since this is a common artifact of foundry cells and almost
always incompatible with magic. Modified the "port" command to
allow "port make <index>" on a label where other labels of the
same text already have the same index. Removed deprecated
documentation and added some missing documentation, such as an
explanation of the "ext2spice subcircuit auto" option.
doing "save", which was missing checks on properties representing
coordinates (e.g., FIXED_BBOX) resulting in those values
potentially getting truncated by scaling the output to an
incompatible common factor. Thanks to Sylvain Munaut for finding
the issue!
prefix when doing a reentrant load, causing include files to not
be found. Fixed a divide-by-zero issue that occurs when some
tech file sections are not loaded to the abovementioned error.
command, but with some critical differences, since the "crash"
command is designed for crash backups. "crash" will save in a
temp file and removes the file after a successful recovery.
"archive" can be used at any time to make a complete snapshot of
a layout in a single file, or to read back that snapshot.
There is a "writeall" option that will make a snapshot including
layout of all read-only (PDK) cells.
command handling in the previous commit, that can cause a crash
whenever "writeall" is called while a cell's filename is still
NULL. Thanks to Daniel Estevez for reporting the error.
a self-intersecting or reversing path is seen in the input.
Added a new feature in which "save <cell>.tcl" will create a
file of magic commands that will re-create the cell when sourced
as a command file.
X and Y values, in which case the selection is moved to the
current pointer position. This allows a different implementation
of the "copy" command as "select move; select keep", which avoids
the problem that "copy" has of modifying the selection with layer
interactions with the existing layout.
label text) vs. non-extended bounding box of a cell when doing
"getcell" (and probably a number of other commands/functions, as
well). A function was always computing the extended bounding box
and then setting both the normal and extended bounding boxes of
the cell to this value, resulting in incorrect cell selections.
(commit 4084a6a246) in which a
misplaced close-brace altered the way that "getcell" handles
some orientation cases. Thanks to Sylvain Munaut for discovering
the error.
nodes which are created due to having multiple ports with different
names on the same wire (electrically connected pins). This prevents
"extresist" from double-counting the wire.
arguments unambiguous, and to allow coordinates to be given in
any units. To do this, the (seldom if ever used) passing of a
label as a reference point was changed to require the "label"
keyword before the label name, avoiding the code having to
disambiguate arguments from label text. This now makes it
possible to specify the coordinates in microns, lambda, etc.,
units.
non-manhattan geometry was not being transformed into the parent
orientation. This went unnoticed for some time due to testing on
PDKs where substrate and well were not allowed to have non-orthogonal
edges.
option "manhattan_dist" that causes corner checks to assume a
manhattan distance measure. This is useful for checking
distances involving generated edges that are created by a CIF
"grow" operator. For "spacing", "manhattan_dist" is equivalent
to "touching_illegal", as a use-case for forcing manhattan
distance measurements in corners has not been found for other
"spacing" options.
to ignore non-Manhattan (split) tiles. This avoids creating false-
positive DRC results on split tiles which are larger in X and Y
than the max-width distance. False negatives are possible but
correctly implementing a "maxwidth" rule for non-orthogonal areas
requires a completely different algorithm.
version 8.3.521 that, due to an argument size mismatch, causes
device parameters in netlist output to be printed as zero.
Also added a small extension to the list of extraction types
to include "device veriloga", which has the same syntax as
"device subcircuit" but generates a component type "N"
(Verilog-A component) in the netlist output.
that all layer types passed to a single "angles" statement in the
tech file are mutually non-interacting. That is, "angles allm1 45"
will check angles on edges between metal1 and space, but not edges
between metal1 and via (which makes sense, given that a via is an
area into which to place contact cuts, and its edge is not a
physical boundary).
changed the behavior from selecting subcells if they are unexpanded
to selecting subcells regardless of the expansion state. The
change was short-sighted and the implementation hard to use. There
is a need, I think, for a selection method that effectively
unexpands instances, selects everything, and re-expands afterward.
But it's not clear if it is better to select all touching instances
or instances inside the box area; maybe only interacting
instances (instances with non-space content inside the box).
Regardless, it needs more thought, and meanwhile, the original
behavior works better.
Manhattan tiles. After splitting a non-Manhattan tile crossing
a search area to paint, the routine was automatically merging
tiles to the right. This is incorrect for tiles inside the search
area, as it can cause the search algorithm to miss unvisited tiles
whose origins are to the left of the split tile, resulting in part
of the area not getting painted.
not to re-process processed tiles made an incorrect assumption,
causing tiles not to be re-processed when the clip area changed,
such that areas would be missed. It is not clear that in its
corrected version, "bloat-all ... [dist]" is any more efficient
than the original implementation of an incremental series of bloat
+ AND. At least the syntax in the tech file is much simplified.
cast as "bool", causing issues with the proper/strict use of
the "bool" type. Changed the version back so that it will
update the github mirror tonight, since the current version
appears to be working.
database corruption discovered recently that was uncovered by a
commit on Jan. 31 and is caused by DBMergeNMTiles0() using a
freed tile (reported in github issue #404).
commands, because the keypad numerical values no longer work
regardless of the setting of Num Lock. The keypad arrow keys
alone implement "move", while Shift + keypad arrow keys
implement "stretch".
keyword in LEF. FOREIGN may take an origin offset, but it is
optional. The routine to check that there were no offset values
in the statement incorrectly checked for a NULL token instead of
a value ";" which would indicate an end-of-statement.
macro. Based on observation of cells in PDKs where ORIGIN and/or
FOREIGN are non-zero, added code that forces a correction of LEF
macro coordinates to match the GDS coordinates, with an
equivalent negative shift of the LEF macro ORIGIN to compensate.
Normally, both ORIGIN and FOREIGN will be zero and the added code
will do nothing. Note that this code does not handle the
additional optional orientation. A LEF macro with a different
coordinate system than its GDS is already weird; a LEF macro
with a different rotation than its GDS is hopefully something
that nobody ever does in practice. If needed, I'll cross that
bridge when I come to it.
manhattan shapes (especially minimum-sized ones) to be eliminated,
as these can survive a shrink-grow operation intended to get rid
of such shapes. This implementation may not be in its final form
but should suffice for now.
instance names in both the selection and in the root edit CellDef,
and then wipes duplicate names from the selection and regenerates
unique IDs. This avoids the unexpected behavior displayed by
magic in which a "copy" function renames the *original* instance
and gives the original name to the copied instance. This is not
only unexpected, but causes an error in which "undo" after
multiple copies fails to remove earlier copies because the name
change was not recorded, and the instance can no longer be found
by name.
interactive wiring into coordinate-based commands. Added new
command extensions for "wire leg", "wire vertical", "wire type",
and "wire horizontal". Modified the command logging such that
"wire show" (which does not modify layout) does not get logged,
which avoids unnecessary logging of mouse movement.
that has since been compressed and given a ".gz" extension.
Removed code that uses a system call to "gunzip" and "gzip" when
the target file is compressed, since the compression is handled
in the code.
"MASK_HINTS" was implemented, it was put into a routine that
scales all coordinate-related properties, including "FIXED_BBOX",
but the original code that scaled only "FIXED_BBOX" was never
removed, resulting in the bounding box getting incorrectly scaled
twice by any grid scaling.
new CIF operator is used in a tech file. Reworked yesterday's
commit to add more related operators, so there are now four new
ones (also renamed them): interacting, noninteracting, overlapping,
and nonoverlapping. "interacting" now means overlapping or touching;
so the four cases allow all variations of adjacency between the two
material types.
regions of a given type and retains only those regions which
interact with (overlap) another given type. Both sets of types
can be either magic database types or CIF temp layers. This will
allow the implementation of rules that were not previously
possible.
connection to a net that has been decomposed into a resistor array
cannot be found. This indicates some fundamental error in the way
extresist works. However, it should not be producing an invalid
and unsimulatable netlist. Instead, it makes an arbitrary connection
from the device terminal to the resistor array and adds an entry in
the output netlist (.res.ext file). This results in a poor
representation of the resistor network to that terminal, but an
otherwise simulatable netlist. A warning is issued to note that an
arbitrary connection has been made. This is most typically a
"garbage in, garbage out" situation in which insufficient information
exists in a layout to inform magic on which direction current is
traveling through a net. However, it should be possible to rewrite
the extresist code so that magic makes somewhat informed decisions
about current paths and produces a halfway decent representation of
the actual net, instead of just giving up on the detailed extraction.
generate a name for an instance that is set to NULL. It is not
clear to me by an instance would have a NULL name, but apparently
it can happen, and should not crash magic.
the former code attempted to determine the precision and generate
an output without unnecessary trailing zeros. Unfortunately there
were counterexamples that fail to be formatted correctly, as found
by Mark Martin, and which generate output that has roundoff error.
Reimplemented the method using code found on StackOverflow, which
appears to solve the problem more robustly.
"bloat-all" which is "bloat-all types1 types2 distance" where the
"distance" value is a maximum amount to grow. It is not (that I
know of) particularly useful for generating output GDS, but it is
very useful for generating temporary layers for DRC checks,
especially things like determining tap distance for latch-up
rules. The alternative (used in the sky130 tech file) is a
tedious step-by-step "grow" followed by "and-not". This rule
option is much cleaner to implement and computes faster (although
it is still a boolean operator and is much slower than an edge
rule).
identifies areas which meet the proper definition of run-length
(both edges are parallel for the run-length distance or more).
Previously, errors were getting triggered for geometry where
only one edge exceeded the run-length distance.
which has the same meaning as the "maxwidth" function ("both"
checks either tile dimension to see if it exceeds the maximum).
This is a simple per-tile check and assumes that violations do
not occur across multiple tiles. This should be sufficient for
most checks.
different from the device (i.e., gate) width, for devices that do
not define a MOS-like gate spanning the width of the device. This
is restricted to the assumption that the terminal is rectangular
and therefore a simple width and length can be derived from the
area and perimeter. Also, length is defined as the smaller
dimension and width as the larger dimension. For additional
restrictions, see the updated documentation. This was added to
allow correct width and length extraction of a bipolar emitter
window, but may be more generally useful.
at a 45 degree angle will shadow the DRC rule for the material
drawn orthogonally (that is, the DRC rule for the distance between
orthogonal shapes will be eliminated from the rule deck).
generated cell is modified multiple times. If the original cell
is orphaned (no longer used anywhere in the design), it is deleted.
However, an instance of the cell may exist in the secondary
select buffer if the cell was previously moved or copied, and
an attempt to do another move or copy will clear the secondary
select buffer, encounter the deleted cell, and crash the program.
previously ignoring the parameters of the entire cell including
the device being overridden by the property, causing the output
to be wrong. The parameters should always be written out to the
.ext file, including the device whose output is being overridden.
mismatch in the SkyWater sky130_fd_io__top_pwrdetv2 circuit
because a resistor with ends shorted together was being assigned
an incorrect length and width. This was due to the similarity
in characteristics of the boundary vector between a shorted
resistor and an annular resistor. The terminals need to be
checked for shorted ends to disambiguate the two cases.
with zero gate error (and was reporting an infinite antenna ratio).
For now, just ignoring the zero-area case. However, since the
procedure is supposed to be looping through nets connected to
specific devices in the .ext file, then every entry is supposed to
have non-zero area, so there is some underlying problem here that
needs to be fixed.
surrounding a device tile may cause the device to be extracted
with the wrong node (picking up the node from the wrong side of
the diagonal tile). Added extra handling to capture the case
where two ports on two different nets are merged when using
"ext2spice short" (previously it was handling only ports on the
same net). Also: Removed the redundant readline-4.3 from the
readline/ directory; only readline/readline is left, which is
version 4.3.
"extract unique notopports" both fail to work correctly because
an attempt some time ago to avoid issuing warnings about shorted
port names when using those options was changed in the wrong
line.
angled edges. This is a rare case and so has never come up before,
but can happen especially on transistors with gate or diffusion
with 45 degree chamfers or flanges. Thanks to Mark Martin for
providing the use case.
request #325 from Daryl Miles. Made a few simple style changes
to conform to (what is vaguely defined as) the overall programming
style for magic (passed down from John Ousterhout).
which prevents the "Ctrl-P" key from raising the .params window;
the underlying error is that the original error in the "property"
command was a disagreement about where "argstart" is placed. The
command was fixed to make the "property" command work correctly.
However, the solution made "argstart" wrong when using the command
"cellname property". Now both uses are handled correctly.
added fairly recently when getting rid of crashes related to
commands operating on non-edit cells. The lack of a cast was
prevening compilation on systems with more rigorous error
checking.
Previously, a file path beginning with "/", "./", or "../" would be
searched for verbatim and no searching would be done over paths.
This behavior now occurs for a leading "/" only. File paths with
"./" or "../" will search for the file with the path verbatim, then
proceed to search for the file with each search path prepended to
the filename as usual. This solves a problem for reusable, non- PDK
IP blocks, where the IP block may have an abstract view pointing to
a GDS file which is specified as being located at "../gds/<file>".
This file would not be found if the IP block was included into
another project. Now it can be done if the path to the IP is given
by "addpath".
unique" on a flattened layout (some labels end up NULL and the
NULL condition needs to be checked). Also remembered to update
VERSION, which was missed on the last commit.
the same: (1) If the inodes of the filename are the same, then the
cells are the same. This avoids treating symbolic links as
different paths with different files; (2) If both layouts are in
git repositories and the git repository commit hashes are the same,
then the cells are considered to be the same. This allows projects
to be cloned into other projects as dependencies and used in
multiple places without magic treating them as different layouts.
assign the result to anything, causing later uses of variable
"lobj" to have an unitialized value and potentially causing a
crash condition. Thanks to Risto Bell for pointing out the
error.
some of the "select" command syntax; the code change caused the
"select top cell" command to behave the same as "select cell".
There was no specific code for handling "top", so now there is.
that no nets will be checked for antenna gate and diffusion area,
and no antenna properties will be output to the LEF file. This can
greatly speed up LEF output file generation for a large design.
Thank you to Tamas Hubai for the code patch.
in the .mag file "properties" list are not handled as being in
database units during .mag file reading and writing (although they
do track internally), making them subject to being scaled incorrectly
and change between a read and a write. Thanks to Sylvain Munaut for
identifying the problem.
a transistor's first tile record is a well or substrate type;
normally this is avoided, but if that's the only device terminal
that connects to the node, it will be used. Also changed the
code to not report a failure when space is found under the
device, when space is allowed as a substrate type. Instead it
will print a message that the substrate is not being extracted as
a resistive network. However, note that the correct solution is
to do what the regular "extract" code does, which is to paint the
substrate type in the cell area first, so that there are valid
tile types to use for extracting the substrate network.
a .mag file to include the original system path. This restores the
ability to find the tech file for any cell created using the old
SCMOS technologies. This method is as problematic as is the way
all SCMOS tech files used the same name "scmos". But at least it
preserves backwards-compatible behavior (behavior prior to 8.3.471).
This commit corrects github issue #306.
report it after "Failure to read in entire sub-tree". This will
not report every failing cell (since it quits reading after the
first failure) but will avoid the existing issue of printing
nothing and leaving the user with no feedback as to which cell
was the problem.
non-Manhattan geometry is analyzed for resistance extraction.
This patch merely prevents the crash condition. It does not
solve the root of the problem, which is that split tiles can
belong to two different nets, but the tile can hold extraction
information for at most one of those nets.
reduce the amount of redundant painting done by the connectivity
search algorithm, but which was preventing composed types (such
as FET gates) from getting added to a net if one of the composing
types (such as poly) was drawn over the device in an ancestor
cell. Removing the "if" statement does not appear to have any
significant performance impact, so this change is being adopted.
argument. For interactive magic in the Tcl/Tk wrapper, the
"-nowindow" option was appended to the command line. But for
Tcl scripts on the command line, all arguments following the
script name are considered arguments of the script. So the
"-nowindow" argument has to be inserted at the beginning of
the command line as the first argument after "magic".
previously, MASTERSLICE layers would not be added to obstruction
layers made by "-hide". However, an nwell, for example, that
stuck outside of a prBoundary *would* be recorded, which was
inconsistent. Resolved this by allowing MASTERSLICE layers
in the OBS block, but only for layers that are not a substrate
type. NOTE: It may be better to just insist that a MASTERSLICE
layer define an obstruction type in the "lef" section of the
tech file, and treat it like routing obstructions. Alternatively,
one may question whether special obstruction types are needed at
all, as one could simply define an obstruction as a type without
a port label.
contains brackets which are not indicating a cell array. Also
fixed a related issue with the PDK toolkit code, in which the
gencell routines fail if an instance name contains brackets which
are not indicating a cell array.
present. Note, however, this fix breaks the use of "ext2spice
subcircuit descend off" because subcircuits are always descended
into. It's not clear that "subcircuit descend off" worked at all
before, anyway. That still needs fixing.
non-default rule. The code was first failing to identify the via
cut type from the generated via record, and then it was failing
to return to the non-default rule width after the route exits the
via. Both issues have been fixed. Thanks to Sylvain Munaut for
providing a reproducible test case.
new port labels are created for an existing port, then they must
take the existing port number. The code was previously causing
collisions between port numbers on different pins.
lefLayer's "via" record even when the layer might not be a via,
causing potential issues with uninitialized variables. Not sure
if this is related to the bug that started this investigation,
but it was the only thing that looked relevant.
used in a re-entrant manner. Applied to an existing layout,
it will no longer keep generating new instances and ports over
top existing ones. Could use improvement by attempting to
retain the location of a device when the instance changes
device type (such as when a device parameter was changed in
the netlist). However, the current set of changes should
help, whether the re-entrant use is purposeful or accidental.
fails in the bplane code for subcell binning. It is still not
clear why this example causes a failure when the bplane code
has been working for so long. However, simply checking for the
BT_ARRAY bit at one additional point in the code prevents the
crash condition and appears not to have caused any issue with
the database.
a method that failed to work on devices with complex shapes on the
device recognition layer, such as snake-geometry resistors. (2)
The use of contact type "xpc" in the sky130 tech file as its own
contact residue caused the contact tracing in extresist to fail.
I opted to keep the unorthodox contact description in the tech
file and wrote an extension to a routine in extresist to handle
the case.
from Ryan Schmidt, changing regular expression strings in the python
preprocessor to raw string types so that they don't produce warnings
in python 3.12.
name. The instance search routine was not rejecting internal cells,
and so would choose, e.g., a selection cell and reject it because it
was not the edit cell.
The fractional part of the rule distance (modulus after scaling)
does not fit in the unsigned char variable unless it is first
divided by the scalefactor (also requires multiplying up by the same
amount when scaling the other direction). The truncation of the
unsigned char value was causing the minimum area value to be off by
a small amount, causing false negatives (no DRC violation is shown
when metal area is slightly smaller than the minimum allowed).
R. Timothy Edwards