llvm-project

4bd80c56 - [AArch64] Fix regression from “Fold scalar-to-vector shuffles into DUP/FMOV

[AArch64] Fix regression from “Fold scalar-to-vector shuffles into DUP/FMOV This patch aims to fix the original compile time regression by restricting the optimisation to run only on non-constant splats. Without the guard, an infinite loop is caused because the CONCAT(SCALAR_TO_VECTOR, zero) folds back into the same BUILD_VECTOR and immediately re-enters LowerBUILD_VECTOR. This patch was tested with the original TensorFlow reproduction provided on the PR and shows a (very) slight improvement on compile-time.