|Version 24 (modified by simonpj, 7 years ago) (diff)|
- GHC Test framework
- Running the testsuite with a compiler other than GHC
- Running individual tests or subdirectories of the testsuite
- Updating tests when the output changes
- Adding a new test
- Format of the test entries
- Sample output files
- The details
- Problems running the testsuite
- The testsuite and branches
GHC Test framework
NOTE: you need GNU make and Python (any version >= 1.5 will probably do) in order to use the testsuite. If you want to run the testsuite in parallel then you need Python 2.5.2 or later. (Avoid Python 2.6.1 as the testsuite tickles a bug in one of the included libraries)
If you have not checked out the test suite, first run:
./darcs-all --testsuite get
To run the test suite against stage 1 of a GHC build in the same source tree:
cd testsuite/tests/ghc-regress make
(from now on, we'll assume that you're in the tests/ghc-regress directory).
To run a fast version of the testsuite, which should complete in under 5 minutes on a fast machine with an optimised GHC build:
To run the testsuite with the stage2 compiler (this is often what you want, because GHCi tests will fail with stage1):
To run the test suite against a different GHC, say ghc-5.04:
To run an individual test or tests (eg. tc054):
(you can also go straight to the directory containing the test and say 'make TEST=tc054' from there, which will save some time).
To run the tests one particular way only (eg. GHCi):
To add specific options to the compiler:
make EXTRA_HC_OPTS='+RTS -K32M -RTS'
To save disk space you can have temporary files deleted after each test:
If you have python 2.5.2 or later then you can run the testsuite in parallel:
For more details, see below.
Running the testsuite with a compiler other than GHC
This doesn't work at the moment, but if it did then it would probably involve something like:
cd testsuite make TEST_HC=nhc98 COMPILER=nhc98
Running individual tests or subdirectories of the testsuite
Most of the subdirectories in the testsuite have a Makefile. In these subdirectories you can use 'make' to run the test driver in two ways:
make -- run all the tests in the current directory make accept -- run the tests, accepting the current output
The following variables may be set on the make command line:
TESTS -- specific tests to run TEST_HC -- compiler to use EXTRA_HC_OPTS -- extra flags to send to the Haskell compiler EXTRA_RUNTEST_OPTS -- extra flags to give the test driver CONFIG -- use a different configuration file COMPILER -- stem of a different configuration file -- from the config directory [default: ghc] WAY -- just this way
The following ways are defined (for GHC, also see the file config/ghc):
normal -- no special options optc -- -O -fvia-C optasm -- -O -fasm profc -- -O -prof -auto-all -fvia-C profasm -- -O -prof -auto-all -fasm ghci -- (run only, not compile) run test under GHCi extcore -- -fext-core optextcore -- -O -fext-core threaded1 -- -threaded -debug threaded2 -- -threaded -O, and +RTS -N2 at run-time hpc -- -fhpc
certain ways are enabled automatically if the GHC build in the local tree supports them. Ways that are enabled this way are optasm, profc, profasm, threaded1, threaded2, and ghci.
Updating tests when the output changes
If the output of a test has changed, but the new output is still correct, you can automatically update the sample output to match the new output like so:
make accept TEST=<test-name>
where <test-name> is the name of the test. In a directory which contains a single test, or if you want to update *all* the tests in the current directory, just omit the 'TEST=<test-name>' part.
Adding a new test
For a test which can be encapsulated in a single source file, follow these steps:
- Find the appropriate place for the test. The GHC regression suite is generally organised in a "white-box" manner: a regression which originally illustrated a bug in a particular part of the compiler is placed in the directory for that part. For example, typechecker regression tests go in the typechecker/ directory, parser tests go in parser/, and so on.
It's not always possible to find a single best place for a test; in those cases just pick one which seems reasonable.
Under each main directory may be up to three subdirectories:
tests which need to compile only
tests which should fail to compile and generate a particular error message
tests which should compile, run with some specific input, and generate a particular output.
We don't always divide the tests up like this, and it's not essential to do so (the directory names have no meaning as far as the test driver is concerned).
- Having found a suitable place for the test, give the test a name. For regression tests, we often just name the test after the bug number (e.g. T2047). Alternatively, follow the convention for the directory in which you place the test: for example, in typecheck/should_compile, tests are named tc001, tc002, and so on. Suppose you name your test T, then you'll have the following files:
The source file containing the test
T.stdin (for tests that run, and optional)
A file to feed the test as standard input when it runs.
T.stdout (for tests that run, and optional)
For tests that run, this file is compared against the standard output generated by the program. If T.stdout does not exist, then the program must not generate anything on stdout.
For tests that run, this file is compared against the standard error generated by the program.
For tests that compile only, this file is compared against the standard error output of the compiler, which is normalised to eliminate bogus differences (eg. absolute pathnames are removed, whitespace differences are ignored, etc.)
- Edit all.T in the relevant directory and add a line for the test. The line is always of the form
test(<name>, <setup>, <test-fn>, <args>)The format of these fields is described in the next section.
A multi-module test is straightforward. It usually goes in a directory of its own (although this isn't essential), and the source files can be named anything you like. The test must have a name, in the same way as a single-module test; and the stdin/stdout/stderr files follow the name of the test as before. In the same directory, place a file 'test.T' containing a line like
test(multimod001, normal, multimod_compile_and_run, \ [ 'Main', '-fglasgow-exts', '', 0 ])
as described above.
For some examples, take a look in tests/ghc-regress/programs.
Format of the test entries
Each test in a test.T file is specified by a line the form
test(<name>, <setup>, <test-fn>, <args>)
The <name> field
<name> is the name of the test, in quotes (' or ").
The <setup> field
<setup> is a function (i.e. any callable object in Python) which allows the options for this test to be changed. There are many pre-defined functions which can be used in this field:
- normal don't change any options from the defaults
- skip skip this test
- skip_if_no_ghci skip unless GHCi is available
- skip_if_fast skip if "fast" is enabled
- omit_ways(ways) skip this test for certain ways
- only_ways(ways) do this test certain ways only
- extra_ways(ways) add some ways which would normally be disabled
- omit_compiler_types(compilers) skip this test for certain compilers
- only_compiler_types(compilers) do this test for certain compilers only
- expect_broken(bug) this test is a expected not to work due to the indicated trac bug number
- expect_broken_for(bug, ways) as expect_broken, but only for the indicated ways
- if_compiler_type(compiler_type, f) Do f, but only for the given compiler type
- if_platform(plat, f) Do f, but only if we are on the specific platform given
- if_tag(tag, f) do f if the compiler has a given tag
- unless_tag(tag, f) do f unless the compiler has a given tag
- set_stdin(file) use a different file for stdin
- no_stdin use no stdin at all (otherwise use /dev/null)
- exit_code(n) expect an exit code of 'n' from the prog
- extra_run_opts(opts) pass some extra opts to the prog
- no_clean don't clean up after this test
- extra_clean(files) extra files to clean after the test has completed
- reqlib(P) requires package P
- req_profiling requires profiling
- ignore_output don't try to compare output
- alone don't run this test in parallel with anything else
- literate look for a .lhs file instead of a .hs file
- c_src look for a .c file
- cmd_prefix(string) prefix this string to the command when run
- normalise_slashes convert backslashes to forward slashes before comparing the output
The following should normally not be used; instead, use the expect_broken* functions above so that the problem doesn't get forgotten about, and when we come back to look at the test later we know whether current behaviour is why we marked it as expected to fail:
- expect_fail this test is an expected failure, i.e. there is a known bug in the compiler, but we don't want to fix it.
- expect_fail_for(ways) expect failure for certain ways
To use more than one modifier on a test, just put them in a list. For example, to expect an exit code of 3 and omit way 'opt', we could use
[ omit_ways(['opt']), exit_code(3) ]
as the <setup> argument.
The <test-fn> field
<test-fn> is a function which describes how the test should be run, and determines the form of <args>. The possible values are:
- compile Just compile the program, the compilation should succeed.
- compile_fail Just compile the program, the compilation should fail (error messages will be in T.stderr). This kind of failure is mandated by the language definition - it does not indicate any bug in the compiler.
- compile_and_run Compile the program and run it, comparing the output against the relevant files.
- multimod_compile Compile a multi-module program (more about multi-module programs below).
- multimod_compile_fail Compile a multi-module program, and expect the compilation to fail with error messages in T.stderr. This kind of failure does not indicate a bug in the compiler.
- multimod_compile_and_run Compile and run a multi-module program.
- compile_and_run_with_prefix Same as compile_and_run, but with command to use to run the execution of the result binary.
- multimod_compile_and_run_with_prefix Same as multimod_compile_and_run, but with command to use to run the execution of the result binary.
- run_command Just run an arbitrary command. The output is checked against T.stdout and T.stderr (unless ignore_output is used), and the stdin and expected exit code can be changed in the same way as for compile_and_run.
- ghci_script Runs the current compiler, passing --interactive and using the specified script as standard input.
The <args> field
<args> is a list of arguments to be passed to <test-fn>.
For compile, compile_fail and compile_and_run, <args> is a list with a single string which contains extra compiler options with which to run the test. eg.
test('tc001', normal, compile, ['-fglasgow-exts'])
would pass the flag -fglasgow-exts to the compiler when compiling tc001.
The multimod_ versions of compile and compile_and_run expect an extra argument on the front of the list: the name of the top module in the program to be compiled (usually this will be 'Main').
Sample output files
Normally, the sample stdout and stderr for a test T go in the files T.stdout and T.stderr respectively. However, sometimes a test may generate different output depending on the platform, compiler, compiler version, or word-size. For this reason the test driver looks for sample output files using this pattern:
Any combination of the optional extensions may be given, but they must be in the order specified. The most specific output file that matches the current configuration will be selected; for example if the platform is i386-unknown-mingw32 then T.stderr-i386-unknown-mingw32 will be picked in preference to T.stderr.
Another common example is to give different sample output for an older compiler version. For example, the sample stderr for GHC 6.8.x would go in the file T.stderr-ghc-6.8.
The test suite driver is just a set of Python scripts, as are all of the .T files in the test suite. The driver (driver/runtests.py) first searches for all the .T files it can find, and then proceeds to execute each one, keeping a track of the number of tests run, and which ones succeeded and failed.
The script runtests.py takes several options:
<file> is just a file containing Python code which is executed. The purpose of this option is so that a file containing settings for the configuration options can be specified on the command line. Multiple --config options may be given.
<dir> is the directory below which to search for .T files to run.
In addition to dumping the test summary to stdout, also put it in <file>. (stdout also gets a lot of other output when running a series of tests, so redirecting it isn't always the right thing).
Only run tests named <test> (multiple --only options can be given). Useful for running a single test from a .T file containing multiple tests.
executes the Python statement <stmt> before running any tests. The main purpose of this option is to allow certain configuration options to be tweaked from the command line; for example, the build system adds '-e config.accept=1' to the command line when 'make accept' is invoked.
Most of the code for running tests is located in driver/testlib.py. Take a look.
There is a single Python class (TestConfig) containing the global configuration for the test suite. It contains information such as the kind of compiler being used, which flags to give it, which platform we're running on, and so on. The idea is that each platform and compiler would have its own file containing assignments for elements of the configuration, which are sourced by passing the appropriate --config options to the test driver. For example, the GHC configuration is contained in the file config/ghc.
A .T file can obviously contain arbitrary Python code, but the general idea is that it contains a sequence of calls to the function test(), which resides in testlib.py. As described above, test() takes four arguments:
test(<name>, <opt-fn>, <test-fn>, <args>)
The function <opt-fn> is allowed to be any Python callable object, which takes a single argument of type TestOptions. TestOptions is a class containing options which affect the way that the current test is run: whether to skip it, whether to expect failure, extra options to pass to the compiler, etc. (see testlib.py for the definition of the TestOptions class). The idea is that the <opt-fn> function modifies the TestOptions object that it is passed. For example, to expect failure for a test, we might do this in the .T file:
def fn(opts): opts.expect = 'fail' test(test001, fn, compile, [''])
so when fn is called, it sets the instance variable "expect" in the instance of TestOptions passed as an argument, to the value 'fail'. This indicates to the test driver that the current test is expected to fail.
Some of these functions, such as the one above, are common, so rather than forcing every .T file to redefine them, we provide canned versions. For example, the provided function expect_fail does the same as fn in the example above. See testlib.py for all the canned functions we provide for <opt-fn>.
The argument <test-fn> is a function which performs the test. It takes three or more arguments:
<test-fn>( <name>, <way>, ... )
where <name> is the name of the test, <way> is the way in which it is to be run (eg. opt, optasm, prof, etc.), and the rest of the arguments are constructed from the list <args> in the original call to test(). The following <test-fn>s are provided at the moment:
compile compile_fail compile_and_run multimod_compile multimod_compile_fail multimod_compile_and_run run_command run_command_ignore_output ghci_script
and obviously others can be defined. The function should return either 'pass' or 'fail' indicating that the test passed or failed respectively.
Problems running the testsuite
- If the test suite fails mysteriously, make sure that the timeout utility is working properly. This Haskell utility is compiled with the stage 1 compiler and invoked by the python driver, which does not print a nice error report if the utility fails. This can happen if, for example, the compiler produces bogus binaries. A workaround is to compile timeout with a stable ghc.
The testsuite and branches
It is not clear what to do with the testsuite when branching a compiler; should the testsuite also be branched?
If it is not branched then we have the problem that, given a set of tests
test(tc1, ...) test(tc2, ...) test(tc3, ...)
if we add first one test, and then another to the HEAD
test(tc1, ...) test(tc2, ...) test(tc3, ...) test(tc4, ...) test(tc5, ...)
and we want to merge tc5 but not tc4 to the branch then the merge has to be done by hand, as the patch for tc5 depends on the patch for tc4, although most of the files in the patches (tc5.hs etc) are disjoint.
On the other hand, if it is not branched then any changes in test output mean we need to add extra logic to the test definitions, e.g.
test(tc5, namebase_if_compiler_lt('ghc','6.9', 'tc5-6.8'), ...)