1 /*
   2  * Copyright (c) 1997, 2017, Oracle and/or its affiliates. All rights reserved.
   3  * Copyright (c) 2012, 2014 SAP SE. 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 #ifndef OS_CPU_AIX_OJDKPPC_VM_ATOMIC_AIX_PPC_HPP
  27 #define OS_CPU_AIX_OJDKPPC_VM_ATOMIC_AIX_PPC_HPP
  28 
  29 #ifndef _LP64
  30 #error "Atomic currently only impleneted for PPC64"
  31 #endif
  32 
  33 // Implementation of class atomic
  34 
  35 inline void Atomic::store    (jbyte    store_value, jbyte*    dest) { *dest = store_value; }
  36 inline void Atomic::store    (jshort   store_value, jshort*   dest) { *dest = store_value; }
  37 inline void Atomic::store    (jint     store_value, jint*     dest) { *dest = store_value; }
  38 inline void Atomic::store    (jlong    store_value, jlong*    dest) { *dest = store_value; }
  39 inline void Atomic::store_ptr(intptr_t store_value, intptr_t* dest) { *dest = store_value; }
  40 inline void Atomic::store_ptr(void*    store_value, void*     dest) { *(void**)dest = store_value; }
  41 
  42 inline void Atomic::store    (jbyte    store_value, volatile jbyte*    dest) { *dest = store_value; }
  43 inline void Atomic::store    (jshort   store_value, volatile jshort*   dest) { *dest = store_value; }
  44 inline void Atomic::store    (jint     store_value, volatile jint*     dest) { *dest = store_value; }
  45 inline void Atomic::store    (jlong    store_value, volatile jlong*    dest) { *dest = store_value; }
  46 inline void Atomic::store_ptr(intptr_t store_value, volatile intptr_t* dest) { *dest = store_value; }
  47 inline void Atomic::store_ptr(void*    store_value, volatile void*     dest) { *(void* volatile *)dest = store_value; }
  48 
  49 inline jlong Atomic::load(const volatile jlong* src) { return *src; }
  50 
  51 //
  52 //   machine barrier instructions:
  53 //
  54 //   - ppc_sync            two-way memory barrier, aka fence
  55 //   - ppc_lwsync          orders  Store|Store,
  56 //                                  Load|Store,
  57 //                                  Load|Load,
  58 //                         but not Store|Load
  59 //   - ppc_eieio           orders memory accesses for device memory (only)
  60 //   - ppc_isync           invalidates speculatively executed instructions
  61 //                         From the POWER ISA 2.06 documentation:
  62 //                          "[...] an isync instruction prevents the execution of
  63 //                         instructions following the isync until instructions
  64 //                         preceding the isync have completed, [...]"
  65 //                         From IBM's AIX assembler reference:
  66 //                          "The isync [...] instructions causes the processor to
  67 //                         refetch any instructions that might have been fetched
  68 //                         prior to the isync instruction. The instruction isync
  69 //                         causes the processor to wait for all previous instructions
  70 //                         to complete. Then any instructions already fetched are
  71 //                         discarded and instruction processing continues in the
  72 //                         environment established by the previous instructions."
  73 //
  74 //   semantic barrier instructions:
  75 //   (as defined in orderAccess.hpp)
  76 //
  77 //   - ppc_release         orders Store|Store,       (maps to ppc_lwsync)
  78 //                                 Load|Store
  79 //   - ppc_acquire         orders  Load|Store,       (maps to ppc_lwsync)
  80 //                                 Load|Load
  81 //   - ppc_fence           orders Store|Store,       (maps to ppc_sync)
  82 //                                 Load|Store,
  83 //                                 Load|Load,
  84 //                                Store|Load
  85 //
  86 
  87 #define strasm_sync                       "\n  sync    \n"
  88 #define strasm_lwsync                     "\n  lwsync  \n"
  89 #define strasm_isync                      "\n  isync   \n"
  90 #define strasm_release                    strasm_lwsync
  91 #define strasm_acquire                    strasm_lwsync
  92 #define strasm_fence                      strasm_sync
  93 #define strasm_nobarrier                  ""
  94 #define strasm_nobarrier_clobber_memory   ""
  95 
  96 inline jint     Atomic::add    (jint     add_value, volatile jint*     dest) {
  97 
  98   unsigned int result;
  99 
 100   __asm__ __volatile__ (
 101     strasm_lwsync
 102     "1: lwarx   %0,  0, %2    \n"
 103     "   add     %0, %0, %1    \n"
 104     "   stwcx.  %0,  0, %2    \n"
 105     "   bne-    1b            \n"
 106     strasm_isync
 107     : /*%0*/"=&r" (result)
 108     : /*%1*/"r" (add_value), /*%2*/"r" (dest)
 109     : "cc", "memory" );
 110 
 111   return (jint) result;
 112 }
 113 
 114 
 115 inline intptr_t Atomic::add_ptr(intptr_t add_value, volatile intptr_t* dest) {
 116 
 117   long result;
 118 
 119   __asm__ __volatile__ (
 120     strasm_lwsync
 121     "1: ldarx   %0,  0, %2    \n"
 122     "   add     %0, %0, %1    \n"
 123     "   stdcx.  %0,  0, %2    \n"
 124     "   bne-    1b            \n"
 125     strasm_isync
 126     : /*%0*/"=&r" (result)
 127     : /*%1*/"r" (add_value), /*%2*/"r" (dest)
 128     : "cc", "memory" );
 129 
 130   return (intptr_t) result;
 131 }
 132 
 133 inline void*    Atomic::add_ptr(intptr_t add_value, volatile void*     dest) {
 134   return (void*)add_ptr(add_value, (volatile intptr_t*)dest);
 135 }
 136 
 137 
 138 inline void Atomic::inc    (volatile jint*     dest) {
 139 
 140   unsigned int temp;
 141 
 142   __asm__ __volatile__ (
 143     strasm_nobarrier
 144     "1: lwarx   %0,  0, %2    \n"
 145     "   addic   %0, %0,  1    \n"
 146     "   stwcx.  %0,  0, %2    \n"
 147     "   bne-    1b            \n"
 148     strasm_nobarrier
 149     : /*%0*/"=&r" (temp), "=m" (*dest)
 150     : /*%2*/"r" (dest), "m" (*dest)
 151     : "cc" strasm_nobarrier_clobber_memory);
 152 
 153 }
 154 
 155 inline void Atomic::inc_ptr(volatile intptr_t* dest) {
 156 
 157   long temp;
 158 
 159   __asm__ __volatile__ (
 160     strasm_nobarrier
 161     "1: ldarx   %0,  0, %2    \n"
 162     "   addic   %0, %0,  1    \n"
 163     "   stdcx.  %0,  0, %2    \n"
 164     "   bne-    1b            \n"
 165     strasm_nobarrier
 166     : /*%0*/"=&r" (temp), "=m" (*dest)
 167     : /*%2*/"r" (dest), "m" (*dest)
 168     : "cc" strasm_nobarrier_clobber_memory);
 169 
 170 }
 171 
 172 inline void Atomic::inc_ptr(volatile void*     dest) {
 173   inc_ptr((volatile intptr_t*)dest);
 174 }
 175 
 176 
 177 inline void Atomic::dec    (volatile jint*     dest) {
 178 
 179   unsigned int temp;
 180 
 181   __asm__ __volatile__ (
 182     strasm_nobarrier
 183     "1: lwarx   %0,  0, %2    \n"
 184     "   addic   %0, %0, -1    \n"
 185     "   stwcx.  %0,  0, %2    \n"
 186     "   bne-    1b            \n"
 187     strasm_nobarrier
 188     : /*%0*/"=&r" (temp), "=m" (*dest)
 189     : /*%2*/"r" (dest), "m" (*dest)
 190     : "cc" strasm_nobarrier_clobber_memory);
 191 
 192 }
 193 
 194 inline void Atomic::dec_ptr(volatile intptr_t* dest) {
 195 
 196   long temp;
 197 
 198   __asm__ __volatile__ (
 199     strasm_nobarrier
 200     "1: ldarx   %0,  0, %2    \n"
 201     "   addic   %0, %0, -1    \n"
 202     "   stdcx.  %0,  0, %2    \n"
 203     "   bne-    1b            \n"
 204     strasm_nobarrier
 205     : /*%0*/"=&r" (temp), "=m" (*dest)
 206     : /*%2*/"r" (dest), "m" (*dest)
 207     : "cc" strasm_nobarrier_clobber_memory);
 208 
 209 }
 210 
 211 inline void Atomic::dec_ptr(volatile void*     dest) {
 212   dec_ptr((volatile intptr_t*)dest);
 213 }
 214 
 215 inline jint Atomic::xchg(jint exchange_value, volatile jint* dest) {
 216 
 217   // Note that xchg_ptr doesn't necessarily do an acquire
 218   // (see synchronizer.cpp).
 219 
 220   unsigned int old_value;
 221   const uint64_t zero = 0;
 222 
 223   __asm__ __volatile__ (
 224     /* lwsync */
 225     strasm_lwsync
 226     /* atomic loop */
 227     "1:                                                 \n"
 228     "   lwarx   %[old_value], %[dest], %[zero]          \n"
 229     "   stwcx.  %[exchange_value], %[dest], %[zero]     \n"
 230     "   bne-    1b                                      \n"
 231     /* isync */
 232     strasm_sync
 233     /* exit */
 234     "2:                                                 \n"
 235     /* out */
 236     : [old_value]       "=&r"   (old_value),
 237                         "=m"    (*dest)
 238     /* in */
 239     : [dest]            "b"     (dest),
 240       [zero]            "r"     (zero),
 241       [exchange_value]  "r"     (exchange_value),
 242                         "m"     (*dest)
 243     /* clobber */
 244     : "cc",
 245       "memory"
 246     );
 247 
 248   return (jint) old_value;
 249 }
 250 
 251 inline intptr_t Atomic::xchg_ptr(intptr_t exchange_value, volatile intptr_t* dest) {
 252 
 253   // Note that xchg_ptr doesn't necessarily do an acquire
 254   // (see synchronizer.cpp).
 255 
 256   long old_value;
 257   const uint64_t zero = 0;
 258 
 259   __asm__ __volatile__ (
 260     /* lwsync */
 261     strasm_lwsync
 262     /* atomic loop */
 263     "1:                                                 \n"
 264     "   ldarx   %[old_value], %[dest], %[zero]          \n"
 265     "   stdcx.  %[exchange_value], %[dest], %[zero]     \n"
 266     "   bne-    1b                                      \n"
 267     /* isync */
 268     strasm_sync
 269     /* exit */
 270     "2:                                                 \n"
 271     /* out */
 272     : [old_value]       "=&r"   (old_value),
 273                         "=m"    (*dest)
 274     /* in */
 275     : [dest]            "b"     (dest),
 276       [zero]            "r"     (zero),
 277       [exchange_value]  "r"     (exchange_value),
 278                         "m"     (*dest)
 279     /* clobber */
 280     : "cc",
 281       "memory"
 282     );
 283 
 284   return (intptr_t) old_value;
 285 }
 286 
 287 inline void* Atomic::xchg_ptr(void* exchange_value, volatile void* dest) {
 288   return (void*)xchg_ptr((intptr_t)exchange_value, (volatile intptr_t*)dest);
 289 }
 290 
 291 inline void cmpxchg_pre_membar(cmpxchg_memory_order order) {
 292   if (order != memory_order_relaxed) {
 293     __asm__ __volatile__ (
 294       /* fence */
 295       strasm_sync
 296       );
 297   }
 298 }
 299 
 300 inline void cmpxchg_post_membar(cmpxchg_memory_order order) {
 301   if (order != memory_order_relaxed) {
 302     __asm__ __volatile__ (
 303       /* fence */
 304       strasm_sync
 305       );
 306   }
 307 }
 308 
 309 template<>
 310 template<typename T>
 311 inline T Atomic::PlatformCmpxchg<1>::operator()(T exchange_value,
 312                                                 T volatile* dest,
 313                                                 T compare_value,
 314                                                 cmpxchg_memory_order order) const {
 315   // Note that cmpxchg guarantees a two-way memory barrier across
 316   // the cmpxchg, so it's really a a 'fence_cmpxchg_fence' if not
 317   // specified otherwise (see atomic.hpp).
 318 
 319   // Using 32 bit internally.
 320   volatile int *dest_base = (volatile int*)((uintptr_t)dest & ~3);
 321 
 322 #ifdef VM_LITTLE_ENDIAN
 323   const unsigned int shift_amount        = ((uintptr_t)dest & 3) * 8;
 324 #else
 325   const unsigned int shift_amount        = ((~(uintptr_t)dest) & 3) * 8;
 326 #endif
 327   const unsigned int masked_compare_val  = ((unsigned int)(unsigned char)compare_value),
 328                      masked_exchange_val = ((unsigned int)(unsigned char)exchange_value),
 329                      xor_value           = (masked_compare_val ^ masked_exchange_val) << shift_amount;
 330 
 331   unsigned int old_value, value32;
 332 
 333   cmpxchg_pre_membar(order);
 334 
 335   __asm__ __volatile__ (
 336     /* simple guard */
 337     "   lbz     %[old_value], 0(%[dest])                  \n"
 338     "   cmpw    %[masked_compare_val], %[old_value]       \n"
 339     "   bne-    2f                                        \n"
 340     /* atomic loop */
 341     "1:                                                   \n"
 342     "   lwarx   %[value32], 0, %[dest_base]               \n"
 343     /* extract byte and compare */
 344     "   srd     %[old_value], %[value32], %[shift_amount] \n"
 345     "   clrldi  %[old_value], %[old_value], 56            \n"
 346     "   cmpw    %[masked_compare_val], %[old_value]       \n"
 347     "   bne-    2f                                        \n"
 348     /* replace byte and try to store */
 349     "   xor     %[value32], %[xor_value], %[value32]      \n"
 350     "   stwcx.  %[value32], 0, %[dest_base]               \n"
 351     "   bne-    1b                                        \n"
 352     /* exit */
 353     "2:                                                   \n"
 354     /* out */
 355     : [old_value]           "=&r"   (old_value),
 356       [value32]             "=&r"   (value32),
 357                             "=m"    (*dest),
 358                             "=m"    (*dest_base)
 359     /* in */
 360     : [dest]                "b"     (dest),
 361       [dest_base]           "b"     (dest_base),
 362       [shift_amount]        "r"     (shift_amount),
 363       [masked_compare_val]  "r"     (masked_compare_val),
 364       [xor_value]           "r"     (xor_value),
 365                             "m"     (*dest),
 366                             "m"     (*dest_base)
 367     /* clobber */
 368     : "cc",
 369       "memory"
 370     );
 371 
 372   cmpxchg_post_membar(order);
 373 
 374   return IntegerTypes::cast<T>((unsigned char)old_value);
 375 }
 376 
 377 template<>
 378 template<typename T>
 379 inline T Atomic::PlatformCmpxchg<4>::operator()(T exchange_value,
 380                                                 T volatile* dest,
 381                                                 T compare_value,
 382                                                 cmpxchg_memory_order order) const {
 383   // Note that cmpxchg guarantees a two-way memory barrier across
 384   // the cmpxchg, so it's really a a 'fence_cmpxchg_fence' if not
 385   // specified otherwise (see atomic.hpp).
 386 
 387   unsigned int old_value;
 388   const uint64_t zero = 0;
 389 
 390   cmpxchg_pre_membar(order);
 391 
 392   __asm__ __volatile__ (
 393     /* simple guard */
 394     "   lwz     %[old_value], 0(%[dest])                \n"
 395     "   cmpw    %[compare_value], %[old_value]          \n"
 396     "   bne-    2f                                      \n"
 397     /* atomic loop */
 398     "1:                                                 \n"
 399     "   lwarx   %[old_value], %[dest], %[zero]          \n"
 400     "   cmpw    %[compare_value], %[old_value]          \n"
 401     "   bne-    2f                                      \n"
 402     "   stwcx.  %[exchange_value], %[dest], %[zero]     \n"
 403     "   bne-    1b                                      \n"
 404     /* exit */
 405     "2:                                                 \n"
 406     /* out */
 407     : [old_value]       "=&r"   (old_value),
 408                         "=m"    (*dest)
 409     /* in */
 410     : [dest]            "b"     (dest),
 411       [zero]            "r"     (zero),
 412       [compare_value]   "r"     (compare_value),
 413       [exchange_value]  "r"     (exchange_value),
 414                         "m"     (*dest)
 415     /* clobber */
 416     : "cc",
 417       "memory"
 418     );
 419 
 420   cmpxchg_post_membar(order);
 421 
 422   return IntegerTypes::cast<T>(old_value);
 423 }
 424 
 425 template<>
 426 template<typename T>
 427 inline T Atomic::PlatformCmpxchg<8>::operator()(T exchange_value,
 428                                                 T volatile* dest,
 429                                                 T compare_value,
 430                                                 cmpxchg_memory_order order) const {
 431   // Note that cmpxchg guarantees a two-way memory barrier across
 432   // the cmpxchg, so it's really a a 'fence_cmpxchg_fence' if not
 433   // specified otherwise (see atomic.hpp).
 434 
 435   long old_value;
 436   const uint64_t zero = 0;
 437 
 438   cmpxchg_pre_membar(order);
 439 
 440   __asm__ __volatile__ (
 441     /* simple guard */
 442     "   ld      %[old_value], 0(%[dest])                \n"
 443     "   cmpd    %[compare_value], %[old_value]          \n"
 444     "   bne-    2f                                      \n"
 445     /* atomic loop */
 446     "1:                                                 \n"
 447     "   ldarx   %[old_value], %[dest], %[zero]          \n"
 448     "   cmpd    %[compare_value], %[old_value]          \n"
 449     "   bne-    2f                                      \n"
 450     "   stdcx.  %[exchange_value], %[dest], %[zero]     \n"
 451     "   bne-    1b                                      \n"
 452     /* exit */
 453     "2:                                                 \n"
 454     /* out */
 455     : [old_value]       "=&r"   (old_value),
 456                         "=m"    (*dest)
 457     /* in */
 458     : [dest]            "b"     (dest),
 459       [zero]            "r"     (zero),
 460       [compare_value]   "r"     (compare_value),
 461       [exchange_value]  "r"     (exchange_value),
 462                         "m"     (*dest)
 463     /* clobber */
 464     : "cc",
 465       "memory"
 466     );
 467 
 468   cmpxchg_post_membar(order);
 469 
 470   return IntegerTypes::cast<T>(old_value);
 471 }
 472 
 473 #undef strasm_sync
 474 #undef strasm_lwsync
 475 #undef strasm_isync
 476 #undef strasm_release
 477 #undef strasm_acquire
 478 #undef strasm_fence
 479 #undef strasm_nobarrier
 480 #undef strasm_nobarrier_clobber_memory
 481 
 482 #endif // OS_CPU_AIX_OJDKPPC_VM_ATOMIC_AIX_PPC_HPP