The previous approach only initialized past_d and past_ad while for FST
cosim we also need to initialize the other past values like past_clk,
etc. Also to properly initialize them, we need to run a combinational
update step in case any of the wires feeding into the FF are private or
otherwise not part of the FST.
Otherwise the final values will not show up in gtkwave waveforms
when looking at the generated traces.
Signed-off-by: Claire Xenia Wolf <claire@clairexen.net>
* Change simlib's $mux cell to use the ternary operator as $_MUX_
already does
* Stop opt_expr -keepdc from changing S=x to S=0
* Change const eval of $mux and $pmux to match the updated simlib
(fixes sim)
* The sat behavior of $mux already matches the updated simlib
The verilog frontend uses $mux for the ternary operators and this
changes all interpreations of the $mux cell (that I found) to match the
verilog simulation behavior for the ternary operator. For 'if' and
'case' expressions the frontend may also use $mux but uses $eqx if the
verilog simulation behavior is requested with the '-ifx' option.
For $pmux there is a remaining mismatch between the sat behavior and the
simlib behavior. Resolving this requires more discussion, as the $pmux
cell does not directly correspond to a specific verilog construct.
This reverts commit 08be796cb8, reversing
changes made to 38dbb44fa0.
This fixes#2728. PR #641 did not actually "fix" #639.
The actual issue in #639 is not equiv_make, but assumptions in equiv_simple
that are not true for the test case provided in #639.
This is a complete rewrite of the FF replacing code.
The previous implementation tried to implement the negative hold time by
wrapping async control signals individually with pulse stretching. This
did not correctly model the interaction between different simultaneously
changing inputs (e.g. a falling ALOAD together with a changing AD would
load the changed AD instead of the value AD had when ALOAD was high; a
falling CLR could mask a raising SET for one cycle; etc.).
The new approach first has the logic for all updates using only sampled
values followed by the logic for all updates using only current values.
That way, e.g., a falling ALOAD will load the sampled AD value but a
still active ALOAD will load the current AD value.
The new code also has deterministic behavior for the initial state: no
operation is active when that operation would depend on a specific
previous signal value. This also means clk2fflogic will no longer
generate any additional uninitialized FFs.
I also documented the negative hold time behavior in the help message,
copying the relevant part from async2sync's help messages.
In 2fcc1ee72e, the following is apparantly added in order to mark any
number of undefined LUT inputs:
lut_a.append(RTLIL::Const(State::Sx, minlut - input_nodes.size()));
However this can only be done if the number of input nodes is less
than minlut.
This fixes#3317
After speaking with the author of ABC he let me know that ifraig is a very old command, and that &get; &fraig -x; &put is over 100x faster than ifraig with improved PPA results.
After making the change I confirmed that this is in fact a major speed up. On our internal designs in O(millions) of standard cells we saw multi hour reductions in runtime.
Also included is an improvement to the dress command. Using AIG based transformations removes the spec it SATs against. Proving the input blif will make sure that no matter what commands are run the dress command can still do its job. I noticed a regression against some LUT mapping jobs that prompted me to fix this.
Uses the regex below to search (using vscode):
^\t\tlog\("(.{10,}(?<!\\n)|.{81,}\\n)"\);
Finds any log messages double indented (which help messages are)
and checks if *either* there are is no newline character at the end,
*or* the number of characters before the newline is more than 80.
Verific generates a lot of FFs with an unused async load and we cannot
always optimize that away before running clk2fflogic, so check for that
special case here.
Find FFs with undefined initialization values for which changing the
initialization does not change the observable behavior and initialize
them. For -ff2anyinit, this reduces the number of generated $anyinit
cells that drive wires with private names.
This attribute can be used by formal backends to indicate which clocks
were mapped to the global clock. Update the btor and smt2 backend which
already handle clock inputs to understand this attribute.
These can be used to protect undefined flip-flop initialization values
from optimizations that are not sound for formal verification and can
help mapping all solver-provided values in witness traces for flows that
use different backends simultaneously.
The formal backends do not support multiple clocks. This includes
constant clocks. Constant clocks do appear in what isn't a proper
multiclock design, for example when mapping not fully initialized ROMs.
As converting FFs with constant clocks to FFs using the global is doable
even in a single clock flow, make async2sync do this.
POSIX defines $TMPDIR as containing the pathname of the directory where
programs can create temporary files. On most systems, this variable points to
"/tmp". However, on some systems it can point to a different location.
Without respecting this variable, yosys fails to run on such systems.
Signed-off-by: Mohamed A. Bamakhrama <mohamed@alumni.tum.de>
Both of these options consider a selection containing only empty modules
as non-empty. This wasn't mentioned in the documentation nor did the
error message when using `select -assert-none` list those empty modules,
which produced a very confusing error message complaining about a
non-empty selection followed by an empty listing of the selection.
This fixes the documentation and changes the `-assert-none` and
`-assert-any` assertion error messages to also output fully selected
modules (this includes selected empty modules).
It doesn't change the messages for `-assert-count` etc. as they don't
count modules.
The `-keepdc` option prevents merging flipflops with dont-care bits in
their initial value, as, in general, this is not a valid transform for
formal verification.
The keepdc option of `opt` is passed along to `opt_merge` now.
Newer GCCs emit the following warning for opt_dff:
passes/opt/opt_dff.cc:560:17: warning: operation on ‘ff.Yosys::FfData::has_clk’ may be undefined [-Wsequence-point]
560 | ff.has_clk = ff.has_ce = ff.has_clk = false;
| ~~~~~~~~~~~^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Which is correct: the order of whether the read or write of has_clk
occurs first is undefined since there is no sequence point between them.
This is almost certainly just a typo/copy paste error and objectively
wrong, so just fix it.
Signed-off-by: Austin Seipp <aseipp@pobox.com>
- Attempt to lookup a derived module if it potentially contains a port
connection with elaboration ambiguities
- Mark the cell if module has not yet been derived
- This can be extended to implement automatic hierarchical port
connections in a future change
- FfData now keeps track of the module and underlying cell, if any (so
calling emit on FfData created from a cell will replace the existing cell)
- FfData implementation is split off to its own .cc file for faster
compilation
- the "flip FF data sense by inserting inverters in front and after"
functionality that zinit uses is moved onto FfData class and beefed up
to have dffsr support, to support more use cases
- *_en is split into *_ce (clock enable) and *_aload (async load aka
latch gate enable), so both can be present at once
- has_d is removed
- has_gclk is added (to have a clear marker for $ff)
- d_is_const and val_d leftovers are removed
- async2sync, clk2fflogic, opt_dff are updated to operate correctly on
FFs with async load
Mutating the SigMap by adding a new connection will throw off FfInitVals
index. Work around this by removing the relevant init values from index
whenever we connect nets, then re-add the new init value.
Should fix#2920.
Previously, opt_clean would reconnect all ports (including FF Q ports)
to a "canonical" SigBit chosen by complex rules, but would leave the
init attribute on the old wire. This change applies the same
canonicalization rules to the init attributes, ensuring that init moves
to wherever the Q port moved.
Part of another jab at #2920.
This also aligns the functionality:
- in all cases, the onehot attribute is used to create appropriate
constraints (previously, opt_dff didn't do it at all, and share
created one-hot constraints based on $pmux presence alone, which
is unsound)
- in all cases, shift and mul/div/pow cells are now skipped when
importing the SAT problem (previously only memory_share did this)
— this avoids creating clauses for hard cells that are unlikely
to help with proving the UNSATness needed for optimization
If width of a case expression was large, explicit patterns could cause
the existing logic to take an extremely long time, or exhaust the
maximum size of the underlying set. For cases where all of the patterns
are fully defined and there are no constants in the case expression,
this change uses a simple set to track which patterns have been seen.
This adds one simple piece of functionality to opt_expr: when a cell
port is connected to a fully-constant signal (as determined by sigmap),
the port is reconnected directly to the constant value. This is just
enough optimization to fix the "non-constant $meminit input" problem
without requiring a full opt_clean or a separate pass.
I think the code is now a bit easier to follow (and has lost some
levels of indentation!).
The only non-trivial change is that I removed the check for
cell->type[0] != '$' when deciding whether to complain if we couldn't
find a module. This will always be true because of the early exit
earlier in the function.
Spotted during compilation:
passes/proc/proc_init.cc: In function ‘void {anonymous}::proc_init(Yosys::RTLIL::Module*, Yosys::SigMap&, Yosys::RTLIL::Process*)’:
passes/proc/proc_init.cc:31:7: warning: variable ‘found_init’ set but not used [-Wunused-but-set-variable]
There should be no functional change, but this splits up the control
flow across functions, using class fields to hold the state that's
being tracked. The result should be a bit easier to read.
This is part of work to add bind support, but I'm doing some
refactoring in the hierarchy pass to make the code a bit easier to
work with. The idea is that (eventually) the IFExpander object will
hold all the logic for expanding interfaces, and then other code can
do bind insertion.
Turns out the code for div by a power of 2 is already almost capable of
optimizing this to a shift-by-0 or and-with-0, which will be further
folded into nothingness; let's beef it up to handle div by 1 as well.
Fixes#2820.
This essentially adds wide port support for free in passes that don't
have a usefully better way of handling wide ports than just breaking
them up to narrow ports, avoiding "please run memory_narrow" annoyance.
When converting a sync transparent read port with const address to async
read port, nothing at all needs to be done other than clk_enable change,
and thus we have no FF cell to return. Handle this case correctly in
the helper and in its users.
There will soon be more (versioned) memory cells, so handle passes that
only care if a cell is memory-related by a simple helper call instead of
a hardcoded list.
* xilinx: add SCC test for DSP48E1
* xilinx: Gate DSP48E1 being a whitebox behind ALLOW_WHITEBOX_DSP48E1
Have a test that checks it works through ABC9 when enabled
* abc9 to break SCCs using $__ABC9_SCC_BREAKER module
* Add test
* abc9_ops: remove refs to (* abc9_keep *) on wires
* abc9_ops: do not bypass cells in an SCC
* Add myself to CODEOWNERS for abc9*
* Fix compile
* abc9_ops: run -prep_hier before scc
* Fix tests
* Remove bug reference pending fix
* abc9: fix for -prep_hier -dff
* xaiger: restore PI handling
* abc9_ops: -prep_xaiger sigmap
* abc9_ops: -mark_scc -> -break_scc
* abc9: eliminate hard-coded abc9.box from tests
Also tidy up
* Address review
Bugpoint's current documentation does specify that the result of a run is stored as the current design,
however it's easy to skim over what that means in practice.
Add a documentation comment to explain specifically that an after bugpoint `write_xyz` pass is required to save
the reduced design.
The already-existing special case for conditionals on clock has been
remade as follows:
- now triggered for the last remaining edge trigger after all others
have been converted to async reset, not just when there is only one
sync rule in the first place
- does not require all contained assignments to be constant, as opposed
to a reset conditional — merely const-folds the condition
In addition, the code has been refactored a bit; as a bonus, the
priority order of async resets found is now preserved in resulting sync
rule ordering (though this is not yet respected by proc_dff).
Fixes#2656.
This change set contains a number of bug fixes and improvements related to
scoping and resolution in generate and procedural blocks. While many of the
frontend changes are interdependent, it may be possible bring the techmap
changes in under a separate PR.
Declarations within unnamed generate blocks previously encountered issues
because the data declarations were left un-prefixed, breaking proper scoping.
The LRM outlines behavior for generating names for unnamed generate blocks. The
original goal was to add this implicit labelling, but doing so exposed a number
of issues downstream. Additional testing highlighted other closely related scope
resolution issues, which have been fixed. This change also adds support for
block item declarations within unnamed blocks in SystemVerilog mode.
1. Unlabled generate blocks are now implicitly named according to the LRM in
`label_genblks`, which is invoked at the beginning of module elaboration
2. The Verilog parser no longer wraps explicitly named generate blocks in a
synthetic unnamed generate block to avoid creating extra hierarchy levels
where they should not exist
3. The techmap phase now allows special control identifiers to be used outside
of the topmost scope, which is necessary because such wires and cells often
appear in unlabeled generate blocks, which now prefix the declarations within
4. Some techlibs required modifications because they relied on the previous
invalid scope resolution behavior
5. `expand_genblock` has been simplified, now only expanding the outermost
scope, completely deferring the inspection and elaboration of nested scopes;
names are now resolved by looking in the innermost scope and stepping outward
6. Loop variables now always become localparams during unrolling, allowing them
to be resolved and shadowed like any other identifier
7. Identifiers in synthetic function call scopes are now prefixed and resolved
in largely the same manner as other blocks
before: `$func$\func_01$tests/simple/scopes.blk.v:60$5$\blk\x`
after: `\func_01$func$tests/simple/scopes.v:60$5.blk.x`
8. Support identifiers referencing a local generate scope nested more
than 1 level deep, i.e. `B.C.x` while within generate scope `A`, or using a
prefix of a current or parent scope, i.e. `B.C.D.x` while in `A.B`, `A.B.C`,
or `A.B.C.D`
9. Variables can now be declared within unnamed blocks in SystemVerilog mode
Addresses the following issues: 656, 2423, 2493
- Signed cell outputs are sign extended when bound to larger wires
- Signed connections are sign extended when bound to larger cell inputs
- Sign extension is performed in hierarchy and flatten phases
- genrtlil indirects signed constants through signed wires
- Other phases producing RTLIL may need to be updated to preserve
signedness information
- Resolves#1418
- Resolves#2265
When the register being merged into the EN signal happens to be a $sdff,
the current code creates a new $mux for every bit, even if they happen
to be identical (as is usually the case), preventing proper grouping
further down the flow. Fix this by adding a simple cache.
Fixes#2409.
Previously, `$memwr` and `$meminit` cells were always preserved (along
with the memory itself). With this change, they are instead part of the
main cell mark-and-sweep pass: a memory (and its `$meminit` and `$memwr`
cells) is only preserved iff any associated `$memrd` cell needs to be
preserved.
When an adffe is being legalized, and is not natively supported,
prioritize unmapping to adff over converting to dffsre if dffsre is not
natively supported itself.
Fixes#2361.
The only difference between "RTLIL" and "ILANG" is that the latter is
the text representation of the former, as opposed to the in-memory
graph representation. This distinction serves no purpose but confuses
people: it is not obvious that the ILANG backend writes RTLIL graphs.
Passes `write_ilang` and `read_ilang` are provided as aliases to
`write_rtlil` and `read_rtlil` for compatibility.
For connection `assign a = b;`, `sigmap(a)` returns `b`. This is
exactly the opposite of the desired canonicalization for driven bits.
Consider the following code:
module foo(inout a, b);
assign a = b;
endmodule
module bar(output c);
foo f(c, 1'b0);
endmodule
Before this commit, the inout ports would be swapped after flattening
(and cause a crash while attempting to drive a constant value).
This issue was introduced in 9f772eb9.
Fixes#2183.
The main part is converting ice40_dsp to recognize the new FF types
created in opt_dff instead of trying to recognize the mux patterns on
its own.
The fsm call has been moved upwards because the passes cannot deal with
$dffe/$sdff*, and other optimizations don't help it much anyway.
The main part is converting xilinx_dsp to recognize the new FF types
created in opt_dff instead of trying to recognize the patterns on its
own.
The fsm call has been moved upwards because the passes cannot deal with
$dffe/$sdff*, and other optimizations don't help it much anyway.
Before this fix, equiv_induct only assumed that one of the following is
true:
- defined value of A is equal to defined value of B
- A is undefined
This lets through valuations where A is defined, B is undefined, and
the defined (meaningless) value of B happens to match the defined value
of A. Instead, tighten this up to OR of the following:
- defined value of A is equal to defined value of B, and B is not
undefined
- A is undefined
This parameter will resolve to the name of the cell being mapped. The
first user of this parameter will be synth_intel_alm's Quartus output,
which requires a unique (and preferably descriptive) name passed as
a cell parameter for the memory cells.
This fixes some dfflegalize equivalence checks, and breaks others — and
I strongly suspect the others are due to bad support for multiple
async inputs in `proc` (in particular, lack of proper support for
dlatchsr and sketchy circuits on dffsr control inputs).
Those can be created by `opt_dff` when optimizing `$adff` with const
clock, or with D == Q. Make dfflegalize do the opposite transform
when such dlatches would be otherwise unimplementable.
This ensures that, when both sync and async FFs are available and abc9
is involved, the sync FFs will be used, and will thus remain available
for sequential synthesis.
I think these were probably missed by accident. Spotted because GCC
spits out lots of messages like this:
passes/techmap/dfflegalize.cc:114:7: warning: zero-length gnu_printf format string [-Wformat-zero-length]
114 | log("");
| ^~
(because we tell GCC that the first argument to log() looks like a
printf control string in log.h, and a zero length such string triggers
a warning).
Rather than assigning specific weights to specific versions of taint tracking logic and summing the weights of all GLIFT cells, sum the following values for each GLIFT cell:
- 0 if the associated hole/$anyconst cell value is non-zero, i.e. reduced-precision taint tracking logic is chosen at this cell
- 1 if the associated hole/$anyconst cell value is zero, i.e. the full-precision taint tracking logic is chosen at this cell
This simplified cost modeling reduces the potential for the QBF-SAT solver to minimize taint tracking logic area but significantly simplifies the QBF-SAT problem.
Look for unique connections in the containing module with the $anyconst port Y SigBit on the RHS and use those. If no such connection is found, fall back to using the name of the $anyconst port Y SigBit.
The new types include:
- FFs with async reset and enable (`$adffe`, `$_DFFE_[NP][NP][01][NP]_`)
- FFs with sync reset (`$sdff`, `$_SDFF_[NP][NP][01]_`)
- FFs with sync reset and enable, reset priority (`$sdffs`, `$_SDFFE_[NP][NP][01][NP]_`)
- FFs with sync reset and enable, enable priority (`$sdffce`, `$_SDFFCE_[NP][NP][01][NP]_`)
- FFs with async reset, set, and enable (`$dffsre`, `$_DFFSRE_[NP][NP][NP][NP]_`)
- latches with reset or set (`$adlatch`, `$_DLATCH_[NP][NP][01]_`)
The new FF types are not actually used anywhere yet (this is left
for future commits).
1. Infinite loop in the optimization procedure when the first solution found while maximizing is at zero.
2. A signed-ness issue when maximizing.
3. Erroneously entering bisection mode with no wire to optimize.
Before this commit, `flatten` matched the template objects with
the newly created objects solely by their name. Because of this,
it could be confused by code such as:
module bar();
$dff a();
endmodule
module foo();
bar b();
$dff \b.a ();
endmodule
After this commit, `flatten` avoids every possible case of name
collision.
Fixes#2106.
log_signal can result in a string with spaces (when bit selection is
involved), which breaks the rule of IdString not containing whitespace.
Instead, remove the sigspec from the name entirely — given that the
resulting wire will have no users, it will be removed later anyway,
so its name doesn't really matter.
Fixes#2118