[Mlir-commits] [mlir] [mlir][linalg] convert arith ops to destination-passing-style. (PR #157854)

[Andrzej Warzyński]

================
@@ -603,6 +612,56 @@ Value linalg::bufferizeToAllocation(
 }
 
 namespace {
+template <typename OpTy>
+FailureOr<Operation *>
+rewriteArithInDestinationPassingStyle(RewriterBase &rewriter, OpTy op) {
+  // reject ops such as `arith.constant` and `arith.select`.
+  auto numOperands = op->getNumOperands();
+  if (numOperands == 0 || numOperands > 2)
+    return failure();
+
+  // destination passing style rewrite is only for ops on tensor types.
+  Type resultType = op->getResult(0).getType();
+  auto tensorType = dyn_cast<RankedTensorType>(resultType);
+  if (!tensorType)
+    return failure();
+
+  auto loc = op.getLoc();
+  OpBuilder::InsertionGuard g(rewriter);
+  auto dynSizes = reifyOrComputeDynamicSizes(rewriter, op->getOperand(0));
+
+  // Create tensor.empty.
+  Value empty = tensor::EmptyOp::create(rewriter, loc, resultType, dynSizes);
----------------
banach-space wrote:

[nit] Use more descriptive variable name rather than `empty` to repeats _how_ the variable is defined and that's available on LHS (i.e. `tensor::EmptyOp::create`). Perhaps `init` or `outs`? Or `initForGeneric`?

https://github.com/llvm/llvm-project/pull/157854