[PATCH] Avoid trying to get the alignment of a FieldDecl in an invalid record

[Richard Smith]

On Tue, Jun 25, 2013 at 2:24 PM, Matt Beaumont-Gay <matthewbg at google.com> wrote:
> While investigating PR16292, I added "assert(!D->isInvalidDecl())" to
> ASTContext::getASTRecordLayout. The problem I fixed in r184751 was one
> of the two paths which tickled that assertion via existing tests; this
> is the other. (The actual testcase which triggers the assertion is
> SemaCXX/alignof.cpp.) This patch changes ASTContext::getDeclAlign to
> skip the actual field alignment calculation (and return 1-byte
> alignment in the general case) if the field's parent record is
> invalid. The patch also adds the assertion in getASTRecordLayout,
> though I can commit that separately if anybody's fussed.
>
> The careful reader will note that I took the opportunity to rename
> local variables in the block which I touched to conform to proper LLVM
> style, since getDeclAlign currently has a mix of styles and it made me
> twitch.

Change LGTM, but please also add a test which would fail even without
the added assertion. Something like:

struct S; struct T { S s; int k; } t; int n = __alignof__(T::k);