[Patch] Adding unit tests to LLVM

(Forwarding this to llvm-dev)

This patch adds a unit test framework to LLVM, along with a sample unit test for DenseMap. I don’t expect this patch to be accepted as-is, this is mainly a trial balloon and proof of concept.

Some notes about the patch:

  1. For the testing framework, I went with Google Test, since it’s the one I have the most experience with. I fully expect an extended bikeshed discussion to result from this.

  2. Both the test framework and the tests are optional build targets, they will not be built with the normal “make all”. To build and run the unit tests, use “make check-unit”.

  3. I did not actually include the testing framework in the patch; It will need to be checked in separately. There are two approaches to this. One approach is to use the svn:external feature to create a link to the googletest svn repository from the LLVM svn repository. The other approach is to take a snapshot of googletest and check it in to the LLVM repository.

The GoogleTest tar archive is here: http://code.google.com/p/googletest/downloads/list. I’ve located it within the LLVM source tree in the location “third-party/googletest”.

  1. I’ve included makefiles for building the test framework and unit tests. I had to do a bit of makefile hacking to get this to work, suggestions on how to improve this are welcome.

unittest.patch (7.42 KB)

Nice work!

(Forwarding this to llvm-dev)

This patch adds a unit test framework to LLVM, along with a sample unit test for DenseMap. I don’t expect this patch to be accepted as-is, this is mainly a trial balloon and proof of concept.

Some notes about the patch:

  1. For the testing framework, I went with Google Test, since it’s the one I have the most experience with. I fully expect an extended bikeshed discussion to result from this.

+1.

Google test is pretty nice (not that I’m unbiased, as I am a gtest contributor :slight_smile:

  1. Both the test framework and the tests are optional build targets, they will not be built with the normal “make all”. To build and run the unit tests, use “make check-unit”.

“make unittest”? I would never guess check-unit.

  1. I did not actually include the testing framework in the patch; It will need to be checked in separately. There are two approaches to this. One approach is to use the svn:external feature to create a link to the googletest svn repository from the LLVM svn repository. The other approach is to take a snapshot of googletest and check it in to the LLVM repository.

The GoogleTest tar archive is here: http://code.google.com/p/googletest/downloads/list. I’ve located it within the LLVM source tree in the location “third-party/googletest”.

I suggest checking a specific version of Google test into the source. There isn’t much clear benfit to tracking upstream gtest. On the other hand, there is benefit to not having mysterious breakages due to upstream (as unlikely as that is due to gtest API guarentees). Also, for projects adding the gtest source directly to their tree, there is a special gtest-all.c which contains the entire gtest codebase in a single file (it’s not that much code).

Keir

So, any response on whether this patch is likely to be accepted? I'm willing to substantially rework things if requested.

I should note that part of my motivation for doing this is that I have been wanting to do some work on the various LLVM container classes; However, I strongly feel that without some kind of unit test framework in place such work would mostly be a waste of time.

Talin wrote:

Hi Talin,

I just got around to looking at this. Sorry for the delay.

I like the idea of unit tests *A LOT*. Adequate testing has been a major sticking point for LLVM for a long time. In my opinion, once the color of the bike shed is agreed upon ;-), this will be a welcome addition. I have never worked with Google's unit testing infrastucture so I don't know how it works or if its license is compatible with ours, etc.

I did a very quick look at your code and it seems fine for a first stab. Let's try to get everyone's input before you do a ton of work. But I for one think that it will be accepted when people are satisfied with the form it takes.

-bw

Thanks for taking a look!

The googletest framework is open source, and licensed under the "new BSD" license. I'll leave it to the license lawyers to determine compatibility.

Introductory docs for googletest are here: http://code.google.com/p/googletest/wiki/GoogleTestPrimer

You might also want to read this (somewhat out of date) comparison of C++ unit testing frameworks:

   "Exploring the C++ Unit Testing Jungle" http://gamesfromwithin.com/?p=29

Note that this doesn't include googletest or UnitTest++, both of which were developed after the article was written. (The latter was developed by the author of the above article.) However, the article does give a good overview of what is desirable in a unit testing framework.

Also note that one additional requirement of a unit test framework for LLVM is that it should not rely on RTTI or exceptions, since LLVM is built with both of those features disabled.

Bill Wendling wrote:

(Forwarding this to llvm-dev)

This patch adds a unit test framework to LLVM, along with a sample unit test
for DenseMap. I don't expect this patch to be accepted as-is, this is mainly
a trial balloon and proof of concept.

I think this is a great idea! As Keir already noted, I would also
agree with LLVM snapshotting a copy of googletest, but I think it
should live in llvm/test/googletest (rather than top-level), since
it's not going to be linked into anything outside of llvm/test.

On the assumption that we'll end up using googletest, here are some
comments on your patch:

* s/Insure/Ensure/

* LLVM uses "llvm/foo.h" for inclusion rather than <llvm/foo.h>
* You should use the same format for gtest headers

* If reverse iteration isn't supported, you should either have an
ASSERT_DEATH() on the decrement, or not have the code there (that's
commented out) at all.

* Instead of this:
EXPECT_TRUE(uintMap[0] == 1);
you should use EXPECT_EQ()

* Instead of this:
EXPECT_TRUE(uintMap.find(0u) == uintMap.begin());
is it possible to use EXPECT_EQ() as well?

* In this test:
TEST_F(DenseMapTest, IterationTest) {
you use the array "int numbers[100];" as an array of booleans; why not
make it "bool visited[100];" to make clear what it for, how it's used,
and maybe be slightly more efficient?

1) For the testing framework, I went with Google Test, since it's the one I
have the most experience with. I fully expect an extended bikeshed
discussion to result from this.

2) Both the test framework and the tests are optional build targets, they
will not be built with the normal "make all". To build and run the unit
tests, use "make check-unit".

I am with Keir on this one: "check-unit" isn't the most intuitive name
to me -- "unittest" sounds fine, but this isn't a big deal to me.

I'll also volunteer to help maintain the LLVM branch of googletest and
review the unittests being checked in, as I also have experience with
using googletest, and I think LLVM could derive a lot of value from
unittesting.

Misha

(Forwarding this to llvm-dev)

This patch adds a unit test framework to LLVM, along with a sample unit test
for DenseMap. I don’t expect this patch to be accepted as-is, this is mainly
a trial balloon and proof of concept.

I think this is a great idea! As Keir already noted, I would also
agree with LLVM snapshotting a copy of googletest, but I think it
should live in llvm/test/googletest (rather than top-level), since
it’s not going to be linked into anything outside of llvm/test.

On the assumption that we’ll end up using googletest, here are some
comments on your patch:

  • s/Insure/Ensure/

  • LLVM uses “llvm/foo.h” for inclusion rather than <llvm/foo.h>

  • You should use the same format for gtest headers

  • If reverse iteration isn’t supported, you should either have an
    ASSERT_DEATH() on the decrement, or not have the code there (that’s
    commented out) at all.

  • Instead of this:
    EXPECT_TRUE(uintMap[0] == 1);
    you should use EXPECT_EQ()

  • Instead of this:
    EXPECT_TRUE(uintMap.find(0u) == uintMap.begin());
    is it possible to use EXPECT_EQ() as well?

  • In this test:
    TEST_F(DenseMapTest, IterationTest) {
    you use the array “int numbers[100];” as an array of booleans; why not
    make it “bool visited[100];” to make clear what it for, how it’s used,
    and maybe be slightly more efficient?

  1. For the testing framework, I went with Google Test, since it’s the one I
    have the most experience with. I fully expect an extended bikeshed
    discussion to result from this.

  2. Both the test framework and the tests are optional build targets, they
    will not be built with the normal “make all”. To build and run the unit
    tests, use “make check-unit”.

I am with Keir on this one: “check-unit” isn’t the most intuitive name
to me – “unittest” sounds fine, but this isn’t a big deal to me.

I’ll also volunteer to help maintain the LLVM branch of googletest and
review the unittests being checked in, as I also have experience with
using googletest, and I think LLVM could derive a lot of value from
unittesting.

If we make useful changes to gtest, I volunteer to push them upstream.

Keir

Just a curiosity question, why push for gtest vs Boost Test or a different test suite?
I normally use Boost, and their test suite, so I’m more familiar with that. So I was wondering is one better then the other, or is it just that someone makes a patch for it?

Regards
Mark Kromis

Misha Brukman wrote:

  

(Forwarding this to llvm-dev)

This patch adds a unit test framework to LLVM, along with a sample unit test
for DenseMap. I don't expect this patch to be accepted as-is, this is mainly
a trial balloon and proof of concept.
    
I think this is a great idea! As Keir already noted, I would also
agree with LLVM snapshotting a copy of googletest, but I think it
should live in llvm/test/googletest (rather than top-level), since
it's not going to be linked into anything outside of llvm/test.
  

So as far as putting things in llvm/test, I have a question - the makefile in that directory contains a whole bunch of rules for interfacing with DejaGNU. The unit tests don't (and, I think, shouldn't) require any dependence on DejaGNU -- in fact, I'm hoping it will be possible to run the unit tests with only the base LLVM package, without all of the additional package installations required to run the other LLVM tests. I'm just wondering if it will be a problem organizing the makefile so that the unit tests don't have any dependence on the other tests.

On the assumption that we'll end up using googletest, here are some
comments on your patch:

* s/Insure/Ensure/

* LLVM uses "llvm/foo.h" for inclusion rather than <llvm/foo.h>
* You should use the same format for gtest headers

* If reverse iteration isn't supported, you should either have an
ASSERT_DEATH() on the decrement, or not have the code there (that's
commented out) at all.
  

I'll probably just remove it.

* Instead of this:
EXPECT_TRUE(uintMap[0] == 1);
you should use EXPECT_EQ()
  

Sure.

* Instead of this:
EXPECT_TRUE(uintMap.find(0u) == uintMap.begin());
is it possible to use EXPECT_EQ() as well?
  

In order to use EXPECT_EQ, both arguments have to be printable, although you can make anything printable by adding the necessary stream operator overloads. In this particular case, I decided that the printed representation of an iterator wouldn't be meaningful, so I didn't bother defining those overloads.

* In this test:
TEST_F(DenseMapTest, IterationTest) {
you use the array "int numbers[100];" as an array of booleans; why not
make it "bool visited[100];" to make clear what it for, how it's used,
and maybe be slightly more efficient?
  

OK

2008/12/27 Mark Kromis <greybird@mac.com>

Just a curiosity question, why push for gtest vs Boost Test or a different test suite?
I normally use Boost, and their test suite, so I’m more familiar with that. So I was wondering is one better then the other, or is it just that someone makes a patch for it?

It seems that the original patch author is most familiar with gtest (as am I). I looked at boost’s testing facilities, as described here: http://gamesfromwithin.com/?p=29#boost and it seems to be quite more verbose for setting up test suites, with what looks like a bit of boilerplate code.

Compare this with the way new tests are written in gtest with TEST_F() macro – see the original patch by Talin. Note that there’s no suite-setup code in your unittest file in gtest, which is required in boost. In fact, the only gtest-related code is the #include statement, not counting the common main() that’s linked into every unittest binary – unless you want to do some custom global-setup code, and hence you’ll write your own main().

Misha

2008/12/27 Mark Kromis <greybird@mac.com>

Just a curiosity question, why push for gtest vs Boost Test or a different test suite?
I normally use Boost, and their test suite, so I’m more familiar with that. So I was wondering is one better then the other, or is it just that someone makes a patch for it?

I looked more into Boost.Test to see what’s in it. Boost.Test doesn’t seem to be stand-alone – I don’t see a way to use Boost.Test without importing some other chunks of Boost that the testing library depends on. While Boost is a fine set of libraries, I don’t think we want to increase the LLVM distribution by sizeof(Boost) just to enable unittesting, nor do we want to spend the time on maintaining a subset of Boost that’s “just enough” to build and use the unittest library, along a modified configure/build process that Boost wants to use (Boost.Build? Boost.Jam?).

Boost also seems to want to use exceptions, and LLVM does not want to. I’m not sure if there would be some difficulties in running a build where some libraries are compiled with no exceptions, some with, and the results are linked together. At the best case, it would complicate our build system to be able to support different set of flags for building LLVM libraries vs. Boost.Test (and the rest of Boost that we import).

Sample usage of Boost.Test: http://svn.boost.org/svn/boost/trunk/libs/test/example/unit_test_example_12.cpp
Note the code at the end setting up the test suite – this is boilerplate code that I think shouldn’t be necessary to setup and run tests.

Google Test, on the other hand, has no external dependencies, and is distributed as a dozen of .h/.cc files; supports Makefile, SCons, and Xcode; and doesn’t use exceptions or RTTI.

Sample usage of GTest: http://code.google.com/p/googletest/source/browse/trunk/samples/sample5_unittest.cc
GTest-specific LOC besides the #include statement: 0.

Note that I’m not counting main() for either Boost or GTest, because both provide a standard main() for use with almost all test files.

Misha

Indeed, Boost.Test requires approximately 500 header files, minimally.

— Gordon

2008/12/27 Mark Kromis <greybird@mac.com>

Just a curiosity question, why push for gtest vs Boost Test or a different test suite?
I normally use Boost, and their test suite, so I’m more familiar with that. So I was wondering is one better then the other, or is it just that someone makes a patch for it?

I looked more into Boost.Test to see what’s in it. Boost.Test doesn’t seem to be stand-alone – I don’t see a way to use Boost.Test without importing some other chunks of Boost that the testing library depends on. While Boost is a fine set of libraries, I don’t think we want to increase the LLVM distribution by sizeof(Boost) just to enable unittesting, nor do we want to spend the time on maintaining a subset of Boost that’s “just enough” to build and use the unittest library, along a modified configure/build process that Boost wants to use (Boost.Build? Boost.Jam?).

So are you planning on maintaining whatever test system, or just have them as a pre-requisite. For example are you going to have the gtest incorporated, or have them install it separately first? I was under the impression that the user would have to install gtest first.

Boost also seems to want to use exceptions, and LLVM does not want to. I’m not sure if there would be some difficulties in running a build where some libraries are compiled with no exceptions, some with, and the results are linked together. At the best case, it would complicate our build system to be able to support different set of flags for building LLVM libraries vs. Boost.Test (and the rest of Boost that we import).

http://www.boost.org/doc/libs/1_37_0/libs/utility/throw_exception.html
#define BOOST_NO_EXCEPTIONS

Sample usage of Boost.Test: http://svn.boost.org/svn/boost/trunk/libs/test/example/unit_test_example_12.cpp
Note the code at the end setting up the test suite – this is boilerplate code that I think shouldn’t be necessary to setup and run tests.

http://svn.boost.org/svn/boost/trunk/libs/test/example/unit_test_example_01.cpp
My test cases are not that in-depth, I’m much closer to sample 1. I haven’t found a reason to go that crazy yet.

Google Test, on the other hand, has no external dependencies, and is distributed as a dozen of .h/.cc files; supports Makefile, SCons, and Xcode; and doesn’t use exceptions or RTTI.

Gtest is much more lightweight, no comparison there. I know that llvm is not very good with exceptions, but should a test case system support that?

Sample usage of GTest: http://code.google.com/p/googletest/source/browse/trunk/samples/sample5_unittest.cc
GTest-specific LOC besides the #include statement: 0.

I think it links to a library as well.

Note that I’m not counting main() for either Boost or GTest, because both provide a standard main() for use with almost all test files.

Misha


LLVM Developers mailing list
LLVMdev@cs.uiuc.edu http://llvm.cs.uiuc.edu
http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev

Also for a note of reference, your links to the examples are the most advanced samples. So boost can do more, thus has more weight/bloat behind it.

Were the other test kits looked at? Is gtest the best solution for the project.

Is this something your planning as putting in the tree, thus require pulling in changes from google (license allowing), or does user need to have the libraries/headers pre-installed?

My question was not to cause a battle, but I wanted to be sure we were using the right test kit, and not just picking one just because. For example gtest is very light weight test kit, that can do the job, but will the tests outgrow what the test kit can do, and cause a conversion to a more advanced one later?

Regards,
Mark Kromis

2008/12/27 Mark Kromis <greybird@mac.com>

Just a curiosity question, why push for gtest vs Boost Test or a different test suite?
I normally use Boost, and their test suite, so I’m more familiar with that. So I was wondering is one better then the other, or is it just that someone makes a patch for it?

I looked more into Boost.Test to see what’s in it. Boost.Test doesn’t seem to be stand-alone – I don’t see a way to use Boost.Test without importing some other chunks of Boost that the testing library depends on. While Boost is a fine set of libraries, I don’t think we want to increase the LLVM distribution by sizeof(Boost) just to enable unittesting, nor do we want to spend the time on maintaining a subset of Boost that’s “just enough” to build and use the unittest library, along a modified configure/build process that Boost wants to use (Boost.Build? Boost.Jam?).

So are you planning on maintaining whatever test system, or just have them as a pre-requisite. For example are you going to have the gtest incorporated, or have them install it separately first? I was under the impression that the user would have to install gtest first.

Boost also seems to want to use exceptions, and LLVM does not want to. I’m not sure if there would be some difficulties in running a build where some libraries are compiled with no exceptions, some with, and the results are linked together. At the best case, it would complicate our build system to be able to support different set of flags for building LLVM libraries vs. Boost.Test (and the rest of Boost that we import).

http://www.boost.org/doc/libs/1_37_0/libs/utility/throw_exception.html
#define BOOST_NO_EXCEPTIONS

Sample usage of Boost.Test: http://svn.boost.org/svn/boost/trunk/libs/test/example/unit_test_example_12.cpp
Note the code at the end setting up the test suite – this is boilerplate code that I think shouldn’t be necessary to setup and run tests.

http://svn.boost.org/svn/boost/trunk/libs/test/example/unit_test_example_01.cpp
My test cases are not that in-depth, I’m much closer to sample 1. I haven’t found a reason to go that crazy yet.

Google Test, on the other hand, has no external dependencies, and is distributed as a dozen of .h/.cc files; supports Makefile, SCons, and Xcode; and doesn’t use exceptions or RTTI.

Gtest is much more lightweight, no comparison there. I know that llvm is not very good with exceptions, but should a test case system support that?

Sample usage of GTest: http://code.google.com/p/googletest/source/browse/trunk/samples/sample5_unittest.cc
GTest-specific LOC besides the #include statement: 0.

I think it links to a library as well.

Note that I’m not counting main() for either Boost or GTest, because both provide a standard main() for use with almost all test files.

Misha


LLVM Developers mailing list
LLVMdev@cs.uiuc.edu http://llvm.cs.uiuc.edu
http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev

Also for a note of reference, your links to the examples are the most advanced samples. So boost can do more, thus has more weight/bloat behind it.

Were the other test kits looked at? Is gtest the best solution for the project.

Is this something your planning as putting in the tree, thus require pulling in changes from google (license allowing), or does user need to have the libraries/headers pre-installed?

Including it in the tree is the most reasonable thing to do. No point in inconveniencing the user over tiny libraries with liberal licenses.

My question was not to cause a battle, but I wanted to be sure we were using the right test kit, and not just picking one just because. For example gtest is very light weight test kit, that can do the job, but will the tests outgrow what the test kit can do, and cause a conversion to a more advanced one later?

I am not sure what advanced features you are thinking of that gtest doesn’t already offer; it’s pretty sophisticated despite it’s small size. Can you give an example of something gtest doesn’t already support? Newer releases added support for data parameterized tests, which is really useful. Type parameterized tests are being discussed on the mailing list. All while maintaining a small codebase!

One other nice feature of using gtest is that it integrates with gmock, one of the only really good C++ mocking libraries available. It was just recently released.

http://code.google.com/p/googlemock

Keir

2008/12/27 Mark Kromis <greybird@mac.com>

So are you planning on maintaining whatever test system, or just have them as a pre-requisite. For example are you going to have the gtest incorporated, or have them install it separately first? I was under the impression that the user would have to install gtest first.

The current plan is to check in the unittest library into LLVM and build it as part of the test process, so the user doesn’t have to do anything separately.

Boost also seems to want to use exceptions, and LLVM does not want to. I’m not sure if there would be some difficulties in running a build where some libraries are compiled with no exceptions, some with, and the results are linked together. At the best case, it would complicate our build system to be able to support different set of flags for building LLVM libraries vs. Boost.Test (and the rest of Boost that we import).

http://www.boost.org/doc/libs/1_37_0/libs/utility/throw_exception.html
#define BOOST_NO_EXCEPTIONS

Thanks for the pointer.

Sample usage of Boost.Test: http://svn.boost.org/svn/boost/trunk/libs/test/example/unit_test_example_12.cpp
Note the code at the end setting up the test suite – this is boilerplate code that I think shouldn’t be necessary to setup and run tests.

http://svn.boost.org/svn/boost/trunk/libs/test/example/unit_test_example_01.cpp
My test cases are not that in-depth, I’m much closer to sample 1. I haven’t found a reason to go that crazy yet.

You would need to use fixtures if you are testing classes and their interactions, otherwise each of your test functions will re-do everything that’s in the fixture’s SetUp() method – why would you want to repeat the same setup code for every test case?

Google Test, on the other hand, has no external dependencies, and is distributed as a dozen of .h/.cc files; supports Makefile, SCons, and Xcode; and doesn’t use exceptions or RTTI.

Gtest is much more lightweight, no comparison there. I know that llvm is not very good with exceptions, but should a test case system support that?

GTest allows testing for exceptions, it just doesn’t require them to work properly:
http://code.google.com/p/googletest/wiki/GoogleTestAdvancedGuide#Exception_Assertions

Sample usage of GTest: http://code.google.com/p/googletest/source/browse/trunk/samples/sample5_unittest.cc
GTest-specific LOC besides the #include statement: 0.

I think it links to a library as well.

Yes, that’s true – just about any unittesting library will have that requirement, that wasn’t part of the comparison. I was pointing out how much C++ a user has to write in their test files for a reasonable test with fixtures.

Also for a note of reference, your links to the examples are the most advanced samples. So boost can do more, thus has more weight/bloat behind it.

Gordon Henriksen pointed out earlier in the thread that Boost.Test will require pulling in 500 headers from Boost just to work, so that’s pretty heavy weight. Also, I was comparing apples-to-apples, not taking the simplest example of GTest and the most complex of Boost – the two examples I chose were doing the same thing, i.e., tests with fixtures.

What does Boost provide that GTest does not, that you think LLVM needs, such that it’s worth importing the entire Boost distribution into the LLVM tree?

Were the other test kits looked at? Is gtest the best solution for the project.

There are dozens of C++ unittesting libraries out there, I admit I haven’t looked at all of them. I don’t know which ones are still under active development or use, but I (and a few other posters) are familiar with and have used GTest enough to say that it’s very well suited for LLVM.

Is this something your planning as putting in the tree, thus require pulling in changes from google (license allowing), or does user need to have the libraries/headers pre-installed?

I already answered this question above (first option: in the tree). GTest is under the BSD license, which is compatible with LLVM (though I am not a copyright lawyer).

For example gtest is very light weight test kit, that can do the job, but will the tests outgrow what the test kit can do, and cause a conversion to a more advanced one later?

As soon as LLVM outgrows what GTest can do, we can easily write wrapper macros to forward from GTest-style macros to the other, more suitable framework; however, I think you’re conflating the two concepts of “light-weight” and “trivial” – GTest is light-weight, but not trivial.

Misha

…snip…

Also for a note of reference, your links to the examples are the most advanced samples. So boost can do more, thus has more weight/bloat behind it.

Were the other test kits looked at? Is gtest the best solution for the project.

Is this something your planning as putting in the tree, thus require pulling in changes from google (license allowing), or does user need to have the libraries/headers pre-installed?

Including it in the tree is the most reasonable thing to do. No point in inconveniencing the user over tiny libraries with liberal licenses.

I didn’t get to look at the license.

My question was not to cause a battle, but I wanted to be sure we were using the right test kit, and not just picking one just because. For example gtest is very light weight test kit, that can do the job, but will the tests outgrow what the test kit can do, and cause a conversion to a more advanced one later?

I am not sure what advanced features you are thinking of that gtest doesn’t already offer; it’s pretty sophisticated despite it’s small size. Can you give an example of something gtest doesn’t already support? Newer releases added support for data parameterized tests, which is really useful. Type parameterized tests are being discussed on the mailing list. All while maintaining a small codebase!

No not really. I was just curious on why this test kit was picked. Thats all.

One other nice feature of using gtest is that it integrates with gmock, one of the only really good C++ mocking libraries available. It was just recently released.

http://code.google.com/p/googlemock

Keir

Just briefly looked at the googlemock site and it sys about needed tr1 library. Third paragraph in System Requirements. It just struck me as amusing that visual c++ people need boost then.

I’ll look into this more.

Thank you for your kind explanations.

Regards
Mark Kromis

2008/12/27 Mark Kromis <greybird@mac.com>

So are you planning on maintaining whatever test system, or just have them as a pre-requisite. For example are you going to have the gtest incorporated, or have them install it separately first? I was under the impression that the user would have to install gtest first.

The current plan is to check in the unittest library into LLVM and build it as part of the test process, so the user doesn’t have to do anything separately.

Boost also seems to want to use exceptions, and LLVM does not want to. I’m not sure if there would be some difficulties in running a build where some libraries are compiled with no exceptions, some with, and the results are linked together. At the best case, it would complicate our build system to be able to support different set of flags for building LLVM libraries vs. Boost.Test (and the rest of Boost that we import).

http://www.boost.org/doc/libs/1_37_0/libs/utility/throw_exception.html
#define BOOST_NO_EXCEPTIONS

Thanks for the pointer.

Sample usage of Boost.Test: http://svn.boost.org/svn/boost/trunk/libs/test/example/unit_test_example_12.cpp
Note the code at the end setting up the test suite – this is boilerplate code that I think shouldn’t be necessary to setup and run tests.

http://svn.boost.org/svn/boost/trunk/libs/test/example/unit_test_example_01.cpp
My test cases are not that in-depth, I’m much closer to sample 1. I haven’t found a reason to go that crazy yet.

You would need to use fixtures if you are testing classes and their interactions, otherwise each of your test functions will re-do everything that’s in the fixture’s SetUp() method – why would you want to repeat the same setup code for every test case?

They have something similar, non obvious though.
http://gamesfromwithin.com/?p=29#comment-47

Google Test, on the other hand, has no external dependencies, and is distributed as a dozen of .h/.cc files; supports Makefile, SCons, and Xcode; and doesn’t use exceptions or RTTI.

Gtest is much more lightweight, no comparison there. I know that llvm is not very good with exceptions, but should a test case system support that?

GTest allows testing for exceptions, it just doesn’t require them to work properly:
http://code.google.com/p/googletest/wiki/GoogleTestAdvancedGuide#Exception_Assertions

ah ok. I misunderstood it then.

Sample usage of GTest: http://code.google.com/p/googletest/source/browse/trunk/samples/sample5_unittest.cc
GTest-specific LOC besides the #include statement: 0.

I think it links to a library as well.

Yes, that’s true – just about any unittesting library will have that requirement, that wasn’t part of the comparison. I was pointing out how much C++ a user has to write in their test files for a reasonable test with fixtures.

Also for a note of reference, your links to the examples are the most advanced samples. So boost can do more, thus has more weight/bloat behind it.

Gordon Henriksen pointed out earlier in the thread that Boost.Test will require pulling in 500 headers from Boost just to work, so that’s pretty heavy weight. Also, I was comparing apples-to-apples, not taking the simplest example of GTest and the most complex of Boost – the two examples I chose were doing the same thing, i.e., tests with fixtures.

ok. I havn’t used fixtures for unit testing. Most of the classes don’t really cross each other that much.

What does Boost provide that GTest does not, that you think LLVM needs, such that it’s worth importing the entire Boost distribution into the LLVM tree?

Were the other test kits looked at? Is gtest the best solution for the project.

There are dozens of C++ unittesting libraries out there, I admit I haven’t looked at all of them. I don’t know which ones are still under active development or use, but I (and a few other posters) are familiar with and have used GTest enough to say that it’s very well suited for LLVM.

Is this something your planning as putting in the tree, thus require pulling in changes from google (license allowing), or does user need to have the libraries/headers pre-installed?

I already answered this question above (first option: in the tree). GTest is under the BSD license, which is compatible with LLVM (though I am not a copyright lawyer).

For example gtest is very light weight test kit, that can do the job, but will the tests outgrow what the test kit can do, and cause a conversion to a more advanced one later?

As soon as LLVM outgrows what GTest can do, we can easily write wrapper macros to forward from GTest-style macros to the other, more suitable framework; however, I think you’re conflating the two concepts of “light-weight” and “trivial” – GTest is light-weight, but not trivial.

I was just concerned that it was too light-weight. Not really dealing with gtest is why I asked the question.

Misha


LLVM Developers mailing list
LLVMdev@cs.uiuc.edu http://llvm.cs.uiuc.edu
http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev

Regards,
Mark Kromis

They recently shipped TR1 for VS2008:
http://www.microsoft.com/downloads/details.aspx?FamilyId=D466226B-8DAB-445F-A7B4-448B326C48E7&displaylang=en

2008/12/27 Mark Kromis <greybird@mac.com>

Is this something your planning as putting in the tree, thus require pulling in changes from google (license allowing), or does user need to have the libraries/headers pre-installed?

Including it in the tree is the most reasonable thing to do. No point in inconveniencing the user over tiny libraries with liberal licenses.

I didn’t get to look at the license.

http://code.google.com/p/googletest/http://www.opensource.org/licenses/bsd-license.php .

One other nice feature of using gtest is that it integrates with gmock, one of the only really good C++ mocking libraries available. It was just recently released.

Just briefly looked at the googlemock site and it sys about needed tr1 library. Third paragraph in System Requirements. It just struck me as amusing that visual c++ people need boost then.

This is getting slightly off-topic (we’re now talking about mocking libraries, rather than unittesting), but Visual Studio 2008 has a free add-on pack to add TR1 support: http://msdn.microsoft.com/en-us/visualc/default.aspx (see the downloads section).

Mark Kromis wrote:

Just a curiosity question, why push for gtest vs Boost Test or a different test suite?
I normally use Boost, and their test suite, so I'm more familiar with that. So I was wondering is one better then the other, or is it just that someone makes a patch for it?

I did a very minimal amount of research into unit testing frameworks before I started using gtest for my own work. Since then I've become quite familiar with gtest's features, which made it a natural choice for me. I presume that other testing frameworks will have their own defenders on this list, and I am certainly willing to listen to what they have to say.

The unit test comparison article mentioned earlier (http://gamesfromwithin.com/?p=29) is a good starting point for evaluating different unit test frameworks. Although gtest is too new to have been mentioned in the article directly, if you actually compare gtest's features against the various criteria used by the author of the article, it comes off quite well.

Note that the author of the article later went and created his own unit test framework (UnitTest++), based on his experience with all of the other frameworks he tested. This was the only framework that I seriously looked at other than gtest, and I found it to be a bit too minimal for what I needed.

I looked more into Boost.Test to see what's in it. Boost.Test doesn't seem to be stand-alone -- I don't see a way to use Boost.Test without importing some other chunks of Boost that the testing library depends on. While Boost is a fine set of libraries, I don't think we want to increase the LLVM distribution by sizeof(Boost) just to enable unittesting, nor do we want to spend the time on maintaining a subset of Boost that's "just enough" to build and use the unittest library, along a modified configure/build process that Boost wants to use (Boost.Build? Boost.Jam?).

Although I haven't actually tried Boost.Test, I kind of figured that this would be the case - that you pretty much have to drink the "Boost Kool-Aid" in order to use it.

So are you planning on maintaining whatever test system, or just have them as a pre-requisite. For example are you going to have the gtest incorporated, or have them install it separately first? I was under the impression that the user would have to install gtest first.

So the plan is to take a snapshot of gtest and check it in to the LLVM tree, rather than have it installed separately. I was able to integrate gtest into LLVM's build system fairly easily, as gtest is designed to be integrated into a foreign build system - basically I just ignored the makefile that comes with gtest, and wrote an LLVM-style makefile rule for it. There's a special source file in gtest which includes all other sources that is intended for just such a purpose. I did not need to modify the gtest sources in any way.

This means that keeping the gtest snapshot up to date will be trivial, since it will only require copying in the latest gtest snapshot and checking it in to LLVM - presuming that gtest remains backwards compatible, which I assume it will.

Licensing-wise, both LLVM and gtest are distributed under a fairly permissive BSD-style license. I don't know who would make the judgement call as to whether or not the licenses can co-exist. However, since neither license is "viral" in the sense of wanting to apply any sort of restriction on derived works or the "work as a whole", I see no barrier to shipping a combined product with different portions falling under different licenses. Thus, the unit tests themselves would still fall under the LLVM license, and linking the unit tests with gtest would not violate either license. Of course, IANAL.

From a maintenance standpoint, we have already heard from several enthusiastic volunteers who are involved in both the development of LLVM and gtest. So I doubt there will be much problems on that score.

My personal goal is that one should be able to check out the head of LLVM on a generic Linux/OS-X system, with only the standard development environment (i.e. make/gcc/etc) and type "make unittest" and have the tests run.

In the longer term, I'd like to see LLVM have an automated build that runs the unit tests as part of the build. I noticed that gtest has an option to output test results in XML form, although I have not played with this personally, it might be useful in this regard.