Maintaining Files on Plan 9 with Mk

Andrew G. Hume
andrew@research.att.com
Bob Flandrena
bobf@plan9.bell-labs.com

ABSTRACT

Mk is a tool for describing and maintaining dependencies between files. It is similar to the UNIX program make, but provides several extensions. Mk's flexible rule specifications, implied dependency derivation, and parallel execution of maintenance actions are well-suited to the Plan 9 environment. Almost all Plan 9 maintenance procedures are automated using mk.

1 Introduction

This document describes how mk, a program functionally similar to make [Feld79], is used to maintain dependencies between files in Plan 9. Mk provides several extensions to the capabilities of its predecessor that work well in Plan 9's distributed, multi-architecture environment. It exploits the power of multiprocessors by executing maintenance actions in parallel and interacts with the Plan 9 command interpreter rc to provide a powerful set of maintenance tools. It accepts pattern-based dependency specifications that are not limited to describing rules for program construction. The result is a tool that is flexible enough to perform many maintenance tasks including database maintenance, hardware design, and document production.

This document begins by discussing the syntax of the control file, the pattern matching capabilities, and the special rules for maintaining archives. A brief description of mk's algorithm for deriving dependencies is followed by a discussion of the conventions used to resolve ambiguous specifications. The final sections describe parallel execution and special features.

An earlier paper [Hume87] provides a detailed discussion of mk's design and an appendix summarizes the differences between mk and make.

2 The `Mkfile`

Mk reads a file describing relationships among files and executes commands to bring the files up to date. The specification file, called a mkfile, contains three types of statements: assignments, includes, and rules. Assignment and include statements are similar to those in C. Rules specify dependencies between a target and its prerequisites. When the target and prerequisites are files, their modification times determine if they are out of date. Rules often contain a recipe, an rc(1) script that produces the target from the prerequisites.

This simple mkfile produces an executable from a C source file:

CC=pcc f1: f1.c $CC -o f1 f1.c

The first line assigns the name of the portable ANSI/POSIX compiler to the mk variable CC; subsequent references of the form $CC select this compiler. The only rule specifies a dependence between the target file f1 and the prerequisite file f1.c. If the target does not exist or if the prerequisite has been modified more recently than the target, mk passes the recipe to rc for execution. Here, f1.c is compiled and loaded to produce f1.

The native Plan 9 environment requires executables for all architectures, not only the current one. The Plan 9 version of the same mkfile looks like:

</$objtype/mkfile f1: f1.$O $LD $LDFLAGS -o f1 f1.$O f1.$O: f1.c $CC $CFLAGS f1.c

The first line is an include statement that replaces itself with the contents of the file /$objtype/mkfile. The variable $objtype is inherited from the environment and contains the name of the target architecture. The prototype mkfile for that architecture defines architecture-specific variables: CC and LD are the names of the compiler and loader, O is the code character of the architecture. The rules compile the source file into an object file and invoke the loader to produce f1. Invoking mk from the command line as follows

% objtype=mips mk vc -w f1.c vl $LDFLAGS -o f1 f1.k %

produces the mips executable of program f1 regardless of the current architecture type.

We can extend the mkfile to build two programs:

</$objtype/mkfile ALL=f1 f2 all:V: $ALL f1: f1.$O $LD $LDFLAGS -o f1 f1.$O f1.$O: f1.c $CC $CFLAGS f1.c f2: f2.$O $LD $LDFLAGS -o f2 f2.$O f2.$O: f2.c $CC $CFLAGS f2.c

The target all, modified by the attribute V, builds both programs. The attribute identifies all as a dummy target that is not related to a file of the same name; its precise effect is explained later. This example describes cascading dependencies: the first target depends on another which depends on a third and so on. Here, individual rules build each program; later we'll see how to do this with a general rule.

3 Variables and the environment

Mk does not distinguish between its internal variables and rc variables in the environment. When mk starts, it imports each environment variable into a mk variable of the same name. Before executing a recipe, mk exports all variables, including those inherited from the environment, to the environment in which rc executes the recipe.

There are several ways for a variable to take a value. It can be set with an assignment statement, inherited from the environment, or specified on the command line. Mk also maintains several special internal variables that are described in mk(1). Assignments have the following decreasing order of precedence:

: 1) Command line assignment
2) Assignment statement
3) Imported from the environment
4) Implicitly set by mk

For example, a command line assignment overrides a value imported from the environment.

All variable values are strings. They can be used for pattern matching and comparison but not for arithmetic. A list is a string containing several values separated by white space. Each member is handled individually during pattern matching, target selection, and prerequisite evaluation.

A namelist is a list produced by transforming the members of an existing list. The transform applies a pattern to each member, replacing each matched string with a new string, much as in the substitute command in sam(1) or ed(1). The syntax is

${var:A%B=C%D}

where var is a variable. The pattern A%B matches a member beginning with the string A and ending with the string B with any string in between; it behaves like the regular expression A.*B. When a member of the var list matches this pattern, the string C replaces A, D replaces B, and the matched string replaces itself. Any of A, B, C, or D may be the empty string. In effect, a namelist is generated by applying the ed(1) substitute command

s/A(.*)B/C\1D/

to each member of a variable list.

Namelists are useful for generating a list based on a predictable transformation. For example,

SRC=a.c b.c c.c OBJ=${SRC:%.c=%.v}

assigns the list (a.v b.v c.v) to OBJ. A namelist may be used anywhere a variable is allowed except in a recipe.

Command output is assigned to a variable using the normal rc syntax:

var=`{rc command}

The command executes in an environment populated with previously assigned variables, including those inherited from mk's execution environment. The command may be arbitrarily complex; for example,

TARG=`{ls -d *.[cy] | sed 's/..$//'}

assigns a list of the C and yacc source files in the current directory, stripped of their suffix, to the variable TARG.

4 The include statement

The include statement replaces itself with the contents of a file. It is functionally similar to the C #include statement but uses a different syntax:

<filename

The contents of the file are evaluated as they are read. An include statement may be used anywhere except in a recipe.

Unlike make, mk has no built-in rules. Instead, the include statement allows generic rules to be imported from a prototype mkfile; most Plan 9 mkfiles use this approach [Flan95].

5 Rules

A rule has four elements: targets, prerequisites, attributes, and a recipe. It has the form:

targets:attributes:prerequisites recipe

The first line, containing the targets, attributes, and prerequisites is the rule header; it must begin at the left margin. The recipe contains zero or more lines, each of which begins with white space. One or more targets must be specified but the attributes, prerequisites, and recipe are optional. A rule specifies a dependency between the target(s) and its prerequisite(s), the recipe brings the target(s) up to date with the prerequisite(s) and attributes modify mk's evaluation of the dependency.

Normally the target is a file that depends on one or more prerequisite files. Mk compares the modification times of each target and each prerequisite; a target is considered out of date when it does not exist or when a prerequisite has been modified more recently. When a target is out of date, mk executes the recipe to bring it up to date. When the recipe completes, the modification time of the target is checked and used in later dependency evaluations. If the recipe does not update the target, evaluation continues with the out of date target.

A prerequisite of one rule may be the target of another. When this happens, the rules cascade to define a multi-step procedure. For example, an executable target depends on prerequisite object files, each of which is a target in a rule with a C source file as the prerequisite. Mk follows a chain of dependencies until it encounters a prerequisite that is not a target of another rule or it finds a target that is up to date. It then executes the recipes in reverse order to produce the desired target.

The rule header is evaluated when the rule is read. Variables are replaced by their values, namelists are generated, and commands are replaced by their output at this time.

Most attributes modify mk's evaluation of a rule. An attribute is usually a single letter but some are more complicated. This paper only discusses commonly used attributes; see mk(1) for a complete list.

The V attribute identifies a virtual target; that is, a target that is not a file. For example,

clean:V: rm *.$O $O.out

removes executables and compiler intermediate files. The target is virtual because it does not refer to a file named clean. Without the attribute, the recipe would not be executed if a file named clean existed. The Q attribute silences the printing of a recipe before execution. It is useful when the output of a recipe is similar to the recipe:

default:QV: echo 'No default target; use mk all or mk install'

The recipe is an rc script. It is optional but when it is missing, the rule is handled specially, as described later. Unlike make, mk executes recipes without interpretation. After stripping the first white space character from each line it passes the entire recipe to rc on standard input. Since mk does not interpret a recipe, escape conventions are exactly those of rc. Scripts for awk and sed commands can be embedded exactly as they would be entered from the command line. Mk invokes rc with the -e flag, which causes rc to stop if any command in the recipe exits with a non-zero status; the E attribute overrides this behavior and allows rc to continue executing in the face of errors. Before a recipe is executed, variables are exported to the environment where they are available to rc. Commands in the recipe may not read from standard input because mk uses it internally.

References to a variable can yield different values depending on the location of the reference in the mkfile. Mk resolves variable references in assignment statements and rule headers when the statement is read. Variable references in recipes are evaluated by rc when the recipe is executed; this happens after the entire mkfile has been read. The value of a variable in a recipe is the last value assigned in the file. For example,

STRING=all all:VQ: echo $STRING STRING=none

produces the message none. A variable assignment in a recipe does not affect the value of the variable in the mkfile for two reasons. First, mk does not import values from the environment when a recipe completes; one recipe cannot pass a value through the environment to another recipe. Second, no recipe is executed until mk has completed its evaluation, so even if a variable were changed, it would not affect the dependency evaluation.

6 Metarules

A metarule is a rule based on a pattern. The pattern selects a class of target(s) and identifies related prerequisites. Mk metarules may select targets and prerequisites based on any criterion that can be described by a pattern, not just the suffix transformations associated with program construction.

Metarule patterns are either intrinsic or regular expressions conforming to the syntax of regexp(6). The intrinsic patterns are shorthand for common regular expressions. The intrinsic pattern % matches one or more of anything; it is equivalent to the regular expression `.+'. The other intrinsic pattern, &, matches one or more of any characters except `/' and `.'. It matches a portion of a path and is equivalent to the regular expression `[^./]+'. An intrinsic pattern in a prerequisite references the string matched by the same intrinsic pattern in the target. For example, the rule

%.v: %.c

says that a file ending in .v depends on a file of the same name with a .c suffix: foo.v depends on foo.c, bar.v depends on bar.c, and so on. The string matched by an intrinsic pattern in the target is supplied to the recipe in the variable $stem. Thus the rule

%.$O: %.c $CC $CFLAGS $stem.c

creates an object file for the target architecture from a similarly named C source file. If several object files are out of date, the rule is applied repeatedly and $stem refers to each file in turn. Since there is only one stem variable, there can only be one % or & pattern in a target; the pattern %-%.c is illegal.

Metarules simplify the mkfile for building programs f1 and f2:

</$objtype/mkfile ALL=f1 f2 all:V: $ALL %: %.$O $LD -o $target $prereq %.$O: %.c $CC $CFLAGS $stem.c clean:V: rm -f $ALL *.[$OS]

(The variable $OS is a list of code characters for all architectures.) Here, metarules specify compile and load steps for all C source files. The loader rule relies on two internal variables set by mk during evaluation of the rule: $target is the name of the target(s) and $prereq the name of all prerequisite(s). Metarules allow this mkfile to be easily extended; a new program is supported by adding its name to the third line.

A regular expression metarule must have an R attribute. Prerequisites may reference matching substrings in the target using the form n where n is a digit from 1 to 9 specifying the nth parenthesized sub-expression. In a recipe, $stemn is the equivalent reference. For example, a compile rule could be specified using regular expressions:

(.+)\.$O:R: \1.c $CC $CFLAGS $stem1.c

Here, 1 and $stem1 refer to the name of the target object file without the suffix. The variable $stem associated with an intrinsic pattern is undefined in a regular expression metarule.

7 Archives

Mk provides a special mechanism for maintaining an archive. An archive member is referenced using the form lib(file) where lib is the name of the archive and file is the name of the member. Two rules define the dependency between an object file and its membership in an archive:

$LIB(foo.8):N: foo.8 $LIB: $LIB(foo.8) ar rv $LIB foo.8

The first rule establishes a dependency between the archive member and the object file. Normally, mk detects an error when a target does not exist and the rule contains no recipe; the N attribute overrides this behavior because the subsequent rule updates the member. The second rule establishes the dependency between the member and the archive; its recipe inserts the member into the archive. This two-step specification allows the modification time of the archive to represent the state of its members. Other rules can then specify the archive as a prerequisite instead of listing each member.

A metarule generalizes library maintenance:

LIB=lib.a OBJS=etoa.$O atoe.$O ebcdic.$O $LIB(%):N: % $LIB: ${OBJS:%=$LIB(%)} ar rv $LIB $OBJS

The namelist prerequisite of the $LIB target generates archive member names for each object file name; for example, etoa.$O becomes lib.a(etoa.$O). This formulation always updates all members. This is acceptable for a small archive, but may be slow for a big one. The rule

$LIB: ${OBJS:%=$LIB(%)} ar rv $LIB `{membername $newprereq}

only updates out of date object files. The internal variable $newprereq contains the names of the out of date prerequisites. The rc script membername transforms an archive member specification into a file name: it translates lib.a(etoa.$O) into etoa.$O.

The mkfile

</$objtype/mkfile LIB=lib.a OBJS=etoa.$O atoe.$O ebcdic.$O prog: main.$O $LIB $LD -o $target $prereq $LIB(%):N: % $LIB: ${OBJS:%=$LIB(%)} ar rv $LIB $OBJS

builds a program by loading it with a library.

8 Evaluation algorithm

For each target of interest, mk uses the rules in a mkfile to build a data structure called a dependency graph. The nodes of the graph represent targets and prerequisites; a directed arc from one node to another indicates that the file associated with the first node depends on the file associated with the second. When the mkfile has been completely read, the graph is analyzed. In the first step, implied dependencies are resolved by computing the transitive closure of the graph. This calculation extends the graph to include all targets that are potentially derivable from the rules in the mkfile. Next the graph is checked for cycles; make accepts cyclic dependencies, but mk does not allow them. Subsequent steps prune subgraphs that are irrelevant for producing the desired target and verify that there is only one way to build it. The recipes associated with the nodes on the longest path between the target and an out of date prerequisite are then executed in reverse order.

The transitive closure calculation is sensitive to metarules; the patterns often select many potential targets and cause the graph to grow rapidly. Fortunately, dependencies associated with the desired target usually form a small part of the graph, so, after pruning, analysis is tractable. For example, the rules

%: x.% recipe1 x.%: %.k recipe2 %.k: %.f recipe3

produce a graph with four nodes for each file in the current directory. If the desired target is foo, mk detects the dependency between it and the original file foo.f through intermediate dependencies on foo.k and x.foo. Nodes associated with other files are deleted during pruning because they are irrelevant to the production of foo.

Mk avoids infinite cycles by evaluating each metarule once. Thus, the rule

%: %.z cp $prereq $prereq.z

copies the prerequisite file once.

9 Conventions for evaluating rules

There must be only one way to build each target. However, during evaluation metarule patterns often select potential targets that conflict with the targets of other rules. Mk uses several conventions to resolve ambiguities and to select the proper dependencies.

When a target selects more than one rule, mk chooses a regular rule over a metarule. For example, the mkfile

</$objtype/mkfile FILES=f1.$O f2.$O f3.$O prog: $FILES $LD -o $target $prereq %.$O: %.c $CC $CFLAGS $stem.c f2.$O: f2.c $CC f2.c

contains two rules that could build f2.$O. Mk selects the last rule because its target, f2.$O, is explicitly specified, while the %.$O rule is a metarule. In effect, the explicit rule for f2.$O overrides the general rule for building object files from C source files.

When a rule has a target and prerequisites but no recipe, those prerequisites are added to all other rules with recipes that have the same target. All prerequisites, regardless of where they were specified, are exported to the recipe in variable $prereq. For example, in

</$objtype/mkfile FILES=f1.$O f2.$O f3.$O prog: $FILES $LD -o $target $prereq %.$O: hdr.h %.$O: %.c $CC $CFLAGS $stem.c

the second rule adds hdr.h as a prerequisite of the compile metarule; an object file produced from a C source file depends on hdr.h as well as the source file. Notice that the recipe of the compile rule uses $stem.c instead of $prereq because the latter specification would attempt to compile hdr.h.

When a target is virtual and there is no other rule with the same target, mk evaluates each prerequisite. For example, adding the rule

all:V: prog

to the preceding example builds the executable when either prog or all is the specified target. In effect, the all target is an alias for prog.

When two rules have identical rule headers and both have recipes, the later rule replaces the former one. For example, if a file named mkrules contains

$O.out: $OFILES $LD $LFLAGS $OFILES %.$O: %.c $CC $CFLAGS $stem.c

the mkfile

OFILES=f1.$O f2.$O f3.$O <mkrules $O.out: $OFILES $LD $LFLAGS -l $OFILES -lbio -lc

overrides the general loader rule with a special rule using a non-standard library search sequence. A rule is neutralized by overriding it with a rule with a null recipe:

<mkrules $O.out:Q: $OFILES ;

The Q attribute suppresses the printing of the semicolon.

When a rule has no prerequisites, the recipe is executed only when the target does not exist. For example,

marker: touch $target

defines a rule to manage a marker file. If the file exists, it is considered up to date regardless of its modification time. When a virtual target has no prerequisites the recipe is always executed. The clean rule is of this type:

clean:V: rm -f [$OS].out *.[$OS]

When a rule without prerequisites has multiple targets, the extra targets are aliases for the rule. For example, in

clean tidy nuke:V: rm -f [$OS].out *.[$OS]

the rule can be invoked by any of three names. The first rule in a mkfile is handled specially: when mk is invoked without a command line target all targets of the first non-metarule are built. If that rule has multiple targets, the recipe is executed once for each target; normally, the recipe of a rule with multiple targets is only executed once.

A rule applies to a target only when its prerequisites exist or can be derived. More than one rule may have the same target as long as only one rule with a recipe remains applicable after the dependency evaluation completes. For example, consider a program built from C and assembler source files. Two rules produce object files:

%.$O: %.c $CC $CFLAGS $stem.c %.$O: %.s $AS $AFLAGS $stem.s

As long as there are not two source files with names like foo.c and foo.s, mk can unambiguously select the proper rule. If both files exist, the rules are ambiguous and mk exits with an error message.

In Plan 9, many programs consist of portable code stored in one directory and architecture-specific source stored in another. For example, the mkfile

</$objtype/mkfile FILES=f1.$O f2.$O f3.$O f3.$O prog: $FILES $LD -o $target $prereq %.$O: %.$c $CC $CFLAGS $stem.c %.$O: ../port/%.c $CC $CFLAGS ../port/$stem.c

builds the program named prog using portable code in directory ../port and architecture-specific code in the current directory. As long as the names of the C source files in ../port do not conflict with the names of files in the current directory, mk selects the appropriate rule to build the object file. If like-named files exist in both directories, the specification is ambiguous and an explicit target must be specified to resolve the ambiguity. For example, adding the rule

f2.$O: f2.c $CC $CFLAGS $f2.c

to the previous mkfile uses the architecture-specific version of f2.c instead of the portable one. Here, the explicit rule unambiguously documents which of the like-named source files is used to build the program.

Mk's heuristics can produce unintended results when rules are not carefully specified. For example, the rules that build object files from C or assembler source files

%.$O: %.c $CC $CFLAGS $stem.c %.$O: %.s $AS $AFLAGS $stem.s

illustrate a subtle pratfall. Adding a header file dependency to the compile rule

%.$O: %.c hdr.h $CC $CFLAGS $stem.c

produces the error message

don't know how to make 'file.c'

when file.s is an assembler source file. This occurs because file.s satisfies the assemble rule and hdr.h satisfies the compile rule, so either rule can potentially produce the target. When a prerequisite exists or can be derived, all other prerequisites in that rule header must exist or be derivable; here, the existence of hdr.h forces the evaluation of a C source file. Specifying the dependencies in different rules avoids this interpretation:

%.$O: hdr.h %.$O: %.c $CC $CFLAGS $stem.c

Although hdr.h is an additional prerequisite of the compile rule, the two rules are evaluated independently and the existence of the C source file is not linked to the existence of the header file. However, this specification describes a different dependency. Originally, only object files derived from C files depended on hdr.h; now all object files, including those built from assembler source, depend on the header file.

Metarule patterns should be as restrictive as possible to prevent conflicts with other rules. Consider the mkfile

</$objtype/mkfile BIN=/$objtype/bin PROG=foo install:V: $BIN/$PROG %: %.c $CC $stem.c $LD -o $target $stem.$O $BIN/%: % mv $stem $target

The first target builds an executable in the local directory; the second installs it in the directory of executables for the architecture. Invoking mk with the install target produces:

mk: ambiguous recipes for /mips/bin/foo: /mips/bin/foo <-(mkfile:8)- /mips/bin/foo.c <-(mkfile:12)- foo.c /mips/bin/foo <-(mkfile:12)- foo <-(mkfile:8)- foo.c

The prerequisite of the install rule, $BIN/$PROG, matches both metarules because the % pattern matches everything. The & pattern restricts the compile rule to files in the current directory and avoids the conflict:

&: &.c $CC $stem.c $LD -o $target $stem.$O

10 Missing intermediates

Mk does not build a missing intermediate file if a target is up to date with the prerequisites of the intermediate. For example, when an executable is up to date with its source file, mk does not compile the source to create a missing object file. The evaluation only applies when a target is considered up to date by pretending that the intermediate exists. Thus, it does not apply when the intermediate is a command line target or when it has no prerequisites.

This capability is useful for maintaining archives. We can modify the archive update recipe to remove object files after they are archived:

$LIB(%):N: % $LIB: ${OBJS:%=$LIB(%)} names=`{membername $newprereq} ar rv $LIB $names rm -f $names

A subsequent mk does not remake the object files as long as the members of the archive remain up to date with the source files. The -i command line option overrides this behavior and causes all intermediates to be built.

11 Alternative out-of-date determination

Sometimes the modification time is not useful for deciding when a target and prerequisite are out of date. The P attribute replaces the default mechanism with the result of a command. The command immediately follows the attribute and is repeatedly executed with each target and each prerequisite as its arguments; if its exit status is non-zero, they are considered out of date and the recipe is executed. Consider the mkfile

foo.ref:Pcmp -s: foo cp $prereq $target

The command

cmp -s foo.ref foo

is executed and if foo.ref differs from foo, the latter file is copied to the former.

12 Parallel processing

When possible, mk executes recipes in parallel. The variable $NPROC specifies the maximum number of simultaneously executing recipes. Normally it is imported from the environment, where the system has set it to the number of available processors. It can be decreased by assigning a new value and can be set to 1 to force single-threaded recipe execution. This is necessary when several targets access a common resource such as a status file or data base. When there is no dependency between targets, mk assumes the recipes can be executed concurrently. Normally, this allows multiple prerequisites to be built simultaneously; for example, the object file prerequisites of a load rule can be produced by compiling the source files in parallel. Mk does not define the order of execution of independent recipes. When the prerequisites of a rule are not independent, the dependencies between them should be specified in a rule or the mkfile should be single-threaded. For example, the archive update rules

$LIB(%):N: % $LIB: ${OBJS:%=$LIB(%)} ar rv $LIB `{membername $newprereq}

compile source files in parallel but update all members of the archive at once. It is a mistake to merge the two rules

$LIB(%): % ar rv $LIB $stem

because an ar command is executed for every member of the library. Not only is this inefficient, but the archive is updated in parallel, making interference likely.

The $nproc environment variable contains a number associated with the processor executing a recipe. It can be used to create unique names when the recipe may be executing simultaneously on several processors. Other maintenance tools provide mechanisms to control recipe scheduling explicitly [Cmel86], but mk's general rules are sufficient for all but the most unusual cases.

13 Deleting target files on errors

The D attribute causes mk to remove the target file when a recipe terminates prematurely. The error message describing the termination condition warns of the deletion. A partially built file is doubly dangerous: it is not only wrong, but is also considered to be up to date so a subsequent mk will not rebuild it. For example,

pic.out:D: mk.ms pic $prereq | tbl | troff -ms > $target

produces the message

mk: pic mk.ms | ... : exit status=rc 685: deleting 'pic.out'

if any program in the recipe exits with an error status.

14 Unspecified dependencies

The -w command line flag forces the files following the flag to be treated as if they were just modified. We can use this flag with a command that selects files to force a build based on the selection criterion. For example, if the declaration of a global variable named var is changed in a header file, all source files that reference it can be rebuilt with the command

$ mk -w`{grep -l var *.[cyl]}

15 Conclusion

There are many programs related to make, each choosing a different balance between specialization and generality. Mk emphasizes generality but allows customization through its pattern specifications and include facilities.

Plan 9 presents a difficult maintenance environment with its heterogeneous architectures and languages. Mk's flexible specification language and simple interaction with rc work well in this environment. As a result, Plan 9 relies on mk to automate almost all maintenance. Tasks as diverse as updating the network data base, producing the manual, or building a release are expressed as mk procedures.

16 References

[Cmel86] R. F. Cmelik, ``Concurrent Make: A Distributed Program in Concurrent C'', AT&T Bell Laboratories Technical Report, 1986.

[Feld79] S. I. Feldman, ``Make a program for maintaining computer programs'', Software Practice & Experience , 1979 Vol 9 #4, pp. 255-266.

[Flan95] Bob Flandrena, ``Plan 9 Mkfiles'', this volume.

[Hume87] A. G. Hume, ``Mk: A Successor to Make'', USENIX Summer Conf. Proc., Phoenix, Az.

17 Appendix: Differences between `make` and `mk`

The differences between mk and make are:

*: Make builds targets when it needs them, allowing systematic use of side effects. Mk constructs the entire dependency graph before building any target.
*: Make supports suffix rules and % metarules. Mk supports % and regular expression metarules. (Older versions of make support only suffix rules.)
*: Mk performs transitive closure on metarules, make does not.
*: Make supports cyclic dependencies, mk does not.
*: Make evaluates recipes one line at a time, replacing variables by their values and executing some commands internally. Mk passes the entire recipe to the shell without interpretation or internal execution.
*: Make supports parallel execution of single-line recipes when building the prerequisites for specified targets. Mk supports parallel execution of all recipes. (Older versions of make did not support parallel execution.)
*: Make uses special targets (beginning with a period) to indicate special processing. Mk uses attributes to modify rule evaluation.
*: Mk supports virtual targets that are independent of the file system.
*: Mk allows non-standard out-of-date determination, make does not.

It is usually easy to convert a makefile to or from an equivalent mkfile.