Mon, 26 Sep 2011 10:24:05 -0700
7081933: Use zeroing elimination optimization for large array
Summary: Don't zero new typeArray during runtime call if the allocation is followed by arraycopy into it.
Reviewed-by: twisti
duke@435 | 1 | /* |
trims@2708 | 2 | * Copyright (c) 1998, 2011, 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 "memory/allocation.inline.hpp" |
stefank@2314 | 27 | #include "opto/block.hpp" |
stefank@2314 | 28 | #include "opto/c2compiler.hpp" |
stefank@2314 | 29 | #include "opto/callnode.hpp" |
stefank@2314 | 30 | #include "opto/cfgnode.hpp" |
stefank@2314 | 31 | #include "opto/machnode.hpp" |
stefank@2314 | 32 | #include "opto/runtime.hpp" |
stefank@2314 | 33 | #ifdef TARGET_ARCH_MODEL_x86_32 |
stefank@2314 | 34 | # include "adfiles/ad_x86_32.hpp" |
stefank@2314 | 35 | #endif |
stefank@2314 | 36 | #ifdef TARGET_ARCH_MODEL_x86_64 |
stefank@2314 | 37 | # include "adfiles/ad_x86_64.hpp" |
stefank@2314 | 38 | #endif |
stefank@2314 | 39 | #ifdef TARGET_ARCH_MODEL_sparc |
stefank@2314 | 40 | # include "adfiles/ad_sparc.hpp" |
stefank@2314 | 41 | #endif |
stefank@2314 | 42 | #ifdef TARGET_ARCH_MODEL_zero |
stefank@2314 | 43 | # include "adfiles/ad_zero.hpp" |
stefank@2314 | 44 | #endif |
roland@2683 | 45 | #ifdef TARGET_ARCH_MODEL_arm |
roland@2683 | 46 | # include "adfiles/ad_arm.hpp" |
roland@2683 | 47 | #endif |
jcoomes@2993 | 48 | #ifdef TARGET_ARCH_MODEL_ppc |
jcoomes@2993 | 49 | # include "adfiles/ad_ppc.hpp" |
jcoomes@2993 | 50 | #endif |
stefank@2314 | 51 | |
duke@435 | 52 | // Optimization - Graph Style |
duke@435 | 53 | |
duke@435 | 54 | //------------------------------implicit_null_check---------------------------- |
duke@435 | 55 | // Detect implicit-null-check opportunities. Basically, find NULL checks |
duke@435 | 56 | // with suitable memory ops nearby. Use the memory op to do the NULL check. |
duke@435 | 57 | // I can generate a memory op if there is not one nearby. |
duke@435 | 58 | // The proj is the control projection for the not-null case. |
kvn@1930 | 59 | // The val is the pointer being checked for nullness or |
kvn@1930 | 60 | // decodeHeapOop_not_null node if it did not fold into address. |
duke@435 | 61 | void Block::implicit_null_check(PhaseCFG *cfg, Node *proj, Node *val, int allowed_reasons) { |
duke@435 | 62 | // Assume if null check need for 0 offset then always needed |
duke@435 | 63 | // Intel solaris doesn't support any null checks yet and no |
duke@435 | 64 | // mechanism exists (yet) to set the switches at an os_cpu level |
duke@435 | 65 | if( !ImplicitNullChecks || MacroAssembler::needs_explicit_null_check(0)) return; |
duke@435 | 66 | |
duke@435 | 67 | // Make sure the ptr-is-null path appears to be uncommon! |
duke@435 | 68 | float f = end()->as_MachIf()->_prob; |
duke@435 | 69 | if( proj->Opcode() == Op_IfTrue ) f = 1.0f - f; |
duke@435 | 70 | if( f > PROB_UNLIKELY_MAG(4) ) return; |
duke@435 | 71 | |
duke@435 | 72 | uint bidx = 0; // Capture index of value into memop |
duke@435 | 73 | bool was_store; // Memory op is a store op |
duke@435 | 74 | |
duke@435 | 75 | // Get the successor block for if the test ptr is non-null |
duke@435 | 76 | Block* not_null_block; // this one goes with the proj |
duke@435 | 77 | Block* null_block; |
duke@435 | 78 | if (_nodes[_nodes.size()-1] == proj) { |
duke@435 | 79 | null_block = _succs[0]; |
duke@435 | 80 | not_null_block = _succs[1]; |
duke@435 | 81 | } else { |
duke@435 | 82 | assert(_nodes[_nodes.size()-2] == proj, "proj is one or the other"); |
duke@435 | 83 | not_null_block = _succs[0]; |
duke@435 | 84 | null_block = _succs[1]; |
duke@435 | 85 | } |
kvn@767 | 86 | while (null_block->is_Empty() == Block::empty_with_goto) { |
kvn@767 | 87 | null_block = null_block->_succs[0]; |
kvn@767 | 88 | } |
duke@435 | 89 | |
duke@435 | 90 | // Search the exception block for an uncommon trap. |
duke@435 | 91 | // (See Parse::do_if and Parse::do_ifnull for the reason |
duke@435 | 92 | // we need an uncommon trap. Briefly, we need a way to |
duke@435 | 93 | // detect failure of this optimization, as in 6366351.) |
duke@435 | 94 | { |
duke@435 | 95 | bool found_trap = false; |
duke@435 | 96 | for (uint i1 = 0; i1 < null_block->_nodes.size(); i1++) { |
duke@435 | 97 | Node* nn = null_block->_nodes[i1]; |
duke@435 | 98 | if (nn->is_MachCall() && |
twisti@2103 | 99 | nn->as_MachCall()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point()) { |
duke@435 | 100 | const Type* trtype = nn->in(TypeFunc::Parms)->bottom_type(); |
duke@435 | 101 | if (trtype->isa_int() && trtype->is_int()->is_con()) { |
duke@435 | 102 | jint tr_con = trtype->is_int()->get_con(); |
duke@435 | 103 | Deoptimization::DeoptReason reason = Deoptimization::trap_request_reason(tr_con); |
duke@435 | 104 | Deoptimization::DeoptAction action = Deoptimization::trap_request_action(tr_con); |
duke@435 | 105 | assert((int)reason < (int)BitsPerInt, "recode bit map"); |
duke@435 | 106 | if (is_set_nth_bit(allowed_reasons, (int) reason) |
duke@435 | 107 | && action != Deoptimization::Action_none) { |
duke@435 | 108 | // This uncommon trap is sure to recompile, eventually. |
duke@435 | 109 | // When that happens, C->too_many_traps will prevent |
duke@435 | 110 | // this transformation from happening again. |
duke@435 | 111 | found_trap = true; |
duke@435 | 112 | } |
duke@435 | 113 | } |
duke@435 | 114 | break; |
duke@435 | 115 | } |
duke@435 | 116 | } |
duke@435 | 117 | if (!found_trap) { |
duke@435 | 118 | // We did not find an uncommon trap. |
duke@435 | 119 | return; |
duke@435 | 120 | } |
duke@435 | 121 | } |
duke@435 | 122 | |
kvn@1930 | 123 | // Check for decodeHeapOop_not_null node which did not fold into address |
kvn@1930 | 124 | bool is_decoden = ((intptr_t)val) & 1; |
kvn@1930 | 125 | val = (Node*)(((intptr_t)val) & ~1); |
kvn@1930 | 126 | |
kvn@1930 | 127 | assert(!is_decoden || (val->in(0) == NULL) && val->is_Mach() && |
kvn@1930 | 128 | (val->as_Mach()->ideal_Opcode() == Op_DecodeN), "sanity"); |
kvn@1930 | 129 | |
duke@435 | 130 | // Search the successor block for a load or store who's base value is also |
duke@435 | 131 | // the tested value. There may be several. |
duke@435 | 132 | Node_List *out = new Node_List(Thread::current()->resource_area()); |
duke@435 | 133 | MachNode *best = NULL; // Best found so far |
duke@435 | 134 | for (DUIterator i = val->outs(); val->has_out(i); i++) { |
duke@435 | 135 | Node *m = val->out(i); |
duke@435 | 136 | if( !m->is_Mach() ) continue; |
duke@435 | 137 | MachNode *mach = m->as_Mach(); |
duke@435 | 138 | was_store = false; |
kvn@2048 | 139 | int iop = mach->ideal_Opcode(); |
kvn@2048 | 140 | switch( iop ) { |
duke@435 | 141 | case Op_LoadB: |
twisti@993 | 142 | case Op_LoadUS: |
duke@435 | 143 | case Op_LoadD: |
duke@435 | 144 | case Op_LoadF: |
duke@435 | 145 | case Op_LoadI: |
duke@435 | 146 | case Op_LoadL: |
duke@435 | 147 | case Op_LoadP: |
coleenp@548 | 148 | case Op_LoadN: |
duke@435 | 149 | case Op_LoadS: |
duke@435 | 150 | case Op_LoadKlass: |
kvn@599 | 151 | case Op_LoadNKlass: |
duke@435 | 152 | case Op_LoadRange: |
duke@435 | 153 | case Op_LoadD_unaligned: |
duke@435 | 154 | case Op_LoadL_unaligned: |
kvn@1586 | 155 | assert(mach->in(2) == val, "should be address"); |
duke@435 | 156 | break; |
duke@435 | 157 | case Op_StoreB: |
duke@435 | 158 | case Op_StoreC: |
duke@435 | 159 | case Op_StoreCM: |
duke@435 | 160 | case Op_StoreD: |
duke@435 | 161 | case Op_StoreF: |
duke@435 | 162 | case Op_StoreI: |
duke@435 | 163 | case Op_StoreL: |
duke@435 | 164 | case Op_StoreP: |
coleenp@548 | 165 | case Op_StoreN: |
duke@435 | 166 | was_store = true; // Memory op is a store op |
duke@435 | 167 | // Stores will have their address in slot 2 (memory in slot 1). |
duke@435 | 168 | // If the value being nul-checked is in another slot, it means we |
duke@435 | 169 | // are storing the checked value, which does NOT check the value! |
duke@435 | 170 | if( mach->in(2) != val ) continue; |
duke@435 | 171 | break; // Found a memory op? |
duke@435 | 172 | case Op_StrComp: |
cfang@1116 | 173 | case Op_StrEquals: |
cfang@1116 | 174 | case Op_StrIndexOf: |
rasbold@604 | 175 | case Op_AryEq: |
duke@435 | 176 | // Not a legit memory op for implicit null check regardless of |
duke@435 | 177 | // embedded loads |
duke@435 | 178 | continue; |
duke@435 | 179 | default: // Also check for embedded loads |
duke@435 | 180 | if( !mach->needs_anti_dependence_check() ) |
duke@435 | 181 | continue; // Not an memory op; skip it |
kvn@2048 | 182 | if( must_clone[iop] ) { |
kvn@2048 | 183 | // Do not move nodes which produce flags because |
kvn@2048 | 184 | // RA will try to clone it to place near branch and |
kvn@2048 | 185 | // it will cause recompilation, see clone_node(). |
kvn@2048 | 186 | continue; |
kvn@2048 | 187 | } |
kvn@1586 | 188 | { |
kvn@1930 | 189 | // Check that value is used in memory address in |
kvn@1930 | 190 | // instructions with embedded load (CmpP val1,(val2+off)). |
kvn@1586 | 191 | Node* base; |
kvn@1586 | 192 | Node* index; |
kvn@1586 | 193 | const MachOper* oper = mach->memory_inputs(base, index); |
kvn@1586 | 194 | if (oper == NULL || oper == (MachOper*)-1) { |
kvn@1586 | 195 | continue; // Not an memory op; skip it |
kvn@1586 | 196 | } |
kvn@1586 | 197 | if (val == base || |
kvn@1586 | 198 | val == index && val->bottom_type()->isa_narrowoop()) { |
kvn@1586 | 199 | break; // Found it |
kvn@1586 | 200 | } else { |
kvn@1586 | 201 | continue; // Skip it |
kvn@1586 | 202 | } |
kvn@1586 | 203 | } |
duke@435 | 204 | break; |
duke@435 | 205 | } |
duke@435 | 206 | // check if the offset is not too high for implicit exception |
duke@435 | 207 | { |
duke@435 | 208 | intptr_t offset = 0; |
duke@435 | 209 | const TypePtr *adr_type = NULL; // Do not need this return value here |
duke@435 | 210 | const Node* base = mach->get_base_and_disp(offset, adr_type); |
duke@435 | 211 | if (base == NULL || base == NodeSentinel) { |
kvn@767 | 212 | // Narrow oop address doesn't have base, only index |
kvn@767 | 213 | if( val->bottom_type()->isa_narrowoop() && |
kvn@767 | 214 | MacroAssembler::needs_explicit_null_check(offset) ) |
kvn@767 | 215 | continue; // Give up if offset is beyond page size |
duke@435 | 216 | // cannot reason about it; is probably not implicit null exception |
duke@435 | 217 | } else { |
kvn@1077 | 218 | const TypePtr* tptr; |
kvn@1077 | 219 | if (UseCompressedOops && Universe::narrow_oop_shift() == 0) { |
kvn@1077 | 220 | // 32-bits narrow oop can be the base of address expressions |
kvn@1077 | 221 | tptr = base->bottom_type()->make_ptr(); |
kvn@1077 | 222 | } else { |
kvn@1077 | 223 | // only regular oops are expected here |
kvn@1077 | 224 | tptr = base->bottom_type()->is_ptr(); |
kvn@1077 | 225 | } |
duke@435 | 226 | // Give up if offset is not a compile-time constant |
duke@435 | 227 | if( offset == Type::OffsetBot || tptr->_offset == Type::OffsetBot ) |
duke@435 | 228 | continue; |
duke@435 | 229 | offset += tptr->_offset; // correct if base is offseted |
duke@435 | 230 | if( MacroAssembler::needs_explicit_null_check(offset) ) |
duke@435 | 231 | continue; // Give up is reference is beyond 4K page size |
duke@435 | 232 | } |
duke@435 | 233 | } |
duke@435 | 234 | |
duke@435 | 235 | // Check ctrl input to see if the null-check dominates the memory op |
duke@435 | 236 | Block *cb = cfg->_bbs[mach->_idx]; |
duke@435 | 237 | cb = cb->_idom; // Always hoist at least 1 block |
duke@435 | 238 | if( !was_store ) { // Stores can be hoisted only one block |
duke@435 | 239 | while( cb->_dom_depth > (_dom_depth + 1)) |
duke@435 | 240 | cb = cb->_idom; // Hoist loads as far as we want |
duke@435 | 241 | // The non-null-block should dominate the memory op, too. Live |
duke@435 | 242 | // range spilling will insert a spill in the non-null-block if it is |
duke@435 | 243 | // needs to spill the memory op for an implicit null check. |
duke@435 | 244 | if (cb->_dom_depth == (_dom_depth + 1)) { |
duke@435 | 245 | if (cb != not_null_block) continue; |
duke@435 | 246 | cb = cb->_idom; |
duke@435 | 247 | } |
duke@435 | 248 | } |
duke@435 | 249 | if( cb != this ) continue; |
duke@435 | 250 | |
duke@435 | 251 | // Found a memory user; see if it can be hoisted to check-block |
duke@435 | 252 | uint vidx = 0; // Capture index of value into memop |
duke@435 | 253 | uint j; |
duke@435 | 254 | for( j = mach->req()-1; j > 0; j-- ) { |
kvn@1930 | 255 | if( mach->in(j) == val ) { |
kvn@1930 | 256 | vidx = j; |
kvn@1930 | 257 | // Ignore DecodeN val which could be hoisted to where needed. |
kvn@1930 | 258 | if( is_decoden ) continue; |
kvn@1930 | 259 | } |
duke@435 | 260 | // Block of memory-op input |
duke@435 | 261 | Block *inb = cfg->_bbs[mach->in(j)->_idx]; |
duke@435 | 262 | Block *b = this; // Start from nul check |
duke@435 | 263 | while( b != inb && b->_dom_depth > inb->_dom_depth ) |
duke@435 | 264 | b = b->_idom; // search upwards for input |
duke@435 | 265 | // See if input dominates null check |
duke@435 | 266 | if( b != inb ) |
duke@435 | 267 | break; |
duke@435 | 268 | } |
duke@435 | 269 | if( j > 0 ) |
duke@435 | 270 | continue; |
duke@435 | 271 | Block *mb = cfg->_bbs[mach->_idx]; |
duke@435 | 272 | // Hoisting stores requires more checks for the anti-dependence case. |
duke@435 | 273 | // Give up hoisting if we have to move the store past any load. |
duke@435 | 274 | if( was_store ) { |
duke@435 | 275 | Block *b = mb; // Start searching here for a local load |
duke@435 | 276 | // mach use (faulting) trying to hoist |
duke@435 | 277 | // n might be blocker to hoisting |
duke@435 | 278 | while( b != this ) { |
duke@435 | 279 | uint k; |
duke@435 | 280 | for( k = 1; k < b->_nodes.size(); k++ ) { |
duke@435 | 281 | Node *n = b->_nodes[k]; |
duke@435 | 282 | if( n->needs_anti_dependence_check() && |
duke@435 | 283 | n->in(LoadNode::Memory) == mach->in(StoreNode::Memory) ) |
duke@435 | 284 | break; // Found anti-dependent load |
duke@435 | 285 | } |
duke@435 | 286 | if( k < b->_nodes.size() ) |
duke@435 | 287 | break; // Found anti-dependent load |
duke@435 | 288 | // Make sure control does not do a merge (would have to check allpaths) |
duke@435 | 289 | if( b->num_preds() != 2 ) break; |
duke@435 | 290 | b = cfg->_bbs[b->pred(1)->_idx]; // Move up to predecessor block |
duke@435 | 291 | } |
duke@435 | 292 | if( b != this ) continue; |
duke@435 | 293 | } |
duke@435 | 294 | |
duke@435 | 295 | // Make sure this memory op is not already being used for a NullCheck |
duke@435 | 296 | Node *e = mb->end(); |
duke@435 | 297 | if( e->is_MachNullCheck() && e->in(1) == mach ) |
duke@435 | 298 | continue; // Already being used as a NULL check |
duke@435 | 299 | |
duke@435 | 300 | // Found a candidate! Pick one with least dom depth - the highest |
duke@435 | 301 | // in the dom tree should be closest to the null check. |
duke@435 | 302 | if( !best || |
duke@435 | 303 | cfg->_bbs[mach->_idx]->_dom_depth < cfg->_bbs[best->_idx]->_dom_depth ) { |
duke@435 | 304 | best = mach; |
duke@435 | 305 | bidx = vidx; |
duke@435 | 306 | |
duke@435 | 307 | } |
duke@435 | 308 | } |
duke@435 | 309 | // No candidate! |
duke@435 | 310 | if( !best ) return; |
duke@435 | 311 | |
duke@435 | 312 | // ---- Found an implicit null check |
duke@435 | 313 | extern int implicit_null_checks; |
duke@435 | 314 | implicit_null_checks++; |
duke@435 | 315 | |
kvn@1930 | 316 | if( is_decoden ) { |
kvn@1930 | 317 | // Check if we need to hoist decodeHeapOop_not_null first. |
kvn@1930 | 318 | Block *valb = cfg->_bbs[val->_idx]; |
kvn@1930 | 319 | if( this != valb && this->_dom_depth < valb->_dom_depth ) { |
kvn@1930 | 320 | // Hoist it up to the end of the test block. |
kvn@1930 | 321 | valb->find_remove(val); |
kvn@1930 | 322 | this->add_inst(val); |
kvn@1930 | 323 | cfg->_bbs.map(val->_idx,this); |
kvn@1930 | 324 | // DecodeN on x86 may kill flags. Check for flag-killing projections |
kvn@1930 | 325 | // that also need to be hoisted. |
kvn@1930 | 326 | for (DUIterator_Fast jmax, j = val->fast_outs(jmax); j < jmax; j++) { |
kvn@1930 | 327 | Node* n = val->fast_out(j); |
kvn@3040 | 328 | if( n->is_MachProj() ) { |
kvn@1930 | 329 | cfg->_bbs[n->_idx]->find_remove(n); |
kvn@1930 | 330 | this->add_inst(n); |
kvn@1930 | 331 | cfg->_bbs.map(n->_idx,this); |
kvn@1930 | 332 | } |
kvn@1930 | 333 | } |
kvn@1930 | 334 | } |
kvn@1930 | 335 | } |
duke@435 | 336 | // Hoist the memory candidate up to the end of the test block. |
duke@435 | 337 | Block *old_block = cfg->_bbs[best->_idx]; |
duke@435 | 338 | old_block->find_remove(best); |
duke@435 | 339 | add_inst(best); |
duke@435 | 340 | cfg->_bbs.map(best->_idx,this); |
duke@435 | 341 | |
duke@435 | 342 | // Move the control dependence |
duke@435 | 343 | if (best->in(0) && best->in(0) == old_block->_nodes[0]) |
duke@435 | 344 | best->set_req(0, _nodes[0]); |
duke@435 | 345 | |
duke@435 | 346 | // Check for flag-killing projections that also need to be hoisted |
duke@435 | 347 | // Should be DU safe because no edge updates. |
duke@435 | 348 | for (DUIterator_Fast jmax, j = best->fast_outs(jmax); j < jmax; j++) { |
duke@435 | 349 | Node* n = best->fast_out(j); |
kvn@3040 | 350 | if( n->is_MachProj() ) { |
duke@435 | 351 | cfg->_bbs[n->_idx]->find_remove(n); |
duke@435 | 352 | add_inst(n); |
duke@435 | 353 | cfg->_bbs.map(n->_idx,this); |
duke@435 | 354 | } |
duke@435 | 355 | } |
duke@435 | 356 | |
duke@435 | 357 | Compile *C = cfg->C; |
duke@435 | 358 | // proj==Op_True --> ne test; proj==Op_False --> eq test. |
duke@435 | 359 | // One of two graph shapes got matched: |
duke@435 | 360 | // (IfTrue (If (Bool NE (CmpP ptr NULL)))) |
duke@435 | 361 | // (IfFalse (If (Bool EQ (CmpP ptr NULL)))) |
duke@435 | 362 | // NULL checks are always branch-if-eq. If we see a IfTrue projection |
duke@435 | 363 | // then we are replacing a 'ne' test with a 'eq' NULL check test. |
duke@435 | 364 | // We need to flip the projections to keep the same semantics. |
duke@435 | 365 | if( proj->Opcode() == Op_IfTrue ) { |
duke@435 | 366 | // Swap order of projections in basic block to swap branch targets |
duke@435 | 367 | Node *tmp1 = _nodes[end_idx()+1]; |
duke@435 | 368 | Node *tmp2 = _nodes[end_idx()+2]; |
duke@435 | 369 | _nodes.map(end_idx()+1, tmp2); |
duke@435 | 370 | _nodes.map(end_idx()+2, tmp1); |
duke@435 | 371 | Node *tmp = new (C, 1) Node(C->top()); // Use not NULL input |
duke@435 | 372 | tmp1->replace_by(tmp); |
duke@435 | 373 | tmp2->replace_by(tmp1); |
duke@435 | 374 | tmp->replace_by(tmp2); |
duke@435 | 375 | tmp->destruct(); |
duke@435 | 376 | } |
duke@435 | 377 | |
duke@435 | 378 | // Remove the existing null check; use a new implicit null check instead. |
duke@435 | 379 | // Since schedule-local needs precise def-use info, we need to correct |
duke@435 | 380 | // it as well. |
duke@435 | 381 | Node *old_tst = proj->in(0); |
duke@435 | 382 | MachNode *nul_chk = new (C) MachNullCheckNode(old_tst->in(0),best,bidx); |
duke@435 | 383 | _nodes.map(end_idx(),nul_chk); |
duke@435 | 384 | cfg->_bbs.map(nul_chk->_idx,this); |
duke@435 | 385 | // Redirect users of old_test to nul_chk |
duke@435 | 386 | for (DUIterator_Last i2min, i2 = old_tst->last_outs(i2min); i2 >= i2min; --i2) |
duke@435 | 387 | old_tst->last_out(i2)->set_req(0, nul_chk); |
duke@435 | 388 | // Clean-up any dead code |
duke@435 | 389 | for (uint i3 = 0; i3 < old_tst->req(); i3++) |
duke@435 | 390 | old_tst->set_req(i3, NULL); |
duke@435 | 391 | |
duke@435 | 392 | cfg->latency_from_uses(nul_chk); |
duke@435 | 393 | cfg->latency_from_uses(best); |
duke@435 | 394 | } |
duke@435 | 395 | |
duke@435 | 396 | |
duke@435 | 397 | //------------------------------select----------------------------------------- |
duke@435 | 398 | // Select a nice fellow from the worklist to schedule next. If there is only |
duke@435 | 399 | // one choice, then use it. Projections take top priority for correctness |
duke@435 | 400 | // reasons - if I see a projection, then it is next. There are a number of |
duke@435 | 401 | // other special cases, for instructions that consume condition codes, et al. |
duke@435 | 402 | // These are chosen immediately. Some instructions are required to immediately |
duke@435 | 403 | // precede the last instruction in the block, and these are taken last. Of the |
duke@435 | 404 | // remaining cases (most), choose the instruction with the greatest latency |
duke@435 | 405 | // (that is, the most number of pseudo-cycles required to the end of the |
duke@435 | 406 | // routine). If there is a tie, choose the instruction with the most inputs. |
duke@435 | 407 | Node *Block::select(PhaseCFG *cfg, Node_List &worklist, int *ready_cnt, VectorSet &next_call, uint sched_slot) { |
duke@435 | 408 | |
duke@435 | 409 | // If only a single entry on the stack, use it |
duke@435 | 410 | uint cnt = worklist.size(); |
duke@435 | 411 | if (cnt == 1) { |
duke@435 | 412 | Node *n = worklist[0]; |
duke@435 | 413 | worklist.map(0,worklist.pop()); |
duke@435 | 414 | return n; |
duke@435 | 415 | } |
duke@435 | 416 | |
duke@435 | 417 | uint choice = 0; // Bigger is most important |
duke@435 | 418 | uint latency = 0; // Bigger is scheduled first |
duke@435 | 419 | uint score = 0; // Bigger is better |
kvn@688 | 420 | int idx = -1; // Index in worklist |
duke@435 | 421 | |
duke@435 | 422 | for( uint i=0; i<cnt; i++ ) { // Inspect entire worklist |
duke@435 | 423 | // Order in worklist is used to break ties. |
duke@435 | 424 | // See caller for how this is used to delay scheduling |
duke@435 | 425 | // of induction variable increments to after the other |
duke@435 | 426 | // uses of the phi are scheduled. |
duke@435 | 427 | Node *n = worklist[i]; // Get Node on worklist |
duke@435 | 428 | |
duke@435 | 429 | int iop = n->is_Mach() ? n->as_Mach()->ideal_Opcode() : 0; |
duke@435 | 430 | if( n->is_Proj() || // Projections always win |
duke@435 | 431 | n->Opcode()== Op_Con || // So does constant 'Top' |
duke@435 | 432 | iop == Op_CreateEx || // Create-exception must start block |
duke@435 | 433 | iop == Op_CheckCastPP |
duke@435 | 434 | ) { |
duke@435 | 435 | worklist.map(i,worklist.pop()); |
duke@435 | 436 | return n; |
duke@435 | 437 | } |
duke@435 | 438 | |
duke@435 | 439 | // Final call in a block must be adjacent to 'catch' |
duke@435 | 440 | Node *e = end(); |
duke@435 | 441 | if( e->is_Catch() && e->in(0)->in(0) == n ) |
duke@435 | 442 | continue; |
duke@435 | 443 | |
duke@435 | 444 | // Memory op for an implicit null check has to be at the end of the block |
duke@435 | 445 | if( e->is_MachNullCheck() && e->in(1) == n ) |
duke@435 | 446 | continue; |
duke@435 | 447 | |
duke@435 | 448 | uint n_choice = 2; |
duke@435 | 449 | |
duke@435 | 450 | // See if this instruction is consumed by a branch. If so, then (as the |
duke@435 | 451 | // branch is the last instruction in the basic block) force it to the |
duke@435 | 452 | // end of the basic block |
duke@435 | 453 | if ( must_clone[iop] ) { |
duke@435 | 454 | // See if any use is a branch |
duke@435 | 455 | bool found_machif = false; |
duke@435 | 456 | |
duke@435 | 457 | for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { |
duke@435 | 458 | Node* use = n->fast_out(j); |
duke@435 | 459 | |
duke@435 | 460 | // The use is a conditional branch, make them adjacent |
duke@435 | 461 | if (use->is_MachIf() && cfg->_bbs[use->_idx]==this ) { |
duke@435 | 462 | found_machif = true; |
duke@435 | 463 | break; |
duke@435 | 464 | } |
duke@435 | 465 | |
duke@435 | 466 | // More than this instruction pending for successor to be ready, |
duke@435 | 467 | // don't choose this if other opportunities are ready |
duke@435 | 468 | if (ready_cnt[use->_idx] > 1) |
duke@435 | 469 | n_choice = 1; |
duke@435 | 470 | } |
duke@435 | 471 | |
duke@435 | 472 | // loop terminated, prefer not to use this instruction |
duke@435 | 473 | if (found_machif) |
duke@435 | 474 | continue; |
duke@435 | 475 | } |
duke@435 | 476 | |
duke@435 | 477 | // See if this has a predecessor that is "must_clone", i.e. sets the |
duke@435 | 478 | // condition code. If so, choose this first |
duke@435 | 479 | for (uint j = 0; j < n->req() ; j++) { |
duke@435 | 480 | Node *inn = n->in(j); |
duke@435 | 481 | if (inn) { |
duke@435 | 482 | if (inn->is_Mach() && must_clone[inn->as_Mach()->ideal_Opcode()] ) { |
duke@435 | 483 | n_choice = 3; |
duke@435 | 484 | break; |
duke@435 | 485 | } |
duke@435 | 486 | } |
duke@435 | 487 | } |
duke@435 | 488 | |
duke@435 | 489 | // MachTemps should be scheduled last so they are near their uses |
duke@435 | 490 | if (n->is_MachTemp()) { |
duke@435 | 491 | n_choice = 1; |
duke@435 | 492 | } |
duke@435 | 493 | |
kvn@2040 | 494 | uint n_latency = cfg->_node_latency->at_grow(n->_idx); |
duke@435 | 495 | uint n_score = n->req(); // Many inputs get high score to break ties |
duke@435 | 496 | |
duke@435 | 497 | // Keep best latency found |
duke@435 | 498 | if( choice < n_choice || |
duke@435 | 499 | ( choice == n_choice && |
duke@435 | 500 | ( latency < n_latency || |
duke@435 | 501 | ( latency == n_latency && |
duke@435 | 502 | ( score < n_score ))))) { |
duke@435 | 503 | choice = n_choice; |
duke@435 | 504 | latency = n_latency; |
duke@435 | 505 | score = n_score; |
duke@435 | 506 | idx = i; // Also keep index in worklist |
duke@435 | 507 | } |
duke@435 | 508 | } // End of for all ready nodes in worklist |
duke@435 | 509 | |
kvn@688 | 510 | assert(idx >= 0, "index should be set"); |
kvn@688 | 511 | Node *n = worklist[(uint)idx]; // Get the winner |
duke@435 | 512 | |
kvn@688 | 513 | worklist.map((uint)idx, worklist.pop()); // Compress worklist |
duke@435 | 514 | return n; |
duke@435 | 515 | } |
duke@435 | 516 | |
duke@435 | 517 | |
duke@435 | 518 | //------------------------------set_next_call---------------------------------- |
duke@435 | 519 | void Block::set_next_call( Node *n, VectorSet &next_call, Block_Array &bbs ) { |
duke@435 | 520 | if( next_call.test_set(n->_idx) ) return; |
duke@435 | 521 | for( uint i=0; i<n->len(); i++ ) { |
duke@435 | 522 | Node *m = n->in(i); |
duke@435 | 523 | if( !m ) continue; // must see all nodes in block that precede call |
duke@435 | 524 | if( bbs[m->_idx] == this ) |
duke@435 | 525 | set_next_call( m, next_call, bbs ); |
duke@435 | 526 | } |
duke@435 | 527 | } |
duke@435 | 528 | |
duke@435 | 529 | //------------------------------needed_for_next_call--------------------------- |
duke@435 | 530 | // Set the flag 'next_call' for each Node that is needed for the next call to |
duke@435 | 531 | // be scheduled. This flag lets me bias scheduling so Nodes needed for the |
duke@435 | 532 | // next subroutine call get priority - basically it moves things NOT needed |
duke@435 | 533 | // for the next call till after the call. This prevents me from trying to |
duke@435 | 534 | // carry lots of stuff live across a call. |
duke@435 | 535 | void Block::needed_for_next_call(Node *this_call, VectorSet &next_call, Block_Array &bbs) { |
duke@435 | 536 | // Find the next control-defining Node in this block |
duke@435 | 537 | Node* call = NULL; |
duke@435 | 538 | for (DUIterator_Fast imax, i = this_call->fast_outs(imax); i < imax; i++) { |
duke@435 | 539 | Node* m = this_call->fast_out(i); |
duke@435 | 540 | if( bbs[m->_idx] == this && // Local-block user |
duke@435 | 541 | m != this_call && // Not self-start node |
kvn@3040 | 542 | m->is_MachCall() ) |
duke@435 | 543 | call = m; |
duke@435 | 544 | break; |
duke@435 | 545 | } |
duke@435 | 546 | if (call == NULL) return; // No next call (e.g., block end is near) |
duke@435 | 547 | // Set next-call for all inputs to this call |
duke@435 | 548 | set_next_call(call, next_call, bbs); |
duke@435 | 549 | } |
duke@435 | 550 | |
duke@435 | 551 | //------------------------------sched_call------------------------------------- |
duke@435 | 552 | uint Block::sched_call( Matcher &matcher, Block_Array &bbs, uint node_cnt, Node_List &worklist, int *ready_cnt, MachCallNode *mcall, VectorSet &next_call ) { |
duke@435 | 553 | RegMask regs; |
duke@435 | 554 | |
duke@435 | 555 | // Schedule all the users of the call right now. All the users are |
duke@435 | 556 | // projection Nodes, so they must be scheduled next to the call. |
duke@435 | 557 | // Collect all the defined registers. |
duke@435 | 558 | for (DUIterator_Fast imax, i = mcall->fast_outs(imax); i < imax; i++) { |
duke@435 | 559 | Node* n = mcall->fast_out(i); |
kvn@3040 | 560 | assert( n->is_MachProj(), "" ); |
duke@435 | 561 | --ready_cnt[n->_idx]; |
duke@435 | 562 | assert( !ready_cnt[n->_idx], "" ); |
duke@435 | 563 | // Schedule next to call |
duke@435 | 564 | _nodes.map(node_cnt++, n); |
duke@435 | 565 | // Collect defined registers |
duke@435 | 566 | regs.OR(n->out_RegMask()); |
duke@435 | 567 | // Check for scheduling the next control-definer |
duke@435 | 568 | if( n->bottom_type() == Type::CONTROL ) |
duke@435 | 569 | // Warm up next pile of heuristic bits |
duke@435 | 570 | needed_for_next_call(n, next_call, bbs); |
duke@435 | 571 | |
duke@435 | 572 | // Children of projections are now all ready |
duke@435 | 573 | for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { |
duke@435 | 574 | Node* m = n->fast_out(j); // Get user |
duke@435 | 575 | if( bbs[m->_idx] != this ) continue; |
duke@435 | 576 | if( m->is_Phi() ) continue; |
duke@435 | 577 | if( !--ready_cnt[m->_idx] ) |
duke@435 | 578 | worklist.push(m); |
duke@435 | 579 | } |
duke@435 | 580 | |
duke@435 | 581 | } |
duke@435 | 582 | |
duke@435 | 583 | // Act as if the call defines the Frame Pointer. |
duke@435 | 584 | // Certainly the FP is alive and well after the call. |
duke@435 | 585 | regs.Insert(matcher.c_frame_pointer()); |
duke@435 | 586 | |
duke@435 | 587 | // Set all registers killed and not already defined by the call. |
duke@435 | 588 | uint r_cnt = mcall->tf()->range()->cnt(); |
duke@435 | 589 | int op = mcall->ideal_Opcode(); |
duke@435 | 590 | MachProjNode *proj = new (matcher.C, 1) MachProjNode( mcall, r_cnt+1, RegMask::Empty, MachProjNode::fat_proj ); |
duke@435 | 591 | bbs.map(proj->_idx,this); |
duke@435 | 592 | _nodes.insert(node_cnt++, proj); |
duke@435 | 593 | |
duke@435 | 594 | // Select the right register save policy. |
duke@435 | 595 | const char * save_policy; |
duke@435 | 596 | switch (op) { |
duke@435 | 597 | case Op_CallRuntime: |
duke@435 | 598 | case Op_CallLeaf: |
duke@435 | 599 | case Op_CallLeafNoFP: |
duke@435 | 600 | // Calling C code so use C calling convention |
duke@435 | 601 | save_policy = matcher._c_reg_save_policy; |
duke@435 | 602 | break; |
duke@435 | 603 | |
duke@435 | 604 | case Op_CallStaticJava: |
duke@435 | 605 | case Op_CallDynamicJava: |
duke@435 | 606 | // Calling Java code so use Java calling convention |
duke@435 | 607 | save_policy = matcher._register_save_policy; |
duke@435 | 608 | break; |
duke@435 | 609 | |
duke@435 | 610 | default: |
duke@435 | 611 | ShouldNotReachHere(); |
duke@435 | 612 | } |
duke@435 | 613 | |
duke@435 | 614 | // When using CallRuntime mark SOE registers as killed by the call |
duke@435 | 615 | // so values that could show up in the RegisterMap aren't live in a |
duke@435 | 616 | // callee saved register since the register wouldn't know where to |
duke@435 | 617 | // find them. CallLeaf and CallLeafNoFP are ok because they can't |
duke@435 | 618 | // have debug info on them. Strictly speaking this only needs to be |
duke@435 | 619 | // done for oops since idealreg2debugmask takes care of debug info |
duke@435 | 620 | // references but there no way to handle oops differently than other |
duke@435 | 621 | // pointers as far as the kill mask goes. |
duke@435 | 622 | bool exclude_soe = op == Op_CallRuntime; |
duke@435 | 623 | |
twisti@1572 | 624 | // If the call is a MethodHandle invoke, we need to exclude the |
twisti@1572 | 625 | // register which is used to save the SP value over MH invokes from |
twisti@1572 | 626 | // the mask. Otherwise this register could be used for |
twisti@1572 | 627 | // deoptimization information. |
twisti@1572 | 628 | if (op == Op_CallStaticJava) { |
twisti@1572 | 629 | MachCallStaticJavaNode* mcallstaticjava = (MachCallStaticJavaNode*) mcall; |
twisti@1572 | 630 | if (mcallstaticjava->_method_handle_invoke) |
twisti@1572 | 631 | proj->_rout.OR(Matcher::method_handle_invoke_SP_save_mask()); |
twisti@1572 | 632 | } |
twisti@1572 | 633 | |
duke@435 | 634 | // Fill in the kill mask for the call |
duke@435 | 635 | for( OptoReg::Name r = OptoReg::Name(0); r < _last_Mach_Reg; r=OptoReg::add(r,1) ) { |
duke@435 | 636 | if( !regs.Member(r) ) { // Not already defined by the call |
duke@435 | 637 | // Save-on-call register? |
duke@435 | 638 | if ((save_policy[r] == 'C') || |
duke@435 | 639 | (save_policy[r] == 'A') || |
duke@435 | 640 | ((save_policy[r] == 'E') && exclude_soe)) { |
duke@435 | 641 | proj->_rout.Insert(r); |
duke@435 | 642 | } |
duke@435 | 643 | } |
duke@435 | 644 | } |
duke@435 | 645 | |
duke@435 | 646 | return node_cnt; |
duke@435 | 647 | } |
duke@435 | 648 | |
duke@435 | 649 | |
duke@435 | 650 | //------------------------------schedule_local--------------------------------- |
duke@435 | 651 | // Topological sort within a block. Someday become a real scheduler. |
duke@435 | 652 | bool Block::schedule_local(PhaseCFG *cfg, Matcher &matcher, int *ready_cnt, VectorSet &next_call) { |
duke@435 | 653 | // Already "sorted" are the block start Node (as the first entry), and |
duke@435 | 654 | // the block-ending Node and any trailing control projections. We leave |
duke@435 | 655 | // these alone. PhiNodes and ParmNodes are made to follow the block start |
duke@435 | 656 | // Node. Everything else gets topo-sorted. |
duke@435 | 657 | |
duke@435 | 658 | #ifndef PRODUCT |
duke@435 | 659 | if (cfg->trace_opto_pipelining()) { |
duke@435 | 660 | tty->print_cr("# --- schedule_local B%d, before: ---", _pre_order); |
duke@435 | 661 | for (uint i = 0;i < _nodes.size();i++) { |
duke@435 | 662 | tty->print("# "); |
duke@435 | 663 | _nodes[i]->fast_dump(); |
duke@435 | 664 | } |
duke@435 | 665 | tty->print_cr("#"); |
duke@435 | 666 | } |
duke@435 | 667 | #endif |
duke@435 | 668 | |
duke@435 | 669 | // RootNode is already sorted |
duke@435 | 670 | if( _nodes.size() == 1 ) return true; |
duke@435 | 671 | |
duke@435 | 672 | // Move PhiNodes and ParmNodes from 1 to cnt up to the start |
duke@435 | 673 | uint node_cnt = end_idx(); |
duke@435 | 674 | uint phi_cnt = 1; |
duke@435 | 675 | uint i; |
duke@435 | 676 | for( i = 1; i<node_cnt; i++ ) { // Scan for Phi |
duke@435 | 677 | Node *n = _nodes[i]; |
duke@435 | 678 | if( n->is_Phi() || // Found a PhiNode or ParmNode |
duke@435 | 679 | (n->is_Proj() && n->in(0) == head()) ) { |
duke@435 | 680 | // Move guy at 'phi_cnt' to the end; makes a hole at phi_cnt |
duke@435 | 681 | _nodes.map(i,_nodes[phi_cnt]); |
duke@435 | 682 | _nodes.map(phi_cnt++,n); // swap Phi/Parm up front |
duke@435 | 683 | } else { // All others |
duke@435 | 684 | // Count block-local inputs to 'n' |
duke@435 | 685 | uint cnt = n->len(); // Input count |
duke@435 | 686 | uint local = 0; |
duke@435 | 687 | for( uint j=0; j<cnt; j++ ) { |
duke@435 | 688 | Node *m = n->in(j); |
duke@435 | 689 | if( m && cfg->_bbs[m->_idx] == this && !m->is_top() ) |
duke@435 | 690 | local++; // One more block-local input |
duke@435 | 691 | } |
duke@435 | 692 | ready_cnt[n->_idx] = local; // Count em up |
duke@435 | 693 | |
never@2780 | 694 | #ifdef ASSERT |
never@2780 | 695 | if( UseConcMarkSweepGC || UseG1GC ) { |
never@2780 | 696 | if( n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_StoreCM ) { |
never@2780 | 697 | // Check the precedence edges |
never@2780 | 698 | for (uint prec = n->req(); prec < n->len(); prec++) { |
never@2780 | 699 | Node* oop_store = n->in(prec); |
never@2780 | 700 | if (oop_store != NULL) { |
never@2780 | 701 | assert(cfg->_bbs[oop_store->_idx]->_dom_depth <= this->_dom_depth, "oop_store must dominate card-mark"); |
never@2780 | 702 | } |
never@2780 | 703 | } |
never@2780 | 704 | } |
never@2780 | 705 | } |
never@2780 | 706 | #endif |
never@2780 | 707 | |
duke@435 | 708 | // A few node types require changing a required edge to a precedence edge |
duke@435 | 709 | // before allocation. |
kvn@1535 | 710 | if( n->is_Mach() && n->req() > TypeFunc::Parms && |
kvn@1535 | 711 | (n->as_Mach()->ideal_Opcode() == Op_MemBarAcquire || |
kvn@1535 | 712 | n->as_Mach()->ideal_Opcode() == Op_MemBarVolatile) ) { |
kvn@688 | 713 | // MemBarAcquire could be created without Precedent edge. |
kvn@688 | 714 | // del_req() replaces the specified edge with the last input edge |
kvn@688 | 715 | // and then removes the last edge. If the specified edge > number of |
kvn@688 | 716 | // edges the last edge will be moved outside of the input edges array |
kvn@688 | 717 | // and the edge will be lost. This is why this code should be |
kvn@688 | 718 | // executed only when Precedent (== TypeFunc::Parms) edge is present. |
duke@435 | 719 | Node *x = n->in(TypeFunc::Parms); |
duke@435 | 720 | n->del_req(TypeFunc::Parms); |
duke@435 | 721 | n->add_prec(x); |
duke@435 | 722 | } |
duke@435 | 723 | } |
duke@435 | 724 | } |
duke@435 | 725 | for(uint i2=i; i2<_nodes.size(); i2++ ) // Trailing guys get zapped count |
duke@435 | 726 | ready_cnt[_nodes[i2]->_idx] = 0; |
duke@435 | 727 | |
duke@435 | 728 | // All the prescheduled guys do not hold back internal nodes |
duke@435 | 729 | uint i3; |
duke@435 | 730 | for(i3 = 0; i3<phi_cnt; i3++ ) { // For all pre-scheduled |
duke@435 | 731 | Node *n = _nodes[i3]; // Get pre-scheduled |
duke@435 | 732 | for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) { |
duke@435 | 733 | Node* m = n->fast_out(j); |
duke@435 | 734 | if( cfg->_bbs[m->_idx] ==this ) // Local-block user |
duke@435 | 735 | ready_cnt[m->_idx]--; // Fix ready count |
duke@435 | 736 | } |
duke@435 | 737 | } |
duke@435 | 738 | |
duke@435 | 739 | Node_List delay; |
duke@435 | 740 | // Make a worklist |
duke@435 | 741 | Node_List worklist; |
duke@435 | 742 | for(uint i4=i3; i4<node_cnt; i4++ ) { // Put ready guys on worklist |
duke@435 | 743 | Node *m = _nodes[i4]; |
duke@435 | 744 | if( !ready_cnt[m->_idx] ) { // Zero ready count? |
duke@435 | 745 | if (m->is_iteratively_computed()) { |
duke@435 | 746 | // Push induction variable increments last to allow other uses |
duke@435 | 747 | // of the phi to be scheduled first. The select() method breaks |
duke@435 | 748 | // ties in scheduling by worklist order. |
duke@435 | 749 | delay.push(m); |
never@560 | 750 | } else if (m->is_Mach() && m->as_Mach()->ideal_Opcode() == Op_CreateEx) { |
never@560 | 751 | // Force the CreateEx to the top of the list so it's processed |
never@560 | 752 | // first and ends up at the start of the block. |
never@560 | 753 | worklist.insert(0, m); |
duke@435 | 754 | } else { |
duke@435 | 755 | worklist.push(m); // Then on to worklist! |
duke@435 | 756 | } |
duke@435 | 757 | } |
duke@435 | 758 | } |
duke@435 | 759 | while (delay.size()) { |
duke@435 | 760 | Node* d = delay.pop(); |
duke@435 | 761 | worklist.push(d); |
duke@435 | 762 | } |
duke@435 | 763 | |
duke@435 | 764 | // Warm up the 'next_call' heuristic bits |
duke@435 | 765 | needed_for_next_call(_nodes[0], next_call, cfg->_bbs); |
duke@435 | 766 | |
duke@435 | 767 | #ifndef PRODUCT |
duke@435 | 768 | if (cfg->trace_opto_pipelining()) { |
duke@435 | 769 | for (uint j=0; j<_nodes.size(); j++) { |
duke@435 | 770 | Node *n = _nodes[j]; |
duke@435 | 771 | int idx = n->_idx; |
duke@435 | 772 | tty->print("# ready cnt:%3d ", ready_cnt[idx]); |
kvn@2040 | 773 | tty->print("latency:%3d ", cfg->_node_latency->at_grow(idx)); |
duke@435 | 774 | tty->print("%4d: %s\n", idx, n->Name()); |
duke@435 | 775 | } |
duke@435 | 776 | } |
duke@435 | 777 | #endif |
duke@435 | 778 | |
duke@435 | 779 | // Pull from worklist and schedule |
duke@435 | 780 | while( worklist.size() ) { // Worklist is not ready |
duke@435 | 781 | |
duke@435 | 782 | #ifndef PRODUCT |
duke@435 | 783 | if (cfg->trace_opto_pipelining()) { |
duke@435 | 784 | tty->print("# ready list:"); |
duke@435 | 785 | for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist |
duke@435 | 786 | Node *n = worklist[i]; // Get Node on worklist |
duke@435 | 787 | tty->print(" %d", n->_idx); |
duke@435 | 788 | } |
duke@435 | 789 | tty->cr(); |
duke@435 | 790 | } |
duke@435 | 791 | #endif |
duke@435 | 792 | |
duke@435 | 793 | // Select and pop a ready guy from worklist |
duke@435 | 794 | Node* n = select(cfg, worklist, ready_cnt, next_call, phi_cnt); |
duke@435 | 795 | _nodes.map(phi_cnt++,n); // Schedule him next |
duke@435 | 796 | |
duke@435 | 797 | #ifndef PRODUCT |
duke@435 | 798 | if (cfg->trace_opto_pipelining()) { |
duke@435 | 799 | tty->print("# select %d: %s", n->_idx, n->Name()); |
kvn@2040 | 800 | tty->print(", latency:%d", cfg->_node_latency->at_grow(n->_idx)); |
duke@435 | 801 | n->dump(); |
duke@435 | 802 | if (Verbose) { |
duke@435 | 803 | tty->print("# ready list:"); |
duke@435 | 804 | for( uint i=0; i<worklist.size(); i++ ) { // Inspect entire worklist |
duke@435 | 805 | Node *n = worklist[i]; // Get Node on worklist |
duke@435 | 806 | tty->print(" %d", n->_idx); |
duke@435 | 807 | } |
duke@435 | 808 | tty->cr(); |
duke@435 | 809 | } |
duke@435 | 810 | } |
duke@435 | 811 | |
duke@435 | 812 | #endif |
duke@435 | 813 | if( n->is_MachCall() ) { |
duke@435 | 814 | MachCallNode *mcall = n->as_MachCall(); |
duke@435 | 815 | phi_cnt = sched_call(matcher, cfg->_bbs, phi_cnt, worklist, ready_cnt, mcall, next_call); |
duke@435 | 816 | continue; |
duke@435 | 817 | } |
duke@435 | 818 | // Children are now all ready |
duke@435 | 819 | for (DUIterator_Fast i5max, i5 = n->fast_outs(i5max); i5 < i5max; i5++) { |
duke@435 | 820 | Node* m = n->fast_out(i5); // Get user |
duke@435 | 821 | if( cfg->_bbs[m->_idx] != this ) continue; |
duke@435 | 822 | if( m->is_Phi() ) continue; |
duke@435 | 823 | if( !--ready_cnt[m->_idx] ) |
duke@435 | 824 | worklist.push(m); |
duke@435 | 825 | } |
duke@435 | 826 | } |
duke@435 | 827 | |
duke@435 | 828 | if( phi_cnt != end_idx() ) { |
duke@435 | 829 | // did not schedule all. Retry, Bailout, or Die |
duke@435 | 830 | Compile* C = matcher.C; |
duke@435 | 831 | if (C->subsume_loads() == true && !C->failing()) { |
duke@435 | 832 | // Retry with subsume_loads == false |
duke@435 | 833 | // If this is the first failure, the sentinel string will "stick" |
duke@435 | 834 | // to the Compile object, and the C2Compiler will see it and retry. |
duke@435 | 835 | C->record_failure(C2Compiler::retry_no_subsuming_loads()); |
duke@435 | 836 | } |
duke@435 | 837 | // assert( phi_cnt == end_idx(), "did not schedule all" ); |
duke@435 | 838 | return false; |
duke@435 | 839 | } |
duke@435 | 840 | |
duke@435 | 841 | #ifndef PRODUCT |
duke@435 | 842 | if (cfg->trace_opto_pipelining()) { |
duke@435 | 843 | tty->print_cr("#"); |
duke@435 | 844 | tty->print_cr("# after schedule_local"); |
duke@435 | 845 | for (uint i = 0;i < _nodes.size();i++) { |
duke@435 | 846 | tty->print("# "); |
duke@435 | 847 | _nodes[i]->fast_dump(); |
duke@435 | 848 | } |
duke@435 | 849 | tty->cr(); |
duke@435 | 850 | } |
duke@435 | 851 | #endif |
duke@435 | 852 | |
duke@435 | 853 | |
duke@435 | 854 | return true; |
duke@435 | 855 | } |
duke@435 | 856 | |
duke@435 | 857 | //--------------------------catch_cleanup_fix_all_inputs----------------------- |
duke@435 | 858 | static void catch_cleanup_fix_all_inputs(Node *use, Node *old_def, Node *new_def) { |
duke@435 | 859 | for (uint l = 0; l < use->len(); l++) { |
duke@435 | 860 | if (use->in(l) == old_def) { |
duke@435 | 861 | if (l < use->req()) { |
duke@435 | 862 | use->set_req(l, new_def); |
duke@435 | 863 | } else { |
duke@435 | 864 | use->rm_prec(l); |
duke@435 | 865 | use->add_prec(new_def); |
duke@435 | 866 | l--; |
duke@435 | 867 | } |
duke@435 | 868 | } |
duke@435 | 869 | } |
duke@435 | 870 | } |
duke@435 | 871 | |
duke@435 | 872 | //------------------------------catch_cleanup_find_cloned_def------------------ |
duke@435 | 873 | static Node *catch_cleanup_find_cloned_def(Block *use_blk, Node *def, Block *def_blk, Block_Array &bbs, int n_clone_idx) { |
duke@435 | 874 | assert( use_blk != def_blk, "Inter-block cleanup only"); |
duke@435 | 875 | |
duke@435 | 876 | // The use is some block below the Catch. Find and return the clone of the def |
duke@435 | 877 | // that dominates the use. If there is no clone in a dominating block, then |
duke@435 | 878 | // create a phi for the def in a dominating block. |
duke@435 | 879 | |
duke@435 | 880 | // Find which successor block dominates this use. The successor |
duke@435 | 881 | // blocks must all be single-entry (from the Catch only; I will have |
duke@435 | 882 | // split blocks to make this so), hence they all dominate. |
duke@435 | 883 | while( use_blk->_dom_depth > def_blk->_dom_depth+1 ) |
duke@435 | 884 | use_blk = use_blk->_idom; |
duke@435 | 885 | |
duke@435 | 886 | // Find the successor |
duke@435 | 887 | Node *fixup = NULL; |
duke@435 | 888 | |
duke@435 | 889 | uint j; |
duke@435 | 890 | for( j = 0; j < def_blk->_num_succs; j++ ) |
duke@435 | 891 | if( use_blk == def_blk->_succs[j] ) |
duke@435 | 892 | break; |
duke@435 | 893 | |
duke@435 | 894 | if( j == def_blk->_num_succs ) { |
duke@435 | 895 | // Block at same level in dom-tree is not a successor. It needs a |
duke@435 | 896 | // PhiNode, the PhiNode uses from the def and IT's uses need fixup. |
duke@435 | 897 | Node_Array inputs = new Node_List(Thread::current()->resource_area()); |
duke@435 | 898 | for(uint k = 1; k < use_blk->num_preds(); k++) { |
duke@435 | 899 | inputs.map(k, catch_cleanup_find_cloned_def(bbs[use_blk->pred(k)->_idx], def, def_blk, bbs, n_clone_idx)); |
duke@435 | 900 | } |
duke@435 | 901 | |
duke@435 | 902 | // Check to see if the use_blk already has an identical phi inserted. |
duke@435 | 903 | // If it exists, it will be at the first position since all uses of a |
duke@435 | 904 | // def are processed together. |
duke@435 | 905 | Node *phi = use_blk->_nodes[1]; |
duke@435 | 906 | if( phi->is_Phi() ) { |
duke@435 | 907 | fixup = phi; |
duke@435 | 908 | for (uint k = 1; k < use_blk->num_preds(); k++) { |
duke@435 | 909 | if (phi->in(k) != inputs[k]) { |
duke@435 | 910 | // Not a match |
duke@435 | 911 | fixup = NULL; |
duke@435 | 912 | break; |
duke@435 | 913 | } |
duke@435 | 914 | } |
duke@435 | 915 | } |
duke@435 | 916 | |
duke@435 | 917 | // If an existing PhiNode was not found, make a new one. |
duke@435 | 918 | if (fixup == NULL) { |
duke@435 | 919 | Node *new_phi = PhiNode::make(use_blk->head(), def); |
duke@435 | 920 | use_blk->_nodes.insert(1, new_phi); |
duke@435 | 921 | bbs.map(new_phi->_idx, use_blk); |
duke@435 | 922 | for (uint k = 1; k < use_blk->num_preds(); k++) { |
duke@435 | 923 | new_phi->set_req(k, inputs[k]); |
duke@435 | 924 | } |
duke@435 | 925 | fixup = new_phi; |
duke@435 | 926 | } |
duke@435 | 927 | |
duke@435 | 928 | } else { |
duke@435 | 929 | // Found the use just below the Catch. Make it use the clone. |
duke@435 | 930 | fixup = use_blk->_nodes[n_clone_idx]; |
duke@435 | 931 | } |
duke@435 | 932 | |
duke@435 | 933 | return fixup; |
duke@435 | 934 | } |
duke@435 | 935 | |
duke@435 | 936 | //--------------------------catch_cleanup_intra_block-------------------------- |
duke@435 | 937 | // Fix all input edges in use that reference "def". The use is in the same |
duke@435 | 938 | // block as the def and both have been cloned in each successor block. |
duke@435 | 939 | static void catch_cleanup_intra_block(Node *use, Node *def, Block *blk, int beg, int n_clone_idx) { |
duke@435 | 940 | |
duke@435 | 941 | // Both the use and def have been cloned. For each successor block, |
duke@435 | 942 | // get the clone of the use, and make its input the clone of the def |
duke@435 | 943 | // found in that block. |
duke@435 | 944 | |
duke@435 | 945 | uint use_idx = blk->find_node(use); |
duke@435 | 946 | uint offset_idx = use_idx - beg; |
duke@435 | 947 | for( uint k = 0; k < blk->_num_succs; k++ ) { |
duke@435 | 948 | // Get clone in each successor block |
duke@435 | 949 | Block *sb = blk->_succs[k]; |
duke@435 | 950 | Node *clone = sb->_nodes[offset_idx+1]; |
duke@435 | 951 | assert( clone->Opcode() == use->Opcode(), "" ); |
duke@435 | 952 | |
duke@435 | 953 | // Make use-clone reference the def-clone |
duke@435 | 954 | catch_cleanup_fix_all_inputs(clone, def, sb->_nodes[n_clone_idx]); |
duke@435 | 955 | } |
duke@435 | 956 | } |
duke@435 | 957 | |
duke@435 | 958 | //------------------------------catch_cleanup_inter_block--------------------- |
duke@435 | 959 | // Fix all input edges in use that reference "def". The use is in a different |
duke@435 | 960 | // block than the def. |
duke@435 | 961 | static void catch_cleanup_inter_block(Node *use, Block *use_blk, Node *def, Block *def_blk, Block_Array &bbs, int n_clone_idx) { |
duke@435 | 962 | if( !use_blk ) return; // Can happen if the use is a precedence edge |
duke@435 | 963 | |
duke@435 | 964 | Node *new_def = catch_cleanup_find_cloned_def(use_blk, def, def_blk, bbs, n_clone_idx); |
duke@435 | 965 | catch_cleanup_fix_all_inputs(use, def, new_def); |
duke@435 | 966 | } |
duke@435 | 967 | |
duke@435 | 968 | //------------------------------call_catch_cleanup----------------------------- |
duke@435 | 969 | // If we inserted any instructions between a Call and his CatchNode, |
duke@435 | 970 | // clone the instructions on all paths below the Catch. |
duke@435 | 971 | void Block::call_catch_cleanup(Block_Array &bbs) { |
duke@435 | 972 | |
duke@435 | 973 | // End of region to clone |
duke@435 | 974 | uint end = end_idx(); |
duke@435 | 975 | if( !_nodes[end]->is_Catch() ) return; |
duke@435 | 976 | // Start of region to clone |
duke@435 | 977 | uint beg = end; |
kvn@3040 | 978 | while(!_nodes[beg-1]->is_MachProj() || |
kvn@3040 | 979 | !_nodes[beg-1]->in(0)->is_MachCall() ) { |
duke@435 | 980 | beg--; |
duke@435 | 981 | assert(beg > 0,"Catch cleanup walking beyond block boundary"); |
duke@435 | 982 | } |
duke@435 | 983 | // Range of inserted instructions is [beg, end) |
duke@435 | 984 | if( beg == end ) return; |
duke@435 | 985 | |
duke@435 | 986 | // Clone along all Catch output paths. Clone area between the 'beg' and |
duke@435 | 987 | // 'end' indices. |
duke@435 | 988 | for( uint i = 0; i < _num_succs; i++ ) { |
duke@435 | 989 | Block *sb = _succs[i]; |
duke@435 | 990 | // Clone the entire area; ignoring the edge fixup for now. |
duke@435 | 991 | for( uint j = end; j > beg; j-- ) { |
kvn@2048 | 992 | // It is safe here to clone a node with anti_dependence |
kvn@2048 | 993 | // since clones dominate on each path. |
duke@435 | 994 | Node *clone = _nodes[j-1]->clone(); |
duke@435 | 995 | sb->_nodes.insert( 1, clone ); |
duke@435 | 996 | bbs.map(clone->_idx,sb); |
duke@435 | 997 | } |
duke@435 | 998 | } |
duke@435 | 999 | |
duke@435 | 1000 | |
duke@435 | 1001 | // Fixup edges. Check the def-use info per cloned Node |
duke@435 | 1002 | for(uint i2 = beg; i2 < end; i2++ ) { |
duke@435 | 1003 | uint n_clone_idx = i2-beg+1; // Index of clone of n in each successor block |
duke@435 | 1004 | Node *n = _nodes[i2]; // Node that got cloned |
duke@435 | 1005 | // Need DU safe iterator because of edge manipulation in calls. |
duke@435 | 1006 | Unique_Node_List *out = new Unique_Node_List(Thread::current()->resource_area()); |
duke@435 | 1007 | for (DUIterator_Fast j1max, j1 = n->fast_outs(j1max); j1 < j1max; j1++) { |
duke@435 | 1008 | out->push(n->fast_out(j1)); |
duke@435 | 1009 | } |
duke@435 | 1010 | uint max = out->size(); |
duke@435 | 1011 | for (uint j = 0; j < max; j++) {// For all users |
duke@435 | 1012 | Node *use = out->pop(); |
duke@435 | 1013 | Block *buse = bbs[use->_idx]; |
duke@435 | 1014 | if( use->is_Phi() ) { |
duke@435 | 1015 | for( uint k = 1; k < use->req(); k++ ) |
duke@435 | 1016 | if( use->in(k) == n ) { |
duke@435 | 1017 | Node *fixup = catch_cleanup_find_cloned_def(bbs[buse->pred(k)->_idx], n, this, bbs, n_clone_idx); |
duke@435 | 1018 | use->set_req(k, fixup); |
duke@435 | 1019 | } |
duke@435 | 1020 | } else { |
duke@435 | 1021 | if (this == buse) { |
duke@435 | 1022 | catch_cleanup_intra_block(use, n, this, beg, n_clone_idx); |
duke@435 | 1023 | } else { |
duke@435 | 1024 | catch_cleanup_inter_block(use, buse, n, this, bbs, n_clone_idx); |
duke@435 | 1025 | } |
duke@435 | 1026 | } |
duke@435 | 1027 | } // End for all users |
duke@435 | 1028 | |
duke@435 | 1029 | } // End of for all Nodes in cloned area |
duke@435 | 1030 | |
duke@435 | 1031 | // Remove the now-dead cloned ops |
duke@435 | 1032 | for(uint i3 = beg; i3 < end; i3++ ) { |
duke@435 | 1033 | _nodes[beg]->disconnect_inputs(NULL); |
duke@435 | 1034 | _nodes.remove(beg); |
duke@435 | 1035 | } |
duke@435 | 1036 | |
duke@435 | 1037 | // If the successor blocks have a CreateEx node, move it back to the top |
duke@435 | 1038 | for(uint i4 = 0; i4 < _num_succs; i4++ ) { |
duke@435 | 1039 | Block *sb = _succs[i4]; |
duke@435 | 1040 | uint new_cnt = end - beg; |
duke@435 | 1041 | // Remove any newly created, but dead, nodes. |
duke@435 | 1042 | for( uint j = new_cnt; j > 0; j-- ) { |
duke@435 | 1043 | Node *n = sb->_nodes[j]; |
duke@435 | 1044 | if (n->outcnt() == 0 && |
duke@435 | 1045 | (!n->is_Proj() || n->as_Proj()->in(0)->outcnt() == 1) ){ |
duke@435 | 1046 | n->disconnect_inputs(NULL); |
duke@435 | 1047 | sb->_nodes.remove(j); |
duke@435 | 1048 | new_cnt--; |
duke@435 | 1049 | } |
duke@435 | 1050 | } |
duke@435 | 1051 | // If any newly created nodes remain, move the CreateEx node to the top |
duke@435 | 1052 | if (new_cnt > 0) { |
duke@435 | 1053 | Node *cex = sb->_nodes[1+new_cnt]; |
duke@435 | 1054 | if( cex->is_Mach() && cex->as_Mach()->ideal_Opcode() == Op_CreateEx ) { |
duke@435 | 1055 | sb->_nodes.remove(1+new_cnt); |
duke@435 | 1056 | sb->_nodes.insert(1,cex); |
duke@435 | 1057 | } |
duke@435 | 1058 | } |
duke@435 | 1059 | } |
duke@435 | 1060 | } |