Merge pull request #3310 from robinsonb5-PRs/master

[yosys.git] / manual / CHAPTER_Optimize.tex

\chapter{Optimizations}
\label{chapter:opt}

Yosys employs a number of optimizations to generate better and cleaner results.
This chapter outlines these optimizations.

\section{Simple Optimizations}

The Yosys pass {\tt opt} runs a number of simple optimizations. This includes removing unused
signals and cells and const folding. It is recommended to run this pass after each major step
in the synthesis script. At the time of this writing the {\tt opt} pass executes the following
passes that each perform a simple optimization:

\begin{itemize}
\item Once at the beginning of {\tt opt}:
\begin{itemize}
\item {\tt opt\_expr}
\item {\tt opt\_merge -nomux}
\end{itemize}
\item Repeat until result is stable:
\begin{itemize}
\item {\tt opt\_muxtree}
\item {\tt opt\_reduce}
\item {\tt opt\_merge}
\item {\tt opt\_rmdff}
\item {\tt opt\_clean}
\item {\tt opt\_expr}
\end{itemize}
\end{itemize}

The following section describes each of the {\tt opt\_*} passes.

\subsection{The opt\_expr pass}

This pass performs const folding on the internal combinational cell types
described in Chap.~\ref{chapter:celllib}. This means a cell with all constant
inputs is replaced with the constant value this cell drives. In some cases
this pass can also optimize cells with some constant inputs.

\begin{table}
        \hfil
        \begin{tabular}{cc|c}
                A-Input & B-Input & Replacement \\
                \hline
                any &   0 &   0 \\
                  0 & any &   0 \\
                  1 &   1 &   1 \\
                \hline
                X/Z & X/Z &   X \\
                  1 & X/Z &   X \\
                X/Z &   1 &   X \\
                \hline
                any & X/Z &   0 \\
                X/Z & any &   0 \\
                \hline
                $a$ &   1 & $a$ \\
                  1 & $b$ & $b$ \\
        \end{tabular}
        \caption{Const folding rules for {\tt\$\_AND\_} cells as used in {\tt opt\_expr}.}
        \label{tab:opt_expr_and}
\end{table}

Table~\ref{tab:opt_expr_and} shows the replacement rules used for optimizing
an {\tt\$\_AND\_} gate. The first three rules implement the obvious const folding
rules. Note that `any' might include dynamic values calculated by other parts
of the circuit. The following three lines propagate undef (X) states.
These are the only three cases in which it is allowed to propagate an undef
according to Sec.~5.1.10 of IEEE Std. 1364-2005 \cite{Verilog2005}.

The next two lines assume the value 0 for undef states. These two rules are only
used if no other substitutions are possible in the current module. If other substitutions
are possible they are performed first, in the hope that the `any' will change to
an undef value or a 1 and therefore the output can be set to undef.

The last two lines simply replace an {\tt\$\_AND\_} gate with one constant-1
input with a buffer.

Besides this basic const folding the {\tt opt\_expr} pass can replace 1-bit wide
{\tt \$eq} and {\tt \$ne} cells with buffers or not-gates if one input is constant.

The {\tt opt\_expr} pass is very conservative regarding optimizing {\tt \$mux} cells,
as these cells are often used to model decision-trees and breaking these trees can
interfere with other optimizations.

\subsection{The opt\_muxtree pass}

This pass optimizes trees of multiplexer cells by analyzing the select inputs.
Consider the following simple example:

\begin{lstlisting}[numbers=left,frame=single,language=Verilog]
module uut(a, y);
input a;
output [1:0] y = a ? (a ? 1 : 2) : 3;
endmodule
\end{lstlisting}

The output can never be 2, as this would require \lstinline[language=Verilog];a;
to be 1 for the outer multiplexer and 0 for the inner multiplexer. The {\tt
opt\_muxtree} pass detects this contradiction and replaces the inner multiplexer
with a constant 1, yielding the logic for \lstinline[language=Verilog];y = a ? 1 : 3;.

\subsection{The opt\_reduce pass}

\begin{sloppypar}
This is a simple optimization pass that identifies and consolidates identical input
bits to {\tt \$reduce\_and} and {\tt \$reduce\_or} cells. It also sorts the input
bits to ease identification of shareable {\tt \$reduce\_and} and {\tt \$reduce\_or} cells
in other passes.
\end{sloppypar}

This pass also identifies and consolidates identical inputs to multiplexer cells. In this
case the new shared select bit is driven using a {\tt \$reduce\_or} cell that combines
the original select bits.

Lastly this pass consolidates trees of {\tt \$reduce\_and} cells and trees of
{\tt \$reduce\_or} cells to single large {\tt \$reduce\_and} or {\tt \$reduce\_or} cells.

These three simple optimizations are performed in a loop until a stable result is
produced.

\subsection{The opt\_rmdff pass}

This pass identifies single-bit d-type flip-flops ({\tt \$\_DFF\_*}, {\tt \$dff}, and {\tt
\$adff} cells) with a constant data input and replaces them with a constant driver.

\subsection{The opt\_clean pass}

This pass identifies unused signals and cells and removes them from the design. It also
creates an \B{unused\_bits} attribute on wires with unused bits. This attribute can be
used for debugging or by other optimization passes.

\subsection{The opt\_merge pass}

This pass performs trivial resource sharing. This means that this pass identifies cells
with identical inputs and replaces them with a single instance of the cell.

The option {\tt -nomux} can be used to disable resource sharing for multiplexer
cells ({\tt \$mux} and {\tt \$pmux}. This can be useful as
it prevents multiplexer trees to be merged, which might prevent {\tt opt\_muxtree}
to identify possible optimizations.

\section{FSM Extraction and Encoding}

The {\tt fsm} pass performs finite-state-machine (FSM) extraction and recoding. The {\tt fsm}
pass simply executes the following other passes:

\begin{itemize}
\item Identify and extract FSMs:
\begin{itemize}
\item {\tt fsm\_detect}
\item {\tt fsm\_extract}
\end{itemize}

\item Basic optimizations:
\begin{itemize}
\item {\tt fsm\_opt}
\item {\tt opt\_clean}
\item {\tt fsm\_opt}
\end{itemize}

\item Expanding to nearby gate-logic (if called with {\tt -expand}):
\begin{itemize}
\item {\tt fsm\_expand}
\item {\tt opt\_clean}
\item {\tt fsm\_opt}
\end{itemize}

\item Re-code FSM states (unless called with {\tt -norecode}):
\begin{itemize}
\item {\tt fsm\_recode}
\end{itemize}

\item Print information about FSMs:
\begin{itemize}
\item {\tt fsm\_info}
\end{itemize}

\item Export FSMs in KISS2 file format (if called with {\tt -export}):
\begin{itemize}
\item {\tt fsm\_export}
\end{itemize}

\item Map FSMs to RTL cells (unless called with {\tt -nomap}):
\begin{itemize}
\item {\tt fsm\_map}
\end{itemize}
\end{itemize}

The {\tt fsm\_detect} pass identifies FSM state registers and marks them using the
\B{fsm\_encoding}{\tt = "auto"} attribute. The {\tt fsm\_extract} extracts all
FSMs marked using the \B{fsm\_encoding} attribute (unless \B{fsm\_encoding} is
set to {\tt "none"}) and replaces the corresponding RTL cells with a {\tt \$fsm}
cell. All other {\tt fsm\_*} passes operate on these {\tt \$fsm} cells. The
{\tt fsm\_map} call finally replaces the {\tt \$fsm} cells with RTL cells.

Note that these optimizations operate on an RTL netlist. I.e.~the {\tt fsm} pass
should be executed after the {\tt proc} pass has transformed all
{\tt RTLIL::Process} objects to RTL cells.

The algorithms used for FSM detection and extraction are influenced by a more
general reported technique \cite{fsmextract}.

\subsection{FSM Detection}

The {\tt fsm\_detect} pass identifies FSM state registers. It sets the
\B{fsm\_encoding}{\tt = "auto"} attribute on any (multi-bit) wire that matches
the following description:

\begin{itemize}
\item Does not already have the \B{fsm\_encoding} attribute.
\item Is not an output of the containing module.
\item Is driven by single {\tt \$dff} or {\tt \$adff} cell.
\item The \B{D}-Input of this {\tt \$dff} or {\tt \$adff} cell is driven by a multiplexer
tree that only has constants or the old state value on its leaves.
\item The state value is only used in the said multiplexer tree or by simple relational
cells that compare the state value to a constant (usually {\tt \$eq} cells).
\end{itemize}

This heuristic has proven to work very well. It is possible to overwrite it by setting
\B{fsm\_encoding}{\tt = "auto"} on registers that should be considered FSM state registers
and setting \B{fsm\_encoding}{\tt = "none"} on registers that match the above criteria
but should not be considered FSM state registers.

Note however that marking state registers with \B{fsm\_encoding} that are not
suitable for FSM recoding can cause synthesis to fail or produce invalid
results.

\subsection{FSM Extraction}

The {\tt fsm\_extract} pass operates on all state signals marked with the
\B{fsm\_encoding} ({\tt != "none"}) attribute. For each state signal the following
information is determined:

\begin{itemize}
\item The state registers
\item The asynchronous reset state if the state registers use asynchronous reset
\item All states and the control input signals used in the state transition functions
\item The control output signals calculated from the state signals and control inputs
\item A table of all state transitions and corresponding control inputs- and outputs
\end{itemize}

The state registers (and asynchronous reset state, if applicable) is simply determined
by identifying the driver for the state signal.

From there the {\tt \$mux}-tree driving the state register inputs is
recursively traversed. All select inputs are control signals and the leaves of the
{\tt \$mux}-tree are the states. The algorithm fails if a non-constant leaf
that is not the state signal itself is found.

The list of control outputs is initialized with the bits from the state signal.
It is then extended by adding all values that are calculated by cells that
compare the state signal with a constant value.

In most cases this will cover all uses of the state register, thus rendering the
state encoding arbitrary. If however a design uses e.g.~a single bit of the state
value to drive a control output directly, this bit of the state signal will be
transformed to a control output of the same value.

Finally, a transition table for the FSM is generated. This is done by using the
{\tt ConstEval} C++ helper class (defined in {\tt kernel/consteval.h}) that can
be used to evaluate parts of the design. The {\tt ConstEval} class can be asked
to calculate a given set of result signals using a set of signal-value
assignments. It can also be passed a list of stop-signals that abort the {\tt
ConstEval} algorithm if the value of a stop-signal is needed in order to
calculate the result signals.

The {\tt fsm\_extract} pass uses the {\tt ConstEval} class in the following way
to create a transition table. For each state:

\begin{enumerate}
\item Create a {\tt ConstEval} object for the module containing the FSM
\item Add all control inputs to the list of stop signals
\item Set the state signal to the current state
\item Try to evaluate the next state and control output \label{enum:fsm_extract_cealg_try}
\item If step~\ref{enum:fsm_extract_cealg_try} was not successful:
\begin{itemize}
\item Recursively goto step~\ref{enum:fsm_extract_cealg_try} with the offending stop-signal set to 0.
\item Recursively goto step~\ref{enum:fsm_extract_cealg_try} with the offending stop-signal set to 1.
\end{itemize}
\item If step~\ref{enum:fsm_extract_cealg_try} was successful: Emit transition
\end{enumerate}

Finally a {\tt \$fsm} cell is created with the generated transition table and added to the
module. This new cell is connected to the control signals and the old drivers for the
control outputs are disconnected.

\subsection{FSM Optimization}

The {\tt fsm\_opt} pass performs basic optimizations on {\tt \$fsm} cells (not including state
recoding). The following optimizations are performed (in this order):

\begin{itemize}
\item Unused control outputs are removed from the {\tt \$fsm} cell. The attribute \B{unused\_bits}
(that is usually set by the {\tt opt\_clean} pass) is used to determine which control
outputs are unused.
\item Control inputs that are connected to the same driver are merged.
\item When a control input is driven by a control output, the control input is removed and the transition
table altered to give the same performance without the external feedback path.
\item Entries in the transition table that yield the same output and only
differ in the value of a single control input bit are merged and the different bit is removed
from the sensitivity list (turned into a don't-care bit).
\item Constant inputs are removed and the transition table is altered to give an unchanged behaviour.
\item Unused inputs are removed.
\end{itemize}

\subsection{FSM Recoding}

The {\tt fsm\_recode} pass assigns new bit pattern to the states. Usually this
also implies a change in the width of the state signal. At the moment of this
writing only one-hot encoding with all-zero for the reset state is supported.

The {\tt fsm\_recode} pass can also write a text file with the changes performed
by it that can be used when verifying designs synthesized by Yosys using Synopsys
Formality \citeweblink{Formality}.

\section{Logic Optimization}

Yosys can perform multi-level combinational logic optimization on gate-level netlists using the
external program ABC \citeweblink{ABC}. The {\tt abc} pass extracts the combinational gate-level
parts of the design, passes it through ABC, and re-integrates the results. The {\tt abc} pass
can also be used to perform other operations using ABC, such as technology mapping (see
Sec.~\ref{sec:techmap_extern} for details).