clang accepts initializer for incomplete type

[efriedma-quic]

Bugzilla Link	3461
Resolution	FIXED
Resolved on	Mar 12, 2010 00:57
Version	unspecified
OS	Linux

Extended Description

Testcase:
struct foo x = {1};

Run through clang, prints no warnings/errors. Should print an error.

[lattner]

clang -S crashes on this.

[llvmbot]

I'm not really worried about the IRgen crash. There are lots of places IRgen assumes that we get "valid" ASTs. My kingdom for an AST verifier.

[llvmbot]

Attempt to check for incomplete initialisation
This looked like a nice bug to get started on, so please forgive any muppetry on my part.

I think this patch requires a complete type at the correct location, and I've added a few lines to the relevant test-cases. I don't think it gives any regressions to any other existing tests either.

I noticed that valid tentative declarations still give an error on codegen. E.g.
struct foo x;
struct foo { int a; };

I plan to work on that next unless anyone thinks I'm on completely the wrong track. Should I file a separate bug in the meantime?

[efriedma-quic]

Looks like generally the right idea, but I think it mishandles stuff like "int a[] = {1};". Did you run "make test" with your patch?

Please file a new bug for the IRGen crash.

[llvmbot]

You are right; clang does have a current codegen bug in dealing with globals
with incomplete type.

We probably need to handle these by waiting until the end of the translation
unit to emit the tentative definition, if the initializer is incomplete.

I haven't completely thought it through, but this may be almost simple as
simple deferring their emission (like we do for statics). However, we will need
to make sure that we do not try to create a common definition for them, instead
we should leave the definition external, and then wait until the end of the
translation unit to emit a common definition if no actual definition was seen.

[llvmbot]

Looks like generally the right idea, but I think it mishandles stuff like "int
a[] = {1};".

Thanks. I've updated the patch. Effectively just shifted the block on a bit so that the array type will have been resolved by CheckInitializerTypes before. I considered moving the check into that function, but it looks all the other callers handle the case already.

I've also added a test-case for the array with unspecified bounds.

Did you run "make test" with your patch?

I thought I had, but having seen what happens with the new patch I think I must have forgotten to recompile or something. This time it really does pass all the existing tests.

Please file a new bug for the IRGen crash.

Will do.

[llvmbot]

Updated patch for incomplete initializers

[efriedma-quic]

Fix committed in r68991.