[Mlir-commits] [clang] [llvm] [mlir] [openmp] [LoopTiling][Clang][MLIR] Canonical Intra-tile Loops (PR #191114)

[Michael Kruse]

Meinersbur wrote:

> But one concern, aren't we over-complicating this by saving execution of remainder iterations in the last tile?

LoopVectorize will require if-conversion/masked execution with such loop bodies which significantly hampers all executions (not just the remainder iterations). Since we do tiling for performance, we cannot make the generated code run slower than without tiling.

If LLVM is not able to optimize it away, another option is to emit the predicate, but optimize it to an smin expression at Clang's CodeGen.

https://github.com/llvm/llvm-project/pull/191114