This page collects together information about the optimisations that GHC does and does not perform.
- GHC experts: Please check that the info in this page is correct.
- Everybody else: Feel free to add questions!
Common subexpression elimination
First of all, common subexpression elemination (CSE) means that if an expression appears in several places, the code is rearranged so that the value of that expression is computed only once. For example:
foo x = (bar x) * (bar x)
might be transformed into
foo x = let x' = bar x in x' * x'
bar function is only called once. (And if
bar is a particularly expensive function, this might save quite a lot of work.)
GHC doesn't actually perform CSE as often as you might expect. The trouble is, performing CSE can affect the strictness/lazyness of the program. So GHC does do CSE, but only in specific circumstances --- see the GHC manual. (Section??)
Long story short: "If you care about CSE, do it by hand."
Inlining is where a function call is replaced by that function's definition. For example, the standard
map function can be defined as
map :: (a -> b) -> [a] -> [b] map f  =  map f (x:xs) = f x : map f xs
Now if you write something like
foo = map bar
it's possible that the compiler might inline the definition of
map, yielding something like
foo  =  foo (x:xs) = bar x : foo xs
which is (hopefully!) faster, because it doesn't involve a call to the
map function any more, it just does the work directly. (This might also expose new optimisations opportunities;
map works for any types, whereas
foo probably works for only one type.)
So, that's what inlining is. By default, GHC will inline things if they are 'small enough'. Every time you inline a function, you are in a sense making a (customised) copy of that function. Do too much of this and the compiled program will be enormous. So it's only worth it for 'small' functions.
(How does GHC determine 'small'? Isn't there a switch that adjusts this?)
Haskell is a lazy language. Calculations are notionally not performed until their results are 'needed'. However, if the result definitely will be needed, it's a waste of time and effort to save up the expression and execute it later; more efficient to just execute it right now.
Strictness analysis is a process by which GHC attempts to determine, at compile-time, which data definitely will 'always be needed'. GHC can then build code to just calculate such data, rather than the normal (higher overhead) process for storing up the calculation and executing it later.
Unfortunately, looking at a program and saying "will this data be needed?" is a bit like looking at a program and saying "this program will never halt" --- see The Halting Problem. (Good link?) But GHC does its best, and can give big speedups in some cases.
In order to understand how to write efficient code, and what GHC does with your code to optimise it, it helps to know a bit about what your compiled code looks like and how it works.
To a first approximation, at any moment your program is a 'graph' of objects in memory. ('Graph' in the graph theory sense --- nodes connected by arcs.) Some of the objects are 'data' --- booleans, integers, strings, lists, etc. Some of those objects are functions (because Haskell lets you pass functions around like data). And some of these are thunks --- unevaluated expressions (because Haskell only evaluates expressions 'as needed').
The program starts off with a single node representing the unevaluated call to
main, and proceeds to execute from there. Each time a thunk is executed, the result (whatever it is) overwrites the thunk data. (It's possible that the result of evaluating a thunk is a new thunk of course.)
GHC compiles to the spineless tagness G-machine (STG). This is a notional graph reduction machine (i.e., a virtual machine that performs graph reductions as described above). 'G-machine' because it does graph reduction. 'Spineless' because it can't stand up to bullies. 'Tagless' because the graph nodes don't have 'tags' on them to say what they are.
Instead of tags, the nodes have access pointers. If the node is a thunk, its pointer points to the code to evaluate the thunk and return the real result. Otherwise the pointer points to some 'do-nothing' code. So to access any type of node, you just do an indirect jump on this pointer; no case analysis is necessary.
(Gosh I hope I got that lot right!)
Internally, GHC uses a kind of 'machine code' that runs on this non-existent G-machine. It does a number of optimisations on that representation, before finally compiling it into real machine code (possibly via C using GCC).
There are a number of optimisations done at the STG level. These mainly involve trying to avoid unnecessary steps:
- Avoid creating a thunk which immediately creates another thunk when executed; make it evaluate all the way down to a final result in one go. (If we 'need' the thunk's value, we're going to evaluate all the way down anyway, so let's leave out the overhead...)