def _gauss_impl(context, builder, sig, args, state):
# The type for all computations (either float or double)
ty = sig.return_type
llty = context.get_data_type(ty)
state_ptr = get_state_ptr(context, builder, state)
_random = {"py": random.random,
"np": np.random.random}[state]
ret = cgutils.alloca_once(builder, llty, name="result")
gauss_ptr = get_gauss_ptr(builder, state_ptr)
has_gauss_ptr = get_has_gauss_ptr(builder, state_ptr)
has_gauss = cgutils.is_true(builder, builder.load(has_gauss_ptr))
with cgutils.ifelse(builder, has_gauss) as (then, otherwise):
with then:
# if has_gauss: return it
builder.store(builder.load(gauss_ptr), ret)
builder.store(const_int(0), has_gauss_ptr)
with otherwise:
# if not has_gauss: compute a pair of numbers using the Box-Muller
# transform; keep one and return the other
pair = context.compile_internal(builder,
_gauss_pair_impl(_random),
signature(types.UniTuple(ty, 2)),
())
first, second = cgutils.unpack_tuple(builder, pair, 2)
builder.store(first, gauss_ptr)
builder.store(second, ret)
builder.store(const_int(1), has_gauss_ptr)
mu, sigma = args
return builder.fadd(mu,
builder.fmul(sigma, builder.load(ret)))
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 _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 cgutils.ifelse(self.builder, 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 to_native_optional(self, obj, typ):
"""
Convert object *obj* to a native optional structure.
"""
noneval = self.context.make_optional_none(self.builder, typ.type)
is_not_none = self.builder.icmp(lc.ICMP_NE, obj, self.borrow_none())
retptr = cgutils.alloca_once(self.builder, noneval.type)
errptr = cgutils.alloca_once_value(self.builder, cgutils.false_bit)
with cgutils.ifelse(self.builder, is_not_none) as (then, orelse):
with then:
native = self.to_native_value(obj, typ.type)
just = self.context.make_optional_value(self.builder,
typ.type, native.value)
self.builder.store(just, retptr)
self.builder.store(native.is_error, errptr)
with orelse:
self.builder.store(ir.Constant(noneval.type, ir.Undefined),
retptr)
self.builder.store(noneval, retptr)
if native.cleanup is not None:
def cleanup():
with cgutils.ifthen(self.builder, is_not_none):
native.cleanup()
else:
cleanup = None
ret = self.builder.load(retptr)
return NativeValue(ret, is_error=self.builder.load(errptr),
cleanup=cleanup)
def _gauss_impl(context, builder, sig, args, state):
# The type for all computations (either float or double)
ty = sig.return_type
llty = context.get_data_type(ty)
state_ptr = get_state_ptr(context, builder, state)
_random = {"py": random.random,
"np": np.random.random}[state]
ret = cgutils.alloca_once(builder, llty, name="result")
gauss_ptr = get_gauss_ptr(builder, state_ptr)
has_gauss_ptr = get_has_gauss_ptr(builder, state_ptr)
has_gauss = cgutils.is_true(builder, builder.load(has_gauss_ptr))
with cgutils.ifelse(builder, has_gauss) as (then, otherwise):
with then:
# if has_gauss: return it
builder.store(builder.load(gauss_ptr), ret)
builder.store(const_int(0), has_gauss_ptr)
with otherwise:
# if not has_gauss: compute a pair of numbers using the Box-Muller
# transform; keep one and return the other
pair = context.compile_internal(builder,
_gauss_pair_impl(_random),
signature(types.UniTuple(ty, 2)),
())
first, second = cgutils.unpack_tuple(builder, pair, 2)
builder.store(first, gauss_ptr)
builder.store(second, ret)
builder.store(const_int(1), has_gauss_ptr)
mu, sigma = args
return builder.fadd(mu,
builder.fmul(sigma, builder.load(ret)))
def getiter_range_generic(context, builder, iterobj, start, stop, step):
diff = builder.sub(stop, start)
intty = start.type
zero = Constant.int(intty, 0)
one = Constant.int(intty, 1)
pos_diff = builder.icmp(lc.ICMP_SGT, diff, zero)
pos_step = builder.icmp(lc.ICMP_SGT, step, zero)
sign_differs = builder.xor(pos_diff, pos_step)
zero_step = builder.icmp(lc.ICMP_EQ, step, zero)
with cgutils.if_unlikely(builder, zero_step):
# step shouldn't be zero
context.return_errcode(builder, 1)
with cgutils.ifelse(builder, sign_differs) as (then, orelse):
with then:
builder.store(zero, iterobj.count)
with orelse:
rem = builder.srem(diff, step)
uneven = builder.icmp(lc.ICMP_SGT, rem, zero)
newcount = builder.add(builder.sdiv(diff, step),
builder.select(uneven, one, zero))
builder.store(newcount, iterobj.count)
return iterobj._getvalue()
def get_next_int(context, builder, state_ptr, nbits):
"""
Get the next integer with width *nbits*.
"""
c32 = ir.Constant(nbits.type, 32)
def get_shifted_int(nbits):
shift = builder.sub(c32, nbits)
y = get_next_int32(context, builder, state_ptr)
return builder.lshr(y, builder.zext(shift, y.type))
ret = cgutils.alloca_once_value(builder, ir.Constant(int64_t, 0))
is_32b = builder.icmp_unsigned('<=', nbits, c32)
with cgutils.ifelse(builder, is_32b) as (ifsmall, iflarge):
with ifsmall:
low = get_shifted_int(nbits)
builder.store(builder.zext(low, int64_t), ret)
with iflarge:
# XXX This assumes nbits <= 64
low = get_next_int32(context, builder, state_ptr)
high = get_shifted_int(builder.sub(nbits, c32))
total = builder.add(
builder.zext(low, int64_t),
builder.shl(builder.zext(high, int64_t), ir.Constant(int64_t, 32)))
builder.store(total, ret)
return builder.load(ret)
def complex128_power_impl(context, builder, sig, args):
[ca, cb] = args
a = Complex128(context, builder, value=ca)
b = Complex128(context, builder, value=cb)
c = Complex128(context, builder)
module = cgutils.get_module(builder)
pa = a._getpointer()
pb = b._getpointer()
pc = c._getpointer()
# Optimize for square because cpow looses a lot of precsiion
TWO = context.get_constant(types.float64, 2)
ZERO = context.get_constant(types.float64, 0)
b_real_is_two = builder.fcmp(lc.FCMP_OEQ, b.real, TWO)
b_imag_is_zero = builder.fcmp(lc.FCMP_OEQ, b.imag, ZERO)
b_is_two = builder.and_(b_real_is_two, b_imag_is_zero)
with cgutils.ifelse(builder, b_is_two) as (then, otherwise):
with then:
# Lower as multiplication
res = complex_mul_impl(context, builder, sig, (ca, ca))
cres = Complex128(context, builder, value=res)
c.real = cres.real
c.imag = cres.imag
with otherwise:
# Lower with call to external function
fnty = Type.function(Type.void(), [pa.type] * 3)
cpow = module.get_or_insert_function(fnty, name="numba.math.cpow")
builder.call(cpow, (pa, pb, pc))
return builder.load(pc)
def get_next_int(context, builder, state_ptr, nbits):
"""
Get the next integer with width *nbits*.
"""
c32 = ir.Constant(nbits.type, 32)
def get_shifted_int(nbits):
shift = builder.sub(c32, nbits)
y = get_next_int32(context, builder, state_ptr)
return builder.lshr(y, builder.zext(shift, y.type))
ret = cgutils.alloca_once_value(builder, ir.Constant(int64_t, 0))
is_32b = builder.icmp_unsigned('<=', nbits, c32)
with cgutils.ifelse(builder, is_32b) as (ifsmall, iflarge):
with ifsmall:
low = get_shifted_int(nbits)
builder.store(builder.zext(low, int64_t), ret)
with iflarge:
# XXX This assumes nbits <= 64
low = get_next_int32(context, builder, state_ptr)
high = get_shifted_int(builder.sub(nbits, c32))
total = builder.add(
builder.zext(low, int64_t),
builder.shl(builder.zext(high, int64_t), ir.Constant(int64_t, 32)))
builder.store(total, ret)
return builder.load(ret)
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 cgutils.ifelse(self.builder, 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 timedelta_abs_impl(context, builder, sig, args):
val, = args
ret = alloc_timedelta_result(builder)
with cgutils.ifelse(builder,
cgutils.is_scalar_neg(builder, val)) as (then, otherwise):
with then:
builder.store(builder.neg(val), ret)
with otherwise:
builder.store(val, ret)
return builder.load(ret)
def year_to_days(builder, year_val):
"""
Given a year *year_val* (offset to 1970), return the number of days
since the 1970 epoch.
"""
# The algorithm below is copied from Numpy's get_datetimestruct_days()
# (src/multiarray/datetime.c)
ret = cgutils.alloca_once(builder, TIMEDELTA64)
# First approximation
days = scale_by_constant(builder, year_val, 365)
# Adjust for leap years
with cgutils.ifelse(builder, cgutils.is_neg_int(builder, year_val)) \
as (if_neg, if_pos):
with if_pos:
# At or after 1970:
# 1968 is the closest leap year before 1970.
# Exclude the current year, so add 1.
from_1968 = add_constant(builder, year_val, 1)
# Add one day for each 4 years
p_days = builder.add(days,
unscale_by_constant(builder, from_1968, 4))
# 1900 is the closest previous year divisible by 100
from_1900 = add_constant(builder, from_1968, 68)
# Subtract one day for each 100 years
p_days = builder.sub(p_days,
unscale_by_constant(builder, from_1900, 100))
# 1600 is the closest previous year divisible by 400
from_1600 = add_constant(builder, from_1900, 300)
# Add one day for each 400 years
p_days = builder.add(p_days,
unscale_by_constant(builder, from_1600, 400))
builder.store(p_days, ret)
with if_neg:
# Before 1970:
# NOTE `year_val` is negative, and so will be `from_1972` and `from_2000`.
# 1972 is the closest later year after 1970.
# Include the current year, so subtract 2.
from_1972 = add_constant(builder, year_val, -2)
# Subtract one day for each 4 years (`from_1972` is negative)
n_days = builder.add(days,
unscale_by_constant(builder, from_1972, 4))
# 2000 is the closest later year divisible by 100
from_2000 = add_constant(builder, from_1972, -28)
# Add one day for each 100 years
n_days = builder.sub(n_days,
unscale_by_constant(builder, from_2000, 100))
# 2000 is also the closest later year divisible by 400
# Subtract one day for each 400 years
n_days = builder.add(n_days,
unscale_by_constant(builder, from_2000, 400))
builder.store(n_days, ret)
return builder.load(ret)
def store(retval):
is_error = cgutils.is_null(builder, retval)
with cgutils.ifelse(builder, is_error) as (if_error, if_ok):
with if_error:
msg = context.insert_const_string(pyapi.module,
"object mode ufunc")
msgobj = pyapi.string_from_string(msg)
pyapi.err_write_unraisable(msgobj)
pyapi.decref(msgobj)
with if_ok:
# Unbox
retval = pyapi.to_native_value(retval, signature.return_type)
# Store
out.store_direct(retval, builder.load(store_offset))
def impl(context, builder, sig, args):
[va, vb] = args
[ta, tb] = sig.args
ret = alloc_boolean_result(builder)
with cgutils.ifelse(builder, are_not_nat(builder, [va, vb])) as (then, otherwise):
with then:
norm_a, norm_b = normalize_timedeltas(context, builder, va, vb, ta, tb)
builder.store(builder.icmp(ll_op, norm_a, norm_b), ret)
with otherwise:
# No scaling when comparing NaT with something else
# (i.e. NaT is <= everything else, since it's the smallest
# int64 value)
builder.store(builder.icmp(ll_op, va, vb), ret)
return builder.load(ret)
def store(retval):
is_error = cgutils.is_null(builder, retval)
with cgutils.ifelse(builder, is_error) as (if_error, if_ok):
with if_error:
msg = context.insert_const_string(pyapi.module,
"object mode ufunc")
msgobj = pyapi.string_from_string(msg)
pyapi.err_write_unraisable(msgobj)
pyapi.decref(msgobj)
with if_ok:
# Unbox
retval = pyapi.to_native_value(retval, signature.return_type)
# Store
out.store_direct(retval, builder.load(store_offset))
def impl(context, builder, sig, args):
[va, vb] = args
[ta, tb] = sig.args
ret = alloc_boolean_result(builder)
with cgutils.ifelse(builder, are_not_nat(builder, [va, vb])) as (then, otherwise):
with then:
try:
norm_a, norm_b = normalize_timedeltas(context, builder, va, vb, ta, tb)
except RuntimeError:
# Cannot normalize units => the values are unequal (except if NaT)
builder.store(default_value, ret)
else:
builder.store(builder.icmp(ll_op, norm_a, norm_b), ret)
with otherwise:
# No scaling when comparing NaTs
builder.store(builder.icmp(ll_op, va, vb), ret)
return builder.load(ret)
def from_range_state(cls, context, builder, state):
"""
Create a RangeIter initialized from the given RangeState *state*.
"""
self = cls(context, builder)
start = state.start
stop = state.stop
step = state.step
startptr = cgutils.alloca_once(builder, start.type)
builder.store(start, startptr)
countptr = cgutils.alloca_once(builder, start.type)
self.iter = startptr
self.stop = stop
self.step = step
self.count = countptr
diff = builder.sub(stop, start)
zero = context.get_constant(int_type, 0)
one = context.get_constant(int_type, 1)
pos_diff = builder.icmp(lc.ICMP_SGT, diff, zero)
pos_step = builder.icmp(lc.ICMP_SGT, step, zero)
sign_differs = builder.xor(pos_diff, pos_step)
zero_step = builder.icmp(lc.ICMP_EQ, step, zero)
with cgutils.if_unlikely(builder, zero_step):
# step shouldn't be zero
context.call_conv.return_user_exc(builder, ValueError,
("range() arg 3 must not be zero",))
with cgutils.ifelse(builder, sign_differs) as (then, orelse):
with then:
builder.store(zero, self.count)
with orelse:
rem = builder.srem(diff, step)
rem = builder.select(pos_diff, rem, builder.neg(rem))
uneven = builder.icmp(lc.ICMP_SGT, rem, zero)
newcount = builder.add(builder.sdiv(diff, step),
builder.select(uneven, one, zero))
builder.store(newcount, self.count)
return self
def from_range_state(cls, context, builder, state):
"""
Create a RangeIter initialized from the given RangeState *state*.
"""
self = cls(context, builder)
start = state.start
stop = state.stop
step = state.step
startptr = cgutils.alloca_once(builder, start.type)
builder.store(start, startptr)
countptr = cgutils.alloca_once(builder, start.type)
self.iter = startptr
self.stop = stop
self.step = step
self.count = countptr
diff = builder.sub(stop, start)
zero = context.get_constant(int_type, 0)
one = context.get_constant(int_type, 1)
pos_diff = builder.icmp(lc.ICMP_SGT, diff, zero)
pos_step = builder.icmp(lc.ICMP_SGT, step, zero)
sign_differs = builder.xor(pos_diff, pos_step)
zero_step = builder.icmp(lc.ICMP_EQ, step, zero)
with cgutils.if_unlikely(builder, zero_step):
# step shouldn't be zero
context.call_conv.return_user_exc(
builder, ValueError, ("range() arg 3 must not be zero", ))
with cgutils.ifelse(builder, sign_differs) as (then, orelse):
with then:
builder.store(zero, self.count)
with orelse:
rem = builder.srem(diff, step)
rem = builder.select(pos_diff, rem, builder.neg(rem))
uneven = builder.icmp(lc.ICMP_SGT, rem, zero)
newcount = builder.add(builder.sdiv(diff, step),
builder.select(uneven, one, zero))
builder.store(newcount, self.count)
return self
def reduce_datetime_for_unit(builder, dt_val, src_unit, dest_unit):
dest_unit_code = npdatetime.DATETIME_UNITS[dest_unit]
src_unit_code = npdatetime.DATETIME_UNITS[src_unit]
if dest_unit_code < 2 or src_unit_code >= 2:
return dt_val, src_unit
# Need to compute the day ordinal for *dt_val*
if src_unit_code == 0:
# Years to days
year_val = dt_val
days_val = year_to_days(builder, year_val)
else:
# Months to days
leap_array = cgutils.global_constant(builder, "leap_year_months_acc",
leap_year_months_acc)
normal_array = cgutils.global_constant(builder, "normal_year_months_acc",
normal_year_months_acc)
days = cgutils.alloca_once(builder, TIMEDELTA64)
# First compute year number and month number
year, month = cgutils.divmod_by_constant(builder, dt_val, 12)
# Then deduce the number of days
with cgutils.ifelse(builder,
is_leap_year(builder, year)) as (then, otherwise):
with then:
addend = builder.load(cgutils.gep(builder, leap_array,
0, month))
builder.store(addend, days)
with otherwise:
addend = builder.load(cgutils.gep(builder, normal_array,
0, month))
builder.store(addend, days)
days_val = year_to_days(builder, year)
days_val = builder.add(days_val, builder.load(days))
if dest_unit_code == 2:
# Need to scale back to weeks
weeks, _ = cgutils.divmod_by_constant(builder, days_val, 7)
return weeks, 'W'
else:
return days_val, 'D'
def impl(context, builder, sig, args):
va, vb = args
ta, tb = sig.args
unit_a = ta.unit
unit_b = tb.unit
ret_unit = npdatetime.get_best_unit(unit_a, unit_b)
ret = alloc_boolean_result(builder)
with cgutils.ifelse(builder,
are_not_nat(builder, [va, vb])) as (then, otherwise):
with then:
norm_a = convert_datetime_for_arith(builder, va, unit_a, ret_unit)
norm_b = convert_datetime_for_arith(builder, vb, unit_b, ret_unit)
ret_val = builder.icmp(ll_op, norm_a, norm_b)
builder.store(ret_val, ret)
with otherwise:
# No scaling when comparing NaTs
ret_val = builder.icmp(ll_op, va, vb)
builder.store(ret_val, ret)
return builder.load(ret)
def build_ufunc_wrapper(library, context, func, signature, objmode, env):
"""
Wrap the scalar function with a loop that iterates over the arguments
"""
byte_t = Type.int(8)
byte_ptr_t = Type.pointer(byte_t)
byte_ptr_ptr_t = Type.pointer(byte_ptr_t)
intp_t = context.get_value_type(types.intp)
intp_ptr_t = Type.pointer(intp_t)
fnty = Type.function(Type.void(),
[byte_ptr_ptr_t, intp_ptr_t, intp_ptr_t, byte_ptr_t])
wrapper_module = library.create_ir_module('')
if objmode:
func_type = context.call_conv.get_function_type(
types.pyobject, [types.pyobject] * len(signature.args))
else:
func_type = context.call_conv.get_function_type(
signature.return_type, signature.args)
oldfunc = func
func = wrapper_module.add_function(func_type, name=func.name)
func.attributes.add("alwaysinline")
wrapper = wrapper_module.add_function(fnty, "__ufunc__." + func.name)
arg_args, arg_dims, arg_steps, arg_data = wrapper.args
arg_args.name = "args"
arg_dims.name = "dims"
arg_steps.name = "steps"
arg_data.name = "data"
builder = Builder.new(wrapper.append_basic_block("entry"))
loopcount = builder.load(arg_dims, name="loopcount")
actual_args = context.call_conv.get_arguments(func)
# Prepare inputs
arrays = []
for i, typ in enumerate(signature.args):
arrays.append(
UArrayArg(context, builder, arg_args, arg_steps, i,
context.get_argument_type(typ)))
# Prepare output
valty = context.get_data_type(signature.return_type)
out = UArrayArg(context, builder, arg_args, arg_steps, len(actual_args),
valty)
# Setup indices
offsets = []
zero = context.get_constant(types.intp, 0)
for _ in arrays:
p = cgutils.alloca_once(builder, intp_t)
offsets.append(p)
builder.store(zero, p)
store_offset = cgutils.alloca_once(builder, intp_t)
builder.store(zero, store_offset)
unit_strided = cgutils.true_bit
for ary in arrays:
unit_strided = builder.and_(unit_strided, ary.is_unit_strided)
if objmode:
# General loop
pyapi = context.get_python_api(builder)
gil = pyapi.gil_ensure()
with cgutils.for_range(builder, loopcount, intp=intp_t):
slowloop = build_obj_loop_body(context, func, builder, arrays, out,
offsets, store_offset, signature,
pyapi, env)
pyapi.gil_release(gil)
builder.ret_void()
else:
with cgutils.ifelse(builder,
unit_strided) as (is_unit_strided, is_strided):
with is_unit_strided:
with cgutils.for_range(builder, loopcount, intp=intp_t) as ind:
fastloop = build_fast_loop_body(context, func, builder,
arrays, out, offsets,
store_offset, signature,
ind)
builder.ret_void()
with is_strided:
# General loop
with cgutils.for_range(builder, loopcount, intp=intp_t):
slowloop = build_slow_loop_body(context, func, builder,
arrays, out, offsets,
store_offset, signature)
builder.ret_void()
builder.ret_void()
del builder
# Run optimizer
library.add_ir_module(wrapper_module)
wrapper = library.get_function(wrapper.name)
oldfunc.linkage = LINKAGE_INTERNAL
return wrapper
def to_native_value(self, obj, typ):
if isinstance(typ, types.Object) or typ == types.pyobject:
return obj
elif typ == types.boolean:
istrue = self.object_istrue(obj)
zero = Constant.null(istrue.type)
return self.builder.icmp(lc.ICMP_NE, istrue, zero)
elif typ in types.unsigned_domain:
longobj = self.number_long(obj)
ullval = self.long_as_ulonglong(longobj)
self.decref(longobj)
return self.builder.trunc(ullval,
self.context.get_argument_type(typ))
elif typ in types.signed_domain:
longobj = self.number_long(obj)
llval = self.long_as_longlong(longobj)
self.decref(longobj)
return self.builder.trunc(llval,
self.context.get_argument_type(typ))
elif typ == types.float32:
fobj = self.number_float(obj)
fval = self.float_as_double(fobj)
self.decref(fobj)
return self.builder.fptrunc(fval,
self.context.get_argument_type(typ))
elif typ == types.float64:
fobj = self.number_float(obj)
fval = self.float_as_double(fobj)
self.decref(fobj)
return fval
elif typ in (types.complex128, types.complex64):
cplxcls = self.context.make_complex(types.complex128)
cplx = cplxcls(self.context, self.builder)
pcplx = cplx._getpointer()
ok = self.complex_adaptor(obj, pcplx)
failed = cgutils.is_false(self.builder, ok)
with cgutils.if_unlikely(self.builder, failed):
self.builder.ret(self.get_null_object())
if typ == types.complex64:
c64cls = self.context.make_complex(typ)
c64 = c64cls(self.context, self.builder)
freal = self.context.cast(self.builder, cplx.real,
types.float64, types.float32)
fimag = self.context.cast(self.builder, cplx.imag,
types.float64, types.float32)
c64.real = freal
c64.imag = fimag
return c64._getvalue()
else:
return cplx._getvalue()
elif isinstance(typ, types.NPDatetime):
val = self.extract_np_datetime(obj)
return val
elif isinstance(typ, types.NPTimedelta):
val = self.extract_np_timedelta(obj)
return val
elif isinstance(typ, types.Array):
return self.to_native_array(typ, obj)
elif isinstance(typ, types.Optional):
isnone = self.builder.icmp(lc.ICMP_EQ, obj, self.borrow_none())
with cgutils.ifelse(self.builder, isnone) as (then, orelse):
with then:
noneval = self.context.make_optional_none(
self.builder, typ.type)
ret = cgutils.alloca_once(self.builder, noneval.type)
self.builder.store(noneval, ret)
with orelse:
val = self.to_native_value(obj, typ.type)
just = self.context.make_optional_value(
self.builder, typ.type, val)
self.builder.store(just, ret)
return ret
raise NotImplementedError(typ)
def build_ufunc_wrapper(library, context, func, signature, objmode, env):
"""
Wrap the scalar function with a loop that iterates over the arguments
"""
byte_t = Type.int(8)
byte_ptr_t = Type.pointer(byte_t)
byte_ptr_ptr_t = Type.pointer(byte_ptr_t)
intp_t = context.get_value_type(types.intp)
intp_ptr_t = Type.pointer(intp_t)
fnty = Type.function(Type.void(), [byte_ptr_ptr_t, intp_ptr_t,
intp_ptr_t, byte_ptr_t])
wrapper_module = library.create_ir_module('')
if objmode:
func_type = context.call_conv.get_function_type(
types.pyobject, [types.pyobject] * len(signature.args))
else:
func_type = context.call_conv.get_function_type(
signature.return_type, signature.args)
oldfunc = func
func = wrapper_module.add_function(func_type,
name=func.name)
func.attributes.add("alwaysinline")
wrapper = wrapper_module.add_function(fnty, "__ufunc__." + func.name)
arg_args, arg_dims, arg_steps, arg_data = wrapper.args
arg_args.name = "args"
arg_dims.name = "dims"
arg_steps.name = "steps"
arg_data.name = "data"
builder = Builder.new(wrapper.append_basic_block("entry"))
loopcount = builder.load(arg_dims, name="loopcount")
# Prepare inputs
arrays = []
for i, typ in enumerate(signature.args):
arrays.append(UArrayArg(context, builder, arg_args, arg_steps, i, typ))
# Prepare output
out = UArrayArg(context, builder, arg_args, arg_steps, len(arrays),
signature.return_type)
# Setup indices
offsets = []
zero = context.get_constant(types.intp, 0)
for _ in arrays:
p = cgutils.alloca_once(builder, intp_t)
offsets.append(p)
builder.store(zero, p)
store_offset = cgutils.alloca_once(builder, intp_t)
builder.store(zero, store_offset)
unit_strided = cgutils.true_bit
for ary in arrays:
unit_strided = builder.and_(unit_strided, ary.is_unit_strided)
if objmode:
# General loop
pyapi = context.get_python_api(builder)
gil = pyapi.gil_ensure()
with cgutils.for_range(builder, loopcount, intp=intp_t):
slowloop = build_obj_loop_body(context, func, builder,
arrays, out, offsets,
store_offset, signature,
pyapi, env)
pyapi.gil_release(gil)
builder.ret_void()
else:
with cgutils.ifelse(builder, unit_strided) as (is_unit_strided,
is_strided):
with is_unit_strided:
with cgutils.for_range(builder, loopcount, intp=intp_t) as ind:
fastloop = build_fast_loop_body(context, func, builder,
arrays, out, offsets,
store_offset, signature,
ind)
builder.ret_void()
with is_strided:
# General loop
with cgutils.for_range(builder, loopcount, intp=intp_t):
slowloop = build_slow_loop_body(context, func, builder,
arrays, out, offsets,
store_offset, signature)
builder.ret_void()
builder.ret_void()
del builder
# Run optimizer
library.add_ir_module(wrapper_module)
wrapper = library.get_function(wrapper.name)
return wrapper
def setitem_array1d_slice(context, builder, sig, args):
aryty, idxty, valty = sig.args
ary, idx, val = args
arystty = make_array(aryty)
ary = arystty(context, builder, ary)
shapes = cgutils.unpack_tuple(builder, ary.shape, aryty.ndim)
slicestruct = Slice(context, builder, value=idx)
# the logic here follows that of Python's Objects/sliceobject.c
# in particular PySlice_GetIndicesEx function
ZERO = Constant.int(slicestruct.step.type, 0)
NEG_ONE = Constant.int(slicestruct.start.type, -1)
b_step_eq_zero = builder.icmp(lc.ICMP_EQ, slicestruct.step, ZERO)
# bail if step is 0
with cgutils.ifthen(builder, b_step_eq_zero):
context.return_errcode(builder, errcode.ASSERTION_ERROR)
# adjust for negative indices for start
start = cgutils.alloca_once_value(builder, slicestruct.start)
b_start_lt_zero = builder.icmp(lc.ICMP_SLT, builder.load(start), ZERO)
with cgutils.ifthen(builder, b_start_lt_zero):
add = builder.add(builder.load(start), shapes[0])
builder.store(add, start)
b_start_lt_zero = builder.icmp(lc.ICMP_SLT, builder.load(start), ZERO)
with cgutils.ifthen(builder, b_start_lt_zero):
b_step_lt_zero = builder.icmp(lc.ICMP_SLT, slicestruct.step, ZERO)
cond = builder.select(b_step_lt_zero, NEG_ONE, ZERO)
builder.store(cond, start)
b_start_geq_len = builder.icmp(lc.ICMP_SGE, builder.load(start), shapes[0])
ONE = Constant.int(shapes[0].type, 1)
with cgutils.ifthen(builder, b_start_geq_len):
b_step_lt_zero = builder.icmp(lc.ICMP_SLT, slicestruct.step, ZERO)
cond = builder.select(b_step_lt_zero, builder.sub(shapes[0], ONE), shapes[0])
builder.store(cond, start)
# adjust stop for negative value
stop = cgutils.alloca_once_value(builder, slicestruct.stop)
b_stop_lt_zero = builder.icmp(lc.ICMP_SLT, builder.load(stop), ZERO)
with cgutils.ifthen(builder, b_stop_lt_zero):
add = builder.add(builder.load(stop), shapes[0])
builder.store(add, stop)
b_stop_lt_zero = builder.icmp(lc.ICMP_SLT, builder.load(stop), ZERO)
with cgutils.ifthen(builder, b_stop_lt_zero):
b_step_lt_zero = builder.icmp(lc.ICMP_SLT, slicestruct.step, ZERO)
cond = builder.select(b_step_lt_zero, NEG_ONE, ZERO)
builder.store(cond, start)
b_stop_geq_len = builder.icmp(lc.ICMP_SGE, builder.load(stop), shapes[0])
ONE = Constant.int(shapes[0].type, 1)
with cgutils.ifthen(builder, b_stop_geq_len):
b_step_lt_zero = builder.icmp(lc.ICMP_SLT, slicestruct.step, ZERO)
cond = builder.select(b_step_lt_zero, builder.sub(shapes[0], ONE), shapes[0])
builder.store(cond, stop)
b_step_gt_zero = builder.icmp(lc.ICMP_SGT, slicestruct.step, ZERO)
with cgutils.ifelse(builder, b_step_gt_zero) as (then0, otherwise0):
with then0:
with cgutils.for_range_slice(builder, builder.load(start), builder.load(stop), slicestruct.step, slicestruct.start.type) as loop_idx1:
ptr = cgutils.get_item_pointer(builder, aryty, ary,
[loop_idx1],
wraparound=True)
context.pack_value(builder, aryty.dtype, val, ptr)
with otherwise0:
with cgutils.for_range_slice(builder, builder.load(start), builder.load(stop), slicestruct.step, slicestruct.start.type, inc=False) as loop_idx2:
ptr = cgutils.get_item_pointer(builder, aryty, ary,
[loop_idx2],
wraparound=True)
context.pack_value(builder, aryty.dtype, val, ptr)
def impl(context, builder, sig, args):
[tyvx, tywy, tyout] = sig.args
[vx, wy, out] = args
assert tyvx.dtype == tywy.dtype
ndim = tyvx.ndim
xary = context.make_array(tyvx)(context, builder, vx)
yary = context.make_array(tywy)(context, builder, wy)
oary = context.make_array(tyout)(context, builder, out)
intpty = context.get_value_type(types.intp)
# TODO handle differing shape by mimicking broadcasting
loopshape = cgutils.unpack_tuple(builder, xary.shape, ndim)
xyo_shape = [cgutils.unpack_tuple(builder, ary.shape, ndim)
for ary in (xary, yary, oary)]
xyo_strides = [cgutils.unpack_tuple(builder, ary.strides, ndim)
for ary in (xary, yary, oary)]
xyo_data = [ary.data for ary in (xary, yary, oary)]
xyo_layout = [ty.layout for ty in (tyvx, tywy, tyout)]
with cgutils.loop_nest(builder, loopshape, intp=intpty) as indices:
[px, py, po] = [cgutils.get_item_pointer2(builder,
data=data, shape=shape,
strides=strides,
layout=layout,
inds=indices)
for data, shape, strides, layout
in zip(xyo_data, xyo_shape, xyo_strides,
xyo_layout)]
x = builder.load(px)
y = builder.load(py)
if divbyzero:
# Handle division
iszero = cgutils.is_scalar_zero(builder, y)
with cgutils.ifelse(builder, iszero, expect=False) as (then,
orelse):
with then:
# Divide by zero
if tyout.dtype in types.real_domain:
# If x is float and is 0 also, return Nan; else
# return Inf
outltype = context.get_data_type(tyout.dtype)
shouldretnan = cgutils.is_scalar_zero(builder, x)
nan = Constant.real(outltype, float("nan"))
inf = Constant.real(outltype, float("inf"))
res = builder.select(shouldretnan, nan, inf)
elif (tyout.dtype in types.signed_domain and
not numpy_support.int_divbyzero_returns_zero):
res = Constant.int(y.type, 0x1 << (y.type.width-1))
else:
res = Constant.null(y.type)
assert res.type == po.type.pointee, \
(str(res.type), str(po.type.pointee))
builder.store(res, po)
with orelse:
# Normal
res = core(builder, (x, y))
assert res.type == po.type.pointee, \
(str(res.type), str(po.type.pointee))
builder.store(res, po)
else:
# Handle other operations
res = core(builder, (x, y))
assert res.type == po.type.pointee, (res.type,
po.type.pointee)
builder.store(res, po)
return out
def lower_expr(self, expr):
if expr.op == 'binop':
return self.lower_binop(expr, inplace=False)
elif expr.op == 'inplace_binop':
return self.lower_binop(expr, inplace=True)
elif expr.op == 'unary':
value = self.loadvar(expr.value.name)
if expr.fn == '-':
res = self.pyapi.number_negative(value)
elif expr.fn == '+':
res = self.pyapi.number_positive(value)
elif expr.fn == 'not':
res = self.pyapi.object_not(value)
self.check_int_status(res)
longval = self.builder.zext(res, self.pyapi.long)
res = self.pyapi.bool_from_long(longval)
elif expr.fn == '~':
res = self.pyapi.number_invert(value)
else:
raise NotImplementedError(expr)
self.check_error(res)
return res
elif expr.op == 'call':
argvals = [self.loadvar(a.name) for a in expr.args]
fn = self.loadvar(expr.func.name)
if not expr.kws:
# No keyword
ret = self.pyapi.call_function_objargs(fn, argvals)
else:
# Have Keywords
keyvalues = [(k, self.loadvar(v.name)) for k, v in expr.kws]
args = self.pyapi.tuple_pack(argvals)
kws = self.pyapi.dict_pack(keyvalues)
ret = self.pyapi.call(fn, args, kws)
self.decref(kws)
self.decref(args)
self.check_error(ret)
return ret
elif expr.op == 'getattr':
obj = self.loadvar(expr.value.name)
res = self.pyapi.object_getattr_string(obj, expr.attr)
self.check_error(res)
return res
elif expr.op == 'build_tuple':
items = [self.loadvar(it.name) for it in expr.items]
res = self.pyapi.tuple_pack(items)
self.check_error(res)
return res
elif expr.op == 'build_list':
items = [self.loadvar(it.name) for it in expr.items]
res = self.pyapi.list_pack(items)
self.check_error(res)
return res
elif expr.op == 'build_map':
res = self.pyapi.dict_new(expr.size)
self.check_error(res)
return res
elif expr.op == 'build_set':
items = [self.loadvar(it.name) for it in expr.items]
res = self.pyapi.set_new()
self.check_error(res)
for it in items:
ok = self.pyapi.set_add(res, it)
self.check_int_status(ok)
return res
elif expr.op == 'getiter':
obj = self.loadvar(expr.value.name)
res = self.pyapi.object_getiter(obj)
self.check_error(res)
return res
elif expr.op == 'iternext':
iterobj = self.loadvar(expr.value.name)
item = self.pyapi.iter_next(iterobj)
is_valid = cgutils.is_not_null(self.builder, item)
pair = self.pyapi.tuple_new(2)
with cgutils.ifelse(self.builder, is_valid) as (then, otherwise):
with then:
self.pyapi.tuple_setitem(pair, 0, item)
with otherwise:
self.check_occurred()
# Make the tuple valid by inserting None as dummy
# iteration "result" (it will be ignored).
self.pyapi.tuple_setitem(pair, 0, self.pyapi.make_none())
self.pyapi.tuple_setitem(pair, 1,
self.pyapi.bool_from_bool(is_valid))
return pair
elif expr.op == 'pair_first':
pair = self.loadvar(expr.value.name)
first = self.pyapi.tuple_getitem(pair, 0)
self.incref(first)
return first
elif expr.op == 'pair_second':
pair = self.loadvar(expr.value.name)
second = self.pyapi.tuple_getitem(pair, 1)
self.incref(second)
return second
elif expr.op == 'exhaust_iter':
iterobj = self.loadvar(expr.value.name)
tup = self.pyapi.sequence_tuple(iterobj)
self.check_error(tup)
# Check tuple size is as expected
tup_size = self.pyapi.tuple_size(tup)
expected_size = self.context.get_constant(types.intp, expr.count)
has_wrong_size = self.builder.icmp(lc.ICMP_NE, tup_size,
expected_size)
with cgutils.if_unlikely(self.builder, has_wrong_size):
excid = self.add_exception(ValueError)
self.context.return_user_exc(self.builder, excid)
return tup
elif expr.op == 'getitem':
value = self.loadvar(expr.value.name)
index = self.loadvar(expr.index.name)
res = self.pyapi.object_getitem(value, index)
self.check_error(res)
return res
elif expr.op == 'static_getitem':
value = self.loadvar(expr.value.name)
index = self.context.get_constant(types.intp, expr.index)
indexobj = self.pyapi.long_from_ssize_t(index)
self.check_error(indexobj)
res = self.pyapi.object_getitem(value, indexobj)
self.decref(indexobj)
self.check_error(res)
return res
elif expr.op == 'getslice':
target = self.loadvar(expr.target.name)
start = self.loadvar(expr.start.name)
stop = self.loadvar(expr.stop.name)
slicefn = self.get_builtin_obj("slice")
sliceobj = self.pyapi.call_function_objargs(slicefn, (start, stop))
self.decref(slicefn)
self.check_error(sliceobj)
res = self.pyapi.object_getitem(target, sliceobj)
self.check_error(res)
return res
elif expr.op == 'cast':
val = self.loadvar(expr.value.name)
self.incref(val)
return val
else:
raise NotImplementedError(expr)
def build_ufunc_wrapper(context, func, signature):
"""
Wrap the scalar function with a loop that iterates over the arguments
"""
module = func.module
byte_t = Type.int(8)
byte_ptr_t = Type.pointer(byte_t)
byte_ptr_ptr_t = Type.pointer(byte_ptr_t)
intp_t = context.get_value_type(types.intp)
intp_ptr_t = Type.pointer(intp_t)
fnty = Type.function(Type.void(),
[byte_ptr_ptr_t, intp_ptr_t, intp_ptr_t, byte_ptr_t])
wrapper = module.add_function(fnty, "__ufunc__." + func.name)
arg_args, arg_dims, arg_steps, arg_data = wrapper.args
arg_args.name = "args"
arg_dims.name = "dims"
arg_steps.name = "steps"
arg_data.name = "data"
builder = Builder.new(wrapper.append_basic_block("entry"))
loopcount = builder.load(arg_dims, name="loopcount")
actual_args = context.get_arguments(func)
# Prepare inputs
arrays = []
for i, typ in enumerate(signature.args):
arrays.append(
UArrayArg(context, builder, arg_args, arg_steps, i,
context.get_argument_type(typ)))
# Prepare output
valty = context.get_data_type(signature.return_type)
out = UArrayArg(context, builder, arg_args, arg_steps, len(actual_args),
valty)
# Setup indices
offsets = []
zero = context.get_constant(types.intp, 0)
for _ in arrays:
p = cgutils.alloca_once(builder, intp_t)
offsets.append(p)
builder.store(zero, p)
store_offset = cgutils.alloca_once(builder, intp_t)
builder.store(zero, store_offset)
unit_strided = cgutils.true_bit
for ary in arrays:
unit_strided = builder.and_(unit_strided, ary.is_unit_strided)
with cgutils.ifelse(builder,
unit_strided) as (is_unit_strided, is_strided):
with is_unit_strided:
with cgutils.for_range(builder, loopcount, intp=intp_t) as ind:
fastloop = build_fast_loop_body(context, func, builder, arrays,
out, offsets, store_offset,
signature, ind)
builder.ret_void()
with is_strided:
# General loop
with cgutils.for_range(builder, loopcount, intp=intp_t):
slowloop = build_slow_loop_body(context, func, builder, arrays,
out, offsets, store_offset,
signature)
builder.ret_void()
builder.ret_void()
del builder
# Set core function to internal so that it is not generated
func.linkage = LINKAGE_INTERNAL
# Force inline of code function
inline_function(slowloop)
inline_function(fastloop)
# Run optimizer
context.optimize(module)
if config.DUMP_OPTIMIZED:
print(module)
return wrapper
def impl(context, builder, sig, args):
[tyinp1, tyinp2, tyout] = sig.args
[inp1, inp2, out] = args
if isinstance(tyinp1, types.Array):
scalar_inp1 = False
scalar_tyinp1 = tyinp1.dtype
inp1_ndim = tyinp1.ndim
elif tyinp1 in types.number_domain:
scalar_inp1 = True
scalar_tyinp1 = tyinp1
inp1_ndim = 1
else:
raise TypeError('unknown type for first input operand')
if isinstance(tyinp2, types.Array):
scalar_inp2 = False
scalar_tyinp2 = tyinp2.dtype
inp2_ndim = tyinp2.ndim
elif tyinp2 in types.number_domain:
scalar_inp2 = True
scalar_tyinp2 = tyinp2
inp2_ndim = 1
else:
raise TypeError('unknown type for second input operand')
out_ndim = tyout.ndim
if asfloat:
promote_type = types.float64
elif scalar_tyinp1 in types.real_domain or \
scalar_tyinp2 in types.real_domain:
promote_type = types.float64
elif scalar_tyinp1 in types.signed_domain or \
scalar_tyinp2 in types.signed_domain:
promote_type = types.int64
else:
promote_type = types.uint64
result_type = promote_type
# Temporary hack for __ftol2 llvm bug. Don't allow storing
# float results in uint64 array on windows.
if result_type in types.real_domain and \
tyout.dtype is types.uint64 and \
sys.platform.startswith('win32'):
raise TypeError('Cannot store result in uint64 array')
sig = typing.signature(result_type, promote_type, promote_type)
if not scalar_inp1:
i1ary = context.make_array(tyinp1)(context, builder, inp1)
if not scalar_inp2:
i2ary = context.make_array(tyinp2)(context, builder, inp2)
oary = context.make_array(tyout)(context, builder, out)
fnwork = context.get_function(funckey, sig)
intpty = context.get_value_type(types.intp)
if not scalar_inp1:
inp1_shape = cgutils.unpack_tuple(builder, i1ary.shape, inp1_ndim)
inp1_strides = cgutils.unpack_tuple(builder, i1ary.strides, inp1_ndim)
inp1_data = i1ary.data
inp1_layout = tyinp1.layout
if not scalar_inp2:
inp2_shape = cgutils.unpack_tuple(builder, i2ary.shape, inp2_ndim)
inp2_strides = cgutils.unpack_tuple(builder, i2ary.strides, inp2_ndim)
inp2_data = i2ary.data
inp2_layout = tyinp2.layout
out_shape = cgutils.unpack_tuple(builder, oary.shape, out_ndim)
out_strides = cgutils.unpack_tuple(builder, oary.strides, out_ndim)
out_data = oary.data
out_layout = tyout.layout
ZERO = Constant.int(Type.int(intpty.width), 0)
ONE = Constant.int(Type.int(intpty.width), 1)
inp1_indices = None
if not scalar_inp1:
inp1_indices = []
for i in range(inp1_ndim):
x = builder.alloca(Type.int(intpty.width))
builder.store(ZERO, x)
inp1_indices.append(x)
inp2_indices = None
if not scalar_inp2:
inp2_indices = []
for i in range(inp2_ndim):
x = builder.alloca(Type.int(intpty.width))
builder.store(ZERO, x)
inp2_indices.append(x)
loopshape = cgutils.unpack_tuple(builder, oary.shape, out_ndim)
with cgutils.loop_nest(builder, loopshape, intp=intpty) as indices:
# Increment input indices.
# Since the output dimensions are already being incremented,
# we'll use that to set the input indices. In order to
# handle broadcasting, any input dimension of size 1 won't be
# incremented.
def build_increment_blocks(inp_indices, inp_shape, inp_ndim, inp_num):
bb_inc_inp_index = [cgutils.append_basic_block(builder,
'.inc_inp{0}_index{1}'.format(inp_num, str(i))) for i in range(inp_ndim)]
bb_end_inc_index = cgutils.append_basic_block(builder,
'.end_inc{0}_index'.format(inp_num))
builder.branch(bb_inc_inp_index[0])
for i in range(inp_ndim):
with cgutils.goto_block(builder, bb_inc_inp_index[i]):
# If the shape of this dimension is 1, then leave the
# index at 0 so that this dimension is broadcasted over
# the corresponding input and output dimensions.
cond = builder.icmp(ICMP_UGT, inp_shape[i], ONE)
with cgutils.ifthen(builder, cond):
builder.store(indices[out_ndim-inp_ndim+i], inp_indices[i])
if i + 1 == inp_ndim:
builder.branch(bb_end_inc_index)
else:
builder.branch(bb_inc_inp_index[i+1])
builder.position_at_end(bb_end_inc_index)
if not scalar_inp1:
build_increment_blocks(inp1_indices, inp1_shape, inp1_ndim, '1')
if not scalar_inp2:
build_increment_blocks(inp2_indices, inp2_shape, inp2_ndim, '2')
if scalar_inp1:
x = inp1
else:
inds = [builder.load(index) for index in inp1_indices]
px = cgutils.get_item_pointer2(builder,
data=inp1_data,
shape=inp1_shape,
strides=inp1_strides,
layout=inp1_layout,
inds=inds)
x = builder.load(px)
if scalar_inp2:
y = inp2
else:
inds = [builder.load(index) for index in inp2_indices]
py = cgutils.get_item_pointer2(builder,
data=inp2_data,
shape=inp2_shape,
strides=inp2_strides,
layout=inp2_layout,
inds=inds)
y = builder.load(py)
po = cgutils.get_item_pointer2(builder,
data=out_data,
shape=out_shape,
strides=out_strides,
layout=out_layout,
inds=indices)
if divbyzero:
# Handle division
iszero = cgutils.is_scalar_zero(builder, y)
with cgutils.ifelse(builder, iszero, expect=False) as (then,
orelse):
with then:
# Divide by zero
if (scalar_tyinp1 in types.real_domain or
scalar_tyinp2 in types.real_domain) or \
not numpy_support.int_divbyzero_returns_zero:
# If y is float and is 0 also, return Nan; else
# return Inf
outltype = context.get_data_type(result_type)
shouldretnan = cgutils.is_scalar_zero(builder, x)
nan = Constant.real(outltype, float("nan"))
inf = Constant.real(outltype, float("inf"))
tempres = builder.select(shouldretnan, nan, inf)
res = context.cast(builder, tempres, result_type,
tyout.dtype)
elif tyout.dtype in types.signed_domain and \
not numpy_support.int_divbyzero_returns_zero:
res = Constant.int(context.get_data_type(tyout.dtype),
0x1 << (y.type.width-1))
else:
res = Constant.null(context.get_data_type(tyout.dtype))
assert res.type == po.type.pointee, \
(str(res.type), str(po.type.pointee))
builder.store(res, po)
with orelse:
# Normal
d_x = context.cast(builder, x, scalar_tyinp1, promote_type)
d_y = context.cast(builder, y, scalar_tyinp2, promote_type)
tempres = fnwork(builder, [d_x, d_y])
res = context.cast(builder, tempres, result_type, tyout.dtype)
assert res.type == po.type.pointee, (res.type,
po.type.pointee)
builder.store(res, po)
else:
# Handle non-division operations
d_x = context.cast(builder, x, scalar_tyinp1, promote_type)
d_y = context.cast(builder, y, scalar_tyinp2, promote_type)
tempres = fnwork(builder, [d_x, d_y])
res = context.cast(builder, tempres, result_type, tyout.dtype)
assert res.type == po.type.pointee, (res.type,
po.type.pointee)
builder.store(res, po)
return out
def setitem_array1d_slice(context, builder, sig, args):
aryty, idxty, valty = sig.args
ary, idx, val = args
arystty = make_array(aryty)
ary = arystty(context, builder, ary)
shapes = cgutils.unpack_tuple(builder, ary.shape, aryty.ndim)
slicestruct = Slice(context, builder, value=idx)
# the logic here follows that of Python's Objects/sliceobject.c
# in particular PySlice_GetIndicesEx function
ZERO = Constant.int(slicestruct.step.type, 0)
NEG_ONE = Constant.int(slicestruct.start.type, -1)
b_step_eq_zero = builder.icmp(lc.ICMP_EQ, slicestruct.step, ZERO)
# bail if step is 0
with cgutils.ifthen(builder, b_step_eq_zero):
context.call_conv.return_user_exc(builder, ValueError,
("slice step cannot be zero", ))
# adjust for negative indices for start
start = cgutils.alloca_once_value(builder, slicestruct.start)
b_start_lt_zero = builder.icmp(lc.ICMP_SLT, builder.load(start), ZERO)
with cgutils.ifthen(builder, b_start_lt_zero):
add = builder.add(builder.load(start), shapes[0])
builder.store(add, start)
b_start_lt_zero = builder.icmp(lc.ICMP_SLT, builder.load(start), ZERO)
with cgutils.ifthen(builder, b_start_lt_zero):
b_step_lt_zero = builder.icmp(lc.ICMP_SLT, slicestruct.step, ZERO)
cond = builder.select(b_step_lt_zero, NEG_ONE, ZERO)
builder.store(cond, start)
b_start_geq_len = builder.icmp(lc.ICMP_SGE, builder.load(start), shapes[0])
ONE = Constant.int(shapes[0].type, 1)
with cgutils.ifthen(builder, b_start_geq_len):
b_step_lt_zero = builder.icmp(lc.ICMP_SLT, slicestruct.step, ZERO)
cond = builder.select(b_step_lt_zero, builder.sub(shapes[0], ONE),
shapes[0])
builder.store(cond, start)
# adjust stop for negative value
stop = cgutils.alloca_once_value(builder, slicestruct.stop)
b_stop_lt_zero = builder.icmp(lc.ICMP_SLT, builder.load(stop), ZERO)
with cgutils.ifthen(builder, b_stop_lt_zero):
add = builder.add(builder.load(stop), shapes[0])
builder.store(add, stop)
b_stop_lt_zero = builder.icmp(lc.ICMP_SLT, builder.load(stop), ZERO)
with cgutils.ifthen(builder, b_stop_lt_zero):
b_step_lt_zero = builder.icmp(lc.ICMP_SLT, slicestruct.step, ZERO)
cond = builder.select(b_step_lt_zero, NEG_ONE, ZERO)
builder.store(cond, start)
b_stop_geq_len = builder.icmp(lc.ICMP_SGE, builder.load(stop), shapes[0])
ONE = Constant.int(shapes[0].type, 1)
with cgutils.ifthen(builder, b_stop_geq_len):
b_step_lt_zero = builder.icmp(lc.ICMP_SLT, slicestruct.step, ZERO)
cond = builder.select(b_step_lt_zero, builder.sub(shapes[0], ONE),
shapes[0])
builder.store(cond, stop)
b_step_gt_zero = builder.icmp(lc.ICMP_SGT, slicestruct.step, ZERO)
with cgutils.ifelse(builder, b_step_gt_zero) as (then0, otherwise0):
with then0:
with cgutils.for_range_slice(builder, builder.load(start),
builder.load(stop), slicestruct.step,
slicestruct.start.type) as loop_idx1:
ptr = cgutils.get_item_pointer(builder,
aryty,
ary, [loop_idx1],
wraparound=True)
context.pack_value(builder, aryty.dtype, val, ptr)
with otherwise0:
with cgutils.for_range_slice(builder,
builder.load(start),
builder.load(stop),
slicestruct.step,
slicestruct.start.type,
inc=False) as loop_idx2:
ptr = cgutils.get_item_pointer(builder,
aryty,
ary, [loop_idx2],
wraparound=True)
context.pack_value(builder, aryty.dtype, val, ptr)
def build_ufunc_wrapper(context, func, signature):
"""
Wrap the scalar function with a loop that iterates over the arguments
"""
module = func.module
byte_t = Type.int(8)
byte_ptr_t = Type.pointer(byte_t)
byte_ptr_ptr_t = Type.pointer(byte_ptr_t)
intp_t = context.get_value_type(types.intp)
intp_ptr_t = Type.pointer(intp_t)
fnty = Type.function(Type.void(), [byte_ptr_ptr_t, intp_ptr_t,
intp_ptr_t, byte_ptr_t])
wrapper = module.add_function(fnty, "__ufunc__." + func.name)
arg_args, arg_dims, arg_steps, arg_data = wrapper.args
arg_args.name = "args"
arg_dims.name = "dims"
arg_steps.name = "steps"
arg_data.name = "data"
builder = Builder.new(wrapper.append_basic_block("entry"))
loopcount = builder.load(arg_dims, name="loopcount")
actual_args = context.get_arguments(func)
# Prepare inputs
arrays = []
for i, typ in enumerate(signature.args):
arrays.append(UArrayArg(context, builder, arg_args, arg_steps, i,
context.get_argument_type(typ)))
# Prepare output
valty = context.get_data_type(signature.return_type)
out = UArrayArg(context, builder, arg_args, arg_steps, len(actual_args),
valty)
# Setup indices
offsets = []
zero = context.get_constant(types.intp, 0)
for _ in arrays:
p = cgutils.alloca_once(builder, intp_t)
offsets.append(p)
builder.store(zero, p)
store_offset = cgutils.alloca_once(builder, intp_t)
builder.store(zero, store_offset)
unit_strided = cgutils.true_bit
for ary in arrays:
unit_strided = builder.and_(unit_strided, ary.is_unit_strided)
with cgutils.ifelse(builder, unit_strided) as (is_unit_strided,
is_strided):
with is_unit_strided:
with cgutils.for_range(builder, loopcount, intp=intp_t) as ind:
fastloop = build_fast_loop_body(context, func, builder,
arrays, out, offsets,
store_offset, signature, ind)
builder.ret_void()
with is_strided:
# General loop
with cgutils.for_range(builder, loopcount, intp=intp_t):
slowloop = build_slow_loop_body(context, func, builder,
arrays, out, offsets,
store_offset, signature)
builder.ret_void()
builder.ret_void()
del builder
# Set core function to internal so that it is not generated
func.linkage = LINKAGE_INTERNAL
# Force inline of code function
inline_function(slowloop)
inline_function(fastloop)
# Run optimizer
context.optimize(module)
if config.DUMP_OPTIMIZED:
print(module)
return wrapper
def impl(context, builder, sig, args):
[tyinp1, tyinp2, tyout] = sig.args
[inp1, inp2, out] = args
if scalar_inputs:
ndim = 1
else:
ndim = tyinp1.ndim
# Temporary hack for __ftol2 llvm bug. Don't allow storing
# float results in uint64 array on windows.
if scalar_inputs and tyinp1 in types.real_domain and \
tyout.dtype is types.uint64 and \
sys.platform.startswith('win32'):
raise TypeError('Cannot store result in uint64 array')
if not scalar_inputs and tyinp1.dtype in types.real_domain and \
tyout.dtype is types.uint64 and \
sys.platform.startswith('win32'):
raise TypeError('Cannot store result in uint64 array')
if not scalar_inputs:
i1ary = context.make_array(tyinp1)(context, builder, inp1)
i2ary = context.make_array(tyinp2)(context, builder, inp2)
oary = context.make_array(tyout)(context, builder, out)
if asfloat and not divbyzero:
sig = typing.signature(types.float64, types.float64, types.float64)
else:
if scalar_inputs:
sig = typing.signature(tyout.dtype, tyinp1, tyinp2)
else:
sig = typing.signature(tyout.dtype, tyinp1.dtype, tyinp2.dtype)
fnwork = context.get_function(funckey, sig)
intpty = context.get_value_type(types.intp)
# TODO handle differing shape by mimicking broadcasting
loopshape = cgutils.unpack_tuple(builder, oary.shape, ndim)
if scalar_inputs:
xyo_shape = [cgutils.unpack_tuple(builder, ary.shape, ndim)
for ary in (oary,)]
xyo_strides = [cgutils.unpack_tuple(builder, ary.strides, ndim)
for ary in (oary,)]
xyo_data = [ary.data for ary in (oary,)]
xyo_layout = [ty.layout for ty in (tyout,)]
else:
xyo_shape = [cgutils.unpack_tuple(builder, ary.shape, ndim)
for ary in (i1ary, i2ary, oary)]
xyo_strides = [cgutils.unpack_tuple(builder, ary.strides, ndim)
for ary in (i1ary, i2ary, oary)]
xyo_data = [ary.data for ary in (i1ary, i2ary, oary)]
xyo_layout = [ty.layout for ty in (tyinp1, tyinp2, tyout)]
with cgutils.loop_nest(builder, loopshape, intp=intpty) as indices:
if scalar_inputs:
[po] = [cgutils.get_item_pointer2(builder,
data=data, shape=shape,
strides=strides,
layout=layout,
inds=indices)
for data, shape, strides, layout
in zip(xyo_data, xyo_shape, xyo_strides,
xyo_layout)]
else:
[px, py, po] = [cgutils.get_item_pointer2(builder,
data=data, shape=shape,
strides=strides,
layout=layout,
inds=indices)
for data, shape, strides, layout
in zip(xyo_data, xyo_shape, xyo_strides,
xyo_layout)]
if scalar_inputs:
x = inp1
y = inp2
else:
x = builder.load(px)
y = builder.load(py)
if divbyzero:
# Handle division
iszero = cgutils.is_scalar_zero(builder, y)
with cgutils.ifelse(builder, iszero, expect=False) as (then,
orelse):
with then:
# Divide by zero
if ((scalar_inputs and tyinp2 in types.real_domain) or
(not scalar_inputs and
tyinp2.dtype in types.real_domain) or
not numpy_support.int_divbyzero_returns_zero):
# If y is float and is 0 also, return Nan; else
# return Inf
outltype = context.get_data_type(tyout.dtype)
shouldretnan = cgutils.is_scalar_zero(builder, x)
nan = Constant.real(outltype, float("nan"))
inf = Constant.real(outltype, float("inf"))
res = builder.select(shouldretnan, nan, inf)
elif (scalar_inputs and tyout in types.signed_domain and
not numpy_support.int_divbyzero_returns_zero):
res = Constant.int(context.get_data_type(tyout),
0x1 << (y.type.width-1))
elif (not scalar_inputs and
tyout.dtype in types.signed_domain and
not numpy_support.int_divbyzero_returns_zero):
res = Constant.int(context.get_data_type(tyout.dtype),
0x1 << (y.type.width-1))
else:
res = Constant.null(context.get_data_type(tyout.dtype))
assert res.type == po.type.pointee, \
(str(res.type), str(po.type.pointee))
builder.store(res, po)
with orelse:
# Normal
tempres = fnwork(builder, (x, y))
if scalar_inputs and tyinp1 in types.real_domain:
res = context.cast(builder, tempres,
tyinp1, tyout.dtype)
elif (not scalar_inputs and
tyinp1.dtype in types.real_domain):
res = context.cast(builder, tempres,
tyinp1.dtype, tyout.dtype)
else:
res = context.cast(builder, tempres,
types.float64, tyout.dtype)
assert res.type == po.type.pointee, \
(str(res.type), str(po.type.pointee))
builder.store(res, po)
else:
# Handle non-division operations
if asfloat:
if scalar_inputs:
d_x = context.cast(builder, x, tyinp1, types.float64)
d_y = context.cast(builder, y, tyinp2, types.float64)
else:
d_x = context.cast(builder, x, tyinp1.dtype,
types.float64)
d_y = context.cast(builder, y, tyinp2.dtype,
types.float64)
tempres = fnwork(builder, [d_x, d_y])
res = context.cast(builder, tempres,
types.float64, tyout.dtype)
elif scalar_inputs:
if tyinp1 != tyout.dtype:
tempres = fnwork(builder, [x, y])
res = context.cast(builder, tempres, tyinp1,
tyout.dtype)
else:
res = fnwork(builder, (x, y))
elif tyinp1.dtype != tyout.dtype:
tempres = fnwork(builder, [x, y])
res = context.cast(builder, tempres, tyinp1.dtype,
tyout.dtype)
else:
res = fnwork(builder, (x, y))
assert res.type == po.type.pointee, (res.type,
po.type.pointee)
builder.store(res, po)
return out
def lower_expr(self, expr):
if expr.op == 'binop':
return self.lower_binop(expr, inplace=False)
elif expr.op == 'inplace_binop':
return self.lower_binop(expr, inplace=True)
elif expr.op == 'unary':
value = self.loadvar(expr.value.name)
if expr.fn == '-':
res = self.pyapi.number_negative(value)
elif expr.fn == '+':
res = self.pyapi.number_positive(value)
elif expr.fn == 'not':
res = self.pyapi.object_not(value)
self.check_int_status(res)
longval = self.builder.zext(res, self.pyapi.long)
res = self.pyapi.bool_from_long(longval)
elif expr.fn == '~':
res = self.pyapi.number_invert(value)
else:
raise NotImplementedError(expr)
self.check_error(res)
return res
elif expr.op == 'call':
argvals = [self.loadvar(a.name) for a in expr.args]
fn = self.loadvar(expr.func.name)
if not expr.kws:
# No keyword
ret = self.pyapi.call_function_objargs(fn, argvals)
else:
# Have Keywords
keyvalues = [(k, self.loadvar(v.name)) for k, v in expr.kws]
args = self.pyapi.tuple_pack(argvals)
kws = self.pyapi.dict_pack(keyvalues)
ret = self.pyapi.call(fn, args, kws)
self.decref(kws)
self.decref(args)
self.check_error(ret)
return ret
elif expr.op == 'getattr':
obj = self.loadvar(expr.value.name)
res = self.pyapi.object_getattr(obj, self._freeze_string(expr.attr))
self.check_error(res)
return res
elif expr.op == 'build_tuple':
items = [self.loadvar(it.name) for it in expr.items]
res = self.pyapi.tuple_pack(items)
self.check_error(res)
return res
elif expr.op == 'build_list':
items = [self.loadvar(it.name) for it in expr.items]
res = self.pyapi.list_pack(items)
self.check_error(res)
return res
elif expr.op == 'build_map':
res = self.pyapi.dict_new(expr.size)
self.check_error(res)
return res
elif expr.op == 'build_set':
items = [self.loadvar(it.name) for it in expr.items]
res = self.pyapi.set_new()
self.check_error(res)
for it in items:
ok = self.pyapi.set_add(res, it)
self.check_int_status(ok)
return res
elif expr.op == 'getiter':
obj = self.loadvar(expr.value.name)
res = self.pyapi.object_getiter(obj)
self.check_error(res)
return res
elif expr.op == 'iternext':
iterobj = self.loadvar(expr.value.name)
item = self.pyapi.iter_next(iterobj)
is_valid = cgutils.is_not_null(self.builder, item)
pair = self.pyapi.tuple_new(2)
with cgutils.ifelse(self.builder, is_valid) as (then, otherwise):
with then:
self.pyapi.tuple_setitem(pair, 0, item)
with otherwise:
self.check_occurred()
# Make the tuple valid by inserting None as dummy
# iteration "result" (it will be ignored).
self.pyapi.tuple_setitem(pair, 0, self.pyapi.make_none())
self.pyapi.tuple_setitem(pair, 1, self.pyapi.bool_from_bool(is_valid))
return pair
elif expr.op == 'pair_first':
pair = self.loadvar(expr.value.name)
first = self.pyapi.tuple_getitem(pair, 0)
self.incref(first)
return first
elif expr.op == 'pair_second':
pair = self.loadvar(expr.value.name)
second = self.pyapi.tuple_getitem(pair, 1)
self.incref(second)
return second
elif expr.op == 'exhaust_iter':
iterobj = self.loadvar(expr.value.name)
tup = self.pyapi.sequence_tuple(iterobj)
self.check_error(tup)
# Check tuple size is as expected
tup_size = self.pyapi.tuple_size(tup)
expected_size = self.context.get_constant(types.intp, expr.count)
has_wrong_size = self.builder.icmp(lc.ICMP_NE,
tup_size, expected_size)
with cgutils.if_unlikely(self.builder, has_wrong_size):
excid = self.add_exception(ValueError)
self.context.return_user_exc(self.builder, excid)
return tup
elif expr.op == 'getitem':
value = self.loadvar(expr.value.name)
index = self.loadvar(expr.index.name)
res = self.pyapi.object_getitem(value, index)
self.check_error(res)
return res
elif expr.op == 'static_getitem':
value = self.loadvar(expr.value.name)
index = self.context.get_constant(types.intp, expr.index)
indexobj = self.pyapi.long_from_ssize_t(index)
self.check_error(indexobj)
res = self.pyapi.object_getitem(value, indexobj)
self.decref(indexobj)
self.check_error(res)
return res
elif expr.op == 'getslice':
target = self.loadvar(expr.target.name)
start = self.loadvar(expr.start.name)
stop = self.loadvar(expr.stop.name)
slicefn = self.get_builtin_obj("slice")
sliceobj = self.pyapi.call_function_objargs(slicefn, (start, stop))
self.decref(slicefn)
self.check_error(sliceobj)
res = self.pyapi.object_getitem(target, sliceobj)
self.check_error(res)
return res
elif expr.op == 'cast':
val = self.loadvar(expr.value.name)
self.incref(val)
return val
else:
raise NotImplementedError(expr)
def to_native_value(self, obj, typ):
if isinstance(typ, types.Object) or typ == types.pyobject:
return obj
elif typ == types.boolean:
istrue = self.object_istrue(obj)
zero = Constant.null(istrue.type)
return self.builder.icmp(lc.ICMP_NE, istrue, zero)
elif typ in types.unsigned_domain:
longobj = self.number_long(obj)
ullval = self.long_as_ulonglong(longobj)
self.decref(longobj)
return self.builder.trunc(ullval,
self.context.get_argument_type(typ))
elif typ in types.signed_domain:
longobj = self.number_long(obj)
llval = self.long_as_longlong(longobj)
self.decref(longobj)
return self.builder.trunc(llval,
self.context.get_argument_type(typ))
elif typ == types.float32:
fobj = self.number_float(obj)
fval = self.float_as_double(fobj)
self.decref(fobj)
return self.builder.fptrunc(fval,
self.context.get_argument_type(typ))
elif typ == types.float64:
fobj = self.number_float(obj)
fval = self.float_as_double(fobj)
self.decref(fobj)
return fval
elif typ in (types.complex128, types.complex64):
cplxcls = self.context.make_complex(types.complex128)
cplx = cplxcls(self.context, self.builder)
pcplx = cplx._getpointer()
ok = self.complex_adaptor(obj, pcplx)
failed = cgutils.is_false(self.builder, ok)
with cgutils.if_unlikely(self.builder, failed):
self.builder.ret(self.get_null_object())
if typ == types.complex64:
c64cls = self.context.make_complex(typ)
c64 = c64cls(self.context, self.builder)
freal = self.context.cast(self.builder, cplx.real,
types.float64, types.float32)
fimag = self.context.cast(self.builder, cplx.imag,
types.float64, types.float32)
c64.real = freal
c64.imag = fimag
return c64._getvalue()
else:
return cplx._getvalue()
elif isinstance(typ, types.NPDatetime):
val = self.extract_np_datetime(obj)
return val
elif isinstance(typ, types.NPTimedelta):
val = self.extract_np_timedelta(obj)
return val
elif isinstance(typ, types.Array):
return self.to_native_array(typ, obj)
elif isinstance(typ, types.Optional):
isnone = self.builder.icmp(lc.ICMP_EQ, obj, self.borrow_none())
with cgutils.ifelse(self.builder, isnone) as (then, orelse):
with then:
noneval = self.context.make_optional_none(self.builder, typ.type)
ret = cgutils.alloca_once(self.builder, noneval.type)
self.builder.store(noneval, ret)
with orelse:
val = self.to_native_value(obj, typ.type)
just = self.context.make_optional_value(self.builder,
typ.type, val)
self.builder.store(just, ret)
return ret
raise NotImplementedError(typ)
