Case 126: SYCL `device` impl pointer — `shared_ptr` → raw pointer¶

Field	Value
Verdict	🔴 BREAKING
Category	Breaking
Platforms	Linux
Flags	ABI break, API break
Detected `ChangeKind`s	`type_size_changed`
Source files	`examples/case126_sycl_device_impl_ptr/`

Category: C++ ABI | Verdict: 🔴 ABI BREAK (exit 4)

Provenance — this is a real DPC++ ABI break¶

This case is distilled from a real change in the Intel DPC++ / oneAPI SYCL runtime (the intel/llvm project):

Breaking change: intel/llvm#20821 — "[SYCL] Use device_impl * instead of shared_ptr<device_impl> inside sycl::device" (carries the abi-break label).
Symptom it produced: intel/llvm#20915 — sycl_symbols_windows.dump test failing in post-commit, fixed up by intel/llvm#21028.
Guarded upstream by: the ABI test suite at sycl/test/abi/ (layout_*.cpp record-layout goldens and the sycl_symbols_{linux,windows}.dump exported-symbol goldens).

The code here is an original, self-contained miniature that reproduces the shape of the change (a shared_ptr member replaced by a raw pointer). It is not copied from intel/llvm and pulls in no SYCL headers.

What breaks¶

sycl::device stored its implementation as a reference-counted std::shared_ptr<detail::device_impl>. On a 64-bit target that member is two pointers wide, so sizeof(sycl::device) == 16.

PR #20821 replaced it with a raw detail::device_impl * (the device_impl is owned by the parent platform for the lifetime of the runtime, so the refcount was unnecessary overhead). A raw pointer is one word, so sizeof(sycl::device) shrinks to 8 bytes.

Every consumer compiled against the old 16-byte layout is now wrong:

code that holds a device by value (on the stack, or embedded in another struct) reserves the wrong amount of storage and computes wrong offsets for anything placed after it;
a base/derived or container-of-device relationship shifts every following member.

The mangled names of device's methods do not change, so a pure exported-symbol dump (sycl_symbols_*.dump) does not see the layout change directly — which is exactly why upstream's symbol-dump guard only caught the downstream symbol churn on Windows and needed follow-up PRs (#20902, #21028) to stabilise. abicheck's layout diff catches the root cause.

Code diff¶

v1 (before #20821)	v2 (after #20821)
`std::shared_ptr<detail::device_impl> impl;`	`detail::device_impl *impl;`
`sizeof(sycl::device) == 16`	`sizeof(sycl::device) == 8`

Why abicheck catches it¶

With DWARF or headers (L1/L2) abicheck reports:

type_size_changed  sycl::device   (16 → 8 bytes)  → BREAKING

Without any DWARF (L0, symbols only) abicheck still flags the related binary-only signals through diff_elf_layout.py: had the change instead added or removed a virtual method or a base class, the _ZTV/_ZTI object sizes would shift and surface as vtable_slot_count_changed / rtti_inheritance_changed — see docs/reference/sycl-test-abi-coverage.md.

Reproduce manually¶

g++ -shared -fPIC -g -O0 -fvisibility=hidden -fvisibility-inlines-hidden v1.cpp -o libdev_v1.so
g++ -shared -fPIC -g -O0 -fvisibility=hidden -fvisibility-inlines-hidden v2.cpp -o libdev_v2.so
abicheck compare libdev_v1.so libdev_v2.so
echo "exit: $?"   # → 4 (ABI break)

How to fix¶

Keep the public class size stable across releases with the PIMPL idiom: store a single opaque pointer whose pointee layout (and ownership model) can change freely, and never change the public class's size. If the ownership model must change from shared to raw, bump the library SONAME so consumers relink.

References¶

intel/llvm#20821 — the breaking change
intel/llvm#20915 — the symptom
intel/llvm#21028 — the fixup
sycl/test/abi/ — upstream ABI test suite
Itanium C++ ABI: Data layout

Source files¶

CMakeLists.txt
app.cpp
v1.cpp
v1.h
v2.cpp
v2.h

Case 126: SYCL device impl pointer — shared_ptr → raw pointer¶