src/cpu/ppc/vm/vm_version_ppc.cpp

Tue, 18 Jun 2019 09:33:34 -0400

author
ogatak
date
Tue, 18 Jun 2019 09:33:34 -0400
changeset 9713
c4567d28f31f
parent 9684
69f33959c27f
child 9730
42118db355f5
permissions
-rw-r--r--

8185979: PPC64: Implement SHA2 intrinsic
Reviewed-by: mdoerr, goetz
Contributed-by: Bruno Rosa <bruno.rosa@eldorado.org.br>, Gustavo Serra Scalet <gustavo.scalet@eldorado.org.br>, Igor Nunes <igor.nunes@eldorado.org.br>, Martin Doerr <martin.doerr@sap.com>

goetz@6458 1 /*
gromero@9662 2 * Copyright (c) 1997, 2018, Oracle and/or its affiliates. All rights reserved.
gromero@9662 3 * Copyright 2012, 2018 SAP AG. All rights reserved.
goetz@6458 4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
goetz@6458 5 *
goetz@6458 6 * This code is free software; you can redistribute it and/or modify it
goetz@6458 7 * under the terms of the GNU General Public License version 2 only, as
goetz@6458 8 * published by the Free Software Foundation.
goetz@6458 9 *
goetz@6458 10 * This code is distributed in the hope that it will be useful, but WITHOUT
goetz@6458 11 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
goetz@6458 12 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
goetz@6458 13 * version 2 for more details (a copy is included in the LICENSE file that
goetz@6458 14 * accompanied this code).
goetz@6458 15 *
goetz@6458 16 * You should have received a copy of the GNU General Public License version
goetz@6458 17 * 2 along with this work; if not, write to the Free Software Foundation,
goetz@6458 18 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
goetz@6458 19 *
goetz@6458 20 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
goetz@6458 21 * or visit www.oracle.com if you need additional information or have any
goetz@6458 22 * questions.
goetz@6458 23 *
goetz@6458 24 */
goetz@6458 25
goetz@6458 26 #include "precompiled.hpp"
goetz@6511 27 #include "asm/assembler.inline.hpp"
goetz@6511 28 #include "asm/macroAssembler.inline.hpp"
goetz@6458 29 #include "compiler/disassembler.hpp"
goetz@6458 30 #include "memory/resourceArea.hpp"
goetz@6458 31 #include "runtime/java.hpp"
goetz@6458 32 #include "runtime/stubCodeGenerator.hpp"
goetz@6458 33 #include "utilities/defaultStream.hpp"
goetz@6458 34 #include "vm_version_ppc.hpp"
goetz@6458 35 #ifdef TARGET_OS_FAMILY_aix
goetz@6458 36 # include "os_aix.inline.hpp"
goetz@6458 37 #endif
goetz@6458 38 #ifdef TARGET_OS_FAMILY_linux
goetz@6458 39 # include "os_linux.inline.hpp"
goetz@6458 40 #endif
goetz@6458 41
goetz@6458 42 # include <sys/sysinfo.h>
goetz@6458 43
goetz@6458 44 int VM_Version::_features = VM_Version::unknown_m;
goetz@6458 45 int VM_Version::_measured_cache_line_size = 128; // default value
goetz@6458 46 const char* VM_Version::_features_str = "";
goetz@6458 47 bool VM_Version::_is_determine_features_test_running = false;
gromero@9662 48 uint64_t VM_Version::_dscr_val = 0;
goetz@6458 49
goetz@6458 50 #define MSG(flag) \
goetz@6458 51 if (flag && !FLAG_IS_DEFAULT(flag)) \
goetz@6458 52 jio_fprintf(defaultStream::error_stream(), \
goetz@6458 53 "warning: -XX:+" #flag " requires -XX:+UseSIGTRAP\n" \
goetz@6458 54 " -XX:+" #flag " will be disabled!\n");
goetz@6458 55
goetz@6458 56 void VM_Version::initialize() {
goetz@6458 57
goetz@6458 58 // Test which instructions are supported and measure cache line size.
goetz@6458 59 determine_features();
goetz@6458 60
goetz@6458 61 // If PowerArchitecturePPC64 hasn't been specified explicitly determine from features.
goetz@6458 62 if (FLAG_IS_DEFAULT(PowerArchitecturePPC64)) {
gromero@9662 63 if (VM_Version::has_lqarx()) {
gromero@9662 64 FLAG_SET_ERGO(uintx, PowerArchitecturePPC64, 8);
gromero@9662 65 } else if (VM_Version::has_popcntw()) {
goetz@6458 66 FLAG_SET_ERGO(uintx, PowerArchitecturePPC64, 7);
goetz@6458 67 } else if (VM_Version::has_cmpb()) {
goetz@6458 68 FLAG_SET_ERGO(uintx, PowerArchitecturePPC64, 6);
goetz@6458 69 } else if (VM_Version::has_popcntb()) {
goetz@6458 70 FLAG_SET_ERGO(uintx, PowerArchitecturePPC64, 5);
goetz@6458 71 } else {
goetz@6458 72 FLAG_SET_ERGO(uintx, PowerArchitecturePPC64, 0);
goetz@6458 73 }
goetz@6458 74 }
goetz@6458 75 guarantee(PowerArchitecturePPC64 == 0 || PowerArchitecturePPC64 == 5 ||
gromero@9662 76 PowerArchitecturePPC64 == 6 || PowerArchitecturePPC64 == 7 ||
gromero@9662 77 PowerArchitecturePPC64 == 8,
gromero@9662 78 "PowerArchitecturePPC64 should be 0, 5, 6, 7, or 8");
gromero@9662 79
gromero@9662 80 // Power 8: Configure Data Stream Control Register.
gromero@9662 81 if (PowerArchitecturePPC64 >= 8) {
gromero@9662 82 config_dscr();
gromero@9662 83 }
goetz@6458 84
goetz@6458 85 if (!UseSIGTRAP) {
goetz@6458 86 MSG(TrapBasedICMissChecks);
goetz@6458 87 MSG(TrapBasedNotEntrantChecks);
goetz@6458 88 MSG(TrapBasedNullChecks);
goetz@6458 89 FLAG_SET_ERGO(bool, TrapBasedNotEntrantChecks, false);
goetz@6458 90 FLAG_SET_ERGO(bool, TrapBasedNullChecks, false);
goetz@6458 91 FLAG_SET_ERGO(bool, TrapBasedICMissChecks, false);
goetz@6458 92 }
goetz@6458 93
goetz@6458 94 #ifdef COMPILER2
goetz@6490 95 if (!UseSIGTRAP) {
goetz@6490 96 MSG(TrapBasedRangeChecks);
goetz@6490 97 FLAG_SET_ERGO(bool, TrapBasedRangeChecks, false);
goetz@6490 98 }
goetz@6490 99
goetz@6458 100 // On Power6 test for section size.
goetz@6495 101 if (PowerArchitecturePPC64 == 6) {
goetz@6458 102 determine_section_size();
goetz@6495 103 // TODO: PPC port } else {
goetz@6458 104 // TODO: PPC port PdScheduling::power6SectorSize = 0x20;
goetz@6495 105 }
goetz@6458 106
goetz@6458 107 MaxVectorSize = 8;
goetz@6458 108 #endif
goetz@6458 109
goetz@6458 110 // Create and print feature-string.
goetz@6495 111 char buf[(num_features+1) * 16]; // Max 16 chars per feature.
goetz@6458 112 jio_snprintf(buf, sizeof(buf),
ogatak@9713 113 "ppc64%s%s%s%s%s%s%s%s%s%s%s%s%s%s",
goetz@6458 114 (has_fsqrt() ? " fsqrt" : ""),
goetz@6458 115 (has_isel() ? " isel" : ""),
goetz@6458 116 (has_lxarxeh() ? " lxarxeh" : ""),
goetz@6458 117 (has_cmpb() ? " cmpb" : ""),
goetz@6458 118 //(has_mftgpr()? " mftgpr" : ""),
goetz@6458 119 (has_popcntb() ? " popcntb" : ""),
goetz@6458 120 (has_popcntw() ? " popcntw" : ""),
goetz@6458 121 (has_fcfids() ? " fcfids" : ""),
simonis@8608 122 (has_vand() ? " vand" : ""),
gromero@9662 123 (has_lqarx() ? " lqarx" : ""),
mdoerr@9497 124 (has_vcipher() ? " aes" : ""),
gromero@9662 125 (has_vpmsumb() ? " vpmsumb" : ""),
gromero@9662 126 (has_mfdscr() ? " mfdscr" : ""),
ogatak@9713 127 (has_vsx() ? " vsx" : ""),
ogatak@9713 128 (has_vshasig() ? " sha" : "")
goetz@6458 129 // Make sure number of %s matches num_features!
goetz@6458 130 );
goetz@6458 131 _features_str = strdup(buf);
gromero@9662 132 if (Verbose) {
gromero@9662 133 print_features();
gromero@9662 134 }
goetz@6458 135
goetz@6458 136 // PPC64 supports 8-byte compare-exchange operations (see
goetz@6458 137 // Atomic::cmpxchg and StubGenerator::generate_atomic_cmpxchg_ptr)
goetz@6458 138 // and 'atomic long memory ops' (see Unsafe_GetLongVolatile).
goetz@6458 139 _supports_cx8 = true;
goetz@6458 140
goetz@6458 141 UseSSE = 0; // Only on x86 and x64
goetz@6458 142
goetz@6458 143 intx cache_line_size = _measured_cache_line_size;
goetz@6458 144
goetz@6458 145 if (FLAG_IS_DEFAULT(AllocatePrefetchStyle)) AllocatePrefetchStyle = 1;
goetz@6458 146
goetz@6458 147 if (AllocatePrefetchStyle == 4) {
goetz@6495 148 AllocatePrefetchStepSize = cache_line_size; // Need exact value.
goetz@6495 149 if (FLAG_IS_DEFAULT(AllocatePrefetchLines)) AllocatePrefetchLines = 12; // Use larger blocks by default.
goetz@6495 150 if (AllocatePrefetchDistance < 0) AllocatePrefetchDistance = 2*cache_line_size; // Default is not defined?
goetz@6458 151 } else {
goetz@6458 152 if (cache_line_size > AllocatePrefetchStepSize) AllocatePrefetchStepSize = cache_line_size;
goetz@6495 153 if (FLAG_IS_DEFAULT(AllocatePrefetchLines)) AllocatePrefetchLines = 3; // Optimistic value.
goetz@6495 154 if (AllocatePrefetchDistance < 0) AllocatePrefetchDistance = 3*cache_line_size; // Default is not defined?
goetz@6458 155 }
goetz@6458 156
goetz@6458 157 assert(AllocatePrefetchLines > 0, "invalid value");
goetz@7424 158 if (AllocatePrefetchLines < 1) { // Set valid value in product VM.
goetz@6495 159 AllocatePrefetchLines = 1; // Conservative value.
goetz@7424 160 }
goetz@6458 161
goetz@7424 162 if (AllocatePrefetchStyle == 3 && AllocatePrefetchDistance < cache_line_size) {
goetz@6495 163 AllocatePrefetchStyle = 1; // Fall back if inappropriate.
goetz@7424 164 }
goetz@6458 165
goetz@6458 166 assert(AllocatePrefetchStyle >= 0, "AllocatePrefetchStyle should be positive");
goetz@7424 167
gromero@9496 168 // Implementation does not use any of the vector instructions
gromero@9496 169 // available with Power8. Their exploitation is still pending.
gromero@9496 170 if (!UseCRC32Intrinsics) {
gromero@9496 171 if (FLAG_IS_DEFAULT(UseCRC32Intrinsics)) {
gromero@9496 172 FLAG_SET_DEFAULT(UseCRC32Intrinsics, true);
gromero@9496 173 }
goetz@7424 174 }
goetz@7424 175
goetz@7424 176 // The AES intrinsic stubs require AES instruction support.
simonis@8608 177 #if defined(VM_LITTLE_ENDIAN)
simonis@8608 178 if (has_vcipher()) {
simonis@8608 179 if (FLAG_IS_DEFAULT(UseAES)) {
simonis@8608 180 UseAES = true;
simonis@8608 181 }
simonis@8608 182 } else if (UseAES) {
simonis@8608 183 if (!FLAG_IS_DEFAULT(UseAES))
simonis@8608 184 warning("AES instructions are not available on this CPU");
simonis@8608 185 FLAG_SET_DEFAULT(UseAES, false);
simonis@8608 186 }
simonis@8608 187
simonis@8608 188 if (UseAES && has_vcipher()) {
simonis@8608 189 if (FLAG_IS_DEFAULT(UseAESIntrinsics)) {
simonis@8608 190 UseAESIntrinsics = true;
simonis@8608 191 }
simonis@8608 192 } else if (UseAESIntrinsics) {
simonis@8608 193 if (!FLAG_IS_DEFAULT(UseAESIntrinsics))
simonis@8608 194 warning("AES intrinsics are not available on this CPU");
simonis@8608 195 FLAG_SET_DEFAULT(UseAESIntrinsics, false);
simonis@8608 196 }
simonis@8608 197
simonis@8608 198 #else
goetz@7424 199 if (UseAES) {
goetz@7424 200 warning("AES instructions are not available on this CPU");
goetz@7424 201 FLAG_SET_DEFAULT(UseAES, false);
goetz@7424 202 }
goetz@7424 203 if (UseAESIntrinsics) {
goetz@7424 204 if (!FLAG_IS_DEFAULT(UseAESIntrinsics))
goetz@7424 205 warning("AES intrinsics are not available on this CPU");
goetz@7424 206 FLAG_SET_DEFAULT(UseAESIntrinsics, false);
goetz@7424 207 }
simonis@8608 208 #endif
goetz@7424 209
ogatak@9713 210 if (has_vshasig()) {
ogatak@9713 211 if (FLAG_IS_DEFAULT(UseSHA)) {
ogatak@9713 212 UseSHA = true;
ogatak@9713 213 }
ogatak@9713 214 } else if (UseSHA) {
ogatak@9713 215 if (!FLAG_IS_DEFAULT(UseSHA))
ogatak@9713 216 warning("SHA instructions are not available on this CPU");
goetz@7424 217 FLAG_SET_DEFAULT(UseSHA, false);
goetz@7424 218 }
ogatak@9713 219
ogatak@9713 220 if (UseSHA1Intrinsics) {
ogatak@9713 221 warning("Intrinsics for SHA-1 crypto hash functions not available on this CPU.");
goetz@7424 222 FLAG_SET_DEFAULT(UseSHA1Intrinsics, false);
ogatak@9713 223 }
ogatak@9713 224
ogatak@9713 225 if (UseSHA && has_vshasig()) {
ogatak@9713 226 if (FLAG_IS_DEFAULT(UseSHA256Intrinsics)) {
ogatak@9713 227 FLAG_SET_DEFAULT(UseSHA256Intrinsics, true);
ogatak@9713 228 }
ogatak@9713 229 } else if (UseSHA256Intrinsics) {
ogatak@9713 230 warning("Intrinsics for SHA-224 and SHA-256 crypto hash functions not available on this CPU.");
goetz@7424 231 FLAG_SET_DEFAULT(UseSHA256Intrinsics, false);
ogatak@9713 232 }
ogatak@9713 233
ogatak@9713 234 if (UseSHA && has_vshasig()) {
ogatak@9713 235 if (FLAG_IS_DEFAULT(UseSHA512Intrinsics)) {
ogatak@9713 236 FLAG_SET_DEFAULT(UseSHA512Intrinsics, true);
ogatak@9713 237 }
ogatak@9713 238 } else if (UseSHA512Intrinsics) {
ogatak@9713 239 warning("Intrinsics for SHA-384 and SHA-512 crypto hash functions not available on this CPU.");
goetz@7424 240 FLAG_SET_DEFAULT(UseSHA512Intrinsics, false);
goetz@7424 241 }
goetz@7424 242
ogatak@9713 243 if (!(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics)) {
ogatak@9713 244 FLAG_SET_DEFAULT(UseSHA, false);
ogatak@9713 245 }
ogatak@9713 246
mdoerr@8903 247 if (FLAG_IS_DEFAULT(UseMontgomeryMultiplyIntrinsic)) {
mdoerr@8903 248 UseMontgomeryMultiplyIntrinsic = true;
mdoerr@8903 249 }
mdoerr@8903 250 if (FLAG_IS_DEFAULT(UseMontgomerySquareIntrinsic)) {
mdoerr@8903 251 UseMontgomerySquareIntrinsic = true;
mdoerr@8903 252 }
goetz@6458 253 }
goetz@6458 254
goetz@6458 255 void VM_Version::print_features() {
drchase@6680 256 tty->print_cr("Version: %s cache_line_size = %d", cpu_features(), (int) get_cache_line_size());
goetz@6458 257 }
goetz@6458 258
goetz@6458 259 #ifdef COMPILER2
goetz@6458 260 // Determine section size on power6: If section size is 8 instructions,
goetz@6458 261 // there should be a difference between the two testloops of ~15 %. If
goetz@6458 262 // no difference is detected the section is assumed to be 32 instructions.
goetz@6458 263 void VM_Version::determine_section_size() {
goetz@6458 264
goetz@6458 265 int unroll = 80;
goetz@6458 266
goetz@6458 267 const int code_size = (2* unroll * 32 + 100)*BytesPerInstWord;
goetz@6458 268
goetz@6495 269 // Allocate space for the code.
goetz@6458 270 ResourceMark rm;
goetz@6458 271 CodeBuffer cb("detect_section_size", code_size, 0);
goetz@6458 272 MacroAssembler* a = new MacroAssembler(&cb);
goetz@6458 273
goetz@6458 274 uint32_t *code = (uint32_t *)a->pc();
goetz@6495 275 // Emit code.
goetz@6511 276 void (*test1)() = (void(*)())(void *)a->function_entry();
goetz@6458 277
goetz@6458 278 Label l1;
goetz@6458 279
goetz@6458 280 a->li(R4, 1);
goetz@6458 281 a->sldi(R4, R4, 28);
goetz@6458 282 a->b(l1);
goetz@6458 283 a->align(CodeEntryAlignment);
goetz@6458 284
goetz@6458 285 a->bind(l1);
goetz@6458 286
goetz@6458 287 for (int i = 0; i < unroll; i++) {
goetz@6458 288 // Schleife 1
goetz@6458 289 // ------- sector 0 ------------
goetz@6458 290 // ;; 0
goetz@6458 291 a->nop(); // 1
goetz@6458 292 a->fpnop0(); // 2
goetz@6458 293 a->fpnop1(); // 3
goetz@6458 294 a->addi(R4,R4, -1); // 4
goetz@6458 295
goetz@6458 296 // ;; 1
goetz@6458 297 a->nop(); // 5
goetz@6495 298 a->fmr(F6, F6); // 6
goetz@6495 299 a->fmr(F7, F7); // 7
goetz@6458 300 a->endgroup(); // 8
goetz@6458 301 // ------- sector 8 ------------
goetz@6458 302
goetz@6458 303 // ;; 2
goetz@6458 304 a->nop(); // 9
goetz@6458 305 a->nop(); // 10
goetz@6495 306 a->fmr(F8, F8); // 11
goetz@6495 307 a->fmr(F9, F9); // 12
goetz@6458 308
goetz@6458 309 // ;; 3
goetz@6458 310 a->nop(); // 13
goetz@6495 311 a->fmr(F10, F10); // 14
goetz@6495 312 a->fmr(F11, F11); // 15
goetz@6458 313 a->endgroup(); // 16
goetz@6458 314 // -------- sector 16 -------------
goetz@6458 315
goetz@6458 316 // ;; 4
goetz@6458 317 a->nop(); // 17
goetz@6458 318 a->nop(); // 18
goetz@6495 319 a->fmr(F15, F15); // 19
goetz@6495 320 a->fmr(F16, F16); // 20
goetz@6458 321
goetz@6458 322 // ;; 5
goetz@6458 323 a->nop(); // 21
goetz@6495 324 a->fmr(F17, F17); // 22
goetz@6495 325 a->fmr(F18, F18); // 23
goetz@6458 326 a->endgroup(); // 24
goetz@6458 327 // ------- sector 24 ------------
goetz@6458 328
goetz@6458 329 // ;; 6
goetz@6458 330 a->nop(); // 25
goetz@6458 331 a->nop(); // 26
goetz@6495 332 a->fmr(F19, F19); // 27
goetz@6495 333 a->fmr(F20, F20); // 28
goetz@6458 334
goetz@6458 335 // ;; 7
goetz@6458 336 a->nop(); // 29
goetz@6495 337 a->fmr(F21, F21); // 30
goetz@6495 338 a->fmr(F22, F22); // 31
goetz@6458 339 a->brnop0(); // 32
goetz@6458 340
goetz@6458 341 // ------- sector 32 ------------
goetz@6458 342 }
goetz@6458 343
goetz@6458 344 // ;; 8
goetz@6495 345 a->cmpdi(CCR0, R4, unroll); // 33
goetz@6495 346 a->bge(CCR0, l1); // 34
goetz@6458 347 a->blr();
goetz@6458 348
goetz@6495 349 // Emit code.
goetz@6511 350 void (*test2)() = (void(*)())(void *)a->function_entry();
goetz@6458 351 // uint32_t *code = (uint32_t *)a->pc();
goetz@6458 352
goetz@6458 353 Label l2;
goetz@6458 354
goetz@6458 355 a->li(R4, 1);
goetz@6458 356 a->sldi(R4, R4, 28);
goetz@6458 357 a->b(l2);
goetz@6458 358 a->align(CodeEntryAlignment);
goetz@6458 359
goetz@6458 360 a->bind(l2);
goetz@6458 361
goetz@6458 362 for (int i = 0; i < unroll; i++) {
goetz@6458 363 // Schleife 2
goetz@6458 364 // ------- sector 0 ------------
goetz@6458 365 // ;; 0
goetz@6458 366 a->brnop0(); // 1
goetz@6458 367 a->nop(); // 2
goetz@6458 368 //a->cmpdi(CCR0, R4, unroll);
goetz@6458 369 a->fpnop0(); // 3
goetz@6458 370 a->fpnop1(); // 4
goetz@6458 371 a->addi(R4,R4, -1); // 5
goetz@6458 372
goetz@6458 373 // ;; 1
goetz@6458 374
goetz@6458 375 a->nop(); // 6
goetz@6458 376 a->fmr(F6, F6); // 7
goetz@6458 377 a->fmr(F7, F7); // 8
goetz@6458 378 // ------- sector 8 ---------------
goetz@6458 379
goetz@6458 380 // ;; 2
goetz@6458 381 a->endgroup(); // 9
goetz@6458 382
goetz@6458 383 // ;; 3
goetz@6458 384 a->nop(); // 10
goetz@6458 385 a->nop(); // 11
goetz@6458 386 a->fmr(F8, F8); // 12
goetz@6458 387
goetz@6458 388 // ;; 4
goetz@6458 389 a->fmr(F9, F9); // 13
goetz@6458 390 a->nop(); // 14
goetz@6458 391 a->fmr(F10, F10); // 15
goetz@6458 392
goetz@6458 393 // ;; 5
goetz@6458 394 a->fmr(F11, F11); // 16
goetz@6458 395 // -------- sector 16 -------------
goetz@6458 396
goetz@6458 397 // ;; 6
goetz@6458 398 a->endgroup(); // 17
goetz@6458 399
goetz@6458 400 // ;; 7
goetz@6458 401 a->nop(); // 18
goetz@6458 402 a->nop(); // 19
goetz@6458 403 a->fmr(F15, F15); // 20
goetz@6458 404
goetz@6458 405 // ;; 8
goetz@6458 406 a->fmr(F16, F16); // 21
goetz@6458 407 a->nop(); // 22
goetz@6458 408 a->fmr(F17, F17); // 23
goetz@6458 409
goetz@6458 410 // ;; 9
goetz@6458 411 a->fmr(F18, F18); // 24
goetz@6458 412 // -------- sector 24 -------------
goetz@6458 413
goetz@6458 414 // ;; 10
goetz@6458 415 a->endgroup(); // 25
goetz@6458 416
goetz@6458 417 // ;; 11
goetz@6458 418 a->nop(); // 26
goetz@6458 419 a->nop(); // 27
goetz@6458 420 a->fmr(F19, F19); // 28
goetz@6458 421
goetz@6458 422 // ;; 12
goetz@6458 423 a->fmr(F20, F20); // 29
goetz@6458 424 a->nop(); // 30
goetz@6458 425 a->fmr(F21, F21); // 31
goetz@6458 426
goetz@6458 427 // ;; 13
goetz@6458 428 a->fmr(F22, F22); // 32
goetz@6458 429 }
goetz@6458 430
goetz@6458 431 // -------- sector 32 -------------
goetz@6458 432 // ;; 14
goetz@6458 433 a->cmpdi(CCR0, R4, unroll); // 33
goetz@6458 434 a->bge(CCR0, l2); // 34
goetz@6458 435
goetz@6458 436 a->blr();
goetz@6458 437 uint32_t *code_end = (uint32_t *)a->pc();
goetz@6458 438 a->flush();
goetz@6458 439
goetz@6458 440 double loop1_seconds,loop2_seconds, rel_diff;
goetz@6458 441 uint64_t start1, stop1;
goetz@6458 442
goetz@6458 443 start1 = os::current_thread_cpu_time(false);
goetz@6458 444 (*test1)();
goetz@6458 445 stop1 = os::current_thread_cpu_time(false);
goetz@6458 446 loop1_seconds = (stop1- start1) / (1000 *1000 *1000.0);
goetz@6458 447
goetz@6458 448
goetz@6458 449 start1 = os::current_thread_cpu_time(false);
goetz@6458 450 (*test2)();
goetz@6458 451 stop1 = os::current_thread_cpu_time(false);
goetz@6458 452
goetz@6458 453 loop2_seconds = (stop1 - start1) / (1000 *1000 *1000.0);
goetz@6458 454
goetz@6458 455 rel_diff = (loop2_seconds - loop1_seconds) / loop1_seconds *100;
goetz@6458 456
goetz@6458 457 if (PrintAssembly) {
goetz@6458 458 ttyLocker ttyl;
coleenp@7358 459 tty->print_cr("Decoding section size detection stub at " INTPTR_FORMAT " before execution:", p2i(code));
goetz@6458 460 Disassembler::decode((u_char*)code, (u_char*)code_end, tty);
goetz@6458 461 tty->print_cr("Time loop1 :%f", loop1_seconds);
goetz@6458 462 tty->print_cr("Time loop2 :%f", loop2_seconds);
goetz@6458 463 tty->print_cr("(time2 - time1) / time1 = %f %%", rel_diff);
goetz@6458 464
goetz@6458 465 if (rel_diff > 12.0) {
goetz@6458 466 tty->print_cr("Section Size 8 Instructions");
goetz@6458 467 } else{
goetz@6458 468 tty->print_cr("Section Size 32 Instructions or Power5");
goetz@6458 469 }
goetz@6458 470 }
goetz@6458 471
goetz@6458 472 #if 0 // TODO: PPC port
goetz@6458 473 // Set sector size (if not set explicitly).
goetz@6458 474 if (FLAG_IS_DEFAULT(Power6SectorSize128PPC64)) {
goetz@6458 475 if (rel_diff > 12.0) {
goetz@6458 476 PdScheduling::power6SectorSize = 0x20;
goetz@6458 477 } else {
goetz@6458 478 PdScheduling::power6SectorSize = 0x80;
goetz@6458 479 }
goetz@6458 480 } else if (Power6SectorSize128PPC64) {
goetz@6458 481 PdScheduling::power6SectorSize = 0x80;
goetz@6458 482 } else {
goetz@6458 483 PdScheduling::power6SectorSize = 0x20;
goetz@6458 484 }
goetz@6458 485 #endif
goetz@6458 486 if (UsePower6SchedulerPPC64) Unimplemented();
goetz@6458 487 }
goetz@6458 488 #endif // COMPILER2
goetz@6458 489
goetz@6458 490 void VM_Version::determine_features() {
goetz@6511 491 #if defined(ABI_ELFv2)
goetz@6511 492 const int code_size = (num_features+1+2*7)*BytesPerInstWord; // TODO(asmundak): calculation is incorrect.
goetz@6511 493 #else
goetz@6495 494 // 7 InstWords for each call (function descriptor + blr instruction).
goetz@6495 495 const int code_size = (num_features+1+2*7)*BytesPerInstWord;
goetz@6511 496 #endif
goetz@6458 497 int features = 0;
goetz@6458 498
goetz@6458 499 // create test area
goetz@6495 500 enum { BUFFER_SIZE = 2*4*K }; // Needs to be >=2* max cache line size (cache line size can't exceed min page size).
goetz@6458 501 char test_area[BUFFER_SIZE];
goetz@6458 502 char *mid_of_test_area = &test_area[BUFFER_SIZE>>1];
goetz@6458 503
goetz@6495 504 // Allocate space for the code.
goetz@6458 505 ResourceMark rm;
goetz@6458 506 CodeBuffer cb("detect_cpu_features", code_size, 0);
goetz@6458 507 MacroAssembler* a = new MacroAssembler(&cb);
goetz@6458 508
goetz@6515 509 // Must be set to true so we can generate the test code.
goetz@6515 510 _features = VM_Version::all_features_m;
goetz@6515 511
goetz@6495 512 // Emit code.
goetz@6511 513 void (*test)(address addr, uint64_t offset)=(void(*)(address addr, uint64_t offset))(void *)a->function_entry();
goetz@6458 514 uint32_t *code = (uint32_t *)a->pc();
goetz@6458 515 // Don't use R0 in ldarx.
goetz@6495 516 // Keep R3_ARG1 unmodified, it contains &field (see below).
goetz@6495 517 // Keep R4_ARG2 unmodified, it contains offset = 0 (see below).
goetz@6495 518 a->fsqrt(F3, F4); // code[0] -> fsqrt_m
goetz@6515 519 a->fsqrts(F3, F4); // code[1] -> fsqrts_m
goetz@6515 520 a->isel(R7, R5, R6, 0); // code[2] -> isel_m
goetz@6515 521 a->ldarx_unchecked(R7, R3_ARG1, R4_ARG2, 1); // code[3] -> lxarx_m
goetz@6515 522 a->cmpb(R7, R5, R6); // code[4] -> bcmp
goetz@6515 523 //a->mftgpr(R7, F3); // code[5] -> mftgpr
goetz@6515 524 a->popcntb(R7, R5); // code[6] -> popcntb
goetz@6515 525 a->popcntw(R7, R5); // code[7] -> popcntw
goetz@6515 526 a->fcfids(F3, F4); // code[8] -> fcfids
goetz@6515 527 a->vand(VR0, VR0, VR0); // code[9] -> vand
gromero@9662 528 a->lqarx_unchecked(R7, R3_ARG1, R4_ARG2, 1); // code[10] -> lqarx_m
gromero@9662 529 a->vcipher(VR0, VR1, VR2); // code[11] -> vcipher
gromero@9662 530 a->vpmsumb(VR0, VR1, VR2); // code[12] -> vpmsumb
gromero@9662 531 a->mfdscr(R0); // code[13] -> mfdscr
gromero@9684 532 a->lxvd2x(VSR0, R3_ARG1); // code[14] -> vsx
ogatak@9713 533 a->vshasigmaw(VR0, VR1, 1, 0xF); // code[15] -> vshasig
goetz@6458 534 a->blr();
goetz@6458 535
goetz@6495 536 // Emit function to set one cache line to zero. Emit function descriptor and get pointer to it.
goetz@6511 537 void (*zero_cacheline_func_ptr)(char*) = (void(*)(char*))(void *)a->function_entry();
goetz@6495 538 a->dcbz(R3_ARG1); // R3_ARG1 = addr
goetz@6458 539 a->blr();
goetz@6458 540
goetz@6458 541 uint32_t *code_end = (uint32_t *)a->pc();
goetz@6458 542 a->flush();
goetz@6515 543 _features = VM_Version::unknown_m;
goetz@6458 544
goetz@6458 545 // Print the detection code.
goetz@6458 546 if (PrintAssembly) {
goetz@6458 547 ttyLocker ttyl;
coleenp@7358 548 tty->print_cr("Decoding cpu-feature detection stub at " INTPTR_FORMAT " before execution:", p2i(code));
goetz@6458 549 Disassembler::decode((u_char*)code, (u_char*)code_end, tty);
goetz@6458 550 }
goetz@6458 551
goetz@6458 552 // Measure cache line size.
goetz@6495 553 memset(test_area, 0xFF, BUFFER_SIZE); // Fill test area with 0xFF.
goetz@6495 554 (*zero_cacheline_func_ptr)(mid_of_test_area); // Call function which executes dcbz to the middle.
goetz@6458 555 int count = 0; // count zeroed bytes
goetz@6458 556 for (int i = 0; i < BUFFER_SIZE; i++) if (test_area[i] == 0) count++;
goetz@6458 557 guarantee(is_power_of_2(count), "cache line size needs to be a power of 2");
goetz@6458 558 _measured_cache_line_size = count;
goetz@6458 559
goetz@6458 560 // Execute code. Illegal instructions will be replaced by 0 in the signal handler.
goetz@6458 561 VM_Version::_is_determine_features_test_running = true;
goetz@6458 562 (*test)((address)mid_of_test_area, (uint64_t)0);
goetz@6458 563 VM_Version::_is_determine_features_test_running = false;
goetz@6458 564
goetz@6458 565 // determine which instructions are legal.
goetz@6458 566 int feature_cntr = 0;
goetz@6458 567 if (code[feature_cntr++]) features |= fsqrt_m;
goetz@6515 568 if (code[feature_cntr++]) features |= fsqrts_m;
goetz@6458 569 if (code[feature_cntr++]) features |= isel_m;
goetz@6458 570 if (code[feature_cntr++]) features |= lxarxeh_m;
goetz@6458 571 if (code[feature_cntr++]) features |= cmpb_m;
goetz@6458 572 //if(code[feature_cntr++])features |= mftgpr_m;
goetz@6458 573 if (code[feature_cntr++]) features |= popcntb_m;
goetz@6458 574 if (code[feature_cntr++]) features |= popcntw_m;
goetz@6458 575 if (code[feature_cntr++]) features |= fcfids_m;
goetz@6458 576 if (code[feature_cntr++]) features |= vand_m;
gromero@9662 577 if (code[feature_cntr++]) features |= lqarx_m;
simonis@8608 578 if (code[feature_cntr++]) features |= vcipher_m;
mdoerr@9497 579 if (code[feature_cntr++]) features |= vpmsumb_m;
gromero@9662 580 if (code[feature_cntr++]) features |= mfdscr_m;
gromero@9662 581 if (code[feature_cntr++]) features |= vsx_m;
ogatak@9713 582 if (code[feature_cntr++]) features |= vshasig_m;
goetz@6458 583
goetz@6458 584 // Print the detection code.
goetz@6458 585 if (PrintAssembly) {
goetz@6458 586 ttyLocker ttyl;
coleenp@7358 587 tty->print_cr("Decoding cpu-feature detection stub at " INTPTR_FORMAT " after execution:", p2i(code));
goetz@6458 588 Disassembler::decode((u_char*)code, (u_char*)code_end, tty);
goetz@6458 589 }
goetz@6458 590
goetz@6458 591 _features = features;
goetz@6458 592 }
goetz@6458 593
gromero@9662 594 // Power 8: Configure Data Stream Control Register.
gromero@9662 595 void VM_Version::config_dscr() {
gromero@9662 596 assert(has_lqarx(), "Only execute on Power 8 or later!");
gromero@9662 597
gromero@9662 598 // 7 InstWords for each call (function descriptor + blr instruction).
gromero@9662 599 const int code_size = (2+2*7)*BytesPerInstWord;
gromero@9662 600
gromero@9662 601 // Allocate space for the code.
gromero@9662 602 ResourceMark rm;
gromero@9662 603 CodeBuffer cb("config_dscr", code_size, 0);
gromero@9662 604 MacroAssembler* a = new MacroAssembler(&cb);
gromero@9662 605
gromero@9662 606 // Emit code.
gromero@9662 607 uint64_t (*get_dscr)() = (uint64_t(*)())(void *)a->function_entry();
gromero@9662 608 uint32_t *code = (uint32_t *)a->pc();
gromero@9662 609 a->mfdscr(R3);
gromero@9662 610 a->blr();
gromero@9662 611
gromero@9662 612 void (*set_dscr)(long) = (void(*)(long))(void *)a->function_entry();
gromero@9662 613 a->mtdscr(R3);
gromero@9662 614 a->blr();
gromero@9662 615
gromero@9662 616 uint32_t *code_end = (uint32_t *)a->pc();
gromero@9662 617 a->flush();
gromero@9662 618
gromero@9662 619 // Print the detection code.
gromero@9662 620 if (PrintAssembly) {
gromero@9662 621 ttyLocker ttyl;
gromero@9662 622 tty->print_cr("Decoding dscr configuration stub at " INTPTR_FORMAT " before execution:", p2i(code));
gromero@9662 623 Disassembler::decode((u_char*)code, (u_char*)code_end, tty);
gromero@9662 624 }
gromero@9662 625
gromero@9662 626 // Apply the configuration if needed.
gromero@9662 627 _dscr_val = (*get_dscr)();
gromero@9662 628 if (Verbose) {
gromero@9662 629 tty->print_cr("dscr value was 0x%lx" , _dscr_val);
gromero@9662 630 }
gromero@9662 631 bool change_requested = false;
gromero@9662 632 if (DSCR_PPC64 != (uintx)-1) {
gromero@9662 633 _dscr_val = DSCR_PPC64;
gromero@9662 634 change_requested = true;
gromero@9662 635 }
gromero@9662 636 if (DSCR_DPFD_PPC64 <= 7) {
gromero@9662 637 uint64_t mask = 0x7;
gromero@9662 638 if ((_dscr_val & mask) != DSCR_DPFD_PPC64) {
gromero@9662 639 _dscr_val = (_dscr_val & ~mask) | (DSCR_DPFD_PPC64);
gromero@9662 640 change_requested = true;
gromero@9662 641 }
gromero@9662 642 }
gromero@9662 643 if (DSCR_URG_PPC64 <= 7) {
gromero@9662 644 uint64_t mask = 0x7 << 6;
gromero@9662 645 if ((_dscr_val & mask) != DSCR_DPFD_PPC64 << 6) {
gromero@9662 646 _dscr_val = (_dscr_val & ~mask) | (DSCR_URG_PPC64 << 6);
gromero@9662 647 change_requested = true;
gromero@9662 648 }
gromero@9662 649 }
gromero@9662 650 if (change_requested) {
gromero@9662 651 (*set_dscr)(_dscr_val);
gromero@9662 652 if (Verbose) {
gromero@9662 653 tty->print_cr("dscr was set to 0x%lx" , (*get_dscr)());
gromero@9662 654 }
gromero@9662 655 }
gromero@9662 656 }
goetz@6458 657
goetz@6458 658 static int saved_features = 0;
goetz@6458 659
goetz@6458 660 void VM_Version::allow_all() {
goetz@6458 661 saved_features = _features;
goetz@6458 662 _features = all_features_m;
goetz@6458 663 }
goetz@6458 664
goetz@6458 665 void VM_Version::revert() {
goetz@6458 666 _features = saved_features;
goetz@6458 667 }

mercurial