问题
I've written a standard Analysis pass in LLVM, by extending the FunctionPass class. Everything seems to make sense.
Now what I'd like to do is write a couple of intermodular passes, that is, passes that allows me to analyze more than one module at a time. The purpose of one such pass is to construct a call graph of the entire application. The purpose of the other such pass is that I have an idea for an optimization involving function calls and their parameters.
I know about interprocedural passes in LLVM, via extending the ModulePass class, but that only allows analysis within a single module.
I know about link time optimization (LTO) in LLVM, but (a) I'm not quite clear if this is what I want and (b) I've found no examples or documentation on how to actually write an LTO pass.
How can I write an intermodular pass, i.e., a pass that has access to all the modules in an application, in LLVM?
回答1:
I found one way to achieve my goal: write a simple program that uses llvm::parseBitcodeFile() to read in a bitcode file and create a Module object that can be traversed and analyzed. It's not ideal, because it's not a Pass that can be run within the LLVM framework. However, it is a way to achieve my goal of analyzing multiple modules at once.
For future readers, here's what I did.
Create a simple tool to read in a bitcode file and produce a Module
//ReadBitcode.cpp
#include <iostream>
#include "llvm/IR/Module.h"
#include "llvm/Support/MemoryBuffer.h"
#include "llvm/Support/SourceMgr.h"
#include "llvm/IR/LLVMContext.h"
#include "llvm/Bitcode/ReaderWriter.h"
using namespace llvm;
int main(int argc, char *argv[])
{
if (argc != 2)
{
std::cerr << "Usage: " << argv[0] << " bitcode_filename" << std::endl;
return 1;
}
StringRef filename = argv[1];
LLVMContext context;
ErrorOr<std::unique_ptr<MemoryBuffer>> fileOrErr = MemoryBuffer::getFileOrSTDIN(filename);
if (std::error_code ec = fileOrErr.getError())
{
std::cerr << "Error opening input file: " + ec.message() << std::endl;
return 2;
}
ErrorOr<llvm::Module *> moduleOrErr = parseBitcodeFile(fileOrErr.get()->getMemBufferRef(), context);
if (std::error_code ec = fileOrErr.getError())
{
std::cerr << "Error reading Module: " + ec.message() << std::endl;
return 3;
}
Module *m = moduleOrErr.get();
std::cout << "Successfully read Module:" << std::endl;
std::cout << " Name: " << m->getName().str() << std::endl;
std::cout << " Target triple: " << m->getTargetTriple() << std::endl;
for (auto iter1 = m->getFunctionList().begin(); iter1 != m->getFunctionList().end(); iter1++)
{
Function &f = *iter1;
std::cout << " Function: " << f.getName().str() << std::endl;
for (auto iter2 = f.getBasicBlockList().begin(); iter2 != f.getBasicBlockList().end();
iter2++)
{
BasicBlock &bb = *iter2;
std::cout << " BasicBlock: " << bb.getName().str() << std::endl;
for (auto iter3 = bb.begin(); iter3 != bb.end(); iter3++)
{
Instruction &i = *iter3;
std::cout << " Instruction: " << i.getOpcodeName() << std::endl;
}
}
}
return 0;
}
Compile the tool
$ clang++ ReadBitcode.cpp -o reader `llvm-config --cxxflags --libs --ldflags --system-libs`
Create a bitcode file to analyze
$ cat foo.c
int my_fun(int arg1){
int x = arg1;
return x+1;
}
int main(){
int a = 11;
int b = 22;
int c = 33;
int d = 44;
if (a > 10){
b = c;
} else {
b = my_fun(d);
}
return b;
}
$ clang -emit-llvm -o foo.bc -c foo.c
Run the reader tool on the bitcode
$ ./reader foo.bc
Successfully read Module:
Name: foo.bc
Target triple: x86_64-pc-linux-gnu
Function: my_fun
BasicBlock:
Instruction: alloca
Instruction: alloca
Instruction: store
Instruction: load
Instruction: store
Instruction: load
Instruction: add
Instruction: ret
Function: main
BasicBlock:
Instruction: alloca
Instruction: alloca
Instruction: alloca
Instruction: alloca
Instruction: alloca
Instruction: store
Instruction: store
Instruction: store
Instruction: store
Instruction: store
Instruction: load
Instruction: icmp
Instruction: br
BasicBlock:
Instruction: load
Instruction: store
Instruction: br
BasicBlock:
Instruction: load
Instruction: call
Instruction: store
Instruction: br
BasicBlock:
Instruction: load
Instruction: ret
回答2:
This can be done using a module pass. Below is my code, and if you need help running it you can look here.
bar.c
int your_fun(int arg2) {
int x = arg2;
return x+2;
}
Skeleton.cpp
#include "llvm/Pass.h"
#include "llvm/IR/Module.h"
#include "llvm/Support/raw_ostream.h"
#include "llvm/IR/LegacyPassManager.h"
#include "llvm/Transforms/IPO/PassManagerBuilder.h"
using namespace llvm;
namespace {
struct SkeletonPass : public ModulePass {
static char ID;
SkeletonPass() : ModulePass(ID) {}
virtual bool runOnModule(Module &M) {
for (auto& F : M) {
errs() << "\tFunction: " << F.getName() << "\n";
for (auto& BB : F) {
errs() << "\t\tBasic Block: " << BB.getName() << "\n";
for (auto& I : BB) {
errs() << "\t\t\tInstruction: " << I.getOpcodeName() << "\n";
}
}
}
return false;
}
};
}
char SkeletonPass::ID = 0;
// Automatically enable the pass.
// http://adriansampson.net/blog/clangpass.html
static void registerSkeletonPass(const PassManagerBuilder &,
legacy::PassManagerBase &PM) {
PM.add(new SkeletonPass());
}
static RegisterStandardPasses RegisterMyPass(PassManagerBuilder::EP_ModuleOptimizerEarly,
registerSkeletonPass);
static RegisterStandardPasses RegisterMyPass1(PassManagerBuilder::EP_EnabledOnOptLevel0,
registerSkeletonPass);
Output:
| => clang -Xclang -load -Xclang build/skeleton/libSkeletonPass.so foo.c bar.c
Module: foo.c!
Function: my_fun!
Basicblock: entry!
Instruction: alloca
Instruction: alloca
Instruction: store
Instruction: load
Instruction: store
Instruction: load
Instruction: add
Instruction: ret
Function: main!
Basicblock: entry!
Instruction: alloca
Instruction: alloca
Instruction: alloca
Instruction: alloca
Instruction: alloca
Instruction: store
Instruction: store
Instruction: store
Instruction: store
Instruction: store
Instruction: load
Instruction: icmp
Instruction: br
Basicblock: if.then!
Instruction: load
Instruction: store
Instruction: br
Basicblock: if.else!
Instruction: load
Instruction: call
Instruction: store
Instruction: br
Basicblock: if.end!
Instruction: load
Instruction: ret
Module: bar.c!
Function: your_fun!
Basicblock: entry!
Instruction: alloca
Instruction: alloca
Instruction: store
Instruction: load
Instruction: store
Instruction: load
Instruction: add
Instruction: ret
Output: If you include header file linking to bar.c
Module: foo.c!
Function: your_fun!
Basicblock: entry!
Instruction: alloca
Instruction: alloca
Instruction: store
Instruction: load
Instruction: store
Instruction: load
Instruction: add
Instruction: ret
Function: my_fun!
Basicblock: entry!
Instruction: alloca
Instruction: alloca
Instruction: store
Instruction: load
Instruction: store
Instruction: load
Instruction: add
Instruction: ret
Function: main!
Basicblock: entry!
Instruction: alloca
Instruction: alloca
Instruction: alloca
Instruction: alloca
Instruction: alloca
Instruction: store
Instruction: store
Instruction: store
Instruction: store
Instruction: store
Instruction: load
Instruction: icmp
Instruction: br
Basicblock: if.then!
Instruction: load
Instruction: store
Instruction: br
Basicblock: if.else!
Instruction: load
Instruction: call
Instruction: store
Instruction: load
Instruction: call
Instruction: store
Instruction: br
Basicblock: if.end!
Instruction: load
Instruction: ret
回答3:
In LTO all the modules are combined and you can see the whole program IR in one module.
You need to write a module pass like any module pass and add it to the list of LTO passes in populateLTOPassManager function in PassManagerBuilder.cpp. Here is the doc for PassManagerBuilder: http://llvm.org/docs/doxygen/html/classllvm_1_1PassManagerBuilder.html
When you do this, your pass will be executed with other LTO passes.
来源：https://stackoverflow.com/questions/30198277/how-to-write-a-custom-intermodular-pass-in-llvm