Version 16 (modified by nomeata, 4 years ago) (diff)


Validating Patches

Nowadays there are many people working on GHC, which is a Good Thing, but it does mean that many people are inconvenienced if a patch is pushed which makes the HEAD unbuildable. To try to reduce the number of times this happens, we now have a stricter policy for testing patches before committing them.

In order to test your patches:

  • Get a repository containing the latest HEAD, the patches you want to push, and no other patches or unrecorded changes. Depending on what you are doing, your working repository might be appropriate; otherwise you might prefer to keep a separate repository just for patch testing (see [Git#Workflowwithvalidate] for a suggested git workflow with a separate repository).
  • In the root directory of the tree, run
    sh validate
    This will do a "quick" build and then run the testsuite in "fast" mode. Do not push if any of the tests give unexpected results!
  • Depending on the nature of the changes, more testing might be sensible. e.g. if possible, build system changes should be tested on Linux, Mac OS X and Windows machines. Look at the full documentation for the test suite.

The validate script should take around 40mins on a fast, dual core machine. If you have more cores, you can use them like this:

CPUS=8 sh validate

Assuming all is well, go ahead and commit your changes! If you have commit access then just push as normal. If not, use "git send-email" and add a note to say what testing you have done, and on which operating system/architecture. Or send a normal email with a pull request to the ghc mailing list.

More details on validation

validate usually starts by make distclean to make sure that everything builds from scratch. Furthermore, it ignores the build settings you have put in mk/, and instead uses those in mk/ (It does not mess up your mk/ file of course.)

After you run validate your tree will continue to use the same settings. The way to get back to using your own is to run make distclean. Less brutally, simply remove the file mk/

In order to save time while debugging problems revealed by validate, the validate script understands several flags. These flags, and others understood by validate are documented in the validate script itself.

  • --no-clean: validate will not clean the tree before starting, so it will continue a previous build.
  • --fast: validate skips the build-a-distribution step, and does not build dynamic libraries.
  • --slow: turns on -DDEBUG for the stage2 compiler, and runs more tests
  • --testsuite-only: then validate will not build the tree at all, but only run the testsuite. This is useful if the problems validate found were only due to the testsuite falling out of sync with the code.

Important: When using either of these flags, be careful that you don't end up pushing patches that have not been properly validated!

Testing different configurations

You may want to validate a different configuration, e.g. with GhcLibWays = p. When validating, mk/, where you would normally put this, is ignored; use mk/ instead.

Finding unrecorded or unpushed patches

In order to find unrecorded changes run:

./sync-all status

from the root of your tree. To find unpushed patches, run:

./sync-all new

also from the root of your tree.

Validate has failing tests without any local patches; what do I do?

The best thing to do is to fix them! This will help make the world a better place, and gain you the admiration and thanks of your colleagues.

Fixing them could mean one of two things: Fix a bug in GHC (or the libraries) that the test is reporting, or fixing a broken test to not report a failure when nothing is actually going wrong.

I tried to fix them, but I got stuck

If you can't fix them yourself, then first file a ticket for the problem so that it doesn't get forgotten about. First run the testsuite for just that test (e.g. make fast TEST=thefailingtest), and include the full testsuite output in the ticket description. If you found any information out that might be useful to someone later fixing the bug, add that too. Also, add the name of the test in the Test Case field.

If several tests are failing, then file a ticket for each one individually, unless it is clear that a single problem is causing all the failures.

Then add expect_broken(123) to the test options for that test, where 123 is the number of the ticket you just opened. If the problem only appplies in certain circumstances, then only mark it broken for those circumstances, e.g. when(opsys('mingw32'), expect_broken(123)).