Tuesday, June 07, 2011

INLINE pragmas in the Safe library

Summary: The Safe library has lots of small functions, none of which have INLINE pragmas on them. The lack of INLINE pragmas probably has no effect on the optimisation.

I was recently asked a question about the Safe library:

I was wondering why you didn't use any INLINE pragmas in the Safe library. I'm not a Haskell expert yet, but I've noticed that in many other libraries most one-liners are annotated by INLINE pragmas, so is there any reason you didn't add them to Safe as well?

When compiling with optimisations, GHC tries to inline functions/values that are either "small enough" or have an INLINE pragma. These functions will be inlined within their module, and will also be placed in their module's interface file, for inlining into other modules. The advantage of inlining is avoiding the function call overhead, and possibly exposing further optimisations. The disadvantage is that the code size may grow, which may result in poor utilisation of the processor's instruction cache.

The Safe library contains wrappers for lots of partial Prelude and Data.List functions (i.e. head), with versions that don't fail, or fail with more debugging information. Using the tail function as an example, the library supplies four additional variants:

  • tail :: [a] -> [a], crashes on tail []

  • tailNote :: String -> [a] -> [a], takes an extra argument which supplements the error message

  • tailDef :: [a] -> [a] -> [a], takes a default to return instead of errors

  • tailMay :: [a] -> Maybe [a], wraps the result in a Maybe

  • tailSafe :: [a] -> [a], returns some sensible default if possible, [] in the case of tail

As the questioner correctly noted, there are no INLINE pragmas in the library. Taking the example of tailSafe, it is defined as:

tailSafe :: [a] -> [a]
tailSafe = liftSafe tail null

liftSafe :: (a -> a) -> (a -> Bool) -> (a -> a)
liftSafe func test val = if test val then val else func val

Given this indirection, and the lack of an INLINE pragma, is calling tailSafe as cheap as you might hope? We can answer this question by looking at the interface file at the standard optimisation level, by running ghc -ddump-hi Safe.hs -O -c. Looking at the definition attached to tailSafe, we see:

tailSafe = \val -> case val of
[] -> []
ds1:ds2 -> ds2

The tailSafe function has been optimised as much as it can be, and will be inlined into other modules. Adding an INLINE pragma would have no effect at standard optimisation. When compiling without optimisation, even with an INLINE pragma, the function is not included in the module's interface. Adding the INLINE pragma to tailSafe is of no benefit.

Let us now look at all functions in the Safe module. When compiled without optimisation (-O0) no functions are placed in the interface file. At all higher levels of optimisation (-O1 and -O2) the interface files are identical. Looking through the functions, only 6 functions are not included in the interface files:


The abort function is defined as:

abort = error

Neither adding an INLINE pragma or eta expanding (adding an additional argument) includes the function in the interface file. I suspect that GHC decides no further optimisation is possible, so doesn't ever inline abort.

atMay and $watMay

The function atMay is recursive, and thus cannot be inlined even when it's definition is available, so GHC sensibly omits it from the interface file. The function $watMay is the worker/wrapper definition of atMay, which is also recursive.

readNote and readMay

The functions readNote and readMay both follow exactly the same pattern, so I describe only readNote. The function readNote is quite big, and GHC has split it into two top-level functions - producing both readNote and readNote1. The function readNote is included in the interface file, but readNote1 is not. The result is that GHC will be able to partially inline the original definition of readNote - however, the read functions tend to be rather complex, so GHC is unlikely to benefit from the additional inlining it has managed to acheive.


The atNote function is bigger than the inlining threshold, so does not appear in the interface file unless an INLINE pragma is given. Since the inner loop of atNote is recursive, it is unlikely that inlining would give any noticable benefit, so GHC's default behaviour is appropriate.


Having analysed the interface files, I do not think it is worth adding INLINE pragmas to the Safe library - GHC does an excellent job without my intervention. My advice is to only add INLINE pragmas after either profiling, or analysing the resulting Core program.


Ketil said...

Nice! I suspect optimization pragmas like INLINE should really only be added after benchmarking or profiling. I also wonder if they shouldn't be kept separate from the code, rather than..er..inline? Different compilers will likely have different optimal choices, and separating optimization hints from the code would make this a lot more manageable. In addition, this would help pave the way for advanced stuff like feedback-directed optimization.

mokus said...

In my experience, GHC is very good at inlining automatically. I have found many cases where benchmarking shows that adding either an INLINE pragma _OR_ a NOINLINE pragma (to the same code) slows things down overall - indicating that GHC does a good job of deciding not only what to inline but when to do so. These days I very rarely use those pragmas, and when I do I always measure to see whether they actually help.