MIPS: Share optimized code for closures.

Port r11817 (616dfcf0)

Original commit message:
Each SharedFunctionInfo gets an optimized code map to store
one optimized code object per context. When allocating a new
closure we consult this map and check if there is optimized code
that can be shared.

BUG=
TEST=

Review URL: https://chromiumcodereview.appspot.com/10557002
Patch from Akos Palfi <palfia@homejinni.com>.

git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@11848 ce2b1a6d-e550-0410-aec6-3dcde31c8c00

diff --git a/src/mips/code-stubs-mips.cc b/src/mips/code-stubs-mips.cc
index a7c2597..2155b18 100644 (file)
--- a/src/mips/code-stubs-mips.cc
+++ b/src/mips/code-stubs-mips.cc
@@ -87,6 +87,8 @@ void ToNumberStub::Generate(MacroAssembler* masm) {
 void FastNewClosureStub::Generate(MacroAssembler* masm) {
   // Create a new closure from the given function info in new
   // space. Set the context to the current context in cp.
+  Counters* counters = masm->isolate()->counters();
+
   Label gc;
 
   // Pop the function info from the stack.
@@ -100,6 +102,8 @@ void FastNewClosureStub::Generate(MacroAssembler* masm) {
                         &gc,
                         TAG_OBJECT);
 
+  __ IncrementCounter(counters->fast_new_closure_total(), 1, t2, t3);
+
   int map_index = (language_mode_ == CLASSIC_MODE)
       ? Context::FUNCTION_MAP_INDEX
       : Context::STRICT_MODE_FUNCTION_MAP_INDEX;
@@ -108,24 +112,34 @@ void FastNewClosureStub::Generate(MacroAssembler* masm) {
   // as the map of the allocated object.
   __ lw(a2, MemOperand(cp, Context::SlotOffset(Context::GLOBAL_INDEX)));
   __ lw(a2, FieldMemOperand(a2, GlobalObject::kGlobalContextOffset));
-  __ lw(a2, MemOperand(a2, Context::SlotOffset(map_index)));
-  __ sw(a2, FieldMemOperand(v0, HeapObject::kMapOffset));
+  __ lw(t1, MemOperand(a2, Context::SlotOffset(map_index)));
+  __ sw(t1, FieldMemOperand(v0, HeapObject::kMapOffset));
 
   // Initialize the rest of the function. We don't have to update the
   // write barrier because the allocated object is in new space.
   __ LoadRoot(a1, Heap::kEmptyFixedArrayRootIndex);
-  __ LoadRoot(a2, Heap::kTheHoleValueRootIndex);
-  __ LoadRoot(t0, Heap::kUndefinedValueRootIndex);
+  __ LoadRoot(t1, Heap::kTheHoleValueRootIndex);
   __ sw(a1, FieldMemOperand(v0, JSObject::kPropertiesOffset));
   __ sw(a1, FieldMemOperand(v0, JSObject::kElementsOffset));
-  __ sw(a2, FieldMemOperand(v0, JSFunction::kPrototypeOrInitialMapOffset));
+  __ sw(t1, FieldMemOperand(v0, JSFunction::kPrototypeOrInitialMapOffset));
   __ sw(a3, FieldMemOperand(v0, JSFunction::kSharedFunctionInfoOffset));
   __ sw(cp, FieldMemOperand(v0, JSFunction::kContextOffset));
   __ sw(a1, FieldMemOperand(v0, JSFunction::kLiteralsOffset));
-  __ sw(t0, FieldMemOperand(v0, JSFunction::kNextFunctionLinkOffset));
 
   // Initialize the code pointer in the function to be the one
   // found in the shared function info object.
+  // But first check if there is an optimized version for our context.
+  Label check_optimized;
+  Label install_unoptimized;
+  if (FLAG_cache_optimized_code) {
+    __ lw(a1,
+          FieldMemOperand(a3, SharedFunctionInfo::kOptimizedCodeMapOffset));
+    __ And(at, a1, a1);
+    __ Branch(&check_optimized, ne, at, Operand(zero_reg));
+  }
+  __ bind(&install_unoptimized);
+  __ LoadRoot(t0, Heap::kUndefinedValueRootIndex);
+  __ sw(t0, FieldMemOperand(v0, JSFunction::kNextFunctionLinkOffset));
   __ lw(a3, FieldMemOperand(a3, SharedFunctionInfo::kCodeOffset));
   __ Addu(a3, a3, Operand(Code::kHeaderSize - kHeapObjectTag));
 
@@ -133,6 +147,73 @@ void FastNewClosureStub::Generate(MacroAssembler* masm) {
   __ sw(a3, FieldMemOperand(v0, JSFunction::kCodeEntryOffset));
   __ Ret();
 
+  __ bind(&check_optimized);
+
+  __ IncrementCounter(counters->fast_new_closure_try_optimized(), 1, t2, t3);
+
+  // a2 holds global context, a1 points to fixed array of 3-element entries
+  // (global context, optimized code, literals).
+  // The optimized code map must never be empty, so check the first elements.
+  Label install_optimized;
+  // Speculatively move code object into t0.
+  __ lw(t0, FieldMemOperand(a1, FixedArray::kHeaderSize + kPointerSize));
+  __ lw(t1, FieldMemOperand(a1, FixedArray::kHeaderSize));
+  __ Branch(&install_optimized, eq, a2, Operand(t1));
+  __ Branch(&install_unoptimized);
+
+  // Iterate through the rest of map backwards.  t0 holds an index as a Smi.
+  Label loop;
+  __ lw(t0, FieldMemOperand(a1, FixedArray::kLengthOffset));
+  __ bind(&loop);
+  // Do not double check first entry.
+
+  __ Branch(&install_unoptimized, eq, t0,
+            Operand(Smi::FromInt(SharedFunctionInfo::kEntryLength)));
+  __ Subu(t0, t0, Operand(
+      Smi::FromInt(SharedFunctionInfo::kEntryLength)));  // Skip an entry.
+  __ Addu(t1, a1, Operand(FixedArray::kHeaderSize - kHeapObjectTag));
+  __ sll(at, t0, kPointerSizeLog2 - kSmiTagSize);
+  __ Addu(t1, t1, Operand(at));
+  __ lw(t1, MemOperand(t1));
+  __ Branch(&loop, ne, a2, Operand(t1));
+  // Hit: fetch the optimized code.
+  __ Addu(t1, a1, Operand(FixedArray::kHeaderSize - kHeapObjectTag));
+  __ sll(at, t0, kPointerSizeLog2 - kSmiTagSize);
+  __ Addu(t1, t1, Operand(at));
+  __ Addu(t1, t1, Operand(kPointerSize));
+  __ lw(t0, MemOperand(t1));
+
+  __ bind(&install_optimized);
+  __ IncrementCounter(counters->fast_new_closure_install_optimized(),
+                      1, t2, t3);
+
+  // TODO(fschneider): Idea: store proper code pointers in the map and either
+  // unmangle them on marking or do nothing as the whole map is discarded on
+  // major GC anyway.
+  __ Addu(t0, t0, Operand(Code::kHeaderSize - kHeapObjectTag));
+  __ sw(t0, FieldMemOperand(v0, JSFunction::kCodeEntryOffset));
+
+  // Now link a function into a list of optimized functions.
+  __ lw(t0, ContextOperand(a2, Context::OPTIMIZED_FUNCTIONS_LIST));
+
+  __ sw(t0, FieldMemOperand(v0, JSFunction::kNextFunctionLinkOffset));
+  // No need for write barrier as JSFunction (eax) is in the new space.
+
+  __ sw(v0, ContextOperand(a2, Context::OPTIMIZED_FUNCTIONS_LIST));
+  // Store JSFunction (eax) into edx before issuing write barrier as
+  // it clobbers all the registers passed.
+  __ mov(t0, v0);
+  __ RecordWriteContextSlot(
+      a2,
+      Context::SlotOffset(Context::OPTIMIZED_FUNCTIONS_LIST),
+      t0,
+      a1,
+      kRAHasNotBeenSaved,
+      kDontSaveFPRegs);
+
+  // Return result. The argument function info has been popped already.
+  __ Ret();
+
   // Create a new closure through the slower runtime call.
   __ bind(&gc);
   __ LoadRoot(t0, Heap::kFalseValueRootIndex);
@@ -7380,6 +7461,8 @@ static const AheadOfTimeWriteBarrierStubList kAheadOfTime[] = {
   { REG(a2), REG(t2), REG(t5), EMIT_REMEMBERED_SET },
   // StoreArrayLiteralElementStub::Generate
   { REG(t1), REG(a0), REG(t2), EMIT_REMEMBERED_SET },
+  // FastNewClosureStub::Generate
+  { REG(a2), REG(t0), REG(a1), EMIT_REMEMBERED_SET },
   // Null termination.
   { REG(no_reg), REG(no_reg), REG(no_reg), EMIT_REMEMBERED_SET}
 };

diff --git a/src/mips/deoptimizer-mips.cc b/src/mips/deoptimizer-mips.cc
index 62f3155..b57d2d3 100644 (file)
--- a/src/mips/deoptimizer-mips.cc
+++ b/src/mips/deoptimizer-mips.cc
@@ -48,6 +48,10 @@ void Deoptimizer::DeoptimizeFunction(JSFunction* function) {
 
   if (!function->IsOptimized()) return;
 
+  // The optimized code is going to be patched, so we cannot use it
+  // any more.  Play safe and reset the whole cache.
+  function->shared()->ClearOptimizedCodeMap();
+
   // Get the optimized code.
   Code* code = function->code();
   Address code_start_address = code->instruction_start();
@@ -96,8 +100,19 @@ void Deoptimizer::DeoptimizeFunction(JSFunction* function) {
   // ignore all slots that might have been recorded on it.
   isolate->heap()->mark_compact_collector()->InvalidateCode(code);
 
-  // Set the code for the function to non-optimized version.
-  function->ReplaceCode(function->shared()->code());
+  // Iterate over all the functions which share the same code object
+  // and make them use unoptimized version.
+  Context* context = function->context()->global_context();
+  Object* element = context->get(Context::OPTIMIZED_FUNCTIONS_LIST);
+  SharedFunctionInfo* shared = function->shared();
+  while (!element->IsUndefined()) {
+    JSFunction* func = JSFunction::cast(element);
+    // Grab element before code replacement as ReplaceCode alters the list.
+    element = func->next_function_link();
+    if (func->code() == code) {
+      func->ReplaceCode(shared->code());
+    }
+  }
 
   if (FLAG_trace_deopt) {
     PrintF("[forced deoptimization: ");
@@ -229,9 +244,9 @@ void Deoptimizer::DoComputeOsrOutputFrame() {
   unsigned node_id = iterator.Next();
   USE(node_id);
   ASSERT(node_id == ast_id);
-  JSFunction* function = JSFunction::cast(ComputeLiteral(iterator.Next()));
-  USE(function);
-  ASSERT(function == function_);
+  int closure_id = iterator.Next();
+  USE(closure_id);
+  ASSERT_EQ(Translation::kSelfLiteralId, closure_id);
   unsigned height = iterator.Next();
   unsigned height_in_bytes = height * kPointerSize;
   USE(height_in_bytes);
@@ -342,8 +357,8 @@ void Deoptimizer::DoComputeOsrOutputFrame() {
   if (FLAG_trace_osr) {
     PrintF("[on-stack replacement translation %s: 0x%08" V8PRIxPTR " ",
            ok ? "finished" : "aborted",
-           reinterpret_cast<intptr_t>(function));
-    function->PrintName();
+           reinterpret_cast<intptr_t>(function_));
+    function_->PrintName();
     PrintF(" => pc=0x%0x]\n", output_[0]->GetPc());
   }
 }
@@ -573,7 +588,15 @@ void Deoptimizer::DoComputeJSFrame(TranslationIterator* iterator,
                                    int frame_index) {
   // Read the ast node id, function, and frame height for this output frame.
   int node_id = iterator->Next();
-  JSFunction* function = JSFunction::cast(ComputeLiteral(iterator->Next()));
+  JSFunction* function;
+  if (frame_index != 0) {
+    function = JSFunction::cast(ComputeLiteral(iterator->Next()));
+  } else {
+    int closure_id = iterator->Next();
+    USE(closure_id);
+    ASSERT_EQ(Translation::kSelfLiteralId, closure_id);
+    function = function_;
+  }
   unsigned height = iterator->Next();
   unsigned height_in_bytes = height * kPointerSize;
   if (FLAG_trace_deopt) {

diff --git a/src/mips/lithium-codegen-mips.cc b/src/mips/lithium-codegen-mips.cc
index 67dbe69..2a022b5 100644 (file)
--- a/src/mips/lithium-codegen-mips.cc
+++ b/src/mips/lithium-codegen-mips.cc
@@ -446,7 +446,10 @@ void LCodeGen::WriteTranslation(LEnvironment* environment,
   int height = translation_size - environment->parameter_count();
 
   WriteTranslation(environment->outer(), translation);
-  int closure_id = DefineDeoptimizationLiteral(environment->closure());
+  int closure_id = *info()->closure() != *environment->closure()
+      ? DefineDeoptimizationLiteral(environment->closure())
+      : Translation::kSelfLiteralId;
+
   switch (environment->frame_type()) {
     case JS_FUNCTION:
       translation->BeginJSFrame(environment->ast_id(), closure_id, height);
@@ -2847,7 +2850,7 @@ void LCodeGen::DoDrop(LDrop* instr) {
 
 void LCodeGen::DoThisFunction(LThisFunction* instr) {
   Register result = ToRegister(instr->result());
-  __ LoadHeapObject(result, instr->hydrogen()->closure());
+  __ lw(result, MemOperand(fp, JavaScriptFrameConstants::kFunctionOffset));
 }