Wrong folding to `fabs`

[bongjunj]

Alive2 report: https://alive2.llvm.org/ce/z/iPnYcf

----------------------------------------
define float @select_nnan_ninf_nsz_fcmp_nnan_nsz_olt_negzero.2(float %x) {
#0:
  %#1 = fcmp oge float %x, -0.000000
  %negx = fsub nnan nsz float -0.000000, %x
  %fabs = select nnan ninf nsz i1 %#1, float %negx, float %x
  %#2 = fabs float %fabs
  ret float %#2
}
=>
define float @select_nnan_ninf_nsz_fcmp_nnan_nsz_olt_negzero.2(float %x) nofree willreturn memory(none) {
#0:
  %#1 = fabs nnan ninf nsz float %x
  ret float %#1
}
Transformation doesn't verify!

ERROR: Value mismatch

Example:
float %x = #x00000000 (+0.0)

Source:
i1 %#1 = #x1 (1)
float %negx = #x00000000 (+0.0)
float %fabs = #x00000000 (+0.0)
float %#2 = #x00000000 (+0.0)

Target:
float %#1 = #x80000000 (-0.0)
Source value: #x00000000 (+0.0)
Target value: #x80000000 (-0.0)

Summary:
  0 correct transformations
  1 incorrect transformations
  0 failed-to-prove transformations
  0 Alive2 errors

Is it an expected behavior for the signbit to flip?
I think this is introduced recently from d13947b

[dtcxzyw]

We cannot fix it under the current (Alive2's) semantics of nsz. This is another variant of fabs(fabs nsz(X)) -> fabs nsz (X) which has been reported in https://discourse.llvm.org/t/rfc-clarify-the-behavior-of-fp-operations-on-bit-strings-with-nsz-flag/85981.

FYI, I don't want to fix these nsz-related issues unless we reach a consensus on the semantics of nsz. Only mutating input zeros instead of mutating both inputs and output looks fine to me. But I still need more time to evaluate potential solutions.

[bongjunj]

Thank for clarifying the issue around FP operations. I wasn't aware of that.