Illegal memory access after aligned_sizeof changes

[maleadt]

MWE from https://discourse.julialang.org/t/check-bounds-no-causes-illegal-memory-access-when-using-rand-in-cuda-kernel/129455, run under --check-bounds=no:

using CUDA
N = 10^5
A = CuArray{Float32}(undef, N)
idx = CUDA.zeros(Int, N)
CUDA.@sync sortperm!(idx, A)

ERROR: LoadError: CUDA error: an illegal memory access was encountered (code 700, ERROR_ILLEGAL_ADDRESS)

Bisected to #2757 (cc @vchuravy).

[christiangnrd]

The issue seems to be that Base.aligned_sizeof isn't optimized away when --check-bounds=no.

Reduced MWE:

using CUDA
function kernel()
    CUDA.aligned_sizeof(Float32)
    return
end
CUDA.@sync @cuda kernel()