Complete support for swapping to uninstantiated modules

[openroad-ci]

1. Enabled OpenSTA linkNetwork change to defer module deletion
2. Enabled replace_design4 test
3. Incorporated code review comments from Andy Fox and Matt Liberty

[github-actions]

clang-tidy review says "All clean, LGTM! 👍"

[oharboe]

This sounds interesting, what is the use-case?

Is there a test-case(doubles as documentation)?

Will there be a test in ORFS to demonstrate the use-case for this?

[QuantamHD]

It's part of a feature to swap different operator implementations. Yosys will provide OR N implementations of an add operator and OR will choose the best one depending on PPA requirements.

[oharboe]

It's part of a feature to swap different operator implementations. Yosys will provide OR N implementations of an add operator and OR will choose the best one depending on PPA requirements.

So if I understand:

late instantiation of addition algorithms and standard cell elements when there is information available as to which one is best.

The alternative is impractical: to specify addition algorithms manually. It can vary across addition usage in the design and there can be a great many additions performed and no single solution is best everywhere.

[QuantamHD]

It's part of a feature to swap different operator implementations. Yosys will provide OR N implementations of an add operator and OR will choose the best one depending on PPA requirements.

So if I understand:

late instantiation of addition algorithms and standard cell elements when there is information available as to which one is best.

The alternative is impractical: to specify addition algorithms manually. It can vary across addition usage in the design and there can be a great many additions performed and no single solution is best everywhere.

You got it! There's no best adder, just the best tradeoff. It's essentially just another dimension resizer has to play with, but instead of single cells you get entire modules.

[maliberty]

There will eventually be usage in ORFS when this is fully automated in timing optimization.

[github-actions]

clang-tidy review says "All clean, LGTM! 👍"

[precisionmoon]

Small comments on the swapMaster which I was asked to review (not sure if my comments on that pull request will still exist as pull request accepted).

@andyfox-rushc , yes I took care of your comments in the previous PR.

Please check that after swapping the master the modbterms reference the moditerms in the instance. (SwapMaster)
This is done using setParentModITerm on each of the new master ModBTerms.

I'm already using setParentModITerm in dbModule::copyModuleBoundaryIO()
if (new_mod_bterm) {
old_mod_iterm->setChildModBTerm(new_mod_bterm);
new_mod_bterm->setParentModITerm(old_mod_iterm);

Is there more places where I should be using the API?:

I read through your code on the SwapMaster and see that you correctly update the flat connections (traverse iterms for each modNet pair (new and old), get the original flat net disconnect from original iterm and then connect to the new iterm). Good work. I respectfully suggest adding a test case in which new ModNet has (a) Different number of iterms. (b) Possibly none (a wire running through the new module or an output resolved to a constant). I see no reason why (a) and (b) won't work but I know that sometimes when replacing IP blocks we see "through wires" and "constants"

andy

OK. I'll add the new test cases, perhaps as a separate PR later.

Thanks.

[github-actions]

clang-tidy review says "All clean, LGTM! 👍"