4. Implementation overview¶
Yosys is an extensible open source hardware synthesis tool. It is aimed at designers who are looking for an easily accessible, universal, and vendor-independent synthesis tool, as well as 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, the openMSP430 CPU, the OpenCores I2C master, and the k68 CPU.
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.
4.1. Simplified data flow¶
Figure 4.1 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 RTLIL 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 RTLIL Backend for writing the RTLIL data in the same format that is understood by the RTLIL Frontend.
With the exception of the AST Frontend, which 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 to represent the design in different stages of the synthesis.
4.2. The RTL Intermediate Language (RTLIL)¶
All frontends, passes and backends in Yosys operate on a design in RTLIL 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 reinventing names for the RTLIL classes, they are simply referred to by their full C++ name, i.e. including the RTLIL:: namespace prefix, in this document.
Figure 4.2 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 which passes operate on, frontends add data to 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.
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:
RTLIL::Cell and RTLIL::Wire objects represent classical netlist data.
RTLIL::Process objects represent the decision trees (if-then-else statements, etc.) and synchronization declarations (clock signals and sensitivity) from Verilog always and VHDL process blocks.
RTLIL::Memory objects represent addressable memories (arrays).
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.
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 rationale behind some of the design decisions.
4.2.1. 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 () or a dollar sign ($).
Identifiers starting with a backslash are public visible identifiers. Usually they originate from one of the HDL input files. For example the signal name “\sig42” is most likely a signal that was declared using the name “sig42” in an HDL input file. On the other hand the signal name “$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:
First, it is impossible that an auto-generated identifier collides with an identifier that was provided by the user.
Second, 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” tries 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.
Third, 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 _integer_.
Whitespace and control characters (any character with an ASCII code 32 or less) are not allowed in RTLIL identifiers; most frontends and backends cannot support these characters in identifiers.
In order to avoid programming errors, the RTLIL data structures check if all identifiers start with either a backslash or a dollar sign, and contain no whitespace or control characters. Violating these rules results in a runtime error.
All RTLIL identifiers are case sensitive.
Some transformations, such as flattening, may have to change identifiers provided by the user to avoid name collisions. When that happens, attribute “hdlname” is attached to the object with the changed identifier. This attribute contains one name (if emitted directly by the frontend, or is a result of disambiguation) or multiple names separated by spaces (if a result of flattening). All names specified in the “hdlname” attribute are public and do not include the leading “".
4.2.2. 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 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. 4.2 an RTLIL::Module object contains the following additional properties:
The module name
A list of attributes
A list of connections between wires
An optional frontend callback used to derive parametrized variations of the module
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 from being generated twice. For modules with only a few parameters, a name directly containing all parameters is generated instead of a hash string.)
4.2.3. 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:
The wire name
A list of attributes
A width (buses are just wires with a width > 1)
Bus direction (MSB to LSB or vice versa)
Lowest valid bit index (LSB or MSB depending on bus direction)
If the wire is a port: port number and direction (input/output/inout)
As with modules, the attributes can be Verilog attributes imported by the Verilog frontend or attributes assigned by passes.
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.
In Verilog and VHDL, busses may have arbitrary bounds, and LSB can have either the lowest or the highest bit index. In RTLIL, bit 0 always corresponds to LSB; however, information from the HDL frontend is preserved so that the bus will be correctly indexed in error messages, backend output, constraint files, etc.
An RTLIL::Cell object has the following properties:
The cell name and type
A list of attributes
A list of parameters (for parametric cells)
Cell ports and the connections of ports to wires and constants
The connections of ports to wires are coded by assigning an RTLIL::SigSpec to each cell port. The RTLIL::SigSpec data type is described in the next section.
A “signal” is everything that can be applied to a cell port. I.e.
- Any constant value of arbitrary bit-width1em For example:
1337, 16'b0000010100111001, 1'b1, 1'bx
- All bits of a wire or a selection of bits from a wire1em For example:
mywire, mywire, mywire[15:8]
- Concatenations of the above1em For example:
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.
When a high-level HDL frontend processes behavioural code it splits it up into data path logic (e.g. the expression a + b is replaced by the output of an adder that takes a and b as inputs) and an RTLIL::Process that models the control logic of the behavioural code. Let’s consider a simple example:
1module ff_with_en_and_async_reset(clock, reset, enable, d, q); 2input clock, reset, enable, d; 3output reg q; 4always @(posedge clock, posedge reset) 5 if (reset) 6 q <= 0; 7 else if (enable) 8 q <= d; 9endmodule
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 RTLIL syntax:
1process $proc$ff_with_en_and_async_reset.v:4$1 2 assign $0\q[0:0] \q 3 switch \reset 4 case 1'1 5 assign $0\q[0:0] 1'0 6 case 7 switch \enable 8 case 1'1 9 assign $0\q[0:0] \d 10 case 11 end 12 end 13 sync posedge \clock 14 update \q $0\q[0:0] 15 sync posedge \reset 16 update \q $0\q[0:0] 17end
This RTLIL::Process contains two RTLIL::SyncRule objects, two RTLIL::SwitchRule objects and five RTLIL::CaseRule objects. The wire $0q[0:0] is an automatically created wire that holds the next value of \q. The lines \(2 \dots 12\) describe how $0q[0:0] should be calculated. The lines \(13 \dots 16\) describe how the value of $0q[0:0] is used to update \q.
An RTLIL::Process is a container for zero or more RTLIL::SyncRule objects and exactly one RTLIL::CaseRule object, which is called the root case.
An RTLIL::SyncRule object contains an (optional) synchronization condition (signal and edge-type), zero or more assignments (RTLIL::SigSig), and zero or more memory writes (RTLIL::MemWriteAction). The always synchronization condition is used to break combinatorial loops when a latch should be inferred instead.
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 $0q[0:0] is first assigned the old value \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 switch statement as it uses a control signal (\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  for \reset == 1 that causes $0q[0:0] to be set (lines 4 and 5) and a default case that in turn contains a switch that sets $0q[0:0] to the value of \d if \enable is active (lines \(6 \dots 11\)).
A case can specify zero or more compare values that will determine whether it matches. Each of the compare values must be the exact same width as the control signal. When more than one compare value is specified, the case matches if any of them matches the control signal; when zero compare values are specified, the case always matches (i.e. it is the default case).
A switch prioritizes cases from first to last: multiple cases can match, but only the first matched case becomes active. This normally synthesizes to a priority encoder. The parallel_case attribute allows passes to assume that no more than one case will match, and full_case attribute allows passes to assume that exactly one case will match; if these invariants are ever dynamically violated, the behavior is undefined. These attributes are useful when an invariant invisible to the synthesizer causes the control signal to never take certain bit patterns.
The lines \(13 \dots 16\) then cause \q to be updated whenever there is a positive clock edge on \clock or \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 proc_arst pass. This pass transforms the above example to the following RTLIL::Process:
1process $proc$ff_with_en_and_async_reset.v:4$1 2 assign $0\q[0:0] \q 3 switch \enable 4 case 1'1 5 assign $0\q[0:0] \d 6 case 7 end 8 sync posedge \clock 9 update \q $0\q[0:0] 10 sync high \reset 11 update \q 1'0 12end
This pass has transformed the outer RTLIL::SwitchRule into a modified RTLIL::SyncRule object for the \reset signal. Further processing converts the RTLIL::Process into e.g. a d-type flip-flop with asynchronous reset and a multiplexer for the enable signal:
1cell $adff $procdff$6 2 parameter \ARST_POLARITY 1'1 3 parameter \ARST_VALUE 1'0 4 parameter \CLK_POLARITY 1'1 5 parameter \WIDTH 1 6 connect \ARST \reset 7 connect \CLK \clock 8 connect \D $0\q[0:0] 9 connect \Q \q 10end 11cell $mux $procmux$3 12 parameter \WIDTH 1 13 connect \A \q 14 connect \B \d 15 connect \S \enable 16 connect \Y $0\q[0:0] 17end
Different combinations of passes may yield different results. Note that $adff and $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 proc pass calls a series of other passes that together perform this conversion in a way that is suitable for most synthesis tasks.
For every array (memory) in the HDL code an RTLIL::Memory object is created. A memory object has the following properties:
The memory name
A list of attributes
The width of an addressable word
The size of the memory in number of words
All read accesses to the memory are transformed to $memrd cells and all write
accesses to $memwr cells by the language frontend. These cells consist of
independent read- and write-ports to the memory. Memory initialization is
transformed to $meminit cells by the language frontend. The
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 $memrd and $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 $memrd and $memwr cells to multiport memory blocks after resource sharing is completed.
The 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 $mem cells).
See Sec. 5.1.5 for details about the memory cell types.
4.3. Command interface and synthesis scripts¶
Yosys reads and processes commands from synthesis scripts, command line arguments and an interactive command prompt. Yosys commands consist of a command name and an optional whitespace separated list of arguments. Commands are terminated using the newline character or a semicolon (;). Empty lines and lines starting with the hash sign (#) are ignored. See Sec. 3.3 for an example synthesis script.
The command help can be used to access the command reference manual.
Most commands can operate not only on the entire design but also specifically on selected parts of the design. For example the command dump will print all selected objects in the current design while dump foobar will only print the module foobar and dump * will print the entire design regardless of the current selection.
dump */t:$add %x:+[A] \*/w:\* %i
The selection mechanism is very powerful. For example the command above will
print all wires that are connected to the
\A port of a
Detailed documentation of the select framework can be found in the command
reference for the
4.4. Source tree and build system¶
The Yosys source tree is organized into the following top-level directories:
- backends/This directory contains a subdirectory for each of the backend modules.
- frontends/This directory contains a subdirectory for each of the frontend modules.
- kernel/This directory contains all the core functionality of Yosys. This includes the functions and definitions for working with the RTLIL data structures (rtlil.h and rtlil.cc), the main() function (driver.cc), the internal framework for generating log messages (log.h and log.cc), the internal framework for registering and calling passes (register.h and register.cc), some core commands that are not really passes (select.cc, show.cc, …) and a couple of other small utility libraries.
- passes/This directory contains a subdirectory for each pass or group of passes. For example as of this writing the directory passes/opt/ contains the code for seven passes: opt, opt_expr, opt_muxtree, opt_reduce, opt_rmdff, opt_rmunused and opt_merge.
- techlibs/This directory contains simulation models and standard implementations for the cells from the internal cell library.
- tests/This directory contains a couple of test cases. Most of the smaller tests are executed automatically when 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.
The top-level Makefile includes frontends/*/Makefile.inc, passes/*/Makefile.inc and backends/*/Makefile.inc. So when extending Yosys it is enough to create a new directory in frontends/, passes/ or backends/ with your sources and a 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.
Good starting points for reading example source code to learn how to write passes are passes/opt/opt_rmdff.cc and passes/opt/opt_merge.cc.
See the top-level README file for a quick Getting Started guide and build instructions. The Yosys build is based solely on Makefiles.
Users of the Qt Creator IDE can generate a QT Creator project file using 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 in order to programming extensions to Yosys or just browse the Yosys code base.