[llvm] [LV] Drop prefetches from vector loops (PR #226895)

[via llvm-commits]

llvmorg-github-actions[bot] wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-vectorizers

Author: Dávid Bolvanský (davidbolvansky)

<details>
<summary>Changes</summary>

`llvm.prefetch` can currently prevent an otherwise vectorizable loop from being vectorized. For example:

```c
void add(int *restrict a, int *restrict b, int n) {
  for (int i = 0; i < n; ++i) {
    a[i] += b[i];
    __builtin_prefetch(&b[i + 8]);
  }
}
```

Permit LoopVectorize to omit prefetches from the vector loop. Ignore them during memory dependence and legality checks, and remove their VPlan recipes before call-widening decisions. VPlan DCE then removes address calculations used only by the hint. The scalar remainder retains its original prefetch.

This is sound because `llvm.prefetch` has no observable behavior and cannot trap, so removing it does not change program semantics. Conditional prefetches can be handled in the same way.

---
Full diff: https://github.com/llvm/llvm-project/pull/226895.diff


4 Files Affected:

- (modified) llvm/lib/Analysis/LoopAccessAnalysis.cpp (+7) 
- (modified) llvm/lib/Transforms/Vectorize/LoopVectorizationLegality.cpp (+9) 
- (modified) llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp (+15) 
- (added) llvm/test/Transforms/LoopVectorize/prefetch.ll (+161) 


``````````diff
The server is unavailable at this time. Please wait a few minutes before you try again.
``````````

</details>


https://github.com/llvm/llvm-project/pull/226895