In the attached test case, profiling assigns a large fraction of run time to a single function. The fraction ranges from 19% to 40% across runs. Although I haven't independently verified the profile, I believe execution time is mis-attributed to this function.

The function in question is:

intersect :: Object -> Ray -> Maybe Intersection
intersect obj@(Object dat mem) = objmIntersect mem dat obj

where objmIntersect is a field selector. In -ddump-simpl output, the function consists of two case statements and a function call. The function it calls has its own SCC.

In profiling, this function is usually listed as the most time-consuming function in the program. Its callee is assigned a similar amount of time. Both functions are called the same number of times. The callee computes a ray-triangle intersection, which is presumably more time-consuming than two case statements and a function call...

COST CENTRE                    MODULE               %time %alloc

intersect                      Datatypes             28.5   11.3
triangle.intersect             Triangle              25.0    2.0
(other functions...)

I've reduced the test case from a somewhat larger ray tracer. I'm willing to help reduce it further if needed. The test case depends on the vector-space library.