| VERILATOR(1) | User Contributed Perl Documentation | VERILATOR(1) |
Verilator - Translate and simulate SystemVerilog code using C++/SystemC
verilator --help
verilator --version
verilator --cc [options] [source_files.v]... [opt_c_files.cpp/c/cc/a/o/so]
verilator --sc [options] [source_files.v]... [opt_c_files.cpp/c/cc/a/o/so]
verilator --lint-only -Wall [source_files.v]...
The "Verilator" package converts all synthesizable, and many behavioral, Verilog and SystemVerilog designs into a C++ or SystemC model that after compiling can be executed. Verilator is not a traditional simulator, but a compiler.
Verilator is typically used as follows:
1. The "verilator" executable is invoked with parameters similar to GCC, Cadence Verilog-XL/NC-Verilog, or Synopsys VCS. "verilator" reads the specified user's SystemVerilog code, lints it, optionally adds coverage and waveform tracing support, and compiles the design into a source level C++ or SystemC "model". The resulting model's C++ or SystemC code is output as .cpp and .h files. This is referred to as "verilating" and the process is "to verilate"; the output is a "verilated" model.
2. For simulation, a small user written C++ wrapper file is required, the "wrapper". This wrapper defines the C++ function `main()` which instantiates the Verilated model as a C++/SystemC object.
3. The user main wrapper, the files created by Verilator, a "runtime library" provided by Verilator, and if applicable the SystemC libraries are then compiled using a C++ compiler to create a simulation executable.
4. The resulting executable will perform the actual simulation, during "simulation runtime".
To get started, jump down to the "EXAMPLE C++ EXECUTION" section.
This is a short summary of the arguments to the "verilator" executable. See "VERILATION ARGUMENTS" for the detailed descriptions of these arguments.
{file.v} Verilog package, module and top module filenames
{file.c/cc/cpp} Optional C++ files to compile in
{file.a/o/so} Optional C++ files to link in
+1364-1995ext+<ext> Use Verilog 1995 with file extension <ext>
+1364-2001ext+<ext> Use Verilog 2001 with file extension <ext>
+1364-2005ext+<ext> Use Verilog 2005 with file extension <ext>
+1800-2005ext+<ext> Use SystemVerilog 2005 with file extension <ext>
+1800-2009ext+<ext> Use SystemVerilog 2009 with file extension <ext>
+1800-2012ext+<ext> Use SystemVerilog 2012 with file extension <ext>
+1800-2017ext+<ext> Use SystemVerilog 2017 with file extension <ext>
--assert Enable all assertions
--autoflush Flush streams after all $displays
--bbox-sys Blackbox unknown $system calls
--bbox-unsup Blackbox unsupported language features
--bin <filename> Override Verilator binary
--build Build model executable/library after Verilation
-CFLAGS <flags> C++ compiler flags for makefile
--cc Create C++ output
--cdc Clock domain crossing analysis
--clk <signal-name> Mark specified signal as clock
--make <build-tool> Generate scripts for specified build tool
--compiler <compiler-name> Tune for specified C++ compiler
--converge-limit <loops> Tune convergence settle time
--coverage Enable all coverage
--coverage-line Enable line coverage
--coverage-toggle Enable toggle coverage
--coverage-user Enable SVL user coverage
--coverage-underscore Enable coverage of _signals
-D<var>[=<value>] Set preprocessor define
--debug Enable debugging
--debug-check Enable debugging assertions
--no-debug-leak Disable leaking memory in --debug mode
--debugi <level> Enable debugging at a specified level
--debugi-<srcfile> <level> Enable debugging a source file at a level
--default-language <lang> Default language to parse
+define+<var>=<value> Set preprocessor define
--dpi-hdr-only Only produce the DPI header file
--dump-defines Show preprocessor defines with -E
--dump-tree Enable dumping .tree files
--dump-treei <level> Enable dumping .tree files at a level
--dump-treei-<srcfile> <level> Enable dumping .tree file at a source file at a level
--dump-tree-addrids Use short identifiers instead of addresses
-E Preprocess, but do not compile
--error-limit <value> Abort after this number of errors
--exe Link to create executable
-F <file> Parse options from a file, relatively
-f <file> Parse options from a file
-FI <file> Force include of a file
--flatten Force inlining of all modules, tasks and functions
-G<name>=<value> Overwrite top-level parameter
--gdb Run Verilator under GDB interactively
--gdbbt Run Verilator under GDB for backtrace
--generate-key Create random key for --protect-key
--getenv <var> Get environment variable with defaults
--help Display this help
-I<dir> Directory to search for includes
-j <jobs> Parallelism for --build
--gate-stmts <value> Tune gate optimizer depth
--if-depth <value> Tune IFDEPTH warning
+incdir+<dir> Directory to search for includes
--inhibit-sim Create function to turn off sim
--inline-mult <value> Tune module inlining
-LDFLAGS <flags> Linker pre-object flags for makefile
--l2-name <value> Verilog scope name of the top module
--language <lang> Default language standard to parse
+libext+<ext>+[ext]... Extensions for finding modules
--lint-only Lint, but do not make output
-MAKEFLAGS <flags> Options to make during --build
--max-num-width <value> Maximum number width (default: 64K)
--MMD Create .d dependency files
--MP Create phony dependency targets
--Mdir <directory> Name of output object directory
--mod-prefix <topname> Name to prepend to lower classes
--no-clk <signal-name> Prevent marking specified signal as clock
--no-decoration Disable comments and symbol decorations
--no-pins64 Don't use vluint64_t's for 33-64 bit sigs
--no-skip-identical Disable skipping identical output
+notimingchecks Ignored
-O0 Disable optimizations
-O3 High performance optimizations
-O<optimization-letter> Selectable optimizations
-o <executable> Name of final executable
--no-order-clock-delay Disable ordering clock enable assignments
--no-verilate Skip verilation and just compile previously Verilated code.
--output-split <statements> Split .cpp files into pieces
--output-split-cfuncs <statements> Split model functions
--output-split-ctrace <statements> Split tracing functions
-P Disable line numbers and blanks with -E
--pins-bv <bits> Specify types for top level ports
--pins-sc-uint Specify types for top level ports
--pins-sc-biguint Specify types for top level ports
--pins-uint8 Specify types for top level ports
--pipe-filter <command> Filter all input through a script
--pp-comments Show preprocessor comments with -E
--prefix <topname> Name of top level class
--prof-cfuncs Name functions for profiling
--prof-threads Enable generating gantt chart data for threads
--protect-key <key> Key for symbol protection
--protect-ids Hash identifier names for obscurity
--protect-lib <name> Create a DPI protected library
--private Debugging; see docs
--public Debugging; see docs
--public-flat-rw Mark all variables, etc as public_flat_rw
-pvalue+<name>=<value> Overwrite toplevel parameter
--quiet-exit Don't print the command on failure
--relative-includes Resolve includes relative to current file
--no-relative-cfuncs Disallow 'this->' in generated functions
--report-unoptflat Extra diagnostics for UNOPTFLAT
--rr Run Verilator and record with rr
--savable Enable model save-restore
--sc Create SystemC output
--stats Create statistics file
--stats-vars Provide statistics on variables
-sv Enable SystemVerilog parsing
+systemverilogext+<ext> Synonym for +1800-2017ext+<ext>
--threads <threads> Enable multithreading
--threads-dpi <mode> Enable multithreaded DPI
--threads-max-mtasks <mtasks> Tune maximum mtask partitioning
--timescale <timescale> Sets default timescale
--timescale-override <timescale> Overrides all timescales
--top-module <topname> Name of top level input module
--trace Enable waveform creation
--trace-coverage Enable tracing of coverage
--trace-depth <levels> Depth of tracing
--trace-fst Enable FST waveform creation
--trace-max-array <depth> Maximum bit width for tracing
--trace-max-width <width> Maximum array depth for tracing
--trace-params Enable tracing of parameters
--trace-structs Enable tracing structure names
--trace-threads <threads> Enable waveform creation on separate threads
--trace-underscore Enable tracing of _signals
-U<var> Undefine preprocessor define
--unroll-count <loops> Tune maximum loop iterations
--unroll-stmts <stmts> Tune maximum loop body size
--unused-regexp <regexp> Tune UNUSED lint signals
-V Verbose version and config
-v <filename> Verilog library
+verilog1995ext+<ext> Synonym for +1364-1995ext+<ext>
+verilog2001ext+<ext> Synonym for +1364-2001ext+<ext>
--version Displays program version and exits
--vpi Enable VPI compiles
--waiver-output <filename> Create a waiver file based on the linter warnings
-Wall Enable all style warnings
-Werror-<message> Convert warnings to errors
-Wfuture-<message> Disable unknown message warnings
-Wno-<message> Disable warning
-Wno-context Disable source context on warnings
-Wno-fatal Disable fatal exit on warnings
-Wno-lint Disable all lint warnings
-Wno-style Disable all style warnings
-Wpedantic Warn on compliance-test issues
--x-assign <mode> Assign non-initial Xs to this value
--x-initial <mode> Assign initial Xs to this value
--x-initial-edge Enable initial X->0 and X->1 edge triggers
--xml-only Create XML parser output
--xml-output XML output filename
-y <dir> Directory to search for modules
This is a short summary of the simulation runtime arguments, i.e. for the final Verilated simulation runtime models. See "SIMULATION RUNTIME ARGUMENTS" for the detailed description of these arguments.
+verilator+debug Enable debugging
+verilator+debugi+<value> Enable debugging at a level
+verilator+help Display help
+verilator+prof+threads+file+I<filename> Set profile filename
+verilator+prof+threads+start+I<value> Set profile starting point
+verilator+prof+threads+window+I<value> Set profile duration
+verilator+rand+reset+I<value> Set random reset technique
+verilator+seed+I<value> Set random seed
+verilator+noassert Disable assert checking
+verilator+V Verbose version and config
+verilator+version Show version and exit
The following are the arguments that may be passed to the "verilator" executable.
See also the -CFLAGS and -LDFLAGS options, which are useful when the C++ files need special compiler flags.
If any files are specified in this way, Verilator will include a make rule that uses these files when linking the module executable. This generally is only useful when used with the --exe option.
For compatibility with other simulators, see also the synonyms "+verilog1995ext+"ext, "+verilog2001ext+"ext, and "+systemverilogext+"ext.
For any source file, the language specified by these options takes precedence over any language specified by the "--default-language" or "--language" options.
These options take effect in the order they are encountered. Thus the following would use Verilog 1995 for "a.v" and Verilog 2001 for "b.v".
verilator ... +1364-1995ext+v a.v +1364-2001ext+v b.v
These flags are only recommended for legacy mixed language designs, as the preferable option is to edit the code to repair new keywords, or add appropriate "`begin_keywords".
Note "`begin_keywords" is a SystemVerilog construct, which specifies only the set of keywords to be recognized. This also controls some error messages that vary between language standards. Note at present Verilator tends to be overly permissive, e.g. it will accept many grammar and other semantic extensions which might not be legal when set to an older standard.
Using this argument will likely cause incorrect simulation.
Using this argument will likely cause incorrect simulation.
When make is run on the generated makefile these will be passed to the C++ compiler (g++/clang++/msvc++).
Currently only checks some items that other CDC tools missed; if you have interest in adding more traditional CDC checks, please contact the authors.
With --clk <signal-name>, user can specified root clock into the model, then Verilator will mark the signal as clocker and propagate the clocker attribute automatically to other signals derived from that. In this way, Verilator will try to avoid taking the clocker signal into checking list.
Note signal-name is specified by the RTL hierarchy path. For example, v.foo.bar. If the signal is the input to top-module, the directly the signal name. If you find it difficult to find the exact name, try to use "/*verilator clocker*/" in RTL file to mark the signal directly.
If clock signals are assigned to vectors and then later used individually, Verilator will attempt to decompose the vector and connect the single-bit clock signals directly. This should be transparent to the user.
Supported values are "gmake" for GNU Make and "cmake" for CMake. Both can be specified together. If no build tool is specified, gmake is assumed. The executable of gmake can be configured via environment variable "MAKE".
When using --build Verilator takes over the responsibility of building the model library/executable. For this reason --make cannot be specified when using --build.
Coverage analysis adds statements at each code flow change point (e.g. at branches). At each such branch a unique counter is incremented. At the end of a test, the counters along with the filename and line number corresponding to each counter are written into logs/coverage.dat.
Verilator automatically disables coverage of branches that have a $stop in them, as it is assumed $stop branches contain an error check that should not occur. A /*verilator coverage_block_off*/ comment will perform a similar function on any code in that block or below, or /*verilator coverage_on/coverage_off*/ will disable coverage around lines of code.
Note Verilator may over-count combinatorial (non-clocked) blocks when those blocks receive signals which have had the UNOPTFLAT warning disabled; for most accurate results do not disable this warning when using coverage.
Every bit of every signal in a module has a counter inserted. The counter will increment on every edge change of the corresponding bit.
Signals that are part of tasks or begin/end blocks are considered local variables and are not covered. Signals that begin with underscores, are integers, or are very wide (>256 bits total storage across all dimensions) are also not covered.
Hierarchy is compressed, such that if a module is instantiated multiple times, coverage will be summed for that bit across ALL instantiations of that module with the same parameter set. A module instantiated with different parameter values is considered a different module, and will get counted separately.
Verilator makes a minimally-intelligent decision about what clock domain the signal goes to, and only looks for edges in that clock domain. This means that edges may be ignored if it is known that the edge could never be seen by the receiving logic. This algorithm may improve in the future. The net result is coverage may be lower than what would be seen by looking at traces, but the coverage is a more accurate representation of the quality of stimulus into the design.
There may be edges counted near time zero while the model stabilizes. It's a good practice to zero all coverage just before releasing reset to prevent counting such behavior.
A /*verilator coverage_off/on */ comment pair can be used around signals that do not need toggle analysis, such as RAMs and register files.
For example, the following statement will add a coverage point, with the comment "DefaultClock":
DefaultClock: cover property (@(posedge clk) cyc==3);
This option disables the leak. This may avoid out-of-memory errors when Verilating large models in --debug mode.
Outside of --debug mode, AstNode instances should never be leaked and this option has no effect.
Any language associated with a particular file extension (see the various +langext+ options) will be used in preference to the language specified by --default-language.
The --default-language flag is only recommended for legacy code using the same language in all source files, as the preferable option is to edit the code to repair new keywords, or add appropriate "`begin_keywords". For legacy mixed language designs, the various +langext+ options should be used.
If no language is specified, either by this flag or +langext+ options, then the latest SystemVerilog language (IEEE 1800-2017) is used.
touch foo.v ; verilator -E --dump-defines foo.v
Does not affect simulation runtime errors, for those see +verilator+error+limit.
The file may contain // comments which are ignored to the end of the line. Any $VAR, $(VAR), or ${VAR} will be replaced with the specified environment variable.
When make is run on the generated makefile these will be passed to the C++ linker (ld) *after* the primary file being linked. This flag is called -LDFLAGS as that's the traditional name in simulators; it's would have been better called LDLIBS as that's the Makefile variable it controls. (In Make, LDFLAGS is before the first object, LDLIBS after. -L libraries need to be in the Make variable LDLIBS, not LDFLAGS.)
For example, the program "module t; initial $display("%m"); endmodule" will show by default "t". With "--l2-name v" it will print "v".
You may also want the -Wall option to enable messages that are considered stylistic and not enabled by default.
If the design is not to be completely Verilated see also the --bbox-sys and --bbox-unsup options.
This is a work around for old compilers. Don't set this if your C++ compiler supports __restrict__ properly, as GCC 4.5.x and newer do. For older compilers, test if this switch gives you better performance or not.
Compilers which don't honor __restrict__ will suspect that 'this->' references and 'vlTOPp->' references may alias, and may write slow code with extra loads and stores to handle the (imaginary) aliasing. Using only 'vlTOPp->' references allows these old compilers to produce tight code.
This option is on by default with a value of 20000. To disable, pass with a value of 0.
Defaults to the value of --output-split, unless explicitly specified.
Defaults to the value of --output-split, unless explicitly specified.
Before reading each Verilog file, Verilator will pass the file name to the subprocess' stdin with 'read "<filename>"'. The filter may then read the file and perform any filtering it desires, and feeds the new file contents back to Verilator on stdout by first emitting a line defining the length in bytes of the filtered output 'Content-Length: <bytes>', followed by the new filtered contents. Output to stderr from the filter feeds through to Verilator's stdout and if the filter exits with non-zero status Verilator terminates. See the t/t_pipe_filter test for an example.
To debug the output of the filter, try using the -E option to see preprocessed output.
Verilator will record the start and end time of each macro-task across a number of calls to eval. (What is a macro-task? See the Verilator internals document.)
When profiling is enabled, the simulation runtime will emit a blurb of profiling data in non-human-friendly form. The "verilator_gantt" script will transform this into a nicer visual format and produce some related statistics.
If --protect-key is not specified and a key is needed, Verilator will generate a new key for every Verilator run. As the key is not saved, this is best for security, but means every Verilator run will give vastly different output even for identical input, perhaps harming compile times (and certainly thrashing any ccache).
Verilator will also create a {prefix}__idmap.xml file which contains the mapping from the hashed identifiers back to the original identifiers. This idmap file is to be kept private, and is to assist mapping any simulation runtime design assertions, coverage, or trace information, which will report the hashed identifiers, back to the original design's identifier names.
Using DPI imports/exports is allowed and generally relatively safe in terms of information disclosed, which is limited to the DPI function prototyptes. Use of the VPI is not recommended as many design details may be exposed, and an INSECURE warning will be issued.
This allows for the secure delivery of sensitive IP without the need for encrypted RTL (i.e. IEEE P1735). See examples/make_protect_lib in the distribution for a demonstration of how to build and use the DPI library.
When using --protect-lib it is advised to also use "--timescale-override /1fs" to ensure the model has a time resolution that is always compatible with the time precision of the upper instantiating module.
Declares all signals and modules public. This will turn off signal optimizations as if all signals had a /*verilator public*/ comments and inlining. This will also turn off inlining as if all modules had a /*verilator public_module*/, unless the module specifically enabled it with /*verilator inline_module*/.
In addition produces a GraphViz DOT file of the entire strongly connected components within the source associated with each loop. This is produced irrespective of whether --dump-tree is set. Such graphs may help in analyzing the problem, but can be very large indeed.
Various commands exist for viewing and manipulating DOT files. For example the dot command can be used to convert a DOT file to a PDF for printing. For example:
dot -Tpdf -O Vt_unoptflat_simple_2_35_unoptflat.dot
will generate a PDF Vt_unoptflat_simple_2_35_unoptflat.dot.pdf from the DOT file.
As an alternative, the xdot command can be used to view DOT files interactively:
xdot Vt_unoptflat_simple_2_35_unoptflat.dot
The user code must create a VerilatedSerialize or VerilatedDeserialze object then calling the << or >> operators on the generated model and any other data the process needs saved/restored. These functions are not thread safe, and are typically called only by a main thread.
For example:
void save_model(const char* filenamep) {
VerilatedSave os;
os.open(filenamep);
os << main_time; // user code must save the timestamp, etc
os << *topp;
}
void restore_model(const char* filenamep) {
VerilatedRestore os;
os.open(filenamep);
os >> main_time;
os >> *topp;
}
With --threads-dpi all, enable Verilator to assume all DPI imports are threadsafe, and to use thread-local storage for communication with DPI, potentially improving performance. Any DPI libraries need appropriate mutexes to avoid undefined behavior.
With --threads-dpi none, Verilator assume DPI imports are not thread safe, and Verilator will serialize calls to DPI imports by default, potentially harming performance.
With --threads-dpi pure, the default, Verilator assumes DPI pure imports are threadsafe, but non-pure DPI imports are not.
The time precision must be consistent with SystemC's sc_set_time_resolution, or the C++ code instantiating the Verilated module. As 1fs is the finest time precision it may be desirable to always use a precision of 1fs.
Verilator will generate additional {prefix}__Trace*.cpp files that will need to be compiled. In addition verilated_vcd_sc.cpp (for SystemC traces) or verilated_vcd_c.cpp (for both) must be compiled and linked in. If using the Verilator generated Makefiles, these files will be added to the source file lists for you. If you are not using the Verilator Makefiles, you will need to add these to your Makefile manually.
Having tracing compiled in may result in some small performance losses, even when tracing is not turned on during model execution.
See also "--trace-threads".
The added signal will be a 32-bit value which will increment on each coverage occurrence. Due to this, this option may greatly increase trace file sizes and reduce simulation speed.
The generated file is in the Verilator Configuration format, see "CONFIGURATION FILES", and can directly be consumed by Verilator. The standard file extension is .vlt.
Having warning messages in builds is sloppy. It is strongly recommended you cleanup your code, use inline lint_off, or use -Wno-... flags rather than using this option.
It is strongly recommended you cleanup your code rather than using this option, it is only intended to be use when running test-cases of code received from third parties.
If using --x-assign unique, you may want to seed your random number generator such that each regression run gets a different randomization sequence. The simplest is to use the +verilator+seed runtime option. Alternatively use the system's srand48() or for Windows srand() function to do this. You'll probably also want to print any seeds selected, and code to enable rerunning with that same seed so you can reproduce bugs.
Note. This option applies only to values which are explicitly written as X in the Verilog source code. Initial values of clocks are set to 0 unless --x-initial-edge is specified. Initial values of all other state holding variables are controlled with --x-initial.
"--x-initial 0", initializes all otherwise uninitialized variables to zero.
"--x-initial unique", the default, initializes variables using a function, which determines the value to use each initialization. This gives greatest flexibility and allows finding reset bugs. See "Unknown states".
"--x-initial fast", is best for performance, and initializes all variables to a state Verilator determines is optimal. This may allow further code optimizations, but will likely hide any code bugs relating to missing resets.
Note. This option applies only to initial values of variables. Initial values of clocks are set to 0 unless --x-initial-edge is specified.
reg res_n = 1'b0;
always @(negedge rst_n) begin
if (rst_n == 1'b0) begin
res_n <= 1'b1;
end
end
In Verilator, by default, uninitialized clocks are given a value of zero, so the above "always" block would not trigger.
While it is not good practice, there are some designs that rely on X X 0 triggering a "negedge", particularly in reset sequences. Using --x-initial-edge with Verilator will replicate this behavior. It will also ensure that X X 1 triggers a "posedge".
Note. Some users have reported that using this option can affect convergence, and that it may be necessary to use --converge-limit to increase the number of convergence iterations. This may be another indication of problems with the modeled design that should be addressed.
The XML format is intended to be used to leverage Verilator's parser and elaboration to feed to other downstream tools. Be aware that the XML format is still evolving; there will be some changes in future versions.
Verilator defaults to the current directory ("-y .") and any specified --Mdir, though these default paths are used after any user specified directories. This allows '-y "$(pwd)"' to be used if absolute filenames are desired for error messages instead of relative filenames.
The following are the arguments that may be passed to a Verilated executable, provided that executable calls Verilated::commandArgs().
All simulation runtime arguments begin with +verilator, so that the user's executable may skip over all +verilator arguments when parsing its command line.
We'll compile this example into C++.
mkdir test_our
cd test_our
cat >our.v <<'EOF'
module our;
initial begin $display("Hello World"); $finish; end
endmodule
EOF
cat >sim_main.cpp <<'EOF'
#include "Vour.h"
#include "verilated.h"
int main(int argc, char** argv, char** env) {
Verilated::commandArgs(argc, argv);
Vour* top = new Vour;
while (!Verilated::gotFinish()) { top->eval(); }
delete top;
exit(0);
}
EOF
See the README in the source kit for various ways to install or point to Verilator binaries. In brief, if you installed Verilator using the package manager of your operating system, or did a "make install" to place Verilator into your default path, you do not need anything special in your environment, and should not have VERILATOR_ROOT set. However, if you installed Verilator from sources and want to run Verilator out of where you compiled Verilator, you need to point to the kit:
# See above; don't do this if using an OS-distributed Verilator
export VERILATOR_ROOT=/path/to/where/verilator/was/installed
export PATH=$VERILATOR_ROOT/bin:$PATH
Now we run Verilator on our little example.
verilator -Wall --cc our.v --exe --build sim_main.cpp
We can see the source code under the "obj_dir" directory. See the FILES section below for descriptions of some of the files that were created.
ls -l obj_dir
(Verilator included a default compile rule and link rule, since we used --exe and passed a .cpp file on the Verilator command line. Verilator also then used "make" to build a final executable. You can also write your own compile rules, and run make yourself as we'll show in the SYSTEMC section.)
And now we run it
obj_dir/Vour
And we get as output
Hello World
- our.v:2: Verilog $finish
Really, you're better off writing a Makefile to do all this for you. Then, when your source changes it will automatically run all of these steps; to aid this Verilator can create a makefile dependency file. See the examples directory in the distribution.
This is an example similar to the above, but using SystemC.
mkdir test_our_sc
cd test_our_sc
cat >our.v <<'EOF'
module our (clk);
input clk; // Clock is required to get initial activation
always @ (posedge clk)
begin $display("Hello World"); $finish; end
endmodule
EOF
cat >sc_main.cpp <<'EOF'
#include "Vour.h"
int sc_main(int argc, char** argv) {
Verilated::commandArgs(argc, argv);
sc_clock clk ("clk", 10, 0.5, 3, true);
Vour* top;
top = new Vour("top");
top->clk(clk);
while (!Verilated::gotFinish()) { sc_start(1, SC_NS); }
delete top;
exit(0);
}
EOF
See the README in the source kit for various ways to install or point to Verilator binaries. In brief, if you installed Verilator using the package manager of your operating system, or did a "make install" to place Verilator into your default path, you do not need anything special in your environment, and should not have VERILATOR_ROOT set. However, if you installed Verilator from sources and want to run Verilator out of where you compiled Verilator, you need to point to the kit:
# See above; don't do this if using an OS-distributed Verilator
export VERILATOR_ROOT=/path/to/where/verilator/was/installed
export PATH=$VERILATOR_ROOT/bin:$PATH
Now we run Verilator on our little example.
verilator -Wall --sc our.v
We then can compile it
make -j -C obj_dir -f Vour.mk Vour__ALL.a
make -j -C obj_dir -f Vour.mk ../sc_main.o verilated.o
And link with SystemC. Note your path to the libraries may vary, depending on the operating system.
cd obj_dir
export SYSTEMC_LIBDIR=/path/to/where/libsystemc.a/exists
export LD_LIBRARY_PATH=$SYSTEMC_LIBDIR:$LD_LIBRARY_PATH
# Might be needed if SystemC 2.3.0
export SYSTEMC_CXX_FLAGS=-pthread
g++ -L$SYSTEMC_LIBDIR ../sc_main.o Vour__ALL.a verilated.o \
-o Vour -lsystemc
And now we run it
cd ..
obj_dir/Vour
And we get the same output as the C++ example:
Hello World
- our.v:2: Verilog $finish
Really, you're better off using a Makefile to do all this for you. Then, when your source changes it will automatically run all of these steps. See the examples directory in the distribution.
When using SystemC, evaluation of the Verilated model is managed by the SystemC kernel, and for the most part can be ignored. When using C++, the user must call eval(), or eval_step() and eval_end_step().
1. When there is a single design instantiated at the C++ level that needs to evaluate, just call designp->eval().
2. When there are multiple designs instantiated at the C++ level that need to evaluate, call first_designp->eval_step() then ->eval_step() on all other designs. Then call ->eval_end_step() on the first design then all other designs. If there is only a single design, you would call eval_step() then eval_end_step(); in fact eval() described above is just a wrapper which calls these two functions.
When eval() is called Verilator looks for changes in clock signals and evaluates related sequential always blocks, such as computing always_ff @ (posedge...) outputs. Then Verilator evaluates combinatorial logic.
Note combinatorial logic is not computed before sequential always blocks are computed (for speed reasons). Therefore it is best to set any non-clock inputs up with a separate eval() call before changing clocks.
Alternatively, if all always_ff statements use only the posedge of clocks, or all inputs go directly to always_ff statements, as is typical, then you can change non-clock inputs on the negative edge of the input clock, which will be faster as there will be fewer eval() calls.
For more information on evaluation, see docs/internals.adoc in the distribution.
For best performance, run Verilator with the "-O3 --x-assign fast --x-initial fast --noassert" flags. The -O3 flag will require longer time to run Verilator, and "--x-assign fast --x-initial fast" may increase the risk of reset bugs in trade for performance; see the above documentation for these flags.
If using Verilated multithreaded, use "numactl" to ensure you are using non-conflicting hardware resources. See "MULTITHREADING".
Minor Verilog code changes can also give big wins. You should not have any UNOPTFLAT warnings from Verilator. Fixing these warnings can result in huge improvements; one user fixed their one UNOPTFLAT warning by making a simple change to a clock latch used to gate clocks and gained a 60% performance improvement.
Beyond that, the performance of a Verilated model depends mostly on your C++ compiler and size of your CPU's caches. Experience shows that large models are often limited by the size of the instruction cache, and as such reducing code size if possible can be beneficial.
The supplied $VERILATOR_ROOT/include/verilated.mk file uses the OPT, OPT_FAST, OPT_SLOW and OPT_GLOBAL variables to control optimization. You can set these when compiling the output of Verilator with Make, for example:
make OPT_FAST="-Os -march=native" -f Vour.mk Vour__ALL.a
OPT_FAST specifies optimization flags for those parts of the model that are on the fast path. This is mostly code that is executed every cycle. OPT_SLOW applies to slow-path code, which executes rarely, often only once at the beginning or end of simulation. Note that OPT_SLOW is ignored if VM_PARALLEL_BUILDS is not 1, in which case all generated code will be compiled in a single compilation unit using OPT_FAST. See also the "--output-split" option. The OPT_GLOBAL variable applies to common code in the runtime library used by Verilated models (shipped in $VERILATOR_ROOT/include). Additional C++ files passed on the verilator command line use OPT_FAST. The OPT variable applies to all compilation units in addition to the specific OPT_* variables described above.
You can also use the -CFLAGS and/or -LDFLAGS options on the verilator command line to pass flags directly to the compiler or linker.
The default values of the OPT_* variables are chosen to yield good simulation speed with reasonable C++ compilation times. To this end, OPT_FAST is set to "-Os" by default. Higher optimization such as "-O2" or "-O3" may help (though often they provide only a very small performance benefit), but compile times may be excessively large even with medium sized designs. Compilation times can be improved at the expense of simulation speed by reducing optimization, for example with OPT_FAST="-O0". Often good simulation speed can be achieved with OPT_FAST="-O1 -fstrict-aliasing" but with improved compilation times. Files controlled by OPT_SLOW have little effect on performance and therefore OPT_SLOW is empty by default (equivalent to "-O0") for improved compilation speed. In common use-cases there should be little benefit in changing OPT_SLOW. OPT_GLOBAL is set to "-Os" by default and there should rarely be a need to change it. As the runtime library is small in comparison to a lot of Verilated models, disabling optimization on the runtime library should not have a serious effect on overall compilation time, but may have detrimental effect on simulation speed, especially with tracing. In addition to the above, for best results use OPT="-march=native", the latest Clang compiler (about 10% faster than GCC), and link statically.
Generally the answer to which optimization level gives the best user experience depends on the use case and some experimentation can pay dividends. For a speedy debug cycle during development, especially on large designs where C++ compilation speed can dominate, consider using lower optimization to get to an executable faster. For throughput oriented use cases, for example regressions, it is usually worth spending extra compilation time to reduce total CPU time.
If you will be running many simulations on a single model, you can investigate profile guided optimization. With GCC, using -fprofile-arcs, then -fbranch-probabilities will yield another 15% or so.
Modern compilers also support link-time optimization (LTO), which can help especially if you link in DPI code. To enable LTO on GCC, pass "-flto" in both compilation and link. Note LTO may cause excessive compile times on large designs.
Unfortunately, using the optimizer with SystemC files can result in compilation taking several minutes. (The SystemC libraries have many little inlined functions that drive the compiler nuts.)
If you are using your own makefiles, you may want to compile the Verilated code with -DVL_INLINE_OPT=inline. This will inline functions, however this requires that all cpp files be compiled in a single compiler run.
You may uncover further tuning possibilities by profiling the Verilog code. Use Verilator's --prof-cfuncs, then GCC's -g -pg. You can then run either oprofile or gprof to see where in the C++ code the time is spent. Run the gprof output through verilator_profcfunc and it will tell you what Verilog line numbers on which most of the time is being spent.
When done, please let the author know the results. We like to keep tabs on how Verilator compares, and may be able to suggest additional improvements.
All output files are placed in the output directory specified with the "--Mdir" option, or "obj_dir" if not specified.
Verilator creates the following files in the output directory:
For --make gmake, it creates:
{prefix}.mk // Make include file for compiling
{prefix}_classes.mk // Make include file with class names
For --make cmake, it creates:
{prefix}.cmake // CMake include script for compiling
For -cc and -sc mode, it also creates:
{prefix}.cpp // Top level C++ file
{prefix}.h // Top level header
{prefix}__Slow{__n}.cpp // Constructors and infrequent cold routines
{prefix}{__n}.cpp // Additional top C++ files (--output-split)
{prefix}{each_verilog_module}.cpp // Lower level internal C++ files
{prefix}{each_verilog_module}.h // Lower level internal header files
{prefix}{each_verilog_module}{__n}.cpp // Additional lower C++ files (--output-split)
In certain debug and other modes, it also creates:
{prefix}.xml // XML tree information (--xml)
{prefix}__Dpi.cpp // DPI import and export wrappers
{prefix}__Dpi.h // DPI import and export declarations
{prefix}__Inlines.h // Inline support functions
{prefix}__Syms.cpp // Global symbol table C++
{prefix}__Syms.h // Global symbol table header
{prefix}__Trace__Slow{__n}.cpp // Wave file generation code (--trace)
{prefix}__Trace{__n}.cpp // Wave file generation code (--trace)
{prefix}__cdc.txt // Clock Domain Crossing checks (--cdc)
{prefix}__stats.txt // Statistics (--stats)
{prefix}__idmap.txt // Symbol demangling (--protect-ids)
It also creates internal files that can be mostly ignored:
{mod_prefix}_{each_verilog_module}{__n}.vpp // Pre-processed verilog
{prefix}__ver.d // Make dependencies (-MMD)
{prefix}__verFiles.dat // Timestamps for skip-identical
{prefix}{misc}.dot // Debugging graph files (--debug)
{prefix}{misc}.tree // Debugging files (--debug)
After running Make, the C++ compiler may produce the following:
verilated{misc}.d // Intermediate dependencies
verilated{misc}.o // Intermediate objects
{mod_prefix}{misc}.d // Intermediate dependencies
{mod_prefix}{misc}.o // Intermediate objects
{prefix} // Final executable (w/--exe argument)
{prefix}__ALL.a // Library of all Verilated objects
{prefix}__ALL.cpp // Include of all code for single compile
{prefix}{misc}.d // Intermediate dependencies
{prefix}{misc}.o // Intermediate objects
Verilator creates a prefix.h and prefix.cpp file for the top level module, together with additional .h and .cpp files for internals. See the examples directory in the kit for examples.
After the model is created, there will be a prefix.mk file that may be used with Make to produce a prefix__ALL.a file with all required objects in it. This is then linked with the user's C++ main loop to create the simulation executable.
The user must write the C++ main loop of the simulation. Here is a simple example:
#include <verilated.h> // Defines common routines
#include <iostream> // Need std::cout
#include "Vtop.h" // From Verilating "top.v"
Vtop *top; // Instantiation of module
vluint64_t main_time = 0; // Current simulation time
// This is a 64-bit integer to reduce wrap over issues and
// allow modulus. This is in units of the timeprecision
// used in Verilog (or from --timescale-override)
double sc_time_stamp () { // Called by $time in Verilog
return main_time; // converts to double, to match
// what SystemC does
}
int main(int argc, char** argv) {
Verilated::commandArgs(argc, argv); // Remember args
top = new Vtop; // Create instance
top->reset_l = 0; // Set some inputs
while (!Verilated::gotFinish()) {
if (main_time > 10) {
top->reset_l = 1; // Deassert reset
}
if ((main_time % 10) == 1) {
top->clk = 1; // Toggle clock
}
if ((main_time % 10) == 6) {
top->clk = 0;
}
top->eval(); // Evaluate model
cout << top->out << endl; // Read a output
main_time++; // Time passes...
}
top->final(); // Done simulating
// // (Though this example doesn't get here)
delete top;
}
Note signals are read and written as member variables of the model. You call the eval() method to evaluate the model. When the simulation is complete call the final() method to execute any SystemVerilog final blocks, and complete any assertions. See "EVALUATION LOOP".
Verilator will convert the top level module to a SC_MODULE. This module will plug directly into a SystemC netlist.
The SC_MODULE gets the same pinout as the Verilog module, with the following type conversions: Pins of a single bit become bool. Pins 2-32 bits wide become uint32_t's. Pins 33-64 bits wide become sc_bv's or vluint64_t's depending on the --no-pins64 switch. Wider pins become sc_bv's. (Uints simulate the fastest so are used where possible.)
Lower modules are not pure SystemC code. This is a feature, as using the SystemC pin interconnect scheme everywhere would reduce performance by an order of magnitude.
Verilator supports SystemVerilog Direct Programming Interface import and export statements. Only the SystemVerilog form ("DPI-C") is supported, not the original Synopsys-only DPI.
In the SYSTEMC example above, if you wanted to import C++ functions into Verilog, put in our.v:
import "DPI-C" function int add (input int a, input int b);
initial begin
$display("%x + %x = %x", 1, 2, add(1,2));
endtask
Then after Verilating, Verilator will create a file Vour__Dpi.h with the prototype to call this function:
extern int add (int a, int b);
From the sc_main.cpp file (or another .cpp file passed to the Verilator command line, or the link), you'd then:
#include "svdpi.h"
#include "Vour__Dpi.h"
int add(int a, int b) { return a+b; }
Verilator extends the DPI format to allow using the same scheme to efficiently add system functions. Simply use a dollar-sign prefixed system function name for the import, but note it must be escaped.
export "DPI-C" function integer \$myRand;
initial $display("myRand=%d", $myRand());
Going the other direction, you can export Verilog tasks so they can be called from C++:
export "DPI-C" task publicSetBool;
task publicSetBool;
input bit in_bool;
var_bool = in_bool;
endtask
Then after Verilating, Verilator will create a file Vour__Dpi.h with the prototype to call this function:
extern void publicSetBool(svBit in_bool);
From the sc_main.cpp file, you'd then:
#include "Vour__Dpi.h"
publicSetBool(value);
Or, alternatively, call the function under the design class. This isn't DPI compatible but is easier to read and better supports multiple designs.
#include "Vour__Dpi.h"
Vour::publicSetBool(value);
// or top->publicSetBool(value);
Note that if the DPI task or function accesses any register or net within the RTL, it will require a scope to be set. This can be done using the standard functions within svdpi.h, after the module is instantiated, but before the task(s) and/or function(s) are called.
For example, if the top level module is instantiated with the name "dut" and the name references within tasks are all hierarchical (dotted) names with respect to that top level module, then the scope could be set with
#include "svdpi.h"
...
svSetScope(svGetScopeFromName("TOP.dut"));
(Remember that Verilator adds a "TOP" to the top of the module hierarchy.)
Scope can also be set from within a DPI imported C function that has been called from Verilog by querying the scope of that function. See the sections on DPI Context Functions and DPI Header Isolation below and the comments within the svdpi.h header for more information.
Verilator allows writing $display like functions using this syntax:
import "DPI-C" function void
\$my_display(input string formatted /*verilator sformat*/ );
The /*verilator sformat*/ indicates that this function accepts a $display like format specifier followed by any number of arguments to satisfy the format.
Verilator supports IEEE DPI Context Functions. Context imports pass the simulator context, including calling scope name, and filename and line number to the C code. For example, in Verilog:
import "DPI-C" context function int dpic_line();
initial $display("This is line %d, again, line %d\n", `line, dpic_line());
This will call C++ code which may then use the svGet* functions to read information, in this case the line number of the Verilog statement that invoked the dpic_line function:
int dpic_line() {
// Get a scope: svScope scope = svGetScope();
const char* scopenamep = svGetNameFromScope(scope);
assert(scopenamep);
const char* filenamep = "";
int lineno = 0;
if (svGetCallerInfo(&filenamep, &lineno)) {
printf("dpic_line called from scope %s on line %d\n",
scopenamep, lineno);
return lineno;
} else {
return 0;
}
}
See the IEEE Standard for more information.
Verilator places the IEEE standard header files such as svdpi.h into a separate include directory, vltstd (VeriLaTor STandarD). When compiling most applications $VERILATOR_ROOT/include/vltstd would be in the include path along with the normal $VERILATOR_ROOT/include. However, when compiling Verilated models into other simulators which have their own svdpi.h and similar standard files with different contents, the vltstd directory should not be included to prevent picking up incompatible definitions.
Instead of DPI exporting, there's also Verilator public functions, which are slightly faster, but less compatible.
Verilator supports a very limited subset of the VPI. This subset allows inspection, examination, value change callbacks, and depositing of values to public signals only.
VPI is enabled with the verilator --vpi switch.
To access signals via the VPI, Verilator must be told exactly which signals are to be accessed. This is done using the Verilator public pragmas documented below.
Verilator has an important difference from an event based simulator; signal values that are changed by the VPI will not immediately propagate their values, instead the top level header file's eval() method must be called. Normally this would be part of the normal evaluation (i.e. the next clock edge), not as part of the value change. This makes the performance of VPI routines extremely fast compared to event based simulators, but can confuse some test-benches that expect immediate propagation.
Note the VPI by its specified implementation will always be much slower than accessing the Verilator values by direct reference (structure->module->signame), as the VPI accessors perform lookup in functions at simulation runtime requiring at best hundreds of instructions, while the direct references are evaluated by the compiler and result in only a couple of instructions.
For signal callbacks to work the main loop of the program must call VerilatedVpi::callValueCbs().
In the below example, we have readme marked read-only, and writeme which if written from outside the model will have the same semantics as if it changed on the specified clock edge.
cat >our.v <<'EOF'
module our (input clk);
reg readme /*verilator public_flat_rd*/;
reg writeme /*verilator public_flat_rw @(posedge clk) */;
initial $finish;
endmodule
EOF
There are many online tutorials and books on the VPI, but an example that accesses the above signal "readme" would be:
cat >sim_main.cpp <<'<<EOF'
#include "Vour.h"
#include "verilated.h"
#include "verilated_vpi.h" // Required to get definitions
vluint64_t main_time = 0; // See comments in first example
double sc_time_stamp() { return main_time; }
void read_and_check() {
vpiHandle vh1 = vpi_handle_by_name((PLI_BYTE8*)"TOP.our.readme", NULL);
if (!vh1) { vl_fatal(__FILE__, __LINE__, "sim_main", "No handle found"); }
const char* name = vpi_get_str(vpiName, vh1);
printf("Module name: %s\n", name); // Prints "readme"
s_vpi_value v;
v.format = vpiIntVal;
vpi_get_value(vh1, &v);
printf("Value of v: %d\n", v.value.integer); // Prints "readme"
}
int main(int argc, char** argv, char** env) {
Verilated::commandArgs(argc, argv);
Vour* top = new Vour;
Verilated::internalsDump(); // See scopes to help debug
while (!Verilated::gotFinish()) {
top->eval();
VerilatedVpi::callValueCbs(); // For signal callbacks
read_and_check();
}
delete top;
exit(0);
}
EOF
Verilator supports cross-compiling Verilated code. This is generally used to run Verilator on a Linux system and produce C++ code that is then compiled on Windows.
Cross compilation involves up to three different OSes. The build system is where you configured and compiled Verilator, the host system where you run Verilator, and the target system where you compile the Verilated code and run the simulation.
Currently, Verilator requires the build and host system type to be the same, though the target system type may be different. To support this, ./configure and make Verilator on the build system. Then, run Verilator on the host system. Finally, the output of Verilator may be compiled on the different target system.
To support this, none of the files that Verilator produces will reference any configure generated build-system specific files, such as config.h (which is renamed in Verilator to config_build.h to reduce confusion.) The disadvantage of this approach is that include/verilatedos.h must self-detect the requirements of the target system, rather than using configure.
The target system may also require edits to the Makefiles, the simple Makefiles produced by Verilator presume the target system is the same type as the build system.
Verilator can be run using CMake, which takes care of both running Verilator and compiling the output. There is a CMake example in the examples/ directory. The following is a minimal CMakeLists.txt that would build the code listed in "EXAMPLE C++ EXECUTION":
project(cmake_example)
find_package(verilator HINTS $ENV{VERILATOR_ROOT})
add_executable(Vour sim_main.cpp)
verilate(Vour SOURCES our.v)
find_package will automatically find an installed copy of Verilator, or use a local build if VERILATOR_ROOT is set.
It is recommended to use CMake >= 3.12 and the Ninja generator, though other combinations should work. To build with CMake, change to the folder containing CMakeLists.txt and run:
mkdir build
cd build
cmake -GNinja ..
ninja
Or to build with your system default generator:
mkdir build
cd build
cmake ..
cmake --build .
If you're building the example you should have an executable to run:
./Vour
The package sets the CMake variables verilator_FOUND, VERILATOR_ROOT and VERILATOR_BIN to the appropriate values, and also creates a verilate() function. verilate() will automatically create custom commands to run Verilator and add the generated C++ sources to the target specified.
verilate(target SOURCES source ... [TOP_MODULE top] [PREFIX name]
[TRACE] [TRACE_FST] [SYSTEMC] [COVERAGE]
[INCLUDE_DIRS dir ...] [OPT_SLOW ...] [OPT_FAST ...]
[OPT_GLOBAL ..] [DIRECTORY dir] [VERILATOR_ARGS ...])
Lowercase and ... should be replaced with arguments, the uppercase parts delimit the arguments and can be passed in any order, or left out entirely if optional.
verilate(target ...) can be called multiple times to add other verilog modules to an executable or library target.
When generating Verilated SystemC sources, you should also include the SystemC include directories and link to the SystemC libraries.
Verilator's CMake support provides a convenience function to automatically find and link to the SystemC library. It can be used as:
verilator_link_systemc(target)
where target is the name of your target.
The search paths can be configured by setting some variables:
- The variables SYSTEMC_INCLUDE and SYSTEMC_LIBDIR to give a direct path to the SystemC include an library path.
- SYSTEMC_ROOT to set the installation prefix of an installed
SystemC
library.
- SYSTEMC to set the installation prefix of an installed SystemC
library
(same as above).
- When using Accellera's SystemC with CMake support, a CMake target is available that simplifies the above steps. This will only work if the SystemC installation can be found by CMake. This can be configured by setting the CMAKE_PREFIX_PATH variable during CMake configuration.
Don't forget to set the same C++ standard for the Verilated sources as the SystemC library. This can be specified using the SYSTEMC_CXX_FLAGS environment variable.
Verilator supports multithreaded simulation models.
With --no-threads, the default, the model is not thread safe, and any use of more than one thread calling into one or even different Verilated models may result in unpredictable behavior. This gives the highest single thread performance.
With --threads 1, the generated model is single threaded, however the support libraries are multithread safe. This allows different instantiations of model(s) to potentially each be run under a different thread. All threading is the responsibility of the user's C++ testbench.
With --threads N, where N is at least 2, the generated model will be designed to run in parallel on N threads. The thread calling eval() provides one of those threads, and the generated model will create and manage the other N-1 threads. It's the client's responsibility not to oversubscribe the available CPU cores. Under CPU oversubscription, the Verilated model should not livelock nor deadlock, however, you can expect performance to be far worse than it would be with proper ratio of threads and CPU cores.
The remainder of this section describe behavior with --threads 1 or --threads N (not --no-threads).
VL_THREADED is defined when compiling a threaded Verilated module, causing the Verilated support classes become threadsafe.
The thread used for constructing a model must be the same thread that calls eval() into the model, this is called the "eval thread". The thread used to perform certain global operations such as saving and tracing must be done by a "main thread". In most cases the eval thread and main thread are the same thread (i.e. the user's top C++ testbench runs on a single thread), but this is not required.
The --trace-threads options can be used to produce trace dumps using multiple threads. If --trace-threads is set without --threads, then --trace-threads will imply --threads 1, i.e.: the support libraries will be thread safe.
With --trace-threads 0, trace dumps are produced on the main thread. This again gives the highest single thread performance.
With --trace-threads N, where N is at least 1, N additional threads will be created and managed by the trace files (e.g.: VerilatedVcdC or VerilatedFstC), to generate the trace dump. The main thread will be released to proceed with execution as soon as possible, though some blocking of the main thread is still necessary while capturing the trace. Different trace formats can utilize a various number of threads. See the --trace-threads option.
When running a multithreaded model, the default Linux task scheduler often works against the model, by assuming threads are short lived, and thus often schedules threads using multiple hyperthreads within the same physical core. For best performance use the "numactl" program to (when the threading count fits) select unique physical cores on the same socket. The same applies for --trace-threads as well.
As an example, if a model was Verilated with "--threads 4", we consult
egrep 'processor|physical id|core id' /proc/cpuinfo
To select cores 0, 1, 2, and 3 that are all located on the same socket (0) but different physical cores. (Also useful is "numactl --hardware", or "lscpu" but those doesn't show Hyperthreading cores.) Then we execute
numactl -m 0 -C 0,1,2,3 -- verilated_executable_name
This will limit memory to socket 0, and threads to cores 0, 1, 2, 3, (presumably on socket 0) optimizing performance. Of course this must be adjusted if you want another simulator using e.g. socket 1, or if you Verilated with a different number of threads. To see what CPUs are actually used, use --prof-threads.
$display/$stop/$finish are delayed until the end of an eval() call in order to maintain ordering between threads. This may result in additional tasks completing after the $stop or $finish.
If using --dpi, Verilator assumes pure DPI imports are thread safe, balancing performance versus safety. See --threads-dpi.
If using --savable, the save/restore classes are not multithreaded and must be called only by the eval thread.
If using --sc, the SystemC kernel is not thread safe, therefore the eval thread and main thread must be the same.
If using --trace, the tracing classes must be constructed and called from the main thread.
If using --vpi, since SystemVerilog VPI was not architected by IEEE to be multithreaded, Verilator requires all VPI calls are only made from the main thread.
In addition to the command line, warnings and other features may be controlled by configuration files, typically named with the .vlt extension. An example:
`verilator_config lint_off -rule WIDTH lint_off -rule CASEX -file "silly_vendor_code.v"
This disables WIDTH warnings globally, and CASEX for a specific file.
Configuration files are fed through the normal Verilog preprocessor prior to parsing, so `ifdefs, `defines, and comments may be used as if it were normal Verilog code.
Note that file or line-specific configuration only applies to files read after the configuration file. It is therefore recommended to pass the configuration file to Verilator as the first file.
The grammar of configuration commands is as follows:
With lint_off using '*' will override any lint_on directives in the source, i.e. the warning will still not be printed.
If the -rule is omitted, all lint warnings (see list in -Wno-lint) are enabled/disabled. This will override all later lint warning enables for the specified region.
If -match is set the linter warnings are matched against this (wildcard) string and are waived in case they match and iff rule and file (with wildcard) also match.
In previous versions -rule was named -msg. The latter is deprecated, but still works with a deprecation info, it may be removed in future versions.
For tracing_off, cells below any module in the files/ranges specified will also not be traced.
Same as /*verilator clock_enable*/, see "LANGUAGE EXTENSIONS" for more information and an example.
Same as /*verilator clocker*/, see "LANGUAGE EXTENSIONS" for more information.
Same as /*verilator coverage_block_off*/, see "LANGUAGE EXTENSIONS" for more information.
Same as /* verilator isolate_assignments */, see "LANGUAGE EXTENSIONS" for more information.
Same as /*verilator no_inline_task*/, see "LANGUAGE EXTENSIONS" for more information.
Same as /*verilator sformat*/, see "LANGUAGE EXTENSIONS" for more information.
Same as /*verilator split_var*/, see "LANGUAGE EXTENSIONS" for more information.
Verilator supports most Verilog 2001 language features. This includes signed numbers, "always @*", generate statements, multidimensional arrays, localparam, and C-style declarations inside port lists.
Verilator supports most Verilog 2005 language features. This includes the `begin_keywords and `end_keywords compiler directives, $clog2, and the uwire keyword.
Verilator supports ==? and !=? operators, ++ and -- in some contexts, $bits, $countbits, $countones, $error, $fatal, $info, $isunknown, $onehot, $onehot0, $unit, $warning, always_comb, always_ff, always_latch, bit, byte, chandle, const, do-while, enum, export, final, import, int, interface, logic, longint, modport, package, program, shortint, struct, time, typedef, union, var, void, priority case/if, and unique case/if.
It also supports .name and .* interconnection.
Verilator partially supports concurrent assert and cover statements; see the enclosed coverage tests for the syntax which is allowed.
Verilator implements a full SystemVerilog 2012 preprocessor, including function call-like preprocessor defines, default define arguments, `__FILE__, `__LINE__ and `undefineall.
Verilator currently has some support for SystemVerilog synthesis constructs. As SystemVerilog features enter common usage they are added; please file a bug if a feature you need is missing.
Verilator supports the 2017 "for" loop constructs, and several minor cleanups made in 1800-2017.
Verilator implements a very small subset of Verilog AMS (Verilog Analog and Mixed-Signal Extensions) with the subset corresponding to those VMS keywords with near equivalents in the Verilog 2005 or SystemVerilog 2009 languages.
AMS parsing is enabled with "--language VAMS" or "--language 1800+VAMS".
At present Verilator implements ceil, exp, floor, ln, log, pow, sqrt, string, and wreal.
With the --assert switch, Verilator reads any "//synopsys full_case" or "//synopsys parallel_case" directives. The same applies to any "//ambit synthesis", "//cadence" or "//pragma" directives of the same form.
When these synthesis directives are discovered, Verilator will either formally prove the directive to be true, or failing that, will insert the appropriate code to detect failing cases at simulation runtime and print an "Assertion failed" error message.
Verilator likewise also asserts any "unique" or "priority" SystemVerilog keywords on case statement, as well as "unique" on if statements. However, "priority if" is currently simply ignored.
The following additional constructs are the extensions Verilator supports on top of standard Verilog code. Using these features outside of comments or `ifdef's may break other tools.
String arguments will be put directly into the output C++ code. Expression arguments will have the code to evaluate the expression inserted. Thus to call a C++ function, $c("func(",a,")") will result in 'func(a)' in the output C++ code. For input arguments, rather than hard-coding variable names in the string $c("func(a)"), instead pass the variable as an expression $c("func(",a,")"). This will allow the call to work inside Verilog functions where the variable is flattened out, and also enable other optimizations.
If you will be reading or writing any Verilog variables inside the C++ functions, the Verilog signals must be declared with /*verilator public*/.
You may also append an arbitrary number to $c, generally the width of the output. [signal_32_bits = $c32("...");] This allows for compatibility with other simulators which require a differently named PLI function name for each different output width.
If you will be reading or writing any Verilog variables in the C++ functions, the Verilog signals must be declared with /*verilator public*/. See also the public task feature; writing an accessor may result in cleaner code.
reg enable_r /*verilator clock_enable*/;
wire gated_clk = clk & enable_r;
always_ff @ (posedge clk)
enable_r <= enable_early;
The clock_enable attribute will cause the clock gate to be ignored in the scheduling algorithm, sometimes required for correct clock behavior, and always improving performance. It's also a good idea to enable the IMPERFECTSCH warning, to ensure all clock enables are properly recognized.
Same as "clock_enable" in configuration files, see "CONFIGURATION FILES" for more information.
Same as "clocker" and "no_clocker" in configuration files, see "CONFIGURATION FILES" for more information.
Same as "coverage_block_off" in configuration files, see "CONFIGURATION FILES" for more information.
Same as "inline" in configuration files, see "CONFIGURATION FILES" for more information.
IE, with the following
reg splitme /* verilator isolate_assignments*/;
// Note the placement of the semicolon above
always @* begin
if (....) begin
splitme = ....;
other assignments
end
end
Verilator will internally split the block that assigns to "splitme" into two blocks:
It would then internally break it into (sort of):
// All assignments excluding those to splitme
always @* begin
if (....) begin
other assignments
end
end
// All assignments to splitme
always @* begin
if (....) begin
splitme = ....;
end
end
Same as "isolate_assignments" in configuration files, see "CONFIGURATION FILES" for more information.
// verilator lint_save
// verilator lint_off SOME_WARNING
... // code needing SOME_WARNING turned off
// verilator lint_restore
If SOME_WARNING was on before the lint_off, it will now be restored to on, and if it was off before the lint_off it will remain off.
Same as "no_inline" in configuration files, see "CONFIGURATION FILES" for more information.
Same as "no_inline" in configuration files, see "CONFIGURATION FILES" for more information.
parameter [2:0] PARAM /*verilator public*/ = 2'b0;
typedef enum logic [2:0] { ZERO = 3'b0 } pub_t /*verilator public*/;
Instead of using public variables, consider instead making a DPI or public function that accesses the variable. This is nicer as it provides an obvious entry point that is also compatible across simulators.
Same as "public" in configuration files, see "CONFIGURATION FILES" for more information.
Wide variables over 64 bits cannot be function returns, to avoid exposing complexities. However, wide variables can be input/outputs; they will be passed as references to an array of 32-bit numbers.
Generally, only the values of stored state (flops) should be written, as the model will NOT notice changes made to variables in these functions. (Same as when a signal is declared public.)
You may want to use DPI exports instead, as it's compatible with other simulators.
Same as "public" in configuration files, see "CONFIGURATION FILES" for more information.
Same as "public_flat" in configuration files, see "CONFIGURATION FILES" for more information.
Same as "public_flat_rd" in configuration files, see "CONFIGURATION FILES" for more information.
Same as "public_flat_rw" in configuration files, see "CONFIGURATION FILES" for more information.
Same as "public" in configuration files, see "CONFIGURATION FILES" for more information.
Same as "sc_bv" in configuration files, see "CONFIGURATION FILES" for more information.
Same as "sformat" in configuration files, see "CONFIGURATION FILES" for more information.
For example, Verilator will internally convert a variable with the metacomment such as:
logic [7:0] x [0:1] /*verilator split_var*/;
To:
logic [7:0] x__BRA__0__KET__ /*verilator split_var*/;
logic [7:0] x__BRA__1__KET__ /*verilator split_var*/;
Note that the generated packed variables retain the split_var metacomment because they may be split into further smaller pieces according to the access patterns.
This only supports unpacked arrays, packed arrays, and packed structs of integer types (reg, logic, bit, byte, int...); otherwise if a split was requested but cannot occur a SPLITVAR warning is issued. Splitting large arrays may slow down the Verilation speed, so use this only on variables that require it.
Same as "split_var" in configuration files, see "CONFIGURATION FILES" for more information.
There are some limitations and lack of features relative to the major closed-source simulators, by intent.
Verilator supports the Synthesis subset with other verification constructs being added over time. Verilator also simulates events as Synopsys's Design Compiler would; namely given a block of the form:
always @ (x) y = x & z;
This will recompute y when there is even a potential for change in x or a change in z, that is when the flops computing x or z evaluate (which is what Design Compiler will synthesize.) A compliant simulator would only calculate y if x changes. We recommend using always_comb to make the code run the same everywhere. Also avoid putting $displays in combo blocks, as they may print multiple times when not desired, even on compliant simulators as event ordering is not specified.
To avoid conflicts with C symbol naming, any character in a signal name that is not alphanumeric nor a single underscore will be replaced by __0hh where hh is the hex code of the character. To avoid conflicts with Verilator's internal symbols, any double underscore are replaced with ___05F (5F is the hex code of an underscore.)
Verilator only supports "bind" to a target module name, not an instance path.
Verilator class support is limited but in active development. Verilator supports members, and methods. Verilator does not support class static members, class extend, or class parameters.
Verilator supports dotted references to variables, functions and tasks in different modules. The portion before the dot must have a constant value; for example a[2].b is acceptable, while a[x].b is generally not.
References into generated and arrayed instances use the instance names specified in the Verilog standard; arrayed instances are named {cellName}[{instanceNumber}] in Verilog, which becomes {cellname}__BRA__{instanceNumber}__KET__ inside the generated C++ code.
Verilator is optimized for edge sensitive (flop based) designs. It will attempt to do the correct thing for latches, but most performance optimizations will be disabled around the latch.
Presently Verilator only supports packed structs and packed unions. Rand and randc tags on members are simply ignored. All structures and unions are represented as a single vector, which means that generating one member of a structure from blocking, and another from non-blocking assignments is unsupported.
All delays (#) are ignored, as they are in synthesis.
Verilator is mostly a two state simulator, not a four state simulator. However, it has two features which uncover most initialization bugs (including many that a four state simulator will miss.)
Identity comparisons (=== or !==) are converted to standard ==/!= when neither side is a constant. This may make the expression yield a different result compared to a four state simulator. An === comparison to X will always be false, so that Verilog code which checks for uninitialized logic will not fire.
Assigning X to a variable will actually assign a constant value as determined by the --x-assign switch. This allows runtime randomization, thus if the value is actually used, the random value should cause downstream errors. Integers also get randomized, even though the Verilog 2001 specification says they initialize to zero. Note however that randomization happens at initialization time and hence during a single simulation run, the same constant (but random) value will be used every time the assignment is executed.
All variables, depending on --x-initial setting, are typically randomly initialized using a function. By running several random simulation runs you can determine that reset is working correctly. On the first run, have the function initialize variables to zero. On the second, have it initialize variables to one. On the third and following runs have it initialize them randomly. If the results match, reset works. (Note this is what the hardware will really do.) In practice, just setting all variables to one at startup finds most problems (since typically control signals are active-high).
--x-assign applies to variables explicitly initialized or assigned an X. Uninitialized clocks are initialized to zero, while all other state holding variables are initialized to a random value. Event driven simulators will generally trigger an edge on a transition from X to 1 ("posedge") or X to 0 ("negedge"). However, by default, since clocks are initialized to zero, Verilator will not trigger an initial negedge. Some code (particularly for reset) may rely on X->0 triggering an edge. The --x-initial-edge switch enables this behavior. Comparing runs with and without this switch will find such problems.
Verilator converts some simple tristate structures into two state. Pullup, pulldown, bufif0, bufif1, notif0, notif1, pmos, nmos, tri0 and tri1 are also supported. Simple comparisons with === 1'bz are also supported.
An assignment of the form:
inout driver;
wire driver = (enable) ? output_value : 1'bz;
Will be converted to
input driver; // Value being driven in from "external" drivers
output driver__en; // True if driven from this module
output driver__out; // Value being driven from this module
External logic will be needed to combine these signals with any external drivers.
Tristate drivers are not supported inside functions and tasks; an inout there will be considered a two state variable that is read and written instead of a four state variable.
All functions and tasks will be inlined (will not become functions in C.) The only support provided is for simple statements in tasks (which may affect global variables).
Recursive functions and tasks are not supported. All inputs and outputs are automatic, as if they had the Verilog 2001 "automatic" keyword prepended. (If you don't know what this means, Verilator will do what you probably expect -- what C does. The default behavior of Verilog is different.)
Verilator attempts to deal with generated and gated clocks correctly, however some cases cause problems in the scheduling algorithm which is optimized for performance. The safest option is to have all clocks as primary inputs to the model, or wires directly attached to primary inputs. For proper behavior clock enables may also need the /*verilator clock_enable*/ attribute.
The 2-state gate primitives (and, buf, nand, nor, not, or, xnor, xor) are directly converted to behavioral equivalents. The 3-state and MOS gate primitives are not supported. Tables are not supported.
All specify blocks and timing checks are ignored. All min:typ:max delays use the typical value.
When initializing a large array, you need to use non-delayed assignments. Verilator will tell you when this needs to be fixed; see the BLKLOOPINIT error for more information.
Writing a memory element that is outside the bounds specified for the array may cause a different memory element inside the array to be written instead. For power-of-2 sized arrays, Verilator will give a width warning and the address. For non-power-of-2-sizes arrays, index 0 will be written.
Reading a memory element that is outside the bounds specified for the array will give a width warning and wrap around the power-of-2 size. For non-power-of-2 sizes, it will return a unspecified constant of the appropriate width.
Verilator is beginning to add support for assertions. Verilator currently only converts assertions to simple "if (...) error" statements, and coverage statements to increment the line counters described in the coverage section.
Verilator does not support SEREs yet. All assertion and coverage statements must be simple expressions that complete in one cycle.
Open source simulators like Verilator are unable to use encrypted RTL (i.e. IEEE P1735). Talk to your IP vendor about delivering IP blocks via Verilator's --protect-lib feature.
This section describes specific limitations for each language keyword.
$dumpvars and $dumpports module identifier is ignored; the traced instances will always start at the top of the design. The levels argument is also ignored, use tracing_on/tracing_off pragmas instead.
$dumpportson/$dumpportsoff/$dumpportsall/$dumpportslimit filename argument is ignored, only a single trace file may be active at once.
$dumpall/$dumpportsall, $dumpon/$dumpportson, $dumpoff/$dumpportsoff, and $dumplimit/$dumpportlimit are currently ignored.
Verilated::commandArgs(argc, argv);
to register the command line before calling $test$plusargs or $value$plusargs.
Warnings may be disabled in three ways. First, when the warning is printed it will include a warning code. Simply surround the offending line with a lint_off/lint_on pair:
// verilator lint_off UNSIGNED
if (`DEF_THAT_IS_EQ_ZERO <= 3) $stop;
// verilator lint_on UNSIGNED
Second, warnings may be disabled using a configuration file with a lint_off command. This is useful when a script is suppressing warnings and the Verilog source should not be changed.
Warnings may also be globally disabled by invoking Verilator with the "-Wno-warning" switch. This should be avoided, as it removes all checking across the designs, and prevents other users from compiling your code without knowing the magic set of disables needed to successfully compile your design.
Warnings and errors printed by Verilator always match this regular expression:
%(Error|Warning)(-[A-Z0-9_]+)?: ((\S+):(\d+):((\d+):)? )?.*
Errors and warning start with a percent sign (historical heritage from Digital Equipment Corporation). Some errors or warning have a code attached, with meanings described below. Some errors also have a filename, line number and optional column number (starting at column 1 to match GCC).
Following the error message, Verilator will typically show the user's source code corresponding to the error, prefixed by the line number and a " | ". Following this is typically an arrow and ~ pointing at the error on the source line directly above.
always_comb begin
a = b;
b = 1;
end
Ignoring this warning will only suppress the lint check, it will simulate correctly.
input a;
assign a = 1'b1;
Ignoring this warning will only suppress the lint check, it will simulate correctly.
a <= #100 b;
assign #100 a = b;
Ignoring this warning may make Verilator simulations differ from other simulators, however at one point this was a common style so disabled by default as a code style warning.
always @ (posedge clk) foo[0] <= ...
always @* foo[1] = ...
Simply use a different register for the flop:
always @ (posedge clk) foo_flopped[0] <= ...
always @* foo[0] = foo_flopped[0];
always @* foo[1] = ...
This is not illegal in SystemVerilog, but a violation of good coding practice. Verilator reports this as an error, because ignoring this warning may make Verilator simulations differ from other simulators.
It is generally safe to disable this error (with a "// verilator lint_off BLKANDNBLK" metacomment or the -Wno-BLKANDNBLK option) when one of the assignments is inside a public task, or when the blocking and non-blocking assignments have non-overlapping bits and structure members.
always @ (posedge clk) foo = ...
Disabled by default as this is a code style warning; it will simulate correctly.
always @ (posedge clk)
if (~reset_l) begin
for (i=0; i<`ARRAY_SIZE; i++) begin
array[i] = 0; // Non-delayed for verilator
end
This message is only seen on large or complicated loops because Verilator generally unrolls small loops. You may want to try increasing --unroll-count (and occasionally --unroll-stmts) which will raise the small loop bar to avoid this error.
Ignoring this warning may make Verilator simulations differ from other simulators.
Ignoring this warning will only suppress the lint check, it will simulate correctly.
Ignoring this warning will only suppress the lint check, it will simulate correctly.
Ignoring this warning will only suppress the lint check, it will simulate correctly.
Ignoring this warning will only suppress the lint check, it will simulate correctly.
Ignoring this warning will only suppress the lint check, it will simulate correctly.
The warning can be disabled without affecting the simulation result. But it is recommended to check the warning as this may degrade the performance of the Verilated model.
Ignoring this warning will only suppress the lint check, it will simulate correctly.
Ignoring this warning will only suppress the lint check, it will simulate correctly.
Ignoring this warning may make Verilator simulations differ from other simulators.
This error is only reported when "--language 1364-1995", "--language 1364-2001", or "--language 1364-2005" is used.
Ignoring this error will only suppress the lint check, it will simulate correctly.
Disabled by default as this is a code style warning; it will simulate correctly.
Disabled by default as this is a code style warning; it will simulate correctly.
The solution is to break the loop, as described for UNOPTFLAT.
Verilator sometimes has to evaluate combinatorial logic multiple times, usually around code where a UNOPTFLAT warning was issued, but disabled. For example:
always_comb b = ~a;
always_comb a = b
This code will toggle forever, and thus to prevent an infinite loop, the executable will give the didn't converge error.
To debug this, first review any UNOPTFLAT warnings that were ignored. Though typically it is safe to ignore UNOPTFLAT (at a performance cost), at the time of issuing a UNOPTFLAT Verilator did not know if the logic would eventually converge and assumed it would.
Next, run Verilator with --prof-cfuncs. Run make on the generated files with "CPP_FLAGS=-DVL_DEBUG", to allow enabling simulation runtime debug messages. Rerun the test. Now just before the convergence error you should see additional output similar to this:
CHANGE: filename.v:1: b
CHANGE: filename.v:2: a
This means that signal b and signal a keep changing, inspect the code that modifies these signals. Note if many signals are getting printed then most likely all of them are oscillating. It may also be that e.g. "a" may be oscillating, then "a" feeds signal "c" which then is also reported as oscillating.
Finally, rare, more difficult cases can be debugged like a "C" program; either enter GDB and use its tracing facilities, or edit the generated C++ code to add appropriate prints to see what is going on.
Ignoring this warning will only suppress the lint check, it will simulate correctly.
Disabled by default as this is a code style warning; it will simulate correctly.
This warning is required by IEEE. The portable way to suppress this warning (in SystemVerilog) is to use a void cast, e.g.
void'(function_being_called_as_task());
Ignoring this warning will only suppress the lint check, it will simulate correctly.
Ignoring this warning will only suppress the lint check, it will simulate correctly.
Disabled by default as this is a code style warning; it will simulate correctly.
Ignoring this warning may make Verilator simulations differ from other simulators.
Disabled by default as this is a code style warning; it will simulate correctly.
This might be unintended behavior if the loop body contains statements that in other simulators would make time pass, which Verilator is ignoring due to e.g. STMTDLY warnings being disabled.
Ignoring this warning will only suppress the lint check, it will simulate correctly (i.e. hang due to the infinite loop).
Ignoring this warning may make Verilator simulations differ from other simulators.
Ignoring this warning will only suppress the lint check, it will simulate correctly.
Also warns that a cell is declared with little endian range (i.e. [0:7] or [7]) and is connected to a N-wide signal. Based on IEEE the bits will likely be backwards from what you expect (i.e. cell [0] will connect to signal bit [N-1] not bit [0]).
Ignoring this warning will only suppress the lint check, it will simulate correctly.
Ignoring this warning will cause the more recent module definition to be discarded.
Ignoring this warning will only slow simulations, it will simulate correctly.
1. A single module is intended to be the top. This warning then occurs because some low level cell is being read in, but is not really needed as part of the design. The best solution for this situation is to ensure that only the top module is put on the command line without any flags, and all remaining library files are read in as libraries with -v, or are automatically resolved by having filenames that match the module names.
2. A single module is intended to be the top, the name of it is known, and all other modules should be ignored if not part of the design. The best solution is to use the --top-module option to specify the top module's name. All other modules that are not part of the design will be for the most part ignored (they must be clean in syntax and their contents will be removed as part of the Verilog module elaboration process.)
3. Multiple modules are intended to be design tops, e.g. when linting a library file. As multiple modules are desired, disable the MULTITOP warning. All input/outputs will go uniquely to each module, with any conflicting and identical signal names being made unique by adding a prefix based on the top module name followed by __02E (a Verilator-encoded ASCII ".'). This renaming is done even if the two modules' signals seem identical, e.g. multiple modules with a "clk" input.
Disabled by default as this is a code style warning; it will simulate correctly.
Ignoring this warning will only suppress the lint check, it will simulate correctly.
Disabled by default as this is a code style warning; it will simulate correctly.
`define MACRO def1
//...
`define MACRO otherdef
The best solution is to use a different name for the second macro. If this is not possible, add a undef to indicate the code is overriding the value:
`define MACRO def1
//...
`undef MACRO
`define MACRO otherdef
wire vec[6:0];
initial out = vec[7]; // There is no 7
Verilator will assume zero for this value, instead of X. Note that in some cases this warning may be false, when a condition upstream or downstream of the access means the access out of bounds will never execute or be used.
wire vec[6:0];
initial begin
seven = 7;
...
if (seven != 7) out = vec[seven]; // Never will use vec[7]
Ignoring this warning may make Verilator simulations differ from other simulators, if the increased precision of real affects your model or DPI calls.
* The datatype of the variable is not supported for splitting. (e.g. is a real).
* The access pattern of the variable can not be determined statically. (e.g. is accessed as a memory).
* The index of the array exceeds the array size.
* The variable is accessed from outside using dotted reference. (e.g. top.instance0.variable0 = 1).
* The variable is not declared in a module, but in a package or an interface.
* The variable is a parameter, localparam, genvar, or queue.
* The variable is tristate or bidirectional. (e.g. inout or ref).
#100 $finish;
Ignoring this warning may make Verilator simulations differ from other simulators.
Disabled by default as this is a code style warning; it will simulate correctly.
task foo; output sig; ... endtask
always @* begin
foo(bus_we_select_from[2]); // Will get TASKNSVAR error
end
Change this to:
reg foo_temp_out;
always @* begin
foo(foo_temp_out);
bus_we_select_from[2] = foo_temp_out;
end
Verilator doesn't do this conversion for you, as some more complicated cases would result in simulator mismatches.
Ignoring this warning will only slow simulations, it will simulate correctly.
`include "timescale.vh"
Then in that file set the timescale.
This is an error due to IEEE specifications, but it may be disabled similar to warnings. Ignoring this error may result in a module having an unexpected timescale.
Disabled by default as this is a code style warning; it will simulate correctly.
A less obvious case of this is when a module instantiates two submodules. Inside submodule A, signal I is input and signal O is output. Likewise in submodule B, signal O is an input and I is an output. A loop exists and a UNOPT warning will result if AI & AO both come from and go to combinatorial blocks in both submodules, even if they are unrelated always blocks. This affects performance because Verilator would have to evaluate each submodule multiple times to stabilize the signals crossing between the modules.
Ignoring this warning will only slow simulations, it will simulate correctly.
Unlike the UNOPT warning, this occurs after flattening the netlist, and indicates a more basic problem, as the less obvious case described under UNOPT does not apply.
Often UNOPTFLAT is caused by logic that isn't truly circular as viewed by synthesis which analyzes interconnection per-bit, but is circular to simulation which analyzes per-bus:
wire [2:0] x = {x[1:0], shift_in};
This statement needs to be evaluated multiple times, as a change in "shift_in" requires "x" to be computed 3 times before it becomes stable. This is because a change in "x" requires "x" itself to change value, which causes the warning.
For significantly better performance, split this into 2 separate signals:
wire [2:0] xout = {x[1:0], shift_in};
and change all receiving logic to instead receive "xout". Alternatively, change it to
wire [2:0] x = {xin[1:0], shift_in};
and change all driving logic to instead drive "xin".
With this change this assignment needs to be evaluated only once. These sort of changes may also speed up your traditional event driven simulator, as it will result in fewer events per cycle.
The most complicated UNOPTFLAT path we've seen was due to low bits of a bus being generated from an always statement that consumed high bits of the same bus processed by another series of always blocks. The fix is the same; split it into two separate signals generated from each block.
Another way to resolve this warning is to add a "split_var" metacomment described above. This will cause the variable to be split internally, potentially resolving the conflict. If you run with --report-unoptflat Verilator will suggest possible candidates for "split_var".
The UNOPTFLAT warning may also be due to clock enables, identified from the reported path going through a clock gating cell. To fix these, use the clock_enable meta comment described above.
The UNOPTFLAT warning may also occur where outputs from a block of logic are independent, but occur in the same always block. To fix this, use the isolate_assignments meta comment described above.
To assist in resolving UNOPTFLAT, the option "--report-unoptflat" can be used, which will provide suggestions for variables that can be split up, and a graph of all the nodes connected in the loop. See the Arguments section for more details.
Ignoring this warning will only slow simulations, it will simulate correctly.
One workaround is to request fewer threads with "--threads".
Another possible workaround is to allow more MTasks in the simulation runtime, by increasing the value of --threads-max-mtasks. More MTasks will result in more communication and synchronization overhead at simulation runtime; the scheduler attempts to minimize the number of MTasks for this reason.
Ignoring this warning will only slow simulations, it will simulate correctly.
Ignoring this warning will make Verilator treat the structure as packed, which may make Verilator simulations differ from other simulators. This downgrading may also result what would normally be a legal unpacked struct/array inside an unpacked struct/array becoming an illegal unpacked struct/array inside a packed struct/array.
Ignoring this warning will only suppress the lint check, it will simulate correctly.
This error may be ignored with --bbox-unsup, however this will make the design simulate incorrectly; see the details under --bbox-unsup.
Disabled by default as this is a code style warning; it will simulate correctly.
A recommended style for unused nets is to put at the bottom of a file code similar to the following:
wire _unused_ok = &{1'b0,
sig_not_used_a,
sig_not_used_yet_b, // To be fixed
1'b0};
The reduction AND and constant zeros mean the net will always be zero, so won't use simulation runtime. The redundant leading and trailing zeros avoid syntax errors if there are no signals between them. The magic name "unused" (-unused-regexp) is recognized by Verilator and suppresses warnings; if using other lint tools, either teach it to the tool to ignore signals with "unused" in the name, or put the appropriate lint_off around the wire. Having unused signals in one place makes it easy to find what is unused, and reduces the number of lint_off pragmas, reducing bugs.
Disabled by default as this is a code style warning; it will simulate correctly.
If it's a constant 0 that's 32 bits or less, simply leave it unwidthed. Verilator considers zero to be any width needed.
Concatenate leading zeros when doing arithmetic. In the statement
wire [5:0] plus_one = from[5:0] + 6'd1 + carry[0];
The best fix, which clarifies intent and will also make all tools happy is:
wire [5:0] plus_one = from[5:0] + 6'd1 + {5'd0, carry[0]};
Ignoring this warning will only suppress the lint check, it will simulate correctly.
wire [63:0] concat = {1, 2};
An example where this is technically legal (though still bad form) is:
parameter PAR = 1;
wire [63:0] concat = {PAR, PAR};
The correct fix is to either size the 1 ("32'h1"), or add the width to the parameter definition ("parameter [31:0]"), or add the width to the parameter usage ("{PAR[31:0],PAR[31:0]}".
The following describes the less obvious errors:
The following deprecated items are scheduled for future removal:
Most simulators have to be compliant with the complete IEEE 1364 (Verilog) and IEEE 1800 (SystemVerilog) standards, meaning they have to be event driven. This prevents them from being able to reorder blocks and make netlist-style optimizations, which are where most of the gains come from.
You should not be scared by non-compliance. Your synthesis tool isn't compliant with the whole standard to start with, so your simulator need not be either. Verilator is closer to the synthesis interpretation, so this is a good thing for getting working silicon.
You also have the option of using the Perl Artistic License, which again does not require you to release your Verilog or generated code, and also allows you to modify Verilator for internal use without distributing the modified version. But please contribute back to the community!
One limit is that you cannot under either license release a commercial Verilog simulation product incorporating Verilator without making the source code available.
As is standard with Open Source, contributions back to Verilator will be placed under the Verilator copyright and LGPL/Artistic license. Small test cases will be released into the public domain so they can be used anywhere, and large tests under the LGPL/Artistic, unless requested otherwise.
A. Add the --trace switch to Verilator, and in your top level C code, call Verilated::traceEverOn(true). Then you may use $dumpfile and $dumpvars to enable traces, same as with any Verilog simulator. See "examples/make_tracing_c".
B. Or, for finer-grained control, or C++ files with multiple Verilated modules you may also create the trace purely from C++. Create a VerilatedVcdC object, and in your main loop call "trace_object->dump(time)" every time step, and finally call "trace_object->close()". You also need to compile verilated_vcd_c.cpp and add it to your link, preferably by adding the dependencies in $(VK_GLOBAL_OBJS) to your Makefile's link rule. This is done for you if using the Verilator --exe flag. Note you can also call ->trace on multiple Verilated objects with the same trace file if you want all data to land in the same output file.
#include "verilated_vcd_c.h"
...
int main(int argc, char** argv, char** env) {
...
Verilated::traceEverOn(true);
VerilatedVcdC* tfp = new VerilatedVcdC;
topp->trace(tfp, 99); // Trace 99 levels of hierarchy
tfp->open("obj_dir/t_trace_ena_cc/simx.vcd");
...
while (sc_time_stamp() < sim_time && !Verilated::gotFinish()) {
main_time += #;
tfp->dump(main_time);
}
tfp->close();
}
B. Or, you may create a trace purely from SystemC, which may trace all Verilated designs in the SystemC model. Create a VerilatedVcdSc object as you would create a normal SystemC trace file. For an example, see the call to VerilatedVcdSc in the examples/make_tracing_sc/sc_main.cpp file of the distribution, and below.
Alternatively you may use the C++ trace mechanism described in the previous question, note the timescale and timeprecision will be inherited from your SystemC settings.
You also need to compile verilated_vcd_sc.cpp and verilated_vcd_c.cpp and add them to your link, preferably by adding the dependencies in $(VK_GLOBAL_OBJS) to your Makefile's link rule. This is done for you if using the Verilator --exe flag.
Note you can also call ->trace on multiple Verilated objects with the same trace file if you want all data to land in the same output file.
When using SystemC 2.3, the SystemC library must have been built with the experimental simulation phase callback based tracing disabled. This is disabled by default when building SystemC with its configure based build system, but when building SystemC with CMake, you must pass -DENABLE_PHASE_CALLBACKS_TRACING=OFF to disable this feature.
#include "verilated_vcd_sc.h"
...
int main(int argc, char** argv, char** env) {
...
Verilated::traceEverOn(true);
VerilatedVcdSc* tfp = new VerilatedVcdSc;
topp->trace(tfp, 99); // Trace 99 levels of hierarchy
tfp->open("obj_dir/t_trace_ena_cc/simx.vcd");
...
sc_start(1);
...
tfp->close();
}
#include "verilated_fst_c.h"
VerilatedFstC* tfp = new VerilatedFstC;
Note that currently supporting both FST and VCD in a single simulation is impossible, but such requirement should be rare. You can however ifdef around the trace format in your C++ main loop, and select VCD or FST at build time, should you require.
Next, add /*verilator tracing_off*/ to any very low level modules you never want to trace (such as perhaps library cells). Finally, use the --trace-depth option to limit the depth of tracing, for example --trace-depth 1 to see only the top level signals.
Also be sure you write your trace files to a local solid-state drive, instead of to a network drive. Network drives are generally far slower.
You can also consider using FST tracing instead of VCD. FST dumps are a fraction of the size of the equivalent VCD. FST tracing can be slower than VCD tracing, but it might be the only option if the VCD file size is prohibitively large.
First, run verilator with the --coverage option. If you are using your own makefile, compile the model with the GCC flag -DVM_COVERAGE (if using the makefile provided by Verilator, it will do this for you).
At the end of your test, call VerilatedCov::write passing the name of the coverage data file (typically "logs/coverage.dat").
Run each of your tests in different directories. Each test will create a logs/coverage.dat file.
After running all of your tests, execute the verilator_coverage tool. The verilator_coverage tool reads the logs/coverage.dat file(s), and creates an annotated source code listing showing code coverage details.
For an example, after running 'make test' in the Verilator distribution, see the examples/make_tracing_c/logs directory. Grep for lines starting with '%' to see what lines Verilator believes need more coverage.
Info files can be written by verilator_coverage for import to "lcov". This enables use of "genhtml" for HTML reports and importing reports to sites such as <https://codecov.io>.
`ifndef VERILATOR
Something_Verilator_Dislikes;
`endif
Most synthesis tools similarly define SYNTHESIS for you.
`begin_keywords "1364-2001"
integer bit; initial bit = 1;
`end_keywords
If you want the whole design to be parsed as Verilog 2001, please see the "--default-language" option.
If you want something more complex, since Verilator emits standard C++ code, you can simply write your own C++ routines that can access and modify signal values without needing any PLI interface code, and call it with $c("{any_c++_statement}").
Verilator emits any infrequently executed "cold" routines into separate __Slow.cpp files. This can accelerate compilation as optimization can be disabled on these routines. See the OPT_FAST and OPT_SLOW make variables and the BENCHMARKING & OPTIMIZATION section of the manual.
Use a recent compiler. Newer compilers tend to be faster.
Compile in parallel on many machines and use caching; see the web for the ccache, distcc and icecream packages. ccache will skip GCC runs between identical source builds, even across different users. If ccache was installed when Verilator was built it is used, or see OBJCACHE environment variable to override this. Also see the --output-split option.
To reduce the compile time of classes that use a Verilated module (e.g. a top CPP file) you may wish to add /*verilator no_inline_module*/ to your top level module. This will decrease the amount of code in the model's Verilated class, improving compile times of any instantiating top level C++ code, at a relatively small cost of execution performance.
If you really want raw access to the signals, declare the signals you will be accessing with a /*verilator public*/ comment before the closing semicolon. Then scope into the C++ class to read the value of the signal, as you would any other member variable.
Signals are the smallest of 8-bit unsigned chars (equivalent to uint8_t), 16-bit unsigned shorts (uint16_t), 32-bit unsigned longs (uint32_t), or 64-bit unsigned long longs (uint64_t) that fits the width of the signal. Generally, you can use just uint32_t's for 1 to 32 bits, or vluint64_t for 1 to 64 bits, and the compiler will properly up-convert smaller entities. Note even signed ports are declared as unsigned; you must sign extend yourself to the appropriate signal width.
Signals wider than 64 bits are stored as an array of 32-bit uint32_t's. Thus to read bits 31:0, access signal[0], and for bits 63:32, access signal[1]. Unused bits (for example bit numbers 65-96 of a 65-bit vector) will always be zero. If you change the value you must make sure to pack zeros in the unused bits or core-dumps may result, because Verilator strips array bound checks where it believes them to be unnecessary to improve performance.
In the SYSTEMC example above, if you had in our.v:
input clk /*verilator public*/;
// Note the placement of the semicolon above
From the sc_main.cpp file, you'd then:
#include "Vour.h"
#include "Vour_our.h"
cout << "clock is " << top->our->clk << endl;
In this example, clk is a bool you can read or set as any other variable. The value of normal signals may be set, though clocks shouldn't be changed by your code or you'll get strange results.
A module with only SystemC cells below must be SystemC.
A module with a mix of Verilog and SystemC cells below must be SystemC. (As Verilator cannot connect to lower-level SystemC cells.)
A module with only Verilog cells below can be either, but for best performance should be Verilog. (The exception is if you have a design that is instantiated many times; in this case Verilating one of the lower modules and instantiating that Verilated cells multiple times into a SystemC module *may* be faster.)
First, check the "LANGUAGE LIMITATIONS" section.
Next, try the --debug switch. This will enable additional internal assertions, and may help identify the problem.
Finally, reduce your code to the smallest possible routine that exhibits the bug. Even better, create a test in the test_regress/t directory, as follows:
cd test_regress
cp -p t/t_EXAMPLE.pl t/t_BUG.pl
cp -p t/t_EXAMPLE.v t/t_BUG.v
There are many hints on how to write a good test in the driver.pl documentation which can be seen by running:
cd $VERILATOR_ROOT # Need the original distribution kit
test_regress/driver.pl --help
Edit t/t_BUG.pl to suit your example; you can do anything you want in the Verilog code there; just make sure it retains the single clk input and no outputs. Now, the following should fail:
cd $VERILATOR_ROOT # Need the original distribution kit
cd test_regress
t/t_BUG.pl # Run on Verilator
t/t_BUG.pl --debug # Run on Verilator, passing --debug to Verilator
t/t_BUG.pl --vcs # Run on VCS simulator
t/t_BUG.pl --nc|--iv|--ghdl # Likewise on other simulators
The test driver accepts a number of options, many of which mirror the main Verilator option. For example the previous test could have been run with debugging enabled. The full set of test options can be seen by running driver.pl --help as shown above.
Finally, report the bug using the bug tracker at <https://verilator.org/issues>. The bug will become publicly visible; if this is unacceptable, mail the bug report to "wsnyder@wsnyder.org".
Verilator was conceived in 1994 by Paul Wasson at the Core Logic Group at Digital Equipment Corporation. The Verilog code that was converted to C was then merged with a C based CPU model of the Alpha processor and simulated in a C based environment called CCLI.
In 1995 Verilator started being used also for Multimedia and Network Processor development inside Digital. Duane Galbi took over active development of Verilator, and added several performance enhancements. CCLI was still being used as the shell.
In 1998, through the efforts of existing DECies, mainly Duane Galbi, Digital graciously agreed to release the source code. (Subject to the code not being resold, which is compatible with the GNU Public License.)
In 2001, Wilson Snyder took the kit, and added a SystemC mode, and called it Verilator2. This was the first packaged public release.
In 2002, Wilson Snyder created Verilator 3.000 by rewriting Verilator from scratch in C++. This added many optimizations, yielding about a 2-5x performance gain.
In 2009, major SystemVerilog and DPI language support was added.
In 2018, Verilator 4.000 was released with multithreaded support.
Currently, various language features and performance enhancements are added as the need arises. Verilator is now about 3x faster than in 2002, and is faster than most (if not every) other simulator.
When possible, please instead report bugs to <https://verilator.org/issues>.
Wilson Snyder <wsnyder@wsnyder.org>
Major concepts by Paul Wasson, Duane Galbi, John Coiner and Jie Xu.
Many people have provided ideas and other assistance with Verilator.
Verilator is receiving major development support from the CHIPS Alliance.
Previous major corporate sponsors of Verilator, by providing significant contributions of time or funds included include Atmel Corporation, Cavium Inc., Compaq Corporation, Digital Equipment Corporation, Embecosm Ltd., Hicamp Systems, Intel Corporation, Mindspeed Technologies Inc., MicroTune Inc., picoChip Designs Ltd., Sun Microsystems Inc., Nauticus Networks Inc., and SiCortex Inc.
The people who have contributed major functionality are Byron Bradley, Jeremy Bennett, Lane Brooks, John Coiner, Duane Galbi, Geza Lore, Todd Strader, Stefan Wallentowitz, Paul Wasson, Jie Xu, and Wilson Snyder. Major testers included Jeff Dutton, Jonathon Donaldson, Ralf Karge, David Hewson, Iztok Jeras, Wim Michiels, Alex Solomatnikov, Sebastien Van Cauwenberghe, Gene Weber, and Clifford Wolf.
Some of the people who have provided ideas, and feedback for Verilator include: David Addison, Tariq B. Ahmad, Nikana Anastasiadis, Hans Van Antwerpen, Vasu Arasanipalai, Jens Arm, Sharad Bagri, Matthew Ballance, Andrew Bardsley, Matthew Barr, Geoff Barrett, Julius Baxter, Jeremy Bennett, Michael Berman, Victor Besyakov, Moinak Bhattacharyya, David Binderman, Piotr Binkowski, Johan Bjork, David Black, Tymoteusz Blazejczyk, Daniel Bone, Gregg Bouchard, Christopher Boumenot, Nick Bowler, Byron Bradley, Bryan Brady, Maarten De Braekeleer, Charlie Brej, J Briquet, Lane Brooks, John Brownlee, Jeff Bush, Lawrence Butcher, Tony Bybell, Ted Campbell, Chris Candler, Lauren Carlson, Donal Casey, Sebastien Van Cauwenberghe, Alex Chadwick, Terry Chen, Yi-Chung Chen, Enzo Chi, Robert A. Clark, Allan Cochrane, John Coiner, Gianfranco Costamagna, Sean Cross, George Cuan, Joe DErrico, Lukasz Dalek, Laurens van Dam, Gunter Dannoritzer, Ashutosh Das, Bernard Deadman, John Demme, Mike Denio, John Deroo, Philip Derrick, John Dickol, Ruben Diez, Danny Ding, Jacko Dirks, Ivan Djordjevic, Jonathon Donaldson, Leendert van Doorn, Sebastian Dressler, Alex Duller, Jeff Dutton, Tomas Dzetkulic, Usuario Eda, Charles Eddleston, Chandan Egbert, Joe Eiler, Ahmed El-Mahmoudy, Trevor Elbourne, Mats Engstrom, Robert Farrell, Eugen Fekete, Fabrizio Ferrandi, Udi Finkelstein, Brian Flachs, Andrea Foletto, Bob Fredieu, Duane Galbi, Benjamin Gartner, Christian Gelinek, Peter Gerst, Glen Gibb, Michael Gielda, Shankar Giri, Dan Gisselquist, Petr Gladkikh, Sam Gladstone, Amir Gonnen, Chitlesh Goorah, Kai Gossner, Sergi Granell, Al Grant, Alexander Grobman, Xuan Guo, Driss Hafdi, Neil Hamilton, James Hanlon, Oyvind Harboe, Jannis Harder, Junji Hashimoto, Thomas Hawkins, Mitch Hayenga, Robert Henry, Stephen Henry, David Hewson, Jamey Hicks, Joel Holdsworth, Andrew Holme, Hiroki Honda, Alex Hornung, David Horton, Peter Horvath, Jae Hossell, Alan Hunter, James Hutchinson, Jamie Iles, Ben Jackson, Shareef Jalloq, Krzysztof Jankowski, HyungKi Jeong, Iztok Jeras, James Johnson, Christophe Joly, Franck Jullien, James Jung, Mike Kagen, Arthur Kahlich, Kaalia Kahn, Guy-Armand Kamendje, Vasu Kandadi, Kanad Kanhere, Patricio Kaplan, Pieter Kapsenberg, Ralf Karge, Dan Katz, Sol Katzman, Ian Kennedy, Jonathan Kimmitt, Olof Kindgren, Kevin Kiningham, Dan Kirkham, Sobhan Klnv, Gernot Koch, Soon Koh, Nathan Kohagen, Steve Kolecki, Brett Koonce, Will Korteland, Wojciech Koszek, Varun Koyyalagunta, David Kravitz, Roland Kruse, Sergey Kvachonok, Charles Eric LaForest, Ed Lander, Steve Lang, Stephane Laurent, Walter Lavino, Christian Leber, Larry Lee, Igor Lesik, John Li, Eivind Liland, Yu Sheng Lin, Charlie Lind, Andrew Ling, Jiuyang Liu, Paul Liu, Derek Lockhart, Jake Longo, Geza Lore, Arthur Low, Stefan Ludwig, Dan Lussier, Fred Ma, Duraid Madina, Affe Mao, Julien Margetts, Mark Marshall, Alfonso Martinez, Yves Mathieu, Patrick Maupin, Jason McMullan, Elliot Mednick, Wim Michiels, Miodrag Milanovic, Wai Sum Mong, Peter Monsson, Sean Moore, Dennis Muhlestein, John Murphy, Matt Myers, Nathan Myers, Richard Myers, Dimitris Nalbantis, Peter Nelson, Bob Newgard, Cong Van Nguyen, Paul Nitza, Yossi Nivin, Pete Nixon, Lisa Noack, Mark Nodine, Kuba Ober, Andreas Olofsson, Aleksander Osman, James Pallister, Vassilis Papaefstathiou, Brad Parker, Dan Petrisko, Maciej Piechotka, David Pierce, Dominic Plunkett, David Poole, Mike Popoloski, Roman Popov, Rich Porter, Niranjan Prabhu, Usha Priyadharshini, Mark Jackson Pulver, Prateek Puri, Marshal Qiao, Danilo Ramos, Chris Randall, Anton Rapp, Josh Redford, Odd Magne Reitan, Frederic Requin, Frederick Requin, Dustin Richmond, Alberto Del Rio, Eric Rippey, Oleg Rodionov, Ludwig Rogiers, Paul Rolfe, Arjen Roodselaar, Tobias Rosenkranz, Huang Rui, Jan Egil Ruud, Denis Rystsov, John Sanguinetti, Galen Seitz, Salman Sheikh, Hao Shi, Mike Shinkarovsky, Rafael Shirakawa, Jeffrey Short, Anderson Ignacio Da Silva, Rodney Sinclair, Steven Slatter, Brian Small, Garrett Smith, Tim Snyder, Maciej Sobkowski, Stan Sokorac, Alex Solomatnikov, Wei Song, Art Stamness, David Stanford, John Stevenson, Pete Stevenson, Patrick Stewart, Rob Stoddard, Todd Strader, John Stroebel, Sven Stucki, Howard Su, Emerson Suguimoto, Gene Sullivan, Qingyao Sun, Renga Sundararajan, Rupert Swarbrick, Yutetsu Takatsukasa, Peter Tengstrand, Wesley Terpstra, Rui Terra, Stefan Thiede, Gary Thomas, Ian Thompson, Kevin Thompson, Mike Thyer, Hans Tichelaar, Viktor Tomov, Steve Tong, Michael Tresidder, Neil Turton, Srini Vemuri, Yuri Victorovich, Bogdan Vukobratovic, Holger Waechtler, Philipp Wagner, Stefan Wallentowitz, Shawn Wang, Paul Wasson, Greg Waters, Thomas Watts, Eugene Weber, David Welch, Thomas J Whatson, Marco Widmer, Leon Wildman, Daniel Wilkerson, Gerald Williams, Trevor Williams, Jan Van Winkel, Jeff Winston, Joshua Wise, Clifford Wolf, Tobias Wolfel, Johan Wouters, Junyi Xi, Ding Xiaoliang, Jie Xu, Mandy Xu, Takatsukasa Y, Luke Yang, and Amir Yazdanbakhsh.
Thanks to them, and all those we've missed including above, or wished to remain anonymous.
The latest version is available from <https://verilator.org>.
Copyright 2003-2020 by Wilson Snyder. This program is free software; you can redistribute it and/or modify the Verilator internals under the terms of either the GNU Lesser General Public License Version 3 or the Perl Artistic License Version 2.0.
All Verilog and C++/SystemC code quoted within this documentation file are released as Creative Commons Public Domain (CC0). Many example files and test files are likewise released under CC0 into effectively the Public Domain as described in the files themselves.
verilator_coverage, verilator_gantt, verilator_profcfunc, make,
"verilator --help" which is the source for this document,
and docs/internals.adoc in the distribution.
| 2020-08-27 | perl v5.30.3 |