Wed, 11 Jun 2014 11:05:10 -0700
8035968: Leverage CPU Instructions to Improve SHA Performance on SPARC
Summary: Add C2 SHA intrinsics on SPARC
Reviewed-by: kvn, roland
Contributed-by: james.cheng@oracle.com
1 /*
2 * Copyright (c) 1997, 2014, 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 */
25 #include "precompiled.hpp"
26 #include "asm/macroAssembler.inline.hpp"
27 #include "memory/resourceArea.hpp"
28 #include "runtime/java.hpp"
29 #include "runtime/stubCodeGenerator.hpp"
30 #include "vm_version_sparc.hpp"
31 #ifdef TARGET_OS_FAMILY_linux
32 # include "os_linux.inline.hpp"
33 #endif
34 #ifdef TARGET_OS_FAMILY_solaris
35 # include "os_solaris.inline.hpp"
36 #endif
38 int VM_Version::_features = VM_Version::unknown_m;
39 const char* VM_Version::_features_str = "";
41 void VM_Version::initialize() {
42 _features = determine_features();
43 PrefetchCopyIntervalInBytes = prefetch_copy_interval_in_bytes();
44 PrefetchScanIntervalInBytes = prefetch_scan_interval_in_bytes();
45 PrefetchFieldsAhead = prefetch_fields_ahead();
47 assert(0 <= AllocatePrefetchInstr && AllocatePrefetchInstr <= 1, "invalid value");
48 if( AllocatePrefetchInstr < 0 ) AllocatePrefetchInstr = 0;
49 if( AllocatePrefetchInstr > 1 ) AllocatePrefetchInstr = 0;
51 // Allocation prefetch settings
52 intx cache_line_size = prefetch_data_size();
53 if( cache_line_size > AllocatePrefetchStepSize )
54 AllocatePrefetchStepSize = cache_line_size;
56 assert(AllocatePrefetchLines > 0, "invalid value");
57 if( AllocatePrefetchLines < 1 ) // set valid value in product VM
58 AllocatePrefetchLines = 3;
59 assert(AllocateInstancePrefetchLines > 0, "invalid value");
60 if( AllocateInstancePrefetchLines < 1 ) // set valid value in product VM
61 AllocateInstancePrefetchLines = 1;
63 AllocatePrefetchDistance = allocate_prefetch_distance();
64 AllocatePrefetchStyle = allocate_prefetch_style();
66 assert((AllocatePrefetchDistance % AllocatePrefetchStepSize) == 0 &&
67 (AllocatePrefetchDistance > 0), "invalid value");
68 if ((AllocatePrefetchDistance % AllocatePrefetchStepSize) != 0 ||
69 (AllocatePrefetchDistance <= 0)) {
70 AllocatePrefetchDistance = AllocatePrefetchStepSize;
71 }
73 if (AllocatePrefetchStyle == 3 && !has_blk_init()) {
74 warning("BIS instructions are not available on this CPU");
75 FLAG_SET_DEFAULT(AllocatePrefetchStyle, 1);
76 }
78 guarantee(VM_Version::has_v9(), "only SPARC v9 is supported");
80 assert(ArraycopySrcPrefetchDistance < 4096, "invalid value");
81 if (ArraycopySrcPrefetchDistance >= 4096)
82 ArraycopySrcPrefetchDistance = 4064;
83 assert(ArraycopyDstPrefetchDistance < 4096, "invalid value");
84 if (ArraycopyDstPrefetchDistance >= 4096)
85 ArraycopyDstPrefetchDistance = 4064;
87 UseSSE = 0; // Only on x86 and x64
89 _supports_cx8 = has_v9();
90 _supports_atomic_getset4 = true; // swap instruction
92 // There are Fujitsu Sparc64 CPUs which support blk_init as well so
93 // we have to take this check out of the 'is_niagara()' block below.
94 if (has_blk_init()) {
95 // When using CMS or G1, we cannot use memset() in BOT updates
96 // because the sun4v/CMT version in libc_psr uses BIS which
97 // exposes "phantom zeros" to concurrent readers. See 6948537.
98 if (FLAG_IS_DEFAULT(UseMemSetInBOT) && (UseConcMarkSweepGC || UseG1GC)) {
99 FLAG_SET_DEFAULT(UseMemSetInBOT, false);
100 }
101 // Issue a stern warning if the user has explicitly set
102 // UseMemSetInBOT (it is known to cause issues), but allow
103 // use for experimentation and debugging.
104 if (UseConcMarkSweepGC || UseG1GC) {
105 if (UseMemSetInBOT) {
106 assert(!FLAG_IS_DEFAULT(UseMemSetInBOT), "Error");
107 warning("Experimental flag -XX:+UseMemSetInBOT is known to cause instability"
108 " on sun4v; please understand that you are using at your own risk!");
109 }
110 }
111 }
113 if (is_niagara()) {
114 // Indirect branch is the same cost as direct
115 if (FLAG_IS_DEFAULT(UseInlineCaches)) {
116 FLAG_SET_DEFAULT(UseInlineCaches, false);
117 }
118 // Align loops on a single instruction boundary.
119 if (FLAG_IS_DEFAULT(OptoLoopAlignment)) {
120 FLAG_SET_DEFAULT(OptoLoopAlignment, 4);
121 }
122 #ifdef _LP64
123 // 32-bit oops don't make sense for the 64-bit VM on sparc
124 // since the 32-bit VM has the same registers and smaller objects.
125 Universe::set_narrow_oop_shift(LogMinObjAlignmentInBytes);
126 Universe::set_narrow_klass_shift(LogKlassAlignmentInBytes);
127 #endif // _LP64
128 #ifdef COMPILER2
129 // Indirect branch is the same cost as direct
130 if (FLAG_IS_DEFAULT(UseJumpTables)) {
131 FLAG_SET_DEFAULT(UseJumpTables, true);
132 }
133 // Single-issue, so entry and loop tops are
134 // aligned on a single instruction boundary
135 if (FLAG_IS_DEFAULT(InteriorEntryAlignment)) {
136 FLAG_SET_DEFAULT(InteriorEntryAlignment, 4);
137 }
138 if (is_niagara_plus()) {
139 if (has_blk_init() && UseTLAB &&
140 FLAG_IS_DEFAULT(AllocatePrefetchInstr)) {
141 // Use BIS instruction for TLAB allocation prefetch.
142 FLAG_SET_ERGO(intx, AllocatePrefetchInstr, 1);
143 if (FLAG_IS_DEFAULT(AllocatePrefetchStyle)) {
144 FLAG_SET_ERGO(intx, AllocatePrefetchStyle, 3);
145 }
146 if (FLAG_IS_DEFAULT(AllocatePrefetchDistance)) {
147 // Use smaller prefetch distance with BIS
148 FLAG_SET_DEFAULT(AllocatePrefetchDistance, 64);
149 }
150 }
151 if (is_T4()) {
152 // Double number of prefetched cache lines on T4
153 // since L2 cache line size is smaller (32 bytes).
154 if (FLAG_IS_DEFAULT(AllocatePrefetchLines)) {
155 FLAG_SET_ERGO(intx, AllocatePrefetchLines, AllocatePrefetchLines*2);
156 }
157 if (FLAG_IS_DEFAULT(AllocateInstancePrefetchLines)) {
158 FLAG_SET_ERGO(intx, AllocateInstancePrefetchLines, AllocateInstancePrefetchLines*2);
159 }
160 }
161 if (AllocatePrefetchStyle != 3 && FLAG_IS_DEFAULT(AllocatePrefetchDistance)) {
162 // Use different prefetch distance without BIS
163 FLAG_SET_DEFAULT(AllocatePrefetchDistance, 256);
164 }
165 if (AllocatePrefetchInstr == 1) {
166 // Need a space at the end of TLAB for BIS since it
167 // will fault when accessing memory outside of heap.
169 // +1 for rounding up to next cache line, +1 to be safe
170 int lines = AllocatePrefetchLines + 2;
171 int step_size = AllocatePrefetchStepSize;
172 int distance = AllocatePrefetchDistance;
173 _reserve_for_allocation_prefetch = (distance + step_size*lines)/(int)HeapWordSize;
174 }
175 }
176 #endif
177 }
179 // Use hardware population count instruction if available.
180 if (has_hardware_popc()) {
181 if (FLAG_IS_DEFAULT(UsePopCountInstruction)) {
182 FLAG_SET_DEFAULT(UsePopCountInstruction, true);
183 }
184 } else if (UsePopCountInstruction) {
185 warning("POPC instruction is not available on this CPU");
186 FLAG_SET_DEFAULT(UsePopCountInstruction, false);
187 }
189 // T4 and newer Sparc cpus have new compare and branch instruction.
190 if (has_cbcond()) {
191 if (FLAG_IS_DEFAULT(UseCBCond)) {
192 FLAG_SET_DEFAULT(UseCBCond, true);
193 }
194 } else if (UseCBCond) {
195 warning("CBCOND instruction is not available on this CPU");
196 FLAG_SET_DEFAULT(UseCBCond, false);
197 }
199 assert(BlockZeroingLowLimit > 0, "invalid value");
200 if (has_block_zeroing()) {
201 if (FLAG_IS_DEFAULT(UseBlockZeroing)) {
202 FLAG_SET_DEFAULT(UseBlockZeroing, true);
203 }
204 } else if (UseBlockZeroing) {
205 warning("BIS zeroing instructions are not available on this CPU");
206 FLAG_SET_DEFAULT(UseBlockZeroing, false);
207 }
209 assert(BlockCopyLowLimit > 0, "invalid value");
210 if (has_block_zeroing()) { // has_blk_init() && is_T4(): core's local L2 cache
211 if (FLAG_IS_DEFAULT(UseBlockCopy)) {
212 FLAG_SET_DEFAULT(UseBlockCopy, true);
213 }
214 } else if (UseBlockCopy) {
215 warning("BIS instructions are not available or expensive on this CPU");
216 FLAG_SET_DEFAULT(UseBlockCopy, false);
217 }
219 #ifdef COMPILER2
220 // T4 and newer Sparc cpus have fast RDPC.
221 if (has_fast_rdpc() && FLAG_IS_DEFAULT(UseRDPCForConstantTableBase)) {
222 FLAG_SET_DEFAULT(UseRDPCForConstantTableBase, true);
223 }
225 // Currently not supported anywhere.
226 FLAG_SET_DEFAULT(UseFPUForSpilling, false);
228 MaxVectorSize = 8;
230 assert((InteriorEntryAlignment % relocInfo::addr_unit()) == 0, "alignment is not a multiple of NOP size");
231 #endif
233 assert((CodeEntryAlignment % relocInfo::addr_unit()) == 0, "alignment is not a multiple of NOP size");
234 assert((OptoLoopAlignment % relocInfo::addr_unit()) == 0, "alignment is not a multiple of NOP size");
236 char buf[512];
237 jio_snprintf(buf, sizeof(buf), "%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s",
238 (has_v9() ? ", v9" : (has_v8() ? ", v8" : "")),
239 (has_hardware_popc() ? ", popc" : ""),
240 (has_vis1() ? ", vis1" : ""),
241 (has_vis2() ? ", vis2" : ""),
242 (has_vis3() ? ", vis3" : ""),
243 (has_blk_init() ? ", blk_init" : ""),
244 (has_cbcond() ? ", cbcond" : ""),
245 (has_aes() ? ", aes" : ""),
246 (has_sha1() ? ", sha1" : ""),
247 (has_sha256() ? ", sha256" : ""),
248 (has_sha512() ? ", sha512" : ""),
249 (is_ultra3() ? ", ultra3" : ""),
250 (is_sun4v() ? ", sun4v" : ""),
251 (is_niagara_plus() ? ", niagara_plus" : (is_niagara() ? ", niagara" : "")),
252 (is_sparc64() ? ", sparc64" : ""),
253 (!has_hardware_mul32() ? ", no-mul32" : ""),
254 (!has_hardware_div32() ? ", no-div32" : ""),
255 (!has_hardware_fsmuld() ? ", no-fsmuld" : ""));
257 // buf is started with ", " or is empty
258 _features_str = strdup(strlen(buf) > 2 ? buf + 2 : buf);
260 // UseVIS is set to the smallest of what hardware supports and what
261 // the command line requires. I.e., you cannot set UseVIS to 3 on
262 // older UltraSparc which do not support it.
263 if (UseVIS > 3) UseVIS=3;
264 if (UseVIS < 0) UseVIS=0;
265 if (!has_vis3()) // Drop to 2 if no VIS3 support
266 UseVIS = MIN2((intx)2,UseVIS);
267 if (!has_vis2()) // Drop to 1 if no VIS2 support
268 UseVIS = MIN2((intx)1,UseVIS);
269 if (!has_vis1()) // Drop to 0 if no VIS1 support
270 UseVIS = 0;
272 // SPARC T4 and above should have support for AES instructions
273 if (has_aes()) {
274 if (UseVIS > 2) { // AES intrinsics use MOVxTOd/MOVdTOx which are VIS3
275 if (FLAG_IS_DEFAULT(UseAES)) {
276 FLAG_SET_DEFAULT(UseAES, true);
277 }
278 if (FLAG_IS_DEFAULT(UseAESIntrinsics)) {
279 FLAG_SET_DEFAULT(UseAESIntrinsics, true);
280 }
281 // we disable both the AES flags if either of them is disabled on the command line
282 if (!UseAES || !UseAESIntrinsics) {
283 FLAG_SET_DEFAULT(UseAES, false);
284 FLAG_SET_DEFAULT(UseAESIntrinsics, false);
285 }
286 } else {
287 if (UseAES || UseAESIntrinsics) {
288 warning("SPARC AES intrinsics require VIS3 instruction support. Intrinsics will be disabled.");
289 if (UseAES) {
290 FLAG_SET_DEFAULT(UseAES, false);
291 }
292 if (UseAESIntrinsics) {
293 FLAG_SET_DEFAULT(UseAESIntrinsics, false);
294 }
295 }
296 }
297 } else if (UseAES || UseAESIntrinsics) {
298 warning("AES instructions are not available on this CPU");
299 if (UseAES) {
300 FLAG_SET_DEFAULT(UseAES, false);
301 }
302 if (UseAESIntrinsics) {
303 FLAG_SET_DEFAULT(UseAESIntrinsics, false);
304 }
305 }
307 // SHA1, SHA256, and SHA512 instructions were added to SPARC T-series at different times
308 if (has_sha1() || has_sha256() || has_sha512()) {
309 if (UseVIS > 0) { // SHA intrinsics use VIS1 instructions
310 if (FLAG_IS_DEFAULT(UseSHA)) {
311 FLAG_SET_DEFAULT(UseSHA, true);
312 }
313 } else {
314 if (UseSHA) {
315 warning("SPARC SHA intrinsics require VIS1 instruction support. Intrinsics will be disabled.");
316 FLAG_SET_DEFAULT(UseSHA, false);
317 }
318 }
319 } else if (UseSHA) {
320 warning("SHA instructions are not available on this CPU");
321 FLAG_SET_DEFAULT(UseSHA, false);
322 }
324 if (!UseSHA) {
325 FLAG_SET_DEFAULT(UseSHA1Intrinsics, false);
326 FLAG_SET_DEFAULT(UseSHA256Intrinsics, false);
327 FLAG_SET_DEFAULT(UseSHA512Intrinsics, false);
328 } else {
329 if (has_sha1()) {
330 if (FLAG_IS_DEFAULT(UseSHA1Intrinsics)) {
331 FLAG_SET_DEFAULT(UseSHA1Intrinsics, true);
332 }
333 } else if (UseSHA1Intrinsics) {
334 warning("SHA1 instruction is not available on this CPU.");
335 FLAG_SET_DEFAULT(UseSHA1Intrinsics, false);
336 }
337 if (has_sha256()) {
338 if (FLAG_IS_DEFAULT(UseSHA256Intrinsics)) {
339 FLAG_SET_DEFAULT(UseSHA256Intrinsics, true);
340 }
341 } else if (UseSHA256Intrinsics) {
342 warning("SHA256 instruction (for SHA-224 and SHA-256) is not available on this CPU.");
343 FLAG_SET_DEFAULT(UseSHA256Intrinsics, false);
344 }
346 if (has_sha512()) {
347 if (FLAG_IS_DEFAULT(UseSHA512Intrinsics)) {
348 FLAG_SET_DEFAULT(UseSHA512Intrinsics, true);
349 }
350 } else if (UseSHA512Intrinsics) {
351 warning("SHA512 instruction (for SHA-384 and SHA-512) is not available on this CPU.");
352 FLAG_SET_DEFAULT(UseSHA512Intrinsics, false);
353 }
354 if (!(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics)) {
355 FLAG_SET_DEFAULT(UseSHA, false);
356 }
357 }
359 if (FLAG_IS_DEFAULT(ContendedPaddingWidth) &&
360 (cache_line_size > ContendedPaddingWidth))
361 ContendedPaddingWidth = cache_line_size;
363 #ifndef PRODUCT
364 if (PrintMiscellaneous && Verbose) {
365 tty->print("Allocation");
366 if (AllocatePrefetchStyle <= 0) {
367 tty->print_cr(": no prefetching");
368 } else {
369 tty->print(" prefetching: ");
370 if (AllocatePrefetchInstr == 0) {
371 tty->print("PREFETCH");
372 } else if (AllocatePrefetchInstr == 1) {
373 tty->print("BIS");
374 }
375 if (AllocatePrefetchLines > 1) {
376 tty->print_cr(" at distance %d, %d lines of %d bytes", (int) AllocatePrefetchDistance, (int) AllocatePrefetchLines, (int) AllocatePrefetchStepSize);
377 } else {
378 tty->print_cr(" at distance %d, one line of %d bytes", (int) AllocatePrefetchDistance, (int) AllocatePrefetchStepSize);
379 }
380 }
381 if (PrefetchCopyIntervalInBytes > 0) {
382 tty->print_cr("PrefetchCopyIntervalInBytes %d", (int) PrefetchCopyIntervalInBytes);
383 }
384 if (PrefetchScanIntervalInBytes > 0) {
385 tty->print_cr("PrefetchScanIntervalInBytes %d", (int) PrefetchScanIntervalInBytes);
386 }
387 if (PrefetchFieldsAhead > 0) {
388 tty->print_cr("PrefetchFieldsAhead %d", (int) PrefetchFieldsAhead);
389 }
390 if (ContendedPaddingWidth > 0) {
391 tty->print_cr("ContendedPaddingWidth %d", (int) ContendedPaddingWidth);
392 }
393 }
394 #endif // PRODUCT
395 }
397 void VM_Version::print_features() {
398 tty->print_cr("Version:%s", cpu_features());
399 }
401 int VM_Version::determine_features() {
402 if (UseV8InstrsOnly) {
403 NOT_PRODUCT(if (PrintMiscellaneous && Verbose) tty->print_cr("Version is Forced-V8");)
404 return generic_v8_m;
405 }
407 int features = platform_features(unknown_m); // platform_features() is os_arch specific
409 if (features == unknown_m) {
410 features = generic_v9_m;
411 warning("Cannot recognize SPARC version. Default to V9");
412 }
414 assert(is_T_family(features) == is_niagara(features), "Niagara should be T series");
415 if (UseNiagaraInstrs) { // Force code generation for Niagara
416 if (is_T_family(features)) {
417 // Happy to accomodate...
418 } else {
419 NOT_PRODUCT(if (PrintMiscellaneous && Verbose) tty->print_cr("Version is Forced-Niagara");)
420 features |= T_family_m;
421 }
422 } else {
423 if (is_T_family(features) && !FLAG_IS_DEFAULT(UseNiagaraInstrs)) {
424 NOT_PRODUCT(if (PrintMiscellaneous && Verbose) tty->print_cr("Version is Forced-Not-Niagara");)
425 features &= ~(T_family_m | T1_model_m);
426 } else {
427 // Happy to accomodate...
428 }
429 }
431 return features;
432 }
434 static int saved_features = 0;
436 void VM_Version::allow_all() {
437 saved_features = _features;
438 _features = all_features_m;
439 }
441 void VM_Version::revert() {
442 _features = saved_features;
443 }
445 unsigned int VM_Version::calc_parallel_worker_threads() {
446 unsigned int result;
447 if (is_M_series()) {
448 // for now, use same gc thread calculation for M-series as for niagara-plus
449 // in future, we may want to tweak parameters for nof_parallel_worker_thread
450 result = nof_parallel_worker_threads(5, 16, 8);
451 } else if (is_niagara_plus()) {
452 result = nof_parallel_worker_threads(5, 16, 8);
453 } else {
454 result = nof_parallel_worker_threads(5, 8, 8);
455 }
456 return result;
457 }