currently each and every device specific instance and model structure
duplicates the memory layout of a struct GENmodel / GENinstance
further the first structure elements behind a GENinstance header
are reinterpreted as a sort of a generic "node" array.
This business is very error prone, and makes it difficult to
extend struct GENmodel / struct GENinstance.
This patch makes this business much more bearable.
loop over DEVmodDelete() and DEVdelete() in CKTdestroy()
instead of doing this business in the DEVdestroy() functions.
As a consequence, most DEVdestroy() functions
collapse completely.
This patch itself is not valid until all invokers of DEVdestroy()
(currently this is CKTdestroy() exclusively)
are rewritten to invoke DEVmodDelete() and DEVdelete()
Instead of searching and then deleting a given device-model,
just delete the given model.
The search shall be done somewhere else.
Don't free the model struct itself,
this shall be done by the invoker himself.
Don't free the device instrance list,
this shall be done by the invoker himself.
As a consequence, most DEVmodDelete() functions
collapse almost completely.
This change is of no consequence,
because DEVmodDelete() is currently nowhere used.
Instead of searching and then deleting a given device-instance,
just delete the given instance.
The search shall be done somewhere else.
Don't free the instance struct itself,
this shall be done by the invoker.
As a consequence most DEVdelete() functions
collapse almost completely.
This change is of no consequence,
because DEVdelete() is currently nowhere used.
A missing update of "prev" would cause endless looping
when searching for a device instance.
introduced in commit:
> commit 38e68869a2
> Date: Wed Mar 1 14:30:24 2006 +0000
>
> Added mdel, del and dest functions to interface.
Dependent on given R and L parameter values,
"txl" models might be transformed to "cpl" models in INPdomodel().
This would cause a "unrecognized parameter" warning in create_model()
when searching for the "txl" flag in the "cpl" model,
which is avoided with some awkward extra processing in inpgmod.c
This commit removes this special processing
by addition of an "alias" parameter "txl" to the
"cpl" module parameter description CPLmPTable[] in "cpl.c"
Note,
setModelParm() is a no-op for "cpl" and "txl"
see case CPL_MOD_R in
function CPLmParam()
and case TXL_MOD_R in
function TXLmParam()
When 'level' is a valid model parameter,
then it was processed like any other parameter,
but additionally the
`if (strcmp(parm, "level") == 0)'
invoked INPgetValue() a second time.
This special processing is meant to allow "level" for all models
whether they make use of it or not.
The excess invocation of INPgetValue() didn't cause harm,
merely because the next token after the "level=number"
almost necessarily is a string (the beginning "name=" of the next assignment)
thus not a parse-able number,
thus the second INPgetValue() didn't modify the 'line' pointer.
To sanitise the code invoke the "level" skipping only if "level"
is not recognised as a valid model parameter.
MIFdelete() might be called with third arg being NULL,
searching for the instance to be deleted by name only.
Need to invoke the callback in this case too.
Prevent a crash of example/FFT_test.cir.
Buffer 'buf' in function drawlingrid() has been too small.
Increase 'buf' size, and add some protection.
The example file examples/plot/test-small-plot.cir
shows the limits of plotting:
difference in two numbers at digit 14 is plotted well.
at digit 15 is plotted with bugs.
at digit 16 is reduced to integer.
consider compiling with
-O1 -D_FORTIFY_SOURCE=2
for gcc, and
/GS
for visual studio.
Fix a very old bug in the implementation of the
gummel-poon model formulas for the junction capacities.
(already in spice3f4)
For unknown reasons, the same formulas in the
diode mos1 mos2 mos6 jfet models
have been implemented correctly right from the beginning,
the typo was bjt specific.
Many Thanks for the concise and precise Bug Report
provided by Schabo Rumanus on the ngspice-users mailing list
> To: ngspice-users@lists.sourceforge.net
> Subject: [Ngspice-users] Temperatursimulation von Sperrschichtkapazitäten
Many Thanks to Dietmar Warning for his thorough verification
of this commit.
overwriting SPICE_DSTRING `dstr_p' may invalidate `s'
This bug was injected with the introduction of SPICE_DSTRING in :
> commit 76feebbbfa
> Date: Thu Feb 25 21:43:03 2010 +0000
> various bug fixes
cider relied on our broken ngspice variant of erfc() which behaved as
erfc_ngspice(x) = erfc(fabs(x))
Note:
apart from cider, erfc is used in
devices/ltra/ltramisc.c
always with positive argument
and
src/frontend/spec.c
again always with positive argument
Thus, the difference in implementation
never mathered anywhere except in cider
which these days is guaranteed to be provided by <math.h>
note,
our own implementation was incorrect anyway.
it evaluated to
erfc_ngspice(x) = erfc(fabs(x))
which moved processing of several parameters into a new function
ACM_SourceDrainResistances()
but did so incompletely.
As a consequence the matrix footprint was not correctly modified
when source or drainresistance is zero
Thanks for the bug report by "Sto" in message
> bug: mosfet models with deliberate NRD=0 (ngspice-27)
https://sourceforge.net/p/ngspice/discussion/133842/thread/fa39228d
This function is invoked several times during initialisation,
possibly with yet incomplete model and instance parameter settings.
And it might be invoked many times during dc sweep.
When "alter" is invoked after a break-point stopped a transient simulation,
then it might be necessary to update RESconductance
before continuing the transient simulation.
Thanks for the bug report to An Zou
in message "alter resistor value" on the ngspice-users forum
https://sourceforge.net/p/ngspice/discussion/133842/thread/33ce6855/
Note:
RESresist and RESacResist are device parameters
they do not reflect effective resistance when "ask"'ed
(don't reflect tc1, tc2, tce, m, scale, temp ...)
RESconduct and RESacConduct are "query-only" parameters
these report the effective conductance when "ask"'ed
(including tc1, tc2, tce, m, scale, temp ...)
This commit did rely on the existence of a parent 'wordlist' element,
usually guranteed by a leading "alter" or "altermod" command word.
But 'inp_evaluate_temper()' passes a wordlist without any parent word
when invoking 'com_altermod()'
initialise param_beg and param_end before they will be used,
instead of "afterwards" in preparation for the next following usage.
and move the "comment out" up some lines, think that way:
comment out original line, append new line, ...
Traversing all linked memory structures to free their memory
can be a somewhat lengthy business, especially in xspice,
which is not worth the effort when we simply want to exit()
Long delays have been reported in certain cases.
For developers and for the usage of such tools as valgrind,
we still free all the memory if 'set ngdebug' is given.
which silently dropped the
here->initialized = MIF_FALSE
aspect of the MIFunsetup() function
which caused segfault in testcase
examples/memristor/memristor_x.sp
and use a more robust test for local node numbers.
That is, transform this pattern :
if (here->Node && ...) {
CKTdltNNum(ckt, here->Node);
here->Node = 0;
}
into this :
if (here->Node > 0 && ...)
CKTdltNNum(ckt, here->Node);
here->Node = 0;
The change of "!= 0" ==> "> 0" accounts for rare cases where "Node"
might have been set to -1, (meaning "unconnected")
The unconditional execution of the zero assignment is for those cases
where "Node" might have been assigned to some external or other local Node.
If so, the variable would not be set to zero, confusing the "guarding" if's
in the corresponding XXXsetup() routine.
The Pattern to follow is:
1) unset and delete *all* local Nodes in XXXunsetup()
2) allocate all of them again in a re-invocation of XXXsetup(),
exactly the same way as in the very first invocation.
for almost all other external nodes (notable exception "txl")
src/spicelib/devices/*/*def*.h, declare external node variables const
1) The compiler shall emit an error message if we still mess around
with external node numbers.
2) To mark which elements of the instance struct are meant to be set
externally when parsing the netlist
These "external" node variables are exclusively set via the
overlay struct GENinstance, member GENnode[]
We shall not mess around with these "external" node variables
because it would get rather difficult to avoid bugs considering
re-invocation of the XXXsetup() routine.
This gets interesting for devices with optional ports,
which get copied around depending on the amount of connected ports.
All locally created nodes (CKTmk..() invocations in XXXsetup())
must be deleted in XXXunsetup()
Otherwise CKTmk..() invocations during a following CKTsetup()
will re-emit still used node numbers,
thus accidentally shorting nodes.
This patch fixes the complex cases,
which are external node variables (ports of the instance),
which might be moved over to other external node variables
to cope with optional port connections.
This is fixed by copying the node numbers to local shadow variables
to avoid messing with the external nodes.
Otherwise a following CKTsetup() might again meddle with the external
node variables, at least causing considerable confusion, probably causing
hard to find bugs.
All locally created nodes (CKTmk..() invocations in XXXsetup())
must be deleted in XXXunsetup()
Otherwise CKTmk..() invocations during a following CKTsetup()
will re-emit still used node numbers,
thus accidentally shorting nodes.
This patch fixes a little bit more complex cases,
which are local node variables which will start with value 0
and eventually be set with the result of a CKTmk..() invocations,
but might as well receive a node number from another node variable.
Here CKTdltNNum() must not be invoked if the node number is merely
a copy from another node variable.
All locally created nodes (CKTmk..() invocations in XXXsetup())
must be deleted in XXXunsetup()
Otherwise CKTmk..() invocations during a following CKTsetup()
will re-emit still used node numbers,
thus accidentally shorting nodes.
This patch fixes the simple cases,
which are local node variables which will start with value 0
and eventually be set with the result of a CKTmk..() invocation.
More complex cases evolve, if such a local variable might receive
a node number which origins from another node, or the netlist itself.
Invoke CKTdltNNum() in reverse order of local node creation in XXXsetup()
This fixes a bug at least in mesa and hfet1,
where already destroyed node variables have still been accessed
in the guarding 'if' of a later CKTdltNNum() stanza.
Invoking them in revers order easily prevents such bugs,
and improves readability.
Consider the following silent contracts:
1)
CKTsetup() invocations must be separated by a CKTunsetup() invocation
But CKTsetup() has an internal flag,
which will prevent re-invocation of DEVsetup()
But DEVsetup() will be called during sensitivity analysis,
bypassing this precaution. It is fatal if this will
cause another node allocation (CKTmk..()).
This commit tries to detect such cases.
(Note: many DEVsetup routines (all ?) have their CKTmk..() invocations
guarded to avoid reallocation of local nodes,
see commit f7f454c0a1
bug fix, fix the guard for device generated internal nodes (via CKTmkVolt())
)
FIXME:
DEVsetup() is seriously obfuscated by these guards.
If would be far better, if the sensitivity analysis
wouldn't sidestep into DEVsetup()
consider a device local variant of the CKTisSetup flag
2)
DEVunsetup() must delete all, each and every,
local allocated node in DEVsetup()
Otherwise CKTmk..() invocations in a following CKTsetup() will
return duplicate node numbers, effectively shorting device nodes.
This commit tries to detect incomplete CKTdltNNum() invocations.
3)
DEVunsetup() must not delete a netlist node.
This can easily happen in those devices which have optional ports,
which have code in DEVsetup() which copies node numbers to
local node variables.
This commit tries to detect such errors.
Allow to register a callback function in the cfunc.mod files,
which will be invoked in MIFdestroy.
Usefull to "free" memory which has been allocated locally in a cfunc.mod file.
For developing purpose only, and bound to HAVE_TSEARCH, HAVE_TDESTROY
IFparm descriptors must obey certain contracts,
1) aliases must be grouped together with their associated main descriptor,
and all must have flag IF_REDUNDANT set
2) there shall be no duplicated .id fields except for aliases.
3) aliased entries shall have consistent .dataType fields
check_ifparm will traverse all IFparm sets and report violations
try to make it more robust with regard to
'!=' '<=' '==' don't misinterpret as '='
"atemperaticvariable" don't misinterpret as 'temper'
multiple temper in one expression
https://sourceforge.net/p/ngspice/bugs/315/
longjmp() creates segfault under Windows 10 (VS and MINGW)
somewhere when MainWindowProc() is processing the WM_CLOSE message.
Instead of raising SIGINT, directly call cp_doquit().
return 0 if "askquit" is set and the user declines to "quit".
This continues the simulation, if still active.
When translating an instance_name in a subcircuit instance 'xfoo'
undconditionally express the exception for 'x' instances, which is:
x42 --> xsub.x42
but for all others the first char (the devtype) is prepended
r42 --> r.xsub.r42
Emit an error message when unable to identify a valid modelname token.
If a default model is needed, then create one explicitly:
m1 1 2 3 4 mosdefault
.model mosdefault mos1()
Allow to override the admsXml executable when invoking
autogen.sh and configure.ac
ADMSXML=/foo/bar/baz/admsXml ./autogen.sh --adms
ADMSXML=/foo/bar/baz/admsXml ./configure --enable-adms
alternatively (and more advertised)
./configure ADMSXML=/foo/bar/baz/admsXml --enable-adms
The generated Makefiles remember the given path.
A relative path doesn't make sense of course.
Which is needed to release the allocated circuit nodes in
the corresponding adms generated DEVsetup() routine.
This is required when invoking CKTsetup() more than once
without an intervening "reset".
Tie ports 4 and up to ground, iff not connected.
Actually, I think the model's DEVsetup routine should do this business,
we should pass -1 to denote the fact of being unconnected
If a default model is needed, then create one explicitly:
q1 1 2 3 npndefault
.model npndefault npn()
Note:
This code was buggy anyway because the last token,
which turned out not to be a valid model name,
was silently swallowed.
Thus PARSECALL() didn't parse the whole thing.
Thats a functional change of ngspice.
So far we did cook up a "default" model when we didn't find a model token.
But this was buggy anyway, because in such cases we silently dropped
a token from beeing processed by PARSECALL()
Now, we emit a message, and skip this line alltogether.
Note:
This commit does not change behaviour in any way,
yet one can instantly see some sickness now.
Some of the models definitely do not have 7 nodes,
Actually we would like to fetch the number of terminals from
struct IFdevice, slot terms or slot numNames
But it turns out several models publish bogus values.
For example "B4SOI", which has 7 nodes, but claims to have 8
rlar
Holger Vogt
Fabio Rossi
h_vogt
dwarning
Tim Edwards
mhx
Francesco Lannutti