Interaction between macro placement blockages, rsz repair_design, and gpl

[gudeh]

Description

Right after a non-virtual timing-driven iteration by gpl, where we keep the resizing and buffers inserted by rsz repair_design we can see instances added on top of macro placement blockages, such as the following image:

Gpl does not allow for instances on top of macros, so it moves the instances outside of the positions set by rsz, which are further away, potentially undoing rsz's work:

I wonder if we should worry about this behavior.

Suggested Solution

No response

Additional Context

No response

[maliberty]

rsz needs to be blockage aware. Right now it just uses Steiner trees and placing instances along the tree edges without such consideration. Its more of an rsz issue than a gpl one.

[maliberty]

@gudeh are you wanting to look into this or pass it to @precisionmoon ?

[gudeh]

If rsz needs to be blockage-aware, I think it would make more sense for someone with more experience in rsz to tackle this issue.

Furthermore, I wonder if this behavior has some impact on gpl divergences. Currently, I suspect that the change in area is the cause for gpl diverging.

[maliberty]

I wouldn't expect it to cause divergence but I can't say for sure.