Summary

Improve tile-fusion planning by exploring competing fusion groups instead of committing to the first acceptable DFG path.

Key changes

• Add whole-group cost evaluation alongside existing append legality checks.

• Reuse shared cost computation for append and group evaluation.

• Explore multiple candidate groups, retaining up to 64 candidates per group size.

• Rank candidates by:

  1. Whole-group cost.
  2. Group size.
  3. Stable block/node order.

• Preserve the existing append acceptance rules and fusion metadata contract.

• Add regression coverage for:

  • Selecting a higher-value multi-step path.
  • Stable equal-cost tie-breaking.
  • Formation of subsequent independent fusion groups.

Testing

llvm-lit -sv build/test/lit
--filter 'fusion_plan_cost_guided_(search|tie)'