def transitive_imply(self, other, opt, loop):
if self.op.getopnum() != other.op.getopnum():
# stronger restriction, intermixing e.g. <= and < would be possible
return None
if self.getleftkey() is not other.getleftkey():
return None
if not self.rhs.is_identity():
# stronger restriction
return None
# this is a valid transitive guard that eliminates the loop guard
opnum = self.transitive_cmpop(self.cmp_op.getopnum())
box_rhs = self.emit_varops(opt, self.rhs, self.cmp_op.getarg(1))
other_rhs = self.emit_varops(opt, other.rhs, other.cmp_op.getarg(1))
compare = ResOperation(opnum, [box_rhs, other_rhs])
opt.emit_operation(compare)
# guard
descr = CompileLoopVersionDescr()
descr.copy_all_attributes_from(self.op.getdescr())
descr.rd_vector_info = None # do not copy the accum list
assert isinstance(descr, AbstractFailDescr)
guard = ResOperation(self.op.getopnum(), [compare], descr=descr)
guard.setfailargs(loop.label.getarglist_copy())
opt.emit_operation(guard)
return guard
def create_op(self, opnum, args, res, descr, fail_args):
res = ResOperation(opnum, args, descr)
if fail_args is not None:
res.setfailargs(fail_args)
if self._postproces:
self._postproces(res)
return res
def produce_into(self, builder, r):
fail_subset = builder.subset_of_intvars(r)
v_cond = builder.get_bool_var(r)
subset = builder.subset_of_intvars(r)[:4]
for i in range(len(subset)):
if r.random() < 0.35:
subset[i] = ConstInt(r.random_integer())
#
seen = []
def call_me(*args):
if len(seen) == 0:
seen.append(args)
else:
assert seen[0] == args
#
TP = lltype.FuncType([lltype.Signed] * len(subset), lltype.Void)
ptr = llhelper(lltype.Ptr(TP), call_me)
c_addr = ConstAddr(llmemory.cast_ptr_to_adr(ptr), builder.cpu)
args = [v_cond, c_addr] + subset
descr = self.getcalldescr(builder, TP)
self.put(builder, args, descr)
op = ResOperation(rop.GUARD_NO_EXCEPTION, [], None,
descr=builder.getfaildescr())
op.setfailargs(fail_subset)
builder.loop.operations.append(op)
def force_box(self, op, optforce):
if not self.is_virtual():
return op
optforce.forget_numberings()
if self.mode is mode_string:
s = self.get_constant_string_spec(optforce, mode_string)
if s is not None:
c_s = get_const_ptr_for_string(s)
optforce.get_box_replacement(op).set_forwarded(c_s)
return c_s
else:
s = self.get_constant_string_spec(optforce, mode_unicode)
if s is not None:
c_s = get_const_ptr_for_unicode(s)
optforce.get_box_replacement(op).set_forwarded(c_s)
return c_s
self._is_virtual = False
lengthbox = self.getstrlen(op, optforce.optimizer.optstring, self.mode)
newop = ResOperation(self.mode.NEWSTR, [lengthbox])
if not we_are_translated():
newop.name = 'FORCE'
optforce.emit_operation(newop)
newop = optforce.getlastop()
newop.set_forwarded(self)
op = optforce.get_box_replacement(op)
op.set_forwarded(newop)
optstring = optforce.optimizer.optstring
self.initialize_forced_string(op, optstring, op, CONST_0, self.mode)
return newop
def has_pure_result(self, opnum, args, descr):
op = ResOperation(opnum, args, None, descr)
key = self.optimizer.make_args_key(op)
op = self.pure_operations.get(key, None)
if op is None:
return False
return op.getdescr() is descr
def store_final_boxes_in_guard(self, op, pendingfields):
assert pendingfields is not None
descr = op.getdescr()
assert isinstance(descr, compile.ResumeGuardDescr)
modifier = resume.ResumeDataVirtualAdder(descr, self.resumedata_memo)
try:
newboxes = modifier.finish(self, pendingfields)
if len(newboxes) > self.metainterp_sd.options.failargs_limit:
raise resume.TagOverflow
except resume.TagOverflow:
raise compile.giveup()
descr.store_final_boxes(op, newboxes)
#
if op.getopnum() == rop.GUARD_VALUE:
if self.getvalue(op.getarg(0)) in self.bool_boxes:
# Hack: turn guard_value(bool) into guard_true/guard_false.
# This is done after the operation is emitted to let
# store_final_boxes_in_guard set the guard_opnum field of the
# descr to the original rop.GUARD_VALUE.
constvalue = op.getarg(1).getint()
if constvalue == 0:
opnum = rop.GUARD_FALSE
elif constvalue == 1:
opnum = rop.GUARD_TRUE
else:
raise AssertionError("uh?")
newop = ResOperation(opnum, [op.getarg(0)], op.result, descr)
newop.setfailargs(op.getfailargs())
return newop
else:
# a real GUARD_VALUE. Make it use one counter per value.
descr.make_a_counter_per_value(op)
return op
def handle_malloc_operation(self, op):
opnum = op.getopnum()
if opnum == rop.NEW:
self.handle_new_fixedsize(op.getdescr(), op)
elif opnum == rop.NEW_WITH_VTABLE:
classint = op.getarg(0).getint()
descr = heaptracker.vtable2descr(self.cpu, classint)
self.handle_new_fixedsize(descr, op)
if self.gc_ll_descr.fielddescr_vtable is not None:
op = ResOperation(rop.SETFIELD_GC,
[op.result, ConstInt(classint)], None,
descr=self.gc_ll_descr.fielddescr_vtable)
self.newops.append(op)
elif opnum == rop.NEW_ARRAY or opnum == rop.NEW_ARRAY_CLEAR:
descr = op.getdescr()
assert isinstance(descr, ArrayDescr)
self.handle_new_array(descr, op)
elif opnum == rop.NEWSTR:
self.handle_new_array(self.gc_ll_descr.str_descr, op,
kind=FLAG_STR)
elif opnum == rop.NEWUNICODE:
self.handle_new_array(self.gc_ll_descr.unicode_descr, op,
kind=FLAG_UNICODE)
else:
raise NotImplementedError(op.getopname())
def next(self):
opnum = self._next()
if oparity[opnum] == -1:
argnum = self._next()
else:
argnum = oparity[opnum]
args = []
for i in range(argnum):
args.append(self._untag(self._next()))
descr_index = -1
if opwithdescr[opnum]:
descr_index = self._next()
if descr_index == 0 or rop.is_guard(opnum):
descr = None
else:
if descr_index < self.all_descr_len + 1:
descr = self.metainterp_sd.all_descrs[descr_index - 1]
else:
descr = self.trace._descrs[descr_index - self.all_descr_len - 1]
else:
descr = None
res = ResOperation(opnum, args, descr=descr)
if rop.is_guard(opnum):
assert isinstance(res, GuardResOp)
res.rd_resume_position = descr_index
if res.type != 'v':
self._cache[self._index] = res
self._index += 1
self._count += 1
return res
def do(self, opnum, argboxes, descr=None):
self.fakemetainterp._got_exc = None
op = ResOperation(opnum, argboxes, descr)
result = _execute_arglist(self.cpu, self.fakemetainterp,
opnum, argboxes, descr)
if result is not None:
c_result = wrap_constant(result)
op.copy_value_from(c_result)
self.loop.operations.append(op)
return op
def gen_guard_true(draw, all_ops, framestack):
arg = get_arg(draw, all_ops, allow_const=False)
res = ResOperation(rop.GUARD_TRUE, [arg])
if draw(strategies.booleans()):
s = []
for i in range(10):
s.append(get_arg(draw, all_ops, allow_const=False))
boxes = list(set(s))
framestack.append(Frame(JitCode(1), 2, boxes))
res.framestack = framestack[:]
return res
def getstrlen(self, op, string_optimizer, mode, create_ops=True):
if self.lgtop is not None:
return self.lgtop
assert not self.is_virtual()
if not create_ops:
return None
lengthop = ResOperation(mode.STRLEN, [op])
lengthop.set_forwarded(self.getlenbound(mode))
self.lgtop = lengthop
string_optimizer.emit_operation(lengthop)
return lengthop
def optimize_NEW_ARRAY(self, op):
sizebox = self.get_constant_box(op.getarg(0))
if sizebox is not None:
# if the original 'op' did not have a ConstInt as argument,
# build a new one with the ConstInt argument
if not isinstance(op.getarg(0), ConstInt):
op = ResOperation(rop.NEW_ARRAY, [sizebox], op.result,
descr=op.getdescr())
self.make_varray(op.getdescr(), sizebox.getint(), op.result, op)
else:
self.getvalue(op.result).ensure_nonnull()
self.emit_operation(op)
def _optimize_NEWSTR(self, op, mode):
length_box = self.get_constant_box(op.getarg(0))
if length_box and length_box.getint() <= MAX_CONST_LEN:
# if the original 'op' did not have a ConstInt as argument,
# build a new one with the ConstInt argument
if not isinstance(op.getarg(0), ConstInt):
op = ResOperation(mode.NEWSTR, [length_box], op.result)
vvalue = self.make_vstring_plain(op.result, op, mode)
vvalue.setup(length_box.getint())
else:
self.getvalue(op.result).ensure_nonnull()
self.emit_operation(op)
self.pure(mode.STRLEN, [op.result], op.getarg(0))
def produce_into(self, builder, r):
fail_subset = builder.subset_of_intvars(r)
subset, f = self.non_raising_func_code(builder, r)
RES = self.getresulttype()
TP = lltype.FuncType([lltype.Signed] * len(subset), RES)
ptr = llhelper(lltype.Ptr(TP), f)
c_addr = ConstAddr(llmemory.cast_ptr_to_adr(ptr), builder.cpu)
args = [c_addr] + subset
descr = self.getcalldescr(builder, TP)
self.put(builder, args, descr)
op = ResOperation(rop.GUARD_NO_EXCEPTION, [],
descr=builder.getfaildescr())
op.setfailargs(fail_subset)
builder.loop.operations.append(op)
def emit_operations(self, opt):
# create trace instructions for the index
lhs = self.emit_varops(opt, self.lhs, self.cmp_op.getarg(0))
rhs = self.emit_varops(opt, self.rhs, self.cmp_op.getarg(1))
opnum = self.cmp_op.getopnum()
cmp_op = ResOperation(opnum, [lhs, rhs])
opt.emit_operation(cmp_op)
# emit that actual guard
guard = ResOperation(self.op.getopnum(), [cmp_op], self.op.getdescr())
guard.setfailargs(self.op.getfailargs()[:])
opt.emit_operation(guard)
self.setindex(opt.operation_position()-1)
self.setoperation(guard)
self.setcmp(cmp_op)
def produce_into(self, builder, r):
fail_subset = builder.subset_of_intvars(r)
subset, f, exc = self.raising_func_code(builder, r)
TP = lltype.FuncType([lltype.Signed] * len(subset), lltype.Void)
ptr = llhelper(lltype.Ptr(TP), f)
c_addr = ConstAddr(llmemory.cast_ptr_to_adr(ptr), builder.cpu)
args = [c_addr] + subset
descr = self.getcalldescr(builder, TP)
self.put(builder, args, descr)
exc_box = ConstAddr(llmemory.cast_ptr_to_adr(exc), builder.cpu)
op = ResOperation(rop.GUARD_EXCEPTION, [exc_box], BoxPtr(),
descr=builder.getfaildescr())
op.setfailargs(fail_subset)
builder.loop.operations.append(op)
def gen_initialize_tid(self, v_newgcobj, tid):
if self.gc_ll_descr.fielddescr_tid is not None:
# produce a SETFIELD to initialize the GC header
op = ResOperation(rop.SETFIELD_GC,
[v_newgcobj, ConstInt(tid)],
None,
descr=self.gc_ll_descr.fielddescr_tid)
self.newops.append(op)
def emit_load_effective_address(self, v_gcptr, v_index, base, itemscale):
if self.cpu.supports_load_effective_address:
i1 = ResOperation(rop.LOAD_EFFECTIVE_ADDRESS,
[v_gcptr, v_index, ConstInt(base),
ConstInt(itemscale)])
self.emit_op(i1)
return i1
else:
if itemscale > 0:
v_index = ResOperation(rop.INT_LSHIFT,
[v_index, ConstInt(itemscale)])
self.emit_op(v_index)
i1b = ResOperation(rop.INT_ADD, [v_gcptr, v_index])
self.emit_op(i1b)
i1 = ResOperation(rop.INT_ADD, [i1b, ConstInt(base)])
self.emit_op(i1)
return i1
def handle_str_equal_level1(self, v1, v2, resultbox, mode):
l2box = v2.getstrlen(None, mode, None)
if isinstance(l2box, ConstInt):
if l2box.value == 0:
lengthbox = v1.getstrlen(self, mode, None)
seo = self.optimizer.send_extra_operation
seo(ResOperation(rop.INT_EQ, [lengthbox, CONST_0], resultbox))
return True
if l2box.value == 1:
l1box = v1.getstrlen(None, mode, None)
if isinstance(l1box, ConstInt) and l1box.value == 1:
# comparing two single chars
vchar1 = self.strgetitem(v1, optimizer.CVAL_ZERO, mode)
vchar2 = self.strgetitem(v2, optimizer.CVAL_ZERO, mode)
seo = self.optimizer.send_extra_operation
seo(
ResOperation(
rop.INT_EQ,
[vchar1.force_box(self),
vchar2.force_box(self)], resultbox))
return True
if isinstance(v1, VStringSliceValue):
vchar = self.strgetitem(v2, optimizer.CVAL_ZERO, mode)
do = EffectInfo.OS_STREQ_SLICE_CHAR
self.generate_modified_call(do, [
v1.vstr.force_box(self),
v1.vstart.force_box(self),
v1.vlength.force_box(self),
vchar.force_box(self)
], resultbox, mode)
return True
#
if v2.is_null():
if v1.is_nonnull():
self.make_constant(resultbox, CONST_0)
return True
if v1.is_null():
self.make_constant(resultbox, CONST_1)
return True
op = ResOperation(rop.PTR_EQ,
[v1.force_box(self), llhelper.CONST_NULL],
resultbox)
self.emit_operation(op)
return True
#
return False
def _force_elements(self, op, optforce, descr):
self.size = -1
# at this point we have just written the
# 'op = CALL_I(..., OS_RAW_MALLOC_VARSIZE_CHAR)'.
# Emit now a CHECK_MEMORY_ERROR resop.
check_op = ResOperation(rop.CHECK_MEMORY_ERROR, [op])
optforce.emit_extra(check_op)
#
buffer = self._get_buffer()
for i in range(len(buffer.offsets)):
# write the value
offset = buffer.offsets[i]
descr = buffer.descrs[i]
itembox = buffer.values[i]
setfield_op = ResOperation(rop.RAW_STORE,
[op, ConstInt(offset), itembox], descr=descr)
optforce.emit_extra(setfield_op)
def generate_modified_call(self, oopspecindex, args, result, mode):
oopspecindex += mode.OS_offset
cic = self.optimizer.metainterp_sd.callinfocollection
calldescr, func = cic.callinfo_for_oopspec(oopspecindex)
op = ResOperation(rop.CALL, [ConstInt(func)] + args,
result,
descr=calldescr)
self.emit_operation(op)
def compare_short(self, short, expected_short):
expected_short = self.parse(expected_short)
remap = {}
exp = ([ResOperation(rop.LABEL, expected_short.inputargs)] +
expected_short.operations)
for k, v in zip(short[0].getarglist(), expected_short.inputargs):
remap[v] = k
equaloplists(short, exp, remap=remap)
def emit_gc_store_or_indexed(self, op, ptr_box, index_box, value_box,
itemsize, factor, offset):
factor, offset, index_box = \
self._emit_mul_if_factor_offset_not_supported(index_box,
factor, offset)
#
if index_box is None:
args = [ptr_box, ConstInt(offset), value_box, ConstInt(itemsize)]
newload = ResOperation(rop.GC_STORE, args)
else:
args = [ptr_box, index_box, value_box, ConstInt(factor),
ConstInt(offset), ConstInt(itemsize)]
newload = ResOperation(rop.GC_STORE_INDEXED, args)
if op is not None:
self.replace_op_with(op, newload)
else:
self.emit_op(newload)
def gen_guard(self, builder, r):
if not builder.ptrvars:
raise CannotProduceOperation
box = r.choice(builder.ptrvars)[0]
op = ResOperation(self.opnum, [box])
passing = ((self.opnum == rop.GUARD_NONNULL and getref_base(box)) or
(self.opnum == rop.GUARD_ISNULL and not getref_base(box)))
return op, passing
def produce_into(self, builder, r):
fail_subset = builder.subset_of_intvars(r)
subset, f = self.non_raising_func_code(builder, r)
if len(subset) == 0:
RES = lltype.Void
else:
RES = lltype.Signed
TP = lltype.FuncType([lltype.Signed] * len(subset), RES)
ptr = llhelper(lltype.Ptr(TP), f)
c_addr = ConstAddr(llmemory.cast_ptr_to_adr(ptr), builder.cpu)
args = [c_addr] + subset
descr = self.getcalldescr(builder, TP)
self.put(builder, args, descr)
op = ResOperation(rop.GUARD_NO_EXCEPTION, [], None,
descr=builder.getfaildescr())
op.setfailargs(fail_subset)
builder.loop.operations.append(op)
def make_guards(self, box, guards, optimizer):
if self.is_constant():
guards.append(ResOperation(rop.GUARD_VALUE,
[box, ConstInt(self.upper)]))
return
if self.has_lower and self.lower > MININT:
bound = self.lower
op = ResOperation(rop.INT_GE, [box, ConstInt(bound)])
guards.append(op)
op = ResOperation(rop.GUARD_TRUE, [op])
guards.append(op)
if self.has_upper and self.upper < MAXINT:
bound = self.upper
op = ResOperation(rop.INT_LE, [box, ConstInt(bound)])
guards.append(op)
op = ResOperation(rop.GUARD_TRUE, [op])
guards.append(op)
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, [], None)
else:
op = ResOperation(rop.GUARD_EXCEPTION, [guard_op._exc_box],
BoxPtr())
op.setdescr(self.builder.getfaildescr())
op.setfailargs([])
return op
def import_state(self, targetop, exported_state):
if not targetop: # Trace did not start with a label
self.inputargs = self.optimizer.loop.inputargs
self.short = None
self.initial_virtual_state = None
return
self.inputargs = targetop.getarglist()
target_token = targetop.getdescr()
assert isinstance(target_token, TargetToken)
if not exported_state:
# No state exported, construct one without virtuals
self.short = None
virtual_state = self.get_virtual_state(self.inputargs)
self.initial_virtual_state = virtual_state
return
self.short = target_token.short_preamble[:]
self.short_seen = {}
self.short_boxes = exported_state.short_boxes
self.initial_virtual_state = target_token.virtual_state
for box in self.inputargs:
preamble_value = exported_state.exported_values[box]
value = self.optimizer.getvalue(box)
value.import_from(preamble_value, self.optimizer)
# Setup the state of the new optimizer by emiting the
# short operations and discarding the result
self.optimizer.emitting_dissabled = True
for op in exported_state.inputarg_setup_ops:
self.optimizer.send_extra_operation(op)
seen = {}
for op in self.short_boxes.operations():
self.ensure_short_op_emitted(op, self.optimizer, seen)
if op and op.result:
preamble_value = exported_state.exported_values[op.result]
value = self.optimizer.getvalue(op.result)
if not value.is_virtual() and not value.is_constant():
imp = ValueImporter(self, preamble_value, op)
self.optimizer.importable_values[value] = imp
newvalue = self.optimizer.getvalue(op.result)
newresult = newvalue.get_key_box()
# note that emitting here SAME_AS should not happen, but
# in case it does, we would prefer to be suboptimal in asm
# to a fatal RPython exception.
if newresult is not op.result and \
not self.short_boxes.has_producer(newresult) and \
not newvalue.is_constant():
op = ResOperation(rop.SAME_AS, [op.result], newresult)
self.optimizer._newoperations.append(op)
#if self.optimizer.loop.logops:
# debug_print(' Falling back to add extra: ' +
# self.optimizer.loop.logops.repr_of_resop(op))
self.optimizer.flush()
self.optimizer.emitting_dissabled = False
def compile_tmp_callback(cpu, jitdriver_sd, greenboxes, redargtypes,
memory_manager=None):
"""Make a LoopToken that corresponds to assembler code that just
calls back the interpreter. Used temporarily: a fully compiled
version of the code may end up replacing it.
"""
jitcell_token = make_jitcell_token(jitdriver_sd)
nb_red_args = jitdriver_sd.num_red_args
assert len(redargtypes) == nb_red_args
inputargs = []
for kind in redargtypes:
if kind == history.INT:
box = InputArgInt()
elif kind == history.REF:
box = InputArgRef()
elif kind == history.FLOAT:
box = InputArgFloat()
else:
raise AssertionError
inputargs.append(box)
k = jitdriver_sd.portal_runner_adr
funcbox = history.ConstInt(heaptracker.adr2int(k))
callargs = [funcbox] + greenboxes + inputargs
#
jd = jitdriver_sd
opnum = OpHelpers.call_for_descr(jd.portal_calldescr)
call_op = ResOperation(opnum, callargs, descr=jd.portal_calldescr)
if call_op.type != 'v' is not None:
finishargs = [call_op]
else:
finishargs = []
#
faildescr = jitdriver_sd.propagate_exc_descr
operations = [
call_op,
ResOperation(rop.GUARD_NO_EXCEPTION, [], descr=faildescr),
ResOperation(rop.FINISH, finishargs, descr=jd.portal_finishtoken)
]
operations[1].setfailargs([])
operations = get_deep_immutable_oplist(operations)
cpu.compile_loop(inputargs, operations, jitcell_token, log=False)
if memory_manager is not None: # for tests
memory_manager.keep_loop_alive(jitcell_token)
return jitcell_token
def gen_write_barrier(self, v_base):
write_barrier_descr = self.gc_ll_descr.write_barrier_descr
args = [v_base]
self.newops.append(
ResOperation(rop.COND_CALL_GC_WB,
args,
None,
descr=write_barrier_descr))
self.write_barrier_applied[v_base] = None
def _int_sub(string_optimizer, box1, box2):
if isinstance(box2, ConstInt):
if box2.value == 0:
return box1
if isinstance(box1, ConstInt):
return ConstInt(box1.value - box2.value)
resbox = BoxInt()
string_optimizer.emit_operation(ResOperation(rop.INT_SUB, [box1, box2], resbox))
return resbox
def gen_malloc_nursery_varsize_frame(self, sizebox):
""" Generate CALL_MALLOC_NURSERY_VARSIZE_FRAME
"""
self.emitting_an_operation_that_can_collect()
op = ResOperation(rop.CALL_MALLOC_NURSERY_VARSIZE_FRAME, [sizebox])
self.emit_op(op)
self.remember_write_barrier(op)
return op
def do(self, opnum, argboxes, descr=None):
self.fakemetainterp._got_exc = None
op = ResOperation(opnum, argboxes, descr)
if opnum != rop.ZERO_PTR_FIELD:
result = _execute_arglist(self.cpu, self.fakemetainterp,
opnum, argboxes, descr)
if result is not None:
c_result = wrap_constant(result)
op.copy_value_from(c_result)
else:
import ctypes
addr = self.cpu.cast_gcref_to_int(argboxes[0].getref_base())
offset = argboxes[1].getint()
assert (offset % ctypes.sizeof(ctypes.c_long)) == 0
ptr = ctypes.cast(addr, ctypes.POINTER(ctypes.c_long))
ptr[offset / ctypes.sizeof(ctypes.c_long)] = 0
self.loop.operations.append(op)
return op
def create_op(self, opnum, args, result, descr, fail_args):
if opnum == ESCAPE_OP.OPNUM:
op = ESCAPE_OP(result)
op.initarglist(args)
assert descr is None
return op
if opnum == FORCE_SPILL.OPNUM:
op = FORCE_SPILL(result)
op.initarglist(args)
assert descr is None
return op
else:
res = ResOperation(opnum, args, result, descr)
if fail_args is not None:
res.setfailargs(fail_args)
if self._postproces:
self._postproces(res)
return res
def produce_into(self, builder, r):
subset, f = self.non_raising_func_code(builder, r)
RES = self.getresulttype()
TP = lltype.FuncType([lltype.Signed] * len(subset), RES)
ptr = llhelper(lltype.Ptr(TP), f)
c_addr = ConstAddr(llmemory.cast_ptr_to_adr(ptr), builder.cpu)
args = [c_addr] + subset
descr = self.getcalldescr(builder, TP)
self.put(builder, args, descr)
_, vtableptr = builder.get_random_structure_type_and_vtable(r)
exc_box = ConstAddr(llmemory.cast_ptr_to_adr(vtableptr), builder.cpu)
op = ResOperation(rop.GUARD_EXCEPTION, [exc_box],
descr=builder.getfaildescr())
op.setfailargs(builder.subset_of_intvars(r))
op._exc_box = None
builder.should_fail_by = op
builder.guard_op = op
builder.loop.operations.append(op)
def create_op(self, opnum, args, result, descr, fail_args):
if opnum == ESCAPE_OP.OPNUM:
op = ESCAPE_OP(result)
op.initarglist(args)
assert descr is None
return op
if opnum == FORCE_SPILL.OPNUM:
op = FORCE_SPILL(result)
op.initarglist(args)
assert descr is None
return op
else:
res = ResOperation(opnum, args, result, descr)
if fail_args is not None:
res.setfailargs(fail_args)
if self._postproces:
self._postproces(res)
return res
def cpu_simplify_scale(cpu, index_box, factor, offset):
# Returns (factor, offset, index_box, [ops]) where index_box is either
# a non-constant BoxInt or None.
if isinstance(index_box, ConstInt):
return 1, index_box.value * factor + offset, None, False
else:
if factor != 1 and factor not in cpu.load_supported_factors:
# the factor is supported by the cpu
# x & (x - 1) == 0 is a quick test for power of 2
assert factor > 0
if (factor & (factor - 1)) == 0:
index_box = ResOperation(rop.INT_LSHIFT,
[index_box, ConstInt(highest_bit(factor))])
else:
index_box = ResOperation(rop.INT_MUL,
[index_box, ConstInt(factor)])
return 1, offset, index_box, True
return factor, offset, index_box, False
def gen_malloc_nursery_varsize_frame(self, sizebox, v_result):
""" Generate CALL_MALLOC_NURSERY_VARSIZE_FRAME
"""
self.emitting_an_operation_that_can_collect()
op = ResOperation(rop.CALL_MALLOC_NURSERY_VARSIZE_FRAME, [sizebox],
v_result)
self.newops.append(op)
self.write_barrier_applied[v_result] = None
def _int_sub(string_optimizer, box1, box2):
if isinstance(box2, ConstInt):
if box2.value == 0:
return box1
if isinstance(box1, ConstInt):
return ConstInt(box1.value - box2.value)
op = ResOperation(rop.INT_SUB, [box1, box2])
string_optimizer.send_extra_operation(op)
return op
def produce_into(self, builder, r):
subset, f = self.non_raising_func_code(builder, r)
RES = self.getresulttype()
TP = lltype.FuncType([lltype.Signed] * len(subset), RES)
ptr = llhelper(lltype.Ptr(TP), f)
c_addr = ConstAddr(llmemory.cast_ptr_to_adr(ptr), builder.cpu)
args = [c_addr] + subset
descr = self.getcalldescr(builder, TP)
self.put(builder, args, descr)
_, vtableptr = builder.get_random_structure_type_and_vtable(r)
exc_box = ConstAddr(llmemory.cast_ptr_to_adr(vtableptr), builder.cpu)
op = ResOperation(rop.GUARD_EXCEPTION, [exc_box],
descr=builder.getfaildescr())
op.setfailargs(builder.subset_of_intvars(r))
op._exc_box = None
builder.should_fail_by = op
builder.guard_op = op
builder.loop.operations.append(op)
def get_operations(self):
var = self.var
tolist = []
if self.coefficient_mul != 1:
args = [var, ConstInt(self.coefficient_mul)]
var = ResOperation(rop.INT_MUL, args)
tolist.append(var)
if self.coefficient_div != 1:
assert 0 # should never be the case with handling
# of INT_PY_DIV commented out in this file...
if self.constant > 0:
args = [var, ConstInt(self.constant)]
var = ResOperation(rop.INT_ADD, args)
tolist.append(var)
if self.constant < 0:
args = [var, ConstInt(-self.constant)]
var = ResOperation(rop.INT_SUB, args)
tolist.append(var)
return tolist
def optimize_VIRTUAL_REF(self, op):
# get some constants
vrefinfo = self.optimizer.metainterp_sd.virtualref_info
c_cls = vrefinfo.jit_virtual_ref_const_class
vref_descr = vrefinfo.descr
descr_virtual_token = vrefinfo.descr_virtual_token
descr_forced = vrefinfo.descr_forced
#
# Replace the VIRTUAL_REF operation with a virtual structure of type
# 'jit_virtual_ref'. The jit_virtual_ref structure may be forced soon,
# but the point is that doing so does not force the original structure.
newop = ResOperation(rop.NEW_WITH_VTABLE, [], descr=vref_descr)
vrefvalue = self.make_virtual(c_cls, newop, vref_descr)
op.set_forwarded(newop)
newop.set_forwarded(vrefvalue)
token = ResOperation(rop.FORCE_TOKEN, [])
vrefvalue.setfield(descr_virtual_token, newop, token)
vrefvalue.setfield(descr_forced, newop, CONST_NULL)
return self.emit(token)
def convert_old_style_to_targets(loop, jump):
newloop = TreeLoop(loop.name)
newloop.inputargs = loop.inputargs
newloop.operations = [ResOperation(rop.LABEL, loop.inputargs, descr=FakeDescr())] + \
loop.operations
if not jump:
assert newloop.operations[-1].getopnum() == rop.JUMP
newloop.operations[-1] = newloop.operations[-1].copy_and_change(
rop.LABEL)
return newloop
def gen_malloc_frame(self, frame_info):
descrs = self.gc_ll_descr.getframedescrs(self.cpu)
if self.gc_ll_descr.kind == 'boehm':
ofs, size, sign = unpack_fielddescr(descrs.jfi_frame_depth)
if sign:
size = -size
args = [ConstInt(frame_info), ConstInt(ofs), ConstInt(size)]
size = ResOperation(rop.GC_LOAD_I, args)
self.emit_op(size)
frame = ResOperation(rop.NEW_ARRAY, [size],
descr=descrs.arraydescr)
self.handle_new_array(descrs.arraydescr, frame)
return self.get_box_replacement(frame)
else:
# we read size in bytes here, not the length
ofs, size, sign = unpack_fielddescr(descrs.jfi_frame_size)
if sign:
size = -size
args = [ConstInt(frame_info), ConstInt(ofs), ConstInt(size)]
size = ResOperation(rop.GC_LOAD_I, args)
self.emit_op(size)
frame = self.gen_malloc_nursery_varsize_frame(size)
self.gen_initialize_tid(frame, descrs.arraydescr.tid)
# we need to explicitely zero all the gc fields, because
# of the unusal malloc pattern
length = self.emit_getfield(ConstInt(frame_info),
descr=descrs.jfi_frame_depth, raw=True)
self.emit_setfield(frame, self.c_zero,
descr=descrs.jf_extra_stack_depth)
self.emit_setfield(frame, self.c_null,
descr=descrs.jf_savedata)
self.emit_setfield(frame, self.c_null,
descr=descrs.jf_force_descr)
self.emit_setfield(frame, self.c_null,
descr=descrs.jf_descr)
self.emit_setfield(frame, self.c_null,
descr=descrs.jf_guard_exc)
self.emit_setfield(frame, self.c_null,
descr=descrs.jf_forward)
self.gen_initialize_len(frame, length,
descrs.arraydescr.lendescr)
return self.get_box_replacement(frame)
def optimize(self,
ops,
bridge_ops,
expected,
expected_loop=None,
inline_short_preamble=True,
jump_values=None,
bridge_values=None):
loop = self.parse(ops)
info = self.unroll_and_optimize(loop, None, jump_values=jump_values)
jitcell_token = compile.make_jitcell_token(None)
mid_label_descr = TargetToken(jitcell_token)
mid_label_descr.short_preamble = info.short_preamble
mid_label_descr.virtual_state = info.virtual_state
start_label_descr = TargetToken(jitcell_token)
jitcell_token.target_tokens = [mid_label_descr, start_label_descr]
loop.operations[0].setdescr(mid_label_descr)
loop.operations[-1].setdescr(mid_label_descr)
info.preamble.operations[0].setdescr(start_label_descr)
guards = [op for op in loop.operations if op.is_guard()]
assert len(guards) == 1, "more than one guard in the loop"
bridge = self.parse(bridge_ops)
bridge.operations[-1].setdescr(jitcell_token)
self.add_guard_future_condition(bridge)
trace = oparser.convert_loop_to_trace(
bridge, FakeMetaInterpStaticData(self.cpu))
data = compile.BridgeCompileData(
trace,
self.convert_values(bridge.operations[-1].getarglist(),
bridge_values),
None,
enable_opts=self.enable_opts,
inline_short_preamble=inline_short_preamble)
bridge_info, ops = self._do_optimize_loop(data)
loop.check_consistency(check_descr=False)
info.preamble.check_consistency(check_descr=False)
bridge.operations = ([ResOperation(rop.LABEL, bridge_info.inputargs)] +
ops)
bridge.inputargs = bridge_info.inputargs
bridge.check_consistency(check_descr=False)
expected = self.parse(expected)
self.assert_equal(bridge,
convert_old_style_to_targets(expected, jump=True))
jump_bridge = bridge.operations[-1]
jump_d = jump_bridge.getdescr()
jump_args = jump_bridge.getarglist()
if loop.operations[0].getdescr() is jump_d:
# jump to loop
label_args = loop.operations[0].getarglist()
else:
assert info.preamble.operations[0].getdescr() is jump_d
label_args = info.preamble.operations[0].getarglist()
assert len(jump_args) == len(label_args)
for a, b in zip(jump_args, label_args):
assert a.type == b.type
def produce_into(self, builder, r):
subset, f, exc = self.raising_func_code(builder, r)
TP = lltype.FuncType([lltype.Signed] * len(subset), lltype.Void)
ptr = llhelper(lltype.Ptr(TP), f)
c_addr = ConstAddr(llmemory.cast_ptr_to_adr(ptr), builder.cpu)
args = [c_addr] + subset
descr = self.getcalldescr(builder, TP)
self.put(builder, args, descr)
while True:
_, vtableptr = builder.get_random_structure_type_and_vtable(r)
if vtableptr != exc:
break
other_box = ConstAddr(llmemory.cast_ptr_to_adr(vtableptr), builder.cpu)
op = ResOperation(rop.GUARD_EXCEPTION, [other_box], BoxPtr(),
descr=builder.getfaildescr())
op._exc_box = ConstAddr(llmemory.cast_ptr_to_adr(exc), builder.cpu)
op.setfailargs(builder.subset_of_intvars(r))
builder.should_fail_by = op
builder.guard_op = op
builder.loop.operations.append(op)
def _emit_mul_if_factor_offset_not_supported(self, index_box,
factor, offset):
# Returns (factor, offset, index_box) where index_box is either
# a non-constant BoxInt or None.
if isinstance(index_box, ConstInt):
return 1, index_box.value * factor + offset, None
else:
if factor != 1 and factor not in self.cpu.load_supported_factors:
# the factor is supported by the cpu
# x & (x - 1) == 0 is a quick test for power of 2
assert factor > 0
if (factor & (factor - 1)) == 0:
index_box = ResOperation(rop.INT_LSHIFT,
[index_box, ConstInt(highest_bit(factor))])
else:
index_box = ResOperation(rop.INT_MUL,
[index_box, ConstInt(factor)])
self.emit_op(index_box)
factor = 1
return factor, offset, index_box
def _int_add(optstring, box1, box2):
if isinstance(box1, ConstInt):
if box1.value == 0:
return box2
if isinstance(box2, ConstInt):
return ConstInt(box1.value + box2.value)
elif isinstance(box2, ConstInt) and box2.value == 0:
return box1
op = ResOperation(rop.INT_ADD, [box1, box2])
optstring.optimizer.send_extra_operation(op)
return op
def _force_elements(self, op, optforce, descr):
if self._fields is None:
return
for i, fielddescr in enumerate(descr.get_all_fielddescrs()):
fld = self._fields[i]
if fld is not None:
subbox = optforce.force_box(fld)
setfieldop = ResOperation(rop.SETFIELD_GC, [op, subbox],
descr=fielddescr)
self._fields[i] = None
optforce.emit_operation(setfieldop)
def optimize_VIRTUAL_REF(self, op):
# get some constants
vrefinfo = self.optimizer.metainterp_sd.virtualref_info
c_cls = vrefinfo.jit_virtual_ref_const_class
vref_descr = vrefinfo.descr
descr_virtual_token = vrefinfo.descr_virtual_token
descr_forced = vrefinfo.descr_forced
#
# Replace the VIRTUAL_REF operation with a virtual structure of type
# 'jit_virtual_ref'. The jit_virtual_ref structure may be forced soon,
# but the point is that doing so does not force the original structure.
newop = ResOperation(rop.NEW_WITH_VTABLE, [], descr=vref_descr)
vrefvalue = self.make_virtual(c_cls, newop, vref_descr)
op.set_forwarded(newop)
newop.set_forwarded(vrefvalue)
token = ResOperation(rop.FORCE_TOKEN, [])
self.emit_operation(token)
vrefvalue.setfield(descr_virtual_token, newop, token)
vrefvalue.setfield(descr_forced, newop,
self.optimizer.cpu.ts.CONST_NULLREF)
def produce_into(self, builder, r):
subset, f, exc = self.raising_func_code(builder, r)
TP = lltype.FuncType([lltype.Signed] * len(subset), lltype.Void)
ptr = llhelper(lltype.Ptr(TP), f)
c_addr = ConstAddr(llmemory.cast_ptr_to_adr(ptr), builder.cpu)
args = [c_addr] + subset
descr = self.getcalldescr(builder, TP)
self.put(builder, args, descr)
while True:
_, vtableptr = builder.get_random_structure_type_and_vtable(r)
if vtableptr != exc:
break
other_box = ConstAddr(llmemory.cast_ptr_to_adr(vtableptr), builder.cpu)
op = ResOperation(rop.GUARD_EXCEPTION, [other_box],
descr=builder.getfaildescr())
op._exc_box = ConstAddr(llmemory.cast_ptr_to_adr(exc), builder.cpu)
op.setfailargs(builder.subset_of_intvars(r))
builder.should_fail_by = op
builder.guard_op = op
builder.loop.operations.append(op)
def _really_force(self, optforce):
if self.mode is mode_string:
s = self.get_constant_string_spec(mode_string)
if s is not None:
c_s = get_const_ptr_for_string(s)
self.make_constant(c_s)
return
else:
s = self.get_constant_string_spec(mode_unicode)
if s is not None:
c_s = get_const_ptr_for_unicode(s)
self.make_constant(c_s)
return
assert self.source_op is not None
self.box = box = self.source_op.result
lengthbox = self.getstrlen(optforce, self.mode, None)
op = ResOperation(self.mode.NEWSTR, [lengthbox], box)
if not we_are_translated():
op.name = 'FORCE'
optforce.emit_operation(op)
self.initialize_forced_string(optforce, box, CONST_0, self.mode)
def parse_loop(self, ops, add_label=True):
loop = self.parse(ops, postprocess=self.postprocess)
loop.operations = filter(lambda op: op.getopnum() != rop.DEBUG_MERGE_POINT, loop.operations)
token = JitCellToken()
if add_label:
label = ResOperation(rop.LABEL, loop.inputargs, descr=TargetToken(token))
else:
label = loop.operations[0]
label.setdescr(TargetToken(token))
jump = loop.operations[-1]
loop = VectorLoop(label, loop.operations[0:-1], jump)
loop.jump.setdescr(token)
class Optimizer(object):
metainterp_sd = FakeMetaInterpStaticData(self.cpu)
jitdriver_sd = FakeJitDriverStaticData()
opt = Optimizer()
opt.jitdriver_sd.vec = True
for op in loop.operations:
if op.is_guard() and not op.getdescr():
descr = invent_fail_descr_for_op(op.getopnum(), opt)
op.setdescr(descr)
return loop
def test_opcount_filling_guard(self):
descr = ArrayDescr(0,4, None, 'S')
vec = ResOperation(rop.VEC_RAW_LOAD_I, ['a','i'], descr=descr)
vec.count = 4
pack = Pack([Node(ResOperation(rop.GUARD_TRUE, [vec]), 0),
Node(ResOperation(rop.GUARD_TRUE, [vec]), 1),
Node(ResOperation(rop.GUARD_TRUE, [vec]), 2),
Node(ResOperation(rop.GUARD_TRUE, [vec]), 3),
Node(ResOperation(rop.GUARD_TRUE, [vec]), 4),
Node(ResOperation(rop.GUARD_TRUE, [vec]), 5),
])
assert pack.pack_load(16) == 24-16
assert pack.pack_load(8) == 24-8
assert pack.pack_load(32) == 24-32
assert pack.opcount_filling_vector_register(16) == 4
ops, newops = pack.slice_operations(16)
assert len(ops) == 4
assert len(newops) == 2
assert pack.opcount_filling_vector_register(8) == 2
ops, newops = pack.slice_operations(8)
assert len(ops) == 2
assert len(newops) == 4
def record(self, opnum, argboxes, value, descr=None):
op = ResOperation(opnum, argboxes, descr)
if value is None:
assert op.type == 'v'
elif isinstance(value, bool):
assert op.type == 'i'
op.setint(int(value))
elif lltype.typeOf(value) == lltype.Signed:
assert op.type == 'i'
op.setint(value)
elif lltype.typeOf(value) is longlong.FLOATSTORAGE:
assert op.type == 'f'
op.setfloatstorage(value)
else:
assert lltype.typeOf(value) == llmemory.GCREF
assert op.type == 'r'
op.setref_base(value)
self.operations.append(op)
return op
def remove_tested_failarg(self, op):
opnum = op.getopnum()
if not (opnum == rop.GUARD_TRUE or opnum == rop.GUARD_FALSE):
return
if op.getarg(0).is_vector():
return
try:
i = op.getfailargs().index(op.getarg(0))
except ValueError:
return
# The checked value is also in the failargs. The front-end
# tries not to produce it, but doesn't always succeed (and
# it's hard to test all cases). Rewrite it away.
value = int(opnum == rop.GUARD_FALSE)
op1 = ResOperation(rop.SAME_AS_I, [ConstInt(value)])
op1.setint(value)
self.emit_op(op1)
lst = op.getfailargs()[:]
lst[i] = op1
newop = op.copy_and_change(opnum)
newop.setfailargs(lst)
self._changed_op = op
self._changed_op_to = newop
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
def handle_malloc_operation(self, op):
opnum = op.getopnum()
if opnum == rop.NEW:
self.handle_new_fixedsize(op.getdescr(), op)
elif opnum == rop.NEW_WITH_VTABLE:
descr = op.getdescr()
self.handle_new_fixedsize(descr, op)
if self.gc_ll_descr.fielddescr_vtable is not None:
op = ResOperation(rop.SETFIELD_GC,
[op, ConstInt(descr.get_vtable())],
descr=self.gc_ll_descr.fielddescr_vtable)
self.emit_op(op)
elif opnum == rop.NEW_ARRAY or opnum == rop.NEW_ARRAY_CLEAR:
descr = op.getdescr()
assert isinstance(descr, ArrayDescr)
self.handle_new_array(descr, op)
elif opnum == rop.NEWSTR:
self.handle_new_array(self.gc_ll_descr.str_descr, op,
kind=FLAG_STR)
elif opnum == rop.NEWUNICODE:
self.handle_new_array(self.gc_ll_descr.unicode_descr, op,
kind=FLAG_UNICODE)
else:
raise NotImplementedError(op.getopname())
def compile_simple_loop(metainterp, greenkey, start, inputargs, ops, jumpargs,
enable_opts):
from rpython.jit.metainterp.optimizeopt import optimize_trace
jitdriver_sd = metainterp.jitdriver_sd
metainterp_sd = metainterp.staticdata
jitcell_token = make_jitcell_token(jitdriver_sd)
label = ResOperation(rop.LABEL, inputargs[:], descr=jitcell_token)
jump_op = ResOperation(rop.JUMP, jumpargs[:], descr=jitcell_token)
call_pure_results = metainterp.call_pure_results
data = SimpleCompileData(label, ops + [jump_op],
call_pure_results=call_pure_results,
enable_opts=enable_opts)
try:
loop_info, ops = optimize_trace(metainterp_sd, jitdriver_sd,
data, metainterp.box_names_memo)
except InvalidLoop:
return None
loop = create_empty_loop(metainterp)
loop.original_jitcell_token = jitcell_token
loop.inputargs = loop_info.inputargs
if loop_info.quasi_immutable_deps:
loop.quasi_immutable_deps = loop_info.quasi_immutable_deps
jump_op = ops[-1]
target_token = TargetToken(jitcell_token)
target_token.original_jitcell_token = jitcell_token
label = ResOperation(rop.LABEL, loop_info.inputargs[:], descr=target_token)
jump_op.setdescr(target_token)
loop.operations = [label] + ops
if not we_are_translated():
loop.check_consistency()
jitcell_token.target_tokens = [target_token]
send_loop_to_backend(greenkey, jitdriver_sd, metainterp_sd, loop, "loop",
inputargs, metainterp.box_names_memo)
record_loop_or_bridge(metainterp_sd, loop)
return target_token
def produce_into(self, builder, r):
fail_subset = builder.subset_of_intvars(r)
original_intvars = builder.intvars[:]
builder.fakemetainterp.ovf_flag = False
super(AbstractOvfOperation, self).produce_into(builder, r)
if builder.fakemetainterp.ovf_flag: # overflow detected
op = ResOperation(rop.GUARD_OVERFLOW, [])
# the overflowed result should not be used any more, but can
# be used on the failure path: recompute fail_subset including
# the result, and then remove it from builder.intvars.
fail_subset = builder.subset_of_intvars(r)
builder.intvars[:] = original_intvars
else:
op = ResOperation(rop.GUARD_NO_OVERFLOW, [])
op.setdescr(builder.getfaildescr())
op.setfailargs(fail_subset)
builder.loop.operations.append(op)