[Haskell-cafe] profiling of functions generated by LLVM

[Henning Thielemann]

I use LLVM to create sub-routines for efficient inner loops. I like to 
profile them. However, it seems that GHC's profiler does not interact well 
with such generated functions.

Say, I have a function that creates an efficient subroutine for filling a 
block of data. It looks like the following:

generateFill :: IO (Word64 -> Ptr Float -> IO ())
generateFill = generateFunction (LLVM code for traversing block)

main :: IO ()
main = do
    let n = 100*10^6
    fill <- generateFill
    allocaArray n $ \ptr ->
       {-# SCC fill #-} fill (fromIntegral n) ptr


I run this with profiling. The resulting .prof report says, that 
'generateFunction' demands all the computation time and 'fill' inherits 
it.

   fill                 Main     1297      1    0.0    0.0    99.9    0.0
    generateFill        Main     1298      0    0.0    0.0    99.9    0.0
     generateFunction   Main     1299      0   99.9    0.0    99.9    0.0

I have a rough idea why GHC treats the generated functions the way it 
does. However, it is pretty counter-intuitive and difficult to understand 
in more complex programs.