Is everything in Haskell stored in thunks, even simple values?

Official answer

It's none of your business. Strictly implementation detail of your compiler.

Short answer

Yes.

Longer answer

To the Haskell program itself, the answer is always yes, but the compiler can and will do things differently if it finds out that it can get away with it, for performance reasons.

For example, for '''add x y = x + y''', a compiler might generate code that works with thunks for x and y and constructs a thunk as a result. But consider the following:

foo :: Int -> Int -> Int
foo x y = x * x + y * y

Here, an optimizing compiler will generate code that first takes x and y out of their boxes, then does all the arithmetic, and then stores the result in a box.

Advanced answer

This paper describes how GHC switched from one way of implementing thunks to another that was actually both simpler and faster: http://research.microsoft.com/en-us/um/people/simonpj/papers/eval-apply/

In general, even primitive values in Haskell (e.g. of type Int and Float) are represented by thunks. This is indeed required by the non-strict semantics; consider the following fragment:

bottom :: Int
bottom = div 1 0

This definition will generate a div-by-zero exception only if the value of bottom is inspected, but not if the value is never used.

Consider now the add function:

add :: Int -> Int -> Int
add x y = x+y

A naive implementation of add must force the thunk for x, force the thunk for y, add the values and create an (evaluated) thunk for the result. This is a huge overhead for arithmetic compared to strict functional languages (not to mention imperative ones).

However, an optimizing compiler such as GHC can mostly avoid this overhead; this is a simplified view of how GHC translates the add function:

add :: Int -> Int -> Int
add (I# x) (I# y) = case# (x +# y) of z -> I# z 

Internally, basic types like Int is seen as datatype with a single constructor. The type Int# is the "raw" machine type for integers and +# is the primitive addition on raw types. Operations on raw types are implemented directly on bit-patterns (e.g. registers) --- not thunks. Boxing and unboxing are then translated as constructor application and pattern matching.

The advantage of this approach (not visible in this simple example) is that the compiler is often capable of inlining such definitions and removing intermediate boxing/unboxing operations, leaving only the outermost ones.