r179251 - [libclang] Expose record layout info via new libclang functions:

[Matt Beaumont-Gay]

On Fri, Jun 21, 2013 at 3:24 PM, Argyrios Kyrtzidis <akyrtzi at gmail.com> wrote:
> +struct Test2 {
> +  struct {
> +    struct {
> +//CHECK64: FieldDecl=foo:[[@LINE+1]]:11 (Definition) [type=int]
> [typekind=Int] [sizeof=4] [alignof=4] [offsetof=0]
> +      int foo;
> +    };
> +    struct {
> +//CHECK64: FieldDecl=bar:[[@LINE+1]]:11 (Definition) [type=int]
> [typekind=Int] [sizeof=4] [alignof=4] [offsetof=32]
> +      int bar;
> +    };
> +    struct {
> +        struct {
> +//CHECK64: FieldDecl=foobar:[[@LINE+1]]:15 (Definition) [type=int]
> [typekind=Int] [sizeof=4] [alignof=4] [offsetof=64]
> +          int foobar;
> +        };
> +    };
> +    struct inner {
> +        struct {
> +//CHECK64: FieldDecl=mybar:[[@LINE+1]]:15 (Definition) [type=int]
> [typekind=Int] [sizeof=4] [alignof=4] [offsetof=0]
> +          int mybar;
> +        };
> +//CHECK64: FieldDecl=mole:[[@LINE+1]]:7 (Definition) [type=struct inner]
> [typekind=Unexposed] [sizeof=4] [alignof=4] [offsetof=96]
> +    } mole;
> +  };
> +};
>
>
> This struct is invalid:
> tools/clang/test/Index/print-type-size.cpp:77:12: error: types cannot
> be declared in an anonymous struct
>    struct inner {
>           ^
>
> I noticed because it trips an assert I added to
> ASTContext::getASTRecordLayout (asserting that the decl is valid). How
> deeply do you care about this particular part of the test case? Can I
> remove it, or is there a particular fix I should apply?
>
> I have a patch which adds some invalid-decl checks to
> clang_Type_getOffsetOf, but it breaks all of the existing test cases
> which explicitly exercise invalid structs (not to mention Test2
> above). It's obviously important that we don't crash on those inputs,
> but I'm not sure I see the value of checking for any particular
> output.
>
> Thoughts?
>
>
> I suggest adding the invalid-decl checks and leaving an invalid struct in
> the test to make sure libclang does not hit the assertion; the rest can be
> turned into valid decls so that the checks can work.

Great. I've taken the proverbial hatchet to the CHECKs in that test
and added a new little case that resembles the original crasher; patch
attached, please review.

-Matt
-------------- next part --------------
A non-text attachment was scrubbed...
Name: c-index-record-layout.patch
Type: application/octet-stream
Size: 6802 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/cfe-commits/attachments/20130621/c823f1f1/attachment.obj>