1 /* 2 * Copyright (c) 1997, 2016, 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 "asm/macroAssembler.hpp" 27 #include "asm/macroAssembler.inline.hpp" 28 #include "logging/log.hpp" 29 #include "memory/resourceArea.hpp" 30 #include "runtime/java.hpp" 31 #include "runtime/os.hpp" 32 #include "runtime/stubCodeGenerator.hpp" 33 #include "vm_version_x86.hpp" 34 35 36 int VM_Version::_cpu; 37 int VM_Version::_model; 38 int VM_Version::_stepping; 39 VM_Version::CpuidInfo VM_Version::_cpuid_info = { 0, }; 40 41 // Address of instruction which causes SEGV 42 address VM_Version::_cpuinfo_segv_addr = 0; 43 // Address of instruction after the one which causes SEGV 44 address VM_Version::_cpuinfo_cont_addr = 0; 45 46 static BufferBlob* stub_blob; 47 static const int stub_size = 1000; 48 49 extern "C" { 50 typedef void (*get_cpu_info_stub_t)(void*); 51 } 52 static get_cpu_info_stub_t get_cpu_info_stub = NULL; 53 54 55 class VM_Version_StubGenerator: public StubCodeGenerator { 56 public: 57 58 VM_Version_StubGenerator(CodeBuffer *c) : StubCodeGenerator(c) {} 59 60 address generate_get_cpu_info() { 61 // Flags to test CPU type. 62 const uint32_t HS_EFL_AC = 0x40000; 63 const uint32_t HS_EFL_ID = 0x200000; 64 // Values for when we don't have a CPUID instruction. 65 const int CPU_FAMILY_SHIFT = 8; 66 const uint32_t CPU_FAMILY_386 = (3 << CPU_FAMILY_SHIFT); 67 const uint32_t CPU_FAMILY_486 = (4 << CPU_FAMILY_SHIFT); 68 bool use_evex = FLAG_IS_DEFAULT(UseAVX) || (UseAVX > 2); 69 70 Label detect_486, cpu486, detect_586, std_cpuid1, std_cpuid4; 71 Label sef_cpuid, ext_cpuid, ext_cpuid1, ext_cpuid5, ext_cpuid7, done, wrapup; 72 Label legacy_setup, save_restore_except, legacy_save_restore, start_simd_check; 73 74 StubCodeMark mark(this, "VM_Version", "get_cpu_info_stub"); 75 # define __ _masm-> 76 77 address start = __ pc(); 78 79 // 80 // void get_cpu_info(VM_Version::CpuidInfo* cpuid_info); 81 // 82 // LP64: rcx and rdx are first and second argument registers on windows 83 84 __ push(rbp); 85 #ifdef _LP64 86 __ mov(rbp, c_rarg0); // cpuid_info address 87 #else 88 __ movptr(rbp, Address(rsp, 8)); // cpuid_info address 89 #endif 90 __ push(rbx); 91 __ push(rsi); 92 __ pushf(); // preserve rbx, and flags 93 __ pop(rax); 94 __ push(rax); 95 __ mov(rcx, rax); 96 // 97 // if we are unable to change the AC flag, we have a 386 98 // 99 __ xorl(rax, HS_EFL_AC); 100 __ push(rax); 101 __ popf(); 102 __ pushf(); 103 __ pop(rax); 104 __ cmpptr(rax, rcx); 105 __ jccb(Assembler::notEqual, detect_486); 106 107 __ movl(rax, CPU_FAMILY_386); 108 __ movl(Address(rbp, in_bytes(VM_Version::std_cpuid1_offset())), rax); 109 __ jmp(done); 110 111 // 112 // If we are unable to change the ID flag, we have a 486 which does 113 // not support the "cpuid" instruction. 114 // 115 __ bind(detect_486); 116 __ mov(rax, rcx); 117 __ xorl(rax, HS_EFL_ID); 118 __ push(rax); 119 __ popf(); 120 __ pushf(); 121 __ pop(rax); 122 __ cmpptr(rcx, rax); 123 __ jccb(Assembler::notEqual, detect_586); 124 125 __ bind(cpu486); 126 __ movl(rax, CPU_FAMILY_486); 127 __ movl(Address(rbp, in_bytes(VM_Version::std_cpuid1_offset())), rax); 128 __ jmp(done); 129 130 // 131 // At this point, we have a chip which supports the "cpuid" instruction 132 // 133 __ bind(detect_586); 134 __ xorl(rax, rax); 135 __ cpuid(); 136 __ orl(rax, rax); 137 __ jcc(Assembler::equal, cpu486); // if cpuid doesn't support an input 138 // value of at least 1, we give up and 139 // assume a 486 140 __ lea(rsi, Address(rbp, in_bytes(VM_Version::std_cpuid0_offset()))); 141 __ movl(Address(rsi, 0), rax); 142 __ movl(Address(rsi, 4), rbx); 143 __ movl(Address(rsi, 8), rcx); 144 __ movl(Address(rsi,12), rdx); 145 146 __ cmpl(rax, 0xa); // Is cpuid(0xB) supported? 147 __ jccb(Assembler::belowEqual, std_cpuid4); 148 149 // 150 // cpuid(0xB) Processor Topology 151 // 152 __ movl(rax, 0xb); 153 __ xorl(rcx, rcx); // Threads level 154 __ cpuid(); 155 156 __ lea(rsi, Address(rbp, in_bytes(VM_Version::tpl_cpuidB0_offset()))); 157 __ movl(Address(rsi, 0), rax); 158 __ movl(Address(rsi, 4), rbx); 159 __ movl(Address(rsi, 8), rcx); 160 __ movl(Address(rsi,12), rdx); 161 162 __ movl(rax, 0xb); 163 __ movl(rcx, 1); // Cores level 164 __ cpuid(); 165 __ push(rax); 166 __ andl(rax, 0x1f); // Determine if valid topology level 167 __ orl(rax, rbx); // eax[4:0] | ebx[0:15] == 0 indicates invalid level 168 __ andl(rax, 0xffff); 169 __ pop(rax); 170 __ jccb(Assembler::equal, std_cpuid4); 171 172 __ lea(rsi, Address(rbp, in_bytes(VM_Version::tpl_cpuidB1_offset()))); 173 __ movl(Address(rsi, 0), rax); 174 __ movl(Address(rsi, 4), rbx); 175 __ movl(Address(rsi, 8), rcx); 176 __ movl(Address(rsi,12), rdx); 177 178 __ movl(rax, 0xb); 179 __ movl(rcx, 2); // Packages level 180 __ cpuid(); 181 __ push(rax); 182 __ andl(rax, 0x1f); // Determine if valid topology level 183 __ orl(rax, rbx); // eax[4:0] | ebx[0:15] == 0 indicates invalid level 184 __ andl(rax, 0xffff); 185 __ pop(rax); 186 __ jccb(Assembler::equal, std_cpuid4); 187 188 __ lea(rsi, Address(rbp, in_bytes(VM_Version::tpl_cpuidB2_offset()))); 189 __ movl(Address(rsi, 0), rax); 190 __ movl(Address(rsi, 4), rbx); 191 __ movl(Address(rsi, 8), rcx); 192 __ movl(Address(rsi,12), rdx); 193 194 // 195 // cpuid(0x4) Deterministic cache params 196 // 197 __ bind(std_cpuid4); 198 __ movl(rax, 4); 199 __ cmpl(rax, Address(rbp, in_bytes(VM_Version::std_cpuid0_offset()))); // Is cpuid(0x4) supported? 200 __ jccb(Assembler::greater, std_cpuid1); 201 202 __ xorl(rcx, rcx); // L1 cache 203 __ cpuid(); 204 __ push(rax); 205 __ andl(rax, 0x1f); // Determine if valid cache parameters used 206 __ orl(rax, rax); // eax[4:0] == 0 indicates invalid cache 207 __ pop(rax); 208 __ jccb(Assembler::equal, std_cpuid1); 209 210 __ lea(rsi, Address(rbp, in_bytes(VM_Version::dcp_cpuid4_offset()))); 211 __ movl(Address(rsi, 0), rax); 212 __ movl(Address(rsi, 4), rbx); 213 __ movl(Address(rsi, 8), rcx); 214 __ movl(Address(rsi,12), rdx); 215 216 // 217 // Standard cpuid(0x1) 218 // 219 __ bind(std_cpuid1); 220 __ movl(rax, 1); 221 __ cpuid(); 222 __ lea(rsi, Address(rbp, in_bytes(VM_Version::std_cpuid1_offset()))); 223 __ movl(Address(rsi, 0), rax); 224 __ movl(Address(rsi, 4), rbx); 225 __ movl(Address(rsi, 8), rcx); 226 __ movl(Address(rsi,12), rdx); 227 228 // 229 // Check if OS has enabled XGETBV instruction to access XCR0 230 // (OSXSAVE feature flag) and CPU supports AVX 231 // 232 __ andl(rcx, 0x18000000); // cpuid1 bits osxsave | avx 233 __ cmpl(rcx, 0x18000000); 234 __ jccb(Assembler::notEqual, sef_cpuid); // jump if AVX is not supported 235 236 // 237 // XCR0, XFEATURE_ENABLED_MASK register 238 // 239 __ xorl(rcx, rcx); // zero for XCR0 register 240 __ xgetbv(); 241 __ lea(rsi, Address(rbp, in_bytes(VM_Version::xem_xcr0_offset()))); 242 __ movl(Address(rsi, 0), rax); 243 __ movl(Address(rsi, 4), rdx); 244 245 // 246 // cpuid(0x7) Structured Extended Features 247 // 248 __ bind(sef_cpuid); 249 __ movl(rax, 7); 250 __ cmpl(rax, Address(rbp, in_bytes(VM_Version::std_cpuid0_offset()))); // Is cpuid(0x7) supported? 251 __ jccb(Assembler::greater, ext_cpuid); 252 253 __ xorl(rcx, rcx); 254 __ cpuid(); 255 __ lea(rsi, Address(rbp, in_bytes(VM_Version::sef_cpuid7_offset()))); 256 __ movl(Address(rsi, 0), rax); 257 __ movl(Address(rsi, 4), rbx); 258 259 // 260 // Extended cpuid(0x80000000) 261 // 262 __ bind(ext_cpuid); 263 __ movl(rax, 0x80000000); 264 __ cpuid(); 265 __ cmpl(rax, 0x80000000); // Is cpuid(0x80000001) supported? 266 __ jcc(Assembler::belowEqual, done); 267 __ cmpl(rax, 0x80000004); // Is cpuid(0x80000005) supported? 268 __ jccb(Assembler::belowEqual, ext_cpuid1); 269 __ cmpl(rax, 0x80000006); // Is cpuid(0x80000007) supported? 270 __ jccb(Assembler::belowEqual, ext_cpuid5); 271 __ cmpl(rax, 0x80000007); // Is cpuid(0x80000008) supported? 272 __ jccb(Assembler::belowEqual, ext_cpuid7); 273 // 274 // Extended cpuid(0x80000008) 275 // 276 __ movl(rax, 0x80000008); 277 __ cpuid(); 278 __ lea(rsi, Address(rbp, in_bytes(VM_Version::ext_cpuid8_offset()))); 279 __ movl(Address(rsi, 0), rax); 280 __ movl(Address(rsi, 4), rbx); 281 __ movl(Address(rsi, 8), rcx); 282 __ movl(Address(rsi,12), rdx); 283 284 // 285 // Extended cpuid(0x80000007) 286 // 287 __ bind(ext_cpuid7); 288 __ movl(rax, 0x80000007); 289 __ cpuid(); 290 __ lea(rsi, Address(rbp, in_bytes(VM_Version::ext_cpuid7_offset()))); 291 __ movl(Address(rsi, 0), rax); 292 __ movl(Address(rsi, 4), rbx); 293 __ movl(Address(rsi, 8), rcx); 294 __ movl(Address(rsi,12), rdx); 295 296 // 297 // Extended cpuid(0x80000005) 298 // 299 __ bind(ext_cpuid5); 300 __ movl(rax, 0x80000005); 301 __ cpuid(); 302 __ lea(rsi, Address(rbp, in_bytes(VM_Version::ext_cpuid5_offset()))); 303 __ movl(Address(rsi, 0), rax); 304 __ movl(Address(rsi, 4), rbx); 305 __ movl(Address(rsi, 8), rcx); 306 __ movl(Address(rsi,12), rdx); 307 308 // 309 // Extended cpuid(0x80000001) 310 // 311 __ bind(ext_cpuid1); 312 __ movl(rax, 0x80000001); 313 __ cpuid(); 314 __ lea(rsi, Address(rbp, in_bytes(VM_Version::ext_cpuid1_offset()))); 315 __ movl(Address(rsi, 0), rax); 316 __ movl(Address(rsi, 4), rbx); 317 __ movl(Address(rsi, 8), rcx); 318 __ movl(Address(rsi,12), rdx); 319 320 // 321 // Check if OS has enabled XGETBV instruction to access XCR0 322 // (OSXSAVE feature flag) and CPU supports AVX 323 // 324 __ lea(rsi, Address(rbp, in_bytes(VM_Version::std_cpuid1_offset()))); 325 __ movl(rcx, 0x18000000); // cpuid1 bits osxsave | avx 326 __ andl(rcx, Address(rsi, 8)); // cpuid1 bits osxsave | avx 327 __ cmpl(rcx, 0x18000000); 328 __ jccb(Assembler::notEqual, done); // jump if AVX is not supported 329 330 __ movl(rax, 0x6); 331 __ andl(rax, Address(rbp, in_bytes(VM_Version::xem_xcr0_offset()))); // xcr0 bits sse | ymm 332 __ cmpl(rax, 0x6); 333 __ jccb(Assembler::equal, start_simd_check); // return if AVX is not supported 334 335 // we need to bridge farther than imm8, so we use this island as a thunk 336 __ bind(done); 337 __ jmp(wrapup); 338 339 __ bind(start_simd_check); 340 // 341 // Some OSs have a bug when upper 128/256bits of YMM/ZMM 342 // registers are not restored after a signal processing. 343 // Generate SEGV here (reference through NULL) 344 // and check upper YMM/ZMM bits after it. 345 // 346 intx saved_useavx = UseAVX; 347 intx saved_usesse = UseSSE; 348 // check _cpuid_info.sef_cpuid7_ebx.bits.avx512f 349 __ lea(rsi, Address(rbp, in_bytes(VM_Version::sef_cpuid7_offset()))); 350 __ movl(rax, 0x10000); 351 __ andl(rax, Address(rsi, 4)); // xcr0 bits sse | ymm 352 __ cmpl(rax, 0x10000); 353 __ jccb(Assembler::notEqual, legacy_setup); // jump if EVEX is not supported 354 // check _cpuid_info.xem_xcr0_eax.bits.opmask 355 // check _cpuid_info.xem_xcr0_eax.bits.zmm512 356 // check _cpuid_info.xem_xcr0_eax.bits.zmm32 357 __ movl(rax, 0xE0); 358 __ andl(rax, Address(rbp, in_bytes(VM_Version::xem_xcr0_offset()))); // xcr0 bits sse | ymm 359 __ cmpl(rax, 0xE0); 360 __ jccb(Assembler::notEqual, legacy_setup); // jump if EVEX is not supported 361 362 // If UseAVX is unitialized or is set by the user to include EVEX 363 if (use_evex) { 364 // EVEX setup: run in lowest evex mode 365 VM_Version::set_evex_cpuFeatures(); // Enable temporary to pass asserts 366 UseAVX = 3; 367 UseSSE = 2; 368 #ifdef _WINDOWS 369 // xmm5-xmm15 are not preserved by caller on windows 370 // https://msdn.microsoft.com/en-us/library/9z1stfyw.aspx 371 __ subptr(rsp, 64); 372 __ evmovdqul(Address(rsp, 0), xmm7, Assembler::AVX_512bit); 373 #ifdef _LP64 374 __ subptr(rsp, 64); 375 __ evmovdqul(Address(rsp, 0), xmm8, Assembler::AVX_512bit); 376 __ subptr(rsp, 64); 377 __ evmovdqul(Address(rsp, 0), xmm31, Assembler::AVX_512bit); 378 #endif // _LP64 379 #endif // _WINDOWS 380 381 // load value into all 64 bytes of zmm7 register 382 __ movl(rcx, VM_Version::ymm_test_value()); 383 __ movdl(xmm0, rcx); 384 __ movl(rcx, 0xffff); 385 __ kmovwl(k1, rcx); 386 __ evpbroadcastd(xmm0, xmm0, Assembler::AVX_512bit); 387 __ evmovdqul(xmm7, xmm0, Assembler::AVX_512bit); 388 #ifdef _LP64 389 __ evmovdqul(xmm8, xmm0, Assembler::AVX_512bit); 390 __ evmovdqul(xmm31, xmm0, Assembler::AVX_512bit); 391 #endif 392 VM_Version::clean_cpuFeatures(); 393 __ jmp(save_restore_except); 394 } 395 396 __ bind(legacy_setup); 397 // AVX setup 398 VM_Version::set_avx_cpuFeatures(); // Enable temporary to pass asserts 399 UseAVX = 1; 400 UseSSE = 2; 401 #ifdef _WINDOWS 402 __ subptr(rsp, 32); 403 __ vmovdqu(Address(rsp, 0), xmm7); 404 #ifdef _LP64 405 __ subptr(rsp, 32); 406 __ vmovdqu(Address(rsp, 0), xmm8); 407 __ subptr(rsp, 32); 408 __ vmovdqu(Address(rsp, 0), xmm15); 409 #endif // _LP64 410 #endif // _WINDOWS 411 412 // load value into all 32 bytes of ymm7 register 413 __ movl(rcx, VM_Version::ymm_test_value()); 414 415 __ movdl(xmm0, rcx); 416 __ pshufd(xmm0, xmm0, 0x00); 417 __ vinsertf128_high(xmm0, xmm0); 418 __ vmovdqu(xmm7, xmm0); 419 #ifdef _LP64 420 __ vmovdqu(xmm8, xmm0); 421 __ vmovdqu(xmm15, xmm0); 422 #endif 423 VM_Version::clean_cpuFeatures(); 424 425 __ bind(save_restore_except); 426 __ xorl(rsi, rsi); 427 VM_Version::set_cpuinfo_segv_addr(__ pc()); 428 // Generate SEGV 429 __ movl(rax, Address(rsi, 0)); 430 431 VM_Version::set_cpuinfo_cont_addr(__ pc()); 432 // Returns here after signal. Save xmm0 to check it later. 433 434 // check _cpuid_info.sef_cpuid7_ebx.bits.avx512f 435 __ lea(rsi, Address(rbp, in_bytes(VM_Version::sef_cpuid7_offset()))); 436 __ movl(rax, 0x10000); 437 __ andl(rax, Address(rsi, 4)); 438 __ cmpl(rax, 0x10000); 439 __ jccb(Assembler::notEqual, legacy_save_restore); 440 // check _cpuid_info.xem_xcr0_eax.bits.opmask 441 // check _cpuid_info.xem_xcr0_eax.bits.zmm512 442 // check _cpuid_info.xem_xcr0_eax.bits.zmm32 443 __ movl(rax, 0xE0); 444 __ andl(rax, Address(rbp, in_bytes(VM_Version::xem_xcr0_offset()))); // xcr0 bits sse | ymm 445 __ cmpl(rax, 0xE0); 446 __ jccb(Assembler::notEqual, legacy_save_restore); 447 448 // If UseAVX is unitialized or is set by the user to include EVEX 449 if (use_evex) { 450 // EVEX check: run in lowest evex mode 451 VM_Version::set_evex_cpuFeatures(); // Enable temporary to pass asserts 452 UseAVX = 3; 453 UseSSE = 2; 454 __ lea(rsi, Address(rbp, in_bytes(VM_Version::zmm_save_offset()))); 455 __ evmovdqul(Address(rsi, 0), xmm0, Assembler::AVX_512bit); 456 __ evmovdqul(Address(rsi, 64), xmm7, Assembler::AVX_512bit); 457 #ifdef _LP64 458 __ evmovdqul(Address(rsi, 128), xmm8, Assembler::AVX_512bit); 459 __ evmovdqul(Address(rsi, 192), xmm31, Assembler::AVX_512bit); 460 #endif 461 462 #ifdef _WINDOWS 463 #ifdef _LP64 464 __ evmovdqul(xmm31, Address(rsp, 0), Assembler::AVX_512bit); 465 __ addptr(rsp, 64); 466 __ evmovdqul(xmm8, Address(rsp, 0), Assembler::AVX_512bit); 467 __ addptr(rsp, 64); 468 #endif // _LP64 469 __ evmovdqul(xmm7, Address(rsp, 0), Assembler::AVX_512bit); 470 __ addptr(rsp, 64); 471 #endif // _WINDOWS 472 VM_Version::clean_cpuFeatures(); 473 UseAVX = saved_useavx; 474 UseSSE = saved_usesse; 475 __ jmp(wrapup); 476 } 477 478 __ bind(legacy_save_restore); 479 // AVX check 480 VM_Version::set_avx_cpuFeatures(); // Enable temporary to pass asserts 481 UseAVX = 1; 482 UseSSE = 2; 483 __ lea(rsi, Address(rbp, in_bytes(VM_Version::ymm_save_offset()))); 484 __ vmovdqu(Address(rsi, 0), xmm0); 485 __ vmovdqu(Address(rsi, 32), xmm7); 486 #ifdef _LP64 487 __ vmovdqu(Address(rsi, 64), xmm8); 488 __ vmovdqu(Address(rsi, 96), xmm15); 489 #endif 490 491 #ifdef _WINDOWS 492 #ifdef _LP64 493 __ vmovdqu(xmm15, Address(rsp, 0)); 494 __ addptr(rsp, 32); 495 __ vmovdqu(xmm8, Address(rsp, 0)); 496 __ addptr(rsp, 32); 497 #endif // _LP64 498 __ vmovdqu(xmm7, Address(rsp, 0)); 499 __ addptr(rsp, 32); 500 #endif // _WINDOWS 501 VM_Version::clean_cpuFeatures(); 502 UseAVX = saved_useavx; 503 UseSSE = saved_usesse; 504 505 __ bind(wrapup); 506 __ popf(); 507 __ pop(rsi); 508 __ pop(rbx); 509 __ pop(rbp); 510 __ ret(0); 511 512 # undef __ 513 514 return start; 515 }; 516 }; 517 518 void VM_Version::get_processor_features() { 519 520 _cpu = 4; // 486 by default 521 _model = 0; 522 _stepping = 0; 523 _features = 0; 524 _logical_processors_per_package = 1; 525 // i486 internal cache is both I&D and has a 16-byte line size 526 _L1_data_cache_line_size = 16; 527 528 // Get raw processor info 529 530 get_cpu_info_stub(&_cpuid_info); 531 532 assert_is_initialized(); 533 _cpu = extended_cpu_family(); 534 _model = extended_cpu_model(); 535 _stepping = cpu_stepping(); 536 537 if (cpu_family() > 4) { // it supports CPUID 538 _features = feature_flags(); 539 // Logical processors are only available on P4s and above, 540 // and only if hyperthreading is available. 541 _logical_processors_per_package = logical_processor_count(); 542 _L1_data_cache_line_size = L1_line_size(); 543 } 544 545 _supports_cx8 = supports_cmpxchg8(); 546 // xchg and xadd instructions 547 _supports_atomic_getset4 = true; 548 _supports_atomic_getadd4 = true; 549 LP64_ONLY(_supports_atomic_getset8 = true); 550 LP64_ONLY(_supports_atomic_getadd8 = true); 551 552 #ifdef _LP64 553 // OS should support SSE for x64 and hardware should support at least SSE2. 554 if (!VM_Version::supports_sse2()) { 555 vm_exit_during_initialization("Unknown x64 processor: SSE2 not supported"); 556 } 557 // in 64 bit the use of SSE2 is the minimum 558 if (UseSSE < 2) UseSSE = 2; 559 #endif 560 561 #ifdef AMD64 562 // flush_icache_stub have to be generated first. 563 // That is why Icache line size is hard coded in ICache class, 564 // see icache_x86.hpp. It is also the reason why we can't use 565 // clflush instruction in 32-bit VM since it could be running 566 // on CPU which does not support it. 567 // 568 // The only thing we can do is to verify that flushed 569 // ICache::line_size has correct value. 570 guarantee(_cpuid_info.std_cpuid1_edx.bits.clflush != 0, "clflush is not supported"); 571 // clflush_size is size in quadwords (8 bytes). 572 guarantee(_cpuid_info.std_cpuid1_ebx.bits.clflush_size == 8, "such clflush size is not supported"); 573 #endif 574 575 // If the OS doesn't support SSE, we can't use this feature even if the HW does 576 if (!os::supports_sse()) 577 _features &= ~(CPU_SSE|CPU_SSE2|CPU_SSE3|CPU_SSSE3|CPU_SSE4A|CPU_SSE4_1|CPU_SSE4_2); 578 579 if (UseSSE < 4) { 580 _features &= ~CPU_SSE4_1; 581 _features &= ~CPU_SSE4_2; 582 } 583 584 if (UseSSE < 3) { 585 _features &= ~CPU_SSE3; 586 _features &= ~CPU_SSSE3; 587 _features &= ~CPU_SSE4A; 588 } 589 590 if (UseSSE < 2) 591 _features &= ~CPU_SSE2; 592 593 if (UseSSE < 1) 594 _features &= ~CPU_SSE; 595 596 // first try initial setting and detect what we can support 597 if (UseAVX > 0) { 598 if (UseAVX > 2 && supports_evex()) { 599 UseAVX = 3; 600 } else if (UseAVX > 1 && supports_avx2()) { 601 UseAVX = 2; 602 } else if (UseAVX > 0 && supports_avx()) { 603 UseAVX = 1; 604 } else { 605 UseAVX = 0; 606 } 607 } else if (UseAVX < 0) { 608 UseAVX = 0; 609 } 610 611 if (UseAVX < 3) { 612 _features &= ~CPU_AVX512F; 613 _features &= ~CPU_AVX512DQ; 614 _features &= ~CPU_AVX512CD; 615 _features &= ~CPU_AVX512BW; 616 _features &= ~CPU_AVX512VL; 617 } 618 619 if (UseAVX < 2) 620 _features &= ~CPU_AVX2; 621 622 if (UseAVX < 1) 623 _features &= ~CPU_AVX; 624 625 if (!UseAES && !FLAG_IS_DEFAULT(UseAES)) 626 _features &= ~CPU_AES; 627 628 if (logical_processors_per_package() == 1) { 629 // HT processor could be installed on a system which doesn't support HT. 630 _features &= ~CPU_HT; 631 } 632 633 char buf[256]; 634 jio_snprintf(buf, sizeof(buf), "(%u cores per cpu, %u threads per core) family %d model %d stepping %d%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s", 635 cores_per_cpu(), threads_per_core(), 636 cpu_family(), _model, _stepping, 637 (supports_cmov() ? ", cmov" : ""), 638 (supports_cmpxchg8() ? ", cx8" : ""), 639 (supports_fxsr() ? ", fxsr" : ""), 640 (supports_mmx() ? ", mmx" : ""), 641 (supports_sse() ? ", sse" : ""), 642 (supports_sse2() ? ", sse2" : ""), 643 (supports_sse3() ? ", sse3" : ""), 644 (supports_ssse3()? ", ssse3": ""), 645 (supports_sse4_1() ? ", sse4.1" : ""), 646 (supports_sse4_2() ? ", sse4.2" : ""), 647 (supports_popcnt() ? ", popcnt" : ""), 648 (supports_avx() ? ", avx" : ""), 649 (supports_avx2() ? ", avx2" : ""), 650 (supports_aes() ? ", aes" : ""), 651 (supports_clmul() ? ", clmul" : ""), 652 (supports_erms() ? ", erms" : ""), 653 (supports_rtm() ? ", rtm" : ""), 654 (supports_mmx_ext() ? ", mmxext" : ""), 655 (supports_3dnow_prefetch() ? ", 3dnowpref" : ""), 656 (supports_lzcnt() ? ", lzcnt": ""), 657 (supports_sse4a() ? ", sse4a": ""), 658 (supports_ht() ? ", ht": ""), 659 (supports_tsc() ? ", tsc": ""), 660 (supports_tscinv_bit() ? ", tscinvbit": ""), 661 (supports_tscinv() ? ", tscinv": ""), 662 (supports_bmi1() ? ", bmi1" : ""), 663 (supports_bmi2() ? ", bmi2" : ""), 664 (supports_adx() ? ", adx" : ""), 665 (supports_evex() ? ", evex" : ""), 666 (supports_sha() ? ", sha" : ""), 667 (supports_fma() ? ", fma" : "")); 668 _features_string = os::strdup(buf); 669 670 // UseSSE is set to the smaller of what hardware supports and what 671 // the command line requires. I.e., you cannot set UseSSE to 2 on 672 // older Pentiums which do not support it. 673 if (UseSSE > 4) UseSSE=4; 674 if (UseSSE < 0) UseSSE=0; 675 if (!supports_sse4_1()) // Drop to 3 if no SSE4 support 676 UseSSE = MIN2((intx)3,UseSSE); 677 if (!supports_sse3()) // Drop to 2 if no SSE3 support 678 UseSSE = MIN2((intx)2,UseSSE); 679 if (!supports_sse2()) // Drop to 1 if no SSE2 support 680 UseSSE = MIN2((intx)1,UseSSE); 681 if (!supports_sse ()) // Drop to 0 if no SSE support 682 UseSSE = 0; 683 684 // Use AES instructions if available. 685 if (supports_aes()) { 686 if (FLAG_IS_DEFAULT(UseAES)) { 687 FLAG_SET_DEFAULT(UseAES, true); 688 } 689 if (!UseAES) { 690 if (UseAESIntrinsics && !FLAG_IS_DEFAULT(UseAESIntrinsics)) { 691 warning("AES intrinsics require UseAES flag to be enabled. Intrinsics will be disabled."); 692 } 693 FLAG_SET_DEFAULT(UseAESIntrinsics, false); 694 } else { 695 if (UseSSE > 2) { 696 if (FLAG_IS_DEFAULT(UseAESIntrinsics)) { 697 FLAG_SET_DEFAULT(UseAESIntrinsics, true); 698 } 699 } else { 700 // The AES intrinsic stubs require AES instruction support (of course) 701 // but also require sse3 mode or higher for instructions it use. 702 if (UseAESIntrinsics && !FLAG_IS_DEFAULT(UseAESIntrinsics)) { 703 warning("X86 AES intrinsics require SSE3 instructions or higher. Intrinsics will be disabled."); 704 } 705 FLAG_SET_DEFAULT(UseAESIntrinsics, false); 706 } 707 708 // --AES-CTR begins-- 709 if (!UseAESIntrinsics) { 710 if (UseAESCTRIntrinsics && !FLAG_IS_DEFAULT(UseAESCTRIntrinsics)) { 711 warning("AES-CTR intrinsics require UseAESIntrinsics flag to be enabled. Intrinsics will be disabled."); 712 FLAG_SET_DEFAULT(UseAESCTRIntrinsics, false); 713 } 714 } else { 715 if(supports_sse4_1()) { 716 if (FLAG_IS_DEFAULT(UseAESCTRIntrinsics)) { 717 FLAG_SET_DEFAULT(UseAESCTRIntrinsics, true); 718 } 719 } else { 720 // The AES-CTR intrinsic stubs require AES instruction support (of course) 721 // but also require sse4.1 mode or higher for instructions it use. 722 if (UseAESCTRIntrinsics && !FLAG_IS_DEFAULT(UseAESCTRIntrinsics)) { 723 warning("X86 AES-CTR intrinsics require SSE4.1 instructions or higher. Intrinsics will be disabled."); 724 } 725 FLAG_SET_DEFAULT(UseAESCTRIntrinsics, false); 726 } 727 } 728 // --AES-CTR ends-- 729 } 730 } else if (UseAES || UseAESIntrinsics || UseAESCTRIntrinsics) { 731 if (UseAES && !FLAG_IS_DEFAULT(UseAES)) { 732 warning("AES instructions are not available on this CPU"); 733 FLAG_SET_DEFAULT(UseAES, false); 734 } 735 if (UseAESIntrinsics && !FLAG_IS_DEFAULT(UseAESIntrinsics)) { 736 warning("AES intrinsics are not available on this CPU"); 737 FLAG_SET_DEFAULT(UseAESIntrinsics, false); 738 } 739 if (UseAESCTRIntrinsics && !FLAG_IS_DEFAULT(UseAESCTRIntrinsics)) { 740 warning("AES-CTR intrinsics are not available on this CPU"); 741 FLAG_SET_DEFAULT(UseAESCTRIntrinsics, false); 742 } 743 } 744 745 // Use CLMUL instructions if available. 746 if (supports_clmul()) { 747 if (FLAG_IS_DEFAULT(UseCLMUL)) { 748 UseCLMUL = true; 749 } 750 } else if (UseCLMUL) { 751 if (!FLAG_IS_DEFAULT(UseCLMUL)) 752 warning("CLMUL instructions not available on this CPU (AVX may also be required)"); 753 FLAG_SET_DEFAULT(UseCLMUL, false); 754 } 755 756 if (UseCLMUL && (UseSSE > 2)) { 757 if (FLAG_IS_DEFAULT(UseCRC32Intrinsics)) { 758 UseCRC32Intrinsics = true; 759 } 760 } else if (UseCRC32Intrinsics) { 761 if (!FLAG_IS_DEFAULT(UseCRC32Intrinsics)) 762 warning("CRC32 Intrinsics requires CLMUL instructions (not available on this CPU)"); 763 FLAG_SET_DEFAULT(UseCRC32Intrinsics, false); 764 } 765 766 if (supports_sse4_2()) { 767 if (FLAG_IS_DEFAULT(UseCRC32CIntrinsics)) { 768 UseCRC32CIntrinsics = true; 769 } 770 } 771 else if (UseCRC32CIntrinsics) { 772 if (!FLAG_IS_DEFAULT(UseCRC32CIntrinsics)) { 773 warning("CRC32C intrinsics are not available on this CPU"); 774 } 775 FLAG_SET_DEFAULT(UseCRC32CIntrinsics, false); 776 } 777 778 // GHASH/GCM intrinsics 779 if (UseCLMUL && (UseSSE > 2)) { 780 if (FLAG_IS_DEFAULT(UseGHASHIntrinsics)) { 781 UseGHASHIntrinsics = true; 782 } 783 } else if (UseGHASHIntrinsics) { 784 if (!FLAG_IS_DEFAULT(UseGHASHIntrinsics)) 785 warning("GHASH intrinsic requires CLMUL and SSE2 instructions on this CPU"); 786 FLAG_SET_DEFAULT(UseGHASHIntrinsics, false); 787 } 788 789 if (supports_fma() && UseSSE >= 2) { 790 if (FLAG_IS_DEFAULT(UseFMA)) { 791 UseFMA = true; 792 } 793 } else if (UseFMA) { 794 warning("FMA instructions are not available on this CPU"); 795 FLAG_SET_DEFAULT(UseFMA, false); 796 } 797 798 if (supports_sha() LP64_ONLY(|| supports_avx2() && supports_bmi2())) { 799 if (FLAG_IS_DEFAULT(UseSHA)) { 800 UseSHA = true; 801 } 802 } else if (UseSHA) { 803 warning("SHA instructions are not available on this CPU"); 804 FLAG_SET_DEFAULT(UseSHA, false); 805 } 806 807 if (supports_sha() && UseSHA) { 808 if (FLAG_IS_DEFAULT(UseSHA1Intrinsics)) { 809 FLAG_SET_DEFAULT(UseSHA1Intrinsics, true); 810 } 811 } else if (UseSHA1Intrinsics) { 812 warning("Intrinsics for SHA-1 crypto hash functions not available on this CPU."); 813 FLAG_SET_DEFAULT(UseSHA1Intrinsics, false); 814 } 815 816 if (UseSHA) { 817 if (FLAG_IS_DEFAULT(UseSHA256Intrinsics)) { 818 FLAG_SET_DEFAULT(UseSHA256Intrinsics, true); 819 } 820 } else if (UseSHA256Intrinsics) { 821 warning("Intrinsics for SHA-224 and SHA-256 crypto hash functions not available on this CPU."); 822 FLAG_SET_DEFAULT(UseSHA256Intrinsics, false); 823 } 824 825 if (UseSHA) { 826 if (FLAG_IS_DEFAULT(UseSHA512Intrinsics)) { 827 FLAG_SET_DEFAULT(UseSHA512Intrinsics, true); 828 } 829 } else if (UseSHA512Intrinsics) { 830 warning("Intrinsics for SHA-384 and SHA-512 crypto hash functions not available on this CPU."); 831 FLAG_SET_DEFAULT(UseSHA512Intrinsics, false); 832 } 833 834 if (!(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics)) { 835 FLAG_SET_DEFAULT(UseSHA, false); 836 } 837 838 if (UseAdler32Intrinsics) { 839 warning("Adler32Intrinsics not available on this CPU."); 840 FLAG_SET_DEFAULT(UseAdler32Intrinsics, false); 841 } 842 843 if (!supports_rtm() && UseRTMLocking) { 844 // Can't continue because UseRTMLocking affects UseBiasedLocking flag 845 // setting during arguments processing. See use_biased_locking(). 846 // VM_Version_init() is executed after UseBiasedLocking is used 847 // in Thread::allocate(). 848 vm_exit_during_initialization("RTM instructions are not available on this CPU"); 849 } 850 851 #if INCLUDE_RTM_OPT 852 if (UseRTMLocking) { 853 if (is_client_compilation_mode_vm()) { 854 // Only C2 does RTM locking optimization. 855 // Can't continue because UseRTMLocking affects UseBiasedLocking flag 856 // setting during arguments processing. See use_biased_locking(). 857 vm_exit_during_initialization("RTM locking optimization is not supported in emulated client VM"); 858 } 859 if (is_intel_family_core()) { 860 if ((_model == CPU_MODEL_HASWELL_E3) || 861 (_model == CPU_MODEL_HASWELL_E7 && _stepping < 3) || 862 (_model == CPU_MODEL_BROADWELL && _stepping < 4)) { 863 // currently a collision between SKL and HSW_E3 864 if (!UnlockExperimentalVMOptions && UseAVX < 3) { 865 vm_exit_during_initialization("UseRTMLocking is only available as experimental option on this platform. It must be enabled via -XX:+UnlockExperimentalVMOptions flag."); 866 } else { 867 warning("UseRTMLocking is only available as experimental option on this platform."); 868 } 869 } 870 } 871 if (!FLAG_IS_CMDLINE(UseRTMLocking)) { 872 // RTM locking should be used only for applications with 873 // high lock contention. For now we do not use it by default. 874 vm_exit_during_initialization("UseRTMLocking flag should be only set on command line"); 875 } 876 if (!is_power_of_2(RTMTotalCountIncrRate)) { 877 warning("RTMTotalCountIncrRate must be a power of 2, resetting it to 64"); 878 FLAG_SET_DEFAULT(RTMTotalCountIncrRate, 64); 879 } 880 if (RTMAbortRatio < 0 || RTMAbortRatio > 100) { 881 warning("RTMAbortRatio must be in the range 0 to 100, resetting it to 50"); 882 FLAG_SET_DEFAULT(RTMAbortRatio, 50); 883 } 884 } else { // !UseRTMLocking 885 if (UseRTMForStackLocks) { 886 if (!FLAG_IS_DEFAULT(UseRTMForStackLocks)) { 887 warning("UseRTMForStackLocks flag should be off when UseRTMLocking flag is off"); 888 } 889 FLAG_SET_DEFAULT(UseRTMForStackLocks, false); 890 } 891 if (UseRTMDeopt) { 892 FLAG_SET_DEFAULT(UseRTMDeopt, false); 893 } 894 if (PrintPreciseRTMLockingStatistics) { 895 FLAG_SET_DEFAULT(PrintPreciseRTMLockingStatistics, false); 896 } 897 } 898 #else 899 if (UseRTMLocking) { 900 // Only C2 does RTM locking optimization. 901 // Can't continue because UseRTMLocking affects UseBiasedLocking flag 902 // setting during arguments processing. See use_biased_locking(). 903 vm_exit_during_initialization("RTM locking optimization is not supported in this VM"); 904 } 905 #endif 906 907 #ifdef COMPILER2 908 if (UseFPUForSpilling) { 909 if (UseSSE < 2) { 910 // Only supported with SSE2+ 911 FLAG_SET_DEFAULT(UseFPUForSpilling, false); 912 } 913 } 914 #endif 915 #if defined(COMPILER2) || INCLUDE_JVMCI 916 if (MaxVectorSize > 0) { 917 if (!is_power_of_2(MaxVectorSize)) { 918 warning("MaxVectorSize must be a power of 2"); 919 FLAG_SET_DEFAULT(MaxVectorSize, 64); 920 } 921 if (MaxVectorSize > 64) { 922 FLAG_SET_DEFAULT(MaxVectorSize, 64); 923 } 924 if (MaxVectorSize > 16 && (UseAVX == 0 || !os_supports_avx_vectors())) { 925 // 32 bytes vectors (in YMM) are only supported with AVX+ 926 FLAG_SET_DEFAULT(MaxVectorSize, 16); 927 } 928 if (UseSSE < 2) { 929 // Vectors (in XMM) are only supported with SSE2+ 930 FLAG_SET_DEFAULT(MaxVectorSize, 0); 931 } 932 #if defined(COMPILER2) && defined(ASSERT) 933 if (supports_avx() && PrintMiscellaneous && Verbose && TraceNewVectors) { 934 tty->print_cr("State of YMM registers after signal handle:"); 935 int nreg = 2 LP64_ONLY(+2); 936 const char* ymm_name[4] = {"0", "7", "8", "15"}; 937 for (int i = 0; i < nreg; i++) { 938 tty->print("YMM%s:", ymm_name[i]); 939 for (int j = 7; j >=0; j--) { 940 tty->print(" %x", _cpuid_info.ymm_save[i*8 + j]); 941 } 942 tty->cr(); 943 } 944 } 945 #endif // COMPILER2 && ASSERT 946 } 947 #endif // COMPILER2 || INCLUDE_JVMCI 948 949 #ifdef COMPILER2 950 #ifdef _LP64 951 if (FLAG_IS_DEFAULT(UseMultiplyToLenIntrinsic)) { 952 UseMultiplyToLenIntrinsic = true; 953 } 954 if (FLAG_IS_DEFAULT(UseSquareToLenIntrinsic)) { 955 UseSquareToLenIntrinsic = true; 956 } 957 if (FLAG_IS_DEFAULT(UseMulAddIntrinsic)) { 958 UseMulAddIntrinsic = true; 959 } 960 if (FLAG_IS_DEFAULT(UseMontgomeryMultiplyIntrinsic)) { 961 UseMontgomeryMultiplyIntrinsic = true; 962 } 963 if (FLAG_IS_DEFAULT(UseMontgomerySquareIntrinsic)) { 964 UseMontgomerySquareIntrinsic = true; 965 } 966 #else 967 if (UseMultiplyToLenIntrinsic) { 968 if (!FLAG_IS_DEFAULT(UseMultiplyToLenIntrinsic)) { 969 warning("multiplyToLen intrinsic is not available in 32-bit VM"); 970 } 971 FLAG_SET_DEFAULT(UseMultiplyToLenIntrinsic, false); 972 } 973 if (UseMontgomeryMultiplyIntrinsic) { 974 if (!FLAG_IS_DEFAULT(UseMontgomeryMultiplyIntrinsic)) { 975 warning("montgomeryMultiply intrinsic is not available in 32-bit VM"); 976 } 977 FLAG_SET_DEFAULT(UseMontgomeryMultiplyIntrinsic, false); 978 } 979 if (UseMontgomerySquareIntrinsic) { 980 if (!FLAG_IS_DEFAULT(UseMontgomerySquareIntrinsic)) { 981 warning("montgomerySquare intrinsic is not available in 32-bit VM"); 982 } 983 FLAG_SET_DEFAULT(UseMontgomerySquareIntrinsic, false); 984 } 985 if (UseSquareToLenIntrinsic) { 986 if (!FLAG_IS_DEFAULT(UseSquareToLenIntrinsic)) { 987 warning("squareToLen intrinsic is not available in 32-bit VM"); 988 } 989 FLAG_SET_DEFAULT(UseSquareToLenIntrinsic, false); 990 } 991 if (UseMulAddIntrinsic) { 992 if (!FLAG_IS_DEFAULT(UseMulAddIntrinsic)) { 993 warning("mulAdd intrinsic is not available in 32-bit VM"); 994 } 995 FLAG_SET_DEFAULT(UseMulAddIntrinsic, false); 996 } 997 #endif 998 #endif // COMPILER2 999 1000 // On new cpus instructions which update whole XMM register should be used 1001 // to prevent partial register stall due to dependencies on high half. 1002 // 1003 // UseXmmLoadAndClearUpper == true --> movsd(xmm, mem) 1004 // UseXmmLoadAndClearUpper == false --> movlpd(xmm, mem) 1005 // UseXmmRegToRegMoveAll == true --> movaps(xmm, xmm), movapd(xmm, xmm). 1006 // UseXmmRegToRegMoveAll == false --> movss(xmm, xmm), movsd(xmm, xmm). 1007 1008 if( is_amd() ) { // AMD cpus specific settings 1009 if( supports_sse2() && FLAG_IS_DEFAULT(UseAddressNop) ) { 1010 // Use it on new AMD cpus starting from Opteron. 1011 UseAddressNop = true; 1012 } 1013 if( supports_sse2() && FLAG_IS_DEFAULT(UseNewLongLShift) ) { 1014 // Use it on new AMD cpus starting from Opteron. 1015 UseNewLongLShift = true; 1016 } 1017 if( FLAG_IS_DEFAULT(UseXmmLoadAndClearUpper) ) { 1018 if (supports_sse4a()) { 1019 UseXmmLoadAndClearUpper = true; // use movsd only on '10h' Opteron 1020 } else { 1021 UseXmmLoadAndClearUpper = false; 1022 } 1023 } 1024 if( FLAG_IS_DEFAULT(UseXmmRegToRegMoveAll) ) { 1025 if( supports_sse4a() ) { 1026 UseXmmRegToRegMoveAll = true; // use movaps, movapd only on '10h' 1027 } else { 1028 UseXmmRegToRegMoveAll = false; 1029 } 1030 } 1031 if( FLAG_IS_DEFAULT(UseXmmI2F) ) { 1032 if( supports_sse4a() ) { 1033 UseXmmI2F = true; 1034 } else { 1035 UseXmmI2F = false; 1036 } 1037 } 1038 if( FLAG_IS_DEFAULT(UseXmmI2D) ) { 1039 if( supports_sse4a() ) { 1040 UseXmmI2D = true; 1041 } else { 1042 UseXmmI2D = false; 1043 } 1044 } 1045 if (supports_sse4_2()) { 1046 if (FLAG_IS_DEFAULT(UseSSE42Intrinsics)) { 1047 FLAG_SET_DEFAULT(UseSSE42Intrinsics, true); 1048 } 1049 } else { 1050 if (UseSSE42Intrinsics && !FLAG_IS_DEFAULT(UseAESIntrinsics)) { 1051 warning("SSE4.2 intrinsics require SSE4.2 instructions or higher. Intrinsics will be disabled."); 1052 } 1053 FLAG_SET_DEFAULT(UseSSE42Intrinsics, false); 1054 } 1055 1056 // some defaults for AMD family 15h 1057 if ( cpu_family() == 0x15 ) { 1058 // On family 15h processors default is no sw prefetch 1059 if (FLAG_IS_DEFAULT(AllocatePrefetchStyle)) { 1060 AllocatePrefetchStyle = 0; 1061 } 1062 // Also, if some other prefetch style is specified, default instruction type is PREFETCHW 1063 if (FLAG_IS_DEFAULT(AllocatePrefetchInstr)) { 1064 AllocatePrefetchInstr = 3; 1065 } 1066 // On family 15h processors use XMM and UnalignedLoadStores for Array Copy 1067 if (supports_sse2() && FLAG_IS_DEFAULT(UseXMMForArrayCopy)) { 1068 UseXMMForArrayCopy = true; 1069 } 1070 if (supports_sse2() && FLAG_IS_DEFAULT(UseUnalignedLoadStores)) { 1071 UseUnalignedLoadStores = true; 1072 } 1073 } 1074 1075 #ifdef COMPILER2 1076 if (MaxVectorSize > 16) { 1077 // Limit vectors size to 16 bytes on current AMD cpus. 1078 FLAG_SET_DEFAULT(MaxVectorSize, 16); 1079 } 1080 #endif // COMPILER2 1081 } 1082 1083 if( is_intel() ) { // Intel cpus specific settings 1084 if( FLAG_IS_DEFAULT(UseStoreImmI16) ) { 1085 UseStoreImmI16 = false; // don't use it on Intel cpus 1086 } 1087 if( cpu_family() == 6 || cpu_family() == 15 ) { 1088 if( FLAG_IS_DEFAULT(UseAddressNop) ) { 1089 // Use it on all Intel cpus starting from PentiumPro 1090 UseAddressNop = true; 1091 } 1092 } 1093 if( FLAG_IS_DEFAULT(UseXmmLoadAndClearUpper) ) { 1094 UseXmmLoadAndClearUpper = true; // use movsd on all Intel cpus 1095 } 1096 if( FLAG_IS_DEFAULT(UseXmmRegToRegMoveAll) ) { 1097 if( supports_sse3() ) { 1098 UseXmmRegToRegMoveAll = true; // use movaps, movapd on new Intel cpus 1099 } else { 1100 UseXmmRegToRegMoveAll = false; 1101 } 1102 } 1103 if( cpu_family() == 6 && supports_sse3() ) { // New Intel cpus 1104 #ifdef COMPILER2 1105 if( FLAG_IS_DEFAULT(MaxLoopPad) ) { 1106 // For new Intel cpus do the next optimization: 1107 // don't align the beginning of a loop if there are enough instructions 1108 // left (NumberOfLoopInstrToAlign defined in c2_globals.hpp) 1109 // in current fetch line (OptoLoopAlignment) or the padding 1110 // is big (> MaxLoopPad). 1111 // Set MaxLoopPad to 11 for new Intel cpus to reduce number of 1112 // generated NOP instructions. 11 is the largest size of one 1113 // address NOP instruction '0F 1F' (see Assembler::nop(i)). 1114 MaxLoopPad = 11; 1115 } 1116 #endif // COMPILER2 1117 if (FLAG_IS_DEFAULT(UseXMMForArrayCopy)) { 1118 UseXMMForArrayCopy = true; // use SSE2 movq on new Intel cpus 1119 } 1120 if (supports_sse4_2() && supports_ht()) { // Newest Intel cpus 1121 if (FLAG_IS_DEFAULT(UseUnalignedLoadStores)) { 1122 UseUnalignedLoadStores = true; // use movdqu on newest Intel cpus 1123 } 1124 } 1125 if (supports_sse4_2()) { 1126 if (FLAG_IS_DEFAULT(UseSSE42Intrinsics)) { 1127 FLAG_SET_DEFAULT(UseSSE42Intrinsics, true); 1128 } 1129 } else { 1130 if (UseSSE42Intrinsics && !FLAG_IS_DEFAULT(UseAESIntrinsics)) { 1131 warning("SSE4.2 intrinsics require SSE4.2 instructions or higher. Intrinsics will be disabled."); 1132 } 1133 FLAG_SET_DEFAULT(UseSSE42Intrinsics, false); 1134 } 1135 } 1136 if ((cpu_family() == 0x06) && 1137 ((extended_cpu_model() == 0x36) || // Centerton 1138 (extended_cpu_model() == 0x37) || // Silvermont 1139 (extended_cpu_model() == 0x4D))) { 1140 #ifdef COMPILER2 1141 if (FLAG_IS_DEFAULT(OptoScheduling)) { 1142 OptoScheduling = true; 1143 } 1144 #endif 1145 if (supports_sse4_2()) { // Silvermont 1146 if (FLAG_IS_DEFAULT(UseUnalignedLoadStores)) { 1147 UseUnalignedLoadStores = true; // use movdqu on newest Intel cpus 1148 } 1149 } 1150 } 1151 if(FLAG_IS_DEFAULT(AllocatePrefetchInstr) && supports_3dnow_prefetch()) { 1152 AllocatePrefetchInstr = 3; 1153 } 1154 } 1155 1156 #ifdef _LP64 1157 if (UseSSE42Intrinsics) { 1158 if (FLAG_IS_DEFAULT(UseVectorizedMismatchIntrinsic)) { 1159 UseVectorizedMismatchIntrinsic = true; 1160 } 1161 } else if (UseVectorizedMismatchIntrinsic) { 1162 if (!FLAG_IS_DEFAULT(UseVectorizedMismatchIntrinsic)) 1163 warning("vectorizedMismatch intrinsics are not available on this CPU"); 1164 FLAG_SET_DEFAULT(UseVectorizedMismatchIntrinsic, false); 1165 } 1166 #else 1167 if (UseVectorizedMismatchIntrinsic) { 1168 if (!FLAG_IS_DEFAULT(UseVectorizedMismatchIntrinsic)) { 1169 warning("vectorizedMismatch intrinsic is not available in 32-bit VM"); 1170 } 1171 FLAG_SET_DEFAULT(UseVectorizedMismatchIntrinsic, false); 1172 } 1173 #endif // _LP64 1174 1175 // Use count leading zeros count instruction if available. 1176 if (supports_lzcnt()) { 1177 if (FLAG_IS_DEFAULT(UseCountLeadingZerosInstruction)) { 1178 UseCountLeadingZerosInstruction = true; 1179 } 1180 } else if (UseCountLeadingZerosInstruction) { 1181 warning("lzcnt instruction is not available on this CPU"); 1182 FLAG_SET_DEFAULT(UseCountLeadingZerosInstruction, false); 1183 } 1184 1185 // Use count trailing zeros instruction if available 1186 if (supports_bmi1()) { 1187 // tzcnt does not require VEX prefix 1188 if (FLAG_IS_DEFAULT(UseCountTrailingZerosInstruction)) { 1189 if (!UseBMI1Instructions && !FLAG_IS_DEFAULT(UseBMI1Instructions)) { 1190 // Don't use tzcnt if BMI1 is switched off on command line. 1191 UseCountTrailingZerosInstruction = false; 1192 } else { 1193 UseCountTrailingZerosInstruction = true; 1194 } 1195 } 1196 } else if (UseCountTrailingZerosInstruction) { 1197 warning("tzcnt instruction is not available on this CPU"); 1198 FLAG_SET_DEFAULT(UseCountTrailingZerosInstruction, false); 1199 } 1200 1201 // BMI instructions (except tzcnt) use an encoding with VEX prefix. 1202 // VEX prefix is generated only when AVX > 0. 1203 if (supports_bmi1() && supports_avx()) { 1204 if (FLAG_IS_DEFAULT(UseBMI1Instructions)) { 1205 UseBMI1Instructions = true; 1206 } 1207 } else if (UseBMI1Instructions) { 1208 warning("BMI1 instructions are not available on this CPU (AVX is also required)"); 1209 FLAG_SET_DEFAULT(UseBMI1Instructions, false); 1210 } 1211 1212 if (supports_bmi2() && supports_avx()) { 1213 if (FLAG_IS_DEFAULT(UseBMI2Instructions)) { 1214 UseBMI2Instructions = true; 1215 } 1216 } else if (UseBMI2Instructions) { 1217 warning("BMI2 instructions are not available on this CPU (AVX is also required)"); 1218 FLAG_SET_DEFAULT(UseBMI2Instructions, false); 1219 } 1220 1221 // Use population count instruction if available. 1222 if (supports_popcnt()) { 1223 if (FLAG_IS_DEFAULT(UsePopCountInstruction)) { 1224 UsePopCountInstruction = true; 1225 } 1226 } else if (UsePopCountInstruction) { 1227 warning("POPCNT instruction is not available on this CPU"); 1228 FLAG_SET_DEFAULT(UsePopCountInstruction, false); 1229 } 1230 1231 // Use fast-string operations if available. 1232 if (supports_erms()) { 1233 if (FLAG_IS_DEFAULT(UseFastStosb)) { 1234 UseFastStosb = true; 1235 } 1236 } else if (UseFastStosb) { 1237 warning("fast-string operations are not available on this CPU"); 1238 FLAG_SET_DEFAULT(UseFastStosb, false); 1239 } 1240 1241 #ifdef COMPILER2 1242 if (FLAG_IS_DEFAULT(AlignVector)) { 1243 // Modern processors allow misaligned memory operations for vectors. 1244 AlignVector = !UseUnalignedLoadStores; 1245 } 1246 #endif // COMPILER2 1247 1248 if( AllocatePrefetchInstr == 3 && !supports_3dnow_prefetch() ) AllocatePrefetchInstr=0; 1249 if( !supports_sse() && supports_3dnow_prefetch() ) AllocatePrefetchInstr = 3; 1250 1251 // Allocation prefetch settings 1252 intx cache_line_size = prefetch_data_size(); 1253 if( cache_line_size > AllocatePrefetchStepSize ) 1254 AllocatePrefetchStepSize = cache_line_size; 1255 1256 AllocatePrefetchDistance = allocate_prefetch_distance(); 1257 AllocatePrefetchStyle = allocate_prefetch_style(); 1258 1259 if (is_intel() && cpu_family() == 6 && supports_sse3()) { 1260 if (AllocatePrefetchStyle == 2) { // watermark prefetching on Core 1261 #ifdef _LP64 1262 AllocatePrefetchDistance = 384; 1263 #else 1264 AllocatePrefetchDistance = 320; 1265 #endif 1266 } 1267 if (supports_sse4_2() && supports_ht()) { // Nehalem based cpus 1268 AllocatePrefetchDistance = 192; 1269 if (FLAG_IS_DEFAULT(AllocatePrefetchLines)) { 1270 FLAG_SET_DEFAULT(AllocatePrefetchLines, 4); 1271 } 1272 } 1273 #ifdef COMPILER2 1274 if (supports_sse4_2()) { 1275 if (FLAG_IS_DEFAULT(UseFPUForSpilling)) { 1276 FLAG_SET_DEFAULT(UseFPUForSpilling, true); 1277 } 1278 } 1279 #endif 1280 } 1281 1282 #ifdef _LP64 1283 // Prefetch settings 1284 PrefetchCopyIntervalInBytes = prefetch_copy_interval_in_bytes(); 1285 PrefetchScanIntervalInBytes = prefetch_scan_interval_in_bytes(); 1286 PrefetchFieldsAhead = prefetch_fields_ahead(); 1287 #endif 1288 1289 if (FLAG_IS_DEFAULT(ContendedPaddingWidth) && 1290 (cache_line_size > ContendedPaddingWidth)) 1291 ContendedPaddingWidth = cache_line_size; 1292 1293 // This machine allows unaligned memory accesses 1294 if (FLAG_IS_DEFAULT(UseUnalignedAccesses)) { 1295 FLAG_SET_DEFAULT(UseUnalignedAccesses, true); 1296 } 1297 1298 #ifndef PRODUCT 1299 if (log_is_enabled(Info, os, cpu)) { 1300 outputStream* log = Log(os, cpu)::info_stream(); 1301 log->print_cr("Logical CPUs per core: %u", 1302 logical_processors_per_package()); 1303 log->print_cr("L1 data cache line size: %u", L1_data_cache_line_size()); 1304 log->print("UseSSE=%d", (int) UseSSE); 1305 if (UseAVX > 0) { 1306 log->print(" UseAVX=%d", (int) UseAVX); 1307 } 1308 if (UseAES) { 1309 log->print(" UseAES=1"); 1310 } 1311 #ifdef COMPILER2 1312 if (MaxVectorSize > 0) { 1313 log->print(" MaxVectorSize=%d", (int) MaxVectorSize); 1314 } 1315 #endif 1316 log->cr(); 1317 log->print("Allocation"); 1318 if (AllocatePrefetchStyle <= 0 || UseSSE == 0 && !supports_3dnow_prefetch()) { 1319 log->print_cr(": no prefetching"); 1320 } else { 1321 log->print(" prefetching: "); 1322 if (UseSSE == 0 && supports_3dnow_prefetch()) { 1323 log->print("PREFETCHW"); 1324 } else if (UseSSE >= 1) { 1325 if (AllocatePrefetchInstr == 0) { 1326 log->print("PREFETCHNTA"); 1327 } else if (AllocatePrefetchInstr == 1) { 1328 log->print("PREFETCHT0"); 1329 } else if (AllocatePrefetchInstr == 2) { 1330 log->print("PREFETCHT2"); 1331 } else if (AllocatePrefetchInstr == 3) { 1332 log->print("PREFETCHW"); 1333 } 1334 } 1335 if (AllocatePrefetchLines > 1) { 1336 log->print_cr(" at distance %d, %d lines of %d bytes", (int) AllocatePrefetchDistance, (int) AllocatePrefetchLines, (int) AllocatePrefetchStepSize); 1337 } else { 1338 log->print_cr(" at distance %d, one line of %d bytes", (int) AllocatePrefetchDistance, (int) AllocatePrefetchStepSize); 1339 } 1340 } 1341 1342 if (PrefetchCopyIntervalInBytes > 0) { 1343 log->print_cr("PrefetchCopyIntervalInBytes %d", (int) PrefetchCopyIntervalInBytes); 1344 } 1345 if (PrefetchScanIntervalInBytes > 0) { 1346 log->print_cr("PrefetchScanIntervalInBytes %d", (int) PrefetchScanIntervalInBytes); 1347 } 1348 if (PrefetchFieldsAhead > 0) { 1349 log->print_cr("PrefetchFieldsAhead %d", (int) PrefetchFieldsAhead); 1350 } 1351 if (ContendedPaddingWidth > 0) { 1352 log->print_cr("ContendedPaddingWidth %d", (int) ContendedPaddingWidth); 1353 } 1354 } 1355 #endif // !PRODUCT 1356 } 1357 1358 bool VM_Version::use_biased_locking() { 1359 #if INCLUDE_RTM_OPT 1360 // RTM locking is most useful when there is high lock contention and 1361 // low data contention. With high lock contention the lock is usually 1362 // inflated and biased locking is not suitable for that case. 1363 // RTM locking code requires that biased locking is off. 1364 // Note: we can't switch off UseBiasedLocking in get_processor_features() 1365 // because it is used by Thread::allocate() which is called before 1366 // VM_Version::initialize(). 1367 if (UseRTMLocking && UseBiasedLocking) { 1368 if (FLAG_IS_DEFAULT(UseBiasedLocking)) { 1369 FLAG_SET_DEFAULT(UseBiasedLocking, false); 1370 } else { 1371 warning("Biased locking is not supported with RTM locking; ignoring UseBiasedLocking flag." ); 1372 UseBiasedLocking = false; 1373 } 1374 } 1375 #endif 1376 return UseBiasedLocking; 1377 } 1378 1379 void VM_Version::initialize() { 1380 ResourceMark rm; 1381 // Making this stub must be FIRST use of assembler 1382 1383 stub_blob = BufferBlob::create("get_cpu_info_stub", stub_size); 1384 if (stub_blob == NULL) { 1385 vm_exit_during_initialization("Unable to allocate get_cpu_info_stub"); 1386 } 1387 CodeBuffer c(stub_blob); 1388 VM_Version_StubGenerator g(&c); 1389 get_cpu_info_stub = CAST_TO_FN_PTR(get_cpu_info_stub_t, 1390 g.generate_get_cpu_info()); 1391 1392 get_processor_features(); 1393 }