[MLIR][Affine] Fixed crash with invalid reduction op (Issue #64073)

[SwapnilGhanshyala]

Updated AffineParallelLowering to check if reduction op value being returned is valid else return failure in matchAndRewrite Updated ArithOps.cpp getIdentityValue method to return nullptr if op is not a valid reduction op Code cleanup in ArithOps.cpp getReductionOp method; removed cases maxnumf and minnumf as not valid reduction ops

Reporting Issue is llvm#64073

[SwapnilGhanshyala]

This crash occurred when getIdentityValue tries to build a value based on the 'attr' recieved, which is a nullptr if the op is not a valid reduction op. In llvm#64073 the reduction op being passed is 'assign' which although is atomicRMW but not reduction op.
The assert in AffineParallelLowering which was meant to catch this invalid reduction op before it was passed to getIdentityValue, fails. This is because the AtomicRMWKind enum/list of ops that was checked for the validity of reduction op. This enum returns 'assign' as a valid reduction op, therefore, the assert passes in the receiver.

In the current fix, the getIdentityValue returns a nullptr instead of building a value if the op is invalid.

Another proposed fix could be to have a separate enum like 'AtomicRMWReductionKind' to hold only reduction ops of atomicRMW kind. It would be easier to work with reduction ops using this enum.
Currently the approach is identify AtomicRMWKind ops and use 'getIdentityValue' as a rough filter.

[bondhugula]

Why is this deleted?

[SwapnilGhanshyala]

Both maxnumf and minnumf are invalid reduction ops and should not be a valid case in getReductionOps.

[bondhugula]

I see.

[SwapnilGhanshyala]

Changes addressed :

1. variable name readability
2. unrelated clang-formatting

[SwapnilGhanshyala]

Addressing comment on lines 2528 to 2530 regarding deletion of maxnumf and minnumf as reduction op cases:

The 'getIdentityAttr()' method called here returns valid attr iff the op passed is a valid reductionOp else returns nullptr.
This method does not consider 'maxnumf' and 'minnumf' as valid reduction ops. If this is correct, then consequently maxnumf and minnumf cases should be safely removable from the "getReductionOp()", as done in this commit.