Okay, I now have a suggested fix hint. PTAL.

FINALLY! found out how to do suggestions for typos.

Add a "FIXME" to improve the diagnostics with a typo hint.

Move attribute checking over to SemaDeclAttr.cpp

The GitHub repo with this and the counted_by change: https://github.com/bwendling/llvm-project/tree/array-bounds-remarks-with-counted-by

Use CBA as an acronym of CountedByAttribute instead of EBA, which is oldspeak.

In D144136#4639907, @kees wrote:

Can you refresh this patch to work with https://reviews.llvm.org/D148381 ? My testing seems to imply that it doesn't know the size of the array. I assume the if (!IsUnboundedArray) check is incomplete now. i.e. for a __counted_by array, I see the "unknown" remark:

array-bounds.c:341:2: remark: accessing unknown sized array by 'index - 1' [-Rarray-bounds]
  341 |         TEST_ACCESS(p, array, index, SHOULD_TRAP);               
      |         ^

which is from the array-bounds.c test cases:

TEST_SIGNAL(counted_by_enforced_by_sanitizer, SIGILL)
{
        struct annotated *p;
        int index = MAX_INDEX + unconst;

        p = alloc_annotated(index);

        REPORT_SIZE(p->array);
        TEST_ACCESS(p, array, index, SHOULD_TRAP);
}

Friendly Ping.

In D148381#4646780, @kees wrote:

In D148381#4645600, @void wrote:

Added more error messages. Changed some code around to align with coding practices. Added some more test cases.

Doing a full "allmodconfig" Linux kernel build with my 200ish __counted_by annotations is building cleanly. Yay! :)

Add release note.

Added more error messages. Changed some code around to align with coding practices. Added some more test cases.

This is ready for review. Please take a look.

Add test output when the attribute isn't used.

Fix docs to use the correct code examples throughout.

I changed more than just the testcase (not a lot, but non-trivial nonetheless). PTAL.

Remove unneded extern decl from test.

• Expand test to show the unsanitized version,
• Unify the ways of finding a specific field based on a predicate,
• Make sure we're using the correct types in a couple of key places.

In D148381#4639909, @kees wrote:

I can generate warnings for anonymous structs were the __counted_by member is reported as "not found". For example:

little.c:7:28: warning: counted_by field 'count' not found
    7 |                 int array[] __counted_by(count);
      |                                          ^~~~~

For this:

#define __counted_by(member)   __attribute__((__counted_by__(member)))

struct anon {
        unsigned long flags;
        struct {
                unsigned char count;
                int array[] __counted_by(count);
        };
};

extern void bar(int input);

void foo(struct anon *p, int index)
{
        bar(p->array[index]);
}

I'll look into this.

This is now ready for a non-WIP review. PTAL. :-)

Update documentation. Cribbed off of Qing Zhao's GCC implementation.

Fix testcase.

Remove un-needed change.

Add a few more tests to the codegen testcase.

Add testcases.

Add more diagnostics.

Fix some issues Nick pointed out.

In D148381#4625462, @nickdesaulniers wrote:

I assume you plan to add some clang CodeGen tests at some point?

In D148381#4626229, @serge-sans-paille wrote:

Jumping a bit late in the thread, apologize in advance if I missed something.

The GCC version of the attributes seems to be element_count (as found in the link https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108896 mentionned by @kees) I guess we could align (I personally prefer the GCC name, FWIW)

Move the isFlexibleArrayMemberLike fully into Decl and fix tests.

Modify sanitize scope to avoid the instruction creation.

Fix clang-ast-dump issue.

Address Nick's comments.

Support when doing a _bdos on the full structure.

Fix a segfault.

Make builtin_dynamic_object_size use the counted_by attribute to help determine the object size.

Update to more match Qing's GCC implementation.

Bueno!

I'm not familiar with this code, but why not place the information into ConstantUser instead? Or maybe make the struct a private member of ConstantHoistingPass.

Nice catch!

Rebasing

Add bounds checking for the "ArrayBounds" saniziter kind.

Use strings for the attribute argument.

Use "Expected" for the SourceRange imports.

Improve error message to use the unknown field's name.

Change sanitizer scope because instructions may be created.

Fix the way the source ranges are imported via the ASTImporter.
Do other minor fixes according to feedback.

Use the new memory location.

This is a precursor to specifying the field with the designated initializer syntax. Therefore, it may be a bit cumbersome.

Fix ICE when the FD isn't found.

Add the ability to specify a "path" to the element count if it resides within a sub-structure.

I think this change is causing build failures. Could you take a look?

Add comment.

Fix failing testcase.

In D148381#4274857, @nickdesaulniers wrote:

Cool!

How about some codegen tests?

I'll add those to the next iteration. :-)

Reformat.

Friendly ping.

Note: This is a work-in-progress, so there's no need for a formal review, though any suggestions will definitely be appreciated. :-)

Fix compilation issue.

In D147580#4250232, @shafik wrote:

I guess I should have waited till I saw: https://reviews.llvm.org/D147673