Tue, 02 Sep 2014 12:48:45 -0700
8055494: Add C2 x86 intrinsic for BigInteger::multiplyToLen() method
Summary: Add new C2 intrinsic for BigInteger::multiplyToLen() on x86 in 64-bit VM.
Reviewed-by: roland
duke@435 | 1 | /* |
drchase@6680 | 2 | * Copyright (c) 1997, 2014, 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" |
stefank@2314 | 26 | #include "libadt/vectset.hpp" |
stefank@2314 | 27 | #include "memory/allocation.inline.hpp" |
stefank@2314 | 28 | #include "opto/block.hpp" |
stefank@2314 | 29 | #include "opto/cfgnode.hpp" |
stefank@2314 | 30 | #include "opto/chaitin.hpp" |
stefank@2314 | 31 | #include "opto/loopnode.hpp" |
stefank@2314 | 32 | #include "opto/machnode.hpp" |
stefank@2314 | 33 | #include "opto/matcher.hpp" |
stefank@2314 | 34 | #include "opto/opcodes.hpp" |
stefank@2314 | 35 | #include "opto/rootnode.hpp" |
stefank@2314 | 36 | #include "utilities/copy.hpp" |
stefank@2314 | 37 | |
duke@435 | 38 | void Block_Array::grow( uint i ) { |
duke@435 | 39 | assert(i >= Max(), "must be an overflow"); |
duke@435 | 40 | debug_only(_limit = i+1); |
duke@435 | 41 | if( i < _size ) return; |
duke@435 | 42 | if( !_size ) { |
duke@435 | 43 | _size = 1; |
duke@435 | 44 | _blocks = (Block**)_arena->Amalloc( _size * sizeof(Block*) ); |
duke@435 | 45 | _blocks[0] = NULL; |
duke@435 | 46 | } |
duke@435 | 47 | uint old = _size; |
duke@435 | 48 | while( i >= _size ) _size <<= 1; // Double to fit |
duke@435 | 49 | _blocks = (Block**)_arena->Arealloc( _blocks, old*sizeof(Block*),_size*sizeof(Block*)); |
duke@435 | 50 | Copy::zero_to_bytes( &_blocks[old], (_size-old)*sizeof(Block*) ); |
duke@435 | 51 | } |
duke@435 | 52 | |
duke@435 | 53 | void Block_List::remove(uint i) { |
duke@435 | 54 | assert(i < _cnt, "index out of bounds"); |
duke@435 | 55 | Copy::conjoint_words_to_lower((HeapWord*)&_blocks[i+1], (HeapWord*)&_blocks[i], ((_cnt-i-1)*sizeof(Block*))); |
duke@435 | 56 | pop(); // shrink list by one block |
duke@435 | 57 | } |
duke@435 | 58 | |
duke@435 | 59 | void Block_List::insert(uint i, Block *b) { |
duke@435 | 60 | push(b); // grow list by one block |
duke@435 | 61 | Copy::conjoint_words_to_higher((HeapWord*)&_blocks[i], (HeapWord*)&_blocks[i+1], ((_cnt-i-1)*sizeof(Block*))); |
duke@435 | 62 | _blocks[i] = b; |
duke@435 | 63 | } |
duke@435 | 64 | |
rasbold@853 | 65 | #ifndef PRODUCT |
rasbold@853 | 66 | void Block_List::print() { |
rasbold@853 | 67 | for (uint i=0; i < size(); i++) { |
rasbold@853 | 68 | tty->print("B%d ", _blocks[i]->_pre_order); |
rasbold@853 | 69 | } |
rasbold@853 | 70 | tty->print("size = %d\n", size()); |
rasbold@853 | 71 | } |
rasbold@853 | 72 | #endif |
duke@435 | 73 | |
duke@435 | 74 | uint Block::code_alignment() { |
duke@435 | 75 | // Check for Root block |
kvn@3049 | 76 | if (_pre_order == 0) return CodeEntryAlignment; |
duke@435 | 77 | // Check for Start block |
kvn@3049 | 78 | if (_pre_order == 1) return InteriorEntryAlignment; |
duke@435 | 79 | // Check for loop alignment |
kvn@3049 | 80 | if (has_loop_alignment()) return loop_alignment(); |
rasbold@853 | 81 | |
kvn@3049 | 82 | return relocInfo::addr_unit(); // no particular alignment |
rasbold@853 | 83 | } |
rasbold@853 | 84 | |
rasbold@853 | 85 | uint Block::compute_loop_alignment() { |
duke@435 | 86 | Node *h = head(); |
kvn@3049 | 87 | int unit_sz = relocInfo::addr_unit(); |
kvn@3049 | 88 | if (h->is_Loop() && h->as_Loop()->is_inner_loop()) { |
duke@435 | 89 | // Pre- and post-loops have low trip count so do not bother with |
duke@435 | 90 | // NOPs for align loop head. The constants are hidden from tuning |
duke@435 | 91 | // but only because my "divide by 4" heuristic surely gets nearly |
duke@435 | 92 | // all possible gain (a "do not align at all" heuristic has a |
duke@435 | 93 | // chance of getting a really tiny gain). |
kvn@3049 | 94 | if (h->is_CountedLoop() && (h->as_CountedLoop()->is_pre_loop() || |
kvn@3049 | 95 | h->as_CountedLoop()->is_post_loop())) { |
kvn@3049 | 96 | return (OptoLoopAlignment > 4*unit_sz) ? (OptoLoopAlignment>>2) : unit_sz; |
kvn@3049 | 97 | } |
duke@435 | 98 | // Loops with low backedge frequency should not be aligned. |
duke@435 | 99 | Node *n = h->in(LoopNode::LoopBackControl)->in(0); |
kvn@3049 | 100 | if (n->is_MachIf() && n->as_MachIf()->_prob < 0.01) { |
kvn@3049 | 101 | return unit_sz; // Loop does not loop, more often than not! |
duke@435 | 102 | } |
duke@435 | 103 | return OptoLoopAlignment; // Otherwise align loop head |
duke@435 | 104 | } |
rasbold@853 | 105 | |
kvn@3049 | 106 | return unit_sz; // no particular alignment |
duke@435 | 107 | } |
duke@435 | 108 | |
duke@435 | 109 | // Compute the size of first 'inst_cnt' instructions in this block. |
duke@435 | 110 | // Return the number of instructions left to compute if the block has |
rasbold@853 | 111 | // less then 'inst_cnt' instructions. Stop, and return 0 if sum_size |
rasbold@853 | 112 | // exceeds OptoLoopAlignment. |
duke@435 | 113 | uint Block::compute_first_inst_size(uint& sum_size, uint inst_cnt, |
duke@435 | 114 | PhaseRegAlloc* ra) { |
adlertz@5635 | 115 | uint last_inst = number_of_nodes(); |
duke@435 | 116 | for( uint j = 0; j < last_inst && inst_cnt > 0; j++ ) { |
adlertz@5635 | 117 | uint inst_size = get_node(j)->size(ra); |
duke@435 | 118 | if( inst_size > 0 ) { |
duke@435 | 119 | inst_cnt--; |
duke@435 | 120 | uint sz = sum_size + inst_size; |
duke@435 | 121 | if( sz <= (uint)OptoLoopAlignment ) { |
duke@435 | 122 | // Compute size of instructions which fit into fetch buffer only |
duke@435 | 123 | // since all inst_cnt instructions will not fit even if we align them. |
duke@435 | 124 | sum_size = sz; |
duke@435 | 125 | } else { |
duke@435 | 126 | return 0; |
duke@435 | 127 | } |
duke@435 | 128 | } |
duke@435 | 129 | } |
duke@435 | 130 | return inst_cnt; |
duke@435 | 131 | } |
duke@435 | 132 | |
duke@435 | 133 | uint Block::find_node( const Node *n ) const { |
adlertz@5635 | 134 | for( uint i = 0; i < number_of_nodes(); i++ ) { |
adlertz@5635 | 135 | if( get_node(i) == n ) |
duke@435 | 136 | return i; |
duke@435 | 137 | } |
duke@435 | 138 | ShouldNotReachHere(); |
duke@435 | 139 | return 0; |
duke@435 | 140 | } |
duke@435 | 141 | |
duke@435 | 142 | // Find and remove n from block list |
duke@435 | 143 | void Block::find_remove( const Node *n ) { |
adlertz@5635 | 144 | remove_node(find_node(n)); |
duke@435 | 145 | } |
duke@435 | 146 | |
goetz@6478 | 147 | bool Block::contains(const Node *n) const { |
goetz@6478 | 148 | return _nodes.contains(n); |
goetz@6478 | 149 | } |
goetz@6478 | 150 | |
duke@435 | 151 | // Return empty status of a block. Empty blocks contain only the head, other |
duke@435 | 152 | // ideal nodes, and an optional trailing goto. |
duke@435 | 153 | int Block::is_Empty() const { |
duke@435 | 154 | |
duke@435 | 155 | // Root or start block is not considered empty |
duke@435 | 156 | if (head()->is_Root() || head()->is_Start()) { |
duke@435 | 157 | return not_empty; |
duke@435 | 158 | } |
duke@435 | 159 | |
duke@435 | 160 | int success_result = completely_empty; |
adlertz@5635 | 161 | int end_idx = number_of_nodes() - 1; |
duke@435 | 162 | |
duke@435 | 163 | // Check for ending goto |
adlertz@5635 | 164 | if ((end_idx > 0) && (get_node(end_idx)->is_MachGoto())) { |
duke@435 | 165 | success_result = empty_with_goto; |
duke@435 | 166 | end_idx--; |
duke@435 | 167 | } |
duke@435 | 168 | |
duke@435 | 169 | // Unreachable blocks are considered empty |
duke@435 | 170 | if (num_preds() <= 1) { |
duke@435 | 171 | return success_result; |
duke@435 | 172 | } |
duke@435 | 173 | |
duke@435 | 174 | // Ideal nodes are allowable in empty blocks: skip them Only MachNodes |
duke@435 | 175 | // turn directly into code, because only MachNodes have non-trivial |
duke@435 | 176 | // emit() functions. |
adlertz@5635 | 177 | while ((end_idx > 0) && !get_node(end_idx)->is_Mach()) { |
duke@435 | 178 | end_idx--; |
duke@435 | 179 | } |
duke@435 | 180 | |
duke@435 | 181 | // No room for any interesting instructions? |
duke@435 | 182 | if (end_idx == 0) { |
duke@435 | 183 | return success_result; |
duke@435 | 184 | } |
duke@435 | 185 | |
duke@435 | 186 | return not_empty; |
duke@435 | 187 | } |
duke@435 | 188 | |
twisti@1040 | 189 | // Return true if the block's code implies that it is likely to be |
duke@435 | 190 | // executed infrequently. Check to see if the block ends in a Halt or |
duke@435 | 191 | // a low probability call. |
duke@435 | 192 | bool Block::has_uncommon_code() const { |
duke@435 | 193 | Node* en = end(); |
duke@435 | 194 | |
kvn@3040 | 195 | if (en->is_MachGoto()) |
duke@435 | 196 | en = en->in(0); |
duke@435 | 197 | if (en->is_Catch()) |
duke@435 | 198 | en = en->in(0); |
kvn@3040 | 199 | if (en->is_MachProj() && en->in(0)->is_MachCall()) { |
duke@435 | 200 | MachCallNode* call = en->in(0)->as_MachCall(); |
duke@435 | 201 | if (call->cnt() != COUNT_UNKNOWN && call->cnt() <= PROB_UNLIKELY_MAG(4)) { |
duke@435 | 202 | // This is true for slow-path stubs like new_{instance,array}, |
duke@435 | 203 | // slow_arraycopy, complete_monitor_locking, uncommon_trap. |
duke@435 | 204 | // The magic number corresponds to the probability of an uncommon_trap, |
duke@435 | 205 | // even though it is a count not a probability. |
duke@435 | 206 | return true; |
duke@435 | 207 | } |
duke@435 | 208 | } |
duke@435 | 209 | |
duke@435 | 210 | int op = en->is_Mach() ? en->as_Mach()->ideal_Opcode() : en->Opcode(); |
duke@435 | 211 | return op == Op_Halt; |
duke@435 | 212 | } |
duke@435 | 213 | |
duke@435 | 214 | // True if block is low enough frequency or guarded by a test which |
duke@435 | 215 | // mostly does not go here. |
adlertz@5639 | 216 | bool PhaseCFG::is_uncommon(const Block* block) { |
duke@435 | 217 | // Initial blocks must never be moved, so are never uncommon. |
adlertz@5639 | 218 | if (block->head()->is_Root() || block->head()->is_Start()) return false; |
duke@435 | 219 | |
duke@435 | 220 | // Check for way-low freq |
adlertz@5639 | 221 | if(block->_freq < BLOCK_FREQUENCY(0.00001f) ) return true; |
duke@435 | 222 | |
duke@435 | 223 | // Look for code shape indicating uncommon_trap or slow path |
adlertz@5639 | 224 | if (block->has_uncommon_code()) return true; |
duke@435 | 225 | |
duke@435 | 226 | const float epsilon = 0.05f; |
duke@435 | 227 | const float guard_factor = PROB_UNLIKELY_MAG(4) / (1.f - epsilon); |
duke@435 | 228 | uint uncommon_preds = 0; |
duke@435 | 229 | uint freq_preds = 0; |
duke@435 | 230 | uint uncommon_for_freq_preds = 0; |
duke@435 | 231 | |
adlertz@5639 | 232 | for( uint i=1; i< block->num_preds(); i++ ) { |
adlertz@5639 | 233 | Block* guard = get_block_for_node(block->pred(i)); |
duke@435 | 234 | // Check to see if this block follows its guard 1 time out of 10000 |
duke@435 | 235 | // or less. |
duke@435 | 236 | // |
duke@435 | 237 | // See list of magnitude-4 unlikely probabilities in cfgnode.hpp which |
duke@435 | 238 | // we intend to be "uncommon", such as slow-path TLE allocation, |
duke@435 | 239 | // predicted call failure, and uncommon trap triggers. |
duke@435 | 240 | // |
duke@435 | 241 | // Use an epsilon value of 5% to allow for variability in frequency |
duke@435 | 242 | // predictions and floating point calculations. The net effect is |
duke@435 | 243 | // that guard_factor is set to 9500. |
duke@435 | 244 | // |
duke@435 | 245 | // Ignore low-frequency blocks. |
duke@435 | 246 | // The next check is (guard->_freq < 1.e-5 * 9500.). |
duke@435 | 247 | if(guard->_freq*BLOCK_FREQUENCY(guard_factor) < BLOCK_FREQUENCY(0.00001f)) { |
duke@435 | 248 | uncommon_preds++; |
duke@435 | 249 | } else { |
duke@435 | 250 | freq_preds++; |
adlertz@5639 | 251 | if(block->_freq < guard->_freq * guard_factor ) { |
duke@435 | 252 | uncommon_for_freq_preds++; |
duke@435 | 253 | } |
duke@435 | 254 | } |
duke@435 | 255 | } |
adlertz@5639 | 256 | if( block->num_preds() > 1 && |
duke@435 | 257 | // The block is uncommon if all preds are uncommon or |
adlertz@5639 | 258 | (uncommon_preds == (block->num_preds()-1) || |
duke@435 | 259 | // it is uncommon for all frequent preds. |
duke@435 | 260 | uncommon_for_freq_preds == freq_preds) ) { |
duke@435 | 261 | return true; |
duke@435 | 262 | } |
duke@435 | 263 | return false; |
duke@435 | 264 | } |
duke@435 | 265 | |
duke@435 | 266 | #ifndef PRODUCT |
kvn@3049 | 267 | void Block::dump_bidx(const Block* orig, outputStream* st) const { |
kvn@3049 | 268 | if (_pre_order) st->print("B%d",_pre_order); |
kvn@3049 | 269 | else st->print("N%d", head()->_idx); |
duke@435 | 270 | |
duke@435 | 271 | if (Verbose && orig != this) { |
duke@435 | 272 | // Dump the original block's idx |
kvn@3049 | 273 | st->print(" ("); |
kvn@3049 | 274 | orig->dump_bidx(orig, st); |
kvn@3049 | 275 | st->print(")"); |
duke@435 | 276 | } |
duke@435 | 277 | } |
duke@435 | 278 | |
adlertz@5509 | 279 | void Block::dump_pred(const PhaseCFG* cfg, Block* orig, outputStream* st) const { |
duke@435 | 280 | if (is_connector()) { |
duke@435 | 281 | for (uint i=1; i<num_preds(); i++) { |
adlertz@5509 | 282 | Block *p = cfg->get_block_for_node(pred(i)); |
adlertz@5509 | 283 | p->dump_pred(cfg, orig, st); |
duke@435 | 284 | } |
duke@435 | 285 | } else { |
kvn@3049 | 286 | dump_bidx(orig, st); |
kvn@3049 | 287 | st->print(" "); |
duke@435 | 288 | } |
duke@435 | 289 | } |
duke@435 | 290 | |
adlertz@5509 | 291 | void Block::dump_head(const PhaseCFG* cfg, outputStream* st) const { |
duke@435 | 292 | // Print the basic block |
kvn@3049 | 293 | dump_bidx(this, st); |
kvn@3049 | 294 | st->print(": #\t"); |
duke@435 | 295 | |
duke@435 | 296 | // Print the incoming CFG edges and the outgoing CFG edges |
duke@435 | 297 | for( uint i=0; i<_num_succs; i++ ) { |
kvn@3049 | 298 | non_connector_successor(i)->dump_bidx(_succs[i], st); |
kvn@3049 | 299 | st->print(" "); |
duke@435 | 300 | } |
kvn@3049 | 301 | st->print("<- "); |
duke@435 | 302 | if( head()->is_block_start() ) { |
duke@435 | 303 | for (uint i=1; i<num_preds(); i++) { |
duke@435 | 304 | Node *s = pred(i); |
adlertz@5509 | 305 | if (cfg != NULL) { |
adlertz@5509 | 306 | Block *p = cfg->get_block_for_node(s); |
adlertz@5509 | 307 | p->dump_pred(cfg, p, st); |
duke@435 | 308 | } else { |
duke@435 | 309 | while (!s->is_block_start()) |
duke@435 | 310 | s = s->in(0); |
kvn@3049 | 311 | st->print("N%d ", s->_idx ); |
duke@435 | 312 | } |
duke@435 | 313 | } |
adlertz@5509 | 314 | } else { |
kvn@3049 | 315 | st->print("BLOCK HEAD IS JUNK "); |
adlertz@5509 | 316 | } |
duke@435 | 317 | |
duke@435 | 318 | // Print loop, if any |
duke@435 | 319 | const Block *bhead = this; // Head of self-loop |
duke@435 | 320 | Node *bh = bhead->head(); |
adlertz@5509 | 321 | |
adlertz@5509 | 322 | if ((cfg != NULL) && bh->is_Loop() && !head()->is_Root()) { |
duke@435 | 323 | LoopNode *loop = bh->as_Loop(); |
adlertz@5509 | 324 | const Block *bx = cfg->get_block_for_node(loop->in(LoopNode::LoopBackControl)); |
duke@435 | 325 | while (bx->is_connector()) { |
adlertz@5509 | 326 | bx = cfg->get_block_for_node(bx->pred(1)); |
duke@435 | 327 | } |
kvn@3049 | 328 | st->print("\tLoop: B%d-B%d ", bhead->_pre_order, bx->_pre_order); |
duke@435 | 329 | // Dump any loop-specific bits, especially for CountedLoops. |
kvn@3049 | 330 | loop->dump_spec(st); |
rasbold@853 | 331 | } else if (has_loop_alignment()) { |
kvn@3049 | 332 | st->print(" top-of-loop"); |
duke@435 | 333 | } |
kvn@3049 | 334 | st->print(" Freq: %g",_freq); |
duke@435 | 335 | if( Verbose || WizardMode ) { |
kvn@3049 | 336 | st->print(" IDom: %d/#%d", _idom ? _idom->_pre_order : 0, _dom_depth); |
kvn@3049 | 337 | st->print(" RegPressure: %d",_reg_pressure); |
kvn@3049 | 338 | st->print(" IHRP Index: %d",_ihrp_index); |
kvn@3049 | 339 | st->print(" FRegPressure: %d",_freg_pressure); |
kvn@3049 | 340 | st->print(" FHRP Index: %d",_fhrp_index); |
duke@435 | 341 | } |
drchase@6680 | 342 | st->cr(); |
duke@435 | 343 | } |
duke@435 | 344 | |
adlertz@5509 | 345 | void Block::dump() const { |
adlertz@5509 | 346 | dump(NULL); |
adlertz@5509 | 347 | } |
duke@435 | 348 | |
adlertz@5509 | 349 | void Block::dump(const PhaseCFG* cfg) const { |
adlertz@5509 | 350 | dump_head(cfg); |
adlertz@5635 | 351 | for (uint i=0; i< number_of_nodes(); i++) { |
adlertz@5635 | 352 | get_node(i)->dump(); |
adlertz@5509 | 353 | } |
duke@435 | 354 | tty->print("\n"); |
duke@435 | 355 | } |
duke@435 | 356 | #endif |
duke@435 | 357 | |
adlertz@5509 | 358 | PhaseCFG::PhaseCFG(Arena* arena, RootNode* root, Matcher& matcher) |
adlertz@5509 | 359 | : Phase(CFG) |
adlertz@5509 | 360 | , _block_arena(arena) |
adlertz@5539 | 361 | , _root(root) |
adlertz@5539 | 362 | , _matcher(matcher) |
adlertz@5509 | 363 | , _node_to_block_mapping(arena) |
adlertz@5509 | 364 | , _node_latency(NULL) |
duke@435 | 365 | #ifndef PRODUCT |
adlertz@5509 | 366 | , _trace_opto_pipelining(TraceOptoPipelining || C->method_has_option("TraceOptoPipelining")) |
duke@435 | 367 | #endif |
kvn@1268 | 368 | #ifdef ASSERT |
adlertz@5509 | 369 | , _raw_oops(arena) |
kvn@1268 | 370 | #endif |
duke@435 | 371 | { |
duke@435 | 372 | ResourceMark rm; |
duke@435 | 373 | // I'll need a few machine-specific GotoNodes. Make an Ideal GotoNode, |
duke@435 | 374 | // then Match it into a machine-specific Node. Then clone the machine |
duke@435 | 375 | // Node on demand. |
kvn@4115 | 376 | Node *x = new (C) GotoNode(NULL); |
duke@435 | 377 | x->init_req(0, x); |
adlertz@5509 | 378 | _goto = matcher.match_tree(x); |
duke@435 | 379 | assert(_goto != NULL, ""); |
duke@435 | 380 | _goto->set_req(0,_goto); |
duke@435 | 381 | |
duke@435 | 382 | // Build the CFG in Reverse Post Order |
adlertz@5539 | 383 | _number_of_blocks = build_cfg(); |
adlertz@5539 | 384 | _root_block = get_block_for_node(_root); |
duke@435 | 385 | } |
duke@435 | 386 | |
duke@435 | 387 | // Build a proper looking CFG. Make every block begin with either a StartNode |
duke@435 | 388 | // or a RegionNode. Make every block end with either a Goto, If or Return. |
duke@435 | 389 | // The RootNode both starts and ends it's own block. Do this with a recursive |
duke@435 | 390 | // backwards walk over the control edges. |
duke@435 | 391 | uint PhaseCFG::build_cfg() { |
duke@435 | 392 | Arena *a = Thread::current()->resource_area(); |
duke@435 | 393 | VectorSet visited(a); |
duke@435 | 394 | |
duke@435 | 395 | // Allocate stack with enough space to avoid frequent realloc |
duke@435 | 396 | Node_Stack nstack(a, C->unique() >> 1); |
duke@435 | 397 | nstack.push(_root, 0); |
duke@435 | 398 | uint sum = 0; // Counter for blocks |
duke@435 | 399 | |
duke@435 | 400 | while (nstack.is_nonempty()) { |
duke@435 | 401 | // node and in's index from stack's top |
duke@435 | 402 | // 'np' is _root (see above) or RegionNode, StartNode: we push on stack |
duke@435 | 403 | // only nodes which point to the start of basic block (see below). |
duke@435 | 404 | Node *np = nstack.node(); |
duke@435 | 405 | // idx > 0, except for the first node (_root) pushed on stack |
duke@435 | 406 | // at the beginning when idx == 0. |
duke@435 | 407 | // We will use the condition (idx == 0) later to end the build. |
duke@435 | 408 | uint idx = nstack.index(); |
duke@435 | 409 | Node *proj = np->in(idx); |
duke@435 | 410 | const Node *x = proj->is_block_proj(); |
duke@435 | 411 | // Does the block end with a proper block-ending Node? One of Return, |
duke@435 | 412 | // If or Goto? (This check should be done for visited nodes also). |
duke@435 | 413 | if (x == NULL) { // Does not end right... |
duke@435 | 414 | Node *g = _goto->clone(); // Force it to end in a Goto |
duke@435 | 415 | g->set_req(0, proj); |
duke@435 | 416 | np->set_req(idx, g); |
duke@435 | 417 | x = proj = g; |
duke@435 | 418 | } |
duke@435 | 419 | if (!visited.test_set(x->_idx)) { // Visit this block once |
duke@435 | 420 | // Skip any control-pinned middle'in stuff |
duke@435 | 421 | Node *p = proj; |
duke@435 | 422 | do { |
duke@435 | 423 | proj = p; // Update pointer to last Control |
duke@435 | 424 | p = p->in(0); // Move control forward |
duke@435 | 425 | } while( !p->is_block_proj() && |
duke@435 | 426 | !p->is_block_start() ); |
duke@435 | 427 | // Make the block begin with one of Region or StartNode. |
duke@435 | 428 | if( !p->is_block_start() ) { |
kvn@4115 | 429 | RegionNode *r = new (C) RegionNode( 2 ); |
duke@435 | 430 | r->init_req(1, p); // Insert RegionNode in the way |
duke@435 | 431 | proj->set_req(0, r); // Insert RegionNode in the way |
duke@435 | 432 | p = r; |
duke@435 | 433 | } |
duke@435 | 434 | // 'p' now points to the start of this basic block |
duke@435 | 435 | |
duke@435 | 436 | // Put self in array of basic blocks |
adlertz@5509 | 437 | Block *bb = new (_block_arena) Block(_block_arena, p); |
adlertz@5509 | 438 | map_node_to_block(p, bb); |
adlertz@5509 | 439 | map_node_to_block(x, bb); |
kvn@3049 | 440 | if( x != p ) { // Only for root is x == p |
adlertz@5635 | 441 | bb->push_node((Node*)x); |
kvn@3049 | 442 | } |
duke@435 | 443 | // Now handle predecessors |
duke@435 | 444 | ++sum; // Count 1 for self block |
duke@435 | 445 | uint cnt = bb->num_preds(); |
duke@435 | 446 | for (int i = (cnt - 1); i > 0; i-- ) { // For all predecessors |
duke@435 | 447 | Node *prevproj = p->in(i); // Get prior input |
duke@435 | 448 | assert( !prevproj->is_Con(), "dead input not removed" ); |
duke@435 | 449 | // Check to see if p->in(i) is a "control-dependent" CFG edge - |
duke@435 | 450 | // i.e., it splits at the source (via an IF or SWITCH) and merges |
duke@435 | 451 | // at the destination (via a many-input Region). |
duke@435 | 452 | // This breaks critical edges. The RegionNode to start the block |
duke@435 | 453 | // will be added when <p,i> is pulled off the node stack |
duke@435 | 454 | if ( cnt > 2 ) { // Merging many things? |
duke@435 | 455 | assert( prevproj== bb->pred(i),""); |
duke@435 | 456 | if(prevproj->is_block_proj() != prevproj) { // Control-dependent edge? |
duke@435 | 457 | // Force a block on the control-dependent edge |
duke@435 | 458 | Node *g = _goto->clone(); // Force it to end in a Goto |
duke@435 | 459 | g->set_req(0,prevproj); |
duke@435 | 460 | p->set_req(i,g); |
duke@435 | 461 | } |
duke@435 | 462 | } |
duke@435 | 463 | nstack.push(p, i); // 'p' is RegionNode or StartNode |
duke@435 | 464 | } |
duke@435 | 465 | } else { // Post-processing visited nodes |
duke@435 | 466 | nstack.pop(); // remove node from stack |
duke@435 | 467 | // Check if it the fist node pushed on stack at the beginning. |
duke@435 | 468 | if (idx == 0) break; // end of the build |
duke@435 | 469 | // Find predecessor basic block |
adlertz@5509 | 470 | Block *pb = get_block_for_node(x); |
duke@435 | 471 | // Insert into nodes array, if not already there |
adlertz@5509 | 472 | if (!has_block(proj)) { |
duke@435 | 473 | assert( x != proj, "" ); |
duke@435 | 474 | // Map basic block of projection |
adlertz@5509 | 475 | map_node_to_block(proj, pb); |
adlertz@5635 | 476 | pb->push_node(proj); |
duke@435 | 477 | } |
duke@435 | 478 | // Insert self as a child of my predecessor block |
adlertz@5509 | 479 | pb->_succs.map(pb->_num_succs++, get_block_for_node(np)); |
adlertz@5635 | 480 | assert( pb->get_node(pb->number_of_nodes() - pb->_num_succs)->is_block_proj(), |
duke@435 | 481 | "too many control users, not a CFG?" ); |
duke@435 | 482 | } |
duke@435 | 483 | } |
duke@435 | 484 | // Return number of basic blocks for all children and self |
duke@435 | 485 | return sum; |
duke@435 | 486 | } |
duke@435 | 487 | |
duke@435 | 488 | // Inserts a goto & corresponding basic block between |
duke@435 | 489 | // block[block_no] and its succ_no'th successor block |
duke@435 | 490 | void PhaseCFG::insert_goto_at(uint block_no, uint succ_no) { |
duke@435 | 491 | // get block with block_no |
adlertz@5539 | 492 | assert(block_no < number_of_blocks(), "illegal block number"); |
adlertz@5539 | 493 | Block* in = get_block(block_no); |
duke@435 | 494 | // get successor block succ_no |
duke@435 | 495 | assert(succ_no < in->_num_succs, "illegal successor number"); |
duke@435 | 496 | Block* out = in->_succs[succ_no]; |
rasbold@743 | 497 | // Compute frequency of the new block. Do this before inserting |
rasbold@743 | 498 | // new block in case succ_prob() needs to infer the probability from |
rasbold@743 | 499 | // surrounding blocks. |
rasbold@743 | 500 | float freq = in->_freq * in->succ_prob(succ_no); |
duke@435 | 501 | // get ProjNode corresponding to the succ_no'th successor of the in block |
adlertz@5635 | 502 | ProjNode* proj = in->get_node(in->number_of_nodes() - in->_num_succs + succ_no)->as_Proj(); |
duke@435 | 503 | // create region for basic block |
kvn@4115 | 504 | RegionNode* region = new (C) RegionNode(2); |
duke@435 | 505 | region->init_req(1, proj); |
duke@435 | 506 | // setup corresponding basic block |
adlertz@5509 | 507 | Block* block = new (_block_arena) Block(_block_arena, region); |
adlertz@5509 | 508 | map_node_to_block(region, block); |
duke@435 | 509 | C->regalloc()->set_bad(region->_idx); |
duke@435 | 510 | // add a goto node |
duke@435 | 511 | Node* gto = _goto->clone(); // get a new goto node |
duke@435 | 512 | gto->set_req(0, region); |
duke@435 | 513 | // add it to the basic block |
adlertz@5635 | 514 | block->push_node(gto); |
adlertz@5509 | 515 | map_node_to_block(gto, block); |
duke@435 | 516 | C->regalloc()->set_bad(gto->_idx); |
duke@435 | 517 | // hook up successor block |
duke@435 | 518 | block->_succs.map(block->_num_succs++, out); |
duke@435 | 519 | // remap successor's predecessors if necessary |
duke@435 | 520 | for (uint i = 1; i < out->num_preds(); i++) { |
duke@435 | 521 | if (out->pred(i) == proj) out->head()->set_req(i, gto); |
duke@435 | 522 | } |
duke@435 | 523 | // remap predecessor's successor to new block |
duke@435 | 524 | in->_succs.map(succ_no, block); |
rasbold@743 | 525 | // Set the frequency of the new block |
rasbold@743 | 526 | block->_freq = freq; |
duke@435 | 527 | // add new basic block to basic block list |
adlertz@5539 | 528 | add_block_at(block_no + 1, block); |
duke@435 | 529 | } |
duke@435 | 530 | |
duke@435 | 531 | // Does this block end in a multiway branch that cannot have the default case |
duke@435 | 532 | // flipped for another case? |
goetz@6490 | 533 | static bool no_flip_branch(Block *b) { |
adlertz@5635 | 534 | int branch_idx = b->number_of_nodes() - b->_num_succs-1; |
goetz@6490 | 535 | if (branch_idx < 1) { |
goetz@6490 | 536 | return false; |
goetz@6490 | 537 | } |
goetz@6490 | 538 | Node *branch = b->get_node(branch_idx); |
goetz@6490 | 539 | if (branch->is_Catch()) { |
rasbold@853 | 540 | return true; |
goetz@6490 | 541 | } |
goetz@6490 | 542 | if (branch->is_Mach()) { |
goetz@6490 | 543 | if (branch->is_MachNullCheck()) { |
rasbold@853 | 544 | return true; |
goetz@6490 | 545 | } |
goetz@6490 | 546 | int iop = branch->as_Mach()->ideal_Opcode(); |
goetz@6490 | 547 | if (iop == Op_FastLock || iop == Op_FastUnlock) { |
duke@435 | 548 | return true; |
goetz@6490 | 549 | } |
goetz@6490 | 550 | // Don't flip if branch has an implicit check. |
goetz@6490 | 551 | if (branch->as_Mach()->is_TrapBasedCheckNode()) { |
goetz@6490 | 552 | return true; |
goetz@6490 | 553 | } |
duke@435 | 554 | } |
duke@435 | 555 | return false; |
duke@435 | 556 | } |
duke@435 | 557 | |
duke@435 | 558 | // Check for NeverBranch at block end. This needs to become a GOTO to the |
duke@435 | 559 | // true target. NeverBranch are treated as a conditional branch that always |
duke@435 | 560 | // goes the same direction for most of the optimizer and are used to give a |
duke@435 | 561 | // fake exit path to infinite loops. At this late stage they need to turn |
duke@435 | 562 | // into Goto's so that when you enter the infinite loop you indeed hang. |
duke@435 | 563 | void PhaseCFG::convert_NeverBranch_to_Goto(Block *b) { |
duke@435 | 564 | // Find true target |
duke@435 | 565 | int end_idx = b->end_idx(); |
adlertz@5635 | 566 | int idx = b->get_node(end_idx+1)->as_Proj()->_con; |
duke@435 | 567 | Block *succ = b->_succs[idx]; |
duke@435 | 568 | Node* gto = _goto->clone(); // get a new goto node |
duke@435 | 569 | gto->set_req(0, b->head()); |
adlertz@5635 | 570 | Node *bp = b->get_node(end_idx); |
adlertz@5635 | 571 | b->map_node(gto, end_idx); // Slam over NeverBranch |
adlertz@5509 | 572 | map_node_to_block(gto, b); |
duke@435 | 573 | C->regalloc()->set_bad(gto->_idx); |
adlertz@5635 | 574 | b->pop_node(); // Yank projections |
adlertz@5635 | 575 | b->pop_node(); // Yank projections |
duke@435 | 576 | b->_succs.map(0,succ); // Map only successor |
duke@435 | 577 | b->_num_succs = 1; |
duke@435 | 578 | // remap successor's predecessors if necessary |
duke@435 | 579 | uint j; |
duke@435 | 580 | for( j = 1; j < succ->num_preds(); j++) |
duke@435 | 581 | if( succ->pred(j)->in(0) == bp ) |
duke@435 | 582 | succ->head()->set_req(j, gto); |
duke@435 | 583 | // Kill alternate exit path |
duke@435 | 584 | Block *dead = b->_succs[1-idx]; |
duke@435 | 585 | for( j = 1; j < dead->num_preds(); j++) |
duke@435 | 586 | if( dead->pred(j)->in(0) == bp ) |
duke@435 | 587 | break; |
duke@435 | 588 | // Scan through block, yanking dead path from |
duke@435 | 589 | // all regions and phis. |
duke@435 | 590 | dead->head()->del_req(j); |
adlertz@5635 | 591 | for( int k = 1; dead->get_node(k)->is_Phi(); k++ ) |
adlertz@5635 | 592 | dead->get_node(k)->del_req(j); |
duke@435 | 593 | } |
duke@435 | 594 | |
duke@435 | 595 | // Helper function to move block bx to the slot following b_index. Return |
duke@435 | 596 | // true if the move is successful, otherwise false |
rasbold@853 | 597 | bool PhaseCFG::move_to_next(Block* bx, uint b_index) { |
duke@435 | 598 | if (bx == NULL) return false; |
duke@435 | 599 | |
duke@435 | 600 | // Return false if bx is already scheduled. |
duke@435 | 601 | uint bx_index = bx->_pre_order; |
adlertz@5539 | 602 | if ((bx_index <= b_index) && (get_block(bx_index) == bx)) { |
duke@435 | 603 | return false; |
duke@435 | 604 | } |
duke@435 | 605 | |
duke@435 | 606 | // Find the current index of block bx on the block list |
duke@435 | 607 | bx_index = b_index + 1; |
adlertz@5539 | 608 | while (bx_index < number_of_blocks() && get_block(bx_index) != bx) { |
adlertz@5539 | 609 | bx_index++; |
adlertz@5539 | 610 | } |
adlertz@5539 | 611 | assert(get_block(bx_index) == bx, "block not found"); |
duke@435 | 612 | |
duke@435 | 613 | // If the previous block conditionally falls into bx, return false, |
duke@435 | 614 | // because moving bx will create an extra jump. |
duke@435 | 615 | for(uint k = 1; k < bx->num_preds(); k++ ) { |
adlertz@5509 | 616 | Block* pred = get_block_for_node(bx->pred(k)); |
adlertz@5539 | 617 | if (pred == get_block(bx_index - 1)) { |
duke@435 | 618 | if (pred->_num_succs != 1) { |
duke@435 | 619 | return false; |
duke@435 | 620 | } |
duke@435 | 621 | } |
duke@435 | 622 | } |
duke@435 | 623 | |
duke@435 | 624 | // Reinsert bx just past block 'b' |
duke@435 | 625 | _blocks.remove(bx_index); |
duke@435 | 626 | _blocks.insert(b_index + 1, bx); |
duke@435 | 627 | return true; |
duke@435 | 628 | } |
duke@435 | 629 | |
duke@435 | 630 | // Move empty and uncommon blocks to the end. |
rasbold@853 | 631 | void PhaseCFG::move_to_end(Block *b, uint i) { |
duke@435 | 632 | int e = b->is_Empty(); |
duke@435 | 633 | if (e != Block::not_empty) { |
duke@435 | 634 | if (e == Block::empty_with_goto) { |
duke@435 | 635 | // Remove the goto, but leave the block. |
adlertz@5635 | 636 | b->pop_node(); |
duke@435 | 637 | } |
duke@435 | 638 | // Mark this block as a connector block, which will cause it to be |
duke@435 | 639 | // ignored in certain functions such as non_connector_successor(). |
duke@435 | 640 | b->set_connector(); |
duke@435 | 641 | } |
duke@435 | 642 | // Move the empty block to the end, and don't recheck. |
duke@435 | 643 | _blocks.remove(i); |
duke@435 | 644 | _blocks.push(b); |
duke@435 | 645 | } |
duke@435 | 646 | |
rasbold@853 | 647 | // Set loop alignment for every block |
rasbold@853 | 648 | void PhaseCFG::set_loop_alignment() { |
adlertz@5539 | 649 | uint last = number_of_blocks(); |
adlertz@5539 | 650 | assert(get_block(0) == get_root_block(), ""); |
rasbold@853 | 651 | |
adlertz@5539 | 652 | for (uint i = 1; i < last; i++) { |
adlertz@5539 | 653 | Block* block = get_block(i); |
adlertz@5539 | 654 | if (block->head()->is_Loop()) { |
adlertz@5539 | 655 | block->set_loop_alignment(block); |
rasbold@853 | 656 | } |
rasbold@853 | 657 | } |
rasbold@853 | 658 | } |
rasbold@853 | 659 | |
rasbold@853 | 660 | // Make empty basic blocks to be "connector" blocks, Move uncommon blocks |
rasbold@853 | 661 | // to the end. |
adlertz@5539 | 662 | void PhaseCFG::remove_empty_blocks() { |
duke@435 | 663 | // Move uncommon blocks to the end |
adlertz@5539 | 664 | uint last = number_of_blocks(); |
adlertz@5539 | 665 | assert(get_block(0) == get_root_block(), ""); |
rasbold@853 | 666 | |
rasbold@853 | 667 | for (uint i = 1; i < last; i++) { |
adlertz@5539 | 668 | Block* block = get_block(i); |
adlertz@5539 | 669 | if (block->is_connector()) { |
adlertz@5539 | 670 | break; |
adlertz@5539 | 671 | } |
duke@435 | 672 | |
duke@435 | 673 | // Check for NeverBranch at block end. This needs to become a GOTO to the |
duke@435 | 674 | // true target. NeverBranch are treated as a conditional branch that |
duke@435 | 675 | // always goes the same direction for most of the optimizer and are used |
duke@435 | 676 | // to give a fake exit path to infinite loops. At this late stage they |
duke@435 | 677 | // need to turn into Goto's so that when you enter the infinite loop you |
duke@435 | 678 | // indeed hang. |
adlertz@5635 | 679 | if (block->get_node(block->end_idx())->Opcode() == Op_NeverBranch) { |
adlertz@5539 | 680 | convert_NeverBranch_to_Goto(block); |
adlertz@5539 | 681 | } |
duke@435 | 682 | |
duke@435 | 683 | // Look for uncommon blocks and move to end. |
rasbold@853 | 684 | if (!C->do_freq_based_layout()) { |
adlertz@5639 | 685 | if (is_uncommon(block)) { |
adlertz@5539 | 686 | move_to_end(block, i); |
rasbold@853 | 687 | last--; // No longer check for being uncommon! |
adlertz@5539 | 688 | if (no_flip_branch(block)) { // Fall-thru case must follow? |
adlertz@5539 | 689 | // Find the fall-thru block |
adlertz@5539 | 690 | block = get_block(i); |
adlertz@5539 | 691 | move_to_end(block, i); |
rasbold@853 | 692 | last--; |
rasbold@853 | 693 | } |
adlertz@5539 | 694 | // backup block counter post-increment |
adlertz@5539 | 695 | i--; |
duke@435 | 696 | } |
duke@435 | 697 | } |
duke@435 | 698 | } |
duke@435 | 699 | |
rasbold@853 | 700 | // Move empty blocks to the end |
adlertz@5539 | 701 | last = number_of_blocks(); |
rasbold@853 | 702 | for (uint i = 1; i < last; i++) { |
adlertz@5539 | 703 | Block* block = get_block(i); |
adlertz@5539 | 704 | if (block->is_Empty() != Block::not_empty) { |
adlertz@5539 | 705 | move_to_end(block, i); |
rasbold@853 | 706 | last--; |
rasbold@853 | 707 | i--; |
duke@435 | 708 | } |
duke@435 | 709 | } // End of for all blocks |
rasbold@853 | 710 | } |
duke@435 | 711 | |
goetz@6490 | 712 | Block *PhaseCFG::fixup_trap_based_check(Node *branch, Block *block, int block_pos, Block *bnext) { |
goetz@6490 | 713 | // Trap based checks must fall through to the successor with |
goetz@6490 | 714 | // PROB_ALWAYS. |
goetz@6490 | 715 | // They should be an If with 2 successors. |
goetz@6490 | 716 | assert(branch->is_MachIf(), "must be If"); |
goetz@6490 | 717 | assert(block->_num_succs == 2, "must have 2 successors"); |
goetz@6490 | 718 | |
goetz@6490 | 719 | // Get the If node and the projection for the first successor. |
goetz@6490 | 720 | MachIfNode *iff = block->get_node(block->number_of_nodes()-3)->as_MachIf(); |
goetz@6490 | 721 | ProjNode *proj0 = block->get_node(block->number_of_nodes()-2)->as_Proj(); |
goetz@6490 | 722 | ProjNode *proj1 = block->get_node(block->number_of_nodes()-1)->as_Proj(); |
goetz@6490 | 723 | ProjNode *projt = (proj0->Opcode() == Op_IfTrue) ? proj0 : proj1; |
goetz@6490 | 724 | ProjNode *projf = (proj0->Opcode() == Op_IfFalse) ? proj0 : proj1; |
goetz@6490 | 725 | |
goetz@6490 | 726 | // Assert that proj0 and succs[0] match up. Similarly for proj1 and succs[1]. |
goetz@6490 | 727 | assert(proj0->raw_out(0) == block->_succs[0]->head(), "Mismatch successor 0"); |
goetz@6490 | 728 | assert(proj1->raw_out(0) == block->_succs[1]->head(), "Mismatch successor 1"); |
goetz@6490 | 729 | |
goetz@6490 | 730 | ProjNode *proj_always; |
goetz@6490 | 731 | ProjNode *proj_never; |
goetz@6490 | 732 | // We must negate the branch if the implicit check doesn't follow |
goetz@6490 | 733 | // the branch's TRUE path. Then, the new TRUE branch target will |
goetz@6490 | 734 | // be the old FALSE branch target. |
goetz@6490 | 735 | if (iff->_prob <= 2*PROB_NEVER) { // There are small rounding errors. |
goetz@6490 | 736 | proj_never = projt; |
goetz@6490 | 737 | proj_always = projf; |
goetz@6490 | 738 | } else { |
goetz@6490 | 739 | // We must negate the branch if the trap doesn't follow the |
goetz@6490 | 740 | // branch's TRUE path. Then, the new TRUE branch target will |
goetz@6490 | 741 | // be the old FALSE branch target. |
goetz@6490 | 742 | proj_never = projf; |
goetz@6490 | 743 | proj_always = projt; |
goetz@6490 | 744 | iff->negate(); |
goetz@6490 | 745 | } |
goetz@6490 | 746 | assert(iff->_prob <= 2*PROB_NEVER, "Trap based checks are expected to trap never!"); |
goetz@6490 | 747 | // Map the successors properly |
goetz@6490 | 748 | block->_succs.map(0, get_block_for_node(proj_never ->raw_out(0))); // The target of the trap. |
goetz@6490 | 749 | block->_succs.map(1, get_block_for_node(proj_always->raw_out(0))); // The fall through target. |
goetz@6490 | 750 | |
goetz@6500 | 751 | if (block->get_node(block->number_of_nodes() - block->_num_succs + 1) != proj_always) { |
goetz@6500 | 752 | block->map_node(proj_never, block->number_of_nodes() - block->_num_succs + 0); |
goetz@6500 | 753 | block->map_node(proj_always, block->number_of_nodes() - block->_num_succs + 1); |
goetz@6500 | 754 | } |
goetz@6500 | 755 | |
goetz@6490 | 756 | // Place the fall through block after this block. |
goetz@6490 | 757 | Block *bs1 = block->non_connector_successor(1); |
goetz@6490 | 758 | if (bs1 != bnext && move_to_next(bs1, block_pos)) { |
goetz@6490 | 759 | bnext = bs1; |
goetz@6490 | 760 | } |
goetz@6490 | 761 | // If the fall through block still is not the next block, insert a goto. |
goetz@6490 | 762 | if (bs1 != bnext) { |
goetz@6490 | 763 | insert_goto_at(block_pos, 1); |
goetz@6490 | 764 | } |
goetz@6490 | 765 | return bnext; |
goetz@6490 | 766 | } |
goetz@6490 | 767 | |
rasbold@853 | 768 | // Fix up the final control flow for basic blocks. |
rasbold@853 | 769 | void PhaseCFG::fixup_flow() { |
duke@435 | 770 | // Fixup final control flow for the blocks. Remove jump-to-next |
goetz@6478 | 771 | // block. If neither arm of an IF follows the conditional branch, we |
duke@435 | 772 | // have to add a second jump after the conditional. We place the |
duke@435 | 773 | // TRUE branch target in succs[0] for both GOTOs and IFs. |
adlertz@5539 | 774 | for (uint i = 0; i < number_of_blocks(); i++) { |
adlertz@5539 | 775 | Block* block = get_block(i); |
adlertz@5539 | 776 | block->_pre_order = i; // turn pre-order into block-index |
duke@435 | 777 | |
duke@435 | 778 | // Connector blocks need no further processing. |
adlertz@5539 | 779 | if (block->is_connector()) { |
adlertz@5539 | 780 | assert((i+1) == number_of_blocks() || get_block(i + 1)->is_connector(), "All connector blocks should sink to the end"); |
duke@435 | 781 | continue; |
duke@435 | 782 | } |
adlertz@5539 | 783 | assert(block->is_Empty() != Block::completely_empty, "Empty blocks should be connectors"); |
duke@435 | 784 | |
adlertz@5539 | 785 | Block* bnext = (i < number_of_blocks() - 1) ? get_block(i + 1) : NULL; |
adlertz@5539 | 786 | Block* bs0 = block->non_connector_successor(0); |
duke@435 | 787 | |
duke@435 | 788 | // Check for multi-way branches where I cannot negate the test to |
duke@435 | 789 | // exchange the true and false targets. |
adlertz@5539 | 790 | if (no_flip_branch(block)) { |
goetz@6490 | 791 | // Find fall through case - if must fall into its target. |
goetz@6490 | 792 | // Get the index of the branch's first successor. |
adlertz@5635 | 793 | int branch_idx = block->number_of_nodes() - block->_num_succs; |
goetz@6490 | 794 | |
goetz@6490 | 795 | // The branch is 1 before the branch's first successor. |
goetz@6490 | 796 | Node *branch = block->get_node(branch_idx-1); |
goetz@6490 | 797 | |
goetz@6490 | 798 | // Handle no-flip branches which have implicit checks and which require |
goetz@6490 | 799 | // special block ordering and individual semantics of the 'fall through |
goetz@6490 | 800 | // case'. |
goetz@6490 | 801 | if ((TrapBasedNullChecks || TrapBasedRangeChecks) && |
goetz@6490 | 802 | branch->is_Mach() && branch->as_Mach()->is_TrapBasedCheckNode()) { |
goetz@6490 | 803 | bnext = fixup_trap_based_check(branch, block, i, bnext); |
goetz@6490 | 804 | } else { |
goetz@6490 | 805 | // Else, default handling for no-flip branches |
goetz@6490 | 806 | for (uint j2 = 0; j2 < block->_num_succs; j2++) { |
goetz@6490 | 807 | const ProjNode* p = block->get_node(branch_idx + j2)->as_Proj(); |
goetz@6490 | 808 | if (p->_con == 0) { |
goetz@6490 | 809 | // successor j2 is fall through case |
goetz@6490 | 810 | if (block->non_connector_successor(j2) != bnext) { |
goetz@6490 | 811 | // but it is not the next block => insert a goto |
goetz@6490 | 812 | insert_goto_at(i, j2); |
goetz@6490 | 813 | } |
goetz@6490 | 814 | // Put taken branch in slot 0 |
goetz@6490 | 815 | if (j2 == 0 && block->_num_succs == 2) { |
goetz@6490 | 816 | // Flip targets in succs map |
goetz@6490 | 817 | Block *tbs0 = block->_succs[0]; |
goetz@6490 | 818 | Block *tbs1 = block->_succs[1]; |
goetz@6490 | 819 | block->_succs.map(0, tbs1); |
goetz@6490 | 820 | block->_succs.map(1, tbs0); |
goetz@6490 | 821 | } |
goetz@6490 | 822 | break; |
duke@435 | 823 | } |
duke@435 | 824 | } |
duke@435 | 825 | } |
adlertz@5539 | 826 | |
duke@435 | 827 | // Remove all CatchProjs |
adlertz@5539 | 828 | for (uint j = 0; j < block->_num_succs; j++) { |
adlertz@5635 | 829 | block->pop_node(); |
adlertz@5539 | 830 | } |
duke@435 | 831 | |
adlertz@5539 | 832 | } else if (block->_num_succs == 1) { |
duke@435 | 833 | // Block ends in a Goto? |
duke@435 | 834 | if (bnext == bs0) { |
duke@435 | 835 | // We fall into next block; remove the Goto |
adlertz@5635 | 836 | block->pop_node(); |
duke@435 | 837 | } |
duke@435 | 838 | |
adlertz@5539 | 839 | } else if(block->_num_succs == 2) { // Block ends in a If? |
duke@435 | 840 | // Get opcode of 1st projection (matches _succs[0]) |
duke@435 | 841 | // Note: Since this basic block has 2 exits, the last 2 nodes must |
duke@435 | 842 | // be projections (in any order), the 3rd last node must be |
duke@435 | 843 | // the IfNode (we have excluded other 2-way exits such as |
duke@435 | 844 | // CatchNodes already). |
adlertz@5635 | 845 | MachNode* iff = block->get_node(block->number_of_nodes() - 3)->as_Mach(); |
adlertz@5635 | 846 | ProjNode* proj0 = block->get_node(block->number_of_nodes() - 2)->as_Proj(); |
adlertz@5635 | 847 | ProjNode* proj1 = block->get_node(block->number_of_nodes() - 1)->as_Proj(); |
duke@435 | 848 | |
duke@435 | 849 | // Assert that proj0 and succs[0] match up. Similarly for proj1 and succs[1]. |
adlertz@5539 | 850 | assert(proj0->raw_out(0) == block->_succs[0]->head(), "Mismatch successor 0"); |
adlertz@5539 | 851 | assert(proj1->raw_out(0) == block->_succs[1]->head(), "Mismatch successor 1"); |
duke@435 | 852 | |
adlertz@5539 | 853 | Block* bs1 = block->non_connector_successor(1); |
duke@435 | 854 | |
duke@435 | 855 | // Check for neither successor block following the current |
duke@435 | 856 | // block ending in a conditional. If so, move one of the |
duke@435 | 857 | // successors after the current one, provided that the |
duke@435 | 858 | // successor was previously unscheduled, but moveable |
duke@435 | 859 | // (i.e., all paths to it involve a branch). |
adlertz@5539 | 860 | if (!C->do_freq_based_layout() && bnext != bs0 && bnext != bs1) { |
duke@435 | 861 | // Choose the more common successor based on the probability |
duke@435 | 862 | // of the conditional branch. |
adlertz@5539 | 863 | Block* bx = bs0; |
adlertz@5539 | 864 | Block* by = bs1; |
duke@435 | 865 | |
duke@435 | 866 | // _prob is the probability of taking the true path. Make |
duke@435 | 867 | // p the probability of taking successor #1. |
duke@435 | 868 | float p = iff->as_MachIf()->_prob; |
adlertz@5539 | 869 | if (proj0->Opcode() == Op_IfTrue) { |
duke@435 | 870 | p = 1.0 - p; |
duke@435 | 871 | } |
duke@435 | 872 | |
duke@435 | 873 | // Prefer successor #1 if p > 0.5 |
duke@435 | 874 | if (p > PROB_FAIR) { |
duke@435 | 875 | bx = bs1; |
duke@435 | 876 | by = bs0; |
duke@435 | 877 | } |
duke@435 | 878 | |
duke@435 | 879 | // Attempt the more common successor first |
rasbold@853 | 880 | if (move_to_next(bx, i)) { |
duke@435 | 881 | bnext = bx; |
rasbold@853 | 882 | } else if (move_to_next(by, i)) { |
duke@435 | 883 | bnext = by; |
duke@435 | 884 | } |
duke@435 | 885 | } |
duke@435 | 886 | |
duke@435 | 887 | // Check for conditional branching the wrong way. Negate |
duke@435 | 888 | // conditional, if needed, so it falls into the following block |
duke@435 | 889 | // and branches to the not-following block. |
duke@435 | 890 | |
duke@435 | 891 | // Check for the next block being in succs[0]. We are going to branch |
duke@435 | 892 | // to succs[0], so we want the fall-thru case as the next block in |
duke@435 | 893 | // succs[1]. |
duke@435 | 894 | if (bnext == bs0) { |
duke@435 | 895 | // Fall-thru case in succs[0], so flip targets in succs map |
adlertz@5539 | 896 | Block* tbs0 = block->_succs[0]; |
adlertz@5539 | 897 | Block* tbs1 = block->_succs[1]; |
adlertz@5539 | 898 | block->_succs.map(0, tbs1); |
adlertz@5539 | 899 | block->_succs.map(1, tbs0); |
duke@435 | 900 | // Flip projection for each target |
adlertz@5539 | 901 | ProjNode* tmp = proj0; |
adlertz@5539 | 902 | proj0 = proj1; |
adlertz@5539 | 903 | proj1 = tmp; |
duke@435 | 904 | |
adlertz@5539 | 905 | } else if(bnext != bs1) { |
rasbold@853 | 906 | // Need a double-branch |
duke@435 | 907 | // The existing conditional branch need not change. |
duke@435 | 908 | // Add a unconditional branch to the false target. |
duke@435 | 909 | // Alas, it must appear in its own block and adding a |
duke@435 | 910 | // block this late in the game is complicated. Sigh. |
duke@435 | 911 | insert_goto_at(i, 1); |
duke@435 | 912 | } |
duke@435 | 913 | |
duke@435 | 914 | // Make sure we TRUE branch to the target |
adlertz@5539 | 915 | if (proj0->Opcode() == Op_IfFalse) { |
kvn@3051 | 916 | iff->as_MachIf()->negate(); |
rasbold@853 | 917 | } |
duke@435 | 918 | |
adlertz@5635 | 919 | block->pop_node(); // Remove IfFalse & IfTrue projections |
adlertz@5635 | 920 | block->pop_node(); |
duke@435 | 921 | |
duke@435 | 922 | } else { |
duke@435 | 923 | // Multi-exit block, e.g. a switch statement |
duke@435 | 924 | // But we don't need to do anything here |
duke@435 | 925 | } |
duke@435 | 926 | } // End of for all blocks |
duke@435 | 927 | } |
duke@435 | 928 | |
duke@435 | 929 | |
goetz@6478 | 930 | // postalloc_expand: Expand nodes after register allocation. |
goetz@6478 | 931 | // |
goetz@6478 | 932 | // postalloc_expand has to be called after register allocation, just |
goetz@6478 | 933 | // before output (i.e. scheduling). It only gets called if |
goetz@6478 | 934 | // Matcher::require_postalloc_expand is true. |
goetz@6478 | 935 | // |
goetz@6478 | 936 | // Background: |
goetz@6478 | 937 | // |
goetz@6478 | 938 | // Nodes that are expandend (one compound node requiring several |
goetz@6478 | 939 | // assembler instructions to be implemented split into two or more |
goetz@6478 | 940 | // non-compound nodes) after register allocation are not as nice as |
goetz@6478 | 941 | // the ones expanded before register allocation - they don't |
goetz@6478 | 942 | // participate in optimizations as global code motion. But after |
goetz@6478 | 943 | // register allocation we can expand nodes that use registers which |
goetz@6478 | 944 | // are not spillable or registers that are not allocated, because the |
goetz@6478 | 945 | // old compound node is simply replaced (in its location in the basic |
goetz@6478 | 946 | // block) by a new subgraph which does not contain compound nodes any |
goetz@6478 | 947 | // more. The scheduler called during output can later on process these |
goetz@6478 | 948 | // non-compound nodes. |
goetz@6478 | 949 | // |
goetz@6478 | 950 | // Implementation: |
goetz@6478 | 951 | // |
goetz@6478 | 952 | // Nodes requiring postalloc expand are specified in the ad file by using |
goetz@6478 | 953 | // a postalloc_expand statement instead of ins_encode. A postalloc_expand |
goetz@6478 | 954 | // contains a single call to an encoding, as does an ins_encode |
goetz@6478 | 955 | // statement. Instead of an emit() function a postalloc_expand() function |
goetz@6478 | 956 | // is generated that doesn't emit assembler but creates a new |
goetz@6478 | 957 | // subgraph. The code below calls this postalloc_expand function for each |
goetz@6478 | 958 | // node with the appropriate attribute. This function returns the new |
goetz@6478 | 959 | // nodes generated in an array passed in the call. The old node, |
goetz@6478 | 960 | // potential MachTemps before and potential Projs after it then get |
goetz@6478 | 961 | // disconnected and replaced by the new nodes. The instruction |
goetz@6478 | 962 | // generating the result has to be the last one in the array. In |
goetz@6478 | 963 | // general it is assumed that Projs after the node expanded are |
goetz@6478 | 964 | // kills. These kills are not required any more after expanding as |
goetz@6478 | 965 | // there are now explicitly visible def-use chains and the Projs are |
goetz@6478 | 966 | // removed. This does not hold for calls: They do not only have |
goetz@6478 | 967 | // kill-Projs but also Projs defining values. Therefore Projs after |
goetz@6478 | 968 | // the node expanded are removed for all but for calls. If a node is |
goetz@6478 | 969 | // to be reused, it must be added to the nodes list returned, and it |
goetz@6478 | 970 | // will be added again. |
goetz@6478 | 971 | // |
goetz@6478 | 972 | // Implementing the postalloc_expand function for a node in an enc_class |
goetz@6478 | 973 | // is rather tedious. It requires knowledge about many node details, as |
goetz@6478 | 974 | // the nodes and the subgraph must be hand crafted. To simplify this, |
goetz@6478 | 975 | // adlc generates some utility variables into the postalloc_expand function, |
goetz@6478 | 976 | // e.g., holding the operands as specified by the postalloc_expand encoding |
goetz@6478 | 977 | // specification, e.g.: |
goetz@6478 | 978 | // * unsigned idx_<par_name> holding the index of the node in the ins |
goetz@6478 | 979 | // * Node *n_<par_name> holding the node loaded from the ins |
goetz@6478 | 980 | // * MachOpnd *op_<par_name> holding the corresponding operand |
goetz@6478 | 981 | // |
goetz@6478 | 982 | // The ordering of operands can not be determined by looking at a |
goetz@6478 | 983 | // rule. Especially if a match rule matches several different trees, |
goetz@6478 | 984 | // several nodes are generated from one instruct specification with |
goetz@6478 | 985 | // different operand orderings. In this case the adlc generated |
goetz@6478 | 986 | // variables are the only way to access the ins and operands |
goetz@6478 | 987 | // deterministically. |
goetz@6478 | 988 | // |
goetz@6478 | 989 | // If assigning a register to a node that contains an oop, don't |
goetz@6478 | 990 | // forget to call ra_->set_oop() for the node. |
goetz@6478 | 991 | void PhaseCFG::postalloc_expand(PhaseRegAlloc* _ra) { |
goetz@6478 | 992 | GrowableArray <Node *> new_nodes(32); // Array with new nodes filled by postalloc_expand function of node. |
goetz@6478 | 993 | GrowableArray <Node *> remove(32); |
goetz@6478 | 994 | GrowableArray <Node *> succs(32); |
goetz@6478 | 995 | unsigned int max_idx = C->unique(); // Remember to distinguish new from old nodes. |
goetz@6478 | 996 | DEBUG_ONLY(bool foundNode = false); |
goetz@6478 | 997 | |
goetz@6478 | 998 | // for all blocks |
goetz@6478 | 999 | for (uint i = 0; i < number_of_blocks(); i++) { |
goetz@6478 | 1000 | Block *b = _blocks[i]; |
goetz@6478 | 1001 | // For all instructions in the current block. |
goetz@6478 | 1002 | for (uint j = 0; j < b->number_of_nodes(); j++) { |
goetz@6478 | 1003 | Node *n = b->get_node(j); |
goetz@6478 | 1004 | if (n->is_Mach() && n->as_Mach()->requires_postalloc_expand()) { |
goetz@6478 | 1005 | #ifdef ASSERT |
goetz@6478 | 1006 | if (TracePostallocExpand) { |
goetz@6478 | 1007 | if (!foundNode) { |
goetz@6478 | 1008 | foundNode = true; |
goetz@6478 | 1009 | tty->print("POSTALLOC EXPANDING %d %s\n", C->compile_id(), |
goetz@6478 | 1010 | C->method() ? C->method()->name()->as_utf8() : C->stub_name()); |
goetz@6478 | 1011 | } |
goetz@6478 | 1012 | tty->print(" postalloc expanding "); n->dump(); |
goetz@6478 | 1013 | if (Verbose) { |
goetz@6478 | 1014 | tty->print(" with ins:\n"); |
goetz@6478 | 1015 | for (uint k = 0; k < n->len(); ++k) { |
goetz@6478 | 1016 | if (n->in(k)) { tty->print(" "); n->in(k)->dump(); } |
goetz@6478 | 1017 | } |
goetz@6478 | 1018 | } |
goetz@6478 | 1019 | } |
goetz@6478 | 1020 | #endif |
goetz@6478 | 1021 | new_nodes.clear(); |
goetz@6478 | 1022 | // Collect nodes that have to be removed from the block later on. |
goetz@6478 | 1023 | uint req = n->req(); |
goetz@6478 | 1024 | remove.clear(); |
goetz@6478 | 1025 | for (uint k = 0; k < req; ++k) { |
goetz@6478 | 1026 | if (n->in(k) && n->in(k)->is_MachTemp()) { |
goetz@6478 | 1027 | remove.push(n->in(k)); // MachTemps which are inputs to the old node have to be removed. |
goetz@6478 | 1028 | n->in(k)->del_req(0); |
goetz@6478 | 1029 | j--; |
goetz@6478 | 1030 | } |
goetz@6478 | 1031 | } |
goetz@6478 | 1032 | |
goetz@6478 | 1033 | // Check whether we can allocate enough nodes. We set a fix limit for |
goetz@6478 | 1034 | // the size of postalloc expands with this. |
goetz@6478 | 1035 | uint unique_limit = C->unique() + 40; |
goetz@6478 | 1036 | if (unique_limit >= _ra->node_regs_max_index()) { |
goetz@6478 | 1037 | Compile::current()->record_failure("out of nodes in postalloc expand"); |
goetz@6478 | 1038 | return; |
goetz@6478 | 1039 | } |
goetz@6478 | 1040 | |
goetz@6478 | 1041 | // Emit (i.e. generate new nodes). |
goetz@6478 | 1042 | n->as_Mach()->postalloc_expand(&new_nodes, _ra); |
goetz@6478 | 1043 | |
goetz@6478 | 1044 | assert(C->unique() < unique_limit, "You allocated too many nodes in your postalloc expand."); |
goetz@6478 | 1045 | |
goetz@6478 | 1046 | // Disconnect the inputs of the old node. |
goetz@6478 | 1047 | // |
goetz@6478 | 1048 | // We reuse MachSpillCopy nodes. If we need to expand them, there |
goetz@6478 | 1049 | // are many, so reusing pays off. If reused, the node already |
goetz@6478 | 1050 | // has the new ins. n must be the last node on new_nodes list. |
goetz@6478 | 1051 | if (!n->is_MachSpillCopy()) { |
goetz@6478 | 1052 | for (int k = req - 1; k >= 0; --k) { |
goetz@6478 | 1053 | n->del_req(k); |
goetz@6478 | 1054 | } |
goetz@6478 | 1055 | } |
goetz@6478 | 1056 | |
goetz@6478 | 1057 | #ifdef ASSERT |
goetz@6478 | 1058 | // Check that all nodes have proper operands. |
goetz@6478 | 1059 | for (int k = 0; k < new_nodes.length(); ++k) { |
goetz@6478 | 1060 | if (new_nodes.at(k)->_idx < max_idx || !new_nodes.at(k)->is_Mach()) continue; // old node, Proj ... |
goetz@6478 | 1061 | MachNode *m = new_nodes.at(k)->as_Mach(); |
goetz@6478 | 1062 | for (unsigned int l = 0; l < m->num_opnds(); ++l) { |
goetz@6478 | 1063 | if (MachOper::notAnOper(m->_opnds[l])) { |
goetz@6478 | 1064 | outputStream *os = tty; |
goetz@6478 | 1065 | os->print("Node %s ", m->Name()); |
goetz@6478 | 1066 | os->print("has invalid opnd %d: %p\n", l, m->_opnds[l]); |
goetz@6478 | 1067 | assert(0, "Invalid operands, see inline trace in hs_err_pid file."); |
goetz@6478 | 1068 | } |
goetz@6478 | 1069 | } |
goetz@6478 | 1070 | } |
goetz@6478 | 1071 | #endif |
goetz@6478 | 1072 | |
goetz@6478 | 1073 | // Collect succs of old node in remove (for projections) and in succs (for |
goetz@6478 | 1074 | // all other nodes) do _not_ collect projections in remove (but in succs) |
goetz@6478 | 1075 | // in case the node is a call. We need the projections for calls as they are |
goetz@6478 | 1076 | // associated with registes (i.e. they are defs). |
goetz@6478 | 1077 | succs.clear(); |
goetz@6478 | 1078 | for (DUIterator k = n->outs(); n->has_out(k); k++) { |
goetz@6478 | 1079 | if (n->out(k)->is_Proj() && !n->is_MachCall() && !n->is_MachBranch()) { |
goetz@6478 | 1080 | remove.push(n->out(k)); |
goetz@6478 | 1081 | } else { |
goetz@6478 | 1082 | succs.push(n->out(k)); |
goetz@6478 | 1083 | } |
goetz@6478 | 1084 | } |
goetz@6478 | 1085 | // Replace old node n as input of its succs by last of the new nodes. |
goetz@6478 | 1086 | for (int k = 0; k < succs.length(); ++k) { |
goetz@6478 | 1087 | Node *succ = succs.at(k); |
goetz@6478 | 1088 | for (uint l = 0; l < succ->req(); ++l) { |
goetz@6478 | 1089 | if (succ->in(l) == n) { |
goetz@6478 | 1090 | succ->set_req(l, new_nodes.at(new_nodes.length() - 1)); |
goetz@6478 | 1091 | } |
goetz@6478 | 1092 | } |
goetz@6478 | 1093 | for (uint l = succ->req(); l < succ->len(); ++l) { |
goetz@6478 | 1094 | if (succ->in(l) == n) { |
goetz@6478 | 1095 | succ->set_prec(l, new_nodes.at(new_nodes.length() - 1)); |
goetz@6478 | 1096 | } |
goetz@6478 | 1097 | } |
goetz@6478 | 1098 | } |
goetz@6478 | 1099 | |
goetz@6478 | 1100 | // Index of old node in block. |
goetz@6478 | 1101 | uint index = b->find_node(n); |
goetz@6478 | 1102 | // Insert new nodes into block and map them in nodes->blocks array |
goetz@6478 | 1103 | // and remember last node in n2. |
goetz@6478 | 1104 | Node *n2 = NULL; |
goetz@6478 | 1105 | for (int k = 0; k < new_nodes.length(); ++k) { |
goetz@6478 | 1106 | n2 = new_nodes.at(k); |
goetz@6478 | 1107 | b->insert_node(n2, ++index); |
goetz@6478 | 1108 | map_node_to_block(n2, b); |
goetz@6478 | 1109 | } |
goetz@6478 | 1110 | |
goetz@6478 | 1111 | // Add old node n to remove and remove them all from block. |
goetz@6478 | 1112 | remove.push(n); |
goetz@6478 | 1113 | j--; |
goetz@6478 | 1114 | #ifdef ASSERT |
goetz@6478 | 1115 | if (TracePostallocExpand && Verbose) { |
goetz@6478 | 1116 | tty->print(" removing:\n"); |
goetz@6478 | 1117 | for (int k = 0; k < remove.length(); ++k) { |
goetz@6478 | 1118 | tty->print(" "); remove.at(k)->dump(); |
goetz@6478 | 1119 | } |
goetz@6478 | 1120 | tty->print(" inserting:\n"); |
goetz@6478 | 1121 | for (int k = 0; k < new_nodes.length(); ++k) { |
goetz@6478 | 1122 | tty->print(" "); new_nodes.at(k)->dump(); |
goetz@6478 | 1123 | } |
goetz@6478 | 1124 | } |
goetz@6478 | 1125 | #endif |
goetz@6478 | 1126 | for (int k = 0; k < remove.length(); ++k) { |
goetz@6478 | 1127 | if (b->contains(remove.at(k))) { |
goetz@6478 | 1128 | b->find_remove(remove.at(k)); |
goetz@6478 | 1129 | } else { |
goetz@6478 | 1130 | assert(remove.at(k)->is_Proj() && (remove.at(k)->in(0)->is_MachBranch()), ""); |
goetz@6478 | 1131 | } |
goetz@6478 | 1132 | } |
goetz@6478 | 1133 | // If anything has been inserted (n2 != NULL), continue after last node inserted. |
goetz@6478 | 1134 | // This does not always work. Some postalloc expands don't insert any nodes, if they |
goetz@6478 | 1135 | // do optimizations (e.g., max(x,x)). In this case we decrement j accordingly. |
goetz@6478 | 1136 | j = n2 ? b->find_node(n2) : j; |
goetz@6478 | 1137 | } |
goetz@6478 | 1138 | } |
goetz@6478 | 1139 | } |
goetz@6478 | 1140 | |
goetz@6478 | 1141 | #ifdef ASSERT |
goetz@6478 | 1142 | if (foundNode) { |
goetz@6478 | 1143 | tty->print("FINISHED %d %s\n", C->compile_id(), |
goetz@6478 | 1144 | C->method() ? C->method()->name()->as_utf8() : C->stub_name()); |
goetz@6478 | 1145 | tty->flush(); |
goetz@6478 | 1146 | } |
goetz@6478 | 1147 | #endif |
goetz@6478 | 1148 | } |
goetz@6478 | 1149 | |
goetz@6478 | 1150 | |
goetz@6478 | 1151 | //------------------------------dump------------------------------------------- |
duke@435 | 1152 | #ifndef PRODUCT |
duke@435 | 1153 | void PhaseCFG::_dump_cfg( const Node *end, VectorSet &visited ) const { |
duke@435 | 1154 | const Node *x = end->is_block_proj(); |
duke@435 | 1155 | assert( x, "not a CFG" ); |
duke@435 | 1156 | |
duke@435 | 1157 | // Do not visit this block again |
duke@435 | 1158 | if( visited.test_set(x->_idx) ) return; |
duke@435 | 1159 | |
duke@435 | 1160 | // Skip through this block |
duke@435 | 1161 | const Node *p = x; |
duke@435 | 1162 | do { |
duke@435 | 1163 | p = p->in(0); // Move control forward |
duke@435 | 1164 | assert( !p->is_block_proj() || p->is_Root(), "not a CFG" ); |
duke@435 | 1165 | } while( !p->is_block_start() ); |
duke@435 | 1166 | |
duke@435 | 1167 | // Recursively visit |
adlertz@5509 | 1168 | for (uint i = 1; i < p->req(); i++) { |
adlertz@5509 | 1169 | _dump_cfg(p->in(i), visited); |
adlertz@5509 | 1170 | } |
duke@435 | 1171 | |
duke@435 | 1172 | // Dump the block |
adlertz@5509 | 1173 | get_block_for_node(p)->dump(this); |
duke@435 | 1174 | } |
duke@435 | 1175 | |
duke@435 | 1176 | void PhaseCFG::dump( ) const { |
adlertz@5539 | 1177 | tty->print("\n--- CFG --- %d BBs\n", number_of_blocks()); |
adlertz@5509 | 1178 | if (_blocks.size()) { // Did we do basic-block layout? |
adlertz@5539 | 1179 | for (uint i = 0; i < number_of_blocks(); i++) { |
adlertz@5539 | 1180 | const Block* block = get_block(i); |
adlertz@5539 | 1181 | block->dump(this); |
adlertz@5509 | 1182 | } |
duke@435 | 1183 | } else { // Else do it with a DFS |
adlertz@5509 | 1184 | VectorSet visited(_block_arena); |
duke@435 | 1185 | _dump_cfg(_root,visited); |
duke@435 | 1186 | } |
duke@435 | 1187 | } |
duke@435 | 1188 | |
duke@435 | 1189 | void PhaseCFG::dump_headers() { |
adlertz@5539 | 1190 | for (uint i = 0; i < number_of_blocks(); i++) { |
adlertz@5539 | 1191 | Block* block = get_block(i); |
adlertz@5539 | 1192 | if (block != NULL) { |
adlertz@5539 | 1193 | block->dump_head(this); |
adlertz@5509 | 1194 | } |
duke@435 | 1195 | } |
duke@435 | 1196 | } |
duke@435 | 1197 | |
adlertz@5539 | 1198 | void PhaseCFG::verify() const { |
kvn@1001 | 1199 | #ifdef ASSERT |
duke@435 | 1200 | // Verify sane CFG |
adlertz@5539 | 1201 | for (uint i = 0; i < number_of_blocks(); i++) { |
adlertz@5539 | 1202 | Block* block = get_block(i); |
adlertz@5635 | 1203 | uint cnt = block->number_of_nodes(); |
duke@435 | 1204 | uint j; |
kvn@3311 | 1205 | for (j = 0; j < cnt; j++) { |
adlertz@5635 | 1206 | Node *n = block->get_node(j); |
adlertz@5539 | 1207 | assert(get_block_for_node(n) == block, ""); |
adlertz@5539 | 1208 | if (j >= 1 && n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_CreateEx) { |
adlertz@5635 | 1209 | assert(j == 1 || block->get_node(j-1)->is_Phi(), "CreateEx must be first instruction in block"); |
duke@435 | 1210 | } |
kvn@3311 | 1211 | for (uint k = 0; k < n->req(); k++) { |
kvn@1001 | 1212 | Node *def = n->in(k); |
kvn@3311 | 1213 | if (def && def != n) { |
adlertz@5509 | 1214 | assert(get_block_for_node(def) || def->is_Con(), "must have block; constants for debug info ok"); |
kvn@1001 | 1215 | // Verify that instructions in the block is in correct order. |
kvn@1001 | 1216 | // Uses must follow their definition if they are at the same block. |
kvn@1001 | 1217 | // Mostly done to check that MachSpillCopy nodes are placed correctly |
kvn@1001 | 1218 | // when CreateEx node is moved in build_ifg_physical(). |
adlertz@5539 | 1219 | if (get_block_for_node(def) == block && !(block->head()->is_Loop() && n->is_Phi()) && |
kvn@1001 | 1220 | // See (+++) comment in reg_split.cpp |
kvn@3311 | 1221 | !(n->jvms() != NULL && n->jvms()->is_monitor_use(k))) { |
kvn@1328 | 1222 | bool is_loop = false; |
kvn@1328 | 1223 | if (n->is_Phi()) { |
kvn@3311 | 1224 | for (uint l = 1; l < def->req(); l++) { |
kvn@1328 | 1225 | if (n == def->in(l)) { |
kvn@1328 | 1226 | is_loop = true; |
kvn@1328 | 1227 | break; // Some kind of loop |
kvn@1328 | 1228 | } |
kvn@1328 | 1229 | } |
kvn@1328 | 1230 | } |
adlertz@5539 | 1231 | assert(is_loop || block->find_node(def) < j, "uses must follow definitions"); |
kvn@1036 | 1232 | } |
duke@435 | 1233 | } |
duke@435 | 1234 | } |
duke@435 | 1235 | } |
duke@435 | 1236 | |
adlertz@5539 | 1237 | j = block->end_idx(); |
adlertz@5635 | 1238 | Node* bp = (Node*)block->get_node(block->number_of_nodes() - 1)->is_block_proj(); |
adlertz@5539 | 1239 | assert(bp, "last instruction must be a block proj"); |
adlertz@5635 | 1240 | assert(bp == block->get_node(j), "wrong number of successors for this block"); |
kvn@3311 | 1241 | if (bp->is_Catch()) { |
adlertz@5635 | 1242 | while (block->get_node(--j)->is_MachProj()) { |
adlertz@5539 | 1243 | ; |
adlertz@5539 | 1244 | } |
adlertz@5635 | 1245 | assert(block->get_node(j)->is_MachCall(), "CatchProj must follow call"); |
kvn@3311 | 1246 | } else if (bp->is_Mach() && bp->as_Mach()->ideal_Opcode() == Op_If) { |
adlertz@5539 | 1247 | assert(block->_num_succs == 2, "Conditional branch must have two targets"); |
duke@435 | 1248 | } |
duke@435 | 1249 | } |
kvn@1001 | 1250 | #endif |
duke@435 | 1251 | } |
duke@435 | 1252 | #endif |
duke@435 | 1253 | |
duke@435 | 1254 | UnionFind::UnionFind( uint max ) : _cnt(max), _max(max), _indices(NEW_RESOURCE_ARRAY(uint,max)) { |
duke@435 | 1255 | Copy::zero_to_bytes( _indices, sizeof(uint)*max ); |
duke@435 | 1256 | } |
duke@435 | 1257 | |
duke@435 | 1258 | void UnionFind::extend( uint from_idx, uint to_idx ) { |
duke@435 | 1259 | _nesting.check(); |
duke@435 | 1260 | if( from_idx >= _max ) { |
duke@435 | 1261 | uint size = 16; |
duke@435 | 1262 | while( size <= from_idx ) size <<=1; |
duke@435 | 1263 | _indices = REALLOC_RESOURCE_ARRAY( uint, _indices, _max, size ); |
duke@435 | 1264 | _max = size; |
duke@435 | 1265 | } |
duke@435 | 1266 | while( _cnt <= from_idx ) _indices[_cnt++] = 0; |
duke@435 | 1267 | _indices[from_idx] = to_idx; |
duke@435 | 1268 | } |
duke@435 | 1269 | |
duke@435 | 1270 | void UnionFind::reset( uint max ) { |
duke@435 | 1271 | assert( max <= max_uint, "Must fit within uint" ); |
duke@435 | 1272 | // Force the Union-Find mapping to be at least this large |
duke@435 | 1273 | extend(max,0); |
duke@435 | 1274 | // Initialize to be the ID mapping. |
rasbold@853 | 1275 | for( uint i=0; i<max; i++ ) map(i,i); |
duke@435 | 1276 | } |
duke@435 | 1277 | |
duke@435 | 1278 | // Straight out of Tarjan's union-find algorithm |
duke@435 | 1279 | uint UnionFind::Find_compress( uint idx ) { |
duke@435 | 1280 | uint cur = idx; |
duke@435 | 1281 | uint next = lookup(cur); |
duke@435 | 1282 | while( next != cur ) { // Scan chain of equivalences |
duke@435 | 1283 | assert( next < cur, "always union smaller" ); |
duke@435 | 1284 | cur = next; // until find a fixed-point |
duke@435 | 1285 | next = lookup(cur); |
duke@435 | 1286 | } |
duke@435 | 1287 | // Core of union-find algorithm: update chain of |
duke@435 | 1288 | // equivalences to be equal to the root. |
duke@435 | 1289 | while( idx != next ) { |
duke@435 | 1290 | uint tmp = lookup(idx); |
duke@435 | 1291 | map(idx, next); |
duke@435 | 1292 | idx = tmp; |
duke@435 | 1293 | } |
duke@435 | 1294 | return idx; |
duke@435 | 1295 | } |
duke@435 | 1296 | |
duke@435 | 1297 | // Like Find above, but no path compress, so bad asymptotic behavior |
duke@435 | 1298 | uint UnionFind::Find_const( uint idx ) const { |
duke@435 | 1299 | if( idx == 0 ) return idx; // Ignore the zero idx |
duke@435 | 1300 | // Off the end? This can happen during debugging dumps |
duke@435 | 1301 | // when data structures have not finished being updated. |
duke@435 | 1302 | if( idx >= _max ) return idx; |
duke@435 | 1303 | uint next = lookup(idx); |
duke@435 | 1304 | while( next != idx ) { // Scan chain of equivalences |
duke@435 | 1305 | idx = next; // until find a fixed-point |
duke@435 | 1306 | next = lookup(idx); |
duke@435 | 1307 | } |
duke@435 | 1308 | return next; |
duke@435 | 1309 | } |
duke@435 | 1310 | |
duke@435 | 1311 | // union 2 sets together. |
duke@435 | 1312 | void UnionFind::Union( uint idx1, uint idx2 ) { |
duke@435 | 1313 | uint src = Find(idx1); |
duke@435 | 1314 | uint dst = Find(idx2); |
duke@435 | 1315 | assert( src, "" ); |
duke@435 | 1316 | assert( dst, "" ); |
duke@435 | 1317 | assert( src < _max, "oob" ); |
duke@435 | 1318 | assert( dst < _max, "oob" ); |
duke@435 | 1319 | assert( src < dst, "always union smaller" ); |
duke@435 | 1320 | map(dst,src); |
duke@435 | 1321 | } |
rasbold@853 | 1322 | |
rasbold@853 | 1323 | #ifndef PRODUCT |
rasbold@853 | 1324 | void Trace::dump( ) const { |
rasbold@853 | 1325 | tty->print_cr("Trace (freq %f)", first_block()->_freq); |
rasbold@853 | 1326 | for (Block *b = first_block(); b != NULL; b = next(b)) { |
rasbold@853 | 1327 | tty->print(" B%d", b->_pre_order); |
rasbold@853 | 1328 | if (b->head()->is_Loop()) { |
rasbold@853 | 1329 | tty->print(" (L%d)", b->compute_loop_alignment()); |
rasbold@853 | 1330 | } |
rasbold@853 | 1331 | if (b->has_loop_alignment()) { |
rasbold@853 | 1332 | tty->print(" (T%d)", b->code_alignment()); |
rasbold@853 | 1333 | } |
rasbold@853 | 1334 | } |
rasbold@853 | 1335 | tty->cr(); |
rasbold@853 | 1336 | } |
rasbold@853 | 1337 | |
rasbold@853 | 1338 | void CFGEdge::dump( ) const { |
rasbold@853 | 1339 | tty->print(" B%d --> B%d Freq: %f out:%3d%% in:%3d%% State: ", |
rasbold@853 | 1340 | from()->_pre_order, to()->_pre_order, freq(), _from_pct, _to_pct); |
rasbold@853 | 1341 | switch(state()) { |
rasbold@853 | 1342 | case connected: |
rasbold@853 | 1343 | tty->print("connected"); |
rasbold@853 | 1344 | break; |
rasbold@853 | 1345 | case open: |
rasbold@853 | 1346 | tty->print("open"); |
rasbold@853 | 1347 | break; |
rasbold@853 | 1348 | case interior: |
rasbold@853 | 1349 | tty->print("interior"); |
rasbold@853 | 1350 | break; |
rasbold@853 | 1351 | } |
rasbold@853 | 1352 | if (infrequent()) { |
rasbold@853 | 1353 | tty->print(" infrequent"); |
rasbold@853 | 1354 | } |
rasbold@853 | 1355 | tty->cr(); |
rasbold@853 | 1356 | } |
rasbold@853 | 1357 | #endif |
rasbold@853 | 1358 | |
rasbold@853 | 1359 | // Comparison function for edges |
rasbold@853 | 1360 | static int edge_order(CFGEdge **e0, CFGEdge **e1) { |
rasbold@853 | 1361 | float freq0 = (*e0)->freq(); |
rasbold@853 | 1362 | float freq1 = (*e1)->freq(); |
rasbold@853 | 1363 | if (freq0 != freq1) { |
rasbold@853 | 1364 | return freq0 > freq1 ? -1 : 1; |
rasbold@853 | 1365 | } |
rasbold@853 | 1366 | |
rasbold@853 | 1367 | int dist0 = (*e0)->to()->_rpo - (*e0)->from()->_rpo; |
rasbold@853 | 1368 | int dist1 = (*e1)->to()->_rpo - (*e1)->from()->_rpo; |
rasbold@853 | 1369 | |
rasbold@853 | 1370 | return dist1 - dist0; |
rasbold@853 | 1371 | } |
rasbold@853 | 1372 | |
rasbold@853 | 1373 | // Comparison function for edges |
kvn@3128 | 1374 | extern "C" int trace_frequency_order(const void *p0, const void *p1) { |
rasbold@853 | 1375 | Trace *tr0 = *(Trace **) p0; |
rasbold@853 | 1376 | Trace *tr1 = *(Trace **) p1; |
rasbold@853 | 1377 | Block *b0 = tr0->first_block(); |
rasbold@853 | 1378 | Block *b1 = tr1->first_block(); |
rasbold@853 | 1379 | |
rasbold@853 | 1380 | // The trace of connector blocks goes at the end; |
rasbold@853 | 1381 | // we only expect one such trace |
rasbold@853 | 1382 | if (b0->is_connector() != b1->is_connector()) { |
rasbold@853 | 1383 | return b1->is_connector() ? -1 : 1; |
rasbold@853 | 1384 | } |
rasbold@853 | 1385 | |
rasbold@853 | 1386 | // Pull more frequently executed blocks to the beginning |
rasbold@853 | 1387 | float freq0 = b0->_freq; |
rasbold@853 | 1388 | float freq1 = b1->_freq; |
rasbold@853 | 1389 | if (freq0 != freq1) { |
rasbold@853 | 1390 | return freq0 > freq1 ? -1 : 1; |
rasbold@853 | 1391 | } |
rasbold@853 | 1392 | |
rasbold@853 | 1393 | int diff = tr0->first_block()->_rpo - tr1->first_block()->_rpo; |
rasbold@853 | 1394 | |
rasbold@853 | 1395 | return diff; |
rasbold@853 | 1396 | } |
rasbold@853 | 1397 | |
rasbold@853 | 1398 | // Find edges of interest, i.e, those which can fall through. Presumes that |
rasbold@853 | 1399 | // edges which don't fall through are of low frequency and can be generally |
rasbold@853 | 1400 | // ignored. Initialize the list of traces. |
adlertz@5539 | 1401 | void PhaseBlockLayout::find_edges() { |
rasbold@853 | 1402 | // Walk the blocks, creating edges and Traces |
rasbold@853 | 1403 | uint i; |
rasbold@853 | 1404 | Trace *tr = NULL; |
adlertz@5539 | 1405 | for (i = 0; i < _cfg.number_of_blocks(); i++) { |
adlertz@5539 | 1406 | Block* b = _cfg.get_block(i); |
rasbold@853 | 1407 | tr = new Trace(b, next, prev); |
rasbold@853 | 1408 | traces[tr->id()] = tr; |
rasbold@853 | 1409 | |
rasbold@853 | 1410 | // All connector blocks should be at the end of the list |
rasbold@853 | 1411 | if (b->is_connector()) break; |
rasbold@853 | 1412 | |
rasbold@853 | 1413 | // If this block and the next one have a one-to-one successor |
rasbold@853 | 1414 | // predecessor relationship, simply append the next block |
rasbold@853 | 1415 | int nfallthru = b->num_fall_throughs(); |
rasbold@853 | 1416 | while (nfallthru == 1 && |
rasbold@853 | 1417 | b->succ_fall_through(0)) { |
rasbold@853 | 1418 | Block *n = b->_succs[0]; |
rasbold@853 | 1419 | |
rasbold@853 | 1420 | // Skip over single-entry connector blocks, we don't want to |
rasbold@853 | 1421 | // add them to the trace. |
rasbold@853 | 1422 | while (n->is_connector() && n->num_preds() == 1) { |
rasbold@853 | 1423 | n = n->_succs[0]; |
rasbold@853 | 1424 | } |
rasbold@853 | 1425 | |
rasbold@853 | 1426 | // We see a merge point, so stop search for the next block |
rasbold@853 | 1427 | if (n->num_preds() != 1) break; |
rasbold@853 | 1428 | |
rasbold@853 | 1429 | i++; |
adlertz@5539 | 1430 | assert(n = _cfg.get_block(i), "expecting next block"); |
rasbold@853 | 1431 | tr->append(n); |
rasbold@853 | 1432 | uf->map(n->_pre_order, tr->id()); |
rasbold@853 | 1433 | traces[n->_pre_order] = NULL; |
rasbold@853 | 1434 | nfallthru = b->num_fall_throughs(); |
rasbold@853 | 1435 | b = n; |
rasbold@853 | 1436 | } |
rasbold@853 | 1437 | |
rasbold@853 | 1438 | if (nfallthru > 0) { |
rasbold@853 | 1439 | // Create a CFGEdge for each outgoing |
rasbold@853 | 1440 | // edge that could be a fall-through. |
rasbold@853 | 1441 | for (uint j = 0; j < b->_num_succs; j++ ) { |
rasbold@853 | 1442 | if (b->succ_fall_through(j)) { |
rasbold@853 | 1443 | Block *target = b->non_connector_successor(j); |
rasbold@853 | 1444 | float freq = b->_freq * b->succ_prob(j); |
rasbold@853 | 1445 | int from_pct = (int) ((100 * freq) / b->_freq); |
rasbold@853 | 1446 | int to_pct = (int) ((100 * freq) / target->_freq); |
rasbold@853 | 1447 | edges->append(new CFGEdge(b, target, freq, from_pct, to_pct)); |
rasbold@853 | 1448 | } |
rasbold@853 | 1449 | } |
rasbold@853 | 1450 | } |
rasbold@853 | 1451 | } |
rasbold@853 | 1452 | |
rasbold@853 | 1453 | // Group connector blocks into one trace |
adlertz@5539 | 1454 | for (i++; i < _cfg.number_of_blocks(); i++) { |
adlertz@5539 | 1455 | Block *b = _cfg.get_block(i); |
rasbold@853 | 1456 | assert(b->is_connector(), "connector blocks at the end"); |
rasbold@853 | 1457 | tr->append(b); |
rasbold@853 | 1458 | uf->map(b->_pre_order, tr->id()); |
rasbold@853 | 1459 | traces[b->_pre_order] = NULL; |
rasbold@853 | 1460 | } |
rasbold@853 | 1461 | } |
rasbold@853 | 1462 | |
rasbold@853 | 1463 | // Union two traces together in uf, and null out the trace in the list |
adlertz@5539 | 1464 | void PhaseBlockLayout::union_traces(Trace* updated_trace, Trace* old_trace) { |
rasbold@853 | 1465 | uint old_id = old_trace->id(); |
rasbold@853 | 1466 | uint updated_id = updated_trace->id(); |
rasbold@853 | 1467 | |
rasbold@853 | 1468 | uint lo_id = updated_id; |
rasbold@853 | 1469 | uint hi_id = old_id; |
rasbold@853 | 1470 | |
rasbold@853 | 1471 | // If from is greater than to, swap values to meet |
rasbold@853 | 1472 | // UnionFind guarantee. |
rasbold@853 | 1473 | if (updated_id > old_id) { |
rasbold@853 | 1474 | lo_id = old_id; |
rasbold@853 | 1475 | hi_id = updated_id; |
rasbold@853 | 1476 | |
rasbold@853 | 1477 | // Fix up the trace ids |
rasbold@853 | 1478 | traces[lo_id] = traces[updated_id]; |
rasbold@853 | 1479 | updated_trace->set_id(lo_id); |
rasbold@853 | 1480 | } |
rasbold@853 | 1481 | |
rasbold@853 | 1482 | // Union the lower with the higher and remove the pointer |
rasbold@853 | 1483 | // to the higher. |
rasbold@853 | 1484 | uf->Union(lo_id, hi_id); |
rasbold@853 | 1485 | traces[hi_id] = NULL; |
rasbold@853 | 1486 | } |
rasbold@853 | 1487 | |
rasbold@853 | 1488 | // Append traces together via the most frequently executed edges |
adlertz@5539 | 1489 | void PhaseBlockLayout::grow_traces() { |
rasbold@853 | 1490 | // Order the edges, and drive the growth of Traces via the most |
rasbold@853 | 1491 | // frequently executed edges. |
rasbold@853 | 1492 | edges->sort(edge_order); |
rasbold@853 | 1493 | for (int i = 0; i < edges->length(); i++) { |
rasbold@853 | 1494 | CFGEdge *e = edges->at(i); |
rasbold@853 | 1495 | |
rasbold@853 | 1496 | if (e->state() != CFGEdge::open) continue; |
rasbold@853 | 1497 | |
rasbold@853 | 1498 | Block *src_block = e->from(); |
rasbold@853 | 1499 | Block *targ_block = e->to(); |
rasbold@853 | 1500 | |
rasbold@853 | 1501 | // Don't grow traces along backedges? |
rasbold@853 | 1502 | if (!BlockLayoutRotateLoops) { |
rasbold@853 | 1503 | if (targ_block->_rpo <= src_block->_rpo) { |
rasbold@853 | 1504 | targ_block->set_loop_alignment(targ_block); |
rasbold@853 | 1505 | continue; |
rasbold@853 | 1506 | } |
rasbold@853 | 1507 | } |
rasbold@853 | 1508 | |
rasbold@853 | 1509 | Trace *src_trace = trace(src_block); |
rasbold@853 | 1510 | Trace *targ_trace = trace(targ_block); |
rasbold@853 | 1511 | |
rasbold@853 | 1512 | // If the edge in question can join two traces at their ends, |
rasbold@853 | 1513 | // append one trace to the other. |
rasbold@853 | 1514 | if (src_trace->last_block() == src_block) { |
rasbold@853 | 1515 | if (src_trace == targ_trace) { |
rasbold@853 | 1516 | e->set_state(CFGEdge::interior); |
rasbold@853 | 1517 | if (targ_trace->backedge(e)) { |
rasbold@853 | 1518 | // Reset i to catch any newly eligible edge |
rasbold@853 | 1519 | // (Or we could remember the first "open" edge, and reset there) |
rasbold@853 | 1520 | i = 0; |
rasbold@853 | 1521 | } |
rasbold@853 | 1522 | } else if (targ_trace->first_block() == targ_block) { |
rasbold@853 | 1523 | e->set_state(CFGEdge::connected); |
rasbold@853 | 1524 | src_trace->append(targ_trace); |
rasbold@853 | 1525 | union_traces(src_trace, targ_trace); |
rasbold@853 | 1526 | } |
rasbold@853 | 1527 | } |
rasbold@853 | 1528 | } |
rasbold@853 | 1529 | } |
rasbold@853 | 1530 | |
rasbold@853 | 1531 | // Embed one trace into another, if the fork or join points are sufficiently |
rasbold@853 | 1532 | // balanced. |
adlertz@5539 | 1533 | void PhaseBlockLayout::merge_traces(bool fall_thru_only) { |
rasbold@853 | 1534 | // Walk the edge list a another time, looking at unprocessed edges. |
rasbold@853 | 1535 | // Fold in diamonds |
rasbold@853 | 1536 | for (int i = 0; i < edges->length(); i++) { |
rasbold@853 | 1537 | CFGEdge *e = edges->at(i); |
rasbold@853 | 1538 | |
rasbold@853 | 1539 | if (e->state() != CFGEdge::open) continue; |
rasbold@853 | 1540 | if (fall_thru_only) { |
rasbold@853 | 1541 | if (e->infrequent()) continue; |
rasbold@853 | 1542 | } |
rasbold@853 | 1543 | |
rasbold@853 | 1544 | Block *src_block = e->from(); |
rasbold@853 | 1545 | Trace *src_trace = trace(src_block); |
rasbold@853 | 1546 | bool src_at_tail = src_trace->last_block() == src_block; |
rasbold@853 | 1547 | |
rasbold@853 | 1548 | Block *targ_block = e->to(); |
rasbold@853 | 1549 | Trace *targ_trace = trace(targ_block); |
rasbold@853 | 1550 | bool targ_at_start = targ_trace->first_block() == targ_block; |
rasbold@853 | 1551 | |
rasbold@853 | 1552 | if (src_trace == targ_trace) { |
rasbold@853 | 1553 | // This may be a loop, but we can't do much about it. |
rasbold@853 | 1554 | e->set_state(CFGEdge::interior); |
rasbold@853 | 1555 | continue; |
rasbold@853 | 1556 | } |
rasbold@853 | 1557 | |
rasbold@853 | 1558 | if (fall_thru_only) { |
rasbold@853 | 1559 | // If the edge links the middle of two traces, we can't do anything. |
rasbold@853 | 1560 | // Mark the edge and continue. |
rasbold@853 | 1561 | if (!src_at_tail & !targ_at_start) { |
rasbold@853 | 1562 | continue; |
rasbold@853 | 1563 | } |
rasbold@853 | 1564 | |
rasbold@853 | 1565 | // Don't grow traces along backedges? |
rasbold@853 | 1566 | if (!BlockLayoutRotateLoops && (targ_block->_rpo <= src_block->_rpo)) { |
rasbold@853 | 1567 | continue; |
rasbold@853 | 1568 | } |
rasbold@853 | 1569 | |
rasbold@853 | 1570 | // If both ends of the edge are available, why didn't we handle it earlier? |
rasbold@853 | 1571 | assert(src_at_tail ^ targ_at_start, "Should have caught this edge earlier."); |
rasbold@853 | 1572 | |
rasbold@853 | 1573 | if (targ_at_start) { |
rasbold@853 | 1574 | // Insert the "targ" trace in the "src" trace if the insertion point |
rasbold@853 | 1575 | // is a two way branch. |
rasbold@853 | 1576 | // Better profitability check possible, but may not be worth it. |
rasbold@853 | 1577 | // Someday, see if the this "fork" has an associated "join"; |
rasbold@853 | 1578 | // then make a policy on merging this trace at the fork or join. |
rasbold@853 | 1579 | // For example, other things being equal, it may be better to place this |
rasbold@853 | 1580 | // trace at the join point if the "src" trace ends in a two-way, but |
rasbold@853 | 1581 | // the insertion point is one-way. |
rasbold@853 | 1582 | assert(src_block->num_fall_throughs() == 2, "unexpected diamond"); |
rasbold@853 | 1583 | e->set_state(CFGEdge::connected); |
rasbold@853 | 1584 | src_trace->insert_after(src_block, targ_trace); |
rasbold@853 | 1585 | union_traces(src_trace, targ_trace); |
rasbold@853 | 1586 | } else if (src_at_tail) { |
adlertz@5539 | 1587 | if (src_trace != trace(_cfg.get_root_block())) { |
rasbold@853 | 1588 | e->set_state(CFGEdge::connected); |
rasbold@853 | 1589 | targ_trace->insert_before(targ_block, src_trace); |
rasbold@853 | 1590 | union_traces(targ_trace, src_trace); |
rasbold@853 | 1591 | } |
rasbold@853 | 1592 | } |
rasbold@853 | 1593 | } else if (e->state() == CFGEdge::open) { |
rasbold@853 | 1594 | // Append traces, even without a fall-thru connection. |
twisti@1040 | 1595 | // But leave root entry at the beginning of the block list. |
adlertz@5539 | 1596 | if (targ_trace != trace(_cfg.get_root_block())) { |
rasbold@853 | 1597 | e->set_state(CFGEdge::connected); |
rasbold@853 | 1598 | src_trace->append(targ_trace); |
rasbold@853 | 1599 | union_traces(src_trace, targ_trace); |
rasbold@853 | 1600 | } |
rasbold@853 | 1601 | } |
rasbold@853 | 1602 | } |
rasbold@853 | 1603 | } |
rasbold@853 | 1604 | |
rasbold@853 | 1605 | // Order the sequence of the traces in some desirable way, and fixup the |
rasbold@853 | 1606 | // jumps at the end of each block. |
adlertz@5539 | 1607 | void PhaseBlockLayout::reorder_traces(int count) { |
rasbold@853 | 1608 | ResourceArea *area = Thread::current()->resource_area(); |
rasbold@853 | 1609 | Trace ** new_traces = NEW_ARENA_ARRAY(area, Trace *, count); |
rasbold@853 | 1610 | Block_List worklist; |
rasbold@853 | 1611 | int new_count = 0; |
rasbold@853 | 1612 | |
rasbold@853 | 1613 | // Compact the traces. |
rasbold@853 | 1614 | for (int i = 0; i < count; i++) { |
rasbold@853 | 1615 | Trace *tr = traces[i]; |
rasbold@853 | 1616 | if (tr != NULL) { |
rasbold@853 | 1617 | new_traces[new_count++] = tr; |
rasbold@853 | 1618 | } |
rasbold@853 | 1619 | } |
rasbold@853 | 1620 | |
rasbold@853 | 1621 | // The entry block should be first on the new trace list. |
adlertz@5539 | 1622 | Trace *tr = trace(_cfg.get_root_block()); |
rasbold@853 | 1623 | assert(tr == new_traces[0], "entry trace misplaced"); |
rasbold@853 | 1624 | |
rasbold@853 | 1625 | // Sort the new trace list by frequency |
rasbold@853 | 1626 | qsort(new_traces + 1, new_count - 1, sizeof(new_traces[0]), trace_frequency_order); |
rasbold@853 | 1627 | |
rasbold@853 | 1628 | // Patch up the successor blocks |
adlertz@5539 | 1629 | _cfg.clear_blocks(); |
rasbold@853 | 1630 | for (int i = 0; i < new_count; i++) { |
rasbold@853 | 1631 | Trace *tr = new_traces[i]; |
rasbold@853 | 1632 | if (tr != NULL) { |
rasbold@853 | 1633 | tr->fixup_blocks(_cfg); |
rasbold@853 | 1634 | } |
rasbold@853 | 1635 | } |
rasbold@853 | 1636 | } |
rasbold@853 | 1637 | |
rasbold@853 | 1638 | // Order basic blocks based on frequency |
adlertz@5539 | 1639 | PhaseBlockLayout::PhaseBlockLayout(PhaseCFG &cfg) |
adlertz@5539 | 1640 | : Phase(BlockLayout) |
adlertz@5539 | 1641 | , _cfg(cfg) { |
rasbold@853 | 1642 | ResourceMark rm; |
rasbold@853 | 1643 | ResourceArea *area = Thread::current()->resource_area(); |
rasbold@853 | 1644 | |
rasbold@853 | 1645 | // List of traces |
adlertz@5539 | 1646 | int size = _cfg.number_of_blocks() + 1; |
rasbold@853 | 1647 | traces = NEW_ARENA_ARRAY(area, Trace *, size); |
rasbold@853 | 1648 | memset(traces, 0, size*sizeof(Trace*)); |
rasbold@853 | 1649 | next = NEW_ARENA_ARRAY(area, Block *, size); |
rasbold@853 | 1650 | memset(next, 0, size*sizeof(Block *)); |
rasbold@853 | 1651 | prev = NEW_ARENA_ARRAY(area, Block *, size); |
rasbold@853 | 1652 | memset(prev , 0, size*sizeof(Block *)); |
rasbold@853 | 1653 | |
rasbold@853 | 1654 | // List of edges |
rasbold@853 | 1655 | edges = new GrowableArray<CFGEdge*>; |
rasbold@853 | 1656 | |
rasbold@853 | 1657 | // Mapping block index --> block_trace |
rasbold@853 | 1658 | uf = new UnionFind(size); |
rasbold@853 | 1659 | uf->reset(size); |
rasbold@853 | 1660 | |
rasbold@853 | 1661 | // Find edges and create traces. |
rasbold@853 | 1662 | find_edges(); |
rasbold@853 | 1663 | |
rasbold@853 | 1664 | // Grow traces at their ends via most frequent edges. |
rasbold@853 | 1665 | grow_traces(); |
rasbold@853 | 1666 | |
rasbold@853 | 1667 | // Merge one trace into another, but only at fall-through points. |
rasbold@853 | 1668 | // This may make diamonds and other related shapes in a trace. |
rasbold@853 | 1669 | merge_traces(true); |
rasbold@853 | 1670 | |
rasbold@853 | 1671 | // Run merge again, allowing two traces to be catenated, even if |
rasbold@853 | 1672 | // one does not fall through into the other. This appends loosely |
rasbold@853 | 1673 | // related traces to be near each other. |
rasbold@853 | 1674 | merge_traces(false); |
rasbold@853 | 1675 | |
rasbold@853 | 1676 | // Re-order all the remaining traces by frequency |
rasbold@853 | 1677 | reorder_traces(size); |
rasbold@853 | 1678 | |
adlertz@5539 | 1679 | assert(_cfg.number_of_blocks() >= (uint) (size - 1), "number of blocks can not shrink"); |
rasbold@853 | 1680 | } |
rasbold@853 | 1681 | |
rasbold@853 | 1682 | |
rasbold@853 | 1683 | // Edge e completes a loop in a trace. If the target block is head of the |
rasbold@853 | 1684 | // loop, rotate the loop block so that the loop ends in a conditional branch. |
rasbold@853 | 1685 | bool Trace::backedge(CFGEdge *e) { |
rasbold@853 | 1686 | bool loop_rotated = false; |
rasbold@853 | 1687 | Block *src_block = e->from(); |
rasbold@853 | 1688 | Block *targ_block = e->to(); |
rasbold@853 | 1689 | |
rasbold@853 | 1690 | assert(last_block() == src_block, "loop discovery at back branch"); |
rasbold@853 | 1691 | if (first_block() == targ_block) { |
rasbold@853 | 1692 | if (BlockLayoutRotateLoops && last_block()->num_fall_throughs() < 2) { |
rasbold@853 | 1693 | // Find the last block in the trace that has a conditional |
rasbold@853 | 1694 | // branch. |
rasbold@853 | 1695 | Block *b; |
rasbold@853 | 1696 | for (b = last_block(); b != NULL; b = prev(b)) { |
rasbold@853 | 1697 | if (b->num_fall_throughs() == 2) { |
rasbold@853 | 1698 | break; |
rasbold@853 | 1699 | } |
rasbold@853 | 1700 | } |
rasbold@853 | 1701 | |
rasbold@853 | 1702 | if (b != last_block() && b != NULL) { |
rasbold@853 | 1703 | loop_rotated = true; |
rasbold@853 | 1704 | |
rasbold@853 | 1705 | // Rotate the loop by doing two-part linked-list surgery. |
rasbold@853 | 1706 | append(first_block()); |
rasbold@853 | 1707 | break_loop_after(b); |
rasbold@853 | 1708 | } |
rasbold@853 | 1709 | } |
rasbold@853 | 1710 | |
rasbold@853 | 1711 | // Backbranch to the top of a trace |
twisti@1040 | 1712 | // Scroll forward through the trace from the targ_block. If we find |
rasbold@853 | 1713 | // a loop head before another loop top, use the the loop head alignment. |
rasbold@853 | 1714 | for (Block *b = targ_block; b != NULL; b = next(b)) { |
rasbold@853 | 1715 | if (b->has_loop_alignment()) { |
rasbold@853 | 1716 | break; |
rasbold@853 | 1717 | } |
rasbold@853 | 1718 | if (b->head()->is_Loop()) { |
rasbold@853 | 1719 | targ_block = b; |
rasbold@853 | 1720 | break; |
rasbold@853 | 1721 | } |
rasbold@853 | 1722 | } |
rasbold@853 | 1723 | |
rasbold@853 | 1724 | first_block()->set_loop_alignment(targ_block); |
rasbold@853 | 1725 | |
rasbold@853 | 1726 | } else { |
rasbold@853 | 1727 | // Backbranch into the middle of a trace |
rasbold@853 | 1728 | targ_block->set_loop_alignment(targ_block); |
rasbold@853 | 1729 | } |
rasbold@853 | 1730 | |
rasbold@853 | 1731 | return loop_rotated; |
rasbold@853 | 1732 | } |
rasbold@853 | 1733 | |
rasbold@853 | 1734 | // push blocks onto the CFG list |
rasbold@853 | 1735 | // ensure that blocks have the correct two-way branch sense |
rasbold@853 | 1736 | void Trace::fixup_blocks(PhaseCFG &cfg) { |
rasbold@853 | 1737 | Block *last = last_block(); |
rasbold@853 | 1738 | for (Block *b = first_block(); b != NULL; b = next(b)) { |
adlertz@5539 | 1739 | cfg.add_block(b); |
rasbold@853 | 1740 | if (!b->is_connector()) { |
rasbold@853 | 1741 | int nfallthru = b->num_fall_throughs(); |
rasbold@853 | 1742 | if (b != last) { |
rasbold@853 | 1743 | if (nfallthru == 2) { |
rasbold@853 | 1744 | // Ensure that the sense of the branch is correct |
rasbold@853 | 1745 | Block *bnext = next(b); |
rasbold@853 | 1746 | Block *bs0 = b->non_connector_successor(0); |
rasbold@853 | 1747 | |
adlertz@5635 | 1748 | MachNode *iff = b->get_node(b->number_of_nodes() - 3)->as_Mach(); |
adlertz@5635 | 1749 | ProjNode *proj0 = b->get_node(b->number_of_nodes() - 2)->as_Proj(); |
adlertz@5635 | 1750 | ProjNode *proj1 = b->get_node(b->number_of_nodes() - 1)->as_Proj(); |
rasbold@853 | 1751 | |
rasbold@853 | 1752 | if (bnext == bs0) { |
rasbold@853 | 1753 | // Fall-thru case in succs[0], should be in succs[1] |
rasbold@853 | 1754 | |
rasbold@853 | 1755 | // Flip targets in _succs map |
rasbold@853 | 1756 | Block *tbs0 = b->_succs[0]; |
rasbold@853 | 1757 | Block *tbs1 = b->_succs[1]; |
rasbold@853 | 1758 | b->_succs.map( 0, tbs1 ); |
rasbold@853 | 1759 | b->_succs.map( 1, tbs0 ); |
rasbold@853 | 1760 | |
rasbold@853 | 1761 | // Flip projections to match targets |
adlertz@5635 | 1762 | b->map_node(proj1, b->number_of_nodes() - 2); |
adlertz@5635 | 1763 | b->map_node(proj0, b->number_of_nodes() - 1); |
rasbold@853 | 1764 | } |
rasbold@853 | 1765 | } |
rasbold@853 | 1766 | } |
rasbold@853 | 1767 | } |
rasbold@853 | 1768 | } |
rasbold@853 | 1769 | } |