path: root/deps/v8/src/arm/deoptimizer-arm.cc

// Copyright 2012 the V8 project authors. All rights reserved.
// Use of this source code is governed by a BSD-style license that can be
// found in the LICENSE file.

#include "src/assembler-inl.h"
#include "src/codegen.h"
#include "src/deoptimizer.h"
#include "src/full-codegen/full-codegen.h"
#include "src/objects-inl.h"
#include "src/register-configuration.h"
#include "src/safepoint-table.h"

namespace v8 {
namespace internal {

const int Deoptimizer::table_entry_size_ = 8;


int Deoptimizer::patch_size() {
  const int kCallInstructionSizeInWords = 3;
  return kCallInstructionSizeInWords * Assembler::kInstrSize;
}


void Deoptimizer::EnsureRelocSpaceForLazyDeoptimization(Handle<Code> code) {
  // Empty because there is no need for relocation information for the code
  // patching in Deoptimizer::PatchCodeForDeoptimization below.
}


void Deoptimizer::PatchCodeForDeoptimization(Isolate* isolate, Code* code) {
  Address code_start_address = code->instruction_start();
  // Invalidate the relocation information, as it will become invalid by the
  // code patching below, and is not needed any more.
  code->InvalidateRelocation();

  // Fail hard and early if we enter this code object again.
  byte* pointer = code->FindCodeAgeSequence();
  if (pointer != NULL) {
    pointer += kNoCodeAgeSequenceLength;
  } else {
    pointer = code->instruction_start();
  }

  {
    PatchingAssembler patcher(Assembler::IsolateData(isolate), pointer, 1);
    patcher.bkpt(0);
    patcher.FlushICache(isolate);
  }

  DeoptimizationInputData* data =
      DeoptimizationInputData::cast(code->deoptimization_data());
  int osr_offset = data->OsrPcOffset()->value();
  if (osr_offset > 0) {
    PatchingAssembler patcher(Assembler::IsolateData(isolate),
                              code_start_address + osr_offset, 1);
    patcher.bkpt(0);
    patcher.FlushICache(isolate);
  }

  DeoptimizationInputData* deopt_data =
      DeoptimizationInputData::cast(code->deoptimization_data());
#ifdef DEBUG
  Address prev_call_address = NULL;
#endif
  // For each LLazyBailout instruction insert a call to the corresponding
  // deoptimization entry.
  for (int i = 0; i < deopt_data->DeoptCount(); i++) {
    if (deopt_data->Pc(i)->value() == -1) continue;
    Address call_address = code_start_address + deopt_data->Pc(i)->value();
    Address deopt_entry = GetDeoptimizationEntry(isolate, i, LAZY);
    // We need calls to have a predictable size in the unoptimized code, but
    // this is optimized code, so we don't have to have a predictable size.
    int call_size_in_bytes = MacroAssembler::CallDeoptimizerSize();
    int call_size_in_words = call_size_in_bytes / Assembler::kInstrSize;
    DCHECK(call_size_in_bytes % Assembler::kInstrSize == 0);
    DCHECK(call_size_in_bytes <= patch_size());
    CodePatcher patcher(isolate, call_address, call_size_in_words);
    patcher.masm()->CallDeoptimizer(deopt_entry);
    DCHECK(prev_call_address == NULL ||
           call_address >= prev_call_address + patch_size());
    DCHECK(call_address + patch_size() <= code->instruction_end());
#ifdef DEBUG
    prev_call_address = call_address;
#endif
  }
}


void Deoptimizer::SetPlatformCompiledStubRegisters(
    FrameDescription* output_frame, CodeStubDescriptor* descriptor) {
  ApiFunction function(descriptor->deoptimization_handler());
  ExternalReference xref(&function, ExternalReference::BUILTIN_CALL, isolate_);
  intptr_t handler = reinterpret_cast<intptr_t>(xref.address());
  int params = descriptor->GetHandlerParameterCount();
  output_frame->SetRegister(r0.code(), params);
  output_frame->SetRegister(r1.code(), handler);
}


void Deoptimizer::CopyDoubleRegisters(FrameDescription* output_frame) {
  for (int i = 0; i < DwVfpRegister::kMaxNumRegisters; ++i) {
    Float64 double_value = input_->GetDoubleRegister(i);
    output_frame->SetDoubleRegister(i, double_value);
  }
}

#define __ masm()->

// This code tries to be close to ia32 code so that any changes can be
// easily ported.
void Deoptimizer::TableEntryGenerator::Generate() {
  GeneratePrologue();

  // Save all general purpose registers before messing with them.
  const int kNumberOfRegisters = Register::kNumRegisters;

  // Everything but pc, lr and ip which will be saved but not restored.
  RegList restored_regs = kJSCallerSaved | kCalleeSaved | ip.bit();

  const int kDoubleRegsSize = kDoubleSize * DwVfpRegister::kMaxNumRegisters;
  const int kFloatRegsSize = kFloatSize * SwVfpRegister::kMaxNumRegisters;

  // Save all allocatable VFP registers before messing with them.
  DCHECK(kDoubleRegZero.code() == 13);
  DCHECK(kScratchDoubleReg.code() == 14);

  {
    // We use a run-time check for VFP32DREGS.
    CpuFeatureScope scope(masm(), VFP32DREGS,
                          CpuFeatureScope::kDontCheckSupported);

    // Check CPU flags for number of registers, setting the Z condition flag.
    __ CheckFor32DRegs(ip);

    // Push registers d0-d15, and possibly d16-d31, on the stack.
    // If d16-d31 are not pushed, decrease the stack pointer instead.
    __ vstm(db_w, sp, d16, d31, ne);
    __ sub(sp, sp, Operand(16 * kDoubleSize), LeaveCC, eq);
    __ vstm(db_w, sp, d0, d15);

    // Push registers s0-s31 on the stack.
    __ vstm(db_w, sp, s0, s31);
  }

  // Push all 16 registers (needed to populate FrameDescription::registers_).
  // TODO(1588) Note that using pc with stm is deprecated, so we should perhaps
  // handle this a bit differently.
  __ stm(db_w, sp, restored_regs  | sp.bit() | lr.bit() | pc.bit());

  __ mov(ip, Operand(ExternalReference(Isolate::kCEntryFPAddress, isolate())));
  __ str(fp, MemOperand(ip));

  const int kSavedRegistersAreaSize =
      (kNumberOfRegisters * kPointerSize) + kDoubleRegsSize + kFloatRegsSize;

  // Get the bailout id from the stack.
  __ ldr(r2, MemOperand(sp, kSavedRegistersAreaSize));

  // Get the address of the location in the code object (r3) (return
  // address for lazy deoptimization) and compute the fp-to-sp delta in
  // register r4.
  __ mov(r3, lr);
  // Correct one word for bailout id.
  __ add(r4, sp, Operand(kSavedRegistersAreaSize + (1 * kPointerSize)));
  __ sub(r4, fp, r4);

  // Allocate a new deoptimizer object.
  // Pass four arguments in r0 to r3 and fifth argument on stack.
  __ PrepareCallCFunction(6, r5);
  __ mov(r0, Operand(0));
  Label context_check;
  __ ldr(r1, MemOperand(fp, CommonFrameConstants::kContextOrFrameTypeOffset));
  __ JumpIfSmi(r1, &context_check);
  __ ldr(r0, MemOperand(fp, JavaScriptFrameConstants::kFunctionOffset));
  __ bind(&context_check);
  __ mov(r1, Operand(type()));  // bailout type,
  // r2: bailout id already loaded.
  // r3: code address or 0 already loaded.
  __ str(r4, MemOperand(sp, 0 * kPointerSize));  // Fp-to-sp delta.
  __ mov(r5, Operand(ExternalReference::isolate_address(isolate())));
  __ str(r5, MemOperand(sp, 1 * kPointerSize));  // Isolate.
  // Call Deoptimizer::New().
  {
    AllowExternalCallThatCantCauseGC scope(masm());
    __ CallCFunction(ExternalReference::new_deoptimizer_function(isolate()), 6);
  }

  // Preserve "deoptimizer" object in register r0 and get the input
  // frame descriptor pointer to r1 (deoptimizer->input_);
  __ ldr(r1, MemOperand(r0, Deoptimizer::input_offset()));

  // Copy core registers into FrameDescription::registers_[kNumRegisters].
  DCHECK(Register::kNumRegisters == kNumberOfRegisters);
  for (int i = 0; i < kNumberOfRegisters; i++) {
    int offset = (i * kPointerSize) + FrameDescription::registers_offset();
    __ ldr(r2, MemOperand(sp, i * kPointerSize));
    __ str(r2, MemOperand(r1, offset));
  }

  // Copy VFP registers to
  // double_registers_[DoubleRegister::kMaxNumAllocatableRegisters]
  int double_regs_offset = FrameDescription::double_registers_offset();
  const RegisterConfiguration* config = RegisterConfiguration::Crankshaft();
  for (int i = 0; i < config->num_allocatable_double_registers(); ++i) {
    int code = config->GetAllocatableDoubleCode(i);
    int dst_offset = code * kDoubleSize + double_regs_offset;
    int src_offset =
        code * kDoubleSize + kNumberOfRegisters * kPointerSize + kFloatRegsSize;
    __ vldr(d0, sp, src_offset);
    __ vstr(d0, r1, dst_offset);
  }

  // Copy VFP registers to
  // float_registers_[FloatRegister::kMaxNumAllocatableRegisters]
  int float_regs_offset = FrameDescription::float_registers_offset();
  for (int i = 0; i < config->num_allocatable_float_registers(); ++i) {
    int code = config->GetAllocatableFloatCode(i);
    int dst_offset = code * kFloatSize + float_regs_offset;
    int src_offset = code * kFloatSize + kNumberOfRegisters * kPointerSize;
    __ ldr(r2, MemOperand(sp, src_offset));
    __ str(r2, MemOperand(r1, dst_offset));
  }

  // Remove the bailout id and the saved registers from the stack.
  __ add(sp, sp, Operand(kSavedRegistersAreaSize + (1 * kPointerSize)));

  // Compute a pointer to the unwinding limit in register r2; that is
  // the first stack slot not part of the input frame.
  __ ldr(r2, MemOperand(r1, FrameDescription::frame_size_offset()));
  __ add(r2, r2, sp);

  // Unwind the stack down to - but not including - the unwinding
  // limit and copy the contents of the activation frame to the input
  // frame description.
  __ add(r3,  r1, Operand(FrameDescription::frame_content_offset()));
  Label pop_loop;
  Label pop_loop_header;
  __ b(&pop_loop_header);
  __ bind(&pop_loop);
  __ pop(r4);
  __ str(r4, MemOperand(r3, 0));
  __ add(r3, r3, Operand(sizeof(uint32_t)));
  __ bind(&pop_loop_header);
  __ cmp(r2, sp);
  __ b(ne, &pop_loop);

  // Compute the output frame in the deoptimizer.
  __ push(r0);  // Preserve deoptimizer object across call.
  // r0: deoptimizer object; r1: scratch.
  __ PrepareCallCFunction(1, r1);
  // Call Deoptimizer::ComputeOutputFrames().
  {
    AllowExternalCallThatCantCauseGC scope(masm());
    __ CallCFunction(
        ExternalReference::compute_output_frames_function(isolate()), 1);
  }
  __ pop(r0);  // Restore deoptimizer object (class Deoptimizer).

  __ ldr(sp, MemOperand(r0, Deoptimizer::caller_frame_top_offset()));

  // Replace the current (input) frame with the output frames.
  Label outer_push_loop, inner_push_loop,
      outer_loop_header, inner_loop_header;
  // Outer loop state: r4 = current "FrameDescription** output_",
  // r1 = one past the last FrameDescription**.
  __ ldr(r1, MemOperand(r0, Deoptimizer::output_count_offset()));
  __ ldr(r4, MemOperand(r0, Deoptimizer::output_offset()));  // r4 is output_.
  __ add(r1, r4, Operand(r1, LSL, 2));
  __ jmp(&outer_loop_header);
  __ bind(&outer_push_loop);
  // Inner loop state: r2 = current FrameDescription*, r3 = loop index.
  __ ldr(r2, MemOperand(r4, 0));  // output_[ix]
  __ ldr(r3, MemOperand(r2, FrameDescription::frame_size_offset()));
  __ jmp(&inner_loop_header);
  __ bind(&inner_push_loop);
  __ sub(r3, r3, Operand(sizeof(uint32_t)));
  __ add(r6, r2, Operand(r3));
  __ ldr(r6, MemOperand(r6, FrameDescription::frame_content_offset()));
  __ push(r6);
  __ bind(&inner_loop_header);
  __ cmp(r3, Operand::Zero());
  __ b(ne, &inner_push_loop);  // test for gt?
  __ add(r4, r4, Operand(kPointerSize));
  __ bind(&outer_loop_header);
  __ cmp(r4, r1);
  __ b(lt, &outer_push_loop);

  __ ldr(r1, MemOperand(r0, Deoptimizer::input_offset()));
  for (int i = 0; i < config->num_allocatable_double_registers(); ++i) {
    int code = config->GetAllocatableDoubleCode(i);
    DwVfpRegister reg = DwVfpRegister::from_code(code);
    int src_offset = code * kDoubleSize + double_regs_offset;
    __ vldr(reg, r1, src_offset);
  }

  // Push state, pc, and continuation from the last output frame.
  __ ldr(r6, MemOperand(r2, FrameDescription::state_offset()));
  __ push(r6);
  __ ldr(r6, MemOperand(r2, FrameDescription::pc_offset()));
  __ push(r6);
  __ ldr(r6, MemOperand(r2, FrameDescription::continuation_offset()));
  __ push(r6);

  // Push the registers from the last output frame.
  for (int i = kNumberOfRegisters - 1; i >= 0; i--) {
    int offset = (i * kPointerSize) + FrameDescription::registers_offset();
    __ ldr(r6, MemOperand(r2, offset));
    __ push(r6);
  }

  // Restore the registers from the stack.
  __ ldm(ia_w, sp, restored_regs);  // all but pc registers.
  __ pop(ip);  // remove sp
  __ pop(ip);  // remove lr

  __ InitializeRootRegister();

  __ pop(ip);  // remove pc
  __ pop(ip);  // get continuation, leave pc on stack
  __ pop(lr);
  __ Jump(ip);
  __ stop("Unreachable.");
}


void Deoptimizer::TableEntryGenerator::GeneratePrologue() {
  // Create a sequence of deoptimization entries.
  // Note that registers are still live when jumping to an entry.

  // We need to be able to generate immediates up to kMaxNumberOfEntries. On
  // ARMv7, we can use movw (with a maximum immediate of 0xffff). On ARMv6, we
  // need two instructions.
  STATIC_ASSERT((kMaxNumberOfEntries - 1) <= 0xffff);
  if (CpuFeatures::IsSupported(ARMv7)) {
    CpuFeatureScope scope(masm(), ARMv7);
    Label done;
    for (int i = 0; i < count(); i++) {
      int start = masm()->pc_offset();
      USE(start);
      __ movw(ip, i);
      __ b(&done);
      DCHECK_EQ(table_entry_size_, masm()->pc_offset() - start);
    }
    __ bind(&done);
  } else {
    // We want to keep table_entry_size_ == 8 (since this is the common case),
    // but we need two instructions to load most immediates over 0xff. To handle
    // this, we set the low byte in the main table, and then set the high byte
    // in a separate table if necessary.
    Label high_fixes[256];
    int high_fix_max = (count() - 1) >> 8;
    DCHECK_GT(arraysize(high_fixes), static_cast<size_t>(high_fix_max));
    for (int i = 0; i < count(); i++) {
      int start = masm()->pc_offset();
      USE(start);
      __ mov(ip, Operand(i & 0xff));  // Set the low byte.
      __ b(&high_fixes[i >> 8]);      // Jump to the secondary table.
      DCHECK_EQ(table_entry_size_, masm()->pc_offset() - start);
    }
    // Generate the secondary table, to set the high byte.
    for (int high = 1; high <= high_fix_max; high++) {
      __ bind(&high_fixes[high]);
      __ orr(ip, ip, Operand(high << 8));
      // If this isn't the last entry, emit a branch to the end of the table.
      // The last entry can just fall through.
      if (high < high_fix_max) __ b(&high_fixes[0]);
    }
    // Bind high_fixes[0] last, for indices like 0x00**. This case requires no
    // fix-up, so for (common) small tables we can jump here, then just fall
    // through with no additional branch.
    __ bind(&high_fixes[0]);
  }
  __ push(ip);
}


void FrameDescription::SetCallerPc(unsigned offset, intptr_t value) {
  SetFrameSlot(offset, value);
}


void FrameDescription::SetCallerFp(unsigned offset, intptr_t value) {
  SetFrameSlot(offset, value);
}


void FrameDescription::SetCallerConstantPool(unsigned offset, intptr_t value) {
  // No embedded constant pool support.
  UNREACHABLE();
}


#undef __

}  // namespace internal
}  // namespace v8