def wrap_constant(value):
if lltype.typeOf(value) == lltype.Signed:
return ConstInt(value)
elif isinstance(value, bool):
return ConstInt(int(value))
elif lltype.typeOf(value) == longlong.FLOATSTORAGE:
return ConstFloat(value)
elif isinstance(value, float):
return ConstFloat(longlong.getfloatstorage(value))
else:
assert lltype.typeOf(value) == llmemory.GCREF
return ConstPtr(value)
def test_ResumeDataLoopMemo_other():
memo = ResumeDataLoopMemo(FakeMetaInterpStaticData())
const = ConstFloat(longlong.getfloatstorage(-1.0))
tagged = memo.getconst(const)
index, tagbits = untag(tagged)
assert tagbits == TAGCONST
assert memo.consts[index - TAG_CONST_OFFSET] is const
def callback(asm):
c = ConstFloat(longlong.getfloatstorage(-42.5))
loc = self.xrm.convert_to_imm(c)
asm.mov(loc, xmm5)
asm.regalloc_push(xmm5)
asm.regalloc_pop(xmm0)
asm.mc.CVTTSD2SI(eax, xmm0)
def produce_into(self, builder, r):
if r.random() < 0.4:
UnaryOperation.produce_into(self, builder, r)
elif r.random() < 0.75 or not builder.cpu.supports_floats:
self.put(builder, [ConstInt(r.random_integer())])
else:
self.put(builder, [ConstFloat(r.random_float_storage())])
def produce_into(self, builder, r):
if not builder.floatvars:
raise CannotProduceOperation
k = r.random()
if k < 0.18:
v_first = ConstFloat(r.random_float_storage())
else:
v_first = r.choice(builder.floatvars)
if k > 0.82:
v_second = ConstFloat(r.random_float_storage())
else:
v_second = r.choice(builder.floatvars)
if abs(v_first.getfloat()) > 1E100 or abs(v_second.getfloat()) > 1E100:
raise CannotProduceOperation # avoid infinities
if abs(v_second.getfloat()) < 1E-100:
raise CannotProduceOperation # e.g. division by zero error
self.put(builder, [v_first, v_second])
def constant_from_op(op):
if op.type == 'i':
return ConstInt(op.getint())
elif op.type == 'r':
return ConstPtr(op.getref_base())
else:
assert op.type == 'f'
return ConstFloat(op.getfloatstorage())
def constbox(v):
if v.type == INT:
return ConstInt(getint(v))
if v.type == FLOAT:
return ConstFloat(getfloatstorage(v))
if v.type == REF:
return ConstPtr(getref_base(v))
assert 0, v.type
def produce_into(self, builder, r):
if not builder.floatvars:
raise CannotProduceOperation
k = r.random()
if k < 0.18:
v_first = ConstFloat(r.random_float_storage())
else:
v_first = r.choice(builder.floatvars)
if k > 0.82:
v_second = ConstFloat(r.random_float_storage())
else:
v_second = r.choice(builder.floatvars)
if abs(getfloat(v_first)) > 1E100 or abs(getfloat(v_second)) > 1E100:
raise CannotProduceOperation # avoid infinities
if abs(getfloat(v_second)) < 1E-100:
raise CannotProduceOperation # e.g. division by zero error
self.put(builder, [v_first, v_second])
def get_current_constant_fieldvalue(self):
struct = self.struct
fielddescr = self.fielddescr
if self.fielddescr.is_pointer_field():
return ConstPtr(self.cpu.bh_getfield_gc_r(struct, fielddescr))
elif self.fielddescr.is_float_field():
return ConstFloat(self.cpu.bh_getfield_gc_f(struct, fielddescr))
else:
return ConstInt(self.cpu.bh_getfield_gc_i(struct, fielddescr))
def accumulate_prepare(self, state):
vec_reg_size = state.vec_reg_size
for pack in self.packs:
if not pack.is_accumulating():
continue
if pack.leftmost().is_guard():
# guard breaks dependencies, thus it is an accumulation pack
continue
for i,node in enumerate(pack.operations):
op = node.getoperation()
state.accumulation[op] = pack
assert isinstance(pack, AccumPack)
datatype = pack.getdatatype()
bytesize = pack.getbytesize()
count = vec_reg_size // bytesize
signed = datatype == 'i'
oplist = state.invariant_oplist
# reset the box to zeros or ones
if pack.reduce_init() == 0:
vecop = OpHelpers.create_vec(datatype, bytesize, signed, count)
oplist.append(vecop)
opnum = rop.VEC_INT_XOR
if datatype == FLOAT:
# see PRECISION loss below
raise NotImplementedError
vecop = VecOperation(opnum, [vecop, vecop],
vecop, count)
oplist.append(vecop)
elif pack.reduce_init() == 1:
# PRECISION loss, because the numbers are accumulated (associative, commutative properties must hold)
# you can end up a small number and a huge number that is finally multiplied. giving an
# inprecision result, thus this is disabled now
raise NotImplementedError
# multiply is only supported by floats
vecop = OpHelpers.create_vec_expand(ConstFloat(1.0), bytesize,
signed, count)
oplist.append(vecop)
else:
raise NotImplementedError("cannot handle %s" % pack.operator)
# pack the scalar value
args = [vecop, pack.getleftmostseed(), ConstInt(0), ConstInt(1)]
vecop = OpHelpers.create_vec_pack(datatype, args, bytesize,
signed, count)
oplist.append(vecop)
seed = pack.getleftmostseed()
state.accumulation[seed] = pack
# rename the variable with the box
state.setvector_of_box(seed, 0, vecop) # prevent it from expansion
state.renamer.start_renaming(seed, vecop)
def _untag(self, tagged):
tag, v = untag(tagged)
if tag == TAGBOX:
return self._get(v)
elif tag == TAGINT:
return ConstInt(v + SMALL_INT_START)
elif tag == TAGCONSTPTR:
return ConstPtr(self.trace._refs[v])
elif tag == TAGCONSTOTHER:
if v & 1:
return ConstFloat(self.trace._floats[v >> 1])
else:
return ConstInt(self.trace._bigints[v >> 1])
else:
assert False
def optimize_FLOAT_TRUEDIV(self, op):
arg1 = op.getarg(0)
arg2 = op.getarg(1)
v2 = self.getvalue(arg2)
# replace "x / const" by "x * (1/const)" if possible
if v2.is_constant():
divisor = v2.box.getfloat()
fraction = math.frexp(divisor)[0]
# This optimization is valid for powers of two
# but not for zeroes, some denormals and NaN:
if fraction == 0.5 or fraction == -0.5:
reciprocal = 1.0 / divisor
rfraction = math.frexp(reciprocal)[0]
if rfraction == 0.5 or rfraction == -0.5:
c = ConstFloat(longlong.getfloatstorage(reciprocal))
op = op.copy_and_change(rop.FLOAT_MUL, args=[arg1, c])
self.emit_operation(op)
def optimize_FLOAT_TRUEDIV(self, op):
arg1 = op.getarg(0)
arg2 = op.getarg(1)
v2 = get_box_replacement(arg2)
# replace "x / const" by "x * (1/const)" if possible
newop = op
if v2.is_constant():
divisor = v2.getfloat()
fraction = math.frexp(divisor)[0]
# This optimization is valid for powers of two
# but not for zeroes, some denormals and NaN:
if fraction == 0.5 or fraction == -0.5:
reciprocal = 1.0 / divisor
rfraction = math.frexp(reciprocal)[0]
if rfraction == 0.5 or rfraction == -0.5:
c = ConstFloat(longlong.getfloatstorage(reciprocal))
newop = self.replace_op_with(op, rop.FLOAT_MUL,
args=[arg1, c])
return self.emit(newop)
def accumulate_prepare(self, state):
vec_reg_size = state.vec_reg_size
for pack in self.packs:
if not pack.is_accumulating():
continue
if pack.leftmost().is_guard():
# guard breaks dependencies, thus it is an accumulation pack
continue
for i,node in enumerate(pack.operations):
op = node.getoperation()
state.accumulation[op] = pack
assert isinstance(pack, AccumPack)
datatype = pack.getdatatype()
bytesize = pack.getbytesize()
count = vec_reg_size // bytesize
signed = datatype == 'i'
oplist = state.invariant_oplist
# reset the box to zeros or ones
if pack.reduce_init() == 0:
vecop = OpHelpers.create_vec(datatype, bytesize, signed, count)
oplist.append(vecop)
vecop = VecOperation(rop.VEC_INT_XOR, [vecop, vecop],
vecop, count)
oplist.append(vecop)
elif pack.reduce_init() == 1:
# multiply is only supported by floats
vecop = OpHelpers.create_vec_expand(ConstFloat(1.0), bytesize,
signed, count)
oplist.append(vecop)
else:
raise NotImplementedError("cannot handle %s" % pack.operator)
# pack the scalar value
args = [vecop, pack.getleftmostseed(), ConstInt(0), ConstInt(1)]
vecop = OpHelpers.create_vec_pack(datatype, args, bytesize,
signed, count)
oplist.append(vecop)
seed = pack.getleftmostseed()
state.accumulation[seed] = pack
# rename the variable with the box
state.setvector_of_box(seed, 0, vecop) # prevent it from expansion
state.renamer.start_renaming(seed, vecop)
def test_wrap():
def _is(box1, box2):
return (box1.__class__ == box2.__class__ and
box1.value == box2.value)
p = lltype.malloc(lltype.GcStruct('S'))
po = lltype.cast_opaque_ptr(llmemory.GCREF, p)
assert _is(wrap(None, 42), BoxInt(42))
assert _is(wrap(None, 42.5), boxfloat(42.5))
assert _is(wrap(None, p), BoxPtr(po))
assert _is(wrap(None, 42, in_const_box=True), ConstInt(42))
assert _is(wrap(None, 42.5, in_const_box=True), constfloat(42.5))
assert _is(wrap(None, p, in_const_box=True), ConstPtr(po))
if longlong.supports_longlong:
import sys
from rpython.rlib.rarithmetic import r_longlong, r_ulonglong
value = r_longlong(-sys.maxint*17)
assert _is(wrap(None, value), BoxFloat(value))
assert _is(wrap(None, value, in_const_box=True), ConstFloat(value))
value_unsigned = r_ulonglong(-sys.maxint*17)
assert _is(wrap(None, value_unsigned), BoxFloat(value))
sfval = r_singlefloat(42.5)
ival = longlong.singlefloat2int(sfval)
assert _is(wrap(None, sfval), BoxInt(ival))
assert _is(wrap(None, sfval, in_const_box=True), ConstInt(ival))
def constfloat(x):
return ConstFloat(longlong.getfloatstorage(x))
def build_bridge(self):
def exc_handling(guard_op):
# operations need to start with correct GUARD_EXCEPTION
if guard_op._exc_box is None:
op = ResOperation(rop.GUARD_NO_EXCEPTION, [])
else:
op = ResOperation(rop.GUARD_EXCEPTION, [guard_op._exc_box])
op.setdescr(self.builder.getfaildescr())
op.setfailargs([])
return op
if self.dont_generate_more:
return False
r = self.r
guard_op = self.guard_op
fail_args = guard_op.getfailargs()
fail_descr = guard_op.getdescr()
op = self.should_fail_by
if not op.getfailargs():
return False
for _fail_box in fail_args:
_fail_box.set_forwarded(None)
# generate the branch: a sequence of operations that ends in a FINISH
subloop = DummyLoop([])
self.subloops.append(subloop) # keep around for debugging
if rop.is_guard_exception(guard_op.getopnum()):
subloop.operations.append(exc_handling(guard_op))
bridge_builder = self.builder.fork(self.builder.cpu, subloop,
op.getfailargs()[:])
self.generate_ops(bridge_builder, r, subloop, op.getfailargs()[:])
# note that 'self.guard_op' now points to the guard that will fail in
# this new bridge, while 'guard_op' still points to the guard that
# has just failed.
if r.random() < 0.1 and self.guard_op is None:
# Occasionally, instead of ending in a FINISH, we end in a jump
# to another loop. We don't do it, however, if the new bridge's
# execution will hit 'self.guard_op', but only if it executes
# to the FINISH normally. (There is no point to the extra
# complexity, as we might get the same effect by two calls
# to build_bridge().)
# First make up the other loop...
#
# New restriction: must have the same argument count and types
# as the original loop
subset = []
for box in self.loop.inputargs:
srcbox = r.choice(fail_args)
if srcbox.type != box.type:
if box.type == INT:
srcbox = ConstInt(r.random_integer())
elif box.type == FLOAT:
srcbox = ConstFloat(r.random_float_storage())
else:
raise AssertionError(box.type)
subset.append(srcbox)
#
args = []
for x in subset:
if x.type == INT:
args.append(InputArgInt(getint(x)))
elif x.type == FLOAT:
args.append(InputArgFloat(getfloatstorage(x)))
else:
assert 0, x.type
rl = RandomLoop(self.builder.cpu, self.builder.fork, r, args)
# done
self.should_fail_by = rl.should_fail_by
self.expected = rl.expected
assert len(rl.loop.inputargs) == len(args)
# The new bridge's execution will end normally at its FINISH.
# Just replace the FINISH with the JUMP to the new loop.
jump_op = ResOperation(rop.JUMP,
subset,
descr=rl.loop._targettoken)
subloop.operations[-1] = jump_op
self.guard_op = rl.guard_op
self.prebuilt_ptr_consts += rl.prebuilt_ptr_consts
self.loop._jitcelltoken.record_jump_to(rl.loop._jitcelltoken)
self.dont_generate_more = True
if r.random() < .05:
return False
dump(subloop)
self.builder.cpu.compile_bridge(fail_descr, fail_args,
subloop.operations,
self.loop._jitcelltoken)
if self.output:
bridge_builder.print_loop(self.output, fail_descr, fail_args)
return True
def produce_into(self, builder, r):
if r.random() < 0.4:
UnaryFloatOperation.produce_into(self, builder, r)
else:
self.put(builder, [ConstFloat(r.random_float_storage())])
