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 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 #define VM_HAS_SPECIALIZED_CMPXCHG_BYTE 310 inline jbyte Atomic::cmpxchg(jbyte exchange_value, volatile jbyte* dest, jbyte compare_value, cmpxchg_memory_order order) { 311 312 // Note that cmpxchg guarantees a two-way memory barrier across 313 // the cmpxchg, so it's really a a 'fence_cmpxchg_fence' if not 314 // specified otherwise (see atomic.hpp). 315 316 // Using 32 bit internally. 317 volatile int *dest_base = (volatile int*)((uintptr_t)dest & ~3); 318 319 #ifdef VM_LITTLE_ENDIAN 320 const unsigned int shift_amount = ((uintptr_t)dest & 3) * 8; 321 #else 322 const unsigned int shift_amount = ((~(uintptr_t)dest) & 3) * 8; 323 #endif 324 const unsigned int masked_compare_val = ((unsigned int)(unsigned char)compare_value), 325 masked_exchange_val = ((unsigned int)(unsigned char)exchange_value), 326 xor_value = (masked_compare_val ^ masked_exchange_val) << shift_amount; 327 328 unsigned int old_value, value32; 329 330 cmpxchg_pre_membar(order); 331 332 __asm__ __volatile__ ( 333 /* simple guard */ 334 " lbz %[old_value], 0(%[dest]) \n" 335 " cmpw %[masked_compare_val], %[old_value] \n" 336 " bne- 2f \n" 337 /* atomic loop */ 338 "1: \n" 339 " lwarx %[value32], 0, %[dest_base] \n" 340 /* extract byte and compare */ 341 " srd %[old_value], %[value32], %[shift_amount] \n" 342 " clrldi %[old_value], %[old_value], 56 \n" 343 " cmpw %[masked_compare_val], %[old_value] \n" 344 " bne- 2f \n" 345 /* replace byte and try to store */ 346 " xor %[value32], %[xor_value], %[value32] \n" 347 " stwcx. %[value32], 0, %[dest_base] \n" 348 " bne- 1b \n" 349 /* exit */ 350 "2: \n" 351 /* out */ 352 : [old_value] "=&r" (old_value), 353 [value32] "=&r" (value32), 354 "=m" (*dest), 355 "=m" (*dest_base) 356 /* in */ 357 : [dest] "b" (dest), 358 [dest_base] "b" (dest_base), 359 [shift_amount] "r" (shift_amount), 360 [masked_compare_val] "r" (masked_compare_val), 361 [xor_value] "r" (xor_value), 362 "m" (*dest), 363 "m" (*dest_base) 364 /* clobber */ 365 : "cc", 366 "memory" 367 ); 368 369 cmpxchg_post_membar(order); 370 371 return (jbyte)(unsigned char)old_value; 372 } 373 374 inline jint Atomic::cmpxchg(jint exchange_value, volatile jint* dest, jint compare_value, cmpxchg_memory_order order) { 375 376 // Note that cmpxchg guarantees a two-way memory barrier across 377 // the cmpxchg, so it's really a a 'fence_cmpxchg_fence' if not 378 // specified otherwise (see atomic.hpp). 379 380 unsigned int old_value; 381 const uint64_t zero = 0; 382 383 cmpxchg_pre_membar(order); 384 385 __asm__ __volatile__ ( 386 /* simple guard */ 387 " lwz %[old_value], 0(%[dest]) \n" 388 " cmpw %[compare_value], %[old_value] \n" 389 " bne- 2f \n" 390 /* atomic loop */ 391 "1: \n" 392 " lwarx %[old_value], %[dest], %[zero] \n" 393 " cmpw %[compare_value], %[old_value] \n" 394 " bne- 2f \n" 395 " stwcx. %[exchange_value], %[dest], %[zero] \n" 396 " bne- 1b \n" 397 /* exit */ 398 "2: \n" 399 /* out */ 400 : [old_value] "=&r" (old_value), 401 "=m" (*dest) 402 /* in */ 403 : [dest] "b" (dest), 404 [zero] "r" (zero), 405 [compare_value] "r" (compare_value), 406 [exchange_value] "r" (exchange_value), 407 "m" (*dest) 408 /* clobber */ 409 : "cc", 410 "memory" 411 ); 412 413 cmpxchg_post_membar(order); 414 415 return (jint) old_value; 416 } 417 418 inline jlong Atomic::cmpxchg(jlong exchange_value, volatile jlong* dest, jlong compare_value, cmpxchg_memory_order order) { 419 420 // Note that cmpxchg guarantees a two-way memory barrier across 421 // the cmpxchg, so it's really a a 'fence_cmpxchg_fence' if not 422 // specified otherwise (see atomic.hpp). 423 424 long old_value; 425 const uint64_t zero = 0; 426 427 cmpxchg_pre_membar(order); 428 429 __asm__ __volatile__ ( 430 /* simple guard */ 431 " ld %[old_value], 0(%[dest]) \n" 432 " cmpd %[compare_value], %[old_value] \n" 433 " bne- 2f \n" 434 /* atomic loop */ 435 "1: \n" 436 " ldarx %[old_value], %[dest], %[zero] \n" 437 " cmpd %[compare_value], %[old_value] \n" 438 " bne- 2f \n" 439 " stdcx. %[exchange_value], %[dest], %[zero] \n" 440 " bne- 1b \n" 441 /* exit */ 442 "2: \n" 443 /* out */ 444 : [old_value] "=&r" (old_value), 445 "=m" (*dest) 446 /* in */ 447 : [dest] "b" (dest), 448 [zero] "r" (zero), 449 [compare_value] "r" (compare_value), 450 [exchange_value] "r" (exchange_value), 451 "m" (*dest) 452 /* clobber */ 453 : "cc", 454 "memory" 455 ); 456 457 cmpxchg_post_membar(order); 458 459 return (jlong) old_value; 460 } 461 462 inline intptr_t Atomic::cmpxchg_ptr(intptr_t exchange_value, volatile intptr_t* dest, intptr_t compare_value, cmpxchg_memory_order order) { 463 return (intptr_t)cmpxchg((jlong)exchange_value, (volatile jlong*)dest, (jlong)compare_value, order); 464 } 465 466 inline void* Atomic::cmpxchg_ptr(void* exchange_value, volatile void* dest, void* compare_value, cmpxchg_memory_order order) { 467 return (void*)cmpxchg((jlong)exchange_value, (volatile jlong*)dest, (jlong)compare_value, order); 468 } 469 470 #undef strasm_sync 471 #undef strasm_lwsync 472 #undef strasm_isync 473 #undef strasm_release 474 #undef strasm_acquire 475 #undef strasm_fence 476 #undef strasm_nobarrier 477 #undef strasm_nobarrier_clobber_memory 478 479 #endif // OS_CPU_LINUX_PPC_VM_ATOMIC_LINUX_PPC_HPP