The Verilog and AST Frontends¶

This chapter provides an overview of the implementation of the Yosys Verilog and AST frontends. The Verilog frontend reads Verilog-2005 code and creates an abstract syntax tree (AST) representation of the input. This AST representation is then passed to the AST frontend that converts it to RTLIL data, as illustrated in Fig. [fig:Verilog_flow].

Transforming Verilog to AST¶

The Verilog frontend converts the Verilog sources to an internal AST representation that closely resembles the structure of the original Verilog code. The Verilog frontend consists of three components, the Preprocessor, the Lexer and the Parser.

The source code to the Verilog frontend can be found in frontends/verilog/ in the Yosys source tree.

The Verilog Preprocessor¶

The Verilog preprocessor scans over the Verilog source code and interprets some of the Verilog compiler directives such as `include, `define and `ifdef.

It is implemented as a C++ function that is passed a file descriptor as input and returns the pre-processed Verilog code as a std::string.

The source code to the Verilog Preprocessor can be found in frontends/verilog/preproc.cc in the Yosys source tree.

The Verilog Lexer¶

The Verilog Lexer is written using the lexer generator flex . Its source code can be found in frontends/verilog/verilog_lexer.l in the Yosys source tree. The lexer does little more than identifying all keywords and literals recognised by the Yosys Verilog frontend.

The lexer keeps track of the current location in the Verilog source code using some global variables. These variables are used by the constructor of AST nodes to annotate each node with the source code location it originated from.

Finally the lexer identifies and handles special comments such as “// synopsys translate_off” and “// synopsys full_case”. (It is recommended to use `ifdef constructs instead of the Synsopsys translate_on/off comments and attributes such as (* full_case *) over “// synopsys full_case” whenever possible.)

The Verilog Parser¶

The Verilog Parser is written using the parser generator bison . Its source code can be found in frontends/verilog/verilog_parser.y in the Yosys source tree.

It generates an AST using the AST::AstNode data structure defined in frontends/ast/ast.h. An AST::AstNode object has the following properties:

constructs. (continued on next page)

AST Node Type

Corresponding Verilog Construct

AST_NONE

This Node type should never be used.

This node type is used for the top node of the AST tree. It has no corresponding Verilog construct.

, AST_TASK, AST_FUNCTION

module, task and function

input, output, wire, reg and integer

Verilog Arrays

Created by the simplifier when an undeclared signal name is used.

, AST_LOCALPARAM

parameter and localparam

Parameter set in cell instantiation

Port connection in cell instantiation

Bit-Index in a signal or element index in array

A literal value

The type of cell in cell instantiation

An Identifier (signal name in expression or cell/task/etc. name in other contexts)

Construct an identifier in the form <prefix>[<index>].<suffix> (used only in advanced generate constructs)

, AST_TCALL

Call to function or task

, AST_TO_UNSIGNED

The $signed() and $unsigned() functions

>p7cm>p8cm AST Node Type & Corresponding Verilog Construct
AST_CONCAT AST_REPLICATE & The {...} and {...{...}}
operators
AST_BIT_NOT, AST_BIT_AND, AST_BIT_OR, AST_BIT_XOR,
AST_BIT_XNOR & The bitwise operators ~, &, |, ^
and ~^
AST_REDUCE_AND, AST_REDUCE_OR, AST_REDUCE_XOR,
AST_REDUCE_XNOR & The unary reduction operators ~, &,
|, ^ and ~^
AST_REDUCE_BOOL & Conversion from multi-bit value to boolean
value (equivalent to AST_REDUCE_OR)
AST_SHIFT_LEFT, AST_SHIFT_RIGHT, AST_SHIFT_SLEFT,
AST_SHIFT_SRIGHT & The shift operators <<, >>, <<<
and >>>
AST_LT, AST_LE, AST_EQ, AST_NE, AST_GE,
AST_GT & The relational operators <, <=, ==,
!=, >= and >
AST_ADD, AST_SUB, AST_MUL, AST_DIV, AST_MOD,
AST_POW & The binary operators +, -, *, /,
% and **
AST_POS, AST_NEG & The prefix operators + and -
AST_LOGIC_AND, AST_LOGIC_OR, AST_LOGIC_NOT & The logic
operators &&, || and !
AST_TERNARY & The ternary ?:-operator
AST_MEMRD AST_MEMWR & Read and write memories. These nodes
are generated by the AST simplifier for writes/reads to/from
Verilog arrays.
AST_ASSIGN & An assign statement
AST_CELL & A cell instantiation
AST_PRIMITIVE & A primitive cell (and, nand, or,
etc.)
AST_ALWAYS, AST_INITIAL & Verilog always- and
initial-blocks
AST_BLOCK & A begin-end-block
AST_ASSIGN_EQ. AST_ASSIGN_LE & Blocking (=) and
nonblocking (<=) assignments within an always- or
initial-block
AST_CASE. AST_COND, AST_DEFAULT & The case (if)
statements, conditions within a case and the default case
respectively
AST_FOR & A for-loop with an always- or
initial-block
AST_GENVAR, AST_GENBLOCK, AST_GENFOR, AST_GENIF &
The genvar and generate keywords and for and if
within a generate block.
AST_POSEDGE, AST_NEGEDGE, AST_EDGE & Event conditions
for always blocks.

The node type

This enum (AST::AstNodeType) specifies the role of the node. Table 1.1 contains a list of all node types.
The child nodes

This is a list of pointers to all children in the abstract syntax tree.
Attributes

As almost every AST node might have Verilog attributes assigned to it, the AST::AstNode has direct support for attributes. Note that the attribute values are again AST nodes.
Node content

Each node might have additional content data. A series of member variables exist to hold such data. For example the member std::string str can hold a string value and is used e.g. in the AST_IDENTIFIER node type to store the identifier name.
Source code location

Each AST::AstNode is automatically annotated with the current source code location by the AST::AstNode constructor. It is stored in the std::string filename and int linenum member variables.

The AST::AstNode constructor can be called with up to two child nodes that are automatically added to the list of child nodes for the new object. This simplifies the creation of AST nodes for simple expressions a bit. For example the bison code for parsing multiplications:

basic_expr '*' attr basic_expr {
                $$ = new AstNode(AST_MUL, $1, $4);
                append_attr($$, $3);
        } |

The generated AST data structure is then passed directly to the AST frontend that performs the actual conversion to RTLIL.

Note that the Yosys command read_verilog provides the options -yydebug and -dump_ast that can be used to print the parse tree or abstract syntax tree respectively.

Transforming AST to RTLIL¶

The AST Frontend converts a set of modules in AST representation to modules in RTLIL representation and adds them to the current design. This is done in two steps: simplification and RTLIL generation.

The source code to the AST frontend can be found in frontends/ast/ in the Yosys source tree.

AST Simplification¶

A full-featured AST is too complex to be transformed into RTLIL directly. Therefore it must first be brought into a simpler form. This is done by calling the AST::AstNode::simplify() method of all AST_MODULE nodes in the AST. This initiates a recursive process that performs the following transformations on the AST data structure:

Inline all task and function calls.
Evaluate all generate-statements and unroll all for-loops.
Perform const folding where it is necessary (e.g. in the value part of AST_PARAMETER, AST_LOCALPARAM, AST_PARASET and AST_RANGE nodes).
Replace AST_PRIMITIVE nodes with appropriate AST_ASSIGN nodes.
Replace dynamic bit ranges in the left-hand-side of assignments with AST_CASE nodes with AST_COND children for each possible case.
Detect array access patterns that are too complicated for the RTLIL::Memory abstraction and replace them with a set of signals and cases for all reads and/or writes.
Otherwise replace array accesses with AST_MEMRD and AST_MEMWR nodes.

In addition to these transformations, the simplifier also annotates the AST with additional information that is needed for the RTLIL generator, namely:

All ranges (width of signals and bit selections) are not only const folded but (when a constant value is found) are also written to member variables in the AST_RANGE node.
All identifiers are resolved and all AST_IDENTIFIER nodes are annotated with a pointer to the AST node that contains the declaration of the identifier. If no declaration has been found, an AST_AUTOWIRE node is created and used for the annotation.

This produces an AST that is fairly easy to convert to the RTLIL format.

Generating RTLIL¶

After AST simplification, the AST::AstNode::genRTLIL() method of each AST_MODULE node in the AST is called. This initiates a recursive process that generates equivalent RTLIL data for the AST data.

The AST::AstNode::genRTLIL() method returns an RTLIL::SigSpec structure. For nodes that represent expressions (operators, constants, signals, etc.), the cells needed to implement the calculation described by the expression are created and the resulting signal is returned. That way it is easy to generate the circuits for large expressions using depth-first recursion. For nodes that do not represent an expression (such as AST_CELL), the corresponding circuit is generated and an empty RTLIL::SigSpec is returned.

Synthesizing Verilog always Blocks¶

For behavioural Verilog code (code utilizing always- and initial-blocks) it is necessary to also generate RTLIL::Process objects. This is done in the following way:

Whenever AST::AstNode::genRTLIL() encounters an always- or initial-block, it creates an instance of AST_INTERNAL::ProcessGenerator. This object then generates the RTLIL::Process object for the block. It also calls AST::AstNode::genRTLIL() for all right-hand-side expressions contained within the block.

First the AST_INTERNAL::ProcessGenerator creates a list of all signals assigned within the block. It then creates a set of temporary signals using the naming scheme $<number> \<original_name> for each of the assigned signals.

Then an RTLIL::Process is created that assigns all intermediate values for each left-hand-side signal to the temporary signal in its RTLIL::CaseRule/RTLIL::SwitchRule tree.

Finally a RTLIL::SyncRule is created for the RTLIL::Process that assigns the temporary signals for the final values to the actual signals.

A process may also contain memory writes. A RTLIL::MemWriteAction is created for each of them.

Calls to AST::AstNode::genRTLIL() are generated for right hand sides as needed. When blocking assignments are used, AST::AstNode::genRTLIL() is configured using global variables to use the temporary signals that hold the correct intermediate values whenever one of the previously assigned signals is used in an expression.

Unfortunately the generation of a correct RTLIL::CaseRule/RTLIL::SwitchRule tree for behavioural code is a non-trivial task. The AST frontend solves the problem using the approach described on the following pages. The following example illustrates what the algorithm is supposed to do. Consider the following Verilog code:

always @(posedge clock) begin
    out1 = in1;
    if (in2)
        out1 = !out1;
    out2 <= out1;
    if (in3)
        out2 <= out2;
    if (in4)
        if (in5)
            out3 <= in6;
        else
            out3 <= in7;
    out1 = out1 ^ out2;
end

This is translated by the Verilog and AST frontends into the following RTLIL code (attributes, cell parameters and wire declarations not included):

cell $logic_not $logic_not$<input>:4$2
  connect \A \in1
  connect \Y $logic_not$<input>:4$2_Y
end
cell $xor $xor$<input>:13$3
  connect \A $1\out1[0:0]
  connect \B \out2
  connect \Y $xor$<input>:13$3_Y
end
process $proc$<input>:1$1
  assign $0\out3[0:0] \out3
  assign $0\out2[0:0] $1\out1[0:0]
  assign $0\out1[0:0] $xor$<input>:13$3_Y
  switch \in2
    case 1'1
      assign $1\out1[0:0] $logic_not$<input>:4$2_Y
    case
      assign $1\out1[0:0] \in1
  end
  switch \in3
    case 1'1
      assign $0\out2[0:0] \out2
    case
  end
  switch \in4
    case 1'1
      switch \in5
        case 1'1
          assign $0\out3[0:0] \in6
        case
          assign $0\out3[0:0] \in7
      end
    case
  end
  sync posedge \clock
    update \out1 $0\out1[0:0]
    update \out2 $0\out2[0:0]
    update \out3 $0\out3[0:0]
end

Note that the two operators are translated into separate cells outside the generated process. The signal out1 is assigned using blocking assignments and therefore out1 has been replaced with a different signal in all expressions after the initial assignment. The signal out2 is assigned using nonblocking assignments and therefore is not substituted on the right-hand-side expressions.

The RTLIL::CaseRule/RTLIL::SwitchRule tree must be interpreted the following way:

On each case level (the body of the process is the root case), first the actions on this level are evaluated and then the switches within the case are evaluated. (Note that the last assignment on line 13 of the Verilog code has been moved to the beginning of the RTLIL process to line 13 of the RTLIL listing.)

I.e. the special cases deeper in the switch hierarchy override the defaults on the upper levels. The assignments in lines 12 and 22 of the RTLIL code serve as an example for this.

Note that in contrast to this, the order within the RTLIL::SwitchRule objects within a RTLIL::CaseRule is preserved with respect to the original AST and Verilog code.
The whole RTLIL::CaseRule/RTLIL::SwitchRule tree describes an asynchronous circuit. I.e. the decision tree formed by the switches can be seen independently for each assigned signal. Whenever one assigned signal changes, all signals that depend on the changed signals are to be updated. For example the assignments in lines 16 and 18 in the RTLIL code in fact influence the assignment in line 12, even though they are in the “wrong order”.

The only synchronous part of the process is in the RTLIL::SyncRule object generated at line 35 in the RTLIL code. The sync rule is the only part of the process where the original signals are assigned. The synchronization event from the original Verilog code has been translated into the synchronization type (posedge) and signal (\clock) for the RTLIL::SyncRule object. In the case of this simple example the RTLIL::SyncRule object is later simply transformed into a set of d-type flip-flops and the RTLIL::CaseRule/RTLIL::SwitchRule tree to a decision tree using multiplexers.

In more complex examples (e.g. asynchronous resets) the part of the RTLIL::CaseRule/RTLIL::SwitchRule tree that describes the asynchronous reset must first be transformed to the correct RTLIL::SyncRule objects. This is done by the proc_adff pass.

The ProcessGenerator Algorithm¶

The AST_INTERNAL::ProcessGenerator uses the following internal state variables:

subst_rvalue_from and subst_rvalue_to

These two variables hold the replacement pattern that should be used by AST::AstNode::genRTLIL() for signals with blocking assignments. After initialization of AST_INTERNAL::ProcessGenerator these two variables are empty.
subst_lvalue_from and subst_lvalue_to

These two variables contain the mapping from left-hand-side signals (\<name>) to the current temporary signal for the same thing (initially $0\<name>).
current_case

A pointer to a RTLIL::CaseRule object. Initially this is the root case of the generated RTLIL::Process.

As the algorithm runs these variables are continuously modified as well as pushed to the stack and later restored to their earlier values by popping from the stack.

On startup the ProcessGenerator generates a new RTLIL::Process object with an empty root case and initializes its state variables as described above. Then the RTLIL::SyncRule objects are created using the synchronization events from the AST_ALWAYS node and the initial values of subst_lvalue_from and subst_lvalue_to. Then the AST for this process is evaluated recursively.

During this recursive evaluation, three different relevant types of AST nodes can be discovered: AST_ASSIGN_LE (nonblocking assignments), AST_ASSIGN_EQ (blocking assignments) and AST_CASE (if or case statement).

Handling of Nonblocking Assignments¶

When an AST_ASSIGN_LE node is discovered, the following actions are performed by the ProcessGenerator:

The left-hand-side is evaluated using AST::AstNode::genRTLIL() and mapped to a temporary signal name using subst_lvalue_from and subst_lvalue_to.
The right-hand-side is evaluated using AST::AstNode::genRTLIL(). For this call, the values of subst_rvalue_from and subst_rvalue_to are used to map blocking-assigned signals correctly.
Remove all assignments to the same left-hand-side as this assignment from the current_case and all cases within it.
Add the new assignment to the current_case.

Handling of Blocking Assignments¶

When an AST_ASSIGN_EQ node is discovered, the following actions are performed by the ProcessGenerator:

Perform all the steps that would be performed for a nonblocking assignment (see above).
Remove the found left-hand-side (before lvalue mapping) from subst_rvalue_from and also remove the respective bits from subst_rvalue_to.
Append the found left-hand-side (before lvalue mapping) to subst_rvalue_from and append the found right-hand-side to subst_rvalue_to.

Handling of Cases and if-Statements¶

When an AST_CASE node is discovered, the following actions are performed by the ProcessGenerator:

The values of subst_rvalue_from, subst_rvalue_to, subst_lvalue_from and subst_lvalue_to are pushed to the stack.
A new RTLIL::SwitchRule object is generated, the selection expression is evaluated using AST::AstNode::genRTLIL() (with the use of subst_rvalue_from and subst_rvalue_to) and added to the RTLIL::SwitchRule object and the object is added to the current_case.
All lvalues assigned to within the AST_CASE node using blocking assignments are collected and saved in the local variable this_case_eq_lvalue.
New temporary signals are generated for all signals in this_case_eq_lvalue and stored in this_case_eq_ltemp.
The signals in this_case_eq_lvalue are mapped using subst_rvalue_from and subst_rvalue_to and the resulting set of signals is stored in this_case_eq_rvalue.

Then the following steps are performed for each AST_COND node within the AST_CASE node:

Set subst_rvalue_from, subst_rvalue_to, subst_lvalue_from and subst_lvalue_to to the values that have been pushed to the stack.
Remove this_case_eq_lvalue from subst_lvalue_from/subst_lvalue_to.
Append this_case_eq_lvalue to subst_lvalue_from and append this_case_eq_ltemp to subst_lvalue_to.
Push the value of current_case.
Create a new RTLIL::CaseRule. Set current_case to the new object and add the new object to the RTLIL::SwitchRule created above.
Add an assignment from this_case_eq_rvalue to this_case_eq_ltemp to the new current_case.
Evaluate the compare value for this case using AST::AstNode::genRTLIL() (with the use of subst_rvalue_from and subst_rvalue_to) modify the new current_case accordingly.
Recursion into the children of the AST_COND node.
Restore current_case by popping the old value from the stack.

Finally the following steps are performed:

The values of subst_rvalue_from, subst_rvalue_to, subst_lvalue_from and subst_lvalue_to are popped from the stack.
The signals from this_case_eq_lvalue are removed from the subst_rvalue_from/subst_rvalue_to-pair.
The value of this_case_eq_lvalue is appended to subst_rvalue_from and the value of this_case_eq_ltemp is appended to subst_rvalue_to.
Map the signals in this_case_eq_lvalue using subst_lvalue_from/subst_lvalue_to.
Remove all assignments to signals in this_case_eq_lvalue in current_case and all cases within it.
Add an assignment from this_case_eq_ltemp to this_case_eq_lvalue to current_case.

Further Analysis of the Algorithm for Cases and if-Statements¶

With respect to nonblocking assignments the algorithm is easy: later assignments invalidate earlier assignments. For each signal assigned using nonblocking assignments exactly one temporary variable is generated (with the $0-prefix) and this variable is used for all assignments of the variable.

Note how all the _eq_-variables become empty when no blocking assignments are used and many of the steps in the algorithm can then be ignored as a result of this.

For a variable with blocking assignments the algorithm shows the following behaviour: First a new temporary variable is created. This new temporary variable is then registered as the assignment target for all assignments for this variable within the cases for this AST_CASE node. Then for each case the new temporary variable is first assigned the old temporary variable. This assignment is overwritten if the variable is actually assigned in this case and is kept as a default value otherwise.

This yields an RTLIL::CaseRule that assigns the new temporary variable in all branches. So when all cases have been processed a final assignment is added to the containing block that assigns the new temporary variable to the old one. Note how this step always overrides a previous assignment to the old temporary variable. Other than nonblocking assignments, the old assignment could still have an effect somewhere in the design, as there have been calls to AST::AstNode::genRTLIL() with a subst_rvalue_from/subst_rvalue_to-tuple that contained the right-hand-side of the old assignment.

The proc pass¶

The ProcessGenerator converts a behavioural model in AST representation to a behavioural model in RTLIL::Process representation. The actual conversion from a behavioural model to an RTL representation is performed by the proc pass and the passes it launches:

proc_clean and proc_rmdead

These two passes just clean up the RTLIL::Process structure. The proc_clean pass removes empty parts (eg. empty assignments) from the process and proc_rmdead detects and removes unreachable branches from the process’s decision trees.
proc_arst

This pass detects processes that describe d-type flip-flops with asynchronous resets and rewrites the process to better reflect what they are modelling: Before this pass, an asynchronous reset has two edge-sensitive sync rules and one top-level for the reset path. After this pass the sync rule for the reset is level-sensitive and the top-level has been removed.
proc_mux

This pass converts the /-tree to a tree of multiplexers per written signal. After this, the structure only contains the s that describe the output registers.
proc_dff

This pass replaces the s to d-type flip-flops (with asynchronous resets if necessary).
proc_dff

This pass replaces the s with $memwr cells.
proc_clean

A final call to proc_clean removes the now empty objects.

Performing these last processing steps in passes instead of in the Verilog frontend has two important benefits:

First it improves the transparency of the process. Everything that happens in a separate pass is easier to debug, as the RTLIL data structures can be easily investigated before and after each of the steps.

Second it improves flexibility. This scheme can easily be extended to support other types of storage-elements, such as sr-latches or d-latches, without having to extend the actual Verilog frontend.

Synthesizing Verilog Arrays¶

Add some information on the generation of $memrd and $memwr cells and how they are processed in the memory pass.

Synthesizing Parametric Designs¶

Add some information on the RTLIL::Module::derive() method and how it is used to synthesize parametric modules via the hierarchy pass.