Example LLVM passes - based on LLVM 10
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
CMakebuild 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)
LLVM implements a very rich, powerful and popular API. However, like many complex technologies, it can be quite daunting and overwhelming to learn and master. The goal of this LLVM tutorial is to showcase that LLVM can in fact be easy and fun to work with. This is demonstrated through a range self-contained, testable LLVM passes, which are implemented using idiomatic LLVM.
This document explains how to set-up your environment, build and run the examples, and go about debugging. It contains a high-level overview of the implemented examples and contains some background information on writing LLVM passes. The source files, apart from the code itself, contain comments that will guide you through the implementation. All examples are complemented with LIT tests and reference input files.
- HelloWorld
- Development Environment
- Building & Testing
- Overview of the Passes
- Debugging
- About Pass Managers in LLVM
- Credits & References
- License
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 every function defined in the input 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/10>
mkdir build
cd build
cmake -DLT_LLVM_INSTALL_DIR=$LLVM_DIR <source/dir/llvm/tutor>/HelloWorld/
makeBefore 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.llFinally, run HelloWorld with opt:
# Run the pass
$LLVM_DIR/bin/opt -load-pass-plugin libHelloWorld.dylib -passes=hello-world -disable-output input_for_hello.ll
# Expected output
(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: fez
(llvm-tutor) number of arguments: 3
(llvm-tutor) Hello from: main
(llvm-tutor) number of arguments: 2The HelloWorld pass doesn't modify the input module. The -disable-output
flag is used to prevent opt from printing the output bitcode file.
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 10
- C++ compiler that supports C++14
- CMake 3.4.3 or higher
In order to run the passes, you will need:
- clang-10 (to generate input LLVM files)
- opt (to run the passes)
There are additional requirements for tests (these will be satisfied by installing LLVM 10):
- lit (aka llvm-lit, LLVM tool for executing the tests)
- FileCheck (LIT requirement, it's used to check whether tests generate the expected output)
On Darwin you can install LLVM 10 with Homebrew:
brew install llvm@10If you already have an older version of LLVM installed, you can upgrade it to LLVM 10 like this:
brew upgrade llvmOnce the installation (or upgrade) is complete, all the required header files,
libraries and tools will be located in /usr/local/opt/llvm/.
If your default linker is older then the one used to build LLVM 10, you may
need to pass -mlinker-version=0 to clang-10 for it to work. Otherwise you
may see errors like this:
ld: unknown option: -platform_version
clang-10: error: linker command failed with exit code 1 (use -v to see invocation)This is a known issue
that will only affect you if installing with brew.
On Ubuntu Bionic, you can install modern LLVM from the official repository:
wget -O - https://apt.llvm.org/llvm-snapshot.gpg.key | sudo apt-key add -
sudo apt-add-repository "deb http://apt.llvm.org/bionic/ llvm-toolchain-bionic-10 main"
sudo apt-get update
sudo apt-get install -y llvm-10 llvm-10-dev clang-10 llvm-10-toolsThis will install all the required header files, libraries and tools in
/usr/lib/llvm-10/.
Building from sources can be slow and tricky to debug. It is not necessary, but might be your preferred way of obtaining LLVM 10. The following steps will work on Linux and Mac OS X:
git clone https://github.com/llvm/llvm-project.git
cd llvm-project
git checkout release/10.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.
You can build llvm-tutor (and all the provided passes) as follows:
cd <build/dir>
cmake -DLT_LLVM_INSTALL_DIR=<installation/dir/of/llvm/10> <source/dir/llvm/tutor>
makeThe LT_LLVM_INSTALL_DIR variable should be set to the root of either the
installation or build directory of LLVM 10. 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 10 packages, but you can install it with pip:
# Install lit - note that this installs lit globally
pip install litRunning the tests is as simple as:
$ lit <build_dir>/testVoilà! You should see all tests passing.
- HelloWorld - prints the name and the number of arguments for each function encounterd in the input module
- OpcodeCounter - prints the summary of the LLVM IR opcodes encountered in the input function
- InjectFuncCall - instruments
the input module by inserting calls to
printf - StaticCallCounter - counts direct function calls at compile-time
- DynamicCallCounter - counts direct function calls at run-time
- MBASub - code transformation for integer
subinstructions - MBAAdd - code transformation for 8-bit integer
addinstructions - RIV - finds reachable integer values for each basic block
- DuplicateBB - duplicates basic blocks, requires RIV analysis results
- MergeBB - merges duplicated basic blocks
Once you've built this project, you can experiment with every pass separately. All passes work with LLVM files. You can generate one like this:
export LLVM_DIR=<installation/dir/of/llvm/10>
# 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.bcIt 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.
OpcodeCounter prints a summary of the LLVM IR opcodes encountered in every function in the input module. This pass is slightly more complicated than HelloWorld and it can be run automatically. Let's use our tried and tested method first.
We will use input_for_cc.c to test OpcodeCounter:
export LLVM_DIR=<installation/dir/of/llvm/10>
# 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/libOpcodeCounter.dylib -legacy-opcode-counter input_for_cc.bcFor main OpcodeCounter, prints the following summary (note that when running the pass,
a summary for other functions defined in input_for_cc.bc is also printed):
=================================================
LLVM-TUTOR: OpcodeCounter results for `main`
=================================================
OPCODE #N TIMES USED
-------------------------------------------------
load 2
br 4
icmp 1
add 1
ret 1
alloca 2
store 4
call 4
-------------------------------------------------
You can configure llvm-tutor so that OpcodeCounter is run automatically
at any optimisation level (i.e. -O{1|2|3|s}). This is achieved through
auto-registration with the existing pipelines, which is enabled with the
LT_OPT_PIPELINE_REG CMake variable. Simply add -DLT_OPT_PIPELINE_REG=On
when building llvm-tutor.
Once built, you can run OpcodeCounter by specifying an optimisation level. Note that you still have to specify the plugin file to be loaded:
$LLVM_DIR/bin/opt -load <build_dir>/lib/libOpcodeCounter.dylib -O1 input_for_cc.bcHere I used the Legacy Pass Manager (the plugin file was specified with
-load rather than -load-pass-plugin), but the auto-registration also works with
the New Pass Manager:
$LLVM_DIR/bin/opt -load-pass-plugin <build_dir>/lib/libOpcodeCounter.dylib --passes='default<O1>' input_for_cc.bcThis is implemented in OpcodeCounter.cpp, on line 82 for the New PM, and on line 111 for the Legacy PM. This section contains more information about the pass managers in LLVM.
This pass is a HelloWorld example for code instrumentation. 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.
We will use input_for_hello.c to test InjectFuncCall:
export LLVM_DIR=<installation/dir/of/llvm/10>
# Generate an LLVM file to analyze
$LLVM_DIR/bin/clang -emit-llvm -c <source_dir>/inputs/input_for_hello.c -o input_for_hello.bc
# Run the pass through opt
$LLVM_DIR/bin/opt -load <build_dir>/lib/libInjectFuncCall.dylib -legacy-inject-func-call input_for_hello.bc -o instrumented.binThis generates instrumented.bin, which is the instrumented version of
input_for_hello.bc. In order to verify that InjectFuncCall worked as
expected, you can either check the output file (and verify that it contains
extra calls to printf) or run it:
$LLVM_DIR/bin/lli instrumented.bin
(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
You might have noticed that InjectFuncCall is somewhat similar to
HelloWorld. In both cases the pass visits all functions,
prints their names and the number of arguments. The difference between the two
passes becomes quite apparent when you compare the output generated for the same
input file, e.g. input_for_hello.c. The number of times Hello from is
printed is either:
- once per every function call in the case of InjectFuncCall, or
- once per function definition in the case of HelloWorld.
This makes perfect sense and hints how different the two passes are. Whether to
print Hello from is determined at either:
- run-time for InjectFuncCall, or
- compile-time for HelloWorld.
Also, note that in the case of InjectFuncCall we had to first run the pass with opt and then execute the instrumented IR module in order to see the output. For HelloWorld it was sufficient to run run the pass with opt.
The StaticCallCounter pass counts the number of static function calls in the input LLVM module. Static refers to the fact that these function calls are compile-time calls (i.e. visible during the compilation). This is in contrast to dynamic function calls, i.e. function calls encountered at run-time (when the compiled module is run). The distinction becomes apparent when analysing functions calls within loops, e.g.:
for (i = 0; i < 10; i++)
foo();Although at run-time foo will be executed 10 times, StaticCallCounter
will report only 1 function call.
This pass will only consider direct functions calls. Functions calls via function pointers are not taken into account.
StaticCallCounter resembles OpcodeCounter - both passes analyse opcodes in the input module. However, only StaticCallCounter is a genuine analysis pass (in LLVM terms). Indeed, StaticCallCounter has the following additional properties:
- it implements the
printmethod (Legacy PM interface) - it inherits from
AnalysisInfoMixin(New PM interface)
We will use input_for_cc.c to test StaticCallCounter:
export LLVM_DIR=<installation/dir/of/llvm/10>
# 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.bcYou should see the following output:
=================================================
LLVM-TUTOR: static analysis results
=================================================
NAME #N DIRECT CALLS
-------------------------------------------------
foo 3
bar 2
fez 1
-------------------------------------------------
Note the extra command line option above: -analyze. This option is only
available when using the Legacy Pass Manager and is used to print the results
of the analysis pass that has just been run. It is enabled by implementing the
print method
mentioned earlier.
You can run StaticCallCounter through a standalone tool called static.
static is an LLVM based tool implemented in
StaticMain.cpp.
It is a command line wrapper that allows you to run StaticCallCounter
without the need for opt:
<build_dir>/bin/static input_for_cc.bcIt is an example of a relatively basic static analysis tool. Its implementation demonstrates how basic pass management in LLVM works (i.e. it handles that for itself instead of relying on opt).
The DynamicCallCounter pass counts the number of run-time (i.e. encountered during the execution) function calls. It does so by inserting call-counting instructions that are executed every time a function is called. Only calls to functions that are defined in the input module are counted. This pass builds on top of ideas presented in InjectFuncCall. You may want to experiment with that example first.
We will use input_for_cc.c to test DynamicCallCounter:
export LLVM_DIR=<installation/dir/of/llvm/10>
# 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_binThis generates instrumented.bin, which is the instrumented version of
input_for_cc.bc. In order to verify that DynamicCallCounter worked as
expected, you can either check the output file (and verify that it contains
new call-counting instructions) or run it:
# Run the instrumented binary
$LLVM_DIR/bin/lli ./instrumented_binYou will see the following output:
=================================================
LLVM-TUTOR: dynamic analysis results
=================================================
NAME #N DIRECT CALLS
-------------------------------------------------
foo 13
bar 2
fez 1
main 1
The number of function calls reported by DynamicCallCounter and StaticCallCounter are different, but both results are correct. They correspond to run-time and compile-time function calls respectively. Note also that for StaticCallCounter it was sufficient to run the pass through opt to have the summary printed. For DynamicCallCounter we had to run the instrumented binary to see the output. This is similar to what we observed when comparing HelloWorld and InjectFuncCall.
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.
We will use input_for_mba_sub.c to test MBASub:
export LLVM_DIR=<installation/dir/of/llvm/10>
$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/libMBASub.so -legacy-mba-sub input_for_sub.ll -o out.llThe 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.
We will use input_for_add.c to test MBAAdd:
export LLVM_DIR=<installation/dir/of/llvm/10>
$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/libMBAAdd.so -legacy-mba-add input_for_mba.ll -o out.llYou 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/libMBAAdd.so -legacy-mba-add -mba-ratio=0.3 inputs/input_for_mba.c -o out.llRIV is an analysis pass that for each basic
block BB in
the input function computes the set reachable integer values, i.e. the integer
values that are visible (i.e. can be used) in BB. Since the pass operates on
the LLVM IR representation of the input file, it takes into account all values
that have integer type in
the LLVM IR sense. In particular, since
at the LLVM IR level booleans are represented as 1-bit wide integers (i.e.
i1), you will notice that booleans are also included in the result.
This pass demonstrates how to request results from other analysis passes in LLVM. In particular, it relies on the Dominator Tree analysis pass from LLVM, which is is used to obtain the dominance tree for the basic blocks in the input function.
We will use input_for_riv.c to test RIV:
export LLVM_DIR=<installation/dir/of/llvm/10>
$LLVM_DIR/bin/clang -emit-llvm -S -O1 inputs/input_for_riv.c -o input_for_riv.ll
$LLVM_DIR/bin/opt -load <build_dir>/lib/libRIV.so -legacy-riv inputs/input_for_riv.llYou will see the following output:
=================================================
LLVM-TUTOR: RIV analysis results
=================================================
BB id Reachable Ineger Values
-------------------------------------------------
BB %entry
i32 %a
i32 %b
i32 %c
BB %if.then
%add = add nsw i32 %a, 123
%cmp = icmp sgt i32 %a, 0
i32 %a
i32 %b
i32 %c
BB %if.end8
%add = add nsw i32 %a, 123
%cmp = icmp sgt i32 %a, 0
i32 %a
i32 %b
i32 %c
BB %if.then2
%mul = mul nsw i32 %b, %a
%div = sdiv i32 %b, %c
%cmp1 = icmp eq i32 %mul, %div
%add = add nsw i32 %a, 123
%cmp = icmp sgt i32 %a, 0
i32 %a
i32 %b
i32 %c
BB %if.else
%mul = mul nsw i32 %b, %a
%div = sdiv i32 %b, %c
%cmp1 = icmp eq i32 %mul, %div
%add = add nsw i32 %a, 123
%cmp = icmp sgt i32 %a, 0
i32 %a
i32 %b
i32 %c
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.
Basic blocks are duplicated by first inserting an if-then-else construct and
then cloning all the instructions from the original basic block (with the
exception of PHI
nodes) into two
new basic blocks (clones of the original basic block). The if-then-else
construct is introduced as a non-trivial mechanism that decides which of the
cloned basic blocks to branch to. This condition is equivalent to:
if (var == 0)
goto clone 1
else
goto clone 2in which:
varis a randomly picked variable from theRIVset for the current basic blockclone 1andclone 2are labels for the cloned basic blocks.
The complete transformation looks like this:
BEFORE: AFTER:
------- ------
[ if-then-else ]
DuplicateBB / \
[ BB ] ------------> [clone 1] [clone 2]
\ /
[ tail ]
LEGEND:
-------
[BB] - the original basic block
[if-then-else] - a new basic block that contains the if-then-else statement (inserted by DuplicateBB)
[clone 1|2] - two new basic blocks that are clones of BB (inserted by DuplicateBB)
[tail] - the new basic block that merges [clone 1] and [clone 2] (inserted by DuplicateBB)As depicted above, DuplicateBB replaces qualifying basic blocks with 4 new
basic blocks. This is implemented through LLVM's
SplitBlockAndInsertIfThenElse.
DuplicateBB does all the necessary preparation and clean-up. In other
words, it's an elaborate wrapper for LLVM's SplitBlockAndInsertIfThenElse.
This pass depends on the RIV pass, which also needs be loaded in order for DuplicateBB to work. Lets use input_for_duplicate_bb.c as our sample input. First, generate the LLVM file:
export LLVM_DIR=<installation/dir/of/llvm/10>
$LLVM_DIR/bin/clang -emit-llvm -S -O1 inputs/input_for_duplicate_bb.c -o input_for_duplicate_bb.llFunction foo in input_for_duplicate_bb.ll should look like this (all metadata has been stripped):
define i32 @foo(i32) {
ret i32 1
}Note that there's only one basic block (the entry block) and that foo takes
one argument (this means that the result from RIV will be a non-empty set).
We will now apply DuplicateBB to foo:
$LLVM_DIR/bin/opt -load <build_dir>/lib/libRIV.so -load <build_dir>/lib/libDuplicateBB.so -legacy-duplicate-bb input_for_duplicate_bb.llAfter the instrumentation foo will look like this (all metadata has been stripped):
define i32 @foo(i32) {
lt-if-then-else-0:
%2 = icmp eq i32 %0, 0
br i1 %2, label %lt-if-then-0, label %lt-else-0
clone-1-0:
br label %lt-tail-0
clone-2-0:
br label %lt-tail-0
lt-tail-0:
ret i32 1
}There are four basic blocks instead of one. All new basic blocks end with a
numeric id of the original basic block (0 in this case). lt-if-then-else-0
contains the new if-then-else condition. clone-1-0 and clone-2-0 are
clones of the original basic block in foo. lt-tail-0 is the extra basic
block that's required to merge clone-1-0 and clone-2-0.
MergeBB will merge qualifying basic blocks that are identical. To some extent, this pass reverts the transformations introduced by DuplicateBB. This is illustrated below:
BEFORE: AFTER DuplicateBB: AFTER MergeBB:
------- ------------------ --------------
[ if-then-else ] [ if-then-else* ]
DuplicateBB / \ MergeBB |
[ BB ] ------------> [clone 1] [clone 2] --------> [ clone ]
\ / |
[ tail ] [ tail* ]
LEGEND:
-------
[BB] - the original basic block
[if-then-else] - a new basic block that contains the if-then-else statement (**DuplicateBB**)
[clone 1|2] - two new basic blocks that are clones of BB (**DuplicateBB**)
[tail] - the new basic block that merges [clone 1] and [clone 2] (**DuplicateBB**)
[clone] - [clone 1] and [clone 2] after merging, this block should be very similar to [BB] (**MergeBB**)
[label*] - [label] after being updated by **MergeBB**Recall that DuplicateBB replaces all qualifying basic block with four new basic blocks, two of which are clones of the original block. MergeBB will merge those two clones back together, but it will not remove the remaining two blocks added by DuplicateBB (it will update them though).
Lets use the following IR implementation of foo as input. Note that basic
blocks 3 and 5 are identical and can safely be merged:
define i32 @foo(i32) {
%2 = icmp eq i32 %0, 19
br i1 %2, label %3, label %5
; <label>:3:
%4 = add i32 %0, 13
br label %7
; <label>:5:
%6 = add i32 %0, 13
br label %7
; <label>:7:
%8 = phi i32 [ %4, %3 ], [ %6, %5 ]
ret i32 %8
}We will now apply MergeBB to foo:
$LLVM_DIR/bin/opt -load <build_dir>/lib/libMergeBB.so -legacy-merge-bb foo.llAfter the instrumentation foo will look like this (all metadata has been stripped):
define i32 @foo(i32) {
%2 = icmp eq i32 %0, 19
br i1 %2, label %3, label %3
3:
%4 = add i32 %0, 13
br label %5
5:
ret i32 %4
}As you can see, basic blocks 3 and 5 from the input module have been merged into one basic block.
export LLVM_DIR=<installation/dir/of/llvm/10>
$LLVM_DIR/bin/clang -emit-llvm -S -O1 inputs/input_for_duplicate_bb.c -o input_for_duplicate_bb.llIt is really interesting to see the effect of MergeBB on the output from DuplicateBB. Lets start with the same input as we used for DuplicateBB:
export LLVM_DIR=<installation/dir/of/llvm/10>
$LLVM_DIR/bin/clang -emit-llvm -S -O1 inputs/input_for_duplicate_bb.c -o input_for_duplicate_bb.llNow we will apply DuplicateBB and MergeBB (in this order) to foo.
Recall that DuplicateBB requires RIV, which means that in total we have
to load three plugins:
$LLVM_DIR/bin/opt -load-pass-plugin <build_dir>/lib/libRIV.so -load-pass-plugin <build_dir>/lib/libMergeBB.so -load-pass-plugin <build-dir>/lib/libDuplicateBB.so -passes=duplicate-bb,merge-bb input_for_duplicate_bb.llAnd here's the output:
define i32 @foo(i32) {
lt-if-then-else-0:
%1 = icmp eq i32 %0, 0
br i1 %1, label %lt-clone-2-0, label %lt-clone-2-0
lt-clone-2-0:
br label %lt-tail-0
lt-tail-0:
ret i32 1
}Compare this with the output generated by DuplicateBB.
Only one of the clones, lt-clone-2-0, has been preserved, and
lt-if-then-else-0 has been updated accordingly. Regardless of the value of of
the if condition (more precisely, variable %1), the control flow jumps to
lt-clone-2-0.
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/10>
$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.llNote 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 instructionsAs 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.
The default debugger on OS X is LLDB. You will normally use it like this:
export LLVM_DIR=<installation/dir/of/llvm/10>
$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 launchor, equivalently, by using LLDBs aliases:
export LLVM_DIR=<installation/dir/of/llvm/10>
$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) rAt 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/10>
$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/libMBAAdd.so -passes=mba-add input_for_mba.ll -o out.ll
(gdb) b MBAAdd.cpp:MBAAdd::run
(gdb) rAt this point, GDB should break at the entry to MBAAdd::run.
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 managers. This is important because depending on which pass manager you decide to use, the implementation of your pass (and in particular how you register it) will look slightly differently.
As I 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 there is no official documentation. Min-Yih Hsu kindly wrote this great blog series that you can refer to instead.
If you are not sure which pass manager to use, it is probably best to make sure that your passes are compatible with both. Fortunately, once you have an implementation that works with one of them, it's relatively straightforward to extend it so that it works with the other one as well.
MBAAdd implements interface for both pass managers. This is how you will use it with the legacy pass manager:
$LLVM_DIR/bin/opt -load <build_dir>/lib/libMBAAdd.so -legacy-mba-add input_for_mba.ll -o out.llAnd this is how you run it with the new pass manager:
$LLVM_DIR/bin/opt -load-pass-plugin <build_dir>/lib/libMBAAdd.so -passes=mba-add input_for_mba.ll -o out.llThere are two differences:
- the way you load your plugin:
-loadvs-load-pass-plugin - the way you specify which pass/plugin to run:
-legacy-mba-addvs-passes=mba-add
These differences stem from the fact that in the case of Legacy Pass Manager you
register a new command line option for opt, whereas New Pass Manager
simply requires you to define a pass pipeline (with -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
- Examples in LLVM
- Control Flow Graph simplifications: llvm/examples/IRTransforms/
- Hello World Pass: llvm/lib/Transforms/Hello/
- Good Bye World Pass: llvm/examples/Bye/
- LLVM Pass Development
- "Getting Started With LLVM: Basics ", J. Paquette, F. Hahn, LLVM Dev Meeting 2019 video
- "Writing an LLVM Pass: 101", A. Warzyński, LLVM Dev Meeting 2019 video
- "Writing LLVM Pass in 2018", Min-Yih Hsu blog
- "Building, Testing and Debugging a Simple out-of-tree LLVM Pass" Serge Guelton, Adrien Guinet, LLVM Dev Meeting 2015 (slides, video)
- Legacy vs New Pass Manager
- LLVM Based Tools Development
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.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.