src/cpu/x86/vm/vm_version_x86.cpp

Thu, 03 Nov 2011 04:12:49 -0700

author
twisti
date
Thu, 03 Nov 2011 04:12:49 -0700
changeset 3252
448691f285a5
parent 3156
f08d439fab8c
child 3378
7ab5f6318694
child 3388
127b3692c168
permissions
-rw-r--r--

7106944: assert(_pc == *pc_addr) failed may be too strong
Reviewed-by: kvn, never

twisti@1020 1 /*
katleman@3058 2 * Copyright (c) 1997, 2011, Oracle and/or its affiliates. All rights reserved.
twisti@1020 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
twisti@1020 4 *
twisti@1020 5 * This code is free software; you can redistribute it and/or modify it
twisti@1020 6 * under the terms of the GNU General Public License version 2 only, as
twisti@1020 7 * published by the Free Software Foundation.
twisti@1020 8 *
twisti@1020 9 * This code is distributed in the hope that it will be useful, but WITHOUT
twisti@1020 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
twisti@1020 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
twisti@1020 12 * version 2 for more details (a copy is included in the LICENSE file that
twisti@1020 13 * accompanied this code).
twisti@1020 14 *
twisti@1020 15 * You should have received a copy of the GNU General Public License version
twisti@1020 16 * 2 along with this work; if not, write to the Free Software Foundation,
twisti@1020 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
twisti@1020 18 *
trims@1907 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
trims@1907 20 * or visit www.oracle.com if you need additional information or have any
trims@1907 21 * questions.
twisti@1020 22 *
twisti@1020 23 */
twisti@1020 24
stefank@2314 25 #include "precompiled.hpp"
stefank@2314 26 #include "assembler_x86.inline.hpp"
stefank@2314 27 #include "memory/resourceArea.hpp"
stefank@2314 28 #include "runtime/java.hpp"
stefank@2314 29 #include "runtime/stubCodeGenerator.hpp"
stefank@2314 30 #include "vm_version_x86.hpp"
stefank@2314 31 #ifdef TARGET_OS_FAMILY_linux
stefank@2314 32 # include "os_linux.inline.hpp"
stefank@2314 33 #endif
stefank@2314 34 #ifdef TARGET_OS_FAMILY_solaris
stefank@2314 35 # include "os_solaris.inline.hpp"
stefank@2314 36 #endif
stefank@2314 37 #ifdef TARGET_OS_FAMILY_windows
stefank@2314 38 # include "os_windows.inline.hpp"
stefank@2314 39 #endif
never@3156 40 #ifdef TARGET_OS_FAMILY_bsd
never@3156 41 # include "os_bsd.inline.hpp"
never@3156 42 #endif
twisti@1020 43
twisti@1020 44
twisti@1020 45 int VM_Version::_cpu;
twisti@1020 46 int VM_Version::_model;
twisti@1020 47 int VM_Version::_stepping;
twisti@1020 48 int VM_Version::_cpuFeatures;
twisti@1020 49 const char* VM_Version::_features_str = "";
twisti@1020 50 VM_Version::CpuidInfo VM_Version::_cpuid_info = { 0, };
twisti@1020 51
twisti@1020 52 static BufferBlob* stub_blob;
kvn@1977 53 static const int stub_size = 400;
twisti@1020 54
twisti@1020 55 extern "C" {
twisti@1020 56 typedef void (*getPsrInfo_stub_t)(void*);
twisti@1020 57 }
twisti@1020 58 static getPsrInfo_stub_t getPsrInfo_stub = NULL;
twisti@1020 59
twisti@1020 60
twisti@1020 61 class VM_Version_StubGenerator: public StubCodeGenerator {
twisti@1020 62 public:
twisti@1020 63
twisti@1020 64 VM_Version_StubGenerator(CodeBuffer *c) : StubCodeGenerator(c) {}
twisti@1020 65
twisti@1020 66 address generate_getPsrInfo() {
twisti@1020 67 // Flags to test CPU type.
twisti@1020 68 const uint32_t EFL_AC = 0x40000;
twisti@1020 69 const uint32_t EFL_ID = 0x200000;
twisti@1020 70 // Values for when we don't have a CPUID instruction.
twisti@1020 71 const int CPU_FAMILY_SHIFT = 8;
twisti@1020 72 const uint32_t CPU_FAMILY_386 = (3 << CPU_FAMILY_SHIFT);
twisti@1020 73 const uint32_t CPU_FAMILY_486 = (4 << CPU_FAMILY_SHIFT);
twisti@1020 74
kvn@1977 75 Label detect_486, cpu486, detect_586, std_cpuid1, std_cpuid4;
twisti@1020 76 Label ext_cpuid1, ext_cpuid5, done;
twisti@1020 77
twisti@1020 78 StubCodeMark mark(this, "VM_Version", "getPsrInfo_stub");
twisti@1020 79 # define __ _masm->
twisti@1020 80
twisti@1020 81 address start = __ pc();
twisti@1020 82
twisti@1020 83 //
twisti@1020 84 // void getPsrInfo(VM_Version::CpuidInfo* cpuid_info);
twisti@1020 85 //
twisti@1020 86 // LP64: rcx and rdx are first and second argument registers on windows
twisti@1020 87
twisti@1020 88 __ push(rbp);
twisti@1020 89 #ifdef _LP64
twisti@1020 90 __ mov(rbp, c_rarg0); // cpuid_info address
twisti@1020 91 #else
twisti@1020 92 __ movptr(rbp, Address(rsp, 8)); // cpuid_info address
twisti@1020 93 #endif
twisti@1020 94 __ push(rbx);
twisti@1020 95 __ push(rsi);
twisti@1020 96 __ pushf(); // preserve rbx, and flags
twisti@1020 97 __ pop(rax);
twisti@1020 98 __ push(rax);
twisti@1020 99 __ mov(rcx, rax);
twisti@1020 100 //
twisti@1020 101 // if we are unable to change the AC flag, we have a 386
twisti@1020 102 //
twisti@1020 103 __ xorl(rax, EFL_AC);
twisti@1020 104 __ push(rax);
twisti@1020 105 __ popf();
twisti@1020 106 __ pushf();
twisti@1020 107 __ pop(rax);
twisti@1020 108 __ cmpptr(rax, rcx);
twisti@1020 109 __ jccb(Assembler::notEqual, detect_486);
twisti@1020 110
twisti@1020 111 __ movl(rax, CPU_FAMILY_386);
twisti@1020 112 __ movl(Address(rbp, in_bytes(VM_Version::std_cpuid1_offset())), rax);
twisti@1020 113 __ jmp(done);
twisti@1020 114
twisti@1020 115 //
twisti@1020 116 // If we are unable to change the ID flag, we have a 486 which does
twisti@1020 117 // not support the "cpuid" instruction.
twisti@1020 118 //
twisti@1020 119 __ bind(detect_486);
twisti@1020 120 __ mov(rax, rcx);
twisti@1020 121 __ xorl(rax, EFL_ID);
twisti@1020 122 __ push(rax);
twisti@1020 123 __ popf();
twisti@1020 124 __ pushf();
twisti@1020 125 __ pop(rax);
twisti@1020 126 __ cmpptr(rcx, rax);
twisti@1020 127 __ jccb(Assembler::notEqual, detect_586);
twisti@1020 128
twisti@1020 129 __ bind(cpu486);
twisti@1020 130 __ movl(rax, CPU_FAMILY_486);
twisti@1020 131 __ movl(Address(rbp, in_bytes(VM_Version::std_cpuid1_offset())), rax);
twisti@1020 132 __ jmp(done);
twisti@1020 133
twisti@1020 134 //
twisti@1020 135 // At this point, we have a chip which supports the "cpuid" instruction
twisti@1020 136 //
twisti@1020 137 __ bind(detect_586);
twisti@1020 138 __ xorl(rax, rax);
twisti@1020 139 __ cpuid();
twisti@1020 140 __ orl(rax, rax);
twisti@1020 141 __ jcc(Assembler::equal, cpu486); // if cpuid doesn't support an input
twisti@1020 142 // value of at least 1, we give up and
twisti@1020 143 // assume a 486
twisti@1020 144 __ lea(rsi, Address(rbp, in_bytes(VM_Version::std_cpuid0_offset())));
twisti@1020 145 __ movl(Address(rsi, 0), rax);
twisti@1020 146 __ movl(Address(rsi, 4), rbx);
twisti@1020 147 __ movl(Address(rsi, 8), rcx);
twisti@1020 148 __ movl(Address(rsi,12), rdx);
twisti@1020 149
kvn@1977 150 __ cmpl(rax, 0xa); // Is cpuid(0xB) supported?
kvn@1977 151 __ jccb(Assembler::belowEqual, std_cpuid4);
kvn@1977 152
kvn@1977 153 //
kvn@1977 154 // cpuid(0xB) Processor Topology
kvn@1977 155 //
kvn@1977 156 __ movl(rax, 0xb);
kvn@1977 157 __ xorl(rcx, rcx); // Threads level
kvn@1977 158 __ cpuid();
kvn@1977 159
kvn@1977 160 __ lea(rsi, Address(rbp, in_bytes(VM_Version::tpl_cpuidB0_offset())));
kvn@1977 161 __ movl(Address(rsi, 0), rax);
kvn@1977 162 __ movl(Address(rsi, 4), rbx);
kvn@1977 163 __ movl(Address(rsi, 8), rcx);
kvn@1977 164 __ movl(Address(rsi,12), rdx);
kvn@1977 165
kvn@1977 166 __ movl(rax, 0xb);
kvn@1977 167 __ movl(rcx, 1); // Cores level
kvn@1977 168 __ cpuid();
kvn@1977 169 __ push(rax);
kvn@1977 170 __ andl(rax, 0x1f); // Determine if valid topology level
kvn@1977 171 __ orl(rax, rbx); // eax[4:0] | ebx[0:15] == 0 indicates invalid level
kvn@1977 172 __ andl(rax, 0xffff);
kvn@1977 173 __ pop(rax);
kvn@1977 174 __ jccb(Assembler::equal, std_cpuid4);
kvn@1977 175
kvn@1977 176 __ lea(rsi, Address(rbp, in_bytes(VM_Version::tpl_cpuidB1_offset())));
kvn@1977 177 __ movl(Address(rsi, 0), rax);
kvn@1977 178 __ movl(Address(rsi, 4), rbx);
kvn@1977 179 __ movl(Address(rsi, 8), rcx);
kvn@1977 180 __ movl(Address(rsi,12), rdx);
kvn@1977 181
kvn@1977 182 __ movl(rax, 0xb);
kvn@1977 183 __ movl(rcx, 2); // Packages level
kvn@1977 184 __ cpuid();
kvn@1977 185 __ push(rax);
kvn@1977 186 __ andl(rax, 0x1f); // Determine if valid topology level
kvn@1977 187 __ orl(rax, rbx); // eax[4:0] | ebx[0:15] == 0 indicates invalid level
kvn@1977 188 __ andl(rax, 0xffff);
kvn@1977 189 __ pop(rax);
kvn@1977 190 __ jccb(Assembler::equal, std_cpuid4);
kvn@1977 191
kvn@1977 192 __ lea(rsi, Address(rbp, in_bytes(VM_Version::tpl_cpuidB2_offset())));
kvn@1977 193 __ movl(Address(rsi, 0), rax);
kvn@1977 194 __ movl(Address(rsi, 4), rbx);
kvn@1977 195 __ movl(Address(rsi, 8), rcx);
kvn@1977 196 __ movl(Address(rsi,12), rdx);
twisti@1020 197
twisti@1020 198 //
twisti@1020 199 // cpuid(0x4) Deterministic cache params
twisti@1020 200 //
kvn@1977 201 __ bind(std_cpuid4);
twisti@1020 202 __ movl(rax, 4);
kvn@1977 203 __ cmpl(rax, Address(rbp, in_bytes(VM_Version::std_cpuid0_offset()))); // Is cpuid(0x4) supported?
kvn@1977 204 __ jccb(Assembler::greater, std_cpuid1);
kvn@1977 205
twisti@1020 206 __ xorl(rcx, rcx); // L1 cache
twisti@1020 207 __ cpuid();
twisti@1020 208 __ push(rax);
twisti@1020 209 __ andl(rax, 0x1f); // Determine if valid cache parameters used
twisti@1020 210 __ orl(rax, rax); // eax[4:0] == 0 indicates invalid cache
twisti@1020 211 __ pop(rax);
twisti@1020 212 __ jccb(Assembler::equal, std_cpuid1);
twisti@1020 213
twisti@1020 214 __ lea(rsi, Address(rbp, in_bytes(VM_Version::dcp_cpuid4_offset())));
twisti@1020 215 __ movl(Address(rsi, 0), rax);
twisti@1020 216 __ movl(Address(rsi, 4), rbx);
twisti@1020 217 __ movl(Address(rsi, 8), rcx);
twisti@1020 218 __ movl(Address(rsi,12), rdx);
twisti@1020 219
twisti@1020 220 //
twisti@1020 221 // Standard cpuid(0x1)
twisti@1020 222 //
twisti@1020 223 __ bind(std_cpuid1);
twisti@1020 224 __ movl(rax, 1);
twisti@1020 225 __ cpuid();
twisti@1020 226 __ lea(rsi, Address(rbp, in_bytes(VM_Version::std_cpuid1_offset())));
twisti@1020 227 __ movl(Address(rsi, 0), rax);
twisti@1020 228 __ movl(Address(rsi, 4), rbx);
twisti@1020 229 __ movl(Address(rsi, 8), rcx);
twisti@1020 230 __ movl(Address(rsi,12), rdx);
twisti@1020 231
twisti@1020 232 __ movl(rax, 0x80000000);
twisti@1020 233 __ cpuid();
twisti@1020 234 __ cmpl(rax, 0x80000000); // Is cpuid(0x80000001) supported?
twisti@1020 235 __ jcc(Assembler::belowEqual, done);
twisti@1020 236 __ cmpl(rax, 0x80000004); // Is cpuid(0x80000005) supported?
twisti@1020 237 __ jccb(Assembler::belowEqual, ext_cpuid1);
twisti@1020 238 __ cmpl(rax, 0x80000007); // Is cpuid(0x80000008) supported?
twisti@1020 239 __ jccb(Assembler::belowEqual, ext_cpuid5);
twisti@1020 240 //
twisti@1020 241 // Extended cpuid(0x80000008)
twisti@1020 242 //
twisti@1020 243 __ movl(rax, 0x80000008);
twisti@1020 244 __ cpuid();
twisti@1020 245 __ lea(rsi, Address(rbp, in_bytes(VM_Version::ext_cpuid8_offset())));
twisti@1020 246 __ movl(Address(rsi, 0), rax);
twisti@1020 247 __ movl(Address(rsi, 4), rbx);
twisti@1020 248 __ movl(Address(rsi, 8), rcx);
twisti@1020 249 __ movl(Address(rsi,12), rdx);
twisti@1020 250
twisti@1020 251 //
twisti@1020 252 // Extended cpuid(0x80000005)
twisti@1020 253 //
twisti@1020 254 __ bind(ext_cpuid5);
twisti@1020 255 __ movl(rax, 0x80000005);
twisti@1020 256 __ cpuid();
twisti@1020 257 __ lea(rsi, Address(rbp, in_bytes(VM_Version::ext_cpuid5_offset())));
twisti@1020 258 __ movl(Address(rsi, 0), rax);
twisti@1020 259 __ movl(Address(rsi, 4), rbx);
twisti@1020 260 __ movl(Address(rsi, 8), rcx);
twisti@1020 261 __ movl(Address(rsi,12), rdx);
twisti@1020 262
twisti@1020 263 //
twisti@1020 264 // Extended cpuid(0x80000001)
twisti@1020 265 //
twisti@1020 266 __ bind(ext_cpuid1);
twisti@1020 267 __ movl(rax, 0x80000001);
twisti@1020 268 __ cpuid();
twisti@1020 269 __ lea(rsi, Address(rbp, in_bytes(VM_Version::ext_cpuid1_offset())));
twisti@1020 270 __ movl(Address(rsi, 0), rax);
twisti@1020 271 __ movl(Address(rsi, 4), rbx);
twisti@1020 272 __ movl(Address(rsi, 8), rcx);
twisti@1020 273 __ movl(Address(rsi,12), rdx);
twisti@1020 274
twisti@1020 275 //
twisti@1020 276 // return
twisti@1020 277 //
twisti@1020 278 __ bind(done);
twisti@1020 279 __ popf();
twisti@1020 280 __ pop(rsi);
twisti@1020 281 __ pop(rbx);
twisti@1020 282 __ pop(rbp);
twisti@1020 283 __ ret(0);
twisti@1020 284
twisti@1020 285 # undef __
twisti@1020 286
twisti@1020 287 return start;
twisti@1020 288 };
twisti@1020 289 };
twisti@1020 290
twisti@1020 291
twisti@1020 292 void VM_Version::get_processor_features() {
twisti@1020 293
twisti@1020 294 _cpu = 4; // 486 by default
twisti@1020 295 _model = 0;
twisti@1020 296 _stepping = 0;
twisti@1020 297 _cpuFeatures = 0;
twisti@1020 298 _logical_processors_per_package = 1;
twisti@1020 299
twisti@1020 300 if (!Use486InstrsOnly) {
twisti@1020 301 // Get raw processor info
twisti@1020 302 getPsrInfo_stub(&_cpuid_info);
twisti@1020 303 assert_is_initialized();
twisti@1020 304 _cpu = extended_cpu_family();
twisti@1020 305 _model = extended_cpu_model();
twisti@1020 306 _stepping = cpu_stepping();
twisti@1020 307
twisti@1020 308 if (cpu_family() > 4) { // it supports CPUID
twisti@1020 309 _cpuFeatures = feature_flags();
twisti@1020 310 // Logical processors are only available on P4s and above,
twisti@1020 311 // and only if hyperthreading is available.
twisti@1020 312 _logical_processors_per_package = logical_processor_count();
twisti@1020 313 }
twisti@1020 314 }
twisti@1020 315
twisti@1020 316 _supports_cx8 = supports_cmpxchg8();
twisti@1020 317
twisti@1020 318 #ifdef _LP64
twisti@1020 319 // OS should support SSE for x64 and hardware should support at least SSE2.
twisti@1020 320 if (!VM_Version::supports_sse2()) {
twisti@1020 321 vm_exit_during_initialization("Unknown x64 processor: SSE2 not supported");
twisti@1020 322 }
roland@1495 323 // in 64 bit the use of SSE2 is the minimum
roland@1495 324 if (UseSSE < 2) UseSSE = 2;
twisti@1020 325 #endif
twisti@1020 326
kvn@2984 327 #ifdef AMD64
kvn@2984 328 // flush_icache_stub have to be generated first.
kvn@2984 329 // That is why Icache line size is hard coded in ICache class,
kvn@2984 330 // see icache_x86.hpp. It is also the reason why we can't use
kvn@2984 331 // clflush instruction in 32-bit VM since it could be running
kvn@2984 332 // on CPU which does not support it.
kvn@2984 333 //
kvn@2984 334 // The only thing we can do is to verify that flushed
kvn@2984 335 // ICache::line_size has correct value.
kvn@2984 336 guarantee(_cpuid_info.std_cpuid1_edx.bits.clflush != 0, "clflush is not supported");
kvn@2984 337 // clflush_size is size in quadwords (8 bytes).
kvn@2984 338 guarantee(_cpuid_info.std_cpuid1_ebx.bits.clflush_size == 8, "such clflush size is not supported");
kvn@2984 339 #endif
kvn@2984 340
twisti@1020 341 // If the OS doesn't support SSE, we can't use this feature even if the HW does
twisti@1020 342 if (!os::supports_sse())
twisti@1020 343 _cpuFeatures &= ~(CPU_SSE|CPU_SSE2|CPU_SSE3|CPU_SSSE3|CPU_SSE4A|CPU_SSE4_1|CPU_SSE4_2);
twisti@1020 344
twisti@1020 345 if (UseSSE < 4) {
twisti@1020 346 _cpuFeatures &= ~CPU_SSE4_1;
twisti@1020 347 _cpuFeatures &= ~CPU_SSE4_2;
twisti@1020 348 }
twisti@1020 349
twisti@1020 350 if (UseSSE < 3) {
twisti@1020 351 _cpuFeatures &= ~CPU_SSE3;
twisti@1020 352 _cpuFeatures &= ~CPU_SSSE3;
twisti@1020 353 _cpuFeatures &= ~CPU_SSE4A;
twisti@1020 354 }
twisti@1020 355
twisti@1020 356 if (UseSSE < 2)
twisti@1020 357 _cpuFeatures &= ~CPU_SSE2;
twisti@1020 358
twisti@1020 359 if (UseSSE < 1)
twisti@1020 360 _cpuFeatures &= ~CPU_SSE;
twisti@1020 361
twisti@1020 362 if (logical_processors_per_package() == 1) {
twisti@1020 363 // HT processor could be installed on a system which doesn't support HT.
twisti@1020 364 _cpuFeatures &= ~CPU_HT;
twisti@1020 365 }
twisti@1020 366
twisti@1020 367 char buf[256];
kvn@2761 368 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",
twisti@1020 369 cores_per_cpu(), threads_per_core(),
twisti@1020 370 cpu_family(), _model, _stepping,
twisti@1020 371 (supports_cmov() ? ", cmov" : ""),
twisti@1020 372 (supports_cmpxchg8() ? ", cx8" : ""),
twisti@1020 373 (supports_fxsr() ? ", fxsr" : ""),
twisti@1020 374 (supports_mmx() ? ", mmx" : ""),
twisti@1020 375 (supports_sse() ? ", sse" : ""),
twisti@1020 376 (supports_sse2() ? ", sse2" : ""),
twisti@1020 377 (supports_sse3() ? ", sse3" : ""),
twisti@1020 378 (supports_ssse3()? ", ssse3": ""),
twisti@1020 379 (supports_sse4_1() ? ", sse4.1" : ""),
twisti@1020 380 (supports_sse4_2() ? ", sse4.2" : ""),
twisti@1078 381 (supports_popcnt() ? ", popcnt" : ""),
twisti@1020 382 (supports_mmx_ext() ? ", mmxext" : ""),
kvn@2761 383 (supports_3dnow_prefetch() ? ", 3dnowpref" : ""),
twisti@1210 384 (supports_lzcnt() ? ", lzcnt": ""),
twisti@1020 385 (supports_sse4a() ? ", sse4a": ""),
twisti@1020 386 (supports_ht() ? ", ht": ""));
twisti@1020 387 _features_str = strdup(buf);
twisti@1020 388
twisti@1020 389 // UseSSE is set to the smaller of what hardware supports and what
twisti@1020 390 // the command line requires. I.e., you cannot set UseSSE to 2 on
twisti@1020 391 // older Pentiums which do not support it.
twisti@1020 392 if( UseSSE > 4 ) UseSSE=4;
twisti@1020 393 if( UseSSE < 0 ) UseSSE=0;
twisti@1020 394 if( !supports_sse4_1() ) // Drop to 3 if no SSE4 support
twisti@1020 395 UseSSE = MIN2((intx)3,UseSSE);
twisti@1020 396 if( !supports_sse3() ) // Drop to 2 if no SSE3 support
twisti@1020 397 UseSSE = MIN2((intx)2,UseSSE);
twisti@1020 398 if( !supports_sse2() ) // Drop to 1 if no SSE2 support
twisti@1020 399 UseSSE = MIN2((intx)1,UseSSE);
twisti@1020 400 if( !supports_sse () ) // Drop to 0 if no SSE support
twisti@1020 401 UseSSE = 0;
twisti@1020 402
twisti@1020 403 // On new cpus instructions which update whole XMM register should be used
twisti@1020 404 // to prevent partial register stall due to dependencies on high half.
twisti@1020 405 //
twisti@1020 406 // UseXmmLoadAndClearUpper == true --> movsd(xmm, mem)
twisti@1020 407 // UseXmmLoadAndClearUpper == false --> movlpd(xmm, mem)
twisti@1020 408 // UseXmmRegToRegMoveAll == true --> movaps(xmm, xmm), movapd(xmm, xmm).
twisti@1020 409 // UseXmmRegToRegMoveAll == false --> movss(xmm, xmm), movsd(xmm, xmm).
twisti@1020 410
twisti@1020 411 if( is_amd() ) { // AMD cpus specific settings
twisti@1020 412 if( supports_sse2() && FLAG_IS_DEFAULT(UseAddressNop) ) {
twisti@1020 413 // Use it on new AMD cpus starting from Opteron.
twisti@1020 414 UseAddressNop = true;
twisti@1020 415 }
twisti@1020 416 if( supports_sse2() && FLAG_IS_DEFAULT(UseNewLongLShift) ) {
twisti@1020 417 // Use it on new AMD cpus starting from Opteron.
twisti@1020 418 UseNewLongLShift = true;
twisti@1020 419 }
twisti@1020 420 if( FLAG_IS_DEFAULT(UseXmmLoadAndClearUpper) ) {
twisti@1020 421 if( supports_sse4a() ) {
twisti@1020 422 UseXmmLoadAndClearUpper = true; // use movsd only on '10h' Opteron
twisti@1020 423 } else {
twisti@1020 424 UseXmmLoadAndClearUpper = false;
twisti@1020 425 }
twisti@1020 426 }
twisti@1020 427 if( FLAG_IS_DEFAULT(UseXmmRegToRegMoveAll) ) {
twisti@1020 428 if( supports_sse4a() ) {
twisti@1020 429 UseXmmRegToRegMoveAll = true; // use movaps, movapd only on '10h'
twisti@1020 430 } else {
twisti@1020 431 UseXmmRegToRegMoveAll = false;
twisti@1020 432 }
twisti@1020 433 }
twisti@1020 434 if( FLAG_IS_DEFAULT(UseXmmI2F) ) {
twisti@1020 435 if( supports_sse4a() ) {
twisti@1020 436 UseXmmI2F = true;
twisti@1020 437 } else {
twisti@1020 438 UseXmmI2F = false;
twisti@1020 439 }
twisti@1020 440 }
twisti@1020 441 if( FLAG_IS_DEFAULT(UseXmmI2D) ) {
twisti@1020 442 if( supports_sse4a() ) {
twisti@1020 443 UseXmmI2D = true;
twisti@1020 444 } else {
twisti@1020 445 UseXmmI2D = false;
twisti@1020 446 }
twisti@1020 447 }
kvn@2688 448 if( FLAG_IS_DEFAULT(UseSSE42Intrinsics) ) {
kvn@2688 449 if( supports_sse4_2() && UseSSE >= 4 ) {
kvn@2688 450 UseSSE42Intrinsics = true;
kvn@2688 451 }
kvn@2688 452 }
twisti@1210 453
twisti@1210 454 // Use count leading zeros count instruction if available.
twisti@1210 455 if (supports_lzcnt()) {
twisti@1210 456 if (FLAG_IS_DEFAULT(UseCountLeadingZerosInstruction)) {
twisti@1210 457 UseCountLeadingZerosInstruction = true;
twisti@1210 458 }
twisti@1210 459 }
kvn@2640 460
kvn@2808 461 // some defaults for AMD family 15h
kvn@2808 462 if ( cpu_family() == 0x15 ) {
kvn@2808 463 // On family 15h processors default is no sw prefetch
kvn@2640 464 if (FLAG_IS_DEFAULT(AllocatePrefetchStyle)) {
kvn@2640 465 AllocatePrefetchStyle = 0;
kvn@2640 466 }
kvn@2808 467 // Also, if some other prefetch style is specified, default instruction type is PREFETCHW
kvn@2808 468 if (FLAG_IS_DEFAULT(AllocatePrefetchInstr)) {
kvn@2808 469 AllocatePrefetchInstr = 3;
kvn@2808 470 }
kvn@2808 471 // On family 15h processors use XMM and UnalignedLoadStores for Array Copy
kvn@2808 472 if( FLAG_IS_DEFAULT(UseXMMForArrayCopy) ) {
kvn@2808 473 UseXMMForArrayCopy = true;
kvn@2808 474 }
kvn@2808 475 if( FLAG_IS_DEFAULT(UseUnalignedLoadStores) && UseXMMForArrayCopy ) {
kvn@2808 476 UseUnalignedLoadStores = true;
kvn@2808 477 }
kvn@2640 478 }
kvn@2808 479
twisti@1020 480 }
twisti@1020 481
twisti@1020 482 if( is_intel() ) { // Intel cpus specific settings
twisti@1020 483 if( FLAG_IS_DEFAULT(UseStoreImmI16) ) {
twisti@1020 484 UseStoreImmI16 = false; // don't use it on Intel cpus
twisti@1020 485 }
twisti@1020 486 if( cpu_family() == 6 || cpu_family() == 15 ) {
twisti@1020 487 if( FLAG_IS_DEFAULT(UseAddressNop) ) {
twisti@1020 488 // Use it on all Intel cpus starting from PentiumPro
twisti@1020 489 UseAddressNop = true;
twisti@1020 490 }
twisti@1020 491 }
twisti@1020 492 if( FLAG_IS_DEFAULT(UseXmmLoadAndClearUpper) ) {
twisti@1020 493 UseXmmLoadAndClearUpper = true; // use movsd on all Intel cpus
twisti@1020 494 }
twisti@1020 495 if( FLAG_IS_DEFAULT(UseXmmRegToRegMoveAll) ) {
twisti@1020 496 if( supports_sse3() ) {
twisti@1020 497 UseXmmRegToRegMoveAll = true; // use movaps, movapd on new Intel cpus
twisti@1020 498 } else {
twisti@1020 499 UseXmmRegToRegMoveAll = false;
twisti@1020 500 }
twisti@1020 501 }
twisti@1020 502 if( cpu_family() == 6 && supports_sse3() ) { // New Intel cpus
twisti@1020 503 #ifdef COMPILER2
twisti@1020 504 if( FLAG_IS_DEFAULT(MaxLoopPad) ) {
twisti@1020 505 // For new Intel cpus do the next optimization:
twisti@1020 506 // don't align the beginning of a loop if there are enough instructions
twisti@1020 507 // left (NumberOfLoopInstrToAlign defined in c2_globals.hpp)
twisti@1020 508 // in current fetch line (OptoLoopAlignment) or the padding
twisti@1020 509 // is big (> MaxLoopPad).
twisti@1020 510 // Set MaxLoopPad to 11 for new Intel cpus to reduce number of
twisti@1020 511 // generated NOP instructions. 11 is the largest size of one
twisti@1020 512 // address NOP instruction '0F 1F' (see Assembler::nop(i)).
twisti@1020 513 MaxLoopPad = 11;
twisti@1020 514 }
twisti@1020 515 #endif // COMPILER2
twisti@1020 516 if( FLAG_IS_DEFAULT(UseXMMForArrayCopy) ) {
twisti@1020 517 UseXMMForArrayCopy = true; // use SSE2 movq on new Intel cpus
twisti@1020 518 }
twisti@1020 519 if( supports_sse4_2() && supports_ht() ) { // Newest Intel cpus
twisti@1020 520 if( FLAG_IS_DEFAULT(UseUnalignedLoadStores) && UseXMMForArrayCopy ) {
twisti@1020 521 UseUnalignedLoadStores = true; // use movdqu on newest Intel cpus
twisti@1020 522 }
twisti@1020 523 }
cfang@1116 524 if( supports_sse4_2() && UseSSE >= 4 ) {
cfang@1116 525 if( FLAG_IS_DEFAULT(UseSSE42Intrinsics)) {
cfang@1116 526 UseSSE42Intrinsics = true;
cfang@1116 527 }
cfang@1116 528 }
twisti@1020 529 }
twisti@1020 530 }
twisti@1020 531
twisti@1078 532 // Use population count instruction if available.
twisti@1078 533 if (supports_popcnt()) {
twisti@1078 534 if (FLAG_IS_DEFAULT(UsePopCountInstruction)) {
twisti@1078 535 UsePopCountInstruction = true;
twisti@1078 536 }
twisti@1078 537 }
twisti@1078 538
never@2085 539 #ifdef COMPILER2
never@2085 540 if (UseFPUForSpilling) {
never@2085 541 if (UseSSE < 2) {
never@2085 542 // Only supported with SSE2+
never@2085 543 FLAG_SET_DEFAULT(UseFPUForSpilling, false);
never@2085 544 }
never@2085 545 }
never@2085 546 #endif
never@2085 547
twisti@1020 548 assert(0 <= ReadPrefetchInstr && ReadPrefetchInstr <= 3, "invalid value");
twisti@1020 549 assert(0 <= AllocatePrefetchInstr && AllocatePrefetchInstr <= 3, "invalid value");
twisti@1020 550
twisti@1020 551 // set valid Prefetch instruction
twisti@1020 552 if( ReadPrefetchInstr < 0 ) ReadPrefetchInstr = 0;
twisti@1020 553 if( ReadPrefetchInstr > 3 ) ReadPrefetchInstr = 3;
kvn@2761 554 if( ReadPrefetchInstr == 3 && !supports_3dnow_prefetch() ) ReadPrefetchInstr = 0;
kvn@2761 555 if( !supports_sse() && supports_3dnow_prefetch() ) ReadPrefetchInstr = 3;
twisti@1020 556
twisti@1020 557 if( AllocatePrefetchInstr < 0 ) AllocatePrefetchInstr = 0;
twisti@1020 558 if( AllocatePrefetchInstr > 3 ) AllocatePrefetchInstr = 3;
kvn@2761 559 if( AllocatePrefetchInstr == 3 && !supports_3dnow_prefetch() ) AllocatePrefetchInstr=0;
kvn@2761 560 if( !supports_sse() && supports_3dnow_prefetch() ) AllocatePrefetchInstr = 3;
twisti@1020 561
twisti@1020 562 // Allocation prefetch settings
kvn@3052 563 intx cache_line_size = prefetch_data_size();
twisti@1020 564 if( cache_line_size > AllocatePrefetchStepSize )
twisti@1020 565 AllocatePrefetchStepSize = cache_line_size;
kvn@3052 566
twisti@1020 567 assert(AllocatePrefetchLines > 0, "invalid value");
kvn@3052 568 if( AllocatePrefetchLines < 1 ) // set valid value in product VM
kvn@3052 569 AllocatePrefetchLines = 3;
kvn@3052 570 assert(AllocateInstancePrefetchLines > 0, "invalid value");
kvn@3052 571 if( AllocateInstancePrefetchLines < 1 ) // set valid value in product VM
kvn@3052 572 AllocateInstancePrefetchLines = 1;
twisti@1020 573
twisti@1020 574 AllocatePrefetchDistance = allocate_prefetch_distance();
twisti@1020 575 AllocatePrefetchStyle = allocate_prefetch_style();
twisti@1020 576
kvn@1977 577 if( is_intel() && cpu_family() == 6 && supports_sse3() ) {
kvn@1977 578 if( AllocatePrefetchStyle == 2 ) { // watermark prefetching on Core
twisti@1020 579 #ifdef _LP64
kvn@1977 580 AllocatePrefetchDistance = 384;
twisti@1020 581 #else
kvn@1977 582 AllocatePrefetchDistance = 320;
twisti@1020 583 #endif
kvn@1977 584 }
kvn@1977 585 if( supports_sse4_2() && supports_ht() ) { // Nehalem based cpus
kvn@1977 586 AllocatePrefetchDistance = 192;
kvn@1977 587 AllocatePrefetchLines = 4;
never@2085 588 #ifdef COMPILER2
never@2085 589 if (AggressiveOpts && FLAG_IS_DEFAULT(UseFPUForSpilling)) {
never@2085 590 FLAG_SET_DEFAULT(UseFPUForSpilling, true);
never@2085 591 }
never@2085 592 #endif
kvn@1977 593 }
twisti@1020 594 }
twisti@1020 595 assert(AllocatePrefetchDistance % AllocatePrefetchStepSize == 0, "invalid value");
twisti@1020 596
twisti@1020 597 #ifdef _LP64
twisti@1020 598 // Prefetch settings
twisti@1020 599 PrefetchCopyIntervalInBytes = prefetch_copy_interval_in_bytes();
twisti@1020 600 PrefetchScanIntervalInBytes = prefetch_scan_interval_in_bytes();
twisti@1020 601 PrefetchFieldsAhead = prefetch_fields_ahead();
twisti@1020 602 #endif
twisti@1020 603
twisti@1020 604 #ifndef PRODUCT
twisti@1020 605 if (PrintMiscellaneous && Verbose) {
twisti@1020 606 tty->print_cr("Logical CPUs per core: %u",
twisti@1020 607 logical_processors_per_package());
twisti@1020 608 tty->print_cr("UseSSE=%d",UseSSE);
kvn@3052 609 tty->print("Allocation");
kvn@2761 610 if (AllocatePrefetchStyle <= 0 || UseSSE == 0 && !supports_3dnow_prefetch()) {
kvn@3052 611 tty->print_cr(": no prefetching");
twisti@1020 612 } else {
kvn@3052 613 tty->print(" prefetching: ");
kvn@2761 614 if (UseSSE == 0 && supports_3dnow_prefetch()) {
twisti@1020 615 tty->print("PREFETCHW");
twisti@1020 616 } else if (UseSSE >= 1) {
twisti@1020 617 if (AllocatePrefetchInstr == 0) {
twisti@1020 618 tty->print("PREFETCHNTA");
twisti@1020 619 } else if (AllocatePrefetchInstr == 1) {
twisti@1020 620 tty->print("PREFETCHT0");
twisti@1020 621 } else if (AllocatePrefetchInstr == 2) {
twisti@1020 622 tty->print("PREFETCHT2");
twisti@1020 623 } else if (AllocatePrefetchInstr == 3) {
twisti@1020 624 tty->print("PREFETCHW");
twisti@1020 625 }
twisti@1020 626 }
twisti@1020 627 if (AllocatePrefetchLines > 1) {
kvn@3052 628 tty->print_cr(" at distance %d, %d lines of %d bytes", AllocatePrefetchDistance, AllocatePrefetchLines, AllocatePrefetchStepSize);
twisti@1020 629 } else {
kvn@3052 630 tty->print_cr(" at distance %d, one line of %d bytes", AllocatePrefetchDistance, AllocatePrefetchStepSize);
twisti@1020 631 }
twisti@1020 632 }
twisti@1020 633
twisti@1020 634 if (PrefetchCopyIntervalInBytes > 0) {
twisti@1020 635 tty->print_cr("PrefetchCopyIntervalInBytes %d", PrefetchCopyIntervalInBytes);
twisti@1020 636 }
twisti@1020 637 if (PrefetchScanIntervalInBytes > 0) {
twisti@1020 638 tty->print_cr("PrefetchScanIntervalInBytes %d", PrefetchScanIntervalInBytes);
twisti@1020 639 }
twisti@1020 640 if (PrefetchFieldsAhead > 0) {
twisti@1020 641 tty->print_cr("PrefetchFieldsAhead %d", PrefetchFieldsAhead);
twisti@1020 642 }
twisti@1020 643 }
twisti@1020 644 #endif // !PRODUCT
twisti@1020 645 }
twisti@1020 646
twisti@1020 647 void VM_Version::initialize() {
twisti@1020 648 ResourceMark rm;
twisti@1020 649 // Making this stub must be FIRST use of assembler
twisti@1020 650
twisti@1020 651 stub_blob = BufferBlob::create("getPsrInfo_stub", stub_size);
twisti@1020 652 if (stub_blob == NULL) {
twisti@1020 653 vm_exit_during_initialization("Unable to allocate getPsrInfo_stub");
twisti@1020 654 }
twisti@2103 655 CodeBuffer c(stub_blob);
twisti@1020 656 VM_Version_StubGenerator g(&c);
twisti@1020 657 getPsrInfo_stub = CAST_TO_FN_PTR(getPsrInfo_stub_t,
twisti@1020 658 g.generate_getPsrInfo());
twisti@1020 659
twisti@1020 660 get_processor_features();
twisti@1020 661 }

mercurial