llvm-project

342de06a - [Reassociate] Distribute multiply over add to enable factorization (#178201)

[Reassociate] Distribute multiply over add to enable factorization (#178201) ### This patch improves ReassociatePass to handle patterns like: (x*C1) - ((y+x)*C2) → x*(C1-C2) - (y*C2) The optimization consists of two changes: 1. Distribution pre-processing: Transform (A+B)*C → A*C + B*C when: - The add has exactly one use (avoids code bloat) - Both add operands are non-constant (avoids unprofitable cases) This exposes common factors that would otherwise be hidden inside the addition, enabling subsequent factorization. 2. Factorization heuristic: Prefer extracting non-constant factors (Instructions/Arguments) over constant factors when occurrence counts are equal. This enables better constant folding opportunities. Note: undef is excluded from this preference to maintain existing test expectations. Example transformation: Input IR: ``` %add = add nsw i16 %y, %x %mul1 = mul nsw i16 %x, 8697 %mul2 = mul nsw i16 %add, 6436 %sub = sub nsw i16 %mul1, %mul2 ret i16 %sub ``` After distribution: ``` %1 = mul nsw i16 %y, 6436 %2 = mul nsw i16 %x, 6436 %3 = add nsw i16 %1, %2 %sub = sub nsw i16 %mul1, %3 ``` After reassociation (factors %x, folds 8697-6436): ``` %neg = mul i16 %y, -6436 %reass.mul = mul i16 %x, 2261 %result = add i16 %reass.mul, %neg ``` The transformation preserves nsw/nuw flags during distribution. Note that flags may be dropped by subsequent reassociation passes; comprehensive flag preservation can be addressed in future work. Fixes #167190