1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000 1.2 +++ b/src/share/vm/opto/parse2.cpp Wed Apr 27 01:25:04 2016 +0800 1.3 @@ -0,0 +1,2403 @@ 1.4 +/* 1.5 + * Copyright (c) 1998, 2014, Oracle and/or its affiliates. All rights reserved. 1.6 + * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 1.7 + * 1.8 + * This code is free software; you can redistribute it and/or modify it 1.9 + * under the terms of the GNU General Public License version 2 only, as 1.10 + * published by the Free Software Foundation. 1.11 + * 1.12 + * This code is distributed in the hope that it will be useful, but WITHOUT 1.13 + * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 1.14 + * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 1.15 + * version 2 for more details (a copy is included in the LICENSE file that 1.16 + * accompanied this code). 1.17 + * 1.18 + * You should have received a copy of the GNU General Public License version 1.19 + * 2 along with this work; if not, write to the Free Software Foundation, 1.20 + * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 1.21 + * 1.22 + * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 1.23 + * or visit www.oracle.com if you need additional information or have any 1.24 + * questions. 1.25 + * 1.26 + */ 1.27 + 1.28 +#include "precompiled.hpp" 1.29 +#include "ci/ciMethodData.hpp" 1.30 +#include "classfile/systemDictionary.hpp" 1.31 +#include "classfile/vmSymbols.hpp" 1.32 +#include "compiler/compileLog.hpp" 1.33 +#include "interpreter/linkResolver.hpp" 1.34 +#include "memory/universe.inline.hpp" 1.35 +#include "opto/addnode.hpp" 1.36 +#include "opto/divnode.hpp" 1.37 +#include "opto/idealGraphPrinter.hpp" 1.38 +#include "opto/matcher.hpp" 1.39 +#include "opto/memnode.hpp" 1.40 +#include "opto/mulnode.hpp" 1.41 +#include "opto/parse.hpp" 1.42 +#include "opto/runtime.hpp" 1.43 +#include "runtime/deoptimization.hpp" 1.44 +#include "runtime/sharedRuntime.hpp" 1.45 + 1.46 +extern int explicit_null_checks_inserted, 1.47 + explicit_null_checks_elided; 1.48 + 1.49 +//---------------------------------array_load---------------------------------- 1.50 +void Parse::array_load(BasicType elem_type) { 1.51 + const Type* elem = Type::TOP; 1.52 + Node* adr = array_addressing(elem_type, 0, &elem); 1.53 + if (stopped()) return; // guaranteed null or range check 1.54 + dec_sp(2); // Pop array and index 1.55 + const TypeAryPtr* adr_type = TypeAryPtr::get_array_body_type(elem_type); 1.56 + Node* ld = make_load(control(), adr, elem, elem_type, adr_type, MemNode::unordered); 1.57 + push(ld); 1.58 +} 1.59 + 1.60 + 1.61 +//--------------------------------array_store---------------------------------- 1.62 +void Parse::array_store(BasicType elem_type) { 1.63 + Node* adr = array_addressing(elem_type, 1); 1.64 + if (stopped()) return; // guaranteed null or range check 1.65 + Node* val = pop(); 1.66 + dec_sp(2); // Pop array and index 1.67 + const TypeAryPtr* adr_type = TypeAryPtr::get_array_body_type(elem_type); 1.68 + store_to_memory(control(), adr, val, elem_type, adr_type, StoreNode::release_if_reference(elem_type)); 1.69 +} 1.70 + 1.71 + 1.72 +//------------------------------array_addressing------------------------------- 1.73 +// Pull array and index from the stack. Compute pointer-to-element. 1.74 +Node* Parse::array_addressing(BasicType type, int vals, const Type* *result2) { 1.75 + Node *idx = peek(0+vals); // Get from stack without popping 1.76 + Node *ary = peek(1+vals); // in case of exception 1.77 + 1.78 + // Null check the array base, with correct stack contents 1.79 + ary = null_check(ary, T_ARRAY); 1.80 + // Compile-time detect of null-exception? 1.81 + if (stopped()) return top(); 1.82 + 1.83 + const TypeAryPtr* arytype = _gvn.type(ary)->is_aryptr(); 1.84 + const TypeInt* sizetype = arytype->size(); 1.85 + const Type* elemtype = arytype->elem(); 1.86 + 1.87 + if (UseUniqueSubclasses && result2 != NULL) { 1.88 + const Type* el = elemtype->make_ptr(); 1.89 + if (el && el->isa_instptr()) { 1.90 + const TypeInstPtr* toop = el->is_instptr(); 1.91 + if (toop->klass()->as_instance_klass()->unique_concrete_subklass()) { 1.92 + // If we load from "AbstractClass[]" we must see "ConcreteSubClass". 1.93 + const Type* subklass = Type::get_const_type(toop->klass()); 1.94 + elemtype = subklass->join_speculative(el); 1.95 + } 1.96 + } 1.97 + } 1.98 + 1.99 + // Check for big class initializers with all constant offsets 1.100 + // feeding into a known-size array. 1.101 + const TypeInt* idxtype = _gvn.type(idx)->is_int(); 1.102 + // See if the highest idx value is less than the lowest array bound, 1.103 + // and if the idx value cannot be negative: 1.104 + bool need_range_check = true; 1.105 + if (idxtype->_hi < sizetype->_lo && idxtype->_lo >= 0) { 1.106 + need_range_check = false; 1.107 + if (C->log() != NULL) C->log()->elem("observe that='!need_range_check'"); 1.108 + } 1.109 + 1.110 + ciKlass * arytype_klass = arytype->klass(); 1.111 + if ((arytype_klass != NULL) && (!arytype_klass->is_loaded())) { 1.112 + // Only fails for some -Xcomp runs 1.113 + // The class is unloaded. We have to run this bytecode in the interpreter. 1.114 + uncommon_trap(Deoptimization::Reason_unloaded, 1.115 + Deoptimization::Action_reinterpret, 1.116 + arytype->klass(), "!loaded array"); 1.117 + return top(); 1.118 + } 1.119 + 1.120 + // Do the range check 1.121 + if (GenerateRangeChecks && need_range_check) { 1.122 + Node* tst; 1.123 + if (sizetype->_hi <= 0) { 1.124 + // The greatest array bound is negative, so we can conclude that we're 1.125 + // compiling unreachable code, but the unsigned compare trick used below 1.126 + // only works with non-negative lengths. Instead, hack "tst" to be zero so 1.127 + // the uncommon_trap path will always be taken. 1.128 + tst = _gvn.intcon(0); 1.129 + } else { 1.130 + // Range is constant in array-oop, so we can use the original state of mem 1.131 + Node* len = load_array_length(ary); 1.132 + 1.133 + // Test length vs index (standard trick using unsigned compare) 1.134 + Node* chk = _gvn.transform( new (C) CmpUNode(idx, len) ); 1.135 + BoolTest::mask btest = BoolTest::lt; 1.136 + tst = _gvn.transform( new (C) BoolNode(chk, btest) ); 1.137 + } 1.138 + // Branch to failure if out of bounds 1.139 + { BuildCutout unless(this, tst, PROB_MAX); 1.140 + if (C->allow_range_check_smearing()) { 1.141 + // Do not use builtin_throw, since range checks are sometimes 1.142 + // made more stringent by an optimistic transformation. 1.143 + // This creates "tentative" range checks at this point, 1.144 + // which are not guaranteed to throw exceptions. 1.145 + // See IfNode::Ideal, is_range_check, adjust_check. 1.146 + uncommon_trap(Deoptimization::Reason_range_check, 1.147 + Deoptimization::Action_make_not_entrant, 1.148 + NULL, "range_check"); 1.149 + } else { 1.150 + // If we have already recompiled with the range-check-widening 1.151 + // heroic optimization turned off, then we must really be throwing 1.152 + // range check exceptions. 1.153 + builtin_throw(Deoptimization::Reason_range_check, idx); 1.154 + } 1.155 + } 1.156 + } 1.157 + // Check for always knowing you are throwing a range-check exception 1.158 + if (stopped()) return top(); 1.159 + 1.160 + Node* ptr = array_element_address(ary, idx, type, sizetype); 1.161 + 1.162 + if (result2 != NULL) *result2 = elemtype; 1.163 + 1.164 + assert(ptr != top(), "top should go hand-in-hand with stopped"); 1.165 + 1.166 + return ptr; 1.167 +} 1.168 + 1.169 + 1.170 +// returns IfNode 1.171 +IfNode* Parse::jump_if_fork_int(Node* a, Node* b, BoolTest::mask mask) { 1.172 + Node *cmp = _gvn.transform( new (C) CmpINode( a, b)); // two cases: shiftcount > 32 and shiftcount <= 32 1.173 + Node *tst = _gvn.transform( new (C) BoolNode( cmp, mask)); 1.174 + IfNode *iff = create_and_map_if( control(), tst, ((mask == BoolTest::eq) ? PROB_STATIC_INFREQUENT : PROB_FAIR), COUNT_UNKNOWN ); 1.175 + return iff; 1.176 +} 1.177 + 1.178 +// return Region node 1.179 +Node* Parse::jump_if_join(Node* iffalse, Node* iftrue) { 1.180 + Node *region = new (C) RegionNode(3); // 2 results 1.181 + record_for_igvn(region); 1.182 + region->init_req(1, iffalse); 1.183 + region->init_req(2, iftrue ); 1.184 + _gvn.set_type(region, Type::CONTROL); 1.185 + region = _gvn.transform(region); 1.186 + set_control (region); 1.187 + return region; 1.188 +} 1.189 + 1.190 + 1.191 +//------------------------------helper for tableswitch------------------------- 1.192 +void Parse::jump_if_true_fork(IfNode *iff, int dest_bci_if_true, int prof_table_index) { 1.193 + // True branch, use existing map info 1.194 + { PreserveJVMState pjvms(this); 1.195 + Node *iftrue = _gvn.transform( new (C) IfTrueNode (iff) ); 1.196 + set_control( iftrue ); 1.197 + profile_switch_case(prof_table_index); 1.198 + merge_new_path(dest_bci_if_true); 1.199 + } 1.200 + 1.201 + // False branch 1.202 + Node *iffalse = _gvn.transform( new (C) IfFalseNode(iff) ); 1.203 + set_control( iffalse ); 1.204 +} 1.205 + 1.206 +void Parse::jump_if_false_fork(IfNode *iff, int dest_bci_if_true, int prof_table_index) { 1.207 + // True branch, use existing map info 1.208 + { PreserveJVMState pjvms(this); 1.209 + Node *iffalse = _gvn.transform( new (C) IfFalseNode (iff) ); 1.210 + set_control( iffalse ); 1.211 + profile_switch_case(prof_table_index); 1.212 + merge_new_path(dest_bci_if_true); 1.213 + } 1.214 + 1.215 + // False branch 1.216 + Node *iftrue = _gvn.transform( new (C) IfTrueNode(iff) ); 1.217 + set_control( iftrue ); 1.218 +} 1.219 + 1.220 +void Parse::jump_if_always_fork(int dest_bci, int prof_table_index) { 1.221 + // False branch, use existing map and control() 1.222 + profile_switch_case(prof_table_index); 1.223 + merge_new_path(dest_bci); 1.224 +} 1.225 + 1.226 + 1.227 +extern "C" { 1.228 + static int jint_cmp(const void *i, const void *j) { 1.229 + int a = *(jint *)i; 1.230 + int b = *(jint *)j; 1.231 + return a > b ? 1 : a < b ? -1 : 0; 1.232 + } 1.233 +} 1.234 + 1.235 + 1.236 +// Default value for methodData switch indexing. Must be a negative value to avoid 1.237 +// conflict with any legal switch index. 1.238 +#define NullTableIndex -1 1.239 + 1.240 +class SwitchRange : public StackObj { 1.241 + // a range of integers coupled with a bci destination 1.242 + jint _lo; // inclusive lower limit 1.243 + jint _hi; // inclusive upper limit 1.244 + int _dest; 1.245 + int _table_index; // index into method data table 1.246 + 1.247 +public: 1.248 + jint lo() const { return _lo; } 1.249 + jint hi() const { return _hi; } 1.250 + int dest() const { return _dest; } 1.251 + int table_index() const { return _table_index; } 1.252 + bool is_singleton() const { return _lo == _hi; } 1.253 + 1.254 + void setRange(jint lo, jint hi, int dest, int table_index) { 1.255 + assert(lo <= hi, "must be a non-empty range"); 1.256 + _lo = lo, _hi = hi; _dest = dest; _table_index = table_index; 1.257 + } 1.258 + bool adjoinRange(jint lo, jint hi, int dest, int table_index) { 1.259 + assert(lo <= hi, "must be a non-empty range"); 1.260 + if (lo == _hi+1 && dest == _dest && table_index == _table_index) { 1.261 + _hi = hi; 1.262 + return true; 1.263 + } 1.264 + return false; 1.265 + } 1.266 + 1.267 + void set (jint value, int dest, int table_index) { 1.268 + setRange(value, value, dest, table_index); 1.269 + } 1.270 + bool adjoin(jint value, int dest, int table_index) { 1.271 + return adjoinRange(value, value, dest, table_index); 1.272 + } 1.273 + 1.274 + void print() { 1.275 + if (is_singleton()) 1.276 + tty->print(" {%d}=>%d", lo(), dest()); 1.277 + else if (lo() == min_jint) 1.278 + tty->print(" {..%d}=>%d", hi(), dest()); 1.279 + else if (hi() == max_jint) 1.280 + tty->print(" {%d..}=>%d", lo(), dest()); 1.281 + else 1.282 + tty->print(" {%d..%d}=>%d", lo(), hi(), dest()); 1.283 + } 1.284 +}; 1.285 + 1.286 + 1.287 +//-------------------------------do_tableswitch-------------------------------- 1.288 +void Parse::do_tableswitch() { 1.289 + Node* lookup = pop(); 1.290 + 1.291 + // Get information about tableswitch 1.292 + int default_dest = iter().get_dest_table(0); 1.293 + int lo_index = iter().get_int_table(1); 1.294 + int hi_index = iter().get_int_table(2); 1.295 + int len = hi_index - lo_index + 1; 1.296 + 1.297 + if (len < 1) { 1.298 + // If this is a backward branch, add safepoint 1.299 + maybe_add_safepoint(default_dest); 1.300 + merge(default_dest); 1.301 + return; 1.302 + } 1.303 + 1.304 + // generate decision tree, using trichotomy when possible 1.305 + int rnum = len+2; 1.306 + bool makes_backward_branch = false; 1.307 + SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum); 1.308 + int rp = -1; 1.309 + if (lo_index != min_jint) { 1.310 + ranges[++rp].setRange(min_jint, lo_index-1, default_dest, NullTableIndex); 1.311 + } 1.312 + for (int j = 0; j < len; j++) { 1.313 + jint match_int = lo_index+j; 1.314 + int dest = iter().get_dest_table(j+3); 1.315 + makes_backward_branch |= (dest <= bci()); 1.316 + int table_index = method_data_update() ? j : NullTableIndex; 1.317 + if (rp < 0 || !ranges[rp].adjoin(match_int, dest, table_index)) { 1.318 + ranges[++rp].set(match_int, dest, table_index); 1.319 + } 1.320 + } 1.321 + jint highest = lo_index+(len-1); 1.322 + assert(ranges[rp].hi() == highest, ""); 1.323 + if (highest != max_jint 1.324 + && !ranges[rp].adjoinRange(highest+1, max_jint, default_dest, NullTableIndex)) { 1.325 + ranges[++rp].setRange(highest+1, max_jint, default_dest, NullTableIndex); 1.326 + } 1.327 + assert(rp < len+2, "not too many ranges"); 1.328 + 1.329 + // Safepoint in case if backward branch observed 1.330 + if( makes_backward_branch && UseLoopSafepoints ) 1.331 + add_safepoint(); 1.332 + 1.333 + jump_switch_ranges(lookup, &ranges[0], &ranges[rp]); 1.334 +} 1.335 + 1.336 + 1.337 +//------------------------------do_lookupswitch-------------------------------- 1.338 +void Parse::do_lookupswitch() { 1.339 + Node *lookup = pop(); // lookup value 1.340 + // Get information about lookupswitch 1.341 + int default_dest = iter().get_dest_table(0); 1.342 + int len = iter().get_int_table(1); 1.343 + 1.344 + if (len < 1) { // If this is a backward branch, add safepoint 1.345 + maybe_add_safepoint(default_dest); 1.346 + merge(default_dest); 1.347 + return; 1.348 + } 1.349 + 1.350 + // generate decision tree, using trichotomy when possible 1.351 + jint* table = NEW_RESOURCE_ARRAY(jint, len*2); 1.352 + { 1.353 + for( int j = 0; j < len; j++ ) { 1.354 + table[j+j+0] = iter().get_int_table(2+j+j); 1.355 + table[j+j+1] = iter().get_dest_table(2+j+j+1); 1.356 + } 1.357 + qsort( table, len, 2*sizeof(table[0]), jint_cmp ); 1.358 + } 1.359 + 1.360 + int rnum = len*2+1; 1.361 + bool makes_backward_branch = false; 1.362 + SwitchRange* ranges = NEW_RESOURCE_ARRAY(SwitchRange, rnum); 1.363 + int rp = -1; 1.364 + for( int j = 0; j < len; j++ ) { 1.365 + jint match_int = table[j+j+0]; 1.366 + int dest = table[j+j+1]; 1.367 + int next_lo = rp < 0 ? min_jint : ranges[rp].hi()+1; 1.368 + int table_index = method_data_update() ? j : NullTableIndex; 1.369 + makes_backward_branch |= (dest <= bci()); 1.370 + if( match_int != next_lo ) { 1.371 + ranges[++rp].setRange(next_lo, match_int-1, default_dest, NullTableIndex); 1.372 + } 1.373 + if( rp < 0 || !ranges[rp].adjoin(match_int, dest, table_index) ) { 1.374 + ranges[++rp].set(match_int, dest, table_index); 1.375 + } 1.376 + } 1.377 + jint highest = table[2*(len-1)]; 1.378 + assert(ranges[rp].hi() == highest, ""); 1.379 + if( highest != max_jint 1.380 + && !ranges[rp].adjoinRange(highest+1, max_jint, default_dest, NullTableIndex) ) { 1.381 + ranges[++rp].setRange(highest+1, max_jint, default_dest, NullTableIndex); 1.382 + } 1.383 + assert(rp < rnum, "not too many ranges"); 1.384 + 1.385 + // Safepoint in case backward branch observed 1.386 + if( makes_backward_branch && UseLoopSafepoints ) 1.387 + add_safepoint(); 1.388 + 1.389 + jump_switch_ranges(lookup, &ranges[0], &ranges[rp]); 1.390 +} 1.391 + 1.392 +//----------------------------create_jump_tables------------------------------- 1.393 +bool Parse::create_jump_tables(Node* key_val, SwitchRange* lo, SwitchRange* hi) { 1.394 + // Are jumptables enabled 1.395 + if (!UseJumpTables) return false; 1.396 + 1.397 + // Are jumptables supported 1.398 + if (!Matcher::has_match_rule(Op_Jump)) return false; 1.399 + 1.400 + // Don't make jump table if profiling 1.401 + if (method_data_update()) return false; 1.402 + 1.403 + // Decide if a guard is needed to lop off big ranges at either (or 1.404 + // both) end(s) of the input set. We'll call this the default target 1.405 + // even though we can't be sure that it is the true "default". 1.406 + 1.407 + bool needs_guard = false; 1.408 + int default_dest; 1.409 + int64 total_outlier_size = 0; 1.410 + int64 hi_size = ((int64)hi->hi()) - ((int64)hi->lo()) + 1; 1.411 + int64 lo_size = ((int64)lo->hi()) - ((int64)lo->lo()) + 1; 1.412 + 1.413 + if (lo->dest() == hi->dest()) { 1.414 + total_outlier_size = hi_size + lo_size; 1.415 + default_dest = lo->dest(); 1.416 + } else if (lo_size > hi_size) { 1.417 + total_outlier_size = lo_size; 1.418 + default_dest = lo->dest(); 1.419 + } else { 1.420 + total_outlier_size = hi_size; 1.421 + default_dest = hi->dest(); 1.422 + } 1.423 + 1.424 + // If a guard test will eliminate very sparse end ranges, then 1.425 + // it is worth the cost of an extra jump. 1.426 + if (total_outlier_size > (MaxJumpTableSparseness * 4)) { 1.427 + needs_guard = true; 1.428 + if (default_dest == lo->dest()) lo++; 1.429 + if (default_dest == hi->dest()) hi--; 1.430 + } 1.431 + 1.432 + // Find the total number of cases and ranges 1.433 + int64 num_cases = ((int64)hi->hi()) - ((int64)lo->lo()) + 1; 1.434 + int num_range = hi - lo + 1; 1.435 + 1.436 + // Don't create table if: too large, too small, or too sparse. 1.437 + if (num_cases < MinJumpTableSize || num_cases > MaxJumpTableSize) 1.438 + return false; 1.439 + if (num_cases > (MaxJumpTableSparseness * num_range)) 1.440 + return false; 1.441 + 1.442 + // Normalize table lookups to zero 1.443 + int lowval = lo->lo(); 1.444 + key_val = _gvn.transform( new (C) SubINode(key_val, _gvn.intcon(lowval)) ); 1.445 + 1.446 + // Generate a guard to protect against input keyvals that aren't 1.447 + // in the switch domain. 1.448 + if (needs_guard) { 1.449 + Node* size = _gvn.intcon(num_cases); 1.450 + Node* cmp = _gvn.transform( new (C) CmpUNode(key_val, size) ); 1.451 + Node* tst = _gvn.transform( new (C) BoolNode(cmp, BoolTest::ge) ); 1.452 + IfNode* iff = create_and_map_if( control(), tst, PROB_FAIR, COUNT_UNKNOWN); 1.453 + jump_if_true_fork(iff, default_dest, NullTableIndex); 1.454 + } 1.455 + 1.456 + // Create an ideal node JumpTable that has projections 1.457 + // of all possible ranges for a switch statement 1.458 + // The key_val input must be converted to a pointer offset and scaled. 1.459 + // Compare Parse::array_addressing above. 1.460 +#ifdef _LP64 1.461 + // Clean the 32-bit int into a real 64-bit offset. 1.462 + // Otherwise, the jint value 0 might turn into an offset of 0x0800000000. 1.463 + const TypeLong* lkeytype = TypeLong::make(CONST64(0), num_cases-1, Type::WidenMin); 1.464 + key_val = _gvn.transform( new (C) ConvI2LNode(key_val, lkeytype) ); 1.465 +#endif 1.466 + // Shift the value by wordsize so we have an index into the table, rather 1.467 + // than a switch value 1.468 + Node *shiftWord = _gvn.MakeConX(wordSize); 1.469 + key_val = _gvn.transform( new (C) MulXNode( key_val, shiftWord)); 1.470 + 1.471 + // Create the JumpNode 1.472 + Node* jtn = _gvn.transform( new (C) JumpNode(control(), key_val, num_cases) ); 1.473 + 1.474 + // These are the switch destinations hanging off the jumpnode 1.475 + int i = 0; 1.476 + for (SwitchRange* r = lo; r <= hi; r++) { 1.477 + for (int64 j = r->lo(); j <= r->hi(); j++, i++) { 1.478 + Node* input = _gvn.transform(new (C) JumpProjNode(jtn, i, r->dest(), (int)(j - lowval))); 1.479 + { 1.480 + PreserveJVMState pjvms(this); 1.481 + set_control(input); 1.482 + jump_if_always_fork(r->dest(), r->table_index()); 1.483 + } 1.484 + } 1.485 + } 1.486 + assert(i == num_cases, "miscount of cases"); 1.487 + stop_and_kill_map(); // no more uses for this JVMS 1.488 + return true; 1.489 +} 1.490 + 1.491 +//----------------------------jump_switch_ranges------------------------------- 1.492 +void Parse::jump_switch_ranges(Node* key_val, SwitchRange *lo, SwitchRange *hi, int switch_depth) { 1.493 + Block* switch_block = block(); 1.494 + 1.495 + if (switch_depth == 0) { 1.496 + // Do special processing for the top-level call. 1.497 + assert(lo->lo() == min_jint, "initial range must exhaust Type::INT"); 1.498 + assert(hi->hi() == max_jint, "initial range must exhaust Type::INT"); 1.499 + 1.500 + // Decrement pred-numbers for the unique set of nodes. 1.501 +#ifdef ASSERT 1.502 + // Ensure that the block's successors are a (duplicate-free) set. 1.503 + int successors_counted = 0; // block occurrences in [hi..lo] 1.504 + int unique_successors = switch_block->num_successors(); 1.505 + for (int i = 0; i < unique_successors; i++) { 1.506 + Block* target = switch_block->successor_at(i); 1.507 + 1.508 + // Check that the set of successors is the same in both places. 1.509 + int successors_found = 0; 1.510 + for (SwitchRange* p = lo; p <= hi; p++) { 1.511 + if (p->dest() == target->start()) successors_found++; 1.512 + } 1.513 + assert(successors_found > 0, "successor must be known"); 1.514 + successors_counted += successors_found; 1.515 + } 1.516 + assert(successors_counted == (hi-lo)+1, "no unexpected successors"); 1.517 +#endif 1.518 + 1.519 + // Maybe prune the inputs, based on the type of key_val. 1.520 + jint min_val = min_jint; 1.521 + jint max_val = max_jint; 1.522 + const TypeInt* ti = key_val->bottom_type()->isa_int(); 1.523 + if (ti != NULL) { 1.524 + min_val = ti->_lo; 1.525 + max_val = ti->_hi; 1.526 + assert(min_val <= max_val, "invalid int type"); 1.527 + } 1.528 + while (lo->hi() < min_val) lo++; 1.529 + if (lo->lo() < min_val) lo->setRange(min_val, lo->hi(), lo->dest(), lo->table_index()); 1.530 + while (hi->lo() > max_val) hi--; 1.531 + if (hi->hi() > max_val) hi->setRange(hi->lo(), max_val, hi->dest(), hi->table_index()); 1.532 + } 1.533 + 1.534 +#ifndef PRODUCT 1.535 + if (switch_depth == 0) { 1.536 + _max_switch_depth = 0; 1.537 + _est_switch_depth = log2_intptr((hi-lo+1)-1)+1; 1.538 + } 1.539 +#endif 1.540 + 1.541 + assert(lo <= hi, "must be a non-empty set of ranges"); 1.542 + if (lo == hi) { 1.543 + jump_if_always_fork(lo->dest(), lo->table_index()); 1.544 + } else { 1.545 + assert(lo->hi() == (lo+1)->lo()-1, "contiguous ranges"); 1.546 + assert(hi->lo() == (hi-1)->hi()+1, "contiguous ranges"); 1.547 + 1.548 + if (create_jump_tables(key_val, lo, hi)) return; 1.549 + 1.550 + int nr = hi - lo + 1; 1.551 + 1.552 + SwitchRange* mid = lo + nr/2; 1.553 + // if there is an easy choice, pivot at a singleton: 1.554 + if (nr > 3 && !mid->is_singleton() && (mid-1)->is_singleton()) mid--; 1.555 + 1.556 + assert(lo < mid && mid <= hi, "good pivot choice"); 1.557 + assert(nr != 2 || mid == hi, "should pick higher of 2"); 1.558 + assert(nr != 3 || mid == hi-1, "should pick middle of 3"); 1.559 + 1.560 + Node *test_val = _gvn.intcon(mid->lo()); 1.561 + 1.562 + if (mid->is_singleton()) { 1.563 + IfNode *iff_ne = jump_if_fork_int(key_val, test_val, BoolTest::ne); 1.564 + jump_if_false_fork(iff_ne, mid->dest(), mid->table_index()); 1.565 + 1.566 + // Special Case: If there are exactly three ranges, and the high 1.567 + // and low range each go to the same place, omit the "gt" test, 1.568 + // since it will not discriminate anything. 1.569 + bool eq_test_only = (hi == lo+2 && hi->dest() == lo->dest()); 1.570 + if (eq_test_only) { 1.571 + assert(mid == hi-1, ""); 1.572 + } 1.573 + 1.574 + // if there is a higher range, test for it and process it: 1.575 + if (mid < hi && !eq_test_only) { 1.576 + // two comparisons of same values--should enable 1 test for 2 branches 1.577 + // Use BoolTest::le instead of BoolTest::gt 1.578 + IfNode *iff_le = jump_if_fork_int(key_val, test_val, BoolTest::le); 1.579 + Node *iftrue = _gvn.transform( new (C) IfTrueNode(iff_le) ); 1.580 + Node *iffalse = _gvn.transform( new (C) IfFalseNode(iff_le) ); 1.581 + { PreserveJVMState pjvms(this); 1.582 + set_control(iffalse); 1.583 + jump_switch_ranges(key_val, mid+1, hi, switch_depth+1); 1.584 + } 1.585 + set_control(iftrue); 1.586 + } 1.587 + 1.588 + } else { 1.589 + // mid is a range, not a singleton, so treat mid..hi as a unit 1.590 + IfNode *iff_ge = jump_if_fork_int(key_val, test_val, BoolTest::ge); 1.591 + 1.592 + // if there is a higher range, test for it and process it: 1.593 + if (mid == hi) { 1.594 + jump_if_true_fork(iff_ge, mid->dest(), mid->table_index()); 1.595 + } else { 1.596 + Node *iftrue = _gvn.transform( new (C) IfTrueNode(iff_ge) ); 1.597 + Node *iffalse = _gvn.transform( new (C) IfFalseNode(iff_ge) ); 1.598 + { PreserveJVMState pjvms(this); 1.599 + set_control(iftrue); 1.600 + jump_switch_ranges(key_val, mid, hi, switch_depth+1); 1.601 + } 1.602 + set_control(iffalse); 1.603 + } 1.604 + } 1.605 + 1.606 + // in any case, process the lower range 1.607 + jump_switch_ranges(key_val, lo, mid-1, switch_depth+1); 1.608 + } 1.609 + 1.610 + // Decrease pred_count for each successor after all is done. 1.611 + if (switch_depth == 0) { 1.612 + int unique_successors = switch_block->num_successors(); 1.613 + for (int i = 0; i < unique_successors; i++) { 1.614 + Block* target = switch_block->successor_at(i); 1.615 + // Throw away the pre-allocated path for each unique successor. 1.616 + target->next_path_num(); 1.617 + } 1.618 + } 1.619 + 1.620 +#ifndef PRODUCT 1.621 + _max_switch_depth = MAX2(switch_depth, _max_switch_depth); 1.622 + if (TraceOptoParse && Verbose && WizardMode && switch_depth == 0) { 1.623 + SwitchRange* r; 1.624 + int nsing = 0; 1.625 + for( r = lo; r <= hi; r++ ) { 1.626 + if( r->is_singleton() ) nsing++; 1.627 + } 1.628 + tty->print(">>> "); 1.629 + _method->print_short_name(); 1.630 + tty->print_cr(" switch decision tree"); 1.631 + tty->print_cr(" %d ranges (%d singletons), max_depth=%d, est_depth=%d", 1.632 + (int) (hi-lo+1), nsing, _max_switch_depth, _est_switch_depth); 1.633 + if (_max_switch_depth > _est_switch_depth) { 1.634 + tty->print_cr("******** BAD SWITCH DEPTH ********"); 1.635 + } 1.636 + tty->print(" "); 1.637 + for( r = lo; r <= hi; r++ ) { 1.638 + r->print(); 1.639 + } 1.640 + tty->cr(); 1.641 + } 1.642 +#endif 1.643 +} 1.644 + 1.645 +void Parse::modf() { 1.646 + Node *f2 = pop(); 1.647 + Node *f1 = pop(); 1.648 + Node* c = make_runtime_call(RC_LEAF, OptoRuntime::modf_Type(), 1.649 + CAST_FROM_FN_PTR(address, SharedRuntime::frem), 1.650 + "frem", NULL, //no memory effects 1.651 + f1, f2); 1.652 + Node* res = _gvn.transform(new (C) ProjNode(c, TypeFunc::Parms + 0)); 1.653 + 1.654 + push(res); 1.655 +} 1.656 + 1.657 +void Parse::modd() { 1.658 + Node *d2 = pop_pair(); 1.659 + Node *d1 = pop_pair(); 1.660 + Node* c = make_runtime_call(RC_LEAF, OptoRuntime::Math_DD_D_Type(), 1.661 + CAST_FROM_FN_PTR(address, SharedRuntime::drem), 1.662 + "drem", NULL, //no memory effects 1.663 + d1, top(), d2, top()); 1.664 + Node* res_d = _gvn.transform(new (C) ProjNode(c, TypeFunc::Parms + 0)); 1.665 + 1.666 +#ifdef ASSERT 1.667 + Node* res_top = _gvn.transform(new (C) ProjNode(c, TypeFunc::Parms + 1)); 1.668 + assert(res_top == top(), "second value must be top"); 1.669 +#endif 1.670 + 1.671 + push_pair(res_d); 1.672 +} 1.673 + 1.674 +void Parse::l2f() { 1.675 + Node* f2 = pop(); 1.676 + Node* f1 = pop(); 1.677 + Node* c = make_runtime_call(RC_LEAF, OptoRuntime::l2f_Type(), 1.678 + CAST_FROM_FN_PTR(address, SharedRuntime::l2f), 1.679 + "l2f", NULL, //no memory effects 1.680 + f1, f2); 1.681 + Node* res = _gvn.transform(new (C) ProjNode(c, TypeFunc::Parms + 0)); 1.682 + 1.683 + push(res); 1.684 +} 1.685 + 1.686 +void Parse::do_irem() { 1.687 + // Must keep both values on the expression-stack during null-check 1.688 + zero_check_int(peek()); 1.689 + // Compile-time detect of null-exception? 1.690 + if (stopped()) return; 1.691 + 1.692 + Node* b = pop(); 1.693 + Node* a = pop(); 1.694 + 1.695 + const Type *t = _gvn.type(b); 1.696 + if (t != Type::TOP) { 1.697 + const TypeInt *ti = t->is_int(); 1.698 + if (ti->is_con()) { 1.699 + int divisor = ti->get_con(); 1.700 + // check for positive power of 2 1.701 + if (divisor > 0 && 1.702 + (divisor & ~(divisor-1)) == divisor) { 1.703 + // yes ! 1.704 + Node *mask = _gvn.intcon((divisor - 1)); 1.705 + // Sigh, must handle negative dividends 1.706 + Node *zero = _gvn.intcon(0); 1.707 + IfNode *ifff = jump_if_fork_int(a, zero, BoolTest::lt); 1.708 + Node *iff = _gvn.transform( new (C) IfFalseNode(ifff) ); 1.709 + Node *ift = _gvn.transform( new (C) IfTrueNode (ifff) ); 1.710 + Node *reg = jump_if_join(ift, iff); 1.711 + Node *phi = PhiNode::make(reg, NULL, TypeInt::INT); 1.712 + // Negative path; negate/and/negate 1.713 + Node *neg = _gvn.transform( new (C) SubINode(zero, a) ); 1.714 + Node *andn= _gvn.transform( new (C) AndINode(neg, mask) ); 1.715 + Node *negn= _gvn.transform( new (C) SubINode(zero, andn) ); 1.716 + phi->init_req(1, negn); 1.717 + // Fast positive case 1.718 + Node *andx = _gvn.transform( new (C) AndINode(a, mask) ); 1.719 + phi->init_req(2, andx); 1.720 + // Push the merge 1.721 + push( _gvn.transform(phi) ); 1.722 + return; 1.723 + } 1.724 + } 1.725 + } 1.726 + // Default case 1.727 + push( _gvn.transform( new (C) ModINode(control(),a,b) ) ); 1.728 +} 1.729 + 1.730 +// Handle jsr and jsr_w bytecode 1.731 +void Parse::do_jsr() { 1.732 + assert(bc() == Bytecodes::_jsr || bc() == Bytecodes::_jsr_w, "wrong bytecode"); 1.733 + 1.734 + // Store information about current state, tagged with new _jsr_bci 1.735 + int return_bci = iter().next_bci(); 1.736 + int jsr_bci = (bc() == Bytecodes::_jsr) ? iter().get_dest() : iter().get_far_dest(); 1.737 + 1.738 + // Update method data 1.739 + profile_taken_branch(jsr_bci); 1.740 + 1.741 + // The way we do things now, there is only one successor block 1.742 + // for the jsr, because the target code is cloned by ciTypeFlow. 1.743 + Block* target = successor_for_bci(jsr_bci); 1.744 + 1.745 + // What got pushed? 1.746 + const Type* ret_addr = target->peek(); 1.747 + assert(ret_addr->singleton(), "must be a constant (cloned jsr body)"); 1.748 + 1.749 + // Effect on jsr on stack 1.750 + push(_gvn.makecon(ret_addr)); 1.751 + 1.752 + // Flow to the jsr. 1.753 + merge(jsr_bci); 1.754 +} 1.755 + 1.756 +// Handle ret bytecode 1.757 +void Parse::do_ret() { 1.758 + // Find to whom we return. 1.759 + assert(block()->num_successors() == 1, "a ret can only go one place now"); 1.760 + Block* target = block()->successor_at(0); 1.761 + assert(!target->is_ready(), "our arrival must be expected"); 1.762 + profile_ret(target->flow()->start()); 1.763 + int pnum = target->next_path_num(); 1.764 + merge_common(target, pnum); 1.765 +} 1.766 + 1.767 +//--------------------------dynamic_branch_prediction-------------------------- 1.768 +// Try to gather dynamic branch prediction behavior. Return a probability 1.769 +// of the branch being taken and set the "cnt" field. Returns a -1.0 1.770 +// if we need to use static prediction for some reason. 1.771 +float Parse::dynamic_branch_prediction(float &cnt) { 1.772 + ResourceMark rm; 1.773 + 1.774 + cnt = COUNT_UNKNOWN; 1.775 + 1.776 + // Use MethodData information if it is available 1.777 + // FIXME: free the ProfileData structure 1.778 + ciMethodData* methodData = method()->method_data(); 1.779 + if (!methodData->is_mature()) return PROB_UNKNOWN; 1.780 + ciProfileData* data = methodData->bci_to_data(bci()); 1.781 + if (!data->is_JumpData()) return PROB_UNKNOWN; 1.782 + 1.783 + // get taken and not taken values 1.784 + int taken = data->as_JumpData()->taken(); 1.785 + int not_taken = 0; 1.786 + if (data->is_BranchData()) { 1.787 + not_taken = data->as_BranchData()->not_taken(); 1.788 + } 1.789 + 1.790 + // scale the counts to be commensurate with invocation counts: 1.791 + taken = method()->scale_count(taken); 1.792 + not_taken = method()->scale_count(not_taken); 1.793 + 1.794 + // Give up if too few (or too many, in which case the sum will overflow) counts to be meaningful. 1.795 + // We also check that individual counters are positive first, overwise the sum can become positive. 1.796 + if (taken < 0 || not_taken < 0 || taken + not_taken < 40) { 1.797 + if (C->log() != NULL) { 1.798 + C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d'", iter().get_dest(), taken, not_taken); 1.799 + } 1.800 + return PROB_UNKNOWN; 1.801 + } 1.802 + 1.803 + // Compute frequency that we arrive here 1.804 + float sum = taken + not_taken; 1.805 + // Adjust, if this block is a cloned private block but the 1.806 + // Jump counts are shared. Taken the private counts for 1.807 + // just this path instead of the shared counts. 1.808 + if( block()->count() > 0 ) 1.809 + sum = block()->count(); 1.810 + cnt = sum / FreqCountInvocations; 1.811 + 1.812 + // Pin probability to sane limits 1.813 + float prob; 1.814 + if( !taken ) 1.815 + prob = (0+PROB_MIN) / 2; 1.816 + else if( !not_taken ) 1.817 + prob = (1+PROB_MAX) / 2; 1.818 + else { // Compute probability of true path 1.819 + prob = (float)taken / (float)(taken + not_taken); 1.820 + if (prob > PROB_MAX) prob = PROB_MAX; 1.821 + if (prob < PROB_MIN) prob = PROB_MIN; 1.822 + } 1.823 + 1.824 + assert((cnt > 0.0f) && (prob > 0.0f), 1.825 + "Bad frequency assignment in if"); 1.826 + 1.827 + if (C->log() != NULL) { 1.828 + const char* prob_str = NULL; 1.829 + if (prob >= PROB_MAX) prob_str = (prob == PROB_MAX) ? "max" : "always"; 1.830 + if (prob <= PROB_MIN) prob_str = (prob == PROB_MIN) ? "min" : "never"; 1.831 + char prob_str_buf[30]; 1.832 + if (prob_str == NULL) { 1.833 + sprintf(prob_str_buf, "%g", prob); 1.834 + prob_str = prob_str_buf; 1.835 + } 1.836 + C->log()->elem("branch target_bci='%d' taken='%d' not_taken='%d' cnt='%g' prob='%s'", 1.837 + iter().get_dest(), taken, not_taken, cnt, prob_str); 1.838 + } 1.839 + return prob; 1.840 +} 1.841 + 1.842 +//-----------------------------branch_prediction------------------------------- 1.843 +float Parse::branch_prediction(float& cnt, 1.844 + BoolTest::mask btest, 1.845 + int target_bci) { 1.846 + float prob = dynamic_branch_prediction(cnt); 1.847 + // If prob is unknown, switch to static prediction 1.848 + if (prob != PROB_UNKNOWN) return prob; 1.849 + 1.850 + prob = PROB_FAIR; // Set default value 1.851 + if (btest == BoolTest::eq) // Exactly equal test? 1.852 + prob = PROB_STATIC_INFREQUENT; // Assume its relatively infrequent 1.853 + else if (btest == BoolTest::ne) 1.854 + prob = PROB_STATIC_FREQUENT; // Assume its relatively frequent 1.855 + 1.856 + // If this is a conditional test guarding a backwards branch, 1.857 + // assume its a loop-back edge. Make it a likely taken branch. 1.858 + if (target_bci < bci()) { 1.859 + if (is_osr_parse()) { // Could be a hot OSR'd loop; force deopt 1.860 + // Since it's an OSR, we probably have profile data, but since 1.861 + // branch_prediction returned PROB_UNKNOWN, the counts are too small. 1.862 + // Let's make a special check here for completely zero counts. 1.863 + ciMethodData* methodData = method()->method_data(); 1.864 + if (!methodData->is_empty()) { 1.865 + ciProfileData* data = methodData->bci_to_data(bci()); 1.866 + // Only stop for truly zero counts, which mean an unknown part 1.867 + // of the OSR-ed method, and we want to deopt to gather more stats. 1.868 + // If you have ANY counts, then this loop is simply 'cold' relative 1.869 + // to the OSR loop. 1.870 + if (data->as_BranchData()->taken() + 1.871 + data->as_BranchData()->not_taken() == 0 ) { 1.872 + // This is the only way to return PROB_UNKNOWN: 1.873 + return PROB_UNKNOWN; 1.874 + } 1.875 + } 1.876 + } 1.877 + prob = PROB_STATIC_FREQUENT; // Likely to take backwards branch 1.878 + } 1.879 + 1.880 + assert(prob != PROB_UNKNOWN, "must have some guess at this point"); 1.881 + return prob; 1.882 +} 1.883 + 1.884 +// The magic constants are chosen so as to match the output of 1.885 +// branch_prediction() when the profile reports a zero taken count. 1.886 +// It is important to distinguish zero counts unambiguously, because 1.887 +// some branches (e.g., _213_javac.Assembler.eliminate) validly produce 1.888 +// very small but nonzero probabilities, which if confused with zero 1.889 +// counts would keep the program recompiling indefinitely. 1.890 +bool Parse::seems_never_taken(float prob) { 1.891 + return prob < PROB_MIN; 1.892 +} 1.893 + 1.894 +// True if the comparison seems to be the kind that will not change its 1.895 +// statistics from true to false. See comments in adjust_map_after_if. 1.896 +// This question is only asked along paths which are already 1.897 +// classifed as untaken (by seems_never_taken), so really, 1.898 +// if a path is never taken, its controlling comparison is 1.899 +// already acting in a stable fashion. If the comparison 1.900 +// seems stable, we will put an expensive uncommon trap 1.901 +// on the untaken path. To be conservative, and to allow 1.902 +// partially executed counted loops to be compiled fully, 1.903 +// we will plant uncommon traps only after pointer comparisons. 1.904 +bool Parse::seems_stable_comparison(BoolTest::mask btest, Node* cmp) { 1.905 + for (int depth = 4; depth > 0; depth--) { 1.906 + // The following switch can find CmpP here over half the time for 1.907 + // dynamic language code rich with type tests. 1.908 + // Code using counted loops or array manipulations (typical 1.909 + // of benchmarks) will have many (>80%) CmpI instructions. 1.910 + switch (cmp->Opcode()) { 1.911 + case Op_CmpP: 1.912 + // A never-taken null check looks like CmpP/BoolTest::eq. 1.913 + // These certainly should be closed off as uncommon traps. 1.914 + if (btest == BoolTest::eq) 1.915 + return true; 1.916 + // A never-failed type check looks like CmpP/BoolTest::ne. 1.917 + // Let's put traps on those, too, so that we don't have to compile 1.918 + // unused paths with indeterminate dynamic type information. 1.919 + if (ProfileDynamicTypes) 1.920 + return true; 1.921 + return false; 1.922 + 1.923 + case Op_CmpI: 1.924 + // A small minority (< 10%) of CmpP are masked as CmpI, 1.925 + // as if by boolean conversion ((p == q? 1: 0) != 0). 1.926 + // Detect that here, even if it hasn't optimized away yet. 1.927 + // Specifically, this covers the 'instanceof' operator. 1.928 + if (btest == BoolTest::ne || btest == BoolTest::eq) { 1.929 + if (_gvn.type(cmp->in(2))->singleton() && 1.930 + cmp->in(1)->is_Phi()) { 1.931 + PhiNode* phi = cmp->in(1)->as_Phi(); 1.932 + int true_path = phi->is_diamond_phi(); 1.933 + if (true_path > 0 && 1.934 + _gvn.type(phi->in(1))->singleton() && 1.935 + _gvn.type(phi->in(2))->singleton()) { 1.936 + // phi->region->if_proj->ifnode->bool->cmp 1.937 + BoolNode* bol = phi->in(0)->in(1)->in(0)->in(1)->as_Bool(); 1.938 + btest = bol->_test._test; 1.939 + cmp = bol->in(1); 1.940 + continue; 1.941 + } 1.942 + } 1.943 + } 1.944 + return false; 1.945 + } 1.946 + } 1.947 + return false; 1.948 +} 1.949 + 1.950 +//-------------------------------repush_if_args-------------------------------- 1.951 +// Push arguments of an "if" bytecode back onto the stack by adjusting _sp. 1.952 +inline int Parse::repush_if_args() { 1.953 +#ifndef PRODUCT 1.954 + if (PrintOpto && WizardMode) { 1.955 + tty->print("defending against excessive implicit null exceptions on %s @%d in ", 1.956 + Bytecodes::name(iter().cur_bc()), iter().cur_bci()); 1.957 + method()->print_name(); tty->cr(); 1.958 + } 1.959 +#endif 1.960 + int bc_depth = - Bytecodes::depth(iter().cur_bc()); 1.961 + assert(bc_depth == 1 || bc_depth == 2, "only two kinds of branches"); 1.962 + DEBUG_ONLY(sync_jvms()); // argument(n) requires a synced jvms 1.963 + assert(argument(0) != NULL, "must exist"); 1.964 + assert(bc_depth == 1 || argument(1) != NULL, "two must exist"); 1.965 + inc_sp(bc_depth); 1.966 + return bc_depth; 1.967 +} 1.968 + 1.969 +//----------------------------------do_ifnull---------------------------------- 1.970 +void Parse::do_ifnull(BoolTest::mask btest, Node *c) { 1.971 + int target_bci = iter().get_dest(); 1.972 + 1.973 + Block* branch_block = successor_for_bci(target_bci); 1.974 + Block* next_block = successor_for_bci(iter().next_bci()); 1.975 + 1.976 + float cnt; 1.977 + float prob = branch_prediction(cnt, btest, target_bci); 1.978 + if (prob == PROB_UNKNOWN) { 1.979 + // (An earlier version of do_ifnull omitted this trap for OSR methods.) 1.980 +#ifndef PRODUCT 1.981 + if (PrintOpto && Verbose) 1.982 + tty->print_cr("Never-taken edge stops compilation at bci %d",bci()); 1.983 +#endif 1.984 + repush_if_args(); // to gather stats on loop 1.985 + // We need to mark this branch as taken so that if we recompile we will 1.986 + // see that it is possible. In the tiered system the interpreter doesn't 1.987 + // do profiling and by the time we get to the lower tier from the interpreter 1.988 + // the path may be cold again. Make sure it doesn't look untaken 1.989 + profile_taken_branch(target_bci, !ProfileInterpreter); 1.990 + uncommon_trap(Deoptimization::Reason_unreached, 1.991 + Deoptimization::Action_reinterpret, 1.992 + NULL, "cold"); 1.993 + if (C->eliminate_boxing()) { 1.994 + // Mark the successor blocks as parsed 1.995 + branch_block->next_path_num(); 1.996 + next_block->next_path_num(); 1.997 + } 1.998 + return; 1.999 + } 1.1000 + 1.1001 + explicit_null_checks_inserted++; 1.1002 + 1.1003 + // Generate real control flow 1.1004 + Node *tst = _gvn.transform( new (C) BoolNode( c, btest ) ); 1.1005 + 1.1006 + // Sanity check the probability value 1.1007 + assert(prob > 0.0f,"Bad probability in Parser"); 1.1008 + // Need xform to put node in hash table 1.1009 + IfNode *iff = create_and_xform_if( control(), tst, prob, cnt ); 1.1010 + assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser"); 1.1011 + // True branch 1.1012 + { PreserveJVMState pjvms(this); 1.1013 + Node* iftrue = _gvn.transform( new (C) IfTrueNode (iff) ); 1.1014 + set_control(iftrue); 1.1015 + 1.1016 + if (stopped()) { // Path is dead? 1.1017 + explicit_null_checks_elided++; 1.1018 + if (C->eliminate_boxing()) { 1.1019 + // Mark the successor block as parsed 1.1020 + branch_block->next_path_num(); 1.1021 + } 1.1022 + } else { // Path is live. 1.1023 + // Update method data 1.1024 + profile_taken_branch(target_bci); 1.1025 + adjust_map_after_if(btest, c, prob, branch_block, next_block); 1.1026 + if (!stopped()) { 1.1027 + merge(target_bci); 1.1028 + } 1.1029 + } 1.1030 + } 1.1031 + 1.1032 + // False branch 1.1033 + Node* iffalse = _gvn.transform( new (C) IfFalseNode(iff) ); 1.1034 + set_control(iffalse); 1.1035 + 1.1036 + if (stopped()) { // Path is dead? 1.1037 + explicit_null_checks_elided++; 1.1038 + if (C->eliminate_boxing()) { 1.1039 + // Mark the successor block as parsed 1.1040 + next_block->next_path_num(); 1.1041 + } 1.1042 + } else { // Path is live. 1.1043 + // Update method data 1.1044 + profile_not_taken_branch(); 1.1045 + adjust_map_after_if(BoolTest(btest).negate(), c, 1.0-prob, 1.1046 + next_block, branch_block); 1.1047 + } 1.1048 +} 1.1049 + 1.1050 +//------------------------------------do_if------------------------------------ 1.1051 +void Parse::do_if(BoolTest::mask btest, Node* c) { 1.1052 + int target_bci = iter().get_dest(); 1.1053 + 1.1054 + Block* branch_block = successor_for_bci(target_bci); 1.1055 + Block* next_block = successor_for_bci(iter().next_bci()); 1.1056 + 1.1057 + float cnt; 1.1058 + float prob = branch_prediction(cnt, btest, target_bci); 1.1059 + float untaken_prob = 1.0 - prob; 1.1060 + 1.1061 + if (prob == PROB_UNKNOWN) { 1.1062 +#ifndef PRODUCT 1.1063 + if (PrintOpto && Verbose) 1.1064 + tty->print_cr("Never-taken edge stops compilation at bci %d",bci()); 1.1065 +#endif 1.1066 + repush_if_args(); // to gather stats on loop 1.1067 + // We need to mark this branch as taken so that if we recompile we will 1.1068 + // see that it is possible. In the tiered system the interpreter doesn't 1.1069 + // do profiling and by the time we get to the lower tier from the interpreter 1.1070 + // the path may be cold again. Make sure it doesn't look untaken 1.1071 + profile_taken_branch(target_bci, !ProfileInterpreter); 1.1072 + uncommon_trap(Deoptimization::Reason_unreached, 1.1073 + Deoptimization::Action_reinterpret, 1.1074 + NULL, "cold"); 1.1075 + if (C->eliminate_boxing()) { 1.1076 + // Mark the successor blocks as parsed 1.1077 + branch_block->next_path_num(); 1.1078 + next_block->next_path_num(); 1.1079 + } 1.1080 + return; 1.1081 + } 1.1082 + 1.1083 + // Sanity check the probability value 1.1084 + assert(0.0f < prob && prob < 1.0f,"Bad probability in Parser"); 1.1085 + 1.1086 + bool taken_if_true = true; 1.1087 + // Convert BoolTest to canonical form: 1.1088 + if (!BoolTest(btest).is_canonical()) { 1.1089 + btest = BoolTest(btest).negate(); 1.1090 + taken_if_true = false; 1.1091 + // prob is NOT updated here; it remains the probability of the taken 1.1092 + // path (as opposed to the prob of the path guarded by an 'IfTrueNode'). 1.1093 + } 1.1094 + assert(btest != BoolTest::eq, "!= is the only canonical exact test"); 1.1095 + 1.1096 + Node* tst0 = new (C) BoolNode(c, btest); 1.1097 + Node* tst = _gvn.transform(tst0); 1.1098 + BoolTest::mask taken_btest = BoolTest::illegal; 1.1099 + BoolTest::mask untaken_btest = BoolTest::illegal; 1.1100 + 1.1101 + if (tst->is_Bool()) { 1.1102 + // Refresh c from the transformed bool node, since it may be 1.1103 + // simpler than the original c. Also re-canonicalize btest. 1.1104 + // This wins when (Bool ne (Conv2B p) 0) => (Bool ne (CmpP p NULL)). 1.1105 + // That can arise from statements like: if (x instanceof C) ... 1.1106 + if (tst != tst0) { 1.1107 + // Canonicalize one more time since transform can change it. 1.1108 + btest = tst->as_Bool()->_test._test; 1.1109 + if (!BoolTest(btest).is_canonical()) { 1.1110 + // Reverse edges one more time... 1.1111 + tst = _gvn.transform( tst->as_Bool()->negate(&_gvn) ); 1.1112 + btest = tst->as_Bool()->_test._test; 1.1113 + assert(BoolTest(btest).is_canonical(), "sanity"); 1.1114 + taken_if_true = !taken_if_true; 1.1115 + } 1.1116 + c = tst->in(1); 1.1117 + } 1.1118 + BoolTest::mask neg_btest = BoolTest(btest).negate(); 1.1119 + taken_btest = taken_if_true ? btest : neg_btest; 1.1120 + untaken_btest = taken_if_true ? neg_btest : btest; 1.1121 + } 1.1122 + 1.1123 + // Generate real control flow 1.1124 + float true_prob = (taken_if_true ? prob : untaken_prob); 1.1125 + IfNode* iff = create_and_map_if(control(), tst, true_prob, cnt); 1.1126 + assert(iff->_prob > 0.0f,"Optimizer made bad probability in parser"); 1.1127 + Node* taken_branch = new (C) IfTrueNode(iff); 1.1128 + Node* untaken_branch = new (C) IfFalseNode(iff); 1.1129 + if (!taken_if_true) { // Finish conversion to canonical form 1.1130 + Node* tmp = taken_branch; 1.1131 + taken_branch = untaken_branch; 1.1132 + untaken_branch = tmp; 1.1133 + } 1.1134 + 1.1135 + // Branch is taken: 1.1136 + { PreserveJVMState pjvms(this); 1.1137 + taken_branch = _gvn.transform(taken_branch); 1.1138 + set_control(taken_branch); 1.1139 + 1.1140 + if (stopped()) { 1.1141 + if (C->eliminate_boxing()) { 1.1142 + // Mark the successor block as parsed 1.1143 + branch_block->next_path_num(); 1.1144 + } 1.1145 + } else { 1.1146 + // Update method data 1.1147 + profile_taken_branch(target_bci); 1.1148 + adjust_map_after_if(taken_btest, c, prob, branch_block, next_block); 1.1149 + if (!stopped()) { 1.1150 + merge(target_bci); 1.1151 + } 1.1152 + } 1.1153 + } 1.1154 + 1.1155 + untaken_branch = _gvn.transform(untaken_branch); 1.1156 + set_control(untaken_branch); 1.1157 + 1.1158 + // Branch not taken. 1.1159 + if (stopped()) { 1.1160 + if (C->eliminate_boxing()) { 1.1161 + // Mark the successor block as parsed 1.1162 + next_block->next_path_num(); 1.1163 + } 1.1164 + } else { 1.1165 + // Update method data 1.1166 + profile_not_taken_branch(); 1.1167 + adjust_map_after_if(untaken_btest, c, untaken_prob, 1.1168 + next_block, branch_block); 1.1169 + } 1.1170 +} 1.1171 + 1.1172 +//----------------------------adjust_map_after_if------------------------------ 1.1173 +// Adjust the JVM state to reflect the result of taking this path. 1.1174 +// Basically, it means inspecting the CmpNode controlling this 1.1175 +// branch, seeing how it constrains a tested value, and then 1.1176 +// deciding if it's worth our while to encode this constraint 1.1177 +// as graph nodes in the current abstract interpretation map. 1.1178 +void Parse::adjust_map_after_if(BoolTest::mask btest, Node* c, float prob, 1.1179 + Block* path, Block* other_path) { 1.1180 + if (stopped() || !c->is_Cmp() || btest == BoolTest::illegal) 1.1181 + return; // nothing to do 1.1182 + 1.1183 + bool is_fallthrough = (path == successor_for_bci(iter().next_bci())); 1.1184 + 1.1185 + if (seems_never_taken(prob) && seems_stable_comparison(btest, c)) { 1.1186 + // If this might possibly turn into an implicit null check, 1.1187 + // and the null has never yet been seen, we need to generate 1.1188 + // an uncommon trap, so as to recompile instead of suffering 1.1189 + // with very slow branches. (We'll get the slow branches if 1.1190 + // the program ever changes phase and starts seeing nulls here.) 1.1191 + // 1.1192 + // We do not inspect for a null constant, since a node may 1.1193 + // optimize to 'null' later on. 1.1194 + // 1.1195 + // Null checks, and other tests which expect inequality, 1.1196 + // show btest == BoolTest::eq along the non-taken branch. 1.1197 + // On the other hand, type tests, must-be-null tests, 1.1198 + // and other tests which expect pointer equality, 1.1199 + // show btest == BoolTest::ne along the non-taken branch. 1.1200 + // We prune both types of branches if they look unused. 1.1201 + repush_if_args(); 1.1202 + // We need to mark this branch as taken so that if we recompile we will 1.1203 + // see that it is possible. In the tiered system the interpreter doesn't 1.1204 + // do profiling and by the time we get to the lower tier from the interpreter 1.1205 + // the path may be cold again. Make sure it doesn't look untaken 1.1206 + if (is_fallthrough) { 1.1207 + profile_not_taken_branch(!ProfileInterpreter); 1.1208 + } else { 1.1209 + profile_taken_branch(iter().get_dest(), !ProfileInterpreter); 1.1210 + } 1.1211 + uncommon_trap(Deoptimization::Reason_unreached, 1.1212 + Deoptimization::Action_reinterpret, 1.1213 + NULL, 1.1214 + (is_fallthrough ? "taken always" : "taken never")); 1.1215 + return; 1.1216 + } 1.1217 + 1.1218 + Node* val = c->in(1); 1.1219 + Node* con = c->in(2); 1.1220 + const Type* tcon = _gvn.type(con); 1.1221 + const Type* tval = _gvn.type(val); 1.1222 + bool have_con = tcon->singleton(); 1.1223 + if (tval->singleton()) { 1.1224 + if (!have_con) { 1.1225 + // Swap, so constant is in con. 1.1226 + con = val; 1.1227 + tcon = tval; 1.1228 + val = c->in(2); 1.1229 + tval = _gvn.type(val); 1.1230 + btest = BoolTest(btest).commute(); 1.1231 + have_con = true; 1.1232 + } else { 1.1233 + // Do we have two constants? Then leave well enough alone. 1.1234 + have_con = false; 1.1235 + } 1.1236 + } 1.1237 + if (!have_con) // remaining adjustments need a con 1.1238 + return; 1.1239 + 1.1240 + sharpen_type_after_if(btest, con, tcon, val, tval); 1.1241 +} 1.1242 + 1.1243 + 1.1244 +static Node* extract_obj_from_klass_load(PhaseGVN* gvn, Node* n) { 1.1245 + Node* ldk; 1.1246 + if (n->is_DecodeNKlass()) { 1.1247 + if (n->in(1)->Opcode() != Op_LoadNKlass) { 1.1248 + return NULL; 1.1249 + } else { 1.1250 + ldk = n->in(1); 1.1251 + } 1.1252 + } else if (n->Opcode() != Op_LoadKlass) { 1.1253 + return NULL; 1.1254 + } else { 1.1255 + ldk = n; 1.1256 + } 1.1257 + assert(ldk != NULL && ldk->is_Load(), "should have found a LoadKlass or LoadNKlass node"); 1.1258 + 1.1259 + Node* adr = ldk->in(MemNode::Address); 1.1260 + intptr_t off = 0; 1.1261 + Node* obj = AddPNode::Ideal_base_and_offset(adr, gvn, off); 1.1262 + if (obj == NULL || off != oopDesc::klass_offset_in_bytes()) // loading oopDesc::_klass? 1.1263 + return NULL; 1.1264 + const TypePtr* tp = gvn->type(obj)->is_ptr(); 1.1265 + if (tp == NULL || !(tp->isa_instptr() || tp->isa_aryptr())) // is obj a Java object ptr? 1.1266 + return NULL; 1.1267 + 1.1268 + return obj; 1.1269 +} 1.1270 + 1.1271 +void Parse::sharpen_type_after_if(BoolTest::mask btest, 1.1272 + Node* con, const Type* tcon, 1.1273 + Node* val, const Type* tval) { 1.1274 + // Look for opportunities to sharpen the type of a node 1.1275 + // whose klass is compared with a constant klass. 1.1276 + if (btest == BoolTest::eq && tcon->isa_klassptr()) { 1.1277 + Node* obj = extract_obj_from_klass_load(&_gvn, val); 1.1278 + const TypeOopPtr* con_type = tcon->isa_klassptr()->as_instance_type(); 1.1279 + if (obj != NULL && (con_type->isa_instptr() || con_type->isa_aryptr())) { 1.1280 + // Found: 1.1281 + // Bool(CmpP(LoadKlass(obj._klass), ConP(Foo.klass)), [eq]) 1.1282 + // or the narrowOop equivalent. 1.1283 + const Type* obj_type = _gvn.type(obj); 1.1284 + const TypeOopPtr* tboth = obj_type->join_speculative(con_type)->isa_oopptr(); 1.1285 + if (tboth != NULL && tboth->klass_is_exact() && tboth != obj_type && 1.1286 + tboth->higher_equal(obj_type)) { 1.1287 + // obj has to be of the exact type Foo if the CmpP succeeds. 1.1288 + int obj_in_map = map()->find_edge(obj); 1.1289 + JVMState* jvms = this->jvms(); 1.1290 + if (obj_in_map >= 0 && 1.1291 + (jvms->is_loc(obj_in_map) || jvms->is_stk(obj_in_map))) { 1.1292 + TypeNode* ccast = new (C) CheckCastPPNode(control(), obj, tboth); 1.1293 + const Type* tcc = ccast->as_Type()->type(); 1.1294 + assert(tcc != obj_type && tcc->higher_equal_speculative(obj_type), "must improve"); 1.1295 + // Delay transform() call to allow recovery of pre-cast value 1.1296 + // at the control merge. 1.1297 + _gvn.set_type_bottom(ccast); 1.1298 + record_for_igvn(ccast); 1.1299 + // Here's the payoff. 1.1300 + replace_in_map(obj, ccast); 1.1301 + } 1.1302 + } 1.1303 + } 1.1304 + } 1.1305 + 1.1306 + int val_in_map = map()->find_edge(val); 1.1307 + if (val_in_map < 0) return; // replace_in_map would be useless 1.1308 + { 1.1309 + JVMState* jvms = this->jvms(); 1.1310 + if (!(jvms->is_loc(val_in_map) || 1.1311 + jvms->is_stk(val_in_map))) 1.1312 + return; // again, it would be useless 1.1313 + } 1.1314 + 1.1315 + // Check for a comparison to a constant, and "know" that the compared 1.1316 + // value is constrained on this path. 1.1317 + assert(tcon->singleton(), ""); 1.1318 + ConstraintCastNode* ccast = NULL; 1.1319 + Node* cast = NULL; 1.1320 + 1.1321 + switch (btest) { 1.1322 + case BoolTest::eq: // Constant test? 1.1323 + { 1.1324 + const Type* tboth = tcon->join_speculative(tval); 1.1325 + if (tboth == tval) break; // Nothing to gain. 1.1326 + if (tcon->isa_int()) { 1.1327 + ccast = new (C) CastIINode(val, tboth); 1.1328 + } else if (tcon == TypePtr::NULL_PTR) { 1.1329 + // Cast to null, but keep the pointer identity temporarily live. 1.1330 + ccast = new (C) CastPPNode(val, tboth); 1.1331 + } else { 1.1332 + const TypeF* tf = tcon->isa_float_constant(); 1.1333 + const TypeD* td = tcon->isa_double_constant(); 1.1334 + // Exclude tests vs float/double 0 as these could be 1.1335 + // either +0 or -0. Just because you are equal to +0 1.1336 + // doesn't mean you ARE +0! 1.1337 + // Note, following code also replaces Long and Oop values. 1.1338 + if ((!tf || tf->_f != 0.0) && 1.1339 + (!td || td->_d != 0.0)) 1.1340 + cast = con; // Replace non-constant val by con. 1.1341 + } 1.1342 + } 1.1343 + break; 1.1344 + 1.1345 + case BoolTest::ne: 1.1346 + if (tcon == TypePtr::NULL_PTR) { 1.1347 + cast = cast_not_null(val, false); 1.1348 + } 1.1349 + break; 1.1350 + 1.1351 + default: 1.1352 + // (At this point we could record int range types with CastII.) 1.1353 + break; 1.1354 + } 1.1355 + 1.1356 + if (ccast != NULL) { 1.1357 + const Type* tcc = ccast->as_Type()->type(); 1.1358 + assert(tcc != tval && tcc->higher_equal_speculative(tval), "must improve"); 1.1359 + // Delay transform() call to allow recovery of pre-cast value 1.1360 + // at the control merge. 1.1361 + ccast->set_req(0, control()); 1.1362 + _gvn.set_type_bottom(ccast); 1.1363 + record_for_igvn(ccast); 1.1364 + cast = ccast; 1.1365 + } 1.1366 + 1.1367 + if (cast != NULL) { // Here's the payoff. 1.1368 + replace_in_map(val, cast); 1.1369 + } 1.1370 +} 1.1371 + 1.1372 +/** 1.1373 + * Use speculative type to optimize CmpP node: if comparison is 1.1374 + * against the low level class, cast the object to the speculative 1.1375 + * type if any. CmpP should then go away. 1.1376 + * 1.1377 + * @param c expected CmpP node 1.1378 + * @return result of CmpP on object casted to speculative type 1.1379 + * 1.1380 + */ 1.1381 +Node* Parse::optimize_cmp_with_klass(Node* c) { 1.1382 + // If this is transformed by the _gvn to a comparison with the low 1.1383 + // level klass then we may be able to use speculation 1.1384 + if (c->Opcode() == Op_CmpP && 1.1385 + (c->in(1)->Opcode() == Op_LoadKlass || c->in(1)->Opcode() == Op_DecodeNKlass) && 1.1386 + c->in(2)->is_Con()) { 1.1387 + Node* load_klass = NULL; 1.1388 + Node* decode = NULL; 1.1389 + if (c->in(1)->Opcode() == Op_DecodeNKlass) { 1.1390 + decode = c->in(1); 1.1391 + load_klass = c->in(1)->in(1); 1.1392 + } else { 1.1393 + load_klass = c->in(1); 1.1394 + } 1.1395 + if (load_klass->in(2)->is_AddP()) { 1.1396 + Node* addp = load_klass->in(2); 1.1397 + Node* obj = addp->in(AddPNode::Address); 1.1398 + const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr(); 1.1399 + if (obj_type->speculative_type() != NULL) { 1.1400 + ciKlass* k = obj_type->speculative_type(); 1.1401 + inc_sp(2); 1.1402 + obj = maybe_cast_profiled_obj(obj, k); 1.1403 + dec_sp(2); 1.1404 + // Make the CmpP use the casted obj 1.1405 + addp = basic_plus_adr(obj, addp->in(AddPNode::Offset)); 1.1406 + load_klass = load_klass->clone(); 1.1407 + load_klass->set_req(2, addp); 1.1408 + load_klass = _gvn.transform(load_klass); 1.1409 + if (decode != NULL) { 1.1410 + decode = decode->clone(); 1.1411 + decode->set_req(1, load_klass); 1.1412 + load_klass = _gvn.transform(decode); 1.1413 + } 1.1414 + c = c->clone(); 1.1415 + c->set_req(1, load_klass); 1.1416 + c = _gvn.transform(c); 1.1417 + } 1.1418 + } 1.1419 + } 1.1420 + return c; 1.1421 +} 1.1422 + 1.1423 +//------------------------------do_one_bytecode-------------------------------- 1.1424 +// Parse this bytecode, and alter the Parsers JVM->Node mapping 1.1425 +void Parse::do_one_bytecode() { 1.1426 + Node *a, *b, *c, *d; // Handy temps 1.1427 + BoolTest::mask btest; 1.1428 + int i; 1.1429 + 1.1430 + assert(!has_exceptions(), "bytecode entry state must be clear of throws"); 1.1431 + 1.1432 + if (C->check_node_count(NodeLimitFudgeFactor * 5, 1.1433 + "out of nodes parsing method")) { 1.1434 + return; 1.1435 + } 1.1436 + 1.1437 +#ifdef ASSERT 1.1438 + // for setting breakpoints 1.1439 + if (TraceOptoParse) { 1.1440 + tty->print(" @"); 1.1441 + dump_bci(bci()); 1.1442 + tty->cr(); 1.1443 + } 1.1444 +#endif 1.1445 + 1.1446 + switch (bc()) { 1.1447 + case Bytecodes::_nop: 1.1448 + // do nothing 1.1449 + break; 1.1450 + case Bytecodes::_lconst_0: 1.1451 + push_pair(longcon(0)); 1.1452 + break; 1.1453 + 1.1454 + case Bytecodes::_lconst_1: 1.1455 + push_pair(longcon(1)); 1.1456 + break; 1.1457 + 1.1458 + case Bytecodes::_fconst_0: 1.1459 + push(zerocon(T_FLOAT)); 1.1460 + break; 1.1461 + 1.1462 + case Bytecodes::_fconst_1: 1.1463 + push(makecon(TypeF::ONE)); 1.1464 + break; 1.1465 + 1.1466 + case Bytecodes::_fconst_2: 1.1467 + push(makecon(TypeF::make(2.0f))); 1.1468 + break; 1.1469 + 1.1470 + case Bytecodes::_dconst_0: 1.1471 + push_pair(zerocon(T_DOUBLE)); 1.1472 + break; 1.1473 + 1.1474 + case Bytecodes::_dconst_1: 1.1475 + push_pair(makecon(TypeD::ONE)); 1.1476 + break; 1.1477 + 1.1478 + case Bytecodes::_iconst_m1:push(intcon(-1)); break; 1.1479 + case Bytecodes::_iconst_0: push(intcon( 0)); break; 1.1480 + case Bytecodes::_iconst_1: push(intcon( 1)); break; 1.1481 + case Bytecodes::_iconst_2: push(intcon( 2)); break; 1.1482 + case Bytecodes::_iconst_3: push(intcon( 3)); break; 1.1483 + case Bytecodes::_iconst_4: push(intcon( 4)); break; 1.1484 + case Bytecodes::_iconst_5: push(intcon( 5)); break; 1.1485 + case Bytecodes::_bipush: push(intcon(iter().get_constant_u1())); break; 1.1486 + case Bytecodes::_sipush: push(intcon(iter().get_constant_u2())); break; 1.1487 + case Bytecodes::_aconst_null: push(null()); break; 1.1488 + case Bytecodes::_ldc: 1.1489 + case Bytecodes::_ldc_w: 1.1490 + case Bytecodes::_ldc2_w: 1.1491 + // If the constant is unresolved, run this BC once in the interpreter. 1.1492 + { 1.1493 + ciConstant constant = iter().get_constant(); 1.1494 + if (constant.basic_type() == T_OBJECT && 1.1495 + !constant.as_object()->is_loaded()) { 1.1496 + int index = iter().get_constant_pool_index(); 1.1497 + constantTag tag = iter().get_constant_pool_tag(index); 1.1498 + uncommon_trap(Deoptimization::make_trap_request 1.1499 + (Deoptimization::Reason_unloaded, 1.1500 + Deoptimization::Action_reinterpret, 1.1501 + index), 1.1502 + NULL, tag.internal_name()); 1.1503 + break; 1.1504 + } 1.1505 + assert(constant.basic_type() != T_OBJECT || constant.as_object()->is_instance(), 1.1506 + "must be java_mirror of klass"); 1.1507 + bool pushed = push_constant(constant, true); 1.1508 + guarantee(pushed, "must be possible to push this constant"); 1.1509 + } 1.1510 + 1.1511 + break; 1.1512 + 1.1513 + case Bytecodes::_aload_0: 1.1514 + push( local(0) ); 1.1515 + break; 1.1516 + case Bytecodes::_aload_1: 1.1517 + push( local(1) ); 1.1518 + break; 1.1519 + case Bytecodes::_aload_2: 1.1520 + push( local(2) ); 1.1521 + break; 1.1522 + case Bytecodes::_aload_3: 1.1523 + push( local(3) ); 1.1524 + break; 1.1525 + case Bytecodes::_aload: 1.1526 + push( local(iter().get_index()) ); 1.1527 + break; 1.1528 + 1.1529 + case Bytecodes::_fload_0: 1.1530 + case Bytecodes::_iload_0: 1.1531 + push( local(0) ); 1.1532 + break; 1.1533 + case Bytecodes::_fload_1: 1.1534 + case Bytecodes::_iload_1: 1.1535 + push( local(1) ); 1.1536 + break; 1.1537 + case Bytecodes::_fload_2: 1.1538 + case Bytecodes::_iload_2: 1.1539 + push( local(2) ); 1.1540 + break; 1.1541 + case Bytecodes::_fload_3: 1.1542 + case Bytecodes::_iload_3: 1.1543 + push( local(3) ); 1.1544 + break; 1.1545 + case Bytecodes::_fload: 1.1546 + case Bytecodes::_iload: 1.1547 + push( local(iter().get_index()) ); 1.1548 + break; 1.1549 + case Bytecodes::_lload_0: 1.1550 + push_pair_local( 0 ); 1.1551 + break; 1.1552 + case Bytecodes::_lload_1: 1.1553 + push_pair_local( 1 ); 1.1554 + break; 1.1555 + case Bytecodes::_lload_2: 1.1556 + push_pair_local( 2 ); 1.1557 + break; 1.1558 + case Bytecodes::_lload_3: 1.1559 + push_pair_local( 3 ); 1.1560 + break; 1.1561 + case Bytecodes::_lload: 1.1562 + push_pair_local( iter().get_index() ); 1.1563 + break; 1.1564 + 1.1565 + case Bytecodes::_dload_0: 1.1566 + push_pair_local(0); 1.1567 + break; 1.1568 + case Bytecodes::_dload_1: 1.1569 + push_pair_local(1); 1.1570 + break; 1.1571 + case Bytecodes::_dload_2: 1.1572 + push_pair_local(2); 1.1573 + break; 1.1574 + case Bytecodes::_dload_3: 1.1575 + push_pair_local(3); 1.1576 + break; 1.1577 + case Bytecodes::_dload: 1.1578 + push_pair_local(iter().get_index()); 1.1579 + break; 1.1580 + case Bytecodes::_fstore_0: 1.1581 + case Bytecodes::_istore_0: 1.1582 + case Bytecodes::_astore_0: 1.1583 + set_local( 0, pop() ); 1.1584 + break; 1.1585 + case Bytecodes::_fstore_1: 1.1586 + case Bytecodes::_istore_1: 1.1587 + case Bytecodes::_astore_1: 1.1588 + set_local( 1, pop() ); 1.1589 + break; 1.1590 + case Bytecodes::_fstore_2: 1.1591 + case Bytecodes::_istore_2: 1.1592 + case Bytecodes::_astore_2: 1.1593 + set_local( 2, pop() ); 1.1594 + break; 1.1595 + case Bytecodes::_fstore_3: 1.1596 + case Bytecodes::_istore_3: 1.1597 + case Bytecodes::_astore_3: 1.1598 + set_local( 3, pop() ); 1.1599 + break; 1.1600 + case Bytecodes::_fstore: 1.1601 + case Bytecodes::_istore: 1.1602 + case Bytecodes::_astore: 1.1603 + set_local( iter().get_index(), pop() ); 1.1604 + break; 1.1605 + // long stores 1.1606 + case Bytecodes::_lstore_0: 1.1607 + set_pair_local( 0, pop_pair() ); 1.1608 + break; 1.1609 + case Bytecodes::_lstore_1: 1.1610 + set_pair_local( 1, pop_pair() ); 1.1611 + break; 1.1612 + case Bytecodes::_lstore_2: 1.1613 + set_pair_local( 2, pop_pair() ); 1.1614 + break; 1.1615 + case Bytecodes::_lstore_3: 1.1616 + set_pair_local( 3, pop_pair() ); 1.1617 + break; 1.1618 + case Bytecodes::_lstore: 1.1619 + set_pair_local( iter().get_index(), pop_pair() ); 1.1620 + break; 1.1621 + 1.1622 + // double stores 1.1623 + case Bytecodes::_dstore_0: 1.1624 + set_pair_local( 0, dstore_rounding(pop_pair()) ); 1.1625 + break; 1.1626 + case Bytecodes::_dstore_1: 1.1627 + set_pair_local( 1, dstore_rounding(pop_pair()) ); 1.1628 + break; 1.1629 + case Bytecodes::_dstore_2: 1.1630 + set_pair_local( 2, dstore_rounding(pop_pair()) ); 1.1631 + break; 1.1632 + case Bytecodes::_dstore_3: 1.1633 + set_pair_local( 3, dstore_rounding(pop_pair()) ); 1.1634 + break; 1.1635 + case Bytecodes::_dstore: 1.1636 + set_pair_local( iter().get_index(), dstore_rounding(pop_pair()) ); 1.1637 + break; 1.1638 + 1.1639 + case Bytecodes::_pop: dec_sp(1); break; 1.1640 + case Bytecodes::_pop2: dec_sp(2); break; 1.1641 + case Bytecodes::_swap: 1.1642 + a = pop(); 1.1643 + b = pop(); 1.1644 + push(a); 1.1645 + push(b); 1.1646 + break; 1.1647 + case Bytecodes::_dup: 1.1648 + a = pop(); 1.1649 + push(a); 1.1650 + push(a); 1.1651 + break; 1.1652 + case Bytecodes::_dup_x1: 1.1653 + a = pop(); 1.1654 + b = pop(); 1.1655 + push( a ); 1.1656 + push( b ); 1.1657 + push( a ); 1.1658 + break; 1.1659 + case Bytecodes::_dup_x2: 1.1660 + a = pop(); 1.1661 + b = pop(); 1.1662 + c = pop(); 1.1663 + push( a ); 1.1664 + push( c ); 1.1665 + push( b ); 1.1666 + push( a ); 1.1667 + break; 1.1668 + case Bytecodes::_dup2: 1.1669 + a = pop(); 1.1670 + b = pop(); 1.1671 + push( b ); 1.1672 + push( a ); 1.1673 + push( b ); 1.1674 + push( a ); 1.1675 + break; 1.1676 + 1.1677 + case Bytecodes::_dup2_x1: 1.1678 + // before: .. c, b, a 1.1679 + // after: .. b, a, c, b, a 1.1680 + // not tested 1.1681 + a = pop(); 1.1682 + b = pop(); 1.1683 + c = pop(); 1.1684 + push( b ); 1.1685 + push( a ); 1.1686 + push( c ); 1.1687 + push( b ); 1.1688 + push( a ); 1.1689 + break; 1.1690 + case Bytecodes::_dup2_x2: 1.1691 + // before: .. d, c, b, a 1.1692 + // after: .. b, a, d, c, b, a 1.1693 + // not tested 1.1694 + a = pop(); 1.1695 + b = pop(); 1.1696 + c = pop(); 1.1697 + d = pop(); 1.1698 + push( b ); 1.1699 + push( a ); 1.1700 + push( d ); 1.1701 + push( c ); 1.1702 + push( b ); 1.1703 + push( a ); 1.1704 + break; 1.1705 + 1.1706 + case Bytecodes::_arraylength: { 1.1707 + // Must do null-check with value on expression stack 1.1708 + Node *ary = null_check(peek(), T_ARRAY); 1.1709 + // Compile-time detect of null-exception? 1.1710 + if (stopped()) return; 1.1711 + a = pop(); 1.1712 + push(load_array_length(a)); 1.1713 + break; 1.1714 + } 1.1715 + 1.1716 + case Bytecodes::_baload: array_load(T_BYTE); break; 1.1717 + case Bytecodes::_caload: array_load(T_CHAR); break; 1.1718 + case Bytecodes::_iaload: array_load(T_INT); break; 1.1719 + case Bytecodes::_saload: array_load(T_SHORT); break; 1.1720 + case Bytecodes::_faload: array_load(T_FLOAT); break; 1.1721 + case Bytecodes::_aaload: array_load(T_OBJECT); break; 1.1722 + case Bytecodes::_laload: { 1.1723 + a = array_addressing(T_LONG, 0); 1.1724 + if (stopped()) return; // guaranteed null or range check 1.1725 + dec_sp(2); // Pop array and index 1.1726 + push_pair(make_load(control(), a, TypeLong::LONG, T_LONG, TypeAryPtr::LONGS, MemNode::unordered)); 1.1727 + break; 1.1728 + } 1.1729 + case Bytecodes::_daload: { 1.1730 + a = array_addressing(T_DOUBLE, 0); 1.1731 + if (stopped()) return; // guaranteed null or range check 1.1732 + dec_sp(2); // Pop array and index 1.1733 + push_pair(make_load(control(), a, Type::DOUBLE, T_DOUBLE, TypeAryPtr::DOUBLES, MemNode::unordered)); 1.1734 + break; 1.1735 + } 1.1736 + case Bytecodes::_bastore: array_store(T_BYTE); break; 1.1737 + case Bytecodes::_castore: array_store(T_CHAR); break; 1.1738 + case Bytecodes::_iastore: array_store(T_INT); break; 1.1739 + case Bytecodes::_sastore: array_store(T_SHORT); break; 1.1740 + case Bytecodes::_fastore: array_store(T_FLOAT); break; 1.1741 + case Bytecodes::_aastore: { 1.1742 + d = array_addressing(T_OBJECT, 1); 1.1743 + if (stopped()) return; // guaranteed null or range check 1.1744 + array_store_check(); 1.1745 + c = pop(); // Oop to store 1.1746 + b = pop(); // index (already used) 1.1747 + a = pop(); // the array itself 1.1748 + const TypeOopPtr* elemtype = _gvn.type(a)->is_aryptr()->elem()->make_oopptr(); 1.1749 + const TypeAryPtr* adr_type = TypeAryPtr::OOPS; 1.1750 + Node* store = store_oop_to_array(control(), a, d, adr_type, c, elemtype, T_OBJECT, MemNode::release); 1.1751 + break; 1.1752 + } 1.1753 + case Bytecodes::_lastore: { 1.1754 + a = array_addressing(T_LONG, 2); 1.1755 + if (stopped()) return; // guaranteed null or range check 1.1756 + c = pop_pair(); 1.1757 + dec_sp(2); // Pop array and index 1.1758 + store_to_memory(control(), a, c, T_LONG, TypeAryPtr::LONGS, MemNode::unordered); 1.1759 + break; 1.1760 + } 1.1761 + case Bytecodes::_dastore: { 1.1762 + a = array_addressing(T_DOUBLE, 2); 1.1763 + if (stopped()) return; // guaranteed null or range check 1.1764 + c = pop_pair(); 1.1765 + dec_sp(2); // Pop array and index 1.1766 + c = dstore_rounding(c); 1.1767 + store_to_memory(control(), a, c, T_DOUBLE, TypeAryPtr::DOUBLES, MemNode::unordered); 1.1768 + break; 1.1769 + } 1.1770 + case Bytecodes::_getfield: 1.1771 + do_getfield(); 1.1772 + break; 1.1773 + 1.1774 + case Bytecodes::_getstatic: 1.1775 + do_getstatic(); 1.1776 + break; 1.1777 + 1.1778 + case Bytecodes::_putfield: 1.1779 + do_putfield(); 1.1780 + break; 1.1781 + 1.1782 + case Bytecodes::_putstatic: 1.1783 + do_putstatic(); 1.1784 + break; 1.1785 + 1.1786 + case Bytecodes::_irem: 1.1787 + do_irem(); 1.1788 + break; 1.1789 + case Bytecodes::_idiv: 1.1790 + // Must keep both values on the expression-stack during null-check 1.1791 + zero_check_int(peek()); 1.1792 + // Compile-time detect of null-exception? 1.1793 + if (stopped()) return; 1.1794 + b = pop(); 1.1795 + a = pop(); 1.1796 + push( _gvn.transform( new (C) DivINode(control(),a,b) ) ); 1.1797 + break; 1.1798 + case Bytecodes::_imul: 1.1799 + b = pop(); a = pop(); 1.1800 + push( _gvn.transform( new (C) MulINode(a,b) ) ); 1.1801 + break; 1.1802 + case Bytecodes::_iadd: 1.1803 + b = pop(); a = pop(); 1.1804 + push( _gvn.transform( new (C) AddINode(a,b) ) ); 1.1805 + break; 1.1806 + case Bytecodes::_ineg: 1.1807 + a = pop(); 1.1808 + push( _gvn.transform( new (C) SubINode(_gvn.intcon(0),a)) ); 1.1809 + break; 1.1810 + case Bytecodes::_isub: 1.1811 + b = pop(); a = pop(); 1.1812 + push( _gvn.transform( new (C) SubINode(a,b) ) ); 1.1813 + break; 1.1814 + case Bytecodes::_iand: 1.1815 + b = pop(); a = pop(); 1.1816 + push( _gvn.transform( new (C) AndINode(a,b) ) ); 1.1817 + break; 1.1818 + case Bytecodes::_ior: 1.1819 + b = pop(); a = pop(); 1.1820 + push( _gvn.transform( new (C) OrINode(a,b) ) ); 1.1821 + break; 1.1822 + case Bytecodes::_ixor: 1.1823 + b = pop(); a = pop(); 1.1824 + push( _gvn.transform( new (C) XorINode(a,b) ) ); 1.1825 + break; 1.1826 + case Bytecodes::_ishl: 1.1827 + b = pop(); a = pop(); 1.1828 + push( _gvn.transform( new (C) LShiftINode(a,b) ) ); 1.1829 + break; 1.1830 + case Bytecodes::_ishr: 1.1831 + b = pop(); a = pop(); 1.1832 + push( _gvn.transform( new (C) RShiftINode(a,b) ) ); 1.1833 + break; 1.1834 + case Bytecodes::_iushr: 1.1835 + b = pop(); a = pop(); 1.1836 + push( _gvn.transform( new (C) URShiftINode(a,b) ) ); 1.1837 + break; 1.1838 + 1.1839 + case Bytecodes::_fneg: 1.1840 + a = pop(); 1.1841 + b = _gvn.transform(new (C) NegFNode (a)); 1.1842 + push(b); 1.1843 + break; 1.1844 + 1.1845 + case Bytecodes::_fsub: 1.1846 + b = pop(); 1.1847 + a = pop(); 1.1848 + c = _gvn.transform( new (C) SubFNode(a,b) ); 1.1849 + d = precision_rounding(c); 1.1850 + push( d ); 1.1851 + break; 1.1852 + 1.1853 + case Bytecodes::_fadd: 1.1854 + b = pop(); 1.1855 + a = pop(); 1.1856 + c = _gvn.transform( new (C) AddFNode(a,b) ); 1.1857 + d = precision_rounding(c); 1.1858 + push( d ); 1.1859 + break; 1.1860 + 1.1861 + case Bytecodes::_fmul: 1.1862 + b = pop(); 1.1863 + a = pop(); 1.1864 + c = _gvn.transform( new (C) MulFNode(a,b) ); 1.1865 + d = precision_rounding(c); 1.1866 + push( d ); 1.1867 + break; 1.1868 + 1.1869 + case Bytecodes::_fdiv: 1.1870 + b = pop(); 1.1871 + a = pop(); 1.1872 + c = _gvn.transform( new (C) DivFNode(0,a,b) ); 1.1873 + d = precision_rounding(c); 1.1874 + push( d ); 1.1875 + break; 1.1876 + 1.1877 + case Bytecodes::_frem: 1.1878 + if (Matcher::has_match_rule(Op_ModF)) { 1.1879 + // Generate a ModF node. 1.1880 + b = pop(); 1.1881 + a = pop(); 1.1882 + c = _gvn.transform( new (C) ModFNode(0,a,b) ); 1.1883 + d = precision_rounding(c); 1.1884 + push( d ); 1.1885 + } 1.1886 + else { 1.1887 + // Generate a call. 1.1888 + modf(); 1.1889 + } 1.1890 + break; 1.1891 + 1.1892 + case Bytecodes::_fcmpl: 1.1893 + b = pop(); 1.1894 + a = pop(); 1.1895 + c = _gvn.transform( new (C) CmpF3Node( a, b)); 1.1896 + push(c); 1.1897 + break; 1.1898 + case Bytecodes::_fcmpg: 1.1899 + b = pop(); 1.1900 + a = pop(); 1.1901 + 1.1902 + // Same as fcmpl but need to flip the unordered case. Swap the inputs, 1.1903 + // which negates the result sign except for unordered. Flip the unordered 1.1904 + // as well by using CmpF3 which implements unordered-lesser instead of 1.1905 + // unordered-greater semantics. Finally, commute the result bits. Result 1.1906 + // is same as using a CmpF3Greater except we did it with CmpF3 alone. 1.1907 + c = _gvn.transform( new (C) CmpF3Node( b, a)); 1.1908 + c = _gvn.transform( new (C) SubINode(_gvn.intcon(0),c) ); 1.1909 + push(c); 1.1910 + break; 1.1911 + 1.1912 + case Bytecodes::_f2i: 1.1913 + a = pop(); 1.1914 + push(_gvn.transform(new (C) ConvF2INode(a))); 1.1915 + break; 1.1916 + 1.1917 + case Bytecodes::_d2i: 1.1918 + a = pop_pair(); 1.1919 + b = _gvn.transform(new (C) ConvD2INode(a)); 1.1920 + push( b ); 1.1921 + break; 1.1922 + 1.1923 + case Bytecodes::_f2d: 1.1924 + a = pop(); 1.1925 + b = _gvn.transform( new (C) ConvF2DNode(a)); 1.1926 + push_pair( b ); 1.1927 + break; 1.1928 + 1.1929 + case Bytecodes::_d2f: 1.1930 + a = pop_pair(); 1.1931 + b = _gvn.transform( new (C) ConvD2FNode(a)); 1.1932 + // This breaks _227_mtrt (speed & correctness) and _222_mpegaudio (speed) 1.1933 + //b = _gvn.transform(new (C) RoundFloatNode(0, b) ); 1.1934 + push( b ); 1.1935 + break; 1.1936 + 1.1937 + case Bytecodes::_l2f: 1.1938 + if (Matcher::convL2FSupported()) { 1.1939 + a = pop_pair(); 1.1940 + b = _gvn.transform( new (C) ConvL2FNode(a)); 1.1941 + // For i486.ad, FILD doesn't restrict precision to 24 or 53 bits. 1.1942 + // Rather than storing the result into an FP register then pushing 1.1943 + // out to memory to round, the machine instruction that implements 1.1944 + // ConvL2D is responsible for rounding. 1.1945 + // c = precision_rounding(b); 1.1946 + c = _gvn.transform(b); 1.1947 + push(c); 1.1948 + } else { 1.1949 + l2f(); 1.1950 + } 1.1951 + break; 1.1952 + 1.1953 + case Bytecodes::_l2d: 1.1954 + a = pop_pair(); 1.1955 + b = _gvn.transform( new (C) ConvL2DNode(a)); 1.1956 + // For i486.ad, rounding is always necessary (see _l2f above). 1.1957 + // c = dprecision_rounding(b); 1.1958 + c = _gvn.transform(b); 1.1959 + push_pair(c); 1.1960 + break; 1.1961 + 1.1962 + case Bytecodes::_f2l: 1.1963 + a = pop(); 1.1964 + b = _gvn.transform( new (C) ConvF2LNode(a)); 1.1965 + push_pair(b); 1.1966 + break; 1.1967 + 1.1968 + case Bytecodes::_d2l: 1.1969 + a = pop_pair(); 1.1970 + b = _gvn.transform( new (C) ConvD2LNode(a)); 1.1971 + push_pair(b); 1.1972 + break; 1.1973 + 1.1974 + case Bytecodes::_dsub: 1.1975 + b = pop_pair(); 1.1976 + a = pop_pair(); 1.1977 + c = _gvn.transform( new (C) SubDNode(a,b) ); 1.1978 + d = dprecision_rounding(c); 1.1979 + push_pair( d ); 1.1980 + break; 1.1981 + 1.1982 + case Bytecodes::_dadd: 1.1983 + b = pop_pair(); 1.1984 + a = pop_pair(); 1.1985 + c = _gvn.transform( new (C) AddDNode(a,b) ); 1.1986 + d = dprecision_rounding(c); 1.1987 + push_pair( d ); 1.1988 + break; 1.1989 + 1.1990 + case Bytecodes::_dmul: 1.1991 + b = pop_pair(); 1.1992 + a = pop_pair(); 1.1993 + c = _gvn.transform( new (C) MulDNode(a,b) ); 1.1994 + d = dprecision_rounding(c); 1.1995 + push_pair( d ); 1.1996 + break; 1.1997 + 1.1998 + case Bytecodes::_ddiv: 1.1999 + b = pop_pair(); 1.2000 + a = pop_pair(); 1.2001 + c = _gvn.transform( new (C) DivDNode(0,a,b) ); 1.2002 + d = dprecision_rounding(c); 1.2003 + push_pair( d ); 1.2004 + break; 1.2005 + 1.2006 + case Bytecodes::_dneg: 1.2007 + a = pop_pair(); 1.2008 + b = _gvn.transform(new (C) NegDNode (a)); 1.2009 + push_pair(b); 1.2010 + break; 1.2011 + 1.2012 + case Bytecodes::_drem: 1.2013 + if (Matcher::has_match_rule(Op_ModD)) { 1.2014 + // Generate a ModD node. 1.2015 + b = pop_pair(); 1.2016 + a = pop_pair(); 1.2017 + // a % b 1.2018 + 1.2019 + c = _gvn.transform( new (C) ModDNode(0,a,b) ); 1.2020 + d = dprecision_rounding(c); 1.2021 + push_pair( d ); 1.2022 + } 1.2023 + else { 1.2024 + // Generate a call. 1.2025 + modd(); 1.2026 + } 1.2027 + break; 1.2028 + 1.2029 + case Bytecodes::_dcmpl: 1.2030 + b = pop_pair(); 1.2031 + a = pop_pair(); 1.2032 + c = _gvn.transform( new (C) CmpD3Node( a, b)); 1.2033 + push(c); 1.2034 + break; 1.2035 + 1.2036 + case Bytecodes::_dcmpg: 1.2037 + b = pop_pair(); 1.2038 + a = pop_pair(); 1.2039 + // Same as dcmpl but need to flip the unordered case. 1.2040 + // Commute the inputs, which negates the result sign except for unordered. 1.2041 + // Flip the unordered as well by using CmpD3 which implements 1.2042 + // unordered-lesser instead of unordered-greater semantics. 1.2043 + // Finally, negate the result bits. Result is same as using a 1.2044 + // CmpD3Greater except we did it with CmpD3 alone. 1.2045 + c = _gvn.transform( new (C) CmpD3Node( b, a)); 1.2046 + c = _gvn.transform( new (C) SubINode(_gvn.intcon(0),c) ); 1.2047 + push(c); 1.2048 + break; 1.2049 + 1.2050 + 1.2051 + // Note for longs -> lo word is on TOS, hi word is on TOS - 1 1.2052 + case Bytecodes::_land: 1.2053 + b = pop_pair(); 1.2054 + a = pop_pair(); 1.2055 + c = _gvn.transform( new (C) AndLNode(a,b) ); 1.2056 + push_pair(c); 1.2057 + break; 1.2058 + case Bytecodes::_lor: 1.2059 + b = pop_pair(); 1.2060 + a = pop_pair(); 1.2061 + c = _gvn.transform( new (C) OrLNode(a,b) ); 1.2062 + push_pair(c); 1.2063 + break; 1.2064 + case Bytecodes::_lxor: 1.2065 + b = pop_pair(); 1.2066 + a = pop_pair(); 1.2067 + c = _gvn.transform( new (C) XorLNode(a,b) ); 1.2068 + push_pair(c); 1.2069 + break; 1.2070 + 1.2071 + case Bytecodes::_lshl: 1.2072 + b = pop(); // the shift count 1.2073 + a = pop_pair(); // value to be shifted 1.2074 + c = _gvn.transform( new (C) LShiftLNode(a,b) ); 1.2075 + push_pair(c); 1.2076 + break; 1.2077 + case Bytecodes::_lshr: 1.2078 + b = pop(); // the shift count 1.2079 + a = pop_pair(); // value to be shifted 1.2080 + c = _gvn.transform( new (C) RShiftLNode(a,b) ); 1.2081 + push_pair(c); 1.2082 + break; 1.2083 + case Bytecodes::_lushr: 1.2084 + b = pop(); // the shift count 1.2085 + a = pop_pair(); // value to be shifted 1.2086 + c = _gvn.transform( new (C) URShiftLNode(a,b) ); 1.2087 + push_pair(c); 1.2088 + break; 1.2089 + case Bytecodes::_lmul: 1.2090 + b = pop_pair(); 1.2091 + a = pop_pair(); 1.2092 + c = _gvn.transform( new (C) MulLNode(a,b) ); 1.2093 + push_pair(c); 1.2094 + break; 1.2095 + 1.2096 + case Bytecodes::_lrem: 1.2097 + // Must keep both values on the expression-stack during null-check 1.2098 + assert(peek(0) == top(), "long word order"); 1.2099 + zero_check_long(peek(1)); 1.2100 + // Compile-time detect of null-exception? 1.2101 + if (stopped()) return; 1.2102 + b = pop_pair(); 1.2103 + a = pop_pair(); 1.2104 + c = _gvn.transform( new (C) ModLNode(control(),a,b) ); 1.2105 + push_pair(c); 1.2106 + break; 1.2107 + 1.2108 + case Bytecodes::_ldiv: 1.2109 + // Must keep both values on the expression-stack during null-check 1.2110 + assert(peek(0) == top(), "long word order"); 1.2111 + zero_check_long(peek(1)); 1.2112 + // Compile-time detect of null-exception? 1.2113 + if (stopped()) return; 1.2114 + b = pop_pair(); 1.2115 + a = pop_pair(); 1.2116 + c = _gvn.transform( new (C) DivLNode(control(),a,b) ); 1.2117 + push_pair(c); 1.2118 + break; 1.2119 + 1.2120 + case Bytecodes::_ladd: 1.2121 + b = pop_pair(); 1.2122 + a = pop_pair(); 1.2123 + c = _gvn.transform( new (C) AddLNode(a,b) ); 1.2124 + push_pair(c); 1.2125 + break; 1.2126 + case Bytecodes::_lsub: 1.2127 + b = pop_pair(); 1.2128 + a = pop_pair(); 1.2129 + c = _gvn.transform( new (C) SubLNode(a,b) ); 1.2130 + push_pair(c); 1.2131 + break; 1.2132 + case Bytecodes::_lcmp: 1.2133 + // Safepoints are now inserted _before_ branches. The long-compare 1.2134 + // bytecode painfully produces a 3-way value (-1,0,+1) which requires a 1.2135 + // slew of control flow. These are usually followed by a CmpI vs zero and 1.2136 + // a branch; this pattern then optimizes to the obvious long-compare and 1.2137 + // branch. However, if the branch is backwards there's a Safepoint 1.2138 + // inserted. The inserted Safepoint captures the JVM state at the 1.2139 + // pre-branch point, i.e. it captures the 3-way value. Thus if a 1.2140 + // long-compare is used to control a loop the debug info will force 1.2141 + // computation of the 3-way value, even though the generated code uses a 1.2142 + // long-compare and branch. We try to rectify the situation by inserting 1.2143 + // a SafePoint here and have it dominate and kill the safepoint added at a 1.2144 + // following backwards branch. At this point the JVM state merely holds 2 1.2145 + // longs but not the 3-way value. 1.2146 + if( UseLoopSafepoints ) { 1.2147 + switch( iter().next_bc() ) { 1.2148 + case Bytecodes::_ifgt: 1.2149 + case Bytecodes::_iflt: 1.2150 + case Bytecodes::_ifge: 1.2151 + case Bytecodes::_ifle: 1.2152 + case Bytecodes::_ifne: 1.2153 + case Bytecodes::_ifeq: 1.2154 + // If this is a backwards branch in the bytecodes, add Safepoint 1.2155 + maybe_add_safepoint(iter().next_get_dest()); 1.2156 + } 1.2157 + } 1.2158 + b = pop_pair(); 1.2159 + a = pop_pair(); 1.2160 + c = _gvn.transform( new (C) CmpL3Node( a, b )); 1.2161 + push(c); 1.2162 + break; 1.2163 + 1.2164 + case Bytecodes::_lneg: 1.2165 + a = pop_pair(); 1.2166 + b = _gvn.transform( new (C) SubLNode(longcon(0),a)); 1.2167 + push_pair(b); 1.2168 + break; 1.2169 + case Bytecodes::_l2i: 1.2170 + a = pop_pair(); 1.2171 + push( _gvn.transform( new (C) ConvL2INode(a))); 1.2172 + break; 1.2173 + case Bytecodes::_i2l: 1.2174 + a = pop(); 1.2175 + b = _gvn.transform( new (C) ConvI2LNode(a)); 1.2176 + push_pair(b); 1.2177 + break; 1.2178 + case Bytecodes::_i2b: 1.2179 + // Sign extend 1.2180 + a = pop(); 1.2181 + a = _gvn.transform( new (C) LShiftINode(a,_gvn.intcon(24)) ); 1.2182 + a = _gvn.transform( new (C) RShiftINode(a,_gvn.intcon(24)) ); 1.2183 + push( a ); 1.2184 + break; 1.2185 + case Bytecodes::_i2s: 1.2186 + a = pop(); 1.2187 + a = _gvn.transform( new (C) LShiftINode(a,_gvn.intcon(16)) ); 1.2188 + a = _gvn.transform( new (C) RShiftINode(a,_gvn.intcon(16)) ); 1.2189 + push( a ); 1.2190 + break; 1.2191 + case Bytecodes::_i2c: 1.2192 + a = pop(); 1.2193 + push( _gvn.transform( new (C) AndINode(a,_gvn.intcon(0xFFFF)) ) ); 1.2194 + break; 1.2195 + 1.2196 + case Bytecodes::_i2f: 1.2197 + a = pop(); 1.2198 + b = _gvn.transform( new (C) ConvI2FNode(a) ) ; 1.2199 + c = precision_rounding(b); 1.2200 + push (b); 1.2201 + break; 1.2202 + 1.2203 + case Bytecodes::_i2d: 1.2204 + a = pop(); 1.2205 + b = _gvn.transform( new (C) ConvI2DNode(a)); 1.2206 + push_pair(b); 1.2207 + break; 1.2208 + 1.2209 + case Bytecodes::_iinc: // Increment local 1.2210 + i = iter().get_index(); // Get local index 1.2211 + set_local( i, _gvn.transform( new (C) AddINode( _gvn.intcon(iter().get_iinc_con()), local(i) ) ) ); 1.2212 + break; 1.2213 + 1.2214 + // Exit points of synchronized methods must have an unlock node 1.2215 + case Bytecodes::_return: 1.2216 + return_current(NULL); 1.2217 + break; 1.2218 + 1.2219 + case Bytecodes::_ireturn: 1.2220 + case Bytecodes::_areturn: 1.2221 + case Bytecodes::_freturn: 1.2222 + return_current(pop()); 1.2223 + break; 1.2224 + case Bytecodes::_lreturn: 1.2225 + return_current(pop_pair()); 1.2226 + break; 1.2227 + case Bytecodes::_dreturn: 1.2228 + return_current(pop_pair()); 1.2229 + break; 1.2230 + 1.2231 + case Bytecodes::_athrow: 1.2232 + // null exception oop throws NULL pointer exception 1.2233 + null_check(peek()); 1.2234 + if (stopped()) return; 1.2235 + // Hook the thrown exception directly to subsequent handlers. 1.2236 + if (BailoutToInterpreterForThrows) { 1.2237 + // Keep method interpreted from now on. 1.2238 + uncommon_trap(Deoptimization::Reason_unhandled, 1.2239 + Deoptimization::Action_make_not_compilable); 1.2240 + return; 1.2241 + } 1.2242 + if (env()->jvmti_can_post_on_exceptions()) { 1.2243 + // check if we must post exception events, take uncommon trap if so (with must_throw = false) 1.2244 + uncommon_trap_if_should_post_on_exceptions(Deoptimization::Reason_unhandled, false); 1.2245 + } 1.2246 + // Here if either can_post_on_exceptions or should_post_on_exceptions is false 1.2247 + add_exception_state(make_exception_state(peek())); 1.2248 + break; 1.2249 + 1.2250 + case Bytecodes::_goto: // fall through 1.2251 + case Bytecodes::_goto_w: { 1.2252 + int target_bci = (bc() == Bytecodes::_goto) ? iter().get_dest() : iter().get_far_dest(); 1.2253 + 1.2254 + // If this is a backwards branch in the bytecodes, add Safepoint 1.2255 + maybe_add_safepoint(target_bci); 1.2256 + 1.2257 + // Update method data 1.2258 + profile_taken_branch(target_bci); 1.2259 + 1.2260 + // Merge the current control into the target basic block 1.2261 + merge(target_bci); 1.2262 + 1.2263 + // See if we can get some profile data and hand it off to the next block 1.2264 + Block *target_block = block()->successor_for_bci(target_bci); 1.2265 + if (target_block->pred_count() != 1) break; 1.2266 + ciMethodData* methodData = method()->method_data(); 1.2267 + if (!methodData->is_mature()) break; 1.2268 + ciProfileData* data = methodData->bci_to_data(bci()); 1.2269 + assert( data->is_JumpData(), "" ); 1.2270 + int taken = ((ciJumpData*)data)->taken(); 1.2271 + taken = method()->scale_count(taken); 1.2272 + target_block->set_count(taken); 1.2273 + break; 1.2274 + } 1.2275 + 1.2276 + case Bytecodes::_ifnull: btest = BoolTest::eq; goto handle_if_null; 1.2277 + case Bytecodes::_ifnonnull: btest = BoolTest::ne; goto handle_if_null; 1.2278 + handle_if_null: 1.2279 + // If this is a backwards branch in the bytecodes, add Safepoint 1.2280 + maybe_add_safepoint(iter().get_dest()); 1.2281 + a = null(); 1.2282 + b = pop(); 1.2283 + c = _gvn.transform( new (C) CmpPNode(b, a) ); 1.2284 + do_ifnull(btest, c); 1.2285 + break; 1.2286 + 1.2287 + case Bytecodes::_if_acmpeq: btest = BoolTest::eq; goto handle_if_acmp; 1.2288 + case Bytecodes::_if_acmpne: btest = BoolTest::ne; goto handle_if_acmp; 1.2289 + handle_if_acmp: 1.2290 + // If this is a backwards branch in the bytecodes, add Safepoint 1.2291 + maybe_add_safepoint(iter().get_dest()); 1.2292 + a = pop(); 1.2293 + b = pop(); 1.2294 + c = _gvn.transform( new (C) CmpPNode(b, a) ); 1.2295 + c = optimize_cmp_with_klass(c); 1.2296 + do_if(btest, c); 1.2297 + break; 1.2298 + 1.2299 + case Bytecodes::_ifeq: btest = BoolTest::eq; goto handle_ifxx; 1.2300 + case Bytecodes::_ifne: btest = BoolTest::ne; goto handle_ifxx; 1.2301 + case Bytecodes::_iflt: btest = BoolTest::lt; goto handle_ifxx; 1.2302 + case Bytecodes::_ifle: btest = BoolTest::le; goto handle_ifxx; 1.2303 + case Bytecodes::_ifgt: btest = BoolTest::gt; goto handle_ifxx; 1.2304 + case Bytecodes::_ifge: btest = BoolTest::ge; goto handle_ifxx; 1.2305 + handle_ifxx: 1.2306 + // If this is a backwards branch in the bytecodes, add Safepoint 1.2307 + maybe_add_safepoint(iter().get_dest()); 1.2308 + a = _gvn.intcon(0); 1.2309 + b = pop(); 1.2310 + c = _gvn.transform( new (C) CmpINode(b, a) ); 1.2311 + do_if(btest, c); 1.2312 + break; 1.2313 + 1.2314 + case Bytecodes::_if_icmpeq: btest = BoolTest::eq; goto handle_if_icmp; 1.2315 + case Bytecodes::_if_icmpne: btest = BoolTest::ne; goto handle_if_icmp; 1.2316 + case Bytecodes::_if_icmplt: btest = BoolTest::lt; goto handle_if_icmp; 1.2317 + case Bytecodes::_if_icmple: btest = BoolTest::le; goto handle_if_icmp; 1.2318 + case Bytecodes::_if_icmpgt: btest = BoolTest::gt; goto handle_if_icmp; 1.2319 + case Bytecodes::_if_icmpge: btest = BoolTest::ge; goto handle_if_icmp; 1.2320 + handle_if_icmp: 1.2321 + // If this is a backwards branch in the bytecodes, add Safepoint 1.2322 + maybe_add_safepoint(iter().get_dest()); 1.2323 + a = pop(); 1.2324 + b = pop(); 1.2325 + c = _gvn.transform( new (C) CmpINode( b, a ) ); 1.2326 + do_if(btest, c); 1.2327 + break; 1.2328 + 1.2329 + case Bytecodes::_tableswitch: 1.2330 + do_tableswitch(); 1.2331 + break; 1.2332 + 1.2333 + case Bytecodes::_lookupswitch: 1.2334 + do_lookupswitch(); 1.2335 + break; 1.2336 + 1.2337 + case Bytecodes::_invokestatic: 1.2338 + case Bytecodes::_invokedynamic: 1.2339 + case Bytecodes::_invokespecial: 1.2340 + case Bytecodes::_invokevirtual: 1.2341 + case Bytecodes::_invokeinterface: 1.2342 + do_call(); 1.2343 + break; 1.2344 + case Bytecodes::_checkcast: 1.2345 + do_checkcast(); 1.2346 + break; 1.2347 + case Bytecodes::_instanceof: 1.2348 + do_instanceof(); 1.2349 + break; 1.2350 + case Bytecodes::_anewarray: 1.2351 + do_anewarray(); 1.2352 + break; 1.2353 + case Bytecodes::_newarray: 1.2354 + do_newarray((BasicType)iter().get_index()); 1.2355 + break; 1.2356 + case Bytecodes::_multianewarray: 1.2357 + do_multianewarray(); 1.2358 + break; 1.2359 + case Bytecodes::_new: 1.2360 + do_new(); 1.2361 + break; 1.2362 + 1.2363 + case Bytecodes::_jsr: 1.2364 + case Bytecodes::_jsr_w: 1.2365 + do_jsr(); 1.2366 + break; 1.2367 + 1.2368 + case Bytecodes::_ret: 1.2369 + do_ret(); 1.2370 + break; 1.2371 + 1.2372 + 1.2373 + case Bytecodes::_monitorenter: 1.2374 + do_monitor_enter(); 1.2375 + break; 1.2376 + 1.2377 + case Bytecodes::_monitorexit: 1.2378 + do_monitor_exit(); 1.2379 + break; 1.2380 + 1.2381 + case Bytecodes::_breakpoint: 1.2382 + // Breakpoint set concurrently to compile 1.2383 + // %%% use an uncommon trap? 1.2384 + C->record_failure("breakpoint in method"); 1.2385 + return; 1.2386 + 1.2387 + default: 1.2388 +#ifndef PRODUCT 1.2389 + map()->dump(99); 1.2390 +#endif 1.2391 + tty->print("\nUnhandled bytecode %s\n", Bytecodes::name(bc()) ); 1.2392 + ShouldNotReachHere(); 1.2393 + } 1.2394 + 1.2395 +#ifndef PRODUCT 1.2396 + IdealGraphPrinter *printer = IdealGraphPrinter::printer(); 1.2397 + if(printer) { 1.2398 + char buffer[256]; 1.2399 + sprintf(buffer, "Bytecode %d: %s", bci(), Bytecodes::name(bc())); 1.2400 + bool old = printer->traverse_outs(); 1.2401 + printer->set_traverse_outs(true); 1.2402 + printer->print_method(C, buffer, 4); 1.2403 + printer->set_traverse_outs(old); 1.2404 + } 1.2405 +#endif 1.2406 +}