[pocl] Back @private Scratchpad with GPUCompiler.alloca

[vchuravy]

Summary

Replaces the POCL back-end's MArray-backed @private scratchpad with a direct per-workitem stack allocation via GPUCompiler.alloca. The returned Ptr is wrapped in a CLDeviceArray over OpenCL "Function" storage (LLVM addrspace 0), which is where the SPIR-V target places allocas.

@device_override @inline function KA.Scratchpad(ctx, ::Type{T}, ::Val{Dims}) where {T, Dims}
    ptr = POCL.GPUCompiler.alloca(T, Val(prod(Dims)))
    CLDeviceArray(Dims, reinterpret(POCL.LLVMPtr{T, POCL.AS.Function}, ptr))
end

This drops the StaticArrays dependency from the POCL back-end (StaticArrays is still used by the CPU back-end).

Why

GPUCompiler.alloca emits a real entry-block alloca that the optimizer can promote, in the target's alloca address space — avoiding the unsoundness of llvmcall + alloca and the overhead/semantics of MArray. See the motivation in the companion GPUCompiler PR.

Alignment

The alloca is aligned to Base.datatype_alignment(T), which is exactly the alignment CLDeviceArray uses for its element loads/stores (alignment(::CLDeviceArray{T})), so accesses are consistent. isbits-union element types are intentionally unsupported (GPUCompiler.alloca guards on isbitstype(T)).

Status

Draft — depends on JuliaGPU/GPUCompiler.jl#859 (adds the alloca intrinsic). Project.toml compat is bumped to GPUCompiler = "1.23"; this can be un-drafted once that is merged and released.

Testing

Verified end-to-end against the local GPUCompiler branch: @private Float32 (4,) lowers to alloca [16 x i8], align 4 in addrspace 0 with no surviving julia.gpu.alloca, and the kernel runs correctly on the POCL CPU device.

🤖 Generated with Claude Code

[codecov]

Codecov Report

❌ Patch coverage is 0% with 2 lines in your changes missing coverage. Please review.
✅ Project coverage is 0.00%. Comparing base (a8022b2) to head (0b58f9b).

Files with missing lines	Patch %	Lines
src/pocl/backend.jl	0.00%	2 Missing ⚠️

❗ There is a different number of reports uploaded between BASE (a8022b2) and HEAD (0b58f9b). Click for more details.

HEAD has 28 uploads less than BASE

Flag	BASE (a8022b2)	HEAD (0b58f9b)
	48	20

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #714       +/-   ##
==========================================
- Coverage   62.51%   0.00%   -62.52%     
==========================================
  Files          23      22        -1     
  Lines        1926    1737      -189     
==========================================
- Hits         1204       0     -1204     
- Misses        722    1737     +1015     

☔ 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.