duke@435: /* xdono@631: * Copyright 1998-2008 Sun Microsystems, Inc. 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: * duke@435: * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, duke@435: * CA 95054 USA or visit www.sun.com if you need additional information or duke@435: * have any questions. duke@435: * duke@435: */ duke@435: duke@435: #include "incls/_precompiled.incl" duke@435: #include "incls/_output.cpp.incl" duke@435: duke@435: extern uint size_java_to_interp(); duke@435: extern uint reloc_java_to_interp(); duke@435: extern uint size_exception_handler(); duke@435: extern uint size_deopt_handler(); duke@435: duke@435: #ifndef PRODUCT duke@435: #define DEBUG_ARG(x) , x duke@435: #else duke@435: #define DEBUG_ARG(x) duke@435: #endif duke@435: duke@435: extern int emit_exception_handler(CodeBuffer &cbuf); duke@435: extern int emit_deopt_handler(CodeBuffer &cbuf); duke@435: duke@435: //------------------------------Output----------------------------------------- duke@435: // Convert Nodes to instruction bits and pass off to the VM duke@435: void Compile::Output() { duke@435: // RootNode goes duke@435: assert( _cfg->_broot->_nodes.size() == 0, "" ); duke@435: duke@435: // Initialize the space for the BufferBlob used to find and verify duke@435: // instruction size in MachNode::emit_size() duke@435: init_scratch_buffer_blob(); kvn@598: if (failing()) return; // Out of memory duke@435: duke@435: // Make sure I can find the Start Node duke@435: Block_Array& bbs = _cfg->_bbs; duke@435: Block *entry = _cfg->_blocks[1]; duke@435: Block *broot = _cfg->_broot; duke@435: duke@435: const StartNode *start = entry->_nodes[0]->as_Start(); duke@435: duke@435: // Replace StartNode with prolog duke@435: MachPrologNode *prolog = new (this) MachPrologNode(); duke@435: entry->_nodes.map( 0, prolog ); duke@435: bbs.map( prolog->_idx, entry ); duke@435: bbs.map( start->_idx, NULL ); // start is no longer in any block duke@435: duke@435: // Virtual methods need an unverified entry point duke@435: duke@435: if( is_osr_compilation() ) { duke@435: if( PoisonOSREntry ) { duke@435: // TODO: Should use a ShouldNotReachHereNode... duke@435: _cfg->insert( broot, 0, new (this) MachBreakpointNode() ); duke@435: } duke@435: } else { duke@435: if( _method && !_method->flags().is_static() ) { duke@435: // Insert unvalidated entry point duke@435: _cfg->insert( broot, 0, new (this) MachUEPNode() ); duke@435: } duke@435: duke@435: } duke@435: duke@435: duke@435: // Break before main entry point duke@435: if( (_method && _method->break_at_execute()) duke@435: #ifndef PRODUCT duke@435: ||(OptoBreakpoint && is_method_compilation()) duke@435: ||(OptoBreakpointOSR && is_osr_compilation()) duke@435: ||(OptoBreakpointC2R && !_method) duke@435: #endif duke@435: ) { duke@435: // checking for _method means that OptoBreakpoint does not apply to duke@435: // runtime stubs or frame converters duke@435: _cfg->insert( entry, 1, new (this) MachBreakpointNode() ); duke@435: } duke@435: duke@435: // Insert epilogs before every return duke@435: for( uint i=0; i<_cfg->_num_blocks; i++ ) { duke@435: Block *b = _cfg->_blocks[i]; duke@435: if( !b->is_connector() && b->non_connector_successor(0) == _cfg->_broot ) { // Found a program exit point? duke@435: Node *m = b->end(); duke@435: if( m->is_Mach() && m->as_Mach()->ideal_Opcode() != Op_Halt ) { duke@435: MachEpilogNode *epilog = new (this) MachEpilogNode(m->as_Mach()->ideal_Opcode() == Op_Return); duke@435: b->add_inst( epilog ); duke@435: bbs.map(epilog->_idx, b); duke@435: //_regalloc->set_bad(epilog->_idx); // Already initialized this way. duke@435: } duke@435: } duke@435: } duke@435: duke@435: # ifdef ENABLE_ZAP_DEAD_LOCALS duke@435: if ( ZapDeadCompiledLocals ) Insert_zap_nodes(); duke@435: # endif duke@435: duke@435: ScheduleAndBundle(); duke@435: duke@435: #ifndef PRODUCT duke@435: if (trace_opto_output()) { duke@435: tty->print("\n---- After ScheduleAndBundle ----\n"); duke@435: for (uint i = 0; i < _cfg->_num_blocks; i++) { duke@435: tty->print("\nBB#%03d:\n", i); duke@435: Block *bb = _cfg->_blocks[i]; duke@435: for (uint j = 0; j < bb->_nodes.size(); j++) { duke@435: Node *n = bb->_nodes[j]; duke@435: OptoReg::Name reg = _regalloc->get_reg_first(n); duke@435: tty->print(" %-6s ", reg >= 0 && reg < REG_COUNT ? Matcher::regName[reg] : ""); duke@435: n->dump(); duke@435: } duke@435: } duke@435: } duke@435: #endif duke@435: duke@435: if (failing()) return; duke@435: duke@435: BuildOopMaps(); duke@435: duke@435: if (failing()) return; duke@435: duke@435: Fill_buffer(); duke@435: } duke@435: duke@435: bool Compile::need_stack_bang(int frame_size_in_bytes) const { duke@435: // Determine if we need to generate a stack overflow check. duke@435: // Do it if the method is not a stub function and duke@435: // has java calls or has frame size > vm_page_size/8. duke@435: return (stub_function() == NULL && duke@435: (has_java_calls() || frame_size_in_bytes > os::vm_page_size()>>3)); duke@435: } duke@435: duke@435: bool Compile::need_register_stack_bang() const { duke@435: // Determine if we need to generate a register stack overflow check. duke@435: // This is only used on architectures which have split register duke@435: // and memory stacks (ie. IA64). duke@435: // Bang if the method is not a stub function and has java calls duke@435: return (stub_function() == NULL && has_java_calls()); duke@435: } duke@435: duke@435: # ifdef ENABLE_ZAP_DEAD_LOCALS duke@435: duke@435: duke@435: // In order to catch compiler oop-map bugs, we have implemented duke@435: // a debugging mode called ZapDeadCompilerLocals. duke@435: // This mode causes the compiler to insert a call to a runtime routine, duke@435: // "zap_dead_locals", right before each place in compiled code duke@435: // that could potentially be a gc-point (i.e., a safepoint or oop map point). duke@435: // The runtime routine checks that locations mapped as oops are really duke@435: // oops, that locations mapped as values do not look like oops, duke@435: // and that locations mapped as dead are not used later duke@435: // (by zapping them to an invalid address). duke@435: duke@435: int Compile::_CompiledZap_count = 0; duke@435: duke@435: void Compile::Insert_zap_nodes() { duke@435: bool skip = false; duke@435: duke@435: duke@435: // Dink with static counts because code code without the extra duke@435: // runtime calls is MUCH faster for debugging purposes duke@435: duke@435: if ( CompileZapFirst == 0 ) ; // nothing special duke@435: else if ( CompileZapFirst > CompiledZap_count() ) skip = true; duke@435: else if ( CompileZapFirst == CompiledZap_count() ) duke@435: warning("starting zap compilation after skipping"); duke@435: duke@435: if ( CompileZapLast == -1 ) ; // nothing special duke@435: else if ( CompileZapLast < CompiledZap_count() ) skip = true; duke@435: else if ( CompileZapLast == CompiledZap_count() ) duke@435: warning("about to compile last zap"); duke@435: duke@435: ++_CompiledZap_count; // counts skipped zaps, too duke@435: duke@435: if ( skip ) return; duke@435: duke@435: duke@435: if ( _method == NULL ) duke@435: return; // no safepoints/oopmaps emitted for calls in stubs,so we don't care duke@435: duke@435: // Insert call to zap runtime stub before every node with an oop map duke@435: for( uint i=0; i<_cfg->_num_blocks; i++ ) { duke@435: Block *b = _cfg->_blocks[i]; duke@435: for ( uint j = 0; j < b->_nodes.size(); ++j ) { duke@435: Node *n = b->_nodes[j]; duke@435: duke@435: // Determining if we should insert a zap-a-lot node in output. duke@435: // We do that for all nodes that has oopmap info, except for calls duke@435: // to allocation. Calls to allocation passes in the old top-of-eden pointer duke@435: // and expect the C code to reset it. Hence, there can be no safepoints between duke@435: // the inlined-allocation and the call to new_Java, etc. duke@435: // We also cannot zap monitor calls, as they must hold the microlock duke@435: // during the call to Zap, which also wants to grab the microlock. duke@435: bool insert = n->is_MachSafePoint() && (n->as_MachSafePoint()->oop_map() != NULL); duke@435: if ( insert ) { // it is MachSafePoint duke@435: if ( !n->is_MachCall() ) { duke@435: insert = false; duke@435: } else if ( n->is_MachCall() ) { duke@435: MachCallNode* call = n->as_MachCall(); duke@435: if (call->entry_point() == OptoRuntime::new_instance_Java() || duke@435: call->entry_point() == OptoRuntime::new_array_Java() || duke@435: call->entry_point() == OptoRuntime::multianewarray2_Java() || duke@435: call->entry_point() == OptoRuntime::multianewarray3_Java() || duke@435: call->entry_point() == OptoRuntime::multianewarray4_Java() || duke@435: call->entry_point() == OptoRuntime::multianewarray5_Java() || duke@435: call->entry_point() == OptoRuntime::slow_arraycopy_Java() || duke@435: call->entry_point() == OptoRuntime::complete_monitor_locking_Java() duke@435: ) { duke@435: insert = false; duke@435: } duke@435: } duke@435: if (insert) { duke@435: Node *zap = call_zap_node(n->as_MachSafePoint(), i); duke@435: b->_nodes.insert( j, zap ); duke@435: _cfg->_bbs.map( zap->_idx, b ); duke@435: ++j; duke@435: } duke@435: } duke@435: } duke@435: } duke@435: } duke@435: duke@435: duke@435: Node* Compile::call_zap_node(MachSafePointNode* node_to_check, int block_no) { duke@435: const TypeFunc *tf = OptoRuntime::zap_dead_locals_Type(); duke@435: CallStaticJavaNode* ideal_node = duke@435: new (this, tf->domain()->cnt()) CallStaticJavaNode( tf, duke@435: OptoRuntime::zap_dead_locals_stub(_method->flags().is_native()), duke@435: "call zap dead locals stub", 0, TypePtr::BOTTOM); duke@435: // We need to copy the OopMap from the site we're zapping at. duke@435: // We have to make a copy, because the zap site might not be duke@435: // a call site, and zap_dead is a call site. duke@435: OopMap* clone = node_to_check->oop_map()->deep_copy(); duke@435: duke@435: // Add the cloned OopMap to the zap node duke@435: ideal_node->set_oop_map(clone); duke@435: return _matcher->match_sfpt(ideal_node); duke@435: } duke@435: duke@435: //------------------------------is_node_getting_a_safepoint-------------------- duke@435: bool Compile::is_node_getting_a_safepoint( Node* n) { duke@435: // This code duplicates the logic prior to the call of add_safepoint duke@435: // below in this file. duke@435: if( n->is_MachSafePoint() ) return true; duke@435: return false; duke@435: } duke@435: duke@435: # endif // ENABLE_ZAP_DEAD_LOCALS duke@435: duke@435: //------------------------------compute_loop_first_inst_sizes------------------ rasbold@853: // Compute the size of first NumberOfLoopInstrToAlign instructions at the top duke@435: // of a loop. When aligning a loop we need to provide enough instructions duke@435: // in cpu's fetch buffer to feed decoders. The loop alignment could be duke@435: // avoided if we have enough instructions in fetch buffer at the head of a loop. duke@435: // By default, the size is set to 999999 by Block's constructor so that duke@435: // a loop will be aligned if the size is not reset here. duke@435: // duke@435: // Note: Mach instructions could contain several HW instructions duke@435: // so the size is estimated only. duke@435: // duke@435: void Compile::compute_loop_first_inst_sizes() { duke@435: // The next condition is used to gate the loop alignment optimization. duke@435: // Don't aligned a loop if there are enough instructions at the head of a loop duke@435: // or alignment padding is larger then MaxLoopPad. By default, MaxLoopPad duke@435: // is equal to OptoLoopAlignment-1 except on new Intel cpus, where it is duke@435: // equal to 11 bytes which is the largest address NOP instruction. duke@435: if( MaxLoopPad < OptoLoopAlignment-1 ) { duke@435: uint last_block = _cfg->_num_blocks-1; duke@435: for( uint i=1; i <= last_block; i++ ) { duke@435: Block *b = _cfg->_blocks[i]; duke@435: // Check the first loop's block which requires an alignment. rasbold@853: if( b->loop_alignment() > (uint)relocInfo::addr_unit() ) { duke@435: uint sum_size = 0; duke@435: uint inst_cnt = NumberOfLoopInstrToAlign; rasbold@853: inst_cnt = b->compute_first_inst_size(sum_size, inst_cnt, _regalloc); rasbold@853: rasbold@853: // Check subsequent fallthrough blocks if the loop's first rasbold@853: // block(s) does not have enough instructions. rasbold@853: Block *nb = b; rasbold@853: while( inst_cnt > 0 && rasbold@853: i < last_block && rasbold@853: !_cfg->_blocks[i+1]->has_loop_alignment() && rasbold@853: !nb->has_successor(b) ) { rasbold@853: i++; rasbold@853: nb = _cfg->_blocks[i]; rasbold@853: inst_cnt = nb->compute_first_inst_size(sum_size, inst_cnt, _regalloc); rasbold@853: } // while( inst_cnt > 0 && i < last_block ) rasbold@853: duke@435: b->set_first_inst_size(sum_size); duke@435: } // f( b->head()->is_Loop() ) duke@435: } // for( i <= last_block ) duke@435: } // if( MaxLoopPad < OptoLoopAlignment-1 ) duke@435: } duke@435: duke@435: //----------------------Shorten_branches--------------------------------------- duke@435: // The architecture description provides short branch variants for some long duke@435: // branch instructions. Replace eligible long branches with short branches. duke@435: void Compile::Shorten_branches(Label *labels, int& code_size, int& reloc_size, int& stub_size, int& const_size) { duke@435: duke@435: // fill in the nop array for bundling computations duke@435: MachNode *_nop_list[Bundle::_nop_count]; duke@435: Bundle::initialize_nops(_nop_list, this); duke@435: duke@435: // ------------------ duke@435: // Compute size of each block, method size, and relocation information size duke@435: uint *jmp_end = NEW_RESOURCE_ARRAY(uint,_cfg->_num_blocks); duke@435: uint *blk_starts = NEW_RESOURCE_ARRAY(uint,_cfg->_num_blocks+1); duke@435: DEBUG_ONLY( uint *jmp_target = NEW_RESOURCE_ARRAY(uint,_cfg->_num_blocks); ) never@850: DEBUG_ONLY( uint *jmp_rule = NEW_RESOURCE_ARRAY(uint,_cfg->_num_blocks); ) duke@435: blk_starts[0] = 0; duke@435: duke@435: // Initialize the sizes to 0 duke@435: code_size = 0; // Size in bytes of generated code duke@435: stub_size = 0; // Size in bytes of all stub entries duke@435: // Size in bytes of all relocation entries, including those in local stubs. duke@435: // Start with 2-bytes of reloc info for the unvalidated entry point duke@435: reloc_size = 1; // Number of relocation entries duke@435: const_size = 0; // size of fp constants in words duke@435: duke@435: // Make three passes. The first computes pessimistic blk_starts, duke@435: // relative jmp_end, reloc_size and const_size information. duke@435: // The second performs short branch substitution using the pessimistic duke@435: // sizing. The third inserts nops where needed. duke@435: duke@435: Node *nj; // tmp duke@435: duke@435: // Step one, perform a pessimistic sizing pass. duke@435: uint i; duke@435: uint min_offset_from_last_call = 1; // init to a positive value duke@435: uint nop_size = (new (this) MachNopNode())->size(_regalloc); duke@435: for( i=0; i<_cfg->_num_blocks; i++ ) { // For all blocks duke@435: Block *b = _cfg->_blocks[i]; duke@435: duke@435: // Sum all instruction sizes to compute block size duke@435: uint last_inst = b->_nodes.size(); duke@435: uint blk_size = 0; duke@435: for( uint j = 0; j_nodes[j]; duke@435: uint inst_size = nj->size(_regalloc); duke@435: blk_size += inst_size; duke@435: // Handle machine instruction nodes duke@435: if( nj->is_Mach() ) { duke@435: MachNode *mach = nj->as_Mach(); duke@435: blk_size += (mach->alignment_required() - 1) * relocInfo::addr_unit(); // assume worst case padding duke@435: reloc_size += mach->reloc(); duke@435: const_size += mach->const_size(); duke@435: if( mach->is_MachCall() ) { duke@435: MachCallNode *mcall = mach->as_MachCall(); duke@435: // This destination address is NOT PC-relative duke@435: duke@435: mcall->method_set((intptr_t)mcall->entry_point()); duke@435: duke@435: if( mcall->is_MachCallJava() && mcall->as_MachCallJava()->_method ) { duke@435: stub_size += size_java_to_interp(); duke@435: reloc_size += reloc_java_to_interp(); duke@435: } duke@435: } else if (mach->is_MachSafePoint()) { duke@435: // If call/safepoint are adjacent, account for possible duke@435: // nop to disambiguate the two safepoints. duke@435: if (min_offset_from_last_call == 0) { duke@435: blk_size += nop_size; duke@435: } duke@435: } duke@435: } duke@435: min_offset_from_last_call += inst_size; duke@435: // Remember end of call offset duke@435: if (nj->is_MachCall() && nj->as_MachCall()->is_safepoint_node()) { duke@435: min_offset_from_last_call = 0; duke@435: } duke@435: } duke@435: duke@435: // During short branch replacement, we store the relative (to blk_starts) duke@435: // end of jump in jmp_end, rather than the absolute end of jump. This duke@435: // is so that we do not need to recompute sizes of all nodes when we compute duke@435: // correct blk_starts in our next sizing pass. duke@435: jmp_end[i] = blk_size; duke@435: DEBUG_ONLY( jmp_target[i] = 0; ) duke@435: duke@435: // When the next block starts a loop, we may insert pad NOP duke@435: // instructions. Since we cannot know our future alignment, duke@435: // assume the worst. duke@435: if( i<_cfg->_num_blocks-1 ) { duke@435: Block *nb = _cfg->_blocks[i+1]; duke@435: int max_loop_pad = nb->code_alignment()-relocInfo::addr_unit(); duke@435: if( max_loop_pad > 0 ) { duke@435: assert(is_power_of_2(max_loop_pad+relocInfo::addr_unit()), ""); duke@435: blk_size += max_loop_pad; duke@435: } duke@435: } duke@435: duke@435: // Save block size; update total method size duke@435: blk_starts[i+1] = blk_starts[i]+blk_size; duke@435: } duke@435: duke@435: // Step two, replace eligible long jumps. duke@435: duke@435: // Note: this will only get the long branches within short branch duke@435: // range. Another pass might detect more branches that became duke@435: // candidates because the shortening in the first pass exposed duke@435: // more opportunities. Unfortunately, this would require duke@435: // recomputing the starting and ending positions for the blocks duke@435: for( i=0; i<_cfg->_num_blocks; i++ ) { duke@435: Block *b = _cfg->_blocks[i]; duke@435: duke@435: int j; duke@435: // Find the branch; ignore trailing NOPs. duke@435: for( j = b->_nodes.size()-1; j>=0; j-- ) { duke@435: nj = b->_nodes[j]; duke@435: if( !nj->is_Mach() || nj->as_Mach()->ideal_Opcode() != Op_Con ) duke@435: break; duke@435: } duke@435: duke@435: if (j >= 0) { duke@435: if( nj->is_Mach() && nj->as_Mach()->may_be_short_branch() ) { duke@435: MachNode *mach = nj->as_Mach(); duke@435: // This requires the TRUE branch target be in succs[0] duke@435: uint bnum = b->non_connector_successor(0)->_pre_order; duke@435: uintptr_t target = blk_starts[bnum]; duke@435: if( mach->is_pc_relative() ) { duke@435: int offset = target-(blk_starts[i] + jmp_end[i]); never@850: if (_matcher->is_short_branch_offset(mach->rule(), offset)) { duke@435: // We've got a winner. Replace this branch. never@850: MachNode* replacement = mach->short_branch_version(this); duke@435: b->_nodes.map(j, replacement); never@657: mach->subsume_by(replacement); duke@435: duke@435: // Update the jmp_end size to save time in our duke@435: // next pass. duke@435: jmp_end[i] -= (mach->size(_regalloc) - replacement->size(_regalloc)); duke@435: DEBUG_ONLY( jmp_target[i] = bnum; ); never@850: DEBUG_ONLY( jmp_rule[i] = mach->rule(); ); duke@435: } duke@435: } else { duke@435: #ifndef PRODUCT duke@435: mach->dump(3); duke@435: #endif duke@435: Unimplemented(); duke@435: } duke@435: } duke@435: } duke@435: } duke@435: duke@435: // Compute the size of first NumberOfLoopInstrToAlign instructions at head duke@435: // of a loop. It is used to determine the padding for loop alignment. duke@435: compute_loop_first_inst_sizes(); duke@435: duke@435: // Step 3, compute the offsets of all the labels duke@435: uint last_call_adr = max_uint; duke@435: for( i=0; i<_cfg->_num_blocks; i++ ) { // For all blocks duke@435: // copy the offset of the beginning to the corresponding label duke@435: assert(labels[i].is_unused(), "cannot patch at this point"); duke@435: labels[i].bind_loc(blk_starts[i], CodeBuffer::SECT_INSTS); duke@435: duke@435: // insert padding for any instructions that need it duke@435: Block *b = _cfg->_blocks[i]; duke@435: uint last_inst = b->_nodes.size(); duke@435: uint adr = blk_starts[i]; duke@435: for( uint j = 0; j_nodes[j]; duke@435: if( nj->is_Mach() ) { duke@435: int padding = nj->as_Mach()->compute_padding(adr); duke@435: // If call/safepoint are adjacent insert a nop (5010568) duke@435: if (padding == 0 && nj->is_MachSafePoint() && !nj->is_MachCall() && duke@435: adr == last_call_adr ) { duke@435: padding = nop_size; duke@435: } duke@435: if(padding > 0) { duke@435: assert((padding % nop_size) == 0, "padding is not a multiple of NOP size"); duke@435: int nops_cnt = padding / nop_size; duke@435: MachNode *nop = new (this) MachNopNode(nops_cnt); duke@435: b->_nodes.insert(j++, nop); duke@435: _cfg->_bbs.map( nop->_idx, b ); duke@435: adr += padding; duke@435: last_inst++; duke@435: } duke@435: } duke@435: adr += nj->size(_regalloc); duke@435: duke@435: // Remember end of call offset duke@435: if (nj->is_MachCall() && nj->as_MachCall()->is_safepoint_node()) { duke@435: last_call_adr = adr; duke@435: } duke@435: } duke@435: duke@435: if ( i != _cfg->_num_blocks-1) { duke@435: // Get the size of the block duke@435: uint blk_size = adr - blk_starts[i]; duke@435: rasbold@853: // When the next block is the top of a loop, we may insert pad NOP duke@435: // instructions. duke@435: Block *nb = _cfg->_blocks[i+1]; duke@435: int current_offset = blk_starts[i] + blk_size; duke@435: current_offset += nb->alignment_padding(current_offset); duke@435: // Save block size; update total method size duke@435: blk_starts[i+1] = current_offset; duke@435: } duke@435: } duke@435: duke@435: #ifdef ASSERT duke@435: for( i=0; i<_cfg->_num_blocks; i++ ) { // For all blocks duke@435: if( jmp_target[i] != 0 ) { duke@435: int offset = blk_starts[jmp_target[i]]-(blk_starts[i] + jmp_end[i]); never@850: if (!_matcher->is_short_branch_offset(jmp_rule[i], offset)) { duke@435: tty->print_cr("target (%d) - jmp_end(%d) = offset (%d), jmp_block B%d, target_block B%d", blk_starts[jmp_target[i]], blk_starts[i] + jmp_end[i], offset, i, jmp_target[i]); duke@435: } never@850: assert(_matcher->is_short_branch_offset(jmp_rule[i], offset), "Displacement too large for short jmp"); duke@435: } duke@435: } duke@435: #endif duke@435: duke@435: // ------------------ duke@435: // Compute size for code buffer duke@435: code_size = blk_starts[i-1] + jmp_end[i-1]; duke@435: duke@435: // Relocation records duke@435: reloc_size += 1; // Relo entry for exception handler duke@435: duke@435: // Adjust reloc_size to number of record of relocation info duke@435: // Min is 2 bytes, max is probably 6 or 8, with a tax up to 25% for duke@435: // a relocation index. duke@435: // The CodeBuffer will expand the locs array if this estimate is too low. duke@435: reloc_size *= 10 / sizeof(relocInfo); duke@435: duke@435: // Adjust const_size to number of bytes duke@435: const_size *= 2*jintSize; // both float and double take two words per entry duke@435: duke@435: } duke@435: duke@435: //------------------------------FillLocArray----------------------------------- duke@435: // Create a bit of debug info and append it to the array. The mapping is from duke@435: // Java local or expression stack to constant, register or stack-slot. For duke@435: // doubles, insert 2 mappings and return 1 (to tell the caller that the next duke@435: // entry has been taken care of and caller should skip it). duke@435: static LocationValue *new_loc_value( PhaseRegAlloc *ra, OptoReg::Name regnum, Location::Type l_type ) { duke@435: // This should never have accepted Bad before duke@435: assert(OptoReg::is_valid(regnum), "location must be valid"); duke@435: return (OptoReg::is_reg(regnum)) duke@435: ? new LocationValue(Location::new_reg_loc(l_type, OptoReg::as_VMReg(regnum)) ) duke@435: : new LocationValue(Location::new_stk_loc(l_type, ra->reg2offset(regnum))); duke@435: } duke@435: kvn@498: kvn@498: ObjectValue* kvn@498: Compile::sv_for_node_id(GrowableArray *objs, int id) { kvn@498: for (int i = 0; i < objs->length(); i++) { kvn@498: assert(objs->at(i)->is_object(), "corrupt object cache"); kvn@498: ObjectValue* sv = (ObjectValue*) objs->at(i); kvn@498: if (sv->id() == id) { kvn@498: return sv; kvn@498: } kvn@498: } kvn@498: // Otherwise.. kvn@498: return NULL; kvn@498: } kvn@498: kvn@498: void Compile::set_sv_for_object_node(GrowableArray *objs, kvn@498: ObjectValue* sv ) { kvn@498: assert(sv_for_node_id(objs, sv->id()) == NULL, "Precondition"); kvn@498: objs->append(sv); kvn@498: } kvn@498: kvn@498: kvn@498: void Compile::FillLocArray( int idx, MachSafePointNode* sfpt, Node *local, kvn@498: GrowableArray *array, kvn@498: GrowableArray *objs ) { duke@435: assert( local, "use _top instead of null" ); duke@435: if (array->length() != idx) { duke@435: assert(array->length() == idx + 1, "Unexpected array count"); duke@435: // Old functionality: duke@435: // return duke@435: // New functionality: duke@435: // Assert if the local is not top. In product mode let the new node duke@435: // override the old entry. duke@435: assert(local == top(), "LocArray collision"); duke@435: if (local == top()) { duke@435: return; duke@435: } duke@435: array->pop(); duke@435: } duke@435: const Type *t = local->bottom_type(); duke@435: kvn@498: // Is it a safepoint scalar object node? kvn@498: if (local->is_SafePointScalarObject()) { kvn@498: SafePointScalarObjectNode* spobj = local->as_SafePointScalarObject(); kvn@498: kvn@498: ObjectValue* sv = Compile::sv_for_node_id(objs, spobj->_idx); kvn@498: if (sv == NULL) { kvn@498: ciKlass* cik = t->is_oopptr()->klass(); kvn@498: assert(cik->is_instance_klass() || kvn@498: cik->is_array_klass(), "Not supported allocation."); kvn@498: sv = new ObjectValue(spobj->_idx, kvn@498: new ConstantOopWriteValue(cik->encoding())); kvn@498: Compile::set_sv_for_object_node(objs, sv); kvn@498: kvn@498: uint first_ind = spobj->first_index(); kvn@498: for (uint i = 0; i < spobj->n_fields(); i++) { kvn@498: Node* fld_node = sfpt->in(first_ind+i); kvn@498: (void)FillLocArray(sv->field_values()->length(), sfpt, fld_node, sv->field_values(), objs); kvn@498: } kvn@498: } kvn@498: array->append(sv); kvn@498: return; kvn@498: } kvn@498: duke@435: // Grab the register number for the local duke@435: OptoReg::Name regnum = _regalloc->get_reg_first(local); duke@435: if( OptoReg::is_valid(regnum) ) {// Got a register/stack? duke@435: // Record the double as two float registers. duke@435: // The register mask for such a value always specifies two adjacent duke@435: // float registers, with the lower register number even. duke@435: // Normally, the allocation of high and low words to these registers duke@435: // is irrelevant, because nearly all operations on register pairs duke@435: // (e.g., StoreD) treat them as a single unit. duke@435: // Here, we assume in addition that the words in these two registers duke@435: // stored "naturally" (by operations like StoreD and double stores duke@435: // within the interpreter) such that the lower-numbered register duke@435: // is written to the lower memory address. This may seem like duke@435: // a machine dependency, but it is not--it is a requirement on duke@435: // the author of the .ad file to ensure that, for every duke@435: // even/odd double-register pair to which a double may be allocated, duke@435: // the word in the even single-register is stored to the first duke@435: // memory word. (Note that register numbers are completely duke@435: // arbitrary, and are not tied to any machine-level encodings.) duke@435: #ifdef _LP64 duke@435: if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon ) { duke@435: array->append(new ConstantIntValue(0)); duke@435: array->append(new_loc_value( _regalloc, regnum, Location::dbl )); duke@435: } else if ( t->base() == Type::Long ) { duke@435: array->append(new ConstantIntValue(0)); duke@435: array->append(new_loc_value( _regalloc, regnum, Location::lng )); duke@435: } else if ( t->base() == Type::RawPtr ) { duke@435: // jsr/ret return address which must be restored into a the full duke@435: // width 64-bit stack slot. duke@435: array->append(new_loc_value( _regalloc, regnum, Location::lng )); duke@435: } duke@435: #else //_LP64 duke@435: #ifdef SPARC duke@435: if (t->base() == Type::Long && OptoReg::is_reg(regnum)) { duke@435: // For SPARC we have to swap high and low words for duke@435: // long values stored in a single-register (g0-g7). duke@435: array->append(new_loc_value( _regalloc, regnum , Location::normal )); duke@435: array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal )); duke@435: } else duke@435: #endif //SPARC duke@435: if( t->base() == Type::DoubleBot || t->base() == Type::DoubleCon || t->base() == Type::Long ) { duke@435: // Repack the double/long as two jints. duke@435: // The convention the interpreter uses is that the second local duke@435: // holds the first raw word of the native double representation. duke@435: // This is actually reasonable, since locals and stack arrays duke@435: // grow downwards in all implementations. duke@435: // (If, on some machine, the interpreter's Java locals or stack duke@435: // were to grow upwards, the embedded doubles would be word-swapped.) duke@435: array->append(new_loc_value( _regalloc, OptoReg::add(regnum,1), Location::normal )); duke@435: array->append(new_loc_value( _regalloc, regnum , Location::normal )); duke@435: } duke@435: #endif //_LP64 duke@435: else if( (t->base() == Type::FloatBot || t->base() == Type::FloatCon) && duke@435: OptoReg::is_reg(regnum) ) { duke@435: array->append(new_loc_value( _regalloc, regnum, Matcher::float_in_double duke@435: ? Location::float_in_dbl : Location::normal )); duke@435: } else if( t->base() == Type::Int && OptoReg::is_reg(regnum) ) { duke@435: array->append(new_loc_value( _regalloc, regnum, Matcher::int_in_long duke@435: ? Location::int_in_long : Location::normal )); kvn@766: } else if( t->base() == Type::NarrowOop ) { kvn@766: array->append(new_loc_value( _regalloc, regnum, Location::narrowoop )); duke@435: } else { duke@435: array->append(new_loc_value( _regalloc, regnum, _regalloc->is_oop(local) ? Location::oop : Location::normal )); duke@435: } duke@435: return; duke@435: } duke@435: duke@435: // No register. It must be constant data. duke@435: switch (t->base()) { duke@435: case Type::Half: // Second half of a double duke@435: ShouldNotReachHere(); // Caller should skip 2nd halves duke@435: break; duke@435: case Type::AnyPtr: duke@435: array->append(new ConstantOopWriteValue(NULL)); duke@435: break; duke@435: case Type::AryPtr: duke@435: case Type::InstPtr: duke@435: case Type::KlassPtr: // fall through duke@435: array->append(new ConstantOopWriteValue(t->isa_oopptr()->const_oop()->encoding())); duke@435: break; kvn@766: case Type::NarrowOop: kvn@766: if (t == TypeNarrowOop::NULL_PTR) { kvn@766: array->append(new ConstantOopWriteValue(NULL)); kvn@766: } else { kvn@766: array->append(new ConstantOopWriteValue(t->make_ptr()->isa_oopptr()->const_oop()->encoding())); kvn@766: } kvn@766: break; duke@435: case Type::Int: duke@435: array->append(new ConstantIntValue(t->is_int()->get_con())); duke@435: break; duke@435: case Type::RawPtr: duke@435: // A return address (T_ADDRESS). duke@435: assert((intptr_t)t->is_ptr()->get_con() < (intptr_t)0x10000, "must be a valid BCI"); duke@435: #ifdef _LP64 duke@435: // Must be restored to the full-width 64-bit stack slot. duke@435: array->append(new ConstantLongValue(t->is_ptr()->get_con())); duke@435: #else duke@435: array->append(new ConstantIntValue(t->is_ptr()->get_con())); duke@435: #endif duke@435: break; duke@435: case Type::FloatCon: { duke@435: float f = t->is_float_constant()->getf(); duke@435: array->append(new ConstantIntValue(jint_cast(f))); duke@435: break; duke@435: } duke@435: case Type::DoubleCon: { duke@435: jdouble d = t->is_double_constant()->getd(); duke@435: #ifdef _LP64 duke@435: array->append(new ConstantIntValue(0)); duke@435: array->append(new ConstantDoubleValue(d)); duke@435: #else duke@435: // Repack the double as two jints. duke@435: // The convention the interpreter uses is that the second local duke@435: // holds the first raw word of the native double representation. duke@435: // This is actually reasonable, since locals and stack arrays duke@435: // grow downwards in all implementations. duke@435: // (If, on some machine, the interpreter's Java locals or stack duke@435: // were to grow upwards, the embedded doubles would be word-swapped.) duke@435: jint *dp = (jint*)&d; duke@435: array->append(new ConstantIntValue(dp[1])); duke@435: array->append(new ConstantIntValue(dp[0])); duke@435: #endif duke@435: break; duke@435: } duke@435: case Type::Long: { duke@435: jlong d = t->is_long()->get_con(); duke@435: #ifdef _LP64 duke@435: array->append(new ConstantIntValue(0)); duke@435: array->append(new ConstantLongValue(d)); duke@435: #else duke@435: // Repack the long as two jints. duke@435: // The convention the interpreter uses is that the second local duke@435: // holds the first raw word of the native double representation. duke@435: // This is actually reasonable, since locals and stack arrays duke@435: // grow downwards in all implementations. duke@435: // (If, on some machine, the interpreter's Java locals or stack duke@435: // were to grow upwards, the embedded doubles would be word-swapped.) duke@435: jint *dp = (jint*)&d; duke@435: array->append(new ConstantIntValue(dp[1])); duke@435: array->append(new ConstantIntValue(dp[0])); duke@435: #endif duke@435: break; duke@435: } duke@435: case Type::Top: // Add an illegal value here duke@435: array->append(new LocationValue(Location())); duke@435: break; duke@435: default: duke@435: ShouldNotReachHere(); duke@435: break; duke@435: } duke@435: } duke@435: duke@435: // Determine if this node starts a bundle duke@435: bool Compile::starts_bundle(const Node *n) const { duke@435: return (_node_bundling_limit > n->_idx && duke@435: _node_bundling_base[n->_idx].starts_bundle()); duke@435: } duke@435: duke@435: //--------------------------Process_OopMap_Node-------------------------------- duke@435: void Compile::Process_OopMap_Node(MachNode *mach, int current_offset) { duke@435: duke@435: // Handle special safepoint nodes for synchronization duke@435: MachSafePointNode *sfn = mach->as_MachSafePoint(); duke@435: MachCallNode *mcall; duke@435: duke@435: #ifdef ENABLE_ZAP_DEAD_LOCALS duke@435: assert( is_node_getting_a_safepoint(mach), "logic does not match; false negative"); duke@435: #endif duke@435: duke@435: int safepoint_pc_offset = current_offset; duke@435: duke@435: // Add the safepoint in the DebugInfoRecorder duke@435: if( !mach->is_MachCall() ) { duke@435: mcall = NULL; duke@435: debug_info()->add_safepoint(safepoint_pc_offset, sfn->_oop_map); duke@435: } else { duke@435: mcall = mach->as_MachCall(); duke@435: safepoint_pc_offset += mcall->ret_addr_offset(); duke@435: debug_info()->add_safepoint(safepoint_pc_offset, mcall->_oop_map); duke@435: } duke@435: duke@435: // Loop over the JVMState list to add scope information duke@435: // Do not skip safepoints with a NULL method, they need monitor info duke@435: JVMState* youngest_jvms = sfn->jvms(); duke@435: int max_depth = youngest_jvms->depth(); duke@435: kvn@498: // Allocate the object pool for scalar-replaced objects -- the map from kvn@498: // small-integer keys (which can be recorded in the local and ostack kvn@498: // arrays) to descriptions of the object state. kvn@498: GrowableArray *objs = new GrowableArray(); kvn@498: duke@435: // Visit scopes from oldest to youngest. duke@435: for (int depth = 1; depth <= max_depth; depth++) { duke@435: JVMState* jvms = youngest_jvms->of_depth(depth); duke@435: int idx; duke@435: ciMethod* method = jvms->has_method() ? jvms->method() : NULL; duke@435: // Safepoints that do not have method() set only provide oop-map and monitor info duke@435: // to support GC; these do not support deoptimization. duke@435: int num_locs = (method == NULL) ? 0 : jvms->loc_size(); duke@435: int num_exps = (method == NULL) ? 0 : jvms->stk_size(); duke@435: int num_mon = jvms->nof_monitors(); duke@435: assert(method == NULL || jvms->bci() < 0 || num_locs == method->max_locals(), duke@435: "JVMS local count must match that of the method"); duke@435: duke@435: // Add Local and Expression Stack Information duke@435: duke@435: // Insert locals into the locarray duke@435: GrowableArray *locarray = new GrowableArray(num_locs); duke@435: for( idx = 0; idx < num_locs; idx++ ) { kvn@498: FillLocArray( idx, sfn, sfn->local(jvms, idx), locarray, objs ); duke@435: } duke@435: duke@435: // Insert expression stack entries into the exparray duke@435: GrowableArray *exparray = new GrowableArray(num_exps); duke@435: for( idx = 0; idx < num_exps; idx++ ) { kvn@498: FillLocArray( idx, sfn, sfn->stack(jvms, idx), exparray, objs ); duke@435: } duke@435: duke@435: // Add in mappings of the monitors duke@435: assert( !method || duke@435: !method->is_synchronized() || duke@435: method->is_native() || duke@435: num_mon > 0 || duke@435: !GenerateSynchronizationCode, duke@435: "monitors must always exist for synchronized methods"); duke@435: duke@435: // Build the growable array of ScopeValues for exp stack duke@435: GrowableArray *monarray = new GrowableArray(num_mon); duke@435: duke@435: // Loop over monitors and insert into array duke@435: for(idx = 0; idx < num_mon; idx++) { duke@435: // Grab the node that defines this monitor kvn@895: Node* box_node = sfn->monitor_box(jvms, idx); kvn@895: Node* obj_node = sfn->monitor_obj(jvms, idx); duke@435: duke@435: // Create ScopeValue for object duke@435: ScopeValue *scval = NULL; kvn@498: kvn@498: if( obj_node->is_SafePointScalarObject() ) { kvn@498: SafePointScalarObjectNode* spobj = obj_node->as_SafePointScalarObject(); kvn@498: scval = Compile::sv_for_node_id(objs, spobj->_idx); kvn@498: if (scval == NULL) { kvn@498: const Type *t = obj_node->bottom_type(); kvn@498: ciKlass* cik = t->is_oopptr()->klass(); kvn@498: assert(cik->is_instance_klass() || kvn@498: cik->is_array_klass(), "Not supported allocation."); kvn@498: ObjectValue* sv = new ObjectValue(spobj->_idx, kvn@498: new ConstantOopWriteValue(cik->encoding())); kvn@498: Compile::set_sv_for_object_node(objs, sv); kvn@498: kvn@498: uint first_ind = spobj->first_index(); kvn@498: for (uint i = 0; i < spobj->n_fields(); i++) { kvn@498: Node* fld_node = sfn->in(first_ind+i); kvn@498: (void)FillLocArray(sv->field_values()->length(), sfn, fld_node, sv->field_values(), objs); kvn@498: } kvn@498: scval = sv; kvn@498: } kvn@498: } else if( !obj_node->is_Con() ) { duke@435: OptoReg::Name obj_reg = _regalloc->get_reg_first(obj_node); kvn@766: if( obj_node->bottom_type()->base() == Type::NarrowOop ) { kvn@766: scval = new_loc_value( _regalloc, obj_reg, Location::narrowoop ); kvn@766: } else { kvn@766: scval = new_loc_value( _regalloc, obj_reg, Location::oop ); kvn@766: } duke@435: } else { kvn@766: const TypePtr *tp = obj_node->bottom_type()->make_ptr(); kvn@766: scval = new ConstantOopWriteValue(tp->is_instptr()->const_oop()->encoding()); duke@435: } duke@435: duke@435: OptoReg::Name box_reg = BoxLockNode::stack_slot(box_node); kvn@501: Location basic_lock = Location::new_stk_loc(Location::normal,_regalloc->reg2offset(box_reg)); kvn@895: while( !box_node->is_BoxLock() ) box_node = box_node->in(1); kvn@501: monarray->append(new MonitorValue(scval, basic_lock, box_node->as_BoxLock()->is_eliminated())); duke@435: } duke@435: kvn@498: // We dump the object pool first, since deoptimization reads it in first. kvn@498: debug_info()->dump_object_pool(objs); kvn@498: duke@435: // Build first class objects to pass to scope duke@435: DebugToken *locvals = debug_info()->create_scope_values(locarray); duke@435: DebugToken *expvals = debug_info()->create_scope_values(exparray); duke@435: DebugToken *monvals = debug_info()->create_monitor_values(monarray); duke@435: duke@435: // Make method available for all Safepoints duke@435: ciMethod* scope_method = method ? method : _method; duke@435: // Describe the scope here duke@435: assert(jvms->bci() >= InvocationEntryBci && jvms->bci() <= 0x10000, "must be a valid or entry BCI"); kvn@498: // Now we can describe the scope. duke@435: debug_info()->describe_scope(safepoint_pc_offset,scope_method,jvms->bci(),locvals,expvals,monvals); duke@435: } // End jvms loop duke@435: duke@435: // Mark the end of the scope set. duke@435: debug_info()->end_safepoint(safepoint_pc_offset); duke@435: } duke@435: duke@435: duke@435: duke@435: // A simplified version of Process_OopMap_Node, to handle non-safepoints. duke@435: class NonSafepointEmitter { duke@435: Compile* C; duke@435: JVMState* _pending_jvms; duke@435: int _pending_offset; duke@435: duke@435: void emit_non_safepoint(); duke@435: duke@435: public: duke@435: NonSafepointEmitter(Compile* compile) { duke@435: this->C = compile; duke@435: _pending_jvms = NULL; duke@435: _pending_offset = 0; duke@435: } duke@435: duke@435: void observe_instruction(Node* n, int pc_offset) { duke@435: if (!C->debug_info()->recording_non_safepoints()) return; duke@435: duke@435: Node_Notes* nn = C->node_notes_at(n->_idx); duke@435: if (nn == NULL || nn->jvms() == NULL) return; duke@435: if (_pending_jvms != NULL && duke@435: _pending_jvms->same_calls_as(nn->jvms())) { duke@435: // Repeated JVMS? Stretch it up here. duke@435: _pending_offset = pc_offset; duke@435: } else { duke@435: if (_pending_jvms != NULL && duke@435: _pending_offset < pc_offset) { duke@435: emit_non_safepoint(); duke@435: } duke@435: _pending_jvms = NULL; duke@435: if (pc_offset > C->debug_info()->last_pc_offset()) { duke@435: // This is the only way _pending_jvms can become non-NULL: duke@435: _pending_jvms = nn->jvms(); duke@435: _pending_offset = pc_offset; duke@435: } duke@435: } duke@435: } duke@435: duke@435: // Stay out of the way of real safepoints: duke@435: void observe_safepoint(JVMState* jvms, int pc_offset) { duke@435: if (_pending_jvms != NULL && duke@435: !_pending_jvms->same_calls_as(jvms) && duke@435: _pending_offset < pc_offset) { duke@435: emit_non_safepoint(); duke@435: } duke@435: _pending_jvms = NULL; duke@435: } duke@435: duke@435: void flush_at_end() { duke@435: if (_pending_jvms != NULL) { duke@435: emit_non_safepoint(); duke@435: } duke@435: _pending_jvms = NULL; duke@435: } duke@435: }; duke@435: duke@435: void NonSafepointEmitter::emit_non_safepoint() { duke@435: JVMState* youngest_jvms = _pending_jvms; duke@435: int pc_offset = _pending_offset; duke@435: duke@435: // Clear it now: duke@435: _pending_jvms = NULL; duke@435: duke@435: DebugInformationRecorder* debug_info = C->debug_info(); duke@435: assert(debug_info->recording_non_safepoints(), "sanity"); duke@435: duke@435: debug_info->add_non_safepoint(pc_offset); duke@435: int max_depth = youngest_jvms->depth(); duke@435: duke@435: // Visit scopes from oldest to youngest. duke@435: for (int depth = 1; depth <= max_depth; depth++) { duke@435: JVMState* jvms = youngest_jvms->of_depth(depth); duke@435: ciMethod* method = jvms->has_method() ? jvms->method() : NULL; duke@435: debug_info->describe_scope(pc_offset, method, jvms->bci()); duke@435: } duke@435: duke@435: // Mark the end of the scope set. duke@435: debug_info->end_non_safepoint(pc_offset); duke@435: } duke@435: duke@435: duke@435: duke@435: // helper for Fill_buffer bailout logic duke@435: static void turn_off_compiler(Compile* C) { duke@435: if (CodeCache::unallocated_capacity() >= CodeCacheMinimumFreeSpace*10) { duke@435: // Do not turn off compilation if a single giant method has duke@435: // blown the code cache size. duke@435: C->record_failure("excessive request to CodeCache"); duke@435: } else { kvn@463: // Let CompilerBroker disable further compilations. duke@435: C->record_failure("CodeCache is full"); duke@435: } duke@435: } duke@435: duke@435: duke@435: //------------------------------Fill_buffer------------------------------------ duke@435: void Compile::Fill_buffer() { duke@435: duke@435: // Set the initially allocated size duke@435: int code_req = initial_code_capacity; duke@435: int locs_req = initial_locs_capacity; duke@435: int stub_req = TraceJumps ? initial_stub_capacity * 10 : initial_stub_capacity; duke@435: int const_req = initial_const_capacity; duke@435: bool labels_not_set = true; duke@435: duke@435: int pad_req = NativeCall::instruction_size; duke@435: // The extra spacing after the code is necessary on some platforms. duke@435: // Sometimes we need to patch in a jump after the last instruction, duke@435: // if the nmethod has been deoptimized. (See 4932387, 4894843.) duke@435: duke@435: uint i; duke@435: // Compute the byte offset where we can store the deopt pc. duke@435: if (fixed_slots() != 0) { duke@435: _orig_pc_slot_offset_in_bytes = _regalloc->reg2offset(OptoReg::stack2reg(_orig_pc_slot)); duke@435: } duke@435: duke@435: // Compute prolog code size duke@435: _method_size = 0; duke@435: _frame_slots = OptoReg::reg2stack(_matcher->_old_SP)+_regalloc->_framesize; duke@435: #ifdef IA64 duke@435: if (save_argument_registers()) { duke@435: // 4815101: this is a stub with implicit and unknown precision fp args. duke@435: // The usual spill mechanism can only generate stfd's in this case, which duke@435: // doesn't work if the fp reg to spill contains a single-precision denorm. duke@435: // Instead, we hack around the normal spill mechanism using stfspill's and duke@435: // ldffill's in the MachProlog and MachEpilog emit methods. We allocate duke@435: // space here for the fp arg regs (f8-f15) we're going to thusly spill. duke@435: // duke@435: // If we ever implement 16-byte 'registers' == stack slots, we can duke@435: // get rid of this hack and have SpillCopy generate stfspill/ldffill duke@435: // instead of stfd/stfs/ldfd/ldfs. duke@435: _frame_slots += 8*(16/BytesPerInt); duke@435: } duke@435: #endif duke@435: assert( _frame_slots >= 0 && _frame_slots < 1000000, "sanity check" ); duke@435: duke@435: // Create an array of unused labels, one for each basic block duke@435: Label *blk_labels = NEW_RESOURCE_ARRAY(Label, _cfg->_num_blocks+1); duke@435: duke@435: for( i=0; i <= _cfg->_num_blocks; i++ ) { duke@435: blk_labels[i].init(); duke@435: } duke@435: duke@435: // If this machine supports different size branch offsets, then pre-compute duke@435: // the length of the blocks never@850: if( _matcher->is_short_branch_offset(-1, 0) ) { duke@435: Shorten_branches(blk_labels, code_req, locs_req, stub_req, const_req); duke@435: labels_not_set = false; duke@435: } duke@435: duke@435: // nmethod and CodeBuffer count stubs & constants as part of method's code. duke@435: int exception_handler_req = size_exception_handler(); duke@435: int deopt_handler_req = size_deopt_handler(); duke@435: exception_handler_req += MAX_stubs_size; // add marginal slop for handler duke@435: deopt_handler_req += MAX_stubs_size; // add marginal slop for handler duke@435: stub_req += MAX_stubs_size; // ensure per-stub margin duke@435: code_req += MAX_inst_size; // ensure per-instruction margin duke@435: if (StressCodeBuffers) duke@435: code_req = const_req = stub_req = exception_handler_req = deopt_handler_req = 0x10; // force expansion duke@435: int total_req = code_req + pad_req + stub_req + exception_handler_req + deopt_handler_req + const_req; duke@435: CodeBuffer* cb = code_buffer(); duke@435: cb->initialize(total_req, locs_req); duke@435: duke@435: // Have we run out of code space? duke@435: if (cb->blob() == NULL) { duke@435: turn_off_compiler(this); duke@435: return; duke@435: } duke@435: // Configure the code buffer. duke@435: cb->initialize_consts_size(const_req); duke@435: cb->initialize_stubs_size(stub_req); duke@435: cb->initialize_oop_recorder(env()->oop_recorder()); duke@435: duke@435: // fill in the nop array for bundling computations duke@435: MachNode *_nop_list[Bundle::_nop_count]; duke@435: Bundle::initialize_nops(_nop_list, this); duke@435: duke@435: // Create oopmap set. duke@435: _oop_map_set = new OopMapSet(); duke@435: duke@435: // !!!!! This preserves old handling of oopmaps for now duke@435: debug_info()->set_oopmaps(_oop_map_set); duke@435: duke@435: // Count and start of implicit null check instructions duke@435: uint inct_cnt = 0; duke@435: uint *inct_starts = NEW_RESOURCE_ARRAY(uint, _cfg->_num_blocks+1); duke@435: duke@435: // Count and start of calls duke@435: uint *call_returns = NEW_RESOURCE_ARRAY(uint, _cfg->_num_blocks+1); duke@435: duke@435: uint return_offset = 0; duke@435: MachNode *nop = new (this) MachNopNode(); duke@435: duke@435: int previous_offset = 0; duke@435: int current_offset = 0; duke@435: int last_call_offset = -1; duke@435: duke@435: // Create an array of unused labels, one for each basic block, if printing is enabled duke@435: #ifndef PRODUCT duke@435: int *node_offsets = NULL; duke@435: uint node_offset_limit = unique(); duke@435: duke@435: duke@435: if ( print_assembly() ) duke@435: node_offsets = NEW_RESOURCE_ARRAY(int, node_offset_limit); duke@435: #endif duke@435: duke@435: NonSafepointEmitter non_safepoints(this); // emit non-safepoints lazily duke@435: duke@435: // ------------------ duke@435: // Now fill in the code buffer duke@435: Node *delay_slot = NULL; duke@435: duke@435: for( i=0; i < _cfg->_num_blocks; i++ ) { duke@435: Block *b = _cfg->_blocks[i]; duke@435: duke@435: Node *head = b->head(); duke@435: duke@435: // If this block needs to start aligned (i.e, can be reached other duke@435: // than by falling-thru from the previous block), then force the duke@435: // start of a new bundle. duke@435: if( Pipeline::requires_bundling() && starts_bundle(head) ) duke@435: cb->flush_bundle(true); duke@435: duke@435: // Define the label at the beginning of the basic block duke@435: if( labels_not_set ) duke@435: MacroAssembler(cb).bind( blk_labels[b->_pre_order] ); duke@435: duke@435: else duke@435: assert( blk_labels[b->_pre_order].loc_pos() == cb->code_size(), duke@435: "label position does not match code offset" ); duke@435: duke@435: uint last_inst = b->_nodes.size(); duke@435: duke@435: // Emit block normally, except for last instruction. duke@435: // Emit means "dump code bits into code buffer". duke@435: for( uint j = 0; j_nodes[j]; duke@435: duke@435: // See if delay slots are supported duke@435: if (valid_bundle_info(n) && duke@435: node_bundling(n)->used_in_unconditional_delay()) { duke@435: assert(delay_slot == NULL, "no use of delay slot node"); duke@435: assert(n->size(_regalloc) == Pipeline::instr_unit_size(), "delay slot instruction wrong size"); duke@435: duke@435: delay_slot = n; duke@435: continue; duke@435: } duke@435: duke@435: // If this starts a new instruction group, then flush the current one duke@435: // (but allow split bundles) duke@435: if( Pipeline::requires_bundling() && starts_bundle(n) ) duke@435: cb->flush_bundle(false); duke@435: duke@435: // The following logic is duplicated in the code ifdeffed for twisti@1040: // ENABLE_ZAP_DEAD_LOCALS which appears above in this file. It duke@435: // should be factored out. Or maybe dispersed to the nodes? duke@435: duke@435: // Special handling for SafePoint/Call Nodes duke@435: bool is_mcall = false; duke@435: if( n->is_Mach() ) { duke@435: MachNode *mach = n->as_Mach(); duke@435: is_mcall = n->is_MachCall(); duke@435: bool is_sfn = n->is_MachSafePoint(); duke@435: duke@435: // If this requires all previous instructions be flushed, then do so duke@435: if( is_sfn || is_mcall || mach->alignment_required() != 1) { duke@435: cb->flush_bundle(true); duke@435: current_offset = cb->code_size(); duke@435: } duke@435: duke@435: // align the instruction if necessary duke@435: int nop_size = nop->size(_regalloc); duke@435: int padding = mach->compute_padding(current_offset); duke@435: // Make sure safepoint node for polling is distinct from a call's duke@435: // return by adding a nop if needed. duke@435: if (is_sfn && !is_mcall && padding == 0 && current_offset == last_call_offset ) { duke@435: padding = nop_size; duke@435: } duke@435: assert( labels_not_set || padding == 0, "instruction should already be aligned") duke@435: duke@435: if(padding > 0) { duke@435: assert((padding % nop_size) == 0, "padding is not a multiple of NOP size"); duke@435: int nops_cnt = padding / nop_size; duke@435: MachNode *nop = new (this) MachNopNode(nops_cnt); duke@435: b->_nodes.insert(j++, nop); duke@435: last_inst++; duke@435: _cfg->_bbs.map( nop->_idx, b ); duke@435: nop->emit(*cb, _regalloc); duke@435: cb->flush_bundle(true); duke@435: current_offset = cb->code_size(); duke@435: } duke@435: duke@435: // Remember the start of the last call in a basic block duke@435: if (is_mcall) { duke@435: MachCallNode *mcall = mach->as_MachCall(); duke@435: duke@435: // This destination address is NOT PC-relative duke@435: mcall->method_set((intptr_t)mcall->entry_point()); duke@435: duke@435: // Save the return address duke@435: call_returns[b->_pre_order] = current_offset + mcall->ret_addr_offset(); duke@435: duke@435: if (!mcall->is_safepoint_node()) { duke@435: is_mcall = false; duke@435: is_sfn = false; duke@435: } duke@435: } duke@435: duke@435: // sfn will be valid whenever mcall is valid now because of inheritance duke@435: if( is_sfn || is_mcall ) { duke@435: duke@435: // Handle special safepoint nodes for synchronization duke@435: if( !is_mcall ) { duke@435: MachSafePointNode *sfn = mach->as_MachSafePoint(); duke@435: // !!!!! Stubs only need an oopmap right now, so bail out duke@435: if( sfn->jvms()->method() == NULL) { duke@435: // Write the oopmap directly to the code blob??!! duke@435: # ifdef ENABLE_ZAP_DEAD_LOCALS duke@435: assert( !is_node_getting_a_safepoint(sfn), "logic does not match; false positive"); duke@435: # endif duke@435: continue; duke@435: } duke@435: } // End synchronization duke@435: duke@435: non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(), duke@435: current_offset); duke@435: Process_OopMap_Node(mach, current_offset); duke@435: } // End if safepoint duke@435: duke@435: // If this is a null check, then add the start of the previous instruction to the list duke@435: else if( mach->is_MachNullCheck() ) { duke@435: inct_starts[inct_cnt++] = previous_offset; duke@435: } duke@435: duke@435: // If this is a branch, then fill in the label with the target BB's label duke@435: else if ( mach->is_Branch() ) { duke@435: duke@435: if ( mach->ideal_Opcode() == Op_Jump ) { duke@435: for (uint h = 0; h < b->_num_succs; h++ ) { duke@435: Block* succs_block = b->_succs[h]; duke@435: for (uint j = 1; j < succs_block->num_preds(); j++) { duke@435: Node* jpn = succs_block->pred(j); duke@435: if ( jpn->is_JumpProj() && jpn->in(0) == mach ) { duke@435: uint block_num = succs_block->non_connector()->_pre_order; duke@435: Label *blkLabel = &blk_labels[block_num]; duke@435: mach->add_case_label(jpn->as_JumpProj()->proj_no(), blkLabel); duke@435: } duke@435: } duke@435: } duke@435: } else { duke@435: // For Branchs duke@435: // This requires the TRUE branch target be in succs[0] duke@435: uint block_num = b->non_connector_successor(0)->_pre_order; duke@435: mach->label_set( blk_labels[block_num], block_num ); duke@435: } duke@435: } duke@435: duke@435: #ifdef ASSERT twisti@1040: // Check that oop-store precedes the card-mark duke@435: else if( mach->ideal_Opcode() == Op_StoreCM ) { duke@435: uint storeCM_idx = j; duke@435: Node *oop_store = mach->in(mach->_cnt); // First precedence edge duke@435: assert( oop_store != NULL, "storeCM expects a precedence edge"); duke@435: uint i4; duke@435: for( i4 = 0; i4 < last_inst; ++i4 ) { duke@435: if( b->_nodes[i4] == oop_store ) break; duke@435: } duke@435: // Note: This test can provide a false failure if other precedence duke@435: // edges have been added to the storeCMNode. duke@435: assert( i4 == last_inst || i4 < storeCM_idx, "CM card-mark executes before oop-store"); duke@435: } duke@435: #endif duke@435: duke@435: else if( !n->is_Proj() ) { twisti@1040: // Remember the beginning of the previous instruction, in case duke@435: // it's followed by a flag-kill and a null-check. Happens on duke@435: // Intel all the time, with add-to-memory kind of opcodes. duke@435: previous_offset = current_offset; duke@435: } duke@435: } duke@435: duke@435: // Verify that there is sufficient space remaining duke@435: cb->insts()->maybe_expand_to_ensure_remaining(MAX_inst_size); duke@435: if (cb->blob() == NULL) { duke@435: turn_off_compiler(this); duke@435: return; duke@435: } duke@435: duke@435: // Save the offset for the listing duke@435: #ifndef PRODUCT duke@435: if( node_offsets && n->_idx < node_offset_limit ) duke@435: node_offsets[n->_idx] = cb->code_size(); duke@435: #endif duke@435: duke@435: // "Normal" instruction case duke@435: n->emit(*cb, _regalloc); duke@435: current_offset = cb->code_size(); duke@435: non_safepoints.observe_instruction(n, current_offset); duke@435: duke@435: // mcall is last "call" that can be a safepoint duke@435: // record it so we can see if a poll will directly follow it duke@435: // in which case we'll need a pad to make the PcDesc sites unique duke@435: // see 5010568. This can be slightly inaccurate but conservative duke@435: // in the case that return address is not actually at current_offset. duke@435: // This is a small price to pay. duke@435: duke@435: if (is_mcall) { duke@435: last_call_offset = current_offset; duke@435: } duke@435: duke@435: // See if this instruction has a delay slot duke@435: if ( valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { duke@435: assert(delay_slot != NULL, "expecting delay slot node"); duke@435: duke@435: // Back up 1 instruction duke@435: cb->set_code_end( duke@435: cb->code_end()-Pipeline::instr_unit_size()); duke@435: duke@435: // Save the offset for the listing duke@435: #ifndef PRODUCT duke@435: if( node_offsets && delay_slot->_idx < node_offset_limit ) duke@435: node_offsets[delay_slot->_idx] = cb->code_size(); duke@435: #endif duke@435: duke@435: // Support a SafePoint in the delay slot duke@435: if( delay_slot->is_MachSafePoint() ) { duke@435: MachNode *mach = delay_slot->as_Mach(); duke@435: // !!!!! Stubs only need an oopmap right now, so bail out duke@435: if( !mach->is_MachCall() && mach->as_MachSafePoint()->jvms()->method() == NULL ) { duke@435: // Write the oopmap directly to the code blob??!! duke@435: # ifdef ENABLE_ZAP_DEAD_LOCALS duke@435: assert( !is_node_getting_a_safepoint(mach), "logic does not match; false positive"); duke@435: # endif duke@435: delay_slot = NULL; duke@435: continue; duke@435: } duke@435: duke@435: int adjusted_offset = current_offset - Pipeline::instr_unit_size(); duke@435: non_safepoints.observe_safepoint(mach->as_MachSafePoint()->jvms(), duke@435: adjusted_offset); duke@435: // Generate an OopMap entry duke@435: Process_OopMap_Node(mach, adjusted_offset); duke@435: } duke@435: duke@435: // Insert the delay slot instruction duke@435: delay_slot->emit(*cb, _regalloc); duke@435: duke@435: // Don't reuse it duke@435: delay_slot = NULL; duke@435: } duke@435: duke@435: } // End for all instructions in block duke@435: rasbold@853: // If the next block is the top of a loop, pad this block out to align rasbold@853: // the loop top a little. Helps prevent pipe stalls at loop back branches. duke@435: int nop_size = (new (this) MachNopNode())->size(_regalloc); duke@435: if( i<_cfg->_num_blocks-1 ) { duke@435: Block *nb = _cfg->_blocks[i+1]; duke@435: uint padding = nb->alignment_padding(current_offset); duke@435: if( padding > 0 ) { duke@435: MachNode *nop = new (this) MachNopNode(padding / nop_size); duke@435: b->_nodes.insert( b->_nodes.size(), nop ); duke@435: _cfg->_bbs.map( nop->_idx, b ); duke@435: nop->emit(*cb, _regalloc); duke@435: current_offset = cb->code_size(); duke@435: } duke@435: } duke@435: duke@435: } // End of for all blocks duke@435: duke@435: non_safepoints.flush_at_end(); duke@435: duke@435: // Offset too large? duke@435: if (failing()) return; duke@435: duke@435: // Define a pseudo-label at the end of the code duke@435: MacroAssembler(cb).bind( blk_labels[_cfg->_num_blocks] ); duke@435: duke@435: // Compute the size of the first block duke@435: _first_block_size = blk_labels[1].loc_pos() - blk_labels[0].loc_pos(); duke@435: duke@435: assert(cb->code_size() < 500000, "method is unreasonably large"); duke@435: duke@435: // ------------------ duke@435: duke@435: #ifndef PRODUCT duke@435: // Information on the size of the method, without the extraneous code duke@435: Scheduling::increment_method_size(cb->code_size()); duke@435: #endif duke@435: duke@435: // ------------------ duke@435: // Fill in exception table entries. duke@435: FillExceptionTables(inct_cnt, call_returns, inct_starts, blk_labels); duke@435: duke@435: // Only java methods have exception handlers and deopt handlers duke@435: if (_method) { duke@435: // Emit the exception handler code. duke@435: _code_offsets.set_value(CodeOffsets::Exceptions, emit_exception_handler(*cb)); duke@435: // Emit the deopt handler code. duke@435: _code_offsets.set_value(CodeOffsets::Deopt, emit_deopt_handler(*cb)); duke@435: } duke@435: duke@435: // One last check for failed CodeBuffer::expand: duke@435: if (cb->blob() == NULL) { duke@435: turn_off_compiler(this); duke@435: return; duke@435: } duke@435: duke@435: #ifndef PRODUCT duke@435: // Dump the assembly code, including basic-block numbers duke@435: if (print_assembly()) { duke@435: ttyLocker ttyl; // keep the following output all in one block duke@435: if (!VMThread::should_terminate()) { // test this under the tty lock duke@435: // This output goes directly to the tty, not the compiler log. duke@435: // To enable tools to match it up with the compilation activity, duke@435: // be sure to tag this tty output with the compile ID. duke@435: if (xtty != NULL) { duke@435: xtty->head("opto_assembly compile_id='%d'%s", compile_id(), duke@435: is_osr_compilation() ? " compile_kind='osr'" : duke@435: ""); duke@435: } duke@435: if (method() != NULL) { duke@435: method()->print_oop(); duke@435: print_codes(); duke@435: } duke@435: dump_asm(node_offsets, node_offset_limit); duke@435: if (xtty != NULL) { duke@435: xtty->tail("opto_assembly"); duke@435: } duke@435: } duke@435: } duke@435: #endif duke@435: duke@435: } duke@435: duke@435: void Compile::FillExceptionTables(uint cnt, uint *call_returns, uint *inct_starts, Label *blk_labels) { duke@435: _inc_table.set_size(cnt); duke@435: duke@435: uint inct_cnt = 0; duke@435: for( uint i=0; i<_cfg->_num_blocks; i++ ) { duke@435: Block *b = _cfg->_blocks[i]; duke@435: Node *n = NULL; duke@435: int j; duke@435: duke@435: // Find the branch; ignore trailing NOPs. duke@435: for( j = b->_nodes.size()-1; j>=0; j-- ) { duke@435: n = b->_nodes[j]; duke@435: if( !n->is_Mach() || n->as_Mach()->ideal_Opcode() != Op_Con ) duke@435: break; duke@435: } duke@435: duke@435: // If we didn't find anything, continue duke@435: if( j < 0 ) continue; duke@435: duke@435: // Compute ExceptionHandlerTable subtable entry and add it duke@435: // (skip empty blocks) duke@435: if( n->is_Catch() ) { duke@435: duke@435: // Get the offset of the return from the call duke@435: uint call_return = call_returns[b->_pre_order]; duke@435: #ifdef ASSERT duke@435: assert( call_return > 0, "no call seen for this basic block" ); duke@435: while( b->_nodes[--j]->Opcode() == Op_MachProj ) ; duke@435: assert( b->_nodes[j]->is_Call(), "CatchProj must follow call" ); duke@435: #endif duke@435: // last instruction is a CatchNode, find it's CatchProjNodes duke@435: int nof_succs = b->_num_succs; duke@435: // allocate space duke@435: GrowableArray handler_bcis(nof_succs); duke@435: GrowableArray handler_pcos(nof_succs); duke@435: // iterate through all successors duke@435: for (int j = 0; j < nof_succs; j++) { duke@435: Block* s = b->_succs[j]; duke@435: bool found_p = false; duke@435: for( uint k = 1; k < s->num_preds(); k++ ) { duke@435: Node *pk = s->pred(k); duke@435: if( pk->is_CatchProj() && pk->in(0) == n ) { duke@435: const CatchProjNode* p = pk->as_CatchProj(); duke@435: found_p = true; duke@435: // add the corresponding handler bci & pco information duke@435: if( p->_con != CatchProjNode::fall_through_index ) { duke@435: // p leads to an exception handler (and is not fall through) duke@435: assert(s == _cfg->_blocks[s->_pre_order],"bad numbering"); duke@435: // no duplicates, please duke@435: if( !handler_bcis.contains(p->handler_bci()) ) { duke@435: uint block_num = s->non_connector()->_pre_order; duke@435: handler_bcis.append(p->handler_bci()); duke@435: handler_pcos.append(blk_labels[block_num].loc_pos()); duke@435: } duke@435: } duke@435: } duke@435: } duke@435: assert(found_p, "no matching predecessor found"); duke@435: // Note: Due to empty block removal, one block may have duke@435: // several CatchProj inputs, from the same Catch. duke@435: } duke@435: duke@435: // Set the offset of the return from the call duke@435: _handler_table.add_subtable(call_return, &handler_bcis, NULL, &handler_pcos); duke@435: continue; duke@435: } duke@435: duke@435: // Handle implicit null exception table updates duke@435: if( n->is_MachNullCheck() ) { duke@435: uint block_num = b->non_connector_successor(0)->_pre_order; duke@435: _inc_table.append( inct_starts[inct_cnt++], blk_labels[block_num].loc_pos() ); duke@435: continue; duke@435: } duke@435: } // End of for all blocks fill in exception table entries duke@435: } duke@435: duke@435: // Static Variables duke@435: #ifndef PRODUCT duke@435: uint Scheduling::_total_nop_size = 0; duke@435: uint Scheduling::_total_method_size = 0; duke@435: uint Scheduling::_total_branches = 0; duke@435: uint Scheduling::_total_unconditional_delays = 0; duke@435: uint Scheduling::_total_instructions_per_bundle[Pipeline::_max_instrs_per_cycle+1]; duke@435: #endif duke@435: duke@435: // Initializer for class Scheduling duke@435: duke@435: Scheduling::Scheduling(Arena *arena, Compile &compile) duke@435: : _arena(arena), duke@435: _cfg(compile.cfg()), duke@435: _bbs(compile.cfg()->_bbs), duke@435: _regalloc(compile.regalloc()), duke@435: _reg_node(arena), duke@435: _bundle_instr_count(0), duke@435: _bundle_cycle_number(0), duke@435: _scheduled(arena), duke@435: _available(arena), duke@435: _next_node(NULL), duke@435: _bundle_use(0, 0, resource_count, &_bundle_use_elements[0]), duke@435: _pinch_free_list(arena) duke@435: #ifndef PRODUCT duke@435: , _branches(0) duke@435: , _unconditional_delays(0) duke@435: #endif duke@435: { duke@435: // Create a MachNopNode duke@435: _nop = new (&compile) MachNopNode(); duke@435: duke@435: // Now that the nops are in the array, save the count duke@435: // (but allow entries for the nops) duke@435: _node_bundling_limit = compile.unique(); duke@435: uint node_max = _regalloc->node_regs_max_index(); duke@435: duke@435: compile.set_node_bundling_limit(_node_bundling_limit); duke@435: twisti@1040: // This one is persistent within the Compile class duke@435: _node_bundling_base = NEW_ARENA_ARRAY(compile.comp_arena(), Bundle, node_max); duke@435: duke@435: // Allocate space for fixed-size arrays duke@435: _node_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max); duke@435: _uses = NEW_ARENA_ARRAY(arena, short, node_max); duke@435: _current_latency = NEW_ARENA_ARRAY(arena, unsigned short, node_max); duke@435: duke@435: // Clear the arrays duke@435: memset(_node_bundling_base, 0, node_max * sizeof(Bundle)); duke@435: memset(_node_latency, 0, node_max * sizeof(unsigned short)); duke@435: memset(_uses, 0, node_max * sizeof(short)); duke@435: memset(_current_latency, 0, node_max * sizeof(unsigned short)); duke@435: duke@435: // Clear the bundling information duke@435: memcpy(_bundle_use_elements, duke@435: Pipeline_Use::elaborated_elements, duke@435: sizeof(Pipeline_Use::elaborated_elements)); duke@435: duke@435: // Get the last node duke@435: Block *bb = _cfg->_blocks[_cfg->_blocks.size()-1]; duke@435: duke@435: _next_node = bb->_nodes[bb->_nodes.size()-1]; duke@435: } duke@435: duke@435: #ifndef PRODUCT duke@435: // Scheduling destructor duke@435: Scheduling::~Scheduling() { duke@435: _total_branches += _branches; duke@435: _total_unconditional_delays += _unconditional_delays; duke@435: } duke@435: #endif duke@435: duke@435: // Step ahead "i" cycles duke@435: void Scheduling::step(uint i) { duke@435: duke@435: Bundle *bundle = node_bundling(_next_node); duke@435: bundle->set_starts_bundle(); duke@435: duke@435: // Update the bundle record, but leave the flags information alone duke@435: if (_bundle_instr_count > 0) { duke@435: bundle->set_instr_count(_bundle_instr_count); duke@435: bundle->set_resources_used(_bundle_use.resourcesUsed()); duke@435: } duke@435: duke@435: // Update the state information duke@435: _bundle_instr_count = 0; duke@435: _bundle_cycle_number += i; duke@435: _bundle_use.step(i); duke@435: } duke@435: duke@435: void Scheduling::step_and_clear() { duke@435: Bundle *bundle = node_bundling(_next_node); duke@435: bundle->set_starts_bundle(); duke@435: duke@435: // Update the bundle record duke@435: if (_bundle_instr_count > 0) { duke@435: bundle->set_instr_count(_bundle_instr_count); duke@435: bundle->set_resources_used(_bundle_use.resourcesUsed()); duke@435: duke@435: _bundle_cycle_number += 1; duke@435: } duke@435: duke@435: // Clear the bundling information duke@435: _bundle_instr_count = 0; duke@435: _bundle_use.reset(); duke@435: duke@435: memcpy(_bundle_use_elements, duke@435: Pipeline_Use::elaborated_elements, duke@435: sizeof(Pipeline_Use::elaborated_elements)); duke@435: } duke@435: duke@435: //------------------------------ScheduleAndBundle------------------------------ duke@435: // Perform instruction scheduling and bundling over the sequence of duke@435: // instructions in backwards order. duke@435: void Compile::ScheduleAndBundle() { duke@435: duke@435: // Don't optimize this if it isn't a method duke@435: if (!_method) duke@435: return; duke@435: duke@435: // Don't optimize this if scheduling is disabled duke@435: if (!do_scheduling()) duke@435: return; duke@435: duke@435: NOT_PRODUCT( TracePhase t2("isched", &_t_instrSched, TimeCompiler); ) duke@435: duke@435: // Create a data structure for all the scheduling information duke@435: Scheduling scheduling(Thread::current()->resource_area(), *this); duke@435: duke@435: // Walk backwards over each basic block, computing the needed alignment duke@435: // Walk over all the basic blocks duke@435: scheduling.DoScheduling(); duke@435: } duke@435: duke@435: //------------------------------ComputeLocalLatenciesForward------------------- duke@435: // Compute the latency of all the instructions. This is fairly simple, duke@435: // because we already have a legal ordering. Walk over the instructions duke@435: // from first to last, and compute the latency of the instruction based twisti@1040: // on the latency of the preceding instruction(s). duke@435: void Scheduling::ComputeLocalLatenciesForward(const Block *bb) { duke@435: #ifndef PRODUCT duke@435: if (_cfg->C->trace_opto_output()) duke@435: tty->print("# -> ComputeLocalLatenciesForward\n"); duke@435: #endif duke@435: duke@435: // Walk over all the schedulable instructions duke@435: for( uint j=_bb_start; j < _bb_end; j++ ) { duke@435: duke@435: // This is a kludge, forcing all latency calculations to start at 1. duke@435: // Used to allow latency 0 to force an instruction to the beginning duke@435: // of the bb duke@435: uint latency = 1; duke@435: Node *use = bb->_nodes[j]; duke@435: uint nlen = use->len(); duke@435: duke@435: // Walk over all the inputs duke@435: for ( uint k=0; k < nlen; k++ ) { duke@435: Node *def = use->in(k); duke@435: if (!def) duke@435: continue; duke@435: duke@435: uint l = _node_latency[def->_idx] + use->latency(k); duke@435: if (latency < l) duke@435: latency = l; duke@435: } duke@435: duke@435: _node_latency[use->_idx] = latency; duke@435: duke@435: #ifndef PRODUCT duke@435: if (_cfg->C->trace_opto_output()) { duke@435: tty->print("# latency %4d: ", latency); duke@435: use->dump(); duke@435: } duke@435: #endif duke@435: } duke@435: duke@435: #ifndef PRODUCT duke@435: if (_cfg->C->trace_opto_output()) duke@435: tty->print("# <- ComputeLocalLatenciesForward\n"); duke@435: #endif duke@435: duke@435: } // end ComputeLocalLatenciesForward duke@435: duke@435: // See if this node fits into the present instruction bundle duke@435: bool Scheduling::NodeFitsInBundle(Node *n) { duke@435: uint n_idx = n->_idx; duke@435: duke@435: // If this is the unconditional delay instruction, then it fits duke@435: if (n == _unconditional_delay_slot) { duke@435: #ifndef PRODUCT duke@435: if (_cfg->C->trace_opto_output()) duke@435: tty->print("# NodeFitsInBundle [%4d]: TRUE; is in unconditional delay slot\n", n->_idx); duke@435: #endif duke@435: return (true); duke@435: } duke@435: duke@435: // If the node cannot be scheduled this cycle, skip it duke@435: if (_current_latency[n_idx] > _bundle_cycle_number) { duke@435: #ifndef PRODUCT duke@435: if (_cfg->C->trace_opto_output()) duke@435: tty->print("# NodeFitsInBundle [%4d]: FALSE; latency %4d > %d\n", duke@435: n->_idx, _current_latency[n_idx], _bundle_cycle_number); duke@435: #endif duke@435: return (false); duke@435: } duke@435: duke@435: const Pipeline *node_pipeline = n->pipeline(); duke@435: duke@435: uint instruction_count = node_pipeline->instructionCount(); duke@435: if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0) duke@435: instruction_count = 0; duke@435: else if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot) duke@435: instruction_count++; duke@435: duke@435: if (_bundle_instr_count + instruction_count > Pipeline::_max_instrs_per_cycle) { duke@435: #ifndef PRODUCT duke@435: if (_cfg->C->trace_opto_output()) duke@435: tty->print("# NodeFitsInBundle [%4d]: FALSE; too many instructions: %d > %d\n", duke@435: n->_idx, _bundle_instr_count + instruction_count, Pipeline::_max_instrs_per_cycle); duke@435: #endif duke@435: return (false); duke@435: } duke@435: duke@435: // Don't allow non-machine nodes to be handled this way duke@435: if (!n->is_Mach() && instruction_count == 0) duke@435: return (false); duke@435: duke@435: // See if there is any overlap duke@435: uint delay = _bundle_use.full_latency(0, node_pipeline->resourceUse()); duke@435: duke@435: if (delay > 0) { duke@435: #ifndef PRODUCT duke@435: if (_cfg->C->trace_opto_output()) duke@435: tty->print("# NodeFitsInBundle [%4d]: FALSE; functional units overlap\n", n_idx); duke@435: #endif duke@435: return false; duke@435: } duke@435: duke@435: #ifndef PRODUCT duke@435: if (_cfg->C->trace_opto_output()) duke@435: tty->print("# NodeFitsInBundle [%4d]: TRUE\n", n_idx); duke@435: #endif duke@435: duke@435: return true; duke@435: } duke@435: duke@435: Node * Scheduling::ChooseNodeToBundle() { duke@435: uint siz = _available.size(); duke@435: duke@435: if (siz == 0) { duke@435: duke@435: #ifndef PRODUCT duke@435: if (_cfg->C->trace_opto_output()) duke@435: tty->print("# ChooseNodeToBundle: NULL\n"); duke@435: #endif duke@435: return (NULL); duke@435: } duke@435: duke@435: // Fast path, if only 1 instruction in the bundle duke@435: if (siz == 1) { duke@435: #ifndef PRODUCT duke@435: if (_cfg->C->trace_opto_output()) { duke@435: tty->print("# ChooseNodeToBundle (only 1): "); duke@435: _available[0]->dump(); duke@435: } duke@435: #endif duke@435: return (_available[0]); duke@435: } duke@435: duke@435: // Don't bother, if the bundle is already full duke@435: if (_bundle_instr_count < Pipeline::_max_instrs_per_cycle) { duke@435: for ( uint i = 0; i < siz; i++ ) { duke@435: Node *n = _available[i]; duke@435: duke@435: // Skip projections, we'll handle them another way duke@435: if (n->is_Proj()) duke@435: continue; duke@435: duke@435: // This presupposed that instructions are inserted into the duke@435: // available list in a legality order; i.e. instructions that duke@435: // must be inserted first are at the head of the list duke@435: if (NodeFitsInBundle(n)) { duke@435: #ifndef PRODUCT duke@435: if (_cfg->C->trace_opto_output()) { duke@435: tty->print("# ChooseNodeToBundle: "); duke@435: n->dump(); duke@435: } duke@435: #endif duke@435: return (n); duke@435: } duke@435: } duke@435: } duke@435: duke@435: // Nothing fits in this bundle, choose the highest priority duke@435: #ifndef PRODUCT duke@435: if (_cfg->C->trace_opto_output()) { duke@435: tty->print("# ChooseNodeToBundle: "); duke@435: _available[0]->dump(); duke@435: } duke@435: #endif duke@435: duke@435: return _available[0]; duke@435: } duke@435: duke@435: //------------------------------AddNodeToAvailableList------------------------- duke@435: void Scheduling::AddNodeToAvailableList(Node *n) { duke@435: assert( !n->is_Proj(), "projections never directly made available" ); duke@435: #ifndef PRODUCT duke@435: if (_cfg->C->trace_opto_output()) { duke@435: tty->print("# AddNodeToAvailableList: "); duke@435: n->dump(); duke@435: } duke@435: #endif duke@435: duke@435: int latency = _current_latency[n->_idx]; duke@435: duke@435: // Insert in latency order (insertion sort) duke@435: uint i; duke@435: for ( i=0; i < _available.size(); i++ ) duke@435: if (_current_latency[_available[i]->_idx] > latency) duke@435: break; duke@435: duke@435: // Special Check for compares following branches duke@435: if( n->is_Mach() && _scheduled.size() > 0 ) { duke@435: int op = n->as_Mach()->ideal_Opcode(); duke@435: Node *last = _scheduled[0]; duke@435: if( last->is_MachIf() && last->in(1) == n && duke@435: ( op == Op_CmpI || duke@435: op == Op_CmpU || duke@435: op == Op_CmpP || duke@435: op == Op_CmpF || duke@435: op == Op_CmpD || duke@435: op == Op_CmpL ) ) { duke@435: duke@435: // Recalculate position, moving to front of same latency duke@435: for ( i=0 ; i < _available.size(); i++ ) duke@435: if (_current_latency[_available[i]->_idx] >= latency) duke@435: break; duke@435: } duke@435: } duke@435: duke@435: // Insert the node in the available list duke@435: _available.insert(i, n); duke@435: duke@435: #ifndef PRODUCT duke@435: if (_cfg->C->trace_opto_output()) duke@435: dump_available(); duke@435: #endif duke@435: } duke@435: duke@435: //------------------------------DecrementUseCounts----------------------------- duke@435: void Scheduling::DecrementUseCounts(Node *n, const Block *bb) { duke@435: for ( uint i=0; i < n->len(); i++ ) { duke@435: Node *def = n->in(i); duke@435: if (!def) continue; duke@435: if( def->is_Proj() ) // If this is a machine projection, then duke@435: def = def->in(0); // propagate usage thru to the base instruction duke@435: duke@435: if( _bbs[def->_idx] != bb ) // Ignore if not block-local duke@435: continue; duke@435: duke@435: // Compute the latency duke@435: uint l = _bundle_cycle_number + n->latency(i); duke@435: if (_current_latency[def->_idx] < l) duke@435: _current_latency[def->_idx] = l; duke@435: duke@435: // If this does not have uses then schedule it duke@435: if ((--_uses[def->_idx]) == 0) duke@435: AddNodeToAvailableList(def); duke@435: } duke@435: } duke@435: duke@435: //------------------------------AddNodeToBundle-------------------------------- duke@435: void Scheduling::AddNodeToBundle(Node *n, const Block *bb) { duke@435: #ifndef PRODUCT duke@435: if (_cfg->C->trace_opto_output()) { duke@435: tty->print("# AddNodeToBundle: "); duke@435: n->dump(); duke@435: } duke@435: #endif duke@435: duke@435: // Remove this from the available list duke@435: uint i; duke@435: for (i = 0; i < _available.size(); i++) duke@435: if (_available[i] == n) duke@435: break; duke@435: assert(i < _available.size(), "entry in _available list not found"); duke@435: _available.remove(i); duke@435: duke@435: // See if this fits in the current bundle duke@435: const Pipeline *node_pipeline = n->pipeline(); duke@435: const Pipeline_Use& node_usage = node_pipeline->resourceUse(); duke@435: duke@435: // Check for instructions to be placed in the delay slot. We duke@435: // do this before we actually schedule the current instruction, duke@435: // because the delay slot follows the current instruction. duke@435: if (Pipeline::_branch_has_delay_slot && duke@435: node_pipeline->hasBranchDelay() && duke@435: !_unconditional_delay_slot) { duke@435: duke@435: uint siz = _available.size(); duke@435: duke@435: // Conditional branches can support an instruction that twisti@1040: // is unconditionally executed and not dependent by the duke@435: // branch, OR a conditionally executed instruction if duke@435: // the branch is taken. In practice, this means that duke@435: // the first instruction at the branch target is duke@435: // copied to the delay slot, and the branch goes to duke@435: // the instruction after that at the branch target duke@435: if ( n->is_Mach() && n->is_Branch() ) { duke@435: duke@435: assert( !n->is_MachNullCheck(), "should not look for delay slot for Null Check" ); duke@435: assert( !n->is_Catch(), "should not look for delay slot for Catch" ); duke@435: duke@435: #ifndef PRODUCT duke@435: _branches++; duke@435: #endif duke@435: duke@435: // At least 1 instruction is on the available list twisti@1040: // that is not dependent on the branch duke@435: for (uint i = 0; i < siz; i++) { duke@435: Node *d = _available[i]; duke@435: const Pipeline *avail_pipeline = d->pipeline(); duke@435: duke@435: // Don't allow safepoints in the branch shadow, that will duke@435: // cause a number of difficulties duke@435: if ( avail_pipeline->instructionCount() == 1 && duke@435: !avail_pipeline->hasMultipleBundles() && duke@435: !avail_pipeline->hasBranchDelay() && duke@435: Pipeline::instr_has_unit_size() && duke@435: d->size(_regalloc) == Pipeline::instr_unit_size() && duke@435: NodeFitsInBundle(d) && duke@435: !node_bundling(d)->used_in_delay()) { duke@435: duke@435: if (d->is_Mach() && !d->is_MachSafePoint()) { duke@435: // A node that fits in the delay slot was found, so we need to duke@435: // set the appropriate bits in the bundle pipeline information so duke@435: // that it correctly indicates resource usage. Later, when we duke@435: // attempt to add this instruction to the bundle, we will skip duke@435: // setting the resource usage. duke@435: _unconditional_delay_slot = d; duke@435: node_bundling(n)->set_use_unconditional_delay(); duke@435: node_bundling(d)->set_used_in_unconditional_delay(); duke@435: _bundle_use.add_usage(avail_pipeline->resourceUse()); duke@435: _current_latency[d->_idx] = _bundle_cycle_number; duke@435: _next_node = d; duke@435: ++_bundle_instr_count; duke@435: #ifndef PRODUCT duke@435: _unconditional_delays++; duke@435: #endif duke@435: break; duke@435: } duke@435: } duke@435: } duke@435: } duke@435: duke@435: // No delay slot, add a nop to the usage duke@435: if (!_unconditional_delay_slot) { duke@435: // See if adding an instruction in the delay slot will overflow duke@435: // the bundle. duke@435: if (!NodeFitsInBundle(_nop)) { duke@435: #ifndef PRODUCT duke@435: if (_cfg->C->trace_opto_output()) duke@435: tty->print("# *** STEP(1 instruction for delay slot) ***\n"); duke@435: #endif duke@435: step(1); duke@435: } duke@435: duke@435: _bundle_use.add_usage(_nop->pipeline()->resourceUse()); duke@435: _next_node = _nop; duke@435: ++_bundle_instr_count; duke@435: } duke@435: duke@435: // See if the instruction in the delay slot requires a duke@435: // step of the bundles duke@435: if (!NodeFitsInBundle(n)) { duke@435: #ifndef PRODUCT duke@435: if (_cfg->C->trace_opto_output()) duke@435: tty->print("# *** STEP(branch won't fit) ***\n"); duke@435: #endif duke@435: // Update the state information duke@435: _bundle_instr_count = 0; duke@435: _bundle_cycle_number += 1; duke@435: _bundle_use.step(1); duke@435: } duke@435: } duke@435: duke@435: // Get the number of instructions duke@435: uint instruction_count = node_pipeline->instructionCount(); duke@435: if (node_pipeline->mayHaveNoCode() && n->size(_regalloc) == 0) duke@435: instruction_count = 0; duke@435: duke@435: // Compute the latency information duke@435: uint delay = 0; duke@435: duke@435: if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) { duke@435: int relative_latency = _current_latency[n->_idx] - _bundle_cycle_number; duke@435: if (relative_latency < 0) duke@435: relative_latency = 0; duke@435: duke@435: delay = _bundle_use.full_latency(relative_latency, node_usage); duke@435: duke@435: // Does not fit in this bundle, start a new one duke@435: if (delay > 0) { duke@435: step(delay); duke@435: duke@435: #ifndef PRODUCT duke@435: if (_cfg->C->trace_opto_output()) duke@435: tty->print("# *** STEP(%d) ***\n", delay); duke@435: #endif duke@435: } duke@435: } duke@435: duke@435: // If this was placed in the delay slot, ignore it duke@435: if (n != _unconditional_delay_slot) { duke@435: duke@435: if (delay == 0) { duke@435: if (node_pipeline->hasMultipleBundles()) { duke@435: #ifndef PRODUCT duke@435: if (_cfg->C->trace_opto_output()) duke@435: tty->print("# *** STEP(multiple instructions) ***\n"); duke@435: #endif duke@435: step(1); duke@435: } duke@435: duke@435: else if (instruction_count + _bundle_instr_count > Pipeline::_max_instrs_per_cycle) { duke@435: #ifndef PRODUCT duke@435: if (_cfg->C->trace_opto_output()) duke@435: tty->print("# *** STEP(%d >= %d instructions) ***\n", duke@435: instruction_count + _bundle_instr_count, duke@435: Pipeline::_max_instrs_per_cycle); duke@435: #endif duke@435: step(1); duke@435: } duke@435: } duke@435: duke@435: if (node_pipeline->hasBranchDelay() && !_unconditional_delay_slot) duke@435: _bundle_instr_count++; duke@435: duke@435: // Set the node's latency duke@435: _current_latency[n->_idx] = _bundle_cycle_number; duke@435: duke@435: // Now merge the functional unit information duke@435: if (instruction_count > 0 || !node_pipeline->mayHaveNoCode()) duke@435: _bundle_use.add_usage(node_usage); duke@435: duke@435: // Increment the number of instructions in this bundle duke@435: _bundle_instr_count += instruction_count; duke@435: duke@435: // Remember this node for later duke@435: if (n->is_Mach()) duke@435: _next_node = n; duke@435: } duke@435: duke@435: // It's possible to have a BoxLock in the graph and in the _bbs mapping but duke@435: // not in the bb->_nodes array. This happens for debug-info-only BoxLocks. duke@435: // 'Schedule' them (basically ignore in the schedule) but do not insert them duke@435: // into the block. All other scheduled nodes get put in the schedule here. duke@435: int op = n->Opcode(); duke@435: if( (op == Op_Node && n->req() == 0) || // anti-dependence node OR duke@435: (op != Op_Node && // Not an unused antidepedence node and duke@435: // not an unallocated boxlock duke@435: (OptoReg::is_valid(_regalloc->get_reg_first(n)) || op != Op_BoxLock)) ) { duke@435: duke@435: // Push any trailing projections duke@435: if( bb->_nodes[bb->_nodes.size()-1] != n ) { duke@435: for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { duke@435: Node *foi = n->fast_out(i); duke@435: if( foi->is_Proj() ) duke@435: _scheduled.push(foi); duke@435: } duke@435: } duke@435: duke@435: // Put the instruction in the schedule list duke@435: _scheduled.push(n); duke@435: } duke@435: duke@435: #ifndef PRODUCT duke@435: if (_cfg->C->trace_opto_output()) duke@435: dump_available(); duke@435: #endif duke@435: duke@435: // Walk all the definitions, decrementing use counts, and duke@435: // if a definition has a 0 use count, place it in the available list. duke@435: DecrementUseCounts(n,bb); duke@435: } duke@435: duke@435: //------------------------------ComputeUseCount-------------------------------- duke@435: // This method sets the use count within a basic block. We will ignore all duke@435: // uses outside the current basic block. As we are doing a backwards walk, duke@435: // any node we reach that has a use count of 0 may be scheduled. This also duke@435: // avoids the problem of cyclic references from phi nodes, as long as phi duke@435: // nodes are at the front of the basic block. This method also initializes duke@435: // the available list to the set of instructions that have no uses within this duke@435: // basic block. duke@435: void Scheduling::ComputeUseCount(const Block *bb) { duke@435: #ifndef PRODUCT duke@435: if (_cfg->C->trace_opto_output()) duke@435: tty->print("# -> ComputeUseCount\n"); duke@435: #endif duke@435: duke@435: // Clear the list of available and scheduled instructions, just in case duke@435: _available.clear(); duke@435: _scheduled.clear(); duke@435: duke@435: // No delay slot specified duke@435: _unconditional_delay_slot = NULL; duke@435: duke@435: #ifdef ASSERT duke@435: for( uint i=0; i < bb->_nodes.size(); i++ ) duke@435: assert( _uses[bb->_nodes[i]->_idx] == 0, "_use array not clean" ); duke@435: #endif duke@435: duke@435: // Force the _uses count to never go to zero for unscheduable pieces duke@435: // of the block duke@435: for( uint k = 0; k < _bb_start; k++ ) duke@435: _uses[bb->_nodes[k]->_idx] = 1; duke@435: for( uint l = _bb_end; l < bb->_nodes.size(); l++ ) duke@435: _uses[bb->_nodes[l]->_idx] = 1; duke@435: duke@435: // Iterate backwards over the instructions in the block. Don't count the duke@435: // branch projections at end or the block header instructions. duke@435: for( uint j = _bb_end-1; j >= _bb_start; j-- ) { duke@435: Node *n = bb->_nodes[j]; duke@435: if( n->is_Proj() ) continue; // Projections handled another way duke@435: duke@435: // Account for all uses duke@435: for ( uint k = 0; k < n->len(); k++ ) { duke@435: Node *inp = n->in(k); duke@435: if (!inp) continue; duke@435: assert(inp != n, "no cycles allowed" ); duke@435: if( _bbs[inp->_idx] == bb ) { // Block-local use? duke@435: if( inp->is_Proj() ) // Skip through Proj's duke@435: inp = inp->in(0); duke@435: ++_uses[inp->_idx]; // Count 1 block-local use duke@435: } duke@435: } duke@435: duke@435: // If this instruction has a 0 use count, then it is available duke@435: if (!_uses[n->_idx]) { duke@435: _current_latency[n->_idx] = _bundle_cycle_number; duke@435: AddNodeToAvailableList(n); duke@435: } duke@435: duke@435: #ifndef PRODUCT duke@435: if (_cfg->C->trace_opto_output()) { duke@435: tty->print("# uses: %3d: ", _uses[n->_idx]); duke@435: n->dump(); duke@435: } duke@435: #endif duke@435: } duke@435: duke@435: #ifndef PRODUCT duke@435: if (_cfg->C->trace_opto_output()) duke@435: tty->print("# <- ComputeUseCount\n"); duke@435: #endif duke@435: } duke@435: duke@435: // This routine performs scheduling on each basic block in reverse order, duke@435: // using instruction latencies and taking into account function unit duke@435: // availability. duke@435: void Scheduling::DoScheduling() { duke@435: #ifndef PRODUCT duke@435: if (_cfg->C->trace_opto_output()) duke@435: tty->print("# -> DoScheduling\n"); duke@435: #endif duke@435: duke@435: Block *succ_bb = NULL; duke@435: Block *bb; duke@435: duke@435: // Walk over all the basic blocks in reverse order duke@435: for( int i=_cfg->_num_blocks-1; i >= 0; succ_bb = bb, i-- ) { duke@435: bb = _cfg->_blocks[i]; duke@435: duke@435: #ifndef PRODUCT duke@435: if (_cfg->C->trace_opto_output()) { duke@435: tty->print("# Schedule BB#%03d (initial)\n", i); duke@435: for (uint j = 0; j < bb->_nodes.size(); j++) duke@435: bb->_nodes[j]->dump(); duke@435: } duke@435: #endif duke@435: duke@435: // On the head node, skip processing duke@435: if( bb == _cfg->_broot ) duke@435: continue; duke@435: duke@435: // Skip empty, connector blocks duke@435: if (bb->is_connector()) duke@435: continue; duke@435: duke@435: // If the following block is not the sole successor of duke@435: // this one, then reset the pipeline information duke@435: if (bb->_num_succs != 1 || bb->non_connector_successor(0) != succ_bb) { duke@435: #ifndef PRODUCT duke@435: if (_cfg->C->trace_opto_output()) { duke@435: tty->print("*** bundle start of next BB, node %d, for %d instructions\n", duke@435: _next_node->_idx, _bundle_instr_count); duke@435: } duke@435: #endif duke@435: step_and_clear(); duke@435: } duke@435: duke@435: // Leave untouched the starting instruction, any Phis, a CreateEx node duke@435: // or Top. bb->_nodes[_bb_start] is the first schedulable instruction. duke@435: _bb_end = bb->_nodes.size()-1; duke@435: for( _bb_start=1; _bb_start <= _bb_end; _bb_start++ ) { duke@435: Node *n = bb->_nodes[_bb_start]; duke@435: // Things not matched, like Phinodes and ProjNodes don't get scheduled. duke@435: // Also, MachIdealNodes do not get scheduled duke@435: if( !n->is_Mach() ) continue; // Skip non-machine nodes duke@435: MachNode *mach = n->as_Mach(); duke@435: int iop = mach->ideal_Opcode(); duke@435: if( iop == Op_CreateEx ) continue; // CreateEx is pinned duke@435: if( iop == Op_Con ) continue; // Do not schedule Top duke@435: if( iop == Op_Node && // Do not schedule PhiNodes, ProjNodes duke@435: mach->pipeline() == MachNode::pipeline_class() && duke@435: !n->is_SpillCopy() ) // Breakpoints, Prolog, etc duke@435: continue; duke@435: break; // Funny loop structure to be sure... duke@435: } duke@435: // Compute last "interesting" instruction in block - last instruction we duke@435: // might schedule. _bb_end points just after last schedulable inst. We duke@435: // normally schedule conditional branches (despite them being forced last duke@435: // in the block), because they have delay slots we can fill. Calls all duke@435: // have their delay slots filled in the template expansions, so we don't duke@435: // bother scheduling them. duke@435: Node *last = bb->_nodes[_bb_end]; duke@435: if( last->is_Catch() || duke@435: (last->is_Mach() && last->as_Mach()->ideal_Opcode() == Op_Halt) ) { duke@435: // There must be a prior call. Skip it. duke@435: while( !bb->_nodes[--_bb_end]->is_Call() ) { duke@435: assert( bb->_nodes[_bb_end]->is_Proj(), "skipping projections after expected call" ); duke@435: } duke@435: } else if( last->is_MachNullCheck() ) { duke@435: // Backup so the last null-checked memory instruction is duke@435: // outside the schedulable range. Skip over the nullcheck, duke@435: // projection, and the memory nodes. duke@435: Node *mem = last->in(1); duke@435: do { duke@435: _bb_end--; duke@435: } while (mem != bb->_nodes[_bb_end]); duke@435: } else { duke@435: // Set _bb_end to point after last schedulable inst. duke@435: _bb_end++; duke@435: } duke@435: duke@435: assert( _bb_start <= _bb_end, "inverted block ends" ); duke@435: duke@435: // Compute the register antidependencies for the basic block duke@435: ComputeRegisterAntidependencies(bb); duke@435: if (_cfg->C->failing()) return; // too many D-U pinch points duke@435: duke@435: // Compute intra-bb latencies for the nodes duke@435: ComputeLocalLatenciesForward(bb); duke@435: duke@435: // Compute the usage within the block, and set the list of all nodes duke@435: // in the block that have no uses within the block. duke@435: ComputeUseCount(bb); duke@435: duke@435: // Schedule the remaining instructions in the block duke@435: while ( _available.size() > 0 ) { duke@435: Node *n = ChooseNodeToBundle(); duke@435: AddNodeToBundle(n,bb); duke@435: } duke@435: duke@435: assert( _scheduled.size() == _bb_end - _bb_start, "wrong number of instructions" ); duke@435: #ifdef ASSERT duke@435: for( uint l = _bb_start; l < _bb_end; l++ ) { duke@435: Node *n = bb->_nodes[l]; duke@435: uint m; duke@435: for( m = 0; m < _bb_end-_bb_start; m++ ) duke@435: if( _scheduled[m] == n ) duke@435: break; duke@435: assert( m < _bb_end-_bb_start, "instruction missing in schedule" ); duke@435: } duke@435: #endif duke@435: duke@435: // Now copy the instructions (in reverse order) back to the block duke@435: for ( uint k = _bb_start; k < _bb_end; k++ ) duke@435: bb->_nodes.map(k, _scheduled[_bb_end-k-1]); duke@435: duke@435: #ifndef PRODUCT duke@435: if (_cfg->C->trace_opto_output()) { duke@435: tty->print("# Schedule BB#%03d (final)\n", i); duke@435: uint current = 0; duke@435: for (uint j = 0; j < bb->_nodes.size(); j++) { duke@435: Node *n = bb->_nodes[j]; duke@435: if( valid_bundle_info(n) ) { duke@435: Bundle *bundle = node_bundling(n); duke@435: if (bundle->instr_count() > 0 || bundle->flags() > 0) { duke@435: tty->print("*** Bundle: "); duke@435: bundle->dump(); duke@435: } duke@435: n->dump(); duke@435: } duke@435: } duke@435: } duke@435: #endif duke@435: #ifdef ASSERT duke@435: verify_good_schedule(bb,"after block local scheduling"); duke@435: #endif duke@435: } duke@435: duke@435: #ifndef PRODUCT duke@435: if (_cfg->C->trace_opto_output()) duke@435: tty->print("# <- DoScheduling\n"); duke@435: #endif duke@435: duke@435: // Record final node-bundling array location duke@435: _regalloc->C->set_node_bundling_base(_node_bundling_base); duke@435: duke@435: } // end DoScheduling duke@435: duke@435: //------------------------------verify_good_schedule--------------------------- duke@435: // Verify that no live-range used in the block is killed in the block by a duke@435: // wrong DEF. This doesn't verify live-ranges that span blocks. duke@435: duke@435: // Check for edge existence. Used to avoid adding redundant precedence edges. duke@435: static bool edge_from_to( Node *from, Node *to ) { duke@435: for( uint i=0; ilen(); i++ ) duke@435: if( from->in(i) == to ) duke@435: return true; duke@435: return false; duke@435: } duke@435: duke@435: #ifdef ASSERT duke@435: //------------------------------verify_do_def---------------------------------- duke@435: void Scheduling::verify_do_def( Node *n, OptoReg::Name def, const char *msg ) { duke@435: // Check for bad kills duke@435: if( OptoReg::is_valid(def) ) { // Ignore stores & control flow duke@435: Node *prior_use = _reg_node[def]; duke@435: if( prior_use && !edge_from_to(prior_use,n) ) { duke@435: tty->print("%s = ",OptoReg::as_VMReg(def)->name()); duke@435: n->dump(); duke@435: tty->print_cr("..."); duke@435: prior_use->dump(); duke@435: assert_msg(edge_from_to(prior_use,n),msg); duke@435: } duke@435: _reg_node.map(def,NULL); // Kill live USEs duke@435: } duke@435: } duke@435: duke@435: //------------------------------verify_good_schedule--------------------------- duke@435: void Scheduling::verify_good_schedule( Block *b, const char *msg ) { duke@435: duke@435: // Zap to something reasonable for the verify code duke@435: _reg_node.clear(); duke@435: duke@435: // Walk over the block backwards. Check to make sure each DEF doesn't duke@435: // kill a live value (other than the one it's supposed to). Add each duke@435: // USE to the live set. duke@435: for( uint i = b->_nodes.size()-1; i >= _bb_start; i-- ) { duke@435: Node *n = b->_nodes[i]; duke@435: int n_op = n->Opcode(); duke@435: if( n_op == Op_MachProj && n->ideal_reg() == MachProjNode::fat_proj ) { duke@435: // Fat-proj kills a slew of registers duke@435: RegMask rm = n->out_RegMask();// Make local copy duke@435: while( rm.is_NotEmpty() ) { duke@435: OptoReg::Name kill = rm.find_first_elem(); duke@435: rm.Remove(kill); duke@435: verify_do_def( n, kill, msg ); duke@435: } duke@435: } else if( n_op != Op_Node ) { // Avoid brand new antidependence nodes duke@435: // Get DEF'd registers the normal way duke@435: verify_do_def( n, _regalloc->get_reg_first(n), msg ); duke@435: verify_do_def( n, _regalloc->get_reg_second(n), msg ); duke@435: } duke@435: duke@435: // Now make all USEs live duke@435: for( uint i=1; ireq(); i++ ) { duke@435: Node *def = n->in(i); duke@435: assert(def != 0, "input edge required"); duke@435: OptoReg::Name reg_lo = _regalloc->get_reg_first(def); duke@435: OptoReg::Name reg_hi = _regalloc->get_reg_second(def); duke@435: if( OptoReg::is_valid(reg_lo) ) { duke@435: assert_msg(!_reg_node[reg_lo] || edge_from_to(_reg_node[reg_lo],def), msg ); duke@435: _reg_node.map(reg_lo,n); duke@435: } duke@435: if( OptoReg::is_valid(reg_hi) ) { duke@435: assert_msg(!_reg_node[reg_hi] || edge_from_to(_reg_node[reg_hi],def), msg ); duke@435: _reg_node.map(reg_hi,n); duke@435: } duke@435: } duke@435: duke@435: } duke@435: duke@435: // Zap to something reasonable for the Antidependence code duke@435: _reg_node.clear(); duke@435: } duke@435: #endif duke@435: duke@435: // Conditionally add precedence edges. Avoid putting edges on Projs. duke@435: static void add_prec_edge_from_to( Node *from, Node *to ) { duke@435: if( from->is_Proj() ) { // Put precedence edge on Proj's input duke@435: assert( from->req() == 1 && (from->len() == 1 || from->in(1)==0), "no precedence edges on projections" ); duke@435: from = from->in(0); duke@435: } duke@435: if( from != to && // No cycles (for things like LD L0,[L0+4] ) duke@435: !edge_from_to( from, to ) ) // Avoid duplicate edge duke@435: from->add_prec(to); duke@435: } duke@435: duke@435: //------------------------------anti_do_def------------------------------------ duke@435: void Scheduling::anti_do_def( Block *b, Node *def, OptoReg::Name def_reg, int is_def ) { duke@435: if( !OptoReg::is_valid(def_reg) ) // Ignore stores & control flow duke@435: return; duke@435: duke@435: Node *pinch = _reg_node[def_reg]; // Get pinch point duke@435: if( !pinch || _bbs[pinch->_idx] != b || // No pinch-point yet? duke@435: is_def ) { // Check for a true def (not a kill) duke@435: _reg_node.map(def_reg,def); // Record def/kill as the optimistic pinch-point duke@435: return; duke@435: } duke@435: duke@435: Node *kill = def; // Rename 'def' to more descriptive 'kill' duke@435: debug_only( def = (Node*)0xdeadbeef; ) duke@435: duke@435: // After some number of kills there _may_ be a later def duke@435: Node *later_def = NULL; duke@435: duke@435: // Finding a kill requires a real pinch-point. duke@435: // Check for not already having a pinch-point. duke@435: // Pinch points are Op_Node's. duke@435: if( pinch->Opcode() != Op_Node ) { // Or later-def/kill as pinch-point? duke@435: later_def = pinch; // Must be def/kill as optimistic pinch-point duke@435: if ( _pinch_free_list.size() > 0) { duke@435: pinch = _pinch_free_list.pop(); duke@435: } else { duke@435: pinch = new (_cfg->C, 1) Node(1); // Pinch point to-be duke@435: } duke@435: if (pinch->_idx >= _regalloc->node_regs_max_index()) { duke@435: _cfg->C->record_method_not_compilable("too many D-U pinch points"); duke@435: return; duke@435: } duke@435: _bbs.map(pinch->_idx,b); // Pretend it's valid in this block (lazy init) duke@435: _reg_node.map(def_reg,pinch); // Record pinch-point duke@435: //_regalloc->set_bad(pinch->_idx); // Already initialized this way. duke@435: if( later_def->outcnt() == 0 || later_def->ideal_reg() == MachProjNode::fat_proj ) { // Distinguish def from kill duke@435: pinch->init_req(0, _cfg->C->top()); // set not NULL for the next call duke@435: add_prec_edge_from_to(later_def,pinch); // Add edge from kill to pinch duke@435: later_def = NULL; // and no later def duke@435: } duke@435: pinch->set_req(0,later_def); // Hook later def so we can find it duke@435: } else { // Else have valid pinch point duke@435: if( pinch->in(0) ) // If there is a later-def duke@435: later_def = pinch->in(0); // Get it duke@435: } duke@435: duke@435: // Add output-dependence edge from later def to kill duke@435: if( later_def ) // If there is some original def duke@435: add_prec_edge_from_to(later_def,kill); // Add edge from def to kill duke@435: duke@435: // See if current kill is also a use, and so is forced to be the pinch-point. duke@435: if( pinch->Opcode() == Op_Node ) { duke@435: Node *uses = kill->is_Proj() ? kill->in(0) : kill; duke@435: for( uint i=1; ireq(); i++ ) { duke@435: if( _regalloc->get_reg_first(uses->in(i)) == def_reg || duke@435: _regalloc->get_reg_second(uses->in(i)) == def_reg ) { duke@435: // Yes, found a use/kill pinch-point duke@435: pinch->set_req(0,NULL); // duke@435: pinch->replace_by(kill); // Move anti-dep edges up duke@435: pinch = kill; duke@435: _reg_node.map(def_reg,pinch); duke@435: return; duke@435: } duke@435: } duke@435: } duke@435: duke@435: // Add edge from kill to pinch-point duke@435: add_prec_edge_from_to(kill,pinch); duke@435: } duke@435: duke@435: //------------------------------anti_do_use------------------------------------ duke@435: void Scheduling::anti_do_use( Block *b, Node *use, OptoReg::Name use_reg ) { duke@435: if( !OptoReg::is_valid(use_reg) ) // Ignore stores & control flow duke@435: return; duke@435: Node *pinch = _reg_node[use_reg]; // Get pinch point duke@435: // Check for no later def_reg/kill in block duke@435: if( pinch && _bbs[pinch->_idx] == b && duke@435: // Use has to be block-local as well duke@435: _bbs[use->_idx] == b ) { duke@435: if( pinch->Opcode() == Op_Node && // Real pinch-point (not optimistic?) duke@435: pinch->req() == 1 ) { // pinch not yet in block? duke@435: pinch->del_req(0); // yank pointer to later-def, also set flag duke@435: // Insert the pinch-point in the block just after the last use duke@435: b->_nodes.insert(b->find_node(use)+1,pinch); duke@435: _bb_end++; // Increase size scheduled region in block duke@435: } duke@435: duke@435: add_prec_edge_from_to(pinch,use); duke@435: } duke@435: } duke@435: duke@435: //------------------------------ComputeRegisterAntidependences----------------- duke@435: // We insert antidependences between the reads and following write of duke@435: // allocated registers to prevent illegal code motion. Hopefully, the duke@435: // number of added references should be fairly small, especially as we duke@435: // are only adding references within the current basic block. duke@435: void Scheduling::ComputeRegisterAntidependencies(Block *b) { duke@435: duke@435: #ifdef ASSERT duke@435: verify_good_schedule(b,"before block local scheduling"); duke@435: #endif duke@435: duke@435: // A valid schedule, for each register independently, is an endless cycle duke@435: // of: a def, then some uses (connected to the def by true dependencies), duke@435: // then some kills (defs with no uses), finally the cycle repeats with a new duke@435: // def. The uses are allowed to float relative to each other, as are the duke@435: // kills. No use is allowed to slide past a kill (or def). This requires duke@435: // antidependencies between all uses of a single def and all kills that duke@435: // follow, up to the next def. More edges are redundant, because later defs duke@435: // & kills are already serialized with true or antidependencies. To keep duke@435: // the edge count down, we add a 'pinch point' node if there's more than duke@435: // one use or more than one kill/def. duke@435: duke@435: // We add dependencies in one bottom-up pass. duke@435: duke@435: // For each instruction we handle it's DEFs/KILLs, then it's USEs. duke@435: duke@435: // For each DEF/KILL, we check to see if there's a prior DEF/KILL for this duke@435: // register. If not, we record the DEF/KILL in _reg_node, the duke@435: // register-to-def mapping. If there is a prior DEF/KILL, we insert a duke@435: // "pinch point", a new Node that's in the graph but not in the block. duke@435: // We put edges from the prior and current DEF/KILLs to the pinch point. duke@435: // We put the pinch point in _reg_node. If there's already a pinch point duke@435: // we merely add an edge from the current DEF/KILL to the pinch point. duke@435: duke@435: // After doing the DEF/KILLs, we handle USEs. For each used register, we duke@435: // put an edge from the pinch point to the USE. duke@435: duke@435: // To be expedient, the _reg_node array is pre-allocated for the whole duke@435: // compilation. _reg_node is lazily initialized; it either contains a NULL, duke@435: // or a valid def/kill/pinch-point, or a leftover node from some prior duke@435: // block. Leftover node from some prior block is treated like a NULL (no duke@435: // prior def, so no anti-dependence needed). Valid def is distinguished by duke@435: // it being in the current block. duke@435: bool fat_proj_seen = false; duke@435: uint last_safept = _bb_end-1; duke@435: Node* end_node = (_bb_end-1 >= _bb_start) ? b->_nodes[last_safept] : NULL; duke@435: Node* last_safept_node = end_node; duke@435: for( uint i = _bb_end-1; i >= _bb_start; i-- ) { duke@435: Node *n = b->_nodes[i]; duke@435: int is_def = n->outcnt(); // def if some uses prior to adding precedence edges duke@435: if( n->Opcode() == Op_MachProj && n->ideal_reg() == MachProjNode::fat_proj ) { duke@435: // Fat-proj kills a slew of registers duke@435: // This can add edges to 'n' and obscure whether or not it was a def, duke@435: // hence the is_def flag. duke@435: fat_proj_seen = true; duke@435: RegMask rm = n->out_RegMask();// Make local copy duke@435: while( rm.is_NotEmpty() ) { duke@435: OptoReg::Name kill = rm.find_first_elem(); duke@435: rm.Remove(kill); duke@435: anti_do_def( b, n, kill, is_def ); duke@435: } duke@435: } else { duke@435: // Get DEF'd registers the normal way duke@435: anti_do_def( b, n, _regalloc->get_reg_first(n), is_def ); duke@435: anti_do_def( b, n, _regalloc->get_reg_second(n), is_def ); duke@435: } duke@435: duke@435: // Check each register used by this instruction for a following DEF/KILL duke@435: // that must occur afterward and requires an anti-dependence edge. duke@435: for( uint j=0; jreq(); j++ ) { duke@435: Node *def = n->in(j); duke@435: if( def ) { duke@435: assert( def->Opcode() != Op_MachProj || def->ideal_reg() != MachProjNode::fat_proj, "" ); duke@435: anti_do_use( b, n, _regalloc->get_reg_first(def) ); duke@435: anti_do_use( b, n, _regalloc->get_reg_second(def) ); duke@435: } duke@435: } duke@435: // Do not allow defs of new derived values to float above GC duke@435: // points unless the base is definitely available at the GC point. duke@435: duke@435: Node *m = b->_nodes[i]; duke@435: duke@435: // Add precedence edge from following safepoint to use of derived pointer duke@435: if( last_safept_node != end_node && duke@435: m != last_safept_node) { duke@435: for (uint k = 1; k < m->req(); k++) { duke@435: const Type *t = m->in(k)->bottom_type(); duke@435: if( t->isa_oop_ptr() && duke@435: t->is_ptr()->offset() != 0 ) { duke@435: last_safept_node->add_prec( m ); duke@435: break; duke@435: } duke@435: } duke@435: } duke@435: duke@435: if( n->jvms() ) { // Precedence edge from derived to safept duke@435: // Check if last_safept_node was moved by pinch-point insertion in anti_do_use() duke@435: if( b->_nodes[last_safept] != last_safept_node ) { duke@435: last_safept = b->find_node(last_safept_node); duke@435: } duke@435: for( uint j=last_safept; j > i; j-- ) { duke@435: Node *mach = b->_nodes[j]; duke@435: if( mach->is_Mach() && mach->as_Mach()->ideal_Opcode() == Op_AddP ) duke@435: mach->add_prec( n ); duke@435: } duke@435: last_safept = i; duke@435: last_safept_node = m; duke@435: } duke@435: } duke@435: duke@435: if (fat_proj_seen) { duke@435: // Garbage collect pinch nodes that were not consumed. duke@435: // They are usually created by a fat kill MachProj for a call. duke@435: garbage_collect_pinch_nodes(); duke@435: } duke@435: } duke@435: duke@435: //------------------------------garbage_collect_pinch_nodes------------------------------- duke@435: duke@435: // Garbage collect pinch nodes for reuse by other blocks. duke@435: // duke@435: // The block scheduler's insertion of anti-dependence duke@435: // edges creates many pinch nodes when the block contains duke@435: // 2 or more Calls. A pinch node is used to prevent a duke@435: // combinatorial explosion of edges. If a set of kills for a duke@435: // register is anti-dependent on a set of uses (or defs), rather duke@435: // than adding an edge in the graph between each pair of kill duke@435: // and use (or def), a pinch is inserted between them: duke@435: // duke@435: // use1 use2 use3 duke@435: // \ | / duke@435: // \ | / duke@435: // pinch duke@435: // / | \ duke@435: // / | \ duke@435: // kill1 kill2 kill3 duke@435: // duke@435: // One pinch node is created per register killed when duke@435: // the second call is encountered during a backwards pass duke@435: // over the block. Most of these pinch nodes are never duke@435: // wired into the graph because the register is never duke@435: // used or def'ed in the block. duke@435: // duke@435: void Scheduling::garbage_collect_pinch_nodes() { duke@435: #ifndef PRODUCT duke@435: if (_cfg->C->trace_opto_output()) tty->print("Reclaimed pinch nodes:"); duke@435: #endif duke@435: int trace_cnt = 0; duke@435: for (uint k = 0; k < _reg_node.Size(); k++) { duke@435: Node* pinch = _reg_node[k]; duke@435: if (pinch != NULL && pinch->Opcode() == Op_Node && duke@435: // no predecence input edges duke@435: (pinch->req() == pinch->len() || pinch->in(pinch->req()) == NULL) ) { duke@435: cleanup_pinch(pinch); duke@435: _pinch_free_list.push(pinch); duke@435: _reg_node.map(k, NULL); duke@435: #ifndef PRODUCT duke@435: if (_cfg->C->trace_opto_output()) { duke@435: trace_cnt++; duke@435: if (trace_cnt > 40) { duke@435: tty->print("\n"); duke@435: trace_cnt = 0; duke@435: } duke@435: tty->print(" %d", pinch->_idx); duke@435: } duke@435: #endif duke@435: } duke@435: } duke@435: #ifndef PRODUCT duke@435: if (_cfg->C->trace_opto_output()) tty->print("\n"); duke@435: #endif duke@435: } duke@435: duke@435: // Clean up a pinch node for reuse. duke@435: void Scheduling::cleanup_pinch( Node *pinch ) { duke@435: assert (pinch && pinch->Opcode() == Op_Node && pinch->req() == 1, "just checking"); duke@435: duke@435: for (DUIterator_Last imin, i = pinch->last_outs(imin); i >= imin; ) { duke@435: Node* use = pinch->last_out(i); duke@435: uint uses_found = 0; duke@435: for (uint j = use->req(); j < use->len(); j++) { duke@435: if (use->in(j) == pinch) { duke@435: use->rm_prec(j); duke@435: uses_found++; duke@435: } duke@435: } duke@435: assert(uses_found > 0, "must be a precedence edge"); duke@435: i -= uses_found; // we deleted 1 or more copies of this edge duke@435: } duke@435: // May have a later_def entry duke@435: pinch->set_req(0, NULL); duke@435: } duke@435: duke@435: //------------------------------print_statistics------------------------------- duke@435: #ifndef PRODUCT duke@435: duke@435: void Scheduling::dump_available() const { duke@435: tty->print("#Availist "); duke@435: for (uint i = 0; i < _available.size(); i++) duke@435: tty->print(" N%d/l%d", _available[i]->_idx,_current_latency[_available[i]->_idx]); duke@435: tty->cr(); duke@435: } duke@435: duke@435: // Print Scheduling Statistics duke@435: void Scheduling::print_statistics() { duke@435: // Print the size added by nops for bundling duke@435: tty->print("Nops added %d bytes to total of %d bytes", duke@435: _total_nop_size, _total_method_size); duke@435: if (_total_method_size > 0) duke@435: tty->print(", for %.2f%%", duke@435: ((double)_total_nop_size) / ((double) _total_method_size) * 100.0); duke@435: tty->print("\n"); duke@435: duke@435: // Print the number of branch shadows filled duke@435: if (Pipeline::_branch_has_delay_slot) { duke@435: tty->print("Of %d branches, %d had unconditional delay slots filled", duke@435: _total_branches, _total_unconditional_delays); duke@435: if (_total_branches > 0) duke@435: tty->print(", for %.2f%%", duke@435: ((double)_total_unconditional_delays) / ((double)_total_branches) * 100.0); duke@435: tty->print("\n"); duke@435: } duke@435: duke@435: uint total_instructions = 0, total_bundles = 0; duke@435: duke@435: for (uint i = 1; i <= Pipeline::_max_instrs_per_cycle; i++) { duke@435: uint bundle_count = _total_instructions_per_bundle[i]; duke@435: total_instructions += bundle_count * i; duke@435: total_bundles += bundle_count; duke@435: } duke@435: duke@435: if (total_bundles > 0) duke@435: tty->print("Average ILP (excluding nops) is %.2f\n", duke@435: ((double)total_instructions) / ((double)total_bundles)); duke@435: } duke@435: #endif