Shapes have long been a part of nMigen, but represented using tuples.
This commit adds a Shape class (using namedtuple for backwards
compatibility), and accepts anything castable to Shape (including
enums, ranges, etc) anywhere a tuple was accepted previously.
In addition, `signed(n)` and `unsigned(n)` are added as aliases for
`Shape(n, signed=True)` and `Shape(n, signed=False)`, transforming
code such as `Signal((8, True))` to `Signal(signed(8))`.
These aliases are also included in prelude.
Preparation for #225.
This can cause confusion:
* If the erroneous object is None, it is printed as 'None', which
appears as a string (and could be the result of converting None
to a string.)
* If the erroneous object is a string, it is printed as ''<val>'',
which is a rather strange combination of quotes.
This change achieves two related goals.
First, default_rst is no longer assumed to be synchronous to
default_clk, which is the safer option, since it can be connected to
e.g. buttons on some evaluation boards.
Second, since the power-on / configuration reset is inherently
asynchronous to any user clock, the default create_missing_domain()
behavior is to use a reset synchronizer with `0` as input. Since,
like all reset synchronizers, it uses Signal(reset=1) for its
synchronization stages, after power-on reset it keeps its subordinate
clock domain in reset, and releases it after fabric flops start
toggling.
The latter change is helpful to architectures that lack an end-of-
configuration signal, i.e. most of them. ECP5 was already using
a similar scheme (and is not changed here). Xilinx devices with EOS
use EOS to drive a BUFGMUX, which is more efficient than using
a global reset when the design does not need one; Xilinx devices
without EOS use the new scheme. iCE40 requires a post-configuration
timer because of BRAM silicon bug, and was changed to add a reset
synchronizer if user clock is provided.
The write port priority in Yosys is derived directly from the order
in which the ports are declared in the Verilog frontend. It is being
removed for several reasons:
1. It is not clear if it works correctly for all cases (FFRAM,
LUTRAM, BRAM).
2. Although it is roundtripped via Verilog with correct simulation
semantics, the resulting code has a high chance of being
interpreted incorrectly by Xilinx tools.
3. It cannot be roundtripped via FIRRTL, which is an alternative
backend that is an interesting future option. (FIRRTL leaves
write collision completely undefined.)
3. It is a niche feature that, if it is needed, can be completely
replaced using an explicit comparator, priority encoder, and
write enable gating circuit. (This is what Xilinx recommends
for handling this case.)
In the future we should extend nMigen's formal verification to assert
that a write collision does not happen.
Although useful for debugging, most external tools often complain
about such attributes (with notable exception of Vivado). As such,
it is better to emit Verilog with these attributes into a separate
file such as `design.debug.v` and only emit the attributes that were
explicitly placed by the user to `design.v`.
This still leaves the (*init*) attribute. See #220 for details.
Since we use hertz elsewhere, this provides for easy conversions.
Also, cast the delay to string before applying it in xilinx_7series,
to avoid stripping the fractional digits.
Closes#234.
Almost no code would specify Signal(_, name) as a positional argument
on purpose, but forgetting parens and accidentally placing signedness
into the name position is so common that we had a test for it.
Unless exact_depth=True is specified.
The logic introduced in this commit is idempotent: that is, if one
uses the depth of one AsyncFIFOBuffered in the constructor of another
AsyncFIFOBuffered, they will end up with the same depth. More naive
logic would result in an unbounded, quadratic growth with each such
step.
Fixes#219.
These functions were originally changed in 3ed51938, in an attempt
to make them take one cycle instead of two. However, this does not
actually work because of drawbacks of the simulator interface.
Avoid committing to any specific implementation for now, and instead
make them compat-only extensions.
Before this commit, it was possible to set and get clock constraints
placed on Pin objects. This was not a very good implementation, since
it relied on matching the identity of the provided Pin object to
a previously requested one. The only reason it worked like that is
deficiencies in nextpnr.
Since then, nextpnr has been fixed to allow setting constraints on
arbitrary nets. Correspondingly, backends that are using Synplify
were changed to use [get_nets] instead of [get_ports] in SDC files.
However, in some situations, Synplify does not allow specifying
ports in [get_nets]. (In fact, nextpnr had a similar problem, but
it has also been fixed.)
The simplest way to address this is to refer to the interior net
(after the input buffer), which always works. The only downside
of this is that requesting a clock as a raw pin using
platform.request("clk", dir="-")
and directly applying a constraint to it could fail in some cases.
This is not a significant issue.
Because of YosysHQ/yosys#1390, using a transparent port in AsyncFIFO,
instead of being a no-op (as the semantics of \TRANSPARENT would
require it to be in this case), results in a failure to infer BRAM.
This can be easily avoided by using a non-transparent port instead,
which produces the desirable result with Yosys. It does not affect
the semantics on Xilinx platforms, since the interaction between
the two ports in case of address collision is undefined in either
transparent (WRITE_FIRST) or non-transparent (READ_FIRST) case, and
the data out of the write port is not used at all.
Fixes#172.
This is necessary for consistency, since for transparent read ports,
we currently do not support .en at all (it is fixed at 1) due to
YosysHQ/yosys#760. Before this commit, changing port transparency
would require adding or removing an assignment to .en, which is
confusing and error-prone.
Also, most read ports are always enabled, so this behavior is also
convenient.
Also, replace `bits, sign = x.shape()` with more idiomatic
`width, signed = x.shape()`.
This unifies all properties corresponding to `len(x)` to `x.width`.
(Not all values have a `width` property.)
Fixes#210.
It's not practical to detect tools within the toolchain environment
for various reasons, so just assume the tools are there if the user
says they are.
Before this commit, the tools would be searched outside the toolchain
environment, which of course would always fail for Vivado, ISE, etc.
This obscure functionality was likely only ever used in old MiSoC
code, and doesn't justify the added complexity. It was also not
provided (and could not be reasonably provided) in SyncFIFOBuffered,
which made its utility extremely marginal.
This is a somewhat obscure use case, but it is possible to use async
functions with pysim by carefully using @asyncio.coroutine. That is,
async functions can call back into pysim if they are declared in
a specific way:
@asyncio.coroutine
def do_something(self, value):
yield self.reg.eq(value)
which may then be called from elsewhere with:
async def test_case(self):
await do_something(0x1234)
This approach is unfortunately limited in that async functions
cannot yield directly. It should likely be improved by using async
generators, but supporting coroutines in pysim is unobtrustive and
allows existing code that made use of this feature in oMigen to work.
Platform.prepare() was completely broken after addition of local
clock domains, and only really worked before by a series of
accidents because there was a circular dependency between creation
of missing domains, fragment preparation, and insertion of pin
subfragments.
This commit untangles the dependency by adding a separate public
method Fragment.create_missing_domains(), used in build.plat.
It also makes DomainCollector consider both used and defined domains,
such that it will work on fragments before domain propagation, since
create_missing_domains() can be called by user code before prepare().
The fragment driving missing clock domain is not flattened anymore,
because flattening does not work well combined with local domains.
Because Fragment.prepare is not (currently) idempotent, it is useful
to be able to avoid calling it when converting. Even if it is made
idempotent, it can be slow on large designs, so it is advantageous
regardless of that.
Before this commit, the TransformedElaboratable of a CompatModule
would be ignored, and .get_fragment() would be used to retrieve
the CompatModule within.
After this commit, the finalization process is reworked to match
oMigen's finalization closely, and all submodules, native and compat,
are added in the same way that preserves applied transforms.
On Xilinx devices, flip-flops are reset to their initial state with
an internal global reset network, but this network is deasserted
asynchronously to user clocks. Use BUFGCE and STARTUP to hold default
clock low until after GWE is deasserted.
Previously changed in 27063a3b.
I haven't realized the .bin file is the same as the .bit file without
a small header. That means generating it is free and it's just easier
to let programming tools to be able to always rely on its existence.
Using 'x is legal RTLIL, in theory, but in practice it crashes Yosys
and when it doesn't, it causes Yosys to produce invalid Verilog.
Using a dummy wire is always safe and is not a major readability
issue as this is a rare corner case.
(It is not trivial to shorten the RHS in this case, because during
expansion of an ArrayProxy, match_shape() could be called in
a context far from the RHS handling logic.)
The elaboratable is already likely driving the clk/rst signals in
some way appropriate for the platform; if we expose them as ports
nevertheless it will cause problems downstream.
This pattern usually produces an extremely hard to notice bug that
will usually break a design when it is triggered, but will also be
hidden unless the pathological value of a boolean switch is used.
Fixes#159.
This might help with propagation of locations through optimizer
passes, since not all of them take care to preserve cells at all,
but usually wires stay intact when possible.
Also fixes incorrect source location on value.part().
Before this commit, it was a print statement, and therefore, command
interpreter options like -Wignore did not affect it. There is no API
to access the warning filter list, so it was turned into a real
warning; and further, since Python 3.6, tracemalloc can be used
as a standard method to display traceback to allocation site instead
of the ad-hoc traceback logic that was used in Elaboratable before.
This gives particularly pathological results on IO buffers, like:
connect \D_OUT_0 \user_led_0_user_led_0__o
Since subfragment signals are name-prefixed because this works well
for signals propagated upwards across hierarchy, this is never
desirable for instances.
This is useful when most attributes in a large composite resource
are the same, but a few signals are different, and also when building
abstractions around resources.
Fixes#128.
This primarily fixes the problem with source location precision in
Module (which used to trace locations from __exit__ of the context
managers, by which point everything interesting has been lost), but
also improves memory port and control inserter source locations.
On the sample of examples/basic/*.py, the only incorrectly inferred
remaining location is clk pointing to hdl/mem.py:166.
This effectively reverts and reimplements half of commit 82903e49.
I was confused and did not realize that RTLIL does, in fact, have
attributes on switches.
After this commit, processes no longer have any source locations.
This is semantically fine, as the processes we emit are purely
artificial (because of LHS grouping), but I have not checked how
downstream tooling handles this.
Right now an array is expected in any _opts overrides, and if it is
actually a string (because it is passed via an environment variable,
usually), awkwardness results as each character is joined with ` `.
Fixes#130.
This is useful for two reasons:
1. nMigen can provide better error messages than the platform and
do it earlier in the build pipeline.
2. Many platforms handle diffpairs by only constraining the P pin;
the N pin is completely ignored. If this is undetected,
downstream users (human or software) can rely on this
information assuming it is correct and introduce more errors.
(Of course, this will not catch every mistake, but the most
common is a copy-paste issue, and that will handle it.)
Fixes#124.
In compat.fhdl.module, we already default to "sync" as the default
clocked domain. Using "sys" in memories only would be inconsistent
and result in more bugs.
According to RTLIL semantics (that was undocumented before today),
the only purpose of `sync always` is to enable inference of latches,
because there is no other way to express them in terms of RTLIL
processes without ending up with a combinatorial loop. But, nMigen
specifically avoids latches, so this is not necessary.
This change results in major improvements in Verilog readability.
See also #98.
Commit 300d47ca introduced the same bug commit 779f3ee9 was trying to
avoid, but now only in the simulator. Since the names in simulator
don't have to make any sense, just use DUID to generate them.
This was rewritten to use Yosys cells in 779f3ee9 to avoid leaking
the interior clock domain, but the simulator doesn't understand Yosys
cells. So, use the old implementation in the simulator.
Such clock domains will "leak" into the enclosing scope, which is
generally undesirable. Also, this is instructive for a platform
overriding the behavior, since it provides guidance on how to
correctly instantiate platform-specific flops.
I've considered also doing this for MultiReg(), but it is very
challenging in presence of non-reset-less CDC FFs, since Yosys'
$dffsr primitive has separate set and clear inputs, and reshuffling
the reset value for those results in quite a bit of additional logic.
(That said, it might have to be done anyway, precisely because
letting Yosys generate this additional logic might prove too much
for the toolchain to cope with, and again, platform-independent
code should provide guidance to platform-specific code.)
This means that instead of:
with m.Case(0b00):
<body>
with m.Case(0b01):
<body>
it is legal to write:
with m.Case(0b00, 0b01):
<body>
with no change in semantics, and slightly nicer RTLIL or Verilog
output.
Fixes#103.
