[mlir][Vector] Do not propagate vector.extract on dynamic position (#148245)

Propagating vector.extract when a dynamic position is present can cause
dominance issues and needs better handling. For now, disable propagation
if there is a dynamic position present.

Author: Groverkss

mlir/lib/Dialect/Vector/Transforms/VectorTransforms.cpp

@@ -1088,6 +1088,12 @@ class ExtractOpFromElementwise final
     if (!llvm::all_equal(eltwise->getOperandTypes()))
       return rewriter.notifyMatchFailure(op, "operand types are different");
 
+    // Dynamic position can cause dominance issues, so conservatively fail for
+    // now.
+    if (!op.getDynamicPosition().empty())
+      return rewriter.notifyMatchFailure(
+          op, "dynamic position not yet implemented");
+
     Type dstType = op.getType();
 
     OpBuilder::InsertionGuard g(rewriter);

mlir/test/Dialect/Vector/vector-sink.mlir

@@ -514,6 +514,18 @@ func.func @negative_extract_vec_fma(%arg0: vector<4xf32>, %arg1: vector<4xf32>,
   return %1 : f32
 }
 
+// CHECK-LABEL: @negative_extract_dynamic_pos
+func.func @negative_extract_dynamic_pos(%arg0: vector<4xf32>, %arg1 : vector<4xf32>, %idx : vector<4xindex>) -> f32 {
+  // CHECK-NOT: vector.extract
+  // CHECK: arith.addf %{{.*}}, %{{.*}} : vector<4xf32>
+  // CHECK: vector.extract
+  // CHECK: vector.extract
+  %0 = arith.addf %arg0, %arg1 : vector<4xf32>
+  %1 = vector.extract %idx[0] : index from vector<4xindex>
+  %2 = vector.extract %0[%1] : f32 from vector<4xf32>
+  return %2 : f32
+}
+
 //-----------------------------------------------------------------------------
 // [Pattern: ExtractOpFromLoad]
 //-----------------------------------------------------------------------------