Author: Richard Plangger <planri...@gmail.com>
Branch: ppc-vsx-support
Changeset: r87028:88be11973c7a
Date: 2016-09-12 12:04 +0200
http://bitbucket.org/pypy/pypy/changeset/88be11973c7a/
Log: fixed vecopt tests with the new delayed execution
diff --git a/rpython/jit/metainterp/optimizeopt/dependency.py
b/rpython/jit/metainterp/optimizeopt/dependency.py
--- a/rpython/jit/metainterp/optimizeopt/dependency.py
+++ b/rpython/jit/metainterp/optimizeopt/dependency.py
@@ -1057,7 +1057,7 @@
var = ResOperation(rop.INT_ADD, args)
tolist.append(var)
if self.constant < 0:
- args = [var, ConstInt(self.constant)]
+ args = [var, ConstInt(-self.constant)]
var = ResOperation(rop.INT_SUB, args)
tolist.append(var)
return tolist
diff --git a/rpython/jit/metainterp/optimizeopt/schedule.py
b/rpython/jit/metainterp/optimizeopt/schedule.py
--- a/rpython/jit/metainterp/optimizeopt/schedule.py
+++ b/rpython/jit/metainterp/optimizeopt/schedule.py
@@ -55,18 +55,29 @@
op = node.getoperation()
if op in needs_resolving:
# either it is a normal operation, or we know that there is a
linear combination
+ del needs_resolving[op]
if op in indexvars:
indexvar = indexvars[op]
for operation in indexvar.get_operations():
- self.oplist.append(operation)
+ self.append_to_oplist(operation)
last = operation
+ indexvars[last] = indexvar
self.renamer.start_renaming(op, last)
- del needs_resolving[op]
else:
- del needs_resolving[op]
self.resolve_delayed(needs_resolving, delayed, op)
- self.oplist.append(op)
+ self.append_to_oplist(op)
+ self.seen[op] = None
+ if len(delayed) > i:
+ del delayed[i]
i -= 1
+ # some times the recursive call can remove several items from
delayed,
+ # thus we correct the index here
+ if len(delayed) <= i:
+ i = len(delayed)-1
+
+ def append_to_oplist(self, op):
+ self.renamer.rename(op)
+ self.oplist.append(op)
def post_schedule(self):
@@ -106,7 +117,79 @@
def try_emit_or_delay(self, node):
# implement me in subclass. e.g. as in VecScheduleState
- raise NotImplementedError
+
+ if not node.is_imaginary() and node.is_pure():
+ # this operation might never be emitted. only if it is really
needed
+ self.delay_emit(node)
+ return
+ # emit a now!
+ self.pre_emit(node)
+ self.mark_emitted(node)
+ if not node.is_imaginary():
+ op = node.getoperation()
+ self.seen[op] = None
+ self.append_to_oplist(op)
+
+ def delay_emit(self, node):
+ """ it has been decided that the operation might be scheduled later """
+ delayed = node.delayed or []
+ delayed.append(node)
+ node.delayed = None
+ provides = node.provides()
+ if len(provides) == 0:
+ self.delayed.append(node)
+ else:
+ for to in node.provides():
+ tnode = to.target_node()
+ self.delegate_delay(tnode, delayed[:])
+ self.mark_emitted(node)
+
+ def delegate_delay(self, node, delayed):
+ """ Chain up delays, this can reduce many more of the operations """
+ if node.delayed is None:
+ node.delayed = delayed
+ else:
+ delayedlist = node.delayed
+ for d in delayed:
+ delayedlist.append(d)
+
+
+ def mark_emitted(state, node, unpack=True):
+ """ An operation has been emitted, adds new operations to the worklist
+ whenever their dependency count drops to zero.
+ Keeps worklist sorted (see priority) """
+ worklist = state.worklist
+ provides = node.provides()[:]
+ for dep in provides: # COPY
+ target = dep.to
+ node.remove_edge_to(target)
+ if not target.emitted and target.depends_count() == 0:
+ # sorts them by priority
+ i = len(worklist)-1
+ while i >= 0:
+ cur = worklist[i]
+ c = (cur.priority - target.priority)
+ if c < 0: # meaning itnode.priority < target.priority:
+ worklist.insert(i+1, target)
+ break
+ elif c == 0:
+ # if they have the same priority, sort them
+ # using the original position in the trace
+ if target.getindex() < cur.getindex():
+ worklist.insert(i+1, target)
+ break
+ i -= 1
+ else:
+ worklist.insert(0, target)
+ node.clear_dependencies()
+ node.emitted = True
+ if not node.is_imaginary():
+ op = node.getoperation()
+ state.renamer.rename(op)
+ if unpack:
+ state.ensure_args_unpacked(op)
+ state.post_emit(node)
+
def delay(self, node):
return False
@@ -120,8 +203,27 @@
def post_emit(self, node):
pass
- def pre_emit(self, node):
- pass
+ def pre_emit(self, node, pack_first=True):
+ delayed = node.delayed
+ if delayed:
+ # there are some nodes that have been delayed just for this
operation
+ if pack_first:
+ op = node.getoperation()
+ self.resolve_delayed({}, delayed, op)
+ for node in delayed:
+ if node in self.seen:
+ continue
+ if node is not None:
+ provides = node.provides()
+ if len(provides) == 0:
+ # add this node to the final delay list
+ # might be emitted before jump!
+ self.delayed.append(node)
+ else:
+ for to in node.provides():
+ tnode = to.target_node()
+ self.delegate_delay(tnode, [node])
+ node.delayed = None
class Scheduler(object):
""" Create an instance of this class to (re)schedule a vector trace. """
@@ -412,7 +514,7 @@
assert vecinfo.count > argvecinfo.count
def expand(state, pack, args, arg, index):
- """ Expand a value into a vector box. useful for arith metic
+ """ Expand a value into a vector box. useful for arithmetic
of one vector with a scalar (either constant/varialbe)
"""
left = pack.leftmost()
@@ -547,23 +649,8 @@
failargs[i] = self.renamer.rename_map.get(seed, seed)
op.setfailargs(failargs)
- delayed = node.delayed
- if delayed:
- # there are some nodes that have been delayed just for this
operation
- if pack_first:
- self.resolve_delayed({}, delayed, op)
- for node in delayed:
- if node is not None:
- provides = node.provides()
- if len(provides) == 0:
- # add this node to the final delay list
- # might be emitted before jumping!
- self.delayed.append(node)
- else:
- for to in node.provides():
- tnode = to.target_node()
- self.delegate_delay(tnode, [node])
- node.delayed = None
+ SchedulerState.pre_emit(self, node, pack_first)
+
def profitable(self):
return self.costmodel.profitable()
@@ -584,76 +671,7 @@
turn_into_vector(self, node.pack)
return
elif not node.emitted:
- if not node.is_imaginary() and node.is_pure():
- # this operation might never be emitted. only if it is really
needed
- self.delay_emit(node)
- return
- # emit a now!
- self.pre_emit(node)
- self.mark_emitted(node)
- if not node.is_imaginary():
- op = node.getoperation()
- self.seen[op] = None
- self.oplist.append(op)
-
- def delay_emit(self, node):
- """ it has been decided that the operation might be scheduled later """
- delayed = node.delayed or []
- delayed.append(node)
- node.delayed = None
- provides = node.provides()
- if len(provides) == 0:
- self.delayed.append(node)
- else:
- for to in node.provides():
- tnode = to.target_node()
- self.delegate_delay(tnode, delayed[:])
- self.mark_emitted(node)
-
- def delegate_delay(self, node, delayed):
- """ Chain up delays, this can reduce many more of the operations """
- if node.delayed is None:
- node.delayed = delayed
- else:
- delayedlist = node.delayed
- for d in delayed:
- delayedlist.append(d)
-
- def mark_emitted(state, node, unpack=True):
- """ An operation has been emitted, adds new operations to the worklist
- whenever their dependency count drops to zero.
- Keeps worklist sorted (see priority) """
- worklist = state.worklist
- provides = node.provides()[:]
- for dep in provides: # COPY
- target = dep.to
- node.remove_edge_to(target)
- if not target.emitted and target.depends_count() == 0:
- # sorts them by priority
- i = len(worklist)-1
- while i >= 0:
- cur = worklist[i]
- c = (cur.priority - target.priority)
- if c < 0: # meaning itnode.priority < target.priority:
- worklist.insert(i+1, target)
- break
- elif c == 0:
- # if they have the same priority, sort them
- # using the original position in the trace
- if target.getindex() < cur.getindex():
- worklist.insert(i+1, target)
- break
- i -= 1
- else:
- worklist.insert(0, target)
- node.clear_dependencies()
- node.emitted = True
- if not node.is_imaginary():
- op = node.getoperation()
- state.renamer.rename(op)
- if unpack:
- state.ensure_args_unpacked(op)
- state.post_emit(node)
+ SchedulerState.try_emit_or_delay(self, node)
def delay(self, node):
if node.pack:
@@ -706,7 +724,7 @@
self.renamer.start_renaming(arg, vecop)
self.seen[vecop] = None
self.costmodel.record_vector_unpack(var, pos, 1)
- self.oplist.append(vecop)
+ self.append_to_oplist(vecop)
return vecop
return arg
diff --git a/rpython/jit/metainterp/optimizeopt/test/test_vecopt.py
b/rpython/jit/metainterp/optimizeopt/test/test_vecopt.py
--- a/rpython/jit/metainterp/optimizeopt/test/test_vecopt.py
+++ b/rpython/jit/metainterp/optimizeopt/test/test_vecopt.py
@@ -271,7 +271,7 @@
def test_opcount_filling_guard(self):
descr = ArrayDescr(0,4, None, 'S')
- vec = ResOperation(rop.VEC_LOAD_I, ['a','i'], descr=descr)
+ vec = ResOperation(rop.VEC_LOAD_I, ['a','i', 8, 0], descr=descr)
vec.count = 4
pack = Pack([Node(ResOperation(rop.GUARD_TRUE, [vec]), 0),
Node(ResOperation(rop.GUARD_TRUE, [vec]), 1),
@@ -319,7 +319,7 @@
self.vectorize(trace)
self.debug_print_operations(trace)
self.ensure_operations([
- 'v10[4xi32] = vec_load_i(p0,i0,descr=int32arraydescr)',
+ 'v10[4xi32] = vec_load_i(p0,i0,4,0,descr=int32arraydescr)',
'v11[4xi32] = vec_int_is_true(v10[4xi32])',
'i100 = vec_unpack_i(v11[4xi32], 0, 1)',
'vec_guard_true(v11[4xi32]) [i100]',
@@ -384,9 +384,8 @@
"""
opt = """
[p0,i0]
- i1 = int_add(i0,1)
+ v3[2xi64] = vec_load_i(p0,i0,8,0,descr=arraydescr)
i2 = int_add(i0,2)
- v3[2xi64] = vec_load_i(p0,i0,descr=arraydescr)
jump(p0,i2)
"""
loop = self.parse_loop(ops)
@@ -799,13 +798,11 @@
for i in range(15):
for j in range(15):
try:
+ mref1 = graph.getmemref(i)
+ mref2 = graph.getmemref(j)
if i-4 == j or i+4 == j:
- mref1 = graph.getmemref(i)
- mref2 = graph.getmemref(j)
assert mref1.is_adjacent_to(mref2)
else:
- mref1 = graph.getmemref(i)
- mref2 = graph.getmemref(j)
assert not mref1.is_adjacent_to(mref2)
except KeyError:
pass
@@ -846,11 +843,13 @@
loop = self.parse_loop(ops)
vopt, graph = self.extend_packset(loop,1)
assert len(graph.memory_refs) == 2
- self.assert_independent(graph, 5,10)
- assert len(vopt.packset.packs) == 2
+ self.assert_independent(graph, 3,7)
+ # the delayed scheduling strips away the vectorized addition,
+ # because it is never used
+ assert len(vopt.packset.packs) == 1
self.assert_packset_empty(vopt.packset,
len(loop.operations),
- [(5,10), (4,9)])
+ [(4,8)])
def test_packset_extend_load_modify_store(self):
ops = """
@@ -1009,11 +1008,11 @@
i1 = int_add(i0, {stride})
i11 = int_le(i1, 128)
guard_true(i11) [p0,p1,p2,i1]
- i12 = int_add(i1, {stride})
- v1 = vec_load{suffix}(p0, i0, descr={descr}arraydescr)
- v2 = vec_load{suffix}(p1, i0, descr={descr}arraydescr)
+ v1 = vec_load{suffix}(p0, i0,8,0, descr={descr}arraydescr)
+ v2 = vec_load{suffix}(p1, i0,8,0, descr={descr}arraydescr)
v3 = {op}(v1,v2)
- vec_store(p2, i0, v3, descr={descr}arraydescr)
+ vec_store(p2, i0, v3,8,0, descr={descr}arraydescr)
+ i12 = int_add(i0, 2)
jump(p0,p1,p2,i12)
""".format(op='vec_'+op,descr=descr,stride=1,suffix=suffix)
loop = self.parse_loop(ops)
@@ -1038,15 +1037,14 @@
i11 = int_add(i0, 1)
i12 = int_lt(i11, i1)
guard_true(i12) [i0,i1,i2,i3,i4]
- i6 = int_mul(i0, 8)
- i13 = int_add(i11, 1)
+ i13 = int_add(i0, 2)
i18 = int_lt(i13, i1)
guard_true(i18) [i11,i1,i2,i3,i4]
- i14 = int_mul(i11, 8)
- v19[2xi64] = vec_load_i(i2, i6, descr=arraydescr)
- v20[2xi64] = vec_load_i(i3, i6, descr=arraydescr)
+ i6 = int_mul(i0, 8)
+ v19[2xi64] = vec_load_i(i2, i6, 1, 0, descr=arraydescr)
+ v20[2xi64] = vec_load_i(i3, i6, 1, 0, descr=arraydescr)
v21[2xi64] = vec_int_add(v19, v20)
- vec_store(i4, i6, v21, descr=arraydescr)
+ vec_store(i4, i6, v21, 1, 0, descr=arraydescr)
jump(i13, i1, i2, i3, i4)
"""
loop = self.parse_loop(ops)
@@ -1076,7 +1074,7 @@
{dead_code}
i500 = int_add(i0, 16)
i501 = int_lt(i500, 102)
- v10[16xi8] = vec_load_i(p0, i0, descr=chararraydescr)
+ v10[16xi8] = vec_load_i(p0, i0, 1, 0, descr=chararraydescr)
jump(p0,i2)
""".format(dead_code=dead_code)
loop = self.parse_loop(ops)
@@ -1105,6 +1103,7 @@
[p0,i0]
i1 = getarrayitem_raw_i(p0, i0, descr=arraydescr)
i4 = int_sub(i1, 42)
+ setarrayitem_raw(p0, i0, i4, descr=arraydescr)
i3 = int_add(i0,1)
i5 = int_lt(i3, 10)
guard_true(i5) [p0, i0]
@@ -1121,8 +1120,9 @@
guard_true(i3) [p0,i0]
i4 = int_add(i0, 2)
i5 = int_lt(i4, 10)
- v1[2xf64] = vec_load_i(p0, i0, descr=arraydescr)
+ v1[2xf64] = vec_load_i(p0, i0, 8, 0, descr=arraydescr)
v2[2xf64] = vec_int_sub(v1[2xf64], v3[2xf64])
+ vec_store(p0, i0, v2[2xf64], 8, 0, descr=arraydescr)
jump(p0,i2,v3[2xf64])
"""
loop = self.parse_loop(ops)
@@ -1134,6 +1134,7 @@
[p0,i0,f3]
f1 = getarrayitem_raw_f(p0, i0, descr=floatarraydescr)
f4 = float_add(f1, f3)
+ setarrayitem_raw(p0, i0, f4, descr=floatarraydescr)
i3 = int_add(i0,1)
i5 = int_lt(i3, 10)
guard_true(i5) [p0, i0]
@@ -1150,8 +1151,9 @@
guard_true(i3) [p0,i0,f3]
i4 = int_add(i0, 2)
i5 = int_lt(i4, 10)
- v1[2xf64] = vec_load_f(p0, i0, descr=floatarraydescr)
+ v1[2xf64] = vec_load_f(p0, i0, 8, 0, descr=floatarraydescr)
v2[2xf64] = vec_float_add(v1[2xf64], v3[2xf64])
+ vec_store(p0, i0, v2[2xf64], 8, 0, descr=floatarraydescr)
jump(p0,i2,f3,v3[2xf64])
"""
loop = self.parse_loop(ops)
@@ -1181,7 +1183,7 @@
guard_true(i2) [p0, i0, v2[2xf64]]
i10 = int_add(i0, 16)
i20 = int_lt(i10, 100)
- v1[2xf64] = vec_load_f(p0, i0, descr=floatarraydescr)
+ v1[2xf64] = vec_load_f(p0, i0, 1, 0, descr=floatarraydescr)
v3[2xf64] = vec_float_add(v2[2xf64], v1[2xf64])
jump(p0, i1, v3[2xf64])
"""
@@ -1213,21 +1215,18 @@
i54 = int_add(i28, 2)
i638 = int_ge(i54, i18)
guard_false(i638) [p36, i28, p9, i37, p14, f34, p12, p38, f35, p39,
i40, i41, p42, i43, i44, i21, i4, i0, i18]
- i12 = int_add(i44, 8)
- i56 = int_add(i41, 8)
- i46 = int_add(i37, 8)
i47 = int_add(i28, 2)
i52 = int_ge(i47, i18)
+ v61[2xf64] = vec_load_f(i21, i44, 1, 0, descr=floatarraydescr)
+ v62[2xf64] = vec_load_f(i4, i41, 1, 0, descr=floatarraydescr)
+ v63[2xf64] = vec_float_add(v61, v62)
+ vec_store(i0, i37, v63, 1, 0, descr=floatarraydescr)
i55 = int_add(i44, 16)
i629 = int_add(i41, 16)
i637 = int_add(i37, 16)
- v61[2xf64] = vec_load_f(i21, i44, descr=floatarraydescr)
- v62[2xf64] = vec_load_f(i4, i41, descr=floatarraydescr)
- v63[2xf64] = vec_float_add(v61, v62)
- vec_store(i0, i37, v63, descr=floatarraydescr)
f100 = vec_unpack_f(v61, 1, 1)
f101 = vec_unpack_f(v62, 1, 1)
- jump(p36, i637, p9, i56, p14, f100, p12, p38, f101, p39, i40, i54,
p42, i43, i55, i21, i4, i0, i18)
+ jump(p36, i637, p9, i629, p14, f100, p12, p38, f101, p39, i40, i54,
p42, i43, i55, i21, i4, i0, i18)
""")
vopt = self.vectorize(trace)
self.assert_equal(trace, trace_opt)
@@ -1256,12 +1255,12 @@
i11 = int_ge(i6, 36)
i7 = int_add(i1, 4)
i14 = int_ge(i7, 36)
- v17 = vec_load_f(p0, i1, descr=floatarraydescr)
+ v17 = vec_load_f(p0, i1, 8, 0, descr=floatarraydescr)
v19 = vec_cast_float_to_singlefloat(v17)
- v18 = vec_load_f(p0, i5, descr=floatarraydescr)
+ v18 = vec_load_f(p0, i5, 8, 0, descr=floatarraydescr)
v20 = vec_cast_float_to_singlefloat(v18)
v21 = vec_pack_i(v19, v20, 2, 2)
- vec_store(p1, i1, v21, descr=float32arraydescr)
+ vec_store(p1, i1, v21, 4, 0, descr=float32arraydescr)
jump(p0, p1, i50)
"""
loop = self.parse_loop(ops)
@@ -1291,21 +1290,18 @@
i500 = int_add(i4, 16)
i501 = int_lt(i500, 100)
guard_true(i501) [p0, p1, p2, i0, i4]
- i189 = int_add(i0, 4)
i187 = int_add(i4, 8)
i188 = int_lt(i187, 100)
- i207 = int_add(i0, 8)
i196 = int_add(i4, 12)
i197 = int_lt(i196, 100)
- i205 = int_add(i0, 12)
i400 = int_add(i4, 16)
i401= int_lt(i400, 100)
- i402 = int_add(i0, 16)
- v228[4xi32] = vec_load_i(p0, i0, descr=float32arraydescr)
+ v228[4xi32] = vec_load_i(p0, i0, 1, 0, descr=float32arraydescr)
v229[2xf64] = vec_cast_singlefloat_to_float(v228)
v230 = vec_unpack_i(v228, 2, 2)
v231 = vec_cast_singlefloat_to_float(v230)
- v232 = vec_load_i(p1, i189, descr=float32arraydescr)
+ i189 = int_add(i0, 4)
+ v232 = vec_load_i(p1, i189, 1, 0, descr=float32arraydescr)
v233 = vec_cast_singlefloat_to_float(v232)
v236 = vec_float_add(v229, v233)
v238 = vec_cast_float_to_singlefloat(v236)
@@ -1314,7 +1310,8 @@
v237 = vec_float_add(v231, v235)
v239 = vec_cast_float_to_singlefloat(v237)
v240 = vec_pack_i(v238, v239, 2, 2)
- vec_store(p2, i4, v240, descr=float32arraydescr)
+ vec_store(p2, i4, v240, 1, 0, descr=float32arraydescr)
+ i207 = int_add(i0, 16)
jump(p0, p1, p2, i207, i500)
""")
vopt = self.vectorize(trace)
@@ -1354,7 +1351,7 @@
""")
vopt = self.schedule(trace)
self.ensure_operations([
- 'v10[2xf64] = vec_load_f(p0,i0,descr=floatarraydescr)',
+ 'v10[2xf64] = vec_load_f(p0,i0,8,0,descr=floatarraydescr)',
'v11[2xf64] = vec_float_mul(v10[2xf64], v9[2xf64])',
'v12[2xf64] = vec_float_eq(v11[2xf64], v11[2xf64])',
'i100 = vec_unpack_f(v12[4xi32], 0, 1)',
_______________________________________________
pypy-commit mailing list
pypy-commit@python.org
https://mail.python.org/mailman/listinfo/pypy-commit
