Confused on how to do a machinefunction pass

Hi everyone,
I am a LLVM newbie. I need to write a machinefunction pass for my project. This should be an analyzer of the machine code to do some profiling. I have written a couple of function passes for front end, but it seems they are completely different. I searched through forum and llvm documents, but couldn’t find anything useful or questions were unanswered. Can anyone reference me to any material on how to do that or help me with my roblem.

I have created a folder in: lib/Transform and put my pass in it.
I am writing a simple pass like this:

using namespace llvm;

namespace {

struct analyzer : public MachineFunctionPass {
static char ID;
analyzer() : MachineFunctionPass(ID) {}

virtual bool runOnMachineFunction(MachineFunction &MF) {
errs() << "hello " ;

return false;
}
};
}
char analyzer::ID = 0;
static RegisterPass X(“analyzer”, “WAW analyzer”);

I make it successfully. But when I load it for my test code I get this error:

Pass ‘WAW analyzer’ is not initialized.
Verify if there is a pass dependency cycle.
Required Passes:
opt: /llvm/llvm-3.7/lib/IR/LegacyPassManager.cpp:635: void llvm::PMTopLevelManager::schedulePass(llvm::Pass*): Assertion `PI && “Expected required passes to be initialized”’ failed.

I don’t know how to solve it? Do we have to run machinefunction passes with OPT command? Is this correct to put machinefunction pass in an external folder or do we have to change some llvm built-in files?

Regards,

Fami

Dear Fami,

A MachineFunctionPass is run by the code generator, so you can only use it in tools like llc, clang, and libLTO. The opt program only manipulates LLVM IR and therefore does not run MachineFunctionPass'es.

Regarding documentation, have you looked at the following?

http://llvm.org/docs/WritingAnLLVMBackend.html
http://llvm.org/docs/CodeGenerator.html
http://llvm.org/docs/MIRLangRef.html

Regards,

John Criswell

Dear John,
Thank you so much for your help. I looked at those documents. Could you kindly answer the following questions:

Does it mean that I have to make my own backend target in order to write a machine pass even if I want to run a simple machinefunction pass? for example,if I want my pass to get MIPS instructions as an input, I have to copy all the files from mips target and add a file to it which implements my pass. and there should be a way to connect the pass to other MIPS backend files?

As a result, is the way that I described in my first post completely wrong because I cannot run a stand alone machine function pass like a frontend function pass?

Also another question, I’m a PhD student and I’m completely new to the llvm backend process. I want to estimate my project time. How long approximately do you think it will take to be get familiar with backend and be able to write machinefunction pass? for further steps I have to implement a register allocation algorithm.

Regards,
Fami

No, you do not need to create a new backend. All you need to do is to add your pass to the list of passes that are run when the MIPS code generator is used. In LLVM 3.3, there was a file in the X86 backend that had code to schedule all the MachineFunctionPass’es when the X86 code generator was used. That was in lib/Target/X86/X86TargetMachine.cpp. You can probably find a similar file for LLVM 3.7 for the MIPS backend. So, to summarize, you’ll add your source file to the MIPS backend, add a line somewhere to run your pass when the MIPS code generator is used, and then recompile llvm/lib and llvm/tools/llc. As far as I know, there is no way to load a MachineFunctionPass and run it using llc, opt, or clang. I believe you need to compile it into the MIPS backend code. It’s difficult to estimate (especially since I haven’t implemented anything as complicated as a register allocator). Your profiling pass may take anywhere from 1 week to 1 month depending on how complicated it is. I think a register allocator would take anywhere from 1 month up to 6 months depending on how robust you need it to be, but since I’ve never built one, I can’t say for certain. Regards, John Criswell

Thank you so much.
That helped alot.

Fami

Hi
So I m trying to write my first backend pass. I created a file in target\x86\wawanalyzer.cpp based on hexagonHardwareLoops.cpp file. Now it’s just a simple file that prints hello to the output. I have attached the file to this post.

Then I did the following:

  1. add “FunctionPass *createwawAnalyzer();” line to x86.h
  2. add file name to CMakelist.txt.
  3. This is going to run after postRAscheduler and before code emission so I changed the x86TrgetMachine.cpp as follow:

void X86PassConfig::addPreEmitPass() {
/////////////add mypass here///////
addPass(createwawAnalyzer());
//////////////////////////////////

if (getOptLevel() != CodeGenOpt::None)
addPass(createExecutionDependencyFixPass(&X86::VR128RegClass));

if (UseVZeroUpper)
addPass(createX86IssueVZeroUpperPass());

if (getOptLevel() != CodeGenOpt::None) {
addPass(createX86PadShortFunctions());
addPass(createX86FixupLEAs());
}

then I run make on mypass and remake llvm3.7\lib and llvm3.7\tool\llc.
but when I run llc -help it doesn’t show my pass. Did I miss some steps or did something wrong? What’s the correct way to run the pass with llc command?

I appreciate any help,
Regards
Fami

wawanalyzer.cpp (1.96 KB)

You cannot run individual passes from llc. Llc -help only shows the available command-line options.

Try compiling some .ll file with the llc. You should see the message that your pass it printing.

-Krzysztof

Hi,
So, I run my pass in X86 target with llc command and it printed out “hello****”. Now I am trying to do the same pass for ARM target. So I did exactly what I did for X86 as mentioned in my previous posts. When I run the following command:
llc -march=arm test.ll -o test
nothing prints out. I did the same for MIPS target too and I got no result. Can anyone tell me what I’m doing wrong. Is there any difference between writing machinefunction passes in x86 and other targets?
Thank you,
Fami

Have you modified the ARM code generator to run your pass (the same way that you modified the X86 code generator to run your pass)? Each backend has code that configures the set of MachineFunctionPass’es to run when that code generator is used. For each backend, you must modify that code to run your MachineFunctionPass. Regards, John Criswell

Yes, I have done exactly the same. The wawanalyzer is the same. I changed ARM.h and ARMTargetMachine.cpp in the tager/arm folder. then I make tool/llc and lib folder.

I noticed something else, I just made a change to my x86 wawanalyzer and make it again. And I can’t see the change in my output. So the problem is with all my targets.
After I ran my first sunctionpass in x86, I needed to crosscopmpile LLVM so I ran the following instructions a couple of days ago:

$ cd <llvm_build_dir>
$ <PATH_TO_SOURCE>/configure --enable-targets=arm --disable-optimized
–prefix=/usr/local/llvm-arm --target=armv7a-unknown-linux-gnueabi
$ make && sudo make install

Does my problem have to do something with these instructions?

I fixed the problem. I have to make clean the target folder everytime, then make it again.