Fri, 11 Mar 2011 21:19:15 -0800
Merge
src/share/vm/oops/methodOop.hpp | file | annotate | diff | comparison | revisions | |
src/share/vm/runtime/arguments.cpp | file | annotate | diff | comparison | revisions |
1.1 --- a/src/share/vm/c1/c1_Canonicalizer.cpp Mon Mar 07 16:03:28 2011 -0500 1.2 +++ b/src/share/vm/c1/c1_Canonicalizer.cpp Fri Mar 11 21:19:15 2011 -0800 1.3 @@ -209,7 +209,7 @@ 1.4 // limit this optimization to current block 1.5 if (value != NULL && in_current_block(conv)) { 1.6 set_canonical(new StoreField(x->obj(), x->offset(), x->field(), value, x->is_static(), 1.7 - x->state_before(), x->is_loaded(), x->is_initialized())); 1.8 + x->state_before(), x->needs_patching())); 1.9 return; 1.10 } 1.11 }
2.1 --- a/src/share/vm/c1/c1_GraphBuilder.cpp Mon Mar 07 16:03:28 2011 -0500 2.2 +++ b/src/share/vm/c1/c1_GraphBuilder.cpp Fri Mar 11 21:19:15 2011 -0800 2.3 @@ -1456,12 +1456,12 @@ 2.4 BasicType field_type = field->type()->basic_type(); 2.5 ValueType* type = as_ValueType(field_type); 2.6 // call will_link again to determine if the field is valid. 2.7 - const bool is_loaded = holder->is_loaded() && 2.8 - field->will_link(method()->holder(), code); 2.9 - const bool is_initialized = is_loaded && holder->is_initialized(); 2.10 + const bool needs_patching = !holder->is_loaded() || 2.11 + !field->will_link(method()->holder(), code) || 2.12 + PatchALot; 2.13 2.14 ValueStack* state_before = NULL; 2.15 - if (!is_initialized || PatchALot) { 2.16 + if (!holder->is_initialized() || needs_patching) { 2.17 // save state before instruction for debug info when 2.18 // deoptimization happens during patching 2.19 state_before = copy_state_before(); 2.20 @@ -1469,10 +1469,6 @@ 2.21 2.22 Value obj = NULL; 2.23 if (code == Bytecodes::_getstatic || code == Bytecodes::_putstatic) { 2.24 - // commoning of class constants should only occur if the class is 2.25 - // fully initialized and resolved in this constant pool. The will_link test 2.26 - // above essentially checks if this class is resolved in this constant pool 2.27 - // so, the is_initialized flag should be suffiect. 2.28 if (state_before != NULL) { 2.29 // build a patching constant 2.30 obj = new Constant(new ClassConstant(holder), state_before); 2.31 @@ -1482,7 +1478,7 @@ 2.32 } 2.33 2.34 2.35 - const int offset = is_loaded ? field->offset() : -1; 2.36 + const int offset = !needs_patching ? field->offset() : -1; 2.37 switch (code) { 2.38 case Bytecodes::_getstatic: { 2.39 // check for compile-time constants, i.e., initialized static final fields 2.40 @@ -1509,7 +1505,7 @@ 2.41 state_before = copy_state_for_exception(); 2.42 } 2.43 push(type, append(new LoadField(append(obj), offset, field, true, 2.44 - state_before, is_loaded, is_initialized))); 2.45 + state_before, needs_patching))); 2.46 } 2.47 break; 2.48 } 2.49 @@ -1518,7 +1514,7 @@ 2.50 if (state_before == NULL) { 2.51 state_before = copy_state_for_exception(); 2.52 } 2.53 - append(new StoreField(append(obj), offset, field, val, true, state_before, is_loaded, is_initialized)); 2.54 + append(new StoreField(append(obj), offset, field, val, true, state_before, needs_patching)); 2.55 } 2.56 break; 2.57 case Bytecodes::_getfield : 2.58 @@ -1526,8 +1522,8 @@ 2.59 if (state_before == NULL) { 2.60 state_before = copy_state_for_exception(); 2.61 } 2.62 - LoadField* load = new LoadField(apop(), offset, field, false, state_before, is_loaded, true); 2.63 - Value replacement = is_loaded ? _memory->load(load) : load; 2.64 + LoadField* load = new LoadField(apop(), offset, field, false, state_before, needs_patching); 2.65 + Value replacement = !needs_patching ? _memory->load(load) : load; 2.66 if (replacement != load) { 2.67 assert(replacement->is_linked() || !replacement->can_be_linked(), "should already by linked"); 2.68 push(type, replacement); 2.69 @@ -1542,8 +1538,8 @@ 2.70 if (state_before == NULL) { 2.71 state_before = copy_state_for_exception(); 2.72 } 2.73 - StoreField* store = new StoreField(apop(), offset, field, val, false, state_before, is_loaded, true); 2.74 - if (is_loaded) store = _memory->store(store); 2.75 + StoreField* store = new StoreField(apop(), offset, field, val, false, state_before, needs_patching); 2.76 + if (!needs_patching) store = _memory->store(store); 2.77 if (store != NULL) { 2.78 append(store); 2.79 }
3.1 --- a/src/share/vm/c1/c1_Instruction.hpp Mon Mar 07 16:03:28 2011 -0500 3.2 +++ b/src/share/vm/c1/c1_Instruction.hpp Fri Mar 11 21:19:15 2011 -0800 3.3 @@ -323,8 +323,6 @@ 3.4 CanTrapFlag, 3.5 DirectCompareFlag, 3.6 IsEliminatedFlag, 3.7 - IsInitializedFlag, 3.8 - IsLoadedFlag, 3.9 IsSafepointFlag, 3.10 IsStaticFlag, 3.11 IsStrictfpFlag, 3.12 @@ -693,7 +691,7 @@ 3.13 public: 3.14 // creation 3.15 AccessField(Value obj, int offset, ciField* field, bool is_static, 3.16 - ValueStack* state_before, bool is_loaded, bool is_initialized) 3.17 + ValueStack* state_before, bool needs_patching) 3.18 : Instruction(as_ValueType(field->type()->basic_type()), state_before) 3.19 , _obj(obj) 3.20 , _offset(offset) 3.21 @@ -701,16 +699,9 @@ 3.22 , _explicit_null_check(NULL) 3.23 { 3.24 set_needs_null_check(!is_static); 3.25 - set_flag(IsLoadedFlag, is_loaded); 3.26 - set_flag(IsInitializedFlag, is_initialized); 3.27 set_flag(IsStaticFlag, is_static); 3.28 + set_flag(NeedsPatchingFlag, needs_patching); 3.29 ASSERT_VALUES 3.30 - if (!is_loaded || (PatchALot && !field->is_volatile())) { 3.31 - // need to patch if the holder wasn't loaded or we're testing 3.32 - // using PatchALot. Don't allow PatchALot for fields which are 3.33 - // known to be volatile they aren't patchable. 3.34 - set_flag(NeedsPatchingFlag, true); 3.35 - } 3.36 // pin of all instructions with memory access 3.37 pin(); 3.38 } 3.39 @@ -721,11 +712,14 @@ 3.40 ciField* field() const { return _field; } 3.41 BasicType field_type() const { return _field->type()->basic_type(); } 3.42 bool is_static() const { return check_flag(IsStaticFlag); } 3.43 - bool is_loaded() const { return check_flag(IsLoadedFlag); } 3.44 - bool is_initialized() const { return check_flag(IsInitializedFlag); } 3.45 NullCheck* explicit_null_check() const { return _explicit_null_check; } 3.46 bool needs_patching() const { return check_flag(NeedsPatchingFlag); } 3.47 3.48 + // Unresolved getstatic and putstatic can cause initialization. 3.49 + // Technically it occurs at the Constant that materializes the base 3.50 + // of the static fields but it's simpler to model it here. 3.51 + bool is_init_point() const { return is_static() && (needs_patching() || !_field->holder()->is_initialized()); } 3.52 + 3.53 // manipulation 3.54 3.55 // Under certain circumstances, if a previous NullCheck instruction 3.56 @@ -745,15 +739,15 @@ 3.57 public: 3.58 // creation 3.59 LoadField(Value obj, int offset, ciField* field, bool is_static, 3.60 - ValueStack* state_before, bool is_loaded, bool is_initialized) 3.61 - : AccessField(obj, offset, field, is_static, state_before, is_loaded, is_initialized) 3.62 + ValueStack* state_before, bool needs_patching) 3.63 + : AccessField(obj, offset, field, is_static, state_before, needs_patching) 3.64 {} 3.65 3.66 ciType* declared_type() const; 3.67 ciType* exact_type() const; 3.68 3.69 // generic 3.70 - HASHING2(LoadField, is_loaded() && !field()->is_volatile(), obj()->subst(), offset()) // cannot be eliminated if not yet loaded or if volatile 3.71 + HASHING2(LoadField, !needs_patching() && !field()->is_volatile(), obj()->subst(), offset()) // cannot be eliminated if needs patching or if volatile 3.72 }; 3.73 3.74 3.75 @@ -764,8 +758,8 @@ 3.76 public: 3.77 // creation 3.78 StoreField(Value obj, int offset, ciField* field, Value value, bool is_static, 3.79 - ValueStack* state_before, bool is_loaded, bool is_initialized) 3.80 - : AccessField(obj, offset, field, is_static, state_before, is_loaded, is_initialized) 3.81 + ValueStack* state_before, bool needs_patching) 3.82 + : AccessField(obj, offset, field, is_static, state_before, needs_patching) 3.83 , _value(value) 3.84 { 3.85 set_flag(NeedsWriteBarrierFlag, as_ValueType(field_type())->is_object());
4.1 --- a/src/share/vm/c1/c1_LIRGenerator.cpp Mon Mar 07 16:03:28 2011 -0500 4.2 +++ b/src/share/vm/c1/c1_LIRGenerator.cpp Fri Mar 11 21:19:15 2011 -0800 4.3 @@ -1559,9 +1559,7 @@ 4.4 (info ? new CodeEmitInfo(info) : NULL)); 4.5 } 4.6 4.7 - if (is_volatile) { 4.8 - assert(!needs_patching && x->is_loaded(), 4.9 - "how do we know it's volatile if it's not loaded"); 4.10 + if (is_volatile && !needs_patching) { 4.11 volatile_field_store(value.result(), address, info); 4.12 } else { 4.13 LIR_PatchCode patch_code = needs_patching ? lir_patch_normal : lir_patch_none; 4.14 @@ -1627,9 +1625,7 @@ 4.15 address = generate_address(object.result(), x->offset(), field_type); 4.16 } 4.17 4.18 - if (is_volatile) { 4.19 - assert(!needs_patching && x->is_loaded(), 4.20 - "how do we know it's volatile if it's not loaded"); 4.21 + if (is_volatile && !needs_patching) { 4.22 volatile_field_load(address, reg, info); 4.23 } else { 4.24 LIR_PatchCode patch_code = needs_patching ? lir_patch_normal : lir_patch_none;
5.1 --- a/src/share/vm/c1/c1_ValueMap.hpp Mon Mar 07 16:03:28 2011 -0500 5.2 +++ b/src/share/vm/c1/c1_ValueMap.hpp Fri Mar 11 21:19:15 2011 -0800 5.3 @@ -141,7 +141,8 @@ 5.4 5.5 // visitor functions 5.6 void do_StoreField (StoreField* x) { 5.7 - if (!x->is_initialized()) { 5.8 + if (x->is_init_point()) { 5.9 + // putstatic is an initialization point so treat it as a wide kill 5.10 kill_memory(); 5.11 } else { 5.12 kill_field(x->field()); 5.13 @@ -159,7 +160,8 @@ 5.14 void do_Local (Local* x) { /* nothing to do */ } 5.15 void do_Constant (Constant* x) { /* nothing to do */ } 5.16 void do_LoadField (LoadField* x) { 5.17 - if (!x->is_initialized()) { 5.18 + if (x->is_init_point()) { 5.19 + // getstatic is an initialization point so treat it as a wide kill 5.20 kill_memory(); 5.21 } 5.22 }
6.1 --- a/src/share/vm/code/codeCache.cpp Mon Mar 07 16:03:28 2011 -0500 6.2 +++ b/src/share/vm/code/codeCache.cpp Fri Mar 11 21:19:15 2011 -0800 6.3 @@ -939,9 +939,16 @@ 6.4 _heap->high(), 6.5 _heap->high_boundary()); 6.6 st->print_cr(" total_blobs=" UINT32_FORMAT " nmethods=" UINT32_FORMAT 6.7 - " adapters=" UINT32_FORMAT " free_code_cache=" SIZE_FORMAT 6.8 + " adapters=" UINT32_FORMAT " free_code_cache=" SIZE_FORMAT "Kb" 6.9 " largest_free_block=" SIZE_FORMAT, 6.10 - CodeCache::nof_blobs(), CodeCache::nof_nmethods(), 6.11 - CodeCache::nof_adapters(), CodeCache::unallocated_capacity(), 6.12 - CodeCache::largest_free_block()); 6.13 + nof_blobs(), nof_nmethods(), nof_adapters(), 6.14 + unallocated_capacity()/K, largest_free_block()); 6.15 } 6.16 + 6.17 +void CodeCache::log_state(outputStream* st) { 6.18 + st->print(" total_blobs='" UINT32_FORMAT "' nmethods='" UINT32_FORMAT "'" 6.19 + " adapters='" UINT32_FORMAT "' free_code_cache='" SIZE_FORMAT "'" 6.20 + " largest_free_block='" SIZE_FORMAT "'", 6.21 + nof_blobs(), nof_nmethods(), nof_adapters(), 6.22 + unallocated_capacity(), largest_free_block()); 6.23 +}
7.1 --- a/src/share/vm/code/codeCache.hpp Mon Mar 07 16:03:28 2011 -0500 7.2 +++ b/src/share/vm/code/codeCache.hpp Fri Mar 11 21:19:15 2011 -0800 7.3 @@ -147,6 +147,7 @@ 7.4 static void verify(); // verifies the code cache 7.5 static void print_trace(const char* event, CodeBlob* cb, int size = 0) PRODUCT_RETURN; 7.6 static void print_bounds(outputStream* st); // Prints a summary of the bounds of the code cache 7.7 + static void log_state(outputStream* st); 7.8 7.9 // The full limits of the codeCache 7.10 static address low_bound() { return (address) _heap->low_boundary(); } 7.11 @@ -159,7 +160,7 @@ 7.12 static size_t max_capacity() { return _heap->max_capacity(); } 7.13 static size_t unallocated_capacity() { return _heap->unallocated_capacity(); } 7.14 static size_t largest_free_block() { return _heap->largest_free_block(); } 7.15 - static bool needs_flushing() { return unallocated_capacity() < CodeCacheFlushingMinimumFreeSpace; } 7.16 + static bool needs_flushing() { return largest_free_block() < CodeCacheFlushingMinimumFreeSpace; } 7.17 7.18 static bool needs_cache_clean() { return _needs_cache_clean; } 7.19 static void set_needs_cache_clean(bool v) { _needs_cache_clean = v; }
8.1 --- a/src/share/vm/code/nmethod.cpp Mon Mar 07 16:03:28 2011 -0500 8.2 +++ b/src/share/vm/code/nmethod.cpp Fri Mar 11 21:19:15 2011 -0800 8.3 @@ -762,7 +762,7 @@ 8.4 8.5 void* nmethod::operator new(size_t size, int nmethod_size) { 8.6 // Always leave some room in the CodeCache for I2C/C2I adapters 8.7 - if (CodeCache::unallocated_capacity() < CodeCacheMinimumFreeSpace) return NULL; 8.8 + if (CodeCache::largest_free_block() < CodeCacheMinimumFreeSpace) return NULL; 8.9 return CodeCache::allocate(nmethod_size); 8.10 } 8.11 8.12 @@ -1881,7 +1881,7 @@ 8.13 8.14 8.15 oop nmethod::embeddedOop_at(u_char* p) { 8.16 - RelocIterator iter(this, p, p + oopSize); 8.17 + RelocIterator iter(this, p, p + 1); 8.18 while (iter.next()) 8.19 if (iter.type() == relocInfo::oop_type) { 8.20 return iter.oop_reloc()->oop_value();
9.1 --- a/src/share/vm/compiler/compileBroker.cpp Mon Mar 07 16:03:28 2011 -0500 9.2 +++ b/src/share/vm/compiler/compileBroker.cpp Fri Mar 11 21:19:15 2011 -0800 9.3 @@ -1364,7 +1364,7 @@ 9.4 // We need this HandleMark to avoid leaking VM handles. 9.5 HandleMark hm(thread); 9.6 9.7 - if (CodeCache::unallocated_capacity() < CodeCacheMinimumFreeSpace) { 9.8 + if (CodeCache::largest_free_block() < CodeCacheMinimumFreeSpace) { 9.9 // the code cache is really full 9.10 handle_full_code_cache(); 9.11 } else if (UseCodeCacheFlushing && CodeCache::needs_flushing()) { 9.12 @@ -1645,11 +1645,13 @@ 9.13 if (UseCompiler || AlwaysCompileLoopMethods ) { 9.14 if (xtty != NULL) { 9.15 xtty->begin_elem("code_cache_full"); 9.16 + CodeCache::log_state(xtty); 9.17 xtty->stamp(); 9.18 xtty->end_elem(); 9.19 } 9.20 warning("CodeCache is full. Compiler has been disabled."); 9.21 warning("Try increasing the code cache size using -XX:ReservedCodeCacheSize="); 9.22 + CodeCache::print_bounds(tty); 9.23 #ifndef PRODUCT 9.24 if (CompileTheWorld || ExitOnFullCodeCache) { 9.25 before_exit(JavaThread::current());
10.1 --- a/src/share/vm/memory/heap.cpp Mon Mar 07 16:03:28 2011 -0500 10.2 +++ b/src/share/vm/memory/heap.cpp Fri Mar 11 21:19:15 2011 -0800 10.3 @@ -316,12 +316,19 @@ 10.4 } 10.5 10.6 size_t CodeHeap::largest_free_block() const { 10.7 + // First check unused space excluding free blocks. 10.8 + size_t free_sz = size(_free_segments); 10.9 + size_t unused = max_capacity() - allocated_capacity() - free_sz; 10.10 + if (unused >= free_sz) 10.11 + return unused; 10.12 + 10.13 + // Now check largest free block. 10.14 size_t len = 0; 10.15 for (FreeBlock* b = _freelist; b != NULL; b = b->link()) { 10.16 if (b->length() > len) 10.17 len = b->length(); 10.18 } 10.19 - return size(len); 10.20 + return MAX2(unused, size(len)); 10.21 } 10.22 10.23 // Free list management
11.1 --- a/src/share/vm/oops/methodKlass.cpp Mon Mar 07 16:03:28 2011 -0500 11.2 +++ b/src/share/vm/oops/methodKlass.cpp Fri Mar 11 21:19:15 2011 -0800 11.3 @@ -103,6 +103,12 @@ 11.4 m->backedge_counter()->init(); 11.5 m->clear_number_of_breakpoints(); 11.6 11.7 +#ifdef TIERED 11.8 + m->set_rate(0); 11.9 + m->set_prev_event_count(0); 11.10 + m->set_prev_time(0); 11.11 +#endif 11.12 + 11.13 assert(m->is_parsable(), "must be parsable here."); 11.14 assert(m->size() == size, "wrong size for object"); 11.15 // We should not publish an uprasable object's reference
12.1 --- a/src/share/vm/oops/methodOop.hpp Mon Mar 07 16:03:28 2011 -0500 12.2 +++ b/src/share/vm/oops/methodOop.hpp Fri Mar 11 21:19:15 2011 -0800 12.3 @@ -84,6 +84,11 @@ 12.4 // | invocation_counter | 12.5 // | backedge_counter | 12.6 // |------------------------------------------------------| 12.7 +// | prev_time (tiered only, 64 bit wide) | 12.8 +// | | 12.9 +// |------------------------------------------------------| 12.10 +// | rate (tiered) | 12.11 +// |------------------------------------------------------| 12.12 // | code (pointer) | 12.13 // | i2i (pointer) | 12.14 // | adapter (pointer) | 12.15 @@ -124,6 +129,11 @@ 12.16 InvocationCounter _invocation_counter; // Incremented before each activation of the method - used to trigger frequency-based optimizations 12.17 InvocationCounter _backedge_counter; // Incremented before each backedge taken - used to trigger frequencey-based optimizations 12.18 12.19 +#ifdef TIERED 12.20 + jlong _prev_time; // Previous time the rate was acquired 12.21 + float _rate; // Events (invocation and backedge counter increments) per millisecond 12.22 +#endif 12.23 + 12.24 #ifndef PRODUCT 12.25 int _compiled_invocation_count; // Number of nmethod invocations so far (for perf. debugging) 12.26 #endif 12.27 @@ -304,6 +314,17 @@ 12.28 InvocationCounter* invocation_counter() { return &_invocation_counter; } 12.29 InvocationCounter* backedge_counter() { return &_backedge_counter; } 12.30 12.31 +#ifdef TIERED 12.32 + // We are reusing interpreter_invocation_count as a holder for the previous event count! 12.33 + // We can do that since interpreter_invocation_count is not used in tiered. 12.34 + int prev_event_count() const { return _interpreter_invocation_count; } 12.35 + void set_prev_event_count(int count) { _interpreter_invocation_count = count; } 12.36 + jlong prev_time() const { return _prev_time; } 12.37 + void set_prev_time(jlong time) { _prev_time = time; } 12.38 + float rate() const { return _rate; } 12.39 + void set_rate(float rate) { _rate = rate; } 12.40 +#endif 12.41 + 12.42 int invocation_count(); 12.43 int backedge_count(); 12.44
13.1 --- a/src/share/vm/opto/output.cpp Mon Mar 07 16:03:28 2011 -0500 13.2 +++ b/src/share/vm/opto/output.cpp Fri Mar 11 21:19:15 2011 -0800 13.3 @@ -1028,7 +1028,7 @@ 13.4 13.5 // helper for Fill_buffer bailout logic 13.6 static void turn_off_compiler(Compile* C) { 13.7 - if (CodeCache::unallocated_capacity() >= CodeCacheMinimumFreeSpace*10) { 13.8 + if (CodeCache::largest_free_block() >= CodeCacheMinimumFreeSpace*10) { 13.9 // Do not turn off compilation if a single giant method has 13.10 // blown the code cache size. 13.11 C->record_failure("excessive request to CodeCache");
14.1 --- a/src/share/vm/opto/type.cpp Mon Mar 07 16:03:28 2011 -0500 14.2 +++ b/src/share/vm/opto/type.cpp Fri Mar 11 21:19:15 2011 -0800 14.3 @@ -3386,7 +3386,22 @@ 14.4 instance_id = InstanceBot; 14.5 tary = TypeAry::make(Type::BOTTOM, tary->_size); 14.6 } 14.7 + } else // Non integral arrays. 14.8 + // Must fall to bottom if exact klasses in upper lattice 14.9 + // are not equal or super klass is exact. 14.10 + if ( above_centerline(ptr) && klass() != tap->klass() && 14.11 + // meet with top[] and bottom[] are processed further down: 14.12 + tap ->_klass != NULL && this->_klass != NULL && 14.13 + // both are exact and not equal: 14.14 + ((tap ->_klass_is_exact && this->_klass_is_exact) || 14.15 + // 'tap' is exact and super or unrelated: 14.16 + (tap ->_klass_is_exact && !tap->klass()->is_subtype_of(klass())) || 14.17 + // 'this' is exact and super or unrelated: 14.18 + (this->_klass_is_exact && !klass()->is_subtype_of(tap->klass())))) { 14.19 + tary = TypeAry::make(Type::BOTTOM, tary->_size); 14.20 + return make( NotNull, NULL, tary, lazy_klass, false, off, InstanceBot ); 14.21 } 14.22 + 14.23 bool xk = false; 14.24 switch (tap->ptr()) { 14.25 case AnyNull: 14.26 @@ -3766,7 +3781,7 @@ 14.27 // Oops, need to compute _klass and cache it 14.28 ciKlass* k_ary = compute_klass(); 14.29 14.30 - if( this != TypeAryPtr::OOPS ) { 14.31 + if( this != TypeAryPtr::OOPS && this->dual() != TypeAryPtr::OOPS ) { 14.32 // The _klass field acts as a cache of the underlying 14.33 // ciKlass for this array type. In order to set the field, 14.34 // we need to cast away const-ness.
15.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000 15.2 +++ b/src/share/vm/runtime/advancedThresholdPolicy.cpp Fri Mar 11 21:19:15 2011 -0800 15.3 @@ -0,0 +1,450 @@ 15.4 +/* 15.5 +* Copyright (c) 2010, 2011 Oracle and/or its affiliates. All rights reserved. 15.6 +* ORACLE PROPRIETARY/CONFIDENTIAL. Use is subject to license terms. 15.7 +*/ 15.8 + 15.9 +#include "precompiled.hpp" 15.10 +#include "runtime/advancedThresholdPolicy.hpp" 15.11 +#include "runtime/simpleThresholdPolicy.inline.hpp" 15.12 + 15.13 +#ifdef TIERED 15.14 +// Print an event. 15.15 +void AdvancedThresholdPolicy::print_specific(EventType type, methodHandle mh, methodHandle imh, 15.16 + int bci, CompLevel level) { 15.17 + tty->print(" rate: "); 15.18 + if (mh->prev_time() == 0) tty->print("n/a"); 15.19 + else tty->print("%f", mh->rate()); 15.20 + 15.21 + tty->print(" k: %.2lf,%.2lf", threshold_scale(CompLevel_full_profile, Tier3LoadFeedback), 15.22 + threshold_scale(CompLevel_full_optimization, Tier4LoadFeedback)); 15.23 + 15.24 +} 15.25 + 15.26 +void AdvancedThresholdPolicy::initialize() { 15.27 + // Turn on ergonomic compiler count selection 15.28 + if (FLAG_IS_DEFAULT(CICompilerCountPerCPU) && FLAG_IS_DEFAULT(CICompilerCount)) { 15.29 + FLAG_SET_DEFAULT(CICompilerCountPerCPU, true); 15.30 + } 15.31 + int count = CICompilerCount; 15.32 + if (CICompilerCountPerCPU) { 15.33 + // Simple log n seems to grow too slowly for tiered, try something faster: log n * log log n 15.34 + int log_cpu = log2_intptr(os::active_processor_count()); 15.35 + int loglog_cpu = log2_intptr(MAX2(log_cpu, 1)); 15.36 + count = MAX2(log_cpu * loglog_cpu, 1) * 3 / 2; 15.37 + } 15.38 + 15.39 + set_c1_count(MAX2(count / 3, 1)); 15.40 + set_c2_count(MAX2(count - count / 3, 1)); 15.41 + 15.42 + // Some inlining tuning 15.43 +#ifdef X86 15.44 + if (FLAG_IS_DEFAULT(InlineSmallCode)) { 15.45 + FLAG_SET_DEFAULT(InlineSmallCode, 2000); 15.46 + } 15.47 +#endif 15.48 + 15.49 +#ifdef SPARC 15.50 + if (FLAG_IS_DEFAULT(InlineSmallCode)) { 15.51 + FLAG_SET_DEFAULT(InlineSmallCode, 2500); 15.52 + } 15.53 +#endif 15.54 + 15.55 + 15.56 + set_start_time(os::javaTimeMillis()); 15.57 +} 15.58 + 15.59 +// update_rate() is called from select_task() while holding a compile queue lock. 15.60 +void AdvancedThresholdPolicy::update_rate(jlong t, methodOop m) { 15.61 + if (is_old(m)) { 15.62 + // We don't remove old methods from the queue, 15.63 + // so we can just zero the rate. 15.64 + m->set_rate(0); 15.65 + return; 15.66 + } 15.67 + 15.68 + // We don't update the rate if we've just came out of a safepoint. 15.69 + // delta_s is the time since last safepoint in milliseconds. 15.70 + jlong delta_s = t - SafepointSynchronize::end_of_last_safepoint(); 15.71 + jlong delta_t = t - (m->prev_time() != 0 ? m->prev_time() : start_time()); // milliseconds since the last measurement 15.72 + // How many events were there since the last time? 15.73 + int event_count = m->invocation_count() + m->backedge_count(); 15.74 + int delta_e = event_count - m->prev_event_count(); 15.75 + 15.76 + // We should be running for at least 1ms. 15.77 + if (delta_s >= TieredRateUpdateMinTime) { 15.78 + // And we must've taken the previous point at least 1ms before. 15.79 + if (delta_t >= TieredRateUpdateMinTime && delta_e > 0) { 15.80 + m->set_prev_time(t); 15.81 + m->set_prev_event_count(event_count); 15.82 + m->set_rate((float)delta_e / (float)delta_t); // Rate is events per millisecond 15.83 + } else 15.84 + if (delta_t > TieredRateUpdateMaxTime && delta_e == 0) { 15.85 + // If nothing happened for 25ms, zero the rate. Don't modify prev values. 15.86 + m->set_rate(0); 15.87 + } 15.88 + } 15.89 +} 15.90 + 15.91 +// Check if this method has been stale from a given number of milliseconds. 15.92 +// See select_task(). 15.93 +bool AdvancedThresholdPolicy::is_stale(jlong t, jlong timeout, methodOop m) { 15.94 + jlong delta_s = t - SafepointSynchronize::end_of_last_safepoint(); 15.95 + jlong delta_t = t - m->prev_time(); 15.96 + if (delta_t > timeout && delta_s > timeout) { 15.97 + int event_count = m->invocation_count() + m->backedge_count(); 15.98 + int delta_e = event_count - m->prev_event_count(); 15.99 + // Return true if there were no events. 15.100 + return delta_e == 0; 15.101 + } 15.102 + return false; 15.103 +} 15.104 + 15.105 +// We don't remove old methods from the compile queue even if they have 15.106 +// very low activity. See select_task(). 15.107 +bool AdvancedThresholdPolicy::is_old(methodOop method) { 15.108 + return method->invocation_count() > 50000 || method->backedge_count() > 500000; 15.109 +} 15.110 + 15.111 +double AdvancedThresholdPolicy::weight(methodOop method) { 15.112 + return (method->rate() + 1) * ((method->invocation_count() + 1) * (method->backedge_count() + 1)); 15.113 +} 15.114 + 15.115 +// Apply heuristics and return true if x should be compiled before y 15.116 +bool AdvancedThresholdPolicy::compare_methods(methodOop x, methodOop y) { 15.117 + if (x->highest_comp_level() > y->highest_comp_level()) { 15.118 + // recompilation after deopt 15.119 + return true; 15.120 + } else 15.121 + if (x->highest_comp_level() == y->highest_comp_level()) { 15.122 + if (weight(x) > weight(y)) { 15.123 + return true; 15.124 + } 15.125 + } 15.126 + return false; 15.127 +} 15.128 + 15.129 +// Is method profiled enough? 15.130 +bool AdvancedThresholdPolicy::is_method_profiled(methodOop method) { 15.131 + methodDataOop mdo = method->method_data(); 15.132 + if (mdo != NULL) { 15.133 + int i = mdo->invocation_count_delta(); 15.134 + int b = mdo->backedge_count_delta(); 15.135 + return call_predicate_helper<CompLevel_full_profile>(i, b, 1); 15.136 + } 15.137 + return false; 15.138 +} 15.139 + 15.140 +// Called with the queue locked and with at least one element 15.141 +CompileTask* AdvancedThresholdPolicy::select_task(CompileQueue* compile_queue) { 15.142 + CompileTask *max_task = NULL; 15.143 + methodOop max_method; 15.144 + jlong t = os::javaTimeMillis(); 15.145 + // Iterate through the queue and find a method with a maximum rate. 15.146 + for (CompileTask* task = compile_queue->first(); task != NULL;) { 15.147 + CompileTask* next_task = task->next(); 15.148 + methodOop method = (methodOop)JNIHandles::resolve(task->method_handle()); 15.149 + methodDataOop mdo = method->method_data(); 15.150 + update_rate(t, method); 15.151 + if (max_task == NULL) { 15.152 + max_task = task; 15.153 + max_method = method; 15.154 + } else { 15.155 + // If a method has been stale for some time, remove it from the queue. 15.156 + if (is_stale(t, TieredCompileTaskTimeout, method) && !is_old(method)) { 15.157 + if (PrintTieredEvents) { 15.158 + print_event(KILL, method, method, task->osr_bci(), (CompLevel)task->comp_level()); 15.159 + } 15.160 + CompileTaskWrapper ctw(task); // Frees the task 15.161 + compile_queue->remove(task); 15.162 + method->clear_queued_for_compilation(); 15.163 + task = next_task; 15.164 + continue; 15.165 + } 15.166 + 15.167 + // Select a method with a higher rate 15.168 + if (compare_methods(method, max_method)) { 15.169 + max_task = task; 15.170 + max_method = method; 15.171 + } 15.172 + } 15.173 + task = next_task; 15.174 + } 15.175 + 15.176 + if (max_task->comp_level() == CompLevel_full_profile && is_method_profiled(max_method)) { 15.177 + max_task->set_comp_level(CompLevel_limited_profile); 15.178 + if (PrintTieredEvents) { 15.179 + print_event(UPDATE, max_method, max_method, max_task->osr_bci(), (CompLevel)max_task->comp_level()); 15.180 + } 15.181 + } 15.182 + 15.183 + return max_task; 15.184 +} 15.185 + 15.186 +double AdvancedThresholdPolicy::threshold_scale(CompLevel level, int feedback_k) { 15.187 + double queue_size = CompileBroker::queue_size(level); 15.188 + int comp_count = compiler_count(level); 15.189 + double k = queue_size / (feedback_k * comp_count) + 1; 15.190 + return k; 15.191 +} 15.192 + 15.193 +// Call and loop predicates determine whether a transition to a higher 15.194 +// compilation level should be performed (pointers to predicate functions 15.195 +// are passed to common()). 15.196 +// Tier?LoadFeedback is basically a coefficient that determines of 15.197 +// how many methods per compiler thread can be in the queue before 15.198 +// the threshold values double. 15.199 +bool AdvancedThresholdPolicy::loop_predicate(int i, int b, CompLevel cur_level) { 15.200 + switch(cur_level) { 15.201 + case CompLevel_none: 15.202 + case CompLevel_limited_profile: { 15.203 + double k = threshold_scale(CompLevel_full_profile, Tier3LoadFeedback); 15.204 + return loop_predicate_helper<CompLevel_none>(i, b, k); 15.205 + } 15.206 + case CompLevel_full_profile: { 15.207 + double k = threshold_scale(CompLevel_full_optimization, Tier4LoadFeedback); 15.208 + return loop_predicate_helper<CompLevel_full_profile>(i, b, k); 15.209 + } 15.210 + default: 15.211 + return true; 15.212 + } 15.213 +} 15.214 + 15.215 +bool AdvancedThresholdPolicy::call_predicate(int i, int b, CompLevel cur_level) { 15.216 + switch(cur_level) { 15.217 + case CompLevel_none: 15.218 + case CompLevel_limited_profile: { 15.219 + double k = threshold_scale(CompLevel_full_profile, Tier3LoadFeedback); 15.220 + return call_predicate_helper<CompLevel_none>(i, b, k); 15.221 + } 15.222 + case CompLevel_full_profile: { 15.223 + double k = threshold_scale(CompLevel_full_optimization, Tier4LoadFeedback); 15.224 + return call_predicate_helper<CompLevel_full_profile>(i, b, k); 15.225 + } 15.226 + default: 15.227 + return true; 15.228 + } 15.229 +} 15.230 + 15.231 +// If a method is old enough and is still in the interpreter we would want to 15.232 +// start profiling without waiting for the compiled method to arrive. 15.233 +// We also take the load on compilers into the account. 15.234 +bool AdvancedThresholdPolicy::should_create_mdo(methodOop method, CompLevel cur_level) { 15.235 + if (cur_level == CompLevel_none && 15.236 + CompileBroker::queue_size(CompLevel_full_optimization) <= 15.237 + Tier3DelayOn * compiler_count(CompLevel_full_optimization)) { 15.238 + int i = method->invocation_count(); 15.239 + int b = method->backedge_count(); 15.240 + double k = Tier0ProfilingStartPercentage / 100.0; 15.241 + return call_predicate_helper<CompLevel_none>(i, b, k) || loop_predicate_helper<CompLevel_none>(i, b, k); 15.242 + } 15.243 + return false; 15.244 +} 15.245 + 15.246 +// Create MDO if necessary. 15.247 +void AdvancedThresholdPolicy::create_mdo(methodHandle mh, TRAPS) { 15.248 + if (mh->is_native() || mh->is_abstract() || mh->is_accessor()) return; 15.249 + if (mh->method_data() == NULL) { 15.250 + methodOopDesc::build_interpreter_method_data(mh, THREAD); 15.251 + if (HAS_PENDING_EXCEPTION) { 15.252 + CLEAR_PENDING_EXCEPTION; 15.253 + } 15.254 + } 15.255 +} 15.256 + 15.257 + 15.258 +/* 15.259 + * Method states: 15.260 + * 0 - interpreter (CompLevel_none) 15.261 + * 1 - pure C1 (CompLevel_simple) 15.262 + * 2 - C1 with invocation and backedge counting (CompLevel_limited_profile) 15.263 + * 3 - C1 with full profiling (CompLevel_full_profile) 15.264 + * 4 - C2 (CompLevel_full_optimization) 15.265 + * 15.266 + * Common state transition patterns: 15.267 + * a. 0 -> 3 -> 4. 15.268 + * The most common path. But note that even in this straightforward case 15.269 + * profiling can start at level 0 and finish at level 3. 15.270 + * 15.271 + * b. 0 -> 2 -> 3 -> 4. 15.272 + * This case occures when the load on C2 is deemed too high. So, instead of transitioning 15.273 + * into state 3 directly and over-profiling while a method is in the C2 queue we transition to 15.274 + * level 2 and wait until the load on C2 decreases. This path is disabled for OSRs. 15.275 + * 15.276 + * c. 0 -> (3->2) -> 4. 15.277 + * In this case we enqueue a method for compilation at level 3, but the C1 queue is long enough 15.278 + * to enable the profiling to fully occur at level 0. In this case we change the compilation level 15.279 + * of the method to 2, because it'll allow it to run much faster without full profiling while c2 15.280 + * is compiling. 15.281 + * 15.282 + * d. 0 -> 3 -> 1 or 0 -> 2 -> 1. 15.283 + * After a method was once compiled with C1 it can be identified as trivial and be compiled to 15.284 + * level 1. These transition can also occur if a method can't be compiled with C2 but can with C1. 15.285 + * 15.286 + * e. 0 -> 4. 15.287 + * This can happen if a method fails C1 compilation (it will still be profiled in the interpreter) 15.288 + * or because of a deopt that didn't require reprofiling (compilation won't happen in this case because 15.289 + * the compiled version already exists). 15.290 + * 15.291 + * Note that since state 0 can be reached from any other state via deoptimization different loops 15.292 + * are possible. 15.293 + * 15.294 + */ 15.295 + 15.296 +// Common transition function. Given a predicate determines if a method should transition to another level. 15.297 +CompLevel AdvancedThresholdPolicy::common(Predicate p, methodOop method, CompLevel cur_level) { 15.298 + if (is_trivial(method)) return CompLevel_simple; 15.299 + 15.300 + CompLevel next_level = cur_level; 15.301 + int i = method->invocation_count(); 15.302 + int b = method->backedge_count(); 15.303 + 15.304 + switch(cur_level) { 15.305 + case CompLevel_none: 15.306 + // If we were at full profile level, would we switch to full opt? 15.307 + if (common(p, method, CompLevel_full_profile) == CompLevel_full_optimization) { 15.308 + next_level = CompLevel_full_optimization; 15.309 + } else if ((this->*p)(i, b, cur_level)) { 15.310 + // C1-generated fully profiled code is about 30% slower than the limited profile 15.311 + // code that has only invocation and backedge counters. The observation is that 15.312 + // if C2 queue is large enough we can spend too much time in the fully profiled code 15.313 + // while waiting for C2 to pick the method from the queue. To alleviate this problem 15.314 + // we introduce a feedback on the C2 queue size. If the C2 queue is sufficiently long 15.315 + // we choose to compile a limited profiled version and then recompile with full profiling 15.316 + // when the load on C2 goes down. 15.317 + if (CompileBroker::queue_size(CompLevel_full_optimization) > 15.318 + Tier3DelayOn * compiler_count(CompLevel_full_optimization)) { 15.319 + next_level = CompLevel_limited_profile; 15.320 + } else { 15.321 + next_level = CompLevel_full_profile; 15.322 + } 15.323 + } 15.324 + break; 15.325 + case CompLevel_limited_profile: 15.326 + if (is_method_profiled(method)) { 15.327 + // Special case: we got here because this method was fully profiled in the interpreter. 15.328 + next_level = CompLevel_full_optimization; 15.329 + } else { 15.330 + methodDataOop mdo = method->method_data(); 15.331 + if (mdo != NULL) { 15.332 + if (mdo->would_profile()) { 15.333 + if (CompileBroker::queue_size(CompLevel_full_optimization) <= 15.334 + Tier3DelayOff * compiler_count(CompLevel_full_optimization) && 15.335 + (this->*p)(i, b, cur_level)) { 15.336 + next_level = CompLevel_full_profile; 15.337 + } 15.338 + } else { 15.339 + next_level = CompLevel_full_optimization; 15.340 + } 15.341 + } 15.342 + } 15.343 + break; 15.344 + case CompLevel_full_profile: 15.345 + { 15.346 + methodDataOop mdo = method->method_data(); 15.347 + if (mdo != NULL) { 15.348 + if (mdo->would_profile()) { 15.349 + int mdo_i = mdo->invocation_count_delta(); 15.350 + int mdo_b = mdo->backedge_count_delta(); 15.351 + if ((this->*p)(mdo_i, mdo_b, cur_level)) { 15.352 + next_level = CompLevel_full_optimization; 15.353 + } 15.354 + } else { 15.355 + next_level = CompLevel_full_optimization; 15.356 + } 15.357 + } 15.358 + } 15.359 + break; 15.360 + } 15.361 + return next_level; 15.362 +} 15.363 + 15.364 +// Determine if a method should be compiled with a normal entry point at a different level. 15.365 +CompLevel AdvancedThresholdPolicy::call_event(methodOop method, CompLevel cur_level) { 15.366 + CompLevel osr_level = (CompLevel) method->highest_osr_comp_level(); 15.367 + CompLevel next_level = common(&AdvancedThresholdPolicy::call_predicate, method, cur_level); 15.368 + 15.369 + // If OSR method level is greater than the regular method level, the levels should be 15.370 + // equalized by raising the regular method level in order to avoid OSRs during each 15.371 + // invocation of the method. 15.372 + if (osr_level == CompLevel_full_optimization && cur_level == CompLevel_full_profile) { 15.373 + methodDataOop mdo = method->method_data(); 15.374 + guarantee(mdo != NULL, "MDO should not be NULL"); 15.375 + if (mdo->invocation_count() >= 1) { 15.376 + next_level = CompLevel_full_optimization; 15.377 + } 15.378 + } else { 15.379 + next_level = MAX2(osr_level, next_level); 15.380 + } 15.381 + 15.382 + return next_level; 15.383 +} 15.384 + 15.385 +// Determine if we should do an OSR compilation of a given method. 15.386 +CompLevel AdvancedThresholdPolicy::loop_event(methodOop method, CompLevel cur_level) { 15.387 + if (cur_level == CompLevel_none) { 15.388 + // If there is a live OSR method that means that we deopted to the interpreter 15.389 + // for the transition. 15.390 + CompLevel osr_level = (CompLevel)method->highest_osr_comp_level(); 15.391 + if (osr_level > CompLevel_none) { 15.392 + return osr_level; 15.393 + } 15.394 + } 15.395 + return common(&AdvancedThresholdPolicy::loop_predicate, method, cur_level); 15.396 +} 15.397 + 15.398 +// Update the rate and submit compile 15.399 +void AdvancedThresholdPolicy::submit_compile(methodHandle mh, int bci, CompLevel level, TRAPS) { 15.400 + int hot_count = (bci == InvocationEntryBci) ? mh->invocation_count() : mh->backedge_count(); 15.401 + update_rate(os::javaTimeMillis(), mh()); 15.402 + CompileBroker::compile_method(mh, bci, level, mh, hot_count, "tiered", THREAD); 15.403 +} 15.404 + 15.405 + 15.406 +// Handle the invocation event. 15.407 +void AdvancedThresholdPolicy::method_invocation_event(methodHandle mh, methodHandle imh, 15.408 + CompLevel level, TRAPS) { 15.409 + if (should_create_mdo(mh(), level)) { 15.410 + create_mdo(mh, THREAD); 15.411 + } 15.412 + if (is_compilation_enabled() && !CompileBroker::compilation_is_in_queue(mh, InvocationEntryBci)) { 15.413 + CompLevel next_level = call_event(mh(), level); 15.414 + if (next_level != level) { 15.415 + compile(mh, InvocationEntryBci, next_level, THREAD); 15.416 + } 15.417 + } 15.418 +} 15.419 + 15.420 +// Handle the back branch event. Notice that we can compile the method 15.421 +// with a regular entry from here. 15.422 +void AdvancedThresholdPolicy::method_back_branch_event(methodHandle mh, methodHandle imh, 15.423 + int bci, CompLevel level, TRAPS) { 15.424 + if (should_create_mdo(mh(), level)) { 15.425 + create_mdo(mh, THREAD); 15.426 + } 15.427 + 15.428 + // If the method is already compiling, quickly bail out. 15.429 + if (is_compilation_enabled() && !CompileBroker::compilation_is_in_queue(mh, bci)) { 15.430 + // Use loop event as an opportinity to also check there's been 15.431 + // enough calls. 15.432 + CompLevel cur_level = comp_level(mh()); 15.433 + CompLevel next_level = call_event(mh(), cur_level); 15.434 + CompLevel next_osr_level = loop_event(mh(), level); 15.435 + if (next_osr_level == CompLevel_limited_profile) { 15.436 + next_osr_level = CompLevel_full_profile; // OSRs are supposed to be for very hot methods. 15.437 + } 15.438 + next_level = MAX2(next_level, 15.439 + next_osr_level < CompLevel_full_optimization ? next_osr_level : cur_level); 15.440 + bool is_compiling = false; 15.441 + if (next_level != cur_level) { 15.442 + compile(mh, InvocationEntryBci, next_level, THREAD); 15.443 + is_compiling = true; 15.444 + } 15.445 + 15.446 + // Do the OSR version 15.447 + if (!is_compiling && next_osr_level != level) { 15.448 + compile(mh, bci, next_osr_level, THREAD); 15.449 + } 15.450 + } 15.451 +} 15.452 + 15.453 +#endif // TIERED
16.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000 16.2 +++ b/src/share/vm/runtime/advancedThresholdPolicy.hpp Fri Mar 11 21:19:15 2011 -0800 16.3 @@ -0,0 +1,207 @@ 16.4 +/* 16.5 +* Copyright (c) 2010, 2011 Oracle and/or its affiliates. All rights reserved. 16.6 +* ORACLE PROPRIETARY/CONFIDENTIAL. Use is subject to license terms. 16.7 +*/ 16.8 + 16.9 +#ifndef SHARE_VM_RUNTIME_ADVANCEDTHRESHOLDPOLICY_HPP 16.10 +#define SHARE_VM_RUNTIME_ADVANCEDTHRESHOLDPOLICY_HPP 16.11 + 16.12 +#include "runtime/simpleThresholdPolicy.hpp" 16.13 + 16.14 +#ifdef TIERED 16.15 +class CompileTask; 16.16 +class CompileQueue; 16.17 + 16.18 +/* 16.19 + * The system supports 5 execution levels: 16.20 + * * level 0 - interpreter 16.21 + * * level 1 - C1 with full optimization (no profiling) 16.22 + * * level 2 - C1 with invocation and backedge counters 16.23 + * * level 3 - C1 with full profiling (level 2 + MDO) 16.24 + * * level 4 - C2 16.25 + * 16.26 + * Levels 0, 2 and 3 periodically notify the runtime about the current value of the counters 16.27 + * (invocation counters and backedge counters). The frequency of these notifications is 16.28 + * different at each level. These notifications are used by the policy to decide what transition 16.29 + * to make. 16.30 + * 16.31 + * Execution starts at level 0 (interpreter), then the policy can decide either to compile the 16.32 + * method at level 3 or level 2. The decision is based on the following factors: 16.33 + * 1. The length of the C2 queue determines the next level. The observation is that level 2 16.34 + * is generally faster than level 3 by about 30%, therefore we would want to minimize the time 16.35 + * a method spends at level 3. We should only spend the time at level 3 that is necessary to get 16.36 + * adequate profiling. So, if the C2 queue is long enough it is more beneficial to go first to 16.37 + * level 2, because if we transitioned to level 3 we would be stuck there until our C2 compile 16.38 + * request makes its way through the long queue. When the load on C2 recedes we are going to 16.39 + * recompile at level 3 and start gathering profiling information. 16.40 + * 2. The length of C1 queue is used to dynamically adjust the thresholds, so as to introduce 16.41 + * additional filtering if the compiler is overloaded. The rationale is that by the time a 16.42 + * method gets compiled it can become unused, so it doesn't make sense to put too much onto the 16.43 + * queue. 16.44 + * 16.45 + * After profiling is completed at level 3 the transition is made to level 4. Again, the length 16.46 + * of the C2 queue is used as a feedback to adjust the thresholds. 16.47 + * 16.48 + * After the first C1 compile some basic information is determined about the code like the number 16.49 + * of the blocks and the number of the loops. Based on that it can be decided that a method 16.50 + * is trivial and compiling it with C1 will yield the same code. In this case the method is 16.51 + * compiled at level 1 instead of 4. 16.52 + * 16.53 + * We also support profiling at level 0. If C1 is slow enough to produce the level 3 version of 16.54 + * the code and the C2 queue is sufficiently small we can decide to start profiling in the 16.55 + * interpreter (and continue profiling in the compiled code once the level 3 version arrives). 16.56 + * If the profiling at level 0 is fully completed before level 3 version is produced, a level 2 16.57 + * version is compiled instead in order to run faster waiting for a level 4 version. 16.58 + * 16.59 + * Compile queues are implemented as priority queues - for each method in the queue we compute 16.60 + * the event rate (the number of invocation and backedge counter increments per unit of time). 16.61 + * When getting an element off the queue we pick the one with the largest rate. Maintaining the 16.62 + * rate also allows us to remove stale methods (the ones that got on the queue but stopped 16.63 + * being used shortly after that). 16.64 +*/ 16.65 + 16.66 +/* Command line options: 16.67 + * - Tier?InvokeNotifyFreqLog and Tier?BackedgeNotifyFreqLog control the frequency of method 16.68 + * invocation and backedge notifications. Basically every n-th invocation or backedge a mutator thread 16.69 + * makes a call into the runtime. 16.70 + * 16.71 + * - Tier?CompileThreshold, Tier?BackEdgeThreshold, Tier?MinInvocationThreshold control 16.72 + * compilation thresholds. 16.73 + * Level 2 thresholds are not used and are provided for option-compatibility and potential future use. 16.74 + * Other thresholds work as follows: 16.75 + * 16.76 + * Transition from interpreter (level 0) to C1 with full profiling (level 3) happens when 16.77 + * the following predicate is true (X is the level): 16.78 + * 16.79 + * i > TierXInvocationThreshold * s || (i > TierXMinInvocationThreshold * s && i + b > TierXCompileThreshold * s), 16.80 + * 16.81 + * where $i$ is the number of method invocations, $b$ number of backedges and $s$ is the scaling 16.82 + * coefficient that will be discussed further. 16.83 + * The intuition is to equalize the time that is spend profiling each method. 16.84 + * The same predicate is used to control the transition from level 3 to level 4 (C2). It should be 16.85 + * noted though that the thresholds are relative. Moreover i and b for the 0->3 transition come 16.86 + * from methodOop and for 3->4 transition they come from MDO (since profiled invocations are 16.87 + * counted separately). 16.88 + * 16.89 + * OSR transitions are controlled simply with b > TierXBackEdgeThreshold * s predicates. 16.90 + * 16.91 + * - Tier?LoadFeedback options are used to automatically scale the predicates described above depending 16.92 + * on the compiler load. The scaling coefficients are computed as follows: 16.93 + * 16.94 + * s = queue_size_X / (TierXLoadFeedback * compiler_count_X) + 1, 16.95 + * 16.96 + * where queue_size_X is the current size of the compiler queue of level X, and compiler_count_X 16.97 + * is the number of level X compiler threads. 16.98 + * 16.99 + * Basically these parameters describe how many methods should be in the compile queue 16.100 + * per compiler thread before the scaling coefficient increases by one. 16.101 + * 16.102 + * This feedback provides the mechanism to automatically control the flow of compilation requests 16.103 + * depending on the machine speed, mutator load and other external factors. 16.104 + * 16.105 + * - Tier3DelayOn and Tier3DelayOff parameters control another important feedback loop. 16.106 + * Consider the following observation: a method compiled with full profiling (level 3) 16.107 + * is about 30% slower than a method at level 2 (just invocation and backedge counters, no MDO). 16.108 + * Normally, the following transitions will occur: 0->3->4. The problem arises when the C2 queue 16.109 + * gets congested and the 3->4 transition is delayed. While the method is the C2 queue it continues 16.110 + * executing at level 3 for much longer time than is required by the predicate and at suboptimal speed. 16.111 + * The idea is to dynamically change the behavior of the system in such a way that if a substantial 16.112 + * load on C2 is detected we would first do the 0->2 transition allowing a method to run faster. 16.113 + * And then when the load decreases to allow 2->3 transitions. 16.114 + * 16.115 + * Tier3Delay* parameters control this switching mechanism. 16.116 + * Tier3DelayOn is the number of methods in the C2 queue per compiler thread after which the policy 16.117 + * no longer does 0->3 transitions but does 0->2 transitions instead. 16.118 + * Tier3DelayOff switches the original behavior back when the number of methods in the C2 queue 16.119 + * per compiler thread falls below the specified amount. 16.120 + * The hysteresis is necessary to avoid jitter. 16.121 + * 16.122 + * - TieredCompileTaskTimeout is the amount of time an idle method can spend in the compile queue. 16.123 + * Basically, since we use the event rate d(i + b)/dt as a value of priority when selecting a method to 16.124 + * compile from the compile queue, we also can detect stale methods for which the rate has been 16.125 + * 0 for some time in the same iteration. Stale methods can appear in the queue when an application 16.126 + * abruptly changes its behavior. 16.127 + * 16.128 + * - TieredStopAtLevel, is used mostly for testing. It allows to bypass the policy logic and stick 16.129 + * to a given level. For example it's useful to set TieredStopAtLevel = 1 in order to compile everything 16.130 + * with pure c1. 16.131 + * 16.132 + * - Tier0ProfilingStartPercentage allows the interpreter to start profiling when the inequalities in the 16.133 + * 0->3 predicate are already exceeded by the given percentage but the level 3 version of the 16.134 + * method is still not ready. We can even go directly from level 0 to 4 if c1 doesn't produce a compiled 16.135 + * version in time. This reduces the overall transition to level 4 and decreases the startup time. 16.136 + * Note that this behavior is also guarded by the Tier3Delay mechanism: when the c2 queue is too long 16.137 + * these is not reason to start profiling prematurely. 16.138 + * 16.139 + * - TieredRateUpdateMinTime and TieredRateUpdateMaxTime are parameters of the rate computation. 16.140 + * Basically, the rate is not computed more frequently than TieredRateUpdateMinTime and is considered 16.141 + * to be zero if no events occurred in TieredRateUpdateMaxTime. 16.142 + */ 16.143 + 16.144 + 16.145 +class AdvancedThresholdPolicy : public SimpleThresholdPolicy { 16.146 + jlong _start_time; 16.147 + 16.148 + // Call and loop predicates determine whether a transition to a higher compilation 16.149 + // level should be performed (pointers to predicate functions are passed to common(). 16.150 + // Predicates also take compiler load into account. 16.151 + typedef bool (AdvancedThresholdPolicy::*Predicate)(int i, int b, CompLevel cur_level); 16.152 + bool call_predicate(int i, int b, CompLevel cur_level); 16.153 + bool loop_predicate(int i, int b, CompLevel cur_level); 16.154 + // Common transition function. Given a predicate determines if a method should transition to another level. 16.155 + CompLevel common(Predicate p, methodOop method, CompLevel cur_level); 16.156 + // Transition functions. 16.157 + // call_event determines if a method should be compiled at a different 16.158 + // level with a regular invocation entry. 16.159 + CompLevel call_event(methodOop method, CompLevel cur_level); 16.160 + // loop_event checks if a method should be OSR compiled at a different 16.161 + // level. 16.162 + CompLevel loop_event(methodOop method, CompLevel cur_level); 16.163 + // Has a method been long around? 16.164 + // We don't remove old methods from the compile queue even if they have 16.165 + // very low activity (see select_task()). 16.166 + inline bool is_old(methodOop method); 16.167 + // Was a given method inactive for a given number of milliseconds. 16.168 + // If it is, we would remove it from the queue (see select_task()). 16.169 + inline bool is_stale(jlong t, jlong timeout, methodOop m); 16.170 + // Compute the weight of the method for the compilation scheduling 16.171 + inline double weight(methodOop method); 16.172 + // Apply heuristics and return true if x should be compiled before y 16.173 + inline bool compare_methods(methodOop x, methodOop y); 16.174 + // Compute event rate for a given method. The rate is the number of event (invocations + backedges) 16.175 + // per millisecond. 16.176 + inline void update_rate(jlong t, methodOop m); 16.177 + // Compute threshold scaling coefficient 16.178 + inline double threshold_scale(CompLevel level, int feedback_k); 16.179 + // If a method is old enough and is still in the interpreter we would want to 16.180 + // start profiling without waiting for the compiled method to arrive. This function 16.181 + // determines whether we should do that. 16.182 + inline bool should_create_mdo(methodOop method, CompLevel cur_level); 16.183 + // Create MDO if necessary. 16.184 + void create_mdo(methodHandle mh, TRAPS); 16.185 + // Is method profiled enough? 16.186 + bool is_method_profiled(methodOop method); 16.187 + 16.188 +protected: 16.189 + void print_specific(EventType type, methodHandle mh, methodHandle imh, int bci, CompLevel level); 16.190 + 16.191 + void set_start_time(jlong t) { _start_time = t; } 16.192 + jlong start_time() const { return _start_time; } 16.193 + 16.194 + // Submit a given method for compilation (and update the rate). 16.195 + virtual void submit_compile(methodHandle mh, int bci, CompLevel level, TRAPS); 16.196 + // event() from SimpleThresholdPolicy would call these. 16.197 + virtual void method_invocation_event(methodHandle method, methodHandle inlinee, 16.198 + CompLevel level, TRAPS); 16.199 + virtual void method_back_branch_event(methodHandle method, methodHandle inlinee, 16.200 + int bci, CompLevel level, TRAPS); 16.201 +public: 16.202 + AdvancedThresholdPolicy() : _start_time(0) { } 16.203 + // Select task is called by CompileBroker. We should return a task or NULL. 16.204 + virtual CompileTask* select_task(CompileQueue* compile_queue); 16.205 + virtual void initialize(); 16.206 +}; 16.207 + 16.208 +#endif // TIERED 16.209 + 16.210 +#endif // SHARE_VM_RUNTIME_ADVANCEDTHRESHOLDPOLICY_HPP
17.1 --- a/src/share/vm/runtime/arguments.cpp Mon Mar 07 16:03:28 2011 -0500 17.2 +++ b/src/share/vm/runtime/arguments.cpp Fri Mar 11 21:19:15 2011 -0800 17.3 @@ -1026,8 +1026,9 @@ 17.4 } 17.5 17.6 void Arguments::set_tiered_flags() { 17.7 + // With tiered, set default policy to AdvancedThresholdPolicy, which is 3. 17.8 if (FLAG_IS_DEFAULT(CompilationPolicyChoice)) { 17.9 - FLAG_SET_DEFAULT(CompilationPolicyChoice, 2); 17.10 + FLAG_SET_DEFAULT(CompilationPolicyChoice, 3); 17.11 } 17.12 if (CompilationPolicyChoice < 2) { 17.13 vm_exit_during_initialization(
18.1 --- a/src/share/vm/runtime/compilationPolicy.cpp Mon Mar 07 16:03:28 2011 -0500 18.2 +++ b/src/share/vm/runtime/compilationPolicy.cpp Fri Mar 11 21:19:15 2011 -0800 18.3 @@ -1,5 +1,5 @@ 18.4 /* 18.5 - * Copyright (c) 2000, 2010, Oracle and/or its affiliates. All rights reserved. 18.6 + * Copyright (c) 2000, 2011, Oracle and/or its affiliates. All rights reserved. 18.7 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 18.8 * 18.9 * This code is free software; you can redistribute it and/or modify it 18.10 @@ -32,6 +32,7 @@ 18.11 #include "oops/methodOop.hpp" 18.12 #include "oops/oop.inline.hpp" 18.13 #include "prims/nativeLookup.hpp" 18.14 +#include "runtime/advancedThresholdPolicy.hpp" 18.15 #include "runtime/compilationPolicy.hpp" 18.16 #include "runtime/frame.hpp" 18.17 #include "runtime/handles.inline.hpp" 18.18 @@ -72,8 +73,15 @@ 18.19 Unimplemented(); 18.20 #endif 18.21 break; 18.22 + case 3: 18.23 +#ifdef TIERED 18.24 + CompilationPolicy::set_policy(new AdvancedThresholdPolicy()); 18.25 +#else 18.26 + Unimplemented(); 18.27 +#endif 18.28 + break; 18.29 default: 18.30 - fatal("CompilationPolicyChoice must be in the range: [0-2]"); 18.31 + fatal("CompilationPolicyChoice must be in the range: [0-3]"); 18.32 } 18.33 CompilationPolicy::policy()->initialize(); 18.34 }
19.1 --- a/src/share/vm/runtime/sweeper.cpp Mon Mar 07 16:03:28 2011 -0500 19.2 +++ b/src/share/vm/runtime/sweeper.cpp Fri Mar 11 21:19:15 2011 -0800 19.3 @@ -426,9 +426,7 @@ 19.4 tty->vprint(format, ap); 19.5 va_end(ap); 19.6 } 19.7 - tty->print_cr(" total_blobs='" UINT32_FORMAT "' nmethods='" UINT32_FORMAT "'" 19.8 - " adapters='" UINT32_FORMAT "' free_code_cache='" SIZE_FORMAT "'", 19.9 - CodeCache::nof_blobs(), CodeCache::nof_nmethods(), CodeCache::nof_adapters(), CodeCache::unallocated_capacity()); 19.10 + CodeCache::log_state(tty); tty->cr(); 19.11 } 19.12 19.13 if (LogCompilation && (xtty != NULL)) { 19.14 @@ -440,9 +438,7 @@ 19.15 xtty->vprint(format, ap); 19.16 va_end(ap); 19.17 } 19.18 - xtty->print(" total_blobs='" UINT32_FORMAT "' nmethods='" UINT32_FORMAT "'" 19.19 - " adapters='" UINT32_FORMAT "' free_code_cache='" SIZE_FORMAT "'", 19.20 - CodeCache::nof_blobs(), CodeCache::nof_nmethods(), CodeCache::nof_adapters(), CodeCache::unallocated_capacity()); 19.21 + CodeCache::log_state(xtty); 19.22 xtty->stamp(); 19.23 xtty->end_elem(); 19.24 }