llvm/utils/lit/TODO

169 lines
7.7 KiB
Plaintext
Raw Normal View History

================
lit TODO Items
================
Infrastructure
==============
1. Change to always load suites, then resolve command line arguments?
Currently we expect each input argument to be a path on disk; we do a
recursive search to find the test suite for each item, but then we only do a
local search based at the input path to find tests. Additionally, for any path
that matches a file on disk we explicitly construct a test instance (bypassing
the formats on discovery implementation).
This has a couple problems:
* The test format doesn't have control over the test instances that result
from file paths.
* It isn't possible to specify virtual tests as inputs. For example, it is not
possible to specify an individual subtest to run with the googletest format.
* The test format doesn't have full control over the discovery of tests in
subdirectories.
Instead, we should move to a model whereby first all of the input specifiers
are resolved to test suites, and then the resolution of the input specifier is
delegated to each test suite. This could take a couple forms:
* We could resolve to test suites, then fully load each test suite, then have
a fixed process to map input specifiers to tests in the test suite
(presumably based on path-in-suite derivations). This has the benefit of
being consistent across all test formats, but the downside of requiring
loading the entire test suite.
* We could delegate all of the resolution of specifiers to the test
suite. This would allow formats that anticipate large test suites to manage
their own resolution for better performance. We could provide a default
resolution strategy that was similar to what we do now (start at subpaths
for directories, but allow the test format control over what happens for
individual tests).
2. Consider move to identifying all tests by path-to-test-suite and then path to
subtest, and don't use test suite names.
Currently the test suite name is presented as part of test names, but it has
no other useful function, and it is something that has to be skipped over to
cut-and-paste a name to subsequently use to rerun a test. If we just
represented each test suite by the path to its suite, then it would allow more
easy cut-and-paste of the test output lines. This has the downside that the
lines might get rather long.
3. Allow 'lit' driver to cooperate with test formats and suites to add options
(or at least sanitize accepted params).
We have started to use the --params method more and more extensively, and it is
cumbersome and error prone. Additionally, there are currently various options
``lit`` honors that should more correctly be specified as belonging to the
ShTest test format.
It would be really nice if we could allow test formats and test suites to add
their own options to be parsed. The difficulty here, of course, is that we
don't know what test formats or test suites are in use until we have parsed the
input specifiers. For test formats we could ostensibly require all the possible
formats to be registered in order to have options, but for test suites we would
certainly have to load the suite before we can query it for what options it
understands.
That leaves us with the following options:
* Currently we could almost get away with parsing the input specifiers without
having done option parsing first (the exception is ``--config-prefix``) but
that isn't a very extensible design.
* We could make a distinction in the command line syntax for test format and
test suite options. For example, we could require something like::
lit -j 1 -sv input-specifier -- --some-format-option
which would be relatively easy to implement with optparser (I think).
* We could allow fully interspersed arguments by first extracting the options
lit knows about and parsing them, then dispatching the remainder to the
formats. This seems the most convenient for users, who are unlikely to care
about (or even be aware of) the distinction between the generic lit
infrastructure and format or suite specific options.
4. Eliminate duplicate execution models for ShTest tests.
Currently, the ShTest format uses tests written with shell-script like syntax,
and executes them in one of two ways. The first way is by converting them into
a bash script and literally executing externally them using bash. The second
way is through the use of an internal shell parser and shell execution code
(built on the subprocess module). The external execution mode is used on most
Unix systems that have bash, the internal execution mode is used on Windows.
Having two ways to do the same thing is error prone and leads to unnecessary
complexity in the testing environment. Additionally, because the mode that
converts scripts to bash doesn't try and validate the syntax, it is possible
to write tests that use bash shell features unsupported by the internal
shell. Such tests won't work on Windows but this may not be obvious to the
developer writing the test.
Another limitation is that when executing the scripts externally, the ShTest
format has no idea which commands fail, or what output comes from which
commands, so this limits how convenient the output of ShTest failures can be
and limits other features (for example, knowing what temporary files were
written).
We should eliminate having two ways of executing the same tests to reduce
platform differences and make it easier to develop new features in the ShTest
module. This is currently blocked on:
* The external execution mode is faster in some situations, because it avoids
being bottlenecked on the GIL. We could fix this by moving to a good
multiprocessing model.
* Some tests in LLVM/Clang are explicitly disabled with the internal shell
(because they use features specific to bash). We would need to rewrite these
tests, or add additional features to the internal shell handling to allow
them to pass.
5. Consider changing core to support setup vs. execute distinction.
Many of the existing test formats are cleanly divided into two phases, once
parses the test format and extracts XFAIL and REQUIRES information, etc., and
the other code actually executes the test.
We could make this distinction part of the core infrastructure and that would
enable a couple things:
* The REQUIREs handling could be lifted to the core, which is nice.
* This would provide a clear place to insert subtest support, because the
setup phase could be responsible for providing subtests back to the
core. That would provide part of the infrastructure to parallelize them, for
example, and would probably interact well with other possible features like
parameterized tests.
* This affords a clean implementation of --no-execute.
* One possible downside could be for test formats that cannot determine their
subtests without having executed the test. Supporting such formats would
either force the test to actually be executed in the setup stage (which
might be ok, as long as the API was explicitly phrased to support that), or
would mean we are forced into supporting subtests as return values from the
execute phase.
Any format can just keep all of its code in execute, presumably, so the only
cost of implementing this is its impact on the API and futures changes.
Miscellaneous
=============
* Move temp directory name into local test config.
* Add --show-unsupported, don't show by default?
* Optionally use multiprocessing.
* Support valgrind in all configs, and LLVM style valgrind.
* Support a timeout / ulimit.
* Create an explicit test suite object (instead of using the top-level
TestingConfig object).