1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/src/cpu/ppc/vm/vm_version_ppc.cpp Wed Apr 27 01:25:04 2016 +0800
1.3 @@ -0,0 +1,488 @@
1.4 +/*
1.5 + * Copyright (c) 1997, 2014, Oracle and/or its affiliates. All rights reserved.
1.6 + * Copyright 2012, 2014 SAP AG. All rights reserved.
1.7 + * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
1.8 + *
1.9 + * This code is free software; you can redistribute it and/or modify it
1.10 + * under the terms of the GNU General Public License version 2 only, as
1.11 + * published by the Free Software Foundation.
1.12 + *
1.13 + * This code is distributed in the hope that it will be useful, but WITHOUT
1.14 + * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
1.15 + * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
1.16 + * version 2 for more details (a copy is included in the LICENSE file that
1.17 + * accompanied this code).
1.18 + *
1.19 + * You should have received a copy of the GNU General Public License version
1.20 + * 2 along with this work; if not, write to the Free Software Foundation,
1.21 + * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
1.22 + *
1.23 + * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
1.24 + * or visit www.oracle.com if you need additional information or have any
1.25 + * questions.
1.26 + *
1.27 + */
1.28 +
1.29 +#include "precompiled.hpp"
1.30 +#include "asm/assembler.inline.hpp"
1.31 +#include "asm/macroAssembler.inline.hpp"
1.32 +#include "compiler/disassembler.hpp"
1.33 +#include "memory/resourceArea.hpp"
1.34 +#include "runtime/java.hpp"
1.35 +#include "runtime/stubCodeGenerator.hpp"
1.36 +#include "utilities/defaultStream.hpp"
1.37 +#include "vm_version_ppc.hpp"
1.38 +#ifdef TARGET_OS_FAMILY_aix
1.39 +# include "os_aix.inline.hpp"
1.40 +#endif
1.41 +#ifdef TARGET_OS_FAMILY_linux
1.42 +# include "os_linux.inline.hpp"
1.43 +#endif
1.44 +
1.45 +# include <sys/sysinfo.h>
1.46 +
1.47 +int VM_Version::_features = VM_Version::unknown_m;
1.48 +int VM_Version::_measured_cache_line_size = 128; // default value
1.49 +const char* VM_Version::_features_str = "";
1.50 +bool VM_Version::_is_determine_features_test_running = false;
1.51 +
1.52 +
1.53 +#define MSG(flag) \
1.54 + if (flag && !FLAG_IS_DEFAULT(flag)) \
1.55 + jio_fprintf(defaultStream::error_stream(), \
1.56 + "warning: -XX:+" #flag " requires -XX:+UseSIGTRAP\n" \
1.57 + " -XX:+" #flag " will be disabled!\n");
1.58 +
1.59 +void VM_Version::initialize() {
1.60 +
1.61 + // Test which instructions are supported and measure cache line size.
1.62 + determine_features();
1.63 +
1.64 + // If PowerArchitecturePPC64 hasn't been specified explicitly determine from features.
1.65 + if (FLAG_IS_DEFAULT(PowerArchitecturePPC64)) {
1.66 + if (VM_Version::has_popcntw()) {
1.67 + FLAG_SET_ERGO(uintx, PowerArchitecturePPC64, 7);
1.68 + } else if (VM_Version::has_cmpb()) {
1.69 + FLAG_SET_ERGO(uintx, PowerArchitecturePPC64, 6);
1.70 + } else if (VM_Version::has_popcntb()) {
1.71 + FLAG_SET_ERGO(uintx, PowerArchitecturePPC64, 5);
1.72 + } else {
1.73 + FLAG_SET_ERGO(uintx, PowerArchitecturePPC64, 0);
1.74 + }
1.75 + }
1.76 + guarantee(PowerArchitecturePPC64 == 0 || PowerArchitecturePPC64 == 5 ||
1.77 + PowerArchitecturePPC64 == 6 || PowerArchitecturePPC64 == 7,
1.78 + "PowerArchitecturePPC64 should be 0, 5, 6 or 7");
1.79 +
1.80 + if (!UseSIGTRAP) {
1.81 + MSG(TrapBasedICMissChecks);
1.82 + MSG(TrapBasedNotEntrantChecks);
1.83 + MSG(TrapBasedNullChecks);
1.84 + FLAG_SET_ERGO(bool, TrapBasedNotEntrantChecks, false);
1.85 + FLAG_SET_ERGO(bool, TrapBasedNullChecks, false);
1.86 + FLAG_SET_ERGO(bool, TrapBasedICMissChecks, false);
1.87 + }
1.88 +
1.89 +#ifdef COMPILER2
1.90 + if (!UseSIGTRAP) {
1.91 + MSG(TrapBasedRangeChecks);
1.92 + FLAG_SET_ERGO(bool, TrapBasedRangeChecks, false);
1.93 + }
1.94 +
1.95 + // On Power6 test for section size.
1.96 + if (PowerArchitecturePPC64 == 6) {
1.97 + determine_section_size();
1.98 + // TODO: PPC port } else {
1.99 + // TODO: PPC port PdScheduling::power6SectorSize = 0x20;
1.100 + }
1.101 +
1.102 + MaxVectorSize = 8;
1.103 +#endif
1.104 +
1.105 + // Create and print feature-string.
1.106 + char buf[(num_features+1) * 16]; // Max 16 chars per feature.
1.107 + jio_snprintf(buf, sizeof(buf),
1.108 + "ppc64%s%s%s%s%s%s%s%s",
1.109 + (has_fsqrt() ? " fsqrt" : ""),
1.110 + (has_isel() ? " isel" : ""),
1.111 + (has_lxarxeh() ? " lxarxeh" : ""),
1.112 + (has_cmpb() ? " cmpb" : ""),
1.113 + //(has_mftgpr()? " mftgpr" : ""),
1.114 + (has_popcntb() ? " popcntb" : ""),
1.115 + (has_popcntw() ? " popcntw" : ""),
1.116 + (has_fcfids() ? " fcfids" : ""),
1.117 + (has_vand() ? " vand" : "")
1.118 + // Make sure number of %s matches num_features!
1.119 + );
1.120 + _features_str = strdup(buf);
1.121 + NOT_PRODUCT(if (Verbose) print_features(););
1.122 +
1.123 + // PPC64 supports 8-byte compare-exchange operations (see
1.124 + // Atomic::cmpxchg and StubGenerator::generate_atomic_cmpxchg_ptr)
1.125 + // and 'atomic long memory ops' (see Unsafe_GetLongVolatile).
1.126 + _supports_cx8 = true;
1.127 +
1.128 + UseSSE = 0; // Only on x86 and x64
1.129 +
1.130 + intx cache_line_size = _measured_cache_line_size;
1.131 +
1.132 + if (FLAG_IS_DEFAULT(AllocatePrefetchStyle)) AllocatePrefetchStyle = 1;
1.133 +
1.134 + if (AllocatePrefetchStyle == 4) {
1.135 + AllocatePrefetchStepSize = cache_line_size; // Need exact value.
1.136 + if (FLAG_IS_DEFAULT(AllocatePrefetchLines)) AllocatePrefetchLines = 12; // Use larger blocks by default.
1.137 + if (AllocatePrefetchDistance < 0) AllocatePrefetchDistance = 2*cache_line_size; // Default is not defined?
1.138 + } else {
1.139 + if (cache_line_size > AllocatePrefetchStepSize) AllocatePrefetchStepSize = cache_line_size;
1.140 + if (FLAG_IS_DEFAULT(AllocatePrefetchLines)) AllocatePrefetchLines = 3; // Optimistic value.
1.141 + if (AllocatePrefetchDistance < 0) AllocatePrefetchDistance = 3*cache_line_size; // Default is not defined?
1.142 + }
1.143 +
1.144 + assert(AllocatePrefetchLines > 0, "invalid value");
1.145 + if (AllocatePrefetchLines < 1) // Set valid value in product VM.
1.146 + AllocatePrefetchLines = 1; // Conservative value.
1.147 +
1.148 + if (AllocatePrefetchStyle == 3 && AllocatePrefetchDistance < cache_line_size)
1.149 + AllocatePrefetchStyle = 1; // Fall back if inappropriate.
1.150 +
1.151 + assert(AllocatePrefetchStyle >= 0, "AllocatePrefetchStyle should be positive");
1.152 +}
1.153 +
1.154 +void VM_Version::print_features() {
1.155 + tty->print_cr("Version: %s cache_line_size = %d", cpu_features(), (int) get_cache_line_size());
1.156 +}
1.157 +
1.158 +#ifdef COMPILER2
1.159 +// Determine section size on power6: If section size is 8 instructions,
1.160 +// there should be a difference between the two testloops of ~15 %. If
1.161 +// no difference is detected the section is assumed to be 32 instructions.
1.162 +void VM_Version::determine_section_size() {
1.163 +
1.164 + int unroll = 80;
1.165 +
1.166 + const int code_size = (2* unroll * 32 + 100)*BytesPerInstWord;
1.167 +
1.168 + // Allocate space for the code.
1.169 + ResourceMark rm;
1.170 + CodeBuffer cb("detect_section_size", code_size, 0);
1.171 + MacroAssembler* a = new MacroAssembler(&cb);
1.172 +
1.173 + uint32_t *code = (uint32_t *)a->pc();
1.174 + // Emit code.
1.175 + void (*test1)() = (void(*)())(void *)a->function_entry();
1.176 +
1.177 + Label l1;
1.178 +
1.179 + a->li(R4, 1);
1.180 + a->sldi(R4, R4, 28);
1.181 + a->b(l1);
1.182 + a->align(CodeEntryAlignment);
1.183 +
1.184 + a->bind(l1);
1.185 +
1.186 + for (int i = 0; i < unroll; i++) {
1.187 + // Schleife 1
1.188 + // ------- sector 0 ------------
1.189 + // ;; 0
1.190 + a->nop(); // 1
1.191 + a->fpnop0(); // 2
1.192 + a->fpnop1(); // 3
1.193 + a->addi(R4,R4, -1); // 4
1.194 +
1.195 + // ;; 1
1.196 + a->nop(); // 5
1.197 + a->fmr(F6, F6); // 6
1.198 + a->fmr(F7, F7); // 7
1.199 + a->endgroup(); // 8
1.200 + // ------- sector 8 ------------
1.201 +
1.202 + // ;; 2
1.203 + a->nop(); // 9
1.204 + a->nop(); // 10
1.205 + a->fmr(F8, F8); // 11
1.206 + a->fmr(F9, F9); // 12
1.207 +
1.208 + // ;; 3
1.209 + a->nop(); // 13
1.210 + a->fmr(F10, F10); // 14
1.211 + a->fmr(F11, F11); // 15
1.212 + a->endgroup(); // 16
1.213 + // -------- sector 16 -------------
1.214 +
1.215 + // ;; 4
1.216 + a->nop(); // 17
1.217 + a->nop(); // 18
1.218 + a->fmr(F15, F15); // 19
1.219 + a->fmr(F16, F16); // 20
1.220 +
1.221 + // ;; 5
1.222 + a->nop(); // 21
1.223 + a->fmr(F17, F17); // 22
1.224 + a->fmr(F18, F18); // 23
1.225 + a->endgroup(); // 24
1.226 + // ------- sector 24 ------------
1.227 +
1.228 + // ;; 6
1.229 + a->nop(); // 25
1.230 + a->nop(); // 26
1.231 + a->fmr(F19, F19); // 27
1.232 + a->fmr(F20, F20); // 28
1.233 +
1.234 + // ;; 7
1.235 + a->nop(); // 29
1.236 + a->fmr(F21, F21); // 30
1.237 + a->fmr(F22, F22); // 31
1.238 + a->brnop0(); // 32
1.239 +
1.240 + // ------- sector 32 ------------
1.241 + }
1.242 +
1.243 + // ;; 8
1.244 + a->cmpdi(CCR0, R4, unroll); // 33
1.245 + a->bge(CCR0, l1); // 34
1.246 + a->blr();
1.247 +
1.248 + // Emit code.
1.249 + void (*test2)() = (void(*)())(void *)a->function_entry();
1.250 + // uint32_t *code = (uint32_t *)a->pc();
1.251 +
1.252 + Label l2;
1.253 +
1.254 + a->li(R4, 1);
1.255 + a->sldi(R4, R4, 28);
1.256 + a->b(l2);
1.257 + a->align(CodeEntryAlignment);
1.258 +
1.259 + a->bind(l2);
1.260 +
1.261 + for (int i = 0; i < unroll; i++) {
1.262 + // Schleife 2
1.263 + // ------- sector 0 ------------
1.264 + // ;; 0
1.265 + a->brnop0(); // 1
1.266 + a->nop(); // 2
1.267 + //a->cmpdi(CCR0, R4, unroll);
1.268 + a->fpnop0(); // 3
1.269 + a->fpnop1(); // 4
1.270 + a->addi(R4,R4, -1); // 5
1.271 +
1.272 + // ;; 1
1.273 +
1.274 + a->nop(); // 6
1.275 + a->fmr(F6, F6); // 7
1.276 + a->fmr(F7, F7); // 8
1.277 + // ------- sector 8 ---------------
1.278 +
1.279 + // ;; 2
1.280 + a->endgroup(); // 9
1.281 +
1.282 + // ;; 3
1.283 + a->nop(); // 10
1.284 + a->nop(); // 11
1.285 + a->fmr(F8, F8); // 12
1.286 +
1.287 + // ;; 4
1.288 + a->fmr(F9, F9); // 13
1.289 + a->nop(); // 14
1.290 + a->fmr(F10, F10); // 15
1.291 +
1.292 + // ;; 5
1.293 + a->fmr(F11, F11); // 16
1.294 + // -------- sector 16 -------------
1.295 +
1.296 + // ;; 6
1.297 + a->endgroup(); // 17
1.298 +
1.299 + // ;; 7
1.300 + a->nop(); // 18
1.301 + a->nop(); // 19
1.302 + a->fmr(F15, F15); // 20
1.303 +
1.304 + // ;; 8
1.305 + a->fmr(F16, F16); // 21
1.306 + a->nop(); // 22
1.307 + a->fmr(F17, F17); // 23
1.308 +
1.309 + // ;; 9
1.310 + a->fmr(F18, F18); // 24
1.311 + // -------- sector 24 -------------
1.312 +
1.313 + // ;; 10
1.314 + a->endgroup(); // 25
1.315 +
1.316 + // ;; 11
1.317 + a->nop(); // 26
1.318 + a->nop(); // 27
1.319 + a->fmr(F19, F19); // 28
1.320 +
1.321 + // ;; 12
1.322 + a->fmr(F20, F20); // 29
1.323 + a->nop(); // 30
1.324 + a->fmr(F21, F21); // 31
1.325 +
1.326 + // ;; 13
1.327 + a->fmr(F22, F22); // 32
1.328 + }
1.329 +
1.330 + // -------- sector 32 -------------
1.331 + // ;; 14
1.332 + a->cmpdi(CCR0, R4, unroll); // 33
1.333 + a->bge(CCR0, l2); // 34
1.334 +
1.335 + a->blr();
1.336 + uint32_t *code_end = (uint32_t *)a->pc();
1.337 + a->flush();
1.338 +
1.339 + double loop1_seconds,loop2_seconds, rel_diff;
1.340 + uint64_t start1, stop1;
1.341 +
1.342 + start1 = os::current_thread_cpu_time(false);
1.343 + (*test1)();
1.344 + stop1 = os::current_thread_cpu_time(false);
1.345 + loop1_seconds = (stop1- start1) / (1000 *1000 *1000.0);
1.346 +
1.347 +
1.348 + start1 = os::current_thread_cpu_time(false);
1.349 + (*test2)();
1.350 + stop1 = os::current_thread_cpu_time(false);
1.351 +
1.352 + loop2_seconds = (stop1 - start1) / (1000 *1000 *1000.0);
1.353 +
1.354 + rel_diff = (loop2_seconds - loop1_seconds) / loop1_seconds *100;
1.355 +
1.356 + if (PrintAssembly) {
1.357 + ttyLocker ttyl;
1.358 + tty->print_cr("Decoding section size detection stub at " INTPTR_FORMAT " before execution:", code);
1.359 + Disassembler::decode((u_char*)code, (u_char*)code_end, tty);
1.360 + tty->print_cr("Time loop1 :%f", loop1_seconds);
1.361 + tty->print_cr("Time loop2 :%f", loop2_seconds);
1.362 + tty->print_cr("(time2 - time1) / time1 = %f %%", rel_diff);
1.363 +
1.364 + if (rel_diff > 12.0) {
1.365 + tty->print_cr("Section Size 8 Instructions");
1.366 + } else{
1.367 + tty->print_cr("Section Size 32 Instructions or Power5");
1.368 + }
1.369 + }
1.370 +
1.371 +#if 0 // TODO: PPC port
1.372 + // Set sector size (if not set explicitly).
1.373 + if (FLAG_IS_DEFAULT(Power6SectorSize128PPC64)) {
1.374 + if (rel_diff > 12.0) {
1.375 + PdScheduling::power6SectorSize = 0x20;
1.376 + } else {
1.377 + PdScheduling::power6SectorSize = 0x80;
1.378 + }
1.379 + } else if (Power6SectorSize128PPC64) {
1.380 + PdScheduling::power6SectorSize = 0x80;
1.381 + } else {
1.382 + PdScheduling::power6SectorSize = 0x20;
1.383 + }
1.384 +#endif
1.385 + if (UsePower6SchedulerPPC64) Unimplemented();
1.386 +}
1.387 +#endif // COMPILER2
1.388 +
1.389 +void VM_Version::determine_features() {
1.390 +#if defined(ABI_ELFv2)
1.391 + const int code_size = (num_features+1+2*7)*BytesPerInstWord; // TODO(asmundak): calculation is incorrect.
1.392 +#else
1.393 + // 7 InstWords for each call (function descriptor + blr instruction).
1.394 + const int code_size = (num_features+1+2*7)*BytesPerInstWord;
1.395 +#endif
1.396 + int features = 0;
1.397 +
1.398 + // create test area
1.399 + enum { BUFFER_SIZE = 2*4*K }; // Needs to be >=2* max cache line size (cache line size can't exceed min page size).
1.400 + char test_area[BUFFER_SIZE];
1.401 + char *mid_of_test_area = &test_area[BUFFER_SIZE>>1];
1.402 +
1.403 + // Allocate space for the code.
1.404 + ResourceMark rm;
1.405 + CodeBuffer cb("detect_cpu_features", code_size, 0);
1.406 + MacroAssembler* a = new MacroAssembler(&cb);
1.407 +
1.408 + // Must be set to true so we can generate the test code.
1.409 + _features = VM_Version::all_features_m;
1.410 +
1.411 + // Emit code.
1.412 + void (*test)(address addr, uint64_t offset)=(void(*)(address addr, uint64_t offset))(void *)a->function_entry();
1.413 + uint32_t *code = (uint32_t *)a->pc();
1.414 + // Don't use R0 in ldarx.
1.415 + // Keep R3_ARG1 unmodified, it contains &field (see below).
1.416 + // Keep R4_ARG2 unmodified, it contains offset = 0 (see below).
1.417 + a->fsqrt(F3, F4); // code[0] -> fsqrt_m
1.418 + a->fsqrts(F3, F4); // code[1] -> fsqrts_m
1.419 + a->isel(R7, R5, R6, 0); // code[2] -> isel_m
1.420 + a->ldarx_unchecked(R7, R3_ARG1, R4_ARG2, 1); // code[3] -> lxarx_m
1.421 + a->cmpb(R7, R5, R6); // code[4] -> bcmp
1.422 + //a->mftgpr(R7, F3); // code[5] -> mftgpr
1.423 + a->popcntb(R7, R5); // code[6] -> popcntb
1.424 + a->popcntw(R7, R5); // code[7] -> popcntw
1.425 + a->fcfids(F3, F4); // code[8] -> fcfids
1.426 + a->vand(VR0, VR0, VR0); // code[9] -> vand
1.427 + a->blr();
1.428 +
1.429 + // Emit function to set one cache line to zero. Emit function descriptor and get pointer to it.
1.430 + void (*zero_cacheline_func_ptr)(char*) = (void(*)(char*))(void *)a->function_entry();
1.431 + a->dcbz(R3_ARG1); // R3_ARG1 = addr
1.432 + a->blr();
1.433 +
1.434 + uint32_t *code_end = (uint32_t *)a->pc();
1.435 + a->flush();
1.436 + _features = VM_Version::unknown_m;
1.437 +
1.438 + // Print the detection code.
1.439 + if (PrintAssembly) {
1.440 + ttyLocker ttyl;
1.441 + tty->print_cr("Decoding cpu-feature detection stub at " INTPTR_FORMAT " before execution:", code);
1.442 + Disassembler::decode((u_char*)code, (u_char*)code_end, tty);
1.443 + }
1.444 +
1.445 + // Measure cache line size.
1.446 + memset(test_area, 0xFF, BUFFER_SIZE); // Fill test area with 0xFF.
1.447 + (*zero_cacheline_func_ptr)(mid_of_test_area); // Call function which executes dcbz to the middle.
1.448 + int count = 0; // count zeroed bytes
1.449 + for (int i = 0; i < BUFFER_SIZE; i++) if (test_area[i] == 0) count++;
1.450 + guarantee(is_power_of_2(count), "cache line size needs to be a power of 2");
1.451 + _measured_cache_line_size = count;
1.452 +
1.453 + // Execute code. Illegal instructions will be replaced by 0 in the signal handler.
1.454 + VM_Version::_is_determine_features_test_running = true;
1.455 + (*test)((address)mid_of_test_area, (uint64_t)0);
1.456 + VM_Version::_is_determine_features_test_running = false;
1.457 +
1.458 + // determine which instructions are legal.
1.459 + int feature_cntr = 0;
1.460 + if (code[feature_cntr++]) features |= fsqrt_m;
1.461 + if (code[feature_cntr++]) features |= fsqrts_m;
1.462 + if (code[feature_cntr++]) features |= isel_m;
1.463 + if (code[feature_cntr++]) features |= lxarxeh_m;
1.464 + if (code[feature_cntr++]) features |= cmpb_m;
1.465 + //if(code[feature_cntr++])features |= mftgpr_m;
1.466 + if (code[feature_cntr++]) features |= popcntb_m;
1.467 + if (code[feature_cntr++]) features |= popcntw_m;
1.468 + if (code[feature_cntr++]) features |= fcfids_m;
1.469 + if (code[feature_cntr++]) features |= vand_m;
1.470 +
1.471 + // Print the detection code.
1.472 + if (PrintAssembly) {
1.473 + ttyLocker ttyl;
1.474 + tty->print_cr("Decoding cpu-feature detection stub at " INTPTR_FORMAT " after execution:", code);
1.475 + Disassembler::decode((u_char*)code, (u_char*)code_end, tty);
1.476 + }
1.477 +
1.478 + _features = features;
1.479 +}
1.480 +
1.481 +
1.482 +static int saved_features = 0;
1.483 +
1.484 +void VM_Version::allow_all() {
1.485 + saved_features = _features;
1.486 + _features = all_features_m;
1.487 +}
1.488 +
1.489 +void VM_Version::revert() {
1.490 + _features = saved_features;
1.491 +}
