duke@435: /* trims@2708: * Copyright (c) 2002, 2011, Oracle and/or its affiliates. All rights reserved. duke@435: * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. duke@435: * duke@435: * This code is free software; you can redistribute it and/or modify it duke@435: * under the terms of the GNU General Public License version 2 only, as duke@435: * published by the Free Software Foundation. duke@435: * duke@435: * This code is distributed in the hope that it will be useful, but WITHOUT duke@435: * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or duke@435: * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License duke@435: * version 2 for more details (a copy is included in the LICENSE file that duke@435: * accompanied this code). duke@435: * duke@435: * You should have received a copy of the GNU General Public License version duke@435: * 2 along with this work; if not, write to the Free Software Foundation, duke@435: * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. duke@435: * trims@1907: * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA trims@1907: * or visit www.oracle.com if you need additional information or have any trims@1907: * questions. duke@435: * duke@435: */ duke@435: stefank@2314: #include "precompiled.hpp" stefank@2314: #include "compiler/oopMap.hpp" stefank@2314: #include "opto/addnode.hpp" stefank@2314: #include "opto/callnode.hpp" stefank@2314: #include "opto/compile.hpp" stefank@2314: #include "opto/machnode.hpp" stefank@2314: #include "opto/matcher.hpp" stefank@2314: #include "opto/phase.hpp" stefank@2314: #include "opto/regalloc.hpp" stefank@2314: #include "opto/rootnode.hpp" stefank@2314: #ifdef TARGET_ARCH_x86 stefank@2314: # include "vmreg_x86.inline.hpp" stefank@2314: #endif stefank@2314: #ifdef TARGET_ARCH_sparc stefank@2314: # include "vmreg_sparc.inline.hpp" stefank@2314: #endif stefank@2314: #ifdef TARGET_ARCH_zero stefank@2314: # include "vmreg_zero.inline.hpp" stefank@2314: #endif bobv@2508: #ifdef TARGET_ARCH_arm bobv@2508: # include "vmreg_arm.inline.hpp" bobv@2508: #endif bobv@2508: #ifdef TARGET_ARCH_ppc bobv@2508: # include "vmreg_ppc.inline.hpp" bobv@2508: #endif duke@435: duke@435: // The functions in this file builds OopMaps after all scheduling is done. duke@435: // duke@435: // OopMaps contain a list of all registers and stack-slots containing oops (so duke@435: // they can be updated by GC). OopMaps also contain a list of derived-pointer duke@435: // base-pointer pairs. When the base is moved, the derived pointer moves to duke@435: // follow it. Finally, any registers holding callee-save values are also duke@435: // recorded. These might contain oops, but only the caller knows. duke@435: // duke@435: // BuildOopMaps implements a simple forward reaching-defs solution. At each duke@435: // GC point we'll have the reaching-def Nodes. If the reaching Nodes are duke@435: // typed as pointers (no offset), then they are oops. Pointers+offsets are duke@435: // derived pointers, and bases can be found from them. Finally, we'll also duke@435: // track reaching callee-save values. Note that a copy of a callee-save value duke@435: // "kills" it's source, so that only 1 copy of a callee-save value is alive at duke@435: // a time. duke@435: // duke@435: // We run a simple bitvector liveness pass to help trim out dead oops. Due to duke@435: // irreducible loops, we can have a reaching def of an oop that only reaches duke@435: // along one path and no way to know if it's valid or not on the other path. duke@435: // The bitvectors are quite dense and the liveness pass is fast. duke@435: // duke@435: // At GC points, we consult this information to build OopMaps. All reaching duke@435: // defs typed as oops are added to the OopMap. Only 1 instance of a duke@435: // callee-save register can be recorded. For derived pointers, we'll have to duke@435: // find and record the register holding the base. duke@435: // duke@435: // The reaching def's is a simple 1-pass worklist approach. I tried a clever duke@435: // breadth-first approach but it was worse (showed O(n^2) in the duke@435: // pick-next-block code). duke@435: // twisti@1040: // The relevant data is kept in a struct of arrays (it could just as well be duke@435: // an array of structs, but the struct-of-arrays is generally a little more duke@435: // efficient). The arrays are indexed by register number (including duke@435: // stack-slots as registers) and so is bounded by 200 to 300 elements in duke@435: // practice. One array will map to a reaching def Node (or NULL for duke@435: // conflict/dead). The other array will map to a callee-saved register or duke@435: // OptoReg::Bad for not-callee-saved. duke@435: duke@435: duke@435: //------------------------------OopFlow---------------------------------------- duke@435: // Structure to pass around duke@435: struct OopFlow : public ResourceObj { duke@435: short *_callees; // Array mapping register to callee-saved duke@435: Node **_defs; // array mapping register to reaching def duke@435: // or NULL if dead/conflict duke@435: // OopFlow structs, when not being actively modified, describe the _end_ of duke@435: // this block. duke@435: Block *_b; // Block for this struct duke@435: OopFlow *_next; // Next free OopFlow kvn@1268: // or NULL if dead/conflict kvn@1268: Compile* C; duke@435: kvn@1268: OopFlow( short *callees, Node **defs, Compile* c ) : _callees(callees), _defs(defs), kvn@1268: _b(NULL), _next(NULL), C(c) { } duke@435: duke@435: // Given reaching-defs for this block start, compute it for this block end duke@435: void compute_reach( PhaseRegAlloc *regalloc, int max_reg, Dict *safehash ); duke@435: duke@435: // Merge these two OopFlows into the 'this' pointer. duke@435: void merge( OopFlow *flow, int max_reg ); duke@435: duke@435: // Copy a 'flow' over an existing flow duke@435: void clone( OopFlow *flow, int max_size); duke@435: duke@435: // Make a new OopFlow from scratch kvn@1268: static OopFlow *make( Arena *A, int max_size, Compile* C ); duke@435: duke@435: // Build an oopmap from the current flow info duke@435: OopMap *build_oop_map( Node *n, int max_reg, PhaseRegAlloc *regalloc, int* live ); duke@435: }; duke@435: duke@435: //------------------------------compute_reach---------------------------------- duke@435: // Given reaching-defs for this block start, compute it for this block end duke@435: void OopFlow::compute_reach( PhaseRegAlloc *regalloc, int max_reg, Dict *safehash ) { duke@435: duke@435: for( uint i=0; i<_b->_nodes.size(); i++ ) { duke@435: Node *n = _b->_nodes[i]; duke@435: duke@435: if( n->jvms() ) { // Build an OopMap here? duke@435: JVMState *jvms = n->jvms(); duke@435: // no map needed for leaf calls duke@435: if( n->is_MachSafePoint() && !n->is_MachCallLeaf() ) { duke@435: int *live = (int*) (*safehash)[n]; duke@435: assert( live, "must find live" ); duke@435: n->as_MachSafePoint()->set_oop_map( build_oop_map(n,max_reg,regalloc, live) ); duke@435: } duke@435: } duke@435: duke@435: // Assign new reaching def's. duke@435: // Note that I padded the _defs and _callees arrays so it's legal duke@435: // to index at _defs[OptoReg::Bad]. duke@435: OptoReg::Name first = regalloc->get_reg_first(n); duke@435: OptoReg::Name second = regalloc->get_reg_second(n); duke@435: _defs[first] = n; duke@435: _defs[second] = n; duke@435: duke@435: // Pass callee-save info around copies duke@435: int idx = n->is_Copy(); duke@435: if( idx ) { // Copies move callee-save info duke@435: OptoReg::Name old_first = regalloc->get_reg_first(n->in(idx)); duke@435: OptoReg::Name old_second = regalloc->get_reg_second(n->in(idx)); duke@435: int tmp_first = _callees[old_first]; duke@435: int tmp_second = _callees[old_second]; duke@435: _callees[old_first] = OptoReg::Bad; // callee-save is moved, dead in old location duke@435: _callees[old_second] = OptoReg::Bad; duke@435: _callees[first] = tmp_first; duke@435: _callees[second] = tmp_second; duke@435: } else if( n->is_Phi() ) { // Phis do not mod callee-saves duke@435: assert( _callees[first] == _callees[regalloc->get_reg_first(n->in(1))], "" ); duke@435: assert( _callees[second] == _callees[regalloc->get_reg_second(n->in(1))], "" ); duke@435: assert( _callees[first] == _callees[regalloc->get_reg_first(n->in(n->req()-1))], "" ); duke@435: assert( _callees[second] == _callees[regalloc->get_reg_second(n->in(n->req()-1))], "" ); duke@435: } else { duke@435: _callees[first] = OptoReg::Bad; // No longer holding a callee-save value duke@435: _callees[second] = OptoReg::Bad; duke@435: duke@435: // Find base case for callee saves duke@435: if( n->is_Proj() && n->in(0)->is_Start() ) { duke@435: if( OptoReg::is_reg(first) && duke@435: regalloc->_matcher.is_save_on_entry(first) ) duke@435: _callees[first] = first; duke@435: if( OptoReg::is_reg(second) && duke@435: regalloc->_matcher.is_save_on_entry(second) ) duke@435: _callees[second] = second; duke@435: } duke@435: } duke@435: } duke@435: } duke@435: duke@435: //------------------------------merge------------------------------------------ duke@435: // Merge the given flow into the 'this' flow duke@435: void OopFlow::merge( OopFlow *flow, int max_reg ) { duke@435: assert( _b == NULL, "merging into a happy flow" ); duke@435: assert( flow->_b, "this flow is still alive" ); duke@435: assert( flow != this, "no self flow" ); duke@435: duke@435: // Do the merge. If there are any differences, drop to 'bottom' which duke@435: // is OptoReg::Bad or NULL depending. duke@435: for( int i=0; i_callees[i] ) duke@435: _callees[i] = OptoReg::Bad; duke@435: // Merge the reaching defs duke@435: if( _defs[i] != flow->_defs[i] ) duke@435: _defs[i] = NULL; duke@435: } duke@435: duke@435: } duke@435: duke@435: //------------------------------clone------------------------------------------ duke@435: void OopFlow::clone( OopFlow *flow, int max_size ) { duke@435: _b = flow->_b; duke@435: memcpy( _callees, flow->_callees, sizeof(short)*max_size); duke@435: memcpy( _defs , flow->_defs , sizeof(Node*)*max_size); duke@435: } duke@435: duke@435: //------------------------------make------------------------------------------- kvn@1268: OopFlow *OopFlow::make( Arena *A, int max_size, Compile* C ) { duke@435: short *callees = NEW_ARENA_ARRAY(A,short,max_size+1); duke@435: Node **defs = NEW_ARENA_ARRAY(A,Node*,max_size+1); duke@435: debug_only( memset(defs,0,(max_size+1)*sizeof(Node*)) ); kvn@1268: OopFlow *flow = new (A) OopFlow(callees+1, defs+1, C); duke@435: assert( &flow->_callees[OptoReg::Bad] == callees, "Ok to index at OptoReg::Bad" ); duke@435: assert( &flow->_defs [OptoReg::Bad] == defs , "Ok to index at OptoReg::Bad" ); duke@435: return flow; duke@435: } duke@435: duke@435: //------------------------------bit twiddlers---------------------------------- duke@435: static int get_live_bit( int *live, int reg ) { duke@435: return live[reg>>LogBitsPerInt] & (1<<(reg&(BitsPerInt-1))); } duke@435: static void set_live_bit( int *live, int reg ) { duke@435: live[reg>>LogBitsPerInt] |= (1<<(reg&(BitsPerInt-1))); } duke@435: static void clr_live_bit( int *live, int reg ) { duke@435: live[reg>>LogBitsPerInt] &= ~(1<<(reg&(BitsPerInt-1))); } duke@435: duke@435: //------------------------------build_oop_map---------------------------------- duke@435: // Build an oopmap from the current flow info duke@435: OopMap *OopFlow::build_oop_map( Node *n, int max_reg, PhaseRegAlloc *regalloc, int* live ) { duke@435: int framesize = regalloc->_framesize; duke@435: int max_inarg_slot = OptoReg::reg2stack(regalloc->_matcher._new_SP); duke@435: debug_only( char *dup_check = NEW_RESOURCE_ARRAY(char,OptoReg::stack0()); duke@435: memset(dup_check,0,OptoReg::stack0()) ); duke@435: duke@435: OopMap *omap = new OopMap( framesize, max_inarg_slot ); duke@435: MachCallNode *mcall = n->is_MachCall() ? n->as_MachCall() : NULL; duke@435: JVMState* jvms = n->jvms(); duke@435: duke@435: // For all registers do... duke@435: for( int reg=0; regis_reg() && !r->is_concrete()) { duke@435: continue; duke@435: } duke@435: duke@435: // See if dead (no reaching def). duke@435: Node *def = _defs[reg]; // Get reaching def duke@435: assert( def, "since live better have reaching def" ); duke@435: duke@435: // Classify the reaching def as oop, derived, callee-save, dead, or other duke@435: const Type *t = def->bottom_type(); duke@435: if( t->isa_oop_ptr() ) { // Oop or derived? duke@435: assert( !OptoReg::is_valid(_callees[reg]), "oop can't be callee save" ); duke@435: #ifdef _LP64 duke@435: // 64-bit pointers record oop-ishness on 2 aligned adjacent registers. duke@435: // Make sure both are record from the same reaching def, but do not duke@435: // put both into the oopmap. duke@435: if( (reg&1) == 1 ) { // High half of oop-pair? duke@435: assert( _defs[reg-1] == _defs[reg], "both halves from same reaching def" ); duke@435: continue; // Do not record high parts in oopmap duke@435: } duke@435: #endif duke@435: duke@435: // Check for a legal reg name in the oopMap and bailout if it is not. duke@435: if (!omap->legal_vm_reg_name(r)) { duke@435: regalloc->C->record_method_not_compilable("illegal oopMap register name"); duke@435: continue; duke@435: } duke@435: if( t->is_ptr()->_offset == 0 ) { // Not derived? duke@435: if( mcall ) { duke@435: // Outgoing argument GC mask responsibility belongs to the callee, duke@435: // not the caller. Inspect the inputs to the call, to see if duke@435: // this live-range is one of them. duke@435: uint cnt = mcall->tf()->domain()->cnt(); duke@435: uint j; duke@435: for( j = TypeFunc::Parms; j < cnt; j++) duke@435: if( mcall->in(j) == def ) duke@435: break; // reaching def is an argument oop duke@435: if( j < cnt ) // arg oops dont go in GC map duke@435: continue; // Continue on to the next register duke@435: } duke@435: omap->set_oop(r); duke@435: } else { // Else it's derived. duke@435: // Find the base of the derived value. duke@435: uint i; duke@435: // Fast, common case, scan duke@435: for( i = jvms->oopoff(); i < n->req(); i+=2 ) duke@435: if( n->in(i) == def ) break; // Common case duke@435: if( i == n->req() ) { // Missed, try a more generous scan duke@435: // Scan again, but this time peek through copies duke@435: for( i = jvms->oopoff(); i < n->req(); i+=2 ) { duke@435: Node *m = n->in(i); // Get initial derived value duke@435: while( 1 ) { duke@435: Node *d = def; // Get initial reaching def duke@435: while( 1 ) { // Follow copies of reaching def to end duke@435: if( m == d ) goto found; // breaks 3 loops duke@435: int idx = d->is_Copy(); duke@435: if( !idx ) break; duke@435: d = d->in(idx); // Link through copy duke@435: } duke@435: int idx = m->is_Copy(); duke@435: if( !idx ) break; duke@435: m = m->in(idx); duke@435: } duke@435: } kvn@1268: guarantee( 0, "must find derived/base pair" ); duke@435: } duke@435: found: ; duke@435: Node *base = n->in(i+1); // Base is other half of pair duke@435: int breg = regalloc->get_reg_first(base); duke@435: VMReg b = OptoReg::as_VMReg(OptoReg::Name(breg), framesize, max_inarg_slot); duke@435: duke@435: // I record liveness at safepoints BEFORE I make the inputs duke@435: // live. This is because argument oops are NOT live at a duke@435: // safepoint (or at least they cannot appear in the oopmap). duke@435: // Thus bases of base/derived pairs might not be in the duke@435: // liveness data but they need to appear in the oopmap. duke@435: if( get_live_bit(live,breg) == 0 ) {// Not live? duke@435: // Flag it, so next derived pointer won't re-insert into oopmap duke@435: set_live_bit(live,breg); duke@435: // Already missed our turn? duke@435: if( breg < reg ) { duke@435: if (b->is_stack() || b->is_concrete() || true ) { duke@435: omap->set_oop( b); duke@435: } duke@435: } duke@435: } duke@435: if (b->is_stack() || b->is_concrete() || true ) { duke@435: omap->set_derived_oop( r, b); duke@435: } duke@435: } duke@435: coleenp@548: } else if( t->isa_narrowoop() ) { coleenp@548: assert( !OptoReg::is_valid(_callees[reg]), "oop can't be callee save" ); coleenp@548: // Check for a legal reg name in the oopMap and bailout if it is not. coleenp@548: if (!omap->legal_vm_reg_name(r)) { coleenp@548: regalloc->C->record_method_not_compilable("illegal oopMap register name"); coleenp@548: continue; coleenp@548: } coleenp@548: if( mcall ) { coleenp@548: // Outgoing argument GC mask responsibility belongs to the callee, coleenp@548: // not the caller. Inspect the inputs to the call, to see if coleenp@548: // this live-range is one of them. coleenp@548: uint cnt = mcall->tf()->domain()->cnt(); coleenp@548: uint j; coleenp@548: for( j = TypeFunc::Parms; j < cnt; j++) coleenp@548: if( mcall->in(j) == def ) coleenp@548: break; // reaching def is an argument oop coleenp@548: if( j < cnt ) // arg oops dont go in GC map coleenp@548: continue; // Continue on to the next register coleenp@548: } coleenp@548: omap->set_narrowoop(r); duke@435: } else if( OptoReg::is_valid(_callees[reg])) { // callee-save? duke@435: // It's a callee-save value duke@435: assert( dup_check[_callees[reg]]==0, "trying to callee save same reg twice" ); duke@435: debug_only( dup_check[_callees[reg]]=1; ) duke@435: VMReg callee = OptoReg::as_VMReg(OptoReg::Name(_callees[reg])); duke@435: if ( callee->is_concrete() || true ) { duke@435: omap->set_callee_saved( r, callee); duke@435: } duke@435: duke@435: } else { duke@435: // Other - some reaching non-oop value duke@435: omap->set_value( r); kvn@1268: #ifdef ASSERT kvn@1268: if( t->isa_rawptr() && C->cfg()->_raw_oops.member(def) ) { kvn@1268: def->dump(); kvn@1268: n->dump(); kvn@1268: assert(false, "there should be a oop in OopMap instead of a live raw oop at safepoint"); kvn@1268: } kvn@1268: #endif duke@435: } duke@435: duke@435: } duke@435: duke@435: #ifdef ASSERT duke@435: /* Nice, Intel-only assert duke@435: int cnt_callee_saves=0; duke@435: int reg2 = 0; duke@435: while (OptoReg::is_reg(reg2)) { duke@435: if( dup_check[reg2] != 0) cnt_callee_saves++; duke@435: assert( cnt_callee_saves==3 || cnt_callee_saves==5, "missed some callee-save" ); duke@435: reg2++; duke@435: } duke@435: */ duke@435: #endif duke@435: kvn@1164: #ifdef ASSERT kvn@1164: for( OopMapStream oms1(omap, OopMapValue::derived_oop_value); !oms1.is_done(); oms1.next()) { kvn@1164: OopMapValue omv1 = oms1.current(); kvn@1164: bool found = false; kvn@1164: for( OopMapStream oms2(omap,OopMapValue::oop_value); !oms2.is_done(); oms2.next()) { kvn@1164: if( omv1.content_reg() == oms2.current().reg() ) { kvn@1164: found = true; kvn@1164: break; kvn@1164: } kvn@1164: } kvn@1164: assert( found, "derived with no base in oopmap" ); kvn@1164: } kvn@1164: #endif kvn@1164: duke@435: return omap; duke@435: } duke@435: duke@435: //------------------------------do_liveness------------------------------------ duke@435: // Compute backwards liveness on registers duke@435: static void do_liveness( PhaseRegAlloc *regalloc, PhaseCFG *cfg, Block_List *worklist, int max_reg_ints, Arena *A, Dict *safehash ) { duke@435: int *live = NEW_ARENA_ARRAY(A, int, (cfg->_num_blocks+1) * max_reg_ints); duke@435: int *tmp_live = &live[cfg->_num_blocks * max_reg_ints]; duke@435: Node *root = cfg->C->root(); duke@435: // On CISC platforms, get the node representing the stack pointer that regalloc duke@435: // used for spills duke@435: Node *fp = NodeSentinel; duke@435: if (UseCISCSpill && root->req() > 1) { duke@435: fp = root->in(1)->in(TypeFunc::FramePtr); duke@435: } duke@435: memset( live, 0, cfg->_num_blocks * (max_reg_ints<req(); i++) { adlertz@5509: Block* block = cfg->get_block_for_node(root->in(i)); adlertz@5509: worklist->push(block); adlertz@5509: } duke@435: duke@435: // ZKM.jar includes tiny infinite loops which are unreached from below. duke@435: // If we missed any blocks, we'll retry here after pushing all missed duke@435: // blocks on the worklist. Normally this outer loop never trips more duke@435: // than once. adlertz@5509: while (1) { duke@435: duke@435: while( worklist->size() ) { // Standard worklist algorithm duke@435: Block *b = worklist->rpop(); duke@435: duke@435: // Copy first successor into my tmp_live space duke@435: int s0num = b->_succs[0]->_pre_order; duke@435: int *t = &live[s0num*max_reg_ints]; duke@435: for( int i=0; i_num_succs; j++ ) { duke@435: uint sjnum = b->_succs[j]->_pre_order; duke@435: int *t = &live[sjnum*max_reg_ints]; duke@435: for( int i=0; i_nodes.size()-1; k>=0; k-- ) { duke@435: Node *n = b->_nodes[k]; duke@435: // KILL def'd bits duke@435: int first = regalloc->get_reg_first(n); duke@435: int second = regalloc->get_reg_second(n); duke@435: if( OptoReg::is_valid(first) ) clr_live_bit(tmp_live,first); duke@435: if( OptoReg::is_valid(second) ) clr_live_bit(tmp_live,second); duke@435: duke@435: MachNode *m = n->is_Mach() ? n->as_Mach() : NULL; duke@435: duke@435: // Check if m is potentially a CISC alternate instruction (i.e, possibly duke@435: // synthesized by RegAlloc from a conventional instruction and a duke@435: // spilled input) duke@435: bool is_cisc_alternate = false; duke@435: if (UseCISCSpill && m) { duke@435: is_cisc_alternate = m->is_cisc_alternate(); duke@435: } duke@435: duke@435: // GEN use'd bits duke@435: for( uint l=1; lreq(); l++ ) { duke@435: Node *def = n->in(l); duke@435: assert(def != 0, "input edge required"); duke@435: int first = regalloc->get_reg_first(def); duke@435: int second = regalloc->get_reg_second(def); duke@435: if( OptoReg::is_valid(first) ) set_live_bit(tmp_live,first); duke@435: if( OptoReg::is_valid(second) ) set_live_bit(tmp_live,second); duke@435: // If we use the stack pointer in a cisc-alternative instruction, duke@435: // check for use as a memory operand. Then reconstruct the RegName duke@435: // for this stack location, and set the appropriate bit in the duke@435: // live vector 4987749. duke@435: if (is_cisc_alternate && def == fp) { duke@435: const TypePtr *adr_type = NULL; duke@435: intptr_t offset; duke@435: const Node* base = m->get_base_and_disp(offset, adr_type); duke@435: if (base == NodeSentinel) { duke@435: // Machnode has multiple memory inputs. We are unable to reason duke@435: // with these, but are presuming (with trepidation) that not any of duke@435: // them are oops. This can be fixed by making get_base_and_disp() duke@435: // look at a specific input instead of all inputs. duke@435: assert(!def->bottom_type()->isa_oop_ptr(), "expecting non-oop mem input"); duke@435: } else if (base != fp || offset == Type::OffsetBot) { duke@435: // Do nothing: the fp operand is either not from a memory use duke@435: // (base == NULL) OR the fp is used in a non-memory context duke@435: // (base is some other register) OR the offset is not constant, duke@435: // so it is not a stack slot. duke@435: } else { duke@435: assert(offset >= 0, "unexpected negative offset"); duke@435: offset -= (offset % jintSize); // count the whole word duke@435: int stack_reg = regalloc->offset2reg(offset); duke@435: if (OptoReg::is_stack(stack_reg)) { duke@435: set_live_bit(tmp_live, stack_reg); duke@435: } else { duke@435: assert(false, "stack_reg not on stack?"); duke@435: } duke@435: } duke@435: } duke@435: } duke@435: duke@435: if( n->jvms() ) { // Record liveness at safepoint duke@435: duke@435: // This placement of this stanza means inputs to calls are duke@435: // considered live at the callsite's OopMap. Argument oops are duke@435: // hence live, but NOT included in the oopmap. See cutout in duke@435: // build_oop_map. Debug oops are live (and in OopMap). duke@435: int *n_live = NEW_ARENA_ARRAY(A, int, max_reg_ints); duke@435: for( int l=0; lInsert(n,n_live); duke@435: } duke@435: duke@435: } duke@435: duke@435: // Now at block top, see if we have any changes. If so, propagate duke@435: // to prior blocks. duke@435: int *old_live = &live[b->_pre_order*max_reg_ints]; duke@435: int l; duke@435: for( l=0; lnum_preds(); l++) { adlertz@5509: Block* block = cfg->get_block_for_node(b->pred(l)); adlertz@5509: worklist->push(block); adlertz@5509: } duke@435: } duke@435: } duke@435: duke@435: // Scan for any missing safepoints. Happens to infinite loops duke@435: // ala ZKM.jar duke@435: uint i; duke@435: for( i=1; i_num_blocks; i++ ) { duke@435: Block *b = cfg->_blocks[i]; duke@435: uint j; duke@435: for( j=1; j_nodes.size(); j++ ) duke@435: if( b->_nodes[j]->jvms() && duke@435: (*safehash)[b->_nodes[j]] == NULL ) duke@435: break; duke@435: if( j_nodes.size() ) break; duke@435: } duke@435: if( i == cfg->_num_blocks ) duke@435: break; // Got 'em all duke@435: #ifndef PRODUCT duke@435: if( PrintOpto && Verbose ) duke@435: tty->print_cr("retripping live calc"); duke@435: #endif duke@435: // Force the issue (expensively): recheck everybody duke@435: for( i=1; i_num_blocks; i++ ) duke@435: worklist->push(cfg->_blocks[i]); duke@435: } duke@435: duke@435: } duke@435: duke@435: //------------------------------BuildOopMaps----------------------------------- duke@435: // Collect GC mask info - where are all the OOPs? duke@435: void Compile::BuildOopMaps() { duke@435: NOT_PRODUCT( TracePhase t3("bldOopMaps", &_t_buildOopMaps, TimeCompiler); ) duke@435: // Can't resource-mark because I need to leave all those OopMaps around, duke@435: // or else I need to resource-mark some arena other than the default. duke@435: // ResourceMark rm; // Reclaim all OopFlows when done duke@435: int max_reg = _regalloc->_max_reg; // Current array extent duke@435: duke@435: Arena *A = Thread::current()->resource_area(); duke@435: Block_List worklist; // Worklist of pending blocks duke@435: duke@435: int max_reg_ints = round_to(max_reg, BitsPerInt)>>LogBitsPerInt; duke@435: Dict *safehash = NULL; // Used for assert only duke@435: // Compute a backwards liveness per register. Needs a bitarray of duke@435: // #blocks x (#registers, rounded up to ints) duke@435: safehash = new Dict(cmpkey,hashkey,A); duke@435: do_liveness( _regalloc, _cfg, &worklist, max_reg_ints, A, safehash ); duke@435: OopFlow *free_list = NULL; // Free, unused duke@435: duke@435: // Array mapping blocks to completed oopflows duke@435: OopFlow **flows = NEW_ARENA_ARRAY(A, OopFlow*, _cfg->_num_blocks); duke@435: memset( flows, 0, _cfg->_num_blocks*sizeof(OopFlow*) ); duke@435: duke@435: duke@435: // Do the first block 'by hand' to prime the worklist duke@435: Block *entry = _cfg->_blocks[1]; kvn@1268: OopFlow *rootflow = OopFlow::make(A,max_reg,this); duke@435: // Initialize to 'bottom' (not 'top') duke@435: memset( rootflow->_callees, OptoReg::Bad, max_reg*sizeof(short) ); duke@435: memset( rootflow->_defs , 0, max_reg*sizeof(Node*) ); duke@435: flows[entry->_pre_order] = rootflow; duke@435: duke@435: // Do the first block 'by hand' to prime the worklist duke@435: rootflow->_b = entry; duke@435: rootflow->compute_reach( _regalloc, max_reg, safehash ); duke@435: for( uint i=0; i_num_succs; i++ ) duke@435: worklist.push(entry->_succs[i]); duke@435: duke@435: // Now worklist contains blocks which have some, but perhaps not all, duke@435: // predecessors visited. duke@435: while( worklist.size() ) { duke@435: // Scan for a block with all predecessors visited, or any randoms slob duke@435: // otherwise. All-preds-visited order allows me to recycle OopFlow duke@435: // structures rapidly and cut down on the memory footprint. duke@435: // Note: not all predecessors might be visited yet (must happen for duke@435: // irreducible loops). This is OK, since every live value must have the duke@435: // SAME reaching def for the block, so any reaching def is OK. duke@435: uint i; duke@435: duke@435: Block *b = worklist.pop(); duke@435: // Ignore root block duke@435: if( b == _cfg->_broot ) continue; duke@435: // Block is already done? Happens if block has several predecessors, duke@435: // he can get on the worklist more than once. duke@435: if( flows[b->_pre_order] ) continue; duke@435: duke@435: // If this block has a visited predecessor AND that predecessor has this duke@435: // last block as his only undone child, we can move the OopFlow from the duke@435: // pred to this block. Otherwise we have to grab a new OopFlow. duke@435: OopFlow *flow = NULL; // Flag for finding optimized flow duke@435: Block *pred = (Block*)0xdeadbeef; duke@435: // Scan this block's preds to find a done predecessor adlertz@5509: for (uint j = 1; j < b->num_preds(); j++) { adlertz@5509: Block* p = _cfg->get_block_for_node(b->pred(j)); duke@435: OopFlow *p_flow = flows[p->_pre_order]; duke@435: if( p_flow ) { // Predecessor is done duke@435: assert( p_flow->_b == p, "cross check" ); duke@435: pred = p; // Record some predecessor duke@435: // If all successors of p are done except for 'b', then we can carry duke@435: // p_flow forward to 'b' without copying, otherwise we have to draw duke@435: // from the free_list and clone data. duke@435: uint k; duke@435: for( k=0; k_num_succs; k++ ) duke@435: if( !flows[p->_succs[k]->_pre_order] && duke@435: p->_succs[k] != b ) duke@435: break; duke@435: duke@435: // Either carry-forward the now-unused OopFlow for b's use duke@435: // or draw a new one from the free list duke@435: if( k==p->_num_succs ) { duke@435: flow = p_flow; duke@435: break; // Found an ideal pred, use him duke@435: } duke@435: } duke@435: } duke@435: duke@435: if( flow ) { duke@435: // We have an OopFlow that's the last-use of a predecessor. duke@435: // Carry it forward. duke@435: } else { // Draw a new OopFlow from the freelist duke@435: if( !free_list ) kvn@1268: free_list = OopFlow::make(A,max_reg,C); duke@435: flow = free_list; duke@435: assert( flow->_b == NULL, "oopFlow is not free" ); duke@435: free_list = flow->_next; duke@435: flow->_next = NULL; duke@435: duke@435: // Copy/clone over the data duke@435: flow->clone(flows[pred->_pre_order], max_reg); duke@435: } duke@435: duke@435: // Mark flow for block. Blocks can only be flowed over once, duke@435: // because after the first time they are guarded from entering duke@435: // this code again. duke@435: assert( flow->_b == pred, "have some prior flow" ); duke@435: flow->_b = NULL; duke@435: duke@435: // Now push flow forward duke@435: flows[b->_pre_order] = flow;// Mark flow for this block duke@435: flow->_b = b; duke@435: flow->compute_reach( _regalloc, max_reg, safehash ); duke@435: duke@435: // Now push children onto worklist duke@435: for( i=0; i_num_succs; i++ ) duke@435: worklist.push(b->_succs[i]); duke@435: duke@435: } duke@435: }