Masked Vectorization & Handling of Conditionals
- Add a `PsMaskedBlock` IR AST class to `ast.vector`, to model masked code blocks where only a part of all vector lanes are active
- Add `PsVecAny` boolean reduction and `PsVecBlend` masked blending operation to `ast.vector`
- Implement transformation of conditionals and loops into the `AstVectorizer` using `PsMaskedBlock` for their bodies
- Implement treatment of trailing iterations using a masked block into the `LoopVectorizer`
- Upgrade `SelectIntrinsics` to detect `PsMaskedBlock`s and pass on the active mask to the platform for selection of masked intrinsics
- Add masked intrinsics to available vector platforms
- The granularity of masking can be decided independently for each platform: Only operations with effects visible to the outside really *need* to be masked, such as memory accesses and variable blends.
- Add extensive test cases for masked vectorization
task