VectorCombinePass in buildScalarOptimizerPipeline produces invalid SPIR-V on OpenCLBackend

[michel2323]

Since v1.12.0, kernels compiled for SPIRVCompilerTarget (OpenCL.jl + pocl_jll) (KA on CPU) can produce SPIR-V that spirv-val rejects with:

error: line 0: The following forward referenced IDs have not been defined:
'308[%308]'

git bisect between v1.11.1 (good) and v1.13.2 (bad) identifies commit c095147 ("Align optimization pipeline with Julia's") as the first bad commit. A sub-bisection within c095147 narrows the cause to a single newly-added pass in buildScalarOptimizerPipeline:

# src/optim.jl, in the opt_level >= 2 branch
add!(fpm, VectorCombinePass())

Maybe guarding it like this is fine?

if can_vectorize(job)
  add!(fpm, VectorCombinePass())
end

[michel2323]

The MWE currently involves ExaModels.jl and is pretty heavy. I can try to reproduce it on a smaller kernel if needed.

[maleadt]

The change is correct, but it seems like a major workaround. Which back-end are you using, LLVM or Khronos?