Wrong result from a reduction with a branching operator on Julia 1.12 with --check-bounds=yes #678

With the GPUArrays 12 port (#665), == between two equal 64×64 oneArrays returns false. This only happens on Julia 1.12 and with bounds checking enabled, which Pkg.test does by default. It makes the strided reductions testset in test/array.jl:223 fail on 1.12. CI only tests 1.10 and 1.13, so it doesn't show up there.

GPUArrays' == reduces over a named tuple of two Bools (one tracking missing, one the result), and its reducer branches on the missing flag. Reduced:

using oneAPI
A = rand(Int32(1):Int32(4), 64, 64); dA = oneArray(A); dB = oneArray(A)
init = (; is_missing=false, is_equal=true)
mapper(a, b) = (; is_missing=false, is_equal=a == b)
branching(a, b) = (a.is_missing || b.is_missing) ? (; is_missing=true, is_equal=false) :
                                                  (; is_missing=false, is_equal=a.is_equal & b.is_equal)
branchless(a, b) = (; is_missing=a.is_missing | b.is_missing, is_equal=a.is_equal & b.is_equal)
@show mapreduce(mapper, branching, dA, dB; init)    # is_equal = false with --check-bounds=yes on 1.12
@show mapreduce(mapper, branchless, dA, dB; init)   # is_equal = true
@show mapreduce(mapper, branching, oneArray(ones(Int32, 4096)), oneArray(ones(Int32, 4096)); init)  # 1-D: true

• Only 2-D inputs fail. 1-D vectors of 1 to 4096 elements are correct.
• The missing-tracking mapper doesn't matter. Replacing the branching reducer with a branchless one fixes it.
• It doesn't depend on AcceleratedKernels (0.5.2, 0.5.3 and 0.5.4 all fail) or GPUCompiler (2.13.0, 2.13.1 and 2.13.2 all fail).
• The same code on CUDA.jl with Julia 1.12 and --check-bounds=yes is correct.
• oneAPI's main branch (GPUArrays 11) isn't affected, but it uses different reduction kernels.

The SPIR-V path for this kernel seems the likely culprit, on Julia 1.12's LLVM 18 or in IGC.

Environment: Intel Iris Xe Graphics (driver 1.3.38308), NEO 26.18.38308, libigc 2.34.4, SPIRV_LLVM_Backend 23.1.1, SPIRV_LLVM_Translator 23.1.1, Julia 1.12.7 (LLVM 18.1.7), oneAPI.jl from #665 (2.9.2 + GPUArrays 12.1.0, AcceleratedKernels 0.5.4, KernelAbstractions 0.9.44, GPUCompiler 2.13.2).