Partial ia32 implementation of optimized try/catch (by Kevin Millikin)

src/ia32/deoptimizer-ia32.cc

Index: src/ia32/deoptimizer-ia32.cc
diff --git a/src/ia32/deoptimizer-ia32.cc b/src/ia32/deoptimizer-ia32.cc
index f50010b9d97b884e2dda6109b5bf9e43ca8c7fd8..6b550c2ddc424a44c28486fc2e0e7d0e76376827 100644
--- a/src/ia32/deoptimizer-ia32.cc
+++ b/src/ia32/deoptimizer-ia32.cc
@@ -351,6 +351,8 @@ void Deoptimizer::DoComputeOsrOutputFrame() {
   unsigned height = iterator.Next();
   unsigned height_in_bytes = height * kPointerSize;
   USE(height_in_bytes);
+  int handler_count = iterator.Next();
+  USE(handler_count);
 
   unsigned fixed_size = ComputeFixedSize(function_);
   unsigned input_frame_size = input_->GetFrameSize();
@@ -480,7 +482,7 @@ void Deoptimizer::DoComputeOsrOutputFrame() {
            ok ? "finished" : "aborted",
            reinterpret_cast<intptr_t>(function_));
     function_->PrintName();
-    PrintF(" => pc=0x%0x]\n", output_[0]->GetPc());
+    PrintF(" => pc=0x%08x]\n", output_[0]->GetPc());
   }
 }
 
@@ -825,19 +827,32 @@ void Deoptimizer::DoComputeJSFrame(TranslationIterator* iterator,
     ASSERT_EQ(Translation::kSelfLiteralId, closure_id);
     function = function_;
   }
+  Code* non_optimized_code = function->shared()->code();
   unsigned height = iterator->Next();
   unsigned height_in_bytes = height * kPointerSize;
-  if (FLAG_trace_deopt) {
-    PrintF("  translating ");
-    function->PrintName();
-    PrintF(" => node=%d, height=%d\n", node_id.ToInt(), height_in_bytes);
-  }
+  int handler_count = iterator->Next();
+  // If we are doing lazy deoptimization for catching in an optimized frame,
+  // one of the handlers in the frame description has already been dropped
+  // by throwing to it.
+  if (isolate_->optimized_handler_patch_buffer() != NULL) {
+    ASSERT(frame_index == 0 && output_count_ == 1);
+    ASSERT(bailout_type_ == LAZY);
+    ASSERT(handler_count > 0);
+    --handler_count;
+  }
+  int handlers_size = handler_count * StackHandlerConstants::kSize;
 
   // The 'fixed' part of the frame consists of the incoming parameters and
   // the part described by JavaScriptFrameConstants.
   unsigned fixed_frame_size = ComputeFixedSize(function);
-  unsigned input_frame_size = input_->GetFrameSize();
-  unsigned output_frame_size = height_in_bytes + fixed_frame_size;
+  unsigned varying_frame_size = height_in_bytes + handlers_size;
+  unsigned output_frame_size = fixed_frame_size + varying_frame_size;
+  if (FLAG_trace_deopt) {
+    PrintF("  translating ");
+    function->PrintName();
+    PrintF(" => node=%d, height=%d\n",
+           node_id.ToInt(), height_in_bytes + handlers_size);
+  }
 
   // Allocate and store the output frame description.
   FrameDescription* output_frame =
@@ -845,6 +860,7 @@ void Deoptimizer::DoComputeJSFrame(TranslationIterator* iterator,
   output_frame->SetFrameType(StackFrame::JAVA_SCRIPT);
 
   bool is_bottommost = (0 == frame_index);
+  ASSERT(is_bottommost || handler_count == 0);  // We do not inline try-catch.
   bool is_topmost = (output_count_ - 1 == frame_index);
   ASSERT(frame_index >= 0 && frame_index < output_count_);
   ASSERT(output_[frame_index] == NULL);
@@ -853,7 +869,7 @@ void Deoptimizer::DoComputeJSFrame(TranslationIterator* iterator,
   // 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 input_offset = input_->GetFrameSize();
 
   unsigned alignment_state_offset =
     input_offset - parameter_count * kPointerSize -
@@ -876,7 +892,7 @@ void Deoptimizer::DoComputeJSFrame(TranslationIterator* iterator,
     // 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;
+        varying_frame_size + has_alignment_padding_ * kPointerSize;
   } else {
     top_address = output_[frame_index - 1]->GetTop() - output_frame_size;
   }
@@ -966,35 +982,141 @@ void Deoptimizer::DoComputeJSFrame(TranslationIterator* iterator,
            top_address + output_offset, output_offset, value);
   }
 
+
+  // Translate the stack-allocated locals.
+  int stack_slot_count = function->shared()->scope_info()->StackSlotCount();
+  for (int i = 0; i < stack_slot_count; ++i) {
+    output_offset -= kPointerSize;
+    DoTranslateCommand(iterator, frame_index, output_offset);
+  }
+
+  // If there are any exception handlers, translate them.
+  if (handler_count > 0) {
+    // Translate the handler blocks.  The output frame needs:
+    //
+    // incoming args | fixed part | locals | handlers | expression stack
+    //
+    // because try occurs as a statement, i.e., with the expression stack
+    // empty.  The input frame has:
+    //
+    // incoming args | fixed part | spill slots | handlers | outgoing args
+    //
+    // also because try is a statement, i.e., there is no pending call to
+    // interleave handlers and args.  This would change if we began inlining
+    // functions containing try/catch.
+    //
+    // TODO(kmillikin): Begin inlining functions containing try/catch.
+    input_offset = ComputeOutgoingArgumentSize() + ComputeHandlersSize();
+    intptr_t next_handler = 0;
+    for (int i = 0; i < handler_count; ++i) {
+      // The first two fields (fp and context) are invariant under
+      // deoptimization.
+      output_offset -= kPointerSize;
+      input_offset -= kPointerSize;
+      value = input_->GetFrameSlot(input_offset);
+      ASSERT(value == fp_value);
+      output_frame->SetFrameSlot(output_offset, value);
+      if (FLAG_trace_deopt) {
+        PrintF("    0x%08x: [top + %d] <- 0x%08x ; handler[%d] fp\n",
+               top_address + output_offset, output_offset, value, i);
+      }
+
+      output_offset -= kPointerSize;
+      input_offset -= kPointerSize;
+      value = input_->GetFrameSlot(input_offset);
+      output_frame->SetFrameSlot(output_offset, value);
+      if (FLAG_trace_deopt) {
+        PrintF("    0x%08x: [top + %d] <- 0x%08x ; handler[%d] context\n",
+               top_address + output_offset, output_offset, value, i);
+      }
+
+      // The state is the same except that the kind is unoptimized.
+      output_offset -= kPointerSize;
+      input_offset -= kPointerSize;
+      value =
+          input_->GetFrameSlot(input_offset) & ~StackHandler::kIsOptimizedMask;
+      output_frame->SetFrameSlot(output_offset, value);
+      if (FLAG_trace_deopt) {
+        PrintF("    0x%08x: [top + %d] <- 0x%08x ; handler[%d] index + kind\n",
+               top_address + output_offset, output_offset, value, i);
+      }
+
+      // For all handlers in the frame, the code is the unoptimized code.
+      output_offset -= kPointerSize;
+      input_offset -= kPointerSize;
+      value = reinterpret_cast<uint32_t>(non_optimized_code);
+      output_frame->SetFrameSlot(output_offset, value);
+      if (FLAG_trace_deopt) {
+          PrintF("    0x%08x: [top + %d] <- 0x%08x ; handler[%d] code\n",
+                 top_address + output_offset, output_offset, value, i);
+      }
+
+      output_offset -= kPointerSize;
+      input_offset -= kPointerSize;
+      if (i == 0) {
+        // Bottommost handler's next link is deoptimization invariant.
+        value = input_->GetFrameSlot(input_offset);
+      } else {
+        value = next_handler;
+      }
+      output_frame->SetFrameSlot(output_offset, value);
+      next_handler = top_address + output_offset;
+      if (FLAG_trace_deopt) {
+          PrintF("    0x%08x: [top + %d] <- 0x%08x ; handler[%d] next\n",
+                 top_address + output_offset, output_offset, value, i);
+      }
+    }
+    *reinterpret_cast<intptr_t*>(isolate_->handler_address()) = next_handler;
+  }
+
   // Translate the rest of the frame.
-  for (unsigned i = 0; i < height; ++i) {
+  while (output_offset > 0) {
     output_offset -= kPointerSize;
     DoTranslateCommand(iterator, frame_index, output_offset);
   }
-  ASSERT(0 == output_offset);
 
   // Compute this frame's PC, state, and continuation.
-  Code* non_optimized_code = function->shared()->code();
-  FixedArray* raw_data = non_optimized_code->deoptimization_data();
-  DeoptimizationOutputData* data = DeoptimizationOutputData::cast(raw_data);
-  Address start = non_optimized_code->instruction_start();
-  unsigned pc_and_state = GetOutputInfo(data, node_id, function->shared());
-  unsigned pc_offset = FullCodeGenerator::PcField::decode(pc_and_state);
-  uint32_t pc_value = reinterpret_cast<uint32_t>(start + pc_offset);
-  output_frame->SetPc(pc_value);
-
-  FullCodeGenerator::State state =
-      FullCodeGenerator::StateField::decode(pc_and_state);
+  int pc_offset;
+  FullCodeGenerator::State state;
+  Builtins::Name continuation;
+  if (isolate_->optimized_handler_patch_buffer() != NULL) {
+    ASSERT(frame_index == 0 && output_count_ == 1);
+    // Catching in optimized code.  Deopt to the corresponding unoptimized
+    // catch handler (not to the instruction following the one that threw).
+    FixedArray* handler_table = non_optimized_code->handler_table();
+    int handler_index = isolate_->optimized_handler_handler_index();
+    pc_offset = Smi::cast(handler_table->get(handler_index))->value();
+    state = FullCodeGenerator::NO_REGISTERS;
+    continuation = Builtins::kNotifyLazyDeoptimized;
+    // The exception value is needed in eax.
+    output_frame->SetRegister(eax.code(),
+        reinterpret_cast<intptr_t>(isolate_->optimized_pending_exception()));
+    isolate_->clear_optimized_pending_exception();
+  } else {
+    FixedArray* raw_data = non_optimized_code->deoptimization_data();
+    DeoptimizationOutputData* data = DeoptimizationOutputData::cast(raw_data);
+    unsigned pc_and_state = GetOutputInfo(data, node_id, function->shared());
+    pc_offset = FullCodeGenerator::PcField::decode(pc_and_state);
+    state = FullCodeGenerator::StateField::decode(pc_and_state);
+    continuation = (bailout_type_ == EAGER)
+        ? Builtins::kNotifyDeoptimized
+        : Builtins::kNotifyLazyDeoptimized;
+    // For lazy deopt we preserve eax in the stack.  Ensure that it is not
+    // the zap value (a probably-invalid HeapObject*).
+    // TODO(mmassi): According to Kevin the test should be "!= EAGER"...
+    if (bailout_type_ == EAGER && is_topmost) {
+      output_frame->SetRegister(eax.code(),
+                                reinterpret_cast<intptr_t>(Smi::FromInt(0)));
+    }
+  }
+  byte* code_start = non_optimized_code->instruction_start();
+  output_frame->SetPc(reinterpret_cast<uint32_t>(code_start + pc_offset));
   output_frame->SetState(Smi::FromInt(state));
-
   // Set the continuation for the topmost frame.
   if (is_topmost && bailout_type_ != DEBUGGER) {
     Builtins* builtins = isolate_->builtins();
-    Code* continuation = (bailout_type_ == EAGER)
-        ? builtins->builtin(Builtins::kNotifyDeoptimized)
-        : builtins->builtin(Builtins::kNotifyLazyDeoptimized);
-    output_frame->SetContinuation(
-        reinterpret_cast<uint32_t>(continuation->entry()));
+    byte* entry = builtins->builtin(continuation)->entry();
+    output_frame->SetContinuation(reinterpret_cast<uint32_t>(entry));
   }
 }