diff --git a/llvm/_core.cpp b/llvm/_core.cpp
index 5acfa83..6f13c61 100644
--- a/llvm/_core.cpp
+++ b/llvm/_core.cpp
@@ -878,6 +878,9 @@ _wrap_objobj2obj(LLVMBuildFree, LLVMBuilderRef, LLVMValueRef, LLVMValueRef)
 _wrap_objobjstr2obj(LLVMBuildLoad, LLVMBuilderRef, LLVMValueRef, LLVMValueRef)
 _wrap_objobjobj2obj(LLVMBuildStore, LLVMBuilderRef, LLVMValueRef, LLVMValueRef, LLVMValueRef)
 _wrap_objobjliststr2obj(LLVMBuildGEP, LLVMBuilderRef, LLVMValueRef, LLVMValueRef, LLVMValueRef)
+_wrap_objobjliststr2obj(LLVMBuildInBoundsGEP, LLVMBuilderRef, LLVMValueRef, LLVMValueRef, LLVMValueRef)
+
+
 
 _wrap_objint2none(LLVMLdSetAlignment, LLVMValueRef)
 _wrap_objint2none(LLVMStSetAlignment, LLVMValueRef)
@@ -1959,6 +1962,7 @@ static PyMethodDef core_methods[] = {
     _method( LLVMBuildLoad )
     _method( LLVMBuildStore )
     _method( LLVMBuildGEP )
+    _method( LLVMBuildInBoundsGEP )
 
     _method( LLVMLdSetAlignment )
     _method( LLVMStSetAlignment )
diff --git a/llvm/core.py b/llvm/core.py
index 4e59675..fa9dc77 100644
--- a/llvm/core.py
+++ b/llvm/core.py
@@ -2066,11 +2066,14 @@ class Builder(object):
             inst.set_volatile(volatile)
         return inst
 
-    def gep(self, ptr, indices, name=""):
+    def gep(self, ptr, indices, name="", inbounds=False):
         check_is_value(ptr)
         index_ptrs = unpack_values(indices)
-        return _make_value(
-            _core.LLVMBuildGEP(self.ptr, ptr.ptr, index_ptrs, name))
+        if inbounds:
+            ret = _core.LLVMBuildInBoundsGEP(self.ptr, ptr.ptr, index_ptrs, name)
+        else:
+            ret = _core.LLVMBuildGEP(self.ptr, ptr.ptr, index_ptrs, name)
+        return _make_value(ret)
 
     # casts and extensions
 
diff --git a/llvm_cbuilder/builder.py b/llvm_cbuilder/builder.py
index 6aa5e1d..64083b1 100644
--- a/llvm_cbuilder/builder.py
+++ b/llvm_cbuilder/builder.py
@@ -818,8 +818,7 @@ class CDefinition(object):
         if optimize:
             fpm = lp.FunctionPassManager.new(module)
             pmb = lp.PassManagerBuilder.new()
-            pmb.opt_level = 3
-            pmb.vectorize = True
+            pmb.opt_level = 2
             pmb.populate(fpm)
             fpm.run(func)
 
@@ -1135,11 +1134,11 @@ class PointerIndexing(OperatorMixin):
             # Case #1: A[idx:] get pointer offset by idx
             if not idx.step and not idx.stop:
                 idx = _auto_coerce_index(self.parent, idx.start)
-                ptr = bldr.gep(self.value, [idx.value])
+                ptr = bldr.gep(self.value, [idx.value], inbounds=True)
                 return CArray(self.parent, ptr)
         else: # return an variable at idx
             idx = _auto_coerce_index(self.parent, idx)
-            ptr = bldr.gep(self.value, [idx.value])
+            ptr = bldr.gep(self.value, [idx.value], inbounds=True)
             return CVar(self.parent, ptr)
 
     def __setitem__(self, idx, val):
@@ -1184,7 +1183,7 @@ class PointerValue(PointerIndexing, PointerCasting):
         inst = self.parent.builder.store(val.value, self.value, **kws)
         if nontemporal:
             self.parent.set_memop_non_temporal(inst)
-
+        return inst
 
     def atomic_load(self, ordering, align=None, crossthread=True):
         '''atomic load memory for pointer types