Repository navigation
Conversation
…its min A parent's CTS balances a hard macro by the max_clock_tree_path on its clock pin: pin arrival plus that number against the parent's registers. A model whose max lies below its min clocks the block late in every parent that reads it, and nothing downstream says so: the parent's timing reads the same model and agrees with it. generate_abstract now reads each model it writes back and stops (FLW-0030) on a clock pin whose max_clock_tree_path, rise or fall, is below its min_clock_tree_path. One block of XiangShan had max 216.3/227.4 ps against min 501.1/525.9: write_timing_model wrote a nested macro's pin arrival without the macro's own tree (fixed in OpenSTA The-OpenROAD-Project#481), and the parent's CTS put the block's clock pin 245 ps late. A clock tree path written in a form the check does not read stops it too (FLW-0031), so a change in the writer cannot turn the check into a silent pass. The check is necessary, not sufficient: a bound that is wrong but on the right side of the other passes it. Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com> Signed-off-by: Øyvind Harboe <oyvind.harboe@zylin.com>
There was a problem hiding this comment.
Code Review
This pull request introduces a new Tcl script, check_clock_tree_paths.tcl, which defines a procedure to verify that the maximum clock tree path is not below the minimum clock tree path in timing models, and integrates this check into the abstract generation process. The review feedback recommends making the regular expressions in the script more robust by allowing optional whitespace around colons and parentheses to prevent potential silent failures due to formatting variations in the timing model files.
… path
The expressions read "timing_type : " and "pin(" with exactly the
spacing write_timing_model uses today. Written with other spacing, a
path was neither read nor counted, and the model passed unchecked.
The colon and the pin group's parentheses now take any spacing, the
count of what was written looks for the bare min/max_clock_tree_path
token so a form the reader misses still stops (FLW-0031), and a path
that sits in no pin group the reader can find stops too (FLW-0032).
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Signed-off-by: Øyvind Harboe <oyvind.harboe@zylin.com>
🔍 QoR checkMetrics reflect the PR merge build — i.e. what will land on the target branch. Commit 62 design(s) checked — 0 with regression(s), 0 without a comparable baseline. |
generate_abstractstops when the timing model it just wrote has a clock pin whosemax_clock_tree_pathis below itsmin_clock_tree_path. A parent's CTS balances a hard macro by that max, so such a model clocks the block late in every parent that reads it, and nothing downstream says so: the parent's timing reads the same model and agrees with it. One block of XiangShan had max 216.3/227.4 ps against min 501.1/525.9, and its parent's CTS put the block's clock pin 245 ps late.The cause of that case is fixed in The-OpenROAD-Project/OpenSTA#481. This is the check that would have stopped it where the model was made, hours before the parent's timing looked wrong.
The change
check_clock_tree_paths.tcl: reads a written.liband stops with FLW-0030, naming the pin, the edge and both values, when a max is below its min.generate_abstract.tclruns it on each model it writes.The check is necessary, not sufficient: a bound that is wrong but on the right side of the other passes it.
Test
No new test in this PR. Run in OpenROAD on a model
write_timing_modelwrote (OpenSTA'swrite_timing_model8.ok) and on edits of it:🤖 Generated with Claude Code