8300002: Performance regression caused by non-inlined hot methods due to post call noop instructions

Reviewed-by: kvn, iveresov, eosterlund

Author: Sergey Kuksenko

src/hotspot/cpu/aarch64/macroAssembler_aarch64.cpp

@@ -1101,6 +1101,7 @@ void MacroAssembler::post_call_nop() {
   }
   InstructionMark im(this);
   relocate(post_call_nop_Relocation::spec());
+  InlineSkippedInstructionsCounter skipCounter(this);
   nop();
   movk(zr, 0);
   movk(zr, 0);

src/hotspot/cpu/ppc/macroAssembler_ppc.cpp

@@ -1185,6 +1185,7 @@ void MacroAssembler::post_call_nop() {
   if (!Continuations::enabled()) {
     return;
   }
+  InlineSkippedInstructionsCounter skipCounter(this);
   nop();
 }
 

src/hotspot/cpu/x86/macroAssembler_x86.cpp

@@ -2035,6 +2035,7 @@ void MacroAssembler::post_call_nop() {
   }
   InstructionMark im(this);
   relocate(post_call_nop_Relocation::spec());
+  InlineSkippedInstructionsCounter skipCounter(this);
   emit_int8((int8_t)0x0f);
   emit_int8((int8_t)0x1f);
   emit_int8((int8_t)0x84);

src/hotspot/share/asm/assembler.hpp

@@ -241,6 +241,19 @@ class AbstractAssembler : public ResourceObj  {
     }
   };
   friend class InstructionMark;
+
+  // count size of instructions which are skipped from inline heuristics
+  class InlineSkippedInstructionsCounter: public StackObj {
+   private:
+    AbstractAssembler* _assm;
+    address _start;
+   public:
+    InlineSkippedInstructionsCounter(AbstractAssembler* assm) : _assm(assm), _start(assm->pc()) {
+    }
+    ~InlineSkippedInstructionsCounter() {
+      _assm->register_skipped(_assm->pc() - _start);
+    }
+  };
 #ifdef ASSERT
   // Make it return true on platforms which need to verify
   // instruction boundaries for some operations.
@@ -333,10 +346,13 @@ class AbstractAssembler : public ResourceObj  {
   OopRecorder*  oop_recorder() const   { return _oop_recorder; }
   void      set_oop_recorder(OopRecorder* r) { _oop_recorder = r; }
 
+  void   register_skipped(int size) { code_section()->register_skipped(size); }
+
   address       inst_mark() const { return code_section()->mark();       }
   void      set_inst_mark()       {        code_section()->set_mark();   }
   void    clear_inst_mark()       {        code_section()->clear_mark(); }
 
+
   // Constants in code
   void relocate(RelocationHolder const& rspec, int format = 0) {
     assert(!pd_check_instruction_mark()

src/hotspot/share/asm/codeBuffer.cpp

@@ -596,6 +596,17 @@ csize_t CodeBuffer::total_offset_of(const CodeSection* cs) const {
   return -1;
 }
 
+int CodeBuffer::total_skipped_instructions_size() const {
+  int total_skipped_size = 0;
+  for (int n = (int) SECT_FIRST; n < (int) SECT_LIMIT; n++) {
+    const CodeSection* cur_cs = code_section(n);
+    if (!cur_cs->is_empty()) {
+      total_skipped_size += cur_cs->_skipped_instructions_size;
+    }
+  }
+  return total_skipped_size;
+}
+
 csize_t CodeBuffer::total_relocation_size() const {
   csize_t total = copy_relocations_to(NULL);  // dry run only
   return (csize_t) align_up(total, HeapWordSize);

src/hotspot/share/asm/codeBuffer.hpp

@@ -98,6 +98,7 @@ class CodeSection {
   address     _locs_point;      // last relocated position (grows upward)
   bool        _locs_own;        // did I allocate the locs myself?
   bool        _scratch_emit;    // Buffer is used for scratch emit, don't relocate.
+  int         _skipped_instructions_size;
   char        _index;           // my section number (SECT_INST, etc.)
   CodeBuffer* _outer;           // enclosing CodeBuffer
 
@@ -114,6 +115,7 @@ class CodeSection {
     _locs_point    = NULL;
     _locs_own      = false;
     _scratch_emit  = false;
+    _skipped_instructions_size = 0;
     debug_only(_index = (char)-1);
     debug_only(_outer = (CodeBuffer*)badAddress);
   }
@@ -144,6 +146,7 @@ class CodeSection {
     _end        = cs->_end;
     _limit      = cs->_limit;
     _locs_point = cs->_locs_point;
+    _skipped_instructions_size = cs->_skipped_instructions_size;
   }
 
  public:
@@ -204,6 +207,10 @@ class CodeSection {
     _locs_point = pc;
   }
 
+  void register_skipped(int size) {
+    _skipped_instructions_size += size;
+  }
+
   // Code emission
   void emit_int8(uint8_t x1) {
     address curr = end();
@@ -638,6 +645,8 @@ class CodeBuffer: public StackObj DEBUG_ONLY(COMMA private Scrubber) {
   // allocated size of all relocation data, including index, rounded up
   csize_t total_relocation_size() const;
 
+  int total_skipped_instructions_size() const;
+
   csize_t copy_relocations_to(address buf, csize_t buf_limit, bool only_inst) const;
 
   // allocated size of any and all recorded oops

src/hotspot/share/ci/ciMethod.cpp

@@ -151,7 +151,7 @@ ciMethod::ciMethod(const methodHandle& h_m, ciInstanceKlass* holder) :
   }
   if (_interpreter_invocation_count == 0)
     _interpreter_invocation_count = 1;
-  _instructions_size = -1;
+  _inline_instructions_size = -1;
   if (ReplayCompiles) {
     ciReplay::initialize(this);
   }
@@ -172,7 +172,7 @@ ciMethod::ciMethod(ciInstanceKlass* holder,
   _method_data(            NULL),
   _method_blocks(          NULL),
   _intrinsic_id(           vmIntrinsics::_none),
-  _instructions_size(-1),
+  _inline_instructions_size(-1),
   _can_be_statically_bound(false),
   _can_omit_stack_trace(true),
   _liveness(               NULL)
@@ -1087,7 +1087,7 @@ bool ciMethod::can_be_compiled() {
 // ------------------------------------------------------------------
 // ciMethod::has_compiled_code
 bool ciMethod::has_compiled_code() {
-  return instructions_size() > 0;
+  return inline_instructions_size() > 0;
 }
 
 int ciMethod::highest_osr_comp_level() {
@@ -1110,25 +1110,28 @@ int ciMethod::code_size_for_inlining() {
 }
 
 // ------------------------------------------------------------------
-// ciMethod::instructions_size
+// ciMethod::inline_instructions_size
 //
 // This is a rough metric for "fat" methods, compared before inlining
 // with InlineSmallCode.  The CodeBlob::code_size accessor includes
 // junk like exception handler, stubs, and constant table, which are
 // not highly relevant to an inlined method.  So we use the more
 // specific accessor nmethod::insts_size.
-int ciMethod::instructions_size() {
-  if (_instructions_size == -1) {
+// Also some instructions inside the code are excluded from inline
+// heuristic (e.g. post call nop instructions; see InlineSkippedInstructionsCounter)
+int ciMethod::inline_instructions_size() {
+  if (_inline_instructions_size == -1) {
     GUARDED_VM_ENTRY(
-                     CompiledMethod* code = get_Method()->code();
-                     if (code != NULL && (code->comp_level() == CompLevel_full_optimization)) {
-                       _instructions_size = code->insts_end() - code->verified_entry_point();
-                     } else {
-                       _instructions_size = 0;
-                     }
-                     );
+      CompiledMethod* code = get_Method()->code();
+      if (code != NULL && (code->comp_level() == CompLevel_full_optimization)) {
+        int isize = code->insts_end() - code->verified_entry_point() - code->skipped_instructions_size();
+        _inline_instructions_size = isize > 0 ? isize : 0;
+      } else {
+        _inline_instructions_size = 0;
+      }
+    );
   }
-  return _instructions_size;
+  return _inline_instructions_size;
 }
 
 // ------------------------------------------------------------------
@@ -1315,7 +1318,7 @@ void ciMethod::dump_replay_data(outputStream* st) {
                mcs == NULL ? 0 : mcs->backedge_counter()->raw_counter(),
                interpreter_invocation_count(),
                interpreter_throwout_count(),
-               _instructions_size);
+               _inline_instructions_size);
 }
 
 // ------------------------------------------------------------------

src/hotspot/share/ci/ciMethod.hpp

@@ -82,7 +82,7 @@ class ciMethod : public ciMetadata {
   int _handler_count;
   int _interpreter_invocation_count;
   int _interpreter_throwout_count;
-  int _instructions_size;
+  int _inline_instructions_size;
   int _size_of_parameters;
 
   bool _uses_monitors;
@@ -315,7 +315,8 @@ class ciMethod : public ciMetadata {
   bool check_call(int refinfo_index, bool is_static) const;
   bool ensure_method_data();  // make sure it exists in the VM also
   MethodCounters* ensure_method_counters();
-  int instructions_size();
+
+  int inline_instructions_size();
   int scale_count(int count, float prof_factor = 1.);  // make MDO count commensurate with IIC
 
   // Stack walking support

src/hotspot/share/ci/ciReplay.cpp

@@ -1529,7 +1529,7 @@ void ciReplay::initialize(ciMethod* m) {
   } else {
     EXCEPTION_CONTEXT;
     // m->_instructions_size = rec->_instructions_size;
-    m->_instructions_size = -1;
+    m->_inline_instructions_size = -1;
     m->_interpreter_invocation_count = rec->_interpreter_invocation_count;
     m->_interpreter_throwout_count = rec->_interpreter_throwout_count;
     MethodCounters* mcs = method->get_method_counters(CHECK_AND_CLEAR);

src/hotspot/share/code/compiledMethod.hpp

@@ -281,6 +281,8 @@ class CompiledMethod : public CodeBlob {
   bool consts_contains(address addr) const { return consts_begin() <= addr && addr < consts_end(); }
   int consts_size() const { return consts_end() - consts_begin(); }
 
+  virtual int skipped_instructions_size() const = 0;
+
   virtual address stub_begin() const = 0;
   virtual address stub_end() const = 0;
   bool stub_contains(address addr) const { return stub_begin() <= addr && addr < stub_end(); }