[CombToDatapath] Lower comb::SubOp #9053
Conversation
Adds support for lowering comb::SubOp in the CombToDatapath conversion pass by refactoring the SubOp-to-AddOp conversion into a shared utility function that can be reused across multiple conversion passes. This commit extracts the subtraction lowering logic from CombToSynth into a dialect-agnostic utility function `comb::convertSubToAdd` that transforms `sub(lhs, rhs)` into `add(lhs, ~rhs, 1)` using two's complement representation. AIG logic depth improved 29 -> 24 for i128 subtraction
There was a problem hiding this comment.
LGTM - thanks for embracing the compressors! I think this is a sensible step. But not entirely sure on the best long term strategy.
I've been dwelling on this for some time as currently the compress operator is unaware of constant ones. The problem is that each adder can accommodate a carry-in (at no logic depth cost) so we don't really need a compress operator at all here - but what we do need is an adder that can represent the carry-in - this is a bigger discussion (and certainly there may be other ways) so non-blocking as this step is clearly sensible!
| auto notRhs = | ||
| comb::createOrFoldNot(subOp.getLoc(), rhs, rewriter, subOp.getTwoState()); |
There was a problem hiding this comment.
Neat - did no realise there was a createOrFoldNot() function - will stop writing create constant one and Xor... Even though it presumably maps to this
Yeah that makes sense. That sounds like clever optimization in DatapathToComb. |
2 participants
Adds support for lowering comb::SubOp in the CombToDatapath conversion pass by refactoring the SubOp-to-AddOp conversion into a shared utility function that can be reused across multiple conversion passes. This change makes SubOp it possible to leverage datapath optimizations
This commit extracts the subtraction lowering logic from CombToSynth into a dialect-agnostic utility function
comb::convertSubToAddthat transformssub(lhs, rhs)intoadd(lhs, ~rhs, 1)using two's complement representation.AIG logic depth improved 29 -> 24 for i128 subtraction