def _emit_method_array(self, llvm_module):
"""
Collect exported methods and emit a PyMethodDef array.
:returns: a pointer to the PyMethodDef array.
"""
method_defs = []
for entry in self.export_entries:
name = entry.symbol
llvm_func_name = self._mangle_method_symbol(name)
fnty = self.exported_function_types[entry]
lfunc = ir.Function(llvm_module, fnty, llvm_func_name)
method_name = self.context.insert_const_string(llvm_module, name)
method_def_const = ir.Constant.literal_struct(
(method_name, ir.Constant.bitcast(lfunc, lt._void_star),
METH_VARARGS_AND_KEYWORDS, NULL))
method_defs.append(method_def_const)
sentinel = ir.Constant.literal_struct([NULL, NULL, ZERO, NULL])
method_defs.append(sentinel)
method_array_init = create_constant_array(self.method_def_ty,
method_defs)
method_array = cgutils.add_global_variable(llvm_module,
method_array_init.type,
'.module_methods')
method_array.initializer = method_array_init
method_array.linkage = 'internal'
method_array_ptr = ir.Constant.gep(method_array, [ZERO, ZERO])
return method_array_ptr
def insert_const_string(self, mod, string):
"""Create a global string from the passed in string argument and return
a void* in the GENERIC address space pointing to that string.
Args:
mod: LLVM module where the global string value is to be inserted.
string: A Python string that will be converted to a global constant
string and inserted into the module.
Returns: A LLVM Constant pointing to the global string value inserted
into the module.
"""
text = cgutils.make_bytearray(string.encode("utf-8") + b"\x00")
name = "$".join(["__conststring__", self.mangler(string, ["str"])])
# Try to reuse existing global
gv = mod.globals.get(name)
if gv is None:
# Not defined yet
gv = cgutils.add_global_variable(mod,
text.type,
name=name,
addrspace=address_space.GENERIC)
gv.linkage = "internal"
gv.global_constant = True
gv.initializer = text
# Cast to a i8* pointer
charty = gv.type.pointee.element
return gv.bitcast(charty.as_pointer(address_space.GENERIC))
def _generic_array(context, builder, shape, dtype, symbol_name, addrspace):
elemcount = reduce(operator.mul, shape, 1)
lldtype = context.get_data_type(dtype)
laryty = Type.array(lldtype, elemcount)
if addrspace == target.SPIR_LOCAL_ADDRSPACE:
lmod = builder.module
# Create global variable in the requested address-space
gvmem = cgutils.add_global_variable(lmod, laryty, symbol_name,
addrspace)
if elemcount <= 0:
raise ValueError("array length <= 0")
else:
gvmem.linkage = lc.LINKAGE_INTERNAL
if dtype not in types.number_domain:
raise TypeError("unsupported type: %s" % dtype)
# Convert to generic address-space
dataptr = context.addrspacecast(builder, gvmem,
target.SPIR_GENERIC_ADDRSPACE)
else:
raise NotImplementedError("addrspace {addrspace}".format(**locals()))
return _make_array(context, builder, dataptr, dtype, shape)
def get_c_value(self, builder, typ, name, dllimport=False):
"""
Get a global value through its C-accessible *name*, with the given
LLVM type.
If *dllimport* is true, the symbol will be marked as imported
from a DLL (necessary for AOT compilation under Windows).
"""
module = builder.function.module
try:
gv = module.globals[name]
except KeyError:
gv = cgutils.add_global_variable(module, typ, name)
if dllimport and self.aot_mode and sys.platform == 'win32':
gv.storage_class = "dllimport"
return gv
def make_constant_array(self, builder, aryty, arr):
"""
Unlike the parent version. This returns a a pointer in the constant
addrspace.
"""
lmod = builder.module
constvals = [
self.get_constant(types.byte, i)
for i in iter(arr.tobytes(order='A'))
]
constaryty = ir.ArrayType(ir.IntType(8), len(constvals))
constary = ir.Constant(constaryty, constvals)
addrspace = nvvm.ADDRSPACE_CONSTANT
gv = cgutils.add_global_variable(lmod,
constary.type,
"_cudapy_cmem",
addrspace=addrspace)
gv.linkage = 'internal'
gv.global_constant = True
gv.initializer = constary
# Preserve the underlying alignment
lldtype = self.get_data_type(aryty.dtype)
align = self.get_abi_sizeof(lldtype)
gv.align = 2**(align - 1).bit_length()
# Convert to generic address-space
conv = nvvmutils.insert_addrspace_conv(lmod, ir.IntType(8), addrspace)
addrspaceptr = gv.bitcast(ir.PointerType(ir.IntType(8), addrspace))
genptr = builder.call(conv, [addrspaceptr])
# Create array object
ary = self.make_array(aryty)(self, builder)
kshape = [self.get_constant(types.intp, s) for s in arr.shape]
kstrides = [self.get_constant(types.intp, s) for s in arr.strides]
self.populate_array(ary,
data=builder.bitcast(genptr, ary.data.type),
shape=kshape,
strides=kstrides,
itemsize=ary.itemsize,
parent=ary.parent,
meminfo=None)
return ary._getvalue()
def declare_string(builder, value):
lmod = builder.basic_block.function.module
cval = cgutils.make_bytearray(value.encode("utf-8") + b"\x00")
gl = cgutils.add_global_variable(lmod,
cval.type,
name="_str",
addrspace=nvvm.ADDRSPACE_CONSTANT)
gl.linkage = 'internal'
gl.global_constant = True
gl.initializer = cval
charty = ir.IntType(8)
constcharptrty = ir.PointerType(charty, nvvm.ADDRSPACE_CONSTANT)
charptr = builder.bitcast(gl, constcharptrty)
conv = insert_addrspace_conv(lmod, charty, nvvm.ADDRSPACE_CONSTANT)
return builder.call(conv, [charptr])
def add_dynamic_addr(self, builder, intaddr, info):
"""
Returns dynamic address as a void pointer `i8*`.
Internally, a global variable is added to inform the lowerer about
the usage of dynamic addresses. Caching will be disabled.
"""
assert self.allow_dynamic_globals, "dyn globals disabled in this target"
assert isinstance(intaddr, int), 'dyn addr not of int type'
mod = builder.module
llvoidptr = self.get_value_type(types.voidptr)
addr = self.get_constant(types.uintp, intaddr).inttoptr(llvoidptr)
# Use a unique name by embedding the address value
symname = 'numba.dynamic.globals.{:x}'.format(intaddr)
gv = cgutils.add_global_variable(mod, llvoidptr, symname)
# Use linkonce linkage to allow merging with other GV of the same name.
# And, avoid optimization from assuming its value.
gv.linkage = 'linkonce'
gv.initializer = addr
return builder.load(gv)
def insert_const_string(self, mod, string):
"""
Unlike the parent version. This returns a a pointer in the constant
addrspace.
"""
text = cgutils.make_bytearray(string.encode("utf-8") + b"\x00")
name = '$'.join(["__conststring__",
itanium_mangler.mangle_identifier(string)])
# Try to reuse existing global
gv = mod.globals.get(name)
if gv is None:
# Not defined yet
gv = cgutils.add_global_variable(mod, text.type, name,
addrspace=nvvm.ADDRSPACE_CONSTANT)
gv.linkage = 'internal'
gv.global_constant = True
gv.initializer = text
# Cast to a i8* pointer
charty = gv.type.pointee.element
return gv.bitcast(charty.as_pointer(nvvm.ADDRSPACE_CONSTANT))
def define_error_gv(postfix):
name = wrapfn.name + postfix
gv = cgutils.add_global_variable(wrapper_module, ir.IntType(32),
name)
gv.initializer = ir.Constant(gv.type.pointee, None)
return gv
def _generic_array(context,
builder,
shape,
dtype,
symbol_name,
addrspace,
can_dynsized=False):
elemcount = reduce(operator.mul, shape, 1)
# Check for valid shape for this type of allocation.
# Only 1d arrays can be dynamic.
dynamic_smem = elemcount <= 0 and can_dynsized and len(shape) == 1
if elemcount <= 0 and not dynamic_smem:
raise ValueError("array length <= 0")
# Check that we support the requested dtype
data_model = context.data_model_manager[dtype]
other_supported_type = (isinstance(dtype, (types.Record, types.Boolean))
or isinstance(data_model, models.StructModel))
if dtype not in types.number_domain and not other_supported_type:
raise TypeError("unsupported type: %s" % dtype)
lldtype = context.get_data_type(dtype)
laryty = ir.ArrayType(lldtype, elemcount)
if addrspace == nvvm.ADDRSPACE_LOCAL:
# Special case local address space allocation to use alloca
# NVVM is smart enough to only use local memory if no register is
# available
dataptr = cgutils.alloca_once(builder, laryty, name=symbol_name)
else:
lmod = builder.module
# Create global variable in the requested address space
gvmem = cgutils.add_global_variable(lmod, laryty, symbol_name,
addrspace)
# Specify alignment to avoid misalignment bug
align = context.get_abi_sizeof(lldtype)
# Alignment is required to be a power of 2 for shared memory. If it is
# not a power of 2 (e.g. for a Record array) then round up accordingly.
gvmem.align = 1 << (align - 1).bit_length()
if dynamic_smem:
gvmem.linkage = 'external'
else:
## Comment out the following line to workaround a NVVM bug
## which generates a invalid symbol name when the linkage
## is internal and in some situation.
## See _get_unique_smem_id()
# gvmem.linkage = lc.LINKAGE_INTERNAL
gvmem.initializer = ir.Constant(laryty, ir.Undefined)
# Convert to generic address-space
conv = nvvmutils.insert_addrspace_conv(lmod, ir.IntType(8), addrspace)
addrspaceptr = gvmem.bitcast(ir.PointerType(ir.IntType(8), addrspace))
dataptr = builder.call(conv, [addrspaceptr])
targetdata = _get_target_data(context)
lldtype = context.get_data_type(dtype)
itemsize = lldtype.get_abi_size(targetdata)
# Compute strides
laststride = itemsize
rstrides = []
for i, lastsize in enumerate(reversed(shape)):
rstrides.append(laststride)
laststride *= lastsize
strides = [s for s in reversed(rstrides)]
kstrides = [context.get_constant(types.intp, s) for s in strides]
# Compute shape
if dynamic_smem:
# Compute the shape based on the dynamic shared memory configuration.
# Unfortunately NVVM does not provide an intrinsic for the
# %dynamic_smem_size register, so we must read it using inline
# assembly.
get_dynshared_size = ir.InlineAsm(ir.FunctionType(ir.IntType(32), []),
"mov.u32 $0, %dynamic_smem_size;",
'=r',
side_effect=True)
dynsmem_size = builder.zext(builder.call(get_dynshared_size, []),
ir.IntType(64))
# Only 1-D dynamic shared memory is supported so the following is a
# sufficient construction of the shape
kitemsize = context.get_constant(types.intp, itemsize)
kshape = [builder.udiv(dynsmem_size, kitemsize)]
else:
kshape = [context.get_constant(types.intp, s) for s in shape]
# Create array object
ndim = len(shape)
aryty = types.Array(dtype=dtype, ndim=ndim, layout='C')
ary = context.make_array(aryty)(context, builder)
context.populate_array(ary,
data=builder.bitcast(dataptr, ary.data.type),
shape=kshape,
strides=kstrides,
itemsize=context.get_constant(types.intp, itemsize),
meminfo=None)
return ary._getvalue()
def _emit_python_wrapper(self, llvm_module):
# Figure out the Python C API module creation function, and
# get a LLVM function for it.
create_module_fn = ir.Function(llvm_module,
*self.module_create_definition)
create_module_fn.linkage = 'external'
# Define a constant string for the module name.
mod_name_const = self.context.insert_const_string(
llvm_module, self.module_name)
mod_def_base_init = ir.Constant.literal_struct((
lt._pyobject_head_init, # PyObject_HEAD
ir.Constant(self.m_init_ty, None), # m_init
ir.Constant(lt._llvm_py_ssize_t, None), # m_index
ir.Constant(lt._pyobject_head_p, None), # m_copy
))
mod_def_base = cgutils.add_global_variable(llvm_module,
mod_def_base_init.type,
'.module_def_base')
mod_def_base.initializer = mod_def_base_init
mod_def_base.linkage = 'internal'
method_array = self._emit_method_array(llvm_module)
mod_def_init = ir.Constant.literal_struct((
mod_def_base_init, # m_base
mod_name_const, # m_name
ir.Constant(self._char_star, None), # m_doc
ir.Constant(lt._llvm_py_ssize_t, -1), # m_size
method_array, # m_methods
ir.Constant(self.inquiry_ty, None), # m_reload
ir.Constant(self.traverseproc_ty, None), # m_traverse
ir.Constant(self.inquiry_ty, None), # m_clear
ir.Constant(self.freefunc_ty, None) # m_free
))
# Define a constant string for the module name.
mod_def = cgutils.add_global_variable(llvm_module, mod_def_init.type,
'.module_def')
mod_def.initializer = mod_def_init
mod_def.linkage = 'internal'
# Define the module initialization function.
mod_init_fn = ir.Function(llvm_module, *self.module_init_definition)
entry = mod_init_fn.append_basic_block('Entry')
builder = ir.IRBuilder(entry)
pyapi = self.context.get_python_api(builder)
mod = builder.call(create_module_fn,
(mod_def, ir.Constant(lt._int32, sys.api_version)))
# Test if module has been created correctly.
# (XXX for some reason comparing with the NULL constant fails llvm
# with an assertion in pydebug mode)
with builder.if_then(cgutils.is_null(builder, mod)):
builder.ret(NULL.bitcast(mod_init_fn.type.pointee.return_type))
env_array = self._emit_environment_array(llvm_module, builder, pyapi)
envgv_array = self._emit_envgvs_array(llvm_module, builder, pyapi)
ret = self._emit_module_init_code(llvm_module, builder, mod,
method_array, env_array, envgv_array)
if ret is not None:
with builder.if_then(cgutils.is_not_null(builder, ret)):
# Init function errored out
builder.ret(ir.Constant(mod.type, None))
builder.ret(mod)
self.dll_exports.append(mod_init_fn.name)