When multiply is done in native words, the conversion to words from the
vp_vector4_t vectors must be done signed. This only matters if the
input operands are different sizes (and themselves signed) but will
not hurt even if we want an unsigned result.
Logical (in)equality needs to look at all the bits of both operands,
and cannot short circuit the test unless defined bits differ. If there
are undefined bits, the equality is undefined at that point, but return
x only if there are not other bits that make the results clearly
unequal.
The double to vvp_vector4_t constructor was not using the correct
declaration for the bit words. This worked as long as unsigned and
unsigned long were the same size (usually).
The new real to int conversion was incorrectly setting the
bits for minus infinity to all ones. This is incorrect in a
two's complement encoding where the largest negative number
would be a leading 1 followed by an infinite number of zeros.
This patch updates the %cvt/vr command to use the new double to vector
constructor. This allows the resulting bit pattern to be larger than
a long. The old method was producing incorrect results without a
warning for large bit values.
This patch makes .part/pv strength aware, resolv vec8_pv
aware. vvp_net_fun_t adds vec8_pv as a virtual function
with an appropriate error default. vvp_fun_signal should
full support vec8_pv (not tested and may not be needed).
This patch adds .cast/int and updates .cast/real to act as a local
(temporary) net and to support either a signed or unsigned input.
The vvp_vector4_t class not can convert an arbitrarily sized double
to a vector value. This removes the restriction of lround().
Also document the new statements.
This handles the general case of a non-real operand to a real-valued
division. This can turn up if only 1 operand of a divide is real. In
this case the division as a whole is real and the other operand must
be cast to real.
This method creates an extra node, but it should be a very compact
node and this node does no evaluation tricks so in the run time should
be no more expensive then folding the cast into the .arith/div.r itself.
This patch changes the base oct to string converter to correctly
pad the top digit. x or xx should display as a single lower case
x when they are located in the top bits. Before these were being
interpreted as 00x or 0xx and displayed X. Also modified the hex
conversion to use this same scheme instead of a loop.
First, handle the trivial (but possibly common) resolution cases in
inlined code, and only call the complete function for the complicated
cases. Then clean up the complex function for readability, and account
for the constraints that the front-end function established.
Arrays of vvp_vector4_t values redundantly store some fields in every
word. Create a special type that stores vvp_vector4_t values in a form
that does not duplicate the width of all the items. This can save a lot
of space when big memories are simulated.
The schedule_assign_plucked_vector is a better way to implement the
schedule_assign_vector, or at least no worse, so remove the now
redundent schedule_assign_vector.
It is legal (though worthy of a warning, I think) for the part select
of an l-value to me out of bounds, so replace the error message with
a warning, and generate the appropriate code. In the process, clean
up some of the code for signal l-values to divide out the various kinds
of processing that can be done. This cleans things up a bit.
The load-and-add for vectors %load/vp0/s can be combined with the
load-and-add for array words, and the %load/avp0/s added to round
out the combinations. This can make for fewer instructions when
words are padded in arithmetic expressions.
The functor_ref_lookup() function fills its argument in with the
vvp_net_t* pointer that matches the var name, so there is no need
to create the vvp_net_t object before then.
The vpiIndex is really just a different view into the same object,
so implement the trickery needed to support a vpiIndex with the
absolute minimum memory cost.
The %load/vp0 instruction adds a signed value to the signal value being
loaded, but it doesn't allow for a signed source vector. Add the
%load/vp0/s instruction that pads the loaded vector, and add the code
generator details to properly use it.
The vvp_net_fun_t objects, and derived objects, are small, and are
created in large quantities. Tightly pack them into permanently
allocated space in order to save on system allocation overhead, and
thus save overall on memory.
The part select of a vector is converted by the compiler during
elaboration to a 0-based canonical address. But since it is legal
to address bits below the LSB, the canonical address can be negative.
So make the part select base for selecting from signals work with
signed arithmetic and make the code generator generate negative
indices when needed.
Scheduler cells are small objects that come and go in great quantities.
Even though they are allocated and deallocated a lot, they tend to a
steady state quantity, so put together a heap that is unique for each
cell type.
This heap actually saves memory overall because cells are allocated in
chunks, thus eliminating allocator overhead, and they are pulled/pushed
from/to a heap very quickly so that what overhead remains is slight and
bounded.
The vvp_net_t objects are never deleted, so overload the new operator
to do a more space efficient permanent allocation.
The %assign/v instruction copied the vvp_vector4_t object needlessly
on its way to the scheduler. Eliminate that duplication.(cherry picked from commit d0f303463d)
The vvp_vector8_t constructor and destructor involve memory allocation
so it is best to pass these objects by reference as much as possible.
Also have the islands take more care not to perform resolution if the
inputs aren't really different.
NOTE: This is a port of commit 2f4e5bf5b6
from the "performance" branch, without the resolver scheduling changes.
This was causing test suite variances with pr1820472.v. It looks like
there might be a race in that program anyhow, but for now leave out the
resolver scheduling changes so that the rest of this commit can go in.
The __vpiPV objects express themselves as vpiPartSelect objects.
Add support for value change callbacks by attaching the callback
to the signal that we part select from.
The %load/v instruction was doing some spurious resizes of the vector
that comes from the signal. Eliminate those resizes that can be
removed, and optimize some that remain.
Stephen Williams
Cary R
Holger Wächtler
Larry Doolittle