Update

Author: hasekawa-takumi

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

@@ -631,11 +631,9 @@ struct ConvertVectorMaskedLoad final
                                        *foldedIntraVectorOffset, origElements);
       }
     } else {
-      auto resultVector = rewriter.create<arith::ConstantOp>(
-          loc, op.getType(), rewriter.getZeroAttr(op.getType()));
       result = dynamicallyExtractSubVector(
-          rewriter, loc, dyn_cast<TypedValue<VectorType>>(result), resultVector,
-          linearizedInfo.intraDataOffset, origElements);
+          rewriter, loc, dyn_cast<TypedValue<VectorType>>(result),
+          op.getPassThru(), linearizedInfo.intraDataOffset, origElements);
     }
     rewriter.replaceOp(op, result);
 

mlir/test/Dialect/Vector/vector-emulate-narrow-type-unaligned.mlir

@@ -205,6 +205,8 @@ func.func @vector_maskedload_i2_dynamic_indexing_mixed(%passthru: vector<3xi2>,
 // CHECK: %[[LINEAR2:.+]] = affine.apply #map1()[%[[IDX]]]
 // CHECK: %[[ONE:.+]] = arith.constant dense<true> : vector<2xi1>
 // CHECK: %[[ZERO:.+]] = arith.constant dense<0> : vector<8xi2>
+
+// extract passthru vector, and insert into zero vector, this is for constructing a new passthru
 // CHECK: %[[EX1:.+]] = vector.extract %[[PTH]][0] : i2 from vector<3xi2>
 // CHECK: %[[IN1:.+]] = vector.insert %[[EX1]], %[[ZERO]] [%[[LINEAR2]]] : i2 into vector<8xi2>
 // CHECK: %[[C1:.+]] = arith.constant 1 : index
@@ -215,21 +217,33 @@ func.func @vector_maskedload_i2_dynamic_indexing_mixed(%passthru: vector<3xi2>,
 // CHECK: %[[INCIDX2:.+]] = arith.addi %[[LINEAR2]], %[[C2]] : index
 // CHECK: %[[EX3:.+]] = vector.extract %[[PTH]][2] : i2 from vector<3xi2>
 // CHECK: %[[IN3:.+]] = vector.insert %[[EX3]], %[[IN2]] [%[[INCIDX2]]] : i2 into vector<8xi2>
+
+// bitcast the new passthru vector to emulated i8 vector
 // CHECK: %[[BITCAST:.+]] = vector.bitcast %[[IN3]] : vector<8xi2> to vector<2xi8>
+
+// use the emulated i8 vector to masked load from the memory
 // CHECK: %[[MASKEDLOAD:.+]] = vector.maskedload %[[ALLOC]][%[[LINEAR1]]], %[[ONE]], %[[BITCAST]]
 // CHECK-SAME: memref<3xi8>, vector<2xi1>, vector<2xi8> into vector<2xi8>
+
+// bitcast back to i2 vector
 // CHECK: %[[BITCAST2:.+]] = vector.bitcast %[[MASKEDLOAD]] : vector<2xi8> to vector<8xi2>
+
 // CHECK: %[[CST1:.+]] = arith.constant dense<false> : vector<8xi1>
+
+// create a mask vector and select passthru part from the loaded vector.
+// note that if indices are known then we can fold the part generating mask.
 // CHECK: %[[EX4:.+]] = vector.extract %[[MASK]][0] : i1 from vector<3xi1>
 // CHECK: %[[IN4:.+]] = vector.insert %[[EX4]], %[[CST1]] [%[[LINEAR2]]] : i1 into vector<8xi1>
 // CHECK: %[[EX5:.+]] = vector.extract %[[MASK]][1] : i1 from vector<3xi1>
 // CHECK: %[[IN5:.+]] = vector.insert %[[EX5]], %[[IN4]] [%[[INCIDX]]] : i1 into vector<8xi1>
 // CHECK: %[[EX6:.+]] = vector.extract %[[MASK]][2] : i1 from vector<3xi1>
 // CHECK: %[[IN6:.+]] = vector.insert %[[EX6]], %[[IN5]] [%[[INCIDX2]]] : i1 into vector<8xi1>
+
 // CHECK: %[[SELECT:.+]] = arith.select %[[IN6]], %[[BITCAST2]], %[[IN3]] : vector<8xi1>, vector<8xi2>
-// CHECK: %[[CST2:.+]] = arith.constant dense<0> : vector<3xi2>
+
+// finally, insert the selected parts into actual passthru vector.
 // CHECK: %[[EX7:.+]] = vector.extract %[[SELECT]][%[[LINEAR2]]] : i2 from vector<8xi2>
-// CHECK: %[[IN7:.+]] = vector.insert %[[EX7]], %[[CST2]] [0] : i2 into vector<3xi2>
+// CHECK: %[[IN7:.+]] = vector.insert %[[EX7]], %[[PTH]] [0] : i2 into vector<3xi2>
 // CHECK: %[[EX8:.+]] = vector.extract %[[SELECT]][%[[INCIDX]]] : i2 from vector<8xi2>
 // CHECK: %[[IN8:.+]] = vector.insert %[[EX8]], %[[IN7]] [1] : i2 into vector<3xi2>
 // CHECK: %[[EX9:.+]] = vector.extract %[[SELECT]][%[[INCIDX2]]] : i2 from vector<8xi2>