Skip to content

8.3.686

Choose a tag to compare

@github-actions github-actions released this 26 Oct 07:03
· 55 commits to master since this release

commit 8.3.686

"Tagging version 8.3.686"

See documentation at https://github.com/RTimothyEdwards/magic/blob/56319630/appimage/7/README.md
See documentation at https://github.com/RTimothyEdwards/magic/blob/56319630/appimage/8/README.md
See documentation at https://github.com/RTimothyEdwards/magic/blob/56319630/appimage/9/README.md
See documentation at https://github.com/RTimothyEdwards/magic/blob/56319630/appimage/10/README.md

Change Log (since previous tag):

56319630 After investigating a problem with generating the SEALBOUND layer over an entire chip for ihp-sg13cmos5l, I discovered that the "maxrect external" operator was not behaving as expected for hierarchical processing, causing the layer to get clipped to interaction areas and therefore erased over most of the chip. The solution offered by Claude Opus 5.5 was to treat certain operators like "maxrect external" as non-local, forcing them to be computed over the entire chip instead of inside interaction areas.  As long as "maxrect external" is used carefully in a tech file, this does not appreciably change GDS generation time, and it solves the problem with maxrect (as well as other related issues that happen not to have come up yet).  I tested the new code on the Chipalooza IHP chip.  The run-time is approximately the same, and the result is the same except for the corrected SEALBOUND layer.  So I am signing off on this code change.
d083bef0 Fix to the handling of instance names (CellUse record cu_id) in a number of contexts.  Some routines had assumed an unlimited length to this record, while others had assumed it to be limited to 1024 characters (or less) and proceeded to write into a similar sized fixed string without checking the length.  Claude Opus 5.5 carefully worked through the several routines that process the cu_id record and resolved all the issues with potential string overflow, and I have reviewed the resulting code.  This fix resolves github issue #564 (while leaving unaddressed the issue of why any tool would produce an instance name that is thousands of characters in length).
c5139c42 Additional work (entirely by Claude Opus 5.5, with my review) to generalize the method of the previous commit to include spacing rules as well as surround rules, as spacing rules are also subject to the problem of the difference between measuring a distance from the original edge to a target edge at a corner.  Discovered an error in the original commit which does not except cases where the rule defintion includes the residue layer type as well as the contact type, since the method only applies to the contact type itself.  Further generalization to non-contact types is possible but will not be implemented unless the need arises.
8c02247b Correction to an issue that has been known for a long time but remained unresolved:  There is a quirk of geometry in which corner- to-corner measurement of a "surround" rule for a derived layer edge like a contact can be incorrect when the derived layer is an expansion of the original layer.  If the derived layer is distance "d" larger than the original layer, then a surround rule of distance "s" to the original layer can be implemented as rule distance "s - d" for the derived layer, but this is not quite correct for a corner distance measured at an angle.  The correction flags rules that are measured to a derived layer edge and adjusts corner measurements to compensate.  There may be additional work to resolve similar spacing rule errors or to identify derived layer edges on something other than a contact type, but as those issues have not come up in practice, it is not urgent.  This correction resolves a handful of false-positive DRC errors in both the GF and sky130 standard cell libraries. The fix was made by Claude Opus 5.5 and reviewed by me.