diff --git a/CHANGELOG b/CHANGELOG
index 5af6169..12bb294 100644
--- a/CHANGELOG
+++ b/CHANGELOG
@@ -1,3 +1,10 @@
+2013-8-28    0.12.0:
+--------------------
+  * update to LLVM 3.3 and maintain compatibility with LLVM 3.2
+  * add LLRT for minimal support for 64-bit divmod on 32-bit platform
+  * start to adopt MCJIT (not quite usable on win32)
+  * various bug fixes
+
 2013-03-05   0.11.1:
 --------------------
   * fix test when cc is not available
diff --git a/buildscripts/condarecipe/bld.bat b/buildscripts/condarecipe/bld.bat
new file mode 100644
index 0000000..e9177c5
--- /dev/null
+++ b/buildscripts/condarecipe/bld.bat
@@ -0,0 +1,6 @@
+set LLVMPY_DYNLINK=0
+set INCLUDE=%LIBRARY_INC%
+set LIBPATH=%LIBRARY_LIB%
+set LIB=%LIBRARY_LIB%
+%PYTHON% setup.py install
+if errorlevel 1 exit 1
diff --git a/buildscripts/condarecipe/build.sh b/buildscripts/condarecipe/build.sh
new file mode 100644
index 0000000..5c98fde
--- /dev/null
+++ b/buildscripts/condarecipe/build.sh
@@ -0,0 +1,13 @@
+#!/bin/bash
+
+if [[ (`uname` == Linux) && (`uname -m` != armv6l) ]]
+then
+    export CC=gcc
+	#gcc44
+    export CXX=g++
+	#g++44
+fi
+
+export LLVMPY_DYNLINK=$DISTRO_BUILD
+
+$PYTHON setup.py install
diff --git a/buildscripts/condarecipe/meta.yaml b/buildscripts/condarecipe/meta.yaml
new file mode 100644
index 0000000..4612840
--- /dev/null
+++ b/buildscripts/condarecipe/meta.yaml
@@ -0,0 +1,28 @@
+package:
+  name: llvmpy
+  version: 99.9.9
+  
+source:
+  git_url: git@github.com:llvmpy/llvmpy.git
+#  git_tag: 0.12.0
+
+requirements:
+  build:
+    - llvm
+    - python
+    #- chrpath        [linux]
+  run:
+    - llvm           [unix]
+    - python
+
+test:
+  imports:
+    - llvm
+    - llvmpy
+    - llvmpy._api
+    - llvmpy._capsule
+    - llpython
+    - llvm_array
+    - llvm_cbuilder
+
+
diff --git a/buildscripts/condarecipe/run_test.py b/buildscripts/condarecipe/run_test.py
new file mode 100644
index 0000000..281a599
--- /dev/null
+++ b/buildscripts/condarecipe/run_test.py
@@ -0,0 +1,19 @@
+import sys
+import platform
+import llvm
+
+from llvm.core import Module
+from llvm.ee import EngineBuilder
+m = Module.new('fjoidajfa')
+eb = EngineBuilder.new(m)
+target = eb.select_target()
+
+print('target.triple=%r' % target.triple)
+if sys.platform == 'darwin':
+    s = {'64bit': 'x86_64', '32bit': 'x86'}[platform.architecture()[0]]
+    assert target.triple.startswith(s + '-apple-darwin')
+
+assert llvm.test(verbosity=2) == 0
+
+print('llvm.__version__: %s' % llvm.__version__)
+#assert llvm.__version__ == '0.12.0'
diff --git a/llrtc/Makefile b/llrtc/Makefile
new file mode 100644
index 0000000..b15eac5
--- /dev/null
+++ b/llrtc/Makefile
@@ -0,0 +1,22 @@
+all:
+	make -C lib
+
+ir:
+	make -C lib ir
+
+test:
+	make -C lib test
+
+clean-test:
+	make -C lib clean-test
+
+clean-temp:
+	make -C lib clean-temp
+
+clean:
+	make -C lib clean
+
+install: ir
+	cp llrt_*.ll ../llvm/llrt
+	make -C lib clean-temp
+	
\ No newline at end of file
diff --git a/llrtc/README.md b/llrtc/README.md
new file mode 100644
index 0000000..539d9c3
--- /dev/null
+++ b/llrtc/README.md
@@ -0,0 +1,25 @@
+# LLRT: Low Level Runtime
+
+## Why?
+
+The same reason for LLVM compiler-rt.  LLVM generates libgcc symbols, such as
+__divdi3 for 64-bit division on 32-bit platform.  They are not also available.
+We need to ship compiler-rt but it is not Windows ready.
+This subproject aims to provide a small portable subset of compiler-rt.
+Start small and add only the things we really needed.
+Performance is not crucial but should not be terrible.
+Functionality and usefullness should be more important than performance.
+
+## Developer Instructions
+
+LLRT implements some functionalities in compiler-rt in ANSI C.
+The C files are compiled using clang to produce LLVM IR which are shipped.
+The IR files are committed in the repository.
+So, remember to build the IR files commit them after modifying the C files.
+
+## Build Requirement
+
+- Make
+- Clang
+- Python
+
diff --git a/llrtc/lib/.gitignore b/llrtc/lib/.gitignore
new file mode 100644
index 0000000..fbcf76a
--- /dev/null
+++ b/llrtc/lib/.gitignore
@@ -0,0 +1,4 @@
+*.o
+*.run
+*.out
+*.ll
diff --git a/llrtc/lib/Makefile b/llrtc/lib/Makefile
new file mode 100644
index 0000000..9dbfb27
--- /dev/null
+++ b/llrtc/lib/Makefile
@@ -0,0 +1,66 @@
+OUTPUT = llrt
+SOURCES = udivmod64.c sdivmod64.c div64.c mod64.c
+TESTS = test_udivmod64.c test_sdivmod64.c
+
+CLANG = clang
+LLVM_LINK = llvm-link
+CF = -Wall -ansi
+CF_TEST = $(CF) -ftrapv
+CF_BUILD = $(CF) -O0 -emit-llvm
+OUTDIR = ..
+STRIPPER = ../tools/striptriple.py
+
+all: ir
+
+ir: $(OUTDIR)/$(OUTPUT)_x86.ll $(OUTDIR)/$(OUTPUT)_x86_64.ll
+
+$(OUTDIR)/$(OUTPUT)_x86.ll: $(SOURCES:.c=_x86.bc)
+	$(LLVM_LINK) -S $+ -o $@
+	python $(STRIPPER) $@
+
+$(OUTDIR)/$(OUTPUT)_x86_64.ll: $(SOURCES:.c=_x86_64.bc)
+	$(LLVM_LINK) -S $+ -o $@
+	python $(STRIPPER) $@
+
+build-test: $(SOURCES:.c=.o) $(TESTS:.c=.run)
+
+lib$(OUTPUT).a: $(SOURCES:.c=.o)
+	$(CLANG) -static $+ -o $@
+
+test: $(TESTS:.c=.run)
+	for src in $+; do \
+		echo "testing $${src}";				 \
+		python $${src%.*}.py > $${src%.*}.out;	 \
+	done;
+
+clean-test:
+	rm -f *.out
+	rm -f *.o
+	rm -f *.run
+
+clean-dist: clean-temp
+	rm -f *.ll
+
+clean-temp:
+	rm -f *.bc
+	rm -f *.o
+	rm -f *.out
+
+clean: clean-test clean-dist
+
+%.c: llrt.h
+
+%_x86.bc: %.c
+	$(CLANG) -m32 $(CF_BUILD) -c $< -o $@
+
+%_x86_64.bc: %.c
+	$(CLANG) -m64 $(CF_BUILD) -c $< -o $@
+
+%.o: %.c
+	$(CLANG) $(CF_TEST) -c $<
+
+%.run: %.c
+	$(CLANG) $(CF_TEST) -o $@ $+
+
+test_udivmod64.run: udivmod64.o
+test_sdivmod64.run: udivmod64.o sdivmod64.o
diff --git a/llrtc/lib/div64.c b/llrtc/lib/div64.c
new file mode 100644
index 0000000..8466965
--- /dev/null
+++ b/llrtc/lib/div64.c
@@ -0,0 +1,11 @@
+#include "llrt.h"
+
+uint64_t udiv64(uint64_t dividend, uint64_t divisor)
+{
+    return udivmod64(dividend, divisor, NULL);
+}
+
+int64_t sdiv64(int64_t dividend, int64_t divisor)
+{
+    return sdivmod64(dividend, divisor, NULL);
+}
diff --git a/llrtc/lib/llrt.h b/llrtc/lib/llrt.h
new file mode 100644
index 0000000..4e9e348
--- /dev/null
+++ b/llrtc/lib/llrt.h
@@ -0,0 +1,19 @@
+#ifndef LLRT_H_
+#define LLRT_H_
+
+#include <stdint.h>
+
+#define NULL 0
+#define BITS_PER_BYTE 8
+
+uint64_t udivmod64(uint64_t dividend, uint64_t divisor, uint64_t *remainder);
+int64_t sdivmod64(int64_t dividend, int64_t divisor, int64_t *remainder);
+
+uint64_t udiv64(uint64_t dividend, uint64_t divisor);
+int64_t sdiv64(int64_t dividend, int64_t divisor);
+
+uint64_t umod64(uint64_t dividend, uint64_t divisor);
+int64_t smod64(int64_t dividend, int64_t divisor);
+
+#endif /* LLRT_H_ */
+
diff --git a/llrtc/lib/mod64.c b/llrtc/lib/mod64.c
new file mode 100644
index 0000000..875f6b9
--- /dev/null
+++ b/llrtc/lib/mod64.c
@@ -0,0 +1,15 @@
+#include "llrt.h"
+
+uint64_t umod64(uint64_t dividend, uint64_t divisor)
+{
+    uint64_t rem;
+    udivmod64(dividend, divisor, &rem);
+    return rem;
+}
+
+int64_t smod64(int64_t dividend, int64_t divisor)
+{
+    int64_t rem;
+    sdivmod64(dividend, divisor, &rem);
+    return rem;
+}
diff --git a/llrtc/lib/sdivmod64.c b/llrtc/lib/sdivmod64.c
new file mode 100644
index 0000000..df294b3
--- /dev/null
+++ b/llrtc/lib/sdivmod64.c
@@ -0,0 +1,40 @@
+#include "llrt.h"
+#include <stdio.h>
+
+/*
+Calls to udivmod64 internally.
+Note: remainder uses sign of divisor.
+*/
+int64_t sdivmod64(int64_t dividend, int64_t divisor, int64_t *remainder)
+{
+    int signbitidx = BITS_PER_BYTE * sizeof(dividend) - 1;
+    int signed_dividend = dividend < 0;
+    int signed_divisor = divisor < 0;
+    int signed_result = signed_divisor ^ signed_dividend;
+
+    int64_t quotient;
+    uint64_t udvd, udvr, uquotient, uremainder;
+
+    udvd = signed_dividend ? -dividend : dividend;
+    udvr = signed_divisor ? -divisor : divisor;
+    uquotient = udivmod64(udvd, udvr, &uremainder);
+
+    if (signed_result){
+        if (uremainder) {
+            quotient = -(int64_t)uquotient - 1;
+        } else {
+            quotient = -(int64_t)uquotient;
+        }
+        if (remainder) {
+            /* if signed, there could be unsigned overflow
+               causing undefined behavior */
+            *remainder = (uint64_t)dividend - (uint64_t)quotient * (uint64_t)divisor;
+        }
+    } else {
+        quotient = (int64_t)uquotient;
+        if (remainder) {
+            *remainder = signed_divisor ? -uremainder : uremainder;
+        }
+    }
+    return quotient;
+}
diff --git a/llrtc/lib/test_sdivmod64.c b/llrtc/lib/test_sdivmod64.c
new file mode 100644
index 0000000..1e804d6
--- /dev/null
+++ b/llrtc/lib/test_sdivmod64.c
@@ -0,0 +1,21 @@
+#include <stdio.h>
+#include <stdint.h>
+#include "llrt.h"
+
+int main(int argc, char * argv[]){
+    int64_t n, d, q, r;
+
+    if (argc != 3) {
+        printf("invalid argument: %s dividend divisor", argv[0]);
+        return 1;
+    }
+    sscanf(argv[1], "%lld", &n);
+    sscanf(argv[2], "%lld", &d);
+
+    q = sdivmod64(n, d, &r);
+
+    printf("%lld\n", q);
+    printf("%lld\n", r);
+
+    return 0;
+}
diff --git a/llrtc/lib/test_sdivmod64.py b/llrtc/lib/test_sdivmod64.py
new file mode 100644
index 0000000..4aa49a9
--- /dev/null
+++ b/llrtc/lib/test_sdivmod64.py
@@ -0,0 +1,56 @@
+import math
+import os
+import subprocess
+udt = os.path.join('.', 'test_sdivmod64.run')
+
+def testcase(dividend, divisor):
+    print 'divmod64(%d, %d)' % (dividend, divisor)
+
+    procargs = ('%s %s %s' % (udt, dividend, divisor)).split()
+    result = subprocess.check_output(procargs)
+    gotQ, gotR = map(int, result.splitlines())
+
+    expectQ = dividend // divisor
+    expectR = dividend % divisor
+
+    print 'Q = %d, R = %d' % (gotQ, gotR)
+
+    if expectQ != gotQ:
+        raise ValueError("invalid quotient: got=%d but expect=%d" %
+                         (gotQ, expectQ))
+    if expectR != gotR:
+        raise ValueError("invalid remainder: got=%d but expect=%d" %
+                         (gotR, expectR))
+    print 'OK'
+
+def testsequence():
+    subjects = [
+        (0, 1),
+        (0, 0xffffffff),
+        (1, 2),
+        (1, 983219),
+        (2, 2),
+        (3, 2),
+        (1024, 2),
+        (2048, 512),
+        (21321, 512),
+        (9329189, 1031),
+        (0xffffffff, 2),
+        (0xffffffff, 0xffff),
+        (0x1ffffffff, 2),
+        (0x1ffffffff, 0xffff),
+        (0xffff, 0xffffffff),
+        (0x0fffffffffffffff, 0xffff),
+        (0x7fffffffffffffff, 0x7fffffffffffffff),
+        (0x7fffffffffffffff, 0x7ffffffffffffff0),
+        (0x7fffffffffffffff, 87655678587161901),
+    ]
+
+    for dvd, dvr in subjects:
+        testcase(dvd, dvr)
+        testcase(dvd, -dvr)
+        testcase(-dvd, dvr)
+        testcase(-dvd, -dvr)
+
+if __name__ == '__main__':
+    testsequence()
diff --git a/llrtc/lib/test_udivmod64.c b/llrtc/lib/test_udivmod64.c
new file mode 100644
index 0000000..70b44aa
--- /dev/null
+++ b/llrtc/lib/test_udivmod64.c
@@ -0,0 +1,20 @@
+#include <stdio.h>
+#include <stdint.h>
+#include "llrt.h"
+
+int main(int argc, char * argv[]){
+    uint64_t n, d, q, r;
+    if (argc != 3) {
+        printf("invalid argument: %s dividend divisor", argv[0]);
+        return 1;
+    }
+    sscanf(argv[1], "%llu", &n);
+    sscanf(argv[2], "%llu", &d);
+
+    q = udivmod64(n, d, &r);
+
+    printf("%llu\n", q);
+    printf("%llu\n", r);
+
+    return 0;
+}
diff --git a/llrtc/lib/test_udivmod64.py b/llrtc/lib/test_udivmod64.py
new file mode 100644
index 0000000..da07ccb
--- /dev/null
+++ b/llrtc/lib/test_udivmod64.py
@@ -0,0 +1,53 @@
+import math
+import os
+import subprocess
+udt = os.path.join('.', 'test_udivmod64.run')
+
+def testcase(dividend, divisor):
+    print 'divmod64(%d, %d)' % (dividend, divisor)
+
+    procargs = ('%s %s %s' % (udt, dividend, divisor)).split()
+    result = subprocess.check_output(procargs)
+    gotQ, gotR = map(int, result.splitlines())
+
+    expectQ = dividend // divisor
+    expectR = dividend % divisor
+
+    print 'Q = %d, R = %d' % (gotQ, gotR)
+
+    if expectQ != gotQ:
+        raise ValueError("invalid quotient: got=%d but expect=%d" %
+                         (gotQ, expectQ))
+    if expectR != gotR:
+        raise ValueError("invalid remainder: got=%d but expect=%d" %
+                         (gotR, expectR))
+    print 'OK'
+
+def testsequence():
+    subjects = [
+        (0, 1),
+        (0, 0xffffffffffffffff),
+        (1, 2),
+        (1, 983219),
+        (2, 2),
+        (3, 2),
+        (1024, 2),
+        (2048, 512),
+        (21321, 512),
+        (9329189, 1031),
+        (0xffffffff, 2),
+        (0xffffffff, 0xffff),
+        (0x1ffffffff, 2),
+        (0x1ffffffff, 0xffff),
+        (0xffff, 0xffffffff),
+        (0xffffffffffffffff, 0xffff),
+        (0xffffffffffffffff, 0x7fffffffffffffff),
+        (0xffffffffffffffff, 0xfffffffffffffff0),
+        (0xffffffffffffffff, 87655678587161901),
+    ]
+
+    for dvd, dvr in subjects:
+        testcase(dvd, dvr)
+
+if __name__ == '__main__':
+    testsequence()
diff --git a/llrtc/lib/udivmod64.c b/llrtc/lib/udivmod64.c
new file mode 100644
index 0000000..5936f98
--- /dev/null
+++ b/llrtc/lib/udivmod64.c
@@ -0,0 +1,84 @@
+/*
+Implements unsigned divmod using for platform missing 64-bit division and/or
+modulo functions.
+*/
+#include "llrt.h"
+
+/*
+count left zero for 64-bit words
+
+*/
+static
+int clz64(uint64_t x)
+{
+    const int total_bits = sizeof(x) * BITS_PER_BYTE;
+    int zc = 0;
+
+    while (zc < total_bits && ((x >> (total_bits - zc - 1)) & 1) == 0) {
+        ++zc;
+    }
+    return zc;
+}
+
+typedef struct div_state_
+{
+    uint64_t tmp, dvd;
+} div_state;
+
+/*
+Left shift div_state by 1 bit
+*/
+static
+void div_state_lshift(div_state *state)
+{
+    state->tmp = (state->tmp << 1) | (state->dvd >> 63);
+    state->dvd = state->dvd << 1;
+}
+
+/*
+Division of unsigned 64-bit word using 64-bit addition and subtration following
+the shift-restore division algorithm.
+For those interested in 32-bit implementation,
+mapping of 64-bit addition and subtraction to 32-bit should be trivial.
+
+Reference: 
+    - IBM. The PowerPC Compiler Writer's Guide
+    - LLVM compiler-rt
+
+Assumptions:
+    - all operands and results are positive
+    - unsigned wrapped around
+*/
+uint64_t udivmod64(uint64_t dividend, uint64_t divisor, uint64_t *remainder)
+{
+    div_state state = {0, dividend};
+    uint64_t quotient = 0;
+    int i;
+    int skipahead;
+
+    if (divisor == 0) {
+        return 1 / 0;               /* intentionally div by zero */
+    }
+
+    /*
+    skipahead to reduce iteration
+    */
+    skipahead = clz64(dividend);
+
+    for (i = 0; i < skipahead; ++i) {
+        div_state_lshift(&state);
+    }
+
+    /*
+    division loop
+    */
+    for (i = skipahead; i < 64; ++i) {
+        div_state_lshift(&state);
+        if (state.tmp >= divisor) {
+            state.tmp = state.tmp - divisor;
+            quotient |= 1ull << (63 - i);
+        }
+    }
+    if (remainder) *remainder = state.tmp;
+    return quotient;
+}
diff --git a/llrtc/tools/striptriple.py b/llrtc/tools/striptriple.py
new file mode 100644
index 0000000..a20ba34
--- /dev/null
+++ b/llrtc/tools/striptriple.py
@@ -0,0 +1,15 @@
+import sys
+import re
+
+buf = []
+with open(sys.argv[1], 'r') as fin:
+    tripleline = re.compile('^target\s+triple\s+=\s+')
+    for line in fin.readlines():
+        if not tripleline.match(line):
+            buf.append(line)
+
+with open(sys.argv[1], 'w') as fout:
+    for line in buf:
+        fout.write(line)
+
+
diff --git a/llvm-config-win32.py b/llvm-config-win32.py
index 5f6c401..8c268cb 100644
--- a/llvm-config-win32.py
+++ b/llvm-config-win32.py
@@ -53,6 +53,7 @@ LLVMInterpreter
 LLVMipa
 LLVMipo
 LLVMJIT
+LLVMMCJIT
 LLVMLinker
 LLVMMC
 LLVMMCParser
@@ -71,6 +72,7 @@ LLVMX86Desc
 LLVMX86Disassembler
 LLVMX86Info
 LLVMX86Utils
+LLVMDebugInfo
 Advapi32
 Shell32
 """.split():
@@ -106,7 +108,7 @@ def main():
 
     elif option == '--includedir':
         incdir = join(find_llvm_prefix(), 'include')
-        ensure_file(join(incdir, 'llvm' , 'BasicBlock.h'))
+        ensure_file(join(incdir, 'llvm' , 'Linker.h'))
         print(incdir)
 
     elif option == '--libdir':
diff --git a/llvm/core.py b/llvm/core.py
index 9c8501e..c7bb49c 100644
--- a/llvm/core.py
+++ b/llvm/core.py
@@ -325,11 +325,17 @@ class VisibilityEnum(Enum):
 
 VisibilityEnum.declare()
 
-# parameter attributes llvm::Attributes::AttrVal (see llvm/Attributes.h)
+# parameter attributes
+#      LLVM 3.2 llvm::Attributes::AttrVal (see llvm/Attributes.h)
+#      LLVM 3.3 llvm::Attribute::AttrKind (see llvm/Attributes.h)
 class AttrEnum(Enum):
     prefix = 'ATTR_'
 
-    AttrVal = api.llvm.Attributes.AttrVal
+    if llvm.version >= (3, 3):
+        AttrVal = api.llvm.Attribute.AttrKind
+    else:
+        AttrVal = api.llvm.Attributes.AttrVal
+
     ATTR_NONE               = AttrVal.None_
     ATTR_ZEXT               = AttrVal.ZExt
     ATTR_SEXT               = AttrVal.SExt
@@ -1031,6 +1037,9 @@ class Value(llvm.Wrapper):
     def __str__(self):
         return str(self._ptr)
 
+    def __hash__(self):
+        return hash(self._ptr)
+
     def __eq__(self, rhs):
         if isinstance(rhs, Value):
             return str(self) == str(rhs)
@@ -1436,29 +1445,55 @@ class Argument(Value):
     _valid_attrs = frozenset([ATTR_BY_VAL, ATTR_NEST, ATTR_NO_ALIAS,
                               ATTR_NO_CAPTURE, ATTR_STRUCT_RET])
 
-    def add_attribute(self, attr):
-        context = api.llvm.getGlobalContext()
-        attrbldr = api.llvm.AttrBuilder.new()
-        attrbldr.addAttribute(attr)
-        attrs = api.llvm.Attributes.get(context, attrbldr)
-        self._ptr.addAttr(attrs)
-        if attr not in self:
-            raise ValueError("Attribute %s is not valid for arg %s" %
-                             (attr, self))
+    if llvm.version >= (3, 3):
+        def add_attribute(self, attr):
+            context = api.llvm.getGlobalContext()
+            attrbldr = api.llvm.AttrBuilder.new()
+            attrbldr.addAttribute(attr)
+            attrs = api.llvm.AttributeSet.get(context, 0, attrbldr)
+            self._ptr.addAttr(attrs)
+        
+            if attr not in self:
+                raise ValueError("Attribute %r is not valid for arg %s" %
+                                 (attr, self))
 
-    def remove_attribute(self, attr):
-        context = api.llvm.getGlobalContext()
-        attrbldr = api.llvm.AttrBuilder.new()
-        attrbldr.addAttribute(attr)
-        attrs = api.llvm.Attributes.get(context, attrbldr)
-        self._ptr.removeAttr(attrs)
+        def remove_attribute(self, attr):
+            context = api.llvm.getGlobalContext()
+            attrbldr = api.llvm.AttrBuilder.new()
+            attrbldr.addAttribute(attr)
+            attrs = api.llvm.AttributeSet.get(context, 0, attrbldr)
+            self._ptr.removeAttr(attrs)
 
-    def _set_alignment(self, align):
-        context = api.llvm.getGlobalContext()
-        attrbldr = api.llvm.AttrBuilder.new()
-        attrbldr.addAlignmentAttr(align)
-        attrs = api.llvm.Attributes.get(context, attrbldr)
-        self._ptr.addAttr(attrs)
+        def _set_alignment(self, align):
+            context = api.llvm.getGlobalContext()
+            attrbldr = api.llvm.AttrBuilder.new()
+            attrbldr.addAlignmentAttr(align)
+            attrs = api.llvm.AttributeSet.get(context, 0, attrbldr)
+            self._ptr.addAttr(attrs)
+    else:
+        def add_attribute(self, attr):
+            context = api.llvm.getGlobalContext()
+            attrbldr = api.llvm.AttrBuilder.new()
+            attrbldr.addAttribute(attr)
+            attrs = api.llvm.Attributes.get(context, attrbldr)
+            self._ptr.addAttr(attrs)
+            if attr not in self:
+                raise ValueError("Attribute %r is not valid for arg %s" %
+                                 (attr, self))
+
+        def remove_attribute(self, attr):
+            context = api.llvm.getGlobalContext()
+            attrbldr = api.llvm.AttrBuilder.new()
+            attrbldr.addAttribute(attr)
+            attrs = api.llvm.Attributes.get(context, attrbldr)
+            self._ptr.removeAttr(attrs)
+
+        def _set_alignment(self, align):
+            context = api.llvm.getGlobalContext()
+            attrbldr = api.llvm.AttrBuilder.new()
+            attrbldr.addAlignmentAttr(align)
+            attrs = api.llvm.Attributes.get(context, attrbldr)
+            self._ptr.addAttr(attrs)
 
     def _get_alignment(self):
         return self._ptr.getParamAlignment()
@@ -1816,6 +1851,9 @@ class Instruction(User):
     def erase_from_parent(self):
         return self._ptr.eraseFromParent()
 
+    def replace_all_uses_with(self, inst):
+        self._ptr.replaceAllUsesWith(inst)
+
 
 class CallOrInvokeInstruction(Instruction):
     _type_ = api.llvm.CallInst, api.llvm.InvokeInst
@@ -2092,29 +2130,34 @@ class Builder(llvm.Wrapper):
 
     # arithmethic, bitwise and logical
 
-    def add(self, lhs, rhs, name=""):
-        return _make_value(self._ptr.CreateAdd(lhs._ptr, rhs._ptr, name))
+    def add(self, lhs, rhs, name="", nuw=False, nsw=False):
+        return _make_value(self._ptr.CreateAdd(lhs._ptr, rhs._ptr, name,
+                                               nuw, nsw))
 
     def fadd(self, lhs, rhs, name=""):
         return _make_value(self._ptr.CreateFAdd(lhs._ptr, rhs._ptr, name))
 
-    def sub(self, lhs, rhs, name=""):
-        return _make_value(self._ptr.CreateSub(lhs._ptr, rhs._ptr, name))
+    def sub(self, lhs, rhs, name="", nuw=False, nsw=False):
+        return _make_value(self._ptr.CreateSub(lhs._ptr, rhs._ptr, name,
+                                               nuw, nsw))
 
     def fsub(self, lhs, rhs, name=""):
         return _make_value(self._ptr.CreateFSub(lhs._ptr, rhs._ptr, name))
 
-    def mul(self, lhs, rhs, name=""):
-        return _make_value(self._ptr.CreateMul(lhs._ptr, rhs._ptr, name))
+    def mul(self, lhs, rhs, name="", nuw=False, nsw=False):
+        return _make_value(self._ptr.CreateMul(lhs._ptr, rhs._ptr, name,
+                                               nuw, nsw))
 
     def fmul(self, lhs, rhs, name=""):
         return _make_value(self._ptr.CreateFMul(lhs._ptr, rhs._ptr, name))
 
-    def udiv(self, lhs, rhs, name=""):
-        return _make_value(self._ptr.CreateUDiv(lhs._ptr, rhs._ptr, name))
+    def udiv(self, lhs, rhs, name="", exact=False):
+        return _make_value(self._ptr.CreateUDiv(lhs._ptr, rhs._ptr, name,
+                                                exact))
 
-    def sdiv(self, lhs, rhs, name=""):
-        return _make_value(self._ptr.CreateSDiv(lhs._ptr, rhs._ptr, name))
+    def sdiv(self, lhs, rhs, name="", exact=False):
+        return _make_value(self._ptr.CreateSDiv(lhs._ptr, rhs._ptr, name,
+                                                exact))
 
     def fdiv(self, lhs, rhs, name=""):
         return _make_value(self._ptr.CreateFDiv(lhs._ptr, rhs._ptr, name))
@@ -2128,14 +2171,17 @@ class Builder(llvm.Wrapper):
     def frem(self, lhs, rhs, name=""):
         return _make_value(self._ptr.CreateFRem(lhs._ptr, rhs._ptr, name))
 
-    def shl(self, lhs, rhs, name=""):
-        return _make_value(self._ptr.CreateShl(lhs._ptr, rhs._ptr, name))
+    def shl(self, lhs, rhs, name="", nuw=False, nsw=False):
+        return _make_value(self._ptr.CreateShl(lhs._ptr, rhs._ptr, name,
+                                               nuw, nsw))
 
-    def lshr(self, lhs, rhs, name=""):
-        return _make_value(self._ptr.CreateLShr(lhs._ptr, rhs._ptr, name))
+    def lshr(self, lhs, rhs, name="", exact=False):
+        return _make_value(self._ptr.CreateLShr(lhs._ptr, rhs._ptr, name,
+                                                exact))
 
-    def ashr(self, lhs, rhs, name=""):
-        return _make_value(self._ptr.CreateAShr(lhs._ptr, rhs._ptr, name))
+    def ashr(self, lhs, rhs, name="", exact=False):
+        return _make_value(self._ptr.CreateAShr(lhs._ptr, rhs._ptr, name,
+                                                exact))
 
     def and_(self, lhs, rhs, name=""):
         return _make_value(self._ptr.CreateAnd(lhs._ptr, rhs._ptr, name))
@@ -2146,8 +2192,8 @@ class Builder(llvm.Wrapper):
     def xor(self, lhs, rhs, name=""):
         return _make_value(self._ptr.CreateXor(lhs._ptr, rhs._ptr, name))
 
-    def neg(self, val, name=""):
-        return _make_value(self._ptr.CreateNeg(val._ptr, name))
+    def neg(self, val, name="", nuw=False, nsw=False):
+        return _make_value(self._ptr.CreateNeg(val._ptr, name, nuw, nsw))
 
     def not_(self, val, name=""):
         return _make_value(self._ptr.CreateNot(val._ptr, name))
@@ -2422,4 +2468,8 @@ if api.llvm.InitializeNativeTargetAsmPrinter():
     # should this be an optional feature?
     # should user trigger the initialization?
     raise llvm.LLVMException("No native asm printer!?")
-
+if api.llvm.InitializeNativeTargetAsmParser():
+    # required by MCJIT?
+    # should this be an optional feature?
+    # should user trigger the initialization?
+    raise llvm.LLVMException("No native asm parser!?")
diff --git a/llvm/ee.py b/llvm/ee.py
index 7ea90d0..b91c4cb 100644
--- a/llvm/ee.py
+++ b/llvm/ee.py
@@ -30,6 +30,7 @@
 
 "Execution Engine and related classes."
 
+import sys
 from io import BytesIO
 import contextlib
 
@@ -150,9 +151,17 @@ class EngineBuilder(llvm.Wrapper):
         '''
         if tm is not None:
             engine = self._ptr.create(tm._ptr)
+        elif (sys.platform.startswith('win32') and
+                    getattr(self, '_use_mcjit', False)):
+            # force ELF generation on MCJIT on win32
+            triple = get_default_triple()
+            tm = TargetMachine.new('%s-elf' % triple)
+            engine = self._ptr.create(tm._ptr)
         else:
             engine = self._ptr.create()
-        return ExecutionEngine(engine)
+        ee = ExecutionEngine(engine)
+        ee.finalize_object()                # no effect for legacy JIT
+        return ee
 
     def select_target(self, *args):
         '''get the corresponding target machine
@@ -171,6 +180,7 @@ class EngineBuilder(llvm.Wrapper):
         '''Enable/disable MCJIT
         '''
         self._ptr.setUseMCJIT(enable)
+        self._use_mcjit = True
         return self
 
 #===----------------------------------------------------------------------===
@@ -193,6 +203,9 @@ class ExecutionEngine(llvm.Wrapper):
         ptr = self._ptr.runFunction(fn._ptr, list(map(lambda x: x._ptr, args)))
         return GenericValue(ptr)
 
+    def get_pointer_to_named_function(self, name, abort=True):
+        return self._ptr.getPointerToNamedFunction(name, abort)
+
     def get_pointer_to_function(self, fn):
         return self._ptr.getPointerToFunction(fn._ptr)
 
@@ -218,6 +231,9 @@ class ExecutionEngine(llvm.Wrapper):
     def remove_module(self, module):
         return self._ptr.removeModule(module._ptr)
 
+    def finalize_object(self):
+        return self._ptr.finalizeObject()
+
     @property
     def target_data(self):
         ptr = self._ptr.getDataLayout()
@@ -232,7 +248,7 @@ def initialize_target(target, noraise=False):
     It is safe to initialize the same target multiple times.
     """
     prefix = 'LLVMInitialize'
-    postfixes = ['Target', 'TargetInfo', 'TargetMC', 'AsmPrinter']
+    postfixes = ['Target', 'TargetInfo', 'TargetMC', 'AsmPrinter', 'AsmParser']
     try:
         for postfix in postfixes:
             getattr(api, '%s%s%s' % (prefix, target, postfix))()
@@ -274,7 +290,7 @@ class TargetMachine(llvm.Wrapper):
         with contextlib.closing(BytesIO()) as error:
             target = api.llvm.TargetRegistry.lookupTarget(triple, error)
             if not target:
-                raise llvm.LLVMException(error)
+                raise llvm.LLVMException(error.getvalue())
             if not target.hasTargetMachine():
                 raise llvm.LLVMException(target, "No target machine.")
             target_options = api.llvm.TargetOptions.new()
@@ -366,3 +382,14 @@ class TargetMachine(llvm.Wrapper):
     def feature_string(self):
         return self._ptr.getTargetFeatureString()
 
+
+
+#===----------------------------------------------------------------------===
+# Dynamic Library
+#===----------------------------------------------------------------------===
+
+def dylib_add_symbol(name, ptr):
+    api.llvm.sys.DynamicLibrary.AddSymbol(name, ptr)
+
+def dylib_address_of_symbol(name):
+    return api.llvm.sys.DynamicLibrary.SearchForAddressOfSymbol(name)
diff --git a/llvm/llrt.py b/llvm/llrt.py
new file mode 100644
index 0000000..13402d6
--- /dev/null
+++ b/llvm/llrt.py
@@ -0,0 +1,79 @@
+import os
+import llvm.core as lc
+import llvm.passes as lp
+import llvm.ee as le
+
+def replace_divmod64(lfunc):
+    '''Replaces all 64-bit integer division (sdiv, udiv) and modulo (srem, urem)
+    '''
+    int64 = lc.Type.int(64)
+    int64ptr = lc.Type.pointer(lc.Type.int(64))
+
+    functy = lc.Type.function(int64, [int64, int64])
+    udiv64 = lfunc.module.get_or_insert_function(functy, '__llrt_udiv64')
+    sdiv64 = lfunc.module.get_or_insert_function(functy, '__llrt_sdiv64')
+    umod64 = lfunc.module.get_or_insert_function(functy, '__llrt_umod64')
+    smod64 = lfunc.module.get_or_insert_function(functy, '__llrt_smod64')
+
+    builder = lc.Builder.new(lfunc.entry_basic_block)
+    for bb in lfunc.basic_blocks:
+        for inst in bb.instructions:
+            if inst.opcode_name == 'sdiv' and inst.type == int64:
+                _replace_with(builder, inst, sdiv64)
+            elif inst.opcode_name == 'udiv' and inst.type == int64:
+                _replace_with(builder, inst, udiv64)
+            elif inst.opcode_name == 'srem' and inst.type == int64:
+                _replace_with(builder, inst, smod64)
+            elif inst.opcode_name == 'urem' and inst.type == int64:
+                _replace_with(builder, inst, umod64)
+
+def _replace_with(builder, inst, func):
+    '''Replace instruction with a call to the function with the same operands
+    as arguments.
+    '''
+    builder.position_before(inst)
+    replacement = builder.call(func, inst.operands)
+    inst.replace_all_uses_with(replacement._ptr)
+    inst.erase_from_parent()
+
+def load(arch):
+    '''Load the LLRT module corresponding to the given architecture
+    Creates a new module and optimizes it using the information from 
+    the host machine.
+    '''
+    if arch != 'x86_64':
+        arch = 'x86'
+    path = os.path.join(os.path.dirname(__file__), 'llrt', 'llrt_%s.ll' % arch)
+    with open(path) as fin:
+        lib = lc.Module.from_assembly(fin)
+
+    # run passes to optimize
+    tm = le.TargetMachine.new()
+    pms = lp.build_pass_managers(tm, opt=3, fpm=False)
+    pms.pm.run(lib)
+    return lib
+
+class LLRT(object):
+    def __init__(self):
+        arch = le.get_default_triple().split('-', 1)[0]
+        self.module = load(arch)
+        self.engine = le.EngineBuilder.new(self.module).opt(3).create()
+        self.installed_symbols = set()
+
+    def install_symbols(self):
+        '''Bind all the external symbols to the global symbol map.
+        Any future reference to these symbols will be automatically resolved
+        by LLVM.
+        '''
+        for lfunc in self.module.functions:
+            if lfunc.linkage == lc.LINKAGE_EXTERNAL:
+                mangled = '__llrt_' + lfunc.name
+                self.installed_symbols.add(mangled)
+                ptr = self.engine.get_pointer_to_function(lfunc)
+                le.dylib_add_symbol(mangled, ptr)
+
+    def uninstall_symbols(self):
+        for sym in self.installed_symbols:
+            le.dylib_add_symbol(sym, 0)
+
+
diff --git a/llvm/llrt/llrt_x86.ll b/llvm/llrt/llrt_x86.ll
new file mode 100644
index 0000000..5096214
--- /dev/null
+++ b/llvm/llrt/llrt_x86.ll
@@ -0,0 +1,371 @@
+; ModuleID = 'udivmod64_x86.bc'
+target datalayout = "e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64-f32:32:32-f64:32:64-v64:64:64-v128:128:128-a0:0:64-f80:128:128-n8:16:32-S128"
+
+%struct.div_state_ = type { i64, i64 }
+
+define i64 @udivmod64(i64 %dividend, i64 %divisor, i64* %remainder) nounwind ssp {
+  %1 = alloca i64, align 4
+  %2 = alloca i64, align 8
+  %3 = alloca i64, align 8
+  %4 = alloca i64*, align 4
+  %state = alloca %struct.div_state_, align 4
+  %quotient = alloca i64, align 8
+  %i = alloca i32, align 4
+  %skipahead = alloca i32, align 4
+  store i64 %dividend, i64* %2, align 8
+  store i64 %divisor, i64* %3, align 8
+  store i64* %remainder, i64** %4, align 4
+  %5 = getelementptr inbounds %struct.div_state_* %state, i32 0, i32 0
+  store i64 0, i64* %5, align 4
+  %6 = getelementptr inbounds %struct.div_state_* %state, i32 0, i32 1
+  %7 = load i64* %2, align 8
+  store i64 %7, i64* %6, align 4
+  store i64 0, i64* %quotient, align 8
+  %8 = load i64* %3, align 8
+  %9 = icmp eq i64 %8, 0
+  br i1 %9, label %10, label %11
+
+; <label>:10                                      ; preds = %0
+  store i64 0, i64* %1
+  br label %57
+
+; <label>:11                                      ; preds = %0
+  %12 = load i64* %2, align 8
+  %13 = call i32 @clz64(i64 %12)
+  store i32 %13, i32* %skipahead, align 4
+  store i32 0, i32* %i, align 4
+  br label %14
+
+; <label>:14                                      ; preds = %19, %11
+  %15 = load i32* %i, align 4
+  %16 = load i32* %skipahead, align 4
+  %17 = icmp slt i32 %15, %16
+  br i1 %17, label %18, label %22
+
+; <label>:18                                      ; preds = %14
+  call void @div_state_lshift(%struct.div_state_* %state)
+  br label %19
+
+; <label>:19                                      ; preds = %18
+  %20 = load i32* %i, align 4
+  %21 = add nsw i32 %20, 1
+  store i32 %21, i32* %i, align 4
+  br label %14
+
+; <label>:22                                      ; preds = %14
+  %23 = load i32* %skipahead, align 4
+  store i32 %23, i32* %i, align 4
+  br label %24
+
+; <label>:24                                      ; preds = %45, %22
+  %25 = load i32* %i, align 4
+  %26 = icmp slt i32 %25, 64
+  br i1 %26, label %27, label %48
+
+; <label>:27                                      ; preds = %24
+  call void @div_state_lshift(%struct.div_state_* %state)
+  %28 = getelementptr inbounds %struct.div_state_* %state, i32 0, i32 0
+  %29 = load i64* %28, align 4
+  %30 = load i64* %3, align 8
+  %31 = icmp uge i64 %29, %30
+  br i1 %31, label %32, label %44
+
+; <label>:32                                      ; preds = %27
+  %33 = getelementptr inbounds %struct.div_state_* %state, i32 0, i32 0
+  %34 = load i64* %33, align 4
+  %35 = load i64* %3, align 8
+  %36 = sub i64 %34, %35
+  %37 = getelementptr inbounds %struct.div_state_* %state, i32 0, i32 0
+  store i64 %36, i64* %37, align 4
+  %38 = load i32* %i, align 4
+  %39 = sub nsw i32 63, %38
+  %40 = zext i32 %39 to i64
+  %41 = shl i64 1, %40
+  %42 = load i64* %quotient, align 8
+  %43 = or i64 %42, %41
+  store i64 %43, i64* %quotient, align 8
+  br label %44
+
+; <label>:44                                      ; preds = %32, %27
+  br label %45
+
+; <label>:45                                      ; preds = %44
+  %46 = load i32* %i, align 4
+  %47 = add nsw i32 %46, 1
+  store i32 %47, i32* %i, align 4
+  br label %24
+
+; <label>:48                                      ; preds = %24
+  %49 = load i64** %4, align 4
+  %50 = icmp ne i64* %49, null
+  br i1 %50, label %51, label %55
+
+; <label>:51                                      ; preds = %48
+  %52 = getelementptr inbounds %struct.div_state_* %state, i32 0, i32 0
+  %53 = load i64* %52, align 4
+  %54 = load i64** %4, align 4
+  store i64 %53, i64* %54, align 4
+  br label %55
+
+; <label>:55                                      ; preds = %51, %48
+  %56 = load i64* %quotient, align 8
+  store i64 %56, i64* %1
+  br label %57
+
+; <label>:57                                      ; preds = %55, %10
+  %58 = load i64* %1
+  ret i64 %58
+}
+
+define internal i32 @clz64(i64 %x) nounwind ssp {
+  %1 = alloca i64, align 8
+  %total_bits = alloca i32, align 4
+  %zc = alloca i32, align 4
+  store i64 %x, i64* %1, align 8
+  store i32 64, i32* %total_bits, align 4
+  store i32 0, i32* %zc, align 4
+  br label %2
+
+; <label>:2                                       ; preds = %16, %0
+  %3 = load i32* %zc, align 4
+  %4 = icmp slt i32 %3, 64
+  br i1 %4, label %5, label %14
+
+; <label>:5                                       ; preds = %2
+  %6 = load i64* %1, align 8
+  %7 = load i32* %zc, align 4
+  %8 = sub nsw i32 64, %7
+  %9 = sub nsw i32 %8, 1
+  %10 = zext i32 %9 to i64
+  %11 = lshr i64 %6, %10
+  %12 = and i64 %11, 1
+  %13 = icmp eq i64 %12, 0
+  br label %14
+
+; <label>:14                                      ; preds = %5, %2
+  %15 = phi i1 [ false, %2 ], [ %13, %5 ]
+  br i1 %15, label %16, label %19
+
+; <label>:16                                      ; preds = %14
+  %17 = load i32* %zc, align 4
+  %18 = add nsw i32 %17, 1
+  store i32 %18, i32* %zc, align 4
+  br label %2
+
+; <label>:19                                      ; preds = %14
+  %20 = load i32* %zc, align 4
+  ret i32 %20
+}
+
+define internal void @div_state_lshift(%struct.div_state_* %state) nounwind ssp {
+  %1 = alloca %struct.div_state_*, align 4
+  store %struct.div_state_* %state, %struct.div_state_** %1, align 4
+  %2 = load %struct.div_state_** %1, align 4
+  %3 = getelementptr inbounds %struct.div_state_* %2, i32 0, i32 0
+  %4 = load i64* %3, align 4
+  %5 = shl i64 %4, 1
+  %6 = load %struct.div_state_** %1, align 4
+  %7 = getelementptr inbounds %struct.div_state_* %6, i32 0, i32 1
+  %8 = load i64* %7, align 4
+  %9 = lshr i64 %8, 63
+  %10 = or i64 %5, %9
+  %11 = load %struct.div_state_** %1, align 4
+  %12 = getelementptr inbounds %struct.div_state_* %11, i32 0, i32 0
+  store i64 %10, i64* %12, align 4
+  %13 = load %struct.div_state_** %1, align 4
+  %14 = getelementptr inbounds %struct.div_state_* %13, i32 0, i32 1
+  %15 = load i64* %14, align 4
+  %16 = shl i64 %15, 1
+  %17 = load %struct.div_state_** %1, align 4
+  %18 = getelementptr inbounds %struct.div_state_* %17, i32 0, i32 1
+  store i64 %16, i64* %18, align 4
+  ret void
+}
+
+define i64 @sdivmod64(i64 %dividend, i64 %divisor, i64* %remainder) nounwind ssp {
+  %1 = alloca i64, align 8
+  %2 = alloca i64, align 8
+  %3 = alloca i64*, align 4
+  %signbitidx = alloca i32, align 4
+  %signed_dividend = alloca i32, align 4
+  %signed_divisor = alloca i32, align 4
+  %signed_result = alloca i32, align 4
+  %quotient = alloca i64, align 8
+  %udvd = alloca i64, align 8
+  %udvr = alloca i64, align 8
+  %uquotient = alloca i64, align 8
+  %uremainder = alloca i64, align 8
+  store i64 %dividend, i64* %1, align 8
+  store i64 %divisor, i64* %2, align 8
+  store i64* %remainder, i64** %3, align 4
+  store i32 63, i32* %signbitidx, align 4
+  %4 = load i64* %1, align 8
+  %5 = icmp slt i64 %4, 0
+  %6 = zext i1 %5 to i32
+  store i32 %6, i32* %signed_dividend, align 4
+  %7 = load i64* %2, align 8
+  %8 = icmp slt i64 %7, 0
+  %9 = zext i1 %8 to i32
+  store i32 %9, i32* %signed_divisor, align 4
+  %10 = load i32* %signed_divisor, align 4
+  %11 = load i32* %signed_dividend, align 4
+  %12 = xor i32 %10, %11
+  store i32 %12, i32* %signed_result, align 4
+  %13 = load i32* %signed_dividend, align 4
+  %14 = icmp ne i32 %13, 0
+  br i1 %14, label %15, label %18
+
+; <label>:15                                      ; preds = %0
+  %16 = load i64* %1, align 8
+  %17 = sub nsw i64 0, %16
+  br label %20
+
+; <label>:18                                      ; preds = %0
+  %19 = load i64* %1, align 8
+  br label %20
+
+; <label>:20                                      ; preds = %18, %15
+  %21 = phi i64 [ %17, %15 ], [ %19, %18 ]
+  store i64 %21, i64* %udvd, align 8
+  %22 = load i32* %signed_divisor, align 4
+  %23 = icmp ne i32 %22, 0
+  br i1 %23, label %24, label %27
+
+; <label>:24                                      ; preds = %20
+  %25 = load i64* %2, align 8
+  %26 = sub nsw i64 0, %25
+  br label %29
+
+; <label>:27                                      ; preds = %20
+  %28 = load i64* %2, align 8
+  br label %29
+
+; <label>:29                                      ; preds = %27, %24
+  %30 = phi i64 [ %26, %24 ], [ %28, %27 ]
+  store i64 %30, i64* %udvr, align 8
+  %31 = load i64* %udvd, align 8
+  %32 = load i64* %udvr, align 8
+  %33 = call i64 @udivmod64(i64 %31, i64 %32, i64* %uremainder)
+  store i64 %33, i64* %uquotient, align 8
+  %34 = load i32* %signed_result, align 4
+  %35 = icmp ne i32 %34, 0
+  br i1 %35, label %36, label %57
+
+; <label>:36                                      ; preds = %29
+  %37 = load i64* %uremainder, align 8
+  %38 = icmp ne i64 %37, 0
+  br i1 %38, label %39, label %43
+
+; <label>:39                                      ; preds = %36
+  %40 = load i64* %uquotient, align 8
+  %41 = sub nsw i64 0, %40
+  %42 = sub nsw i64 %41, 1
+  store i64 %42, i64* %quotient, align 8
+  br label %46
+
+; <label>:43                                      ; preds = %36
+  %44 = load i64* %uquotient, align 8
+  %45 = sub nsw i64 0, %44
+  store i64 %45, i64* %quotient, align 8
+  br label %46
+
+; <label>:46                                      ; preds = %43, %39
+  %47 = load i64** %3, align 4
+  %48 = icmp ne i64* %47, null
+  br i1 %48, label %49, label %56
+
+; <label>:49                                      ; preds = %46
+  %50 = load i64* %1, align 8
+  %51 = load i64* %quotient, align 8
+  %52 = load i64* %2, align 8
+  %53 = mul i64 %51, %52
+  %54 = sub i64 %50, %53
+  %55 = load i64** %3, align 4
+  store i64 %54, i64* %55, align 4
+  br label %56
+
+; <label>:56                                      ; preds = %49, %46
+  br label %73
+
+; <label>:57                                      ; preds = %29
+  %58 = load i64* %uquotient, align 8
+  store i64 %58, i64* %quotient, align 8
+  %59 = load i64** %3, align 4
+  %60 = icmp ne i64* %59, null
+  br i1 %60, label %61, label %72
+
+; <label>:61                                      ; preds = %57
+  %62 = load i32* %signed_divisor, align 4
+  %63 = icmp ne i32 %62, 0
+  br i1 %63, label %64, label %67
+
+; <label>:64                                      ; preds = %61
+  %65 = load i64* %uremainder, align 8
+  %66 = sub i64 0, %65
+  br label %69
+
+; <label>:67                                      ; preds = %61
+  %68 = load i64* %uremainder, align 8
+  br label %69
+
+; <label>:69                                      ; preds = %67, %64
+  %70 = phi i64 [ %66, %64 ], [ %68, %67 ]
+  %71 = load i64** %3, align 4
+  store i64 %70, i64* %71, align 4
+  br label %72
+
+; <label>:72                                      ; preds = %69, %57
+  br label %73
+
+; <label>:73                                      ; preds = %72, %56
+  %74 = load i64* %quotient, align 8
+  ret i64 %74
+}
+
+define i64 @udiv64(i64 %dividend, i64 %divisor) nounwind ssp {
+  %1 = alloca i64, align 8
+  %2 = alloca i64, align 8
+  store i64 %dividend, i64* %1, align 8
+  store i64 %divisor, i64* %2, align 8
+  %3 = load i64* %1, align 8
+  %4 = load i64* %2, align 8
+  %5 = call i64 @udivmod64(i64 %3, i64 %4, i64* null)
+  ret i64 %5
+}
+
+define i64 @sdiv64(i64 %dividend, i64 %divisor) nounwind ssp {
+  %1 = alloca i64, align 8
+  %2 = alloca i64, align 8
+  store i64 %dividend, i64* %1, align 8
+  store i64 %divisor, i64* %2, align 8
+  %3 = load i64* %1, align 8
+  %4 = load i64* %2, align 8
+  %5 = call i64 @sdivmod64(i64 %3, i64 %4, i64* null)
+  ret i64 %5
+}
+
+define i64 @umod64(i64 %dividend, i64 %divisor) nounwind ssp {
+  %1 = alloca i64, align 8
+  %2 = alloca i64, align 8
+  %rem = alloca i64, align 8
+  store i64 %dividend, i64* %1, align 8
+  store i64 %divisor, i64* %2, align 8
+  %3 = load i64* %1, align 8
+  %4 = load i64* %2, align 8
+  %5 = call i64 @udivmod64(i64 %3, i64 %4, i64* %rem)
+  %6 = load i64* %rem, align 8
+  ret i64 %6
+}
+
+define i64 @smod64(i64 %dividend, i64 %divisor) nounwind ssp {
+  %1 = alloca i64, align 8
+  %2 = alloca i64, align 8
+  %rem = alloca i64, align 8
+  store i64 %dividend, i64* %1, align 8
+  store i64 %divisor, i64* %2, align 8
+  %3 = load i64* %1, align 8
+  %4 = load i64* %2, align 8
+  %5 = call i64 @sdivmod64(i64 %3, i64 %4, i64* %rem)
+  %6 = load i64* %rem, align 8
+  ret i64 %6
+}
diff --git a/llvm/llrt/llrt_x86_64.ll b/llvm/llrt/llrt_x86_64.ll
new file mode 100644
index 0000000..6d9af01
--- /dev/null
+++ b/llvm/llrt/llrt_x86_64.ll
@@ -0,0 +1,371 @@
+; ModuleID = 'udivmod64_x86_64.bc'
+target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64-S128"
+
+%struct.div_state_ = type { i64, i64 }
+
+define i64 @udivmod64(i64 %dividend, i64 %divisor, i64* %remainder) nounwind uwtable ssp {
+  %1 = alloca i64, align 8
+  %2 = alloca i64, align 8
+  %3 = alloca i64, align 8
+  %4 = alloca i64*, align 8
+  %state = alloca %struct.div_state_, align 8
+  %quotient = alloca i64, align 8
+  %i = alloca i32, align 4
+  %skipahead = alloca i32, align 4
+  store i64 %dividend, i64* %2, align 8
+  store i64 %divisor, i64* %3, align 8
+  store i64* %remainder, i64** %4, align 8
+  %5 = getelementptr inbounds %struct.div_state_* %state, i32 0, i32 0
+  store i64 0, i64* %5, align 8
+  %6 = getelementptr inbounds %struct.div_state_* %state, i32 0, i32 1
+  %7 = load i64* %2, align 8
+  store i64 %7, i64* %6, align 8
+  store i64 0, i64* %quotient, align 8
+  %8 = load i64* %3, align 8
+  %9 = icmp eq i64 %8, 0
+  br i1 %9, label %10, label %11
+
+; <label>:10                                      ; preds = %0
+  store i64 0, i64* %1
+  br label %57
+
+; <label>:11                                      ; preds = %0
+  %12 = load i64* %2, align 8
+  %13 = call i32 @clz64(i64 %12)
+  store i32 %13, i32* %skipahead, align 4
+  store i32 0, i32* %i, align 4
+  br label %14
+
+; <label>:14                                      ; preds = %19, %11
+  %15 = load i32* %i, align 4
+  %16 = load i32* %skipahead, align 4
+  %17 = icmp slt i32 %15, %16
+  br i1 %17, label %18, label %22
+
+; <label>:18                                      ; preds = %14
+  call void @div_state_lshift(%struct.div_state_* %state)
+  br label %19
+
+; <label>:19                                      ; preds = %18
+  %20 = load i32* %i, align 4
+  %21 = add nsw i32 %20, 1
+  store i32 %21, i32* %i, align 4
+  br label %14
+
+; <label>:22                                      ; preds = %14
+  %23 = load i32* %skipahead, align 4
+  store i32 %23, i32* %i, align 4
+  br label %24
+
+; <label>:24                                      ; preds = %45, %22
+  %25 = load i32* %i, align 4
+  %26 = icmp slt i32 %25, 64
+  br i1 %26, label %27, label %48
+
+; <label>:27                                      ; preds = %24
+  call void @div_state_lshift(%struct.div_state_* %state)
+  %28 = getelementptr inbounds %struct.div_state_* %state, i32 0, i32 0
+  %29 = load i64* %28, align 8
+  %30 = load i64* %3, align 8
+  %31 = icmp uge i64 %29, %30
+  br i1 %31, label %32, label %44
+
+; <label>:32                                      ; preds = %27
+  %33 = getelementptr inbounds %struct.div_state_* %state, i32 0, i32 0
+  %34 = load i64* %33, align 8
+  %35 = load i64* %3, align 8
+  %36 = sub i64 %34, %35
+  %37 = getelementptr inbounds %struct.div_state_* %state, i32 0, i32 0
+  store i64 %36, i64* %37, align 8
+  %38 = load i32* %i, align 4
+  %39 = sub nsw i32 63, %38
+  %40 = zext i32 %39 to i64
+  %41 = shl i64 1, %40
+  %42 = load i64* %quotient, align 8
+  %43 = or i64 %42, %41
+  store i64 %43, i64* %quotient, align 8
+  br label %44
+
+; <label>:44                                      ; preds = %32, %27
+  br label %45
+
+; <label>:45                                      ; preds = %44
+  %46 = load i32* %i, align 4
+  %47 = add nsw i32 %46, 1
+  store i32 %47, i32* %i, align 4
+  br label %24
+
+; <label>:48                                      ; preds = %24
+  %49 = load i64** %4, align 8
+  %50 = icmp ne i64* %49, null
+  br i1 %50, label %51, label %55
+
+; <label>:51                                      ; preds = %48
+  %52 = getelementptr inbounds %struct.div_state_* %state, i32 0, i32 0
+  %53 = load i64* %52, align 8
+  %54 = load i64** %4, align 8
+  store i64 %53, i64* %54, align 8
+  br label %55
+
+; <label>:55                                      ; preds = %51, %48
+  %56 = load i64* %quotient, align 8
+  store i64 %56, i64* %1
+  br label %57
+
+; <label>:57                                      ; preds = %55, %10
+  %58 = load i64* %1
+  ret i64 %58
+}
+
+define internal i32 @clz64(i64 %x) nounwind uwtable ssp {
+  %1 = alloca i64, align 8
+  %total_bits = alloca i32, align 4
+  %zc = alloca i32, align 4
+  store i64 %x, i64* %1, align 8
+  store i32 64, i32* %total_bits, align 4
+  store i32 0, i32* %zc, align 4
+  br label %2
+
+; <label>:2                                       ; preds = %16, %0
+  %3 = load i32* %zc, align 4
+  %4 = icmp slt i32 %3, 64
+  br i1 %4, label %5, label %14
+
+; <label>:5                                       ; preds = %2
+  %6 = load i64* %1, align 8
+  %7 = load i32* %zc, align 4
+  %8 = sub nsw i32 64, %7
+  %9 = sub nsw i32 %8, 1
+  %10 = zext i32 %9 to i64
+  %11 = lshr i64 %6, %10
+  %12 = and i64 %11, 1
+  %13 = icmp eq i64 %12, 0
+  br label %14
+
+; <label>:14                                      ; preds = %5, %2
+  %15 = phi i1 [ false, %2 ], [ %13, %5 ]
+  br i1 %15, label %16, label %19
+
+; <label>:16                                      ; preds = %14
+  %17 = load i32* %zc, align 4
+  %18 = add nsw i32 %17, 1
+  store i32 %18, i32* %zc, align 4
+  br label %2
+
+; <label>:19                                      ; preds = %14
+  %20 = load i32* %zc, align 4
+  ret i32 %20
+}
+
+define internal void @div_state_lshift(%struct.div_state_* %state) nounwind uwtable ssp {
+  %1 = alloca %struct.div_state_*, align 8
+  store %struct.div_state_* %state, %struct.div_state_** %1, align 8
+  %2 = load %struct.div_state_** %1, align 8
+  %3 = getelementptr inbounds %struct.div_state_* %2, i32 0, i32 0
+  %4 = load i64* %3, align 8
+  %5 = shl i64 %4, 1
+  %6 = load %struct.div_state_** %1, align 8
+  %7 = getelementptr inbounds %struct.div_state_* %6, i32 0, i32 1
+  %8 = load i64* %7, align 8
+  %9 = lshr i64 %8, 63
+  %10 = or i64 %5, %9
+  %11 = load %struct.div_state_** %1, align 8
+  %12 = getelementptr inbounds %struct.div_state_* %11, i32 0, i32 0
+  store i64 %10, i64* %12, align 8
+  %13 = load %struct.div_state_** %1, align 8
+  %14 = getelementptr inbounds %struct.div_state_* %13, i32 0, i32 1
+  %15 = load i64* %14, align 8
+  %16 = shl i64 %15, 1
+  %17 = load %struct.div_state_** %1, align 8
+  %18 = getelementptr inbounds %struct.div_state_* %17, i32 0, i32 1
+  store i64 %16, i64* %18, align 8
+  ret void
+}
+
+define i64 @sdivmod64(i64 %dividend, i64 %divisor, i64* %remainder) nounwind uwtable ssp {
+  %1 = alloca i64, align 8
+  %2 = alloca i64, align 8
+  %3 = alloca i64*, align 8
+  %signbitidx = alloca i32, align 4
+  %signed_dividend = alloca i32, align 4
+  %signed_divisor = alloca i32, align 4
+  %signed_result = alloca i32, align 4
+  %quotient = alloca i64, align 8
+  %udvd = alloca i64, align 8
+  %udvr = alloca i64, align 8
+  %uquotient = alloca i64, align 8
+  %uremainder = alloca i64, align 8
+  store i64 %dividend, i64* %1, align 8
+  store i64 %divisor, i64* %2, align 8
+  store i64* %remainder, i64** %3, align 8
+  store i32 63, i32* %signbitidx, align 4
+  %4 = load i64* %1, align 8
+  %5 = icmp slt i64 %4, 0
+  %6 = zext i1 %5 to i32
+  store i32 %6, i32* %signed_dividend, align 4
+  %7 = load i64* %2, align 8
+  %8 = icmp slt i64 %7, 0
+  %9 = zext i1 %8 to i32
+  store i32 %9, i32* %signed_divisor, align 4
+  %10 = load i32* %signed_divisor, align 4
+  %11 = load i32* %signed_dividend, align 4
+  %12 = xor i32 %10, %11
+  store i32 %12, i32* %signed_result, align 4
+  %13 = load i32* %signed_dividend, align 4
+  %14 = icmp ne i32 %13, 0
+  br i1 %14, label %15, label %18
+
+; <label>:15                                      ; preds = %0
+  %16 = load i64* %1, align 8
+  %17 = sub nsw i64 0, %16
+  br label %20
+
+; <label>:18                                      ; preds = %0
+  %19 = load i64* %1, align 8
+  br label %20
+
+; <label>:20                                      ; preds = %18, %15
+  %21 = phi i64 [ %17, %15 ], [ %19, %18 ]
+  store i64 %21, i64* %udvd, align 8
+  %22 = load i32* %signed_divisor, align 4
+  %23 = icmp ne i32 %22, 0
+  br i1 %23, label %24, label %27
+
+; <label>:24                                      ; preds = %20
+  %25 = load i64* %2, align 8
+  %26 = sub nsw i64 0, %25
+  br label %29
+
+; <label>:27                                      ; preds = %20
+  %28 = load i64* %2, align 8
+  br label %29
+
+; <label>:29                                      ; preds = %27, %24
+  %30 = phi i64 [ %26, %24 ], [ %28, %27 ]
+  store i64 %30, i64* %udvr, align 8
+  %31 = load i64* %udvd, align 8
+  %32 = load i64* %udvr, align 8
+  %33 = call i64 @udivmod64(i64 %31, i64 %32, i64* %uremainder)
+  store i64 %33, i64* %uquotient, align 8
+  %34 = load i32* %signed_result, align 4
+  %35 = icmp ne i32 %34, 0
+  br i1 %35, label %36, label %57
+
+; <label>:36                                      ; preds = %29
+  %37 = load i64* %uremainder, align 8
+  %38 = icmp ne i64 %37, 0
+  br i1 %38, label %39, label %43
+
+; <label>:39                                      ; preds = %36
+  %40 = load i64* %uquotient, align 8
+  %41 = sub nsw i64 0, %40
+  %42 = sub nsw i64 %41, 1
+  store i64 %42, i64* %quotient, align 8
+  br label %46
+
+; <label>:43                                      ; preds = %36
+  %44 = load i64* %uquotient, align 8
+  %45 = sub nsw i64 0, %44
+  store i64 %45, i64* %quotient, align 8
+  br label %46
+
+; <label>:46                                      ; preds = %43, %39
+  %47 = load i64** %3, align 8
+  %48 = icmp ne i64* %47, null
+  br i1 %48, label %49, label %56
+
+; <label>:49                                      ; preds = %46
+  %50 = load i64* %1, align 8
+  %51 = load i64* %quotient, align 8
+  %52 = load i64* %2, align 8
+  %53 = mul i64 %51, %52
+  %54 = sub i64 %50, %53
+  %55 = load i64** %3, align 8
+  store i64 %54, i64* %55, align 8
+  br label %56
+
+; <label>:56                                      ; preds = %49, %46
+  br label %73
+
+; <label>:57                                      ; preds = %29
+  %58 = load i64* %uquotient, align 8
+  store i64 %58, i64* %quotient, align 8
+  %59 = load i64** %3, align 8
+  %60 = icmp ne i64* %59, null
+  br i1 %60, label %61, label %72
+
+; <label>:61                                      ; preds = %57
+  %62 = load i32* %signed_divisor, align 4
+  %63 = icmp ne i32 %62, 0
+  br i1 %63, label %64, label %67
+
+; <label>:64                                      ; preds = %61
+  %65 = load i64* %uremainder, align 8
+  %66 = sub i64 0, %65
+  br label %69
+
+; <label>:67                                      ; preds = %61
+  %68 = load i64* %uremainder, align 8
+  br label %69
+
+; <label>:69                                      ; preds = %67, %64
+  %70 = phi i64 [ %66, %64 ], [ %68, %67 ]
+  %71 = load i64** %3, align 8
+  store i64 %70, i64* %71, align 8
+  br label %72
+
+; <label>:72                                      ; preds = %69, %57
+  br label %73
+
+; <label>:73                                      ; preds = %72, %56
+  %74 = load i64* %quotient, align 8
+  ret i64 %74
+}
+
+define i64 @udiv64(i64 %dividend, i64 %divisor) nounwind uwtable ssp {
+  %1 = alloca i64, align 8
+  %2 = alloca i64, align 8
+  store i64 %dividend, i64* %1, align 8
+  store i64 %divisor, i64* %2, align 8
+  %3 = load i64* %1, align 8
+  %4 = load i64* %2, align 8
+  %5 = call i64 @udivmod64(i64 %3, i64 %4, i64* null)
+  ret i64 %5
+}
+
+define i64 @sdiv64(i64 %dividend, i64 %divisor) nounwind uwtable ssp {
+  %1 = alloca i64, align 8
+  %2 = alloca i64, align 8
+  store i64 %dividend, i64* %1, align 8
+  store i64 %divisor, i64* %2, align 8
+  %3 = load i64* %1, align 8
+  %4 = load i64* %2, align 8
+  %5 = call i64 @sdivmod64(i64 %3, i64 %4, i64* null)
+  ret i64 %5
+}
+
+define i64 @umod64(i64 %dividend, i64 %divisor) nounwind uwtable ssp {
+  %1 = alloca i64, align 8
+  %2 = alloca i64, align 8
+  %rem = alloca i64, align 8
+  store i64 %dividend, i64* %1, align 8
+  store i64 %divisor, i64* %2, align 8
+  %3 = load i64* %1, align 8
+  %4 = load i64* %2, align 8
+  %5 = call i64 @udivmod64(i64 %3, i64 %4, i64* %rem)
+  %6 = load i64* %rem, align 8
+  ret i64 %6
+}
+
+define i64 @smod64(i64 %dividend, i64 %divisor) nounwind uwtable ssp {
+  %1 = alloca i64, align 8
+  %2 = alloca i64, align 8
+  %rem = alloca i64, align 8
+  store i64 %dividend, i64* %1, align 8
+  store i64 %divisor, i64* %2, align 8
+  %3 = load i64* %1, align 8
+  %4 = load i64* %2, align 8
+  %5 = call i64 @sdivmod64(i64 %3, i64 %4, i64* %rem)
+  %6 = load i64* %rem, align 8
+  ret i64 %6
+}
diff --git a/llvm/passes.py b/llvm/passes.py
index 068a7a3..4753027 100644
--- a/llvm/passes.py
+++ b/llvm/passes.py
@@ -70,13 +70,24 @@ class PassManagerBuilder(llvm.Wrapper):
     def size_level(self, sizelevel):
         self._ptr.SizeLevel = sizelevel
 
-    @property
-    def vectorize(self):
-        return self._ptr.Vectorize
+    if llvm.version >= (3, 3):
+        @property
+        def bbvectorize(self):
+            return self._ptr.BBVectorize
 
-    @vectorize.setter
-    def vectorize(self, enable):
-        self._ptr.Vectorize = enable
+        @bbvectorize.setter
+        def bbvectorize(self, enable):
+            self._ptr.BBVectorize = enable
+
+        vectorize = bbvectorize
+    else:
+        @property
+        def vectorize(self):
+            return self._ptr.Vectorize
+
+        @vectorize.setter
+        def vectorize(self, enable):
+            self._ptr.Vectorize = enable
 
 
     @property
@@ -331,14 +342,14 @@ def build_pass_managers(tm, opt=2, loop_vectorize=False, vectorize=False,
     if pm:
         pm.add(tm.target_data.clone())
         pm.add(TargetLibraryInfo.new(tm.triple))
-        if llvm.version >= (3, 2):
+        if llvm.version == (3, 2):
             pm.add(TargetTransformInfo.new(tm))
         pmb.populate(pm)
 
     if fpm:
         fpm.add(tm.target_data.clone())
         fpm.add(TargetLibraryInfo.new(tm.triple))
-        if llvm.version >= (3, 2):
+        if llvm.version == (3, 2):
             fpm.add(TargetTransformInfo.new(tm))
         pmb.populate(fpm)
         fpm.initialize()
diff --git a/llvm/test_llvmpy.py b/llvm/test_llvmpy.py
index e21a825..2764cf2 100644
--- a/llvm/test_llvmpy.py
+++ b/llvm/test_llvmpy.py
@@ -1353,6 +1353,235 @@ class TestCmp(TestCase):
 tests.append(TestCmp)
 
 
+# ---------------------------------------------------------------------------
+
+class TestMCJIT(TestCase):
+    def test_mcjit(self):
+        m = Module.new('oidfjs')
+        fnty = Type.function(Type.int(), [Type.int(), Type.int()])
+        func = m.add_function(fnty, 'foo')
+        bb = func.append_basic_block('')
+        bldr = Builder.new(bb)
+        bldr.ret(bldr.add(*func.args))
+
+        func.verify()
+
+        engine = EngineBuilder.new(m).mcjit(True).create()
+        ptr = engine.get_pointer_to_function(func)
+
+        from ctypes import c_int, CFUNCTYPE
+        callee = CFUNCTYPE(c_int, c_int, c_int)(ptr)
+        self.assertEqual(321 + 123, callee(321, 123))
+
+    def test_multi_module_linking(self):
+        # generate external library module
+        m = Module.new('external-library-module')
+        fnty = Type.function(Type.int(), [Type.int(), Type.int()])
+        libfname = 'myadd'
+        func = m.add_function(fnty, libfname)
+        bb = func.append_basic_block('')
+        bldr = Builder.new(bb)
+        bldr.ret(bldr.add(*func.args))
+        func.verify()
+
+        # JIT the lib module and bind dynamic symbol
+        libengine = EngineBuilder.new(m).mcjit(True).create()
+        myadd_ptr = libengine.get_pointer_to_function(func)
+        le.dylib_add_symbol(libfname, myadd_ptr)
+
+        # reference external library
+        m = Module.new('user')
+        fnty = Type.function(Type.int(), [Type.int(), Type.int()])
+        func = m.add_function(fnty, 'foo')
+        bb = func.append_basic_block('')
+        bldr = Builder.new(bb)
+        extadd = m.get_or_insert_function(fnty, name=libfname)
+        bldr.ret(bldr.call(extadd, func.args))
+        func.verify()
+
+        # JIT the user module
+        engine = EngineBuilder.new(m).mcjit(True).create()
+        ptr = engine.get_pointer_to_function(func)
+        self.assertEqual(myadd_ptr,
+                         engine.get_pointer_to_named_function(libfname))
+
+        from ctypes import c_int, CFUNCTYPE
+        callee = CFUNCTYPE(c_int, c_int, c_int)(ptr)
+        self.assertEqual(321 + 123, callee(321, 123))
+
+
+if (llvm.version >= (3, 3) and
+    not (sys.platform.startswith('win32') and BITS == 64)):
+    # MCJIT broken in 3.2, the test will segfault in OSX?
+    # Compatbility problem on windows 7 64-bit?
+    tests.append(TestMCJIT)
+
+
+class TestLLRT(TestCase):
+    def test_llrt_divmod(self):
+        from llvm import llrt
+        m = lc.Module.new('testllrt')
+        longlong = lc.Type.int(64)
+        lfunc = m.add_function(lc.Type.function(longlong, [longlong, longlong]), 'foo')
+        bldr = lc.Builder.new(lfunc.append_basic_block(''))
+        bldr.ret(bldr.udiv(*lfunc.args))
+
+        llrt.replace_divmod64(lfunc)
+
+        rt = llrt.LLRT()
+        rt.install_symbols()
+
+        engine = le.EngineBuilder.new(m).create()
+        pointer = engine.get_pointer_to_function(lfunc)
+
+        from ctypes import CFUNCTYPE, c_uint64, c_int64
+        func = CFUNCTYPE(c_uint64, c_uint64, c_uint64)(pointer)
+        a, b = 98342, 2231
+        self.assertEqual(func(98342, 2231), 98342 // 2231)
+
+        rt.uninstall_symbols()
+
+tests.append(TestLLRT)
+
+class TestArith(TestCase):
+    '''
+    Test basic arithmetic support with LLVM MCJIT
+    '''
+    def func_template(self, ty, op):
+        m = Module.new('dofjaa')
+        fnty = Type.function(ty, [ty, ty])
+        fn = m.add_function(fnty, 'foo')
+        bldr = Builder.new(fn.append_basic_block(''))
+        bldr.ret(getattr(bldr, op)(*fn.args))
+
+        engine = EngineBuilder.new(m).mcjit(True).create()
+        ptr = engine.get_pointer_to_function(fn)
+
+        from ctypes import c_uint32, c_uint64, c_float, c_double, CFUNCTYPE
+
+        maptypes = {
+            Type.int(32): c_uint32,
+            Type.int(64): c_uint64,
+            Type.float(): c_float,
+            Type.double(): c_double,
+        }
+        cty = maptypes[ty]
+        prototype = CFUNCTYPE(*[cty] * 3)
+        callee = prototype(ptr)
+        callee(12, 23) 
+
+    def template(self, iop, fop):
+        inttys = [Type.int(32), Type.int(64)]
+        flttys = [Type.float(), Type.double()]
+
+        if iop:
+            for ty in inttys:
+                self.func_template(ty, iop)
+        if fop:
+            for ty in flttys:
+                self.func_template(ty, fop)
+
+    def test_add(self):
+        self.template('add', 'fadd')
+
+    def test_sub(self):
+        self.template('sub', 'fsub')
+        
+    def test_mul(self):
+        self.template('mul', 'fmul')
+
+    def test_div(self):
+        if BITS == 32:
+            print('skipped test for div')
+            print('known failure due to unresolved external symbol __udivdi3')
+            return
+        self.template('udiv', None) # 'fdiv')
+
+    def test_rem(self):
+        if BITS == 32:
+            print('skipped test for rem')
+            print('known failure due to unresolved external symbol __umoddi3')
+            return
+        self.template('urem', None) # 'frem')
+
+if llvm.version >= (3, 3):
+    # MCJIT is broken in 3.2
+    tests.append(TestArith)
+
+class TestNUWNSW(TestCase):
+    def make_module(self):
+        mod = Module.new('asdfa')
+        fnty = Type.function(Type.void(), [Type.int()] * 2)
+        func = mod.add_function(fnty, 'foo')
+        bldr = Builder.new(func.append_basic_block(''))
+        return mod, func, bldr
+
+    def has_nsw(self, inst, op):
+        self.assertTrue(('%s nsw' % op) in str(inst), "NSW flag does not work")
+
+    def has_nuw(self, inst, op):
+        self.assertTrue(('%s nuw' % op) in str(inst), "NUW flag does not work")
+
+    def _test_template(self, opf, opname):
+        mod, func, bldr = self.make_module()
+        a, b = func.args
+        self.has_nsw(opf(bldr, a, b, nsw=True), opname)
+        self.has_nuw(opf(bldr, a, b, nuw=True), opname)
+
+    def test_add_nuw_nsw(self):
+        self._test_template(Builder.add, 'add')
+
+    def test_sub_nuw_nsw(self):
+        self._test_template(Builder.sub, 'sub')
+
+    def test_mul_nuw_nsw(self):
+        self._test_template(Builder.mul, 'mul')
+
+    def test_shl_nuw_nsw(self):
+        self._test_template(Builder.shl, 'shl')
+
+    def test_neg_nuw_nsw(self):
+        mod, func, bldr = self.make_module()
+        a, b = func.args
+        self.has_nsw(bldr.neg(a, nsw=True), 'sub')
+        self.has_nuw(bldr.neg(a, nuw=True), 'sub')
+
+
+tests.append(TestNUWNSW)
+
+
+class TestExact(TestCase):
+    def make_module(self):
+        mod = Module.new('asdfa')
+        fnty = Type.function(Type.void(), [Type.int()] * 2)
+        func = mod.add_function(fnty, 'foo')
+        bldr = Builder.new(func.append_basic_block(''))
+        return mod, func, bldr
+
+    def has_exact(self, inst, op):
+        self.assertTrue(('%s exact' % op) in str(inst), "exact flag does not work")
+
+    def _test_template(self, opf, opname):
+        mod, func, bldr = self.make_module()
+        a, b = func.args
+        self.has_exact(opf(bldr, a, b, exact=True), opname)
+
+    def test_udiv_exact(self):
+        self._test_template(Builder.udiv, 'udiv')
+
+    def test_sdiv_exact(self):
+        self._test_template(Builder.sdiv, 'sdiv')
+
+    def test_lshr_exact(self):
+        self._test_template(Builder.lshr, 'lshr')
+
+    def test_ashr_exact(self):
+        self._test_template(Builder.ashr, 'ashr')
+
+tests.append(TestExact)
+
+
+
 # ---------------------------------------------------------------------------
 
 def run(verbosity=1):
diff --git a/llvm/workaround/avx_support.py b/llvm/workaround/avx_support.py
index 81cb936..0b1c2da 100644
--- a/llvm/workaround/avx_support.py
+++ b/llvm/workaround/avx_support.py
@@ -50,8 +50,7 @@ def detect_osx_like():
                                 stdout=subprocess.PIPE)
     except OSError:
         return False
-
-    features = info.stdout.read()
+    features = info.stdout.read().decode('UTF8')
     features = features.split()
     return 'AVX1.0' in features and 'OSXSAVE' in features and 'XSAVE' in features
 
diff --git a/llvmpy/README b/llvmpy/README
deleted file mode 100644
index 2bfabcd..0000000
--- a/llvmpy/README
+++ /dev/null
@@ -1,5 +0,0 @@
-# README
-
-This is a reimplementation of the LLVM binding, aiming to provide a more familiar interface to the C++ API whenever possible.
-
-The implementation uses a custom DSL in python to describe the interface (under binding directory). The DSL serves as input to *gen.py* for generation of the .cpp and .py files for the actual binding.
diff --git a/llvmpy/README.md b/llvmpy/README.md
new file mode 100644
index 0000000..4e444f8
--- /dev/null
+++ b/llvmpy/README.md
@@ -0,0 +1,151 @@
+# README
+
+This is a reimplementation of the LLVM binding, aiming to provide a more 
+familiar interface to the C++ API whenever possible.
+
+The implementation uses a custom DSL in python to describe the interface. 
+The DSL serves as input to *gen/gen.py* for generation of the .cpp and .py files 
+for the actual binding.
+
+# How to Add New Class
+
+Let's use the `llvm::Module` as an example because it should be one of the most
+familiar class in LLVM.
+
+Reference code in https://github.com/llvmpy/llvmpy/blob/master/llvmpy/src/Module.py
+and see LLVM documentation at http://llvm.org/docs/doxygen/html/classllvm_1_1Module.html
+
+1) Import binding helpers
+
+```python
+from binding import *
+# Yes, it is bad practice to import star.
+# I will fix it one day.
+```
+
+2) Import LLVM namespace
+
+```python
+from .namespace import llvm
+```
+
+3) Declare the class
+
+```python
+Module = llvm.Class()
+```
+
+4) Import all the dependencies for the definition
+
+5) Define the class
+
+```python
+@Module
+class Module:
+    ...
+```
+
+## Inside the definition...
+
+5.1) Use the ``_include_`` attribute to add include files.
+
+5.2) Use ``Enum`` to create an enumerator type.
+
+5.3) Make constructor 
+
+Not every class needs to have a binding for the constructor.
+Only add things that will be used.
+
+```python
+new = Constructor(cast(str, StringRef), ref(LLVMContext))
+```
+
+The constructor must be named as "new".
+
+The args to ``Constructor`` are parameters of the signature.
+The first parameter means cast Python string to a StringRef.
+The second parameter means pass LLVMContext object as a value reference.
+
+5.4) Make destructor
+
+Not every class needs to have a binding for the destructor.
+Only add things that will be used.
+If it is always owned by another object, it usually does not need to have one.
+
+```python
+delete = Destructor()
+```
+
+The destructor must be named as "delete".
+
+5.5) Add Simple Methods
+
+```python
+getFunction = Method(ptr(Function), cast(str, StringRef))
+```
+
+The first arg is the return type: a ponter to Function.
+The rest of the args are for the parameters.
+
+Note: ``cast(fromtype, totype)`` can be used as return-type as well.
+In that case, the ``fromtype`` will usually refer to a LLVM object
+and the ``totype`` will refer to the Python object.
+
+5.6) Add custom methods defined in C++
+
+The ``list_functions`` is created as a ``CustomMethod``.
+
+```python
+list_functions = CustomMethod('Module_list_functions', PyObjectPtr)
+```
+
+The first arg is the name that appears in C++.
+The second argument is the return-type.
+The rest of the arguments are parameters.
+
+The definition of ``Module_list_functions`` is located in 
+"include/llvm_binding/extra.h".
+
+5.7) Add custom python method
+
+One can also add custom methods in Python, e.g. ``__str__``.
+
+```python
+@CustomPythonMethod
+def __str__(self):
+    from llvmpy import extra
+    os = extra.make_raw_ostream_for_printing()
+    self.print_(os, None)
+    return os.str()
+```
+
+The body of ``__str__`` is directly copied to the Python output file.
+Thus, it can't reference to anything in current file scope.
+
+# Static Methods
+
+``StaticMethod`` https://github.com/llvmpy/llvmpy/blob/master/llvmpy/src/PassRegistry.py
+
+``CustomStaticMethod`` https://github.com/llvmpy/llvmpy/blob/master/llvmpy/src/Support/TargetRegistry.py
+
+
+# Namespace
+
+``Namespace`` https://github.com/llvmpy/llvmpy/blob/master/llvmpy/src/Support/CodeGen.py
+
+# Functions
+
+``Function`` https://github.com/llvmpy/llvmpy/blob/master/llvmpy/src/Assembly/Parser.py
+
+``CustomFunction`` https://github.com/llvmpy/llvmpy/blob/master/llvmpy/src/Bitcode/ReaderWriter.py
+
+
+
+
+
+# Other Things from the binding.py
+
+https://github.com/llvmpy/llvmpy/blob/master/llvmpy/gen/binding.py
+
+The list of C++ types: https://github.com/llvmpy/llvmpy/blob/master/llvmpy/gen/binding.py#L218
+
diff --git a/llvmpy/capsule.py b/llvmpy/capsule.py
index 5efed13..9de0571 100644
--- a/llvmpy/capsule.py
+++ b/llvmpy/capsule.py
@@ -3,11 +3,13 @@ from collections import defaultdict
 import logging
 logger = logging.getLogger(__name__)
 
+NO_DEBUG = False
 def silent_logger():
     '''
     Silent logger for unless we have a error message.
     '''
     logger.setLevel(logging.ERROR)
+    NO_DEBUG = True
 
 # comment out the line below to re-enable logging at DEBUG level.
 silent_logger()
@@ -35,7 +37,11 @@ def _capsule_weakref_dtor(item):
     if refct == 0:
         dtor = _addr2dtor.pop((name, addr), None)
         if dtor is not None:
-            logger.debug('Destroy %s %s', name, hex(addr))
+            if not NO_DEBUG:
+                # Some globals in logger could be removed by python GC
+                # at interpreter teardown.
+                # That can cause exception raised and ignored message.
+                logger.debug('Destroy %s %s', name, hex(addr))
             dtor(item.capsule)
 
 class Capsule(object):
diff --git a/llvmpy/gen/binding.py b/llvmpy/gen/binding.py
index 8006fd9..30e1614 100644
--- a/llvmpy/gen/binding.py
+++ b/llvmpy/gen/binding.py
@@ -247,7 +247,8 @@ class Class(SubModule, _Type):
     def __call__(self, defn):
         assert not self._is_defined
         # process the definition in "defn"
-        self.name = defn.__name__
+        self.name = getattr(defn, '_name_', defn.__name__)
+
         for k, v in defn.__dict__.items():
             if isinstance(v, Method):
                 self.methods.append(v)
@@ -332,7 +333,7 @@ class Class(SubModule, _Type):
         writer.die_if_false(raw, verbose=name)
         ptrty = ptr(self).fullname
         ty = self.fullname
-        fmt = 'typecast< %(ty)s >::from(%(raw)s)'
+        fmt = 'unwrap_as<%(ty)s, %(name)s >::from(%(raw)s)'
         casted = writer.declare(ptrty, fmt % locals())
         writer.die_if_false(casted)
         return casted
diff --git a/llvmpy/gen/codegen.py b/llvmpy/gen/codegen.py
index f9585fd..6e36683 100644
--- a/llvmpy/gen/codegen.py
+++ b/llvmpy/gen/codegen.py
@@ -202,7 +202,8 @@ class CppCodeWriter(CodeWriterBase):
 
     def pycapsule_new(self, ptr, name, clsname):
         name_soften = mangle(name)
-        ret = self.call('pycapsule_new', 'PyObject*', ptr, quote(name),
+        cast_to_base = 'cast_to_base<%s >::from(%s)' % (name, ptr)
+        ret = self.call('pycapsule_new', 'PyObject*', cast_to_base, quote(name),
                         quote(clsname))
         with self.block('if (!%(ret)s)' % locals()):
             self.return_null()
diff --git a/llvmpy/include/llvm_binding/conversion.h b/llvmpy/include/llvm_binding/conversion.h
index a9f073c..b37f618 100644
--- a/llvmpy/include/llvm_binding/conversion.h
+++ b/llvmpy/include/llvm_binding/conversion.h
@@ -256,6 +256,7 @@ template<class Td>
 struct typecast {
     template<class Ts> static
     Td* from(Ts* src) {
+        // check why this is only used in Python3
         return llvm::dyn_cast<Td>(src);
     }
 
@@ -265,3 +266,24 @@ struct typecast {
     }
 };
 
+template<class Td, class Tbase>
+struct unwrap_as {
+    static
+    Td* from(void* src) {
+        Tbase* base = static_cast<Tbase*>(src);
+        return static_cast<Td*>(base);
+    }
+};
+
+template<class Td>
+struct cast_to_base {
+    template<class Ts> static
+    Td* from(Ts* src){
+        return static_cast<Td*>(src);
+    }
+    
+    template<class Ts> static
+    const Td* from(const Ts* src){
+        return static_cast<const Td*>(src);
+    }
+};
diff --git a/llvmpy/include/llvm_binding/extra.h b/llvmpy/include/llvm_binding/extra.h
index 456f75e..161c52e 100644
--- a/llvmpy/include/llvm_binding/extra.h
+++ b/llvmpy/include/llvm_binding/extra.h
@@ -1,8 +1,22 @@
 #include <Python.h>
 #include <llvm/ADT/SmallVector.h>
-#include <llvm/Value.h>
-#include <llvm/DerivedTypes.h>
-#include <llvm/Function.h>
+#if LLVM_VERSION_MAJOR >= 3 and LLVM_VERSION_MINOR >= 3
+    #include <llvm/IR/Value.h>
+    #include <llvm/IR/DerivedTypes.h>
+    #include <llvm/IR/Function.h>
+    #include <llvm/IR/Module.h>
+    #include <llvm/IR/Constants.h>
+    #include <llvm/IR/Intrinsics.h>
+    #include <llvm/IR/IRBuilder.h>
+#else
+    #include <llvm/Value.h>
+    #include <llvm/DerivedTypes.h>
+    #include <llvm/Function.h>
+    #include <llvm/Module.h>
+    #include <llvm/Constants.h>
+    #include <llvm/Intrinsics.h>
+    #include <llvm/IRBuilder.h>
+#endif
 #include <llvm/Support/raw_ostream.h>
 #include <llvm/Support/FormattedStream.h>
 #include <llvm/Support/MemoryBuffer.h>
@@ -12,14 +26,11 @@
 #include <llvm/ExecutionEngine/ExecutionEngine.h>
 #include <llvm/ExecutionEngine/GenericValue.h>
 #include <llvm/Linker.h>
-#include <llvm/Module.h>
 #include <llvm/Analysis/Verifier.h>
-#include <llvm/Constants.h>
-#include <llvm/Intrinsics.h>
-#include <llvm/IRBuilder.h>
 #include <llvm/PassRegistry.h>
 #include <llvm/Support/Host.h>
 
+#include <llvm/ExecutionEngine/MCJIT.h> // to make MCJIT working
 
 #include "auto_pyobject.h"
 
@@ -56,6 +67,15 @@ namespace extra{
 
 }
 
+static
+PyObject* callwrite(PyObject* self, PyObject* arg)
+{
+    char meth[] = "write";
+    char fmt[] = "O";
+    return PyObject_CallMethod(self, meth, fmt, arg);
+}
+
+
 static
 PyObject* make_raw_ostream_for_printing(PyObject* self, PyObject* args)
 {
@@ -271,7 +291,7 @@ llvm::ExecutionEngine* ExecutionEngine_create(
                                                   &ErrorStr, OptLevel,
                                                   GVsWithCode);
     auto_pyobject buf = PyBytes_FromString(ErrorStr.c_str());
-    if (errout && NULL == PyObject_CallMethod(errout, "write", "O", *buf)){
+    if (errout && NULL == callwrite(errout, *buf)){
         return NULL;
     }
 
@@ -298,7 +318,7 @@ llvm::ExecutionEngine* ExecutionEngine_createJIT(
     ExecutionEngine *ee = ExecutionEngine::createJIT(M, &ErrorStr, JMM, OL,
                                                      GCsWithCode, RM, CMM);
     auto_pyobject buf = PyBytes_FromString(ErrorStr.c_str());
-    if (errout && NULL == PyObject_CallMethod(errout, "write", "O", *buf)){
+    if (errout && NULL == callwrite(errout, *buf)){
         return NULL;
     }
 //    PyFile_WriteString(ErrorStr.c_str(), errout);
@@ -483,7 +503,7 @@ PyObject* llvm_ParseBitCodeFile(llvm::StringRef Buf, llvm::LLVMContext& Ctx,
         std::string ErrStr;
         M = ParseBitcodeFile(MB, Ctx, &ErrStr);
         auto_pyobject buf = PyBytes_FromString(ErrStr.c_str());
-        if (NULL == PyObject_CallMethod(FObj, "write", "O", *buf)){
+        if (NULL == callwrite(FObj, *buf)){
             return NULL;
         }
 //        if (-1 == PyFile_WriteString(ErrStr.c_str(), FObj)) {
@@ -507,7 +527,7 @@ PyObject* llvm_WriteBitcodeToFile(const llvm::Module *M, PyObject* FObj)
     rso.flush();
     StringRef ref = rso.str();
     auto_pyobject buf = PyBytes_FromStringAndSize(ref.data(), ref.size());
-    return PyObject_CallMethod(FObj, "write", "O", *buf);
+    return callwrite(FObj, *buf);
 }
 
 static
@@ -522,7 +542,7 @@ PyObject* llvm_getBitcodeTargetTriple(llvm::StringRef Buf,
         std::string ErrStr;
         Triple = getBitcodeTargetTriple(MB, Ctx, &ErrStr);
         auto_pyobject buf = PyBytes_FromString(ErrStr.c_str());
-        if (NULL == PyObject_CallMethod(FObj, "write", "O", *buf)){
+        if (NULL == callwrite(FObj, *buf)){
             return NULL;
         }
 //        if (-1 == PyFile_WriteString(ErrStr.c_str(), FObj)) {
@@ -594,19 +614,19 @@ PyObject* Linker_LinkInModule(llvm::Linker* Linker,
                                PyObject* ErrMsg)
 {
     std::string errmsg;
+#if LLVM_VERSION_MAJOR >= 3 && LLVM_VERSION_MINOR >= 3
+    bool failed = Linker->linkInModule(Mod, &errmsg);
+#else
     bool failed = Linker->LinkInModule(Mod, &errmsg);
+#endif
     if (! failed) {
         Py_RETURN_FALSE;
     } else {
     
         auto_pyobject buf = PyBytes_FromString(errmsg.c_str());
-        if (NULL == PyObject_CallMethod(ErrMsg, "write", "O", *buf)){
+        if (NULL == callwrite(ErrMsg, *buf)){
             return NULL;
         }
-
-//        if (-1 == PyFile_WriteString(errmsg.c_str(), ErrMsg)) {
-//            return NULL;
-//        }
         Py_RETURN_TRUE;
     }
 }
@@ -623,7 +643,7 @@ PyObject* Linker_LinkModules(llvm::Module* Dest,
         Py_RETURN_FALSE;
     } else {
         auto_pyobject buf = PyBytes_FromString(errmsg.c_str());
-        if (NULL == PyObject_CallMethod(ErrMsg, "write", "O", *buf)){
+        if (NULL == callwrite(ErrMsg, *buf)){
             return NULL;
         }
 //        if (-1 == PyFile_WriteString(errmsg.c_str(), ErrMsg)) {
@@ -689,7 +709,7 @@ PyObject* llvm_verifyModule(const llvm::Module& Fn,
 
     if (failed) {
         auto_pyobject buf = PyBytes_FromString(errmsg.c_str());
-        if (NULL == PyObject_CallMethod(ErrMsg, "write", "O", *buf)){
+        if (NULL == callwrite(ErrMsg, *buf)){
             return NULL;
         }
 
@@ -819,7 +839,7 @@ PyObject* DynamicLibrary_LoadLibraryPermanently(const char * Filename,
         failed = DynamicLibrary::LoadLibraryPermanently(Filename, &errmsg);
         if (failed) {
             auto_pyobject buf = PyBytes_FromString(errmsg.c_str());
-            if (NULL == PyObject_CallMethod(ErrMsg, "write", "O", *buf)){
+            if (NULL == callwrite(ErrMsg, *buf)){
                 return NULL;
             }
 //            if (-1 == PyFile_WriteString(errmsg.c_str(), ErrMsg)) {
@@ -869,7 +889,7 @@ PyObject* TargetRegistry_lookupTarget(const std::string &Triple,
     if (!target) {
 //        PyFile_WriteString(error.c_str(), Error);
         auto_pyobject buf = PyBytes_FromString(error.c_str());
-        if (NULL == PyObject_CallMethod(Error, "write", "O", *buf)){
+        if (NULL == callwrite(Error, *buf)){
             return NULL;
         }
 
@@ -891,7 +911,7 @@ PyObject* TargetRegistry_lookupTarget(const std::string &Arch,
     if (!target) {
 //        PyFile_WriteString(error.c_str(), Error);
         auto_pyobject buf = PyBytes_FromString(error.c_str());
-        if (NULL == PyObject_CallMethod(Error, "write", "O", *buf)){
+        if (NULL == callwrite(Error, *buf)){
             return NULL;
         }
 
@@ -910,7 +930,7 @@ PyObject* TargetRegistry_getClosestTargetForJIT(PyObject* Error)
     const Target* target = TargetRegistry::getClosestTargetForJIT(error);
     if (!target) {
         auto_pyobject buf = PyBytes_FromString(error.c_str());
-        if (NULL == PyObject_CallMethod(Error, "write", "O", *buf)){
+        if (NULL == callwrite(Error, *buf)){
             return NULL;
         }
 
@@ -945,3 +965,22 @@ PyObject* llvm_sys_getHostCPUFeatures(PyObject* Features)
     }
 }
 
+#if LLVM_VERSION_MAJOR >= 3 && LLVM_VERSION_MINOR >= 3
+static
+PyObject* llvm_sys_isLittleEndianHost()
+{
+    if (llvm::sys::IsLittleEndianHost)
+        Py_RETURN_TRUE;
+    else
+        Py_RETURN_FALSE;
+}
+
+static
+PyObject* llvm_sys_isBigEndianHost()
+{
+    if (llvm::sys::IsBigEndianHost)
+        Py_RETURN_TRUE;
+    else
+        Py_RETURN_FALSE;
+}
+#endif 
diff --git a/llvmpy/src/Argument.py b/llvmpy/src/Argument.py
index 6262147..a08cca1 100644
--- a/llvmpy/src/Argument.py
+++ b/llvmpy/src/Argument.py
@@ -1,16 +1,27 @@
 from binding import *
 from .namespace import llvm
 from .Value import Argument, Value
-from .Attributes import Attributes
+if LLVM_VERSION >= (3, 3):
+    from .Attributes import AttributeSet, Attribute
+else:
+    from .Attributes import Attributes
 
 @Argument
 class Argument:
-    _include_ = 'llvm/Argument.h'
+    if LLVM_VERSION >= (3, 3):
+        _include_ = 'llvm/IR/Argument.h'
+    else:
+        _include_ = 'llvm/Argument.h'
 
     _downcast_ = Value
 
-    addAttr = Method(Void, ref(Attributes))
-    removeAttr = Method(Void, ref(Attributes))
+    if LLVM_VERSION >= (3, 3):
+        addAttr = Method(Void, ref(AttributeSet))
+        removeAttr = Method(Void, ref(AttributeSet))
+    else:
+        addAttr = Method(Void, ref(Attributes))
+        removeAttr = Method(Void, ref(Attributes))
+
     getParamAlignment = Method(cast(Unsigned, int))
 
     getArgNo = Method(cast(Unsigned, int))
diff --git a/llvmpy/src/Attributes.py b/llvmpy/src/Attributes.py
index 4fe7651..303d61d 100644
--- a/llvmpy/src/Attributes.py
+++ b/llvmpy/src/Attributes.py
@@ -2,38 +2,91 @@ from binding import *
 from .namespace import llvm
 from .LLVMContext import LLVMContext
 
-llvm.includes.add('llvm/Attributes.h')
+if LLVM_VERSION >= (3, 3):
+    llvm.includes.add('llvm/IR/Attributes.h')
+else:
+    llvm.includes.add('llvm/Attributes.h')
+
 
-Attributes = llvm.Class()
 AttrBuilder = llvm.Class()
 
-
-@Attributes
-class Attributes:
-    AttrVal = Enum('''None, AddressSafety, Alignment, AlwaysInline,
-        ByVal, InlineHint, InReg, MinSize,
-        Naked, Nest, NoAlias, NoCapture,
-        NoImplicitFloat, NoInline, NonLazyBind, NoRedZone,
-        NoReturn, NoUnwind, OptimizeForSize, ReadNone,
-        ReadOnly, ReturnsTwice, SExt, StackAlignment,
-        StackProtect, StackProtectReq, StructRet, UWTable, ZExt''')
-
-    delete = Destructor()
-
-    get = StaticMethod(Attributes, ref(LLVMContext), ref(AttrBuilder))
+if LLVM_VERSION >= (3, 3):
+    Attribute = llvm.Class()
+    AttributeSet = llvm.Class()
+else:
+    Attributes = llvm.Class()
 
 
-@AttrBuilder
-class AttrBuilder:
+if LLVM_VERSION >= (3, 3):
+    @Attribute
+    class Attribute:
+        AttrKind = Enum('''None, Alignment, AlwaysInline,
+                  ByVal, InlineHint, InReg, 
+                  MinSize, Naked, Nest, NoAlias, 
+                  NoBuiltin, NoCapture, NoDuplicate, NoImplicitFloat, 
+                  NoInline, NonLazyBind, NoRedZone, NoReturn, 
+                  NoUnwind, OptimizeForSize, ReadNone, ReadOnly, 
+                  Returned, ReturnsTwice, SExt, StackAlignment, 
+                  StackProtect, StackProtectReq, StackProtectStrong, StructRet, 
+                  SanitizeAddress, SanitizeThread, SanitizeMemory, UWTable, 
+                  ZExt, EndAttrKinds''')
 
-    new = Constructor()
-    delete = Destructor()
+        delete = Destructor()
 
-    clear = Method()
+        get = StaticMethod(Attribute,
+                           ref(LLVMContext),
+                           AttrKind,
+                           cast(int, Uint64)).require_only(2)
 
-    addAttribute = Method(ref(AttrBuilder), Attributes.AttrVal)
-    removeAttribute = Method(ref(AttrBuilder), Attributes.AttrVal)
+    @AttrBuilder
+    class AttrBuilder:
 
-    addAlignmentAttr = Method(ref(AttrBuilder), cast(int, Unsigned))
+        new = Constructor()
+        delete = Destructor()
+
+        clear = Method()
+
+        addAttribute = Method(ref(AttrBuilder), Attribute.AttrKind)
+        removeAttribute = Method(ref(AttrBuilder), Attribute.AttrKind)
+
+        addAlignmentAttr = Method(ref(AttrBuilder), cast(int, Unsigned))
+
+    @AttributeSet
+    class AttributeSet:
+        delete = Destructor()
+
+        get = StaticMethod(AttributeSet,
+                           ref(LLVMContext),
+                           cast(int, Unsigned),
+                           ref(AttrBuilder))
+
+else:
+    @Attributes
+    class Attributes:
+        AttrVal = Enum('''None, AddressSafety, Alignment, AlwaysInline,
+            ByVal, InlineHint, InReg, MinSize,
+            Naked, Nest, NoAlias, NoCapture,
+            NoImplicitFloat, NoInline, NonLazyBind, NoRedZone,
+            NoReturn, NoUnwind, OptimizeForSize, ReadNone,
+            ReadOnly, ReturnsTwice, SExt, StackAlignment,
+            StackProtect, StackProtectReq, StructRet, UWTable, ZExt''')
+
+        delete = Destructor()
+
+        get = StaticMethod(Attributes, ref(LLVMContext), ref(AttrBuilder))
+
+
+    @AttrBuilder
+    class AttrBuilder:
+
+        new = Constructor()
+        delete = Destructor()
+
+        clear = Method()
+
+        addAttribute = Method(ref(AttrBuilder), Attributes.AttrVal)
+        removeAttribute = Method(ref(AttrBuilder), Attributes.AttrVal)
+
+        addAlignmentAttr = Method(ref(AttrBuilder), cast(int, Unsigned))
 
 
diff --git a/llvmpy/src/DIBuilder.py b/llvmpy/src/DIBuilder.py
index 9671963..cb934d4 100644
--- a/llvmpy/src/DIBuilder.py
+++ b/llvmpy/src/DIBuilder.py
@@ -124,17 +124,31 @@ class DIBuilder:
                              ptr(MDNode),           # TemplateParms = 0
                              ).require_only(10)
 
-    createStructType = Method(DIType,
-                              ref(DIDescriptor),   # Scope
-                              stringref_arg,        # Name
-                              ref(DIFile),          # File
-                              unsigned_arg,         # LineNumber
-                              uint64_arg,           # SizeInBits
-                              uint64_arg,           # AlignInBits
-                              unsigned_arg,         # Flags
-                              ref(DIArray),         # Elements
-                              unsigned_arg,         # RunTimeLang = 0
-                              ).require_only(8)
+    if LLVM_VERSION >= (3, 3):
+        createStructType = Method(DIType,
+                                  ref(DIDescriptor),   # Scope
+                                  stringref_arg,        # Name
+                                  ref(DIFile),          # File
+                                  unsigned_arg,         # LineNumber
+                                  uint64_arg,           # SizeInBits
+                                  uint64_arg,           # AlignInBits
+                                  unsigned_arg,         # Flags
+                                  ref(DIType),          # DerivedFrom
+                                  ref(DIArray),         # Elements
+                                  unsigned_arg,         # RunTimeLang = 0
+                                  ).require_only(9)
+    else:
+        createStructType = Method(DIType,
+                                  ref(DIDescriptor),   # Scope
+                                  stringref_arg,        # Name
+                                  ref(DIFile),          # File
+                                  unsigned_arg,         # LineNumber
+                                  uint64_arg,           # SizeInBits
+                                  uint64_arg,           # AlignInBits
+                                  unsigned_arg,         # Flags
+                                  ref(DIArray),         # Elements
+                                  unsigned_arg,         # RunTimeLang = 0
+                                  ).require_only(8)
 
     createUnionType = Method(DIType,
                              ref(DIDescriptor),     # Scope
@@ -186,7 +200,7 @@ class DIBuilder:
                                 ref(DIType),      # Ty
                                 )
     
-    createTemporaryType = Method(DIType, ref(DIFile)).require_only(0)
+    #createTemporaryType = Method(DIType, ref(DIFile)).require_only(0)
 
     createForwardDecl = Method(DIType,
                                unsigned_arg,         # Tag
diff --git a/llvmpy/src/DataLayout.py b/llvmpy/src/DataLayout.py
index 812a83d..6741cda 100644
--- a/llvmpy/src/DataLayout.py
+++ b/llvmpy/src/DataLayout.py
@@ -15,7 +15,10 @@ from .GlobalVariable import GlobalVariable
 
 @DataLayout
 class DataLayout:
-    _include_ = 'llvm/DataLayout.h'
+    if LLVM_VERSION >= (3, 3):
+        _include_ = 'llvm/IR/DataLayout.h'
+    else:
+        _include_ = 'llvm/DataLayout.h'
 
     _new_string = Constructor(cast(str, StringRef))
     _new_module = Constructor(ptr(Module))
diff --git a/llvmpy/src/DerivedTypes.py b/llvmpy/src/DerivedTypes.py
index 02eec8c..fde01f3 100644
--- a/llvmpy/src/DerivedTypes.py
+++ b/llvmpy/src/DerivedTypes.py
@@ -8,7 +8,10 @@ FunctionType = llvm.Class(Type)
 
 @FunctionType
 class FunctionType:
-    _include_ = 'llvm/DerivedTypes.h'
+    if LLVM_VERSION >= (3, 3):
+        _include_ = 'llvm/IR/DerivedTypes.h'
+    else:
+        _include_ = 'llvm/DerivedTypes.h'
     _downcast_ = Type
 
     _get = StaticMethod(ptr(FunctionType), ptr(Type), cast(bool, Bool))
diff --git a/llvmpy/src/ExecutionEngine/ExecutionEngine.py b/llvmpy/src/ExecutionEngine/ExecutionEngine.py
index 3208269..d6dfe43 100644
--- a/llvmpy/src/ExecutionEngine/ExecutionEngine.py
+++ b/llvmpy/src/ExecutionEngine/ExecutionEngine.py
@@ -101,3 +101,5 @@ class ExecutionEngine:
         unwrapped = list(map(capsule.unwrap, args))
         return self._runFunction(fn, tuple(unwrapped))
 
+    finalizeObject = Method(Void)
+
diff --git a/llvmpy/src/Function.py b/llvmpy/src/Function.py
index 04c3e65..9c2c8b7 100644
--- a/llvmpy/src/Function.py
+++ b/llvmpy/src/Function.py
@@ -4,7 +4,10 @@ from .Value import GlobalValue, Constant, Function, Argument, Value
 from .Module import Module
 from .BasicBlock import BasicBlock
 from .ValueSymbolTable import ValueSymbolTable
-from .Attributes import Attributes
+if LLVM_VERSION >= (3, 3):
+    from .Attributes import Attribute, AttributeSet
+else:
+    from .Attributes import Attributes
 from .Type import Type
 from .DerivedTypes import FunctionType
 from .LLVMContext import LLVMContext
@@ -12,7 +15,11 @@ from .CallingConv import CallingConv
 
 @Function
 class Function:
-    _include_ = 'llvm/Function.h'
+    if LLVM_VERSION >= (3, 3):
+        _include_ = 'llvm/IR/Function.h'
+    else:
+        _include_ = 'llvm/Function.h'
+        
     _downcast_  = GlobalValue, Constant, Value
 
     getReturnType = Method(ptr(Type))
@@ -48,10 +55,17 @@ class Function:
 
     deleteBody = Method()
     viewCFG = Method()
+
     viewCFGOnly = Method()
 
-    addFnAttr = Method(Void, Attributes.AttrVal)
-    removeFnAttr = Method(Void, ref(Attributes))
+    if LLVM_VERSION >= (3, 3):
+        addFnAttr = Method(Void, Attribute.AttrKind)
+        addAttributes = Method(Void, cast(int, Unsigned), ref(AttributeSet))
+        removeAttributes = Method(Void, cast(int, Unsigned), ref(AttributeSet))
+        #removeFnAttr = Method(Void, Attribute.AttrKind) # 3.4?
+    else:
+        addFnAttr = Method(Void, Attributes.AttrVal)
+        removeFnAttr = Method(Void, ref(Attributes))
     #hasFnAttribute = Method(cast(Bool, bool), Attributes.AttrVal)
 
     Create = StaticMethod(ptr(Function),
diff --git a/llvmpy/src/GlobalValue.py b/llvmpy/src/GlobalValue.py
index d954ed0..49071e3 100644
--- a/llvmpy/src/GlobalValue.py
+++ b/llvmpy/src/GlobalValue.py
@@ -6,7 +6,10 @@ from .ADT.StringRef import StringRef
 
 @GlobalValue
 class GlobalValue:
-    _include_ = 'llvm/GlobalValue.h'
+    if LLVM_VERSION >= (3, 3):
+        _include_ = 'llvm/IR/GlobalValue.h'
+    else:
+        _include_ = 'llvm/GlobalValue.h'
 
     LinkageTypes = Enum('''
         ExternalLinkage, AvailableExternallyLinkage, LinkOnceAnyLinkage,
diff --git a/llvmpy/src/IRBuilder.py b/llvmpy/src/IRBuilder.py
index c0766e9..c4d8a86 100644
--- a/llvmpy/src/IRBuilder.py
+++ b/llvmpy/src/IRBuilder.py
@@ -18,7 +18,11 @@ IRBuilder = llvm.Class()
 
 @IRBuilder
 class IRBuilder:
-    _include_ = 'llvm/IRBuilder.h'
+    if LLVM_VERSION >= (3, 3):
+        _include_ = 'llvm/IR/IRBuilder.h'
+    else:
+        _include_ = 'llvm/IRBuilder.h'
+
     _realname_ = 'IRBuilder<>'
 
     new = Constructor(ref(LLVMContext))
diff --git a/llvmpy/src/InlineAsm.py b/llvmpy/src/InlineAsm.py
index 8793343..c96cef0 100644
--- a/llvmpy/src/InlineAsm.py
+++ b/llvmpy/src/InlineAsm.py
@@ -4,7 +4,10 @@ from .Value import Value
 from .DerivedTypes import FunctionType
 from .ADT.StringRef import StringRef
 
-llvm.includes.add('llvm/InlineAsm.h')
+if LLVM_VERSION >= (3, 3):
+    llvm.includes.add('llvm/IR/InlineAsm.h')
+else:
+    llvm.includes.add('llvm/InlineAsm.h')
 
 InlineAsm = llvm.Class(Value)
 
diff --git a/llvmpy/src/Instruction.py b/llvmpy/src/Instruction.py
index e17e8a1..2771015 100644
--- a/llvmpy/src/Instruction.py
+++ b/llvmpy/src/Instruction.py
@@ -65,7 +65,10 @@ SynchronizationScope = llvm.Enum('SynchronizationScope',
 
 from .ADT.StringRef import StringRef
 from .CallingConv import CallingConv
-from .Attributes import Attributes
+if LLVM_VERSION >= (3, 3):
+    from .Attributes import AttributeSet, Attribute
+else:
+    from .Attributes import Attributes
 from .Type import Type
 
 
@@ -142,8 +145,12 @@ class CallInst:
     getCallingConv = Method(CallingConv.ID)
     setCallingConv = Method(Void, CallingConv.ID)
     getParamAlignment = Method(cast(Unsigned, int), cast(int, Unsigned))
-    addAttribute = Method(Void, cast(int, Unsigned), ref(Attributes))
-    removeAttribute = Method(Void, cast(int, Unsigned), ref(Attributes))
+    if LLVM_VERSION >= (3, 3):
+        addAttribute = Method(Void, cast(int, Unsigned), Attribute.AttrKind)
+        removeAttribute = Method(Void, cast(int, Unsigned), ref(Attribute))
+    else:
+        addAttribute = Method(Void, cast(int, Unsigned), ref(Attributes))
+        removeAttribute = Method(Void, cast(int, Unsigned), ref(Attributes))
     getCalledFunction = Method(ptr(Function))
     getCalledValue = Method(ptr(Value))
     setCalledFunction = Method(Void, ptr(Function))
@@ -264,7 +271,10 @@ class UnaryInstruction:
 
 @IntrinsicInst
 class IntrinsicInst:
-    _include_ = 'llvm/IntrinsicInst.h'
+    if LLVM_VERSION >= (3, 3):
+        _include_ = 'llvm/IR/IntrinsicInst.h'
+    else:
+        _include_ = 'llvm/IntrinsicInst.h'
     _downcast_ = Value, User, Instruction
 
 #compare
@@ -292,8 +302,12 @@ class InvokeInst:
     getCallingConv = Method(CallingConv.ID)
     setCallingConv = Method(Void, CallingConv.ID)
     getParamAlignment = Method(cast(Unsigned, int), cast(int, Unsigned))
-    addAttribute = Method(Void, cast(int, Unsigned), ref(Attributes))
-    removeAttribute = Method(Void, cast(int, Unsigned), ref(Attributes))
+    if LLVM_VERSION >= (3, 3):
+        addAttribute = Method(Void, cast(int, Unsigned), Attribute.AttrKind)
+        removeAttribute = Method(Void, cast(int, Unsigned), ref(Attribute))
+    else:
+        addAttribute = Method(Void, cast(int, Unsigned), ref(Attributes))
+        removeAttribute = Method(Void, cast(int, Unsigned), ref(Attributes))
     getCalledFunction = Method(ptr(Function))
     getCalledValue = Method(ptr(Value))
     setCalledFunction = Method(Void, ptr(Function))
diff --git a/llvmpy/src/LLVMContext.py b/llvmpy/src/LLVMContext.py
index 4a6b49a..8ec1607 100644
--- a/llvmpy/src/LLVMContext.py
+++ b/llvmpy/src/LLVMContext.py
@@ -3,6 +3,9 @@ from .namespace import llvm
 
 @llvm.Class()
 class LLVMContext:
-    _include_ = "llvm/LLVMContext.h"
+    if LLVM_VERSION >= (3, 3):
+        _include_ = "llvm/IR/LLVMContext.h"
+    else:
+        _include_ = "llvm/LLVMContext.h"
 
 llvm.Function('getGlobalContext', ref(LLVMContext))
diff --git a/llvmpy/src/Linker.py b/llvmpy/src/Linker.py
index 960bd90..5f1b7af 100644
--- a/llvmpy/src/Linker.py
+++ b/llvmpy/src/Linker.py
@@ -10,30 +10,33 @@ Linker = llvm.Class()
 
 @Linker
 class Linker:
-    ControlFlags = Enum('Verbose, QuietWarnings, QuietErrors')
+    #ControlFlags = Enum('Verbose, QuietWarnings, QuietErrors')
     LinkerMode = Enum('DestroySource, PreserveSource')
 
-    _new_w_empty = Constructor(cast(str, StringRef),
-                               cast(str, StringRef),
-                               ref(LLVMContext),
-                               cast(int, Unsigned)).require_only(3)
+    if LLVM_VERSION >= (3, 3):
+        new = Constructor(ptr(Module))
+    else:
+        _new_w_empty = Constructor(cast(str, StringRef),
+                                   cast(str, StringRef),
+                                   ref(LLVMContext),
+                                   cast(int, Unsigned)).require_only(3)
 
-    _new_w_existing = Constructor(cast(str, StringRef),
-                                  ptr(Module),
-                                  cast(int, Unsigned)).require_only(2)
+        _new_w_existing = Constructor(cast(str, StringRef),
+                                      ptr(Module),
+                                      cast(int, Unsigned)).require_only(2)
 
-    @CustomPythonStaticMethod
-    def new(progname, module_or_name, *args):
-        if isinstance(module_or_name, Module):
-            return _new_w_existing(progname, module_or_name, *args)
-        else:
-            return _new_w_empty(progname, module_or_name, *args)
+        @CustomPythonStaticMethod
+        def new(progname, module_or_name, *args):
+            if isinstance(module_or_name, Module):
+                return _new_w_existing(progname, module_or_name, *args)
+            else:
+                return _new_w_empty(progname, module_or_name, *args)
 
     delete = Destructor()
 
     getModule = Method(ptr(Module))
-    releaseModule = Method(ptr(Module))
-    getLastError = Method(cast(ConstStdString, str))
+    #releaseModule = Method(ptr(Module))
+    #getLastError = Method(cast(ConstStdString, str))
 
     LinkInModule = CustomMethod('Linker_LinkInModule',
                                 PyObjectPtr, # boolean
diff --git a/llvmpy/src/Module.py b/llvmpy/src/Module.py
index 46fad5a..9dbe754 100644
--- a/llvmpy/src/Module.py
+++ b/llvmpy/src/Module.py
@@ -16,7 +16,10 @@ from .Metadata import NamedMDNode
 
 @Module
 class Module:
-    _include_ = "llvm/Module.h"
+    if LLVM_VERSION >= (3, 3):
+        _include_ = "llvm/IR/Module.h"
+    else:
+        _include_ = "llvm/Module.h"
     # Enumerators
     Endianness = Enum('AnyEndianness', 'LittleEndian', 'BigEndian')
     PointerSize = Enum('AnyPointerSize', 'Pointer32', 'Pointer64')
diff --git a/llvmpy/src/Support/Host.py b/llvmpy/src/Support/Host.py
index 8163ebe..cd063e8 100644
--- a/llvmpy/src/Support/Host.py
+++ b/llvmpy/src/Support/Host.py
@@ -1,18 +1,29 @@
 from binding import *
 from src.namespace import sys
 
-isLittleEndianHost = sys.Function('isLittleEndianHost',
-                                  cast(Bool, bool))
-
-isBigEndianHost = sys.Function('isBigEndianHost',
-                               cast(Bool, bool))
-
 getDefaultTargetTriple = sys.Function('getDefaultTargetTriple',
                                       cast(ConstStdString, str))
 
-# llvm 3.3
-#getProcessTriple = sys.Function('getProcessTriple',
-#                                cast(ConstStdString, str))
+if LLVM_VERSION >= (3, 3):
+    getProcessTriple = sys.Function('getProcessTriple',
+                                    cast(ConstStdString, str))
+
+    isLittleEndianHost = sys.CustomFunction('isLittleEndianHost',
+                                            'llvm_sys_isLittleEndianHost',
+                                            cast(Bool, bool))
+
+    isBigEndianHost = sys.CustomFunction('isBigEndianHost',
+                                         'llvm_sys_isBigEndianHost',
+                                         cast(Bool, bool))
+
+else:
+    
+    isLittleEndianHost = sys.Function('isLittleEndianHost',
+                                      cast(Bool, bool))
+
+    isBigEndianHost = sys.Function('isBigEndianHost',
+                                   cast(Bool, bool))
+
 
 getHostCPUName = sys.Function('getHostCPUName',
                               cast(ConstStdString, str))
diff --git a/llvmpy/src/Support/TargetRegistry.py b/llvmpy/src/Support/TargetRegistry.py
index 2a445e7..04e08c7 100644
--- a/llvmpy/src/Support/TargetRegistry.py
+++ b/llvmpy/src/Support/TargetRegistry.py
@@ -14,7 +14,7 @@ from src.Support.CodeGen import Reloc, CodeModel, CodeGenOpt
 
 @Target
 class Target:
-    getNext = Method(const(ptr(Target)))
+    getNext = Method(const(ownedptr(Target)))
 
     getName = Method(cast(StdString, str))
     getShortDescription = Method(cast(StdString, str))
diff --git a/llvmpy/src/Target/TargetMachine.py b/llvmpy/src/Target/TargetMachine.py
index 1bd6843..84c0d5d 100644
--- a/llvmpy/src/Target/TargetMachine.py
+++ b/llvmpy/src/Target/TargetMachine.py
@@ -8,8 +8,10 @@ from src.ADT.StringRef import StringRef
 from src.Support.CodeGen import CodeModel, TLSModel, CodeGenOpt, Reloc
 from src.GlobalValue import GlobalValue
 from src.DataLayout import DataLayout
-from src.TargetTransformInfo import (ScalarTargetTransformInfo,
-                                     VectorTargetTransformInfo)
+
+if LLVM_VERSION < (3, 3):
+    from src.TargetTransformInfo import (ScalarTargetTransformInfo,
+                                         VectorTargetTransformInfo)
 from src.PassManager import PassManagerBase
 from src.Support.FormattedStream import formatted_raw_ostream
 
@@ -39,9 +41,11 @@ class TargetMachine:
     setMCUseDwarfDirectory = Method(Void, cast(bool, Bool))
 
     getDataLayout = Method(const(ownedptr(DataLayout)))
-    getScalarTargetTransformInfo = Method(const(
+
+    if LLVM_VERSION < (3, 3):
+        getScalarTargetTransformInfo = Method(const(
                                           ownedptr(ScalarTargetTransformInfo)))
-    getVectorTargetTransformInfo = Method(const(
+        getVectorTargetTransformInfo = Method(const(
                                           ownedptr(VectorTargetTransformInfo)))
 
     addPassesToEmitFile = Method(cast(bool, Bool),
diff --git a/llvmpy/src/TargetTransformInfo.py b/llvmpy/src/TargetTransformInfo.py
index 1208050..d51bc52 100644
--- a/llvmpy/src/TargetTransformInfo.py
+++ b/llvmpy/src/TargetTransformInfo.py
@@ -2,7 +2,10 @@ from binding import *
 from src.namespace import llvm
 from src.Pass import ImmutablePass
 
-llvm.includes.add('llvm/TargetTransformInfo.h')
+if LLVM_VERSION >= (3, 3):
+    llvm.includes.add('llvm/Analysis/TargetTransformInfo.h')
+else:
+    llvm.includes.add('llvm/TargetTransformInfo.h')
 
 TargetTransformInfo = llvm.Class(ImmutablePass)
 ScalarTargetTransformInfo = llvm.Class()
@@ -11,14 +14,17 @@ VectorTargetTransformInfo = llvm.Class()
 
 @ScalarTargetTransformInfo
 class ScalarTargetTransformInfo:
-    delete = Destructor()
+    if LLVM_VERSION < (3, 3):
+        delete = Destructor()
 
 @VectorTargetTransformInfo
 class VectorTargetTransformInfo:
-    delete = Destructor()
+    if LLVM_VERSION < (3, 3):
+        delete = Destructor()
 
 @TargetTransformInfo
 class TargetTransformInfo:
-    new = Constructor(ptr(ScalarTargetTransformInfo),
-                      ptr(VectorTargetTransformInfo))
+    if LLVM_VERSION < (3, 3):
+        new = Constructor(ptr(ScalarTargetTransformInfo),
+                          ptr(VectorTargetTransformInfo))
 
diff --git a/llvmpy/src/Transforms/PassManagerBuilder.py b/llvmpy/src/Transforms/PassManagerBuilder.py
index 86be4fb..146d409 100644
--- a/llvmpy/src/Transforms/PassManagerBuilder.py
+++ b/llvmpy/src/Transforms/PassManagerBuilder.py
@@ -36,7 +36,11 @@ class PassManagerBuilder:
     DisableSimplifyLibCalls = _attr_bool()
     DisableUnitAtATime = _attr_bool()
     DisableUnrollLoops = _attr_bool()
-    Vectorize = _attr_bool()
+    if LLVM_VERSION >= (3, 3):
+        BBVectorize = _attr_bool()
+        SLPVectorize = _attr_bool()
+    else:
+        Vectorize = _attr_bool()
     LoopVectorize = _attr_bool()
 
     LibraryInfo = Attr(getter=ownedptr(TargetLibraryInfo),
diff --git a/llvmpy/src/Type.py b/llvmpy/src/Type.py
index afc148c..50c4862 100644
--- a/llvmpy/src/Type.py
+++ b/llvmpy/src/Type.py
@@ -15,7 +15,10 @@ VectorType = llvm.Class(SequentialType)
 
 @Type
 class Type:
-    _include_ = 'llvm/Type.h'
+    if LLVM_VERSION >= (3, 3):
+        _include_ = 'llvm/IR/Type.h'
+    else:
+        _include_ = 'llvm/Type.h'
 
     TypeID = Enum('''
         VoidTyID, HalfTyID, FloatTyID, DoubleTyID,
diff --git a/llvmpy/src/ValueSymbolTable.py b/llvmpy/src/ValueSymbolTable.py
index 7e8890c..8762156 100644
--- a/llvmpy/src/ValueSymbolTable.py
+++ b/llvmpy/src/ValueSymbolTable.py
@@ -4,7 +4,10 @@ from .ADT.StringRef import StringRef
 
 @ValueSymbolTable
 class ValueSymbolTable:
-    _include_ = 'llvm/ValueSymbolTable.h'
+    if LLVM_VERSION >= (3, 3):
+        _include_ = 'llvm/IR/ValueSymbolTable.h'
+    else:
+        _include_ = 'llvm/ValueSymbolTable.h'
     new = Constructor()
     delete = Destructor()
     lookup = Method(ptr(Value), cast(str, StringRef))
diff --git a/setup.py b/setup.py
index 1dc8af8..15b98e6 100644
--- a/setup.py
+++ b/setup.py
@@ -36,6 +36,7 @@ def run_llvm_config(extra_args):
     return stdout.decode().strip()
 
 llvm_version = run_llvm_config(['--version'])
+
 print('LLVM version = %r' % llvm_version)
 
 targets_built = run_llvm_config(['--targets-built'])
@@ -61,7 +62,12 @@ def auto_intrinsic_gen(incdir):
     # let's do auto intrinsic generation
     print("Generate intrinsic IDs")
     from tools import intrgen
-    path = "%s/llvm/Intrinsics.gen" % incdir
+
+    if llvm_version.startswith('3.3'):
+        path = "%s/llvm/IR/Intrinsics.gen" % incdir
+    else:
+        path = "%s/llvm/Intrinsics.gen" % incdir
+
     with open('llvm/_intrinsic_ids.py', 'w') as fout:
         intrgen.gen(path, fout)
 
@@ -119,7 +125,7 @@ else:
     extra_components = list(extra_components & enabled_components)
 
     libs_core, objs_core = get_libs_and_objs(
-        ['core', 'analysis', 'scalaropts', 'executionengine',
+        ['core', 'analysis', 'scalaropts', 'executionengine', 'mcjit',
          'jit',  'native', 'interpreter', 'bitreader', 'bitwriter',
          'instrumentation', 'ipa', 'ipo', 'transformutils',
          'asmparser', 'linker', 'support', 'vectorize']
@@ -181,7 +187,8 @@ setup(
                 'llvm_cbuilder',
                 'llpython',
                 'llvm_array',
-                'llvmpy.api', 'llvmpy.api.llvm'],
+                'llvmpy.api', 'llvmpy.api.llvm',],
+    package_data = {'llvm': ['llrt/*.ll']},
     py_modules = ['llvmpy',
                   'llvmpy._capsule',
                   'llvmpy._api',