[Mlir-commits] [mlir] [mlir][bytecode] Fix crashes when reading bytecode with unsupported types (PR #186354)

[Mehdi Amini]

================
@@ -0,0 +1,15 @@
+// RUN: not mlir-opt %s --test-bytecode-roundtrip="test-kind=2" 2>&1 | FileCheck %s
+
+// CHECK: expected integer or index type for IntegerAttr, but got: '!test.i32'
+// CHECK: failed to read bytecode
+
+// This test verifies that a proper error is emitted (rather than crashing with
+// an APInt assertion) when the type callback replaces an integer type with one
+// that does not implement IntegerType or IndexType.
+module {
+  func.func @combined_operations() -> () {
+    %cst = arith.constant dense<[[true, true, true], [true, false, true]]> : tensor<2x3xi1>
+    %reduction_output = tosa.reduce_all %cst {axis = 1 : i32} : (tensor<2x3xi1>) -> tensor<2x1xi1>
+    return
----------------
joker-eph wrote:

The TOSA ops add unnecessary complexity. The same path (IntegerAttr whose type gets replaced by the callback) is triggered by any `i32` attribute — `arith.constant 1 : i32` would suffice and avoids a TOSA dialect dependency.

https://github.com/llvm/llvm-project/pull/186354