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 __ vzeroupper(); 473 VM_Version::clean_cpuFeatures(); 474 UseAVX = saved_useavx; 475 UseSSE = saved_usesse; 476 __ jmp(wrapup); 477 } 478 479 __ bind(legacy_save_restore); 480 // AVX check 481 VM_Version::set_avx_cpuFeatures(); // Enable temporary to pass asserts 482 UseAVX = 1; 483 UseSSE = 2; 484 __ lea(rsi, Address(rbp, in_bytes(VM_Version::ymm_save_offset()))); 485 __ vmovdqu(Address(rsi, 0), xmm0); 486 __ vmovdqu(Address(rsi, 32), xmm7); 487 #ifdef _LP64 488 __ vmovdqu(Address(rsi, 64), xmm8); 489 __ vmovdqu(Address(rsi, 96), xmm15); 490 #endif 491 492 #ifdef _WINDOWS 493 #ifdef _LP64 494 __ vmovdqu(xmm15, Address(rsp, 0)); 495 __ addptr(rsp, 32); 496 __ vmovdqu(xmm8, Address(rsp, 0)); 497 __ addptr(rsp, 32); 498 #endif // _LP64 499 __ vmovdqu(xmm7, Address(rsp, 0)); 500 __ addptr(rsp, 32); 501 #endif // _WINDOWS 502 __ vzeroupper(); 503 VM_Version::clean_cpuFeatures(); 504 UseAVX = saved_useavx; 505 UseSSE = saved_usesse; 506 507 __ bind(wrapup); 508 __ popf(); 509 __ pop(rsi); 510 __ pop(rbx); 511 __ pop(rbp); 512 __ ret(0); 513 514 # undef __ 515 516 return start; 517 }; 518 }; 519 520 void VM_Version::get_processor_features() { 521 522 _cpu = 4; // 486 by default 523 _model = 0; 524 _stepping = 0; 525 _features = 0; 526 _logical_processors_per_package = 1; 527 // i486 internal cache is both I&D and has a 16-byte line size 528 _L1_data_cache_line_size = 16; 529 530 // Get raw processor info 531 532 get_cpu_info_stub(&_cpuid_info); 533 534 assert_is_initialized(); 535 _cpu = extended_cpu_family(); 536 _model = extended_cpu_model(); 537 _stepping = cpu_stepping(); 538 539 if (cpu_family() > 4) { // it supports CPUID 540 _features = feature_flags(); 541 // Logical processors are only available on P4s and above, 542 // and only if hyperthreading is available. 543 _logical_processors_per_package = logical_processor_count(); 544 _L1_data_cache_line_size = L1_line_size(); 545 } 546 547 _supports_cx8 = supports_cmpxchg8(); 548 // xchg and xadd instructions 549 _supports_atomic_getset4 = true; 550 _supports_atomic_getadd4 = true; 551 LP64_ONLY(_supports_atomic_getset8 = true); 552 LP64_ONLY(_supports_atomic_getadd8 = true); 553 554 #ifdef _LP64 555 // OS should support SSE for x64 and hardware should support at least SSE2. 556 if (!VM_Version::supports_sse2()) { 557 vm_exit_during_initialization("Unknown x64 processor: SSE2 not supported"); 558 } 559 // in 64 bit the use of SSE2 is the minimum 560 if (UseSSE < 2) UseSSE = 2; 561 #endif 562 563 #ifdef AMD64 564 // flush_icache_stub have to be generated first. 565 // That is why Icache line size is hard coded in ICache class, 566 // see icache_x86.hpp. It is also the reason why we can't use 567 // clflush instruction in 32-bit VM since it could be running 568 // on CPU which does not support it. 569 // 570 // The only thing we can do is to verify that flushed 571 // ICache::line_size has correct value. 572 guarantee(_cpuid_info.std_cpuid1_edx.bits.clflush != 0, "clflush is not supported"); 573 // clflush_size is size in quadwords (8 bytes). 574 guarantee(_cpuid_info.std_cpuid1_ebx.bits.clflush_size == 8, "such clflush size is not supported"); 575 #endif 576 577 // If the OS doesn't support SSE, we can't use this feature even if the HW does 578 if (!os::supports_sse()) 579 _features &= ~(CPU_SSE|CPU_SSE2|CPU_SSE3|CPU_SSSE3|CPU_SSE4A|CPU_SSE4_1|CPU_SSE4_2); 580 581 if (UseSSE < 4) { 582 _features &= ~CPU_SSE4_1; 583 _features &= ~CPU_SSE4_2; 584 } 585 586 if (UseSSE < 3) { 587 _features &= ~CPU_SSE3; 588 _features &= ~CPU_SSSE3; 589 _features &= ~CPU_SSE4A; 590 } 591 592 if (UseSSE < 2) 593 _features &= ~CPU_SSE2; 594 595 if (UseSSE < 1) 596 _features &= ~CPU_SSE; 597 598 // first try initial setting and detect what we can support 599 if (UseAVX > 0) { 600 if (UseAVX > 2 && supports_evex()) { 601 UseAVX = 3; 602 } else if (UseAVX > 1 && supports_avx2()) { 603 UseAVX = 2; 604 } else if (UseAVX > 0 && supports_avx()) { 605 UseAVX = 1; 606 } else { 607 UseAVX = 0; 608 } 609 } else if (UseAVX < 0) { 610 UseAVX = 0; 611 } 612 613 if (UseAVX < 3) { 614 _features &= ~CPU_AVX512F; 615 _features &= ~CPU_AVX512DQ; 616 _features &= ~CPU_AVX512CD; 617 _features &= ~CPU_AVX512BW; 618 _features &= ~CPU_AVX512VL; 619 } 620 621 if (UseAVX < 2) 622 _features &= ~CPU_AVX2; 623 624 if (UseAVX < 1) { 625 _features &= ~CPU_AVX; 626 _features &= ~CPU_VZEROUPPER; 627 } 628 629 if (!UseAES && !FLAG_IS_DEFAULT(UseAES)) 630 _features &= ~CPU_AES; 631 632 if (logical_processors_per_package() == 1) { 633 // HT processor could be installed on a system which doesn't support HT. 634 _features &= ~CPU_HT; 635 } 636 637 if( is_intel() ) { // Intel cpus specific settings 638 if ((cpu_family() == 0x06) && 639 ((extended_cpu_model() == 0x57) || // Xeon Phi 3200/5200/7200 640 (extended_cpu_model() == 0x85))) { // Future Xeon Phi 641 _features &= ~CPU_VZEROUPPER; 642 } 643 } 644 645 char buf[256]; 646 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", 647 cores_per_cpu(), threads_per_core(), 648 cpu_family(), _model, _stepping, 649 (supports_cmov() ? ", cmov" : ""), 650 (supports_cmpxchg8() ? ", cx8" : ""), 651 (supports_fxsr() ? ", fxsr" : ""), 652 (supports_mmx() ? ", mmx" : ""), 653 (supports_sse() ? ", sse" : ""), 654 (supports_sse2() ? ", sse2" : ""), 655 (supports_sse3() ? ", sse3" : ""), 656 (supports_ssse3()? ", ssse3": ""), 657 (supports_sse4_1() ? ", sse4.1" : ""), 658 (supports_sse4_2() ? ", sse4.2" : ""), 659 (supports_popcnt() ? ", popcnt" : ""), 660 (supports_avx() ? ", avx" : ""), 661 (supports_avx2() ? ", avx2" : ""), 662 (supports_aes() ? ", aes" : ""), 663 (supports_clmul() ? ", clmul" : ""), 664 (supports_erms() ? ", erms" : ""), 665 (supports_rtm() ? ", rtm" : ""), 666 (supports_mmx_ext() ? ", mmxext" : ""), 667 (supports_3dnow_prefetch() ? ", 3dnowpref" : ""), 668 (supports_lzcnt() ? ", lzcnt": ""), 669 (supports_sse4a() ? ", sse4a": ""), 670 (supports_ht() ? ", ht": ""), 671 (supports_tsc() ? ", tsc": ""), 672 (supports_tscinv_bit() ? ", tscinvbit": ""), 673 (supports_tscinv() ? ", tscinv": ""), 674 (supports_bmi1() ? ", bmi1" : ""), 675 (supports_bmi2() ? ", bmi2" : ""), 676 (supports_adx() ? ", adx" : ""), 677 (supports_evex() ? ", evex" : ""), 678 (supports_sha() ? ", sha" : ""), 679 (supports_fma() ? ", fma" : "")); 680 _features_string = os::strdup(buf); 681 682 // UseSSE is set to the smaller of what hardware supports and what 683 // the command line requires. I.e., you cannot set UseSSE to 2 on 684 // older Pentiums which do not support it. 685 if (UseSSE > 4) UseSSE=4; 686 if (UseSSE < 0) UseSSE=0; 687 if (!supports_sse4_1()) // Drop to 3 if no SSE4 support 688 UseSSE = MIN2((intx)3,UseSSE); 689 if (!supports_sse3()) // Drop to 2 if no SSE3 support 690 UseSSE = MIN2((intx)2,UseSSE); 691 if (!supports_sse2()) // Drop to 1 if no SSE2 support 692 UseSSE = MIN2((intx)1,UseSSE); 693 if (!supports_sse ()) // Drop to 0 if no SSE support 694 UseSSE = 0; 695 696 // Use AES instructions if available. 697 if (supports_aes()) { 698 if (FLAG_IS_DEFAULT(UseAES)) { 699 FLAG_SET_DEFAULT(UseAES, true); 700 } 701 if (!UseAES) { 702 if (UseAESIntrinsics && !FLAG_IS_DEFAULT(UseAESIntrinsics)) { 703 warning("AES intrinsics require UseAES flag to be enabled. Intrinsics will be disabled."); 704 } 705 FLAG_SET_DEFAULT(UseAESIntrinsics, false); 706 } else { 707 if (UseSSE > 2) { 708 if (FLAG_IS_DEFAULT(UseAESIntrinsics)) { 709 FLAG_SET_DEFAULT(UseAESIntrinsics, true); 710 } 711 } else { 712 // The AES intrinsic stubs require AES instruction support (of course) 713 // but also require sse3 mode or higher for instructions it use. 714 if (UseAESIntrinsics && !FLAG_IS_DEFAULT(UseAESIntrinsics)) { 715 warning("X86 AES intrinsics require SSE3 instructions or higher. Intrinsics will be disabled."); 716 } 717 FLAG_SET_DEFAULT(UseAESIntrinsics, false); 718 } 719 720 // --AES-CTR begins-- 721 if (!UseAESIntrinsics) { 722 if (UseAESCTRIntrinsics && !FLAG_IS_DEFAULT(UseAESCTRIntrinsics)) { 723 warning("AES-CTR intrinsics require UseAESIntrinsics flag to be enabled. Intrinsics will be disabled."); 724 FLAG_SET_DEFAULT(UseAESCTRIntrinsics, false); 725 } 726 } else { 727 if(supports_sse4_1()) { 728 if (FLAG_IS_DEFAULT(UseAESCTRIntrinsics)) { 729 FLAG_SET_DEFAULT(UseAESCTRIntrinsics, true); 730 } 731 } else { 732 // The AES-CTR intrinsic stubs require AES instruction support (of course) 733 // but also require sse4.1 mode or higher for instructions it use. 734 if (UseAESCTRIntrinsics && !FLAG_IS_DEFAULT(UseAESCTRIntrinsics)) { 735 warning("X86 AES-CTR intrinsics require SSE4.1 instructions or higher. Intrinsics will be disabled."); 736 } 737 FLAG_SET_DEFAULT(UseAESCTRIntrinsics, false); 738 } 739 } 740 // --AES-CTR ends-- 741 } 742 } else if (UseAES || UseAESIntrinsics || UseAESCTRIntrinsics) { 743 if (UseAES && !FLAG_IS_DEFAULT(UseAES)) { 744 warning("AES instructions are not available on this CPU"); 745 FLAG_SET_DEFAULT(UseAES, false); 746 } 747 if (UseAESIntrinsics && !FLAG_IS_DEFAULT(UseAESIntrinsics)) { 748 warning("AES intrinsics are not available on this CPU"); 749 FLAG_SET_DEFAULT(UseAESIntrinsics, false); 750 } 751 if (UseAESCTRIntrinsics && !FLAG_IS_DEFAULT(UseAESCTRIntrinsics)) { 752 warning("AES-CTR intrinsics are not available on this CPU"); 753 FLAG_SET_DEFAULT(UseAESCTRIntrinsics, false); 754 } 755 } 756 757 // Use CLMUL instructions if available. 758 if (supports_clmul()) { 759 if (FLAG_IS_DEFAULT(UseCLMUL)) { 760 UseCLMUL = true; 761 } 762 } else if (UseCLMUL) { 763 if (!FLAG_IS_DEFAULT(UseCLMUL)) 764 warning("CLMUL instructions not available on this CPU (AVX may also be required)"); 765 FLAG_SET_DEFAULT(UseCLMUL, false); 766 } 767 768 if (UseCLMUL && (UseSSE > 2)) { 769 if (FLAG_IS_DEFAULT(UseCRC32Intrinsics)) { 770 UseCRC32Intrinsics = true; 771 } 772 } else if (UseCRC32Intrinsics) { 773 if (!FLAG_IS_DEFAULT(UseCRC32Intrinsics)) 774 warning("CRC32 Intrinsics requires CLMUL instructions (not available on this CPU)"); 775 FLAG_SET_DEFAULT(UseCRC32Intrinsics, false); 776 } 777 778 if (supports_sse4_2()) { 779 if (FLAG_IS_DEFAULT(UseCRC32CIntrinsics)) { 780 UseCRC32CIntrinsics = true; 781 } 782 } 783 else if (UseCRC32CIntrinsics) { 784 if (!FLAG_IS_DEFAULT(UseCRC32CIntrinsics)) { 785 warning("CRC32C intrinsics are not available on this CPU"); 786 } 787 FLAG_SET_DEFAULT(UseCRC32CIntrinsics, false); 788 } 789 790 // GHASH/GCM intrinsics 791 if (UseCLMUL && (UseSSE > 2)) { 792 if (FLAG_IS_DEFAULT(UseGHASHIntrinsics)) { 793 UseGHASHIntrinsics = true; 794 } 795 } else if (UseGHASHIntrinsics) { 796 if (!FLAG_IS_DEFAULT(UseGHASHIntrinsics)) 797 warning("GHASH intrinsic requires CLMUL and SSE2 instructions on this CPU"); 798 FLAG_SET_DEFAULT(UseGHASHIntrinsics, false); 799 } 800 801 if (supports_fma() && UseSSE >= 2) { 802 if (FLAG_IS_DEFAULT(UseFMA)) { 803 UseFMA = true; 804 } 805 } else if (UseFMA) { 806 warning("FMA instructions are not available on this CPU"); 807 FLAG_SET_DEFAULT(UseFMA, false); 808 } 809 810 if (supports_sha() LP64_ONLY(|| supports_avx2() && supports_bmi2())) { 811 if (FLAG_IS_DEFAULT(UseSHA)) { 812 UseSHA = true; 813 } 814 } else if (UseSHA) { 815 warning("SHA instructions are not available on this CPU"); 816 FLAG_SET_DEFAULT(UseSHA, false); 817 } 818 819 if (supports_sha() && UseSHA) { 820 if (FLAG_IS_DEFAULT(UseSHA1Intrinsics)) { 821 FLAG_SET_DEFAULT(UseSHA1Intrinsics, true); 822 } 823 } else if (UseSHA1Intrinsics) { 824 warning("Intrinsics for SHA-1 crypto hash functions not available on this CPU."); 825 FLAG_SET_DEFAULT(UseSHA1Intrinsics, false); 826 } 827 828 if (UseSHA) { 829 if (FLAG_IS_DEFAULT(UseSHA256Intrinsics)) { 830 FLAG_SET_DEFAULT(UseSHA256Intrinsics, true); 831 } 832 } else if (UseSHA256Intrinsics) { 833 warning("Intrinsics for SHA-224 and SHA-256 crypto hash functions not available on this CPU."); 834 FLAG_SET_DEFAULT(UseSHA256Intrinsics, false); 835 } 836 837 if (UseSHA) { 838 if (FLAG_IS_DEFAULT(UseSHA512Intrinsics)) { 839 FLAG_SET_DEFAULT(UseSHA512Intrinsics, true); 840 } 841 } else if (UseSHA512Intrinsics) { 842 warning("Intrinsics for SHA-384 and SHA-512 crypto hash functions not available on this CPU."); 843 FLAG_SET_DEFAULT(UseSHA512Intrinsics, false); 844 } 845 846 if (!(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics)) { 847 FLAG_SET_DEFAULT(UseSHA, false); 848 } 849 850 if (UseAdler32Intrinsics) { 851 warning("Adler32Intrinsics not available on this CPU."); 852 FLAG_SET_DEFAULT(UseAdler32Intrinsics, false); 853 } 854 855 if (!supports_rtm() && UseRTMLocking) { 856 // Can't continue because UseRTMLocking affects UseBiasedLocking flag 857 // setting during arguments processing. See use_biased_locking(). 858 // VM_Version_init() is executed after UseBiasedLocking is used 859 // in Thread::allocate(). 860 vm_exit_during_initialization("RTM instructions are not available on this CPU"); 861 } 862 863 #if INCLUDE_RTM_OPT 864 if (UseRTMLocking) { 865 if (is_client_compilation_mode_vm()) { 866 // Only C2 does RTM locking optimization. 867 // Can't continue because UseRTMLocking affects UseBiasedLocking flag 868 // setting during arguments processing. See use_biased_locking(). 869 vm_exit_during_initialization("RTM locking optimization is not supported in emulated client VM"); 870 } 871 if (is_intel_family_core()) { 872 if ((_model == CPU_MODEL_HASWELL_E3) || 873 (_model == CPU_MODEL_HASWELL_E7 && _stepping < 3) || 874 (_model == CPU_MODEL_BROADWELL && _stepping < 4)) { 875 // currently a collision between SKL and HSW_E3 876 if (!UnlockExperimentalVMOptions && UseAVX < 3) { 877 vm_exit_during_initialization("UseRTMLocking is only available as experimental option on this platform. It must be enabled via -XX:+UnlockExperimentalVMOptions flag."); 878 } else { 879 warning("UseRTMLocking is only available as experimental option on this platform."); 880 } 881 } 882 } 883 if (!FLAG_IS_CMDLINE(UseRTMLocking)) { 884 // RTM locking should be used only for applications with 885 // high lock contention. For now we do not use it by default. 886 vm_exit_during_initialization("UseRTMLocking flag should be only set on command line"); 887 } 888 if (!is_power_of_2(RTMTotalCountIncrRate)) { 889 warning("RTMTotalCountIncrRate must be a power of 2, resetting it to 64"); 890 FLAG_SET_DEFAULT(RTMTotalCountIncrRate, 64); 891 } 892 if (RTMAbortRatio < 0 || RTMAbortRatio > 100) { 893 warning("RTMAbortRatio must be in the range 0 to 100, resetting it to 50"); 894 FLAG_SET_DEFAULT(RTMAbortRatio, 50); 895 } 896 } else { // !UseRTMLocking 897 if (UseRTMForStackLocks) { 898 if (!FLAG_IS_DEFAULT(UseRTMForStackLocks)) { 899 warning("UseRTMForStackLocks flag should be off when UseRTMLocking flag is off"); 900 } 901 FLAG_SET_DEFAULT(UseRTMForStackLocks, false); 902 } 903 if (UseRTMDeopt) { 904 FLAG_SET_DEFAULT(UseRTMDeopt, false); 905 } 906 if (PrintPreciseRTMLockingStatistics) { 907 FLAG_SET_DEFAULT(PrintPreciseRTMLockingStatistics, false); 908 } 909 } 910 #else 911 if (UseRTMLocking) { 912 // Only C2 does RTM locking optimization. 913 // Can't continue because UseRTMLocking affects UseBiasedLocking flag 914 // setting during arguments processing. See use_biased_locking(). 915 vm_exit_during_initialization("RTM locking optimization is not supported in this VM"); 916 } 917 #endif 918 919 #ifdef COMPILER2 920 if (UseFPUForSpilling) { 921 if (UseSSE < 2) { 922 // Only supported with SSE2+ 923 FLAG_SET_DEFAULT(UseFPUForSpilling, false); 924 } 925 } 926 #endif 927 #if defined(COMPILER2) || INCLUDE_JVMCI 928 if (MaxVectorSize > 0) { 929 if (!is_power_of_2(MaxVectorSize)) { 930 warning("MaxVectorSize must be a power of 2"); 931 FLAG_SET_DEFAULT(MaxVectorSize, 64); 932 } 933 if (UseSSE < 2) { 934 // Vectors (in XMM) are only supported with SSE2+ 935 if (MaxVectorSize > 0) { 936 if (!FLAG_IS_DEFAULT(MaxVectorSize)) 937 warning("MaxVectorSize must be 0"); 938 FLAG_SET_DEFAULT(MaxVectorSize, 0); 939 } 940 } 941 else if (UseAVX == 0 || !os_supports_avx_vectors()) { 942 // 32 bytes vectors (in YMM) are only supported with AVX+ 943 if (MaxVectorSize > 16) { 944 if (!FLAG_IS_DEFAULT(MaxVectorSize)) 945 warning("MaxVectorSize must be <= 16"); 946 FLAG_SET_DEFAULT(MaxVectorSize, 16); 947 } 948 } 949 else if (UseAVX == 1 || UseAVX == 2) { 950 // 64 bytes vectors (in ZMM) are only supported with AVX 3 951 if (MaxVectorSize > 32) { 952 if (!FLAG_IS_DEFAULT(MaxVectorSize)) 953 warning("MaxVectorSize must be <= 32"); 954 FLAG_SET_DEFAULT(MaxVectorSize, 32); 955 } 956 } 957 else if (UseAVX > 2 ) { 958 if (MaxVectorSize > 64) { 959 if (!FLAG_IS_DEFAULT(MaxVectorSize)) 960 warning("MaxVectorSize must be <= 64"); 961 FLAG_SET_DEFAULT(MaxVectorSize, 64); 962 } 963 } 964 #if defined(COMPILER2) && defined(ASSERT) 965 if (supports_avx() && PrintMiscellaneous && Verbose && TraceNewVectors) { 966 tty->print_cr("State of YMM registers after signal handle:"); 967 int nreg = 2 LP64_ONLY(+2); 968 const char* ymm_name[4] = {"0", "7", "8", "15"}; 969 for (int i = 0; i < nreg; i++) { 970 tty->print("YMM%s:", ymm_name[i]); 971 for (int j = 7; j >=0; j--) { 972 tty->print(" %x", _cpuid_info.ymm_save[i*8 + j]); 973 } 974 tty->cr(); 975 } 976 } 977 #endif // COMPILER2 && ASSERT 978 } 979 #endif // COMPILER2 || INCLUDE_JVMCI 980 981 #ifdef COMPILER2 982 #ifdef _LP64 983 if (FLAG_IS_DEFAULT(UseMultiplyToLenIntrinsic)) { 984 UseMultiplyToLenIntrinsic = true; 985 } 986 if (FLAG_IS_DEFAULT(UseSquareToLenIntrinsic)) { 987 UseSquareToLenIntrinsic = true; 988 } 989 if (FLAG_IS_DEFAULT(UseMulAddIntrinsic)) { 990 UseMulAddIntrinsic = true; 991 } 992 if (FLAG_IS_DEFAULT(UseMontgomeryMultiplyIntrinsic)) { 993 UseMontgomeryMultiplyIntrinsic = true; 994 } 995 if (FLAG_IS_DEFAULT(UseMontgomerySquareIntrinsic)) { 996 UseMontgomerySquareIntrinsic = true; 997 } 998 #else 999 if (UseMultiplyToLenIntrinsic) { 1000 if (!FLAG_IS_DEFAULT(UseMultiplyToLenIntrinsic)) { 1001 warning("multiplyToLen intrinsic is not available in 32-bit VM"); 1002 } 1003 FLAG_SET_DEFAULT(UseMultiplyToLenIntrinsic, false); 1004 } 1005 if (UseMontgomeryMultiplyIntrinsic) { 1006 if (!FLAG_IS_DEFAULT(UseMontgomeryMultiplyIntrinsic)) { 1007 warning("montgomeryMultiply intrinsic is not available in 32-bit VM"); 1008 } 1009 FLAG_SET_DEFAULT(UseMontgomeryMultiplyIntrinsic, false); 1010 } 1011 if (UseMontgomerySquareIntrinsic) { 1012 if (!FLAG_IS_DEFAULT(UseMontgomerySquareIntrinsic)) { 1013 warning("montgomerySquare intrinsic is not available in 32-bit VM"); 1014 } 1015 FLAG_SET_DEFAULT(UseMontgomerySquareIntrinsic, false); 1016 } 1017 if (UseSquareToLenIntrinsic) { 1018 if (!FLAG_IS_DEFAULT(UseSquareToLenIntrinsic)) { 1019 warning("squareToLen intrinsic is not available in 32-bit VM"); 1020 } 1021 FLAG_SET_DEFAULT(UseSquareToLenIntrinsic, false); 1022 } 1023 if (UseMulAddIntrinsic) { 1024 if (!FLAG_IS_DEFAULT(UseMulAddIntrinsic)) { 1025 warning("mulAdd intrinsic is not available in 32-bit VM"); 1026 } 1027 FLAG_SET_DEFAULT(UseMulAddIntrinsic, false); 1028 } 1029 #endif 1030 #endif // COMPILER2 1031 1032 // On new cpus instructions which update whole XMM register should be used 1033 // to prevent partial register stall due to dependencies on high half. 1034 // 1035 // UseXmmLoadAndClearUpper == true --> movsd(xmm, mem) 1036 // UseXmmLoadAndClearUpper == false --> movlpd(xmm, mem) 1037 // UseXmmRegToRegMoveAll == true --> movaps(xmm, xmm), movapd(xmm, xmm). 1038 // UseXmmRegToRegMoveAll == false --> movss(xmm, xmm), movsd(xmm, xmm). 1039 1040 if( is_amd() ) { // AMD cpus specific settings 1041 if( supports_sse2() && FLAG_IS_DEFAULT(UseAddressNop) ) { 1042 // Use it on new AMD cpus starting from Opteron. 1043 UseAddressNop = true; 1044 } 1045 if( supports_sse2() && FLAG_IS_DEFAULT(UseNewLongLShift) ) { 1046 // Use it on new AMD cpus starting from Opteron. 1047 UseNewLongLShift = true; 1048 } 1049 if( FLAG_IS_DEFAULT(UseXmmLoadAndClearUpper) ) { 1050 if (supports_sse4a()) { 1051 UseXmmLoadAndClearUpper = true; // use movsd only on '10h' Opteron 1052 } else { 1053 UseXmmLoadAndClearUpper = false; 1054 } 1055 } 1056 if( FLAG_IS_DEFAULT(UseXmmRegToRegMoveAll) ) { 1057 if( supports_sse4a() ) { 1058 UseXmmRegToRegMoveAll = true; // use movaps, movapd only on '10h' 1059 } else { 1060 UseXmmRegToRegMoveAll = false; 1061 } 1062 } 1063 if( FLAG_IS_DEFAULT(UseXmmI2F) ) { 1064 if( supports_sse4a() ) { 1065 UseXmmI2F = true; 1066 } else { 1067 UseXmmI2F = false; 1068 } 1069 } 1070 if( FLAG_IS_DEFAULT(UseXmmI2D) ) { 1071 if( supports_sse4a() ) { 1072 UseXmmI2D = true; 1073 } else { 1074 UseXmmI2D = false; 1075 } 1076 } 1077 if (supports_sse4_2()) { 1078 if (FLAG_IS_DEFAULT(UseSSE42Intrinsics)) { 1079 FLAG_SET_DEFAULT(UseSSE42Intrinsics, true); 1080 } 1081 } else { 1082 if (UseSSE42Intrinsics && !FLAG_IS_DEFAULT(UseAESIntrinsics)) { 1083 warning("SSE4.2 intrinsics require SSE4.2 instructions or higher. Intrinsics will be disabled."); 1084 } 1085 FLAG_SET_DEFAULT(UseSSE42Intrinsics, false); 1086 } 1087 1088 // some defaults for AMD family 15h 1089 if ( cpu_family() == 0x15 ) { 1090 // On family 15h processors default is no sw prefetch 1091 if (FLAG_IS_DEFAULT(AllocatePrefetchStyle)) { 1092 AllocatePrefetchStyle = 0; 1093 } 1094 // Also, if some other prefetch style is specified, default instruction type is PREFETCHW 1095 if (FLAG_IS_DEFAULT(AllocatePrefetchInstr)) { 1096 AllocatePrefetchInstr = 3; 1097 } 1098 // On family 15h processors use XMM and UnalignedLoadStores for Array Copy 1099 if (supports_sse2() && FLAG_IS_DEFAULT(UseXMMForArrayCopy)) { 1100 UseXMMForArrayCopy = true; 1101 } 1102 if (supports_sse2() && FLAG_IS_DEFAULT(UseUnalignedLoadStores)) { 1103 UseUnalignedLoadStores = true; 1104 } 1105 } 1106 1107 #ifdef COMPILER2 1108 if (MaxVectorSize > 16) { 1109 // Limit vectors size to 16 bytes on current AMD cpus. 1110 FLAG_SET_DEFAULT(MaxVectorSize, 16); 1111 } 1112 #endif // COMPILER2 1113 } 1114 1115 if( is_intel() ) { // Intel cpus specific settings 1116 if( FLAG_IS_DEFAULT(UseStoreImmI16) ) { 1117 UseStoreImmI16 = false; // don't use it on Intel cpus 1118 } 1119 if( cpu_family() == 6 || cpu_family() == 15 ) { 1120 if( FLAG_IS_DEFAULT(UseAddressNop) ) { 1121 // Use it on all Intel cpus starting from PentiumPro 1122 UseAddressNop = true; 1123 } 1124 } 1125 if( FLAG_IS_DEFAULT(UseXmmLoadAndClearUpper) ) { 1126 UseXmmLoadAndClearUpper = true; // use movsd on all Intel cpus 1127 } 1128 if( FLAG_IS_DEFAULT(UseXmmRegToRegMoveAll) ) { 1129 if( supports_sse3() ) { 1130 UseXmmRegToRegMoveAll = true; // use movaps, movapd on new Intel cpus 1131 } else { 1132 UseXmmRegToRegMoveAll = false; 1133 } 1134 } 1135 if( cpu_family() == 6 && supports_sse3() ) { // New Intel cpus 1136 #ifdef COMPILER2 1137 if( FLAG_IS_DEFAULT(MaxLoopPad) ) { 1138 // For new Intel cpus do the next optimization: 1139 // don't align the beginning of a loop if there are enough instructions 1140 // left (NumberOfLoopInstrToAlign defined in c2_globals.hpp) 1141 // in current fetch line (OptoLoopAlignment) or the padding 1142 // is big (> MaxLoopPad). 1143 // Set MaxLoopPad to 11 for new Intel cpus to reduce number of 1144 // generated NOP instructions. 11 is the largest size of one 1145 // address NOP instruction '0F 1F' (see Assembler::nop(i)). 1146 MaxLoopPad = 11; 1147 } 1148 #endif // COMPILER2 1149 if (FLAG_IS_DEFAULT(UseXMMForArrayCopy)) { 1150 UseXMMForArrayCopy = true; // use SSE2 movq on new Intel cpus 1151 } 1152 if (supports_sse4_2() && supports_ht()) { // Newest Intel cpus 1153 if (FLAG_IS_DEFAULT(UseUnalignedLoadStores)) { 1154 UseUnalignedLoadStores = true; // use movdqu on newest Intel cpus 1155 } 1156 } 1157 if (supports_sse4_2()) { 1158 if (FLAG_IS_DEFAULT(UseSSE42Intrinsics)) { 1159 FLAG_SET_DEFAULT(UseSSE42Intrinsics, true); 1160 } 1161 } else { 1162 if (UseSSE42Intrinsics && !FLAG_IS_DEFAULT(UseAESIntrinsics)) { 1163 warning("SSE4.2 intrinsics require SSE4.2 instructions or higher. Intrinsics will be disabled."); 1164 } 1165 FLAG_SET_DEFAULT(UseSSE42Intrinsics, false); 1166 } 1167 } 1168 if ((cpu_family() == 0x06) && 1169 ((extended_cpu_model() == 0x36) || // Centerton 1170 (extended_cpu_model() == 0x37) || // Silvermont 1171 (extended_cpu_model() == 0x4D))) { 1172 #ifdef COMPILER2 1173 if (FLAG_IS_DEFAULT(OptoScheduling)) { 1174 OptoScheduling = true; 1175 } 1176 #endif 1177 if (supports_sse4_2()) { // Silvermont 1178 if (FLAG_IS_DEFAULT(UseUnalignedLoadStores)) { 1179 UseUnalignedLoadStores = true; // use movdqu on newest Intel cpus 1180 } 1181 } 1182 } 1183 if(FLAG_IS_DEFAULT(AllocatePrefetchInstr) && supports_3dnow_prefetch()) { 1184 AllocatePrefetchInstr = 3; 1185 } 1186 } 1187 1188 #ifdef _LP64 1189 if (UseSSE42Intrinsics) { 1190 if (FLAG_IS_DEFAULT(UseVectorizedMismatchIntrinsic)) { 1191 UseVectorizedMismatchIntrinsic = true; 1192 } 1193 } else if (UseVectorizedMismatchIntrinsic) { 1194 if (!FLAG_IS_DEFAULT(UseVectorizedMismatchIntrinsic)) 1195 warning("vectorizedMismatch intrinsics are not available on this CPU"); 1196 FLAG_SET_DEFAULT(UseVectorizedMismatchIntrinsic, false); 1197 } 1198 #else 1199 if (UseVectorizedMismatchIntrinsic) { 1200 if (!FLAG_IS_DEFAULT(UseVectorizedMismatchIntrinsic)) { 1201 warning("vectorizedMismatch intrinsic is not available in 32-bit VM"); 1202 } 1203 FLAG_SET_DEFAULT(UseVectorizedMismatchIntrinsic, false); 1204 } 1205 #endif // _LP64 1206 1207 // Use count leading zeros count instruction if available. 1208 if (supports_lzcnt()) { 1209 if (FLAG_IS_DEFAULT(UseCountLeadingZerosInstruction)) { 1210 UseCountLeadingZerosInstruction = true; 1211 } 1212 } else if (UseCountLeadingZerosInstruction) { 1213 warning("lzcnt instruction is not available on this CPU"); 1214 FLAG_SET_DEFAULT(UseCountLeadingZerosInstruction, false); 1215 } 1216 1217 // Use count trailing zeros instruction if available 1218 if (supports_bmi1()) { 1219 // tzcnt does not require VEX prefix 1220 if (FLAG_IS_DEFAULT(UseCountTrailingZerosInstruction)) { 1221 if (!UseBMI1Instructions && !FLAG_IS_DEFAULT(UseBMI1Instructions)) { 1222 // Don't use tzcnt if BMI1 is switched off on command line. 1223 UseCountTrailingZerosInstruction = false; 1224 } else { 1225 UseCountTrailingZerosInstruction = true; 1226 } 1227 } 1228 } else if (UseCountTrailingZerosInstruction) { 1229 warning("tzcnt instruction is not available on this CPU"); 1230 FLAG_SET_DEFAULT(UseCountTrailingZerosInstruction, false); 1231 } 1232 1233 // BMI instructions (except tzcnt) use an encoding with VEX prefix. 1234 // VEX prefix is generated only when AVX > 0. 1235 if (supports_bmi1() && supports_avx()) { 1236 if (FLAG_IS_DEFAULT(UseBMI1Instructions)) { 1237 UseBMI1Instructions = true; 1238 } 1239 } else if (UseBMI1Instructions) { 1240 warning("BMI1 instructions are not available on this CPU (AVX is also required)"); 1241 FLAG_SET_DEFAULT(UseBMI1Instructions, false); 1242 } 1243 1244 if (supports_bmi2() && supports_avx()) { 1245 if (FLAG_IS_DEFAULT(UseBMI2Instructions)) { 1246 UseBMI2Instructions = true; 1247 } 1248 } else if (UseBMI2Instructions) { 1249 warning("BMI2 instructions are not available on this CPU (AVX is also required)"); 1250 FLAG_SET_DEFAULT(UseBMI2Instructions, false); 1251 } 1252 1253 // Use population count instruction if available. 1254 if (supports_popcnt()) { 1255 if (FLAG_IS_DEFAULT(UsePopCountInstruction)) { 1256 UsePopCountInstruction = true; 1257 } 1258 } else if (UsePopCountInstruction) { 1259 warning("POPCNT instruction is not available on this CPU"); 1260 FLAG_SET_DEFAULT(UsePopCountInstruction, false); 1261 } 1262 1263 // Use fast-string operations if available. 1264 if (supports_erms()) { 1265 if (FLAG_IS_DEFAULT(UseFastStosb)) { 1266 UseFastStosb = true; 1267 } 1268 } else if (UseFastStosb) { 1269 warning("fast-string operations are not available on this CPU"); 1270 FLAG_SET_DEFAULT(UseFastStosb, false); 1271 } 1272 1273 #ifdef COMPILER2 1274 if (FLAG_IS_DEFAULT(AlignVector)) { 1275 // Modern processors allow misaligned memory operations for vectors. 1276 AlignVector = !UseUnalignedLoadStores; 1277 } 1278 #endif // COMPILER2 1279 1280 if( AllocatePrefetchInstr == 3 && !supports_3dnow_prefetch() ) AllocatePrefetchInstr=0; 1281 if( !supports_sse() && supports_3dnow_prefetch() ) AllocatePrefetchInstr = 3; 1282 1283 // Allocation prefetch settings 1284 intx cache_line_size = prefetch_data_size(); 1285 if( cache_line_size > AllocatePrefetchStepSize ) 1286 AllocatePrefetchStepSize = cache_line_size; 1287 1288 AllocatePrefetchDistance = allocate_prefetch_distance(); 1289 AllocatePrefetchStyle = allocate_prefetch_style(); 1290 1291 if (is_intel() && cpu_family() == 6 && supports_sse3()) { 1292 if (AllocatePrefetchStyle == 2) { // watermark prefetching on Core 1293 #ifdef _LP64 1294 AllocatePrefetchDistance = 384; 1295 #else 1296 AllocatePrefetchDistance = 320; 1297 #endif 1298 } 1299 if (supports_sse4_2() && supports_ht()) { // Nehalem based cpus 1300 AllocatePrefetchDistance = 192; 1301 if (FLAG_IS_DEFAULT(AllocatePrefetchLines)) { 1302 FLAG_SET_DEFAULT(AllocatePrefetchLines, 4); 1303 } 1304 } 1305 #ifdef COMPILER2 1306 if (supports_sse4_2()) { 1307 if (FLAG_IS_DEFAULT(UseFPUForSpilling)) { 1308 FLAG_SET_DEFAULT(UseFPUForSpilling, true); 1309 } 1310 } 1311 #endif 1312 } 1313 1314 #ifdef _LP64 1315 // Prefetch settings 1316 PrefetchCopyIntervalInBytes = prefetch_copy_interval_in_bytes(); 1317 PrefetchScanIntervalInBytes = prefetch_scan_interval_in_bytes(); 1318 PrefetchFieldsAhead = prefetch_fields_ahead(); 1319 #endif 1320 1321 if (FLAG_IS_DEFAULT(ContendedPaddingWidth) && 1322 (cache_line_size > ContendedPaddingWidth)) 1323 ContendedPaddingWidth = cache_line_size; 1324 1325 // This machine allows unaligned memory accesses 1326 if (FLAG_IS_DEFAULT(UseUnalignedAccesses)) { 1327 FLAG_SET_DEFAULT(UseUnalignedAccesses, true); 1328 } 1329 1330 #ifndef PRODUCT 1331 if (log_is_enabled(Info, os, cpu)) { 1332 outputStream* log = Log(os, cpu)::info_stream(); 1333 log->print_cr("Logical CPUs per core: %u", 1334 logical_processors_per_package()); 1335 log->print_cr("L1 data cache line size: %u", L1_data_cache_line_size()); 1336 log->print("UseSSE=%d", (int) UseSSE); 1337 if (UseAVX > 0) { 1338 log->print(" UseAVX=%d", (int) UseAVX); 1339 } 1340 if (UseAES) { 1341 log->print(" UseAES=1"); 1342 } 1343 #ifdef COMPILER2 1344 if (MaxVectorSize > 0) { 1345 log->print(" MaxVectorSize=%d", (int) MaxVectorSize); 1346 } 1347 #endif 1348 log->cr(); 1349 log->print("Allocation"); 1350 if (AllocatePrefetchStyle <= 0 || UseSSE == 0 && !supports_3dnow_prefetch()) { 1351 log->print_cr(": no prefetching"); 1352 } else { 1353 log->print(" prefetching: "); 1354 if (UseSSE == 0 && supports_3dnow_prefetch()) { 1355 log->print("PREFETCHW"); 1356 } else if (UseSSE >= 1) { 1357 if (AllocatePrefetchInstr == 0) { 1358 log->print("PREFETCHNTA"); 1359 } else if (AllocatePrefetchInstr == 1) { 1360 log->print("PREFETCHT0"); 1361 } else if (AllocatePrefetchInstr == 2) { 1362 log->print("PREFETCHT2"); 1363 } else if (AllocatePrefetchInstr == 3) { 1364 log->print("PREFETCHW"); 1365 } 1366 } 1367 if (AllocatePrefetchLines > 1) { 1368 log->print_cr(" at distance %d, %d lines of %d bytes", (int) AllocatePrefetchDistance, (int) AllocatePrefetchLines, (int) AllocatePrefetchStepSize); 1369 } else { 1370 log->print_cr(" at distance %d, one line of %d bytes", (int) AllocatePrefetchDistance, (int) AllocatePrefetchStepSize); 1371 } 1372 } 1373 1374 if (PrefetchCopyIntervalInBytes > 0) { 1375 log->print_cr("PrefetchCopyIntervalInBytes %d", (int) PrefetchCopyIntervalInBytes); 1376 } 1377 if (PrefetchScanIntervalInBytes > 0) { 1378 log->print_cr("PrefetchScanIntervalInBytes %d", (int) PrefetchScanIntervalInBytes); 1379 } 1380 if (PrefetchFieldsAhead > 0) { 1381 log->print_cr("PrefetchFieldsAhead %d", (int) PrefetchFieldsAhead); 1382 } 1383 if (ContendedPaddingWidth > 0) { 1384 log->print_cr("ContendedPaddingWidth %d", (int) ContendedPaddingWidth); 1385 } 1386 } 1387 #endif // !PRODUCT 1388 } 1389 1390 bool VM_Version::use_biased_locking() { 1391 #if INCLUDE_RTM_OPT 1392 // RTM locking is most useful when there is high lock contention and 1393 // low data contention. With high lock contention the lock is usually 1394 // inflated and biased locking is not suitable for that case. 1395 // RTM locking code requires that biased locking is off. 1396 // Note: we can't switch off UseBiasedLocking in get_processor_features() 1397 // because it is used by Thread::allocate() which is called before 1398 // VM_Version::initialize(). 1399 if (UseRTMLocking && UseBiasedLocking) { 1400 if (FLAG_IS_DEFAULT(UseBiasedLocking)) { 1401 FLAG_SET_DEFAULT(UseBiasedLocking, false); 1402 } else { 1403 warning("Biased locking is not supported with RTM locking; ignoring UseBiasedLocking flag." ); 1404 UseBiasedLocking = false; 1405 } 1406 } 1407 #endif 1408 return UseBiasedLocking; 1409 } 1410 1411 void VM_Version::initialize() { 1412 ResourceMark rm; 1413 // Making this stub must be FIRST use of assembler 1414 1415 stub_blob = BufferBlob::create("get_cpu_info_stub", stub_size); 1416 if (stub_blob == NULL) { 1417 vm_exit_during_initialization("Unable to allocate get_cpu_info_stub"); 1418 } 1419 CodeBuffer c(stub_blob); 1420 VM_Version_StubGenerator g(&c); 1421 get_cpu_info_stub = CAST_TO_FN_PTR(get_cpu_info_stub_t, 1422 g.generate_get_cpu_info()); 1423 1424 get_processor_features(); 1425 }