From YobiWiki
Jump to navigation Jump to search



Tagged version on git failed, so working from the tarballs:

cd /path/to
mkdir llvm32
cd llvm32
wget http://llvm.org/releases/3.2/llvm-3.2.src.tar.gz
wget http://llvm.org/releases/3.2/clang-3.2.src.tar.gz
tar xzf llvm-3.2.src.tar.gz
cd llvm-3.2.src/tools/
tar xzf ../../clang-3.2.src.tar.gz
cd /path/to/llvm32
mkdir llvm-build
cd llvm-build
../llvm-3.2.src/configure --enable-optimized --prefix=/path/to/llvm32/llvm-install
REQUIRES_RTTI=1 make -j3
make install


export PATH=/path/to/llvm32/llvm-install/bin:$PATH


Tagged version on git failed, so working from the tarballs:

cd /path/to
mkdir llvm33
cd llvm33
wget http://llvm.org/releases/3.3/llvm-3.3.src.tar.gz
wget http://llvm.org/releases/3.3/cfe-3.3.src.tar.gz
tar xzf llvm-3.3.src.tar.gz
cd llvm-3.3.src/tools/
tar xzf ../../cfe-3.3.src.tar.gz
mv cfe-3.3.src clang
cd /path/to/llvm33
mkdir llvm-build
cd llvm-build
../llvm-3.3.src/configure --enable-optimized --prefix=/path/to/llvm33/llvm-install
REQUIRES_RTTI=1 make -j3
make install


export PATH=/path/to/llvm33/llvm-install/bin:$PATH



cd /path/to
mkdir llvm34
cd llvm34
wget http://llvm.org/releases/3.4/llvm-3.4.src.tar.gz
wget http://llvm.org/releases/3.4/clang-3.4.src.tar.gz
tar xzf llvm-3.4.src.tar.gz
cd llvm-3.4/tools/
tar xzf ../../clang-3.4.src.tar.gz
mv clang-3.4 clang
cd /path/to/llvm34
mkdir llvm-build
cd llvm-build
../llvm-3.4/configure --enable-optimized --prefix=/path/to/llvm34/llvm-install
REQUIRES_RTTI=1 make -j3
make install


export PATH=/path/to/llvm34/llvm-install/bin:$PATH

LLVM basics


clang -Os -S -emit-llvm foo.c -o foo.ll


clang -Os -c -emit-llvm foo.c -o foo.bc


clang -Os -c -emit-llvm foo.ll -o foo.bc


llvm-dis -f foo.bc


llc foo.bc

bc analysis

opt -stats  -instcount  foo.bc
llvm-bcanalyzer [-dump] foo.bc

See also bc.py for another bc parser written in Python

with Python


Install with LLVM 3.2:

export PATH=/path/to/llvm32/llvm-install/bin:$PATH
virtualenv llvmpy_llvm32.env
. llvmpy_llvm32.env/bin/activate
git clone git@github.com:llvmpy/llvmpy.git llvmpy.git
cd llvmpy.git
git checkout 0.12.5
LLVM_CONFIG_PATH=/path/to/llvm32/llvm-install/bin/llvm-config python setup.py install
cd ..
python -c "import llvm; llvm.test()"

Usage with LLVM 3.2:

export PATH=/path/to/llvm32/llvm-install/bin:$PATH
. llvmpy_llvm32.env/bin/activate
# Do your stuff... python -c "import llvm; llvm.test()"

Install with LLVM 3.3:
idem, just replace llvm32 by llvm33

Install with LLVM 3.4
v0.12.5 doesn't support LLVM 3.4 apparently, compil fails


set of Python-contexts you can use to write C-like constructs in Python which generates llvmpy code directly.
Similar to llpython it allows you to build llvm IR without using the llvmpy interface directly.


The primary goal of the llpython package is to provide a Python dialect/subset that maps directly to LLVM code.
LLPython differs from its originating LLVM translator, Numba, in the following aspects:

  • LLPython code is not intended to work in Python if not translated and wrapped.
  • The LLPython translator only uses LLVM types.
  • LLPython is explicitly typed, and does not support type inference. LLPython does not support implicit casts, all casts must be explicit.
  • LLPython supports code that directly calls the C API, the Python C API, and the llvm.core.Builder methods.

We developed LLPython with the initial goal of simplifying writing LLVM code.

Tests as at the moment there is very little user doc:

import llvm.core as lc
from llpython import bytetype

ll_ipow = lc.Type.function(bytetype.li32, (bytetype.li32, bytetype.li32))

def ipow (val, exp):
    ret_val = 1
    temp = val
    w = exp
    while w > 0:
        if (w & 1) != 0:
            ret_val *= temp
            # TODO: Overflow check on ret_val
        w >>= 1
        if w == 0: break
        temp *= temp
        # TODO: Overflow check on temp
    return ret_val

ll_pymod = lc.Type.function(bytetype.li32, (bytetype.li32, bytetype.li32))

def pymod (arg1, arg2):
    ret_val = arg1 % arg2
    if ret_val < 0:
        if arg2 > 0:
            ret_val += arg2
    elif arg2 < 0:
        ret_val += arg2
    return ret_val

lm = lc.Module.new('test_module')
from llpython import byte_translator
byte_translator.translate_function(ipow, ll_ipow, lm)
byte_translator.translate_function(pymod, ll_pymod, lm)

# Ok let's JIT & run the LLVM IR code:
from llvm.ee import ExecutionEngine, GenericValue
ee = ExecutionEngine.new(lm)

arg1 = GenericValue.int(bytetype.li32, 235)
arg2 = GenericValue.int(bytetype.li32, 17)
retval = ee.run_function(lm.get_function_named('pymod'), [arg1, arg2])
print "pymod(235, 17) =", retval.as_int()

# Tried ipow() but it loops infinitely
#arg1 = GenericValue.int(bytetype.li32, 2)
#arg2 = GenericValue.int(bytetype.li32, 8)
#retval = ee.run_function(lm.get_function_named('ipow'), [arg1, arg2])
#print "returned", retval.as_int()


Numba is a NumPy aware dynamic compiler for Python. It creates LLVM bit-code from Python syntax and then creates a wrapper around that bitcode to call from Python.

pycc allows users to compile Numba functions into a shared library. The user writes the functions, exports them and the compiler will import the module, collect the exported functions and compile them to a shared library. Below is an example:

from numba import *
def mult(a, b):
    return a * b
export('mult f8(f8, f8)'))(mult)
export(['multf f4(f4, f4)', 'multi i4(i4, i4)'])(mult)
export('multc c16(c16, c16)'))(mult)


llvmlite is the lightweight Python binding developed for Numba


by Dropbox

Not mature yet

old attempts