[Synth] Add structural hashing pass for AIG/MIG operations #8962
Conversation
This commit implements a domain-specific structural hashing (CSE) pass for Synth dialect operations (AIG/MIG), providing more aggressive optimization than MLIR's general CSE pass. The pass performs structural hashing for synth.aig.and_inv and synth.mig.maj_inv operations, with operand sorting based on IR postions for better canonicalization and inversion-aware hashing that properly handles AIG inversion flags.
| /// Maps inverted values to their non-inverted equivalents for propagation. | ||
| DenseMap<Value, Value> inversion; |
There was a problem hiding this comment.
Nit: struggled to understand what is meant by "for propagation" - would it be possible to add a small example in the comment?
There was a problem hiding this comment.
LGTM - cannot say I've fully understood how the algorithm works so will have to read the papers from the field. Only comment would be to see if there's a way to provide examples to guide understanding - although the tests are very useful to see the capabilities.
Non-Blocking: As the available passes grow - would be good to expand documentation to understand the capabilities of different passes?
| // Topologically sort target ops within the block. | ||
| return !isa<circt::synth::aig::AndInverterOp, | ||
| circt::synth::mig::MajorityInverterOp>(op); |
There was a problem hiding this comment.
Nit: Comment doesn't seem to match function?
| pm.addPass(createStructuralHash()); | ||
| pm.addPass(createSimpleCanonicalizerPass()); |
There was a problem hiding this comment.
Q: is there an argument for running strash after canonicalization - does canonicalization expose more opportunities? e.g. constant folding 1 & x -> x?
There was a problem hiding this comment.
Yes, strash would help canonicalization (vice-versa). I added another run of strash after canonicalization.
|
Thank you for the review! Section 2.3 in https://people.eecs.berkeley.edu/~alanmi/publications/2005/tech05_fraigs.pdf briefly mentions Structural hashing. The current implementation probably corresponds to the most basic one-level structural hashing. This is still stronger than MLIR's CSE pass, e.g. MLIR one doesn't use commutativity (a+b and b+a are not CSE'd). https://circt.llvm.org/docs/Passes/ has a list of passes in CIRCT FYI (Synth was missing but should appear tomorrow f6b70b9). Also I'll describe Synth rational doc. |
This commit implements a domain-specific structural hashing (CSE) pass for Synth dialect operations (AIG/MIG), providing more aggressive optimization than MLIR's general CSE pass, e.g:
The pass performs structural hashing for synth.aig.and_inv and synth.mig.maj_inv operations, with operand sorting based on IR positions for better canonicalization and inversion-aware hashing that properly handles AIG/MIG inversion flags.
Basically this corresponds to ABC's
strashcommand.