LLVM 18 / Julia 1.12: Invalid vectorization leads to `ptxas` error

[maleadt]

MWE:

julia> kron(adjoint(CuArray(rand(Complex{Int16}, 32, 64))), adjoint(CuArray(rand(Complex{Int16}, 128, 16))))
ERROR: Failed to compile PTX code (ptxas exited with code 255)
Invocation arguments: --generate-line-info --verbose --gpu-name sm_120 --output-file /tmp/jl_A7CvIrOtW5.cubin /tmp/jl_KlexYHAjwJ.ptx
ptxas /tmp/jl_KlexYHAjwJ.ptx, line 336; error   : Unexpected instruction types specified for 'sub'
ptxas fatal   : Ptx assembly aborted due to errors

This because LLVM emits sub.s16x2 %r18, %r16, %r17; which doesn't exist.
Fixed in LLVM 19 by llvm/llvm-project@2089596

Not easy to work around since this is LLVM vectorizing things:

  %49 = insertelement <2 x i16> <i16 poison, i16 0>, i16 %40, i64 0
  %50 = insertelement <2 x i16> poison, i16 %41, i64 0
  %51 = insertelement <2 x i16> %50, i16 %44, i64 1
  %52 = sub <2 x i16> %49, %51