Wed, 18 Sep 2013 14:34:56 -0700
8024342: PPC64 (part 111): Support for C calling conventions that require 64-bit ints.
Summary: Some platforms, as ppc and s390x/zArch require that 32-bit ints are passed as 64-bit values to C functions. This change adds support to adapt the signature and to issue proper casts to c2-compiled stubs. The functions are used in generate_native_wrapper(). Adapt signature used by the compiler as in PhaseIdealLoop::intrinsify_fill().
Reviewed-by: kvn
duke@435 | 1 | /* |
mikael@4153 | 2 | * Copyright (c) 1997, 2012, Oracle and/or its affiliates. All rights reserved. |
duke@435 | 3 | * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. |
duke@435 | 4 | * |
duke@435 | 5 | * This code is free software; you can redistribute it and/or modify it |
duke@435 | 6 | * under the terms of the GNU General Public License version 2 only, as |
duke@435 | 7 | * published by the Free Software Foundation. |
duke@435 | 8 | * |
duke@435 | 9 | * This code is distributed in the hope that it will be useful, but WITHOUT |
duke@435 | 10 | * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or |
duke@435 | 11 | * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
duke@435 | 12 | * version 2 for more details (a copy is included in the LICENSE file that |
duke@435 | 13 | * accompanied this code). |
duke@435 | 14 | * |
duke@435 | 15 | * You should have received a copy of the GNU General Public License version |
duke@435 | 16 | * 2 along with this work; if not, write to the Free Software Foundation, |
duke@435 | 17 | * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. |
duke@435 | 18 | * |
trims@1907 | 19 | * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA |
trims@1907 | 20 | * or visit www.oracle.com if you need additional information or have any |
trims@1907 | 21 | * questions. |
duke@435 | 22 | * |
duke@435 | 23 | */ |
duke@435 | 24 | |
stefank@2314 | 25 | #include "precompiled.hpp" |
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) { |
duke@435 | 115 | uint last_inst = _nodes.size(); |
duke@435 | 116 | for( uint j = 0; j < last_inst && inst_cnt > 0; j++ ) { |
duke@435 | 117 | uint inst_size = _nodes[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 { |
duke@435 | 134 | for( uint i = 0; i < _nodes.size(); i++ ) { |
duke@435 | 135 | if( _nodes[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 ) { |
duke@435 | 144 | _nodes.remove(find_node(n)); |
duke@435 | 145 | } |
duke@435 | 146 | |
duke@435 | 147 | // Return empty status of a block. Empty blocks contain only the head, other |
duke@435 | 148 | // ideal nodes, and an optional trailing goto. |
duke@435 | 149 | int Block::is_Empty() const { |
duke@435 | 150 | |
duke@435 | 151 | // Root or start block is not considered empty |
duke@435 | 152 | if (head()->is_Root() || head()->is_Start()) { |
duke@435 | 153 | return not_empty; |
duke@435 | 154 | } |
duke@435 | 155 | |
duke@435 | 156 | int success_result = completely_empty; |
duke@435 | 157 | int end_idx = _nodes.size()-1; |
duke@435 | 158 | |
duke@435 | 159 | // Check for ending goto |
kvn@3040 | 160 | if ((end_idx > 0) && (_nodes[end_idx]->is_MachGoto())) { |
duke@435 | 161 | success_result = empty_with_goto; |
duke@435 | 162 | end_idx--; |
duke@435 | 163 | } |
duke@435 | 164 | |
duke@435 | 165 | // Unreachable blocks are considered empty |
duke@435 | 166 | if (num_preds() <= 1) { |
duke@435 | 167 | return success_result; |
duke@435 | 168 | } |
duke@435 | 169 | |
duke@435 | 170 | // Ideal nodes are allowable in empty blocks: skip them Only MachNodes |
duke@435 | 171 | // turn directly into code, because only MachNodes have non-trivial |
duke@435 | 172 | // emit() functions. |
duke@435 | 173 | while ((end_idx > 0) && !_nodes[end_idx]->is_Mach()) { |
duke@435 | 174 | end_idx--; |
duke@435 | 175 | } |
duke@435 | 176 | |
duke@435 | 177 | // No room for any interesting instructions? |
duke@435 | 178 | if (end_idx == 0) { |
duke@435 | 179 | return success_result; |
duke@435 | 180 | } |
duke@435 | 181 | |
duke@435 | 182 | return not_empty; |
duke@435 | 183 | } |
duke@435 | 184 | |
twisti@1040 | 185 | // Return true if the block's code implies that it is likely to be |
duke@435 | 186 | // executed infrequently. Check to see if the block ends in a Halt or |
duke@435 | 187 | // a low probability call. |
duke@435 | 188 | bool Block::has_uncommon_code() const { |
duke@435 | 189 | Node* en = end(); |
duke@435 | 190 | |
kvn@3040 | 191 | if (en->is_MachGoto()) |
duke@435 | 192 | en = en->in(0); |
duke@435 | 193 | if (en->is_Catch()) |
duke@435 | 194 | en = en->in(0); |
kvn@3040 | 195 | if (en->is_MachProj() && en->in(0)->is_MachCall()) { |
duke@435 | 196 | MachCallNode* call = en->in(0)->as_MachCall(); |
duke@435 | 197 | if (call->cnt() != COUNT_UNKNOWN && call->cnt() <= PROB_UNLIKELY_MAG(4)) { |
duke@435 | 198 | // This is true for slow-path stubs like new_{instance,array}, |
duke@435 | 199 | // slow_arraycopy, complete_monitor_locking, uncommon_trap. |
duke@435 | 200 | // The magic number corresponds to the probability of an uncommon_trap, |
duke@435 | 201 | // even though it is a count not a probability. |
duke@435 | 202 | return true; |
duke@435 | 203 | } |
duke@435 | 204 | } |
duke@435 | 205 | |
duke@435 | 206 | int op = en->is_Mach() ? en->as_Mach()->ideal_Opcode() : en->Opcode(); |
duke@435 | 207 | return op == Op_Halt; |
duke@435 | 208 | } |
duke@435 | 209 | |
duke@435 | 210 | // True if block is low enough frequency or guarded by a test which |
duke@435 | 211 | // mostly does not go here. |
adlertz@5509 | 212 | bool Block::is_uncommon(PhaseCFG* cfg) const { |
duke@435 | 213 | // Initial blocks must never be moved, so are never uncommon. |
duke@435 | 214 | if (head()->is_Root() || head()->is_Start()) return false; |
duke@435 | 215 | |
duke@435 | 216 | // Check for way-low freq |
duke@435 | 217 | if( _freq < BLOCK_FREQUENCY(0.00001f) ) return true; |
duke@435 | 218 | |
duke@435 | 219 | // Look for code shape indicating uncommon_trap or slow path |
duke@435 | 220 | if (has_uncommon_code()) return true; |
duke@435 | 221 | |
duke@435 | 222 | const float epsilon = 0.05f; |
duke@435 | 223 | const float guard_factor = PROB_UNLIKELY_MAG(4) / (1.f - epsilon); |
duke@435 | 224 | uint uncommon_preds = 0; |
duke@435 | 225 | uint freq_preds = 0; |
duke@435 | 226 | uint uncommon_for_freq_preds = 0; |
duke@435 | 227 | |
duke@435 | 228 | for( uint i=1; i<num_preds(); i++ ) { |
adlertz@5509 | 229 | Block* guard = cfg->get_block_for_node(pred(i)); |
duke@435 | 230 | // Check to see if this block follows its guard 1 time out of 10000 |
duke@435 | 231 | // or less. |
duke@435 | 232 | // |
duke@435 | 233 | // See list of magnitude-4 unlikely probabilities in cfgnode.hpp which |
duke@435 | 234 | // we intend to be "uncommon", such as slow-path TLE allocation, |
duke@435 | 235 | // predicted call failure, and uncommon trap triggers. |
duke@435 | 236 | // |
duke@435 | 237 | // Use an epsilon value of 5% to allow for variability in frequency |
duke@435 | 238 | // predictions and floating point calculations. The net effect is |
duke@435 | 239 | // that guard_factor is set to 9500. |
duke@435 | 240 | // |
duke@435 | 241 | // Ignore low-frequency blocks. |
duke@435 | 242 | // The next check is (guard->_freq < 1.e-5 * 9500.). |
duke@435 | 243 | if(guard->_freq*BLOCK_FREQUENCY(guard_factor) < BLOCK_FREQUENCY(0.00001f)) { |
duke@435 | 244 | uncommon_preds++; |
duke@435 | 245 | } else { |
duke@435 | 246 | freq_preds++; |
duke@435 | 247 | if( _freq < guard->_freq * guard_factor ) { |
duke@435 | 248 | uncommon_for_freq_preds++; |
duke@435 | 249 | } |
duke@435 | 250 | } |
duke@435 | 251 | } |
duke@435 | 252 | if( num_preds() > 1 && |
duke@435 | 253 | // The block is uncommon if all preds are uncommon or |
duke@435 | 254 | (uncommon_preds == (num_preds()-1) || |
duke@435 | 255 | // it is uncommon for all frequent preds. |
duke@435 | 256 | uncommon_for_freq_preds == freq_preds) ) { |
duke@435 | 257 | return true; |
duke@435 | 258 | } |
duke@435 | 259 | return false; |
duke@435 | 260 | } |
duke@435 | 261 | |
duke@435 | 262 | #ifndef PRODUCT |
kvn@3049 | 263 | void Block::dump_bidx(const Block* orig, outputStream* st) const { |
kvn@3049 | 264 | if (_pre_order) st->print("B%d",_pre_order); |
kvn@3049 | 265 | else st->print("N%d", head()->_idx); |
duke@435 | 266 | |
duke@435 | 267 | if (Verbose && orig != this) { |
duke@435 | 268 | // Dump the original block's idx |
kvn@3049 | 269 | st->print(" ("); |
kvn@3049 | 270 | orig->dump_bidx(orig, st); |
kvn@3049 | 271 | st->print(")"); |
duke@435 | 272 | } |
duke@435 | 273 | } |
duke@435 | 274 | |
adlertz@5509 | 275 | void Block::dump_pred(const PhaseCFG* cfg, Block* orig, outputStream* st) const { |
duke@435 | 276 | if (is_connector()) { |
duke@435 | 277 | for (uint i=1; i<num_preds(); i++) { |
adlertz@5509 | 278 | Block *p = cfg->get_block_for_node(pred(i)); |
adlertz@5509 | 279 | p->dump_pred(cfg, orig, st); |
duke@435 | 280 | } |
duke@435 | 281 | } else { |
kvn@3049 | 282 | dump_bidx(orig, st); |
kvn@3049 | 283 | st->print(" "); |
duke@435 | 284 | } |
duke@435 | 285 | } |
duke@435 | 286 | |
adlertz@5509 | 287 | void Block::dump_head(const PhaseCFG* cfg, outputStream* st) const { |
duke@435 | 288 | // Print the basic block |
kvn@3049 | 289 | dump_bidx(this, st); |
kvn@3049 | 290 | st->print(": #\t"); |
duke@435 | 291 | |
duke@435 | 292 | // Print the incoming CFG edges and the outgoing CFG edges |
duke@435 | 293 | for( uint i=0; i<_num_succs; i++ ) { |
kvn@3049 | 294 | non_connector_successor(i)->dump_bidx(_succs[i], st); |
kvn@3049 | 295 | st->print(" "); |
duke@435 | 296 | } |
kvn@3049 | 297 | st->print("<- "); |
duke@435 | 298 | if( head()->is_block_start() ) { |
duke@435 | 299 | for (uint i=1; i<num_preds(); i++) { |
duke@435 | 300 | Node *s = pred(i); |
adlertz@5509 | 301 | if (cfg != NULL) { |
adlertz@5509 | 302 | Block *p = cfg->get_block_for_node(s); |
adlertz@5509 | 303 | p->dump_pred(cfg, p, st); |
duke@435 | 304 | } else { |
duke@435 | 305 | while (!s->is_block_start()) |
duke@435 | 306 | s = s->in(0); |
kvn@3049 | 307 | st->print("N%d ", s->_idx ); |
duke@435 | 308 | } |
duke@435 | 309 | } |
adlertz@5509 | 310 | } else { |
kvn@3049 | 311 | st->print("BLOCK HEAD IS JUNK "); |
adlertz@5509 | 312 | } |
duke@435 | 313 | |
duke@435 | 314 | // Print loop, if any |
duke@435 | 315 | const Block *bhead = this; // Head of self-loop |
duke@435 | 316 | Node *bh = bhead->head(); |
adlertz@5509 | 317 | |
adlertz@5509 | 318 | if ((cfg != NULL) && bh->is_Loop() && !head()->is_Root()) { |
duke@435 | 319 | LoopNode *loop = bh->as_Loop(); |
adlertz@5509 | 320 | const Block *bx = cfg->get_block_for_node(loop->in(LoopNode::LoopBackControl)); |
duke@435 | 321 | while (bx->is_connector()) { |
adlertz@5509 | 322 | bx = cfg->get_block_for_node(bx->pred(1)); |
duke@435 | 323 | } |
kvn@3049 | 324 | st->print("\tLoop: B%d-B%d ", bhead->_pre_order, bx->_pre_order); |
duke@435 | 325 | // Dump any loop-specific bits, especially for CountedLoops. |
kvn@3049 | 326 | loop->dump_spec(st); |
rasbold@853 | 327 | } else if (has_loop_alignment()) { |
kvn@3049 | 328 | st->print(" top-of-loop"); |
duke@435 | 329 | } |
kvn@3049 | 330 | st->print(" Freq: %g",_freq); |
duke@435 | 331 | if( Verbose || WizardMode ) { |
kvn@3049 | 332 | st->print(" IDom: %d/#%d", _idom ? _idom->_pre_order : 0, _dom_depth); |
kvn@3049 | 333 | st->print(" RegPressure: %d",_reg_pressure); |
kvn@3049 | 334 | st->print(" IHRP Index: %d",_ihrp_index); |
kvn@3049 | 335 | st->print(" FRegPressure: %d",_freg_pressure); |
kvn@3049 | 336 | st->print(" FHRP Index: %d",_fhrp_index); |
duke@435 | 337 | } |
kvn@3049 | 338 | st->print_cr(""); |
duke@435 | 339 | } |
duke@435 | 340 | |
adlertz@5509 | 341 | void Block::dump() const { |
adlertz@5509 | 342 | dump(NULL); |
adlertz@5509 | 343 | } |
duke@435 | 344 | |
adlertz@5509 | 345 | void Block::dump(const PhaseCFG* cfg) const { |
adlertz@5509 | 346 | dump_head(cfg); |
adlertz@5509 | 347 | for (uint i=0; i< _nodes.size(); i++) { |
duke@435 | 348 | _nodes[i]->dump(); |
adlertz@5509 | 349 | } |
duke@435 | 350 | tty->print("\n"); |
duke@435 | 351 | } |
duke@435 | 352 | #endif |
duke@435 | 353 | |
adlertz@5509 | 354 | PhaseCFG::PhaseCFG(Arena* arena, RootNode* root, Matcher& matcher) |
adlertz@5509 | 355 | : Phase(CFG) |
adlertz@5509 | 356 | , _block_arena(arena) |
adlertz@5539 | 357 | , _root(root) |
adlertz@5539 | 358 | , _matcher(matcher) |
adlertz@5509 | 359 | , _node_to_block_mapping(arena) |
adlertz@5509 | 360 | , _node_latency(NULL) |
duke@435 | 361 | #ifndef PRODUCT |
adlertz@5509 | 362 | , _trace_opto_pipelining(TraceOptoPipelining || C->method_has_option("TraceOptoPipelining")) |
duke@435 | 363 | #endif |
kvn@1268 | 364 | #ifdef ASSERT |
adlertz@5509 | 365 | , _raw_oops(arena) |
kvn@1268 | 366 | #endif |
duke@435 | 367 | { |
duke@435 | 368 | ResourceMark rm; |
duke@435 | 369 | // I'll need a few machine-specific GotoNodes. Make an Ideal GotoNode, |
duke@435 | 370 | // then Match it into a machine-specific Node. Then clone the machine |
duke@435 | 371 | // Node on demand. |
kvn@4115 | 372 | Node *x = new (C) GotoNode(NULL); |
duke@435 | 373 | x->init_req(0, x); |
adlertz@5509 | 374 | _goto = matcher.match_tree(x); |
duke@435 | 375 | assert(_goto != NULL, ""); |
duke@435 | 376 | _goto->set_req(0,_goto); |
duke@435 | 377 | |
duke@435 | 378 | // Build the CFG in Reverse Post Order |
adlertz@5539 | 379 | _number_of_blocks = build_cfg(); |
adlertz@5539 | 380 | _root_block = get_block_for_node(_root); |
duke@435 | 381 | } |
duke@435 | 382 | |
duke@435 | 383 | // Build a proper looking CFG. Make every block begin with either a StartNode |
duke@435 | 384 | // or a RegionNode. Make every block end with either a Goto, If or Return. |
duke@435 | 385 | // The RootNode both starts and ends it's own block. Do this with a recursive |
duke@435 | 386 | // backwards walk over the control edges. |
duke@435 | 387 | uint PhaseCFG::build_cfg() { |
duke@435 | 388 | Arena *a = Thread::current()->resource_area(); |
duke@435 | 389 | VectorSet visited(a); |
duke@435 | 390 | |
duke@435 | 391 | // Allocate stack with enough space to avoid frequent realloc |
duke@435 | 392 | Node_Stack nstack(a, C->unique() >> 1); |
duke@435 | 393 | nstack.push(_root, 0); |
duke@435 | 394 | uint sum = 0; // Counter for blocks |
duke@435 | 395 | |
duke@435 | 396 | while (nstack.is_nonempty()) { |
duke@435 | 397 | // node and in's index from stack's top |
duke@435 | 398 | // 'np' is _root (see above) or RegionNode, StartNode: we push on stack |
duke@435 | 399 | // only nodes which point to the start of basic block (see below). |
duke@435 | 400 | Node *np = nstack.node(); |
duke@435 | 401 | // idx > 0, except for the first node (_root) pushed on stack |
duke@435 | 402 | // at the beginning when idx == 0. |
duke@435 | 403 | // We will use the condition (idx == 0) later to end the build. |
duke@435 | 404 | uint idx = nstack.index(); |
duke@435 | 405 | Node *proj = np->in(idx); |
duke@435 | 406 | const Node *x = proj->is_block_proj(); |
duke@435 | 407 | // Does the block end with a proper block-ending Node? One of Return, |
duke@435 | 408 | // If or Goto? (This check should be done for visited nodes also). |
duke@435 | 409 | if (x == NULL) { // Does not end right... |
duke@435 | 410 | Node *g = _goto->clone(); // Force it to end in a Goto |
duke@435 | 411 | g->set_req(0, proj); |
duke@435 | 412 | np->set_req(idx, g); |
duke@435 | 413 | x = proj = g; |
duke@435 | 414 | } |
duke@435 | 415 | if (!visited.test_set(x->_idx)) { // Visit this block once |
duke@435 | 416 | // Skip any control-pinned middle'in stuff |
duke@435 | 417 | Node *p = proj; |
duke@435 | 418 | do { |
duke@435 | 419 | proj = p; // Update pointer to last Control |
duke@435 | 420 | p = p->in(0); // Move control forward |
duke@435 | 421 | } while( !p->is_block_proj() && |
duke@435 | 422 | !p->is_block_start() ); |
duke@435 | 423 | // Make the block begin with one of Region or StartNode. |
duke@435 | 424 | if( !p->is_block_start() ) { |
kvn@4115 | 425 | RegionNode *r = new (C) RegionNode( 2 ); |
duke@435 | 426 | r->init_req(1, p); // Insert RegionNode in the way |
duke@435 | 427 | proj->set_req(0, r); // Insert RegionNode in the way |
duke@435 | 428 | p = r; |
duke@435 | 429 | } |
duke@435 | 430 | // 'p' now points to the start of this basic block |
duke@435 | 431 | |
duke@435 | 432 | // Put self in array of basic blocks |
adlertz@5509 | 433 | Block *bb = new (_block_arena) Block(_block_arena, p); |
adlertz@5509 | 434 | map_node_to_block(p, bb); |
adlertz@5509 | 435 | map_node_to_block(x, bb); |
kvn@3049 | 436 | if( x != p ) { // Only for root is x == p |
duke@435 | 437 | bb->_nodes.push((Node*)x); |
kvn@3049 | 438 | } |
duke@435 | 439 | // Now handle predecessors |
duke@435 | 440 | ++sum; // Count 1 for self block |
duke@435 | 441 | uint cnt = bb->num_preds(); |
duke@435 | 442 | for (int i = (cnt - 1); i > 0; i-- ) { // For all predecessors |
duke@435 | 443 | Node *prevproj = p->in(i); // Get prior input |
duke@435 | 444 | assert( !prevproj->is_Con(), "dead input not removed" ); |
duke@435 | 445 | // Check to see if p->in(i) is a "control-dependent" CFG edge - |
duke@435 | 446 | // i.e., it splits at the source (via an IF or SWITCH) and merges |
duke@435 | 447 | // at the destination (via a many-input Region). |
duke@435 | 448 | // This breaks critical edges. The RegionNode to start the block |
duke@435 | 449 | // will be added when <p,i> is pulled off the node stack |
duke@435 | 450 | if ( cnt > 2 ) { // Merging many things? |
duke@435 | 451 | assert( prevproj== bb->pred(i),""); |
duke@435 | 452 | if(prevproj->is_block_proj() != prevproj) { // Control-dependent edge? |
duke@435 | 453 | // Force a block on the control-dependent edge |
duke@435 | 454 | Node *g = _goto->clone(); // Force it to end in a Goto |
duke@435 | 455 | g->set_req(0,prevproj); |
duke@435 | 456 | p->set_req(i,g); |
duke@435 | 457 | } |
duke@435 | 458 | } |
duke@435 | 459 | nstack.push(p, i); // 'p' is RegionNode or StartNode |
duke@435 | 460 | } |
duke@435 | 461 | } else { // Post-processing visited nodes |
duke@435 | 462 | nstack.pop(); // remove node from stack |
duke@435 | 463 | // Check if it the fist node pushed on stack at the beginning. |
duke@435 | 464 | if (idx == 0) break; // end of the build |
duke@435 | 465 | // Find predecessor basic block |
adlertz@5509 | 466 | Block *pb = get_block_for_node(x); |
duke@435 | 467 | // Insert into nodes array, if not already there |
adlertz@5509 | 468 | if (!has_block(proj)) { |
duke@435 | 469 | assert( x != proj, "" ); |
duke@435 | 470 | // Map basic block of projection |
adlertz@5509 | 471 | map_node_to_block(proj, pb); |
duke@435 | 472 | pb->_nodes.push(proj); |
duke@435 | 473 | } |
duke@435 | 474 | // Insert self as a child of my predecessor block |
adlertz@5509 | 475 | pb->_succs.map(pb->_num_succs++, get_block_for_node(np)); |
duke@435 | 476 | assert( pb->_nodes[ pb->_nodes.size() - pb->_num_succs ]->is_block_proj(), |
duke@435 | 477 | "too many control users, not a CFG?" ); |
duke@435 | 478 | } |
duke@435 | 479 | } |
duke@435 | 480 | // Return number of basic blocks for all children and self |
duke@435 | 481 | return sum; |
duke@435 | 482 | } |
duke@435 | 483 | |
duke@435 | 484 | // Inserts a goto & corresponding basic block between |
duke@435 | 485 | // block[block_no] and its succ_no'th successor block |
duke@435 | 486 | void PhaseCFG::insert_goto_at(uint block_no, uint succ_no) { |
duke@435 | 487 | // get block with block_no |
adlertz@5539 | 488 | assert(block_no < number_of_blocks(), "illegal block number"); |
adlertz@5539 | 489 | Block* in = get_block(block_no); |
duke@435 | 490 | // get successor block succ_no |
duke@435 | 491 | assert(succ_no < in->_num_succs, "illegal successor number"); |
duke@435 | 492 | Block* out = in->_succs[succ_no]; |
rasbold@743 | 493 | // Compute frequency of the new block. Do this before inserting |
rasbold@743 | 494 | // new block in case succ_prob() needs to infer the probability from |
rasbold@743 | 495 | // surrounding blocks. |
rasbold@743 | 496 | float freq = in->_freq * in->succ_prob(succ_no); |
duke@435 | 497 | // get ProjNode corresponding to the succ_no'th successor of the in block |
duke@435 | 498 | ProjNode* proj = in->_nodes[in->_nodes.size() - in->_num_succs + succ_no]->as_Proj(); |
duke@435 | 499 | // create region for basic block |
kvn@4115 | 500 | RegionNode* region = new (C) RegionNode(2); |
duke@435 | 501 | region->init_req(1, proj); |
duke@435 | 502 | // setup corresponding basic block |
adlertz@5509 | 503 | Block* block = new (_block_arena) Block(_block_arena, region); |
adlertz@5509 | 504 | map_node_to_block(region, block); |
duke@435 | 505 | C->regalloc()->set_bad(region->_idx); |
duke@435 | 506 | // add a goto node |
duke@435 | 507 | Node* gto = _goto->clone(); // get a new goto node |
duke@435 | 508 | gto->set_req(0, region); |
duke@435 | 509 | // add it to the basic block |
duke@435 | 510 | block->_nodes.push(gto); |
adlertz@5509 | 511 | map_node_to_block(gto, block); |
duke@435 | 512 | C->regalloc()->set_bad(gto->_idx); |
duke@435 | 513 | // hook up successor block |
duke@435 | 514 | block->_succs.map(block->_num_succs++, out); |
duke@435 | 515 | // remap successor's predecessors if necessary |
duke@435 | 516 | for (uint i = 1; i < out->num_preds(); i++) { |
duke@435 | 517 | if (out->pred(i) == proj) out->head()->set_req(i, gto); |
duke@435 | 518 | } |
duke@435 | 519 | // remap predecessor's successor to new block |
duke@435 | 520 | in->_succs.map(succ_no, block); |
rasbold@743 | 521 | // Set the frequency of the new block |
rasbold@743 | 522 | block->_freq = freq; |
duke@435 | 523 | // add new basic block to basic block list |
adlertz@5539 | 524 | add_block_at(block_no + 1, block); |
duke@435 | 525 | } |
duke@435 | 526 | |
duke@435 | 527 | // Does this block end in a multiway branch that cannot have the default case |
duke@435 | 528 | // flipped for another case? |
duke@435 | 529 | static bool no_flip_branch( Block *b ) { |
duke@435 | 530 | int branch_idx = b->_nodes.size() - b->_num_succs-1; |
duke@435 | 531 | if( branch_idx < 1 ) return false; |
duke@435 | 532 | Node *bra = b->_nodes[branch_idx]; |
rasbold@853 | 533 | if( bra->is_Catch() ) |
rasbold@853 | 534 | return true; |
duke@435 | 535 | if( bra->is_Mach() ) { |
rasbold@853 | 536 | if( bra->is_MachNullCheck() ) |
rasbold@853 | 537 | return true; |
duke@435 | 538 | int iop = bra->as_Mach()->ideal_Opcode(); |
duke@435 | 539 | if( iop == Op_FastLock || iop == Op_FastUnlock ) |
duke@435 | 540 | return true; |
duke@435 | 541 | } |
duke@435 | 542 | return false; |
duke@435 | 543 | } |
duke@435 | 544 | |
duke@435 | 545 | // Check for NeverBranch at block end. This needs to become a GOTO to the |
duke@435 | 546 | // true target. NeverBranch are treated as a conditional branch that always |
duke@435 | 547 | // goes the same direction for most of the optimizer and are used to give a |
duke@435 | 548 | // fake exit path to infinite loops. At this late stage they need to turn |
duke@435 | 549 | // into Goto's so that when you enter the infinite loop you indeed hang. |
duke@435 | 550 | void PhaseCFG::convert_NeverBranch_to_Goto(Block *b) { |
duke@435 | 551 | // Find true target |
duke@435 | 552 | int end_idx = b->end_idx(); |
duke@435 | 553 | int idx = b->_nodes[end_idx+1]->as_Proj()->_con; |
duke@435 | 554 | Block *succ = b->_succs[idx]; |
duke@435 | 555 | Node* gto = _goto->clone(); // get a new goto node |
duke@435 | 556 | gto->set_req(0, b->head()); |
duke@435 | 557 | Node *bp = b->_nodes[end_idx]; |
duke@435 | 558 | b->_nodes.map(end_idx,gto); // Slam over NeverBranch |
adlertz@5509 | 559 | map_node_to_block(gto, b); |
duke@435 | 560 | C->regalloc()->set_bad(gto->_idx); |
duke@435 | 561 | b->_nodes.pop(); // Yank projections |
duke@435 | 562 | b->_nodes.pop(); // Yank projections |
duke@435 | 563 | b->_succs.map(0,succ); // Map only successor |
duke@435 | 564 | b->_num_succs = 1; |
duke@435 | 565 | // remap successor's predecessors if necessary |
duke@435 | 566 | uint j; |
duke@435 | 567 | for( j = 1; j < succ->num_preds(); j++) |
duke@435 | 568 | if( succ->pred(j)->in(0) == bp ) |
duke@435 | 569 | succ->head()->set_req(j, gto); |
duke@435 | 570 | // Kill alternate exit path |
duke@435 | 571 | Block *dead = b->_succs[1-idx]; |
duke@435 | 572 | for( j = 1; j < dead->num_preds(); j++) |
duke@435 | 573 | if( dead->pred(j)->in(0) == bp ) |
duke@435 | 574 | break; |
duke@435 | 575 | // Scan through block, yanking dead path from |
duke@435 | 576 | // all regions and phis. |
duke@435 | 577 | dead->head()->del_req(j); |
duke@435 | 578 | for( int k = 1; dead->_nodes[k]->is_Phi(); k++ ) |
duke@435 | 579 | dead->_nodes[k]->del_req(j); |
duke@435 | 580 | } |
duke@435 | 581 | |
duke@435 | 582 | // Helper function to move block bx to the slot following b_index. Return |
duke@435 | 583 | // true if the move is successful, otherwise false |
rasbold@853 | 584 | bool PhaseCFG::move_to_next(Block* bx, uint b_index) { |
duke@435 | 585 | if (bx == NULL) return false; |
duke@435 | 586 | |
duke@435 | 587 | // Return false if bx is already scheduled. |
duke@435 | 588 | uint bx_index = bx->_pre_order; |
adlertz@5539 | 589 | if ((bx_index <= b_index) && (get_block(bx_index) == bx)) { |
duke@435 | 590 | return false; |
duke@435 | 591 | } |
duke@435 | 592 | |
duke@435 | 593 | // Find the current index of block bx on the block list |
duke@435 | 594 | bx_index = b_index + 1; |
adlertz@5539 | 595 | while (bx_index < number_of_blocks() && get_block(bx_index) != bx) { |
adlertz@5539 | 596 | bx_index++; |
adlertz@5539 | 597 | } |
adlertz@5539 | 598 | assert(get_block(bx_index) == bx, "block not found"); |
duke@435 | 599 | |
duke@435 | 600 | // If the previous block conditionally falls into bx, return false, |
duke@435 | 601 | // because moving bx will create an extra jump. |
duke@435 | 602 | for(uint k = 1; k < bx->num_preds(); k++ ) { |
adlertz@5509 | 603 | Block* pred = get_block_for_node(bx->pred(k)); |
adlertz@5539 | 604 | if (pred == get_block(bx_index - 1)) { |
duke@435 | 605 | if (pred->_num_succs != 1) { |
duke@435 | 606 | return false; |
duke@435 | 607 | } |
duke@435 | 608 | } |
duke@435 | 609 | } |
duke@435 | 610 | |
duke@435 | 611 | // Reinsert bx just past block 'b' |
duke@435 | 612 | _blocks.remove(bx_index); |
duke@435 | 613 | _blocks.insert(b_index + 1, bx); |
duke@435 | 614 | return true; |
duke@435 | 615 | } |
duke@435 | 616 | |
duke@435 | 617 | // Move empty and uncommon blocks to the end. |
rasbold@853 | 618 | void PhaseCFG::move_to_end(Block *b, uint i) { |
duke@435 | 619 | int e = b->is_Empty(); |
duke@435 | 620 | if (e != Block::not_empty) { |
duke@435 | 621 | if (e == Block::empty_with_goto) { |
duke@435 | 622 | // Remove the goto, but leave the block. |
duke@435 | 623 | b->_nodes.pop(); |
duke@435 | 624 | } |
duke@435 | 625 | // Mark this block as a connector block, which will cause it to be |
duke@435 | 626 | // ignored in certain functions such as non_connector_successor(). |
duke@435 | 627 | b->set_connector(); |
duke@435 | 628 | } |
duke@435 | 629 | // Move the empty block to the end, and don't recheck. |
duke@435 | 630 | _blocks.remove(i); |
duke@435 | 631 | _blocks.push(b); |
duke@435 | 632 | } |
duke@435 | 633 | |
rasbold@853 | 634 | // Set loop alignment for every block |
rasbold@853 | 635 | void PhaseCFG::set_loop_alignment() { |
adlertz@5539 | 636 | uint last = number_of_blocks(); |
adlertz@5539 | 637 | assert(get_block(0) == get_root_block(), ""); |
rasbold@853 | 638 | |
adlertz@5539 | 639 | for (uint i = 1; i < last; i++) { |
adlertz@5539 | 640 | Block* block = get_block(i); |
adlertz@5539 | 641 | if (block->head()->is_Loop()) { |
adlertz@5539 | 642 | block->set_loop_alignment(block); |
rasbold@853 | 643 | } |
rasbold@853 | 644 | } |
rasbold@853 | 645 | } |
rasbold@853 | 646 | |
rasbold@853 | 647 | // Make empty basic blocks to be "connector" blocks, Move uncommon blocks |
rasbold@853 | 648 | // to the end. |
adlertz@5539 | 649 | void PhaseCFG::remove_empty_blocks() { |
duke@435 | 650 | // Move uncommon blocks to the end |
adlertz@5539 | 651 | uint last = number_of_blocks(); |
adlertz@5539 | 652 | assert(get_block(0) == get_root_block(), ""); |
rasbold@853 | 653 | |
rasbold@853 | 654 | for (uint i = 1; i < last; i++) { |
adlertz@5539 | 655 | Block* block = get_block(i); |
adlertz@5539 | 656 | if (block->is_connector()) { |
adlertz@5539 | 657 | break; |
adlertz@5539 | 658 | } |
duke@435 | 659 | |
duke@435 | 660 | // Check for NeverBranch at block end. This needs to become a GOTO to the |
duke@435 | 661 | // true target. NeverBranch are treated as a conditional branch that |
duke@435 | 662 | // always goes the same direction for most of the optimizer and are used |
duke@435 | 663 | // to give a fake exit path to infinite loops. At this late stage they |
duke@435 | 664 | // need to turn into Goto's so that when you enter the infinite loop you |
duke@435 | 665 | // indeed hang. |
adlertz@5539 | 666 | if (block->_nodes[block->end_idx()]->Opcode() == Op_NeverBranch) { |
adlertz@5539 | 667 | convert_NeverBranch_to_Goto(block); |
adlertz@5539 | 668 | } |
duke@435 | 669 | |
duke@435 | 670 | // Look for uncommon blocks and move to end. |
rasbold@853 | 671 | if (!C->do_freq_based_layout()) { |
adlertz@5539 | 672 | if (block->is_uncommon(this)) { |
adlertz@5539 | 673 | move_to_end(block, i); |
rasbold@853 | 674 | last--; // No longer check for being uncommon! |
adlertz@5539 | 675 | if (no_flip_branch(block)) { // Fall-thru case must follow? |
adlertz@5539 | 676 | // Find the fall-thru block |
adlertz@5539 | 677 | block = get_block(i); |
adlertz@5539 | 678 | move_to_end(block, i); |
rasbold@853 | 679 | last--; |
rasbold@853 | 680 | } |
adlertz@5539 | 681 | // backup block counter post-increment |
adlertz@5539 | 682 | i--; |
duke@435 | 683 | } |
duke@435 | 684 | } |
duke@435 | 685 | } |
duke@435 | 686 | |
rasbold@853 | 687 | // Move empty blocks to the end |
adlertz@5539 | 688 | last = number_of_blocks(); |
rasbold@853 | 689 | for (uint i = 1; i < last; i++) { |
adlertz@5539 | 690 | Block* block = get_block(i); |
adlertz@5539 | 691 | if (block->is_Empty() != Block::not_empty) { |
adlertz@5539 | 692 | move_to_end(block, i); |
rasbold@853 | 693 | last--; |
rasbold@853 | 694 | i--; |
duke@435 | 695 | } |
duke@435 | 696 | } // End of for all blocks |
rasbold@853 | 697 | } |
duke@435 | 698 | |
rasbold@853 | 699 | // Fix up the final control flow for basic blocks. |
rasbold@853 | 700 | void PhaseCFG::fixup_flow() { |
duke@435 | 701 | // Fixup final control flow for the blocks. Remove jump-to-next |
duke@435 | 702 | // block. If neither arm of a IF follows the conditional branch, we |
duke@435 | 703 | // have to add a second jump after the conditional. We place the |
duke@435 | 704 | // TRUE branch target in succs[0] for both GOTOs and IFs. |
adlertz@5539 | 705 | for (uint i = 0; i < number_of_blocks(); i++) { |
adlertz@5539 | 706 | Block* block = get_block(i); |
adlertz@5539 | 707 | block->_pre_order = i; // turn pre-order into block-index |
duke@435 | 708 | |
duke@435 | 709 | // Connector blocks need no further processing. |
adlertz@5539 | 710 | if (block->is_connector()) { |
adlertz@5539 | 711 | assert((i+1) == number_of_blocks() || get_block(i + 1)->is_connector(), "All connector blocks should sink to the end"); |
duke@435 | 712 | continue; |
duke@435 | 713 | } |
adlertz@5539 | 714 | assert(block->is_Empty() != Block::completely_empty, "Empty blocks should be connectors"); |
duke@435 | 715 | |
adlertz@5539 | 716 | Block* bnext = (i < number_of_blocks() - 1) ? get_block(i + 1) : NULL; |
adlertz@5539 | 717 | Block* bs0 = block->non_connector_successor(0); |
duke@435 | 718 | |
duke@435 | 719 | // Check for multi-way branches where I cannot negate the test to |
duke@435 | 720 | // exchange the true and false targets. |
adlertz@5539 | 721 | if (no_flip_branch(block)) { |
duke@435 | 722 | // Find fall through case - if must fall into its target |
adlertz@5539 | 723 | int branch_idx = block->_nodes.size() - block->_num_succs; |
adlertz@5539 | 724 | for (uint j2 = 0; j2 < block->_num_succs; j2++) { |
adlertz@5539 | 725 | const ProjNode* p = block->_nodes[branch_idx + j2]->as_Proj(); |
duke@435 | 726 | if (p->_con == 0) { |
duke@435 | 727 | // successor j2 is fall through case |
adlertz@5539 | 728 | if (block->non_connector_successor(j2) != bnext) { |
duke@435 | 729 | // but it is not the next block => insert a goto |
duke@435 | 730 | insert_goto_at(i, j2); |
duke@435 | 731 | } |
duke@435 | 732 | // Put taken branch in slot 0 |
adlertz@5539 | 733 | if (j2 == 0 && block->_num_succs == 2) { |
duke@435 | 734 | // Flip targets in succs map |
adlertz@5539 | 735 | Block *tbs0 = block->_succs[0]; |
adlertz@5539 | 736 | Block *tbs1 = block->_succs[1]; |
adlertz@5539 | 737 | block->_succs.map(0, tbs1); |
adlertz@5539 | 738 | block->_succs.map(1, tbs0); |
duke@435 | 739 | } |
duke@435 | 740 | break; |
duke@435 | 741 | } |
duke@435 | 742 | } |
adlertz@5539 | 743 | |
duke@435 | 744 | // Remove all CatchProjs |
adlertz@5539 | 745 | for (uint j = 0; j < block->_num_succs; j++) { |
adlertz@5539 | 746 | block->_nodes.pop(); |
adlertz@5539 | 747 | } |
duke@435 | 748 | |
adlertz@5539 | 749 | } else if (block->_num_succs == 1) { |
duke@435 | 750 | // Block ends in a Goto? |
duke@435 | 751 | if (bnext == bs0) { |
duke@435 | 752 | // We fall into next block; remove the Goto |
adlertz@5539 | 753 | block->_nodes.pop(); |
duke@435 | 754 | } |
duke@435 | 755 | |
adlertz@5539 | 756 | } else if(block->_num_succs == 2) { // Block ends in a If? |
duke@435 | 757 | // Get opcode of 1st projection (matches _succs[0]) |
duke@435 | 758 | // Note: Since this basic block has 2 exits, the last 2 nodes must |
duke@435 | 759 | // be projections (in any order), the 3rd last node must be |
duke@435 | 760 | // the IfNode (we have excluded other 2-way exits such as |
duke@435 | 761 | // CatchNodes already). |
adlertz@5539 | 762 | MachNode* iff = block->_nodes[block->_nodes.size() - 3]->as_Mach(); |
adlertz@5539 | 763 | ProjNode* proj0 = block->_nodes[block->_nodes.size() - 2]->as_Proj(); |
adlertz@5539 | 764 | ProjNode* proj1 = block->_nodes[block->_nodes.size() - 1]->as_Proj(); |
duke@435 | 765 | |
duke@435 | 766 | // Assert that proj0 and succs[0] match up. Similarly for proj1 and succs[1]. |
adlertz@5539 | 767 | assert(proj0->raw_out(0) == block->_succs[0]->head(), "Mismatch successor 0"); |
adlertz@5539 | 768 | assert(proj1->raw_out(0) == block->_succs[1]->head(), "Mismatch successor 1"); |
duke@435 | 769 | |
adlertz@5539 | 770 | Block* bs1 = block->non_connector_successor(1); |
duke@435 | 771 | |
duke@435 | 772 | // Check for neither successor block following the current |
duke@435 | 773 | // block ending in a conditional. If so, move one of the |
duke@435 | 774 | // successors after the current one, provided that the |
duke@435 | 775 | // successor was previously unscheduled, but moveable |
duke@435 | 776 | // (i.e., all paths to it involve a branch). |
adlertz@5539 | 777 | if (!C->do_freq_based_layout() && bnext != bs0 && bnext != bs1) { |
duke@435 | 778 | // Choose the more common successor based on the probability |
duke@435 | 779 | // of the conditional branch. |
adlertz@5539 | 780 | Block* bx = bs0; |
adlertz@5539 | 781 | Block* by = bs1; |
duke@435 | 782 | |
duke@435 | 783 | // _prob is the probability of taking the true path. Make |
duke@435 | 784 | // p the probability of taking successor #1. |
duke@435 | 785 | float p = iff->as_MachIf()->_prob; |
adlertz@5539 | 786 | if (proj0->Opcode() == Op_IfTrue) { |
duke@435 | 787 | p = 1.0 - p; |
duke@435 | 788 | } |
duke@435 | 789 | |
duke@435 | 790 | // Prefer successor #1 if p > 0.5 |
duke@435 | 791 | if (p > PROB_FAIR) { |
duke@435 | 792 | bx = bs1; |
duke@435 | 793 | by = bs0; |
duke@435 | 794 | } |
duke@435 | 795 | |
duke@435 | 796 | // Attempt the more common successor first |
rasbold@853 | 797 | if (move_to_next(bx, i)) { |
duke@435 | 798 | bnext = bx; |
rasbold@853 | 799 | } else if (move_to_next(by, i)) { |
duke@435 | 800 | bnext = by; |
duke@435 | 801 | } |
duke@435 | 802 | } |
duke@435 | 803 | |
duke@435 | 804 | // Check for conditional branching the wrong way. Negate |
duke@435 | 805 | // conditional, if needed, so it falls into the following block |
duke@435 | 806 | // and branches to the not-following block. |
duke@435 | 807 | |
duke@435 | 808 | // Check for the next block being in succs[0]. We are going to branch |
duke@435 | 809 | // to succs[0], so we want the fall-thru case as the next block in |
duke@435 | 810 | // succs[1]. |
duke@435 | 811 | if (bnext == bs0) { |
duke@435 | 812 | // Fall-thru case in succs[0], so flip targets in succs map |
adlertz@5539 | 813 | Block* tbs0 = block->_succs[0]; |
adlertz@5539 | 814 | Block* tbs1 = block->_succs[1]; |
adlertz@5539 | 815 | block->_succs.map(0, tbs1); |
adlertz@5539 | 816 | block->_succs.map(1, tbs0); |
duke@435 | 817 | // Flip projection for each target |
adlertz@5539 | 818 | ProjNode* tmp = proj0; |
adlertz@5539 | 819 | proj0 = proj1; |
adlertz@5539 | 820 | proj1 = tmp; |
duke@435 | 821 | |
adlertz@5539 | 822 | } else if(bnext != bs1) { |
rasbold@853 | 823 | // Need a double-branch |
duke@435 | 824 | // The existing conditional branch need not change. |
duke@435 | 825 | // Add a unconditional branch to the false target. |
duke@435 | 826 | // Alas, it must appear in its own block and adding a |
duke@435 | 827 | // block this late in the game is complicated. Sigh. |
duke@435 | 828 | insert_goto_at(i, 1); |
duke@435 | 829 | } |
duke@435 | 830 | |
duke@435 | 831 | // Make sure we TRUE branch to the target |
adlertz@5539 | 832 | if (proj0->Opcode() == Op_IfFalse) { |
kvn@3051 | 833 | iff->as_MachIf()->negate(); |
rasbold@853 | 834 | } |
duke@435 | 835 | |
adlertz@5539 | 836 | block->_nodes.pop(); // Remove IfFalse & IfTrue projections |
adlertz@5539 | 837 | block->_nodes.pop(); |
duke@435 | 838 | |
duke@435 | 839 | } else { |
duke@435 | 840 | // Multi-exit block, e.g. a switch statement |
duke@435 | 841 | // But we don't need to do anything here |
duke@435 | 842 | } |
duke@435 | 843 | } // End of for all blocks |
duke@435 | 844 | } |
duke@435 | 845 | |
duke@435 | 846 | |
duke@435 | 847 | #ifndef PRODUCT |
duke@435 | 848 | void PhaseCFG::_dump_cfg( const Node *end, VectorSet &visited ) const { |
duke@435 | 849 | const Node *x = end->is_block_proj(); |
duke@435 | 850 | assert( x, "not a CFG" ); |
duke@435 | 851 | |
duke@435 | 852 | // Do not visit this block again |
duke@435 | 853 | if( visited.test_set(x->_idx) ) return; |
duke@435 | 854 | |
duke@435 | 855 | // Skip through this block |
duke@435 | 856 | const Node *p = x; |
duke@435 | 857 | do { |
duke@435 | 858 | p = p->in(0); // Move control forward |
duke@435 | 859 | assert( !p->is_block_proj() || p->is_Root(), "not a CFG" ); |
duke@435 | 860 | } while( !p->is_block_start() ); |
duke@435 | 861 | |
duke@435 | 862 | // Recursively visit |
adlertz@5509 | 863 | for (uint i = 1; i < p->req(); i++) { |
adlertz@5509 | 864 | _dump_cfg(p->in(i), visited); |
adlertz@5509 | 865 | } |
duke@435 | 866 | |
duke@435 | 867 | // Dump the block |
adlertz@5509 | 868 | get_block_for_node(p)->dump(this); |
duke@435 | 869 | } |
duke@435 | 870 | |
duke@435 | 871 | void PhaseCFG::dump( ) const { |
adlertz@5539 | 872 | tty->print("\n--- CFG --- %d BBs\n", number_of_blocks()); |
adlertz@5509 | 873 | if (_blocks.size()) { // Did we do basic-block layout? |
adlertz@5539 | 874 | for (uint i = 0; i < number_of_blocks(); i++) { |
adlertz@5539 | 875 | const Block* block = get_block(i); |
adlertz@5539 | 876 | block->dump(this); |
adlertz@5509 | 877 | } |
duke@435 | 878 | } else { // Else do it with a DFS |
adlertz@5509 | 879 | VectorSet visited(_block_arena); |
duke@435 | 880 | _dump_cfg(_root,visited); |
duke@435 | 881 | } |
duke@435 | 882 | } |
duke@435 | 883 | |
duke@435 | 884 | void PhaseCFG::dump_headers() { |
adlertz@5539 | 885 | for (uint i = 0; i < number_of_blocks(); i++) { |
adlertz@5539 | 886 | Block* block = get_block(i); |
adlertz@5539 | 887 | if (block != NULL) { |
adlertz@5539 | 888 | block->dump_head(this); |
adlertz@5509 | 889 | } |
duke@435 | 890 | } |
duke@435 | 891 | } |
duke@435 | 892 | |
adlertz@5539 | 893 | void PhaseCFG::verify() const { |
kvn@1001 | 894 | #ifdef ASSERT |
duke@435 | 895 | // Verify sane CFG |
adlertz@5539 | 896 | for (uint i = 0; i < number_of_blocks(); i++) { |
adlertz@5539 | 897 | Block* block = get_block(i); |
adlertz@5539 | 898 | uint cnt = block->_nodes.size(); |
duke@435 | 899 | uint j; |
kvn@3311 | 900 | for (j = 0; j < cnt; j++) { |
adlertz@5539 | 901 | Node *n = block->_nodes[j]; |
adlertz@5539 | 902 | assert(get_block_for_node(n) == block, ""); |
adlertz@5539 | 903 | if (j >= 1 && n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_CreateEx) { |
adlertz@5539 | 904 | assert(j == 1 || block->_nodes[j-1]->is_Phi(), "CreateEx must be first instruction in block"); |
duke@435 | 905 | } |
kvn@3311 | 906 | for (uint k = 0; k < n->req(); k++) { |
kvn@1001 | 907 | Node *def = n->in(k); |
kvn@3311 | 908 | if (def && def != n) { |
adlertz@5509 | 909 | assert(get_block_for_node(def) || def->is_Con(), "must have block; constants for debug info ok"); |
kvn@1001 | 910 | // Verify that instructions in the block is in correct order. |
kvn@1001 | 911 | // Uses must follow their definition if they are at the same block. |
kvn@1001 | 912 | // Mostly done to check that MachSpillCopy nodes are placed correctly |
kvn@1001 | 913 | // when CreateEx node is moved in build_ifg_physical(). |
adlertz@5539 | 914 | if (get_block_for_node(def) == block && !(block->head()->is_Loop() && n->is_Phi()) && |
kvn@1001 | 915 | // See (+++) comment in reg_split.cpp |
kvn@3311 | 916 | !(n->jvms() != NULL && n->jvms()->is_monitor_use(k))) { |
kvn@1328 | 917 | bool is_loop = false; |
kvn@1328 | 918 | if (n->is_Phi()) { |
kvn@3311 | 919 | for (uint l = 1; l < def->req(); l++) { |
kvn@1328 | 920 | if (n == def->in(l)) { |
kvn@1328 | 921 | is_loop = true; |
kvn@1328 | 922 | break; // Some kind of loop |
kvn@1328 | 923 | } |
kvn@1328 | 924 | } |
kvn@1328 | 925 | } |
adlertz@5539 | 926 | assert(is_loop || block->find_node(def) < j, "uses must follow definitions"); |
kvn@1036 | 927 | } |
duke@435 | 928 | } |
duke@435 | 929 | } |
duke@435 | 930 | } |
duke@435 | 931 | |
adlertz@5539 | 932 | j = block->end_idx(); |
adlertz@5539 | 933 | Node* bp = (Node*)block->_nodes[block->_nodes.size() - 1]->is_block_proj(); |
adlertz@5539 | 934 | assert(bp, "last instruction must be a block proj"); |
adlertz@5539 | 935 | assert(bp == block->_nodes[j], "wrong number of successors for this block"); |
kvn@3311 | 936 | if (bp->is_Catch()) { |
adlertz@5539 | 937 | while (block->_nodes[--j]->is_MachProj()) { |
adlertz@5539 | 938 | ; |
adlertz@5539 | 939 | } |
adlertz@5539 | 940 | assert(block->_nodes[j]->is_MachCall(), "CatchProj must follow call"); |
kvn@3311 | 941 | } else if (bp->is_Mach() && bp->as_Mach()->ideal_Opcode() == Op_If) { |
adlertz@5539 | 942 | assert(block->_num_succs == 2, "Conditional branch must have two targets"); |
duke@435 | 943 | } |
duke@435 | 944 | } |
kvn@1001 | 945 | #endif |
duke@435 | 946 | } |
duke@435 | 947 | #endif |
duke@435 | 948 | |
duke@435 | 949 | UnionFind::UnionFind( uint max ) : _cnt(max), _max(max), _indices(NEW_RESOURCE_ARRAY(uint,max)) { |
duke@435 | 950 | Copy::zero_to_bytes( _indices, sizeof(uint)*max ); |
duke@435 | 951 | } |
duke@435 | 952 | |
duke@435 | 953 | void UnionFind::extend( uint from_idx, uint to_idx ) { |
duke@435 | 954 | _nesting.check(); |
duke@435 | 955 | if( from_idx >= _max ) { |
duke@435 | 956 | uint size = 16; |
duke@435 | 957 | while( size <= from_idx ) size <<=1; |
duke@435 | 958 | _indices = REALLOC_RESOURCE_ARRAY( uint, _indices, _max, size ); |
duke@435 | 959 | _max = size; |
duke@435 | 960 | } |
duke@435 | 961 | while( _cnt <= from_idx ) _indices[_cnt++] = 0; |
duke@435 | 962 | _indices[from_idx] = to_idx; |
duke@435 | 963 | } |
duke@435 | 964 | |
duke@435 | 965 | void UnionFind::reset( uint max ) { |
duke@435 | 966 | assert( max <= max_uint, "Must fit within uint" ); |
duke@435 | 967 | // Force the Union-Find mapping to be at least this large |
duke@435 | 968 | extend(max,0); |
duke@435 | 969 | // Initialize to be the ID mapping. |
rasbold@853 | 970 | for( uint i=0; i<max; i++ ) map(i,i); |
duke@435 | 971 | } |
duke@435 | 972 | |
duke@435 | 973 | // Straight out of Tarjan's union-find algorithm |
duke@435 | 974 | uint UnionFind::Find_compress( uint idx ) { |
duke@435 | 975 | uint cur = idx; |
duke@435 | 976 | uint next = lookup(cur); |
duke@435 | 977 | while( next != cur ) { // Scan chain of equivalences |
duke@435 | 978 | assert( next < cur, "always union smaller" ); |
duke@435 | 979 | cur = next; // until find a fixed-point |
duke@435 | 980 | next = lookup(cur); |
duke@435 | 981 | } |
duke@435 | 982 | // Core of union-find algorithm: update chain of |
duke@435 | 983 | // equivalences to be equal to the root. |
duke@435 | 984 | while( idx != next ) { |
duke@435 | 985 | uint tmp = lookup(idx); |
duke@435 | 986 | map(idx, next); |
duke@435 | 987 | idx = tmp; |
duke@435 | 988 | } |
duke@435 | 989 | return idx; |
duke@435 | 990 | } |
duke@435 | 991 | |
duke@435 | 992 | // Like Find above, but no path compress, so bad asymptotic behavior |
duke@435 | 993 | uint UnionFind::Find_const( uint idx ) const { |
duke@435 | 994 | if( idx == 0 ) return idx; // Ignore the zero idx |
duke@435 | 995 | // Off the end? This can happen during debugging dumps |
duke@435 | 996 | // when data structures have not finished being updated. |
duke@435 | 997 | if( idx >= _max ) return idx; |
duke@435 | 998 | uint next = lookup(idx); |
duke@435 | 999 | while( next != idx ) { // Scan chain of equivalences |
duke@435 | 1000 | idx = next; // until find a fixed-point |
duke@435 | 1001 | next = lookup(idx); |
duke@435 | 1002 | } |
duke@435 | 1003 | return next; |
duke@435 | 1004 | } |
duke@435 | 1005 | |
duke@435 | 1006 | // union 2 sets together. |
duke@435 | 1007 | void UnionFind::Union( uint idx1, uint idx2 ) { |
duke@435 | 1008 | uint src = Find(idx1); |
duke@435 | 1009 | uint dst = Find(idx2); |
duke@435 | 1010 | assert( src, "" ); |
duke@435 | 1011 | assert( dst, "" ); |
duke@435 | 1012 | assert( src < _max, "oob" ); |
duke@435 | 1013 | assert( dst < _max, "oob" ); |
duke@435 | 1014 | assert( src < dst, "always union smaller" ); |
duke@435 | 1015 | map(dst,src); |
duke@435 | 1016 | } |
rasbold@853 | 1017 | |
rasbold@853 | 1018 | #ifndef PRODUCT |
rasbold@853 | 1019 | void Trace::dump( ) const { |
rasbold@853 | 1020 | tty->print_cr("Trace (freq %f)", first_block()->_freq); |
rasbold@853 | 1021 | for (Block *b = first_block(); b != NULL; b = next(b)) { |
rasbold@853 | 1022 | tty->print(" B%d", b->_pre_order); |
rasbold@853 | 1023 | if (b->head()->is_Loop()) { |
rasbold@853 | 1024 | tty->print(" (L%d)", b->compute_loop_alignment()); |
rasbold@853 | 1025 | } |
rasbold@853 | 1026 | if (b->has_loop_alignment()) { |
rasbold@853 | 1027 | tty->print(" (T%d)", b->code_alignment()); |
rasbold@853 | 1028 | } |
rasbold@853 | 1029 | } |
rasbold@853 | 1030 | tty->cr(); |
rasbold@853 | 1031 | } |
rasbold@853 | 1032 | |
rasbold@853 | 1033 | void CFGEdge::dump( ) const { |
rasbold@853 | 1034 | tty->print(" B%d --> B%d Freq: %f out:%3d%% in:%3d%% State: ", |
rasbold@853 | 1035 | from()->_pre_order, to()->_pre_order, freq(), _from_pct, _to_pct); |
rasbold@853 | 1036 | switch(state()) { |
rasbold@853 | 1037 | case connected: |
rasbold@853 | 1038 | tty->print("connected"); |
rasbold@853 | 1039 | break; |
rasbold@853 | 1040 | case open: |
rasbold@853 | 1041 | tty->print("open"); |
rasbold@853 | 1042 | break; |
rasbold@853 | 1043 | case interior: |
rasbold@853 | 1044 | tty->print("interior"); |
rasbold@853 | 1045 | break; |
rasbold@853 | 1046 | } |
rasbold@853 | 1047 | if (infrequent()) { |
rasbold@853 | 1048 | tty->print(" infrequent"); |
rasbold@853 | 1049 | } |
rasbold@853 | 1050 | tty->cr(); |
rasbold@853 | 1051 | } |
rasbold@853 | 1052 | #endif |
rasbold@853 | 1053 | |
rasbold@853 | 1054 | // Comparison function for edges |
rasbold@853 | 1055 | static int edge_order(CFGEdge **e0, CFGEdge **e1) { |
rasbold@853 | 1056 | float freq0 = (*e0)->freq(); |
rasbold@853 | 1057 | float freq1 = (*e1)->freq(); |
rasbold@853 | 1058 | if (freq0 != freq1) { |
rasbold@853 | 1059 | return freq0 > freq1 ? -1 : 1; |
rasbold@853 | 1060 | } |
rasbold@853 | 1061 | |
rasbold@853 | 1062 | int dist0 = (*e0)->to()->_rpo - (*e0)->from()->_rpo; |
rasbold@853 | 1063 | int dist1 = (*e1)->to()->_rpo - (*e1)->from()->_rpo; |
rasbold@853 | 1064 | |
rasbold@853 | 1065 | return dist1 - dist0; |
rasbold@853 | 1066 | } |
rasbold@853 | 1067 | |
rasbold@853 | 1068 | // Comparison function for edges |
kvn@3128 | 1069 | extern "C" int trace_frequency_order(const void *p0, const void *p1) { |
rasbold@853 | 1070 | Trace *tr0 = *(Trace **) p0; |
rasbold@853 | 1071 | Trace *tr1 = *(Trace **) p1; |
rasbold@853 | 1072 | Block *b0 = tr0->first_block(); |
rasbold@853 | 1073 | Block *b1 = tr1->first_block(); |
rasbold@853 | 1074 | |
rasbold@853 | 1075 | // The trace of connector blocks goes at the end; |
rasbold@853 | 1076 | // we only expect one such trace |
rasbold@853 | 1077 | if (b0->is_connector() != b1->is_connector()) { |
rasbold@853 | 1078 | return b1->is_connector() ? -1 : 1; |
rasbold@853 | 1079 | } |
rasbold@853 | 1080 | |
rasbold@853 | 1081 | // Pull more frequently executed blocks to the beginning |
rasbold@853 | 1082 | float freq0 = b0->_freq; |
rasbold@853 | 1083 | float freq1 = b1->_freq; |
rasbold@853 | 1084 | if (freq0 != freq1) { |
rasbold@853 | 1085 | return freq0 > freq1 ? -1 : 1; |
rasbold@853 | 1086 | } |
rasbold@853 | 1087 | |
rasbold@853 | 1088 | int diff = tr0->first_block()->_rpo - tr1->first_block()->_rpo; |
rasbold@853 | 1089 | |
rasbold@853 | 1090 | return diff; |
rasbold@853 | 1091 | } |
rasbold@853 | 1092 | |
rasbold@853 | 1093 | // Find edges of interest, i.e, those which can fall through. Presumes that |
rasbold@853 | 1094 | // edges which don't fall through are of low frequency and can be generally |
rasbold@853 | 1095 | // ignored. Initialize the list of traces. |
adlertz@5539 | 1096 | void PhaseBlockLayout::find_edges() { |
rasbold@853 | 1097 | // Walk the blocks, creating edges and Traces |
rasbold@853 | 1098 | uint i; |
rasbold@853 | 1099 | Trace *tr = NULL; |
adlertz@5539 | 1100 | for (i = 0; i < _cfg.number_of_blocks(); i++) { |
adlertz@5539 | 1101 | Block* b = _cfg.get_block(i); |
rasbold@853 | 1102 | tr = new Trace(b, next, prev); |
rasbold@853 | 1103 | traces[tr->id()] = tr; |
rasbold@853 | 1104 | |
rasbold@853 | 1105 | // All connector blocks should be at the end of the list |
rasbold@853 | 1106 | if (b->is_connector()) break; |
rasbold@853 | 1107 | |
rasbold@853 | 1108 | // If this block and the next one have a one-to-one successor |
rasbold@853 | 1109 | // predecessor relationship, simply append the next block |
rasbold@853 | 1110 | int nfallthru = b->num_fall_throughs(); |
rasbold@853 | 1111 | while (nfallthru == 1 && |
rasbold@853 | 1112 | b->succ_fall_through(0)) { |
rasbold@853 | 1113 | Block *n = b->_succs[0]; |
rasbold@853 | 1114 | |
rasbold@853 | 1115 | // Skip over single-entry connector blocks, we don't want to |
rasbold@853 | 1116 | // add them to the trace. |
rasbold@853 | 1117 | while (n->is_connector() && n->num_preds() == 1) { |
rasbold@853 | 1118 | n = n->_succs[0]; |
rasbold@853 | 1119 | } |
rasbold@853 | 1120 | |
rasbold@853 | 1121 | // We see a merge point, so stop search for the next block |
rasbold@853 | 1122 | if (n->num_preds() != 1) break; |
rasbold@853 | 1123 | |
rasbold@853 | 1124 | i++; |
adlertz@5539 | 1125 | assert(n = _cfg.get_block(i), "expecting next block"); |
rasbold@853 | 1126 | tr->append(n); |
rasbold@853 | 1127 | uf->map(n->_pre_order, tr->id()); |
rasbold@853 | 1128 | traces[n->_pre_order] = NULL; |
rasbold@853 | 1129 | nfallthru = b->num_fall_throughs(); |
rasbold@853 | 1130 | b = n; |
rasbold@853 | 1131 | } |
rasbold@853 | 1132 | |
rasbold@853 | 1133 | if (nfallthru > 0) { |
rasbold@853 | 1134 | // Create a CFGEdge for each outgoing |
rasbold@853 | 1135 | // edge that could be a fall-through. |
rasbold@853 | 1136 | for (uint j = 0; j < b->_num_succs; j++ ) { |
rasbold@853 | 1137 | if (b->succ_fall_through(j)) { |
rasbold@853 | 1138 | Block *target = b->non_connector_successor(j); |
rasbold@853 | 1139 | float freq = b->_freq * b->succ_prob(j); |
rasbold@853 | 1140 | int from_pct = (int) ((100 * freq) / b->_freq); |
rasbold@853 | 1141 | int to_pct = (int) ((100 * freq) / target->_freq); |
rasbold@853 | 1142 | edges->append(new CFGEdge(b, target, freq, from_pct, to_pct)); |
rasbold@853 | 1143 | } |
rasbold@853 | 1144 | } |
rasbold@853 | 1145 | } |
rasbold@853 | 1146 | } |
rasbold@853 | 1147 | |
rasbold@853 | 1148 | // Group connector blocks into one trace |
adlertz@5539 | 1149 | for (i++; i < _cfg.number_of_blocks(); i++) { |
adlertz@5539 | 1150 | Block *b = _cfg.get_block(i); |
rasbold@853 | 1151 | assert(b->is_connector(), "connector blocks at the end"); |
rasbold@853 | 1152 | tr->append(b); |
rasbold@853 | 1153 | uf->map(b->_pre_order, tr->id()); |
rasbold@853 | 1154 | traces[b->_pre_order] = NULL; |
rasbold@853 | 1155 | } |
rasbold@853 | 1156 | } |
rasbold@853 | 1157 | |
rasbold@853 | 1158 | // Union two traces together in uf, and null out the trace in the list |
adlertz@5539 | 1159 | void PhaseBlockLayout::union_traces(Trace* updated_trace, Trace* old_trace) { |
rasbold@853 | 1160 | uint old_id = old_trace->id(); |
rasbold@853 | 1161 | uint updated_id = updated_trace->id(); |
rasbold@853 | 1162 | |
rasbold@853 | 1163 | uint lo_id = updated_id; |
rasbold@853 | 1164 | uint hi_id = old_id; |
rasbold@853 | 1165 | |
rasbold@853 | 1166 | // If from is greater than to, swap values to meet |
rasbold@853 | 1167 | // UnionFind guarantee. |
rasbold@853 | 1168 | if (updated_id > old_id) { |
rasbold@853 | 1169 | lo_id = old_id; |
rasbold@853 | 1170 | hi_id = updated_id; |
rasbold@853 | 1171 | |
rasbold@853 | 1172 | // Fix up the trace ids |
rasbold@853 | 1173 | traces[lo_id] = traces[updated_id]; |
rasbold@853 | 1174 | updated_trace->set_id(lo_id); |
rasbold@853 | 1175 | } |
rasbold@853 | 1176 | |
rasbold@853 | 1177 | // Union the lower with the higher and remove the pointer |
rasbold@853 | 1178 | // to the higher. |
rasbold@853 | 1179 | uf->Union(lo_id, hi_id); |
rasbold@853 | 1180 | traces[hi_id] = NULL; |
rasbold@853 | 1181 | } |
rasbold@853 | 1182 | |
rasbold@853 | 1183 | // Append traces together via the most frequently executed edges |
adlertz@5539 | 1184 | void PhaseBlockLayout::grow_traces() { |
rasbold@853 | 1185 | // Order the edges, and drive the growth of Traces via the most |
rasbold@853 | 1186 | // frequently executed edges. |
rasbold@853 | 1187 | edges->sort(edge_order); |
rasbold@853 | 1188 | for (int i = 0; i < edges->length(); i++) { |
rasbold@853 | 1189 | CFGEdge *e = edges->at(i); |
rasbold@853 | 1190 | |
rasbold@853 | 1191 | if (e->state() != CFGEdge::open) continue; |
rasbold@853 | 1192 | |
rasbold@853 | 1193 | Block *src_block = e->from(); |
rasbold@853 | 1194 | Block *targ_block = e->to(); |
rasbold@853 | 1195 | |
rasbold@853 | 1196 | // Don't grow traces along backedges? |
rasbold@853 | 1197 | if (!BlockLayoutRotateLoops) { |
rasbold@853 | 1198 | if (targ_block->_rpo <= src_block->_rpo) { |
rasbold@853 | 1199 | targ_block->set_loop_alignment(targ_block); |
rasbold@853 | 1200 | continue; |
rasbold@853 | 1201 | } |
rasbold@853 | 1202 | } |
rasbold@853 | 1203 | |
rasbold@853 | 1204 | Trace *src_trace = trace(src_block); |
rasbold@853 | 1205 | Trace *targ_trace = trace(targ_block); |
rasbold@853 | 1206 | |
rasbold@853 | 1207 | // If the edge in question can join two traces at their ends, |
rasbold@853 | 1208 | // append one trace to the other. |
rasbold@853 | 1209 | if (src_trace->last_block() == src_block) { |
rasbold@853 | 1210 | if (src_trace == targ_trace) { |
rasbold@853 | 1211 | e->set_state(CFGEdge::interior); |
rasbold@853 | 1212 | if (targ_trace->backedge(e)) { |
rasbold@853 | 1213 | // Reset i to catch any newly eligible edge |
rasbold@853 | 1214 | // (Or we could remember the first "open" edge, and reset there) |
rasbold@853 | 1215 | i = 0; |
rasbold@853 | 1216 | } |
rasbold@853 | 1217 | } else if (targ_trace->first_block() == targ_block) { |
rasbold@853 | 1218 | e->set_state(CFGEdge::connected); |
rasbold@853 | 1219 | src_trace->append(targ_trace); |
rasbold@853 | 1220 | union_traces(src_trace, targ_trace); |
rasbold@853 | 1221 | } |
rasbold@853 | 1222 | } |
rasbold@853 | 1223 | } |
rasbold@853 | 1224 | } |
rasbold@853 | 1225 | |
rasbold@853 | 1226 | // Embed one trace into another, if the fork or join points are sufficiently |
rasbold@853 | 1227 | // balanced. |
adlertz@5539 | 1228 | void PhaseBlockLayout::merge_traces(bool fall_thru_only) { |
rasbold@853 | 1229 | // Walk the edge list a another time, looking at unprocessed edges. |
rasbold@853 | 1230 | // Fold in diamonds |
rasbold@853 | 1231 | for (int i = 0; i < edges->length(); i++) { |
rasbold@853 | 1232 | CFGEdge *e = edges->at(i); |
rasbold@853 | 1233 | |
rasbold@853 | 1234 | if (e->state() != CFGEdge::open) continue; |
rasbold@853 | 1235 | if (fall_thru_only) { |
rasbold@853 | 1236 | if (e->infrequent()) continue; |
rasbold@853 | 1237 | } |
rasbold@853 | 1238 | |
rasbold@853 | 1239 | Block *src_block = e->from(); |
rasbold@853 | 1240 | Trace *src_trace = trace(src_block); |
rasbold@853 | 1241 | bool src_at_tail = src_trace->last_block() == src_block; |
rasbold@853 | 1242 | |
rasbold@853 | 1243 | Block *targ_block = e->to(); |
rasbold@853 | 1244 | Trace *targ_trace = trace(targ_block); |
rasbold@853 | 1245 | bool targ_at_start = targ_trace->first_block() == targ_block; |
rasbold@853 | 1246 | |
rasbold@853 | 1247 | if (src_trace == targ_trace) { |
rasbold@853 | 1248 | // This may be a loop, but we can't do much about it. |
rasbold@853 | 1249 | e->set_state(CFGEdge::interior); |
rasbold@853 | 1250 | continue; |
rasbold@853 | 1251 | } |
rasbold@853 | 1252 | |
rasbold@853 | 1253 | if (fall_thru_only) { |
rasbold@853 | 1254 | // If the edge links the middle of two traces, we can't do anything. |
rasbold@853 | 1255 | // Mark the edge and continue. |
rasbold@853 | 1256 | if (!src_at_tail & !targ_at_start) { |
rasbold@853 | 1257 | continue; |
rasbold@853 | 1258 | } |
rasbold@853 | 1259 | |
rasbold@853 | 1260 | // Don't grow traces along backedges? |
rasbold@853 | 1261 | if (!BlockLayoutRotateLoops && (targ_block->_rpo <= src_block->_rpo)) { |
rasbold@853 | 1262 | continue; |
rasbold@853 | 1263 | } |
rasbold@853 | 1264 | |
rasbold@853 | 1265 | // If both ends of the edge are available, why didn't we handle it earlier? |
rasbold@853 | 1266 | assert(src_at_tail ^ targ_at_start, "Should have caught this edge earlier."); |
rasbold@853 | 1267 | |
rasbold@853 | 1268 | if (targ_at_start) { |
rasbold@853 | 1269 | // Insert the "targ" trace in the "src" trace if the insertion point |
rasbold@853 | 1270 | // is a two way branch. |
rasbold@853 | 1271 | // Better profitability check possible, but may not be worth it. |
rasbold@853 | 1272 | // Someday, see if the this "fork" has an associated "join"; |
rasbold@853 | 1273 | // then make a policy on merging this trace at the fork or join. |
rasbold@853 | 1274 | // For example, other things being equal, it may be better to place this |
rasbold@853 | 1275 | // trace at the join point if the "src" trace ends in a two-way, but |
rasbold@853 | 1276 | // the insertion point is one-way. |
rasbold@853 | 1277 | assert(src_block->num_fall_throughs() == 2, "unexpected diamond"); |
rasbold@853 | 1278 | e->set_state(CFGEdge::connected); |
rasbold@853 | 1279 | src_trace->insert_after(src_block, targ_trace); |
rasbold@853 | 1280 | union_traces(src_trace, targ_trace); |
rasbold@853 | 1281 | } else if (src_at_tail) { |
adlertz@5539 | 1282 | if (src_trace != trace(_cfg.get_root_block())) { |
rasbold@853 | 1283 | e->set_state(CFGEdge::connected); |
rasbold@853 | 1284 | targ_trace->insert_before(targ_block, src_trace); |
rasbold@853 | 1285 | union_traces(targ_trace, src_trace); |
rasbold@853 | 1286 | } |
rasbold@853 | 1287 | } |
rasbold@853 | 1288 | } else if (e->state() == CFGEdge::open) { |
rasbold@853 | 1289 | // Append traces, even without a fall-thru connection. |
twisti@1040 | 1290 | // But leave root entry at the beginning of the block list. |
adlertz@5539 | 1291 | if (targ_trace != trace(_cfg.get_root_block())) { |
rasbold@853 | 1292 | e->set_state(CFGEdge::connected); |
rasbold@853 | 1293 | src_trace->append(targ_trace); |
rasbold@853 | 1294 | union_traces(src_trace, targ_trace); |
rasbold@853 | 1295 | } |
rasbold@853 | 1296 | } |
rasbold@853 | 1297 | } |
rasbold@853 | 1298 | } |
rasbold@853 | 1299 | |
rasbold@853 | 1300 | // Order the sequence of the traces in some desirable way, and fixup the |
rasbold@853 | 1301 | // jumps at the end of each block. |
adlertz@5539 | 1302 | void PhaseBlockLayout::reorder_traces(int count) { |
rasbold@853 | 1303 | ResourceArea *area = Thread::current()->resource_area(); |
rasbold@853 | 1304 | Trace ** new_traces = NEW_ARENA_ARRAY(area, Trace *, count); |
rasbold@853 | 1305 | Block_List worklist; |
rasbold@853 | 1306 | int new_count = 0; |
rasbold@853 | 1307 | |
rasbold@853 | 1308 | // Compact the traces. |
rasbold@853 | 1309 | for (int i = 0; i < count; i++) { |
rasbold@853 | 1310 | Trace *tr = traces[i]; |
rasbold@853 | 1311 | if (tr != NULL) { |
rasbold@853 | 1312 | new_traces[new_count++] = tr; |
rasbold@853 | 1313 | } |
rasbold@853 | 1314 | } |
rasbold@853 | 1315 | |
rasbold@853 | 1316 | // The entry block should be first on the new trace list. |
adlertz@5539 | 1317 | Trace *tr = trace(_cfg.get_root_block()); |
rasbold@853 | 1318 | assert(tr == new_traces[0], "entry trace misplaced"); |
rasbold@853 | 1319 | |
rasbold@853 | 1320 | // Sort the new trace list by frequency |
rasbold@853 | 1321 | qsort(new_traces + 1, new_count - 1, sizeof(new_traces[0]), trace_frequency_order); |
rasbold@853 | 1322 | |
rasbold@853 | 1323 | // Patch up the successor blocks |
adlertz@5539 | 1324 | _cfg.clear_blocks(); |
rasbold@853 | 1325 | for (int i = 0; i < new_count; i++) { |
rasbold@853 | 1326 | Trace *tr = new_traces[i]; |
rasbold@853 | 1327 | if (tr != NULL) { |
rasbold@853 | 1328 | tr->fixup_blocks(_cfg); |
rasbold@853 | 1329 | } |
rasbold@853 | 1330 | } |
rasbold@853 | 1331 | } |
rasbold@853 | 1332 | |
rasbold@853 | 1333 | // Order basic blocks based on frequency |
adlertz@5539 | 1334 | PhaseBlockLayout::PhaseBlockLayout(PhaseCFG &cfg) |
adlertz@5539 | 1335 | : Phase(BlockLayout) |
adlertz@5539 | 1336 | , _cfg(cfg) { |
rasbold@853 | 1337 | ResourceMark rm; |
rasbold@853 | 1338 | ResourceArea *area = Thread::current()->resource_area(); |
rasbold@853 | 1339 | |
rasbold@853 | 1340 | // List of traces |
adlertz@5539 | 1341 | int size = _cfg.number_of_blocks() + 1; |
rasbold@853 | 1342 | traces = NEW_ARENA_ARRAY(area, Trace *, size); |
rasbold@853 | 1343 | memset(traces, 0, size*sizeof(Trace*)); |
rasbold@853 | 1344 | next = NEW_ARENA_ARRAY(area, Block *, size); |
rasbold@853 | 1345 | memset(next, 0, size*sizeof(Block *)); |
rasbold@853 | 1346 | prev = NEW_ARENA_ARRAY(area, Block *, size); |
rasbold@853 | 1347 | memset(prev , 0, size*sizeof(Block *)); |
rasbold@853 | 1348 | |
rasbold@853 | 1349 | // List of edges |
rasbold@853 | 1350 | edges = new GrowableArray<CFGEdge*>; |
rasbold@853 | 1351 | |
rasbold@853 | 1352 | // Mapping block index --> block_trace |
rasbold@853 | 1353 | uf = new UnionFind(size); |
rasbold@853 | 1354 | uf->reset(size); |
rasbold@853 | 1355 | |
rasbold@853 | 1356 | // Find edges and create traces. |
rasbold@853 | 1357 | find_edges(); |
rasbold@853 | 1358 | |
rasbold@853 | 1359 | // Grow traces at their ends via most frequent edges. |
rasbold@853 | 1360 | grow_traces(); |
rasbold@853 | 1361 | |
rasbold@853 | 1362 | // Merge one trace into another, but only at fall-through points. |
rasbold@853 | 1363 | // This may make diamonds and other related shapes in a trace. |
rasbold@853 | 1364 | merge_traces(true); |
rasbold@853 | 1365 | |
rasbold@853 | 1366 | // Run merge again, allowing two traces to be catenated, even if |
rasbold@853 | 1367 | // one does not fall through into the other. This appends loosely |
rasbold@853 | 1368 | // related traces to be near each other. |
rasbold@853 | 1369 | merge_traces(false); |
rasbold@853 | 1370 | |
rasbold@853 | 1371 | // Re-order all the remaining traces by frequency |
rasbold@853 | 1372 | reorder_traces(size); |
rasbold@853 | 1373 | |
adlertz@5539 | 1374 | assert(_cfg.number_of_blocks() >= (uint) (size - 1), "number of blocks can not shrink"); |
rasbold@853 | 1375 | } |
rasbold@853 | 1376 | |
rasbold@853 | 1377 | |
rasbold@853 | 1378 | // Edge e completes a loop in a trace. If the target block is head of the |
rasbold@853 | 1379 | // loop, rotate the loop block so that the loop ends in a conditional branch. |
rasbold@853 | 1380 | bool Trace::backedge(CFGEdge *e) { |
rasbold@853 | 1381 | bool loop_rotated = false; |
rasbold@853 | 1382 | Block *src_block = e->from(); |
rasbold@853 | 1383 | Block *targ_block = e->to(); |
rasbold@853 | 1384 | |
rasbold@853 | 1385 | assert(last_block() == src_block, "loop discovery at back branch"); |
rasbold@853 | 1386 | if (first_block() == targ_block) { |
rasbold@853 | 1387 | if (BlockLayoutRotateLoops && last_block()->num_fall_throughs() < 2) { |
rasbold@853 | 1388 | // Find the last block in the trace that has a conditional |
rasbold@853 | 1389 | // branch. |
rasbold@853 | 1390 | Block *b; |
rasbold@853 | 1391 | for (b = last_block(); b != NULL; b = prev(b)) { |
rasbold@853 | 1392 | if (b->num_fall_throughs() == 2) { |
rasbold@853 | 1393 | break; |
rasbold@853 | 1394 | } |
rasbold@853 | 1395 | } |
rasbold@853 | 1396 | |
rasbold@853 | 1397 | if (b != last_block() && b != NULL) { |
rasbold@853 | 1398 | loop_rotated = true; |
rasbold@853 | 1399 | |
rasbold@853 | 1400 | // Rotate the loop by doing two-part linked-list surgery. |
rasbold@853 | 1401 | append(first_block()); |
rasbold@853 | 1402 | break_loop_after(b); |
rasbold@853 | 1403 | } |
rasbold@853 | 1404 | } |
rasbold@853 | 1405 | |
rasbold@853 | 1406 | // Backbranch to the top of a trace |
twisti@1040 | 1407 | // Scroll forward through the trace from the targ_block. If we find |
rasbold@853 | 1408 | // a loop head before another loop top, use the the loop head alignment. |
rasbold@853 | 1409 | for (Block *b = targ_block; b != NULL; b = next(b)) { |
rasbold@853 | 1410 | if (b->has_loop_alignment()) { |
rasbold@853 | 1411 | break; |
rasbold@853 | 1412 | } |
rasbold@853 | 1413 | if (b->head()->is_Loop()) { |
rasbold@853 | 1414 | targ_block = b; |
rasbold@853 | 1415 | break; |
rasbold@853 | 1416 | } |
rasbold@853 | 1417 | } |
rasbold@853 | 1418 | |
rasbold@853 | 1419 | first_block()->set_loop_alignment(targ_block); |
rasbold@853 | 1420 | |
rasbold@853 | 1421 | } else { |
rasbold@853 | 1422 | // Backbranch into the middle of a trace |
rasbold@853 | 1423 | targ_block->set_loop_alignment(targ_block); |
rasbold@853 | 1424 | } |
rasbold@853 | 1425 | |
rasbold@853 | 1426 | return loop_rotated; |
rasbold@853 | 1427 | } |
rasbold@853 | 1428 | |
rasbold@853 | 1429 | // push blocks onto the CFG list |
rasbold@853 | 1430 | // ensure that blocks have the correct two-way branch sense |
rasbold@853 | 1431 | void Trace::fixup_blocks(PhaseCFG &cfg) { |
rasbold@853 | 1432 | Block *last = last_block(); |
rasbold@853 | 1433 | for (Block *b = first_block(); b != NULL; b = next(b)) { |
adlertz@5539 | 1434 | cfg.add_block(b); |
rasbold@853 | 1435 | if (!b->is_connector()) { |
rasbold@853 | 1436 | int nfallthru = b->num_fall_throughs(); |
rasbold@853 | 1437 | if (b != last) { |
rasbold@853 | 1438 | if (nfallthru == 2) { |
rasbold@853 | 1439 | // Ensure that the sense of the branch is correct |
rasbold@853 | 1440 | Block *bnext = next(b); |
rasbold@853 | 1441 | Block *bs0 = b->non_connector_successor(0); |
rasbold@853 | 1442 | |
rasbold@853 | 1443 | MachNode *iff = b->_nodes[b->_nodes.size()-3]->as_Mach(); |
rasbold@853 | 1444 | ProjNode *proj0 = b->_nodes[b->_nodes.size()-2]->as_Proj(); |
rasbold@853 | 1445 | ProjNode *proj1 = b->_nodes[b->_nodes.size()-1]->as_Proj(); |
rasbold@853 | 1446 | |
rasbold@853 | 1447 | if (bnext == bs0) { |
rasbold@853 | 1448 | // Fall-thru case in succs[0], should be in succs[1] |
rasbold@853 | 1449 | |
rasbold@853 | 1450 | // Flip targets in _succs map |
rasbold@853 | 1451 | Block *tbs0 = b->_succs[0]; |
rasbold@853 | 1452 | Block *tbs1 = b->_succs[1]; |
rasbold@853 | 1453 | b->_succs.map( 0, tbs1 ); |
rasbold@853 | 1454 | b->_succs.map( 1, tbs0 ); |
rasbold@853 | 1455 | |
rasbold@853 | 1456 | // Flip projections to match targets |
rasbold@853 | 1457 | b->_nodes.map(b->_nodes.size()-2, proj1); |
rasbold@853 | 1458 | b->_nodes.map(b->_nodes.size()-1, proj0); |
rasbold@853 | 1459 | } |
rasbold@853 | 1460 | } |
rasbold@853 | 1461 | } |
rasbold@853 | 1462 | } |
rasbold@853 | 1463 | } |
rasbold@853 | 1464 | } |