def get_constant(self, ty, val):
assert not self.is_struct_type(ty)
lty = self.get_value_type(ty)
if ty == types.none:
assert val is None
return self.get_dummy_value()
elif ty == types.boolean:
return Constant.int(Type.int(1), int(val))
elif ty in types.signed_domain:
return Constant.int_signextend(lty, val)
elif ty in types.unsigned_domain:
return Constant.int(lty, val)
elif ty in types.real_domain:
return Constant.real(lty, val)
elif isinstance(ty, (types.NPDatetime, types.NPTimedelta)):
return Constant.real(lty, val.astype(numpy.int64))
elif isinstance(ty, (types.UniTuple, types.NamedUniTuple)):
consts = [self.get_constant(ty.dtype, v) for v in val]
return Constant.array(consts[0].type, consts)
raise NotImplementedError("cannot lower constant of type '%s'" % (ty,))
def _long_from_native_int(self, ival, func_name, native_int_type,
signed):
fnty = Type.function(self.pyobj, [native_int_type])
fn = self._get_function(fnty, name=func_name)
resptr = cgutils.alloca_once(self.builder, self.pyobj)
if PYVERSION < (3, 0):
# Under Python 2, we try to return a PyInt object whenever
# the given number fits in a C long.
pyint_fnty = Type.function(self.pyobj, [self.long])
pyint_fn = self._get_function(pyint_fnty, name="PyInt_FromLong")
long_max = Constant.int(native_int_type, _helperlib.long_max)
if signed:
long_min = Constant.int(native_int_type, _helperlib.long_min)
use_pyint = self.builder.and_(
self.builder.icmp(lc.ICMP_SGE, ival, long_min),
self.builder.icmp(lc.ICMP_SLE, ival, long_max),
)
else:
use_pyint = self.builder.icmp(lc.ICMP_ULE, ival, long_max)
with self.builder.if_else(use_pyint) as (then, otherwise):
with then:
downcast_ival = self.builder.trunc(ival, self.long)
res = self.builder.call(pyint_fn, [downcast_ival])
self.builder.store(res, resptr)
with otherwise:
res = self.builder.call(fn, [ival])
self.builder.store(res, resptr)
else:
fn = self._get_function(fnty, name=func_name)
self.builder.store(self.builder.call(fn, [ival]), resptr)
return self.builder.load(resptr)
def for_range(builder, count, intp):
start = Constant.int(intp, 0)
stop = count
bbcond = append_basic_block(builder, "for.cond")
bbbody = append_basic_block(builder, "for.body")
bbend = append_basic_block(builder, "for.end")
bbstart = builder.basic_block
builder.branch(bbcond)
ONE = Constant.int(intp, 1)
with goto_block(builder, bbcond):
index = builder.phi(intp, name="loop.index")
pred = builder.icmp(lc.ICMP_SLT, index, stop)
builder.cbranch(pred, bbbody, bbend)
with goto_block(builder, bbbody):
yield index
bbbody = builder.basic_block
incr = builder.add(index, ONE)
terminate(builder, bbcond)
index.add_incoming(start, bbstart)
index.add_incoming(incr, bbbody)
builder.position_at_end(bbend)
def __init__(self, context, builder, value=None, ref=None, cast_ref=False):
self._type = context.get_struct_type(self)
self._context = context
self._builder = builder
if ref is None:
self._value = alloca_once(builder, self._type)
if value is not None:
assert not is_pointer(value.type)
assert value.type == self._type, (value.type, self._type)
builder.store(value, self._value)
else:
assert value is None
assert is_pointer(ref.type)
if self._type != ref.type.pointee:
if cast_ref:
ref = builder.bitcast(ref, Type.pointer(self._type))
else:
raise TypeError(
"mismatching pointer type: got %s, expected %s"
% (ref.type.pointee, self._type))
self._value = ref
self._namemap = {}
self._fdmap = []
self._typemap = []
base = Constant.int(Type.int(), 0)
for i, (k, tp) in enumerate(self._fields):
self._namemap[k] = i
self._fdmap.append((base, Constant.int(Type.int(), i)))
self._typemap.append(tp)
def __init__(self, context, builder, value=None, ref=None, cast_ref=False):
self._type = context.get_struct_type(self)
self._context = context
self._builder = builder
if ref is None:
self._value = alloca_once(builder, self._type)
if value is not None:
assert not is_pointer(value.type)
assert value.type == self._type, (value.type, self._type)
builder.store(value, self._value)
else:
assert value is None
assert is_pointer(ref.type)
if self._type != ref.type.pointee:
if cast_ref:
ref = builder.bitcast(ref, Type.pointer(self._type))
else:
raise TypeError(
"mismatching pointer type: got %s, expected %s"
% (ref.type.pointee, self._type))
self._value = ref
self._namemap = {}
self._fdmap = []
self._typemap = []
base = Constant.int(Type.int(), 0)
for i, (k, tp) in enumerate(self._fields):
self._namemap[k] = i
self._fdmap.append((base, Constant.int(Type.int(), i)))
self._typemap.append(tp)
def timedelta_floor_div_timedelta(context, builder, sig, args):
[va, vb] = args
[ta, tb] = sig.args
ll_ret_type = context.get_value_type(sig.return_type)
not_nan = are_not_nat(builder, [va, vb])
ret = cgutils.alloca_once(builder, ll_ret_type, name='ret')
zero = Constant.int(ll_ret_type, 0)
one = Constant.int(ll_ret_type, 1)
builder.store(zero, ret)
with cgutils.if_likely(builder, not_nan):
va, vb = normalize_timedeltas(context, builder, va, vb, ta, tb)
# is the denominator zero or NaT?
denom_ok = builder.not_(builder.icmp_signed('==', vb, zero))
with cgutils.if_likely(builder, denom_ok):
# is either arg negative?
vaneg = builder.icmp_signed('<', va, zero)
neg = builder.or_(vaneg, builder.icmp_signed('<', vb, zero))
with builder.if_else(neg) as (then, otherwise):
with then: # one or more value negative
with builder.if_else(vaneg) as (negthen, negotherwise):
with negthen:
top = builder.sub(va, one)
div = builder.sdiv(top, vb)
builder.store(div, ret)
with negotherwise:
top = builder.add(va, one)
div = builder.sdiv(top, vb)
builder.store(div, ret)
with otherwise:
div = builder.sdiv(va, vb)
builder.store(div, ret)
res = builder.load(ret)
return impl_ret_untracked(context, builder, sig.return_type, res)
def divmod_by_constant(builder, val, divisor):
"""
Compute the (quotient, remainder) of *val* divided by the constant
positive *divisor*. The semantics reflects those of Python integer
floor division, rather than C's / LLVM's signed division and modulo.
The difference lies with a negative *val*.
"""
assert divisor > 0
divisor = Constant.int(val.type, divisor)
one = Constant.int(val.type, 1)
quot = alloca_once(builder, val.type)
with ifelse(builder, is_neg_int(builder, val)) as (if_neg, if_pos):
with if_pos:
# quot = val / divisor
quot_val = builder.sdiv(val, divisor)
builder.store(quot_val, quot)
with if_neg:
# quot = -1 + (val + 1) / divisor
val_plus_one = builder.add(val, one)
quot_val = builder.sdiv(val_plus_one, divisor)
builder.store(builder.sub(quot_val, one), quot)
# rem = val - quot * divisor
# (should be slightly faster than a separate modulo operation)
quot_val = builder.load(quot)
rem_val = builder.sub(val, builder.mul(quot_val, divisor))
return quot_val, rem_val
def get_constant(self, ty, val):
assert not self.is_struct_type(ty)
lty = self.get_value_type(ty)
if ty == types.none:
assert val is None
return self.get_dummy_value()
elif ty == types.boolean:
return Constant.int(Type.int(1), int(val))
elif ty in types.signed_domain:
return Constant.int_signextend(lty, val)
elif ty in types.unsigned_domain:
return Constant.int(lty, val)
elif ty in types.real_domain:
return Constant.real(lty, val)
elif isinstance(ty, types.UniTuple):
consts = [self.get_constant(ty.dtype, v) for v in val]
return Constant.array(consts[0].type, consts)
raise NotImplementedError("cannot lower constant of type '%s'" % (ty,))
def make_exception_switch(self, api, builder, code):
"""Handle user defined exceptions.
Build a switch to check which exception class was raised.
"""
nexc = len(self.exceptions)
elseblk = cgutils.append_basic_block(builder, ".invalid.user.exception")
swt = builder.switch(code, elseblk, n=nexc)
for num, exc in self.exceptions.items():
bb = cgutils.append_basic_block(builder,
".user.exception.%d" % num)
swt.add_case(Constant.int(code.type, num), bb)
builder.position_at_end(bb)
api.raise_exception(exc, exc)
builder.ret(api.get_null_object())
builder.position_at_end(elseblk)
# Handle native error
elseblk = cgutils.append_basic_block(builder, ".invalid.native.error")
swt = builder.switch(code, elseblk, n=len(errcode.error_names))
msgfmt = "{error} in native function: {fname}"
for errnum, errname in errcode.error_names.items():
bb = cgutils.append_basic_block(builder,
".native.error.%d" % errnum)
swt.add_case(Constant.int(code.type, errnum), bb)
builder.position_at_end(bb)
api.raise_native_error(msgfmt.format(error=errname,
fname=self.fndesc.mangled_name))
builder.ret(api.get_null_object())
builder.position_at_end(elseblk)
msg = "unknown error in native function: %s" % self.fndesc.mangled_name
api.raise_native_error(msg)
def for_range(builder, count, intp):
start = Constant.int(intp, 0)
stop = count
bbcond = builder.append_basic_block("for.cond")
bbbody = builder.append_basic_block("for.body")
bbend = builder.append_basic_block("for.end")
bbstart = builder.basic_block
builder.branch(bbcond)
ONE = Constant.int(intp, 1)
with builder.goto_block(bbcond):
index = builder.phi(intp, name="loop.index")
pred = builder.icmp(lc.ICMP_SLT, index, stop)
builder.cbranch(pred, bbbody, bbend)
with builder.goto_block(bbbody):
yield index
bbbody = builder.basic_block
incr = builder.add(index, ONE)
terminate(builder, bbcond)
index.add_incoming(start, bbstart)
index.add_incoming(incr, bbbody)
builder.position_at_end(bbend)
def int_power_func_body(context, builder, x, y):
pcounter = cgutils.alloca_once(builder, y.type)
presult = cgutils.alloca_once(builder, x.type)
result = Constant.int(x.type, 1)
counter = y
builder.store(counter, pcounter)
builder.store(result, presult)
bbcond = cgutils.append_basic_block(builder, ".cond")
bbbody = cgutils.append_basic_block(builder, ".body")
bbexit = cgutils.append_basic_block(builder, ".exit")
del counter
del result
builder.branch(bbcond)
with cgutils.goto_block(builder, bbcond):
counter = builder.load(pcounter)
ONE = Constant.int(counter.type, 1)
ZERO = Constant.null(counter.type)
builder.store(builder.sub(counter, ONE), pcounter)
pred = builder.icmp(lc.ICMP_SGT, counter, ZERO)
builder.cbranch(pred, bbbody, bbexit)
with cgutils.goto_block(builder, bbbody):
result = builder.load(presult)
builder.store(builder.mul(result, x), presult)
builder.branch(bbcond)
builder.position_at_end(bbexit)
return builder.load(presult)
def int_power_func_body(context, builder, x, y):
pcounter = cgutils.alloca_once(builder, y.type)
presult = cgutils.alloca_once(builder, x.type)
result = Constant.int(x.type, 1)
counter = y
builder.store(counter, pcounter)
builder.store(result, presult)
bbcond = cgutils.append_basic_block(builder, ".cond")
bbbody = cgutils.append_basic_block(builder, ".body")
bbexit = cgutils.append_basic_block(builder, ".exit")
del counter
del result
builder.branch(bbcond)
with cgutils.goto_block(builder, bbcond):
counter = builder.load(pcounter)
ONE = Constant.int(counter.type, 1)
ZERO = Constant.null(counter.type)
builder.store(builder.sub(counter, ONE), pcounter)
pred = builder.icmp(lc.ICMP_SGT, counter, ZERO)
builder.cbranch(pred, bbbody, bbexit)
with cgutils.goto_block(builder, bbbody):
result = builder.load(presult)
builder.store(builder.mul(result, x), presult)
builder.branch(bbcond)
builder.position_at_end(bbexit)
return builder.load(presult)
def divmod_by_constant(builder, val, divisor):
"""
Compute the (quotient, remainder) of *val* divided by the constant
positive *divisor*. The semantics reflects those of Python integer
floor division, rather than C's / LLVM's signed division and modulo.
The difference lies with a negative *val*.
"""
assert divisor > 0
divisor = Constant.int(val.type, divisor)
one = Constant.int(val.type, 1)
quot = alloca_once(builder, val.type)
with builder.if_else(is_neg_int(builder, val)) as (if_neg, if_pos):
with if_pos:
# quot = val / divisor
quot_val = builder.sdiv(val, divisor)
builder.store(quot_val, quot)
with if_neg:
# quot = -1 + (val + 1) / divisor
val_plus_one = builder.add(val, one)
quot_val = builder.sdiv(val_plus_one, divisor)
builder.store(builder.sub(quot_val, one), quot)
# rem = val - quot * divisor
# (should be slightly faster than a separate modulo operation)
quot_val = builder.load(quot)
rem_val = builder.sub(val, builder.mul(quot_val, divisor))
return quot_val, rem_val
def timedelta_floor_div_timedelta(context, builder, sig, args):
[va, vb] = args
[ta, tb] = sig.args
ll_ret_type = context.get_value_type(sig.return_type)
not_nan = are_not_nat(builder, [va, vb])
ret = cgutils.alloca_once(builder, ll_ret_type, name='ret')
zero = Constant.int(ll_ret_type, 0)
one = Constant.int(ll_ret_type, 1)
builder.store(zero, ret)
with cgutils.if_likely(builder, not_nan):
va, vb = normalize_timedeltas(context, builder, va, vb, ta, tb)
# is the denominator zero or NaT?
denom_ok = builder.not_(builder.icmp_signed('==', vb, zero))
with cgutils.if_likely(builder, denom_ok):
# is either arg negative?
vaneg = builder.icmp_signed('<', va, zero)
neg = builder.or_(vaneg, builder.icmp_signed('<', vb, zero))
with builder.if_else(neg) as (then, otherwise):
with then: # one or more value negative
with builder.if_else(vaneg) as (negthen, negotherwise):
with negthen:
top = builder.sub(va, one)
div = builder.sdiv(top, vb)
builder.store(div, ret)
with negotherwise:
top = builder.add(va, one)
div = builder.sdiv(top, vb)
builder.store(div, ret)
with otherwise:
div = builder.sdiv(va, vb)
builder.store(div, ret)
res = builder.load(ret)
return impl_ret_untracked(context, builder, sig.return_type, res)
def get_constant(self, ty, val):
assert not self.is_struct_type(ty)
lty = self.get_value_type(ty)
if ty == types.none:
assert val is None
return self.get_dummy_value()
elif ty == types.boolean:
return Constant.int(Type.int(1), int(val))
elif ty in types.signed_domain:
return Constant.int_signextend(lty, val)
elif ty in types.unsigned_domain:
return Constant.int(lty, val)
elif ty in types.real_domain:
return Constant.real(lty, val)
elif isinstance(ty, types.UniTuple):
consts = [self.get_constant(ty.dtype, v) for v in val]
return Constant.array(consts[0].type, consts)
raise NotImplementedError(ty)
def make_exception_switch(self, api, builder, code):
"""Handle user defined exceptions.
Build a switch to check which exception class was raised.
"""
nexc = len(self.exceptions)
elseblk = cgutils.append_basic_block(builder, ".invalid.user.exception")
swt = builder.switch(code, elseblk, n=nexc)
for num, exc in self.exceptions.items():
bb = cgutils.append_basic_block(builder,
".user.exception.%d" % num)
swt.add_case(Constant.int(code.type, num), bb)
builder.position_at_end(bb)
api.raise_exception(exc, exc)
builder.ret(api.get_null_object())
builder.position_at_end(elseblk)
# Handle native error
elseblk = cgutils.append_basic_block(builder, ".invalid.native.error")
swt = builder.switch(code, elseblk, n=len(errcode.error_names))
msgfmt = "{error} in native function: {fname}"
for errnum, errname in errcode.error_names.items():
bb = cgutils.append_basic_block(builder,
".native.error.%d" % errnum)
swt.add_case(Constant.int(code.type, errnum), bb)
builder.position_at_end(bb)
api.raise_native_error(msgfmt.format(error=errname,
fname=self.fndesc.mangled_name))
builder.ret(api.get_null_object())
builder.position_at_end(elseblk)
msg = "unknown error in native function: %s" % self.fndesc.mangled_name
api.raise_native_error(msg)
def _long_from_native_int(self, ival, func_name, native_int_type,
signed):
fnty = Type.function(self.pyobj, [native_int_type])
fn = self._get_function(fnty, name=func_name)
resptr = cgutils.alloca_once(self.builder, self.pyobj)
if PYVERSION < (3, 0):
# Under Python 2, we try to return a PyInt object whenever
# the given number fits in a C long.
pyint_fnty = Type.function(self.pyobj, [self.long])
pyint_fn = self._get_function(pyint_fnty, name="PyInt_FromLong")
long_max = Constant.int(native_int_type, _helperlib.long_max)
if signed:
long_min = Constant.int(native_int_type, _helperlib.long_min)
use_pyint = self.builder.and_(
self.builder.icmp(lc.ICMP_SGE, ival, long_min),
self.builder.icmp(lc.ICMP_SLE, ival, long_max),
)
else:
use_pyint = self.builder.icmp(lc.ICMP_ULE, ival, long_max)
with self.builder.if_else(use_pyint) as (then, otherwise):
with then:
downcast_ival = self.builder.trunc(ival, self.long)
res = self.builder.call(pyint_fn, [downcast_ival])
self.builder.store(res, resptr)
with otherwise:
res = self.builder.call(fn, [ival])
self.builder.store(res, resptr)
else:
fn = self._get_function(fnty, name=func_name)
self.builder.store(self.builder.call(fn, [ival]), resptr)
return self.builder.load(resptr)
def unpack_tuple(self, args, name, n_min, n_max, *objs):
charptr = Type.pointer(Type.int(8))
argtypes = [self.pyobj, charptr, self.py_ssize_t, self.py_ssize_t]
fnty = Type.function(Type.int(), argtypes, var_arg=True)
fn = self._get_function(fnty, name="PyArg_UnpackTuple")
n_min = Constant.int(self.py_ssize_t, n_min)
n_max = Constant.int(self.py_ssize_t, n_max)
if isinstance(name, str):
name = self.context.insert_const_string(self.builder.module, name)
return self.builder.call(fn, [args, name, n_min, n_max] + list(objs))
def unpack_tuple(self, args, name, n_min, n_max, *objs):
charptr = Type.pointer(Type.int(8))
argtypes = [self.pyobj, charptr, self.py_ssize_t, self.py_ssize_t]
fnty = Type.function(Type.int(), argtypes, var_arg=True)
fn = self._get_function(fnty, name="PyArg_UnpackTuple")
n_min = Constant.int(self.py_ssize_t, n_min)
n_max = Constant.int(self.py_ssize_t, n_max)
if isinstance(name, str):
name = self.context.insert_const_string(self.builder.module, name)
return self.builder.call(fn, [args, name, n_min, n_max] + list(objs))
def get_item_pointer2(builder,
data,
shape,
strides,
layout,
inds,
wraparound=False):
if wraparound:
# Wraparound
indices = []
for ind, dimlen in zip(inds, shape):
ZERO = Constant.null(ind.type)
negative = builder.icmp(lc.ICMP_SLT, ind, ZERO)
wrapped = builder.add(dimlen, ind)
selected = builder.select(negative, wrapped, ind)
indices.append(selected)
else:
indices = inds
if not indices:
# Indexing with empty tuple
return builder.gep(data, [get_null_value(Type.int(32))])
intp = indices[0].type
# Indexing code
if layout in 'CF':
steps = []
# Compute steps for each dimension
if layout == 'C':
# C contiguous
for i in range(len(shape)):
last = Constant.int(intp, 1)
for j in shape[i + 1:]:
last = builder.mul(last, j)
steps.append(last)
elif layout == 'F':
# F contiguous
for i in range(len(shape)):
last = Constant.int(intp, 1)
for j in shape[:i]:
last = builder.mul(last, j)
steps.append(last)
else:
raise Exception("unreachable")
# Compute index
loc = Constant.int(intp, 0)
for i, s in zip(indices, steps):
tmp = builder.mul(i, s)
loc = builder.add(loc, tmp)
ptr = builder.gep(data, [loc])
return ptr
else:
# Any layout
dimoffs = [builder.mul(s, i) for s, i in zip(strides, indices)]
offset = functools.reduce(builder.add, dimoffs)
return pointer_add(builder, data, offset)
def is_leap_year(builder, year_val):
"""
Return a predicate indicating whether *year_val* (offset by 1970) is a
leap year.
"""
actual_year = builder.add(year_val, Constant.int(DATETIME64, 1970))
multiple_of_4 = cgutils.is_null(
builder, builder.and_(actual_year, Constant.int(DATETIME64, 3)))
not_multiple_of_100 = cgutils.is_not_null(
builder, builder.srem(actual_year, Constant.int(DATETIME64, 100)))
multiple_of_400 = cgutils.is_null(
builder, builder.srem(actual_year, Constant.int(DATETIME64, 400)))
return builder.and_(multiple_of_4,
builder.or_(not_multiple_of_100, multiple_of_400))
def get_item_pointer2(builder, data, shape, strides, layout, inds,
wraparound=False):
if wraparound:
# Wraparound
indices = []
for ind, dimlen in zip(inds, shape):
ZERO = Constant.null(ind.type)
negative = builder.icmp(lc.ICMP_SLT, ind, ZERO)
wrapped = builder.add(dimlen, ind)
selected = builder.select(negative, wrapped, ind)
indices.append(selected)
else:
indices = inds
if not indices:
# Indexing with empty tuple
return builder.gep(data, [get_null_value(Type.int(32))])
intp = indices[0].type
# Indexing code
if layout in 'CF':
steps = []
# Compute steps for each dimension
if layout == 'C':
# C contiguous
for i in range(len(shape)):
last = Constant.int(intp, 1)
for j in shape[i + 1:]:
last = builder.mul(last, j)
steps.append(last)
elif layout == 'F':
# F contiguous
for i in range(len(shape)):
last = Constant.int(intp, 1)
for j in shape[:i]:
last = builder.mul(last, j)
steps.append(last)
else:
raise Exception("unreachable")
# Compute index
loc = Constant.int(intp, 0)
for i, s in zip(indices, steps):
tmp = builder.mul(i, s)
loc = builder.add(loc, tmp)
ptr = builder.gep(data, [loc])
return ptr
else:
# Any layout
dimoffs = [builder.mul(s, i) for s, i in zip(strides, indices)]
offset = functools.reduce(builder.add, dimoffs)
return pointer_add(builder, data, offset)
def is_leap_year(builder, year_val):
"""
Return a predicate indicating whether *year_val* (offset by 1970) is a
leap year.
"""
actual_year = builder.add(year_val, Constant.int(DATETIME64, 1970))
multiple_of_4 = cgutils.is_null(
builder, builder.and_(actual_year, Constant.int(DATETIME64, 3)))
not_multiple_of_100 = cgutils.is_not_null(
builder, builder.srem(actual_year, Constant.int(DATETIME64, 100)))
multiple_of_400 = cgutils.is_null(
builder, builder.srem(actual_year, Constant.int(DATETIME64, 400)))
return builder.and_(multiple_of_4,
builder.or_(not_multiple_of_100, multiple_of_400))
def memset(builder, ptr, size, value):
"""
Fill *size* bytes starting from *ptr* with *value*.
"""
sizety = size.type
memset = "llvm.memset.p0i8.i%d" % (sizety.width)
i32 = lc.Type.int(32)
i8 = lc.Type.int(8)
i8_star = i8.as_pointer()
i1 = lc.Type.int(1)
fn = builder.module.declare_intrinsic('llvm.memset', (i8_star, size.type))
ptr = builder.bitcast(ptr, i8_star)
if isinstance(value, int):
value = Constant.int(i8, value)
builder.call(fn, [ptr, value, size,
Constant.int(i32, 0), Constant.int(i1, 0)])
def lower_finalize_func_body(self, builder, genptr):
"""
Lower the body of the generator's finalizer: decref all live
state variables.
"""
pyapi = self.context.get_python_api(builder)
resume_index_ptr = self.get_resume_index_ptr(builder, genptr)
resume_index = builder.load(resume_index_ptr)
# If resume_index is 0, next() was never called
# If resume_index is -1, generator terminated cleanly
# (note function arguments are saved in state variables,
# so they don't need a separate cleanup step)
need_cleanup = builder.icmp_signed('>', resume_index,
Constant.int(resume_index.type, 0))
with cgutils.if_unlikely(builder, need_cleanup):
# Decref all live vars (some may be NULL)
gen_state_ptr = self.get_state_ptr(builder, genptr)
for state_index in range(len(self.gentype.state_types)):
state_slot = cgutils.gep_inbounds(builder, gen_state_ptr, 0,
state_index)
ty = self.gentype.state_types[state_index]
val = self.context.unpack_value(builder, ty, state_slot)
pyapi.decref(val)
builder.ret_void()
def return_from_generator(self, lower):
"""
Emit a StopIteration at generator end and mark the generator exhausted.
"""
indexval = Constant.int(self.resume_index_ptr.type.pointee, -1)
lower.builder.store(indexval, self.resume_index_ptr)
self.call_conv.return_stop_iteration(lower.builder)
def lower_finalize_func_body(self, builder, genptr):
"""
Lower the body of the generator's finalizer: decref all live
state variables.
"""
pyapi = self.context.get_python_api(builder)
resume_index_ptr = self.get_resume_index_ptr(builder, genptr)
resume_index = builder.load(resume_index_ptr)
# If resume_index is 0, next() was never called
# If resume_index is -1, generator terminated cleanly
# (note function arguments are saved in state variables,
# so they don't need a separate cleanup step)
need_cleanup = builder.icmp_signed(
'>', resume_index, Constant.int(resume_index.type, 0))
with cgutils.if_unlikely(builder, need_cleanup):
# Decref all live vars (some may be NULL)
gen_state_ptr = self.get_state_ptr(builder, genptr)
for state_index in range(len(self.gentype.state_types)):
state_slot = cgutils.gep_inbounds(builder, gen_state_ptr,
0, state_index)
ty = self.gentype.state_types[state_index]
val = self.context.unpack_value(builder, ty, state_slot)
pyapi.decref(val)
builder.ret_void()
def return_from_generator(self, lower):
"""
Emit a StopIteration at generator end and mark the generator exhausted.
"""
indexval = Constant.int(self.resume_index_ptr.type.pointee, -1)
lower.builder.store(indexval, self.resume_index_ptr)
self.call_conv.return_stop_iteration(lower.builder)
def timedelta_sign_impl(context, builder, sig, args):
val, = args
ret = alloc_timedelta_result(builder)
zero = Constant.int(TIMEDELTA64, 0)
with cgutils.ifelse(builder, builder.icmp(lc.ICMP_SGT, val, zero)
) as (gt_zero, le_zero):
with gt_zero:
builder.store(Constant.int(TIMEDELTA64, 1), ret)
with le_zero:
with cgutils.ifelse(builder, builder.icmp(lc.ICMP_EQ, val, zero)
) as (eq_zero, lt_zero):
with eq_zero:
builder.store(Constant.int(TIMEDELTA64, 0), ret)
with lt_zero:
builder.store(Constant.int(TIMEDELTA64, -1), ret)
return builder.load(ret)
def timedelta_sign_impl(context, builder, sig, args):
val, = args
ret = alloc_timedelta_result(builder)
zero = Constant.int(TIMEDELTA64, 0)
with builder.if_else(builder.icmp(lc.ICMP_SGT, val,
zero)) as (gt_zero, le_zero):
with gt_zero:
builder.store(Constant.int(TIMEDELTA64, 1), ret)
with le_zero:
with builder.if_else(builder.icmp(lc.ICMP_EQ, val,
zero)) as (eq_zero, lt_zero):
with eq_zero:
builder.store(Constant.int(TIMEDELTA64, 0), ret)
with lt_zero:
builder.store(Constant.int(TIMEDELTA64, -1), ret)
return builder.load(ret)
def timedelta_mod_timedelta(context, builder, sig, args):
# inspired by https://github.com/numpy/numpy/blob/fe8072a12d65e43bd2e0b0f9ad67ab0108cc54b3/numpy/core/src/umath/loops.c.src#L1424
# alg is basically as `a % b`:
# if a or b is NaT return NaT
# elseif b is 0 return NaT
# else pretend a and b are int and do pythonic int modulus
[va, vb] = args
[ta, tb] = sig.args
not_nan = are_not_nat(builder, [va, vb])
ll_ret_type = context.get_value_type(sig.return_type)
ret = alloc_timedelta_result(builder)
builder.store(NAT, ret)
zero = Constant.int(ll_ret_type, 0)
with cgutils.if_likely(builder, not_nan):
va, vb = normalize_timedeltas(context, builder, va, vb, ta, tb)
# is the denominator zero or NaT?
denom_ok = builder.not_(builder.icmp_signed('==', vb, zero))
with cgutils.if_likely(builder, denom_ok):
# is either arg negative?
vapos = builder.icmp_signed('>', va, zero)
vbpos = builder.icmp_signed('>', vb, zero)
rem = builder.srem(va, vb)
cond = builder.or_(builder.and_(vapos, vbpos),
builder.icmp_signed('==', rem, zero))
with builder.if_else(cond) as (then, otherwise):
with then:
builder.store(rem, ret)
with otherwise:
builder.store(builder.add(rem, vb), ret)
res = builder.load(ret)
return impl_ret_untracked(context, builder, sig.return_type, res)
def timedelta_mod_timedelta(context, builder, sig, args):
# inspired by https://github.com/numpy/numpy/blob/fe8072a12d65e43bd2e0b0f9ad67ab0108cc54b3/numpy/core/src/umath/loops.c.src#L1424
# alg is basically as `a % b`:
# if a or b is NaT return NaT
# elseif b is 0 return NaT
# else pretend a and b are int and do pythonic int modulus
[va, vb] = args
[ta, tb] = sig.args
not_nan = are_not_nat(builder, [va, vb])
ll_ret_type = context.get_value_type(sig.return_type)
ret = alloc_timedelta_result(builder)
builder.store(NAT, ret)
zero = Constant.int(ll_ret_type, 0)
with cgutils.if_likely(builder, not_nan):
va, vb = normalize_timedeltas(context, builder, va, vb, ta, tb)
# is the denominator zero or NaT?
denom_ok = builder.not_(builder.icmp_signed('==', vb, zero))
with cgutils.if_likely(builder, denom_ok):
# is either arg negative?
vapos = builder.icmp_signed('>', va, zero)
vbpos = builder.icmp_signed('>', vb, zero)
rem = builder.srem(va, vb)
cond = builder.or_(builder.and_(vapos, vbpos),
builder.icmp_signed('==', rem, zero))
with builder.if_else(cond) as (then, otherwise):
with then:
builder.store(rem, ret)
with otherwise:
builder.store(builder.add(rem, vb), ret)
res = builder.load(ret)
return impl_ret_untracked(context, builder, sig.return_type, res)
def gen_prologue(self, builder, pyapi):
# Get an environment object for the function
ll_intp = self.context.get_value_type(types.intp)
ll_pyobj = self.context.get_value_type(types.pyobject)
self.envptr = Constant.int(ll_intp, id(self.env)).inttoptr(ll_pyobj)
# Acquire the GIL
self.gil = pyapi.gil_ensure()
def gen_prologue(self, builder, pyapi):
# Get an environment object for the function
ll_intp = self.context.get_value_type(types.intp)
ll_pyobj = self.context.get_value_type(types.pyobject)
self.envptr = Constant.int(ll_intp, id(self.env)).inttoptr(ll_pyobj)
# Acquire the GIL
self.gil = pyapi.gil_ensure()
def gep(builder, ptr, *inds):
idx = []
for i in inds:
if isinstance(i, int):
ind = Constant.int(Type.int(64), i)
else:
ind = i
idx.append(ind)
return builder.gep(ptr, idx)
def inbound_gep(builder, ptr, *inds):
idx = []
for i in inds:
if isinstance(i, int):
ind = Constant.int(Type.int(32), i)
else:
ind = i
idx.append(ind)
return builder.gep(ptr, idx, inbounds=True)
def increment_index(builder, val):
"""
Increment an index *val*.
"""
one = Constant.int(val.type, 1)
# We pass the "nsw" flag in the hope that LLVM understands the index
# never changes sign. Unfortunately this doesn't always work
# (e.g. ndindex()).
return builder.add(val, one, flags=['nsw'])
def inbound_gep(builder, ptr, *inds):
idx = []
for i in inds:
if isinstance(i, int):
ind = Constant.int(Type.int(32), i)
else:
ind = i
idx.append(ind)
return builder.gep(ptr, idx, inbounds=True)
def increment_index(builder, val):
"""
Increment an index *val*.
"""
one = Constant.int(val.type, 1)
# We pass the "nsw" flag in the hope that LLVM understands the index
# never changes sign. Unfortunately this doesn't always work
# (e.g. ndindex()).
return builder.add(val, one, flags=['nsw'])
def gep(builder, ptr, *inds):
idx = []
for i in inds:
if isinstance(i, int):
ind = Constant.int(Type.int(64), i)
else:
ind = i
idx.append(ind)
return builder.gep(ptr, idx)
def for_range(builder, count, start=None, intp=None):
"""
Generate LLVM IR for a for-loop in [start, count).
*start* is equal to 0 by default.
Yields a Loop namedtuple with the following members:
- `index` is the loop index's value
- `do_break` is a no-argument callable to break out of the loop
"""
if intp is None:
intp = count.type
if start is None:
start = Constant.int(intp, 0)
stop = count
bbcond = builder.append_basic_block("for.cond")
bbbody = builder.append_basic_block("for.body")
bbend = builder.append_basic_block("for.end")
def do_break():
builder.branch(bbend)
bbstart = builder.basic_block
builder.branch(bbcond)
ONE = Constant.int(intp, 1)
with builder.goto_block(bbcond):
index = builder.phi(intp, name="loop.index")
pred = builder.icmp(lc.ICMP_SLT, index, stop)
builder.cbranch(pred, bbbody, bbend)
with builder.goto_block(bbbody):
yield Loop(index, do_break)
# Update bbbody as a new basic block may have been activated
bbbody = builder.basic_block
incr = builder.add(index, ONE)
terminate(builder, bbcond)
index.add_incoming(start, bbstart)
index.add_incoming(incr, bbbody)
builder.position_at_end(bbend)
def dict_new(self, presize=0):
if presize == 0:
fnty = Type.function(self.pyobj, ())
fn = self._get_function(fnty, name="PyDict_New")
return self.builder.call(fn, ())
else:
fnty = Type.function(self.pyobj, [self.py_ssize_t])
fn = self._get_function(fnty, name="_PyDict_NewPresized")
return self.builder.call(fn,
[Constant.int(self.py_ssize_t, presize)])
def dict_new(self, presize=0):
if presize == 0:
fnty = Type.function(self.pyobj, ())
fn = self._get_function(fnty, name="PyDict_New")
return self.builder.call(fn, ())
else:
fnty = Type.function(self.pyobj, [self.py_ssize_t])
fn = self._get_function(fnty, name="_PyDict_NewPresized")
return self.builder.call(fn,
[Constant.int(self.py_ssize_t, presize)])
def set_cuda_kernel(lfunc):
from llvmlite.llvmpy.core import MetaData, MetaDataString, Constant, Type
m = lfunc.module
ops = lfunc, MetaDataString.get(m, "kernel"), Constant.int(Type.int(), 1)
md = MetaData.get(m, ops)
nmd = m.get_or_insert_named_metadata('nvvm.annotations')
nmd.add(md)
def lower_yield_suspend(self):
# Save live vars in state
for state_index, name in zip(self.live_var_indices, self.live_vars):
state_slot = cgutils.gep(self.builder, self.gen_state_ptr, 0, state_index)
ty = self.gentype.state_types[state_index]
val = self.lower.loadvar(name)
self.context.pack_value(self.builder, ty, val, state_slot)
# Save resume index
indexval = Constant.int(self.resume_index_ptr.type.pointee, self.inst.index)
self.builder.store(indexval, self.resume_index_ptr)
def gep(builder, ptr, *inds):
idx = []
for i in inds:
if isinstance(i, int):
# NOTE: llvm only accepts int32 inside structs, not int64
ind = Constant.int(Type.int(32), i)
else:
ind = i
idx.append(ind)
return builder.gep(ptr, idx)
def printf(builder, format_string, *values):
str_const = Constant.stringz(format_string)
global_str_const = get_module(builder).add_global_variable(str_const.type,
'')
global_str_const.initializer = str_const
idx = [Constant.int(Type.int(32), 0), Constant.int(Type.int(32), 0)]
str_addr = global_str_const.gep(idx)
args = []
for v in values:
if isinstance(v, int):
args.append(Constant.int(Type.int(), v))
elif isinstance(v, float):
args.append(Constant.real(Type.double(), v))
else:
args.append(v)
functype = Type.function(Type.int(32), [Type.pointer(Type.int(8))], True)
fn = get_module(builder).get_or_insert_function(functype, 'printf')
builder.call(fn, [str_addr] + args)
def printf(builder, format_string, *values):
str_const = Constant.stringz(format_string)
global_str_const = get_module(builder).add_global_variable(
str_const.type, '')
global_str_const.initializer = str_const
idx = [Constant.int(Type.int(32), 0), Constant.int(Type.int(32), 0)]
str_addr = global_str_const.gep(idx)
args = []
for v in values:
if isinstance(v, int):
args.append(Constant.int(Type.int(), v))
elif isinstance(v, float):
args.append(Constant.real(Type.double(), v))
else:
args.append(v)
functype = Type.function(Type.int(32), [Type.pointer(Type.int(8))], True)
fn = get_module(builder).get_or_insert_function(functype, 'printf')
builder.call(fn, [str_addr] + args)
def gep(builder, ptr, *inds, **kws):
name = kws.pop('name', '')
idx = []
for i in inds:
if isinstance(i, int):
# NOTE: llvm only accepts int32 inside structs, not int64
ind = Constant.int(Type.int(32), i)
else:
ind = i
idx.append(ind)
return builder.gep(ptr, idx, name=name)
def memset(builder, ptr, size, value):
"""
Fill *size* bytes starting from *ptr* with *value*.
"""
sizety = size.type
memset = "llvm.memset.p0i8.i%d" % (sizety.width)
module = get_module(builder)
i32 = lc.Type.int(32)
i8 = lc.Type.int(8)
i1 = lc.Type.int(1)
# void @llvm.memset.p0i8.iXY(i8* <dest>, i8 <val>,
# iXY <len>, i32 <align>, i1 <isvolatile>)
ptr = builder.bitcast(ptr, lc.Type.pointer(i8))
fnty = lc.Type.function(lc.Type.void(),
[ptr.type, i8, sizety, i32, i1])
fn = module.get_or_insert_function(fnty, name=memset)
if isinstance(value, int):
value = Constant.int(i8, value)
builder.call(fn, [ptr, value, size,
Constant.int(i32, 0), Constant.int(i1, 0)])
def timedelta_sign_impl(context, builder, sig, args):
"""
np.sign(timedelta64)
"""
val, = args
ret = alloc_timedelta_result(builder)
zero = Constant.int(TIMEDELTA64, 0)
with builder.if_else(builder.icmp(lc.ICMP_SGT, val, zero)
) as (gt_zero, le_zero):
with gt_zero:
builder.store(Constant.int(TIMEDELTA64, 1), ret)
with le_zero:
with builder.if_else(builder.icmp(lc.ICMP_EQ, val, zero)
) as (eq_zero, lt_zero):
with eq_zero:
builder.store(Constant.int(TIMEDELTA64, 0), ret)
with lt_zero:
builder.store(Constant.int(TIMEDELTA64, -1), ret)
res = builder.load(ret)
return impl_ret_untracked(context, builder, sig.return_type, res)
def lower_yield_suspend(self):
# Save live vars in state
for state_index, name in zip(self.live_var_indices, self.live_vars):
state_slot = cgutils.gep(self.builder, self.gen_state_ptr,
0, state_index)
ty = self.gentype.state_types[state_index]
val = self.lower.loadvar(name)
self.context.pack_value(self.builder, ty, val, state_slot)
# Save resume index
indexval = Constant.int(self.resume_index_ptr.type.pointee,
self.inst.index)
self.builder.store(indexval, self.resume_index_ptr)
def timedelta_sign_impl(context, builder, sig, args):
"""
np.sign(timedelta64)
"""
(val,) = args
ret = alloc_timedelta_result(builder)
zero = Constant.int(TIMEDELTA64, 0)
with builder.if_else(builder.icmp(lc.ICMP_SGT, val, zero)) as (gt_zero, le_zero):
with gt_zero:
builder.store(Constant.int(TIMEDELTA64, 1), ret)
with le_zero:
with builder.if_else(builder.icmp(lc.ICMP_EQ, val, zero)) as (
eq_zero,
lt_zero,
):
with eq_zero:
builder.store(Constant.int(TIMEDELTA64, 0), ret)
with lt_zero:
builder.store(Constant.int(TIMEDELTA64, -1), ret)
res = builder.load(ret)
return impl_ret_untracked(context, builder, sig.return_type, res)
def gen_prologue(self, builder):
# Create an environment object for the function
moduledict = self.fndesc.lookup_module().__dict__
self.env = _dynfunc.Environment(globals=moduledict)
ll_intp = self.context.get_value_type(types.intp)
ll_pyobj = self.context.get_value_type(types.pyobject)
self.envptr = Constant.int(ll_intp, id(self.env)).inttoptr(ll_pyobj)
# Acquire the GIL
self.pyapi = self.context.get_python_api(builder)
self.gil = self.pyapi.gil_ensure()
def int_sign_impl(context, builder, sig, args):
"""
np.sign(int)
"""
[x] = args
POS = Constant.int(x.type, 1)
NEG = Constant.int(x.type, -1)
ZERO = Constant.int(x.type, 0)
cmp_zero = builder.icmp(lc.ICMP_EQ, x, ZERO)
cmp_pos = builder.icmp(lc.ICMP_SGT, x, ZERO)
presult = cgutils.alloca_once(builder, x.type)
bb_zero = builder.append_basic_block(".zero")
bb_postest = builder.append_basic_block(".postest")
bb_pos = builder.append_basic_block(".pos")
bb_neg = builder.append_basic_block(".neg")
bb_exit = builder.append_basic_block(".exit")
builder.cbranch(cmp_zero, bb_zero, bb_postest)
with builder.goto_block(bb_zero):
builder.store(ZERO, presult)
builder.branch(bb_exit)
with builder.goto_block(bb_postest):
builder.cbranch(cmp_pos, bb_pos, bb_neg)
with builder.goto_block(bb_pos):
builder.store(POS, presult)
builder.branch(bb_exit)
with builder.goto_block(bb_neg):
builder.store(NEG, presult)
builder.branch(bb_exit)
builder.position_at_end(bb_exit)
res = builder.load(presult)
return impl_ret_untracked(context, builder, sig.return_type, res)
def for_range(builder, count, intp=None):
"""
Generate LLVM IR for a for-loop in [0, count). Yields a
Loop namedtuple with the following members:
- `index` is the loop index's value
- `do_break` is a no-argument callable to break out of the loop
"""
if intp is None:
intp = count.type
start = Constant.int(intp, 0)
stop = count
bbcond = builder.append_basic_block("for.cond")
bbbody = builder.append_basic_block("for.body")
bbend = builder.append_basic_block("for.end")
def do_break():
builder.branch(bbend)
bbstart = builder.basic_block
builder.branch(bbcond)
ONE = Constant.int(intp, 1)
with builder.goto_block(bbcond):
index = builder.phi(intp, name="loop.index")
pred = builder.icmp(lc.ICMP_SLT, index, stop)
builder.cbranch(pred, bbbody, bbend)
with builder.goto_block(bbbody):
yield Loop(index, do_break)
# Update bbbody as a new basic block may have been activated
bbbody = builder.basic_block
incr = builder.add(index, ONE)
terminate(builder, bbcond)
index.add_incoming(start, bbstart)
index.add_incoming(incr, bbbody)
builder.position_at_end(bbend)
def int_sign_impl(context, builder, sig, args):
"""
np.sign(int)
"""
[x] = args
POS = Constant.int(x.type, 1)
NEG = Constant.int(x.type, -1)
ZERO = Constant.int(x.type, 0)
cmp_zero = builder.icmp(lc.ICMP_EQ, x, ZERO)
cmp_pos = builder.icmp(lc.ICMP_SGT, x, ZERO)
presult = cgutils.alloca_once(builder, x.type)
bb_zero = builder.append_basic_block(".zero")
bb_postest = builder.append_basic_block(".postest")
bb_pos = builder.append_basic_block(".pos")
bb_neg = builder.append_basic_block(".neg")
bb_exit = builder.append_basic_block(".exit")
builder.cbranch(cmp_zero, bb_zero, bb_postest)
with builder.goto_block(bb_zero):
builder.store(ZERO, presult)
builder.branch(bb_exit)
with builder.goto_block(bb_postest):
builder.cbranch(cmp_pos, bb_pos, bb_neg)
with builder.goto_block(bb_pos):
builder.store(POS, presult)
builder.branch(bb_exit)
with builder.goto_block(bb_neg):
builder.store(NEG, presult)
builder.branch(bb_exit)
builder.position_at_end(bb_exit)
res = builder.load(presult)
return impl_ret_untracked(context, builder, sig.return_type, res)
def pointer_add(builder, ptr, offset, return_type=None):
"""
Add an integral *offset* to pointer *ptr*, and return a pointer
of *return_type* (or, if omitted, the same type as *ptr*).
Note the computation is done in bytes, and ignores the width of
the pointed item type.
"""
intptr = builder.ptrtoint(ptr, intp_t)
if isinstance(offset, int):
offset = Constant.int(intp_t, offset)
intptr = builder.add(intptr, offset)
return builder.inttoptr(intptr, return_type or ptr.type)
