\chapter{Implementation Overview} \label{chapter:overview} Yosys is an extensible open source hardware synthesis tool. It is aimed at designers who are looking for an easy accessible, universal, and vendor independent synthesis tool, and scientists who do research in electronic design automation (EDA) and are looking for an open synthesis framework that can be used to test algorithms on complex real-world designs. Yosys can synthesize a large subset of Verilog 2005 and has been tested with a wide range of real-world designs, including the OpenRISC 1200 CPU \citeweblink{OR1200}, the openMSP430 CPU \citeweblink{openMSP430}, the OpenCores I$^2$C master \citeweblink{i2cmaster} and the k68 CPU \citeweblink{k68}. As of this writing a Yosys VHDL frontend is in development. Yosys is written in C++ (using some features from the new C++11 standard). This chapter describes some of the fundamental Yosys data structures. For the sake of simplicity the C++ type names used in the Yosys implementation are used in this chapter, even though the chapter only explains the conceptual idea behind it and can be used as reference to implement a similar system in any language. \section{Simplified Data Flow} Figure~\ref{fig:Overview_flow} shows the simplified data flow within Yosys. Rectangles in the figure represent program modules and ellipses internal data structures that are used to exchange design data between the program modules. Design data is read in using one of the frontend modules. The high-level HDL frontends for Verilog and VHDL code generate an abstract syntax tree (AST) that is then passed to the AST frontend. Note that both HDL frontends use the same AST representation that is powerful enough to cover the Verilog HDL and VHDL language. The AST Frontend then compiles the AST to Yosys's main internal data format, the RTL Intermediate Language (RTLIL). A more detailed description of this format is given in the next section. There is also a text representation of the RTLIL data structure that can be parsed using the ILANG Frontend. The design data may then be transformed using a series of passes that all operate on the RTLIL representation of the design. Finally the design in RTLIL representation is converted back to text by one of the backends, namely the Verilog Backend for generating Verilog netlists and the ILANG Backend for writing the RTLIL data in the same format that is understood by the ILANG Frontend. With the exception of the AST Frontend, that is called by the high-level HDL frontends and can't be called directly by the user, all program modules are called by the user (usually using a synthesis script that contains text commands for Yosys). By combining passes in different ways and/or adding additional passes to Yosys it is possible to adapt Yosys to a wide range of applications. For this to be possible it is key that (1) all passes operate on the same data structure (RTLIL) and (2) that this data structure is powerful enough represent the design in different stages of the synthesis. \begin{figure}[t] \hfil \begin{tikzpicture} \tikzstyle{process} = [draw, fill=green!10, rectangle, minimum height=3em, minimum width=10em, node distance=15em] \tikzstyle{data} = [draw, fill=blue!10, ellipse, minimum height=3em, minimum width=7em, node distance=15em] \node[process] (vlog) {Verilog Frontend}; \node[process, dashed, fill=green!5] (vhdl) [right of=vlog] {VHDL Frontend}; \node[process] (ilang) [right of=vhdl] {ILANG Frontend}; \node[data] (ast) [below of=vlog, node distance=5em, xshift=7.5em] {AST}; \node[process] (astfe) [below of=ast, node distance=5em] {AST Frontend}; \node[data] (rtlil) [below of=astfe, node distance=5em, xshift=7.5em] {RTLIL}; \node[process] (pass) [right of=rtlil, node distance=5em, xshift=7.5em] {Passes}; \node[process] (vlbe) [below of=rtlil, node distance=7em, xshift=-13em] {Verilog Backend}; \node[process] (ilangbe) [below of=rtlil, node distance=7em, xshift=0em] {ILANG Backend}; \node[process, dashed, fill=green!5] (otherbe) [below of=rtlil, node distance=7em, xshift=+13em] {Other Backends}; \draw[-latex] (vlog) -- (ast); \draw[-latex] (vhdl) -- (ast); \draw[-latex] (ast) -- (astfe); \draw[-latex] (astfe) -- (rtlil); \draw[-latex] (ilang) -- (rtlil); \draw[latex-latex] (rtlil) -- (pass); \draw[-latex] (rtlil) -- (vlbe); \draw[-latex] (rtlil) -- (ilangbe); \draw[-latex] (rtlil) -- (otherbe); \end{tikzpicture} \caption{Yosys simplified data flow (ellipses: data structures, rectangles: program modules)} \label{fig:Overview_flow} \end{figure} \section{The RTL Intermediate Language} All frontends, passes and backends in Yosys operate on a design in RTLIL\footnote{The {\it Language} in {\it RTL Intermediate Language} refers to the fact, that RTLIL also has a text representation, usually referred to as {\it Intermediate Language} (ILANG).} representation. The only exception are the high-level frontends that use the AST representation as an intermediate step before generating RTLIL data. In order to avoid re-inventing names for the RTLIL classes, they are simply referred to by their full C++ name, i.e.~including the {\tt RTLIL::} namespace prefix, in this document. Figure~\ref{fig:Overview_RTLIL} shows a simplified Entity-Relationship Diagram (ER Diagram) of RTLIL. In $1:N$ relationships the arrow points from the $N$ side to the $1$. For example one RTLIL::Design contains $N$ (zero to many) instances of RTLIL::Module. A two-pointed arrow indicates a $1:1$ relationship. The RTLIL::Design is the root object of the RTLIL data structure. There is always one ``current design'' in memory on which passes operate, frontends add data to it and backends convert to exportable formats. But in some cases passes internally generate additional RTLIL::Design objects. For example when a pass is reading an auxiliary Verilog file such as a cell library, it might create an additional RTLIL::Design object and call the Verilog frontend with this other object to parse the cell library. \begin{figure}[t] \hfil \begin{tikzpicture} \tikzstyle{entity} = [draw, fill=gray!10, rectangle, minimum height=3em, minimum width=7em, node distance=5em, font={\ttfamily}] \node[entity] (design) {RTLIL::Design}; \node[entity] (module) [right of=design, node distance=11em] {RTLIL::Module} edge [-latex] node[above] {\tiny 1 \hskip3em N} (design); \node[entity] (process) [fill=green!10, right of=module, node distance=10em] {RTLIL::Process} (process.west) edge [-latex] (module); \node[entity] (memory) [fill=red!10, below of=process] {RTLIL::Memory} edge [-latex] (module); \node[entity] (wire) [fill=blue!10, above of=process] {RTLIL::Wire} (wire.west) edge [-latex] (module); \node[entity] (cell) [fill=blue!10, above of=wire] {RTLIL::Cell} (cell.west) edge [-latex] (module); \node[entity] (case) [fill=green!10, right of=process, node distance=10em] {RTLIL::CaseRule} edge [latex-latex] (process); \node[entity] (sync) [fill=green!10, above of=case] {RTLIL::SyncRule} edge [-latex] (process); \node[entity] (switch) [fill=green!10, below of=case] {RTLIL::SwitchRule} edge [-latex] (case); \draw[latex-] (switch.east) -- ++(1em,0) |- (case.east); \end{tikzpicture} \caption{Simplified RTLIL Entity-Relationship Diagram} \label{fig:Overview_RTLIL} \end{figure} There is only one active RTLIL::Design object that is used by all frontends, passes and backends called by the user, e.g.~using a synthesis script. The RTLIL::Design then contains zero to many RTLIL::Module objects. This corresponds to modules in Verilog or entities in VHDL. Each module in turn contains objects from three different categories: \begin{itemize} \item RTLIL::Cell and RTLIL::Wire objects represent classical netlist data. \item RTLIL::Process objects represent the decision trees (if-then-else statements, etc.) and synchronization declarations (clock signals and sensitivity) from Verilog {\tt always} and VHDL {\tt process} blocks. \item RTLIL::Memory objects represent addressable memories (arrays). \end{itemize} \begin{sloppypar} Usually the output of the synthesis procedure is a netlist, i.e. all RTLIL::Process and RTLIL::Memory objects must be replaced by RTLIL::Cell and RTLIL::Wire objects by synthesis passes. \end{sloppypar} All features of the HDL that cannot be mapped directly to these RTLIL classes must be transformed to an RTLIL-compatible representation by the HDL frontend. This includes Verilog-features such as generate-blocks, loops and parameters. The following sections contain a more detailed description of the different parts of RTLIL and rationales behind some of the design decisions. \subsection{RTLIL Identifiers} All identifiers in RTLIL (such as module names, port names, signal names, cell types, etc.) follow the following naming convention: They must either start with a backslash (\textbackslash) or a dollar sign (\$). Identifiers starting with a backslash are public visible identifiers. Usually they originate from one of the HDL input files. For example the signal name ``{\tt \textbackslash sig42}'' is most likely a signal that was declared using the name ``{\tt sig42}'' in an HDL input file. On the other hand the signal name ``{\tt \$sig42}'' is an auto-generated signal name. The backends convert all identifiers that start with a dollar sign to identifiers that do not collide with identifiers that start with a backslash. This has three advantages: \begin{itemize} \item Firstly it is impossible that an auto-generated identifier collides with an identifier that was provided by the user. \item Secondly the information about which identifiers were originally provided by the user is always available which can help guide some optimizations. For example the ``opt\_rmunused'' is trying to preserve signals with a user-provided name but doesn't hesitate to delete signals that have auto-generated names when they just duplicate other signals. \item Thirdly the delicate job of finding suitable auto-generated public visible names is deferred to one central location. Internally auto-generated names that may hold important information for Yosys developers can be used without disturbing external tools. For example the Verilog backend assigns names in the form {\tt \_{\it integer}\_}. \end{itemize} In order to avoid programming errors, the RTLIL data structures check if all identifiers start with either a backslash or a dollar sign and generate a runtime error if this rule is violated. All RTLIL identifiers are case sensitive. \subsection{RTLIL::Design and RTLIL::Module} The RTLIL::Design object is basically just a container for RTLIL::Module objects. In addition to a list of RTLIL::Module objects the RTLIL::Design also keeps a list of {\it selected objects}, i.e. the objects that passes should operate on. In most cases the whole design is selected and therefore passes operate on the whole design. But this mechanism can be useful for more complex synthesis jobs in which only parts of the design should be affected by certain passes. Besides the objects shown in the ER diagram in Fig.~\ref{fig:Overview_RTLIL} an RTLIL::Module object contains the following additional properties: \begin{itemize} \item The module name \item A list of attributes \item A list of connections between wires \item An optional frontend callback used to derive parametrized variations of the module \end{itemize} The attributes can be Verilog attributes imported by the Verilog frontend or attributes assigned by passes. They can be used to store additional metadata about modules or just mark them to be used by certain part of the synthesis script but not by others. Verilog and VHDL both support parametric modules (known as ``generic entities'' in VHDL). The RTLIL format does not support parametric modules itself. Instead each module contains a callback function into the AST frontend to generate a parametrized variation of the RTLIL::Module as needed. This callback then returns the auto-generated name of the parametrized variation of the module. (A hash over the parameters and the module name is used to prohibit the same parametrized variation to be generated twice. For modules with only a few parameters, a name directly containing all parameters is generated instead of a hash string.) \subsection{RTLIL::Cell and RTLIL::Wire} A module contains zero to many RTLIL::Cell and RTLIL::Wire objects. Objects of these types are used to model netlists. Usually the goal of all synthesis efforts is to convert all modules to a state where the functionality of the module is implemented only by cells from a given cell library and wires to connect these cells with each other. Note that module ports are just wires with a special property. An RTLIL::Wire object has the following properties: \begin{itemize} \item The wire name \item A list of attributes \item A width (busses are just wires with a width > 1) \item If the wire is a port: port number and direction (input/output/inout) \end{itemize} As with modules, the attributes can be Verilog attributes imported by the Verilog frontend or attributes assigned by passees. In Yosys, busses (signal vectors) are represented using a single wire object with a width > 1. So Yosys does not convert signal vectors to individual signals. This makes some aspects of RTLIL more complex but enables Yosys to be used for coarse grain synthesis where the cells of the target architecture operate on entire signal vectors instead of single bit wires. An RTLIL::Cell object has the following properties: \begin{itemize} \item The cell name and type \item A list of attributes \item A list of parameters (for parametric cells) \item Cell ports and the connections of ports to wires and constants \end{itemize} The connections of ports to wires are coded by assigning an RTLIL::SigSpec to each cell ports. The RTLIL::SigSpec data type is described in the next section. \subsection{RTLIL::SigSpec} A ``signal'' is everything that can be applied to a cell port. I.e. \begin{itemize} \item Any constant value of arbitrary bit-width \\ \null\hskip1em For example: \lstinline[language=Verilog]{1337, 16'b0000010100111001, 1'b1, 1'bx} \item All bits of a wire or a selection of bits from a wire \\ \null\hskip1em For example: \lstinline[language=Verilog]{mywire, mywire[24], mywire[15:8]} \item Concatenations of the above \\ \null\hskip1em For example: \lstinline[language=Verilog]|{16'd1337, mywire[15:8]}| \end{itemize} The RTLIL::SigSpec data type is used to represent signals. The RTLIL::Cell object contains one RTLIL::SigSpec for each cell port. In addition, connections between wires are represented using a pair of RTLIL::SigSpec objects. Such pairs are needed in different locations. Therefore the type name RTLIL::SigSig was defined for such a pair. \subsection{RTLIL::Process} When a high-level HDL frontend processes behavioural code it splits it up into data path logic (e.g.~the expression {\tt a + b} is replaced by the output of an adder that takes {\tt a} and {\tt b} as inputs) and an RTLIL::Process that models the control logic of the behavioural code. Let's consider a simple example: \begin{lstlisting}[numbers=left,frame=single,language=Verilog] module ff_with_en_and_async_reset(clock, reset, enable, d, q); input clock, reset, enable, d; output reg q; always @(posedge clock, posedge reset) if (reset) q <= 0; else if (enable) q <= d; endmodule \end{lstlisting} In this example there is no data path and therefore the RTLIL::Module generated by the frontend only contains a few RTLIL::Wire objects and an RTLIL::Process. The RTLIL::Process in ILANG syntax: \begin{lstlisting}[numbers=left,frame=single,language=rtlil] process $proc$ff_with_en_and_async_reset.v:4$1 assign $0\q[0:0] \q switch \reset case 1'1 assign $0\q[0:0] 1'0 case switch \enable case 1'1 assign $0\q[0:0] \d case end end sync posedge \clock update \q $0\q[0:0] sync posedge \reset update \q $0\q[0:0] end \end{lstlisting} This RTLIL::Process contains two RTLIL::SyncRule objects, two RTLIL::SwitchRule objects and five RTLIL::CaseRule objects. The wire {\tt \$0\textbackslash{}q[0:0]} is an automatically created wire that holds the next value of {\tt \textbackslash{}q}. The lines $2 \dots 12$ describe how {\tt \$0\textbackslash{}q[0:0]} should be calculated. The lines $13 \dots 16$ describe how the value of {\tt \$0\textbackslash{}q[0:0]} is used to update {\tt \textbackslash{}q}. An RTLIL::Process is a container for zero or more RTLIL::SyncRule objects and exactly one RTLIL::CaseRule object, which is called the {\it root case}. An RTLIL::SyncRule object contains an (optional) synchronization condition (signal and edge-type) and zero or more assignments (RTLIL::SigSig). An RTLIL::CaseRule is a container for zero or more assignments (RTLIL::SigSig) and zero or more RTLIL::SwitchRule objects. An RTLIL::SwitchRule objects is a container for zero or more RTLIL::CaseRule objects. In the above example the lines $2 \dots 12$ are the root case. Here {\tt \$0\textbackslash{}q[0:0]} is first assigned the old value {\tt \textbackslash{}q} as default value (line 2). The root case also contains an RTLIL::SwitchRule object (lines $3 \dots 12$). Such an object is very similar to the C {\tt switch} statement as it uses a control signal ({\tt \textbackslash{}reset} in this case) to determine which of its cases should be active. The RTLIL::SwitchRule object then contains one RTLIL::CaseRule object per case. In this example there is a case\footnote{The syntax {\tt 1'1} in the ILANG code specifies a constant with a length of one bit (the first ``1''), and this bit is a one (the second ``1'').} for {\tt \textbackslash{}reset == 1} that causes {\tt \$0\textbackslash{}q[0:0]} to be set (lines 4 and 5) and a default case that in turn contains a switch that sets {\tt \$0\textbackslash{}q[0:0]} to the value of {\tt \textbackslash{}d} if {\tt \textbackslash{}enable} is active (lines $6 \dots 11$). The lines $13 \dots 16$ then cause {\tt \textbackslash{}q} to be updated whenever there is a positive clock edge on {\tt \textbackslash{}clock} or {\tt \textbackslash{}reset}. In order to generate such a representation, the language frontend must be able to handle blocking and nonblocking assignments correctly. However, the language frontend does not need to identify the correct type of storage element for the output signal or generate multiplexers for the decision tree. This is done by passes that work on the RTLIL representation. Therefore it is relatively easy to substitute these steps with other algorithms that target different target architectures or perform optimizations or other transformations on the decision trees before further processing them. One of the first actions performed on a design in RTLIL representation in most synthesis scripts is identifying asynchronous resets. This is usually done using the {\tt proc\_arst} pass. This pass transforms the above example to the following RTLIL::Process: \begin{lstlisting}[numbers=left,frame=single,language=rtlil] process $proc$ff_with_en_and_async_reset.v:4$1 assign $0\q[0:0] \q switch \enable case 1'1 assign $0\q[0:0] \d case end sync posedge \clock update \q $0\q[0:0] sync high \reset update \q 1'0 end \end{lstlisting} This pass has transformed the outer RTLIL::SwitchRule into a modified RTLIL::SyncRule object for the {\tt \textbackslash{}reset} signal. Further processing converts the RTLIL::Process e.g.~into a d-type flip-flop with asynchronous reset and a multiplexer for the enable signal: \begin{lstlisting}[numbers=left,frame=single,language=rtlil] cell $adff $procdff$6 parameter \ARST_POLARITY 1'1 parameter \ARST_VALUE 1'0 parameter \CLK_POLARITY 1'1 parameter \WIDTH 1 connect \ARST \reset connect \CLK \clock connect \D $0\q[0:0] connect \Q \q end cell $mux $procmux$3 parameter \WIDTH 1 connect \A \q connect \B \d connect \S \enable connect \Y $0\q[0:0] end \end{lstlisting} Different combinations of passes may yield different results. Note that {\tt \$adff} and {\tt \$mux} are internal cell types that still need to be mapped to cell types from the target cell library. Some passes refuse to operate on modules that still contain RTLIL::Process objects as the presence of these objects in a module increases the complexity. Therefore the passes to translate processes to a netlist of cells are usually called early in a synthesis script. The {\tt proc} pass calls a series of other passes that together perform this conversion in a way that is suitable for most synthesis taks. \subsection{RTLIL::Memory} For every array (memory) in the HDL code an RTLIL::Memory object is created. A memory object has the following properties: \begin{itemize} \item The memory name \item A list of attributes \item The width of an addressable word \item The size of the memory in number of words \end{itemize} All read accesses to the memory are transformed to {\tt \$memrd} cells and all write accesses to {\tt \$memwr} cells by the language frontend. These cells consist of independent read- and write-ports to the memory. The \B{MEMID} parameter on these cells is used to link them together and to the RTLIL::Memory object they belong to. The rationale behind using separate cells for the individual ports versus creating a large multiport memory cell right in the language frontend is that the separate {\tt \$memrd} and {\tt \$memwr} cells can be consolidated using resource sharing. As resource sharing is a non-trivial optimization problem where different synthesis tasks can have different requirements it lends itself to do the optimisation in separate passes and merge the RTLIL::Memory objects and {\tt \$memrd} and {\tt \$memwr} cells to multiport memory blocks after resource sharing is completed. The {\tt memory} pass performs this conversion and can (depending on the options passed to it) transform the memories directly to d-type flip-flops and address logic or yield multiport memory blocks (represented using {\tt \$mem} cells). See Sec.~\ref{sec:memcells} for details on the memory cell types. \section{Command Interface and Synthesis Scripts} Yosys reads and processes commands from synthesis scripts, command line arguments and an interactive command prompt. Yosys commands consist of a command name and an optional whitespace sparated list of arguments. Commands are terminated using the newline character or a semicolon ({\tt ;}). Empty lines and lines starting with the hash sign ({\tt \#}) are ignored. See Sec.~\ref{sec:typusecase} for an example synthesis script. The command {\tt help} can be used to access the command reference manual. Most commands can operate not only on the entire design but also only on {\it selected} parts of the design. For example the command {\tt dump} will print all selected objects in the current design while {\tt dump foobar} will only print the module {\tt foobar} and {\tt dump *} will print the entire design regardless of the current selection. The selection mechanism is very powerful. For example the command {\tt dump */t:\$add \%x:+[A] */w:* \%i} will print all wires that are connected to the \B{A} port of a {\tt \$add} cell. A detailed documentation of the select framework can be found in the command reference for the {\tt select} command. \section{Source Tree and Build System} The Yosys source tree is organized in the following top-level directories: \begin{itemize} \item {\tt backends/} \\ This directory contains a subdirectory for each of the backend modules. \item {\tt frontends/} \\ This directory contains a subdirectory for each of the frontend modules. \item {\tt kernel/} \\ This directory contains all the core functionality of Yosys. This includes the functions and definitions for working with the RTLIL data structures ({\tt rtlil.h} and {\tt rtlil.cc}), the main() function ({\tt driver.cc}), the internal framework for generating log messages ({\tt log.h} and {\tt log.cc}), the internal framework for registering and calling passes ({\tt register.h} and {\tt register.cc}), some core commands that are not really passes ({\tt select.cc}, {\tt show.cc}, \dots) and a couple of other small utility libraries. \item {\tt passes/} \\ This directory contains a subdirectory for each pass or group of passes. For example as of this writing the directory {\tt passes/opt/} contains the code for seven passes: {\tt opt}, {\tt opt\_const}, {\tt opt\_muxtree}, {\tt opt\_reduce}, {\tt opt\_rmdff}, {\tt opt\_rmunused} and {\tt opt\_share}. \item {\tt techlibs/} \\ This directory contains simulation models and standard implementations for the cells from the internal cell library. \item {\tt tests/} \\ This directory contains a couple of test cases. Most of the smaller tests are executed automatically when {\tt make test} is called. The larger tests must be executed manually. Most of the larger tests require downloading external HDL source code and/or external tools. The tests range from comparing simulation results of the synthesized design to the original sources to logic equivalence checking of entire CPU cores. \end{itemize} \begin{sloppypar} The top-level Makefile includes {\tt frontends/*/Makefile.inc}, {\tt passes/*/Makefile.inc} and {\tt backends/*/Makefile.inc}. So when extending Yosys it is enough to create a new directory in {\tt frontends/}, {\tt passes/} or {\tt backends/} with your sources and a {\tt Makefile.inc}. The Yosys kernel automatically detects all commands linked with Yosys. So it is not needed to add additional commands to a central list of commands. \end{sloppypar} A good starting point for reading example source code for learning how to write passes are {\tt passes/opt/opt\_rmdff.cc} and {\tt passes/opt/opt\_share.cc}. See the top-level README file for a quick {\it Getting Started} guide and build instructions. Yosys is a pure Makefile based project. Users of the Qt Creator IDE can generate a QT Creator project file using {\tt make qtcreator}. Users of the Eclipse IDE can use the ``Makefile Project with Existing Code'' project type in the Eclipse ``New Project'' dialog (only available after the CDT plugin has been installed) to create an Eclipse Project for programming extensions to Yosys or just browsing the Yosys code base.