[Mlir-commits] [mlir] [mlir][arith] Restrict mulsi_extended(x, 1) rewrite to integer types (PR #233234)

[Prakhar Dixit]

Prakhar-Dixit wrote:

> I think we should consider disabling extended ops with index types in the first place -- I can't imagine how this could be used correctly

That makes complete sense. Because the width of index is unknown until lowering, computing a high half or a carry/overflow bit at the IR level is inherently ambiguous.

This causes silent miscompiles today. The folder assumes a 64-bit index, so:
```
%c = arith.constant 65536 : index
%low, %high = arith.mului_extended %c, %c : index
```
folds to high = 0 with --canonicalize. But lowering the same IR to SPIR-V first, where index becomes i32 by default, gives high = 1 (--convert-arith-to-spirv --canonicalize).

Disallowing index here would also be consistent with #121455, which removed index support from arith.bitcast, another op that is only defined for a fixed bit width.

https://github.com/llvm/llvm-project/pull/233234