Question

我使用LLVM C ++ API生成一个非常简单的函数。这是IR输出：

target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"

define i64 @"foo"(i64) {
  %2 = alloca i64
  store i64 %0, i64* %2
  %3 = load i64, i64* %2
  ret i64 %3
}

我希望LLVM在将其编译为目标文件时摆脱多余的存储/加载，但这不是我观察到的。如果我使用llvm::CodeGenOpt::None，我会

$ objdump -M intel -d out.o | grep -A 10 foo
0000000000000000 <foo>:
   0:   48 89 7c 24 f8          mov    QWORD PTR [rsp-0x8],rdi
   5:   48 8b 44 24 f8          mov    rax,QWORD PTR [rsp-0x8]
   a:   c3                      ret

如果我使用llvm::CodeGenOpt::Aggressive，我会

$ objdump -M intel -d out.o | grep -A 10 foo
0000000000000000 <foo>:
   0:   48 89 7c 24 f8          mov    QWORD PTR [rsp-0x8],rdi
   5:   48 89 f8                mov    rax,rdi
   8:   c3                      ret

我希望LLVM能够优化mov QWORD PTR [rsp-0x8],rdi。我错过了一些选择吗？这是目标文件生成代码：

llvm::InitializeAllTargetInfos();
llvm::InitializeAllTargets();
llvm::InitializeAllTargetMCs();
llvm::InitializeAllAsmPrinters();
llvm::InitializeAllAsmParsers();
llvm::InitializeAllDisassemblers();

auto target_triple = llvm::sys::getDefaultTargetTriple();

string error;
auto* target = llvm::TargetRegistry::lookupTarget(target_triple, error);
if (target == nullptr) {
  return error;
}

auto cpu = llvm::sys::getHostCPUName();

llvm::SubtargetFeatures subtarget_features;
llvm::StringMap<bool> feature_map;
if (llvm::sys::getHostCPUFeatures(feature_map)) {
  for (auto &feature : feature_map) {
    subtarget_features.AddFeature(feature.first(), feature.second);
  }
}
auto features = subtarget_features.getString();

llvm::TargetOptions target_options;

auto reloc_model = llvm::Optional<llvm::Reloc::Model>();

auto* target_machine = target->createTargetMachine(
    target_triple,
    cpu,
    features,
    target_options,
    reloc_model,
    llvm::CodeModel::Default,
    //      llvm::CodeGenOpt::Default,
    llvm::CodeGenOpt::Aggressive);
if (target_machine == nullptr) {
  error = "Failed to create target machine";
  return error;
}

llvm::legacy::PassManager pass_manager;

module.setDataLayout(target_machine->createDataLayout());

module.print(llvm::errs(), nullptr);

auto file = "out.o";
error_code ec;
llvm::raw_fd_ostream out(llvm::StringRef(file), ec, llvm::sys::fs::F_None);
if (ec) {
  error = ec.message();
  return error;
}

llvm::MachineModuleInfo* mmi = new llvm::MachineModuleInfo(
    reinterpret_cast<const llvm::LLVMTargetMachine*>(target_machine));
if (mmi == nullptr) {
  error = "Failed to create machine module info";
  return error;
}

if (target_machine->addPassesToEmitFile(pass_manager, out, llvm::TargetMachine::CGFT_ObjectFile)) {
  error = "Failed to emit file";
  return error;
}
pass_manager.run(module);
out.close();

return error;

LLVM无法优化简单函数

0 个答案: