llvm中如何利用分支概率和基本块频率估计

1. 背景

llvm自2.9版以后，已经集成了对分支概率和基本块频率的静态分析。

分支概率（branch probability）是指在程序的控制流图中，从控制流从一个基本块A到其任意后继基本块Si的概率。控制流从基本块A到其所有后继基本块的概率之和为1. 基本块频率（block frequency）是指在程序的控制流图中，任意基本块的执行次数。这两种信息都可以通过静态分析得到。其原理如下【1】【2】：

An alternative is static profiling, in which a compiler estimates execution frequencies (not absolute counts) with static program analysis. A static profile eliminates the drawbacks of dynamic profiling— if it accurately captures a program’s dynamic behavior. Recent work suggests that static analysis can predict dynamic program behavior. Fisher and Freudenberger [Fisher-92] observed that many programs’ dynamic branching behavior is independent of their input data. Ball and Larus developed a simple algorithm that statically predicts the outcome of a conditional branch with good accuracy [Ball-93]. Wagner et al. usedl simple estimates of branch probabilities to compute static profiles [Wagner-94]. （见文献【2】）

2. llvm3.3中的相关文件

Support/BranchProbability.cpp(.h): 实现一个用来表示分支概率的数据结构
Analysis/BranchProbabilityInfo.cpp(.h): 实现一个在basic block级别进行分支概率估计的FunctionPass
CodeGen/MachineBranchProbabilityInfo.cpp(.h): 实现一个在machine basic block级别进行分支概率估计的ImmutablePass

Support/BlockFrequency.cpp(.h): 实现一个用来表示基本块频率的数据结构
Analysis/BlockFrequencyInfo.cpp(.h): 实现一个在basic block级别进行基本块频率估计的FunctionPass
CodeGen/MachineBlockFrequency.cpp(.h): 实现一个在machine basic block级别进行基本块频率估计的MachineFunctionPass
Analysis/BlockFrequencyImpl.h: 在basic block级别和machine basic block级别共用的基本块频率估计的实现

3. llvm3.3中的相关实现

3.1 分支概率分析

3.1.1 在basic block级别，分支概率分析的实现主要参考文献【2】的方法，利用几个基本启发式来给分支加权。

for (po_iterator<BasicBlock *> I = po_begin(&F.getEntryBlock()), E = po_end(&F.getEntryBlock()); I != E; ++I) {
DEBUG(dbgs() << "Computing probabilities for " << I->getName() << "\n");
if (calcUnreachableHeuristics(*I))
continue;
if (calcMetadataWeights(*I))
continue;
if (calcColdCallHeuristics(*I))
continue;
if (calcLoopBranchHeuristics(*I))
continue;
if (calcPointerHeuristics(*I))
continue;
if (calcZeroHeuristics(*I))
continue;
if (calcFloatingPointHeuristics(*I))
continue;
calcInvokeHeuristics(*I);
}
return false;
}

3.1.2 在machine basic block级别，分支概率的实现实际上依赖于basic block级别的分支概率分析结果，所以MachineBranchProbabilityInfo并不是一个独立的MachineFunctionPass.

3.2 基本块频率分析

3.2.1 在basic block级别和machine basic block级别共用基本块频率估计的实现

bool BlockFrequencyInfo::runOnFunction(Function &F) {
BranchProbabilityInfo &BPI = getAnalysis<BranchProbabilityInfo>();
BFI->doFunction(&F, &BPI);
return false;
}

bool MachineBlockFrequencyInfo::runOnMachineFunction(MachineFunction &F) {
MachineBranchProbabilityInfo &MBPI = getAnalysis<MachineBranchProbabilityInfo>();
MBFI->doFunction(&F, &MBPI);
return false;
}

3.2.2 上面的代码还可以看出，基本块频率分析依赖于分支概率分析。因此，如果要利用这两种分析结果，只需要在自己的FunctionPass或者MachineFunctionPass里面进行类似如下的修改（建议参考CodeGen/MachineBlockPlacement.cpp）：

1）修改getAnalysisUsage函数如下：

void getAnalysisUsage(AnalysisUsage &AU) const {
AU.addRequired<MachineBranchProbabilityInfo>();
AU.addRequired<MachineBlockFrequencyInfo>();
MachineFunctionPass::getAnalysisUsage(AU);
}

2）修改runOnMachineFunction函数如下

bool MachineBlockPlacement::runOnMachineFunction(MachineFunction &F) {
MBPI = &getAnalysis<MachineBranchProbabilityInfo>();
MBFI = &getAnalysis<MachineBlockFrequencyInfo>();
...
//打印分支概率信息
std::string szInfo;
raw_fd_ostream S("machinBranchProbs.txt", szInfo, raw_fd_ostream::F_Append);
for (MachineFunction::iterator BI = F.begin(), BE = F.end(); BI != BE; ++BI) {
MachineBasicBlock *MBB = BI;
for (MachineBasicBlock::const_succ_iterator SI =MBB->succ_begin(), EI = MBB->succ_end(); SI != EI; ++SI) {
MachineBasicBlock *mBlock = *SI;
MBPI->printEdgeProbability(S << "  ", MBB, mBlock);
}
}
S.close();
//打印基本块频率信息
raw_fd_ostream S1("machinBlockFreq.txt", szInfo, raw_fd_ostream::F_Append);
if (MBFI) MBFI->print(S1);;
S1.close();
...
return false;
}

4. 参考文献：

【1】. Hashemi, A., Kaeli, D., Calder, B.: Procedure mapping using static call graph estimation. In: Workshop on Interaction between Compiler and Computer Architecture, San Antonio, TX (1997)

【2】. Youfeng Wu, James R. Larus: Static branch frequency and program profile analysis. MICRO 1994: 1-11