openharmony/arkcompiler_runtime_core
1
0
Fork 0
mirror of https://gitee.com/openharmony/arkcompiler_runtime_core synced 2024-10-06 21:34:16 +00:00
Code Issues Actions Packages Projects Releases Wiki Activity

Remove unused code

Browse Source 
Remove unused code

Issue: #I638QE
Signed-off-by: qiuyu <qiuyu22@huawei.com>
Change-Id: I33908e556a638806a5bb2a06e023ba449ba910ef
This commit is contained in:
qiuyu 2022-11-26 11:13:39 +08:00
parent 49c02e36ac
commit 2b028ad277
1170 changed files with 1445 additions and 240282 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

40
BUILD.gn
View File

@ -220,7 +220,6 @@ config("ark_config") {
}
plugins_yamls = []
runtime_options_yamls = []
foreach(plugin, enabled_plugins) {
plugin_dir = "$ark_root/plugins/$plugin"
source_files = read_file("$plugin_dir/subproject_sources.gn", "scope")
@ -229,24 +228,14 @@ foreach(plugin, enabled_plugins) {
plugins_yamls += [ "$plugin_dir/${source_files.option_yaml_path}" ]
}
if (defined(source_files.runtime_option_yaml_path)) {
runtime_options_yamls +=
[ "$plugin_dir/${source_files.runtime_option_yaml_path}" ]
}
source_files = {
}
}
entrypoints_yamls = []
foreach(plugin, enabled_plugins) {
plugin_dir = "$ark_root/plugins/$plugin"
source_files = read_file("$plugin_dir/subproject_sources.gn", "scope")
if (defined(source_files.entrypoints_yaml_path)) {
entrypoints_yamls += [ "$plugin_dir/${source_files.entrypoints_yaml_path}" ]
}
source_files = {
}
concat_yamls("concat_plugins_yamls") {
output_file = "$target_gen_dir/plugin_options.yaml"
default_file = "$ark_root/templates/plugin_options.yaml"
add_yamls = plugins_yamls
}
inst_templates_yamls = []
@ -261,29 +250,12 @@ foreach(plugin, enabled_plugins) {
}
}
concat_yamls("concat_plugins_yamls") {
output_file = "$target_gen_dir/plugin_options.yaml"
default_file = "$ark_root/templates/plugin_options.yaml"
add_yamls = plugins_yamls
}
concat_yamls("concat_entrypoints_yamls") {
output_file = "$target_gen_dir/runtime/entrypoints.yaml"
default_file = "$ark_root/runtime/entrypoints/entrypoints.yaml"
add_yamls = entrypoints_yamls
}
concat_yamls("concat_inst_templates_yamls") {
output_file = "$target_gen_dir/compiler/generated/inst_templates.yaml"
default_file = "$ark_root/compiler/optimizer/ir_builder/inst_templates.yaml"
add_yamls = inst_templates_yamls
}
merge_yamls("merge_runtime_options_yamls") {
output_file = "$target_gen_dir/runtime_options.yaml"
add_yamls = [ "$ark_root/runtime/options.yaml" ] + runtime_options_yamls
}
if (!ark_standalone_build) {
group("bcopt_type_adapter_unit_test") {
if (host_os == "mac") {
@ -321,8 +293,6 @@ if (!ark_standalone_build) {
"$ark_root/libpandabase/tests:unittest",
"$ark_root/libpandafile/tests:unittest",
"$ark_root/libziparchive/tests:unittest",
"$ark_root/plugins/ecmascript/tests:unittest",
"$ark_root/runtime/tests:unittest",
]
}
@ -332,8 +302,6 @@ if (!ark_standalone_build) {
"$ark_root/libpandabase/tests:host_unittest",
"$ark_root/libpandafile/tests:host_unittest",
"$ark_root/libziparchive/tests:host_unittest",
"$ark_root/plugins/ecmascript/tests:host_unittest",
"$ark_root/runtime/tests:host_unittest",
]
}
}

4
assembler/BUILD.gn
View File

@ -25,7 +25,6 @@ config("arkassembler_public_config") {
}
libarkassembler_sources = [
"$ark_root/plugins/ecmascript/assembler/extension/ecmascript_meta.cpp",
"$target_gen_dir/ins_to_string.cpp",
"annotation.cpp",
"assembly-emitter.cpp",
@ -34,6 +33,7 @@ libarkassembler_sources = [
"assembly-program.cpp",
"assembly-type.cpp",
"context.cpp",
"extensions/ecmascript_meta.cpp",
"extensions/extensions.cpp",
"lexer.cpp",
"meta.cpp",
@ -147,12 +147,10 @@ source_set("ark_asm_static") {
"$ark_root:ark_config",
"$ark_root/libpandabase:arkbase_public_config",
"$ark_root/libpandafile:arkfile_public_config",
"$ark_root/runtime:arkruntime_public_config",
]
deps = [
":libarkassembler_frontend_static",
"$ark_root/bytecode_optimizer:libarkbytecodeopt_frontend_static",
"$ark_root/libpandabase:libarkbase_frontend_static",
"$ark_root/libpandafile:libarkfile_frontend_static",
]

0
plugins/ecmascript/assembler/extension/ecmascript_meta.cpp → assembler/extensions/ecmascript_meta.cpp
View File

0
plugins/ecmascript/assembler/extension/ecmascript_meta.h → assembler/extensions/ecmascript_meta.h
View File

2
assembler/extensions/register_extensions.h.erb
View File

@ -17,7 +17,7 @@
#define PANDA_ASSEMBLER_EXTENSIONS_REGISTER_EXTENSIONS_H
#include "extensions/extensions.h"
#include "extension/ecmascript_meta.h"
#include "extensions/ecmascript_meta.h"
namespace panda::pandasm::extensions {
std::unique_ptr<panda::pandasm::RecordMetadata> MetadataExtension::CreateRecordMetadata(panda::panda_file::SourceLang lang)

17
assembler/pandasm.cpp
View File

@ -25,9 +25,6 @@
#include "assembly-emitter.h"
#include "assembly-parser.h"
#ifdef PANDA_WITH_BYTECODE_OPTIMIZER
#include "bytecode_optimizer/optimize_bytecode.h"
#endif
#include "file_format_version.h"
#include "error.h"
#include "lexer.h"
@ -174,20 +171,6 @@ bool EmitProgramInBinary(panda::pandasm::Program &program, panda::PandArgParser
return false;
}
#ifdef PANDA_WITH_BYTECODE_OPTIMIZER
if (optimize.GetValue()) {
bool is_optimized = panda::bytecodeopt::OptimizeBytecode(&program, mapsp, output_file.GetValue());
if (!panda::pandasm::AsmEmitter::Emit(output_file.GetValue(), program, statp, mapsp, emit_debug_info)) {
std::cerr << "Failed to emit binary data: " << panda::pandasm::AsmEmitter::GetLastError() << std::endl;
return false;
}
if (!is_optimized) {
std::cerr << "Bytecode optimizer reported internal errors" << std::endl;
return false;
}
}
#endif
if (size_stat.GetValue()) {
size_t total_size = 0;
std::cout << "Panda file size statistic:" << std::endl;

1
bytecode_optimizer/BUILD.gn
View File

@ -39,7 +39,6 @@ libarkbytecodeopt_configs = [
"$ark_root/libpandabase:arkbase_public_config",
"$ark_root/libpandafile:arkfile_public_config",
"$ark_root/assembler:arkassembler_public_config",
"$ark_root/runtime:arkruntime_public_config",
]
ohos_shared_library("libarkbytecodeopt") {

8
bytecode_optimizer/bytecode_optimizer_isapi.rb
View File

@ -228,10 +228,6 @@ def if_fcmpg?
'static_cast<int>(inst->IsFcmpg())'
end
def if_inci?
"static_cast<int>(CanConvertToIncI(inst))"
end
# Operand printers
def dst_r
'inst->GetDstReg()'
@ -313,9 +309,7 @@ end
def call_me_from_template
# Empty visitors for IR instructions we want to ignore
# (Add missing IRs on demand)
%w[NullCheck BoundsCheck ZeroCheck NegativeCheck SafePoint
InitClass SaveStateDeoptimize RefTypeCheck Phi
Try SaveState LoadClass LoadAndInitClass Parameter].each do |op|
%w[Phi Try SaveState Parameter].each do |op|
visit(op) do
empty
end

112
bytecode_optimizer/codegen.cpp
View File

@ -14,8 +14,8 @@
*/
#include "codegen.h"
#include "common.h"
#include "runtime/include/coretypes/tagged_value.h"
#include "generate_ecma.inl"
#include "tagged_value.h"
namespace panda::bytecodeopt {
@ -251,11 +251,8 @@ void BytecodeGen::EncodeSpillFillData(const compiler::SpillFillData &sf)
}
pandasm::Ins move;
if (GetGraph()->IsDynamicMethod()) {
result_.emplace_back(pandasm::Create_MOV(sf.DstValue(), sf.SrcValue()));
return;
}
UNREACHABLE();
result_.emplace_back(pandasm::Create_MOV(sf.DstValue(), sf.SrcValue()));
return;
}
void BytecodeGen::VisitSpillFill(GraphVisitor *visitor, Inst *inst)
@ -285,11 +282,9 @@ void BytecodeGen::VisitConstant(GraphVisitor *visitor, Inst *inst)
auto type = inst->GetType();
/* Do not emit unused code for Const -> CastValueToAnyType chains */
if (enc->GetGraph()->IsDynamicMethod()) {
if (!HasUserPredicate(inst,
[](Inst const *i) { return i->GetOpcode() != compiler::Opcode::CastValueToAnyType; })) {
return;
}
if (!HasUserPredicate(inst,
[](Inst const *i) { return i->GetOpcode() != compiler::Opcode::CastValueToAnyType; })) {
return;
}
pandasm::Ins movi;
@ -297,42 +292,17 @@ void BytecodeGen::VisitConstant(GraphVisitor *visitor, Inst *inst)
switch (type) {
case compiler::DataType::INT64:
case compiler::DataType::UINT64:
if (enc->GetGraph()->IsDynamicMethod()) {
enc->result_.emplace_back(pandasm::Create_LDAI(inst->CastToConstant()->GetInt64Value()));
DoSta(inst->GetDstReg(), enc->result_);
} else {
UNREACHABLE();
}
enc->result_.emplace_back(pandasm::Create_LDAI(inst->CastToConstant()->GetInt64Value()));
DoSta(inst->GetDstReg(), enc->result_);
break;
case compiler::DataType::FLOAT64:
if (enc->GetGraph()->IsDynamicMethod()) {
enc->result_.emplace_back(pandasm::Create_FLDAI(inst->CastToConstant()->GetDoubleValue()));
DoSta(inst->GetDstReg(), enc->result_);
} else {
UNREACHABLE();
}
enc->result_.emplace_back(pandasm::Create_FLDAI(inst->CastToConstant()->GetDoubleValue()));
DoSta(inst->GetDstReg(), enc->result_);
break;
case compiler::DataType::BOOL:
case compiler::DataType::INT8:
case compiler::DataType::UINT8:
case compiler::DataType::INT16:
case compiler::DataType::UINT16:
case compiler::DataType::INT32:
case compiler::DataType::UINT32:
if (enc->GetGraph()->IsDynamicMethod()) {
enc->result_.emplace_back(pandasm::Create_LDAI(inst->CastToConstant()->GetInt32Value()));
DoSta(inst->GetDstReg(), enc->result_);
} else {
UNREACHABLE();
}
break;
case compiler::DataType::FLOAT32:
if (enc->GetGraph()->IsDynamicMethod()) {
enc->result_.emplace_back(pandasm::Create_FLDAI(inst->CastToConstant()->GetFloatValue()));
DoSta(inst->GetDstReg(), enc->result_);
} else {
UNREACHABLE();
}
enc->result_.emplace_back(pandasm::Create_LDAI(inst->CastToConstant()->GetInt32Value()));
DoSta(inst->GetDstReg(), enc->result_);
break;
default:
UNREACHABLE();
@ -367,21 +337,16 @@ void BytecodeGen::VisitIf(GraphVisitor *v, Inst *inst_base)
auto inst = inst_base->CastToIf();
switch (inst->GetInputType(0)) {
case compiler::DataType::ANY: {
if (enc->GetGraph()->IsDynamicMethod()) {
#if defined(ENABLE_BYTECODE_OPT) && defined(PANDA_WITH_ECMASCRIPT) && defined(ARK_INTRINSIC_SET)
IfEcma(v, inst);
break;
#endif
}
LOG(ERROR, BYTECODE_OPTIMIZER)
<< "Codegen for " << compiler::GetOpcodeString(inst->GetOpcode()) << " failed";
enc->success_ = false;
IfEcma(v, inst);
break;
#endif
}
default:
LOG(ERROR, BYTECODE_OPTIMIZER)
<< "Codegen for " << compiler::GetOpcodeString(inst->GetOpcode()) << " failed";
enc->success_ = false;
break;
}
}
@ -476,14 +441,12 @@ void BytecodeGen::VisitIfImm(GraphVisitor *v, Inst *inst_base)
IfImmZero(v, inst_base);
return;
}
IfImmNonZero(v, inst_base);
}
void BytecodeGen::IfImmZero(GraphVisitor *v, Inst *inst_base)
{
auto enc = static_cast<BytecodeGen *>(v);
auto inst = inst_base->CastToIfImm();
ASSERT(enc->GetGraph()->IsDynamicMethod());
DoLda(inst->GetSrcReg(0), enc->result_);
auto label = LabelName(inst->GetBasicBlock()->GetTrueSuccessor()->GetId());
switch (inst->GetCc()) {
@ -498,45 +461,20 @@ void BytecodeGen::IfImmZero(GraphVisitor *v, Inst *inst_base)
}
}
// NOLINTNEXTLINE(readability-function-size)
void BytecodeGen::IfImmNonZero(GraphVisitor *v, Inst *inst_base)
{
UNREACHABLE();
}
// NOLINTNEXTLINE(readability-function-size)
void BytecodeGen::IfImm64(GraphVisitor *v, Inst *inst_base)
{
UNREACHABLE();
}
// NOLINTNEXTLINE(readability-function-size)
void BytecodeGen::VisitCast(GraphVisitor *v, Inst *inst_base)
{
UNREACHABLE();
}
void BytecodeGen::VisitLoadString(GraphVisitor *v, Inst *inst_base)
{
pandasm::Ins ins;
auto enc = static_cast<BytecodeGen *>(v);
auto inst = inst_base->CastToLoadString();
/* Do not emit unused code for Str -> CastValueToAnyType chains */
if (enc->GetGraph()->IsDynamicMethod()) {
if (!HasUserPredicate(inst,
[](Inst const *i) { return i->GetOpcode() != compiler::Opcode::CastValueToAnyType; })) {
return;
}
if (!HasUserPredicate(inst,
[](Inst const *i) { return i->GetOpcode() != compiler::Opcode::CastValueToAnyType; })) {
return;
}
enc->result_.emplace_back(pandasm::Create_LDA_STR(enc->ir_interface_->GetStringIdByOffset(inst->GetTypeId())));
if (inst->GetDstReg() != compiler::ACC_REG_ID) {
if (enc->GetGraph()->IsDynamicMethod()) {
enc->result_.emplace_back(pandasm::Create_STA(inst->GetDstReg()));
} else {
UNREACHABLE();
}
enc->result_.emplace_back(pandasm::Create_STA(inst->GetDstReg()));
}
}
@ -628,17 +566,7 @@ void BytecodeGen::VisitCastValueToAnyType([[maybe_unused]] GraphVisitor *v, [[ma
void BytecodeGen::VisitIntrinsic(GraphVisitor *visitor, Inst *inst_base)
{
ASSERT(inst_base->IsIntrinsic());
auto inst = inst_base->CastToIntrinsic();
auto enc = static_cast<BytecodeGen *>(visitor);
if (!enc->GetGraph()->IsDynamicMethod()) {
LOG(ERROR, BYTECODE_OPTIMIZER) << "Codegen for " << compiler::GetOpcodeString(inst->GetOpcode()) << " failed";
enc->success_ = false;
} else {
#ifdef ENABLE_BYTECODE_OPT
VisitEcma(visitor, inst_base);
#endif
}
VisitEcma(visitor, inst_base);
}
void BytecodeGen::VisitCatchPhi(GraphVisitor *v, Inst *inst)

3
bytecode_optimizer/codegen.h
View File

@ -106,10 +106,7 @@ public:
static void VisitIf(GraphVisitor *v, Inst *inst_base);
static void VisitIfImm(GraphVisitor *v, Inst *inst_base);
static void VisitCast(GraphVisitor *v, Inst *inst_base);
static void IfImmZero(GraphVisitor *v, Inst *inst_base);
static void IfImmNonZero(GraphVisitor *v, Inst *inst_base);
static void IfImm64(GraphVisitor *v, Inst *inst_base);
static void VisitIntrinsic(GraphVisitor *v, Inst *inst_base);
static void VisitLoadString(GraphVisitor *v, Inst *inst_base);
static void VisitReturn(GraphVisitor *v, Inst *inst_base);

72
bytecode_optimizer/common.cpp
View File

@ -21,73 +21,11 @@ namespace panda::bytecodeopt {
uint8_t AccReadIndex(const compiler::Inst *inst)
{
// For calls we cannot tell static index for acc position, thus
// ensure that we don't invoke this for calls
ASSERT(!inst->IsCall());
switch (inst->GetOpcode()) {
case compiler::Opcode::LoadArray:
case compiler::Opcode::StoreObject:
case compiler::Opcode::StoreStatic:
case compiler::Opcode::NewArray:
return 1U;
case compiler::Opcode::StoreArray:
return 2U;
default: {
if (inst->IsIntrinsic() && inst->IsAccRead()) {
ASSERT(inst->GetBasicBlock()->GetGraph()->IsDynamicMethod());
ASSERT(inst->GetInputsCount() >= 2U);
return inst->GetInputsCount() - 2U;
}
return 0;
}
if (inst->IsIntrinsic() && inst->IsAccRead()) {
ASSERT(inst->GetBasicBlock()->GetGraph()->IsDynamicMethod());
ASSERT(inst->GetInputsCount() >= 2U);
return inst->GetInputsCount() - 2U;
}
return 0;
}
// This method is used by bytecode optimizer's codegen.
bool CanConvertToIncI(const compiler::BinaryImmOperation *binop)
{
ASSERT(binop->GetBasicBlock()->GetGraph()->IsRegAllocApplied());
ASSERT(binop->GetOpcode() == compiler::Opcode::AddI || binop->GetOpcode() == compiler::Opcode::SubI);
// IncI works on the same register.
if (binop->GetSrcReg(0) != binop->GetDstReg()) {
return false;
}
// IncI cannot write accumulator.
if (binop->GetSrcReg(0) == compiler::ACC_REG_ID) {
return false;
}
// IncI users cannot read from accumulator.
// While Addi/SubI stores the output in accumulator, IncI works directly on registers.
for (const auto &user : binop->GetUsers()) {
const auto *uinst = user.GetInst();
if (uinst->IsCall()) {
continue;
}
const uint8_t index = AccReadIndex(uinst);
if (uinst->GetInput(index).GetInst() == binop && uinst->GetSrcReg(index) == compiler::ACC_REG_ID) {
return false;
}
}
constexpr uint64_t bitmask = 0xffffffff;
// Define min and max values of i4 type.
constexpr int32_t min = -8;
constexpr int32_t max = 7;
int32_t imm = binop->GetImm() & bitmask;
// Note: subi 3 is the same as inci v2, -3.
if (binop->GetOpcode() == compiler::Opcode::SubI) {
imm = -imm;
}
// IncI works only with 4 bits immediates.
return imm >= min && imm <= max;
}
} // namespace panda::bytecodeopt

5
bytecode_optimizer/common.h
View File

@ -19,10 +19,6 @@
#include "compiler/optimizer/ir/constants.h"
#include "compiler/optimizer/ir/inst.h"
namespace panda::compiler {
class BinaryImmOperation;
} // namespace panda::compiler
namespace panda::bytecodeopt {
static constexpr compiler::Register MIN_REGISTER_NUMBER = 0;
static constexpr compiler::Register MAX_NUM_SHORT_CALL_ARGS = 2;
@ -33,7 +29,6 @@ static constexpr panda::compiler::Register NUM_COMPACTLY_ENCODED_REGS = 16;
// Get the position where accumulator read happens.
uint8_t AccReadIndex(const compiler::Inst *inst);
bool CanConvertToIncI(const compiler::BinaryImmOperation *binop);
} // namespace panda::bytecodeopt
#endif // PANDA_BYTECODE_OPTIMIZER_COMMON_H

8
bytecode_optimizer/optimize_bytecode.cpp
View File

@ -45,17 +45,17 @@ namespace panda::bytecodeopt {
panda::bytecodeopt::Options options("");
template <typename T>
constexpr void RunOpts(compiler::Graph *graph, [[maybe_unused]] BytecodeOptIrInterface *iface)
constexpr void RunOpts(compiler::Graph *graph)
{
graph->RunPass<compiler::Cleanup>();
graph->RunPass<T>();
}
template <typename First, typename Second, typename... Rest>
constexpr void RunOpts(compiler::Graph *graph, BytecodeOptIrInterface *iface = nullptr)
constexpr void RunOpts(compiler::Graph *graph)
{
RunOpts<First>(graph, iface);
RunOpts<Second, Rest...>(graph, iface);
RunOpts<First>(graph);
RunOpts<Second, Rest...>(graph);
}
bool RunOptimizations(compiler::Graph *graph, BytecodeOptIrInterface *iface)

41
bytecode_optimizer/reg_acc_alloc.cpp
View File

@ -88,14 +88,6 @@ bool RegAccAlloc::IsAccRead(compiler::Inst *inst) const
return UNLIKELY(inst->IsPhi()) ? IsPhiOptimizable(inst) : inst->IsAccRead();
}
bool UserNeedSwapInputs(compiler::Inst *inst, compiler::Inst *user)
{
if (!user->IsCommutative()) {
return false;
}
return user->GetInput(AccReadIndex(user)).GetInst() != inst;
}
/**
* Return true if instruction can write the accumulator.
*/
@ -164,21 +156,13 @@ bool RegAccAlloc::IsPhiAccReady(compiler::Inst *phi) const
}
}
std::unordered_set<compiler::Inst *> users_that_required_swap_inputs;
for (auto &user : phi->GetUsers()) {
compiler::Inst *uinst = user.GetInst();
if (!CanUserReadAcc(phi, uinst)) {
return false;
}
if (UserNeedSwapInputs(phi, uinst)) {
users_that_required_swap_inputs.insert(uinst);
}
}
for (auto uinst : users_that_required_swap_inputs) {
uinst->SwapInputs();
}
return true;
}
@ -223,25 +207,13 @@ bool RegAccAlloc::RunImpl()
}
for (size_t i = 0; i < inst->GetInputsCount(); ++i) {
inst->SetSrcReg(i, compiler::INVALID_REG);
if ((inst->GetOpcode() == compiler::Opcode::LoadObject) || (inst->IsConst())) {
if (inst->IsConst()) {
inst->SetDstReg(compiler::INVALID_REG);
}
}
}
}
// Drop the pass if the function contains unsupported opcodes
// TODO(rtakacs): support these opcodes.
if (!GetGraph()->IsDynamicMethod()) {
for (auto block : GetGraph()->GetBlocksRPO()) {
for (auto inst : block->AllInsts()) {
if (inst->GetOpcode() == compiler::Opcode::Builtin) {
return false;
}
}
}
}
// Mark Phi instructions if they can be optimized for acc.
for (auto block : GetGraph()->GetBlocksRPO()) {
for (auto phi : block->PhiInsts()) {
@ -260,28 +232,21 @@ bool RegAccAlloc::RunImpl()
bool use_acc_dst_reg = true;
std::unordered_set<compiler::Inst *> users_that_required_swap_inputs;
for (auto &user : inst->GetUsers()) {
compiler::Inst *uinst = user.GetInst();
if (uinst->IsSaveState()) {
continue;
}
if (CanUserReadAcc(inst, uinst)) {
if (UserNeedSwapInputs(inst, uinst)) {
users_that_required_swap_inputs.insert(uinst);
}
SetNeedLda(uinst, false);
} else {
use_acc_dst_reg = false;
}
}
for (auto uinst : users_that_required_swap_inputs) {
uinst->SwapInputs();
}
if (use_acc_dst_reg) {
inst->SetDstReg(compiler::ACC_REG_ID);
} else if ((inst->GetOpcode() == compiler::Opcode::LoadObject) || inst->IsConst()) {
} else if (inst->IsConst()) {
inst->ClearFlag(compiler::inst_flags::ACC_WRITE);
for (auto &user : inst->GetUsers()) {
compiler::Inst *uinst = user.GetInst();
@ -308,7 +273,7 @@ bool RegAccAlloc::RunImpl()
input->SetDstReg(compiler::INVALID_REG);
SetNeedLda(inst, true);
if ((input->GetOpcode() == compiler::Opcode::LoadObject) || (input->IsConst())) {
if (input->IsConst()) {
input->ClearFlag(compiler::inst_flags::ACC_WRITE);
for (auto &user : input->GetUsers()) {
compiler::Inst *uinst = user.GetInst();

4
bytecode_optimizer/reg_encoder.cpp
View File

@ -342,7 +342,7 @@ static bool IsAccReadPosition(compiler::Inst *inst, size_t pos)
void RegEncoder::InsertSpillsForDynInputsInst(compiler::Inst *inst)
{
ASSERT(state_ == RegEncoderState::INSERT_SPILLS);
ASSERT(inst->IsStaticCall() || inst->IsVirtualCall() || inst->IsInitObject() || inst->IsIntrinsic());
ASSERT(inst->IsIntrinsic());
RegContentMap spill_map(GetGraph()->GetLocalAllocator()->Adapter()); // src -> (dst, src_type), non-callrange
RegContentVec spill_vec(GetGraph()->GetLocalAllocator()->Adapter()); // spill_vec is used to handle callrange
@ -435,7 +435,7 @@ void RegEncoder::CalculateNumNeededTempsForInst(compiler::Inst *inst)
if (IsIntrinsicRange(inst)) {
return;
}
ASSERT(inst->IsStaticCall() || inst->IsVirtualCall() || inst->IsInitObject() || inst->IsIntrinsic());
ASSERT(inst->IsIntrinsic());
auto nargs = inst->GetInputsCount() - (inst->RequireState() ? 1 : 0);
size_t start = 0;

7
bytecode_optimizer/reg_encoder.h
View File

@ -127,12 +127,7 @@ public:
#include "generated/check_width.h"
void VisitDefault(Inst *inst) override
{
LOG(ERROR, BYTECODE_OPTIMIZER) << "Opcode " << compiler::GetOpcodeString(inst->GetOpcode())
<< " not yet implemented in RegEncoder";
success_ = false;
}
void VisitDefault(Inst *inst) override {}
#include "compiler/optimizer/ir/visitor.inc"

225
bytecode_optimizer/runtime_adapter.h
View File

@ -41,69 +41,16 @@ public:
return const_cast<panda_file::File *>(&panda_file_);
}
MethodId ResolveMethodIndex(MethodPtr parent_method, MethodIndex index) const override
{
return panda_file_.ResolveMethodIndex(MethodCast(parent_method), index).GetOffset();
}
uint32_t ResolveOffsetByIndex(MethodPtr parent_method, uint16_t index) const override
{
return panda_file_.ResolveOffsetByIndex(MethodCast(parent_method), index).GetOffset();
}
FieldId ResolveFieldIndex(MethodPtr parent_method, FieldIndex index) const override
{
return panda_file_.ResolveFieldIndex(MethodCast(parent_method), index).GetOffset();
}
IdType ResolveTypeIndex(MethodPtr parent_method, TypeIndex index) const override
{
return panda_file_.ResolveClassIndex(MethodCast(parent_method), index).GetOffset();
}
MethodPtr GetMethodById([[maybe_unused]] MethodPtr caller, MethodId id) const override
{
return reinterpret_cast<MethodPtr>(id);
}
MethodId GetMethodId(MethodPtr method) const override
{
return static_cast<MethodId>(reinterpret_cast<uintptr_t>(method));
}
compiler::DataType::Type GetMethodReturnType(MethodPtr method) const override
{
panda_file::MethodDataAccessor mda(panda_file_, MethodCast(method));
panda_file::ProtoDataAccessor pda(panda_file_, mda.GetProtoId());
return ToCompilerType(panda_file::GetEffectiveType(pda.GetReturnType()));
}
compiler::DataType::Type GetMethodTotalArgumentType(MethodPtr method, size_t index) const override
{
panda_file::MethodDataAccessor mda(panda_file_, MethodCast(method));
if (!mda.IsStatic()) {
if (index == 0) {
return ToCompilerType(
panda_file::GetEffectiveType(panda_file::Type(panda_file::Type::TypeId::REFERENCE)));
}
--index;
}
panda_file::ProtoDataAccessor pda(panda_file_, mda.GetProtoId());
return ToCompilerType(panda_file::GetEffectiveType(pda.GetArgType(index)));
}
compiler::DataType::Type GetMethodArgumentType([[maybe_unused]] MethodPtr caller, MethodId id,
size_t index) const override
{
panda_file::MethodDataAccessor mda(panda_file_, panda_file::File::EntityId(id));
panda_file::ProtoDataAccessor pda(panda_file_, mda.GetProtoId());
return ToCompilerType(panda_file::GetEffectiveType(pda.GetArgType(index)));
}
size_t GetMethodTotalArgumentsCount(MethodPtr method) const override
{
panda_file::MethodDataAccessor mda(panda_file_, MethodCast(method));
@ -122,11 +69,6 @@ public:
return pda.GetNumArgs();
}
compiler::DataType::Type GetMethodReturnType(MethodPtr caller, MethodId id) const override
{
return GetMethodReturnType(GetMethodById(caller, id));
}
size_t GetMethodRegistersCount(MethodPtr method) const override
{
panda_file::MethodDataAccessor mda(panda_file_, MethodCast(method));
@ -169,20 +111,6 @@ public:
return static_cast<compiler::SourceLanguage>(source_lang.value());
}
size_t GetClassIdForField([[maybe_unused]] MethodPtr method, size_t field_id) const override
{
panda_file::FieldDataAccessor fda(panda_file_, panda_file::File::EntityId(field_id));
return static_cast<size_t>(fda.GetClassId().GetOffset());
}
ClassPtr GetClassForField(FieldPtr field) const override
{
panda_file::FieldDataAccessor fda(panda_file_, FieldCast(field));
return reinterpret_cast<ClassPtr>(fda.GetClassId().GetOffset());
}
size_t GetClassIdForMethod(MethodPtr method) const override
{
panda_file::MethodDataAccessor mda(panda_file_, MethodCast(method));
@ -190,50 +118,6 @@ public:
return static_cast<size_t>(mda.GetClassId().GetOffset());
}
size_t GetClassIdForMethod([[maybe_unused]] MethodPtr caller, size_t method_id) const override
{
panda_file::MethodDataAccessor mda(panda_file_, panda_file::File::EntityId(method_id));
return static_cast<size_t>(mda.GetClassId().GetOffset());
}
bool IsMethodExternal([[maybe_unused]] MethodPtr caller, MethodPtr callee) const override
{
panda_file::MethodDataAccessor mda(panda_file_, MethodCast(callee));
return mda.IsExternal();
}
bool IsMethodIntrinsic([[maybe_unused]] MethodPtr method) const override
{
return false;
}
bool IsMethodIntrinsic([[maybe_unused]] MethodPtr caller, [[maybe_unused]] MethodId id) const override
{
return false;
}
bool IsMethodStatic(MethodPtr method) const override
{
panda_file::MethodDataAccessor mda(panda_file_, MethodCast(method));
return mda.IsStatic();
}
bool IsMethodStatic([[maybe_unused]] MethodPtr caller, MethodId id) const override
{
panda_file::MethodDataAccessor mda(panda_file_, panda_file::File::EntityId(id));
return mda.IsStatic();
}
// return true if the method is Jni with exception
bool HasNativeException([[maybe_unused]] MethodPtr method) const override
{
return false;
}
std::string GetClassNameFromMethod(MethodPtr method) const override
{
panda_file::MethodDataAccessor mda(panda_file_, MethodCast(method));
@ -243,13 +127,6 @@ public:
return std::string(reinterpret_cast<const char *>(string_data.data));
}
std::string GetClassName(ClassPtr cls) const override
{
auto string_data = panda_file_.GetStringData(ClassCast(cls));
return std::string(reinterpret_cast<const char *>(string_data.data));
}
std::string GetMethodName(MethodPtr method) const override
{
panda_file::MethodDataAccessor mda(panda_file_, MethodCast(method));
@ -259,23 +136,6 @@ public:
return std::string(reinterpret_cast<const char *>(string_data.data));
}
bool IsConstructor(MethodPtr method, uint32_t class_id) override
{
if (GetClassIdForMethod(method) != class_id) {
return false;
}
panda_file::File::EntityId entity_id(class_id);
panda_file::SourceLang lang = panda_file::SourceLang::PANDA_ASSEMBLY;
if (!panda_file_.IsExternal(entity_id)) {
panda_file::ClassDataAccessor cda(panda_file_, entity_id);
lang = cda.GetSourceLang().value_or(lang);
}
return GetMethodName(method) == GetCtorName(lang);
}
std::string GetMethodFullName(MethodPtr method, bool /* with_signature */) const override
{
auto class_name = GetClassNameFromMethod(method);
@ -284,69 +144,6 @@ public:
return class_name + "::" + method_name;
}
ClassPtr GetClass(MethodPtr method) const override
{
panda_file::MethodDataAccessor mda(panda_file_, MethodCast(method));
return reinterpret_cast<ClassPtr>(mda.GetClassId().GetOffset());
}
std::string GetBytecodeString(MethodPtr method, uintptr_t pc) const override
{
// NOLINTNEXTLINE(cppcoreguidelines-pro-bounds-pointer-arithmetic)
BytecodeInstruction inst(GetMethodCode(method) + pc);
std::stringstream ss;
ss << inst;
return ss.str();
}
bool IsArrayClass([[maybe_unused]] MethodPtr method, IdType id) const override
{
panda_file::File::EntityId cid(id);
return panda_file::IsArrayDescriptor(panda_file_.GetStringData(cid).data);
}
FieldPtr ResolveField([[maybe_unused]] MethodPtr method, size_t id, [[maybe_unused]] bool allow_external,
uint32_t * /* class_id */) override
{
return reinterpret_cast<FieldPtr>(id);
}
compiler::DataType::Type GetFieldType(FieldPtr field) const override
{
panda_file::FieldDataAccessor fda(panda_file_, FieldCast(field));
return ToCompilerType(panda_file::Type::GetTypeFromFieldEncoding(fda.GetType()));
}
compiler::DataType::Type GetFieldTypeById([[maybe_unused]] MethodPtr parent_method, IdType id) const override
{
panda_file::FieldDataAccessor fda(panda_file_, panda_file::File::EntityId(id));
return ToCompilerType(panda_file::Type::GetTypeFromFieldEncoding(fda.GetType()));
}
bool IsFieldVolatile(FieldPtr field) const override
{
panda_file::FieldDataAccessor fda(panda_file_, FieldCast(field));
return fda.IsVolatile();
}
ClassPtr ResolveType([[maybe_unused]] MethodPtr method, size_t id) const override
{
return reinterpret_cast<ClassPtr>(id);
}
std::string GetFieldName(FieldPtr field) const override
{
panda_file::FieldDataAccessor fda(panda_file_, FieldCast(field));
auto string_data = panda_file_.GetStringData(fda.GetNameId());
return utf::Mutf8AsCString(string_data.data);
}
TypeInfoIndex GetTypeInfoIndexByInstId(size_t id) const override
{
const auto it = instid_type_map_.find(id);
@ -414,16 +211,6 @@ private:
switch (type.GetId()) {
case panda_file::Type::TypeId::VOID:
return compiler::DataType::VOID;
case panda_file::Type::TypeId::U1:
return compiler::DataType::BOOL;
case panda_file::Type::TypeId::I8:
return compiler::DataType::INT8;
case panda_file::Type::TypeId::U8:
return compiler::DataType::UINT8;
case panda_file::Type::TypeId::I16:
return compiler::DataType::INT16;
case panda_file::Type::TypeId::U16:
return compiler::DataType::UINT16;
case panda_file::Type::TypeId::I32:
return compiler::DataType::INT32;
case panda_file::Type::TypeId::U32:
@ -432,8 +219,6 @@ private:
return compiler::DataType::INT64;
case panda_file::Type::TypeId::U64:
return compiler::DataType::UINT64;
case panda_file::Type::TypeId::F32:
return compiler::DataType::FLOAT32;
case panda_file::Type::TypeId::F64:
return compiler::DataType::FLOAT64;
case panda_file::Type::TypeId::REFERENCE:
@ -452,16 +237,6 @@ private:
return panda_file::File::EntityId(reinterpret_cast<uintptr_t>(method));
}
static panda_file::File::EntityId ClassCast(RuntimeInterface::ClassPtr cls)
{
return panda_file::File::EntityId(reinterpret_cast<uintptr_t>(cls));
}
static panda_file::File::EntityId FieldCast(RuntimeInterface::FieldPtr field)
{
return panda_file::File::EntityId(reinterpret_cast<uintptr_t>(field));
}
const panda_file::File &panda_file_;
std::unordered_map<size_t, TypeInfoIndex> instid_type_map_;
std::unordered_map<int32_t, TypeInfoIndex> pc_type_map_;

192
runtime/include/coretypes/tagged_value.h → bytecode_optimizer/tagged_value.h
View File

@ -22,10 +22,6 @@
#include "libpandabase/mem/mem.h"
#include "utils/bit_utils.h"
namespace panda {
class ObjectHeader;
} // namespace panda
namespace panda::coretypes {
// Every double with all of its exponent bits set and its highest mantissa bit set is a quiet NaN.
@ -51,8 +47,6 @@ namespace panda::coretypes {
using TaggedType = uint64_t;
static const TaggedType NULL_POINTER = 0;
inline TaggedType ReinterpretDoubleToTaggedType(double value)
{
return bit_cast<TaggedType>(value);
@ -94,6 +88,7 @@ public:
TaggedValue(void *) = delete;
static const TaggedType NULL_POINTER = 0;
constexpr TaggedValue() : value_(NULL_POINTER) {}
constexpr explicit TaggedValue(TaggedType v) : value_(v) {}
@ -109,31 +104,6 @@ public:
value_ = TaggedValue(static_cast<int32_t>(v)).GetRawData();
}
static uint64_t GetIntTaggedValue(uint64_t v)
{
ASSERT(INT32_MIN <= static_cast<int32_t>(bit_cast<int64_t>(v)));
ASSERT(static_cast<int32_t>(bit_cast<int64_t>(v)) <= INT32_MAX);
return static_cast<uint32_t>(v) | TAG_INT;
}
static uint64_t GetDoubleTaggedValue(uint64_t v)
{
return v + DOUBLE_ENCODE_OFFSET;
}
static uint64_t GetBoolTaggedValue(uint64_t v)
{
ASSERT(v == 0 || v == 1);
return (v == 0) ? static_cast<uint64_t>(coretypes::TaggedValue::False().GetRawData())
: static_cast<uint64_t>(coretypes::TaggedValue::True().GetRawData());
}
static uint64_t GetObjectTaggedValue(uint64_t v)
{
ASSERT(static_cast<uint32_t>(v) == v);
return v;
}
explicit TaggedValue(int64_t v)
{
if (UNLIKELY(static_cast<int32_t>(v) != v)) {
@ -159,181 +129,21 @@ public:
explicit TaggedValue(const ObjectHeader *v) : value_(static_cast<TaggedType>(ToUintPtr(v))) {}
inline void CreateWeakRef()
{
ASSERT_PRINT(IsHeapObject() && ((value_ & TAG_WEAK_FILTER) == 0U),
"The least significant two bits of TaggedValue are not zero.");
value_ = value_ | TAG_WEAK_MASK;
}
inline void RemoveWeakTag()
{
ASSERT_PRINT(IsHeapObject() && ((value_ & TAG_WEAK_MASK) == 1U), "The tagged value is not a weak ref.");
value_ = value_ & (~TAG_WEAK_FILTER);
}
inline TaggedValue CreateAndGetWeakRef()
{
ASSERT_PRINT(IsHeapObject() && ((value_ & TAG_WEAK_FILTER) == 0U),
"The least significant two bits of TaggedValue are not zero.");
return TaggedValue(value_ | TAG_WEAK_MASK);
}
inline bool IsWeak() const
{
return IsHeapObject() && ((value_ & TAG_WEAK_MASK) == 1U);
}
inline bool IsDouble() const
{
return !IsInt() && !IsObject();
}
inline bool IsInt() const
{
return (value_ & TAG_MASK) == TAG_INT;
}
inline bool IsSpecial() const
{
return ((value_ & (~TAG_SPECIAL_MASK)) == 0U) && (((value_ & TAG_SPECIAL_VALUE) != 0U) || IsHole());
}
inline bool IsObject() const
{
return ((value_ & TAG_MASK) == TAG_OBJECT);
}
inline bool IsHeapObject() const
{
return IsObject() && !IsSpecial();
}
inline bool IsNumber() const
{
return !IsObject();
}
inline bool IsBoolean() const
{
return value_ == VALUE_FALSE || value_ == VALUE_TRUE;
}
inline double GetDouble() const
{
ASSERT_PRINT(IsDouble(), "can not convert TaggedValue to Double : " << std::hex << value_);
return ReinterpretTaggedTypeToDouble(value_ - DOUBLE_ENCODE_OFFSET);
}
inline int GetInt() const
{
ASSERT_PRINT(IsInt(), "can not convert TaggedValue to Int :" << std::hex << value_);
return static_cast<int>(value_ & (~TAG_MASK));
}
inline constexpr TaggedType GetRawData() const
{
return value_;
}
inline ObjectHeader *GetHeapObject() const
{
ASSERT_PRINT(IsHeapObject(), "can not convert TaggedValue to HeapObject :" << std::hex << value_);
// TODO(vpukhov): weakref ignored
// ASSERT_PRINT((value_ & TAG_WEAK_FILTER) == 0U,
// "can not convert TaggedValue to HeapObject :" << std::hex << value_);
return reinterpret_cast<ObjectHeader *>(value_ & (~TAG_WEAK_MASK));
}
// This function returns the heap object pointer which may have the weak tag.
inline ObjectHeader *GetRawHeapObject() const
{
ASSERT_PRINT(IsHeapObject(), "can not convert TaggedValue to HeapObject :" << std::hex << value_);
return reinterpret_cast<ObjectHeader *>(value_);
}
inline ObjectHeader *GetWeakReferent() const
{
ASSERT_PRINT(IsWeak(), "can not convert TaggedValue to WeakRef HeapObject :" << std::hex << value_);
return reinterpret_cast<ObjectHeader *>(value_ & (~TAG_WEAK_MASK));
}
static inline TaggedType Cast(void *ptr)
{
ASSERT_PRINT(sizeof(void *) == TaggedTypeSize(), "32bit platform is not support yet");
return static_cast<TaggedType>(ToUintPtr(ptr));
}
inline bool IsFalse() const
{
return value_ == VALUE_FALSE;
}
inline bool IsTrue() const
{
return value_ == VALUE_TRUE;
}
inline bool IsUndefined() const
{
return value_ == VALUE_UNDEFINED;
}
inline bool IsNull() const
{
return value_ == VALUE_NULL;
}
inline bool IsUndefinedOrNull() const
{
return IsNull() || IsUndefined();
}
inline bool IsHole() const
{
return value_ == VALUE_HOLE;
}
inline bool IsException() const
{
return value_ == VALUE_EXCEPTION;
}
static inline constexpr TaggedValue False()
{
return TaggedValue(VALUE_FALSE);
}
static inline constexpr TaggedValue True()
{
return TaggedValue(VALUE_TRUE);
}
static inline constexpr TaggedValue Undefined()
{
return TaggedValue(VALUE_UNDEFINED);
}
static inline constexpr TaggedValue Null()
{
return TaggedValue(VALUE_NULL);
}
static inline constexpr TaggedValue Hole()
{
return TaggedValue(VALUE_HOLE);
}
static inline constexpr TaggedValue Exception()
{
return TaggedValue(VALUE_EXCEPTION);
}
static inline constexpr size_t TaggedTypeSize()
{
return sizeof(TaggedType);
}
static inline bool IsImpureNaN(double value)
{
// Tests if the double value would break tagged double encoding.

5
bytecode_optimizer/templates/check_width.cpp.erb
View File

@ -50,11 +50,6 @@ void RegEncoder::VisitIfImm([[maybe_unused]] GraphVisitor* v, Inst* inst_base) {
re->success_ = false;
}
}
void RegEncoder::VisitCast([[maybe_unused]] GraphVisitor* v, Inst* inst_base) {
[[maybe_unused]] auto re = static_cast<RegEncoder*>(v);
[[maybe_unused]] auto inst = inst_base->CastToCast();
return;
}
% call_me_from_template
% visitors.each do |visitor|

1
bytecode_optimizer/tests/BUILD.gn
View File

@ -19,7 +19,6 @@ ohos_executable("bcopt_type_adapter_unit_test") {
configs = [
"$ark_root:ark_config",
"$ark_root/assembler:arkassembler_public_config",
"$ark_root/runtime:arkruntime_public_config",
"$ark_root/libpandabase:arkbase_public_config",
"$ark_root/libpandafile:arkfile_public_config",
"$ark_root/compiler:arkcompiler_public_config",

122
bytecode_optimizer/tests/bytecodeopt_peepholes_runtime_test.cpp
View File

@ -1,122 +0,0 @@
/**
* Copyright (c) 2021-2022 Huawei Device Co., Ltd.
* 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 <gtest/gtest.h>
#include "assembler/assembly-emitter.h"
#include "assembler/assembly-parser.h"
#include "bytecode_optimizer/optimize_bytecode.h"
#include "runtime/include/class_linker.h"
#include "runtime/include/runtime.h"
#include "mangling.h"
namespace panda::bytecodeopt::test {
class BytecodeOptPeepholes : public testing::Test {
public:
BytecodeOptPeepholes()
{
RuntimeOptions options;
options.SetHeapSizeLimit(128_MB);
options.SetShouldLoadBootPandaFiles(false);
options.SetShouldInitializeIntrinsics(false);
Logger::InitializeDummyLogging();
Runtime::Create(options);
thread_ = panda::MTManagedThread::GetCurrent();
thread_->ManagedCodeBegin();
}
~BytecodeOptPeepholes()
{
thread_->ManagedCodeEnd();
Runtime::Destroy();
}
protected:
panda::MTManagedThread *thread_;
};
TEST_F(BytecodeOptPeepholes, TryBlock)
{
pandasm::Parser p;
auto source = R"(
.record E {}
.record R {
u1 field
}
.function void R.ctor(R a0) <ctor> {
newobj v0, E
throw v0
}
.function u8 try_catch() {
try_begin:
movi v1, 0x1
newobj v0, R
movi v1, 0x2
call.short R.ctor, v0
try_end:
ldai 0x0
return
catch_all:
lda v1
return
.catchall try_begin, try_end, catch_all
}
)";
auto res = p.Parse(source);
auto &program = res.Value();
pandasm::AsmEmitter::PandaFileToPandaAsmMaps maps;
std::string file_name = "bc_peepholes";
auto piece = pandasm::AsmEmitter::Emit(file_name, program, nullptr, &maps);
ASSERT_NE(piece, false);
EXPECT_TRUE(OptimizeBytecode(&program, &maps, file_name, false, true));
// Check if there is initobj instruction in the bytecode
bool contains_initobj = false;
const auto sig_try_catch = pandasm::GetFunctionSignatureFromName("try_catch", {});
for (const auto &inst : program.function_table.at(sig_try_catch).ins) {
if (inst.opcode == pandasm::Opcode::INITOBJ) {
contains_initobj = true;
}
}
EXPECT_FALSE(contains_initobj);
auto pf = pandasm::AsmEmitter::Emit(program);
ASSERT_NE(pf, nullptr);
ClassLinker *class_linker = Runtime::GetCurrent()->GetClassLinker();
class_linker->AddPandaFile(std::move(pf));
auto *extension = class_linker->GetExtension(panda_file::SourceLang::PANDA_ASSEMBLY);
PandaString descriptor;
auto *klass = extension->GetClass(ClassHelper::GetDescriptor(utf::CStringAsMutf8("_GLOBAL"), &descriptor));
ASSERT_NE(klass, nullptr);
Method *method = klass->GetDirectMethod(utf::CStringAsMutf8("try_catch"));
ASSERT_NE(method, nullptr);
std::vector<Value> args;
args.emplace_back(Value(1, interpreter::TypeTag::INT));
Value v = method->Invoke(ManagedThread::GetCurrent(), args.data());
EXPECT_EQ(v.GetAsLong(), 0x2);
}
} // namespace panda::bytecodeopt::test

230
bytecode_optimizer/tests/bytecodeopt_peepholes_test.cpp
View File

@ -1,230 +0,0 @@
/**
* Copyright (c) 2021-2022 Huawei Device Co., Ltd.
* 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 "bytecodeopt_peepholes.h"
#include "common.h"
#include "compiler/optimizer/optimizations/cleanup.h"
namespace panda::bytecodeopt::test {
TEST_F(IrBuilderTest, PeepholesTryBlockInstBetween)
{
auto source = R"(
.record E {}
.record R {
u1 field
}
.function void R.ctor(R a0) <ctor> {
newobj v0, E
throw v0
}
.function u8 main() {
try_begin:
movi v1, 0x1
newobj v0, R
movi v1, 0x2
call.short R.ctor, v0
try_end:
ldai 0x0
return
catch_all:
lda v1
return
.catchall try_begin, try_end, catch_all
}
)";
ASSERT_TRUE(ParseToGraph(source, "main"));
EXPECT_FALSE(GetGraph()->RunPass<BytecodeOptPeepholes>());
}
TEST_F(IrBuilderTest, PeepholesTryBlockNoInstBetween)
{
auto source = R"(
.record E {}
.record R {
u1 field
}
.function void R.ctor(R a0) <ctor> {
newobj v0, E
throw v0
}
.function u8 main() {
try_begin:
movi v1, 0x1
newobj v0, R
call.short R.ctor, v0
try_end:
ldai 0x0
return
catch_all:
lda v1
return
.catchall try_begin, try_end, catch_all
}
)";
ASSERT_TRUE(ParseToGraph(source, "main"));
EXPECT_TRUE(GetGraph()->RunPass<BytecodeOptPeepholes>());
}
// TODO(aromanov): enable
TEST_F(CommonTest, DISABLED_NoNullCheck)
{
RuntimeInterfaceMock runtime(0);
auto graph = CreateEmptyGraph();
graph->SetRuntime(&runtime);
GRAPH(graph)
{
using namespace compiler::DataType;
CONSTANT(6, 0);
BASIC_BLOCK(2, -1)
{
INST(0, Opcode::SaveState).NoVregs();
INST(1, Opcode::LoadAndInitClass).ref().Inputs(0).TypeId(68);
INST(2, Opcode::NewObject).ref().Inputs(1, 0).TypeId(68);
INST(3, Opcode::SaveState).NoVregs();
INST(5, Opcode::CallStatic).v0id().Inputs({{REFERENCE, 2}, {NO_TYPE, 3}});
INST(7, Opcode::Return).s32().Inputs(6);
}
}
EXPECT_TRUE(graph->RunPass<BytecodeOptPeepholes>());
EXPECT_TRUE(graph->RunPass<compiler::Cleanup>());
auto after = CreateEmptyGraph();
GRAPH(after)
{
using namespace compiler::DataType;
CONSTANT(6, 0);
BASIC_BLOCK(2, -1)
{
INST(0, Opcode::SaveState).NoVregs();
INST(1, Opcode::LoadAndInitClass).ref().Inputs(0).TypeId(68);
INST(3, Opcode::SaveState).NoVregs();
INST(8, Opcode::InitObject).ref().Inputs({{REFERENCE, 1}, {NO_TYPE, 3}});
INST(7, Opcode::Return).s32().Inputs(6);
}
}
EXPECT_TRUE(GraphComparator().Compare(graph, after));
}
// TODO(aromanov): enable
TEST_F(CommonTest, DISABLED_NotRelatedNullCheck)
{
RuntimeInterfaceMock runtime(1);
auto graph = CreateEmptyGraph();
graph->SetRuntime(&runtime);
GRAPH(graph)
{
using namespace compiler::DataType;
PARAMETER(10, 0).ref();
CONSTANT(6, 0);
BASIC_BLOCK(2, -1)
{
INST(0, Opcode::SaveState).NoVregs();
INST(1, Opcode::LoadAndInitClass).ref().Inputs(0);
INST(2, Opcode::NewObject).ref().Inputs(1, 0);
INST(3, Opcode::SaveState).NoVregs();
INST(4, Opcode::NullCheck).ref().Inputs(10, 3);
INST(5, Opcode::CallStatic).v0id().Inputs({{REFERENCE, 2}, {NO_TYPE, 3}});
INST(7, Opcode::Return).s32().Inputs(6);
}
}
EXPECT_FALSE(graph->RunPass<BytecodeOptPeepholes>());
}
TEST_F(CommonTest, CallStaticOtherBasicBlock)
{
RuntimeInterfaceMock runtime(1);
auto graph = CreateEmptyGraph();
graph->SetRuntime(&runtime);
GRAPH(graph)
{
using namespace compiler::DataType;
PARAMETER(10, 0).ref();
CONSTANT(6, 0);
BASIC_BLOCK(2, 3)
{
INST(0, Opcode::SaveState).NoVregs();
INST(1, Opcode::LoadAndInitClass).ref().Inputs(0);
INST(2, Opcode::NewObject).ref().Inputs(1, 0);
INST(3, Opcode::SaveState).NoVregs();
}
BASIC_BLOCK(3, -1)
{
INST(5, Opcode::CallStatic).v0id().Inputs({{REFERENCE, 2}, {NO_TYPE, 3}});
INST(7, Opcode::Return).s32().Inputs(6);
}
}
EXPECT_FALSE(graph->RunPass<BytecodeOptPeepholes>());
}
// TODO(aromanov): enable
TEST_F(CommonTest, DISABLED_NoSaveStateNullCheckAfterNewObject)
{
RuntimeInterfaceMock runtime(0);
auto graph = CreateEmptyGraph();
graph->SetRuntime(&runtime);
GRAPH(graph)
{
using namespace compiler::DataType;
BASIC_BLOCK(2, -1)
{
INST(0, Opcode::SaveState).NoVregs();
INST(1, Opcode::LoadAndInitClass).ref().Inputs(0);
INST(2, Opcode::NewObject).ref().Inputs(1, 0);
CONSTANT(3, 0).s32();
INST(4, Opcode::SaveState).NoVregs();
INST(5, Opcode::CallStatic).v0id().Inputs({{REFERENCE, 2}, {NO_TYPE, 4}});
INST(6, Opcode::ReturnVoid).v0id();
}
}
EXPECT_FALSE(graph->RunPass<BytecodeOptPeepholes>());
}
TEST_F(CommonTest, CallConstructorOtherClass)
{
RuntimeInterfaceMock runtime(1, false);
auto graph = CreateEmptyGraph();
graph->SetRuntime(&runtime);
GRAPH(graph)
{
using namespace compiler::DataType;
PARAMETER(10, 0).ref();
CONSTANT(6, 0);
BASIC_BLOCK(2, -1)
{
INST(0, Opcode::SaveState).NoVregs();
INST(1, Opcode::LoadAndInitClass).ref().Inputs(0);
INST(2, Opcode::NewObject).ref().Inputs(1, 0);
INST(3, Opcode::SaveState).NoVregs();
INST(5, Opcode::CallStatic).v0id().Inputs({{REFERENCE, 2}, {NO_TYPE, 3}});
INST(7, Opcode::Return).s32().Inputs(6);
}
}
EXPECT_FALSE(graph->RunPass<BytecodeOptPeepholes>());
}
} // namespace panda::bytecodeopt::test

172
compiler/BUILD.gn
View File

@ -18,10 +18,6 @@ config("arkcompiler_public_config") {
"$ark_root/compiler/code_info",
"$ark_root/compiler/optimizer/ir",
"$target_gen_dir/generated",
get_label_info("$ark_root/cross_values:cross_values_getters_generate",
"target_gen_dir"),
get_label_info("$ark_root/runtime:plugins_defines_h", "target_gen_dir") +
"/asm_defines/generated",
]
include_dirs += platform_include_dirs
@ -50,20 +46,11 @@ config("arkcompiler_public_config") {
libarkcompiler_sources = [
"compiler_logger.cpp",
"compiler_options.cpp",
"optimizer/analysis/alias_analysis.cpp",
"optimizer/analysis/bounds_analysis.cpp",
"optimizer/analysis/countable_loop_parser.cpp",
"optimizer/analysis/dominators_tree.cpp",
"optimizer/analysis/linear_order.cpp",
"optimizer/analysis/live_registers.cpp",
"optimizer/analysis/liveness_analyzer.cpp",
"optimizer/analysis/liveness_use_table.cpp",
"optimizer/analysis/loop_analyzer.cpp",
"optimizer/analysis/object_type_propagation.cpp",
"optimizer/analysis/rpo.cpp",
"optimizer/analysis/types_analysis.cpp",
"optimizer/ir/analysis.cpp",
"optimizer/ir/aot_data.cpp",
"optimizer/ir/basicblock.cpp",
"optimizer/ir/dump.cpp",
"optimizer/ir/graph.cpp",
@ -73,20 +60,9 @@ libarkcompiler_sources = [
"optimizer/ir/locations.cpp",
"optimizer/ir_builder/inst_builder.cpp",
"optimizer/ir_builder/ir_builder.cpp",
"optimizer/optimizations/adjust_arefs.cpp",
"optimizer/optimizations/balance_expressions.cpp",
"optimizer/optimizations/cleanup.cpp",
"optimizer/optimizations/code_sink.cpp",
"optimizer/optimizations/deoptimize_elimination.cpp",
"optimizer/optimizations/if_conversion.cpp",
"optimizer/optimizations/loop_peeling.cpp",
"optimizer/optimizations/loop_unroll.cpp",
"optimizer/optimizations/lowering.cpp",
"optimizer/optimizations/memory_barriers.cpp",
"optimizer/optimizations/memory_coalescing.cpp",
"optimizer/optimizations/move_constants.cpp",
"optimizer/optimizations/object_type_check_elimination.cpp",
"optimizer/optimizations/redundant_loop_elimination.cpp",
"optimizer/optimizations/regalloc/interference_graph.cpp",
"optimizer/optimizations/regalloc/reg_alloc.cpp",
"optimizer/optimizations/regalloc/reg_alloc_base.cpp",
@ -96,44 +72,12 @@ libarkcompiler_sources = [
"optimizer/optimizations/regalloc/reg_map.cpp",
"optimizer/optimizations/regalloc/spill_fills_resolver.cpp",
"optimizer/optimizations/regalloc/split_resolver.cpp",
"optimizer/optimizations/scheduler.cpp",
"optimizer/optimizations/try_catch_resolving.cpp",
"optimizer/optimizations/types_resolving.cpp",
"optimizer/optimizations/vn.cpp",
"optimizer/pass.cpp",
"optimizer/pass_manager.cpp",
"optimizer/pass_manager_statistics.cpp",
]
plugin_irts = []
foreach(plugin, enabled_plugins) {
print("add compiler plugin: $plugin")
plugin_dir = "$ark_root/plugins/$plugin"
source_files = read_file("$plugin_dir/subproject_sources.gn", "scope")
if (defined(source_files.srcs_compiler_path)) {
source_file = "$plugin_dir/${source_files.srcs_compiler_path}"
src_scope = read_file(source_file, "scope")
if (defined(src_scope.srcs)) {
foreach(src, src_scope.srcs) {
libarkcompiler_sources += [ "$plugin_dir/compiler/$src" ]
}
}
if (defined(source_files.irtoc_plugins)) {
scripts = read_file("$plugin_dir/${source_files.irtoc_plugins}", "scope")
foreach(script, scripts.srcs) {
plugin_irts +=
[ rebase_path("$plugin_dir/irtoc_scripts/$script", ark_root) ]
}
}
src_scope = {
}
}
source_files = []
}
libarkcompiler_sources += [ "$target_gen_dir/generated/inst_builder_gen.cpp" ]
libarkcompiler_configs = [
@ -142,7 +86,6 @@ libarkcompiler_configs = [
"$ark_root:ark_config",
"$ark_root/libpandabase:arkbase_public_config",
"$ark_root/libpandafile:arkfile_public_config",
"$ark_root/runtime:arkruntime_public_config",
]
ohos_shared_library("libarkcompiler") {
@ -157,7 +100,6 @@ ohos_shared_library("libarkcompiler") {
":compiler_logger_components_inc",
":compiler_options_gen_h",
":cpu_features_gen_inc",
":intrinsics_can_encode_inl",
":intrinsics_codegen_ext_inl_h",
":intrinsics_inline_inl",
":intrinsics_ir_build_inl_h",
@ -165,14 +107,10 @@ ohos_shared_library("libarkcompiler") {
":intrinsics_ir_build_virtual_call_inl",
":intrinsics_stub_inl",
":intrinsics_stub_inl_h",
":intrinsics_types_resolving_inl_h",
":ir_dyn_base_types_h",
":irtoc_builder_cpp",
":irtoc_generate_ir_inline",
":isa_gen_libarkcompiler_inst_builder_gen_cpp",
":libarkcompiler_ecma_intrinsics_enum_inl",
":libarkcompiler_generate_ecma_inl",
":libarkcompiler_intrinsics_gen_inl_can_encode_builtin_inl",
":libarkcompiler_intrinsics_gen_inl_generate_operations_intrinsic_graph_inl",
":libarkcompiler_intrinsics_gen_inl_generate_operations_intrinsic_inst_inl",
":libarkcompiler_intrinsics_gen_inl_get_intrinsics_inl",
@ -187,26 +125,14 @@ ohos_shared_library("libarkcompiler") {
":libarkcompiler_opcodes_h_IR-instructions_md",
":libarkcompiler_opcodes_h_arch_info_gen_h",
":libarkcompiler_opcodes_h_codegen_arm64_gen_inc",
":libarkcompiler_opcodes_h_deoptimize_elimination_call_visitors_inl",
":libarkcompiler_opcodes_h_inst_checker_gen_h",
":libarkcompiler_opcodes_h_opcodes_h",
":source_languages_h",
"$ark_root/cross_values:cross_values_getters_generate(${default_toolchain})",
"$ark_root/libpandabase:libarkbase",
"$ark_root/libpandafile:isa_gen_libarkfile_bytecode_instruction-inl_gen_h",
"$ark_root/libpandafile:isa_gen_libarkfile_bytecode_instruction_enum_gen_h",
"$ark_root/libpandafile:libarkfile",
"$ark_root/libpandafile:libarkfile_type_gen_h",
"$ark_root/runtime:arkruntime_gen_entrypoints_compiler_checksum_entrypoints_compiler_checksum_inl",
"$ark_root/runtime:arkruntime_gen_entrypoints_entrypoints_compiler_inl",
"$ark_root/runtime:arkruntime_gen_entrypoints_entrypoints_gen_S",
"$ark_root/runtime:arkruntime_gen_entrypoints_entrypoints_gen_h",
"$ark_root/runtime:arkruntime_gen_intrinsics_intrinsics_enum_h",
"$ark_root/runtime:arkruntime_gen_intrinsics_intrinsics_gen_h",
"$ark_root/runtime:arkruntime_gen_intrinsics_intrinsics_h",
"$ark_root/runtime:arkruntime_gen_intrinsics_unimplemented_intrinsics-inl_cpp",
"$ark_root/runtime:plugins_entrypoints_gen_h",
"$ark_root/runtime/asm_defines:asm_defines_generator",
sdk_libc_secshared_dep,
]
@ -230,7 +156,6 @@ ohos_static_library("libarkcompiler_frontend_static") {
":compiler_logger_components_inc",
":compiler_options_gen_h",
":cpu_features_gen_inc",
":intrinsics_can_encode_inl",
":intrinsics_codegen_ext_inl_h",
":intrinsics_inline_inl",
":intrinsics_ir_build_inl_h",
@ -238,14 +163,10 @@ ohos_static_library("libarkcompiler_frontend_static") {
":intrinsics_ir_build_virtual_call_inl",
":intrinsics_stub_inl",
":intrinsics_stub_inl_h",
":intrinsics_types_resolving_inl_h",
":ir_dyn_base_types_h",
":irtoc_builder_cpp",
":irtoc_generate_ir_inline",
":isa_gen_libarkcompiler_inst_builder_gen_cpp",
":libarkcompiler_ecma_intrinsics_enum_inl",
":libarkcompiler_generate_ecma_inl",
":libarkcompiler_intrinsics_gen_inl_can_encode_builtin_inl",
":libarkcompiler_intrinsics_gen_inl_generate_operations_intrinsic_graph_inl",
":libarkcompiler_intrinsics_gen_inl_generate_operations_intrinsic_inst_inl",
":libarkcompiler_intrinsics_gen_inl_get_intrinsics_inl",
@ -260,26 +181,14 @@ ohos_static_library("libarkcompiler_frontend_static") {
":libarkcompiler_opcodes_h_IR-instructions_md",
":libarkcompiler_opcodes_h_arch_info_gen_h",
":libarkcompiler_opcodes_h_codegen_arm64_gen_inc",
":libarkcompiler_opcodes_h_deoptimize_elimination_call_visitors_inl",
":libarkcompiler_opcodes_h_inst_checker_gen_h",
":libarkcompiler_opcodes_h_opcodes_h",
":source_languages_h",
"$ark_root/cross_values:cross_values_getters_generate(${default_toolchain})",
"$ark_root/libpandabase:libarkbase_frontend_static",
"$ark_root/libpandafile:isa_gen_libarkfile_bytecode_instruction-inl_gen_h",
"$ark_root/libpandafile:isa_gen_libarkfile_bytecode_instruction_enum_gen_h",
"$ark_root/libpandafile:libarkfile_frontend_static",
"$ark_root/libpandafile:libarkfile_type_gen_h",
"$ark_root/runtime:arkruntime_gen_entrypoints_compiler_checksum_entrypoints_compiler_checksum_inl",
"$ark_root/runtime:arkruntime_gen_entrypoints_entrypoints_compiler_inl",
"$ark_root/runtime:arkruntime_gen_entrypoints_entrypoints_gen_S",
"$ark_root/runtime:arkruntime_gen_entrypoints_entrypoints_gen_h",
"$ark_root/runtime:arkruntime_gen_intrinsics_intrinsics_enum_h",
"$ark_root/runtime:arkruntime_gen_intrinsics_intrinsics_gen_h",
"$ark_root/runtime:arkruntime_gen_intrinsics_intrinsics_h",
"$ark_root/runtime:arkruntime_gen_intrinsics_unimplemented_intrinsics-inl_cpp",
"$ark_root/runtime:plugins_entrypoints_gen_h",
"$ark_root/runtime/asm_defines:asm_defines_generator",
sdk_libc_secshared_dep,
]
@ -293,7 +202,7 @@ ark_isa_gen("isa_gen_libarkcompiler") {
}
ark_gen("libarkcompiler_intrinsics_gen_inl") {
data = "$target_gen_dir/../runtime/intrinsics.yaml"
data = "intrinsics.yaml"
template_files = [
"intrinsics_enum.inl.erb",
"get_intrinsics.inl.erb",
@ -304,7 +213,6 @@ ark_gen("libarkcompiler_intrinsics_gen_inl") {
"generate_operations_intrinsic_graph.inl.erb",
"intrinsic_codegen_test.inl.erb",
"intrinsic_flags_test.inl.erb",
"can_encode_builtin.inl.erb",
"intrinsics_codegen.inl.h.erb",
"intrinsics_codegen.inl.erb",
"intrinsics_flags.inl.erb",
@ -314,7 +222,6 @@ ark_gen("libarkcompiler_intrinsics_gen_inl") {
requires = [
"$ark_root/compiler/optimizer/templates/intrinsics/compiler_intrinsics.rb",
]
extra_dependencies = [ "$ark_root/runtime:arkruntime_gen_intrinsics_yaml" ]
}
ark_isa_gen("libarkcompiler") {
@ -335,7 +242,6 @@ ark_gen("libarkcompiler_opcodes_h") {
"inst_checker_gen.h.erb",
"IR-instructions.md.erb",
"codegen_arm64_gen.inc.erb",
"deoptimize_elimination_call_visitors.inl.erb",
]
sources = "optimizer/templates"
destination = "$target_gen_dir/generated"
@ -399,14 +305,6 @@ ark_gen_file("intrinsics_inline_inl") {
requires = [ "$ark_root/templates/plugin_options.rb" ]
output_file = "$target_gen_dir/generated/intrinsics_inline.inl"
}
ark_gen_file("intrinsics_types_resolving_inl_h") {
extra_dependencies = [ "$ark_root:concat_plugins_yamls" ]
template_file =
"optimizer/templates/intrinsics/intrinsics_types_resolving.inl.h.erb"
data_file = "$target_gen_dir/../plugin_options.yaml"
requires = [ "$ark_root/templates/plugin_options.rb" ]
output_file = "$target_gen_dir/generated/intrinsics_types_resolving.inl.h"
}
ark_gen_file("intrinsics_ir_build_inl_h") {
extra_dependencies = [ "$ark_root:concat_plugins_yamls" ]
template_file = "optimizer/templates/intrinsics/intrinsics_ir_build.inl.h.erb"
@ -430,13 +328,6 @@ ark_gen_file("intrinsics_ir_build_virtual_call_inl") {
requires = [ "$ark_root/templates/plugin_options.rb" ]
output_file = "$target_gen_dir/generated/intrinsics_ir_build_virtual_call.inl"
}
ark_gen_file("intrinsics_can_encode_inl") {
extra_dependencies = [ "$ark_root:concat_plugins_yamls" ]
template_file = "optimizer/templates/intrinsics/intrinsics_can_encode.inl.erb"
data_file = "$target_gen_dir/../plugin_options.yaml"
requires = [ "$ark_root/templates/plugin_options.rb" ]
output_file = "$target_gen_dir/generated/intrinsics_can_encode.inl"
}
ark_gen_file("ir_dyn_base_types_h") {
extra_dependencies = [ "$ark_root:concat_plugins_yamls" ]
@ -469,64 +360,3 @@ ark_gen_file("compiler_interface_extensions_inl_h") {
requires = [ "$ark_root/templates/plugin_options.rb" ]
output_file = "$target_gen_dir/generated/compiler_interface_extensions.inl.h"
}
irtoc_file_plugin = "$root_gen_dir/generated/irtoc_plugin_builder.txt"
write_file(irtoc_file_plugin, plugin_irts)
action("irtoc_builder_cpp") {
script = "$ark_root/irtoc/lang/irtoc.rb"
outputs = [ "$target_gen_dir/generated/irtoc_builder.cpp" ]
args = [
"--input",
rebase_path("$ark_root/irtoc/scripts/interpreter.irt", root_build_dir),
"--output",
rebase_path("$target_gen_dir/generated/irtoc_builder.cpp", root_build_dir),
"--ark_source_dir",
rebase_path("$ark_root", root_build_dir),
"--isa",
rebase_path("$root_gen_dir/isa/isa.yaml", root_build_dir),
"--definitions",
"NDEBUG",
"--arch",
target_cpu,
"--ir-api",
"ir-builder",
"--plugins",
rebase_path(irtoc_file_plugin, root_build_dir),
]
deps = [
"$ark_root/isa:isa_combine",
"$ark_root/runtime:plugins_asm_defines_def",
"$ark_root/runtime:plugins_defines_h",
]
}
action("irtoc_generate_ir_inline") {
script = "$ark_root/irtoc/lang/irtoc.rb"
outputs = [ "$target_gen_dir/generated/irtoc_ir_inline.h" ]
args = [
"--input",
rebase_path("$ark_root/irtoc/scripts/interpreter.irt", root_build_dir),
"--output",
rebase_path("$target_gen_dir/generated/irtoc_ir_inline.h", root_build_dir),
"--ark_source_dir",
rebase_path("$ark_root", root_build_dir),
"--isa",
rebase_path("$root_gen_dir/isa/isa.yaml", root_build_dir),
"--definitions",
"NDEBUG",
"--arch",
target_cpu,
"--ir-api",
"ir-inline",
"--plugins",
rebase_path(irtoc_file_plugin, root_build_dir),
]
deps = [
"$ark_root/isa:isa_combine",
"$ark_root/runtime:plugins_asm_defines_def",
"$ark_root/runtime:plugins_defines_h",
]
}

643
compiler/CMakeLists.txt
View File

@ -1,643 +0,0 @@
# Copyright (c) 2021-2022 Huawei Device Co., Ltd.
# 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.
project(compiler)
include(cmake/coverage.cmake)
if(PANDA_COMPILER_ENABLE)
include(cmake/benchmark_coverage.cmake)
include(cmake/ir_builder_coverage.cmake)
endif()
include(cmake/target.cmake)
include_directories(
.
${PANDA_ROOT}
)
include_directories(AFTER ${PANDA_BINARY_ROOT}/cross_values)
set(GENERATED_DIR ${CMAKE_CURRENT_BINARY_DIR}/generated)
file(MAKE_DIRECTORY ${GENERATED_DIR})
set(IRTOC_BUILDER_GEN ${GENERATED_DIR}/irtoc_builder.cpp)
irtoc_generate(
TARGET irtoc_generate_inst_builder
IR_API ir-builder
INPUT_FILES ${IRTOC_SOURCE_DIR}/scripts/interpreter.irt
OUTPUT_FILE ${IRTOC_BUILDER_GEN}
WORKING_DIRECTORY ${IRTOC_BUILD_DIR}/inst_builder
)
add_dependencies(irtoc_generate_inst_builder asm_defines)
set(IRTOC_IR_INLINE_GEN ${GENERATED_DIR}/irtoc_ir_inline.h)
irtoc_generate(
TARGET irtoc_generate_ir_inline
IR_API ir-inline
INPUT_FILES ${IRTOC_SOURCE_DIR}/scripts/interpreter.irt
OUTPUT_FILE ${IRTOC_IR_INLINE_GEN}
WORKING_DIRECTORY ${IRTOC_BUILD_DIR}/ir_inline
)
add_dependencies(irtoc_generate_ir_inline asm_defines)
set(COMPILER_TEMPLATES_DIR ${CMAKE_CURRENT_LIST_DIR}/optimizer/templates)
add_custom_target(inst_templates_gen)
set_target_properties(inst_templates_gen PROPERTIES INST_TEMPLATES_YAML_FILES ${CMAKE_CURRENT_LIST_DIR}/optimizer/ir_builder/inst_templates.yaml)
set(INST_TEMPLATES_GEN_YAML ${GENERATED_DIR}/inst_templates.yaml)
set_target_properties(inst_templates_gen PROPERTIES INST_TEMPLATES_GEN_YAML ${INST_TEMPLATES_GEN_YAML})
function(add_inst_templates YAML_FILE_PATH)
get_target_property(YAML_FILES inst_templates_gen INST_TEMPLATES_YAML_FILES)
list(APPEND YAML_FILES ${YAML_FILE_PATH})
set_target_properties(inst_templates_gen PROPERTIES INST_TEMPLATES_YAML_FILES "${YAML_FILES}")
endfunction()
set(INST_BUILDER_GEN ${GENERATED_DIR}/inst_builder_gen.cpp)
configure_file(${COMPILER_TEMPLATES_DIR}/inst_builder_gen.cpp.erb ${GENERATED_DIR}/inst_builder_gen.cpp.erb @ONLY)
panda_isa_gen(
TEMPLATES
"inst_builder_gen.cpp.erb"
SOURCE ${GENERATED_DIR}
REQUIRES ${PANDA_ROOT}/assembler/asm_isapi.rb
DESTINATION ${GENERATED_DIR}
EXTRA_DEPENDENCIES inst_templates_merge
)
panda_gen(DATA ${PANDA_BINARY_ROOT}/runtime/intrinsics.yaml
TEMPLATES
intrinsics_enum.inl.erb
get_intrinsics.inl.erb
entrypoints_bridge_asm_macro.inl.erb
intrinsics_ir_build.inl.erb
intrinsics_flags.inl.erb
get_intrinsics_names.inl.erb
generate_operations_intrinsic_inst.inl.erb
generate_operations_intrinsic_graph.inl.erb
intrinsic_codegen_test.inl.erb
intrinsic_flags_test.inl.erb
can_encode_builtin.inl.erb
intrinsics_codegen.inl.h.erb
intrinsics_codegen.inl.erb
SOURCE ${COMPILER_TEMPLATES_DIR}/intrinsics
DESTINATION ${GENERATED_DIR}
REQUIRES ${COMPILER_TEMPLATES_DIR}/intrinsics/compiler_intrinsics.rb
EXTRA_DEPENDENCIES arkruntime_gen_intrinsics_yaml
)
panda_gen(DATA ${CMAKE_CURRENT_LIST_DIR}/optimizer/ir/instructions.yaml
TEMPLATES
opcodes.h.erb
arch_info_gen.h.erb
inst_checker_gen.h.erb
IR-instructions.md.erb
codegen_arm64_gen.inc.erb
deoptimize_elimination_call_visitors.inl.erb
SOURCE ${COMPILER_TEMPLATES_DIR}
DESTINATION ${GENERATED_DIR}
REQUIRES ${COMPILER_TEMPLATES_DIR}/instructions.rb
)
set(COMPILER_OPTIONS_GEN_H ${GENERATED_DIR}/compiler_options_gen.h)
panda_gen_file(
DATAFILE ${CMAKE_CURRENT_LIST_DIR}/compiler.yaml
TEMPLATE ${PANDA_ROOT}/templates/options/options.h.erb
OUTPUTFILE ${COMPILER_OPTIONS_GEN_H}
REQUIRES ${PANDA_ROOT}/templates/common.rb
)
add_custom_target(compiler_options_gen DEPENDS ${COMPILER_OPTIONS_GEN_H})
set(COMPILER_EVENTS_GEN_H ${GENERATED_DIR}/compiler_events_gen.h)
panda_gen_file(
DATAFILE ${CMAKE_CURRENT_LIST_DIR}/compiler.yaml
TEMPLATE ${PANDA_ROOT}/templates/events/events.h.erb
OUTPUTFILE ${COMPILER_EVENTS_GEN_H}
REQUIRES ${PANDA_ROOT}/templates/common.rb
)
add_custom_target(compiler_events_gen DEPENDS ${COMPILER_EVENTS_GEN_H})
set(COMPILER_LOGGER_COMPONENTS_GEN_H ${GENERATED_DIR}/compiler_logger_components.inc)
panda_gen_file(
DATAFILE ${CMAKE_CURRENT_LIST_DIR}/compiler.yaml
TEMPLATE ${PANDA_ROOT}/templates/logger_components/logger_components.inc.erb
OUTPUTFILE ${COMPILER_LOGGER_COMPONENTS_GEN_H}
REQUIRES ${PANDA_ROOT}/templates/common.rb
)
add_custom_target(compiler_logger_components_gen DEPENDS ${COMPILER_LOGGER_COMPONENTS_GEN_H})
set(CPU_FEATURES_GEN_H ${GENERATED_DIR}/cpu_features.inc)
panda_gen_file(
DATAFILE ${CMAKE_CURRENT_LIST_DIR}/compiler.yaml
TEMPLATE ${PANDA_ROOT}/templates/cpu_features.inc.erb
OUTPUTFILE ${CPU_FEATURES_GEN_H}
REQUIRES ${PANDA_ROOT}/templates/common.rb
)
add_custom_target(cpu_features_gen DEPENDS ${CPU_FEATURES_GEN_H})
set(OPTIMIZER_SOURCES
optimizer/pass.cpp
optimizer/pass_manager.cpp
optimizer/pass_manager_statistics.cpp
optimizer/analysis/alias_analysis.cpp
optimizer/analysis/bounds_analysis.cpp
optimizer/analysis/countable_loop_parser.cpp
optimizer/analysis/dominators_tree.cpp
optimizer/analysis/linear_order.cpp
optimizer/analysis/liveness_analyzer.cpp
optimizer/analysis/liveness_use_table.cpp
optimizer/analysis/live_registers.cpp
optimizer/analysis/loop_analyzer.cpp
optimizer/analysis/monitor_analysis.cpp
optimizer/analysis/object_type_propagation.cpp
optimizer/analysis/rpo.cpp
optimizer/analysis/reg_alloc_verifier.cpp
optimizer/analysis/types_analysis.cpp
optimizer/ir/analysis.cpp
optimizer/ir/basicblock.cpp
optimizer/ir/dump.cpp
optimizer/ir/graph.cpp
optimizer/ir/inst.cpp
optimizer/ir/locations.cpp
optimizer/ir/visualizer_printer.cpp
optimizer/ir/graph_checker.cpp
optimizer/ir/graph_cloner.cpp
optimizer/optimizations/adjust_arefs.cpp
optimizer/optimizations/balance_expressions.cpp
optimizer/optimizations/branch_elimination.cpp
optimizer/optimizations/checks_elimination.cpp
optimizer/optimizations/code_sink.cpp
optimizer/optimizations/const_folding.cpp
optimizer/optimizations/deoptimize_elimination.cpp
optimizer/optimizations/cleanup.cpp
optimizer/optimizations/if_conversion.cpp
optimizer/optimizations/licm.cpp
optimizer/optimizations/locations_builder.cpp
optimizer/optimizations/loop_peeling.cpp
optimizer/optimizations/loop_unroll.cpp
optimizer/optimizations/lse.cpp
optimizer/optimizations/memory_barriers.cpp
optimizer/optimizations/memory_coalescing.cpp
optimizer/optimizations/object_type_check_elimination.cpp
optimizer/optimizations/peepholes.cpp
optimizer/optimizations/redundant_loop_elimination.cpp
optimizer/optimizations/scheduler.cpp
optimizer/optimizations/try_catch_resolving.cpp
optimizer/optimizations/types_resolving.cpp
optimizer/optimizations/vn.cpp
optimizer/optimizations/cse.cpp
tools/debug/jit_writer.cpp
compiler_logger.cpp
compiler_options.cpp
)
set(COMPILER_SOURCES
optimizer/ir/aot_data.cpp
optimizer/ir_builder/ir_builder.cpp
optimizer/ir_builder/inst_builder.cpp
optimizer/optimizations/inlining.cpp
optimizer/optimizations/lowering.cpp
optimizer/optimizations/move_constants.cpp
optimizer/optimizations/regalloc/reg_alloc_base.cpp
optimizer/optimizations/regalloc/interference_graph.cpp
optimizer/optimizations/regalloc/reg_alloc.cpp
optimizer/optimizations/regalloc/reg_alloc_stat.cpp
optimizer/optimizations/regalloc/reg_alloc_graph_coloring.cpp
optimizer/optimizations/regalloc/reg_map.cpp
optimizer/optimizations/regalloc/reg_alloc_linear_scan.cpp
optimizer/optimizations/regalloc/spill_fills_resolver.cpp
optimizer/optimizations/regalloc/split_resolver.cpp
optimizer/optimizations/regalloc/reg_alloc_resolver.cpp
${OPTIMIZER_SOURCES}
${INST_BUILDER_GEN}
${IRTOC_IR_INLINE_GEN}
)
if (NOT PANDA_TARGET_WINDOWS AND NOT PANDA_TARGET_MACOS)
list(APPEND COMPILER_SOURCES
compile_method.cpp
optimizer_run.cpp
optimizer/code_generator/disassembly.cpp
optimizer/code_generator/codegen.cpp
optimizer/code_generator/codegen_native.cpp
optimizer/code_generator/spill_fill_encoder.cpp
optimizer/code_generator/slow_path.cpp
optimizer/code_generator/method_properties.cpp
code_info/code_info.cpp
code_info/code_info_builder.cpp
)
endif()
add_library(arkcompiler ${PANDA_DEFAULT_LIB_TYPE} ${COMPILER_SOURCES})
panda_add_to_clang_tidy(TARGET arkcompiler)
add_dependencies(arkcompiler isa_gen_${PROJECT_NAME})
add_dependencies(arkcompiler instructions_gen_${PROJECT_NAME})
add_dependencies(arkcompiler compiler_events_gen)
add_dependencies(arkcompiler compiler_logger_components_gen)
add_dependencies(arkcompiler compiler_options_gen)
add_dependencies(arkcompiler irtoc_generate_inst_builder)
add_dependencies(arkcompiler irtoc_generate_ir_inline)
add_dependencies(arkcompiler intrinsics_gen_compiler)
add_dependencies(arkcompiler intrinsics_gen_arkruntime)
add_dependencies(arkcompiler entrypoints_gen)
add_dependencies(arkcompiler entrypoints_compiler_checksum_gen)
add_dependencies(arkcompiler cross_values)
add_dependencies(arkcompiler cpu_features_gen)
add_dependencies(arkcompiler asm_defines_generator)
if (PANDA_TARGET_MOBILE OR PANDA_TARGET_OHOS)
add_dependencies(host_tools_depends arkcompiler)
endif()
target_link_libraries(arkcompiler arkbase arkfile)
if (NOT PANDA_TARGET_WINDOWS AND NOT PANDA_TARGET_MACOS)
target_link_libraries(arkcompiler arkencoder)
endif()
target_include_directories(arkcompiler
PUBLIC ${PANDA_ROOT}
PUBLIC ${PANDA_ROOT}/runtime
PUBLIC ${PANDA_BINARY_ROOT}/runtime/include
PUBLIC ${PANDA_BINARY_ROOT}/cross_values
PUBLIC ${CMAKE_CURRENT_SOURCE_DIR}
PUBLIC ${GENERATED_DIR}
)
# Disable warning about offsetof usage for non-standard layout types.
# In C++17 offsetof is conditionally implemented for such types, so
# compiler should issue error if it doesn't implemented offsetof for
# them. Also we use static asserts to ensure that offsetof works correcly
# for non-standard layout types.
target_compile_options(arkcompiler PUBLIC -Wno-invalid-offsetof)
set(PANDA_COMPILER_TESTS_SOURCES
tests/adjust_arefs_test.cpp
tests/alias_analysis_test.cpp
tests/aot_test.cpp
tests/balance_expressions_test.cpp
tests/basicblock_test.cpp
tests/bounds_analysis_test.cpp
tests/branch_elimination_test.cpp
tests/call_input_types_test.cpp
tests/checks_elimination_test.cpp
tests/cleanup_test.cpp
tests/codegen_runner_test.cpp
tests/code_info_test.cpp
tests/code_sink_test.cpp
tests/const_folding_test.cpp
tests/cse_test.cpp
tests/deoptimize_elimination_test.cpp
tests/dom_tree_test.cpp
tests/graph_cloner_test.cpp
tests/graph_comparator_test.cpp
tests/graph_creation_test.cpp
tests/if_conversion_test.cpp
tests/inlining_test.cpp
tests/inst_test.cpp
tests/iterators_test.cpp
tests/licm_test.cpp
tests/liveness_analyzer_test.cpp
tests/live_registers_test.cpp
tests/loop_analyzer_test.cpp
tests/loop_peeling_test.cpp
tests/lse_test.cpp
tests/memory_barriers_test.cpp
tests/memory_coalescing_test.cpp
tests/method_properties_test.cpp
tests/monitor_analysis_test.cpp
tests/move_constants_test.cpp
tests/osr_test.cpp
tests/peepholes_test.cpp
tests/redundant_loop_elimination_test.cpp
tests/reg_alloc_common_test.cpp
tests/reg_alloc_graph_coloring_test.cpp
tests/reg_alloc_interference_graph_test.cpp
tests/rpo_test.cpp
tests/scheduler_test.cpp
tests/split_resolver_test.cpp
tests/try_catch_resolving_test.cpp
tests/unit_test.cpp
tests/vn_test.cpp
tests/class_hash_table_test.cpp
tests/linear_order_test.cpp
tests/profiling_runner_test.cpp
)
set_source_files_properties(tests/loop_unroll_test.cpp PROPERTIES COMPILE_FLAGS -Wno-shadow)
# Distinguish 'PANDA_COMPILER_TARGET_..' and 'PANDA_TARGET_..' because for PANDA_TARGET_AMD64 tests
# are being executed for Arch::AARCH64
if(PANDA_COMPILER_TARGET_AARCH64 OR PANDA_TARGET_ARM32)
if(PANDA_COMPILER_TARGET_AARCH64)
list(APPEND PANDA_COMPILER_TESTS_SOURCES
tests/aarch64/codegen_test.cpp
)
endif()
list(APPEND PANDA_COMPILER_TESTS_SOURCES
tests/life_intervals_test.cpp
tests/loop_unroll_test.cpp
tests/reg_alloc_linear_scan_test.cpp
tests/reg_alloc_verifier_test.cpp
)
list(APPEND PANDA_COMPILER_TESTS_SOURCES
tests/lowering_test.cpp)
endif()
if(PANDA_TARGET_AMD64 OR PANDA_NIGHTLY_TEST_ON)
list(APPEND PANDA_COMPILER_TESTS_SOURCES
tests/ir_builder_test.cpp)
endif()
set(PANDA_COMPILER_TESTS_LIBRARIES arkcompiler arkbase arkassembler arkruntime arkaotmanager aot_builder)
set(ENCODER_TESTS_SOURCES
tests/encoder_operands.cpp
tests/constructor_test.cpp
)
if(PANDA_COMPILER_ENABLE)
list(APPEND ENCODER_TESTS_SOURCES tests/asm_printer_test.cpp)
endif()
if (PANDA_TARGET_ARM32 AND PANDA_COMPILER_TARGET_AARCH32)
# Append to PANDA_COMPILER_TESTS_SOURCES modified version of inst_generator_test
list(APPEND ENCODER_TESTS_SOURCES
tests/aarch32/callconv32_test.cpp
tests/aarch32/encoder32_test.cpp
tests/aarch32/register32_test.cpp
)
endif (PANDA_TARGET_ARM32 AND PANDA_COMPILER_TARGET_AARCH32)
if (PANDA_TARGET_ARM64)
list(APPEND ENCODER_TESTS_SOURCES
tests/aarch64/callconv64_test.cpp
tests/aarch64/encoder64_test.cpp
tests/aarch64/register64_test.cpp
)
endif (PANDA_TARGET_ARM64)
if (PANDA_TARGET_X86)
list(APPEND ENCODER_TESTS_SOURCES
tests/x86/asmjit_test.cpp
)
list(APPEND PANDA_COMPILER_TESTS_LIBRARIES asmjit)
endif (PANDA_TARGET_X86)
if (PANDA_TARGET_AMD64)
list(APPEND ENCODER_TESTS_SOURCES
tests/amd64/asmjit_test.cpp
tests/amd64/callconv64_test.cpp
tests/amd64/encoder64_test.cpp
tests/amd64/register64_test.cpp
)
if (PANDA_COMPILER_TARGET_AARCH64)
list(APPEND ENCODER_TESTS_SOURCES
tests/encoders_test.cpp
)
endif()
list(APPEND PANDA_COMPILER_TESTS_LIBRARIES asmjit)
endif (PANDA_TARGET_AMD64)
if (NOT (PANDA_TARGET_MOBILE OR PANDA_TARGET_OHOS OR PANDA_ENABLE_FUZZBENCH))
list(APPEND PANDA_COMPILER_TESTS_LIBRARIES stdc++fs)
endif()
if(NOT PANDA_MINIMAL_VIXL AND PANDA_COMPILER_ENABLE)
panda_add_gtest(
CONTAINS_MAIN
NAME compiler_unit_tests
SOURCES
${PANDA_COMPILER_TESTS_SOURCES}
LIBRARIES
${PANDA_COMPILER_TESTS_LIBRARIES}
SANITIZERS
${PANDA_SANITIZERS_LIST}
)
endif()
# AMD64 and X86 - for unit tests
if(NOT PANDA_MINIMAL_VIXL AND PANDA_TARGET_AMD64 AND PANDA_WITH_TESTS)
set(PANDA_CODEGEN_TESTS_SOURCES
tests/unit_test.cpp
tests/spill_fill_encoder_test.cpp
)
if(PANDA_COMPILER_TARGET_AARCH64)
list(APPEND PANDA_CODEGEN_TESTS_SOURCES
tests/codegen_test.cpp
)
endif()
set_source_files_properties(tests/codegen_test.cpp PROPERTIES COMPILE_FLAGS -Wno-shadow)
panda_add_gtest(
CONTAINS_MAIN
NAME compiler_codegen_tests
SOURCES
${PANDA_CODEGEN_TESTS_SOURCES}
LIBRARIES
${PANDA_COMPILER_TESTS_LIBRARIES}
SANITIZERS
${PANDA_SANITIZERS_LIST}
)
if (PANDA_WITH_TESTS AND PANDA_COMPILER_TARGET_AARCH64)
set(PANDA_INST_GEN_TESTS_SOURCES
tests/unit_test.cpp
tests/inst_generator.cpp
tests/inst_generator_test.cpp
)
set_source_files_properties(tests/inst_generator_test.cpp PROPERTIES COMPILE_FLAGS -Wno-shadow)
panda_add_gtest(
CONTAINS_MAIN
NAME compiler_inst_gen_tests
SOURCES
${PANDA_INST_GEN_TESTS_SOURCES}
LIBRARIES
${PANDA_COMPILER_TESTS_LIBRARIES}
SANITIZERS
${PANDA_SANITIZERS_LIST}
)
set(PANDA_INTRINSIC_CODEGEN_TESTS_SOURCES
tests/unit_test.cpp
tests/inst_generator.cpp
tests/intrinsic_codegen_test.cpp
)
set_source_files_properties(tests/intrinsic_codegen_test.cpp PROPERTIES COMPILE_FLAGS -Wno-shadow)
panda_add_gtest(
CONTAINS_MAIN
NAME compiler_intrinsic_codegen_arm64_tests
SOURCES
${PANDA_INTRINSIC_CODEGEN_TESTS_SOURCES}
LIBRARIES
${PANDA_COMPILER_TESTS_LIBRARIES}
SANITIZERS
${PANDA_SANITIZERS_LIST}
)
target_compile_options(compiler_intrinsic_codegen_arm64_tests PUBLIC "-DINTRINSIC_CODEGEN_TEST_ARM64")
panda_add_gtest(
CONTAINS_MAIN
NAME compiler_intrinsic_codegen_amd64_tests
SOURCES
${PANDA_INTRINSIC_CODEGEN_TESTS_SOURCES}
LIBRARIES
${PANDA_COMPILER_TESTS_LIBRARIES}
SANITIZERS
${PANDA_SANITIZERS_LIST}
)
target_compile_options(compiler_intrinsic_codegen_amd64_tests PUBLIC "-DINTRINSIC_CODEGEN_TEST_AMD64")
panda_add_gtest(
CONTAINS_MAIN
NAME compiler_intrinsic_codegen_arm32_tests
SOURCES
${PANDA_INTRINSIC_CODEGEN_TESTS_SOURCES}
LIBRARIES
${PANDA_COMPILER_TESTS_LIBRARIES}
SANITIZERS
${PANDA_SANITIZERS_LIST}
)
target_compile_options(compiler_intrinsic_codegen_arm32_tests PUBLIC "-DINTRINSIC_CODEGEN_TEST_ARM32")
endif()
endif()
if(NOT PANDA_MINIMAL_VIXL AND PANDA_WITH_TESTS AND PANDA_COMPILER_ENABLE)
add_dependencies(compiler_unit_tests ark_aot)
# TODO(igorban): Enable build in other platforms
if (PANDA_TARGET_AMD64 OR PANDA_TARGET_ARM64)
add_dependencies(compiler_unit_tests ark_aotdump)
endif()
target_include_directories(compiler_unit_tests
PUBLIC "$<TARGET_PROPERTY:arkruntime,INTERFACE_INCLUDE_DIRECTORIES>"
)
if(PANDA_TARGET_AMD64 AND PANDA_COMPILER_TARGET_AARCH64)
target_include_directories(compiler_inst_gen_tests
PUBLIC "$<TARGET_PROPERTY:arkruntime,INTERFACE_INCLUDE_DIRECTORIES>"
)
endif()
if(PANDA_TARGET_AMD64)
target_include_directories(compiler_codegen_tests
PUBLIC "$<TARGET_PROPERTY:arkruntime,INTERFACE_INCLUDE_DIRECTORIES>"
)
endif()
endif()
# Encoder gtests are separate from other, because they are test the library, not whole compiler
# Please, do not build them together - to do not make additional dependencies!
if(NOT PANDA_MINIMAL_VIXL AND PANDA_WITH_TESTS)
panda_add_gtest(
NAME encoder_unit_tests
SOURCES
${ENCODER_TESTS_SOURCES}
LIBRARIES
${PANDA_COMPILER_TESTS_LIBRARIES}
SANITIZERS
${PANDA_SANITIZERS_LIST}
)
target_compile_options(encoder_unit_tests PUBLIC "-Wno-unused-variable" "-Wno-uninitialized" "-Wno-shadow")
endif()
panda_add_sanitizers(TARGET arkcompiler SANITIZERS ${PANDA_SANITIZERS_LIST})
# Special target to create compile_commands.json in right dir.
add_check_style(".")
# Enable documentation
add_doxygen(
NAME "compiler"
PATH "."
)
# Support mobile execution
if(NOT PANDA_MINIMAL_VIXL AND PANDA_WITH_TESTS AND NOT (PANDA_TARGET_MOBILE OR PANDA_TARGET_OHOS) AND PANDA_COMPILER_ENABLE)
set(ASM_TEST_LIST
"mov"
"neg"
"abs"
"not"
"add"
"sub"
"mul"
"shl"
"shr"
"ashr"
"and"
"or"
"xor"
)
if (PANDA_TARGET_AMD64)
set(LIBGENERATED_DIR ${CMAKE_BINARY_DIR}/bin-gtests/asm_output/amd64/)
elseif (PANDA_TARGET_ARM64)
set(LIBGENERATED_DIR ${CMAKE_BINARY_DIR}/bin-gtests/asm_output/aarch64/)
elseif(PANDA_TARGET_ARM32)
set(LIBGENERATED_DIR ${CMAKE_BINARY_DIR}/bin-gtests/asm_output/aarch32/)
else()
set(LIBGENERATED_FILES "")
endif()
# Main target
add_custom_target(asm_test)
# Pseudo-target for generate asm
add_custom_target(asm_generate)
add_dependencies(asm_generate encoder_unit_tests_gtests)
foreach(TEST ${ASM_TEST_LIST})
set(TEST_ASM ${LIBGENERATED_DIR}/${TEST}.S)
add_custom_command(
OUTPUT ${TEST_ASM}
COMMAND echo " Pseudo-command for generate asm ${TEST_ASM}" > /dev/null
DEPENDS asm_generate)
set_property(SOURCE ${TEST_ASM} PROPERTY GENERATED TRUE)
add_custom_target(${TEST} DEPENDS ${TEST_ASM})
add_dependencies(asm_test ${TEST})
list(APPEND LIBGENERATED_FILES ${LIBGENERATED_DIR}/${TEST}.S)
endforeach()
enable_language(ASM)
set(ASM_CALL_TEST
tests/asm_caller.cpp
${LIBGENERATED_FILES}
)
common_add_gtest(
NAME compiler_asm_tests
SOURCES
${ASM_CALL_TEST}
LIBRARIES
${PANDA_COMPILER_TESTS_LIBRARIES}
SANITIZERS
${PANDA_SANITIZERS_LIST}
OUTPUT_DIRECTORY
${PANDA_BINARY_ROOT}/bin-gtests
)
if(PANDA_WITH_TESTS)
add_dependencies(asm_test compiler_asm_tests)
endif()
# PANDA_TARGET_MOBILE
endif()
add_subdirectory(tools/paoc ark_aot)
if (NOT PANDA_MINIMAL_VIXL AND (PANDA_TARGET_AMD64 OR PANDA_TARGET_ARM64))
add_subdirectory(tools/aotdump aotdump)
endif()
add_subdirectory(aot)
add_subdirectory(aot/aot_builder aot_builder)

22
compiler/CompilerInstTemplatesPostPlugins.cmake
View File

@ -1,22 +0,0 @@
# Copyright (c) 2021-2022 Huawei Device Co., Ltd.
# 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.
get_target_property(INST_TEMPLATES_YAML_FILES inst_templates_gen INST_TEMPLATES_YAML_FILES)
get_target_property(INST_TEMPLATES_GEN_YAML inst_templates_gen INST_TEMPLATES_GEN_YAML)
add_custom_command(OUTPUT ${INST_TEMPLATES_GEN_YAML}
COMMENT "Merge yaml files: ${INST_TEMPLATES_YAML_FILES}"
COMMAND ${PANDA_ROOT}/templates/concat_yamls.sh "${INST_TEMPLATES_GEN_YAML}" ${INST_TEMPLATES_YAML_FILES}
DEPENDS ${INST_TEMPLATES_YAML_FILES}
)
add_custom_target(inst_templates_merge DEPENDS ${INST_TEMPLATES_GEN_YAML})

172
compiler/CompilerIntrinsicsPostPlugins.cmake
View File

@ -1,172 +0,0 @@
# Copyright (c) 2021-2022 Huawei Device Co., Ltd.
# 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.
set(INTRINSICS_STUB_INL ${PANDA_BINARY_ROOT}/compiler/generated/intrinsics_stub.inl)
panda_gen_file(
DATAFILE ${GEN_PLUGIN_OPTIONS_YAML}
TEMPLATE ${PANDA_ROOT}/compiler/optimizer/templates/intrinsics/intrinsics_stub.inl.erb
REQUIRES ${PANDA_ROOT}/templates/plugin_options.rb
EXTRA_DEPENDENCIES plugin_options_merge
OUTPUTFILE ${INTRINSICS_STUB_INL}
)
set(INTRINSICS_STUB_INL_H ${PANDA_BINARY_ROOT}/compiler/generated/intrinsics_stub.inl.h)
panda_gen_file(
DATAFILE ${GEN_PLUGIN_OPTIONS_YAML}
TEMPLATE ${PANDA_ROOT}/compiler/optimizer/templates/intrinsics/intrinsics_stub.inl.h.erb
REQUIRES ${PANDA_ROOT}/templates/plugin_options.rb
EXTRA_DEPENDENCIES plugin_options_merge
OUTPUTFILE ${INTRINSICS_STUB_INL_H}
)
set(INTRINSICS_CODEGEN_EXT_INL_H ${PANDA_BINARY_ROOT}/compiler/generated/intrinsics_codegen_ext.inl.h)
panda_gen_file(
DATAFILE ${GEN_PLUGIN_OPTIONS_YAML}
TEMPLATE ${PANDA_ROOT}/compiler/optimizer/templates/intrinsics/intrinsics_codegen_ext.inl.h.erb
REQUIRES ${PANDA_ROOT}/templates/plugin_options.rb
EXTRA_DEPENDENCIES plugin_options_merge
OUTPUTFILE ${INTRINSICS_CODEGEN_EXT_INL_H}
)
set(INTRINSICS_IR_BUILD_STATIC_CALL_INL ${PANDA_BINARY_ROOT}/compiler/generated/intrinsics_ir_build_static_call.inl)
panda_gen_file(
DATAFILE ${GEN_PLUGIN_OPTIONS_YAML}
TEMPLATE ${PANDA_ROOT}/compiler/optimizer/templates/intrinsics/intrinsics_ir_build_static_call.inl.erb
REQUIRES ${PANDA_ROOT}/templates/plugin_options.rb
EXTRA_DEPENDENCIES plugin_options_merge
OUTPUTFILE ${INTRINSICS_IR_BUILD_STATIC_CALL_INL}
)
set(INTRINSICS_IR_BUILD_VIRTUAL_CALL_INL ${PANDA_BINARY_ROOT}/compiler/generated/intrinsics_ir_build_virtual_call.inl)
panda_gen_file(
DATAFILE ${GEN_PLUGIN_OPTIONS_YAML}
TEMPLATE ${PANDA_ROOT}/compiler/optimizer/templates/intrinsics/intrinsics_ir_build_virtual_call.inl.erb
REQUIRES ${PANDA_ROOT}/templates/plugin_options.rb
EXTRA_DEPENDENCIES plugin_options_merge
OUTPUTFILE ${INTRINSICS_IR_BUILD_VIRTUAL_CALL_INL}
)
set(INTRINSICS_IR_BUILD_INL_H ${PANDA_BINARY_ROOT}/compiler/generated/intrinsics_ir_build.inl.h)
panda_gen_file(
DATAFILE ${GEN_PLUGIN_OPTIONS_YAML}
TEMPLATE ${PANDA_ROOT}/compiler/optimizer/templates/intrinsics/intrinsics_ir_build.inl.h.erb
REQUIRES ${PANDA_ROOT}/templates/plugin_options.rb
EXTRA_DEPENDENCIES plugin_options_merge
OUTPUTFILE ${INTRINSICS_IR_BUILD_INL_H}
)
set(INTRINSICS_CAN_ENCODE_INL ${PANDA_BINARY_ROOT}/compiler/generated/intrinsics_can_encode.inl)
panda_gen_file(
DATAFILE ${GEN_PLUGIN_OPTIONS_YAML}
TEMPLATE ${PANDA_ROOT}/compiler/optimizer/templates/intrinsics/intrinsics_can_encode.inl.erb
REQUIRES ${PANDA_ROOT}/templates/plugin_options.rb
EXTRA_DEPENDENCIES plugin_options_merge
OUTPUTFILE ${INTRINSICS_CAN_ENCODE_INL}
)
set(IR_DYN_BASE_TYPES_H ${PANDA_BINARY_ROOT}/compiler/generated/ir-dyn-base-types.h)
panda_gen_file(
DATAFILE ${GEN_PLUGIN_OPTIONS_YAML}
TEMPLATE ${PANDA_ROOT}/compiler/optimizer/templates/ir-dyn-base-types.h.erb
REQUIRES ${PANDA_ROOT}/templates/plugin_options.rb
EXTRA_DEPENDENCIES ${YAML_FILES}
OUTPUTFILE ${IR_DYN_BASE_TYPES_H}
)
add_custom_target(ir_dyn_base_types_h DEPENDS ${IR_DYN_BASE_TYPES_H})
set(SOURCE_LANGUAGES_H ${PANDA_BINARY_ROOT}/compiler/generated/source_languages.h)
panda_gen_file(
DATAFILE ${GEN_PLUGIN_OPTIONS_YAML}
TEMPLATE ${PANDA_ROOT}/compiler/optimizer/templates/source_languages.h.erb
REQUIRES ${PANDA_ROOT}/templates/plugin_options.rb
EXTRA_DEPENDENCIES ${YAML_FILES}
OUTPUTFILE ${SOURCE_LANGUAGES_H}
)
add_custom_target(source_languages_h DEPENDS ${SOURCE_LANGUAGES_H})
set(CODEGEN_LANGUAGE_EXTENSIONS_H ${PANDA_BINARY_ROOT}/compiler/generated/codegen_language_extensions.h)
panda_gen_file(
DATAFILE ${GEN_PLUGIN_OPTIONS_YAML}
TEMPLATE ${PANDA_ROOT}/compiler/optimizer/templates/codegen_language_extensions.h.erb
REQUIRES ${PANDA_ROOT}/templates/plugin_options.rb
EXTRA_DEPENDENCIES ${YAML_FILES}
OUTPUTFILE ${CODEGEN_LANGUAGE_EXTENSIONS_H}
)
set(COMPILER_INTERFACE_EXTENSIONS_H ${PANDA_BINARY_ROOT}/compiler/generated/compiler_interface_extensions.inl.h)
panda_gen_file(
DATAFILE ${GEN_PLUGIN_OPTIONS_YAML}
TEMPLATE ${PANDA_ROOT}/compiler/optimizer/templates/compiler_interface_extensions.inl.h.erb
REQUIRES ${PANDA_ROOT}/templates/plugin_options.rb
EXTRA_DEPENDENCIES ${YAML_FILES}
OUTPUTFILE ${COMPILER_INTERFACE_EXTENSIONS_H}
)
set(INST_BUILDER_EXTENSIONS_H ${PANDA_BINARY_ROOT}/compiler/generated/inst_builder_extensions.inl.h)
panda_gen_file(
DATAFILE ${GEN_PLUGIN_OPTIONS_YAML}
TEMPLATE ${PANDA_ROOT}/compiler/optimizer/templates/inst_builder_extensions.inl.h.erb
REQUIRES ${PANDA_ROOT}/templates/plugin_options.rb
EXTRA_DEPENDENCIES ${YAML_FILES}
OUTPUTFILE ${INST_BUILDER_EXTENSIONS_H}
)
set(INTRINSICS_EXTENSIONS_H ${PANDA_BINARY_ROOT}/compiler/generated/intrinsics_extensions.inl.h)
panda_gen_file(
DATAFILE ${GEN_PLUGIN_OPTIONS_YAML}
TEMPLATE ${PANDA_ROOT}/compiler/optimizer/templates/intrinsics_extensions.inl.h.erb
REQUIRES ${PANDA_ROOT}/templates/plugin_options.rb
EXTRA_DEPENDENCIES ${YAML_FILES}
OUTPUTFILE ${INTRINSICS_EXTENSIONS_H}
)
set(INTRINSICS_INLINE_INL ${PANDA_BINARY_ROOT}/compiler/generated/intrinsics_inline.inl)
panda_gen_file(
DATAFILE ${GEN_PLUGIN_OPTIONS_YAML}
TEMPLATE ${PANDA_ROOT}/compiler/optimizer/templates/intrinsics/intrinsics_inline.inl.erb
REQUIRES ${PANDA_ROOT}/templates/plugin_options.rb
EXTRA_DEPENDENCIES plugin_options_merge
OUTPUTFILE ${INTRINSICS_INLINE_INL}
)
set(INTRINSICS_TYPES_RESOLVING_INL_H ${PANDA_BINARY_ROOT}/compiler/generated/intrinsics_types_resolving.inl.h)
panda_gen_file(
DATAFILE ${GEN_PLUGIN_OPTIONS_YAML}
TEMPLATE ${PANDA_ROOT}/compiler/optimizer/templates/intrinsics/intrinsics_types_resolving.inl.h.erb
REQUIRES ${PANDA_ROOT}/templates/plugin_options.rb
EXTRA_DEPENDENCIES plugin_options_merge
OUTPUTFILE ${INTRINSICS_TYPES_RESOLVING_INL_H}
)
add_custom_target(compiler_intrinsics DEPENDS
plugin_options_gen
${INTRINSICS_STUB_INL}
${INTRINSICS_STUB_INL_H}
${INTRINSICS_CODEGEN_EXT_INL_H}
${INTRINSICS_IR_BUILD_STATIC_CALL_INL}
${INTRINSICS_IR_BUILD_VIRTUAL_CALL_INL}
${INTRINSICS_IR_BUILD_INL_H}
${INTRINSICS_CAN_ENCODE_INL}
${IR_DYN_BASE_TYPES_H}
${SOURCE_LANGUAGES_H}
${CODEGEN_LANGUAGE_EXTENSIONS_H}
${COMPILER_INTERFACE_EXTENSIONS_H}
${INST_BUILDER_EXTENSIONS_H}
${INTRINSICS_EXTENSIONS_H}
${INTRINSICS_INLINE_INL}
${INTRINSICS_TYPES_RESOLVING_INL_H}
)
add_dependencies(arkcompiler compiler_intrinsics)

67
compiler/aot/BUILD.gn
View File

@ -1,67 +0,0 @@
# Copyright (c) 2021-2022 Huawei Device Co., Ltd.
# 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.
import("//arkcompiler/runtime_core/ark_config.gni")
config("aot_manager_config") {
include_dirs = [
"$ark_root/compiler",
"$ark_root/compiler/aot",
"$ark_root/libpandabase",
"$ark_root/libpandafile",
"$ark_root/runtime",
"$target_gen_dir/../generated",
"$target_gen_dir/include",
]
}
ohos_shared_library("libarkaotmanager") {
sources = [
"aot_file.cpp",
"aot_manager.cpp",
]
configs = [
":aot_manager_config",
"$ark_root:ark_config",
sdk_libc_secshared_config,
"$ark_root/libpandabase:arkbase_public_config",
"$ark_root/libpandafile:arkfile_public_config",
"$ark_root/runtime:arkruntime_public_config",
]
deps = [
"$ark_root/compiler:intrinsics_stub_inl_h",
"$ark_root/compiler:ir_dyn_base_types_h",
"$ark_root/compiler:libarkcompiler",
"$ark_root/compiler:libarkcompiler_intrinsics_gen_inl_generate_operations_intrinsic_graph_inl",
"$ark_root/compiler:libarkcompiler_intrinsics_gen_inl_generate_operations_intrinsic_inst_inl",
"$ark_root/compiler:libarkcompiler_intrinsics_gen_inl_get_intrinsics_inl",
"$ark_root/compiler:libarkcompiler_intrinsics_gen_inl_get_intrinsics_names_inl",
"$ark_root/compiler:libarkcompiler_intrinsics_gen_inl_intrinsic_codegen_test_inl",
"$ark_root/compiler:libarkcompiler_intrinsics_gen_inl_intrinsics_enum_inl",
"$ark_root/compiler:libarkcompiler_intrinsics_gen_inl_intrinsics_ir_build_inl",
"$ark_root/libpandabase:libarkbase",
"$ark_root/libpandafile:libarkfile",
"$ark_root/runtime:arkruntime_gen_entrypoints_compiler_checksum_entrypoints_compiler_checksum_inl",
"$ark_root/runtime:arkruntime_gen_entrypoints_entrypoints_compiler_inl",
"$ark_root/runtime:arkruntime_gen_entrypoints_entrypoints_gen_h",
"$ark_root/runtime:plugins_entrypoints_gen_h",
sdk_libc_secshared_dep,
]
output_extension = "so"
relative_install_dir = "ark"
part_name = "runtime_core"
subsystem_name = "arkcompiler"
}

41
compiler/aot/CMakeLists.txt
View File

@ -1,41 +0,0 @@
# Copyright (c) 2021-2022 Huawei Device Co., Ltd.
# 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.
project(aot_manager)
add_library(arkaotmanager SHARED aot_file.cpp aot_manager.cpp)
target_link_libraries(arkaotmanager arkbase arkfile arkcompiler)
add_dependencies(arkaotmanager intrinsics_gen_compiler isa_gen_compiler entrypoints_gen arkbase)
target_include_directories(arkaotmanager
PUBLIC ${GENERATED_DIR}
PUBLIC ${PANDA_ROOT}/runtime
)
if(PANDA_WITH_TESTS AND TARGET arkruntime_test_interpreter_impl)
target_include_directories(arkaotmanager
PUBLIC "$<TARGET_PROPERTY:arkruntime,INTERFACE_INCLUDE_DIRECTORIES>"
PUBLIC "$<TARGET_PROPERTY:arkruntime_test_interpreter_impl,INTERFACE_INCLUDE_DIRECTORIES>"
)
else()
target_include_directories(arkaotmanager
PUBLIC "$<TARGET_PROPERTY:arkruntime,INTERFACE_INCLUDE_DIRECTORIES>"
)
endif()
panda_add_sanitizers(TARGET arkaotmanager SANITIZERS ${PANDA_SANITIZERS_LIST})
panda_add_to_clang_tidy(TARGET arkaotmanager)
add_check_style(".")

46
compiler/aot/aot_builder/BUILD.gn
View File

@ -1,46 +0,0 @@
# Copyright (c) 2021-2022 Huawei Device Co., Ltd.
# 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.
import("//arkcompiler/runtime_core/ark_config.gni")
config("aotbuilder_public_config") {
include_dirs = [
"$ark_root/compiler",
"$ark_root/compiler/aot",
"$ark_root/libpandabase",
"$ark_root/libpandafile",
"$ark_root/runtime",
"$ark_root",
"$target_gen_dir/../../../libpandafile",
"$target_gen_dir/../../../libpandabase/include",
"$target_gen_dir/../../../runtime",
]
}
ohos_static_library("aotbuilder") {
sources = [ "aot_builder.cpp" ]
configs = [
"$ark_root:ark_config",
"$ark_root/compiler:arkcompiler_public_config",
":aotbuilder_public_config",
sdk_libc_secshared_config,
"$ark_root/runtime:arkruntime_public_config",
"$ark_root/libpandafile:arkfile_public_config",
]
deps = [
"$ark_root/compiler:libarkcompiler",
"$ark_root/libpandafile:isa_gen_libarkfile_bytecode_instruction_enum_gen_h",
]
}

26
compiler/aot/aot_builder/CMakeLists.txt
View File

@ -1,26 +0,0 @@
# Copyright (c) 2021-2022 Huawei Device Co., Ltd.
# 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.
project(aot_builder)
add_library(aot_builder STATIC
aot_builder.cpp
)
target_link_libraries(aot_builder arkbase arkfile arkcompiler)
panda_set_lib_32bit_property(aot_builder)
add_check_style(".")
panda_add_to_clang_tidy(TARGET aot_builder)

348
compiler/aot/aot_builder/aot_builder.cpp
View File

@ -1,348 +0,0 @@
/**
* Copyright (c) 2021-2022 Huawei Device Co., Ltd.
* 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 "aot_builder.h"
#include "aot/aot_file.h"
#include "elf_builder.h"
#include "include/class.h"
#include "include/method.h"
#include "optimizer/code_generator/encode.h"
#include "code_info/code_info.h"
#include <numeric>
namespace panda::compiler {
/**
* Fills text section in the ELF builder by the code from the methods in AotBuilder.
*/
class CodeDataProvider : public ElfSectionDataProvider {
public:
explicit CodeDataProvider(AotBuilder *aot_builder) : aot_builder_(aot_builder) {}
void FillData(Span<uint8_t> stream, size_t stream_begin) const override
{
const size_t code_offset = CodeInfo::GetCodeOffset(aot_builder_->GetArch());
CodePrefix prefix;
size_t curr_pos = stream_begin;
for (size_t i = 0; i < aot_builder_->methods_.size(); i++) {
auto &method = aot_builder_->methods_[i];
auto &method_header = aot_builder_->method_headers_[i];
prefix.code_size = method.GetCode().size();
prefix.code_info_offset = code_offset + RoundUp(method.GetCode().size(), CodeInfo::ALIGNMENT);
prefix.code_info_size = method.GetCodeInfo().size();
// Prefix
curr_pos = stream_begin + method_header.code_offset;
const char *data = reinterpret_cast<char *>(&prefix);
CopyToSpan(stream, data, sizeof(prefix), curr_pos);
curr_pos += sizeof(prefix);
// Code
curr_pos += code_offset - sizeof(prefix);
data = reinterpret_cast<const char *>(method.GetCode().data());
CopyToSpan(stream, data, method.GetCode().size(), curr_pos);
curr_pos += method.GetCode().size();
// CodeInfo
curr_pos += RoundUp(method.GetCode().size(), CodeInfo::ALIGNMENT) - method.GetCode().size();
data = reinterpret_cast<const char *>(method.GetCodeInfo().data());
CopyToSpan(stream, data, method.GetCodeInfo().size(), curr_pos);
}
}
size_t GetDataSize() const override
{
return aot_builder_->current_code_size_;
}
private:
AotBuilder *aot_builder_;
};
void AotBuilder::StartFile(const std::string &name, uint32_t checksum)
{
auto &file_header = file_headers_.emplace_back();
file_header.classes_offset = class_headers_.size();
file_header.file_checksum = checksum;
file_header.file_offset = 0;
file_header.file_name_str = AddString(name);
file_header.methods_offset = method_headers_.size();
}
void AotBuilder::EndFile()
{
ASSERT(!file_headers_.empty());
auto &file_header = file_headers_.back();
file_header.classes_count = class_headers_.size() - file_header.classes_offset;
if (file_header.classes_count == 0 && (class_hash_tables_size_.empty() || class_hash_tables_size_.back() == 0)) {
/* Just return, if there is nothing compiled in the file */
CHECK_EQ(file_header.methods_count, 0U);
file_headers_.pop_back();
return;
}
ASSERT(!class_hash_tables_size_.empty());
file_header.class_hash_table_offset =
(entity_pair_headers_.size() - class_hash_tables_size_.back()) * sizeof(panda_file::EntityPairHeader);
file_header.class_hash_table_size = class_hash_tables_size_.back();
file_header.methods_count = method_headers_.size() - file_header.methods_offset;
// We should keep class headers sorted, since AOT manager uses binary search to find classes.
std::sort(class_headers_.begin() + file_header.classes_offset, class_headers_.end(),
[](const auto &a, const auto &b) { return a.class_id < b.class_id; });
}
int AotBuilder::Write(const std::string &cmdline, const std::string &file_name)
{
switch (arch_) {
case Arch::AARCH32:
return WriteImpl<Arch::AARCH32>(cmdline, file_name);
case Arch::AARCH64:
return WriteImpl<Arch::AARCH64>(cmdline, file_name);
case Arch::X86:
return WriteImpl<Arch::X86>(cmdline, file_name);
case Arch::X86_64:
return WriteImpl<Arch::X86_64>(cmdline, file_name);
default:
LOG(ERROR, COMPILER) << "AotBuilder: Unsupported arch";
return 1;
}
}
void AotBuilder::FillHeader(const std::string &cmdline, const std::string &file_name)
{
aot_header_.magic = compiler::AotFile::MAGIC;
aot_header_.version = compiler::AotFile::VERSION;
aot_header_.checksum = 0; // TODO(msherstennikov)
ASSERT(GetRuntime() != nullptr);
aot_header_.environment_checksum = GetRuntime()->GetEnvironmentChecksum(arch_);
aot_header_.arch = static_cast<uint32_t>(arch_);
aot_header_.gc_type = gc_type_;
aot_header_.files_offset = sizeof(aot_header_);
aot_header_.files_count = file_headers_.size();
aot_header_.class_hash_tables_offset =
aot_header_.files_offset + aot_header_.files_count * sizeof(compiler::PandaFileHeader);
size_t class_hash_tables_size = entity_pair_headers_.size() * sizeof(panda_file::EntityPairHeader);
aot_header_.classes_offset = aot_header_.class_hash_tables_offset + class_hash_tables_size;
aot_header_.methods_offset = aot_header_.classes_offset + class_headers_.size() * sizeof(compiler::ClassHeader);
aot_header_.bitmap_offset = aot_header_.methods_offset + methods_.size() * sizeof(compiler::MethodHeader);
size_t bitmaps_size =
std::accumulate(class_methods_bitmaps_.begin(), class_methods_bitmaps_.end(), 0U,
[](size_t sum, const auto &vec) { return vec.GetContainerSizeInBytes() + sum; });
aot_header_.strtab_offset = aot_header_.bitmap_offset + bitmaps_size;
aot_header_.file_name_str = AddString(file_name);
aot_header_.cmdline_str = AddString(cmdline);
aot_header_.boot_aot = static_cast<uint32_t>(boot_aot_);
aot_header_.with_cha = static_cast<uint32_t>(with_cha_);
aot_header_.class_ctx_str = AddString(class_ctx_);
}
template <Arch arch>
int AotBuilder::WriteImpl(const std::string &cmdline, const std::string &file_name)
{
constexpr size_t PAGE_SIZE_BYTES = 0x1000;
constexpr size_t CALL_STATIC_SLOT_SIZE = 3;
constexpr size_t CALL_VIRTUAL_SLOT_SIZE = 2;
constexpr size_t STRING_SLOT_SIZE = 2;
constexpr size_t INLINE_CACHE_SLOT_SIZE = 1;
ElfBuilder<arch> builder;
auto text_section = builder.GetTextSection();
auto aot_section = builder.GetAotSection();
auto got_section = builder.GetGotSection();
std::vector<uint8_t> &got_data = got_section->GetVector();
// +1 is the extra slot that indicates the end of the aot table
auto got_data_size = static_cast<size_t>(RuntimeInterface::IntrinsicId::COUNT) + 1 +
CALL_STATIC_SLOT_SIZE * (got_plt_.size() + got_class_.size()) +
CALL_VIRTUAL_SLOT_SIZE * got_virt_indexes_.size() + STRING_SLOT_SIZE * got_string_.size() +
INLINE_CACHE_SLOT_SIZE * got_intf_inline_cache_.size();
// We need to fill the whole segment with aot_got section because it is filled from the end.
got_data.resize(RoundUp(PointerSize(arch) * got_data_size, PAGE_SIZE_BYTES), 0);
GenerateSymbols(builder);
FillHeader(cmdline, file_name);
aot_section->AppendData(&aot_header_, sizeof(aot_header_));
aot_section->AppendData(file_headers_.data(), file_headers_.size() * sizeof(compiler::PandaFileHeader));
aot_section->AppendData(entity_pair_headers_.data(),
entity_pair_headers_.size() * sizeof(panda_file::EntityPairHeader));
aot_section->AppendData(class_headers_.data(), class_headers_.size() * sizeof(compiler::ClassHeader));
aot_section->AppendData(method_headers_.data(), method_headers_.size() * sizeof(compiler::MethodHeader));
for (auto &bitmap : class_methods_bitmaps_) {
aot_section->AppendData(bitmap.data(), bitmap.GetContainerSizeInBytes());
}
aot_section->AppendData(string_table_.data(), string_table_.size());
CodeDataProvider code_provider(this);
text_section->SetDataProvider(&code_provider);
using PtrType = typename ArchTraits<arch>::WordType;
auto ptr_view = Span(got_data).template SubSpan<PtrType>(0, got_data.size() / sizeof(PtrType));
EmitPlt<arch>(ptr_view, got_data_size);
#ifdef PANDA_COMPILER_CFI
builder.SetFrameData(&frame_data_);
#endif
builder.Build(file_name);
builder.Write(file_name);
return 0;
}
template <Arch arch>
void AotBuilder::EmitPlt(Span<typename ArchTraits<arch>::WordType> ptr_view, size_t got_data_size)
{
if (!got_plt_.empty() || !got_virt_indexes_.empty() || !got_class_.empty() || !got_string_.empty() ||
!got_intf_inline_cache_.empty()) {
ASSERT(PointerSize(arch) >= sizeof(uint32_t));
auto ptr_cnt = ptr_view.Size();
auto end = static_cast<size_t>(RuntimeInterface::IntrinsicId::COUNT);
ptr_view[ptr_cnt - got_data_size] = 0;
constexpr size_t IMM_2 = 2;
for (auto [method, idx] : got_plt_) {
ASSERT(idx <= 0);
ptr_view[ptr_cnt - end + idx] = AotFile::AotSlotType::PLT_SLOT;
ptr_view[ptr_cnt - end + idx - IMM_2] = method.second;
}
for (auto [method, idx] : got_virt_indexes_) {
ASSERT(idx <= 0);
ptr_view[ptr_cnt - end + idx] = AotFile::AotSlotType::VTABLE_INDEX;
ptr_view[ptr_cnt - end + idx - 1] = method.second;
}
for (auto [klass, idx] : got_class_) {
ASSERT(idx <= 0);
ptr_view[ptr_cnt - end + idx] = AotFile::AotSlotType::CLASS_SLOT;
ptr_view[ptr_cnt - end + idx - IMM_2] = klass.second;
}
for (auto [string_id, idx] : got_string_) {
ASSERT(idx <= 0);
ptr_view[ptr_cnt - end + idx] = AotFile::AotSlotType::STRING_SLOT;
ptr_view[ptr_cnt - end + idx - 1] = string_id.second;
}
for (auto [cache, idx] : got_intf_inline_cache_) {
(void)cache;
ASSERT(idx < 0);
ptr_view[ptr_cnt - end + idx] = AotFile::AotSlotType::INLINECACHE_SLOT;
}
}
}
/**
* Add method names to the symbol table
*/
template <Arch arch>
void AotBuilder::GenerateSymbols(ElfBuilder<arch> &builder)
{
if (generate_symbols_) {
auto text_section = builder.GetTextSection();
size_t offset = 0;
std::string method_name;
for (auto &method : methods_) {
if (method.GetMethod()->GetPandaFile() == nullptr) {
method_name = "Error: method doesn't belong to any panda file";
} else {
auto method_casted = reinterpret_cast<RuntimeInterface::MethodPtr>(method.GetMethod());
method_name = runtime_->GetMethodFullName(method_casted, true);
}
builder.template AddSymbol<true>(
method_name, method.GetOverallSize(), *text_section, [offset, text_section]() {
return text_section->GetAddress() + offset + CodeInfo::GetCodeOffset(arch);
});
offset += RoundUp(method.GetOverallSize(), ArchTraits<arch>::CODE_ALIGNMENT);
}
}
}
void AotBuilder::AddClassHashTable(const panda_file::File &panda_file)
{
const panda_file::File::Header *header = panda_file.GetHeader();
uint32_t num_classes = header->num_classes;
if (num_classes == 0) {
return;
}
size_t hash_table_size = panda::helpers::math::GetPowerOfTwoValue32(num_classes);
std::vector<panda_file::EntityPairHeader> entity_pairs;
std::vector<unsigned int> conflict_entity_table;
entity_pairs.resize(hash_table_size);
conflict_entity_table.resize(hash_table_size);
size_t conflict_num = 0;
auto classes = panda_file.GetClasses();
for (size_t i = 0; i < num_classes; ++i) {
auto entity_id = panda_file::File::EntityId(classes[i]);
auto name = panda_file.GetStringData(entity_id).data;
uint32_t hash = GetHash32String(name);
uint32_t pos = hash & (hash_table_size - 1);
auto &entity_pair = entity_pairs[pos];
if (entity_pair.descriptor_hash == 0) {
entity_pair.descriptor_hash = hash;
entity_pair.entity_id_offset = entity_id.GetOffset();
} else {
conflict_entity_table[conflict_num] = i;
conflict_num++;
}
}
if (conflict_num == 0) {
entity_pair_headers_.insert(entity_pair_headers_.end(), entity_pairs.begin(), entity_pairs.end());
class_hash_tables_size_.emplace_back(entity_pairs.size());
} else {
ResolveConflictClassHashTable(panda_file, std::move(conflict_entity_table), conflict_num, entity_pairs);
}
}
void AotBuilder::ResolveConflictClassHashTable(const panda_file::File &panda_file,
std::vector<unsigned int> conflict_entity_table, size_t conflict_num,
std::vector<panda_file::EntityPairHeader> &entity_pairs)
{
auto classes = panda_file.GetClasses();
auto hash_table_size = entity_pairs.size();
for (size_t j = 0; j < conflict_num; ++j) {
if (j > 0 && conflict_entity_table[j - 1] == conflict_entity_table[j]) {
break; // Exit for loop if there is no conlict elements anymore
}
auto i = conflict_entity_table[j];
auto entity_id = panda_file::File::EntityId(classes[i]);
auto name = panda_file.GetStringData(entity_id).data;
uint32_t hash = GetHash32String(name);
uint32_t theory_pos = hash & (hash_table_size - 1);
ASSERT(entity_pairs[theory_pos].descriptor_hash != 0);
uint32_t actual_pos = theory_pos;
while (actual_pos < (hash_table_size - 1) && entity_pairs[actual_pos].descriptor_hash != 0) {
actual_pos++;
}
if (actual_pos == (hash_table_size - 1) && entity_pairs[actual_pos].descriptor_hash != 0) {
actual_pos = 0;
while (actual_pos < theory_pos && entity_pairs[actual_pos].descriptor_hash != 0) {
actual_pos++;
}
}
ASSERT(entity_pairs[actual_pos].descriptor_hash == 0);
auto &entity_pair = entity_pairs[actual_pos];
entity_pair.descriptor_hash = hash;
entity_pair.entity_id_offset = entity_id.GetOffset();
while (entity_pairs[theory_pos].next_pos != 0) {
theory_pos = entity_pairs[theory_pos].next_pos - 1;
}
// add 1 is to distinguish the initial value 0 of next_pos and the situation that the next pos is really 0
entity_pairs[theory_pos].next_pos = actual_pos + 1;
}
entity_pair_headers_.insert(entity_pair_headers_.end(), entity_pairs.begin(), entity_pairs.end());
class_hash_tables_size_.emplace_back(entity_pairs.size());
}
} // namespace panda::compiler

163
compiler/aot/aot_builder/aot_builder.h
View File

@ -1,163 +0,0 @@
/**
* Copyright (c) 2021-2022 Huawei Device Co., Ltd.
* 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.
*/
#ifndef COMPILER_AOT_AOT_BULDER_AOT_FILE_BUILDER_H
#define COMPILER_AOT_AOT_BULDER_AOT_FILE_BUILDER_H
#include <string>
#include <vector>
#include "aot/compiled_method.h"
#include "aot/aot_file.h"
#include "elf_builder.h"
#include "utils/arch.h"
#include "utils/arena_containers.h"
#include "utils/bit_vector.h"
#include "optimizer/ir/runtime_interface.h"
#include "mem/gc/gc_types.h"
namespace panda {
class Class;
} // namespace panda
namespace panda::compiler {
template <Arch arch, bool is_jit_mode>
class ElfBuilder;
class AotBuilder : public ElfWriter {
public:
void SetGcType(uint32_t gc_type)
{
gc_type_ = gc_type;
}
uint32_t GetGcType() const
{
return gc_type_;
}
uint64_t *GetIntfInlineCacheIndex()
{
return &intf_inline_cache_index_;
}
int Write(const std::string &cmdline, const std::string &file_name);
void StartFile(const std::string &name, uint32_t checksum);
void EndFile();
auto *GetGotPlt()
{
return &got_plt_;
}
auto *GetGotVirtIndexes()
{
return &got_virt_indexes_;
}
auto *GetGotClass()
{
return &got_class_;
}
auto *GetGotString()
{
return &got_string_;
}
auto *GetGotIntfInlineCache()
{
return &got_intf_inline_cache_;
}
void SetBootAot(bool boot_aot)
{
boot_aot_ = boot_aot;
}
void SetWithCha(bool with_cha)
{
with_cha_ = with_cha;
}
void SetGenerateSymbols(bool generate_symbols)
{
generate_symbols_ = generate_symbols;
}
void AddClassHashTable(const panda_file::File &panda_file);
void InsertEntityPairHeader(uint32_t class_hash, uint32_t class_id)
{
entity_pair_headers_.emplace_back();
auto &entity_pair = entity_pair_headers_.back();
entity_pair.descriptor_hash = class_hash;
entity_pair.entity_id_offset = class_id;
}
auto *GetEntityPairHeaders() const
{
return &entity_pair_headers_;
}
void InsertClassHashTableSize(uint32_t size)
{
class_hash_tables_size_.emplace_back(size);
}
auto *GetClassHashTableSize() const
{
return &class_hash_tables_size_;
}
private:
template <Arch arch>
int WriteImpl(const std::string &cmdline, const std::string &file_name);
template <Arch arch>
void GenerateSymbols(ElfBuilder<arch> &builder);
template <Arch arch>
void EmitPlt(Span<typename ArchTraits<arch>::WordType> ptr_view, size_t got_data_size);
void FillHeader(const std::string &cmdline, const std::string &file_name);
void ResolveConflictClassHashTable(const panda_file::File &panda_file,
std::vector<unsigned int> conflict_entity_table, size_t conflict_num,
std::vector<panda_file::EntityPairHeader> &entity_pairs);
private:
std::string file_name_;
compiler::AotHeader aot_header_ {};
uint32_t gc_type_ {static_cast<uint32_t>(mem::GCType::INVALID_GC)};
uint64_t intf_inline_cache_index_ {0};
std::map<std::pair<const panda_file::File *, uint32_t>, int32_t> got_plt_;
std::map<std::pair<const panda_file::File *, uint32_t>, int32_t> got_virt_indexes_;
std::map<std::pair<const panda_file::File *, uint32_t>, int32_t> got_class_;
std::map<std::pair<const panda_file::File *, uint32_t>, int32_t> got_string_;
std::map<std::pair<const panda_file::File *, uint64_t>, int32_t> got_intf_inline_cache_;
bool boot_aot_ {false};
bool with_cha_ {true};
bool generate_symbols_ {false};
std::vector<panda_file::EntityPairHeader> entity_pair_headers_;
std::vector<uint32_t> class_hash_tables_size_;
friend class CodeDataProvider;
friend class JitCodeDataProvider;
};
} // namespace panda::compiler
#endif // COMPILER_AOT_AOT_BULDER_AOT_FILE_BUILDER_H

1316
compiler/aot/aot_builder/elf_builder.h
View File

File diff suppressed because it is too large Load Diff

194
compiler/aot/aot_file.cpp
View File

@ -1,194 +0,0 @@
/**
* Copyright (c) 2021-2022 Huawei Device Co., Ltd.
* 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 "aot_file.h"
#include "compiler/optimizer/ir/inst.h"
#include "optimizer/ir/runtime_interface.h"
#include "utils/logger.h"
#include "code_info/code_info.h"
#include "mem/gc/gc_types.h"
#include "trace/trace.h"
#include "entrypoints/entrypoints.h"
// In some targets, runtime library is not linked, so linker will fail while searching CallStaticPltResolver symbol.
// To solve this issue, we define this function as weak.
// TODO(msherstennikov): find a better way instead of weak function, e.g. make aot_manager library static.
extern "C" void CallStaticPltResolver([[maybe_unused]] void *slot) __attribute__((weak));
extern "C" void CallStaticPltResolver([[maybe_unused]] void *slot) {}
namespace panda::compiler {
static inline Expected<const uint8_t *, std::string> LoadSymbol(const panda::os::library_loader::LibraryHandle &handle,
const char *name)
{
auto sym = panda::os::library_loader::ResolveSymbol(handle, name);
if (!sym) {
return Unexpected(sym.Error().ToString());
}
return reinterpret_cast<uint8_t *>(sym.Value());
}
// NOLINTNEXTLINE(cppcoreguidelines-macro-usage)
#define LOAD_AOT_SYMBOL(name) \
auto name = LoadSymbol(handle, #name); \
if (!name) { \
return Unexpected("Cannot load name section: " + name.Error()); \
}
Expected<std::unique_ptr<AotFile>, std::string> AotFile::Open(const std::string &file_name, uint32_t gc_type,
bool for_dump)
{
trace::ScopedTrace scoped_trace("Open aot file " + file_name);
auto handle_load = panda::os::library_loader::Load(file_name);
if (!handle_load) {
return Unexpected("AOT elf library open failed: " + handle_load.Error().ToString());
}
auto handle = std::move(handle_load.Value());
LOAD_AOT_SYMBOL(aot);
LOAD_AOT_SYMBOL(aot_end);
LOAD_AOT_SYMBOL(code);
LOAD_AOT_SYMBOL(code_end);
if (code_end.Value() < code.Value() || aot_end.Value() <= aot.Value()) {
return Unexpected(std::string("Invalid symbols"));
}
auto aot_header = reinterpret_cast<const AotHeader *>(aot.Value());
if (aot_header->magic != MAGIC) {
return Unexpected(std::string("Wrong AotHeader magic"));
}
if (aot_header->version != VERSION) {
return Unexpected(std::string("Wrong AotHeader version"));
}
if (!for_dump && aot_header->environment_checksum != RuntimeInterface::GetEnvironmentChecksum(RUNTIME_ARCH)) {
return Unexpected(std::string("Compiler environment checksum mismatch"));
}
if (!for_dump && aot_header->gc_type != gc_type) {
return Unexpected(std::string("Wrong AotHeader gc-type: ") +
std::string(mem::GCStringFromType(static_cast<mem::GCType>(aot_header->gc_type))) + " vs " +
std::string(mem::GCStringFromType(static_cast<mem::GCType>(gc_type))));
}
return std::make_unique<AotFile>(std::move(handle), Span(aot.Value(), aot_end.Value() - aot.Value()),
Span(code.Value(), code_end.Value() - code.Value()));
}
void AotFile::InitializeGot(RuntimeInterface *runtime)
{
size_t minus_first_slot = static_cast<size_t>(RuntimeInterface::IntrinsicId::COUNT) + 1;
auto *table = const_cast<uintptr_t *>(
reinterpret_cast<const uintptr_t *>(code_.data() - minus_first_slot * PointerSize(RUNTIME_ARCH)));
while (*table != 0) {
switch (*table) {
case AotSlotType::PLT_SLOT:
table -= 2U; // NOLINT(cppcoreguidelines-pro-bounds-pointer-arithmetic)
// NOLINTNEXTLINE(cppcoreguidelines-pro-bounds-pointer-arithmetic)
table[1] = reinterpret_cast<uintptr_t>(CallStaticPltResolver);
// NOLINTNEXTLINE(cppcoreguidelines-pro-bounds-pointer-arithmetic)
table[2U] = reinterpret_cast<uintptr_t>(
// NOLINTNEXTLINE(cppcoreguidelines-pro-bounds-pointer-arithmetic)
table + 1 - runtime->GetCompiledEntryPointOffset(RUNTIME_ARCH) / sizeof(uintptr_t));
break;
case AotSlotType::VTABLE_INDEX:
table--; // NOLINT(cppcoreguidelines-pro-bounds-pointer-arithmetic)
table[1] = 0; // NOLINT(cppcoreguidelines-pro-bounds-pointer-arithmetic)
break;
case AotSlotType::CLASS_SLOT:
table -= 2U; // NOLINT(cppcoreguidelines-pro-bounds-pointer-arithmetic)
table[1] = 0; // NOLINT(cppcoreguidelines-pro-bounds-pointer-arithmetic)
table[2U] = 0; // NOLINT(cppcoreguidelines-pro-bounds-pointer-arithmetic)
break;
case AotSlotType::STRING_SLOT:
table--; // NOLINT(cppcoreguidelines-pro-bounds-pointer-arithmetic)
table[1] = 0; // NOLINT(cppcoreguidelines-pro-bounds-pointer-arithmetic)
break;
case AotSlotType::INLINECACHE_SLOT:
break;
default:
UNREACHABLE();
break;
}
table--; // NOLINT(cppcoreguidelines-pro-bounds-pointer-arithmetic)
}
}
void AotFile::PatchTable(RuntimeInterface *runtime)
{
auto *table = const_cast<uintptr_t *>(reinterpret_cast<const uintptr_t *>(
code_.data() - static_cast<size_t>(RuntimeInterface::IntrinsicId::COUNT) * PointerSize(RUNTIME_ARCH)));
for (size_t i = 0; i < static_cast<size_t>(RuntimeInterface::IntrinsicId::COUNT); i++) {
IntrinsicInst inst(Opcode::Intrinsic, static_cast<RuntimeInterface::IntrinsicId>(i));
// NOLINTNEXTLINE(cppcoreguidelines-pro-bounds-pointer-arithmetic)
table[i] = runtime->GetIntrinsicAddress(inst.IsRuntimeCall(), static_cast<RuntimeInterface::IntrinsicId>(i));
}
}
AotClass AotPandaFile::GetClass(uint32_t class_id) const
{
auto classes = aot_file_->GetClassHeaders(*header_);
auto it = std::lower_bound(classes.begin(), classes.end(), class_id,
[](const auto &a, uintptr_t klass_id) { return a.class_id < klass_id; });
if (it == classes.end() || it->class_id != class_id) {
return {};
}
ASSERT(it->methods_count != 0 && "AOT file shall not contain empty classes");
return AotClass(aot_file_, &*it);
}
const void *AotClass::FindMethodCodeEntry(size_t index) const
{
auto method_header = FindMethodHeader(index);
if (method_header == nullptr) {
return nullptr;
}
// NOLINTNEXTLINE(cppcoreguidelines-pro-bounds-pointer-arithmetic)
return aot_file_->GetMethodCode(method_header) + CodeInfo::GetCodeOffset(RUNTIME_ARCH);
}
Span<const uint8_t> AotClass::FindMethodCodeSpan(size_t index) const
{
auto method_header = FindMethodHeader(index);
if (method_header == nullptr) {
return {};
}
auto code = Span(aot_file_->GetMethodCode(method_header), method_header->code_size);
return CodeInfo(code).GetCodeSpan();
}
const MethodHeader *AotClass::FindMethodHeader(size_t index) const
{
auto bitmap = GetBitmap();
CHECK_LT(index, bitmap.size());
if (!bitmap[index]) {
return nullptr;
}
auto method_index = bitmap.PopCount(index);
ASSERT(method_index < header_->methods_count);
return aot_file_->GetMethodHeader(header_->methods_offset + method_index);
}
BitVectorSpan AotClass::GetBitmap() const
{
// TODO(msherstennikov): remove const_cast once BitVector support constant storage
auto bitmap_base = const_cast<uint32_t *>(reinterpret_cast<const uint32_t *>(aot_file_->GetMethodsBitmap()));
// NOLINTNEXTLINE(cppcoreguidelines-pro-bounds-pointer-arithmetic)
return BitVectorSpan(bitmap_base + header_->methods_bitmap_offset, header_->methods_bitmap_size);
}
} // namespace panda::compiler

254
compiler/aot/aot_file.h
View File

@ -1,254 +0,0 @@
/**
* Copyright (c) 2021-2022 Huawei Device Co., Ltd.
* 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.
*/
#ifndef COMPILER_AOT_AOT_FILE_H
#define COMPILER_AOT_AOT_FILE_H
#include "aot_headers.h"
#include "compiler/code_info/code_info.h"
#include "os/library_loader.h"
#include "utils/span.h"
#include "libpandafile/file.h"
#include <string>
#include <array>
#include <memory>
#include <algorithm>
namespace panda::compiler {
class RuntimeInterface;
class AotFile {
public:
static constexpr std::array MAGIC = {'.', 'a', 'n', '\0'};
static constexpr std::array VERSION = {'0', '0', '6', '\0'};
enum AotSlotType { PLT_SLOT = 1, VTABLE_INDEX = 2, CLASS_SLOT = 3, STRING_SLOT = 4, INLINECACHE_SLOT = 5 };
AotFile(panda::os::library_loader::LibraryHandle &&handle, Span<const uint8_t> aot_data, Span<const uint8_t> code)
: handle_(std::move(handle)), aot_data_(aot_data), code_(code)
{
}
NO_MOVE_SEMANTIC(AotFile);
NO_COPY_SEMANTIC(AotFile);
~AotFile() = default;
public:
static Expected<std::unique_ptr<AotFile>, std::string> Open(const std::string &file_name, uint32_t gc_type,
bool for_dump = false);
const void *GetCode() const
{
return code_.data();
}
size_t GetCodeSize() const
{
return code_.size();
}
auto FileHeaders() const
{
return aot_data_.SubSpan<const PandaFileHeader>(GetAotHeader()->files_offset, GetFilesCount());
}
auto GetMethodHeader(size_t index) const
{
// NOLINTNEXTLINE(cppcoreguidelines-pro-bounds-pointer-arithmetic)
return reinterpret_cast<const MethodHeader *>(&aot_data_[GetAotHeader()->methods_offset]) + index;
}
const MethodHeader *GetMethodHeadersPtr() const
{
return reinterpret_cast<const MethodHeader *>(&aot_data_[GetAotHeader()->methods_offset]);
}
auto GetClassHeaders(const PandaFileHeader &file_header) const
{
return aot_data_.SubSpan<const ClassHeader>(GetAotHeader()->classes_offset +
file_header.classes_offset * sizeof(ClassHeader),
file_header.classes_count);
}
auto GetClassHashTable(const PandaFileHeader &file_header) const
{
return aot_data_.SubSpan<const panda::panda_file::EntityPairHeader>(GetAotHeader()->class_hash_tables_offset +
file_header.class_hash_table_offset,
file_header.class_hash_table_size);
}
const uint8_t *GetMethodsBitmap() const
{
return &aot_data_[GetAotHeader()->bitmap_offset];
}
size_t GetFilesCount() const
{
return GetAotHeader()->files_count;
}
const PandaFileHeader *FindPandaFile(const std::string &file_name) const
{
auto file_headers = FileHeaders();
auto res = std::find_if(file_headers.begin(), file_headers.end(), [this, &file_name](auto &header) {
return file_name == GetString(header.file_name_str);
});
return res == file_headers.end() ? nullptr : res;
}
const uint8_t *GetMethodCode(const MethodHeader *method_header) const
{
return code_.data() + method_header->code_offset;
}
const char *GetString(size_t offset) const
{
return reinterpret_cast<const char *>(aot_data_.data() + GetAotHeader()->strtab_offset + offset);
}
const AotHeader *GetAotHeader() const
{
return reinterpret_cast<const AotHeader *>(aot_data_.data());
}
const char *GetFileName() const
{
return GetString(GetAotHeader()->file_name_str);
}
const char *GetCommandLine() const
{
return GetString(GetAotHeader()->cmdline_str);
}
const char *GetClassContext() const
{
return GetString(GetAotHeader()->class_ctx_str);
}
bool IsCompiledWithCha() const
{
return GetAotHeader()->with_cha != 0U;
}
bool IsBootPandaFile() const
{
return GetAotHeader()->boot_aot != 0U;
}
void InitializeGot(RuntimeInterface *runtime);
void PatchTable(RuntimeInterface *runtime);
private:
panda::os::library_loader::LibraryHandle handle_ {nullptr};
Span<const uint8_t> aot_data_;
Span<const uint8_t> code_;
};
class AotClass final {
public:
AotClass() = default;
AotClass(const AotFile *file, const ClassHeader *header) : aot_file_(file), header_(header) {}
~AotClass() = default;
DEFAULT_COPY_SEMANTIC(AotClass);
DEFAULT_MOVE_SEMANTIC(AotClass);
const void *FindMethodCodeEntry(size_t index) const;
Span<const uint8_t> FindMethodCodeSpan(size_t index) const;
const MethodHeader *FindMethodHeader(size_t index) const;
BitVectorSpan GetBitmap() const;
auto GetMethodHeaders() const
{
// NOLINTNEXTLINE(cppcoreguidelines-pro-bounds-pointer-arithmetic)
return Span(aot_file_->GetMethodHeadersPtr() + header_->methods_offset, header_->methods_count);
}
static AotClass Invalid()
{
return AotClass();
}
bool IsValid() const
{
return header_ != nullptr;
}
private:
const AotFile *aot_file_ {nullptr};
const ClassHeader *header_ {nullptr};
};
class AotPandaFile {
public:
AotPandaFile() = default;
AotPandaFile(AotFile *file, const PandaFileHeader *header) : aot_file_(file), header_(header)
{
LoadClassHashTable();
}
DEFAULT_MOVE_SEMANTIC(AotPandaFile);
DEFAULT_COPY_SEMANTIC(AotPandaFile);
~AotPandaFile() = default;
const AotFile *GetAotFile() const
{
return aot_file_;
}
const PandaFileHeader *GetHeader()
{
return header_;
}
const PandaFileHeader *GetHeader() const
{
return header_;
}
std::string GetFileName() const
{
return GetAotFile()->GetString(GetHeader()->file_name_str);
}
AotClass GetClass(uint32_t class_id) const;
Span<const ClassHeader> GetClassHeaders() const
{
return aot_file_->GetClassHeaders(*header_);
}
CodeInfo GetMethodCodeInfo(const MethodHeader *method_header) const
{
return CodeInfo(GetAotFile()->GetMethodCode(method_header), method_header->code_size);
}
void LoadClassHashTable()
{
class_hash_table_ = GetAotFile()->GetClassHashTable(*header_);
}
panda::Span<const panda::panda_file::EntityPairHeader> GetClassHashTable() const
{
return class_hash_table_;
}
private:
AotFile *aot_file_ {nullptr};
const PandaFileHeader *header_ {nullptr};
panda::Span<const panda::panda_file::EntityPairHeader> class_hash_table_;
};
} // namespace panda::compiler
#endif // COMPILER_AOT_AOT_FILE_H

83
compiler/aot/aot_headers.h
View File

@ -1,83 +0,0 @@
/**
* Copyright (c) 2021-2022 Huawei Device Co., Ltd.
* 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.
*/
#ifndef COMPILER_AOT_AOT_HEADERS_H
#define COMPILER_AOT_AOT_HEADERS_H
#include <array>
#include <cstddef>
#include <cstdint>
namespace panda::compiler {
constexpr size_t AOT_HEADER_MAGIC_SIZE = 4;
constexpr size_t AOT_HEADER_VERSION_SIZE = 4;
struct AotHeader {
alignas(alignof(uint32_t)) std::array<char, AOT_HEADER_MAGIC_SIZE> magic;
alignas(alignof(uint32_t)) std::array<char, AOT_HEADER_VERSION_SIZE> version;
uint32_t checksum;
uint32_t environment_checksum;
uint32_t arch;
uint32_t gc_type;
uint32_t files_count;
uint32_t files_offset;
uint32_t class_hash_tables_offset;
uint32_t classes_offset;
uint32_t methods_offset;
uint32_t bitmap_offset;
uint32_t strtab_offset;
uint32_t file_name_str;
uint32_t cmdline_str;
uint32_t boot_aot;
uint32_t with_cha;
uint32_t class_ctx_str;
};
static_assert((sizeof(AotHeader) % sizeof(uint32_t)) == 0);
static_assert(alignof(AotHeader) == alignof(uint32_t));
struct PandaFileHeader {
uint32_t class_hash_table_size;
uint32_t class_hash_table_offset;
uint32_t classes_count;
uint32_t classes_offset;
uint32_t methods_count;
uint32_t methods_offset;
uint32_t file_checksum;
uint32_t file_offset;
uint32_t file_name_str;
};
struct ClassHeader {
uint32_t class_id;
uint32_t pab_offset;
uint32_t methods_count;
uint32_t methods_offset;
// Offset to the methods bitmap (aligned as uint32_t)
uint32_t methods_bitmap_offset;
// Size of bitmap in bits
uint32_t methods_bitmap_size;
};
struct MethodHeader {
uint32_t method_id;
uint32_t code_offset;
uint32_t code_size;
};
} // namespace panda::compiler
#endif // COMPILER_AOT_AOT_HEADERS_H

186
compiler/aot/aot_manager.cpp
View File

@ -1,186 +0,0 @@
/**
* Copyright (c) 2021-2022 Huawei Device Co., Ltd.
* 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 "aot_manager.h"
#include "os/filesystem.h"
#include "events/events.h"
namespace panda::compiler {
class RuntimeInterface;
Expected<bool, std::string> AotManager::AddFile(const std::string &file_name, RuntimeInterface *runtime,
uint32_t gc_type, bool force)
{
if (GetFile(file_name) != nullptr) {
LOG(DEBUG, AOT) << "Trying to add already existing AOT file: '" << file_name << "'";
return true;
}
auto aot_file = AotFile::Open(file_name, gc_type);
if (!aot_file) {
EVENT_AOT_MANAGER(file_name, events::AotManagerAction::OPEN_FAILED);
return Unexpected("AotFile::Open failed: " + aot_file.Error());
}
if (runtime != nullptr) {
aot_file.Value()->PatchTable(runtime);
aot_file.Value()->InitializeGot(runtime);
}
LOG(DEBUG, AOT) << "AOT file '" << file_name << "' has been loaded, code=" << aot_file.Value()->GetCode()
<< ", code_size=" << aot_file.Value()->GetCodeSize();
LOG(DEBUG, AOT) << " It contains the following panda files:";
for (auto header : aot_file.Value()->FileHeaders()) {
LOG(DEBUG, AOT) << " " << aot_file.Value()->GetString(header.file_name_str);
}
aot_files_.push_back(std::move(aot_file.Value()));
auto &aot_insert_file = aot_files_[aot_files_.size() - 1];
for (auto header : aot_insert_file->FileHeaders()) {
auto pf_name = aot_insert_file->GetString(header.file_name_str);
auto file_header = aot_insert_file->FindPandaFile(pf_name);
if (force) {
files_map_[pf_name] = AotPandaFile(aot_insert_file.get(), file_header);
} else {
files_map_.emplace(std::make_pair(pf_name, AotPandaFile(aot_insert_file.get(), file_header)));
}
}
EVENT_AOT_MANAGER(file_name, events::AotManagerAction::ADDED);
return true;
}
const AotPandaFile *AotManager::FindPandaFile(const std::string &file_name)
{
if (file_name.empty()) {
return nullptr;
}
auto it = files_map_.find(file_name);
if (it != files_map_.end()) {
return &it->second;
}
return nullptr;
}
const AotFile *AotManager::GetFile(const std::string &file_name) const
{
auto res = std::find_if(aot_files_.begin(), aot_files_.end(),
[&file_name](auto &file) { return file_name == file->GetFileName(); });
return res == aot_files_.end() ? nullptr : (*res).get();
}
/* We need such kind of complex print because line length of some tool is limited by 4000 characters */
static void FancyClassContextPrint(std::string_view context)
{
constexpr char DELIMITER = ':';
size_t start = 0;
size_t end = context.find(DELIMITER, start);
while (end != std::string::npos) {
LOG(ERROR, AOT) << "\t\t" << context.substr(start, end - start);
start = end + 1;
end = context.find(DELIMITER, start);
}
LOG(ERROR, AOT) << "\t\t" << context.substr(start);
}
static bool CheckFilesInClassContext(std::string_view context, std::string_view aot_context)
{
constexpr char DELIMITER = ':';
size_t start = 0;
size_t end = aot_context.find(DELIMITER, start);
while (end != std::string::npos) {
auto file_context = aot_context.substr(start, end - start);
if (context.find(file_context) == std::string::npos) {
LOG(ERROR, AOT) << "Cannot found file " << file_context << " in runtime context";
return false;
}
start = end + 1;
end = aot_context.find(DELIMITER, start);
}
return true;
}
void AotManager::VerifyClassHierarchy(bool only_boot)
{
auto complete_context = boot_class_context_;
if (!only_boot && !app_class_context_.empty()) {
if (!complete_context.empty()) {
complete_context.append(":");
}
complete_context.append(app_class_context_);
}
auto verify_aot = [this, &complete_context](auto &aot_file) {
auto context = aot_file->IsBootPandaFile() ? boot_class_context_ : complete_context;
bool is_check = true;
if (aot_file->IsCompiledWithCha()) {
// Aot file context must be prefix of current runtime context
if (context.rfind(aot_file->GetClassContext(), 0) != 0) {
is_check = false;
EVENT_AOT_MANAGER(aot_file->GetFileName(), events::AotManagerAction::CHA_VERIFY_FAILED);
}
} else {
// Aot file context must be contained in current runtime context
if (!CheckFilesInClassContext(context, aot_file->GetClassContext())) {
is_check = false;
EVENT_AOT_MANAGER(aot_file->GetFileName(), events::AotManagerAction::FILE_VERIFY_FAILED);
}
}
if (!is_check) {
auto boot_pref = aot_file->IsBootPandaFile() ? "boot " : "";
LOG(ERROR, AOT) << "Cannot use " << boot_pref << "AOT file '" << aot_file->GetFileName() << '\'';
LOG(ERROR, AOT) << "\tRuntime " << boot_pref << "class context: ";
FancyClassContextPrint(context);
LOG(ERROR, AOT) << "\tAOT class context: ";
FancyClassContextPrint(aot_file->GetClassContext());
LOG(FATAL, AOT) << "Aborting due to mismatched class hierarchy";
return true;
}
EVENT_AOT_MANAGER(aot_file->GetFileName(), events::AotManagerAction::VERIFIED);
return false;
};
for (auto &cur_aot_file : aot_files_) {
verify_aot(cur_aot_file);
}
}
void AotManager::RegisterAotStringRoot(ObjectHeader **slot, bool is_young)
{
os::memory::LockHolder lock(aot_string_roots_lock_);
aot_string_gc_roots_.push_back(slot);
// Atomic with acquire order reason: data race with aot_string_gc_roots_count_ with dependecies on reads after the
// load which should become visible
size_t roots_count = aot_string_gc_roots_count_.load(std::memory_order_acquire);
if (aot_string_young_set_.size() <= roots_count / MASK_WIDTH) {
aot_string_young_set_.push_back(0);
}
if (is_young) {
has_young_aot_string_refs_ = true;
aot_string_young_set_[roots_count / MASK_WIDTH] |= 1ULL << (roots_count % MASK_WIDTH);
}
// Atomic with acq_rel order reason: data race with aot_string_gc_roots_count_ with dependecies on reads after the
// load and on writes before the store
aot_string_gc_roots_count_.fetch_add(1, std::memory_order_acq_rel);
}
bool AotClassContextCollector::operator()(const panda_file::File &pf)
{
if (!acc_->empty()) {
acc_->append(":");
}
acc_->append(os::GetAbsolutePath(pf.GetFilename()));
acc_->append("*");
acc_->append(std::to_string(pf.GetHeader()->checksum));
return true;
}
} // namespace panda::compiler

179
compiler/aot/aot_manager.h
View File

@ -1,179 +0,0 @@
/**
* Copyright (c) 2021-2022 Huawei Device Co., Ltd.
* 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.
*/
#ifndef COMPILER_AOT_AOT_MANAGER_H
#define COMPILER_AOT_AOT_MANAGER_H
#include "aot_file.h"
#include "file.h"
#include "utils/arena_containers.h"
#include "runtime/include/mem/panda_containers.h"
#include "runtime/include/mem/panda_string.h"
#include "utils/expected.h"
namespace panda::compiler {
class RuntimeInterface;
class AotManager {
using BitSetElement = uint32_t;
static constexpr size_t MASK_WIDTH = BITS_PER_BYTE * sizeof(BitSetElement);
public:
explicit AotManager() = default;
NO_MOVE_SEMANTIC(AotManager);
NO_COPY_SEMANTIC(AotManager);
~AotManager() = default;
Expected<bool, std::string> AddFile(const std::string &file_name, RuntimeInterface *runtime, uint32_t gc_type,
bool force = false);
const AotFile *GetFile(const std::string &file_name) const;
const AotPandaFile *FindPandaFile(const std::string &file_name);
PandaString GetBootClassContext() const
{
return boot_class_context_;
}
void SetBootClassContext(PandaString context)
{
boot_class_context_ = std::move(context);
}
PandaString GetAppClassContext() const
{
return app_class_context_;
}
void SetAppClassContext(PandaString context)
{
app_class_context_ = std::move(context);
}
void VerifyClassHierarchy(bool only_boot = false);
uint32_t GetAotStringRootsCount()
{
// use counter to get roots count without acquiring vector's lock
// Atomic with acquire order reason: data race with aot_string_gc_roots_count_ with dependecies on reads after
// the load which should become visible
return aot_string_gc_roots_count_.load(std::memory_order_acquire);
}
void RegisterAotStringRoot(ObjectHeader **slot, bool is_young);
template <typename Callback>
void VisitAotStringRoots(Callback cb, bool visit_only_young)
{
ASSERT(aot_string_gc_roots_.empty() ||
(aot_string_young_set_.size() - 1) == (aot_string_gc_roots_.size() - 1) / MASK_WIDTH);
if (!visit_only_young) {
for (auto root : aot_string_gc_roots_) {
cb(root);
}
return;
}
if (!has_young_aot_string_refs_) {
return;
}
// Atomic with acquire order reason: data race with aot_string_gc_roots_count_ with dependecies on reads after
// the load which should become visible
size_t total_roots = aot_string_gc_roots_count_.load(std::memory_order_acquire);
for (size_t idx = 0; idx < aot_string_young_set_.size(); idx++) {
auto mask = aot_string_young_set_[idx];
if (mask == 0) {
continue;
}
for (size_t offset = 0; offset < MASK_WIDTH && idx * MASK_WIDTH + offset < total_roots; offset++) {
if ((mask & (1ULL << offset)) != 0) {
cb(aot_string_gc_roots_[idx * MASK_WIDTH + offset]);
}
}
}
}
template <typename Callback, typename IsYoungPredicate>
void UpdateAotStringRoots(Callback cb, IsYoungPredicate p)
{
ASSERT(aot_string_gc_roots_.empty() ||
(aot_string_young_set_.size() - 1) == (aot_string_gc_roots_.size() - 1) / MASK_WIDTH);
has_young_aot_string_refs_ = false;
size_t idx = 0;
for (auto root : aot_string_gc_roots_) {
cb(root);
uint64_t bitmask = 1ULL << (idx % MASK_WIDTH);
if ((aot_string_young_set_[idx / MASK_WIDTH] & bitmask) != 0) {
bool is_young = p(*root);
has_young_aot_string_refs_ |= is_young;
if (!is_young) {
aot_string_young_set_[idx / MASK_WIDTH] &= ~bitmask;
}
}
idx++;
}
}
bool InAotFileRange(uintptr_t pc)
{
for (auto &aot_file : aot_files_) {
auto code = reinterpret_cast<uintptr_t>(aot_file->GetCode());
if (pc >= code && pc < code + reinterpret_cast<uintptr_t>(aot_file->GetCodeSize())) {
return true;
}
}
return false;
}
bool HasAotFiles()
{
return !aot_files_.empty();
}
private:
PandaVector<std::unique_ptr<AotFile>> aot_files_;
PandaUnorderedMap<std::string, AotPandaFile> files_map_;
PandaString boot_class_context_;
PandaString app_class_context_;
os::memory::RecursiveMutex aot_string_roots_lock_;
PandaVector<ObjectHeader **> aot_string_gc_roots_;
std::atomic_uint32_t aot_string_gc_roots_count_ {0};
bool has_young_aot_string_refs_ {false};
PandaVector<BitSetElement> aot_string_young_set_;
};
class AotClassContextCollector {
public:
explicit AotClassContextCollector(PandaString *acc) : acc_(acc) {};
bool operator()(const panda_file::File &pf);
DEFAULT_MOVE_SEMANTIC(AotClassContextCollector);
DEFAULT_COPY_SEMANTIC(AotClassContextCollector);
~AotClassContextCollector() = default;
private:
PandaString *acc_;
};
} // namespace panda::compiler
#endif // COMPILER_AOT_AOT_MANAGER_H

106
compiler/aot/compiled_method.h
View File

@ -1,106 +0,0 @@
/**
* Copyright (c) 2021-2022 Huawei Device Co., Ltd.
* 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.
*/
#ifndef COMPILER_AOT_COMPILED_METHOD_H
#define COMPILER_AOT_COMPILED_METHOD_H
#include "utils/arch.h"
#include "utils/span.h"
#include "compiler/code_info/code_info.h"
#include "compiler/optimizer/code_generator/callconv.h"
#include <cstdint>
#include <vector>
namespace panda {
class Method;
} // namespace panda
namespace panda::compiler {
class CompiledMethod {
public:
CompiledMethod(Arch arch, Method *method) : arch_(arch), method_(method) {}
NO_COPY_OPERATOR(CompiledMethod);
DEFAULT_COPY_CTOR(CompiledMethod)
DEFAULT_MOVE_SEMANTIC(CompiledMethod);
~CompiledMethod() = default;
void SetCode(Span<const uint8_t> data)
{
code_.reserve(data.size());
std::copy(data.begin(), data.end(), std::back_inserter(code_));
}
void SetCodeInfo(Span<const uint8_t> data)
{
code_info_.reserve(data.size());
std::copy(data.begin(), data.end(), std::back_inserter(code_info_));
}
Method *GetMethod()
{
return method_;
}
const Method *GetMethod() const
{
return method_;
}
Span<const uint8_t> GetCode() const
{
return Span(code_);
}
Span<const uint8_t> GetCodeInfo() const
{
return Span(code_info_);
}
size_t GetOverallSize() const
{
return RoundUp(CodePrefix::STRUCT_SIZE, GetCodeAlignment(arch_)) + RoundUp(code_.size(), CodeInfo::ALIGNMENT) +
RoundUp(code_info_.size(), CodeInfo::SIZE_ALIGNMENT);
}
#ifdef PANDA_COMPILER_CFI
CfiInfo &GetCfiInfo()
{
return cfi_info_;
}
const CfiInfo &GetCfiInfo() const
{
return cfi_info_;
}
void SetCfiInfo(const CfiInfo &cfi_info)
{
cfi_info_ = cfi_info;
}
#endif
private:
Arch arch_ {RUNTIME_ARCH};
Method *method_ {nullptr};
std::vector<uint8_t> code_;
std::vector<uint8_t> code_info_;
#ifdef PANDA_COMPILER_CFI
CfiInfo cfi_info_;
#endif
};
} // namespace panda::compiler
#endif // COMPILER_AOT_COMPILED_METHOD_H

61
compiler/cmake/benchmark_coverage.cmake
View File

@ -1,61 +0,0 @@
# Copyright (c) 2021-2022 Huawei Device Co., Ltd.
# 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.
add_custom_target(benchmark_coverage DEPENDS benchmark_coverage_arm64 benchmark_coverage_x86_64 benchmark_coverage_arm)
# arm64 targets:
add_custom_target(benchmark_coverage_arm64 DEPENDS benchmark_coverage_arm64_aot benchmark_coverage_arm64_jit benchmark_coverage_arm64_osr)
add_custom_target(benchmark_coverage_arm64_aot DEPENDS benchmark_coverage_arm64_aot pandasm paoc)
add_custom_target(benchmark_coverage_arm64_jit DEPENDS benchmark_coverage_arm64_jit pandasm paoc)
add_custom_target(benchmark_coverage_arm64_osr DEPENDS benchmark_coverage_arm64_osr pandasm paoc)
# x86_64 targets:
add_custom_target(benchmark_coverage_x86_64 DEPENDS benchmark_coverage_x86_64_aot benchmark_coverage_x86_64_jit)
add_custom_target(benchmark_coverage_x86_64_aot DEPENDS benchmark_coverage_x86_64_aot pandasm paoc)
add_custom_target(benchmark_coverage_x86_64_jit DEPENDS benchmark_coverage_x86_64_jit pandasm paoc)
# arm32 targets:
add_custom_target(benchmark_coverage_arm DEPENDS benchmark_coverage_arm_jit)
add_custom_target(benchmark_coverage_arm_jit DEPENDS benchmark_coverage_arm_jit pandasm paoc)
# Build rules:
add_custom_command(TARGET benchmark_coverage_arm64_aot POST_BUILD
WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}
COMMAND bash ${PANDA_ROOT}/compiler/tools/benchmark_coverage.sh --binary-dir=${PANDA_BINARY_ROOT} --root-dir=${PANDA_ROOT} --target-arch=arm64 --paoc-mode=aot 2>/dev/null
)
add_custom_command(TARGET benchmark_coverage_arm64_jit POST_BUILD
WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}
COMMAND bash ${PANDA_ROOT}/compiler/tools/benchmark_coverage.sh --binary-dir=${PANDA_BINARY_ROOT} --root-dir=${PANDA_ROOT} --target-arch=arm64 --paoc-mode=jit 2>/dev/null
)
add_custom_command(TARGET benchmark_coverage_arm64_osr POST_BUILD
WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}
COMMAND bash ${PANDA_ROOT}/compiler/tools/benchmark_coverage.sh --binary-dir=${PANDA_BINARY_ROOT} --root-dir=${PANDA_ROOT} --target-arch=arm64 --paoc-mode=osr 2>/dev/null
)
add_custom_command(TARGET benchmark_coverage_x86_64_jit POST_BUILD
WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}
COMMAND bash ${PANDA_ROOT}/compiler/tools/benchmark_coverage.sh --binary-dir=${PANDA_BINARY_ROOT} --root-dir=${PANDA_ROOT} --target-arch=x86_64 --paoc-mode=jit 2>/dev/null
)
add_custom_command(TARGET benchmark_coverage_x86_64_aot POST_BUILD
WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}
COMMAND bash ${PANDA_ROOT}/compiler/tools/benchmark_coverage.sh --binary-dir=${PANDA_BINARY_ROOT} --root-dir=${PANDA_ROOT} --target-arch=x86_64 --paoc-mode=aot 2>/dev/null
)
# TODO (asidorov): enable after AOT support on arm32
#add_custom_command(TARGET benchmark_coverage_arm_jit POST_BUILD
# WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}
# COMMAND bash ${PANDA_ROOT}/compiler/tools/benchmark_coverage.sh --binary-dir=${PANDA_BINARY_ROOT} --root-dir=${PANDA_ROOT} --target-arch=arm --paoc-mode=jit
#)

50
compiler/cmake/coverage.cmake
View File

@ -1,50 +0,0 @@
# Copyright (c) 2021-2022 Huawei Device Co., Ltd.
# 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.
option(ENABLE_COMPILER_COVERAGE "Enable coverage-calculation for compiler" false)
find_program(
LCOV
NAMES "lcov"
DOC "Path to lcov executable")
if(NOT LCOV)
set(ENABLE_COMPILER_COVERAGE false)
endif()
find_program(
GENHTML
NAMES "genhtml"
DOC "Path to genhtml executable")
if(NOT GENHTML)
set(ENABLE_COMPILER_COVERAGE false)
endif()
if(ENABLE_COMPILER_COVERAGE)
# Set coverage options
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fprofile-arcs -ftest-coverage")
add_custom_target(coverage DEPENDS cts-assembly tests benchmarks) # Execute tests
set(ADD_COV_FLAGS --quiet --rc lcov_branch_coverage=1)
add_custom_command(TARGET coverage POST_BUILD
WORKING_DIRECTORY ${PANDA_BINARY_ROOT}
# Update current coverage info
COMMAND lcov --no-external -b ${PANDA_ROOT}/compiler -d ${CMAKE_CURRENT_BINARY_DIR} -c -o compiler_coverage.info ${ADD_COV_FLAGS}
# Generating an html report
COMMAND genhtml -o compiler_coverage_report compiler_coverage.info --ignore-errors source ${ADD_COV_FLAGS}
COMMAND echo "Coverage report: ${PANDA_BINARY_ROOT}/compiler_coverage_report"
# Delete temporary files to collect statistics
COMMAND rm compiler_coverage.info
COMMAND find ${PANDA_BINARY_ROOT}/* -iname "*.gcda" -delete
)
else()
message(STATUS "Coverage will not be calculated (may be enabled by -DENABLE_COMPILER_COVERAGE=true ).")
endif(ENABLE_COMPILER_COVERAGE)

25
compiler/cmake/ir_builder_coverage.cmake
View File

@ -1,25 +0,0 @@
# Copyright (c) 2021-2022 Huawei Device Co., Ltd.
# 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.
add_custom_target(ir_builder_coverage DEPENDS compiler_unit_tests)
add_custom_target(pbc_2_ir_doc_gen)
add_custom_command(TARGET ir_builder_coverage POST_BUILD
WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}
COMMAND bash ${PANDA_ROOT}/compiler/tools/ir_builder_coverage.sh --binary-dir=${PANDA_BINARY_ROOT} --root-dir=${PANDA_ROOT}
)
add_custom_command(TARGET pbc_2_ir_doc_gen POST_BUILD
WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}
COMMAND bash ${PANDA_ROOT}/compiler/tools/pbc_2_ir_doc_gen.sh --binary-dir=${PANDA_BINARY_ROOT} --root-dir=${PANDA_ROOT}
)

60
compiler/cmake/target.cmake
View File

@ -1,60 +0,0 @@
# Copyright (c) 2021-2022 Huawei Device Co., Ltd.
# 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.
set(PANDA_COMPILER_MULTI_TARGET OFF)
# Define function just to restrict the scope of variables
function(panda_check_compiler_multi_target)
set(compiler_any_target OFF)
foreach(target X86 X86_64 AARCH32 AARCH64)
set(target_name "PANDA_COMPILER_TARGET_${target}")
if(compiler_any_target AND ${target_name})
set(PANDA_COMPILER_MULTI_TARGET ON PARENT_SCOPE)
endif()
if(${target_name})
set(compiler_any_target ON)
endif()
endforeach()
endfunction()
panda_check_compiler_multi_target()
panda_promote_to_definitions(
PANDA_COMPILER_MULTI_TARGET
PANDA_COMPILER_CFI
)
if (PANDA_COMPILER_TARGET_AARCH32)
if (PANDA_TARGET_ARM64)
message(ERROR "Unimplemented multi-build aarch32 on aarch64-target")
endif()
if (PANDA_TARGET_X86)
message(ERROR "Unimplemented multi-build aarch32 on x86-build-target")
endif()
# ABI must be defined for build on non-arm archs
if (NOT(PANDA_TARGET_ARM32_ABI_SOFT OR PANDA_TARGET_ARM32_ABI_SOFTFP OR PANDA_TARGET_ARM32_ABI_HARD))
panda_set_flag(PANDA_TARGET_ARM32_ABI_SOFTFP)
endif()
endif()
if (PANDA_COMPILER_TARGET_AARCH64)
if (PANDA_TARGET_ARM32)
message(ERROR "Unimplemented multi-build aarch64 on aarch32-target")
endif()
if (PANDA_TARGET_X86)
message(ERROR "Unimplemented multi-build aarch64 on x86-build-target")
endif()
endif()
message(STATUS "Compiler backends status:")
message(STATUS "PANDA_COMPILER_MULTI_TARGET = ${PANDA_COMPILER_MULTI_TARGET}")
message(STATUS "PANDA_COMPILER_TARGET: X86=${PANDA_COMPILER_TARGET_X86} X86_64=${PANDA_COMPILER_TARGET_X86_64} AARCH32=${PANDA_COMPILER_TARGET_AARCH32} AARCH64=${PANDA_COMPILER_TARGET_AARCH64}")

65
compiler/code_info/code_info.cpp
View File

@ -1,65 +0,0 @@
/**
* Copyright (c) 2021-2022 Huawei Device Co., Ltd.
* 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 "code_info.h"
#include "utils/bit_memory_region-inl.h"
namespace panda::compiler {
void CodeInfo::Dump(std::ostream &stream) const
{
stream << "CodeInfo: vregs_num=" << GetHeader().GetVRegsCount() << ", frame_size=" << GetHeader().GetFrameSize()
<< std::endl;
EnumerateTables([this, &stream](size_t index, auto member) {
if (HasTable(index)) {
const auto &table = this->*member;
table.Dump(stream);
}
});
}
void CodeInfo::Dump(std::ostream &stream, const StackMap &stack_map, Arch arch) const
{
stream << "Stackmap #" << stack_map.GetRow() << ": npc=0x" << std::hex << stack_map.GetNativePcUnpacked(arch)
<< ", bpc=0x" << std::hex << stack_map.GetBytecodePc();
if (stack_map.HasInlineInfoIndex()) {
stream << ", inline_depth=" << GetInlineDepth(stack_map) + 1;
}
if (stack_map.HasRootsRegMaskIndex() || stack_map.HasRootsStackMaskIndex()) {
stream << ", roots=[";
const char *sep = "";
if (stack_map.HasRootsRegMaskIndex()) {
stream << "r:0x" << std::hex << GetRootsRegMask(stack_map);
sep = ",";
}
if (stack_map.HasRootsStackMaskIndex()) {
auto region = GetRootsStackMask(stack_map);
stream << sep << "s:" << region;
}
stream << "]";
}
if (stack_map.HasVRegMaskIndex()) {
stream << ", vregs=" << GetVRegMask(stack_map);
}
}
void CodeInfo::DumpInlineInfo(std::ostream &stream, const StackMap &stack_map, int depth) const
{
auto ii = GetInlineInfo(stack_map, depth);
stream << "InlineInfo #" << depth << ": bpc=0x" << std::hex << ii.GetBytecodePc() << std::dec
<< ", vregs_num: " << ii.GetVRegsCount();
}
} // namespace panda::compiler

610
compiler/code_info/code_info.h
View File

@ -1,610 +0,0 @@
/**
* Copyright (c) 2021-2022 Huawei Device Co., Ltd.
* 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.
*/
#ifndef PANDA_CODE_INFO_H
#define PANDA_CODE_INFO_H
#include "code_info_tables.h"
#include "utils/bit_field.h"
#include "utils/bit_table.h"
#include "utils/cframe_layout.h"
#include "utils/small_vector.h"
#include "utils/span.h"
namespace panda::compiler {
/*
*
* Compiled code layout:
* +-------------+
* | CodePrefix |
* +-------------+ <- Method::CompiledCodeEntrypoint
* | Code |
* +-------------+-----------------+
* | CodeInfo | CodeInfoHeader |
* | |-----------------+----------------------+
* | | | StackMap |
* | | | InlineInfo |
* | | | Roots Reg Mask |
* | | | Roots Stack Mask |
* | | Bit Tables | Method indexes |
* | | | VRegs mask |
* | | | VRegs map |
* | | | VRegs catalogue |
* | | | Implicit Nullchecks |
* | | | Constants |
* |-------------+-----------------+----------------------+
*/
struct CodePrefix {
static constexpr uint32_t MAGIC = 0xaccadeca;
uint32_t magic {MAGIC};
uint32_t code_size {};
uint32_t code_info_offset {};
uint32_t code_info_size {};
static constexpr size_t STRUCT_SIZE = 16;
};
static_assert(sizeof(CodePrefix) == CodePrefix::STRUCT_SIZE);
class CodeInfoHeader {
public:
enum Elements { PROPERTIES, CALLEE_REG_MASK, CALLEE_FP_REG_MASK, TABLE_MASK, VREGS_COUNT, SIZE };
void SetFrameSize(uint32_t size)
{
ASSERT(MinimumBitsToStore(size) <= FRAME_SIZE_FIELD_WIDTH);
FieldFrameSize::Set(size, &data_[PROPERTIES]);
}
uint32_t GetFrameSize() const
{
return FieldFrameSize::Get(data_[PROPERTIES]);
}
void SetCalleeRegMask(uint32_t value)
{
data_[CALLEE_REG_MASK] = value;
}
uint32_t GetCalleeRegMask() const
{
return data_[CALLEE_REG_MASK];
}
void SetCalleeFpRegMask(uint32_t value)
{
data_[CALLEE_FP_REG_MASK] = value;
}
uint32_t GetCalleeFpRegMask() const
{
return data_[CALLEE_FP_REG_MASK];
}
void SetTableMask(uint32_t value)
{
data_[TABLE_MASK] = value;
}
uint32_t GetTableMask() const
{
return data_[TABLE_MASK];
}
void SetVRegsCount(uint32_t value)
{
data_[VREGS_COUNT] = value;
}
uint32_t GetVRegsCount() const
{
return data_[VREGS_COUNT];
}
void SetHasFloatRegs(bool value)
{
HasFloatRegsFlag::Set(value, &data_[PROPERTIES]);
}
bool HasFloatRegs() const
{
return HasFloatRegsFlag::Get(data_[PROPERTIES]);
}
template <typename Container>
void Encode(BitMemoryStreamOut<Container> &out)
{
VarintPack::Write(out, data_);
}
void Decode(BitMemoryStreamIn *in)
{
data_ = VarintPack::Read<SIZE>(in);
}
private:
std::array<uint32_t, SIZE> data_;
static constexpr size_t FRAME_SIZE_FIELD_WIDTH = 16;
using FieldFrameSize = BitField<uint32_t, 0, FRAME_SIZE_FIELD_WIDTH>;
using HasFloatRegsFlag = FieldFrameSize::NextFlag;
};
class CodeInfo final {
public:
static constexpr size_t TABLES_COUNT = 10;
static constexpr size_t VREG_LIST_STATIC_SIZE = 16;
static constexpr size_t ALIGNMENT = sizeof(uint64_t);
static constexpr size_t SIZE_ALIGNMENT = sizeof(uint64_t);
template <typename Allocator>
using VRegList = SmallVector<VRegInfo, VREG_LIST_STATIC_SIZE, Allocator, true>;
NO_COPY_SEMANTIC(CodeInfo);
NO_MOVE_SEMANTIC(CodeInfo);
CodeInfo() = default;
CodeInfo(const void *data, size_t size)
: CodeInfo(Span<const uint8_t>(reinterpret_cast<const uint8_t *>(data), size))
{
}
explicit CodeInfo(Span<const uint8_t> code) : CodeInfo(code.data())
{
ASSERT(GetDataSize() <= code.size());
}
explicit CodeInfo(Span<uint8_t> code) : CodeInfo(code.data())
{
ASSERT(GetDataSize() <= code.size());
}
explicit CodeInfo(const void *code_entry)
{
ASSERT(code_entry != nullptr);
auto prefix = reinterpret_cast<const CodePrefix *>(code_entry);
ASSERT(prefix->magic == CodePrefix::MAGIC);
data_ = Span(reinterpret_cast<const uint8_t *>(code_entry), prefix->code_info_offset + prefix->code_info_size);
auto code_info = Span<const uint8_t>(&data_[prefix->code_info_offset], prefix->code_info_size);
Decode(code_info);
}
virtual ~CodeInfo() = default;
static const void *GetCodeOriginFromEntryPoint(const void *data)
{
return reinterpret_cast<const void *>(reinterpret_cast<uintptr_t>(data) -
CodeInfo::GetCodeOffset(RUNTIME_ARCH));
}
static CodeInfo CreateFromCodeEntryPoint(const void *data)
{
ASSERT(data != nullptr);
// NOLINTNEXTLINE(cppcoreguidelines-pro-bounds-pointer-arithmetic)
return CodeInfo(reinterpret_cast<const uint8_t *>(data) - CodeInfo::GetCodeOffset(RUNTIME_ARCH));
}
void Decode(Span<const uint8_t> code_info)
{
BitMemoryStreamIn stream(const_cast<uint8_t *>(code_info.data()), code_info.size() * BITS_PER_BYTE);
header_.Decode(&stream);
EnumerateTables([this, &stream](size_t index, auto member) {
if (HasTable(index)) {
(this->*member).Decode(&stream);
}
});
}
const CodeInfoHeader &GetHeader() const
{
return header_;
}
CodeInfoHeader &GetHeader()
{
return header_;
}
const CodePrefix *GetPrefix() const
{
return reinterpret_cast<const CodePrefix *>(data_.data());
}
uint32_t GetFrameSize() const
{
return GetHeader().GetFrameSize();
}
const uint8_t *GetData()
{
return data_.data();
}
size_t GetDataSize()
{
return data_.size();
}
const uint8_t *GetCode() const
{
return &data_[CodeInfo::GetCodeOffset(RUNTIME_ARCH)];
}
size_t GetCodeSize() const
{
return GetPrefix()->code_size;
}
Span<const uint8_t> GetCodeSpan() const
{
return {&data_[CodeInfo::GetCodeOffset(RUNTIME_ARCH)], GetCodeSize()};
}
size_t GetInfoSize() const
{
return GetPrefix()->code_info_size;
}
bool HasTable(size_t index) const
{
return (GetHeader().GetTableMask() & (1U << index)) != 0;
}
std::variant<void *, uint32_t> GetMethod(const StackMap &stackmap, int inline_depth)
{
ASSERT(inline_depth >= 0);
auto inline_info = inline_infos_.GetRow(stackmap.GetInlineInfoIndex() + inline_depth);
if (inline_info.HasMethodLow()) {
if constexpr (ArchTraits<RUNTIME_ARCH>::IS_64_BITS) {
uintptr_t val =
inline_info.GetMethodLow() | (static_cast<uint64_t>(inline_info.GetMethodHi()) << BITS_PER_UINT32);
return reinterpret_cast<void *>(val);
} else {
return reinterpret_cast<void *>(inline_info.GetMethodLow());
}
}
return method_ids_.GetRow(inline_info.GetMethodIdIndex()).GetId();
}
uint64_t GetConstant(const VRegInfo &vreg) const
{
ASSERT(vreg.GetLocation() == VRegInfo::Location::CONSTANT);
uint64_t low = constant_table_.GetRow(vreg.GetConstantLowIndex()).GetValue();
uint64_t hi = constant_table_.GetRow(vreg.GetConstantHiIndex()).GetValue();
return low | (hi << BITS_PER_UINT32);
}
static size_t GetCodeOffset(Arch arch)
{
return RoundUp(CodePrefix::STRUCT_SIZE, GetCodeAlignment(arch));
}
uint32_t GetSavedCalleeRegsMask(bool is_fp) const
{
return is_fp ? GetHeader().GetCalleeFpRegMask() : GetHeader().GetCalleeRegMask();
}
auto GetVRegMask(const StackMap &stack_map)
{
return stack_map.HasVRegMaskIndex() ? vreg_masks_.GetBitMemoryRegion(stack_map.GetVRegMaskIndex())
: BitMemoryRegion<const uint8_t>();
}
auto GetVRegMask(const StackMap &stack_map) const
{
return const_cast<CodeInfo *>(this)->GetVRegMask(stack_map);
}
size_t GetVRegCount(const StackMap &stack_map) const
{
return GetVRegMask(stack_map).Popcount();
}
uint32_t GetRootsRegMask(const StackMap &stack_map) const
{
return stack_map.HasRootsRegMaskIndex() ? roots_reg_masks_.GetRow(stack_map.GetRootsRegMaskIndex()).GetMask()
: 0;
}
auto GetRootsStackMask(const StackMap &stack_map) const
{
return stack_map.HasRootsStackMaskIndex()
? roots_stack_masks_.GetBitMemoryRegion(stack_map.GetRootsStackMaskIndex())
: BitMemoryRegion<const uint8_t>();
}
auto GetInlineInfos(const StackMap &stack_map)
{
if (!stack_map.HasInlineInfoIndex()) {
return inline_infos_.GetRangeReversed(0, 0);
}
auto index = stack_map.GetInlineInfoIndex();
uint32_t size = index;
for (; inline_infos_.GetRow(size).GetIsLast() == 0; size++) {
}
return inline_infos_.GetRangeReversed(index, helpers::ToSigned(size) + 1);
}
auto GetInlineInfo(const StackMap &stack_map, int inline_depth) const
{
ASSERT(stack_map.HasInlineInfoIndex());
CHECK_GE(GetInlineDepth(stack_map), inline_depth);
return inline_infos_.GetRow(stack_map.GetInlineInfoIndex() + inline_depth);
}
int GetInlineDepth(const StackMap &stack_map) const
{
if (!stack_map.HasInlineInfoIndex()) {
return -1;
}
int index = stack_map.GetInlineInfoIndex();
int depth = index;
for (; inline_infos_.GetRow(depth).GetIsLast() == 0; depth++) {
}
return depth - index;
}
StackMap FindStackMapForNativePc(uint32_t pc, Arch arch = RUNTIME_ARCH) const
{
auto it =
std::lower_bound(stack_maps_.begin(), stack_maps_.end(), pc, [arch](const auto &a, uintptr_t counter) {
return a.GetNativePcUnpacked(arch) < counter;
});
return (it == stack_maps_.end() || it->GetNativePcUnpacked(arch) != pc) ? stack_maps_.GetInvalidRow() : *it;
}
StackMap FindOsrStackMap(uint32_t pc) const
{
auto it = std::find_if(stack_maps_.begin(), stack_maps_.end(),
[pc](const auto &a) { return a.GetBytecodePc() == pc && a.IsOsr(); });
return it == stack_maps_.end() ? stack_maps_.GetInvalidRow() : *it;
}
auto GetStackMap(size_t index) const
{
return StackMap(&stack_maps_, index);
}
auto &GetStackMaps()
{
return stack_maps_;
}
auto &GetVRegCatalogue()
{
return vregs_catalogue_;
}
auto &GetVRegMapTable()
{
return vregs_map_;
}
auto &GetVRegMaskTable()
{
return vreg_masks_;
}
auto &GetInlineInfosTable()
{
return inline_infos_;
}
auto &GetConstantTable()
{
return constant_table_;
}
const auto &GetImplicitNullChecksTable() const
{
return implicit_nullchecks_;
}
bool HasFloatRegs() const
{
return GetHeader().HasFloatRegs();
}
template <typename Func>
static void EnumerateTables(Func func)
{
size_t index = 0;
func(index++, &CodeInfo::stack_maps_);
func(index++, &CodeInfo::inline_infos_);
func(index++, &CodeInfo::roots_reg_masks_);
func(index++, &CodeInfo::roots_stack_masks_);
func(index++, &CodeInfo::method_ids_);
func(index++, &CodeInfo::vreg_masks_);
func(index++, &CodeInfo::vregs_map_);
func(index++, &CodeInfo::vregs_catalogue_);
func(index++, &CodeInfo::implicit_nullchecks_);
func(index++, &CodeInfo::constant_table_);
ASSERT(index == TABLES_COUNT);
}
template <typename Callback>
void EnumerateStaticRoots(const StackMap &stack_map, Callback callback)
{
return EnumerateRoots<Callback, false>(stack_map, callback);
}
template <typename Callback>
void EnumerateDynamicRoots(const StackMap &stack_map, Callback callback)
{
return EnumerateRoots<Callback, true>(stack_map, callback);
}
template <typename Allocator>
VRegList<Allocator> GetVRegList(StackMap stack_map, uint32_t first_vreg, uint32_t vregs_count,
Allocator *allocator = nullptr) const
{
if (vregs_count == 0 || !stack_map.HasRegMap()) {
return CodeInfo::VRegList<Allocator>(allocator);
}
VRegList<Allocator> vreg_list(allocator);
vreg_list.resize(vregs_count, VRegInfo());
ASSERT(!vreg_list[0].IsLive());
std::vector<bool> reg_set(vregs_count);
uint32_t remaining_registers = vregs_count;
for (int sindex = stack_map.GetRow(); sindex >= 0 && remaining_registers > 0; sindex--) {
stack_map = GetStackMap(sindex);
if (!stack_map.HasVRegMaskIndex()) {
continue;
}
// Skip stackmaps that are not in the same inline depth
auto vreg_mask = GetVRegMask(stack_map);
if (vreg_mask.Size() <= first_vreg) {
continue;
}
ASSERT(stack_map.HasVRegMapIndex());
uint32_t map_index = stack_map.GetVRegMapIndex();
map_index += vreg_mask.Popcount(0, first_vreg);
vreg_mask = vreg_mask.Subregion(first_vreg, vreg_mask.Size() - first_vreg);
uint32_t end = std::min<uint32_t>(vreg_mask.Size(), vregs_count);
for (size_t i = 0; i < end; i += BITS_PER_UINT32) {
uint32_t mask = vreg_mask.Read(i, std::min<uint32_t>(end - i, BITS_PER_UINT32));
while (mask != 0) {
uint32_t reg_idx = Ctz(mask);
if (!reg_set[i + reg_idx]) {
auto vreg_index = vregs_map_.GetRow(map_index);
if (vreg_index.GetIndex() != StackMap::NO_VALUE) {
ASSERT(!vreg_list[i + reg_idx].IsLive());
vreg_list[i + reg_idx] = vregs_catalogue_.GetRow(vreg_index.GetIndex()).GetVRegInfo();
vreg_list[i + reg_idx].SetIndex(i + reg_idx);
}
remaining_registers--;
reg_set[i + reg_idx] = true;
}
map_index++;
mask ^= 1U << reg_idx;
}
}
}
return vreg_list;
}
template <typename Allocator>
VRegList<Allocator> GetVRegList(StackMap stack_map, int inline_depth, Allocator *allocator = nullptr) const
{
if (inline_depth < 0) {
return GetVRegList<Allocator>(stack_map, 0, GetHeader().GetVRegsCount(), allocator);
}
ASSERT(stack_map.HasInlineInfoIndex());
auto inline_info = GetInlineInfo(stack_map, inline_depth);
if (inline_info.GetVRegsCount() == 0) {
return VRegList<Allocator>(allocator);
}
auto depth = inline_info.GetRow() - stack_map.GetInlineInfoIndex();
uint32_t first =
depth == 0 ? GetHeader().GetVRegsCount() : inline_infos_.GetRow(inline_info.GetRow() - 1).GetVRegsCount();
ASSERT(inline_info.GetVRegsCount() >= first);
return GetVRegList<Allocator>(stack_map, first, inline_info.GetVRegsCount() - first, allocator);
}
template <typename Allocator>
VRegList<Allocator> GetVRegList(StackMap stack_map, Allocator *allocator = nullptr) const
{
return GetVRegList<Allocator>(stack_map, -1, allocator);
}
static bool VerifyCompiledEntry(uintptr_t compiled_entry)
{
auto codeheader = compiled_entry - GetCodeOffset(RUNTIME_ARCH);
return (*reinterpret_cast<const uint32_t *>(codeheader) == CodePrefix::MAGIC);
}
void Dump(std::ostream &stream) const;
void Dump(std::ostream &stream, const StackMap &stack_map, Arch arch = RUNTIME_ARCH) const;
void DumpInlineInfo(std::ostream &stream, const StackMap &stack_map, int depth) const;
size_t CountSpillSlots()
{
auto frame_slots = GetFrameSize() / PointerSize(RUNTIME_ARCH);
auto spills_count = frame_slots - (static_cast<size_t>(CFrameSlots::Start()) + GetRegsCount(RUNTIME_ARCH) + 1U);
// Reverse 'CFrameLayout::AlignSpillCount' counting
if (RUNTIME_ARCH == Arch::AARCH32) {
spills_count = spills_count / 2U - 1;
}
if (spills_count % 2U != 0) {
spills_count--;
}
return spills_count;
}
private:
template <typename Callback, bool is_dynamic>
void EnumerateRoots(const StackMap &stack_map, Callback callback);
BitTable<StackMap> stack_maps_;
BitTable<InlineInfo> inline_infos_;
BitTable<RegisterMask> roots_reg_masks_;
BitTable<StackMask> roots_stack_masks_;
BitTable<MethodId> method_ids_;
BitTable<VRegisterInfo> vregs_catalogue_;
BitTable<VRegisterCatalogueIndex> vregs_map_;
BitTable<VRegisterMask> vreg_masks_;
BitTable<ImplicitNullChecks> implicit_nullchecks_;
BitTable<ConstantTable> constant_table_;
CodeInfoHeader header_ {};
Span<const uint8_t> data_;
};
template <typename Callback, bool is_dynamic>
void CodeInfo::EnumerateRoots(const StackMap &stack_map, Callback callback)
{
auto root_type = is_dynamic ? VRegInfo::Type::ANY : VRegInfo::Type::OBJECT;
if (stack_map.HasRootsRegMaskIndex()) {
auto reg_mask = roots_reg_masks_.GetRow(stack_map.GetRootsRegMaskIndex()).GetMask();
ArenaBitVectorSpan vec(&reg_mask, BITS_PER_UINT32);
for (auto reg_idx : vec.GetSetBitsIndices()) {
if (!callback(VRegInfo(reg_idx, VRegInfo::Location::REGISTER, root_type, false))) {
return;
}
}
}
// Simplify after renumbering stack slots
if (stack_map.HasRootsStackMaskIndex()) {
auto stack_slots_count = CountSpillSlots();
auto reg_mask = roots_stack_masks_.GetBitMemoryRegion(stack_map.GetRootsStackMaskIndex());
for (auto reg_idx : reg_mask) {
if (reg_idx >= stack_slots_count) {
// Parameter-slots' indexes are added to the root-mask with `stack_slots_count` offset to distinct them
// from spill-slots
auto param_slot_idx = reg_idx - stack_slots_count;
reg_idx = static_cast<size_t>(CFrameLayout::StackArgSlot::Start()) - param_slot_idx -
static_cast<size_t>(CFrameSlots::Start());
} else {
if constexpr (!ArchTraits<RUNTIME_ARCH>::IS_64_BITS) { // NOLINT
reg_idx = (reg_idx << 1U) + 1;
}
// Stack roots are began from spill/fill stack origin, so we need to adjust it according to registers
// buffer
reg_idx += GetRegsCount(RUNTIME_ARCH);
}
VRegInfo vreg(reg_idx, VRegInfo::Location::SLOT, root_type, false);
if (!callback(vreg)) {
return;
}
}
}
}
} // namespace panda::compiler
#endif // PANDA_CODE_INFO_H

232
compiler/code_info/code_info_builder.cpp
View File

@ -1,232 +0,0 @@
/**
* Copyright (c) 2021-2022 Huawei Device Co., Ltd.
* 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 "code_info_builder.h"
#include "utils/bit_memory_region-inl.h"
namespace panda::compiler {
void CodeInfoBuilder::BeginMethod(uint32_t frame_size, uint32_t vregs_count)
{
#ifndef NDEBUG
ASSERT(!was_method_begin_);
ASSERT(!was_stack_map_begin_);
ASSERT(!was_inline_info_begin_);
was_method_begin_ = true;
#endif
header_.SetFrameSize(frame_size);
vregs_count_ = vregs_count;
constant_table_.Add({0});
}
void CodeInfoBuilder::EndMethod()
{
#ifndef NDEBUG
ASSERT(was_method_begin_);
ASSERT(!was_stack_map_begin_);
ASSERT(!was_inline_info_begin_);
was_method_begin_ = false;
#endif
}
void CodeInfoBuilder::BeginStackMap(uint32_t bpc, uint32_t npc, ArenaBitVector *stack_roots, uint32_t regs_roots,
bool require_vreg_map, bool is_osr)
{
#ifndef NDEBUG
ASSERT(was_method_begin_);
ASSERT(!was_stack_map_begin_);
ASSERT(!was_inline_info_begin_);
was_stack_map_begin_ = true;
#endif
inline_info_stack_.clear();
current_vregs_.clear();
ASSERT(stack_maps_.GetSize() == 0 || npc >= stack_maps_.GetLast()[StackMap::COLUMN_NATIVE_PC]);
current_vregs_count_ = require_vreg_map ? vregs_count_ : 0;
current_stack_map_ = BitTableBuilder<StackMap>::Entry();
current_stack_map_[StackMap::COLUMN_PROPERTIES] = StackMap::CreateProperties(is_osr, require_vreg_map);
current_stack_map_[StackMap::COLUMN_BYTECODE_PC] = bpc;
current_stack_map_[StackMap::COLUMN_NATIVE_PC] = StackMap::PackAddress(npc, arch_);
if (regs_roots != 0) {
current_stack_map_[StackMap::COLUMN_ROOTS_REG_MASK_INDEX] = roots_reg_masks_.Add({regs_roots});
}
if (stack_roots != nullptr && !stack_roots->empty()) {
current_stack_map_[StackMap::COLUMN_ROOTS_STACK_MASK_INDEX] = roots_stack_masks_.Add(stack_roots->GetFixed());
}
// Ensure that stackmaps are inserted in sorted order
if (stack_maps_.GetRowsCount() != 0) {
ASSERT(current_stack_map_[StackMap::COLUMN_NATIVE_PC] >= stack_maps_.GetLast()[StackMap::COLUMN_NATIVE_PC]);
}
}
void CodeInfoBuilder::EndStackMap()
{
#ifndef NDEBUG
ASSERT(was_method_begin_);
ASSERT(was_stack_map_begin_);
ASSERT(!was_inline_info_begin_);
was_stack_map_begin_ = false;
#endif
if (!inline_info_stack_.empty()) {
inline_info_stack_.back()[InlineInfo::COLUMN_IS_LAST] = static_cast<uint32_t>(true);
current_stack_map_[StackMap::COLUMN_INLINE_INFO_INDEX] = inline_infos_.AddArray(Span(inline_info_stack_));
}
EmitVRegs();
stack_maps_.Add(current_stack_map_);
}
void CodeInfoBuilder::DumpCurrentStackMap(std::ostream &stream) const
{
stream << "Stackmap #" << stack_maps_.GetRowsCount() - 1 << ": npc=0x" << std::hex
<< StackMap::UnpackAddress(current_stack_map_[StackMap::COLUMN_NATIVE_PC], arch_) << ", bpc=0x" << std::hex
<< current_stack_map_[StackMap::COLUMN_BYTECODE_PC];
if (current_stack_map_[StackMap::COLUMN_INLINE_INFO_INDEX] != StackMap::NO_VALUE) {
stream << ", inline_depth=" << inline_info_stack_.size();
}
if (current_stack_map_[StackMap::COLUMN_ROOTS_REG_MASK_INDEX] != StackMap::NO_VALUE ||
current_stack_map_[StackMap::COLUMN_ROOTS_STACK_MASK_INDEX] != StackMap::NO_VALUE) {
stream << ", roots=[";
const char *sep = "";
if (current_stack_map_[StackMap::COLUMN_ROOTS_REG_MASK_INDEX] != StackMap::NO_VALUE) {
auto &entry = roots_reg_masks_.GetEntry(current_stack_map_[StackMap::COLUMN_ROOTS_REG_MASK_INDEX]);
stream << "r:0x" << std::hex << entry[RegisterMask::COLUMN_MASK];
sep = ",";
}
if (current_stack_map_[StackMap::COLUMN_ROOTS_STACK_MASK_INDEX] != StackMap::NO_VALUE) {
auto region = roots_stack_masks_.GetEntry(current_stack_map_[StackMap::COLUMN_ROOTS_STACK_MASK_INDEX]);
stream << sep << "s:" << region;
}
stream << "]";
}
if (current_stack_map_[StackMap::COLUMN_VREG_MASK_INDEX] != StackMap::NO_VALUE) {
stream << ", vregs=" << vreg_masks_.GetEntry(current_stack_map_[StackMap::COLUMN_VREG_MASK_INDEX]);
}
}
void CodeInfoBuilder::BeginInlineInfo(void *method, uint32_t method_id, uint32_t bpc, uint32_t vregs_count)
{
#ifndef NDEBUG
ASSERT(was_method_begin_);
ASSERT(was_stack_map_begin_);
was_inline_info_begin_ = true;
#endif
BitTableBuilder<InlineInfo>::Entry inline_info;
current_vregs_count_ += vregs_count;
inline_info[InlineInfo::COLUMN_IS_LAST] = static_cast<uint32_t>(false);
inline_info[InlineInfo::COLUMN_BYTECODE_PC] = bpc;
inline_info[InlineInfo::COLUMN_VREGS_COUNT] = current_vregs_count_;
if (method != nullptr) {
inline_info[InlineInfo::COLUMN_METHOD_HI] = High32Bits(method);
inline_info[InlineInfo::COLUMN_METHOD_LOW] = Low32Bits(method);
} else {
ASSERT(method_id != 0);
inline_info[InlineInfo::COLUMN_METHOD_ID_INDEX] = method_ids_.Add({method_id});
}
inline_info_stack_.push_back(inline_info);
}
void CodeInfoBuilder::EndInlineInfo()
{
#ifndef NDEBUG
ASSERT(was_method_begin_);
ASSERT(was_stack_map_begin_);
ASSERT(was_inline_info_begin_);
was_inline_info_begin_ = false;
#endif
ASSERT(current_vregs_.size() == current_vregs_count_);
}
void CodeInfoBuilder::AddConstant(uint64_t value, VRegInfo::Type type, bool is_acc)
{
VRegInfo vreg(0, VRegInfo::Location::CONSTANT, type, is_acc);
uint32_t low = value & ((1LLU << BITS_PER_UINT32) - 1);
uint32_t hi = (value >> BITS_PER_UINT32) & ((1LLU << BITS_PER_UINT32) - 1);
vreg.SetConstantIndices(constant_table_.Add({low}), constant_table_.Add({hi}));
current_vregs_.push_back(vreg);
}
void CodeInfoBuilder::EmitVRegs()
{
ASSERT(current_vregs_.size() == current_vregs_count_);
if (current_vregs_.empty()) {
return;
}
if (current_vregs_.size() > last_vregs_.size()) {
last_vregs_.resize(current_vregs_.size(), VRegInfo::Invalid());
vregs_last_change_.resize(current_vregs_.size());
}
ArenaVector<BitTableBuilder<VRegisterCatalogueIndex>::Entry> &vregs_map = vregs_map_storage_;
ArenaBitVector &vregs_mask = vregs_mask_storage_;
vregs_map.clear();
vregs_mask.clear();
for (size_t i = 0; i < current_vregs_.size(); i++) {
auto &vreg = current_vregs_[i];
uint32_t distatnce = stack_maps_.GetRowsCount() - vregs_last_change_[i];
if (last_vregs_[i] != vreg || distatnce > MAX_VREG_LIVE_DISTANCE) {
BitTableBuilder<VRegisterInfo>::Entry vreg_entry;
vreg_entry[VRegisterInfo::COLUMN_INFO] = vreg.GetInfo();
vreg_entry[VRegisterInfo::COLUMN_VALUE] = vreg.GetValue();
uint32_t index = vreg.IsLive() ? vregs_catalogue_.Add(vreg_entry) : decltype(vregs_catalogue_)::NO_VALUE;
vregs_map.push_back({index});
vregs_mask.SetBit(i);
last_vregs_[i] = vreg;
vregs_last_change_[i] = stack_maps_.GetRowsCount();
}
}
BitMemoryRegion rgn(vregs_mask.data(), vregs_mask.size());
ASSERT(vregs_mask.PopCount() == vregs_map.size());
if (vregs_mask.PopCount() != 0) {
current_stack_map_[StackMap::COLUMN_VREG_MASK_INDEX] = vreg_masks_.Add(vregs_mask.GetFixed());
}
if (!current_vregs_.empty()) {
current_stack_map_[StackMap::COLUMN_VREG_MAP_INDEX] = vregs_map_.AddArray(Span(vregs_map));
}
}
void CodeInfoBuilder::Encode(ArenaVector<uint8_t> *stream, size_t offset)
{
BitMemoryStreamOut out(stream, offset);
uint32_t tables_mask = 0;
EnumerateTables([&tables_mask](size_t index, const auto &table) {
if (table->GetRowsCount() != 0) {
tables_mask |= (1U << index);
}
});
header_.SetTableMask(tables_mask);
header_.SetVRegsCount(vregs_count_);
header_.Encode(out);
EnumerateTables([&out]([[maybe_unused]] size_t index, const auto &table) {
if (table->GetRowsCount() != 0) {
table->Encode(out);
}
});
stream->resize(RoundUp(stream->size(), CodeInfo::SIZE_ALIGNMENT));
}
} // namespace panda::compiler

158
compiler/code_info/code_info_builder.h
View File

@ -1,158 +0,0 @@
/**
* Copyright (c) 2021-2022 Huawei Device Co., Ltd.
* 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.
*/
#ifndef PANDA_CODE_INFO_BUILDER_H
#define PANDA_CODE_INFO_BUILDER_H
#include "code_info.h"
#include "utils/arena_containers.h"
#include "utils/bit_vector.h"
namespace panda::compiler {
class CodeInfoBuilder {
public:
CodeInfoBuilder(Arch arch, ArenaAllocator *allocator)
: arch_(arch),
stack_maps_(allocator),
inline_infos_(allocator),
roots_reg_masks_(allocator),
roots_stack_masks_(allocator),
method_ids_(allocator),
vregs_catalogue_(allocator),
vregs_map_(allocator),
vreg_masks_(allocator),
implicit_nullchecks_(allocator),
constant_table_(allocator),
current_vregs_(allocator->Adapter()),
last_vregs_(allocator->Adapter()),
vregs_last_change_(allocator->Adapter()),
inline_info_stack_(allocator->Adapter()),
vregs_map_storage_(allocator->Adapter()),
vregs_mask_storage_(allocator)
{
}
NO_COPY_SEMANTIC(CodeInfoBuilder);
NO_MOVE_SEMANTIC(CodeInfoBuilder);
~CodeInfoBuilder() = default;
void BeginMethod(uint32_t frame_size, uint32_t vregs_count);
void EndMethod();
void BeginStackMap(uint32_t bpc, uint32_t npc, ArenaBitVector *stack_roots, uint32_t regs_roots,
bool require_vreg_map, bool is_osr);
void EndStackMap();
void BeginInlineInfo(void *method, uint32_t method_id, uint32_t bpc, uint32_t vregs_count);
void EndInlineInfo();
void AddVReg(VRegInfo reg)
{
// Constant should be added via `AddConstant` method
ASSERT(reg.GetLocation() != VRegInfo::Location::CONSTANT);
current_vregs_.push_back(reg);
}
void AddConstant(uint64_t value, VRegInfo::Type type, bool is_acc);
void SetFrameSize(uint32_t size)
{
header_.SetFrameSize(size);
}
void Encode(ArenaVector<uint8_t> *stream, size_t offset = 0);
void SetSavedCalleeRegsMask(uint32_t mask, uint32_t vmask)
{
header_.SetCalleeRegMask(mask);
header_.SetCalleeFpRegMask(vmask);
}
void AddImplicitNullCheck(uint32_t instruction_native_pc, uint32_t offset)
{
implicit_nullchecks_.Add({instruction_native_pc, offset});
}
void SetHasFloatRegs(bool has)
{
header_.SetHasFloatRegs(has);
}
template <typename Func>
constexpr void EnumerateTables(Func func)
{
size_t index = 0;
func(index++, &stack_maps_);
func(index++, &inline_infos_);
func(index++, &roots_reg_masks_);
func(index++, &roots_stack_masks_);
func(index++, &method_ids_);
func(index++, &vreg_masks_);
func(index++, &vregs_map_);
func(index++, &vregs_catalogue_);
func(index++, &implicit_nullchecks_);
func(index++, &constant_table_);
ASSERT(index == CodeInfo::TABLES_COUNT);
}
void DumpCurrentStackMap(std::ostream &stream) const;
private:
void EmitVRegs();
private:
Arch arch_;
uint32_t vregs_count_ {0};
uint32_t current_vregs_count_ {0};
CodeInfoHeader header_ {};
// Tables
BitTableBuilder<StackMap> stack_maps_;
BitTableBuilder<InlineInfo> inline_infos_;
BitTableBuilder<RegisterMask> roots_reg_masks_;
BitmapTableBuilder roots_stack_masks_;
BitTableBuilder<MethodId> method_ids_;
BitTableBuilder<VRegisterInfo> vregs_catalogue_;
BitTableBuilder<VRegisterCatalogueIndex> vregs_map_;
BitmapTableBuilder vreg_masks_;
BitTableBuilder<ImplicitNullChecks> implicit_nullchecks_;
BitTableBuilder<ConstantTable> constant_table_;
// Auxiliary containers
BitTableBuilder<StackMap>::Entry current_stack_map_;
ArenaVector<VRegInfo> current_vregs_;
ArenaVector<VRegInfo> last_vregs_;
ArenaVector<uint32_t> vregs_last_change_;
ArenaVector<BitTableBuilder<InlineInfo>::Entry> inline_info_stack_;
ArenaVector<BitTableBuilder<VRegisterCatalogueIndex>::Entry> vregs_map_storage_;
ArenaBitVector vregs_mask_storage_;
#ifndef NDEBUG
bool was_method_begin_ {false};
bool was_stack_map_begin_ {false};
bool was_inline_info_begin_ {false};
#endif
static constexpr size_t MAX_VREG_LIVE_DISTANCE = 32;
};
} // namespace panda::compiler
#endif // PANDA_CODE_INFO_BUILDER_H

214
compiler/code_info/code_info_tables.h
View File

@ -1,214 +0,0 @@
/**
* Copyright (c) 2021-2022 Huawei Device Co., Ltd.
* 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.
*/
#ifndef PANDA_CODE_INFO_TABLES_H
#define PANDA_CODE_INFO_TABLES_H
#include "utils/bit_field.h"
#include "utils/bit_table.h"
#include "utils/arch.h"
#include "vreg_info.h"
namespace panda::compiler {
class StackMap : public BitTableRow<8U, StackMap> {
public:
BIT_TABLE_HEADER(8, StackMap)
BIT_TABLE_COLUMN(0, Properties, PROPERTIES)
BIT_TABLE_COLUMN(1, NativePc, NATIVE_PC)
BIT_TABLE_COLUMN(2, BytecodePc, BYTECODE_PC)
BIT_TABLE_COLUMN(3, RootsRegMaskIndex, ROOTS_REG_MASK_INDEX)
BIT_TABLE_COLUMN(4, RootsStackMaskIndex, ROOTS_STACK_MASK_INDEX)
BIT_TABLE_COLUMN(5, InlineInfoIndex, INLINE_INFO_INDEX)
BIT_TABLE_COLUMN(6, VRegMaskIndex, VREG_MASK_INDEX)
BIT_TABLE_COLUMN(7, VRegMapIndex, VREG_MAP_INDEX)
DEFAULT_MOVE_SEMANTIC(StackMap);
DEFAULT_COPY_SEMANTIC(StackMap);
~StackMap() = default;
std::string GetColumnStr(size_t column) const
{
if (column != COLUMN_NATIVE_PC) {
return Base::GetColumnStr(column);
}
if (Get(column) == NO_VALUE) {
return "-";
}
return std::to_string(GetNativePcUnpacked());
}
static constexpr uintptr_t PackAddress(uintptr_t address, Arch arch)
{
ASSERT(IsAligned(address, GetInstructionAlignment(arch)));
return address / GetInstructionAlignment(arch);
}
uint32_t GetNativePcUnpacked(Arch arch = RUNTIME_ARCH) const
{
return UnpackAddress(GetNativePc(), arch);
}
static constexpr uintptr_t UnpackAddress(uintptr_t address, Arch arch)
{
return address * GetInstructionAlignment(arch);
}
static constexpr uint32_t CreateProperties(bool is_osr, bool has_regmap)
{
return FieldIsOsr::Encode(is_osr) | FieldHasRegMap::Encode(has_regmap);
}
bool IsOsr() const
{
return FieldIsOsr::Get(GetProperties());
}
bool HasRegMap() const
{
return FieldHasRegMap::Get(GetProperties());
}
private:
using FieldIsOsr = BitField<bool, 0, 1>;
using FieldHasRegMap = FieldIsOsr::NextFlag;
};
class InlineInfo : public BitTableRow<6U, InlineInfo> {
public:
BIT_TABLE_HEADER(6, InlineInfo)
BIT_TABLE_COLUMN(0, IsLast, IS_LAST)
BIT_TABLE_COLUMN(1, BytecodePc, BYTECODE_PC)
BIT_TABLE_COLUMN(2, MethodIdIndex, METHOD_ID_INDEX)
BIT_TABLE_COLUMN(3, MethodHi, METHOD_HI)
BIT_TABLE_COLUMN(4, MethodLow, METHOD_LOW)
BIT_TABLE_COLUMN(5, VRegsCount, VREGS_COUNT)
DEFAULT_MOVE_SEMANTIC(InlineInfo);
DEFAULT_COPY_SEMANTIC(InlineInfo);
~InlineInfo() = default;
};
class RegisterMask : public BitTableRow<1, RegisterMask> {
public:
BIT_TABLE_HEADER(1, RegisterMask)
BIT_TABLE_COLUMN(0, Mask, MASK)
DEFAULT_MOVE_SEMANTIC(RegisterMask);
DEFAULT_COPY_SEMANTIC(RegisterMask);
~RegisterMask() = default;
};
class StackMask : public BitTableRow<1, StackMask> {
public:
BIT_TABLE_HEADER(1, StackMask)
BIT_TABLE_COLUMN(0, Mask, MASK)
DEFAULT_MOVE_SEMANTIC(StackMask);
DEFAULT_COPY_SEMANTIC(StackMask);
~StackMask() = default;
};
class VRegisterMask : public BitTableRow<1, VRegisterMask> {
public:
BIT_TABLE_HEADER(1, VRegisterMask)
BIT_TABLE_COLUMN(0, Mask, MASK)
DEFAULT_MOVE_SEMANTIC(VRegisterMask);
DEFAULT_COPY_SEMANTIC(VRegisterMask);
~VRegisterMask() = default;
};
class MethodId : public BitTableRow<1, MethodId> {
public:
BIT_TABLE_HEADER(1, MethodId)
BIT_TABLE_COLUMN(0, Id, ID)
DEFAULT_MOVE_SEMANTIC(MethodId);
DEFAULT_COPY_SEMANTIC(MethodId);
~MethodId() = default;
};
class VRegisterCatalogueIndex : public BitTableRow<1, VRegisterCatalogueIndex> {
public:
BIT_TABLE_HEADER(1, VRegisterCatalogueIndex)
BIT_TABLE_COLUMN(0, Index, INDEX)
DEFAULT_MOVE_SEMANTIC(VRegisterCatalogueIndex);
DEFAULT_COPY_SEMANTIC(VRegisterCatalogueIndex);
~VRegisterCatalogueIndex() = default;
};
class VRegisterInfo : public BitTableRow<2U, VRegisterInfo> {
public:
BIT_TABLE_HEADER(2, VRegisterInfo)
BIT_TABLE_COLUMN(0, Info, INFO)
BIT_TABLE_COLUMN(1, Value, VALUE)
DEFAULT_MOVE_SEMANTIC(VRegisterInfo);
DEFAULT_COPY_SEMANTIC(VRegisterInfo);
~VRegisterInfo() = default;
std::string GetColumnStr(size_t column) const
{
if (column != COLUMN_INFO || Get(column) == NO_VALUE) {
return Base::GetColumnStr(column);
}
auto vreg = GetVRegInfo();
return std::string(vreg.GetLocationString()) + ":" + vreg.GetTypeString();
}
VRegInfo GetVRegInfo() const
{
return VRegInfo(GetValue(), GetInfo());
}
uint32_t GetConstantLowIndex() const
{
ASSERT(GetVRegInfo().GetLocation() == VRegInfo::Location::CONSTANT);
return GetValue() & ((1U << BITS_PER_UINT16) - 1);
}
uint32_t GetConstantHiIndex() const
{
ASSERT(GetVRegInfo().GetLocation() == VRegInfo::Location::CONSTANT);
return (GetValue() >> BITS_PER_UINT16) & ((1U << BITS_PER_UINT16) - 1);
}
};
class ImplicitNullChecks : public BitTableRow<2U, ImplicitNullChecks> {
public:
BIT_TABLE_HEADER(2, ImplicitNullChecks)
BIT_TABLE_COLUMN(0, InstNativePc, INST_NATIVE_PC)
BIT_TABLE_COLUMN(1, Offset, OFFSET)
DEFAULT_MOVE_SEMANTIC(ImplicitNullChecks);
DEFAULT_COPY_SEMANTIC(ImplicitNullChecks);
~ImplicitNullChecks() = default;
};
class ConstantTable : public BitTableRow<1, ConstantTable> {
public:
BIT_TABLE_HEADER(1, ConstantTable)
BIT_TABLE_COLUMN(0, Value, VALUE)
DEFAULT_MOVE_SEMANTIC(ConstantTable);
DEFAULT_COPY_SEMANTIC(ConstantTable);
~ConstantTable() = default;
};
} // namespace panda::compiler
#endif // PANDA_CODE_INFO_TABLES_H

220
compiler/code_info/vreg_info.h
View File

@ -1,220 +0,0 @@
/**
* Copyright (c) 2021-2022 Huawei Device Co., Ltd.
* 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.
*/
#ifndef COMPILER_CODE_INFO_VREG_INFO_H
#define COMPILER_CODE_INFO_VREG_INFO_H
#include "utils/bit_field.h"
namespace panda::compiler {
class VRegInfo final {
public:
enum class Location : int8_t { NONE, SLOT, REGISTER, FP_REGISTER, CONSTANT, COUNT = CONSTANT };
static constexpr size_t INVALID_LOCATION = (1U << MinimumBitsToStore(static_cast<size_t>(Location::COUNT))) - 1;
enum class Type : uint8_t { UNDEFINED, OBJECT, INT32, INT64, FLOAT32, FLOAT64, BOOL, ANY, COUNT = ANY };
VRegInfo()
{
FieldLocation::Set(Location::NONE, &info_);
ASSERT(!IsLive());
}
VRegInfo(uint32_t value, VRegInfo::Location location, Type type, bool is_acc)
: value_(value),
info_(FieldLocation::Encode(location) | FieldType::Encode(type) | FieldIsAccumulator::Encode(is_acc))
{
}
VRegInfo(uint32_t value, VRegInfo::Location location, Type type, bool is_acc, uint32_t index)
: VRegInfo(value, location, type, is_acc)
{
FieldVRegIndex::Set(index, &info_);
}
VRegInfo(uint32_t value, uint32_t packed_info) : value_(value), info_(packed_info) {}
static VRegInfo Invalid()
{
return VRegInfo(0, static_cast<Location>(INVALID_LOCATION), Type::UNDEFINED, false);
}
~VRegInfo() = default;
DEFAULT_COPY_SEMANTIC(VRegInfo);
DEFAULT_MOVE_SEMANTIC(VRegInfo);
uint32_t GetValue() const
{
return value_;
}
void SetValue(uint32_t value)
{
value_ = value;
}
Location GetLocation() const
{
return FieldLocation::Get(info_);
}
Type GetType() const
{
return FieldType::Get(info_);
}
uint16_t GetIndex() const
{
return FieldVRegIndex::Get(info_);
}
void SetIndex(uint16_t value)
{
FieldVRegIndex::Set(value, &info_);
}
bool IsAccumulator() const
{
return FieldIsAccumulator::Get(info_);
}
bool IsLive() const
{
return GetLocation() != Location::NONE;
}
bool IsObject() const
{
return GetType() == Type::OBJECT;
}
bool IsFloat() const
{
return GetType() == Type::FLOAT32 || GetType() == Type::FLOAT64;
}
bool Has64BitValue() const
{
return GetType() == VRegInfo::Type::FLOAT64 || GetType() == VRegInfo::Type::INT64;
}
bool IsLocationRegister() const
{
auto location = GetLocation();
return location == Location::REGISTER || location == Location::FP_REGISTER;
}
uint32_t GetConstantLowIndex() const
{
ASSERT(GetLocation() == Location::CONSTANT);
return GetValue() & ((1U << BITS_PER_UINT16) - 1);
}
uint32_t GetConstantHiIndex() const
{
ASSERT(GetLocation() == Location::CONSTANT);
return (GetValue() >> BITS_PER_UINT16) & ((1U << BITS_PER_UINT16) - 1);
}
void SetConstantIndices(uint16_t low, uint16_t hi)
{
value_ = low | (static_cast<uint32_t>(hi) << BITS_PER_UINT16);
}
bool operator==(const VRegInfo &rhs) const
{
return value_ == rhs.value_ && info_ == rhs.info_;
}
bool operator!=(const VRegInfo &rhs) const
{
return !(*this == rhs);
}
uint32_t GetInfo()
{
return info_;
}
const char *GetTypeString() const
{
switch (GetType()) {
case Type::OBJECT:
return "OBJECT";
case Type::INT64:
return "INT64";
case Type::INT32:
return "INT32";
case Type::FLOAT32:
return "FLOAT32";
case Type::FLOAT64:
return "FLOAT64";
case Type::BOOL:
return "BOOL";
case Type::ANY:
return "ANY";
default:
break;
}
UNREACHABLE();
}
const char *GetLocationString() const
{
switch (GetLocation()) {
case Location::NONE:
return "NONE";
case Location::SLOT:
return "SLOT";
case Location::REGISTER:
return "REGISTER";
case Location::FP_REGISTER:
return "FP_REGISTER";
case Location::CONSTANT:
return "CONSTANT";
default:
break;
}
UNREACHABLE();
}
void Dump(std::ostream &os) const
{
os << "VReg #" << GetIndex() << ":" << GetTypeString() << ", " << GetLocationString() << "="
<< helpers::ToSigned(GetValue());
if (IsAccumulator()) {
os << ", ACC";
}
}
private:
uint32_t value_ {0};
uint32_t info_ {0};
using FieldLocation = BitField<Location, 0, MinimumBitsToStore(static_cast<uint32_t>(Location::COUNT))>;
using FieldType = FieldLocation::NextField<Type, MinimumBitsToStore(static_cast<uint32_t>(Type::COUNT))>;
using FieldIsAccumulator = FieldType::NextFlag;
using FieldVRegIndex = FieldIsAccumulator::NextField<uint16_t, BITS_PER_UINT16>;
};
static_assert(sizeof(VRegInfo) <= sizeof(uint64_t));
inline std::ostream &operator<<(std::ostream &os, const VRegInfo &vreg)
{
vreg.Dump(os);
return os;
}
} // namespace panda::compiler
#endif // COMPILER_CODE_INFO_VREG_INFO_H

345
compiler/compile_method.cpp
View File

@ -1,345 +0,0 @@
/**
* Copyright (c) 2021-2022 Huawei Device Co., Ltd.
* 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 "optimizer_run.h"
#include "mem/pool_manager.h"
#include "mem/code_allocator.h"
#include "include/class.h"
#include "include/method.h"
#include "optimizer/ir/ir_constructor.h"
#include "optimizer/ir/runtime_interface.h"
#include "optimizer/analysis/loop_analyzer.h"
#include "optimizer/pass.h"
#include "optimizer/ir_builder/ir_builder.h"
#include "utils/logger.h"
#include "code_info/code_info.h"
#include "events/events.h"
#include "trace/trace.h"
#include "optimizer/code_generator/codegen.h"
#include "compile_method.h"
namespace panda::compiler {
#ifdef PANDA_COMPILER_CFI
static Span<uint8_t> EmitElf(Graph *graph, CodeAllocator *code_allocator, ArenaAllocator *gdb_debug_info_allocator,
const std::string &method_name);
#endif
static Arch ChooseArch(Arch arch)
{
if (arch != Arch::NONE) {
return arch;
}
arch = RUNTIME_ARCH;
if (RUNTIME_ARCH == Arch::X86_64 && options.WasSetCompilerCrossArch()) {
arch = GetArchFromString(options.GetCompilerCrossArch());
}
return arch;
}
static bool CheckSingleImplementation(Graph *graph)
{
// Check that all methods that were inlined due to its single implementation property, still have this property,
// otherwise we must drop compiled code.
// TODO(compiler): we need to reset hotness counter hereby avoid yet another warmup phase.
auto cha = graph->GetRuntime()->GetCha();
for (auto si_method : graph->GetSingleImplementationList()) {
if (!cha->IsSingleImplementation(si_method)) {
LOG(WARNING, COMPILER)
<< "Method lost single-implementation property after compilation, so we need to drop "
"whole compiled code: "
<< graph->GetRuntime()->GetMethodFullName(si_method);
return false;
}
}
return true;
}
static Span<uint8_t> EmitCode(const Graph *graph, CodeAllocator *allocator)
{
size_t code_offset = RoundUp(CodePrefix::STRUCT_SIZE, GetCodeAlignment(graph->GetArch()));
CodePrefix prefix;
prefix.code_size = graph->GetData().size();
prefix.code_info_offset = code_offset + RoundUp(graph->GetData().size(), sizeof(uint32_t));
prefix.code_info_size = graph->GetCodeInfoData().size();
size_t code_size = prefix.code_info_offset + prefix.code_info_size;
auto mem_range = allocator->AllocateCodeUnprotected(code_size);
if (mem_range.GetSize() == 0) {
return Span<uint8_t> {};
}
auto data = reinterpret_cast<uint8_t *>(mem_range.GetData());
memcpy_s(data, sizeof(CodePrefix), &prefix, sizeof(CodePrefix));
// NOLINTNEXTLINE(cppcoreguidelines-pro-bounds-pointer-arithmetic)
memcpy_s(&data[code_offset], graph->GetData().size(), graph->GetData().data(), graph->GetData().size());
// NOLINTNEXTLINE(cppcoreguidelines-pro-bounds-pointer-arithmetic)
memcpy_s(&data[prefix.code_info_offset], graph->GetCodeInfoData().size(), graph->GetCodeInfoData().data(),
graph->GetCodeInfoData().size());
allocator->ProtectCode(mem_range);
return Span<uint8_t>(reinterpret_cast<uint8_t *>(mem_range.GetData()), code_size);
}
uint8_t *GetEntryPoint(Graph *graph, [[maybe_unused]] Method *method, const std::string &method_name,
[[maybe_unused]] bool is_osr, CodeAllocator *code_allocator,
[[maybe_unused]] ArenaAllocator *gdb_debug_info_allocator)
{
#ifdef PANDA_COMPILER_CFI
auto generated_data = options.IsCompilerEmitDebugInfo()
? EmitElf(graph, code_allocator, gdb_debug_info_allocator, method_name)
: EmitCode(graph, code_allocator);
#else
auto generated_data = EmitCode(graph, code_allocator);
#endif
if (generated_data.Empty()) {
LOG(INFO, COMPILER) << "Compilation failed due to memory allocation fail: " << method_name;
return nullptr;
}
CodeInfo code_info(generated_data);
LOG(INFO, COMPILER) << "Compiled code for '" << method_name << "' has been installed to "
<< bit_cast<void *>(code_info.GetCode()) << ", code size " << code_info.GetCodeSize();
auto entry_point = const_cast<uint8_t *>(code_info.GetCode());
EVENT_COMPILATION(method_name, is_osr, method->GetCodeSize(), reinterpret_cast<uintptr_t>(entry_point),
code_info.GetCodeSize(), code_info.GetInfoSize(), events::CompilationStatus::COMPILED);
return entry_point;
}
bool JITCompileMethod(RuntimeInterface *runtime, Method *method, bool is_osr, CodeAllocator *code_allocator,
ArenaAllocator *allocator, ArenaAllocator *local_allocator,
ArenaAllocator *gdb_debug_info_allocator)
{
std::string method_name = runtime->GetMethodFullName(method, false);
SCOPED_TRACE_STREAM << "JIT compiling " << method_name;
if (!options.MatchesRegex(method_name)) {
LOG(DEBUG, COMPILER) << "Skip the method due to regexp mismatch: " << method_name;
return false;
}
Graph *graph {nullptr};
auto finalizer = [&graph]([[maybe_unused]] void *ptr) {
if (graph != nullptr) {
graph->~Graph();
}
};
std::unique_ptr<void, decltype(finalizer)> fin(&finalizer, finalizer);
auto arch {Arch::NONE};
bool is_dynamic = panda::panda_file::IsDynamicLanguage(method->GetClass()->GetSourceLang());
if (!CompileInGraph(runtime, method, is_osr, allocator, local_allocator, is_dynamic, &arch, method_name, &graph)) {
return false;
}
ASSERT(graph != nullptr && graph->GetData().data() != nullptr);
if (!is_dynamic && !CheckSingleImplementation(graph)) {
EVENT_COMPILATION(method_name, is_osr, method->GetCodeSize(), 0, 0, 0,
events::CompilationStatus::FAILED_SINGLE_IMPL);
return false;
}
// Drop non-native code in any case
if (arch != RUNTIME_ARCH) {
EVENT_COMPILATION(method_name, is_osr, method->GetCodeSize(), 0, 0, 0, events::CompilationStatus::DROPPED);
return false;
}
auto entry_point = GetEntryPoint(graph, method, method_name, is_osr, code_allocator, gdb_debug_info_allocator);
if (entry_point == nullptr) {
return false;
}
if (is_osr) {
if (runtime->HasCompiledCode(method)) {
runtime->SetOsrCode(method, entry_point);
ASSERT(runtime->GetOsrCode(method) != nullptr);
} else {
// Compiled code has been deoptimized, so we shouldn't install osr code.
// TODO(compiler): release compiled code memory, when CodeAllocator supports freeing the memory.
return false;
}
} else {
runtime->SetCompiledEntryPoint(method, entry_point);
}
ASSERT(graph != nullptr);
return true;
}
bool CompileInGraph(RuntimeInterface *runtime, Method *method, bool is_osr, ArenaAllocator *allocator,
ArenaAllocator *local_allocator, bool is_dynamic, Arch *arch, const std::string &method_name,
Graph **graph)
{
LOG(INFO, COMPILER) << "Compile method" << (is_osr ? "(OSR)" : "") << ": " << method_name << " ("
<< runtime->GetFileName(method) << ')';
*arch = ChooseArch(*arch);
if (*arch == Arch::NONE || !BackendSupport(*arch)) {
LOG(DEBUG, COMPILER) << "Compilation unsupported for this platform!";
return false;
}
ASSERT(*graph == nullptr);
*graph = allocator->New<Graph>(allocator, local_allocator, *arch, method, runtime, is_osr, nullptr, is_dynamic);
if (*graph == nullptr) {
LOG(ERROR, COMPILER) << "Creating graph failed!";
EVENT_COMPILATION(method_name, is_osr, method->GetCodeSize(), 0, 0, 0, events::CompilationStatus::FAILED);
return false;
}
if (!(*graph)->RunPass<IrBuilder>()) {
if (!compiler::options.IsCompilerIgnoreFailures()) {
LOG(FATAL, COMPILER) << "IrBuilder failed!";
}
LOG(WARNING, COMPILER) << "IrBuilder failed!";
EVENT_COMPILATION(method_name, is_osr, method->GetCodeSize(), 0, 0, 0, events::CompilationStatus::FAILED);
return false;
}
// Run compiler optimizations over created graph
bool res = RunOptimizations(*graph);
if (!res) {
if (!compiler::options.IsCompilerIgnoreFailures()) {
LOG(FATAL, COMPILER) << "RunOptimizations failed!";
}
LOG(WARNING, COMPILER) << "RunOptimizations failed!";
EVENT_COMPILATION(method_name, is_osr, method->GetCodeSize(), 0, 0, 0, events::CompilationStatus::FAILED);
return false;
}
LOG(DEBUG, COMPILER) << "The method is compiled";
return true;
}
} // namespace panda::compiler
#ifdef PANDA_COMPILER_CFI
#include "optimizer/ir/aot_data.h"
#include "tools/debug/jit_writer.h"
// Next "C"-code need for enable interaction with gdb
// Please read "JIT Compilation Interface" from gdb-documentation for more information
extern "C" {
// Gdb will replace implementation of this function
void NO_INLINE __jit_debug_register_code(void)
{
// NOLINTNEXTLINE(hicpp-no-assembler)
asm("");
}
// Default version for descriptor (may be checked before register code)
// NOLINTNEXTLINE(modernize-use-nullptr)
jit_descriptor __jit_debug_descriptor = {1, JIT_NOACTION, NULL, NULL};
} // extern "C"
namespace panda::compiler {
// NOLINTNEXTLINE(fuchsia-statically-constructed-objects)
static os::memory::Mutex jit_debug_lock;
// Will register jit-elf description in linked list
static void RegisterJitCode(jit_code_entry *entry)
{
ASSERT(options.IsCompilerEmitDebugInfo());
os::memory::LockHolder lock(jit_debug_lock);
// Re-link list
entry->next_entry = __jit_debug_descriptor.first_entry;
if (__jit_debug_descriptor.first_entry != nullptr) {
__jit_debug_descriptor.first_entry->prev_entry = entry;
}
__jit_debug_descriptor.first_entry = entry;
// Fill last entry
__jit_debug_descriptor.relevant_entry = entry;
__jit_debug_descriptor.action_flag = JIT_REGISTER_FN;
// Call gdb-callback
__jit_debug_register_code();
__jit_debug_descriptor.action_flag = JIT_NOACTION;
__jit_debug_descriptor.relevant_entry = nullptr;
}
// When code allocator cleaned - also will clean entry
void CleanJitDebugCode()
{
ASSERT(options.IsCompilerEmitDebugInfo());
os::memory::LockHolder lock(jit_debug_lock);
__jit_debug_descriptor.action_flag = JIT_UNREGISTER_FN;
while (__jit_debug_descriptor.first_entry != nullptr) {
__jit_debug_descriptor.first_entry->prev_entry = nullptr;
__jit_debug_descriptor.relevant_entry = __jit_debug_descriptor.first_entry;
// Call gdb-callback
__jit_debug_register_code();
__jit_debug_descriptor.first_entry = __jit_debug_descriptor.first_entry->next_entry;
}
__jit_debug_descriptor.action_flag = JIT_NOACTION;
__jit_debug_descriptor.relevant_entry = nullptr;
}
// For each jit code - will generate small elf description and put them in gdb-special linked list.
static Span<uint8_t> EmitElf(Graph *graph, CodeAllocator *code_allocator, ArenaAllocator *gdb_debug_info_allocator,
const std::string &method_name)
{
ASSERT(options.IsCompilerEmitDebugInfo());
if (graph->GetData().Empty()) {
return {};
}
JitDebugWriter jit_writer(graph->GetArch(), graph->GetRuntime(), code_allocator, method_name);
jit_writer.Start();
auto method = reinterpret_cast<Method *>(graph->GetMethod());
auto klass = reinterpret_cast<Class *>(graph->GetRuntime()->GetClass(method));
jit_writer.StartClass(*klass);
CompiledMethod compiled_method(graph->GetArch(), method);
compiled_method.SetCode(graph->GetData().ToConst());
compiled_method.SetCodeInfo(graph->GetCodeInfoData());
compiled_method.SetCfiInfo(graph->GetCallingConvention()->GetCfiInfo());
jit_writer.AddMethod(compiled_method, 0);
jit_writer.EndClass();
jit_writer.End();
if (!jit_writer.Write()) {
return {};
}
auto gdb_entry {gdb_debug_info_allocator->New<jit_code_entry>()};
if (gdb_entry == nullptr) {
return {};
}
auto elf_file {jit_writer.GetElf()};
// Pointer to Elf-file entry
gdb_entry->symfile_addr = reinterpret_cast<const char *>(elf_file.Data());
// Elf-in-memory file size
gdb_entry->symfile_size = elf_file.Size();
gdb_entry->prev_entry = nullptr;
RegisterJitCode(gdb_entry);
return jit_writer.GetCode();
}
} // namespace panda::compiler
#endif

38
compiler/compile_method.h
View File

@ -1,38 +0,0 @@
/**
* Copyright (c) 2021-2022 Huawei Device Co., Ltd.
* 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.
*/
#ifndef COMPILER_COMPILE_METHOD_H_
#define COMPILER_COMPILE_METHOD_H_
#include "compiler_options.h"
#include "mem/arena_allocator.h"
#include "mem/code_allocator.h"
#include "include/method.h"
#include "utils/arch.h"
namespace panda::compiler {
class Graph;
class RuntimeInterface;
bool JITCompileMethod(RuntimeInterface *runtime, Method *method, bool is_osr, CodeAllocator *code_allocator,
ArenaAllocator *allocator, ArenaAllocator *local_allocator,
ArenaAllocator *gdb_debug_info_allocator);
bool CompileInGraph(RuntimeInterface *runtime, Method *method, bool is_osr, ArenaAllocator *allocator,
ArenaAllocator *local_allocator, bool is_dynamic, Arch *arch, const std::string &method_name,
Graph **graph);
bool CheckMethodInLists(const std::string &method_name);
} // namespace panda::compiler
#endif // COMPILER_COMPILE_METHOD_H_

145
compiler/docs/balance_expressions_doc.md
View File

@ -1,145 +0,0 @@
# Expression balancing
## Overview
**Expression balancing** - optimization that reorganises computation of algebraic expressions.
The optimization is applied to expressions of the form of a long chain of the same binary associative and commutative operator, such as `ADD`, `MUL`, `AND`, `OR`, and etc. It calculates expression's critical path and, if it can be decreased, reorganises expression so it would be optimal. For example: critical path of `(((a + b) + c) + d)` is 3, whereas critical path of `((a + b) + (c + d))` is 2.
## Rationality
Increases instruction-level parallelism.
## Dependence
* RPO analyze.
## Algorithm
Visit all basic blocks in PRO order.
For each block iterate over instructions in reverse order looking for suitable instruction (i.e. operator).
If such instruction is found, it is the last one in an expression and it is necessary to determine the whole chain and its critical path by recursive [analysis](#operator-analysis) of operator inputs (`lhs` and `rhs`).
If the critical path isn't [optimal](#optimal-critical-path), delete expression's operators from the basic block, allocate them to expression's terms in an [optimal way](#operators-allocation) and insert new operators in the basic block.
### Note
#### Operator analysis
`Analysis of operator inputs` is a check if `lhs`(`rhs`) has the same opcode and has the only user (the operator itself). If so, the input is an operator of the expression, the analysis is called for it too and it should be saved to array of operators (`operators_`); otherwise it is an expression's term and it should be saved to array of terms (`sources_`).
If inputs belong to different basic blocks but satisfy the conditions above (have single user and are arithmetic operations), they should be moved to a single (dominatee) basic block (similiar to CodeSink).
#### Optimal critical path
The `optimal critical path` of an expression is `ceil[log2(n_terms)]` (`ceil[x]` is rounding `x` up).
#### Operators allocation
`Allocation in an optimal way` is an algorithm that creates expression and guarantees that it would have an [optimal](#optimal-critical-path) critical path:
Assume `terms[]` is an array of expression terms, algorithm called for two arguments `first_idx` and `last_idx` creates expression with terms in range from `terms[first_idx]` to `terms[last_idx]`.
The algorithm is:
- If range `first_idx:last_idx` covers `1` element then return this element.
- If range `first_idx:last_idx` covers `2` elements then create and return operator with `lhs` and `rhs` equal to `terms[first_idx]` and `terms[last_idx]`.
- Else calculate `split_idx` so that `split_idx` is strictly less than `last_idx` and size of `first_idx:split_idx` is the greatest possible power of 2, create and return operator with `lhs` and `rhs` equal to results of recursive calls `allocate(first_idx, split_idx)` and `allocate(split_idx + 1, last_idx)`.
## Pseudocode
```
for (auto basic_block : GetGraph()->GetBlocksRPO()) {
for (auto instruction : basic_block->InstsReverse()) {
if (instruction is a suitable operator) {
// Recursively check and save inputs, determine critical_path_length;
...
// Calculate optimal_critical_path_length;
...
if (optimal < current) {
AllocateSources(0, size(terms[]) - 1);
insert new expression to the basic block;
}
}
}
}
auto AllocateSources(size_t first_idx, size_t last_idx) {
if (first_idx == last_idx) {
return terms[first_idx];
}
if (first_idx == last_idx + 1) {
auto operator = operators[free_op_idx++];
operator->GetBasicBlock()->RemoveInst(operator);
operator.lhs = terms[first_idx];
operator.rhs = terms[last_idx];
basic_block.Insert(operator);
return operator;
}
else {
size_t split_idx = calculate split_idx;
auto lhs = AllocateSources(first_idx, split_idx);
auto rhs = AllocateSources(split_idx + 1, last_idx);
auto operator = operators[free_op_idx++];
basic_block->RemoveInst(operator);
operator.lhs = lhs;
operator.rhs = rhs;
basic_block.Insert(operator);
return operator;
}
}
```
## Examples
Before expressions balancing:
```
0.i64 Parameter -> v8 // a
1.i64 Parameter -> v8 // b
2.i64 Parameter -> v9 // c
3.i64 Parameter -> v10 // d
4.i64 Parameter -> v11 // e
5.i64 Parameter -> v12 // f
6.i64 Parameter -> v12 // g
8. Add v0, v1 -> v9 // a + b
9. Add v2, v8 -> v10 // c + (a + b)
// As soon as v10 has more than one users it has side-effects, so the algorithm considers it as a term:
10. Add v3, v9 -> v11, v14 // s10 = d + (c + (a + b))
11. Add v4, v10 -> v13 // e + s10
12. Add v5, v6 -> v13 // f + g
13. Add v11, v12 -> v14 // (e + s10) + (f + g)
14. Add v10, v13 -> v15 // s10 + ((e + s10) + (f + g))
15. Return v14
```
The code above contains two expressions: `v8-v10` (critical path is 3) and `v11-v14` (critical path is 3). Moreover, `v11-v14` is [optimal](#optimal-critical-path).
After expressions balancing:
```
0.i64 Parameter -> v8 // a
1.i64 Parameter -> v8 // b
2.i64 Parameter -> v9 // c
3.i64 Parameter -> v10 // d
4.i64 Parameter -> v11 // e
5.i64 Parameter -> v12 // f
6.i64 Parameter -> v12 // g
8. Add v3, v2 -> v10 // d + c
9. Add v0, v1 -> v10 // a + b
// As soon as v10 has more than one users it has side-effects, so the algorithm considers it as a term:
10. Add v8, v9 -> v11, v14 // s10 = ((d + c) + (a + b))
11. Add v4, v10 -> v13 // e + s10
12. Add v5, v6 -> v13 // f + g
13. Add v11, v12 -> v14 // (e + s10) + (f + g)
14. Add v10, v13 -> v15 // s10 + ((e + s10) + (f + g))
15. Return v14
```
## Links
Source code:
[balance_expressions.cpp](../optimizer/optimizations/balance_expressions.cpp)
[balance_expressions.h](../optimizer/optimizations/balance_expressions.h)
Tests:
[balance_expressions_test.cpp](../tests/balance_expressions_test.cpp)

231
compiler/docs/code_sink_doc.md
View File

@ -1,231 +0,0 @@
# Code Sink
## Overview
The optimization moves instructions into successor blocks, when possible, so that they are not executed on paths where their results are not needed.
## Rationality
This optimization allows to avoid execution of statements that are not used on execution path. This should speedup the execution.
Motivational example:
```
BB 3
0.i32 Parameter arg 0 -> (v5, v8, v7)
succs: [bb 0]
BB 0 preds: [bb 3]
8.i32 AddI v0, 0x1 -> (v6)
5. IfImm GT i32 v0, 0x0
succs: [bb 1, bb 2]
BB 2 preds: [bb 0]
6.i32 Return v8
BB 1 preds: [bb 0]
7.i32 Return v0
```
In this example `v8` is used only in one branch however it is always executed. The code sinking optimization suggests to move `v8` into `BB 2`.
## Dependence
* AliasAnalysis
* DominatorsTree
* LoopAnalysis
* Reverse Post Order (RPO)
## Algorithm
The iterative approach is used. On each iteration the optimization tries to sink each instruction to one of its immediately dominated blocks. It is possible if all users of the instruction is dominated by a block that the instruction is sunk into. Instructions in a basic block are iterated in reverse order to decrease the number of iterations. Iterating finishes when no instruction was sunk.
Instructions that cannot sink:
* Instructions allocating memory
* Control flow instructions
* Instructions that can throw an exception
* Barrier instructions (calls, monitors, volatile, SafePoints, etc.)
* Store instructions
* Load instructions if they dominate in scope of current basic block:
* an aliased store instruction
* a Monitor instruction
* a volatile store instruction
To determine which load instruction can be sunk we keep a list of store instructions that have been met so far (we are iterating in reverse order; therefore, when we meet load instruction, we have already collected all stores after this load and can easily check on aliases).
Blocks that instruction cannot be sunk into:
* Blocks that do not dominate all users of the instruction
* Loads cannot be sunk into blocks with more than one predecessors (because other predecessors might have aliased stores)
* Do not sunk instructions into loops
## Pseudocode
```
void CodeSink::RunImpl() {
// Iteratively sink instructions. On each iteration an instruction can be
// sunk to its basic block dominatee. Iterate sinking until no changes
// happens.
bool changed = true;
while(changed) {
changed = false;
for (auto block : GetGraph()->GetBlocksRPO()) {
bool barriered = false;
ArenaVector<Inst *> stores;
for (auto inst : block->InstsSafeReverse()) {
barriered |= inst->IsMonitor() || (inst->IsStore && inst->IsVolatile());
candidate = SinkInstruction(inst, &stores, barriered);
if (candidate != nullptr) {
block->EraseInst(inst);
candidate->PrependInst(inst);
changed = true;
}
}
}
}
}
BasicBlock *CodeSink::SinkInstruction(Inst *inst, ArenaVector<Inst *> *stores, bool barriered) {
// Save stores to be sure we do not sink a load instruction that may be aliased
if (inst->IsStore()) {
stores->push_back(inst);
return nullptr;
}
// Check that instruction can be sunk
if (inst->IsAllocation() || inst->IsControlFlow() || inst->CanThrow() || inst->IsBarrier()) {
return nullptr;
}
if (inst->IsLoad()) {
// Do not sink over monitor
// Do not sink over volatile store
if (barriered) {
return nullptr;
}
for (auto store : *stores) {
if (GetGraph()->CheckInstAlias(inst, store) != AliasType::NO_ALIAS) {
return nullptr;
}
}
}
// Iterate over dominated blocks
for (auto cand : inst->GetBasicBlock()->GetDominatedBlocks()) {
if (IsAcceptableTarget(inst, cand)) {
return cand;
}
}
return nullptr;
}
bool CodeSink::IsAcceptableTarget(Inst *inst, BasicBlock *candidate) {
BasicBlock *block = inst->GetBasicBlock();
Loop *loop = block->GetLoop();
Loop *cand_loop = candidate->GetLoop();
if (candidate->GetPredsBlocks().size() > 1) {
// Do not sink loads across a critical edge there may be stores in other code paths.
if (inst->IsLoad()) {
return false;
}
// Do not sink into loops
if (loop != cand_loop) {
return false;
}
}
// Check that all uses are dominated by the candidate
for (auto &user : inst->GetUsers()) {
Inst *uinst = user.GetInst();
if (!candidate->IsDominate(uinst->GetBasicBlock())) {
return false;
}
}
return true;
}
```
## Examples
### Regular sinking
```
BB 2 preds: [bb 0]
5.i32 Add v3, v2 -> (v15)
8.i64 LoadObject 243 v0 -> (v9, v13)
9.b Compare NE i64 v8, v4 -> (v10)
10. IfImm NE b v9, 0x0
succs: [bb 3, bb 4]
BB 4 preds: [bb 2]
13.i64 StoreObject 243 v1, v8
succs: [bb 3]
BB 3 preds: [bb 2, bb 4]
15.i32 Return v5
succs: [bb 1]
```
Sink arithmetic operation `v5` but do not `v8` because `BB 3` has several predecessors.
```
BB 2 preds: [bb 0]
8.i64 LoadObject 243 v0 -> (v9, v13)
9.b Compare NE i64 v8, v4 -> (v10)
10. IfImm NE b v9, 0x0
succs: [bb 3, bb 4]
BB 4 preds: [bb 2]
13.i64 StoreObject 243 v1, v8
succs: [bb 3]
BB 3 preds: [bb 2, bb 4]
5.i32 Add v3, v2 -> (v15)
15.i32 Return v5
succs: [bb 1]
```
### Loop Sinking
```
BB 2 preds: [bb 0]
6.i64 Add v1, v5 -> (v21)
succs: [bb 3]
BB 3 preds: [bb 2, bb 3]
prop: head, loop 1
10p.i64 Phi v4(bb2), v22(bb3) -> (v23, v22, v20)
20.i64 LoadArray v2, v10p -> (v21)
21.i64 Add v20, v6 -> (v26, v22)
22.i64 Add v21, v10p -> (v10p, v26)
23.i32 Add v10p, v3 -> (v24)
26.i64 Add v21, v22 -> (v27)
24.b Compare LT i32 v23, v0 -> (v25)
25. IfImm NE b v24, 0x0
succs: [bb 3, bb 4]
BB 4 preds: [bb 3]
27.i64 Return v26
succs: [bb 1]
```
Sinking `v6` into loop and `v26` out of loop
```
BB 2 preds: [bb 0]
6.i64 Add v1, v5 -> (v21)
succs: [bb 3]
BB 3 preds: [bb 2, bb 3]
prop: head, loop 1
10p.i64 Phi v4(bb2), v22(bb3) -> (v23, v20, v22)
20.i64 LoadArray v2, v10p -> (v21)
21.i64 Add v20, v6 -> (v22, v26)
22.i64 Add v21, v10p -> (v10p, v26)
23.i32 Add v10p, v3 -> (v24)
24.b Compare LT i32 v23, v0 -> (v25)
25. IfImm NE b v24, 0x0
succs: [bb 3, bb 4]
BB 4 preds: [bb 3]
26.i64 Add v21, v22 -> (v27)
27.i64 Return v26
succs: [bb 1]
```
## Links
Source code:
[code_sink.cpp](../optimizer/optimizations/code_sink.cpp)
[code_sink.h](../optimizer/optimizations/code_sink.h)
Tests:
[code_sink_test.cpp](../tests/code_sink_test.cpp)

153
compiler/docs/codegen_doc.md
View File

@ -1,153 +0,0 @@
# Code generator
## Overview
Codegen is the backend for current compiler implementation. It was implemented with the main idea - to be independent of specific encoder-architecture and have the possibility to work with different calling-conventions. It was designed to be independence from target-architecture - also, for this target - was created special compiler-independent library - encoder ([encoder.md](../optimizer/code_generator/encoder.md)).
## Dependence
It is needed Regalloc-pass and LinearOrder-analysis for codegen-work. First - because it is needed to fill register for each operand, second - because of needed jump-instruction generation for major edges.
## Implementation details
List of major codegen-dependency:
1. Encoder library (calling convention, encoder, and register description).
2. CodeBuilder - contain binary data and hold header with information for stack-walker.
3. CFrameLayout - is responsible for stack-frame layout in runtime(class described in `libpandabase/utils`).
4. SlowPath - class, which is responsible for side exits.
Codegen internal-implementation is responsibility for:
1. Conversions from IR to encoder structures (like conditions, immediate, registers, etc).
2. Filling meta-info for recovery frame-information in exceptions (state stamps, which used in stack-walker).
3. Getting information from runtime about current methods and objects.
4. Supporting logic for build AOT-code (there are needed runtime calls during aot-code execution).
5. Correct parameter transferring.
## Logic of work
Constructor makes additional work - it walks through call-instructions and calculates parameters count (they need for reserve stack-slots).
The main logic was made in RunImpl-method:
1. Generates Calling-Convention prologue.
2. Creates encode-visitor and visits each instruction in linear order by blocks.
3. After that - side exits (SlowPath) are emitted (also for OSR).
4. Generate epilog for correctness exit from method
5. At the end - there is the finalization of the encoder(emit real offsets for each branch) and CodeBuilder-header filling.
## Additional features
Also, there is implemented ScopedDisasmPrinter(for print disassembly), EncodeVisitor (for generating special opcodes).
Example of disasm-dump:
```
- START_METHOD
0000: stp x0, x20, [sp, #-32]!
0004: stp x29, x30, [sp, #16]
0008: add x29, sp, #0x10 (16)
...
- CallRuntime
0014: mov x16, #0xcbd4
0018: movk x16, #0x52, lsl #16
001c: blr x16
0020: ldp x30, xzr, [sp], #16
...
```
## Pseudocode
codegen.cpp:
```
void Codegen::RunImpl() {
GetCallingConvention()->BeginMethod(); // emit moves from parameters to dst-regs etc.
for (auto bb : GetGraph()->GetBlocksLinearOrder()) {
GetEncoder()->BindLabel(bb->GetId());
for (auto inst : bb->AllInsts()) {
visitor.VisitInstruction(inst);
}
}
EmitSlowPaths(); // Emit code, which responsibility for side exits.
GetEncoder()->Finalize(); // After that - it is possible to use generated code.
GetGraph()->SetData(EncodeDataType(code_entry, code_size)); // This data(entry and code) was used in code-cache
}
void Codegen::CreateCall(inst) {
auto callconv = GetCallingConvention();
auto dst_reg = ConvertRegister(call_inst->GetDstReg(), call_inst->GetType());
// 1-st parameter register
Reg param = GetRegfile()->GetSpecialRegister(SpecialReg::DEFAULT_PARAMETER);
if (GetGraph()->IsAotMode()) {
... // AOT code
} else {
GetEncoder()->SaveCallerRegisters();
SetCallParameters(call_inst);
auto method = ...->GetMethodById(...->GetMethod(), call_inst->GetCallMethodId());
// Move immediate-value to parameter
GetEncoder()->EncodeMov(param_0, Imm(method));
}
// Get offset of
size_t entry_point_offset = ...->GetCompiledEntryPointOffset();
// Major call instruction
GetEncoder()->MakeCall(MemRef(param_0, entry_point_offset));
...
GetEncoder()->LoadCallerRegisters(dst_reg);
}
class EncodeVisitor : public GraphVisitor {
/* VisitAdd, VisitSub, VisitMul etc... */
VisitInstruction(inst) {
switch (inst->Opc()) {
case (Opcode::Add) VisitAdd(inst); break;
case (Opcode::Cmp) VisitCmp(inst); break;
case (Opcode::Call) VisitCallStatic(inst); break;
...
}
}
// Full logic for generate one instruction:
void VisitAdd(inst) {
auto dst = GetCodegen()->ConvertRegister(inst->GetDstReg(), type);
auto src0 = GetCodegen()->ConvertRegister(inst->GetSrcReg(0), type);
enc->GetEncoder()->EncodeAdd(dst, src0);
}
void EncodeVisitor::VisitCallStatic(inst) {
GetCodegen()->CreateCall(inst);
}
void EncodeVisitor::VisitCmp(GraphVisitor* visitor, Inst* inst) {
auto* enc = static_cast<EncodeVisitor*>(visitor);
auto cmp_inst = inst->CastToCmp();
auto dst = GetCodegen()->ConvertRegister(inst->GetDstReg());
auto src0 = GetCodegen()->ConvertRegister(inst->GetSrcReg(0));
auto src1 = GetCodegen()->ConvertRegister(inst->GetSrcReg(1));
Condition cc = ...
GetEncoder()->EncodeCmp(dst, src0, src1, cc);
}
}
```
And them will have different encoding for each architecture:
```
target/aarch32/encode.cpp:
void Aarch32Encoder::EncodeCompare(Reg dst, Reg src0, Reg src1, Condition cc) {
CompareHelper(src0, src1, &cc); // Method for calculate flags for each src1 and src2 types
__ Mov(Convert(cc), VixlReg(dst), 0x1);
__ Mov(Convert(cc).Negate(), VixlReg(dst), 0x0);
}
target/aarch64/encode.cpp:
void Aarch64Encoder::EncodeCompare(Reg dst, Reg src0, Reg src1, Condition cc) {
__ Cmp(VixlReg(src0), VixlReg(src1));
__ Cset(VixlReg(dst), Convert(cc));
}
```
## Links
[codegen.cpp](../optimizer/code_generator/codegen.cpp)
[codegen.h](../optimizer/code_generator/codegen.h)
[slow_path.cpp](../optimizer/code_generator/slow_path.cpp)
[slow_path.h](../optimizer/code_generator/slow_path.h)
Tests:
[codegen_test.cpp](../tests/codegen_test.cpp), [inst_generator_test.cpp](../tests/inst_generator_test.cpp) - test with generation each instruction.

155
compiler/docs/if_conversion_doc.md
View File

@ -1,155 +0,0 @@
# IfConversion
## Overview
`IfConversion` tries to remove branches in executable code by creating linear sections with predicate instructions.
## Rationality
Hardware executes the program speculatively. It tries to predict the branch conditional(true or false) and starts executing instructions after the branch speculatively before executing the branch itself. If the prediction is incorrect(named branch misprediction), the pipeline stops and the state is restored. As result, several dozen cycles are lost. `IfConversion` can add several instructions but remove branch misprediction.
## Dependence
* Dead Code Elimination(DCE)
* Remove Empty Blocks
* Remove Linear blocks
* Reverse Post Order(RPO)
## Algorithm
Optimization makes a pass through the blocks by post order traversal.
Two patterns are checked for each block: `Triangle` and `Diamond`.
### Triangle
The pattern:
```
[BB]
| \
| [JBB]
| /
[PBB]
```
`BB` -- basic block the recognition starts from
`JBB`(Join BB) -- true or false successor of `BB`, which will be joined to BB
`PBB`(Phi BB) -- false or true successor of `BB`, which contain PHI instruction for BB and JBB
### Diamond
```
[BB]
/ \
[JBB] [JBB 2]
\ /
[PBB]
```
`BB` -- basic block the recognition starts from
`JBB`(Join BB), `JBB 2` -- true and false successors of BB, which will be joined to `BB `
`PBB`(Phi BB) -- the successor of `JBB` and `JBB 2`, which contain PHI instruction for `JBB` and `JBB 2`
### Conditions to check
1. `JBB`(and `JBB 2` for Diamond) must have only one predecessor and one successor
2. `PBB` must have 2 or more predecessors
3. `JBB`(and `JBB 2` for Diamond) is the predecessor of the PBB
4. `JBB`(and `JBB 2` for Diamond) doesn't contain instruction with `no_ifcvt` property(for example memory instruction, call instruction, instruction with a call to runtime)
5. The number of instructions in `JBB`(and `JBB 2` for Diamond) less than the limit(set by the option `--compiler-if-conversion-limit=N` with the default value 2)
6. The number of Phi instruction in `PBB`, which have different inputs from corresponding predecessor blocks, should also be less than the limit(each of them would be converted into Select)
7. `PBB` doesn't contain float Phi with different inputs for `JBB` and `BB`(`JBB 2` for Diamond)
### Transformation
1. `If` instructions removed from `BB`(the necessary information, such as the CC, is saved)
2. Edges `BB` -> `JBB` and `JBB` -> `PBB` are removed
3. All instruction from `JBB` are copied to `BB`
4. Select instructions are constructed at the end of `BB`(`JBB 2` for Diamond)
5. All Phi instructions in `PBB` are edited:
a. If `PBB` has other predecessors, we check if inputs from `JBB` and `BB`(`JBB 2` for Diamond) are equal, then input from `JBB` is removed. Otherwise, it is also removed, but input from `BB`(`JBB 2` for Diamond) is changed to corresponding Select instruction.
b. If `PBB` doesn't have other predecessors, all Phi inputs are copied to Select instructions and Phi instruction is deleted.
6. For Diamond `BB` and `JBB 2` are merged
7. If `PBB` doesn't have other predecessors, `BB` and `PBB` are merged
8. Loop information is fixed
## Pseudocode
TODO
## Examples
**Triangle**:
Before:
```
BB 2 preds: [bb 0]
3.b Compare B u64 v0, v1 -> (v4)
4. IfImm NE b v3, 0x0
succs: [bb 3, bb 4]
BB 3 preds: [bb 2]
5.u64 Mul v0, v2 -> (v6p)
succs: [bb 4]
BB 4 preds: [bb 2, bb 3]
6p.u64 Phi v0(bb2), v5(bb3) -> (v7)
7.u64 Return v6p
succs: [bb 1]
```
After:
```
BB 2 preds: [bb 0]
3.b Compare B u64 v0, v1 -> (v8)
5.u64 Mul v0, v2 -> (v8)
8.u64 SelectImm NE b v5, v0, v3, 0x0 -> (v7)
7.u64 Return v8
succs: [bb 1]
```
**Diamond**:
Before:
```
BB 2 preds: [bb 0]
3.b Compare EQ u32 v1, v2 -> (v4)
4. IfImm NE b v3, 0x0
succs: [bb 3, bb 4]
BB 4 preds: [bb 2]
5.u32 Add v0, v1 -> (v8p)
succs: [bb 5]
BB 3 preds: [bb 2]
7.u32 Sub v0, v1 -> (v8p)
succs: [bb 5]
BB 5 preds: [bb 4, bb 3]
8p.u32 Phi v5(bb4), v7(bb3) -> (v9)
9.u32 Return v8p
succs: [bb 1]
```
After:
```
BB 2 preds: [bb 0]
3.b Compare EQ u32 v1, v2 -> (v10)
7.u32 Sub v0, v1 -> (v10)
5.u32 Add v0, v1 -> (v10)
10.u32 SelectImm NE b v7, v5, v3, 0x0 -> (v9)
9.u32 Return v10
succs: [bb 1]
```
## Links
Source code:
[if_conversion.cpp](../optimizer/optimizations/if_conversion.cpp)
[if_conversion.h](../optimizer/optimizations/if_conversion.h)
Tests:
[if_conversion_test.cpp](../tests/if_conversion_test.cpp)

179
compiler/docs/loop_peeling.md
View File

@ -1,179 +0,0 @@
# Loop Peeling
## Overview
`Loop peeling` optimization modifies the loops with exit-point at loop-header to the loops with exit-point at loop-backedge.
## Rationality
Simplify the loop and allow further loop optimizations.
## Dependence
* Loop Analysis
* Dominators Tree
* Reverse Post Order (RPO)
## Algorithm
`Loop peeling` modifies loops with the following requirements:
- loop is not irreducible;
- loop-header is not OSR-entry;
- there is only 1 back-edge;
- loop-header is a single loop-exit point;
- there are no inner loops;
```cpp
[pre-header]
|
v
/---->[header]--------\
| | |
| v v
\----[back-edge] [outer]
```
There are two stages of the algorithm:
### 1. Insert pre-loop
```cpp
[pre-header]
|
v
[pre-loop]--------\
| |
v v
/---->[header]-------->|
| | |
| v v
\----[back-edge] [resolver]
|
v
[outer]
```
Pre-loop basic block is a loop-header clone with all instructions, excluding `SafePoint`.
### 2. Move exit-point form the loop-header to the loop-backedge block
```cpp
[pre-header]
|
v
[pre-loop]--------\
| |
v v
/---->[header] |
| | |
| v v
\----[back-edge]-->[resolver]
|
v
[outer]
```
All instructions from loop-header are moving to the loop-backedge block. Also control-flow edge between loop-header and resolver-block is moving to the loop-backedge.
## Pseudocode
```cpp
auto header = loop->GetHeader();
auto pre-loop = LoopPeeling->CreatePreLoop();
for (auto inst : header->GetInstructions()) {
auto clone_inst = Clone(inst);
pre-loop.AppendInst(clone_inst);
}
auto exit_block = LoopPeeling->CreateExitBlock();
for (auto inst : header->GetInstructionsReverse()) {
header->EraseInst(inst);
exit_block->PrependInst(inst);
}
```
## Examples
[0]
|
v
/---->[2]-----\
| | |
| v v
\-----[3] [4]
|
[exit]
GRAPH(GetGraph()) {
PARAMETER(0, 0).u64();
PARAMETER(1, 1).u64();
PARAMETER(2, 2).u64();
BASIC_BLOCK(2, 3, 4) {
INST(3, Opcode::Phi).u64().Inputs(1, 5);
INST(4, Opcode::Phi).u64().Inputs(2, 10);
INST(5, Opcode::Sub).u64().Inputs(3, 2);
INST(6, Opcode::SafePoint).Inputs(0, 3, 4).SrcVregs({0, 1, 2});
INST(7, Opcode::Compare).CC(CC_EQ).b().Inputs(5, 0);
INST(8, Opcode::IfImm).SrcType(DataType::BOOL).CC(CC_NE).Imm(0).Inputs(7);
}
BASIC_BLOCK(3, 2) {
INST(9, Opcode::And).u64().Inputs(4, 5);
INST(10, Opcode::Add).u64().Inputs(9, 4);
}
BASIC_BLOCK(4, -1) {
INST(11, Opcode::Return).u64().Inputs(4);
}
}
`Loop peeling` transforms to:
[0]
|
v
[pre-loop]---------\
| |
/---->[2] |
| | |
| v |
| [3] |
| | |
| v v
\--[loop-exit]--->[loop-outer]
|
v
[4]
|
v
[exit]
GRAPH(expected_graph) {
PARAMETER(0, 0).u64();
PARAMETER(1, 1).u64();
PARAMETER(2, 2).u64();
BASIC_BLOCK(5, 2, 4) {
INST(12, Opcode::Sub).u64().Inputs(1, 2);
INST(13, Opcode::Compare).CC(CC_EQ).b().Inputs(12, 0);
INST(14, Opcode::IfImm).SrcType(DataType::BOOL).CC(CC_NE).Imm(0).Inputs(13);
}
BASIC_BLOCK(2, 2, 4) {
INST(3, Opcode::Phi).u64().Inputs({{5, 12}, {2, 5}});
INST(4, Opcode::Phi).u64().Inputs({{5, 2}, {2, 10}});
INST(15, Opcode::Phi).u64().Inputs({{5, 12}, {2, 5}});
INST(9, Opcode::And).u64().Inputs(4, 15);
INST(10, Opcode::Add).u64().Inputs(9, 4);
INST(5, Opcode::Sub).u64().Inputs(3, 2);
INST(6, Opcode::SafePoint).Inputs(0, 5, 10).SrcVregs({0, 1, 2});
INST(7, Opcode::Compare).CC(CC_EQ).b().Inputs(5, 0);
INST(8, Opcode::IfImm).SrcType(DataType::BOOL).CC(CC_NE).Imm(0).Inputs(7);
}
BASIC_BLOCK(4, -1) {
INST(16, Opcode::Phi).u64().Inputs({{5, 2}, {2, 10}});
INST(11, Opcode::Return).u64().Inputs(16);
}
}
## Links
Source code:
[loop_peeling.cpp](../optimizer/optimizations/loop_peeling.cpp)
[loop_peeling.h](../optimizer/optimizations/loop_peeling.h)
Tests:
[loop_peeling_test.cpp](../tests/loop_peeling_test.cpp)

212
compiler/docs/loop_unrolling.md
View File

@ -1,212 +0,0 @@
# Loop Unrolling
## Overview
`Loop unrolling` optimization increases loop body by copying instructions of the original loop body.
## Rationality
Increase number of instructions for each loop iteration, reduce branch penalty.
## Dependence
* Loop Analysis
* Dominators Tree
* Reverse Post Order (RPO)
* Loop Peeling (to make loop with exit-point from backedge)
## Algorithm
`Loop unrolling` modifies loops with the following requirements:
* loop is not irreducible;
* loop-header is not OSR-entry;
* there is only 1 back-edge;
* loop-backedge is a single loop-exit point;
* there are no inner loops;
Optimization settings:
**Instructions limit** - the maximum number of loop instructions after its unrolling;
**Unroll factor** - the number of loop body copies including the original one;
There two types of unrolling: with side-exits and without them. Unrolling without side-exits is applied for countable loops.
### Countable loops
Loop is countable if it contains compare between loop-index instruction and test-instruction defined outside loop.
Loop-index should be incremented or decremented by a constant. Currently signed integer indexes are supported.
```
[Loop-header]
Phi(init, update)
...
[Loop-backedge]
update(phi, constant)
Compare(update, test)
...
where `update` is Add or Sub instruction
```
### Unrolling without side-exits
There are 3 stages of unrolling loop without side-exits:
1. Clone loop-body without loop-backedge `factor` times;
2. Fix loop-backedge compare by incrementing/decrementing test input with constant, counted using formula: `factor * loop_step`. If compare has `not-equal` condition code, replace it by `less-than`/`greater-than`
3. Clone loop-body with the original loop-backedge `factor` times, but replace edge to the loop-header with edge the loop-outer block;
Here `factor` means number of cloned loop bodies.
```cpp
/---->[header]
| |
| v
| [loop-body]
| |
| v
\-----[backedge]----> ...
```
```
/---->[header]
| |
| v
| [loop-body]
| |
| v
| [loop-body']
| |
| v
| [loop-body'']
| |
| v
\-----[backedge]----> ...
|
v
[loop-body]
|
v
[backedge]------\
| |
v |
[loop-body] |
| |
v |
[outer-block]<---/
```
### Unrolling with side-exits
For this case both loop-body and loop-backedge are cloned:
```cpp
/---->[header]
| |
| v
| [loop-body]
| |
| v
| [backedge]------------\ << exit-block
| | |
| v |
| [loop-body-clone] |
| | |
| v |
\-----[backedge-clone]----->| << last-block
|
v
[outer]-----> ...
```
## Pseudocode
```cpp
if (IsLoopCountable(loop)) {
auto clone_loop = CloneLoop(loop);
UnrollLoopBodyWithoutSideExits(loop);
FixCompareInst(loop);
UnrollLoopBodyWithSideExits(clone_loop);
RemoveEdgeToLoopHeader(clone_loop);
} else {
UnrollLoopBodyWithSideExits(loop);
}
```
## Examples
Countable loop unrolling:
```cpp
auto graph = CreateEmptyGraph();
GRAPH(graph) {
CONSTANT(0, stop);
CONSTANT(1, 0); // a = 0, b = 0
CONSTANT(2, step);
BASIC_BLOCK(2, 3, 4) {
INST(3, Opcode::Compare).b().SrcType(DataType::INT32).CC(CC_LT).Inputs(1, 0);
INST(4, Opcode::IfImm).SrcType(DataType::BOOL).CC(CC_NE).Imm(0).Inputs(3); // if a < stop
}
BASIC_BLOCK(3, 3, 4) {
INST(5, Opcode::Phi).s32().Inputs(1, 7); // a
INST(6, Opcode::Phi).s32().Inputs(1, 8); // b
INST(7, Opcode::Add).s32().Inputs(5, 2); // a += step
INST(8, Opcode::Add).s32().Inputs(6, 7); // b += a
INST(9, Opcode::Compare).b().SrcType(DataType::INT32).CC(CC_LT).Inputs(7, 0);
INST(10, Opcode::IfImm).SrcType(DataType::BOOL).CC(CC_NE).Imm(0).Inputs(9); // if a < stop
}
BASIC_BLOCK(4, -1) {
INST(11, Opcode::Phi).s32().Inputs(1, 6);
INST(12, Opcode::Return).s32().Inputs(11); // return b;
}
}
return graph;
```
```cpp
uint32_t UNROLL_FACTOR = 2;
GRAPH(graph_unroll) {
CONSTANT(0, 10);
CONSTANT(1, 0); // a = 0, b = 0
CONSTANT(2, 1);
BASIC_BLOCK(2, 3, 5) {
INST(20, Opcode::SubI).s32().Inputs(0).Imm(UNROLL_FACTOR - 1);
INST(3, Opcode::Compare).b().SrcType(DataType::INT32).CC(CC_LT).Inputs(1, 20); // if (a < 10 - (UNROLL_FACTOR - 1))
INST(4, Opcode::IfImm).SrcType(DataType::BOOL).CC(CC_NE).Imm(0).Inputs(3);
}
BASIC_BLOCK(3, 3, 5) {
INST(5, Opcode::Phi).s32().Inputs(1, 21); // a
INST(6, Opcode::Phi).s32().Inputs(1, 22); // b
INST(7, Opcode::Add).s32().Inputs(5, 2); // a + 1
INST(8, Opcode::Add).s32().Inputs(6, 7); // b + 1
INST(21, Opcode::Add).s32().Inputs(7, 2); // a + 1
INST(22, Opcode::Add).s32().Inputs(8, 21); // b + 1
INST(9, Opcode::Compare).b().SrcType(DataType::INT32).CC(CC_LT).Inputs(21, 20);
INST(10, Opcode::IfImm).SrcType(DataType::BOOL).CC(CC_NE).Imm(0).Inputs(9); // if (a < 10 - (UNROLL_FACTOR - 1))
}
BASIC_BLOCK(5, 6, 4) {
INST(11, Opcode::Phi).s32().Inputs(1, 8);
INST(25, Opcode::Phi).s32().Inputs(1, 21); // a
INST(26, Opcode::Phi).s32().Inputs(1, 22); // b
INST(27, Opcode::Compare).b().SrcType(DataType::INT32).CC(CC_LT).Inputs(25, 0); // if (a < 10)
INST(28, Opcode::IfImm).SrcType(DataType::BOOL).CC(CC_NE).Imm(0).Inputs(27);
}
BASIC_BLOCK(6, 4) {
INST(29, Opcode::Add).s32().Inputs(25, 2); // a + 1
INST(30, Opcode::Add).s32().Inputs(26, 29); // b + 1
}
BASIC_BLOCK(4, -1) {
INST(31, Opcode::Phi).s32().Inputs(11, 26);
INST(12, Opcode::Return).s32().Inputs(31); // return b
}
}
```
## Links
Source code:
[loop_unroll.cpp](../optimizer/optimizations/loop_unroll.cpp)
[loop_unroll.h](../optimizer/optimizations/loop_unroll.h)
Tests:
[loop_unroll_test.cpp](../tests/loop_unroll_test.cpp)

104
compiler/docs/memory_barriers_doc.md
View File

@ -1,104 +0,0 @@
# Optimize memory barriers
## Overview
We need to encode barriers after the instructions NewArray, NewObject, NewMultiArray so that if the created objects are used in another thread, the initialization is fully completed.
We can remove the barrier if we prove that the created object cannot be passed to another thread before the next barrier.
This can happen if we save the object to memory or pass it to another method
## Rationality
Reducing the number of instructions and speed up execution.
## Dependence
RPO analysis
## Algorithm
There is instruction flag `MEM_BARRIER`. The flag is set to `true` for the instructions NewObject, NewArray and NewMultiArray.
The pass `OptimizeMemoryBarriers` try remove the flag(set false) from the instruction.
We pass through all instructions in PRO order. If the instruction has flag `MEM_BARRIER` we add the instruction in special vector `barriers_insts_`.
If we visit an instruction that can pass an object to another thread(Store instruction, Call instruction e.t.c) we check the instruction inputs.
If the instruction has input from the `barriers_insts_`, we call function `MergeBarriers`.
The function set `false` for the flag `MEM_BARRIER`, exclude last instruction from the vector.
So we will only set the barrier in the last instruction before potentially passing the created objects to another thread
The function `MergeBarriers` also is called at end of the basic block.
Codegen checks the flag `MEM_BARRIER` for the instructions NewObject, NewArray and NewMultiArray and encode memory barrier if the flag `true`
## Pseudocode
```
bool OptimizeMemoryBarriers::RunImpl()
{
barriers_insts.clear();
for (auto bb : GetGraph()->GetBlocksRPO()) {
for (auto inst : bb->Insts()) {
if (inst->GetFlag(inst_flags::MEM_BARRIER)) {
barriers_insts.push_back(inst);
}
if (InstCanMoveObjectInAnotherthread(inst) && InstHasBarrierInput(inst, barriers_insts)) {
MergeBarriers(barriers_insts);
}
}
MergeBarriers(barriers_insts);
}
return true;
}
void MemoryBarriersVisitor::MergeBarriers(InstVector& barriers_insts)
{
if (barriers_insts.empty()) {
return;
}
auto last_barrier_inst = barriers_insts.back();
for (auto inst : barriers_insts) {
inst->ClearFlag(inst_flags::MEM_BARRIER);
}
last_barrier_inst->SetFlag(inst_flags::MEM_BARRIER);
barriers_insts.clear();
}
```
## Examples
```
BB 0
prop: start
0.i64 Constant 0x2a -> (v6, v3, v1, v2, v8, v11)
succs: [bb 2]
BB 2 preds: [bb 0]
1. SaveState v0(vr0) -> (v2)
2.ref NewArray 1 v0, v1 -> (v6, v3, v8, v11, v12)
3. SaveState v0(vr0), v2(vr1) -> (v5, v4)
4.ref LoadAndInitClass 'A' v3 -> (v5)
5.ref NewObject 2 v4, v3 -> (v6, v8, v11, v12)
6. SaveState v0(vr0), v2(vr1), v5(vr2) -> (v7, v12)
7.void CallStatic 3 v6
8. SaveState v0(vr0), v2(vr1), v5(vr2) -> (v9, v10)
9.ref LoadAndInitClass 'B' v8 -> (v10)
10.ref NewObject 4 v9, v8 -> (v11, v13)
11. SaveState v0(vr0), v2(vr1), v5(vr2), v10(vr3)
12.i64 CallVirtual 5 v2, v5, v6
13.ref Return v10
succs: [bb 1]
BB 1 preds: [bb 2]
prop: end
```
Instructions `2.ref NewArray`, `5.ref NewObject` and `10.ref NewObject` have flag `MEM_BARRIER` by default.
`7.void CallStatic` don't have the instructions `2.ref NewArray`, `5.ref NewObject` as inputs.
So the pass `OptimizeMemoryBarriers` will remove the flag from these instructions and skip in `10.ref NewObject`.
## Links
Source code:
[memory_barriers.cpp](../optimizer/optimizations/memory_barriers.cpp)
[memory_barriers.h](../optimizer/optimizations/memory_barriers.h)
Tests:
[memory_barriers_test.cpp](../tests/memory_barriers_test.cpp)

250
compiler/docs/memory_coalescing_doc.md
View File

@ -1,250 +0,0 @@
# Memory Coalescing
## Overview
The optimization is based on the fact that some architectures (`AArch64` particularly) support simultaneous load and store operations for consecutive addresses instead of several separate operations.
## Rationality
Replacing two memory operations with one generally reduces the number of long latency memory instructions.
| Code | Regular | Optimized |
| ------ | ------ | ------|
| `num[0] * num[1];` | `ldr x1, [x0]` <br> `ldr x0, [x0, 8]` <br> `mul x0, x1, x0` | `ldp x1, x0, [x0]` <br> `mul x0, x1, x0` |
## Dependence
* DominatorsTree
* LoopAnalyzer
* AliasAnalysis
* Reverse Post Order (RPO)
## Assumptions
The optimization was implemented for arrays' accesses for `AArch64` architecture.
Array accesses cannot be volatile.
`AArch64` has `32`-bit and `64`-bit versions of coalescing operations – `ldp` and `stp`. As a result it is possible only for following Panda's types: `INT32`, `UINT32`, `INT64`, `UINT64`, `REFERENCE`.
## Algorithm
To implement such kind of optimization the extra support from IR is required
### IR Support
The following actions are required
* Separate instructions that will represent coalesced memory accesses from regular accesses.
* Handle multiple output from one instruction in terms of SSA
The case with a coalesced store is quite straightforward: having two consecutive stores we replace them by one instruction that accepts index and two values to store.
| Consecutive Stores | Coalesced Store |
| --- | --- |
| `248.i64 StoreArrayI v2, 0x0, v53` <br> `250.i64 StoreArrayI v2, 0x1, v53` | `251.i64 StoreArrayPairI v2, 0x0, v53, v53` |
The problem occurs with a coalesced load because a load instruction of multiple values produces multiple assignment that is not a part of SSA form. By this reason, we need additional pseudo instructions as `LoadPairPart` to divide multiple values into single values. The type of `LoadArrayPair` instruction corresponds to the type of a single element, so there is an assumption that `LoadArrayPair` loads only multiple values of the same type.
| Consecutive Loads | Coalesced Load |
| --- | --- |
| `58.i64 LoadArrayI v2, 0x0 -> (v37)` <br> `61.i64 LoadArrayI v2, 0x1 -> (v43)` | `62.i64 LoadArrayPairI v2, 0x0 -> (v63, v64)` <br> `63.i64 LoadPairPart v62, 0x0 -> (v37)` <br> `64.i64 LoadPairPart v62, 0x1 -> (v43)` |
### Transformation
The optimization tries to coalesce memory operations in a scope of a basic block. It needs that two consecutive memory operations are placed near each other without intermediate instructions. By this reason we need to find a place to sunk an upper memory operation and to hoist the lower according to reordering rules.
During hoisting and sinking of memory operations we use rules for memory instruction scheduling: do not move over monitors, calls, save states, save points and etc.
Memory coalescing was implemented for array accesses. We process instructions of basic block in order. To find accesses of consecutive memory addresses we keep a queue of candidates. Each instruction that may be coalesced is inserted into this queue. A candidate is marked as invalid in the following conditions:
* it has been paired already
* store candidates are invalid if SaveState instruction has been met
* all candidates are invalid if a barrier is met: calls, control flow instructions, monitors, exceptions, intrinsic and etc.
To track indices we use basic implementation of scalar evolution that allows to track how variables evolves: basic value (variable or constant), difference from the basic value (if basic value is a variable) and evolution (if basic value is a variable incremented on each iteration of a loop). It is a simple graph bypass by collecting assignments including Phi evolutions (supported only addition and subtraction).
Processing each instruction in basic block we do the following:
1) If the instruction cannot be coalesced.
1) If the instruction is a barrier – invalidate all candidates.
2) If the instruction is a SaveState – invalidate all store candidates.
3) If the instruction is a memory operation – add it as invalid candidate.
2) If we can't determine anything about index variable, we add this instruction as a candidate and move on next instruction
3) Iterate candidates in backward order
2) If a candidate is invalid **or** candidate cannot be coalesced with the instruction **or** both refer to different objects **or** we have no information about candidate index, move on next candidate
3) If indices differs by one and there is a place to sunk the candidate instruction and hoist the currently processing instruction – add this candidate and the instruction as a pair for coalescing and invalidate both.
4) Add the instruction as a candidate.
Finally, we replace collected pairs by coalesced instructions.
To find a place for candidate and current instruction:
1) find the lowest position the candidate can be sunk
2) find the highest position the instruction can be hoisted
3) The place can be any between highest and lowest position. If the intersection is empty, coalescing is not possible.
## Pseudocode
```
void MemoryCoalescing::RunImpl() {
VariableAnalysis variables(GetGraph());
for (auto block : GetGraph()->GetBlocksRPO()) {
for (auto inst : block.Insts()) {
if (IsArrayAccess(inst)) {
HandleArrayAccess(inst, variables);
} else if (inst->IsMemory()) {
inst->SetMarker(mrk_invalid_);
candidates.push_back(inst);
} else if (inst->IsBarrier()) {
// Remove all candidates -- do not move anything across barriers
candidates.clear();
}
}
// Work in scope of basic block
candidates.clear();
}
for (auto pair : pairs) {
// Replace a pair of instructions by a coalesced instruction
ReplacePair(pair);
}
}
void HandleArrayAccess(Inst *inst, VariableAnalysis &vars) {
Inst *obj = inst->GetObject();
Inst *idx = inst->GetIndex();
// If we don't know anything about index, do nothing
if (!vars.IsAnalyzed(idx)) {
candidates.push_back(inst);
return;
}
// Last candidates more likely to be coalesced
for (auto iter = candidates.rbegin(); iter != candidates.rend(); iter++) {
auto cand = *iter;
// Skip not interesting candidates: invalid and that cannot be coalesced with current inst
if (cand->IsMarked(invalid) || cand->GetOpcode() != inst->GetOpcode()) {
continue;
}
Inst *cand_obj = cand->GetObject();
auto cand_idx = cand->GetIndex();
// We need to have info about candidate's index and array objects must alias each other
if (!vars.IsAnalyzed(cand_idx) || obj.IsAlias(cand_obj) != MUST_ALIAS) {
continue;
}
// If both indices differs by one
Inst *position = FindBetterPlace(cand, inst);
if (poisition && vars.DiffersByConst(idx, cand_idx, 1)) {
pairs.push_back({cand, inst, position});
cand->SetMarker(invalid);
inst->SetMarket(invalid);
}
}
candidates.push_back(inst);
}
```
## Examples
### Loads and Stores with immediate indices
Before optimization
```
BB 0
prop: start
0.i64 Constant 0x2a -> (v3)
succs: [bb 2]
BB 2 preds: [bb 0]
3.ref NewArray 77 v0 -> (v42, v41)
41. SaveState v3(vr7) -> (v42)
42.ref NullCheck v3, v41 -> (v225, v229, v227, v230)
225.i64 LoadArrayI v42, 0x0 -> (v51)
227.i64 LoadArrayI v42, 0x1 -> (v51)
51.i64 Add v225, v227 -> (v229, v40, v230)
229.i64 StoreArrayI v42, 0x0, v51
230.i64 StoreArrayI v42, 0x1, v51
40.i64 Return v51
succs: [bb 1]
```
After optimization
```
BB 0
prop: start
0.i64 Constant 0x2a -> (v3)
succs: [bb 2]
BB 2 preds: [bb 0]
3.ref NewArray 77 v0 -> (v41, v42)
41. SaveState v3(vr7) -> (v42)
42.ref NullCheck v3, v41 -> (v231, v234)
231.i64 LoadArrayPairI v42, 0x0 -> (v232, v233)
232.i64 LoadPairPart v231, 0x0 -> (v51)
233.i64 LoadPairPart v231, 0x1 -> (v51)
51.i64 Add v232, v233 -> (v234, v234, v40)
234.i64 StoreArrayPairI v42, 0x0, v51, v51
40.i64 Return v51
succs: [bb 1]
```
### Coalescing inside loop
Before optimization
```
BB 2 preds: [bb 0]
3.i32 LenArray v0 -> (v35)
succs: [bb 3]
BB 3 preds: [bb 2, bb 3]
prop: head, loop 1
6p.i32 Phi v4(bb2), v34(bb3) -> (v33, v17, v34)
7p.i32 Phi v5(bb2), v24(bb3) -> (v24, v17)
8p.i32 Phi v5(bb2), v25(bb3) -> (v25, v23, v24)
17.i32 StoreArray v0, v6p, v7p
33.i32 AddI v6p, 0x1 -> (v23)
23.i32 StoreArray v0, v33, v8p
24.i32 Add v7p, v8p -> (v7p, v25)
25.i32 Add v8p, v24 -> (v8p)
34.i32 AddI v6p, 0x2 -> (v6p, v35)
35. If LT i32 v34, v3
succs: [bb 3, bb 4]
BB 4 preds: [bb 3]
29.void ReturnVoid
succs: [bb 1]
```
After optimization
```
BB 2 preds: [bb 0]
3.i32 LenArray v0 -> (v35)
succs: [bb 3]
BB 3 preds: [bb 2, bb 3]
prop: head, loop 1
6p.i32 Phi v4(bb2), v34(bb3) -> (v33, v36, v34)
7p.i32 Phi v5(bb2), v24(bb3) -> (v36, v24)
8p.i32 Phi v5(bb2), v25(bb3) -> (v36, v24, v25)
33.i32 AddI v6p, 0x1
36.i32 StoreArrayPair v0, v6p, v7p, v8p
24.i32 Add v7p, v8p -> (v7p, v25)
25.i32 Add v8p, v24 -> (v8p)
34.i32 AddI v6p, 0x2 -> (v6p, v35)
35. If LT i32 v34, v3
succs: [bb 3, bb 4]
BB 4 preds: [bb 3]
29.void ReturnVoid
succs: [bb 1]
```
## Options
| Option | Description | Default value |
| --- | --- | --- |
| `--compiler-memory-coalescing` | Enables optimization | `true` |
| `--compiler-memory-coalescing-objects` | Allows coalescing of operations with `ref`s | `true` |
| `--compiler-memory-coalescing-aligned` | Coalesces only aligned accesses (starting with even indices e.g. 0-1, 4-5 etc.) | `false` |
## Links
Source code:
[memory_coalescing.cpp](../optimizer/optimizations/memory_coalescing.cpp)
[memory_coalescing.h](../optimizer/optimizations/memory_coalescing.h)
Tests:
[memory_coalescing_test.cpp](../tests/memory_coalescing_test.cpp)

57
compiler/docs/object_type_check_elimination_doc.md
View File

@ -1,57 +0,0 @@
# Object Type Check Elimination
## Overview
**Object Type Check Elimination** - optimization which try to reduce number of IsInstance/CheckCast instructions.
## Rationality
Reduce number of instructions and remove unnecessary data-flow dependencies.
## Dependences
* RPO
* ObjectTypePropagation
## Algorithm
Visit `IsInstance` and `CheckCast` instructions in RPO order and try to elimiate them.
If instruction couldn't be eliminated, `ObjectTypeInfo` for input is dropped.
### IsInstance
`IsInstance` is replaced by 1 if an object in input can be cast to the resolved type, else replaced by 0. 'null' object is not an instance of any class.
`IsInstanceVisitor` also used in `Peephole` optimizations.
### CheckCast
If an object in input can't be cast to the resolved type `CheckCast` is replaced by deoptimize, else removed. 'null' object reference can be cast to every type.
`CheckCastVisitor` also used in `CheckCast` optimizations.
## Pseudocode
TODO
## Examples
```
.record A {}
.record B <extends=A> {}
.record C {}
...
newobj v0, B
lda.obj v0
isinstance A // will replaced by 1
newobj v0, C
lda.obj v0
isinstance A // will replaced by 0
...
newobj v0, B
lda.obj v0
checkcast A // will removed
checkcast C // will replaced by deoptimze
```
## Links
Source code:
[object_type_check_elimination.h](../optimizer/optimizations/object_type_check_elimination.h)
[object_type_check_elimination.cpp](../optimizer/optimizations/object_type_check_elimination.cpp)
Tests:
[isinstance_elimination_test.cpp](../../tests/checked/isinstance_elimination_test.pa)
[checkcast_elimination_test.cpp](../../tests/checked/checkcast_elimination_test.pa)

81
compiler/docs/redundant_loop_elimination_doc.md
View File

@ -1,81 +0,0 @@
# Redundant Loop Elimination
## Overview
**Redundant Loop Elimination(RLE)** - optimization which find and remove useless loops.
## Rationality
Reducing number of basic blocks and instructions.
## Dependence
* Loop analysis
## Algorithm
Visit loops in LRN order (first children, then parent).
For each loop check that:
* RLE is applied for all children loops.
* Loop doesn't contain instructions with side effect (ex. call instructions).
* Loop doesn't contain instructions with users out of the loop.
If all checks are true then loop is removing:
1. Loop pre-header connect with loop outer block.
2. Loop inner blocks disconnect from graph.
## Pseudocode
```
LoopVisitLRN(Loop* loop) {
for (auto inner_loop : loop->GetInnerLoops()) {
LoopVisitLRN(inner_loop);
}
if (Check(loop)) {
Remove(loop);
}
}
```
## Examples
Before RLE:
```
BB 0
prop: start
0.i64 Constant 0x0 -> (v4p)
1.i64 Constant 0x1 -> (v10)
2.i64 Constant 0xa -> (v5)
succs: [bb 3]
BB 3 preds: [bb 0, bb 4]
prop: head, loop 1
4p.i32 Phi v0(bb8), v10(bb4) -> (v5, v10)
5.b Compare LT i32 v4p, v2 -> (v6)
6. IfImm NE b v5, 0x0
succs: [bb 4, bb 5]
BB 5 preds: [bb 3]
12. ReturnVoid
succs: [bb 1]
BB 1 preds: [bb 5]
prop: end
BB 4 preds: [bb 3]
prop: loop 1
10.i32 Add v4p, v1 -> (v4p)
succs: [bb 3]
```
After RLE:
```
BB 0
prop: start
0.i64 Constant 0x0
1.i64 Constant 0x1
2.i64 Constant 0xa
succs: [bb 5]
BB 5 preds: [bb 0]
12. ReturnVoid
succs: [bb 1]
BB 1 preds: [bb 5]
prop: end
```
## Links
Source code:
[redundant_loop_elimination.cpp](../optimizer/optimizations/redundant_loop_elimination.cpp)
[redundant_loop_elimination.h](../optimizer/optimizations/redundant_loop_elimination.h)
Tests:
[redundant_loop_elimination_test.cpp](../tests/redundant_loop_elimination_test.cpp)

272
compiler/docs/scheduler_doc.md
View File

@ -1,272 +0,0 @@
# Instruction scheduling
## Overview
Rearrange adjacent instructions for better performance.
## Rationality
When instructions are executed on CPU they may stall the processor pipeline when input registers are not ready yet, because they are written by one of the previous instructions. Scheduling allows to reduce the amount of such stalls in pipeline.
## Dependence
* Dead Code Elimination(DCE)
* Remove Empty Blocks
* Remove Linear blocks
* Reverse Post Order(RPO)
## Algorithm
Current decisions/limitations:
* Scheduler pass is placed immediately before register allocation
* It rearranges instructions only inside the basic block, but not between them
* No liveness analysis, only calculating dependencies using barrier/users/alias information
* No CPU pipeline/resource modeling, only having dependency costs
* Forward list scheduling algorithm with standard critical-path-based priority
For each basic block we first scan instructions in reverse order marking barriers and calculating the dependencies.
Together with dependencies we calculate priority as a longest (critical) path to leaf instructions in basic block dependency graph.
Than we schedule each interval between barriers using the following algorithm.
There are two priority queues, `waiting` and `ready`. `ready` queue is sorted based on priority calculated previously, while `waiting` queue is based on so-called `ASAP` (as soon as possible) values. In initialization, `ready` is empty and `waiting` contains all leaf instructions (without incoming dependencies), their `ASAP` is 1.
`ASAP` value for each instruction is changed only before it enters the `waiting` queue, and remains unchanged since that time.
Algorithm starts from tick `cycle` 1. If `ready` queue is empty we look through "soonest" instruction from `waiting` queue and if we need to skip some ticks without scheduling any instruction we have to adjust `cycle` value.
Next, we move all already available instructions (`ASAP` <= `cycle`) from `waiting` queue into `ready` queue.
Finally, we extract top instruction from `ready` queue and add it into new schedule. At this moment we adjust `ASAP` value for all dependent instructions and add some of them (which depend only on already scheduled instructions) into `waiting` queue.
## Pseudocode
```c++
Scheduler::RunImpl() {
for (auto bb : GetGraph()->GetBlocksRPO())
ScheduleBasicBlock(bb);
}
// Dependency helper function
void Scheduler::AddDep(uint32_t* prio, Inst* from, Inst* to, uint32_t latency, Inst* barrier) {
// Update instruction priority - "how high instruction is in dependency tree"
*prio = std::max(*prio, latency + prio_[to]);
// Do not add cross-barrier dependencies into deps_
if (barrier == nullptr || old_[to] > old_[barrier]) {
if (deps_.at(from).count(to) == 1) {
uint32_t old_latency = deps_.at(from).at(to);
if (old_latency >= latency)
return;
} else
num_deps_[to]++;
deps_.at(from)[to] = latency;
}
}
// Rearranges instructions in the basic block using list scheduling algorithm.
Scheduler::ScheduleBasicBlock(BasicBlock* bb) {
// Calculate priority and dependencies
uint32_t num_inst = 0;
Inst* last_barrier = nullptr;
for (auto inst : bb->InstsSafeReverse()) {
uint32_t prio = 0;
old_.insert({inst, num_inst++});
num_deps_.insert({inst, 0U});
deps_.emplace(inst, GetGraph()->GetLocalAllocator()->Adapter());
// Dependency to the barrier
if (last_barrier != nullptr)
AddDep(&prio, inst, last_barrier, 1U, last_barrier);
// Dependency from barrier
if (barrier) {
Inst* old_last_barrier = last_barrier;
last_barrier = inst;
num_barriers++;
for (auto user = inst->GetNext(); user != old_last_barrier; user = user->GetNext())
AddDep(&prio, inst, user, 1U, last_barrier);
}
// Users
for (auto& user_item : inst->GetUsers()) {
auto user = user_item.GetInst();
AddDep(&prio, inst, user, inst_latency, last_barrier);
}
.... // Memory dependencies calculation
... // CanThrow or SaveState can't be rearranged, and stores can't be moved over them
prio_.insert({inst, prio});
}
// Schedule intervals between barriers
uint32_t cycle = 0;
num_inst = 0;
Inst* first = nullptr;
for (auto inst = bb->GetFirstInst(); inst != nullptr; inst = inst->GetNext()) {
bool barrier = inst->IsBarrier();
num_inst++;
if (first == nullptr)
first = inst;
if (barrier || inst == bb->GetLastInst()) {
Inst* last = nullptr;
if (barrier) {
last = inst->GetPrev();
num_inst--;
} else
last = inst;
if (num_inst > 1)
cycle += ScheduleInstsBetweenBarriers(first, last);
else if (num_inst == 1) {
sched_.push_back(first);
cycle++;
}
if (barrier) {
sched_.push_back(inst);
cycle++;
}
num_inst = 0;
first = nullptr;
}
}
... // Here we rearrange instructions in basic block according to sched_
}
// Schedule instructions between [first..last] inclusive, none of them are barriers.
uint32_t Scheduler::ScheduleInstsBetweenBarriers(Inst* first, Inst* last) {
// Compare function for 'waiting' queue
auto cmp_asap = [this](Inst* left, Inst* right) {
return asap_[left] > asap_[right] || (asap_[left] == asap_[right] && old_[left] < old_[right]);
};
// Queue of instructions, which dependencies are scheduled already, but they are still not finished yet
std::priority_queue<Inst*, ArenaVector<Inst*>, decltype(cmp_asap)> waiting(
cmp_asap, GetGraph()->GetLocalAllocator()->Adapter());
// Compare function for 'ready' queue
auto cmp_prio = [this](Inst* left, Inst* right) {
return prio_[left] < prio_[right] || (prio_[left] == prio_[right] && old_[left] < old_[right]);
};
// Queue of ready instructions
std::priority_queue<Inst*, ArenaVector<Inst*>, decltype(cmp_prio)> ready(
cmp_prio, GetGraph()->GetLocalAllocator()->Adapter());
// Initialization, add leafs into 'waiting' queue
uint32_t num_inst = 0;
for (auto inst = first; inst != last->GetNext(); inst = inst->GetNext()) {
asap_.insert({inst, 1U});
if (num_deps_[inst] == 0)
waiting.push(inst);
num_inst++;
}
// Scheduling
uint32_t cycle = 1;
while (num_inst > 0) {
if (ready.empty()) {
uint32_t nearest = asap_[waiting.top()];
// Skipping cycles where we can't schedule any instruction
if (nearest > cycle)
cycle = nearest;
}
// Move from 'waiting' to 'ready'
while (!waiting.empty()) {
Inst* soonest = waiting.top();
if (asap_[soonest] <= cycle) {
waiting.pop();
ready.push(soonest);
} else
break;
}
// Extract top 'ready' instruction
auto cur = ready.top();
ready.pop();
// Adjust all dependent instructions
for (auto pair : deps_.at(cur)) {
// Adjust asap
uint32_t asap = asap_[pair.first];
asap = std::max(asap, cycle + pair.second);
asap_[pair.first] = asap;
// Adjust num_deps
uint32_t num_deps = num_deps_[pair.first];
num_deps--;
num_deps_[pair.first] = num_deps;
if (num_deps == 0 && pair.first->GetOpcode() != Opcode::LoadPairPart)
waiting.push(pair.first);
}
// Add into schedule
sched_.push_back(cur);
num_inst--;
cycle++;
}
asap_.clear();
return cycle;
}
```
## Examples
IR Before optimization:
```
BB 0
prop: start
0.i64 Constant 0x2a -> (v8)
1.i64 Constant 0x2b -> (v8)
2.i64 Constant 0x2c -> (v9)
3.i64 Constant 0x2d -> (v9)
4.i64 Constant 0x2e -> (v11)
5.i64 Constant 0x2f -> (v11)
6.i64 Constant 0x30 -> (v12)
7.i64 Constant 0x31 -> (v12)
succs: [bb 2]
BB 2 preds: [bb 0]
8.u64 Add v0, v1 -> (v10)
9.u64 Add v2, v3 -> (v10)
10.u64 Add v8, v9 -> (v14)
11.u64 Add v4, v5 -> (v13)
12.u64 Add v6, v7 -> (v13)
13.u64 Add v11, v12 -> (v14)
14.u64 Add v10, v13 -> (v15)
15.u64 Return v14
succs: [bb 1]
BB 1 preds: [bb 2]
prop: end
```
IR after optimization:
```
BB 0
prop: start
0.i64 Constant 0x2a -> (v8)
1.i64 Constant 0x2b -> (v8)
2.i64 Constant 0x2c -> (v9)
3.i64 Constant 0x2d -> (v9)
4.i64 Constant 0x2e -> (v11)
5.i64 Constant 0x2f -> (v11)
6.i64 Constant 0x30 -> (v12)
7.i64 Constant 0x31 -> (v12)
succs: [bb 2]
BB 2 preds: [bb 0]
8.u64 Add v0, v1 -> (v10)
9.u64 Add v2, v3 -> (v10)
11.u64 Add v4, v5 -> (v13)
12.u64 Add v6, v7 -> (v13)
10.u64 Add v8, v9 -> (v14)
13.u64 Add v11, v12 -> (v14)
14.u64 Add v10, v13 -> (v15)
15.u64 Return v14
succs: [bb 1]
BB 1 preds: [bb 2]
prop: end
```
Instruction 10 was moved down.
## Links
Algorithm: [article](http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.211.7673&rep=rep1&type=pdf)
Source code:
[scheduler.cpp](../optimizer/optimizations/scheduler.cpp)
[scheduler.h](../optimizer/optimizations/scheduler.h)
Tests:
[scheduler_test.cpp](../tests/scheduler_test.cpp)

1134
compiler/intrinsics.yaml Normal file
View File

File diff suppressed because it is too large Load Diff

342
compiler/optimizer/analysis/alias_analysis.cpp
View File

@ -1,342 +0,0 @@
/**
* Copyright (c) 2021-2022 Huawei Device Co., Ltd.
* 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 "optimizer/ir/basicblock.h"
#include "optimizer/ir/graph.h"
#include "compiler_logger.h"
#include "optimizer/analysis/alias_analysis.h"
/**
* See "Efficient Field-sensitive pointer analysis for C" by "David J. Pearce
* and Paul H. J. Kelly and Chris Hankin
*
* We treat each IR Inst as a constraint that may be applied to a set of
* aliases of some virtual register. Virtual registers are used as constraint
* variables as well.
*
* In order to solve the system of set constraints, the following is done:
*
* 1. Each constraint variable x has a solution set associated with it, Sol(x).
* Implemented through AliasAnalysis::points_to_ that contains mapping of
* virtual register to possible aliases.
*
* 2. Constraints are separated into direct, copy.
*
* - Direct constraints are constraints that require no extra processing, such
* as P = &Q.
*
* - Copy constraints are those of the form P = Q. Such semantic can be
* obtained through NullCheck, Mov, and Phi instructions.
*
* 3. All direct constraints of the form P = &Q are processed, such that Q is
* added to Sol(P).
*
* 4. A directed graph is built out of the copy constraints. Each constraint
* variable is a node in the graph, and an edge from Q to P is added for each
* copy constraint of the form P = Q.
*
* 5. The graph is then walked, and solution sets are propagated along the copy
* edges, such that an edge from Q to P causes Sol(P) <- Sol(P) union Sol(Q).
*
* 6. The process of walking the graph is iterated until no solution sets
* change.
*
* To add new instructions to alias analysis please consider following:
* - AliasAnalysis class: to add a visitor for a new instruction that should be analyzed
*
* TODO(Evgenii Kudriashov): Prior to walking the graph in steps 5 and 6, We
* need to perform static cycle elimination on the constraint graph, as well as
* off-line variable substitution.
*
* TODO(Evgenii Kudriashov): To add flow-sensitivity the "Flow-sensitive
* pointer analysis for millions of lines of code" by Ben Hardekopf and Calvin
* Lin may be considered.
*
* TODO(Evgenii Kudriashov): After implementing VRP and SCEV the "Loop-Oriented
* Array- and Field-Sensitive Pointer Analysis for Automatic SIMD
* Vectorization" by Yulei Sui, Xiaokang Fan, Hao Zhou, and Jingling Xue may be
* considered to add advanced analysis of array indices.
*/
namespace panda::compiler {
AliasAnalysis::AliasAnalysis(Graph *graph) : Analysis(graph), points_to_(graph->GetAllocator()->Adapter()) {}
const ArenaVector<BasicBlock *> &AliasAnalysis::GetBlocksToVisit() const
{
return GetGraph()->GetBlocksRPO();
}
bool AliasAnalysis::RunImpl()
{
Init();
VisitGraph();
// Initialize solution sets
for (auto pair : *direct_) {
auto it = points_to_.try_emplace(pair.first, GetGraph()->GetAllocator()->Adapter());
ASSERT(pair.first.GetBase() == nullptr || pair.first.GetBase()->GetOpcode() != Opcode::NullCheck);
ASSERT(pair.second.GetBase() == nullptr || pair.second.GetBase()->GetOpcode() != Opcode::NullCheck);
it.first->second.insert(pair.second);
}
SolveConstraints();
#ifndef NDEBUG
if (CompilerLogger::IsComponentEnabled(CompilerLoggerComponents::ALIAS_ANALYSIS)) {
std::ostringstream out;
DumpChains(&out);
Dump(&out);
COMPILER_LOG(DEBUG, ALIAS_ANALYSIS) << out.str();
}
#endif
return true;
}
void AliasAnalysis::Init()
{
auto allocator = GetGraph()->GetLocalAllocator();
chains_ = allocator->New<PointerMap<ArenaVector<Pointer>>>(allocator->Adapter());
direct_ = allocator->New<PointerPairVector>(allocator->Adapter());
inputs_set_ = allocator->New<ArenaSet<Inst *>>(allocator->Adapter());
ASSERT(chains_ != nullptr);
ASSERT(direct_ != nullptr);
ASSERT(inputs_set_ != nullptr);
points_to_.clear();
}
void Pointer::Dump(std::ostream *out) const
{
switch (type_) {
case OBJECT:
(*out) << "v" << base_->GetId();
break;
case STATIC_FIELD:
(*out) << "SF #" << imm_;
break;
case POOL_CONSTANT:
(*out) << "PC #" << imm_;
break;
case OBJECT_FIELD:
(*out) << "v" << base_->GetId() << " #" << imm_;
break;
case ARRAY_ELEMENT:
(*out) << "v" << base_->GetId() << "[";
if (idx_ != nullptr) {
(*out) << "v" << idx_->GetId();
if (imm_ != 0) {
(*out) << "+" << imm_;
}
} else {
(*out) << imm_;
}
(*out) << "]";
break;
default:
UNREACHABLE();
}
if (local_) {
(*out) << "(local)";
}
if (volatile_) {
(*out) << "(v)";
}
}
static bool PointerLess(const Pointer &lhs, const Pointer &rhs)
{
if (lhs.GetBase() == rhs.GetBase()) {
return lhs.GetImm() < rhs.GetImm();
}
if (lhs.GetBase() == nullptr) {
return true;
}
if (rhs.GetBase() == nullptr) {
return false;
}
return lhs.GetBase()->GetId() < rhs.GetBase()->GetId();
}
void AliasAnalysis::DumpChains(std::ostream *out) const
{
ArenaVector<Pointer> sorted_keys(GetGraph()->GetLocalAllocator()->Adapter());
for (auto &pair : *chains_) {
sorted_keys.push_back(pair.first);
}
std::sort(sorted_keys.begin(), sorted_keys.end(), PointerLess);
(*out) << "The chains are the following:" << std::endl;
for (auto &p : sorted_keys) {
(*out) << "\t";
p.Dump(out);
(*out) << ": {";
// Sort by instruction ID to add more readability to logs
ArenaVector<Pointer> sorted(chains_->at(p), GetGraph()->GetLocalAllocator()->Adapter());
std::sort(sorted.begin(), sorted.end(), PointerLess);
auto edge = sorted.begin();
if (edge != sorted.end()) {
edge->Dump(out);
while (++edge != sorted.end()) {
(*out) << ", ";
edge->Dump(out);
}
}
(*out) << "}" << std::endl;
}
}
void AliasAnalysis::Dump(std::ostream *out) const
{
ArenaVector<Pointer> sorted_keys(GetGraph()->GetLocalAllocator()->Adapter());
for (auto &pair : points_to_) {
sorted_keys.push_back(pair.first);
}
std::sort(sorted_keys.begin(), sorted_keys.end(), PointerLess);
(*out) << "The solution set is the following:" << std::endl;
for (auto &p : sorted_keys) {
(*out) << "\t";
p.Dump(out);
(*out) << ": {";
// Sort by instruction ID to add more readability to logs
auto values = points_to_.at(p);
ArenaVector<Pointer> sorted(values.begin(), values.end(), GetGraph()->GetLocalAllocator()->Adapter());
std::sort(sorted.begin(), sorted.end(), PointerLess);
auto iter = sorted.begin();
if (iter != sorted.end()) {
iter->Dump(out);
while (++iter != sorted.end()) {
(*out) << ", ";
iter->Dump(out);
}
}
(*out) << "}" << std::endl;
}
}
AliasType AliasAnalysis::CheckInstAlias(Inst *mem1, Inst *mem2) const
{
return MAY_ALIAS;
}
/**
* Here we propagate solutions obtained from direct constraints through copy
* constraints e.g: we have a node A with solution {a} and the node A was
* copied to B and C (this->chains_ maintains these links), and C was copied to
* D.
*
* A{a} -> B
* \-> C -> D
*
* After first iteration (iterating A node) we will obtain
*
* A{a} -> B{a}
* \-> C{a} -> D
*
* After second iteration (iterating B node) nothing changes
*
* After third iteration (iterating C node):
*
* A{a} -> B{a}
* \-> C{a} -> D{a}
*
* For complex nodes (OBJECT_FIELD, ARRAY_ELEMENT) we create auxiliary nodes e.g.
* if a field F was accessed from object A then we have two nodes:
*
* A{a} -> A.F
*
* And solutions from A would be propagated as following:
*
* A{a} -> A.F{a.F}
*
* The function works using worklist to process only updated nodes.
*/
void AliasAnalysis::SolveConstraints()
{
ArenaQueue<Pointer> worklist(GetGraph()->GetLocalAllocator()->Adapter());
for (auto &pair : *direct_) {
if (chains_->find(pair.first) != chains_->end()) {
worklist.push(pair.first);
}
}
while (!worklist.empty()) {
Pointer &ref = worklist.front();
ASSERT(ref.GetBase() == nullptr || ref.GetBase()->GetOpcode() != Opcode::NullCheck);
for (auto &edge : chains_->at(ref)) {
// POOL_CONSTANT cannot be assignee
ASSERT(edge.GetType() != POOL_CONSTANT);
auto &sols = points_to_.try_emplace(edge, GetGraph()->GetAllocator()->Adapter()).first->second;
bool added = false;
for (auto &alias : points_to_.at(ref)) {
ASSERT(alias.GetBase() == nullptr || alias.GetBase()->GetOpcode() != Opcode::NullCheck);
if (edge.GetType() == OBJECT_FIELD && ref.GetBase() == edge.GetBase()) {
// Propagating from object to fields: A{a} -> A.F{a.f}
if (alias.GetType() == OBJECT) {
Pointer p = Pointer::CreateObjectField(alias.GetBase(), edge.GetImm(), edge.GetTypePtr());
p.SetLocalVolatile(alias.IsLocal(), edge.IsVolatile());
added |= sols.insert(p).second;
continue;
}
// In case A{a.g} -> A.F we propagate symbolic name: A{a.g} -> A.F{A.F}
Pointer p = Pointer::CreateObjectField(ref.GetBase(), edge.GetImm(), edge.GetTypePtr());
p.SetLocalVolatile(alias.IsLocal(), edge.IsVolatile());
added |= sols.insert(p).second;
continue;
}
if (edge.GetType() == ARRAY_ELEMENT && ref.GetBase() == edge.GetBase()) {
// Propagating from object to elements: A{a} -> A[i]{a[i]}
if (alias.GetType() == OBJECT) {
Pointer p = Pointer::CreateArrayElement(alias.GetBase(), edge.GetIdx(), edge.GetImm());
p.SetLocalVolatile(alias.IsLocal(), edge.IsVolatile());
added |= sols.insert(p).second;
continue;
}
// In case A{a[j]} -> A[i] we propagate symbolic name: A{a[j]} -> A[i]{A[i]}
Pointer p = Pointer::CreateArrayElement(ref.GetBase(), edge.GetIdx(), edge.GetImm());
p.SetLocalVolatile(alias.IsLocal(), edge.IsVolatile());
added |= sols.insert(p).second;
continue;
}
added |= sols.insert(alias).second;
}
if (added && chains_->find(edge) != chains_->end()) {
worklist.push(edge);
}
ASSERT(!sols.empty());
}
worklist.pop();
}
}
/**
* Instructions that introduce aliases.
*/
void AliasAnalysis::VisitCastAnyTypeValue(GraphVisitor *v, Inst *inst)
{
if (inst->GetType() == DataType::REFERENCE) {
static_cast<AliasAnalysis *>(v)->AddDirectEdge(Pointer::CreateObject(inst));
}
}
} // namespace panda::compiler

235
compiler/optimizer/analysis/alias_analysis.h
View File

@ -1,235 +0,0 @@
/**
* Copyright (c) 2021-2022 Huawei Device Co., Ltd.
* 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.
*/
#ifndef COMPILER_OPTIMIZER_ANALYSIS_ALIAS_ANALYSIS_H_
#define COMPILER_OPTIMIZER_ANALYSIS_ALIAS_ANALYSIS_H_
#include <unordered_map>
#include "optimizer/ir/graph_visitor.h"
#include "optimizer/pass.h"
#include "utils/arena_containers.h"
#include "utils/hash.h"
namespace panda::compiler {
class BasicBlock;
class Graph;
enum AliasType : uint8_t {
// Proved that references are not aliases
NO_ALIAS,
// References may or may not alias each other (cannot be proven statically)
MAY_ALIAS,
// References are proven aliases
MUST_ALIAS
};
enum PointerType {
// Reference to unknown object.
// Valid fields: base
OBJECT,
// Constant from pool
// Valid fields: imm
POOL_CONSTANT,
// Object's field
// Valid fields: base, imm, type_ptr
OBJECT_FIELD,
// Static field of the object
// Valid fields: imm, type_ptr
STATIC_FIELD,
// Array pointer
// Valid fields: base, idx
ARRAY_ELEMENT
};
class Pointer {
public:
Pointer() = default;
Pointer(PointerType type, const Inst *base, const Inst *idx, uint64_t imm, const void *type_ptr)
: type_(type), base_(base), idx_(idx), imm_(imm), type_ptr_(type_ptr), volatile_(false)
{
local_ = false;
};
static Pointer CreateObject(const Inst *base)
{
return Pointer(OBJECT, base, nullptr, 0, nullptr);
}
static Pointer CreateObjectField(const Inst *base, uint32_t type_id, const void *type_ptr = nullptr)
{
return Pointer(OBJECT_FIELD, base, nullptr, type_id, type_ptr);
}
static Pointer CreateArrayElement(const Inst *array, const Inst *idx, uint64_t imm = 0)
{
return Pointer(ARRAY_ELEMENT, array, idx, imm, nullptr);
}
PointerType GetType() const
{
return type_;
}
const Inst *GetBase() const
{
return base_;
}
const Inst *GetIdx() const
{
return idx_;
}
uint64_t GetImm() const
{
return imm_;
}
const void *GetTypePtr() const
{
return type_ptr_;
}
bool IsLocal() const
{
return local_;
}
void SetLocalVolatile(bool local, bool is_volatile)
{
local_ = local;
volatile_ = is_volatile;
}
bool IsVolatile() const
{
return volatile_;
}
void Dump(std::ostream *out) const;
bool HasSameOffset(const Pointer &p) const
{
if (type_ptr_ == nullptr && p.type_ptr_ == nullptr) {
return imm_ == p.imm_;
}
return type_ptr_ == p.type_ptr_;
}
private:
PointerType type_;
const Inst *base_;
const Inst *idx_;
uint64_t imm_;
const void *type_ptr_;
bool local_;
bool volatile_;
};
struct PointerEqual {
bool operator()(Pointer const &p1, Pointer const &p2) const
{
return p1.GetType() == p2.GetType() && p1.GetBase() == p2.GetBase() && p1.GetIdx() == p2.GetIdx() &&
p1.HasSameOffset(p2);
}
};
struct PointerHash {
uint32_t operator()(Pointer const &p) const
{
auto inst_hasher = std::hash<const Inst *> {};
uint32_t hash = inst_hasher(p.GetBase());
hash += inst_hasher(p.GetIdx());
if (p.GetTypePtr() == nullptr) {
hash += std::hash<uint64_t> {}(p.GetImm());
} else {
hash += std::hash<const void *> {}(p.GetTypePtr());
}
return hash;
}
};
// NOLINTNEXTLINE(fuchsia-multiple-inheritance)
class AliasAnalysis : public Analysis, public GraphVisitor {
public:
enum class Trilean {
TRUE,
UNKNOWN,
FALSE,
};
using PointerPairVector = ArenaVector<std::pair<Pointer, Pointer>>;
explicit AliasAnalysis(Graph *graph);
NO_MOVE_SEMANTIC(AliasAnalysis);
NO_COPY_SEMANTIC(AliasAnalysis);
~AliasAnalysis() override = default;
bool RunImpl() override;
const char *GetPassName() const override
{
return "AliasAnalysis";
}
AliasType CheckInstAlias(Inst *mem1, Inst *mem2) const;
void Dump(std::ostream *out) const;
/**
* Sort IR instructions into two constraint groups:
* Direct: introduce the alias
* Copy: copy one alias to another
*/
const ArenaVector<BasicBlock *> &GetBlocksToVisit() const override;
/**
* Instructions that introduce aliases.
*/
static void VisitCastAnyTypeValue(GraphVisitor *v, Inst *inst);
void AddDirectEdge(const Pointer &p)
{
direct_->push_back({p, p});
}
ArenaSet<Inst *> *GetClearInputsSet()
{
inputs_set_->clear();
return inputs_set_;
}
#include "optimizer/ir/visitor.inc"
private:
void Init();
using PointerSet = ArenaUnorderedSet<Pointer, PointerHash, PointerEqual>;
template <class T>
using PointerMap = ArenaUnorderedMap<Pointer, T, PointerHash, PointerEqual>;
void SolveConstraints();
void DumpChains(std::ostream *out) const;
private:
PointerMap<PointerSet> points_to_;
// Local containers:
PointerMap<ArenaVector<Pointer>> *chains_ {nullptr};
PointerPairVector *direct_ {nullptr};
ArenaSet<Inst *> *inputs_set_ {nullptr};
};
} // namespace panda::compiler
#endif // COMPILER_OPTIMIZER_ANALYSIS_ALIAS_ANALYSIS_H_

537
compiler/optimizer/analysis/bounds_analysis.cpp
View File

@ -1,537 +0,0 @@
/**
* Copyright (c) 2021-2022 Huawei Device Co., Ltd.
* 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 "bounds_analysis.h"
#include "dominators_tree.h"
#include "optimizer/ir/graph.h"
#include "optimizer/ir/graph_visitor.h"
#include "optimizer/ir/basicblock.h"
#include "optimizer/ir/inst.h"
#include "compiler/optimizer/ir/analysis.h"
#include "optimizer/analysis/loop_analyzer.h"
namespace panda::compiler {
BoundsRange::BoundsRange(int64_t val, DataType::Type type) : BoundsRange(val, val, nullptr, type) {}
BoundsRange::BoundsRange(int64_t left, int64_t right, const Inst *inst, [[maybe_unused]] DataType::Type type)
: left_(left), right_(right), len_array_(inst)
{
ASSERT(inst == nullptr);
ASSERT(left <= right);
ASSERT(GetMin(type) <= left);
ASSERT(right <= GetMax(type));
}
int64_t BoundsRange::GetLeft() const
{
return left_;
}
int64_t BoundsRange::GetRight() const
{
return right_;
}
bool BoundsRange::IsConst() const
{
return left_ == right_;
}
bool BoundsRange::IsMaxRange(DataType::Type type) const
{
return left_ <= GetMin(type) && right_ >= GetMax(type);
}
bool BoundsRange::IsEqual(const BoundsRange &range) const
{
return left_ == range.GetLeft() && right_ == range.GetRight();
}
bool BoundsRange::IsLess(const BoundsRange &range) const
{
return right_ < range.GetLeft();
}
bool BoundsRange::IsLess(const Inst *inst) const
{
return false;
}
bool BoundsRange::IsMore(const BoundsRange &range) const
{
return left_ > range.GetRight();
}
bool BoundsRange::IsMoreOrEqual(const BoundsRange &range) const
{
return left_ >= range.GetRight();
}
bool BoundsRange::IsNotNegative() const
{
return left_ >= 0;
}
bool BoundsRange::IsNegative() const
{
return right_ < 0;
}
/**
* Return the minimal value for a type.
*
* We consider that REFERENCE type has only non-negative address values
*/
int64_t BoundsRange::GetMin(DataType::Type type)
{
ASSERT(!IsFloatType(type));
switch (type) {
case DataType::BOOL:
case DataType::UINT8:
case DataType::UINT16:
case DataType::UINT32:
case DataType::UINT64:
case DataType::REFERENCE:
return 0;
case DataType::INT8:
return INT8_MIN;
case DataType::INT16:
return INT16_MIN;
case DataType::INT32:
return INT32_MIN;
case DataType::INT64:
return INT64_MIN;
default:
UNREACHABLE();
}
}
/**
* Return the maximal value for a type.
*
* For REFERENCE we are interested in whether it is NULL or not. Set the
* maximum to INT64_MAX regardless the real architecture bitness.
*/
int64_t BoundsRange::GetMax(DataType::Type type)
{
ASSERT(!IsFloatType(type));
ASSERT(type != DataType::UINT64);
switch (type) {
case DataType::BOOL:
return 1;
case DataType::UINT8:
return UINT8_MAX;
case DataType::UINT16:
return UINT16_MAX;
case DataType::UINT32:
return UINT32_MAX;
case DataType::INT8:
return INT8_MAX;
case DataType::INT16:
return INT16_MAX;
case DataType::INT32:
return INT32_MAX;
// NOLINTNEXTLINE(bugprone-branch-clone)
case DataType::INT64:
return INT64_MAX;
case DataType::REFERENCE:
return INT64_MAX;
default:
UNREACHABLE();
}
}
BoundsRange BoundsRange::FitInType(DataType::Type type) const
{
auto type_min = BoundsRange::GetMin(type);
auto type_max = BoundsRange::GetMax(type);
if (left_ < type_min || left_ > type_max || right_ < type_min || right_ > type_max) {
return BoundsRange(type_min, type_max);
}
return *this;
}
BoundsRange BoundsRange::Union(const ArenaVector<BoundsRange> &ranges)
{
int64_t min = MAX_RANGE_VALUE;
int64_t max = MIN_RANGE_VALUE;
for (const auto &range : ranges) {
if (range.GetLeft() < min) {
min = range.GetLeft();
}
if (range.GetRight() > max) {
max = range.GetRight();
}
}
return BoundsRange(min, max);
}
BoundsRange::RangePair BoundsRange::NarrowBoundsByNE(BoundsRange::RangePair const &ranges)
{
auto &[left_range, right_range] = ranges;
int64_t ll = left_range.GetLeft();
int64_t lr = left_range.GetRight();
int64_t rl = right_range.GetLeft();
int64_t rr = right_range.GetRight();
// We can narrow bounds of a range if another is a constant and matches one of the bounds
// Mostly needed for a reference comparison with null
if (left_range.IsConst() && !right_range.IsConst()) {
if (ll == rl) {
return {left_range, BoundsRange(rl + 1, rr)};
}
if (ll == rr) {
return {left_range, BoundsRange(rl, rr - 1)};
}
}
if (!left_range.IsConst() && right_range.IsConst()) {
if (rl == ll) {
return {BoundsRange(ll + 1, lr), right_range};
}
if (rl == lr) {
return {BoundsRange(ll, lr - 1), right_range};
}
}
return ranges;
}
BoundsRange::RangePair BoundsRange::NarrowBoundsCase1(ConditionCode cc, BoundsRange::RangePair const &ranges)
{
auto &[left_range, right_range] = ranges;
int64_t lr = left_range.GetRight();
int64_t rl = right_range.GetLeft();
if (cc == ConditionCode::CC_GT || cc == ConditionCode::CC_A) {
// With equal rl and lr left_range cannot be greater than right_range
if (rl == lr) {
return {BoundsRange(), BoundsRange()};
}
return {BoundsRange(rl + 1, lr), BoundsRange(rl, lr - 1)};
}
if (cc == ConditionCode::CC_GE || cc == ConditionCode::CC_AE || cc == ConditionCode::CC_EQ) {
return {BoundsRange(rl, lr), BoundsRange(rl, lr)};
}
return ranges;
}
BoundsRange::RangePair BoundsRange::NarrowBoundsCase2(ConditionCode cc, BoundsRange::RangePair const &ranges)
{
if (cc == ConditionCode::CC_GT || cc == ConditionCode::CC_GE || cc == ConditionCode::CC_EQ ||
cc == ConditionCode::CC_A || cc == ConditionCode::CC_AE) {
return {BoundsRange(), BoundsRange()};
}
return ranges;
}
BoundsRange::RangePair BoundsRange::NarrowBoundsCase3(ConditionCode cc, BoundsRange::RangePair const &ranges)
{
auto &[left_range, right_range] = ranges;
int64_t ll = left_range.GetLeft();
int64_t lr = left_range.GetRight();
int64_t rl = right_range.GetLeft();
int64_t rr = right_range.GetRight();
if (cc == ConditionCode::CC_GT || cc == ConditionCode::CC_A) {
// rl == lr handled in case 1
return {BoundsRange(rl + 1, lr), right_range};
}
if (cc == ConditionCode::CC_GE || cc == ConditionCode::CC_AE) {
return {BoundsRange(rl, lr), right_range};
}
if (cc == ConditionCode::CC_LT || cc == ConditionCode::CC_B) {
// With equal ll and rr left_range cannot be less than right_range
if (ll == rr) {
return {BoundsRange(), BoundsRange()};
}
return {BoundsRange(ll, rr - 1), right_range};
}
if (cc == ConditionCode::CC_LE || cc == ConditionCode::CC_BE) {
return {BoundsRange(ll, rr), right_range};
}
if (cc == ConditionCode::CC_EQ) {
return {BoundsRange(rl, rr), right_range};
}
return ranges;
}
BoundsRange::RangePair BoundsRange::NarrowBoundsCase4(ConditionCode cc, BoundsRange::RangePair const &ranges)
{
auto &[left_range, right_range] = ranges;
int64_t ll = left_range.GetLeft();
int64_t rr = right_range.GetRight();
if (cc == ConditionCode::CC_LT || cc == ConditionCode::CC_B) {
// With equal ll and rr left_range cannot be less than right_range
if (ll == rr) {
return {BoundsRange(), BoundsRange()};
}
return {BoundsRange(ll, rr - 1), BoundsRange(ll + 1, rr)};
}
if (cc == ConditionCode::CC_LE || cc == ConditionCode::CC_BE || cc == ConditionCode::CC_EQ) {
return {BoundsRange(ll, rr), BoundsRange(ll, rr)};
}
return ranges;
}
BoundsRange::RangePair BoundsRange::NarrowBoundsCase5(ConditionCode cc, BoundsRange::RangePair const &ranges)
{
if (cc == ConditionCode::CC_LT || cc == ConditionCode::CC_LE || cc == ConditionCode::CC_EQ ||
cc == ConditionCode::CC_B || cc == ConditionCode::CC_BE) {
return {BoundsRange(), BoundsRange()};
}
return ranges;
}
BoundsRange::RangePair BoundsRange::NarrowBoundsCase6(ConditionCode cc, BoundsRange::RangePair const &ranges)
{
auto &[left_range, right_range] = ranges;
int64_t ll = left_range.GetLeft();
int64_t lr = left_range.GetRight();
int64_t rl = right_range.GetLeft();
int64_t rr = right_range.GetRight();
if (cc == ConditionCode::CC_GT || cc == ConditionCode::CC_A) {
// rl == lr handled in case 1
return {left_range, BoundsRange(rl, lr - 1)};
}
if (cc == ConditionCode::CC_GE || cc == ConditionCode::CC_AE) {
return {left_range, BoundsRange(rl, lr)};
}
if (cc == ConditionCode::CC_LT || cc == ConditionCode::CC_B) {
// ll == rr handled in case 4
return {left_range, BoundsRange(ll + 1, rr)};
}
if (cc == ConditionCode::CC_LE || cc == ConditionCode::CC_BE) {
return {left_range, BoundsRange(ll, rr)};
}
if (cc == ConditionCode::CC_EQ) {
return {left_range, BoundsRange(ll, lr)};
}
return ranges;
}
/**
* Try narrow bounds range for <if (left CC right)> situation
* Return a pair of narrowed left and right intervals
*/
BoundsRange::RangePair BoundsRange::TryNarrowBoundsByCC(ConditionCode cc, BoundsRange::RangePair const &ranges)
{
if (cc == ConditionCode::CC_NE) {
return NarrowBoundsByNE(ranges);
}
auto &[left_range, right_range] = ranges;
int64_t ll = left_range.GetLeft();
int64_t lr = left_range.GetRight();
int64_t rl = right_range.GetLeft();
int64_t rr = right_range.GetRight();
// For further description () is for left_range bounds and [] is for right_range bounds
// case 1: ( [ ) ]
if (ll <= rl && rl <= lr && lr <= rr) {
return NarrowBoundsCase1(cc, ranges);
}
// case 2: ( ) [ ]
if (ll <= lr && lr < rl && rl <= rr) {
return NarrowBoundsCase2(cc, ranges);
}
// case 3: ( [ ] )
if (ll <= rl && rl <= rr && rr <= lr) {
return NarrowBoundsCase3(cc, ranges);
}
// case 4: [ ( ] )
if (rl <= ll && ll <= rr && rr <= lr) {
return NarrowBoundsCase4(cc, ranges);
}
// case 5: [ ] ( )
if (rl <= rr && rr < ll && ll <= lr) {
return NarrowBoundsCase5(cc, ranges);
}
// case 6: [ ( ) ]
if (rl <= ll && ll <= lr && lr <= rr) {
return NarrowBoundsCase6(cc, ranges);
}
return ranges;
}
BoundsRange BoundsRangeInfo::FindBoundsRange(const BasicBlock *block, Inst *inst) const
{
ASSERT(block != nullptr && inst != nullptr);
ASSERT(!IsFloatType(inst->GetType()));
ASSERT(inst->GetType() == DataType::REFERENCE || DataType::GetCommonType(inst->GetType()) == DataType::INT64);
if (inst->GetOpcode() == Opcode::NullPtr) {
ASSERT(inst->GetType() == DataType::REFERENCE);
return BoundsRange(0);
}
if (IsInstNotNull(inst)) {
ASSERT(inst->GetType() == DataType::REFERENCE);
return BoundsRange(1, BoundsRange::GetMax(DataType::REFERENCE));
}
while (block != nullptr) {
if (bounds_range_info_.find(block) != bounds_range_info_.end() &&
bounds_range_info_.at(block).find(inst) != bounds_range_info_.at(block).end()) {
return bounds_range_info_.at(block).at(inst);
}
block = block->GetDominator();
}
if (inst->IsConst()) {
ASSERT(inst->GetType() == DataType::INT64);
auto val = static_cast<int64_t>(inst->CastToConstant()->GetIntValue());
return BoundsRange(val);
}
// if we know nothing about inst return the complete range of type
return BoundsRange(inst->GetType());
}
void BoundsRangeInfo::SetBoundsRange(const BasicBlock *block, const Inst *inst, BoundsRange range)
{
if (inst->IsConst() && range.GetLenArray() == nullptr) {
return;
}
if (inst->IsConst()) {
auto val = static_cast<int64_t>(static_cast<const ConstantInst *>(inst)->GetIntValue());
range = BoundsRange(val, val, range.GetLenArray());
}
ASSERT(inst->GetType() == DataType::REFERENCE || DataType::GetCommonType(inst->GetType()) == DataType::INT64);
ASSERT(range.GetLeft() >= BoundsRange::GetMin(inst->GetType()));
ASSERT(range.GetRight() <= BoundsRange::GetMax(inst->GetType()));
if (!range.IsMaxRange() || range.GetLenArray() != nullptr) {
if (bounds_range_info_.find(block) == bounds_range_info_.end()) {
auto it1 = bounds_range_info_.emplace(block, aa_.Adapter());
ASSERT(it1.second);
it1.first->second.emplace(inst, range);
} else if (bounds_range_info_.at(block).find(inst) == bounds_range_info_.at(block).end()) {
bounds_range_info_.at(block).emplace(inst, range);
} else {
bounds_range_info_.at(block).at(inst) = range;
}
}
}
BoundsAnalysis::BoundsAnalysis(Graph *graph) : Analysis(graph), bounds_range_info_(graph->GetAllocator()) {}
bool BoundsAnalysis::RunImpl()
{
GetGraph()->RunPass<DominatorsTree>();
GetGraph()->RunPass<LoopAnalyzer>();
VisitGraph();
return true;
}
const ArenaVector<BasicBlock *> &BoundsAnalysis::GetBlocksToVisit() const
{
return GetGraph()->GetBlocksRPO();
}
void BoundsAnalysis::VisitIf([[maybe_unused]] GraphVisitor *v, [[maybe_unused]] Inst *inst)
{
UNREACHABLE();
}
void BoundsAnalysis::VisitIfImm(GraphVisitor *v, Inst *inst)
{
auto if_inst = inst->CastToIfImm();
ASSERT(if_inst->GetOperandsType() == DataType::BOOL);
ASSERT(if_inst->GetCc() == ConditionCode::CC_NE || if_inst->GetCc() == ConditionCode::CC_EQ);
ASSERT(if_inst->GetImm() == 0);
auto input = inst->GetInput(0).GetInst();
if (input->GetOpcode() != Opcode::Compare) {
return;
}
auto compare = input->CastToCompare();
if (compare->GetOperandsType() == DataType::UINT64) {
return;
}
auto op0 = compare->GetInput(0).GetInst();
auto op1 = compare->GetInput(1).GetInst();
if ((DataType::GetCommonType(op0->GetType()) != DataType::INT64 && op0->GetType() != DataType::REFERENCE) ||
(DataType::GetCommonType(op1->GetType()) != DataType::INT64 && op1->GetType() != DataType::REFERENCE)) {
return;
}
auto cc = compare->GetCc();
auto block = inst->GetBasicBlock();
BasicBlock *true_block;
BasicBlock *false_block;
if (if_inst->GetCc() == ConditionCode::CC_NE) {
// Corresponds to Compare result
true_block = block->GetTrueSuccessor();
false_block = block->GetFalseSuccessor();
} else if (if_inst->GetCc() == ConditionCode::CC_EQ) {
// Corresponds to inversion of Compare result
true_block = block->GetFalseSuccessor();
false_block = block->GetTrueSuccessor();
} else {
UNREACHABLE();
}
CalcNewBoundsRangeForCompare(v, block, cc, op0, op1, true_block);
CalcNewBoundsRangeForCompare(v, block, GetInverseConditionCode(cc), op0, op1, false_block);
}
void BoundsAnalysis::VisitPhi(GraphVisitor *v, Inst *inst)
{
if (IsFloatType(inst->GetType()) || inst->GetType() == DataType::REFERENCE || inst->GetType() == DataType::UINT64) {
return;
}
auto bri = static_cast<BoundsAnalysis *>(v)->GetBoundsRangeInfo();
auto phi = inst->CastToPhi();
auto phi_block = inst->GetBasicBlock();
auto phi_type = phi->GetType();
ArenaVector<BoundsRange> ranges(phi_block->GetGraph()->GetLocalAllocator()->Adapter());
for (auto &block : phi_block->GetPredsBlocks()) {
ranges.emplace_back(bri->FindBoundsRange(block, phi->GetPhiInput(block)));
}
bri->SetBoundsRange(phi_block, phi, BoundsRange::Union(ranges).FitInType(phi_type));
}
bool BoundsAnalysis::CheckTriangleCase(const BasicBlock *block, const BasicBlock *tgt_block)
{
auto &preds_blocks = tgt_block->GetPredsBlocks();
auto loop = tgt_block->GetLoop();
auto &back_edges = loop->GetBackEdges();
if (preds_blocks.size() == 1) {
return true;
}
if (!loop->IsRoot() && back_edges.size() == 1 && preds_blocks.size() == 2U) {
if (preds_blocks[0] == block && preds_blocks[1] == back_edges[0]) {
return true;
}
if (preds_blocks[1] == block && preds_blocks[0] == back_edges[0]) {
return true;
}
return false;
}
return false;
}
void BoundsAnalysis::CalcNewBoundsRangeForCompare(GraphVisitor *v, BasicBlock *block, ConditionCode cc, Inst *left,
Inst *right, BasicBlock *tgt_block)
{
auto bri = static_cast<BoundsAnalysis *>(v)->GetBoundsRangeInfo();
auto left_range = bri->FindBoundsRange(block, left);
auto right_range = bri->FindBoundsRange(block, right);
// try to skip triangle:
/* [block]
* | \
* | \
* | [BB]
* | /
* | /
* [tgt_block]
*/
if (CheckTriangleCase(block, tgt_block)) {
auto ranges = BoundsRange::TryNarrowBoundsByCC(cc, {left_range, right_range});
ASSERT(left_range.GetLenArray() == nullptr);
ASSERT(right_range.GetLenArray() == nullptr);
bri->SetBoundsRange(tgt_block, left, ranges.first.FitInType(left->GetType()));
bri->SetBoundsRange(tgt_block, right, ranges.second.FitInType(right->GetType()));
}
}
} // namespace panda::compiler

160
compiler/optimizer/analysis/bounds_analysis.h
View File

@ -1,160 +0,0 @@
/**
* Copyright (c) 2021-2022 Huawei Device Co., Ltd.
* 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.
*/
#ifndef COMPILER_OPTIMIZER_ANALYSIS_BOUNDSRANGE_ANALYSIS_H
#define COMPILER_OPTIMIZER_ANALYSIS_BOUNDSRANGE_ANALYSIS_H
#include "optimizer/ir/graph_visitor.h"
#include "optimizer/ir/datatype.h"
#include "optimizer/ir/inst.h"
#include "optimizer/pass.h"
#include "utils/arena_containers.h"
namespace panda::compiler {
/**
* Represents a range of values that a variable might have.
*
* It is used to represent variables of integral types according to their size
* and sign.
* It is used for REFERENCE type as well but only for reasoning whether a
* variable is NULL or not.
*/
class BoundsRange {
public:
using RangePair = std::pair<BoundsRange, BoundsRange>;
explicit BoundsRange(DataType::Type type = DataType::INT64) : left_(GetMin(type)), right_(GetMax(type)) {};
explicit BoundsRange(int64_t left, int64_t right, const Inst *inst = nullptr,
DataType::Type type = DataType::INT64);
explicit BoundsRange(int64_t val, DataType::Type type = DataType::INT64);
DEFAULT_COPY_SEMANTIC(BoundsRange);
DEFAULT_MOVE_SEMANTIC(BoundsRange);
~BoundsRange() = default;
const Inst *GetLenArray()
{
return len_array_;
}
int64_t GetLeft() const;
int64_t GetRight() const;
BoundsRange FitInType(DataType::Type type) const;
bool IsConst() const;
bool IsMaxRange(DataType::Type type = DataType::INT64) const;
bool IsEqual(const BoundsRange &range) const;
bool IsLess(const BoundsRange &range) const;
bool IsLess(const Inst *inst) const;
bool IsMore(const BoundsRange &range) const;
bool IsMoreOrEqual(const BoundsRange &range) const;
bool IsNotNegative() const;
bool IsNegative() const;
static int64_t GetMin(DataType::Type type);
static int64_t GetMax(DataType::Type type);
static BoundsRange Union(const ArenaVector<BoundsRange> &ranges);
static RangePair NarrowBoundsByNE(RangePair const &ranges);
static RangePair NarrowBoundsCase1(ConditionCode cc, RangePair const &ranges);
static RangePair NarrowBoundsCase2(ConditionCode cc, RangePair const &ranges);
static RangePair NarrowBoundsCase3(ConditionCode cc, RangePair const &ranges);
static RangePair NarrowBoundsCase4(ConditionCode cc, RangePair const &ranges);
static RangePair NarrowBoundsCase5(ConditionCode cc, RangePair const &ranges);
static RangePair NarrowBoundsCase6(ConditionCode cc, RangePair const &ranges);
static RangePair TryNarrowBoundsByCC(ConditionCode cc, RangePair const &ranges);
static constexpr int64_t MAX_RANGE_VALUE = INT64_MAX;
static constexpr int64_t MIN_RANGE_VALUE = INT64_MIN;
private:
int64_t left_ = MIN_RANGE_VALUE;
int64_t right_ = MAX_RANGE_VALUE;
const Inst *len_array_ {nullptr};
};
class BoundsRangeInfo {
public:
explicit BoundsRangeInfo(ArenaAllocator *aa) : aa_(*aa), bounds_range_info_(aa->Adapter()) {}
NO_COPY_SEMANTIC(BoundsRangeInfo);
NO_MOVE_SEMANTIC(BoundsRangeInfo);
~BoundsRangeInfo() = default;
BoundsRange FindBoundsRange(const BasicBlock *block, Inst *inst) const;
void SetBoundsRange(const BasicBlock *block, const Inst *inst, BoundsRange range);
private:
ArenaAllocator &aa_;
ArenaDoubleUnorderedMap<const BasicBlock *, const Inst *, BoundsRange> bounds_range_info_;
};
// NOLINTNEXTLINE(fuchsia-multiple-inheritance)
class BoundsAnalysis : public Analysis, public GraphVisitor {
public:
explicit BoundsAnalysis(Graph *graph);
NO_MOVE_SEMANTIC(BoundsAnalysis);
NO_COPY_SEMANTIC(BoundsAnalysis);
~BoundsAnalysis() override = default;
const ArenaVector<BasicBlock *> &GetBlocksToVisit() const override;
bool RunImpl() override;
const char *GetPassName() const override
{
return "BoundsAnalysis";
}
BoundsRangeInfo *GetBoundsRangeInfo()
{
return &bounds_range_info_;
}
const BoundsRangeInfo *GetBoundsRangeInfo() const
{
return &bounds_range_info_;
}
static void VisitIf([[maybe_unused]] GraphVisitor *v, [[maybe_unused]] Inst *inst);
static void VisitIfImm(GraphVisitor *v, Inst *inst);
static void VisitPhi(GraphVisitor *v, Inst *inst);
#include "optimizer/ir/visitor.inc"
private:
static bool CheckTriangleCase(const BasicBlock *block, const BasicBlock *tgt_block);
static void CalcNewBoundsRangeForCompare(GraphVisitor *v, BasicBlock *block, ConditionCode cc, Inst *left,
Inst *right, BasicBlock *tgt_block);
private:
BoundsRangeInfo bounds_range_info_;
};
} // namespace panda::compiler
#endif // COMPILER_OPTIMIZER_ANALYSIS_BOUNDS_ANALYSIS_H

38
compiler/optimizer/analysis/countable_loop_parser.cpp
View File

@ -1,38 +0,0 @@
/**
* Copyright (c) 2021-2022 Huawei Device Co., Ltd.
* 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 "optimizer/analysis/countable_loop_parser.h"
#include "optimizer/analysis/loop_analyzer.h"
#include "optimizer/ir/basicblock.h"
#include "optimizer/ir/graph.h"
namespace panda::compiler {
/**
* Check if loop is countable
*
* [Loop]
* Phi(init, update)
* ...
* update(phi, 1)
* Compare(Add/Sub, test)
*
* where `update` is Add or Sub instruction
*/
std::optional<CountableLoopInfo> CountableLoopParser::Parse()
{
return std::nullopt;
}
} // namespace panda::compiler

61
compiler/optimizer/analysis/countable_loop_parser.h
View File

@ -1,61 +0,0 @@
/**
* Copyright (c) 2021-2022 Huawei Device Co., Ltd.
* 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.
*/
#ifndef COMPILER_OPTIMIZER_ANALYSIS_COUNTABLE_LOOP_PARSER_H_
#define COMPILER_OPTIMIZER_ANALYSIS_COUNTABLE_LOOP_PARSER_H_
#include "optimizer/ir/inst.h"
namespace panda::compiler {
class Loop;
/**
* Example of code
* ---------------
* for (init(a); if_imm(compare(a,test)); update(a)) {...}
*/
struct CountableLoopInfo {
Inst *if_imm;
Inst *init;
Inst *test;
Inst *update;
Inst *index;
uint64_t const_step;
ConditionCode normalized_cc; // cc between `update` and `test`
};
/**
* Helper class to check if loop is countable and to get its parameters
*/
class CountableLoopParser {
public:
explicit CountableLoopParser(const Loop &loop) : loop_(loop) {}
NO_MOVE_SEMANTIC(CountableLoopParser);
NO_COPY_SEMANTIC(CountableLoopParser);
~CountableLoopParser() = default;
std::optional<CountableLoopInfo> Parse();
private:
Inst *SetIndexAndRetrunConstInst();
private:
const Loop &loop_;
CountableLoopInfo loop_info_ {};
bool is_head_loop_exit_ = false;
};
} // namespace panda::compiler
#endif // COMPILER_OPTIMIZER_ANALYSIS_COUNTABLE_LOOP_PARSER_H_

9
compiler/optimizer/analysis/linear_order.cpp
View File

@ -37,10 +37,6 @@ void LinearOrder::HandleIfBlock(BasicBlock *if_true_block, BasicBlock *next_bloc
if_inst->CastToIfImm()->InverseConditionCode();
} else if (if_inst->GetOpcode() == Opcode::If) {
if_inst->CastToIf()->InverseConditionCode();
} else if (if_inst->GetOpcode() == Opcode::AddOverflow) {
if_inst->CastToAddOverflow()->InverseConditionCode();
} else if (if_inst->GetOpcode() == Opcode::SubOverflow) {
if_inst->CastToSubOverflow()->InverseConditionCode();
} else {
LOG(FATAL, COMPILER) << "Unexpected `If` instruction: " << *if_inst;
}
@ -58,15 +54,10 @@ void LinearOrder::HandlePrevInstruction(BasicBlock *block, BasicBlock *prev_bloc
switch (prev_inst->GetOpcode()) {
case Opcode::IfImm:
case Opcode::If:
case Opcode::AddOverflow:
case Opcode::SubOverflow:
ASSERT(prev_block->GetSuccsBlocks().size() == MAX_SUCCS_NUM);
HandleIfBlock(prev_block, block);
break;
case Opcode::Throw:
break;
default:
ASSERT(prev_block->GetSuccsBlocks().size() == 1 || prev_block->IsTryBegin() || prev_block->IsTryEnd());
if (block != prev_block->GetSuccessor(0) && !prev_block->GetLastInst()->IsControlFlow()) {

152
compiler/optimizer/analysis/live_registers.cpp
View File

@ -1,152 +0,0 @@
/**
* Copyright (c) 2021-2022 Huawei Device Co., Ltd.
* 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 <algorithm>
#include "liveness_analyzer.h"
#include "live_registers.h"
namespace panda::compiler {
namespace {
struct Split {
Split(LifeIntervalsIt p_begin, LifeIntervalsIt p_end, LifeNumber p_min, LifeNumber p_max,
LifeIntervalsTreeNode *p_parent)
: begin(p_begin), end(p_end), min(p_min), max(p_max), parent(p_parent)
{
ASSERT(p_begin < p_end);
ASSERT(p_min <= p_max);
}
LifeIntervalsIt begin; // NOLINT(misc-non-private-member-variables-in-classes)
LifeIntervalsIt end; // NOLINT(misc-non-private-member-variables-in-classes)
LifeNumber min; // NOLINT(misc-non-private-member-variables-in-classes)
LifeNumber max; // NOLINT(misc-non-private-member-variables-in-classes)
LifeIntervalsTreeNode *parent; // NOLINT(misc-non-private-member-variables-in-classes)
};
// copy intervals with assigned registers and compute min and max life numbers covered by all these intervals
std::pair<LifeNumber, LifeNumber> CopyIntervals(const ArenaVector<LifeIntervals *> &source,
ArenaVector<LifeIntervals *> *destination)
{
LifeNumber min_ln = std::numeric_limits<LifeNumber>::max();
LifeNumber max_ln = 0;
for (auto &interval : source) {
for (auto split = interval; !interval->IsPhysical() && split != nullptr; split = split->GetSibling()) {
if (split->HasReg()) {
min_ln = std::min(min_ln, split->GetBegin());
max_ln = std::max(max_ln, split->GetEnd());
destination->push_back(split);
}
}
}
return std::make_pair(min_ln, max_ln);
}
LifeIntervalsIt PartitionLeftSplit(const LifeIntervalsIt &left, const LifeIntervalsIt &right, LifeNumber midpoint,
LifeNumber *min_ln, LifeNumber *max_ln)
{
LifeNumber left_min_ln = std::numeric_limits<LifeNumber>::max();
LifeNumber left_max_ln = 0;
auto result = std::partition(left, right, [&midpoint, &left_min_ln, &left_max_ln](const auto &em) {
if (em->GetEnd() < midpoint) {
left_min_ln = std::min(left_min_ln, em->GetBegin());
left_max_ln = std::max(left_max_ln, em->GetEnd());
return true;
}
return false;
});
*min_ln = left_min_ln;
*max_ln = left_max_ln;
return result;
}
LifeIntervalsIt PartitionRightSplit(const LifeIntervalsIt &left, const LifeIntervalsIt &right, LifeNumber midpoint,
LifeNumber *min_ln, LifeNumber *max_ln)
{
LifeNumber right_min_ln = std::numeric_limits<LifeNumber>::max();
LifeNumber right_max_ln = 0;
auto result = std::partition(left, right, [&midpoint, &right_min_ln, &right_max_ln](const auto &em) {
if (em->GetBegin() > midpoint) {
right_min_ln = std::min(right_min_ln, em->GetBegin());
right_max_ln = std::max(right_max_ln, em->GetEnd());
return false;
}
return true;
});
*min_ln = right_min_ln;
*max_ln = right_max_ln;
return result;
}
} // namespace
LifeIntervalsTree *LifeIntervalsTree::BuildIntervalsTree(const ArenaVector<LifeIntervals *> &life_intervals,
const Graph *graph)
{
auto alloc = graph->GetAllocator();
auto lalloc = graph->GetLocalAllocator();
auto intervals = alloc->New<ArenaVector<LifeIntervals *>>(alloc->Adapter());
ArenaQueue<const Split *> queue(lalloc->Adapter());
auto ln_range = CopyIntervals(life_intervals, intervals);
if (intervals->empty()) {
return nullptr;
}
queue.push(lalloc->New<Split>(intervals->begin(), intervals->end(), ln_range.first, ln_range.second, nullptr));
LifeIntervalsTreeNode *root {nullptr};
// Split each interval into three parts:
// 1) intervals covering mid point;
// 2) intervals ended before mid point;
// 3) intervals started after mid point.
// Allocate tree node for (1), recursively process (2) and (3).
while (!queue.empty()) {
auto split = queue.front();
queue.pop();
if (split->end - split->begin <= 0) {
continue;
}
auto midpoint = split->min + (split->max - split->min) / 2U;
LifeNumber left_min_ln;
LifeNumber left_max_ln;
auto left_midpoint = PartitionLeftSplit(split->begin, split->end, midpoint, &left_min_ln, &left_max_ln);
LifeNumber right_min_ln;
LifeNumber right_max_ln;
auto right_midpoint = PartitionRightSplit(left_midpoint, split->end, midpoint, &right_min_ln, &right_max_ln);
std::sort(left_midpoint, right_midpoint,
[](LifeIntervals *l, LifeIntervals *r) { return l->GetEnd() > r->GetEnd(); });
auto node = alloc->New<LifeIntervalsTreeNode>(split->min, split->max, left_midpoint, right_midpoint);
if (split->parent == nullptr) {
root = node;
} else if (split->parent->GetMidpoint() > midpoint) {
split->parent->SetLeft(node);
} else {
split->parent->SetRight(node);
}
if (split->begin < left_midpoint) {
queue.push(lalloc->New<Split>(split->begin, left_midpoint, left_min_ln, left_max_ln, node));
}
if (right_midpoint < split->end) {
queue.push(lalloc->New<Split>(right_midpoint, split->end, right_min_ln, right_max_ln, node));
}
}
return alloc->New<LifeIntervalsTree>(root);
}
} // namespace panda::compiler

188
compiler/optimizer/analysis/live_registers.h
View File

@ -1,188 +0,0 @@
/**
* Copyright (c) 2021-2022 Huawei Device Co., Ltd.
* 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.
*/
#ifndef COMPILER_OPTIMIZER_ANALYSIS_LIVE_REGISTERS_H
#define COMPILER_OPTIMIZER_ANALYSIS_LIVE_REGISTERS_H
#include <optimizer/ir/inst.h>
#include <optimizer/ir/graph.h>
#include "utils/arena_containers.h"
#include "optimizer/pass.h"
#include "optimizer/analysis/liveness_analyzer.h"
namespace panda::compiler {
using LifeIntervalsIt = ArenaVector<LifeIntervals *>::iterator;
class LifeIntervalsTreeNode {
public:
explicit LifeIntervalsTreeNode(LifeNumber min_value, LifeNumber max_value, LifeIntervalsIt begin,
LifeIntervalsIt end)
: min_value_(min_value), max_value_(max_value), begin_(begin), end_(end)
{
}
DEFAULT_MOVE_SEMANTIC(LifeIntervalsTreeNode);
DEFAULT_COPY_SEMANTIC(LifeIntervalsTreeNode);
~LifeIntervalsTreeNode() = default;
LifeNumber GetMidpoint() const
{
return (min_value_ + max_value_) / 2;
}
LifeNumber GetMinValue() const
{
return min_value_;
}
LifeNumber GetMaxValue() const
{
return max_value_;
}
LifeIntervalsIt GetBegin() const
{
return begin_;
}
LifeIntervalsIt GetEnd() const
{
return end_;
}
LifeIntervalsTreeNode *GetLeft() const
{
return left_;
}
void SetLeft(LifeIntervalsTreeNode *left)
{
ASSERT(left_ == nullptr);
left_ = left;
}
LifeIntervalsTreeNode *GetRight() const
{
return right_;
}
void SetRight(LifeIntervalsTreeNode *right)
{
ASSERT(right_ == nullptr);
right_ = right;
}
private:
LifeNumber min_value_;
LifeNumber max_value_;
LifeIntervalsIt begin_;
LifeIntervalsIt end_;
LifeIntervalsTreeNode *left_ {nullptr};
LifeIntervalsTreeNode *right_ {nullptr};
};
// Simplified intervals tree implementation.
// Each LifeIntervalsTreeNode stores intervals covering the mid point associated with a node, these intervals
// sorted by the life range end in descending order. Every left child stores intervals ended before current node's
// mid point and every right child stores intervals started after current node's mid point.
class LifeIntervalsTree {
public:
static LifeIntervalsTree *BuildIntervalsTree(Graph *graph)
{
ASSERT(graph->IsAnalysisValid<LivenessAnalyzer>());
return LifeIntervalsTree::BuildIntervalsTree(graph->GetAnalysis<LivenessAnalyzer>().GetLifeIntervals(), graph);
}
static LifeIntervalsTree *BuildIntervalsTree(const ArenaVector<LifeIntervals *> &life_intervals,
const Graph *graph);
explicit LifeIntervalsTree(LifeIntervalsTreeNode *root) : root_(root) {};
DEFAULT_MOVE_SEMANTIC(LifeIntervalsTree);
DEFAULT_COPY_SEMANTIC(LifeIntervalsTree);
~LifeIntervalsTree() = default;
template <bool live_inputs = true, typename Func>
void VisitIntervals(LifeNumber ln, Func func, const Inst *skip_inst = nullptr) const
{
for (auto node = root_; node != nullptr; node = ln < node->GetMidpoint() ? node->GetLeft() : node->GetRight()) {
if (ln < node->GetMinValue() || ln > node->GetMaxValue()) {
// current node does not contain intervals covering target life number
continue;
}
for (auto i = node->GetBegin(); i < node->GetEnd(); i++) {
auto interval = *i;
if (interval->GetInst() == skip_inst) {
continue;
}
if (ln > interval->GetEnd()) {
// intervals are ordered by its end in descending order, so we can stop on first interval
// that ends before target life number
break;
}
if (interval->SplitCover<live_inputs>(ln)) {
func(interval);
}
}
}
}
private:
LifeIntervalsTreeNode *root_;
};
// Analysis collecting live intervals with assigned registers and
// allowing to visit those of them which are intersecting with specific instruction.
class LiveRegisters : public Analysis {
public:
explicit LiveRegisters(Graph *graph) : Analysis(graph) {};
NO_MOVE_SEMANTIC(LiveRegisters);
NO_COPY_SEMANTIC(LiveRegisters);
~LiveRegisters() override = default;
bool RunImpl() override
{
inst_life_intervals_tree_ = LifeIntervalsTree::BuildIntervalsTree(GetGraph());
return true;
}
// Visit all live intervals with assigned registers intersecting with specified instruction
// (excluding the interval for that instruction).
template <bool live_inputs = true, typename Func>
void VisitIntervalsWithLiveRegisters(Inst *inst, Func func)
{
ASSERT(GetGraph()->IsAnalysisValid<LivenessAnalyzer>());
if (inst_life_intervals_tree_ == nullptr) {
return;
}
auto &la = GetGraph()->GetAnalysis<LivenessAnalyzer>();
auto li = la.GetInstLifeIntervals(inst);
inst_life_intervals_tree_->VisitIntervals<live_inputs, Func>(li->GetBegin(), func, inst);
}
const char *GetPassName() const override
{
return "Live Registers";
}
private:
LifeIntervalsTree *inst_life_intervals_tree_ {nullptr};
};
} // namespace panda::compiler
#endif // COMPILER_OPTIMIZER_ANALYSIS_LIVE_REGISTERS_H

49
compiler/optimizer/analysis/liveness_analyzer.cpp
View File

@ -32,9 +32,7 @@ LivenessAnalyzer::LivenessAnalyzer(Graph *graph)
block_live_sets_(graph->GetLocalAllocator()->Adapter()),
pending_catch_phi_inputs_(graph->GetAllocator()->Adapter()),
physical_general_intervals_(graph->GetAllocator()->Adapter()),
physical_vector_intervals_(graph->GetAllocator()->Adapter()),
use_table_(graph->GetAllocator()),
has_safepoint_during_call_(graph->GetRuntime()->HasSafepointDuringCall())
physical_vector_intervals_(graph->GetAllocator()->Adapter())
{
}
@ -224,12 +222,6 @@ void LivenessAnalyzer::BuildInstLifeNumbers()
for (auto inst : block->Insts()) {
inst->SetLinearNumber(linear_number++);
CreateLifeIntervals(inst);
if (IsPseudoUserOfMultiOutput(inst)) {
// Should be the same life number as pseudo-user, since actually they have the same definition
SetInstLifeNumber(inst, life_number);
GetInstLifeIntervals(inst)->AddUsePosition(life_number);
continue;
}
life_number += LIFE_NUMBER_GAP;
SetInstLifeNumber(inst, life_number);
insts_by_life_number_.push_back(inst);
@ -324,9 +316,6 @@ void LivenessAnalyzer::ProcessBlockLiveInstructions(BasicBlock *block, InstLiveS
interval->StartFrom(inst_life_number);
AdjustCatchPhiInputsLifetime(inst);
} else {
if (inst->GetOpcode() == Opcode::LiveOut) {
interval->AppendRange({inst_life_number, GetBlockLiveRange(GetGraph()->GetEndBlock()).GetBegin()});
}
auto current_live_range = LiveRange {GetBlockLiveRange(block).GetBegin(), inst_life_number};
AdjustInputsLifetime(inst, current_live_range, live_set);
}
@ -354,38 +343,12 @@ void LivenessAnalyzer::ProcessBlockLiveInstructions(BasicBlock *block, InstLiveS
/* static */
LiveRange LivenessAnalyzer::GetPropagatedLiveRange(Inst *inst, LiveRange live_range)
{
/*
* Implicit null check encoded as no-op and if the reference to check is null
* then SIGSEGV will be raised at the first (closest) user. Regmap generated for
* NullCheck's SaveState should be valid at that user so we need to extend
* life intervals of SaveState's inputs until NullCheck user.
*/
if (inst->IsNullCheck() && !inst->GetUsers().Empty() && inst->CastToNullCheck()->IsImplicit()) {
auto extend_until = std::numeric_limits<LifeNumber>::max();
for (auto &user : inst->GetUsers()) {
auto li = GetInstLifeIntervals(user.GetInst());
ASSERT(li != nullptr);
extend_until = std::min<LifeNumber>(extend_until, li->GetBegin() + 1);
}
live_range.SetEnd(extend_until);
return live_range;
}
/*
* We need to propagate liveness for instruction with CallRuntime to save registers before call;
* Otherwise, we will not be able to restore the value of the virtual registers
*/
if (inst->IsPropagateLiveness()) {
live_range.SetEnd(live_range.GetEnd() + 1);
} else if (inst->GetOpcode() == Opcode::ReturnInlined && inst->CastToReturnInlined()->IsExtendedLiveness()) {
/*
* [ReturnInlined]
* [ReturnInlined]
* ...
* [Deoptimize/Throw]
*
* In this case we propagate ReturnInlined inputs liveness up to the end of basic block
*/
live_range.SetEnd(GetBlockLiveRange(inst->GetBasicBlock()).GetEnd());
}
return live_range;
}
@ -428,13 +391,9 @@ void LivenessAnalyzer::AdjustInputsLifetime(Inst *inst, LiveRange live_range, In
* Increase ref-input liveness in the 'no-async-jit' mode, since GC can be triggered and delete ref during callee-method
* compilation
*/
void LivenessAnalyzer::SetInputRange(const Inst *inst, const Inst *input, LiveRange live_range) const
void LivenessAnalyzer::SetInputRange([[maybe_unused]] const Inst *inst, const Inst *input, LiveRange live_range) const
{
if (has_safepoint_during_call_ && inst->IsCall() && DataType::IsReference(input->GetType())) {
GetInstLifeIntervals(input)->AppendRange(live_range.GetBegin(), live_range.GetEnd() + 1U);
} else {
GetInstLifeIntervals(input)->AppendRange(live_range);
}
GetInstLifeIntervals(input)->AppendRange(live_range);
}
/*
@ -615,7 +574,7 @@ void LivenessAnalyzer::BlockReg(Register reg, LifeNumber block_from)
bool LivenessAnalyzer::IsCallBlockingRegisters(Inst *inst) const
{
if (inst->IsCall() && !static_cast<CallInst *>(inst)->IsInlined()) {
if (inst->IsCall()) {
return true;
}
if (inst->IsIntrinsic() && inst->CastToIntrinsic()->IsNativeCall()) {

12
compiler/optimizer/analysis/liveness_analyzer.h
View File

@ -17,7 +17,6 @@
#define COMPILER_OPTIMIZER_ANALYSIS_LIVENESS_ANALIZER_H
#include "utils/arena_containers.h"
#include "optimizer/analysis/liveness_use_table.h"
#include "optimizer/ir/constants.h"
#include "optimizer/ir/inst.h"
#include "optimizer/ir/marker.h"
@ -331,9 +330,6 @@ public:
bool NoDest() const
{
if (IsPseudoUserOfMultiOutput(inst_)) {
return false;
}
return inst_->NoDest();
}
@ -543,11 +539,6 @@ public:
void DumpLifeIntervals(std::ostream &out = std::cout) const;
void DumpLocationsUsage(std::ostream &out = std::cout) const;
const UseTable &GetUseTable() const
{
return use_table_;
}
private:
ArenaAllocator *GetAllocator()
{
@ -601,9 +592,6 @@ private:
ArenaMultiMap<Inst *, Inst *> pending_catch_phi_inputs_;
ArenaVector<LifeIntervals *> physical_general_intervals_;
ArenaVector<LifeIntervals *> physical_vector_intervals_;
UseTable use_table_;
bool has_safepoint_during_call_;
Marker marker_ {UNDEF_MARKER};
};
} // namespace panda::compiler

64
compiler/optimizer/analysis/liveness_use_table.cpp
View File

@ -1,64 +0,0 @@
/*
* Copyright (c) 2022 Huawei Device Co., Ltd.
* 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 "optimizer/analysis/liveness_use_table.h"
namespace panda::compiler {
UseTable::UseTable(ArenaAllocator *allocator) : table_(allocator->Adapter()), allocator_(allocator) {}
void UseTable::AddUseOnFixedLocation(const Inst *inst, Location location, LifeNumber ln)
{
auto res = table_.try_emplace(inst, allocator_->Adapter());
auto &uses = res.first->second;
ASSERT(location.IsRegisterValid());
uses[ln] = location.GetValue();
}
bool UseTable::HasUseOnFixedLocation(const Inst *inst, LifeNumber ln) const
{
auto it = table_.find(inst);
if (it == table_.end()) {
return false;
}
const auto &uses = it->second;
return uses.count(ln) > 0;
}
Register UseTable::GetNextUseOnFixedLocation(const Inst *inst, LifeNumber ln) const
{
auto it = table_.find(inst);
if (it == table_.end()) {
return INVALID_REG;
}
const auto &uses = it->second;
auto uses_it = uses.lower_bound(ln);
return uses_it == uses.end() ? INVALID_REG : uses_it->second;
}
void UseTable::Dump(std::ostream &out, Arch arch) const
{
out << "UseTable" << std::endl;
for (auto [inst, uses] : table_) {
out << "Inst v" << inst->GetId() << ": ";
auto sep = "";
for (auto [ln, r] : uses) {
out << sep << "{" << std::to_string(ln) << ", " << Location::MakeRegister(r, inst->GetType()).ToString(arch)
<< "}";
sep = ", ";
}
out << std::endl;
}
}
} // namespace panda::compiler

44
compiler/optimizer/analysis/liveness_use_table.h
View File

@ -1,44 +0,0 @@
/*
* Copyright (c) 2022 Huawei Device Co., Ltd.
* 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.
*/
#ifndef COMPILER_OPTIMIZER_ANALYSIS_USE_TABLE_H
#define COMPILER_OPTIMIZER_ANALYSIS_USE_TABLE_H
#include "utils/arena_containers.h"
#include "optimizer/ir/inst.h"
namespace panda::compiler {
using FixedUses = ArenaMap<LifeNumber, Register>;
/**
* For each added instruction holds its uses on the fixed locations
*/
class UseTable {
public:
explicit UseTable(ArenaAllocator *allocator);
void AddUseOnFixedLocation(const Inst *inst, Location location, LifeNumber ln);
bool HasUseOnFixedLocation(const Inst *inst, LifeNumber ln) const;
Register GetNextUseOnFixedLocation(const Inst *inst, LifeNumber ln) const;
void Dump(std::ostream &out, Arch arch) const;
private:
ArenaUnorderedMap<const Inst *, FixedUses> table_;
ArenaAllocator *allocator_;
};
} // namespace panda::compiler
#endif // COMPILER_OPTIMIZER_ANALYSIS_USE_TABLE_H

1
compiler/optimizer/analysis/loop_analyzer.h
View File

@ -17,7 +17,6 @@
#include "optimizer/ir/inst.h"
#include "optimizer/pass.h"
#include "optimizer/analysis/countable_loop_parser.h"
namespace panda::compiler {
class BasicBlock;

63
compiler/optimizer/analysis/object_type_propagation.cpp
View File

@ -1,63 +0,0 @@
/**
* Copyright (c) 2021-2022 Huawei Device Co., Ltd.
* 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 "object_type_propagation.h"
#include "optimizer/ir/basicblock.h"
#include "optimizer/ir/inst.h"
namespace panda::compiler {
bool ObjectTypePropagation::RunImpl()
{
VisitGraph();
return true;
}
void ObjectTypePropagation::VisitNewObject(GraphVisitor *v, Inst *i)
{
auto self = static_cast<ObjectTypePropagation *>(v);
auto inst = i->CastToNewObject();
auto klass = self->GetGraph()->GetRuntime()->GetClass(inst->GetMethod(), inst->GetTypeId());
if (klass != nullptr) {
inst->SetObjectTypeInfo(ObjectTypeInfo(static_cast<ObjectTypeInfo::ClassType>(klass)));
}
}
void ObjectTypePropagation::VisitNewArray(GraphVisitor *v, Inst *i)
{
auto self = static_cast<ObjectTypePropagation *>(v);
auto inst = i->CastToNewArray();
auto klass = self->GetGraph()->GetRuntime()->GetClass(inst->GetMethod(), inst->GetTypeId());
if (klass != nullptr) {
inst->SetObjectTypeInfo(ObjectTypeInfo(static_cast<ObjectTypeInfo::ClassType>(klass)));
}
}
void ObjectTypePropagation::VisitLoadString(GraphVisitor *v, Inst *i)
{
auto self = static_cast<ObjectTypePropagation *>(v);
auto inst = i->CastToLoadString();
auto klass = self->GetGraph()->GetRuntime()->GetStringClass(inst->GetMethod());
if (klass != nullptr) {
inst->SetObjectTypeInfo(ObjectTypeInfo(static_cast<ObjectTypeInfo::ClassType>(klass)));
}
}
void ObjectTypePropagation::VisitLoadArray([[maybe_unused]] GraphVisitor *v, [[maybe_unused]] Inst *i)
{
// LoadArray should be processed more carefully, because it may contain object of the derived class with own method
// implementation. We need to check all array stores and method calls between NewArray and LoadArray.
// TODO(mshertennikov): Support it.
}
} // namespace panda::compiler

54
compiler/optimizer/analysis/object_type_propagation.h
View File

@ -1,54 +0,0 @@
/**
* Copyright (c) 2021-2022 Huawei Device Co., Ltd.
* 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.
*/
#ifndef PANDA_OBJECT_TYPE_PROPAGATION_H
#define PANDA_OBJECT_TYPE_PROPAGATION_H
#include "optimizer/pass.h"
#include "optimizer/ir/graph.h"
#include "optimizer/ir/graph_visitor.h"
namespace panda::compiler {
// NOLINTNEXTLINE(fuchsia-multiple-inheritance)
class ObjectTypePropagation final : public Analysis, public GraphVisitor {
public:
explicit ObjectTypePropagation(Graph *graph) : Analysis(graph) {}
NO_MOVE_SEMANTIC(ObjectTypePropagation);
NO_COPY_SEMANTIC(ObjectTypePropagation);
~ObjectTypePropagation() override = default;
const ArenaVector<BasicBlock *> &GetBlocksToVisit() const override
{
return GetGraph()->GetBlocksRPO();
}
const char *GetPassName() const override
{
return "ObjectTypePropagation";
}
bool RunImpl() override;
protected:
static void VisitNewObject(GraphVisitor *v, Inst *inst);
static void VisitNewArray(GraphVisitor *v, Inst *inst);
static void VisitLoadArray(GraphVisitor *v, Inst *inst);
static void VisitLoadString(GraphVisitor *v, Inst *inst);
#include "optimizer/ir/visitor.inc"
};
} // namespace panda::compiler
#endif // PANDA_OBJECT_TYPE_PROPAGATION_H

73
compiler/optimizer/analysis/types_analysis.cpp
View File

@ -1,73 +0,0 @@
/**
* Copyright (c) 2021-2022 Huawei Device Co., Ltd.
* 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 "types_analysis.h"
#include "optimizer/ir/inst.h"
namespace panda::compiler {
bool TypesAnalysis::RunImpl()
{
marker_ = GetGraph()->NewMarker();
VisitGraph();
GetGraph()->EraseMarker(marker_);
return true;
}
void TypesAnalysis::MarkedPhiRec(PhiInst *phi, AnyBaseType type)
{
if (phi->SetMarker(marker_)) {
auto phi_type = phi->GetAnyType();
// Phi has 2 inputs or users with different types
if (phi_type != type) {
phi->SetAssumedAnyType(AnyBaseType::UNDEFINED_TYPE);
return;
}
return;
}
phi->SetAssumedAnyType(type);
for (auto &user : phi->GetUsers()) {
auto user_inst = user.GetInst();
if (user_inst->GetOpcode() == Opcode::Phi) {
MarkedPhiRec(user_inst->CastToPhi(), type);
}
}
}
void TypesAnalysis::VisitCastValueToAnyType(GraphVisitor *v, Inst *inst)
{
auto self = static_cast<TypesAnalysis *>(v);
auto type = inst->CastToCastValueToAnyType()->GetAnyType();
ASSERT(type != AnyBaseType::UNDEFINED_TYPE);
for (auto &user : inst->GetUsers()) {
auto user_inst = user.GetInst();
if (user_inst->GetOpcode() == Opcode::Phi) {
self->MarkedPhiRec(user_inst->CastToPhi(), type);
}
}
}
void TypesAnalysis::VisitAnyTypeCheck(GraphVisitor *v, Inst *inst)
{
auto self = static_cast<TypesAnalysis *>(v);
auto type = inst->CastToAnyTypeCheck()->GetAnyType();
if (type == AnyBaseType::UNDEFINED_TYPE) {
return;
}
auto input = inst->GetInput(0).GetInst();
if (input->GetOpcode() == Opcode::Phi) {
self->MarkedPhiRec(input->CastToPhi(), type);
}
}
} // namespace panda::compiler

62
compiler/optimizer/analysis/types_analysis.h
View File

@ -1,62 +0,0 @@
/**
* Copyright (c) 2021-2022 Huawei Device Co., Ltd.
* 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.
*/
#ifndef COMPILER_OPTIMIZER_ANALYSIS_TYPES_ANALISYS_H
#define COMPILER_OPTIMIZER_ANALYSIS_TYPES_ANALISYS_H
#include "optimizer/pass.h"
#include "optimizer/ir/graph.h"
#include "optimizer/ir/graph_visitor.h"
namespace panda::compiler {
/**
* TypesAnalysis sets assumed types(dynamic type) to phi instruction.
* If a PHI is user of a CastValueToAnyType, the type of CastValueToAnyType is set as assumed type of the PHI
* If a PHI is input of AnyTypeCheck, the type of CastValueToAnyType is set as assumed type of the PHI
* If a PHI has assumed type and we try to set another type, we change assumed type to undefine.
* If an assumed type is set for PHI, then we try to assign the type for all PHI that are inputs of the current PHI
*/
// NOLINTNEXTLINE(fuchsia-multiple-inheritance)
class TypesAnalysis final : public Analysis, public GraphVisitor {
public:
explicit TypesAnalysis(Graph *graph) : Analysis(graph) {}
NO_MOVE_SEMANTIC(TypesAnalysis);
NO_COPY_SEMANTIC(TypesAnalysis);
~TypesAnalysis() override = default;
const ArenaVector<BasicBlock *> &GetBlocksToVisit() const override
{
return GetGraph()->GetBlocksRPO();
}
const char *GetPassName() const override
{
return "TypesAnalysis";
}
bool RunImpl() override;
protected:
static void VisitCastValueToAnyType(GraphVisitor *v, Inst *inst);
static void VisitAnyTypeCheck(GraphVisitor *v, Inst *inst);
#include "optimizer/ir/visitor.inc"
private:
void MarkedPhiRec(PhiInst *phi, AnyBaseType type);
Marker marker_ {UNDEF_MARKER};
};
} // namespace panda::compiler
#endif // COMPILER_OPTIMIZER_ANALYSIS_TYPES_ANALISYS_H

213
compiler/optimizer/code_generator/BUILD.gn
View File

@ -1,213 +0,0 @@
# Copyright (c) 2021-2022 Huawei Device Co., Ltd.
# 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.
import("//arkcompiler/runtime_core/ark_config.gni")
config("arkencoder_config") {
include_dirs = [ "$ark_root/compiler/optimizer/code_generator" ]
cflags_cc = [ "-DPANDA_BUILD" ]
if (ark_standalone_build) {
ldflags = [ "-lrt" ]
}
}
ohos_shared_library("libarkencoder") {
sources = [
"target/asm_printer.cpp",
"target/target.cpp",
]
deps = []
if (current_cpu == "x86") {
sources += [
"target/x86/callconv.cpp",
"target/x86/encode.cpp",
"target/x86/regfile.cpp",
]
include_dirs = [ "$ark_root/compiler/optimizer/code_generator/target/x86" ]
} else if (current_cpu == "amd64" || current_cpu == "x64" ||
current_cpu == "x86_64") {
if (!is_mingw) {
sources += [
"target/aarch64/callconv.cpp",
"target/aarch64/encode.cpp",
"target/aarch64/regfile.cpp",
]
cflags_cc = [
"-DUSE_VIXL_ARM64",
"-DVIXL_CODE_BUFFER_MMAP",
]
include_dirs =
[ "$ark_root/compiler/optimizer/code_generator/target/aarch64" ]
if (ark_enable_compiler_x64) {
sources += [
"target/amd64/callconv.cpp",
"target/amd64/encode.cpp",
"target/amd64/regfile.cpp",
]
include_dirs += [
"$ark_third_party_root/asmjit/src",
"$ark_third_party_root/zydis/include",
"$ark_third_party_root/zydis/zycore/include",
]
deps += [
"$ark_third_party_root/asmjit:libasmjit_frontend_static",
"$ark_third_party_root/zydis:libZydis",
]
}
}
} else if (current_cpu == "arm") {
sources += [
"target/aarch32/callconv.cpp",
"target/aarch32/encode.cpp",
"target/aarch32/regfile.cpp",
]
cflags_cc = [
"-DUSE_VIXL_ARM32",
"-DVIXL_CODE_BUFFER_MMAP",
]
include_dirs =
[ "$ark_root/compiler/optimizer/code_generator/target/aarch32" ]
} else if (current_cpu == "arm64") {
sources += [
"target/aarch64/callconv.cpp",
"target/aarch64/encode.cpp",
"target/aarch64/regfile.cpp",
]
cflags_cc = [
"-DUSE_VIXL_ARM64",
"-DVIXL_CODE_BUFFER_MMAP",
]
include_dirs =
[ "$ark_root/compiler/optimizer/code_generator/target/aarch64" ]
}
configs = [
"$ark_root:ark_config",
"$ark_root/libpandabase:arkbase_public_config",
"$ark_root/compiler:arkcompiler_public_config",
":arkencoder_config",
]
deps += [
"$ark_root/compiler:libarkcompiler_intrinsics_gen_inl_can_encode_builtin_inl",
"$ark_root/compiler:libarkcompiler_intrinsics_gen_inl_generate_operations_intrinsic_graph_inl",
"$ark_root/compiler:libarkcompiler_intrinsics_gen_inl_generate_operations_intrinsic_inst_inl",
"$ark_root/compiler:libarkcompiler_intrinsics_gen_inl_get_intrinsics_inl",
"$ark_root/compiler:libarkcompiler_intrinsics_gen_inl_get_intrinsics_names_inl",
"$ark_root/compiler:libarkcompiler_intrinsics_gen_inl_intrinsic_codegen_test_inl",
"$ark_root/compiler:libarkcompiler_intrinsics_gen_inl_intrinsics_enum_inl",
"$ark_root/compiler:libarkcompiler_intrinsics_gen_inl_intrinsics_ir_build_inl",
"$ark_root/libpandabase:libarkbase",
]
output_extension = "so"
relative_install_dir = "ark"
part_name = "runtime_core"
subsystem_name = "arkcompiler"
}
ohos_static_library("libarkencoder_frontend_static") {
sources = [
"target/asm_printer.cpp",
"target/target.cpp",
]
deps = []
if (current_cpu == "x86") {
sources += [
"target/x86/callconv.cpp",
"target/x86/encode.cpp",
"target/x86/regfile.cpp",
]
include_dirs = [ "$ark_root/compiler/optimizer/code_generator/target/x86" ]
} else if (current_cpu == "amd64" || current_cpu == "x64" ||
current_cpu == "x86_64") {
if (!is_mingw) {
sources += [
"target/aarch64/callconv.cpp",
"target/aarch64/encode.cpp",
"target/aarch64/regfile.cpp",
]
cflags_cc = [
"-DUSE_VIXL_ARM64",
"-DVIXL_CODE_BUFFER_MMAP",
]
include_dirs =
[ "$ark_root/compiler/optimizer/code_generator/target/aarch64" ]
if (ark_enable_compiler_x64) {
sources += [
"target/amd64/callconv.cpp",
"target/amd64/encode.cpp",
"target/amd64/regfile.cpp",
]
include_dirs += [
"$ark_third_party_root/asmjit/src",
"$ark_third_party_root/zydis/include",
"$ark_third_party_root/zydis/zycore/include",
]
deps += [
"$ark_third_party_root/asmjit:libasmjit_frontend_static",
"$ark_third_party_root/zydis:libZydis",
]
}
}
} else if (current_cpu == "arm") {
sources += [
"target/aarch32/callconv.cpp",
"target/aarch32/encode.cpp",
"target/aarch32/regfile.cpp",
]
cflags_cc = [
"-DUSE_VIXL_ARM32",
"-DVIXL_CODE_BUFFER_MMAP",
]
include_dirs =
[ "$ark_root/compiler/optimizer/code_generator/target/aarch32" ]
} else if (current_cpu == "arm64") {
sources += [
"target/aarch64/callconv.cpp",
"target/aarch64/encode.cpp",
"target/aarch64/regfile.cpp",
]
cflags_cc = [
"-DUSE_VIXL_ARM64",
"-DVIXL_CODE_BUFFER_MMAP",
]
include_dirs =
[ "$ark_root/compiler/optimizer/code_generator/target/aarch64" ]
}
configs = [
"$ark_root:ark_config",
"$ark_root/libpandabase:arkbase_public_config",
"$ark_root/compiler:arkcompiler_public_config",
":arkencoder_config",
]
deps += [
"$ark_root/compiler:intrinsics_stub_inl_h",
"$ark_root/compiler:libarkcompiler_intrinsics_gen_inl_can_encode_builtin_inl",
"$ark_root/compiler:libarkcompiler_intrinsics_gen_inl_generate_operations_intrinsic_graph_inl",
"$ark_root/compiler:libarkcompiler_intrinsics_gen_inl_generate_operations_intrinsic_inst_inl",
"$ark_root/compiler:libarkcompiler_intrinsics_gen_inl_get_intrinsics_inl",
"$ark_root/compiler:libarkcompiler_intrinsics_gen_inl_get_intrinsics_names_inl",
"$ark_root/compiler:libarkcompiler_intrinsics_gen_inl_intrinsics_enum_inl",
"$ark_root/compiler:libarkcompiler_intrinsics_gen_inl_intrinsics_ir_build_inl",
"$ark_root/compiler:libarkcompiler_opcodes_h_arch_info_gen_h",
"$ark_root/libpandabase:libarkbase_frontend_static",
]
}

96
compiler/optimizer/code_generator/CMakeLists.txt
View File

@ -1,96 +0,0 @@
# Copyright (c) 2021-2022 Huawei Device Co., Ltd.
# 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.
project(encoder)
include(${PANDA_ROOT}/compiler/cmake/target.cmake)
include_directories(
.
${PANDA_ROOT}
)
set(GENERATED_DIR ${PANDA_BINARY_ROOT}/compiler/generated)
function(append_sources dir)
list(APPEND ENCODE_SOURCES target/${dir}/encode.cpp)
list(APPEND ENCODE_SOURCES target/${dir}/callconv.cpp)
list(APPEND ENCODE_SOURCES target/${dir}/regfile.cpp)
set(ENCODE_SOURCES ${ENCODE_SOURCES} PARENT_SCOPE)
endfunction()
set(ENCODE_SOURCES
target/target.cpp
target/asm_printer.cpp
)
if (PANDA_COMPILER_TARGET_X86_64)
message(STATUS "Encoder: Build amd64 target")
append_sources("amd64")
endif()
if (PANDA_COMPILER_TARGET_X86)
message(STATUS "Encoder: Build x86 target")
append_sources("x86")
endif()
if (PANDA_COMPILER_TARGET_AARCH64)
message(STATUS "Encoder: Build aarch64 target")
append_sources("aarch64")
endif()
if (PANDA_COMPILER_TARGET_AARCH32)
if (NOT(PANDA_TARGET_ARM32_ABI_SOFT OR PANDA_TARGET_ARM32_ABI_SOFTFP OR PANDA_TARGET_ARM32_ABI_HARD))
message(ERROR "Undefined ABI for aarch32 architecture - please set PANDA_TARGET_ARM32_ABI_SOFT or PANDA_TARGET_ARM32_ABI_SOFTFP or PANDA_TARGET_ARM32_ABI_HARD")
endif()
message(STATUS "Encoder: Build aarch32 target")
append_sources("aarch32")
endif()
add_library(arkencoder STATIC ${ENCODE_SOURCES})
set_property(TARGET arkencoder PROPERTY POSITION_INDEPENDENT_CODE ON)
target_compile_options(arkencoder PRIVATE -Wno-shadow)
target_include_directories(arkencoder INTERFACE .)
target_include_directories(arkencoder
PRIVATE ${PANDA_ROOT}/compiler
)
target_link_libraries(arkencoder arkbase)
if (PANDA_COMPILER_TARGET_AARCH64)
target_link_libraries(arkencoder vixl)
add_dependencies(arkencoder vixl)
target_compile_options(arkencoder PUBLIC "-DUSE_VIXL_ARM64" "-DVIXL_CODE_BUFFER_MALLOC")
endif()
if (PANDA_COMPILER_TARGET_AARCH32)
target_link_libraries(arkencoder vixl)
add_dependencies(arkencoder vixl)
target_compile_options(arkencoder PUBLIC "-DUSE_VIXL_ARM32" "-DVIXL_CODE_BUFFER_MALLOC")
endif()
if (PANDA_COMPILER_TARGET_X86)
target_link_libraries(arkencoder asmjit Zydis)
add_dependencies(arkencoder asmjit Zydis)
endif()
if (PANDA_COMPILER_TARGET_X86_64)
target_link_libraries(arkencoder asmjit Zydis)
add_dependencies(arkencoder asmjit Zydis)
endif()
panda_add_to_clang_tidy(TARGET arkencoder)
panda_add_sanitizers(TARGET arkencoder SANITIZERS ${PANDA_SANITIZERS_LIST})
add_dependencies(arkencoder intrinsics_gen_compiler)
add_dependencies(arkencoder isa_gen_compiler)
add_dependencies(arkencoder instructions_gen_compiler)
target_include_directories(arkencoder PUBLIC ${GENERATED_DIR})
add_check_style(.)

265
compiler/optimizer/code_generator/callconv.h
View File

@ -1,265 +0,0 @@
/**
* Copyright (c) 2021-2022 Huawei Device Co., Ltd.
* 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.
*/
#ifndef COMPILER_OPTIMIZER_CODEGEN_CALLCONV_H_
#define COMPILER_OPTIMIZER_CODEGEN_CALLCONV_H_
/*
Codegen Hi-Level calling-convention interface
Also contains branches targets(labels)
Responsible for
Branches and jump-encoding
Labels (points for jump)
Conditional instructions
*/
#include <functional>
#include "encode.h"
#include "compiler/optimizer/ir/datatype.h"
#include "compiler/optimizer/ir/locations.h"
#include "compiler/optimizer/code_generator/frame_info.h"
namespace panda::compiler {
class ParameterInfo {
public:
using SlotID = uint8_t;
ParameterInfo() = default;
virtual ~ParameterInfo() = default;
// Get next native parameter, on condition, what previous list - in vector
// Push data in Reg
// Return register or stack_slot
virtual std::variant<Reg, SlotID> GetNativeParam(const TypeInfo &) = 0;
virtual Location GetNextLocation([[maybe_unused]] DataType::Type type) = 0;
void Reset()
{
current_scalar_number_ = 0;
current_vector_number_ = 0;
current_stack_offset_ = 0;
}
NO_COPY_SEMANTIC(ParameterInfo);
NO_MOVE_SEMANTIC(ParameterInfo);
protected:
uint32_t current_scalar_number_ {0}; // NOLINT(misc-non-private-member-variables-in-classes)
uint32_t current_vector_number_ {0}; // NOLINT(misc-non-private-member-variables-in-classes)
uint8_t current_stack_offset_ {0}; // NOLINT(misc-non-private-member-variables-in-classes)
};
#ifdef PANDA_COMPILER_CFI
struct CfiOffsets {
size_t push_fplr {0};
size_t set_fp {0};
size_t push_callees {0};
size_t pop_callees {0};
size_t pop_fplr {0};
};
struct CfiInfo {
CfiOffsets offsets;
RegMask callee_regs;
VRegMask callee_vregs;
};
// NOLINTNEXTLINE(cppcoreguidelines-macro-usage)
#define SET_CFI_OFFSET(field, value) GetCfiInfo().offsets.field = value
// NOLINTNEXTLINE(cppcoreguidelines-macro-usage)
#define SET_CFI_CALLEE_REGS(value) GetCfiInfo().callee_regs = value
// NOLINTNEXTLINE(cppcoreguidelines-macro-usage)
#define SET_CFI_CALLEE_VREGS(value) GetCfiInfo().callee_vregs = value
#else
// NOLINTNEXTLINE(cppcoreguidelines-macro-usage)
#define SET_CFI_OFFSET(field, value)
// NOLINTNEXTLINE(cppcoreguidelines-macro-usage)
#define SET_CFI_CALLEE_REGS(value)
// NOLINTNEXTLINE(cppcoreguidelines-macro-usage)
#define SET_CFI_CALLEE_VREGS(value)
#endif
/**
* Specifies CallingConvention mode.
*/
class CallConvMode final {
public:
explicit CallConvMode(uint32_t value) : value_(value) {}
DEFAULT_COPY_SEMANTIC(CallConvMode);
DEFAULT_MOVE_SEMANTIC(CallConvMode);
~CallConvMode() = default;
// NOLINTNEXTLINE(cppcoreguidelines-macro-usage)
#define DECLARE_CALLCONV_MODE(name) \
static CallConvMode name(bool set = true) \
{ \
return CallConvMode(Flag##name ::Encode(set)); \
} \
void Set##name(bool v) \
{ \
Flag##name ::Set(v, &value_); \
} \
bool Is##name() const \
{ \
return Flag##name ::Get(value_); \
}
// Panda ABI convention (native - otherwise)
DECLARE_CALLCONV_MODE(Panda);
// Compile for osr (jit - otherwise)
DECLARE_CALLCONV_MODE(Osr);
// The method from dynamic language
DECLARE_CALLCONV_MODE(Dyn);
#undef DECLARE_CALLCONV_MODE
private:
using FlagPanda = BitField<bool, 0, 1>;
using FlagOsr = FlagPanda::NextFlag;
using FlagDyn = FlagOsr::NextFlag;
uint32_t value_ {0};
friend CallConvMode operator|(CallConvMode a, CallConvMode b);
};
inline CallConvMode operator|(CallConvMode a, CallConvMode b)
{
return CallConvMode(a.value_ | b.value_);
}
/**
* CallConv - just holds information about calling convention in current architecture.
*/
class CallingConvention {
public:
virtual ~CallingConvention() = default;
// All possible reasons for call and return
enum Reason {
// Reason for save/restore registers
FUNCTION, // Function inside programm
NATIVE, // native function
PROGRAMM // Enter/exit from programm (UNSUPPORTED)
};
// Implemented in target.cpp
static CallingConvention *Create(ArenaAllocator *arena_allocator, Encoder *enc, RegistersDescription *descr,
Arch arch, bool is_panda_abi = false, bool is_osr = false, bool is_dyn = false,
bool print_asm = false);
public:
CallingConvention(ArenaAllocator *allocator, Encoder *enc, RegistersDescription *descr, CallConvMode mode)
: allocator_(allocator), encoder_(enc), regfile_(descr), mode_(mode)
{
}
ArenaAllocator *GetAllocator() const
{
return allocator_;
}
Encoder *GetEncoder() const
{
return encoder_;
}
void SetEncoder(Encoder *enc)
{
encoder_ = enc;
}
RegistersDescription *GetRegfile() const
{
return regfile_;
}
virtual bool IsValid() const
{
return false;
}
CallConvMode GetMode() const
{
return mode_;
}
bool IsPandaMode() const
{
return mode_.IsPanda();
}
bool IsOsrMode() const
{
return mode_.IsOsr();
}
bool IsDynMode() const
{
return mode_.IsDyn();
}
#ifdef PANDA_COMPILER_CFI
CfiInfo &GetCfiInfo()
{
return cfi_info_;
}
const CfiInfo &GetCfiInfo() const
{
return cfi_info_;
}
static constexpr bool ProvideCFI()
{
return true;
}
#else
static constexpr bool ProvideCFI()
{
return false;
}
#endif
// Prologue/Epilogue interfaces
virtual void GeneratePrologue(const FrameInfo &frame_info) = 0;
virtual void GenerateEpilogue(const FrameInfo &frame_info, std::function<void()> post_job) = 0;
virtual void GenerateNativePrologue(const FrameInfo &frame_info) = 0;
virtual void GenerateNativeEpilogue(const FrameInfo &frame_info, std::function<void()> post_job) = 0;
// Code generation completion interfaces
virtual void *GetCodeEntry() = 0;
virtual uint32_t GetCodeSize() = 0;
// Calculating information about parameters and save regs_offset registers for special needs
virtual ParameterInfo *GetParameterInfo(uint8_t regs_offset) = 0;
NO_COPY_SEMANTIC(CallingConvention);
NO_MOVE_SEMANTIC(CallingConvention);
private:
// Must not use ExecModel!
ArenaAllocator *allocator_ {nullptr};
Encoder *encoder_ {nullptr};
RegistersDescription *regfile_ {nullptr};
#ifdef PANDA_COMPILER_CFI
CfiInfo cfi_info_;
#endif
CallConvMode mode_ {0};
};
} // namespace panda::compiler
#endif // COMPILER_OPTIMIZER_CODEGEN_CALLCONV_H_

5419
compiler/optimizer/code_generator/codegen.cpp
View File

File diff suppressed because it is too large Load Diff

879
compiler/optimizer/code_generator/codegen.h
View File

@ -1,879 +0,0 @@
/**
* Copyright (c) 2021-2022 Huawei Device Co., Ltd.
* 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.
*/
#ifndef COMPILER_OPTIMIZER_CODEGEN_CODEGEN_H_
#define COMPILER_OPTIMIZER_CODEGEN_CODEGEN_H_
/*
Codegen interface for compiler
! Do not use this file in runtime
*/
#include "code_info/code_info_builder.h"
#include "compiler_logger.h"
#include "disassembly.h"
#include "frame_info.h"
#include "optimizer/analysis/live_registers.h"
#include "optimizer/code_generator/callconv.h"
#include "optimizer/code_generator/encode.h"
#include "optimizer/code_generator/registers_description.h"
#include "optimizer/code_generator/slow_path.h"
#include "optimizer/code_generator/spill_fill_encoder.h"
#include "optimizer/code_generator/target_info.h"
#include "optimizer/ir/analysis.h"
#include "optimizer/ir/graph.h"
#include "optimizer/ir/graph_visitor.h"
#include "optimizer/optimizations/regalloc/spill_fills_resolver.h"
#include "optimizer/pass_manager.h"
#include "utils/cframe_layout.h"
namespace panda::compiler {
// Maximum size in bytes
constexpr size_t INST_IN_SLOW_PATH = 64;
class Encoder;
class CodeBuilder;
class OsrEntryStub;
class Codegen : public Optimization {
using EntrypointId = RuntimeInterface::EntrypointId;
public:
explicit Codegen(Graph *graph);
NO_MOVE_SEMANTIC(Codegen);
NO_COPY_SEMANTIC(Codegen);
~Codegen() override = default;
bool RunImpl() override;
const char *GetPassName() const override;
bool AbortIfFailed() const override;
static bool Run(Graph *graph);
ArenaAllocator *GetAllocator() const
{
return allocator_;
}
ArenaAllocator *GetLocalAllocator() const
{
return local_allocator_;
}
FrameInfo *GetFrameInfo() const
{
return frame_info_;
}
void SetFrameInfo(FrameInfo *frame_info)
{
frame_info_ = frame_info;
}
virtual void CreateFrameInfo();
RuntimeInterface *GetRuntime() const
{
return runtime_;
}
RegistersDescription *GetRegfile() const
{
return regfile_;
}
Encoder *GetEncoder() const
{
return enc_;
}
CallingConvention *GetCallingConvention() const
{
return callconv_;
}
GraphVisitor *GetGraphVisitor() const
{
return visitor_;
}
LabelHolder::LabelId GetLabelEntry() const
{
return label_entry_;
}
LabelHolder::LabelId GetLabelExit() const
{
return label_exit_;
}
RuntimeInterface::MethodId GetMethodId()
{
return method_id_;
}
void SetStartCodeOffset(size_t offset)
{
start_code_offset_ = offset;
}
size_t GetStartCodeOffset() const
{
return start_code_offset_;
}
size_t GetLanguageExtensionOffsetFromSpInBytes();
void Convert(ArenaVector<Reg> *regs_usage, const ArenaVector<bool> *mask, TypeInfo type_info);
Reg ConvertRegister(Register ref, DataType::Type type = DataType::Type::INT64);
Imm ConvertImm(uint64_t imm, DataType::Type type);
Imm ConvertImmWithExtend(uint64_t imm, DataType::Type type);
Imm ConvertImm(ConstantInst *const_inst, DataType::Type type);
Condition ConvertCc(ConditionCode cc);
Condition ConvertCcOverflow(ConditionCode cc);
static inline TypeInfo ConvertDataType(DataType::Type type, Arch arch)
{
return TypeInfo::FromDataType(type, arch);
}
Arch GetArch() const
{
return GetTarget().GetArch();
}
Target GetTarget() const
{
return target_;
}
TypeInfo GetPtrRegType() const
{
return target_.GetPtrRegType();
}
CodeInfoBuilder *GetCodeBuilder() const
{
return code_builder_;
}
void CreateStackMap(Inst *inst, Inst *user = nullptr);
void CreateStackMapRec(SaveStateInst *save_state, bool require_vreg_map, Inst *target_site);
void CreateVRegMap(SaveStateInst *save_state, size_t vregs_count, Inst *target_site);
void CreateVreg(const Location &location, Inst *inst, const VirtualRegister &vreg);
void FillVregIndices(SaveStateInst *save_state);
void CreateOsrEntry(SaveStateInst *save_state);
void CreateVRegForRegister(const Location &location, Inst *inst, const VirtualRegister &vreg);
/**
* 'live_inputs' shows that inst's source registers should be added the the mask
*/
template <bool live_inputs = false>
std::pair<RegMask, VRegMask> GetLiveRegisters(Inst *inst)
{
RegMask live_regs;
VRegMask live_fp_regs;
if (!options.IsCompilerSaveOnlyLiveRegisters() || inst == nullptr) {
live_regs.set();
live_fp_regs.set();
return {live_regs, live_fp_regs};
}
// Run LiveRegisters pass only if it is actually required
if (!GetGraph()->IsAnalysisValid<LiveRegisters>()) {
GetGraph()->RunPass<LiveRegisters>();
}
// Add registers from intervals that are live at inst's definition
auto &lr = GetGraph()->GetAnalysis<LiveRegisters>();
lr.VisitIntervalsWithLiveRegisters<live_inputs>(inst, [&live_regs, &live_fp_regs, this](const auto &li) {
auto reg = ConvertRegister(li->GetReg(), li->GetType());
GetEncoder()->SetRegister(&live_regs, &live_fp_regs, reg);
});
// Add live temp registers
live_regs |= GetEncoder()->GetLiveTmpRegMask();
live_fp_regs |= GetEncoder()->GetLiveTmpFpRegMask();
return {live_regs, live_fp_regs};
}
// Limits live register set to a number of registers used to pass parameters to the runtime call:
// 1) these ones are saved/restored by caller
// 2) the remaining ones are saved/restored by the bridge function (aarch only)
void FillOnlyParameters(RegMask *live_regs, uint32_t num_params) const;
template <typename T, typename... Args>
T *CreateSlowPath(Inst *inst, Args &&... args)
{
static_assert(std::is_base_of_v<SlowPathBase, T>);
auto label = GetEncoder()->CreateLabel();
auto slow_path = GetLocalAllocator()->New<T>(label, inst, std::forward<Args>(args)...);
slow_paths_.push_back(slow_path);
return slow_path;
}
void EmitSlowPaths();
void InsertTrace(std::initializer_list<std::variant<Reg, Imm>> params);
void CallIntrinsic(Inst *inst, RuntimeInterface::IntrinsicId id);
// The function is used for calling runtime functions through special bridges.
// !NOTE Don't use the function for calling runtime without bridges(it save only parameters on stack)
void CallRuntime(Inst *inst, EntrypointId id, Reg dst_reg, std::initializer_list<std::variant<Reg, Imm>> params,
RegMask preserved_regs = {});
template <typename... Args>
void CallRuntimeWithMethod(Inst *inst, void *method, EntrypointId eid, Reg dst_reg, Args &&... params)
{
if (GetGraph()->IsAotMode()) {
ScopedTmpReg method_reg(GetEncoder());
LoadMethod(method_reg);
CallRuntime(inst, eid, dst_reg, {method_reg, params...});
} else {
if (Is64BitsArch(GetArch())) {
CallRuntime(inst, eid, dst_reg, {Imm(reinterpret_cast<uint64_t>(method)), params...});
} else {
// uintptr_t causes problems on host cross-jit compilation
CallRuntime(inst, eid, dst_reg, {Imm(down_cast<uint32_t>(method)), params...});
}
}
}
void SaveRegistersForImplicitRuntime(Inst *inst, RegMask *params_mask, RegMask *mask);
void VisitNewArray(Inst *inst);
void LoadClassFromObject(Reg class_reg, Reg obj_reg);
void CreateCall(CallInst *call_inst);
void VisitCallIndirect(CallIndirectInst *inst);
void VisitCall(CallInst *inst);
void CreateUnresolvedVirtualMethodLoad(CallInst *vcall, Reg method);
void CreateVirtualCall(CallInst *call_inst);
void CreateDynamicCall(CallInst *call_inst);
void CreateCallIntrinsic(IntrinsicInst *inst);
void CreateMultiArrayCall(CallInst *call_inst);
void CreateNewObjCall(NewObjectInst *new_obj);
void CreateNewObjCallOld(NewObjectInst *new_obj);
void CreateMonitorCall(MonitorInst *inst);
void CreateMonitorCallOld(MonitorInst *inst);
void CreateCheckCastInterfaceCall(Inst *inst);
void CreateNonDefaultInitClass(ClassInst *init_inst);
void CreatePreWRB(Inst *inst, MemRef mem, bool store_pair = false);
void CreatePostWRB(Inst *inst, MemRef mem, Reg reg1, Reg reg2 = INVALID_REGISTER);
void EncodePostWRB(Inst *inst, MemRef mem, Reg reg1, Reg reg2, bool check_nullptr = true);
void CreatePostInterRegionBarrier(Inst *inst, MemRef mem, Reg reg1, Reg reg2, bool check_nullptr);
void CreatePostInterGenerationalBarrier(MemRef mem);
void CallBarrier(RegMask live_regs, VRegMask live_vregs, EntrypointId id,
const std::initializer_list<std::variant<Reg, Imm>> &params);
void CreateLoadClassFromPLT(Inst *inst, Reg tmp_reg, Reg dst, size_t class_id);
void CreateJumpToClassResolverPltShared(Inst *inst, Reg tmp_reg, RuntimeInterface::EntrypointId id);
void CreateLoadTLABInformation(Reg reg_tlab_start, Reg reg_tlab_size);
void CreateCheckForTLABWithConstSize(Inst *inst, Reg reg_tlab_start, Reg reg_tlab_size, size_t size,
LabelHolder::LabelId label);
void CreateDebugRuntimeCallsForNewObject(Inst *inst, Reg reg_tlab_start, size_t alloc_size, RegMask preserved);
void CreateDebugRuntimeCallsForCreateString(Inst *inst, Reg dst);
void CreateReturn(const Inst *inst);
// The function alignment up the value from alignment_reg using tmp_reg.
void CreateAlignmentValue(Reg alignment_reg, Reg tmp_reg, size_t alignment);
void TryInsertImplicitNullCheck(Inst *inst, size_t prevOffset);
const CFrameLayout &GetFrameLayout() const
{
return frame_layout_;
}
bool RegisterKeepCallArgument(CallInst *call_inst, Reg reg);
void LoadMethod(Reg dst);
void LoadFreeSlot(Reg dst);
void StoreFreeSlot(Reg src);
ssize_t GetStackOffset(Location location)
{
if (location.GetKind() == LocationType::STACK_ARGUMENT) {
return location.GetValue() * GetFrameLayout().GetSlotSize();
}
if (location.GetKind() == LocationType::STACK_PARAMETER) {
return GetFrameLayout().GetFrameSize<CFrameLayout::BYTES>() +
(location.GetValue() * GetFrameLayout().GetSlotSize());
}
ASSERT(location.GetKind() == LocationType::STACK);
return GetFrameLayout().GetSpillOffsetFromSpInBytes(location.GetValue());
}
MemRef GetMemRefForSlot(Location location)
{
ASSERT(location.IsAnyStack());
return MemRef(SpReg(), GetStackOffset(location));
}
Reg SpReg() const
{
return GetTarget().GetStackReg();
}
Reg FpReg() const
{
return GetTarget().GetFrameReg();
}
bool HasLiveCallerSavedRegs(Inst *inst);
void SaveCallerRegisters(RegMask live_regs, VRegMask live_vregs, bool adjust_regs);
void LoadCallerRegisters(RegMask live_regs, VRegMask live_vregs, bool adjust_regs);
// Initialization internal variables
void Initialize();
const Disassembly *GetDisasm() const
{
return &disasm_;
}
Disassembly *GetDisasm()
{
return &disasm_;
}
void AddLiveOut(const BasicBlock *bb, const Register reg)
{
live_outs_[bb].Set(reg);
}
RegMask GetLiveOut(const BasicBlock *bb) const
{
auto it = live_outs_.find(bb);
return it != live_outs_.end() ? it->second : RegMask();
}
Reg ThreadReg() const
{
return Reg(GetThreadReg(GetArch()), GetTarget().GetPtrRegType());
}
static bool InstEncodedWithLibCall(const Inst *inst, Arch arch);
protected:
virtual void GeneratePrologue();
virtual void GenerateEpilogue();
// Main logic steps
bool BeginMethod();
bool VisitGraph();
void EndMethod();
bool CopyToCodeCache();
void DumpCode();
RegMask GetUsedRegs() const
{
return used_regs_;
}
RegMask GetUsedVRegs() const
{
return used_vregs_;
}
void FillCallParams(const std::initializer_list<std::variant<Reg, Imm>> &params);
void EmitJump(const BasicBlock *bb);
bool EmitCallRuntimeCode(Inst *inst, EntrypointId id);
void PrepareAndEmitCallVirtual(CallInst *call_inst);
void IntfInlineCachePass(CallInst *call_inst, Reg method_reg, Reg tmp_reg, Reg obj_reg);
void EmitCallVirtual(Reg method_reg);
void PrepareCallVirtualAot(CallInst *call_inst, Reg method_reg);
void PrepareCallVirtual(CallInst *call_inst, Reg method_reg);
uint32_t GetVtableShift();
void CalculateCardIndex(MemRef mem, ScopedTmpReg *tmp, ScopedTmpReg *tmp1);
void EmitGetUnresolvedCalleeMethod(CallInst *call_inst);
void EmitCreateCallCode(CallInst *call_inst);
void EmitEpilogueForCreateCall(CallInst *call_inst);
void CreateBuiltinIntrinsic(IntrinsicInst *inst);
static constexpr int32_t NUM_OF_SRC_BUILTIN = 6;
static constexpr uint8_t FIRST_OPERAND = 0;
static constexpr uint8_t SECOND_OPERAND = 1;
static constexpr uint8_t THIRD_OPERAND = 2;
static constexpr uint8_t FOURTH_OPERAND = 3;
static constexpr uint8_t FIFTH_OPERAND = 4;
using SRCREGS = std::array<Reg, NUM_OF_SRC_BUILTIN>;
// implementation is generated with compiler/optimizer/templates/intrinsics/intrinsics_codegen.inl.erb
void FillBuiltin(IntrinsicInst *inst, SRCREGS src, Reg dst, RegMask *mask);
static Reg AcquireNonLiveReg(RegMask *mask);
void AddParamRegsInLiveMasks(RegMask *live_regs, VRegMask *live_vregs,
const std::initializer_list<std::variant<Reg, Imm>> &params);
void CreateStubCall(Inst *inst, RuntimeInterface::IntrinsicId intrinsicId, Reg dst,
const std::initializer_list<std::variant<Reg, Imm>> &params);
ScopedTmpReg CalculatePreviousTLABAllocSize(Reg reg, LabelHolder::LabelId label);
friend class IntrinsicCodegenTest;
virtual void IntrinsicSlowPathEntry([[maybe_unused]] IntrinsicInst *inst)
{
GetEncoder()->SetFalseResult();
}
virtual void IntrinsicCallRuntimeSaveAll([[maybe_unused]] IntrinsicInst *inst)
{
GetEncoder()->SetFalseResult();
}
virtual void IntrinsicSaveRegisters([[maybe_unused]] IntrinsicInst *inst)
{
GetEncoder()->SetFalseResult();
}
virtual void IntrinsicRestoreRegisters([[maybe_unused]] IntrinsicInst *inst)
{
GetEncoder()->SetFalseResult();
}
virtual void IntrinsicTailCall([[maybe_unused]] IntrinsicInst *inst)
{
GetEncoder()->SetFalseResult();
}
#include "codegen_language_extensions.h"
#include "intrinsics_codegen.inl.h"
private:
template <typename T>
void EncodeImms(const T &imms)
{
auto param_info = GetCallingConvention()->GetParameterInfo(0);
auto imm_type = DataType::INT32;
for (auto imm : imms) {
auto location = param_info->GetNextLocation(imm_type);
ASSERT(location.IsFixedRegister());
auto dst_reg = ConvertRegister(location.GetValue(), imm_type);
GetEncoder()->EncodeMov(dst_reg, Imm(imm));
}
}
private:
ArenaAllocator *allocator_;
ArenaAllocator *local_allocator_;
// Register description
RegistersDescription *regfile_;
// Encoder implementation
Encoder *enc_;
// Target architecture calling convention model
CallingConvention *callconv_;
// Current execution model implementation
// Visitor for instructions
GraphVisitor *visitor_ {};
CodeInfoBuilder *code_builder_ {nullptr};
ArenaVector<SlowPathBase *> slow_paths_;
ArenaUnorderedMap<RuntimeInterface::EntrypointId, SlowPathShared *> slow_paths_map_;
const CFrameLayout frame_layout_; // NOLINT(readability-identifier-naming)
ArenaVector<OsrEntryStub *> osr_entries_;
RuntimeInterface::MethodId method_id_ {INVALID_ID};
size_t start_code_offset_ {0};
ArenaVector<std::pair<int16_t, int16_t>> vreg_indices_;
RuntimeInterface *runtime_ {nullptr};
LabelHolder::LabelId label_entry_ {};
LabelHolder::LabelId label_exit_ {};
FrameInfo *frame_info_ {nullptr};
const Target target_;
/* Registers that have been allocated by regalloc */
RegMask used_regs_ {0};
RegMask used_vregs_ {0};
/* Map of BasicBlock to live-out regsiters mask. It is needed in epilogue encoding to avoid overwriting of the
* live-out registers */
ArenaUnorderedMap<const BasicBlock *, RegMask> live_outs_;
Disassembly disasm_;
SpillFillsResolver spill_fills_resolver_;
friend class EncodeVisitor;
friend class BaselineCodegen;
void CreateStubCall(RuntimeInterface::IntrinsicId intrinsicId, Reg dst,
const std::initializer_list<std::variant<Reg, Imm>> &params);
}; // Codegen
class EncodeVisitor : public GraphVisitor {
using EntrypointId = RuntimeInterface::EntrypointId;
public:
explicit EncodeVisitor(Codegen *cg) : cg_(cg), arch_(cg->GetArch()) {}
EncodeVisitor() = delete;
const ArenaVector<BasicBlock *> &GetBlocksToVisit() const override
{
return cg_->GetGraph()->GetBlocksRPO();
}
Codegen *GetCodegen() const
{
return cg_;
}
Encoder *GetEncoder()
{
return cg_->GetEncoder();
}
Arch GetArch() const
{
return arch_;
}
CallingConvention *GetCallingConvention()
{
return cg_->GetCallingConvention();
}
RegistersDescription *GetRegfile()
{
return cg_->GetRegfile();
}
bool GetResult()
{
return success_ && cg_->GetEncoder()->GetResult();
}
// For each group of SpillFillData representing spill or fill operations and
// sharing the same source and destination types order by stack slot number in descending order.
static void SortSpillFillData(ArenaVector<SpillFillData> *spill_fills);
// Checks if two spill-fill operations could be coalesced into single operation over pair of arguments.
static bool CanCombineSpillFills(SpillFillData pred, SpillFillData succ, const CFrameLayout &fl,
const Graph *graph);
protected:
// UnaryOperation
static void VisitMov(GraphVisitor *visitor, Inst *inst);
static void VisitNeg(GraphVisitor *visitor, Inst *inst);
static void VisitAbs(GraphVisitor *visitor, Inst *inst);
static void VisitNot(GraphVisitor *visitor, Inst *inst);
static void VisitSqrt(GraphVisitor *visitor, Inst *inst);
// BinaryOperation
static void VisitAdd(GraphVisitor *visitor, Inst *inst);
static void VisitSub(GraphVisitor *visitor, Inst *inst);
static void VisitMul(GraphVisitor *visitor, Inst *inst);
static void VisitShl(GraphVisitor *visitor, Inst *inst);
static void VisitAShr(GraphVisitor *visitor, Inst *inst);
static void VisitAnd(GraphVisitor *visitor, Inst *inst);
static void VisitOr(GraphVisitor *visitor, Inst *inst);
static void VisitXor(GraphVisitor *visitor, Inst *inst);
// Binary Overflow Operation
static void VisitAddOverflow(GraphVisitor *v, Inst *inst);
static void VisitAddOverflowCheck(GraphVisitor *v, Inst *inst);
static void VisitSubOverflow(GraphVisitor *v, Inst *inst);
static void VisitSubOverflowCheck(GraphVisitor *v, Inst *inst);
// NOLINTNEXTLINE(cppcoreguidelines-macro-usage)
#define BinaryImmOperation(opc) static void Visit##opc##I(GraphVisitor *visitor, Inst *inst);
// NOLINTNEXTLINE(cppcoreguidelines-macro-usage)
#define BINARRY_IMM_OPS(DEF) DEF(Add) DEF(Sub) DEF(Shl) DEF(AShr) DEF(And) DEF(Or) DEF(Xor)
BINARRY_IMM_OPS(BinaryImmOperation)
#undef BINARRY_IMM_OPS
#undef BinaryImmOperation
// NOLINTNEXTLINE(cppcoreguidelines-macro-usage)
#define BinarySignUnsignOperation(opc) static void Visit##opc(GraphVisitor *visitor, Inst *inst);
// NOLINTNEXTLINE(cppcoreguidelines-macro-usage)
#define SIGN_UNSIGN_OPS(DEF) DEF(Div) DEF(Mod) DEF(Min) DEF(Max) DEF(Shr)
SIGN_UNSIGN_OPS(BinarySignUnsignOperation)
#undef SIGN_UNSIGN_OPS
#undef BinarySignUnsignOperation
// NOLINTNEXTLINE(cppcoreguidelines-macro-usage)
#define BinaryShiftedRegisterOperationDef(opc, ignored) static void Visit##opc##SR(GraphVisitor *visitor, Inst *inst);
ENCODE_INST_WITH_SHIFTED_OPERAND(BinaryShiftedRegisterOperationDef)
#undef BinaryShiftedRegisterOperationDef
static void VisitShrI(GraphVisitor *visitor, Inst *inst);
static void VisitCast(GraphVisitor *visitor, Inst *inst);
static void VisitPhi([[maybe_unused]] GraphVisitor *visitor, [[maybe_unused]] Inst *inst);
static void VisitConstant(GraphVisitor *visitor, Inst *inst);
static void VisitNullPtr(GraphVisitor *visitor, Inst *inst);
// Next visitors use calling convention
static void VisitIndirectJump(GraphVisitor *visitor, Inst *inst);
static void VisitIf(GraphVisitor *visitor, Inst *inst);
static void VisitIfImm(GraphVisitor *visitor, Inst *inst);
static void VisitCompare(GraphVisitor *visitor, Inst *inst);
static void VisitCmp(GraphVisitor *visitor, Inst *inst);
// All next visitors use execution model for implementation
static void VisitReturnVoid(GraphVisitor *visitor, Inst * /* unused */);
static void VisitReturn(GraphVisitor *visitor, Inst *inst);
static void VisitReturnI(GraphVisitor *visitor, Inst *inst);
static void VisitReturnInlined(GraphVisitor *visitor, Inst * /* unused */);
static void VisitNewArray(GraphVisitor *visitor, Inst *inst);
static void VisitLoadConstArray(GraphVisitor *visitor, Inst *inst);
static void VisitFillConstArray(GraphVisitor *visitor, Inst *inst);
static void VisitParameter(GraphVisitor *visitor, Inst *inst);
static void VisitStoreArray(GraphVisitor *visitor, Inst *inst);
static void VisitSpillFill(GraphVisitor *visitor, Inst *inst);
static void VisitSaveState(GraphVisitor *visitor, Inst *inst);
static void VisitSaveStateDeoptimize(GraphVisitor *visitor, Inst *inst);
static void VisitSaveStateOsr(GraphVisitor *visitor, Inst *inst);
static void VisitLoadArray(GraphVisitor *visitor, Inst *inst);
static void VisitLoadCompressedStringChar(GraphVisitor *visitor, Inst *inst);
static void VisitLenArray(GraphVisitor *visitor, Inst *inst);
static void VisitNullCheck(GraphVisitor *visitor, Inst *inst);
static void VisitBoundsCheck(GraphVisitor *visitor, Inst *inst);
static void VisitZeroCheck(GraphVisitor *visitor, Inst *inst);
static void VisitRefTypeCheck(GraphVisitor *visitor, Inst *inst);
static void VisitNegativeCheck(GraphVisitor *visitor, Inst *inst);
static void VisitLoadString(GraphVisitor *visitor, Inst *inst);
static void VisitLoadObject(GraphVisitor *visitor, Inst *inst);
static void VisitUnresolvedLoadObject(GraphVisitor *visitor, Inst *inst);
static void VisitLoad(GraphVisitor *visitor, Inst *inst);
static void VisitStoreObject(GraphVisitor *visitor, Inst *inst);
static void VisitUnresolvedStoreObject(GraphVisitor *visitor, Inst *inst);
static void VisitStore(GraphVisitor *visitor, Inst *inst);
static void VisitLoadStatic(GraphVisitor *visitor, Inst *inst);
static void VisitUnresolvedLoadStatic(GraphVisitor *visitor, Inst *inst);
static void VisitStoreStatic(GraphVisitor *visitor, Inst *inst);
static void VisitUnresolvedStoreStatic(GraphVisitor *visitor, Inst *inst);
static void VisitNewObject(GraphVisitor *visitor, Inst *inst);
static void VisitLoadClass(GraphVisitor *visitor, Inst *inst);
static void VisitLoadAndInitClass(GraphVisitor *visitor, Inst *inst);
static void VisitUnresolvedLoadAndInitClass(GraphVisitor *visitor, Inst *inst);
static void VisitInitClass(GraphVisitor *visitor, Inst *inst);
static void VisitUnresolvedInitClass(GraphVisitor *visitor, Inst *inst);
static void VisitLoadType(GraphVisitor *visitor, Inst *inst);
static void VisitUnresolvedLoadType(GraphVisitor *visitor, Inst *inst);
static void VisitCheckCast(GraphVisitor *visitor, Inst *inst);
static void VisitIsInstance(GraphVisitor *visitor, Inst *inst);
static void VisitMonitor(GraphVisitor *visitor, Inst *inst);
static void VisitIntrinsic(GraphVisitor *visitor, Inst *inst);
static void VisitBuiltin(GraphVisitor *visitor, Inst *inst);
static void VisitBoundsCheckI(GraphVisitor *visitor, Inst *inst);
static void VisitStoreArrayI(GraphVisitor *visitor, Inst *inst);
static void VisitLoadArrayI(GraphVisitor *visitor, Inst *inst);
static void VisitLoadCompressedStringCharI(GraphVisitor *visitor, Inst *inst);
static void VisitLoadI(GraphVisitor *visitor, Inst *inst);
static void VisitStoreI(GraphVisitor *visitor, Inst *inst);
static void VisitMultiArray(GraphVisitor *visitor, Inst *inst);
static void VisitCallStatic(GraphVisitor *visitor, Inst *inst);
static void VisitUnresolvedCallStatic(GraphVisitor *visitor, Inst *inst);
static void VisitCallVirtual(GraphVisitor *visitor, Inst *inst);
static void VisitUnresolvedCallVirtual(GraphVisitor *visitor, Inst *inst);
static void VisitCallDynamic(GraphVisitor *visitor, Inst *inst);
static void VisitSafePoint(GraphVisitor *visitor, Inst *inst);
static void VisitSelect(GraphVisitor *visitor, Inst *inst);
static void VisitSelectImm(GraphVisitor *visitor, Inst *inst);
static void VisitLoadArrayPair(GraphVisitor *visitor, Inst *inst);
static void VisitLoadArrayPairI(GraphVisitor *visitor, Inst *inst);
static void VisitLoadPairPart(GraphVisitor *visitor, Inst *inst);
static void VisitStoreArrayPair(GraphVisitor *visitor, Inst *inst);
static void VisitStoreArrayPairI(GraphVisitor *visitor, Inst *inst);
static void VisitLoadExclusive(GraphVisitor *visitor, Inst *inst);
static void VisitStoreExclusive(GraphVisitor *visitor, Inst *inst);
static void VisitNOP(GraphVisitor *visitor, Inst *inst);
static void VisitThrow(GraphVisitor *visitor, Inst *inst);
static void VisitDeoptimizeIf(GraphVisitor *visitor, Inst *inst);
static void VisitDeoptimizeCompare(GraphVisitor *visitor, Inst *inst);
static void VisitDeoptimizeCompareImm(GraphVisitor *visitor, Inst *inst);
static void VisitDeoptimize(GraphVisitor *visitor, Inst *inst);
static void VisitIsMustDeoptimize(GraphVisitor *visitor, Inst *inst);
static void VisitMAdd(GraphVisitor *visitor, Inst *inst);
static void VisitMSub(GraphVisitor *visitor, Inst *inst);
static void VisitMNeg(GraphVisitor *visitor, Inst *inst);
static void VisitOrNot(GraphVisitor *visitor, Inst *inst);
static void VisitAndNot(GraphVisitor *visitor, Inst *inst);
static void VisitXorNot(GraphVisitor *visitor, Inst *inst);
static void VisitNegSR(GraphVisitor *visitor, Inst *inst);
static void VisitGetInstanceClass(GraphVisitor *visitor, Inst *inst);
static void VisitGetManagedClassObject(GraphVisitor *visito, Inst *inst);
static void VisitClassImmediate(GraphVisitor *visitor, Inst *inst);
static void VisitRegDef(GraphVisitor *visitor, Inst *inst);
static void VisitLiveIn(GraphVisitor *visitor, Inst *inst);
static void VisitLiveOut(GraphVisitor *visitor, Inst *inst);
static void VisitCallIndirect(GraphVisitor *visitor, Inst *inst);
static void VisitCall(GraphVisitor *visitor, Inst *inst);
// Dyn inst.
static void VisitCompareAnyType(GraphVisitor *visitor, Inst *inst);
static void VisitCastAnyTypeValue(GraphVisitor *visitor, Inst *inst);
static void VisitCastValueToAnyType(GraphVisitor *visitor, Inst *inst);
static void VisitAnyTypeCheck(GraphVisitor *visitor, Inst *inst);
void VisitDefault([[maybe_unused]] Inst *inst) override
{
#ifndef NDEBUG
COMPILER_LOG(DEBUG, CODEGEN) << "Can't encode instruction " << GetOpcodeString(inst->GetOpcode())
<< " with type " << DataType::ToString(inst->GetType());
#endif
success_ = false;
}
// Helper functions
static void FillUnresolvedClass(GraphVisitor *visitor, Inst *inst);
static void FillObjectClass(GraphVisitor *visitor, Reg tmp_reg, LabelHolder::LabelId throw_label);
static void FillOtherClass(GraphVisitor *visitor, Inst *inst, Reg tmp_reg, LabelHolder::LabelId throw_label);
static void FillArrayObjectClass(GraphVisitor *visitor, Reg tmp_reg, LabelHolder::LabelId throw_label);
static void FillArrayClass(GraphVisitor *visitor, Inst *inst, Reg tmp_reg, LabelHolder::LabelId throw_label);
static void FillInterfaceClass(GraphVisitor *visitor, Inst *inst);
static void FillLoadClassUnresolved(GraphVisitor *visitor, Inst *inst);
static void FillCheckCast(GraphVisitor *visitor, Inst *inst, Reg src, LabelHolder::LabelId end_label,
compiler::ClassType klass_type);
static void FillIsInstanceUnresolved(GraphVisitor *visitor, Inst *inst);
static void FillIsInstanceCaseObject(GraphVisitor *visitor, Inst *inst, Reg tmp_reg);
static void FillIsInstanceCaseOther(GraphVisitor *visitor, Inst *inst, Reg tmp_reg, LabelHolder::LabelId end_label);
static void FillIsInstanceCaseArrayObject(GraphVisitor *visitor, Inst *inst, Reg tmp_reg,
LabelHolder::LabelId end_label);
static void FillIsInstanceCaseArrayClass(GraphVisitor *visitor, Inst *inst, Reg tmp_reg,
LabelHolder::LabelId end_label);
static void FillIsInstanceCaseInterface(GraphVisitor *visitor, Inst *inst);
static void FillIsInstance(GraphVisitor *visitor, Inst *inst, Reg tmp_reg, LabelHolder::LabelId end_label);
#include "optimizer/ir/visitor.inc"
private:
static void VisitDynamicMethodParameter(GraphVisitor *visitor, Inst *inst);
static void HandleDynParamPassed(const SpillFillData &sf, EncodeVisitor *enc);
static void HandleDynParamNotPassed(const SpillFillData &sf, EncodeVisitor *enc);
static void CastToAny(GraphVisitor *visitor, Inst *inst);
private:
Codegen *cg_;
Arch arch_;
bool success_ {true};
}; // EncodeVisitor
} // namespace panda::compiler
#endif // COMPILER_OPTIMIZER_CODEGEN_CODEGEN_H_

105
compiler/optimizer/code_generator/codegen_native.cpp
View File

@ -1,105 +0,0 @@
/**
* Copyright (c) 2021-2022 Huawei Device Co., Ltd.
* 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 "codegen_native.h"
#include "optimizer/code_generator/method_properties.h"
namespace panda::compiler {
void CodegenNative::CreateFrameInfo()
{
auto &fl = GetFrameLayout();
auto frame = GetGraph()->GetLocalAllocator()->New<FrameInfo>(
FrameInfo::PositionedCallers::Encode(true) | FrameInfo::PositionedCallees::Encode(true) |
FrameInfo::CallersRelativeFp::Encode(false) | FrameInfo::CalleesRelativeFp::Encode(true));
frame->SetFrameSize(fl.GetFrameSize<CFrameLayout::BYTES>());
frame->SetSpillsCount(fl.GetSpillsCount());
frame->SetCallersOffset(
fl.GetOffset<CFrameLayout::SP, CFrameLayout::SLOTS>(fl.GetStackStartSlot() + fl.GetCallerLastSlot(false)));
frame->SetFpCallersOffset(
fl.GetOffset<CFrameLayout::SP, CFrameLayout::SLOTS>(fl.GetStackStartSlot() + fl.GetCallerLastSlot(true)));
frame->SetCalleesOffset(
-fl.GetOffset<CFrameLayout::FP, CFrameLayout::SLOTS>(fl.GetStackStartSlot() + fl.GetCalleeLastSlot(false)));
frame->SetFpCalleesOffset(
-fl.GetOffset<CFrameLayout::FP, CFrameLayout::SLOTS>(fl.GetStackStartSlot() + fl.GetCalleeLastSlot(true)));
ASSERT(!GetGraph()->GetMethodProperties().GetRequireFrameSetup());
// we don't need to setup frame in native mode
frame->SetSetupFrame(false);
// we don't need to save FP and LR registers only for leaf methods
frame->SetSaveFrameAndLinkRegs(!GetGraph()->GetMethodProperties().IsLeaf());
// we never need to save unused registers in native mode
frame->SetSaveUnusedCalleeRegs(false);
// we have to sub/add SP in prologue/epilogue in the following cases:
// - non-leaf method
// - leaf method and there are spills or parameters on stack
frame->SetAdjustSpReg(!GetGraph()->GetMethodProperties().IsLeaf() || GetGraph()->GetStackSlotsCount() != 0 ||
GetGraph()->GetMethodProperties().GetHasParamsOnStack());
SetFrameInfo(frame);
}
void CodegenNative::GeneratePrologue()
{
SCOPED_DISASM_STR(this, "Method Prologue");
GetCallingConvention()->GenerateNativePrologue(*GetFrameInfo());
if (GetGraph()->IsDynamicMethod()) {
GenerateExtensionsForPrologue();
}
#if defined(EVENT_METHOD_ENTER_ENABLED) && EVENT_METHOD_ENTER_ENABLED != 0
if (GetGraph()->IsAotMode()) {
SCOPED_DISASM_STR(this, "LoadMethod for trace");
ScopedTmpReg method_reg(GetEncoder());
LoadMethod(method_reg);
InsertTrace({Imm(static_cast<size_t>(TraceId::METHOD_ENTER)), method_reg,
Imm(static_cast<size_t>(events::MethodEnterKind::COMPILED))});
} else {
InsertTrace({Imm(static_cast<size_t>(TraceId::METHOD_ENTER)),
Imm(reinterpret_cast<size_t>(GetGraph()->GetMethod())),
Imm(static_cast<size_t>(events::MethodEnterKind::COMPILED))});
}
#endif
}
void CodegenNative::GenerateEpilogue()
{
ASSERT(GetGraph()->GetMethodProperties().IsLeaf());
SCOPED_DISASM_STR(this, "Method Epilogue");
if (GetGraph()->IsDynamicMethod()) {
GenerateExtensionsForEpilogue();
}
#if defined(EVENT_METHOD_EXIT_ENABLED) && EVENT_METHOD_EXIT_ENABLED != 0
GetCallingConvention()->GenerateNativeEpilogue(*GetFrameInfo(), [this]() {
if (GetGraph()->IsAotMode()) {
ScopedTmpReg method_reg(GetEncoder());
LoadMethod(method_reg);
InsertTrace({Imm(static_cast<size_t>(TraceId::METHOD_EXIT)), method_reg,
Imm(static_cast<size_t>(events::MethodExitKind::COMPILED))});
} else {
InsertTrace({Imm(static_cast<size_t>(TraceId::METHOD_EXIT)),
Imm(reinterpret_cast<size_t>(GetGraph()->GetMethod())),
Imm(static_cast<size_t>(events::MethodExitKind::COMPILED))});
}
});
#else
GetCallingConvention()->GenerateNativeEpilogue(*GetFrameInfo(), []() {});
#endif
}
} // namespace panda::compiler

45
compiler/optimizer/code_generator/codegen_native.h
View File

@ -1,45 +0,0 @@
/**
* Copyright (c) 2021-2022 Huawei Device Co., Ltd.
* 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.
*/
#ifndef COMPILER_OPTIMIZER_CODEGEN_CODEGEN_NATIVE_H
#define COMPILER_OPTIMIZER_CODEGEN_CODEGEN_NATIVE_H
#include "optimizer/code_generator/codegen.h"
namespace panda::compiler {
/**
* CodegenNative provides support for 'Native' calling convention.
*/
class CodegenNative : public Codegen {
public:
explicit CodegenNative(Graph *graph) : Codegen(graph) {}
NO_MOVE_SEMANTIC(CodegenNative);
NO_COPY_SEMANTIC(CodegenNative);
~CodegenNative() override = default;
const char *GetPassName() const override
{
return "CodegenNative";
}
void CreateFrameInfo() override;
protected:
void GeneratePrologue() override;
void GenerateEpilogue() override;
}; // CodegenNative
} // namespace panda::compiler
#endif // COMPILER_OPTIMIZER_CODEGEN_CODEGEN_NATIVE_H

213
compiler/optimizer/code_generator/disassembly.cpp
View File

@ -1,213 +0,0 @@
/**
* Copyright (c) 2021-2022 Huawei Device Co., Ltd.
* 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 <mutex>
#include "codegen.h"
#include "disassembly.h"
namespace panda::compiler {
// clang-format off
static constexpr std::array INDENT_STRINGS = {
" ",
" ",
" ",
" ",
" ",
" ",
" ",
" ",
" ",
};
// clang-format on
static const char *GetIndent(uint32_t depth)
{
return INDENT_STRINGS[depth];
}
static constexpr auto StreamDeleter = [](std::ostream *stream) {
if (!options.IsCompilerDisasmDumpStdout()) {
delete stream;
}
};
Disassembly::Disassembly(const Codegen *codegen)
: codegen_(codegen),
encoder_(codegen->GetEncoder()),
stream_(nullptr, StreamDeleter),
is_enabled_(options.IsCompilerDisasmDump()),
is_code_enabled_(is_enabled_ && options.IsCompilerDisasmDumpCode())
{
}
void Disassembly::Init()
{
if (!IsEnabled()) {
return;
}
if (options.IsCompilerDisasmDumpStdout()) {
stream_.reset(&std::cout);
} else if (options.IsCompilerDisasmDumpSingleFile()) {
auto stm = new std::ofstream;
if (stm == nullptr) {
UNREACHABLE();
}
static std::once_flag flag;
auto file_name = options.GetCompilerDisasmDumpFileName();
std::call_once(flag, [&file_name]() { std::remove(file_name.c_str()); });
stm->open(file_name, std::ios_base::app);
if (!stm->is_open()) {
LOG(FATAL, COMPILER) << "Cannot open 'disasm.txt'";
}
stream_.reset(stm);
} else {
auto stm = new std::ofstream;
if (stm == nullptr) {
UNREACHABLE();
}
std::stringstream ss;
auto graph = codegen_->GetGraph();
auto exec_num = graph->GetPassManager()->GetExecutionCounter();
ss << "disasm_" << exec_num << '_' << codegen_->GetRuntime()->GetClassNameFromMethod(graph->GetMethod()) << '_'
<< codegen_->GetRuntime()->GetMethodName(graph->GetMethod()) << (graph->IsOsrMode() ? "_osr" : "") << ".txt";
stm->open(ss.str());
if (!stm->is_open()) {
LOG(FATAL, COMPILER) << "Cannot open '" << ss.str() << "'";
}
stream_.reset(stm);
}
}
static void PrintChapter(std::ostream &stream, const char *name)
{
stream << name << ":\n";
}
void Disassembly::IncreaseDepth()
{
if ((depth_ + 1) < INDENT_STRINGS.size()) {
depth_++;
}
}
void Disassembly::PrintMethodEntry(const Codegen *codegen)
{
static constexpr const char *indent = " ";
auto &stream = GetStream();
auto graph = codegen->GetGraph();
stream << "======================================================================\n";
PrintChapter(stream, "METHOD_INFO");
stream << indent << "name: " << codegen->GetRuntime()->GetMethodFullName(graph->GetMethod(), true) << std::endl;
stream << indent << "mode: ";
graph->GetMode().Dump(stream);
stream << std::endl;
stream << indent << "id: " << codegen->GetRuntime()->GetMethodId(graph->GetMethod()) << std::endl;
if (graph->IsAotMode()) {
stream << indent << "code_offset: " << reinterpret_cast<void *>(graph->GetAotData()->GetCodeOffset())
<< std::endl;
}
auto arch = codegen->GetArch();
auto frame = codegen->GetFrameInfo();
stream << indent << "frame_size: " << frame->GetFrameSize() << std::endl;
stream << indent << "spills_count: " << frame->GetSpillsCount() << std::endl;
stream << indent << "Callees: " << (frame->GetCalleesRelativeFp() ? "fp" : "sp") << std::showpos
<< frame->GetCalleesOffset() << std::noshowpos << " (" << GetCalleeRegsCount(arch, false) << ")"
<< std::endl;
stream << indent << "FpCallees: " << (frame->GetCalleesRelativeFp() ? "fp" : "sp") << std::showpos
<< frame->GetFpCalleesOffset() << std::noshowpos << " (" << GetCalleeRegsCount(arch, true) << ")"
<< std::endl;
stream << indent << "Callers: " << (frame->GetCallersRelativeFp() ? "fp" : "sp") << std::showpos
<< frame->GetCallersOffset() << std::noshowpos << " (" << GetCallerRegsCount(arch, false) << ")"
<< std::endl;
stream << indent << "FpCallers: " << (frame->GetCallersRelativeFp() ? "fp" : "sp") << std::showpos
<< frame->GetFpCallersOffset() << std::noshowpos << " (" << GetCallerRegsCount(arch, true) << ")"
<< std::endl;
if (IsCodeEnabled()) {
PrintChapter(stream, "DISASSEMBLY");
}
}
void Disassembly::PrintCodeInfo(const Codegen *codegen)
{
auto &stream = GetStream();
auto graph = codegen->GetGraph();
CodeInfo code_info;
ASSERT(!graph->GetCodeInfoData().empty());
code_info.Decode(graph->GetCodeInfoData());
PrintChapter(stream, "CODE_INFO");
code_info.Dump(stream);
}
void Disassembly::PrintCodeStatistics(const Codegen *codegen)
{
auto &stream = GetStream();
auto graph = codegen->GetGraph();
PrintChapter(stream, "CODE_STATS");
stream << " code_size: " << std::dec << graph->GetData().Size() << std::endl;
}
void Disassembly::FlushDisasm([[maybe_unused]] const Codegen *codegen)
{
auto encoder = GetEncoder();
auto &stream = GetStream();
for (size_t pc = GetPosition(); pc < (encoder->GetCursorOffset());) {
stream << GetIndent(GetDepth());
auto new_pc = encoder->DisasmInstr(stream, pc, 0);
stream << std::endl;
pc = new_pc;
}
SetPosition(encoder->GetCursorOffset());
}
void Disassembly::PrintStackMap(const Codegen *codegen)
{
FlushDisasm(codegen);
auto &stream = GetStream();
stream << GetIndent(GetDepth());
codegen->GetCodeBuilder()->DumpCurrentStackMap(stream);
stream << std::endl;
}
ScopedDisasmPrinter::ScopedDisasmPrinter(Codegen *codegen, const Inst *inst) : disasm_(codegen->GetDisasm())
{
if (disasm_->IsCodeEnabled()) {
disasm_->FlushDisasm(codegen);
disasm_->GetStream() << GetIndent(disasm_->GetDepth()) << "# [inst] " << *inst << std::endl;
disasm_->IncreaseDepth();
}
}
ScopedDisasmPrinter::ScopedDisasmPrinter(Codegen *codegen, const std::string &msg) : disasm_(codegen->GetDisasm())
{
if (disasm_->IsCodeEnabled()) {
disasm_->FlushDisasm(codegen);
disasm_->GetStream() << GetIndent(disasm_->GetDepth()) << "# " << msg << std::endl;
disasm_->IncreaseDepth();
}
}
ScopedDisasmPrinter::~ScopedDisasmPrinter()
{
if (disasm_->IsCodeEnabled()) {
disasm_->FlushDisasm(nullptr);
disasm_->DecreaseDepth();
}
}
} // namespace panda::compiler

130
compiler/optimizer/code_generator/disassembly.h
View File

@ -1,130 +0,0 @@
/**
* Copyright (c) 2021-2022 Huawei Device Co., Ltd.
* 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.
*/
#ifndef PANDA_DISASSEMBLY_H
#define PANDA_DISASSEMBLY_H
#include <fstream>
#include <string>
#include <variant>
#include <memory>
#include "compiler_options.h"
#include "macros.h"
namespace panda::compiler {
class Codegen;
class Encoder;
class Inst;
class StackMap;
class Disassembly {
public:
explicit Disassembly(const Codegen *codegen);
~Disassembly() = default;
NO_COPY_SEMANTIC(Disassembly);
NO_MOVE_SEMANTIC(Disassembly);
void Init();
std::ostream &GetStream()
{
return *stream_;
}
uint32_t GetDepth() const
{
return depth_;
}
uint32_t GetPosition() const
{
return position_;
}
void SetPosition(uint32_t pos)
{
position_ = pos;
}
const Encoder *GetEncoder() const
{
return encoder_;
}
void SetEncoder(const Encoder *encoder)
{
encoder_ = encoder;
}
void IncreaseDepth();
void DecreaseDepth()
{
depth_--;
}
bool IsEnabled() const
{
return is_enabled_;
}
bool IsCodeEnabled() const
{
return is_code_enabled_;
}
void PrintMethodEntry(const Codegen *codegen);
void PrintCodeInfo(const Codegen *codegen);
void PrintCodeStatistics(const Codegen *codegen);
void PrintStackMap(const Codegen *codegen);
private:
void FlushDisasm(const Codegen *codegen);
private:
using StreamDeleterType = void (*)(std::ostream *stream);
const Codegen *codegen_ {nullptr};
const Encoder *encoder_ {nullptr};
std::unique_ptr<std::ostream, StreamDeleterType> stream_;
uint32_t depth_ {0};
uint32_t position_ {0};
bool is_enabled_ {false};
bool is_code_enabled_ {false};
friend class ScopedDisasmPrinter;
};
class ScopedDisasmPrinter {
public:
ScopedDisasmPrinter(Codegen *codegen, const std::string &msg);
ScopedDisasmPrinter(Codegen *codegen, const Inst *inst);
~ScopedDisasmPrinter();
NO_COPY_SEMANTIC(ScopedDisasmPrinter);
NO_MOVE_SEMANTIC(ScopedDisasmPrinter);
private:
Disassembly *disasm_ {nullptr};
};
// NOLINTNEXTLINE(cppcoreguidelines-macro-usage)
#define DISASM_VAR_CONCAT2(a, b) a##b
// NOLINTNEXTLINE(cppcoreguidelines-macro-usage)
#define DISASM_VAR_CONCAT(a, b) DISASM_VAR_CONCAT2(a, b)
// NOLINTNEXTLINE(cppcoreguidelines-macro-usage)
#define SCOPED_DISASM_INST(codegen, inst) ScopedDisasmPrinter DISASM_VAR_CONCAT(disasm_, __LINE__)(codegen, inst)
#ifndef NDEBUG
// NOLINTNEXTLINE(cppcoreguidelines-macro-usage)
#define SCOPED_DISASM_STR(codegen, str) ScopedDisasmPrinter DISASM_VAR_CONCAT(disasm_, __LINE__)(codegen, str)
#else
// NOLINTNEXTLINE(cppcoreguidelines-macro-usage)
#define SCOPED_DISASM_STR(codegen, str) (void)codegen
#endif
} // namespace panda::compiler
#endif // PANDA_DISASSEMBLY_H

1201
compiler/optimizer/code_generator/encode.h
View File

File diff suppressed because it is too large Load Diff

209
compiler/optimizer/code_generator/encoder.md
View File

@ -1,209 +0,0 @@
# Encoder library
## Building
Current realization is not building outside compiler.
## Sub-modules description
There are two major variable-parts - architecture and execution model. Current architecture has aim to make easy switching of this two directions.
### Register file (concrete implementation is architecture dependent)
RegistersDescription - is class for get access for all possible information about registers, which are used in the target architecture: amount of available registers, theirs type and size. Description for special registers and temps. It also may holds low-level information about callee and caller-saved registers. Also it is useful to have conversation from abstract register to arch-special registers(for example for vixlReg).
This file must not use other interfaces.
### Encoder (concrete implementation is architecture dependent)
Encoder - is class for emit main list of instructions - math and memory-usage.
This class uses RegistersDescription.
### Calling Convention (concrete implementation is architecture dependent)
CallingConvention - is class for control calls/jumps, for spill/fill registers and access for native parameters. That's why it contains Labels list for have possibility to know concrete address for special branch-target.
This class uses Encoder and RegistersDescription classes.
Current stack layout must be like next one:
```
----------+----------------------+
Caller | Default calling convention frame:
Frame | Caller saved + parameters, if they
| don't fit in registers
|
----------+-Encoder-Frame-Start--+ For CallingConvention::Begin(FUNCTION):
--------+----------------------+ (CallConv::PushHeader) {
Pre | lr | |
Header | fp | |
| 1-st param | - (method) |
| allignment reg | (push one reg for padding)
--------+----------------------+ } // (CallConv::PushHeader)
--------+----------------------+ (CallConv::PushRegs(GetCalleeSavedR + V)) {
Header | scalar registers | |
callee | + optional allign-reg| |
| vector registers | |
| + optional allign-reg| |
--------+----------------------+ } // (CallConv::PushRegs(GetCalleeSavedR + V))
--------+----------------------+ CallConv::IncrementStack(SIZE_IN_BYTES)
| |
| | Memory, avaliable for user
| |
--------+----------------------+ CallingConvention::Begin(NATIVE):
--------+----------------------+ (CallConv::PushRegs(GetCallerSavedR + V)) {
Header | scalar registers | |
caller | + optional allign-reg| |
| vector registers | |
| + optional allign-reg| |
--------+----------------------+ } // (CallConv::PushRegs(GetCallerSavedR + V))
Param | Here will be |
| parameters, if they | Must be manually filled
| don't fit in regs |
| |
----------+-Encoder-Frame-End------+
----------+------------------------+
Native | Default calling convention frame:
Frame | Callee saved ... e.t.c
```
### Execution model (concrete implementation - must be architecture independent)
ExecModel - is class for control emitting instructions. It must to be able to create default headers for emitted code and implement execution-specific code, e.g interpreter-ExecState access values.
This class uses all above classes.
### Codegen (concrete implementation - must be architecture independent)
Codegen - must not be in Encode-library, not to make dependency from compiler. It also must not to make dependency from special architecture.
This class also must use all of above classes and have possibility to switch between arc or models.
## Operands
### TypeInfo
Class **TypeInfo** contains information about supported types: BOOL, INT8, INT16, INT32, INT64, FLOAT32, FLOAT64
You can get next information about type: size, scalar or vector.
Example:
```
auto type = TypeInfo(FLOAT32); // TypeInfo for float type
ASSERT(type->GetSize() == 32 && type->IsFloat());
```
### Register
Class **Reg** contains number of register(id) and **TypeInfo**.
You can get next information about register: type, id, size, scalar or vector.
Example:
```
auto reg = Reg(0, TypeInfo(INT32)); // scalar word regster
ASSERT(reg->GetId() == 0 && reg->GetType() == TypeInfo(INT32) && reg->GetSize() == 32 && reg->IsScalar());
```
### Immediate
Class **Imm** contains value of the following types: int8_t, int16_t, int32_t, int64_t, float, double
You can get next information about immediate: type, value, size.
Example:
```
double value = 123.456;
auto imm = Imm(value); // double immediate
ASSERT(imm.GetValue<double>() == value && imm->GetType() == TypeInfo(FLOAT64) && imm->GetSize() == 64 &&
!imm->IsScalar());
```
### Memory
Class **MemRef** contains base **Reg**, index **Reg**, scale **Imm** and disp **Imm**.
The memory address is calculated using the following formula: `base + (index << scale) + disp`
If a parameter is not defined, it is considered equal to 0. The base must be defined.
Example:
```
// memory with base register and disp
auto base_reg = Reg(5, TypeInfo(INT64));
auto disp = Imm(static_cast<int642>(16));
auto mem_disp = MemRef(base_reg, disp); // base_reg + disp
ASSERT(mem_disp.HasBase() && !mem_disp.HasIndex() && !mem_disp.HasScale() && mem_disp.HasDisp());
ASSERT(mem_disp.GetBase() == base_reg && mem_disp.GetDisp() == disp);
ASSERT(mem_disp.GetIndex() == INVALID_REGISTER && mem_disp.GetScale() == INVALID_IMM);
// memory with base , index registers and scale
auto base_reg = Reg(5, TypeInfo(INT64));
auto index_reg = Reg(6, TypeInfo(INT64));
auto scale = Imm(static_cast<int32>(3));
auto mem_scale = MemRef(base_reg, index_reg, scale); // base_reg + (index_reg << scale)
ASSERT(mem_scale.HasBase() && mem_scale.HasIndex() && mem_scale.HasScale() && !mem_scale.HasDisp());
ASSERT(mem_scale.GetBase() == base_reg && mem_scale.GetIndex() == index_reg && mem_scale.GetScale() == scale);
ASSERT(mem_scale.GetDisp() == INVALID_IMM);
```
## Code Example (WIP)
```
// Example of usage encode-library
/* == Sequence for create and configure encoder ==*/
// 1. Create encoder for special-architecture
Encoder *enc= buildEncoder(Arch);
// 2. Create execution model - must be target independent
ExecModel* exec = buildExecutionModel(Model::JIT);
// 3. Create Calling convention-model - architecture-dependent
CallingConvention *callconv = buildCallingConvention(exec, Arch);
// 4. Fill Encoder
enc->Configure(exec, callconv);
/*======= Here encoder is fully configured =======*/
// Usage:
// Method start header
callconv->Begin(CallingConvention::Reason::StartMethod);
callconv->CallerStart(CallingConvention::Reason::StartMethod); // (?)
// Use encoding:
// Get register information
auto register_description = enc->GetRegisterDescription();
// Get special registers data
auto sp = register_description->GetSpecialRegister(SpecialReg::stack_pointer);
ASSERT(sp.IsValid());
// Get all un-reserved register list
auto available_regs = register_description->GetRegAllocatable()
// Build memory operand
auto memory = enc->BuildMem(sp, callconv->GetParamOffset(2));
auto tmp1 = available_regs[0];
enc->EncodeLoad(tmp1, memory);
auto tmp2 = available_regs[1];
enc->EncodeMov(tmp2, Imm(0x123));
enc->EncodeAdd(tmp1, tmp2, tmp1);
auto label = callconv->CreateLabel();
// or CreateLabel<type>
callconv->EncodeJump(label, CallingConvention::Reason::Jump);
callconv->BindLabel(label);
enc->EncodeAdd(tmp1, tmp2, tmp1);
enc->EncodeStore(tmp1, memory);
callconv->EncodeBranch(label);
callconv->CallerEnd(CallingConvention::Reason::ExitMethod); // (?)
callconv->EncodeReturn(CallingConvention::Reason::ExitMethod);
// End of method
callconv->End(CallingConvention::Reason::EndMethod);
enc->Finalize();
auto code = enc->GetEncodeData();
code_allocator->Allocate(code.start(), code.size());
///Example from encoder-part
void Aarch64Encoder::EncodeAShr(Reg dst, Reg src0, Reg src1) {
if (dst.GetSize() < MAX_REG_SIZE) {
__ And(VixlReg(src1), (VixlReg(src1), VixlImm(dsr.GetSize() - 1));
}
__ Asr(VixlReg(dst), VixlReg(src0), VixlReg(src1));
}
```

154
compiler/optimizer/code_generator/frame_info.h
View File

@ -1,154 +0,0 @@
/**
* Copyright (c) 2021-2022 Huawei Device Co., Ltd.
* 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.
*/
#ifndef PANDA_FRAME_INFO_H
#define PANDA_FRAME_INFO_H
#include "libpandabase/utils/cframe_layout.h"
#include "libpandabase/utils/bit_field.h"
#include "libpandabase/mem/mem.h"
namespace panda::compiler {
class Encoder;
class Graph;
/**
* This class describes layout of the frame being compiled.
*/
class FrameInfo {
public:
explicit FrameInfo(uint32_t fields) : fields_(fields) {}
~FrameInfo() = default;
NO_COPY_SEMANTIC(FrameInfo);
NO_MOVE_SEMANTIC(FrameInfo);
// NOLINTNEXTLINE(cppcoreguidelines-macro-usage)
#define FRAME_INFO_GET_ATTR(name, var) \
auto Get##name() const \
{ \
return var; \
}
// NOLINTNEXTLINE(cppcoreguidelines-macro-usage)
#define FRAME_INFO_SET_ATTR(name, var) \
void Set##name(ssize_t val) \
{ \
ASSERT(val <= std::numeric_limits<decltype(var)>::max()); \
ASSERT(val >= std::numeric_limits<decltype(var)>::min()); \
var = val; \
}
// NOLINTNEXTLINE(cppcoreguidelines-macro-usage)
#define FRAME_INFO_ATTR(name, var) \
FRAME_INFO_GET_ATTR(name, var) \
FRAME_INFO_SET_ATTR(name, var)
// NOLINTNEXTLINE(cppcoreguidelines-macro-usage)
#define FRAME_INFO_GET_FIELD(name, type) \
type Get##name() const \
{ \
return name::Get(fields_); \
}
// NOLINTNEXTLINE(cppcoreguidelines-macro-usage)
#define FRAME_INFO_SET_FIELD(name, type) \
void Set##name(type val) \
{ \
name::Set(val, &fields_); \
}
// NOLINTNEXTLINE(cppcoreguidelines-macro-usage)
#define FRAME_INFO_FIELD(name, type) \
FRAME_INFO_GET_FIELD(name, type) \
FRAME_INFO_SET_FIELD(name, type)
FRAME_INFO_ATTR(FrameSize, frame_size_);
FRAME_INFO_ATTR(SpillsCount, spills_count_);
FRAME_INFO_ATTR(CallersOffset, callers_offset_);
FRAME_INFO_ATTR(CalleesOffset, callees_offset_);
FRAME_INFO_ATTR(FpCallersOffset, fp_callers_offset_);
FRAME_INFO_ATTR(FpCalleesOffset, fp_callees_offset_);
FRAME_INFO_FIELD(PositionedCallers, bool);
FRAME_INFO_FIELD(PositionedCallees, bool);
FRAME_INFO_FIELD(CallersRelativeFp, bool);
FRAME_INFO_FIELD(CalleesRelativeFp, bool);
// SaveFrameAndLinkRegs - save/restore FP and LR registers in prologue/epilogue.
FRAME_INFO_FIELD(SaveFrameAndLinkRegs, bool);
// SetupFrame - setup CFrame (aka. 'managed' frame).
// Namely, set FP reg, method and flags in prologue.
FRAME_INFO_FIELD(SetupFrame, bool);
// SaveUnusedCalleeRegs - save/restore used+unused callee-saved registers in prologue/epilogue.
FRAME_INFO_FIELD(SaveUnusedCalleeRegs, bool);
// AdjustSpReg - sub SP,#framesize in prologue and add SP,#framesize in epilogue.
FRAME_INFO_FIELD(AdjustSpReg, bool);
FRAME_INFO_FIELD(HasFloatRegs, bool);
using PositionedCallers = BitField<bool, 0, 1>;
using PositionedCallees = PositionedCallers::NextFlag;
using CallersRelativeFp = PositionedCallees::NextFlag;
using CalleesRelativeFp = CallersRelativeFp::NextFlag;
using SaveFrameAndLinkRegs = CalleesRelativeFp::NextFlag;
using SetupFrame = SaveFrameAndLinkRegs::NextFlag;
using SaveUnusedCalleeRegs = SetupFrame::NextFlag;
using AdjustSpReg = SaveUnusedCalleeRegs::NextFlag;
using HasFloatRegs = AdjustSpReg::NextFlag;
// The following static 'constructors' are for situations
// when we have to generate prologue/epilogue but there is
// no codegen at hand (some tests etc.)
// 'Leaf' means a prologue for a function which does not call
// any other functions (library, runtime etc.)
static FrameInfo LeafPrologue()
{
return FrameInfo(AdjustSpReg::Encode(true));
}
// 'Native' means just a regular prologue, that is used for native functions.
// 'Native' is also used for Irtoc.
static FrameInfo NativePrologue()
{
return FrameInfo(AdjustSpReg::Encode(true) | SaveFrameAndLinkRegs::Encode(true) |
SaveUnusedCalleeRegs::Encode(true));
}
// 'Full' means NativePrologue + setting up frame (set FP, method and flags),
// i.e. a prologue for managed code.
static FrameInfo FullPrologue()
{
return FrameInfo(AdjustSpReg::Encode(true) | SaveFrameAndLinkRegs::Encode(true) |
SaveUnusedCalleeRegs::Encode(true) | SetupFrame::Encode(true));
}
#undef FRAME_INFO_GET_ATTR
#undef FRAME_INFO_SET_ATTR
#undef FRAME_INFO_ATTR
#undef FRAME_INFO_GET_FIELD
#undef FRAME_INFO_SET_FIELD
#undef FRAME_INFO_FIELD
private:
uint32_t fields_ {0};
int32_t frame_size_ {0};
int16_t spills_count_ {0};
// Offset to caller registers storage (in words)
int16_t callers_offset_ {0};
int16_t callees_offset_ {0};
int16_t fp_callers_offset_ {0};
int16_t fp_callees_offset_ {0};
};
} // namespace panda::compiler
#endif // PANDA_FRAME_INFO_H

59
compiler/optimizer/code_generator/lib_call_inst.h
View File

@ -1,59 +0,0 @@
/**
* Copyright (c) 2022 Huawei Device Co., Ltd.
* 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.
*/
#ifndef COMPILER_OPTIMIZER_CODEGEN_LIB_CALL_INST_H
#define COMPILER_OPTIMIZER_CODEGEN_LIB_CALL_INST_H
#include "compiler/optimizer/ir/graph.h"
#include "compiler/optimizer/ir/inst.h"
namespace panda::compiler {
inline bool HasLibCall(Inst *inst, Arch arch)
{
auto opcode = inst->GetOpcode();
auto type = inst->GetType();
switch (arch) {
case Arch::X86_64:
case Arch::AARCH64: {
if (opcode == Opcode::Mod) {
return DataType::IsFloatType(type);
}
return false;
}
case Arch::AARCH32: {
if (opcode == Opcode::Mod) {
return true;
}
if (opcode == Opcode::Div) {
return type == DataType::INT64 || type == DataType::UINT64;
}
if (opcode == Opcode::Cast) {
auto src_type = inst->GetInputType(0);
if (DataType::IsFloatType(type)) {
return src_type == DataType::INT64 || src_type == DataType::UINT64;
}
if (DataType::IsFloatType(src_type)) {
return type == DataType::INT64 || type == DataType::UINT64;
}
}
return false;
}
default:
UNREACHABLE();
}
}
} // namespace panda::compiler
#endif // COMPILER_OPTIMIZER_CODEGEN_LIB_CALL_INST_H

82
compiler/optimizer/code_generator/method_properties.cpp
View File

@ -1,82 +0,0 @@
/**
* Copyright (c) 2021-2022 Huawei Device Co., Ltd.
* 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 "optimizer/code_generator/codegen.h"
#include "optimizer/ir/graph.h"
#include "optimizer/ir/graph_visitor.h"
#include "optimizer/code_generator/method_properties.h"
namespace panda::compiler {
MethodProperties::MethodProperties(const Graph *graph)
{
for (auto bb : graph->GetBlocksRPO()) {
// Calls may be in the middle of method
for (auto inst : bb->Insts()) {
if (inst->IsInitObject()) {
ASSERT(options.IsCompilerSupportInitObjectInst());
}
if (inst->GetFlag(inst_flags::CAN_DEOPTIMIZE)) {
SetHasDeopt(true);
}
if (inst->GetOpcode() == Opcode::Return || inst->GetOpcode() == Opcode::ReturnI ||
inst->GetOpcode() == Opcode::ReturnVoid) {
last_return_ = inst;
}
if (!GetHasParamsOnStack() && inst->GetOpcode() == Opcode::Parameter) {
auto sf = static_cast<const ParameterInst *>(inst)->GetLocationData();
if (sf.DstValue() != INVALID_REG && sf.SrcType() == LocationType::STACK_PARAMETER) {
SetHasParamsOnStack(true);
}
}
if (inst->GetOpcode() == Opcode::SafePoint) {
SetHasSafepoints(true);
}
if (inst->IsCall() || inst->IsIntrinsic()) {
SetHasCalls(true);
}
if (Codegen::InstEncodedWithLibCall(inst, graph->GetArch())) {
SetHasLibCalls(true);
}
if (inst->IsRuntimeCall()) {
SetHasRuntimeCalls(true);
}
if (inst->RequireState()) {
SetHasRequireState(true);
}
if (inst->CanThrow()) {
SetCanThrow(true);
}
}
}
/**
* "Compact" prologue/epilogue means that unused callee-saved registers
* are not saved in the prologue and restored in the epilogue.
*
* 1. We do support compact prologue/epilogue only for AARCH64.
* The reasons are as follows:
* - for X86_64 we're reserving almost all callee-saved registers for temporaries
* - for AARCH32 we're treating all callee-saved registers as "used".
* Thus there is no sense in supporting compact prologue for these targets.
*
* 2. We don't support compact prologue/epilogue for OSR to simplify OSR entry bridge.
*/
SetCompactPrologueAllowed(graph->GetArch() == Arch::AARCH64 && !graph->IsOsrMode() &&
options.IsCompilerCompactPrologue());
SetRequireFrameSetup(!IsLeaf() || graph->IsOsrMode());
}
} // namespace panda::compiler

104
compiler/optimizer/code_generator/method_properties.h
View File

@ -1,104 +0,0 @@
/**
* Copyright (c) 2021-2022 Huawei Device Co., Ltd.
* 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.
*/
#ifndef COMPILER_OPTIMIZER_CODEGEN_METHOD_PROPERTIES_H
#define COMPILER_OPTIMIZER_CODEGEN_METHOD_PROPERTIES_H
#include <cstddef>
#include <iostream>
#include "libpandabase/mem/arena_allocator.h"
#include "libpandabase/utils/bit_field.h"
namespace panda::compiler {
class Graph;
class MethodProperties {
public:
explicit MethodProperties(const Graph *graph);
MethodProperties(const MethodProperties &) = default;
MethodProperties &operator=(const MethodProperties &) = default;
MethodProperties(MethodProperties &&) = default;
MethodProperties &operator=(MethodProperties &&) = default;
~MethodProperties() = default;
static MethodProperties *Create(ArenaAllocator *arena_allocator, const Graph *graph)
{
return arena_allocator->New<MethodProperties>(graph);
}
// NOLINTNEXTLINE(cppcoreguidelines-macro-usage)
#define MPROP_GET_FIELD(name, type) \
type Get##name() const \
{ \
return name::Get(fields_); \
}
// NOLINTNEXTLINE(cppcoreguidelines-macro-usage)
#define MPROP_SET_FIELD(name, type) \
void Set##name(type val) \
{ \
name::Set(val, &fields_); \
}
// NOLINTNEXTLINE(cppcoreguidelines-macro-usage)
#define MPROP_FIELD(name, type) \
MPROP_GET_FIELD(name, type) \
MPROP_SET_FIELD(name, type)
using FieldsTy = uint32_t;
MPROP_FIELD(CanThrow, bool);
MPROP_FIELD(HasCalls, bool);
MPROP_FIELD(HasLibCalls, bool);
MPROP_FIELD(HasRuntimeCalls, bool);
MPROP_FIELD(HasSafepoints, bool);
MPROP_FIELD(HasRequireState, bool);
MPROP_FIELD(HasDeopt, bool);
MPROP_FIELD(HasParamsOnStack, bool);
MPROP_FIELD(CompactPrologueAllowed, bool);
MPROP_FIELD(RequireFrameSetup, bool);
using CanThrow = BitField<bool, 0, 1>;
using HasCalls = CanThrow::NextFlag;
using HasLibCalls = HasCalls::NextFlag;
using HasRuntimeCalls = HasLibCalls::NextFlag;
using HasSafepoints = HasRuntimeCalls::NextFlag;
using HasRequireState = HasSafepoints::NextFlag;
using HasDeopt = HasRequireState::NextFlag;
using HasParamsOnStack = HasDeopt::NextFlag;
using CompactPrologueAllowed = HasParamsOnStack::NextFlag;
using RequireFrameSetup = CompactPrologueAllowed::NextFlag;
bool IsLeaf() const
{
return !GetHasCalls() && !GetHasRuntimeCalls() && !GetHasLibCalls() && !GetHasRequireState() && !GetCanThrow();
}
Inst *GetLastReturn() const
{
return last_return_;
}
#undef MPROP_GET_FIELD
#undef MPROP_SET_FIELD
#undef MPROP_FIELD
private:
Inst *last_return_ {nullptr};
FieldsTy fields_ {0};
};
} // namespace panda::compiler
#endif // COMPILER_OPTIMIZER_CODEGEN_METHOD_PROPERTIES_H

940
compiler/optimizer/code_generator/operands.h
View File

@ -1,940 +0,0 @@
/**
* Copyright (c) 2021-2022 Huawei Device Co., Ltd.
* 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.
*/
#ifndef COMPILER_OPTIMIZER_CODEGEN_OPERANDS_H_
#define COMPILER_OPTIMIZER_CODEGEN_OPERANDS_H_
/*
Arch-feature definitions
*/
#include <bitset>
#include <cstdint>
#include "utils/arch.h"
#include "utils/arena_containers.h"
#include "utils/bit_field.h"
#include "utils/bit_utils.h"
#include "utils/regmask.h"
#include "compiler/optimizer/ir/constants.h"
#include "compiler/optimizer/ir/datatype.h"
#include "utils/type_helpers.h"
#ifdef __clang_analyzer__
#ifdef PANDA_TARGET_ARM32
#define __arm__
#endif
#endif
namespace panda::compiler {
constexpr uint8_t BYTE_SIZE = 8;
constexpr uint8_t HALF_SIZE = 16;
constexpr uint8_t WORD_SIZE = 32;
constexpr uint8_t DOUBLE_WORD_SIZE = 64;
constexpr uint8_t WORD_SIZE_BYTE = 4;
constexpr uint8_t DOUBLE_WORD_SIZE_BYTE = 8;
constexpr uint8_t QUAD_WORD_SIZE_BYTE = 16;
/// Maximum possible registers count (for scalar and for vector):
constexpr uint8_t MAX_NUM_REGS = 32;
constexpr uint8_t MAX_NUM_VREGS = 32;
constexpr uint64_t NAN_DOUBLE = uint64_t(0x7ff8000000000000);
constexpr uint32_t NAN_FLOAT = uint32_t(0x7fc00000);
constexpr uint32_t NAN_FLOAT_BITS = NAN_FLOAT >> 16U;
// Constants for cast from float to int64:
// The number of the bit from which exponential part starts in float
constexpr uint8_t START_EXP_FLOAT = 23;
// Size exponential part in float
constexpr uint8_t SIZE_EXP_FLOAT = 8;
// The maximum exponential part of float that can be loaded in int64
constexpr uint32_t POSSIBLE_EXP_FLOAT = 0xbe;
// Mask say that float number is NaN by IEEE 754
constexpr uint32_t UP_BITS_NAN_FLOAT = 0xff;
// Constants for cast from double to int64:
// The number of the bit from which exponential part starts in double
constexpr uint8_t START_EXP_DOUBLE = 20;
// Size exponential part in double
constexpr uint8_t SIZE_EXP_DOUBLE = 11;
// The maximum exponential part of double that can be loaded in int64
constexpr uint32_t POSSIBLE_EXP_DOUBLE = 0x43e;
// Mask say that double number is NaN by IEEE 754
constexpr uint32_t UP_BITS_NAN_DOUBLE = 0x7ff;
constexpr uint32_t SHIFT_BITS_DOUBLE = 12;
constexpr uint32_t SHIFT_BITS_FLOAT = 9;
// Return true, if architecture can be encoded.
bool BackendSupport(Arch arch);
// NOLINTNEXTLINE(cppcoreguidelines-macro-usage)
#define ENCODE_MATH_LIST(DEF) \
DEF(Mov, UnaryOperation) \
DEF(Neg, UnaryOperation) \
DEF(Abs, UnaryOperation) \
DEF(Not, UnaryOperation) \
DEF(Add, BinaryOperation) \
DEF(Sub, BinaryOperation) \
DEF(Mul, BinaryOperation) \
DEF(Shl, BinaryOperation) \
DEF(Shr, BinaryOperation) \
DEF(AShr, BinaryOperation) \
DEF(And, BinaryOperation) \
DEF(Or, BinaryOperation) \
DEF(Xor, BinaryOperation) \
DEF(Sqrt, UnaryOperation)
// NOLINTNEXTLINE(cppcoreguidelines-macro-usage)
#define ENCODE_INST_WITH_SHIFTED_OPERAND(DEF) \
DEF(And, BinaryShiftedRegisterOperation) \
DEF(Or, BinaryShiftedRegisterOperation) \
DEF(Xor, BinaryShiftedRegisterOperation) \
DEF(OrNot, BinaryShiftedRegisterOperation) \
DEF(AndNot, BinaryShiftedRegisterOperation) \
DEF(XorNot, BinaryShiftedRegisterOperation) \
DEF(Add, BinaryShiftedRegisterOperation) \
DEF(Sub, BinaryShiftedRegisterOperation)
// Arch-independent access types
/**
* Template class for identify types compile-time (nortti - can't use typeid).
* Used in register class. Immediate class support conversion to it.
*/
class TypeInfo final {
public:
enum TypeId : uint8_t { INT8 = 0, INT16 = 1, INT32 = 2, INT64 = 3, FLOAT32 = 4, FLOAT64 = 5, INVALID = 6 };
/**
* Template constructor - use template parameter for create object.
*/
template <class T>
constexpr explicit TypeInfo(T /* unused */)
{
#ifndef __clang_analyzer__
if constexpr (std::is_same<T, uint8_t>()) {
type_id_ = INT8;
} else if constexpr (std::is_same<T, int8_t>()) {
type_id_ = INT8;
} else if constexpr (std::is_same<T, uint16_t>()) {
type_id_ = INT16;
} else if constexpr (std::is_same<T, int16_t>()) {
type_id_ = INT16;
} else if constexpr (std::is_same<T, uint32_t>()) {
type_id_ = INT32;
} else if constexpr (std::is_same<T, int32_t>()) {
type_id_ = INT32;
} else if constexpr (std::is_same<T, uint64_t>()) {
type_id_ = INT64;
} else if constexpr (std::is_same<T, int64_t>()) {
type_id_ = INT64;
} else if constexpr (std::is_same<T, float>()) {
type_id_ = FLOAT32;
} else if constexpr (std::is_same<T, double>()) {
type_id_ = FLOAT64;
} else {
type_id_ = INVALID;
}
#endif
}
constexpr explicit TypeInfo(TypeId type) : type_id_(type) {}
DEFAULT_MOVE_SEMANTIC(TypeInfo);
DEFAULT_COPY_SEMANTIC(TypeInfo);
~TypeInfo() = default;
/**
* Constructor for create invalid TypeInfo
*/
constexpr TypeInfo() = default;
/**
* Validation check
*/
constexpr bool IsValid() const
{
return type_id_ != INVALID;
}
/**
* Type expected size
*/
constexpr size_t GetSize() const
{
ASSERT(IsValid());
switch (type_id_) {
case INT8:
return BYTE_SIZE;
case INT16:
return HALF_SIZE;
case INT32:
case FLOAT32:
return WORD_SIZE;
case INT64:
case FLOAT64:
return DOUBLE_WORD_SIZE;
default:
return 0;
}
return 0;
}
constexpr bool IsFloat() const
{
ASSERT(IsValid());
return type_id_ == FLOAT32 || type_id_ == FLOAT64;
}
constexpr bool IsScalar() const
{
// VOID - is scalar type here
return !IsFloat();
}
constexpr bool operator==(const TypeInfo &other) const
{
return (type_id_ == other.type_id_);
}
constexpr bool operator!=(const TypeInfo &other) const
{
return !operator==(other);
}
static TypeInfo FromDataType(DataType::Type type, Arch arch)
{
switch (type) {
case DataType::BOOL:
case DataType::UINT8:
case DataType::INT8: {
return TypeInfo(INT8);
}
case DataType::UINT16:
case DataType::INT16: {
return TypeInfo(INT16);
}
case DataType::UINT32:
case DataType::INT32: {
return TypeInfo(INT32);
}
case DataType::UINT64:
case DataType::INT64:
case DataType::ANY: {
return TypeInfo(INT64);
}
case DataType::FLOAT32: {
return TypeInfo(FLOAT32);
}
case DataType::FLOAT64: {
return TypeInfo(FLOAT64);
}
case DataType::REFERENCE: {
return FromDataType(DataType::GetIntTypeForReference(arch), arch);
}
case DataType::POINTER: {
return Is64BitsArch(arch) ? TypeInfo(INT64) : TypeInfo(INT32);
}
default:
UNREACHABLE();
}
}
DataType::Type ToDataType() const
{
switch (type_id_) {
case INT8:
return DataType::INT8;
case INT16:
return DataType::INT16;
case INT32:
return DataType::INT32;
case INT64:
return DataType::INT64;
case FLOAT32:
return DataType::FLOAT32;
case FLOAT64:
return DataType::FLOAT64;
default:
UNREACHABLE();
}
}
static constexpr TypeInfo GetScalarTypeBySize(size_t size);
void Dump()
{
std::cerr << "TypeInfo:";
switch (type_id_) {
case INT8:
std::cerr << "INT8";
break;
case INT16:
std::cerr << "INT16";
break;
case INT32:
std::cerr << "INT32";
break;
case FLOAT32:
std::cerr << "FLOAT32";
break;
case INT64:
std::cerr << "INT64";
break;
case FLOAT64:
std::cerr << "FLOAT64";
break;
default:
std::cerr << "INVALID";
break;
}
std::cerr << ", size = " << GetSize();
}
private:
TypeId type_id_ {INVALID};
};
constexpr TypeInfo INT8_TYPE {TypeInfo::INT8};
constexpr TypeInfo INT16_TYPE {TypeInfo::INT16};
constexpr TypeInfo INT32_TYPE {TypeInfo::INT32};
constexpr TypeInfo INT64_TYPE {TypeInfo::INT64};
constexpr TypeInfo FLOAT32_TYPE {TypeInfo::FLOAT32};
constexpr TypeInfo FLOAT64_TYPE {TypeInfo::FLOAT64};
constexpr TypeInfo INVALID_TYPE;
constexpr TypeInfo TypeInfo::GetScalarTypeBySize(size_t size)
{
auto type = INT64_TYPE;
if (size == BYTE_SIZE) {
type = INT8_TYPE;
} else if (size == HALF_SIZE) {
type = INT16_TYPE;
} else if (size == WORD_SIZE) {
type = INT32_TYPE;
}
return type;
}
// Mapping model for registers:
// reg-reg - support getters for small parts of registers
// reg-other - mapping between types of registers
enum RegMapping : uint32_t {
SCALAR_SCALAR = 1UL << 0UL,
SCALAR_VECTOR = 1UL << 1UL,
SCALAR_FLOAT = 1UL << 2UL,
VECTOR_VECTOR = 1UL << 3UL,
VECTOR_FLOAT = 1UL << 4UL,
FLOAT_FLOAT = 1UL << 5UL
};
constexpr uint8_t INVALID_REG_ID = std::numeric_limits<uint8_t>::max();
constexpr uint8_t ACC_REG_ID = INVALID_REG_ID - 1U;
class Reg final {
public:
using RegIDType = uint8_t;
using RegSizeType = size_t;
constexpr Reg() = default;
DEFAULT_MOVE_SEMANTIC(Reg);
DEFAULT_COPY_SEMANTIC(Reg);
~Reg() = default;
// Default register constructor
constexpr Reg(RegIDType id, TypeInfo type) : id_(id), type_(type) {}
constexpr RegIDType GetId() const
{
return id_;
}
constexpr size_t GetMask() const
{
return (1U << id_);
}
constexpr TypeInfo GetType() const
{
return type_;
}
RegSizeType GetSize() const
{
return GetType().GetSize();
}
bool IsScalar() const
{
return GetType().IsScalar();
}
bool IsFloat() const
{
return GetType().IsFloat();
}
constexpr bool IsValid() const
{
return type_ != INVALID_TYPE && id_ != INVALID_REG_ID;
}
Reg As(TypeInfo type) const
{
return Reg(GetId(), type);
}
constexpr bool operator==(Reg other) const
{
return (GetId() == other.GetId()) && (GetType() == other.GetType());
}
constexpr bool operator!=(Reg other) const
{
return !operator==(other);
}
void Dump()
{
std::cerr << " Reg: id = " << static_cast<int64_t>(id_) << ", ";
type_.Dump();
std::cerr << "\n";
}
private:
RegIDType id_ {INVALID_REG_ID};
TypeInfo type_ {INVALID_TYPE};
}; // Reg
constexpr Reg INVALID_REGISTER = Reg();
static_assert(!INVALID_REGISTER.IsValid());
static_assert(sizeof(Reg) <= sizeof(uintptr_t));
/**
* Immediate class may hold only int or float values (maybe vectors in future).
* It knows nothing about pointers and bools (bools maybe be in future).
*/
class Imm final {
static inline constexpr uint8_t BITS_PER_BYTE = 8;
static constexpr size_t UNDEFINED_SIZE = 0;
static constexpr size_t INT8_SIZE = 8;
static constexpr size_t INT16_SIZE = 16;
static constexpr size_t INT32_SIZE = 32;
static constexpr size_t INT64_SIZE = 64;
static constexpr size_t FLOAT32_SIZE = 32;
static constexpr size_t FLOAT64_SIZE = 64;
enum VariantID {
// Pointer used for invalidate variants
V_INT8 = 1,
V_INT16 = 2,
V_INT32 = 3,
V_INT64 = 4,
V_FLOAT32 = 5,
V_FLOAT64 = 6,
};
template <class T>
constexpr bool CheckVariantID() const
{
#ifndef __clang_analyzer__
// Immediate could be only signed (int/float)
// look at value_-type.
static_assert(std::is_signed<T>::value);
if constexpr (std::is_same<T, int8_t>()) {
return value_.index() == V_INT8;
}
if constexpr (std::is_same<T, int16_t>()) {
return value_.index() == V_INT16;
}
if constexpr (std::is_same<T, int32_t>()) {
return value_.index() == V_INT32;
}
if constexpr (std::is_same<T, int64_t>()) {
return value_.index() == V_INT64;
}
if constexpr (std::is_same<T, float>()) {
return value_.index() == V_FLOAT32;
}
if constexpr (std::is_same<T, double>()) {
return value_.index() == V_FLOAT64;
}
return false;
#else
return true;
#endif
}
public:
// Invalid constructor
constexpr Imm() = default;
// Special type constructor
template <class T>
constexpr explicit Imm(T value) : value_(value)
{
}
// Partial template specialization
constexpr explicit Imm(uint8_t value) : value_(static_cast<int8_t>(value)) {};
constexpr explicit Imm(uint16_t value) : value_(static_cast<int16_t>(value)) {};
constexpr explicit Imm(uint32_t value) : value_(static_cast<int32_t>(value)) {};
constexpr explicit Imm(uint64_t value) : value_(static_cast<int64_t>(value)) {};
#if (PANDA_TARGET_MACOS)
constexpr explicit Imm(size_t value) : value_(static_cast<int64_t>(value)) {};
constexpr explicit Imm(long value) : value_(static_cast<int64_t>(value)) {};
#endif
DEFAULT_MOVE_SEMANTIC(Imm);
DEFAULT_COPY_SEMANTIC(Imm);
~Imm() = default;
template <class T>
T GetValue() const
{
ASSERT(CheckVariantID<T>());
ASSERT(sizeof(T) * BITS_PER_BYTE == GetSize());
return std::get<T>(value_);
}
void Inc(size_t value)
{
switch (value_.index()) {
case V_INT8:
value_ = static_cast<int8_t>(std::get<int8_t>(value_) + value);
break;
case V_INT16:
value_ = static_cast<int16_t>(std::get<int16_t>(value_) + value);
break;
case V_INT32:
value_ = static_cast<int32_t>(std::get<int32_t>(value_) + value);
break;
case V_INT64:
value_ = static_cast<int64_t>(std::get<int64_t>(value_) + value);
break;
case V_FLOAT32:
value_ = static_cast<float>(std::get<float>(value_) + value);
break;
case V_FLOAT64:
value_ = static_cast<double>(std::get<double>(value_) + value);
break;
default:
// Check before increment
UNREACHABLE();
break;
}
}
void Dec(size_t value)
{
switch (value_.index()) {
case V_INT8:
value_ = static_cast<int8_t>(std::get<int8_t>(value_) - value);
break;
case V_INT16:
value_ = static_cast<int16_t>(std::get<int16_t>(value_) - value);
break;
case V_INT32:
value_ = static_cast<int32_t>(std::get<int32_t>(value_) - value);
break;
case V_INT64:
value_ = static_cast<int64_t>(std::get<int64_t>(value_) - value);
break;
case V_FLOAT32:
value_ = static_cast<float>(std::get<float>(value_) - value);
break;
case V_FLOAT64:
value_ = static_cast<double>(std::get<double>(value_) - value);
break;
default:
// Check before decrement
UNREACHABLE();
break;
}
}
TypeInfo GetType() const
{
switch (value_.index()) {
case V_INT8:
return INT8_TYPE;
case V_INT16:
return INT16_TYPE;
case V_INT32:
return INT32_TYPE;
case V_INT64:
return INT64_TYPE;
case V_FLOAT32:
return FLOAT32_TYPE;
case V_FLOAT64:
return FLOAT64_TYPE;
default:
UNREACHABLE();
return INVALID_TYPE;
}
}
constexpr size_t GetSize() const
{
switch (value_.index()) {
case V_INT8:
return INT8_SIZE;
case V_INT16:
return INT16_SIZE;
case V_INT32:
return INT32_SIZE;
case V_INT64:
return INT64_SIZE;
case V_FLOAT32:
return FLOAT32_SIZE;
case V_FLOAT64:
return FLOAT64_SIZE;
default:
return UNDEFINED_SIZE;
}
}
bool IsZero() const
{
if (std::holds_alternative<float>(value_)) {
return std::get<float>(value_) == 0.0;
}
if (std::holds_alternative<double>(value_)) {
return std::get<double>(value_) == 0.0;
}
if (std::holds_alternative<int8_t>(value_)) {
return std::get<int8_t>(value_) == 0;
}
if (std::holds_alternative<int16_t>(value_)) {
return std::get<int16_t>(value_) == 0;
}
if (std::holds_alternative<int32_t>(value_)) {
return std::get<int32_t>(value_) == 0;
}
if (std::holds_alternative<int64_t>(value_)) {
return std::get<int64_t>(value_) == 0;
}
return true;
}
bool IsFloat() const
{
return std::holds_alternative<float>(value_) || std::holds_alternative<double>(value_);
}
bool IsScalar() const
{
return std::holds_alternative<int8_t>(value_) || std::holds_alternative<int16_t>(value_) ||
std::holds_alternative<int32_t>(value_) || std::holds_alternative<int64_t>(value_);
}
bool IsValid() const
{
bool hold_data = std::holds_alternative<int8_t>(value_) || std::holds_alternative<int16_t>(value_) ||
std::holds_alternative<int32_t>(value_) || std::holds_alternative<int64_t>(value_) ||
std::holds_alternative<float>(value_) || std::holds_alternative<double>(value_);
return (GetSize() != 0) && hold_data;
}
unsigned GetShift()
{
if (GetType() == INT64_TYPE) {
return GetValue<int64_t>();
}
if (GetType() == INT32_TYPE) {
return GetValue<int32_t>();
}
if (GetType() == INT16_TYPE) {
return GetValue<int16_t>();
}
if (GetType() == INT8_TYPE) {
return GetValue<int8_t>();
}
UNREACHABLE();
return 0;
}
bool operator==(Imm other) const
{
return value_ == other.value_;
}
bool operator!=(Imm other) const
{
return !(operator==(other));
}
private:
std::variant<void *, int8_t, int16_t, int32_t, int64_t, float, double> value_ {nullptr};
}; // Imm
constexpr Imm INVALID_IMM = Imm();
// Why memory ref - because you may create one link for one encode-session
// And when you see this one - you can easy understand, what type of memory
// you use. But if you load/store dirrectly address - you need to decode it
// each time, when you read code
// model -> base + index<<scale + disp
class MemRef final {
public:
MemRef() = default;
explicit MemRef(Reg base) : MemRef(base, 0) {}
MemRef(Reg base, ssize_t disp) : MemRef(base, INVALID_REGISTER, 0, disp) {}
MemRef(Reg base, Reg index, uint16_t scale) : MemRef(base, index, scale, 0) {}
MemRef(Reg base, Reg index, uint16_t scale, ssize_t disp) : disp_(disp), scale_(scale), base_(base), index_(index)
{
CHECK_LE(disp, std::numeric_limits<decltype(disp_)>::max());
CHECK_LE(scale, std::numeric_limits<decltype(scale_)>::max());
}
DEFAULT_MOVE_SEMANTIC(MemRef);
DEFAULT_COPY_SEMANTIC(MemRef);
~MemRef() = default;
Reg GetBase() const
{
return base_;
}
Reg GetIndex() const
{
return index_;
}
auto GetScale() const
{
return scale_;
}
auto GetDisp() const
{
return disp_;
}
bool HasBase() const
{
return base_.IsValid();
}
bool HasIndex() const
{
return index_.IsValid();
}
bool HasScale() const
{
return HasIndex() && scale_ != 0;
}
bool HasDisp() const
{
return disp_ != 0;
}
// Ref must contain at least one of field
bool IsValid() const
{
return HasBase() || HasIndex() || HasScale() || HasDisp();
}
// return true if mem doesn't has index and scalar
bool IsOffsetMem() const
{
return !HasIndex() && !HasScale();
}
bool operator==(MemRef other) const
{
return (base_ == other.base_) && (index_ == other.index_) && (scale_ == other.scale_) && (disp_ == other.disp_);
}
bool operator!=(MemRef other) const
{
return !(operator==(other));
}
private:
ssize_t disp_ {0};
uint16_t scale_ {0};
Reg base_ {INVALID_REGISTER};
Reg index_ {INVALID_REGISTER};
}; // MemRef
inline ArenaVector<std::pair<uint8_t, uint8_t>> ResoveParameterSequence(
ArenaVector<std::pair<uint8_t, uint8_t>> *moved_registers, uint8_t tmp, ArenaAllocator *allocator)
{
constexpr uint8_t INVALID_FIST = -1;
constexpr uint8_t INVALID_SECOND = -2;
moved_registers->emplace_back(std::pair<uint8_t, uint8_t>(INVALID_FIST, INVALID_SECOND));
/*
Example:
1. mov x0 <- x3
2. mov x1 <- x0
3. mov x2 <- x3
4. mov x3 <- x2
Agreement - in dst can't be multipy same registers (double move to one register)
- src for movs can hold same register multiply times
Algorithm:
1. Find handing edges (x1 - just in dst)
emit "2. mov x1 <- x0"
goto 1.
emit "1. mov x0 <- x3"
2. Assert all registers used just one time (loop from registers sequence)
All multiply-definitions must be resolved on previous step
emit ".. mov xtmp <- x2" (strore xtmp == x3)
emit "3. mov x2 <- x3"
emit "4. mov x3 <- xtmp" (ASSERT(4->GetReg == x3) - there is no other possible situations here)
*/
// Calculate weigth
ArenaVector<std::pair<uint8_t, uint8_t>> result(allocator->Adapter());
// --moved_registers->end() - for remove marker-element
for (auto pair = moved_registers->begin(); pair != --moved_registers->end();) {
auto conflict = std::find_if(moved_registers->begin(), moved_registers->end(), [pair](auto in_pair) {
return (in_pair.second == pair->first && (in_pair != *pair));
});
if (conflict == moved_registers->end()) {
// emit immediate - there are no another possible combinations
result.emplace_back(*pair);
moved_registers->erase(pair);
pair = moved_registers->begin();
} else {
++pair;
}
}
// Here just loops
for (;;) {
/* Need support single mov x1 <- x1:
ASSERT(moved_registers->size() != 1);
*/
auto curr_pair = moved_registers->begin();
if (curr_pair->first == INVALID_FIST && curr_pair->second == INVALID_SECOND) {
moved_registers->erase(curr_pair);
break;
// Finish algorithm - only marker in vector
}
auto saved_reg = curr_pair->first;
result.emplace_back(std::pair<uint8_t, uint8_t>(tmp, curr_pair->first));
result.emplace_back(*curr_pair); // we already save dst_register
// Remove current instruction
auto curr_reg = curr_pair->second;
moved_registers->erase(curr_pair);
for (; curr_pair != moved_registers->end();) {
curr_pair = std::find_if(moved_registers->begin(), moved_registers->end(),
[curr_reg](auto in_pair) { return in_pair.first == curr_reg; });
if (curr_pair != moved_registers->end()) {
if (curr_pair->second == saved_reg) {
result.emplace_back(std::pair<uint8_t, uint8_t>(curr_pair->first, tmp));
moved_registers->erase(curr_pair);
break;
// exit from loop
};
result.emplace_back(*curr_pair);
curr_reg = curr_pair->second;
moved_registers->erase(curr_pair);
} else {
ASSERT(curr_pair != moved_registers->end());
}
}
}
return result;
}
// Condition also used for tell comparison registers type
enum Condition {
EQ, // equal to 0
NE, // not equal to 0
// signed
LT, // less
LE, // less than or equal
GT, // greater
GE, // greater than or equal
// unsigned - checked from registers
LO, // less
LS, // less than or equal
HI, // greater
HS, // greater than or equal
// Special arch-dependecy TODO (igorban) Fix them
MI, // N set Negative
PL, // N clear Positive or zero
VS, // V set Overflow.
VC, // V clear No overflow.
AL, // Always.
NV, // Behaves as always/al.
TST_EQ,
TST_NE,
INVALID_COND
};
static inline bool IsTestCc(Condition cond)
{
return cond == TST_EQ || cond == TST_NE;
}
class Shift final {
public:
explicit Shift(Reg base, ShiftType type, uint32_t scale) : scale_(scale), base_(base), type_(type) {}
explicit Shift(Reg base, uint32_t scale) : Shift(base, ShiftType::LSL, scale) {}
DEFAULT_MOVE_SEMANTIC(Shift);
DEFAULT_COPY_SEMANTIC(Shift);
~Shift() = default;
Reg GetBase() const
{
return base_;
}
ShiftType GetType() const
{
return type_;
}
uint32_t GetScale() const
{
return scale_;
}
private:
uint32_t scale_ {0};
Reg base_;
ShiftType type_ {INVALID_SHIFT};
};
inline int64_t GetIntValue(Imm imm)
{
int64_t value {0};
auto type = imm.GetType();
if (type == INT32_TYPE) {
value = imm.GetValue<int32_t>();
} else if (type == INT64_TYPE) {
value = imm.GetValue<int64_t>();
} else if (type == INT16_TYPE) {
value = imm.GetValue<int16_t>();
} else if (type == INT8_TYPE) {
value = imm.GetValue<int8_t>();
} else {
// Inconsistent int-type
UNREACHABLE();
}
return value;
}
} // namespace panda::compiler
#endif // COMPILER_OPTIMIZER_CODEGEN_REGISTERS_H_

152
compiler/optimizer/code_generator/registers_description.h
View File

@ -1,152 +0,0 @@
/**
* Copyright (c) 2021-2022 Huawei Device Co., Ltd.
* 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.
*/
#ifndef COMPILER_OPTIMIZER_CODEGEN_REGFILE_H_
#define COMPILER_OPTIMIZER_CODEGEN_REGFILE_H_
#include "operands.h"
#include "utils/arch.h"
/*
Register file wrapper used for get major data and for Regalloc
*/
namespace panda::compiler {
#ifdef PANDA_COMPILER_TARGET_X86_64
namespace amd64 {
static constexpr size_t RENAMING_MASK_3_5_OR_9_11 {0xE38};
static constexpr size_t RENAMING_CONST {14U};
// There is a problem with callee/caller register numbers with amd64.
// For example, take a look at
// caller reg mask: 0000111111000111 and
// callee reg mask: 1111000000001000
// Stack walker requires this mask to be densed, so the decision is to
// rename regs number 3, 4, 5 to 11, 10, 9 (and vice versa).
// Resulting
// caller mask is 0000000111111111 and
// callee mask is 1111100000000000.
static inline constexpr size_t ConvertRegNumber(size_t reg_id)
{
ASSERT(reg_id < MAX_NUM_REGS);
// NOLINTNEXTLINE(clang-analyzer-core.UndefinedBinaryOperatorResult)
if ((RENAMING_MASK_3_5_OR_9_11 & (size_t(1) << reg_id)) != 0) {
return RENAMING_CONST - reg_id;
}
return reg_id;
}
} // namespace amd64
#endif // PANDA_COMPILER_TARGET_X86_64
class RegistersDescription {
public:
explicit RegistersDescription(ArenaAllocator *aa, Arch arch) : arena_allocator_(aa), arch_(arch) {}
virtual ~RegistersDescription() = default;
virtual ArenaVector<Reg> GetCalleeSaved() = 0;
virtual void SetCalleeSaved(const ArenaVector<Reg> &) = 0;
// Set used regs - change GetCallee
virtual void SetUsedRegs(const ArenaVector<Reg> &) = 0;
// Return zero register. If target architecture doesn't support zero register, it should return INVALID_REGISTER.
virtual Reg GetZeroReg() const = 0;
virtual bool IsZeroReg(Reg reg) const = 0;
virtual Reg::RegIDType GetTempReg() = 0;
virtual Reg::RegIDType GetTempVReg() = 0;
// Return RegMapping bitset
virtual bool SupportMapping(uint32_t) = 0;
virtual bool IsValid() const
{
return false;
};
virtual bool IsCalleeRegister(Reg reg) = 0;
ArenaAllocator *GetAllocator() const
{
return arena_allocator_;
};
// May be re-define to ignore some cases
virtual bool IsRegUsed(ArenaVector<Reg> vec_reg, Reg reg)
{
// size ignored in arm64
auto equality = [reg](Reg in) {
return ((reg.GetId() == in.GetId()) && (reg.GetType() == in.GetType()) &&
(reg.GetSize() == in.GetSize())) ||
(!reg.IsValid() && !in.IsValid());
};
return (std::find_if(vec_reg.begin(), vec_reg.end(), equality) != vec_reg.end());
}
static RegistersDescription *Create(ArenaAllocator *arena_allocator, Arch arch);
RegMask GetRegMask() const
{
return reg_mask_.None() ? GetDefaultRegMask() : reg_mask_;
}
void SetRegMask(const RegMask &mask)
{
reg_mask_ = mask;
}
// Get registers mask which used in codegen, runtime e.t.c
// 0 means - available, 1 - unavailable to use
// Note that it is a default architecture-specific registers mask.
virtual RegMask GetDefaultRegMask() const = 0;
// Get vector registers mask which used in codegen, runtime e.t.c
virtual VRegMask GetVRegMask() = 0;
virtual RegMask GetCallerSavedRegMask() const = 0;
virtual RegMask GetCallerSavedVRegMask() const = 0;
void FillUsedCalleeSavedRegisters(RegMask *callee_regs, VRegMask *callee_vregs, bool set_all_callee_registers)
{
if (set_all_callee_registers) {
*callee_regs = RegMask(panda::GetCalleeRegsMask(arch_, false));
*callee_vregs = VRegMask(panda::GetCalleeRegsMask(arch_, true));
} else {
*callee_regs = GetUsedRegsMask<RegMask, false>(GetCalleeSaved());
*callee_vregs = GetUsedRegsMask<VRegMask, true>(GetCalleeSaved());
}
}
NO_COPY_SEMANTIC(RegistersDescription);
NO_MOVE_SEMANTIC(RegistersDescription);
private:
ArenaAllocator *arena_allocator_ {nullptr};
Arch arch_;
RegMask reg_mask_ {0};
template <typename M, bool is_fp>
M GetUsedRegsMask(const ArenaVector<Reg> &regs)
{
M mask;
for (auto reg : regs) {
// NOLINTNEXTLINE(bugprone-branch-clone,-warnings-as-errors)
if (reg.IsFloat() && is_fp) {
mask.set(reg.GetId());
} else if (reg.IsScalar() && !is_fp) {
mask.set(reg.GetId());
}
}
return mask;
}
};
} // namespace panda::compiler
#endif // COMPILER_OPTIMIZER_CODEGEN_REGFILE_H_

36
compiler/optimizer/code_generator/relocations.h
View File

@ -1,36 +0,0 @@
/**
* Copyright (c) 2021-2022 Huawei Device Co., Ltd.
* 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.
*/
#ifndef PANDA_RELOCATIONS_H
#define PANDA_RELOCATIONS_H
namespace panda::compiler {
class RelocationInfo {
public:
uint32_t data {0};
uint32_t offset {0};
uint32_t type {0};
int8_t addend {0};
};
class RelocationHandler {
public:
virtual void AddRelocation(const RelocationInfo &info) = 0;
};
} // namespace panda::compiler
#endif // PANDA_RELOCATIONS_H

288
compiler/optimizer/code_generator/slow_path.cpp
View File

@ -1,288 +0,0 @@
/**
* Copyright (c) 2021-2022 Huawei Device Co., Ltd.
* 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 "slow_path.h"
#include "codegen.h"
namespace panda::compiler {
void SlowPathBase::Generate(Codegen *codegen)
{
ASSERT(!generated_);
SCOPED_DISASM_STR(codegen, std::string("SlowPath for inst ") + std::to_string(GetInst()->GetId()) + ". " +
GetInst()->GetOpcodeStr());
Encoder *encoder = codegen->GetEncoder();
ASSERT(encoder->IsValid());
encoder->BindLabel(GetLabel());
GenerateImpl(codegen);
if (encoder->IsLabelValid(label_back_)) {
codegen->GetEncoder()->EncodeJump(GetBackLabel());
}
#ifndef NDEBUG
generated_ = true;
#endif
}
// ARRAY_INDEX_OUT_OF_BOUNDS_EXCEPTION, STRING_INDEX_OUT_OF_BOUNDS_EXCEPTION
bool SlowPathEntrypoint::GenerateThrowOutOfBoundsException(Codegen *codegen)
{
auto len_reg = codegen->ConvertRegister(GetInst()->GetSrcReg(0), GetInst()->GetInputType(0));
if (GetInst()->GetOpcode() == Opcode::BoundsCheckI) {
ScopedTmpReg index_reg(codegen->GetEncoder());
codegen->GetEncoder()->EncodeMov(index_reg, Imm(GetInst()->CastToBoundsCheckI()->GetImm()));
codegen->CallRuntime(GetInst(), GetEntrypoint(), INVALID_REGISTER, {index_reg, len_reg});
} else {
ASSERT(GetInst()->GetOpcode() == Opcode::BoundsCheck);
auto index_reg = codegen->ConvertRegister(GetInst()->GetSrcReg(1), GetInst()->GetInputType(1));
codegen->CallRuntime(GetInst(), GetEntrypoint(), INVALID_REGISTER, {index_reg, len_reg});
}
return true;
}
// INITIALIZE_CLASS
bool SlowPathEntrypoint::GenerateInitializeClass(Codegen *codegen)
{
auto inst = GetInst();
if (GetInst()->GetDstReg() != INVALID_REG) {
ASSERT(inst->GetOpcode() == Opcode::LoadAndInitClass);
Reg klass_reg {codegen->ConvertRegister(GetInst()->GetDstReg(), DataType::REFERENCE)};
RegMask preserved_regs;
codegen->GetEncoder()->SetRegister(&preserved_regs, nullptr, klass_reg);
codegen->CallRuntime(GetInst(), GetEntrypoint(), INVALID_REGISTER, {klass_reg}, preserved_regs);
} else {
ASSERT(inst->GetOpcode() == Opcode::InitClass);
ASSERT(!codegen->GetGraph()->IsAotMode());
auto klass = reinterpret_cast<uintptr_t>(inst->CastToInitClass()->GetClass());
codegen->CallRuntime(GetInst(), GetEntrypoint(), INVALID_REGISTER, {Imm(klass)});
}
return true;
}
// IS_INSTANCE
bool SlowPathEntrypoint::GenerateIsInstance(Codegen *codegen)
{
auto src = codegen->ConvertRegister(GetInst()->GetSrcReg(0), DataType::REFERENCE); // obj
auto klass = codegen->ConvertRegister(GetInst()->GetSrcReg(1), DataType::REFERENCE);
auto dst = codegen->ConvertRegister(GetInst()->GetDstReg(), GetInst()->GetType());
codegen->CallRuntime(GetInst(), EntrypointId::IS_INSTANCE, dst, {src, klass});
return true;
}
// CHECK_CAST
bool SlowPathEntrypoint::GenerateCheckCast(Codegen *codegen)
{
auto src = codegen->ConvertRegister(GetInst()->GetSrcReg(0), DataType::REFERENCE); // obj
auto klass = codegen->ConvertRegister(GetInst()->GetSrcReg(1), DataType::REFERENCE);
codegen->CallRuntime(GetInst(), EntrypointId::CHECK_CAST, INVALID_REGISTER, {src, klass});
return true;
}
// DEOPTIMIZE
bool SlowPathEntrypoint::GenerateDeoptimize(Codegen *codegen)
{
DeoptimizeType type = DeoptimizeType::INVALID;
if (GetInst()->GetOpcode() == Opcode::Deoptimize) {
type = GetInst()->CastToDeoptimize()->GetDeoptimizeType();
} else if (GetInst()->GetOpcode() == Opcode::DeoptimizeIf) {
type = GetInst()->CastToDeoptimizeIf()->GetDeoptimizeType();
} else if (GetInst()->GetOpcode() == Opcode::DeoptimizeCompare) {
type = GetInst()->CastToDeoptimizeCompare()->GetDeoptimizeType();
} else if (GetInst()->GetOpcode() == Opcode::DeoptimizeCompareImm) {
type = GetInst()->CastToDeoptimizeCompareImm()->GetDeoptimizeType();
} else if (GetInst()->GetOpcode() == Opcode::AnyTypeCheck) {
type = DeoptimizeType::ANY_TYPE_CHECK;
} else if (GetInst()->GetOpcode() == Opcode::AddOverflowCheck) {
type = DeoptimizeType::DEOPT_OVERFLOW;
} else if (GetInst()->GetOpcode() == Opcode::SubOverflowCheck) {
type = DeoptimizeType::DEOPT_OVERFLOW;
} else {
UNREACHABLE();
}
codegen->CallRuntime(GetInst(), GetEntrypoint(), INVALID_REGISTER, {Imm(static_cast<uint8_t>(type))});
return true;
}
// CREATE_OBJECT
bool SlowPathEntrypoint::GenerateCreateObject(Codegen *codegen)
{
auto inst = GetInst();
auto dst = codegen->ConvertRegister(inst->GetDstReg(), inst->GetType());
auto src = codegen->ConvertRegister(inst->GetSrcReg(0), inst->GetInputType(0));
codegen->CallRuntime(inst, EntrypointId::CREATE_OBJECT_BY_CLASS, dst, {src});
return true;
}
bool SlowPathEntrypoint::GenerateByEntry(Codegen *codegen)
{
switch (GetEntrypoint()) {
case EntrypointId::THROW_EXCEPTION: {
auto src = codegen->ConvertRegister(GetInst()->GetSrcReg(0), DataType::Type::REFERENCE);
codegen->CallRuntime(GetInst(), GetEntrypoint(), INVALID_REGISTER, {src});
return true;
}
case EntrypointId::NULL_POINTER_EXCEPTION:
case EntrypointId::ARITHMETIC_EXCEPTION:
codegen->CallRuntime(GetInst(), GetEntrypoint(), INVALID_REGISTER, {});
return true;
case EntrypointId::ARRAY_INDEX_OUT_OF_BOUNDS_EXCEPTION:
case EntrypointId::STRING_INDEX_OUT_OF_BOUNDS_EXCEPTION:
return GenerateThrowOutOfBoundsException(codegen);
case EntrypointId::NEGATIVE_ARRAY_SIZE_EXCEPTION: {
auto size = codegen->ConvertRegister(GetInst()->GetSrcReg(0), GetInst()->GetInputType(0));
codegen->CallRuntime(GetInst(), GetEntrypoint(), INVALID_REGISTER, {size});
return true;
}
case EntrypointId::INITIALIZE_CLASS:
return GenerateInitializeClass(codegen);
case EntrypointId::IS_INSTANCE:
return GenerateIsInstance(codegen);
case EntrypointId::CHECK_CAST:
return GenerateCheckCast(codegen);
case EntrypointId::CREATE_OBJECT_BY_CLASS:
return GenerateCreateObject(codegen);
case EntrypointId::SAFEPOINT:
codegen->CallRuntime(GetInst(), GetEntrypoint(), INVALID_REGISTER, {});
return true;
case EntrypointId::DEOPTIMIZE: {
return GenerateDeoptimize(codegen);
}
default:
return false;
}
}
void SlowPathEntrypoint::GenerateImpl(Codegen *codegen)
{
if (!GenerateByEntry(codegen)) {
switch (GetEntrypoint()) {
case EntrypointId::GET_UNKNOWN_CALLEE_METHOD:
case EntrypointId::RESOLVE_UNKNOWN_VIRTUAL_CALL:
case EntrypointId::GET_FIELD_OFFSET:
case EntrypointId::GET_UNKNOWN_STATIC_FIELD_MEMORY_ADDRESS:
case EntrypointId::GET_UNKNOWN_STATIC_FIELD_PTR:
case EntrypointId::RESOLVE_CLASS_OBJECT:
case EntrypointId::RESOLVE_CLASS:
case EntrypointId::ABSTRACT_METHOD_ERROR:
case EntrypointId::INITIALIZE_CLASS_BY_ID:
case EntrypointId::CHECK_STORE_ARRAY_REFERENCE:
case EntrypointId::RESOLVE_STRING_AOT:
case EntrypointId::CLASS_CAST_EXCEPTION:
break;
default:
LOG(FATAL, COMPILER) << "Unsupported entrypoint!";
UNREACHABLE();
break;
}
}
}
void SlowPathIntrinsic::GenerateImpl(Codegen *codegen)
{
codegen->CreateCallIntrinsic(GetInst()->CastToIntrinsic());
}
void SlowPathImplicitNullCheck::GenerateImpl(Codegen *codegen)
{
ASSERT(!GetInst()->CastToNullCheck()->IsImplicit());
SlowPathEntrypoint::GenerateImpl(codegen);
}
void SlowPathShared::GenerateImpl(Codegen *codegen)
{
ASSERT(tmp_reg_ != INVALID_REGISTER);
[[maybe_unused]] ScopedTmpReg tmp_reg(codegen->GetEncoder(), tmp_reg_);
ASSERT(tmp_reg.GetReg().GetId() == tmp_reg_.GetId());
auto graph = codegen->GetGraph();
ASSERT(graph->IsAotMode());
auto aot_data = graph->GetAotData();
aot_data->SetSharedSlowPathOffset(GetEntrypoint(), codegen->GetEncoder()->GetCursorOffset());
MemRef entry(codegen->ThreadReg(), graph->GetRuntime()->GetEntrypointTlsOffset(graph->GetArch(), GetEntrypoint()));
ScopedTmpReg tmp1_reg(codegen->GetEncoder());
codegen->GetEncoder()->EncodeLdr(tmp1_reg, false, entry);
codegen->GetEncoder()->EncodeJump(tmp1_reg);
}
void SlowPathResolveStringAot::GenerateImpl(Codegen *codegen)
{
ScopedTmpRegU64 tmp_addr_reg(codegen->GetEncoder());
// Slot address was loaded into temporary register before we jumped into slow path, but it is already released
// because temporary registers are scoped. Try to allocate a new one and check that it is the same register
// as was allocated in codegen. If it is a different register then copy the slot address into it.
if (tmp_addr_reg.GetReg() != addr_reg_) {
codegen->GetEncoder()->EncodeMov(tmp_addr_reg, addr_reg_);
}
codegen->CallRuntimeWithMethod(GetInst(), method_, GetEntrypoint(), dst_reg_, Imm(string_id_), tmp_addr_reg);
}
void SlowPathRefCheck::GenerateImpl(Codegen *codegen)
{
ASSERT(array_reg_ != INVALID_REGISTER);
ASSERT(ref_reg_ != INVALID_REGISTER);
codegen->CallRuntime(GetInst(), GetEntrypoint(), INVALID_REGISTER, {array_reg_, ref_reg_});
}
void SlowPathAbstract::GenerateImpl(Codegen *codegen)
{
SCOPED_DISASM_STR(codegen, std::string("SlowPath for Abstract method ") + std::to_string(GetInst()->GetId()));
ASSERT(method_reg_ != INVALID_REGISTER);
ScopedTmpReg method_reg(codegen->GetEncoder(), method_reg_);
ASSERT(method_reg.GetReg().GetId() == method_reg_.GetId());
codegen->CallRuntime(GetInst(), GetEntrypoint(), INVALID_REGISTER, {method_reg.GetReg()});
}
void SlowPathCheckCast::GenerateImpl(Codegen *codegen)
{
SCOPED_DISASM_STR(codegen, std::string("SlowPath for CheckCast exception") + std::to_string(GetInst()->GetId()));
auto inst = GetInst();
auto src = codegen->ConvertRegister(inst->GetSrcReg(0), inst->GetInputType(0));
codegen->CallRuntime(GetInst(), GetEntrypoint(), INVALID_REGISTER, {class_reg_, src});
}
void SlowPathUnresolved::GenerateImpl(Codegen *codegen)
{
SlowPathEntrypoint::GenerateImpl(codegen);
ASSERT(method_ != nullptr);
ASSERT(type_id_ != 0);
ASSERT(slot_addr_ != 0);
ScopedTmpReg value_reg(codegen->GetEncoder());
if (GetInst()->GetOpcode() == Opcode::UnresolvedCallVirtual) {
codegen->CallRuntimeWithMethod(GetInst(), method_, GetEntrypoint(), value_reg, arg_reg_, Imm(type_id_),
Imm(slot_addr_));
} else if (GetEntrypoint() == EntrypointId::GET_UNKNOWN_CALLEE_METHOD ||
GetEntrypoint() == EntrypointId::GET_UNKNOWN_STATIC_FIELD_MEMORY_ADDRESS ||
GetEntrypoint() == EntrypointId::GET_UNKNOWN_STATIC_FIELD_PTR) {
codegen->CallRuntimeWithMethod(GetInst(), method_, GetEntrypoint(), value_reg, Imm(type_id_), Imm(slot_addr_));
} else {
codegen->CallRuntimeWithMethod(GetInst(), method_, GetEntrypoint(), value_reg, Imm(type_id_));
ScopedTmpReg addr_reg(codegen->GetEncoder());
codegen->GetEncoder()->EncodeMov(addr_reg, Imm(slot_addr_));
codegen->GetEncoder()->EncodeStr(value_reg, MemRef(addr_reg));
}
if (dst_reg_.IsValid()) {
codegen->GetEncoder()->EncodeMov(dst_reg_, value_reg);
}
}
} // namespace panda::compiler

Some files were not shown because too many files have changed in this diff Show More