Reject allocations of objects that contain references

[maleadt]

There is no device GC, and gc_pool_alloc does not provide the object header expected by Julia's runtime. Only statically typed uses of plain-data objects work, such as an escaping Ref{Int}, a mutable struct with bits fields, or a boxed bits value. Objects that reference other objects also get GC orderings (unordered, release) on their loads and stores, which Metal and SPIR-V cannot express. After AllocOpt, GPULowerGCFrame now resolves the type of each remaining allocation. For types that are not pointer-free, it emits a marker that check_ir reports with a backtrace:

InvalidIRError: compiling MethodInstance for refs(::MtlDeviceVector{Int32, 1}) resulted in invalid LLVM IR
Reason: unsupported allocation of an object with references (Holder)
Stacktrace:
 [1] Holder
   @ ./example.jl:3
 [2] refs
   @ ./example.jl:7

referenced_object now also looks through cast instructions. Metal.jl's :table relocations produce those around the type operand.

This depends on #940, which removes dead exception objects before this check. On main with only this change, kernels containing Bool(x), Int32(::Float32), or throw(DomainError(x, LazyString(...))) fail to compile on Julia 1.11 and 1.12 for both Metal and SPIR-V (Tuple{DataType, Int64}, Tuple{DataType, Float32}, LazyString).

Scope

The check runs for every target that does not use the Julia runtime, including PTX and GCN. The suites below cover Metal, oneAPI, OpenCL, and KernelAbstractions (POCL); the CUDA.jl and AMDGPU.jl suites have not been run with this change.

Known gaps

• A failing type assertion on an isbits-union value still boxes integer members for jl_type_error through the box_* runtime. That box contains plain data on the failure path, so it is allowed.
• Enzyme's interpreter does not run julia_ir_passes. Enzyme-differentiated GPU code with dead exception objects containing references, from throw paths not covered by a back-end override, is now rejected instead of compiled. Fixing this requires a follow-up in Enzyme that calls the interpreter-independent GPUCompiler.run_julia_ir_passes.

[maleadt]

check_allocation! crashes instead of rejecting or handling an allocation whose type operand is a small type tag:

using Metal
k(a) = (Base.donotdelete(Core.svec(a[1])); nothing)
Metal.code_llvm(devnull, k, Tuple{MtlDeviceVector{Float32,1}}; kernel = true)
# signal 11: referenced_object at src/relocation.jl:934, from check_allocation! (LowerGCFrame)

Reproduced with Julia 1.12.7 on this branch through the #951 stack and Metal.jl #977. GPUCompiler 2.8.2 compiles the same kernel.

The allocation is:

call ptr @julia.gc_alloc_obj(ptr %current_task, i64 16, ptr inttoptr (i64 144 to ptr))

Here, 144 is SimpleVector’s small type tag—an index into jl_small_typeof, as used for tags below jl_max_tags << 4—not a type address. The inttoptr branch in referenced_object passes it to unsafe_pointer_to_objref; dispatching on the resulting value then segfaults.

Small tags need to be resolved through jl_small_typeof instead.

I found this with a kernel that forwards keyword arguments through a non-inlined kwargs... wrapper, which allocates an svec.

[codecov]

Codecov Report

❌ Patch coverage is 0% with 36 lines in your changes missing coverage. Please review.
✅ Project coverage is 0.00%. Comparing base (946c1d7) to head (634b0b2).

Files with missing lines	Patch %	Lines
src/optim.jl	0.00%	21 Missing ⚠️
src/relocation.jl	0.00%	11 Missing ⚠️
src/validation.jl	0.00%	4 Missing ⚠️

Additional details and impacted files

@@                  Coverage Diff                   @@
##           tb/throw-arguments    #949       +/-   ##
======================================================
- Coverage               86.28%   0.00%   -86.29%     
======================================================
  Files                      29      29               
  Lines                    5783    5682      -101     
======================================================
- Hits                     4990       0     -4990     
- Misses                    793    5682     +4889     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.