1 /* 2 * Copyright (c) 1997, 2017, 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 "jvm.h" 27 #include "asm/macroAssembler.hpp" 28 #include "asm/macroAssembler.inline.hpp" 29 #include "logging/log.hpp" 30 #include "logging/logStream.hpp" 31 #include "memory/resourceArea.hpp" 32 #include "runtime/java.hpp" 33 #include "runtime/os.hpp" 34 #include "runtime/stubCodeGenerator.hpp" 35 #include "vm_version_x86.hpp" 36 37 38 int VM_Version::_cpu; 39 int VM_Version::_model; 40 int VM_Version::_stepping; 41 VM_Version::CpuidInfo VM_Version::_cpuid_info = { 0, }; 42 43 // Address of instruction which causes SEGV 44 address VM_Version::_cpuinfo_segv_addr = 0; 45 // Address of instruction after the one which causes SEGV 46 address VM_Version::_cpuinfo_cont_addr = 0; 47 48 static BufferBlob* stub_blob; 49 static const int stub_size = 1100; 50 51 extern "C" { 52 typedef void (*get_cpu_info_stub_t)(void*); 53 } 54 static get_cpu_info_stub_t get_cpu_info_stub = NULL; 55 56 57 class VM_Version_StubGenerator: public StubCodeGenerator { 58 public: 59 60 VM_Version_StubGenerator(CodeBuffer *c) : StubCodeGenerator(c) {} 61 62 address generate_get_cpu_info() { 63 // Flags to test CPU type. 64 const uint32_t HS_EFL_AC = 0x40000; 65 const uint32_t HS_EFL_ID = 0x200000; 66 // Values for when we don't have a CPUID instruction. 67 const int CPU_FAMILY_SHIFT = 8; 68 const uint32_t CPU_FAMILY_386 = (3 << CPU_FAMILY_SHIFT); 69 const uint32_t CPU_FAMILY_486 = (4 << CPU_FAMILY_SHIFT); 70 bool use_evex = FLAG_IS_DEFAULT(UseAVX) || (UseAVX > 2); 71 72 Label detect_486, cpu486, detect_586, std_cpuid1, std_cpuid4; 73 Label sef_cpuid, ext_cpuid, ext_cpuid1, ext_cpuid5, ext_cpuid7, ext_cpuid8, done, wrapup; 74 Label legacy_setup, save_restore_except, legacy_save_restore, start_simd_check; 75 76 StubCodeMark mark(this, "VM_Version", "get_cpu_info_stub"); 77 # define __ _masm-> 78 79 address start = __ pc(); 80 81 // 82 // void get_cpu_info(VM_Version::CpuidInfo* cpuid_info); 83 // 84 // LP64: rcx and rdx are first and second argument registers on windows 85 86 __ push(rbp); 87 #ifdef _LP64 88 __ mov(rbp, c_rarg0); // cpuid_info address 89 #else 90 __ movptr(rbp, Address(rsp, 8)); // cpuid_info address 91 #endif 92 __ push(rbx); 93 __ push(rsi); 94 __ pushf(); // preserve rbx, and flags 95 __ pop(rax); 96 __ push(rax); 97 __ mov(rcx, rax); 98 // 99 // if we are unable to change the AC flag, we have a 386 100 // 101 __ xorl(rax, HS_EFL_AC); 102 __ push(rax); 103 __ popf(); 104 __ pushf(); 105 __ pop(rax); 106 __ cmpptr(rax, rcx); 107 __ jccb(Assembler::notEqual, detect_486); 108 109 __ movl(rax, CPU_FAMILY_386); 110 __ movl(Address(rbp, in_bytes(VM_Version::std_cpuid1_offset())), rax); 111 __ jmp(done); 112 113 // 114 // If we are unable to change the ID flag, we have a 486 which does 115 // not support the "cpuid" instruction. 116 // 117 __ bind(detect_486); 118 __ mov(rax, rcx); 119 __ xorl(rax, HS_EFL_ID); 120 __ push(rax); 121 __ popf(); 122 __ pushf(); 123 __ pop(rax); 124 __ cmpptr(rcx, rax); 125 __ jccb(Assembler::notEqual, detect_586); 126 127 __ bind(cpu486); 128 __ movl(rax, CPU_FAMILY_486); 129 __ movl(Address(rbp, in_bytes(VM_Version::std_cpuid1_offset())), rax); 130 __ jmp(done); 131 132 // 133 // At this point, we have a chip which supports the "cpuid" instruction 134 // 135 __ bind(detect_586); 136 __ xorl(rax, rax); 137 __ cpuid(); 138 __ orl(rax, rax); 139 __ jcc(Assembler::equal, cpu486); // if cpuid doesn't support an input 140 // value of at least 1, we give up and 141 // assume a 486 142 __ lea(rsi, Address(rbp, in_bytes(VM_Version::std_cpuid0_offset()))); 143 __ movl(Address(rsi, 0), rax); 144 __ movl(Address(rsi, 4), rbx); 145 __ movl(Address(rsi, 8), rcx); 146 __ movl(Address(rsi,12), rdx); 147 148 __ cmpl(rax, 0xa); // Is cpuid(0xB) supported? 149 __ jccb(Assembler::belowEqual, std_cpuid4); 150 151 // 152 // cpuid(0xB) Processor Topology 153 // 154 __ movl(rax, 0xb); 155 __ xorl(rcx, rcx); // Threads level 156 __ cpuid(); 157 158 __ lea(rsi, Address(rbp, in_bytes(VM_Version::tpl_cpuidB0_offset()))); 159 __ movl(Address(rsi, 0), rax); 160 __ movl(Address(rsi, 4), rbx); 161 __ movl(Address(rsi, 8), rcx); 162 __ movl(Address(rsi,12), rdx); 163 164 __ movl(rax, 0xb); 165 __ movl(rcx, 1); // Cores level 166 __ cpuid(); 167 __ push(rax); 168 __ andl(rax, 0x1f); // Determine if valid topology level 169 __ orl(rax, rbx); // eax[4:0] | ebx[0:15] == 0 indicates invalid level 170 __ andl(rax, 0xffff); 171 __ pop(rax); 172 __ jccb(Assembler::equal, std_cpuid4); 173 174 __ lea(rsi, Address(rbp, in_bytes(VM_Version::tpl_cpuidB1_offset()))); 175 __ movl(Address(rsi, 0), rax); 176 __ movl(Address(rsi, 4), rbx); 177 __ movl(Address(rsi, 8), rcx); 178 __ movl(Address(rsi,12), rdx); 179 180 __ movl(rax, 0xb); 181 __ movl(rcx, 2); // Packages level 182 __ cpuid(); 183 __ push(rax); 184 __ andl(rax, 0x1f); // Determine if valid topology level 185 __ orl(rax, rbx); // eax[4:0] | ebx[0:15] == 0 indicates invalid level 186 __ andl(rax, 0xffff); 187 __ pop(rax); 188 __ jccb(Assembler::equal, std_cpuid4); 189 190 __ lea(rsi, Address(rbp, in_bytes(VM_Version::tpl_cpuidB2_offset()))); 191 __ movl(Address(rsi, 0), rax); 192 __ movl(Address(rsi, 4), rbx); 193 __ movl(Address(rsi, 8), rcx); 194 __ movl(Address(rsi,12), rdx); 195 196 // 197 // cpuid(0x4) Deterministic cache params 198 // 199 __ bind(std_cpuid4); 200 __ movl(rax, 4); 201 __ cmpl(rax, Address(rbp, in_bytes(VM_Version::std_cpuid0_offset()))); // Is cpuid(0x4) supported? 202 __ jccb(Assembler::greater, std_cpuid1); 203 204 __ xorl(rcx, rcx); // L1 cache 205 __ cpuid(); 206 __ push(rax); 207 __ andl(rax, 0x1f); // Determine if valid cache parameters used 208 __ orl(rax, rax); // eax[4:0] == 0 indicates invalid cache 209 __ pop(rax); 210 __ jccb(Assembler::equal, std_cpuid1); 211 212 __ lea(rsi, Address(rbp, in_bytes(VM_Version::dcp_cpuid4_offset()))); 213 __ movl(Address(rsi, 0), rax); 214 __ movl(Address(rsi, 4), rbx); 215 __ movl(Address(rsi, 8), rcx); 216 __ movl(Address(rsi,12), rdx); 217 218 // 219 // Standard cpuid(0x1) 220 // 221 __ bind(std_cpuid1); 222 __ movl(rax, 1); 223 __ cpuid(); 224 __ lea(rsi, Address(rbp, in_bytes(VM_Version::std_cpuid1_offset()))); 225 __ movl(Address(rsi, 0), rax); 226 __ movl(Address(rsi, 4), rbx); 227 __ movl(Address(rsi, 8), rcx); 228 __ movl(Address(rsi,12), rdx); 229 230 // 231 // Check if OS has enabled XGETBV instruction to access XCR0 232 // (OSXSAVE feature flag) and CPU supports AVX 233 // 234 __ andl(rcx, 0x18000000); // cpuid1 bits osxsave | avx 235 __ cmpl(rcx, 0x18000000); 236 __ jccb(Assembler::notEqual, sef_cpuid); // jump if AVX is not supported 237 238 // 239 // XCR0, XFEATURE_ENABLED_MASK register 240 // 241 __ xorl(rcx, rcx); // zero for XCR0 register 242 __ xgetbv(); 243 __ lea(rsi, Address(rbp, in_bytes(VM_Version::xem_xcr0_offset()))); 244 __ movl(Address(rsi, 0), rax); 245 __ movl(Address(rsi, 4), rdx); 246 247 // 248 // cpuid(0x7) Structured Extended Features 249 // 250 __ bind(sef_cpuid); 251 __ movl(rax, 7); 252 __ cmpl(rax, Address(rbp, in_bytes(VM_Version::std_cpuid0_offset()))); // Is cpuid(0x7) supported? 253 __ jccb(Assembler::greater, ext_cpuid); 254 255 __ xorl(rcx, rcx); 256 __ cpuid(); 257 __ lea(rsi, Address(rbp, in_bytes(VM_Version::sef_cpuid7_offset()))); 258 __ movl(Address(rsi, 0), rax); 259 __ movl(Address(rsi, 4), rbx); 260 261 // 262 // Extended cpuid(0x80000000) 263 // 264 __ bind(ext_cpuid); 265 __ movl(rax, 0x80000000); 266 __ cpuid(); 267 __ cmpl(rax, 0x80000000); // Is cpuid(0x80000001) supported? 268 __ jcc(Assembler::belowEqual, done); 269 __ cmpl(rax, 0x80000004); // Is cpuid(0x80000005) supported? 270 __ jcc(Assembler::belowEqual, ext_cpuid1); 271 __ cmpl(rax, 0x80000006); // Is cpuid(0x80000007) supported? 272 __ jccb(Assembler::belowEqual, ext_cpuid5); 273 __ cmpl(rax, 0x80000007); // Is cpuid(0x80000008) supported? 274 __ jccb(Assembler::belowEqual, ext_cpuid7); 275 __ cmpl(rax, 0x80000008); // Is cpuid(0x80000009 and above) supported? 276 __ jccb(Assembler::belowEqual, ext_cpuid8); 277 __ cmpl(rax, 0x8000001E); // Is cpuid(0x8000001E) supported? 278 __ jccb(Assembler::below, ext_cpuid8); 279 // 280 // Extended cpuid(0x8000001E) 281 // 282 __ movl(rax, 0x8000001E); 283 __ cpuid(); 284 __ lea(rsi, Address(rbp, in_bytes(VM_Version::ext_cpuid1E_offset()))); 285 __ movl(Address(rsi, 0), rax); 286 __ movl(Address(rsi, 4), rbx); 287 __ movl(Address(rsi, 8), rcx); 288 __ movl(Address(rsi,12), rdx); 289 290 // 291 // Extended cpuid(0x80000008) 292 // 293 __ bind(ext_cpuid8); 294 __ movl(rax, 0x80000008); 295 __ cpuid(); 296 __ lea(rsi, Address(rbp, in_bytes(VM_Version::ext_cpuid8_offset()))); 297 __ movl(Address(rsi, 0), rax); 298 __ movl(Address(rsi, 4), rbx); 299 __ movl(Address(rsi, 8), rcx); 300 __ movl(Address(rsi,12), rdx); 301 302 // 303 // Extended cpuid(0x80000007) 304 // 305 __ bind(ext_cpuid7); 306 __ movl(rax, 0x80000007); 307 __ cpuid(); 308 __ lea(rsi, Address(rbp, in_bytes(VM_Version::ext_cpuid7_offset()))); 309 __ movl(Address(rsi, 0), rax); 310 __ movl(Address(rsi, 4), rbx); 311 __ movl(Address(rsi, 8), rcx); 312 __ movl(Address(rsi,12), rdx); 313 314 // 315 // Extended cpuid(0x80000005) 316 // 317 __ bind(ext_cpuid5); 318 __ movl(rax, 0x80000005); 319 __ cpuid(); 320 __ lea(rsi, Address(rbp, in_bytes(VM_Version::ext_cpuid5_offset()))); 321 __ movl(Address(rsi, 0), rax); 322 __ movl(Address(rsi, 4), rbx); 323 __ movl(Address(rsi, 8), rcx); 324 __ movl(Address(rsi,12), rdx); 325 326 // 327 // Extended cpuid(0x80000001) 328 // 329 __ bind(ext_cpuid1); 330 __ movl(rax, 0x80000001); 331 __ cpuid(); 332 __ lea(rsi, Address(rbp, in_bytes(VM_Version::ext_cpuid1_offset()))); 333 __ movl(Address(rsi, 0), rax); 334 __ movl(Address(rsi, 4), rbx); 335 __ movl(Address(rsi, 8), rcx); 336 __ movl(Address(rsi,12), rdx); 337 338 // 339 // Check if OS has enabled XGETBV instruction to access XCR0 340 // (OSXSAVE feature flag) and CPU supports AVX 341 // 342 __ lea(rsi, Address(rbp, in_bytes(VM_Version::std_cpuid1_offset()))); 343 __ movl(rcx, 0x18000000); // cpuid1 bits osxsave | avx 344 __ andl(rcx, Address(rsi, 8)); // cpuid1 bits osxsave | avx 345 __ cmpl(rcx, 0x18000000); 346 __ jccb(Assembler::notEqual, done); // jump if AVX is not supported 347 348 __ movl(rax, 0x6); 349 __ andl(rax, Address(rbp, in_bytes(VM_Version::xem_xcr0_offset()))); // xcr0 bits sse | ymm 350 __ cmpl(rax, 0x6); 351 __ jccb(Assembler::equal, start_simd_check); // return if AVX is not supported 352 353 // we need to bridge farther than imm8, so we use this island as a thunk 354 __ bind(done); 355 __ jmp(wrapup); 356 357 __ bind(start_simd_check); 358 // 359 // Some OSs have a bug when upper 128/256bits of YMM/ZMM 360 // registers are not restored after a signal processing. 361 // Generate SEGV here (reference through NULL) 362 // and check upper YMM/ZMM bits after it. 363 // 364 intx saved_useavx = UseAVX; 365 intx saved_usesse = UseSSE; 366 // check _cpuid_info.sef_cpuid7_ebx.bits.avx512f 367 __ lea(rsi, Address(rbp, in_bytes(VM_Version::sef_cpuid7_offset()))); 368 __ movl(rax, 0x10000); 369 __ andl(rax, Address(rsi, 4)); // xcr0 bits sse | ymm 370 __ cmpl(rax, 0x10000); 371 __ jccb(Assembler::notEqual, legacy_setup); // jump if EVEX is not supported 372 // check _cpuid_info.xem_xcr0_eax.bits.opmask 373 // check _cpuid_info.xem_xcr0_eax.bits.zmm512 374 // check _cpuid_info.xem_xcr0_eax.bits.zmm32 375 __ movl(rax, 0xE0); 376 __ andl(rax, Address(rbp, in_bytes(VM_Version::xem_xcr0_offset()))); // xcr0 bits sse | ymm 377 __ cmpl(rax, 0xE0); 378 __ jccb(Assembler::notEqual, legacy_setup); // jump if EVEX is not supported 379 380 // If UseAVX is unitialized or is set by the user to include EVEX 381 if (use_evex) { 382 // EVEX setup: run in lowest evex mode 383 VM_Version::set_evex_cpuFeatures(); // Enable temporary to pass asserts 384 UseAVX = 3; 385 UseSSE = 2; 386 #ifdef _WINDOWS 387 // xmm5-xmm15 are not preserved by caller on windows 388 // https://msdn.microsoft.com/en-us/library/9z1stfyw.aspx 389 __ subptr(rsp, 64); 390 __ evmovdqul(Address(rsp, 0), xmm7, Assembler::AVX_512bit); 391 #ifdef _LP64 392 __ subptr(rsp, 64); 393 __ evmovdqul(Address(rsp, 0), xmm8, Assembler::AVX_512bit); 394 __ subptr(rsp, 64); 395 __ evmovdqul(Address(rsp, 0), xmm31, Assembler::AVX_512bit); 396 #endif // _LP64 397 #endif // _WINDOWS 398 399 // load value into all 64 bytes of zmm7 register 400 __ movl(rcx, VM_Version::ymm_test_value()); 401 __ movdl(xmm0, rcx); 402 __ movl(rcx, 0xffff); 403 __ kmovwl(k1, rcx); 404 __ evpbroadcastd(xmm0, xmm0, Assembler::AVX_512bit); 405 __ evmovdqul(xmm7, xmm0, Assembler::AVX_512bit); 406 #ifdef _LP64 407 __ evmovdqul(xmm8, xmm0, Assembler::AVX_512bit); 408 __ evmovdqul(xmm31, xmm0, Assembler::AVX_512bit); 409 #endif 410 VM_Version::clean_cpuFeatures(); 411 __ jmp(save_restore_except); 412 } 413 414 __ bind(legacy_setup); 415 // AVX setup 416 VM_Version::set_avx_cpuFeatures(); // Enable temporary to pass asserts 417 UseAVX = 1; 418 UseSSE = 2; 419 #ifdef _WINDOWS 420 __ subptr(rsp, 32); 421 __ vmovdqu(Address(rsp, 0), xmm7); 422 #ifdef _LP64 423 __ subptr(rsp, 32); 424 __ vmovdqu(Address(rsp, 0), xmm8); 425 __ subptr(rsp, 32); 426 __ vmovdqu(Address(rsp, 0), xmm15); 427 #endif // _LP64 428 #endif // _WINDOWS 429 430 // load value into all 32 bytes of ymm7 register 431 __ movl(rcx, VM_Version::ymm_test_value()); 432 433 __ movdl(xmm0, rcx); 434 __ pshufd(xmm0, xmm0, 0x00); 435 __ vinsertf128_high(xmm0, xmm0); 436 __ vmovdqu(xmm7, xmm0); 437 #ifdef _LP64 438 __ vmovdqu(xmm8, xmm0); 439 __ vmovdqu(xmm15, xmm0); 440 #endif 441 VM_Version::clean_cpuFeatures(); 442 443 __ bind(save_restore_except); 444 __ xorl(rsi, rsi); 445 VM_Version::set_cpuinfo_segv_addr(__ pc()); 446 // Generate SEGV 447 __ movl(rax, Address(rsi, 0)); 448 449 VM_Version::set_cpuinfo_cont_addr(__ pc()); 450 // Returns here after signal. Save xmm0 to check it later. 451 452 // check _cpuid_info.sef_cpuid7_ebx.bits.avx512f 453 __ lea(rsi, Address(rbp, in_bytes(VM_Version::sef_cpuid7_offset()))); 454 __ movl(rax, 0x10000); 455 __ andl(rax, Address(rsi, 4)); 456 __ cmpl(rax, 0x10000); 457 __ jcc(Assembler::notEqual, legacy_save_restore); 458 // check _cpuid_info.xem_xcr0_eax.bits.opmask 459 // check _cpuid_info.xem_xcr0_eax.bits.zmm512 460 // check _cpuid_info.xem_xcr0_eax.bits.zmm32 461 __ movl(rax, 0xE0); 462 __ andl(rax, Address(rbp, in_bytes(VM_Version::xem_xcr0_offset()))); // xcr0 bits sse | ymm 463 __ cmpl(rax, 0xE0); 464 __ jcc(Assembler::notEqual, legacy_save_restore); 465 466 // If UseAVX is unitialized or is set by the user to include EVEX 467 if (use_evex) { 468 // EVEX check: run in lowest evex mode 469 VM_Version::set_evex_cpuFeatures(); // Enable temporary to pass asserts 470 UseAVX = 3; 471 UseSSE = 2; 472 __ lea(rsi, Address(rbp, in_bytes(VM_Version::zmm_save_offset()))); 473 __ evmovdqul(Address(rsi, 0), xmm0, Assembler::AVX_512bit); 474 __ evmovdqul(Address(rsi, 64), xmm7, Assembler::AVX_512bit); 475 #ifdef _LP64 476 __ evmovdqul(Address(rsi, 128), xmm8, Assembler::AVX_512bit); 477 __ evmovdqul(Address(rsi, 192), xmm31, Assembler::AVX_512bit); 478 #endif 479 480 #ifdef _WINDOWS 481 #ifdef _LP64 482 __ evmovdqul(xmm31, Address(rsp, 0), Assembler::AVX_512bit); 483 __ addptr(rsp, 64); 484 __ evmovdqul(xmm8, Address(rsp, 0), Assembler::AVX_512bit); 485 __ addptr(rsp, 64); 486 #endif // _LP64 487 __ evmovdqul(xmm7, Address(rsp, 0), Assembler::AVX_512bit); 488 __ addptr(rsp, 64); 489 #endif // _WINDOWS 490 generate_vzeroupper(wrapup); 491 VM_Version::clean_cpuFeatures(); 492 UseAVX = saved_useavx; 493 UseSSE = saved_usesse; 494 __ jmp(wrapup); 495 } 496 497 __ bind(legacy_save_restore); 498 // AVX check 499 VM_Version::set_avx_cpuFeatures(); // Enable temporary to pass asserts 500 UseAVX = 1; 501 UseSSE = 2; 502 __ lea(rsi, Address(rbp, in_bytes(VM_Version::ymm_save_offset()))); 503 __ vmovdqu(Address(rsi, 0), xmm0); 504 __ vmovdqu(Address(rsi, 32), xmm7); 505 #ifdef _LP64 506 __ vmovdqu(Address(rsi, 64), xmm8); 507 __ vmovdqu(Address(rsi, 96), xmm15); 508 #endif 509 510 #ifdef _WINDOWS 511 #ifdef _LP64 512 __ vmovdqu(xmm15, Address(rsp, 0)); 513 __ addptr(rsp, 32); 514 __ vmovdqu(xmm8, Address(rsp, 0)); 515 __ addptr(rsp, 32); 516 #endif // _LP64 517 __ vmovdqu(xmm7, Address(rsp, 0)); 518 __ addptr(rsp, 32); 519 #endif // _WINDOWS 520 generate_vzeroupper(wrapup); 521 VM_Version::clean_cpuFeatures(); 522 UseAVX = saved_useavx; 523 UseSSE = saved_usesse; 524 525 __ bind(wrapup); 526 __ popf(); 527 __ pop(rsi); 528 __ pop(rbx); 529 __ pop(rbp); 530 __ ret(0); 531 532 # undef __ 533 534 return start; 535 }; 536 void generate_vzeroupper(Label& L_wrapup) { 537 # define __ _masm-> 538 __ lea(rsi, Address(rbp, in_bytes(VM_Version::std_cpuid0_offset()))); 539 __ cmpl(Address(rsi, 4), 0x756e6547); // 'uneG' 540 __ jcc(Assembler::notEqual, L_wrapup); 541 __ movl(rcx, 0x0FFF0FF0); 542 __ lea(rsi, Address(rbp, in_bytes(VM_Version::std_cpuid1_offset()))); 543 __ andl(rcx, Address(rsi, 0)); 544 __ cmpl(rcx, 0x00050670); // If it is Xeon Phi 3200/5200/7200 545 __ jcc(Assembler::equal, L_wrapup); 546 __ cmpl(rcx, 0x00080650); // If it is Future Xeon Phi 547 __ jcc(Assembler::equal, L_wrapup); 548 __ vzeroupper(); 549 # undef __ 550 } 551 }; 552 553 void VM_Version::get_processor_features() { 554 555 _cpu = 4; // 486 by default 556 _model = 0; 557 _stepping = 0; 558 _features = 0; 559 _logical_processors_per_package = 1; 560 // i486 internal cache is both I&D and has a 16-byte line size 561 _L1_data_cache_line_size = 16; 562 563 // Get raw processor info 564 565 get_cpu_info_stub(&_cpuid_info); 566 567 assert_is_initialized(); 568 _cpu = extended_cpu_family(); 569 _model = extended_cpu_model(); 570 _stepping = cpu_stepping(); 571 572 if (cpu_family() > 4) { // it supports CPUID 573 _features = feature_flags(); 574 // Logical processors are only available on P4s and above, 575 // and only if hyperthreading is available. 576 _logical_processors_per_package = logical_processor_count(); 577 _L1_data_cache_line_size = L1_line_size(); 578 } 579 580 _supports_cx8 = supports_cmpxchg8(); 581 // xchg and xadd instructions 582 _supports_atomic_getset4 = true; 583 _supports_atomic_getadd4 = true; 584 LP64_ONLY(_supports_atomic_getset8 = true); 585 LP64_ONLY(_supports_atomic_getadd8 = true); 586 587 #ifdef _LP64 588 // OS should support SSE for x64 and hardware should support at least SSE2. 589 if (!VM_Version::supports_sse2()) { 590 vm_exit_during_initialization("Unknown x64 processor: SSE2 not supported"); 591 } 592 // in 64 bit the use of SSE2 is the minimum 593 if (UseSSE < 2) UseSSE = 2; 594 #endif 595 596 #ifdef AMD64 597 // flush_icache_stub have to be generated first. 598 // That is why Icache line size is hard coded in ICache class, 599 // see icache_x86.hpp. It is also the reason why we can't use 600 // clflush instruction in 32-bit VM since it could be running 601 // on CPU which does not support it. 602 // 603 // The only thing we can do is to verify that flushed 604 // ICache::line_size has correct value. 605 guarantee(_cpuid_info.std_cpuid1_edx.bits.clflush != 0, "clflush is not supported"); 606 // clflush_size is size in quadwords (8 bytes). 607 guarantee(_cpuid_info.std_cpuid1_ebx.bits.clflush_size == 8, "such clflush size is not supported"); 608 #endif 609 610 // If the OS doesn't support SSE, we can't use this feature even if the HW does 611 if (!os::supports_sse()) 612 _features &= ~(CPU_SSE|CPU_SSE2|CPU_SSE3|CPU_SSSE3|CPU_SSE4A|CPU_SSE4_1|CPU_SSE4_2); 613 614 if (UseSSE < 4) { 615 _features &= ~CPU_SSE4_1; 616 _features &= ~CPU_SSE4_2; 617 } 618 619 if (UseSSE < 3) { 620 _features &= ~CPU_SSE3; 621 _features &= ~CPU_SSSE3; 622 _features &= ~CPU_SSE4A; 623 } 624 625 if (UseSSE < 2) 626 _features &= ~CPU_SSE2; 627 628 if (UseSSE < 1) 629 _features &= ~CPU_SSE; 630 631 // first try initial setting and detect what we can support 632 int use_avx_limit = 0; 633 if (UseAVX > 0) { 634 if (UseAVX > 2 && supports_evex()) { 635 use_avx_limit = 3; 636 } else if (UseAVX > 1 && supports_avx2()) { 637 use_avx_limit = 2; 638 } else if (UseAVX > 0 && supports_avx()) { 639 use_avx_limit = 1; 640 } else { 641 use_avx_limit = 0; 642 } 643 } 644 if (FLAG_IS_DEFAULT(UseAVX)) { 645 FLAG_SET_DEFAULT(UseAVX, use_avx_limit); 646 } else if (UseAVX > use_avx_limit) { 647 warning("UseAVX=%d is not supported on this CPU, setting it to UseAVX=%d", (int) UseAVX, use_avx_limit); 648 FLAG_SET_DEFAULT(UseAVX, use_avx_limit); 649 } else if (UseAVX < 0) { 650 warning("UseAVX=%d is not valid, setting it to UseAVX=0", (int) UseAVX); 651 FLAG_SET_DEFAULT(UseAVX, 0); 652 } 653 654 if (UseAVX < 3) { 655 _features &= ~CPU_AVX512F; 656 _features &= ~CPU_AVX512DQ; 657 _features &= ~CPU_AVX512CD; 658 _features &= ~CPU_AVX512BW; 659 _features &= ~CPU_AVX512VL; 660 } 661 662 if (UseAVX < 2) 663 _features &= ~CPU_AVX2; 664 665 if (UseAVX < 1) { 666 _features &= ~CPU_AVX; 667 _features &= ~CPU_VZEROUPPER; 668 } 669 670 if (logical_processors_per_package() == 1) { 671 // HT processor could be installed on a system which doesn't support HT. 672 _features &= ~CPU_HT; 673 } 674 675 if( is_intel() ) { // Intel cpus specific settings 676 if (is_knights_family()) { 677 _features &= ~CPU_VZEROUPPER; 678 } 679 } 680 681 char buf[256]; 682 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", 683 cores_per_cpu(), threads_per_core(), 684 cpu_family(), _model, _stepping, 685 (supports_cmov() ? ", cmov" : ""), 686 (supports_cmpxchg8() ? ", cx8" : ""), 687 (supports_fxsr() ? ", fxsr" : ""), 688 (supports_mmx() ? ", mmx" : ""), 689 (supports_sse() ? ", sse" : ""), 690 (supports_sse2() ? ", sse2" : ""), 691 (supports_sse3() ? ", sse3" : ""), 692 (supports_ssse3()? ", ssse3": ""), 693 (supports_sse4_1() ? ", sse4.1" : ""), 694 (supports_sse4_2() ? ", sse4.2" : ""), 695 (supports_popcnt() ? ", popcnt" : ""), 696 (supports_avx() ? ", avx" : ""), 697 (supports_avx2() ? ", avx2" : ""), 698 (supports_aes() ? ", aes" : ""), 699 (supports_clmul() ? ", clmul" : ""), 700 (supports_erms() ? ", erms" : ""), 701 (supports_rtm() ? ", rtm" : ""), 702 (supports_mmx_ext() ? ", mmxext" : ""), 703 (supports_3dnow_prefetch() ? ", 3dnowpref" : ""), 704 (supports_lzcnt() ? ", lzcnt": ""), 705 (supports_sse4a() ? ", sse4a": ""), 706 (supports_ht() ? ", ht": ""), 707 (supports_tsc() ? ", tsc": ""), 708 (supports_tscinv_bit() ? ", tscinvbit": ""), 709 (supports_tscinv() ? ", tscinv": ""), 710 (supports_bmi1() ? ", bmi1" : ""), 711 (supports_bmi2() ? ", bmi2" : ""), 712 (supports_adx() ? ", adx" : ""), 713 (supports_evex() ? ", evex" : ""), 714 (supports_sha() ? ", sha" : ""), 715 (supports_fma() ? ", fma" : "")); 716 _features_string = os::strdup(buf); 717 718 // UseSSE is set to the smaller of what hardware supports and what 719 // the command line requires. I.e., you cannot set UseSSE to 2 on 720 // older Pentiums which do not support it. 721 int use_sse_limit = 0; 722 if (UseSSE > 0) { 723 if (UseSSE > 3 && supports_sse4_1()) { 724 use_sse_limit = 4; 725 } else if (UseSSE > 2 && supports_sse3()) { 726 use_sse_limit = 3; 727 } else if (UseSSE > 1 && supports_sse2()) { 728 use_sse_limit = 2; 729 } else if (UseSSE > 0 && supports_sse()) { 730 use_sse_limit = 1; 731 } else { 732 use_sse_limit = 0; 733 } 734 } 735 if (FLAG_IS_DEFAULT(UseSSE)) { 736 FLAG_SET_DEFAULT(UseSSE, use_sse_limit); 737 } else if (UseSSE > use_sse_limit) { 738 warning("UseSSE=%d is not supported on this CPU, setting it to UseSSE=%d", (int) UseSSE, use_sse_limit); 739 FLAG_SET_DEFAULT(UseSSE, use_sse_limit); 740 } else if (UseSSE < 0) { 741 warning("UseSSE=%d is not valid, setting it to UseSSE=0", (int) UseSSE); 742 FLAG_SET_DEFAULT(UseSSE, 0); 743 } 744 745 // Use AES instructions if available. 746 if (supports_aes()) { 747 if (FLAG_IS_DEFAULT(UseAES)) { 748 FLAG_SET_DEFAULT(UseAES, true); 749 } 750 if (!UseAES) { 751 if (UseAESIntrinsics && !FLAG_IS_DEFAULT(UseAESIntrinsics)) { 752 warning("AES intrinsics require UseAES flag to be enabled. Intrinsics will be disabled."); 753 } 754 FLAG_SET_DEFAULT(UseAESIntrinsics, false); 755 } else { 756 if (UseSSE > 2) { 757 if (FLAG_IS_DEFAULT(UseAESIntrinsics)) { 758 FLAG_SET_DEFAULT(UseAESIntrinsics, true); 759 } 760 } else { 761 // The AES intrinsic stubs require AES instruction support (of course) 762 // but also require sse3 mode or higher for instructions it use. 763 if (UseAESIntrinsics && !FLAG_IS_DEFAULT(UseAESIntrinsics)) { 764 warning("X86 AES intrinsics require SSE3 instructions or higher. Intrinsics will be disabled."); 765 } 766 FLAG_SET_DEFAULT(UseAESIntrinsics, false); 767 } 768 769 // --AES-CTR begins-- 770 if (!UseAESIntrinsics) { 771 if (UseAESCTRIntrinsics && !FLAG_IS_DEFAULT(UseAESCTRIntrinsics)) { 772 warning("AES-CTR intrinsics require UseAESIntrinsics flag to be enabled. Intrinsics will be disabled."); 773 FLAG_SET_DEFAULT(UseAESCTRIntrinsics, false); 774 } 775 } else { 776 if(supports_sse4_1()) { 777 if (FLAG_IS_DEFAULT(UseAESCTRIntrinsics)) { 778 FLAG_SET_DEFAULT(UseAESCTRIntrinsics, true); 779 } 780 } else { 781 // The AES-CTR intrinsic stubs require AES instruction support (of course) 782 // but also require sse4.1 mode or higher for instructions it use. 783 if (UseAESCTRIntrinsics && !FLAG_IS_DEFAULT(UseAESCTRIntrinsics)) { 784 warning("X86 AES-CTR intrinsics require SSE4.1 instructions or higher. Intrinsics will be disabled."); 785 } 786 FLAG_SET_DEFAULT(UseAESCTRIntrinsics, false); 787 } 788 } 789 // --AES-CTR ends-- 790 } 791 } else if (UseAES || UseAESIntrinsics || UseAESCTRIntrinsics) { 792 if (UseAES && !FLAG_IS_DEFAULT(UseAES)) { 793 warning("AES instructions are not available on this CPU"); 794 FLAG_SET_DEFAULT(UseAES, false); 795 } 796 if (UseAESIntrinsics && !FLAG_IS_DEFAULT(UseAESIntrinsics)) { 797 warning("AES intrinsics are not available on this CPU"); 798 FLAG_SET_DEFAULT(UseAESIntrinsics, false); 799 } 800 if (UseAESCTRIntrinsics && !FLAG_IS_DEFAULT(UseAESCTRIntrinsics)) { 801 warning("AES-CTR intrinsics are not available on this CPU"); 802 FLAG_SET_DEFAULT(UseAESCTRIntrinsics, false); 803 } 804 } 805 806 // Use CLMUL instructions if available. 807 if (supports_clmul()) { 808 if (FLAG_IS_DEFAULT(UseCLMUL)) { 809 UseCLMUL = true; 810 } 811 } else if (UseCLMUL) { 812 if (!FLAG_IS_DEFAULT(UseCLMUL)) 813 warning("CLMUL instructions not available on this CPU (AVX may also be required)"); 814 FLAG_SET_DEFAULT(UseCLMUL, false); 815 } 816 817 if (UseCLMUL && (UseSSE > 2)) { 818 if (FLAG_IS_DEFAULT(UseCRC32Intrinsics)) { 819 UseCRC32Intrinsics = true; 820 } 821 } else if (UseCRC32Intrinsics) { 822 if (!FLAG_IS_DEFAULT(UseCRC32Intrinsics)) 823 warning("CRC32 Intrinsics requires CLMUL instructions (not available on this CPU)"); 824 FLAG_SET_DEFAULT(UseCRC32Intrinsics, false); 825 } 826 827 if (supports_sse4_2() && supports_clmul()) { 828 if (FLAG_IS_DEFAULT(UseCRC32CIntrinsics)) { 829 UseCRC32CIntrinsics = true; 830 } 831 } else if (UseCRC32CIntrinsics) { 832 if (!FLAG_IS_DEFAULT(UseCRC32CIntrinsics)) { 833 warning("CRC32C intrinsics are not available on this CPU"); 834 } 835 FLAG_SET_DEFAULT(UseCRC32CIntrinsics, false); 836 } 837 838 // GHASH/GCM intrinsics 839 if (UseCLMUL && (UseSSE > 2)) { 840 if (FLAG_IS_DEFAULT(UseGHASHIntrinsics)) { 841 UseGHASHIntrinsics = true; 842 } 843 } else if (UseGHASHIntrinsics) { 844 if (!FLAG_IS_DEFAULT(UseGHASHIntrinsics)) 845 warning("GHASH intrinsic requires CLMUL and SSE2 instructions on this CPU"); 846 FLAG_SET_DEFAULT(UseGHASHIntrinsics, false); 847 } 848 849 if (supports_fma() && UseSSE >= 2) { // Check UseSSE since FMA code uses SSE instructions 850 if (FLAG_IS_DEFAULT(UseFMA)) { 851 UseFMA = true; 852 } 853 } else if (UseFMA) { 854 warning("FMA instructions are not available on this CPU"); 855 FLAG_SET_DEFAULT(UseFMA, false); 856 } 857 858 if (supports_sha() LP64_ONLY(|| supports_avx2() && supports_bmi2())) { 859 if (FLAG_IS_DEFAULT(UseSHA)) { 860 UseSHA = true; 861 } 862 } else if (UseSHA) { 863 warning("SHA instructions are not available on this CPU"); 864 FLAG_SET_DEFAULT(UseSHA, false); 865 } 866 867 if (supports_sha() && UseSHA) { 868 if (FLAG_IS_DEFAULT(UseSHA1Intrinsics)) { 869 FLAG_SET_DEFAULT(UseSHA1Intrinsics, true); 870 } 871 } else if (UseSHA1Intrinsics) { 872 warning("Intrinsics for SHA-1 crypto hash functions not available on this CPU."); 873 FLAG_SET_DEFAULT(UseSHA1Intrinsics, false); 874 } 875 876 if (UseSHA) { 877 if (FLAG_IS_DEFAULT(UseSHA256Intrinsics)) { 878 FLAG_SET_DEFAULT(UseSHA256Intrinsics, true); 879 } 880 } else if (UseSHA256Intrinsics) { 881 warning("Intrinsics for SHA-224 and SHA-256 crypto hash functions not available on this CPU."); 882 FLAG_SET_DEFAULT(UseSHA256Intrinsics, false); 883 } 884 885 if (UseSHA LP64_ONLY(&& supports_avx2() && supports_bmi2())) { 886 if (FLAG_IS_DEFAULT(UseSHA512Intrinsics)) { 887 FLAG_SET_DEFAULT(UseSHA512Intrinsics, true); 888 } 889 } else if (UseSHA512Intrinsics) { 890 warning("Intrinsics for SHA-384 and SHA-512 crypto hash functions not available on this CPU."); 891 FLAG_SET_DEFAULT(UseSHA512Intrinsics, false); 892 } 893 894 if (!(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics)) { 895 FLAG_SET_DEFAULT(UseSHA, false); 896 } 897 898 if (UseAdler32Intrinsics) { 899 warning("Adler32Intrinsics not available on this CPU."); 900 FLAG_SET_DEFAULT(UseAdler32Intrinsics, false); 901 } 902 903 if (!supports_rtm() && UseRTMLocking) { 904 // Can't continue because UseRTMLocking affects UseBiasedLocking flag 905 // setting during arguments processing. See use_biased_locking(). 906 // VM_Version_init() is executed after UseBiasedLocking is used 907 // in Thread::allocate(). 908 vm_exit_during_initialization("RTM instructions are not available on this CPU"); 909 } 910 911 #if INCLUDE_RTM_OPT 912 if (UseRTMLocking) { 913 if (is_client_compilation_mode_vm()) { 914 // Only C2 does RTM locking optimization. 915 // Can't continue because UseRTMLocking affects UseBiasedLocking flag 916 // setting during arguments processing. See use_biased_locking(). 917 vm_exit_during_initialization("RTM locking optimization is not supported in emulated client VM"); 918 } 919 if (is_intel_family_core()) { 920 if ((_model == CPU_MODEL_HASWELL_E3) || 921 (_model == CPU_MODEL_HASWELL_E7 && _stepping < 3) || 922 (_model == CPU_MODEL_BROADWELL && _stepping < 4)) { 923 // currently a collision between SKL and HSW_E3 924 if (!UnlockExperimentalVMOptions && UseAVX < 3) { 925 vm_exit_during_initialization("UseRTMLocking is only available as experimental option on this " 926 "platform. It must be enabled via -XX:+UnlockExperimentalVMOptions flag."); 927 } else { 928 warning("UseRTMLocking is only available as experimental option on this platform."); 929 } 930 } 931 } 932 if (!FLAG_IS_CMDLINE(UseRTMLocking)) { 933 // RTM locking should be used only for applications with 934 // high lock contention. For now we do not use it by default. 935 vm_exit_during_initialization("UseRTMLocking flag should be only set on command line"); 936 } 937 } else { // !UseRTMLocking 938 if (UseRTMForStackLocks) { 939 if (!FLAG_IS_DEFAULT(UseRTMForStackLocks)) { 940 warning("UseRTMForStackLocks flag should be off when UseRTMLocking flag is off"); 941 } 942 FLAG_SET_DEFAULT(UseRTMForStackLocks, false); 943 } 944 if (UseRTMDeopt) { 945 FLAG_SET_DEFAULT(UseRTMDeopt, false); 946 } 947 if (PrintPreciseRTMLockingStatistics) { 948 FLAG_SET_DEFAULT(PrintPreciseRTMLockingStatistics, false); 949 } 950 } 951 #else 952 if (UseRTMLocking) { 953 // Only C2 does RTM locking optimization. 954 // Can't continue because UseRTMLocking affects UseBiasedLocking flag 955 // setting during arguments processing. See use_biased_locking(). 956 vm_exit_during_initialization("RTM locking optimization is not supported in this VM"); 957 } 958 #endif 959 960 #ifdef COMPILER2 961 if (UseFPUForSpilling) { 962 if (UseSSE < 2) { 963 // Only supported with SSE2+ 964 FLAG_SET_DEFAULT(UseFPUForSpilling, false); 965 } 966 } 967 #endif 968 #if COMPILER2_OR_JVMCI 969 if (MaxVectorSize > 0) { 970 if (!is_power_of_2(MaxVectorSize)) { 971 warning("MaxVectorSize must be a power of 2"); 972 FLAG_SET_DEFAULT(MaxVectorSize, 64); 973 } 974 if (UseSSE < 2) { 975 // Vectors (in XMM) are only supported with SSE2+ 976 if (MaxVectorSize > 0) { 977 if (!FLAG_IS_DEFAULT(MaxVectorSize)) 978 warning("MaxVectorSize must be 0"); 979 FLAG_SET_DEFAULT(MaxVectorSize, 0); 980 } 981 } 982 else if (UseAVX == 0 || !os_supports_avx_vectors()) { 983 // 32 bytes vectors (in YMM) are only supported with AVX+ 984 if (MaxVectorSize > 16) { 985 if (!FLAG_IS_DEFAULT(MaxVectorSize)) 986 warning("MaxVectorSize must be <= 16"); 987 FLAG_SET_DEFAULT(MaxVectorSize, 16); 988 } 989 } 990 else if (UseAVX == 1 || UseAVX == 2) { 991 // 64 bytes vectors (in ZMM) are only supported with AVX 3 992 if (MaxVectorSize > 32) { 993 if (!FLAG_IS_DEFAULT(MaxVectorSize)) 994 warning("MaxVectorSize must be <= 32"); 995 FLAG_SET_DEFAULT(MaxVectorSize, 32); 996 } 997 } 998 else if (UseAVX > 2 ) { 999 if (MaxVectorSize > 64) { 1000 if (!FLAG_IS_DEFAULT(MaxVectorSize)) 1001 warning("MaxVectorSize must be <= 64"); 1002 FLAG_SET_DEFAULT(MaxVectorSize, 64); 1003 } 1004 } 1005 #if defined(COMPILER2) && defined(ASSERT) 1006 if (supports_avx() && PrintMiscellaneous && Verbose && TraceNewVectors) { 1007 tty->print_cr("State of YMM registers after signal handle:"); 1008 int nreg = 2 LP64_ONLY(+2); 1009 const char* ymm_name[4] = {"0", "7", "8", "15"}; 1010 for (int i = 0; i < nreg; i++) { 1011 tty->print("YMM%s:", ymm_name[i]); 1012 for (int j = 7; j >=0; j--) { 1013 tty->print(" %x", _cpuid_info.ymm_save[i*8 + j]); 1014 } 1015 tty->cr(); 1016 } 1017 } 1018 #endif // COMPILER2 && ASSERT 1019 } 1020 #endif // COMPILER2_OR_JVMCI 1021 1022 #ifdef COMPILER2 1023 #ifdef _LP64 1024 if (FLAG_IS_DEFAULT(UseMultiplyToLenIntrinsic)) { 1025 UseMultiplyToLenIntrinsic = true; 1026 } 1027 if (FLAG_IS_DEFAULT(UseSquareToLenIntrinsic)) { 1028 UseSquareToLenIntrinsic = true; 1029 } 1030 if (FLAG_IS_DEFAULT(UseMulAddIntrinsic)) { 1031 UseMulAddIntrinsic = true; 1032 } 1033 if (FLAG_IS_DEFAULT(UseMontgomeryMultiplyIntrinsic)) { 1034 UseMontgomeryMultiplyIntrinsic = true; 1035 } 1036 if (FLAG_IS_DEFAULT(UseMontgomerySquareIntrinsic)) { 1037 UseMontgomerySquareIntrinsic = true; 1038 } 1039 #else 1040 if (UseMultiplyToLenIntrinsic) { 1041 if (!FLAG_IS_DEFAULT(UseMultiplyToLenIntrinsic)) { 1042 warning("multiplyToLen intrinsic is not available in 32-bit VM"); 1043 } 1044 FLAG_SET_DEFAULT(UseMultiplyToLenIntrinsic, false); 1045 } 1046 if (UseMontgomeryMultiplyIntrinsic) { 1047 if (!FLAG_IS_DEFAULT(UseMontgomeryMultiplyIntrinsic)) { 1048 warning("montgomeryMultiply intrinsic is not available in 32-bit VM"); 1049 } 1050 FLAG_SET_DEFAULT(UseMontgomeryMultiplyIntrinsic, false); 1051 } 1052 if (UseMontgomerySquareIntrinsic) { 1053 if (!FLAG_IS_DEFAULT(UseMontgomerySquareIntrinsic)) { 1054 warning("montgomerySquare intrinsic is not available in 32-bit VM"); 1055 } 1056 FLAG_SET_DEFAULT(UseMontgomerySquareIntrinsic, false); 1057 } 1058 if (UseSquareToLenIntrinsic) { 1059 if (!FLAG_IS_DEFAULT(UseSquareToLenIntrinsic)) { 1060 warning("squareToLen intrinsic is not available in 32-bit VM"); 1061 } 1062 FLAG_SET_DEFAULT(UseSquareToLenIntrinsic, false); 1063 } 1064 if (UseMulAddIntrinsic) { 1065 if (!FLAG_IS_DEFAULT(UseMulAddIntrinsic)) { 1066 warning("mulAdd intrinsic is not available in 32-bit VM"); 1067 } 1068 FLAG_SET_DEFAULT(UseMulAddIntrinsic, false); 1069 } 1070 #endif 1071 #endif // COMPILER2 1072 1073 // On new cpus instructions which update whole XMM register should be used 1074 // to prevent partial register stall due to dependencies on high half. 1075 // 1076 // UseXmmLoadAndClearUpper == true --> movsd(xmm, mem) 1077 // UseXmmLoadAndClearUpper == false --> movlpd(xmm, mem) 1078 // UseXmmRegToRegMoveAll == true --> movaps(xmm, xmm), movapd(xmm, xmm). 1079 // UseXmmRegToRegMoveAll == false --> movss(xmm, xmm), movsd(xmm, xmm). 1080 1081 if( is_amd() ) { // AMD cpus specific settings 1082 if( supports_sse2() && FLAG_IS_DEFAULT(UseAddressNop) ) { 1083 // Use it on new AMD cpus starting from Opteron. 1084 UseAddressNop = true; 1085 } 1086 if( supports_sse2() && FLAG_IS_DEFAULT(UseNewLongLShift) ) { 1087 // Use it on new AMD cpus starting from Opteron. 1088 UseNewLongLShift = true; 1089 } 1090 if( FLAG_IS_DEFAULT(UseXmmLoadAndClearUpper) ) { 1091 if (supports_sse4a()) { 1092 UseXmmLoadAndClearUpper = true; // use movsd only on '10h' Opteron 1093 } else { 1094 UseXmmLoadAndClearUpper = false; 1095 } 1096 } 1097 if( FLAG_IS_DEFAULT(UseXmmRegToRegMoveAll) ) { 1098 if( supports_sse4a() ) { 1099 UseXmmRegToRegMoveAll = true; // use movaps, movapd only on '10h' 1100 } else { 1101 UseXmmRegToRegMoveAll = false; 1102 } 1103 } 1104 if( FLAG_IS_DEFAULT(UseXmmI2F) ) { 1105 if( supports_sse4a() ) { 1106 UseXmmI2F = true; 1107 } else { 1108 UseXmmI2F = false; 1109 } 1110 } 1111 if( FLAG_IS_DEFAULT(UseXmmI2D) ) { 1112 if( supports_sse4a() ) { 1113 UseXmmI2D = true; 1114 } else { 1115 UseXmmI2D = false; 1116 } 1117 } 1118 if (supports_sse4_2()) { 1119 if (FLAG_IS_DEFAULT(UseSSE42Intrinsics)) { 1120 FLAG_SET_DEFAULT(UseSSE42Intrinsics, true); 1121 } 1122 } else { 1123 if (UseSSE42Intrinsics && !FLAG_IS_DEFAULT(UseAESIntrinsics)) { 1124 warning("SSE4.2 intrinsics require SSE4.2 instructions or higher. Intrinsics will be disabled."); 1125 } 1126 FLAG_SET_DEFAULT(UseSSE42Intrinsics, false); 1127 } 1128 1129 // some defaults for AMD family 15h 1130 if ( cpu_family() == 0x15 ) { 1131 // On family 15h processors default is no sw prefetch 1132 if (FLAG_IS_DEFAULT(AllocatePrefetchStyle)) { 1133 FLAG_SET_DEFAULT(AllocatePrefetchStyle, 0); 1134 } 1135 // Also, if some other prefetch style is specified, default instruction type is PREFETCHW 1136 if (FLAG_IS_DEFAULT(AllocatePrefetchInstr)) { 1137 FLAG_SET_DEFAULT(AllocatePrefetchInstr, 3); 1138 } 1139 // On family 15h processors use XMM and UnalignedLoadStores for Array Copy 1140 if (supports_sse2() && FLAG_IS_DEFAULT(UseXMMForArrayCopy)) { 1141 FLAG_SET_DEFAULT(UseXMMForArrayCopy, true); 1142 } 1143 if (supports_sse2() && FLAG_IS_DEFAULT(UseUnalignedLoadStores)) { 1144 FLAG_SET_DEFAULT(UseUnalignedLoadStores, true); 1145 } 1146 } 1147 1148 #ifdef COMPILER2 1149 if (cpu_family() < 0x17 && MaxVectorSize > 16) { 1150 // Limit vectors size to 16 bytes on AMD cpus < 17h. 1151 FLAG_SET_DEFAULT(MaxVectorSize, 16); 1152 } 1153 #endif // COMPILER2 1154 1155 // Some defaults for AMD family 17h 1156 if ( cpu_family() == 0x17 ) { 1157 // On family 17h processors use XMM and UnalignedLoadStores for Array Copy 1158 if (supports_sse2() && FLAG_IS_DEFAULT(UseXMMForArrayCopy)) { 1159 FLAG_SET_DEFAULT(UseXMMForArrayCopy, true); 1160 } 1161 if (supports_sse2() && FLAG_IS_DEFAULT(UseUnalignedLoadStores)) { 1162 FLAG_SET_DEFAULT(UseUnalignedLoadStores, true); 1163 } 1164 #ifdef COMPILER2 1165 if (supports_sse4_2() && FLAG_IS_DEFAULT(UseFPUForSpilling)) { 1166 FLAG_SET_DEFAULT(UseFPUForSpilling, true); 1167 } 1168 #endif 1169 } 1170 } 1171 1172 if( is_intel() ) { // Intel cpus specific settings 1173 if( FLAG_IS_DEFAULT(UseStoreImmI16) ) { 1174 UseStoreImmI16 = false; // don't use it on Intel cpus 1175 } 1176 if( cpu_family() == 6 || cpu_family() == 15 ) { 1177 if( FLAG_IS_DEFAULT(UseAddressNop) ) { 1178 // Use it on all Intel cpus starting from PentiumPro 1179 UseAddressNop = true; 1180 } 1181 } 1182 if( FLAG_IS_DEFAULT(UseXmmLoadAndClearUpper) ) { 1183 UseXmmLoadAndClearUpper = true; // use movsd on all Intel cpus 1184 } 1185 if( FLAG_IS_DEFAULT(UseXmmRegToRegMoveAll) ) { 1186 if( supports_sse3() ) { 1187 UseXmmRegToRegMoveAll = true; // use movaps, movapd on new Intel cpus 1188 } else { 1189 UseXmmRegToRegMoveAll = false; 1190 } 1191 } 1192 if( cpu_family() == 6 && supports_sse3() ) { // New Intel cpus 1193 #ifdef COMPILER2 1194 if( FLAG_IS_DEFAULT(MaxLoopPad) ) { 1195 // For new Intel cpus do the next optimization: 1196 // don't align the beginning of a loop if there are enough instructions 1197 // left (NumberOfLoopInstrToAlign defined in c2_globals.hpp) 1198 // in current fetch line (OptoLoopAlignment) or the padding 1199 // is big (> MaxLoopPad). 1200 // Set MaxLoopPad to 11 for new Intel cpus to reduce number of 1201 // generated NOP instructions. 11 is the largest size of one 1202 // address NOP instruction '0F 1F' (see Assembler::nop(i)). 1203 MaxLoopPad = 11; 1204 } 1205 #endif // COMPILER2 1206 if (FLAG_IS_DEFAULT(UseXMMForArrayCopy)) { 1207 UseXMMForArrayCopy = true; // use SSE2 movq on new Intel cpus 1208 } 1209 if (supports_sse4_2() && supports_ht()) { // Newest Intel cpus 1210 if (FLAG_IS_DEFAULT(UseUnalignedLoadStores)) { 1211 UseUnalignedLoadStores = true; // use movdqu on newest Intel cpus 1212 } 1213 } 1214 if (supports_sse4_2()) { 1215 if (FLAG_IS_DEFAULT(UseSSE42Intrinsics)) { 1216 FLAG_SET_DEFAULT(UseSSE42Intrinsics, true); 1217 } 1218 } else { 1219 if (UseSSE42Intrinsics && !FLAG_IS_DEFAULT(UseAESIntrinsics)) { 1220 warning("SSE4.2 intrinsics require SSE4.2 instructions or higher. Intrinsics will be disabled."); 1221 } 1222 FLAG_SET_DEFAULT(UseSSE42Intrinsics, false); 1223 } 1224 } 1225 if (is_atom_family() || is_knights_family()) { 1226 #ifdef COMPILER2 1227 if (FLAG_IS_DEFAULT(OptoScheduling)) { 1228 OptoScheduling = true; 1229 } 1230 #endif 1231 if (supports_sse4_2()) { // Silvermont 1232 if (FLAG_IS_DEFAULT(UseUnalignedLoadStores)) { 1233 UseUnalignedLoadStores = true; // use movdqu on newest Intel cpus 1234 } 1235 } 1236 if (FLAG_IS_DEFAULT(UseIncDec)) { 1237 FLAG_SET_DEFAULT(UseIncDec, false); 1238 } 1239 } 1240 if(FLAG_IS_DEFAULT(AllocatePrefetchInstr) && supports_3dnow_prefetch()) { 1241 FLAG_SET_DEFAULT(AllocatePrefetchInstr, 3); 1242 } 1243 } 1244 1245 #ifdef _LP64 1246 if (UseSSE42Intrinsics) { 1247 if (FLAG_IS_DEFAULT(UseVectorizedMismatchIntrinsic)) { 1248 UseVectorizedMismatchIntrinsic = true; 1249 } 1250 } else if (UseVectorizedMismatchIntrinsic) { 1251 if (!FLAG_IS_DEFAULT(UseVectorizedMismatchIntrinsic)) 1252 warning("vectorizedMismatch intrinsics are not available on this CPU"); 1253 FLAG_SET_DEFAULT(UseVectorizedMismatchIntrinsic, false); 1254 } 1255 #else 1256 if (UseVectorizedMismatchIntrinsic) { 1257 if (!FLAG_IS_DEFAULT(UseVectorizedMismatchIntrinsic)) { 1258 warning("vectorizedMismatch intrinsic is not available in 32-bit VM"); 1259 } 1260 FLAG_SET_DEFAULT(UseVectorizedMismatchIntrinsic, false); 1261 } 1262 #endif // _LP64 1263 1264 // Use count leading zeros count instruction if available. 1265 if (supports_lzcnt()) { 1266 if (FLAG_IS_DEFAULT(UseCountLeadingZerosInstruction)) { 1267 UseCountLeadingZerosInstruction = true; 1268 } 1269 } else if (UseCountLeadingZerosInstruction) { 1270 warning("lzcnt instruction is not available on this CPU"); 1271 FLAG_SET_DEFAULT(UseCountLeadingZerosInstruction, false); 1272 } 1273 1274 // Use count trailing zeros instruction if available 1275 if (supports_bmi1()) { 1276 // tzcnt does not require VEX prefix 1277 if (FLAG_IS_DEFAULT(UseCountTrailingZerosInstruction)) { 1278 if (!UseBMI1Instructions && !FLAG_IS_DEFAULT(UseBMI1Instructions)) { 1279 // Don't use tzcnt if BMI1 is switched off on command line. 1280 UseCountTrailingZerosInstruction = false; 1281 } else { 1282 UseCountTrailingZerosInstruction = true; 1283 } 1284 } 1285 } else if (UseCountTrailingZerosInstruction) { 1286 warning("tzcnt instruction is not available on this CPU"); 1287 FLAG_SET_DEFAULT(UseCountTrailingZerosInstruction, false); 1288 } 1289 1290 // BMI instructions (except tzcnt) use an encoding with VEX prefix. 1291 // VEX prefix is generated only when AVX > 0. 1292 if (supports_bmi1() && supports_avx()) { 1293 if (FLAG_IS_DEFAULT(UseBMI1Instructions)) { 1294 UseBMI1Instructions = true; 1295 } 1296 } else if (UseBMI1Instructions) { 1297 warning("BMI1 instructions are not available on this CPU (AVX is also required)"); 1298 FLAG_SET_DEFAULT(UseBMI1Instructions, false); 1299 } 1300 1301 if (supports_bmi2() && supports_avx()) { 1302 if (FLAG_IS_DEFAULT(UseBMI2Instructions)) { 1303 UseBMI2Instructions = true; 1304 } 1305 } else if (UseBMI2Instructions) { 1306 warning("BMI2 instructions are not available on this CPU (AVX is also required)"); 1307 FLAG_SET_DEFAULT(UseBMI2Instructions, false); 1308 } 1309 1310 // Use population count instruction if available. 1311 if (supports_popcnt()) { 1312 if (FLAG_IS_DEFAULT(UsePopCountInstruction)) { 1313 UsePopCountInstruction = true; 1314 } 1315 } else if (UsePopCountInstruction) { 1316 warning("POPCNT instruction is not available on this CPU"); 1317 FLAG_SET_DEFAULT(UsePopCountInstruction, false); 1318 } 1319 1320 // Use fast-string operations if available. 1321 if (supports_erms()) { 1322 if (FLAG_IS_DEFAULT(UseFastStosb)) { 1323 UseFastStosb = true; 1324 } 1325 } else if (UseFastStosb) { 1326 warning("fast-string operations are not available on this CPU"); 1327 FLAG_SET_DEFAULT(UseFastStosb, false); 1328 } 1329 1330 #ifdef COMPILER2 1331 if (FLAG_IS_DEFAULT(AlignVector)) { 1332 // Modern processors allow misaligned memory operations for vectors. 1333 AlignVector = !UseUnalignedLoadStores; 1334 } 1335 #endif // COMPILER2 1336 1337 if (FLAG_IS_DEFAULT(AllocatePrefetchInstr)) { 1338 if (AllocatePrefetchInstr == 3 && !supports_3dnow_prefetch()) { 1339 FLAG_SET_DEFAULT(AllocatePrefetchInstr, 0); 1340 } else if (!supports_sse() && supports_3dnow_prefetch()) { 1341 FLAG_SET_DEFAULT(AllocatePrefetchInstr, 3); 1342 } 1343 } 1344 1345 // Allocation prefetch settings 1346 intx cache_line_size = prefetch_data_size(); 1347 if (FLAG_IS_DEFAULT(AllocatePrefetchStepSize) && 1348 (cache_line_size > AllocatePrefetchStepSize)) { 1349 FLAG_SET_DEFAULT(AllocatePrefetchStepSize, cache_line_size); 1350 } 1351 1352 if ((AllocatePrefetchDistance == 0) && (AllocatePrefetchStyle != 0)) { 1353 assert(!FLAG_IS_DEFAULT(AllocatePrefetchDistance), "default value should not be 0"); 1354 if (!FLAG_IS_DEFAULT(AllocatePrefetchStyle)) { 1355 warning("AllocatePrefetchDistance is set to 0 which disable prefetching. Ignoring AllocatePrefetchStyle flag."); 1356 } 1357 FLAG_SET_DEFAULT(AllocatePrefetchStyle, 0); 1358 } 1359 1360 if (FLAG_IS_DEFAULT(AllocatePrefetchDistance)) { 1361 bool use_watermark_prefetch = (AllocatePrefetchStyle == 2); 1362 FLAG_SET_DEFAULT(AllocatePrefetchDistance, allocate_prefetch_distance(use_watermark_prefetch)); 1363 } 1364 1365 if (is_intel() && cpu_family() == 6 && supports_sse3()) { 1366 if (FLAG_IS_DEFAULT(AllocatePrefetchLines) && 1367 supports_sse4_2() && supports_ht()) { // Nehalem based cpus 1368 FLAG_SET_DEFAULT(AllocatePrefetchLines, 4); 1369 } 1370 #ifdef COMPILER2 1371 if (FLAG_IS_DEFAULT(UseFPUForSpilling) && supports_sse4_2()) { 1372 FLAG_SET_DEFAULT(UseFPUForSpilling, true); 1373 } 1374 #endif 1375 } 1376 1377 #ifdef _LP64 1378 // Prefetch settings 1379 1380 // Prefetch interval for gc copy/scan == 9 dcache lines. Derived from 1381 // 50-warehouse specjbb runs on a 2-way 1.8ghz opteron using a 4gb heap. 1382 // Tested intervals from 128 to 2048 in increments of 64 == one cache line. 1383 // 256 bytes (4 dcache lines) was the nearest runner-up to 576. 1384 1385 // gc copy/scan is disabled if prefetchw isn't supported, because 1386 // Prefetch::write emits an inlined prefetchw on Linux. 1387 // Do not use the 3dnow prefetchw instruction. It isn't supported on em64t. 1388 // The used prefetcht0 instruction works for both amd64 and em64t. 1389 1390 if (FLAG_IS_DEFAULT(PrefetchCopyIntervalInBytes)) { 1391 FLAG_SET_DEFAULT(PrefetchCopyIntervalInBytes, 576); 1392 } 1393 if (FLAG_IS_DEFAULT(PrefetchScanIntervalInBytes)) { 1394 FLAG_SET_DEFAULT(PrefetchScanIntervalInBytes, 576); 1395 } 1396 if (FLAG_IS_DEFAULT(PrefetchFieldsAhead)) { 1397 FLAG_SET_DEFAULT(PrefetchFieldsAhead, 1); 1398 } 1399 #endif 1400 1401 if (FLAG_IS_DEFAULT(ContendedPaddingWidth) && 1402 (cache_line_size > ContendedPaddingWidth)) 1403 ContendedPaddingWidth = cache_line_size; 1404 1405 // This machine allows unaligned memory accesses 1406 if (FLAG_IS_DEFAULT(UseUnalignedAccesses)) { 1407 FLAG_SET_DEFAULT(UseUnalignedAccesses, true); 1408 } 1409 1410 #ifndef PRODUCT 1411 if (log_is_enabled(Info, os, cpu)) { 1412 LogStream ls(Log(os, cpu)::info()); 1413 outputStream* log = &ls; 1414 log->print_cr("Logical CPUs per core: %u", 1415 logical_processors_per_package()); 1416 log->print_cr("L1 data cache line size: %u", L1_data_cache_line_size()); 1417 log->print("UseSSE=%d", (int) UseSSE); 1418 if (UseAVX > 0) { 1419 log->print(" UseAVX=%d", (int) UseAVX); 1420 } 1421 if (UseAES) { 1422 log->print(" UseAES=1"); 1423 } 1424 #ifdef COMPILER2 1425 if (MaxVectorSize > 0) { 1426 log->print(" MaxVectorSize=%d", (int) MaxVectorSize); 1427 } 1428 #endif 1429 log->cr(); 1430 log->print("Allocation"); 1431 if (AllocatePrefetchStyle <= 0 || (UseSSE == 0 && !supports_3dnow_prefetch())) { 1432 log->print_cr(": no prefetching"); 1433 } else { 1434 log->print(" prefetching: "); 1435 if (UseSSE == 0 && supports_3dnow_prefetch()) { 1436 log->print("PREFETCHW"); 1437 } else if (UseSSE >= 1) { 1438 if (AllocatePrefetchInstr == 0) { 1439 log->print("PREFETCHNTA"); 1440 } else if (AllocatePrefetchInstr == 1) { 1441 log->print("PREFETCHT0"); 1442 } else if (AllocatePrefetchInstr == 2) { 1443 log->print("PREFETCHT2"); 1444 } else if (AllocatePrefetchInstr == 3) { 1445 log->print("PREFETCHW"); 1446 } 1447 } 1448 if (AllocatePrefetchLines > 1) { 1449 log->print_cr(" at distance %d, %d lines of %d bytes", (int) AllocatePrefetchDistance, (int) AllocatePrefetchLines, (int) AllocatePrefetchStepSize); 1450 } else { 1451 log->print_cr(" at distance %d, one line of %d bytes", (int) AllocatePrefetchDistance, (int) AllocatePrefetchStepSize); 1452 } 1453 } 1454 1455 if (PrefetchCopyIntervalInBytes > 0) { 1456 log->print_cr("PrefetchCopyIntervalInBytes %d", (int) PrefetchCopyIntervalInBytes); 1457 } 1458 if (PrefetchScanIntervalInBytes > 0) { 1459 log->print_cr("PrefetchScanIntervalInBytes %d", (int) PrefetchScanIntervalInBytes); 1460 } 1461 if (PrefetchFieldsAhead > 0) { 1462 log->print_cr("PrefetchFieldsAhead %d", (int) PrefetchFieldsAhead); 1463 } 1464 if (ContendedPaddingWidth > 0) { 1465 log->print_cr("ContendedPaddingWidth %d", (int) ContendedPaddingWidth); 1466 } 1467 } 1468 #endif // !PRODUCT 1469 } 1470 1471 bool VM_Version::use_biased_locking() { 1472 #if INCLUDE_RTM_OPT 1473 // RTM locking is most useful when there is high lock contention and 1474 // low data contention. With high lock contention the lock is usually 1475 // inflated and biased locking is not suitable for that case. 1476 // RTM locking code requires that biased locking is off. 1477 // Note: we can't switch off UseBiasedLocking in get_processor_features() 1478 // because it is used by Thread::allocate() which is called before 1479 // VM_Version::initialize(). 1480 if (UseRTMLocking && UseBiasedLocking) { 1481 if (FLAG_IS_DEFAULT(UseBiasedLocking)) { 1482 FLAG_SET_DEFAULT(UseBiasedLocking, false); 1483 } else { 1484 warning("Biased locking is not supported with RTM locking; ignoring UseBiasedLocking flag." ); 1485 UseBiasedLocking = false; 1486 } 1487 } 1488 #endif 1489 return UseBiasedLocking; 1490 } 1491 1492 void VM_Version::initialize() { 1493 ResourceMark rm; 1494 // Making this stub must be FIRST use of assembler 1495 1496 stub_blob = BufferBlob::create("get_cpu_info_stub", stub_size); 1497 if (stub_blob == NULL) { 1498 vm_exit_during_initialization("Unable to allocate get_cpu_info_stub"); 1499 } 1500 CodeBuffer c(stub_blob); 1501 VM_Version_StubGenerator g(&c); 1502 get_cpu_info_stub = CAST_TO_FN_PTR(get_cpu_info_stub_t, 1503 g.generate_get_cpu_info()); 1504 1505 get_processor_features(); 1506 }