Constant folding for bit casts between word and float types

Summary

Bit casts between word and float types are not constant folded.

Steps to reproduce

import GHC.Float
import GHC.Word

foo :: Word64
foo = castDoubleToWord64 1.0   

❯ ghc-9.8.1 -ddump-simpl -dno-typeable-binds -fforce-recomp -O2 Test.hs
[1 of 1] Compiling Test             ( Test.hs, Test.o )

==================== Tidy Core ====================
Result size of Tidy Core
  = {terms: 7, types: 2, coercions: 0, joins: 0/0}

-- RHS size: {terms: 6, types: 1, coercions: 0, joins: 0/0}
foo :: Word64
[GblId,
 Unf=Unf{Src=<vanilla>, TopLvl=True,
         Value=False, ConLike=False, WorkFree=False, Expandable=False,
         Guidance=IF_ARGS [] 40 10}]
foo
  = case {__ffi_static_ccall_safe base:stg_doubleToWord64zh "stg_doubleToWord64zh" :: Double#
                                                                             -> Word64#}_aD3
           1.0##
    of ds1_aD6
    { __DEFAULT ->
    GHC.Word.W64# ds1_aD6
    }

Expected behavior

Core dump should contain:

foo = 4607182418800017408

Also we could use inline assembly instead of a call to a Cmm function (e.g. MOVD instruction on x86-64).

What I would do:

• Move these bitcast Cmm functions from base:GHC.Float to ghc-prim Not needed now: already in ghc-internal
• Replace foreign imports with real primops that are lowered to these Cmm calls
• Implement constant folding for the new primops when they are applied to literals
• Lower these primops to inline assembly in the native code generators (x86-64, aarch64...)
  • NCG x86-64
  • NCG aarch64
  • LLVM
  • C backend?

Environment

• GHC version used: 9.8.1