Skip to content
A collection of LLVM passes (with tests and build scripts)
C++ LLVM C CMake Python
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Type Name Latest commit message Commit time
Failed to load latest commit information.


Build Status Build Status Build Status

Example LLVM passes - based on LLVM 9

llvm-tutor is a collection of self-contained reference LLVM passes. It's a tutorial that targets novice and aspiring LLVM developers. Key features:

  • Complete - includes CMake build scripts, LIT tests and CI set-up
  • Out of source - builds against a binary LLVM installation (no need to build LLVM from sources)
  • Modern - based on the latest version of LLVM (and updated with every release)

The source files contain comments that will guide you through the implementation and the LIT tests verify that each pass works as expected. This document explains how to get started.

Table of Contents


The HelloWorld pass from HelloWorld.cpp is a self-contained reference example. The corresponding CMakeLists.txt implements the minimum set-up for an out-of-source pass.

For each function in a module, HelloWord prints its name and the number of arguments that it takes. You can build it like this:

export LLVM_DIR=<installation/dir/of/llvm/9>
mkdir build
cd build
cmake -DLT_LLVM_INSTALL_DIR=$LLVM_DIR <source/dir/llvm/tutor>/HelloWorld/

Before you can test it, you need to prepare an input file:

# Generate an llvm test file
$LLVM_DIR/bin/clang -S -emit-llvm <source/dir/llvm/tutor/>inputs/input_for_hello.c -o input_for_hello.ll

Finally, run HelloWorld with opt:

# Run the pass on the llvm file
$LLVM_DIR/bin/opt -load-pass-plugin libHelloWorld.dylib -passes=hello-world -disable-output input_for_hello.ll
# The expected output
Visiting: foo (takes 1 args)
Visiting: bar (takes 2 args)
Visiting: fez (takes 3 args)
Visiting: main (takes 2 args)

The HelloWorld pass doesn't modify the input module. The -disable-output flag is used to prevent opt from printing the output bitcode file.

Development Environment

Platform Support And Requirements

This project has been tested on Linux 18.04 and Mac OS X 10.14.4. In order to build llvm-tutor you will need:

  • LLVM 9
  • C++ compiler that supports C++14
  • CMake 3.4.3 or higher

In order to run the passes, you will need:

  • clang-9 (to generate input LLVM files)
  • opt (to run the passes)

There are additional requirements for tests (these will be satisfied by installing LLVM-9):

  • lit (aka llvm-lit, LLVM tool for executing the tests)
  • FileCheck (LIT requirement, it's used to check whether tests generate the expected output)

Installing LLVM-9 on Mac OS X

On Darwin you can install LLVM 9 with Homebrew:

brew install llvm@9

This will install all the required header files, libraries and tools in /usr/local/opt/llvm/.

Installing LLVM-9 on Ubuntu

On Ubuntu Bionic, you can install modern LLVM from the official repository:

wget -O - | sudo apt-key add -
sudo apt-add-repository "deb llvm-toolchain-bionic-9 main"
sudo apt-get update
sudo apt-get install -y llvm-9 llvm-9-dev clang-9 llvm-9-tools

This will install all the required header files, libraries and tools in /usr/lib/llvm-9/.

Building LLVM-9 From Sources

Building from sources can be slow and tricky to debug. It is not necessary, but might be your preferred way of obtaining LLVM-9. The following steps will work on Linux and Mac OS X:

git clone
cd llvm-project
git checkout release/9.x
mkdir build
cd build
cmake -DCMAKE_BUILD_TYPE=Release -DLLVM_TARGETS_TO_BUILD=X86 <llvm-project/root/dir>/llvm/
cmake --build .

For more details read the official documentation.

Building & Testing

You can build llvm-tutor (and all the provided passes) as follows:

cd <build/dir>
cmake -DLT_LLVM_INSTALL_DIR=<installation/dir/of/llvm/9> <source/dir/llvm/tutor>

The LT_LLVM_INSTALL_DIR variable should be set to the root of either the installation or build directory of LLVM 9. It is used to locate the corresponding LLVMConfig.cmake script that is used to set the include and library paths.

In order to run the tests, you need to install llvm-lit (aka lit). It's not bundled with LLVM 9 packages, but you can install it with pip:

# Install lit - note that this installs lit globally
pip install lit

Running the tests is as simple as:

$ lit <build_dir>/test

Voilà! You should see all tests passing.

Overview of The Passes

  • HelloWorld - prints the functions in the input module and prints the number of arguments for each
  • InjectFuncCall - instruments the input module by inserting calls to printf (instrumentation pass)
  • StaticCallCounter - counts direct function calls at compile time (only static calls, pure analysis pass)
  • DynamicCallCounter - counts direct function calls at run-time ( analysis through instrumentation)
  • MBASub - code transformation for integer sub instructions (transformation pass, parametrisable)
  • MBAAdd - code transformation for 8-bit integer add instructions (transformation pass, parametrisable)
  • RIV - finds reachable integer values for each basic block (analysis pass)
  • DuplicateBB - duplicates basic blocks, requires RIV analysis results (transformation pass, parametrisable)

Once you've built this project, you can experiment with every pass separately. It is assumed that you have clang and opt available in your PATH. All passes work with LLVM files. You can generate one like this:

export LLVM_DIR=<installation/dir/of/llvm/9>
# Textual form
$LLVM_DIR/bin/clang  -emit-llvm input.c -S -o out.ll
# Binary/bit-code form
$LLVM_DIR/bin/clang  -emit-llvm input.c -o out.bc

It doesn't matter whether you choose the binary (without -S) or textual form (with -S), but obviously the latter is more human-friendly. All passes, except for HelloWorld, are described below.

Inject Calls To Printf (InjectFuncCall)

For every function defined in the input module, InjectFuncCall will add (inject) the following call to printf:

printf("(llvm-tutor) Hello from: %s\n(llvm-tutor)   number of arguments: %d\n", FuncName, FuncNumArgs)

This call is added at the beginning of each function (i.e. before any other instruction). FuncName is the name of the function and FuncNumArgs is the number of arguments that the function takes.

This pass is a HelloWorld example for code instrumentation. You can test it as follows:

export LLVM_DIR=<installation/dir/of/llvm/9>
# Generate an LLVM file to analyze
$LLVM_DIR/bin/clang  -emit-llvm -c <source_dir>/inputs/input_for_inject.c -o input_for_inject.bc
# Run the pass through opt
$LLVM_DIR/bin/opt -load <build_dir>/lib/libInjectFuncCall.dylib -legacy-inject-func-call input_for_inject.bc -o instrumented.bin
$LLVM_DIR/bin/lli instrumented.bin 123

instrumented.bin will print the following output:

(llvm-tutor) Hello from: main
(llvm-tutor)   number of arguments: 2
(llvm-tutor) Hello from: foo
(llvm-tutor)   number of arguments: 1
(llvm-tutor) Hello from: bar
(llvm-tutor)   number of arguments: 2
(llvm-tutor) Hello from: foo
(llvm-tutor)   number of arguments: 1
(llvm-tutor) Hello from: fez
(llvm-tutor)   number of arguments: 3
(llvm-tutor) Hello from: bar
(llvm-tutor)   number of arguments: 2
(llvm-tutor) Hello from: foo
(llvm-tutor)   number of arguments: 1

As you can see the output is similar to what HelloWorld generates, but the implementation is completely different.

Count Compile Time Function Calls (StaticCallCounter)

StaticCallCounter will count the number of function calls in the input LLVM file that are visible during the compilation (i.e. if a function is called within a loop, that counts as one call). Only direct function calls are considered (TODO: Expand).

export LLVM_DIR=<installation/dir/of/llvm/9>
# Generate an LLVM file to analyze
$LLVM_DIR/bin/clang  -emit-llvm -c <source_dir>/inputs/input_for_cc.c -o input_for_cc.bc
# Run the pass through opt
$LLVM_DIR/bin/opt -load <build_dir>/lib/libStaticCallCounter.dylib -legacy-static-cc -analyze input_for_cc.bc

The static executable is a command line wrapper that allows you to run StaticCallCounter without the need for opt:

<build_dir>/bin/static input_for_cc.bc

Count Run-Time Function Calls (DynamicCallCounter)

DynamicCallCounter will count the number of run-time function calls. It does so by instrumenting the input LLVM file - it inserts call-counting instructions that are executed every time a function is called. This pass will only count calls to functions that are defined in the input module are counted.

Although the primary goal of this pass is to analyse function calls, it also modifies the input file. Therefore it is a transformation/instrumentation pass. You can test it with one of the provided examples, e.g.:

export LLVM_DIR=<installation/dir/of/llvm/9>
# Generate an LLVM file to analyze
$LLVM_DIR/bin/clang  -emit-llvm -c <source_dir>/inputs/input_for_cc.c -o input_for_cc.bc
# Instrument the input file
$LLVM_DIR/bin/opt -load <build_dir>/lib/libDynamicCallCounter.dylib -legacy-dynamic-cc input_for_cc.bc -o instrumented_bin
# Run the instrumented binary

You will see the following output:

LLVM-TUTOR: dynamic analysis results
NAME                 #N DIRECT CALLS
foo                  13
bar                  2
fez                  1
main                 1

If you are interested in a more introductory example for code instrumentation, you may want experiment with InjectFuncCall first.

Mixed Boolean Arithmetic Transformations

These passes implement mixed boolean arithmetic transformations. Similar transformation are often used in code obfuscation (you may also know them from Hacker's Delight) and are a great illustration of what and how LLVM passes can be used for.


The MBASub pass implements this rather basic expression:

a - b == (a + ~b) + 1

Basically, it replaces all instances of integer sub according to the above formula. The corresponding LIT tests verify that both the formula and that the implementation are correct. You can run this pass as follows:

export LLVM_DIR=<installation/dir/of/llvm/9>
$LLVM_DIR/bin/clang -emit-llvm -S inputs/input_for_mba_sub.c -o input_for_sub.ll
$LLVM_DIR/bin/opt -load <build_dir>/lib/ -legacy-mba-sub input_for_sub.ll -o out.ll


The MBAAdd pass implements a slightly more involved formula that is only valid for 8 bit integers:

a + b == (((a ^ b) + 2 * (a & b)) * 39 + 23) * 151 + 111

Similarly to MBASub, it replaces all instances of integer add according to the above identity, but only for 8-bit integers. The LIT tests verify that both the formula and the implementation are correct. You can run MBAAdd like this:

export LLVM_DIR=<installation/dir/of/llvm/9>
$LLVM_DIR/bin/clang -O1 -emit-llvm -S inputs/input_for_mba.c -o input_for_mba.ll
$LLVM_DIR/bin/opt -load <build_dir>/lib/ -legacy-mba-add input_for_mba.ll -o out.ll

You can also specify the level of obfuscation on a scale of 0.0 to 1.0, with 0 corresponding to no obfuscation and 1 meaning that all add instructions are to be replaced with (((a ^ b) + 2 * (a & b)) * 39 + 23) * 151 + 111, e.g.:

$LLVM_DIR/bin/opt -load <build_dir>/lib/ -legacy-mba-add -mba-ratio=0.3 inputs/input_for_mba.c -o out.ll

Reachable Integer Values (RIV)

For each basic block in a module, RIV calculates the reachable integer values (i.e. values that can be used in the particular basic block). There are a few LIT tests that verify that indeed this is correct. You can run this pass as follows:

export LLVM_DIR=<installation/dir/of/llvm/9>
$LLVM_DIR/bin/opt -load <build_dir>/lib/ -riv inputs/input_for_riv.c

Note that this pass, unlike previous passes, will produce information only about the IR representation of the original module. It won't be very useful if trying to understand the original C or C++ input file.

Duplicate Basic Blocks (DuplicateBB)

This pass will duplicate all basic blocks in a module, with the exception of basic blocks for which there are no reachable integer values (identified through the RIV pass). An example of such a basic block is the entry block in a function that:

  • takes no arguments and
  • is embedded in a module that defines no global values.

This pass depends on the RIV pass, hence you need to load it too in order for DuplicateBB to work:

export LLVM_DIR=<installation/dir/of/llvm/9>
$LLVM_DIR/bin/opt -load <build_dir>/lib/ -load <build_dir>/lib/ -riv inputs/input_for_duplicate_bb.c

Basic blocks are duplicated by inserting an if-then-else construct and cloning all the instructions (with the exception of PHI nodes) into the new blocks.


Before running a debugger, you may want to analyze the output from LLVM_DEBUG and STATISTIC macros. For example, for MBAAdd:

export LLVM_DIR=<installation/dir/of/llvm/9>
$LLVM_DIR/bin/clang -emit-llvm -S -O1 inputs/input_for_mba.c -o input_for_mba.ll
$LLVM_DIR/bin/opt -load-pass-plugin <build_dir>/lib/libMBAAdd.dylib -passes=mba-add input_for_mba.ll -debug-only=mba-add -stats -o out.ll

Note the -debug-only=mba-add and -stats flags in the command line - that's what enables the following output:

  %12 = add i8 %1, %0 ->   <badref> = add i8 111, %11
  %20 = add i8 %12, %2 ->   <badref> = add i8 111, %19
  %28 = add i8 %20, %3 ->   <badref> = add i8 111, %27
                          ... Statistics Collected ...

3 mba-add - The # of substituted instructions

As you can see, you get a nice summary from MBAAdd. In many cases this will be sufficient to understand what might be going wrong.

For tricker issues just use a debugger. Below I demonstrate how to debug MBAAdd. More specifically, how to set up a breakpoint on entry to MBAAdd::run. Hopefully that will be sufficient for you to start.

Mac OS X

The default debugger on OS X is LLDB. You will normally use it like this:

export LLVM_DIR=<installation/dir/of/llvm/9>
$LLVM_DIR/bin/clang -emit-llvm -S -O1 inputs/input_for_mba.c -o input_for_mba.ll
lldb -- $LLVM_DIR/bin/opt -load-pass-plugin <build_dir>/lib/libMBAAdd.dylib -passes=mba-add input_for_mba.ll -o out.ll
(lldb) breakpoint set --name MBAAdd::run
(lldb) process launch

or, equivalently, by using LLDBs aliases:

export LLVM_DIR=<installation/dir/of/llvm/9>
$LLVM_DIR/bin/clang -emit-llvm -S -O1 inputs/input_for_mba.c -o input_for_mba.ll
lldb -- $LLVM_DIR/bin/opt -load-pass-plugin <build_dir>/lib/libMBAAdd.dylib -passes=mba-add input_for_mba.ll -o out.ll
(lldb) b MBAAdd::run
(lldb) r

At this point, LLDB should break at the entry to MBAAdd::run.


On most Linux systems, GDB is the most popular debugger. A typical session will look like this:

export LLVM_DIR=<installation/dir/of/llvm/9>
$LLVM_DIR/bin/clang -emit-llvm -S -O1 inputs/input_for_mba.c -o input_for_mba.ll
gdb --args $LLVM_DIR/bin/opt -load-pass-plugin <build_dir>/lib/ -passes=mba-add input_for_mba.ll -o out.ll
(gdb) b MBAAdd.cpp:MBAAdd::run
(gdb) r

At this point, GDB should break at the entry to MBAAdd::run.

About Pass Managers in LLVM

LLVM is a quite complex project (to put it mildly) and passes lay at its center - this is true for any multi-pass compiler. In order to manage the passes, a compiler needs a pass manager. LLVM currently enjoys not one, but two pass manager. This is important, because depending on which pass manager you decide to use, the implementation (and in particular pass registration) will look slightly differently. I have tried my best to make the distinction in the source code very clear.

Overview of Pass Managers in LLVM

As mentioned earlier, there are two pass managers in LLVM:

  • Legacy Pass Manager which currently is the default pass manager
    • It is implemented in the legacy namespace
    • It is very well documented (more specifically, writing and registering a pass withing the Legacy PM is very well documented)
  • New Pass Manager aka Pass Manager (that's how it's referred to in the code base)
    • I understand that it is soon to become the default pass manager in LLVM
    • The source code is very throughly commented, but otherwise I am only aware of this great blog series by Min-Yih Hsu.

The best approach is to implement your passes for both pass managers. Fortunately, once you have an implementation that works for one of them, it's relatively straightforward to extend it so that it works for the other one as well. All passes in LLVM provide an interface for both and that's what I've been trying to achieve here as well.

New vs Legacy PM When Running Opt

MBAAdd implements interface for both pass managers. This is how you will use it via the legacy pass manager:

$LLVM_DIR/bin/opt -load <build_dir>/lib/ -legacy-mba-add input_for_mba.ll -o out.ll

And this is how you run it with the new pass manager:

$LLVM_DIR/bin/opt -load-pass-plugin <build_dir>/lib/ -passes=mba-add input_for_mba.ll -o out.ll

There are two differences:

  • the way you load your plugins: -load vs -load-pass-plugin
  • the way you specify which pass/plugin to run: -legacy-mba-add vs -passes=mba-add

The command line option is different because with the legacy pass manager you register a new command line option with opt and with the new pass manager you just define the pass pipeline (via -passes=).


This is first and foremost a community effort. This project wouldn't be possible without the amazing LLVM online documentation, the plethora of great comments in the source code, and the llvm-dev mailing list. Thank you!

It goes without saying that there's plenty of great presentations on YouTube, blog posts and GitHub projects that cover similar subjects. I've learnt a great deal from them - thank you all for sharing! There's one presentation/tutorial that has been particularly important in my journey as an aspiring LLVM developer and that helped to democratise out-of-source pass development:

  • "Building, Testing and Debugging a Simple out-of-tree LLVM Pass" Serge Guelton, Adrien Guinet (slides, video)

Adrien and Serge came up with some great, illustrative and self-contained examples that are great for learning and tutoring LLVM pass development. You'll notice that there are similar transformation and analysis passes available in this project. The implementations available here reflect what I (aka banach-space) found most challenging while studying them.

I also want to thank Min-Yih Hsu for his blog series "Writing LLVM Pass in 2018". It was invaluable in understanding how the new pass manager works and how to use it. Last, but not least I am very grateful to Nick Sunmer (e.g. llvm-demo) and Mike Shah (see Mike's Fosdem 2018 talk) for sharing their knowledge online. I have learnt a great deal from it, thank you! I always look-up to those of us brave and bright enough to work in academia - thank you for driving the education and research forward!


Below is a list of LLVM resources available outside the official online documentation that I have found very helpful. Where possible, the items are sorted by date.

    • ”LLVM IR Tutorial-Phis,GEPs and other things, ohmy!”, V.Bridgers, F. Piovezan, EuroLLVM, (slides, video)
    • "Mapping High Level Constructs to LLVM IR", M. Rodler (link)
  • Legacy vs New Pass Manager
    • "New PM: taming a custom pipeline of Falcon JIT", F. Sergeev, EuroLLVM 2018 (slides, video)
    • "The LLVM Pass Manager Part 2", Ch. Carruth, LLVM Dev Meeting 2014 (slides, video)a
    • ”Passes in LLVM, Part 1”, Ch. Carruth, EuroLLVM 2014 (slides, video)
  • Examples in LLVM
commit 7d0b1d77b3d4d47df477519fd1bf099b3df6f899
Author: Florian Hahn <>
Date:   Tue Nov 12 14:06:12 2019 +0000

[Examples] Add IRTransformations directory to examples.
  • LLVM Pass Development
    • "Getting Started With LLVM: Basics ", J. Paquette, F. Hahn, LLVM Dev Meeting 2019 (not yet uploaded)
    • "Writing an LLVM Pass: 101", A. Warzyński, LLVM Dev Meeting 2019 (not yet uploaded)
    • "Writing LLVM Pass in 2018", Min-Yih Hsu, blog series
    • "Building, Testing and Debugging a Simple out-of-tree LLVM Pass" Serge Guelton, Adrien Guinet, LLVM Dev Meeting 2015 (slides, video)
  • LLVM Based Tools Development
    • "Introduction to LLVM", M. Shah, Fosdem 2018, link
    • llvm-demo, by N Sumner
    • "Building an LLVM-based tool. Lessons learned", A. Denisov, blog post, video


The MIT License (MIT)

Copyright (c) 2019 Andrzej Warzyński

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.


You can’t perform that action at this time.