s390x: `vfchsbs` and `vfchesbs` error with only `+vector`

[folkertdev]

These LLVM functions should implement the semantics of float compare (even if they are inefficient).

https://godbolt.org/z/38PYEWW77

Most of the functions do indeed emulate for vector_float when only vector (and not vector-enhancements-1 is available, but these don't.

rustc-LLVM ERROR: Cannot select: 0x7da6699a12a0: v4f32,i32 = SystemZISD::VFCMPHS 0x7da6699a1070, 0x7da6699a1150, example.rs:37:5
  0x7da6699a1070: v4f32,ch = CopyFromReg 0x7da669811520, Register:v4f32 %0
    0x7da6699a1000: v4f32 = Register %0
  0x7da6699a1150: v4f32,ch = CopyFromReg 0x7da669811520, Register:v4f32 %1
    0x7da6699a10e0: v4f32 = Register %1
In function: vec_float_compare

[uweigand]

The low-level intrinsics are intended to map 1:1 to the instruction with the same name. It is intentional that they are not supported on architecture levels that do not have the instruction. Fully emulating something like vfchsbs, in particular the condition code result, would be quite inefficient, and not something anyone would want to use.

Note that you can still generate code for vector float comparisons on z13 via regular IR, if that's what you want. (Also, reductions of vector comparisons should now work better, even in the case of float on z13.)