A testing system for C, contained in 1 header file.
-
Small, Portable, Lightweight
greatest doesn't depend on anything beyond ANSI C89, and the test scaffolding should build without warnings when compiled with
-Wall -Wextra -pedantic
. It is under 1,000 LOC (SLOCCount), and does no dynamic allocation. -
Permissive License
greatest is released under the ISC License. You can use it freely, even for commercial purposes.
-
Easy To Set Up
To use, just
#include "greatest.h"
in your project. There is very little boilerplate. Most features are optional. -
Un-Opinionated
When a command-line test runner is useful, greatest can provide one, but it can also run as part of other programs. It doesn't depend on a particular build system or other tooling, and should accommodate a variety of testing approaches. It actively avoids imposing architectural choices on code under test. While greatest was designed with C in mind, it attempts to be usable from C++.
-
Modular
Tests can be run individually, or grouped into suites. Suites can share common setup, and can be in distinct compilation units.
-
Low Friction
Specific tests or suites can be run by name, for focused and rapid iteration during development. greatest adds very little startup latency.
There are some compile-time options, and slightly nicer syntax for parametric testing (running tests with arguments) if compiled with a C99 or later language standard.
I wrote a blog post with more information. While it's several years old, it's still accurate about the main functionality.
theft, a related project, adds property-based testing.
#include "greatest.h"
/* A test runs various assertions, then calls PASS(), FAIL(), or SKIP(). */
TEST x_should_equal_1(void) {
int x = 1;
/* Compare, with an automatic "1 != x" failure message */
ASSERT_EQ(1, x);
/* Compare, with a custom failure message */
ASSERT_EQm("Yikes, x doesn't equal 1", 1, x);
/* Compare, and if they differ, print both values,
* formatted like `printf("Expected: %d\nGot: %d\n", 1, x);` */
ASSERT_EQ_FMT(1, x, "%d");
PASS();
}
/* Suites can group multiple tests with common setup. */
SUITE(the_suite) {
RUN_TEST(x_should_equal_1);
}
/* Add definitions that need to be in the test runner's main file. */
GREATEST_MAIN_DEFS();
int main(int argc, char **argv) {
GREATEST_MAIN_BEGIN(); /* command-line options, initialization. */
/* Individual tests can be run directly in main, outside of suites. */
/* RUN_TEST(x_should_equal_1); */
/* Tests can also be gathered into test suites. */
RUN_SUITE(the_suite);
GREATEST_MAIN_END(); /* display results */
}
Output:
$ make simple && ./simple
cc -g -Wall -Werror -pedantic simple.c -o simple
* Suite the_suite:
.
1 test - 1 passed, 0 failed, 0 skipped (5 ticks, 0.000 sec)
Total: 1 test (47 ticks, 0.000 sec), 3 assertions
Pass: 1, fail: 0, skip: 0.
Tests are run with RUN_TEST(test_name)
, which can be called directly
from the test runner's main
function or grouped into suites (which are
run with RUN_SUITE(suite_name)
). (Calls to RUN_TEST
from inside
another test are ignored.)
Test cases can be run with arguments: RUN_TEST1(test_name, arg)
passes
a single argument, and if C99 features are supported, then
RUN_TESTp(test_name, ...)
uses __VA_ARGS__
to run a test case with
one or mare arguments. greatest_set_test_suffix
sets a name suffix, so
output from the test runner can include info about arguments.
Test cases should call assertions and then end with PASS()
, SKIP()
,
FAIL()
, or their custom message variants (e.g. SKIPm("TODO");
).
If there are any test failures, the test runner will return 1,
otherwise it will return 0. (Skips do not cause the test runner to
report failure.)
PASS()
, SKIP()
, FAIL()
, and their custom message variants are
macros that updating internal bookkeeping and then returning and enum
value, such as GREATEST_TEST_RES_FAIL
. They all return
from the
current test case function.
PASS()
/PASSm("msg")
prints as a dot when verbosity is zero, or
the test name and custom message (if any) with verbosity >= 1.
FAIL()
/FAILm("msg")
always prints "FAIL test_name: msg file:line".
SKIP()
/SKIPm("msg")
prints as an 's' when verbosity is zero, or the
test name and custom message (if any) with verbosity >= 1. Because skips
are not treated as a failure by the test runner, they can be used to
skip test cases that aren't relevant in a particular build or
environment, a way to temporarily disable broken tests, or as a sort of
todo list for tests and functionality under active development.
Tests and suites are just functions, so normal C scoping rules apply.
For example, a test or suite named main
will have a name collision.
(For more examples, look at example.c
and example_suite.c
.)
greatest runs all tests by default, but can be configured to only run suites or tests whose names contain a filter string, and/or exclude tests whose name contains a filter string. When test name filtering and exclusion are used together, exclusion takes precedence.
void greatest_set_suite_filter(const char *name);
void greatest_set_test_filter(const char *name);
void greatest_set_test_exclude(const char *name);
These correspond to the following command line test runner options:
`-s SUITE`: Only run suites whose names contain the string "SUITE"
`-t TEST`: Only run tests whose names contain the string "TEST"
`-x EXCLUDE`: Exclude tests whose names contain the string "EXCLUDE"
For example, to run any tests with "tree" in the name, in suites with "pars" in the name (such as "parser"), but exclude any tests whose names also contain "slow":
./test_project -s pars -t tree -x slow
The string matching includes optional test name suffixes.
The greatest_set_exact_name_match()
function and corresponding -e
command line runner flag can be used to only run tests and/or suites
whose names exactly match the name filter(s). Note: exact-match suite
filtering by name will not skip tests that are run outside of any suite.
Assertions fail the current test unless some condition holds. All
assertions have a custom message variant (with an m
suffix), which
takes a message string as its first argument. For example, the
assertion ASSERT_EQ(apple, orange);
could instead be used like
ASSERT_EQm("these should match", apple, orange)
. Non-message
assertions create a default message.
Assert that COND
evaluates to a true (non-zero) value.
Assert that COND
evaluates to a false (zero) value.
Assert that EXPECTED == ACTUAL
. To print the values if they
differ, use ASSERT_EQ_FMT
. To compare with custom equality test
and print functions, use ASSERT_EQUAL_T
instead.
Assert that EXPECTED != ACTUAL
.
Assert that EXPECTED > ACTUAL
.
Assert that EXPECTED >= ACTUAL
.
Assert that EXPECTED < ACTUAL
.
Assert that EXPECTED <= ACTUAL
.
Assert that EXPECTED == ACTUAL
. If they are not equal, print their
values using FORMAT
as the printf
format string.
For example: ASSERT_EQ_FMT(123, result, "%d");
will call printf
like printf("Expected: %d\nGot: %d\n", 123, result);
if its
EXPECTED
and ACTUAL
arguments don't match.
Note: EXPECTED
and ACTUAL
will be evaluated more than once on
failure, so they should not be a function call with side effects.
(Since their type is not known by the macro, they cannot be
captured in a local variable. typeof
is a GCC extension.)
Assert that ACTUAL
is within EXPECTED
+/- TOLERANCE
, once the
values have been converted to a configurable floating point type
(GREATEST_FLOAT
).
greatest does not depent on floating point math.
It is only used within ASSERT_IN_RANGE
's macro expansion.
Assert that the strings are equal
(i.e., strcmp(EXPECTED, ACTUAL) == 0
).
Assert that the first SIZE
bytes of the strings are equal
(i.e., strncmp(EXPECTED, ACTUAL, SIZE) == 0
).
Assert that the first SIZE
bytes of memory pointed to by EXPECTED
and ACTUAL
are equal. If their memory differs, print a hexdump and
highlight the lines and individual bytes which do not match.
Assert that the enum value EXPECTED
is equal to ACTUAL
. If not,
convert each enum value to a string using ENUM_STR_FUN
before printing
them.
ENUM_STR_FUN
should have a type like:
const char *FUN(int x);
Assert that EXPECTED
and ACTUAL
are equal, using the
greatest_equal_cb
function pointed to by TYPE_INFO->equal
to compare
them. The assertion's UDATA
argument can be used to pass in arbitrary
user data (or NULL
) to the callbacks. If the values are not equal and
the TYPE_INFO->print
function is defined, it will be used to print an
"Expected: X, Got: Y" message.
Assert that COND
evaluates to a true value. If not, then use
longjmp(3)
to immediately return from the test case and any
intermediate function calls.
If built with GREATEST_USE_LONGJMP
#define
d to 0, then all
setjmp/longjmp-related functionality will be compiled out. This also
reduces memory usage by sizeof jmp_buf
, which may be several hundred
bytes, depending on the platform.
Groups of suites or tests can be run in shuffled order by using
GREATEST_SHUFFLE_SUITES
and GREATEST_SHUFFLE_TESTS
, respectively.
This can help find and eliminate accidental coupling between tests.
The shuffling depends on the seed and the test/suite count, so a consistent seed will only lead to reproducible ordering until the group's count changes.
Shuffling suites:
SHUFFLE_SUITES(seed, {
RUN_SUITE(suite1);
RUN_SUITE(suite2);
RUN_SUITE(suite3);
RUN_SUITE(suite4);
RUN_SUITE(suite5);
});
Shuffling tests:
SHUFFLE_TESTS(seed, {
RUN_TEST(test_a);
RUN_TEST1(test_b, 12345);
RUN_TEST(test_c);
RUN_TESTp(test_d, "some_argument");
RUN_TEST(test_e);
});
Note: Any other code inside the block will be executed several times.
The shuffling macro expands to a loop with (count + 1) iterations -- the
first pass counts, and the following passes only execute the next chosen
suite/test. In particular, avoid running tests directly inside of a
SHUFFLE_SUITES
block without a suite, because the test will be run
on each iteration.
greatest_set_test_suffix
can be used to set an optional name suffix
for the next test:
for (i = 0; i < row_count; i++) {
const struct table_row *row = &table[row_count];
greatest_set_test_suffix(row->name);
RUN_TEST1(test_with_arg, row);
}
This will cause the test name to print with a _
separator and the
suffix in all pass/fail/skip messages (i.e., test_with_arg_KEY
). This
is especially useful when running a test several times with different
arguments, in shuffled order. The name suffix is included when using
name-based filtering.
The test name and optional suffix are copied into an internal buffer.
Its size can be configured by #define
ing the constant
GREATEST_TESTNAME_BUF_SIZE
. (If not #define
d, it defaults to 128
bytes.) If the buffer is not large enough for the name and suffix, it
will truncate after size - 1
bytes, to ensure that it is properly
\0
-terminated.
The name suffix pointer is cleared after each RUN_TEST*
call, so a
suffix can be constructed in a stack allocated buffer without later
dereferencing a pointer that has gone out of scope.
Because of how PASS()
, ASSERT()
, FAIL()
, etc. are implemented
(returning a test result enum value), calls to functions that use them
directly from test functions must be wrapped in CHECK_CALL
:
TEST example_using_subfunctions(void) {
CHECK_CALL(less_than_three(5));
PASS();
}
This is only necessary if the called function can cause test failures.
The function should have a return type of enum greatest_test_res
.
Test runners build with the following command line options:
Usage: (test_runner) [-hlfave] [-s SUITE] [-t TEST] [-x EXCLUDE]
-h, --help print this Help
-l List suites and tests, then exit (dry run)
-f Stop runner after first failure
-a Abort on first failure (implies -f)
-v Verbose output
-s SUITE only run suite w/ name containing substring SUITE
-t TEST only run test w/ name containing substring TEST
-e only run exact name match for -s or -t
-x EXCLUDE exclude tests containing string substring EXCLUDE
Any arguments after --
will be ignored.
If you want to run multiple test suites in parallel, look at parade.
These command line options are processed by GREATEST_MAIN_BEGIN();
.
Rather than producing a command line test runner (which checks the
command line options, and exits with a pass/fail return code after
running tests), greatest can be used more like a library. Instead of
using GREATEST_MAIN_BEGIN()
, use GREATEST_INIT()
to (re-)initialize
greatest, then use either GREATEST_PRINT_REPORT()
to print the report
to GREATEST_STDOUT
, or use greatest_get_report(&report)
to get the
pass, fail, skip, and assertion counters.
The command line flags above have corresponding functions:
greatest_stop_at_first_fail()
greatest_abort_on_fail()
greatest_list_only()
greatest_set_exact_name_match()
greatest_set_suite_filter(const char *filter)
greatest_set_test_filter(const char *filter)
greatest_set_test_exclude(const char *filter)
greatest_get_verbosity()
greatest_set_verbosity(unsigned int verbosity)
Most of the macros have prefixed and unprefixed forms. For example,
SUITE
is the same as GREATEST_SUITE
.
Check the source for the list -- search for GREATEST_USE_ABBREVS
.
These aliases can be disabled by #define
-ing GREATEST_USE_ABBREVS
to 0.
If you want color output (PASS
in green, FAIL
in red, etc.), you can
pipe the output through the included greenest
script in contrib/
:
$ ./example -v | greenest
(Note that greenest
depends on a Unix-like environment.)
greatest itself doesn't have built-in coloring to stay small and portable.
There is an awk script provided, contrib/entapment
, that converts the
verbose output from the default CLI test runner to TAP version 13
format:
./example -v | contrib/entapment