DebugInfo: No alignment info for "transitive" alignment-enforced structures.

[hokein]

Split from #72913.

Consider the following case:

struct alignas(8) Base {} a;

struct Derived : Base {} b;

Debug info emitted by clang ./bin/clang -cc1 -emit-llvm -debug-info-kind=limited /tmp/t.cpp -o -

!8 = distinct !DICompositeType(tag: DW_TAG_structure_type, name: "Derived", file: !7, line: 4, size: 64, flags: DIFlagTypePassByValue, elements: !9, identifier: "_ZTS7Derived")
!9 = !{!10}
!10 = !DIDerivedType(tag: DW_TAG_inheritance, scope: !8, baseType: !11, extraData: i32 0)
!11 = distinct !DICompositeType(tag: DW_TAG_structure_type, name: "Base", file: !7, line: 2, size: 64, align: 64, flags: DIFlagTypePassByValue, elements: !12, identifier: "_ZTS4Base")

The DICompositeType for Derived is missing the align attribute, it is a critical information for lldb to get correct layout for the structure.

This is a regression caused by https://reviews.llvm.org/D24426.

Comments from @Michael137:

FWIW, GCC does do this "alignment propagation" and attaches DW_AT_alignment to the derived class too: https://godbolt.org/z/j5MWdWYaf

[llvmbot]

@llvm/issue-subscribers-debuginfo

Author: Haojian Wu (hokein)

Split from #72913.

Consider the following case:

struct alignas(8) Base {} a;

struct Derived : Base {} b;

Debug info emitted by clang ./bin/clang -cc1 -emit-llvm -debug-info-kind=limited /tmp/t.cpp -o -

!8 = distinct !DICompositeType(tag: DW_TAG_structure_type, name: "Derived", file: !7, line: 4, size: 64, flags: DIFlagTypePassByValue, elements: !9, identifier: "_ZTS7Derived")
!9 = !{!10}
!10 = !DIDerivedType(tag: DW_TAG_inheritance, scope: !8, baseType: !11, extraData: i32 0)
!11 = distinct !DICompositeType(tag: DW_TAG_structure_type, name: "Base", file: !7, line: 2, size: 64, align: 64, flags: DIFlagTypePassByValue, elements: !12, identifier: "_ZTS4Base")

The DICompositeType for Derived is missing the align attribute, it is a critical information for lldb to get correct layout for the structure.

This is a regression caused by https://reviews.llvm.org/D24426.

[llvmbot]

@llvm/issue-subscribers-lldb

Author: Haojian Wu (hokein)

Split from #72913.

Consider the following case:

struct alignas(8) Base {} a;

struct Derived : Base {} b;

Debug info emitted by clang ./bin/clang -cc1 -emit-llvm -debug-info-kind=limited /tmp/t.cpp -o -

!8 = distinct !DICompositeType(tag: DW_TAG_structure_type, name: "Derived", file: !7, line: 4, size: 64, flags: DIFlagTypePassByValue, elements: !9, identifier: "_ZTS7Derived")
!9 = !{!10}
!10 = !DIDerivedType(tag: DW_TAG_inheritance, scope: !8, baseType: !11, extraData: i32 0)
!11 = distinct !DICompositeType(tag: DW_TAG_structure_type, name: "Base", file: !7, line: 2, size: 64, align: 64, flags: DIFlagTypePassByValue, elements: !12, identifier: "_ZTS4Base")

The DICompositeType for Derived is missing the align attribute, it is a critical information for lldb to get correct layout for the structure.

This is a regression caused by https://reviews.llvm.org/D24426.

[dwblaikie]

GCC Does include the alignment, Clang does not: https://godbolt.org/z/q5rGMvEMn

I could see an argument for Clang being correct & leaving it up to the consumer to do the layout - but DWARF does encode the member offsets, etc, when those could be derived with sufficient ABI knowledge, etc, so explicitly including the alignment when it's non-trivial seems maybe suitable.
(on the other hand neither GCC nor Clang include the alignment when it isn't overriden - and consumers derive it themselves by looking at the member types, etc, presumably? So maybe it's not that much of a burden for them to derive it in this case too?)

[Michael137]

(on the other hand neither GCC nor Clang include the alignment when it isn't overriden - and consumers derive it themselves by looking at the member types, etc, presumably? So maybe it's not that much of a burden for them to derive it in this case too?

I had the same initial thought, since when alignment isn't specified, the ItaniumRecordLayoutBuilder has an InferAlignment path for each record. But in practice the inference isn't setting the alignment on derived classes appropriately (given #72913 still crashes after LLDB started respecting DW_AT_alignment in 439b16e). I guess an alternative to propagating this in DWARF could be to "fix" the InferAlignment path of the layout builder, if feasible.

[Michael137]

Interestingly, I hacked something together that addresses #73563. With that in place we end up inferring the alignment correctly, presumably because the alignment calculations rely on data-size being set correctly.

So I suggest we focus on the other issue first, and see if that resolves all our problems

[hokein]

Hi, @Michael137

You seem to take a look on this issue and #73563, is there any progress or update on them?

[Michael137]

Hi, @Michael137

You seem to take a look on this issue and #73563, is there any progress or update on them?

I've been meaning to follow-up on https://discourse.llvm.org/t/rfc-lldb-handling-no-unique-address-in-lldb/77483 but was sidetracked with other things. I think the direction we're taking is to make the Clang layout-builder assertions less restrictive for external sources (such as LLDB). This would resolve the original issue raised in #72913 (but wouldn't address the "DW_AT_alignment isn't propagated problem" raised here).