X86 jump tables for PackedSwitch
Implement X86PackedSwitch using a jump table of offsets to blocks. The
X86PackedSwitch version just adds an input to address the constant area.
Change-Id: Id2752a1ee79222493040c6fd0e59aee9a544b76a
Bug: 21119474
Signed-off-by: Mark Mendell <mark.p.mendell@intel.com>
diff --git a/compiler/optimizing/code_generator_x86.cc b/compiler/optimizing/code_generator_x86.cc
index f8be21a..b60eebf 100644
--- a/compiler/optimizing/code_generator_x86.cc
+++ b/compiler/optimizing/code_generator_x86.cc
@@ -521,7 +521,8 @@
move_resolver_(graph->GetArena(), this),
isa_features_(isa_features),
method_patches_(graph->GetArena()->Adapter(kArenaAllocCodeGenerator)),
- relative_call_patches_(graph->GetArena()->Adapter(kArenaAllocCodeGenerator)) {
+ relative_call_patches_(graph->GetArena()->Adapter(kArenaAllocCodeGenerator)),
+ fixups_to_jump_tables_(graph->GetArena()->Adapter(kArenaAllocCodeGenerator)) {
// Use a fake return address register to mimic Quick.
AddAllocatedRegister(Location::RegisterLocation(kFakeReturnRegister));
}
@@ -5669,6 +5670,51 @@
}
}
+void LocationsBuilderX86::VisitX86PackedSwitch(HX86PackedSwitch* switch_instr) {
+ LocationSummary* locations =
+ new (GetGraph()->GetArena()) LocationSummary(switch_instr, LocationSummary::kNoCall);
+ locations->SetInAt(0, Location::RequiresRegister());
+
+ // Constant area pointer.
+ locations->SetInAt(1, Location::RequiresRegister());
+
+ // And the temporary we need.
+ locations->AddTemp(Location::RequiresRegister());
+}
+
+void InstructionCodeGeneratorX86::VisitX86PackedSwitch(HX86PackedSwitch* switch_instr) {
+ int32_t lower_bound = switch_instr->GetStartValue();
+ int32_t num_entries = switch_instr->GetNumEntries();
+ LocationSummary* locations = switch_instr->GetLocations();
+ Register value_reg = locations->InAt(0).AsRegister<Register>();
+ HBasicBlock* default_block = switch_instr->GetDefaultBlock();
+
+ // Optimizing has a jump area.
+ Register temp_reg = locations->GetTemp(0).AsRegister<Register>();
+ Register constant_area = locations->InAt(1).AsRegister<Register>();
+
+ // Remove the bias, if needed.
+ if (lower_bound != 0) {
+ __ leal(temp_reg, Address(value_reg, -lower_bound));
+ value_reg = temp_reg;
+ }
+
+ // Is the value in range?
+ DCHECK_GE(num_entries, 1);
+ __ cmpl(value_reg, Immediate(num_entries - 1));
+ __ j(kAbove, codegen_->GetLabelOf(default_block));
+
+ // We are in the range of the table.
+ // Load (target-constant_area) from the jump table, indexing by the value.
+ __ movl(temp_reg, codegen_->LiteralCaseTable(switch_instr, constant_area, value_reg));
+
+ // Compute the actual target address by adding in constant_area.
+ __ addl(temp_reg, constant_area);
+
+ // And jump.
+ __ jmp(temp_reg);
+}
+
void LocationsBuilderX86::VisitX86ComputeBaseMethodAddress(
HX86ComputeBaseMethodAddress* insn) {
LocationSummary* locations =
@@ -5752,28 +5798,18 @@
}
}
-void CodeGeneratorX86::Finalize(CodeAllocator* allocator) {
- // Generate the constant area if needed.
- X86Assembler* assembler = GetAssembler();
- if (!assembler->IsConstantAreaEmpty()) {
- // Align to 4 byte boundary to reduce cache misses, as the data is 4 and 8
- // byte values.
- assembler->Align(4, 0);
- constant_area_start_ = assembler->CodeSize();
- assembler->AddConstantArea();
- }
-
- // And finish up.
- CodeGenerator::Finalize(allocator);
-}
-
/**
* Class to handle late fixup of offsets into constant area.
*/
class RIPFixup : public AssemblerFixup, public ArenaObject<kArenaAllocCodeGenerator> {
public:
- RIPFixup(const CodeGeneratorX86& codegen, int offset)
- : codegen_(codegen), offset_into_constant_area_(offset) {}
+ RIPFixup(CodeGeneratorX86& codegen, size_t offset)
+ : codegen_(&codegen), offset_into_constant_area_(offset) {}
+
+ protected:
+ void SetOffset(size_t offset) { offset_into_constant_area_ = offset; }
+
+ CodeGeneratorX86* codegen_;
private:
void Process(const MemoryRegion& region, int pos) OVERRIDE {
@@ -5781,19 +5817,77 @@
// last 4 bytes of the instruction.
// The value to patch is the distance from the offset in the constant area
// from the address computed by the HX86ComputeBaseMethodAddress instruction.
- int32_t constant_offset = codegen_.ConstantAreaStart() + offset_into_constant_area_;
- int32_t relative_position = constant_offset - codegen_.GetMethodAddressOffset();;
+ int32_t constant_offset = codegen_->ConstantAreaStart() + offset_into_constant_area_;
+ int32_t relative_position = constant_offset - codegen_->GetMethodAddressOffset();;
// Patch in the right value.
region.StoreUnaligned<int32_t>(pos - 4, relative_position);
}
- const CodeGeneratorX86& codegen_;
-
// Location in constant area that the fixup refers to.
- int offset_into_constant_area_;
+ int32_t offset_into_constant_area_;
};
+/**
+ * Class to handle late fixup of offsets to a jump table that will be created in the
+ * constant area.
+ */
+class JumpTableRIPFixup : public RIPFixup {
+ public:
+ JumpTableRIPFixup(CodeGeneratorX86& codegen, HX86PackedSwitch* switch_instr)
+ : RIPFixup(codegen, static_cast<size_t>(-1)), switch_instr_(switch_instr) {}
+
+ void CreateJumpTable() {
+ X86Assembler* assembler = codegen_->GetAssembler();
+
+ // Ensure that the reference to the jump table has the correct offset.
+ const int32_t offset_in_constant_table = assembler->ConstantAreaSize();
+ SetOffset(offset_in_constant_table);
+
+ // The label values in the jump table are computed relative to the
+ // instruction addressing the constant area.
+ const int32_t relative_offset = codegen_->GetMethodAddressOffset();
+
+ // Populate the jump table with the correct values for the jump table.
+ int32_t num_entries = switch_instr_->GetNumEntries();
+ HBasicBlock* block = switch_instr_->GetBlock();
+ const ArenaVector<HBasicBlock*>& successors = block->GetSuccessors();
+ // The value that we want is the target offset - the position of the table.
+ for (int32_t i = 0; i < num_entries; i++) {
+ HBasicBlock* b = successors[i];
+ Label* l = codegen_->GetLabelOf(b);
+ DCHECK(l->IsBound());
+ int32_t offset_to_block = l->Position() - relative_offset;
+ assembler->AppendInt32(offset_to_block);
+ }
+ }
+
+ private:
+ const HX86PackedSwitch* switch_instr_;
+};
+
+void CodeGeneratorX86::Finalize(CodeAllocator* allocator) {
+ // Generate the constant area if needed.
+ X86Assembler* assembler = GetAssembler();
+ if (!assembler->IsConstantAreaEmpty() || !fixups_to_jump_tables_.empty()) {
+ // Align to 4 byte boundary to reduce cache misses, as the data is 4 and 8
+ // byte values.
+ assembler->Align(4, 0);
+ constant_area_start_ = assembler->CodeSize();
+
+ // Populate any jump tables.
+ for (auto jump_table : fixups_to_jump_tables_) {
+ jump_table->CreateJumpTable();
+ }
+
+ // And now add the constant area to the generated code.
+ assembler->AddConstantArea();
+ }
+
+ // And finish up.
+ CodeGenerator::Finalize(allocator);
+}
+
Address CodeGeneratorX86::LiteralDoubleAddress(double v, Register reg) {
AssemblerFixup* fixup = new (GetGraph()->GetArena()) RIPFixup(*this, __ AddDouble(v));
return Address(reg, kDummy32BitOffset, fixup);
@@ -5814,6 +5908,20 @@
return Address(reg, kDummy32BitOffset, fixup);
}
+Address CodeGeneratorX86::LiteralCaseTable(HX86PackedSwitch* switch_instr,
+ Register reg,
+ Register value) {
+ // Create a fixup to be used to create and address the jump table.
+ JumpTableRIPFixup* table_fixup =
+ new (GetGraph()->GetArena()) JumpTableRIPFixup(*this, switch_instr);
+
+ // We have to populate the jump tables.
+ fixups_to_jump_tables_.push_back(table_fixup);
+
+ // We want a scaled address, as we are extracting the correct offset from the table.
+ return Address(reg, value, TIMES_4, kDummy32BitOffset, table_fixup);
+}
+
/**
* Finds instructions that need the constant area base as an input.
*/
@@ -5864,6 +5972,21 @@
}
}
+ void VisitPackedSwitch(HPackedSwitch* switch_insn) OVERRIDE {
+ // We need to replace the HPackedSwitch with a HX86PackedSwitch in order to
+ // address the constant area.
+ InitializeConstantAreaPointer(switch_insn);
+ HGraph* graph = GetGraph();
+ HBasicBlock* block = switch_insn->GetBlock();
+ HX86PackedSwitch* x86_switch = new (graph->GetArena()) HX86PackedSwitch(
+ switch_insn->GetStartValue(),
+ switch_insn->GetNumEntries(),
+ switch_insn->InputAt(0),
+ base_,
+ switch_insn->GetDexPc());
+ block->ReplaceAndRemoveInstructionWith(switch_insn, x86_switch);
+ }
+
void InitializeConstantAreaPointer(HInstruction* user) {
// Ensure we only initialize the pointer once.
if (base_ != nullptr) {
diff --git a/compiler/optimizing/code_generator_x86.h b/compiler/optimizing/code_generator_x86.h
index ae2d84f..fdfc5ab 100644
--- a/compiler/optimizing/code_generator_x86.h
+++ b/compiler/optimizing/code_generator_x86.h
@@ -245,6 +245,8 @@
DISALLOW_COPY_AND_ASSIGN(InstructionCodeGeneratorX86);
};
+class JumpTableRIPFixup;
+
class CodeGeneratorX86 : public CodeGenerator {
public:
CodeGeneratorX86(HGraph* graph,
@@ -385,6 +387,8 @@
Address LiteralInt32Address(int32_t v, Register reg);
Address LiteralInt64Address(int64_t v, Register reg);
+ Address LiteralCaseTable(HX86PackedSwitch* switch_instr, Register reg, Register value);
+
void Finalize(CodeAllocator* allocator) OVERRIDE;
private:
@@ -405,6 +409,9 @@
// Used for fixups to the constant area.
int32_t constant_area_start_;
+ // Fixups for jump tables that need to be patched after the constant table is generated.
+ ArenaVector<JumpTableRIPFixup*> fixups_to_jump_tables_;
+
// If there is a HX86ComputeBaseMethodAddress instruction in the graph
// (which shall be the sole instruction of this kind), subtracting this offset
// from the value contained in the out register of this HX86ComputeBaseMethodAddress
diff --git a/compiler/optimizing/nodes.cc b/compiler/optimizing/nodes.cc
index 24a89bc..ed401b6 100644
--- a/compiler/optimizing/nodes.cc
+++ b/compiler/optimizing/nodes.cc
@@ -606,8 +606,23 @@
void HBasicBlock::ReplaceAndRemoveInstructionWith(HInstruction* initial,
HInstruction* replacement) {
DCHECK(initial->GetBlock() == this);
- InsertInstructionBefore(replacement, initial);
- initial->ReplaceWith(replacement);
+ if (initial->IsControlFlow()) {
+ // We can only replace a control flow instruction with another control flow instruction.
+ DCHECK(replacement->IsControlFlow());
+ DCHECK_EQ(replacement->GetId(), -1);
+ DCHECK_EQ(replacement->GetType(), Primitive::kPrimVoid);
+ DCHECK_EQ(initial->GetBlock(), this);
+ DCHECK_EQ(initial->GetType(), Primitive::kPrimVoid);
+ DCHECK(initial->GetUses().IsEmpty());
+ DCHECK(initial->GetEnvUses().IsEmpty());
+ replacement->SetBlock(this);
+ replacement->SetId(GetGraph()->GetNextInstructionId());
+ instructions_.InsertInstructionBefore(replacement, initial);
+ UpdateInputsUsers(replacement);
+ } else {
+ InsertInstructionBefore(replacement, initial);
+ initial->ReplaceWith(replacement);
+ }
RemoveInstruction(initial);
}
diff --git a/compiler/optimizing/nodes.h b/compiler/optimizing/nodes.h
index 82909c4..0d668e8 100644
--- a/compiler/optimizing/nodes.h
+++ b/compiler/optimizing/nodes.h
@@ -1074,7 +1074,8 @@
#define FOR_EACH_CONCRETE_INSTRUCTION_X86(M) \
M(X86ComputeBaseMethodAddress, Instruction) \
- M(X86LoadFromConstantTable, Instruction)
+ M(X86LoadFromConstantTable, Instruction) \
+ M(X86PackedSwitch, Instruction)
#define FOR_EACH_CONCRETE_INSTRUCTION_X86_64(M)
diff --git a/compiler/optimizing/nodes_x86.h b/compiler/optimizing/nodes_x86.h
index f7cc872..556217b 100644
--- a/compiler/optimizing/nodes_x86.h
+++ b/compiler/optimizing/nodes_x86.h
@@ -62,6 +62,45 @@
DISALLOW_COPY_AND_ASSIGN(HX86LoadFromConstantTable);
};
+// X86 version of HPackedSwitch that holds a pointer to the base method address.
+class HX86PackedSwitch : public HTemplateInstruction<2> {
+ public:
+ HX86PackedSwitch(int32_t start_value,
+ int32_t num_entries,
+ HInstruction* input,
+ HX86ComputeBaseMethodAddress* method_base,
+ uint32_t dex_pc)
+ : HTemplateInstruction(SideEffects::None(), dex_pc),
+ start_value_(start_value),
+ num_entries_(num_entries) {
+ SetRawInputAt(0, input);
+ SetRawInputAt(1, method_base);
+ }
+
+ bool IsControlFlow() const OVERRIDE { return true; }
+
+ int32_t GetStartValue() const { return start_value_; }
+
+ int32_t GetNumEntries() const { return num_entries_; }
+
+ HX86ComputeBaseMethodAddress* GetBaseMethodAddress() const {
+ return InputAt(1)->AsX86ComputeBaseMethodAddress();
+ }
+
+ HBasicBlock* GetDefaultBlock() const {
+ // Last entry is the default block.
+ return GetBlock()->GetSuccessors()[num_entries_];
+ }
+
+ DECLARE_INSTRUCTION(X86PackedSwitch);
+
+ private:
+ const int32_t start_value_;
+ const int32_t num_entries_;
+
+ DISALLOW_COPY_AND_ASSIGN(HX86PackedSwitch);
+};
+
} // namespace art
#endif // ART_COMPILER_OPTIMIZING_NODES_X86_H_
diff --git a/compiler/utils/x86/assembler_x86.cc b/compiler/utils/x86/assembler_x86.cc
index 04e815a..5347bf0 100644
--- a/compiler/utils/x86/assembler_x86.cc
+++ b/compiler/utils/x86/assembler_x86.cc
@@ -2369,44 +2369,48 @@
}
}
-int ConstantArea::AddInt32(int32_t v) {
- for (size_t i = 0, e = buffer_.size(); i < e; i++) {
- if (v == buffer_[i]) {
- return i * kEntrySize;
- }
- }
-
- // Didn't match anything.
- int result = buffer_.size() * kEntrySize;
+size_t ConstantArea::AppendInt32(int32_t v) {
+ size_t result = buffer_.size() * elem_size_;
buffer_.push_back(v);
return result;
}
-int ConstantArea::AddInt64(int64_t v) {
+size_t ConstantArea::AddInt32(int32_t v) {
+ for (size_t i = 0, e = buffer_.size(); i < e; i++) {
+ if (v == buffer_[i]) {
+ return i * elem_size_;
+ }
+ }
+
+ // Didn't match anything.
+ return AppendInt32(v);
+}
+
+size_t ConstantArea::AddInt64(int64_t v) {
int32_t v_low = Low32Bits(v);
int32_t v_high = High32Bits(v);
if (buffer_.size() > 1) {
// Ensure we don't pass the end of the buffer.
for (size_t i = 0, e = buffer_.size() - 1; i < e; i++) {
if (v_low == buffer_[i] && v_high == buffer_[i + 1]) {
- return i * kEntrySize;
+ return i * elem_size_;
}
}
}
// Didn't match anything.
- int result = buffer_.size() * kEntrySize;
+ size_t result = buffer_.size() * elem_size_;
buffer_.push_back(v_low);
buffer_.push_back(v_high);
return result;
}
-int ConstantArea::AddDouble(double v) {
+size_t ConstantArea::AddDouble(double v) {
// Treat the value as a 64-bit integer value.
return AddInt64(bit_cast<int64_t, double>(v));
}
-int ConstantArea::AddFloat(float v) {
+size_t ConstantArea::AddFloat(float v) {
// Treat the value as a 32-bit integer value.
return AddInt32(bit_cast<int32_t, float>(v));
}
diff --git a/compiler/utils/x86/assembler_x86.h b/compiler/utils/x86/assembler_x86.h
index 93ecdf5..b50fda9 100644
--- a/compiler/utils/x86/assembler_x86.h
+++ b/compiler/utils/x86/assembler_x86.h
@@ -166,21 +166,6 @@
Init(base_in, disp.Int32Value());
}
- void Init(Register base_in, int32_t disp) {
- if (disp == 0 && base_in != EBP) {
- SetModRM(0, base_in);
- if (base_in == ESP) SetSIB(TIMES_1, ESP, base_in);
- } else if (disp >= -128 && disp <= 127) {
- SetModRM(1, base_in);
- if (base_in == ESP) SetSIB(TIMES_1, ESP, base_in);
- SetDisp8(disp);
- } else {
- SetModRM(2, base_in);
- if (base_in == ESP) SetSIB(TIMES_1, ESP, base_in);
- SetDisp32(disp);
- }
- }
-
Address(Register index_in, ScaleFactor scale_in, int32_t disp) {
CHECK_NE(index_in, ESP); // Illegal addressing mode.
SetModRM(0, ESP);
@@ -189,19 +174,15 @@
}
Address(Register base_in, Register index_in, ScaleFactor scale_in, int32_t disp) {
- CHECK_NE(index_in, ESP); // Illegal addressing mode.
- if (disp == 0 && base_in != EBP) {
- SetModRM(0, ESP);
- SetSIB(scale_in, index_in, base_in);
- } else if (disp >= -128 && disp <= 127) {
- SetModRM(1, ESP);
- SetSIB(scale_in, index_in, base_in);
- SetDisp8(disp);
- } else {
- SetModRM(2, ESP);
- SetSIB(scale_in, index_in, base_in);
- SetDisp32(disp);
- }
+ Init(base_in, index_in, scale_in, disp);
+ }
+
+ Address(Register base_in,
+ Register index_in,
+ ScaleFactor scale_in,
+ int32_t disp, AssemblerFixup *fixup) {
+ Init(base_in, index_in, scale_in, disp);
+ SetFixup(fixup);
}
static Address Absolute(uintptr_t addr) {
@@ -217,6 +198,37 @@
private:
Address() {}
+
+ void Init(Register base_in, int32_t disp) {
+ if (disp == 0 && base_in != EBP) {
+ SetModRM(0, base_in);
+ if (base_in == ESP) SetSIB(TIMES_1, ESP, base_in);
+ } else if (disp >= -128 && disp <= 127) {
+ SetModRM(1, base_in);
+ if (base_in == ESP) SetSIB(TIMES_1, ESP, base_in);
+ SetDisp8(disp);
+ } else {
+ SetModRM(2, base_in);
+ if (base_in == ESP) SetSIB(TIMES_1, ESP, base_in);
+ SetDisp32(disp);
+ }
+ }
+
+ void Init(Register base_in, Register index_in, ScaleFactor scale_in, int32_t disp) {
+ CHECK_NE(index_in, ESP); // Illegal addressing mode.
+ if (disp == 0 && base_in != EBP) {
+ SetModRM(0, ESP);
+ SetSIB(scale_in, index_in, base_in);
+ } else if (disp >= -128 && disp <= 127) {
+ SetModRM(1, ESP);
+ SetSIB(scale_in, index_in, base_in);
+ SetDisp8(disp);
+ } else {
+ SetModRM(2, ESP);
+ SetSIB(scale_in, index_in, base_in);
+ SetDisp32(disp);
+ }
+ }
};
@@ -252,40 +264,39 @@
// Add a double to the constant area, returning the offset into
// the constant area where the literal resides.
- int AddDouble(double v);
+ size_t AddDouble(double v);
// Add a float to the constant area, returning the offset into
// the constant area where the literal resides.
- int AddFloat(float v);
+ size_t AddFloat(float v);
// Add an int32_t to the constant area, returning the offset into
// the constant area where the literal resides.
- int AddInt32(int32_t v);
+ size_t AddInt32(int32_t v);
+
+ // Add an int32_t to the end of the constant area, returning the offset into
+ // the constant area where the literal resides.
+ size_t AppendInt32(int32_t v);
// Add an int64_t to the constant area, returning the offset into
// the constant area where the literal resides.
- int AddInt64(int64_t v);
+ size_t AddInt64(int64_t v);
bool IsEmpty() const {
return buffer_.size() == 0;
}
+ size_t GetSize() const {
+ return buffer_.size() * elem_size_;
+ }
+
const std::vector<int32_t>& GetBuffer() const {
return buffer_;
}
- void AddFixup(AssemblerFixup* fixup) {
- fixups_.push_back(fixup);
- }
-
- const std::vector<AssemblerFixup*>& GetFixups() const {
- return fixups_;
- }
-
private:
- static constexpr size_t kEntrySize = sizeof(int32_t);
+ static constexpr size_t elem_size_ = sizeof(int32_t);
std::vector<int32_t> buffer_;
- std::vector<AssemblerFixup*> fixups_;
};
class X86Assembler FINAL : public Assembler {
@@ -740,26 +751,36 @@
// Add a double to the constant area, returning the offset into
// the constant area where the literal resides.
- int AddDouble(double v) { return constant_area_.AddDouble(v); }
+ size_t AddDouble(double v) { return constant_area_.AddDouble(v); }
// Add a float to the constant area, returning the offset into
// the constant area where the literal resides.
- int AddFloat(float v) { return constant_area_.AddFloat(v); }
+ size_t AddFloat(float v) { return constant_area_.AddFloat(v); }
// Add an int32_t to the constant area, returning the offset into
// the constant area where the literal resides.
- int AddInt32(int32_t v) { return constant_area_.AddInt32(v); }
+ size_t AddInt32(int32_t v) {
+ return constant_area_.AddInt32(v);
+ }
+
+ // Add an int32_t to the end of the constant area, returning the offset into
+ // the constant area where the literal resides.
+ size_t AppendInt32(int32_t v) {
+ return constant_area_.AppendInt32(v);
+ }
// Add an int64_t to the constant area, returning the offset into
// the constant area where the literal resides.
- int AddInt64(int64_t v) { return constant_area_.AddInt64(v); }
+ size_t AddInt64(int64_t v) { return constant_area_.AddInt64(v); }
// Add the contents of the constant area to the assembler buffer.
void AddConstantArea();
// Is the constant area empty? Return true if there are no literals in the constant area.
bool IsConstantAreaEmpty() const { return constant_area_.IsEmpty(); }
- void AddConstantAreaFixup(AssemblerFixup* fixup) { constant_area_.AddFixup(fixup); }
+
+ // Return the current size of the constant area.
+ size_t ConstantAreaSize() const { return constant_area_.GetSize(); }
private:
inline void EmitUint8(uint8_t value);