How to enable +zfinx in llvm 14?

[hughperkins]

I need to compile with +zfinx, so that float operations use int registers.

In llvm 14, it looks like this is available:

/usr/local/opt/llvm-14.0.0/bin/llc --march riscv32 -mattr=help 2>&1 | grep zfinx
  zfinx            - 'Zfinx' (Float in Integer).

However, if I try to compile LLVM IR to riscv source-code with it, It continues to use float registers.

In conjueciton with +f, it uses float variables:

$ llc-14 sum_floats-device.ll -o sum_floats-device.s --march=riscv32 -mattr=+m,+zfinx,+f 2>/dev/null && grep ft sum_floats-device.s
        flw     ft0, 0(a0)
        flw     ft1, -28(s0)
        fadd.s  ft0, ft1, ft0
        fsw     ft0, -28(s0)
        flw     ft0, -28(s0)
        fsw     ft0, 0(a0)

conversely, if I use without +f, the float add becomes software emulated:

$ llc-14 sum_floats-device.ll -o sum_floats-device.s --march=riscv32 -mattr=+m,+zfinx 2>/dev/null && grep '@plt' sum_floats-device.s
        call    __addsf3@plt

finally, using some flag that doesnt exist has the same effect as using zfinx, i.e. none, so I wonder if I'm passing in the right parameter?

$ llc-14 sum_floats-device.ll -o sum_floats-device.s --march=riscv32 -mattr=+m,+blahblah 2>/dev/null && grep '@plt' sum_floats-device.s
        call    __addsf3@plt

Thoughts on how I can enable zfinx in llvm 14?