There are a few functions that handle implement different kinds of vector
writes that have to handle that the assigned value partially or completely
out-of-bounds.
Each function has similar, but not identical, code for this, sometimes with
small bugs for corner cases.
Add a helper function that takes care of handling of updating the width and
offset of the assigned value if necessary.
This ensure consistent and correct behavior and allow to remove some
duplicated code.
Signed-off-by: Lars-Peter Clausen <lars@metafoo.de>
The `%assign/vec4/a/{d,e}` instructions, when checking for a full
out-of-bounds write on the low side, uses the target signal width, while it
should use the assigned value width.
This can lead to a fully out-of-bounds write to be assumed to be a partial
out-of-bounds access, which will trigger an assert later on.
E.g.
```
integer a[1:0];
integer i = -4;
a[0][i+:4] <= 4'h0; // Triggers assert
```
Signed-off-by: Lars-Peter Clausen <lars@metafoo.de>
A non-blocking event controlled assignment consists of 3 steps.
* Setup event
* Perform assignment
* Clear event
This works fine if the lvalue is a singular value. If the lvalue is a
concatenation multiple assignments are generated and the event is cleared
after each assignment. As a result only the first assignment is event
controlled. All other assignments will be regular non-blocking assignments.
E.g.
```
reg x, y;
event e;
{x,y} <= @e 2'b11;
$display(x, y); // x will be 1'b1, y will be 1'bx
```
To resolve this the event needs to be cleared after all assignments have
been done. This requires changes to both tgt-vvp and the vvp runtime.
tgt-vvp is updated to only insert a single `%evctl/c` instruction for each
event controlled non-blocking assignment.
The vvp runtime is not updated to implicitly clear the event in the
`%assign/vec4/e` instruction and instead rely on the explicit `%evctl/c`.
Signed-off-by: Lars-Peter Clausen <lars@metafoo.de>
Both the `%concati/vec4` and `%pushi/vec4` instructions need to construct a
vector from the immediate value encoded in the instruction. Currently both
these instructions have a custom implementation for that.
Remove the custom implementations from those functions and use the
`get_immediate_rval()` helper function. This removes a bit of duplicated
code.
Signed-off-by: Lars-Peter Clausen <lars@metafoo.de>
The `%cmp/ws` and `%cmp/wu` instructions compare two index registers. They
are currently unused. Since the index registers are not used for data there
is not really a need to compare them. Values can be compared before loading
them into an index register.
So remove these two instructions.
Signed-off-by: Lars-Peter Clausen <lars@metafoo.de>
The `%mov/wu` instruction moves data from one index register to another.
The instruction is not used. It also does the same as `%ix/mov`. So remove
it.
Signed-off-by: Lars-Peter Clausen <lars@metafoo.de>
The `recv_vec{4,8}_pv()` functions are used to implement a partial write to
a vector. As parameters they take both the value and the width of the
value.
All callers of of these functions pass `val.size()` or a variation thereof
as the width of the value. And all implementations that do anything with
the data have an assert that `val.size() == wid`.
Remove the `wid` parameter from these functions and just use `val.size()`
directly where needed. This allows to simplify the interface and also
to remove the asserts.
Signed-off-by: Lars-Peter Clausen <lars@metafoo.de>
Writing parameters into VCD files makes the values available to waveform
tools. This can be done easily enough by writing out a $dumpadd section
at the beginning of the file that sets the parameter values. We don't need
to track the values over change, because by definition they do not change.
This changes the typical vcd output as well, so a few of the regression tests
need to be adjusted to account for this.
Also, while tracking this down, found and fixed the vvp/README.txt documention
for the .param/x records.
The `%disable` instruction will stop the execution of all active
threads of a specific scope. This is what is required to implement
the semantics of the Verilog `disable` statement.
But it is not suited to implement the SystemVerilog flow control
statements such as `return`, `continue` and `break`. These only
affect the thread hierarchy from which it is called, but not other
concurrently running threads from the same scope.
Add a new `%disable/flow` instruction that will only disable the thread
closest to the current thread in the thread hierarchy. This can either be
the thread itself or one of its parents. This will leave other concurrent
threads of the same scope untouched and also allows function recursion
since only the closest parent thread is disabled.
Note that it is not possible to implement this using `%jmp` instructions
since a block in a function with variable declarations will be its own
sub-thread, but using flow control instructions it is possible to exit from
that thread to the parent scope, which is not possible with `%jmp`
instructions.
Signed-off-by: Lars-Peter Clausen <lars@metafoo.de>
Previously both ':' and ';' were recognised as path separators on all
platforms, but ':' can't be used in Windows. So now we only recognise
';' when running in Windows and ':' when running in any other OS.
Array ports are created via a resolve list. We need to detect and
record whether they need to be created in an automatic context at
the point they are declared, not at the time they are created.
String literals may have escaped special characters in them. Make sure
we are processing all the special characters that the standard supports,
and also fix the handling of the assignment to strings. Note that the
vvp input has string literals sanitized so that the parser can handle
the various binary values. desanitize the strings when pushing string
literals into the string stack. This fixes string assignments, and other
string operations.
The old implementation connected all inputs to the same vvp_net_t port,
on the basis that we don't care about the data values or what port they
arrived on. But if one or more of the inputs fans out to multiple nets,
the chains get tangled, which either results in connections being lost
or inappropriate connections being made, depending on the order that
the inputs are linked.
This could have been fixed by using a standard wide functor. But as we
don't care about the data values, that would be unnecessary overhead.
We just need separate vvp_net_t objects to handle the input connectivity
and can keep using a single shared functor.
This is needed for the waveform dumpers now that vpi_iterate(vpiModule, NULL)
has been ficed to only return modules.
This includes recognising vpiProgram and vpiInterface, although the compiler
and vvp currently incorrectly classify them as modules.
The code assumed all handles stored in vpip_root_table were modules. This
is true for traditional Verilog, but not so for SystemVerilog.
Whilst here, also add support for iterating over packages.
This might come to pass for example when calling the function
$ivl_string_method$len(<expression>) where the expression is
calculated, and not simply a variable name.
In the rare cases where DARRAY signals are in the network, handle
them by creating the proper ivl_nexus_t node. This also implements
the receive of vvp_object_t objects bu vvp_fun_anyedge_sa. This
together makes it possible for IVL_VT_DQUEUE objects to be in
wait lists.
This fixes#412
sync_cb::run_run() attempts to cast the obj field of the callback data to a
__vpiSysTaskCall pointer. But a sync_cb object is only used for simulation
time callbacks, where the obj field is (mostly) not used, so I can't see
that would ever succeed. As the obj field is not required to be set by the
user, the dynamic cast results in undefined behaviour, so mingw32 is not to
blame.
This code was introduced by Johann Klammer in commit c79df7c44, but the user
klammerj and all trace of that pull request have vanished from GitHub, and
there is no associated regression test, so I can't establish the rationale
for it.
IEEE Std 1800-2017 Section 7.6 Array assignments
Assignment of a dynamic array creates a duplicate of the source,
so that assignments to the copy don't impact the original. Handle
all sorts of dynamic array base types.
Since commit 15cda5fe, forked threads are never embedded in the parent
thread, so we no longer need the special case code that ensured that
task/function calls were reaped before a join.
This also fixes GitHub issue #368.
When -pfileline=1 is used the queue procedural warnings have file
and line information added to the messages. Also switch the trace
debugging to be off by default.
Also, Add some preliminary missing darray functionality.
With a pattern rule, the recipe will only be executed once, even when
the rule has multiple targets. Using this to handle the output from
bison is included as an example in the GNU make manual.
This fixes the makefiles so that bison-generated header files will be
regenerated if they are deleted.
The current bison (3.7) generates a *.cc file that includes the header
it generated. For parse.cc this would be parse.hh. Right now, we rename
this header to have a common name used in other files, but this results
in a compile error for the parse.cc file:
parse.cc:462:10: fatal error: parse.hh: No such file or directory
462 | #include "parse.hh"
| ^~~~~~~~~~
Fix this by telling bison to output the header file to the correct
filename in the first place so that we don't have to rename it.
(using the --defines instead of -d option).
This looks like a bison specific option not available in Posix yacc;
but looks like we're requiring bison anyway.
Signed-off-by: Henner Zeller <h.zeller@acm.org>
__CYGWIN32__ is not defined when building with the 64-bit Cygwin
toolchain. According to the Cygwin FAQ, __CYGWIN__ has been defined
since 1998, so this should still work for users on 32-bit systems.
In PR #300, @xdch47 pointed out a stable way to fix parallel
installation problems.
This fix applied the method, thanks!
Signed-off-by: Huang Rui <vowstar@gmail.com>
The old scheme of linking the VPI modules with the vvp exports meant they
did not work when loaded by the compiler. Instead, let each client create
a jump table for the VPI routines and pass that to each VPI module as it
is loaded.
This allows any always processes that are sensitive to the UDP output
to start first. This fixes a time 0 race that was found in a Lattice
Semiconductor simulation library (reported on iverilog-devel).
If the initial value is 'x', propagate the value to the UDP output
before the start of simulation, to avoid unwanted update events on
the z -> x transition on downstream nets.
In 64-bit Windows, an unsigned long is 32 bits, so initialising the num
member of the union did not properly initialise the ptr member. The num
member isn't actually needed, so eliminate it.
- type is vpiPartSelect, not vpiMemoryWord
- left/right range is for part, not full word
- index is not valid for a part select
The user will now get a sensible error message if they pass part of an
array word to $dumpvars (GitHub issue #230).
bison 3.4.1 writes the header file before the c++ file. Our makefile
rules make the header files depend on the c++ files, so we need to
fix the timestamps accordingly.
When vpi_handle_by_name() iterates over the VPI objects in a scope, handle
the case that vpi_get_str() returns a null value. This currently occurs if
the scope contains an enum type definition, as vpi_get_str() is not
implemented for __vpiEnumTypespec.
Note we only want these enabled when building the compiler and runtime
binaries. If we included the ICARUS_VPI_CONST definition in the global
CPP_FLAGS, that would propagate to the flags used by iverilog-vpi, so
would affect compilation of user VPI code.
When performing a signed division or modulus operation using native
arithmetic, trap the special case that the numerator is the minimum
integer value and the denominator is -1, as this gives an undefined
result in C++.
In vvp, the __vpiSignal object holds a pointer to the scope containing
the signal. This was getting set to the current scope when the net was
finally resolved, rather than to the scope where the net was declared.
If a thread becomes detached due to a join_any statement, that
thread must not attempt to join its parent, even if the parent
is waiting on a subsequent join statement.
Normally there is at most one signal attached to a vvp functor, but
due to port collapsing, there can be more than one. If these signals
are array words, we need to trigger vpi callbacks on all the associated
arrays when the functor value changes.
Change cbReadWriteSynch and cbReadOnlySynch to interpret the passed
time value as a relative delay, not an absolute time. This matches
the behaviour of other simulators.
SystemVerilog allows tasks, functions, and classes to be defined at the
root level or inside packages, so we can't rely on an enclosing module
being present to provide the timescale.
SystemVerilog allows a mixture of procedural and continuous assignments
to be applied to different parts of the same vector. The previous attempt
to make this work for non-blocking assignments was flawed (see preceding
fix for vvp_fun_part_pv::recv_vec4_pv). Instead, handle this case by
converting the non-blocking assignment into a delayed force statement,
which matches the way mixed continuous and blocking assignments are
handled.
Most vvp functors need to support recv_vec4_pv. Any that are strength-aware
also need to support recv_vec8_pv. Note the simplifying assumption that is
documented in the base class recv_vec4_pv_ implementation.
Lars-Peter Clausen
Stephen Williams
Martin Whitaker
Cary R
Martin Whitaker
Miodrag Milanovic
Marlon James
Henner Zeller
Tim Pambor
Huang Rui
Greg Steuck
Alex Forencich
Scott Mansell
Johann Klammer
Udi Finkelstein
Maciej Suminski
Yury Gribov
Michele Castellana
Larry Doolittle