src/share/vm/opto/compile.cpp

Mon, 27 May 2013 12:56:34 +0200

author
stefank
date
Mon, 27 May 2013 12:56:34 +0200
changeset 5195
95c00927be11
parent 5110
6f3fd5150b67
child 5237
f2110083203d
child 6441
d2907f74462e
permissions
-rw-r--r--

8015428: Remove unused CDS support from StringTable
Summary: The string in StringTable is not used by CDS anymore. Remove the unnecessary code in preparation for 8015422: Large performance hit when the StringTable is walked twice in Parallel Scavenge
Reviewed-by: pliden, tschatzl, coleenp

duke@435 1 /*
kvn@3651 2 * Copyright (c) 1997, 2012, Oracle and/or its affiliates. All rights reserved.
duke@435 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
duke@435 4 *
duke@435 5 * This code is free software; you can redistribute it and/or modify it
duke@435 6 * under the terms of the GNU General Public License version 2 only, as
duke@435 7 * published by the Free Software Foundation.
duke@435 8 *
duke@435 9 * This code is distributed in the hope that it will be useful, but WITHOUT
duke@435 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
duke@435 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
duke@435 12 * version 2 for more details (a copy is included in the LICENSE file that
duke@435 13 * accompanied this code).
duke@435 14 *
duke@435 15 * You should have received a copy of the GNU General Public License version
duke@435 16 * 2 along with this work; if not, write to the Free Software Foundation,
duke@435 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
duke@435 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.
duke@435 22 *
duke@435 23 */
duke@435 24
stefank@2314 25 #include "precompiled.hpp"
twisti@4318 26 #include "asm/macroAssembler.hpp"
twisti@4318 27 #include "asm/macroAssembler.inline.hpp"
stefank@2314 28 #include "classfile/systemDictionary.hpp"
stefank@2314 29 #include "code/exceptionHandlerTable.hpp"
stefank@2314 30 #include "code/nmethod.hpp"
stefank@2314 31 #include "compiler/compileLog.hpp"
twisti@4318 32 #include "compiler/disassembler.hpp"
stefank@2314 33 #include "compiler/oopMap.hpp"
stefank@2314 34 #include "opto/addnode.hpp"
stefank@2314 35 #include "opto/block.hpp"
stefank@2314 36 #include "opto/c2compiler.hpp"
stefank@2314 37 #include "opto/callGenerator.hpp"
stefank@2314 38 #include "opto/callnode.hpp"
stefank@2314 39 #include "opto/cfgnode.hpp"
stefank@2314 40 #include "opto/chaitin.hpp"
stefank@2314 41 #include "opto/compile.hpp"
stefank@2314 42 #include "opto/connode.hpp"
stefank@2314 43 #include "opto/divnode.hpp"
stefank@2314 44 #include "opto/escape.hpp"
stefank@2314 45 #include "opto/idealGraphPrinter.hpp"
stefank@2314 46 #include "opto/loopnode.hpp"
stefank@2314 47 #include "opto/machnode.hpp"
stefank@2314 48 #include "opto/macro.hpp"
stefank@2314 49 #include "opto/matcher.hpp"
stefank@2314 50 #include "opto/memnode.hpp"
stefank@2314 51 #include "opto/mulnode.hpp"
stefank@2314 52 #include "opto/node.hpp"
stefank@2314 53 #include "opto/opcodes.hpp"
stefank@2314 54 #include "opto/output.hpp"
stefank@2314 55 #include "opto/parse.hpp"
stefank@2314 56 #include "opto/phaseX.hpp"
stefank@2314 57 #include "opto/rootnode.hpp"
stefank@2314 58 #include "opto/runtime.hpp"
stefank@2314 59 #include "opto/stringopts.hpp"
stefank@2314 60 #include "opto/type.hpp"
stefank@2314 61 #include "opto/vectornode.hpp"
stefank@2314 62 #include "runtime/arguments.hpp"
stefank@2314 63 #include "runtime/signature.hpp"
stefank@2314 64 #include "runtime/stubRoutines.hpp"
stefank@2314 65 #include "runtime/timer.hpp"
stefank@2314 66 #include "utilities/copy.hpp"
stefank@2314 67 #ifdef TARGET_ARCH_MODEL_x86_32
stefank@2314 68 # include "adfiles/ad_x86_32.hpp"
stefank@2314 69 #endif
stefank@2314 70 #ifdef TARGET_ARCH_MODEL_x86_64
stefank@2314 71 # include "adfiles/ad_x86_64.hpp"
stefank@2314 72 #endif
stefank@2314 73 #ifdef TARGET_ARCH_MODEL_sparc
stefank@2314 74 # include "adfiles/ad_sparc.hpp"
stefank@2314 75 #endif
stefank@2314 76 #ifdef TARGET_ARCH_MODEL_zero
stefank@2314 77 # include "adfiles/ad_zero.hpp"
stefank@2314 78 #endif
bobv@2508 79 #ifdef TARGET_ARCH_MODEL_arm
bobv@2508 80 # include "adfiles/ad_arm.hpp"
bobv@2508 81 #endif
bobv@2508 82 #ifdef TARGET_ARCH_MODEL_ppc
bobv@2508 83 # include "adfiles/ad_ppc.hpp"
bobv@2508 84 #endif
duke@435 85
twisti@2350 86
twisti@2350 87 // -------------------- Compile::mach_constant_base_node -----------------------
twisti@2350 88 // Constant table base node singleton.
twisti@2350 89 MachConstantBaseNode* Compile::mach_constant_base_node() {
twisti@2350 90 if (_mach_constant_base_node == NULL) {
twisti@2350 91 _mach_constant_base_node = new (C) MachConstantBaseNode();
twisti@2350 92 _mach_constant_base_node->add_req(C->root());
twisti@2350 93 }
twisti@2350 94 return _mach_constant_base_node;
twisti@2350 95 }
twisti@2350 96
twisti@2350 97
duke@435 98 /// Support for intrinsics.
duke@435 99
duke@435 100 // Return the index at which m must be inserted (or already exists).
duke@435 101 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
duke@435 102 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) {
duke@435 103 #ifdef ASSERT
duke@435 104 for (int i = 1; i < _intrinsics->length(); i++) {
duke@435 105 CallGenerator* cg1 = _intrinsics->at(i-1);
duke@435 106 CallGenerator* cg2 = _intrinsics->at(i);
duke@435 107 assert(cg1->method() != cg2->method()
duke@435 108 ? cg1->method() < cg2->method()
duke@435 109 : cg1->is_virtual() < cg2->is_virtual(),
duke@435 110 "compiler intrinsics list must stay sorted");
duke@435 111 }
duke@435 112 #endif
duke@435 113 // Binary search sorted list, in decreasing intervals [lo, hi].
duke@435 114 int lo = 0, hi = _intrinsics->length()-1;
duke@435 115 while (lo <= hi) {
duke@435 116 int mid = (uint)(hi + lo) / 2;
duke@435 117 ciMethod* mid_m = _intrinsics->at(mid)->method();
duke@435 118 if (m < mid_m) {
duke@435 119 hi = mid-1;
duke@435 120 } else if (m > mid_m) {
duke@435 121 lo = mid+1;
duke@435 122 } else {
duke@435 123 // look at minor sort key
duke@435 124 bool mid_virt = _intrinsics->at(mid)->is_virtual();
duke@435 125 if (is_virtual < mid_virt) {
duke@435 126 hi = mid-1;
duke@435 127 } else if (is_virtual > mid_virt) {
duke@435 128 lo = mid+1;
duke@435 129 } else {
duke@435 130 return mid; // exact match
duke@435 131 }
duke@435 132 }
duke@435 133 }
duke@435 134 return lo; // inexact match
duke@435 135 }
duke@435 136
duke@435 137 void Compile::register_intrinsic(CallGenerator* cg) {
duke@435 138 if (_intrinsics == NULL) {
roland@4409 139 _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL);
duke@435 140 }
duke@435 141 // This code is stolen from ciObjectFactory::insert.
duke@435 142 // Really, GrowableArray should have methods for
duke@435 143 // insert_at, remove_at, and binary_search.
duke@435 144 int len = _intrinsics->length();
duke@435 145 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual());
duke@435 146 if (index == len) {
duke@435 147 _intrinsics->append(cg);
duke@435 148 } else {
duke@435 149 #ifdef ASSERT
duke@435 150 CallGenerator* oldcg = _intrinsics->at(index);
duke@435 151 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice");
duke@435 152 #endif
duke@435 153 _intrinsics->append(_intrinsics->at(len-1));
duke@435 154 int pos;
duke@435 155 for (pos = len-2; pos >= index; pos--) {
duke@435 156 _intrinsics->at_put(pos+1,_intrinsics->at(pos));
duke@435 157 }
duke@435 158 _intrinsics->at_put(index, cg);
duke@435 159 }
duke@435 160 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
duke@435 161 }
duke@435 162
duke@435 163 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
duke@435 164 assert(m->is_loaded(), "don't try this on unloaded methods");
duke@435 165 if (_intrinsics != NULL) {
duke@435 166 int index = intrinsic_insertion_index(m, is_virtual);
duke@435 167 if (index < _intrinsics->length()
duke@435 168 && _intrinsics->at(index)->method() == m
duke@435 169 && _intrinsics->at(index)->is_virtual() == is_virtual) {
duke@435 170 return _intrinsics->at(index);
duke@435 171 }
duke@435 172 }
duke@435 173 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
jrose@1291 174 if (m->intrinsic_id() != vmIntrinsics::_none &&
jrose@1291 175 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
duke@435 176 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
duke@435 177 if (cg != NULL) {
duke@435 178 // Save it for next time:
duke@435 179 register_intrinsic(cg);
duke@435 180 return cg;
duke@435 181 } else {
duke@435 182 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
duke@435 183 }
duke@435 184 }
duke@435 185 return NULL;
duke@435 186 }
duke@435 187
duke@435 188 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined
duke@435 189 // in library_call.cpp.
duke@435 190
duke@435 191
duke@435 192 #ifndef PRODUCT
duke@435 193 // statistics gathering...
duke@435 194
duke@435 195 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0};
duke@435 196 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0};
duke@435 197
duke@435 198 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
duke@435 199 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
duke@435 200 int oflags = _intrinsic_hist_flags[id];
duke@435 201 assert(flags != 0, "what happened?");
duke@435 202 if (is_virtual) {
duke@435 203 flags |= _intrinsic_virtual;
duke@435 204 }
duke@435 205 bool changed = (flags != oflags);
duke@435 206 if ((flags & _intrinsic_worked) != 0) {
duke@435 207 juint count = (_intrinsic_hist_count[id] += 1);
duke@435 208 if (count == 1) {
duke@435 209 changed = true; // first time
duke@435 210 }
duke@435 211 // increment the overall count also:
duke@435 212 _intrinsic_hist_count[vmIntrinsics::_none] += 1;
duke@435 213 }
duke@435 214 if (changed) {
duke@435 215 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
duke@435 216 // Something changed about the intrinsic's virtuality.
duke@435 217 if ((flags & _intrinsic_virtual) != 0) {
duke@435 218 // This is the first use of this intrinsic as a virtual call.
duke@435 219 if (oflags != 0) {
duke@435 220 // We already saw it as a non-virtual, so note both cases.
duke@435 221 flags |= _intrinsic_both;
duke@435 222 }
duke@435 223 } else if ((oflags & _intrinsic_both) == 0) {
duke@435 224 // This is the first use of this intrinsic as a non-virtual
duke@435 225 flags |= _intrinsic_both;
duke@435 226 }
duke@435 227 }
duke@435 228 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags);
duke@435 229 }
duke@435 230 // update the overall flags also:
duke@435 231 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags;
duke@435 232 return changed;
duke@435 233 }
duke@435 234
duke@435 235 static char* format_flags(int flags, char* buf) {
duke@435 236 buf[0] = 0;
duke@435 237 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
duke@435 238 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
duke@435 239 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
duke@435 240 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
duke@435 241 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
duke@435 242 if (buf[0] == 0) strcat(buf, ",");
duke@435 243 assert(buf[0] == ',', "must be");
duke@435 244 return &buf[1];
duke@435 245 }
duke@435 246
duke@435 247 void Compile::print_intrinsic_statistics() {
duke@435 248 char flagsbuf[100];
duke@435 249 ttyLocker ttyl;
duke@435 250 if (xtty != NULL) xtty->head("statistics type='intrinsic'");
duke@435 251 tty->print_cr("Compiler intrinsic usage:");
duke@435 252 juint total = _intrinsic_hist_count[vmIntrinsics::_none];
duke@435 253 if (total == 0) total = 1; // avoid div0 in case of no successes
duke@435 254 #define PRINT_STAT_LINE(name, c, f) \
duke@435 255 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
duke@435 256 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) {
duke@435 257 vmIntrinsics::ID id = (vmIntrinsics::ID) index;
duke@435 258 int flags = _intrinsic_hist_flags[id];
duke@435 259 juint count = _intrinsic_hist_count[id];
duke@435 260 if ((flags | count) != 0) {
duke@435 261 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
duke@435 262 }
duke@435 263 }
duke@435 264 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf));
duke@435 265 if (xtty != NULL) xtty->tail("statistics");
duke@435 266 }
duke@435 267
duke@435 268 void Compile::print_statistics() {
duke@435 269 { ttyLocker ttyl;
duke@435 270 if (xtty != NULL) xtty->head("statistics type='opto'");
duke@435 271 Parse::print_statistics();
duke@435 272 PhaseCCP::print_statistics();
duke@435 273 PhaseRegAlloc::print_statistics();
duke@435 274 Scheduling::print_statistics();
duke@435 275 PhasePeephole::print_statistics();
duke@435 276 PhaseIdealLoop::print_statistics();
duke@435 277 if (xtty != NULL) xtty->tail("statistics");
duke@435 278 }
duke@435 279 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) {
duke@435 280 // put this under its own <statistics> element.
duke@435 281 print_intrinsic_statistics();
duke@435 282 }
duke@435 283 }
duke@435 284 #endif //PRODUCT
duke@435 285
duke@435 286 // Support for bundling info
duke@435 287 Bundle* Compile::node_bundling(const Node *n) {
duke@435 288 assert(valid_bundle_info(n), "oob");
duke@435 289 return &_node_bundling_base[n->_idx];
duke@435 290 }
duke@435 291
duke@435 292 bool Compile::valid_bundle_info(const Node *n) {
duke@435 293 return (_node_bundling_limit > n->_idx);
duke@435 294 }
duke@435 295
duke@435 296
never@1515 297 void Compile::gvn_replace_by(Node* n, Node* nn) {
never@1515 298 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
never@1515 299 Node* use = n->last_out(i);
never@1515 300 bool is_in_table = initial_gvn()->hash_delete(use);
never@1515 301 uint uses_found = 0;
never@1515 302 for (uint j = 0; j < use->len(); j++) {
never@1515 303 if (use->in(j) == n) {
never@1515 304 if (j < use->req())
never@1515 305 use->set_req(j, nn);
never@1515 306 else
never@1515 307 use->set_prec(j, nn);
never@1515 308 uses_found++;
never@1515 309 }
never@1515 310 }
never@1515 311 if (is_in_table) {
never@1515 312 // reinsert into table
never@1515 313 initial_gvn()->hash_find_insert(use);
never@1515 314 }
never@1515 315 record_for_igvn(use);
never@1515 316 i -= uses_found; // we deleted 1 or more copies of this edge
never@1515 317 }
never@1515 318 }
never@1515 319
never@1515 320
bharadwaj@4315 321 static inline bool not_a_node(const Node* n) {
bharadwaj@4315 322 if (n == NULL) return true;
bharadwaj@4315 323 if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc.
bharadwaj@4315 324 if (*(address*)n == badAddress) return true; // kill by Node::destruct
bharadwaj@4315 325 return false;
bharadwaj@4315 326 }
never@1515 327
duke@435 328 // Identify all nodes that are reachable from below, useful.
duke@435 329 // Use breadth-first pass that records state in a Unique_Node_List,
duke@435 330 // recursive traversal is slower.
duke@435 331 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
duke@435 332 int estimated_worklist_size = unique();
duke@435 333 useful.map( estimated_worklist_size, NULL ); // preallocate space
duke@435 334
duke@435 335 // Initialize worklist
duke@435 336 if (root() != NULL) { useful.push(root()); }
duke@435 337 // If 'top' is cached, declare it useful to preserve cached node
duke@435 338 if( cached_top_node() ) { useful.push(cached_top_node()); }
duke@435 339
duke@435 340 // Push all useful nodes onto the list, breadthfirst
duke@435 341 for( uint next = 0; next < useful.size(); ++next ) {
duke@435 342 assert( next < unique(), "Unique useful nodes < total nodes");
duke@435 343 Node *n = useful.at(next);
duke@435 344 uint max = n->len();
duke@435 345 for( uint i = 0; i < max; ++i ) {
duke@435 346 Node *m = n->in(i);
bharadwaj@4315 347 if (not_a_node(m)) continue;
duke@435 348 useful.push(m);
duke@435 349 }
duke@435 350 }
duke@435 351 }
duke@435 352
bharadwaj@4315 353 // Update dead_node_list with any missing dead nodes using useful
bharadwaj@4315 354 // list. Consider all non-useful nodes to be useless i.e., dead nodes.
bharadwaj@4315 355 void Compile::update_dead_node_list(Unique_Node_List &useful) {
bharadwaj@4315 356 uint max_idx = unique();
bharadwaj@4315 357 VectorSet& useful_node_set = useful.member_set();
bharadwaj@4315 358
bharadwaj@4315 359 for (uint node_idx = 0; node_idx < max_idx; node_idx++) {
bharadwaj@4315 360 // If node with index node_idx is not in useful set,
bharadwaj@4315 361 // mark it as dead in dead node list.
bharadwaj@4315 362 if (! useful_node_set.test(node_idx) ) {
bharadwaj@4315 363 record_dead_node(node_idx);
bharadwaj@4315 364 }
bharadwaj@4315 365 }
bharadwaj@4315 366 }
bharadwaj@4315 367
roland@4409 368 void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
roland@4409 369 int shift = 0;
roland@4409 370 for (int i = 0; i < inlines->length(); i++) {
roland@4409 371 CallGenerator* cg = inlines->at(i);
roland@4409 372 CallNode* call = cg->call_node();
roland@4409 373 if (shift > 0) {
roland@4409 374 inlines->at_put(i-shift, cg);
roland@4409 375 }
roland@4409 376 if (!useful.member(call)) {
roland@4409 377 shift++;
roland@4409 378 }
roland@4409 379 }
roland@4409 380 inlines->trunc_to(inlines->length()-shift);
roland@4409 381 }
roland@4409 382
duke@435 383 // Disconnect all useless nodes by disconnecting those at the boundary.
duke@435 384 void Compile::remove_useless_nodes(Unique_Node_List &useful) {
duke@435 385 uint next = 0;
kvn@3260 386 while (next < useful.size()) {
duke@435 387 Node *n = useful.at(next++);
duke@435 388 // Use raw traversal of out edges since this code removes out edges
duke@435 389 int max = n->outcnt();
kvn@3260 390 for (int j = 0; j < max; ++j) {
duke@435 391 Node* child = n->raw_out(j);
kvn@3260 392 if (! useful.member(child)) {
kvn@3260 393 assert(!child->is_top() || child != top(),
kvn@3260 394 "If top is cached in Compile object it is in useful list");
duke@435 395 // Only need to remove this out-edge to the useless node
duke@435 396 n->raw_del_out(j);
duke@435 397 --j;
duke@435 398 --max;
duke@435 399 }
duke@435 400 }
duke@435 401 if (n->outcnt() == 1 && n->has_special_unique_user()) {
kvn@3260 402 record_for_igvn(n->unique_out());
kvn@3260 403 }
kvn@3260 404 }
kvn@3260 405 // Remove useless macro and predicate opaq nodes
kvn@3260 406 for (int i = C->macro_count()-1; i >= 0; i--) {
kvn@3260 407 Node* n = C->macro_node(i);
kvn@3260 408 if (!useful.member(n)) {
kvn@3260 409 remove_macro_node(n);
duke@435 410 }
duke@435 411 }
roland@4589 412 // Remove useless expensive node
roland@4589 413 for (int i = C->expensive_count()-1; i >= 0; i--) {
roland@4589 414 Node* n = C->expensive_node(i);
roland@4589 415 if (!useful.member(n)) {
roland@4589 416 remove_expensive_node(n);
roland@4589 417 }
roland@4589 418 }
roland@4409 419 // clean up the late inline lists
roland@4409 420 remove_useless_late_inlines(&_string_late_inlines, useful);
kvn@5110 421 remove_useless_late_inlines(&_boxing_late_inlines, useful);
roland@4409 422 remove_useless_late_inlines(&_late_inlines, useful);
duke@435 423 debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
duke@435 424 }
duke@435 425
duke@435 426 //------------------------------frame_size_in_words-----------------------------
duke@435 427 // frame_slots in units of words
duke@435 428 int Compile::frame_size_in_words() const {
duke@435 429 // shift is 0 in LP32 and 1 in LP64
duke@435 430 const int shift = (LogBytesPerWord - LogBytesPerInt);
duke@435 431 int words = _frame_slots >> shift;
duke@435 432 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
duke@435 433 return words;
duke@435 434 }
duke@435 435
duke@435 436 // ============================================================================
duke@435 437 //------------------------------CompileWrapper---------------------------------
duke@435 438 class CompileWrapper : public StackObj {
duke@435 439 Compile *const _compile;
duke@435 440 public:
duke@435 441 CompileWrapper(Compile* compile);
duke@435 442
duke@435 443 ~CompileWrapper();
duke@435 444 };
duke@435 445
duke@435 446 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
duke@435 447 // the Compile* pointer is stored in the current ciEnv:
duke@435 448 ciEnv* env = compile->env();
duke@435 449 assert(env == ciEnv::current(), "must already be a ciEnv active");
duke@435 450 assert(env->compiler_data() == NULL, "compile already active?");
duke@435 451 env->set_compiler_data(compile);
duke@435 452 assert(compile == Compile::current(), "sanity");
duke@435 453
duke@435 454 compile->set_type_dict(NULL);
duke@435 455 compile->set_type_hwm(NULL);
duke@435 456 compile->set_type_last_size(0);
duke@435 457 compile->set_last_tf(NULL, NULL);
duke@435 458 compile->set_indexSet_arena(NULL);
duke@435 459 compile->set_indexSet_free_block_list(NULL);
duke@435 460 compile->init_type_arena();
duke@435 461 Type::Initialize(compile);
duke@435 462 _compile->set_scratch_buffer_blob(NULL);
duke@435 463 _compile->begin_method();
duke@435 464 }
duke@435 465 CompileWrapper::~CompileWrapper() {
duke@435 466 _compile->end_method();
duke@435 467 if (_compile->scratch_buffer_blob() != NULL)
duke@435 468 BufferBlob::free(_compile->scratch_buffer_blob());
duke@435 469 _compile->env()->set_compiler_data(NULL);
duke@435 470 }
duke@435 471
duke@435 472
duke@435 473 //----------------------------print_compile_messages---------------------------
duke@435 474 void Compile::print_compile_messages() {
duke@435 475 #ifndef PRODUCT
duke@435 476 // Check if recompiling
duke@435 477 if (_subsume_loads == false && PrintOpto) {
duke@435 478 // Recompiling without allowing machine instructions to subsume loads
duke@435 479 tty->print_cr("*********************************************************");
duke@435 480 tty->print_cr("** Bailout: Recompile without subsuming loads **");
duke@435 481 tty->print_cr("*********************************************************");
duke@435 482 }
kvn@473 483 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) {
kvn@473 484 // Recompiling without escape analysis
kvn@473 485 tty->print_cr("*********************************************************");
kvn@473 486 tty->print_cr("** Bailout: Recompile without escape analysis **");
kvn@473 487 tty->print_cr("*********************************************************");
kvn@473 488 }
kvn@5110 489 if (_eliminate_boxing != EliminateAutoBox && PrintOpto) {
kvn@5110 490 // Recompiling without boxing elimination
kvn@5110 491 tty->print_cr("*********************************************************");
kvn@5110 492 tty->print_cr("** Bailout: Recompile without boxing elimination **");
kvn@5110 493 tty->print_cr("*********************************************************");
kvn@5110 494 }
duke@435 495 if (env()->break_at_compile()) {
twisti@1040 496 // Open the debugger when compiling this method.
duke@435 497 tty->print("### Breaking when compiling: ");
duke@435 498 method()->print_short_name();
duke@435 499 tty->cr();
duke@435 500 BREAKPOINT;
duke@435 501 }
duke@435 502
duke@435 503 if( PrintOpto ) {
duke@435 504 if (is_osr_compilation()) {
duke@435 505 tty->print("[OSR]%3d", _compile_id);
duke@435 506 } else {
duke@435 507 tty->print("%3d", _compile_id);
duke@435 508 }
duke@435 509 }
duke@435 510 #endif
duke@435 511 }
duke@435 512
duke@435 513
kvn@2414 514 //-----------------------init_scratch_buffer_blob------------------------------
kvn@2414 515 // Construct a temporary BufferBlob and cache it for this compile.
twisti@2350 516 void Compile::init_scratch_buffer_blob(int const_size) {
kvn@2414 517 // If there is already a scratch buffer blob allocated and the
kvn@2414 518 // constant section is big enough, use it. Otherwise free the
kvn@2414 519 // current and allocate a new one.
kvn@2414 520 BufferBlob* blob = scratch_buffer_blob();
kvn@2414 521 if ((blob != NULL) && (const_size <= _scratch_const_size)) {
kvn@2414 522 // Use the current blob.
kvn@2414 523 } else {
kvn@2414 524 if (blob != NULL) {
kvn@2414 525 BufferBlob::free(blob);
kvn@2414 526 }
duke@435 527
kvn@2414 528 ResourceMark rm;
kvn@2414 529 _scratch_const_size = const_size;
kvn@2414 530 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size);
kvn@2414 531 blob = BufferBlob::create("Compile::scratch_buffer", size);
kvn@2414 532 // Record the buffer blob for next time.
kvn@2414 533 set_scratch_buffer_blob(blob);
kvn@2414 534 // Have we run out of code space?
kvn@2414 535 if (scratch_buffer_blob() == NULL) {
kvn@2414 536 // Let CompilerBroker disable further compilations.
kvn@2414 537 record_failure("Not enough space for scratch buffer in CodeCache");
kvn@2414 538 return;
kvn@2414 539 }
kvn@598 540 }
duke@435 541
duke@435 542 // Initialize the relocation buffers
twisti@2103 543 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
duke@435 544 set_scratch_locs_memory(locs_buf);
duke@435 545 }
duke@435 546
duke@435 547
duke@435 548 //-----------------------scratch_emit_size-------------------------------------
duke@435 549 // Helper function that computes size by emitting code
duke@435 550 uint Compile::scratch_emit_size(const Node* n) {
twisti@2350 551 // Start scratch_emit_size section.
twisti@2350 552 set_in_scratch_emit_size(true);
twisti@2350 553
duke@435 554 // Emit into a trash buffer and count bytes emitted.
duke@435 555 // This is a pretty expensive way to compute a size,
duke@435 556 // but it works well enough if seldom used.
duke@435 557 // All common fixed-size instructions are given a size
duke@435 558 // method by the AD file.
duke@435 559 // Note that the scratch buffer blob and locs memory are
duke@435 560 // allocated at the beginning of the compile task, and
duke@435 561 // may be shared by several calls to scratch_emit_size.
duke@435 562 // The allocation of the scratch buffer blob is particularly
duke@435 563 // expensive, since it has to grab the code cache lock.
duke@435 564 BufferBlob* blob = this->scratch_buffer_blob();
duke@435 565 assert(blob != NULL, "Initialize BufferBlob at start");
duke@435 566 assert(blob->size() > MAX_inst_size, "sanity");
duke@435 567 relocInfo* locs_buf = scratch_locs_memory();
twisti@2103 568 address blob_begin = blob->content_begin();
duke@435 569 address blob_end = (address)locs_buf;
twisti@2103 570 assert(blob->content_contains(blob_end), "sanity");
duke@435 571 CodeBuffer buf(blob_begin, blob_end - blob_begin);
twisti@2350 572 buf.initialize_consts_size(_scratch_const_size);
duke@435 573 buf.initialize_stubs_size(MAX_stubs_size);
duke@435 574 assert(locs_buf != NULL, "sanity");
twisti@2350 575 int lsize = MAX_locs_size / 3;
twisti@2350 576 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
twisti@2350 577 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
twisti@2350 578 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
twisti@2350 579
twisti@2350 580 // Do the emission.
kvn@3037 581
kvn@3037 582 Label fakeL; // Fake label for branch instructions.
kvn@3051 583 Label* saveL = NULL;
kvn@3051 584 uint save_bnum = 0;
kvn@3051 585 bool is_branch = n->is_MachBranch();
kvn@3037 586 if (is_branch) {
kvn@3037 587 MacroAssembler masm(&buf);
kvn@3037 588 masm.bind(fakeL);
kvn@3051 589 n->as_MachBranch()->save_label(&saveL, &save_bnum);
kvn@3051 590 n->as_MachBranch()->label_set(&fakeL, 0);
kvn@3037 591 }
duke@435 592 n->emit(buf, this->regalloc());
kvn@3051 593 if (is_branch) // Restore label.
kvn@3051 594 n->as_MachBranch()->label_set(saveL, save_bnum);
twisti@2350 595
twisti@2350 596 // End scratch_emit_size section.
twisti@2350 597 set_in_scratch_emit_size(false);
twisti@2350 598
twisti@2103 599 return buf.insts_size();
duke@435 600 }
duke@435 601
duke@435 602
duke@435 603 // ============================================================================
duke@435 604 //------------------------------Compile standard-------------------------------
duke@435 605 debug_only( int Compile::_debug_idx = 100000; )
duke@435 606
duke@435 607 // Compile a method. entry_bci is -1 for normal compilations and indicates
duke@435 608 // the continuation bci for on stack replacement.
duke@435 609
duke@435 610
kvn@5110 611 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci,
kvn@5110 612 bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing )
duke@435 613 : Phase(Compiler),
duke@435 614 _env(ci_env),
duke@435 615 _log(ci_env->log()),
duke@435 616 _compile_id(ci_env->compile_id()),
duke@435 617 _save_argument_registers(false),
duke@435 618 _stub_name(NULL),
duke@435 619 _stub_function(NULL),
duke@435 620 _stub_entry_point(NULL),
duke@435 621 _method(target),
duke@435 622 _entry_bci(osr_bci),
duke@435 623 _initial_gvn(NULL),
duke@435 624 _for_igvn(NULL),
duke@435 625 _warm_calls(NULL),
duke@435 626 _subsume_loads(subsume_loads),
kvn@473 627 _do_escape_analysis(do_escape_analysis),
kvn@5110 628 _eliminate_boxing(eliminate_boxing),
duke@435 629 _failure_reason(NULL),
duke@435 630 _code_buffer("Compile::Fill_buffer"),
duke@435 631 _orig_pc_slot(0),
duke@435 632 _orig_pc_slot_offset_in_bytes(0),
twisti@1700 633 _has_method_handle_invokes(false),
twisti@2350 634 _mach_constant_base_node(NULL),
duke@435 635 _node_bundling_limit(0),
duke@435 636 _node_bundling_base(NULL),
kvn@1294 637 _java_calls(0),
kvn@1294 638 _inner_loops(0),
twisti@2350 639 _scratch_const_size(-1),
twisti@2350 640 _in_scratch_emit_size(false),
bharadwaj@4315 641 _dead_node_list(comp_arena()),
bharadwaj@4315 642 _dead_node_count(0),
duke@435 643 #ifndef PRODUCT
duke@435 644 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")),
duke@435 645 _printer(IdealGraphPrinter::printer()),
duke@435 646 #endif
roland@4357 647 _congraph(NULL),
roland@4409 648 _late_inlines(comp_arena(), 2, 0, NULL),
roland@4409 649 _string_late_inlines(comp_arena(), 2, 0, NULL),
kvn@5110 650 _boxing_late_inlines(comp_arena(), 2, 0, NULL),
roland@4409 651 _late_inlines_pos(0),
roland@4409 652 _number_of_mh_late_inlines(0),
roland@4409 653 _inlining_progress(false),
roland@4409 654 _inlining_incrementally(false),
roland@4357 655 _print_inlining_list(NULL),
roland@4357 656 _print_inlining(0) {
duke@435 657 C = this;
duke@435 658
duke@435 659 CompileWrapper cw(this);
duke@435 660 #ifndef PRODUCT
duke@435 661 if (TimeCompiler2) {
duke@435 662 tty->print(" ");
duke@435 663 target->holder()->name()->print();
duke@435 664 tty->print(".");
duke@435 665 target->print_short_name();
duke@435 666 tty->print(" ");
duke@435 667 }
duke@435 668 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2);
duke@435 669 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false);
jrose@535 670 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly");
jrose@535 671 if (!print_opto_assembly) {
jrose@535 672 bool print_assembly = (PrintAssembly || _method->should_print_assembly());
jrose@535 673 if (print_assembly && !Disassembler::can_decode()) {
jrose@535 674 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly");
jrose@535 675 print_opto_assembly = true;
jrose@535 676 }
jrose@535 677 }
jrose@535 678 set_print_assembly(print_opto_assembly);
never@802 679 set_parsed_irreducible_loop(false);
duke@435 680 #endif
duke@435 681
duke@435 682 if (ProfileTraps) {
duke@435 683 // Make sure the method being compiled gets its own MDO,
duke@435 684 // so we can at least track the decompile_count().
iveresov@2349 685 method()->ensure_method_data();
duke@435 686 }
duke@435 687
duke@435 688 Init(::AliasLevel);
duke@435 689
duke@435 690
duke@435 691 print_compile_messages();
duke@435 692
duke@435 693 if (UseOldInlining || PrintCompilation NOT_PRODUCT( || PrintOpto) )
duke@435 694 _ilt = InlineTree::build_inline_tree_root();
duke@435 695 else
duke@435 696 _ilt = NULL;
duke@435 697
duke@435 698 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
duke@435 699 assert(num_alias_types() >= AliasIdxRaw, "");
duke@435 700
duke@435 701 #define MINIMUM_NODE_HASH 1023
duke@435 702 // Node list that Iterative GVN will start with
duke@435 703 Unique_Node_List for_igvn(comp_arena());
duke@435 704 set_for_igvn(&for_igvn);
duke@435 705
duke@435 706 // GVN that will be run immediately on new nodes
duke@435 707 uint estimated_size = method()->code_size()*4+64;
duke@435 708 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
duke@435 709 PhaseGVN gvn(node_arena(), estimated_size);
duke@435 710 set_initial_gvn(&gvn);
duke@435 711
kvn@4448 712 if (PrintInlining || PrintIntrinsics NOT_PRODUCT( || PrintOptoInlining)) {
roland@4357 713 _print_inlining_list = new (comp_arena())GrowableArray<PrintInliningBuffer>(comp_arena(), 1, 1, PrintInliningBuffer());
roland@4357 714 }
duke@435 715 { // Scope for timing the parser
duke@435 716 TracePhase t3("parse", &_t_parser, true);
duke@435 717
duke@435 718 // Put top into the hash table ASAP.
duke@435 719 initial_gvn()->transform_no_reclaim(top());
duke@435 720
duke@435 721 // Set up tf(), start(), and find a CallGenerator.
johnc@2781 722 CallGenerator* cg = NULL;
duke@435 723 if (is_osr_compilation()) {
duke@435 724 const TypeTuple *domain = StartOSRNode::osr_domain();
duke@435 725 const TypeTuple *range = TypeTuple::make_range(method()->signature());
duke@435 726 init_tf(TypeFunc::make(domain, range));
kvn@4115 727 StartNode* s = new (this) StartOSRNode(root(), domain);
duke@435 728 initial_gvn()->set_type_bottom(s);
duke@435 729 init_start(s);
duke@435 730 cg = CallGenerator::for_osr(method(), entry_bci());
duke@435 731 } else {
duke@435 732 // Normal case.
duke@435 733 init_tf(TypeFunc::make(method()));
kvn@4115 734 StartNode* s = new (this) StartNode(root(), tf()->domain());
duke@435 735 initial_gvn()->set_type_bottom(s);
duke@435 736 init_start(s);
johnc@2781 737 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) {
johnc@2781 738 // With java.lang.ref.reference.get() we must go through the
johnc@2781 739 // intrinsic when G1 is enabled - even when get() is the root
johnc@2781 740 // method of the compile - so that, if necessary, the value in
johnc@2781 741 // the referent field of the reference object gets recorded by
johnc@2781 742 // the pre-barrier code.
johnc@2781 743 // Specifically, if G1 is enabled, the value in the referent
johnc@2781 744 // field is recorded by the G1 SATB pre barrier. This will
johnc@2781 745 // result in the referent being marked live and the reference
johnc@2781 746 // object removed from the list of discovered references during
johnc@2781 747 // reference processing.
johnc@2781 748 cg = find_intrinsic(method(), false);
johnc@2781 749 }
johnc@2781 750 if (cg == NULL) {
johnc@2781 751 float past_uses = method()->interpreter_invocation_count();
johnc@2781 752 float expected_uses = past_uses;
johnc@2781 753 cg = CallGenerator::for_inline(method(), expected_uses);
johnc@2781 754 }
duke@435 755 }
duke@435 756 if (failing()) return;
duke@435 757 if (cg == NULL) {
duke@435 758 record_method_not_compilable_all_tiers("cannot parse method");
duke@435 759 return;
duke@435 760 }
duke@435 761 JVMState* jvms = build_start_state(start(), tf());
duke@435 762 if ((jvms = cg->generate(jvms)) == NULL) {
duke@435 763 record_method_not_compilable("method parse failed");
duke@435 764 return;
duke@435 765 }
duke@435 766 GraphKit kit(jvms);
duke@435 767
duke@435 768 if (!kit.stopped()) {
duke@435 769 // Accept return values, and transfer control we know not where.
duke@435 770 // This is done by a special, unique ReturnNode bound to root.
duke@435 771 return_values(kit.jvms());
duke@435 772 }
duke@435 773
duke@435 774 if (kit.has_exceptions()) {
duke@435 775 // Any exceptions that escape from this call must be rethrown
duke@435 776 // to whatever caller is dynamically above us on the stack.
duke@435 777 // This is done by a special, unique RethrowNode bound to root.
duke@435 778 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
duke@435 779 }
duke@435 780
roland@4409 781 assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
roland@4409 782
roland@4409 783 if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
roland@4409 784 inline_string_calls(true);
never@1515 785 }
roland@4409 786
roland@4409 787 if (failing()) return;
never@1515 788
never@852 789 print_method("Before RemoveUseless", 3);
never@802 790
duke@435 791 // Remove clutter produced by parsing.
duke@435 792 if (!failing()) {
duke@435 793 ResourceMark rm;
duke@435 794 PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
duke@435 795 }
duke@435 796 }
duke@435 797
duke@435 798 // Note: Large methods are capped off in do_one_bytecode().
duke@435 799 if (failing()) return;
duke@435 800
duke@435 801 // After parsing, node notes are no longer automagic.
duke@435 802 // They must be propagated by register_new_node_with_optimizer(),
duke@435 803 // clone(), or the like.
duke@435 804 set_default_node_notes(NULL);
duke@435 805
duke@435 806 for (;;) {
duke@435 807 int successes = Inline_Warm();
duke@435 808 if (failing()) return;
duke@435 809 if (successes == 0) break;
duke@435 810 }
duke@435 811
duke@435 812 // Drain the list.
duke@435 813 Finish_Warm();
duke@435 814 #ifndef PRODUCT
duke@435 815 if (_printer) {
duke@435 816 _printer->print_inlining(this);
duke@435 817 }
duke@435 818 #endif
duke@435 819
duke@435 820 if (failing()) return;
duke@435 821 NOT_PRODUCT( verify_graph_edges(); )
duke@435 822
duke@435 823 // Now optimize
duke@435 824 Optimize();
duke@435 825 if (failing()) return;
duke@435 826 NOT_PRODUCT( verify_graph_edges(); )
duke@435 827
duke@435 828 #ifndef PRODUCT
duke@435 829 if (PrintIdeal) {
duke@435 830 ttyLocker ttyl; // keep the following output all in one block
duke@435 831 // This output goes directly to the tty, not the compiler log.
duke@435 832 // To enable tools to match it up with the compilation activity,
duke@435 833 // be sure to tag this tty output with the compile ID.
duke@435 834 if (xtty != NULL) {
duke@435 835 xtty->head("ideal compile_id='%d'%s", compile_id(),
duke@435 836 is_osr_compilation() ? " compile_kind='osr'" :
duke@435 837 "");
duke@435 838 }
duke@435 839 root()->dump(9999);
duke@435 840 if (xtty != NULL) {
duke@435 841 xtty->tail("ideal");
duke@435 842 }
duke@435 843 }
duke@435 844 #endif
duke@435 845
duke@435 846 // Now that we know the size of all the monitors we can add a fixed slot
duke@435 847 // for the original deopt pc.
duke@435 848
duke@435 849 _orig_pc_slot = fixed_slots();
duke@435 850 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size);
duke@435 851 set_fixed_slots(next_slot);
duke@435 852
duke@435 853 // Now generate code
duke@435 854 Code_Gen();
duke@435 855 if (failing()) return;
duke@435 856
duke@435 857 // Check if we want to skip execution of all compiled code.
duke@435 858 {
duke@435 859 #ifndef PRODUCT
duke@435 860 if (OptoNoExecute) {
duke@435 861 record_method_not_compilable("+OptoNoExecute"); // Flag as failed
duke@435 862 return;
duke@435 863 }
duke@435 864 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler);
duke@435 865 #endif
duke@435 866
duke@435 867 if (is_osr_compilation()) {
duke@435 868 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
duke@435 869 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
duke@435 870 } else {
duke@435 871 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
duke@435 872 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
duke@435 873 }
duke@435 874
duke@435 875 env()->register_method(_method, _entry_bci,
duke@435 876 &_code_offsets,
duke@435 877 _orig_pc_slot_offset_in_bytes,
duke@435 878 code_buffer(),
duke@435 879 frame_size_in_words(), _oop_map_set,
duke@435 880 &_handler_table, &_inc_table,
duke@435 881 compiler,
duke@435 882 env()->comp_level(),
kvn@4103 883 has_unsafe_access(),
kvn@4103 884 SharedRuntime::is_wide_vector(max_vector_size())
duke@435 885 );
vlivanov@4154 886
vlivanov@4154 887 if (log() != NULL) // Print code cache state into compiler log
vlivanov@4154 888 log()->code_cache_state();
duke@435 889 }
duke@435 890 }
duke@435 891
duke@435 892 //------------------------------Compile----------------------------------------
duke@435 893 // Compile a runtime stub
duke@435 894 Compile::Compile( ciEnv* ci_env,
duke@435 895 TypeFunc_generator generator,
duke@435 896 address stub_function,
duke@435 897 const char *stub_name,
duke@435 898 int is_fancy_jump,
duke@435 899 bool pass_tls,
duke@435 900 bool save_arg_registers,
duke@435 901 bool return_pc )
duke@435 902 : Phase(Compiler),
duke@435 903 _env(ci_env),
duke@435 904 _log(ci_env->log()),
neliasso@4730 905 _compile_id(0),
duke@435 906 _save_argument_registers(save_arg_registers),
duke@435 907 _method(NULL),
duke@435 908 _stub_name(stub_name),
duke@435 909 _stub_function(stub_function),
duke@435 910 _stub_entry_point(NULL),
duke@435 911 _entry_bci(InvocationEntryBci),
duke@435 912 _initial_gvn(NULL),
duke@435 913 _for_igvn(NULL),
duke@435 914 _warm_calls(NULL),
duke@435 915 _orig_pc_slot(0),
duke@435 916 _orig_pc_slot_offset_in_bytes(0),
duke@435 917 _subsume_loads(true),
kvn@473 918 _do_escape_analysis(false),
kvn@5110 919 _eliminate_boxing(false),
duke@435 920 _failure_reason(NULL),
duke@435 921 _code_buffer("Compile::Fill_buffer"),
twisti@1700 922 _has_method_handle_invokes(false),
twisti@2350 923 _mach_constant_base_node(NULL),
duke@435 924 _node_bundling_limit(0),
duke@435 925 _node_bundling_base(NULL),
kvn@1294 926 _java_calls(0),
kvn@1294 927 _inner_loops(0),
duke@435 928 #ifndef PRODUCT
duke@435 929 _trace_opto_output(TraceOptoOutput),
duke@435 930 _printer(NULL),
duke@435 931 #endif
bharadwaj@4315 932 _dead_node_list(comp_arena()),
bharadwaj@4315 933 _dead_node_count(0),
roland@4357 934 _congraph(NULL),
roland@4409 935 _number_of_mh_late_inlines(0),
roland@4409 936 _inlining_progress(false),
roland@4409 937 _inlining_incrementally(false),
roland@4357 938 _print_inlining_list(NULL),
roland@4357 939 _print_inlining(0) {
duke@435 940 C = this;
duke@435 941
duke@435 942 #ifndef PRODUCT
duke@435 943 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false);
duke@435 944 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false);
duke@435 945 set_print_assembly(PrintFrameConverterAssembly);
never@802 946 set_parsed_irreducible_loop(false);
duke@435 947 #endif
duke@435 948 CompileWrapper cw(this);
duke@435 949 Init(/*AliasLevel=*/ 0);
duke@435 950 init_tf((*generator)());
duke@435 951
duke@435 952 {
duke@435 953 // The following is a dummy for the sake of GraphKit::gen_stub
duke@435 954 Unique_Node_List for_igvn(comp_arena());
duke@435 955 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this
duke@435 956 PhaseGVN gvn(Thread::current()->resource_area(),255);
duke@435 957 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
duke@435 958 gvn.transform_no_reclaim(top());
duke@435 959
duke@435 960 GraphKit kit;
duke@435 961 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
duke@435 962 }
duke@435 963
duke@435 964 NOT_PRODUCT( verify_graph_edges(); )
duke@435 965 Code_Gen();
duke@435 966 if (failing()) return;
duke@435 967
duke@435 968
duke@435 969 // Entry point will be accessed using compile->stub_entry_point();
duke@435 970 if (code_buffer() == NULL) {
duke@435 971 Matcher::soft_match_failure();
duke@435 972 } else {
duke@435 973 if (PrintAssembly && (WizardMode || Verbose))
duke@435 974 tty->print_cr("### Stub::%s", stub_name);
duke@435 975
duke@435 976 if (!failing()) {
duke@435 977 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs");
duke@435 978
duke@435 979 // Make the NMethod
duke@435 980 // For now we mark the frame as never safe for profile stackwalking
duke@435 981 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
duke@435 982 code_buffer(),
duke@435 983 CodeOffsets::frame_never_safe,
duke@435 984 // _code_offsets.value(CodeOffsets::Frame_Complete),
duke@435 985 frame_size_in_words(),
duke@435 986 _oop_map_set,
duke@435 987 save_arg_registers);
duke@435 988 assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
duke@435 989
duke@435 990 _stub_entry_point = rs->entry_point();
duke@435 991 }
duke@435 992 }
duke@435 993 }
duke@435 994
duke@435 995 //------------------------------Init-------------------------------------------
duke@435 996 // Prepare for a single compilation
duke@435 997 void Compile::Init(int aliaslevel) {
duke@435 998 _unique = 0;
duke@435 999 _regalloc = NULL;
duke@435 1000
duke@435 1001 _tf = NULL; // filled in later
duke@435 1002 _top = NULL; // cached later
duke@435 1003 _matcher = NULL; // filled in later
duke@435 1004 _cfg = NULL; // filled in later
duke@435 1005
duke@435 1006 set_24_bit_selection_and_mode(Use24BitFP, false);
duke@435 1007
duke@435 1008 _node_note_array = NULL;
duke@435 1009 _default_node_notes = NULL;
duke@435 1010
duke@435 1011 _immutable_memory = NULL; // filled in at first inquiry
duke@435 1012
duke@435 1013 // Globally visible Nodes
duke@435 1014 // First set TOP to NULL to give safe behavior during creation of RootNode
duke@435 1015 set_cached_top_node(NULL);
kvn@4115 1016 set_root(new (this) RootNode());
duke@435 1017 // Now that you have a Root to point to, create the real TOP
kvn@4115 1018 set_cached_top_node( new (this) ConNode(Type::TOP) );
duke@435 1019 set_recent_alloc(NULL, NULL);
duke@435 1020
duke@435 1021 // Create Debug Information Recorder to record scopes, oopmaps, etc.
coleenp@4037 1022 env()->set_oop_recorder(new OopRecorder(env()->arena()));
duke@435 1023 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
duke@435 1024 env()->set_dependencies(new Dependencies(env()));
duke@435 1025
duke@435 1026 _fixed_slots = 0;
duke@435 1027 set_has_split_ifs(false);
duke@435 1028 set_has_loops(has_method() && method()->has_loops()); // first approximation
never@1515 1029 set_has_stringbuilder(false);
kvn@5110 1030 set_has_boxed_value(false);
duke@435 1031 _trap_can_recompile = false; // no traps emitted yet
duke@435 1032 _major_progress = true; // start out assuming good things will happen
duke@435 1033 set_has_unsafe_access(false);
kvn@4103 1034 set_max_vector_size(0);
duke@435 1035 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
duke@435 1036 set_decompile_count(0);
duke@435 1037
rasbold@853 1038 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency"));
iveresov@2138 1039 set_num_loop_opts(LoopOptsCount);
iveresov@2138 1040 set_do_inlining(Inline);
iveresov@2138 1041 set_max_inline_size(MaxInlineSize);
iveresov@2138 1042 set_freq_inline_size(FreqInlineSize);
iveresov@2138 1043 set_do_scheduling(OptoScheduling);
iveresov@2138 1044 set_do_count_invocations(false);
iveresov@2138 1045 set_do_method_data_update(false);
duke@435 1046
duke@435 1047 if (debug_info()->recording_non_safepoints()) {
duke@435 1048 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
duke@435 1049 (comp_arena(), 8, 0, NULL));
duke@435 1050 set_default_node_notes(Node_Notes::make(this));
duke@435 1051 }
duke@435 1052
duke@435 1053 // // -- Initialize types before each compile --
duke@435 1054 // // Update cached type information
duke@435 1055 // if( _method && _method->constants() )
duke@435 1056 // Type::update_loaded_types(_method, _method->constants());
duke@435 1057
duke@435 1058 // Init alias_type map.
kvn@473 1059 if (!_do_escape_analysis && aliaslevel == 3)
duke@435 1060 aliaslevel = 2; // No unique types without escape analysis
duke@435 1061 _AliasLevel = aliaslevel;
duke@435 1062 const int grow_ats = 16;
duke@435 1063 _max_alias_types = grow_ats;
duke@435 1064 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
duke@435 1065 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
duke@435 1066 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
duke@435 1067 {
duke@435 1068 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
duke@435 1069 }
duke@435 1070 // Initialize the first few types.
duke@435 1071 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
duke@435 1072 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
duke@435 1073 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
duke@435 1074 _num_alias_types = AliasIdxRaw+1;
duke@435 1075 // Zero out the alias type cache.
duke@435 1076 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
duke@435 1077 // A NULL adr_type hits in the cache right away. Preload the right answer.
duke@435 1078 probe_alias_cache(NULL)->_index = AliasIdxTop;
duke@435 1079
duke@435 1080 _intrinsics = NULL;
kvn@2040 1081 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
kvn@2040 1082 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
roland@4589 1083 _expensive_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
duke@435 1084 register_library_intrinsics();
duke@435 1085 }
duke@435 1086
duke@435 1087 //---------------------------init_start----------------------------------------
duke@435 1088 // Install the StartNode on this compile object.
duke@435 1089 void Compile::init_start(StartNode* s) {
duke@435 1090 if (failing())
duke@435 1091 return; // already failing
duke@435 1092 assert(s == start(), "");
duke@435 1093 }
duke@435 1094
duke@435 1095 StartNode* Compile::start() const {
duke@435 1096 assert(!failing(), "");
duke@435 1097 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
duke@435 1098 Node* start = root()->fast_out(i);
duke@435 1099 if( start->is_Start() )
duke@435 1100 return start->as_Start();
duke@435 1101 }
duke@435 1102 ShouldNotReachHere();
duke@435 1103 return NULL;
duke@435 1104 }
duke@435 1105
duke@435 1106 //-------------------------------immutable_memory-------------------------------------
duke@435 1107 // Access immutable memory
duke@435 1108 Node* Compile::immutable_memory() {
duke@435 1109 if (_immutable_memory != NULL) {
duke@435 1110 return _immutable_memory;
duke@435 1111 }
duke@435 1112 StartNode* s = start();
duke@435 1113 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
duke@435 1114 Node *p = s->fast_out(i);
duke@435 1115 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
duke@435 1116 _immutable_memory = p;
duke@435 1117 return _immutable_memory;
duke@435 1118 }
duke@435 1119 }
duke@435 1120 ShouldNotReachHere();
duke@435 1121 return NULL;
duke@435 1122 }
duke@435 1123
duke@435 1124 //----------------------set_cached_top_node------------------------------------
duke@435 1125 // Install the cached top node, and make sure Node::is_top works correctly.
duke@435 1126 void Compile::set_cached_top_node(Node* tn) {
duke@435 1127 if (tn != NULL) verify_top(tn);
duke@435 1128 Node* old_top = _top;
duke@435 1129 _top = tn;
duke@435 1130 // Calling Node::setup_is_top allows the nodes the chance to adjust
duke@435 1131 // their _out arrays.
duke@435 1132 if (_top != NULL) _top->setup_is_top();
duke@435 1133 if (old_top != NULL) old_top->setup_is_top();
duke@435 1134 assert(_top == NULL || top()->is_top(), "");
duke@435 1135 }
duke@435 1136
bharadwaj@4315 1137 #ifdef ASSERT
bharadwaj@4315 1138 uint Compile::count_live_nodes_by_graph_walk() {
bharadwaj@4315 1139 Unique_Node_List useful(comp_arena());
bharadwaj@4315 1140 // Get useful node list by walking the graph.
bharadwaj@4315 1141 identify_useful_nodes(useful);
bharadwaj@4315 1142 return useful.size();
bharadwaj@4315 1143 }
bharadwaj@4315 1144
bharadwaj@4315 1145 void Compile::print_missing_nodes() {
bharadwaj@4315 1146
bharadwaj@4315 1147 // Return if CompileLog is NULL and PrintIdealNodeCount is false.
bharadwaj@4315 1148 if ((_log == NULL) && (! PrintIdealNodeCount)) {
bharadwaj@4315 1149 return;
bharadwaj@4315 1150 }
bharadwaj@4315 1151
bharadwaj@4315 1152 // This is an expensive function. It is executed only when the user
bharadwaj@4315 1153 // specifies VerifyIdealNodeCount option or otherwise knows the
bharadwaj@4315 1154 // additional work that needs to be done to identify reachable nodes
bharadwaj@4315 1155 // by walking the flow graph and find the missing ones using
bharadwaj@4315 1156 // _dead_node_list.
bharadwaj@4315 1157
bharadwaj@4315 1158 Unique_Node_List useful(comp_arena());
bharadwaj@4315 1159 // Get useful node list by walking the graph.
bharadwaj@4315 1160 identify_useful_nodes(useful);
bharadwaj@4315 1161
bharadwaj@4315 1162 uint l_nodes = C->live_nodes();
bharadwaj@4315 1163 uint l_nodes_by_walk = useful.size();
bharadwaj@4315 1164
bharadwaj@4315 1165 if (l_nodes != l_nodes_by_walk) {
bharadwaj@4315 1166 if (_log != NULL) {
bharadwaj@4315 1167 _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk)));
bharadwaj@4315 1168 _log->stamp();
bharadwaj@4315 1169 _log->end_head();
bharadwaj@4315 1170 }
bharadwaj@4315 1171 VectorSet& useful_member_set = useful.member_set();
bharadwaj@4315 1172 int last_idx = l_nodes_by_walk;
bharadwaj@4315 1173 for (int i = 0; i < last_idx; i++) {
bharadwaj@4315 1174 if (useful_member_set.test(i)) {
bharadwaj@4315 1175 if (_dead_node_list.test(i)) {
bharadwaj@4315 1176 if (_log != NULL) {
bharadwaj@4315 1177 _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i);
bharadwaj@4315 1178 }
bharadwaj@4315 1179 if (PrintIdealNodeCount) {
bharadwaj@4315 1180 // Print the log message to tty
bharadwaj@4315 1181 tty->print_cr("mismatched_node idx='%d' both live and dead'", i);
bharadwaj@4315 1182 useful.at(i)->dump();
bharadwaj@4315 1183 }
bharadwaj@4315 1184 }
bharadwaj@4315 1185 }
bharadwaj@4315 1186 else if (! _dead_node_list.test(i)) {
bharadwaj@4315 1187 if (_log != NULL) {
bharadwaj@4315 1188 _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i);
bharadwaj@4315 1189 }
bharadwaj@4315 1190 if (PrintIdealNodeCount) {
bharadwaj@4315 1191 // Print the log message to tty
bharadwaj@4315 1192 tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i);
bharadwaj@4315 1193 }
bharadwaj@4315 1194 }
bharadwaj@4315 1195 }
bharadwaj@4315 1196 if (_log != NULL) {
bharadwaj@4315 1197 _log->tail("mismatched_nodes");
bharadwaj@4315 1198 }
bharadwaj@4315 1199 }
bharadwaj@4315 1200 }
bharadwaj@4315 1201 #endif
bharadwaj@4315 1202
duke@435 1203 #ifndef PRODUCT
duke@435 1204 void Compile::verify_top(Node* tn) const {
duke@435 1205 if (tn != NULL) {
duke@435 1206 assert(tn->is_Con(), "top node must be a constant");
duke@435 1207 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
duke@435 1208 assert(tn->in(0) != NULL, "must have live top node");
duke@435 1209 }
duke@435 1210 }
duke@435 1211 #endif
duke@435 1212
duke@435 1213
duke@435 1214 ///-------------------Managing Per-Node Debug & Profile Info-------------------
duke@435 1215
duke@435 1216 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
duke@435 1217 guarantee(arr != NULL, "");
duke@435 1218 int num_blocks = arr->length();
duke@435 1219 if (grow_by < num_blocks) grow_by = num_blocks;
duke@435 1220 int num_notes = grow_by * _node_notes_block_size;
duke@435 1221 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
duke@435 1222 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
duke@435 1223 while (num_notes > 0) {
duke@435 1224 arr->append(notes);
duke@435 1225 notes += _node_notes_block_size;
duke@435 1226 num_notes -= _node_notes_block_size;
duke@435 1227 }
duke@435 1228 assert(num_notes == 0, "exact multiple, please");
duke@435 1229 }
duke@435 1230
duke@435 1231 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
duke@435 1232 if (source == NULL || dest == NULL) return false;
duke@435 1233
duke@435 1234 if (dest->is_Con())
duke@435 1235 return false; // Do not push debug info onto constants.
duke@435 1236
duke@435 1237 #ifdef ASSERT
duke@435 1238 // Leave a bread crumb trail pointing to the original node:
duke@435 1239 if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
duke@435 1240 dest->set_debug_orig(source);
duke@435 1241 }
duke@435 1242 #endif
duke@435 1243
duke@435 1244 if (node_note_array() == NULL)
duke@435 1245 return false; // Not collecting any notes now.
duke@435 1246
duke@435 1247 // This is a copy onto a pre-existing node, which may already have notes.
duke@435 1248 // If both nodes have notes, do not overwrite any pre-existing notes.
duke@435 1249 Node_Notes* source_notes = node_notes_at(source->_idx);
duke@435 1250 if (source_notes == NULL || source_notes->is_clear()) return false;
duke@435 1251 Node_Notes* dest_notes = node_notes_at(dest->_idx);
duke@435 1252 if (dest_notes == NULL || dest_notes->is_clear()) {
duke@435 1253 return set_node_notes_at(dest->_idx, source_notes);
duke@435 1254 }
duke@435 1255
duke@435 1256 Node_Notes merged_notes = (*source_notes);
duke@435 1257 // The order of operations here ensures that dest notes will win...
duke@435 1258 merged_notes.update_from(dest_notes);
duke@435 1259 return set_node_notes_at(dest->_idx, &merged_notes);
duke@435 1260 }
duke@435 1261
duke@435 1262
duke@435 1263 //--------------------------allow_range_check_smearing-------------------------
duke@435 1264 // Gating condition for coalescing similar range checks.
duke@435 1265 // Sometimes we try 'speculatively' replacing a series of a range checks by a
duke@435 1266 // single covering check that is at least as strong as any of them.
duke@435 1267 // If the optimization succeeds, the simplified (strengthened) range check
duke@435 1268 // will always succeed. If it fails, we will deopt, and then give up
duke@435 1269 // on the optimization.
duke@435 1270 bool Compile::allow_range_check_smearing() const {
duke@435 1271 // If this method has already thrown a range-check,
duke@435 1272 // assume it was because we already tried range smearing
duke@435 1273 // and it failed.
duke@435 1274 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
duke@435 1275 return !already_trapped;
duke@435 1276 }
duke@435 1277
duke@435 1278
duke@435 1279 //------------------------------flatten_alias_type-----------------------------
duke@435 1280 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
duke@435 1281 int offset = tj->offset();
duke@435 1282 TypePtr::PTR ptr = tj->ptr();
duke@435 1283
kvn@682 1284 // Known instance (scalarizable allocation) alias only with itself.
kvn@682 1285 bool is_known_inst = tj->isa_oopptr() != NULL &&
kvn@682 1286 tj->is_oopptr()->is_known_instance();
kvn@682 1287
duke@435 1288 // Process weird unsafe references.
duke@435 1289 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
duke@435 1290 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
kvn@682 1291 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
duke@435 1292 tj = TypeOopPtr::BOTTOM;
duke@435 1293 ptr = tj->ptr();
duke@435 1294 offset = tj->offset();
duke@435 1295 }
duke@435 1296
duke@435 1297 // Array pointers need some flattening
duke@435 1298 const TypeAryPtr *ta = tj->isa_aryptr();
kvn@682 1299 if( ta && is_known_inst ) {
kvn@682 1300 if ( offset != Type::OffsetBot &&
kvn@682 1301 offset > arrayOopDesc::length_offset_in_bytes() ) {
kvn@682 1302 offset = Type::OffsetBot; // Flatten constant access into array body only
kvn@682 1303 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id());
kvn@682 1304 }
kvn@682 1305 } else if( ta && _AliasLevel >= 2 ) {
duke@435 1306 // For arrays indexed by constant indices, we flatten the alias
duke@435 1307 // space to include all of the array body. Only the header, klass
duke@435 1308 // and array length can be accessed un-aliased.
duke@435 1309 if( offset != Type::OffsetBot ) {
coleenp@4037 1310 if( ta->const_oop() ) { // MethodData* or Method*
duke@435 1311 offset = Type::OffsetBot; // Flatten constant access into array body
kvn@682 1312 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
duke@435 1313 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
duke@435 1314 // range is OK as-is.
duke@435 1315 tj = ta = TypeAryPtr::RANGE;
duke@435 1316 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
duke@435 1317 tj = TypeInstPtr::KLASS; // all klass loads look alike
duke@435 1318 ta = TypeAryPtr::RANGE; // generic ignored junk
duke@435 1319 ptr = TypePtr::BotPTR;
duke@435 1320 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
duke@435 1321 tj = TypeInstPtr::MARK;
duke@435 1322 ta = TypeAryPtr::RANGE; // generic ignored junk
duke@435 1323 ptr = TypePtr::BotPTR;
duke@435 1324 } else { // Random constant offset into array body
duke@435 1325 offset = Type::OffsetBot; // Flatten constant access into array body
kvn@682 1326 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
duke@435 1327 }
duke@435 1328 }
duke@435 1329 // Arrays of fixed size alias with arrays of unknown size.
duke@435 1330 if (ta->size() != TypeInt::POS) {
duke@435 1331 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
kvn@682 1332 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset);
duke@435 1333 }
duke@435 1334 // Arrays of known objects become arrays of unknown objects.
coleenp@548 1335 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
coleenp@548 1336 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
kvn@682 1337 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
coleenp@548 1338 }
duke@435 1339 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
duke@435 1340 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
kvn@682 1341 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
duke@435 1342 }
duke@435 1343 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
duke@435 1344 // cannot be distinguished by bytecode alone.
duke@435 1345 if (ta->elem() == TypeInt::BOOL) {
duke@435 1346 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
duke@435 1347 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
kvn@682 1348 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
duke@435 1349 }
duke@435 1350 // During the 2nd round of IterGVN, NotNull castings are removed.
duke@435 1351 // Make sure the Bottom and NotNull variants alias the same.
duke@435 1352 // Also, make sure exact and non-exact variants alias the same.
duke@435 1353 if( ptr == TypePtr::NotNull || ta->klass_is_exact() ) {
kvn@2986 1354 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset);
duke@435 1355 }
duke@435 1356 }
duke@435 1357
duke@435 1358 // Oop pointers need some flattening
duke@435 1359 const TypeInstPtr *to = tj->isa_instptr();
duke@435 1360 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
never@2658 1361 ciInstanceKlass *k = to->klass()->as_instance_klass();
duke@435 1362 if( ptr == TypePtr::Constant ) {
never@2658 1363 if (to->klass() != ciEnv::current()->Class_klass() ||
never@2658 1364 offset < k->size_helper() * wordSize) {
never@2658 1365 // No constant oop pointers (such as Strings); they alias with
never@2658 1366 // unknown strings.
never@2658 1367 assert(!is_known_inst, "not scalarizable allocation");
never@2658 1368 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
never@2658 1369 }
kvn@682 1370 } else if( is_known_inst ) {
kvn@598 1371 tj = to; // Keep NotNull and klass_is_exact for instance type
duke@435 1372 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
duke@435 1373 // During the 2nd round of IterGVN, NotNull castings are removed.
duke@435 1374 // Make sure the Bottom and NotNull variants alias the same.
duke@435 1375 // Also, make sure exact and non-exact variants alias the same.
kvn@682 1376 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
duke@435 1377 }
duke@435 1378 // Canonicalize the holder of this field
coleenp@548 1379 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
duke@435 1380 // First handle header references such as a LoadKlassNode, even if the
duke@435 1381 // object's klass is unloaded at compile time (4965979).
kvn@682 1382 if (!is_known_inst) { // Do it only for non-instance types
kvn@682 1383 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset);
kvn@682 1384 }
duke@435 1385 } else if (offset < 0 || offset >= k->size_helper() * wordSize) {
never@2658 1386 // Static fields are in the space above the normal instance
never@2658 1387 // fields in the java.lang.Class instance.
never@2658 1388 if (to->klass() != ciEnv::current()->Class_klass()) {
never@2658 1389 to = NULL;
never@2658 1390 tj = TypeOopPtr::BOTTOM;
never@2658 1391 offset = tj->offset();
never@2658 1392 }
duke@435 1393 } else {
duke@435 1394 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
duke@435 1395 if (!k->equals(canonical_holder) || tj->offset() != offset) {
kvn@682 1396 if( is_known_inst ) {
kvn@682 1397 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id());
kvn@682 1398 } else {
kvn@682 1399 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset);
kvn@682 1400 }
duke@435 1401 }
duke@435 1402 }
duke@435 1403 }
duke@435 1404
duke@435 1405 // Klass pointers to object array klasses need some flattening
duke@435 1406 const TypeKlassPtr *tk = tj->isa_klassptr();
duke@435 1407 if( tk ) {
duke@435 1408 // If we are referencing a field within a Klass, we need
duke@435 1409 // to assume the worst case of an Object. Both exact and
never@3389 1410 // inexact types must flatten to the same alias class so
never@3389 1411 // use NotNull as the PTR.
duke@435 1412 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
duke@435 1413
never@3389 1414 tj = tk = TypeKlassPtr::make(TypePtr::NotNull,
duke@435 1415 TypeKlassPtr::OBJECT->klass(),
duke@435 1416 offset);
duke@435 1417 }
duke@435 1418
duke@435 1419 ciKlass* klass = tk->klass();
duke@435 1420 if( klass->is_obj_array_klass() ) {
duke@435 1421 ciKlass* k = TypeAryPtr::OOPS->klass();
duke@435 1422 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs
duke@435 1423 k = TypeInstPtr::BOTTOM->klass();
duke@435 1424 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset );
duke@435 1425 }
duke@435 1426
duke@435 1427 // Check for precise loads from the primary supertype array and force them
duke@435 1428 // to the supertype cache alias index. Check for generic array loads from
duke@435 1429 // the primary supertype array and also force them to the supertype cache
duke@435 1430 // alias index. Since the same load can reach both, we need to merge
duke@435 1431 // these 2 disparate memories into the same alias class. Since the
duke@435 1432 // primary supertype array is read-only, there's no chance of confusion
duke@435 1433 // where we bypass an array load and an array store.
stefank@3391 1434 int primary_supers_offset = in_bytes(Klass::primary_supers_offset());
never@3389 1435 if (offset == Type::OffsetBot ||
never@3389 1436 (offset >= primary_supers_offset &&
never@3389 1437 offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) ||
stefank@3391 1438 offset == (int)in_bytes(Klass::secondary_super_cache_offset())) {
stefank@3391 1439 offset = in_bytes(Klass::secondary_super_cache_offset());
duke@435 1440 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset );
duke@435 1441 }
duke@435 1442 }
duke@435 1443
duke@435 1444 // Flatten all Raw pointers together.
duke@435 1445 if (tj->base() == Type::RawPtr)
duke@435 1446 tj = TypeRawPtr::BOTTOM;
duke@435 1447
duke@435 1448 if (tj->base() == Type::AnyPtr)
duke@435 1449 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
duke@435 1450
duke@435 1451 // Flatten all to bottom for now
duke@435 1452 switch( _AliasLevel ) {
duke@435 1453 case 0:
duke@435 1454 tj = TypePtr::BOTTOM;
duke@435 1455 break;
duke@435 1456 case 1: // Flatten to: oop, static, field or array
duke@435 1457 switch (tj->base()) {
duke@435 1458 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break;
duke@435 1459 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break;
duke@435 1460 case Type::AryPtr: // do not distinguish arrays at all
duke@435 1461 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break;
duke@435 1462 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
duke@435 1463 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it
duke@435 1464 default: ShouldNotReachHere();
duke@435 1465 }
duke@435 1466 break;
twisti@1040 1467 case 2: // No collapsing at level 2; keep all splits
twisti@1040 1468 case 3: // No collapsing at level 3; keep all splits
duke@435 1469 break;
duke@435 1470 default:
duke@435 1471 Unimplemented();
duke@435 1472 }
duke@435 1473
duke@435 1474 offset = tj->offset();
duke@435 1475 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
duke@435 1476
duke@435 1477 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
duke@435 1478 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
duke@435 1479 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
duke@435 1480 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
duke@435 1481 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
duke@435 1482 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
duke@435 1483 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) ,
duke@435 1484 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
duke@435 1485 assert( tj->ptr() != TypePtr::TopPTR &&
duke@435 1486 tj->ptr() != TypePtr::AnyNull &&
duke@435 1487 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
duke@435 1488 // assert( tj->ptr() != TypePtr::Constant ||
duke@435 1489 // tj->base() == Type::RawPtr ||
duke@435 1490 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
duke@435 1491
duke@435 1492 return tj;
duke@435 1493 }
duke@435 1494
duke@435 1495 void Compile::AliasType::Init(int i, const TypePtr* at) {
duke@435 1496 _index = i;
duke@435 1497 _adr_type = at;
duke@435 1498 _field = NULL;
duke@435 1499 _is_rewritable = true; // default
duke@435 1500 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
kvn@658 1501 if (atoop != NULL && atoop->is_known_instance()) {
kvn@658 1502 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
duke@435 1503 _general_index = Compile::current()->get_alias_index(gt);
duke@435 1504 } else {
duke@435 1505 _general_index = 0;
duke@435 1506 }
duke@435 1507 }
duke@435 1508
duke@435 1509 //---------------------------------print_on------------------------------------
duke@435 1510 #ifndef PRODUCT
duke@435 1511 void Compile::AliasType::print_on(outputStream* st) {
duke@435 1512 if (index() < 10)
duke@435 1513 st->print("@ <%d> ", index());
duke@435 1514 else st->print("@ <%d>", index());
duke@435 1515 st->print(is_rewritable() ? " " : " RO");
duke@435 1516 int offset = adr_type()->offset();
duke@435 1517 if (offset == Type::OffsetBot)
duke@435 1518 st->print(" +any");
duke@435 1519 else st->print(" +%-3d", offset);
duke@435 1520 st->print(" in ");
duke@435 1521 adr_type()->dump_on(st);
duke@435 1522 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
duke@435 1523 if (field() != NULL && tjp) {
duke@435 1524 if (tjp->klass() != field()->holder() ||
duke@435 1525 tjp->offset() != field()->offset_in_bytes()) {
duke@435 1526 st->print(" != ");
duke@435 1527 field()->print();
duke@435 1528 st->print(" ***");
duke@435 1529 }
duke@435 1530 }
duke@435 1531 }
duke@435 1532
duke@435 1533 void print_alias_types() {
duke@435 1534 Compile* C = Compile::current();
duke@435 1535 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
duke@435 1536 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
duke@435 1537 C->alias_type(idx)->print_on(tty);
duke@435 1538 tty->cr();
duke@435 1539 }
duke@435 1540 }
duke@435 1541 #endif
duke@435 1542
duke@435 1543
duke@435 1544 //----------------------------probe_alias_cache--------------------------------
duke@435 1545 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
duke@435 1546 intptr_t key = (intptr_t) adr_type;
duke@435 1547 key ^= key >> logAliasCacheSize;
duke@435 1548 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
duke@435 1549 }
duke@435 1550
duke@435 1551
duke@435 1552 //-----------------------------grow_alias_types--------------------------------
duke@435 1553 void Compile::grow_alias_types() {
duke@435 1554 const int old_ats = _max_alias_types; // how many before?
duke@435 1555 const int new_ats = old_ats; // how many more?
duke@435 1556 const int grow_ats = old_ats+new_ats; // how many now?
duke@435 1557 _max_alias_types = grow_ats;
duke@435 1558 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
duke@435 1559 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
duke@435 1560 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
duke@435 1561 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
duke@435 1562 }
duke@435 1563
duke@435 1564
duke@435 1565 //--------------------------------find_alias_type------------------------------
never@2658 1566 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
duke@435 1567 if (_AliasLevel == 0)
duke@435 1568 return alias_type(AliasIdxBot);
duke@435 1569
duke@435 1570 AliasCacheEntry* ace = probe_alias_cache(adr_type);
duke@435 1571 if (ace->_adr_type == adr_type) {
duke@435 1572 return alias_type(ace->_index);
duke@435 1573 }
duke@435 1574
duke@435 1575 // Handle special cases.
duke@435 1576 if (adr_type == NULL) return alias_type(AliasIdxTop);
duke@435 1577 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
duke@435 1578
duke@435 1579 // Do it the slow way.
duke@435 1580 const TypePtr* flat = flatten_alias_type(adr_type);
duke@435 1581
duke@435 1582 #ifdef ASSERT
duke@435 1583 assert(flat == flatten_alias_type(flat), "idempotent");
duke@435 1584 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr");
duke@435 1585 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
duke@435 1586 const TypeOopPtr* foop = flat->is_oopptr();
kvn@682 1587 // Scalarizable allocations have exact klass always.
kvn@682 1588 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
kvn@682 1589 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
duke@435 1590 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type");
duke@435 1591 }
duke@435 1592 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter");
duke@435 1593 #endif
duke@435 1594
duke@435 1595 int idx = AliasIdxTop;
duke@435 1596 for (int i = 0; i < num_alias_types(); i++) {
duke@435 1597 if (alias_type(i)->adr_type() == flat) {
duke@435 1598 idx = i;
duke@435 1599 break;
duke@435 1600 }
duke@435 1601 }
duke@435 1602
duke@435 1603 if (idx == AliasIdxTop) {
duke@435 1604 if (no_create) return NULL;
duke@435 1605 // Grow the array if necessary.
duke@435 1606 if (_num_alias_types == _max_alias_types) grow_alias_types();
duke@435 1607 // Add a new alias type.
duke@435 1608 idx = _num_alias_types++;
duke@435 1609 _alias_types[idx]->Init(idx, flat);
duke@435 1610 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
duke@435 1611 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
duke@435 1612 if (flat->isa_instptr()) {
duke@435 1613 if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
duke@435 1614 && flat->is_instptr()->klass() == env()->Class_klass())
duke@435 1615 alias_type(idx)->set_rewritable(false);
duke@435 1616 }
duke@435 1617 if (flat->isa_klassptr()) {
stefank@3391 1618 if (flat->offset() == in_bytes(Klass::super_check_offset_offset()))
duke@435 1619 alias_type(idx)->set_rewritable(false);
stefank@3391 1620 if (flat->offset() == in_bytes(Klass::modifier_flags_offset()))
duke@435 1621 alias_type(idx)->set_rewritable(false);
stefank@3391 1622 if (flat->offset() == in_bytes(Klass::access_flags_offset()))
duke@435 1623 alias_type(idx)->set_rewritable(false);
stefank@3391 1624 if (flat->offset() == in_bytes(Klass::java_mirror_offset()))
duke@435 1625 alias_type(idx)->set_rewritable(false);
duke@435 1626 }
duke@435 1627 // %%% (We would like to finalize JavaThread::threadObj_offset(),
duke@435 1628 // but the base pointer type is not distinctive enough to identify
duke@435 1629 // references into JavaThread.)
duke@435 1630
never@2658 1631 // Check for final fields.
duke@435 1632 const TypeInstPtr* tinst = flat->isa_instptr();
coleenp@548 1633 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
never@2658 1634 ciField* field;
never@2658 1635 if (tinst->const_oop() != NULL &&
never@2658 1636 tinst->klass() == ciEnv::current()->Class_klass() &&
never@2658 1637 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) {
never@2658 1638 // static field
never@2658 1639 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
never@2658 1640 field = k->get_field_by_offset(tinst->offset(), true);
never@2658 1641 } else {
never@2658 1642 ciInstanceKlass *k = tinst->klass()->as_instance_klass();
never@2658 1643 field = k->get_field_by_offset(tinst->offset(), false);
never@2658 1644 }
never@2658 1645 assert(field == NULL ||
never@2658 1646 original_field == NULL ||
never@2658 1647 (field->holder() == original_field->holder() &&
never@2658 1648 field->offset() == original_field->offset() &&
never@2658 1649 field->is_static() == original_field->is_static()), "wrong field?");
duke@435 1650 // Set field() and is_rewritable() attributes.
duke@435 1651 if (field != NULL) alias_type(idx)->set_field(field);
duke@435 1652 }
duke@435 1653 }
duke@435 1654
duke@435 1655 // Fill the cache for next time.
duke@435 1656 ace->_adr_type = adr_type;
duke@435 1657 ace->_index = idx;
duke@435 1658 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
duke@435 1659
duke@435 1660 // Might as well try to fill the cache for the flattened version, too.
duke@435 1661 AliasCacheEntry* face = probe_alias_cache(flat);
duke@435 1662 if (face->_adr_type == NULL) {
duke@435 1663 face->_adr_type = flat;
duke@435 1664 face->_index = idx;
duke@435 1665 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
duke@435 1666 }
duke@435 1667
duke@435 1668 return alias_type(idx);
duke@435 1669 }
duke@435 1670
duke@435 1671
duke@435 1672 Compile::AliasType* Compile::alias_type(ciField* field) {
duke@435 1673 const TypeOopPtr* t;
duke@435 1674 if (field->is_static())
never@2658 1675 t = TypeInstPtr::make(field->holder()->java_mirror());
duke@435 1676 else
duke@435 1677 t = TypeOopPtr::make_from_klass_raw(field->holder());
never@2658 1678 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
duke@435 1679 assert(field->is_final() == !atp->is_rewritable(), "must get the rewritable bits correct");
duke@435 1680 return atp;
duke@435 1681 }
duke@435 1682
duke@435 1683
duke@435 1684 //------------------------------have_alias_type--------------------------------
duke@435 1685 bool Compile::have_alias_type(const TypePtr* adr_type) {
duke@435 1686 AliasCacheEntry* ace = probe_alias_cache(adr_type);
duke@435 1687 if (ace->_adr_type == adr_type) {
duke@435 1688 return true;
duke@435 1689 }
duke@435 1690
duke@435 1691 // Handle special cases.
duke@435 1692 if (adr_type == NULL) return true;
duke@435 1693 if (adr_type == TypePtr::BOTTOM) return true;
duke@435 1694
never@2658 1695 return find_alias_type(adr_type, true, NULL) != NULL;
duke@435 1696 }
duke@435 1697
duke@435 1698 //-----------------------------must_alias--------------------------------------
duke@435 1699 // True if all values of the given address type are in the given alias category.
duke@435 1700 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
duke@435 1701 if (alias_idx == AliasIdxBot) return true; // the universal category
duke@435 1702 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP
duke@435 1703 if (alias_idx == AliasIdxTop) return false; // the empty category
duke@435 1704 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
duke@435 1705
duke@435 1706 // the only remaining possible overlap is identity
duke@435 1707 int adr_idx = get_alias_index(adr_type);
duke@435 1708 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
duke@435 1709 assert(adr_idx == alias_idx ||
duke@435 1710 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
duke@435 1711 && adr_type != TypeOopPtr::BOTTOM),
duke@435 1712 "should not be testing for overlap with an unsafe pointer");
duke@435 1713 return adr_idx == alias_idx;
duke@435 1714 }
duke@435 1715
duke@435 1716 //------------------------------can_alias--------------------------------------
duke@435 1717 // True if any values of the given address type are in the given alias category.
duke@435 1718 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
duke@435 1719 if (alias_idx == AliasIdxTop) return false; // the empty category
duke@435 1720 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP
duke@435 1721 if (alias_idx == AliasIdxBot) return true; // the universal category
duke@435 1722 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins
duke@435 1723
duke@435 1724 // the only remaining possible overlap is identity
duke@435 1725 int adr_idx = get_alias_index(adr_type);
duke@435 1726 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
duke@435 1727 return adr_idx == alias_idx;
duke@435 1728 }
duke@435 1729
duke@435 1730
duke@435 1731
duke@435 1732 //---------------------------pop_warm_call-------------------------------------
duke@435 1733 WarmCallInfo* Compile::pop_warm_call() {
duke@435 1734 WarmCallInfo* wci = _warm_calls;
duke@435 1735 if (wci != NULL) _warm_calls = wci->remove_from(wci);
duke@435 1736 return wci;
duke@435 1737 }
duke@435 1738
duke@435 1739 //----------------------------Inline_Warm--------------------------------------
duke@435 1740 int Compile::Inline_Warm() {
duke@435 1741 // If there is room, try to inline some more warm call sites.
duke@435 1742 // %%% Do a graph index compaction pass when we think we're out of space?
duke@435 1743 if (!InlineWarmCalls) return 0;
duke@435 1744
duke@435 1745 int calls_made_hot = 0;
duke@435 1746 int room_to_grow = NodeCountInliningCutoff - unique();
duke@435 1747 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
duke@435 1748 int amount_grown = 0;
duke@435 1749 WarmCallInfo* call;
duke@435 1750 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
duke@435 1751 int est_size = (int)call->size();
duke@435 1752 if (est_size > (room_to_grow - amount_grown)) {
duke@435 1753 // This one won't fit anyway. Get rid of it.
duke@435 1754 call->make_cold();
duke@435 1755 continue;
duke@435 1756 }
duke@435 1757 call->make_hot();
duke@435 1758 calls_made_hot++;
duke@435 1759 amount_grown += est_size;
duke@435 1760 amount_to_grow -= est_size;
duke@435 1761 }
duke@435 1762
duke@435 1763 if (calls_made_hot > 0) set_major_progress();
duke@435 1764 return calls_made_hot;
duke@435 1765 }
duke@435 1766
duke@435 1767
duke@435 1768 //----------------------------Finish_Warm--------------------------------------
duke@435 1769 void Compile::Finish_Warm() {
duke@435 1770 if (!InlineWarmCalls) return;
duke@435 1771 if (failing()) return;
duke@435 1772 if (warm_calls() == NULL) return;
duke@435 1773
duke@435 1774 // Clean up loose ends, if we are out of space for inlining.
duke@435 1775 WarmCallInfo* call;
duke@435 1776 while ((call = pop_warm_call()) != NULL) {
duke@435 1777 call->make_cold();
duke@435 1778 }
duke@435 1779 }
duke@435 1780
cfang@1607 1781 //---------------------cleanup_loop_predicates-----------------------
cfang@1607 1782 // Remove the opaque nodes that protect the predicates so that all unused
cfang@1607 1783 // checks and uncommon_traps will be eliminated from the ideal graph
cfang@1607 1784 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
cfang@1607 1785 if (predicate_count()==0) return;
cfang@1607 1786 for (int i = predicate_count(); i > 0; i--) {
cfang@1607 1787 Node * n = predicate_opaque1_node(i-1);
cfang@1607 1788 assert(n->Opcode() == Op_Opaque1, "must be");
cfang@1607 1789 igvn.replace_node(n, n->in(1));
cfang@1607 1790 }
cfang@1607 1791 assert(predicate_count()==0, "should be clean!");
cfang@1607 1792 }
duke@435 1793
roland@4409 1794 // StringOpts and late inlining of string methods
roland@4409 1795 void Compile::inline_string_calls(bool parse_time) {
roland@4409 1796 {
roland@4409 1797 // remove useless nodes to make the usage analysis simpler
roland@4409 1798 ResourceMark rm;
roland@4409 1799 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
roland@4409 1800 }
roland@4409 1801
roland@4409 1802 {
roland@4409 1803 ResourceMark rm;
roland@4409 1804 print_method("Before StringOpts", 3);
roland@4409 1805 PhaseStringOpts pso(initial_gvn(), for_igvn());
roland@4409 1806 print_method("After StringOpts", 3);
roland@4409 1807 }
roland@4409 1808
roland@4409 1809 // now inline anything that we skipped the first time around
roland@4409 1810 if (!parse_time) {
roland@4409 1811 _late_inlines_pos = _late_inlines.length();
roland@4409 1812 }
roland@4409 1813
roland@4409 1814 while (_string_late_inlines.length() > 0) {
roland@4409 1815 CallGenerator* cg = _string_late_inlines.pop();
roland@4409 1816 cg->do_late_inline();
roland@4409 1817 if (failing()) return;
roland@4409 1818 }
roland@4409 1819 _string_late_inlines.trunc_to(0);
roland@4409 1820 }
roland@4409 1821
kvn@5110 1822 // Late inlining of boxing methods
kvn@5110 1823 void Compile::inline_boxing_calls(PhaseIterGVN& igvn) {
kvn@5110 1824 if (_boxing_late_inlines.length() > 0) {
kvn@5110 1825 assert(has_boxed_value(), "inconsistent");
kvn@5110 1826
kvn@5110 1827 PhaseGVN* gvn = initial_gvn();
kvn@5110 1828 set_inlining_incrementally(true);
kvn@5110 1829
kvn@5110 1830 assert( igvn._worklist.size() == 0, "should be done with igvn" );
kvn@5110 1831 for_igvn()->clear();
kvn@5110 1832 gvn->replace_with(&igvn);
kvn@5110 1833
kvn@5110 1834 while (_boxing_late_inlines.length() > 0) {
kvn@5110 1835 CallGenerator* cg = _boxing_late_inlines.pop();
kvn@5110 1836 cg->do_late_inline();
kvn@5110 1837 if (failing()) return;
kvn@5110 1838 }
kvn@5110 1839 _boxing_late_inlines.trunc_to(0);
kvn@5110 1840
kvn@5110 1841 {
kvn@5110 1842 ResourceMark rm;
kvn@5110 1843 PhaseRemoveUseless pru(gvn, for_igvn());
kvn@5110 1844 }
kvn@5110 1845
kvn@5110 1846 igvn = PhaseIterGVN(gvn);
kvn@5110 1847 igvn.optimize();
kvn@5110 1848
kvn@5110 1849 set_inlining_progress(false);
kvn@5110 1850 set_inlining_incrementally(false);
kvn@5110 1851 }
kvn@5110 1852 }
kvn@5110 1853
roland@4409 1854 void Compile::inline_incrementally_one(PhaseIterGVN& igvn) {
roland@4409 1855 assert(IncrementalInline, "incremental inlining should be on");
roland@4409 1856 PhaseGVN* gvn = initial_gvn();
roland@4409 1857
roland@4409 1858 set_inlining_progress(false);
roland@4409 1859 for_igvn()->clear();
roland@4409 1860 gvn->replace_with(&igvn);
roland@4409 1861
roland@4409 1862 int i = 0;
roland@4409 1863
roland@4409 1864 for (; i <_late_inlines.length() && !inlining_progress(); i++) {
roland@4409 1865 CallGenerator* cg = _late_inlines.at(i);
roland@4409 1866 _late_inlines_pos = i+1;
roland@4409 1867 cg->do_late_inline();
roland@4409 1868 if (failing()) return;
roland@4409 1869 }
roland@4409 1870 int j = 0;
roland@4409 1871 for (; i < _late_inlines.length(); i++, j++) {
roland@4409 1872 _late_inlines.at_put(j, _late_inlines.at(i));
roland@4409 1873 }
roland@4409 1874 _late_inlines.trunc_to(j);
roland@4409 1875
roland@4409 1876 {
roland@4409 1877 ResourceMark rm;
kvn@5110 1878 PhaseRemoveUseless pru(gvn, for_igvn());
roland@4409 1879 }
roland@4409 1880
roland@4409 1881 igvn = PhaseIterGVN(gvn);
roland@4409 1882 }
roland@4409 1883
roland@4409 1884 // Perform incremental inlining until bound on number of live nodes is reached
roland@4409 1885 void Compile::inline_incrementally(PhaseIterGVN& igvn) {
roland@4409 1886 PhaseGVN* gvn = initial_gvn();
roland@4409 1887
roland@4409 1888 set_inlining_incrementally(true);
roland@4409 1889 set_inlining_progress(true);
roland@4409 1890 uint low_live_nodes = 0;
roland@4409 1891
roland@4409 1892 while(inlining_progress() && _late_inlines.length() > 0) {
roland@4409 1893
roland@4409 1894 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
roland@4409 1895 if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
roland@4409 1896 // PhaseIdealLoop is expensive so we only try it once we are
roland@4409 1897 // out of loop and we only try it again if the previous helped
roland@4409 1898 // got the number of nodes down significantly
roland@4409 1899 PhaseIdealLoop ideal_loop( igvn, false, true );
roland@4409 1900 if (failing()) return;
roland@4409 1901 low_live_nodes = live_nodes();
roland@4409 1902 _major_progress = true;
roland@4409 1903 }
roland@4409 1904
roland@4409 1905 if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
roland@4409 1906 break;
roland@4409 1907 }
roland@4409 1908 }
roland@4409 1909
roland@4409 1910 inline_incrementally_one(igvn);
roland@4409 1911
roland@4409 1912 if (failing()) return;
roland@4409 1913
roland@4409 1914 igvn.optimize();
roland@4409 1915
roland@4409 1916 if (failing()) return;
roland@4409 1917 }
roland@4409 1918
roland@4409 1919 assert( igvn._worklist.size() == 0, "should be done with igvn" );
roland@4409 1920
roland@4409 1921 if (_string_late_inlines.length() > 0) {
roland@4409 1922 assert(has_stringbuilder(), "inconsistent");
roland@4409 1923 for_igvn()->clear();
roland@4409 1924 initial_gvn()->replace_with(&igvn);
roland@4409 1925
roland@4409 1926 inline_string_calls(false);
roland@4409 1927
roland@4409 1928 if (failing()) return;
roland@4409 1929
roland@4409 1930 {
roland@4409 1931 ResourceMark rm;
roland@4409 1932 PhaseRemoveUseless pru(initial_gvn(), for_igvn());
roland@4409 1933 }
roland@4409 1934
roland@4409 1935 igvn = PhaseIterGVN(gvn);
roland@4409 1936
roland@4409 1937 igvn.optimize();
roland@4409 1938 }
roland@4409 1939
roland@4409 1940 set_inlining_incrementally(false);
roland@4409 1941 }
roland@4409 1942
roland@4409 1943
duke@435 1944 //------------------------------Optimize---------------------------------------
duke@435 1945 // Given a graph, optimize it.
duke@435 1946 void Compile::Optimize() {
duke@435 1947 TracePhase t1("optimizer", &_t_optimizer, true);
duke@435 1948
duke@435 1949 #ifndef PRODUCT
duke@435 1950 if (env()->break_at_compile()) {
duke@435 1951 BREAKPOINT;
duke@435 1952 }
duke@435 1953
duke@435 1954 #endif
duke@435 1955
duke@435 1956 ResourceMark rm;
duke@435 1957 int loop_opts_cnt;
duke@435 1958
duke@435 1959 NOT_PRODUCT( verify_graph_edges(); )
duke@435 1960
never@657 1961 print_method("After Parsing");
duke@435 1962
duke@435 1963 {
duke@435 1964 // Iterative Global Value Numbering, including ideal transforms
duke@435 1965 // Initialize IterGVN with types and values from parse-time GVN
duke@435 1966 PhaseIterGVN igvn(initial_gvn());
duke@435 1967 {
duke@435 1968 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); )
duke@435 1969 igvn.optimize();
duke@435 1970 }
duke@435 1971
duke@435 1972 print_method("Iter GVN 1", 2);
duke@435 1973
duke@435 1974 if (failing()) return;
duke@435 1975
kvn@5110 1976 {
kvn@5110 1977 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); )
kvn@5110 1978 inline_incrementally(igvn);
kvn@5110 1979 }
roland@4409 1980
roland@4409 1981 print_method("Incremental Inline", 2);
roland@4409 1982
roland@4409 1983 if (failing()) return;
roland@4409 1984
kvn@5110 1985 if (eliminate_boxing()) {
kvn@5110 1986 NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); )
kvn@5110 1987 // Inline valueOf() methods now.
kvn@5110 1988 inline_boxing_calls(igvn);
kvn@5110 1989
kvn@5110 1990 print_method("Incremental Boxing Inline", 2);
kvn@5110 1991
kvn@5110 1992 if (failing()) return;
kvn@5110 1993 }
kvn@5110 1994
roland@4589 1995 // No more new expensive nodes will be added to the list from here
roland@4589 1996 // so keep only the actual candidates for optimizations.
roland@4589 1997 cleanup_expensive_nodes(igvn);
roland@4589 1998
kvn@1989 1999 // Perform escape analysis
kvn@1989 2000 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
kvn@3260 2001 if (has_loops()) {
kvn@3260 2002 // Cleanup graph (remove dead nodes).
kvn@3260 2003 TracePhase t2("idealLoop", &_t_idealLoop, true);
kvn@3260 2004 PhaseIdealLoop ideal_loop( igvn, false, true );
kvn@3260 2005 if (major_progress()) print_method("PhaseIdealLoop before EA", 2);
kvn@3260 2006 if (failing()) return;
kvn@3260 2007 }
kvn@1989 2008 ConnectionGraph::do_analysis(this, &igvn);
kvn@1989 2009
kvn@1989 2010 if (failing()) return;
kvn@1989 2011
kvn@3311 2012 // Optimize out fields loads from scalar replaceable allocations.
kvn@1989 2013 igvn.optimize();
kvn@3260 2014 print_method("Iter GVN after EA", 2);
kvn@1989 2015
kvn@1989 2016 if (failing()) return;
kvn@1989 2017
kvn@3311 2018 if (congraph() != NULL && macro_count() > 0) {
kvn@3651 2019 NOT_PRODUCT( TracePhase t2("macroEliminate", &_t_macroEliminate, TimeCompiler); )
kvn@3311 2020 PhaseMacroExpand mexp(igvn);
kvn@3311 2021 mexp.eliminate_macro_nodes();
kvn@3311 2022 igvn.set_delay_transform(false);
kvn@3311 2023
kvn@3311 2024 igvn.optimize();
kvn@3311 2025 print_method("Iter GVN after eliminating allocations and locks", 2);
kvn@3311 2026
kvn@3311 2027 if (failing()) return;
kvn@3311 2028 }
kvn@1989 2029 }
kvn@1989 2030
duke@435 2031 // Loop transforms on the ideal graph. Range Check Elimination,
duke@435 2032 // peeling, unrolling, etc.
duke@435 2033
duke@435 2034 // Set loop opts counter
duke@435 2035 loop_opts_cnt = num_loop_opts();
duke@435 2036 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
duke@435 2037 {
duke@435 2038 TracePhase t2("idealLoop", &_t_idealLoop, true);
kvn@2727 2039 PhaseIdealLoop ideal_loop( igvn, true );
duke@435 2040 loop_opts_cnt--;
duke@435 2041 if (major_progress()) print_method("PhaseIdealLoop 1", 2);
duke@435 2042 if (failing()) return;
duke@435 2043 }
duke@435 2044 // Loop opts pass if partial peeling occurred in previous pass
duke@435 2045 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) {
duke@435 2046 TracePhase t3("idealLoop", &_t_idealLoop, true);
kvn@2727 2047 PhaseIdealLoop ideal_loop( igvn, false );
duke@435 2048 loop_opts_cnt--;
duke@435 2049 if (major_progress()) print_method("PhaseIdealLoop 2", 2);
duke@435 2050 if (failing()) return;
duke@435 2051 }
duke@435 2052 // Loop opts pass for loop-unrolling before CCP
duke@435 2053 if(major_progress() && (loop_opts_cnt > 0)) {
duke@435 2054 TracePhase t4("idealLoop", &_t_idealLoop, true);
kvn@2727 2055 PhaseIdealLoop ideal_loop( igvn, false );
duke@435 2056 loop_opts_cnt--;
duke@435 2057 if (major_progress()) print_method("PhaseIdealLoop 3", 2);
duke@435 2058 }
never@1356 2059 if (!failing()) {
never@1356 2060 // Verify that last round of loop opts produced a valid graph
never@1356 2061 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
never@1356 2062 PhaseIdealLoop::verify(igvn);
never@1356 2063 }
duke@435 2064 }
duke@435 2065 if (failing()) return;
duke@435 2066
duke@435 2067 // Conditional Constant Propagation;
duke@435 2068 PhaseCCP ccp( &igvn );
duke@435 2069 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
duke@435 2070 {
duke@435 2071 TracePhase t2("ccp", &_t_ccp, true);
duke@435 2072 ccp.do_transform();
duke@435 2073 }
duke@435 2074 print_method("PhaseCPP 1", 2);
duke@435 2075
duke@435 2076 assert( true, "Break here to ccp.dump_old2new_map()");
duke@435 2077
duke@435 2078 // Iterative Global Value Numbering, including ideal transforms
duke@435 2079 {
duke@435 2080 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); )
duke@435 2081 igvn = ccp;
duke@435 2082 igvn.optimize();
duke@435 2083 }
duke@435 2084
duke@435 2085 print_method("Iter GVN 2", 2);
duke@435 2086
duke@435 2087 if (failing()) return;
duke@435 2088
duke@435 2089 // Loop transforms on the ideal graph. Range Check Elimination,
duke@435 2090 // peeling, unrolling, etc.
duke@435 2091 if(loop_opts_cnt > 0) {
duke@435 2092 debug_only( int cnt = 0; );
duke@435 2093 while(major_progress() && (loop_opts_cnt > 0)) {
duke@435 2094 TracePhase t2("idealLoop", &_t_idealLoop, true);
duke@435 2095 assert( cnt++ < 40, "infinite cycle in loop optimization" );
kvn@2727 2096 PhaseIdealLoop ideal_loop( igvn, true);
duke@435 2097 loop_opts_cnt--;
duke@435 2098 if (major_progress()) print_method("PhaseIdealLoop iterations", 2);
duke@435 2099 if (failing()) return;
duke@435 2100 }
duke@435 2101 }
never@1356 2102
never@1356 2103 {
never@1356 2104 // Verify that all previous optimizations produced a valid graph
never@1356 2105 // at least to this point, even if no loop optimizations were done.
never@1356 2106 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
never@1356 2107 PhaseIdealLoop::verify(igvn);
never@1356 2108 }
never@1356 2109
duke@435 2110 {
duke@435 2111 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); )
duke@435 2112 PhaseMacroExpand mex(igvn);
duke@435 2113 if (mex.expand_macro_nodes()) {
duke@435 2114 assert(failing(), "must bail out w/ explicit message");
duke@435 2115 return;
duke@435 2116 }
duke@435 2117 }
duke@435 2118
duke@435 2119 } // (End scope of igvn; run destructor if necessary for asserts.)
duke@435 2120
kvn@4448 2121 dump_inlining();
duke@435 2122 // A method with only infinite loops has no edges entering loops from root
duke@435 2123 {
duke@435 2124 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); )
duke@435 2125 if (final_graph_reshaping()) {
duke@435 2126 assert(failing(), "must bail out w/ explicit message");
duke@435 2127 return;
duke@435 2128 }
duke@435 2129 }
duke@435 2130
duke@435 2131 print_method("Optimize finished", 2);
duke@435 2132 }
duke@435 2133
duke@435 2134
duke@435 2135 //------------------------------Code_Gen---------------------------------------
duke@435 2136 // Given a graph, generate code for it
duke@435 2137 void Compile::Code_Gen() {
duke@435 2138 if (failing()) return;
duke@435 2139
duke@435 2140 // Perform instruction selection. You might think we could reclaim Matcher
duke@435 2141 // memory PDQ, but actually the Matcher is used in generating spill code.
duke@435 2142 // Internals of the Matcher (including some VectorSets) must remain live
duke@435 2143 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
duke@435 2144 // set a bit in reclaimed memory.
duke@435 2145
duke@435 2146 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
duke@435 2147 // nodes. Mapping is only valid at the root of each matched subtree.
duke@435 2148 NOT_PRODUCT( verify_graph_edges(); )
duke@435 2149
duke@435 2150 Node_List proj_list;
duke@435 2151 Matcher m(proj_list);
duke@435 2152 _matcher = &m;
duke@435 2153 {
duke@435 2154 TracePhase t2("matcher", &_t_matcher, true);
duke@435 2155 m.match();
duke@435 2156 }
duke@435 2157 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
duke@435 2158 // nodes. Mapping is only valid at the root of each matched subtree.
duke@435 2159 NOT_PRODUCT( verify_graph_edges(); )
duke@435 2160
duke@435 2161 // If you have too many nodes, or if matching has failed, bail out
duke@435 2162 check_node_count(0, "out of nodes matching instructions");
duke@435 2163 if (failing()) return;
duke@435 2164
duke@435 2165 // Build a proper-looking CFG
duke@435 2166 PhaseCFG cfg(node_arena(), root(), m);
duke@435 2167 _cfg = &cfg;
duke@435 2168 {
duke@435 2169 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); )
duke@435 2170 cfg.Dominators();
duke@435 2171 if (failing()) return;
duke@435 2172
duke@435 2173 NOT_PRODUCT( verify_graph_edges(); )
duke@435 2174
duke@435 2175 cfg.Estimate_Block_Frequency();
duke@435 2176 cfg.GlobalCodeMotion(m,unique(),proj_list);
never@3654 2177 if (failing()) return;
duke@435 2178
duke@435 2179 print_method("Global code motion", 2);
duke@435 2180
duke@435 2181 NOT_PRODUCT( verify_graph_edges(); )
duke@435 2182
duke@435 2183 debug_only( cfg.verify(); )
duke@435 2184 }
duke@435 2185 NOT_PRODUCT( verify_graph_edges(); )
duke@435 2186
neliasso@4949 2187 PhaseChaitin regalloc(unique(), cfg, m);
duke@435 2188 _regalloc = &regalloc;
duke@435 2189 {
duke@435 2190 TracePhase t2("regalloc", &_t_registerAllocation, true);
duke@435 2191 // Perform register allocation. After Chaitin, use-def chains are
duke@435 2192 // no longer accurate (at spill code) and so must be ignored.
duke@435 2193 // Node->LRG->reg mappings are still accurate.
duke@435 2194 _regalloc->Register_Allocate();
duke@435 2195
duke@435 2196 // Bail out if the allocator builds too many nodes
neliasso@4949 2197 if (failing()) {
neliasso@4949 2198 return;
neliasso@4949 2199 }
duke@435 2200 }
duke@435 2201
duke@435 2202 // Prior to register allocation we kept empty basic blocks in case the
duke@435 2203 // the allocator needed a place to spill. After register allocation we
duke@435 2204 // are not adding any new instructions. If any basic block is empty, we
duke@435 2205 // can now safely remove it.
duke@435 2206 {
rasbold@853 2207 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); )
rasbold@853 2208 cfg.remove_empty();
rasbold@853 2209 if (do_freq_based_layout()) {
rasbold@853 2210 PhaseBlockLayout layout(cfg);
rasbold@853 2211 } else {
rasbold@853 2212 cfg.set_loop_alignment();
rasbold@853 2213 }
rasbold@853 2214 cfg.fixup_flow();
duke@435 2215 }
duke@435 2216
duke@435 2217 // Apply peephole optimizations
duke@435 2218 if( OptoPeephole ) {
duke@435 2219 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); )
duke@435 2220 PhasePeephole peep( _regalloc, cfg);
duke@435 2221 peep.do_transform();
duke@435 2222 }
duke@435 2223
duke@435 2224 // Convert Nodes to instruction bits in a buffer
duke@435 2225 {
duke@435 2226 // %%%% workspace merge brought two timers together for one job
duke@435 2227 TracePhase t2a("output", &_t_output, true);
duke@435 2228 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); )
duke@435 2229 Output();
duke@435 2230 }
duke@435 2231
never@657 2232 print_method("Final Code");
duke@435 2233
duke@435 2234 // He's dead, Jim.
duke@435 2235 _cfg = (PhaseCFG*)0xdeadbeef;
duke@435 2236 _regalloc = (PhaseChaitin*)0xdeadbeef;
duke@435 2237 }
duke@435 2238
duke@435 2239
duke@435 2240 //------------------------------dump_asm---------------------------------------
duke@435 2241 // Dump formatted assembly
duke@435 2242 #ifndef PRODUCT
duke@435 2243 void Compile::dump_asm(int *pcs, uint pc_limit) {
duke@435 2244 bool cut_short = false;
duke@435 2245 tty->print_cr("#");
duke@435 2246 tty->print("# "); _tf->dump(); tty->cr();
duke@435 2247 tty->print_cr("#");
duke@435 2248
duke@435 2249 // For all blocks
duke@435 2250 int pc = 0x0; // Program counter
duke@435 2251 char starts_bundle = ' ';
duke@435 2252 _regalloc->dump_frame();
duke@435 2253
duke@435 2254 Node *n = NULL;
duke@435 2255 for( uint i=0; i<_cfg->_num_blocks; i++ ) {
duke@435 2256 if (VMThread::should_terminate()) { cut_short = true; break; }
duke@435 2257 Block *b = _cfg->_blocks[i];
duke@435 2258 if (b->is_connector() && !Verbose) continue;
duke@435 2259 n = b->_nodes[0];
duke@435 2260 if (pcs && n->_idx < pc_limit)
duke@435 2261 tty->print("%3.3x ", pcs[n->_idx]);
duke@435 2262 else
duke@435 2263 tty->print(" ");
duke@435 2264 b->dump_head( &_cfg->_bbs );
duke@435 2265 if (b->is_connector()) {
duke@435 2266 tty->print_cr(" # Empty connector block");
duke@435 2267 } else if (b->num_preds() == 2 && b->pred(1)->is_CatchProj() && b->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
duke@435 2268 tty->print_cr(" # Block is sole successor of call");
duke@435 2269 }
duke@435 2270
duke@435 2271 // For all instructions
duke@435 2272 Node *delay = NULL;
duke@435 2273 for( uint j = 0; j<b->_nodes.size(); j++ ) {
duke@435 2274 if (VMThread::should_terminate()) { cut_short = true; break; }
duke@435 2275 n = b->_nodes[j];
duke@435 2276 if (valid_bundle_info(n)) {
duke@435 2277 Bundle *bundle = node_bundling(n);
duke@435 2278 if (bundle->used_in_unconditional_delay()) {
duke@435 2279 delay = n;
duke@435 2280 continue;
duke@435 2281 }
duke@435 2282 if (bundle->starts_bundle())
duke@435 2283 starts_bundle = '+';
duke@435 2284 }
duke@435 2285
coleenp@548 2286 if (WizardMode) n->dump();
coleenp@548 2287
duke@435 2288 if( !n->is_Region() && // Dont print in the Assembly
duke@435 2289 !n->is_Phi() && // a few noisely useless nodes
duke@435 2290 !n->is_Proj() &&
duke@435 2291 !n->is_MachTemp() &&
kvn@1535 2292 !n->is_SafePointScalarObject() &&
duke@435 2293 !n->is_Catch() && // Would be nice to print exception table targets
duke@435 2294 !n->is_MergeMem() && // Not very interesting
duke@435 2295 !n->is_top() && // Debug info table constants
duke@435 2296 !(n->is_Con() && !n->is_Mach())// Debug info table constants
duke@435 2297 ) {
duke@435 2298 if (pcs && n->_idx < pc_limit)
duke@435 2299 tty->print("%3.3x", pcs[n->_idx]);
duke@435 2300 else
duke@435 2301 tty->print(" ");
duke@435 2302 tty->print(" %c ", starts_bundle);
duke@435 2303 starts_bundle = ' ';
duke@435 2304 tty->print("\t");
duke@435 2305 n->format(_regalloc, tty);
duke@435 2306 tty->cr();
duke@435 2307 }
duke@435 2308
duke@435 2309 // If we have an instruction with a delay slot, and have seen a delay,
duke@435 2310 // then back up and print it
duke@435 2311 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
duke@435 2312 assert(delay != NULL, "no unconditional delay instruction");
coleenp@548 2313 if (WizardMode) delay->dump();
coleenp@548 2314
duke@435 2315 if (node_bundling(delay)->starts_bundle())
duke@435 2316 starts_bundle = '+';
duke@435 2317 if (pcs && n->_idx < pc_limit)
duke@435 2318 tty->print("%3.3x", pcs[n->_idx]);
duke@435 2319 else
duke@435 2320 tty->print(" ");
duke@435 2321 tty->print(" %c ", starts_bundle);
duke@435 2322 starts_bundle = ' ';
duke@435 2323 tty->print("\t");
duke@435 2324 delay->format(_regalloc, tty);
duke@435 2325 tty->print_cr("");
duke@435 2326 delay = NULL;
duke@435 2327 }
duke@435 2328
duke@435 2329 // Dump the exception table as well
duke@435 2330 if( n->is_Catch() && (Verbose || WizardMode) ) {
duke@435 2331 // Print the exception table for this offset
duke@435 2332 _handler_table.print_subtable_for(pc);
duke@435 2333 }
duke@435 2334 }
duke@435 2335
duke@435 2336 if (pcs && n->_idx < pc_limit)
duke@435 2337 tty->print_cr("%3.3x", pcs[n->_idx]);
duke@435 2338 else
duke@435 2339 tty->print_cr("");
duke@435 2340
duke@435 2341 assert(cut_short || delay == NULL, "no unconditional delay branch");
duke@435 2342
duke@435 2343 } // End of per-block dump
duke@435 2344 tty->print_cr("");
duke@435 2345
duke@435 2346 if (cut_short) tty->print_cr("*** disassembly is cut short ***");
duke@435 2347 }
duke@435 2348 #endif
duke@435 2349
duke@435 2350 //------------------------------Final_Reshape_Counts---------------------------
duke@435 2351 // This class defines counters to help identify when a method
duke@435 2352 // may/must be executed using hardware with only 24-bit precision.
duke@435 2353 struct Final_Reshape_Counts : public StackObj {
duke@435 2354 int _call_count; // count non-inlined 'common' calls
duke@435 2355 int _float_count; // count float ops requiring 24-bit precision
duke@435 2356 int _double_count; // count double ops requiring more precision
duke@435 2357 int _java_call_count; // count non-inlined 'java' calls
kvn@1294 2358 int _inner_loop_count; // count loops which need alignment
duke@435 2359 VectorSet _visited; // Visitation flags
duke@435 2360 Node_List _tests; // Set of IfNodes & PCTableNodes
duke@435 2361
duke@435 2362 Final_Reshape_Counts() :
kvn@1294 2363 _call_count(0), _float_count(0), _double_count(0),
kvn@1294 2364 _java_call_count(0), _inner_loop_count(0),
duke@435 2365 _visited( Thread::current()->resource_area() ) { }
duke@435 2366
duke@435 2367 void inc_call_count () { _call_count ++; }
duke@435 2368 void inc_float_count () { _float_count ++; }
duke@435 2369 void inc_double_count() { _double_count++; }
duke@435 2370 void inc_java_call_count() { _java_call_count++; }
kvn@1294 2371 void inc_inner_loop_count() { _inner_loop_count++; }
duke@435 2372
duke@435 2373 int get_call_count () const { return _call_count ; }
duke@435 2374 int get_float_count () const { return _float_count ; }
duke@435 2375 int get_double_count() const { return _double_count; }
duke@435 2376 int get_java_call_count() const { return _java_call_count; }
kvn@1294 2377 int get_inner_loop_count() const { return _inner_loop_count; }
duke@435 2378 };
duke@435 2379
mikael@4889 2380 #ifdef ASSERT
duke@435 2381 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
duke@435 2382 ciInstanceKlass *k = tp->klass()->as_instance_klass();
duke@435 2383 // Make sure the offset goes inside the instance layout.
coleenp@548 2384 return k->contains_field_offset(tp->offset());
duke@435 2385 // Note that OffsetBot and OffsetTop are very negative.
duke@435 2386 }
mikael@4889 2387 #endif
duke@435 2388
never@2780 2389 // Eliminate trivially redundant StoreCMs and accumulate their
never@2780 2390 // precedence edges.
bharadwaj@4315 2391 void Compile::eliminate_redundant_card_marks(Node* n) {
never@2780 2392 assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
never@2780 2393 if (n->in(MemNode::Address)->outcnt() > 1) {
never@2780 2394 // There are multiple users of the same address so it might be
never@2780 2395 // possible to eliminate some of the StoreCMs
never@2780 2396 Node* mem = n->in(MemNode::Memory);
never@2780 2397 Node* adr = n->in(MemNode::Address);
never@2780 2398 Node* val = n->in(MemNode::ValueIn);
never@2780 2399 Node* prev = n;
never@2780 2400 bool done = false;
never@2780 2401 // Walk the chain of StoreCMs eliminating ones that match. As
never@2780 2402 // long as it's a chain of single users then the optimization is
never@2780 2403 // safe. Eliminating partially redundant StoreCMs would require
never@2780 2404 // cloning copies down the other paths.
never@2780 2405 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
never@2780 2406 if (adr == mem->in(MemNode::Address) &&
never@2780 2407 val == mem->in(MemNode::ValueIn)) {
never@2780 2408 // redundant StoreCM
never@2780 2409 if (mem->req() > MemNode::OopStore) {
never@2780 2410 // Hasn't been processed by this code yet.
never@2780 2411 n->add_prec(mem->in(MemNode::OopStore));
never@2780 2412 } else {
never@2780 2413 // Already converted to precedence edge
never@2780 2414 for (uint i = mem->req(); i < mem->len(); i++) {
never@2780 2415 // Accumulate any precedence edges
never@2780 2416 if (mem->in(i) != NULL) {
never@2780 2417 n->add_prec(mem->in(i));
never@2780 2418 }
never@2780 2419 }
never@2780 2420 // Everything above this point has been processed.
never@2780 2421 done = true;
never@2780 2422 }
never@2780 2423 // Eliminate the previous StoreCM
never@2780 2424 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
never@2780 2425 assert(mem->outcnt() == 0, "should be dead");
bharadwaj@4315 2426 mem->disconnect_inputs(NULL, this);
never@2780 2427 } else {
never@2780 2428 prev = mem;
never@2780 2429 }
never@2780 2430 mem = prev->in(MemNode::Memory);
never@2780 2431 }
never@2780 2432 }
never@2780 2433 }
never@2780 2434
duke@435 2435 //------------------------------final_graph_reshaping_impl----------------------
duke@435 2436 // Implement items 1-5 from final_graph_reshaping below.
bharadwaj@4315 2437 void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) {
duke@435 2438
kvn@603 2439 if ( n->outcnt() == 0 ) return; // dead node
duke@435 2440 uint nop = n->Opcode();
duke@435 2441
duke@435 2442 // Check for 2-input instruction with "last use" on right input.
duke@435 2443 // Swap to left input. Implements item (2).
duke@435 2444 if( n->req() == 3 && // two-input instruction
duke@435 2445 n->in(1)->outcnt() > 1 && // left use is NOT a last use
duke@435 2446 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
duke@435 2447 n->in(2)->outcnt() == 1 &&// right use IS a last use
duke@435 2448 !n->in(2)->is_Con() ) { // right use is not a constant
duke@435 2449 // Check for commutative opcode
duke@435 2450 switch( nop ) {
duke@435 2451 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
duke@435 2452 case Op_MaxI: case Op_MinI:
duke@435 2453 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
duke@435 2454 case Op_AndL: case Op_XorL: case Op_OrL:
duke@435 2455 case Op_AndI: case Op_XorI: case Op_OrI: {
duke@435 2456 // Move "last use" input to left by swapping inputs
duke@435 2457 n->swap_edges(1, 2);
duke@435 2458 break;
duke@435 2459 }
duke@435 2460 default:
duke@435 2461 break;
duke@435 2462 }
duke@435 2463 }
duke@435 2464
kvn@1964 2465 #ifdef ASSERT
kvn@1964 2466 if( n->is_Mem() ) {
bharadwaj@4315 2467 int alias_idx = get_alias_index(n->as_Mem()->adr_type());
kvn@1964 2468 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
kvn@1964 2469 // oop will be recorded in oop map if load crosses safepoint
kvn@1964 2470 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
kvn@1964 2471 LoadNode::is_immutable_value(n->in(MemNode::Address))),
kvn@1964 2472 "raw memory operations should have control edge");
kvn@1964 2473 }
kvn@1964 2474 #endif
duke@435 2475 // Count FPU ops and common calls, implements item (3)
duke@435 2476 switch( nop ) {
duke@435 2477 // Count all float operations that may use FPU
duke@435 2478 case Op_AddF:
duke@435 2479 case Op_SubF:
duke@435 2480 case Op_MulF:
duke@435 2481 case Op_DivF:
duke@435 2482 case Op_NegF:
duke@435 2483 case Op_ModF:
duke@435 2484 case Op_ConvI2F:
duke@435 2485 case Op_ConF:
duke@435 2486 case Op_CmpF:
duke@435 2487 case Op_CmpF3:
duke@435 2488 // case Op_ConvL2F: // longs are split into 32-bit halves
kvn@1294 2489 frc.inc_float_count();
duke@435 2490 break;
duke@435 2491
duke@435 2492 case Op_ConvF2D:
duke@435 2493 case Op_ConvD2F:
kvn@1294 2494 frc.inc_float_count();
kvn@1294 2495 frc.inc_double_count();
duke@435 2496 break;
duke@435 2497
duke@435 2498 // Count all double operations that may use FPU
duke@435 2499 case Op_AddD:
duke@435 2500 case Op_SubD:
duke@435 2501 case Op_MulD:
duke@435 2502 case Op_DivD:
duke@435 2503 case Op_NegD:
duke@435 2504 case Op_ModD:
duke@435 2505 case Op_ConvI2D:
duke@435 2506 case Op_ConvD2I:
duke@435 2507 // case Op_ConvL2D: // handled by leaf call
duke@435 2508 // case Op_ConvD2L: // handled by leaf call
duke@435 2509 case Op_ConD:
duke@435 2510 case Op_CmpD:
duke@435 2511 case Op_CmpD3:
kvn@1294 2512 frc.inc_double_count();
duke@435 2513 break;
duke@435 2514 case Op_Opaque1: // Remove Opaque Nodes before matching
duke@435 2515 case Op_Opaque2: // Remove Opaque Nodes before matching
bharadwaj@4315 2516 n->subsume_by(n->in(1), this);
duke@435 2517 break;
duke@435 2518 case Op_CallStaticJava:
duke@435 2519 case Op_CallJava:
duke@435 2520 case Op_CallDynamicJava:
kvn@1294 2521 frc.inc_java_call_count(); // Count java call site;
duke@435 2522 case Op_CallRuntime:
duke@435 2523 case Op_CallLeaf:
duke@435 2524 case Op_CallLeafNoFP: {
duke@435 2525 assert( n->is_Call(), "" );
duke@435 2526 CallNode *call = n->as_Call();
duke@435 2527 // Count call sites where the FP mode bit would have to be flipped.
duke@435 2528 // Do not count uncommon runtime calls:
duke@435 2529 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
duke@435 2530 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
duke@435 2531 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) {
kvn@1294 2532 frc.inc_call_count(); // Count the call site
duke@435 2533 } else { // See if uncommon argument is shared
duke@435 2534 Node *n = call->in(TypeFunc::Parms);
duke@435 2535 int nop = n->Opcode();
duke@435 2536 // Clone shared simple arguments to uncommon calls, item (1).
duke@435 2537 if( n->outcnt() > 1 &&
duke@435 2538 !n->is_Proj() &&
duke@435 2539 nop != Op_CreateEx &&
duke@435 2540 nop != Op_CheckCastPP &&
kvn@766 2541 nop != Op_DecodeN &&
roland@4159 2542 nop != Op_DecodeNKlass &&
duke@435 2543 !n->is_Mem() ) {
duke@435 2544 Node *x = n->clone();
duke@435 2545 call->set_req( TypeFunc::Parms, x );
duke@435 2546 }
duke@435 2547 }
duke@435 2548 break;
duke@435 2549 }
duke@435 2550
duke@435 2551 case Op_StoreD:
duke@435 2552 case Op_LoadD:
duke@435 2553 case Op_LoadD_unaligned:
kvn@1294 2554 frc.inc_double_count();
duke@435 2555 goto handle_mem;
duke@435 2556 case Op_StoreF:
duke@435 2557 case Op_LoadF:
kvn@1294 2558 frc.inc_float_count();
duke@435 2559 goto handle_mem;
duke@435 2560
never@2780 2561 case Op_StoreCM:
never@2780 2562 {
never@2780 2563 // Convert OopStore dependence into precedence edge
never@2780 2564 Node* prec = n->in(MemNode::OopStore);
never@2780 2565 n->del_req(MemNode::OopStore);
never@2780 2566 n->add_prec(prec);
never@2780 2567 eliminate_redundant_card_marks(n);
never@2780 2568 }
never@2780 2569
never@2780 2570 // fall through
never@2780 2571
duke@435 2572 case Op_StoreB:
duke@435 2573 case Op_StoreC:
duke@435 2574 case Op_StorePConditional:
duke@435 2575 case Op_StoreI:
duke@435 2576 case Op_StoreL:
kvn@855 2577 case Op_StoreIConditional:
duke@435 2578 case Op_StoreLConditional:
duke@435 2579 case Op_CompareAndSwapI:
duke@435 2580 case Op_CompareAndSwapL:
duke@435 2581 case Op_CompareAndSwapP:
coleenp@548 2582 case Op_CompareAndSwapN:
roland@4106 2583 case Op_GetAndAddI:
roland@4106 2584 case Op_GetAndAddL:
roland@4106 2585 case Op_GetAndSetI:
roland@4106 2586 case Op_GetAndSetL:
roland@4106 2587 case Op_GetAndSetP:
roland@4106 2588 case Op_GetAndSetN:
duke@435 2589 case Op_StoreP:
coleenp@548 2590 case Op_StoreN:
roland@4159 2591 case Op_StoreNKlass:
duke@435 2592 case Op_LoadB:
twisti@1059 2593 case Op_LoadUB:
twisti@993 2594 case Op_LoadUS:
duke@435 2595 case Op_LoadI:
duke@435 2596 case Op_LoadKlass:
kvn@599 2597 case Op_LoadNKlass:
duke@435 2598 case Op_LoadL:
duke@435 2599 case Op_LoadL_unaligned:
duke@435 2600 case Op_LoadPLocked:
duke@435 2601 case Op_LoadP:
coleenp@548 2602 case Op_LoadN:
duke@435 2603 case Op_LoadRange:
duke@435 2604 case Op_LoadS: {
duke@435 2605 handle_mem:
duke@435 2606 #ifdef ASSERT
duke@435 2607 if( VerifyOptoOopOffsets ) {
duke@435 2608 assert( n->is_Mem(), "" );
duke@435 2609 MemNode *mem = (MemNode*)n;
duke@435 2610 // Check to see if address types have grounded out somehow.
duke@435 2611 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
duke@435 2612 assert( !tp || oop_offset_is_sane(tp), "" );
duke@435 2613 }
duke@435 2614 #endif
duke@435 2615 break;
duke@435 2616 }
duke@435 2617
duke@435 2618 case Op_AddP: { // Assert sane base pointers
kvn@617 2619 Node *addp = n->in(AddPNode::Address);
duke@435 2620 assert( !addp->is_AddP() ||
duke@435 2621 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
duke@435 2622 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
duke@435 2623 "Base pointers must match" );
kvn@617 2624 #ifdef _LP64
roland@4159 2625 if ((UseCompressedOops || UseCompressedKlassPointers) &&
kvn@617 2626 addp->Opcode() == Op_ConP &&
kvn@617 2627 addp == n->in(AddPNode::Base) &&
kvn@617 2628 n->in(AddPNode::Offset)->is_Con()) {
kvn@617 2629 // Use addressing with narrow klass to load with offset on x86.
kvn@617 2630 // On sparc loading 32-bits constant and decoding it have less
kvn@617 2631 // instructions (4) then load 64-bits constant (7).
kvn@617 2632 // Do this transformation here since IGVN will convert ConN back to ConP.
kvn@617 2633 const Type* t = addp->bottom_type();
roland@4159 2634 if (t->isa_oopptr() || t->isa_klassptr()) {
kvn@617 2635 Node* nn = NULL;
kvn@617 2636
roland@4159 2637 int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass;
roland@4159 2638
kvn@617 2639 // Look for existing ConN node of the same exact type.
bharadwaj@4315 2640 Node* r = root();
kvn@617 2641 uint cnt = r->outcnt();
kvn@617 2642 for (uint i = 0; i < cnt; i++) {
kvn@617 2643 Node* m = r->raw_out(i);
roland@4159 2644 if (m!= NULL && m->Opcode() == op &&
kvn@656 2645 m->bottom_type()->make_ptr() == t) {
kvn@617 2646 nn = m;
kvn@617 2647 break;
kvn@617 2648 }
kvn@617 2649 }
kvn@617 2650 if (nn != NULL) {
kvn@617 2651 // Decode a narrow oop to match address
kvn@617 2652 // [R12 + narrow_oop_reg<<3 + offset]
roland@4159 2653 if (t->isa_oopptr()) {
bharadwaj@4315 2654 nn = new (this) DecodeNNode(nn, t);
roland@4159 2655 } else {
bharadwaj@4315 2656 nn = new (this) DecodeNKlassNode(nn, t);
roland@4159 2657 }
kvn@617 2658 n->set_req(AddPNode::Base, nn);
kvn@617 2659 n->set_req(AddPNode::Address, nn);
kvn@617 2660 if (addp->outcnt() == 0) {
bharadwaj@4315 2661 addp->disconnect_inputs(NULL, this);
kvn@617 2662 }
kvn@617 2663 }
kvn@617 2664 }
kvn@617 2665 }
kvn@617 2666 #endif
duke@435 2667 break;
duke@435 2668 }
duke@435 2669
kvn@599 2670 #ifdef _LP64
kvn@803 2671 case Op_CastPP:
kvn@1930 2672 if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
kvn@803 2673 Node* in1 = n->in(1);
kvn@803 2674 const Type* t = n->bottom_type();
kvn@803 2675 Node* new_in1 = in1->clone();
kvn@803 2676 new_in1->as_DecodeN()->set_type(t);
kvn@803 2677
kvn@1930 2678 if (!Matcher::narrow_oop_use_complex_address()) {
kvn@803 2679 //
kvn@803 2680 // x86, ARM and friends can handle 2 adds in addressing mode
kvn@803 2681 // and Matcher can fold a DecodeN node into address by using
kvn@803 2682 // a narrow oop directly and do implicit NULL check in address:
kvn@803 2683 //
kvn@803 2684 // [R12 + narrow_oop_reg<<3 + offset]
kvn@803 2685 // NullCheck narrow_oop_reg
kvn@803 2686 //
kvn@803 2687 // On other platforms (Sparc) we have to keep new DecodeN node and
kvn@803 2688 // use it to do implicit NULL check in address:
kvn@803 2689 //
kvn@803 2690 // decode_not_null narrow_oop_reg, base_reg
kvn@803 2691 // [base_reg + offset]
kvn@803 2692 // NullCheck base_reg
kvn@803 2693 //
twisti@1040 2694 // Pin the new DecodeN node to non-null path on these platform (Sparc)
kvn@803 2695 // to keep the information to which NULL check the new DecodeN node
kvn@803 2696 // corresponds to use it as value in implicit_null_check().
kvn@803 2697 //
kvn@803 2698 new_in1->set_req(0, n->in(0));
kvn@803 2699 }
kvn@803 2700
bharadwaj@4315 2701 n->subsume_by(new_in1, this);
kvn@803 2702 if (in1->outcnt() == 0) {
bharadwaj@4315 2703 in1->disconnect_inputs(NULL, this);
kvn@803 2704 }
kvn@803 2705 }
kvn@803 2706 break;
kvn@803 2707
kvn@599 2708 case Op_CmpP:
kvn@603 2709 // Do this transformation here to preserve CmpPNode::sub() and
kvn@603 2710 // other TypePtr related Ideal optimizations (for example, ptr nullness).
roland@4159 2711 if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) {
kvn@766 2712 Node* in1 = n->in(1);
kvn@766 2713 Node* in2 = n->in(2);
roland@4159 2714 if (!in1->is_DecodeNarrowPtr()) {
kvn@766 2715 in2 = in1;
kvn@766 2716 in1 = n->in(2);
kvn@766 2717 }
roland@4159 2718 assert(in1->is_DecodeNarrowPtr(), "sanity");
kvn@766 2719
kvn@766 2720 Node* new_in2 = NULL;
roland@4159 2721 if (in2->is_DecodeNarrowPtr()) {
roland@4159 2722 assert(in2->Opcode() == in1->Opcode(), "must be same node type");
kvn@766 2723 new_in2 = in2->in(1);
kvn@766 2724 } else if (in2->Opcode() == Op_ConP) {
kvn@766 2725 const Type* t = in2->bottom_type();
kvn@1930 2726 if (t == TypePtr::NULL_PTR) {
roland@4159 2727 assert(in1->is_DecodeN(), "compare klass to null?");
kvn@1930 2728 // Don't convert CmpP null check into CmpN if compressed
kvn@1930 2729 // oops implicit null check is not generated.
kvn@1930 2730 // This will allow to generate normal oop implicit null check.
kvn@1930 2731 if (Matcher::gen_narrow_oop_implicit_null_checks())
bharadwaj@4315 2732 new_in2 = ConNode::make(this, TypeNarrowOop::NULL_PTR);
kvn@803 2733 //
kvn@803 2734 // This transformation together with CastPP transformation above
kvn@803 2735 // will generated code for implicit NULL checks for compressed oops.
kvn@803 2736 //
kvn@803 2737 // The original code after Optimize()
kvn@803 2738 //
kvn@803 2739 // LoadN memory, narrow_oop_reg
kvn@803 2740 // decode narrow_oop_reg, base_reg
kvn@803 2741 // CmpP base_reg, NULL
kvn@803 2742 // CastPP base_reg // NotNull
kvn@803 2743 // Load [base_reg + offset], val_reg
kvn@803 2744 //
kvn@803 2745 // after these transformations will be
kvn@803 2746 //
kvn@803 2747 // LoadN memory, narrow_oop_reg
kvn@803 2748 // CmpN narrow_oop_reg, NULL
kvn@803 2749 // decode_not_null narrow_oop_reg, base_reg
kvn@803 2750 // Load [base_reg + offset], val_reg
kvn@803 2751 //
kvn@803 2752 // and the uncommon path (== NULL) will use narrow_oop_reg directly
kvn@803 2753 // since narrow oops can be used in debug info now (see the code in
kvn@803 2754 // final_graph_reshaping_walk()).
kvn@803 2755 //
kvn@803 2756 // At the end the code will be matched to
kvn@803 2757 // on x86:
kvn@803 2758 //
kvn@803 2759 // Load_narrow_oop memory, narrow_oop_reg
kvn@803 2760 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
kvn@803 2761 // NullCheck narrow_oop_reg
kvn@803 2762 //
kvn@803 2763 // and on sparc:
kvn@803 2764 //
kvn@803 2765 // Load_narrow_oop memory, narrow_oop_reg
kvn@803 2766 // decode_not_null narrow_oop_reg, base_reg
kvn@803 2767 // Load [base_reg + offset], val_reg
kvn@803 2768 // NullCheck base_reg
kvn@803 2769 //
kvn@599 2770 } else if (t->isa_oopptr()) {
bharadwaj@4315 2771 new_in2 = ConNode::make(this, t->make_narrowoop());
roland@4159 2772 } else if (t->isa_klassptr()) {
bharadwaj@4315 2773 new_in2 = ConNode::make(this, t->make_narrowklass());
kvn@599 2774 }
kvn@599 2775 }
kvn@766 2776 if (new_in2 != NULL) {
bharadwaj@4315 2777 Node* cmpN = new (this) CmpNNode(in1->in(1), new_in2);
bharadwaj@4315 2778 n->subsume_by(cmpN, this);
kvn@766 2779 if (in1->outcnt() == 0) {
bharadwaj@4315 2780 in1->disconnect_inputs(NULL, this);
kvn@766 2781 }
kvn@766 2782 if (in2->outcnt() == 0) {
bharadwaj@4315 2783 in2->disconnect_inputs(NULL, this);
kvn@766 2784 }
kvn@599 2785 }
kvn@599 2786 }
kvn@728 2787 break;
kvn@803 2788
kvn@803 2789 case Op_DecodeN:
roland@4159 2790 case Op_DecodeNKlass:
roland@4159 2791 assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out");
kvn@1930 2792 // DecodeN could be pinned when it can't be fold into
kvn@927 2793 // an address expression, see the code for Op_CastPP above.
roland@4159 2794 assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control");
kvn@803 2795 break;
kvn@803 2796
roland@4159 2797 case Op_EncodeP:
roland@4159 2798 case Op_EncodePKlass: {
kvn@803 2799 Node* in1 = n->in(1);
roland@4159 2800 if (in1->is_DecodeNarrowPtr()) {
bharadwaj@4315 2801 n->subsume_by(in1->in(1), this);
kvn@803 2802 } else if (in1->Opcode() == Op_ConP) {
kvn@803 2803 const Type* t = in1->bottom_type();
kvn@803 2804 if (t == TypePtr::NULL_PTR) {
roland@4159 2805 assert(t->isa_oopptr(), "null klass?");
bharadwaj@4315 2806 n->subsume_by(ConNode::make(this, TypeNarrowOop::NULL_PTR), this);
kvn@803 2807 } else if (t->isa_oopptr()) {
bharadwaj@4315 2808 n->subsume_by(ConNode::make(this, t->make_narrowoop()), this);
roland@4159 2809 } else if (t->isa_klassptr()) {
bharadwaj@4315 2810 n->subsume_by(ConNode::make(this, t->make_narrowklass()), this);
kvn@803 2811 }
kvn@803 2812 }
kvn@803 2813 if (in1->outcnt() == 0) {
bharadwaj@4315 2814 in1->disconnect_inputs(NULL, this);
kvn@803 2815 }
kvn@803 2816 break;
kvn@803 2817 }
kvn@803 2818
never@1515 2819 case Op_Proj: {
never@1515 2820 if (OptimizeStringConcat) {
never@1515 2821 ProjNode* p = n->as_Proj();
never@1515 2822 if (p->_is_io_use) {
never@1515 2823 // Separate projections were used for the exception path which
never@1515 2824 // are normally removed by a late inline. If it wasn't inlined
never@1515 2825 // then they will hang around and should just be replaced with
never@1515 2826 // the original one.
never@1515 2827 Node* proj = NULL;
never@1515 2828 // Replace with just one
never@1515 2829 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) {
never@1515 2830 Node *use = i.get();
never@1515 2831 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) {
never@1515 2832 proj = use;
never@1515 2833 break;
never@1515 2834 }
never@1515 2835 }
kvn@3396 2836 assert(proj != NULL, "must be found");
bharadwaj@4315 2837 p->subsume_by(proj, this);
never@1515 2838 }
never@1515 2839 }
never@1515 2840 break;
never@1515 2841 }
never@1515 2842
kvn@803 2843 case Op_Phi:
roland@4159 2844 if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) {
kvn@803 2845 // The EncodeP optimization may create Phi with the same edges
kvn@803 2846 // for all paths. It is not handled well by Register Allocator.
kvn@803 2847 Node* unique_in = n->in(1);
kvn@803 2848 assert(unique_in != NULL, "");
kvn@803 2849 uint cnt = n->req();
kvn@803 2850 for (uint i = 2; i < cnt; i++) {
kvn@803 2851 Node* m = n->in(i);
kvn@803 2852 assert(m != NULL, "");
kvn@803 2853 if (unique_in != m)
kvn@803 2854 unique_in = NULL;
kvn@803 2855 }
kvn@803 2856 if (unique_in != NULL) {
bharadwaj@4315 2857 n->subsume_by(unique_in, this);
kvn@803 2858 }
kvn@803 2859 }
kvn@803 2860 break;
kvn@803 2861
kvn@599 2862 #endif
kvn@599 2863
duke@435 2864 case Op_ModI:
duke@435 2865 if (UseDivMod) {
duke@435 2866 // Check if a%b and a/b both exist
duke@435 2867 Node* d = n->find_similar(Op_DivI);
duke@435 2868 if (d) {
duke@435 2869 // Replace them with a fused divmod if supported
duke@435 2870 if (Matcher::has_match_rule(Op_DivModI)) {
bharadwaj@4315 2871 DivModINode* divmod = DivModINode::make(this, n);
bharadwaj@4315 2872 d->subsume_by(divmod->div_proj(), this);
bharadwaj@4315 2873 n->subsume_by(divmod->mod_proj(), this);
duke@435 2874 } else {
duke@435 2875 // replace a%b with a-((a/b)*b)
bharadwaj@4315 2876 Node* mult = new (this) MulINode(d, d->in(2));
bharadwaj@4315 2877 Node* sub = new (this) SubINode(d->in(1), mult);
bharadwaj@4315 2878 n->subsume_by(sub, this);
duke@435 2879 }
duke@435 2880 }
duke@435 2881 }
duke@435 2882 break;
duke@435 2883
duke@435 2884 case Op_ModL:
duke@435 2885 if (UseDivMod) {
duke@435 2886 // Check if a%b and a/b both exist
duke@435 2887 Node* d = n->find_similar(Op_DivL);
duke@435 2888 if (d) {
duke@435 2889 // Replace them with a fused divmod if supported
duke@435 2890 if (Matcher::has_match_rule(Op_DivModL)) {
bharadwaj@4315 2891 DivModLNode* divmod = DivModLNode::make(this, n);
bharadwaj@4315 2892 d->subsume_by(divmod->div_proj(), this);
bharadwaj@4315 2893 n->subsume_by(divmod->mod_proj(), this);
duke@435 2894 } else {
duke@435 2895 // replace a%b with a-((a/b)*b)
bharadwaj@4315 2896 Node* mult = new (this) MulLNode(d, d->in(2));
bharadwaj@4315 2897 Node* sub = new (this) SubLNode(d->in(1), mult);
bharadwaj@4315 2898 n->subsume_by(sub, this);
duke@435 2899 }
duke@435 2900 }
duke@435 2901 }
duke@435 2902 break;
duke@435 2903
kvn@3882 2904 case Op_LoadVector:
kvn@3882 2905 case Op_StoreVector:
duke@435 2906 break;
duke@435 2907
duke@435 2908 case Op_PackB:
duke@435 2909 case Op_PackS:
duke@435 2910 case Op_PackI:
duke@435 2911 case Op_PackF:
duke@435 2912 case Op_PackL:
duke@435 2913 case Op_PackD:
duke@435 2914 if (n->req()-1 > 2) {
duke@435 2915 // Replace many operand PackNodes with a binary tree for matching
duke@435 2916 PackNode* p = (PackNode*) n;
bharadwaj@4315 2917 Node* btp = p->binary_tree_pack(this, 1, n->req());
bharadwaj@4315 2918 n->subsume_by(btp, this);
duke@435 2919 }
duke@435 2920 break;
kvn@1294 2921 case Op_Loop:
kvn@1294 2922 case Op_CountedLoop:
kvn@1294 2923 if (n->as_Loop()->is_inner_loop()) {
kvn@1294 2924 frc.inc_inner_loop_count();
kvn@1294 2925 }
kvn@1294 2926 break;
roland@2683 2927 case Op_LShiftI:
roland@2683 2928 case Op_RShiftI:
roland@2683 2929 case Op_URShiftI:
roland@2683 2930 case Op_LShiftL:
roland@2683 2931 case Op_RShiftL:
roland@2683 2932 case Op_URShiftL:
roland@2683 2933 if (Matcher::need_masked_shift_count) {
roland@2683 2934 // The cpu's shift instructions don't restrict the count to the
roland@2683 2935 // lower 5/6 bits. We need to do the masking ourselves.
roland@2683 2936 Node* in2 = n->in(2);
roland@2683 2937 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
roland@2683 2938 const TypeInt* t = in2->find_int_type();
roland@2683 2939 if (t != NULL && t->is_con()) {
roland@2683 2940 juint shift = t->get_con();
roland@2683 2941 if (shift > mask) { // Unsigned cmp
bharadwaj@4315 2942 n->set_req(2, ConNode::make(this, TypeInt::make(shift & mask)));
roland@2683 2943 }
roland@2683 2944 } else {
roland@2683 2945 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
bharadwaj@4315 2946 Node* shift = new (this) AndINode(in2, ConNode::make(this, TypeInt::make(mask)));
roland@2683 2947 n->set_req(2, shift);
roland@2683 2948 }
roland@2683 2949 }
roland@2683 2950 if (in2->outcnt() == 0) { // Remove dead node
bharadwaj@4315 2951 in2->disconnect_inputs(NULL, this);
roland@2683 2952 }
roland@2683 2953 }
roland@2683 2954 break;
roland@4694 2955 case Op_MemBarStoreStore:
kvn@5110 2956 case Op_MemBarRelease:
roland@4694 2957 // Break the link with AllocateNode: it is no longer useful and
roland@4694 2958 // confuses register allocation.
roland@4694 2959 if (n->req() > MemBarNode::Precedent) {
roland@4694 2960 n->set_req(MemBarNode::Precedent, top());
roland@4694 2961 }
roland@4694 2962 break;
duke@435 2963 default:
duke@435 2964 assert( !n->is_Call(), "" );
duke@435 2965 assert( !n->is_Mem(), "" );
duke@435 2966 break;
duke@435 2967 }
never@562 2968
never@562 2969 // Collect CFG split points
never@562 2970 if (n->is_MultiBranch())
kvn@1294 2971 frc._tests.push(n);
duke@435 2972 }
duke@435 2973
duke@435 2974 //------------------------------final_graph_reshaping_walk---------------------
duke@435 2975 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
duke@435 2976 // requires that the walk visits a node's inputs before visiting the node.
bharadwaj@4315 2977 void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
kvn@766 2978 ResourceArea *area = Thread::current()->resource_area();
kvn@766 2979 Unique_Node_List sfpt(area);
kvn@766 2980
kvn@1294 2981 frc._visited.set(root->_idx); // first, mark node as visited
duke@435 2982 uint cnt = root->req();
duke@435 2983 Node *n = root;
duke@435 2984 uint i = 0;
duke@435 2985 while (true) {
duke@435 2986 if (i < cnt) {
duke@435 2987 // Place all non-visited non-null inputs onto stack
duke@435 2988 Node* m = n->in(i);
duke@435 2989 ++i;
kvn@1294 2990 if (m != NULL && !frc._visited.test_set(m->_idx)) {
kvn@766 2991 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL)
kvn@766 2992 sfpt.push(m);
duke@435 2993 cnt = m->req();
duke@435 2994 nstack.push(n, i); // put on stack parent and next input's index
duke@435 2995 n = m;
duke@435 2996 i = 0;
duke@435 2997 }
duke@435 2998 } else {
duke@435 2999 // Now do post-visit work
kvn@1294 3000 final_graph_reshaping_impl( n, frc );
duke@435 3001 if (nstack.is_empty())
duke@435 3002 break; // finished
duke@435 3003 n = nstack.node(); // Get node from stack
duke@435 3004 cnt = n->req();
duke@435 3005 i = nstack.index();
duke@435 3006 nstack.pop(); // Shift to the next node on stack
duke@435 3007 }
duke@435 3008 }
kvn@766 3009
kvn@1930 3010 // Skip next transformation if compressed oops are not used.
roland@4159 3011 if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) ||
roland@4159 3012 (!UseCompressedOops && !UseCompressedKlassPointers))
kvn@1930 3013 return;
kvn@1930 3014
roland@4159 3015 // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges.
kvn@766 3016 // It could be done for an uncommon traps or any safepoints/calls
roland@4159 3017 // if the DecodeN/DecodeNKlass node is referenced only in a debug info.
kvn@766 3018 while (sfpt.size() > 0) {
kvn@766 3019 n = sfpt.pop();
kvn@766 3020 JVMState *jvms = n->as_SafePoint()->jvms();
kvn@766 3021 assert(jvms != NULL, "sanity");
kvn@766 3022 int start = jvms->debug_start();
kvn@766 3023 int end = n->req();
kvn@766 3024 bool is_uncommon = (n->is_CallStaticJava() &&
kvn@766 3025 n->as_CallStaticJava()->uncommon_trap_request() != 0);
kvn@766 3026 for (int j = start; j < end; j++) {
kvn@766 3027 Node* in = n->in(j);
roland@4159 3028 if (in->is_DecodeNarrowPtr()) {
kvn@766 3029 bool safe_to_skip = true;
kvn@766 3030 if (!is_uncommon ) {
kvn@766 3031 // Is it safe to skip?
kvn@766 3032 for (uint i = 0; i < in->outcnt(); i++) {
kvn@766 3033 Node* u = in->raw_out(i);
kvn@766 3034 if (!u->is_SafePoint() ||
kvn@766 3035 u->is_Call() && u->as_Call()->has_non_debug_use(n)) {
kvn@766 3036 safe_to_skip = false;
kvn@766 3037 }
kvn@766 3038 }
kvn@766 3039 }
kvn@766 3040 if (safe_to_skip) {
kvn@766 3041 n->set_req(j, in->in(1));
kvn@766 3042 }
kvn@766 3043 if (in->outcnt() == 0) {
bharadwaj@4315 3044 in->disconnect_inputs(NULL, this);
kvn@766 3045 }
kvn@766 3046 }
kvn@766 3047 }
kvn@766 3048 }
duke@435 3049 }
duke@435 3050
duke@435 3051 //------------------------------final_graph_reshaping--------------------------
duke@435 3052 // Final Graph Reshaping.
duke@435 3053 //
duke@435 3054 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
duke@435 3055 // and not commoned up and forced early. Must come after regular
duke@435 3056 // optimizations to avoid GVN undoing the cloning. Clone constant
duke@435 3057 // inputs to Loop Phis; these will be split by the allocator anyways.
duke@435 3058 // Remove Opaque nodes.
duke@435 3059 // (2) Move last-uses by commutative operations to the left input to encourage
duke@435 3060 // Intel update-in-place two-address operations and better register usage
duke@435 3061 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
duke@435 3062 // calls canonicalizing them back.
duke@435 3063 // (3) Count the number of double-precision FP ops, single-precision FP ops
duke@435 3064 // and call sites. On Intel, we can get correct rounding either by
duke@435 3065 // forcing singles to memory (requires extra stores and loads after each
duke@435 3066 // FP bytecode) or we can set a rounding mode bit (requires setting and
duke@435 3067 // clearing the mode bit around call sites). The mode bit is only used
duke@435 3068 // if the relative frequency of single FP ops to calls is low enough.
duke@435 3069 // This is a key transform for SPEC mpeg_audio.
duke@435 3070 // (4) Detect infinite loops; blobs of code reachable from above but not
duke@435 3071 // below. Several of the Code_Gen algorithms fail on such code shapes,
duke@435 3072 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
duke@435 3073 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
duke@435 3074 // Detection is by looking for IfNodes where only 1 projection is
duke@435 3075 // reachable from below or CatchNodes missing some targets.
duke@435 3076 // (5) Assert for insane oop offsets in debug mode.
duke@435 3077
duke@435 3078 bool Compile::final_graph_reshaping() {
duke@435 3079 // an infinite loop may have been eliminated by the optimizer,
duke@435 3080 // in which case the graph will be empty.
duke@435 3081 if (root()->req() == 1) {
duke@435 3082 record_method_not_compilable("trivial infinite loop");
duke@435 3083 return true;
duke@435 3084 }
duke@435 3085
roland@4589 3086 // Expensive nodes have their control input set to prevent the GVN
roland@4589 3087 // from freely commoning them. There's no GVN beyond this point so
roland@4589 3088 // no need to keep the control input. We want the expensive nodes to
roland@4589 3089 // be freely moved to the least frequent code path by gcm.
roland@4589 3090 assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?");
roland@4589 3091 for (int i = 0; i < expensive_count(); i++) {
roland@4589 3092 _expensive_nodes->at(i)->set_req(0, NULL);
roland@4589 3093 }
roland@4589 3094
kvn@1294 3095 Final_Reshape_Counts frc;
duke@435 3096
duke@435 3097 // Visit everybody reachable!
duke@435 3098 // Allocate stack of size C->unique()/2 to avoid frequent realloc
duke@435 3099 Node_Stack nstack(unique() >> 1);
kvn@1294 3100 final_graph_reshaping_walk(nstack, root(), frc);
duke@435 3101
duke@435 3102 // Check for unreachable (from below) code (i.e., infinite loops).
kvn@1294 3103 for( uint i = 0; i < frc._tests.size(); i++ ) {
kvn@1294 3104 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
never@562 3105 // Get number of CFG targets.
duke@435 3106 // Note that PCTables include exception targets after calls.
never@562 3107 uint required_outcnt = n->required_outcnt();
never@562 3108 if (n->outcnt() != required_outcnt) {
duke@435 3109 // Check for a few special cases. Rethrow Nodes never take the
duke@435 3110 // 'fall-thru' path, so expected kids is 1 less.
duke@435 3111 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
duke@435 3112 if (n->in(0)->in(0)->is_Call()) {
duke@435 3113 CallNode *call = n->in(0)->in(0)->as_Call();
duke@435 3114 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
never@562 3115 required_outcnt--; // Rethrow always has 1 less kid
duke@435 3116 } else if (call->req() > TypeFunc::Parms &&
duke@435 3117 call->is_CallDynamicJava()) {
duke@435 3118 // Check for null receiver. In such case, the optimizer has
duke@435 3119 // detected that the virtual call will always result in a null
duke@435 3120 // pointer exception. The fall-through projection of this CatchNode
duke@435 3121 // will not be populated.
duke@435 3122 Node *arg0 = call->in(TypeFunc::Parms);
duke@435 3123 if (arg0->is_Type() &&
duke@435 3124 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
never@562 3125 required_outcnt--;
duke@435 3126 }
duke@435 3127 } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
duke@435 3128 call->req() > TypeFunc::Parms+1 &&
duke@435 3129 call->is_CallStaticJava()) {
duke@435 3130 // Check for negative array length. In such case, the optimizer has
duke@435 3131 // detected that the allocation attempt will always result in an
duke@435 3132 // exception. There is no fall-through projection of this CatchNode .
duke@435 3133 Node *arg1 = call->in(TypeFunc::Parms+1);
duke@435 3134 if (arg1->is_Type() &&
duke@435 3135 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
never@562 3136 required_outcnt--;
duke@435 3137 }
duke@435 3138 }
duke@435 3139 }
duke@435 3140 }
never@562 3141 // Recheck with a better notion of 'required_outcnt'
never@562 3142 if (n->outcnt() != required_outcnt) {
duke@435 3143 record_method_not_compilable("malformed control flow");
duke@435 3144 return true; // Not all targets reachable!
duke@435 3145 }
duke@435 3146 }
duke@435 3147 // Check that I actually visited all kids. Unreached kids
duke@435 3148 // must be infinite loops.
duke@435 3149 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
kvn@1294 3150 if (!frc._visited.test(n->fast_out(j)->_idx)) {
duke@435 3151 record_method_not_compilable("infinite loop");
duke@435 3152 return true; // Found unvisited kid; must be unreach
duke@435 3153 }
duke@435 3154 }
duke@435 3155
duke@435 3156 // If original bytecodes contained a mixture of floats and doubles
duke@435 3157 // check if the optimizer has made it homogenous, item (3).
never@1364 3158 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 &&
kvn@1294 3159 frc.get_float_count() > 32 &&
kvn@1294 3160 frc.get_double_count() == 0 &&
kvn@1294 3161 (10 * frc.get_call_count() < frc.get_float_count()) ) {
duke@435 3162 set_24_bit_selection_and_mode( false, true );
duke@435 3163 }
duke@435 3164
kvn@1294 3165 set_java_calls(frc.get_java_call_count());
kvn@1294 3166 set_inner_loops(frc.get_inner_loop_count());
duke@435 3167
duke@435 3168 // No infinite loops, no reason to bail out.
duke@435 3169 return false;
duke@435 3170 }
duke@435 3171
duke@435 3172 //-----------------------------too_many_traps----------------------------------
duke@435 3173 // Report if there are too many traps at the current method and bci.
duke@435 3174 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
duke@435 3175 bool Compile::too_many_traps(ciMethod* method,
duke@435 3176 int bci,
duke@435 3177 Deoptimization::DeoptReason reason) {
duke@435 3178 ciMethodData* md = method->method_data();
duke@435 3179 if (md->is_empty()) {
duke@435 3180 // Assume the trap has not occurred, or that it occurred only
duke@435 3181 // because of a transient condition during start-up in the interpreter.
duke@435 3182 return false;
duke@435 3183 }
duke@435 3184 if (md->has_trap_at(bci, reason) != 0) {
duke@435 3185 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
duke@435 3186 // Also, if there are multiple reasons, or if there is no per-BCI record,
duke@435 3187 // assume the worst.
duke@435 3188 if (log())
duke@435 3189 log()->elem("observe trap='%s' count='%d'",
duke@435 3190 Deoptimization::trap_reason_name(reason),
duke@435 3191 md->trap_count(reason));
duke@435 3192 return true;
duke@435 3193 } else {
duke@435 3194 // Ignore method/bci and see if there have been too many globally.
duke@435 3195 return too_many_traps(reason, md);
duke@435 3196 }
duke@435 3197 }
duke@435 3198
duke@435 3199 // Less-accurate variant which does not require a method and bci.
duke@435 3200 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
duke@435 3201 ciMethodData* logmd) {
duke@435 3202 if (trap_count(reason) >= (uint)PerMethodTrapLimit) {
duke@435 3203 // Too many traps globally.
duke@435 3204 // Note that we use cumulative trap_count, not just md->trap_count.
duke@435 3205 if (log()) {
duke@435 3206 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
duke@435 3207 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
duke@435 3208 Deoptimization::trap_reason_name(reason),
duke@435 3209 mcount, trap_count(reason));
duke@435 3210 }
duke@435 3211 return true;
duke@435 3212 } else {
duke@435 3213 // The coast is clear.
duke@435 3214 return false;
duke@435 3215 }
duke@435 3216 }
duke@435 3217
duke@435 3218 //--------------------------too_many_recompiles--------------------------------
duke@435 3219 // Report if there are too many recompiles at the current method and bci.
duke@435 3220 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
duke@435 3221 // Is not eager to return true, since this will cause the compiler to use
duke@435 3222 // Action_none for a trap point, to avoid too many recompilations.
duke@435 3223 bool Compile::too_many_recompiles(ciMethod* method,
duke@435 3224 int bci,
duke@435 3225 Deoptimization::DeoptReason reason) {
duke@435 3226 ciMethodData* md = method->method_data();
duke@435 3227 if (md->is_empty()) {
duke@435 3228 // Assume the trap has not occurred, or that it occurred only
duke@435 3229 // because of a transient condition during start-up in the interpreter.
duke@435 3230 return false;
duke@435 3231 }
duke@435 3232 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
duke@435 3233 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
duke@435 3234 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
duke@435 3235 Deoptimization::DeoptReason per_bc_reason
duke@435 3236 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
duke@435 3237 if ((per_bc_reason == Deoptimization::Reason_none
duke@435 3238 || md->has_trap_at(bci, reason) != 0)
duke@435 3239 // The trap frequency measure we care about is the recompile count:
duke@435 3240 && md->trap_recompiled_at(bci)
duke@435 3241 && md->overflow_recompile_count() >= bc_cutoff) {
duke@435 3242 // Do not emit a trap here if it has already caused recompilations.
duke@435 3243 // Also, if there are multiple reasons, or if there is no per-BCI record,
duke@435 3244 // assume the worst.
duke@435 3245 if (log())
duke@435 3246 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
duke@435 3247 Deoptimization::trap_reason_name(reason),
duke@435 3248 md->trap_count(reason),
duke@435 3249 md->overflow_recompile_count());
duke@435 3250 return true;
duke@435 3251 } else if (trap_count(reason) != 0
duke@435 3252 && decompile_count() >= m_cutoff) {
duke@435 3253 // Too many recompiles globally, and we have seen this sort of trap.
duke@435 3254 // Use cumulative decompile_count, not just md->decompile_count.
duke@435 3255 if (log())
duke@435 3256 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
duke@435 3257 Deoptimization::trap_reason_name(reason),
duke@435 3258 md->trap_count(reason), trap_count(reason),
duke@435 3259 md->decompile_count(), decompile_count());
duke@435 3260 return true;
duke@435 3261 } else {
duke@435 3262 // The coast is clear.
duke@435 3263 return false;
duke@435 3264 }
duke@435 3265 }
duke@435 3266
duke@435 3267
duke@435 3268 #ifndef PRODUCT
duke@435 3269 //------------------------------verify_graph_edges---------------------------
duke@435 3270 // Walk the Graph and verify that there is a one-to-one correspondence
duke@435 3271 // between Use-Def edges and Def-Use edges in the graph.
duke@435 3272 void Compile::verify_graph_edges(bool no_dead_code) {
duke@435 3273 if (VerifyGraphEdges) {
duke@435 3274 ResourceArea *area = Thread::current()->resource_area();
duke@435 3275 Unique_Node_List visited(area);
duke@435 3276 // Call recursive graph walk to check edges
duke@435 3277 _root->verify_edges(visited);
duke@435 3278 if (no_dead_code) {
duke@435 3279 // Now make sure that no visited node is used by an unvisited node.
duke@435 3280 bool dead_nodes = 0;
duke@435 3281 Unique_Node_List checked(area);
duke@435 3282 while (visited.size() > 0) {
duke@435 3283 Node* n = visited.pop();
duke@435 3284 checked.push(n);
duke@435 3285 for (uint i = 0; i < n->outcnt(); i++) {
duke@435 3286 Node* use = n->raw_out(i);
duke@435 3287 if (checked.member(use)) continue; // already checked
duke@435 3288 if (visited.member(use)) continue; // already in the graph
duke@435 3289 if (use->is_Con()) continue; // a dead ConNode is OK
duke@435 3290 // At this point, we have found a dead node which is DU-reachable.
duke@435 3291 if (dead_nodes++ == 0)
duke@435 3292 tty->print_cr("*** Dead nodes reachable via DU edges:");
duke@435 3293 use->dump(2);
duke@435 3294 tty->print_cr("---");
duke@435 3295 checked.push(use); // No repeats; pretend it is now checked.
duke@435 3296 }
duke@435 3297 }
duke@435 3298 assert(dead_nodes == 0, "using nodes must be reachable from root");
duke@435 3299 }
duke@435 3300 }
duke@435 3301 }
duke@435 3302 #endif
duke@435 3303
duke@435 3304 // The Compile object keeps track of failure reasons separately from the ciEnv.
duke@435 3305 // This is required because there is not quite a 1-1 relation between the
duke@435 3306 // ciEnv and its compilation task and the Compile object. Note that one
duke@435 3307 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
duke@435 3308 // to backtrack and retry without subsuming loads. Other than this backtracking
duke@435 3309 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
duke@435 3310 // by the logic in C2Compiler.
duke@435 3311 void Compile::record_failure(const char* reason) {
duke@435 3312 if (log() != NULL) {
duke@435 3313 log()->elem("failure reason='%s' phase='compile'", reason);
duke@435 3314 }
duke@435 3315 if (_failure_reason == NULL) {
duke@435 3316 // Record the first failure reason.
duke@435 3317 _failure_reason = reason;
duke@435 3318 }
never@657 3319 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
never@657 3320 C->print_method(_failure_reason);
never@657 3321 }
duke@435 3322 _root = NULL; // flush the graph, too
duke@435 3323 }
duke@435 3324
duke@435 3325 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog)
bharadwaj@4315 3326 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false),
bharadwaj@4315 3327 _phase_name(name), _dolog(dolog)
duke@435 3328 {
duke@435 3329 if (dolog) {
duke@435 3330 C = Compile::current();
duke@435 3331 _log = C->log();
duke@435 3332 } else {
duke@435 3333 C = NULL;
duke@435 3334 _log = NULL;
duke@435 3335 }
duke@435 3336 if (_log != NULL) {
bharadwaj@4315 3337 _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
duke@435 3338 _log->stamp();
duke@435 3339 _log->end_head();
duke@435 3340 }
duke@435 3341 }
duke@435 3342
duke@435 3343 Compile::TracePhase::~TracePhase() {
bharadwaj@4315 3344
bharadwaj@4315 3345 C = Compile::current();
bharadwaj@4315 3346 if (_dolog) {
bharadwaj@4315 3347 _log = C->log();
bharadwaj@4315 3348 } else {
bharadwaj@4315 3349 _log = NULL;
bharadwaj@4315 3350 }
bharadwaj@4315 3351
bharadwaj@4315 3352 #ifdef ASSERT
bharadwaj@4315 3353 if (PrintIdealNodeCount) {
bharadwaj@4315 3354 tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'",
bharadwaj@4315 3355 _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk());
bharadwaj@4315 3356 }
bharadwaj@4315 3357
bharadwaj@4315 3358 if (VerifyIdealNodeCount) {
bharadwaj@4315 3359 Compile::current()->print_missing_nodes();
bharadwaj@4315 3360 }
bharadwaj@4315 3361 #endif
bharadwaj@4315 3362
duke@435 3363 if (_log != NULL) {
bharadwaj@4315 3364 _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes());
duke@435 3365 }
duke@435 3366 }
twisti@2350 3367
twisti@2350 3368 //=============================================================================
twisti@2350 3369 // Two Constant's are equal when the type and the value are equal.
twisti@2350 3370 bool Compile::Constant::operator==(const Constant& other) {
twisti@2350 3371 if (type() != other.type() ) return false;
twisti@2350 3372 if (can_be_reused() != other.can_be_reused()) return false;
twisti@2350 3373 // For floating point values we compare the bit pattern.
twisti@2350 3374 switch (type()) {
coleenp@4037 3375 case T_FLOAT: return (_v._value.i == other._v._value.i);
twisti@2350 3376 case T_LONG:
coleenp@4037 3377 case T_DOUBLE: return (_v._value.j == other._v._value.j);
twisti@2350 3378 case T_OBJECT:
coleenp@4037 3379 case T_ADDRESS: return (_v._value.l == other._v._value.l);
coleenp@4037 3380 case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries
kvn@4199 3381 case T_METADATA: return (_v._metadata == other._v._metadata);
twisti@2350 3382 default: ShouldNotReachHere();
twisti@2350 3383 }
twisti@2350 3384 return false;
twisti@2350 3385 }
twisti@2350 3386
twisti@2350 3387 static int type_to_size_in_bytes(BasicType t) {
twisti@2350 3388 switch (t) {
twisti@2350 3389 case T_LONG: return sizeof(jlong );
twisti@2350 3390 case T_FLOAT: return sizeof(jfloat );
twisti@2350 3391 case T_DOUBLE: return sizeof(jdouble);
coleenp@4037 3392 case T_METADATA: return sizeof(Metadata*);
twisti@2350 3393 // We use T_VOID as marker for jump-table entries (labels) which
twisti@3310 3394 // need an internal word relocation.
twisti@2350 3395 case T_VOID:
twisti@2350 3396 case T_ADDRESS:
twisti@2350 3397 case T_OBJECT: return sizeof(jobject);
twisti@2350 3398 }
twisti@2350 3399
twisti@2350 3400 ShouldNotReachHere();
twisti@2350 3401 return -1;
twisti@2350 3402 }
twisti@2350 3403
twisti@3310 3404 int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) {
twisti@3310 3405 // sort descending
twisti@3310 3406 if (a->freq() > b->freq()) return -1;
twisti@3310 3407 if (a->freq() < b->freq()) return 1;
twisti@3310 3408 return 0;
twisti@3310 3409 }
twisti@3310 3410
twisti@2350 3411 void Compile::ConstantTable::calculate_offsets_and_size() {
twisti@3310 3412 // First, sort the array by frequencies.
twisti@3310 3413 _constants.sort(qsort_comparator);
twisti@3310 3414
twisti@3310 3415 #ifdef ASSERT
twisti@3310 3416 // Make sure all jump-table entries were sorted to the end of the
twisti@3310 3417 // array (they have a negative frequency).
twisti@3310 3418 bool found_void = false;
twisti@3310 3419 for (int i = 0; i < _constants.length(); i++) {
twisti@3310 3420 Constant con = _constants.at(i);
twisti@3310 3421 if (con.type() == T_VOID)
twisti@3310 3422 found_void = true; // jump-tables
twisti@3310 3423 else
twisti@3310 3424 assert(!found_void, "wrong sorting");
twisti@3310 3425 }
twisti@3310 3426 #endif
twisti@3310 3427
twisti@3310 3428 int offset = 0;
twisti@3310 3429 for (int i = 0; i < _constants.length(); i++) {
twisti@3310 3430 Constant* con = _constants.adr_at(i);
twisti@3310 3431
twisti@3310 3432 // Align offset for type.
twisti@3310 3433 int typesize = type_to_size_in_bytes(con->type());
twisti@3310 3434 offset = align_size_up(offset, typesize);
twisti@3310 3435 con->set_offset(offset); // set constant's offset
twisti@3310 3436
twisti@3310 3437 if (con->type() == T_VOID) {
twisti@3310 3438 MachConstantNode* n = (MachConstantNode*) con->get_jobject();
twisti@3310 3439 offset = offset + typesize * n->outcnt(); // expand jump-table
twisti@3310 3440 } else {
twisti@3310 3441 offset = offset + typesize;
twisti@2350 3442 }
twisti@2350 3443 }
twisti@2350 3444
twisti@2350 3445 // Align size up to the next section start (which is insts; see
twisti@2350 3446 // CodeBuffer::align_at_start).
twisti@2350 3447 assert(_size == -1, "already set?");
twisti@3310 3448 _size = align_size_up(offset, CodeEntryAlignment);
twisti@2350 3449 }
twisti@2350 3450
twisti@2350 3451 void Compile::ConstantTable::emit(CodeBuffer& cb) {
twisti@2350 3452 MacroAssembler _masm(&cb);
twisti@3310 3453 for (int i = 0; i < _constants.length(); i++) {
twisti@3310 3454 Constant con = _constants.at(i);
twisti@3310 3455 address constant_addr;
twisti@3310 3456 switch (con.type()) {
twisti@3310 3457 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break;
twisti@3310 3458 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break;
twisti@3310 3459 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break;
twisti@3310 3460 case T_OBJECT: {
twisti@3310 3461 jobject obj = con.get_jobject();
twisti@3310 3462 int oop_index = _masm.oop_recorder()->find_index(obj);
twisti@3310 3463 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index));
twisti@3310 3464 break;
twisti@3310 3465 }
twisti@3310 3466 case T_ADDRESS: {
twisti@3310 3467 address addr = (address) con.get_jobject();
twisti@3310 3468 constant_addr = _masm.address_constant(addr);
twisti@3310 3469 break;
twisti@3310 3470 }
twisti@3310 3471 // We use T_VOID as marker for jump-table entries (labels) which
twisti@3310 3472 // need an internal word relocation.
twisti@3310 3473 case T_VOID: {
twisti@3310 3474 MachConstantNode* n = (MachConstantNode*) con.get_jobject();
twisti@3310 3475 // Fill the jump-table with a dummy word. The real value is
twisti@3310 3476 // filled in later in fill_jump_table.
twisti@3310 3477 address dummy = (address) n;
twisti@3310 3478 constant_addr = _masm.address_constant(dummy);
twisti@3310 3479 // Expand jump-table
twisti@3310 3480 for (uint i = 1; i < n->outcnt(); i++) {
twisti@3310 3481 address temp_addr = _masm.address_constant(dummy + i);
twisti@3310 3482 assert(temp_addr, "consts section too small");
twisti@2350 3483 }
twisti@3310 3484 break;
twisti@2350 3485 }
coleenp@4037 3486 case T_METADATA: {
coleenp@4037 3487 Metadata* obj = con.get_metadata();
coleenp@4037 3488 int metadata_index = _masm.oop_recorder()->find_index(obj);
coleenp@4037 3489 constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index));
coleenp@4037 3490 break;
coleenp@4037 3491 }
twisti@3310 3492 default: ShouldNotReachHere();
twisti@3310 3493 }
twisti@3310 3494 assert(constant_addr, "consts section too small");
kvn@3971 3495 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), err_msg_res("must be: %d == %d", constant_addr - _masm.code()->consts()->start(), con.offset()));
twisti@2350 3496 }
twisti@2350 3497 }
twisti@2350 3498
twisti@2350 3499 int Compile::ConstantTable::find_offset(Constant& con) const {
twisti@2350 3500 int idx = _constants.find(con);
twisti@2350 3501 assert(idx != -1, "constant must be in constant table");
twisti@2350 3502 int offset = _constants.at(idx).offset();
twisti@2350 3503 assert(offset != -1, "constant table not emitted yet?");
twisti@2350 3504 return offset;
twisti@2350 3505 }
twisti@2350 3506
twisti@2350 3507 void Compile::ConstantTable::add(Constant& con) {
twisti@2350 3508 if (con.can_be_reused()) {
twisti@2350 3509 int idx = _constants.find(con);
twisti@2350 3510 if (idx != -1 && _constants.at(idx).can_be_reused()) {
twisti@3310 3511 _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value
twisti@2350 3512 return;
twisti@2350 3513 }
twisti@2350 3514 }
twisti@2350 3515 (void) _constants.append(con);
twisti@2350 3516 }
twisti@2350 3517
twisti@3310 3518 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) {
twisti@3310 3519 Block* b = Compile::current()->cfg()->_bbs[n->_idx];
twisti@3310 3520 Constant con(type, value, b->_freq);
twisti@2350 3521 add(con);
twisti@2350 3522 return con;
twisti@2350 3523 }
twisti@2350 3524
coleenp@4037 3525 Compile::Constant Compile::ConstantTable::add(Metadata* metadata) {
coleenp@4037 3526 Constant con(metadata);
coleenp@4037 3527 add(con);
coleenp@4037 3528 return con;
coleenp@4037 3529 }
coleenp@4037 3530
twisti@3310 3531 Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) {
twisti@2350 3532 jvalue value;
twisti@2350 3533 BasicType type = oper->type()->basic_type();
twisti@2350 3534 switch (type) {
twisti@2350 3535 case T_LONG: value.j = oper->constantL(); break;
twisti@2350 3536 case T_FLOAT: value.f = oper->constantF(); break;
twisti@2350 3537 case T_DOUBLE: value.d = oper->constantD(); break;
twisti@2350 3538 case T_OBJECT:
twisti@2350 3539 case T_ADDRESS: value.l = (jobject) oper->constant(); break;
coleenp@4037 3540 case T_METADATA: return add((Metadata*)oper->constant()); break;
coleenp@4037 3541 default: guarantee(false, err_msg_res("unhandled type: %s", type2name(type)));
twisti@2350 3542 }
twisti@3310 3543 return add(n, type, value);
twisti@2350 3544 }
twisti@2350 3545
twisti@3310 3546 Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) {
twisti@2350 3547 jvalue value;
twisti@2350 3548 // We can use the node pointer here to identify the right jump-table
twisti@2350 3549 // as this method is called from Compile::Fill_buffer right before
twisti@2350 3550 // the MachNodes are emitted and the jump-table is filled (means the
twisti@2350 3551 // MachNode pointers do not change anymore).
twisti@2350 3552 value.l = (jobject) n;
twisti@3310 3553 Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused.
twisti@3310 3554 add(con);
twisti@2350 3555 return con;
twisti@2350 3556 }
twisti@2350 3557
twisti@2350 3558 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const {
twisti@2350 3559 // If called from Compile::scratch_emit_size do nothing.
twisti@2350 3560 if (Compile::current()->in_scratch_emit_size()) return;
twisti@2350 3561
twisti@2350 3562 assert(labels.is_nonempty(), "must be");
kvn@3971 3563 assert((uint) labels.length() == n->outcnt(), err_msg_res("must be equal: %d == %d", labels.length(), n->outcnt()));
twisti@2350 3564
twisti@2350 3565 // Since MachConstantNode::constant_offset() also contains
twisti@2350 3566 // table_base_offset() we need to subtract the table_base_offset()
twisti@2350 3567 // to get the plain offset into the constant table.
twisti@2350 3568 int offset = n->constant_offset() - table_base_offset();
twisti@2350 3569
twisti@2350 3570 MacroAssembler _masm(&cb);
twisti@2350 3571 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset);
twisti@2350 3572
twisti@3310 3573 for (uint i = 0; i < n->outcnt(); i++) {
twisti@2350 3574 address* constant_addr = &jump_table_base[i];
kvn@3971 3575 assert(*constant_addr == (((address) n) + i), err_msg_res("all jump-table entries must contain adjusted node pointer: " INTPTR_FORMAT " == " INTPTR_FORMAT, *constant_addr, (((address) n) + i)));
twisti@2350 3576 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr);
twisti@2350 3577 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type);
twisti@2350 3578 }
twisti@2350 3579 }
roland@4357 3580
roland@4357 3581 void Compile::dump_inlining() {
kvn@4448 3582 if (PrintInlining || PrintIntrinsics NOT_PRODUCT( || PrintOptoInlining)) {
roland@4409 3583 // Print inlining message for candidates that we couldn't inline
roland@4409 3584 // for lack of space or non constant receiver
roland@4409 3585 for (int i = 0; i < _late_inlines.length(); i++) {
roland@4409 3586 CallGenerator* cg = _late_inlines.at(i);
roland@4409 3587 cg->print_inlining_late("live nodes > LiveNodeCountInliningCutoff");
roland@4409 3588 }
roland@4409 3589 Unique_Node_List useful;
roland@4409 3590 useful.push(root());
roland@4409 3591 for (uint next = 0; next < useful.size(); ++next) {
roland@4409 3592 Node* n = useful.at(next);
roland@4409 3593 if (n->is_Call() && n->as_Call()->generator() != NULL && n->as_Call()->generator()->call_node() == n) {
roland@4409 3594 CallNode* call = n->as_Call();
roland@4409 3595 CallGenerator* cg = call->generator();
roland@4409 3596 cg->print_inlining_late("receiver not constant");
roland@4409 3597 }
roland@4409 3598 uint max = n->len();
roland@4409 3599 for ( uint i = 0; i < max; ++i ) {
roland@4409 3600 Node *m = n->in(i);
roland@4409 3601 if ( m == NULL ) continue;
roland@4409 3602 useful.push(m);
roland@4409 3603 }
roland@4409 3604 }
roland@4357 3605 for (int i = 0; i < _print_inlining_list->length(); i++) {
roland@4357 3606 tty->print(_print_inlining_list->at(i).ss()->as_string());
roland@4357 3607 }
roland@4357 3608 }
roland@4357 3609 }
roland@4589 3610
roland@4589 3611 int Compile::cmp_expensive_nodes(Node* n1, Node* n2) {
roland@4589 3612 if (n1->Opcode() < n2->Opcode()) return -1;
roland@4589 3613 else if (n1->Opcode() > n2->Opcode()) return 1;
roland@4589 3614
roland@4589 3615 assert(n1->req() == n2->req(), err_msg_res("can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req()));
roland@4589 3616 for (uint i = 1; i < n1->req(); i++) {
roland@4589 3617 if (n1->in(i) < n2->in(i)) return -1;
roland@4589 3618 else if (n1->in(i) > n2->in(i)) return 1;
roland@4589 3619 }
roland@4589 3620
roland@4589 3621 return 0;
roland@4589 3622 }
roland@4589 3623
roland@4589 3624 int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) {
roland@4589 3625 Node* n1 = *n1p;
roland@4589 3626 Node* n2 = *n2p;
roland@4589 3627
roland@4589 3628 return cmp_expensive_nodes(n1, n2);
roland@4589 3629 }
roland@4589 3630
roland@4589 3631 void Compile::sort_expensive_nodes() {
roland@4589 3632 if (!expensive_nodes_sorted()) {
roland@4589 3633 _expensive_nodes->sort(cmp_expensive_nodes);
roland@4589 3634 }
roland@4589 3635 }
roland@4589 3636
roland@4589 3637 bool Compile::expensive_nodes_sorted() const {
roland@4589 3638 for (int i = 1; i < _expensive_nodes->length(); i++) {
roland@4589 3639 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) {
roland@4589 3640 return false;
roland@4589 3641 }
roland@4589 3642 }
roland@4589 3643 return true;
roland@4589 3644 }
roland@4589 3645
roland@4589 3646 bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) {
roland@4589 3647 if (_expensive_nodes->length() == 0) {
roland@4589 3648 return false;
roland@4589 3649 }
roland@4589 3650
roland@4589 3651 assert(OptimizeExpensiveOps, "optimization off?");
roland@4589 3652
roland@4589 3653 // Take this opportunity to remove dead nodes from the list
roland@4589 3654 int j = 0;
roland@4589 3655 for (int i = 0; i < _expensive_nodes->length(); i++) {
roland@4589 3656 Node* n = _expensive_nodes->at(i);
roland@4589 3657 if (!n->is_unreachable(igvn)) {
roland@4589 3658 assert(n->is_expensive(), "should be expensive");
roland@4589 3659 _expensive_nodes->at_put(j, n);
roland@4589 3660 j++;
roland@4589 3661 }
roland@4589 3662 }
roland@4589 3663 _expensive_nodes->trunc_to(j);
roland@4589 3664
roland@4589 3665 // Then sort the list so that similar nodes are next to each other
roland@4589 3666 // and check for at least two nodes of identical kind with same data
roland@4589 3667 // inputs.
roland@4589 3668 sort_expensive_nodes();
roland@4589 3669
roland@4589 3670 for (int i = 0; i < _expensive_nodes->length()-1; i++) {
roland@4589 3671 if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) {
roland@4589 3672 return true;
roland@4589 3673 }
roland@4589 3674 }
roland@4589 3675
roland@4589 3676 return false;
roland@4589 3677 }
roland@4589 3678
roland@4589 3679 void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) {
roland@4589 3680 if (_expensive_nodes->length() == 0) {
roland@4589 3681 return;
roland@4589 3682 }
roland@4589 3683
roland@4589 3684 assert(OptimizeExpensiveOps, "optimization off?");
roland@4589 3685
roland@4589 3686 // Sort to bring similar nodes next to each other and clear the
roland@4589 3687 // control input of nodes for which there's only a single copy.
roland@4589 3688 sort_expensive_nodes();
roland@4589 3689
roland@4589 3690 int j = 0;
roland@4589 3691 int identical = 0;
roland@4589 3692 int i = 0;
roland@4589 3693 for (; i < _expensive_nodes->length()-1; i++) {
roland@4589 3694 assert(j <= i, "can't write beyond current index");
roland@4589 3695 if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) {
roland@4589 3696 identical++;
roland@4589 3697 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
roland@4589 3698 continue;
roland@4589 3699 }
roland@4589 3700 if (identical > 0) {
roland@4589 3701 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
roland@4589 3702 identical = 0;
roland@4589 3703 } else {
roland@4589 3704 Node* n = _expensive_nodes->at(i);
roland@4589 3705 igvn.hash_delete(n);
roland@4589 3706 n->set_req(0, NULL);
roland@4589 3707 igvn.hash_insert(n);
roland@4589 3708 }
roland@4589 3709 }
roland@4589 3710 if (identical > 0) {
roland@4589 3711 _expensive_nodes->at_put(j++, _expensive_nodes->at(i));
roland@4589 3712 } else if (_expensive_nodes->length() >= 1) {
roland@4589 3713 Node* n = _expensive_nodes->at(i);
roland@4589 3714 igvn.hash_delete(n);
roland@4589 3715 n->set_req(0, NULL);
roland@4589 3716 igvn.hash_insert(n);
roland@4589 3717 }
roland@4589 3718 _expensive_nodes->trunc_to(j);
roland@4589 3719 }
roland@4589 3720
roland@4589 3721 void Compile::add_expensive_node(Node * n) {
roland@4589 3722 assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list");
roland@4589 3723 assert(n->is_expensive(), "expensive nodes with non-null control here only");
roland@4589 3724 assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here");
roland@4589 3725 if (OptimizeExpensiveOps) {
roland@4589 3726 _expensive_nodes->append(n);
roland@4589 3727 } else {
roland@4589 3728 // Clear control input and let IGVN optimize expensive nodes if
roland@4589 3729 // OptimizeExpensiveOps is off.
roland@4589 3730 n->set_req(0, NULL);
roland@4589 3731 }
roland@4589 3732 }
shade@4691 3733
shade@4691 3734 // Auxiliary method to support randomized stressing/fuzzing.
shade@4691 3735 //
shade@4691 3736 // This method can be called the arbitrary number of times, with current count
shade@4691 3737 // as the argument. The logic allows selecting a single candidate from the
shade@4691 3738 // running list of candidates as follows:
shade@4691 3739 // int count = 0;
shade@4691 3740 // Cand* selected = null;
shade@4691 3741 // while(cand = cand->next()) {
shade@4691 3742 // if (randomized_select(++count)) {
shade@4691 3743 // selected = cand;
shade@4691 3744 // }
shade@4691 3745 // }
shade@4691 3746 //
shade@4691 3747 // Including count equalizes the chances any candidate is "selected".
shade@4691 3748 // This is useful when we don't have the complete list of candidates to choose
shade@4691 3749 // from uniformly. In this case, we need to adjust the randomicity of the
shade@4691 3750 // selection, or else we will end up biasing the selection towards the latter
shade@4691 3751 // candidates.
shade@4691 3752 //
shade@4691 3753 // Quick back-envelope calculation shows that for the list of n candidates
shade@4691 3754 // the equal probability for the candidate to persist as "best" can be
shade@4691 3755 // achieved by replacing it with "next" k-th candidate with the probability
shade@4691 3756 // of 1/k. It can be easily shown that by the end of the run, the
shade@4691 3757 // probability for any candidate is converged to 1/n, thus giving the
shade@4691 3758 // uniform distribution among all the candidates.
shade@4691 3759 //
shade@4691 3760 // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large.
shade@4691 3761 #define RANDOMIZED_DOMAIN_POW 29
shade@4691 3762 #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW)
shade@4691 3763 #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1)
shade@4691 3764 bool Compile::randomized_select(int count) {
shade@4691 3765 assert(count > 0, "only positive");
shade@4691 3766 return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count);
shade@4691 3767 }

mercurial