Reimplement dynamic frame alignment for frames that are compiled via OSR or have more than 2 double spill slots.

The first spill slot is now reserved on all optimized frames to distinguish frames that were aligned.

Review URL: https://chromiumcodereview.appspot.com/10532066

git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@11772 ce2b1a6d-e550-0410-aec6-3dcde31c8c00

src/debug.cc

@@ -1451,7 +1451,7 @@ void Debug::PrepareStep(StepAction step_action, int step_count) {
     // Remember source position and frame to handle step next.
     thread_local_.last_statement_position_ =
         debug_info->code()->SourceStatementPosition(frame->pc());
-    thread_local_.last_fp_ = frame->fp();
+    thread_local_.last_fp_ = frame->UnpaddedFP();
   } else {
     // If there's restarter frame on top of the stack, just get the pointer
     // to function which is going to be restarted.
@@ -1520,7 +1520,7 @@ void Debug::PrepareStep(StepAction step_action, int step_count) {
       // propagated on the next Debug::Break.
       thread_local_.last_statement_position_ =
           debug_info->code()->SourceStatementPosition(frame->pc());
-      thread_local_.last_fp_ = frame->fp();
+      thread_local_.last_fp_ = frame->UnpaddedFP();
     }
 
     // Step in or Step in min
@@ -1555,7 +1555,7 @@ bool Debug::StepNextContinue(BreakLocationIterator* break_location_iterator,
     // Continue if we are still on the same frame and in the same statement.
     int current_statement_position =
         break_location_iterator->code()->SourceStatementPosition(frame->pc());
-    return thread_local_.last_fp_ == frame->fp() &&
+    return thread_local_.last_fp_ == frame->UnpaddedFP() &&
         thread_local_.last_statement_position_ == current_statement_position;
   }
 
@@ -1756,7 +1756,7 @@ void Debug::ClearOneShot() {
 
 void Debug::ActivateStepIn(StackFrame* frame) {
   ASSERT(!StepOutActive());
-  thread_local_.step_into_fp_ = frame->fp();
+  thread_local_.step_into_fp_ = frame->UnpaddedFP();
 }
 
 
@@ -1767,7 +1767,7 @@ void Debug::ClearStepIn() {
 
 void Debug::ActivateStepOut(StackFrame* frame) {
   ASSERT(!StepInActive());
-  thread_local_.step_out_fp_ = frame->fp();
+  thread_local_.step_out_fp_ = frame->UnpaddedFP();
 }
 
 

src/deoptimizer.cc

@@ -354,6 +354,7 @@ Deoptimizer::Deoptimizer(Isolate* isolate,
       bailout_type_(type),
       from_(from),
       fp_to_sp_delta_(fp_to_sp_delta),
+      has_alignment_padding_(0),
       input_(NULL),
       output_count_(0),
       jsframe_count_(0),
@@ -594,12 +595,14 @@ void Deoptimizer::DoComputeOutputFrames() {
     PrintF("[deoptimizing: end 0x%08" V8PRIxPTR " ",
            reinterpret_cast<intptr_t>(function));
     function->PrintName();
-    PrintF(" => node=%u, pc=0x%08" V8PRIxPTR ", state=%s, took %0.3f ms]\n",
+    PrintF(" => node=%u, pc=0x%08" V8PRIxPTR ", state=%s, alignment=%s,"
+           " took %0.3f ms]\n",
            node_id,
            output_[index]->GetPc(),
            FullCodeGenerator::State2String(
                static_cast<FullCodeGenerator::State>(
                    output_[index]->GetState()->value())),
+           has_alignment_padding_ ? "with padding" : "no padding",
            ms);
   }
 }
@@ -770,7 +773,7 @@ void Deoptimizer::DoTranslateCommand(TranslationIterator* iterator,
       if (FLAG_trace_deopt) {
         PrintF("    0x%08" V8PRIxPTR ": ",
                output_[frame_index]->GetTop() + output_offset);
-        PrintF("[top + %d] <- 0x%08" V8PRIxPTR " ; [esp + %d] ",
+        PrintF("[top + %d] <- 0x%08" V8PRIxPTR " ; [sp + %d] ",
                output_offset,
                input_value,
                input_offset);
@@ -790,7 +793,7 @@ void Deoptimizer::DoTranslateCommand(TranslationIterator* iterator,
       if (FLAG_trace_deopt) {
         PrintF("    0x%08" V8PRIxPTR ": ",
                output_[frame_index]->GetTop() + output_offset);
-        PrintF("[top + %d] <- %" V8PRIdPTR " ; [esp + %d] (%s)\n",
+        PrintF("[top + %d] <- %" V8PRIdPTR " ; [sp + %d] (%s)\n",
                output_offset,
                value,
                input_offset,
@@ -816,7 +819,7 @@ void Deoptimizer::DoTranslateCommand(TranslationIterator* iterator,
           input_->GetOffsetFromSlotIndex(input_slot_index);
       double value = input_->GetDoubleFrameSlot(input_offset);
       if (FLAG_trace_deopt) {
-        PrintF("    0x%08" V8PRIxPTR ": [top + %d] <- %e ; [esp + %d]\n",
+        PrintF("    0x%08" V8PRIxPTR ": [top + %d] <- %e ; [sp + %d]\n",
                output_[frame_index]->GetTop() + output_offset,
                output_offset,
                value,

src/deoptimizer.h

@@ -221,6 +221,10 @@ class Deoptimizer : public Malloced {
   }
   static int output_offset() { return OFFSET_OF(Deoptimizer, output_); }
 
+  static int has_alignment_padding_offset() {
+    return OFFSET_OF(Deoptimizer, has_alignment_padding_);
+  }
+
   static int GetDeoptimizedCodeCount(Isolate* isolate);
 
   static const int kNotDeoptimizationEntry = -1;
@@ -322,6 +326,7 @@ class Deoptimizer : public Malloced {
   BailoutType bailout_type_;
   Address from_;
   int fp_to_sp_delta_;
+  int has_alignment_padding_;
 
   // Input frame description.
   FrameDescription* input_;

src/frames.cc

@@ -469,6 +469,20 @@ StackFrame::Type StackFrame::GetCallerState(State* state) const {
 }
 
 
+Address StackFrame::UnpaddedFP() const {
+#if defined(V8_TARGET_ARCH_IA32) || defined(V8_HOST_ARCH_IA32)
+  if (!is_optimized()) return fp();
+  int32_t alignment_state = Memory::int32_at(
+    fp() + JavaScriptFrameConstants::kDynamicAlignmentStateOffset);
+
+  return (alignment_state == kAlignmentPaddingPushed) ?
+    (fp() + kPointerSize) : fp();
+#else
+  return fp();
+#endif
+}
+
+
 Code* EntryFrame::unchecked_code() const {
   return HEAP->raw_unchecked_js_entry_code();
 }

src/frames.h

@@ -206,6 +206,11 @@ class StackFrame BASE_EMBEDDED {
   Address fp() const { return state_.fp; }
   Address caller_sp() const { return GetCallerStackPointer(); }
 
+  // If this frame is optimized and was dynamically aligned return its old
+  // unaligned frame pointer.  When the frame is deoptimized its FP will shift
+  // up one word and become unaligned.
+  Address UnpaddedFP() const;
+
   Address pc() const { return *pc_address(); }
   void set_pc(Address pc) { *pc_address() = pc; }
 

src/hydrogen.cc

@@ -681,7 +681,8 @@ HGraph::HGraph(CompilationInfo* info, Zone* zone)
       blocks_(8, zone),
       values_(16, zone),
       phi_list_(NULL),
-      zone_(zone) {
+      zone_(zone),
+      is_recursive_(false) {
   start_environment_ =
       new(zone) HEnvironment(NULL, info->scope(), info->closure(), zone);
   start_environment_->set_ast_id(AstNode::kFunctionEntryId);
@@ -6927,6 +6928,11 @@ void HGraphBuilder::VisitCall(Call* expr) {
           return;
         }
         if (TryInlineCall(expr)) return;
+
+        if (expr->target().is_identical_to(info()->closure())) {
+          graph()->MarkRecursive();
+        }
+
         call = PreProcessCall(new(zone()) HCallKnownGlobal(expr->target(),
                                                            argument_count));
       } else {

src/hydrogen.h

@@ -336,6 +336,14 @@ class HGraph: public ZoneObject {
     osr_values_.set(values);
   }
 
+  void MarkRecursive() {
+    is_recursive_ = true;
+  }
+
+  bool is_recursive() const {
+    return is_recursive_;
+  }
+
  private:
   void Postorder(HBasicBlock* block,
                  BitVector* visited,
@@ -382,6 +390,8 @@ class HGraph: public ZoneObject {
 
   Zone* zone_;
 
+  bool is_recursive_;
+
   DISALLOW_COPY_AND_ASSIGN(HGraph);
 };
 

src/ia32/deoptimizer-ia32.cc

@@ -351,10 +351,12 @@ void Deoptimizer::DoComputeOsrOutputFrame() {
     PrintF("[on-stack replacement: begin 0x%08" V8PRIxPTR " ",
            reinterpret_cast<intptr_t>(function_));
     function_->PrintName();
-    PrintF(" => node=%u, frame=%d->%d]\n",
+    PrintF(" => node=%u, frame=%d->%d, ebp:esp=0x%08x:0x%08x]\n",
            ast_id,
            input_frame_size,
-           output_frame_size);
+           output_frame_size,
+           input_->GetRegister(ebp.code()),
+           input_->GetRegister(esp.code()));
   }
 
   // There's only one output frame in the OSR case.
@@ -404,7 +406,7 @@ void Deoptimizer::DoComputeOsrOutputFrame() {
           name = "function";
           break;
       }
-      PrintF("    [esp + %d] <- 0x%08x ; [esp + %d] (fixed part - %s)\n",
+      PrintF("    [sp + %d] <- 0x%08x ; [sp + %d] (fixed part - %s)\n",
              output_offset,
              input_value,
              input_offset,
@@ -415,6 +417,24 @@ void Deoptimizer::DoComputeOsrOutputFrame() {
     output_offset -= kPointerSize;
   }
 
+  // All OSR stack frames are dynamically aligned to an 8-byte boundary.
+  int frame_pointer = input_->GetRegister(ebp.code());
+  if ((frame_pointer & kPointerSize) != 0) {
+    frame_pointer -= kPointerSize;
+    has_alignment_padding_ = 1;
+  }
+
+  int32_t alignment_state = (has_alignment_padding_ == 1) ?
+    kAlignmentPaddingPushed :
+    kNoAlignmentPadding;
+  if (FLAG_trace_osr) {
+    PrintF("    [sp + %d] <- 0x%08x ; (alignment state)\n",
+           output_offset,
+           alignment_state);
+  }
+  output_[0]->SetFrameSlot(output_offset, alignment_state);
+  output_offset -= kPointerSize;
+
   // Translate the rest of the frame.
   while (ok && input_offset >= 0) {
     ok = DoOsrTranslateCommand(&iterator, &input_offset);
@@ -427,7 +447,7 @@ void Deoptimizer::DoComputeOsrOutputFrame() {
     output_[0]->SetPc(reinterpret_cast<uint32_t>(from_));
   } else {
     // Set up the frame pointer and the context pointer.
-    output_[0]->SetRegister(ebp.code(), input_->GetRegister(ebp.code()));
+    output_[0]->SetRegister(ebp.code(), frame_pointer);
     output_[0]->SetRegister(esi.code(), input_->GetRegister(esi.code()));
 
     unsigned pc_offset = data->OsrPcOffset()->value();
@@ -688,24 +708,38 @@ void Deoptimizer::DoComputeJSFrame(TranslationIterator* iterator,
   ASSERT(output_[frame_index] == NULL);
   output_[frame_index] = output_frame;
 
+  // Compute the incoming parameter translation.
+  int parameter_count = function->shared()->formal_parameter_count() + 1;
+  unsigned output_offset = output_frame_size;
+  unsigned input_offset = input_frame_size;
+
+  unsigned alignment_state_offset =
+    input_offset - parameter_count * kPointerSize -
+    StandardFrameConstants::kFixedFrameSize -
+    kPointerSize;
+  ASSERT(JavaScriptFrameConstants::kDynamicAlignmentStateOffset ==
+    JavaScriptFrameConstants::kLocal0Offset);
+
   // The top address for the bottommost output frame can be computed from
   // the input frame pointer and the output frame's height.  For all
   // subsequent output frames, it can be computed from the previous one's
   // top address and the current frame's size.
   uint32_t top_address;
   if (is_bottommost) {
+    int32_t alignment_state = input_->GetFrameSlot(alignment_state_offset);
+    has_alignment_padding_ =
+      (alignment_state == kAlignmentPaddingPushed) ? 1 : 0;
     // 2 = context and function in the frame.
-    top_address =
+    // If the optimized frame had alignment padding, adjust the frame pointer
-        input_->GetRegister(ebp.code()) - (2 * kPointerSize) - height_in_bytes;
+    // to point to the new position of the old frame pointer after padding
+    // is removed. Subtract 2 * kPointerSize for the context and function slots.
+    top_address = input_->GetRegister(ebp.code()) - (2 * kPointerSize) -
+        height_in_bytes + has_alignment_padding_ * kPointerSize;
   } else {
     top_address = output_[frame_index - 1]->GetTop() - output_frame_size;
   }
   output_frame->SetTop(top_address);
 
-  // Compute the incoming parameter translation.
-  int parameter_count = function->shared()->formal_parameter_count() + 1;
-  unsigned output_offset = output_frame_size;
-  unsigned input_offset = input_frame_size;
   for (int i = 0; i < parameter_count; ++i) {
     output_offset -= kPointerSize;
     DoTranslateCommand(iterator, frame_index, output_offset);
@@ -747,13 +781,17 @@ void Deoptimizer::DoComputeJSFrame(TranslationIterator* iterator,
   }
   output_frame->SetFrameSlot(output_offset, value);
   intptr_t fp_value = top_address + output_offset;
-  ASSERT(!is_bottommost || input_->GetRegister(ebp.code()) == fp_value);
+  ASSERT(!is_bottommost ||
+    (input_->GetRegister(ebp.code()) + has_alignment_padding_ * kPointerSize) ==
+    fp_value);
   output_frame->SetFp(fp_value);
   if (is_topmost) output_frame->SetRegister(ebp.code(), fp_value);
   if (FLAG_trace_deopt) {
     PrintF("    0x%08x: [top + %d] <- 0x%08x ; caller's fp\n",
            fp_value, output_offset, value);
   }
+  ASSERT(!is_bottommost || !has_alignment_padding_ ||
+    (fp_value & kPointerSize) != 0);
 
   // For the bottommost output frame the context can be gotten from the input
   // frame. For all subsequent output frames it can be gotten from the function
@@ -948,6 +986,28 @@ void Deoptimizer::EntryGenerator::Generate() {
   }
   __ pop(eax);
 
+  if (type() != OSR) {
+    // If frame was dynamically aligned, pop padding.
+    Label no_padding;
+    __ cmp(Operand(eax, Deoptimizer::has_alignment_padding_offset()),
+           Immediate(0));
+    __ j(equal, &no_padding);
+    __ pop(ecx);
+    if (FLAG_debug_code) {
+      __ cmp(ecx, Immediate(kAlignmentZapValue));
+      __ Assert(equal, "alignment marker expected");
+    }
+    __ bind(&no_padding);
+  } else {
+    // If frame needs dynamic alignment push padding.
+    Label no_padding;
+    __ cmp(Operand(eax, Deoptimizer::has_alignment_padding_offset()),
+           Immediate(0));
+    __ j(equal, &no_padding);
+    __ push(Immediate(kAlignmentZapValue));
+    __ bind(&no_padding);
+  }
+
   // Replace the current frame with the output frames.
   Label outer_push_loop, inner_push_loop;
   // Outer loop state: eax = current FrameDescription**, edx = one past the

src/ia32/frames-ia32.h

@@ -53,6 +53,10 @@ typedef Object* JSCallerSavedBuffer[kNumJSCallerSaved];
 // Number of registers for which space is reserved in safepoints.
 const int kNumSafepointRegisters = 8;
 
+const int kNoAlignmentPadding = 0;
+const int kAlignmentPaddingPushed = 2;
+const int kAlignmentZapValue = 0x12345678;  // Not heap object tagged.
+
 // ----------------------------------------------------
 
 
@@ -119,6 +123,8 @@ class JavaScriptFrameConstants : public AllStatic {
   // Caller SP-relative.
   static const int kParam0Offset   = -2 * kPointerSize;
   static const int kReceiverOffset = -1 * kPointerSize;
+
+  static const int kDynamicAlignmentStateOffset = kLocal0Offset;
 };
 
 

src/ia32/lithium-codegen-ia32.cc

@@ -79,6 +79,10 @@ bool LCodeGen::GenerateCode() {
   // the frame (that is done in GeneratePrologue).
   FrameScope frame_scope(masm_, StackFrame::MANUAL);
 
+  dynamic_frame_alignment_ = (chunk()->num_double_slots() > 2 &&
+                              !chunk()->graph()->is_recursive()) ||
+                             info()->osr_ast_id() != AstNode::kNoNumber;
+
   return GeneratePrologue() &&
       GenerateBody() &&
       GenerateDeferredCode() &&
@@ -153,14 +157,52 @@ bool LCodeGen::GeneratePrologue() {
     __ bind(&ok);
   }
 
+
+  if (dynamic_frame_alignment_) {
+    // Move state of dynamic frame alignment into edx.
+    __ mov(edx, Immediate(kNoAlignmentPadding));
+
+    Label do_not_pad, align_loop;
+    STATIC_ASSERT(kDoubleSize == 2 * kPointerSize);
+    // Align esp + 4 to a multiple of 2 * kPointerSize.
+    __ test(esp, Immediate(kPointerSize));
+    __ j(not_zero, &do_not_pad, Label::kNear);
+    __ push(Immediate(0));
+    __ mov(ebx, esp);
+    __ mov(edx, Immediate(kAlignmentPaddingPushed));
+    // Copy arguments, receiver, and return address.
+    __ mov(ecx, Immediate(scope()->num_parameters() + 2));
+
+    __ bind(&align_loop);
+    __ mov(eax, Operand(ebx, 1 * kPointerSize));
+    __ mov(Operand(ebx, 0), eax);
+    __ add(Operand(ebx), Immediate(kPointerSize));
+    __ dec(ecx);
+    __ j(not_zero, &align_loop, Label::kNear);
+    __ mov(Operand(ebx, 0), Immediate(kAlignmentZapValue));
+    __ bind(&do_not_pad);
+  }
+
   __ push(ebp);  // Caller's frame pointer.
   __ mov(ebp, esp);
   __ push(esi);  // Callee's context.
   __ push(edi);  // Callee's JS function.
 
+  if (dynamic_frame_alignment_ && FLAG_debug_code) {
+    __ test(esp, Immediate(kPointerSize));
+    __ Assert(zero, "frame is expected to be aligned");
+  }
+
   // Reserve space for the stack slots needed by the code.
   int slots = GetStackSlotCount();
-  if (slots > 0) {
+  ASSERT_GE(slots, 1);
+  if (slots == 1) {
+    if (dynamic_frame_alignment_) {
+      __ push(edx);
+    } else {
+      __ push(Immediate(kNoAlignmentPadding));
+    }
+  } else {
     if (FLAG_debug_code) {
       __ mov(Operand(eax), Immediate(slots));
       Label loop;
@@ -170,7 +212,7 @@ bool LCodeGen::GeneratePrologue() {
       __ j(not_zero, &loop);
     } else {
       __ sub(Operand(esp), Immediate(slots * kPointerSize));
-#ifdef _MSC_VER
+  #ifdef _MSC_VER
       // On windows, you may not access the stack more than one page below
       // the most recently mapped page. To make the allocated area randomly
       // accessible, we write to each page in turn (the value is irrelevant).
@@ -180,7 +222,18 @@ bool LCodeGen::GeneratePrologue() {
            offset -= kPageSize) {
         __ mov(Operand(esp, offset), eax);
       }
-#endif
+  #endif
+    }
+
+    // Store dynamic frame alignment state in the first local.
+    if (dynamic_frame_alignment_) {
+      __ mov(Operand(ebp,
+                     JavaScriptFrameConstants::kDynamicAlignmentStateOffset),
+             edx);
+    } else {
+      __ mov(Operand(ebp,
+                     JavaScriptFrameConstants::kDynamicAlignmentStateOffset),
+             Immediate(kNoAlignmentPadding));
     }
   }
 
@@ -2101,8 +2154,25 @@ void LCodeGen::DoReturn(LReturn* instr) {
     __ mov(esi, Operand(ebp, StandardFrameConstants::kContextOffset));
     __ CallRuntime(Runtime::kTraceExit, 1);
   }
+  if (dynamic_frame_alignment_) {
+    // Fetch the state of the dynamic frame alignment.
+    __ mov(edx, Operand(ebp,
+      JavaScriptFrameConstants::kDynamicAlignmentStateOffset));
+  }
   __ mov(esp, ebp);
   __ pop(ebp);
+  if (dynamic_frame_alignment_) {
+    Label no_padding;
+    __ cmp(edx, Immediate(kNoAlignmentPadding));
+    __ j(equal, &no_padding);
+    if (FLAG_debug_code) {
+      __ cmp(Operand(esp, (GetParameterCount() + 2) * kPointerSize),
+             Immediate(kAlignmentZapValue));
+      __ Assert(equal, "expected alignment marker");
+    }
+    __ Ret((GetParameterCount() + 2) * kPointerSize, ecx);
+    __ bind(&no_padding);
+  }
   __ Ret((GetParameterCount() + 1) * kPointerSize, ecx);
 }
 

src/ia32/lithium-codegen-ia32.h

@@ -61,6 +61,7 @@ class LCodeGen BASE_EMBEDDED {
         status_(UNUSED),
         translations_(zone),
         deferred_(8, zone),
+        dynamic_frame_alignment_(false),
         osr_pc_offset_(-1),
         last_lazy_deopt_pc_(0),
         safepoints_(zone),
@@ -342,6 +343,7 @@ class LCodeGen BASE_EMBEDDED {
   Status status_;
   TranslationBuffer translations_;
   ZoneList<LDeferredCode*> deferred_;
+  bool dynamic_frame_alignment_;
   int osr_pc_offset_;
   int last_lazy_deopt_pc_;
 

src/ia32/lithium-ia32.cc

@@ -368,7 +368,11 @@ void LAccessArgumentsAt::PrintDataTo(StringStream* stream) {
 
 int LChunk::GetNextSpillIndex(bool is_double) {
   // Skip a slot if for a double-width slot.
-  if (is_double) spill_slot_count_++;
+  if (is_double) {
+    spill_slot_count_++;
+    spill_slot_count_ |= 1;
+    num_double_slots_++;
+  }
   return spill_slot_count_++;
 }
 
@@ -550,6 +554,12 @@ LChunk* LChunkBuilder::Build() {
   chunk_ = new(zone()) LChunk(info(), graph());
   HPhase phase("L_Building chunk", chunk_);
   status_ = BUILDING;
+
+  // Reserve the first spill slot for the state of dynamic alignment.
+  int alignment_state_index = chunk_->GetNextSpillIndex(false);
+  ASSERT_EQ(alignment_state_index, 0);
+  USE(alignment_state_index);
+
   const ZoneList<HBasicBlock*>* blocks = graph()->blocks();
   for (int i = 0; i < blocks->length(); i++) {
     HBasicBlock* next = NULL;

src/ia32/lithium-ia32.h

@@ -2307,6 +2307,7 @@ class LChunk: public ZoneObject {
  public:
   LChunk(CompilationInfo* info, HGraph* graph)
     : spill_slot_count_(0),
+      num_double_slots_(0),
       info_(info),
       graph_(graph),
       instructions_(32, graph->zone()),
@@ -2324,6 +2325,7 @@ class LChunk: public ZoneObject {
   int ParameterAt(int index);
   int GetParameterStackSlot(int index) const;
   int spill_slot_count() const { return spill_slot_count_; }
+  int num_double_slots() const { return num_double_slots_; }
   CompilationInfo* info() const { return info_; }
   HGraph* graph() const { return graph_; }
   const ZoneList<LInstruction*>* instructions() const { return &instructions_; }
@@ -2363,6 +2365,7 @@ class LChunk: public ZoneObject {
 
  private:
   int spill_slot_count_;
+  int num_double_slots_;
   CompilationInfo* info_;
   HGraph* const graph_;
   ZoneList<LInstruction*> instructions_;