[oharboe]

A placement that no legalizer search can legalize within its displacement window now fails in seconds, before the diamond search, instead of after it:

	before	after
CTS stage, hierarchical asap7 design, clock buffers in channels narrower than they needed	82 min of diamond search, then failure	stops before the first move, worst windows in microns

On a placement that legalizes today nothing changes: the check passes and the search runs as before.

Why it is a proof

The diamond search puts a cell's origin within +/- max_displacement of where it starts (legalGridPt), and rip-up-and-replace places the cells it displaces from their own starts. Legalization is a matching of cells to free sites, so by Hall's condition every set of cells needs, in the union of the sites it can cover, as many free sites as it covers. One failing set is proof. The core is tiled into boxes the size of the window; the cells starting in a box are compared with the free sites in the box grown by one window on every side, and on the high sides by the widest and tallest cell beyond that (a cell whose origin sits at the window's far edge covers sites past it). Only a cell's own sites are demand: padding may overhang the core edge or a row hole, where no cell can be.

The negotiation legalizer only gets a warning

The negotiation legalizer, the default, also searches around a cell's current position (NegotiationLegalizer::findBestLocation, the second window built at cell.x, cell.y) and nothing bounds the distance from the start, so the proof does not hold for it. Before the negotiation the same report is a warning (DPL-1108) and legalization goes on.

Messages

• DPL-1106, DPL-1109: how many windows fall short, and the worst ten with sites needed / sites reachable.
• DPL-1107 (error, diamond search): the placement cannot be legalized within the window.
• DPL-1108 (warning, negotiation): the same windows, legalization continues.

Tests

• WindowSupplyTest.DiamondStopsWhenWindowsCannotHoldTheCells: DPL-1107. Without the check the search runs and fails on its own (DPL-0036).
• WindowSupplyTest.NegotiationIsNotStoppedByTheCheck: the same placement is not stopped with the negotiation legalizer; with the check made fatal there it is.
• WindowSupplyTest.DiamondLegalizesWhenWindowsHoldTheCells: a window with room passes the check and legalizes.
• fragmented_row03 and report_failures goldens gain the warning (their cores hold 2 free sites for cells needing 7 and 8); nothing else in their logs changes.
• All dpl tests pass.

🤖 Generated with Claude Code

[gemini-code-assist]

Code Review

This pull request introduces a pre-search check, checkWindowSupply, to verify if cells can be legalized within the displacement window before starting detailed placement. It tiles the core into boxes and compares cell demand against free site supply, failing early for diamond search or warning for the negotiation legalizer if a shortfall is detected. The review feedback correctly identifies a potential integer overflow and out-of-memory risk when calculating the total number of boxes on large designs with small displacement values, and suggests using int64_t with a safety threshold check to prevent crashes.