7.3. Optimisation (code improvement)

The -O* options specify convenient “packages” of optimisation flags; the -f* options described later on specify individual optimisations to be turned on/off; the -m* options specify machine-specific optimisations to be turned on/off.

Most of these options are boolean and have options to turn them both “on” and “off” (beginning with the prefix no-). For instance, while -fspecialise enables specialisation, -fno-specialise disables it. When multiple flags for the same option appear in the command-line they are evaluated from left to right. For instance, -fno-specialise -fspecialise will enable specialisation.

It is important to note that the -O* flags are roughly equivalent to combinations of -f* flags. For this reason, the effect of the -O* and -f* flags is dependent upon the order in which they occur on the command line.

For instance, take the example of -fno-specialise -O1. Despite the -fno-specialise appearing in the command line, specialisation will still be enabled. This is the case as -O1 implies -fspecialise, overriding the previous flag. By contrast, -O1 -fno-specialise will compile without specialisation, as one would expect.

7.3.1. -O*: convenient “packages” of optimisation flags.

There are many options that affect the quality of code produced by GHC. Most people only have a general goal, something like “Compile quickly” or “Make my program run like greased lightning.” The following “packages” of optimisations (or lack thereof) should suffice.

Note that higher optimisation levels cause more cross-module optimisation to be performed, which can have an impact on how much of your program needs to be recompiled when you change something. This is one reason to stick to no-optimisation when developing code.

No ``-O*``-type option specified: This is taken to mean “Please compile quickly; I’m not over-bothered about compiled-code quality.” So, for example, ghc -c Foo.hs

-O0

Means “turn off all optimisation”, reverting to the same settings as if no -O options had been specified. Saying -O0 can be useful if e.g. make has inserted a -O on the command line already.

-O
-O1

Means: “Generate good-quality code without taking too long about it.” Thus, for example: ghc -c -O Main.lhs

-O2

Means: “Apply every non-dangerous optimisation, even if it means significantly longer compile times.”

The avoided “dangerous” optimisations are those that can make runtime or space worse if you’re unlucky. They are normally turned on or off individually.

-Odph

Enables all -O2 optimisation, sets -fmax-simplifier-iterations=20 and -fsimplifier-phases=3. Designed for use with Data Parallel Haskell (DPH).

We don’t use a -O* flag for day-to-day work. We use -O to get respectable speed; e.g., when we want to measure something. When we want to go for broke, we tend to use -O2 (and we go for lots of coffee breaks).

The easiest way to see what -O (etc.) “really mean” is to run with -v, then stand back in amazement.

7.3.2. -f*: platform-independent flags

These flags turn on and off individual optimisations. Flags marked as on by default are enabled by -O, and as such you shouldn’t need to set any of them explicitly. A flag -fwombat can be negated by saying -fno-wombat.

-fcase-merge
Default:on

Merge immediately-nested case expressions that scrutinise the same variable. For example,

case x of
   Red -> e1
   _   -> case x of
            Blue -> e2
            Green -> e3

Is transformed to,

case x of
   Red -> e1
   Blue -> e2
   Green -> e2
-fcase-folding
Default:on

Allow constant folding in case expressions that scrutinise some primops: For example,

case x `minusWord#` 10## of
   10## -> e1
   20## -> e2
   v    -> e3

Is transformed to,

case x of
   20## -> e1
   30## -> e2
   _    -> let v = x `minusWord#` 10## in e3
-fcall-arity
Default:on

Enable call-arity analysis.

-fexitification
Default:on

Enables the floating of exit paths out of recursive functions.

-fcmm-elim-common-blocks
Default:on

Enables the common block elimination optimisation in the code generator. This optimisation attempts to find identical Cmm blocks and eliminate the duplicates.

-fcmm-sink
Default:on

Enables the sinking pass in the code generator. This optimisation attempts to find identical Cmm blocks and eliminate the duplicates attempts to move variable bindings closer to their usage sites. It also inlines simple expressions like literals or registers.

-fasm-shortcutting
Default:off

This enables shortcutting at the assembly stage of the code generator. In simpler terms shortcutting means if a block of instructions A only consists of a unconditionally jump, we replace all jumps to A by jumps to the successor of A.

This is mostly done during Cmm passes. However this can miss corner cases. So at -O2 we run the pass again at the asm stage to catch these.

-fcpr-anal
Default:on

Turn on CPR analysis in the demand analyser.

-fcse
Default:on

Enables the common-sub-expression elimination optimisation. Switching this off can be useful if you have some unsafePerformIO expressions that you don’t want commoned-up.

-fstg-cse
Default:on

Enables the common-sub-expression elimination optimisation on the STG intermediate language, where it is able to common up some subexpressions that differ in their types, but not their represetation.

-fdicts-cheap
Default:off

A very experimental flag that makes dictionary-valued expressions seem cheap to the optimiser.

-fdicts-strict
Default:off

Make dictionaries strict.

-fdmd-tx-dict-sel
Default:on

Use a special demand transformer for dictionary selectors.

-fdo-eta-reduction
Default:on

Eta-reduce lambda expressions, if doing so gets rid of a whole group of lambdas.

-fdo-lambda-eta-expansion
Default:on

Eta-expand let-bindings to increase their arity.

-feager-blackholing
Default:off

Usually GHC black-holes a thunk only when it switches threads. This flag makes it do so as soon as the thunk is entered. See Haskell on a shared-memory multiprocessor.

See Compile-time options for SMP parallelism for a dicussion on its use.

-fexcess-precision
Default:off

When this option is given, intermediate floating point values can have a greater precision/range than the final type. Generally this is a good thing, but some programs may rely on the exact precision/range of Float/Double values and should not use this option for their compilation.

Note that the 32-bit x86 native code generator only supports excess-precision mode, so neither -fexcess-precision nor -fno-excess-precision has any effect. This is a known bug, see Bugs in GHC.

-fexpose-all-unfoldings
Default:off

An experimental flag to expose all unfoldings, even for very large or recursive functions. This allows for all functions to be inlined while usually GHC would avoid inlining larger functions.

-ffloat-in
Default:on

Float let-bindings inwards, nearer their binding site. See Let-floating: moving bindings to give faster programs (ICFP‘96).

This optimisation moves let bindings closer to their use site. The benefit here is that this may avoid unnecessary allocation if the branch the let is now on is never executed. It also enables other optimisation passes to work more effectively as they have more information locally.

This optimisation isn’t always beneficial though (so GHC applies some heuristics to decide when to apply it). The details get complicated but a simple example is that it is often beneficial to move let bindings outwards so that multiple let bindings can be grouped into a larger single let binding, effectively batching their allocation and helping the garbage collector and allocator.

-ffull-laziness
Default:on

Run the full laziness optimisation (also known as let-floating), which floats let-bindings outside enclosing lambdas, in the hope they will be thereby be computed less often. See Let-floating: moving bindings to give faster programs (ICFP‘96). Full laziness increases sharing, which can lead to increased memory residency.

Note

GHC doesn’t implement complete full-laziness. When optimisation in on, and -fno-full-laziness is not given, some transformations that increase sharing are performed, such as extracting repeated computations from a loop. These are the same transformations that a fully lazy implementation would do, the difference is that GHC doesn’t consistently apply full-laziness, so don’t rely on it.

-ffun-to-thunk
Default:off

Worker-wrapper removes unused arguments, but usually we do not remove them all, lest it turn a function closure into a thunk, thereby perhaps creating a space leak and/or disrupting inlining. This flag allows worker/wrapper to remove all value lambdas.

-fignore-asserts
Default:on

Causes GHC to ignore uses of the function Exception.assert in source code (in other words, rewriting Exception.assert p e to e (see Assertions).

-fignore-interface-pragmas
Default:off

Tells GHC to ignore all inessential information when reading interface files. That is, even if M.hi contains unfolding or strictness information for a function, GHC will ignore that information.

-flate-dmd-anal
Default:off

Run demand analysis again, at the end of the simplification pipeline. We found some opportunities for discovering strictness that were not visible earlier; and optimisations like -fspec-constr can create functions with unused arguments which are eliminated by late demand analysis. Improvements are modest, but so is the cost. See notes on the Trac wiki page.

-fliberate-case
Default:off but enabled with -O2.

Turn on the liberate-case transformation. This unrolls recursive function once in its own RHS, to avoid repeated case analysis of free variables. It’s a bit like the call-pattern specialiser (-fspec-constr) but for free variables rather than arguments.

-fliberate-case-threshold=⟨n⟩
Default:2000

Set the size threshold for the liberate-case transformation.

-floopification
Default:on

When this optimisation is enabled the code generator will turn all self-recursive saturated tail calls into local jumps rather than function calls.

-fllvm-pass-vectors-in-regs
Default:on

Instructs GHC to use the platform’s native vector registers to pass vector arguments during function calls. As with all vector support, this requires -fllvm.

-fmax-inline-alloc-size=⟨n⟩
Default:128

Set the maximum size of inline array allocations to n bytes. GHC will allocate non-pinned arrays of statically known size in the current nursery block if they’re no bigger than n bytes, ignoring GC overheap. This value should be quite a bit smaller than the block size (typically: 4096).

-fmax-inline-memcpy-insns=⟨n⟩
Default:32

Inline memcpy calls if they would generate no more than ⟨n⟩ pseudo-instructions.

-fmax-inline-memset-insns=⟨n⟩
Default:32

Inline memset calls if they would generate no more than n pseudo instructions.

-fmax-relevant-binds=⟨n⟩
Default:6

The type checker sometimes displays a fragment of the type environment in error messages, but only up to some maximum number, set by this flag. Turning it off with -fno-max-relevant-bindings gives an unlimited number. Syntactically top-level bindings are also usually excluded (since they may be numerous), but -fno-max-relevant-bindings includes them too.

-fmax-uncovered-patterns=⟨n⟩
Default:4

Maximum number of unmatched patterns to be shown in warnings generated by -Wincomplete-patterns and -Wincomplete-uni-patterns.

-fmax-simplifier-iterations=⟨n⟩
Default:4

Sets the maximal number of iterations for the simplifier.

-fmax-worker-args=⟨n⟩
Default:10

If a worker has that many arguments, none will be unpacked anymore.

-fno-opt-coercion
Default:coercion optimisation enabled.

Turn off the coercion optimiser.

-fno-pre-inlining
Default:pre-inlining enabled

Turn off pre-inlining.

-fno-state-hack
Default:state hack is enabled

Turn off the “state hack” whereby any lambda with a State# token as argument is considered to be single-entry, hence it is considered okay to inline things inside it. This can improve performance of IO and ST monad code, but it runs the risk of reducing sharing.

-fomit-interface-pragmas
Default:Implied by -O0, otherwise off.

Tells GHC to omit all inessential information from the interface file generated for the module being compiled (say M). This means that a module importing M will see only the types of the functions that M exports, but not their unfoldings, strictness info, etc. Hence, for example, no function exported by M will be inlined into an importing module. The benefit is that modules that import M will need to be recompiled less often (only when M’s exports change their type, not when they change their implementation).

-fomit-yields
Default:yield points enabled

Tells GHC to omit heap checks when no allocation is being performed. While this improves binary sizes by about 5%, it also means that threads run in tight non-allocating loops will not get preempted in a timely fashion. If it is important to always be able to interrupt such threads, you should turn this optimization off. Consider also recompiling all libraries with this optimization turned off, if you need to guarantee interruptibility.

-fpedantic-bottoms
Default:off

Make GHC be more precise about its treatment of bottom (but see also -fno-state-hack). In particular, stop GHC eta-expanding through a case expression, which is good for performance, but bad if you are using seq on partial applications.

-fregs-graph
Default:off due to a performance regression bug (Trac #7679)

Only applies in combination with the native code generator. Use the graph colouring register allocator for register allocation in the native code generator. By default, GHC uses a simpler, faster linear register allocator. The downside being that the linear register allocator usually generates worse code.

Note that the graph colouring allocator is a bit experimental and may fail when faced with code with high register pressure Trac #8657.

-fregs-iterative
Default:off

Only applies in combination with the native code generator. Use the iterative coalescing graph colouring register allocator for register allocation in the native code generator. This is the same register allocator as the -fregs-graph one but also enables iterative coalescing during register allocation.

-fsimplifier-phases=⟨n⟩
Default:2

Set the number of phases for the simplifier. Ignored with -O0.

-fsimpl-tick-factor=⟨n⟩
Default:100

GHC’s optimiser can diverge if you write rewrite rules (Rewrite rules) that don’t terminate, or (less satisfactorily) if you code up recursion through data types (Bugs in GHC). To avoid making the compiler fall into an infinite loop, the optimiser carries a “tick count” and stops inlining and applying rewrite rules when this count is exceeded. The limit is set as a multiple of the program size, so bigger programs get more ticks. The -fsimpl-tick-factor flag lets you change the multiplier. The default is 100; numbers larger than 100 give more ticks, and numbers smaller than 100 give fewer.

If the tick-count expires, GHC summarises what simplifier steps it has done; you can use -fddump-simpl-stats to generate a much more detailed list. Usually that identifies the loop quite accurately, because some numbers are very large.

-fspec-constr
Default:off but enabled by -O2.

Turn on call-pattern specialisation; see Call-pattern specialisation for Haskell programs.

This optimisation specializes recursive functions according to their argument “shapes”. This is best explained by example so consider:

last :: [a] -> a
last [] = error "last"
last (x : []) = x
last (x : xs) = last xs

In this code, once we pass the initial check for an empty list we know that in the recursive case this pattern match is redundant. As such -fspec-constr will transform the above code to:

last :: [a] -> a
last []       = error "last"
last (x : xs) = last' x xs
    where
      last' x []       = x
      last' x (y : ys) = last' y ys

As well avoid unnecessary pattern matching it also helps avoid unnecessary allocation. This applies when a argument is strict in the recursive call to itself but not on the initial entry. As strict recursive branch of the function is created similar to the above example.

It is also possible for library writers to instruct GHC to perform call-pattern specialisation extremely aggressively. This is necessary for some highly optimized libraries, where we may want to specialize regardless of the number of specialisations, or the size of the code. As an example, consider a simplified use-case from the vector library:

import GHC.Types (SPEC(..))

foldl :: (a -> b -> a) -> a -> Stream b -> a
{-# INLINE foldl #-}
foldl f z (Stream step s _) = foldl_loop SPEC z s
  where
    foldl_loop !sPEC z s = case step s of
                            Yield x s' -> foldl_loop sPEC (f z x) s'
                            Skip       -> foldl_loop sPEC z s'
                            Done       -> z

Here, after GHC inlines the body of foldl to a call site, it will perform call-pattern specialisation very aggressively on foldl_loop due to the use of SPEC in the argument of the loop body. SPEC from GHC.Types is specifically recognised by the compiler.

(NB: it is extremely important you use seq or a bang pattern on the SPEC argument!)

In particular, after inlining this will expose f to the loop body directly, allowing heavy specialisation over the recursive cases.

-fspec-constr-keen
Default:off

If this flag is on, call-pattern specialisation will specialise a call (f (Just x)) with an explicit constructor argument, even if the argument is not scrutinised in the body of the function. This is sometimes beneficial; e.g. the argument might be given to some other function that can itself be specialised.

-fspec-constr-count=⟨n⟩
Default:3

Set the maximum number of specialisations that will be created for any one function by the SpecConstr transformation.

-fspec-constr-threshold=⟨n⟩
Default:2000

Set the size threshold for the SpecConstr transformation.

-fspecialise
Default:on

Specialise each type-class-overloaded function defined in this module for the types at which it is called in this module. If -fcross-module-specialise is set imported functions that have an INLINABLE pragma (INLINABLE pragma) will be specialised as well.

-fspecialise-aggressively
Default:off

By default only type class methods and methods marked INLINABLE or INLINE are specialised. This flag will specialise any overloaded function regardless of size if its unfolding is available. This flag is not included in any optimisation level as it can massively increase code size. It can be used in conjunction with -fexpose-all-unfoldings if you want to ensure all calls are specialised.

-fcross-module-specialise
Default:on

Specialise INLINABLE (INLINABLE pragma) type-class-overloaded functions imported from other modules for the types at which they are called in this module. Note that specialisation must be enabled (by -fspecialise) for this to have any effect.

-flate-specialise
Default:off

Runs another specialisation pass towards the end of the optimisation pipeline. This can catch specialisation opportunities which arose from the previous specialisation pass or other inlining.

You might want to use this if you are you have a type class method which returns a constrained type. For example, a type class where one of the methods implements a traversal.

-fsolve-constant-dicts
Default:on

When solving constraints, try to eagerly solve super classes using available dictionaries.

For example:

class M a b where m :: a -> b

type C a b = (Num a, M a b)

f :: C Int b => b -> Int -> Int
f _ x = x + 1

The body of f requires a Num Int instance. We could solve this constraint from the context because we have C Int b and that provides us a solution for Num Int. However, we can often produce much better code by directly solving for an available Num Int dictionary we might have at hand. This removes potentially many layers of indirection and crucially allows other optimisations to fire as the dictionary will be statically known and selector functions can be inlined.

The optimisation also works for GADTs which bind dictionaries. If we statically know which class dictionary we need then we will solve it directly rather than indirectly using the one passed in at run time.

-fstatic-argument-transformation
Default:off

Turn on the static argument transformation, which turns a recursive function into a non-recursive one with a local recursive loop. See Chapter 7 of Andre Santos’s PhD thesis

-fstrictness
Default:on

Switch on the strictness analyser. The implementation is described in the paper `Theory and Practice of Demand Analysis in Haskell`<https://www.microsoft.com/en-us/research/wp-content/uploads/2017/03/demand-jfp-draft.pdf>`__.

The strictness analyser figures out when arguments and variables in a function can be treated ‘strictly’ (that is they are always evaluated in the function at some point). This allow GHC to apply certain optimisations such as unboxing that otherwise don’t apply as they change the semantics of the program when applied to lazy arguments.

-fstrictness-before=⟨n⟩

Run an additional strictness analysis before simplifier phase ⟨n⟩.

-funbox-small-strict-fields
Default:on

This option causes all constructor fields which are marked strict (i.e. “!”) and which representation is smaller or equal to the size of a pointer to be unpacked, if possible. It is equivalent to adding an UNPACK pragma (see UNPACK pragma) to every strict constructor field that fulfils the size restriction.

For example, the constructor fields in the following data types

data A = A !Int
data B = B !A
newtype C = C B
data D = D !C

would all be represented by a single Int# (see Unboxed types and primitive operations) value with -funbox-small-strict-fields enabled.

This option is less of a sledgehammer than -funbox-strict-fields: it should rarely make things worse. If you use -funbox-small-strict-fields to turn on unboxing by default you can disable it for certain constructor fields using the NOUNPACK pragma (see NOUNPACK pragma).

Note that for consistency Double, Word64, and Int64 constructor fields are unpacked on 32-bit platforms, even though they are technically larger than a pointer on those platforms.

-funbox-strict-fields
Default:off

This option causes all constructor fields which are marked strict (i.e. !) to be unpacked if possible. It is equivalent to adding an UNPACK pragma to every strict constructor field (see UNPACK pragma).

This option is a bit of a sledgehammer: it might sometimes make things worse. Selectively unboxing fields by using UNPACK pragmas might be better. An alternative is to use -funbox-strict-fields to turn on unboxing by default but disable it for certain constructor fields using the NOUNPACK pragma (see NOUNPACK pragma).

Alternatively you can use -funbox-small-strict-fields to only unbox strict fields which are “small”.

-funfolding-creation-threshold=⟨n⟩
Default:750

Governs the maximum size that GHC will allow a function unfolding to be. (An unfolding has a “size” that reflects the cost in terms of “code bloat” of expanding (aka inlining) that unfolding at a call site. A bigger function would be assigned a bigger cost.)

Consequences:

  1. nothing larger than this will be inlined (unless it has an INLINE pragma)
  2. nothing larger than this will be spewed into an interface file.

Increasing this figure is more likely to result in longer compile times than faster code. The -funfolding-use-threshold=⟨n⟩ is more useful.

-funfolding-dict-discount=⟨n⟩
Default:30

How eager should the compiler be to inline dictionaries?

-funfolding-fun-discount=⟨n⟩
Default:60

How eager should the compiler be to inline functions?

-funfolding-keeness-factor=⟨n⟩
Default:1.5

How eager should the compiler be to inline functions?

-funfolding-use-threshold=⟨n⟩
Default:60

This is the magic cut-off figure for unfolding (aka inlining): below this size, a function definition will be unfolded at the call-site, any bigger and it won’t. The size computed for a function depends on two things: the actual size of the expression minus any discounts that apply depending on the context into which the expression is to be inlined.

The difference between this and -funfolding-creation-threshold=⟨n⟩ is that this one determines if a function definition will be inlined at a call site. The other option determines if a function definition will be kept around at all for potential inlining.

-fvectorisation-avoidance
Default:on

Part of Data Parallel Haskell (DPH).

Enable the vectorisation avoidance optimisation. This optimisation only works when used in combination with the -fvectorise transformation.

While vectorisation of code using DPH is often a big win, it can also produce worse results for some kinds of code. This optimisation modifies the vectorisation transformation to try to determine if a function would be better of unvectorised and if so, do just that.

-fvectorise
Default:off

Part of Data Parallel Haskell (DPH).

Enable the vectorisation optimisation transformation. This optimisation transforms the nested data parallelism code of programs using DPH into flat data parallelism. Flat data parallel programs should have better load balancing, enable SIMD parallelism and friendlier cache behaviour.