[llvm] SelectionDAG: Use qNaN constant if FCANONICALIZE not LegalOrCustom (PR #104564)

[Pawan Nirpal via llvm-commits]

pawan-nirpal-031 wrote:

> @pawan-nirpal-031 Can we just mul * 1.0 as the default expansion for ISD::FCANONICALIZE, adding a TargetLowering::expandFCANONICALIZE method?

Hi SImon, Yes we could, I was doing this in instruction selection, and canonicalizing subnormals in legalizer phase, ftz and stuff, replacing snans by qnans, undefs to qnans. I had plans to handle these cases along with mul * 1.0 too. But I wanted to know if this PR does that, if It does than my PR may not be required. My PR aims to handle this intrinsic for constants/variables and subnormals. 

https://github.com/llvm/llvm-project/pull/104564