Add explicit length limit to sscanf in TclmagicRegisterCommands: %92s
instead of %s prevents a potential stack overwrite if a command name
were ever longer than the buffer. Matches the available space (keyword[100]
minus the 7-byte "magic::" prefix minus null).
Extend the CI output-display step to also iterate over output-tcl/ so
that TCL-variant test regressions are visible in the job log without
downloading artifacts.
tclmagic.c: remove stray /*-----*/ line left over from a previous edit
that left a duplicate comment opener before TclmagicRegisterCommands.
magicWasm.c: move TxSetPoint inside the #else (non-TCL) branch of
magic_wasm_source_file and restore its explanation comment. TxSetPoint
routes TxDispatch commands to the layout window; it is irrelevant and
misleading in the Tcl_EvalFile path.
magic/Makefile: guard the TCL linker flags in the magic.js link rule
with ifneq (${TCL_LIB_DIR},). When building the non-TCL WASM variant
TCL_LIB_DIR is empty, so the unconditional -L${TCL_LIB_DIR} -ltclstub
expanded to a bare -L flag and a missing library, breaking the notcl
build.
Replace the last three intubun/tcl mentions with tcltk/tcl in
npm/tcl.js, toolchains/emscripten/build-tcl-wasm.sh, and npm/tcl.ref.
Also remove the stale comment in tcl.ref that referenced a non-existent
update-tcl GitHub Actions workflow.
Rename *-magic.tcl scripts to *-tcl.tcl to match the -tcl.js test naming
convention. Delete the old bare-command pcell.tcl (used cellname create,
no magic:: prefix) and promote pcell-magic.tcl to pcell.tcl.
Fix misleading comment in magicWasm.c: Tclmagic_Init only bootstraps the
interpreter; magic:: commands are registered separately by
TclmagicRegisterCommands after magicMainInit. Align comment block
dashes in TclmagicRegisterCommands to match tclmagic.c style (62 dashes).
Replace intubun/tcl references in build scripts and CI with tcltk/tcl to
match the actual pinned repo in npm/tcl.ref. Also run both test suites
when build.sh is invoked with --test.
All non-TCL tests (extract, gds, drc, cif) now also run against the TCL
variant using magic::-prefixed Tcl scripts. A new PCell test generates
two parameterized rectangle cells and verifies their GDS output.
The CI workflow runs npm run test:tcl after the TCL build and saves the
generated output files as an artifact.
magic_wasm_init() called Tclmagic_Init() and magicMainInit() but never
ran the command-registration loop in _magic_initialize(), so magic::tech,
magic::load, magic::gds and all other Magic commands were missing from
the Tcl interpreter.
Add TclmagicRegisterCommands() in tclmagic.c containing the
WindNextClient/WindGetCommandTable loop and call it from magic_wasm_init()
after magicMainInit() succeeds.
Also change magic_wasm_source_file() to use Tcl_EvalFile in MAGIC_WRAPPER
mode so scripts with magic:: commands are evaluated through the Tcl
interpreter instead of the plain text dispatcher.
Bump VERSION to 8.3.645.
magic.wasm can now be built as two variants packaged in the same npm
release: notcl/ (legacy, magic's own parser) and tcl/ (intubun/tcl 9.x
statically linked, commands evaluated by Tcl_EvalEx). The TCL fork is
pinned via npm/tcl.ref and cloned/built by magic itself — the tcl/
checkout is treated as read-only and built out-of-source into
magic/build-tcl-wasm/.
Configure layer:
- New usingTk variable decoupled from usingTcl in scripts/configure.in
+ scripts/configure, so --with-tcl --without-tk is finally a valid
combination. Native Linux Tcl+Tk builds keep their previous behaviour
(both flags default to enabled).
- When usingTk is empty, configure passes -DMAGIC_NO_TK so the small
number of remaining Tk callsites in tcltk/tclmagic.{h,c} compile out,
and TKCOMMON_SRCS / USE_TK_STUBS are omitted from the link.
WASM build orchestration:
- toolchains/emscripten/build-tcl-wasm.sh builds libtcl9.x.a + libtclstub.a
+ tclConfig.sh out-of-source from a pristine intubun/tcl checkout.
- npm/build.sh grew a --variant=<tcl|notcl|both> flag and writes its
outputs into npm/tcl/ and npm/notcl/. It also clones intubun/tcl with
autocrlf=false at the SHA pinned by npm/tcl.ref.
- magic/Makefile (WASM block only): magicWasm.o is now compiled with
DFLAGS_NOSTUB so Tcl_CreateInterp resolves to libtcl9.x directly
before tclStubsPtr is set. magic.js link pulls in LIB_SPECS_NOSTUB
and -ltclstub. After rules.mak include, magic: is a phony alias for
magic.js so the generic ${MODULE} recipe doesn't fight it.
- toolchains/emscripten/defs.mak: add -sUSE_ZLIB=1 (libtcl9 references
zlib), replace -sSTACK_SIZE=N with -Wl,-z,stack-size=N (emcc >=5
rejects the setting form).
- magic/magicWasm.c bootstraps the embedded interp under MAGIC_WRAPPER
(Tcl_CreateInterp -> Tcl_Init -> Tclmagic_Init) and routes
run_command through Tcl_EvalEx.
- magic/magicTop.c: gate MagicVersion/Revision/CompileTime on
!MAGIC_WRAPPER so they don't collide with the copies in
tcltk/tclmagic.c when both objects land in the same wasm binary.
npm package:
- Subpath exports: ".", "./tcl", "./notcl". Default import keeps the
pre-existing non-TCL behaviour for backward compatibility.
- examples/smoke-tcl.mjs exercises the TCL variant.
CI:
- main-wasm.yml clones intubun/tcl at the pinned ref, builds both
variants via npm/build.sh --variant=both, runs the existing notcl
test suite and the new TCL smoke test, and publishes only on a
v<x.y.z>... git tag. Tag name (minus the leading v) becomes the
npm version.
a warning appropriate for duplicate points and not a path
reversal warning. Modified the output for path reversal or
duplicate point errors so that the coordinate is in GDS units
and not path units (which are multiplied by 2).
This uses environment variable PDK_ROOT to search for available
technologies, and loads them by sourcing the .magicrc file instead
of using "tech load". However, there are currently some caveats:
(1) The device generators do not immediately appear in the window
menu; a new window must be created for the extra menu buttons to
be added; (2) Once a PDK sets up the menus, they cannot be
revised. Switching to another technology causes the extra menu
items to not appear at all. Both issues need to be addressed.
Corrected an error that was caused by moving the handling of the
"Short"-type device from "extract" to "ext2spice". When running
with "ext2spice extresist on", the "Short" device is read twice.
Since it's being treated like an "equiv" statement, it's very
important to do the node merge only for the device as it appears
in the ".res.ext" file, and ignore the entry in ".ext", or the
wrong nodes may get merged. This appears to be the last major
error lurking in the "extresist" code, as I am now able to run
full R-C extraction on an entire chip, in reasonable time.
ambiguity in the tech file and therefore also in the .ext file.
This is easy to resolve and it was only necessary to change the
type after finding the tile belonging to the device. There are
pathological cases where this might still go wrong, like defining
a "short" device name for metal resistors on any plane; then,
if a "short" device overlaps something on another plane, it might
be given the wrong type. Avoiding all ambiguity would require
adding an extra item to "device" lines in the .ext file output;
the short plane name would be most efficient.
and not being used. Corrected one other obscure error in which
reading the .ext file picks up the wrong device type because it
is possible to have multiple types map to the same device name.
This is easy to detect and fix in place.
that the check for the "starting device" was not checking if a
terminal was substrate, and so was treating it as a terminal,
which would cause the extraction starting point to be set on the
wrong plane, along with probably other unwanted behavior.
Ignoring substrate terminals when checking for current drivers
solved the problem. Also: Removed some code that had been
reworked but was still sitting in an "#ifdef 0" block.
time-consuming (note the commit message "There are still likely much
better ways to do this"). Worked out a method of hashing the location
of terminals on other planes so that they can be looked up from the
address of the tile. This requires only one pass through the devices
to locate and record the terminals, and no loops through devices when
processing tiles. Also corrected the positioning of the substrate
terminal to be the center position of the device tile, not the substrate
tile. This not only better represents the substrate connections, but
it also eliminates the condition in which many thousands of resistors
converge on the same point, which causes stack overflow. With the
current scheme, stack overflow is unlikely to occur.
several more questionable methods and allows extresist to find
devices above nets that are part of substrate planes or terminals
that exist on planes other than the device (like diodes, often).
There are still likely much better ways to do this.
checking device terminals in "extresist", leading to incorrect
assignments for devices with only 3 terminals, or for asymmetric
FETs. Corrected "select area label" which had stopped working from
a handful of commits back when the "select" command was corrected for
visible/invisible labels and cells. Corrected the "extresist mindelay"
command option parsing, which was not allowing mindelay to be set to 0.
"tagged <text>" without any additional arguments, to indicate
that the operator should look for text that overlaps the
current layer. That allows handling the case were a text tag
is sitting over space inside a cell, which otherwise would not
be able to be searched. Also corrected a likely issue in which
point labels could not be used as tags because they have no
area to search under.
the output if they reference a killed node. This will result in an
invalid netlist, but at least now all devices are represented in the
output, and a warning is issued about the error. There are still
multiple problems that need to be tracked down and fixed. Most
problems are stemming from the unusual "Short" device handling.
long-standing problem, in which the resistance extraction attempts
to determine the device type of each device. Which is stupid,
because the original extraction process has already done this, and
the information is readily available in the .ext file. This update
allows "extresist" to get all the device information from the .ext
file, so there is no guesswork and no extra time spent trying.
Unfortunately, it does not solve the problem I expected it to solve,
which is the disappearance of _dss and moscap devices from gf180mcu
netlists. More work to do. . .
"lasttop" after a new heap has been allocated and "top" is reset.
This is a minor issue, as even random data at the top of the
newly allocated heap is vanishingly unlikely to make a match,
but valgrind caught it as a use of unallocated memory, and it
should be fixed.
one place as ohms and another as milliohms. Changed it to be
milliohms always, except when written as output to the lumped
resistance file or as diagnostic output. The result makes more
sense, but now has the issue that the maximum resistance is always
an over-estimate and may need to account for resistances in
parallel.
determining what does and doesn't get into the R-C extraction
output. Added option "debug" and moved all excess diagnostic
information to be gated by this flag. Corrected an issue in
the standard extraction in which the substrate is always output
with zero lumped resistance, preventing it from being checked
by "extresist" for inclusion or exclusion in the .res.ext output.
Reinstated the computation of minimum resistance transistor but
excluded other devices. Added more sane handling of exceptional
conditions like loops. The result is a more consistent check of
nets in the design. However, I still need to recheck the
"rg_maxres" calculation because it appears to be too low.
(1) An incorrect use of ExtResetTiles() was found in "extresist"
which impacts performance, especially for small nets.
(2) Corrected units for resistance tolerance in extresist, and
handled output printing in fs when delay values get below ps
size.
(3) Added command option "extract do unique notopports" to be the
extraction option equivalent of the standalone command
"extract unique notopports".
(4) Changed the "extresist" default for "mindelay" to 1ps from 0,
in response to the observation that lumped resistance from
"extract" can be an extreme overestimate, and the extracted
time delay from "extesist" calculations should be used as a
better determination of whether a net should be output as a
resistor network or not.
(5) Added documentation for both "extract do unique notopports"
and the change to the "extresist" default values.
was checking for device terminals that were pointing to the
deallocated node and updating them. However, it failed to check
for whether or not the substrate terminal was pointing to the
deallocated node, potentially causing a crash condition. That has
now been fixed.
that "extresist" selects or rejects networks for resistance
extraction. The "tolerance" value, which nobody (including
myself) seems to understand, is now deprecated, replaced by
a handful of more meaningful options: "threshold" to set a
lumped resistance estimate threshold to determined whether or
not to extract a network; "minres" to specify when to prune
small resistors from the resulting tree network, and "mindelay"
to gate the selection and output on the estimated signal
propagation delay. With these settings, I am much better able
to control the size and complexity of the resulting output.
overhaul was to introduce the concept of a "connection point"
between two cells to the ".ext" file format so that it becomes
possible to do proper hierarchical R-C extraction. Previously
this information was not kept in the .ext file, making it
impossible for "extresist" to know what sub-net of a net connects
up or down the hierarchy. However, additional changes were made
to the extresist code (in the resis/ directory) to add the
handling of connections up and down the hierarchy, and to make
the code a lot clearer and fix ambiguous variable and subroutine
names.
The original version of extresist only worked with 4-terminal FET
devices, and handling of devices with fewer terminals was ignored,
and never properly dealt with. This commit fixes the issues with
devices such as diodes that have fewer terminals.
with respect to visible/invisible labels and cells. The change
inadvertently made the "select area" command option stop selecting
labels on the same type as the layers being selected. This has
been fixed.
"getnode", as there is no limit on the length of a hierarchical
node name, and no check on the string length when copying back into
a fixed-length character array.
The previous code would iterate over `proplen` chunks of strings
but would not initialize `proplen` for plane properties.
Technically for plane properties you don't need to "pre-count"
the string chunks but given how the code is currently written
it's easier to do the counting in all cases.
Also makes sure proprec is init to NULL so that if argument is
empty, property is set to NULL and not random value on the stack.
Fixes#520
Signed-off-by: Sylvain Munaut <tnt@246tNt.com>
in "extresist", which when missing causes the plane to be searched
repeatedly from the same point instead of from the last place searched.
Can knock down the "extresist" time by up to 50%.
this had two issues: (1) "getcell" was not being called with
"child 0 0", leading to child cells that are not placed where
expected in relation to a grid snap setting; and (2) "cursor"
was assuming internal units and needed that to be explicitly
stated using "cursor internal".
DBPaint() and DBErase(), which are meant as high-level calls from
the "paint" and "erase" commands and are not appropriate for low-
level tasks like erasing all contacts out of a database. This
cut maybe 40 to 50% time off of the resistance extraction.
avoid re-parsing a linked list from the current breakpoint
to the end to find all entries pointing to the same node.
Instead, the change to be made is saved in a hash table and
applied at the following link, so only one pass through the
linked list is required.
Changed the "ResDissolveContacts()" routine to avoid running both
redundant and useless code. Changed the breakpoint sorting
routine from a bubble sort to a merge sort for linked lists
longer than 16, as merge sort is more efficient for long lists.
This file exists in the magic directory, and is the one the build system uses, this file is a duplicate that has one different line "/ms { rectfill } bind def", and is not used anywhere
Keeping two duplicate files causes issues like them being out of sync, therefore it would be great if it was deleted
The postscript template had a hardcoded stipple length of 29, any more stipples would cause problems with a postscript parser. This was particularly a problem since the default mos.24bit.dstyle has 36 stipples already.
It also used a base-30 encoding.
I have updated the template to fix both issues, it now uses base-10 encoding, and can support up to 255 stipples (which is the maximum length allowed by the current approach of using fonts)
I was also wondering if it made more sense to encode this file directly into the C program as a literal, as half of it already is in the C program. (some "begin"s have an "end" in strings directly embedded in magic, so this file basically serves as a large string that has to be loaded each time from disk, while other parts of the string are stored normally in memory)
ResTriangleCheck(). By using hash tables, I reduced
ResParallelCheck() from O(N^2) to O(N), and I reduced
ResTriangleCheck() from O(N^3) to O(N^2); however, it is a bit
better than that because previously there was an N^2 loop over
items in the same list, so if one list happened to be very long,
then N^2 would be huge; whereas now the N^2 is the length of two
lists multiplied together, where both lists would have to be very
long to have the same performance impact. It appears that long
resistor lists are pathological and rare. For an example
pathological case, the extresist runtime has been reduced from
hours to seconds.
Intubun
Intubun
R. Timothy Edwards
Torleif Skår
Sylvain Munaut
Ahmed Nematallah