aoqi@0: /*
aoqi@0:  * Copyright (c) 1997, 2013, Oracle and/or its affiliates. All rights reserved.
aoqi@0:  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
aoqi@0:  *
aoqi@0:  * This code is free software; you can redistribute it and/or modify it
aoqi@0:  * under the terms of the GNU General Public License version 2 only, as
aoqi@0:  * published by the Free Software Foundation.
aoqi@0:  *
aoqi@0:  * This code is distributed in the hope that it will be useful, but WITHOUT
aoqi@0:  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
aoqi@0:  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
aoqi@0:  * version 2 for more details (a copy is included in the LICENSE file that
aoqi@0:  * accompanied this code).
aoqi@0:  *
aoqi@0:  * You should have received a copy of the GNU General Public License version
aoqi@0:  * 2 along with this work; if not, write to the Free Software Foundation,
aoqi@0:  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
aoqi@0:  *
aoqi@0:  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
aoqi@0:  * or visit www.oracle.com if you need additional information or have any
aoqi@0:  * questions.
aoqi@0:  *
aoqi@0:  */
aoqi@0: 
aoqi@0: #ifndef SHARE_VM_RUNTIME_FPROFILER_HPP
aoqi@0: #define SHARE_VM_RUNTIME_FPROFILER_HPP
aoqi@0: 
aoqi@0: #include "utilities/macros.hpp"
aoqi@0: #include "runtime/timer.hpp"
aoqi@0: 
aoqi@0: // a simple flat profiler for Java
aoqi@0: 
aoqi@0: 
aoqi@0: // Forward declaration of classes defined in this header file
aoqi@0: class ThreadProfiler;
aoqi@0: class ThreadProfilerMark;
aoqi@0: class FlatProfiler;
aoqi@0: class IntervalData;
aoqi@0: 
aoqi@0: // Declarations of classes defined only in the implementation.
aoqi@0: class ProfilerNode;
aoqi@0: class FlatProfilerTask;
aoqi@0: 
aoqi@0: enum TickPosition {
aoqi@0:   tp_code,
aoqi@0:   tp_native
aoqi@0: };
aoqi@0: 
aoqi@0: // One of these guys is constructed as we enter interesting regions
aoqi@0: // and destructed as we exit the region.  While we are in the region
aoqi@0: // ticks are allotted to the region.
aoqi@0: class ThreadProfilerMark: public StackObj {
aoqi@0: public:
aoqi@0:   // For now, the only thread-specific region is the class loader.
aoqi@0:   enum Region { noRegion, classLoaderRegion, extraRegion, maxRegion };
aoqi@0: 
aoqi@0:   ThreadProfilerMark(Region)  NOT_FPROF_RETURN;
aoqi@0:   ~ThreadProfilerMark()       NOT_FPROF_RETURN;
aoqi@0: 
aoqi@0: private:
aoqi@0:   ThreadProfiler* _pp;
aoqi@0:   Region _r;
aoqi@0: };
aoqi@0: 
aoqi@0: #if INCLUDE_FPROF
aoqi@0: 
aoqi@0: class IntervalData VALUE_OBJ_CLASS_SPEC {
aoqi@0:   // Just to keep these things all together
aoqi@0: private:
aoqi@0:   int _interpreted;
aoqi@0:   int _compiled;
aoqi@0:   int _native;
aoqi@0:   int _compiling;
aoqi@0: public:
aoqi@0:   int interpreted() {
aoqi@0:     return _interpreted;
aoqi@0:   }
aoqi@0:   int compiled() {
aoqi@0:     return _compiled;
aoqi@0:   }
aoqi@0:   int native() {
aoqi@0:     return _native;
aoqi@0:   }
aoqi@0:   int compiling() {
aoqi@0:     return _compiling;
aoqi@0:   }
aoqi@0:   int total() {
aoqi@0:     return (interpreted() + compiled() + native() + compiling());
aoqi@0:   }
aoqi@0:   void inc_interpreted() {
aoqi@0:     _interpreted += 1;
aoqi@0:   }
aoqi@0:   void inc_compiled() {
aoqi@0:     _compiled += 1;
aoqi@0:   }
aoqi@0:   void inc_native() {
aoqi@0:     _native += 1;
aoqi@0:   }
aoqi@0:   void inc_compiling() {
aoqi@0:     _compiling += 1;
aoqi@0:   }
aoqi@0:   void reset() {
aoqi@0:     _interpreted = 0;
aoqi@0:     _compiled = 0;
aoqi@0:     _native = 0;
aoqi@0:     _compiling = 0;
aoqi@0:   }
aoqi@0:   static void print_header(outputStream* st);
aoqi@0:   void print_data(outputStream* st);
aoqi@0: };
aoqi@0: #endif // INCLUDE_FPROF
aoqi@0: 
aoqi@0: class ThreadProfiler: public CHeapObj<mtInternal> {
aoqi@0: public:
aoqi@0:   ThreadProfiler()    NOT_FPROF_RETURN;
aoqi@0:   ~ThreadProfiler()   NOT_FPROF_RETURN;
aoqi@0: 
aoqi@0:   // Resets the profiler
aoqi@0:   void reset()        NOT_FPROF_RETURN;
aoqi@0: 
aoqi@0:   // Activates the profiler for a certain thread
aoqi@0:   void engage()       NOT_FPROF_RETURN;
aoqi@0: 
aoqi@0:   // Deactivates the profiler
aoqi@0:   void disengage()    NOT_FPROF_RETURN;
aoqi@0: 
aoqi@0:   // Prints the collected profiling information
aoqi@0:   void print(const char* thread_name) NOT_FPROF_RETURN;
aoqi@0: 
aoqi@0:   // Garbage Collection Support
aoqi@0:   void oops_do(OopClosure* f)         NOT_FPROF_RETURN;
aoqi@0: 
aoqi@0: #if INCLUDE_FPROF
aoqi@0: private:
aoqi@0:   // for recording ticks.
aoqi@0:   friend class ProfilerNode;
aoqi@0:   char* area_bottom; // preallocated area for pnodes
aoqi@0:   char* area_top;
aoqi@0:   char* area_limit;
aoqi@0:   static int            table_size;
aoqi@0:   ProfilerNode** table;
aoqi@0: 
aoqi@0: private:
aoqi@0:   void record_interpreted_tick(JavaThread* thread, frame fr, TickPosition where, int* ticks);
aoqi@0:   void record_compiled_tick   (JavaThread* thread, frame fr, TickPosition where);
aoqi@0:   void interpreted_update(Method* method, TickPosition where);
aoqi@0:   void compiled_update   (Method* method, TickPosition where);
aoqi@0:   void stub_update       (Method* method, const char* name, TickPosition where);
aoqi@0:   void adapter_update    (TickPosition where);
aoqi@0: 
aoqi@0:   void runtime_stub_update(const CodeBlob* stub, const char* name, TickPosition where);
aoqi@0:   void unknown_compiled_update    (const CodeBlob* cb, TickPosition where);
aoqi@0: 
aoqi@0:   void vm_update    (TickPosition where);
aoqi@0:   void vm_update    (const char* name, TickPosition where);
aoqi@0: 
aoqi@0:   void record_tick_for_running_frame(JavaThread* thread, frame fr);
aoqi@0:   void record_tick_for_calling_frame(JavaThread* thread, frame fr);
aoqi@0: 
aoqi@0:   void initialize();
aoqi@0: 
aoqi@0:   static int  entry(int value);
aoqi@0: 
aoqi@0: 
aoqi@0: private:
aoqi@0:   friend class FlatProfiler;
aoqi@0:   void record_tick(JavaThread* thread);
aoqi@0:   bool engaged;
aoqi@0:   // so we can do percentages for this thread, and quick checks for activity
aoqi@0:   int thread_ticks;
aoqi@0:   int compiler_ticks;
aoqi@0:   int interpreter_ticks;
aoqi@0: 
aoqi@0: public:
aoqi@0:   void inc_thread_ticks() { thread_ticks += 1; }
aoqi@0: 
aoqi@0: private:
aoqi@0:   friend class ThreadProfilerMark;
aoqi@0:   // counters for thread-specific regions
aoqi@0:   bool region_flag[ThreadProfilerMark::maxRegion];
aoqi@0:   int class_loader_ticks;
aoqi@0:   int extra_ticks;
aoqi@0: 
aoqi@0: private:
aoqi@0:   // other thread-specific regions
aoqi@0:   int blocked_ticks;
aoqi@0:   enum UnknownTickSites {
aoqi@0:       ut_null_method,
aoqi@0:       ut_vtable_stubs,
aoqi@0:       ut_running_frame,
aoqi@0:       ut_calling_frame,
aoqi@0:       ut_no_pc,
aoqi@0:       ut_no_last_Java_frame,
aoqi@0:       ut_unknown_thread_state,
aoqi@0:       ut_end
aoqi@0:   };
aoqi@0:   int unknown_ticks_array[ut_end];
aoqi@0:   int unknown_ticks() {
aoqi@0:     int result = 0;
aoqi@0:     for (int ut = 0; ut < ut_end; ut += 1) {
aoqi@0:       result += unknown_ticks_array[ut];
aoqi@0:     }
aoqi@0:     return result;
aoqi@0:   }
aoqi@0: 
aoqi@0:   elapsedTimer timer;
aoqi@0: 
aoqi@0:   // For interval timing
aoqi@0: private:
aoqi@0:   IntervalData _interval_data;
aoqi@0:   IntervalData interval_data() {
aoqi@0:     return _interval_data;
aoqi@0:   }
aoqi@0:   IntervalData* interval_data_ref() {
aoqi@0:     return &_interval_data;
aoqi@0:   }
aoqi@0: #endif // INCLUDE_FPROF
aoqi@0: };
aoqi@0: 
aoqi@0: class FlatProfiler: AllStatic {
aoqi@0: public:
aoqi@0:   static void reset() NOT_FPROF_RETURN ;
aoqi@0:   static void engage(JavaThread* mainThread, bool fullProfile) NOT_FPROF_RETURN ;
aoqi@0:   static void disengage() NOT_FPROF_RETURN ;
aoqi@0:   static void print(int unused) NOT_FPROF_RETURN ;
aoqi@0:   static bool is_active() NOT_FPROF_RETURN_(false) ;
aoqi@0: 
aoqi@0:   // This is NULL if each thread has its own thread profiler,
aoqi@0:   // else this is the single thread profiler used by all threads.
aoqi@0:   // In particular it makes a difference during garbage collection,
aoqi@0:   // where you only want to traverse each thread profiler once.
aoqi@0:   static ThreadProfiler* get_thread_profiler() NOT_FPROF_RETURN_(NULL);
aoqi@0: 
aoqi@0:   // Garbage Collection Support
aoqi@0:   static void oops_do(OopClosure* f) NOT_FPROF_RETURN ;
aoqi@0: 
aoqi@0:   // Support for disassembler to inspect the PCRecorder
aoqi@0: 
aoqi@0:   // Returns the start address for a given pc
aoqi@0:   // NULL is returned if the PCRecorder is inactive
aoqi@0:   static address bucket_start_for(address pc) NOT_FPROF_RETURN_(NULL);
aoqi@0: 
aoqi@0:   enum { MillisecsPerTick = 10 };   // ms per profiling ticks
aoqi@0: 
aoqi@0:   // Returns the number of ticks recorded for the bucket
aoqi@0:   // pc belongs to.
aoqi@0:   static int bucket_count_for(address pc) NOT_FPROF_RETURN_(0);
aoqi@0: 
aoqi@0: #if INCLUDE_FPROF
aoqi@0: 
aoqi@0:  private:
aoqi@0:   static bool full_profile() {
aoqi@0:     return full_profile_flag;
aoqi@0:   }
aoqi@0: 
aoqi@0:   friend class ThreadProfiler;
aoqi@0:   // the following group of ticks cover everything that's not attributed to individual Java methods
aoqi@0:   static int  received_gc_ticks;      // ticks during which gc was active
aoqi@0:   static int vm_operation_ticks;      // total ticks in vm_operations other than GC
aoqi@0:   static int threads_lock_ticks;      // the number of times we couldn't get the Threads_lock without blocking
aoqi@0:   static int      blocked_ticks;      // ticks when the thread was blocked.
aoqi@0:   static int class_loader_ticks;      // total ticks in class loader
aoqi@0:   static int        extra_ticks;      // total ticks an extra temporary measuring
aoqi@0:   static int     compiler_ticks;      // total ticks in compilation
aoqi@0:   static int  interpreter_ticks;      // ticks in unknown interpreted method
aoqi@0:   static int        deopt_ticks;      // ticks in deoptimization
aoqi@0:   static int      unknown_ticks;      // ticks that cannot be categorized
aoqi@0:   static int     received_ticks;      // ticks that were received by task
aoqi@0:   static int    delivered_ticks;      // ticks that were delivered by task
aoqi@0:   static int non_method_ticks() {
aoqi@0:     return
aoqi@0:       ( received_gc_ticks
aoqi@0:       + vm_operation_ticks
aoqi@0:       + deopt_ticks
aoqi@0:       + threads_lock_ticks
aoqi@0:       + blocked_ticks
aoqi@0:       + compiler_ticks
aoqi@0:       + interpreter_ticks
aoqi@0:       + unknown_ticks );
aoqi@0:   }
aoqi@0:   static elapsedTimer timer;
aoqi@0: 
aoqi@0:   // Counts of each of the byte codes
aoqi@0:   static int*           bytecode_ticks;
aoqi@0:   static int*           bytecode_ticks_stub;
aoqi@0:   static void print_byte_code_statistics();
aoqi@0: 
aoqi@0:   // the ticks below are for continuous profiling (to adjust recompilation, etc.)
aoqi@0:   static int          all_ticks;      // total count of ticks received so far
aoqi@0:   static int      all_int_ticks;      // ticks in interpreter
aoqi@0:   static int     all_comp_ticks;      // ticks in compiled code (+ native)
aoqi@0:   static bool full_profile_flag;      // collecting full profile?
aoqi@0: 
aoqi@0:   // to accumulate thread-specific data
aoqi@0:   // if we aren't profiling individual threads.
aoqi@0:   static ThreadProfiler* thread_profiler;
aoqi@0:   static ThreadProfiler* vm_thread_profiler;
aoqi@0: 
aoqi@0:   static void allocate_table();
aoqi@0: 
aoqi@0:   // The task that periodically interrupts things.
aoqi@0:   friend class FlatProfilerTask;
aoqi@0:   static FlatProfilerTask* task;
aoqi@0:   static void record_vm_operation();
aoqi@0:   static void record_vm_tick();
aoqi@0:   static void record_thread_ticks();
aoqi@0: 
aoqi@0:   // For interval analysis
aoqi@0:  private:
aoqi@0:   static int interval_ticks_previous;  // delivered_ticks from the last interval
aoqi@0:   static void interval_record_thread(ThreadProfiler* tp); // extract ticks from ThreadProfiler.
aoqi@0:   static void interval_print();       // print interval data.
aoqi@0:   static void interval_reset();       // reset interval data.
aoqi@0:   enum {interval_print_size = 10};
aoqi@0:   static IntervalData* interval_data;
aoqi@0: #endif // INCLUDE_FPROF
aoqi@0: };
aoqi@0: 
aoqi@0: #endif // SHARE_VM_RUNTIME_FPROFILER_HPP