Whole Program LLVM in Go

TL; DR: A drop-in replacement for wllvm, that builds the bitcode in parallel, and is faster. A comparison between the two tools can be gleaned from building the Linux kernel.

Quick Start Comparison Table

This project, gllvm, provides tools for building whole-program (or whole-library) LLVM bitcode files from an unmodified C or C++ source package. It currently runs on *nix platforms such as Linux, FreeBSD, and Mac OS X. It is a Go port of wllvm.

gllvm provides compiler wrappers that work in two phases. The wrappers first invoke the compiler as normal. Then, for each object file, they call a bitcode compiler to produce LLVM bitcode. The wrappers then store the location of the generated bitcode file in a dedicated section of the object file. When object files are linked together, the contents of the dedicated sections are concatenated (so we don't lose the locations of any of the constituent bitcode files). After the build completes, one can use a gllvm utility to read the contents of the dedicated section and link all of the bitcode into a single whole-program bitcode file. This utility works for both executable and native libraries.

For more details see wllvm.

Prerequisites

To install gllvm you need the go language tool.

To use gllvm you need clang/clang++/flang and the llvm tools llvm-link and llvm-ar. gllvm is agnostic to the actual llvm version. gllvm also relies on standard build tools such as objcopy and ld.

Installation

To install, simply do (making sure to include those ...)

go install github.com/SRI-CSL/gllvm/cmd/...@latest

This should install six binaries: gclang, gclang++, gflang, get-bc, gparse, and gsanity-check in the $GOPATH/bin directory.

Usage

gclang and gclang++ are the wrappers used to compile C and C++.
gflang is the wrapper used to compile Fortran. get-bc is used for extracting the bitcode from a build product (either an object file, executable, library or archive). gsanity-check can be used for detecting configuration errors. gparse can be used to examine how gllvm parses compiler/linker lines.

Here is a simple example. Assuming that clang is in your PATH, you can build bitcode for pkg-config as follows:

tar xf pkg-config-0.26.tar.gz
cd pkg-config-0.26
CC=gclang ./configure
make

This should produce the executable pkg-config. To extract the bitcode:

get-bc pkg-config

which will produce the bitcode module pkg-config.bc. For more on this example see here.

Advanced Configuration

If clang and the llvm tools are not in your PATH, you will need to set some environment variables.

• LLVM_COMPILER_PATH can be set to the absolute path of the directory that contains the compiler and the other LLVM tools to be used.

• LLVM_CC_NAME can be set if your clang compiler is not called clang but something like clang-3.7. Similarly LLVM_CXX_NAME and LLVM_F_NAME can be used to describe what the C++ and Fortran compilers are called, respectively. We also pay attention to the environment variables LLVM_LINK_NAME and LLVM_AR_NAME in an analogous way.

Another useful, and sometimes necessary, environment variable is WLLVM_CONFIGURE_ONLY.

• WLLVM_CONFIGURE_ONLY can be set to anything. If it is set, gclang and gclang++ behave like a normal C or C++ compiler. They do not produce bitcode. Setting WLLVM_CONFIGURE_ONLY may prevent configuration errors caused by the unexpected production of hidden bitcode files. It is sometimes required when configuring a build. For example:

  WLLVM_CONFIGURE_ONLY=1 CC=gclang ./configure
  make

Extracting the Bitcode

The get-bc tool is used to extract the bitcode from a build artifact, such as an executable, object file, thin archive, archive, or library. In the simplest use case, as seen above, one simply does:

get-bc -o <name of bitcode file> <path to executable>

This will produce the desired bitcode file. The situation is similar for an object file. For an archive or library, there is a choice as to whether you produce a bitcode module or a bitcode archive. This choice is made by using the -b switch.

Another useful switch is the -m switch which will, in addition to producing the bitcode, will also produce a manifest of the bitcode files that made up the final product. As is typical

get-bc -h

will list all the commandline switches. Since we use the golang flag module, the switches must precede the artifact path.

Preserving bitcode files in a store

Sometimes, because of pathological build systems, it can be useful to preserve the bitcode files produced in a build, either to prevent deletion or to retrieve it later. If the environment variable WLLVM_BC_STORE is set to the absolute path of an existing directory, then WLLVM will copy the produced bitcode file into that directory. The name of the copied bitcode file is the hash of the path to the original bitcode file. For convenience, when using both the manifest feature of get-bc and the store, the manifest will contain both the original path, and the store path.

Debugging

The gllvm tools can show various levels of output to aid with debugging. To show this output set the WLLVM_OUTPUT_LEVEL environment variable to one of the following levels:

• ERROR
• WARNING
• AUDIT
• INFO
• DEBUG

For example:

    export WLLVM_OUTPUT_LEVEL=DEBUG

Output will be directed to the standard error stream, unless you specify the path of a logfile via the WLLVM_OUTPUT_FILE environment variable. The AUDIT level, new in 2022, logs only the calls to the compiler, and indicates whether each call is compiling or linking, the compiler used, and the arguments provided.

For example:

    export WLLVM_OUTPUT_FILE=/tmp/gllvm.log

Dragons Begone

gllvm does not support the dragonegg plugin.

Sanity Checking

Too many environment variables? Try doing a sanity check:

gsanity-check

it might point out what is wrong.

Under the hoods

Both wllvm and gllvm toolsets do much the same thing, but the way they do it is slightly different. The gllvm toolset's code base is written in golang, and is largely derived from the wllvm's python codebase.

Both generate object files and bitcode files using the compiler. wllvm can use gcc and dragonegg, gllvm can only use clang. The gllvm toolset does these two tasks in parallel, while wllvm does them sequentially. This together with the slowness of python's fork exec-ing, and it's interpreted nature accounts for the large efficiency gap between the two toolsets.

Both inject the path of the bitcode version of the .o file into a dedicated segment of the .o file itself. This segment is the same across toolsets, so extracting the bitcode can be done by the appropriate tool in either toolset. On *nix both toolsets use objcopy to add the segment, while on OS X they use ld.

When the object files are linked into the resulting library or executable, the bitcode path segments are appended, so the resulting binary contains the paths of all the bitcode files that constitute the binary. To extract the sections the gllvm toolset uses the golang packages "debug/elf" and "debug/macho", while the wllvm toolset uses objdump on *nix, and otool on OS X.

Both tools then use llvm-link or llvm-ar to combine the bitcode files into the desired form.

Customization under the hood.

You can specify the exact version of objcopy and ld that gllvm uses to manipulate the artifacts by setting the GLLVM_OBJCOPY and GLLVM_LD environment variables. For more details of what's under the gllvm hood, try

gsanity-check -e

Customizing the BitCode Generation (e.g. LTO)

In some situations it is desirable to pass certain flags to clang in the step that produces the bitcode. This can be fulfilled by setting the LLVM_BITCODE_GENERATION_FLAGS environment variable to the desired flags, for example "-flto -fwhole-program-vtables".

In other situations it is desirable to pass certain flags to llvm-link in the step that merges multiple individual bitcode files together (i.e., within get-bc). This can be fulfilled by setting the LLVM_LINK_FLAGS environment variable to the desired flags, for example "-internalize -only-needed".

Beware of link time optimization.

If the package you are building happens to take advantage of recent clang developments such as link time optimization (indicated by the presence of compiler flag -flto), then your build is unlikely to produce anything that get-bc will work on. This is to be expected. When working under these flags, the compiler actually produces object files that are bitcode, your only recourse here is to try and save these object files, and retrieve them yourself. This can be done by setting the LTO_LINKING_FLAGS to be something like "-g -Wl,-plugin-opt=save-temps" which will be appended to the flags at link time. This will at least preserve the bitcode files, even if get-bc will not be able to retrieve them for you.

Cross-compilation notes

When cross-compiling a project (i.e. you pass the --target= or -target flag to the compiler), you'll need to set the GLLVM_OBJCOPY variable to either

• llvm-objcopy to use LLVM's objcopy, which naturally supports all targets that clang does.
• YOUR-TARGET-TRIPLE-objcopy to use GNU's objcopy, since objcopy only supports the native architecture.

Example:

# test program
echo 'int main() { return 0; }' > a.c 
clang --target=aarch64-linux-gnu a.c # works
gclang --target=aarch64-linux-gnu a.c # breaks
GLLVM_OBJCOPY=llvm-objcopy gclang --target=aarch64-linux-gnu a.c # works
GLLVM_OBJCOPY=aarch64-linux-gnu-objcopy gclang --target=aarch64-linux-gnu a.c # works if you have GNU's arm64 toolchain

Developer tools

Debugging usually boils down to looking in the logs, maybe adding a print statement or two. There is an additional executable, not mentioned above, called gparse that gets installed along with gclang, gclang++, gflang, get-bc and gsanity-check. gparse takes the command line arguments to the compiler, and outputs how it parsed them. This can sometimes be helpful.

License

gllvm is released under a BSD license. See the file LICENSE for details.

This material is based upon work supported by the National Science Foundation under Grant ACI-1440800. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation.