Glossary

Arity

The arity of a function is the number of arguments the function must take to conclude to a result.

Algebraic Data Type

First implemented in the Hope programming language Hudak et al. [9], Algebraic Data Types are composite, meaning that they are data types made from other data types. For example,

data Foo = One Int Int
         | Two Bool Int
         | Three Float Bool Char

Here Foo is an algebraic data type made from the types Int, Bool, Float, and Char. In general, these data types are composed from product types and sum types.

Product types are the types on finds in most other imperative and object oriented programming languages, such as struct in C or tuples in Haskell. Product types are called so because they form the cartesian product on the set of elements represented by the type, for example the type (Int, Bool) would be written \(Int \times Bool\), and would represent the set of all pairs of elements of the sets represented by the types Int and Bool.

Sum types are often not found in imperative and object oriented languages, but are a tagged union or disjoint union of other types. Again, thinking in terms of sets, a sum type represents a disjoint union of two or more types. For example Foo is the union of three product types that are each tagged with a constructor: One, Two and Three. Thus in terms of sets on might write the type Foo as \(f \in Foo = (Int \times Int) + (Bool \times Int) + (Float \times Bool \times Char)\). Notice also that the type Foo with elements f are structural (by its definition we know a Foo can only be a One, Two, or Three and how many fields each of these constructors have) as opposed to nominal. Nominal types are the kind of types created by a newtype. They can have the same representation but are treated as a wholly unique type.

Help Wanted

I’ve tried to give a thorough description of algebraic data types without diving into too much type theory, but I find the explanation a bit unsatisfying. For example, it is not clear why these are called Inductive or Algebraic because I deemed this was too much depth for a glossary entry. If you have a good resource or would like to take a stab at this entry then please make an issue and have at it!

Boxed

A Boxed value is a value that is represented by a pointer to the heap. For example, a value such as 1729 :: Int is represented as:

Cardinality Analysis

A static analysis that GHC performs to determine: (1) How many times a lambda-expression is called, (2) Which components of a data structure are never evaluated, (3) How many times a particular thunk is evaluated. See Graf [10] and Sergey et al. [11] for more.

Closure

A closure is value that pairs a function with an environment, where the environment maps every free variable in the function with a value or reference to which the free variable was bound when the closure was created. Closure’s are the canonical way to realize lexical scoping in languages with first-class functions, such a Haskell. See the wikipedia entry for more.

Closure Conversion

Closure conversion is the default way GHC treats free variables in a function body. Closure Conversion creates a top level record for the original function, called the function environment, whose fields are the free variables of the function. The environment is passed to the function as an implicit parameter and the free variable call sites are rewritten as field accesses. Then the function and the record are grouped in a tuple, i.e., a closure (pair of environment and function) is created causing some extra heap allocation. Finally the call sites of the original function are rewritten to pass the environment with the original function. Consider this example:

...
let f = foldl (\acc _ -> acc + x) y xs
in  f [1..100]
...

In this example x and y are free variables in the function f . Closure conversion will capture them and transform this function to:

...
-- the function environment
data EnvF = EnvF { x :: Int, y :: Int }

-- the new function
f_cc env xs = foldl (\acc _ -> acc + x env) (y env) xs

-- the closure that replaces the original function in the same scope
let f = (f_cc, EnvF x y)
in (fst f) (snd f) [1..100]
...

Notice closure conversion has added an extra let expression for the closure and the reference to x and y have been replaced with accesses to env . The let expression can be a source of extra heap allocations and is one of the costs of closure conversion. However, the benefits are uniformity; every function can be treated as a closure. Closure conversion is often contrasted with Lambda Lifting which is another strategy to handle free variables that does not incur extra heap allocation. See Johnsson [6] and Graf and Jones [7] for more.

CAF

A CAF, or Constant Applicative Form, is a Haskell value which contains no free variables and is not a function. Consider these examples:

-- these are CAFs
-- A static literal is a CAF
foo :: Int
foo = 12

-- A reducible expression that requires no input is a CAF
bar :: (Int, [Int])
bar = ((*) 10 10, [1..])

-- not a lambda, curried functions that can be reduced when given an
-- input are CAFs
baz :: Int -> Int
baz = (*) 3

-- not CAFs
qux :: Int -> Int
qux e = e * 3     -- equivalent to baz but is a lambda so not a CAF

quux :: Int -> Int
quux = (*) x      -- x is free thus not a CAF

These values are constant because they don’t bind any variables or have any free variables. Because they are constant they are floated (see Let Floating) to the top of the program, and statically allocated during compile time. Since they are statically allocated at compile time CAFs are pinned memory and special treatment in the runtime system. Thus, heavily allocating CAFs can increase memory residency. See Jones et al. [5] Section 10.8 for more details.

DWARF

DWARF symbols are a widely used and standardized data format used to provide source level debugging. For more, see the official webpage.

Entry Code

The entry code for a closure on the heap is the code that will evaluate that closure. There are some nuances and exceptions: For functions the entry code applies the function to its arguments, which the entry code assumes are all present; that is, the entry code assumes all arguments are either loaded into registers or are already on the stack. Should the function be applied to too few arguments or should the function be an Unknown function then a generic apply is used. For a PAP, there is no entry code. PAPs can only be applied to more arguments using the generic apply functions. Lastly, Unlifted Objects cannot be evaluated and thus have no entry code.

Full Laziness transformation

A form of Let Floating which moves let bindings out of lambda abstractions to avoid unnecessary allocation and computation. See Peyton Jones and Santos [2] Section 7.2.

Fusion

See What is Fusion.

HNF

An expression that is in head normal form is a value which contains at least one thunk. If the value does not contain any thunks, then it is said to be in normal form (NF). See Jones et al. [5] Section 3.1 for more.

Info Table

Every heap allocated object in the runtime system keeps an information table that stores data such as: the object type (function, data constructor, thunk etc.) before the payload of the object. This is called the Info Table. See Marlow et al. [4], wiki, and Peyton Jones and Salkild [12] Section 7.1 for more details.

Info Table Address

The memory address for heap object descriptors info table.

Join Point

A join point is a place where different execution paths come together or join. Consider this example slightly modified from Maurer et al. [13]:

let join1 _ = some_large_expression
    join2 _ = some_other_large_expr
in if e1 then (if e2 then join1 () else join2 ())
         else (if e3 then join1 () else join2 ())

In this example, join1 and join2 are join points because the branches described by each if-expression conclude by calling them. Thus, the control flow described by the if-expressions joins at specifically join1 and join2. Join points are an important optimization technique that GHC performs automatically to remove redundant allocations. Had we not wrapped some_large_expression and some_other_large_expr in a let, then these expressions would be duplicated and would be captured in an additionally allocated closure unnecessarily. Join points avoid these problems and are particularly relevant for Stream Fusion performance. For more see the join points paper: Maurer et al. [13].

Known Function

A known function is a function in the STG machine of which GHC statically knows the Entry Code pointer and the Arity of. This means that the function binding site is statically visible, that is, the function is Top-Level, or the function is bound by an enclosing let. With this information the STG machine can use a faster function application procedure because the function pointer does not need to be scrutinized. See also Unknown Function.

Levity Polymorphism

A kind of polymorphism that abstracts over calling conventions which allows levity polymorphic functions to be abstracted over memory layout. See Eisenberg and Peyton Jones [14] for a more precise technical definition and discussion.

Let Floating

A group of optimizing transformation’s that move let bindings to reduce heap allocations. See Partain et al. [15] and Peyton Jones and Santos [2] Section 7 for more details.

Lifted

A Lifted type is a type that contains the value \(\bot\); which means the type is lazy and capable of representing non-terminating computation. For example, the Bool type is a set with three values: True, False, and \(\bot\). Therefore Bool is a Lifted type.

Loop Fusion

Loop fusion is a classic optimization technique that reduces the number of loops in a program, thereby reducing the number of memory accesses and the number of looping constructs. In Haskell, loop fusion transforms many traversals over the same data structure to a single traversal. A classic example of this is map fusion.

-- two traversals, one for f, one for g on the result of f
map g . map f $ [1..100]

-- after map fusion:
-- only one traversal
map (g . f) [1..100]

This can also appear in list comprehensions, for example:

...
-- three traversals: two to project elements, 1 to fold
let foo = foldl + 0 [ i | (i,_) <- args ]
let res = bar foo   [ j | (_,j) <- args ]

-- after loop fusion on the list comprehensions
-- 2 traversals: one for the arguments, one to fold
let (is, js) = unzip args
let foo = foldl + 0 is
let bar = bar foo js

Multi-Shot Lambda

A multi-shot lambda is a lambda that is called more than once. In contrast to a one-shot lambda, a multi-shot lambda has a high risk of destroying sharing if subject to certain optimizations, such as Inlining. GHC determines whether a lambda is one-shot or multi-shot during Cardinality Analysis. See Sergey et al. [11] and Graf [10] for more.

NF

An expression that is in normal form is a fully evaluated expression and is a value which contains no thunks. This is in contrast to weak head normal form (WHNF) and head normal form (HNF), both of which may contain thunks. See Jones et al. [5] Section 3.1 for more.

Occurrence Name

An Occurrence name is a name GHC assigns to an entity to disambiguate multiple occurrences of that name. Disambiguation allows GHC to distinguish by name a type constructor from a data constructor, which often occurs due to punning, or from local variables in separate functions with the same name, such as x or xs. Occurrence names are a pair of the original name (as a FastString, a GHC internal type) and a NameSpace; they are ubiquitous in GHC and in the intermediate representations. For example, the occurrence name for the function f x y = ... will be similar to f_r17p. The exact occurrence name will change, but parts are static. For example, the f before the underscore always comes from the name of the function. Had f been name fancyFunction then the ocurrence name would have been fancyFunction_r17p. Similarly, leading character in the suffix; the r in r17p is static and meaningful. In this case, the r indicates that the name f is an element in the NameCache, meaning that all references to f share a single Unique ID in every GHC invocation (See the Note [The Name Cache] for more). When occurrence names are generated, the leading character is a hint for what kind of name is being generated. You can find an incomplete list of tags and their meanings in Note [Uniques for wired-in prelude things and known tags]. For more on names see Note [Choosing external Ids] and this wiki page on GHC’s Reader names.

One-Shot Lambda

A one-shot lambda is a lambda that is called exactly once. These lambda’s are common in functional programming and can be subject to more aggressive optimizations due to their one-shot nature. For example, there is no risk of losing sharing in a one-shot lambda as a result of inlining free variables or floating let expressions into the lambda; something that GHC usually avoids. See Sergey et al. [11] and Graf [10] for more background. See the magic oneShot function in GHC.Exts for an unsafe way to instruct GHC that you have a one-shot lambda.

PAP

A PAP is a partial application. PAPs are heap objects and thus a type of closure that represents a function applied to too few arguments. PAPs should never be entered, and are only applied using the generic apply functions in the STG machine. See the file rts/Apply.cmm in GHC or the heap object wiki page for more.

Pinned

Pinned memory is memory that is guaranteed to not be moved by GHC’s garbage collector. This is most often useful for interfacing with foreign code. Note that pinned memory may lead to memory fragmentation and increased slop because it never moves. See Well Typed’s post and the wiki for more.

Reproducer

A reproducer is the smallest known program that induces incorrect behavior in the system. See Make it Fail for more.

Sharing

Consider the following program:

foo :: Int -> Int
foo n = let x = [1..n]
            in zip (fmap (* (last x)) x) x

We say that x is shared in this program because each of the three references of x refer to the x defined in the let. If x is not shared that the list [1..n] would be allocated for each reference of x. Thus, sharing is fundamental to performance oriented Haskell because it reduces allocations, leverages call-by-need, and saves work.

Shotgun Debugging

Debugging with hope instead of process and measurement. See its Wikepedia entry.

SRT

Static reference tables are how GHC’s garbage collector determines the live CAF’s of a program. SRTs are stored in a heap object’s Info Table and are simply an object in the compiled programs data segment. See The SRT Note in GHC.Cmm.Info.Build for more details.

Thunk

A thunk is a special kind of Closure that represents a suspended computation. Thunks reside on the heap and are the key feature that provides Haskell’s laziness. See Peyton Jones and Salkild [12] Section 3.1.2 for more details.

Thread State Object (TSO)

A thread state object is a heap object that represents a Haskell thread in GHC’s runtime system. For the precise contents please see its definition in GHC’s source code and this description by Ben Gamari.

Top-Level

The most outer-most or global scope of the program.

Unboxed

An UnBoxed value is a value that is represented by the value itself. UnBoxed values therefore cannot be lazy, like boxed values.

Unlifted

An Unlifted type is a type where \(\bot\) is not an element of that type. See Levity Polymorphism and Lifted types for more.

Unknown function

An unknown function is a function in the STG machine whose Entry Code pointer and Arity are not statically known by GHC. Unknown functions require GHC to generate code that first scrutinizes the function pointer to determine its arity and then dispatch to the normal function call handling procedures. This in known has a generic apply in the STG machine and is slower (due to needing to scrutinize the function) than a Known function. See Marlow and Jones [3] for more details on STG calling conventions.

Unfolding

An Unfolding of an identifier, as defined in GHC.Core.Unfold, is the approximate form the identifier would have if the identifier’s definition was substituted for the identifier. That is, Unfoldings are generally the right hand sides or bodies of function definitions untouched by optimizations. Unfoldings appear in Core and Interface files to enable cross-module inlining and optimizations. See the Reading Core chapter for more.

WHNF

An expression is in weak head normal form if it has been evaluated to its’ outermost data constructor or lambda abstraction (i.e., the head). See this post, the wiki , and wikipedia for more.