[pypy-commit] pypy numpy-data-buffer: merge in separate-applevel-numpy
timo_jbo
noreply at buildbot.pypy.org
Tue Oct 4 00:40:57 CEST 2011
Author: Timo Paulssen <timonator at perpetuum-immobile.de>
Branch: numpy-data-buffer
Changeset: r47814:0698ee9b73e2
Date: 2011-10-04 00:35 +0200
http://bitbucket.org/pypy/pypy/changeset/0698ee9b73e2/
Log: merge in separate-applevel-numpy
diff --git a/lib_pypy/numpy/__init__.py b/lib_pypy/numpy/__init__.py
--- a/lib_pypy/numpy/__init__.py
+++ b/lib_pypy/numpy/__init__.py
@@ -13,6 +13,8 @@
arccos,
arcsin,
arctan,
+ arcsinh,
+ arctanh,
copysign,
cos,
divide,
@@ -36,6 +38,8 @@
tan,
)
+inf = float("inf")
+
def average(a):
# This implements a weighted average, for now we don't implement the
# weighting, just the average part!
diff --git a/pypy/interpreter/executioncontext.py b/pypy/interpreter/executioncontext.py
--- a/pypy/interpreter/executioncontext.py
+++ b/pypy/interpreter/executioncontext.py
@@ -307,7 +307,11 @@
self._nonperiodic_actions = []
self.has_bytecode_counter = False
self.fired_actions = None
- self.checkinterval_scaled = 100 * TICK_COUNTER_STEP
+ # the default value is not 100, unlike CPython 2.7, but a much
+ # larger value, because we use a technique that not only allows
+ # but actually *forces* another thread to run whenever the counter
+ # reaches zero.
+ self.checkinterval_scaled = 10000 * TICK_COUNTER_STEP
self._rebuild_action_dispatcher()
def fire(self, action):
diff --git a/pypy/jit/codewriter/effectinfo.py b/pypy/jit/codewriter/effectinfo.py
--- a/pypy/jit/codewriter/effectinfo.py
+++ b/pypy/jit/codewriter/effectinfo.py
@@ -74,6 +74,7 @@
OS_LLONG_UGE = 91
OS_LLONG_URSHIFT = 92
OS_LLONG_FROM_UINT = 93
+ OS_LLONG_U_TO_FLOAT = 94
#
OS_MATH_SQRT = 100
diff --git a/pypy/jit/codewriter/jtransform.py b/pypy/jit/codewriter/jtransform.py
--- a/pypy/jit/codewriter/jtransform.py
+++ b/pypy/jit/codewriter/jtransform.py
@@ -440,6 +440,7 @@
rewrite_op_ullong_mod_zer = _do_builtin_call
rewrite_op_gc_identityhash = _do_builtin_call
rewrite_op_gc_id = _do_builtin_call
+ rewrite_op_uint_mod = _do_builtin_call
# ----------
# getfield/setfield/mallocs etc.
@@ -883,9 +884,15 @@
v = v_arg
oplist = []
if unsigned1:
- opname = 'cast_uint_to_longlong'
+ if unsigned2:
+ opname = 'cast_uint_to_ulonglong'
+ else:
+ opname = 'cast_uint_to_longlong'
else:
- opname = 'cast_int_to_longlong'
+ if unsigned2:
+ opname = 'cast_int_to_ulonglong'
+ else:
+ opname = 'cast_int_to_longlong'
op2 = self.rewrite_operation(
SpaceOperation(opname, [v], v_result)
)
@@ -995,6 +1002,21 @@
return op2
''' % (_op, _oopspec.lower(), _oopspec, _oopspec)).compile()
+ for _op, _oopspec in [('cast_int_to_ulonglong', 'FROM_INT'),
+ ('cast_uint_to_ulonglong', 'FROM_UINT'),
+ ('cast_float_to_ulonglong', 'FROM_FLOAT'),
+ ('cast_ulonglong_to_float', 'U_TO_FLOAT'),
+ ]:
+ exec py.code.Source('''
+ def rewrite_op_%s(self, op):
+ args = op.args
+ op1 = self.prepare_builtin_call(op, "ullong_%s", args)
+ op2 = self._handle_oopspec_call(op1, args,
+ EffectInfo.OS_LLONG_%s,
+ EffectInfo.EF_ELIDABLE_CANNOT_RAISE)
+ return op2
+ ''' % (_op, _oopspec.lower(), _oopspec)).compile()
+
def _normalize(self, oplist):
if isinstance(oplist, SpaceOperation):
return [oplist]
diff --git a/pypy/jit/codewriter/support.py b/pypy/jit/codewriter/support.py
--- a/pypy/jit/codewriter/support.py
+++ b/pypy/jit/codewriter/support.py
@@ -315,18 +315,30 @@
def _ll_1_llong_from_int(x):
return r_longlong(intmask(x))
+def _ll_1_ullong_from_int(x):
+ return r_ulonglong(intmask(x))
+
def _ll_1_llong_from_uint(x):
return r_longlong(r_uint(x))
+def _ll_1_ullong_from_uint(x):
+ return r_ulonglong(r_uint(x))
+
def _ll_1_llong_to_int(xll):
return intmask(xll)
def _ll_1_llong_from_float(xf):
return r_longlong(xf)
+def _ll_1_ullong_from_float(xf):
+ return r_ulonglong(xf)
+
def _ll_1_llong_to_float(xll):
return float(rffi.cast(lltype.SignedLongLong, xll))
+def _ll_1_ullong_u_to_float(xull):
+ return float(rffi.cast(lltype.UnsignedLongLong, xull))
+
def _ll_1_llong_abs(xll):
if xll < 0:
@@ -351,20 +363,23 @@
return llop.llong_mod(lltype.SignedLongLong, xll, yll)
def _ll_2_ullong_floordiv(xll, yll):
- return llop.ullong_floordiv(lltype.SignedLongLong, xll, yll)
+ return llop.ullong_floordiv(lltype.UnsignedLongLong, xll, yll)
def _ll_2_ullong_floordiv_zer(xll, yll):
if yll == 0:
raise ZeroDivisionError
- return llop.ullong_floordiv(lltype.SignedLongLong, xll, yll)
+ return llop.ullong_floordiv(lltype.UnsignedLongLong, xll, yll)
def _ll_2_ullong_mod(xll, yll):
- return llop.ullong_mod(lltype.SignedLongLong, xll, yll)
+ return llop.ullong_mod(lltype.UnsignedLongLong, xll, yll)
def _ll_2_ullong_mod_zer(xll, yll):
if yll == 0:
raise ZeroDivisionError
- return llop.ullong_mod(lltype.SignedLongLong, xll, yll)
+ return llop.ullong_mod(lltype.UnsignedLongLong, xll, yll)
+
+def _ll_2_uint_mod(xll, yll):
+ return llop.uint_mod(lltype.Unsigned, xll, yll)
# libffi support
diff --git a/pypy/jit/codewriter/test/test_flatten.py b/pypy/jit/codewriter/test/test_flatten.py
--- a/pypy/jit/codewriter/test/test_flatten.py
+++ b/pypy/jit/codewriter/test/test_flatten.py
@@ -829,14 +829,15 @@
self.encoding_test(f, [rffi.cast(FROM, 42)], expectedstr,
transform=True)
elif TO in (rffi.LONG, rffi.ULONG):
+ if rffi.cast(FROM, -1) < 0:
+ fnname = "llong_from_int"
+ else:
+ fnname = "llong_from_uint"
if TO == rffi.LONG:
TO = rffi.LONGLONG
else:
TO = rffi.ULONGLONG
- if rffi.cast(FROM, -1) < 0:
- fnname = "llong_from_int"
- else:
- fnname = "llong_from_uint"
+ fnname = "u" + fnname
expected.pop() # remove int_return
expected.append(
"residual_call_irf_f $<* fn %s>, <Descr>, I[%s], R[], F[] -> %%f0"
diff --git a/pypy/jit/codewriter/test/test_longlong.py b/pypy/jit/codewriter/test/test_longlong.py
--- a/pypy/jit/codewriter/test/test_longlong.py
+++ b/pypy/jit/codewriter/test/test_longlong.py
@@ -57,7 +57,8 @@
assert op1.opname == 'residual_call_irf_f'
else:
assert op1.opname == 'residual_call_irf_i'
- gotindex = getattr(EffectInfo, 'OS_' + op1.args[0].value.upper())
+ gotindex = getattr(EffectInfo,
+ 'OS_' + op1.args[0].value.upper().lstrip('U'))
assert gotindex == oopspecindex
assert op1.args[1] == 'calldescr-%d' % oopspecindex
assert list(op1.args[2]) == [v for v in vlist
@@ -192,8 +193,12 @@
[lltype.SignedLongLong], lltype.Signed)
self.do_check('cast_float_to_longlong', EffectInfo.OS_LLONG_FROM_FLOAT,
[lltype.Float], lltype.SignedLongLong)
+ self.do_check('cast_float_to_ulonglong', EffectInfo.OS_LLONG_FROM_FLOAT,
+ [lltype.Float], lltype.UnsignedLongLong)
self.do_check('cast_longlong_to_float', EffectInfo.OS_LLONG_TO_FLOAT,
[lltype.SignedLongLong], lltype.Float)
+ self.do_check('cast_ulonglong_to_float', EffectInfo.OS_LLONG_U_TO_FLOAT,
+ [lltype.UnsignedLongLong], lltype.Float)
for T1 in [lltype.SignedLongLong, lltype.UnsignedLongLong]:
for T2 in [lltype.Signed, lltype.Unsigned]:
self.do_check('cast_primitive', EffectInfo.OS_LLONG_TO_INT,
diff --git a/pypy/jit/metainterp/optimizeopt/__init__.py b/pypy/jit/metainterp/optimizeopt/__init__.py
--- a/pypy/jit/metainterp/optimizeopt/__init__.py
+++ b/pypy/jit/metainterp/optimizeopt/__init__.py
@@ -7,6 +7,8 @@
from pypy.jit.metainterp.optimizeopt.unroll import optimize_unroll, OptInlineShortPreamble
from pypy.jit.metainterp.optimizeopt.fficall import OptFfiCall
from pypy.jit.metainterp.optimizeopt.simplify import OptSimplify
+from pypy.jit.metainterp.optimizeopt.pure import OptPure
+from pypy.jit.metainterp.optimizeopt.earlyforce import OptEarlyForce
from pypy.rlib.jit import PARAMETERS
from pypy.rlib.unroll import unrolling_iterable
@@ -14,6 +16,8 @@
('rewrite', OptRewrite),
('virtualize', OptVirtualize),
('string', OptString),
+ ('earlyforce', OptEarlyForce),
+ ('pure', OptPure),
('heap', OptHeap),
('ffi', None),
('unroll', None)]
diff --git a/pypy/jit/metainterp/optimizeopt/earlyforce.py b/pypy/jit/metainterp/optimizeopt/earlyforce.py
new file mode 100644
--- /dev/null
+++ b/pypy/jit/metainterp/optimizeopt/earlyforce.py
@@ -0,0 +1,24 @@
+from pypy.jit.metainterp.optimizeopt.optimizer import Optimization
+from pypy.jit.metainterp.optimizeopt.vstring import VAbstractStringValue
+from pypy.jit.metainterp.resoperation import rop, ResOperation
+
+class OptEarlyForce(Optimization):
+ def propagate_forward(self, op):
+ opnum = op.getopnum()
+ if (opnum != rop.SETFIELD_GC and
+ opnum != rop.SETARRAYITEM_GC and
+ opnum != rop.QUASIIMMUT_FIELD):
+
+ for arg in op.getarglist():
+ if arg in self.optimizer.values:
+ value = self.getvalue(arg)
+ value.force_box(self)
+ self.emit_operation(op)
+
+ def new(self):
+ return OptEarlyForce()
+
+ def setup(self):
+ self.optimizer.optearlyforce = self
+
+
diff --git a/pypy/jit/metainterp/optimizeopt/fficall.py b/pypy/jit/metainterp/optimizeopt/fficall.py
--- a/pypy/jit/metainterp/optimizeopt/fficall.py
+++ b/pypy/jit/metainterp/optimizeopt/fficall.py
@@ -177,10 +177,10 @@
funcinfo.descr is None):
return [op] # cannot optimize
funcsymval = self.getvalue(op.getarg(2))
- arglist = [funcsymval.force_box()]
+ arglist = [funcsymval.get_key_box()]
for push_op in funcinfo.opargs:
argval = self.getvalue(push_op.getarg(2))
- arglist.append(argval.force_box())
+ arglist.append(argval.get_key_box())
newop = ResOperation(rop.CALL_RELEASE_GIL, arglist, op.result,
descr=funcinfo.descr)
self.commit_optimization()
diff --git a/pypy/jit/metainterp/optimizeopt/heap.py b/pypy/jit/metainterp/optimizeopt/heap.py
--- a/pypy/jit/metainterp/optimizeopt/heap.py
+++ b/pypy/jit/metainterp/optimizeopt/heap.py
@@ -50,6 +50,7 @@
if not self._lazy_setfield_registered:
optheap._lazy_setfields_and_arrayitems.append(self)
self._lazy_setfield_registered = True
+
else:
# this is the case where the pending setfield ends up
# storing precisely the value that is already there,
@@ -158,12 +159,17 @@
self._lazy_setfields_and_arrayitems = []
self._remove_guard_not_invalidated = False
self._seen_guard_not_invalidated = False
+ self.posponedop = None
def force_at_end_of_preamble(self):
self.force_all_lazy_setfields_and_arrayitems()
def flush(self):
self.force_all_lazy_setfields_and_arrayitems()
+ if self.posponedop:
+ posponedop = self.posponedop
+ self.posponedop = None
+ self.next_optimization.propagate_forward(posponedop)
def new(self):
return OptHeap()
@@ -211,7 +217,15 @@
def emit_operation(self, op):
self.emitting_operation(op)
- self.next_optimization.propagate_forward(op)
+ if self.posponedop:
+ posponedop = self.posponedop
+ self.posponedop = None
+ self.next_optimization.propagate_forward(posponedop)
+ if (op.is_comparison() or op.getopnum() == rop.CALL_MAY_FORCE
+ or op.is_ovf()):
+ self.posponedop = op
+ else:
+ self.next_optimization.propagate_forward(op)
def emitting_operation(self, op):
if op.has_no_side_effect():
@@ -293,30 +307,6 @@
if indexvalue is None or indexvalue.intbound.contains(idx):
cf.force_lazy_setfield(self, can_cache)
- def fixup_guard_situation(self):
- # hackish: reverse the order of the last two operations if it makes
- # sense to avoid a situation like "int_eq/setfield_gc/guard_true",
- # which the backend (at least the x86 backend) does not handle well.
- newoperations = self.optimizer.newoperations
- if len(newoperations) < 2:
- return
- lastop = newoperations[-1]
- if (lastop.getopnum() != rop.SETFIELD_GC and
- lastop.getopnum() != rop.SETARRAYITEM_GC):
- return
- # - is_comparison() for cases like "int_eq/setfield_gc/guard_true"
- # - CALL_MAY_FORCE: "call_may_force/setfield_gc/guard_not_forced"
- # - is_ovf(): "int_add_ovf/setfield_gc/guard_no_overflow"
- prevop = newoperations[-2]
- opnum = prevop.getopnum()
- if not (prevop.is_comparison() or opnum == rop.CALL_MAY_FORCE
- or prevop.is_ovf()):
- return
- if prevop.result in lastop.getarglist():
- return
- newoperations[-2] = lastop
- newoperations[-1] = prevop
-
def _assert_valid_cf(self, cf):
# check that 'cf' is in cached_fields or cached_arrayitems
if not we_are_translated():
@@ -362,7 +352,6 @@
fieldvalue.get_key_box(), itemindex))
else:
cf.force_lazy_setfield(self)
- self.fixup_guard_situation()
return pendingfields
def optimize_GETFIELD_GC(self, op):
@@ -374,12 +363,22 @@
return
# default case: produce the operation
structvalue.ensure_nonnull()
- ###self.optimizer.optimize_default(op)
self.emit_operation(op)
# then remember the result of reading the field
fieldvalue = self.getvalue(op.result)
cf.remember_field_value(structvalue, fieldvalue, op)
+ def optimize_GETFIELD_GC_PURE(self, op):
+ structvalue = self.getvalue(op.getarg(0))
+ cf = self.field_cache(op.getdescr())
+ fieldvalue = cf.getfield_from_cache(self, structvalue)
+ if fieldvalue is not None:
+ self.make_equal_to(op.result, fieldvalue)
+ return
+ # default case: produce the operation
+ structvalue.ensure_nonnull()
+ self.emit_operation(op)
+
def optimize_SETFIELD_GC(self, op):
if self.has_pure_result(rop.GETFIELD_GC_PURE, [op.getarg(0)],
op.getdescr()):
@@ -389,6 +388,7 @@
#
cf = self.field_cache(op.getdescr())
cf.do_setfield(self, op)
+
def optimize_GETARRAYITEM_GC(self, op):
arrayvalue = self.getvalue(op.getarg(0))
@@ -413,6 +413,25 @@
fieldvalue = self.getvalue(op.result)
cf.remember_field_value(arrayvalue, fieldvalue, op)
+ def optimize_GETARRAYITEM_GC_PURE(self, op):
+ arrayvalue = self.getvalue(op.getarg(0))
+ indexvalue = self.getvalue(op.getarg(1))
+ cf = None
+ if indexvalue.is_constant():
+ arrayvalue.make_len_gt(MODE_ARRAY, op.getdescr(), indexvalue.box.getint())
+ # use the cache on (arraydescr, index), which is a constant
+ cf = self.arrayitem_cache(op.getdescr(), indexvalue.box.getint())
+ fieldvalue = cf.getfield_from_cache(self, arrayvalue)
+ if fieldvalue is not None:
+ self.make_equal_to(op.result, fieldvalue)
+ return
+ else:
+ # variable index, so make sure the lazy setarrayitems are done
+ self.force_lazy_setarrayitem(op.getdescr(), indexvalue=indexvalue)
+ # default case: produce the operation
+ arrayvalue.ensure_nonnull()
+ self.emit_operation(op)
+
def optimize_SETARRAYITEM_GC(self, op):
if self.has_pure_result(rop.GETARRAYITEM_GC_PURE, [op.getarg(0),
op.getarg(1)],
diff --git a/pypy/jit/metainterp/optimizeopt/optimizer.py b/pypy/jit/metainterp/optimizeopt/optimizer.py
--- a/pypy/jit/metainterp/optimizeopt/optimizer.py
+++ b/pypy/jit/metainterp/optimizeopt/optimizer.py
@@ -31,8 +31,8 @@
class OptValue(object):
__metaclass__ = extendabletype
- _attrs_ = ('box', 'known_class', 'last_guard_index', 'level', 'intbound', 'lenbound')
- last_guard_index = -1
+ _attrs_ = ('box', 'known_class', 'last_guard', 'level', 'intbound', 'lenbound')
+ last_guard = None
level = LEVEL_UNKNOWN
known_class = None
@@ -100,7 +100,7 @@
self.make_constant(other.get_key_box())
optimizer.turned_constant(self)
elif other.level == LEVEL_KNOWNCLASS:
- self.make_constant_class(other.known_class, -1)
+ self.make_constant_class(other.known_class, None)
else:
if other.level == LEVEL_NONNULL:
self.ensure_nonnull()
@@ -114,13 +114,13 @@
self.lenbound = other.lenbound.clone()
- def force_box(self):
+ def force_box(self, optforce):
return self.box
def get_key_box(self):
return self.box
- def force_at_end_of_preamble(self, already_forced):
+ def force_at_end_of_preamble(self, already_forced, optforce):
return self
def get_args_for_fail(self, modifier):
@@ -162,16 +162,16 @@
else:
return None
- def make_constant_class(self, classbox, opindex):
+ def make_constant_class(self, classbox, guardop):
assert self.level < LEVEL_KNOWNCLASS
self.known_class = classbox
self.level = LEVEL_KNOWNCLASS
- self.last_guard_index = opindex
+ self.last_guard = guardop
- def make_nonnull(self, opindex):
+ def make_nonnull(self, guardop):
assert self.level < LEVEL_NONNULL
self.level = LEVEL_NONNULL
- self.last_guard_index = opindex
+ self.last_guard = guardop
def is_nonnull(self):
level = self.level
@@ -240,26 +240,12 @@
def __init__(self):
pass # make rpython happy
- def propagate_begin_forward(self):
- if self.next_optimization:
- self.next_optimization.propagate_begin_forward()
-
- def propagate_end_forward(self):
- if self.next_optimization:
- self.next_optimization.propagate_end_forward()
-
def propagate_forward(self, op):
raise NotImplementedError
def emit_operation(self, op):
self.next_optimization.propagate_forward(op)
- def test_emittable(self, op):
- return self.is_emittable(op)
-
- def is_emittable(self, op):
- return self.next_optimization.test_emittable(op)
-
# FIXME: Move some of these here?
def getvalue(self, box):
return self.optimizer.getvalue(box)
@@ -289,19 +275,19 @@
return self.optimizer.new_const_item(arraydescr)
def pure(self, opnum, args, result):
- op = ResOperation(opnum, args, result)
- key = self.optimizer.make_args_key(op)
- if key not in self.optimizer.pure_operations:
- self.optimizer.pure_operations[key] = op
+ if self.optimizer.optpure:
+ self.optimizer.optpure.pure(opnum, args, result)
def has_pure_result(self, opnum, args, descr):
- op = ResOperation(opnum, args, None, descr)
- key = self.optimizer.make_args_key(op)
- op = self.optimizer.pure_operations.get(key, None)
- if op is None:
- return False
- return op.getdescr() is descr
+ if self.optimizer.optpure:
+ return self.optimizer.optpure.has_pure_result(opnum, args, descr)
+ return False
+ def get_pure_result(self, key):
+ if self.optimizer.optpure:
+ return self.optimizer.optpure.get_pure_result(key)
+ return None
+
def setup(self):
pass
@@ -322,6 +308,9 @@
def produce_potential_short_preamble_ops(self, potential_ops):
pass
+ def forget_numberings(self, box):
+ self.optimizer.forget_numberings(box)
+
class Optimizer(Optimization):
def __init__(self, metainterp_sd, loop, optimizations=None, bridge=False):
@@ -333,14 +322,16 @@
self.interned_refs = self.cpu.ts.new_ref_dict()
self.resumedata_memo = resume.ResumeDataLoopMemo(metainterp_sd)
self.bool_boxes = {}
- self.pure_operations = args_dict()
self.producer = {}
self.pendingfields = []
- self.posponedop = None
self.exception_might_have_happened = False
self.quasi_immutable_deps = None
self.opaque_pointers = {}
- self.newoperations = []
+ self.replaces_guard = {}
+ self._newoperations = []
+ self.optimizer = self
+ self.optpure = None
+ self.optearlyforce = None
if loop is not None:
self.call_pure_results = loop.call_pure_results
@@ -369,21 +360,20 @@
def flush(self):
for o in self.optimizations:
o.flush()
- assert self.posponedop is None
def new(self):
new = Optimizer(self.metainterp_sd, self.loop)
return self._new(new)
def _new(self, new):
- assert self.posponedop is None
optimizations = [o.new() for o in self.optimizations]
new.set_optimizations(optimizations)
new.quasi_immutable_deps = self.quasi_immutable_deps
return new
def produce_potential_short_preamble_ops(self, sb):
- raise NotImplementedError('This is implemented in unroll.UnrollableOptimizer')
+ for opt in self.optimizations:
+ opt.produce_potential_short_preamble_ops(sb)
def turned_constant(self, value):
for o in self.optimizations:
@@ -433,6 +423,13 @@
return constbox
return None
+ def get_newoperations(self):
+ self.flush()
+ return self._newoperations
+
+ def clear_newoperations(self):
+ self._newoperations = []
+
def make_equal_to(self, box, value, replace=False):
assert isinstance(value, OptValue)
assert replace or box not in self.values
@@ -481,15 +478,10 @@
def propagate_all_forward(self):
self.exception_might_have_happened = self.bridge
- self.newoperations = []
- self.first_optimization.propagate_begin_forward()
- self.i = 0
- while self.i < len(self.loop.operations):
- op = self.loop.operations[self.i]
+ self.clear_newoperations()
+ for op in self.loop.operations:
self.first_optimization.propagate_forward(op)
- self.i += 1
- self.first_optimization.propagate_end_forward()
- self.loop.operations = self.newoperations
+ self.loop.operations = self.get_newoperations()
self.loop.quasi_immutable_deps = self.quasi_immutable_deps
# accumulate counters
self.resumedata_memo.update_counters(self.metainterp_sd.profiler)
@@ -501,9 +493,6 @@
self.producer[op.result] = op
dispatch_opt(self, op)
- def test_emittable(self, op):
- return True
-
def emit_operation(self, op):
if op.returns_bool_result():
self.bool_boxes[self.getvalue(op.result)] = None
@@ -519,14 +508,30 @@
pass
else:
self.ensure_imported(value)
- op.setarg(i, value.force_box())
+ op.setarg(i, value.force_box(self))
self.metainterp_sd.profiler.count(jitprof.OPT_OPS)
if op.is_guard():
self.metainterp_sd.profiler.count(jitprof.OPT_GUARDS)
- op = self.store_final_boxes_in_guard(op)
+ if self.replaces_guard and op in self.replaces_guard:
+ self.replace_op(self.replaces_guard[op], op)
+ del self.replaces_guard[op]
+ return
+ else:
+ op = self.store_final_boxes_in_guard(op)
elif op.can_raise():
self.exception_might_have_happened = True
- self.newoperations.append(op)
+ self._newoperations.append(op)
+
+ def replace_op(self, old_op, new_op):
+ # XXX: Do we want to cache indexes to prevent search?
+ i = len(self._newoperations)
+ while i > 0:
+ i -= 1
+ if self._newoperations[i] is old_op:
+ self._newoperations[i] = new_op
+ break
+ else:
+ assert False
def store_final_boxes_in_guard(self, op):
descr = op.getdescr()
@@ -574,51 +579,8 @@
args[n + 1] = op.getdescr()
return args
- @specialize.argtype(0)
def optimize_default(self, op):
- canfold = op.is_always_pure()
- if op.is_ovf():
- self.posponedop = op
- return
- if self.posponedop:
- nextop = op
- op = self.posponedop
- self.posponedop = None
- canfold = nextop.getopnum() == rop.GUARD_NO_OVERFLOW
- else:
- nextop = None
-
- if canfold:
- for i in range(op.numargs()):
- if self.get_constant_box(op.getarg(i)) is None:
- break
- else:
- # all constant arguments: constant-fold away
- resbox = self.constant_fold(op)
- # note that INT_xxx_OVF is not done from here, and the
- # overflows in the INT_xxx operations are ignored
- self.make_constant(op.result, resbox)
- return
-
- # did we do the exact same operation already?
- args = self.make_args_key(op)
- oldop = self.pure_operations.get(args, None)
- if oldop is not None and oldop.getdescr() is op.getdescr():
- assert oldop.getopnum() == op.getopnum()
- self.make_equal_to(op.result, self.getvalue(oldop.result),
- True)
- return
- else:
- self.pure_operations[args] = op
- self.remember_emitting_pure(op)
-
- # otherwise, the operation remains
self.emit_operation(op)
- if nextop:
- self.emit_operation(nextop)
-
- def remember_emitting_pure(self, op):
- pass
def constant_fold(self, op):
argboxes = [self.get_constant_box(op.getarg(i))
@@ -636,11 +598,6 @@
def optimize_DEBUG_MERGE_POINT(self, op):
self.emit_operation(op)
- def optimize_CAST_OPAQUE_PTR(self, op):
- value = self.getvalue(op.getarg(0))
- self.opaque_pointers[value] = True
- self.make_equal_to(op.result, value)
-
def optimize_GETARRAYITEM_GC_PURE(self, op):
indexvalue = self.getvalue(op.getarg(1))
if indexvalue.is_constant():
diff --git a/pypy/jit/metainterp/optimizeopt/pure.py b/pypy/jit/metainterp/optimizeopt/pure.py
new file mode 100644
--- /dev/null
+++ b/pypy/jit/metainterp/optimizeopt/pure.py
@@ -0,0 +1,115 @@
+from pypy.jit.metainterp.optimizeopt.optimizer import Optimization
+from pypy.jit.metainterp.resoperation import rop, ResOperation
+from pypy.jit.metainterp.optimizeopt.util import (make_dispatcher_method,
+ args_dict)
+
+class OptPure(Optimization):
+ def __init__(self):
+ self.posponedop = None
+ self.pure_operations = args_dict()
+ self.emitted_pure_operations = {}
+
+ def propagate_forward(self, op):
+ dispatch_opt(self, op)
+
+ def optimize_default(self, op):
+ canfold = op.is_always_pure()
+ if op.is_ovf():
+ self.posponedop = op
+ return
+ if self.posponedop:
+ nextop = op
+ op = self.posponedop
+ self.posponedop = None
+ canfold = nextop.getopnum() == rop.GUARD_NO_OVERFLOW
+ else:
+ nextop = None
+
+ if canfold:
+ for i in range(op.numargs()):
+ if self.get_constant_box(op.getarg(i)) is None:
+ break
+ else:
+ # all constant arguments: constant-fold away
+ resbox = self.optimizer.constant_fold(op)
+ # note that INT_xxx_OVF is not done from here, and the
+ # overflows in the INT_xxx operations are ignored
+ self.optimizer.make_constant(op.result, resbox)
+ return
+
+ # did we do the exact same operation already?
+ args = self.optimizer.make_args_key(op)
+ oldop = self.pure_operations.get(args, None)
+ if oldop is not None and oldop.getdescr() is op.getdescr():
+ assert oldop.getopnum() == op.getopnum()
+ self.optimizer.make_equal_to(op.result, self.getvalue(oldop.result),
+ True)
+ return
+ else:
+ self.pure_operations[args] = op
+ self.remember_emitting_pure(op)
+
+ # otherwise, the operation remains
+ self.emit_operation(op)
+ if op.returns_bool_result():
+ self.optimizer.bool_boxes[self.getvalue(op.result)] = None
+ if nextop:
+ self.emit_operation(nextop)
+
+ def optimize_CALL_PURE(self, op):
+ args = self.optimizer.make_args_key(op)
+ oldop = self.pure_operations.get(args, None)
+ if oldop is not None and oldop.getdescr() is op.getdescr():
+ assert oldop.getopnum() == op.getopnum()
+ self.make_equal_to(op.result, self.getvalue(oldop.result))
+ return
+ else:
+ self.pure_operations[args] = op
+ self.remember_emitting_pure(op)
+
+ # replace CALL_PURE with just CALL
+ args = op.getarglist()
+ self.emit_operation(ResOperation(rop.CALL, args, op.result,
+ op.getdescr()))
+
+ def flush(self):
+ assert self.posponedop is None
+
+ def new(self):
+ assert self.posponedop is None
+ return OptPure()
+
+ def setup(self):
+ self.optimizer.optpure = self
+
+ def pure(self, opnum, args, result):
+ op = ResOperation(opnum, args, result)
+ key = self.optimizer.make_args_key(op)
+ if key not in self.pure_operations:
+ self.pure_operations[key] = op
+
+ 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 get_pure_result(self, key):
+ return self.pure_operations.get(key, None)
+
+ def remember_emitting_pure(self, op):
+ self.emitted_pure_operations[op] = True
+
+ def produce_potential_short_preamble_ops(self, sb):
+ for op in self.emitted_pure_operations:
+ if op.getopnum() == rop.GETARRAYITEM_GC_PURE or \
+ op.getopnum() == rop.STRGETITEM or \
+ op.getopnum() == rop.UNICODEGETITEM:
+ if not self.getvalue(op.getarg(1)).is_constant():
+ continue
+ sb.add_potential(op)
+
+dispatch_opt = make_dispatcher_method(OptPure, 'optimize_',
+ default=OptPure.optimize_default)
diff --git a/pypy/jit/metainterp/optimizeopt/rewrite.py b/pypy/jit/metainterp/optimizeopt/rewrite.py
--- a/pypy/jit/metainterp/optimizeopt/rewrite.py
+++ b/pypy/jit/metainterp/optimizeopt/rewrite.py
@@ -31,21 +31,8 @@
dispatch_opt(self, op)
- def test_emittable(self, op):
- opnum = op.getopnum()
- for value, cls, func in optimize_guards:
- if opnum == value:
- assert isinstance(op, cls)
- try:
- func(self, op, dryrun=True)
- return self.is_emittable(op)
- except InvalidLoop:
- return False
- return self.is_emittable(op)
-
-
def try_boolinvers(self, op, targs):
- oldop = self.optimizer.pure_operations.get(targs, None)
+ oldop = self.get_pure_result(targs)
if oldop is not None and oldop.getdescr() is op.getdescr():
value = self.getvalue(oldop.result)
if value.is_constant():
@@ -73,7 +60,7 @@
oldopnum = opboolreflex[op.getopnum()] # FIXME: add INT_ADD, INT_MUL
targs = self.optimizer.make_args_key(ResOperation(oldopnum, [args[1], args[0]],
None))
- oldop = self.optimizer.pure_operations.get(targs, None)
+ oldop = self.get_pure_result(targs)
if oldop is not None and oldop.getdescr() is op.getdescr():
self.make_equal_to(op.result, self.getvalue(oldop.result))
return True
@@ -214,40 +201,7 @@
self.emit_operation(op)
self.pure(rop.FLOAT_NEG, [op.result], v1)
- def optimize_CALL_PURE(self, op):
- arg_consts = []
- for i in range(op.numargs()):
- arg = op.getarg(i)
- const = self.get_constant_box(arg)
- if const is None:
- break
- arg_consts.append(const)
- else:
- # all constant arguments: check if we already know the result
- try:
- result = self.optimizer.call_pure_results[arg_consts]
- except KeyError:
- pass
- else:
- self.make_constant(op.result, result)
- return
-
- args = self.optimizer.make_args_key(op)
- oldop = self.optimizer.pure_operations.get(args, None)
- if oldop is not None and oldop.getdescr() is op.getdescr():
- assert oldop.getopnum() == op.getopnum()
- self.make_equal_to(op.result, self.getvalue(oldop.result))
- return
- else:
- self.optimizer.pure_operations[args] = op
- self.optimizer.remember_emitting_pure(op)
-
- # replace CALL_PURE with just CALL
- args = op.getarglist()
- self.emit_operation(ResOperation(rop.CALL, args, op.result,
- op.getdescr()))
-
- def optimize_guard(self, op, constbox, emit_operation=True, dryrun=False):
+ def optimize_guard(self, op, constbox, emit_operation=True):
value = self.getvalue(op.getarg(0))
if value.is_constant():
box = value.box
@@ -255,62 +209,57 @@
if not box.same_constant(constbox):
raise InvalidLoop
return
- if dryrun: return
if emit_operation:
self.emit_operation(op)
value.make_constant(constbox)
self.optimizer.turned_constant(value)
- def optimize_GUARD_ISNULL(self, op, dryrun=False):
+ def optimize_GUARD_ISNULL(self, op):
value = self.getvalue(op.getarg(0))
if value.is_null():
return
elif value.is_nonnull():
raise InvalidLoop
- if dryrun: return
self.emit_operation(op)
value.make_constant(self.optimizer.cpu.ts.CONST_NULL)
- def optimize_GUARD_NONNULL(self, op, dryrun=False):
+ def optimize_GUARD_NONNULL(self, op):
value = self.getvalue(op.getarg(0))
if value.is_nonnull():
return
elif value.is_null():
raise InvalidLoop
- if dryrun: return
self.emit_operation(op)
- value.make_nonnull(len(self.optimizer.newoperations) - 1)
+ value.make_nonnull(op)
- def optimize_GUARD_VALUE(self, op, dryrun=False):
+ def optimize_GUARD_VALUE(self, op):
value = self.getvalue(op.getarg(0))
- emit_operation = True
- if not dryrun and value.last_guard_index != -1:
+ if value.last_guard:
# there already has been a guard_nonnull or guard_class or
# guard_nonnull_class on this value, which is rather silly.
# replace the original guard with a guard_value
- old_guard_op = self.optimizer.newoperations[value.last_guard_index]
- new_guard_op = old_guard_op.copy_and_change(rop.GUARD_VALUE,
- args = [old_guard_op.getarg(0), op.getarg(1)])
- self.optimizer.newoperations[value.last_guard_index] = new_guard_op
+ old_guard_op = value.last_guard
+ op = old_guard_op.copy_and_change(rop.GUARD_VALUE,
+ args = [old_guard_op.getarg(0), op.getarg(1)])
+ self.optimizer.replaces_guard[op] = old_guard_op
# hack hack hack. Change the guard_opnum on
# new_guard_op.getdescr() so that when resuming,
# the operation is not skipped by pyjitpl.py.
- descr = new_guard_op.getdescr()
+ descr = op.getdescr()
assert isinstance(descr, compile.ResumeGuardDescr)
descr.guard_opnum = rop.GUARD_VALUE
- descr.make_a_counter_per_value(new_guard_op)
- emit_operation = False
+ descr.make_a_counter_per_value(op)
constbox = op.getarg(1)
assert isinstance(constbox, Const)
- self.optimize_guard(op, constbox, emit_operation, dryrun)
+ self.optimize_guard(op, constbox)
- def optimize_GUARD_TRUE(self, op, dryrun=False):
- self.optimize_guard(op, CONST_1, dryrun=dryrun)
+ def optimize_GUARD_TRUE(self, op):
+ self.optimize_guard(op, CONST_1)
- def optimize_GUARD_FALSE(self, op, dryrun=False):
- self.optimize_guard(op, CONST_0, dryrun=dryrun)
+ def optimize_GUARD_FALSE(self, op):
+ self.optimize_guard(op, CONST_0)
- def optimize_GUARD_CLASS(self, op, dryrun=False):
+ def optimize_GUARD_CLASS(self, op):
value = self.getvalue(op.getarg(0))
expectedclassbox = op.getarg(1)
assert isinstance(expectedclassbox, Const)
@@ -319,38 +268,32 @@
if realclassbox.same_constant(expectedclassbox):
return
raise InvalidLoop
- if dryrun: return
- emit_operation = True
- if value.last_guard_index != -1:
+ if value.last_guard:
# there already has been a guard_nonnull or guard_class or
# guard_nonnull_class on this value.
- old_guard_op = self.optimizer.newoperations[value.last_guard_index]
+ old_guard_op = value.last_guard
if old_guard_op.getopnum() == rop.GUARD_NONNULL:
# it was a guard_nonnull, which we replace with a
# guard_nonnull_class.
- new_guard_op = old_guard_op.copy_and_change (rop.GUARD_NONNULL_CLASS,
+ op = old_guard_op.copy_and_change (rop.GUARD_NONNULL_CLASS,
args = [old_guard_op.getarg(0), op.getarg(1)])
- self.optimizer.newoperations[value.last_guard_index] = new_guard_op
+ self.optimizer.replaces_guard[op] = old_guard_op
# hack hack hack. Change the guard_opnum on
# new_guard_op.getdescr() so that when resuming,
# the operation is not skipped by pyjitpl.py.
- descr = new_guard_op.getdescr()
+ descr = op.getdescr()
assert isinstance(descr, compile.ResumeGuardDescr)
descr.guard_opnum = rop.GUARD_NONNULL_CLASS
- emit_operation = False
- if emit_operation:
- self.emit_operation(op)
- last_guard_index = len(self.optimizer.newoperations) - 1
- else:
- last_guard_index = value.last_guard_index
- value.make_constant_class(expectedclassbox, last_guard_index)
+ self.emit_operation(op)
+ value.make_constant_class(expectedclassbox, op)
- def optimize_GUARD_NONNULL_CLASS(self, op, dryrun=False):
- self.optimize_GUARD_NONNULL(op, True)
- self.optimize_GUARD_CLASS(op, dryrun)
+ def optimize_GUARD_NONNULL_CLASS(self, op):
+ value = self.getvalue(op.getarg(0))
+ if value.is_null():
+ raise InvalidLoop
+ self.optimize_GUARD_CLASS(op)
- def optimize_GUARD_NO_EXCEPTION(self, op, dryrun=False):
- if dryrun: return
+ def optimize_GUARD_NO_EXCEPTION(self, op):
if not self.optimizer.exception_might_have_happened:
return
self.emit_operation(op)
@@ -470,7 +413,7 @@
newop = ResOperation(rop.SETARRAYITEM_GC,
[op.getarg(2),
ConstInt(index + dest_start),
- val.force_box()], None,
+ val.get_key_box()], None,
descr=source_value.arraydescr)
self.emit_operation(newop)
return True
@@ -478,6 +421,25 @@
return True # 0-length arraycopy
return False
+ def optimize_CALL_PURE(self, op):
+ arg_consts = []
+ for i in range(op.numargs()):
+ arg = op.getarg(i)
+ const = self.get_constant_box(arg)
+ if const is None:
+ break
+ arg_consts.append(const)
+ else:
+ # all constant arguments: check if we already know the result
+ try:
+ result = self.optimizer.call_pure_results[arg_consts]
+ except KeyError:
+ pass
+ else:
+ self.make_constant(op.result, result)
+ return
+ self.emit_operation(op)
+
def optimize_INT_FLOORDIV(self, op):
v1 = self.getvalue(op.getarg(0))
v2 = self.getvalue(op.getarg(1))
@@ -492,6 +454,10 @@
args = [op.getarg(0), ConstInt(highest_bit(val))])
self.emit_operation(op)
+ def optimize_CAST_OPAQUE_PTR(self, op):
+ value = self.getvalue(op.getarg(0))
+ self.optimizer.opaque_pointers[value] = True
+ self.make_equal_to(op.result, value)
dispatch_opt = make_dispatcher_method(OptRewrite, 'optimize_',
default=OptRewrite.emit_operation)
diff --git a/pypy/jit/metainterp/optimizeopt/test/test_optimizebasic.py b/pypy/jit/metainterp/optimizeopt/test/test_optimizebasic.py
--- a/pypy/jit/metainterp/optimizeopt/test/test_optimizebasic.py
+++ b/pypy/jit/metainterp/optimizeopt/test/test_optimizebasic.py
@@ -39,8 +39,9 @@
def test_sharing_field_lists_of_virtual():
class FakeOptimizer(object):
- class cpu(object):
- pass
+ class optimizer(object):
+ class cpu(object):
+ pass
opt = FakeOptimizer()
virt1 = virtualize.AbstractVirtualStructValue(opt, None)
lst1 = virt1._get_field_descr_list()
@@ -69,7 +70,7 @@
class FakeVirtualValue(virtualize.AbstractVirtualValue):
def _make_virtual(self, *args):
return FakeVInfo()
- v1 = FakeVirtualValue(None, None, None)
+ v1 = FakeVirtualValue(None, None)
vinfo1 = v1.make_virtual_info(None, [1, 2, 4])
vinfo2 = v1.make_virtual_info(None, [1, 2, 4])
assert vinfo1 is vinfo2
@@ -111,7 +112,7 @@
class BaseTestBasic(BaseTest):
- enable_opts = "intbounds:rewrite:virtualize:string:heap"
+ enable_opts = "intbounds:rewrite:virtualize:string:earlyforce:pure:heap"
def optimize_loop(self, ops, optops, call_pure_results=None):
@@ -651,8 +652,8 @@
i3 = getfield_gc(p3, descr=valuedescr)
escape(i3)
p1 = new_with_vtable(ConstClass(node_vtable))
+ p1sub = new_with_vtable(ConstClass(node_vtable2))
setfield_gc(p1, i1, descr=valuedescr)
- p1sub = new_with_vtable(ConstClass(node_vtable2))
setfield_gc(p1sub, i1, descr=valuedescr)
setfield_gc(p1, p1sub, descr=nextdescr)
jump(i1, p1, p2)
@@ -667,10 +668,10 @@
p3sub = getfield_gc(p3, descr=nextdescr)
i3 = getfield_gc(p3sub, descr=valuedescr)
escape(i3)
+ p1 = new_with_vtable(ConstClass(node_vtable))
p2sub = new_with_vtable(ConstClass(node_vtable2))
setfield_gc(p2sub, i1, descr=valuedescr)
setfield_gc(p2, p2sub, descr=nextdescr)
- p1 = new_with_vtable(ConstClass(node_vtable))
jump(i1, p1, p2)
"""
# The same as test_p123_simple, but in the end the "old" p2 contains
diff --git a/pypy/jit/metainterp/optimizeopt/test/test_optimizefficall.py b/pypy/jit/metainterp/optimizeopt/test/test_optimizefficall.py
--- a/pypy/jit/metainterp/optimizeopt/test/test_optimizefficall.py
+++ b/pypy/jit/metainterp/optimizeopt/test/test_optimizefficall.py
@@ -36,7 +36,7 @@
class TestFfiCall(BaseTestBasic, LLtypeMixin):
- enable_opts = "intbounds:rewrite:virtualize:string:heap:ffi"
+ enable_opts = "intbounds:rewrite:virtualize:string:pure:earlyforce:heap:ffi"
class namespace:
cpu = LLtypeMixin.cpu
diff --git a/pypy/jit/metainterp/optimizeopt/test/test_optimizeopt.py b/pypy/jit/metainterp/optimizeopt/test/test_optimizeopt.py
--- a/pypy/jit/metainterp/optimizeopt/test/test_optimizeopt.py
+++ b/pypy/jit/metainterp/optimizeopt/test/test_optimizeopt.py
@@ -68,7 +68,7 @@
class BaseTestWithUnroll(BaseTest):
- enable_opts = "intbounds:rewrite:virtualize:string:heap:unroll"
+ enable_opts = "intbounds:rewrite:virtualize:string:earlyforce:pure:heap:unroll"
def optimize_loop(self, ops, expected, expected_preamble=None,
call_pure_results=None, expected_short=None):
@@ -825,8 +825,8 @@
i3 = getfield_gc(p2, descr=valuedescr)
escape(i3)
p4 = new_with_vtable(ConstClass(node_vtable))
+ p1sub = new_with_vtable(ConstClass(node_vtable2))
setfield_gc(p4, i1, descr=valuedescr)
- p1sub = new_with_vtable(ConstClass(node_vtable2))
setfield_gc(p1sub, i1, descr=valuedescr)
setfield_gc(p4, p1sub, descr=nextdescr)
jump(i1, p4)
@@ -865,13 +865,7 @@
p3sub = new_with_vtable(ConstClass(node_vtable2))
setfield_gc(p3sub, i1, descr=valuedescr)
setfield_gc(p1, p3sub, descr=nextdescr)
- # XXX: We get two extra operations here because the setfield
- # above is the result of forcing p1 and thus not
- # registered with the heap optimizer. I've makred tests
- # below with VIRTUALHEAP if they suffer from this issue
- p3sub2 = getfield_gc(p1, descr=nextdescr)
- guard_nonnull_class(p3sub2, ConstClass(node_vtable2)) []
- jump(i1, p1, p3sub2)
+ jump(i1, p1, p3sub)
"""
self.optimize_loop(ops, expected, preamble)
@@ -902,9 +896,7 @@
guard_true(i2b) []
p3 = new_with_vtable(ConstClass(node_vtable))
setfield_gc(p3, i2, descr=nextdescr)
- # XXX: VIRTUALHEAP (see above)
- i3 = getfield_gc(p3, descr=nextdescr)
- jump(p3, i3)
+ jump(p3, i2)
"""
self.optimize_loop(ops, expected, preamble)
@@ -1219,7 +1211,15 @@
setfield_gc(p3, p30, descr=valuedescr)
jump(i29, p30, p3)
"""
- expected = preamble
+ expected = """
+ [i0, p1, p3]
+ i28 = int_add(i0, 1)
+ i29 = int_add(i28, 1)
+ p30 = new_with_vtable(ConstClass(node_vtable))
+ setfield_gc(p3, p30, descr=valuedescr)
+ setfield_gc(p30, i28, descr=nextdescr)
+ jump(i29, p30, p3)
+ """
self.optimize_loop(ops, expected, preamble)
def test_nonvirtual_1(self):
@@ -2408,8 +2408,8 @@
guard_class(p2, ConstClass(node_vtable)) []
p3 = getfield_gc(p1, descr=otherdescr)
guard_class(p3, ConstClass(node_vtable)) []
+ p3a = new_with_vtable(ConstClass(node_vtable))
setfield_gc(p3, p2, descr=otherdescr)
- p3a = new_with_vtable(ConstClass(node_vtable))
escape(p3a)
jump(p3a)
"""
@@ -2421,9 +2421,9 @@
# setfield_gc(p3, p2, descr=otherdescr) # p3a.other = p2a
# p1a = new_with_vtable(ConstClass(node_vtable2))
# p2a = new_with_vtable(ConstClass(node_vtable))
+ p3anew = new_with_vtable(ConstClass(node_vtable))
p2 = new_with_vtable(ConstClass(node_vtable))
setfield_gc(p3a, p2, descr=otherdescr) # p3a.other = p2a
- p3anew = new_with_vtable(ConstClass(node_vtable))
escape(p3anew)
jump(p3anew)
"""
@@ -2458,9 +2458,9 @@
p3 = getfield_gc(p1, descr=otherdescr)
guard_class(p3, ConstClass(node_vtable)) []
# p1a = new_with_vtable(ConstClass(node_vtable2))
+ p3a = new_with_vtable(ConstClass(node_vtable))
p2a = new_with_vtable(ConstClass(node_vtable))
setfield_gc(p3, p2a, descr=otherdescr)
- p3a = new_with_vtable(ConstClass(node_vtable))
escape(p3a)
# setfield_gc(p1a, p2a, descr=nextdescr)
# setfield_gc(p1a, p3a, descr=otherdescr)
@@ -2468,9 +2468,9 @@
"""
expected = """
[p2, p3]
+ p3a = new_with_vtable(ConstClass(node_vtable))
p2a = new_with_vtable(ConstClass(node_vtable))
setfield_gc(p3, p2a, descr=otherdescr)
- p3a = new_with_vtable(ConstClass(node_vtable))
escape(p3a)
jump(p2a, p3a)
"""
@@ -2790,7 +2790,6 @@
self.optimize_loop(ops, expected, preamble)
def test_remove_duplicate_pure_op_ovf_with_lazy_setfield(self):
- py.test.skip('this optimization is not yet supprted')
ops = """
[i1, p1]
i3 = int_add_ovf(i1, 1)
@@ -5961,13 +5960,18 @@
escape(i0)
jump(p1)
"""
- expected = """
+ preamble = """
[p1]
i0 = ptr_eq(p1, NULL)
escape(i0)
- jump(p1)
- """
- self.optimize_strunicode_loop_extradescrs(ops, expected, expected)
+ jump(p1, i0)
+ """
+ expected = """
+ [p1, i0]
+ escape(i0)
+ jump(p1, i0)
+ """
+ self.optimize_strunicode_loop_extradescrs(ops, expected, preamble)
def test_str_equal_none2(self):
ops = """
@@ -5976,13 +5980,18 @@
escape(i0)
jump(p1)
"""
- expected = """
+ preamble = """
[p1]
i0 = ptr_eq(p1, NULL)
escape(i0)
- jump(p1)
- """
- self.optimize_strunicode_loop_extradescrs(ops, expected, expected)
+ jump(p1, i0)
+ """
+ expected = """
+ [p1, i0]
+ escape(i0)
+ jump(p1, i0)
+ """
+ self.optimize_strunicode_loop_extradescrs(ops, expected, preamble)
def test_str_equal_nonnull1(self):
ops = """
@@ -7220,6 +7229,29 @@
"""
self.optimize_loop(ops, expected)
+ def test_heap_cache_forced_virtuals(self):
+ ops = """
+ [i1, i2, p0]
+ p1 = new(descr=ssize)
+ setfield_gc(p1, i1, descr=adescr)
+ setfield_gc(p1, i2, descr=bdescr)
+ call(p0, p1, descr=writeadescr)
+ i3 = getfield_gc(p1, descr=adescr)
+ i4 = getfield_gc(p1, descr=bdescr)
+ jump(i3, i4, p0)
+ """
+ expected = """
+ [i1, i2, p0]
+ p1 = new(descr=ssize)
+ setfield_gc(p1, i1, descr=adescr)
+ call(p0, p1, descr=writeadescr)
+ i3 = getfield_gc(p1, descr=adescr)
+ setfield_gc(p1, i2, descr=bdescr)
+ jump(i3, i2, p0)
+ """
+ self.optimize_loop(ops, expected)
+
+
def test_setarrayitem_followed_by_arraycopy(self):
ops = """
[p1, p2]
@@ -7230,6 +7262,27 @@
"""
self.optimize_loop(ops, ops)
+ def test_heap_cache_virtuals_forced_by_delayed_setfield(self):
+ py.test.skip('not yet supoprted')
+ ops = """
+ [i1, p0]
+ p1 = new(descr=ssize)
+ setfield_gc(p1, i1, descr=valuedescr)
+ setfield_gc(p0, p1, descr=adescr)
+ call(p0, descr=writeadescr)
+ i2 = getfield_gc(p1, descr=valuedescr)
+ jump(i2, p0)
+ """
+ expected = """
+ [i1, p0]
+ p1 = new(descr=ssize)
+ setfield_gc(p1, i1, descr=valuedescr)
+ setfield_gc(p0, p1, descr=adescr)
+ call(p0, descr=writeadescr)
+ jump(i1, p0)
+ """
+ self.optimize_loop(ops, expected)
+
class TestLLtype(OptimizeOptTest, LLtypeMixin):
pass
diff --git a/pypy/jit/metainterp/optimizeopt/unroll.py b/pypy/jit/metainterp/optimizeopt/unroll.py
--- a/pypy/jit/metainterp/optimizeopt/unroll.py
+++ b/pypy/jit/metainterp/optimizeopt/unroll.py
@@ -75,7 +75,6 @@
self.importable_values = {}
self.emitting_dissabled = False
self.emitted_guards = 0
- self.emitted_pure_operations = {}
def ensure_imported(self, value):
if not self.emitting_dissabled and value in self.importable_values:
@@ -96,21 +95,6 @@
new = UnrollableOptimizer(self.metainterp_sd, self.loop)
return self._new(new)
- def remember_emitting_pure(self, op):
- self.emitted_pure_operations[op] = True
-
- def produce_potential_short_preamble_ops(self, sb):
- for op in self.emitted_pure_operations:
- if op.getopnum() == rop.GETARRAYITEM_GC_PURE or \
- op.getopnum() == rop.STRGETITEM or \
- op.getopnum() == rop.UNICODEGETITEM:
- if not self.getvalue(op.getarg(1)).is_constant():
- continue
- sb.add_potential(op)
- for opt in self.optimizations:
- opt.produce_potential_short_preamble_ops(sb)
-
-
class UnrollOptimizer(Optimization):
"""Unroll the loop into two iterations. The first one will
@@ -154,7 +138,7 @@
KillHugeIntBounds(self.optimizer).apply()
- loop.preamble.operations = self.optimizer.newoperations
+ loop.preamble.operations = self.optimizer.get_newoperations()
jump_args = [self.getvalue(a).get_key_box() for a in jump_args]
start_resumedescr = loop.preamble.start_resumedescr.clone_if_mutable()
@@ -167,8 +151,9 @@
virtual_state = modifier.get_virtual_state(jump_args)
values = [self.getvalue(arg) for arg in jump_args]
- inputargs = virtual_state.make_inputargs(values)
- short_inputargs = virtual_state.make_inputargs(values, keyboxes=True)
+ inputargs = virtual_state.make_inputargs(values, self.optimizer)
+ short_inputargs = virtual_state.make_inputargs(values, self.optimizer,
+ keyboxes=True)
self.constant_inputargs = {}
for box in jump_args:
@@ -197,7 +182,7 @@
# operations needed to setup the proper state of those virtuals
# in the peeled loop
inputarg_setup_ops = []
- preamble_optimizer.newoperations = []
+ preamble_optimizer.clear_newoperations()
seen = {}
for box in inputargs:
if box in seen:
@@ -211,9 +196,8 @@
continue
seen[box] = True
value = preamble_optimizer.getvalue(box)
- value.force_box()
- preamble_optimizer.flush()
- inputarg_setup_ops += preamble_optimizer.newoperations
+ value.force_box(preamble_optimizer)
+ inputarg_setup_ops += preamble_optimizer.get_newoperations()
# Setup the state of the new optimizer by emiting the
# short preamble operations and discarding the result
@@ -244,13 +228,13 @@
virtual_state)
loop.inputargs = inputargs
- args = [preamble_optimizer.getvalue(self.short_boxes.original(a)).force_box()\
+ args = [preamble_optimizer.getvalue(self.short_boxes.original(a)).force_box(preamble_optimizer)\
for a in inputargs]
jmp = ResOperation(rop.JUMP, args, None)
jmp.setdescr(loop.token)
loop.preamble.operations.append(jmp)
- loop.operations = self.optimizer.newoperations
+ loop.operations = self.optimizer.get_newoperations()
maxguards = self.optimizer.metainterp_sd.warmrunnerdesc.memory_manager.max_retrace_guards
if self.optimizer.emitted_guards > maxguards:
@@ -335,9 +319,10 @@
assert jumpop
original_jumpargs = jumpop.getarglist()[:]
values = [self.getvalue(arg) for arg in jumpop.getarglist()]
- jumpargs = virtual_state.make_inputargs(values)
+ jumpargs = virtual_state.make_inputargs(values, self.optimizer)
jumpop.initarglist(jumpargs)
- jmp_to_short_args = virtual_state.make_inputargs(values, keyboxes=True)
+ jmp_to_short_args = virtual_state.make_inputargs(values, self.optimizer,
+ keyboxes=True)
self.short_inliner = Inliner(short_inputargs, jmp_to_short_args)
for box, const in self.constant_inputargs.items():
@@ -347,11 +332,11 @@
newop = self.short_inliner.inline_op(op)
self.optimizer.send_extra_operation(newop)
- self.optimizer.flush()
+ newoperations = self.optimizer.get_newoperations()
i = j = 0
- while i < len(self.optimizer.newoperations) or j < len(jumpargs):
- if i == len(self.optimizer.newoperations):
+ while i < len(newoperations) or j < len(jumpargs):
+ if i == len(newoperations):
while j < len(jumpargs):
a = jumpargs[j]
if self.optimizer.loop.logops:
@@ -360,7 +345,7 @@
jumpargs, short_seen)
j += 1
else:
- op = self.optimizer.newoperations[i]
+ op = newoperations[i]
self.boxes_created_this_iteration[op.result] = True
args = op.getarglist()
@@ -375,6 +360,7 @@
self.import_box(a, inputargs, short, short_jumpargs,
jumpargs, short_seen)
i += 1
+ newoperations = self.optimizer.get_newoperations()
jumpop.initarglist(jumpargs)
self.optimizer.send_extra_operation(jumpop)
@@ -468,7 +454,7 @@
inputargs.append(box)
box = newresult
if box in self.optimizer.values:
- box = self.optimizer.values[box].force_box()
+ box = self.optimizer.values[box].force_box(self.optimizer)
jumpargs.append(box)
@@ -483,11 +469,6 @@
if op.getopnum() == rop.JUMP:
loop_token = op.getdescr()
assert isinstance(loop_token, LoopToken)
- # FIXME: Use a tree, similar to the tree formed by the full
- # preamble and it's bridges, instead of a list to save time and
- # memory. This should also allow better behaviour in
- # situations that the is_emittable() chain currently cant
- # handle and the inlining fails unexpectedly belwo.
short = loop_token.short_preamble
if short:
args = op.getarglist()
@@ -523,7 +504,7 @@
values = [self.getvalue(arg)
for arg in op.getarglist()]
- args = sh.virtual_state.make_inputargs(values,
+ args = sh.virtual_state.make_inputargs(values, self.optimizer,
keyboxes=True)
inliner = Inliner(sh.inputargs, args)
diff --git a/pypy/jit/metainterp/optimizeopt/virtualize.py b/pypy/jit/metainterp/optimizeopt/virtualize.py
--- a/pypy/jit/metainterp/optimizeopt/virtualize.py
+++ b/pypy/jit/metainterp/optimizeopt/virtualize.py
@@ -10,13 +10,12 @@
class AbstractVirtualValue(optimizer.OptValue):
- _attrs_ = ('optimizer', 'keybox', 'source_op', '_cached_vinfo')
+ _attrs_ = ('keybox', 'source_op', '_cached_vinfo')
box = None
level = optimizer.LEVEL_NONNULL
_cached_vinfo = None
- def __init__(self, optimizer, keybox, source_op=None):
- self.optimizer = optimizer
+ def __init__(self, keybox, source_op=None):
self.keybox = keybox # only used as a key in dictionaries
self.source_op = source_op # the NEW_WITH_VTABLE/NEW_ARRAY operation
# that builds this box
@@ -29,17 +28,17 @@
return self.keybox
return self.box
- def force_box(self):
+ def force_box(self, optforce):
if self.box is None:
- self.optimizer.forget_numberings(self.keybox)
- self._really_force()
+ optforce.forget_numberings(self.keybox)
+ self._really_force(optforce)
return self.box
- def force_at_end_of_preamble(self, already_forced):
+ def force_at_end_of_preamble(self, already_forced, optforce):
value = already_forced.get(self, None)
if value:
return value
- return OptValue(self.force_box())
+ return OptValue(self.force_box(optforce))
def make_virtual_info(self, modifier, fieldnums):
if fieldnums is None:
@@ -55,7 +54,7 @@
def _make_virtual(self, modifier):
raise NotImplementedError("abstract base")
- def _really_force(self):
+ def _really_force(self, optforce):
raise NotImplementedError("abstract base")
def import_from(self, other, optimizer):
@@ -70,10 +69,11 @@
get_fielddescrlist_cache._annspecialcase_ = "specialize:memo"
class AbstractVirtualStructValue(AbstractVirtualValue):
- _attrs_ = ('_fields', '_cached_sorted_fields')
+ _attrs_ = ('_fields', 'cpu', '_cached_sorted_fields')
- def __init__(self, optimizer, keybox, source_op=None):
- AbstractVirtualValue.__init__(self, optimizer, keybox, source_op)
+ def __init__(self, cpu, keybox, source_op=None):
+ AbstractVirtualValue.__init__(self, keybox, source_op)
+ self.cpu = cpu
self._fields = {}
self._cached_sorted_fields = None
@@ -87,45 +87,44 @@
def _get_descr(self):
raise NotImplementedError
- def _is_immutable_and_filled_with_constants(self):
+ def _is_immutable_and_filled_with_constants(self, optforce):
count = self._get_descr().count_fields_if_immutable()
if count != len(self._fields): # always the case if count == -1
return False
for value in self._fields.itervalues():
- subbox = value.force_box()
+ subbox = value.force_box(optforce)
if not isinstance(subbox, Const):
return False
return True
- def force_at_end_of_preamble(self, already_forced):
+ def force_at_end_of_preamble(self, already_forced, optforce):
if self in already_forced:
return self
already_forced[self] = self
if self._fields:
for ofs in self._fields.keys():
- self._fields[ofs] = self._fields[ofs].force_at_end_of_preamble(already_forced)
+ self._fields[ofs] = self._fields[ofs].force_at_end_of_preamble(already_forced, optforce)
return self
- def _really_force(self):
+ def _really_force(self, optforce):
op = self.source_op
assert op is not None
# ^^^ This case should not occur any more (see test_bug_3).
#
if not we_are_translated():
op.name = 'FORCE ' + self.source_op.name
-
- if self._is_immutable_and_filled_with_constants():
- box = self.optimizer.constant_fold(op)
+
+ if self._is_immutable_and_filled_with_constants(optforce):
+ box = optforce.optimizer.constant_fold(op)
self.make_constant(box)
for ofs, value in self._fields.iteritems():
- subbox = value.force_box()
+ subbox = value.force_box(optforce)
assert isinstance(subbox, Const)
- execute(self.optimizer.cpu, None, rop.SETFIELD_GC,
+ execute(optforce.optimizer.cpu, None, rop.SETFIELD_GC,
ofs, box, subbox)
# keep self._fields, because it's all immutable anyway
else:
- newoperations = self.optimizer.newoperations
- newoperations.append(op)
+ optforce.emit_operation(op)
self.box = box = op.result
#
iteritems = self._fields.iteritems()
@@ -135,10 +134,11 @@
for ofs, value in iteritems:
if value.is_null():
continue
- subbox = value.force_box()
+ subbox = value.force_box(optforce)
op = ResOperation(rop.SETFIELD_GC, [box, subbox], None,
descr=ofs)
- newoperations.append(op)
+
+ optforce.emit_operation(op)
def _get_field_descr_list(self):
_cached_sorted_fields = self._cached_sorted_fields
@@ -155,7 +155,7 @@
else:
lst = self._fields.keys()
sort_descrs(lst)
- cache = get_fielddescrlist_cache(self.optimizer.cpu)
+ cache = get_fielddescrlist_cache(self.cpu)
result = cache.get(lst, None)
if result is None:
cache[lst] = lst
@@ -180,8 +180,8 @@
class VirtualValue(AbstractVirtualStructValue):
level = optimizer.LEVEL_KNOWNCLASS
- def __init__(self, optimizer, known_class, keybox, source_op=None):
- AbstractVirtualStructValue.__init__(self, optimizer, keybox, source_op)
+ def __init__(self, cpu, known_class, keybox, source_op=None):
+ AbstractVirtualStructValue.__init__(self, cpu, keybox, source_op)
assert isinstance(known_class, Const)
self.known_class = known_class
@@ -190,7 +190,7 @@
return modifier.make_virtual(self.known_class, fielddescrs)
def _get_descr(self):
- return vtable2descr(self.optimizer.cpu, self.known_class.getint())
+ return vtable2descr(self.cpu, self.known_class.getint())
def __repr__(self):
cls_name = self.known_class.value.adr.ptr._obj._TYPE._name
@@ -201,8 +201,8 @@
class VStructValue(AbstractVirtualStructValue):
- def __init__(self, optimizer, structdescr, keybox, source_op=None):
- AbstractVirtualStructValue.__init__(self, optimizer, keybox, source_op)
+ def __init__(self, cpu, structdescr, keybox, source_op=None):
+ AbstractVirtualStructValue.__init__(self, cpu, keybox, source_op)
self.structdescr = structdescr
def _make_virtual(self, modifier):
@@ -215,10 +215,10 @@
class VArrayValue(AbstractVirtualValue):
- def __init__(self, optimizer, arraydescr, size, keybox, source_op=None):
- AbstractVirtualValue.__init__(self, optimizer, keybox, source_op)
+ def __init__(self, arraydescr, constvalue, size, keybox, source_op=None):
+ AbstractVirtualValue.__init__(self, keybox, source_op)
self.arraydescr = arraydescr
- self.constvalue = optimizer.new_const_item(arraydescr)
+ self.constvalue = constvalue
self._items = [self.constvalue] * size
def getlength(self):
@@ -232,31 +232,30 @@
assert isinstance(itemvalue, optimizer.OptValue)
self._items[index] = itemvalue
- def force_at_end_of_preamble(self, already_forced):
+ def force_at_end_of_preamble(self, already_forced, optforce):
if self in already_forced:
return self
already_forced[self] = self
for index in range(len(self._items)):
- self._items[index] = self._items[index].force_at_end_of_preamble(already_forced)
+ self._items[index] = self._items[index].force_at_end_of_preamble(already_forced, optforce)
return self
- def _really_force(self):
+ def _really_force(self, optforce):
assert self.source_op is not None
if not we_are_translated():
self.source_op.name = 'FORCE ' + self.source_op.name
- newoperations = self.optimizer.newoperations
- newoperations.append(self.source_op)
+ optforce.emit_operation(self.source_op)
self.box = box = self.source_op.result
for index in range(len(self._items)):
subvalue = self._items[index]
if subvalue is not self.constvalue:
if subvalue.is_null():
continue
- subbox = subvalue.force_box()
+ subbox = subvalue.force_box(optforce)
op = ResOperation(rop.SETARRAYITEM_GC,
[box, ConstInt(index), subbox], None,
descr=self.arraydescr)
- newoperations.append(op)
+ optforce.emit_operation(op)
def get_args_for_fail(self, modifier):
if self.box is None and not modifier.already_seen_virtual(self.keybox):
@@ -279,17 +278,18 @@
return OptVirtualize()
def make_virtual(self, known_class, box, source_op=None):
- vvalue = VirtualValue(self.optimizer, known_class, box, source_op)
+ vvalue = VirtualValue(self.optimizer.cpu, known_class, box, source_op)
self.make_equal_to(box, vvalue)
return vvalue
def make_varray(self, arraydescr, size, box, source_op=None):
- vvalue = VArrayValue(self.optimizer, arraydescr, size, box, source_op)
+ constvalue = self.new_const_item(arraydescr)
+ vvalue = VArrayValue(arraydescr, constvalue, size, box, source_op)
self.make_equal_to(box, vvalue)
return vvalue
def make_vstruct(self, structdescr, box, source_op=None):
- vvalue = VStructValue(self.optimizer, structdescr, box, source_op)
+ vvalue = VStructValue(self.optimizer.cpu, structdescr, box, source_op)
self.make_equal_to(box, vvalue)
return vvalue
@@ -362,7 +362,6 @@
self.make_equal_to(op.result, fieldvalue)
else:
value.ensure_nonnull()
- ###self.heap_op_optimizer.optimize_GETFIELD_GC(op, value)
self.emit_operation(op)
# note: the following line does not mean that the two operations are
@@ -377,7 +376,6 @@
value.setfield(op.getdescr(), fieldvalue)
else:
value.ensure_nonnull()
- ###self.heap_op_optimizer.optimize_SETFIELD_GC(op, value, fieldvalue)
self.emit_operation(op)
def optimize_NEW_WITH_VTABLE(self, op):
@@ -417,7 +415,6 @@
self.make_equal_to(op.result, itemvalue)
return
value.ensure_nonnull()
- ###self.heap_op_optimizer.optimize_GETARRAYITEM_GC(op, value)
self.emit_operation(op)
# note: the following line does not mean that the two operations are
@@ -432,7 +429,6 @@
value.setitem(indexbox.getint(), self.getvalue(op.getarg(2)))
return
value.ensure_nonnull()
- ###self.heap_op_optimizer.optimize_SETARRAYITEM_GC(op, value, fieldvalue)
self.emit_operation(op)
diff --git a/pypy/jit/metainterp/optimizeopt/virtualstate.py b/pypy/jit/metainterp/optimizeopt/virtualstate.py
--- a/pypy/jit/metainterp/optimizeopt/virtualstate.py
+++ b/pypy/jit/metainterp/optimizeopt/virtualstate.py
@@ -30,7 +30,7 @@
def _generate_guards(self, other, box, cpu, extra_guards):
raise InvalidLoop
- def enum_forced_boxes(self, boxes, value):
+ def enum_forced_boxes(self, boxes, value, optimizer):
raise NotImplementedError
def enum(self, virtual_state):
@@ -100,14 +100,14 @@
def _generalization_of(self, other):
raise NotImplementedError
- def enum_forced_boxes(self, boxes, value):
+ def enum_forced_boxes(self, boxes, value, optimizer):
assert isinstance(value, virtualize.AbstractVirtualStructValue)
assert value.is_virtual()
for i in range(len(self.fielddescrs)):
v = value._fields[self.fielddescrs[i]]
s = self.fieldstate[i]
if s.position > self.position:
- s.enum_forced_boxes(boxes, v)
+ s.enum_forced_boxes(boxes, v, optimizer)
def _enum(self, virtual_state):
for s in self.fieldstate:
@@ -177,14 +177,14 @@
return False
return True
- def enum_forced_boxes(self, boxes, value):
+ def enum_forced_boxes(self, boxes, value, optimizer):
assert isinstance(value, virtualize.VArrayValue)
assert value.is_virtual()
for i in range(len(self.fieldstate)):
v = value._items[i]
s = self.fieldstate[i]
if s.position > self.position:
- s.enum_forced_boxes(boxes, v)
+ s.enum_forced_boxes(boxes, v, optimizer)
def _enum(self, virtual_state):
for s in self.fieldstate:
@@ -316,11 +316,11 @@
import pdb; pdb.set_trace()
raise NotImplementedError
- def enum_forced_boxes(self, boxes, value):
+ def enum_forced_boxes(self, boxes, value, optimizer):
if self.level == LEVEL_CONSTANT:
return
assert 0 <= self.position_in_notvirtuals
- boxes[self.position_in_notvirtuals] = value.force_box()
+ boxes[self.position_in_notvirtuals] = value.force_box(optimizer)
def _enum(self, virtual_state):
if self.level == LEVEL_CONSTANT:
@@ -377,11 +377,13 @@
self.state[i].generate_guards(other.state[i], args[i],
cpu, extra_guards, renum)
- def make_inputargs(self, values, keyboxes=False):
+ def make_inputargs(self, values, optimizer, keyboxes=False):
+ if optimizer.optearlyforce:
+ optimizer = optimizer.optearlyforce
assert len(values) == len(self.state)
inputargs = [None] * len(self.notvirtuals)
for i in range(len(values)):
- self.state[i].enum_forced_boxes(inputargs, values[i])
+ self.state[i].enum_forced_boxes(inputargs, values[i], optimizer)
if keyboxes:
for i in range(len(values)):
@@ -434,7 +436,12 @@
def get_virtual_state(self, jump_args):
self.optimizer.force_at_end_of_preamble()
already_forced = {}
- values = [self.getvalue(box).force_at_end_of_preamble(already_forced)
+ if self.optimizer.optearlyforce:
+ opt = self.optimizer.optearlyforce
+ else:
+ opt = self.optimizer
+ values = [self.getvalue(box).force_at_end_of_preamble(already_forced,
+ opt)
for box in jump_args]
for value in values:
diff --git a/pypy/jit/metainterp/optimizeopt/vstring.py b/pypy/jit/metainterp/optimizeopt/vstring.py
--- a/pypy/jit/metainterp/optimizeopt/vstring.py
+++ b/pypy/jit/metainterp/optimizeopt/vstring.py
@@ -43,7 +43,7 @@
class __extend__(optimizer.OptValue):
"""New methods added to the base class OptValue for this file."""
- def getstrlen(self, optimization, mode):
+ def getstrlen(self, string_optimizer, mode):
if mode is mode_string:
s = self.get_constant_string_spec(mode_string)
if s is not None:
@@ -52,12 +52,12 @@
s = self.get_constant_string_spec(mode_unicode)
if s is not None:
return ConstInt(len(s))
- if optimization is None:
+ if string_optimizer is None:
return None
self.ensure_nonnull()
- box = self.force_box()
+ box = self.force_box(string_optimizer)
lengthbox = BoxInt()
- optimization.propagate_forward(ResOperation(mode.STRLEN, [box], lengthbox))
+ string_optimizer.emit_operation(ResOperation(mode.STRLEN, [box], lengthbox))
return lengthbox
@specialize.arg(1)
@@ -68,25 +68,25 @@
else:
return None
- def string_copy_parts(self, optimizer, targetbox, offsetbox, mode):
+ def string_copy_parts(self, string_optimizer, targetbox, offsetbox, mode):
# Copies the pointer-to-string 'self' into the target string
# given by 'targetbox', at the specified offset. Returns the offset
# at the end of the copy.
- lengthbox = self.getstrlen(optimizer, mode)
- srcbox = self.force_box()
- return copy_str_content(optimizer, srcbox, targetbox,
+ lengthbox = self.getstrlen(string_optimizer, mode)
+ srcbox = self.force_box(string_optimizer)
+ return copy_str_content(string_optimizer, srcbox, targetbox,
CONST_0, offsetbox, lengthbox, mode)
class VAbstractStringValue(virtualize.AbstractVirtualValue):
_attrs_ = ('mode',)
- def __init__(self, optimizer, keybox, source_op, mode):
- virtualize.AbstractVirtualValue.__init__(self, optimizer, keybox,
+ def __init__(self, keybox, source_op, mode):
+ virtualize.AbstractVirtualValue.__init__(self, keybox,
source_op)
self.mode = mode
- def _really_force(self):
+ def _really_force(self, optforce):
if self.mode is mode_string:
s = self.get_constant_string_spec(mode_string)
if s is not None:
@@ -101,12 +101,12 @@
return
assert self.source_op is not None
self.box = box = self.source_op.result
- lengthbox = self.getstrlen(self.optimizer, self.mode)
+ lengthbox = self.getstrlen(optforce, self.mode)
op = ResOperation(self.mode.NEWSTR, [lengthbox], box)
if not we_are_translated():
op.name = 'FORCE'
- self.optimizer.emit_operation(op)
- self.string_copy_parts(self.optimizer, box, CONST_0, self.mode)
+ optforce.emit_operation(op)
+ self.string_copy_parts(optforce, box, CONST_0, self.mode)
class VStringPlainValue(VAbstractStringValue):
@@ -140,20 +140,20 @@
return mode.emptystr.join([mode.chr(c.box.getint())
for c in self._chars])
- def string_copy_parts(self, optimizer, targetbox, offsetbox, mode):
+ def string_copy_parts(self, string_optimizer, targetbox, offsetbox, mode):
if not self.is_virtual() and targetbox is not self.box:
- lengthbox = self.getstrlen(optimizer, mode)
- srcbox = self.force_box()
- return copy_str_content(optimizer, srcbox, targetbox,
+ lengthbox = self.getstrlen(string_optimizer, mode)
+ srcbox = self.force_box(string_optimizer)
+ return copy_str_content(string_optimizer, srcbox, targetbox,
CONST_0, offsetbox, lengthbox, mode)
for i in range(len(self._chars)):
- charbox = self._chars[i].force_box()
+ charbox = self._chars[i].force_box(string_optimizer)
if not (isinstance(charbox, Const) and charbox.same_constant(CONST_0)):
- optimizer.emit_operation(ResOperation(mode.STRSETITEM, [targetbox,
- offsetbox,
- charbox],
+ string_optimizer.emit_operation(ResOperation(mode.STRSETITEM, [targetbox,
+ offsetbox,
+ charbox],
None))
- offsetbox = _int_add(optimizer, offsetbox, CONST_1)
+ offsetbox = _int_add(string_optimizer, offsetbox, CONST_1)
return offsetbox
def get_args_for_fail(self, modifier):
@@ -187,16 +187,16 @@
self.left = left
self.right = right
- def getstrlen(self, optimizer, mode):
+ def getstrlen(self, string_optimizer, mode):
if self.lengthbox is None:
- len1box = self.left.getstrlen(optimizer, mode)
+ len1box = self.left.getstrlen(string_optimizer, mode)
if len1box is None:
return None
- len2box = self.right.getstrlen(optimizer, mode)
+ len2box = self.right.getstrlen(string_optimizer, mode)
if len2box is None:
return None
- self.lengthbox = _int_add(optimizer, len1box, len2box)
- # ^^^ may still be None, if optimizer is None
+ self.lengthbox = _int_add(string_optimizer, len1box, len2box)
+ # ^^^ may still be None, if string_optimizer is None
return self.lengthbox
@specialize.arg(1)
@@ -209,10 +209,10 @@
return None
return s1 + s2
- def string_copy_parts(self, optimizer, targetbox, offsetbox, mode):
- offsetbox = self.left.string_copy_parts(optimizer, targetbox,
+ def string_copy_parts(self, string_optimizer, targetbox, offsetbox, mode):
+ offsetbox = self.left.string_copy_parts(string_optimizer, targetbox,
offsetbox, mode)
- offsetbox = self.right.string_copy_parts(optimizer, targetbox,
+ offsetbox = self.right.string_copy_parts(string_optimizer, targetbox,
offsetbox, mode)
return offsetbox
@@ -251,8 +251,8 @@
self.vstart = vstart
self.vlength = vlength
- def getstrlen(self, _, mode):
- return self.vlength.force_box()
+ def getstrlen(self, optforce, mode):
+ return self.vlength.force_box(optforce)
@specialize.arg(1)
def get_constant_string_spec(self, mode):
@@ -267,11 +267,11 @@
return s1[start : start + length]
return None
- def string_copy_parts(self, optimizer, targetbox, offsetbox, mode):
- lengthbox = self.getstrlen(optimizer, mode)
- return copy_str_content(optimizer,
- self.vstr.force_box(), targetbox,
- self.vstart.force_box(), offsetbox,
+ def string_copy_parts(self, string_optimizer, targetbox, offsetbox, mode):
+ lengthbox = self.getstrlen(string_optimizer, mode)
+ return copy_str_content(string_optimizer,
+ self.vstr.force_box(string_optimizer), targetbox,
+ self.vstart.force_box(string_optimizer), offsetbox,
lengthbox, mode)
def get_args_for_fail(self, modifier):
@@ -300,7 +300,7 @@
return modifier.make_vstrslice(self.mode is mode_unicode)
-def copy_str_content(optimizer, srcbox, targetbox,
+def copy_str_content(string_optimizer, srcbox, targetbox,
srcoffsetbox, offsetbox, lengthbox, mode, need_next_offset=True):
if isinstance(srcbox, ConstPtr) and isinstance(srcoffsetbox, Const):
M = 5
@@ -310,26 +310,26 @@
# up to M characters are done "inline", i.e. with STRGETITEM/STRSETITEM
# instead of just a COPYSTRCONTENT.
for i in range(lengthbox.value):
- charbox = _strgetitem(optimizer, srcbox, srcoffsetbox, mode)
- srcoffsetbox = _int_add(optimizer, srcoffsetbox, CONST_1)
- optimizer.emit_operation(ResOperation(mode.STRSETITEM, [targetbox,
- offsetbox,
- charbox],
+ charbox = _strgetitem(string_optimizer, srcbox, srcoffsetbox, mode)
+ srcoffsetbox = _int_add(string_optimizer, srcoffsetbox, CONST_1)
+ string_optimizer.emit_operation(ResOperation(mode.STRSETITEM, [targetbox,
+ offsetbox,
+ charbox],
None))
- offsetbox = _int_add(optimizer, offsetbox, CONST_1)
+ offsetbox = _int_add(string_optimizer, offsetbox, CONST_1)
else:
if need_next_offset:
- nextoffsetbox = _int_add(optimizer, offsetbox, lengthbox)
+ nextoffsetbox = _int_add(string_optimizer, offsetbox, lengthbox)
else:
nextoffsetbox = None
op = ResOperation(mode.COPYSTRCONTENT, [srcbox, targetbox,
srcoffsetbox, offsetbox,
lengthbox], None)
- optimizer.emit_operation(op)
+ string_optimizer.emit_operation(op)
offsetbox = nextoffsetbox
return offsetbox
-def _int_add(optimizer, box1, box2):
+def _int_add(string_optimizer, box1, box2):
if isinstance(box1, ConstInt):
if box1.value == 0:
return box2
@@ -337,23 +337,23 @@
return ConstInt(box1.value + box2.value)
elif isinstance(box2, ConstInt) and box2.value == 0:
return box1
- if optimizer is None:
+ if string_optimizer is None:
return None
resbox = BoxInt()
- optimizer.propagate_forward(ResOperation(rop.INT_ADD, [box1, box2], resbox))
+ string_optimizer.emit_operation(ResOperation(rop.INT_ADD, [box1, box2], resbox))
return resbox
-def _int_sub(optimizer, box1, box2):
+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()
- optimizer.propagate_forward(ResOperation(rop.INT_SUB, [box1, box2], resbox))
+ string_optimizer.emit_operation(ResOperation(rop.INT_SUB, [box1, box2], resbox))
return resbox
-def _strgetitem(optimizer, strbox, indexbox, mode):
+def _strgetitem(string_optimizer, strbox, indexbox, mode):
if isinstance(strbox, ConstPtr) and isinstance(indexbox, ConstInt):
if mode is mode_string:
s = strbox.getref(lltype.Ptr(rstr.STR))
@@ -362,30 +362,28 @@
s = strbox.getref(lltype.Ptr(rstr.UNICODE))
return ConstInt(ord(s.chars[indexbox.getint()]))
resbox = BoxInt()
- optimizer.propagate_forward(ResOperation(mode.STRGETITEM, [strbox, indexbox],
- resbox))
+ string_optimizer.emit_operation(ResOperation(mode.STRGETITEM, [strbox, indexbox],
+ resbox))
return resbox
class OptString(optimizer.Optimization):
"Handling of strings and unicodes."
- enabled = True
-
def new(self):
return OptString()
def make_vstring_plain(self, box, source_op, mode):
- vvalue = VStringPlainValue(self.optimizer, box, source_op, mode)
+ vvalue = VStringPlainValue(box, source_op, mode)
self.make_equal_to(box, vvalue)
return vvalue
def make_vstring_concat(self, box, source_op, mode):
- vvalue = VStringConcatValue(self.optimizer, box, source_op, mode)
+ vvalue = VStringConcatValue(box, source_op, mode)
self.make_equal_to(box, vvalue)
return vvalue
def make_vstring_slice(self, box, source_op, mode):
- vvalue = VStringSliceValue(self.optimizer, box, source_op, mode)
+ vvalue = VStringSliceValue(box, source_op, mode)
self.make_equal_to(box, vvalue)
return vvalue
@@ -435,9 +433,9 @@
value.ensure_nonnull()
#
if value.is_virtual() and isinstance(value, VStringSliceValue):
- fullindexbox = _int_add(self.optimizer,
- value.vstart.force_box(),
- vindex.force_box())
+ fullindexbox = _int_add(self,
+ value.vstart.force_box(self),
+ vindex.force_box(self))
value = value.vstr
vindex = self.getvalue(fullindexbox)
#
@@ -449,7 +447,7 @@
if res is not optimizer.CVAL_UNINITIALIZED_ZERO:
return res
#
- resbox = _strgetitem(self.optimizer, value.force_box(), vindex.force_box(), mode)
+ resbox = _strgetitem(self, value.force_box(self), vindex.force_box(self), mode)
return self.getvalue(resbox)
def optimize_STRLEN(self, op):
@@ -459,7 +457,7 @@
def _optimize_STRLEN(self, op, mode):
value = self.getvalue(op.getarg(0))
- lengthbox = value.getstrlen(self.optimizer, mode)
+ lengthbox = value.getstrlen(self, mode)
self.make_equal_to(op.result, self.getvalue(lengthbox))
def optimize_COPYSTRCONTENT(self, op):
@@ -477,12 +475,12 @@
if length.is_constant() and length.box.getint() == 0:
return
- copy_str_content(self.optimizer,
- src.force_box(),
- dst.force_box(),
- srcstart.force_box(),
- dststart.force_box(),
- length.force_box(),
+ copy_str_content(self,
+ src.force_box(self),
+ dst.force_box(self),
+ srcstart.force_box(self),
+ dststart.force_box(self),
+ length.force_box(self),
mode, need_next_offset=False
)
@@ -508,6 +506,8 @@
return
self.emit_operation(op)
+ optimize_CALL_PURE = optimize_CALL
+
def opt_call_str_STR2UNICODE(self, op):
# Constant-fold unicode("constant string").
# More generally, supporting non-constant but virtual cases is
@@ -547,16 +547,16 @@
return True
#
vstr.ensure_nonnull()
- lengthbox = _int_sub(self.optimizer, vstop.force_box(),
- vstart.force_box())
+ lengthbox = _int_sub(self, vstop.force_box(self),
+ vstart.force_box(self))
#
if isinstance(vstr, VStringSliceValue):
# double slicing s[i:j][k:l]
vintermediate = vstr
vstr = vintermediate.vstr
- startbox = _int_add(self.optimizer,
- vintermediate.vstart.force_box(),
- vstart.force_box())
+ startbox = _int_add(self,
+ vintermediate.vstart.force_box(self),
+ vstart.force_box(self))
vstart = self.getvalue(startbox)
#
value = self.make_vstring_slice(op.result, op, mode)
@@ -594,8 +594,8 @@
do = EffectInfo.OS_STREQ_LENGTHOK
else:
do = EffectInfo.OS_STREQ_NONNULL
- self.generate_modified_call(do, [v1.force_box(),
- v2.force_box()], op.result, mode)
+ self.generate_modified_call(do, [v1.force_box(self),
+ v2.force_box(self)], op.result, mode)
return True
return False
@@ -603,7 +603,7 @@
l2box = v2.getstrlen(None, mode)
if isinstance(l2box, ConstInt):
if l2box.value == 0:
- lengthbox = v1.getstrlen(self.optimizer, mode)
+ lengthbox = v1.getstrlen(self, mode)
seo = self.optimizer.send_extra_operation
seo(ResOperation(rop.INT_EQ, [lengthbox, CONST_0], resultbox))
return True
@@ -614,17 +614,17 @@
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(),
- vchar2.force_box()],
+ 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(),
- v1.vstart.force_box(),
- v1.vlength.force_box(),
- vchar.force_box()],
+ 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
#
@@ -635,10 +635,10 @@
if v1.is_null():
self.make_constant(resultbox, CONST_1)
return True
- op = ResOperation(rop.PTR_EQ, [v1.force_box(),
+ op = ResOperation(rop.PTR_EQ, [v1.force_box(self),
llhelper.CONST_NULL],
resultbox)
- self.optimizer.emit_operation(op)
+ self.emit_operation(op)
return True
#
return False
@@ -652,8 +652,8 @@
do = EffectInfo.OS_STREQ_NONNULL_CHAR
else:
do = EffectInfo.OS_STREQ_CHECKNULL_CHAR
- self.generate_modified_call(do, [v1.force_box(),
- vchar.force_box()], resultbox,
+ self.generate_modified_call(do, [v1.force_box(self),
+ vchar.force_box(self)], resultbox,
mode)
return True
#
@@ -662,10 +662,10 @@
do = EffectInfo.OS_STREQ_SLICE_NONNULL
else:
do = EffectInfo.OS_STREQ_SLICE_CHECKNULL
- self.generate_modified_call(do, [v1.vstr.force_box(),
- v1.vstart.force_box(),
- v1.vlength.force_box(),
- v2.force_box()], resultbox, mode)
+ self.generate_modified_call(do, [v1.vstr.force_box(self),
+ v1.vstart.force_box(self),
+ v1.vlength.force_box(self),
+ v2.force_box(self)], resultbox, mode)
return True
return False
@@ -675,16 +675,11 @@
calldescr, func = cic.callinfo_for_oopspec(oopspecindex)
op = ResOperation(rop.CALL, [ConstInt(func)] + args, result,
descr=calldescr)
- self.optimizer.emit_operation(op)
+ self.emit_operation(op)
def propagate_forward(self, op):
- if not self.enabled:
- self.emit_operation(op)
- return
-
dispatch_opt(self, op)
-
dispatch_opt = make_dispatcher_method(OptString, 'optimize_',
default=OptString.emit_operation)
diff --git a/pypy/jit/metainterp/resoperation.py b/pypy/jit/metainterp/resoperation.py
--- a/pypy/jit/metainterp/resoperation.py
+++ b/pypy/jit/metainterp/resoperation.py
@@ -221,7 +221,6 @@
newop.setfailargs(self.getfailargs())
return newop
-
# ============
# arity mixins
# ============
diff --git a/pypy/jit/metainterp/test/test_ajit.py b/pypy/jit/metainterp/test/test_ajit.py
--- a/pypy/jit/metainterp/test/test_ajit.py
+++ b/pypy/jit/metainterp/test/test_ajit.py
@@ -2956,6 +2956,18 @@
assert res == f(32)
self.check_loops(arraylen_gc=2)
+ def test_ulonglong_mod(self):
+ myjitdriver = JitDriver(greens = [], reds = ['n', 'sa', 'i'])
+ def f(n):
+ sa = i = rffi.cast(rffi.ULONGLONG, 1)
+ while i < rffi.cast(rffi.ULONGLONG, n):
+ myjitdriver.jit_merge_point(sa=sa, n=n, i=i)
+ sa += sa % i
+ i += 1
+ res = self.meta_interp(f, [32])
+ assert res == f(32)
+
+
class TestOOtype(BasicTests, OOJitMixin):
def test_oohash(self):
diff --git a/pypy/jit/metainterp/test/test_longlong.py b/pypy/jit/metainterp/test/test_longlong.py
--- a/pypy/jit/metainterp/test/test_longlong.py
+++ b/pypy/jit/metainterp/test/test_longlong.py
@@ -118,6 +118,26 @@
res = self.interp_operations(f, [1000000000])
assert res == 123500000000.0
+ def test_floats_negative(self):
+ def f(i):
+ # i == 1000000000
+ f = i * -123.5
+ n = r_longlong(f)
+ compare(n, -29, 1054051584)
+ return float(n)
+ res = self.interp_operations(f, [1000000000])
+ assert res == -123500000000.0
+
+ def test_floats_ulonglong(self):
+ def f(i):
+ # i == 1000000000
+ f = i * 12350000000.0
+ n = r_ulonglong(f)
+ compare(n, -1419508847, 538116096)
+ return float(n)
+ res = self.interp_operations(f, [1000000000])
+ assert res == 12350000000000000000.0
+
def test_unsigned_compare_ops(self):
def f(n1, n2):
# n == 30002000000000
diff --git a/pypy/jit/metainterp/test/test_resume.py b/pypy/jit/metainterp/test/test_resume.py
--- a/pypy/jit/metainterp/test/test_resume.py
+++ b/pypy/jit/metainterp/test/test_resume.py
@@ -576,8 +576,9 @@
class FakeOptimizer_VirtualValue(object):
- class cpu:
- pass
+ class optimizer:
+ class cpu:
+ pass
fakeoptimizer = FakeOptimizer_VirtualValue()
def ConstAddr(addr, cpu): # compatibility
@@ -1135,12 +1136,7 @@
modifier.liveboxes = {}
modifier.vfieldboxes = {}
- class FakeOptimizer(object):
- class cpu:
- pass
- def new_const_item(self, descr):
- return None
- v2 = VArrayValue(FakeOptimizer(), LLtypeMixin.arraydescr, 2, b2s)
+ v2 = VArrayValue(LLtypeMixin.arraydescr, None, 2, b2s)
v2._items = [b4s, c1s]
modifier.register_virtual_fields(b2s, [b4s, c1s])
liveboxes = []
diff --git a/pypy/jit/metainterp/test/test_string.py b/pypy/jit/metainterp/test/test_string.py
--- a/pypy/jit/metainterp/test/test_string.py
+++ b/pypy/jit/metainterp/test/test_string.py
@@ -484,6 +484,30 @@
return len(sa.val)
assert self.meta_interp(f, ['a']) == f('a')
+ def test_string_comepare_quasiimmutable(self):
+ class Sys(object):
+ _immutable_fields_ = ["defaultencoding?"]
+ def __init__(self, s):
+ self.defaultencoding = s
+ _str = self._str
+ sys = Sys(_str('ascii'))
+ mydriver = JitDriver(reds = ['n', 'sa'], greens = [])
+ def f(n):
+ sa = 0
+ sys.defaultencoding = _str('ascii')
+ while n:
+ mydriver.jit_merge_point(n=n, sa=sa)
+ if sys.defaultencoding == _str('ascii'):
+ sa += 1
+ n -= 1
+ sys.defaultencoding = _str('utf-8')
+ return sa
+ assert self.meta_interp(f, [8]) == f(8)
+ self.check_loops({'int_add': 1, 'guard_true': 1, 'int_sub': 1,
+ 'jump': 1, 'int_is_true': 1,
+ 'guard_not_invalidated': 1})
+
+
#class TestOOtype(StringTests, OOJitMixin):
# CALL = "oosend"
# CALL_PURE = "oosend_pure"
diff --git a/pypy/jit/metainterp/test/test_virtualstate.py b/pypy/jit/metainterp/test/test_virtualstate.py
--- a/pypy/jit/metainterp/test/test_virtualstate.py
+++ b/pypy/jit/metainterp/test/test_virtualstate.py
@@ -431,7 +431,7 @@
class BaseTestBridges(BaseTest):
- enable_opts = "intbounds:rewrite:virtualize:string:heap:unroll"
+ enable_opts = "intbounds:rewrite:virtualize:string:pure:heap:unroll"
def _do_optimize_bridge(self, bridge, call_pure_results):
from pypy.jit.metainterp.optimizeopt import optimize_bridge_1, build_opt_chain
diff --git a/pypy/module/_numpy/__init__.py b/pypy/module/_numpy/__init__.py
--- a/pypy/module/_numpy/__init__.py
+++ b/pypy/module/_numpy/__init__.py
@@ -23,6 +23,7 @@
("arcsin", "arcsin"),
("arctan", "arctan"),
("arcsinh", "arcsinh"),
+ ("arctanh", "arctanh"),
("copysign", "copysign"),
("cos", "cos"),
("divide", "divide"),
diff --git a/pypy/module/_numpy/app_numpy.py b/pypy/module/_numpy/app_numpy.py
deleted file mode 100644
diff --git a/pypy/module/_numpy/interp_dtype.py b/pypy/module/_numpy/interp_dtype.py
--- a/pypy/module/_numpy/interp_dtype.py
+++ b/pypy/module/_numpy/interp_dtype.py
@@ -7,13 +7,14 @@
from pypy.interpreter.typedef import TypeDef, interp_attrproperty, GetSetProperty
from pypy.module._numpy import signature
from pypy.objspace.std.floatobject import float2string
-from pypy.rlib import rfloat
-from pypy.rlib.rarithmetic import widen
+from pypy.rlib import rarithmetic, rfloat
+from pypy.rlib.rarithmetic import LONG_BIT, widen
from pypy.rlib.objectmodel import specialize, enforceargs
from pypy.rlib.unroll import unrolling_iterable
from pypy.rpython.lltypesystem import lltype, rffi
+UNSIGNEDLTR = "u"
SIGNEDLTR = "i"
BOOLLTR = "b"
FLOATINGLTR = "f"
@@ -61,7 +62,10 @@
self.val = val
def wrap(self, space):
- return space.wrap(self.val)
+ val = self.val
+ if valtype is rarithmetic.r_singlefloat:
+ val = float(val)
+ return space.wrap(val)
def convert_to(self, dtype):
return dtype.adapt_val(self.val)
@@ -145,7 +149,7 @@
return self.adapt_val(func(self, self.for_computation(self.unbox(v))))
return impl
-class ArithmaticTypeMixin(object):
+class ArithmeticTypeMixin(object):
_mixin_ = True
@binop
@@ -200,11 +204,17 @@
return v1 >= v2
-class FloatArithmeticDtype(ArithmaticTypeMixin):
+class FloatArithmeticDtype(ArithmeticTypeMixin):
_mixin_ = True
+ def unwrap(self, space, w_item):
+ return self.adapt_val(space.float_w(space.float(w_item)))
+
def for_computation(self, v):
- return v
+ return float(v)
+
+ def str_format(self, item):
+ return float2string(self.for_computation(self.unbox(item)), 'g', rfloat.DTSF_STR_PRECISION)
@binop
def mod(self, v1, v2):
@@ -250,12 +260,12 @@
return math.tan(v)
@unaryop
def arcsin(self, v):
- if v < -1.0 or v > 1.0:
+ if not -1.0 <= v <= 1.0:
return rfloat.NAN
return math.asin(v)
@unaryop
def arccos(self, v):
- if v < -1.0 or v > 1.0:
+ if not -1.0 <= v <= 1.0:
return rfloat.NAN
return math.acos(v)
@unaryop
@@ -266,7 +276,15 @@
def arcsinh(self, v):
return math.asinh(v)
-class IntegerArithmeticDtype(ArithmaticTypeMixin):
+ @unaryop
+ def arctanh(self, v):
+ if v == 1.0 or v == -1.0:
+ return math.copysign(rfloat.INFINITY, v)
+ if not -1.0 < v < 1.0:
+ return rfloat.NAN
+ return math.atanh(v)
+
+class IntegerArithmeticDtype(ArithmeticTypeMixin):
_mixin_ = True
def unwrap(self, space, w_item):
@@ -275,10 +293,16 @@
def for_computation(self, v):
return widen(v)
+ def str_format(self, item):
+ return str(widen(self.unbox(item)))
+
@binop
def mod(self, v1, v2):
return v1 % v2
+class SignedIntegerArithmeticDtype(IntegerArithmeticDtype):
+ _mixin_ = True
+
@unaryop
def sign(self, v):
if v > 0:
@@ -289,17 +313,21 @@
assert v == 0
return 0
- def str_format(self, item):
- return str(widen(self.unbox(item)))
+class UnsignedIntegerArithmeticDtype(IntegerArithmeticDtype):
+ _mixin_ = True
+
+ @unaryop
+ def sign(self, v):
+ return int(v != 0)
W_BoolDtype = create_low_level_dtype(
num = 0, kind = BOOLLTR, name = "bool",
- aliases = ["?"],
+ aliases = ["?", "bool", "bool8"],
applevel_types = ["bool"],
T = lltype.Bool,
valtype = bool,
)
-class W_BoolDtype(IntegerArithmeticDtype, W_BoolDtype):
+class W_BoolDtype(SignedIntegerArithmeticDtype, W_BoolDtype):
def unwrap(self, space, w_item):
return self.adapt_val(space.is_true(w_item))
@@ -312,69 +340,140 @@
W_Int8Dtype = create_low_level_dtype(
num = 1, kind = SIGNEDLTR, name = "int8",
- aliases = ["int8"],
+ aliases = ["b", "int8", "i1"],
applevel_types = [],
T = rffi.SIGNEDCHAR,
valtype = rffi.SIGNEDCHAR._type,
expected_size = 1,
)
-class W_Int8Dtype(IntegerArithmeticDtype, W_Int8Dtype):
+class W_Int8Dtype(SignedIntegerArithmeticDtype, W_Int8Dtype):
+ pass
+
+W_UInt8Dtype = create_low_level_dtype(
+ num = 2, kind = UNSIGNEDLTR, name = "uint8",
+ aliases = ["B", "uint8", "I1"],
+ applevel_types = [],
+ T = rffi.UCHAR,
+ valtype = rffi.UCHAR._type,
+ expected_size = 1,
+)
+class W_UInt8Dtype(UnsignedIntegerArithmeticDtype, W_UInt8Dtype):
pass
W_Int16Dtype = create_low_level_dtype(
num = 3, kind = SIGNEDLTR, name = "int16",
- aliases = ["int16"],
+ aliases = ["h", "int16", "i2"],
applevel_types = [],
T = rffi.SHORT,
valtype = rffi.SHORT._type,
expected_size = 2,
)
-class W_Int16Dtype(IntegerArithmeticDtype, W_Int16Dtype):
+class W_Int16Dtype(SignedIntegerArithmeticDtype, W_Int16Dtype):
+ pass
+
+W_UInt16Dtype = create_low_level_dtype(
+ num = 4, kind = UNSIGNEDLTR, name = "uint16",
+ aliases = ["H", "uint16", "I2"],
+ applevel_types = [],
+ T = rffi.USHORT,
+ valtype = rffi.USHORT._type,
+ expected_size = 2,
+)
+class W_UInt16Dtype(UnsignedIntegerArithmeticDtype, W_UInt16Dtype):
pass
W_Int32Dtype = create_low_level_dtype(
num = 5, kind = SIGNEDLTR, name = "int32",
- aliases = ["i"],
+ aliases = ["i", "int32", "i4"],
applevel_types = [],
T = rffi.INT,
valtype = rffi.INT._type,
expected_size = 4,
)
-class W_Int32Dtype(IntegerArithmeticDtype, W_Int32Dtype):
+class W_Int32Dtype(SignedIntegerArithmeticDtype, W_Int32Dtype):
+ pass
+
+W_UInt32Dtype = create_low_level_dtype(
+ num = 6, kind = UNSIGNEDLTR, name = "uint32",
+ aliases = ["I", "uint32", "I4"],
+ applevel_types = [],
+ T = rffi.UINT,
+ valtype = rffi.UINT._type,
+ expected_size = 4,
+)
+class W_UInt32Dtype(UnsignedIntegerArithmeticDtype, W_UInt32Dtype):
pass
W_Int64Dtype = create_low_level_dtype(
num = 9, kind = SIGNEDLTR, name = "int64",
- aliases = [],
+ aliases = ["q", "int64", "i8"],
applevel_types = ["long"],
T = rffi.LONGLONG,
valtype = rffi.LONGLONG._type,
expected_size = 8,
)
-class W_Int64Dtype(IntegerArithmeticDtype, W_Int64Dtype):
+class W_Int64Dtype(SignedIntegerArithmeticDtype, W_Int64Dtype):
+ pass
+
+W_UInt64Dtype = create_low_level_dtype(
+ num = 10, kind = UNSIGNEDLTR, name = "uint64",
+ aliases = ["Q", "uint64", "I8"],
+ applevel_types = [],
+ T = rffi.ULONGLONG,
+ valtype = rffi.ULONGLONG._type,
+ expected_size = 8,
+)
+class W_UInt64Dtype(UnsignedIntegerArithmeticDtype, W_UInt64Dtype):
+ pass
+
+if LONG_BIT == 32:
+ class W_LongDtype(W_Int32Dtype):
+ pass
+
+ class W_ULongDtype(W_UInt32Dtype):
+ pass
+else:
+ class W_LongDtype(W_Int64Dtype):
+ pass
+
+ class W_ULongDtype(W_UInt64Dtype):
+ pass
+
+W_LongDtype.num = 7
+W_LongDtype.aliases = ["l"]
+W_LongDtype.applevel_types = ["int"]
+W_ULongDtype.num = 8
+W_ULongDtype.aliases = ["L"]
+
+W_Float32Dtype = create_low_level_dtype(
+ num = 11, kind = FLOATINGLTR, name = "float32",
+ aliases = ["f", "float32", "f4"],
+ applevel_types = [],
+ T = lltype.SingleFloat,
+ valtype = rarithmetic.r_singlefloat,
+ expected_size = 4,
+)
+class W_Float32Dtype(FloatArithmeticDtype, W_Float32Dtype):
pass
W_Float64Dtype = create_low_level_dtype(
num = 12, kind = FLOATINGLTR, name = "float64",
- aliases = [],
+ aliases = ["d", "float64", "f8"],
applevel_types = ["float"],
T = lltype.Float,
valtype = float,
expected_size = 8,
)
class W_Float64Dtype(FloatArithmeticDtype, W_Float64Dtype):
- def unwrap(self, space, w_item):
- return self.adapt_val(space.float_w(space.float(w_item)))
-
- def str_format(self, item):
- return float2string(self.unbox(item), 'g', rfloat.DTSF_STR_PRECISION)
+ pass
ALL_DTYPES = [
W_BoolDtype,
- W_Int8Dtype, W_Int16Dtype, W_Int32Dtype, W_Int64Dtype,
- W_Float64Dtype
+ W_Int8Dtype, W_UInt8Dtype, W_Int16Dtype, W_UInt16Dtype,
+ W_Int32Dtype, W_UInt32Dtype, W_LongDtype, W_ULongDtype,
+ W_Int64Dtype, W_UInt64Dtype,
+ W_Float32Dtype, W_Float64Dtype,
]
-
dtypes_by_alias = unrolling_iterable([
(alias, dtype)
for dtype in ALL_DTYPES
@@ -399,7 +498,7 @@
num = interp_attrproperty("num", cls=W_Dtype),
kind = interp_attrproperty("kind", cls=W_Dtype),
+ itemsize = interp_attrproperty("num_bytes", cls=W_Dtype),
shape = GetSetProperty(W_Dtype.descr_get_shape),
- itemsize = interp_attrproperty("num_bytes", cls=W_Dtype),
)
W_Dtype.typedef.acceptable_as_base_class = False
diff --git a/pypy/module/_numpy/interp_ufuncs.py b/pypy/module/_numpy/interp_ufuncs.py
--- a/pypy/module/_numpy/interp_ufuncs.py
+++ b/pypy/module/_numpy/interp_ufuncs.py
@@ -4,6 +4,7 @@
from pypy.interpreter.typedef import TypeDef, GetSetProperty, interp_attrproperty
from pypy.module._numpy import interp_dtype, signature
from pypy.rlib import jit
+from pypy.rlib.rarithmetic import LONG_BIT
from pypy.tool.sourcetools import func_with_new_name
@@ -180,23 +181,56 @@
# Everything promotes to float, and bool promotes to everything.
if dt2.kind == interp_dtype.FLOATINGLTR or dt1.kind == interp_dtype.BOOLLTR:
+ # Float32 + 8-bit int = Float64
+ if dt2.num == 11 and dt1.num_bytes >= 4:
+ return space.fromcache(interp_dtype.W_Float64Dtype)
return dt2
- assert False
+ # for now this means mixing signed and unsigned
+ if dt2.kind == interp_dtype.SIGNEDLTR:
+ # if dt2 has a greater number of bytes, then just go with it
+ if dt1.num_bytes < dt2.num_bytes:
+ return dt2
+ # we need to promote both dtypes
+ dtypenum = dt2.num + 2
+ else:
+ # increase to the next signed type (or to float)
+ dtypenum = dt2.num + 1
+ # UInt64 + signed = Float64
+ if dt2.num == 10:
+ dtypenum += 1
+ newdtype = interp_dtype.ALL_DTYPES[dtypenum]
+
+ if newdtype.num_bytes > dt2.num_bytes or newdtype.kind == interp_dtype.FLOATINGLTR:
+ return space.fromcache(newdtype)
+ else:
+ # we only promoted to long on 32-bit or to longlong on 64-bit
+ # this is really for dealing with the Long and Ulong dtypes
+ if LONG_BIT == 32:
+ dtypenum += 2
+ else:
+ dtypenum += 3
+ return space.fromcache(interp_dtype.ALL_DTYPES[dtypenum])
def find_unaryop_result_dtype(space, dt, promote_to_float=False,
promote_bools=False, promote_to_largest=False):
if promote_bools and (dt.kind == interp_dtype.BOOLLTR):
return space.fromcache(interp_dtype.W_Int8Dtype)
if promote_to_float:
+ if dt.kind == interp_dtype.FLOATINGLTR:
+ return dt
+ if dt.num >= 5:
+ return space.fromcache(interp_dtype.W_Float64Dtype)
for bytes, dtype in interp_dtype.dtypes_by_num_bytes:
- if dtype.kind == interp_dtype.FLOATINGLTR and dtype.num_bytes >= dt.num_bytes:
+ if dtype.kind == interp_dtype.FLOATINGLTR and dtype.num_bytes > dt.num_bytes:
return space.fromcache(dtype)
if promote_to_largest:
if dt.kind == interp_dtype.BOOLLTR or dt.kind == interp_dtype.SIGNEDLTR:
return space.fromcache(interp_dtype.W_Int64Dtype)
elif dt.kind == interp_dtype.FLOATINGLTR:
return space.fromcache(interp_dtype.W_Float64Dtype)
+ elif dt.kind == interp_dtype.UNSIGNEDLTR:
+ return space.fromcache(interp_dtype.W_UInt64Dtype)
else:
assert False
return dt
@@ -205,15 +239,23 @@
w_type = space.type(w_obj)
bool_dtype = space.fromcache(interp_dtype.W_BoolDtype)
+ long_dtype = space.fromcache(interp_dtype.W_LongDtype)
int64_dtype = space.fromcache(interp_dtype.W_Int64Dtype)
if space.is_w(w_type, space.w_bool):
- if current_guess is None:
+ if current_guess is None or current_guess is bool_dtype:
return bool_dtype
+ return current_guess
elif space.is_w(w_type, space.w_int):
if (current_guess is None or current_guess is bool_dtype or
- current_guess is int64_dtype):
- return int64_dtype
+ current_guess is long_dtype):
+ return long_dtype
+ return current_guess
+ elif space.is_w(w_type, space.w_long):
+ if (current_guess is None or current_guess is bool_dtype or
+ current_guess is long_dtype or current_guess is int64_dtype):
+ return int64_dtype
+ return current_guess
return space.fromcache(interp_dtype.W_Float64Dtype)
@@ -225,7 +267,9 @@
def impl(res_dtype, lvalue, rvalue):
res = getattr(res_dtype, op_name)(lvalue, rvalue)
if comparison_func:
- res = space.fromcache(interp_dtype.W_BoolDtype).box(res)
+ booldtype = space.fromcache(interp_dtype.W_BoolDtype)
+ assert isinstance(booldtype, interp_dtype.W_BoolDtype)
+ res = booldtype.box(res)
return res
return func_with_new_name(impl, ufunc_name)
@@ -269,6 +313,7 @@
("arccos", "arccos", 1, {"promote_to_float": True}),
("arctan", "arctan", 1, {"promote_to_float": True}),
("arcsinh", "arcsinh", 1, {"promote_to_float": True}),
+ ("arctanh", "arctanh", 1, {"promote_to_float": True}),
]:
self.add_ufunc(space, *ufunc_def)
@@ -278,7 +323,7 @@
identity = extra_kwargs.get("identity")
if identity is not None:
- identity = space.fromcache(interp_dtype.W_Int64Dtype).adapt_val(identity)
+ identity = space.fromcache(interp_dtype.W_LongDtype).adapt_val(identity)
extra_kwargs["identity"] = identity
func = ufunc_dtype_caller(space, ufunc_name, op_name, argcount,
diff --git a/pypy/module/_numpy/test/test_base.py b/pypy/module/_numpy/test/test_base.py
--- a/pypy/module/_numpy/test/test_base.py
+++ b/pypy/module/_numpy/test/test_base.py
@@ -64,18 +64,46 @@
def test_unaryops(self, space):
bool_dtype = space.fromcache(interp_dtype.W_BoolDtype)
int8_dtype = space.fromcache(interp_dtype.W_Int8Dtype)
+ uint8_dtype = space.fromcache(interp_dtype.W_UInt8Dtype)
+ int16_dtype = space.fromcache(interp_dtype.W_Int16Dtype)
+ uint16_dtype = space.fromcache(interp_dtype.W_UInt16Dtype)
int32_dtype = space.fromcache(interp_dtype.W_Int32Dtype)
+ uint32_dtype = space.fromcache(interp_dtype.W_UInt32Dtype)
+ long_dtype = space.fromcache(interp_dtype.W_LongDtype)
+ ulong_dtype = space.fromcache(interp_dtype.W_ULongDtype)
+ int64_dtype = space.fromcache(interp_dtype.W_Int64Dtype)
+ uint64_dtype = space.fromcache(interp_dtype.W_UInt64Dtype)
+ float32_dtype = space.fromcache(interp_dtype.W_Float32Dtype)
float64_dtype = space.fromcache(interp_dtype.W_Float64Dtype)
- # Normal rules, everythign returns itself
+ # Normal rules, everything returns itself
assert find_unaryop_result_dtype(space, bool_dtype) is bool_dtype
assert find_unaryop_result_dtype(space, int8_dtype) is int8_dtype
+ assert find_unaryop_result_dtype(space, uint8_dtype) is uint8_dtype
+ assert find_unaryop_result_dtype(space, int16_dtype) is int16_dtype
+ assert find_unaryop_result_dtype(space, uint16_dtype) is uint16_dtype
assert find_unaryop_result_dtype(space, int32_dtype) is int32_dtype
+ assert find_unaryop_result_dtype(space, uint32_dtype) is uint32_dtype
+ assert find_unaryop_result_dtype(space, long_dtype) is long_dtype
+ assert find_unaryop_result_dtype(space, ulong_dtype) is ulong_dtype
+ assert find_unaryop_result_dtype(space, int64_dtype) is int64_dtype
+ assert find_unaryop_result_dtype(space, uint64_dtype) is uint64_dtype
+ assert find_unaryop_result_dtype(space, float32_dtype) is float32_dtype
assert find_unaryop_result_dtype(space, float64_dtype) is float64_dtype
# Coerce to floats, some of these will eventually be float16, or
# whatever our smallest float type is.
- assert find_unaryop_result_dtype(space, bool_dtype, promote_to_float=True) is float64_dtype
- assert find_unaryop_result_dtype(space, int8_dtype, promote_to_float=True) is float64_dtype
+ assert find_unaryop_result_dtype(space, bool_dtype, promote_to_float=True) is float32_dtype # will be float16 if we ever put that in
+ assert find_unaryop_result_dtype(space, int8_dtype, promote_to_float=True) is float32_dtype # will be float16 if we ever put that in
+ assert find_unaryop_result_dtype(space, uint8_dtype, promote_to_float=True) is float32_dtype # will be float16 if we ever put that in
+ assert find_unaryop_result_dtype(space, int16_dtype, promote_to_float=True) is float32_dtype
+ assert find_unaryop_result_dtype(space, uint16_dtype, promote_to_float=True) is float32_dtype
assert find_unaryop_result_dtype(space, int32_dtype, promote_to_float=True) is float64_dtype
+ assert find_unaryop_result_dtype(space, uint32_dtype, promote_to_float=True) is float64_dtype
+ assert find_unaryop_result_dtype(space, int64_dtype, promote_to_float=True) is float64_dtype
+ assert find_unaryop_result_dtype(space, uint64_dtype, promote_to_float=True) is float64_dtype
+ assert find_unaryop_result_dtype(space, float32_dtype, promote_to_float=True) is float32_dtype
assert find_unaryop_result_dtype(space, float64_dtype, promote_to_float=True) is float64_dtype
+
+ # promote bools, happens with sign ufunc
+ assert find_unaryop_result_dtype(space, bool_dtype, promote_bools=True) is int8_dtype
diff --git a/pypy/module/_numpy/test/test_dtypes.py b/pypy/module/_numpy/test/test_dtypes.py
--- a/pypy/module/_numpy/test/test_dtypes.py
+++ b/pypy/module/_numpy/test/test_dtypes.py
@@ -17,6 +17,7 @@
from _numpy import dtype
assert dtype(bool).num == 0
+ assert dtype(int).num == 7
assert dtype(long).num == 9
assert dtype(float).num == 12
@@ -81,6 +82,48 @@
assert isinstance(a[i], (int, long))
assert a[1] == 1
+ def test_overflow(self):
+ from _numpy import array, dtype
+ assert array([128], 'b')[0] == -128
+ assert array([256], 'B')[0] == 0
+ assert array([32768], 'h')[0] == -32768
+ assert array([65536], 'H')[0] == 0
+ if dtype('l').itemsize == 4: # 32-bit
+ raises(OverflowError, "array([2**32/2], 'i')")
+ raises(OverflowError, "array([2**32], 'I')")
+ raises(OverflowError, "array([2**64/2], 'q')")
+ raises(OverflowError, "array([2**64], 'Q')")
+
+ def test_bool_binop_types(self):
+ from _numpy import array, dtype
+ types = ('?','b','B','h','H','i','I','l','L','q','Q','f','d')
+ N = len(types)
+ a = array([True], '?')
+ for t in types:
+ assert (a + array([0], t)).dtype is dtype(t)
+
+ def test_binop_types(self):
+ from _numpy import array, dtype
+ tests = [('b','B','h'), ('b','h','h'), ('b','H','i'), ('b','i','i'),
+ ('b','l','l'), ('b','q','q'), ('b','Q','d'), ('B','h','h'),
+ ('B','H','H'), ('B','i','i'), ('B','I','I'), ('B','l','l'),
+ ('B','L','L'), ('B','q','q'), ('B','Q','Q'), ('h','H','i'),
+ ('h','i','i'), ('h','l','l'), ('h','q','q'), ('h','Q','d'),
+ ('H','i','i'), ('H','I','I'), ('H','l','l'), ('H','L','L'),
+ ('H','q','q'), ('H','Q','Q'), ('i','l','l'), ('i','q','q'),
+ ('i','Q','d'), ('I','L','L'), ('I','q','q'), ('I','Q','Q'),
+ ('q','Q','d'), ('b','f','f'), ('B','f','f'), ('h','f','f'),
+ ('H','f','f'), ('i','f','d'), ('I','f','d'), ('l','f','d'),
+ ('L','f','d'), ('q','f','d'), ('Q','f','d'), ('q','d','d')]
+ if dtype('i').itemsize == dtype('l').itemsize: # 32-bit
+ tests.extend([('b','I','q'), ('b','L','q'), ('h','I','q'),
+ ('h','L','q'), ('i','I','q'), ('i','L','q')])
+ else:
+ tests.extend([('b','I','l'), ('b','L','d'), ('h','I','l'),
+ ('h','L','d'), ('i','I','l'), ('i','L','d')])
+ for d1, d2, dout in tests:
+ assert (array([1], d1) + array([1], d2)).dtype is dtype(dout)
+
def test_add_int8(self):
from _numpy import array, dtype
@@ -99,6 +142,15 @@
for i in range(5):
assert b[i] == i * 2
+ def test_add_uint32(self):
+ from _numpy import array, dtype
+
+ a = array(range(5), dtype="I")
+ b = a + a
+ assert b.dtype is dtype("I")
+ for i in range(5):
+ assert b[i] == i * 2
+
def test_shape(self):
from _numpy import dtype
diff --git a/pypy/module/_numpy/test/test_numarray.py b/pypy/module/_numpy/test/test_numarray.py
--- a/pypy/module/_numpy/test/test_numarray.py
+++ b/pypy/module/_numpy/test/test_numarray.py
@@ -18,7 +18,7 @@
assert a[13] == 5.3
def test_zeros_tuple_shape(self):
- from numpy import zeros, ones, empty
+ from _numpy import zeros, ones, empty
a = zeros((15,))
assert len(a) == 15
b = ones((3,))
@@ -560,8 +560,10 @@
from _numpy import array, dtype
assert array([True]).dtype is dtype(bool)
- assert array([True, 1]).dtype is dtype(long)
- assert array([1, 2, 3]).dtype is dtype(long)
+ assert array([True, False]).dtype is dtype(bool)
+ assert array([True, 1]).dtype is dtype(int)
+ assert array([1, 2, 3]).dtype is dtype(int)
+ assert array([1L, 2, 3]).dtype is dtype(long)
assert array([1.2, True]).dtype is dtype(float)
assert array([1.2, 5]).dtype is dtype(float)
assert array([]).dtype is dtype(float)
diff --git a/pypy/module/_numpy/test/test_ufuncs.py b/pypy/module/_numpy/test/test_ufuncs.py
--- a/pypy/module/_numpy/test/test_ufuncs.py
+++ b/pypy/module/_numpy/test/test_ufuncs.py
@@ -234,7 +234,7 @@
assert b[i] == math.sin(a[i])
a = sin(array([True, False], dtype=bool))
- assert a[0] == sin(1)
+ assert abs(a[0] - sin(1)) < 1e-7 # a[0] will be less precise
assert a[1] == 0.0
def test_cos(self):
@@ -300,12 +300,24 @@
def test_arcsinh(self):
import math
- from numpy import arcsinh, inf
+ from _numpy import arcsinh
+ from numpy import inf
for v in [inf, -inf, 1.0, math.e]:
assert math.asinh(v) == arcsinh(v)
assert math.isnan(arcsinh(float("nan")))
+ def test_arctanh(self):
+ import math
+ from numpy import arctanh
+
+ for v in [.99, .5, 0, -.5, -.99]:
+ assert math.atanh(v) == arctanh(v)
+ for v in [2.0, -2.0]:
+ assert math.isnan(arctanh(v))
+ for v in [1.0, -1.0]:
+ assert arctanh(v) == math.copysign(float("inf"), v)
+
def test_reduce_errors(self):
from _numpy import sin, add
diff --git a/pypy/module/_numpy/test/test_zjit.py b/pypy/module/_numpy/test/test_zjit.py
--- a/pypy/module/_numpy/test/test_zjit.py
+++ b/pypy/module/_numpy/test/test_zjit.py
@@ -1,20 +1,24 @@
from pypy.jit.metainterp.test.support import LLJitMixin
from pypy.module._numpy import interp_ufuncs, signature
from pypy.module._numpy.compile import (numpy_compile, FakeSpace,
- FloatObject)
-from pypy.module._numpy.interp_dtype import W_Float64Dtype, W_Int64Dtype
+ FloatObject, IntObject)
+from pypy.module._numpy.interp_dtype import W_Int32Dtype, W_Float64Dtype, W_Int64Dtype, W_UInt64Dtype
from pypy.module._numpy.interp_numarray import (BaseArray, SingleDimArray,
SingleDimSlice, scalar_w)
from pypy.rlib.nonconst import NonConstant
from pypy.rpython.annlowlevel import llstr
from pypy.rpython.test.test_llinterp import interpret
+import py
+
class TestNumpyJIt(LLJitMixin):
def setup_class(cls):
cls.space = FakeSpace()
cls.float64_dtype = cls.space.fromcache(W_Float64Dtype)
cls.int64_dtype = cls.space.fromcache(W_Int64Dtype)
+ cls.uint64_dtype = cls.space.fromcache(W_UInt64Dtype)
+ cls.int32_dtype = cls.space.fromcache(W_Int32Dtype)
def test_add(self):
def f(i):
@@ -303,6 +307,31 @@
'int_lt': 1, 'guard_true': 1, 'jump': 1})
assert result == 11.0
+ def test_int32_sum(self):
+ py.test.skip("pypy/jit/backend/llimpl.py needs to be changed to "
+ "deal correctly with int dtypes for this test to "
+ "work. skip for now until someone feels up to the task")
+ space = self.space
+ float64_dtype = self.float64_dtype
+ int32_dtype = self.int32_dtype
+
+ def f(n):
+ if NonConstant(False):
+ dtype = float64_dtype
+ else:
+ dtype = int32_dtype
+ ar = SingleDimArray(n, dtype=dtype)
+ i = 0
+ while i < n:
+ ar.get_concrete().setitem(i, int32_dtype.box(7))
+ i += 1
+ v = ar.descr_add(space, ar).descr_sum(space)
+ assert isinstance(v, IntObject)
+ return v.intval
+
+ result = self.meta_interp(f, [5], listops=True, backendopt=True)
+ assert result == f(5)
+
class TestTranslation(object):
def test_compile(self):
x = numpy_compile('aa+f*f/a-', 10)
diff --git a/pypy/module/pypyjit/test_pypy_c/test_00_model.py b/pypy/module/pypyjit/test_pypy_c/test_00_model.py
--- a/pypy/module/pypyjit/test_pypy_c/test_00_model.py
+++ b/pypy/module/pypyjit/test_pypy_c/test_00_model.py
@@ -50,8 +50,7 @@
cmdline.append(str(self.filepath))
#
print cmdline, logfile
- env={'PYPYLOG': 'jit-log-opt,jit-summary:' + str(logfile)}
- #env={'PYPYLOG': ':' + str(logfile)}
+ env={'PYPYLOG': 'jit-log-opt,jit-log-noopt,jit-summary:' + str(logfile)}
pipe = subprocess.Popen(cmdline,
env=env,
stdout=subprocess.PIPE,
diff --git a/pypy/module/pypyjit/test_pypy_c/test_call.py b/pypy/module/pypyjit/test_pypy_c/test_call.py
--- a/pypy/module/pypyjit/test_pypy_c/test_call.py
+++ b/pypy/module/pypyjit/test_pypy_c/test_call.py
@@ -366,12 +366,12 @@
# make sure that the "block" is not allocated
...
i20 = force_token()
- setfield_gc(p0, i20, descr=<SignedFieldDescr .*PyFrame.vable_token .*>)
p22 = new_with_vtable(19511408)
p24 = new_array(1, descr=<GcPtrArrayDescr>)
p26 = new_with_vtable(ConstClass(W_ListObject))
p27 = new(descr=<SizeDescr .*>)
p29 = new_array(0, descr=<GcPtrArrayDescr>)
+ setfield_gc(p0, i20, descr=<SignedFieldDescr .*PyFrame.vable_token .*>)
setfield_gc(p27, p29, descr=<GcPtrFieldDescr list.items .*>)
setfield_gc(p26, p27, descr=<.* .*W_ListObject.inst_wrappeditems .*>)
setarrayitem_gc(p24, 0, p26, descr=<GcPtrArrayDescr>)
diff --git a/pypy/module/pypyjit/test_pypy_c/test_generators.py b/pypy/module/pypyjit/test_pypy_c/test_generators.py
--- a/pypy/module/pypyjit/test_pypy_c/test_generators.py
+++ b/pypy/module/pypyjit/test_pypy_c/test_generators.py
@@ -19,8 +19,8 @@
assert loop.match_by_id("generator", """
i16 = force_token()
p45 = new_with_vtable(ConstClass(W_IntObject))
- setfield_gc(p45, i29, descr=<SignedFieldDescr .*>)
i47 = arraylen_gc(p8, descr=<GcPtrArrayDescr>) # Should be removed by backend
setarrayitem_gc(p8, 0, p45, descr=<GcPtrArrayDescr>)
+ setfield_gc(p45, i29, descr=<SignedFieldDescr .*>)
jump(..., descr=...)
""")
diff --git a/pypy/module/pypyjit/test_pypy_c/test_instance.py b/pypy/module/pypyjit/test_pypy_c/test_instance.py
--- a/pypy/module/pypyjit/test_pypy_c/test_instance.py
+++ b/pypy/module/pypyjit/test_pypy_c/test_instance.py
@@ -125,8 +125,8 @@
i12 = force_token()
--TICK--
p20 = new_with_vtable(ConstClass(W_IntObject))
+ setfield_gc(ConstPtr(ptr21), p20, descr=<GcPtrFieldDescr .*TypeCell.inst_w_value .*>)
setfield_gc(p20, i11, descr=<SignedFieldDescr.*W_IntObject.inst_intval .*>)
- setfield_gc(ConstPtr(ptr21), p20, descr=<GcPtrFieldDescr .*TypeCell.inst_w_value .*>)
jump(p0, p1, p2, p3, p4, p20, p6, i7, p20, descr=<Loop.>)
""")
diff --git a/pypy/module/thread/gil.py b/pypy/module/thread/gil.py
--- a/pypy/module/thread/gil.py
+++ b/pypy/module/thread/gil.py
@@ -16,7 +16,7 @@
class GILThreadLocals(OSThreadLocals):
"""A version of OSThreadLocals that enforces a GIL."""
- ll_GIL = thread.null_ll_lock
+ gil_ready = False
def initialize(self, space):
# add the GIL-releasing callback as an action on the space
@@ -25,12 +25,10 @@
def setup_threads(self, space):
"""Enable threads in the object space, if they haven't already been."""
- if not self.ll_GIL:
- try:
- self.ll_GIL = thread.allocate_ll_lock()
- except thread.error:
+ if not self.gil_ready:
+ if not thread.gil_allocate():
raise wrap_thread_error(space, "can't allocate GIL")
- thread.acquire_NOAUTO(self.ll_GIL, True)
+ self.gil_ready = True
self.enter_thread(space) # setup the main thread
result = True
else:
@@ -44,19 +42,16 @@
# test_lock_again after the global state was cleared by
# test_compile_lock. As a workaround, we repatch these global
# fields systematically.
- spacestate.ll_GIL = self.ll_GIL
invoke_around_extcall(before_external_call, after_external_call)
return result
def reinit_threads(self, space):
- if self.ll_GIL:
- self.ll_GIL = thread.allocate_ll_lock()
- thread.acquire_NOAUTO(self.ll_GIL, True)
- self.enter_thread(space)
+ if self.gil_ready:
+ self.gil_ready = False
+ self.setup_threads(space)
def yield_thread(self):
- thread.yield_thread() # explicitly release the gil (used by test_gil)
-
+ do_yield_thread()
class GILReleaseAction(PeriodicAsyncAction):
"""An action called every sys.checkinterval bytecodes. It releases
@@ -64,16 +59,12 @@
"""
def perform(self, executioncontext, frame):
- # Other threads can run between the release() and the acquire()
- # implicit in the following external function call (which has
- # otherwise no effect).
- thread.yield_thread()
+ do_yield_thread()
class SpaceState:
def _freeze_(self):
- self.ll_GIL = thread.null_ll_lock
self.action_after_thread_switch = None
# ^^^ set by AsyncAction.fire_after_thread_switch()
return False
@@ -95,14 +86,14 @@
# this function must not raise, in such a way that the exception
# transformer knows that it cannot raise!
e = get_errno()
- thread.release_NOAUTO(spacestate.ll_GIL)
+ thread.gil_release()
set_errno(e)
before_external_call._gctransformer_hint_cannot_collect_ = True
before_external_call._dont_reach_me_in_del_ = True
def after_external_call():
e = get_errno()
- thread.acquire_NOAUTO(spacestate.ll_GIL, True)
+ thread.gil_acquire()
thread.gc_thread_run()
spacestate.after_thread_switch()
set_errno(e)
@@ -115,3 +106,18 @@
# pointers in the shadow stack. This is necessary because the GIL is
# not held after the call to before_external_call() or before the call
# to after_external_call().
+
+def do_yield_thread():
+ # explicitly release the gil, in a way that tries to give more
+ # priority to other threads (as opposed to continuing to run in
+ # the same thread).
+ if thread.gil_yield_thread():
+ thread.gc_thread_run()
+ spacestate.after_thread_switch()
+do_yield_thread._gctransformer_hint_close_stack_ = True
+do_yield_thread._dont_reach_me_in_del_ = True
+do_yield_thread._dont_inline_ = True
+
+# do_yield_thread() needs a different hint: _gctransformer_hint_close_stack_.
+# The *_external_call() functions are themselves called only from the rffi
+# module from a helper function that also has this hint.
diff --git a/pypy/module/thread/ll_thread.py b/pypy/module/thread/ll_thread.py
--- a/pypy/module/thread/ll_thread.py
+++ b/pypy/module/thread/ll_thread.py
@@ -17,7 +17,8 @@
include_dirs = [str(py.path.local(autopath.pypydir).join('translator', 'c'))],
export_symbols = ['RPyThreadGetIdent', 'RPyThreadLockInit',
'RPyThreadAcquireLock', 'RPyThreadReleaseLock',
- 'RPyThreadYield',
+ 'RPyGilAllocate', 'RPyGilYieldThread',
+ 'RPyGilRelease', 'RPyGilAcquire',
'RPyThreadGetStackSize', 'RPyThreadSetStackSize',
'RPyOpaqueDealloc_ThreadLock',
'RPyThreadAfterFork']
@@ -69,8 +70,16 @@
[TLOCKP], lltype.Void,
_nowrapper=True)
-# this function does nothing apart from releasing the GIL temporarily.
-yield_thread = llexternal('RPyThreadYield', [], lltype.Void, threadsafe=True)
+# these functions manipulate directly the GIL, whose definition does not
+# escape the C code itself
+gil_allocate = llexternal('RPyGilAllocate', [], lltype.Signed,
+ _nowrapper=True)
+gil_yield_thread = llexternal('RPyGilYieldThread', [], lltype.Signed,
+ _nowrapper=True)
+gil_release = llexternal('RPyGilRelease', [], lltype.Void,
+ _nowrapper=True)
+gil_acquire = llexternal('RPyGilAcquire', [], lltype.Void,
+ _nowrapper=True)
def allocate_lock():
return Lock(allocate_ll_lock())
diff --git a/pypy/module/thread/test/test_gil.py b/pypy/module/thread/test/test_gil.py
--- a/pypy/module/thread/test/test_gil.py
+++ b/pypy/module/thread/test/test_gil.py
@@ -30,19 +30,34 @@
use_threads = True
bigtest = False
- def test_one_thread(self):
+ def test_one_thread(self, skew=+1):
+ from pypy.rlib.debug import debug_print
if self.bigtest:
- N = 1000000
+ N = 100000
+ skew *= 25000
else:
N = 100
+ skew *= 25
space = FakeSpace()
class State:
pass
state = State()
- def runme():
- for i in range(N):
+ def runme(main=False):
+ j = 0
+ for i in range(N + [-skew, skew][main]):
+ state.datalen1 += 1 # try to crash if the GIL is not
+ state.datalen2 += 1 # correctly acquired
state.data.append((thread.get_ident(), i))
+ state.datalen3 += 1
+ state.datalen4 += 1
+ assert state.datalen1 == len(state.data)
+ assert state.datalen2 == len(state.data)
+ assert state.datalen3 == len(state.data)
+ assert state.datalen4 == len(state.data)
+ debug_print(main, i, state.datalen4)
state.threadlocals.yield_thread()
+ assert i == j
+ j += 1
def bootstrap():
try:
runme()
@@ -50,20 +65,26 @@
thread.gc_thread_die()
def f():
state.data = []
+ state.datalen1 = 0
+ state.datalen2 = 0
+ state.datalen3 = 0
+ state.datalen4 = 0
state.threadlocals = gil.GILThreadLocals()
state.threadlocals.setup_threads(space)
thread.gc_thread_prepare()
subident = thread.start_new_thread(bootstrap, ())
mainident = thread.get_ident()
- runme()
+ runme(True)
still_waiting = 3000
while len(state.data) < 2*N:
+ debug_print(len(state.data))
if not still_waiting:
raise ValueError("time out")
still_waiting -= 1
if not we_are_translated(): gil.before_external_call()
time.sleep(0.01)
if not we_are_translated(): gil.after_external_call()
+ debug_print("leaving!")
i1 = i2 = 0
for tid, i in state.data:
if tid == mainident:
@@ -72,14 +93,17 @@
assert i == i2; i2 += 1
else:
assert 0
- assert i1 == N
- assert i2 == N
+ assert i1 == N + skew
+ assert i2 == N - skew
return len(state.data)
fn = self.getcompiled(f, [])
res = fn()
assert res == 2*N
+ def test_one_thread_rev(self):
+ self.test_one_thread(skew=-1)
+
class TestRunDirectly(GILTests):
def getcompiled(self, f, argtypes):
diff --git a/pypy/module/thread/test/test_thread.py b/pypy/module/thread/test/test_thread.py
--- a/pypy/module/thread/test/test_thread.py
+++ b/pypy/module/thread/test/test_thread.py
@@ -225,7 +225,8 @@
def busy_wait():
for x in range(1000):
- time.sleep(0.01)
+ print 'tick...', x # <-force the GIL to be released, as
+ time.sleep(0.01) # time.sleep doesn't do non-translated
# This is normally called by app_main.py
signal.signal(signal.SIGINT, signal.default_int_handler)
diff --git a/pypy/rpython/lltypesystem/opimpl.py b/pypy/rpython/lltypesystem/opimpl.py
--- a/pypy/rpython/lltypesystem/opimpl.py
+++ b/pypy/rpython/lltypesystem/opimpl.py
@@ -357,7 +357,7 @@
def op_cast_float_to_uint(f):
assert type(f) is float
- return r_uint(int(f))
+ return r_uint(long(f))
def op_cast_float_to_longlong(f):
assert type(f) is float
@@ -369,7 +369,7 @@
def op_cast_float_to_ulonglong(f):
assert type(f) is float
- return r_ulonglong(r_longlong(f))
+ return r_ulonglong(long(f))
def op_cast_char_to_int(b):
assert type(b) is str and len(b) == 1
diff --git a/pypy/rpython/lltypesystem/test/test_lloperation.py b/pypy/rpython/lltypesystem/test/test_lloperation.py
--- a/pypy/rpython/lltypesystem/test/test_lloperation.py
+++ b/pypy/rpython/lltypesystem/test/test_lloperation.py
@@ -5,6 +5,7 @@
from pypy.rpython.llinterp import LLFrame
from pypy.rpython.test.test_llinterp import interpret
from pypy.rpython import rclass
+from pypy.rlib.rarithmetic import LONGLONG_MASK, r_longlong, r_ulonglong
LL_INTERP_OPERATIONS = [name[3:] for name in LLFrame.__dict__.keys()
if name.startswith('op_')]
@@ -133,6 +134,14 @@
py.test.raises(TypeError, llop.getinteriorfield,
lltype.Signed, s3, 'y')
+def test_cast_float_to_ulonglong():
+ f = 12350000000000000000.0
+ py.test.raises(OverflowError, r_longlong, f)
+ r_longlong(f / 2) # does not raise OverflowError
+ #
+ x = llop.cast_float_to_ulonglong(lltype.UnsignedLongLong, f)
+ assert x == r_ulonglong(f)
+
# ___________________________________________________________________________
# This tests that the LLInterpreter and the LL_OPERATIONS tables are in sync.
diff --git a/pypy/translator/c/gcc/trackgcroot.py b/pypy/translator/c/gcc/trackgcroot.py
--- a/pypy/translator/c/gcc/trackgcroot.py
+++ b/pypy/translator/c/gcc/trackgcroot.py
@@ -486,6 +486,8 @@
'paddq', 'pinsr',
# zero-extending moves should not produce GC pointers
'movz',
+ # locked operations should not move GC pointers, at least so far
+ 'lock',
])
# a partial list is hopefully good enough for now; it's all to support
diff --git a/pypy/translator/c/src/thread.h b/pypy/translator/c/src/thread.h
--- a/pypy/translator/c/src/thread.h
+++ b/pypy/translator/c/src/thread.h
@@ -37,14 +37,9 @@
#endif
-/* common helper: this does nothing, but is called with the GIL released.
- This gives other threads a chance to grab the GIL and run. */
-void RPyThreadYield(void);
-
-#ifndef PYPY_NOT_MAIN_FILE
-void RPyThreadYield(void)
-{
-}
-#endif
+long RPyGilAllocate(void);
+long RPyGilYieldThread(void);
+void RPyGilRelease(void);
+void RPyGilAcquire(void);
#endif
diff --git a/pypy/translator/c/src/thread_nt.h b/pypy/translator/c/src/thread_nt.h
--- a/pypy/translator/c/src/thread_nt.h
+++ b/pypy/translator/c/src/thread_nt.h
@@ -221,4 +221,57 @@
#define RPyThreadTLS_Set(key, value) TlsSetValue(key, value)
+/************************************************************/
+/* GIL code */
+/************************************************************/
+
+static volatile LONG pending_acquires = -1;
+static CRITICAL_SECTION mutex_gil;
+static HANDLE cond_gil;
+
+long RPyGilAllocate(void)
+{
+ pending_acquires = 0;
+ InitializeCriticalSection(&mutex_gil);
+ EnterCriticalSection(&mutex_gil);
+ cond_gil = CreateEvent (NULL, FALSE, FALSE, NULL);
+ return 1;
+}
+
+long RPyGilYieldThread(void)
+{
+ /* can be called even before RPyGilAllocate(), but in this case,
+ pending_acquires will be -1 */
+ if (pending_acquires <= 0)
+ return 0;
+ InterlockedIncrement(&pending_acquires);
+ PulseEvent(&cond_gil);
+
+ /* hack: the three following lines do a pthread_cond_wait(), and
+ normally specifying a timeout of INFINITE would be fine. But the
+ first and second operations are not done atomically, so there is a
+ (small) risk that PulseEvent misses the WaitForSingleObject().
+ In this case the process will just sleep a few milliseconds. */
+ LeaveCriticalSection(&mutex_gil);
+ WaitForSingleObject(&cond_gil, 15);
+ EnterCriticalSection(&mutex_gil);
+
+ InterlockedDecrement(&pending_acquires);
+ return 1;
+}
+
+void RPyGilRelease(void)
+{
+ LeaveCriticalSection(&mutex_gil);
+ PulseEvent(&cond_gil);
+}
+
+void RPyGilAcquire(void)
+{
+ InterlockedIncrement(&pending_acquires);
+ EnterCriticalSection(&mutex_gil);
+ InterlockedDecrement(&pending_acquires);
+}
+
+
#endif /* PYPY_NOT_MAIN_FILE */
diff --git a/pypy/translator/c/src/thread_pthread.h b/pypy/translator/c/src/thread_pthread.h
--- a/pypy/translator/c/src/thread_pthread.h
+++ b/pypy/translator/c/src/thread_pthread.h
@@ -12,6 +12,7 @@
#include <signal.h>
#include <stdio.h>
#include <errno.h>
+#include <assert.h>
/* The following is hopefully equivalent to what CPython does
(which is trying to compile a snippet of code using it) */
@@ -459,4 +460,113 @@
#define RPyThreadTLS_Set(key, value) pthread_setspecific(key, value)
+/************************************************************/
+/* GIL code */
+/************************************************************/
+
+#ifdef __llvm__
+# define HAS_ATOMIC_ADD
+#endif
+
+#ifdef __GNUC__
+# if __GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 1)
+# define HAS_ATOMIC_ADD
+# endif
+#endif
+
+#ifdef HAS_ATOMIC_ADD
+# define atomic_add __sync_fetch_and_add
+#else
+# if defined(__amd64__)
+# define atomic_add(ptr, value) asm volatile ("lock addq %0, %1" \
+ : : "ri"(value), "m"(*(ptr)) : "memory")
+# elif defined(__i386__)
+# define atomic_add(ptr, value) asm volatile ("lock addl %0, %1" \
+ : : "ri"(value), "m"(*(ptr)) : "memory")
+# else
+# error "Please use gcc >= 4.1 or write a custom 'asm' for your CPU."
+# endif
+#endif
+
+#define ASSERT_STATUS(call) \
+ if (call != 0) { \
+ fprintf(stderr, "Fatal error: " #call "\n"); \
+ abort(); \
+ }
+
+static void _debug_print(const char *msg)
+{
+#if 0
+ int col = (int)pthread_self();
+ col = 31 + ((col / 8) % 8);
+ fprintf(stderr, "\033[%dm%s\033[0m", col, msg);
+#endif
+}
+
+static volatile long pending_acquires = -1;
+static pthread_mutex_t mutex_gil = PTHREAD_MUTEX_INITIALIZER;
+static pthread_cond_t cond_gil = PTHREAD_COND_INITIALIZER;
+
+static void assert_has_the_gil(void)
+{
+#ifdef RPY_ASSERT
+ assert(pthread_mutex_trylock(&mutex_gil) != 0);
+ assert(pending_acquires >= 0);
+#endif
+}
+
+long RPyGilAllocate(void)
+{
+ _debug_print("RPyGilAllocate\n");
+ pending_acquires = 0;
+ pthread_mutex_trylock(&mutex_gil);
+ assert_has_the_gil();
+ return 1;
+}
+
+long RPyGilYieldThread(void)
+{
+ /* can be called even before RPyGilAllocate(), but in this case,
+ pending_acquires will be -1 */
+#ifdef RPY_ASSERT
+ if (pending_acquires >= 0)
+ assert_has_the_gil();
+#endif
+ if (pending_acquires <= 0)
+ return 0;
+ atomic_add(&pending_acquires, 1L);
+ _debug_print("{");
+ ASSERT_STATUS(pthread_cond_signal(&cond_gil));
+ ASSERT_STATUS(pthread_cond_wait(&cond_gil, &mutex_gil));
+ _debug_print("}");
+ atomic_add(&pending_acquires, -1L);
+ assert_has_the_gil();
+ return 1;
+}
+
+void RPyGilRelease(void)
+{
+ _debug_print("RPyGilRelease\n");
+#ifdef RPY_ASSERT
+ assert(pending_acquires >= 0);
+#endif
+ assert_has_the_gil();
+ ASSERT_STATUS(pthread_mutex_unlock(&mutex_gil));
+ ASSERT_STATUS(pthread_cond_signal(&cond_gil));
+}
+
+void RPyGilAcquire(void)
+{
+ _debug_print("about to RPyGilAcquire...\n");
+#ifdef RPY_ASSERT
+ assert(pending_acquires >= 0);
+#endif
+ atomic_add(&pending_acquires, 1L);
+ ASSERT_STATUS(pthread_mutex_lock(&mutex_gil));
+ atomic_add(&pending_acquires, -1L);
+ assert_has_the_gil();
+ _debug_print("RPyGilAcquire\n");
+}
+
+
#endif /* PYPY_NOT_MAIN_FILE */
More information about the pypy-commit
mailing list