Fix crash in DependentGenericTypeResolver::resolveDependentMemberType

[jtbandes]

An erroneous baseTy would cause resolveArchetype to return nullptr, resulting in a null dereference.

Apologies — the test suite is prohibitively large for me to run it fully before submitting this patch. The file 00626-swift-lexer-lexidentifier.swift no longer crashes the compiler, so I've moved it to compiler_crashes_fixed. Before merging, though, someone will need to convince themselves that everything else still works.

[jtbandes]

Note: the same sort of resolveArchetype call is made a few lines down, in DependentGenericTypeResolver::resolveSelfAssociatedType(). However, I couldn't find a case that would trigger the same crash, so I didn't modify that function. Based on git-blame, it looks like @DougGregor might be able to confirm this, or help find a crashing case.

[gribozavr]

Apologies — the test suite is prohibitively large for me to run it fully before submitting this patch.

What configuration (hardware and bulid) are you using? The full test suite runs within less than 10 minutes when the compiler is built in release mode, try utils/build-script -RT.

[jtbandes]

Just kidding — this fixed 335 of the crashing tests 😃

[jckarter]

Nice!

[lattner]

Wow, fantastic! @DougGregor, please review this when you have a chance, thanks!

[jtbandes]

Updated with new fixed crashers.

[DougGregor]

Please squash this into a single commit; we like to have the test suite in sync with the code.

[AlexDenisov]

Hm, why all the fixed crashers removed? Shouldn't they go into fixed_crashers?

[jtbandes]

@DougGregor: done.
@AlexDenisov: I think GitHub is just struggling to show the diff. The files were moved, but it decided they were deleted+added instead of just renamed. 🤷

[yas375]

@jtbandes it looks like the moved files were not commited.

[jtbandes]

@yas375: They actually were. GitHub shows it incorrectly; I've sent a message to support. If you pull jtbandes/swift@d781605a6d, you can see them locally.

[yas375]

@jtbandes you are right! I apologize.

I bet we can help git to understand that those are actually just file moves by using git mv. And the most important outcome from using git mv is that reviewer will be sure that those are file moves and there are no hidden changes in there.

I went ahead and did it locally and here is a list of steps to follow:

Given: you are on top of your branch with the fix.

git reset --soft HEAD~1
git checkout validation-test/compiler_crashers/
git clean -fd
git status # you should see only `modified:   lib/Sema/TypeCheckGeneric.cpp`

I've created a list of commands using string replacement on the output of the removed files to use git mv:

bash <(curl -s https://gist.githubusercontent.com/yas375/ca9667526056c92afd05/raw/d11107bc807ba9e718ad3031334ade9e588b4116/gistfile1.txt)

Now git status should report all file moves in the right way:

➜  swift git:(072a459) ✗ git st
HEAD detached from d781605
Changes to be committed:
  (use "git reset HEAD <file>..." to unstage)

  renamed:    validation-test/compiler_crashers/00066-diagnoseunknowntype.swift -> validation-test/compiler_crashers_fixed/00066-diagnoseunknowntype.swift
  renamed:    validation-test/compiler_crashers/00626-swift-lexer-lexidentifier.swift -> validation-test/compiler_crashers_fixed/00626-swift-lexer-lexidentifier.swift
  renamed:    validation-test/compiler_crashers/01908-std-function-func-mapsignaturetype.swift -> validation-test/compiler_crashers_fixed/01908-std-function-func-mapsignaturetype.swift
  ...
  renamed:    validation-test/compiler_crashers/28179-void.swift -> validation-test/compiler_crashers_fixed/28179-void.swift

Changes not staged for commit:
  (use "git add <file>..." to update what will be committed)
  (use "git checkout -- <file>..." to discard changes in working directory)

  modified:   lib/Sema/TypeCheckGeneric.cpp

Please rerun the tests to make sure I didn't break anything :)

git add lib/Sema/TypeCheckGeneric.cpp
git commit
git push --force

UPD: here is how changes will be shown on GitHub: http://screencast.com/t/NO5nEJDw

Thanks for your contribution! 👍

---

I've also reported the issue with wrong statistics to GitHub support.

[jtbandes]

Thanks @yas375! I wasn't aware that git mv did anything different from mv + git add. I'll update the diff.

[jtbandes]

Actually, that didn't help. (jtbandes/swift@3cbabd0). I think git mv is basically the same as mv + git add.

The problem is that in all of the moved files, I also changed one line, to remove the --crash flag.

I think this is as good as it will get. It's only a GitHub problem, not a git problem :)

[yas375]

@jtbandes this is weird. git mv should produce the right result. It works for me. See my temporary pull request and commit itself where I used git mv. On both pages the changes displayed properly.

I don't think it matters, but just in case:

➜ git --version
git version 2.6.2

[jtbandes]

@yas375 I'm using 2.6.3.

On your PR I see "File renamed without changes" for each of the test files. That seems wrong to me. The files are supposed to be modified as well as renamed.

[yas375]

@jtbandes oh, I'm sorry then. I thought those were just moves. Sorry for the confusion and for taking you time.

Great job! 👍

[jtbandes]

No worries, thanks for trying! It would have been nice if it worked :)

[jtbandes]

@DougGregor please let me know if there's anything else to be done here!

[jtbandes]

@DougGregor @gribozavr I've updated this PR to avoid conflicts from #398 and #405