rev 12409 : 8169177: aarch64: SIGSEGV when "-XX:+ZeroTLAB" is specified along with GC options
Summary: Add zero-initialization to C1 for fast TLAB refills
Reviewed-by: aph, drwhite
Contributed-by: kavitha.natarajan@linaro.org

   1 /*
   2  * Copyright (c) 1999, 2015, Oracle and/or its affiliates. All rights reserved.
   3  * Copyright (c) 2014, Red Hat Inc. All rights reserved.
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This code is free software; you can redistribute it and/or modify it
   7  * under the terms of the GNU General Public License version 2 only, as
   8  * published by the Free Software Foundation.
   9  *
  10  * This code is distributed in the hope that it will be useful, but WITHOUT
  11  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  12  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  13  * version 2 for more details (a copy is included in the LICENSE file that
  14  * accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License version
  17  * 2 along with this work; if not, write to the Free Software Foundation,
  18  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  19  *
  20  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  21  * or visit www.oracle.com if you need additional information or have any
  22  * questions.
  23  *
  24  */
  25 
  26 #include "precompiled.hpp"
  27 #include "c1/c1_MacroAssembler.hpp"
  28 #include "c1/c1_Runtime1.hpp"
  29 #include "classfile/systemDictionary.hpp"
  30 #include "gc/shared/collectedHeap.hpp"
  31 #include "interpreter/interpreter.hpp"
  32 #include "oops/arrayOop.hpp"
  33 #include "oops/markOop.hpp"
  34 #include "runtime/basicLock.hpp"
  35 #include "runtime/biasedLocking.hpp"
  36 #include "runtime/os.hpp"
  37 #include "runtime/sharedRuntime.hpp"
  38 #include "runtime/stubRoutines.hpp"
  39 
  40 void C1_MacroAssembler::float_cmp(bool is_float, int unordered_result,
  41                                   FloatRegister f0, FloatRegister f1,
  42                                   Register result)
  43 {
  44   Label done;
  45   if (is_float) {
  46     fcmps(f0, f1);
  47   } else {
  48     fcmpd(f0, f1);
  49   }
  50   if (unordered_result < 0) {
  51     // we want -1 for unordered or less than, 0 for equal and 1 for
  52     // greater than.
  53     cset(result, NE);  // Not equal or unordered
  54     cneg(result, result, LT);  // Less than or unordered
  55   } else {
  56     // we want -1 for less than, 0 for equal and 1 for unordered or
  57     // greater than.
  58     cset(result, NE);  // Not equal or unordered
  59     cneg(result, result, LO);  // Less than
  60   }
  61 }
  62 
  63 int C1_MacroAssembler::lock_object(Register hdr, Register obj, Register disp_hdr, Register scratch, Label& slow_case) {
  64   const int aligned_mask = BytesPerWord -1;
  65   const int hdr_offset = oopDesc::mark_offset_in_bytes();
  66   assert(hdr != obj && hdr != disp_hdr && obj != disp_hdr, "registers must be different");
  67   Label done, fail;
  68   int null_check_offset = -1;
  69 
  70   verify_oop(obj);
  71 
  72   // save object being locked into the BasicObjectLock
  73   str(obj, Address(disp_hdr, BasicObjectLock::obj_offset_in_bytes()));
  74 
  75   if (UseBiasedLocking) {
  76     assert(scratch != noreg, "should have scratch register at this point");
  77     null_check_offset = biased_locking_enter(disp_hdr, obj, hdr, scratch, false, done, &slow_case);
  78   } else {
  79     null_check_offset = offset();
  80   }
  81 
  82   // Load object header
  83   ldr(hdr, Address(obj, hdr_offset));
  84   // and mark it as unlocked
  85   orr(hdr, hdr, markOopDesc::unlocked_value);
  86   // save unlocked object header into the displaced header location on the stack
  87   str(hdr, Address(disp_hdr, 0));
  88   // test if object header is still the same (i.e. unlocked), and if so, store the
  89   // displaced header address in the object header - if it is not the same, get the
  90   // object header instead
  91   lea(rscratch2, Address(obj, hdr_offset));
  92   cmpxchgptr(hdr, disp_hdr, rscratch2, rscratch1, done, /*fallthough*/NULL);
  93   // if the object header was the same, we're done
  94   // if the object header was not the same, it is now in the hdr register
  95   // => test if it is a stack pointer into the same stack (recursive locking), i.e.:
  96   //
  97   // 1) (hdr & aligned_mask) == 0
  98   // 2) sp <= hdr
  99   // 3) hdr <= sp + page_size
 100   //
 101   // these 3 tests can be done by evaluating the following expression:
 102   //
 103   // (hdr - sp) & (aligned_mask - page_size)
 104   //
 105   // assuming both the stack pointer and page_size have their least
 106   // significant 2 bits cleared and page_size is a power of 2
 107   mov(rscratch1, sp);
 108   sub(hdr, hdr, rscratch1);
 109   ands(hdr, hdr, aligned_mask - os::vm_page_size());
 110   // for recursive locking, the result is zero => save it in the displaced header
 111   // location (NULL in the displaced hdr location indicates recursive locking)
 112   str(hdr, Address(disp_hdr, 0));
 113   // otherwise we don't care about the result and handle locking via runtime call
 114   cbnz(hdr, slow_case);
 115   // done
 116   bind(done);
 117   if (PrintBiasedLockingStatistics) {
 118     lea(rscratch2, ExternalAddress((address)BiasedLocking::fast_path_entry_count_addr()));
 119     addmw(Address(rscratch2, 0), 1, rscratch1);
 120   }
 121   return null_check_offset;
 122 }
 123 
 124 
 125 void C1_MacroAssembler::unlock_object(Register hdr, Register obj, Register disp_hdr, Label& slow_case) {
 126   const int aligned_mask = BytesPerWord -1;
 127   const int hdr_offset = oopDesc::mark_offset_in_bytes();
 128   assert(hdr != obj && hdr != disp_hdr && obj != disp_hdr, "registers must be different");
 129   Label done;
 130 
 131   if (UseBiasedLocking) {
 132     // load object
 133     ldr(obj, Address(disp_hdr, BasicObjectLock::obj_offset_in_bytes()));
 134     biased_locking_exit(obj, hdr, done);
 135   }
 136 
 137   // load displaced header
 138   ldr(hdr, Address(disp_hdr, 0));
 139   // if the loaded hdr is NULL we had recursive locking
 140   // if we had recursive locking, we are done
 141   cbz(hdr, done);
 142   if (!UseBiasedLocking) {
 143     // load object
 144     ldr(obj, Address(disp_hdr, BasicObjectLock::obj_offset_in_bytes()));
 145   }
 146   verify_oop(obj);
 147   // test if object header is pointing to the displaced header, and if so, restore
 148   // the displaced header in the object - if the object header is not pointing to
 149   // the displaced header, get the object header instead
 150   // if the object header was not pointing to the displaced header,
 151   // we do unlocking via runtime call
 152   if (hdr_offset) {
 153     lea(rscratch1, Address(obj, hdr_offset));
 154     cmpxchgptr(disp_hdr, hdr, rscratch1, rscratch2, done, &slow_case);
 155   } else {
 156     cmpxchgptr(disp_hdr, hdr, obj, rscratch2, done, &slow_case);
 157   }
 158   // done
 159   bind(done);
 160 }
 161 
 162 
 163 // Defines obj, preserves var_size_in_bytes
 164 void C1_MacroAssembler::try_allocate(Register obj, Register var_size_in_bytes, int con_size_in_bytes, Register t1, Register t2, Label& slow_case) {
 165   if (UseTLAB) {
 166     tlab_allocate(obj, var_size_in_bytes, con_size_in_bytes, t1, t2, slow_case);
 167   } else {
 168     eden_allocate(obj, var_size_in_bytes, con_size_in_bytes, t1, slow_case);
 169     incr_allocated_bytes(noreg, var_size_in_bytes, con_size_in_bytes, t1);
 170   }
 171 }
 172 
 173 void C1_MacroAssembler::initialize_header(Register obj, Register klass, Register len, Register t1, Register t2) {
 174   assert_different_registers(obj, klass, len);
 175   if (UseBiasedLocking && !len->is_valid()) {
 176     assert_different_registers(obj, klass, len, t1, t2);
 177     ldr(t1, Address(klass, Klass::prototype_header_offset()));
 178   } else {
 179     // This assumes that all prototype bits fit in an int32_t
 180     mov(t1, (int32_t)(intptr_t)markOopDesc::prototype());
 181   }
 182   str(t1, Address(obj, oopDesc::mark_offset_in_bytes()));
 183 
 184   if (UseCompressedClassPointers) { // Take care not to kill klass
 185     encode_klass_not_null(t1, klass);
 186     strw(t1, Address(obj, oopDesc::klass_offset_in_bytes()));
 187   } else {
 188     str(klass, Address(obj, oopDesc::klass_offset_in_bytes()));
 189   }
 190 
 191   if (len->is_valid()) {
 192     strw(len, Address(obj, arrayOopDesc::length_offset_in_bytes()));
 193   } else if (UseCompressedClassPointers) {
 194     store_klass_gap(obj, zr);
 195   }
 196 }
 197 
 198 // Zero words; len is in bytes
 199 // Destroys all registers except addr
 200 // len must be a nonzero multiple of wordSize
 201 void C1_MacroAssembler::zero_memory(Register addr, Register len, Register t1) {
 202   assert_different_registers(addr, len, t1, rscratch1, rscratch2);
 203 
 204 #ifdef ASSERT
 205   { Label L;
 206     tst(len, BytesPerWord - 1);
 207     br(Assembler::EQ, L);
 208     stop("len is not a multiple of BytesPerWord");
 209     bind(L);
 210   }
 211 #endif
 212 
 213 #ifndef PRODUCT
 214   block_comment("zero memory");
 215 #endif
 216 
 217   Label loop;
 218   Label entry;
 219 
 220 //  Algorithm:
 221 //
 222 //    scratch1 = cnt & 7;
 223 //    cnt -= scratch1;
 224 //    p += scratch1;
 225 //    switch (scratch1) {
 226 //      do {
 227 //        cnt -= 8;
 228 //          p[-8] = 0;
 229 //        case 7:
 230 //          p[-7] = 0;
 231 //        case 6:
 232 //          p[-6] = 0;
 233 //          // ...
 234 //        case 1:
 235 //          p[-1] = 0;
 236 //        case 0:
 237 //          p += 8;
 238 //      } while (cnt);
 239 //    }
 240 
 241   const int unroll = 8; // Number of str(zr) instructions we'll unroll
 242 
 243   lsr(len, len, LogBytesPerWord);
 244   andr(rscratch1, len, unroll - 1);  // tmp1 = cnt % unroll
 245   sub(len, len, rscratch1);      // cnt -= unroll
 246   // t1 always points to the end of the region we're about to zero
 247   add(t1, addr, rscratch1, Assembler::LSL, LogBytesPerWord);
 248   adr(rscratch2, entry);
 249   sub(rscratch2, rscratch2, rscratch1, Assembler::LSL, 2);
 250   br(rscratch2);
 251   bind(loop);
 252   sub(len, len, unroll);
 253   for (int i = -unroll; i < 0; i++)
 254     str(zr, Address(t1, i * wordSize));
 255   bind(entry);
 256   add(t1, t1, unroll * wordSize);
 257   cbnz(len, loop);
 258 }
 259 
 260 // preserves obj, destroys len_in_bytes
 261 void C1_MacroAssembler::initialize_body(Register obj, Register len_in_bytes, int hdr_size_in_bytes, Register t1) {

 262   Label done;
 263   assert(obj != len_in_bytes && obj != t1 && t1 != len_in_bytes, "registers must be different");
 264   assert((hdr_size_in_bytes & (BytesPerWord - 1)) == 0, "header size is not a multiple of BytesPerWord");
 265   Register index = len_in_bytes;
 266   // index is positive and ptr sized
 267   subs(index, index, hdr_size_in_bytes);
 268   br(Assembler::EQ, done);
 269   // note: for the remaining code to work, index must be a multiple of BytesPerWord
 270 #ifdef ASSERT
 271   { Label L;
 272     tst(index, BytesPerWord - 1);
 273     br(Assembler::EQ, L);
 274     stop("index is not a multiple of BytesPerWord");
 275     bind(L);
 276   }
 277 #endif
 278 
 279   // Preserve obj
 280   if (hdr_size_in_bytes)
 281     add(obj, obj, hdr_size_in_bytes);
 282   zero_memory(obj, index, t1);
 283   if (hdr_size_in_bytes)
 284     sub(obj, obj, hdr_size_in_bytes);
 285 
 286   // done
 287   bind(done);
 288 }
 289 
 290 
 291 void C1_MacroAssembler::allocate_object(Register obj, Register t1, Register t2, int header_size, int object_size, Register klass, Label& slow_case) {
 292   assert_different_registers(obj, t1, t2); // XXX really?
 293   assert(header_size >= 0 && object_size >= header_size, "illegal sizes");
 294 
 295   try_allocate(obj, noreg, object_size * BytesPerWord, t1, t2, slow_case);
 296 
 297   initialize_object(obj, klass, noreg, object_size * HeapWordSize, t1, t2);
 298 }
 299 
 300 void C1_MacroAssembler::initialize_object(Register obj, Register klass, Register var_size_in_bytes, int con_size_in_bytes, Register t1, Register t2) {
 301   assert((con_size_in_bytes & MinObjAlignmentInBytesMask) == 0,
 302          "con_size_in_bytes is not multiple of alignment");
 303   const int hdr_size_in_bytes = instanceOopDesc::header_size() * HeapWordSize;
 304 
 305   initialize_header(obj, klass, noreg, t1, t2);
 306 

 307   // clear rest of allocated space
 308   const Register index = t2;
 309   const int threshold = 16 * BytesPerWord;   // approximate break even point for code size (see comments below)
 310   if (var_size_in_bytes != noreg) {
 311     mov(index, var_size_in_bytes);
 312     initialize_body(obj, index, hdr_size_in_bytes, t1);
 313   } else if (con_size_in_bytes <= threshold) {
 314     // use explicit null stores
 315     int i = hdr_size_in_bytes;
 316     if (i < con_size_in_bytes && (con_size_in_bytes % (2 * BytesPerWord))) {
 317       str(zr, Address(obj, i));
 318       i += BytesPerWord;
 319     }
 320     for (; i < con_size_in_bytes; i += 2 * BytesPerWord)
 321       stp(zr, zr, Address(obj, i));
 322   } else if (con_size_in_bytes > hdr_size_in_bytes) {
 323     block_comment("zero memory");
 324     // use loop to null out the fields
 325 
 326     int words = (con_size_in_bytes - hdr_size_in_bytes) / BytesPerWord;
 327     mov(index,  words / 8);
 328 
 329     const int unroll = 8; // Number of str(zr) instructions we'll unroll
 330     int remainder = words % unroll;
 331     lea(rscratch1, Address(obj, hdr_size_in_bytes + remainder * BytesPerWord));
 332 
 333     Label entry_point, loop;
 334     b(entry_point);
 335 
 336     bind(loop);
 337     sub(index, index, 1);
 338     for (int i = -unroll; i < 0; i++) {
 339       if (-i == remainder)
 340         bind(entry_point);
 341       str(zr, Address(rscratch1, i * wordSize));
 342     }
 343     if (remainder == 0)
 344       bind(entry_point);
 345     add(rscratch1, rscratch1, unroll * wordSize);
 346     cbnz(index, loop);
 347 

 348   }
 349 
 350   membar(StoreStore);
 351 
 352   if (CURRENT_ENV->dtrace_alloc_probes()) {
 353     assert(obj == r0, "must be");
 354     far_call(RuntimeAddress(Runtime1::entry_for(Runtime1::dtrace_object_alloc_id)));
 355   }
 356 
 357   verify_oop(obj);
 358 }
 359 void C1_MacroAssembler::allocate_array(Register obj, Register len, Register t1, Register t2, int header_size, int f, Register klass, Label& slow_case) {
 360   assert_different_registers(obj, len, t1, t2, klass);
 361 
 362   // determine alignment mask
 363   assert(!(BytesPerWord & 1), "must be a multiple of 2 for masking code to work");
 364 
 365   // check for negative or excessive length
 366   mov(rscratch1, (int32_t)max_array_allocation_length);
 367   cmp(len, rscratch1);
 368   br(Assembler::HS, slow_case);
 369 
 370   const Register arr_size = t2; // okay to be the same
 371   // align object end
 372   mov(arr_size, (int32_t)header_size * BytesPerWord + MinObjAlignmentInBytesMask);
 373   add(arr_size, arr_size, len, ext::uxtw, f);
 374   andr(arr_size, arr_size, ~MinObjAlignmentInBytesMask);
 375 
 376   try_allocate(obj, arr_size, 0, t1, t2, slow_case);
 377 
 378   initialize_header(obj, klass, len, t1, t2);
 379 
 380   // clear rest of allocated space
 381   const Register len_zero = len;
 382   initialize_body(obj, arr_size, header_size * BytesPerWord, len_zero);
 383 
 384   membar(StoreStore);
 385 
 386   if (CURRENT_ENV->dtrace_alloc_probes()) {
 387     assert(obj == r0, "must be");
 388     far_call(RuntimeAddress(Runtime1::entry_for(Runtime1::dtrace_object_alloc_id)));
 389   }
 390 
 391   verify_oop(obj);
 392 }
 393 
 394 
 395 void C1_MacroAssembler::inline_cache_check(Register receiver, Register iCache) {
 396   verify_oop(receiver);
 397   // explicit NULL check not needed since load from [klass_offset] causes a trap
 398   // check against inline cache
 399   assert(!MacroAssembler::needs_explicit_null_check(oopDesc::klass_offset_in_bytes()), "must add explicit null check");
 400 
 401   cmp_klass(receiver, iCache, rscratch1);
 402 }
 403 
 404 
 405 void C1_MacroAssembler::build_frame(int framesize, int bang_size_in_bytes) {
 406   // If we have to make this method not-entrant we'll overwrite its
 407   // first instruction with a jump.  For this action to be legal we
 408   // must ensure that this first instruction is a B, BL, NOP, BKPT,
 409   // SVC, HVC, or SMC.  Make it a NOP.
 410   nop();
 411   assert(bang_size_in_bytes >= framesize, "stack bang size incorrect");
 412   // Make sure there is enough stack space for this method's activation.
 413   // Note that we do this before doing an enter().
 414   generate_stack_overflow_check(bang_size_in_bytes);
 415   MacroAssembler::build_frame(framesize + 2 * wordSize);
 416   if (NotifySimulator) {
 417     notify(Assembler::method_entry);
 418   }
 419 }
 420 
 421 void C1_MacroAssembler::remove_frame(int framesize) {
 422   MacroAssembler::remove_frame(framesize + 2 * wordSize);
 423   if (NotifySimulator) {
 424     notify(Assembler::method_reentry);
 425   }
 426 }
 427 
 428 
 429 void C1_MacroAssembler::verified_entry() {
 430 }
 431 
 432 #ifndef PRODUCT
 433 
 434 void C1_MacroAssembler::verify_stack_oop(int stack_offset) {
 435   if (!VerifyOops) return;
 436   verify_oop_addr(Address(sp, stack_offset), "oop");
 437 }
 438 
 439 void C1_MacroAssembler::verify_not_null_oop(Register r) {
 440   if (!VerifyOops) return;
 441   Label not_null;
 442   cbnz(r, not_null);
 443   stop("non-null oop required");
 444   bind(not_null);
 445   verify_oop(r);
 446 }
 447 
 448 void C1_MacroAssembler::invalidate_registers(bool inv_r0, bool inv_r19, bool inv_r2, bool inv_r3, bool inv_r4, bool inv_r5) {
 449 #ifdef ASSERT
 450   static int nn;
 451   if (inv_r0) mov(r0, 0xDEAD);
 452   if (inv_r19) mov(r19, 0xDEAD);
 453   if (inv_r2) mov(r2, nn++);
 454   if (inv_r3) mov(r3, 0xDEAD);
 455   if (inv_r4) mov(r4, 0xDEAD);
 456   if (inv_r5) mov(r5, 0xDEAD);
 457 #endif
 458 }
 459 #endif // ifndef PRODUCT
--- EOF ---