duke@435: /* sla@5237: * Copyright (c) 1997, 2013, Oracle and/or its affiliates. All rights reserved. duke@435: * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. duke@435: * duke@435: * This code is free software; you can redistribute it and/or modify it duke@435: * under the terms of the GNU General Public License version 2 only, as duke@435: * published by the Free Software Foundation. duke@435: * duke@435: * This code is distributed in the hope that it will be useful, but WITHOUT duke@435: * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or duke@435: * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License duke@435: * version 2 for more details (a copy is included in the LICENSE file that duke@435: * accompanied this code). duke@435: * duke@435: * You should have received a copy of the GNU General Public License version duke@435: * 2 along with this work; if not, write to the Free Software Foundation, duke@435: * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. duke@435: * trims@1907: * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA trims@1907: * or visit www.oracle.com if you need additional information or have any trims@1907: * questions. duke@435: * duke@435: */ duke@435: stefank@2314: #include "precompiled.hpp" twisti@4318: #include "asm/macroAssembler.hpp" twisti@4318: #include "asm/macroAssembler.inline.hpp" stefank@2314: #include "classfile/systemDictionary.hpp" stefank@2314: #include "code/exceptionHandlerTable.hpp" stefank@2314: #include "code/nmethod.hpp" stefank@2314: #include "compiler/compileLog.hpp" twisti@4318: #include "compiler/disassembler.hpp" stefank@2314: #include "compiler/oopMap.hpp" stefank@2314: #include "opto/addnode.hpp" stefank@2314: #include "opto/block.hpp" stefank@2314: #include "opto/c2compiler.hpp" stefank@2314: #include "opto/callGenerator.hpp" stefank@2314: #include "opto/callnode.hpp" stefank@2314: #include "opto/cfgnode.hpp" stefank@2314: #include "opto/chaitin.hpp" stefank@2314: #include "opto/compile.hpp" stefank@2314: #include "opto/connode.hpp" stefank@2314: #include "opto/divnode.hpp" stefank@2314: #include "opto/escape.hpp" stefank@2314: #include "opto/idealGraphPrinter.hpp" stefank@2314: #include "opto/loopnode.hpp" stefank@2314: #include "opto/machnode.hpp" stefank@2314: #include "opto/macro.hpp" stefank@2314: #include "opto/matcher.hpp" stefank@2314: #include "opto/memnode.hpp" stefank@2314: #include "opto/mulnode.hpp" stefank@2314: #include "opto/node.hpp" stefank@2314: #include "opto/opcodes.hpp" stefank@2314: #include "opto/output.hpp" stefank@2314: #include "opto/parse.hpp" stefank@2314: #include "opto/phaseX.hpp" stefank@2314: #include "opto/rootnode.hpp" stefank@2314: #include "opto/runtime.hpp" stefank@2314: #include "opto/stringopts.hpp" stefank@2314: #include "opto/type.hpp" stefank@2314: #include "opto/vectornode.hpp" stefank@2314: #include "runtime/arguments.hpp" stefank@2314: #include "runtime/signature.hpp" stefank@2314: #include "runtime/stubRoutines.hpp" stefank@2314: #include "runtime/timer.hpp" sla@5237: #include "trace/tracing.hpp" stefank@2314: #include "utilities/copy.hpp" stefank@2314: #ifdef TARGET_ARCH_MODEL_x86_32 stefank@2314: # include "adfiles/ad_x86_32.hpp" stefank@2314: #endif stefank@2314: #ifdef TARGET_ARCH_MODEL_x86_64 stefank@2314: # include "adfiles/ad_x86_64.hpp" stefank@2314: #endif stefank@2314: #ifdef TARGET_ARCH_MODEL_sparc stefank@2314: # include "adfiles/ad_sparc.hpp" stefank@2314: #endif stefank@2314: #ifdef TARGET_ARCH_MODEL_zero stefank@2314: # include "adfiles/ad_zero.hpp" stefank@2314: #endif bobv@2508: #ifdef TARGET_ARCH_MODEL_arm bobv@2508: # include "adfiles/ad_arm.hpp" bobv@2508: #endif bobv@2508: #ifdef TARGET_ARCH_MODEL_ppc bobv@2508: # include "adfiles/ad_ppc.hpp" bobv@2508: #endif duke@435: twisti@2350: twisti@2350: // -------------------- Compile::mach_constant_base_node ----------------------- twisti@2350: // Constant table base node singleton. twisti@2350: MachConstantBaseNode* Compile::mach_constant_base_node() { twisti@2350: if (_mach_constant_base_node == NULL) { twisti@2350: _mach_constant_base_node = new (C) MachConstantBaseNode(); twisti@2350: _mach_constant_base_node->add_req(C->root()); twisti@2350: } twisti@2350: return _mach_constant_base_node; twisti@2350: } twisti@2350: twisti@2350: duke@435: /// Support for intrinsics. duke@435: duke@435: // Return the index at which m must be inserted (or already exists). duke@435: // The sort order is by the address of the ciMethod, with is_virtual as minor key. duke@435: int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) { duke@435: #ifdef ASSERT duke@435: for (int i = 1; i < _intrinsics->length(); i++) { duke@435: CallGenerator* cg1 = _intrinsics->at(i-1); duke@435: CallGenerator* cg2 = _intrinsics->at(i); duke@435: assert(cg1->method() != cg2->method() duke@435: ? cg1->method() < cg2->method() duke@435: : cg1->is_virtual() < cg2->is_virtual(), duke@435: "compiler intrinsics list must stay sorted"); duke@435: } duke@435: #endif duke@435: // Binary search sorted list, in decreasing intervals [lo, hi]. duke@435: int lo = 0, hi = _intrinsics->length()-1; duke@435: while (lo <= hi) { duke@435: int mid = (uint)(hi + lo) / 2; duke@435: ciMethod* mid_m = _intrinsics->at(mid)->method(); duke@435: if (m < mid_m) { duke@435: hi = mid-1; duke@435: } else if (m > mid_m) { duke@435: lo = mid+1; duke@435: } else { duke@435: // look at minor sort key duke@435: bool mid_virt = _intrinsics->at(mid)->is_virtual(); duke@435: if (is_virtual < mid_virt) { duke@435: hi = mid-1; duke@435: } else if (is_virtual > mid_virt) { duke@435: lo = mid+1; duke@435: } else { duke@435: return mid; // exact match duke@435: } duke@435: } duke@435: } duke@435: return lo; // inexact match duke@435: } duke@435: duke@435: void Compile::register_intrinsic(CallGenerator* cg) { duke@435: if (_intrinsics == NULL) { roland@4409: _intrinsics = new (comp_arena())GrowableArray(comp_arena(), 60, 0, NULL); duke@435: } duke@435: // This code is stolen from ciObjectFactory::insert. duke@435: // Really, GrowableArray should have methods for duke@435: // insert_at, remove_at, and binary_search. duke@435: int len = _intrinsics->length(); duke@435: int index = intrinsic_insertion_index(cg->method(), cg->is_virtual()); duke@435: if (index == len) { duke@435: _intrinsics->append(cg); duke@435: } else { duke@435: #ifdef ASSERT duke@435: CallGenerator* oldcg = _intrinsics->at(index); duke@435: assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice"); duke@435: #endif duke@435: _intrinsics->append(_intrinsics->at(len-1)); duke@435: int pos; duke@435: for (pos = len-2; pos >= index; pos--) { duke@435: _intrinsics->at_put(pos+1,_intrinsics->at(pos)); duke@435: } duke@435: _intrinsics->at_put(index, cg); duke@435: } duke@435: assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked"); duke@435: } duke@435: duke@435: CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) { duke@435: assert(m->is_loaded(), "don't try this on unloaded methods"); duke@435: if (_intrinsics != NULL) { duke@435: int index = intrinsic_insertion_index(m, is_virtual); duke@435: if (index < _intrinsics->length() duke@435: && _intrinsics->at(index)->method() == m duke@435: && _intrinsics->at(index)->is_virtual() == is_virtual) { duke@435: return _intrinsics->at(index); duke@435: } duke@435: } duke@435: // Lazily create intrinsics for intrinsic IDs well-known in the runtime. jrose@1291: if (m->intrinsic_id() != vmIntrinsics::_none && jrose@1291: m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) { duke@435: CallGenerator* cg = make_vm_intrinsic(m, is_virtual); duke@435: if (cg != NULL) { duke@435: // Save it for next time: duke@435: register_intrinsic(cg); duke@435: return cg; duke@435: } else { duke@435: gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled); duke@435: } duke@435: } duke@435: return NULL; duke@435: } duke@435: duke@435: // Compile:: register_library_intrinsics and make_vm_intrinsic are defined duke@435: // in library_call.cpp. duke@435: duke@435: duke@435: #ifndef PRODUCT duke@435: // statistics gathering... duke@435: duke@435: juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0}; duke@435: jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0}; duke@435: duke@435: bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) { duke@435: assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob"); duke@435: int oflags = _intrinsic_hist_flags[id]; duke@435: assert(flags != 0, "what happened?"); duke@435: if (is_virtual) { duke@435: flags |= _intrinsic_virtual; duke@435: } duke@435: bool changed = (flags != oflags); duke@435: if ((flags & _intrinsic_worked) != 0) { duke@435: juint count = (_intrinsic_hist_count[id] += 1); duke@435: if (count == 1) { duke@435: changed = true; // first time duke@435: } duke@435: // increment the overall count also: duke@435: _intrinsic_hist_count[vmIntrinsics::_none] += 1; duke@435: } duke@435: if (changed) { duke@435: if (((oflags ^ flags) & _intrinsic_virtual) != 0) { duke@435: // Something changed about the intrinsic's virtuality. duke@435: if ((flags & _intrinsic_virtual) != 0) { duke@435: // This is the first use of this intrinsic as a virtual call. duke@435: if (oflags != 0) { duke@435: // We already saw it as a non-virtual, so note both cases. duke@435: flags |= _intrinsic_both; duke@435: } duke@435: } else if ((oflags & _intrinsic_both) == 0) { duke@435: // This is the first use of this intrinsic as a non-virtual duke@435: flags |= _intrinsic_both; duke@435: } duke@435: } duke@435: _intrinsic_hist_flags[id] = (jubyte) (oflags | flags); duke@435: } duke@435: // update the overall flags also: duke@435: _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags; duke@435: return changed; duke@435: } duke@435: duke@435: static char* format_flags(int flags, char* buf) { duke@435: buf[0] = 0; duke@435: if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked"); duke@435: if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed"); duke@435: if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled"); duke@435: if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual"); duke@435: if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual"); duke@435: if (buf[0] == 0) strcat(buf, ","); duke@435: assert(buf[0] == ',', "must be"); duke@435: return &buf[1]; duke@435: } duke@435: duke@435: void Compile::print_intrinsic_statistics() { duke@435: char flagsbuf[100]; duke@435: ttyLocker ttyl; duke@435: if (xtty != NULL) xtty->head("statistics type='intrinsic'"); duke@435: tty->print_cr("Compiler intrinsic usage:"); duke@435: juint total = _intrinsic_hist_count[vmIntrinsics::_none]; duke@435: if (total == 0) total = 1; // avoid div0 in case of no successes duke@435: #define PRINT_STAT_LINE(name, c, f) \ duke@435: tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f); duke@435: for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) { duke@435: vmIntrinsics::ID id = (vmIntrinsics::ID) index; duke@435: int flags = _intrinsic_hist_flags[id]; duke@435: juint count = _intrinsic_hist_count[id]; duke@435: if ((flags | count) != 0) { duke@435: PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf)); duke@435: } duke@435: } duke@435: PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf)); duke@435: if (xtty != NULL) xtty->tail("statistics"); duke@435: } duke@435: duke@435: void Compile::print_statistics() { duke@435: { ttyLocker ttyl; duke@435: if (xtty != NULL) xtty->head("statistics type='opto'"); duke@435: Parse::print_statistics(); duke@435: PhaseCCP::print_statistics(); duke@435: PhaseRegAlloc::print_statistics(); duke@435: Scheduling::print_statistics(); duke@435: PhasePeephole::print_statistics(); duke@435: PhaseIdealLoop::print_statistics(); duke@435: if (xtty != NULL) xtty->tail("statistics"); duke@435: } duke@435: if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) { duke@435: // put this under its own element. duke@435: print_intrinsic_statistics(); duke@435: } duke@435: } duke@435: #endif //PRODUCT duke@435: duke@435: // Support for bundling info duke@435: Bundle* Compile::node_bundling(const Node *n) { duke@435: assert(valid_bundle_info(n), "oob"); duke@435: return &_node_bundling_base[n->_idx]; duke@435: } duke@435: duke@435: bool Compile::valid_bundle_info(const Node *n) { duke@435: return (_node_bundling_limit > n->_idx); duke@435: } duke@435: duke@435: never@1515: void Compile::gvn_replace_by(Node* n, Node* nn) { never@1515: for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) { never@1515: Node* use = n->last_out(i); never@1515: bool is_in_table = initial_gvn()->hash_delete(use); never@1515: uint uses_found = 0; never@1515: for (uint j = 0; j < use->len(); j++) { never@1515: if (use->in(j) == n) { never@1515: if (j < use->req()) never@1515: use->set_req(j, nn); never@1515: else never@1515: use->set_prec(j, nn); never@1515: uses_found++; never@1515: } never@1515: } never@1515: if (is_in_table) { never@1515: // reinsert into table never@1515: initial_gvn()->hash_find_insert(use); never@1515: } never@1515: record_for_igvn(use); never@1515: i -= uses_found; // we deleted 1 or more copies of this edge never@1515: } never@1515: } never@1515: never@1515: bharadwaj@4315: static inline bool not_a_node(const Node* n) { bharadwaj@4315: if (n == NULL) return true; bharadwaj@4315: if (((intptr_t)n & 1) != 0) return true; // uninitialized, etc. bharadwaj@4315: if (*(address*)n == badAddress) return true; // kill by Node::destruct bharadwaj@4315: return false; bharadwaj@4315: } never@1515: duke@435: // Identify all nodes that are reachable from below, useful. duke@435: // Use breadth-first pass that records state in a Unique_Node_List, duke@435: // recursive traversal is slower. duke@435: void Compile::identify_useful_nodes(Unique_Node_List &useful) { duke@435: int estimated_worklist_size = unique(); duke@435: useful.map( estimated_worklist_size, NULL ); // preallocate space duke@435: duke@435: // Initialize worklist duke@435: if (root() != NULL) { useful.push(root()); } duke@435: // If 'top' is cached, declare it useful to preserve cached node duke@435: if( cached_top_node() ) { useful.push(cached_top_node()); } duke@435: duke@435: // Push all useful nodes onto the list, breadthfirst duke@435: for( uint next = 0; next < useful.size(); ++next ) { duke@435: assert( next < unique(), "Unique useful nodes < total nodes"); duke@435: Node *n = useful.at(next); duke@435: uint max = n->len(); duke@435: for( uint i = 0; i < max; ++i ) { duke@435: Node *m = n->in(i); bharadwaj@4315: if (not_a_node(m)) continue; duke@435: useful.push(m); duke@435: } duke@435: } duke@435: } duke@435: bharadwaj@4315: // Update dead_node_list with any missing dead nodes using useful bharadwaj@4315: // list. Consider all non-useful nodes to be useless i.e., dead nodes. bharadwaj@4315: void Compile::update_dead_node_list(Unique_Node_List &useful) { bharadwaj@4315: uint max_idx = unique(); bharadwaj@4315: VectorSet& useful_node_set = useful.member_set(); bharadwaj@4315: bharadwaj@4315: for (uint node_idx = 0; node_idx < max_idx; node_idx++) { bharadwaj@4315: // If node with index node_idx is not in useful set, bharadwaj@4315: // mark it as dead in dead node list. bharadwaj@4315: if (! useful_node_set.test(node_idx) ) { bharadwaj@4315: record_dead_node(node_idx); bharadwaj@4315: } bharadwaj@4315: } bharadwaj@4315: } bharadwaj@4315: roland@4409: void Compile::remove_useless_late_inlines(GrowableArray* inlines, Unique_Node_List &useful) { roland@4409: int shift = 0; roland@4409: for (int i = 0; i < inlines->length(); i++) { roland@4409: CallGenerator* cg = inlines->at(i); roland@4409: CallNode* call = cg->call_node(); roland@4409: if (shift > 0) { roland@4409: inlines->at_put(i-shift, cg); roland@4409: } roland@4409: if (!useful.member(call)) { roland@4409: shift++; roland@4409: } roland@4409: } roland@4409: inlines->trunc_to(inlines->length()-shift); roland@4409: } roland@4409: duke@435: // Disconnect all useless nodes by disconnecting those at the boundary. duke@435: void Compile::remove_useless_nodes(Unique_Node_List &useful) { duke@435: uint next = 0; kvn@3260: while (next < useful.size()) { duke@435: Node *n = useful.at(next++); duke@435: // Use raw traversal of out edges since this code removes out edges duke@435: int max = n->outcnt(); kvn@3260: for (int j = 0; j < max; ++j) { duke@435: Node* child = n->raw_out(j); kvn@3260: if (! useful.member(child)) { kvn@3260: assert(!child->is_top() || child != top(), kvn@3260: "If top is cached in Compile object it is in useful list"); duke@435: // Only need to remove this out-edge to the useless node duke@435: n->raw_del_out(j); duke@435: --j; duke@435: --max; duke@435: } duke@435: } duke@435: if (n->outcnt() == 1 && n->has_special_unique_user()) { kvn@3260: record_for_igvn(n->unique_out()); kvn@3260: } kvn@3260: } kvn@3260: // Remove useless macro and predicate opaq nodes kvn@3260: for (int i = C->macro_count()-1; i >= 0; i--) { kvn@3260: Node* n = C->macro_node(i); kvn@3260: if (!useful.member(n)) { kvn@3260: remove_macro_node(n); duke@435: } duke@435: } roland@4589: // Remove useless expensive node roland@4589: for (int i = C->expensive_count()-1; i >= 0; i--) { roland@4589: Node* n = C->expensive_node(i); roland@4589: if (!useful.member(n)) { roland@4589: remove_expensive_node(n); roland@4589: } roland@4589: } roland@4409: // clean up the late inline lists roland@4409: remove_useless_late_inlines(&_string_late_inlines, useful); kvn@5110: remove_useless_late_inlines(&_boxing_late_inlines, useful); roland@4409: remove_useless_late_inlines(&_late_inlines, useful); duke@435: debug_only(verify_graph_edges(true/*check for no_dead_code*/);) duke@435: } duke@435: duke@435: //------------------------------frame_size_in_words----------------------------- duke@435: // frame_slots in units of words duke@435: int Compile::frame_size_in_words() const { duke@435: // shift is 0 in LP32 and 1 in LP64 duke@435: const int shift = (LogBytesPerWord - LogBytesPerInt); duke@435: int words = _frame_slots >> shift; duke@435: assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" ); duke@435: return words; duke@435: } duke@435: duke@435: // ============================================================================ duke@435: //------------------------------CompileWrapper--------------------------------- duke@435: class CompileWrapper : public StackObj { duke@435: Compile *const _compile; duke@435: public: duke@435: CompileWrapper(Compile* compile); duke@435: duke@435: ~CompileWrapper(); duke@435: }; duke@435: duke@435: CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) { duke@435: // the Compile* pointer is stored in the current ciEnv: duke@435: ciEnv* env = compile->env(); duke@435: assert(env == ciEnv::current(), "must already be a ciEnv active"); duke@435: assert(env->compiler_data() == NULL, "compile already active?"); duke@435: env->set_compiler_data(compile); duke@435: assert(compile == Compile::current(), "sanity"); duke@435: duke@435: compile->set_type_dict(NULL); duke@435: compile->set_type_hwm(NULL); duke@435: compile->set_type_last_size(0); duke@435: compile->set_last_tf(NULL, NULL); duke@435: compile->set_indexSet_arena(NULL); duke@435: compile->set_indexSet_free_block_list(NULL); duke@435: compile->init_type_arena(); duke@435: Type::Initialize(compile); duke@435: _compile->set_scratch_buffer_blob(NULL); duke@435: _compile->begin_method(); duke@435: } duke@435: CompileWrapper::~CompileWrapper() { duke@435: _compile->end_method(); duke@435: if (_compile->scratch_buffer_blob() != NULL) duke@435: BufferBlob::free(_compile->scratch_buffer_blob()); duke@435: _compile->env()->set_compiler_data(NULL); duke@435: } duke@435: duke@435: duke@435: //----------------------------print_compile_messages--------------------------- duke@435: void Compile::print_compile_messages() { duke@435: #ifndef PRODUCT duke@435: // Check if recompiling duke@435: if (_subsume_loads == false && PrintOpto) { duke@435: // Recompiling without allowing machine instructions to subsume loads duke@435: tty->print_cr("*********************************************************"); duke@435: tty->print_cr("** Bailout: Recompile without subsuming loads **"); duke@435: tty->print_cr("*********************************************************"); duke@435: } kvn@473: if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) { kvn@473: // Recompiling without escape analysis kvn@473: tty->print_cr("*********************************************************"); kvn@473: tty->print_cr("** Bailout: Recompile without escape analysis **"); kvn@473: tty->print_cr("*********************************************************"); kvn@473: } kvn@5110: if (_eliminate_boxing != EliminateAutoBox && PrintOpto) { kvn@5110: // Recompiling without boxing elimination kvn@5110: tty->print_cr("*********************************************************"); kvn@5110: tty->print_cr("** Bailout: Recompile without boxing elimination **"); kvn@5110: tty->print_cr("*********************************************************"); kvn@5110: } duke@435: if (env()->break_at_compile()) { twisti@1040: // Open the debugger when compiling this method. duke@435: tty->print("### Breaking when compiling: "); duke@435: method()->print_short_name(); duke@435: tty->cr(); duke@435: BREAKPOINT; duke@435: } duke@435: duke@435: if( PrintOpto ) { duke@435: if (is_osr_compilation()) { duke@435: tty->print("[OSR]%3d", _compile_id); duke@435: } else { duke@435: tty->print("%3d", _compile_id); duke@435: } duke@435: } duke@435: #endif duke@435: } duke@435: duke@435: kvn@2414: //-----------------------init_scratch_buffer_blob------------------------------ kvn@2414: // Construct a temporary BufferBlob and cache it for this compile. twisti@2350: void Compile::init_scratch_buffer_blob(int const_size) { kvn@2414: // If there is already a scratch buffer blob allocated and the kvn@2414: // constant section is big enough, use it. Otherwise free the kvn@2414: // current and allocate a new one. kvn@2414: BufferBlob* blob = scratch_buffer_blob(); kvn@2414: if ((blob != NULL) && (const_size <= _scratch_const_size)) { kvn@2414: // Use the current blob. kvn@2414: } else { kvn@2414: if (blob != NULL) { kvn@2414: BufferBlob::free(blob); kvn@2414: } duke@435: kvn@2414: ResourceMark rm; kvn@2414: _scratch_const_size = const_size; kvn@2414: int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size); kvn@2414: blob = BufferBlob::create("Compile::scratch_buffer", size); kvn@2414: // Record the buffer blob for next time. kvn@2414: set_scratch_buffer_blob(blob); kvn@2414: // Have we run out of code space? kvn@2414: if (scratch_buffer_blob() == NULL) { kvn@2414: // Let CompilerBroker disable further compilations. kvn@2414: record_failure("Not enough space for scratch buffer in CodeCache"); kvn@2414: return; kvn@2414: } kvn@598: } duke@435: duke@435: // Initialize the relocation buffers twisti@2103: relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size; duke@435: set_scratch_locs_memory(locs_buf); duke@435: } duke@435: duke@435: duke@435: //-----------------------scratch_emit_size------------------------------------- duke@435: // Helper function that computes size by emitting code duke@435: uint Compile::scratch_emit_size(const Node* n) { twisti@2350: // Start scratch_emit_size section. twisti@2350: set_in_scratch_emit_size(true); twisti@2350: duke@435: // Emit into a trash buffer and count bytes emitted. duke@435: // This is a pretty expensive way to compute a size, duke@435: // but it works well enough if seldom used. duke@435: // All common fixed-size instructions are given a size duke@435: // method by the AD file. duke@435: // Note that the scratch buffer blob and locs memory are duke@435: // allocated at the beginning of the compile task, and duke@435: // may be shared by several calls to scratch_emit_size. duke@435: // The allocation of the scratch buffer blob is particularly duke@435: // expensive, since it has to grab the code cache lock. duke@435: BufferBlob* blob = this->scratch_buffer_blob(); duke@435: assert(blob != NULL, "Initialize BufferBlob at start"); duke@435: assert(blob->size() > MAX_inst_size, "sanity"); duke@435: relocInfo* locs_buf = scratch_locs_memory(); twisti@2103: address blob_begin = blob->content_begin(); duke@435: address blob_end = (address)locs_buf; twisti@2103: assert(blob->content_contains(blob_end), "sanity"); duke@435: CodeBuffer buf(blob_begin, blob_end - blob_begin); twisti@2350: buf.initialize_consts_size(_scratch_const_size); duke@435: buf.initialize_stubs_size(MAX_stubs_size); duke@435: assert(locs_buf != NULL, "sanity"); twisti@2350: int lsize = MAX_locs_size / 3; twisti@2350: buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize); twisti@2350: buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize); twisti@2350: buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize); twisti@2350: twisti@2350: // Do the emission. kvn@3037: kvn@3037: Label fakeL; // Fake label for branch instructions. kvn@3051: Label* saveL = NULL; kvn@3051: uint save_bnum = 0; kvn@3051: bool is_branch = n->is_MachBranch(); kvn@3037: if (is_branch) { kvn@3037: MacroAssembler masm(&buf); kvn@3037: masm.bind(fakeL); kvn@3051: n->as_MachBranch()->save_label(&saveL, &save_bnum); kvn@3051: n->as_MachBranch()->label_set(&fakeL, 0); kvn@3037: } duke@435: n->emit(buf, this->regalloc()); kvn@3051: if (is_branch) // Restore label. kvn@3051: n->as_MachBranch()->label_set(saveL, save_bnum); twisti@2350: twisti@2350: // End scratch_emit_size section. twisti@2350: set_in_scratch_emit_size(false); twisti@2350: twisti@2103: return buf.insts_size(); duke@435: } duke@435: duke@435: duke@435: // ============================================================================ duke@435: //------------------------------Compile standard------------------------------- duke@435: debug_only( int Compile::_debug_idx = 100000; ) duke@435: duke@435: // Compile a method. entry_bci is -1 for normal compilations and indicates duke@435: // the continuation bci for on stack replacement. duke@435: duke@435: kvn@5110: Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, kvn@5110: bool subsume_loads, bool do_escape_analysis, bool eliminate_boxing ) duke@435: : Phase(Compiler), duke@435: _env(ci_env), duke@435: _log(ci_env->log()), duke@435: _compile_id(ci_env->compile_id()), duke@435: _save_argument_registers(false), duke@435: _stub_name(NULL), duke@435: _stub_function(NULL), duke@435: _stub_entry_point(NULL), duke@435: _method(target), duke@435: _entry_bci(osr_bci), duke@435: _initial_gvn(NULL), duke@435: _for_igvn(NULL), duke@435: _warm_calls(NULL), duke@435: _subsume_loads(subsume_loads), kvn@473: _do_escape_analysis(do_escape_analysis), kvn@5110: _eliminate_boxing(eliminate_boxing), duke@435: _failure_reason(NULL), duke@435: _code_buffer("Compile::Fill_buffer"), duke@435: _orig_pc_slot(0), duke@435: _orig_pc_slot_offset_in_bytes(0), twisti@1700: _has_method_handle_invokes(false), twisti@2350: _mach_constant_base_node(NULL), duke@435: _node_bundling_limit(0), duke@435: _node_bundling_base(NULL), kvn@1294: _java_calls(0), kvn@1294: _inner_loops(0), twisti@2350: _scratch_const_size(-1), twisti@2350: _in_scratch_emit_size(false), bharadwaj@4315: _dead_node_list(comp_arena()), bharadwaj@4315: _dead_node_count(0), duke@435: #ifndef PRODUCT duke@435: _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")), duke@435: _printer(IdealGraphPrinter::printer()), duke@435: #endif roland@4357: _congraph(NULL), roland@4409: _late_inlines(comp_arena(), 2, 0, NULL), roland@4409: _string_late_inlines(comp_arena(), 2, 0, NULL), kvn@5110: _boxing_late_inlines(comp_arena(), 2, 0, NULL), roland@4409: _late_inlines_pos(0), roland@4409: _number_of_mh_late_inlines(0), roland@4409: _inlining_progress(false), roland@4409: _inlining_incrementally(false), roland@4357: _print_inlining_list(NULL), roland@4357: _print_inlining(0) { duke@435: C = this; duke@435: duke@435: CompileWrapper cw(this); duke@435: #ifndef PRODUCT duke@435: if (TimeCompiler2) { duke@435: tty->print(" "); duke@435: target->holder()->name()->print(); duke@435: tty->print("."); duke@435: target->print_short_name(); duke@435: tty->print(" "); duke@435: } duke@435: TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2); duke@435: TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false); jrose@535: bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly"); jrose@535: if (!print_opto_assembly) { jrose@535: bool print_assembly = (PrintAssembly || _method->should_print_assembly()); jrose@535: if (print_assembly && !Disassembler::can_decode()) { jrose@535: tty->print_cr("PrintAssembly request changed to PrintOptoAssembly"); jrose@535: print_opto_assembly = true; jrose@535: } jrose@535: } jrose@535: set_print_assembly(print_opto_assembly); never@802: set_parsed_irreducible_loop(false); duke@435: #endif duke@435: duke@435: if (ProfileTraps) { duke@435: // Make sure the method being compiled gets its own MDO, duke@435: // so we can at least track the decompile_count(). iveresov@2349: method()->ensure_method_data(); duke@435: } duke@435: duke@435: Init(::AliasLevel); duke@435: duke@435: duke@435: print_compile_messages(); duke@435: duke@435: if (UseOldInlining || PrintCompilation NOT_PRODUCT( || PrintOpto) ) duke@435: _ilt = InlineTree::build_inline_tree_root(); duke@435: else duke@435: _ilt = NULL; duke@435: duke@435: // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice duke@435: assert(num_alias_types() >= AliasIdxRaw, ""); duke@435: duke@435: #define MINIMUM_NODE_HASH 1023 duke@435: // Node list that Iterative GVN will start with duke@435: Unique_Node_List for_igvn(comp_arena()); duke@435: set_for_igvn(&for_igvn); duke@435: duke@435: // GVN that will be run immediately on new nodes duke@435: uint estimated_size = method()->code_size()*4+64; duke@435: estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size); duke@435: PhaseGVN gvn(node_arena(), estimated_size); duke@435: set_initial_gvn(&gvn); duke@435: kvn@4448: if (PrintInlining || PrintIntrinsics NOT_PRODUCT( || PrintOptoInlining)) { roland@4357: _print_inlining_list = new (comp_arena())GrowableArray(comp_arena(), 1, 1, PrintInliningBuffer()); roland@4357: } duke@435: { // Scope for timing the parser duke@435: TracePhase t3("parse", &_t_parser, true); duke@435: duke@435: // Put top into the hash table ASAP. duke@435: initial_gvn()->transform_no_reclaim(top()); duke@435: duke@435: // Set up tf(), start(), and find a CallGenerator. johnc@2781: CallGenerator* cg = NULL; duke@435: if (is_osr_compilation()) { duke@435: const TypeTuple *domain = StartOSRNode::osr_domain(); duke@435: const TypeTuple *range = TypeTuple::make_range(method()->signature()); duke@435: init_tf(TypeFunc::make(domain, range)); kvn@4115: StartNode* s = new (this) StartOSRNode(root(), domain); duke@435: initial_gvn()->set_type_bottom(s); duke@435: init_start(s); duke@435: cg = CallGenerator::for_osr(method(), entry_bci()); duke@435: } else { duke@435: // Normal case. duke@435: init_tf(TypeFunc::make(method())); kvn@4115: StartNode* s = new (this) StartNode(root(), tf()->domain()); duke@435: initial_gvn()->set_type_bottom(s); duke@435: init_start(s); johnc@2781: if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) { johnc@2781: // With java.lang.ref.reference.get() we must go through the johnc@2781: // intrinsic when G1 is enabled - even when get() is the root johnc@2781: // method of the compile - so that, if necessary, the value in johnc@2781: // the referent field of the reference object gets recorded by johnc@2781: // the pre-barrier code. johnc@2781: // Specifically, if G1 is enabled, the value in the referent johnc@2781: // field is recorded by the G1 SATB pre barrier. This will johnc@2781: // result in the referent being marked live and the reference johnc@2781: // object removed from the list of discovered references during johnc@2781: // reference processing. johnc@2781: cg = find_intrinsic(method(), false); johnc@2781: } johnc@2781: if (cg == NULL) { johnc@2781: float past_uses = method()->interpreter_invocation_count(); johnc@2781: float expected_uses = past_uses; johnc@2781: cg = CallGenerator::for_inline(method(), expected_uses); johnc@2781: } duke@435: } duke@435: if (failing()) return; duke@435: if (cg == NULL) { duke@435: record_method_not_compilable_all_tiers("cannot parse method"); duke@435: return; duke@435: } duke@435: JVMState* jvms = build_start_state(start(), tf()); duke@435: if ((jvms = cg->generate(jvms)) == NULL) { duke@435: record_method_not_compilable("method parse failed"); duke@435: return; duke@435: } duke@435: GraphKit kit(jvms); duke@435: duke@435: if (!kit.stopped()) { duke@435: // Accept return values, and transfer control we know not where. duke@435: // This is done by a special, unique ReturnNode bound to root. duke@435: return_values(kit.jvms()); duke@435: } duke@435: duke@435: if (kit.has_exceptions()) { duke@435: // Any exceptions that escape from this call must be rethrown duke@435: // to whatever caller is dynamically above us on the stack. duke@435: // This is done by a special, unique RethrowNode bound to root. duke@435: rethrow_exceptions(kit.transfer_exceptions_into_jvms()); duke@435: } duke@435: roland@4409: assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off"); roland@4409: roland@4409: if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) { roland@4409: inline_string_calls(true); never@1515: } roland@4409: roland@4409: if (failing()) return; never@1515: sla@5237: print_method(PHASE_BEFORE_REMOVEUSELESS, 3); never@802: duke@435: // Remove clutter produced by parsing. duke@435: if (!failing()) { duke@435: ResourceMark rm; duke@435: PhaseRemoveUseless pru(initial_gvn(), &for_igvn); duke@435: } duke@435: } duke@435: duke@435: // Note: Large methods are capped off in do_one_bytecode(). duke@435: if (failing()) return; duke@435: duke@435: // After parsing, node notes are no longer automagic. duke@435: // They must be propagated by register_new_node_with_optimizer(), duke@435: // clone(), or the like. duke@435: set_default_node_notes(NULL); duke@435: duke@435: for (;;) { duke@435: int successes = Inline_Warm(); duke@435: if (failing()) return; duke@435: if (successes == 0) break; duke@435: } duke@435: duke@435: // Drain the list. duke@435: Finish_Warm(); duke@435: #ifndef PRODUCT duke@435: if (_printer) { duke@435: _printer->print_inlining(this); duke@435: } duke@435: #endif duke@435: duke@435: if (failing()) return; duke@435: NOT_PRODUCT( verify_graph_edges(); ) duke@435: duke@435: // Now optimize duke@435: Optimize(); duke@435: if (failing()) return; duke@435: NOT_PRODUCT( verify_graph_edges(); ) duke@435: duke@435: #ifndef PRODUCT duke@435: if (PrintIdeal) { duke@435: ttyLocker ttyl; // keep the following output all in one block 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("ideal compile_id='%d'%s", compile_id(), duke@435: is_osr_compilation() ? " compile_kind='osr'" : duke@435: ""); duke@435: } duke@435: root()->dump(9999); duke@435: if (xtty != NULL) { duke@435: xtty->tail("ideal"); duke@435: } duke@435: } duke@435: #endif duke@435: duke@435: // Now that we know the size of all the monitors we can add a fixed slot duke@435: // for the original deopt pc. duke@435: duke@435: _orig_pc_slot = fixed_slots(); duke@435: int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size); duke@435: set_fixed_slots(next_slot); duke@435: duke@435: // Now generate code duke@435: Code_Gen(); duke@435: if (failing()) return; duke@435: duke@435: // Check if we want to skip execution of all compiled code. duke@435: { duke@435: #ifndef PRODUCT duke@435: if (OptoNoExecute) { duke@435: record_method_not_compilable("+OptoNoExecute"); // Flag as failed duke@435: return; duke@435: } duke@435: TracePhase t2("install_code", &_t_registerMethod, TimeCompiler); duke@435: #endif duke@435: duke@435: if (is_osr_compilation()) { duke@435: _code_offsets.set_value(CodeOffsets::Verified_Entry, 0); duke@435: _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size); duke@435: } else { duke@435: _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size); duke@435: _code_offsets.set_value(CodeOffsets::OSR_Entry, 0); duke@435: } duke@435: duke@435: env()->register_method(_method, _entry_bci, duke@435: &_code_offsets, duke@435: _orig_pc_slot_offset_in_bytes, duke@435: code_buffer(), duke@435: frame_size_in_words(), _oop_map_set, duke@435: &_handler_table, &_inc_table, duke@435: compiler, duke@435: env()->comp_level(), kvn@4103: has_unsafe_access(), kvn@4103: SharedRuntime::is_wide_vector(max_vector_size()) duke@435: ); vlivanov@4154: vlivanov@4154: if (log() != NULL) // Print code cache state into compiler log vlivanov@4154: log()->code_cache_state(); duke@435: } duke@435: } duke@435: duke@435: //------------------------------Compile---------------------------------------- duke@435: // Compile a runtime stub duke@435: Compile::Compile( ciEnv* ci_env, duke@435: TypeFunc_generator generator, duke@435: address stub_function, duke@435: const char *stub_name, duke@435: int is_fancy_jump, duke@435: bool pass_tls, duke@435: bool save_arg_registers, duke@435: bool return_pc ) duke@435: : Phase(Compiler), duke@435: _env(ci_env), duke@435: _log(ci_env->log()), neliasso@4730: _compile_id(0), duke@435: _save_argument_registers(save_arg_registers), duke@435: _method(NULL), duke@435: _stub_name(stub_name), duke@435: _stub_function(stub_function), duke@435: _stub_entry_point(NULL), duke@435: _entry_bci(InvocationEntryBci), duke@435: _initial_gvn(NULL), duke@435: _for_igvn(NULL), duke@435: _warm_calls(NULL), duke@435: _orig_pc_slot(0), duke@435: _orig_pc_slot_offset_in_bytes(0), duke@435: _subsume_loads(true), kvn@473: _do_escape_analysis(false), kvn@5110: _eliminate_boxing(false), duke@435: _failure_reason(NULL), duke@435: _code_buffer("Compile::Fill_buffer"), twisti@1700: _has_method_handle_invokes(false), twisti@2350: _mach_constant_base_node(NULL), duke@435: _node_bundling_limit(0), duke@435: _node_bundling_base(NULL), kvn@1294: _java_calls(0), kvn@1294: _inner_loops(0), duke@435: #ifndef PRODUCT duke@435: _trace_opto_output(TraceOptoOutput), duke@435: _printer(NULL), duke@435: #endif bharadwaj@4315: _dead_node_list(comp_arena()), bharadwaj@4315: _dead_node_count(0), roland@4357: _congraph(NULL), roland@4409: _number_of_mh_late_inlines(0), roland@4409: _inlining_progress(false), roland@4409: _inlining_incrementally(false), roland@4357: _print_inlining_list(NULL), roland@4357: _print_inlining(0) { duke@435: C = this; duke@435: duke@435: #ifndef PRODUCT duke@435: TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false); duke@435: TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false); duke@435: set_print_assembly(PrintFrameConverterAssembly); never@802: set_parsed_irreducible_loop(false); duke@435: #endif duke@435: CompileWrapper cw(this); duke@435: Init(/*AliasLevel=*/ 0); duke@435: init_tf((*generator)()); duke@435: duke@435: { duke@435: // The following is a dummy for the sake of GraphKit::gen_stub duke@435: Unique_Node_List for_igvn(comp_arena()); duke@435: set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this duke@435: PhaseGVN gvn(Thread::current()->resource_area(),255); duke@435: set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively duke@435: gvn.transform_no_reclaim(top()); duke@435: duke@435: GraphKit kit; duke@435: kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc); duke@435: } duke@435: duke@435: NOT_PRODUCT( verify_graph_edges(); ) duke@435: Code_Gen(); duke@435: if (failing()) return; duke@435: duke@435: duke@435: // Entry point will be accessed using compile->stub_entry_point(); duke@435: if (code_buffer() == NULL) { duke@435: Matcher::soft_match_failure(); duke@435: } else { duke@435: if (PrintAssembly && (WizardMode || Verbose)) duke@435: tty->print_cr("### Stub::%s", stub_name); duke@435: duke@435: if (!failing()) { duke@435: assert(_fixed_slots == 0, "no fixed slots used for runtime stubs"); duke@435: duke@435: // Make the NMethod duke@435: // For now we mark the frame as never safe for profile stackwalking duke@435: RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name, duke@435: code_buffer(), duke@435: CodeOffsets::frame_never_safe, duke@435: // _code_offsets.value(CodeOffsets::Frame_Complete), duke@435: frame_size_in_words(), duke@435: _oop_map_set, duke@435: save_arg_registers); duke@435: assert(rs != NULL && rs->is_runtime_stub(), "sanity check"); duke@435: duke@435: _stub_entry_point = rs->entry_point(); duke@435: } duke@435: } duke@435: } duke@435: duke@435: //------------------------------Init------------------------------------------- duke@435: // Prepare for a single compilation duke@435: void Compile::Init(int aliaslevel) { duke@435: _unique = 0; duke@435: _regalloc = NULL; duke@435: duke@435: _tf = NULL; // filled in later duke@435: _top = NULL; // cached later duke@435: _matcher = NULL; // filled in later duke@435: _cfg = NULL; // filled in later duke@435: duke@435: set_24_bit_selection_and_mode(Use24BitFP, false); duke@435: duke@435: _node_note_array = NULL; duke@435: _default_node_notes = NULL; duke@435: duke@435: _immutable_memory = NULL; // filled in at first inquiry duke@435: duke@435: // Globally visible Nodes duke@435: // First set TOP to NULL to give safe behavior during creation of RootNode duke@435: set_cached_top_node(NULL); kvn@4115: set_root(new (this) RootNode()); duke@435: // Now that you have a Root to point to, create the real TOP kvn@4115: set_cached_top_node( new (this) ConNode(Type::TOP) ); duke@435: set_recent_alloc(NULL, NULL); duke@435: duke@435: // Create Debug Information Recorder to record scopes, oopmaps, etc. coleenp@4037: env()->set_oop_recorder(new OopRecorder(env()->arena())); duke@435: env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder())); duke@435: env()->set_dependencies(new Dependencies(env())); duke@435: duke@435: _fixed_slots = 0; duke@435: set_has_split_ifs(false); duke@435: set_has_loops(has_method() && method()->has_loops()); // first approximation never@1515: set_has_stringbuilder(false); kvn@5110: set_has_boxed_value(false); duke@435: _trap_can_recompile = false; // no traps emitted yet duke@435: _major_progress = true; // start out assuming good things will happen duke@435: set_has_unsafe_access(false); kvn@4103: set_max_vector_size(0); duke@435: Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist)); duke@435: set_decompile_count(0); duke@435: rasbold@853: set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency")); iveresov@2138: set_num_loop_opts(LoopOptsCount); iveresov@2138: set_do_inlining(Inline); iveresov@2138: set_max_inline_size(MaxInlineSize); iveresov@2138: set_freq_inline_size(FreqInlineSize); iveresov@2138: set_do_scheduling(OptoScheduling); iveresov@2138: set_do_count_invocations(false); iveresov@2138: set_do_method_data_update(false); duke@435: duke@435: if (debug_info()->recording_non_safepoints()) { duke@435: set_node_note_array(new(comp_arena()) GrowableArray duke@435: (comp_arena(), 8, 0, NULL)); duke@435: set_default_node_notes(Node_Notes::make(this)); duke@435: } duke@435: duke@435: // // -- Initialize types before each compile -- duke@435: // // Update cached type information duke@435: // if( _method && _method->constants() ) duke@435: // Type::update_loaded_types(_method, _method->constants()); duke@435: duke@435: // Init alias_type map. kvn@473: if (!_do_escape_analysis && aliaslevel == 3) duke@435: aliaslevel = 2; // No unique types without escape analysis duke@435: _AliasLevel = aliaslevel; duke@435: const int grow_ats = 16; duke@435: _max_alias_types = grow_ats; duke@435: _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats); duke@435: AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats); duke@435: Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats); duke@435: { duke@435: for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i]; duke@435: } duke@435: // Initialize the first few types. duke@435: _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL); duke@435: _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM); duke@435: _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM); duke@435: _num_alias_types = AliasIdxRaw+1; duke@435: // Zero out the alias type cache. duke@435: Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache)); duke@435: // A NULL adr_type hits in the cache right away. Preload the right answer. duke@435: probe_alias_cache(NULL)->_index = AliasIdxTop; duke@435: duke@435: _intrinsics = NULL; kvn@2040: _macro_nodes = new(comp_arena()) GrowableArray(comp_arena(), 8, 0, NULL); kvn@2040: _predicate_opaqs = new(comp_arena()) GrowableArray(comp_arena(), 8, 0, NULL); roland@4589: _expensive_nodes = new(comp_arena()) GrowableArray(comp_arena(), 8, 0, NULL); duke@435: register_library_intrinsics(); duke@435: } duke@435: duke@435: //---------------------------init_start---------------------------------------- duke@435: // Install the StartNode on this compile object. duke@435: void Compile::init_start(StartNode* s) { duke@435: if (failing()) duke@435: return; // already failing duke@435: assert(s == start(), ""); duke@435: } duke@435: duke@435: StartNode* Compile::start() const { duke@435: assert(!failing(), ""); duke@435: for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) { duke@435: Node* start = root()->fast_out(i); duke@435: if( start->is_Start() ) duke@435: return start->as_Start(); duke@435: } duke@435: ShouldNotReachHere(); duke@435: return NULL; duke@435: } duke@435: duke@435: //-------------------------------immutable_memory------------------------------------- duke@435: // Access immutable memory duke@435: Node* Compile::immutable_memory() { duke@435: if (_immutable_memory != NULL) { duke@435: return _immutable_memory; duke@435: } duke@435: StartNode* s = start(); duke@435: for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) { duke@435: Node *p = s->fast_out(i); duke@435: if (p != s && p->as_Proj()->_con == TypeFunc::Memory) { duke@435: _immutable_memory = p; duke@435: return _immutable_memory; duke@435: } duke@435: } duke@435: ShouldNotReachHere(); duke@435: return NULL; duke@435: } duke@435: duke@435: //----------------------set_cached_top_node------------------------------------ duke@435: // Install the cached top node, and make sure Node::is_top works correctly. duke@435: void Compile::set_cached_top_node(Node* tn) { duke@435: if (tn != NULL) verify_top(tn); duke@435: Node* old_top = _top; duke@435: _top = tn; duke@435: // Calling Node::setup_is_top allows the nodes the chance to adjust duke@435: // their _out arrays. duke@435: if (_top != NULL) _top->setup_is_top(); duke@435: if (old_top != NULL) old_top->setup_is_top(); duke@435: assert(_top == NULL || top()->is_top(), ""); duke@435: } duke@435: bharadwaj@4315: #ifdef ASSERT bharadwaj@4315: uint Compile::count_live_nodes_by_graph_walk() { bharadwaj@4315: Unique_Node_List useful(comp_arena()); bharadwaj@4315: // Get useful node list by walking the graph. bharadwaj@4315: identify_useful_nodes(useful); bharadwaj@4315: return useful.size(); bharadwaj@4315: } bharadwaj@4315: bharadwaj@4315: void Compile::print_missing_nodes() { bharadwaj@4315: bharadwaj@4315: // Return if CompileLog is NULL and PrintIdealNodeCount is false. bharadwaj@4315: if ((_log == NULL) && (! PrintIdealNodeCount)) { bharadwaj@4315: return; bharadwaj@4315: } bharadwaj@4315: bharadwaj@4315: // This is an expensive function. It is executed only when the user bharadwaj@4315: // specifies VerifyIdealNodeCount option or otherwise knows the bharadwaj@4315: // additional work that needs to be done to identify reachable nodes bharadwaj@4315: // by walking the flow graph and find the missing ones using bharadwaj@4315: // _dead_node_list. bharadwaj@4315: bharadwaj@4315: Unique_Node_List useful(comp_arena()); bharadwaj@4315: // Get useful node list by walking the graph. bharadwaj@4315: identify_useful_nodes(useful); bharadwaj@4315: bharadwaj@4315: uint l_nodes = C->live_nodes(); bharadwaj@4315: uint l_nodes_by_walk = useful.size(); bharadwaj@4315: bharadwaj@4315: if (l_nodes != l_nodes_by_walk) { bharadwaj@4315: if (_log != NULL) { bharadwaj@4315: _log->begin_head("mismatched_nodes count='%d'", abs((int) (l_nodes - l_nodes_by_walk))); bharadwaj@4315: _log->stamp(); bharadwaj@4315: _log->end_head(); bharadwaj@4315: } bharadwaj@4315: VectorSet& useful_member_set = useful.member_set(); bharadwaj@4315: int last_idx = l_nodes_by_walk; bharadwaj@4315: for (int i = 0; i < last_idx; i++) { bharadwaj@4315: if (useful_member_set.test(i)) { bharadwaj@4315: if (_dead_node_list.test(i)) { bharadwaj@4315: if (_log != NULL) { bharadwaj@4315: _log->elem("mismatched_node_info node_idx='%d' type='both live and dead'", i); bharadwaj@4315: } bharadwaj@4315: if (PrintIdealNodeCount) { bharadwaj@4315: // Print the log message to tty bharadwaj@4315: tty->print_cr("mismatched_node idx='%d' both live and dead'", i); bharadwaj@4315: useful.at(i)->dump(); bharadwaj@4315: } bharadwaj@4315: } bharadwaj@4315: } bharadwaj@4315: else if (! _dead_node_list.test(i)) { bharadwaj@4315: if (_log != NULL) { bharadwaj@4315: _log->elem("mismatched_node_info node_idx='%d' type='neither live nor dead'", i); bharadwaj@4315: } bharadwaj@4315: if (PrintIdealNodeCount) { bharadwaj@4315: // Print the log message to tty bharadwaj@4315: tty->print_cr("mismatched_node idx='%d' type='neither live nor dead'", i); bharadwaj@4315: } bharadwaj@4315: } bharadwaj@4315: } bharadwaj@4315: if (_log != NULL) { bharadwaj@4315: _log->tail("mismatched_nodes"); bharadwaj@4315: } bharadwaj@4315: } bharadwaj@4315: } bharadwaj@4315: #endif bharadwaj@4315: duke@435: #ifndef PRODUCT duke@435: void Compile::verify_top(Node* tn) const { duke@435: if (tn != NULL) { duke@435: assert(tn->is_Con(), "top node must be a constant"); duke@435: assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type"); duke@435: assert(tn->in(0) != NULL, "must have live top node"); duke@435: } duke@435: } duke@435: #endif duke@435: duke@435: duke@435: ///-------------------Managing Per-Node Debug & Profile Info------------------- duke@435: duke@435: void Compile::grow_node_notes(GrowableArray* arr, int grow_by) { duke@435: guarantee(arr != NULL, ""); duke@435: int num_blocks = arr->length(); duke@435: if (grow_by < num_blocks) grow_by = num_blocks; duke@435: int num_notes = grow_by * _node_notes_block_size; duke@435: Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes); duke@435: Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes)); duke@435: while (num_notes > 0) { duke@435: arr->append(notes); duke@435: notes += _node_notes_block_size; duke@435: num_notes -= _node_notes_block_size; duke@435: } duke@435: assert(num_notes == 0, "exact multiple, please"); duke@435: } duke@435: duke@435: bool Compile::copy_node_notes_to(Node* dest, Node* source) { duke@435: if (source == NULL || dest == NULL) return false; duke@435: duke@435: if (dest->is_Con()) duke@435: return false; // Do not push debug info onto constants. duke@435: duke@435: #ifdef ASSERT duke@435: // Leave a bread crumb trail pointing to the original node: duke@435: if (dest != NULL && dest != source && dest->debug_orig() == NULL) { duke@435: dest->set_debug_orig(source); duke@435: } duke@435: #endif duke@435: duke@435: if (node_note_array() == NULL) duke@435: return false; // Not collecting any notes now. duke@435: duke@435: // This is a copy onto a pre-existing node, which may already have notes. duke@435: // If both nodes have notes, do not overwrite any pre-existing notes. duke@435: Node_Notes* source_notes = node_notes_at(source->_idx); duke@435: if (source_notes == NULL || source_notes->is_clear()) return false; duke@435: Node_Notes* dest_notes = node_notes_at(dest->_idx); duke@435: if (dest_notes == NULL || dest_notes->is_clear()) { duke@435: return set_node_notes_at(dest->_idx, source_notes); duke@435: } duke@435: duke@435: Node_Notes merged_notes = (*source_notes); duke@435: // The order of operations here ensures that dest notes will win... duke@435: merged_notes.update_from(dest_notes); duke@435: return set_node_notes_at(dest->_idx, &merged_notes); duke@435: } duke@435: duke@435: duke@435: //--------------------------allow_range_check_smearing------------------------- duke@435: // Gating condition for coalescing similar range checks. duke@435: // Sometimes we try 'speculatively' replacing a series of a range checks by a duke@435: // single covering check that is at least as strong as any of them. duke@435: // If the optimization succeeds, the simplified (strengthened) range check duke@435: // will always succeed. If it fails, we will deopt, and then give up duke@435: // on the optimization. duke@435: bool Compile::allow_range_check_smearing() const { duke@435: // If this method has already thrown a range-check, duke@435: // assume it was because we already tried range smearing duke@435: // and it failed. duke@435: uint already_trapped = trap_count(Deoptimization::Reason_range_check); duke@435: return !already_trapped; duke@435: } duke@435: duke@435: duke@435: //------------------------------flatten_alias_type----------------------------- duke@435: const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const { duke@435: int offset = tj->offset(); duke@435: TypePtr::PTR ptr = tj->ptr(); duke@435: kvn@682: // Known instance (scalarizable allocation) alias only with itself. kvn@682: bool is_known_inst = tj->isa_oopptr() != NULL && kvn@682: tj->is_oopptr()->is_known_instance(); kvn@682: duke@435: // Process weird unsafe references. duke@435: if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) { duke@435: assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops"); kvn@682: assert(!is_known_inst, "scalarizable allocation should not have unsafe references"); duke@435: tj = TypeOopPtr::BOTTOM; duke@435: ptr = tj->ptr(); duke@435: offset = tj->offset(); duke@435: } duke@435: duke@435: // Array pointers need some flattening duke@435: const TypeAryPtr *ta = tj->isa_aryptr(); kvn@682: if( ta && is_known_inst ) { kvn@682: if ( offset != Type::OffsetBot && kvn@682: offset > arrayOopDesc::length_offset_in_bytes() ) { kvn@682: offset = Type::OffsetBot; // Flatten constant access into array body only kvn@682: tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id()); kvn@682: } kvn@682: } else if( ta && _AliasLevel >= 2 ) { duke@435: // For arrays indexed by constant indices, we flatten the alias duke@435: // space to include all of the array body. Only the header, klass duke@435: // and array length can be accessed un-aliased. duke@435: if( offset != Type::OffsetBot ) { coleenp@4037: if( ta->const_oop() ) { // MethodData* or Method* duke@435: offset = Type::OffsetBot; // Flatten constant access into array body kvn@682: tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset); duke@435: } else if( offset == arrayOopDesc::length_offset_in_bytes() ) { duke@435: // range is OK as-is. duke@435: tj = ta = TypeAryPtr::RANGE; duke@435: } else if( offset == oopDesc::klass_offset_in_bytes() ) { duke@435: tj = TypeInstPtr::KLASS; // all klass loads look alike duke@435: ta = TypeAryPtr::RANGE; // generic ignored junk duke@435: ptr = TypePtr::BotPTR; duke@435: } else if( offset == oopDesc::mark_offset_in_bytes() ) { duke@435: tj = TypeInstPtr::MARK; duke@435: ta = TypeAryPtr::RANGE; // generic ignored junk duke@435: ptr = TypePtr::BotPTR; duke@435: } else { // Random constant offset into array body duke@435: offset = Type::OffsetBot; // Flatten constant access into array body kvn@682: tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset); duke@435: } duke@435: } duke@435: // Arrays of fixed size alias with arrays of unknown size. duke@435: if (ta->size() != TypeInt::POS) { duke@435: const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS); kvn@682: tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset); duke@435: } duke@435: // Arrays of known objects become arrays of unknown objects. coleenp@548: if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) { coleenp@548: const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size()); kvn@682: tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); coleenp@548: } duke@435: if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) { duke@435: const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size()); kvn@682: tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset); duke@435: } duke@435: // Arrays of bytes and of booleans both use 'bastore' and 'baload' so duke@435: // cannot be distinguished by bytecode alone. duke@435: if (ta->elem() == TypeInt::BOOL) { duke@435: const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size()); duke@435: ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE); kvn@682: tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset); duke@435: } duke@435: // During the 2nd round of IterGVN, NotNull castings are removed. duke@435: // Make sure the Bottom and NotNull variants alias the same. duke@435: // Also, make sure exact and non-exact variants alias the same. duke@435: if( ptr == TypePtr::NotNull || ta->klass_is_exact() ) { kvn@2986: tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset); duke@435: } duke@435: } duke@435: duke@435: // Oop pointers need some flattening duke@435: const TypeInstPtr *to = tj->isa_instptr(); duke@435: if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) { never@2658: ciInstanceKlass *k = to->klass()->as_instance_klass(); duke@435: if( ptr == TypePtr::Constant ) { never@2658: if (to->klass() != ciEnv::current()->Class_klass() || never@2658: offset < k->size_helper() * wordSize) { never@2658: // No constant oop pointers (such as Strings); they alias with never@2658: // unknown strings. never@2658: assert(!is_known_inst, "not scalarizable allocation"); never@2658: tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); never@2658: } kvn@682: } else if( is_known_inst ) { kvn@598: tj = to; // Keep NotNull and klass_is_exact for instance type duke@435: } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) { duke@435: // During the 2nd round of IterGVN, NotNull castings are removed. duke@435: // Make sure the Bottom and NotNull variants alias the same. duke@435: // Also, make sure exact and non-exact variants alias the same. kvn@682: tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset); duke@435: } duke@435: // Canonicalize the holder of this field coleenp@548: if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) { duke@435: // First handle header references such as a LoadKlassNode, even if the duke@435: // object's klass is unloaded at compile time (4965979). kvn@682: if (!is_known_inst) { // Do it only for non-instance types kvn@682: tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset); kvn@682: } duke@435: } else if (offset < 0 || offset >= k->size_helper() * wordSize) { never@2658: // Static fields are in the space above the normal instance never@2658: // fields in the java.lang.Class instance. never@2658: if (to->klass() != ciEnv::current()->Class_klass()) { never@2658: to = NULL; never@2658: tj = TypeOopPtr::BOTTOM; never@2658: offset = tj->offset(); never@2658: } duke@435: } else { duke@435: ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset); duke@435: if (!k->equals(canonical_holder) || tj->offset() != offset) { kvn@682: if( is_known_inst ) { kvn@682: tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id()); kvn@682: } else { kvn@682: tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset); kvn@682: } duke@435: } duke@435: } duke@435: } duke@435: duke@435: // Klass pointers to object array klasses need some flattening duke@435: const TypeKlassPtr *tk = tj->isa_klassptr(); duke@435: if( tk ) { duke@435: // If we are referencing a field within a Klass, we need duke@435: // to assume the worst case of an Object. Both exact and never@3389: // inexact types must flatten to the same alias class so never@3389: // use NotNull as the PTR. duke@435: if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) { duke@435: never@3389: tj = tk = TypeKlassPtr::make(TypePtr::NotNull, duke@435: TypeKlassPtr::OBJECT->klass(), duke@435: offset); duke@435: } duke@435: duke@435: ciKlass* klass = tk->klass(); duke@435: if( klass->is_obj_array_klass() ) { duke@435: ciKlass* k = TypeAryPtr::OOPS->klass(); duke@435: if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs duke@435: k = TypeInstPtr::BOTTOM->klass(); duke@435: tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset ); duke@435: } duke@435: duke@435: // Check for precise loads from the primary supertype array and force them duke@435: // to the supertype cache alias index. Check for generic array loads from duke@435: // the primary supertype array and also force them to the supertype cache duke@435: // alias index. Since the same load can reach both, we need to merge duke@435: // these 2 disparate memories into the same alias class. Since the duke@435: // primary supertype array is read-only, there's no chance of confusion duke@435: // where we bypass an array load and an array store. stefank@3391: int primary_supers_offset = in_bytes(Klass::primary_supers_offset()); never@3389: if (offset == Type::OffsetBot || never@3389: (offset >= primary_supers_offset && never@3389: offset < (int)(primary_supers_offset + Klass::primary_super_limit() * wordSize)) || stefank@3391: offset == (int)in_bytes(Klass::secondary_super_cache_offset())) { stefank@3391: offset = in_bytes(Klass::secondary_super_cache_offset()); duke@435: tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset ); duke@435: } duke@435: } duke@435: duke@435: // Flatten all Raw pointers together. duke@435: if (tj->base() == Type::RawPtr) duke@435: tj = TypeRawPtr::BOTTOM; duke@435: duke@435: if (tj->base() == Type::AnyPtr) duke@435: tj = TypePtr::BOTTOM; // An error, which the caller must check for. duke@435: duke@435: // Flatten all to bottom for now duke@435: switch( _AliasLevel ) { duke@435: case 0: duke@435: tj = TypePtr::BOTTOM; duke@435: break; duke@435: case 1: // Flatten to: oop, static, field or array duke@435: switch (tj->base()) { duke@435: //case Type::AryPtr: tj = TypeAryPtr::RANGE; break; duke@435: case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break; duke@435: case Type::AryPtr: // do not distinguish arrays at all duke@435: case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break; duke@435: case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break; duke@435: case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it duke@435: default: ShouldNotReachHere(); duke@435: } duke@435: break; twisti@1040: case 2: // No collapsing at level 2; keep all splits twisti@1040: case 3: // No collapsing at level 3; keep all splits duke@435: break; duke@435: default: duke@435: Unimplemented(); duke@435: } duke@435: duke@435: offset = tj->offset(); duke@435: assert( offset != Type::OffsetTop, "Offset has fallen from constant" ); duke@435: duke@435: assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) || duke@435: (offset == Type::OffsetBot && tj->base() == Type::AryPtr) || duke@435: (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) || duke@435: (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) || duke@435: (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) || duke@435: (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) || duke@435: (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) , duke@435: "For oops, klasses, raw offset must be constant; for arrays the offset is never known" ); duke@435: assert( tj->ptr() != TypePtr::TopPTR && duke@435: tj->ptr() != TypePtr::AnyNull && duke@435: tj->ptr() != TypePtr::Null, "No imprecise addresses" ); duke@435: // assert( tj->ptr() != TypePtr::Constant || duke@435: // tj->base() == Type::RawPtr || duke@435: // tj->base() == Type::KlassPtr, "No constant oop addresses" ); duke@435: duke@435: return tj; duke@435: } duke@435: duke@435: void Compile::AliasType::Init(int i, const TypePtr* at) { duke@435: _index = i; duke@435: _adr_type = at; duke@435: _field = NULL; duke@435: _is_rewritable = true; // default duke@435: const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL; kvn@658: if (atoop != NULL && atoop->is_known_instance()) { kvn@658: const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot); duke@435: _general_index = Compile::current()->get_alias_index(gt); duke@435: } else { duke@435: _general_index = 0; duke@435: } duke@435: } duke@435: duke@435: //---------------------------------print_on------------------------------------ duke@435: #ifndef PRODUCT duke@435: void Compile::AliasType::print_on(outputStream* st) { duke@435: if (index() < 10) duke@435: st->print("@ <%d> ", index()); duke@435: else st->print("@ <%d>", index()); duke@435: st->print(is_rewritable() ? " " : " RO"); duke@435: int offset = adr_type()->offset(); duke@435: if (offset == Type::OffsetBot) duke@435: st->print(" +any"); duke@435: else st->print(" +%-3d", offset); duke@435: st->print(" in "); duke@435: adr_type()->dump_on(st); duke@435: const TypeOopPtr* tjp = adr_type()->isa_oopptr(); duke@435: if (field() != NULL && tjp) { duke@435: if (tjp->klass() != field()->holder() || duke@435: tjp->offset() != field()->offset_in_bytes()) { duke@435: st->print(" != "); duke@435: field()->print(); duke@435: st->print(" ***"); duke@435: } duke@435: } duke@435: } duke@435: duke@435: void print_alias_types() { duke@435: Compile* C = Compile::current(); duke@435: tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1); duke@435: for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) { duke@435: C->alias_type(idx)->print_on(tty); duke@435: tty->cr(); duke@435: } duke@435: } duke@435: #endif duke@435: duke@435: duke@435: //----------------------------probe_alias_cache-------------------------------- duke@435: Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) { duke@435: intptr_t key = (intptr_t) adr_type; duke@435: key ^= key >> logAliasCacheSize; duke@435: return &_alias_cache[key & right_n_bits(logAliasCacheSize)]; duke@435: } duke@435: duke@435: duke@435: //-----------------------------grow_alias_types-------------------------------- duke@435: void Compile::grow_alias_types() { duke@435: const int old_ats = _max_alias_types; // how many before? duke@435: const int new_ats = old_ats; // how many more? duke@435: const int grow_ats = old_ats+new_ats; // how many now? duke@435: _max_alias_types = grow_ats; duke@435: _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats); duke@435: AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats); duke@435: Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats); duke@435: for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i]; duke@435: } duke@435: duke@435: duke@435: //--------------------------------find_alias_type------------------------------ never@2658: Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) { duke@435: if (_AliasLevel == 0) duke@435: return alias_type(AliasIdxBot); duke@435: duke@435: AliasCacheEntry* ace = probe_alias_cache(adr_type); duke@435: if (ace->_adr_type == adr_type) { duke@435: return alias_type(ace->_index); duke@435: } duke@435: duke@435: // Handle special cases. duke@435: if (adr_type == NULL) return alias_type(AliasIdxTop); duke@435: if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot); duke@435: duke@435: // Do it the slow way. duke@435: const TypePtr* flat = flatten_alias_type(adr_type); duke@435: duke@435: #ifdef ASSERT duke@435: assert(flat == flatten_alias_type(flat), "idempotent"); duke@435: assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr"); duke@435: if (flat->isa_oopptr() && !flat->isa_klassptr()) { duke@435: const TypeOopPtr* foop = flat->is_oopptr(); kvn@682: // Scalarizable allocations have exact klass always. kvn@682: bool exact = !foop->klass_is_exact() || foop->is_known_instance(); kvn@682: const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr(); duke@435: assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type"); duke@435: } duke@435: assert(flat == flatten_alias_type(flat), "exact bit doesn't matter"); duke@435: #endif duke@435: duke@435: int idx = AliasIdxTop; duke@435: for (int i = 0; i < num_alias_types(); i++) { duke@435: if (alias_type(i)->adr_type() == flat) { duke@435: idx = i; duke@435: break; duke@435: } duke@435: } duke@435: duke@435: if (idx == AliasIdxTop) { duke@435: if (no_create) return NULL; duke@435: // Grow the array if necessary. duke@435: if (_num_alias_types == _max_alias_types) grow_alias_types(); duke@435: // Add a new alias type. duke@435: idx = _num_alias_types++; duke@435: _alias_types[idx]->Init(idx, flat); duke@435: if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false); duke@435: if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false); duke@435: if (flat->isa_instptr()) { duke@435: if (flat->offset() == java_lang_Class::klass_offset_in_bytes() duke@435: && flat->is_instptr()->klass() == env()->Class_klass()) duke@435: alias_type(idx)->set_rewritable(false); duke@435: } duke@435: if (flat->isa_klassptr()) { stefank@3391: if (flat->offset() == in_bytes(Klass::super_check_offset_offset())) duke@435: alias_type(idx)->set_rewritable(false); stefank@3391: if (flat->offset() == in_bytes(Klass::modifier_flags_offset())) duke@435: alias_type(idx)->set_rewritable(false); stefank@3391: if (flat->offset() == in_bytes(Klass::access_flags_offset())) duke@435: alias_type(idx)->set_rewritable(false); stefank@3391: if (flat->offset() == in_bytes(Klass::java_mirror_offset())) duke@435: alias_type(idx)->set_rewritable(false); duke@435: } duke@435: // %%% (We would like to finalize JavaThread::threadObj_offset(), duke@435: // but the base pointer type is not distinctive enough to identify duke@435: // references into JavaThread.) duke@435: never@2658: // Check for final fields. duke@435: const TypeInstPtr* tinst = flat->isa_instptr(); coleenp@548: if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) { never@2658: ciField* field; never@2658: if (tinst->const_oop() != NULL && never@2658: tinst->klass() == ciEnv::current()->Class_klass() && never@2658: tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) { never@2658: // static field never@2658: ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass(); never@2658: field = k->get_field_by_offset(tinst->offset(), true); never@2658: } else { never@2658: ciInstanceKlass *k = tinst->klass()->as_instance_klass(); never@2658: field = k->get_field_by_offset(tinst->offset(), false); never@2658: } never@2658: assert(field == NULL || never@2658: original_field == NULL || never@2658: (field->holder() == original_field->holder() && never@2658: field->offset() == original_field->offset() && never@2658: field->is_static() == original_field->is_static()), "wrong field?"); duke@435: // Set field() and is_rewritable() attributes. duke@435: if (field != NULL) alias_type(idx)->set_field(field); duke@435: } duke@435: } duke@435: duke@435: // Fill the cache for next time. duke@435: ace->_adr_type = adr_type; duke@435: ace->_index = idx; duke@435: assert(alias_type(adr_type) == alias_type(idx), "type must be installed"); duke@435: duke@435: // Might as well try to fill the cache for the flattened version, too. duke@435: AliasCacheEntry* face = probe_alias_cache(flat); duke@435: if (face->_adr_type == NULL) { duke@435: face->_adr_type = flat; duke@435: face->_index = idx; duke@435: assert(alias_type(flat) == alias_type(idx), "flat type must work too"); duke@435: } duke@435: duke@435: return alias_type(idx); duke@435: } duke@435: duke@435: duke@435: Compile::AliasType* Compile::alias_type(ciField* field) { duke@435: const TypeOopPtr* t; duke@435: if (field->is_static()) never@2658: t = TypeInstPtr::make(field->holder()->java_mirror()); duke@435: else duke@435: t = TypeOopPtr::make_from_klass_raw(field->holder()); never@2658: AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field); duke@435: assert(field->is_final() == !atp->is_rewritable(), "must get the rewritable bits correct"); duke@435: return atp; duke@435: } duke@435: duke@435: duke@435: //------------------------------have_alias_type-------------------------------- duke@435: bool Compile::have_alias_type(const TypePtr* adr_type) { duke@435: AliasCacheEntry* ace = probe_alias_cache(adr_type); duke@435: if (ace->_adr_type == adr_type) { duke@435: return true; duke@435: } duke@435: duke@435: // Handle special cases. duke@435: if (adr_type == NULL) return true; duke@435: if (adr_type == TypePtr::BOTTOM) return true; duke@435: never@2658: return find_alias_type(adr_type, true, NULL) != NULL; duke@435: } duke@435: duke@435: //-----------------------------must_alias-------------------------------------- duke@435: // True if all values of the given address type are in the given alias category. duke@435: bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) { duke@435: if (alias_idx == AliasIdxBot) return true; // the universal category duke@435: if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP duke@435: if (alias_idx == AliasIdxTop) return false; // the empty category duke@435: if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins duke@435: duke@435: // the only remaining possible overlap is identity duke@435: int adr_idx = get_alias_index(adr_type); duke@435: assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); duke@435: assert(adr_idx == alias_idx || duke@435: (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM duke@435: && adr_type != TypeOopPtr::BOTTOM), duke@435: "should not be testing for overlap with an unsafe pointer"); duke@435: return adr_idx == alias_idx; duke@435: } duke@435: duke@435: //------------------------------can_alias-------------------------------------- duke@435: // True if any values of the given address type are in the given alias category. duke@435: bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) { duke@435: if (alias_idx == AliasIdxTop) return false; // the empty category duke@435: if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP duke@435: if (alias_idx == AliasIdxBot) return true; // the universal category duke@435: if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins duke@435: duke@435: // the only remaining possible overlap is identity duke@435: int adr_idx = get_alias_index(adr_type); duke@435: assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, ""); duke@435: return adr_idx == alias_idx; duke@435: } duke@435: duke@435: duke@435: duke@435: //---------------------------pop_warm_call------------------------------------- duke@435: WarmCallInfo* Compile::pop_warm_call() { duke@435: WarmCallInfo* wci = _warm_calls; duke@435: if (wci != NULL) _warm_calls = wci->remove_from(wci); duke@435: return wci; duke@435: } duke@435: duke@435: //----------------------------Inline_Warm-------------------------------------- duke@435: int Compile::Inline_Warm() { duke@435: // If there is room, try to inline some more warm call sites. duke@435: // %%% Do a graph index compaction pass when we think we're out of space? duke@435: if (!InlineWarmCalls) return 0; duke@435: duke@435: int calls_made_hot = 0; duke@435: int room_to_grow = NodeCountInliningCutoff - unique(); duke@435: int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep); duke@435: int amount_grown = 0; duke@435: WarmCallInfo* call; duke@435: while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) { duke@435: int est_size = (int)call->size(); duke@435: if (est_size > (room_to_grow - amount_grown)) { duke@435: // This one won't fit anyway. Get rid of it. duke@435: call->make_cold(); duke@435: continue; duke@435: } duke@435: call->make_hot(); duke@435: calls_made_hot++; duke@435: amount_grown += est_size; duke@435: amount_to_grow -= est_size; duke@435: } duke@435: duke@435: if (calls_made_hot > 0) set_major_progress(); duke@435: return calls_made_hot; duke@435: } duke@435: duke@435: duke@435: //----------------------------Finish_Warm-------------------------------------- duke@435: void Compile::Finish_Warm() { duke@435: if (!InlineWarmCalls) return; duke@435: if (failing()) return; duke@435: if (warm_calls() == NULL) return; duke@435: duke@435: // Clean up loose ends, if we are out of space for inlining. duke@435: WarmCallInfo* call; duke@435: while ((call = pop_warm_call()) != NULL) { duke@435: call->make_cold(); duke@435: } duke@435: } duke@435: cfang@1607: //---------------------cleanup_loop_predicates----------------------- cfang@1607: // Remove the opaque nodes that protect the predicates so that all unused cfang@1607: // checks and uncommon_traps will be eliminated from the ideal graph cfang@1607: void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) { cfang@1607: if (predicate_count()==0) return; cfang@1607: for (int i = predicate_count(); i > 0; i--) { cfang@1607: Node * n = predicate_opaque1_node(i-1); cfang@1607: assert(n->Opcode() == Op_Opaque1, "must be"); cfang@1607: igvn.replace_node(n, n->in(1)); cfang@1607: } cfang@1607: assert(predicate_count()==0, "should be clean!"); cfang@1607: } duke@435: roland@4409: // StringOpts and late inlining of string methods roland@4409: void Compile::inline_string_calls(bool parse_time) { roland@4409: { roland@4409: // remove useless nodes to make the usage analysis simpler roland@4409: ResourceMark rm; roland@4409: PhaseRemoveUseless pru(initial_gvn(), for_igvn()); roland@4409: } roland@4409: roland@4409: { roland@4409: ResourceMark rm; sla@5237: print_method(PHASE_BEFORE_STRINGOPTS, 3); roland@4409: PhaseStringOpts pso(initial_gvn(), for_igvn()); sla@5237: print_method(PHASE_AFTER_STRINGOPTS, 3); roland@4409: } roland@4409: roland@4409: // now inline anything that we skipped the first time around roland@4409: if (!parse_time) { roland@4409: _late_inlines_pos = _late_inlines.length(); roland@4409: } roland@4409: roland@4409: while (_string_late_inlines.length() > 0) { roland@4409: CallGenerator* cg = _string_late_inlines.pop(); roland@4409: cg->do_late_inline(); roland@4409: if (failing()) return; roland@4409: } roland@4409: _string_late_inlines.trunc_to(0); roland@4409: } roland@4409: kvn@5110: // Late inlining of boxing methods kvn@5110: void Compile::inline_boxing_calls(PhaseIterGVN& igvn) { kvn@5110: if (_boxing_late_inlines.length() > 0) { kvn@5110: assert(has_boxed_value(), "inconsistent"); kvn@5110: kvn@5110: PhaseGVN* gvn = initial_gvn(); kvn@5110: set_inlining_incrementally(true); kvn@5110: kvn@5110: assert( igvn._worklist.size() == 0, "should be done with igvn" ); kvn@5110: for_igvn()->clear(); kvn@5110: gvn->replace_with(&igvn); kvn@5110: kvn@5110: while (_boxing_late_inlines.length() > 0) { kvn@5110: CallGenerator* cg = _boxing_late_inlines.pop(); kvn@5110: cg->do_late_inline(); kvn@5110: if (failing()) return; kvn@5110: } kvn@5110: _boxing_late_inlines.trunc_to(0); kvn@5110: kvn@5110: { kvn@5110: ResourceMark rm; kvn@5110: PhaseRemoveUseless pru(gvn, for_igvn()); kvn@5110: } kvn@5110: kvn@5110: igvn = PhaseIterGVN(gvn); kvn@5110: igvn.optimize(); kvn@5110: kvn@5110: set_inlining_progress(false); kvn@5110: set_inlining_incrementally(false); kvn@5110: } kvn@5110: } kvn@5110: roland@4409: void Compile::inline_incrementally_one(PhaseIterGVN& igvn) { roland@4409: assert(IncrementalInline, "incremental inlining should be on"); roland@4409: PhaseGVN* gvn = initial_gvn(); roland@4409: roland@4409: set_inlining_progress(false); roland@4409: for_igvn()->clear(); roland@4409: gvn->replace_with(&igvn); roland@4409: roland@4409: int i = 0; roland@4409: roland@4409: for (; i <_late_inlines.length() && !inlining_progress(); i++) { roland@4409: CallGenerator* cg = _late_inlines.at(i); roland@4409: _late_inlines_pos = i+1; roland@4409: cg->do_late_inline(); roland@4409: if (failing()) return; roland@4409: } roland@4409: int j = 0; roland@4409: for (; i < _late_inlines.length(); i++, j++) { roland@4409: _late_inlines.at_put(j, _late_inlines.at(i)); roland@4409: } roland@4409: _late_inlines.trunc_to(j); roland@4409: roland@4409: { roland@4409: ResourceMark rm; kvn@5110: PhaseRemoveUseless pru(gvn, for_igvn()); roland@4409: } roland@4409: roland@4409: igvn = PhaseIterGVN(gvn); roland@4409: } roland@4409: roland@4409: // Perform incremental inlining until bound on number of live nodes is reached roland@4409: void Compile::inline_incrementally(PhaseIterGVN& igvn) { roland@4409: PhaseGVN* gvn = initial_gvn(); roland@4409: roland@4409: set_inlining_incrementally(true); roland@4409: set_inlining_progress(true); roland@4409: uint low_live_nodes = 0; roland@4409: roland@4409: while(inlining_progress() && _late_inlines.length() > 0) { roland@4409: roland@4409: if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { roland@4409: if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) { roland@4409: // PhaseIdealLoop is expensive so we only try it once we are roland@4409: // out of loop and we only try it again if the previous helped roland@4409: // got the number of nodes down significantly roland@4409: PhaseIdealLoop ideal_loop( igvn, false, true ); roland@4409: if (failing()) return; roland@4409: low_live_nodes = live_nodes(); roland@4409: _major_progress = true; roland@4409: } roland@4409: roland@4409: if (live_nodes() > (uint)LiveNodeCountInliningCutoff) { roland@4409: break; roland@4409: } roland@4409: } roland@4409: roland@4409: inline_incrementally_one(igvn); roland@4409: roland@4409: if (failing()) return; roland@4409: roland@4409: igvn.optimize(); roland@4409: roland@4409: if (failing()) return; roland@4409: } roland@4409: roland@4409: assert( igvn._worklist.size() == 0, "should be done with igvn" ); roland@4409: roland@4409: if (_string_late_inlines.length() > 0) { roland@4409: assert(has_stringbuilder(), "inconsistent"); roland@4409: for_igvn()->clear(); roland@4409: initial_gvn()->replace_with(&igvn); roland@4409: roland@4409: inline_string_calls(false); roland@4409: roland@4409: if (failing()) return; roland@4409: roland@4409: { roland@4409: ResourceMark rm; roland@4409: PhaseRemoveUseless pru(initial_gvn(), for_igvn()); roland@4409: } roland@4409: roland@4409: igvn = PhaseIterGVN(gvn); roland@4409: roland@4409: igvn.optimize(); roland@4409: } roland@4409: roland@4409: set_inlining_incrementally(false); roland@4409: } roland@4409: roland@4409: duke@435: //------------------------------Optimize--------------------------------------- duke@435: // Given a graph, optimize it. duke@435: void Compile::Optimize() { duke@435: TracePhase t1("optimizer", &_t_optimizer, true); duke@435: duke@435: #ifndef PRODUCT duke@435: if (env()->break_at_compile()) { duke@435: BREAKPOINT; duke@435: } duke@435: duke@435: #endif duke@435: duke@435: ResourceMark rm; duke@435: int loop_opts_cnt; duke@435: duke@435: NOT_PRODUCT( verify_graph_edges(); ) duke@435: sla@5237: print_method(PHASE_AFTER_PARSING); duke@435: duke@435: { duke@435: // Iterative Global Value Numbering, including ideal transforms duke@435: // Initialize IterGVN with types and values from parse-time GVN duke@435: PhaseIterGVN igvn(initial_gvn()); duke@435: { duke@435: NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); ) duke@435: igvn.optimize(); duke@435: } duke@435: sla@5237: print_method(PHASE_ITER_GVN1, 2); duke@435: duke@435: if (failing()) return; duke@435: kvn@5110: { kvn@5110: NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); ) kvn@5110: inline_incrementally(igvn); kvn@5110: } roland@4409: sla@5237: print_method(PHASE_INCREMENTAL_INLINE, 2); roland@4409: roland@4409: if (failing()) return; roland@4409: kvn@5110: if (eliminate_boxing()) { kvn@5110: NOT_PRODUCT( TracePhase t2("incrementalInline", &_t_incrInline, TimeCompiler); ) kvn@5110: // Inline valueOf() methods now. kvn@5110: inline_boxing_calls(igvn); kvn@5110: sla@5237: print_method(PHASE_INCREMENTAL_BOXING_INLINE, 2); kvn@5110: kvn@5110: if (failing()) return; kvn@5110: } kvn@5110: roland@4589: // No more new expensive nodes will be added to the list from here roland@4589: // so keep only the actual candidates for optimizations. roland@4589: cleanup_expensive_nodes(igvn); roland@4589: kvn@1989: // Perform escape analysis kvn@1989: if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) { kvn@3260: if (has_loops()) { kvn@3260: // Cleanup graph (remove dead nodes). kvn@3260: TracePhase t2("idealLoop", &_t_idealLoop, true); kvn@3260: PhaseIdealLoop ideal_loop( igvn, false, true ); sla@5237: if (major_progress()) print_method(PHASE_PHASEIDEAL_BEFORE_EA, 2); kvn@3260: if (failing()) return; kvn@3260: } kvn@1989: ConnectionGraph::do_analysis(this, &igvn); kvn@1989: kvn@1989: if (failing()) return; kvn@1989: kvn@3311: // Optimize out fields loads from scalar replaceable allocations. kvn@1989: igvn.optimize(); sla@5237: print_method(PHASE_ITER_GVN_AFTER_EA, 2); kvn@1989: kvn@1989: if (failing()) return; kvn@1989: kvn@3311: if (congraph() != NULL && macro_count() > 0) { kvn@3651: NOT_PRODUCT( TracePhase t2("macroEliminate", &_t_macroEliminate, TimeCompiler); ) kvn@3311: PhaseMacroExpand mexp(igvn); kvn@3311: mexp.eliminate_macro_nodes(); kvn@3311: igvn.set_delay_transform(false); kvn@3311: kvn@3311: igvn.optimize(); sla@5237: print_method(PHASE_ITER_GVN_AFTER_ELIMINATION, 2); kvn@3311: kvn@3311: if (failing()) return; kvn@3311: } kvn@1989: } kvn@1989: duke@435: // Loop transforms on the ideal graph. Range Check Elimination, duke@435: // peeling, unrolling, etc. duke@435: duke@435: // Set loop opts counter duke@435: loop_opts_cnt = num_loop_opts(); duke@435: if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) { duke@435: { duke@435: TracePhase t2("idealLoop", &_t_idealLoop, true); kvn@2727: PhaseIdealLoop ideal_loop( igvn, true ); duke@435: loop_opts_cnt--; sla@5237: if (major_progress()) print_method(PHASE_PHASEIDEALLOOP1, 2); duke@435: if (failing()) return; duke@435: } duke@435: // Loop opts pass if partial peeling occurred in previous pass duke@435: if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) { duke@435: TracePhase t3("idealLoop", &_t_idealLoop, true); kvn@2727: PhaseIdealLoop ideal_loop( igvn, false ); duke@435: loop_opts_cnt--; sla@5237: if (major_progress()) print_method(PHASE_PHASEIDEALLOOP2, 2); duke@435: if (failing()) return; duke@435: } duke@435: // Loop opts pass for loop-unrolling before CCP duke@435: if(major_progress() && (loop_opts_cnt > 0)) { duke@435: TracePhase t4("idealLoop", &_t_idealLoop, true); kvn@2727: PhaseIdealLoop ideal_loop( igvn, false ); duke@435: loop_opts_cnt--; sla@5237: if (major_progress()) print_method(PHASE_PHASEIDEALLOOP3, 2); duke@435: } never@1356: if (!failing()) { never@1356: // Verify that last round of loop opts produced a valid graph never@1356: NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) never@1356: PhaseIdealLoop::verify(igvn); never@1356: } duke@435: } duke@435: if (failing()) return; duke@435: duke@435: // Conditional Constant Propagation; duke@435: PhaseCCP ccp( &igvn ); duke@435: assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)"); duke@435: { duke@435: TracePhase t2("ccp", &_t_ccp, true); duke@435: ccp.do_transform(); duke@435: } sla@5237: print_method(PHASE_CPP1, 2); duke@435: duke@435: assert( true, "Break here to ccp.dump_old2new_map()"); duke@435: duke@435: // Iterative Global Value Numbering, including ideal transforms duke@435: { duke@435: NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); ) duke@435: igvn = ccp; duke@435: igvn.optimize(); duke@435: } duke@435: sla@5237: print_method(PHASE_ITER_GVN2, 2); duke@435: duke@435: if (failing()) return; duke@435: duke@435: // Loop transforms on the ideal graph. Range Check Elimination, duke@435: // peeling, unrolling, etc. duke@435: if(loop_opts_cnt > 0) { duke@435: debug_only( int cnt = 0; ); duke@435: while(major_progress() && (loop_opts_cnt > 0)) { duke@435: TracePhase t2("idealLoop", &_t_idealLoop, true); duke@435: assert( cnt++ < 40, "infinite cycle in loop optimization" ); kvn@2727: PhaseIdealLoop ideal_loop( igvn, true); duke@435: loop_opts_cnt--; sla@5237: if (major_progress()) print_method(PHASE_PHASEIDEALLOOP_ITERATIONS, 2); duke@435: if (failing()) return; duke@435: } duke@435: } never@1356: never@1356: { never@1356: // Verify that all previous optimizations produced a valid graph never@1356: // at least to this point, even if no loop optimizations were done. never@1356: NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); ) never@1356: PhaseIdealLoop::verify(igvn); never@1356: } never@1356: duke@435: { duke@435: NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); ) duke@435: PhaseMacroExpand mex(igvn); duke@435: if (mex.expand_macro_nodes()) { duke@435: assert(failing(), "must bail out w/ explicit message"); duke@435: return; duke@435: } duke@435: } duke@435: duke@435: } // (End scope of igvn; run destructor if necessary for asserts.) duke@435: kvn@4448: dump_inlining(); duke@435: // A method with only infinite loops has no edges entering loops from root duke@435: { duke@435: NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); ) duke@435: if (final_graph_reshaping()) { duke@435: assert(failing(), "must bail out w/ explicit message"); duke@435: return; duke@435: } duke@435: } duke@435: sla@5237: print_method(PHASE_OPTIMIZE_FINISHED, 2); duke@435: } duke@435: duke@435: duke@435: //------------------------------Code_Gen--------------------------------------- duke@435: // Given a graph, generate code for it duke@435: void Compile::Code_Gen() { adlertz@5539: if (failing()) { adlertz@5539: return; adlertz@5539: } duke@435: duke@435: // Perform instruction selection. You might think we could reclaim Matcher duke@435: // memory PDQ, but actually the Matcher is used in generating spill code. duke@435: // Internals of the Matcher (including some VectorSets) must remain live duke@435: // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage duke@435: // set a bit in reclaimed memory. duke@435: duke@435: // In debug mode can dump m._nodes.dump() for mapping of ideal to machine duke@435: // nodes. Mapping is only valid at the root of each matched subtree. duke@435: NOT_PRODUCT( verify_graph_edges(); ) duke@435: adlertz@5539: Matcher matcher; adlertz@5539: _matcher = &matcher; duke@435: { duke@435: TracePhase t2("matcher", &_t_matcher, true); adlertz@5539: matcher.match(); duke@435: } duke@435: // In debug mode can dump m._nodes.dump() for mapping of ideal to machine duke@435: // nodes. Mapping is only valid at the root of each matched subtree. duke@435: NOT_PRODUCT( verify_graph_edges(); ) duke@435: duke@435: // If you have too many nodes, or if matching has failed, bail out duke@435: check_node_count(0, "out of nodes matching instructions"); adlertz@5539: if (failing()) { adlertz@5539: return; adlertz@5539: } duke@435: duke@435: // Build a proper-looking CFG adlertz@5539: PhaseCFG cfg(node_arena(), root(), matcher); duke@435: _cfg = &cfg; duke@435: { duke@435: NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); ) adlertz@5539: bool success = cfg.do_global_code_motion(); adlertz@5539: if (!success) { adlertz@5539: return; adlertz@5539: } adlertz@5539: adlertz@5539: print_method(PHASE_GLOBAL_CODE_MOTION, 2); duke@435: NOT_PRODUCT( verify_graph_edges(); ) duke@435: debug_only( cfg.verify(); ) duke@435: } adlertz@5539: adlertz@5539: PhaseChaitin regalloc(unique(), cfg, matcher); duke@435: _regalloc = ®alloc; duke@435: { duke@435: TracePhase t2("regalloc", &_t_registerAllocation, true); duke@435: // Perform register allocation. After Chaitin, use-def chains are duke@435: // no longer accurate (at spill code) and so must be ignored. duke@435: // Node->LRG->reg mappings are still accurate. duke@435: _regalloc->Register_Allocate(); duke@435: duke@435: // Bail out if the allocator builds too many nodes neliasso@4949: if (failing()) { neliasso@4949: return; neliasso@4949: } duke@435: } duke@435: duke@435: // Prior to register allocation we kept empty basic blocks in case the duke@435: // the allocator needed a place to spill. After register allocation we duke@435: // are not adding any new instructions. If any basic block is empty, we duke@435: // can now safely remove it. duke@435: { rasbold@853: NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); ) adlertz@5539: cfg.remove_empty_blocks(); rasbold@853: if (do_freq_based_layout()) { rasbold@853: PhaseBlockLayout layout(cfg); rasbold@853: } else { rasbold@853: cfg.set_loop_alignment(); rasbold@853: } rasbold@853: cfg.fixup_flow(); duke@435: } duke@435: duke@435: // Apply peephole optimizations duke@435: if( OptoPeephole ) { duke@435: NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); ) duke@435: PhasePeephole peep( _regalloc, cfg); duke@435: peep.do_transform(); duke@435: } duke@435: duke@435: // Convert Nodes to instruction bits in a buffer duke@435: { duke@435: // %%%% workspace merge brought two timers together for one job duke@435: TracePhase t2a("output", &_t_output, true); duke@435: NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); ) duke@435: Output(); duke@435: } duke@435: sla@5237: print_method(PHASE_FINAL_CODE); duke@435: duke@435: // He's dead, Jim. duke@435: _cfg = (PhaseCFG*)0xdeadbeef; duke@435: _regalloc = (PhaseChaitin*)0xdeadbeef; duke@435: } duke@435: duke@435: duke@435: //------------------------------dump_asm--------------------------------------- duke@435: // Dump formatted assembly duke@435: #ifndef PRODUCT duke@435: void Compile::dump_asm(int *pcs, uint pc_limit) { duke@435: bool cut_short = false; duke@435: tty->print_cr("#"); duke@435: tty->print("# "); _tf->dump(); tty->cr(); duke@435: tty->print_cr("#"); duke@435: duke@435: // For all blocks duke@435: int pc = 0x0; // Program counter duke@435: char starts_bundle = ' '; duke@435: _regalloc->dump_frame(); duke@435: duke@435: Node *n = NULL; adlertz@5539: for (uint i = 0; i < _cfg->number_of_blocks(); i++) { adlertz@5539: if (VMThread::should_terminate()) { adlertz@5539: cut_short = true; adlertz@5539: break; adlertz@5539: } adlertz@5539: Block* block = _cfg->get_block(i); adlertz@5539: if (block->is_connector() && !Verbose) { adlertz@5539: continue; adlertz@5539: } adlertz@5539: n = block->_nodes[0]; adlertz@5539: if (pcs && n->_idx < pc_limit) { duke@435: tty->print("%3.3x ", pcs[n->_idx]); adlertz@5539: } else { duke@435: tty->print(" "); adlertz@5539: } adlertz@5539: block->dump_head(_cfg); adlertz@5539: if (block->is_connector()) { duke@435: tty->print_cr(" # Empty connector block"); adlertz@5539: } else if (block->num_preds() == 2 && block->pred(1)->is_CatchProj() && block->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) { duke@435: tty->print_cr(" # Block is sole successor of call"); duke@435: } duke@435: duke@435: // For all instructions duke@435: Node *delay = NULL; adlertz@5539: for (uint j = 0; j < block->_nodes.size(); j++) { adlertz@5539: if (VMThread::should_terminate()) { adlertz@5539: cut_short = true; adlertz@5539: break; adlertz@5539: } adlertz@5539: n = block->_nodes[j]; duke@435: if (valid_bundle_info(n)) { adlertz@5539: Bundle* bundle = node_bundling(n); duke@435: if (bundle->used_in_unconditional_delay()) { duke@435: delay = n; duke@435: continue; duke@435: } adlertz@5539: if (bundle->starts_bundle()) { duke@435: starts_bundle = '+'; adlertz@5539: } duke@435: } duke@435: adlertz@5539: if (WizardMode) { adlertz@5539: n->dump(); adlertz@5539: } coleenp@548: duke@435: if( !n->is_Region() && // Dont print in the Assembly duke@435: !n->is_Phi() && // a few noisely useless nodes duke@435: !n->is_Proj() && duke@435: !n->is_MachTemp() && kvn@1535: !n->is_SafePointScalarObject() && duke@435: !n->is_Catch() && // Would be nice to print exception table targets duke@435: !n->is_MergeMem() && // Not very interesting duke@435: !n->is_top() && // Debug info table constants duke@435: !(n->is_Con() && !n->is_Mach())// Debug info table constants duke@435: ) { duke@435: if (pcs && n->_idx < pc_limit) duke@435: tty->print("%3.3x", pcs[n->_idx]); duke@435: else duke@435: tty->print(" "); duke@435: tty->print(" %c ", starts_bundle); duke@435: starts_bundle = ' '; duke@435: tty->print("\t"); duke@435: n->format(_regalloc, tty); duke@435: tty->cr(); duke@435: } duke@435: duke@435: // If we have an instruction with a delay slot, and have seen a delay, duke@435: // then back up and print it duke@435: if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) { duke@435: assert(delay != NULL, "no unconditional delay instruction"); coleenp@548: if (WizardMode) delay->dump(); coleenp@548: duke@435: if (node_bundling(delay)->starts_bundle()) duke@435: starts_bundle = '+'; duke@435: if (pcs && n->_idx < pc_limit) duke@435: tty->print("%3.3x", pcs[n->_idx]); duke@435: else duke@435: tty->print(" "); duke@435: tty->print(" %c ", starts_bundle); duke@435: starts_bundle = ' '; duke@435: tty->print("\t"); duke@435: delay->format(_regalloc, tty); duke@435: tty->print_cr(""); duke@435: delay = NULL; duke@435: } duke@435: duke@435: // Dump the exception table as well duke@435: if( n->is_Catch() && (Verbose || WizardMode) ) { duke@435: // Print the exception table for this offset duke@435: _handler_table.print_subtable_for(pc); duke@435: } duke@435: } duke@435: duke@435: if (pcs && n->_idx < pc_limit) duke@435: tty->print_cr("%3.3x", pcs[n->_idx]); duke@435: else duke@435: tty->print_cr(""); duke@435: duke@435: assert(cut_short || delay == NULL, "no unconditional delay branch"); duke@435: duke@435: } // End of per-block dump duke@435: tty->print_cr(""); duke@435: duke@435: if (cut_short) tty->print_cr("*** disassembly is cut short ***"); duke@435: } duke@435: #endif duke@435: duke@435: //------------------------------Final_Reshape_Counts--------------------------- duke@435: // This class defines counters to help identify when a method duke@435: // may/must be executed using hardware with only 24-bit precision. duke@435: struct Final_Reshape_Counts : public StackObj { duke@435: int _call_count; // count non-inlined 'common' calls duke@435: int _float_count; // count float ops requiring 24-bit precision duke@435: int _double_count; // count double ops requiring more precision duke@435: int _java_call_count; // count non-inlined 'java' calls kvn@1294: int _inner_loop_count; // count loops which need alignment duke@435: VectorSet _visited; // Visitation flags duke@435: Node_List _tests; // Set of IfNodes & PCTableNodes duke@435: duke@435: Final_Reshape_Counts() : kvn@1294: _call_count(0), _float_count(0), _double_count(0), kvn@1294: _java_call_count(0), _inner_loop_count(0), duke@435: _visited( Thread::current()->resource_area() ) { } duke@435: duke@435: void inc_call_count () { _call_count ++; } duke@435: void inc_float_count () { _float_count ++; } duke@435: void inc_double_count() { _double_count++; } duke@435: void inc_java_call_count() { _java_call_count++; } kvn@1294: void inc_inner_loop_count() { _inner_loop_count++; } duke@435: duke@435: int get_call_count () const { return _call_count ; } duke@435: int get_float_count () const { return _float_count ; } duke@435: int get_double_count() const { return _double_count; } duke@435: int get_java_call_count() const { return _java_call_count; } kvn@1294: int get_inner_loop_count() const { return _inner_loop_count; } duke@435: }; duke@435: mikael@4889: #ifdef ASSERT duke@435: static bool oop_offset_is_sane(const TypeInstPtr* tp) { duke@435: ciInstanceKlass *k = tp->klass()->as_instance_klass(); duke@435: // Make sure the offset goes inside the instance layout. coleenp@548: return k->contains_field_offset(tp->offset()); duke@435: // Note that OffsetBot and OffsetTop are very negative. duke@435: } mikael@4889: #endif duke@435: never@2780: // Eliminate trivially redundant StoreCMs and accumulate their never@2780: // precedence edges. bharadwaj@4315: void Compile::eliminate_redundant_card_marks(Node* n) { never@2780: assert(n->Opcode() == Op_StoreCM, "expected StoreCM"); never@2780: if (n->in(MemNode::Address)->outcnt() > 1) { never@2780: // There are multiple users of the same address so it might be never@2780: // possible to eliminate some of the StoreCMs never@2780: Node* mem = n->in(MemNode::Memory); never@2780: Node* adr = n->in(MemNode::Address); never@2780: Node* val = n->in(MemNode::ValueIn); never@2780: Node* prev = n; never@2780: bool done = false; never@2780: // Walk the chain of StoreCMs eliminating ones that match. As never@2780: // long as it's a chain of single users then the optimization is never@2780: // safe. Eliminating partially redundant StoreCMs would require never@2780: // cloning copies down the other paths. never@2780: while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) { never@2780: if (adr == mem->in(MemNode::Address) && never@2780: val == mem->in(MemNode::ValueIn)) { never@2780: // redundant StoreCM never@2780: if (mem->req() > MemNode::OopStore) { never@2780: // Hasn't been processed by this code yet. never@2780: n->add_prec(mem->in(MemNode::OopStore)); never@2780: } else { never@2780: // Already converted to precedence edge never@2780: for (uint i = mem->req(); i < mem->len(); i++) { never@2780: // Accumulate any precedence edges never@2780: if (mem->in(i) != NULL) { never@2780: n->add_prec(mem->in(i)); never@2780: } never@2780: } never@2780: // Everything above this point has been processed. never@2780: done = true; never@2780: } never@2780: // Eliminate the previous StoreCM never@2780: prev->set_req(MemNode::Memory, mem->in(MemNode::Memory)); never@2780: assert(mem->outcnt() == 0, "should be dead"); bharadwaj@4315: mem->disconnect_inputs(NULL, this); never@2780: } else { never@2780: prev = mem; never@2780: } never@2780: mem = prev->in(MemNode::Memory); never@2780: } never@2780: } never@2780: } never@2780: duke@435: //------------------------------final_graph_reshaping_impl---------------------- duke@435: // Implement items 1-5 from final_graph_reshaping below. bharadwaj@4315: void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) { duke@435: kvn@603: if ( n->outcnt() == 0 ) return; // dead node duke@435: uint nop = n->Opcode(); duke@435: duke@435: // Check for 2-input instruction with "last use" on right input. duke@435: // Swap to left input. Implements item (2). duke@435: if( n->req() == 3 && // two-input instruction duke@435: n->in(1)->outcnt() > 1 && // left use is NOT a last use duke@435: (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop duke@435: n->in(2)->outcnt() == 1 &&// right use IS a last use duke@435: !n->in(2)->is_Con() ) { // right use is not a constant duke@435: // Check for commutative opcode duke@435: switch( nop ) { duke@435: case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL: duke@435: case Op_MaxI: case Op_MinI: duke@435: case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL: duke@435: case Op_AndL: case Op_XorL: case Op_OrL: duke@435: case Op_AndI: case Op_XorI: case Op_OrI: { duke@435: // Move "last use" input to left by swapping inputs duke@435: n->swap_edges(1, 2); duke@435: break; duke@435: } duke@435: default: duke@435: break; duke@435: } duke@435: } duke@435: kvn@1964: #ifdef ASSERT kvn@1964: if( n->is_Mem() ) { bharadwaj@4315: int alias_idx = get_alias_index(n->as_Mem()->adr_type()); kvn@1964: assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw || kvn@1964: // oop will be recorded in oop map if load crosses safepoint kvn@1964: n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() || kvn@1964: LoadNode::is_immutable_value(n->in(MemNode::Address))), kvn@1964: "raw memory operations should have control edge"); kvn@1964: } kvn@1964: #endif duke@435: // Count FPU ops and common calls, implements item (3) duke@435: switch( nop ) { duke@435: // Count all float operations that may use FPU duke@435: case Op_AddF: duke@435: case Op_SubF: duke@435: case Op_MulF: duke@435: case Op_DivF: duke@435: case Op_NegF: duke@435: case Op_ModF: duke@435: case Op_ConvI2F: duke@435: case Op_ConF: duke@435: case Op_CmpF: duke@435: case Op_CmpF3: duke@435: // case Op_ConvL2F: // longs are split into 32-bit halves kvn@1294: frc.inc_float_count(); duke@435: break; duke@435: duke@435: case Op_ConvF2D: duke@435: case Op_ConvD2F: kvn@1294: frc.inc_float_count(); kvn@1294: frc.inc_double_count(); duke@435: break; duke@435: duke@435: // Count all double operations that may use FPU duke@435: case Op_AddD: duke@435: case Op_SubD: duke@435: case Op_MulD: duke@435: case Op_DivD: duke@435: case Op_NegD: duke@435: case Op_ModD: duke@435: case Op_ConvI2D: duke@435: case Op_ConvD2I: duke@435: // case Op_ConvL2D: // handled by leaf call duke@435: // case Op_ConvD2L: // handled by leaf call duke@435: case Op_ConD: duke@435: case Op_CmpD: duke@435: case Op_CmpD3: kvn@1294: frc.inc_double_count(); duke@435: break; duke@435: case Op_Opaque1: // Remove Opaque Nodes before matching duke@435: case Op_Opaque2: // Remove Opaque Nodes before matching bharadwaj@4315: n->subsume_by(n->in(1), this); duke@435: break; duke@435: case Op_CallStaticJava: duke@435: case Op_CallJava: duke@435: case Op_CallDynamicJava: kvn@1294: frc.inc_java_call_count(); // Count java call site; duke@435: case Op_CallRuntime: duke@435: case Op_CallLeaf: duke@435: case Op_CallLeafNoFP: { duke@435: assert( n->is_Call(), "" ); duke@435: CallNode *call = n->as_Call(); duke@435: // Count call sites where the FP mode bit would have to be flipped. duke@435: // Do not count uncommon runtime calls: duke@435: // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking, duke@435: // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ... duke@435: if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) { kvn@1294: frc.inc_call_count(); // Count the call site duke@435: } else { // See if uncommon argument is shared duke@435: Node *n = call->in(TypeFunc::Parms); duke@435: int nop = n->Opcode(); duke@435: // Clone shared simple arguments to uncommon calls, item (1). duke@435: if( n->outcnt() > 1 && duke@435: !n->is_Proj() && duke@435: nop != Op_CreateEx && duke@435: nop != Op_CheckCastPP && kvn@766: nop != Op_DecodeN && roland@4159: nop != Op_DecodeNKlass && duke@435: !n->is_Mem() ) { duke@435: Node *x = n->clone(); duke@435: call->set_req( TypeFunc::Parms, x ); duke@435: } duke@435: } duke@435: break; duke@435: } duke@435: duke@435: case Op_StoreD: duke@435: case Op_LoadD: duke@435: case Op_LoadD_unaligned: kvn@1294: frc.inc_double_count(); duke@435: goto handle_mem; duke@435: case Op_StoreF: duke@435: case Op_LoadF: kvn@1294: frc.inc_float_count(); duke@435: goto handle_mem; duke@435: never@2780: case Op_StoreCM: never@2780: { never@2780: // Convert OopStore dependence into precedence edge never@2780: Node* prec = n->in(MemNode::OopStore); never@2780: n->del_req(MemNode::OopStore); never@2780: n->add_prec(prec); never@2780: eliminate_redundant_card_marks(n); never@2780: } never@2780: never@2780: // fall through never@2780: duke@435: case Op_StoreB: duke@435: case Op_StoreC: duke@435: case Op_StorePConditional: duke@435: case Op_StoreI: duke@435: case Op_StoreL: kvn@855: case Op_StoreIConditional: duke@435: case Op_StoreLConditional: duke@435: case Op_CompareAndSwapI: duke@435: case Op_CompareAndSwapL: duke@435: case Op_CompareAndSwapP: coleenp@548: case Op_CompareAndSwapN: roland@4106: case Op_GetAndAddI: roland@4106: case Op_GetAndAddL: roland@4106: case Op_GetAndSetI: roland@4106: case Op_GetAndSetL: roland@4106: case Op_GetAndSetP: roland@4106: case Op_GetAndSetN: duke@435: case Op_StoreP: coleenp@548: case Op_StoreN: roland@4159: case Op_StoreNKlass: duke@435: case Op_LoadB: twisti@1059: case Op_LoadUB: twisti@993: case Op_LoadUS: duke@435: case Op_LoadI: duke@435: case Op_LoadKlass: kvn@599: case Op_LoadNKlass: duke@435: case Op_LoadL: duke@435: case Op_LoadL_unaligned: duke@435: case Op_LoadPLocked: duke@435: case Op_LoadP: coleenp@548: case Op_LoadN: duke@435: case Op_LoadRange: duke@435: case Op_LoadS: { duke@435: handle_mem: duke@435: #ifdef ASSERT duke@435: if( VerifyOptoOopOffsets ) { duke@435: assert( n->is_Mem(), "" ); duke@435: MemNode *mem = (MemNode*)n; duke@435: // Check to see if address types have grounded out somehow. duke@435: const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr(); duke@435: assert( !tp || oop_offset_is_sane(tp), "" ); duke@435: } duke@435: #endif duke@435: break; duke@435: } duke@435: duke@435: case Op_AddP: { // Assert sane base pointers kvn@617: Node *addp = n->in(AddPNode::Address); duke@435: assert( !addp->is_AddP() || duke@435: addp->in(AddPNode::Base)->is_top() || // Top OK for allocation duke@435: addp->in(AddPNode::Base) == n->in(AddPNode::Base), duke@435: "Base pointers must match" ); kvn@617: #ifdef _LP64 roland@4159: if ((UseCompressedOops || UseCompressedKlassPointers) && kvn@617: addp->Opcode() == Op_ConP && kvn@617: addp == n->in(AddPNode::Base) && kvn@617: n->in(AddPNode::Offset)->is_Con()) { kvn@617: // Use addressing with narrow klass to load with offset on x86. kvn@617: // On sparc loading 32-bits constant and decoding it have less kvn@617: // instructions (4) then load 64-bits constant (7). kvn@617: // Do this transformation here since IGVN will convert ConN back to ConP. kvn@617: const Type* t = addp->bottom_type(); roland@4159: if (t->isa_oopptr() || t->isa_klassptr()) { kvn@617: Node* nn = NULL; kvn@617: roland@4159: int op = t->isa_oopptr() ? Op_ConN : Op_ConNKlass; roland@4159: kvn@617: // Look for existing ConN node of the same exact type. bharadwaj@4315: Node* r = root(); kvn@617: uint cnt = r->outcnt(); kvn@617: for (uint i = 0; i < cnt; i++) { kvn@617: Node* m = r->raw_out(i); roland@4159: if (m!= NULL && m->Opcode() == op && kvn@656: m->bottom_type()->make_ptr() == t) { kvn@617: nn = m; kvn@617: break; kvn@617: } kvn@617: } kvn@617: if (nn != NULL) { kvn@617: // Decode a narrow oop to match address kvn@617: // [R12 + narrow_oop_reg<<3 + offset] roland@4159: if (t->isa_oopptr()) { bharadwaj@4315: nn = new (this) DecodeNNode(nn, t); roland@4159: } else { bharadwaj@4315: nn = new (this) DecodeNKlassNode(nn, t); roland@4159: } kvn@617: n->set_req(AddPNode::Base, nn); kvn@617: n->set_req(AddPNode::Address, nn); kvn@617: if (addp->outcnt() == 0) { bharadwaj@4315: addp->disconnect_inputs(NULL, this); kvn@617: } kvn@617: } kvn@617: } kvn@617: } kvn@617: #endif duke@435: break; duke@435: } duke@435: kvn@599: #ifdef _LP64 kvn@803: case Op_CastPP: kvn@1930: if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) { kvn@803: Node* in1 = n->in(1); kvn@803: const Type* t = n->bottom_type(); kvn@803: Node* new_in1 = in1->clone(); kvn@803: new_in1->as_DecodeN()->set_type(t); kvn@803: kvn@1930: if (!Matcher::narrow_oop_use_complex_address()) { kvn@803: // kvn@803: // x86, ARM and friends can handle 2 adds in addressing mode kvn@803: // and Matcher can fold a DecodeN node into address by using kvn@803: // a narrow oop directly and do implicit NULL check in address: kvn@803: // kvn@803: // [R12 + narrow_oop_reg<<3 + offset] kvn@803: // NullCheck narrow_oop_reg kvn@803: // kvn@803: // On other platforms (Sparc) we have to keep new DecodeN node and kvn@803: // use it to do implicit NULL check in address: kvn@803: // kvn@803: // decode_not_null narrow_oop_reg, base_reg kvn@803: // [base_reg + offset] kvn@803: // NullCheck base_reg kvn@803: // twisti@1040: // Pin the new DecodeN node to non-null path on these platform (Sparc) kvn@803: // to keep the information to which NULL check the new DecodeN node kvn@803: // corresponds to use it as value in implicit_null_check(). kvn@803: // kvn@803: new_in1->set_req(0, n->in(0)); kvn@803: } kvn@803: bharadwaj@4315: n->subsume_by(new_in1, this); kvn@803: if (in1->outcnt() == 0) { bharadwaj@4315: in1->disconnect_inputs(NULL, this); kvn@803: } kvn@803: } kvn@803: break; kvn@803: kvn@599: case Op_CmpP: kvn@603: // Do this transformation here to preserve CmpPNode::sub() and kvn@603: // other TypePtr related Ideal optimizations (for example, ptr nullness). roland@4159: if (n->in(1)->is_DecodeNarrowPtr() || n->in(2)->is_DecodeNarrowPtr()) { kvn@766: Node* in1 = n->in(1); kvn@766: Node* in2 = n->in(2); roland@4159: if (!in1->is_DecodeNarrowPtr()) { kvn@766: in2 = in1; kvn@766: in1 = n->in(2); kvn@766: } roland@4159: assert(in1->is_DecodeNarrowPtr(), "sanity"); kvn@766: kvn@766: Node* new_in2 = NULL; roland@4159: if (in2->is_DecodeNarrowPtr()) { roland@4159: assert(in2->Opcode() == in1->Opcode(), "must be same node type"); kvn@766: new_in2 = in2->in(1); kvn@766: } else if (in2->Opcode() == Op_ConP) { kvn@766: const Type* t = in2->bottom_type(); kvn@1930: if (t == TypePtr::NULL_PTR) { roland@4159: assert(in1->is_DecodeN(), "compare klass to null?"); kvn@1930: // Don't convert CmpP null check into CmpN if compressed kvn@1930: // oops implicit null check is not generated. kvn@1930: // This will allow to generate normal oop implicit null check. kvn@1930: if (Matcher::gen_narrow_oop_implicit_null_checks()) bharadwaj@4315: new_in2 = ConNode::make(this, TypeNarrowOop::NULL_PTR); kvn@803: // kvn@803: // This transformation together with CastPP transformation above kvn@803: // will generated code for implicit NULL checks for compressed oops. kvn@803: // kvn@803: // The original code after Optimize() kvn@803: // kvn@803: // LoadN memory, narrow_oop_reg kvn@803: // decode narrow_oop_reg, base_reg kvn@803: // CmpP base_reg, NULL kvn@803: // CastPP base_reg // NotNull kvn@803: // Load [base_reg + offset], val_reg kvn@803: // kvn@803: // after these transformations will be kvn@803: // kvn@803: // LoadN memory, narrow_oop_reg kvn@803: // CmpN narrow_oop_reg, NULL kvn@803: // decode_not_null narrow_oop_reg, base_reg kvn@803: // Load [base_reg + offset], val_reg kvn@803: // kvn@803: // and the uncommon path (== NULL) will use narrow_oop_reg directly kvn@803: // since narrow oops can be used in debug info now (see the code in kvn@803: // final_graph_reshaping_walk()). kvn@803: // kvn@803: // At the end the code will be matched to kvn@803: // on x86: kvn@803: // kvn@803: // Load_narrow_oop memory, narrow_oop_reg kvn@803: // Load [R12 + narrow_oop_reg<<3 + offset], val_reg kvn@803: // NullCheck narrow_oop_reg kvn@803: // kvn@803: // and on sparc: kvn@803: // kvn@803: // Load_narrow_oop memory, narrow_oop_reg kvn@803: // decode_not_null narrow_oop_reg, base_reg kvn@803: // Load [base_reg + offset], val_reg kvn@803: // NullCheck base_reg kvn@803: // kvn@599: } else if (t->isa_oopptr()) { bharadwaj@4315: new_in2 = ConNode::make(this, t->make_narrowoop()); roland@4159: } else if (t->isa_klassptr()) { bharadwaj@4315: new_in2 = ConNode::make(this, t->make_narrowklass()); kvn@599: } kvn@599: } kvn@766: if (new_in2 != NULL) { bharadwaj@4315: Node* cmpN = new (this) CmpNNode(in1->in(1), new_in2); bharadwaj@4315: n->subsume_by(cmpN, this); kvn@766: if (in1->outcnt() == 0) { bharadwaj@4315: in1->disconnect_inputs(NULL, this); kvn@766: } kvn@766: if (in2->outcnt() == 0) { bharadwaj@4315: in2->disconnect_inputs(NULL, this); kvn@766: } kvn@599: } kvn@599: } kvn@728: break; kvn@803: kvn@803: case Op_DecodeN: roland@4159: case Op_DecodeNKlass: roland@4159: assert(!n->in(1)->is_EncodeNarrowPtr(), "should be optimized out"); kvn@1930: // DecodeN could be pinned when it can't be fold into kvn@927: // an address expression, see the code for Op_CastPP above. roland@4159: assert(n->in(0) == NULL || (UseCompressedOops && !Matcher::narrow_oop_use_complex_address()), "no control"); kvn@803: break; kvn@803: roland@4159: case Op_EncodeP: roland@4159: case Op_EncodePKlass: { kvn@803: Node* in1 = n->in(1); roland@4159: if (in1->is_DecodeNarrowPtr()) { bharadwaj@4315: n->subsume_by(in1->in(1), this); kvn@803: } else if (in1->Opcode() == Op_ConP) { kvn@803: const Type* t = in1->bottom_type(); kvn@803: if (t == TypePtr::NULL_PTR) { roland@4159: assert(t->isa_oopptr(), "null klass?"); bharadwaj@4315: n->subsume_by(ConNode::make(this, TypeNarrowOop::NULL_PTR), this); kvn@803: } else if (t->isa_oopptr()) { bharadwaj@4315: n->subsume_by(ConNode::make(this, t->make_narrowoop()), this); roland@4159: } else if (t->isa_klassptr()) { bharadwaj@4315: n->subsume_by(ConNode::make(this, t->make_narrowklass()), this); kvn@803: } kvn@803: } kvn@803: if (in1->outcnt() == 0) { bharadwaj@4315: in1->disconnect_inputs(NULL, this); kvn@803: } kvn@803: break; kvn@803: } kvn@803: never@1515: case Op_Proj: { never@1515: if (OptimizeStringConcat) { never@1515: ProjNode* p = n->as_Proj(); never@1515: if (p->_is_io_use) { never@1515: // Separate projections were used for the exception path which never@1515: // are normally removed by a late inline. If it wasn't inlined never@1515: // then they will hang around and should just be replaced with never@1515: // the original one. never@1515: Node* proj = NULL; never@1515: // Replace with just one never@1515: for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) { never@1515: Node *use = i.get(); never@1515: if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) { never@1515: proj = use; never@1515: break; never@1515: } never@1515: } kvn@3396: assert(proj != NULL, "must be found"); bharadwaj@4315: p->subsume_by(proj, this); never@1515: } never@1515: } never@1515: break; never@1515: } never@1515: kvn@803: case Op_Phi: roland@4159: if (n->as_Phi()->bottom_type()->isa_narrowoop() || n->as_Phi()->bottom_type()->isa_narrowklass()) { kvn@803: // The EncodeP optimization may create Phi with the same edges kvn@803: // for all paths. It is not handled well by Register Allocator. kvn@803: Node* unique_in = n->in(1); kvn@803: assert(unique_in != NULL, ""); kvn@803: uint cnt = n->req(); kvn@803: for (uint i = 2; i < cnt; i++) { kvn@803: Node* m = n->in(i); kvn@803: assert(m != NULL, ""); kvn@803: if (unique_in != m) kvn@803: unique_in = NULL; kvn@803: } kvn@803: if (unique_in != NULL) { bharadwaj@4315: n->subsume_by(unique_in, this); kvn@803: } kvn@803: } kvn@803: break; kvn@803: kvn@599: #endif kvn@599: duke@435: case Op_ModI: duke@435: if (UseDivMod) { duke@435: // Check if a%b and a/b both exist duke@435: Node* d = n->find_similar(Op_DivI); duke@435: if (d) { duke@435: // Replace them with a fused divmod if supported duke@435: if (Matcher::has_match_rule(Op_DivModI)) { bharadwaj@4315: DivModINode* divmod = DivModINode::make(this, n); bharadwaj@4315: d->subsume_by(divmod->div_proj(), this); bharadwaj@4315: n->subsume_by(divmod->mod_proj(), this); duke@435: } else { duke@435: // replace a%b with a-((a/b)*b) bharadwaj@4315: Node* mult = new (this) MulINode(d, d->in(2)); bharadwaj@4315: Node* sub = new (this) SubINode(d->in(1), mult); bharadwaj@4315: n->subsume_by(sub, this); duke@435: } duke@435: } duke@435: } duke@435: break; duke@435: duke@435: case Op_ModL: duke@435: if (UseDivMod) { duke@435: // Check if a%b and a/b both exist duke@435: Node* d = n->find_similar(Op_DivL); duke@435: if (d) { duke@435: // Replace them with a fused divmod if supported duke@435: if (Matcher::has_match_rule(Op_DivModL)) { bharadwaj@4315: DivModLNode* divmod = DivModLNode::make(this, n); bharadwaj@4315: d->subsume_by(divmod->div_proj(), this); bharadwaj@4315: n->subsume_by(divmod->mod_proj(), this); duke@435: } else { duke@435: // replace a%b with a-((a/b)*b) bharadwaj@4315: Node* mult = new (this) MulLNode(d, d->in(2)); bharadwaj@4315: Node* sub = new (this) SubLNode(d->in(1), mult); bharadwaj@4315: n->subsume_by(sub, this); duke@435: } duke@435: } duke@435: } duke@435: break; duke@435: kvn@3882: case Op_LoadVector: kvn@3882: case Op_StoreVector: duke@435: break; duke@435: duke@435: case Op_PackB: duke@435: case Op_PackS: duke@435: case Op_PackI: duke@435: case Op_PackF: duke@435: case Op_PackL: duke@435: case Op_PackD: duke@435: if (n->req()-1 > 2) { duke@435: // Replace many operand PackNodes with a binary tree for matching duke@435: PackNode* p = (PackNode*) n; bharadwaj@4315: Node* btp = p->binary_tree_pack(this, 1, n->req()); bharadwaj@4315: n->subsume_by(btp, this); duke@435: } duke@435: break; kvn@1294: case Op_Loop: kvn@1294: case Op_CountedLoop: kvn@1294: if (n->as_Loop()->is_inner_loop()) { kvn@1294: frc.inc_inner_loop_count(); kvn@1294: } kvn@1294: break; roland@2683: case Op_LShiftI: roland@2683: case Op_RShiftI: roland@2683: case Op_URShiftI: roland@2683: case Op_LShiftL: roland@2683: case Op_RShiftL: roland@2683: case Op_URShiftL: roland@2683: if (Matcher::need_masked_shift_count) { roland@2683: // The cpu's shift instructions don't restrict the count to the roland@2683: // lower 5/6 bits. We need to do the masking ourselves. roland@2683: Node* in2 = n->in(2); roland@2683: juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1); roland@2683: const TypeInt* t = in2->find_int_type(); roland@2683: if (t != NULL && t->is_con()) { roland@2683: juint shift = t->get_con(); roland@2683: if (shift > mask) { // Unsigned cmp bharadwaj@4315: n->set_req(2, ConNode::make(this, TypeInt::make(shift & mask))); roland@2683: } roland@2683: } else { roland@2683: if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) { bharadwaj@4315: Node* shift = new (this) AndINode(in2, ConNode::make(this, TypeInt::make(mask))); roland@2683: n->set_req(2, shift); roland@2683: } roland@2683: } roland@2683: if (in2->outcnt() == 0) { // Remove dead node bharadwaj@4315: in2->disconnect_inputs(NULL, this); roland@2683: } roland@2683: } roland@2683: break; roland@4694: case Op_MemBarStoreStore: kvn@5110: case Op_MemBarRelease: roland@4694: // Break the link with AllocateNode: it is no longer useful and roland@4694: // confuses register allocation. roland@4694: if (n->req() > MemBarNode::Precedent) { roland@4694: n->set_req(MemBarNode::Precedent, top()); roland@4694: } roland@4694: break; duke@435: default: duke@435: assert( !n->is_Call(), "" ); duke@435: assert( !n->is_Mem(), "" ); duke@435: break; duke@435: } never@562: never@562: // Collect CFG split points never@562: if (n->is_MultiBranch()) kvn@1294: frc._tests.push(n); duke@435: } duke@435: duke@435: //------------------------------final_graph_reshaping_walk--------------------- duke@435: // Replacing Opaque nodes with their input in final_graph_reshaping_impl(), duke@435: // requires that the walk visits a node's inputs before visiting the node. bharadwaj@4315: void Compile::final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) { kvn@766: ResourceArea *area = Thread::current()->resource_area(); kvn@766: Unique_Node_List sfpt(area); kvn@766: kvn@1294: frc._visited.set(root->_idx); // first, mark node as visited duke@435: uint cnt = root->req(); duke@435: Node *n = root; duke@435: uint i = 0; duke@435: while (true) { duke@435: if (i < cnt) { duke@435: // Place all non-visited non-null inputs onto stack duke@435: Node* m = n->in(i); duke@435: ++i; kvn@1294: if (m != NULL && !frc._visited.test_set(m->_idx)) { kvn@766: if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL) kvn@766: sfpt.push(m); duke@435: cnt = m->req(); duke@435: nstack.push(n, i); // put on stack parent and next input's index duke@435: n = m; duke@435: i = 0; duke@435: } duke@435: } else { duke@435: // Now do post-visit work kvn@1294: final_graph_reshaping_impl( n, frc ); duke@435: if (nstack.is_empty()) duke@435: break; // finished duke@435: n = nstack.node(); // Get node from stack duke@435: cnt = n->req(); duke@435: i = nstack.index(); duke@435: nstack.pop(); // Shift to the next node on stack duke@435: } duke@435: } kvn@766: kvn@1930: // Skip next transformation if compressed oops are not used. roland@4159: if ((UseCompressedOops && !Matcher::gen_narrow_oop_implicit_null_checks()) || roland@4159: (!UseCompressedOops && !UseCompressedKlassPointers)) kvn@1930: return; kvn@1930: roland@4159: // Go over safepoints nodes to skip DecodeN/DecodeNKlass nodes for debug edges. kvn@766: // It could be done for an uncommon traps or any safepoints/calls roland@4159: // if the DecodeN/DecodeNKlass node is referenced only in a debug info. kvn@766: while (sfpt.size() > 0) { kvn@766: n = sfpt.pop(); kvn@766: JVMState *jvms = n->as_SafePoint()->jvms(); kvn@766: assert(jvms != NULL, "sanity"); kvn@766: int start = jvms->debug_start(); kvn@766: int end = n->req(); kvn@766: bool is_uncommon = (n->is_CallStaticJava() && kvn@766: n->as_CallStaticJava()->uncommon_trap_request() != 0); kvn@766: for (int j = start; j < end; j++) { kvn@766: Node* in = n->in(j); roland@4159: if (in->is_DecodeNarrowPtr()) { kvn@766: bool safe_to_skip = true; kvn@766: if (!is_uncommon ) { kvn@766: // Is it safe to skip? kvn@766: for (uint i = 0; i < in->outcnt(); i++) { kvn@766: Node* u = in->raw_out(i); kvn@766: if (!u->is_SafePoint() || kvn@766: u->is_Call() && u->as_Call()->has_non_debug_use(n)) { kvn@766: safe_to_skip = false; kvn@766: } kvn@766: } kvn@766: } kvn@766: if (safe_to_skip) { kvn@766: n->set_req(j, in->in(1)); kvn@766: } kvn@766: if (in->outcnt() == 0) { bharadwaj@4315: in->disconnect_inputs(NULL, this); kvn@766: } kvn@766: } kvn@766: } kvn@766: } duke@435: } duke@435: duke@435: //------------------------------final_graph_reshaping-------------------------- duke@435: // Final Graph Reshaping. duke@435: // duke@435: // (1) Clone simple inputs to uncommon calls, so they can be scheduled late duke@435: // and not commoned up and forced early. Must come after regular duke@435: // optimizations to avoid GVN undoing the cloning. Clone constant duke@435: // inputs to Loop Phis; these will be split by the allocator anyways. duke@435: // Remove Opaque nodes. duke@435: // (2) Move last-uses by commutative operations to the left input to encourage duke@435: // Intel update-in-place two-address operations and better register usage duke@435: // on RISCs. Must come after regular optimizations to avoid GVN Ideal duke@435: // calls canonicalizing them back. duke@435: // (3) Count the number of double-precision FP ops, single-precision FP ops duke@435: // and call sites. On Intel, we can get correct rounding either by duke@435: // forcing singles to memory (requires extra stores and loads after each duke@435: // FP bytecode) or we can set a rounding mode bit (requires setting and duke@435: // clearing the mode bit around call sites). The mode bit is only used duke@435: // if the relative frequency of single FP ops to calls is low enough. duke@435: // This is a key transform for SPEC mpeg_audio. duke@435: // (4) Detect infinite loops; blobs of code reachable from above but not duke@435: // below. Several of the Code_Gen algorithms fail on such code shapes, duke@435: // so we simply bail out. Happens a lot in ZKM.jar, but also happens duke@435: // from time to time in other codes (such as -Xcomp finalizer loops, etc). duke@435: // Detection is by looking for IfNodes where only 1 projection is duke@435: // reachable from below or CatchNodes missing some targets. duke@435: // (5) Assert for insane oop offsets in debug mode. duke@435: duke@435: bool Compile::final_graph_reshaping() { duke@435: // an infinite loop may have been eliminated by the optimizer, duke@435: // in which case the graph will be empty. duke@435: if (root()->req() == 1) { duke@435: record_method_not_compilable("trivial infinite loop"); duke@435: return true; duke@435: } duke@435: roland@4589: // Expensive nodes have their control input set to prevent the GVN roland@4589: // from freely commoning them. There's no GVN beyond this point so roland@4589: // no need to keep the control input. We want the expensive nodes to roland@4589: // be freely moved to the least frequent code path by gcm. roland@4589: assert(OptimizeExpensiveOps || expensive_count() == 0, "optimization off but list non empty?"); roland@4589: for (int i = 0; i < expensive_count(); i++) { roland@4589: _expensive_nodes->at(i)->set_req(0, NULL); roland@4589: } roland@4589: kvn@1294: Final_Reshape_Counts frc; duke@435: duke@435: // Visit everybody reachable! duke@435: // Allocate stack of size C->unique()/2 to avoid frequent realloc duke@435: Node_Stack nstack(unique() >> 1); kvn@1294: final_graph_reshaping_walk(nstack, root(), frc); duke@435: duke@435: // Check for unreachable (from below) code (i.e., infinite loops). kvn@1294: for( uint i = 0; i < frc._tests.size(); i++ ) { kvn@1294: MultiBranchNode *n = frc._tests[i]->as_MultiBranch(); never@562: // Get number of CFG targets. duke@435: // Note that PCTables include exception targets after calls. never@562: uint required_outcnt = n->required_outcnt(); never@562: if (n->outcnt() != required_outcnt) { duke@435: // Check for a few special cases. Rethrow Nodes never take the duke@435: // 'fall-thru' path, so expected kids is 1 less. duke@435: if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) { duke@435: if (n->in(0)->in(0)->is_Call()) { duke@435: CallNode *call = n->in(0)->in(0)->as_Call(); duke@435: if (call->entry_point() == OptoRuntime::rethrow_stub()) { never@562: required_outcnt--; // Rethrow always has 1 less kid duke@435: } else if (call->req() > TypeFunc::Parms && duke@435: call->is_CallDynamicJava()) { duke@435: // Check for null receiver. In such case, the optimizer has duke@435: // detected that the virtual call will always result in a null duke@435: // pointer exception. The fall-through projection of this CatchNode duke@435: // will not be populated. duke@435: Node *arg0 = call->in(TypeFunc::Parms); duke@435: if (arg0->is_Type() && duke@435: arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) { never@562: required_outcnt--; duke@435: } duke@435: } else if (call->entry_point() == OptoRuntime::new_array_Java() && duke@435: call->req() > TypeFunc::Parms+1 && duke@435: call->is_CallStaticJava()) { duke@435: // Check for negative array length. In such case, the optimizer has duke@435: // detected that the allocation attempt will always result in an duke@435: // exception. There is no fall-through projection of this CatchNode . duke@435: Node *arg1 = call->in(TypeFunc::Parms+1); duke@435: if (arg1->is_Type() && duke@435: arg1->as_Type()->type()->join(TypeInt::POS)->empty()) { never@562: required_outcnt--; duke@435: } duke@435: } duke@435: } duke@435: } never@562: // Recheck with a better notion of 'required_outcnt' never@562: if (n->outcnt() != required_outcnt) { duke@435: record_method_not_compilable("malformed control flow"); duke@435: return true; // Not all targets reachable! duke@435: } duke@435: } duke@435: // Check that I actually visited all kids. Unreached kids duke@435: // must be infinite loops. duke@435: for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) kvn@1294: if (!frc._visited.test(n->fast_out(j)->_idx)) { duke@435: record_method_not_compilable("infinite loop"); duke@435: return true; // Found unvisited kid; must be unreach duke@435: } duke@435: } duke@435: duke@435: // If original bytecodes contained a mixture of floats and doubles duke@435: // check if the optimizer has made it homogenous, item (3). never@1364: if( Use24BitFPMode && Use24BitFP && UseSSE == 0 && kvn@1294: frc.get_float_count() > 32 && kvn@1294: frc.get_double_count() == 0 && kvn@1294: (10 * frc.get_call_count() < frc.get_float_count()) ) { duke@435: set_24_bit_selection_and_mode( false, true ); duke@435: } duke@435: kvn@1294: set_java_calls(frc.get_java_call_count()); kvn@1294: set_inner_loops(frc.get_inner_loop_count()); duke@435: duke@435: // No infinite loops, no reason to bail out. duke@435: return false; duke@435: } duke@435: duke@435: //-----------------------------too_many_traps---------------------------------- duke@435: // Report if there are too many traps at the current method and bci. duke@435: // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded. duke@435: bool Compile::too_many_traps(ciMethod* method, duke@435: int bci, duke@435: Deoptimization::DeoptReason reason) { duke@435: ciMethodData* md = method->method_data(); duke@435: if (md->is_empty()) { duke@435: // Assume the trap has not occurred, or that it occurred only duke@435: // because of a transient condition during start-up in the interpreter. duke@435: return false; duke@435: } duke@435: if (md->has_trap_at(bci, reason) != 0) { duke@435: // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic. duke@435: // Also, if there are multiple reasons, or if there is no per-BCI record, duke@435: // assume the worst. duke@435: if (log()) duke@435: log()->elem("observe trap='%s' count='%d'", duke@435: Deoptimization::trap_reason_name(reason), duke@435: md->trap_count(reason)); duke@435: return true; duke@435: } else { duke@435: // Ignore method/bci and see if there have been too many globally. duke@435: return too_many_traps(reason, md); duke@435: } duke@435: } duke@435: duke@435: // Less-accurate variant which does not require a method and bci. duke@435: bool Compile::too_many_traps(Deoptimization::DeoptReason reason, duke@435: ciMethodData* logmd) { duke@435: if (trap_count(reason) >= (uint)PerMethodTrapLimit) { duke@435: // Too many traps globally. duke@435: // Note that we use cumulative trap_count, not just md->trap_count. duke@435: if (log()) { duke@435: int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason); duke@435: log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'", duke@435: Deoptimization::trap_reason_name(reason), duke@435: mcount, trap_count(reason)); duke@435: } duke@435: return true; duke@435: } else { duke@435: // The coast is clear. duke@435: return false; duke@435: } duke@435: } duke@435: duke@435: //--------------------------too_many_recompiles-------------------------------- duke@435: // Report if there are too many recompiles at the current method and bci. duke@435: // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff. duke@435: // Is not eager to return true, since this will cause the compiler to use duke@435: // Action_none for a trap point, to avoid too many recompilations. duke@435: bool Compile::too_many_recompiles(ciMethod* method, duke@435: int bci, duke@435: Deoptimization::DeoptReason reason) { duke@435: ciMethodData* md = method->method_data(); duke@435: if (md->is_empty()) { duke@435: // Assume the trap has not occurred, or that it occurred only duke@435: // because of a transient condition during start-up in the interpreter. duke@435: return false; duke@435: } duke@435: // Pick a cutoff point well within PerBytecodeRecompilationCutoff. duke@435: uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8; duke@435: uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero duke@435: Deoptimization::DeoptReason per_bc_reason duke@435: = Deoptimization::reason_recorded_per_bytecode_if_any(reason); duke@435: if ((per_bc_reason == Deoptimization::Reason_none duke@435: || md->has_trap_at(bci, reason) != 0) duke@435: // The trap frequency measure we care about is the recompile count: duke@435: && md->trap_recompiled_at(bci) duke@435: && md->overflow_recompile_count() >= bc_cutoff) { duke@435: // Do not emit a trap here if it has already caused recompilations. duke@435: // Also, if there are multiple reasons, or if there is no per-BCI record, duke@435: // assume the worst. duke@435: if (log()) duke@435: log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'", duke@435: Deoptimization::trap_reason_name(reason), duke@435: md->trap_count(reason), duke@435: md->overflow_recompile_count()); duke@435: return true; duke@435: } else if (trap_count(reason) != 0 duke@435: && decompile_count() >= m_cutoff) { duke@435: // Too many recompiles globally, and we have seen this sort of trap. duke@435: // Use cumulative decompile_count, not just md->decompile_count. duke@435: if (log()) duke@435: log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'", duke@435: Deoptimization::trap_reason_name(reason), duke@435: md->trap_count(reason), trap_count(reason), duke@435: md->decompile_count(), decompile_count()); duke@435: return true; duke@435: } else { duke@435: // The coast is clear. duke@435: return false; duke@435: } duke@435: } duke@435: duke@435: duke@435: #ifndef PRODUCT duke@435: //------------------------------verify_graph_edges--------------------------- duke@435: // Walk the Graph and verify that there is a one-to-one correspondence duke@435: // between Use-Def edges and Def-Use edges in the graph. duke@435: void Compile::verify_graph_edges(bool no_dead_code) { duke@435: if (VerifyGraphEdges) { duke@435: ResourceArea *area = Thread::current()->resource_area(); duke@435: Unique_Node_List visited(area); duke@435: // Call recursive graph walk to check edges duke@435: _root->verify_edges(visited); duke@435: if (no_dead_code) { duke@435: // Now make sure that no visited node is used by an unvisited node. duke@435: bool dead_nodes = 0; duke@435: Unique_Node_List checked(area); duke@435: while (visited.size() > 0) { duke@435: Node* n = visited.pop(); duke@435: checked.push(n); duke@435: for (uint i = 0; i < n->outcnt(); i++) { duke@435: Node* use = n->raw_out(i); duke@435: if (checked.member(use)) continue; // already checked duke@435: if (visited.member(use)) continue; // already in the graph duke@435: if (use->is_Con()) continue; // a dead ConNode is OK duke@435: // At this point, we have found a dead node which is DU-reachable. duke@435: if (dead_nodes++ == 0) duke@435: tty->print_cr("*** Dead nodes reachable via DU edges:"); duke@435: use->dump(2); duke@435: tty->print_cr("---"); duke@435: checked.push(use); // No repeats; pretend it is now checked. duke@435: } duke@435: } duke@435: assert(dead_nodes == 0, "using nodes must be reachable from root"); duke@435: } duke@435: } duke@435: } duke@435: #endif duke@435: duke@435: // The Compile object keeps track of failure reasons separately from the ciEnv. duke@435: // This is required because there is not quite a 1-1 relation between the duke@435: // ciEnv and its compilation task and the Compile object. Note that one duke@435: // ciEnv might use two Compile objects, if C2Compiler::compile_method decides duke@435: // to backtrack and retry without subsuming loads. Other than this backtracking duke@435: // behavior, the Compile's failure reason is quietly copied up to the ciEnv duke@435: // by the logic in C2Compiler. duke@435: void Compile::record_failure(const char* reason) { duke@435: if (log() != NULL) { duke@435: log()->elem("failure reason='%s' phase='compile'", reason); duke@435: } duke@435: if (_failure_reason == NULL) { duke@435: // Record the first failure reason. duke@435: _failure_reason = reason; duke@435: } sla@5237: sla@5237: EventCompilerFailure event; sla@5237: if (event.should_commit()) { sla@5237: event.set_compileID(Compile::compile_id()); sla@5237: event.set_failure(reason); sla@5237: event.commit(); sla@5237: } sla@5237: never@657: if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) { sla@5237: C->print_method(PHASE_FAILURE); never@657: } duke@435: _root = NULL; // flush the graph, too duke@435: } duke@435: duke@435: Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog) bharadwaj@4315: : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false), bharadwaj@4315: _phase_name(name), _dolog(dolog) duke@435: { duke@435: if (dolog) { duke@435: C = Compile::current(); duke@435: _log = C->log(); duke@435: } else { duke@435: C = NULL; duke@435: _log = NULL; duke@435: } duke@435: if (_log != NULL) { bharadwaj@4315: _log->begin_head("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); duke@435: _log->stamp(); duke@435: _log->end_head(); duke@435: } duke@435: } duke@435: duke@435: Compile::TracePhase::~TracePhase() { bharadwaj@4315: bharadwaj@4315: C = Compile::current(); bharadwaj@4315: if (_dolog) { bharadwaj@4315: _log = C->log(); bharadwaj@4315: } else { bharadwaj@4315: _log = NULL; bharadwaj@4315: } bharadwaj@4315: bharadwaj@4315: #ifdef ASSERT bharadwaj@4315: if (PrintIdealNodeCount) { bharadwaj@4315: tty->print_cr("phase name='%s' nodes='%d' live='%d' live_graph_walk='%d'", bharadwaj@4315: _phase_name, C->unique(), C->live_nodes(), C->count_live_nodes_by_graph_walk()); bharadwaj@4315: } bharadwaj@4315: bharadwaj@4315: if (VerifyIdealNodeCount) { bharadwaj@4315: Compile::current()->print_missing_nodes(); bharadwaj@4315: } bharadwaj@4315: #endif bharadwaj@4315: duke@435: if (_log != NULL) { bharadwaj@4315: _log->done("phase name='%s' nodes='%d' live='%d'", _phase_name, C->unique(), C->live_nodes()); duke@435: } duke@435: } twisti@2350: twisti@2350: //============================================================================= twisti@2350: // Two Constant's are equal when the type and the value are equal. twisti@2350: bool Compile::Constant::operator==(const Constant& other) { twisti@2350: if (type() != other.type() ) return false; twisti@2350: if (can_be_reused() != other.can_be_reused()) return false; twisti@2350: // For floating point values we compare the bit pattern. twisti@2350: switch (type()) { coleenp@4037: case T_FLOAT: return (_v._value.i == other._v._value.i); twisti@2350: case T_LONG: coleenp@4037: case T_DOUBLE: return (_v._value.j == other._v._value.j); twisti@2350: case T_OBJECT: coleenp@4037: case T_ADDRESS: return (_v._value.l == other._v._value.l); coleenp@4037: case T_VOID: return (_v._value.l == other._v._value.l); // jump-table entries kvn@4199: case T_METADATA: return (_v._metadata == other._v._metadata); twisti@2350: default: ShouldNotReachHere(); twisti@2350: } twisti@2350: return false; twisti@2350: } twisti@2350: twisti@2350: static int type_to_size_in_bytes(BasicType t) { twisti@2350: switch (t) { twisti@2350: case T_LONG: return sizeof(jlong ); twisti@2350: case T_FLOAT: return sizeof(jfloat ); twisti@2350: case T_DOUBLE: return sizeof(jdouble); coleenp@4037: case T_METADATA: return sizeof(Metadata*); twisti@2350: // We use T_VOID as marker for jump-table entries (labels) which twisti@3310: // need an internal word relocation. twisti@2350: case T_VOID: twisti@2350: case T_ADDRESS: twisti@2350: case T_OBJECT: return sizeof(jobject); twisti@2350: } twisti@2350: twisti@2350: ShouldNotReachHere(); twisti@2350: return -1; twisti@2350: } twisti@2350: twisti@3310: int Compile::ConstantTable::qsort_comparator(Constant* a, Constant* b) { twisti@3310: // sort descending twisti@3310: if (a->freq() > b->freq()) return -1; twisti@3310: if (a->freq() < b->freq()) return 1; twisti@3310: return 0; twisti@3310: } twisti@3310: twisti@2350: void Compile::ConstantTable::calculate_offsets_and_size() { twisti@3310: // First, sort the array by frequencies. twisti@3310: _constants.sort(qsort_comparator); twisti@3310: twisti@3310: #ifdef ASSERT twisti@3310: // Make sure all jump-table entries were sorted to the end of the twisti@3310: // array (they have a negative frequency). twisti@3310: bool found_void = false; twisti@3310: for (int i = 0; i < _constants.length(); i++) { twisti@3310: Constant con = _constants.at(i); twisti@3310: if (con.type() == T_VOID) twisti@3310: found_void = true; // jump-tables twisti@3310: else twisti@3310: assert(!found_void, "wrong sorting"); twisti@3310: } twisti@3310: #endif twisti@3310: twisti@3310: int offset = 0; twisti@3310: for (int i = 0; i < _constants.length(); i++) { twisti@3310: Constant* con = _constants.adr_at(i); twisti@3310: twisti@3310: // Align offset for type. twisti@3310: int typesize = type_to_size_in_bytes(con->type()); twisti@3310: offset = align_size_up(offset, typesize); twisti@3310: con->set_offset(offset); // set constant's offset twisti@3310: twisti@3310: if (con->type() == T_VOID) { twisti@3310: MachConstantNode* n = (MachConstantNode*) con->get_jobject(); twisti@3310: offset = offset + typesize * n->outcnt(); // expand jump-table twisti@3310: } else { twisti@3310: offset = offset + typesize; twisti@2350: } twisti@2350: } twisti@2350: twisti@2350: // Align size up to the next section start (which is insts; see twisti@2350: // CodeBuffer::align_at_start). twisti@2350: assert(_size == -1, "already set?"); twisti@3310: _size = align_size_up(offset, CodeEntryAlignment); twisti@2350: } twisti@2350: twisti@2350: void Compile::ConstantTable::emit(CodeBuffer& cb) { twisti@2350: MacroAssembler _masm(&cb); twisti@3310: for (int i = 0; i < _constants.length(); i++) { twisti@3310: Constant con = _constants.at(i); twisti@3310: address constant_addr; twisti@3310: switch (con.type()) { twisti@3310: case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break; twisti@3310: case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break; twisti@3310: case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break; twisti@3310: case T_OBJECT: { twisti@3310: jobject obj = con.get_jobject(); twisti@3310: int oop_index = _masm.oop_recorder()->find_index(obj); twisti@3310: constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index)); twisti@3310: break; twisti@3310: } twisti@3310: case T_ADDRESS: { twisti@3310: address addr = (address) con.get_jobject(); twisti@3310: constant_addr = _masm.address_constant(addr); twisti@3310: break; twisti@3310: } twisti@3310: // We use T_VOID as marker for jump-table entries (labels) which twisti@3310: // need an internal word relocation. twisti@3310: case T_VOID: { twisti@3310: MachConstantNode* n = (MachConstantNode*) con.get_jobject(); twisti@3310: // Fill the jump-table with a dummy word. The real value is twisti@3310: // filled in later in fill_jump_table. twisti@3310: address dummy = (address) n; twisti@3310: constant_addr = _masm.address_constant(dummy); twisti@3310: // Expand jump-table twisti@3310: for (uint i = 1; i < n->outcnt(); i++) { twisti@3310: address temp_addr = _masm.address_constant(dummy + i); twisti@3310: assert(temp_addr, "consts section too small"); twisti@2350: } twisti@3310: break; twisti@2350: } coleenp@4037: case T_METADATA: { coleenp@4037: Metadata* obj = con.get_metadata(); coleenp@4037: int metadata_index = _masm.oop_recorder()->find_index(obj); coleenp@4037: constant_addr = _masm.address_constant((address) obj, metadata_Relocation::spec(metadata_index)); coleenp@4037: break; coleenp@4037: } twisti@3310: default: ShouldNotReachHere(); twisti@3310: } twisti@3310: assert(constant_addr, "consts section too small"); kvn@3971: assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), err_msg_res("must be: %d == %d", constant_addr - _masm.code()->consts()->start(), con.offset())); twisti@2350: } twisti@2350: } twisti@2350: twisti@2350: int Compile::ConstantTable::find_offset(Constant& con) const { twisti@2350: int idx = _constants.find(con); twisti@2350: assert(idx != -1, "constant must be in constant table"); twisti@2350: int offset = _constants.at(idx).offset(); twisti@2350: assert(offset != -1, "constant table not emitted yet?"); twisti@2350: return offset; twisti@2350: } twisti@2350: twisti@2350: void Compile::ConstantTable::add(Constant& con) { twisti@2350: if (con.can_be_reused()) { twisti@2350: int idx = _constants.find(con); twisti@2350: if (idx != -1 && _constants.at(idx).can_be_reused()) { twisti@3310: _constants.adr_at(idx)->inc_freq(con.freq()); // increase the frequency by the current value twisti@2350: return; twisti@2350: } twisti@2350: } twisti@2350: (void) _constants.append(con); twisti@2350: } twisti@2350: twisti@3310: Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, BasicType type, jvalue value) { adlertz@5509: Block* b = Compile::current()->cfg()->get_block_for_node(n); twisti@3310: Constant con(type, value, b->_freq); twisti@2350: add(con); twisti@2350: return con; twisti@2350: } twisti@2350: coleenp@4037: Compile::Constant Compile::ConstantTable::add(Metadata* metadata) { coleenp@4037: Constant con(metadata); coleenp@4037: add(con); coleenp@4037: return con; coleenp@4037: } coleenp@4037: twisti@3310: Compile::Constant Compile::ConstantTable::add(MachConstantNode* n, MachOper* oper) { twisti@2350: jvalue value; twisti@2350: BasicType type = oper->type()->basic_type(); twisti@2350: switch (type) { twisti@2350: case T_LONG: value.j = oper->constantL(); break; twisti@2350: case T_FLOAT: value.f = oper->constantF(); break; twisti@2350: case T_DOUBLE: value.d = oper->constantD(); break; twisti@2350: case T_OBJECT: twisti@2350: case T_ADDRESS: value.l = (jobject) oper->constant(); break; coleenp@4037: case T_METADATA: return add((Metadata*)oper->constant()); break; coleenp@4037: default: guarantee(false, err_msg_res("unhandled type: %s", type2name(type))); twisti@2350: } twisti@3310: return add(n, type, value); twisti@2350: } twisti@2350: twisti@3310: Compile::Constant Compile::ConstantTable::add_jump_table(MachConstantNode* n) { twisti@2350: jvalue value; twisti@2350: // We can use the node pointer here to identify the right jump-table twisti@2350: // as this method is called from Compile::Fill_buffer right before twisti@2350: // the MachNodes are emitted and the jump-table is filled (means the twisti@2350: // MachNode pointers do not change anymore). twisti@2350: value.l = (jobject) n; twisti@3310: Constant con(T_VOID, value, next_jump_table_freq(), false); // Labels of a jump-table cannot be reused. twisti@3310: add(con); twisti@2350: return con; twisti@2350: } twisti@2350: twisti@2350: void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray labels) const { twisti@2350: // If called from Compile::scratch_emit_size do nothing. twisti@2350: if (Compile::current()->in_scratch_emit_size()) return; twisti@2350: twisti@2350: assert(labels.is_nonempty(), "must be"); kvn@3971: assert((uint) labels.length() == n->outcnt(), err_msg_res("must be equal: %d == %d", labels.length(), n->outcnt())); twisti@2350: twisti@2350: // Since MachConstantNode::constant_offset() also contains twisti@2350: // table_base_offset() we need to subtract the table_base_offset() twisti@2350: // to get the plain offset into the constant table. twisti@2350: int offset = n->constant_offset() - table_base_offset(); twisti@2350: twisti@2350: MacroAssembler _masm(&cb); twisti@2350: address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset); twisti@2350: twisti@3310: for (uint i = 0; i < n->outcnt(); i++) { twisti@2350: address* constant_addr = &jump_table_base[i]; kvn@3971: assert(*constant_addr == (((address) n) + i), err_msg_res("all jump-table entries must contain adjusted node pointer: " INTPTR_FORMAT " == " INTPTR_FORMAT, *constant_addr, (((address) n) + i))); twisti@2350: *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr); twisti@2350: cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type); twisti@2350: } twisti@2350: } roland@4357: roland@4357: void Compile::dump_inlining() { kvn@4448: if (PrintInlining || PrintIntrinsics NOT_PRODUCT( || PrintOptoInlining)) { roland@4409: // Print inlining message for candidates that we couldn't inline roland@4409: // for lack of space or non constant receiver roland@4409: for (int i = 0; i < _late_inlines.length(); i++) { roland@4409: CallGenerator* cg = _late_inlines.at(i); roland@4409: cg->print_inlining_late("live nodes > LiveNodeCountInliningCutoff"); roland@4409: } roland@4409: Unique_Node_List useful; roland@4409: useful.push(root()); roland@4409: for (uint next = 0; next < useful.size(); ++next) { roland@4409: Node* n = useful.at(next); roland@4409: if (n->is_Call() && n->as_Call()->generator() != NULL && n->as_Call()->generator()->call_node() == n) { roland@4409: CallNode* call = n->as_Call(); roland@4409: CallGenerator* cg = call->generator(); roland@4409: cg->print_inlining_late("receiver not constant"); roland@4409: } roland@4409: uint max = n->len(); roland@4409: for ( uint i = 0; i < max; ++i ) { roland@4409: Node *m = n->in(i); roland@4409: if ( m == NULL ) continue; roland@4409: useful.push(m); roland@4409: } roland@4409: } roland@4357: for (int i = 0; i < _print_inlining_list->length(); i++) { roland@4357: tty->print(_print_inlining_list->at(i).ss()->as_string()); roland@4357: } roland@4357: } roland@4357: } roland@4589: roland@4589: int Compile::cmp_expensive_nodes(Node* n1, Node* n2) { roland@4589: if (n1->Opcode() < n2->Opcode()) return -1; roland@4589: else if (n1->Opcode() > n2->Opcode()) return 1; roland@4589: roland@4589: assert(n1->req() == n2->req(), err_msg_res("can't compare %s nodes: n1->req() = %d, n2->req() = %d", NodeClassNames[n1->Opcode()], n1->req(), n2->req())); roland@4589: for (uint i = 1; i < n1->req(); i++) { roland@4589: if (n1->in(i) < n2->in(i)) return -1; roland@4589: else if (n1->in(i) > n2->in(i)) return 1; roland@4589: } roland@4589: roland@4589: return 0; roland@4589: } roland@4589: roland@4589: int Compile::cmp_expensive_nodes(Node** n1p, Node** n2p) { roland@4589: Node* n1 = *n1p; roland@4589: Node* n2 = *n2p; roland@4589: roland@4589: return cmp_expensive_nodes(n1, n2); roland@4589: } roland@4589: roland@4589: void Compile::sort_expensive_nodes() { roland@4589: if (!expensive_nodes_sorted()) { roland@4589: _expensive_nodes->sort(cmp_expensive_nodes); roland@4589: } roland@4589: } roland@4589: roland@4589: bool Compile::expensive_nodes_sorted() const { roland@4589: for (int i = 1; i < _expensive_nodes->length(); i++) { roland@4589: if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i-1)) < 0) { roland@4589: return false; roland@4589: } roland@4589: } roland@4589: return true; roland@4589: } roland@4589: roland@4589: bool Compile::should_optimize_expensive_nodes(PhaseIterGVN &igvn) { roland@4589: if (_expensive_nodes->length() == 0) { roland@4589: return false; roland@4589: } roland@4589: roland@4589: assert(OptimizeExpensiveOps, "optimization off?"); roland@4589: roland@4589: // Take this opportunity to remove dead nodes from the list roland@4589: int j = 0; roland@4589: for (int i = 0; i < _expensive_nodes->length(); i++) { roland@4589: Node* n = _expensive_nodes->at(i); roland@4589: if (!n->is_unreachable(igvn)) { roland@4589: assert(n->is_expensive(), "should be expensive"); roland@4589: _expensive_nodes->at_put(j, n); roland@4589: j++; roland@4589: } roland@4589: } roland@4589: _expensive_nodes->trunc_to(j); roland@4589: roland@4589: // Then sort the list so that similar nodes are next to each other roland@4589: // and check for at least two nodes of identical kind with same data roland@4589: // inputs. roland@4589: sort_expensive_nodes(); roland@4589: roland@4589: for (int i = 0; i < _expensive_nodes->length()-1; i++) { roland@4589: if (cmp_expensive_nodes(_expensive_nodes->adr_at(i), _expensive_nodes->adr_at(i+1)) == 0) { roland@4589: return true; roland@4589: } roland@4589: } roland@4589: roland@4589: return false; roland@4589: } roland@4589: roland@4589: void Compile::cleanup_expensive_nodes(PhaseIterGVN &igvn) { roland@4589: if (_expensive_nodes->length() == 0) { roland@4589: return; roland@4589: } roland@4589: roland@4589: assert(OptimizeExpensiveOps, "optimization off?"); roland@4589: roland@4589: // Sort to bring similar nodes next to each other and clear the roland@4589: // control input of nodes for which there's only a single copy. roland@4589: sort_expensive_nodes(); roland@4589: roland@4589: int j = 0; roland@4589: int identical = 0; roland@4589: int i = 0; roland@4589: for (; i < _expensive_nodes->length()-1; i++) { roland@4589: assert(j <= i, "can't write beyond current index"); roland@4589: if (_expensive_nodes->at(i)->Opcode() == _expensive_nodes->at(i+1)->Opcode()) { roland@4589: identical++; roland@4589: _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); roland@4589: continue; roland@4589: } roland@4589: if (identical > 0) { roland@4589: _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); roland@4589: identical = 0; roland@4589: } else { roland@4589: Node* n = _expensive_nodes->at(i); roland@4589: igvn.hash_delete(n); roland@4589: n->set_req(0, NULL); roland@4589: igvn.hash_insert(n); roland@4589: } roland@4589: } roland@4589: if (identical > 0) { roland@4589: _expensive_nodes->at_put(j++, _expensive_nodes->at(i)); roland@4589: } else if (_expensive_nodes->length() >= 1) { roland@4589: Node* n = _expensive_nodes->at(i); roland@4589: igvn.hash_delete(n); roland@4589: n->set_req(0, NULL); roland@4589: igvn.hash_insert(n); roland@4589: } roland@4589: _expensive_nodes->trunc_to(j); roland@4589: } roland@4589: roland@4589: void Compile::add_expensive_node(Node * n) { roland@4589: assert(!_expensive_nodes->contains(n), "duplicate entry in expensive list"); roland@4589: assert(n->is_expensive(), "expensive nodes with non-null control here only"); roland@4589: assert(!n->is_CFG() && !n->is_Mem(), "no cfg or memory nodes here"); roland@4589: if (OptimizeExpensiveOps) { roland@4589: _expensive_nodes->append(n); roland@4589: } else { roland@4589: // Clear control input and let IGVN optimize expensive nodes if roland@4589: // OptimizeExpensiveOps is off. roland@4589: n->set_req(0, NULL); roland@4589: } roland@4589: } shade@4691: shade@4691: // Auxiliary method to support randomized stressing/fuzzing. shade@4691: // shade@4691: // This method can be called the arbitrary number of times, with current count shade@4691: // as the argument. The logic allows selecting a single candidate from the shade@4691: // running list of candidates as follows: shade@4691: // int count = 0; shade@4691: // Cand* selected = null; shade@4691: // while(cand = cand->next()) { shade@4691: // if (randomized_select(++count)) { shade@4691: // selected = cand; shade@4691: // } shade@4691: // } shade@4691: // shade@4691: // Including count equalizes the chances any candidate is "selected". shade@4691: // This is useful when we don't have the complete list of candidates to choose shade@4691: // from uniformly. In this case, we need to adjust the randomicity of the shade@4691: // selection, or else we will end up biasing the selection towards the latter shade@4691: // candidates. shade@4691: // shade@4691: // Quick back-envelope calculation shows that for the list of n candidates shade@4691: // the equal probability for the candidate to persist as "best" can be shade@4691: // achieved by replacing it with "next" k-th candidate with the probability shade@4691: // of 1/k. It can be easily shown that by the end of the run, the shade@4691: // probability for any candidate is converged to 1/n, thus giving the shade@4691: // uniform distribution among all the candidates. shade@4691: // shade@4691: // We don't care about the domain size as long as (RANDOMIZED_DOMAIN / count) is large. shade@4691: #define RANDOMIZED_DOMAIN_POW 29 shade@4691: #define RANDOMIZED_DOMAIN (1 << RANDOMIZED_DOMAIN_POW) shade@4691: #define RANDOMIZED_DOMAIN_MASK ((1 << (RANDOMIZED_DOMAIN_POW + 1)) - 1) shade@4691: bool Compile::randomized_select(int count) { shade@4691: assert(count > 0, "only positive"); shade@4691: return (os::random() & RANDOMIZED_DOMAIN_MASK) < (RANDOMIZED_DOMAIN / count); shade@4691: }