Version 3.6.5

src/ia32/deoptimizer-ia32.cc

 // Copyright 2011 the V8 project authors. All rights reserved.
 // Redistribution and use in source and binary forms, with or without
 // modification, are permitted provided that the following conditions are
 // met:
 //
 //     * Redistributions of source code must retain the above copyright
 //       notice, this list of conditions and the following disclaimer.
 //     * Redistributions in binary form must reproduce the above
 //       copyright notice, this list of conditions and the following
 //       disclaimer in the documentation and/or other materials provided
@@ skipping 98 matching lines @@
         factory->NewByteArray(reloc_length + padding, TENURED);
     memcpy(new_reloc->GetDataStartAddress() + padding,
            code->relocation_info()->GetDataStartAddress(),
            reloc_length);
     // Create a relocation writer to write the comments in the padding
     // space. Use position 0 for everything to ensure short encoding.
     RelocInfoWriter reloc_info_writer(
         new_reloc->GetDataStartAddress() + padding, 0);
     intptr_t comment_string
         = reinterpret_cast<intptr_t>(RelocInfo::kFillerCommentString);
-    RelocInfo rinfo(0, RelocInfo::COMMENT, comment_string);
+    RelocInfo rinfo(0, RelocInfo::COMMENT, comment_string, NULL);
     for (int i = 0; i < additional_comments; ++i) {
 #ifdef DEBUG
       byte* pos_before = reloc_info_writer.pos();
 #endif
       reloc_info_writer.Write(&rinfo);
       ASSERT(RelocInfo::kMinRelocCommentSize ==
              pos_before - reloc_info_writer.pos());
     }
     // Replace relocation information on the code object.
     code->set_relocation_info(*new_reloc);
@@ skipping 37 matching lines @@
       // The gap code is needed to get to the state expected at the bailout.
       curr_address += safepoint_entry.gap_code_size();
 
       CodePatcher patcher(curr_address, patch_size());
       Address deopt_entry = GetDeoptimizationEntry(deoptimization_index, LAZY);
       patcher.masm()->call(deopt_entry, RelocInfo::NONE);
 
       // We use RUNTIME_ENTRY for deoptimization bailouts.
       RelocInfo rinfo(curr_address + 1,  // 1 after the call opcode.
                       RelocInfo::RUNTIME_ENTRY,
-                      reinterpret_cast<intptr_t>(deopt_entry));
+                      reinterpret_cast<intptr_t>(deopt_entry),
+                      NULL);
       reloc_info_writer.Write(&rinfo);
       ASSERT_GE(reloc_info_writer.pos(),
                 reloc_info->address() + ByteArray::kHeaderSize);
       curr_address += patch_size();
     }
     prev_address = curr_address;
   }
   ZapCodeRange(prev_address,
                code_start_address + code->safepoint_table_offset());
 
@@ skipping 10 matching lines @@
   ASSERT(junk_address <= reloc_end_address);
   isolate->heap()->CreateFillerObjectAt(junk_address,
                                         reloc_end_address - junk_address);
 
   // Add the deoptimizing code to the list.
   DeoptimizingCodeListNode* node = new DeoptimizingCodeListNode(code);
   DeoptimizerData* data = isolate->deoptimizer_data();
   node->set_next(data->deoptimizing_code_list_);
   data->deoptimizing_code_list_ = node;
 
+  // We might be in the middle of incremental marking with compaction.
+  // Tell collector to treat this code object in a special way and
+  // 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());
 
   if (FLAG_trace_deopt) {
     PrintF("[forced deoptimization: ");
     function->PrintName();
     PrintF(" / %x]\n", reinterpret_cast<uint32_t>(function));
 #ifdef DEBUG
     if (FLAG_print_code) {
       code->PrintLn();
     }
 #endif
   }
 }
 
 
-void Deoptimizer::PatchStackCheckCodeAt(Address pc_after,
+void Deoptimizer::PatchStackCheckCodeAt(Code* unoptimized_code,
+                                        Address pc_after,
                                         Code* check_code,
                                         Code* replacement_code) {
   Address call_target_address = pc_after - kIntSize;
   ASSERT(check_code->entry() ==
          Assembler::target_address_at(call_target_address));
   // The stack check code matches the pattern:
   //
   //     cmp esp, <limit>
   //     jae ok
   //     call <stack guard>
   //     test eax, <loop nesting depth>
   // ok: ...
   //
   // We will patch away the branch so the code is:
   //
   //     cmp esp, <limit>  ;; Not changed
   //     nop
   //     nop
   //     call <on-stack replacment>
   //     test eax, <loop nesting depth>
   // ok:
   ASSERT(*(call_target_address - 3) == 0x73 &&  // jae
          *(call_target_address - 2) == 0x07 &&  // offset
          *(call_target_address - 1) == 0xe8);   // call
   *(call_target_address - 3) = 0x90;  // nop
   *(call_target_address - 2) = 0x90;  // nop
   Assembler::set_target_address_at(call_target_address,
                                    replacement_code->entry());
+
+  RelocInfo rinfo(call_target_address,
+                  RelocInfo::CODE_TARGET,
+                  0,
+                  unoptimized_code);
+  unoptimized_code->GetHeap()->incremental_marking()->RecordWriteIntoCode(
+      unoptimized_code, &rinfo, replacement_code);
 }
 
 
 void Deoptimizer::RevertStackCheckCodeAt(Address pc_after,
                                          Code* check_code,
                                          Code* replacement_code) {
   Address call_target_address = pc_after - kIntSize;
   ASSERT(replacement_code->entry() ==
          Assembler::target_address_at(call_target_address));
   // Replace the nops from patching (Deoptimizer::PatchStackCheckCode) to
   // restore the conditional branch.
   ASSERT(*(call_target_address - 3) == 0x90 &&  // nop
          *(call_target_address - 2) == 0x90 &&  // nop
          *(call_target_address - 1) == 0xe8);   // call
   *(call_target_address - 3) = 0x73;  // jae
   *(call_target_address - 2) = 0x07;  // offset
   Assembler::set_target_address_at(call_target_address,
                                    check_code->entry());
+
+  check_code->GetHeap()->incremental_marking()->
+      RecordCodeTargetPatch(call_target_address, check_code);
 }
 
 
 static int LookupBailoutId(DeoptimizationInputData* data, unsigned ast_id) {
   ByteArray* translations = data->TranslationByteArray();
   int length = data->DeoptCount();
   for (int i = 0; i < length; i++) {
     if (static_cast<unsigned>(data->AstId(i)->value()) == ast_id) {
       TranslationIterator it(translations,  data->TranslationIndex(i)->value());
       int value = it.Next();
@@ skipping 127 matching lines @@
     ok = DoOsrTranslateCommand(&iterator, &input_offset);
   }
 
   // If translation of any command failed, continue using the input frame.
   if (!ok) {
     delete output_[0];
     output_[0] = input_;
     output_[0]->SetPc(reinterpret_cast<uint32_t>(from_));
   } else {
     // Setup the frame pointer and the context pointer.
-    output_[0]->SetRegister(ebp.code(), input_->GetRegister(ebp.code()));
+    // All OSR stack frames are dynamically aligned to an 8-byte boundary.
+    int frame_pointer = input_->GetRegister(ebp.code());
+    if ((frame_pointer & 0x4) == 0) {
+      // Return address at FP + 4 should be aligned, so FP mod 8 should be 4.
+      frame_pointer -= kPointerSize;
+      has_alignment_padding_ = 1;
+    }
+    output_[0]->SetRegister(ebp.code(), frame_pointer);
     output_[0]->SetRegister(esi.code(), input_->GetRegister(esi.code()));
 
     unsigned pc_offset = data->OsrPcOffset()->value();
     uint32_t pc = reinterpret_cast<uint32_t>(
         optimized_code_->entry() + pc_offset);
     output_[0]->SetPc(pc);
   }
   Code* continuation =
       function->GetIsolate()->builtins()->builtin(Builtins::kNotifyOSR);
   output_[0]->SetContinuation(
@@ skipping 44 matching lines @@
   ASSERT(frame_index >= 0 && frame_index < output_count_);
   ASSERT(output_[frame_index] == NULL);
   output_[frame_index] = output_frame;
 
   // 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) {
-    // 2 = context and function in the frame.
-    top_address =
-        input_->GetRegister(ebp.code()) - (2 * kPointerSize) - height_in_bytes;
+    // If the optimized frame had alignment padding, adjust the frame pointer
+    // 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) {
@@ skipping 30 matching lines @@
   // pointer.
   output_offset -= kPointerSize;
   input_offset -= kPointerSize;
   if (is_bottommost) {
     value = input_->GetFrameSlot(input_offset);
   } else {
     value = output_[frame_index - 1]->GetFp();
   }
   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);
   }
 
   // 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
   // so long as we don't inline functions that need local contexts.
@@ skipping 84 matching lines @@
   GeneratePrologue();
   CpuFeatures::Scope scope(SSE2);
 
   Isolate* isolate = masm()->isolate();
 
   // Save all general purpose registers before messing with them.
   const int kNumberOfRegisters = Register::kNumRegisters;
 
   const int kDoubleRegsSize = kDoubleSize *
                               XMMRegister::kNumAllocatableRegisters;
-  __ sub(Operand(esp), Immediate(kDoubleRegsSize));
+  __ sub(esp, Immediate(kDoubleRegsSize));
   for (int i = 0; i < XMMRegister::kNumAllocatableRegisters; ++i) {
     XMMRegister xmm_reg = XMMRegister::FromAllocationIndex(i);
     int offset = i * kDoubleSize;
     __ movdbl(Operand(esp, offset), xmm_reg);
   }
 
   __ pushad();
 
   const int kSavedRegistersAreaSize = kNumberOfRegisters * kPointerSize +
                                       kDoubleRegsSize;
 
   // Get the bailout id from the stack.
   __ mov(ebx, Operand(esp, kSavedRegistersAreaSize));
 
   // Get the address of the location in the code object if possible
   // and compute the fp-to-sp delta in register edx.
   if (type() == EAGER) {
     __ Set(ecx, Immediate(0));
     __ lea(edx, Operand(esp, kSavedRegistersAreaSize + 1 * kPointerSize));
   } else {
     __ mov(ecx, Operand(esp, kSavedRegistersAreaSize + 1 * kPointerSize));
     __ lea(edx, Operand(esp, kSavedRegistersAreaSize + 2 * kPointerSize));
   }
-  __ sub(edx, Operand(ebp));
+  __ sub(edx, ebp);
   __ neg(edx);
 
   // Allocate a new deoptimizer object.
   __ PrepareCallCFunction(6, eax);
   __ mov(eax, Operand(ebp, JavaScriptFrameConstants::kFunctionOffset));
   __ mov(Operand(esp, 0 * kPointerSize), eax);  // Function.
   __ mov(Operand(esp, 1 * kPointerSize), Immediate(type()));  // Bailout type.
   __ mov(Operand(esp, 2 * kPointerSize), ebx);  // Bailout id.
   __ mov(Operand(esp, 3 * kPointerSize), ecx);  // Code address or 0.
   __ mov(Operand(esp, 4 * kPointerSize), edx);  // Fp-to-sp delta.
   __ mov(Operand(esp, 5 * kPointerSize),
          Immediate(ExternalReference::isolate_address()));
-  __ CallCFunction(ExternalReference::new_deoptimizer_function(isolate), 6);
+  {
+    AllowExternalCallThatCantCauseGC scope(masm());
+    __ CallCFunction(ExternalReference::new_deoptimizer_function(isolate), 6);
+  }
 
   // Preserve deoptimizer object in register eax and get the input
   // frame descriptor pointer.
   __ mov(ebx, Operand(eax, Deoptimizer::input_offset()));
 
   // Fill in the input registers.
   for (int i = kNumberOfRegisters - 1; i >= 0; i--) {
     int offset = (i * kPointerSize) + FrameDescription::registers_offset();
     __ pop(Operand(ebx, offset));
   }
 
   // Fill in the double input registers.
   int double_regs_offset = FrameDescription::double_registers_offset();
   for (int i = 0; i < XMMRegister::kNumAllocatableRegisters; ++i) {
     int dst_offset = i * kDoubleSize + double_regs_offset;
     int src_offset = i * kDoubleSize;
     __ movdbl(xmm0, Operand(esp, src_offset));
     __ movdbl(Operand(ebx, dst_offset), xmm0);
   }
 
   // Remove the bailout id and the double registers from the stack.
   if (type() == EAGER) {
-    __ add(Operand(esp), Immediate(kDoubleRegsSize + kPointerSize));
+    __ add(esp, Immediate(kDoubleRegsSize + kPointerSize));
   } else {
-    __ add(Operand(esp), Immediate(kDoubleRegsSize + 2 * kPointerSize));
+    __ add(esp, Immediate(kDoubleRegsSize + 2 * kPointerSize));
   }
 
   // Compute a pointer to the unwinding limit in register ecx; that is
   // the first stack slot not part of the input frame.
   __ mov(ecx, Operand(ebx, FrameDescription::frame_size_offset()));
-  __ add(ecx, Operand(esp));
+  __ add(ecx, esp);
 
   // Unwind the stack down to - but not including - the unwinding
   // limit and copy the contents of the activation frame to the input
   // frame description.
   __ lea(edx, Operand(ebx, FrameDescription::frame_content_offset()));
   Label pop_loop;
   __ bind(&pop_loop);
   __ pop(Operand(edx, 0));
-  __ add(Operand(edx), Immediate(sizeof(uint32_t)));
-  __ cmp(ecx, Operand(esp));
+  __ add(edx, Immediate(sizeof(uint32_t)));
+  __ cmp(ecx, esp);
   __ j(not_equal, &pop_loop);
 
+  // If frame was dynamically aligned, pop padding.
+  Label sentinel, sentinel_done;
+  __ pop(ecx);
+  __ cmp(ecx, Operand(eax, Deoptimizer::frame_alignment_marker_offset()));
+  __ j(equal, &sentinel);
+  __ push(ecx);
+  __ jmp(&sentinel_done);
+  __ bind(&sentinel);
+  __ mov(Operand(eax, Deoptimizer::has_alignment_padding_offset()),
+         Immediate(1));
+  __ bind(&sentinel_done);
   // Compute the output frame in the deoptimizer.
   __ push(eax);
   __ PrepareCallCFunction(1, ebx);
   __ mov(Operand(esp, 0 * kPointerSize), eax);
-  __ CallCFunction(
-      ExternalReference::compute_output_frames_function(isolate), 1);
+  {
+    AllowExternalCallThatCantCauseGC scope(masm());
+    __ CallCFunction(
+        ExternalReference::compute_output_frames_function(isolate), 1);
+  }
   __ pop(eax);
 
+  if (type() == OSR) {
+    // If alignment padding is added, push the sentinel.
+    Label no_osr_padding;
+    __ cmp(Operand(eax, Deoptimizer::has_alignment_padding_offset()),
+           Immediate(0));
+    __ j(equal, &no_osr_padding, Label::kNear);
+    __ push(Operand(eax, Deoptimizer::frame_alignment_marker_offset()));
+    __ bind(&no_osr_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
   // last FrameDescription**.
   __ mov(edx, Operand(eax, Deoptimizer::output_count_offset()));
   __ mov(eax, Operand(eax, Deoptimizer::output_offset()));
   __ lea(edx, Operand(eax, edx, times_4, 0));
   __ bind(&outer_push_loop);
   // Inner loop state: ebx = current FrameDescription*, ecx = loop index.
   __ mov(ebx, Operand(eax, 0));
   __ mov(ecx, Operand(ebx, FrameDescription::frame_size_offset()));
   __ bind(&inner_push_loop);
-  __ sub(Operand(ecx), Immediate(sizeof(uint32_t)));
+  __ sub(ecx, Immediate(sizeof(uint32_t)));
   __ push(Operand(ebx, ecx, times_1, FrameDescription::frame_content_offset()));
-  __ test(ecx, Operand(ecx));
+  __ test(ecx, ecx);
   __ j(not_zero, &inner_push_loop);
-  __ add(Operand(eax), Immediate(kPointerSize));
-  __ cmp(eax, Operand(edx));
+  __ add(eax, Immediate(kPointerSize));
+  __ cmp(eax, edx);
   __ j(below, &outer_push_loop);
 
   // In case of OSR, we have to restore the XMM registers.
   if (type() == OSR) {
     for (int i = 0; i < XMMRegister::kNumAllocatableRegisters; ++i) {
       XMMRegister xmm_reg = XMMRegister::FromAllocationIndex(i);
       int src_offset = i * kDoubleSize + double_regs_offset;
       __ movdbl(xmm_reg, Operand(ebx, src_offset));
     }
   }
@@ skipping 32 matching lines @@
   }
   __ bind(&done);
 }
 
 #undef __
 
 
 } }  // namespace v8::internal
 
 #endif  // V8_TARGET_ARCH_IA32