[llvm] [LAA] Add stencil group merging to reduce runtime pointer checks (PR #187252)

David Sherwood via llvm-commits llvm-commits at lists.llvm.org
Wed Sep 16 05:24:50 PDT 2026


================
@@ -1098,11 +1128,18 @@ static std::pair<unsigned, unsigned> computeStencilMergeCost(
     }
   }
 
-  // Each not-yet-committed positive-stride predicate becomes one extra runtime
-  // check, so it counts against the saving.
+  // Charge once per stride, even when both limits need runtime predicates.
+  // Later DepSets can narrow an upper limit without adding another check.
+  const auto NeedsNewCheck = [&](const SCEV *Stride) {
+    return !StrideLowerLimits.contains(Stride) &&
+           !StrideUpperLimits.contains(Stride);
+  };
   unsigned NewPredicates = 0;
-  for (const SCEV *Stride : LocalStridesNeedingPreds)
-    if (!CommittedStridePredicates.contains(Stride))
+  for (const SCEV *Stride : LocalStrideLowerLimits)
----------------
david-arm wrote:

This logic is quite complicated and hard to follow. Would it be simpler if we had a single `LocalStrideLimits` of the form:

```
  const SmallMapVector<std::pair<const SCEV *, const SCEV *>, APInt, 4> LocalStrideLimits;
```

where the second element of the pair is nullptr if there is no upper limit? This way you only have to walk through one array.

https://github.com/llvm/llvm-project/pull/187252


More information about the llvm-commits mailing list