// file      : doc/testscript.cli
// copyright : Copyright (c) 2014-2017 Code Synthesis Ltd
// license   : MIT; see accompanying LICENSE file

"\name=build2-testscript-language"
"\subject=Testscript language"
"\title=Testscript Language"

// NOTES
//
// - Maximum <pre> line is 70 characters.
//

"
\h0#preface|Preface|

This document describes the \c{build2} Testscript language. It starts with a
discussion of the motivations behind a separate domain-specific language for
running tests and then introduces a number of Testscript concepts with
examples. The remainder of the document provides a more formal specification
of the language, including its integration into the build system, parsing and
execution model, lexical structure, as well as grammar and semantics. The
final chapter describes the recommended Testscript style as used in the
\c{build2} project.

In this document we use the term \i{Testscript} (capitalized) to refer to the
Testscript language. Just \i{testscript} means some code written in this
language. For example: \"We can pass addition information to testscripts using
target-specific variables.\" Finally, \c{testscript} refers to the file name.

We also use the equivalent distinction between \i{Buildfile} (language),
\i{buildfile} (code), and \c{buildfile} (file).

\h1#intro|Introduction|

The \c{build2} \c{test} module provides the ability to run an executable
target as a test, including passing options and arguments, providing
\c{stdin} input, as well as comparing the \c{stdout} output to the expected
result. For example:

\
exe{hello}: test.options   = --greeting 'Hi'
exe{hello}: test.arguments = -               # Read names from stdin.
exe{hello}: test.input     = names.txt
exe{hello}: test.output    = greetings.txt
\

This works well for simple, single-run tests. If, however, your testing
requires multiple runs with varying input and/or analyzing output,
traditionally, you would resort to using a scripting language, for instance
Bash or Python. This, however, has a number of drawbacks. Firstly, this
approach is usually not portable (there is no Bash or Python on Windows \i{out
of the box}). It is also hard to write concise tests in a general-purpose
scripting language. The result is often a test suite that has grown
incomprehensible with everyone dreading adding new tests. Secondly, it is hard
to run such tests in parallel without a major effort, for example, by having a
separate script for each test and implementing some kind of a test harness.

Testscript is a domain-specific language for running tests. It vaguely
resembles Bash and is optimized for concise test description and fast
execution by focusing on the following functionality:

\ul|

\li|Supplying input via command line and \c{stdin}.|

\li|Comparing to expected exit status.|

\li|Comparing to expected output for \c{stdout}/\c{stderr}, including
    using regex.|

\li|Setup/teardown commands and automatic file/directory cleanups.|

\li|Simple (single-command) and compound (multi-command) tests.|

\li|Test groups with common setup/teardown.|

\li|Test isolation for parallel execution.|

\li|Portable POSIX-like builtins subset.|

\li|Test documentation.||

Note that Testscript is a \i{test runner}, not a testing framework for a
particular programming language. It does not concern itself with how the test
executables themselves are implemented. As a result, it is mostly geared
towards functional testing but can also be used for unit testing if
external input/output is required. Testscript is an extension of the
\c{build2} build system and is implemented by its \c{test} module.

As a quick introduction to Testscript's capabilities, let's test a \"Hello,
World\" program. For a simple implementation the corresponding \c{buildfile}
might look like this:

\
exe{hello}: cxx{hello}
\

We also assume that the project's \c{bootstrap.build} loads the \c{test}
module which implements the execution of testscripts.

To start, we create an empty file called \c{testscript}. To indicate that a
testscript file tests a specific target we simply list it as a target's
prerequisite, for example:

\
exe{hello}: cxx{hello} test{testscript}
\

Let's assume our \c{hello} program expects us to pass the name to greet on the
command line. And if we don't pass anything, it prints an error following by
usage and terminates with a non-zero exit status. We can test this failure
case by adding the following line to the \c{testscript} file:

\
$* 2>- != 0
\

While it sure is concise, it may look cryptic without some explanation. When
the \c{test} module runs tests, it (by default) passes to each testscript the
target path of which this testscript is a prerequisite. So in our case the
testscript will receive the path to our \c{hello} executable. The buildfile
can also pass along additional options and arguments. Inside the testscript,
all of this (target path, options, and arguments) are bound to the \c{$*}
variable. So in our case, if we expand the above line, it would be something
like this:

\
/tmp/hello/hello 2>- != 0
\

Or, if we are on Windows, something like this:

\
C:\projects\hello\hello.exe 2>- != 0
\

The \c{2>-} redirect is the Testscript equivalent of \c{2>/dev/null} that is
both portable and more concise (\c{2} here is the \c{stderr} file
descriptor). If we don't specify it and our program prints anything to
\c{stderr}, then the test will fail (unexpected output).

The remainder of the command (\c{!= 0}) is the exit status check. If we don't
specify it, then the test is expected to exit with zero status (which is
equivalent to specifying \c{== 0}).

If we run our test, it will pass provided our program behaves as expected.
One thing our test doesn't verify, however, is the diagnostics that gets
printed to \c{stderr} (remember, we ignored it with \c{2>-}). Let's fix that
assuming this is the code that prints it:

\
cerr << \"error: missing name\" << endl
     << \"usage: \" << argv[0] << \" <name>\" << endl;
\

In testscripts you can compare output to the expected result for both
\c{stdout} and \c{stderr}. We can supply the expected result as either
\i{here-string} or \i{here-document}, both which can be either literal or
regex. The here-string approach works best for short, single-line output and
we will use it for another test in a minute. For this test let's use
here-document since the expected diagnostics has two lines:

\
$* 2>>EOE != 0
error: missing name
usage: hello <name>
EOE
\

Let's decrypt this: the \c{2>>EOE} is a here-document redirect with \c{EOE}
(stands for End-Of-Error) being the string we chose to mark the end of
here-document. Next comes the here-document fragment followed by the end
marker.

Now, when executing this test, the \c{test} module will check two things: it
will compare the \c{stderr} output to the expected result using the \c{diff}
tool and it will make sure the test exits with a non-zero status. Let's give
it a go:

\
$ b test
testscript:1:1: error: stderr doesn't match expected output
  info: produced stderr: test-hello/1/stderr
  info: expected stderr: test-hello/1/stderr.orig
  info: stderr diff (test-hello/1/stderr.diff):

--- test-hello/1/stderr.orig
+++ test-hello/1/stderr
@@ -1,2 +1,2 @@
 error: missing name
-usage: hello <name>
+usage: /tmp/hello/hello <name>
\

While not what we expected, at least the problem is clear: the program name
varies at runtime so we cannot just hardcode \c{hello} in our expected output.
How do we solve this? The best fix would be to use the actual path to the
target; after all, we know it's the first element in \c{$*}:

\
$* 2>>\"EOE\" != 0
error: missing name
usage: $0 <name>
EOE
\

You can probably guess what \c{$0} expands to. But did you notice another
change? Yes, those double quotes in \c{2>>\"EOE\"}. Here is what's going on:
similar to Bash, single-quoted strings (\c{'foo'}) are taken literally while
double-quoted ones (\c{\"foo\"}) have variable expansion, escaping, etc. This
semantics is extended to here-documents in a curious way: if the end marker is
single-quoted then the here-document lines are taken literally and if it is
double-quoted, then there can be variable expansions, etc.  An unquoted end
marker is treated as single-quoted (note that this is unlike Bash where
here-documents always have variable expansions).

This example illustrated a fairly common testing problem: output variability.
In this case we could fix it perfectly since we could easily calculate the
varying parts exactly. But often figuring out the varying part is difficult of
outright impossible. A good example would be a system error message based on
the \c{errno} code, such as file not being found. Different C runtimes can
phrase the message slightly differently or it can be localized. Worse, it can
be a slightly different error code, for example \c{ENOENT} vs \c{ENOTDIR}.

To handle output variability, Testscript allows us to specify the expected
output as regular expressions. For example, this is an alternative fix to our
usage problem that simply ignores the program name:

\
$* 2>>~/EOE/ != 0
error: missing name
/usage: .+ <name>/
EOE
\

Let's explain what's going here: to use a regex here-string or here-document
we add the \c{~} redirect modifier. In this case the here-document end marker
must start and end with the regex introducer character of your choice (\c{/}
in our case). Any line inside the here-document fragment that begins with this
introducer is then treated as a regular expression rather than a literal (see
\l{#syntax-regex Output Regex} for details).

While this was a fairly deep rabbit hole for a first example, it is a good
illustration of how quickly things get complicated when testing real-world
software.

Now that we have tested the failure case, let's test the normal functionality.
While we could have used here-document, in this case here-string will be more
concise:

\
$* 'World' >'Hello, World!'
\

Nothing new here. It's also a good idea to document our tests. Testscript
has a formalized test description that can capture the test \i{id},
\i{summary}, and \i{details}. All three components are optional and how
thoroughly you document your tests is up to you.

The description lines precede the test command. They start with a colon
(\c{:}), and have the following layout:

\
: <id>
: <summary>
:
: <details>
: ...
\

The recommended format for \c{<id>} is \c{<keyword>-<keyword>...} with at
least two keywords. The id is used in diagnostics, to name the test working
directory, as well as to run individual tests. The recommended style for
\c{<summary>} is that of the \c{git(1)} commit summary. The detailed
description is free-form. Here are some examples (\c{#} starts a comment):

\
# Only id.
#
: missing-name
$* 2>>\"EOE\" != 0
...

# Only summary.
#
: Test handling of missing name
...

# Both id and summary.
#
: missing-name
: Test handling of missing name
...

# All three: id, summary, and a detailed description.
#
: missing-name
: Test handling of missing name
:
: This test makes sure the program detects that the name to greet
: was not specified on the command line and both prints usage and
: exits with non-zero status.
...
\

The recommended way to come up with an id is to distill the summary to its
essential keywords by removing generic words like \"test\", \"handle\", and so
on. If you do this, then both the id and summary will convey essentially the
same information. As a result, to keep things concise, you may choose to drop
the summary and only have the id. This is what we often do in \c{build2}.
Note that if the id is not provided, then it will be automatically derived
from the line number in testscript.

Either the id or summary (but not both) can alternatively be specified inline
in the test command after a colon (\c{:}), for example:

\
$* 'World' >'Hello, World!' : command-name
\

Similar to handling output, Testscript provides a convenient way to supply
input to the test's \c{stdin}. Let's say our \c{hello} program recognizes the
special \c{-} name as an instruction to read the names from \c{stdin}. This
is how we could test this functionality:

\
$* - <<EOI >>EOO : stdin-names
Jane
John
EOI
Hello, Jane!
Hello, John!
EOO
\

As you might suspect, we can also use here-string to supply \c{stdin}, for
example:

\
$* - <'World' >'Hello, World!' : stdin-name
\

Let's say our \c{hello} program has a configuration file that captures custom
name-to-greeting mappings. A path to this file can be passed as a second
command line argument. To test this functionality we first need to create a
sample configuration file. We do these non-test actions with \i{setup} and
\i{teardown} commands, for example:

\
+cat <<EOI >>>hello.conf;
John = Howdy
Jane = Good day
EOI
$* 'Jane' hello.conf >'Good day, Jane!' : config-greet
\

The setup commands start with the plus sign (\c{+}) while teardown \- with
minus (\c{-}). Notice also the semicolon (\c{;}) at the end of the setup
command: it indicates that the following command is part of the same test \-
what we call a multi-command or \i{compound} test.

Other than that it should all look familiar. You may be wondering why we don't
have a teardown command that removes \c{hello.conf}? It is not necessary
because this file will be automatically registered for cleanup that
happens at the end of the test. We can also register our own files and
directories for automatic cleanup. For example, if the \c{hello} program
created the \c{hello.log} file on unsuccessful runs, then here is how we could
have cleaned it up:

\
$* ... &hello.log != 0
\

What if we wanted to run two tests for this configuration file functionality?
For example, we may want to test the custom greeting as above but also make
sure the default greeting is not affected. One way to do this would be to
repeat the setup command in each test. But there is a better way: testscripts
can define test groups. For example:

\
: config
{
  conf = $~/hello.conf

  +cat <<EOI >>>$conf
  John = Howdy
  Jane = Good day
  EOI

  $* 'John' $conf >'Howdy, John!' : custom-greet
  $* 'Jack' $conf >'Hello, Jack!' : default-greet
}
\

A test group is a scope that contains several test/setup/teardown commands.
Variables set inside a scope (like our \c{conf}) are only in effect until the
end of the scope. Plus, setup and teardown commands that are not part of any
test (notice the lack of \c{;} after \c{+cat}) are associated with the scope;
their automatic cleanup only happens at the end of the scope (so our
\c{hello.conf} will only be removed after all the tests in the group have
completed). Note also that a scope can have a description. In particular,
assigning a test group an id allows us to run tests only from this specific
group.

Other than that the two other things we need to discuss in this example are
\c{$~} and \c{cat}. The \c{$~} variable is easy: it stands for the test/group
working directory.

But what is \c{cat} exactly? While most POSIX systems will have a program with
this name, there is no such thing in vanilla Windows. To help with this
Testscript provides a subset (both in terms of the number and supported
features) of POSIX utilities, such as, \c{echo}, \c{touch}, \c{cat},
\c{mkdir}, \c{rm}, and so on (see \l{#builtins Builtins} for details).

Besides explicit group scopes each test is automatically placed in its own
implicit test scope. However, we can make the test scope explicit, for
example, for better visual separation of complex tests:

\
: config-greet
{
  conf = hello.conf

  +cat <'Jane = Good day' >>>$conf;
  $* 'Jane' $conf >'Good day, Jane!'
}
\

We can conditionally exclude sections of testscripts using \c{if-else}
branching. This can be done both at the scope level to exclude test or
group scopes as well as at the command level to exclude individual
commands or variable assignments. Let's start with a scope example by
providing a Windows-specific implementation of a test:

\
: config-empty
:
if ($cxx.target.class != windows)
{
  $* 'Jane' /dev/null >'Hello, Jane!'
}
else
{
  $* 'Jane' nul >'Hello, Jane!'
}
\

Note the \c{if-else} chain is treated as variants of the same test thus
the single description at the beginning.

Let's now see an example of command-level \c{if-else} by reimplementing
the above as a single test with some branching and without using the
\c{nul} device on Windows (notice the semicolon after \c{end}):

\
: config-empty
:
if ($cxx.target.class != windows)
  conf = /dev/null
else
  conf = empty
  touch $conf
end;
$* 'Jane' $conf >'Hello, Jane!'
\

You may have noticed that in the above examples we referenced the
\c{cxx.target.class} variable as if we were in a buildfile. We could do that
because the testscript variable lookup continues in the buildfile starting
from the testscript target and continuing with the standard buildfile variable
lookup. In particular, this means we can pass arbitrary information to
testscripts using target-specific variables. For example, this how we can move
the above platform test to \c{buildfile}:

\
# buildfile

exe{hello}: cxx{hello} test{testscript}

test{*}: windows = ($cxx.target.class == windows)
\

\
# testscript

if! $windows
  conf = /dev/null
else
  ...
\

Note also that in cases when you simply need to conditionally pick a value for
a variable, the \c{build2} evaluation context will often be more concise than
\c{if-else}. For example:

\
: config-empty
:
conf = ($windows ? nul : /dev/null);
$* 'Jane' $conf >'Hello, Jane!'
\

Similar to Bash, test commands can be chained with pipes (\c{|}) and combined
with logical operators (\c{||} and \c{&&}). Let's say our \c{hello} program
provides the \c{-o} option to write the result to a file instead of
\c{stdout}. Here is how we could test it:

\
$* -o hello.out - <<EOI &hello.out && cat hello.out >>EOO
John
Jane
EOI
Hello, John!
Hello, Jane!
EOO
\

Similarly, if it has the \c{-r} option to reverse the greetings back to
their names (as every \c{hello} program should), then we could write a
test like this:

\
$* - <<EOI | $* -d - >>EOO
John
Jane
EOI
John
Jane
EOO
\

To conclude, let's put all our (sensible) tests together so that we can have a
complete picture:

\
$* 'World' >'Hello, World!' : command-name

$* 'Jonh' 'Jane' >EOO       : command-names
Hello, Jane!
Hello, John!
EOO

$* - <<EOI >>EOO            : stdin-names
Jane
John
EOI
Hello, Jane!
Hello, John!
EOO

: config
{
  conf = $~/hello.conf

  +cat <<EOI >>>$conf
  John = Howdy
  Jane = Good day
  EOI

  $* 'John' $conf >'Howdy, John!' : custom-greet
  $* 'Jack' $conf >'Hello, Jack!' : default-greet
}

$* 2>>\"EOE\" != 0            : missing-name
error: missing name
usage: $0 <name>
EOE
\

Testscript isolates tests from each other by running each test in its own
temporary working directory under \c{out_base}. For the above \c{testscript}
the working directory structure will be as follows:

\
$out_base/
└── test-hello/
      ├── command-name/
      ├── command-names/
      ├── stdin-names/
      ├── config/
      │    ├── hello.conf
      │    ├── custom-greet/
      │    └── default-greet/
      └── missing-name/
\

If all the test succeed then this working directory structure is automatically
removed. In case of a failure, however, it is left behind in case you need to
examine the output of the failed tests. It will also be automatically cleaned
on the subsequent run, before executing any tests.

The execution of tests happens in parallel. In the above case Testscript will
start running all the top-level tests as well as the \c{config} group
immediately.  Inside \c{config}, once the setup command (\c{cat}) is
completed, the two inner tests are executed in parallel as well. Refer to
\l{#model Model and Execution} for details on the test directory structure and
execution.


\h1#integration|Build System Integration|

The integration of testscripts into buildfiles is done using the standard
\i{target-prerequisite} mechanism. In this sense, a testscript is a
prerequisite that describes how to test the target similar to how, for
example, the \c{INSTALL} file describes how to install it. For example:

\
exe{hello}: test{testscript} doc{INSTALL README}
\

By convention the testscript file should be either called \c{testscript} if
you only have one or have the \c{.test} extension, for example,
\c{basics.test}. The \c{test} module registers the \c{test{\}} target type
for testscript files.

A testscript prerequisite can be specified for any target. For example, if
our directory contains a bunch of shell scripts that we want to test together,
then it makes sense to specify the testscript prerequisite for the directory
target:

\
./: test{basics}
\

During variable lookup if a variable is not found in one of the testscript
scopes, then its search continues in the buildfile starting with the
target-specific variables of the target being tested (e.g., \c{exe{hello\}};
called \i{test target}), then target-specific variables of the testscript
target (e.g., \c{test{basics\}}; called \i{script target}), and then
continuing with the scopes starting from the one containing the testscript
target. This means a testscript can \"see\" all the existing buildfile
variables plus we can use target-specific variables to pass additional
information to testscrips, for example:

\
# basics.test

if ($cxx.target.class == windows)
  test.arguments += $foo
end

if $windows
  test.arguments += $bar
end
\

\
# buildfile

exe{hello}: test{basics}

# All testscripts in this scope.
#
test{*}: windows = ($cxx.target.class == windows)

# All testscripts for target exe{hello}.
#
exe{hello}: bar = BAR

# Only basics.test.
#
test{basics}@./: foo = FOO
\

Additionally, a number of \c{test.*} variables are used by convention to pass
commonly required information to testscripts.

Unless set manually as a test or script target-specific variable, the \c{test}
variable is automatically set to the target path being tested. For example,
given this \c{buildfile}:

\
exe{hello}: test{testscript}
\

The value of \c{test} inside the testscript will be the absolute path to the
\c{hello} executable.

If the \c{test} variable is set manually to a name of a target, then it is
automatically converted to the target path. This can be useful when testing
a program that is built in another subdirectory of a project. For example,
our \c{hello} may reside in the \c{hello/} subdirectory while we may want
to keep the tests in \c{tests/}:

\
hello/
├── hello/
│   └── hello*
└── tests/
    ├── buildfile
    └── testscript
\

This is how we can implement \c{tests/buildfile} for this setup:

\
hello = ../hello/exe{hello}

./: $hello test{testscript}
./: test = $hello

include ../hello/
\

The other special \c{test.} variables are \c{test.options},
\c{test.arguments}, \c{test.redirects}, and \c{test.cleanups}. You can use
them to pass additional command line options, arguments, redirects, and
cleanups to your test scripts and together with \c{test} they form the test
target command line which, for conciseness, is bound to the following
read-only variable aliases:

\
$* - $test $test.options $test.arguments $test.redirects $test.cleanups
$0 - $test
$N - (N-1)-th element in the {$test.options $test.arguments} array
\

Note that these aliases are read-only; if you need to modify any of these
values from within testscripts, then you should use the original variable
names, for example:

\
test.options += --foo

$* bar # Includes --foo.
\

Note also that these \c{test.} variables only establish a convention. You
could also put everything into, say \c{test.arguments}, and it will still work
as expected.

Finally, the \c{test.target} variable can be used to specify the test target
platform when cross-testing (for example, when running Windows test on Linux
under Wine). Normally, you would set it in your \c{build/root.build} to the
cross-compilation target of your toolchain, for example:

\
# root.build
#

using cxx                 # Load the C++ module.
test.target = $cxx.target # Set test target to the C++ compiler target.
\

If this variable is not set explicitly, then it default to \c{build.host}
(which is the platform on which the build system is running) and only
native testing will be supported.

All the testscripts for a particular test target are executed in a
subdirectory of \c{out_base} (or, more precisely, in subdirectories of this
subdirectory, as discussed below). If the test target is a directory, then the
subdirectory is called \c{test}. Otherwise, it is the name of the target
prefixed with\c{test-}. For example:

\
./:         test{foo}   # $out_base/test/
exe{hello}: test{bar}   # $out_base/test-hello/
\


\h1#model|Model and Execution|

A testscript file is a set of nested scopes. A scope is either a group scope
or a test scope. Group scopes contain nested group and test scopes. Test
scopes only contain test commands.

Group scopes are used to organize related tests with potentially shared
variables as well as setup and teardown commands. Explicit test scopes are
normally used for better visual separation of complex tests.

The top level scope is always an implicit group scope corresponding to the
entire script file. If there is no explicit scope for a test, one is
established implicitly. As a result, a testscript file always starts with a
group scope which then contains other group scopes and/or test scopes,
recursively.

A scope (both group and test) has an \i{id}. If not specified explicitly, it
is automatically derived from the group/test location in the testscript file
(see \l{#syntax-description Description} for details). The id of the
implicit outermost scope is the script file name without the \c{.test}
extension. If the file name is \c{testscript}, then the id is empty.

Based on the ids each nested group and test has an \i{id path} that uniquely
identifies it. It starts with the id of the implied outermost group (unless
empty), may include a number of intermediate group ids, and ends with the
final test or group id. The ids in the path are separated with a forward slash
(\c{/}). Note that this also happens to be the relative filesystem path to the
temporary directory where the test is executed (as discussed below). Inside a
scope its id path is available via the special \c{$@} variable (read-only).

As an example, consider the following testscript file which we assume is
called \c{basics.test}:

\
test: test

: group
{
  test1

  : test2
  {
    test2a;
    test2b
  }
}
\

Below is its version annotated with id paths that also shows all the implicit
scopes:

\
# basics
{
  # basics/test
  {
    test
  }

  # basics/group
  {
    # basics/group/5
    {
      test1
    }

    # basics/group/test2
    {
      test2a;
      test2b
    }
  }
}
\

A scope establishes a nested variable context. A variable set within a scope
will only have effect until the end of this scope. Variable lookup is
performed starting from the scope of the expansion, continuing with the outer
testscript scopes, and then continuing in the buildfile.

A scope also establishes a cleanup context. All cleanups registered in a
certain scope are performed at the end of that scope's execution.

Prior to executing a scope, a nested temporary directory is created with the
scope id as its name. This directory then becomes the scope's working
directory. After executing the scope (and after performing cleanups) this
temporary directory is automatically removed provided that it is empty. If it
is not empty, the test is considered failed (unexpected output).  Inside a
scope its working directory is available via the special \c{$~} variable
(read-only).

As an example, consider the following version of \c{basics.test}. We also
assume that its test target is a directory.

\
: group
{
  foo = FOO
  bar = BAR

  +setup &out-setup

  : test1
  {
    bar = BAZ
    test1 $foo $bar
  }

  test2 $bar: test2
}

test $foo &out-test
\

Below is its annotated version:

\
{                         # $~ = $out_base/test/basics/
  {                       # $~ = .../test/basics/group/
    foo = FOO
    bar = BAR

    +setup &out-setup

    {                     # $~ = .../basics/group/test1/
      bar = BAZ
      test1 $foo $bar     # test1 FOO BAZ
    }

    {                     # $~ = .../basics/group/test2/
      test2 $bar          # test2 BAR
    }
  }                       # Remove out-setup.

  {                       # $~ = .../test/basics/17/
    test $foo &out-test   # test
  }                       # Remove out-test.
}
\

A test should normally create any files or directories in its working
directory to ensure test isolation. A test can, however, access (but normally
not modify) files created by an outer group's setup commands. Because of this
nested directory structure this can be done using \c{../}-based relative
paths, for example:

\
{
  +setup >>>test.conf

  test1 ../test.conf
  test2 ../test.conf
}
\

Alternative, one can use an absolute path:

\
{
  conf = $~/test.conf
  +setup >>>$conf

  test1 $conf
  test2 $conf
}
\

Inside the scope working directory names that start with \c{stdin},
\c{stdout}, \c{stderr}, as well as, \c{cmd-} are reserved.

To executing a test scope its commands (including variable assignments) are
are executed sequentially and in order specified. If any of the commands
fails, nor further commands are executed and the test fails.

Executing a group scope starts with performing its setup commands (including
variable assignments) sequentially and in order specified. If any of them
fail, the group execution is terminated.

After completing the setup, inner scopes (both group and test) are
executed. Because scopes are isolated and test should not depend on each
other, the execution can be performed in parallel.

After executing the inner scopes, if all of them succeeded, the teardown
commands are executed sequentially and in order specified. Again, if any of
them fail, the group execution is terminated.

As an example, consider the following version of \c{basics.test}:

\
test0

: group
{
  +setup1
  +setup2

  test1
  test2
  test3

  -teardown2
  -teardown1
}
\

At the top level, both \c{test0} and \c{group} can start executing in
parallel. Inside \c{group}, first the two setup command are executed
sequentially. Once the setup is completed, \c{test1}, \c{test2}, \c{test3}
can all be executed in parallel (along with \c{test0} which may still be
running). Once the three inner tests complete successfully, the \c{group}'s
teardown command are executed sequentially. At the top level, the script
is completed only when both \c{test0} and \c{group} complete.

The following annotated version illustrated a possible thread scheduling
for this example:

\
{               # thread 1

  test0         # thread 2

  : group       # thread 1
  {
    +setup1     # thread 1
    +setup2     # thread 1

    test1       # thread 3
    test2       # thread 4
    test3       # thread 1

                # thread 1 (wait for 3 & 4)

    -teardown2  # thread 1
    -teardown1  # thread 1
  }
                # thread 1 (wait for 2)
}
\

A testscript would normally contain multiple tests and sometimes it is
desirable to only execute a specific test or a group of tests. For example,
you may be debugging a failing test and would like to re-run it. As an
example, consider the following testscript file called \c{basics.test}:

\
$* foo : foo

: fox
{
  $* fox bar : bar
  $* fox baz : baz
}
\

The id paths for these three test will then be:

\
basics/foo
basics/fox/bar
basics/fox/baz
\

To only run individual tests, test groups, or testscript files we can specify
their id paths in the \c{config.test} variable, for example:

\
$ b test config.test=basics     # All tests in basics.test.
$ b test config.test=basics/fox # All tests in fox (bar and baz).
$ b test config.test=basics/foo # Test foo.
$ b test \"config.test=basics/foo basics/fox/bar\" # Tests foo and bar.
\


\h1#lexical|Lexical Structure|

Testscript is a line-oriented language with a context-dependent lexical
structure. It \"borrows\" several building blocks (variable expansion,
function calls, and evaluation contexts; collectively called \i{expansions}
from now on) from the Buildfile language. In a sense, testscripts are
specialized (for testing) continuations of buildfiles.

Except in here-document fragments, leading whitespaces and blank lines are
ignored except for the line/column counts. A non-empty testscript must
end with a newline.

Except in single-quoted strings and single-quoted here-document fragments,
the backslash (\c{\\}) character followed by a newline signals the line
continuation. Both this character and the newline are removed (note: not
replaced with a whitespace) and the following line is read as if it was part
of the first line. Note that \c{'\\'} followed by EOF is invalid. For example:

\
$* foo | \
$* bar
\

Except in here-document fragments, an unquoted and unescaped \c{'#'} character
starts a comment; everything from this character until the end of line is
ignored. For example:

\
# Setup foo.
$* foo

$* bar # Setup bar.
\

There is no line continuation support in comments; the trailing \c{'\\'} is
ignored except in one case: if the comment is just \c{'#\\'} followed by the
newline, then it starts a multi-line comment that spans until closing
\c{'#\\'} is encountered. For example:

\
#\
$* foo
$* bar
#\

$* foo #\
$* bar
$* baz #\
\

Similar to Buildfile, the Testscript language supports two types of quoting:
single (\c{'}) and double (\c{\"}). Both can span multiple lines.

The single-quoted strings and single-quoted here-document fragments do not
recognize any escape sequences (not even for the single quote itself or line
continuations) or expansions with all the characters taken literally until the
closing single quote or here-document end marker is encountered.

The double-quoted strings and double-quoted here-document fragments recognize
escape sequences (including line continuations) and expansions. For example:

\
foo = FOO

# 'FOO true'
#
bar = \"$foo ($foo == FOO)\"

# 'FOO bool'
#
$* <<\"EOI\"
$foo $type($foo == FOO)
EOI
\

Characters that have special syntactic meaning (for example \c{'$'}) can be
escaped with a backslash (\c{\\}) to preserve their literal meaning (to
specify literal backslash you need to escape it as well). For example:

\
foo = \$foo\\bar # '$foo\bar'
\

Note that quoting could often be a more readable way to achieve the same
result, for example:

\
foo = '$foo\bar'
\

Inside double-quoted strings only the \c{\"\\$(} character set needs to be
escaped. Inside double-quoted here-document fragments \- only \c{\\$(}.

The lexical structure of a line depends on its type. The line type may be
dictated by the preceding construct, as is the case for here-document
fragments. Otherwise the line type is determined by examining the leading
character and, if that fails to determine the line type, leading tokens,
as described next.

A character is said to be \i{unquoted} and \i{unescaped} if it is not escaped
and is not part of a quoted string. A token is said to be unquoted and
unescaped if all its characters are unquoted and unescaped.

The following characters determine the line type if they appear unquoted and
unescaped at the beginning of the line:

\
':'  - description line
'.'  - directive line
'{'  - block start
'}'  - block end
'+'  - setup command line
'-'  - teardown command line
\

If the line doesn't start with any of these characters then the first token of
the line is examined in the \c{first_token} mode (see below). If the first
token is an unquoted word, then the second token of the line is examined in
the \c{second_token} mode (see below). If it is a variable assignment (either
\c{+=}, \c{=+}, or \c{=}), then the line type is a variable line. Otherwise,
it is a test command line. Note that this means computed variable names are
not supported.

The Testscript language defines the following distinct lexing modes (or
contexts):

\dl|

\li|\n\n\cb{command_line}\n

  Whitespaces are token separators. The following characters and character
  sequences (read vertically) are recognized as tokens:

  \
  :;=!|&<>$(#
    ==
  \

  |

\li|\n\n\cb{first_token}\n

  Like \c{command_line} but recognizes variable assignments as separators.|

\li|\n\n\cb{second_token}\n

  Like \c{command_line} but recognizes variable assignments as tokens.|

\li|\n\n\cb{command_expansion}\n

  Subset of \c{command_line} used for re-lexing expansions (see below). Only
  the \c{|&<>} characters are recognized as tokens. Note that whitespaces are
  not separators in this mode.|

\li|\n\n\cb{variable_line}\n

  Similar to the Buildfile value mode. The \c{;$([]} characters are recognized
  as tokens.|

\li|\n\n\cb{description_line}\n

  Like a single-quoted string.|

\li|\n\n\cb{here_line_single}\n

  Like a single-quoted string except it treats newlines as a separator and
  quotes as literals.|

\li|\n\n\cb{here_line_double}\n

  Like a double-quoted string except it treats newlines as a separator and
  quotes as literals. The \c{$(} characters are recognized as tokens.||

Besides having varying lexical structure, parsing some line types involves
performing expansions (variable expansions, function calls, and evaluations
contexts). The following table summarizes the mapping of line types to lexing
modes and indicates whether they are parsed with expansions:

\
variable line                 variable_line
directive line                command_line       expansions
description line              description_line

test command line             command_line       expansions
setup command line            command_line       expansions
teardown command line         command_line       expansions

here-document single-quoted   here_line_single
here-document double-quoted   here_line_double   expansions
\

Finally, unquoted expansions in command lines (test, setup, and teardown) are
re-lexed in the \c{command_expansion} mode in order to recognize command line
syntax tokens (redirects, pipes, etc). To illustrate why this re-lexing is
necessary, consider the following example of a \"canned\" command line:

\
x = echo >-
$x foo
\

The command line token sequence will be \c{$}, \c{x}, \c{foo}. After the
expansion we get \c{echo}, \c{>-}, \c{foo}, however, the second string is not
(yet) recognized as a redirect. To achieve this we need to re-lex the result
of the expansion.

Note that besides the few command line syntax characters, re-lexing will also
\"consume\" quotes and escapes, for example:

\
args = \"'foo'\"  # 'foo'
echo $args      # echo foo
\

To preserve quotes in this context we need to escape them:

\
args = \"\'foo\'\"  # \'foo\'
echo $args        # echo 'foo'
\

Alternatively, for a single value, we could quote the expansion:

\
arg = \"'foo'\"  # 'foo'
echo \"$arg\"    # echo 'foo'
\

To minimize unhelpful consumptions of escape sequences (e.g., in Windows
paths), re-lexing performs only \"effective escaping\" for the \c{'\"\\}
characters. All other escape sequences are passed through uninterpreted. Note
that this means there is no way to escape command line syntax characters. The
idea is to use quoting except for passing literal quotes, for example:

\
args = \'&foo\'  # '&foo'
echo $args       # echo &foo
\

\h1#syntax|Syntax and Semantics|

\h#syntax-notation|Notation|

The formal grammar of the Testscript language is specified using an EBNF-like
notation with the following elements:

\
foo: ...   - production rule
foo        - non-terminal
<foo>      - terminal
'foo'      - literal
foo*       - zero or more multiplier
foo+       - one or more multiplier
foo?       - zero or one multiplier
foo bar    - concatenation (foo then bar)
foo | bar  - alternation   (foo or bar)
(foo bar)  - grouping
{foo bar}  - grouping in any order (foo then bar or bar then foo)
foo\
bar        - line continuation
# foo      - comment
\

Rule right-hand-sides that start on a new line describe the line-level syntax
and ones that start on the same line describes the syntax inside the line. If
a rule contains multiple lines, then each line matches a separate line in the
input.

If a multiplier appears in from on the line then it specifies the number of
repetitions for the whole line. For example, from the following three rules,
the first describes a single line of multiple literals, the second \- multiple
lines of a single literal, and the third \- multiple lines of multiple
literals.

\
# foofoofoo
#
text-line: 'foo'+

# foo
# foo
# foo
#
text-lines:
  +'foo'

# foo
# foofoo
# foofoofoo
#
text-lines:
  +('foo'+)
\

A newline in the grammar matches any standard newline separator sequence
(CR/LF combinations). An unquoted space in the grammar matches zero or more
non-newline whitespaces (spaces and tabs). A quoted space matches exactly one
non-newline whitespace. Note also that in some cases components within lines
may not be whitespace-separated in which case they will be written without any
spaces between them, for example:

\
foo: 'foo' ';'    # 'foo;' or 'foo ;' or 'foo      ;'
bar: 'bar'';'     # 'bar;'
baz: 'baz'' '+';' # 'baz ;' or 'baz      ;'

fox: bar''bar     # 'bar;bar;'
\

You may also notice that several production rules below end with \c{-line}
while potentially spanning several physical lines. In this case they represent
\i{logical lines}, for example, a command line and its here-document
fragments.

\h#syntax-grammar|Grammar|

The complete grammar of the Testscript language is presented next with the
following sections discussing the semantics of each production rule.

\
script:
  scope-body

scope-body:
  *setup
  *(scope|directive|test)
  *tdown

scope:
  ?description
  scope-block|scope-if

scope-block:
  '{'
  scope-body
  '}'

scope-if:
  ('if'|'if!') command-line
  scope-block
  *scope-elif
  ?scope-else

scope-elif:
  ('elif'|'elif!') command-line
  scope-block

scope-else:
  'else'
  scope-block

directive:
  '.' include

include: 'include' (' '+'--once')*(' '+<path>)*

setup:
  variable-like|setup-line

tdown:
  variable-like|tdown-line

setup-line: '+' command-like
tdown-line: '-' command-like

test:
  ?description
  +(variable-line|command-like)

variable-like:
  variable-line|variable-if

variable-line:
  <variable-name> ('='|'+='|'=+') value-attributes? <value>

value-attributes: '[' <key-value-pairs> ']'

variable-if:
  ('if'|'if!') command-line
  variable-if-body
  *variable-elif
  ?variable-else
  'end'

variable-elif:
  ('elif'|'elif!') command-line
  variable-if-body

variable-else:
  'else'
  variable-if-body

variable-if-body:
  *variable-like

command-like:
  command-line|command-if

command-line: command-expr (';'|(':' <text>))?
  *here-document

command-expr: command-pipe (('||'|'&&') command-pipe)*

command-pipe: command ('|' command)*

command: <path>(' '+(<arg>|redirect|cleanup))* command-exit?

command-exit: ('=='|'!=') <exit-status>

command-if:
  ('if'|'if!') command-line
  command-if-body
  *command-elif
  ?command-else
  'end' (';'|(':' <text>))?

command-elif:
  ('elif'|'elif!') command-line
  command-if-body

command-else:
  'else'
  command-if-body

command-if-body:
  *(variable-line|command-like)

redirect: stdin|stdout|stderr

stdin:  '0'?(in-redirect)
stdout: '1'?(out-redirect)
stderr: '2'(out-redirect)

in-redirect:  '<-'|\
              '<+'|\
              '<'{':'?'/'?} <text>|\
              '<<'{':'?'/'?} <here-end>|\
              '<<<' <file>

out-redirect: '>-'|\
              '>+'|\
              '>&' ('1'|'2')|\
              '>'{':'?'/'?}'~'? <text>|\
              '>>'{':'?'/'?}'~'? <here-end>|\
              '>>>'{'&'?} <file>

here-document:
  *<text>
  <here-end>

cleanup: ('&'|'&?'|'&!') (<file>|<dir>)

description:
  +(':' <text>)
\

\h#syntax-script|Script|

\
script:
  scope-body
\

A testscript file is an implicit group scope.

\h#syntax-scope|Scope|

\
scope-body:
  *setup
  *(scope|directive|test)
  *tdown

scope:
  ?description
  scope-block|scope-if

scope-block:
  '{'
  scope-body
  '}'
\

A scope is either a test group scope or an explicit test scope. An explicit
scope is a test scope if it contains a single test, only variable assignments
in setup commands, no teardown commands, and only the scope having the
description, if any. Otherwise, it is a group scope. If there is no explicit
scope for a test, one is established implicitly.


\h#syntax-scope-if|Scope-If|

\
scope-if:
  ('if'|'if!') command-line
  scope-block
  *scope-elif
  ?scope-else

scope-elif:
  ('elif'|'elif!') command-line
  scope-block

scope-else:
  'else'
  scope-block
\

A scope, either test or group, can be executed conditionally. The condition
\c{command-line} is executed in the context of the outer scope. Note that all
the scopes in an \c{if-else} chain are alternative implementations of the same
group/test (thus the single description). If at least one of them is a group
scope, then all the others are treated as groups as well.

\h#syntax-directive|Directive|

\
directive:
  '.' include
\

A line that starts with \c{.} is a Testscript directive. Note that directives
are evaluated during parsing, before any command is executed or testscript
variable assigned. You can, however, use variables assigned in the
buildfile. For example:

\
include common-$(cxx.target.class).test
\

\h2#syntax-directive-include|Include|

\
include: 'include' (' '+'--once')*(' '+<path>)*
\

While in the grammar the \c{include} directive is shown to only appear
interleaving with scopes and tests, it can be used anywhere in the scope
body. It can also contain several parts of a scope, for example, setup and
test lines.

The \c{--once} option signals that files that have already been included in
this scope should not be included again. The implementation is not required to
handle links when determining if two paths are to the same file. Relative
paths are assumed to be relative to the including testscript file.

\h#syntax-setup-teardown|Setup and Teardown|

\
setup:
  variable-like|setup-line

tdown:
  variable-like|tdown-line

setup-line: '+' command-like
tdown-line: '-' command-like
\

Setup and teardown commands are executed sequentially in the order
specified. Note that variable assignments (including \c{variable-if}) do not
use the \c{'+'} and \c{'-'} prefixes. A standalone (not part of a test)
variable assignment is automatically treated as setup if no tests have yet
been encountered in this scope and as teardown otherwise.

\h#syntax-test|Test|

\
test:
  ?description
  +(variable-line|command-like)
\

A test that contains multiple lines is called \i{compound}. In this case each
(logical) line except the last must end with a semicolon to signal the test
continuation. For example:

\
conf = test.conf;
cat <'verbose = true' >>>$conf;
test $conf
\

\h#syntax-variable|Variable|

\
variable-like:
  variable-line|variable-if

variable-line:
  <variable-name> ('='|'+='|'=+') value-attributes? <value>

value-attributes: '[' <key-value-pairs> ']'
\

The Testscript variable assignment semantics is equivalent to Buildfile except
that no \c{{\}}-based name-generation is performed. For example:

\
args = [strings] foo   bar 'fox   baz'
echo $args # foo bar fox   baz
\

\h#syntax-variable-if|Variable-If|

\
variable-if:
  ('if'|'if!') command-line
  variable-if-body
  *variable-elif
  ?variable-else
  'end'

variable-elif:
  ('elif'|'elif!') command-line
  variable-if-body

variable-else:
  'else'
  variable-if-body

variable-if-body:
  *variable-like
\

A group of variables can be set conditionally. The condition \c{command-line}
semantics is the same as in \c{scope-if}. For example:

\
if ($cxx.target.class == 'windows')
  slash = \\
  case = false
else
  slash = /
  case = true
end
\

When conditionally setting a single variable, using the evaluation context
with a ternary operator is often more concise:

\
slash = ($cxx.target.class == 'windows' ? \\ : /)
\

Note also that the only purpose of having a separate (from \c{command-if})
variable-only if-block is to remove the error-prone requirement of having to
specify \c{+} and \c{-} prefixes in group setup/teardown.

\h#syntax-command|Command|

\
command-like:
  command-line|command-if

command-line: command-expr (';'|(':' <text>))?
  *here-document

command-expr: command-pipe (('||'|'&&') command-pipe)*

command-pipe: command ('|' command)*

command: <path>(' '+(<arg>|redirect|cleanup))* command-exit?

command-exit: ('=='|'!=') <exit-status>
\

A \c{command-line} is \c{command-expr}. If it appears directly (as opposed to
inside \c{command-if}) in a test, then it can be followed by \c{;} to signal
the test continuation or by \c{:} and the trailing description.

A \c{command-expr} can combine several \c{command-pipe}'s with logical AND and
OR operators. Note that the evaluation order is always from left to right
(left-associative), both operators have the same precedence, and are
short-circuiting. Note, however, that short-circuiting does not apply to
expansions (variable, function calls, evaluation contexts). The logical result
of a \c{command-expr} is the result of the last \c{command-pipe} executed.

A \c{command-pipe} can combine several \c{command}'s with a pipe (\c{stdout}
of the left-hand-side command is connected to \c{stdin} of the
right-hand-side). The logical result of a \c{command-pipe} is the logical
AND of all its \c{command}'s.

A \c{command} begins with a command path following by options/arguments,
redirects, and cleanups, all optional and in any order.

A \c{command} may specify an exist code check. If executing a \c{command}
results in an abnormal process termination, then the whole outer construct
(e.g., test, setup/teardown, etc) summarily fails. Otherwise (that is, in case
of a normal termination) the exit code is checked. If omitted, then the test
is expected to succeed (0 exit code). The logical result of executing a
\c{command} is therefore a boolean value which is used in the higher-level
constructs (pipe and expression).

\h#syntax-command-if|Command-If|

\
command-if:
  ('if'|'if!') command-line
  command-if-body
  *command-elif
  ?command-else
  'end' (';'|(':' <text>))?

command-elif:
  ('elif'|'elif!') command-line
  command-if-body

command-else:
  'else'
  command-if-body

command-if-body:
  *(variable-line|command-like)
\

A group of commands can be executed conditionally. The condition
\c{command-line} semantics is the same as in \c{scope-if}. Note that in a
compound test commands inside \c{command-if} must not end with \c{;}. Rather,
\c{;} may follow \c{end}. For example:

\
if ($cxx.target.class == 'windows')
  foo = windows
  setup1
  setup2
else
  foo = posix
end;
test $foo
\

\h#syntax-redirect|Redirect|

\
redirect: stdin|stdout|stderr

stdin:  '0'?(in-redirect)
stdout: '1'?(out-redirect)
stderr: '2'(out-redirect)
\

The file descriptors must not be separated from the redirect operators with
whitespaces. And if leading text is not separated from the redirect operators,
then it is expected to be the file descriptor. As an example, the first command
below has \c{2} as an argument and redirects \c{stdout}, not \c{stderr}. While
the second is invalid since \c{a1} is not a valid file descriptor.

\
$* 2 >-
$* a1>-
\


\h#syntax-in-redirect|Input Redirect|

\
in-redirect:  '<-'|\
              '<+'|\
              '<'{':'?'/'?} <text>|\
              '<<'{':'?'/'?} <here-end>|\
              '<<<' <file>
\

The \c{stdin} data can come from a pipe, here-string (\c{<}), here-document
(\c{<<}), a file (\c{<<<}), or \c{/dev/null}-equivalent (\c{<-}). Specifying
both a pipe and a redirect is an error. If no pipe or \c{stdin} redirect is
specified and the test tries to read from \c{stdin}, it is considered to have
failed (unexpected input). However, whether this is detected and diagnosed is
implementation-defined. To allow reading from the default \c{stdin} (for
instance, if the test is really an example), the \c{<+} redirect is used.

Here-string and here-document redirects may specify the following modifiers.
The \c{:} modifier is used to suppress the otherwise automatically-added
terminating newline.

The \c{/} modifier causes all the forward slashes in the here-string or
here-document to be translated to the test target platform's directory
separator.

A here-document redirect must be specified \i{literally} on the command
line. Specifically, it must not be the result of an expansion (which rarely
makes sense anyway since the following here-document fragment itself cannot be
the result of an expansion either).


\h#syntax-in-output|Output Redirect|

\
out-redirect: '>-'|\
              '>+'|\
              '>&' ('1'|'2')|\
              '>'{':'?'/'?}'~'? <text>|\
              '>>'{':'?'/'?}'~'? <here-end>|\
              '>>>'{'&'?} <file>
\

The \c{stdout} and \c{stderr} data can go to a pipe (\c{stdout} only), file
(\c{>>>&} or \c{>>>&} to append), or \c{/dev/null}-equivalent (\c{>-}). It can
also be compared to here-string (\c{>}) or here-document (\c{>>}). For
\c{stdout} specifying both a pipe and a redirect is an error. A test that ties
to write to un-redirected stream (either \c{stdout} or \c{stderr}) it is
considered to have failed (unexpected output).

To allow writing to the default \c{stdout} or \c{stderr} (for instance, if the
test is really an example), the \c{>+} redirect is used.

It is also possible to merge \c{stderr} to \c{stdout} or vice versa with a
merge redirect (\c{>&}). In this case the left-hand-side descriptor (implied
or explicit) must not be the same as the right-hand-side. Having both merge
redirects at the same time is illegal.

The \c{:/} redirect modifiers have the same semantics as in the input
redirects. The \c{~} modifier is used to indicate that the following
here-string/here-document is a regular expression (discussed below) rather
than a literal. Note that if present, it must be specified last.

Similar to the input redirects, an output here-document redirect must be
specified literally on the command line.

\h#syntax-here-document|Here-Document|

\
here-document:
  *<text>
  <here-end>
\

A here-document fragments can be used to supply data to \c{stdin} or to
compare output to the expected result for \c{stdout} and \c{stderr}. Note that
the order of here-document fragments must match the order of redirects, for
example:

\
: select-no-table-error
$* --interactive >>EOO <<EOI 2>>EOE
enter query:
EOO
SELECT * FROM no_such_table
EOI
error: no such table 'no_such_table'
EOE
\

Here-strings can be single-quoted literals or double-quoted with expansion.
This semantics is extended to here-documents as follows. If the end marker
on the command line is single-quoted, then the here-document lines are
parsed as if they were single-quoted except that the single quote itself
is not treated as special. In this mode there are no expansions, escape
sequences, not even line continuations \- each line is taken literally.

If the end marker on the command line is double-quoted, then the here-document
lines are parsed as if they were double-quoted except that the double quote
itself is not treated as special. In this mode we can use variables
expansions, function calls, and evaluation contexts. However, we have to
escape the \c{$(\\} character set.

If the end marker is not quoted then it is treated as if it were
single-quoted. Note also that quoted end markers must be quoted \i{wholly},
that is, from the beginning and until the end and without any interruptions.

If the preceding command line starts with leading whitespaces, then the
equivalent number is stripped (if present) from each here-document line
(including the end marker). For example, the following two testscript
fragments are equivalent:

\
{
  $* <<EOI
  foo
  bar
  EOI
}
\

\
{
  $* <<EOI
foo
bar
EOI
}
\

The leading whitespace stripping does not apply to line continuations.


\h#syntax-regex|Output Regex|

The expected result in output here-strings and here-documents can be specified
as a regular expression instead of plain text. To signal the use of regular
expressions the redirect must include the \c{~} modifier, for example:

\
$* >~'/fo+/' 2>>~/EOE/
/ba+r/
baz
EOE
\

The regular expression used for output matching has two levels. At the outer
level the expression is over lines with each line treated as a single
character. We will refer to this outer expression as \i{line-regex} and
to its characters as \i{line-char}.

A line-char can be a literal line (like \c{baz} in the example above) in
which case it will only be equal to an identical line in the output. Or a
line-char can be an inner level regex (like \c{ba+r} above) in which
case it will be equal to any line in the output that matches this regex.
Where not clear from context we will refer to this inner expression as
\i{char-regex} and its characters as \i{char}.

A line is treated as literal unless it starts with the \i{regex introducer
character} (\c{/} in the above example). In contrast, the line-regex is always
in effect (in a sense, the \c{~} modifier is its introducer). Note that the
here-string regex naturally must always start with an introducer.

A char-regex line that starts with an introducer must also end with one
optionally followed by \i{match flags}, for example:

\
$* >>~/EOO/
/ba+r/i
/ba+z/i
EOO
\

The following match flag are recognized:

\dl|

\li|\n\c{i}

  Perform case-insensitive match.|

\li|\n\c{d}

  Invert the dot character (\c{.}) escaping. With this flag unescaped dots
  are treated as literal characters while the escaped ones (\c{\\.}) \-
  as matching any character. Note that dots specified within character
  classes (\c{[.]}) are not affected.||

Any character can act as a regex introducer. For here-strings it is the first
character in the string. For here-documents the introducer is specified as
part of the end marker. In this case the first character is the introducer,
everything after that and until the second occurrence of the introducer is the
actual end marker, and everything after that are global match flags. Global
match flags apply to every char-regex (but not literal lines or the line-regex
itself) in this here-document. Note that there is no way to escape the
introducer character inside the regex.

As an example, here is a shorter version of the previous example that also
uses a different introducer character.

\
$* >>~%EOO%i
%ba+r%
%ba+z%
EOO
\

By default a line-char is treated as an ordinary, non-syntax character with
regards to line-regex. Lines that start with a regex introducer but do not end
with one are used to specify syntax line-chars. Such syntax line-chars can
also be specified after (or instead of) match flags. For example:

\
$* >>~/EOO/
/(
/fo+x/|
/ba+r/|
/ba+z/
/)+
EOO
\

As an illustration, if we call the \c{/fo+x/} expression \c{A}, \c{/ba+r/} \-
\c{B}, and \c{/ba+z/} \- C, then we can represent the above line-regex in
the following more traditional form:

\
(A|B|C)+
\

Only characters from the \c{.()|*+?{\}\\0123456789,=!} set are allowed as
syntax line-chars with presence of any other character being an error.

A blank line as well as the \c{//} sequence (assuming \c{/} is the introducer)
are treated as an empty line-char. For the purpose of matching, newlines are
viewed as separators rather than being part of a line. In particular, in this
model, the customary trailing newline at the end of the output introduces a
trailing empty line-char. As a result, unless the \c{:} (no newline) redirect
modifier is used, an empty line-char is implicitly added to line-regex.


\h#syntax-cleanup|Cleanup|

\
cleanup: ('&'|'&?'|'&!') (<file>|<dir>)
\

If a command creates extra files or directories then they can be register for
automatic cleanup at the end of the scope (test or group). Files mentioned in
redirects are registered automatically. Additionally, certain builtints (for
example \c{touch} and \c{mkdir}) also register their output files/directories
automatically (as described in each builtin's documentation).

If the path ends with a directory separator (slash), then it is assumed to be
a directory. Otherwise \- a file. A directory about to be removed must be
empty (no unexpected output).

The \c{&} syntax registers a normal or \i{always} cleanup: the test fails if
the file/directory does not exist. The \c{&?} syntax is a \i{maybe} cleanup:
the file/directory is removed if it exists. Finally, \c{&!} is a \i{never}
cleanup: it disables a previously registered cleanup for this file/directory
(primarily used to disable automatic cleanups registered by builtins).

The last component in the path may contain a wildcard with the following
semantics:

\
dir/*   - all immediate files
dir/*/  - all immediate sub-directories (which must be empty)
dir/**  - all files recursively
dir/**/ - all sub-directories recursively (which must be empty)
dir/*** - all files and sub-directories recursively and dir/
\

Registering a path for cleanup that is outside the script working directory is
an error.


\h#syntax-description|Description|

\
description:
  +(':' <text>)
\

Description lines start with a colon (\c{:}) and are used to document tests
and test groups. In a sense they are formalized comments.

A description can be \i{leading}, that is, specified before the test or
group. For tests it can also be \i{trailing} \- specified as a single line
after the (last) command of the test. It is an error to specify both leading
and trailing descriptions.

By convention the leading description has the following format with all three
components being optional.

\
: <id>
: <summary>
:
: <details>
\

If the first line in the description does not contain any whitespaces, then it
is assumed to be the test or test group id. If the next line is followed by a
blank line, then it is assume to be the test or test group summary. After the
blank line come optional details which are free-form.

The trailing description can only be used to specify the id or summary (but
not both).

If an id is not specified then it is automatically derived from the test or
test group location. If the test or test group is contained directly in the
top-level testscript file, then just its start line number is used as an id.
Otherwise, if the test or test group reside in an included file, then the
start line number (inside the included file) is prefixed with the line number
of the \c{include} directive followed by the included file name (without the
extension) in the form \c{<line>-<file>-}. This process is repeated
recursively in case of nested inclusion.

The start line for a scope (either test or group) is the line containing its
opening brace (\c{{}) and for a test \- the first test line.


\h1#builtins|Builtins|

The Testscript language provides a portable subset of POSIX utilities. Each
utility normally implements the commonly used subset of the corresponding
POSIX specification, though there are deviations (e.g., in option handling)
and extensions, as described in this chapter. Note also that the builtins are
implemented in-process with some of the simple ones (e.g., \c{true/false},
\c{mkdir}, etc) being just function calls.


\h#builtins-cat|\c{cat}|

\
cat <file>...
\

Read files in order and write their contents to \c{stdout}. Read from
\c{stdin} if no file is specified or \c{-} is specified as a file name.


\h#builtins-echo|\c{echo}|

\
echo <string>...
\

Write strings to \c{stdout} separating them with a single space followed
by a newline.

\h#builtins-false|\c{false}|

\
false
\

Do nothing and terminate normally with 1 exit code indicating failure.

\h#builtins-mkdir|\c{mkdir}|

\
mkdir [-p] <dir>...
\

Create directories. Unless the \c{-p} option is specified, all the leading
directories must exist and the directory itself must not exist.

\dl|

\li|\n\c{-p}

  Create missing leading directories and ignore directories that already
  exist.||

Created directories (including the leading ones) that are inside the script
working directory are automatically registered for cleanup.


\h#builtins-rm|\c{rm}|

\
rm [-r] [-f] <path>...
\

Remove filesystem entries. To remove a directory (even empty) the \c{-r}
option must be specified.

The path must not be the test working directory or its parent directory. It
also must not be outside the script working directory unless the \c{-f} option
is specified.

\dl|

\li|\n\c{-r}

  Remove directories and their contents recursively.|

\li|\n\c{-f}

  Do not fail if no path is specified, the path does not exist, or is outside
  the script working directory.||

Note that the implementation deviates from POSIX in a number of ways. It never
interacts with the user and fails immediately if unable to act on an
argument. It does not check for dot containment in the path nor considers
filesystem permissions. In essence, it simply tries to remove the filesystem
entry. It also always fails if an empty path is specified.


\h#builtins-rmdir|\c{rmdir}|

\
rmdir [-f] <dir>...
\

Remove directories. The directory must be empty and not be the test working
directory or its parent directory. It also must not be outside the script
working directory unless the \c{-f} option is specified.

\dl|

\li|\n\c{-f}

  Do not fail if no directory is specified, the directory does not exist, or
  is outside the script working directory.||


\h#builtins-touch|\c{touch}|

\
touch <file>...
\

Change file access and modification times to the current time. Create files
that do not exist. Fail if a file system entry other than the file exists for
the specified name.

Created files that are inside the script working directory are automatically
registered for cleanup.

\h#builtins-true|\c{true}|

\
true
\

Do nothing and terminate normally with 0 exit code indicating success.

\h1#style|Style Guide|

This section describes the Testscript style that is used in the \c{build2}
project. The primary goal of testing in \c{build2} is not to exhaustively
test every possible situation. Rather, it is to keep tests comprehensible
and maintainable in the long run.

To this effect, don't try to test every possible combination; this striving
will quickly lead to everyone drowning in hundreds of tests that are only
slight variations of each other. Sometimes combination tests are useful but
generally keep things simple and test one thing at a time. The believe here is
that real-world usage will uncover much more interesting interactions (which
must become regression tests) that you would never have thought of yourself.
To quote a famous physicist, \"\i{...  the imagination of nature is far, far
greater than the imagination of man.}\"

To expand on combination tests, don't confuse them with corner case tests. As
an example, say you have tests for feature A and B. Now you wonder what if for
some reason they don't work together. Note that you don't have a clear
understanding of why they might not work together; you just want to add one
more test, \i{for good measure}. We don't do that. To put it another way, for
each test you should have a clear understanding of what logic in the code you
are testing.

One approach that we found works well is to look at the diff of changes you
would like to commit and make sure you at least have a test that exercises
each \i{happy} (non-error) \i{logic branch}. For critical code you may also
want to do so for unhappy logic branches.

It is also a good idea to keep testing in mind as you implement things. When
tempted to add a small special case just to make the result \i{nicer},
remember that you will also have to test this special case.

If the functionality is well exposed in the program, prefer functional to unit
tests since the former test the end result rather than something intermediate
and possibly mocked. If unit-testing a complex piece of functionality,
consider designing a concise textual \i{mini-format} for input (either via
command line or \c{stdin}) and output rather than constructing the test data
and expected results programmatically.

Documentation-wise, each test should at least include explicit id that
adequately summarizes what it tests. Add summary or even details for more
complex tests. Failure tests usually fall into this category.

Use the leading description for multi-line tests, for example:

\
: multi-name
:
$* 'John' 'Jane' >>EOO
Hello, John!
Hello, Jane!
EOO
\

Here is an example of a description that includes all three components:

\
: multi-name
: Test multiple name arguments
:
: This test makes sure we properly handle multiple names passed as
: separate command line arguments.
:
$* 'John' 'Jane' >>EOO
Hello, John!
Hello, Jane!
EOO
\

Separate multi-line tests with blank lines. You may want to place larger tests
into explicit test scopes for better visual separation (this is especially
helpful if the test contains blank lines, for example, in here-document
fragments). In this case the description should come before the scope. Note
that here-documents are indented as well. For example:

\
: multi-name
:
{
  $* 'John' 'Jane' >>EOO
  Hello, John!

  Hello, Jane!

  EOO
}
\

One-line tests may use the trailing description (which must always be
the test id). Within a test block (one-liners without a blank between
them), the ids should be aligned, for example:

\
$* John >'Hi, John!'       : custom-john
$* World >'Hello, World!'  : custom-world
\

Note that you are free to put multiple spaces between the end of the command
line and the trailing description. But don't try to align ids between blocks
\- this is a maintenance pain.

If multiple tests belong to the same group, consider placing them into an
explicit group scope. A good indication that tests form a group is if their
ids start with the same prefix, as in the above example. If placing tests into
a group scope, use the prefix as the group's id and don't repeat it in the
tests. It is also a good idea to give the summary of the group, for example:

\
: custom
: Test custom greetings
:
{
  $* John >'Hi, John!'       : john
  $* World >'Hello, World!'  : world
}
\

In the same vein, don't repeat the testscript id in group or test ids. For
example, if the above tests were in the \c{greeting.test} testscript, then
using \c{custom-greeting} as the group id would be unnecessarily repetitive
since the id path would become, for example,
\c{greeting/custom-greeting/john}.

We quote values that are \i{strings} as opposed to options, file names, paths
(unless contain spaces), integers, or boolean. When quoting, use the single
quote unless you use expansion (or single quotes) inside. Note that unlike
Bash you do not need to quote variable expansions in order to preserve
whitespaces.  For example:

\
arg = 'Hello   Spaces'
echo $arg              # Hello   Spaces
\

For further reading on testing that we (mostly) agree with, see:

\dl|

\li|\n\n\l{https://blog.nelhage.com/2016/12/how-i-test/ How I Write Tests} by Nelson Elhage\n

  The only part we don't agree on is the (somewhat implied) suggestion to
  write as many tests as possible.||
"