ASSERT when building record layout for anonymous structures.


lldb asserts when inspecting a simple structure.

Here is the simple test case. I already filed a bug

In my project a main base structure has anonymous field which makes
lldb unusable(since it asserts for any structure access). Can somebody
please take a look at this bug?

$ cat test.c

typedef struct {
    int dummy;

typedef struct {
    type_y y;

int main(int argc, char *argv[])
    type_z *z = 0;
    return 0;

$ gcc -O0 -g3 ./test.c

$ lldb a.out
Current executable set to 'a.out' (x86_64).
(lldb) p *((type_z *)0)
lldb: llvm/tools/clang/lib/AST/RecordLayoutBuilder.cpp:2271: uint64_t
<anonymous namespace>::RecordLayoutBuilder::updateExternalFieldOffset(const
clang::FieldDecl *, uint64_t): Assertion
`ExternalFieldOffsets.find(Field) != ExternalFieldOffsets.end() &&
"Field does not have an external offset"' failed.


Are you on the trunk revision? I'm unable to reproduce your crash
(although I do recall seeing a crash like that a few months ago). Here's
what I see:

(lldb) p *((type_z*)0)
error: Couldn't dematerialize struct : Couldn't read a composite type from
the target: Input/output error
Errored out in Execute: Couldn't FinalizeJITExpression

Which makes sense, although the error message is not too user friendly,
the phrase 'couldn't read a composite type from the target' implies the
pointer was not dereferenced properly.

BTW, I am using GCC 4.7.2 (from Ubuntu 12.10)


Hi Samuel,

In the attached patch, I redefined z as a struct variable and added a test for "p z" that does trigger the assert below. I believe this is already captured by, so I cross-referenced the failure accordingly.

I also added a passing test to cover the error path when trying to deference a null pointer to a struct (*pz in the attached patch), and a passing test for "p z.y" so we have some test coverage.

In general, the assert occurs because lldb fails to generate the external AST layout that clang expects while parsing the DIE tree. This could have to do with the order in which the DIE tree is parsed. The attached tests also exercise different nests of anonymous structs and break variously with clang, gcc and icc. I've commented on bugzilla 15036 which mentioned other related failures.

I expect that some of these tests will fail on Darwin as well. Again, the test case is much appreciated,

- Ashok

test-anonymous.patch (6.53 KB)