[llvm] [X86][FP16][BF16] Improve vectorization of fcmp (PR #116153)

[Reid Kleckner via llvm-commits]

rnk wrote:

This appears to have caused an regression on our internal codebase, generating this error message:
```
fatal error: error in backend: Cannot select: 0x13ef7f4be070: v16f32 = X86ISD::CMPP 0x13ef7f4bd2a0, 0x13ef7f5343f0, TargetConstant:i8<7>
  0x13ef7f4bd2a0: v16f32 = X86ISD::CVTPH2PS 0x13ef7d2eb0e0
    0x13ef7d2eb0e0: v16i16 = bitcast 0x13ef7f4bec40
      0x13ef7f4bec40: v16f16,ch = load<(load (s256) from %ir.142, align 1)> 0x13ef6dcf52a0, 0x13ef7f4bd230, undef:i64
        0x13ef7f4bd230: i64 = add 0x13ef7f534770, Constant:i64<96>
          0x13ef7f534770: i64 = add 0x13ef7d2eb3f0, 0x13ef7f4be540
            0x13ef7d2eb3f0: i64,ch = CopyFromReg 0x13ef6dd3d270, Register:i64 %1
              0x13ef6dcf5e70: i64 = Register %1
            0x13ef7f4be540: i64 = shl 0x13ef7f534a80, Constant:i8<1>
              0x13ef7f534a80: i64,ch = CopyFromReg 0x13ef6dd3d270, Register:i64 %4
                0x13ef7f4bd700: i64 = Register %4
              0x13ef7f4bd850: i8 = Constant<1>
          0x13ef6dcf58c0: i64 = Constant<96>
        0x13ef646602a0: i64 = undef
  0x13ef7f5343f0: v16f32 = BUILD_VECTOR ConstantFP:f32<0.000000e+00>, ConstantFP:f32<0.000000e+00>, ConstantFP:f32<0.000000e+00>, ConstantFP:f32<0.000000e+00>, ConstantFP:f32<0.000000e+00>, ConstantFP:f32<0.000000e+00>, ConstantFP:f32<0.000000e+00>, ConstantFP:f32<0.000000e+00>, ConstantFP:f32<0.000000e+00>, ConstantFP:f32<0.000000e+00>, ConstantFP:f32<0.000000e+00>, ConstantFP:f32<0.000000e+00>, ConstantFP:f32<0.000000e+00>, ConstantFP:f32<0.000000e+00>, ConstantFP:f32<0.000000e+00>, ConstantFP:f32<0.000000e+00>
    0x13ef7f4bd1c0: f32 = ConstantFP<0.000000e+00>
    0x13ef7f4bd1c0: f32 = ConstantFP<0.000000e+00>
    0x13ef7f4bd1c0: f32 = ConstantFP<0.000000e+00>
    0x13ef7f4bd1c0: f32 = ConstantFP<0.000000e+00>
    0x13ef7f4bd1c0: f32 = ConstantFP<0.000000e+00>
    0x13ef7f4bd1c0: f32 = ConstantFP<0.000000e+00>
    0x13ef7f4bd1c0: f32 = ConstantFP<0.000000e+00>
    0x13ef7f4bd1c0: f32 = ConstantFP<0.000000e+00>
    0x13ef7f4bd1c0: f32 = ConstantFP<0.000000e+00>
    0x13ef7f4bd1c0: f32 = ConstantFP<0.000000e+00>
    0x13ef7f4bd1c0: f32 = ConstantFP<0.000000e+00>
    0x13ef7f4bd1c0: f32 = ConstantFP<0.000000e+00>
    0x13ef7f4bd1c0: f32 = ConstantFP<0.000000e+00>
    0x13ef7f4bd1c0: f32 = ConstantFP<0.000000e+00>
    0x13ef7f4bd1c0: f32 = ConstantFP<0.000000e+00>
    0x13ef7f4bd1c0: f32 = ConstantFP<0.000000e+00>
  0x13ef6dcf54d0: i8 = TargetConstant<7>
```

I will start a reduction, but I wanted to give you a heads up that we may need to revert or fix forward quickly.

https://github.com/llvm/llvm-project/pull/116153