Merge

.hgtags

@@ -153,3 +153,4 @@ e9aa2ca89ad6c53420623d579765f9706ec523d7 jdk7-b130
 e9aa2ca89ad6c53420623d579765f9706ec523d7 hs21-b02
 0e531ab5ba04967a0e9aa6aef65e6eb3a0dcf632 jdk7-b132
 a8d643a4db47c7b58e0bcb49c77b5c3610de86a8 hs21-b03
+1b3a350709e4325d759bb453ff3fb6a463270488 jdk7-b133

src/share/vm/code/nmethod.cpp

@@ -170,7 +170,7 @@ struct nmethod_stats_struct {
   int pc_desc_resets;   // number of resets (= number of caches)
   int pc_desc_queries;  // queries to nmethod::find_pc_desc
   int pc_desc_approx;   // number of those which have approximate true
-  int pc_desc_repeats;  // number of _last_pc_desc hits
+  int pc_desc_repeats;  // number of _pc_descs[0] hits
   int pc_desc_hits;     // number of LRU cache hits
   int pc_desc_tests;    // total number of PcDesc examinations
   int pc_desc_searches; // total number of quasi-binary search steps
@@ -278,13 +278,12 @@ static inline bool match_desc(PcDesc* pc, int pc_offset, bool approximate) {
 
 void PcDescCache::reset_to(PcDesc* initial_pc_desc) {
   if (initial_pc_desc == NULL) {
-    _last_pc_desc = NULL;  // native method
+    _pc_descs[0] = NULL; // native method; no PcDescs at all
     return;
   }
   NOT_PRODUCT(++nmethod_stats.pc_desc_resets);
   // reset the cache by filling it with benign (non-null) values
   assert(initial_pc_desc->pc_offset() < 0, "must be sentinel");
-  _last_pc_desc = initial_pc_desc + 1;  // first valid one is after sentinel
   for (int i = 0; i < cache_size; i++)
     _pc_descs[i] = initial_pc_desc;
 }
@@ -293,25 +292,30 @@ PcDesc* PcDescCache::find_pc_desc(int pc_offset, bool approximate) {
   NOT_PRODUCT(++nmethod_stats.pc_desc_queries);
   NOT_PRODUCT(if (approximate) ++nmethod_stats.pc_desc_approx);
 
+  // Note: one might think that caching the most recently
+  // read value separately would be a win, but one would be
+  // wrong.  When many threads are updating it, the cache
+  // line it's in would bounce between caches, negating
+  // any benefit.
+
   // In order to prevent race conditions do not load cache elements
   // repeatedly, but use a local copy:
   PcDesc* res;
 
-  // Step one:  Check the most recently returned value.
-  res = _last_pc_desc;
+  // Step one:  Check the most recently added value.
+  res = _pc_descs[0];
   if (res == NULL) return NULL;  // native method; no PcDescs at all
   if (match_desc(res, pc_offset, approximate)) {
     NOT_PRODUCT(++nmethod_stats.pc_desc_repeats);
     return res;
   }
 
-  // Step two:  Check the LRU cache.
-  for (int i = 0; i < cache_size; i++) {
+  // Step two:  Check the rest of the LRU cache.
+  for (int i = 1; i < cache_size; ++i) {
     res = _pc_descs[i];
     if (res->pc_offset() < 0) break;  // optimization: skip empty cache
     if (match_desc(res, pc_offset, approximate)) {
       NOT_PRODUCT(++nmethod_stats.pc_desc_hits);
-      _last_pc_desc = res;  // record this cache hit in case of repeat
       return res;
     }
   }
@@ -322,13 +326,12 @@ PcDesc* PcDescCache::find_pc_desc(int pc_offset, bool approximate) {
 
 void PcDescCache::add_pc_desc(PcDesc* pc_desc) {
   NOT_PRODUCT(++nmethod_stats.pc_desc_adds);
-  // Update the LRU cache by shifting pc_desc forward:
+  // Update the LRU cache by shifting pc_desc forward.
   for (int i = 0; i < cache_size; i++)  {
     PcDesc* next = _pc_descs[i];
     _pc_descs[i] = pc_desc;
     pc_desc = next;
   }
-  // Note:  Do not update _last_pc_desc.  It fronts for the LRU cache.
 }
 
 // adjust pcs_size so that it is a multiple of both oopSize and
@@ -1180,14 +1183,17 @@ void nmethod::mark_as_seen_on_stack() {
   set_stack_traversal_mark(NMethodSweeper::traversal_count());
 }
 
-// Tell if a non-entrant method can be converted to a zombie (i.e., there is no activations on the stack)
+// Tell if a non-entrant method can be converted to a zombie (i.e.,
+// there are no activations on the stack, not in use by the VM,
+// and not in use by the ServiceThread)
 bool nmethod::can_not_entrant_be_converted() {
   assert(is_not_entrant(), "must be a non-entrant method");
 
   // Since the nmethod sweeper only does partial sweep the sweeper's traversal
   // count can be greater than the stack traversal count before it hits the
   // nmethod for the second time.
-  return stack_traversal_mark()+1 < NMethodSweeper::traversal_count();
+  return stack_traversal_mark()+1 < NMethodSweeper::traversal_count() &&
+         !is_locked_by_vm();
 }
 
 void nmethod::inc_decompile_count() {
@@ -1294,6 +1300,7 @@ void nmethod::log_state_change() const {
 // Common functionality for both make_not_entrant and make_zombie
 bool nmethod::make_not_entrant_or_zombie(unsigned int state) {
   assert(state == zombie || state == not_entrant, "must be zombie or not_entrant");
+  assert(!is_zombie(), "should not already be a zombie");
 
   // Make sure neither the nmethod nor the method is flushed in case of a safepoint in code below.
   nmethodLocker nml(this);
@@ -1301,11 +1308,6 @@ bool nmethod::make_not_entrant_or_zombie(unsigned int state) {
   No_Safepoint_Verifier nsv;
 
   {
-    // If the method is already zombie there is nothing to do
-    if (is_zombie()) {
-      return false;
-    }
-
     // invalidate osr nmethod before acquiring the patching lock since
     // they both acquire leaf locks and we don't want a deadlock.
     // This logic is equivalent to the logic below for patching the
@@ -1375,13 +1377,12 @@ bool nmethod::make_not_entrant_or_zombie(unsigned int state) {
       flush_dependencies(NULL);
     }
 
-    {
-      // zombie only - if a JVMTI agent has enabled the CompiledMethodUnload event
-      // and it hasn't already been reported for this nmethod then report it now.
-      // (the event may have been reported earilier if the GC marked it for unloading).
-      Pause_No_Safepoint_Verifier pnsv(&nsv);
+    // zombie only - if a JVMTI agent has enabled the CompiledMethodUnload
+    // event and it hasn't already been reported for this nmethod then
+    // report it now. The event may have been reported earilier if the GC
+    // marked it for unloading). JvmtiDeferredEventQueue support means
+    // we no longer go to a safepoint here.
     post_compiled_method_unload();
-    }
 
 #ifdef ASSERT
     // It's no longer safe to access the oops section since zombie
@@ -1566,7 +1567,7 @@ void nmethod::post_compiled_method_unload() {
   if (_jmethod_id != NULL && JvmtiExport::should_post_compiled_method_unload()) {
     assert(!unload_reported(), "already unloaded");
     JvmtiDeferredEvent event =
-      JvmtiDeferredEvent::compiled_method_unload_event(
+      JvmtiDeferredEvent::compiled_method_unload_event(this,
           _jmethod_id, insts_begin());
     if (SafepointSynchronize::is_at_safepoint()) {
       // Don't want to take the queueing lock. Add it as pending and
@@ -2171,10 +2172,12 @@ nmethodLocker::nmethodLocker(address pc) {
   lock_nmethod(_nm);
 }
 
-void nmethodLocker::lock_nmethod(nmethod* nm) {
+// Only JvmtiDeferredEvent::compiled_method_unload_event()
+// should pass zombie_ok == true.
+void nmethodLocker::lock_nmethod(nmethod* nm, bool zombie_ok) {
   if (nm == NULL)  return;
   Atomic::inc(&nm->_lock_count);
-  guarantee(!nm->is_zombie(), "cannot lock a zombie method");
+  guarantee(zombie_ok || !nm->is_zombie(), "cannot lock a zombie method");
 }
 
 void nmethodLocker::unlock_nmethod(nmethod* nm) {

src/share/vm/code/nmethod.hpp

@@ -69,14 +69,13 @@ class PcDescCache VALUE_OBJ_CLASS_SPEC {
   friend class VMStructs;
  private:
   enum { cache_size = 4 };
-  PcDesc* _last_pc_desc;         // most recent pc_desc found
   PcDesc* _pc_descs[cache_size]; // last cache_size pc_descs found
  public:
-  PcDescCache() { debug_only(_last_pc_desc = NULL); }
+  PcDescCache() { debug_only(_pc_descs[0] = NULL); }
   void    reset_to(PcDesc* initial_pc_desc);
   PcDesc* find_pc_desc(int pc_offset, bool approximate);
   void    add_pc_desc(PcDesc* pc_desc);
-  PcDesc* last_pc_desc() { return _last_pc_desc; }
+  PcDesc* last_pc_desc() { return _pc_descs[0]; }
 };
 
 
@@ -178,7 +177,7 @@ class nmethod : public CodeBlob {
   unsigned int _has_method_handle_invokes:1; // Has this method MethodHandle invokes?
 
   // Protected by Patching_lock
-  unsigned char _state;                      // {alive, not_entrant, zombie, unloaded)
+  unsigned char _state;                      // {alive, not_entrant, zombie, unloaded}
 
 #ifdef ASSERT
   bool _oops_are_stale;  // indicates that it's no longer safe to access oops section
@@ -194,7 +193,10 @@ class nmethod : public CodeBlob {
 
   NOT_PRODUCT(bool _has_debug_info; )
 
-  // Nmethod Flushing lock (if non-zero, then the nmethod is not removed)
+  // Nmethod Flushing lock. If non-zero, then the nmethod is not removed
+  // and is not made into a zombie. However, once the nmethod is made into
+  // a zombie, it will be locked one final time if CompiledMethodUnload
+  // event processing needs to be done.
   jint  _lock_count;
 
   // not_entrant method removal. Each mark_sweep pass will update
@@ -522,8 +524,9 @@ public:
   void flush();
 
  public:
-  // If returning true, it is unsafe to remove this nmethod even though it is a zombie
-  // nmethod, since the VM might have a reference to it. Should only be called from a  safepoint.
+  // When true is returned, it is unsafe to remove this nmethod even if
+  // it is a zombie, since the VM or the ServiceThread might still be
+  // using it.
   bool is_locked_by_vm() const                    { return _lock_count >0; }
 
   // See comment at definition of _last_seen_on_stack
@@ -689,13 +692,20 @@ public:
 
 };
 
-// Locks an nmethod so its code will not get removed, even if it is a zombie/not_entrant method
+// Locks an nmethod so its code will not get removed and it will not
+// be made into a zombie, even if it is a not_entrant method. After the
+// nmethod becomes a zombie, if CompiledMethodUnload event processing
+// needs to be done, then lock_nmethod() is used directly to keep the
+// generated code from being reused too early.
 class nmethodLocker : public StackObj {
   nmethod* _nm;
 
  public:
 
-  static void lock_nmethod(nmethod* nm);   // note: nm can be NULL
+  // note: nm can be NULL
+  // Only JvmtiDeferredEvent::compiled_method_unload_event()
+  // should pass zombie_ok == true.
+  static void lock_nmethod(nmethod* nm, bool zombie_ok = false);
   static void unlock_nmethod(nmethod* nm); // (ditto)
 
   nmethodLocker(address pc); // derive nm from pc

src/share/vm/prims/jvmtiImpl.cpp

@@ -919,15 +919,24 @@ JvmtiDeferredEvent JvmtiDeferredEvent::compiled_method_load_event(
     nmethod* nm) {
   JvmtiDeferredEvent event = JvmtiDeferredEvent(TYPE_COMPILED_METHOD_LOAD);
   event._event_data.compiled_method_load = nm;
-  nmethodLocker::lock_nmethod(nm); // will be unlocked when posted
+  // Keep the nmethod alive until the ServiceThread can process
+  // this deferred event.
+  nmethodLocker::lock_nmethod(nm);
   return event;
 }
 
 JvmtiDeferredEvent JvmtiDeferredEvent::compiled_method_unload_event(
-    jmethodID id, const void* code) {
+    nmethod* nm, jmethodID id, const void* code) {
   JvmtiDeferredEvent event = JvmtiDeferredEvent(TYPE_COMPILED_METHOD_UNLOAD);
+  event._event_data.compiled_method_unload.nm = nm;
   event._event_data.compiled_method_unload.method_id = id;
   event._event_data.compiled_method_unload.code_begin = code;
+  // Keep the nmethod alive until the ServiceThread can process
+  // this deferred event. This will keep the memory for the
+  // generated code from being reused too early. We pass
+  // zombie_ok == true here so that our nmethod that was just
+  // made into a zombie can be locked.
+  nmethodLocker::lock_nmethod(nm, true /* zombie_ok */);
   return event;
 }
 JvmtiDeferredEvent JvmtiDeferredEvent::dynamic_code_generated_event(
@@ -946,14 +955,19 @@ void JvmtiDeferredEvent::post() {
     case TYPE_COMPILED_METHOD_LOAD: {
       nmethod* nm = _event_data.compiled_method_load;
       JvmtiExport::post_compiled_method_load(nm);
+      // done with the deferred event so unlock the nmethod
       nmethodLocker::unlock_nmethod(nm);
       break;
     }
-    case TYPE_COMPILED_METHOD_UNLOAD:
+    case TYPE_COMPILED_METHOD_UNLOAD: {
+      nmethod* nm = _event_data.compiled_method_unload.nm;
       JvmtiExport::post_compiled_method_unload(
         _event_data.compiled_method_unload.method_id,
         _event_data.compiled_method_unload.code_begin);
+      // done with the deferred event so unlock the nmethod
+      nmethodLocker::unlock_nmethod(nm);
       break;
+    }
     case TYPE_DYNAMIC_CODE_GENERATED:
       JvmtiExport::post_dynamic_code_generated_internal(
         _event_data.dynamic_code_generated.name,

src/share/vm/prims/jvmtiImpl.hpp

@@ -458,6 +458,7 @@ class JvmtiDeferredEvent VALUE_OBJ_CLASS_SPEC {
   union {
     nmethod* compiled_method_load;
     struct {
+      nmethod* nm;
       jmethodID method_id;
       const void* code_begin;
     } compiled_method_unload;
@@ -477,7 +478,7 @@ class JvmtiDeferredEvent VALUE_OBJ_CLASS_SPEC {
   // Factory methods
   static JvmtiDeferredEvent compiled_method_load_event(nmethod* nm)
     KERNEL_RETURN_(JvmtiDeferredEvent());
-  static JvmtiDeferredEvent compiled_method_unload_event(
+  static JvmtiDeferredEvent compiled_method_unload_event(nmethod* nm,
       jmethodID id, const void* code) KERNEL_RETURN_(JvmtiDeferredEvent());
   static JvmtiDeferredEvent dynamic_code_generated_event(
       const char* name, const void* begin, const void* end)

src/share/vm/runtime/deoptimization.cpp

@@ -101,9 +101,9 @@ Deoptimization::UnrollBlock::UnrollBlock(int  size_of_deoptimized_frame,
   _frame_pcs                 = frame_pcs;
   _register_block            = NEW_C_HEAP_ARRAY(intptr_t, RegisterMap::reg_count * 2);
   _return_type               = return_type;
+  _initial_fp                = 0;
   // PD (x86 only)
   _counter_temp              = 0;
-  _initial_fp                = 0;
   _unpack_kind               = 0;
   _sender_sp_temp            = 0;
 
@@ -459,18 +459,9 @@ Deoptimization::UnrollBlock* Deoptimization::fetch_unroll_info_helper(JavaThread
                                       frame_sizes,
                                       frame_pcs,
                                       return_type);
-#if defined(IA32) || defined(AMD64)
-  // We need a way to pass fp to the unpacking code so the skeletal frames
-  // come out correct. This is only needed for x86 because of c2 using ebp
-  // as an allocatable register. So this update is useless (and harmless)
-  // on the other platforms. It would be nice to do this in a different
-  // way but even the old style deoptimization had a problem with deriving
-  // this value. NEEDS_CLEANUP
-  // Note: now that c1 is using c2's deopt blob we must do this on all
-  // x86 based platforms
-  intptr_t** fp_addr = (intptr_t**) (((address)info) + info->initial_fp_offset_in_bytes());
-  *fp_addr = array->sender().fp(); // was adapter_caller
-#endif /* IA32 || AMD64 */
+  // On some platforms, we need a way to pass fp to the unpacking code
+  // so the skeletal frames come out correct.
+  info->set_initial_fp((intptr_t) array->sender().fp());
 
   if (array->frames() > 1) {
     if (VerifyStack && TraceDeoptimization) {

src/share/vm/runtime/deoptimization.hpp

@@ -136,12 +136,12 @@ class Deoptimization : AllStatic {
     address*  _frame_pcs;                 // Array of frame pc's, in bytes, for unrolling the stack
     intptr_t* _register_block;            // Block for storing callee-saved registers.
     BasicType _return_type;               // Tells if we have to restore double or long return value
+    intptr_t  _initial_fp;                // FP of the sender frame
     // The following fields are used as temps during the unpacking phase
     // (which is tight on registers, especially on x86). They really ought
     // to be PD variables but that involves moving this class into its own
     // file to use the pd include mechanism. Maybe in a later cleanup ...
     intptr_t  _counter_temp;              // SHOULD BE PD VARIABLE (x86 frame count temp)
-    intptr_t  _initial_fp;                // SHOULD BE PD VARIABLE (x86/c2 initial ebp)
     intptr_t  _unpack_kind;               // SHOULD BE PD VARIABLE (x86 unpack kind)
     intptr_t  _sender_sp_temp;            // SHOULD BE PD VARIABLE (x86 sender_sp)
    public:
@@ -165,6 +165,8 @@ class Deoptimization : AllStatic {
     // Returns the total size of frames
     int size_of_frames() const;
 
+    void set_initial_fp(intptr_t fp) { _initial_fp = fp; }
+
     // Accessors used by the code generator for the unpack stub.
     static int size_of_deoptimized_frame_offset_in_bytes() { return offset_of(UnrollBlock, _size_of_deoptimized_frame); }
     static int caller_adjustment_offset_in_bytes()         { return offset_of(UnrollBlock, _caller_adjustment);         }

src/share/vm/runtime/serviceThread.cpp

@@ -70,11 +70,10 @@ void ServiceThread::initialize() {
     java_lang_Thread::set_priority(thread_oop(), NearMaxPriority);
     java_lang_Thread::set_daemon(thread_oop());
     thread->set_threadObj(thread_oop());
+    _instance = thread;
 
     Threads::add(thread);
     Thread::start(thread);
-
-    _instance = thread;
   }
 }