The Pure Manual

Author: Albert Gräf <>
Date: 2010-09-30

Copyright (c) 2009-2010 by Albert Gräf. This document is available under the GNU Free Documentation License. Also see the Copying section for licensing information of the software.

This manual describes the Pure programming language and how to invoke the Pure interpreter program. To read the manual inside the interpreter, just type help at the command prompt. See the Online Help section for details.

There is a companion to this manual, the Pure Library Manual which contains the description of the standard library operations. More information about Pure can be found under the following URLs:


1   Introduction

Pure is a modern-style functional programming language based on term rewriting. Pure programs are basically collections of equational rules used to evaluate expressions in a symbolic fashion by reducing them to normal form. An overview of the language can be found in the Pure Overview section below, and subsequent sections discuss most language features in detail.

The Pure interpreter has an LLVM backend which JIT-compiles Pure programs to machine code, hence programs run blazingly fast and interfacing to C modules is easy, while the interpreter still provides a convenient, fully interactive environment for running Pure scripts and evaluating expressions. You can also compile your scripts to standalone executables if you prefer that.

Pure programs (a.k.a. scripts) are just ordinary text files containing Pure code. They must be encoded in UTF-8 (which subsumes 7 bit ASCII), other encodings such as Latin-1 are not supported. A bunch of syntax highlighting files and programming modes for various popular text editors are included in the Pure sources. There's no difference between the Pure programming language and the input language accepted by the interpreter, except that the interpreter also understands some special commands when running in interactive mode; see the Interactive Usage section for details.

(In case you're wondering, the name "Pure" actually refers to the adjective. But you can also write it as "PURE" and take this as a recursive acronym for the "Pure Universal Rewriting Engine".)

2   Invoking the Pure Interpreter

The Pure interpreter is invoked as follows:

pure [options ...] [script ...] [-- args ...]
pure [options ...] -x script [args ...]

2.1   Options

The interpreter accepts various options which are described in more detail below:

Batch compilation.
--ctags, --etags
Create a tags file in ctags (vi) or etags (emacs) format.
Enable eager JIT compilation. This requires LLVM 2.7 or later, otherwise this flag will be ignored.
-fPIC, -fpic
Create position-independent code (batch compilation).
Enable symbolic debugging.
--help, -h
Print help message and exit.
Force interactive mode (read commands from stdin).
-I directory
Add a directory to be searched for included source scripts.
-L directory
Add a directory to be searched for dynamic libraries.
-l libname
Library to be linked in batch compilation.
Disable command-line editing.
--noprelude, -n
Do not load the prelude.
Do not run the interactive startup files.
-o filename
Output filename for batch compilation.
Quiet startup (suppresses sign-on message in interactive mode).
-T filename
Tags file to be written by --ctags or --etags.
Do not strip unused functions in batch compilation.
Set verbosity level. See below for details.
Print version information and exit.
Enable compiler warnings.
Execute script with given command line arguments.
Stop option processing and pass the remaining command line arguments in the argv variable.

(Besides these, the interpreter also understands a number of other command line switches for setting various code generation options; please see Code Generation Options below for details.)

2.2   Overview of Operation

If any source scripts are specified on the command line, they are loaded and executed, after which the interpreter exits. Otherwise the interpreter enters the interactive read-eval-print loop, see Running Interactively below. You can also use the -i option to enter the interactive loop (continue reading from stdin) even after processing some source scripts.

Options and source files are processed in the order in which they are given on the command line. Processing of options and source files ends when either the -- or the -x option is encountered. The -x option must be followed by the name of a script to be executed, which becomes the "main script" of the application. In either case, any remaining parameters are passed to the executing script by means of the global argc and argv variables, denoting the number of arguments and the list of the actual parameter strings, respectively. In the case of -x this also includes the script name as argv!0. The -x option is useful, in particular, to turn Pure scripts into executable programs by including a "shebang" like the following as the first line in your main script. (This trick only works with Unix shells, though.)

#!/usr/local/bin/pure -x

On startup, the interpreter also defines the version variable, which is set to the version string of the Pure interpreter, and the sysinfo variable, which provides a string identifying the host system. These are useful if parts of your script depend on the particular version of the interpreter and the system it runs on. (Moreover, Pure 0.21 and later also define the variable compiling which indicates whether the program is executed in a batch compilation, see Compiling Scripts below.)

If available, the prelude script prelude.pure is loaded by the interpreter prior to any other definitions, unless the -n or --noprelude option is specified. The prelude is searched for in the directory specified with the PURELIB environment variable. If the PURELIB variable is not set, a system-specific default is used. Relative pathnames of other source scripts specified on the command line are interpreted relative to the current working directory. In addition, the executed program may load other scripts and libraries via a using declaration in the source, which are searched for in a number of locations, including the directories named with the -I and -L options; see the Declarations and C Interface sections for details.

2.3   Compiling Scripts

The interpreter compiles scripts, as well as definitions that you enter interactively, automatically. This is done in an incremental fashion, as the code is needed, and is therefore known as JIT (just in time) compilation. Thus the interpreter never really "interprets" the source program or some intermediate representation, it just acts as a frontend to the compiler, taking care of compiling source code to native machine code before it gets executed.

Pure's LLVM backend does "lazy JIT compilation" by default, meaning that each function (global or local) is compiled no sooner than it is run for the first time. With the --eager-jit option, however, it will also compile all other (global or local) functions that may be called by the compiled function. (The PURE_EAGER_JIT environment variable, when set to any value, has the same effect, so that you do not have to specify the --eager--jit option each time you run the interpreter.) Eager JIT compilation may be more efficient in some cases (since bigger chunks of compilation work can be done in one go) and less efficient in others (e.g., eager JITing may compile large chunks of code which aren't actually called later, except in rare circumstances).

Note that the eager JIT mode is only available with LLVM 2.7 or later; otherwise this option will be ignored.

It is also possible to compile your scripts to native code beforehand, using the -c batch compilation option. This options forces the interpreter to non-interactive mode (unless -i is specified as well, which overrides -c). Any scripts specified on the command line are then executed as usual, but after execution the interpreter takes a snapshot of the program and compiles it to one of several supported output formats, LLVM assembler (.ll) or bitcode (.bc), native assembler (.s) or object (.o), or a native executable, depending on the output filename specified with -o. If the output filename ends in the .ll extension, an LLVM assembler file is created which can then be processed with the LLVM toolchain. If the output filename is just '-', the assembler file is written to standard output, which is useful if you want to pass the generated code to the LLVM tools in a pipeline. If the output filename ends in the .bc extension, an LLVM bitcode file is created instead.

The .ll and .bc formats are supported natively by the Pure interpreter, no external tools are required to generate these. If the target is an .s, .o or executable file, the Pure interpreter creates a temporary bitcode file on which it invokes the LLVM tools opt and llc to create a native assembler file, and then uses gcc to assemble and link the resulting program (if requested). You can also specify additional libraries to be linked into the executable with the -l option. If the output filename is omitted, it defaults to a.out (a.exe on Windows).

The -c option provides a convenient way to quickly turn a Pure script into a standalone executable which can be invoked directly from the shell. One advantage of compiling your script is that this eliminates the JIT compilation time and thus considerably reduces the startup time of the program. Another reason to prefer a standalone executable is that it lets you deploy the program on systems without a full Pure installation (usually only the runtime library is required on the target system). On the other hand, compiled scripts also have some limitations, mostly concerning the use of the built-in eval function. Please see the Batch Compilation section for details.

The -v64 (or -v0100) verbosity option can be used to have the interpreter print the commands it executes during compilation, see Verbosity and Debugging Options below. When creating an object file, this also prints the suggested linker command (including all the dynamic modules loaded by the script, which also have to be linked in to create a working executable), to which you only have to add the options describing the desired output file.

2.4   Tagging Scripts

Pure programs often have declarations and definitions of global symbols scattered out over many different source files. The --ctags and --etags options let you create a tags file which allows you to quickly locate these items in text editors such as vi and emacs which support this feature.

If --ctags or --etags is specified, the interpreter enters a special mode in which it only parses source files without executing them and collects information about the locations of global symbol declarations and definitions. The collected information is then written to a tags file in the ctags or etags format used by vi and emacs, respectively. The desired name of the tags file can be specified with the -T option; it defaults to tags for --ctags and TAGS for --etags (which matches the default tags file names used by vi and emacs, respectively).

The tags file contains information about the global constant, variable, macro, function and operator symbols of all scripts specified on the command line, as well as the prelude and other scripts included via a using clause. Tagged scripts which are located in the same directory as the tags file (or, recursively, in one of its subdirectories) are specified using relative pathnames, while scripts outside this hierarchy (such as included scripts from the standard library) are denoted with absolute pathnames. This scheme makes it possible to move an entire directory together with its tags file and have the tags information still work in the new location.

2.5   Running Interactively

If the interpreter runs in interactive mode, it repeatedly prompts you for input (which may be any legal Pure code or some special interpreter commands provided for interactive usage), and prints computed results. This is also known as the read-eval-print loop and is described in much more detail in the Interactive Usage section. To exit the interpreter, just type the quit command or the end-of-file character (^D on Unix) at the beginning of the command line.

The interpreter may also source a few additional interactive startup files immediately before entering the interactive loop, unless the --norc option is specified. First .purerc in the user's home directory is read, then .purerc in the current working directory. These are ordinary Pure scripts which can be used to provide additional definitions for interactive usage. Finally, a .pure file in the current directory (containing a dump from a previous interactive session) is loaded if it is present.

When the interpreter is in interactive mode and reads from a tty, unless the --noediting option is specified, commands are usually read using readline(3) or some compatible replacement, providing completion for all commands listed under Interactive Usage, as well as for symbols defined in the running program. When exiting the interpreter, the command history is stored in ~/.pure_history, from where it is restored the next time you run the interpreter.

As of Pure 0.22, the interpreter also provides a simple source level debugger when run in interactive mode, see Debugging for details. To enable the interactive debugger, you need to specify the -g option when invoking the interpreter. This option causes your script to run much slower and also disables tail call optimization, so you should only use this option if you want to run the debugger.

2.6   Verbosity and Debugging Options

The -v option is useful for debugging the interpreter, or if you are interested in the code your program gets compiled to. The level argument is optional; it defaults to 1. Seven different levels are implemented at this time (one more bit is reserved for future extensions). For most purposes, only the first two levels will be useful for the average Pure programmer; the remaining levels are most likely to be used by the Pure interpreter developers.

1 (0x1, 001)
denotes echoing of parsed definitions and expressions.
2 (0x2, 002)
adds special annotations concerning local bindings (de Bruijn indices, subterm paths; this can be helpful to debug tricky variable binding issues).
4 (0x4, 004)
adds descriptions of the matching automata for the left-hand sides of equations (you probably want to see this only when working on the guts of the interpreter).
8 (0x8, 010)
dumps the "real" output code (LLVM assembler, which is as close to the native machine code for your program as it gets; you definitely don't want to see this unless you have to inspect the generated code for bugs or performance issues).
16 (0x10, 020)
adds debugging messages from the bison(1) parser; useful for debugging the parser.
32 (0x20, 040)
adds debugging messages from the flex(1) lexer; useful for debugging the lexer.
64 (0x40, 0100)
turns on verbose batch compilation; this is useful if you want to see exactly which commands get executed during batch compilation (-c).

These values can be or'ed together, and, for convenience, can be specified in either decimal, hexadecimal or octal. Thus 0xff or 0777 always gives you full debugging output (which isn't likely to be used by anyone but the Pure developers). Some useful flag combinations for experts are (in octal) 007 (echo definitions along with de Bruijn indices and matching automata), 011 (definitions and assembler code) and 021 (parser debugging output along with parsed definitions).

Note that the -v option is only applied after the prelude has been loaded. If you want to debug the prelude, use the -n option and specify the prelude.pure file explicitly on the command line. Verbose output is also suppressed for modules imported through a using clause. As a remedy, you can use the interactive show command (see the Interactive Usage section) to list definitions along with additional debugging information.

The -w option enables some additional warnings which are useful to check your scripts for possible errors. Right now it will report implicit declarations of function symbols which might indicate missing or mistyped symbols that need to be fixed, see Symbol Lookup and Creation for details.

2.7   Code Generation Options

Besides the options listed above, the interpreter also understands some additional command line switches and corresponding environment variables to control various code generation options. The options take the form --opt and --noopt, respectively, where opt denotes the option name (see below for a list of supported options). By default, these options are all enabled; --noopt disables the option, --opt reenables it. In addition, for each option opt there is also a corresponding environment variable PURE_NOOPT (with the option name in uppercase) which, when set, disables the option by default. (Setting this variable to any value will do, the interpreter only checks whether the variable exists in the environment.)

For instance, the checks option controls stack and signal checks. Thus --nochecks on the command line disables the option, and setting the PURE_NOCHECKS environment variable makes this the default, in which case you can use --checks on the command line to reenable the option.

Each code generation option can also be used as a pragma (compiler directive) in source code so that you can control it on a per-rule basis. The pragma must be on a line by itself, starting in column 1, and takes the following form (using --nochecks as an example):

#! --nochecks // line-oriented comment may go here

Currently, the following code generation options are recognized:

--checks, --nochecks
Enable or disable various extra stack and signal checks. By default, the interpreter checks for stack overflows (if the PURE_STACK environment variable is set) and pending signals on entry to every function, see Stack Size and Tail Recursion and Handling of Asynchronous Signals for details. This is needed to catch these conditions in a reliable way, so we recommend to leave this enabled. However, these checks also make programs run a little slower (typically some 5%, YMMV). If performance is critical then you can disable the checks with the --nochecks option. (Even then, a minimal amount of checking will be done, usually on entry to every global function.)
--const, --noconst
Enable or disable the precomputing of constant values in batch compilation (cf. Compiling Scripts). If enabled (which is the default), the values of constants in const definitions are precomputed at compile time (if possible) and then stored in the generated executable. This usually yields faster startup times but bigger executables. You can disable this option with --noconst to get smaller executables at the expense of slower startup times. Please see the Batch Compilation section for an example.
--fold, --nofold

Enable or disable constant folding in the compiler frontend. This means that constant expressions involving int and double values and the usual arithmetic and logical operations on these are precomputed at compile time. (This is mostly for cosmetic purposes; the LLVM backend will perform this optimization anyway when generating machine code.) For instance:

> foo x = 2*3*x;
> show foo
foo x = 6*x;

Disabling constant folding in the frontend causes constant expressions to be shown as you entered them:

> #! --nofold
> bar x = 2*3*x;
> show bar
bar x = 2*3*x;
--tc, --notc
Enable or disable tail call optimization (TCO). TCO is needed to make tail-recursive functions execute in constant stack space, so we recommend to leave this enabled. However, at the time of this writing LLVM's TCO support is still bug-ridden on some platforms, so the --notc option allows you to disable it. (Note that TCO can also be disabled when compiling the Pure interpreter, in which case these options have no effect; see the installation documentation for details.)

Besides these, there's also a special pragma for telling the batch compiler about "required" functions:

--required fun
Inform the batch compiler (cf. Compiling Scripts) that the given function symbol fun should never be stripped from the program. This is useful, e.g., if a function is never called explicitly but only through eval. Adding a --required pragma for the function then makes sure that the function is always linked into the program.

Note that the --required pragma can only be used in source code, there's no command line option for it. An example showing how to use this option can be found in the Batch Compilation section.

2.8   Startup Files

The interpreter may source various files during its startup. These are:

Interactive command history.
~/.purerc, .purerc, .pure
Interactive startup files. The latter is usually a dump from a previous interactive session.
Standard prelude. If available, this script is loaded before any other definitions, unless -n was specified.

2.9   Environment

Various aspects of the interpreter can be configured through the following shell environment variables:

If the PURE_HELP variable is not set (see below), this specifies a colon-separated list of browsers to try for reading the online documentation. See
Directory to search for library scripts, including the prelude. If PURELIB is not set, it defaults to some location specified at installation time.
Enable eager JIT compilation (same as --eager-jit), see Compiling Scripts for details.
Command used to browse the Pure manual. This must be a browser capable of displaying html files. Default is w3m(1).
Additional directories (in colon-separated format) to be searched for included scripts.
Additional directories (in colon-separated format) to be searched for dynamic libraries.
Shell command to be used for paging through output of the show command, when the interpreter runs in interactive mode.
Command prompt used in the interactive command loop ("> " by default).
Maximum stack size in kilobytes (default: 0 = unlimited).

(Besides these, the interpreter also understands a number of other environment variables for controlling various code generation options; please see Code Generation Options above for details.)

3   Pure Overview

Pure is a fairly simple yet powerful language. Programs are basically collections of rewriting rules and expressions to be evaluated. For convenience, it is also possible to define global variables and constants, and for advanced uses Pure offers macro functions as a kind of preprocessing facility. These are all described below and in the following sections.

Here's a first example which demonstrates how to define a simple recursive function in Pure, entered interactively in the interpreter (note that the '>' symbol at the beginning of each input line is the interpreter's default command prompt):

> // my first Pure example
> fact 0 = 1;
> fact n::int = n*fact (n-1) if n>0;
> let x = fact 10; x;

3.1   Lexical Matters

Pure is a free-format language; whitespace is insignificant, except if it serves to delimit other symbols. Hence, as shown above, definitions and expressions at the toplevel have to be terminated with a semicolon, even in interactive mode.

Comments have the same syntax as in C++ (using // for line-oriented and /* ... */ for multiline comments; the latter may not be nested). Lines beginning with #! are treated as comments, too; as already discussed above, on Unix-like systems this allows you to add a "shebang" to your main script in order to turn it into an executable program.

There are a few reserved keywords which cannot be used as identifiers:

case const def else end extern if infix infixl infixr let namespace
nonfix of otherwise outfix postfix prefix private public then using
when with

The customary notations for identifiers, numbers and strings are all provided. In addition, Pure also allows you to define your own operator symbols. Pure fully supports Unicode, so that you can write your programs in almost any language and make good use of the special symbols in the Unicode character set, provided that you encode your scripts in UTF-8. To keep this simple, besides the ASCII punctuation characters, Pure also considers the following code points in the Unicode repertoire as punctuation: U+00A1 through U+00BF, U+00D7, U+00F7, and U+20D0 through through U+2BFF. This comprises the special symbols in the Latin-1 repertoire, as well as the Combining Diacritical Marks for Symbols, Letterlike Symbols, Number Forms, Arrows, Mathematical Symbols, Miscellaneous Technical Symbols, Control Pictures, OCR, Enclosed Alphanumerics, Box Drawing, Blocks, Geometric Shapes, Miscellaneous Symbols, Dingbats, Miscellaneous Mathematical Symbols A, Supplemental Arrows A, Supplemental Arrows B, Miscellaneous Mathematical Symbols B, Supplemental Mathematical Operators, and Miscellaneous Symbols and Arrows. This should cover almost everything you'd ever want to use in an operator symbol. All other extended Unicode characters are effectively treated as "letters" which can be used as identifier constituents.

3.2   Definitions and Expression Evaluation

On the surface, Pure is quite similar to other modern functional languages like Haskell and ML. But under the hood it is a much more dynamic language, more akin to Lisp. In particular, Pure is dynamically typed, so functions can be fully polymorphic and you can add to the definition of an existing function at any time. For instance, we can extend the example above to make the fact function work with floating point numbers, too:

> fact 0.0 = 1.0;
> fact n::double = n*fact (n-1) if n>0;
> fact 10.0;
> fact 10;

Note the n::double construct on the left-hand side of the second equation, which means that the equation is only to be applied for (double precision) floating point values n. This construct is also called a "type tag" in Pure parlance, which is actually a simple form of pattern matching (see below). Similarly, our previous definition at the beginning of this section employed the int tag to indicate that the n parameter is an integer value. The int and double types are built into the Pure language, but it is also possible to introduce your own type tags for user-defined data structures. This will be explained in more detail under Type Tags in the Rule Syntax section below.

Expressions are generally evaluated from left to right, innermost expressions first, i.e., using call by value semantics. Pure also has a few built-in special forms (most notably, conditional expressions, the short-circuit logical connectives && and ||, the sequencing operator $$, the lazy evaluation operator &, and the quote) which take some or all of their arguments unevaluated, using call by name.

Like in Haskell and ML, functions are often defined by pattern matching, i.e., the left-hand side of a definition is compared to the target expression, binding the variables in the pattern to their actual values accordingly:

> foo (bar x) = x-1;
> foo (bar 99);

Due to its term rewriting semantics, Pure goes beyond most other functional languages in that it can do symbolic evaluations just as well as "normal" computations:

> square x = x*x;
> square 4;
> square (a+b);

Leaving aside the built-in support for some common data structures such as numbers and strings, all the Pure interpreter really does is evaluate expressions in a symbolic fashion, rewriting expressions using the equations supplied by the programmer, until no more equations are applicable. The result of this process is called a normal form which represents the "value" of the original expression. Keeping with the tradition of term rewriting, there's no distinction between "defined" and "constructor" function symbols in Pure. Consequently, any function symbol or operator can be used anywhere on the left-hand side of an equation, and may act as a constructor symbol if it happens to occur in a normal form term. This enables you to work with algebraic rules like associativity and distributivity in a direct fashion:

> (x+y)*z = x*z+y*z; x*(y+z) = x*y+x*z;
> x*(y*z) = (x*y)*z; x+(y+z) = (x+y)+z;
> square (a+b);

Note that, in contrast, languages like Haskell and ML always enforce the so-called "constructor discipline", which stipulates that only pure constructor symbols (without any defining equations) may occur as a subterm on the left-hand side of a definition. Thus equational definitions like the above are forbidden in these languages. In Pure such definitions are just normal business, which makes the language directly usable as a tool for symbolic algebra.

3.3   Parameters in Equations

Taking a look at the above examples, you might have been wondering how the Pure interpreter figures out what the parameters (a.k.a. "variables") in an equation are. This is quite obvious in rules involving just variables and special operator symbols, such as (x+y)*z = x*z+y*z. However, what about an equation like foo (foo bar) = bar? Since most of the time we don't declare any symbols in Pure, how does the interpreter know that foo is a literal function symbol here, while bar is a variable?

The answer is that the interpreter considers the different positions in the left-hand side expression of an equation. Basically, a Pure expression is just a tree formed by applying expressions to other expressions, with the atomic subexpressions like numbers and symbols at the leaves of the tree. (This is even true for infix expressions like x+y, since in Pure these are always equivalent to a function application of the form (+) x y which has the atomic subterms (+), x and y at its leaves.)

Now the interpreter divides the leaves of the expression tree into "head" (or "function") and "parameter" (or "variable") positions based on which leaves are leftmost in a function application or not. Thus, in an expression like f x y z, f is in the head or function position, while x, y and z are in parameter or variable positions. (Note that in an infix expression like x+y, (+) is the head symbol, not x, as the expression is really parsed as (+) x y, see above.)

Identifiers in head positions are taken as literal function symbols by the interpreter, while identifiers in variable positions denote, well, variables. We also refer to this convention as the head = function rule. It is quite intuitive and lets us get away without declaring the variables in equations. (There are some corner cases not covered here, however. In particular, Pure allows you to declare special constant symbols, if you need a symbol to be recognized as a literal even if it occurs in a variable position. This is done by means of a nonfix declaration, see Symbol Declarations for details.)

3.4   Expression Syntax

The Pure language provides built-in support for machine integers (32 bit), bigints (implemented using GMP), floating point values (double precision IEEE 754), character strings (UTF-8 encoded) and generic C pointers (these don't have a syntactic representation in Pure, though, so they need to be created with external C functions). Truth values are encoded as machine integers (as you might expect, zero denotes false and any non-zero value true). Pure also provides some built-in support for lists and matrices, although most of the corresponding operations are actually defined in the prelude.

Expressions consist of the following elements:

Numbers: 4711, 4711L, 1.2e-3
The usual C notations for integers (decimal: 1000, hexadecimal: 0x3e8, octal: 01750) and floating point values are all provided. Integers can also be denoted in base 2 by using the 0b or 0B prefix: 0b1111101000. Integer constants that are too large to fit into machine integers are promoted to bigints automatically. Moreover, integer literals immediately followed by the uppercase letter L are always interpreted as bigint constants, even if they fit into machine integers. This notation is also used when printing bigint constants, to distinguish them from machine integers.
Strings: "Hello, world!\n"
String constants are double-quoted and terminated with a null character, like in C. In difference to C, strings are always encoded in UTF-8, and character escapes in Pure strings have a more flexible syntax (borrowed from the author's Q language) which provides notations to specify any Unicode character. In particular, the notation \n, where n is an integer literal written in decimal (no prefix), hexadecimal (0x prefix), octal (0 prefix) or binary (0b prefix) notation, denotes the Unicode character (code point) #n. Since these escapes may consist of a varying number of digits, parentheses may be used for disambiguation purposes; thus, e.g. "\(123)4" denotes character #123 followed by the ASCII character 4. The usual C-like escapes for special non-printable characters such as \n are also supported. Moreover, you can use symbolic character escapes of the form \&name;, where name is any of the XML single character entity names specified in the XML Entity definitions for Characters. Thus, e.g., "\&copy;" denotes the copyright character (code point 0x000A9).
Function and variable symbols: foo, foo_bar, BAR, foo::bar
These consist of the usual sequence of letters (including the underscore) and digits, starting with a letter. Case is significant, thus foo, Foo and FOO are distinct identifiers. The '_' symbol, when occurring on the left-hand side of an equation, is special; it denotes the anonymous variable which matches any value without actually binding a variable. Identifiers can also be prefixed with a namespace identifier, like in foo::bar. (This requires that the given namespace has already been created, as explained under Namespaces in the Declarations section.)
Operator and constant symbols: x+y, x==y, not x

For convenience, Pure also provides you with a limited means to extend the syntax of the language with special operator and constant symbols by means of a corresponding fixity declaration, as discussed in section Symbol Declarations. Besides the usual infix, prefix and postfix operators, Pure also provides outfix (bracket) and nonfix (constant) symbols. (Nonfix symbols actually work more or less like ordinary identifiers, but the nonfix attribute tells the compiler that when such a symbol occurs on the left-hand side of an equation, it is always to be interpreted as a literal constant, cf. Parameters in Equations.)

Operator and constant symbols may take the form of an identifier or a sequence of punctuation characters. They must always be declared before use. Once declared, they are always special, and can't be used as ordinary identifiers any more. However, like in Haskell, by enclosing an operator in parentheses, such as (+) or (not), you can turn it into an ordinary function symbol. Also, operators and constant symbols can be qualified with a namespace just like normal identifiers.

Lists: [x,y,z], x:xs, x..y, x:y..z

Pure's basic list syntax is the same as in Haskell, thus [] is the empty list and x:xs denotes a list with head element x and tail list xs. (The infix constructor symbol ':' is declared in the prelude.) The usual syntactic sugar for list values in brackets is provided, thus [x,y,z] is exactly the same as x:y:z:[].

There's also a way to denote arithmetic sequences such as 1..5, which denotes the list [1,2,3,4,5]. (Haskell users should note the missing brackets. In difference to Haskell, Pure doesn't use any special syntax for arithmetic sequences, the '..' symbol is just an ordinary infix operator declared and defined in the prelude.) Sequences with arbitrary stepsizes can be written by denoting the first two sequence elements using the ':' operator, as in 1.0:1.2..3.0. (To prevent unwanted artifacts due to rounding errors, the upper bound in a floating point sequence is always rounded to the nearest grid point. Thus, e.g., 0.0:0.1..0.29 actually yields [0.0,0.1,0.2,0.3], as does 0.0:0.1..0.31.)

Tuples: x,y,z
Pure's tuples are a bit unusual: They are constructed by just "pairing" things using the ',' operator, for which the empty tuple () acts as a neutral element (i.e., (),x is just x, as is x,()). Pairs always associate to the right, meaning that x,y,z == x,(y,z) == (x,y),z, where x,(y,z) is the normalized representation. This implies that tuples are always flat, i.e., there are no nested tuples (tuples of tuples); if you need such constructs then you should use lists instead. Also note that parentheses are generally only used to group expressions and are not part of the tuple syntax in Pure, except if you need to include a tuple in a list or matrix. E.g., [(1,2),3,(4,5)] is a three element list consisting of the tuple 1,2, the integer 3, and another tuple 4,5. Likewise, [(1,2,3)] is a list with a single element, the tuple 1,2,3.
Matrices: {1.0,2.0,3.0}, {1,2;3,4}, {1L,y+1;foo,bar}

Pure also offers matrices, a kind of two-dimensional arrays, as a built-in data structure which provides efficient storage and element access. These work more or less like their Octave/MATLAB equivalents, but using curly braces instead of brackets. As indicated, commas are used to separate the columns of a matrix, semicolons for its rows. In fact, the {...} construct is rather general and allows you to construct new matrices from any collection of individual elements ("scalars") and submatrices, provided that all dimensions match up. Here, any expression which doesn't yield a matrix denotes a scalar, which is considered to be a 1x1 matrix for the purpose of matrix construction. The comma arranges submatrices in columns, while the semicolon arranges them in rows. So, if both x and y are nxm matrices, then {x,y} becomes an n x 2*m matrix consisting of all the columns of x followed by all the columns of y. Likewise, {x;y} becomes a 2*n x m matrix (all the rows of x above of all rows of y). In addition, {...} constructs can be nested to an arbitrary depth. Thus {{1;3},{2;4}} is another way to write the 2x2 matrix {1,2;3,4} in a kind of "column-major" format (however, internally all matrices are stored in C's row-major format).

It is important to note here that, while the [...] and {...} constructs look superficially similar, they work in very different ways. The former is just syntactic sugar for a corresponding constructor term and can thus be used as a pattern on the left-hand side of an equation. In contrast, the latter is special syntax for a built-in operation which creates objects of a special matrix type. Thus matrix expressions can not be used as patterns (instead, matrix values can be matched as a whole using the special matrix type tag, see the Rule Syntax section for details).

Pure supports both numeric and symbolic matrices. The former are homogeneous arrays of double, complex double or (machine) int matrices, while the latter can contain any mixture of Pure expressions. Pure will pick the appropriate type for the data at hand. If a matrix contains values of different types, or Pure values which cannot be stored in a numeric matrix, then a symbolic matrix is created instead (this also includes the case of bigints, which are considered as symbolic values as far as matrix construction is concerned). Numeric matrices use an internal data layout that is fully compatible with the GNU Scientific Library (GSL), and can readily be passed to GSL routines via the C interface. (The Pure interpreter does not require GSL, however, so numeric matrices will work even if GSL is not installed.)

More information about matrices and corresponding examples can be found in the Examples section below.

Comprehensions: [x,y | x=1..n; y=1..m; x<y], {f x | x=1..n}

Pure provides both list and matrix comprehensions as a convenient means to construct list and matrix values from a "template" expression and one or more "generator" and "filter" clauses. The former bind a pattern to values drawn from a list or matrix, the latter are just predicates determining which generated elements should actually be added to the result. Both list and matrix comprehensions are in fact syntactic sugar for a combination of nested lambdas, conditional expressions and "catmaps" (a collection of operations which combine list or matrix construction and mapping a function over a list or matrix, defined in the prelude), but they are often much easier to write.

Matrix comprehensions work pretty much like list comprehensions, but produce matrices instead of lists. List generators in matrix comprehensions alternate between row and column generation so that most common mathematical abbreviations carry over quite easily. Examples of both kinds of comprehensions can be found in the Examples section below.

Function and operator applications: foo x y z, -x, x+y, (+x)

As in other modern FPLs, function applications are written simply as juxtaposition (i.e., in "curried" form) and associate to the left. This means that in fact all functions only take a single argument. Multi-argument functions are represented as chains of single-argument functions. For instance, in f x y = (f x) y first the function f is applied to the first argument x, yielding the function f x which in turn gets applied to the second argument y. This makes it possible to derive new functions from existing ones using partial applications which only specify some but not all arguments of a function. For instance, taking the max function from the prelude as an example, max 0 is the function which, for a given x, returns x itself if it is nonnegative and zero otherwise. This works because (max 0) x = max 0 x is the maximum of 0 and x.

Operator applications are written using prefix, postfix, outfix or infix notation, as the declaration of the operator demands, but are just ordinary function applications in disguise. As already mentioned, enclosing an operator in parentheses turns it into an ordinary function symbol, thus x+y is exactly the same as (+) x y. For convenience, partial applications of infix operators can also be written using so-called operator sections. A left section takes the form (x+) which is equivalent to the partial application (+) x. A right section takes the form (+x) and is equivalent to the term flip (+) x. (This uses the flip combinator from the prelude which is defined as flip f x y = f y x.) Thus (x+) y is equivalent to x+y, while (+x) y reduces to y+x. For instance, (1/) denotes the reciprocal and (+1) the successor function. (Note that, in contrast, (-x) always denotes an application of unary minus; the section (+-x) can be used to indicate a function which subtracts x from its argument.)

Conditional expressions: if x then y else z
Evaluates to y or z depending on whether x is "true" (i.e., a nonzero integer). An exception is raised if the condition is not an integer.
Lambdas: \x -> y
These denote anonymous functions and work pretty much like in Haskell. Pure supports multiple-argument lambdas (e.g, \x y -> x*y), as well as pattern-matching lambda abstractions which match one or more patterns against the lambda arguments, such as \(x,y) -> x*y. An exception is raised if the actual lambda arguments do not match the given patterns.
Case expressions: case x of rule; ... end
Matches an expression, discriminating over a number of different cases, similar to the Haskell case construct. The expression x is matched in turn against each left-hand side pattern in the rule list, and the first pattern which matches x gives the value of the entire expression, by evaluating the corresponding right-hand side with the variables in the pattern bound to their corresponding values. An exception is raised if the target expression doesn't match any of the patterns.
When expressions: x when rule; ... end
An alternative way to bind local variables by matching a collection of subject terms against corresponding patterns, similar to Aardappel's when construct. A single binding such as x when u = v end is equivalent to case v of u = x end, but the former is often more convenient to write. A when clause may contain multiple definitions, which are processed from left to right, so that later definitions may refer to the variables in earlier ones. This is exactly the same as several nested single definitions, with the first binding being the "outermost" one.
With expressions: x with rule; ... end
Defines local functions. Like Haskell's where construct, but it can be used anywhere inside an expression (just like Aardappel's where, but Pure uses the keyword with which better lines up with case and when). Several functions can be defined in a single with clause, and the definitions can be mutually recursive and consist of as many equations as you want.

3.5   Operators and Precedence

Expressions are parsed according to the following precedence rules: Lambda binds most weakly, followed by when, with and case, followed by conditional expressions (if-then-else), followed by the simple expressions, i.e., all other kinds of expressions involving operators, function applications, constants, symbols and other primary expressions. Precedence and associativity of operator symbols are given by their declarations (cf. Symbol Declarations), and function application binds stronger than all operators. Parentheses can be used to override default precedences and associativities as usual.

The common operator symbols like +, -, *, / etc. are all declared at the beginning of the prelude, see the Pure Library Manual for a list of these. Arithmetic and relational operators mostly follow C conventions. However, out of necessity (!, & and | are used for other purposes in Pure) the logical and bitwise operations, as well as the negated equality predicates are named a bit differently: ~, && and || denote logical negation, conjunction and disjunction, while the corresponding bitwise operations are named not, and and or. Moreover, following these conventions, inequality is denoted ~=. Also note that && and || are special forms which are evaluated in short-circuit mode (see below), whereas the bitwise connectives receive their arguments using call-by-value, just like the other arithmetic operations.

3.6   Special Forms

As already mentioned, some operations are actually implemented as special forms. In particular, the conditional expression if x then y else z is a special form with call-by-name arguments y and z; only one of the branches is actually evaluated, depending on the value of x. Similarly, the logical connectives && and || evaluate their operands in short-circuit mode. Thus, e.g., x&&y immediately becomes false if x evaluates to false, without ever evaluating y. The built-in definitions of these operations work as if they were defined by the following equations (but note that the second operand is indeed passed "by name", so that it isn't evaluated unless its value is actually needed):

x::int && y = if x then y else x;
x::int || y = if x then x else y;

(Note that this isn't quite the same as in C, as the results of these operations are not normalized, i.e., they may return nonzero values other than 1 to denote "true". This has the advantage that && and || can be implemented tail-recursively, see Stack Size and Tail Recursion. Thus, if you need a normalized truth value then you'll have to make sure that either both operands are already normalized, or you'll have to normalize the result yourself.)

If the built-in definitions of && and || fail then both operands are evaluated and the resulting symbolic term is reduced using the equations specified by the programmer, which allows you to extend the definitions of && and || just as with other builtins. Also note that if the result term is in normal form then it is returned as is, thus && and || can be used in symbolic evaluations as usual. For instance, the following equations let you compute the disjunctive normal form of logical expressions:

// eliminate double negations:
~~a           = a;

// de Morgan's laws:
~(a || b)     = ~a && ~b;
~(a && b)     = ~a || ~b;

// distributivity:
a && (b || c) = a && b || a && c;
(a || b) && c = a && c || b && c;

// associativity:
(a && b) && c = a && (b && c);
(a || b) || c = a || (b || c);


> a || ~(b || (c && ~d));

The sequencing operator $$ evaluates its left operand, immediately throws the result away and then goes on to evaluate the right operand which gives the result of the entire expression. This operator is useful to write imperative-style code such as the following prompt-input interaction:

> using system;
> puts "Enter a number:" $$ scanf "%g";
Enter a number:

We mention in passing here that the same effect can be achieved with a when clause, which also allows you to execute a function solely for its side-effects and just ignore the return value:

> scanf "%g" when puts "Enter a number:" end;
Enter a number:

The & operator does lazy evaluation. This is the only postfix operator defined in the standard prelude, written as x&, where x is an arbitrary Pure expression. It turns its operand into a kind of parameterless anonymous closure, deferring its evaluation. These kinds of objects are also commonly known as thunks or futures. When the value of a future is actually needed (during pattern-matching, or when the value becomes an argument of a C call), it is evaluated automatically and gets memoized, i.e., the computed result replaces the thunk so that it only has to be computed once. Futures are useful to implement all kinds of lazy data structures in Pure, in particular: lazy lists a.k.a. streams. A stream is simply a list with a thunked tail, which allows it to be infinite. The Pure prelude defines many functions for creating and manipulating these kinds of objects; further details and examples can be found in the Examples section below.

Last but not least, the special form quote quotes an expression, i.e., quote x (or, equivalently, 'x) returns just x itself without evaluating it. The prelude also provides a function eval which can be used to evaluate a quoted expression at a later time. For instance:

> let x = '(2*42+2^12); x;
> eval x;

The quote also inhibits evaluation inside matrix values, including the "splicing" of embedded submatrices:

> '{1,2+3,2*3};
> '{1,{2,3},4};

The quote should be well familiar to Lisp programmers. However, there are some notable differences, please see The Quote in the Caveats and Notes section for details and more examples.

3.7   Toplevel

At the toplevel, a Pure program basically consists of rewriting rules (which are used to define functions and macros), constant and variable definitions, and expressions to be evaluated:

Rules: lhs = rhs;
Rewriting rules always combine a left-hand side pattern (which must be a simple expression) and a right-hand side (which can be any kind of Pure expression described above). The same format is also used in with, when and case expressions. In toplevel rules, with and case expressions, this basic form can also be augmented with a condition if guard tacked on to the end of the rule, where guard is an integer expression which determines whether the rule is applicable. Moreover, the keyword otherwise may be used to denote an empty guard which is always true (this is syntactic sugar to point out the "default" case of a definition; the interpreter just treats this as a comment). Pure also provides some abbreviations for factoring out common left-hand or right-hand sides in collections of rules; see the Rule Syntax section for details.
Macro rules: def lhs = rhs;
A rule starting with the keyword def defines a macro function. No guards or multiple left-hand and right-hand sides are permitted here. Macro rules are used to preprocess expressions on the right-hand side of other definitions at compile time, and are typically employed to implement user-defined special forms and simple kinds of optimization rules. See the Macros section below for details and examples.
Global variable bindings: let lhs = rhs;
Binds every variable in the left-hand side pattern to the corresponding subterm of the right-hand side (after evaluating it). This works like a when clause, but serves to bind global variables occurring free on the right-hand side of other function and variable definitions.
Constant bindings: const lhs = rhs;
An alternative form of let which defines constants rather than variables. (These are not to be confused with nonfix symbols which simply stand for themselves!) Like let, this construct binds the variable symbols on the left-hand side to the corresponding values on the right-hand side (after evaluation). The difference is that const symbols can only be defined once, and thus their values do not change during program execution. This also allows the compiler to apply some special optimizations such as constant folding.
Toplevel expressions: expr;
A singleton expression at the toplevel, terminated with a semicolon, simply causes the given value to be evaluated (and the result to be printed, when running in interactive mode).

3.8   Scoping Rules

A few remarks about the scope of identifiers and other symbols are in order here. Like most modern functional languages, Pure uses lexical or static binding for local functions and variables. What this means is that the binding of a local name is completely determined at compile time by the surrounding program text, and does not change as the program is being executed. In particular, if a function returns another (anonymous or local) function, the returned function captures the environment it was created in, i.e., it becomes a (lexical) closure. For instance, the following function, when invoked with a single argument x, returns another function which adds x to its argument:

> foo x = bar with bar y = x+y end;
> let f = foo 99; f;
> f 10, f 20;

This works the same no matter what other bindings of x may be in effect when the closure is invoked:

> let x = 77; f 10, (f 20 when x = 88 end);

Global bindings of variable and function symbols work a bit differently, though. Like many languages which are to be used interactively, Pure binds global symbols dynamically, so that the bindings can be changed easily at any time during an interactive session. This is mainly a convenience for interactive usage, but works the same no matter whether the source code is entered interactively or being read from a script, in order to ensure consistent behaviour between interactive and batch mode operation.

So, for instance, you can easily bind a global variable to a new value by just entering a corresponding let command:

> foo x = c*x;
> foo 99;
> let c = 2; foo 99;
> let c = 3; foo 99;

This works pretty much like global variables in imperative languages, but note that in Pure the value of a global variable can only be changed with a let command at the toplevel. Thus referential transparency is unimpaired; while the value of a global variable may change between different toplevel expressions, it will always take the same value in a single evaluation.

Similarly, you can also add new equations to an existing function at any time:

> fact 0 = 1;
> fact n::int = n*fact (n-1) if n>0;
> fact 10;
> fact 10.0;
fact 10.0
> fact 1.0 = 1.0;
> fact n::double = n*fact (n-1) if n>1;
> fact 10.0;
> fact 10;

(In interactive mode, it is even possible to completely erase a definition, see section Interactive Usage for details.)

So, while the meaning of a local symbol never changes once its definition has been processed, toplevel definitions may well evolve while the program is being processed, and the interpreter will always use the latest definitions at a given point in the source when an expression is evaluated. This means that, even in a script file, you have to define all symbols needed in an evaluation before entering the expression to be evaluated.

4   Rule Syntax

Basically, the same rule syntax is used in all kinds of global and local definitions. However, some constructs (specifically, when, let, const and def) use a restricted rule syntax where no guards or multiple left-hand and right-hand sides are permitted. When matching against a function or macro call, or the subject term in a case expression, the rules are always considered in the order in which they are written, and the first matching rule (whose guard evaluates to a nonzero value, if applicable) is picked. (Again, the when construct is treated differently, because each rule is actually a separate definition.)

4.1   Patterns

The left-hand side of a rule is a special kind of simple expression, called a pattern. Patterns consist of function and operator applications as well as any of the "atomic" expression types (symbols, numbers, strings and list values). Not permitted are any of the special expression types (lambda, case, when, with, conditional expressions, as well as list and matrix comprehensions). For technical reasons, the current implementation also forbids matrix values in patterns, but it is possible to match a matrix value as a whole using the matrix type tag, see below.

As already mentioned, the '_' symbol is special in patterns; it denotes the anonymous variable which matches an arbitrary value (independently for all occurrences) without actually binding a variable. For instance:

foo _ _ = 0;

This will match the application of foo to any combination of two arguments (and just ignore the values of these arguments).

Constants in patterns must be matched literally. For instance:

foo 0 = 1;

This will only match an application of foo to the machine integer 0, not 0.0 or 0L (even though these compare equal to 0 using the '==' operator).

In difference to Haskell, patterns may contain repeated variables (other than the anonymous variable), i.e., they may be non-linear. Thus rules like the following are legal in Pure, and will only be matched if all occurrences of the same variable in the left-hand side pattern are matched to the same value:

> foo x x = x;
> foo 1 1;
> foo 1 2;
foo 1 2

Non-linear patterns are particularly useful for computer algebra where you will frequently encounter rules such as the following:

> x*y+x*z = x*(y+z);
> a*(3*4)+a*5;

The notion of "sameness" employed here is that of syntactical identity, which means that the matched subterms must be identical in structure and content. The prelude provides syntactic equality as a function same and a comparison predicate '==='. Thus the above definition of foo is roughly equivalent to the following:

foo x y = x if same x y;

It is important to note the differences between syntactic equality embodied by same and '===', and the "semantic" equality operator '=='. The former are always defined on all terms, whereas '==' is only available on data where it has been defined explicitly, either in the prelude or by the programmer. Also note that '==' may assert that two terms are equal even if they are syntactically different. Consider, e.g.:

> 0==0.0;
> 0===0.0;

This distinction is actually quite useful. It gives the programmer the flexibility to define '==' in any way that he sees fit, which is consistent with the way the other comparison operators like '<' and '>' are handled in Pure.

Patterns may also contain the following special elements which are not permitted in right-hand side expressions:

  • A Haskell-style "as" pattern of the form variable @ pattern binds the given variable to the expression matched by the subpattern pattern (in addition to the variables bound by pattern itself). This is convenient if the value matched by the subpattern is to be used on the right-hand side of an equation.
  • A left-hand side variable (including the anonymous variable) may be followed by a type tag of the form :: name, where name is either one of the built-in type symbols int, bigint, double, string, matrix, pointer, or an existing identifier denoting a custom constructor symbol for a user-defined data type. The variable can then match only values of the designated type. Thus, for instance, 'x::int' only matches machine integers. See the Type Tags section below for details.

Syntactically, both "as" patterns and type tags are primary expressions. If the subpattern of an "as" pattern is a compound expression, it must be parenthesized. For instance, the following function duplicates the head element of a list:

foo xs@(x:_) = x:xs;

Note that if you accidentally forget the parentheses around the subpattern x:_, you still get a syntactically correct definition:

foo xs@x:_ = x:xs;

But this gets parsed as (foo xs@x):_ = x:xs, which is most certainly not what you want. It is thus a good idea to just always enclose the subpattern with parentheses in order to prevent such glitches.

Another potential pitfall is that the notation foo::bar is also used to denote "qualified symbols" in Pure, cf. Namespaces. Usually this will be resolved correctly, but if foo happens to also be a valid namespace then most likely you'll get an error message about an undeclared symbol. You can always work around this by adding spaces around the '::' symbol, as in foo :: bar. Spaces are never permitted in qualified symbols, so this makes it clear that the construct denotes a type tag.

4.2   Type Tags

Type tags are really nothing but a special form of "as" patterns which restrict variables to given data "domains". Like Lisp, Pure is essentially a typeless language and doesn't really have a notion of "data types"; all data belongs to the same universe of terms. Thus the only way to restrict the type of values which can be matched by a variable is to provide an "as" pattern which matches objects of the desired domain and nothing else. However, in the special case of the built-in types (machine and big integers, double values, strings, matrices and pointers) there is no way to spell out all "constructors", as there are infinitely many (or none, as in the case of matrix and pointer which are constructed and inspected using special primitives, but are otherwise "opaque" at the Pure level). As a remedy, an appropriate type tag makes it possible to match these values.

In order to generalize this to user-defined domains of data, Pure adopts the convention that any other tagged variable x::bar is just a shorthand for the "as" pattern x@(bar _). Thus a custom data type can be represented by designating a special constructor symbol (bar in the above example) which takes the actual data as its single argument. This works the same no matter what the internal structure of the data is (which in many cases you wouldn't want to depend on anyway, for the sake of data abstraction).

Note that this is merely a convention, but it works reasonably well and makes up for the fact that Pure doesn't support data types at the language level. For instance, we might represent points in the plane using a constructor symbol Point which gets applied to pairs of coordinates. We equip this data type with an operation point to construct a point from its coordinates, and two operations xcoord and ycoord to retrieve the coordinates:

point x y = Point (x,y);
xcoord (Point (x,y)) = x;
ycoord (Point (x,y)) = y;

Now we might define a function translate which shifts the coordinates of a point by a given amount in the x and y directions as follows:

translate (x,y) p::Point = point (xcoord p+x) (ycoord p+y);

Note the use of Point as a type tag on the p variable. By these means, we can ensure that the argument is actually an instance of the point data type, without knowing anything about the internal representation. We can use these definitions as follows:

> let p::Point = point 3 3;
> p; translate (1,2) p;
Point (3,3)
Point (4,5)

Some data types in Pure's standard library (specifically, the container data types) are actually represented in this fashion, see the Pure Library Manual for details.

4.3   General Rules

The most general type of rule, used in function definitions and case expressions, consists of a left-hand side pattern, a right-hand side expression and an optional guard. The left-hand side of a rule can be omitted if it is the same as for the previous rule. This provides a convenient means to write out a collection of equations for the same left-hand side which discriminates over different conditions:

lhs       = rhs if guard;
          = rhs if guard;
          = rhs otherwise;

For instance:

fact n  = n*fact (n-1) if n>0;
        = 1 otherwise;

Pure also allows a collection of rules with different left-hand sides but the same right-hand side(s) to be abbreviated as follows:

lhs       |
lhs       = rhs;

This is useful if you need different specializations of the same rule which use different type tags on the left-hand side variables. For instance:

fact n::int    |
fact n::double |
fact n         = n*fact(n-1) if n>0;
               = 1 otherwise;

In fact, the left-hand sides don't have to be related at all, so that you can also write something like:

foo x | bar y = x*y;

However, this construct is most useful when using an "as" pattern to bind a common variable to a parameter value after checking that it matches one of several possible argument patterns (which is slightly more efficient than using an equivalent type-checking guard). E.g., the following definition binds the xs variable to the parameter of foo, if it is either the empty list or a list starting with an integer:

foo xs@[] | foo xs@(_::int:_) = ... xs ...;

The same construct also works in case expressions, which is convenient if different cases should be mapped to the same value, e.g.:

case ans of "y" | "Y" = 1; _ = 0; end;

Sometimes it is useful if local definitions (when and with) can be shared by the right-hand side and the guard of a rule. This can be done by placing the local definitions behind the guard, as follows (we only show the case of a single when clause here, but of course there may be any number of when and with clauses behind the guard):

lhs = rhs if guard when defns end;

Note that this is different from the following, which indicates that the definitions only apply to the guard but not the right-hand side of the rule:

lhs = rhs if (guard when defns end);

Conversely, definitions placed before the guard only apply to the right-hand side but not the guard (no parentheses are required in this case):

lhs = rhs when defns end if guard;

An example showing the use of a local variable binding spanning both the right-hand side and the guard of a rule is the following quadratic equation solver, which returns the (real) solutions of the equation x^2+p*x+q = 0 if the discriminant d = p^2/4-q is nonnegative:

> using math;
> solve p q = -p/2+sqrt d,-p/2-sqrt d if d>=0 when d = p^2/4-q end;
> solve 4 2; solve 2 4;
solve 2 4

Note that the above definition leaves the case of a negative discriminant undefined.

4.4   Simple Rules

As already mentioned, when, let and const use a simplified kind of rule syntax which just consists of a left-hand and a right-hand side separated by the equals sign. In this case the meaning of the rule is to bind the variables in the left-hand side of the rule to the corresponding subterms of the value of the right-hand side. This is also called a pattern binding.

Guards or multiple left-hand or right-hand sides are not permitted in these rules. However, it is possible to omit the left-hand side if it is just the anonymous variable '_' by itself, indicating that you don't care about the result. The right-hand side is still evaluated, if only for its side-effects, which is handy, e.g., for adding debugging statements to your code. For instance, here is a variation of the quadratic equation solver which also prints the discriminant after it has been computed:

> using math, system;
> solve p q = -p/2+sqrt d,-p/2-sqrt d if d>=0
> when d = p^2/4-q; printf "The discriminant is: %g\n" d; end;
> solve 4 2;
The discriminant is: 2
> solve 2 4;
The discriminant is: -3
solve 2 4

Note that simple rules of the same form lhs = rhs are also used in macro definitions (def), to be discussed in the Macros section. In this case, however, the rule denotes a real rewriting rule, not a pattern binding, hence the left-hand side is mandatory in these rules.

5   Examples

Here are a few examples of simple Pure programs.

The factorial:

fact n = n*fact (n-1) if n>0;
       = 1 otherwise;
let facts = map fact (1..10); facts;

The Fibonacci numbers:

fib n = a when a,b = fibs n end
          with fibs n = 0,1 if n<=0;
                      = case fibs (n-1) of
                          a,b = b,a+b;
let fibs = map fib (1..30); fibs;

It is worth noting here that Pure performs tail call optimization so that tail-recursive definitions like the following will be executed in constant stack space (see Stack Size and Tail Recursion in the Caveats and Notes section for more details on this):

// tail-recursive factorial using an "accumulating parameter"
fact n = loop 1 n with
  loop p n = if n>0 then loop (p*n) (n-1) else p;

Here is an example showing how constants are defined and used. Constant definitions take pretty much the same form as variable definitions with let (see above), but work more like the definition of a parameterless function whose value is precomputed at compile time:

> extern double atan(double);
> const pi = 4*atan 1.0;
> pi;
> foo x = 2*pi*x;
> show foo
foo x = 6.28318530717959*x;

Note that the compiler normally computes constant subexpression at compile time, such as 2*pi in the foo function. This works with all simple scalars (machine ints and doubles), see Constant Definitions for details.

5.1   List Comprehensions

List comprehensions are Pure's main workhorse for generating and processing all kinds of list values. Here's a well-known example, a variation of Erathosthenes' classical prime sieve:

primes n        = sieve (2..n) with
  sieve []      = [];
  sieve (p:qs)  = p : sieve [q | q = qs; q mod p];

(This definition is actually rather inefficient, there are much better albeit more complicated implementations of this sieve.)

For instance:

> primes 100;

If you dare, you can actually have a look at the catmap-lambda-if-then-else expression the comprehension expanded to:

> show primes
primes n = sieve (2..n) with sieve [] = []; sieve (p:qs) = p:sieve
(catmap (\q -> if q mod p then [q] else []) qs) end;

List comprehensions are also a useful device to organize backtracking searches. For instance, here's an algorithm for the n queens problem, which returns the list of all placements of n queens on an n x n board (encoded as lists of n pairs (i,j) with i = 1..n), so that no two queens hold each other in check:

queens n       = search n 1 [] with
  search n i p = [reverse p] if i>n;
               = cat [search n (i+1) ((i,j):p) | j = 1..n; safe (i,j) p];
  safe (i,j) p = ~any (check (i,j)) p;
  check (i1,j1) (i2,j2)
               = i1==i2 || j1==j2 || i1+j1==i2+j2 || i1-j1==i2-j2;

(Again, this algorithm is rather inefficient, see the examples included in the Pure distribution for a much better algorithm by Libor Spacek.)

5.2   Lazy Evaluation and Streams

As already mentioned, lists can also be evaluated in a "lazy" fashion, by just turning the tail of a list into a future. This special kind of list is also called a stream. Streams enable you to work with infinite lists (or finite lists which are so huge that you would never want to keep them in memory in their entirety). E.g., here's one way to define the infinite stream of all Fibonacci numbers:

> let fibs = fibs 0L 1L with fibs a b = a : fibs b (a+b) & end;
> fibs;
0L:#<thunk 0xb5d54320>

Note the & on the tail of the list in the definition of the local fibs function. This turns the result of fibs into a stream, which is required to prevent the function from recursing into samadhi. Also note that we work with bigints in this example because the Fibonacci numbers grow quite rapidly, so with machine integers the values would soon start wrapping around to negative integers.

Streams like these can be worked with in pretty much the same way as with lists. Of course, care must be taken not to invoke "eager" operations such as # (which computes the size of a list) on infinite streams, to prevent infinite recursion. However, many list operations work with infinite streams just fine, and return the appropriate stream results. E.g., the take function (which retrieves a given number of elements from the front of a list) works with streams just as well as with "eager" lists:

> take 10 fibs;
0L:#<thunk 0xb5d54350>

Hmm, not much progress there, but that's just how streams work (or rather they don't, they're lazy bums indeed!). Nevertheless, the stream computed with take is in fact finite and we can readily convert it to an ordinary list, forcing its evaluation:

> list (take 10 fibs);

An easier way to achieve this is to cut a "slice" from the stream:

> fibs!!(0..10);

Also note that since we bound the stream to a variable, the already computed prefix of the stream has been memoized, so that this portion of the stream is now readily available in case we need to have another look at it later. By these means, possibly costly reevaluations are avoided, trading memory for execution speed:

> fibs;
0L:1L:1L:2L:3L:5L:8L:13L:21L:34L:55L:#<thunk 0xb5d54590>

Let's take a look at some of the other convenience operations for generating stream values. The prelude defines infinite arithmetic sequences, using inf or -inf to denote an upper (or lower) infinite bound for the sequence, e.g.:

> let u = 1..inf; let v = -1.0:-1.2..-inf;
> u!!(0..10); v!!(0..10);

Other useful stream generator functions are iterate, which keeps applying the same function over and over again, repeat, which just repeats its argument forever, and cycle, which cycles through the elements of the given list:

> iterate (*2) 1!!(0..10);
> repeat 1!!(0..10);
> cycle [0,1]!!(0..10);

Moreover, list comprehensions can draw values from streams and return the appropriate stream result:

> let rats = [m,n-m | n=2..inf; m=1..n-1; gcd m (n-m) == 1]; rats;
(1,1):#<thunk 0xb5d54950>
> rats!!(0..10);

Finally, let's rewrite our prime sieve so that it generates the infinite stream of all prime numbers:

all_primes      = sieve (2..inf) with
  sieve (p:qs)  = p : sieve [q | q = qs; q mod p] &;

Note that we can omit the empty list case of sieve here, since the sieve now never becomes empty. Example:

> let P = all_primes;
> P!!(0..20);
> P!299;

You can also just print the entire stream. This will run forever, so hit Ctrl-C when you get bored:

> using system;
> do (printf "%d\n") all_primes;

(Make sure that you really use the all_primes function instead of the P variable to print the stream. Otherwise, because of memoization the stream stored in P will grow with the number of elements printed until memory is exhausted. Calling do on a fresh instance of the stream of primes allows do to get rid of each "cons" cell after having printed the corresponding stream element.)

5.3   Matrix Computations

Pure offers a number of basic matrix operations, such as matrix construction, indexing, slicing, as well as getting the size and dimensions of a matrix (these are briefly described in the Standard Library section below). However, it does not supply built-in support for matrix arithmetic and other linear algebra algorithms. The idea is that these can and should be provided through separate libraries (please check the Pure website for the pure-gsl module which is an ongoing project to provide a full GSL interface for the Pure language).

But Pure's facilities for matrix and list processing also make it easy to roll your own, if desired. First, the prelude provides matrix versions of the common list operations like map, fold, zip etc., which provide a way to implement common matrix operations. E.g., multiplying a matrix x with a scalar a amounts to mapping the function (a*) to x, which can be done as follows:

> a * x::matrix = map (a*) x if ~matrixp a;
> 2*{1,2,3;4,5,6};

Likewise, matrix addition and other element-wise operations can be realized using zipwith, which combines corresponding elements of two matrices using a given binary function:

> x::matrix + y::matrix = zipwith (+) x y;
> {1,2,3;4,5,6}+{1,2,1;3,2,3};

Second, matrix comprehensions make it easy to express a variety of algorithms which would typically be implemented using for loops in conventional programming languages. To illustrate the use of matrix comprehensions, here is how we can define an operation to create a square identity matrix of a given dimension:

> eye n = {i==j | i = 1..n; j = 1..n};
> eye 3;

Note that the i==j term is just a Pure idiom for the Kronecker symbol. Another point worth mentioning here is that the generator clauses of matrix comprehensions alternate between row and column generation automatically, if values are drawn from lists as in the example above. (More precisely, the last generator, which varies most quickly, yields a row, the next-to-last one a column of these row vectors, and so on.) This makes matrix comprehensions resemble customary mathematical notation very closely.

Of course, matrix comprehensions can also draw values from other matrices instead of lists. In this case the block layout of the component matrices is preserved. For instance:

> {x,y|x={1,2};y={a,b;c,d}};

Note that a matrix comprehension involving filters may fail because the filtered result isn't a rectangular matrix any more. E.g., {2*x|x={1,2,3,-4};x>0} works, as does {2*x|x={-1,2;3,-4};x>0}, but {2*x|x={1,2;3,-4};x>0} doesn't because the rows of the result matrix have different lengths.

As a slightly more comprehensive example (no pun intended!), here is a definition of matrix multiplication in Pure. The building block here is the "dot" product of two vectors which can be defined as follows:

> sum = foldl (+) 0;
> dot x::matrix y::matrix = sum $ zipwith (*) (rowvector x) (rowvector y);
> dot {1,2,3} {1,0,1};

The general matrix product now boils down to a simple matrix comprehension which just computes the dot product of all rows of x with all columns of y (the rows and cols functions are prelude operations found in matrices.pure):

> x::matrix * y::matrix = {dot u v | u = rows x; v = cols y};
> {0,1;1,0;1,1}*{1,2,3;4,5,6};

(For the sake of simplicity, this doesn't do much error checking. In production code you'd check at least the conformance of matrix dimensions, of course.)

Well, that was easy. So let's take a look at a more challenging example, Gaussian elimination, which can be used to solve systems of linear equations. The algorithm brings a matrix into "row echelon" form, a generalization of triangular matrices. The resulting system can then be solved quite easily using back substitution.

Here is a Pure implementation of the algorithm. Note that the real meat is in the pivoting and elimination step (step function) which is iterated over all columns of the input matrix. In each step, x is the current matrix, i the current row index, j the current column index, and p keeps track of the current permutation of the row indices performed during pivoting. The algorithm returns the updated matrix x, row index i and row permutation p.

gauss_elimination x::matrix = p,x
when n,m = dim x; p,_,x = foldl step (0..n-1,0,x) (0..m-1) end;

// One pivoting and elimination step in column j of the matrix:
step (p,i,x) j
= if max_x==0 then p,i,x
    // updated row permutation and index:
    transp i max_i p, i+1,
    {// the top rows of the matrix remain unchanged:
     // the pivot row, divided by the pivot element:
     {x!(i,l)/x!(i,j)                 | l=0..m-1};
     // subtract suitable multiples of the pivot row:
     {x!(k,l)-x!(k,j)*x!(i,l)/x!(i,j) | k=i+1..n-1; l=0..m-1}}
  n,m = dim x; max_i, max_x = pivot i (col x j);
  x = if max_x>0 then swap x i max_i else x;
end with
  pivot i x       = foldl max (0,0) [j,abs (x!j)|j=i..#x-1];
  max (i,x) (j,y) = if x<y then j,y else i,x;

Please refer to any good textbook on numerical mathematics for a closer description of the algorithm. But here is a brief rundown of what happens in each elimination step: First we find the pivot element in column j of the matrix. (We're doing partial pivoting here, i.e., we only look for the element with the largest absolute value in column j, starting at row i. That's usually good enough to achieve numerical stability.) If the pivot is zero then we're done (the rest of the pivot column is already zeroed out). Otherwise, we bring it into the pivot position (swapping row i and the pivot row), divide the pivot row by the pivot, and subtract suitable multiples of the pivot row to eliminate the elements of the pivot column in all subsequent rows. Finally we update i and p accordingly and return the result.

In order to complete the implementation, we still need the following little helper functions to swap two rows of a matrix (this is used in the pivoting step) and to apply a transposition to a permutation (represented as a list):

swap x i j = x!!(transp i j (0..n-1),0..m-1) when n,m = dim x end;
transp i j p = [p!tr k | k=0..#p-1]
with tr k = if k==i then j else if k==j then i else k end;

Finally, let us define a convenient print representation of double matrices a la Octave (the meaning of the __show__ function is explained in the Caveats and Notes section):

using system;
__show__ x::matrix
= strcat [printd j (x!(i,j))|i=0..n-1; j=0..m-1] + "\n"
with printd 0 = sprintf "\n%10.5f"; printd _ = sprintf "%10.5f" end
when n,m = dim x end if dmatrixp x;


> let x = dmatrix {2,1,-1,8; -3,-1,2,-11; -2,1,2,-3};
> x; gauss_elimination x;
   2.00000   1.00000  -1.00000   8.00000
  -3.00000  -1.00000   2.00000 -11.00000
  -2.00000   1.00000   2.00000  -3.00000
   1.00000   0.33333  -0.66667   3.66667
   0.00000   1.00000   0.40000   2.60000
   0.00000   0.00000   1.00000  -1.00000

5.4   Symbolic Matrices

As already mentioned, matrices may contain not just numbers but any kind of Pure value, in which case they become symbolic matrices. Symbolic matrices are a convenient data structure for storing arbitrary collections of values which provides fast random access to its members. In particular, symbolic matrices can also be nested, and thus arrays of arbitrary dimension can be realized as nested symbolic vectors. However, you have to be careful when constructing such values, as the {...} construct normally combines submatrices to larger matrices. For instance:

> {{1,2},{3,4}};

As already mentioned, one way to inhibit this "splicing" of the submatrices in a larger matrix is to use the "quote" operator (cf. Special Forms):

> '{{1,2},{3,4}};

(Note that this result is really different from {1,2;3,4}. The latter is a 2x2 integer matrix, while the former is a symbolic vector a.k.a. 1x2 matrix whose elements happen to be two integer vectors.)

Unfortunately, the quote operator in fact inhibits evaluation of all embedded subterms which may be undesirable if the matrix expression contains arithmetic (as in '{{1+1,2*3}}), so this method works best for constant matrices. A more general way to create a symbolic vector of matrices is provided by the vector function from the prelude, which is applied to a list of the vector elements as follows:

> vector [{1,2},{3,4}];

Calls to the vector function can be nested to an arbitrary depth to obtain higher-dimensional "arrays":

> vector [vector [{1,2}],vector [{3,4}]];

This obviously becomes a bit unwieldy for higher dimensions, but in Pure you can easily define yourself some more convenient notation if you like. For instance, the following macro may be used to define a pair of "non-splicing" vector brackets:

> outfix {: :};
> def {: xs@(_,_) :} = vector (list xs);
> def {: x :} = vector [x];
> {:{:{1,2}:},{:{3,4}:}:};

(Both macros and outfix symbol declarations are described later in the appropriate sections, see Macros and Symbol Declarations.)

5.5   Record Data

Symbolic matrices also provide a means to represent simple record-like data, by encoding records as symbolic vectors consisting of "hash pairs" of the form key => value. This kind of data structure is very convenient to represent aggregates with lots of different components. Since the components of records can be accessed by indexing with key values, you don't have to remember which components are stored in which order, just knowing the keys of the required members is enough. In contrast, tuples, lists and other kinds of constructor terms quickly become unwieldy for such purposes.

The keys used for indexing the record data must be either symbols or strings, while the corresponding values may be arbitrary Pure values. The prelude provides some operations on these special kinds of matrices, which let you retrieve vector elements by indexing and perform non-destructive updates, see the Record Functions section in the Pure Library Manual for details. Here are a few examples which illustrate how to create records and work with them:

> let r = {x=>5, y=>12};
> recordp r, member r x;
> r!y; r!![y,x];
> insert r (x=>99);
> insert ans (z=>77);
> delete ans z;

Note the use of the "hash rocket" => which denotes the key=>value associations in a record. The hash rocket is a constructor declared as an infix operator in the prelude, see the Prelude section in the Pure Library Manual. There's one caveat here, however. Since neither => nor ! treat their key operand in a special way, symbols used as keys must not be bound to a value, or you'll have to protect them from being evaluated by quoting them. Thus, to be on the safe side, you should actually write:

> let r = {'x=>5, 'y=>12};
> r!'y; r!!['y,'x];

It's also possible to use strings as keys instead, which may actually be more convenient in some cases:

> let r = {"x"=>5, "y"=>12};
> keys r; vals r;
> update r "y" (r!"y"+1);

You can also mix strings and symbols as keys in the same record (but note that strings and symbols are always distinct, so y and "y" are really two different keys here):

> insert r (y=>99);

As records are in fact just special kinds of matrices, the standard matrix operations can be used on record values as well. For instance, the matrix constructor provides an alternative way to quickly augment a record with a collection of new key=>value associations:

> let r = {x=>5, y=>12};
> let r = {r, x=>7, z=>3}; r;
> r!x, r!z;
> delete r x;
> ans!x;

As the example shows, this may produce duplicate keys, but these are handled gracefully; indexing and updates will always work with the last association for a given key in the record. If necessary, you can remove duplicate entries from a record as follows; this will only keep the last association for each key:

> record r;

In fact, the record operation not only removes duplicates, but also orders the record entries by keys. This produces a kind of normalized representation which is useful if you want to compare or combine two record values irrespective of the ordering of the fields. For instance:

> record {x=>5, y=>12} === record {y=>12, x=>5};

The record function can also be used to construct a normalized record directly from a list or tuple of hash pairs:

> record [x=>5, x=>7, y=>12];

Other matrix operations such as map, foldl, etc., and matrix comprehensions can be applied to records just as easily. This enables you to perform bulk updates of record data in a straightforward way. For instance, here's how you can define a function maprec which applies a function to all values stored in a record:

> maprec f = map (\(u=>v) -> u=>f v);
> maprec (*2) {x=>5,y=>12};

Another example: The following ziprec function collects pairs of values stored under common keys in two records (we also normalize the result here so that duplicate keys are always removed):

> ziprec x y = record {u=>(x!u,y!u) | u = keys x; member y u};
> ziprec {a=>3,x=>5,y=>12} {x=>10,y=>24,z=>7};

Thus the full power of generic matrix operations is available for records, which turns them into a very versatile data structure, much more powerful than records in conventional programming languages which are usually limited to constructing records and accessing or modifying their components. Note that since the values stored in records can be arbitrary Pure values, you can also have mutable records by making use of Pure's expression references (see Expression References in the library manual). And of course records can be nested, too:

> let r = {a => {b=>1,c=>2}, b => 2};
> r!a, r!b, r!a!b;

6   Macros

Macros are a special type of functions to be executed as a kind of "preprocessing stage" at compile time. In Pure these are typically used to define custom special forms and to perform inlining of function calls and other simple kinds of source-level optimizations.

Whereas the macro facilities of most programming languages simply provide a kind of textual substitution mechanism, Pure macros operate on symbolic expressions and are implemented by the same kind of rewriting rules that are also used to define ordinary functions in Pure. In difference to these, macro rules start out with the keyword def, and only simple kinds of rules without any guards or multiple left-hand and right-hand sides are permitted.

Syntactically, a macro definition looks just like a variable or constant definition, using def in lieu of let or const, but they are processed in a different way. Macros are substituted into the right-hand sides of function, constant and variable definitions. All macro substitution happens before constant substitutions and the actual compilation step. Macros can be defined in terms of other macros (also recursively), and are evaluated using call by value (i.e., macro calls in macro arguments are expanded before the macro gets applied to its parameters).

6.1   Optimization Rules

Here is a simple example, showing a rule which expands saturated calls of the succ function (defined in the prelude) at compile time:

> def succ x = x+1;
> foo x::int = succ (succ x);
> show foo
foo x::int = x+1+1;

Rules like these can be useful to help the compiler generate better code. Note that a macro may have the same name as an ordinary Pure function, which is essential if you want to optimize calls to an existing function, as in the previous example. (Just like ordinary functions, the number of parameters in each rule for a given macro must be the same, but a macro may have a different number of arguments than the corresponding function.)

A somewhat more practical example is the following rule from the prelude, which eliminates saturated instances of the right-associative function application operator:

def f $ x = f x;

Like in Haskell, this low-priority operator is handy to write cascading function calls. With the above macro rule, these will be "inlined" as ordinary function applications automatically. Example:

> foo x = bar $ bar $ 2*x;
> show foo
foo x = bar (bar (2*x));

Here are two slightly more tricky rules from the prelude, which optimize the case of "throwaway" list comprehensions. This is useful if a list comprehension is evaluated solely for its side effects:

def void (catmap f x) = do f x;
def void (listmap f x) = do f x;

Note that the void function simply throws away its argument and returns () instead. The do function applies a function to every member of a list (like map), but throws away all intermediate results and just returns (), which is much more efficient if you don't need those results anyway. These are both defined in the prelude.

Before we delve into this example, a few remarks are in order about the way list comprehensions are implemented in Pure. As already mentioned, list comprehensions are just syntactic sugar; the compiler immediately transforms them to an equivalent expression involving only lambdas and a few other list operations. Note that list comprehensions are essentially equivalent to piles of nested lambdas, filters and maps, but for various reasons they are actually implemented using two special helper operations, catmap and listmap.

The catmap operation combines map and cat; this is needed, in particular, to accumulate the results of nested generators, such as [i,j | i = 1..n; j = 1..m]. The same operation is also used to implement filter clauses, you can see this below in the examples. However, for efficiency simple generators like [2*i | i = 1..n] are translated to a listmap instead (which is basically just map, but works with different aggregate types, so that list comprehensions can draw values from aggregates other than lists, such as matrices).

Now let's see how the rules above transform a list comprehension if we "voidify" it:

> using system;
> f = [printf "%g\n" (2^x+1) | x=1..5; x mod 2];
> g = void [printf "%g\n" (2^x+1) | x=1..5; x mod 2];
> show f g
f = catmap (\x -> if x mod 2 then [printf "%g\n" (2^x+1)] else []) (1..5);
g = do (\x -> if x mod 2 then [printf "%g\n" (2^x+1)] else []) (1..5);

Ok, so the catmap got replaced with a do which is just what we need to make this code go essentially as fast as a for loop in conventional programming languages (up to constant factors, of course). Here's how it looks like when we run the g function:

> g;

It's not all roses, however, since the above macro rules will only get rid of the outermost catmap if the list comprehension binds multiple variables:

> u = void [puts $ str (x,y) | x=1..2; y=1..3];
> show u
u = do (\x -> listmap (\y -> puts (str (x,y))) (1..3)) (1..2);

If you're bothered by this, you'll have to apply void recursively, creating a nested list comprehension which expands to a nested do:

> v = void [void [puts $ str (x,y) | y=1..3] | x=1..2];
> show v
v = do (\x -> do (\y -> puts (str (x,y))) (1..3)) (1..2);

(It would be nice to have this handled automatically, but the left-hand side of a macro definition must be a simple expression, and thus it's not possible to write a macro which descends recursively into the lambda argument of catmap.)

6.2   Recursive Macros

Macros can also be recursive, in which case they usually consist of multiple rules and make use of pattern-matching like ordinary function definitions. As a simple example, let's implement a Pure version of Lisp's quasiquote which allows you to create a quoted expression from a "template" while substituting variable parts of the template. (For the sake of brevity, our definition is somewhat simplified and does not cover some corner cases. See the Pure distribution for a full version of this example.)

def quasiquote (unquote x)      = x;
def quasiquote (f@_ (splice x)) = foldl ($) (quasiquote f) x;
def quasiquote (f@_ x)          = quasiquote f (quasiquote x);
def quasiquote x                = quote x;

(Note the f@_, which is an anonymous "as" pattern forcing the compiler to recognize f as a function variable, rather than a literal function symbol. See Head = Function in the Caveats and Notes section for an explanation of this trick.)

The first rule above takes care of "unquoting" embedded subterms. The second rule "splices" an argument list into an enclosing function application. The third rule recurses into subterms of a function application, and the fourth and last rule takes care of quoting the "atomic" subterms. Note that unquote and splice themselves are just passive constructor symbols, the real work is done by quasiquote, using foldl at runtime to actually perform the splicing. (Putting off the splicing until runtime makes it possible to splice argument lists computed at runtime.)

If we want, we can also add some syntactic sugar for Lisp weenies. (Note that we cannot have ',' for unquoting, so we use ',$' instead.)

prefix 9 ` ,$ ,@ ;
def `x = quasiquote x; def ,$x = unquote x; def ,@x = splice x;


> `(2*42+2^12);
> `(2*42+,$(2^12));
> `foo 1 2 (,@'[2/3,3/4]) (5/6);
foo 1 2 (2/3) (3/4) (5/6)
> `foo 1 2 (,@'args) (5/6) when args = '[2/3,3/4] end;
foo 1 2 (2/3) (3/4) (5/6)

We mention in passing here that, technically, Pure macros are just as powerful as (unconditional) term rewriting systems and thus they are Turing-complete. This implies that a badly written macro may well send the Pure compiler into an infinite recursion, which results in a stack overflow at compile time. See the Caveats and Notes section for information on how to deal with these by setting the PURE_STACK environment variable.

6.3   User-Defined Special Forms

The quasiquote macro in the preceding subsection also provides an example of how you can use macros to define your own special forms. This works because the actual evaluation of macro arguments is put off until runtime, and thus we can safely pass them to built-in special forms and other constructs which defer their evaluation at runtime. In fact, the right-hand side of a macro rule may be an arbitrary Pure expression involving conditional expressions, lambdas, binding clauses, etc. These are never evaluated during macro substitution, they just become part of the macro expansion (after substituting the macro parameters).

Here is another useful example of a user-defined special form, the macro timex which employs the system function clock to report the cpu time in seconds needed to evaluate a given expression, along with the computed result:

> using system;
> def timex x = (clock-t0)/CLOCKS_PER_SEC,y when t0 = clock; y = x end;
> sum = foldl (+) 0L;
> timex $ sum (1L..100000L);

Note that the above definition of timex wouldn't work as an ordinary function definition, since by virtue of Pure's basic eager evaluation strategy the x parameter would have been evaluated already before it is passed to timex, making timex always return a zero time value. Try it!

Here's yet another example, which is handy if you need to trace function calls. (As of Pure 0.22, the interpreter now has its own built-in debugging facility, see Debugging. However, the following macro allows you to trace functions using your own custom output format, and may thus be useful in situations where the built-in debugger is not appropriate.)

using system;
def trace f x y = printf "** exit %s: %s -> %s\n" (str f,str x,str y) $$ y
when y = printf "** call %s: %s\n: " (str f,str x) $$ gets $$ y end;

This macro is invoked with the function to be traced, the arguments (or whatever you want to be printed as additional debugging information) and the actual function call as parameters. (This is a rather simplistic version, which just prints a prompt on function entry and the final reduction after the call. You can easily make this as elaborate as you like. E.g., you might want to keep track of recursive levels and profiling information, add various interactive commands to selectively enable and disable tracing during the evaluation, etc.)

We can still make this a bit more convenient by introducing the following ordinary function definition:

trace f x = trace f x (f x);

This lets us patch up a call to trace a given function, as shown below, without having to change the definition of the function at all. This trick only works with global functions; for local functions you'll have to add an explicit call of the trace macro to the local definition yourself. Also note that the definition above only works with functions taking a single parameter; see the trace.pure example in the distribution for the full version which can deal with any number of arguments.

// Uncomment this line to trace calls to the 'fact' function.
def fact n = trace fact n;
// Sample function to be traced.
fact n = if n>0 then n*fact(n-1) else 1;

Here's a trace of the fact function obtained in this fashion (hit carriage return after each ':' prompt to proceed with the computation):

> fact 2;
** call fact: 2
** call fact: 1
** call fact: 0
** exit fact: 0 -> 1
** exit fact: 1 -> 1
** exit fact: 2 -> 2

Note that by just removing the macro definition for fact above, you can make the function run untraced as usual again. This scheme is quite flexible, the only real drawback is that you have to explicitly add some code for each function you want to trace.

6.4   Macro Hygiene

Pure macros are lexically scoped, i.e., the binding of symbols in the right-hand-side of a macro definition is determined statically by the text of the definition, and macro parameter substitution also takes into account binding constructs, such as with and when clauses, in the right-hand side of the definition. Macro facilities with these pleasant properties are also known as hygienic macros. They are not susceptible to so-called "name capture," which makes macros in less sophisticated languages bug-ridden and hard to use. (This is explained in more detail in the Hygienic Macros section.)

Pure macros also have their limitations. Specifically, the left-hand side of a macro rule must be a simple expression, just like in ordinary function definitions. This restricts the kinds of expressions which can be rewritten by a macro. But Pure macros are certainly powerful enough for most common preprocessing purposes, while still being robust and easy to use.

7   Declarations

Pure is a very terse language by design. Usually you don't declare much stuff, you just define it and be done with it. However, there are a few toplevel constructs which let you declare symbols with special attributes and manage programs consisting of several source modules:

7.1   Symbol Declarations

Scope declarations take the following form:

public symbol ...;
private symbol ...;

This declares the listed symbols as public or private, respectively. Each symbol must either be an identifier or a sequence of punctuation characters. The latter kind of symbols must always be declared before use, whereas ordinary identifiers can be used without a prior declaration in which case they are declared implicitly and default to public scope, meaning that they are visible everywhere in a program. An explicit public declaration of ordinary identifiers is thus rarely needed (unless you want to declare symbols as members of a specific namespace, see Namespaces below). Symbols can also be declared private, meaning that the symbol is visible only in the namespace it belongs to. This is explained in more detail under Private Symbols in the Namespaces section below.

Note that to declare several symbols in a single declaration, you can list them all with whitespace in between. The same syntax applies to the other types of symbol declarations discussed below. (Commas are not allowed as delimiters here, as they may occur as legal symbol constituents in the list of symbols.) The public and private keywords can also be used as a prefix in any of the special symbol declarations discussed below, to specify the scope of the declared symbols (if the scope prefix is omitted, it defaults to public).

The following "fixity" declarations are available for introducing special operator and constant symbols. This changes the way that these symbols are parsed and thus provides you with a limited means to extend the Pure language at the lexical and syntactical level.

Operator declarations: infix level symbol ...;

Pure provides you with a theoretically unlimited number of different precedence levels for user-defined infix, prefix and postfix operators. Precedence levels are numbered starting at 0; larger numbers indicate higher precedence. (For practical reasons, the current implementation does require that precedence numbers can be encoded as 24 bit unsigned machine integers, giving you a range from 0 to 16777215, but this should be large enough to incur no real limitations on applications. Also, the operator declarations in the prelude have been set up to leave enough "space" between the "standard" levels so that you can easily sneak in new operator symbols at low, high or intermediate precedences.)

On each precedence level, you can declare (in order of increasing precedence) infix (binary non-associative), infixl (binary left-associative), infixr (binary right-associative), prefix (unary prefix) and postfix (unary postfix) operators. For instance, here is a typical excerpt from the prelude (the full table can be found in the Prelude section of the Pure Library Manual):

infix  1800 < > <= >= == ~= ;
infixl 2200 + - ;
infixl 2300 * / div mod ;
infixr 2500 ^ ;
prefix 2600 # ;

Instead of denoting the precedence by an explicit integer value, you can also specify an existing operator symbol enclosed in parentheses. Thus the following declaration gives the ++ operator the same precedence as +:

infixl (+) ++ ;

The given symbol may be of a different fixity than the declaration, but it must have a proper precedence level (i.e., it must be an infix, prefix or postfix symbol). E.g., the following declaration gives ^^ the same precedence level as the infix ^ symbol, but turns it into a postfix operator:

postfix (^) ^^ ;
Outfix symbol declarations: outfix left right ...;

Pure also provides unary outfix operators, which work like in Wm Leler's constraint programming language Bertrand. Outfix operators let you define your own bracket structures. The operators must be given as pairs of matching left and right symbols (which must be distinct). For instance:

outfix |: :| BEGIN END;

After this declaration you can write bracketed expressions like |:x:| or BEGIN foo, bar END. These are always at the highest precedence level (i.e., syntactically they work like parenthesized expressions). Just like other operators, you can turn outfix symbols into ordinary functions by enclosing them in parentheses, but you have to specify the symbols in matching pairs, such as (BEGIN END).

Constant symbol declarations: nonfix symbol ...;

Pure also has a notation for "nullary" operators, i.e., "operators without operands", which are used to denote special constants. These are introduced using a nonfix declaration, e.g.:

nonfix red green blue;

Syntactically, these work just like ordinary identifiers, so they may stand whereever an identifier is allowed (no parentheses are required to "escape" them). The difference to ordinary identifiers is that nonfix symbols are always interpreted as literals, even if they occur in a variable position on the left-hand side of a rule. So, with the above declaration, you can write something like:

> foo x = case x of red = green; green = blue; blue = red end;
> map foo [red,green,blue];

Thus nonfix symbols are pretty much like nullary constructor symbols in languages like Haskell. Non-fixity is just a syntactic attribute, however. Pure doesn't enforce that such values are really "constant", so you can still write a "constructor equation" like the following:

> red = blue;
> map foo [red,green,blue];

Examples for all types of symbol declarations can be found in the prelude which declares a bunch of standard (arithmetic, relational, logical) operator symbols as well as the list and pair constructors ':' and ',', and a few nonfix symbols (mostly for denoting different kinds of exceptions).

One final thing worth noting here is that unary minus plays a special role in the syntax. Like in Haskell and following mathematical tradition, unary minus is the only prefix operator symbol which is also used as an infix operator, and is always on the same precedence level as binary minus, whose precedence may be chosen freely in the prelude. (The minus operator is the only symbol which gets that special treatment; all other operators must have distinct lexical representations.) Thus, with the standard prelude, -x+y will be parsed as (-x)+y, whereas -x*y is the same as -(x*y). Also note that the notation (-) always denotes the binary minus operator; the unary minus operation can be denoted using the built-in neg function.

7.2   Modules and Imports

While Pure doesn't offer separate compilation, the using declaration provides a simple but effective way to assemble a Pure program from several source modules. It takes the following form (note that in contrast to symbol declarations, the comma is used as a delimiter symbol here):

using name, ...;

This causes each given script to be included in the Pure program at the given point (if it wasn't already included before), which makes available all the definitions of the included script in your program. Note that each included script is loaded only once, when the first using clause for the script is encountered. Nested imports are allowed, i.e., an imported module may itself import other modules, etc. A Pure program then basically is the concatenation of all the source modules given as command line arguments, with other modules listed in using clauses inserted at the corresponding source locations.

(The using clause also has an alternative form which allows dynamic libraries to be loaded, this will be discussed in the C Interface section.)

For instance, the following declaration causes the math.pure script from the standard library to be included in your program:

using math;

You can also import multiple scripts in one go:

using array, dict, set;

Moreover, Pure provides a notation for qualified module names which can be used to denote scripts located in specific package directories, e.g.:

using examples::libor::bits;

In fact this is equivalent to the following using clause which spells out the real filename of the script between double quotes (the .pure suffix can also be omitted in which case it is added automatically):

using "examples/libor/bits.pure";

Both notations can be used interchangeably; the former is usually more convenient, but the latter allows you to denote scripts whose names aren't valid Pure identifiers.

Script identifiers are translated to the corresponding filenames by replacing the '::' symbol with the pathname separator '/' and tacking on the '.pure' suffix. The following table illustrates this with a few examples.

Script identifier Filename
math "math.pure"
examples::libor::bits "examples/libor/bits.pure"
::pure::examples::hello "/pure/examples/hello.pure"

Note the last example, which shows how an absolute pathname can be denoted using a qualifier starting with '::'.

Unless an absolute pathname is given, the interpreter performs a search to locate the script. The search algorithm considers the following directories in the given order:

  • the directory of the current script, which is the directory of the script containing the using clause, or the current working directory if the clause was read from standard input (as is the case, e.g., in an interactive session);
  • the directories named in -I options on the command line (in the given order);
  • the colon-separated list of directories in the PURE_INCLUDE environment variable (in the given order);
  • finally the directory named by the PURELIB environment variable.

Note that the current working directory is not searched by default (unless the using clause is read from standard input), but of course you can force this by adding the option -I. to the command line, or by including '.' in the PURE_INCLUDE variable.

The directory of the current script (the first item above) can be skipped by specifying the script to be loaded as a filename in double quotes, prefixed with the special sys: tag. The search then starts with the "system" directories (-I, PURE_INCLUDE and PURELIB) instead. This is useful, e.g., if you want to provide your own custom version of a standard library script which in turn imports that library script. For instance, a custom version of math.pure might employ the following using clause to load the math.pure script from the Pure library:

using "sys:math";
// custom definitions go here
log2 x = ln x/ln 2;

The interpreter compares script names (to determine whether two scripts are actually the same) by using the canonicalized full pathname of the script, following symbolic links to the destination file (albeit only one level). Thus different scripts with the same basename, such as foo/utils.pure and bar/utils.pure can both be included in the same program (unless they link to the same file).

More precisely, canonicalizing a pathname involves the following steps:

  • relative pathnames are expanded to absolute ones, using the search rules discussed above;
  • the directory part of the pathname is normalized to the form returned by the getcwd system call;
  • the ".pure" suffix is added if needed;
  • if the resulting script name is actually a symbolic link, the interpreter follows that link to its destination, albeit only one level. (This is only done on Unix-like systems.)

The directory of the canonicalized pathname is also used when searching other scripts included in a script. This makes it possible to have an executable script with a shebang line in its own directory, which is then executed via a symbolic link placed on the system PATH. In this case the script search performed in using clauses will use the real script directory and thus other required scripts can be located there. This is the recommended practice for installing standalone Pure applications in source form which are to be run directly from the shell.

7.3   Namespaces

To facilitate modular development, Pure also provides namespaces as a means to avoid name clashes between symbols, and to keep the global namespace tidy and clean. Namespaces serve as containers holding groups of related identifiers and other symbols. Inside each namespace, symbols must be unique, but the same symbol may be used to denote different objects (variables, functions, etc.) in different namespaces. (Pure's namespace system was heavily inspired by C++ and works in a very similar fashion. So if you know C++ you should feel right at home and skimming this section to pick up Pure's syntax of the namespace constructs should be enough to start using it.)

The global namespace is always available. By default, new symbols are created in this namespace, which is also called the default namespace. Additional namespaces can be created with the namespace declaration, which also switches to the given namespace (makes it the current namespace), so that new symbols are then created in that namespace rather than the default one. The current namespace also applies to all kinds of symbol declarations, including operator and constant symbol declarations, as well as extern declarations (the latter are described in the C Interface section).

The basic form of the namespace declaration has the following syntax (there's also a "scoped" form of the namespace declaration which will be discussed in Scoped Namespaces at the end of this section):

namespace name;
// declarations and definitions in namespace 'name'

The second form switches back to the default namespace. For instance, in order to define two symbols with the same print name foo in two different namespaces foo and bar, you can write:

namespace foo;
foo x = x+1;
namespace bar;
foo x = x-1;

We can now refer to the symbols we just defined using qualified symbols of the form namespace::symbol:

> foo::foo 99;
> bar::foo 99;

This avoids any potential name clashes, since the qualified identifier notation always makes it clear which namespace the given identifier belongs to.

A namespace can be "reopened" at any time to add new symbols and definitions to it. This allows namespaces to be created that span several source modules. You can also create several different namespaces in the same module.

Similar to the using declaration, a namespace declaration accepts either identifiers or double-quoted strings as namespace names. E.g., the following two declarations are equivalent:

namespace foo;
namespace "foo";

The latter form also allows more descriptive labels which aren't identifiers, e.g.:

namespace "Private stuff, keep out!";

Note that the namespace prefix in a qualified identifier must be a legal identifier, so it isn't possible to access symbols in namespaces with such descriptive labels in a direct fashion. The only way to get at the symbols in this case is to use a namespace or using namespace declaration (for the latter see Using Namespaces below).

7.3.1   Using Namespaces

Since it is rather inconvenient if you always have to write identifiers in their qualified form outside of their "home" namespace, Pure allows you to specify a list of search namespaces which are used to look up symbols not in the default or the current namespace. This is done with the using namespace declaration, which takes the following form:

using namespace name1, name2, ...;
// ...
using namespace;

(As with namespace declarations, the second form without any namespace arguments gets you back to the default empty list of search namespaces.)

For instance, consider this example:

namespace foo;
foo x = x+1;
namespace bar;
foo x = x-1;
bar x = x+1;

The symbols in these namespaces can be accessed unqualified as follows:

> using namespace foo;
> foo 99;
> using namespace bar;
> foo 99;
> bar 99;

This method is often to be preferred over opening a namespace with the namespace declaration, since using namespace only gives you "read access" to the imported symbols, so you can't accidentally mess up the definitions of the namespace you're using. Another advantage is that the using namespace declaration also lets you search multiple namespaces at once:

using namespace foo, bar;

Be warned, however, that this brings up the very same issue of name clashes again:

> using namespace foo, bar;
> foo 99;
<stdin>, line 15: symbol 'foo' is ambiguous here

In such a case you'll have to resort to using namespace qualifiers again, in order to resolve the name clash:

> foo::foo 99;

To avoid this kind of mishap, you can also selectively import just a few symbols from a namespace instead. This can be done with a declaration of the following form:

using namespace name1 ( sym1 sym2 ... ), name2 ... ;

As indicated, the symbols to be imported can optionally be placed as a whitespace-delimited list inside parentheses, following the corresponding namespace name. For instance:

> using namespace foo, bar (bar);
> foo 99;
> bar 99;
> bar::foo 99;

Note that now we have no clash on the foo symbol any more, because we restricted the import from the bar namespace to the bar symbol, so that bar::foo has to be denoted with a qualified symbol now.

7.3.2   Symbol Lookup and Creation

Pure's rules for looking up and creating symbols are fairly straightforward and akin to those in other languages featuring namespaces. However, there are some intricacies involved, because the rewriting rule format of definitions allows "referential" use of symbols not only in the "body" (right-hand side) of a definition, but also in the left-hand side patterns. We discuss this in detail below.

The compiler searches for symbols first in the current namespace (if any), then in the currently active search namespaces (if any), and finally in the default (i.e., the global) namespace, in that order. This automatic lookup can be bypassed by using an absolute namespace qualifier of the form ::foo::bar. In particular, ::bar always denotes the symbol bar in the default namespace, while ::foo::bar denotes the symbol bar in the foo namespace. (Normally, the latter kind of notation is only needed if you have to deal with nested namespaces, see Hierarchical Namespaces below.)

If no existing symbol is found, a new symbol is created automatically, by implicitly declaring a public symbol with default attributes. New unqualified symbols are always created in the current namespace, while new qualified symbols are created in the namespace given by the namespace prefix of the symbol. However, note that in the latter case the compiler always checks that the given namespace prefix matches the current namespace:

> namespace foo;
> namespace;
> foo::bar x = 1/x;
<stdin>, line 3: undeclared symbol 'foo::bar'

Thus it's only possible to introduce a new symbol in a given namespace if that namespace is the current one. These error messages are somewhat annoying, but they provide at least some protection against typos and other silly mistakes and prevent you from accidentally clobbering the contents of other namespaces. To make these errors go away it's enough to just declare the symbols in their proper namespaces.

New symbols are also created if a global unqualified (and yet undeclared) symbol is being "defined" in a rewriting rule or let/const definition, even if a symbol with the same print name from another namespace is already visible in the current scope. To distinguish "defining" from "referring" uses of a global symbol, Pure uses the following (purely syntactic) notions:

  • A defining occurrence of a global function or macro symbol is any occurrence of the symbol as the head symbol on the left-hand side of a rewriting rule.
  • A defining occurrence of a global variable or constant symbol is any occurrence of the symbol in a variable position (as given by the "head = function" rule, cf. Parameters in Equations) on the left-hand side of a let or const definition.
  • All other occurrences of global symbols on the left-hand side, as well as all symbol occurrences on the right-hand side of a definition are referring occurrences.

The following example illustrates these notions:

namespace foo;
bar (bar x) = bar x;
let x,y = 1,2;

Here, the first occurrence of bar on the left-hand side bar (bar x) of the first rule is a defining occurrence, as are the occurrences of x and y on the left-hand side of the let definition. Hence these symbols are created as new symbols in the namespace foo. On the other hand, the other occurrences of bar in the first rule, as well as the ',' symbol on the left-hand side of the let definition are referring occurrences. In the former case, bar refers to the bar symbol defined by the rule, while in the latter case the ',' operator is actually declared in the prelude and thus imported from the global namespace.

As an additional safety measure against missing or mistyped symbols, the interpreter provides the option -w (see Options) to check your scripts for non-defining uses of undeclared unqualified function symbols. For instance:

$ pure -w
> puts "bla"; // missing import of system module
<stdin>, line 1: warning: implicit declaration of 'puts'
puts "bla"

For legitimate uses (such as forward uses of a symbol which is defined later), you can make these warnings go away by declaring the symbol before using it.

Note that special operator (and nonfix) symbols always require an explicit declaration. This works as already discussed in the Symbol Declarations section, except that you first switch to the appropriate namespace before declaring the symbols. For instance, here is how you can create a new + operation which multiplies its operands rather than adding them:

> namespace my;
> infixl 2200 +;
> x+y = x*y;
> 5+7;

Note that the new + operation really belongs to the namespace we created. The + operation in the default namespace works as before, and in fact you can use qualified symbols to pick the version that you need:

> namespace;
> 5+7;
> 5 ::+ 7;
> 5 my::+ 7;

Here's what you get if you happen to forget the declaration of the + operator:

> namespace my;
> x+y = x*y;
<stdin>, line 2: infixl symbol '+' was not declared in this namespace

Thus the compiler will never create a new instance of an operator symbol on the fly, an explicit declaration is always needed in such cases.

Note that if you really wanted to redefine the global + operator, you can do this even while the my namespace is current. You just have to use a qualified identifier in this case, as follows:

> namespace my;
> x ::+ y = x*y;
> a+b;

This should rarely be necessary (in the above example you might just as well enter this rule while in the global namespace), but it can be useful in some circumstances. Specifically, you might want to "overload" a global function or operator with a definition that makes use of private symbols of a namespace (which are only visible inside that namespace; see Private Symbols below). For instance:

> namespace my;
> private bar;
> bar x y = x*y;
> x ::+ y = bar x y;
> a+b;

(The above is a rather contrived example, since the very same functionality can be accomplished much easier, but there are some situations where this approach is necessary.)

7.3.3   Private Symbols

Pure also allows you to have private symbols, as a means to hide away internal operations which shouldn't be accessed directly outside the namespace in which they are declared. The scope of a private symbol is confined to its namespace, i.e., the symbol is only visible when its "home" namespace is current. Symbols are declared private by using the private keyword in the symbol declaration:

> namespace secret;
> private baz;
> // 'baz' is a private symbol in namespace 'secret' here
> baz x = 2*x;
> // you can use 'baz' just like any other symbol here
> baz 99;
> namespace;

Note that, at this point, secret::baz is now invisible, even if you have secret in the search namespace list:

> using namespace secret;
> // this actually creates a 'baz' symbol in the default namespace:
> baz 99;
baz 99
> secret::baz 99;
<stdin>, line 27: symbol 'secret::baz' is private here

The only way to bring the symbol back into scope is to make the secret namespace current again:

> namespace secret;
> baz 99;
> secret::baz 99;

7.3.4   Hierarchical Namespaces

Namespace identifiers can themselves be qualified identifiers in Pure, which enables you to introduce a hierarchy of namespaces. This is useful, e.g., to group related namespaces together under a common "umbrella" namespace:

namespace my;
namespace my::old;
foo x = x+1;
namespace my::new;
foo x = x-1;

Note that the namespace my, which serves as the parent namespace, must be created before the my::old and my::new namespaces, even if it does not contain any symbols of its own. After these declarations, the my::old and my::new namespaces are part of the my namespace and will be considered in name lookup accordingly, so that you can write:

> using namespace my;
> old::foo 99;
> new::foo 99;

This works pretty much like a hierarchy of directories and files, where the namespaces play the role of the directories (with the default namespace as the root directory), the symbols in each namespace correspond to the files in a directory, and the using namespace declaration functions similar to the shell's PATH variable.

Sometimes it is necessary to tell the compiler to use a symbol in a specific namespace, bypassing the usual symbol lookup mechanism. For instance, suppose that we introduce another global old namespace and define yet another version of foo in that namespace:

namespace old;
foo x = 2*x;

Now, if we want to access that function, with my still active as the search namespace, we cannot simply refer to the new function as old::foo, since this name will resolve to my::old::foo instead. As a remedy, the compiler accepts an absolute qualified identifier of the form ::old::foo. This bypasses name lookup and thus always yields exactly the symbol in the given namespace (if it exists; as mentioned previously, the compiler will complain about an undeclared symbol otherwise):

> old::foo 99;
> ::old::foo 99;

Also note that, as a special case of the absolute qualifier notation, ::foo always denotes the symbol foo in the default namespace.

7.3.5   Scoped Namespaces

Pure also provides an alternative scoped namespace construct which makes nested namespace definitions more convenient. This construct takes the following form:

namespace name with ... end;

The part between with and end may contain arbitrary declarations and definitions, using the same syntax as the toplevel. These are processed in the context of the given namespace, as if you had written:

namespace name;

However, the scoped namespace construct always returns you to the namespace which was active before, and thus these declarations may be nested:

namespace foo with
  // declarations and definitions in namespace foo
  namespace bar with
    // declarations and definitions in namespace bar
  // more declarations and definitions in namespace foo

Note that this kind of nesting does not necessarily imply a namespace hierarchy as discussed in Hierarchical Namespaces. However, you can achieve this by using the appropriate qualified namespace names:

namespace foo with
  // ...
  namespace foo::bar with
    // ...
  // ...

Another special feature of the scoped namespace construct is that using namespace declarations are always local to the current namespace scope (and other nested namespace scopes inside it). Thus the previous setting is restored at the end of each scope:

using namespace foo;
namespace foo with
  // still using namespace foo here
  using namespace bar;
  // now using namespace bar
  namespace bar with
    // still using namespace bar here
    using namespace foo;
    // now using namespace foo
  // back to using namespace bar
// back to using namespace foo at toplevel

Finally, here's a more concrete example which shows how scoped namespaces might be used to declare two namespaces and populate them with various functions and operators:

namespace foo with
  infixr (::^) ^;
  foo x = x+1;
  bar x = x-1;
  x^y = 2*x+y;

namespace bar with
  outfix <: :>;
  foo x = x+2;
  bar x = x-2;

using namespace foo(^ foo), bar(bar <: :>);

// namespace foo
foo x;

// namespace bar
bar x;
<: x,y :>;

Pure's namespaces can thus be used pretty much like "modules" or "packages" in languages like Ada or Modula-2. They provide a structured way to describe program components offering collections of related data and operations, which can be brought into scope in a controlled way by making judicious use of using namespace declarations. They also provide an abstraction barrier, since internal operations and data structures can be hidden away employing private symbols.

Please note that these facilities are not Pure's main focus and thus they are somewhat limited compared to programming languages specifically designed for big projects and large teams of developers. Nevertheless they should be useful if your programs grow beyond a small collection of simple source modules, and enable you to manage most Pure projects with ease.

8   Exception Handling

Pure also offers a useful exception handling facility. To raise an exception, you just invoke the built-in function throw with the value to be thrown as the argument. To catch an exception, you use the built-in special form catch with the exception handler (a function to be applied to the exception value) as the first and the expression to be evaluated as the second (call-by-name) argument. For instance:

> catch error (throw hello_world);
error hello_world

Exceptions are also generated by the runtime system if the program runs out of stack space, when a guard does not evaluate to a truth value, and when the subject term fails to match the pattern in a pattern-matching lambda abstraction, or a let, case or when construct. These types of exceptions are reported using the symbols stack_fault, failed_cond and failed_match, respectively, which are declared as constant symbols in the standard prelude. You can use catch to handle these kinds of exceptions just like any other. For instance:

> fact n = if n>0 then n*fact(n-1) else 1;
> catch error (fact foo);
error failed_cond
> catch error (fact 100000);
error stack_fault

(You'll only get the latter kind of exception if the interpreter does stack checks, see the discussion of the PURE_STACK environment variable in the Caveats and Notes section.)

Note that unhandled exceptions are reported by the interpreter with a corresponding error message:

> fact foo;
<stdin>, line 2: unhandled exception 'failed_cond' while evaluating 'fact foo'

Exceptions also provide a way to handle asynchronous signals. Pure's system module provides symbolic constants for common POSIX signals and also defines the operation trap which lets you rebind any signal to a signal exception. For instance, the following lets you handle the SIGQUIT signal:

> using system;

You can also use trap to just ignore a signal or revert to the system's default handler (which might take different actions depending on the type of signal, see signal(7) for details):

> trap SIG_IGN SIGQUIT; // signal is ignored
> trap SIG_DFL SIGQUIT; // reinstalls the default signal handler

Note that when the interpreter runs interactively, for convenience most standard termination signals (SIGINT, SIGTERM, etc.) are already set up to produce corresponding Pure exceptions of the form signal SIG where SIG is the signal number. If a script is to be run non-interactively then you'll have to do this yourself (otherwise most signals will terminate the program).

Last but not least, exceptions can also be used to implement non-local value returns. For instance, here's a variation of our n queens algorithm which only returns the first solution. Note the use of throw in the recursive search routine to bail out with a solution as soon as we found one. The value thrown there is caught in the main routine. Also note the use of void in the second equation of search. This effectively turns the list comprehension into a simple loop which suppresses the normal list result and just returns () instead. Thus, if no value gets thrown then the function regularly returns with () to indicate that there is no solution.

queens n       = catch reverse (search n 1 []) with
  search n i p = throw p if i>n;
               = void [search n (i+1) ((i,j):p) | j = 1..n; safe (i,j) p];
  safe (i,j) p = ~any (check (i,j)) p;
  check (i1,j1) (i2,j2)
               = i1==i2 || j1==j2 || i1+j1==i2+j2 || i1-j1==i2-j2;

E.g., let's compute a solution for a standard 8x8 board:

> queens 8;

9   C Interface

Accessing C functions from Pure programs is dead simple. You just need an extern declaration of the function, which is a simplified kind of C prototype. The function can then be called in Pure just like any other. For instance, the following commands, entered interactively in the interpreter, let you use the sin function from the C library (of course you could just as well put the extern declaration into a script):

> extern double sin(double);
> sin 0.3;

An extern declaration can also be prefixed with a public/private scope specifier:

private extern double sin(double);

Multiple prototypes can be given in one extern declaration, separating them with commas:

extern double sin(double), double cos(double), double tan(double);

For clarity, the parameter types can also be annotated with parameter names, e.g.:

extern double sin(double x);

Parameter names in prototypes only serve informational purposes and are for the human reader; they are effectively treated as comments by the compiler.

The interpreter makes sure that the parameters in a call match; if not, the call is treated as a normal form expression by default, which gives you the opportunity to extend the external function with your own Pure equations (see below). The range of supported C types is a bit limited right now (void, bool, char, short, int, long, float, double, as well as arbitrary pointer types, i.e.: void*, char*, etc.), but in practice these should cover most kinds of calls that need to be done when interfacing to C libraries.

Single precision float arguments and return values are converted from/to Pure's double precision floating point numbers automatically.

A variety of C integer types (char, short, int, long) are provided which are converted from/to the available Pure integer types in a straightforward way. In addition, the synonyms int8, int16 and int32 are provided for char, short and int, respectively, and int64 denotes 64 bit integers (a.k.a. ISO C99 long long). Note that long is equivalent to int32 on 32 bit systems, whereas it is the same as int64 on most 64 bit systems. All integer parameters take both Pure ints and bigints as actual arguments; truncation or sign extension is performed as needed, so that the C interface behaves as if the argument was "cast" to the C target type. Returned integers use the smallest Pure type capable of holding the result, i.e., int for the C char, short and int types, bigint for int64.

To make it easier to interface to various system routines, there's also a special size_t integer type which usually is 4 bytes on 32 bit and 8 bytes on 64 bit systems.

Pure considers all integers as signed quantities, but it is possible to pass unsigned integers as well (if necessary, you can use a bigint to pass positive values which are too big to fit into a machine int). Also note that when an unsigned integer is returned by a C routine, which is too big to fit into the corresponding signed integer type, it will "wrap around" and become negative. In this case, depending on the target type, you can use the ubyte, ushort, uint, ulong and uint64 functions provided by the prelude to convert the result back to an unsigned quantity.

Concerning the pointer types, char* is for string arguments and return values which need translation between Pure's internal utf-8 representation and the system encoding, while void* is for any generic kind of pointer (including strings, which are not translated when passed/returned as void*). Any other kind of pointer (except expr* and the GSL matrix pointer types, which are discussed below) is effectively treated as void* right now, although in a future version the interpreter may keep track of the type names for the purpose of checking parameter types.

The expr* pointer type is special; it indicates a Pure expression parameter or return value which is just passed through unchanged. All other types of values have to be "unboxed" when they are passed as arguments (i.e., from Pure to C) and "boxed" again when they are returned as function results (from C to Pure). All of this is handled by the runtime system in a transparent way, of course.

The matrix pointer types dmatrix*, cmatrix* and imatrix* can be used to pass double, complex double and int matrices to GSL functions taking pointers to the corresponding GSL types (gsl_matrix, gsl_matrix_complex and gsl_matrix_int) as arguments or returning them as results. Note that there is no marshalling of Pure's symbolic matrix type, as these aren't supported by GSL anyway. Also note that matrices are always passed by reference. If you need to pass a matrix as an output parameter of a GSL matrix routine, you can either create a zero matrix or a copy of an existing matrix. The prelude provides various operations for that purpose (in particular, see the dmatrix, cmatrix, imatrix and pack functions in matrices.pure). For instance, here is how you can quickly wrap up GSL's double matrix addition function in a way that preserves value semantics:

> using "lib:gsl";
> extern int gsl_matrix_add(dmatrix*, dmatrix*);
> x::matrix + y::matrix = gsl_matrix_add x y $$ x when x = pack x end;
> let x = dmatrix {1,2,3}; let y = dmatrix {2,3,2}; x; y; x+y;

Most GSL matrix routines can be wrapped in this fashion quite easily. A ready-made GSL interface providing access to all of GSL's numeric functions is in the works; please check the Pure website for details.

For convenience, it is also possible to pass a numeric matrix for a short*, int*, float* or double* parameter. The required conversions are done automatically, on the fly, and the matrix data is copied to temporary storage in order to preserve value sematics.

In addition, any kind of matrix (including symbolic matrices) can also be passed for a generic void* pointer. In this case no conversions are done and a pointer to the raw matrix data is passed, which allows the matrix to be modified in-place. Similarly, void* also allows you to pass a bigint argument as a raw mpz_t value. This makes it possible to call most GMP integer routines directly from Pure.

As already mentioned, it is possible to augment an external C function with ordinary Pure equations, but in this case you have to make sure that the extern declaration of the function comes first. For instance, we might want to extend our imported sin function with a rule to handle integers:

> extern double sin(double);
> sin 0.3;
> sin 0;
sin 0
> sin x::int = sin (double x);
> sin 0;

Sometimes it is preferable to replace a C function with a wrapper function written in Pure. In such a case you can specify an alias under which the original C function is known to the Pure program, so that you can still call the C function from the wrapper. An alias is introduced by terminating the extern declaration with a clause of the form = alias. For instance:

> extern double sin(double) = c_sin;
> sin x::double = c_sin x;
> sin x::int = c_sin (double x);
> sin 0.3; sin 0;

As an alternative, you can also declare the C function in a special namespace (cf. Namespaces in the Declarations section):

> namespace c;
> extern double sin(double);
> c::sin 0.3;

Note that the namespace qualification only affects the Pure side; the underlying C function is still called under the unqualified name as usual. The way in which such qualified externs are accessed is the same as for ordinary qualified symbols. In particular, the using namespace declaration applies as usual, and you can declare such symbols as private if needed. It is also possible to combine a namespace qualifier with an alias:

> namespace c;
> extern double sin(double) = mysin;
> c::mysin 0.3;

External C functions are resolved by the LLVM runtime, which first looks for the symbol in the C library and Pure's runtime library (or the interpreter executable, if the interpreter was linked statically). Thus all C library and Pure runtime functions are readily available in Pure programs. Other functions can be provided by adding them to the runtime, or by linking them into the runtime or the interpreter executable. Better yet, you can just "dlopen" shared libraries at runtime with a special form of the using clause:

using "lib:libname[.ext]";

For instance, if you want to call the functions from library libxyz directly from Pure:

using "lib:libxyz";

After this declaration the functions from the given library will be ready to be imported into your Pure program by means of corresponding extern declarations.

Shared libraries opened with using clauses are searched for in the same way as source scripts (see section Modules and Imports above), using the -L option and the PURE_LIBRARY environment variable in place of -I and PURE_INCLUDE. If the library isn't found by these means, the interpreter will also consider other platform-specific locations searched by the dynamic linker, such as the system library directories and LD_LIBRARY_PATH on Linux. The necessary filename suffix (e.g., .so on Linux or .dll on Windows) will be supplied automatically when needed. Of course you can also specify a full pathname for the library if you prefer that. If a library file cannot be found, or if an extern declaration names a function symbol which cannot be resolved, an appropriate error message is printed.

9.1   LLVM Bitcode Interface

As of Pure 0.44, the interpreter now also provides a direct way to import LLVM bitcode modules in Pure scripts. The main advantage of this method over the "plain" C interface explained above is that the bitcode loader knows all the call interfaces and can generate the necessary extern declarations automatically.

LLVM bitcode is loaded in a Pure script using the following special format of the using clause:

using "bc:mycfun";

(Here the bc tag indicates a bitcode file, and the default .bc bitcode filename extension is supplied automatically. Also, the bitcode file is searched for on the usual library search path.)

That's it, no explicit extern declarations are required on the Pure side. The Pure interpreter automatically creates extern declarations (in the current namespace) for all the external functions defined in the LLVM bitcode module, and generates the corresponding wrappers to make the functions callable from Pure. (This also works when batch-compiling a Pure script. In this case, the bitcode file actually gets linked into the output code, so the loaded bitcode module only needs to be present at compile time.)

Let's take a look at a concrete example to see how this actually works. Consider the following C code which defines a little function to compute the greatest common divisor of two (machine) integers:

int mygcd(int x, int y)
  if (y == 0)
    return x;
    return mygcd(y, x%y);

Let's say that this code is in the file mygcd.c, then you'd compile it to a bitcode module using llvm-gcc as follows:

llvm-gcc -emit-llvm -c mygcd.c -o mygcd.bc

Note that the -emit-llvm -c options instruct llvm-gcc to build an LLVM bitcode module. (Of course, you can also add optimizations and other options to the compile command as desired.)

You can now load the resulting bitcode module and run the mygcd function in the Pure interpreter simply as follows:

> using "bc:mygcd";
> mygcd 75 105;

To actually see the generated extern declaration of the imported function, you can use the interactive show command:

> show mygcd
extern int mygcd(int, int);

Some more examples showing how to use the bitcode interface can be found in the Pure sources. In particular, the interface also works with Fortran (using llvm-gfortran), and there is special support for interfacing to Grame's functional DSP programming language Faust (the latter uses a special variant of the bitcode loader, which is selected with the dsp tag in the using clause). Please refer to the corresponding examples in the distribution for further details.

Caveat: The LLVM bitcode interface is still considered somewhat experimental. It already seems to works fairly well with simple C and Fortran code, but there are some known limitations:

  • LLVM doesn't distinguish between char* and void* in bitcode, so all char* parameters and return values in C code will be promoted to void* on the Pure side. This means that string values won't be converted between Pure's internal UTF-8 encoding and the system encoding. If this is required then you'll have to do the conversions manually (e.g., using the appropriate functions from the Pure runtime).
  • The bitcode interface is limited to the same range of C types as Pure's plain C interface. In practice, this should cover most C code, but it's certainly possible that you run into unsupported types for arguments and return values (such as C structs passed by value). The compiler will then print a warning; the affected functions will still be linked in, but they will not be callable from Pure.
  • The bitcode linker doesn't verify loaded modules for self-containedness. The Pure interpreter will try to resolve external references from loaded libraries and other bitcode modules when the code is JIT-compiled, but if this fails then you'll get a fatal error message from the LLVM runtime. So you'll have to make sure that all required libraries and bitcode modules have been loaded before trying to execute the loaded functions.

10   Standard Library

Pure comes with a collection of Pure library modules, which includes the standard prelude (loaded automatically at startup time) and some other modules which can be loaded explicitly with a using clause. The prelude offers the necessary functions to work with the built-in types (including arithmetic and logical operations) and to do most kind of list processing you can find in ML- and Haskell-like languages. It also provides a collection of basic string and matrix operations. Please refer to the Pure Library Manual for details on the provided operations. Here is a very brief summary of some of the prelude operations which, besides the usual arithmetic and logical operators, are probably used most frequently:

The arithmetic + operation is also used to denote list and string concatenation in Pure.
This is the list-consing operation. x becomes the head of the list, y its tail. As ':' is a constructor symbol, you can use it in patterns on the left hand side of rewriting rules.
Constructs arithmetic sequences. x:y..z can be used to denote sequences with arbitrary stepsize y-x. Infinite sequences can be constructed using an infinite bound (i.e., inf or -inf). E.g., 1:3..inf denotes the stream of all odd integers starting at 1.
This is the pair constructor, used to create tuples of arbitrary sizes. Tuples provide an alternative way to represent aggregate values in Pure. In difference to lists, tuples are always "flat", so that (x,y),z and x,(y,z) denote the same triple x,y,z. (This is explained in more detail in the Pure Overview section.)
The size (number of elements) of the list, tuple, matrix or string x. In addition, dim x yields the dimensions (number of rows and columns) of a matrix.
This is Pure's indexing operation, which applies to lists, tuples, matrices and strings. Note that all indices in Pure are zero-based, thus x!0 and x!(#x-1) are the first and last element of x. In the case of matrices, the subscript may also be a pair of row and column indices, such as x!(1,2).
This is the "slicing" operation, which returns the list, tuple, matrix or string of all x!y while y runs through the (list or matrix) ys. Thus, e.g., x!!(i..j) returns all the elements between i and j (inclusive). Indices which fall outside the valid index range are quietly discarded. The index range ys may contain any number of indices (also duplicates), in any order. Thus x!![0|i=1..n] returns the first element of x n times, and, if ys is a permutation of the range 0..#x-1, then x!!ys yields the corresponding permutation of the elements of x. In the case of matrices the index range may also contain two-dimensional subscripts, or the index range itself may be specified as a pair of row/column index lists such as x!!(i..j,k..l).

The prelude also offers support operations for the implementation of list and matrix comprehensions, as well as the customary list operations like head, tail, drop, take, filter, map, foldl, foldr, scanl, scanr, zip, unzip, etc., which make list programming so much fun in modern FPLs. In Pure, these also work on strings as well as matrices, although, for reasons of efficiency, these data structures are internally represented as arrays.

Besides the prelude, Pure's standard library also comprises a growing number of additional library modules which we can only mention in passing here. In particular, the math.pure module provides additional mathematical functions as well as Pure's complex and rational number data types. Common container data structures like sets and dictionaries are implemented in the set.pure and dict.pure modules, among others. Moreover, the (beginnings of a) system interface can be found in the system.pure module. In particular, this module also provides operations to do basic C-style I/O, including printf and scanf. These are all described in much more detail in the Pure Library Manual.

11   Interactive Usage

In interactive mode, the interpreter reads definitions and expressions and processes them as usual. You can use the -i option to force interactive mode when invoking the interpreter with some script files. Additional scripts can be loaded interactively using either a using declaration or the interactive run command (see the description of the run command below for the differences between these). Or you can just start typing away, entering your own definitions and expressions to be evaluated.

The input language is just the same as for source scripts, and hence individual definitions and expressions must be terminated with a semicolon before they are processed. For instance, here is a simple interaction which defines the factorial and then uses that definition in some evaluations. Input lines begin with '>', which is the interpreter's default command prompt:

> fact 1 = 1;
> fact n = n*fact (n-1) if n>1;
> let x = fact 10; x;
> map fact (1..10);

As indicated, in interactive mode the normal forms of toplevel expressions are printed after each expression is entered. We also call this the read-eval-print loop. Normal form expressions are usually printed in the same form as you'd enter them. However, there are a few special kinds of objects like anonymous closures, thunks ("lazy" values to be evaluated when needed) and pointers which don't have a textual representation in the Pure syntax and will be printed in the format #<object description> by default. It is also possible to override the print representation of any kind of expression by means of the __show__ function, see the Caveats and Notes section for details.

11.1   Online Help

Online help is available in the interpreter with the interactive help command, see Interactive Commands below. You need to have a html browser installed for that. By default, the help command uses w3m(1), but you can change this by setting either the PURE_HELP or the BROWSER environment variable accordingly.

When invoked without arguments, the help command displays this manual:

> help

The help command also accepts a parameter which lets you specify a topic to show in the accompanying Pure Library manual, e.g.:

> help foldl

The help files distributed with the Pure interpreter are located in the Pure library directory (/usr/local/lib/pure by default). You can install additional documentation in html format in this directory, and look up topics in those help files with a command like the following:

> help mydoc#foo

Here mydoc is the basename of your help file (library path and .html suffix are supplied automatically), and foo can be any link target in the document (as specified with a <a name=...> tag or an id attribute in the html source). To just read the mydoc.html file without specifying a target, type the following:

> help mydoc#

Note that just help mydoc wouldn't work, since it would look for an entry mydoc in the standard library documentation.

If the basename of the help file is missing, it defaults to the library manual, so help #foldl does the same as help foldl, and just help # is a convenient shorthand to bring up the library manual without a specific topic. Of course, this syntax also works for looking up sections in the Pure manual:

> help pure#declarations

Note that the docutils tools used to generate the html source of the Pure manual mangle the section titles so that they are in lowercase and blanks are replaced with hyphens. So to look up the present section in this manual you'd have to type:

> help pure#online-help

You can also point the help browser to a proper URL, either a local file or some website (provided that your browser program can handle these). For instance:

> help file:mydoc.html#foo
> help

11.2   Interactive Commands

When running interactively, the interpreter accepts a number of special commands useful for interactive purposes. Here is a quick rundown of the currently supported operations:

! command
Shell escape.
break [symbol ...]
Sets breakpoints on the given function or operator symbols. All symbols must be specified in fully qualified form, see the remarks below. If invoked without arguments, prints all currently defined breakpoints. This requires that the interpreter was invoked with the -g option to enable debugging support. See Debugging below for details.
Prints a full backtrace of the call sequence of the most recent evaluation, if that evaluation ended with an unhandled exception. This requires that the interpreter was invoked with the -g option to enable debugging support. See Debugging below for details.
cd dir
Change the current working dir.
clear [option ...] [symbol ...]
Purge the definitions of the given symbols (functions, macros, constants or global variables). All symbols must be specified in fully qualified form, see the remarks below. If invoked as clear ans, clears the ans value (see Last Result below). When invoked without any arguments, clear purges all definitions at the current interactive "level" (after confirmation) and returns you to the previous level, if any. (It might be a good idea to first check your current definitions with show or back them up with dump before you do that.) The desired level can be specified with the -t option. See the description of the save command and Definition Levels below for further details. A description of the common options accepted by the clear, dump and show commands can be found in Specifying Symbol Selections below.
del [symbol ...]
Deletes breakpoints and tracepoints on the given function or operator symbols. All symbols must be specified in fully qualified form, see the remarks below. If invoked without arguments, clears all currently defined breakpoints and tracepoints (after confirmation). See Debugging below for details.
dump [-n filename] [option ...] [symbol ...]

Dump a snapshot of the current function, macro, constant and variable definitions in Pure syntax to a text file. All symbols must be specified in fully qualified form, see the remarks below. This works similar to the show command (see below), but writes the definitions to a file. The default output file is .pure in the current directory, which is then reloaded automatically the next time the interpreter starts up in interactive mode in the same directory. This provides a quick-and-dirty way to save an interactive session and have it restored later, but note that this isn't perfect yet. In particular, declarations of extern symbols won't be saved unless they're specified explicitly, and some objects like closures, thunks and pointers don't have a textual representation from which they could be reconstructed. To handle these, you'll probably have to prepare a corresponding .purerc file yourself, see Interactive Startup below.

A different filename can be specified with the -n option, which expects the name of the script to be written in the next argument, e.g: dump -n myscript.pure. You can then edit that file and use it as a starting point for an ordinary script or a .purerc file, or you can just run the file with the run command (see below) to restore the definitions in a subsequent interpreter session.

help [target]
Display the Pure manual or some other bit of documentation. In particular, help foo looks up the symbol foo in the Pure Library manual. See Online Help above for details.
ls [args]
List files (shell ls(1) command).
Print current memory usage. This reports the number of expression cells currently in use by the program, along with the size of the freelist (the number of allocated but currently unused expression cells). Note that the actual size of the expression storage may be somewhat larger than this, since the runtime always allocates expression memory in bigger chunks. Also, this figure does not reflect other heap-allocated memory in use by the program, such as strings or malloc'ed pointers.
Enter "override" mode. This allows you to add equations "above" existing definitions in the source script, possibly overriding existing equations. See Definition Levels below for details.
Print the current working dir (shell pwd(1) command).
Exits the interpreter.
run script

Loads the given script file and adds its definitions to the current environment. This works more or less like a using clause, but only searches for the script in the current directory and places the definitions in the script at the current temporary level, so that clear can be used to remove them again. In particular, this makes it possible to quickly reload a script without exiting the interpreter, by issuing the clear command followed by run. (This works best if you start out from a clean environment, with no scripts loaded on the command line.)

Also note that namespace and pragma settings of scripts loaded with run stick around after loading the script. This allows you to quickly set up your environment by just running a script containing the necessary namespace declarations and compiler directives. (Alternatively, you can also use the interpreter's startup files for that purpose, see Interactive Startup below.)

Begin a new level of temporary definitions. A subsequent clear command (see above) will purge the definitions made since the most recent save command. See Definition Levels below for details.
show [option ...] [symbol ...]
Show the definitions of symbols in various formats. See The show Command below for details. All symbols must be specified in fully qualified form, see the remarks below. A description of the common options accepted by the clear, dump and show commands can be found in Specifying Symbol Selections below.
stats [-m] [on|off]
Enables (default) or disables "stats" mode, in which some statistics are printed after an expression has been evaluated. Invoking just stats or stats on only prints the cpu time in seconds for each evaluation. If the -m option is specified, memory usage is printed along with the cpu time, which indicates the maximum amount of expression memory (in terms of expression cells) used during the computation. Invoking stats off disables stats mode, while stats -m off just disables the printing of the memory usage statistics.
trace [symbol ...]
Sets tracepoints on the given function or operator symbols. This works like the the break command (see above) but only prints rule invocations and reductions without actually interrupting the evaluation. See Debugging below for details.
Exits "override" mode. This returns you to the normal mode of operation, where new equations are added "below" previous rules of an existing function. See Definition Levels below for details.

Note that these special commands are only recognized at the beginning of the interactive command line (they are not reserved keywords of the Pure language). Thus it's possible to "escape" identifiers looking like commands by entering a space at the beginning of the line.

Also note that symbols (identifiers, operators etc.) must always be specified in fully qualified form. No form of namespace lookup is performed by these commands, so they always work the same no matter what namespace and using namespace declarations are currently in effect.

11.3   Last Result

Another convenience for interactive usage is the ans function, which retrieves the most recent result printed in interactive mode. For instance:

> fact n = if n<=1 then 1 else n*fact (n-1);
> map fact (1..10);
> scanl (+) 0 ans;

Note that ans is just an ordinary function, defined in the prelude, not a special command. However, there is a special clear ans command which purges the ans value. This is useful, e.g., if you got a huge result which you want to erase from memory before starting the next computation.

> clear ans
> ans;

11.4   Specifying Symbol Selections

The clear, dump and show commands all accept the following options for specifying a subset of symbols and definitions on which to operate. All symbols must be specified in fully qualified form. Options may be combined, thus, e.g., show -mft is the same as show -m -f -t. Some options specify optional numeric parameters; these must follow immediately behind the option character if present, as in -t0.

-c Selects defined constants.
-f Selects defined functions.
-g Indicates that the following symbols are actually shell glob patterns and that all matching symbols should be selected.
-m Select defined macros.
-pflag Select only private symbols if flag is nonzero (the default), otherwise (flag is zero) select only public symbols. If this option is omitted then both private and public symbols are selected.
-tlevel Select symbols and definitions at the given "level" of definitions and above. This is described in more detail below. Briefly, the executing program and all imported modules (including the prelude) are at level 0, while "temporary" definitions made interactively in the interpreter are at level 1 and above. Thus a level of 1 restricts the selection to all temporary definitions, whereas 0 indicates all definitions (i.e., everything, including the prelude). If level is omitted, it defaults to the current definitions level.
-v Select defined variables.

In addition, the -h option prints a short help message describing all available options of the command at hand.

If none of the -c, -f, -m and -v options are specified, then all kinds of symbols (constants, functions, macros and variables) are selected, otherwise only the specified categories will be considered.

A reasonable default is used if the -t option is omitted. By default, if no symbols are specified, only temporary definitions are considered, which corresponds to -t1. Otherwise the command applies to all corresponding definitions, no matter whether they belong to the executing program, the prelude, or some temporary level, which has the same effect as -t0. This default choice can be overridden by specifying the desired level explicitly.

As a special case, just clear (without any other options or symbol arguments) always backs out to the previous definitions level (instead of level #1). This is inconsistent with the rules set out above, but is implemented this way for convenience and backward compatibility. Thus, if you really want to delete all your temporary definitions, use clear -t1 instead. When used in this way, the clear command will only remove temporary definitions; if you need to remove definitions at level #0, you must specify those symbols explicitly.

Note that clear -g * will have pretty much the same disastrous consequences as the Unix command rm -rf *, so don't do that. Also note that a macro or function symbol may well have defining equations at different levels, in which case a command like clear -tn foo might only affect some part of foo's definition. The dump and show commands work analogously (albeit less destructively). See Definition Levels below for some examples.

11.5   The show Command

The show command can be used to obtain information about defined symbols in various formats. Besides the common selection options discussed above, this command recognizes the following additional options for specifying the content to be listed and the format to use.

-a Disassembles pattern matching automata. Works like the -v4 option of the interpreter.
-d Disassembles LLVM IR, showing the generated LLVM assembler code of a function. Works like the -v8 option of the interpreter.
-e Annotate printed definitions with lexical environment information (de Bruijn indices, subterm paths). Works like the -v2 option of the interpreter.
-l Long format, prints definitions along with the summary symbol information. This implies -s.
-s Summary format, print just summary information about listed symbols.

Symbols are always listed in lexicographic order. Note that some of the options (in particular, -a and -d) may produce excessive amounts of information. By setting the PURE_MORE environment variable, you can specify a shell command to be used for paging, usually more(1) or less(1).

For instance, to list all temporary definitions made in an interactive session, simply say:

> show

You can also list a specific symbol, no matter whether it comes from the interactive command line, the executing script or the prelude:

> show foldl
foldl f a x::matrix = foldl f a (list x);
foldl f a s::string = foldl f a (chars s);
foldl f a [] = a;
foldl f a (x:xs) = foldl f (f a x) xs;

Wildcards can be used with the -g option, which is useful if you want to print an entire family of related functions, e.g.:

> show -g foldl*
foldl f a x::matrix = foldl f a (list x);
foldl f a s::string = foldl f a (chars s);
foldl f a [] = a;
foldl f a (x:xs) = foldl f (f a x) xs;
foldl1 f x::matrix = foldl1 f (list x);
foldl1 f s::string = foldl1 f (chars s);
foldl1 f (x:xs) = foldl f x xs;

Or you can just specify multiple symbols as follows (this also works with multiple glob patterns when you add the -g option):

> show min max
max x y = if x>=y then x else y;
min x y = if x<=y then x else y;

You can also select symbols by category. E.g., the following command shows summary information about all the variable symbols along with their current values (using the "long" format):

> show -lvg *
argc       var  argc = 0;
argv       var  argv = [];
compiling  var  compiling = 0;
sysinfo    var  sysinfo = "x86_64-unknown-linux-gnu";
version    var  version = "0.37";
5 variables

Or you can list just private symbols of the namespace foo, as follows:

> show -pg foo::*

The following command will list each and every symbol that's currently defined (instead of -g * you can also use the -t0 option):

> show -g *

This usually produces a lot of output and is rarely needed, unless you'd like to browse through an entire program including all library imports. (In that case you might consider to use the dump command instead, which writes the definitions to a file which can then be loaded into a text editor for easier viewing. This may occasionally be useful for debugging purposes.)

Finally, there are two alternate forms of the show command: show namespace which lists the current and search namespaces, and show namespaces which lists all declared namespaces. These come in handy if you have forgotten what namespaces are currently active and which other namespaces are available in your program. For instance:

> show namespace
> show namespaces
namespace C;
namespace matrix;
> using namespace C;
> namespace my;
> show namespace
namespace my;
using namespace C;

11.6   Definition Levels

To help with incremental development, the interpreter offers some facilities to manipulate the current set of definitions interactively. To these ends, definitions are organized into different subsets called levels. As already mentioned, the prelude, as well as other source programs specified when invoking the interpreter, are always at level 0, while the interactive environment starts at level 1.

Each save command introduces a new temporary level, and each subsequent clear command (without any arguments) "pops" the definitions on the current level and returns you to the previous one (if any). This gives you a "stack" of temporary environments which enables you to "plug and play" in a (more or less) safe fashion, without affecting the rest of your program. For all practical purposes, this stack is unlimited, so that you can create as many levels as you like. Example:

> foo (x:xs) = x+foo xs;
> foo [] = 0;
> show
foo (x:xs) = x+foo xs;
foo [] = 0;
> foo (1..10);
> clear
This will clear all temporary definitions at level #1.
Continue (y/n)? y
> show
> foo (1..10);
foo [1,2,3,4,5,6,7,8,9,10]

We've seen already that normally, if you enter a sequence of equations, they will be recorded in the order in which they were written. However, it is also possible to override definitions in lower levels with the override command:

> foo (x:xs) = x+foo xs;
> foo [] = 0;
> show
foo (x:xs) = x+foo xs;
foo [] = 0;
> foo (1..10);
> save
save: now at temporary definitions level #2
> override
> foo (x:xs) = x*foo xs;
> show
foo (x:xs) = x*foo xs;
foo (x:xs) = x+foo xs;
foo [] = 0;
> foo (1..10);
warning: rule never reduced: foo (x:xs) = x+foo xs;

Note that the equation foo (x:xs) = x*foo xs; was inserted before the previous foo (x:xs) = x+foo xs; rule, which is at level #1. (The latter equation is now "shadowed" by the rule we just entered, hence the compiler warns us that this rule can't be reduced any more.)

Even in override mode, new definitions will be added after other definitions at the current level. This allows us to just continue adding more high-priority definitions overriding lower-priority ones:

> foo [] = 1;
> show
foo (x:xs) = x*foo xs;
foo [] = 1;
foo (x:xs) = x+foo xs;
foo [] = 0;
> foo (1..10);
warning: rule never reduced: foo (x:xs) = x+foo xs;
warning: rule never reduced: foo [] = 0;

Again, the new equation was inserted above the existing lower-priority rules, but below our previous foo (x:xs) = x*foo xs; equation entered at the same level. As you can see, we have now effectively replaced our original definition of foo with a version that calculates list products instead of sums, but of course we can easily go back one level to restore the previous definition:

> clear
This will clear all temporary definitions at level #2.
Continue (y/n)? y
clear: now at temporary definitions level #1
clear: override mode is on
> show
foo (x:xs) = x+foo xs;
foo [] = 0;
> foo (1..10);

Note that clear reminded us that override mode is still enabled (save will do the same if override mode is on while pushing a new definitions level). To turn it off again, use the underride command. This will revert to the normal behaviour of adding new equations below existing ones:

> underride

Finally, it's also possible to use clear to back out multiple levels at once, if you specify the target level to be cleared with the -t option. For instance:

> save
save: now at temporary definitions level #2
> let bar = 99;
> show
let bar = 99;
foo (x:xs) = x+foo xs;
foo [] = 0;
> clear -t1 // this scraps all our scribblings!
This will clear all temporary definitions at level #1 and above.
Continue (y/n)? y
clear: now at temporary definitions level #1
> show

11.7   Debugging

The interpreter provides a simple but reasonably convenient symbolic debugging facility when running interactively. To make this work, you have to specify the -g option when invoking the interpreter:

$ pure -g

This option disables tail call optimization (see Stack Size and Tail Recursion) to make it easier to debug programs. It also causes special debugging code to be generated which will make your program run much slower. Therefore the -g option should only be used if you actually need the debugger.

One common use of the debugger is "post mortem" debugging after an evaluation ended with an unhandled exception. In such a case, the bt command of the interpreter prints a backtrace of the call sequence which caused the exception. Note that this only works if debugging mode was enabled. For instance:

> [1,2]!3;
<stdin>, line 2: unhandled exception 'out_of_bounds' while evaluating '[1,2]!3'
> bt
   [1] (!): (x:xs)!n::int = xs!(n-1) if n>0;
     n = 3; x = 1; xs = [2]
   [2] (!): (x:xs)!n::int = xs!(n-1) if n>0;
     n = 2; x = 2; xs = []
   [3] (!): []!n::int = throw out_of_bounds;
     n = 1
>> [4] throw: extern void pure_throw(expr*) = throw;
     x1 = out_of_bounds

The last call, which is also marked with the >> symbol, is the call that raised the exception. The format is similar to the p command of the debugger, see below, but bt always prints a full backtrace. (As with the show command of the interpreter, you can set the PURE_MORE environment variable to pipe the output through the corresponding command, or use evalcmd to capture the output of bt in a string, cf. Reflection.)

The debugger can also be used interactively. To these ends, you can set breakpoints on functions with the break command. The debugger then gets invoked as soon as a rule for one of the given functions is executed. Example:

> fact n::int = if n>0 then n*fact (n-1) else 1;
> break fact
> fact 1;
** [1] fact: fact n::int = if n>0 then n*fact (n-1) else 1;
     n = 1
(Type 'h' for help.)
** [2] fact: fact n::int = if n>0 then n*fact (n-1) else 1;
     n = 0
++ [2] fact: fact n::int = if n>0 then n*fact (n-1) else 1;
     n = 0
     --> 1
** [2] (*): x::int*y::int = x*y;
     x = 1; y = 1
++ [2] (*): x::int*y::int = x*y;
     x = 1; y = 1
     --> 1
++ [1] fact: fact n::int = if n>0 then n*fact (n-1) else 1;
     n = 1
     --> 1

Lines beginning with ** indicate that the evaluation was interrupted to show the rule (or external) which is currently being considered, along with the current depth of the call stack, the invoked function and the values of parameters and other local variables in the current lexical environment. In contrast, the prefix ++ denotes reductions which were actually performed during the evaluation and the results that were returned by the function call (printed as --> return value).

Sometimes you might also see funny symbols like #<case>, #<when> or #<closure> instead of the function name. These indicate lambdas and the special variable-binding environments, which are all implemented as anonymous closures in Pure. Also note that the debugger doesn't know about the argument names of external functions (which are optional in Pure and not recorded anywhere), so it will display the generic names x1, x2 etc. instead.

At the debugger prompt ':' you can enter various special debugger commands, or just keep on hitting the carriage return key to walk through an evaluation step by step, as we did in the example above. (Command line editing works as usual at the debugger prompt, if it is enabled.) The usual commands are provided to walk through an evaluation, print and navigate the call stack, step over the current call, or continue the evaluation unattended until you hit another breakpoint. If you know other source level debuggers like gdb then you should feel right at home. You can type h at the debugger prompt to print the following list:

: h
Debugger commands:
a       auto: step through the entire program, run unattended
c [f]   continue until next breakpoint, or given function f
h       help: print this list
n       next step: step over reduction
p [n]   print rule stack (n = number of frames)
r       run: finish evaluation without debugger
s       single step: step into reduction
t, b    move to the top or bottom of the rule stack
u, d    move up or down one level in the rule stack
x       exit the interpreter (after confirmation)
.       reprint current rule
! cmd   shell escape
? expr  evaluate expression
<cr>    single step (same as 's')
<eof>   step through program, run unattended (same as 'a')

The command syntax is very simple. Besides the commands listed above you can also enter comment lines (// comment text) which will just be ignored. Extra arguments on commands which don't expect any will generally be ignored as well. The single letter commands all have to be separated from any additional parameters with whitespace, whereas the '!', '?' and '.' commands count as word delimiters and can thus be followed immediately by an argument. For convenience, the '?' command can also be omitted if the expression to be evaluated doesn't start with a single letter or one of the special punctuation commands.

The debugger can be exited or suspended in the following ways:

  • You can type c to continue the evaluation until the next breakpoint, or c foo in order to proceed until the debugger hits an invokation of the function foo.
  • You can type r to run the rest of the evaluation without the debugger.
  • The a ("auto") command single-steps through the rest of the evaluation, running unattended. This command can also be entered by just hitting the end-of-file key (Ctrl-D on Unix systems) at the debugger prompt.
  • You can also type x to exit from the debugger and the interpreter immediately (after confirmation).

At the debugger prompt, you can use the u ("up"), d ("down"), t ("top") and b ("bottom") commands to move around on the current call stack. The p command prints a range of the call stack centered around the currently selected stack frame, which is indicated with the >> tag, whereas ** denotes the current bottom of the stack (which is the rule to be executed with the single step command). The p command can also be followed by a numeric argument which indicates the number of stack frames to be printed (this will then become the default for subsequent invocations of p). The n command steps over the call selected with the stack navigation commands. For instance:

> fact 3;
** [1] fact: fact n::int = if n>0 then n*fact (n-1) else 1;
     n = 3
: c *
** [4] (*): x::int*y::int = x*y;
     x = 1; y = 1
: p
   [1] fact: fact n::int = if n>0 then n*fact (n-1) else 1;
     n = 3
   [2] fact: fact n::int = if n>0 then n*fact (n-1) else 1;
     n = 2
   [3] fact: fact n::int = if n>0 then n*fact (n-1) else 1;
     n = 1
** [4] (*): x::int*y::int = x*y;
     x = 1; y = 1
: u
>> [3] fact: fact n::int = if n>0 then n*fact (n-1) else 1;
     n = 1
: u
>> [2] fact: fact n::int = if n>0 then n*fact (n-1) else 1;
     n = 2
: p
   [1] fact: fact n::int = if n>0 then n*fact (n-1) else 1;
     n = 3
>> [2] fact: fact n::int = if n>0 then n*fact (n-1) else 1;
     n = 2
   [3] fact: fact n::int = if n>0 then n*fact (n-1) else 1;
     n = 1
** [4] (*): x::int*y::int = x*y;
     x = 1; y = 1
: n
++ [2] fact: fact n::int = if n>0 then n*fact (n-1) else 1;
     n = 2
     --> 2
** [2] (*): x::int*y::int = x*y;
     x = 3; y = 2

If you ever get lost, you can reprint the current rule with the '.' command:

: .
** [2] (*): x::int*y::int = x*y;
     x = 3; y = 2

Another useful feature is the ? command which lets you evaluate any Pure expression, with the local variables of the current rule bound to their corresponding values. Like the n command, ? applies to the current stack frame as selected with the stack navigation commands. The expression must be entered on a single line, and the trailing semicolon is optional. For instance:

> fact 3;
** [1] fact: fact n::int = if n>0 then n*fact (n-1) else 1;
     n = 3
: c *
** [4] (*): x::int*y::int = x*y;
     x = 1; y = 1
: ?x+y
: u
>> [3] fact: fact n::int = if n>0 then n*fact (n-1) else 1;
     n = 1
: n>0, fact n

A third use of the debugger is to trace function calls. For that the interpreter provides the trace command which works similarly to break, but sets so-called "tracepoints" which only print rule invocations and reductions instead of actually interrupting the evaluation. For instance, assuming the same example as above, let's first remove the breakpoint on fact (using the del command) and then set it as a tracepoint instead:

> del fact
> trace fact
> fact 1;
** [1] fact: fact n::int = if n>0 then n*fact (n-1) else 1;
     n = 1
** [2] fact: fact n::int = if n>0 then n*fact (n-1) else 1;
     n = 0
++ [2] fact: fact n::int = if n>0 then n*fact (n-1) else 1;
     n = 0
     --> 1
++ [1] fact: fact n::int = if n>0 then n*fact (n-1) else 1;
     n = 1
     --> 1

(The break and trace commands can also be used in concert if you want to debug some functions while only tracing others.)

Note that the current sets of breakpoints and tracepoints can only be changed with the break, trace and del commands of the interpreter, see Interactive Commands above. Use break (or trace) without arguments to list the currently defined breakpoints (or tracepoints). Breakpoints and tracepoints can be deleted with the del command, which is followed by the function and operator symbols to be removed from the breakpoint and tracepoint lists. If del is invoked without arguments, it clears all breakpoints and tracepoints (after confirmation).

11.8   Interactive Startup

In interactive mode, the interpreter also runs some additional scripts at startup, after loading the prelude and the scripts specified on the command line. This lets you tailor the interactive environment to your liking.

The interpreter first looks for a .purerc file in the user's home directory (as given by the HOME environment variable) and then for a .purerc file in the current working directory. These are just ordinary Pure scripts which may contain any additional definitions that you need. The .purerc file in the home directory is for global definitions which should always be available when running interactively, while the .purerc file in the current directory can be used for project-specific definitions.

Finally, you can also have a .pure initialization file in the current directory, which is usually created with the dump command (see above). This file is loaded after the .purerc files if it is present.

The interpreter processes all these files in the same way as with the run command (see above). When invoking the interpreter, you can specify the --norc option on the command line if you wish to skip these initializations.

12   Batch Compilation

As of Pure 0.21, the interpreter has a new -c option which provides a convenient means to turn Pure scripts into standalone executables. This feature is still a bit experimental right now. In particular, note that the compiled executable is essentially a static snapshot of your program which is executed on the "bare metal", without a hosting interpreter. Only a minimal runtime system is provided. This considerably reduces startup times, but also implies the following quirks and limitations:

What all this boils down to is that anything which requires the compile time or interactive facilities of the interpreter, is unavailable. These restrictions only apply at run time, of course. At compile time the program is being executed by the interpreter so you can use eval and evalcmd in any desired way. See the description of the compiling variable below for how to distinguish these cases in your script.

For most kinds of scripts, the above restrictions aren't really that much of an obstacle, or can easily be worked around. For the few scripts which actually need the full dynamic capabilities of Pure you'll just have to run the script with the interpreter. This isn't a big deal either, only the startup will be somewhat slower because the script is compiled on the fly. Once the JIT has done its thing the "interpreted" script will run every bit as fast as the "compiled" one, since in fact both are compiled (only at different times) to exactly the same code!

Also note that during a batch compilation, the compiled program is actually executed as usual, i.e., the script is also run at compile time. This might first seem to be a big annoyance, but it actually opens the door for some powerful programming techniques like partial evaluation. It is also a necessity because of Pure's highly dynamic nature. For instance, Pure allows you to define constants by evaluating an arbitrary expression (see Constant Definitions below), and using eval a program can easily modify itself in even more unforeseeable ways. Therefore pretty much anything in your program can actually depend on previous computations performed while the program is being executed.

12.1   Example

For the sake of a concrete example, consider the following little script:

using system;

fact n = if n>0 then n*fact (n-1) else 1;

main n = do puts ["Hello, world!", str (map fact (1..n))];

if argc<=1 then () else main (sscanf (argv!1) "%d");

When invoked from the command line, with the number n as the first parameter, this program will print the string "Hello, world!" and the list of the first n factorials:

$ pure -x hello.pure 10
Hello, world!

Note the condition on argc in the last line of the script. This prevents the program from producing an exception if no command line parameters are specified, so that the program can also be run interactively:

$ pure -i -q hello.pure
> main 10;
Hello, world!
> quit

To turn the script into an executable, we just invoke the Pure interpreter with the -c option, using the -o option to specify the desired output file name:

$ pure -c hello.pure -o hello
$ ./hello 10
Hello, world!

Next suppose that we'd like to supply the value n at compile rather than run time. To these ends we want to turn the value passed to the main function into a compile time constant, which can be done as follows:

const n = if argc>1 then sscanf (argv!1) "%d" else 10;

(Note that we provide 10 as a default if n isn't specified on the command line.)

Moreover, in such a case we usually want to skip the execution of the main function at compile time. The Pure runtime provides a special system variable compiling which holds a truth value indicating whether the program is actually running under the auspices of the batch compiler, so that it can adjust accordingly. In our example, the evaluation of main becomes:

if compiling then () else main n;

Our program now looks as follows:

using system;

fact n = if n>0 then n*fact (n-1) else 1;

main n = do puts ["Hello, world!", str (map fact (1..n))];

const n = if argc>1 then sscanf (argv!1) "%d" else 10;
if compiling then () else main n;

This script "specializes" n to the first (compile time) parameter when being batch-compiled, and it still works as before when we run it through the interpreter in both batch and interactive mode, too:

$ pure -i -q hello.pure
Hello, world!
> main 5;
Hello, world!
> quit

$ pure -x hello.pure 7
Hello, world!

$ pure -o hello -c -x hello.pure 7
$ ./hello
Hello, world!

You'll rarely need an elaborate setup like this, most of the time something like our simple first example will do the trick. But, as you've seen, Pure can easily do it.

12.2   Code Size and Unstripped Executables

By default, the batch compiler strips unused functions from the output code, to keep the code size small. You can disable this with the -u option, in which case the output code includes all functions defined in the compiled program or imported through a using clause, even if they don't seem to be used anywhere. This considerably increases compilation times and makes the compiled executable much larger. For instance, on a 64 bit Linux systems with ELF binaries the executable of our hello.pure example is about thrice as large:

$ pure -o hello -c -x hello.pure 7 && ls -l hello
-rwxr-xr-x 1 ag users 178484 2010-01-12 06:21 hello
$ pure -o hello -c -u -x hello.pure 7 && ls -l hello
-rwxr-xr-x 1 ag users 541941 2010-01-12 06:21 hello

(Note that even the stripped executable is fairly large when compared to compiled C code, as it still contains the symbol table of the entire program, which is needed by the runtime environment.)

Stripped executables should be fine for most purposes, but you have to be careful when using eval in your compiled program. The compiler only does a static analysis of which functions might be reached from the initialization code (i.e., toplevel expressions and let bindings). It does not take into account code run via the eval routine. Thus, functions used only in evaled code will be stripped from the executable, as if they were never defined at all. If such a function is then being called using eval at runtime, it will evaluate to a plain constructor symbol.

If this is a problem then you can either use the -u option to produce an unstripped executable, or you can force functions to be included in the stripped executable with the --required pragma (cf. Code Generation Options). For instance:

#! --required foo
foo x = bar (x-1);
eval "foo 99";

There is another code generation option which may have a substantial effect on code size, namely the --noconst option. Normally, constant values defined in a const definition are precomputed at compile time and then stored in the generated executable; this reduces startup times but may increase the code size considerably if your program contains big constant values such as lists. If you prefer smaller executables then you can use the --noconst option to force the value of the constant to be recomputed at run time (which effectively turns the constant into a kind of read-only variable). For instance:

#! --noconst
const xs = 1L..100000L;
sum = foldl (+) 0;

using system;
puts $ str $ sum xs;

On my 64 bit Linux system this produces a 187115 bytes executable. Without --noconst the code becomes almost an order of magnitude larger in this case (1788699 bytes). On the other hand, the smaller executable also takes a little longer to run since it must first recompute the value of the list constant at startup. So you have to consider the tradeoffs in a given situation. Usually big executables aren't much of a problem on modern operating systems, but if your program contains a lot of big constants then this may become an important consideration. However, if a constant value takes a long time to compute then you'll be better off with the default behaviour of precomputing the value at compile time.

12.3   Other Output Code Formats

Note that while the batch compiler generates native executables by default, it can just as well create object files which can be linked into other C/C++ programs and libraries:

$ pure -o hello.o -c -x hello.pure 7

The .o extension tells the compiler that you want an object file. When linking the object module, you also need to supply an initialization routine which calls the __pure_main__ function in hello.o to initialize the compiled module. This routine is declared in C/C++ code as follows:

extern "C" void __pure_main__(int argc, char** argv);

As indicated, __pure_main__ is to be invoked with two parameters, the argument count and NULL-terminated argument vector which become the argc and the argv of the Pure program, respectively. (You can also just pass 0 for both arguments if you don't need to supply command line parameters.) The purpose of __pure_main__ is to initialize a shell instance of the Pure interpreter which provides the minimal runtime support necessary to execute the Pure program, and to invoke all "initialization code" (variable definitions and toplevel expressions) of the program itself.

A minimal C main function which does the job of initializing the Pure module looks as follows:

extern void __pure_main__(int argc, char** argv);

int main(int argc, char** argv)
  __pure_main__(argc, argv);
  return 0;

If you link the main routine with the Pure module, don't forget to also pull in the Pure runtime library. Assuming that the above C code is in pure_main.c:

$ gcc -c pure_main.c -o pure_main.o
$ g++ -o hello hello.o pure_main.o -lpure
$ ./hello
Hello, world!

(The C++ compiler is used as the linker here so that the standard C++ library gets linked in, too. This is necessary because Pure's runtime library is actually written in C++.)

In fact, this is pretty much what pure -c actually does for you when creating an executable.

If your script loads dynamic libraries (using "lib:...";) then you'll also have to link with those; all external references have to be resolved at compile time. This is taken care of automatically when creating executables. Otherwise it is a good idea to run pure -c with the -v0100 verbosity option so that it prints the libraries to be linked (in addition to the commands which are invoked in the compilation process):

$ pure -v0100 -c hello.pure -o hello.o
opt -f -std-compile-opts hello.o.bc | llc -f -o hello.o.s
gcc -c hello.o.s -o hello.o
Link with: g++ hello.o -lpure

Well, we already knew that, so let's consider a slightly more interesting example from Pure's ODBC module:

$ pure -v0100 -c pure-odbc/examples/menagerie.pure -o menagerie.o
opt -f -std-compile-opts menagerie.o.bc | llc -f -o menagerie.o.s
gcc -c menagerie.o.s -o menagerie.o
Link with: g++ menagerie.o /usr/local/lib/pure/ -lpure
$ g++ -shared -o menagerie.o /usr/local/lib/pure/ -lpure

Note that the listed link options are necessary but might not be sufficient; pure -c just makes a best guess based on the Pure source. On most systems this will be good enough, but if it isn't, you can just add options to the linker command as needed to pull in additional required libraries.

As this last example shows, you can also create shared libraries from Pure modules. However, on some systems (most notably x86_64), this requires that you pass the -fPIC option when batch-compiling the module, so that position-independent code is generated:

$ pure -c -fPIC pure-odbc/examples/menagerie.pure -o menagerie.o

Also note that even when building a shared module, you'll have to supply an initialization routine which calls __pure_main__ somewhere.

Last but not least, pure -c can also generate just plain LLVM assembler code:

pure -c hello.pure -o hello.ll

Note the .ll extension; this tells the compiler that you want an LLVM assembler file. An LLVM bitcode file can be created just as easily:

pure -c hello.pure -o hello.bc

In these cases you'll have to have to handle the rest of the compilation yourself. This gives you the opportunity, e.g., to play with special optimization and code generation options provided by the LLVM toolchain. Please refer to the LLVM documentation (in particular, the description of the opt and llc programs) for details.

12.4   Calling Pure Functions From C

Another point worth mentioning here is that you can't just call Pure functions in a batch-compiled module directly. That's because in order to call a Pure function, at least in the current implementation, you have to set up a Pure stack frame for the function. However, there's a convenience function called pure_funcall in the runtime API to handle this. This function takes a pointer to the Pure function, the argument count and the arguments themselves (as pure_expr* objects) as parameters. For instance, here is a pure_main.c module which can be linked against the hello.pure program from above, which calls the fact function from the Pure program:

#include <stdio.h>
#include <pure/runtime.h>

extern void __pure_main__(int argc, char** argv);
extern pure_expr *fact(pure_expr *x);

int main()
  int n = 10, m;
  __pure_main__(0, NULL);
  if (pure_is_int(pure_funcall(fact, 1, pure_int(n)), &m))
    printf("fact %d = %d\n", n, m);
  return 0;

And here's how you can compile, link and run this program:

$ pure -o hello.o -c -x hello.pure 7
$ gcc -o pure_main.o -c pure_main.c
$ g++ -o myhello hello.o pure_main.o -lpure
$ ./myhello
Hello, world!
fact 10 = 3628800

Note that the first two lines are output from the Pure program; the last line is what gets printed by the main routine in pure_main.c.

13   Caveats and Notes

This section is a grab bag of casual remarks, useful tips and tricks, and information on common pitfalls, quirks and limitations of the current implementation and how to deal with them.

13.1   Purity

People keep asking me what's so "pure" about Pure. The long and apologetic answer is that Pure tries to stay as close as possible to the spirit of term rewriting without sacrificing practicality. It's possible and in fact quite easy to write purely functional programs in Pure, and you're encouraged to use it that way when possible. On the other hand, Pure doesn't get in your way if you want to call external operations with side effects; it does allow you to call any C function after all.

The short answer is that I simply liked the name, and there wasn't any programming language named "Pure" yet (quite a feat nowadays), so there's one now. :)

13.2   Backward Compatibility

Pure 0.7 introduced built-in matrix structures, which called for some minor changes in the syntax of comprehensions and arithmetic sequences. Specifically, the template expression and generator/filter clauses of a comprehension are now separated with | instead of ;. Moreover, arithmetic sequences with arbitrary stepsize are now written x:y..z instead of x,y..z, and the '..' operator now has a higher precedence than the ',' operator. This makes writing matrix slices like x!!(i..j,k..l) much more convenient.

In Pure 0.13 the naming of the logical and bitwise operations was changed, so that these are now called ~, &&, || and not/and/or, respectively. (Previously, ~ was used for bitwise, not for logical negation, which was rather inconsistent, albeit compatible with the naming of the not operation in Haskell and ML.) Also, to stay in line with this naming scheme, inequality was renamed to ~= (previously !=).

Pure 0.14 introduced the namespaces feature. Consequently, the scope of private symbols is now confined to a namespace rather than a source module; scripts making use of private symbols need to be adapted accordingly. Also note that syntax like foo::int may now also denote a qualified symbol rather than a tagged variable, if foo has been declared as a namespace. You can work around such ambiguities by renaming the variable, or by placing spaces around the '::' delimiter (these aren't permitted in a qualified symbol, so the construct foo :: int is always interpreted as a tagged variable, no matter whether foo is also a valid namespace).

Pure 0.26 extended the namespaces feature to add support for hierarchical namespaces. This means that name lookup works in a slightly different fashion now (see Hierarchical Namespaces for details), but old code which doesn't use the new feature should continue to work unchanged.

Pure 0.26 also changed the nullary keyword to nonfix, which is more consistent with the other kinds of fixity declarations. Moreover, the parser was enhanced so that it can cope with a theoretically unbounded number of precedence levels, and the system of standard operators in the prelude was modified so that it becomes possible to sneak in new operator symbols with ease; details can be found in the Symbol Declarations section.

Pure 0.41 added support for optimization of indirect tail calls, so that any previous restrictions on the use of tail recursion in indirect function calls and mutually recursive globals have been removed. Moreover, the logical operators && and || are now tail-recursive in their second operand and can also be extended with user-defined equations, just like the other builtins. Note that this implies that the values returned by && and || aren't normalized to the values 0 and 1 any more (this isn't possible with tail call semantics). If you need this then you'll have to make sure that either the operands are already normalized, or you'll have to normalize the result yourself.

Also, as of Pure 0.41 the batch compiler produces stripped executables by default. To create unstripped executables you now have to use the -u option, see Code Size and Unstripped Executables for details. The -s option to produce stripped executables is still provided for backward compatibility, but it won't have any effect unless you use it to override a previous -u option.

Pure 0.43 changed the rules for looking up symbols in user-defined namespaces. Unqualified symbols are now created in the current (rather than the global) namespace by default, see Symbol Lookup and Creation for details. The -w option can be used to get warnings about unqualified symbols which are resolved to a different namespace than previously. It also provides a means to check your scripts for implicit declarations which might indicate missing or mistyped function symbols.

13.3   Error Recovery

The parser uses a fairly simplistic panic mode error recovery which tries to catch syntax errors at the toplevel only. This seems to work reasonably well, but might catch some errors much too late. Unfortunately, Pure's terseness makes it rather difficult to design a better scheme. As a remedy, the parser accepts an empty definition (just ; by itself) at the toplevel only. Thus, in interactive usage, if the parser seems to eat away your input without doing anything, entering an extra semicolon or two should break the spell, putting you back at the toplevel where you can start typing the definition again.

13.4   The __show__ Function

As of Pure 0.6, the interpreter provides a "hook" to override the print representations of expressions at runtime by means of the __show__ function, which works in a fashion similar to Haskell's show function. This feature is still a bit experimental, but seems to work reasonably well for the purposes for which it is intended.

__show__ is just an ordinary Pure function expected to return a string with the desired custom representation of a normal form value given as the function's single argument. This function is not defined by default, so you are free to add any rules that you want. The interpreter prints the strings returned by __show__ just as they are. It will not check whether they conform to Pure syntax and/or semantics, or modify them in any way.

Custom print representations are most useful for interactive purposes, if you're not happy with the default print syntax of some kinds of objects. One particularly useful application of __show__ is to change the format of numeric values. Here are some examples:

> using system;
> __show__ x::double = sprintf "%0.6f" x;
> 1/7;
> __show__ x::int = sprintf "0x%0x" x;
> 1786;
> using math;
> __show__ (x::double +: y::double) = sprintf "%0.6f+%0.6fi" (x,y);
> cis (-pi/2);

The prelude function str, which returns the print representation of any Pure expression, calls __show__ as well:

> str (1/7);

Conversely, you can call the str function from __show__, but in this case it always returns the default representation of an expression. This prevents the expression printer from going recursive, and allows you to define your custom representation in terms of the default one. E.g., the following rule removes the L suffixes from bigint values:

> __show__ x::bigint = init (str x);
> fact n = foldl (*) 1L (1..n);
> fact 30;

Of course, your definition of __show__ can also call __show__ itself recursively to determine the custom representation of an object.

One case which needs special consideration are thunks (futures). The printer will never use __show__ for those, to prevent them from being forced inadvertently. In fact, you can use __show__ to define custom representations for thunks, but only in the context of a rule for other kinds of objects, such as lists. For instance:

> nonfix ...;
> __show__ (x:xs) = str (x:...) if thunkp xs;
> 1:2:(3..inf);

Another case which needs special consideration are numeric matrices. For efficiency, the expression printer will always use the default representation for these, unless you override the representation of the matrix as a whole. E.g., the following rule for double matrices mimics Octave's default output format (for the sake of simplicity, this isn't perfect, but you get the idea):

> __show__ x::matrix =
>   strcat [printd j (x!(i,j))|i=0..n-1; j=0..m-1] + "\n"
> with printd 0 = sprintf "\n%10.5f"; printd _ = sprintf "%10.5f" end
> when n,m = dim x end if dmatrixp x;
> {1.0,1/2;1/3,4.0};
   1.00000   0.50000
   0.33333   4.00000

Finally, by just purging the definition of the __show__ function you can easily go back to the standard print syntax:

> clear __show__
> 1/7; 1786; cis (-pi/2);

Note that if you have a set of definitions for the __show__ function which should always be loaded at startup, you can put them into the interpreter's interactive startup files, see Interactive Usage.

13.5   Non-Linear Patterns

As explained in section Patterns, Pure allows multiple occurrences of the same variable in a pattern (so-called non-linearities):

foo x x = x;

This rule will only be matched if both occurrences of x are bound to the same value. More precisely, the two instances of x will checked for syntactic equality during pattern matching, using the same primitive provided by the prelude. This may need time proportional to the sizes of both argument terms, and thus become quite costly for big terms. In fact, same might not even terminate at all if the compared terms are both infinite lazy data structures, such as in foo (1..inf) (1..inf). So you have to be careful to avoid such uses.

When using non-linearities in conjunction with "as" patterns, you also have to make sure that the "as" variable does not occur inside the corresponding subpattern. Thus a definition like the following is illegal:

> foo xs@(x:xs) = x;
<stdin>, line 1: error in pattern (recursive variable 'xs')

The explanation is that such a pattern couldn't possibly be matched by a finite list anyway. Indeed, the only match for xs@(x:xs) would be an infinite list of x's, and there's no way that this condition could be verified in a finite amount of time. Therefore the interpreter reports a "recursive variable" error in such situations.

13.6   "As" Patterns

In the current implementation, "as" patterns cannot be placed on the "spine" of a function definition. Thus rules like the following, which have the pattern somewhere in the head of the left-hand side, will all provoke an error message from the compiler:

a@foo x y   = a,x,y;
a@(foo x) y = a,x,y;
a@(foo x y) = a,x,y;

This is because the spine of a function application is not available when the function is called at runtime. "As" patterns in pattern bindings (let, const, case, when) are not affected by this restriction since the entire value to be matched is available at runtime. For instance:

> case bar 99 of y@(bar x) = y,x+1; end;
bar 99,100

13.7   Head = Function

"As" patterns are also a useful device if you need to manipulate function applications in a generic way. Note that the "head = function" rule means that the head symbol f of an application f x1 ... xn occurring on (or inside) the left-hand side of an equation, variable binding, or pattern-matching lambda expression, is always interpreted as a literal function symbol (not a variable). This implies that you cannot match the "function" component of an application against a variable, at least not directly. An anonymous "as" pattern like f@_ does the trick, however, since the anonymous variable is always recognized, even if it occurs as the head symbol of a function application. Here's a little example which demonstrates how you can convert a function application to a list containing the function and all arguments:

> foo x = a [] x with a xs (x@_ y) = a (y:xs) x; a xs x = x:xs end;
> foo (a b c d);

This may seem a little awkward, but as a matter of fact the "head = function" rule is quite useful since it covers the common cases without forcing the programmer to declare "constructor" symbols (except nonfix symbols). On the other hand, generic rules operating on arbitrary function applications are not all that common, so having to "escape" a variable using the anonymous "as" pattern trick is a small price to pay for that convenience.

Sometimes you may also run into the complementary problem, i.e., to match a function argument against a given function. Consider this code fragment:

foo x = x+1;
foop f = case f of foo = 1; _ = 0 end;

You might expect foop to return true for foo, and false on all other values. Better think again, because in reality foop will always return true! In fact, the Pure compiler will warn you about the second rule of the case expression not being used at all:

> foop 99;
warning: rule never reduced: _ = 0;

This happens because an identifier on the left-hand side of a rule, which is neither the head symbol of a function application nor a nonfix symbol, is always considered to be a variable (cf. Parameters in Equations), even if that symbol is defined as a global function elsewhere. So foo isn't a literal name in the above case expression, it's a variable! (As a matter of fact, this is rather useful, since otherwise a rule like f g = g+1 would suddenly change meaning if you happen to add a definition like g x = x-1 somewhere else in your program, which certainly isn't desirable.)

A possible workaround is to "escape" the function symbol using an empty namespace qualifier:

foop f = case f of ::foo = 1; _ = 0 end;

This trick works in case expressions and function definitions, but fails in circumstances in which qualified variable symbols are permitted (i.e., in variable and constant definitions). A better solution is to employ the syntactic equality operator === defined in the prelude to match the target value against the function symbol. This allows you to define the foop predicate as follows:

> foop f = f===foo;
> foop foo, foop 99;

Another way to deal with the situation would be to just declare foo as a nonfix symbol. However, this makes the foo symbol "precious", i.e., after such a declaration it cannot be used as a local variable anymore. It's usually a good idea to avoid that kind of thing, at least for generic symbols, so the above solution is preferred in this case.

13.8   With and when

A common source of confusion is that Pure provides two different constructs to bind local function and variable symbols, respectively. This distinction is necessary because Pure does not segregate defined functions and constructors, and thus there is no magic to figure out whether an equation like foo x = y by itself is meant as a definition of a function foo with formal parameter x and return value y, or a pattern binding defining the local variable x by matching the pattern foo x against the value of y. The with construct does the former, when the latter. (As a mnemonic, you may consider that when conveys a sense of time, as the individual variable definitions in a when clause are executed in order, while the function definitions in a with clause are all done simultaneously.)

Another speciality is that with and when clauses are tacked on to the end of the expression they belong to. This mimics mathematical language and makes it easy to read and understand a definition in a "top-down" fashion. This style differs considerably from other block-structured programming languages, however, which often place local definitions in front of the code they apply to. To grasp the operational meaning of such nested definitions, it can be helpful to read the nested scopes "in reverse" (from bottom to top). Some people also prefer to write their programs that way. In difference to Haskell and ML which have let expressions to support that kind of notation, Pure doesn't provide any special syntax for this. But note that you can always write when clauses in the following style which places the "body" at the bottom of the clause:

result when
  y = foo (x+1);
  z = bar y;
  result = baz z;

This doesn't incur any overhead, since the compiler will always eliminate the trivial "tail binding" for the result value. E.g., the above will compile to exactly the same code as:

baz z when
  y = foo (x+1);
  z = bar y;

13.9   Numeric Calculations

If possible, you should decorate numeric variables on the left-hand sides of function definitions with the appropriate type tags, like int or double. This often helps the compiler to generate better code and makes your programs run faster. The | syntax makes it easy to add the necessary specializations of existing rules to your program. E.g., taking the polymorphic implementation of the factorial as an example, you only have to add a left-hand side with the appropriate type tag to make that definition go as fast as possible for the special case of machine integers:

fact n::int    |
fact n         = n*fact(n-1) if n>0;
               = 1 otherwise;

(This obviously becomes unwieldy if you have to deal with several numeric arguments of different types, however, so in this case it is usually better to just use a polymorphic rule.)

Also note that int (the machine integers), bigint (the GMP "big" integers) and double (floating point numbers) are all different kinds of objects. While they can be used in mixed operations (such as multiplying an int with a bigint which produces a bigint, or a bigint with a double which produces a double), the int tag will only ever match a machine int, not a bigint or a double. Likewise, bigint only matches bigints (never int or double values), and double only doubles. Thus, if you want to define a function operating on different kinds of numbers, you'll also have to provide equations for all the types that you need (or a polymorphic rule which catches them all). This also applies to equations matching against constant values of these types. In particular, a small integer constant like 0 only matches machine integers, not bigints; for the latter you'll have to use the "big L" notation 0L. Similarly, the constant 0.0 only matches doubles, but not ints or bigints.

13.10   Constant Definitions

Constants differ from variables in that they cannot be redefined (that's their main purpose after all) so that their values, once defined, can be substituted into other definitions which use them. For instance:

> const c = 2;
> foo x = c*x;
> show foo
foo x = 2*x;
> foo 99;

While a variable can be rebound to a new value at any time, you will get an error message if you try to do this with a constant:

> const c = 3;
<stdin>, line 5: symbol 'c' is already defined as a constant

Note that in interactive mode you can work around this by purging the old definition with the clear command. However, this won't affect any earlier uses of the symbol:

> clear c
> const c = 3;
> bar x = c*x;
> show foo bar
bar x = 3*x;
foo x = 2*x;

(You'll also have to purge any existing definition of a variable if you want to redefine it as a constant, or vice versa, since Pure won't let you redefine an existing constant or variable as a different kind of symbol. The same also holds if a symbol is currently defined as a function or a macro.)

Constants can also be used in patterns (i.e., on the left-hand side of a rule in a definition or a case expression), but only if you also declare the corresponding symbol as nonfix. This is useful, e.g., if you'd like to use constants such as true and false on the left-hand side of a definition, just like other nonfix symbols:

> show false true
const false = 0;
const true = 1;
> nonfix false true;
> check false = "no"; check true = "yes";
> show check
check 0 = "no";
check 1 = "yes";
> check (5>0);

Note that without the nonfix declaration, the above definition of check wouldn't work as intended, because the true and false symbols on the left-hand side of the two equations would be interpreted as local variables. Also note that the standard library never declares any constant symbols as nonfix, since once a symbol is nonfix there's no going back. Thus the library leaves this to the programmer to decide.

As the value of a constant is known at compile time, the compiler can apply various optimizations to uses of such values. In particular, the Pure compiler inlines constant scalars (numbers, strings and pointers) by literally substituting their values into the output code, and it also precomputes simple constant expressions involving only (machine) integer and double values. Example:

> extern double atan(double);
> const pi = 4*atan 1.0;
> show pi
const pi = 3.14159265358979;
> foo x = 2*pi*x;
> show foo
foo x = 6.28318530717959*x;

In addition, the LLVM backend eliminates dead code automatically, so you can employ a constant to configure your code for different environments, without any runtime penalties:

const win = index sysinfo "mingw32" >= 0;
check boy = bad boy if win;
          = good boy otherwise;

In this case the code for one of the branches of check will be completely eliminated, depending on the outcome of the configuration check.

For efficiency, constant aggregates (lists, tuples, matrices and other kinds of non-scalar terms) receive special treatment. Here, the constant is computed once and stored in a read-only variable which then gets looked up at runtime, just like an ordinary global variable. However, there's an important difference: If a script is batch-compiled (cf. Batch Compilation), the constant value is normally computed at compile time only; when running the compiled executable, the constant value is simply reconstructed, which is often much more efficient than recomputing its value. For instance, you might use this to precompute a large table whose computation may be costly or involve functions with side effects:

const table = [foo x | x = 1..1000000];
process table;

Note that this only works with const values which are completely determined at compile time. If a constant contains run time objects such as pointers and (local) functions, this is impossible, and the batch compiler will instead create code to recompute the value of the constant at run time. For instance, consider:

using system;
const p = malloc 100;
foo p;

Here, the value of the pointer p of course critically depends on its computation (involving a side effect which sets aside a corresponding chunk of memory). It would become unusable without actually executing the initialization, so the compiler generates the appropriate run time initialization code in this case. For all practical purposes, this turns the constant into a read-only variable. (There's also a code generation option to force this behaviour even for "normal" constants for which it's not strictly necessary, in order to create smaller executables; see Code Size and Unstripped Executables for details.)

13.11   External C Functions

The interpreter always takes your extern declarations of C routines at face value. It will not go and read any C header files to determine whether you actually declared the function correctly! So you have to be careful to give the proper declarations, otherwise your program will probably segfault calling the function.

You also have to be careful when passing generic pointer values to external C routines, since currently there is no type checking for these; any pointer type other than char*, expr* and the matrix pointer types is effectively treated as void*. This considerably simplifies lowlevel programming and interfacing to C libraries, but also makes it very easy to have your program segfault all over the place. Therefore it is highly recommended that you wrap your lowlevel code in Pure routines and data structures which do all the checks necessary to ensure that only the right kind of data is passed to C routines.

Another limitation of the C interface is that it does not offer any special support for C structs and C function parameters. However, an optional addon module is available which interfaces to the libffi library to provide that kind of functionality, please see the description of the pure-ffi module on the Pure website for details.

Last but not least, to make it easier to create Pure interfaces to large C libraries, there's a separate pure-gen program available at the Pure website. This program takes a C header (.h) file and creates a corresponding Pure module with definitions and extern declarations for the constants and functions declared in the header. Please refer to the pure-gen(1) manual page for details.

13.12   Calling Special Forms

Special forms are recognized at compile time only. Thus the catch function, as well as quote and the operators &&, ||, $$ and &, are only treated as special forms in direct (saturated) calls. They can still be used if you pass them around as function values or in partial applications, but in this case they lose all their special call-by-name argument processing.

13.13   The Quote

Like in Lisp, the quote special form can be used to construct literal expressions which can then be manipulated before actually evaluating them using the built-in eval function. However, there are some notable differences to the Lisp version of quote:

  • Local variables (i.e., variables bound by the left-hand side of an equation, a lambda expression or a case or when construct) can never be quoted, so the corresponding value gets substituted even into quoted expressions. Thus, e.g., '(2*42+2^n) when n = 12 end evaluates to 2*42+2^12 rather than 2*42+2^n.
  • Only simple expressions can be quoted in Pure. Special constructs embedded in a quoted expression, such as conditionals and local bindings, are evaluated as usual. For instance, '(2*42+(2^n when n = 2*6 end)) evaluates to 2*42+4096.0 rather than quoting the embedded when expression.

While it's possible to define a Lisp-like quasiquote in Pure (see the Recursive Macros section for a simplified version, a full implementation can be found in the Pure library sources), it's usually not needed. As discussed above, substitution of local variables is performed even in quoted expressions, so that it is quite easy to fill in variable parts in a quoted "template" expression. For instance:

> (\x -> '(2*x+1)) 99;
> foo x = '(2*x+1);
> foo 99; foo $ '(7/y);

In fact, it is possible to perform arbitrary computations right in the middle of a quoted expression, using one of the special if-then-else, case, when and with expressions, since these are never quoted either. Example:

> '(x+1 when x = '(2*3) end);

Another useful feature of Lisp's quasiquote is the capability to splice arguments into a function application. It is possible to achieve pretty much the same in Pure with the following variation of the $ operator which "curries" its second (tuple) operand:

infixr 0 $@ ;
f $@ ()     = f;
f $@ (x,xs) = f x $@ xs;
f $@ x      = f x;

Now you can write, e.g.:

> '(foo 1 2) $@ '(2/3,3/4);
foo 1 2 (2/3) (3/4)

One shortcoming of Pure's quote is that there is no way to quote special constructs such as lambdas. Macro expansion is inhibited in quoted expressions, however, so it is possible to work around this limitation by defining a custom special form to be used as a symbolic representation for, say, a lambda expression, which reduces to a real lambda when evaluated. To these ends, the eval function can be invoked with a string argument as follows:

def lambda x y = eval $ "\\ "+str ('x)+" -> "+str ('y);


> let l = 'lambda x (x+1); l;
lambda x (x+1)
> let f = eval l; f; f 9;
#<closure 0x7fdc3ca45be8>

Other special constructs, such as case, when and with can be handled in a similar fashion.

13.14   Laziness

Pure does lazy evaluation in the same way as Alice ML, providing an explicit operation (&) to defer evaluation and create a "future" which is called by need. However, note that like any language with a basically eager evaluation strategy, Pure cannot really support lazy evaluation in a fully automatic way. That is, coding an operation so that it works with infinite data structures usually requires additional thought, and sometimes special code will be needed to recognize futures in the input and handle them accordingly. This can be hard, but of course in the case of the prelude operations this work has already been done for you, so as long as you stick to these, you'll never have to think about these issues. (It should be noted here that lazy evaluation has its pitfalls even in fully lazy FPLs, such as hidden memory leaks and other kinds of subtle inefficiencies or non-termination issues resulting from definitions being too lazy or not lazy enough. You can read about that in any good textbook on Haskell.)

The prelude goes to great lengths to implement all standard list operations in a way that properly deals with streams (a.k.a. lazy lists). What this all boils down to is that all list operations which can reasonably be expected to operate in a lazy way on streams, will do so. (Exceptions are inherently eager operations such as #, reverse and foldl.) Only those portions of an input stream will be traversed which are strictly required to produce the result. For most purposes, this works just like in fully lazy FPLs such as Haskell. However, there are some notable differences:

  • Since Pure uses dynamic typing, some of the list functions may have to peek ahead one element in input streams to check their arguments for validity, meaning that these functions will be slightly more eager than their Haskell counterparts.
  • Pure's list functions never produce truly cyclic list structures such as the ones you get, e.g., with Haskell's cycle operation. (This is actually a good thing, because the current implementation of the interpreter cannot garbage-collect cyclic expression data.) Cyclic streams such as cycle [1] or fix (1:) will of course work as expected, but, depending on the algorithm, memory usage may increase linearly as they are traversed.
  • Pattern matching is always refutable (and therefore eager) in Pure. If you need something like Haskell's irrefutable matches, you'll have to code them explicitly using futures. See the definition of the unzip function in the prelude for an example showing how to do this.

Here are some common pitfalls with lazy data structures in Pure that you should be aware of:

  • Laziness and side effects don't go well together, as most of the time you can't be sure when a given thunk will be executed. So as a general guideline you should avoid side effects in thunked data structures. If you can't avoid them, then at least make sure that all accesses to the affected resources are done through a single instance of the thunked data structure. E.g., the following definition lets you create a stream of random numbers:

    > using math;
    > let xs = [random | _ = 1..inf];

    This works as expected if only a single stream created with random exists in your program. However, as the random function in the math module modifies an internal data structure to produce a sequence of pseudorandom numbers, using two or more such streams in your program will in fact modify the same underlying data structure and thus produce two disjoint subsequences of the same underlying pseudorandom sequence which might not be distributed uniformly any more.

  • You should avoid keeping references to potentially big (or even infinite) thunked data structures when traversing them (unless you specifically need to memoize the entire data structure). In particular, if you assign such a data structure to a local variable, the traversal of the data structure should then be invoked as a tail call. If you fail to do this, it forces the entire memoized part of the data structure to stay in main memory while it is being traversed, leading to rather nasty memory leaks. Please see the all_primes function in Lazy Evaluation and Streams for an example.

13.15   Reflection

Pure versions since 0.12 offer some basic reflection capabilities via the evalcmd primitive. This function provides access to interactive commands like clear, save and show, which enable you to inspect and modify the running program. The only "canonical" way to represent an entire Pure program in Pure itself is the program text, hence evalcmd only provides a textual interface at this time. But of course custom higher-level representations can be built on top of that, similar to those discussed in section The Quote.

Here's an example showing what can be done using the show command and a little bit of trivial text processing. The following sym_info function retrieves information about a given collection of global symbols in a way which can be processed in a Pure program. The cat argument can be any combination of the letters "c", "v", "f" and "m" denoting the categories of constants, variables, functions and macros, respectively. (You can also just leave this empty if you don't care about the type of symbol.) The pat argument is a shell-like glob pattern for the name of symbols which should be listed (just "*" matches all symbols). The result is a list of tuples (name, value, cat, descr) with the name of the symbol and its value, as well as the category and description of the symbol, as provided by show -s.

using system;
sym_info cat::string pat::string
= [name,eval ("("+name+")"),descr | name,descr = info]
  // Get the info about matching symbols from the 'show' command.
  info = evalcmd $ sprintf "show -sg%s %s" (cat,pat);
  // Split into lines.
  info = if null info then [""] else split "\n" $ init info;
  // Get rid of the last line with the summary information.
  info = init info;
  // Retrieve the information that we need.
  info = [x | x@(s,_) = map fields info;
  // Get rid of extra lines containing extern and fixity declarations.
          s ~= "extern" && s ~= "nonfix" && s ~= "outfix" &&
          s ~= "prefix" && s ~= "postfix" && ~fnmatch "infix*" s 0];
end with
  // Regex call to split the summary information about one symbol, as
  // returned by 'show -s', into the name and description parts.
  fields s::string = tuple $
          [info!2 | info = tail $ regs $ reg_info $
           regex "([^ ]+)[ ]+([a-z]*)[ ]*(.*)" REG_EXTENDED s 0];

E.g., this call retrieves information about all defined macros:

> sym_info "m" "*";
[("$",($),"mac","2 args, 1 rules"),(".",(.),"mac","3 args, 1 rules"),
("void",void,"mac","1 args, 6 rules")]

13.16   Hygienic Macros

As mentioned in the Macro Hygiene section, Pure macros are lexically scoped and thus "hygienic". Macro hygiene is a somewhat esoteric topic for most programmers, so let us take a brief look at what it's all about. The problem avoided by hygienic macros is that of name capture. There are actually two kinds of name capture which may occur in unhygienic macro systems:

  • A free symbol in the macro body inadvertently becomes bound to the value of a local symbol in the context in which the macro is called.
  • A free symbol in the macro call inadvertently becomes bound to the value of a local symbol in the macro body.

Pure's hygienic macros avoid both pitfalls. Here is an example for the first form of name capture:

> def G x = x+y;
> G 10 when y = 99 end;

Note that the expansion of the G macro correctly uses the global instance of y, even though y is locally defined in the context of the macro call. (Sometimes this form of name capture is actually used deliberately in order to make the macro use the binding of the symbol which is active at the point of the macro call. This never works in Pure, hence in such cases you will have to explicitly pass such symbols to the macro.)

In contrast, the second form of name capture is usually not intended, and is therefore more dangerous. Consider the following example:

> def F x = x+y when y = x+1 end;
> F y;

Pure again gives the correct result here. You'd have to be worried if you got (y+1)+(y+1) instead, which would result from the literal expansion y+y when y = y+1 end, where the (free) variable y passed to F gets captured by the local binding of y. In fact, that's exactly what you get with C macros:

#define F(x) { int y = x+1; return x+y; }

Here F(y) expands to { int y = y+1; return y+y; } which is usually not what you want.

There is also one Pure-related caveat here. The expression printer currently doesn't check for different bindings of the same variable identifier when it prints a (compile time) expression. E.g.:

> foo y = F y;
> show foo
foo y = y+y when y = y+1 end;

This looks as if y got captured, but in fact it's not, it's just the show command which displays the definition in an incorrect way. You can add the -e option to show which prints the deBruijn indices of locally bound symbols, then you see that the actual bindings are all right anyway (note that the number before the colon is the actual deBruijn index, the sequence of bits behind it is the subterm path):

> show -e foo
foo y/*0:1*/ = y/*1:1*/+y/*0:*/ when y/*0:*/ = y/*0:1*/+1 end;

Alas, this means that if you use dump to write such a definition to a text file and read it back with run later, then you'll get the wrong definition. Currently you will have to correct this manually.

13.17   Stack Size and Tail Recursion

Pure programs may need a considerable amount of stack space to handle recursive function and macro calls, and the interpreter itself also takes its toll. So you should configure your system accordingly (8 MB of stack space is recommended for 32 bit systems, systems with 64 bit pointers probably need more). If the PURE_STACK environment variable is defined, the interpreter performs advisory stack checks on function entry and raises a Pure exception if the current stack size exceeds the given limit. The value of PURE_STACK should be the maximum stack size in kilobytes. Please note that this is only an advisory limit which does not change the program's physical stack size. Your operating system should supply you with a command such as ulimit(1) to set the real process stack size. (The PURE_STACK limit should be a little less than that, to account for temporary stack usage by the interpreter itself.)

Like Scheme, Pure does proper tail calls (if LLVM provides that feature on the platform at hand), so tail-recursive definitions should work fine in limited stack space. For instance, the following little program will loop forever if your platform supports the required optimizations:

loop = loop;

This also works if your definition involves function parameters, guards and multiple equations, of course. Moreover, conditional expressions (if-then-else) are tail-recursive in both branches, and the logical operators && and ||, as well as the sequence operator $$, are tail-recursive in their second operand.

Also note that tail call optimization is always disabled if the debugger is enabled (-g). This makes it much easier to debug programs, but means that you may run into stack overflows when debugging a program that does deep tail recursion.

13.18   Handling of Asynchronous Signals

As described in section Exception Handling, signals delivered to the process can be caught and handled with Pure's exception handling facilities. This has its limitations, however. Since Pure code cannot be executed directly from a C signal handler, checks for pending signals are only done on function entry. This means that in certain situations (such as the execution of an external C routine), delivery of a signal may be delayed by an arbitrary amount of time. Moreover, if more than one signal arrives between two successive signal checks, only the last one will be reported in the current implementation.

When delivering a signal which has been remapped to a Pure exception, the corresponding exception handler (if any) will be invoked as usual. Further signals are blocked while the exception handler is being executed.

A fairly typical case is that you have to handle signals in a tail-recursive function. This can be done with code like the following:

using system;

// Remap some common POSIX signals.

loop = catch handler process $$ loop
with handler (signal k) = printf "Hey, I got signal %d.\n" k end;
process = sleep 1; // do something

Running the above loop function enters an endless loop reporting all signals delivered to the process. Note that to make this work, the tail-recursive invocation of loop must immediately follow the signal-handling code, so that signals don't escape the exception handler.

Of course, in a real application you'd probably want the loop function to carry around some data to be processed by the process routine, which then returns an updated value for the next iteration. This can be implemented as follows:

loop x = loop (catch handler (process x))
with handler (signal k) = printf "Hey, I got signal %d.\n" k $$ 0 end;
process x = printf "counting: %d\n" x $$ sleep 1 $$ x+1;

14   Copying

Pure comes with a fairly liberal license which lets you distribute your own Pure programs and extensions under a license of your choice and permits linking of commercial applications against the Pure runtime and the Pure standard library without requiring special permission. Moreover, the Pure interpreter (the pure main program), the Pure runtime library (libpure) and the Pure standard library (the Pure scripts in the lib folder distributed with the software) are distributed as free software, and you are welcome to modify and redistribute them under the appropriate license terms, as detailed below.

(The above explanations are not legal advice. Please read the full text of the licenses and consult qualified professional counsel for an interpretation of the license terms as they apply to you.)

The Pure interpreter is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation, either version 3 of the License, or (at your option) any later version.

The Pure runtime library and the Pure standard library are also free software: you can redistribute them and/or modify them under the terms of the GNU Lesser General Public License as published by the Free Software Foundation, either version 3 of the License, or (at your option) any later version.

Pure is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.

Please see the GNU General Public License and the GNU Lesser General Public License for the precise license terms. You can also find the license conditions in the COPYING and COPYING.LESSER files accompanying the software. Also, please see the source code for the copyright and license notes pertaining to individual source files which are part of this software.

Pure uses LLVM as its compiler backend. LLVM is under Copyright (c) 2003-2009 by the University of Illinois at Urbana-Champaign, and is licensed under a 3-clause BSD-style license, please read COPYING.LLVM included in the distribution for the exact licensing terms. You can also find the LLVM license at the LLVM website.

15   Author

Albert Gräf <>, Dept. of Computer Music, Johannes Gutenberg University of Mainz, Germany.

The author gratefully acknowledges the contributions by Scott E. Dillard, Rooslan S. Khayrov, Eddie Rucker, Libor Spacek, Jiri Spitz and Sergei Winitzki, as well as Toni Graffy, Michel Salim and Ryan Schmidt who maintain the SUSE Linux, Fedora Core and OSX packages, respectively. Thanks are also due to Vili Aapro, Alvaro Castro Castilla, John Cowan, Chris Double, Tim Haynes, Roman Neuhauser, Wm Leler, John Lunney and Max Wolf.

16   See Also

Wouter van Oortmerssen's functional programming language based on term rewriting,
Alice ML
A version of ML (see below) from which Pure borrows its model of lazy evaluation,
Wm Leler's constraint programming language based on term rewriting, See Wm Leler: Constraint Programming Languages: Their Specification and Generation. Addison-Wesley, 1988.
Grame's functional DSP programming language,
GNU Multiprecision Library
Free library for arbitrary precision arithmetic,
GNU Octave
A popular high-level language for numeric applications and free MATLAB replacement,
GNU Scientific Library
A free software library for numeric applications, can be used with Pure's numeric matrices,
A popular non-strict FPL,
The LLVM code generator framework,
A popular strict FPL. See Robin Milner, Mads Tofte, Robert Harper, D. MacQueen: The Definition of Standard ML (Revised). MIT Press, 1997.
Find the latest releases and the mailing list at the Pure website,
Another term rewriting language by yours truly,