1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000 1.2 +++ b/src/share/vm/opto/buildOopMap.cpp Sat Dec 01 00:00:00 2007 +0000 1.3 @@ -0,0 +1,623 @@ 1.4 +/* 1.5 + * Copyright 2002-2007 Sun Microsystems, Inc. All Rights Reserved. 1.6 + * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 1.7 + * 1.8 + * This code is free software; you can redistribute it and/or modify it 1.9 + * under the terms of the GNU General Public License version 2 only, as 1.10 + * published by the Free Software Foundation. 1.11 + * 1.12 + * This code is distributed in the hope that it will be useful, but WITHOUT 1.13 + * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 1.14 + * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 1.15 + * version 2 for more details (a copy is included in the LICENSE file that 1.16 + * accompanied this code). 1.17 + * 1.18 + * You should have received a copy of the GNU General Public License version 1.19 + * 2 along with this work; if not, write to the Free Software Foundation, 1.20 + * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 1.21 + * 1.22 + * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, 1.23 + * CA 95054 USA or visit www.sun.com if you need additional information or 1.24 + * have any questions. 1.25 + * 1.26 + */ 1.27 + 1.28 +#include "incls/_precompiled.incl" 1.29 +#include "incls/_buildOopMap.cpp.incl" 1.30 + 1.31 +// The functions in this file builds OopMaps after all scheduling is done. 1.32 +// 1.33 +// OopMaps contain a list of all registers and stack-slots containing oops (so 1.34 +// they can be updated by GC). OopMaps also contain a list of derived-pointer 1.35 +// base-pointer pairs. When the base is moved, the derived pointer moves to 1.36 +// follow it. Finally, any registers holding callee-save values are also 1.37 +// recorded. These might contain oops, but only the caller knows. 1.38 +// 1.39 +// BuildOopMaps implements a simple forward reaching-defs solution. At each 1.40 +// GC point we'll have the reaching-def Nodes. If the reaching Nodes are 1.41 +// typed as pointers (no offset), then they are oops. Pointers+offsets are 1.42 +// derived pointers, and bases can be found from them. Finally, we'll also 1.43 +// track reaching callee-save values. Note that a copy of a callee-save value 1.44 +// "kills" it's source, so that only 1 copy of a callee-save value is alive at 1.45 +// a time. 1.46 +// 1.47 +// We run a simple bitvector liveness pass to help trim out dead oops. Due to 1.48 +// irreducible loops, we can have a reaching def of an oop that only reaches 1.49 +// along one path and no way to know if it's valid or not on the other path. 1.50 +// The bitvectors are quite dense and the liveness pass is fast. 1.51 +// 1.52 +// At GC points, we consult this information to build OopMaps. All reaching 1.53 +// defs typed as oops are added to the OopMap. Only 1 instance of a 1.54 +// callee-save register can be recorded. For derived pointers, we'll have to 1.55 +// find and record the register holding the base. 1.56 +// 1.57 +// The reaching def's is a simple 1-pass worklist approach. I tried a clever 1.58 +// breadth-first approach but it was worse (showed O(n^2) in the 1.59 +// pick-next-block code). 1.60 +// 1.61 +// The relevent data is kept in a struct of arrays (it could just as well be 1.62 +// an array of structs, but the struct-of-arrays is generally a little more 1.63 +// efficient). The arrays are indexed by register number (including 1.64 +// stack-slots as registers) and so is bounded by 200 to 300 elements in 1.65 +// practice. One array will map to a reaching def Node (or NULL for 1.66 +// conflict/dead). The other array will map to a callee-saved register or 1.67 +// OptoReg::Bad for not-callee-saved. 1.68 + 1.69 + 1.70 +//------------------------------OopFlow---------------------------------------- 1.71 +// Structure to pass around 1.72 +struct OopFlow : public ResourceObj { 1.73 + short *_callees; // Array mapping register to callee-saved 1.74 + Node **_defs; // array mapping register to reaching def 1.75 + // or NULL if dead/conflict 1.76 + // OopFlow structs, when not being actively modified, describe the _end_ of 1.77 + // this block. 1.78 + Block *_b; // Block for this struct 1.79 + OopFlow *_next; // Next free OopFlow 1.80 + 1.81 + OopFlow( short *callees, Node **defs ) : _callees(callees), _defs(defs), 1.82 + _b(NULL), _next(NULL) { } 1.83 + 1.84 + // Given reaching-defs for this block start, compute it for this block end 1.85 + void compute_reach( PhaseRegAlloc *regalloc, int max_reg, Dict *safehash ); 1.86 + 1.87 + // Merge these two OopFlows into the 'this' pointer. 1.88 + void merge( OopFlow *flow, int max_reg ); 1.89 + 1.90 + // Copy a 'flow' over an existing flow 1.91 + void clone( OopFlow *flow, int max_size); 1.92 + 1.93 + // Make a new OopFlow from scratch 1.94 + static OopFlow *make( Arena *A, int max_size ); 1.95 + 1.96 + // Build an oopmap from the current flow info 1.97 + OopMap *build_oop_map( Node *n, int max_reg, PhaseRegAlloc *regalloc, int* live ); 1.98 +}; 1.99 + 1.100 +//------------------------------compute_reach---------------------------------- 1.101 +// Given reaching-defs for this block start, compute it for this block end 1.102 +void OopFlow::compute_reach( PhaseRegAlloc *regalloc, int max_reg, Dict *safehash ) { 1.103 + 1.104 + for( uint i=0; i<_b->_nodes.size(); i++ ) { 1.105 + Node *n = _b->_nodes[i]; 1.106 + 1.107 + if( n->jvms() ) { // Build an OopMap here? 1.108 + JVMState *jvms = n->jvms(); 1.109 + // no map needed for leaf calls 1.110 + if( n->is_MachSafePoint() && !n->is_MachCallLeaf() ) { 1.111 + int *live = (int*) (*safehash)[n]; 1.112 + assert( live, "must find live" ); 1.113 + n->as_MachSafePoint()->set_oop_map( build_oop_map(n,max_reg,regalloc, live) ); 1.114 + } 1.115 + } 1.116 + 1.117 + // Assign new reaching def's. 1.118 + // Note that I padded the _defs and _callees arrays so it's legal 1.119 + // to index at _defs[OptoReg::Bad]. 1.120 + OptoReg::Name first = regalloc->get_reg_first(n); 1.121 + OptoReg::Name second = regalloc->get_reg_second(n); 1.122 + _defs[first] = n; 1.123 + _defs[second] = n; 1.124 + 1.125 + // Pass callee-save info around copies 1.126 + int idx = n->is_Copy(); 1.127 + if( idx ) { // Copies move callee-save info 1.128 + OptoReg::Name old_first = regalloc->get_reg_first(n->in(idx)); 1.129 + OptoReg::Name old_second = regalloc->get_reg_second(n->in(idx)); 1.130 + int tmp_first = _callees[old_first]; 1.131 + int tmp_second = _callees[old_second]; 1.132 + _callees[old_first] = OptoReg::Bad; // callee-save is moved, dead in old location 1.133 + _callees[old_second] = OptoReg::Bad; 1.134 + _callees[first] = tmp_first; 1.135 + _callees[second] = tmp_second; 1.136 + } else if( n->is_Phi() ) { // Phis do not mod callee-saves 1.137 + assert( _callees[first] == _callees[regalloc->get_reg_first(n->in(1))], "" ); 1.138 + assert( _callees[second] == _callees[regalloc->get_reg_second(n->in(1))], "" ); 1.139 + assert( _callees[first] == _callees[regalloc->get_reg_first(n->in(n->req()-1))], "" ); 1.140 + assert( _callees[second] == _callees[regalloc->get_reg_second(n->in(n->req()-1))], "" ); 1.141 + } else { 1.142 + _callees[first] = OptoReg::Bad; // No longer holding a callee-save value 1.143 + _callees[second] = OptoReg::Bad; 1.144 + 1.145 + // Find base case for callee saves 1.146 + if( n->is_Proj() && n->in(0)->is_Start() ) { 1.147 + if( OptoReg::is_reg(first) && 1.148 + regalloc->_matcher.is_save_on_entry(first) ) 1.149 + _callees[first] = first; 1.150 + if( OptoReg::is_reg(second) && 1.151 + regalloc->_matcher.is_save_on_entry(second) ) 1.152 + _callees[second] = second; 1.153 + } 1.154 + } 1.155 + } 1.156 +} 1.157 + 1.158 +//------------------------------merge------------------------------------------ 1.159 +// Merge the given flow into the 'this' flow 1.160 +void OopFlow::merge( OopFlow *flow, int max_reg ) { 1.161 + assert( _b == NULL, "merging into a happy flow" ); 1.162 + assert( flow->_b, "this flow is still alive" ); 1.163 + assert( flow != this, "no self flow" ); 1.164 + 1.165 + // Do the merge. If there are any differences, drop to 'bottom' which 1.166 + // is OptoReg::Bad or NULL depending. 1.167 + for( int i=0; i<max_reg; i++ ) { 1.168 + // Merge the callee-save's 1.169 + if( _callees[i] != flow->_callees[i] ) 1.170 + _callees[i] = OptoReg::Bad; 1.171 + // Merge the reaching defs 1.172 + if( _defs[i] != flow->_defs[i] ) 1.173 + _defs[i] = NULL; 1.174 + } 1.175 + 1.176 +} 1.177 + 1.178 +//------------------------------clone------------------------------------------ 1.179 +void OopFlow::clone( OopFlow *flow, int max_size ) { 1.180 + _b = flow->_b; 1.181 + memcpy( _callees, flow->_callees, sizeof(short)*max_size); 1.182 + memcpy( _defs , flow->_defs , sizeof(Node*)*max_size); 1.183 +} 1.184 + 1.185 +//------------------------------make------------------------------------------- 1.186 +OopFlow *OopFlow::make( Arena *A, int max_size ) { 1.187 + short *callees = NEW_ARENA_ARRAY(A,short,max_size+1); 1.188 + Node **defs = NEW_ARENA_ARRAY(A,Node*,max_size+1); 1.189 + debug_only( memset(defs,0,(max_size+1)*sizeof(Node*)) ); 1.190 + OopFlow *flow = new (A) OopFlow(callees+1, defs+1); 1.191 + assert( &flow->_callees[OptoReg::Bad] == callees, "Ok to index at OptoReg::Bad" ); 1.192 + assert( &flow->_defs [OptoReg::Bad] == defs , "Ok to index at OptoReg::Bad" ); 1.193 + return flow; 1.194 +} 1.195 + 1.196 +//------------------------------bit twiddlers---------------------------------- 1.197 +static int get_live_bit( int *live, int reg ) { 1.198 + return live[reg>>LogBitsPerInt] & (1<<(reg&(BitsPerInt-1))); } 1.199 +static void set_live_bit( int *live, int reg ) { 1.200 + live[reg>>LogBitsPerInt] |= (1<<(reg&(BitsPerInt-1))); } 1.201 +static void clr_live_bit( int *live, int reg ) { 1.202 + live[reg>>LogBitsPerInt] &= ~(1<<(reg&(BitsPerInt-1))); } 1.203 + 1.204 +//------------------------------build_oop_map---------------------------------- 1.205 +// Build an oopmap from the current flow info 1.206 +OopMap *OopFlow::build_oop_map( Node *n, int max_reg, PhaseRegAlloc *regalloc, int* live ) { 1.207 + int framesize = regalloc->_framesize; 1.208 + int max_inarg_slot = OptoReg::reg2stack(regalloc->_matcher._new_SP); 1.209 + debug_only( char *dup_check = NEW_RESOURCE_ARRAY(char,OptoReg::stack0()); 1.210 + memset(dup_check,0,OptoReg::stack0()) ); 1.211 + 1.212 + OopMap *omap = new OopMap( framesize, max_inarg_slot ); 1.213 + MachCallNode *mcall = n->is_MachCall() ? n->as_MachCall() : NULL; 1.214 + JVMState* jvms = n->jvms(); 1.215 + 1.216 + // For all registers do... 1.217 + for( int reg=0; reg<max_reg; reg++ ) { 1.218 + if( get_live_bit(live,reg) == 0 ) 1.219 + continue; // Ignore if not live 1.220 + 1.221 + // %%% C2 can use 2 OptoRegs when the physical register is only one 64bit 1.222 + // register in that case we'll get an non-concrete register for the second 1.223 + // half. We only need to tell the map the register once! 1.224 + // 1.225 + // However for the moment we disable this change and leave things as they 1.226 + // were. 1.227 + 1.228 + VMReg r = OptoReg::as_VMReg(OptoReg::Name(reg), framesize, max_inarg_slot); 1.229 + 1.230 + if (false && r->is_reg() && !r->is_concrete()) { 1.231 + continue; 1.232 + } 1.233 + 1.234 + // See if dead (no reaching def). 1.235 + Node *def = _defs[reg]; // Get reaching def 1.236 + assert( def, "since live better have reaching def" ); 1.237 + 1.238 + // Classify the reaching def as oop, derived, callee-save, dead, or other 1.239 + const Type *t = def->bottom_type(); 1.240 + if( t->isa_oop_ptr() ) { // Oop or derived? 1.241 + assert( !OptoReg::is_valid(_callees[reg]), "oop can't be callee save" ); 1.242 +#ifdef _LP64 1.243 + // 64-bit pointers record oop-ishness on 2 aligned adjacent registers. 1.244 + // Make sure both are record from the same reaching def, but do not 1.245 + // put both into the oopmap. 1.246 + if( (reg&1) == 1 ) { // High half of oop-pair? 1.247 + assert( _defs[reg-1] == _defs[reg], "both halves from same reaching def" ); 1.248 + continue; // Do not record high parts in oopmap 1.249 + } 1.250 +#endif 1.251 + 1.252 + // Check for a legal reg name in the oopMap and bailout if it is not. 1.253 + if (!omap->legal_vm_reg_name(r)) { 1.254 + regalloc->C->record_method_not_compilable("illegal oopMap register name"); 1.255 + continue; 1.256 + } 1.257 + if( t->is_ptr()->_offset == 0 ) { // Not derived? 1.258 + if( mcall ) { 1.259 + // Outgoing argument GC mask responsibility belongs to the callee, 1.260 + // not the caller. Inspect the inputs to the call, to see if 1.261 + // this live-range is one of them. 1.262 + uint cnt = mcall->tf()->domain()->cnt(); 1.263 + uint j; 1.264 + for( j = TypeFunc::Parms; j < cnt; j++) 1.265 + if( mcall->in(j) == def ) 1.266 + break; // reaching def is an argument oop 1.267 + if( j < cnt ) // arg oops dont go in GC map 1.268 + continue; // Continue on to the next register 1.269 + } 1.270 + omap->set_oop(r); 1.271 + } else { // Else it's derived. 1.272 + // Find the base of the derived value. 1.273 + uint i; 1.274 + // Fast, common case, scan 1.275 + for( i = jvms->oopoff(); i < n->req(); i+=2 ) 1.276 + if( n->in(i) == def ) break; // Common case 1.277 + if( i == n->req() ) { // Missed, try a more generous scan 1.278 + // Scan again, but this time peek through copies 1.279 + for( i = jvms->oopoff(); i < n->req(); i+=2 ) { 1.280 + Node *m = n->in(i); // Get initial derived value 1.281 + while( 1 ) { 1.282 + Node *d = def; // Get initial reaching def 1.283 + while( 1 ) { // Follow copies of reaching def to end 1.284 + if( m == d ) goto found; // breaks 3 loops 1.285 + int idx = d->is_Copy(); 1.286 + if( !idx ) break; 1.287 + d = d->in(idx); // Link through copy 1.288 + } 1.289 + int idx = m->is_Copy(); 1.290 + if( !idx ) break; 1.291 + m = m->in(idx); 1.292 + } 1.293 + } 1.294 + guarantee( 0, "must find derived/base pair" ); 1.295 + } 1.296 + found: ; 1.297 + Node *base = n->in(i+1); // Base is other half of pair 1.298 + int breg = regalloc->get_reg_first(base); 1.299 + VMReg b = OptoReg::as_VMReg(OptoReg::Name(breg), framesize, max_inarg_slot); 1.300 + 1.301 + // I record liveness at safepoints BEFORE I make the inputs 1.302 + // live. This is because argument oops are NOT live at a 1.303 + // safepoint (or at least they cannot appear in the oopmap). 1.304 + // Thus bases of base/derived pairs might not be in the 1.305 + // liveness data but they need to appear in the oopmap. 1.306 + if( get_live_bit(live,breg) == 0 ) {// Not live? 1.307 + // Flag it, so next derived pointer won't re-insert into oopmap 1.308 + set_live_bit(live,breg); 1.309 + // Already missed our turn? 1.310 + if( breg < reg ) { 1.311 + if (b->is_stack() || b->is_concrete() || true ) { 1.312 + omap->set_oop( b); 1.313 + } 1.314 + } 1.315 + } 1.316 + if (b->is_stack() || b->is_concrete() || true ) { 1.317 + omap->set_derived_oop( r, b); 1.318 + } 1.319 + } 1.320 + 1.321 + } else if( OptoReg::is_valid(_callees[reg])) { // callee-save? 1.322 + // It's a callee-save value 1.323 + assert( dup_check[_callees[reg]]==0, "trying to callee save same reg twice" ); 1.324 + debug_only( dup_check[_callees[reg]]=1; ) 1.325 + VMReg callee = OptoReg::as_VMReg(OptoReg::Name(_callees[reg])); 1.326 + if ( callee->is_concrete() || true ) { 1.327 + omap->set_callee_saved( r, callee); 1.328 + } 1.329 + 1.330 + } else { 1.331 + // Other - some reaching non-oop value 1.332 + omap->set_value( r); 1.333 + } 1.334 + 1.335 + } 1.336 + 1.337 +#ifdef ASSERT 1.338 + /* Nice, Intel-only assert 1.339 + int cnt_callee_saves=0; 1.340 + int reg2 = 0; 1.341 + while (OptoReg::is_reg(reg2)) { 1.342 + if( dup_check[reg2] != 0) cnt_callee_saves++; 1.343 + assert( cnt_callee_saves==3 || cnt_callee_saves==5, "missed some callee-save" ); 1.344 + reg2++; 1.345 + } 1.346 + */ 1.347 +#endif 1.348 + 1.349 + return omap; 1.350 +} 1.351 + 1.352 +//------------------------------do_liveness------------------------------------ 1.353 +// Compute backwards liveness on registers 1.354 +static void do_liveness( PhaseRegAlloc *regalloc, PhaseCFG *cfg, Block_List *worklist, int max_reg_ints, Arena *A, Dict *safehash ) { 1.355 + int *live = NEW_ARENA_ARRAY(A, int, (cfg->_num_blocks+1) * max_reg_ints); 1.356 + int *tmp_live = &live[cfg->_num_blocks * max_reg_ints]; 1.357 + Node *root = cfg->C->root(); 1.358 + // On CISC platforms, get the node representing the stack pointer that regalloc 1.359 + // used for spills 1.360 + Node *fp = NodeSentinel; 1.361 + if (UseCISCSpill && root->req() > 1) { 1.362 + fp = root->in(1)->in(TypeFunc::FramePtr); 1.363 + } 1.364 + memset( live, 0, cfg->_num_blocks * (max_reg_ints<<LogBytesPerInt) ); 1.365 + // Push preds onto worklist 1.366 + for( uint i=1; i<root->req(); i++ ) 1.367 + worklist->push(cfg->_bbs[root->in(i)->_idx]); 1.368 + 1.369 + // ZKM.jar includes tiny infinite loops which are unreached from below. 1.370 + // If we missed any blocks, we'll retry here after pushing all missed 1.371 + // blocks on the worklist. Normally this outer loop never trips more 1.372 + // than once. 1.373 + while( 1 ) { 1.374 + 1.375 + while( worklist->size() ) { // Standard worklist algorithm 1.376 + Block *b = worklist->rpop(); 1.377 + 1.378 + // Copy first successor into my tmp_live space 1.379 + int s0num = b->_succs[0]->_pre_order; 1.380 + int *t = &live[s0num*max_reg_ints]; 1.381 + for( int i=0; i<max_reg_ints; i++ ) 1.382 + tmp_live[i] = t[i]; 1.383 + 1.384 + // OR in the remaining live registers 1.385 + for( uint j=1; j<b->_num_succs; j++ ) { 1.386 + uint sjnum = b->_succs[j]->_pre_order; 1.387 + int *t = &live[sjnum*max_reg_ints]; 1.388 + for( int i=0; i<max_reg_ints; i++ ) 1.389 + tmp_live[i] |= t[i]; 1.390 + } 1.391 + 1.392 + // Now walk tmp_live up the block backwards, computing live 1.393 + for( int k=b->_nodes.size()-1; k>=0; k-- ) { 1.394 + Node *n = b->_nodes[k]; 1.395 + // KILL def'd bits 1.396 + int first = regalloc->get_reg_first(n); 1.397 + int second = regalloc->get_reg_second(n); 1.398 + if( OptoReg::is_valid(first) ) clr_live_bit(tmp_live,first); 1.399 + if( OptoReg::is_valid(second) ) clr_live_bit(tmp_live,second); 1.400 + 1.401 + MachNode *m = n->is_Mach() ? n->as_Mach() : NULL; 1.402 + 1.403 + // Check if m is potentially a CISC alternate instruction (i.e, possibly 1.404 + // synthesized by RegAlloc from a conventional instruction and a 1.405 + // spilled input) 1.406 + bool is_cisc_alternate = false; 1.407 + if (UseCISCSpill && m) { 1.408 + is_cisc_alternate = m->is_cisc_alternate(); 1.409 + } 1.410 + 1.411 + // GEN use'd bits 1.412 + for( uint l=1; l<n->req(); l++ ) { 1.413 + Node *def = n->in(l); 1.414 + assert(def != 0, "input edge required"); 1.415 + int first = regalloc->get_reg_first(def); 1.416 + int second = regalloc->get_reg_second(def); 1.417 + if( OptoReg::is_valid(first) ) set_live_bit(tmp_live,first); 1.418 + if( OptoReg::is_valid(second) ) set_live_bit(tmp_live,second); 1.419 + // If we use the stack pointer in a cisc-alternative instruction, 1.420 + // check for use as a memory operand. Then reconstruct the RegName 1.421 + // for this stack location, and set the appropriate bit in the 1.422 + // live vector 4987749. 1.423 + if (is_cisc_alternate && def == fp) { 1.424 + const TypePtr *adr_type = NULL; 1.425 + intptr_t offset; 1.426 + const Node* base = m->get_base_and_disp(offset, adr_type); 1.427 + if (base == NodeSentinel) { 1.428 + // Machnode has multiple memory inputs. We are unable to reason 1.429 + // with these, but are presuming (with trepidation) that not any of 1.430 + // them are oops. This can be fixed by making get_base_and_disp() 1.431 + // look at a specific input instead of all inputs. 1.432 + assert(!def->bottom_type()->isa_oop_ptr(), "expecting non-oop mem input"); 1.433 + } else if (base != fp || offset == Type::OffsetBot) { 1.434 + // Do nothing: the fp operand is either not from a memory use 1.435 + // (base == NULL) OR the fp is used in a non-memory context 1.436 + // (base is some other register) OR the offset is not constant, 1.437 + // so it is not a stack slot. 1.438 + } else { 1.439 + assert(offset >= 0, "unexpected negative offset"); 1.440 + offset -= (offset % jintSize); // count the whole word 1.441 + int stack_reg = regalloc->offset2reg(offset); 1.442 + if (OptoReg::is_stack(stack_reg)) { 1.443 + set_live_bit(tmp_live, stack_reg); 1.444 + } else { 1.445 + assert(false, "stack_reg not on stack?"); 1.446 + } 1.447 + } 1.448 + } 1.449 + } 1.450 + 1.451 + if( n->jvms() ) { // Record liveness at safepoint 1.452 + 1.453 + // This placement of this stanza means inputs to calls are 1.454 + // considered live at the callsite's OopMap. Argument oops are 1.455 + // hence live, but NOT included in the oopmap. See cutout in 1.456 + // build_oop_map. Debug oops are live (and in OopMap). 1.457 + int *n_live = NEW_ARENA_ARRAY(A, int, max_reg_ints); 1.458 + for( int l=0; l<max_reg_ints; l++ ) 1.459 + n_live[l] = tmp_live[l]; 1.460 + safehash->Insert(n,n_live); 1.461 + } 1.462 + 1.463 + } 1.464 + 1.465 + // Now at block top, see if we have any changes. If so, propagate 1.466 + // to prior blocks. 1.467 + int *old_live = &live[b->_pre_order*max_reg_ints]; 1.468 + int l; 1.469 + for( l=0; l<max_reg_ints; l++ ) 1.470 + if( tmp_live[l] != old_live[l] ) 1.471 + break; 1.472 + if( l<max_reg_ints ) { // Change! 1.473 + // Copy in new value 1.474 + for( l=0; l<max_reg_ints; l++ ) 1.475 + old_live[l] = tmp_live[l]; 1.476 + // Push preds onto worklist 1.477 + for( l=1; l<(int)b->num_preds(); l++ ) 1.478 + worklist->push(cfg->_bbs[b->pred(l)->_idx]); 1.479 + } 1.480 + } 1.481 + 1.482 + // Scan for any missing safepoints. Happens to infinite loops 1.483 + // ala ZKM.jar 1.484 + uint i; 1.485 + for( i=1; i<cfg->_num_blocks; i++ ) { 1.486 + Block *b = cfg->_blocks[i]; 1.487 + uint j; 1.488 + for( j=1; j<b->_nodes.size(); j++ ) 1.489 + if( b->_nodes[j]->jvms() && 1.490 + (*safehash)[b->_nodes[j]] == NULL ) 1.491 + break; 1.492 + if( j<b->_nodes.size() ) break; 1.493 + } 1.494 + if( i == cfg->_num_blocks ) 1.495 + break; // Got 'em all 1.496 +#ifndef PRODUCT 1.497 + if( PrintOpto && Verbose ) 1.498 + tty->print_cr("retripping live calc"); 1.499 +#endif 1.500 + // Force the issue (expensively): recheck everybody 1.501 + for( i=1; i<cfg->_num_blocks; i++ ) 1.502 + worklist->push(cfg->_blocks[i]); 1.503 + } 1.504 + 1.505 +} 1.506 + 1.507 +//------------------------------BuildOopMaps----------------------------------- 1.508 +// Collect GC mask info - where are all the OOPs? 1.509 +void Compile::BuildOopMaps() { 1.510 + NOT_PRODUCT( TracePhase t3("bldOopMaps", &_t_buildOopMaps, TimeCompiler); ) 1.511 + // Can't resource-mark because I need to leave all those OopMaps around, 1.512 + // or else I need to resource-mark some arena other than the default. 1.513 + // ResourceMark rm; // Reclaim all OopFlows when done 1.514 + int max_reg = _regalloc->_max_reg; // Current array extent 1.515 + 1.516 + Arena *A = Thread::current()->resource_area(); 1.517 + Block_List worklist; // Worklist of pending blocks 1.518 + 1.519 + int max_reg_ints = round_to(max_reg, BitsPerInt)>>LogBitsPerInt; 1.520 + Dict *safehash = NULL; // Used for assert only 1.521 + // Compute a backwards liveness per register. Needs a bitarray of 1.522 + // #blocks x (#registers, rounded up to ints) 1.523 + safehash = new Dict(cmpkey,hashkey,A); 1.524 + do_liveness( _regalloc, _cfg, &worklist, max_reg_ints, A, safehash ); 1.525 + OopFlow *free_list = NULL; // Free, unused 1.526 + 1.527 + // Array mapping blocks to completed oopflows 1.528 + OopFlow **flows = NEW_ARENA_ARRAY(A, OopFlow*, _cfg->_num_blocks); 1.529 + memset( flows, 0, _cfg->_num_blocks*sizeof(OopFlow*) ); 1.530 + 1.531 + 1.532 + // Do the first block 'by hand' to prime the worklist 1.533 + Block *entry = _cfg->_blocks[1]; 1.534 + OopFlow *rootflow = OopFlow::make(A,max_reg); 1.535 + // Initialize to 'bottom' (not 'top') 1.536 + memset( rootflow->_callees, OptoReg::Bad, max_reg*sizeof(short) ); 1.537 + memset( rootflow->_defs , 0, max_reg*sizeof(Node*) ); 1.538 + flows[entry->_pre_order] = rootflow; 1.539 + 1.540 + // Do the first block 'by hand' to prime the worklist 1.541 + rootflow->_b = entry; 1.542 + rootflow->compute_reach( _regalloc, max_reg, safehash ); 1.543 + for( uint i=0; i<entry->_num_succs; i++ ) 1.544 + worklist.push(entry->_succs[i]); 1.545 + 1.546 + // Now worklist contains blocks which have some, but perhaps not all, 1.547 + // predecessors visited. 1.548 + while( worklist.size() ) { 1.549 + // Scan for a block with all predecessors visited, or any randoms slob 1.550 + // otherwise. All-preds-visited order allows me to recycle OopFlow 1.551 + // structures rapidly and cut down on the memory footprint. 1.552 + // Note: not all predecessors might be visited yet (must happen for 1.553 + // irreducible loops). This is OK, since every live value must have the 1.554 + // SAME reaching def for the block, so any reaching def is OK. 1.555 + uint i; 1.556 + 1.557 + Block *b = worklist.pop(); 1.558 + // Ignore root block 1.559 + if( b == _cfg->_broot ) continue; 1.560 + // Block is already done? Happens if block has several predecessors, 1.561 + // he can get on the worklist more than once. 1.562 + if( flows[b->_pre_order] ) continue; 1.563 + 1.564 + // If this block has a visited predecessor AND that predecessor has this 1.565 + // last block as his only undone child, we can move the OopFlow from the 1.566 + // pred to this block. Otherwise we have to grab a new OopFlow. 1.567 + OopFlow *flow = NULL; // Flag for finding optimized flow 1.568 + Block *pred = (Block*)0xdeadbeef; 1.569 + uint j; 1.570 + // Scan this block's preds to find a done predecessor 1.571 + for( j=1; j<b->num_preds(); j++ ) { 1.572 + Block *p = _cfg->_bbs[b->pred(j)->_idx]; 1.573 + OopFlow *p_flow = flows[p->_pre_order]; 1.574 + if( p_flow ) { // Predecessor is done 1.575 + assert( p_flow->_b == p, "cross check" ); 1.576 + pred = p; // Record some predecessor 1.577 + // If all successors of p are done except for 'b', then we can carry 1.578 + // p_flow forward to 'b' without copying, otherwise we have to draw 1.579 + // from the free_list and clone data. 1.580 + uint k; 1.581 + for( k=0; k<p->_num_succs; k++ ) 1.582 + if( !flows[p->_succs[k]->_pre_order] && 1.583 + p->_succs[k] != b ) 1.584 + break; 1.585 + 1.586 + // Either carry-forward the now-unused OopFlow for b's use 1.587 + // or draw a new one from the free list 1.588 + if( k==p->_num_succs ) { 1.589 + flow = p_flow; 1.590 + break; // Found an ideal pred, use him 1.591 + } 1.592 + } 1.593 + } 1.594 + 1.595 + if( flow ) { 1.596 + // We have an OopFlow that's the last-use of a predecessor. 1.597 + // Carry it forward. 1.598 + } else { // Draw a new OopFlow from the freelist 1.599 + if( !free_list ) 1.600 + free_list = OopFlow::make(A,max_reg); 1.601 + flow = free_list; 1.602 + assert( flow->_b == NULL, "oopFlow is not free" ); 1.603 + free_list = flow->_next; 1.604 + flow->_next = NULL; 1.605 + 1.606 + // Copy/clone over the data 1.607 + flow->clone(flows[pred->_pre_order], max_reg); 1.608 + } 1.609 + 1.610 + // Mark flow for block. Blocks can only be flowed over once, 1.611 + // because after the first time they are guarded from entering 1.612 + // this code again. 1.613 + assert( flow->_b == pred, "have some prior flow" ); 1.614 + flow->_b = NULL; 1.615 + 1.616 + // Now push flow forward 1.617 + flows[b->_pre_order] = flow;// Mark flow for this block 1.618 + flow->_b = b; 1.619 + flow->compute_reach( _regalloc, max_reg, safehash ); 1.620 + 1.621 + // Now push children onto worklist 1.622 + for( i=0; i<b->_num_succs; i++ ) 1.623 + worklist.push(b->_succs[i]); 1.624 + 1.625 + } 1.626 +}