[mlir][linalg] Treat quant dialect type as unsupported in named conversion

[Lewuathe]

Summary

Since the tosa-to-linalg conversion does not support the quant dialect type, we can treat it as unsupported instead of crash. Issue was reported https://github.com/llvm/llvm-project/issues/62367

[springerm]

Nit: Such test cases are usually named ...-invalid.mlir, so tosa-to-linalg-invalid.mlir or just invalid.mlir would be consistent with the other tests.

mlir/lib/Conversion/TosaToLinalg/TosaToLinalgNamed.cpp

LogicalResult matchAndRewrite(tosa::AvgPool2dOp op,
     SmallVector<Value> dynamicDims = *dynamicDimsOr;
 
     // Apply padding as necessary.
     llvm::SmallVector<int64_t> pad;
     pad.resize(2, 0);
     llvm::append_range(pad, op.getPad());
     pad.resize(pad.size() + 2, 0);
     TypedAttr padAttr = rewriter.getZeroAttr(inElementTy);
+    // Unsupported element type
+    if (!padAttr)
+      return failure();
     Value paddedInput = applyPad(loc, input, pad, padAttr, rewriter);
 
     auto initialAttr = rewriter.getZeroAttr(accETy);
     Value initialValue = rewriter.create<arith::ConstantOp>(loc, initialAttr);
 
     ArrayRef<int64_t> kernel = op.getKernel();
     ArrayRef<int64_t> stride = op.getStride();
 

mlir/test/Conversion/TosaToLinalg/tosa-to-linalg-invalid.mlir

+// RUN: mlir-opt --split-input-file -pass-pipeline="builtin.module(func.func(tosa-to-linalg-named))" %s -verify-diagnostics
+
+// CHECK-LABEL: @avg_pool2d_with_unsupported_quant_type
+func.func @avg_pool2d_with_unsupported_quant_type(%arg0: tensor<1x7x7x9x!quant.uniform<i8:f32, 0.01>>) -> tensor<1x7x7x9x!quant.uniform<i8:f32, 0.01>> {
+  // expected-error@+1 {{failed to legalize operation 'tosa.avg_pool2d'}}
+  %0 = "tosa.avg_pool2d"(%arg0) {acc_type = i32, kernel = array<i64: 2, 2>, pad = array<i64: 0, 1, 0, 1>, stride = array<i64: 1, 1>} : (tensor<1x7x7x9x!quant.uniform<i8:f32, 0.01>>) -> tensor<1x7x7x9x!quant.uniform<i8:f32, 0.01>>
+  return %0 : tensor<1x7x7x9x!quant.uniform<i8:f32, 0.01>>
+}