| /* |
| * Copyright (C) 2014 The Android Open Source Project |
| * |
| * Licensed under the Apache License, Version 2.0 (the "License"); |
| * you may not use this file except in compliance with the License. |
| * You may obtain a copy of the License at |
| * |
| * http://www.apache.org/licenses/LICENSE-2.0 |
| * |
| * Unless required by applicable law or agreed to in writing, software |
| * distributed under the License is distributed on an "AS IS" BASIS, |
| * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| * See the License for the specific language governing permissions and |
| * limitations under the License. |
| */ |
| |
| #include "register_allocator.h" |
| |
| #include "arch/x86/instruction_set_features_x86.h" |
| #include "base/arena_allocator.h" |
| #include "base/macros.h" |
| #include "builder.h" |
| #include "code_generator.h" |
| #include "com_android_art_flags.h" |
| #include "dex/dex_file.h" |
| #include "dex/dex_file_types.h" |
| #include "dex/dex_instruction.h" |
| #include "driver/compiler_options.h" |
| #include "nodes.h" |
| #include "optimizing_unit_test.h" |
| #include "register_allocator_linear_scan.h" |
| #include "ssa_liveness_analysis.h" |
| #include "ssa_phi_elimination.h" |
| |
| namespace art HIDDEN { |
| |
| // Note: the register allocator tests rely on the fact that constants have live |
| // intervals and registers get allocated to them. |
| |
| class RegisterAllocatorTest : public CommonCompilerTest, public OptimizingUnitTestHelper { |
| protected: |
| // Define a shortcut for the `kLivenessPositionsPerInstruction`. |
| static constexpr size_t kLppi = kLivenessPositionsPerInstruction; |
| |
| std::unique_ptr<CodeGenerator> CreateCodeGenerator( |
| HGraph* graph, InstructionSet instruction_set = InstructionSet::kNone) { |
| if (instruction_set == InstructionSet::kNone) { |
| instruction_set = kRuntimeISA; |
| } |
| if (instruction_set == InstructionSet::kArm) { |
| instruction_set = InstructionSet::kThumb2; |
| } |
| compiler_options_ = CommonCompilerTest::CreateCompilerOptions(instruction_set, "default"); |
| CHECK(compiler_options_ != nullptr); |
| CHECK_EQ(instruction_set, compiler_options_->GetInstructionSet()); |
| std::unique_ptr<CodeGenerator> codegen = CodeGenerator::Create(graph, *compiler_options_); |
| CHECK_IMPLIES(codegen != nullptr, codegen->GetInstructionSet() == instruction_set); |
| return codegen; |
| } |
| |
| // Helper functions that make use of the OptimizingUnitTest's members. |
| bool Check(const std::vector<uint16_t>& data); |
| void BuildIfElseWithPhi(HPhi** phi, HInstruction** input1, HInstruction** input2); |
| void BuildFieldReturn(HInstruction** field, HInstruction** ret); |
| void BuildTwoSubs(HInstruction** first_sub, HInstruction** second_sub); |
| void BuildDiv(HInstruction** div); |
| |
| bool ValidateIntervals(const ScopedArenaVector<LiveInterval*>& intervals, |
| const CodeGenerator& codegen) { |
| return RegisterAllocator::ValidateIntervals(ArrayRef<LiveInterval* const>(intervals), |
| /* number_of_spill_slots= */ 0u, |
| /* number_of_out_slots= */ 0u, |
| codegen, |
| /*liveness=*/ nullptr, |
| PhysicalRegisterType::kCoreRegister, |
| /* log_fatal_on_failure= */ false); |
| } |
| |
| void OverrideOutput(LocationSummary* locations, |
| Location out, |
| Location::OutputOverlap output_overlaps = Location::kOutputOverlap) { |
| locations->output_ = out; |
| locations->output_overlaps_ = output_overlaps; |
| } |
| |
| template <typename RegType> |
| void BlockCoreRegistersExcept(CodeGenerator* codegen, std::initializer_list<RegType> allowed) { |
| size_t number_of_core_registers = codegen->GetNumberOfCoreRegisters(); |
| uint32_t blocked_core_registers = MaxInt<uint32_t>(number_of_core_registers); |
| for (RegType reg : allowed) { |
| CHECK_LT(reg, number_of_core_registers); |
| blocked_core_registers &= ~(1u << reg); |
| } |
| RegisterSet blocked_registers = RegisterSet::Empty(); |
| blocked_registers.AddCoreRegisterSet(blocked_core_registers); |
| blocked_registers.AddFpuRegisterSet(codegen->blocked_registers_.GetFpuRegisterSet()); |
| blocked_registers.AddVecRegisterSet(codegen->blocked_registers_.GetVecRegisterSet()); |
| codegen->blocked_registers_ = blocked_registers; |
| } |
| |
| void TestFreeUntil(bool special_first); |
| void TestSpillInactive(); |
| void TestNoOutputOverlap(); |
| void TestNoOutputOverlapAndTemp(); |
| |
| std::unique_ptr<CompilerOptions> compiler_options_; |
| }; |
| |
| bool RegisterAllocatorTest::Check(const std::vector<uint16_t>& data) { |
| HGraph* graph = CreateCFG(data); |
| std::unique_ptr<CodeGenerator> codegen = CreateCodeGenerator(graph); |
| CHECK(codegen != nullptr); |
| SsaLivenessAnalysis liveness(graph, codegen.get(), GetScopedAllocator()); |
| liveness.Analyze(); |
| std::unique_ptr<RegisterAllocator> register_allocator = |
| RegisterAllocator::Create(GetScopedAllocator(), codegen.get(), liveness); |
| register_allocator->AllocateRegisters(); |
| return register_allocator->Validate(false); |
| } |
| |
| /** |
| * Unit testing of RegisterAllocator::ValidateIntervals. Register allocator |
| * tests are based on this validation method. |
| */ |
| TEST_F(RegisterAllocatorTest, ValidateIntervals) { |
| CreateGraph(); |
| std::unique_ptr<CodeGenerator> codegen = CreateCodeGenerator(graph_, InstructionSet::kX86); |
| if (codegen == nullptr) { |
| GTEST_SKIP() << "X86 codegen is unavailable."; |
| } |
| ScopedArenaVector<LiveInterval*> intervals(GetScopedAllocator()->Adapter()); |
| |
| // Test with two intervals of the same range. |
| { |
| static constexpr size_t ranges[][2] = {{0, 42}}; |
| intervals.push_back(BuildInterval(ranges, arraysize(ranges), GetScopedAllocator(), 1u << 0)); |
| intervals.push_back(BuildInterval(ranges, arraysize(ranges), GetScopedAllocator(), 1u << 1)); |
| ASSERT_TRUE(ValidateIntervals(intervals, *codegen)); |
| |
| intervals[1]->SetRegisters(1u << 0); |
| ASSERT_FALSE(ValidateIntervals(intervals, *codegen)); |
| intervals.clear(); |
| } |
| |
| // Test with two non-intersecting intervals. |
| { |
| static constexpr size_t ranges1[][2] = {{0, 42}}; |
| intervals.push_back(BuildInterval(ranges1, arraysize(ranges1), GetScopedAllocator(), 1u << 0)); |
| static constexpr size_t ranges2[][2] = {{42, 43}}; |
| intervals.push_back(BuildInterval(ranges2, arraysize(ranges2), GetScopedAllocator(), 1u << 1)); |
| ASSERT_TRUE(ValidateIntervals(intervals, *codegen)); |
| |
| intervals[1]->SetRegisters(1u << 0); |
| ASSERT_TRUE(ValidateIntervals(intervals, *codegen)); |
| intervals.clear(); |
| } |
| |
| // Test with two non-intersecting intervals, with one with a lifetime hole. |
| { |
| static constexpr size_t ranges1[][2] = {{0, 42}, {45, 48}}; |
| intervals.push_back(BuildInterval(ranges1, arraysize(ranges1), GetScopedAllocator(), 1u << 0)); |
| static constexpr size_t ranges2[][2] = {{42, 43}}; |
| intervals.push_back(BuildInterval(ranges2, arraysize(ranges2), GetScopedAllocator(), 1u << 1)); |
| ASSERT_TRUE(ValidateIntervals(intervals, *codegen)); |
| |
| intervals[1]->SetRegisters(1u << 0); |
| ASSERT_TRUE(ValidateIntervals(intervals, *codegen)); |
| intervals.clear(); |
| } |
| |
| // Test with intersecting intervals. |
| { |
| static constexpr size_t ranges1[][2] = {{0, 42}, {44, 48}}; |
| intervals.push_back(BuildInterval(ranges1, arraysize(ranges1), GetScopedAllocator(), 1u << 0)); |
| static constexpr size_t ranges2[][2] = {{42, 47}}; |
| intervals.push_back(BuildInterval(ranges2, arraysize(ranges2), GetScopedAllocator(), 1u << 1)); |
| ASSERT_TRUE(ValidateIntervals(intervals, *codegen)); |
| |
| intervals[1]->SetRegisters(1u << 0); |
| ASSERT_FALSE(ValidateIntervals(intervals, *codegen)); |
| intervals.clear(); |
| } |
| |
| // Test with siblings. |
| { |
| static constexpr size_t ranges1[][2] = {{0, 42}, {44, 48}}; |
| intervals.push_back(BuildInterval(ranges1, arraysize(ranges1), GetScopedAllocator(), 1u << 0)); |
| intervals[0]->SplitAt(43); |
| static constexpr size_t ranges2[][2] = {{42, 47}}; |
| intervals.push_back(BuildInterval(ranges2, arraysize(ranges2), GetScopedAllocator(), 1u << 1)); |
| ASSERT_TRUE(ValidateIntervals(intervals, *codegen)); |
| |
| intervals[1]->SetRegisters(1u << 0); |
| // Sibling of the first interval has no register allocated to it. |
| ASSERT_TRUE(ValidateIntervals(intervals, *codegen)); |
| |
| intervals[0]->GetNextSibling()->SetRegisters(1u << 0); |
| ASSERT_FALSE(ValidateIntervals(intervals, *codegen)); |
| } |
| } |
| |
| TEST_F(RegisterAllocatorTest, CFG1) { |
| /* |
| * Test the following snippet: |
| * return 0; |
| * |
| * Which becomes the following graph: |
| * constant0 |
| * goto |
| * | |
| * return |
| * | |
| * exit |
| */ |
| const std::vector<uint16_t> data = ONE_REGISTER_CODE_ITEM( |
| Instruction::CONST_4 | 0 | 0, |
| Instruction::RETURN); |
| |
| ASSERT_TRUE(Check(data)); |
| } |
| |
| TEST_F(RegisterAllocatorTest, Loop1) { |
| /* |
| * Test the following snippet: |
| * int a = 0; |
| * while (a == a) { |
| * a = 4; |
| * } |
| * return 5; |
| * |
| * Which becomes the following graph: |
| * constant0 |
| * constant4 |
| * constant5 |
| * goto |
| * | |
| * goto |
| * | |
| * phi |
| * equal |
| * if +++++ |
| * | \ + |
| * | goto |
| * | |
| * return |
| * | |
| * exit |
| */ |
| |
| const std::vector<uint16_t> data = TWO_REGISTERS_CODE_ITEM( |
| Instruction::CONST_4 | 0 | 0, |
| Instruction::IF_EQ, 4, |
| Instruction::CONST_4 | 4 << 12 | 0, |
| Instruction::GOTO | 0xFD00, |
| Instruction::CONST_4 | 5 << 12 | 1 << 8, |
| Instruction::RETURN | 1 << 8); |
| |
| ASSERT_TRUE(Check(data)); |
| } |
| |
| TEST_F(RegisterAllocatorTest, Loop2) { |
| /* |
| * Test the following snippet: |
| * int a = 0; |
| * while (a == 8) { |
| * a = 4 + 5; |
| * } |
| * return 6 + 7; |
| * |
| * Which becomes the following graph: |
| * constant0 |
| * constant4 |
| * constant5 |
| * constant6 |
| * constant7 |
| * constant8 |
| * goto |
| * | |
| * goto |
| * | |
| * phi |
| * equal |
| * if +++++ |
| * | \ + |
| * | 4 + 5 |
| * | goto |
| * | |
| * 6 + 7 |
| * return |
| * | |
| * exit |
| */ |
| |
| const std::vector<uint16_t> data = TWO_REGISTERS_CODE_ITEM( |
| Instruction::CONST_4 | 0 | 0, |
| Instruction::CONST_4 | 8 << 12 | 1 << 8, |
| Instruction::IF_EQ | 1 << 8, 7, |
| Instruction::CONST_4 | 4 << 12 | 0 << 8, |
| Instruction::CONST_4 | 5 << 12 | 1 << 8, |
| Instruction::ADD_INT, 1 << 8 | 0, |
| Instruction::GOTO | 0xFA00, |
| Instruction::CONST_4 | 6 << 12 | 1 << 8, |
| Instruction::CONST_4 | 7 << 12 | 1 << 8, |
| Instruction::ADD_INT, 1 << 8 | 0, |
| Instruction::RETURN | 1 << 8); |
| |
| ASSERT_TRUE(Check(data)); |
| } |
| |
| TEST_F(RegisterAllocatorTest, Loop3) { |
| /* |
| * Test the following snippet: |
| * int a = 0 |
| * do { |
| * b = a; |
| * a++; |
| * } while (a != 5) |
| * return b; |
| * |
| * Which becomes the following graph: |
| * constant0 |
| * constant1 |
| * constant5 |
| * goto |
| * | |
| * goto |
| * |++++++++++++ |
| * phi + |
| * a++ + |
| * equals + |
| * if + |
| * |++++++++++++ |
| * return |
| * | |
| * exit |
| */ |
| |
| const std::vector<uint16_t> data = THREE_REGISTERS_CODE_ITEM( |
| Instruction::CONST_4 | 0 | 0, |
| Instruction::ADD_INT_LIT8 | 1 << 8, 1 << 8, |
| Instruction::CONST_4 | 5 << 12 | 2 << 8, |
| Instruction::IF_NE | 1 << 8 | 2 << 12, 3, |
| Instruction::RETURN | 0 << 8, |
| Instruction::MOVE | 1 << 12 | 0 << 8, |
| Instruction::GOTO | 0xF900); |
| |
| HGraph* graph = CreateCFG(data); |
| std::unique_ptr<CodeGenerator> codegen = CreateCodeGenerator(graph_); |
| CHECK(codegen != nullptr); |
| SsaLivenessAnalysis liveness(graph, codegen.get(), GetScopedAllocator()); |
| liveness.Analyze(); |
| std::unique_ptr<RegisterAllocator> register_allocator = |
| RegisterAllocator::Create(GetScopedAllocator(), codegen.get(), liveness); |
| register_allocator->AllocateRegisters(); |
| ASSERT_TRUE(register_allocator->Validate(false)); |
| |
| HBasicBlock* loop_header = graph->GetBlocks()[2]; |
| HPhi* phi = loop_header->GetFirstPhi()->AsPhi(); |
| |
| LiveInterval* phi_interval = phi->GetLiveInterval(); |
| LiveInterval* loop_update = phi->InputAt(1)->GetLiveInterval(); |
| ASSERT_TRUE(phi_interval->HasRegisters()); |
| ASSERT_TRUE(loop_update->HasRegisters()); |
| ASSERT_NE(phi_interval->GetRegisters(), loop_update->GetRegisters()); |
| |
| HBasicBlock* return_block = graph->GetBlocks()[3]; |
| HReturn* ret = return_block->GetLastInstruction()->AsReturn(); |
| ASSERT_EQ(phi_interval->GetRegisters(), ret->InputAt(0)->GetLiveInterval()->GetRegisters()); |
| } |
| |
| TEST_F(RegisterAllocatorTest, FirstRegisterUse) { |
| const std::vector<uint16_t> data = THREE_REGISTERS_CODE_ITEM( |
| Instruction::CONST_4 | 0 | 0, |
| Instruction::XOR_INT_LIT8 | 1 << 8, 1 << 8, |
| Instruction::XOR_INT_LIT8 | 0 << 8, 1 << 8, |
| Instruction::XOR_INT_LIT8 | 1 << 8, 1 << 8 | 1, |
| Instruction::RETURN_VOID); |
| |
| HGraph* graph = CreateCFG(data); |
| std::unique_ptr<CodeGenerator> codegen = CreateCodeGenerator(graph_, InstructionSet::kX86); |
| if (codegen == nullptr) { |
| GTEST_SKIP() << "X86 codegen is unavailable."; |
| } |
| SsaLivenessAnalysis liveness(graph, codegen.get(), GetScopedAllocator()); |
| liveness.Analyze(); |
| |
| HXor* first_xor = graph->GetBlocks()[1]->GetFirstInstruction()->AsXor(); |
| HXor* last_xor = graph->GetBlocks()[1]->GetLastInstruction()->GetPrevious()->AsXor(); |
| ASSERT_EQ(last_xor->InputAt(0), first_xor); |
| LiveInterval* interval = first_xor->GetLiveInterval(); |
| ASSERT_EQ(interval->GetEnd(), last_xor->GetLifetimePosition()); |
| ASSERT_TRUE(interval->GetNextSibling() == nullptr); |
| |
| // We need a register for the output of the instruction. |
| ASSERT_TRUE(interval->RequiresRegisterForDefinitionAt(interval->GetStart())); |
| ASSERT_EQ(interval->FirstRegisterUse(), last_xor->GetLifetimePosition()); |
| |
| // Split at the next instruction. |
| interval = interval->SplitAt(first_xor->GetLifetimePosition() + 2); |
| // The user of the split is the last add. |
| ASSERT_EQ(interval->FirstRegisterUse(), last_xor->GetLifetimePosition()); |
| |
| // Split before the last add. |
| LiveInterval* new_interval = interval->SplitAt(last_xor->GetLifetimePosition() - 1); |
| // Ensure the current interval has no register use... |
| ASSERT_EQ(interval->FirstRegisterUse(), kNoLifetime); |
| // And the new interval has it for the last add. |
| ASSERT_EQ(new_interval->FirstRegisterUse(), last_xor->GetLifetimePosition()); |
| } |
| |
| TEST_F(RegisterAllocatorTest, FpuSameAsFirstInput) { |
| HBasicBlock* block = InitEntryMainExitGraphWithReturnVoid(); |
| HInstruction* param1 = MakeParam(DataType::Type::kFloat32); |
| HNeg* neg = MakeUnOp<HNeg>(block, DataType::Type::kFloat32, param1); |
| LocationSummary* locations = LocationSummary::CreateNoCall(GetAllocator(), neg); |
| locations->SetInAt(0, Location::FpuRegister(0)); |
| locations->SetOut(Location::SameAsFirstInput()); |
| static constexpr size_t kLifetimePosition = 32u; |
| neg->SetLifetimePosition(kLifetimePosition); |
| LiveInterval* live_interval = LiveInterval::MakeInterval( |
| GetScopedAllocator(), DataType::Type::kFloat32, /*is_pair=*/ false, neg); |
| live_interval->AddRange(kLifetimePosition, kLifetimePosition + 1); |
| // This used to fail a `DCHECK()`. |
| bool requires_reg = live_interval->RequiresRegisterForDefinitionAt(kLifetimePosition); |
| EXPECT_TRUE(requires_reg); |
| } |
| |
| TEST_F(RegisterAllocatorTest, DeadPhi) { |
| /* Test for a dead loop phi taking as back-edge input a phi that also has |
| * this loop phi as input. Walking backwards in SsaDeadPhiElimination |
| * does not solve the problem because the loop phi will be visited last. |
| * |
| * Test the following snippet: |
| * int a = 0 |
| * do { |
| * if (true) { |
| * a = 2; |
| * } |
| * } while (true); |
| */ |
| |
| const std::vector<uint16_t> data = TWO_REGISTERS_CODE_ITEM( |
| Instruction::CONST_4 | 0 | 0, |
| Instruction::CONST_4 | 1 << 8 | 0, |
| Instruction::IF_NE | 1 << 8 | 1 << 12, 3, |
| Instruction::CONST_4 | 2 << 12 | 0 << 8, |
| Instruction::GOTO | 0xFD00, |
| Instruction::RETURN_VOID); |
| |
| HGraph* graph = CreateCFG(data); |
| SsaDeadPhiElimination(graph).Run(); |
| std::unique_ptr<CodeGenerator> codegen = CreateCodeGenerator(graph); |
| CHECK(codegen != nullptr); |
| SsaLivenessAnalysis liveness(graph, codegen.get(), GetScopedAllocator()); |
| liveness.Analyze(); |
| std::unique_ptr<RegisterAllocator> register_allocator = |
| RegisterAllocator::Create(GetScopedAllocator(), codegen.get(), liveness); |
| register_allocator->AllocateRegisters(); |
| ASSERT_TRUE(register_allocator->Validate(false)); |
| } |
| |
| /** |
| * Test that the TryAllocateFreeReg method works in the presence of inactive intervals |
| * that share the same register. It should split the interval it is currently |
| * allocating for at the minimum lifetime position between the two inactive intervals. |
| * This test only applies to the linear scan allocator. |
| */ |
| void RegisterAllocatorTest::TestFreeUntil(bool special_first) { |
| HBasicBlock* block = InitEntryMainExitGraphWithReturnVoid(); |
| HInstruction* const0 = graph_->GetIntConstant(0); |
| |
| HAdd* add = MakeBinOp<HAdd>(block, DataType::Type::kInt32, const0, const0); |
| HInstruction* placeholder1 = MakeUnOp<HNeg>(block, DataType::Type::kInt32, const0); |
| HInstruction* placeholder2 = MakeUnOp<HNeg>(block, DataType::Type::kInt32, const0); |
| MakeReturn(block, add); |
| |
| graph_->ComputeDominanceInformation(); |
| std::unique_ptr<CodeGenerator> codegen = CreateCodeGenerator(graph_, InstructionSet::kX86); |
| if (codegen == nullptr) { |
| GTEST_SKIP() << "X86 codegen is unavailable."; |
| } |
| SsaLivenessAnalysis liveness(graph_, codegen.get(), GetScopedAllocator()); |
| liveness.Analyze(); |
| |
| // Avoid allocating the register for the `const0` when used by the `add`. |
| add->GetLocations()->SetInAt(0, Location::ConstantLocation(const0)); |
| ASSERT_TRUE(add->GetLocations()->InAt(1).IsConstant()); |
| |
| // Record placeholder positions for blocking intervals and remove placeholders. |
| size_t blocking_pos1 = placeholder1->GetLiveInterval()->GetStart(); |
| size_t blocking_pos2 = placeholder2->GetLiveInterval()->GetStart(); |
| auto& const0_uses = const0->GetLiveInterval()->uses_; |
| ASSERT_EQ(4, std::distance(const0_uses.begin(), const0_uses.end())); |
| auto add_it = std::next(const0_uses.begin()); |
| ASSERT_TRUE(add_it->GetUser() == add); |
| for (HInstruction* placeholder : {placeholder1, placeholder2}) { |
| block->RemoveInstruction(placeholder); |
| ASSERT_TRUE(std::next(add_it)->GetUser() == placeholder); |
| const0_uses.erase_after(add_it); |
| // Set the block again in the dead placeholders to allow `liveness` to retrieve the block. |
| placeholder->SetBlock(block); |
| } |
| |
| // Set just one register available to make all intervals compete for the same. |
| BlockCoreRegistersExcept(codegen.get(), {x86::EAX}); |
| |
| RegisterAllocatorLinearScan register_allocator(GetScopedAllocator(), codegen.get(), liveness); |
| |
| // Test two variants, so that we hit the desired configuration once, no matter the order |
| // in which the register allocator inserts the blocking intervals into inactive intervals. |
| size_t call_pos = special_first ? blocking_pos2 : blocking_pos1; |
| size_t special_pos = special_first ? blocking_pos1 : blocking_pos2; |
| register_allocator.block_registers_for_call_interval_->AddRange(call_pos, call_pos + 1); |
| register_allocator.block_registers_special_interval_->AddRange(special_pos, special_pos + 1); |
| |
| register_allocator.AllocateRegistersInternal(); |
| |
| std::pair<size_t, int> expected_add_start_and_regs[] = { |
| {add->GetLifetimePosition() + kLivenessPositionOfNonOverlappingOutput, 1u << 0}, |
| {blocking_pos1, kNoRegisters}, |
| }; |
| LiveInterval* li = add->GetLiveInterval(); |
| for (const std::pair<size_t, int>& expected_start_and_regs : expected_add_start_and_regs) { |
| ASSERT_TRUE(li != nullptr); |
| ASSERT_EQ(expected_start_and_regs.first, li->GetStart()); |
| ASSERT_EQ(expected_start_and_regs.second, li->GetRegisters()); |
| li = li->GetNextSibling(); |
| } |
| ASSERT_TRUE(li == nullptr); |
| } |
| |
| TEST_F(RegisterAllocatorTest, FreeUntilCallFirst) { |
| TestFreeUntil(/*special_first=*/ false); |
| } |
| |
| TEST_F(RegisterAllocatorTest, FreeUntilSpecialFirst) { |
| TestFreeUntil(/*special_first=*/ true); |
| } |
| |
| void RegisterAllocatorTest::BuildIfElseWithPhi(HPhi** phi, |
| HInstruction** input1, |
| HInstruction** input2) { |
| HBasicBlock* join = InitEntryMainExitGraph(); |
| auto [if_block, then, else_] = CreateDiamondPattern(join); |
| HInstruction* parameter = MakeParam(DataType::Type::kReference); |
| HInstruction* test = MakeIFieldGet(if_block, parameter, DataType::Type::kBool, MemberOffset(22)); |
| MakeIf(if_block, test); |
| |
| *input1 = MakeIFieldGet(then, parameter, DataType::Type::kInt32, MemberOffset(42)); |
| *input2 = MakeIFieldGet(else_, parameter, DataType::Type::kInt32, MemberOffset(42)); |
| *phi = MakePhi(join, {*input1, *input2}); |
| MakeReturn(join, *phi); |
| |
| graph_->BuildDominatorTree(); |
| graph_->AnalyzeLoops(); |
| } |
| |
| TEST_F(RegisterAllocatorTest, PhiHint) { |
| HPhi *phi; |
| HInstruction *input1, *input2; |
| |
| { |
| BuildIfElseWithPhi(&phi, &input1, &input2); |
| std::unique_ptr<CodeGenerator> codegen = CreateCodeGenerator(graph_, InstructionSet::kX86); |
| if (codegen == nullptr) { |
| GTEST_SKIP() << "X86 codegen is unavailable."; |
| } |
| CHECK(codegen != nullptr); |
| SsaLivenessAnalysis liveness(graph_, codegen.get(), GetScopedAllocator()); |
| liveness.Analyze(); |
| |
| // Check that the register allocator is deterministic. |
| std::unique_ptr<RegisterAllocator> register_allocator = |
| RegisterAllocator::Create(GetScopedAllocator(), codegen.get(), liveness); |
| register_allocator->AllocateRegisters(); |
| |
| ASSERT_EQ(input1->GetLiveInterval()->GetRegisters(), 1u << 0); |
| ASSERT_EQ(input2->GetLiveInterval()->GetRegisters(), 1u << 0); |
| ASSERT_EQ(phi->GetLiveInterval()->GetRegisters(), 1u << 0); |
| } |
| |
| { |
| BuildIfElseWithPhi(&phi, &input1, &input2); |
| std::unique_ptr<CodeGenerator> codegen = CreateCodeGenerator(graph_, InstructionSet::kX86); |
| CHECK(codegen != nullptr); |
| SsaLivenessAnalysis liveness(graph_, codegen.get(), GetScopedAllocator()); |
| liveness.Analyze(); |
| |
| // Set the phi to a specific register, and check that the inputs get allocated |
| // the same register. |
| phi->GetLocations()->UpdateOut(Location::CoreRegister(2)); |
| std::unique_ptr<RegisterAllocator> register_allocator = |
| RegisterAllocator::Create(GetScopedAllocator(), codegen.get(), liveness); |
| register_allocator->AllocateRegisters(); |
| |
| ASSERT_EQ(input1->GetLiveInterval()->GetRegisters(), 1u << 2); |
| ASSERT_EQ(input2->GetLiveInterval()->GetRegisters(), 1u << 2); |
| ASSERT_EQ(phi->GetLiveInterval()->GetRegisters(), 1u << 2); |
| } |
| |
| { |
| BuildIfElseWithPhi(&phi, &input1, &input2); |
| std::unique_ptr<CodeGenerator> codegen = CreateCodeGenerator(graph_, InstructionSet::kX86); |
| CHECK(codegen != nullptr); |
| SsaLivenessAnalysis liveness(graph_, codegen.get(), GetScopedAllocator()); |
| liveness.Analyze(); |
| |
| // Set input1 to a specific register, and check that the phi and other input get allocated |
| // the same register. |
| input1->GetLocations()->UpdateOut(Location::CoreRegister(2)); |
| std::unique_ptr<RegisterAllocator> register_allocator = |
| RegisterAllocator::Create(GetScopedAllocator(), codegen.get(), liveness); |
| register_allocator->AllocateRegisters(); |
| |
| ASSERT_EQ(input1->GetLiveInterval()->GetRegisters(), 1u << 2); |
| ASSERT_EQ(input2->GetLiveInterval()->GetRegisters(), 1u << 2); |
| ASSERT_EQ(phi->GetLiveInterval()->GetRegisters(), 1u << 2); |
| } |
| |
| { |
| BuildIfElseWithPhi(&phi, &input1, &input2); |
| std::unique_ptr<CodeGenerator> codegen = CreateCodeGenerator(graph_, InstructionSet::kX86); |
| CHECK(codegen != nullptr); |
| SsaLivenessAnalysis liveness(graph_, codegen.get(), GetScopedAllocator()); |
| liveness.Analyze(); |
| |
| // Set input2 to a specific register, and check that the phi and other input get allocated |
| // the same register. |
| input2->GetLocations()->UpdateOut(Location::CoreRegister(2)); |
| std::unique_ptr<RegisterAllocator> register_allocator = |
| RegisterAllocator::Create(GetScopedAllocator(), codegen.get(), liveness); |
| register_allocator->AllocateRegisters(); |
| |
| ASSERT_EQ(input1->GetLiveInterval()->GetRegisters(), 1u << 2); |
| ASSERT_EQ(input2->GetLiveInterval()->GetRegisters(), 1u << 2); |
| ASSERT_EQ(phi->GetLiveInterval()->GetRegisters(), 1u << 2); |
| } |
| } |
| |
| void RegisterAllocatorTest::BuildFieldReturn(HInstruction** field, HInstruction** ret) { |
| HBasicBlock* block = InitEntryMainExitGraph(); |
| HInstruction* parameter = MakeParam(DataType::Type::kReference); |
| |
| *field = MakeIFieldGet(block, parameter, DataType::Type::kInt32, MemberOffset(42)); |
| *ret = MakeReturn(block, *field); |
| |
| graph_->BuildDominatorTree(); |
| } |
| |
| TEST_F(RegisterAllocatorTest, ExpectedInRegisterHint) { |
| HInstruction *field, *ret; |
| |
| { |
| BuildFieldReturn(&field, &ret); |
| std::unique_ptr<CodeGenerator> codegen = CreateCodeGenerator(graph_, InstructionSet::kX86); |
| if (codegen == nullptr) { |
| GTEST_SKIP() << "X86 codegen is unavailable."; |
| } |
| SsaLivenessAnalysis liveness(graph_, codegen.get(), GetScopedAllocator()); |
| liveness.Analyze(); |
| |
| std::unique_ptr<RegisterAllocator> register_allocator = |
| RegisterAllocator::Create(GetScopedAllocator(), codegen.get(), liveness); |
| register_allocator->AllocateRegisters(); |
| |
| // Check the validity that in normal conditions, the register should be hinted to 0 (EAX). |
| ASSERT_EQ(field->GetLiveInterval()->GetRegisters(), 1u << 0); |
| } |
| |
| { |
| BuildFieldReturn(&field, &ret); |
| std::unique_ptr<CodeGenerator> codegen = CreateCodeGenerator(graph_, InstructionSet::kX86); |
| CHECK(codegen != nullptr); |
| SsaLivenessAnalysis liveness(graph_, codegen.get(), GetScopedAllocator()); |
| liveness.Analyze(); |
| |
| // Check that the field gets put in the register expected by its use. |
| // Don't use SetInAt because we are overriding an already allocated location. |
| ret->GetLocations()->Inputs()[0] = Location::CoreRegister(2); |
| |
| std::unique_ptr<RegisterAllocator> register_allocator = |
| RegisterAllocator::Create(GetScopedAllocator(), codegen.get(), liveness); |
| register_allocator->AllocateRegisters(); |
| |
| ASSERT_EQ(field->GetLiveInterval()->GetRegisters(), 1u << 2); |
| } |
| } |
| |
| void RegisterAllocatorTest::BuildTwoSubs(HInstruction** first_sub, HInstruction** second_sub) { |
| HBasicBlock* block = InitEntryMainExitGraph(); |
| HInstruction* parameter = MakeParam(DataType::Type::kInt32); |
| HInstruction* constant1 = graph_->GetIntConstant(1); |
| HInstruction* constant2 = graph_->GetIntConstant(2); |
| |
| *first_sub = new (GetAllocator()) HSub(DataType::Type::kInt32, parameter, constant1); |
| block->AddInstruction(*first_sub); |
| *second_sub = new (GetAllocator()) HSub(DataType::Type::kInt32, *first_sub, constant2); |
| block->AddInstruction(*second_sub); |
| MakeReturn(block, *second_sub); |
| |
| graph_->BuildDominatorTree(); |
| } |
| |
| TEST_F(RegisterAllocatorTest, SameAsFirstInputHint) { |
| HInstruction *first_sub, *second_sub; |
| |
| { |
| BuildTwoSubs(&first_sub, &second_sub); |
| std::unique_ptr<CodeGenerator> codegen = CreateCodeGenerator(graph_, InstructionSet::kX86); |
| if (codegen == nullptr) { |
| GTEST_SKIP() << "X86 codegen is unavailable."; |
| } |
| SsaLivenessAnalysis liveness(graph_, codegen.get(), GetScopedAllocator()); |
| liveness.Analyze(); |
| |
| std::unique_ptr<RegisterAllocator> register_allocator = |
| RegisterAllocator::Create(GetScopedAllocator(), codegen.get(), liveness); |
| register_allocator->AllocateRegisters(); |
| |
| // Check the validity that in normal conditions, the registers are the same. |
| ASSERT_EQ(first_sub->GetLiveInterval()->GetRegisters(), 1u << 1); |
| ASSERT_EQ(second_sub->GetLiveInterval()->GetRegisters(), 1u << 1); |
| } |
| |
| { |
| BuildTwoSubs(&first_sub, &second_sub); |
| std::unique_ptr<CodeGenerator> codegen = CreateCodeGenerator(graph_, InstructionSet::kX86); |
| CHECK(codegen != nullptr); |
| SsaLivenessAnalysis liveness(graph_, codegen.get(), GetScopedAllocator()); |
| liveness.Analyze(); |
| |
| // check that both adds get the same register. |
| // Don't use UpdateOutput because output is already allocated. |
| OverrideOutput(first_sub->InputAt(0)->GetLocations(), Location::CoreRegister(2)); |
| ASSERT_EQ(first_sub->GetLocations()->Out().GetPolicy(), Location::kSameAsFirstInput); |
| ASSERT_EQ(second_sub->GetLocations()->Out().GetPolicy(), Location::kSameAsFirstInput); |
| |
| std::unique_ptr<RegisterAllocator> register_allocator = |
| RegisterAllocator::Create(GetScopedAllocator(), codegen.get(), liveness); |
| register_allocator->AllocateRegisters(); |
| |
| ASSERT_EQ(first_sub->GetLiveInterval()->GetRegisters(), 1u << 2); |
| ASSERT_EQ(second_sub->GetLiveInterval()->GetRegisters(), 1u << 2); |
| } |
| } |
| |
| void RegisterAllocatorTest::BuildDiv(HInstruction** div) { |
| HBasicBlock* block = InitEntryMainExitGraph(); |
| HInstruction* first = MakeParam(DataType::Type::kInt32); |
| HInstruction* second = MakeParam(DataType::Type::kInt32); |
| |
| *div = new (GetAllocator()) HDiv( |
| DataType::Type::kInt32, first, second, 0); // don't care about dex_pc. |
| block->AddInstruction(*div); |
| MakeReturn(block, *div); |
| |
| graph_->BuildDominatorTree(); |
| } |
| |
| TEST_F(RegisterAllocatorTest, ExpectedExactInRegisterAndSameOutputHint) { |
| HInstruction *div; |
| BuildDiv(&div); |
| std::unique_ptr<CodeGenerator> codegen = CreateCodeGenerator(graph_, InstructionSet::kX86); |
| if (codegen == nullptr) { |
| GTEST_SKIP() << "X86 codegen is unavailable."; |
| } |
| SsaLivenessAnalysis liveness(graph_, codegen.get(), GetScopedAllocator()); |
| liveness.Analyze(); |
| |
| std::unique_ptr<RegisterAllocator> register_allocator = |
| RegisterAllocator::Create(GetScopedAllocator(), codegen.get(), liveness); |
| register_allocator->AllocateRegisters(); |
| |
| // div on x86 requires its first input in eax and the output be the same as the first input. |
| ASSERT_EQ(div->GetLiveInterval()->GetRegisters(), 1u << 0); |
| } |
| |
| // Test a bug in the register allocator, where allocating a blocked |
| // register would lead to spilling an inactive interval at the wrong |
| // position. |
| // This test only applies to the linear scan allocator. |
| void RegisterAllocatorTest::TestSpillInactive() { |
| HBasicBlock* block = InitEntryMainExitGraphWithReturnVoid(); |
| HInstruction* one = MakeParam(DataType::Type::kInt32); |
| HInstruction* two = MakeParam(DataType::Type::kInt32); |
| HInstruction* three = MakeParam(DataType::Type::kInt32); |
| HInstruction* four = MakeParam(DataType::Type::kInt32); |
| |
| // We create a synthesized user requesting a register, to avoid just spilling the |
| // intervals. |
| HPhi* user = new (GetAllocator()) HPhi(GetAllocator(), 0, 1, DataType::Type::kInt32); |
| user->SetBlock(block); |
| user->AddInput(one); |
| LocationSummary* locations = LocationSummary::CreateNoCall(GetAllocator(), user); |
| locations->SetInAt(0, Location::RequiresCoreRegister()); |
| static constexpr size_t phi_ranges[][2] = {{10 * kLppi, 15 * kLppi}}; |
| BuildInterval(phi_ranges, arraysize(phi_ranges), GetScopedAllocator(), kNoRegisters, user); |
| |
| // Create an interval with lifetime holes. |
| static constexpr size_t ranges1[][2] = |
| {{0u * kLppi, 2u * kLppi}, {4u * kLppi, 5u * kLppi}, {7u * kLppi, 8u * kLppi}}; |
| LiveInterval* first = |
| BuildInterval(ranges1, arraysize(ranges1), GetScopedAllocator(), kNoRegisters, one); |
| first->uses_.push_front(*new (GetScopedAllocator()) UsePosition(user, 0u, 7u * kLppi)); |
| first->uses_.push_front(*new (GetScopedAllocator()) UsePosition(user, 0u, 6u * kLppi)); |
| first->uses_.push_front(*new (GetScopedAllocator()) UsePosition(user, 0u, 5u * kLppi)); |
| |
| locations = LocationSummary::CreateNoCall(GetAllocator(), first->GetDefinedBy()); |
| locations->SetOut(Location::RequiresCoreRegister()); |
| first = first->SplitAt(1u * kLppi); |
| |
| // Create an interval that conflicts with the next interval, to force the next |
| // interval to call `AllocateBlockedReg`. |
| static constexpr size_t ranges2[][2] = {{2u * kLppi, 4u * kLppi}}; |
| LiveInterval* second = |
| BuildInterval(ranges2, arraysize(ranges2), GetScopedAllocator(), kNoRegisters, two); |
| locations = LocationSummary::CreateNoCall(GetAllocator(), second->GetDefinedBy()); |
| locations->SetOut(Location::RequiresCoreRegister()); |
| |
| // Create an interval that will lead to splitting the first interval. The bug occured |
| // by splitting at a wrong position, in this case at the next intersection between |
| // this interval and the first interval. We would have then put the interval with ranges |
| // "[0, 2(, [4, 6(" in the list of handled intervals, even though we haven't processed intervals |
| // before lifetime position 6 yet. |
| static constexpr size_t ranges3[][2] = {{2u * kLppi, 4u * kLppi}, {7u * kLppi, 8u * kLppi}}; |
| LiveInterval* third = |
| BuildInterval(ranges3, arraysize(ranges3), GetScopedAllocator(), kNoRegisters, three); |
| third->uses_.push_front(*new (GetScopedAllocator()) UsePosition(user, 0u, 7u * kLppi)); |
| third->uses_.push_front(*new (GetScopedAllocator()) UsePosition(user, 0u, 4u * kLppi)); |
| third->uses_.push_front(*new (GetScopedAllocator()) UsePosition(user, 0u, 3u * kLppi)); |
| locations = LocationSummary::CreateNoCall(GetAllocator(), third->GetDefinedBy()); |
| locations->SetOut(Location::RequiresCoreRegister()); |
| third = third->SplitAt(3u * kLppi); |
| |
| // Because the first part of the split interval was considered handled, this interval |
| // was free to allocate the same register, even though it conflicts with it. |
| static constexpr size_t ranges4[][2] = {{4u * kLppi, 5u * kLppi}}; |
| LiveInterval* fourth = |
| BuildInterval(ranges4, arraysize(ranges4), GetScopedAllocator(), kNoRegisters, four); |
| locations = LocationSummary::CreateNoCall(GetAllocator(), fourth->GetDefinedBy()); |
| locations->SetOut(Location::RequiresCoreRegister()); |
| |
| std::unique_ptr<CodeGenerator> codegen = CreateCodeGenerator(graph_, InstructionSet::kX86); |
| if (codegen == nullptr) { |
| GTEST_SKIP() << "X86 codegen is unavailable."; |
| } |
| SsaLivenessAnalysis liveness(graph_, codegen.get(), GetScopedAllocator()); |
| // Populate the instructions in the liveness object, to please the register allocator. |
| liveness.instructions_from_lifetime_position_.assign(16, user); |
| |
| // Set just one register available to make all intervals compete for the same. |
| BlockCoreRegistersExcept(codegen.get(), {x86::EAX}); |
| |
| RegisterAllocatorLinearScan register_allocator(GetScopedAllocator(), codegen.get(), liveness); |
| register_allocator.unhandled_core_intervals_.assign({fourth, third, second, first}); |
| |
| // We have set up all intervals manually and we want `AllocateRegistersInternal()` to run |
| // the linear scan without processing instructions - check that the linear order is empty. |
| ASSERT_TRUE(codegen->GetGraph()->GetLinearOrder().empty()); |
| register_allocator.AllocateRegistersInternal(); |
| |
| // Test that there is no conflicts between intervals. |
| ScopedArenaVector<LiveInterval*> intervals({first, second, third, fourth}, |
| GetScopedAllocator()->Adapter()); |
| ASSERT_TRUE(ValidateIntervals(intervals, *codegen)); |
| } |
| |
| TEST_F(RegisterAllocatorTest, SpillInactive) { |
| TestSpillInactive(); |
| } |
| |
| TEST_F(RegisterAllocatorTest, ReuseSpillSlots) { |
| HBasicBlock* return_block = InitEntryMainExitGraph(); |
| auto [start, left, right] = CreateDiamondPattern(return_block); |
| HInstruction* obj = MakeParam(DataType::Type::kReference); |
| HInstruction* cond = MakeIFieldGet(start, obj, DataType::Type::kBool, MemberOffset(32)); |
| MakeIf(start, cond); |
| |
| // Load two values from fields. Both shall be used as Phi inputs later. |
| HInstruction* left_get1 = MakeIFieldGet(left, obj, DataType::Type::kInt32, MemberOffset(36)); |
| HInstruction* left_get2 = MakeIFieldGet(left, obj, DataType::Type::kInt32, MemberOffset(40)); |
| // Convert one of the values to `Int64` to spill the loaded values. |
| HInstruction* left_conv1 = MakeUnOp<HTypeConversion>(left, DataType::Type::kInt64, left_get1); |
| // Convert the `Int64` value back to `Int32`. x86 codegen uses EAX and EDX for conversion |
| // which is not a normal pair, so avoid using this odd explicit pair for a Phi. |
| HInstruction* left_conv2 = MakeUnOp<HTypeConversion>(left, DataType::Type::kInt32, left_conv1); |
| |
| // Repeat the sequence from `left` block in the `right` block (with different offsets). Without |
| // spill slot hints, spill slots should be assigned the same way as in the `left` block. |
| HInstruction* right_get1 = MakeIFieldGet(right, obj, DataType::Type::kInt32, MemberOffset(44)); |
| HInstruction* right_get2 = MakeIFieldGet(right, obj, DataType::Type::kInt32, MemberOffset(48)); |
| HInstruction* right_conv1 = MakeUnOp<HTypeConversion>(right, DataType::Type::kInt64, right_get1); |
| HInstruction* right_conv2 = MakeUnOp<HTypeConversion>(right, DataType::Type::kInt32, right_conv1); |
| |
| // Add Phis that tie the first field load in `left` to the second field load in `right` and |
| // vice versa, to check that the hints can align the spill slots assigned to inputs. |
| HPhi* phi1 = MakePhi(return_block, {left_get1, right_get2}); |
| HPhi* phi2 = MakePhi(return_block, {left_get2, right_get1}); |
| |
| // Add some instructions that use the `phi1`, `phi2` and even the converted values |
| // to derive some return value. |
| HPhi* phi_conv = MakePhi(return_block, {left_conv2, right_conv2}); |
| HMin* min1 = MakeBinOp<HMin>(return_block, DataType::Type::kInt32, phi1, phi2); |
| HMin* min2 = MakeBinOp<HMin>(return_block, DataType::Type::kInt32, min1, phi_conv); |
| MakeReturn(return_block, min2); |
| |
| graph_->ComputeDominanceInformation(); |
| std::unique_ptr<CodeGenerator> codegen = CreateCodeGenerator(graph_, InstructionSet::kX86); |
| if (codegen == nullptr) { |
| GTEST_SKIP() << "X86 codegen is unavailable."; |
| } |
| SsaLivenessAnalysis liveness(graph_, codegen.get(), GetScopedAllocator()); |
| liveness.Analyze(); |
| |
| // Set just two registers available to make it easy to force spills. |
| // Choose EAX and EDX which are used by type conversion from Int32 to Int64, so that |
| // we can use the type conversion to spill all live intervals wherever we want. |
| BlockCoreRegistersExcept(codegen.get(), {x86::EAX, x86::EDX}); |
| |
| // Change the `obj` parameter to come in EDX. |
| OverrideOutput(obj->GetLocations(), Location::CoreRegister(x86::EDX)); |
| |
| std::unique_ptr<RegisterAllocator> register_allocator = |
| RegisterAllocator::Create(GetScopedAllocator(), codegen.get(), liveness); |
| register_allocator->AllocateRegisters(); |
| |
| // Field loads would be spilled even without using spill slot hints. |
| ASSERT_TRUE(left_get1->GetLiveInterval()->HasSpillSlot()); |
| ASSERT_TRUE(left_get2->GetLiveInterval()->HasSpillSlot()); |
| ASSERT_TRUE(right_get1->GetLiveInterval()->HasSpillSlot()); |
| ASSERT_TRUE(right_get2->GetLiveInterval()->HasSpillSlot()); |
| |
| // Input spill slots are aligned thanks to the spill slot hints. |
| EXPECT_EQ(left_get1->GetLiveInterval()->GetSpillSlot(), |
| right_get2->GetLiveInterval()->GetSpillSlot()); |
| EXPECT_EQ(left_get2->GetLiveInterval()->GetSpillSlot(), |
| right_get1->GetLiveInterval()->GetSpillSlot()); |
| |
| // Check that `phi1` and `phi2` use the spill slots used by their inputs. |
| EXPECT_TRUE(phi1->GetLiveInterval()->HasSpillSlot()); |
| EXPECT_EQ(left_get1->GetLiveInterval()->GetSpillSlot(), phi1->GetLiveInterval()->GetSpillSlot()); |
| EXPECT_TRUE(phi2->GetLiveInterval()->HasSpillSlot()); |
| EXPECT_EQ(left_get2->GetLiveInterval()->GetSpillSlot(), phi2->GetLiveInterval()->GetSpillSlot()); |
| |
| // Check that `phi1` and `phi2` are split and don't have a register in the first sibling. |
| EXPECT_TRUE(phi1->GetLiveInterval()->GetNextSibling() != nullptr); |
| EXPECT_TRUE(!phi1->GetLiveInterval()->HasRegisters()); |
| EXPECT_TRUE(phi2->GetLiveInterval()->GetNextSibling() != nullptr); |
| EXPECT_TRUE(!phi2->GetLiveInterval()->HasRegisters()); |
| } |
| |
| TEST_F(RegisterAllocatorTest, ReuseSpillSlotGaps) { |
| HBasicBlock* return_block = InitEntryMainExitGraph(); |
| auto [pre_header, header, body] = CreateWhileLoop(return_block); |
| |
| HInstruction* const0 = graph_->GetIntConstant(0); |
| HInstruction* const10 = graph_->GetIntConstant(10); |
| |
| HPhi* phi1 = MakePhi(header, {const0, /* placeholder */ const0}); |
| HNeg* neg1 = MakeUnOp<HNeg>(body, DataType::Type::kInt32, phi1); |
| phi1->ReplaceInput(neg1, 1u); // Update back-edge input. |
| |
| HPhi* phi2 = MakePhi(header, {const0, /* placeholder */ const0}); |
| HNeg* neg2 = MakeUnOp<HNeg>(body, DataType::Type::kInt32, phi2); |
| phi2->ReplaceInput(neg2, 1u); // Update back-edge input. |
| |
| // Loop variable and condition. This is added after `neg1` and `neg2` to spill both. |
| HPhi* phi = MakePhi(header, {const0, /* placeholder */ const0}); |
| HNeg* neg = MakeUnOp<HNeg>(body, DataType::Type::kInt32, phi); |
| phi->ReplaceInput(neg, 1u); // Update back-edge input. |
| HCondition* cond = MakeCondition(header, kCondGE, phi, const10); |
| MakeIf(header, cond); |
| |
| // Add an environment use of `phi1` and a normal use of `phi2`. |
| HCondition* deopt_cond = MakeCondition(header, kCondLT, phi, const0); |
| HDeoptimize* deopt = new (GetAllocator()) HDeoptimize( |
| GetAllocator(), deopt_cond, DeoptimizationKind::kDebugging, /*dex_pc=*/ 0u); |
| AddOrInsertInstruction(return_block, deopt); |
| ManuallyBuildEnvFor(deopt, {phi1}); |
| MakeReturn(return_block, phi2); |
| |
| graph_->BuildDominatorTree(); |
| std::unique_ptr<CodeGenerator> codegen = CreateCodeGenerator(graph_, InstructionSet::kX86); |
| if (codegen == nullptr) { |
| GTEST_SKIP() << "X86 codegen is unavailable."; |
| } |
| SsaLivenessAnalysis liveness(graph_, codegen.get(), GetScopedAllocator()); |
| liveness.Analyze(); |
| |
| // Set just one register available to make all intervals compete for the same. |
| BlockCoreRegistersExcept(codegen.get(), {x86::EAX}); |
| // Rewrite condition locations to work with the single register EAX. |
| for (HCondition* c : {cond, deopt_cond}) { |
| ASSERT_TRUE(c->GetLocations()->Out().Equals(Location::CoreRegister(x86::ECX))); |
| OverrideOutput(c->GetLocations(), Location::CoreRegister(x86::EAX)); |
| c->GetLocations()->SetInAt(0, Location::Any()); |
| ASSERT_TRUE(c->GetLocations()->InAt(1).Equals(Location::Any())); |
| } |
| |
| std::unique_ptr<RegisterAllocator> register_allocator = |
| RegisterAllocator::Create(GetScopedAllocator(), codegen.get(), liveness); |
| register_allocator->AllocateRegisters(); |
| |
| ASSERT_TRUE(phi1->GetLiveInterval()->HasSpillSlot()); |
| ASSERT_TRUE(neg1->GetLiveInterval()->HasSpillSlot()); |
| EXPECT_EQ(phi1->GetLiveInterval()->GetSpillSlot(), neg1->GetLiveInterval()->GetSpillSlot()); |
| ASSERT_TRUE(phi2->GetLiveInterval()->HasSpillSlot()); |
| ASSERT_TRUE(neg2->GetLiveInterval()->HasSpillSlot()); |
| EXPECT_EQ(phi2->GetLiveInterval()->GetSpillSlot(), neg2->GetLiveInterval()->GetSpillSlot()); |
| } |
| |
| // Regression test for wrongly assuming that a Phi interval with a spill slot hint |
| // is not split when checking if the spill slot can be used. Indeed, it can be split |
| // and we must use the sibling to determine the lifetime end. |
| TEST_F(RegisterAllocatorTest, ReuseSpillSlotsUnavailableWithSplitPhiInterval) { |
| if (!com::android::art::flags::reg_alloc_no_output_overlap()) { |
| GTEST_SKIP() << "Improved `Location::kNoOutputOverlap` handling disabled."; |
| } |
| HBasicBlock* return_block = InitEntryMainExitGraph(); |
| auto [start, left, right] = CreateDiamondPattern(return_block); |
| HInstruction* const0 = graph_->GetIntConstant(0); |
| HInstruction* obj = MakeParam(DataType::Type::kReference); |
| HInstruction* cond = MakeIFieldGet(start, obj, DataType::Type::kBool, MemberOffset(32)); |
| MakeIf(start, cond); |
| |
| // Add a load followed by `HNeg`, so that the loaded value is spilled. |
| HInstruction* left_get = MakeIFieldGet(left, obj, DataType::Type::kInt32, MemberOffset(36)); |
| HNeg* left_neg = MakeUnOp<HNeg>(left, DataType::Type::kInt32, left_get); |
| |
| // Repeat the sequence from `left` block in the `right` block (with a different offset). |
| HInstruction* right_get = MakeIFieldGet(right, obj, DataType::Type::kInt32, MemberOffset(40)); |
| HNeg* right_neg = MakeUnOp<HNeg>(right, DataType::Type::kInt32, right_get); |
| |
| // Phis shall be processed in the order in which they are inserted. |
| // The first Phi shall initially be allocated the only available register. |
| HPhi* phi1 = MakePhi(return_block, {const0, right_neg}); |
| // The second Phi has no register use, so it shall be spilled. |
| // The spill slot used by `left_get` and `right_get` shall be reused for this unrelated Phi. |
| HPhi* phi2 = MakePhi(return_block, {left_neg, const0}); |
| // The third Phi has a hint that would put it to the same spill slot as both of its inputs |
| // if it was not already taken by `phi2`. But the spill slot is no longer available, so we |
| // try to allocate a register. Since `get_phi`'s first register use is before the `phi1`'s |
| // first register use, we allocate the register for `get_phi` and re-insert `phi1` to the |
| // unhandled intervals. However, since the last use of `get_phi` is after the `invoke` |
| // which blocks the register, `get_phi`'s interval shall be split in the process. |
| HPhi* get_phi = MakePhi(return_block, {left_get, right_get}); |
| // The fourth Phi has an earlier register use than `get_phi`, so it shall be allocated |
| // the register and `get_phi`'s interval shall be re-inserted to unhandled intervals. |
| HPhi* neg_phi = MakePhi(return_block, {left_neg, right_neg}); |
| // Then we shall re-process `phi1`, assigning the next spill slot. |
| // Then we shall re-process `get_phi` and try to assign the hint slot where we previously |
| // wrongly assumed that it has no sibling and triggered a `DCHECK()`. |
| |
| // Add a register use for the `neg_phi`. |
| HNeg* neg_neg = MakeUnOp<HNeg>(return_block, DataType::Type::kInt32, neg_phi); |
| // Add a register use for the `get_phi`. |
| // Use `HSub` which can have the second operand on the stack for x86. |
| HSub* sub1 = MakeBinOp<HSub>(return_block, DataType::Type::kInt32, get_phi, neg_neg); |
| // Add an invoke that forces the `get_phi` interval to be split when initially allocated. |
| MakeInvokeStatic(return_block, DataType::Type::kVoid, {}, {}); |
| // Add another register use for the `get_phi` after the `invoke`. |
| HSub* sub2 = MakeBinOp<HSub>(return_block, DataType::Type::kInt32, get_phi, sub1); |
| |
| // Add some instructions that use all the values to derive some return value. |
| HSub* sub3 = MakeBinOp<HSub>(return_block, DataType::Type::kInt32, sub2, phi1); |
| HSub* sub4 = MakeBinOp<HSub>(return_block, DataType::Type::kInt32, sub3, phi2); |
| MakeReturn(return_block, sub4); |
| |
| graph_->ComputeDominanceInformation(); |
| std::unique_ptr<CodeGenerator> codegen = CreateCodeGenerator(graph_, InstructionSet::kX86); |
| if (codegen == nullptr) { |
| GTEST_SKIP() << "X86 codegen is unavailable."; |
| } |
| SsaLivenessAnalysis liveness(graph_, codegen.get(), GetScopedAllocator()); |
| liveness.Analyze(); |
| |
| // Set just one register available to make all intervals compete for the same. |
| BlockCoreRegistersExcept(codegen.get(), {x86::EAX}); |
| // Change the `obj` parameter to come in EAX. |
| OverrideOutput(obj->GetLocations(), Location::CoreRegister(x86::EAX)); |
| |
| std::unique_ptr<RegisterAllocator> register_allocator = |
| RegisterAllocator::Create(GetScopedAllocator(), codegen.get(), liveness); |
| register_allocator->AllocateRegisters(); |
| |
| ASSERT_TRUE(left_get->GetLiveInterval()->HasSpillSlot()); |
| ASSERT_TRUE(right_get->GetLiveInterval()->HasSpillSlot()); |
| ASSERT_EQ(left_get->GetLiveInterval()->GetSpillSlot(), |
| right_get->GetLiveInterval()->GetSpillSlot()); |
| |
| ASSERT_TRUE(phi2->GetLiveInterval()->HasSpillSlot()); |
| ASSERT_EQ(left_get->GetLiveInterval()->GetSpillSlot(), phi2->GetLiveInterval()->GetSpillSlot()); |
| |
| ASSERT_TRUE(get_phi->GetLiveInterval()->HasSpillSlot()); |
| ASSERT_NE(left_get->GetLiveInterval()->GetSpillSlot(), |
| get_phi->GetLiveInterval()->GetSpillSlot()); |
| } |
| |
| // Regression test for splitting a live range that's recorded as a search hint in the |
| // `SpillSlotData`. Previously, a new range was created for the first part of the live |
| // range and the original `LiveRange` object had its start adjusted but that means the |
| // search hint was pointing to the second part of the split live range and we missed |
| // overlaps when trying to reuse a spill slot. Bug: 426785078 |
| TEST_F(RegisterAllocatorTest, SplitSpillSlotLiveRangeHint) { |
| // Create a graph with a three-way switch to blocks `left`, `mid` and `right`, |
| // and a diamond pattern in the `left` block with branch blocks `left_left` |
| // and `left_right` and merging to `left_end`. The linear order shall have |
| // the block `right` followed by `mid` and `mid` followed by `left`. |
| HBasicBlock* return_block = InitEntryMainExitGraph(); |
| auto [start, left_end, mid] = CreateDiamondPattern(return_block); |
| HBasicBlock* right = AddNewBlock(); |
| start->AddSuccessor(right); |
| right->AddSuccessor(return_block); |
| auto [left, left_left, left_right] = CreateDiamondPattern(left_end); |
| |
| // Create `iget_*` instructions in `start`, `right`, `mid` and `left_right` |
| // (listed in linear order). Merge `iget_start` and `iget_left_right` with |
| // a `phi_left_end` in `left_end` and merge `phi_left_end`, `iget_mid` and |
| // `iget_right` with the final `phi` in `return_block`. Use invokes to force |
| // spilling `iget_*`s; in `left_right` use `HMin` instead to avoid blocking |
| // registers too early. When the `iget_start` is spilled, the spill slot is |
| // propagated to the Phis as a hint that's seen by all the other `iget_*`s. |
| // |
| // For `iget_start`, add a register use in `left`, an `HDeoptimize` in |
| // `left_end` to extend the `iget_start` lifetime across `left_right` and |
| // most of `left_end` and an invoke in `left_end` before the `HDeoptimize` |
| // which blocks registers and forces the splitting of the `iget_start` |
| // interval when allocating a register for the use in `left`. |
| |
| HInstruction* obj = MakeParam(DataType::Type::kReference); |
| HInstruction* const1 = graph_->GetIntConstant(1); |
| |
| HInstruction* iget_start = MakeIFieldGet(start, obj, DataType::Type::kInt32, MemberOffset(40)); |
| HInstruction* switch_input = MakeIFieldGet(start, obj, DataType::Type::kInt32, MemberOffset(32)); |
| MakeInvokeStatic(start, DataType::Type::kVoid, {}); // Spill all. |
| start->AddInstruction( |
| new (GetAllocator()) HPackedSwitch(/*start_value=*/ 0, /*num_entries=*/ 2, switch_input)); |
| |
| HInstruction* iget_right = MakeIFieldGet(right, obj, DataType::Type::kInt32, MemberOffset(44)); |
| MakeInvokeStatic(right, DataType::Type::kVoid, {}); // Spill all. |
| MakeGoto(right); // This block was constructed with `AddNewBlock()` without `HGoto`. |
| |
| HInstruction* iget_mid = MakeIFieldGet(mid, obj, DataType::Type::kInt32, MemberOffset(48)); |
| MakeInvokeStatic(mid, DataType::Type::kVoid, {}); // Spill all. |
| |
| HSub* sub = MakeBinOp<HSub>(left, DataType::Type::kInt32, iget_start, const1); |
| HInstruction* left_cond = MakeIFieldGet(left, obj, DataType::Type::kBool, MemberOffset(36)); |
| MakeIf(left, left_cond); |
| |
| HInstruction* iget_left_right = |
| MakeIFieldGet(left_right, obj, DataType::Type::kInt32, MemberOffset(40)); |
| // Force spilling `iget_left_right` without using an invoke. |
| HMin* min = MakeBinOp<HMin>(left_right, DataType::Type::kInt32, sub, const1); |
| |
| HPhi* phi_left_end = MakePhi(left_end, {iget_start, iget_left_right}); |
| HPhi* phi_min = MakePhi(left_end, {sub, min}); |
| // Block registers to cause splitting of the `iget_start` interval when allocating |
| // a register at the start of the `left` block. |
| MakeInvokeStatic(left_end, DataType::Type::kVoid, {phi_min}); // Use `phi_min`, spill all. |
| HInstruction* deopt_cond = MakeIFieldGet(left_end, obj, DataType::Type::kBool, MemberOffset(37)); |
| HDeoptimize* deopt = new (GetAllocator()) HDeoptimize( |
| GetAllocator(), deopt_cond, DeoptimizationKind::kFullFrame, kNoDexPc); |
| AddOrInsertInstruction(left_end, deopt); |
| ManuallyBuildEnvFor(deopt, {iget_start}); // Keep `iget_start` alive up to this point. |
| |
| HPhi* phi = MakePhi(return_block, {phi_left_end, iget_mid, iget_right}); |
| MakeReturn(return_block, phi); |
| |
| graph_->ComputeDominanceInformation(); |
| std::unique_ptr<CodeGenerator> codegen = CreateCodeGenerator(graph_, InstructionSet::kX86); |
| if (codegen == nullptr) { |
| GTEST_SKIP() << "X86 codegen is unavailable."; |
| } |
| SsaLivenessAnalysis liveness(graph_, codegen.get(), GetScopedAllocator()); |
| liveness.Analyze(); |
| |
| ASSERT_LT(right->GetLifetimeStart(), mid->GetLifetimeStart()); |
| ASSERT_LT(mid->GetLifetimeStart(), left->GetLifetimeStart()); |
| |
| // Set just two register available, including the `obj` input register ECX. |
| BlockCoreRegistersExcept(codegen.get(), {x86::EAX, x86::ECX}); |
| |
| // Before the bug was fixed, the interval validation at the end of `AllocateRegisters()` |
| // would report a spill slot conflict for the `iget_left_right` because it was allocated |
| // the same spill slot as `iget_start` despite overlapping lifetime. This would cause |
| // clobbering of `iget_start` for its use in the `HDeoptimize` environment. |
| std::unique_ptr<RegisterAllocator> register_allocator = |
| RegisterAllocator::Create(GetScopedAllocator(), codegen.get(), liveness); |
| register_allocator->AllocateRegisters(); |
| |
| // While `iget_right` and `iget_mid` use the same spill slot as `iget_start` thanks to the |
| // hint, the `iget_left_right` must use a different spill slot due to lifetime overlap. |
| for (HInstruction* iget : {iget_start, iget_right, iget_mid, iget_left_right}) { |
| ASSERT_TRUE(iget->GetLiveInterval()->HasSpillSlot()); |
| } |
| int iget_start_spill_slot = iget_start->GetLiveInterval()->GetSpillSlot(); |
| EXPECT_EQ(iget_start_spill_slot, iget_right->GetLiveInterval()->GetSpillSlot()); |
| EXPECT_EQ(iget_start_spill_slot, iget_mid->GetLiveInterval()->GetSpillSlot()); |
| EXPECT_NE(iget_start_spill_slot, iget_left_right->GetLiveInterval()->GetSpillSlot()); |
| } |
| |
| void RegisterAllocatorTest::TestNoOutputOverlap() { |
| if (!com::android::art::flags::reg_alloc_no_output_overlap()) { |
| GTEST_SKIP() << "Improved `Location::kNoOutputOverlap` handling disabled."; |
| } |
| HBasicBlock* block = InitEntryMainExitGraph(); |
| HInstruction* const1 = graph_->GetIntConstant(1); |
| HNeg* neg = MakeUnOp<HNeg>(block, DataType::Type::kInt32, const1); |
| HAdd* add = MakeBinOp<HAdd>(block, DataType::Type::kInt32, neg, const1); |
| // Add another use of `neg` after `add`. Previously this would have prevented |
| // using the register allocated for `neg` for the `add` output. |
| HAdd* add2 = MakeBinOp<HAdd>(block, DataType::Type::kInt32, neg, const1); |
| // Insert `HNop` to test the end of the range of the unused instruction `add2`. |
| AddOrInsertInstruction( |
| block, new (GetAllocator()) HNop(/*dex_pc=*/ 0u, /*needs_environment=*/ false)); |
| HReturn* ret = MakeReturn(block, add); |
| |
| graph_->ComputeDominanceInformation(); |
| std::unique_ptr<CodeGenerator> codegen = CreateCodeGenerator(graph_, InstructionSet::kX86); |
| if (codegen == nullptr) { |
| GTEST_SKIP() << "X86 codegen is unavailable."; |
| } |
| SsaLivenessAnalysis liveness(graph_, codegen.get(), GetScopedAllocator()); |
| liveness.Analyze(); |
| |
| ASSERT_TRUE(neg->GetLocations()->OutputCanOverlapWithInputs()); |
| ASSERT_FALSE(add->GetLocations()->OutputCanOverlapWithInputs()); |
| ASSERT_FALSE(add2->GetLocations()->OutputCanOverlapWithInputs()); |
| |
| // Set just one register available to make all intervals compete for the same. |
| BlockCoreRegistersExcept(codegen.get(), {x86::EAX}); |
| |
| RegisterAllocatorLinearScan register_allocator(GetScopedAllocator(), codegen.get(), liveness); |
| |
| register_allocator.AllocateRegistersInternal(); |
| |
| // Check that the register 0 was allocated at the right positions for `neg`, `add` and `add2`. |
| static constexpr int kExpectedRegister = 0; |
| LiveInterval* neg_li = neg->GetLiveInterval(); |
| ASSERT_EQ(neg->GetLifetimePosition(), neg_li->GetStart()); |
| ASSERT_EQ(1u << kExpectedRegister, neg_li->GetRegisters()); |
| LiveInterval* add_li = add->GetLiveInterval(); |
| ASSERT_EQ(add->GetLifetimePosition() + kLivenessPositionOfNonOverlappingOutput, |
| add_li->GetStart()); |
| ASSERT_EQ(1u << kExpectedRegister, add_li->GetRegisters()); |
| LiveInterval* add2_li = add2->GetLiveInterval(); |
| ASSERT_EQ(add2->GetLifetimePosition() + kLivenessPositionOfNonOverlappingOutput, |
| add2_li->GetStart()); |
| ASSERT_EQ(1u << kExpectedRegister, add2_li->GetRegisters()); |
| |
| // Check the splits of the `neg` interval. |
| ASSERT_EQ(add_li->GetStart(), neg_li->GetEnd()); |
| LiveInterval* neg_li2 = neg_li->GetNextSibling(); |
| ASSERT_TRUE(neg_li2 != nullptr); |
| ASSERT_EQ(add_li->GetStart(), neg_li2->GetStart()); |
| ASSERT_EQ(kNoRegisters, neg_li2->GetRegisters()); |
| ASSERT_EQ(add2->GetLifetimePosition(), neg_li2->GetEnd()); |
| LiveInterval* neg_li3 = neg_li2->GetNextSibling(); |
| ASSERT_TRUE(neg_li3 != nullptr); |
| ASSERT_EQ(add2->GetLifetimePosition(), neg_li3->GetStart()); |
| ASSERT_EQ(1u << kExpectedRegister, neg_li3->GetRegisters()); |
| ASSERT_EQ(add2->GetLifetimePosition() + kLivenessPositionOfNormalUse, neg_li3->GetEnd()); |
| ASSERT_TRUE(neg_li3->GetNextSibling() == nullptr); |
| |
| // Check the splits of the `add` interval. |
| ASSERT_EQ(add2->GetLifetimePosition(), add_li->GetEnd()); |
| LiveInterval* add_li2 = add_li->GetNextSibling(); |
| ASSERT_TRUE(add_li2 != nullptr); |
| ASSERT_EQ(add2->GetLifetimePosition(), add_li2->GetStart()); |
| ASSERT_EQ(kNoRegisters, add_li2->GetRegisters()); |
| ASSERT_EQ(ret->GetLifetimePosition(), add_li2->GetEnd()); |
| ASSERT_TRUE(add_li2->GetNextSibling() == nullptr); |
| |
| // Check that the interval for the unused `add2` was not split and ends after the `add2`. |
| ASSERT_TRUE(add2_li->GetNextSibling() == nullptr); |
| ASSERT_EQ(add2->GetLifetimePosition() + kLppi, add2_li->GetEnd()); |
| } |
| |
| TEST_F(RegisterAllocatorTest, NoOutputOverlap) { |
| TestNoOutputOverlap(); |
| } |
| |
| void RegisterAllocatorTest::TestNoOutputOverlapAndTemp() { |
| if (!com::android::art::flags::reg_alloc_no_output_overlap()) { |
| GTEST_SKIP() << "Improved `Location::kNoOutputOverlap` handling disabled."; |
| } |
| HBasicBlock* block = InitEntryMainExitGraph(); |
| HInstruction* const1 = graph_->GetIntConstant(1); |
| HNeg* neg = MakeUnOp<HNeg>(block, DataType::Type::kInt32, const1); |
| HAdd* add = MakeBinOp<HAdd>(block, DataType::Type::kInt32, neg, const1); |
| HSub* sub = MakeBinOp<HSub>(block, DataType::Type::kInt32, add, neg); |
| |
| graph_->ComputeDominanceInformation(); |
| std::unique_ptr<CodeGenerator> codegen = CreateCodeGenerator(graph_, InstructionSet::kX86); |
| if (codegen == nullptr) { |
| GTEST_SKIP() << "X86 codegen is unavailable."; |
| } |
| SsaLivenessAnalysis liveness(graph_, codegen.get(), GetScopedAllocator()); |
| liveness.Analyze(); |
| |
| ASSERT_TRUE(neg->GetLocations()->OutputCanOverlapWithInputs()); |
| ASSERT_FALSE(add->GetLocations()->OutputCanOverlapWithInputs()); |
| ASSERT_TRUE(sub->GetLocations()->OutputCanOverlapWithInputs()); |
| ASSERT_TRUE(sub->GetLocations()->InAt(1).Equals(Location::Any())); |
| |
| // Add a temp for `add`. We want to test that the temp interval shall not be split. |
| // Trying to split it would trigger a `DCHECK(!IsTemp())`. |
| ASSERT_EQ(0, add->GetLocations()->GetTempCount()); |
| add->GetLocations()->AddTemp(Location::RequiresCoreRegister()); |
| |
| // Set just two registers available to avoid adding more instructions |
| // to reproduce the situation where we could try to split the temp. |
| BlockCoreRegistersExcept(codegen.get(), {x86::EAX, x86::ECX}); |
| |
| RegisterAllocatorLinearScan register_allocator(GetScopedAllocator(), codegen.get(), liveness); |
| |
| register_allocator.AllocateRegistersInternal(); |
| |
| // Check the splits of the `neg` interval. |
| LiveInterval* neg_li = neg->GetLiveInterval(); |
| ASSERT_EQ(neg->GetLifetimePosition(), neg_li->GetStart()); |
| ASSERT_TRUE(neg_li->HasRegisters()); |
| ASSERT_EQ(add->GetLifetimePosition() + kLivenessPositionOfNonOverlappingOutput, neg_li->GetEnd()); |
| LiveInterval* neg_li2 = neg_li->GetNextSibling(); |
| ASSERT_TRUE(neg_li2 != nullptr); |
| ASSERT_EQ(neg_li->GetEnd(), neg_li2->GetStart()); |
| ASSERT_EQ(kNoRegisters, neg_li2->GetRegisters()); |
| ASSERT_EQ(sub->GetLifetimePosition() + kLivenessPositionOfNormalUse, neg_li2->GetEnd()); |
| ASSERT_TRUE(neg_li2->GetNextSibling() == nullptr); |
| } |
| |
| TEST_F(RegisterAllocatorTest, NoOutputOverlapAndTemp) { |
| TestNoOutputOverlapAndTemp(); |
| } |
| |
| TEST_F(RegisterAllocatorTest, NoOutputOverlapImmediateSpill) { |
| if (!com::android::art::flags::reg_alloc_no_output_overlap()) { |
| GTEST_SKIP() << "Improved `Location::kNoOutputOverlap` handling disabled."; |
| } |
| HBasicBlock* block = InitEntryMainExitGraph(); |
| HInstruction* param = MakeParam(DataType::Type::kReference); |
| HInvoke* invoke = MakeInvokeStatic(block, DataType::Type::kVoid, {}); // Spill `param`. |
| // After the `invoke`, all registers are free and the `param` input to `get1` is allocated |
| // register 0. With only registers 0-2 available, the only available pair is 0-1, so the |
| // `get1` with no-output-overlap is allocated that pair. Now we need to do something about |
| // the `param` but trying to allocate another register now is not easy. We cannot put moves |
| // in the middle of an instruction, so we'd have to split the interval at the start of the |
| // instruction and move back to that position, potentially bringing back already handled |
| // intervals for inputs that die at this instruction and somehow handle the output as live |
| // at the start of the instructions even though if its lifetime has not yet started. It's |
| // much easier to just force spilling active intervals when we need to take their register |
| // for the no-output-overlap output. |
| // |
| // This test case exposes a situation that led to a compiler crash during development when |
| // we didn't force the spill and ended up with two parallel moves for `param` at the start |
| // of `get1`, hitting a debug-mode check that prevents such move collisions. |
| HInstruction* get1 = MakeIFieldGet(block, param, DataType::Type::kInt64, MemberOffset(32)); |
| HInstruction* get2 = MakeIFieldGet(block, param, DataType::Type::kInt32, MemberOffset(40)); |
| HInstruction* conv = MakeUnOp<HTypeConversion>(block, DataType::Type::kInt32, get1); |
| HAdd* add = MakeBinOp<HAdd>(block, DataType::Type::kInt32, conv, get2); |
| MakeReturn(block, add); |
| |
| graph_->BuildDominatorTree(); |
| std::unique_ptr<CodeGenerator> codegen = CreateCodeGenerator(graph_, InstructionSet::kX86); |
| if (codegen == nullptr) { |
| GTEST_SKIP() << "X86 codegen is unavailable."; |
| } |
| SsaLivenessAnalysis liveness(graph_, codegen.get(), GetScopedAllocator()); |
| liveness.Analyze(); |
| |
| // There is no instruction that would get the locations required for this test on x86. |
| // We need no output overlap for an instruction that takes a single-register input and |
| // produces two-register output. This happens for 64-bit `HInstanceFieldGet` on arm32. |
| // Note: The processing of "no output overlap" happens at the start of the register |
| // allocation. It could conceivably be moved to the liveness analysis which would make |
| // this adjustment too late and we'd need to intercept the location creation which is |
| // doable as long as the `CodeGeneratorX86` is not `final`. |
| LocationSummary* get1_locs = get1->GetLocations(); |
| ASSERT_TRUE(get1_locs->OutputCanOverlapWithInputs()); |
| ASSERT_TRUE(get1_locs->Out().Equals(Location::RequiresCoreRegister())); |
| get1_locs->UpdateOut(Location()); // Invalidate output to work around `DCHECK()` in `SetOut()`. |
| get1_locs->SetOut(Location::RequiresCoreRegister(), Location::kNoOutputOverlap); |
| |
| // Make three registers available. |
| BlockCoreRegistersExcept(codegen.get(), {x86::EAX, x86::ECX, x86::EDX}); |
| |
| std::unique_ptr<RegisterAllocator> register_allocator = |
| RegisterAllocator::Create(GetScopedAllocator(), codegen.get(), liveness); |
| register_allocator->AllocateRegisters(); |
| |
| // Check the splits of the `param`. |
| LiveInterval* param_li = param->GetLiveInterval(); |
| ASSERT_TRUE(param_li != nullptr); |
| ASSERT_TRUE(param_li->HasSpillSlot()); |
| EXPECT_GT(block->GetLifetimeStart(), param_li->GetStart()); |
| EXPECT_EQ(block->GetLifetimeStart(), param_li->GetEnd()); |
| EXPECT_EQ(1u << 1, param_li->GetRegisters()); // Initial argument register. |
| LiveInterval* param_li2 = param_li->GetNextSibling(); |
| ASSERT_TRUE(param_li2 != nullptr); |
| EXPECT_EQ(block->GetLifetimeStart(), param_li2->GetStart()); |
| EXPECT_EQ(invoke->GetLifetimePosition(), param_li2->GetEnd()); |
| EXPECT_EQ(1u << 0, param_li2->GetRegisters()); // Moved to register 0. |
| LiveInterval* param_li3 = param_li2->GetNextSibling(); |
| ASSERT_TRUE(param_li3 != nullptr); |
| EXPECT_EQ(invoke->GetLifetimePosition(), param_li3->GetStart()); |
| EXPECT_EQ(get1->GetLifetimePosition(), param_li3->GetEnd()); |
| EXPECT_FALSE(param_li3->HasRegisters()); // No register. |
| LiveInterval* param_li4 = param_li3->GetNextSibling(); |
| ASSERT_TRUE(param_li4 != nullptr); |
| EXPECT_EQ(get1->GetLifetimePosition(), param_li4->GetStart()); |
| EXPECT_EQ(get1->GetLifetimePosition() + kLivenessPositionOfNonOverlappingOutput, |
| param_li4->GetEnd()); |
| EXPECT_EQ(1u << 0, param_li4->GetRegisters()); // Allocated register 0. |
| LiveInterval* param_li5 = param_li4->GetNextSibling(); |
| ASSERT_TRUE(param_li5 != nullptr); |
| EXPECT_EQ(get1->GetLifetimePosition() + kLivenessPositionOfNonOverlappingOutput, |
| param_li5->GetStart()); |
| EXPECT_EQ(get2->GetLifetimePosition(), param_li5->GetEnd()); |
| EXPECT_FALSE(param_li5->HasRegisters()); // No register. |
| LiveInterval* param_li6 = param_li5->GetNextSibling(); |
| ASSERT_TRUE(param_li6 != nullptr); |
| EXPECT_EQ(get2->GetLifetimePosition(), param_li6->GetStart()); |
| EXPECT_EQ(get2->GetLifetimePosition() + kLivenessPositionOfNormalUse, param_li6->GetEnd()); |
| EXPECT_EQ(1u << 2, param_li6->GetRegisters()); // Allocated reg 2, avoiding `get1` result in 0-1. |
| ASSERT_TRUE(param_li6->GetNextSibling() == nullptr); |
| } |
| |
| } // namespace art |