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_LINUX_PPC_VM_ATOMIC_LINUX_PPC_HPP 27 #define OS_CPU_LINUX_PPC_VM_ATOMIC_LINUX_PPC_HPP 28 29 #ifndef PPC64 30 #error "Atomic currently only implemented 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 // - sync two-way memory barrier, aka fence 55 // - lwsync orders Store|Store, 56 // Load|Store, 57 // Load|Load, 58 // but not Store|Load 59 // - eieio orders memory accesses for device memory (only) 60 // - 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 // - release orders Store|Store, (maps to lwsync) 78 // Load|Store 79 // - acquire orders Load|Store, (maps to lwsync) 80 // Load|Load 81 // - fence orders Store|Store, (maps to 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 template<size_t byte_size> 97 struct Atomic::PlatformAdd 98 : Atomic::AddAndFetch<Atomic::PlatformAdd<byte_size> > 99 { 100 template<typename I, typename D> 101 D add_and_fetch(I add_value, D volatile* dest) const; 102 }; 103 104 template<> 105 template<typename I, typename D> 106 inline D Atomic::PlatformAdd<4>::add_and_fetch(I add_value, D volatile* dest) const { 107 STATIC_ASSERT(4 == sizeof(I)); 108 STATIC_ASSERT(4 == sizeof(D)); 109 110 D result; 111 112 __asm__ __volatile__ ( 113 strasm_lwsync 114 "1: lwarx %0, 0, %2 \n" 115 " add %0, %0, %1 \n" 116 " stwcx. %0, 0, %2 \n" 117 " bne- 1b \n" 118 strasm_isync 119 : /*%0*/"=&r" (result) 120 : /*%1*/"r" (add_value), /*%2*/"r" (dest) 121 : "cc", "memory" ); 122 123 return result; 124 } 125 126 127 template<> 128 template<typename I, typename D> 129 inline D Atomic::PlatformAdd<8>::add_and_fetch(I add_value, D volatile* dest) const { 130 STATIC_ASSERT(8 == sizeof(I)); 131 STATIC_ASSERT(8 == sizeof(D)); 132 133 D result; 134 135 __asm__ __volatile__ ( 136 strasm_lwsync 137 "1: ldarx %0, 0, %2 \n" 138 " add %0, %0, %1 \n" 139 " stdcx. %0, 0, %2 \n" 140 " bne- 1b \n" 141 strasm_isync 142 : /*%0*/"=&r" (result) 143 : /*%1*/"r" (add_value), /*%2*/"r" (dest) 144 : "cc", "memory" ); 145 146 return result; 147 } 148 149 150 inline void Atomic::inc (volatile jint* dest) { 151 152 unsigned int temp; 153 154 __asm__ __volatile__ ( 155 strasm_nobarrier 156 "1: lwarx %0, 0, %2 \n" 157 " addic %0, %0, 1 \n" 158 " stwcx. %0, 0, %2 \n" 159 " bne- 1b \n" 160 strasm_nobarrier 161 : /*%0*/"=&r" (temp), "=m" (*dest) 162 : /*%2*/"r" (dest), "m" (*dest) 163 : "cc" strasm_nobarrier_clobber_memory); 164 165 } 166 167 inline void Atomic::inc_ptr(volatile intptr_t* dest) { 168 169 long temp; 170 171 __asm__ __volatile__ ( 172 strasm_nobarrier 173 "1: ldarx %0, 0, %2 \n" 174 " addic %0, %0, 1 \n" 175 " stdcx. %0, 0, %2 \n" 176 " bne- 1b \n" 177 strasm_nobarrier 178 : /*%0*/"=&r" (temp), "=m" (*dest) 179 : /*%2*/"r" (dest), "m" (*dest) 180 : "cc" strasm_nobarrier_clobber_memory); 181 182 } 183 184 inline void Atomic::inc_ptr(volatile void* dest) { 185 inc_ptr((volatile intptr_t*)dest); 186 } 187 188 189 inline void Atomic::dec (volatile jint* dest) { 190 191 unsigned int temp; 192 193 __asm__ __volatile__ ( 194 strasm_nobarrier 195 "1: lwarx %0, 0, %2 \n" 196 " addic %0, %0, -1 \n" 197 " stwcx. %0, 0, %2 \n" 198 " bne- 1b \n" 199 strasm_nobarrier 200 : /*%0*/"=&r" (temp), "=m" (*dest) 201 : /*%2*/"r" (dest), "m" (*dest) 202 : "cc" strasm_nobarrier_clobber_memory); 203 204 } 205 206 inline void Atomic::dec_ptr(volatile intptr_t* dest) { 207 208 long temp; 209 210 __asm__ __volatile__ ( 211 strasm_nobarrier 212 "1: ldarx %0, 0, %2 \n" 213 " addic %0, %0, -1 \n" 214 " stdcx. %0, 0, %2 \n" 215 " bne- 1b \n" 216 strasm_nobarrier 217 : /*%0*/"=&r" (temp), "=m" (*dest) 218 : /*%2*/"r" (dest), "m" (*dest) 219 : "cc" strasm_nobarrier_clobber_memory); 220 221 } 222 223 inline void Atomic::dec_ptr(volatile void* dest) { 224 dec_ptr((volatile intptr_t*)dest); 225 } 226 227 inline jint Atomic::xchg(jint exchange_value, volatile jint* dest) { 228 229 // Note that xchg_ptr doesn't necessarily do an acquire 230 // (see synchronizer.cpp). 231 232 unsigned int old_value; 233 const uint64_t zero = 0; 234 235 __asm__ __volatile__ ( 236 /* lwsync */ 237 strasm_lwsync 238 /* atomic loop */ 239 "1: \n" 240 " lwarx %[old_value], %[dest], %[zero] \n" 241 " stwcx. %[exchange_value], %[dest], %[zero] \n" 242 " bne- 1b \n" 243 /* isync */ 244 strasm_sync 245 /* exit */ 246 "2: \n" 247 /* out */ 248 : [old_value] "=&r" (old_value), 249 "=m" (*dest) 250 /* in */ 251 : [dest] "b" (dest), 252 [zero] "r" (zero), 253 [exchange_value] "r" (exchange_value), 254 "m" (*dest) 255 /* clobber */ 256 : "cc", 257 "memory" 258 ); 259 260 return (jint) old_value; 261 } 262 263 inline intptr_t Atomic::xchg_ptr(intptr_t exchange_value, volatile intptr_t* dest) { 264 265 // Note that xchg_ptr doesn't necessarily do an acquire 266 // (see synchronizer.cpp). 267 268 long old_value; 269 const uint64_t zero = 0; 270 271 __asm__ __volatile__ ( 272 /* lwsync */ 273 strasm_lwsync 274 /* atomic loop */ 275 "1: \n" 276 " ldarx %[old_value], %[dest], %[zero] \n" 277 " stdcx. %[exchange_value], %[dest], %[zero] \n" 278 " bne- 1b \n" 279 /* isync */ 280 strasm_sync 281 /* exit */ 282 "2: \n" 283 /* out */ 284 : [old_value] "=&r" (old_value), 285 "=m" (*dest) 286 /* in */ 287 : [dest] "b" (dest), 288 [zero] "r" (zero), 289 [exchange_value] "r" (exchange_value), 290 "m" (*dest) 291 /* clobber */ 292 : "cc", 293 "memory" 294 ); 295 296 return (intptr_t) old_value; 297 } 298 299 inline void* Atomic::xchg_ptr(void* exchange_value, volatile void* dest) { 300 return (void*)xchg_ptr((intptr_t)exchange_value, (volatile intptr_t*)dest); 301 } 302 303 inline void cmpxchg_pre_membar(cmpxchg_memory_order order) { 304 if (order != memory_order_relaxed) { 305 __asm__ __volatile__ ( 306 /* fence */ 307 strasm_sync 308 ); 309 } 310 } 311 312 inline void cmpxchg_post_membar(cmpxchg_memory_order order) { 313 if (order != memory_order_relaxed) { 314 __asm__ __volatile__ ( 315 /* fence */ 316 strasm_sync 317 ); 318 } 319 } 320 321 template<> 322 template<typename T> 323 inline T Atomic::PlatformCmpxchg<1>::operator()(T exchange_value, 324 T volatile* dest, 325 T compare_value, 326 cmpxchg_memory_order order) const { 327 STATIC_ASSERT(1 == sizeof(T)); 328 329 // Note that cmpxchg guarantees a two-way memory barrier across 330 // the cmpxchg, so it's really a a 'fence_cmpxchg_fence' if not 331 // specified otherwise (see atomic.hpp). 332 333 // Using 32 bit internally. 334 volatile int *dest_base = (volatile int*)((uintptr_t)dest & ~3); 335 336 #ifdef VM_LITTLE_ENDIAN 337 const unsigned int shift_amount = ((uintptr_t)dest & 3) * 8; 338 #else 339 const unsigned int shift_amount = ((~(uintptr_t)dest) & 3) * 8; 340 #endif 341 const unsigned int masked_compare_val = ((unsigned int)(unsigned char)compare_value), 342 masked_exchange_val = ((unsigned int)(unsigned char)exchange_value), 343 xor_value = (masked_compare_val ^ masked_exchange_val) << shift_amount; 344 345 unsigned int old_value, value32; 346 347 cmpxchg_pre_membar(order); 348 349 __asm__ __volatile__ ( 350 /* simple guard */ 351 " lbz %[old_value], 0(%[dest]) \n" 352 " cmpw %[masked_compare_val], %[old_value] \n" 353 " bne- 2f \n" 354 /* atomic loop */ 355 "1: \n" 356 " lwarx %[value32], 0, %[dest_base] \n" 357 /* extract byte and compare */ 358 " srd %[old_value], %[value32], %[shift_amount] \n" 359 " clrldi %[old_value], %[old_value], 56 \n" 360 " cmpw %[masked_compare_val], %[old_value] \n" 361 " bne- 2f \n" 362 /* replace byte and try to store */ 363 " xor %[value32], %[xor_value], %[value32] \n" 364 " stwcx. %[value32], 0, %[dest_base] \n" 365 " bne- 1b \n" 366 /* exit */ 367 "2: \n" 368 /* out */ 369 : [old_value] "=&r" (old_value), 370 [value32] "=&r" (value32), 371 "=m" (*dest), 372 "=m" (*dest_base) 373 /* in */ 374 : [dest] "b" (dest), 375 [dest_base] "b" (dest_base), 376 [shift_amount] "r" (shift_amount), 377 [masked_compare_val] "r" (masked_compare_val), 378 [xor_value] "r" (xor_value), 379 "m" (*dest), 380 "m" (*dest_base) 381 /* clobber */ 382 : "cc", 383 "memory" 384 ); 385 386 cmpxchg_post_membar(order); 387 388 return PrimitiveConversions::cast<T>((unsigned char)old_value); 389 } 390 391 template<> 392 template<typename T> 393 inline T Atomic::PlatformCmpxchg<4>::operator()(T exchange_value, 394 T volatile* dest, 395 T compare_value, 396 cmpxchg_memory_order order) const { 397 STATIC_ASSERT(4 == sizeof(T)); 398 399 // Note that cmpxchg guarantees a two-way memory barrier across 400 // the cmpxchg, so it's really a a 'fence_cmpxchg_fence' if not 401 // specified otherwise (see atomic.hpp). 402 403 T old_value; 404 const uint64_t zero = 0; 405 406 cmpxchg_pre_membar(order); 407 408 __asm__ __volatile__ ( 409 /* simple guard */ 410 " lwz %[old_value], 0(%[dest]) \n" 411 " cmpw %[compare_value], %[old_value] \n" 412 " bne- 2f \n" 413 /* atomic loop */ 414 "1: \n" 415 " lwarx %[old_value], %[dest], %[zero] \n" 416 " cmpw %[compare_value], %[old_value] \n" 417 " bne- 2f \n" 418 " stwcx. %[exchange_value], %[dest], %[zero] \n" 419 " bne- 1b \n" 420 /* exit */ 421 "2: \n" 422 /* out */ 423 : [old_value] "=&r" (old_value), 424 "=m" (*dest) 425 /* in */ 426 : [dest] "b" (dest), 427 [zero] "r" (zero), 428 [compare_value] "r" (compare_value), 429 [exchange_value] "r" (exchange_value), 430 "m" (*dest) 431 /* clobber */ 432 : "cc", 433 "memory" 434 ); 435 436 cmpxchg_post_membar(order); 437 438 return old_value; 439 } 440 441 template<> 442 template<typename T> 443 inline T Atomic::PlatformCmpxchg<8>::operator()(T exchange_value, 444 T volatile* dest, 445 T compare_value, 446 cmpxchg_memory_order order) const { 447 STATIC_ASSERT(8 == sizeof(T)); 448 449 // Note that cmpxchg guarantees a two-way memory barrier across 450 // the cmpxchg, so it's really a a 'fence_cmpxchg_fence' if not 451 // specified otherwise (see atomic.hpp). 452 453 T old_value; 454 const uint64_t zero = 0; 455 456 cmpxchg_pre_membar(order); 457 458 __asm__ __volatile__ ( 459 /* simple guard */ 460 " ld %[old_value], 0(%[dest]) \n" 461 " cmpd %[compare_value], %[old_value] \n" 462 " bne- 2f \n" 463 /* atomic loop */ 464 "1: \n" 465 " ldarx %[old_value], %[dest], %[zero] \n" 466 " cmpd %[compare_value], %[old_value] \n" 467 " bne- 2f \n" 468 " stdcx. %[exchange_value], %[dest], %[zero] \n" 469 " bne- 1b \n" 470 /* exit */ 471 "2: \n" 472 /* out */ 473 : [old_value] "=&r" (old_value), 474 "=m" (*dest) 475 /* in */ 476 : [dest] "b" (dest), 477 [zero] "r" (zero), 478 [compare_value] "r" (compare_value), 479 [exchange_value] "r" (exchange_value), 480 "m" (*dest) 481 /* clobber */ 482 : "cc", 483 "memory" 484 ); 485 486 cmpxchg_post_membar(order); 487 488 return old_value; 489 } 490 491 #undef strasm_sync 492 #undef strasm_lwsync 493 #undef strasm_isync 494 #undef strasm_release 495 #undef strasm_acquire 496 #undef strasm_fence 497 #undef strasm_nobarrier 498 #undef strasm_nobarrier_clobber_memory 499 500 #endif // OS_CPU_LINUX_PPC_VM_ATOMIC_LINUX_PPC_HPP