usb-tools/amaranth
1
0
Fork 0
Code Issues Pull requests Actions Packages Projects Releases Wiki Activity

hdl.ast: remove Shape<>tuple casts.

Browse source 
Closes #691.
This commit is contained in:
Catherine 2023-01-31 12:24:56 +00:00
parent 309f647c0e
commit 29502442fb
6 changed files with 116 additions and 144 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

126
amaranth/back/rtlil.py
View file

@ -430,7 +430,7 @@ class _ValueCompiler(xfrm.ValueVisitor):
elem = value.elems[index.value]
else:
elem = value.elems[-1]
return self.match_shape(elem, *value.shape())
return self.match_shape(elem, value.shape())
else:
max_index = 1 << len(value.index)
max_elem = len(value.elems)
@ -475,12 +475,12 @@ class _RHSValueCompiler(_ValueCompiler):
if value in self.s.anys:
return self.s.anys[value]
res_bits, res_sign = value.shape()
res = self.s.rtlil.wire(width=res_bits, src=_src(value.src_loc))
res_shape = value.shape()
res = self.s.rtlil.wire(width=res_shape.width, src=_src(value.src_loc))
self.s.rtlil.cell("$anyconst", ports={
"\\Y": res,
}, params={
"WIDTH": res_bits,
"WIDTH": res_shape.width,
}, src=_src(value.src_loc))
self.s.anys[value] = res
return res
@ -489,12 +489,12 @@ class _RHSValueCompiler(_ValueCompiler):
if value in self.s.anys:
return self.s.anys[value]
res_bits, res_sign = value.shape()
res = self.s.rtlil.wire(width=res_bits, src=_src(value.src_loc))
res_shape = value.shape()
res = self.s.rtlil.wire(width=res_shape.width, src=_src(value.src_loc))
self.s.rtlil.cell("$anyseq", ports={
"\\Y": res,
}, params={
"WIDTH": res_bits,
"WIDTH": res_shape.width,
}, src=_src(value.src_loc))
self.s.anys[value] = res
return res
@ -509,74 +509,71 @@ class _RHSValueCompiler(_ValueCompiler):
# These operators don't change the bit pattern, only its interpretation.
return self(arg)
arg_bits, arg_sign = arg.shape()
res_bits, res_sign = value.shape()
res = self.s.rtlil.wire(width=res_bits, src=_src(value.src_loc))
arg_shape, res_shape = arg.shape(), value.shape()
res = self.s.rtlil.wire(width=res_shape.width, src=_src(value.src_loc))
self.s.rtlil.cell(self.operator_map[(1, value.operator)], ports={
"\\A": self(arg),
"\\Y": res,
}, params={
"A_SIGNED": arg_sign,
"A_WIDTH": arg_bits,
"Y_WIDTH": res_bits,
"A_SIGNED": arg_shape.signed,
"A_WIDTH": arg_shape.width,
"Y_WIDTH": res_shape.width,
}, src=_src(value.src_loc))
return res
def match_shape(self, value, new_bits, new_sign):
def match_shape(self, value, new_shape):
if isinstance(value, ast.Const):
return self(ast.Const(value.value, ast.Shape(new_bits, new_sign)))
return self(ast.Const(value.value, new_shape))
value_bits, value_sign = value.shape()
if new_bits <= value_bits:
return self(ast.Slice(value, 0, new_bits))
value_shape = value.shape()
if new_shape.width <= value_shape.width:
return self(ast.Slice(value, 0, new_shape.width))
res = self.s.rtlil.wire(width=new_bits, src=_src(value.src_loc))
res = self.s.rtlil.wire(width=new_shape.width, src=_src(value.src_loc))
self.s.rtlil.cell("$pos", ports={
"\\A": self(value),
"\\Y": res,
}, params={
"A_SIGNED": value_sign,
"A_WIDTH": value_bits,
"Y_WIDTH": new_bits,
"A_SIGNED": value_shape.signed,
"A_WIDTH": value_shape.width,
"Y_WIDTH": new_shape.width,
}, src=_src(value.src_loc))
return res
def on_Operator_binary(self, value):
lhs, rhs = value.operands
lhs_bits, lhs_sign = lhs.shape()
rhs_bits, rhs_sign = rhs.shape()
if lhs_sign == rhs_sign or value.operator in ("<<", ">>", "**"):
lhs_shape, rhs_shape, res_shape = lhs.shape(), rhs.shape(), value.shape()
if lhs_shape.signed == rhs_shape.signed or value.operator in ("<<", ">>", "**"):
lhs_wire = self(lhs)
rhs_wire = self(rhs)
else:
lhs_bits = rhs_bits = max(lhs_bits + rhs_sign, rhs_bits + lhs_sign)
lhs_sign = rhs_sign = True
lhs_wire = self.match_shape(lhs, lhs_bits, lhs_sign)
rhs_wire = self.match_shape(rhs, rhs_bits, rhs_sign)
res_bits, res_sign = value.shape()
res = self.s.rtlil.wire(width=res_bits, src=_src(value.src_loc))
lhs_shape = rhs_shape = ast.signed(max(lhs_shape.width + rhs_shape.signed,
rhs_shape.width + lhs_shape.signed))
lhs_wire = self.match_shape(lhs, lhs_shape)
rhs_wire = self.match_shape(rhs, rhs_shape)
res = self.s.rtlil.wire(width=res_shape.width, src=_src(value.src_loc))
self.s.rtlil.cell(self.operator_map[(2, value.operator)], ports={
"\\A": lhs_wire,
"\\B": rhs_wire,
"\\Y": res,
}, params={
"A_SIGNED": lhs_sign,
"A_WIDTH": lhs_bits,
"B_SIGNED": rhs_sign,
"B_WIDTH": rhs_bits,
"Y_WIDTH": res_bits,
"A_SIGNED": lhs_shape.signed,
"A_WIDTH": lhs_shape.width,
"B_SIGNED": rhs_shape.signed,
"B_WIDTH": rhs_shape.width,
"Y_WIDTH": res_shape.width,
}, src=_src(value.src_loc))
if value.operator in ("//", "%"):
# RTLIL leaves division by zero undefined, but we require it to return zero.
divmod_res = res
res = self.s.rtlil.wire(width=res_bits, src=_src(value.src_loc))
res = self.s.rtlil.wire(width=res_shape.width, src=_src(value.src_loc))
self.s.rtlil.cell("$mux", ports={
"\\A": divmod_res,
"\\B": self(ast.Const(0, ast.Shape(res_bits, res_sign))),
"\\B": self(ast.Const(0, res_shape)),
"\\S": self(rhs == 0),
"\\Y": res,
}, params={
"WIDTH": res_bits
"WIDTH": res_shape.width
}, src=_src(value.src_loc))
return res
@ -584,20 +581,17 @@ class _RHSValueCompiler(_ValueCompiler):
sel, val1, val0 = value.operands
if len(sel) != 1:
sel = sel.bool()
val1_bits, val1_sign = val1.shape()
val0_bits, val0_sign = val0.shape()
res_bits, res_sign = value.shape()
val1_bits = val0_bits = res_bits = max(val1_bits, val0_bits, res_bits)
val1_wire = self.match_shape(val1, val1_bits, val1_sign)
val0_wire = self.match_shape(val0, val0_bits, val0_sign)
res = self.s.rtlil.wire(width=res_bits, src=_src(value.src_loc))
res_shape = value.shape()
val1_wire = self.match_shape(val1, res_shape)
val0_wire = self.match_shape(val0, res_shape)
res = self.s.rtlil.wire(width=res_shape.width, src=_src(value.src_loc))
self.s.rtlil.cell("$mux", ports={
"\\A": val0_wire,
"\\B": val1_wire,
"\\S": self(sel),
"\\Y": res,
}, params={
"WIDTH": res_bits
"WIDTH": res_shape.width
}, src=_src(value.src_loc))
return res
@ -624,10 +618,8 @@ class _RHSValueCompiler(_ValueCompiler):
lhs, rhs = value.value, value.offset
if value.stride != 1:
rhs *= value.stride
lhs_bits, lhs_sign = lhs.shape()
rhs_bits, rhs_sign = rhs.shape()
res_bits, res_sign = value.shape()
res = self.s.rtlil.wire(width=res_bits, src=_src(value.src_loc))
lhs_shape, rhs_shape, res_shape = lhs.shape(), rhs.shape(), value.shape()
res = self.s.rtlil.wire(width=res_shape.width, src=_src(value.src_loc))
# Note: Verilog's x[o+:w] construct produces a $shiftx cell, not a $shift cell.
# However, Amaranth's semantics defines the out-of-range bits to be zero, so it is correct
# to use a $shift cell here instead, even though it produces less idiomatic Verilog.
@ -636,11 +628,11 @@ class _RHSValueCompiler(_ValueCompiler):
"\\B": self(rhs),
"\\Y": res,
}, params={
"A_SIGNED": lhs_sign,
"A_WIDTH": lhs_bits,
"B_SIGNED": rhs_sign,
"B_WIDTH": rhs_bits,
"Y_WIDTH": res_bits,
"A_SIGNED": lhs_shape.signed,
"A_WIDTH": lhs_shape.width,
"B_SIGNED": rhs_shape.signed,
"B_WIDTH": rhs_shape.width,
"Y_WIDTH": res_shape.width,
}, src=_src(value.src_loc))
return res
@ -666,14 +658,14 @@ class _LHSValueCompiler(_ValueCompiler):
raise TypeError # :nocov:
def match_shape(self, value, new_bits, new_sign):
value_bits, value_sign = value.shape()
if new_bits == value_bits:
def match_shape(self, value, new_shape):
value_shape = value.shape()
if new_shape.width == value_shape.width:
return self(value)
elif new_bits < value_bits:
return self(ast.Slice(value, 0, new_bits))
else: # new_bits > value_bits
dummy_bits = new_bits - value_bits
elif new_shape.width < value_shape.width:
return self(ast.Slice(value, 0, new_shape.width))
else: # new_shape.width > value_shape.width
dummy_bits = new_shape.width - value_shape.width
dummy_wire = self.s.rtlil.wire(dummy_bits)
return "{{ {} {} }}".format(dummy_wire, self(value))
@ -738,14 +730,12 @@ class _StatementCompiler(xfrm.StatementVisitor):
def on_Assign(self, stmt):
self._check_rhs(stmt.rhs)
lhs_bits, lhs_sign = stmt.lhs.shape()
rhs_bits, rhs_sign = stmt.rhs.shape()
if lhs_bits == rhs_bits:
lhs_shape, rhs_shape = stmt.lhs.shape(), stmt.rhs.shape()
if lhs_shape.width == rhs_shape.width:
rhs_sigspec = self.rhs_compiler(stmt.rhs)
else:
# In RTLIL, LHS and RHS of assignment must have exactly same width.
rhs_sigspec = self.rhs_compiler.match_shape(
stmt.rhs, lhs_bits, lhs_sign)
rhs_sigspec = self.rhs_compiler.match_shape(stmt.rhs, lhs_shape)
if self._wrap_assign:
# In RTLIL, all assigns are logically sequenced before all switches, even if they are
# interleaved in the source. In Amaranth, the source ordering is used. To handle this

5
amaranth/compat/fhdl/structure.py
View file

@ -51,7 +51,10 @@ class CompatSignal(NativeSignal):
else:
if not (min is None and max is None):
raise ValueError("Only one of bits/signedness or bounds may be specified")
shape = bits_sign
if isinstance(bits_sign, tuple):
shape = Shape(*bits_sign)
else:
shape = Shape.cast(bits_sign)
super().__init__(shape=shape, name=name_override or name,
reset=reset, reset_less=reset_less,

97
amaranth/hdl/ast.py
View file

@ -78,10 +78,6 @@ class Shape:
self.width = width
self.signed = signed
# TODO(nmigen-0.4): remove
def __iter__(self):
return iter((self.width, self.signed))
@staticmethod
def cast(obj, *, src_loc_at=0):
while True:
@ -89,14 +85,6 @@ class Shape:
return obj
elif isinstance(obj, int):
return Shape(obj)
# TODO(nmigen-0.4): remove
elif isinstance(obj, tuple):
width, signed = obj
warnings.warn("instead of `{tuple}`, use `{constructor}({width})`"
.format(constructor="signed" if signed else "unsigned", width=width,
tuple=obj),
DeprecationWarning, stacklevel=2 + src_loc_at)
return Shape(width, signed)
elif isinstance(obj, range):
if len(obj) == 0:
return Shape(0, obj.start < 0)
@ -216,8 +204,7 @@ class Value(metaclass=ABCMeta):
return Operator("//", [other, self])
def __check_shamt(self):
width, signed = self.shape()
if signed:
if self.shape().signed:
# Neither Python nor HDLs implement shifts by negative values; prohibit any shifts
# by a signed value to make sure the shift amount can always be interpreted as
# an unsigned value.
@ -264,8 +251,7 @@ class Value(metaclass=ABCMeta):
return Operator(">=", [self, other])
def __abs__(self):
width, signed = self.shape()
if signed:
if self.shape().signed:
return Mux(self >= 0, self, -self)
else:
return self
@ -607,10 +593,9 @@ class Const(Value):
@staticmethod
def normalize(value, shape):
width, signed = shape
mask = (1 << width) - 1
mask = (1 << shape.width) - 1
value &= mask
if signed and value >> (width - 1):
if shape.signed and value >> (shape.width - 1):
value |= ~mask
return value
@ -623,7 +608,8 @@ class Const(Value):
shape = Shape(shape, signed=self.value < 0)
else:
shape = Shape.cast(shape, src_loc_at=1 + src_loc_at)
self.width, self.signed = shape
self.width = shape.width
self.signed = shape.signed
self.value = self.normalize(self.value, shape)
def shape(self):
@ -645,10 +631,9 @@ C = Const # shorthand
class AnyValue(Value, DUID):
def __init__(self, shape, *, src_loc_at=0):
super().__init__(src_loc_at=src_loc_at)
self.width, self.signed = Shape.cast(shape, src_loc_at=1 + src_loc_at)
if not isinstance(self.width, int) or self.width < 0:
raise TypeError("Width must be a non-negative integer, not {!r}"
.format(self.width))
shape = Shape.cast(shape, src_loc_at=1 + src_loc_at)
self.width = shape.width
self.signed = shape.signed
def shape(self):
return Shape(self.width, self.signed)
@ -678,55 +663,53 @@ class Operator(Value):
def shape(self):
def _bitwise_binary_shape(a_shape, b_shape):
a_bits, a_sign = a_shape
b_bits, b_sign = b_shape
if not a_sign and not b_sign:
if not a_shape.signed and not b_shape.signed:
# both operands unsigned
return Shape(max(a_bits, b_bits), False)
elif a_sign and b_sign:
return unsigned(max(a_shape.width, b_shape.width))
elif a_shape.signed and b_shape.signed:
# both operands signed
return Shape(max(a_bits, b_bits), True)
elif not a_sign and b_sign:
return signed(max(a_shape.width, b_shape.width))
elif not a_shape.signed and b_shape.signed:
# first operand unsigned (add sign bit), second operand signed
return Shape(max(a_bits + 1, b_bits), True)
return signed(max(a_shape.width + 1, b_shape.width))
else:
# first signed, second operand unsigned (add sign bit)
return Shape(max(a_bits, b_bits + 1), True)
return signed(max(a_shape.width, b_shape.width + 1))
op_shapes = list(map(lambda x: x.shape(), self.operands))
if len(op_shapes) == 1:
(a_width, a_signed), = op_shapes
a_shape, = op_shapes
if self.operator in ("+", "~"):
return Shape(a_width, a_signed)
return Shape(a_shape.width, a_shape.signed)
if self.operator == "-":
return Shape(a_width + 1, True)
return Shape(a_shape.width + 1, True)
if self.operator in ("b", "r|", "r&", "r^"):
return Shape(1, False)
if self.operator == "u":
return Shape(a_width, False)
return Shape(a_shape.width, False)
if self.operator == "s":
return Shape(a_width, True)
return Shape(a_shape.width, True)
elif len(op_shapes) == 2:
(a_width, a_signed), (b_width, b_signed) = op_shapes
a_shape, b_shape = op_shapes
if self.operator in ("+", "-"):
width, signed = _bitwise_binary_shape(*op_shapes)
return Shape(width + 1, signed)
o_shape = _bitwise_binary_shape(*op_shapes)
return Shape(o_shape.width + 1, o_shape.signed)
if self.operator == "*":
return Shape(a_width + b_width, a_signed or b_signed)
return Shape(a_shape.width + b_shape.width, a_shape.signed or b_shape.signed)
if self.operator == "//":
return Shape(a_width + b_signed, a_signed or b_signed)
return Shape(a_shape.width + b_shape.signed, a_shape.signed or b_shape.signed)
if self.operator == "%":
return Shape(b_width, b_signed)
return Shape(b_shape.width, b_shape.signed)
if self.operator in ("<", "<=", "==", "!=", ">", ">="):
return Shape(1, False)
if self.operator in ("&", "^", "|"):
return _bitwise_binary_shape(*op_shapes)
if self.operator == "<<":
assert not b_signed
return Shape(a_width + 2 ** b_width - 1, a_signed)
assert not b_shape.signed
return Shape(a_shape.width + 2 ** b_shape.width - 1, a_shape.signed)
if self.operator == ">>":
assert not b_signed
return Shape(a_width, a_signed)
assert not b_shape.signed
return Shape(a_shape.width, a_shape.signed)
elif len(op_shapes) == 3:
if self.operator == "m":
s_shape, a_shape, b_shape = op_shapes
@ -982,9 +965,13 @@ class Signal(Value, DUID):
raise TypeError("Name must be a string, not {!r}".format(name))
self.name = name or tracer.get_var_name(depth=2 + src_loc_at, default="$signal")
orig_shape = shape
if shape is None:
shape = unsigned(1)
self.width, self.signed = Shape.cast(shape, src_loc_at=1 + src_loc_at)
else:
shape = Shape.cast(shape, src_loc_at=1 + src_loc_at)
self.width = shape.width
self.signed = shape.signed
if isinstance(reset, Enum):
reset = reset.value
@ -1003,8 +990,8 @@ class Signal(Value, DUID):
self.attrs = OrderedDict(() if attrs is None else attrs)
if decoder is None and isinstance(shape, type) and issubclass(shape, Enum):
decoder = shape
if decoder is None and isinstance(orig_shape, type) and issubclass(orig_shape, Enum):
decoder = orig_shape
if isinstance(decoder, type) and issubclass(decoder, Enum):
def enum_decoder(value):
try:
@ -1231,13 +1218,13 @@ class ArrayProxy(Value):
def shape(self):
unsigned_width = signed_width = 0
has_unsigned = has_signed = False
for elem_width, elem_signed in (elem.shape() for elem in self._iter_as_values()):
if elem_signed:
for elem_shape in (elem.shape() for elem in self._iter_as_values()):
if elem_shape.signed:
has_signed = True
signed_width = max(signed_width, elem_width)
signed_width = max(signed_width, elem_shape.width)
else:
has_unsigned = True
unsigned_width = max(unsigned_width, elem_width)
unsigned_width = max(unsigned_width, elem_shape.width)
# The shape of the proxy must be such that it preserves the mathematical value of the array
# elements. I.e., shape-wise, an array proxy must be identical to an equivalent mux tree.
# To ensure this holds, if the array contains both signed and unsigned values, make sure

3
amaranth/hdl/dsl.py
View file

@ -210,8 +210,7 @@ class Module(_ModuleBuilderRoot, Elaboratable):
def _check_signed_cond(self, cond):
cond = Value.cast(cond)
width, signed = cond.shape()
if signed:
if cond.shape().signed:
warnings.warn("Signed values in If/Elif conditions usually result from inverting "
"Python booleans with ~, which leads to unexpected results. "
"Replace `~flag` with `not flag`. (If this is a false positive, "

1
docs/changes.rst
View file

@ -24,6 +24,7 @@ Language changes
.. currentmodule:: amaranth.hdl
* Removed: casting of :class:`Shape` to and from a ``(width, signed)`` tuple.
* Added: :class:`ShapeCastable`, similar to :class:`ValueCastable`.
* Added: :meth:`Value.as_signed` and :meth:`Value.as_unsigned` can be used on left-hand side of assignment (with no difference in behavior).
* Changed: :meth:`Value.cast` casts :class:`ValueCastable` objects recursively.

22
tests/test_hdl_ast.py
View file

@ -53,17 +53,18 @@ class ShapeTestCase(FHDLTestCase):
def test_compare_tuple_wrong(self):
with self.assertRaisesRegex(TypeError,
r"^Shapes may be compared with other Shapes and \(int, bool\) tuples, not \(2, 3\)$"):
r"^Shapes may be compared with other Shapes and \(int, bool\) tuples, "
r"not \(2, 3\)$"):
Shape(1, True) == (2, 3)
def test_repr(self):
self.assertEqual(repr(Shape()), "unsigned(1)")
self.assertEqual(repr(Shape(2, True)), "signed(2)")
def test_tuple(self):
def test_convert_tuple_wrong(self):
with self.assertRaisesRegex(TypeError,
r"^cannot unpack non-iterable Shape object$"):
width, signed = Shape()
self.assertEqual(width, 1)
self.assertEqual(signed, False)
def test_unsigned(self):
s1 = unsigned(2)
@ -95,19 +96,10 @@ class ShapeTestCase(FHDLTestCase):
r"^Width must be a non-negative integer, not -1$"):
Shape.cast(-1)
def test_cast_tuple(self):
with warnings.catch_warnings():
warnings.filterwarnings(action="ignore", category=DeprecationWarning)
s1 = Shape.cast((1, True))
self.assertEqual(s1.width, 1)
self.assertEqual(s1.signed, True)
def test_cast_tuple_wrong(self):
with warnings.catch_warnings():
warnings.filterwarnings(action="ignore", category=DeprecationWarning)
with self.assertRaisesRegex(TypeError,
r"^Width must be a non-negative integer, not -1$"):
Shape.cast((-1, True))
r"^Object \(1, True\) cannot be converted to an Amaranth shape$"):
Shape.cast((1, True))
def test_cast_range(self):
s1 = Shape.cast(range(0, 8))