llvm-project

ad5a7609 - [SLP]Do not cache sentinel position for SplitVectorize nodes

[SLP]Do not cache sentinel position for SplitVectorize nodes setInsertPointAfterBundle records LastInstructionToPos[LastInst] = Res so later bundles reuse the sentinel as an insertion point. For a SplitVectorize bundle this entry is stale: operand entries are materialized lazily via recursive vectorizeTree calls and can land in the IR after the cached sentinel. A later bundle that hits the cache then inserts its shuffles ahead of values they consume, tripping the verifier with "Instruction does not dominate all uses". Fixes #193919 Reviewers: Pull Request: https://github.com/llvm/llvm-project/pull/194038