Issue in conversion from index to i64

MLIR
1

Has there been any change in the way the index type gets converted to i64 type? Should this now be explicitly performed by a cast operation?

Previously an FIR to LLVM conversion pattern worked. The constants 3 and 5 are of index type in the input FIR code. The conversion llvm.sub operation in the pattern expects i64 type. The conversion produced LLVM dialect code like the following.

    %2 = llvm.mlir.constant(3 : index) : i64
    %4 = llvm.mlir.constant(5 : index) : i64

    %57 = llvm.sub %4, %2  : i64
    %58 = llvm.mul %57, %37  : i64

But now I get the following error.
error: ‘llvm.sub’ op requires the same type for all operands and results
note: see current operation: %63 = “llvm.sub”(%8, %4) : (index, index) → i64

    %3 = "llvm.mlir.constant"() {value = 3 : index} : () -> i64
    %4 = "builtin.unrealized_conversion_cast"(%3) : (i64) -> index
    %7 = "llvm.mlir.constant"() {value = 5 : index} : () -> i64
    %8 = "builtin.unrealized_conversion_cast"(%7) : (i64) -> index
...
    %63 = "llvm.sub"(%8, %4) : (index, index) -> i64
    %64 = "llvm.mul"(%63, %43) : (i64, i64) -> i64

Any suggestions? Should we use an index_cast or unrealized_conversion_cast explicitly for these kinds of conversions? Is there any utility function that does this?

2

What does your pattern look like? There is a pretty good chance that whatever is generating the llvm.sub is not using the updated conversion operands.

– River

1 Like
3

Thanks for the hint. From a quick glance that is probably the issue. I will confirm tomorrow.

4

Thanks. Confirmed.