Note [RuntimeRep and PrimRep]

GHC/Types/RepType.hs:363 compiler

This Note describes the relationship between GHC.Types.RuntimeRep
(of levity/representation polymorphism fame) and GHC.Core.TyCon.PrimRep,
as these types are closely related.

A "primitive entity" is one that can be
 * stored in one register
 * manipulated with one machine instruction


Examples include:
 * a 32-bit integer
 * a 32-bit float
 * a 64-bit float
 * a machine address (heap pointer), etc.
 * a quad-float (on a machine with SIMD register and instructions)
 * ...etc...

The "representation or a primitive entity" specifies what kind of register is
needed and how many bits are required. The data type GHC.Core.TyCon.PrimRep
enumerates all the possibilities.

data PrimRep
  = LiftedRep     -- ^ Lifted pointer
  | UnliftedRep   -- ^ Unlifted pointer
  | Int8Rep       -- ^ Signed, 8-bit value
  | Int16Rep      -- ^ Signed, 16-bit value
  ...etc...
  | VecRep Int PrimElemRep  -- ^ SIMD fixed-width vector

The Haskell source language is a bit more flexible: a single value may need multiple PrimReps.
For example

  utup :: (# Int, Int #) -> Bool
  utup x = ...

Here x :: (# Int, Int #), and that takes two registers, and two instructions to move around.
Unboxed sums are similar.

Every Haskell expression e has a type ty, whose kind is of form TYPE rep
   e :: ty :: TYPE rep
where rep :: RuntimeRep. Here rep describes the runtime representation for e's value,
but RuntimeRep has some extra cases:

data RuntimeRep = VecRep VecCount VecElem   -- ^ a SIMD vector type
                | TupleRep [RuntimeRep]     -- ^ An unboxed tuple of the given reps
                | SumRep [RuntimeRep]       -- ^ An unboxed sum of the given reps
                | BoxedRep Levity -- ^ boxed; represented by a pointer
                | IntRep          -- ^ signed, word-sized value
                ...etc...
data Levity     = Lifted
                | Unlifted

It's all in 1-1 correspondence with PrimRep except for TupleRep and SumRep,
which describe unboxed products and sums respectively. RuntimeRep is defined
in the library ghc-prim:GHC.Types. It is also "wired-in" to GHC: see
GHC.Builtin.Types.runtimeRepTyCon. The unarisation pass, in GHC.Stg.Unarise, transforms the
program, so that every variable has a type that has a PrimRep. For
example, unarisation transforms our utup function above, to take two Int
arguments instead of one (# Int, Int #) argument.

Also, note that boxed types are represented slightly differently in RuntimeRep
and PrimRep. PrimRep just has the nullary LiftedRep and UnliftedRep data
constructors. RuntimeRep has a BoxedRep data constructor, which accepts a
Levity. The subtle distinction is that since BoxedRep can accept a variable
argument, RuntimeRep can talk about levity polymorphic types. PrimRep, by
contrast, cannot.

See also Note [Getting from RuntimeRep to PrimRep] and Note [VoidRep].

References 2

Referenced by 6