Don't drop in_bounds when vector-transfer-collapse-inner-most-dims

[asaadaldien]

Summary

When operand is a subview we don't infer in_bounds and some default cases (e.g case in the tests) will crash with operand is NULL when converting to LLVM

[ThomasRaoux]

mlir/lib/Dialect/Vector/VectorTransforms.cpp

3577–3583

nit: you could also do this:

ArrayAttr inBounds =  readOp.in_bounds() ? rewriter.getArrayAttr( readOp.in_boundsAttr().getValue().drop_back(dimsToDrop))
            : ArrayAttr()

[ThomasRaoux]

mlir/test/Dialect/Vector/vector-transfer-collapse-inner-most-dims.mlir

34–35

please remove the extra lines

[asaadaldien]

comments

mlir/lib/Dialect/Vector/VectorTransforms.cpp

if (srcType.getLayout().getAffineMap().isIdentity()) {
       resultMemrefType = MemRefType::get(
           srcType.getShape().drop_back(dimsToDrop), srcType.getElementType(),
           map, srcType.getMemorySpaceAsInt());
     }
 
     auto loc = readOp.getLoc();
     SmallVector<int64_t> offsets(srcType.getRank(), 0);
     SmallVector<int64_t> strides(srcType.getRank(), 1);
+
+    ArrayAttr inBounds =
+        readOp.in_bounds()
+            ? rewriter.getArrayAttr(
+                  readOp.in_boundsAttr().getValue().drop_back(dimsToDrop))
+            : ArrayAttr();
     Value rankedReducedView = rewriter.create<memref::SubViewOp>(
         loc, resultMemrefType, readOp.source(), offsets, srcType.getShape(),
         strides);
+    auto permMap = getTransferMinorIdentityMap(
+        rankedReducedView.getType().cast<ShapedType>(), resultTargetVecType);
     Value result = rewriter.create<vector::TransferReadOp>(
         loc, resultTargetVecType, rankedReducedView,
-        readOp.indices().drop_back(dimsToDrop));
+        readOp.indices().drop_back(dimsToDrop), permMap, readOp.padding(),
+        inBounds);
     rewriter.replaceOpWithNewOp<vector::ShapeCastOp>(readOp, targetType,
                                                      result);
-
     return success();
   }
 };
 
 void mlir::vector::populateVectorMaskMaterializationPatterns(
     RewritePatternSet &patterns, bool indexOptimizations) {
   patterns.add<VectorCreateMaskOpConversion,
                MaterializeTransferMask<vector::TransferReadOp>,

mlir/test/Dialect/Vector/vector-transfer-collapse-inner-most-dims.mlir

func @contiguous_inner_most_dim(%A: memref<16x1xf32>, %i:index, %j:index) -> (vector<8x1xf32>) {
   return %1 : vector<8x1xf32>
 }
 //      CHECK: func @contiguous_inner_most_dim(%[[SRC:.+]]: memref<16x1xf32>, %[[I:.+]]: index, %[[J:.+]]: index) -> vector<8x1xf32>
 //      CHECK:   %[[SRC_0:.+]] = memref.subview %[[SRC]]
 // CHECK-SAME:     memref<16x1xf32> to memref<16xf32>
 //      CHECK:   %[[V:.+]] = vector.transfer_read %[[SRC_0]]
 //      CHECK:   %[[RESULT]] = vector.shape_cast %[[V]] : vector<8xf32> to vector<8x1xf32>
 //      CHECK:   return %[[RESULT]]
+
+// -----
+
+func @contiguous_inner_most_dim_bounds(%A: memref<1000x1xf32>, %i:index, %ii:index) -> (vector<4x1xf32>) {
+  %c0 = arith.constant 0 : index
+  %cst = arith.constant 0.0 : f32
+  %0 = memref.subview %A[%i, 0] [40, 1] [1, 1] : memref<1000x1xf32> to memref<40x1xf32, affine_map<(d0, d1)[s0] -> (d0 + s0 + d1)>>
+  %1 = vector.transfer_read %0[%ii, %c0], %cst {in_bounds = [true, true]} : memref<40x1xf32, affine_map<(d0, d1)[s0] -> (d0 + s0 + d1)>>, vector<4x1xf32>
+  return %1 : vector<4x1xf32>
+}
+//      CHECK: func @contiguous_inner_most_dim_bounds(%[[SRC:.+]]: memref<1000x1xf32>, %[[II:.+]]: index, %[[J:.+]]: index) -> vector<4x1xf32>
+//      CHECK:   %[[SRC_0:.+]] = memref.subview %[[SRC]]
+//      CHECK:   %[[SRC_1:.+]] = memref.subview %[[SRC_0]]
+//      CHECK:   %[[V:.+]] = vector.transfer_read %[[SRC_1]]
+// CHECK-SAME:       {in_bounds = [true]}
+// CHECK-SAME:       vector<4xf32>