[mlir][linalg] Linalg.fill on tensor should not have side-effects

[nicolasvasilache]

Committing for you to speed up the integrate, thanks @ThomasRaoux !

mlir/lib/Dialect/Linalg/IR/LinalgOps.cpp

static LogicalResult verify(linalg::YieldOp op) {
   return op.emitOpError("expected parent op with LinalgOp interface");
 }
 
 /////// Operations corresponding to library calls defined with Tablegen ////////
 
 void FillOp::getEffects(
     SmallVectorImpl<SideEffects::EffectInstance<MemoryEffects::Effect>>
         &effects) {
+  if (output().getType().isa<MemRefType>())
     effects.emplace_back(MemoryEffects::Write::get(), output(),
                          SideEffects::DefaultResource::get());
 }
 
 static LogicalResult verify(FillOp op) {
   auto viewType = op.getOutputShapedType(0);
   auto fillType = op.value().getType();
   if (viewType.getElementType() != fillType)
     return op.emitOpError("expects fill type to match view elemental type");
   if (!op.getNumResults() && !viewType.isa<MemRefType>()) {

mlir/test/Dialect/Linalg/canonicalize.mlir

%1 = subtensor %0[0, 0, 0] [5, %arg1, 20] [1, 1, 1]
     : tensor<?x10x40xf32> to tensor<5x?x20xf32>
   return %1 : tensor<5x?x20xf32>
 }
 //      CHECK: func @fold_init_tensor_with_subtensor
 // CHECK-SAME:   %[[ARG0:[a-zA-Z0-9_]+]]: index
 // CHECK-SAME:   %[[ARG1:[a-zA-Z0-9_]+]]: index
 //      CHECK:   %[[T0:.+]] = linalg.init_tensor [5, %[[ARG1]], 20]
 //      CHECK:   return %[[T0]]
+
+// -----
+
+#accesses = [
+  affine_map<(i, j) -> (i, j)>
+]
+
+#trait = {
+  indexing_maps = #accesses,
+  iterator_types = ["parallel", "parallel"]
+}
+
+// CHECK-LABEL: func @dead_linalg_tensor
+//   CHECK-NOT:   linalg.fill
+//   CHECK-NOT:   linalg.matmul
+//   CHECK-NOT:   linalg.generic
+//       CHECK:   return
+func @dead_linalg_tensor(%arg0 : tensor<7x7xi32>, %arg1 : tensor<7x7xf32>) {
+  %c0_i32 = constant 0 : i32
+  %0 = linalg.fill(%arg0, %c0_i32) : tensor<7x7xi32>, i32 -> tensor<7x7xi32>
+  %1 = linalg.matmul ins(%arg1, %arg1: tensor<7x7xf32>, tensor<7x7xf32>)
+                     outs(%arg1: tensor<7x7xf32>) -> tensor<7x7xf32>
+  %2 = linalg.generic #trait outs(%arg0 : tensor<7x7xi32>) {
+  ^bb(%3: i32) :
+    linalg.yield %3 : i32
+  } -> tensor<7x7xi32>
+  return
+}