[Mlir-commits] [mlir] [MLIR][Affine] Guard maximal slice mismatch (PR #206236)

[llvmlistbot at llvm.org]

llvmorg-github-actions[bot] wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-mlir-affine

Author: lianjinfeng2003 (mygitljf)

<details>
<summary>Changes</summary>

I added a conservative guard for affine loop fusion when maximal slice analysis reaches constraints with incompatible Presburger spaces. In that case the pass now treats maximality as unknown instead of continuing into the path that asserts.
Added a regression with a dynamic producer bound to cover the crash path while keeping the expected fusion behavior for the independent constant-bound producer.
Fixes #<!-- -->206081

---
Full diff: https://github.com/llvm/llvm-project/pull/206236.diff


2 Files Affected:

- (modified) mlir/lib/Dialect/Affine/Analysis/Utils.cpp (+3) 
- (modified) mlir/test/Dialect/Affine/loop-fusion-4.mlir (+22) 


``````````diff
diff --git a/mlir/lib/Dialect/Affine/Analysis/Utils.cpp b/mlir/lib/Dialect/Affine/Analysis/Utils.cpp
index cac305df8ba75..1a0dfa95e205d 100644
--- a/mlir/lib/Dialect/Affine/Analysis/Utils.cpp
+++ b/mlir/lib/Dialect/Affine/Analysis/Utils.cpp
@@ -1142,6 +1142,9 @@ std::optional<bool> ComputationSliceState::isMaximal() const {
     // computed so we don't know if the slice is maximal.
     return std::nullopt;
 
+  if (!srcConstraints.getSpace().isCompatible(sliceConstraints.getSpace()))
+    return std::nullopt;
+
   // Compute the difference between the src loop nest and the slice integer
   // sets.
   PresburgerSet srcSet(srcConstraints);
diff --git a/mlir/test/Dialect/Affine/loop-fusion-4.mlir b/mlir/test/Dialect/Affine/loop-fusion-4.mlir
index cf530016c201a..9ce1cdb203ae0 100644
--- a/mlir/test/Dialect/Affine/loop-fusion-4.mlir
+++ b/mlir/test/Dialect/Affine/loop-fusion-4.mlir
@@ -831,6 +831,28 @@ func.func @fusion_non_constant_bounds_1(%N: index, %M: memref<?xf32>, %cst: f32)
   return
 }
 
+// PRODUCER-CONSUMER-MAXIMAL-LABEL: func @maximal_fusion_dynamic_source_bound
+func.func @maximal_fusion_dynamic_source_bound(%N: index, %M: memref<100xf64>) {
+  %cst = arith.constant 0.000000e+00 : f64
+  affine.for %i = 0 to 10 {
+    affine.store %cst, %M[%i] : memref<100xf64>
+  }
+  affine.for %i = 0 to %N {
+    affine.store %cst, %M[%i] : memref<100xf64>
+  }
+  affine.for %i = 0 to 10 {
+    affine.load %M[%i] : memref<100xf64>
+  }
+  return
+}
+// PRODUCER-CONSUMER-MAXIMAL:      affine.for %{{.*}} = 0 to %{{.*}} {
+// PRODUCER-CONSUMER-MAXIMAL-NEXT:   affine.store
+// PRODUCER-CONSUMER-MAXIMAL-NEXT: }
+// PRODUCER-CONSUMER-MAXIMAL:      affine.for %{{.*}} = 0 to 10 {
+// PRODUCER-CONSUMER-MAXIMAL-NEXT:   affine.store
+// PRODUCER-CONSUMER-MAXIMAL-NEXT:   affine.load
+// PRODUCER-CONSUMER-MAXIMAL-NEXT: }
+
 // No fusion here as the cost models computing slice costs run out of 64-bit precision.
 
 // PRODUCER-CONSUMER-LABEL: func @high_trip_count

``````````

</details>


https://github.com/llvm/llvm-project/pull/206236