);
 173   assert(_length->as_register() == rbx, "length must in rbx,");
 174   assert(_klass_reg->as_register() == rdx, "klass_reg must in rdx");
 175   __ call(RuntimeAddress(Runtime1::entry_for(Runtime1::new_type_array_id)));
 176   ce->add_call_info_here(_info);
 177   ce->verify_oop_map(_info);
 178   assert(_result->as_register() == rax, "result must in rax,");
 179   __ jmp(_continuation);
 180 }
 181 
 182 
 183 // Implementation of NewObjectArrayStub
 184 
 185 NewObjectArrayStub::NewObjectArrayStub(LIR_Opr klass_reg, LIR_Opr length, LIR_Opr result, CodeEmitInfo* info) {
 186   _klass_reg = klass_reg;
 187   _result = result;
 188   _length = length;
 189   _info = new CodeEmitInfo(info);
 190 }
 191 
 192 
 193 void NewObjectArrayStub::emit_code(LIR_Assembler* ce) {
 194   assert(__ rsp_offset() == 0, "frame size should be fixed");
 195   __ bind(_entry);
 196   assert(_length->as_register() == rbx, "length must in rbx,");
 197   assert(_klass_reg->as_register() == rdx, "klass_reg must in rdx");
 198   __ call(RuntimeAddress(Runtime1::entry_for(Runtime1::new_object_array_id)));
 199   ce->add_call_info_here(_info);
 200   ce->verify_oop_map(_info);
 201   assert(_result->as_register() == rax, "result must in rax,");
 202   __ jmp(_continuation);
 203 }
 204 
 205 
 206 // Implementation of MonitorAccessStubs
 207 
 208 MonitorEnterStub::MonitorEnterStub(LIR_Opr obj_reg, LIR_Opr lock_reg, CodeEmitInfo* info)
 209 : MonitorAccessStub(obj_reg, lock_reg)
 210 {
 211   _info = new CodeEmitInfo(info);
 212 }
 213 
 214 
 215 void MonitorEnterStub::emit_code(LIR_Assembler* ce) {
 216   assert(__ rsp_offset() == 0, "frame size should be fixed");
 217   __ bind(_entry);
 218   ce->store_parameter(_obj_reg->as_register(),  1);
 219   ce->store_parameter(_lock_reg->as_register(), 0);
 220   Runtime1::StubID enter_id;
 221   if (ce->compilation()->has_fpu_code()) {
 222     enter_id = Runtime1::monitorenter_id;
 223   } else {
 224     enter_id = Runtime1::monitorenter_nofpu_id;
 225   }
 226   __ call(RuntimeAddress(Runtime1::entry_for(enter_id)));
 227   ce->add_call_info_here(_info);
 228   ce->verify_oop_map(_info);
 229   __ jmp(_continuation);
 230 }
 231 
 232 
 233 void MonitorExitStub::emit_code(LIR_Assembler* ce) {
 234   __ bind(_entry);
 235   if (_compute_lock) {
 236     // lock_reg was destroyed by fast unlocking attempt => recompute it
 237     ce->monitor_address(_monitor_ix, _lock_reg);
 238   }
 239   ce->store_parameter(_lock_reg->as_register(), 0);
 240   // note: non-blocking leaf routine => no call info needed
 241   Runtime1::StubID exit_id;
 242   if (ce->compilation()->has_fpu_code()) {
 243     exit_id = Runtime1::monitorexit_id;
 244   } else {
 245     exit_id = Runtime1::monitorexit_nofpu_id;
 246   }
 247   __ call(RuntimeAddress(Runtime1::entry_for(exit_id)));
 248   __ jmp(_continuation);
 249 }
 250 
 251 
 252 // Implementation of patching:
 253 // - Copy the code at given offset to an inlined buffer (first the bytes, then the number of bytes)
 254 // - Replace original code with a call to the stub
 255 // At Runtime:
 256 // - call to stub, jump to runtime
 257 // - in runtime: preserve all registers (rspecially objects, i.e., source and destination object)
 258 // - in runtime: after initializing class, restore original code, reexecute instruction
 259 
 260 int PatchingStub::_patch_info_offset = -NativeGeneralJump::instruction_size;
 261 
 262 void PatchingStub::align_patch_site(MacroAssembler* masm) {
 263   // We're patching a 5-7 byte instruction on intel and we need to
 264   // make sure that we don't see a piece of the instruction.  It
 265   // appears mostly impossible on Intel to simply invalidate other
 266   // processors caches and since they may do aggressive prefetch it's
 267   // very hard to make a guess about what code might be in the icache.
 268   // Force the instruction to be double word aligned so that it
 269   // doesn't span a cache line.
 270   masm->align(round_to(NativeGeneralJump::instruction_size, wordSize));
 271 }
 272 
 273 void PatchingStub::emit_code(LIR_Assembler* ce) {
 274   assert(NativeCall::instruction_size <= _bytes_to_copy && _bytes_to_copy <= 0xFF, "not enough room for call");
 275 
 276   Label call_patch;
 277 
 278   // static field accesses have special semantics while the class
 279   // initializer is being run so we emit a test which can be used to
 280   // check that this code is being executed by the initializing
 281   // thread.
 282   address being_initialized_entry = __ pc();
 283   if (CommentedAssembly) {
 284     __ block_comment(" patch template");
 285   }
 286   if (_id == load_klass_id) {
 287     // produce a copy of the load klass instruction for use by the being initialized case
 288 #ifdef ASSERT
 289     address start = __ pc();
 290 #endif
 291     Metadata* o = NULL;
 292     __ mov_metadata(_obj, o);
 293 #ifdef ASSERT
 294     for (int i = 0; i < _bytes_to_copy; i++) {
 295       address ptr = (address)(_pc_start + i);
 296       int a_byte = (*ptr) & 0xFF;
 297       assert(a_byte == *start++, "should be the same code");
 298     }
 299 #endif
 300   } else if (_id == load_mirror_id) {
 301     // produce a copy of the load mirror instruction for use by the being
 302     // initialized case
 303 #ifdef ASSERT
 304     address start = __ pc();
 305 #endif
 306     jobject o = NULL;
 307     __ movoop(_obj, o);
 308 #ifdef ASSERT
 309     for (int i = 0; i < _bytes_to_copy; i++) {
 310       address ptr = (address)(_pc_start + i);
 311       int a_byte = (*ptr) & 0xFF;
 312       assert(a_byte == *start++, "should be the same code");
 313     }
 314 #endif
 315   } else {
 316     // make a copy the code which is going to be patched.
 317     for (int i = 0; i < _bytes_to_copy; i++) {
 318       address ptr = (address)(_pc_start + i);
 319       int a_byte = (*ptr) & 0xFF;
 320       __ emit_int8(a_byte);
 321       *ptr = 0x90; // make the site look like a nop
 322     }
 323   }
 324 
 325   address end_of_patch = __ pc();
 326   int bytes_to_skip = 0;
 327   if (_id == load_mirror_id) {
 328     int offset = __ offset();
 329     if (CommentedAssembly) {
 330       __ block_comment(" being_initialized check");
 331     }
 332     assert(_obj != noreg, "must be a valid register");
 333     Register tmp = rax;
 334     Register tmp2 = rbx;
 335     __ push(tmp);
 336     __ push(tmp2);
 337     // Load without verification to keep code size small. We need it because
 338     // begin_initialized_entry_offset has to fit in a byte. Also, we know it's not null.
 339     __ movptr(tmp2, Address(_obj, java_lang_Class::klass_offset_in_bytes()));
 340     __ get_thread(tmp);
 341     __ cmpptr(tmp, Address(tmp2, InstanceKlass::init_thread_offset()));
 342     __ pop(tmp2);
 343     __ pop(tmp);
 344     __ jcc(Assembler::notEqual, call_patch);
 345 
 346     // access_field patches may execute the patched code before it's
 347     // copied back into place so we need to jump back into the main
 348     // code of the nmethod to continue execution.
 349     __ jmp(_patch_site_continuation);
 350 
 351     // make sure this extra code gets skipped
 352     bytes_to_skip += __ offset() - offset;
 353   }
 354   if (CommentedAssembly) {
 355     __ block_comment("patch data encoded as movl");
 356   }
 357   // Now emit the patch record telling the runtime how to find the
 358   // pieces of the patch.  We only need 3 bytes but for readability of
 359   // the disassembly we make the data look like a movl reg, imm32,
 360   // which requires 5 bytes
 361   int sizeof_patch_record = 5;
 362   bytes_to_skip += sizeof_patch_record;
 363 
 364   // emit the offsets needed to find the code to patch
 365   int being_initialized_entry_offset = __ pc() - being_initialized_entry + sizeof_patch_record;
 366 
 367   __ emit_int8((unsigned char)0xB8);
 368   __ emit_int8(0);
 369   __ emit_int8(being_initialized_entry_offset);
 370   __ emit_int8(bytes_to_skip);
 371   __ emit_int8(_bytes_to_copy);
 372   address patch_info_pc = __ pc();
 373   assert(patch_info_pc - end_of_patch == bytes_to_skip, "incorrect patch info");
 374 
 375   address entry = __ pc();
 376   NativeGeneralJump::insert_unconditional((address)_pc_start, entry);
 377   address target = NULL;
 378   relocInfo::relocType reloc_type = relocInfo::none;
 379   switch (_id) {
 380     case access_field_id:  target = Runtime1::entry_for(Runtime1::access_field_patching_id); break;
 381     case load_klass_id:    target = Runtime1::entry_for(Runtime1::load_klass_patching_id); reloc_type = relocInfo::metadata_type; break;
 382     case load_mirror_id:   target = Runtime1::entry_for(Runtime1::load_mirror_patching_id); reloc_type = relocInfo::oop_type; break;
 383     default: ShouldNotReachHere();
 384   }
 385   __ bind(call_patch);
 386 
 387   if (CommentedAssembly) {
 388     __ block_comment("patch entry point");
 389   }
 390   __ call(RuntimeAddress(target));
 391   assert(_patch_info_offset == (patch_info_pc - __ pc()), "must not change");
 392   ce->add_call_info_here(_info);
 393   int jmp_off = __ offset();
 394   __ jmp(_patch_site_entry);
 395   // Add enough nops so deoptimization can overwrite the jmp above with a call
 396   // and not destroy the world.
 397   for (int j = __ offset() ; j < jmp_off + 5 ; j++ ) {
 398     __ nop();
 399   }
 400   if (_id == load_klass_id || _id == load_mirror_id) {
 401     CodeSection* cs = __ code_section();
 402     RelocIterator iter(cs, (address)_pc_start, (address)(_pc_start + 1));
 403     relocInfo::change_reloc_info_for_address(&iter, (address) _pc_start, reloc_type, relocInfo::none);
 404   }
 405 }
 406 
 407 
 408 void DeoptimizeStub::emit_code(LIR_Assembler* ce) {
 409   __ bind(_entry);
 410   __ call(RuntimeAddress(Runtime1::entry_for(Runtime1::deoptimize_id)));
 411   ce->add_call_info_here(_info);
 412   DEBUG_ONLY(__ should_not_reach_here());
 413 }
 414 
 415 
 416 void ImplicitNullCheckStub::emit_code(LIR_Assembler* ce) {









 417   ce->compilation()->implicit_exception_table()->append(_offset, __ offset());
 418   __ bind(_entry);
 419   __ call(RuntimeAddress(Runtime1::entry_for(Runtime1::throw_null_pointer_exception_id)));
 420   ce->add_call_info_here(_info);
 421   debug_only(__ should_not_reach_here());
 422 }
 423 
 424 
 425 void SimpleExceptionStub::emit_code(LIR_Assembler* ce) {
 426   assert(__ rsp_offset() == 0, "frame size should be fixed");
 427 
 428   __ bind(_entry);
 429   // pass the object on stack because all registers must be preserved
 430   if (_obj->is_cpu_register()) {
 431     ce->store_parameter(_obj->as_register(), 0);
 432   }
 433   __ call(RuntimeAddress(Runtime1::entry_for(_stub)));
 434   ce->add_call_info_here(_info);
 435   debug_only(__ should_not_reach_here());
 436 }
 437 
 438 
 439 void ArrayCopyStub::emit_code(LIR_Assembler* ce) {
 440   //---------------slow case: call to native-----------------
 441   __ bind(_entry);
 442   // Figure out where the args should go
 443   // This should really convert the IntrinsicID to the Method* and signature
 444   // but I don't know how to do that.
 445   //
 446   VMRegPair args[5];
 447   BasicType signature[5] = { T_OBJECT, T_INT, T_OBJECT, T_INT, T_INT};
 448   SharedRuntime::java_calling_convention(signature, args, 5, true);
 449 
 450   // push parameters
 451   // (src, src_pos, dest, destPos, length)
 452   Register r[5];
 453   r[0] = src()->as_register();
 454   r[1] = src_pos()->as_register();
 455   r[2] = dst()->as_register();
 456   r[3] = dst_pos()->as_register();
 457   r[4] = length()->as_register();
 458 
 459   // next registers will get stored on the stack
 460   for (int i = 0; i < 5 ; i++ ) {
 461     VMReg r_1 = args[i].first();
 462     if (r_1->is_stack()) {
 463       int st_off = r_1->reg2stack() * wordSize;
 464       __ movptr (Address(rsp, st_off), r[i]);
 465     } else {
 466       assert(r[i] == args[i].first()->as_Register(), "Wrong register for arg ");
 467     }
 468   }
 469 
 470   ce->align_call(lir_static_call);
 471 
 472   ce->emit_static_call_stub();
 473   AddressLiteral resolve(SharedRuntime::get_resolve_static_call_stub(),
 474                          relocInfo::static_call_type);
 475   __ call(resolve);
 476   ce->add_call_info_here(info());
 477 
 478 #ifndef PRODUCT
 479   __ incrementl(ExternalAddress((address)&Runtime1::_arraycopy_slowcase_cnt));
 480 #endif
 481 
 482   __ jmp(_continuation);
 483 }
 484 
 485 /////////////////////////////////////////////////////////////////////////////
 486 #if INCLUDE_ALL_GCS
 487 
 488 void G1PreBarrierStub::emit_code(LIR_Assembler* ce) {
 489   // At this point we know that marking is in progress.
 490   // If do_load() is true then we have to emit the
 491   // load of the previous value; otherwise it has already
 492   // been loaded into _pre_val.
 493 
 494   __ bind(_entry);
 495   assert(pre_val()->is_register(), "Precondition.");
 496 
 497   Register pre_val_reg = pre_val()->as_register();
 498 
 499   if (do_load()) {
 500     ce->mem2reg(addr(), pre_val(), T_OBJECT, patch_code(), info(), false /*wide*/, false /*unaligned*/);
 501   }
 502 
 503   __ cmpptr(pre_val_reg, (int32_t) NULL_WORD);
 504   __ jcc(Assembler::equal, _continuation);
 505   ce->store_parameter(pre_val()->as_register(), 0);
 506   __ call(RuntimeAddress(Runtime1::entry_for(Runtime1::g1_pre_barrier_slow_id)));
 507   __ jmp(_continuation);
 508 
 509 }
 510 
 511 jbyte* G1PostBarrierStub::_byte_map_base = NULL;
 512 
 513 jbyte* G1PostBarrierStub::byte_map_base_slow() {
 514   BarrierSet* bs = Universe::heap()->barrier_set();
 515   assert(bs->is_a(BarrierSet::G1SATBCTLogging),
 516          "Must be if we're using this.");
 517   return ((G1SATBCardTableModRefBS*)bs)->byte_map_base;
 518 }
 519 
 520 void G1PostBarrierStub::emit_code(LIR_Assembler* ce) {
 521   __ bind(_entry);
 522   assert(addr()->is_register(), "Precondition.");
 523   assert(new_val()->is_register(), "Precondition.");
 524   Register new_val_reg = new_val()->as_register();
 525   __ cmpptr(new_val_reg, (int32_t) NULL_WORD);
 526   __ jcc(Assembler::equal, _continuation);
 527   ce->store_parameter(addr()->as_pointer_register(), 0);
 528   __ call(RuntimeAddress(Runtime1::entry_for(Runtime1::g1_post_barrier_slow_id)));
 529   __ jmp(_continuation);
 530 }
 531 
 532 #endif // INCLUDE_ALL_GCS
 533 /////////////////////////////////////////////////////////////////////////////
 534 
 535 #undef __
--- EOF ---