Parameterized Machine Outlining for Similar Instruction Sequences - Shonie Caplan
Machine outlining is a backend pass which identifies exact repeated code sequences and replaces them with calls to a shared function. Differential outlining generalizes this to similar, non-identical instruction sequences by normalizing equivalent patterns and parameterizing the differences that remain. Earlier prototype differential outliners relied on sequence alignment or similarity-preserving hashing approaches, which limited them to quadratic or cubic time costs. This talk introduces a novel MIR-level method that identifies and outlines similar instruction sequences in linear time.

Optimizing Across the Differentiation Boundary with EnzymeMLIR - Vimarsh Sathia
EnzymeMLIR represents Automatic Differentiation(AD) as first-class MLIR operations rather than immediately expanding derivative calls into generated code. This talk shows how that representation enables AD-aware canonicalization, automatic batching of repeated derivative computations, and motion of inactive primal work across differentiation regions. We describe the relevant MLIR transformations, their legality and pass-ordering constraints, their integration into Reactant, and their evaluation on a Jacobian-chunk kernel extracted from a two-dimensional Brusselator application.

Hand-in-Hand-in-Hand: Let's Share LLVM-libc with Compiler-rt! - Mohamed Emad
compiler-rt and LLVM libc both implement IEEE-754 soft-float arithmetic, and the copy nobody tests is where bugs hide for years. This talk covers my GSoC project to delete one of them, making compiler-rt the fourth client of Hand-in-Hand, libc's mechanism for sharing header-only internals across LLVM projects. I walk through how we will use LLVM-libc's floating-point inside compiler-rt builtins.