Inconsistent tolerance to poison values between compareFunctions or constantSynthesis

[Misasasa]

In enumerator.cpp, compareFunctions or constantSynthesis will be invoked according to whether the sketch contains any constant:

    if (!HaveC) {
        AliveEngine AE(TLI);
        Good = AE.compareFunctions(*Src, *Tgt);
    } else {
        AliveEngine AE(TLI);
        Good = AE.constantSynthesis(*Src, *Tgt, ConstantResults);
    }

However, we find that their tolerance to poison is different.
Considering Example 3 in the paper:

%0 = shufflevector <32 x i8> %x, poison, <3, 7, 11, 15, 19, 23, 27, 31>
%1 = lshr %0, <6, 6, 6, 6, 6, 6, 6, 6>
%2 = zext 8 x i8> %1 to <8 x i32>
ret <8 x i32> %2
=>
%0 = bitcast <32 x i8> %x to <8 x i32>
%1 = call @llvm.x86.avx2.psrli.d(<8 x i32> %0, 30)
ret <8 x i32> %1

Minotaur can successfully generate this rewrite by constantSynthesis:

define <8 x i32> @cut.256.257(<32 x i8> %__n3, i16 %__n4, i32 %_reservedc_99) {
#0:
  %__n0 = shufflevector <32 x i8> %__n3, <32 x i8> poison, 3, 7, 11, 15, 19, 23, 27, 31
  %__n1 = lshr <8 x i8> %__n0, { 6, 6, 6, 6, 6, 6, 6, 6 }
  %__n2 = zext <8 x i8> %__n1 to <8 x i32>
  ret <8 x i32> %__n2
}
=>
define <8 x i32> @cut.256(<32 x i8> %__n3, i16 %__n4, i32 %_reservedc_99) {
#0:
  %#1 = bitcast <32 x i8> %__n3 to <8 x i32>
  %intr = x86_avx2_psrli_d <8 x i32> %#1, i32 %_reservedc_99
  ret <8 x i32> %intr
}
;result
i32 %_reservedc_99 = #x0000001e (30)

However, when I proceed to check whether they are really equivalent using compareFunctions:

define <8 x i32> @cut.256.257(<32 x i8> %__n3, i16 %__n4, i32 %_reservedc_99) {
#0:
  %__n0 = shufflevector <32 x i8> %__n3, <32 x i8> poison, 3, 7, 11, 15, 19, 23, 27, 31
  %__n1 = lshr <8 x i8> %__n0, { 6, 6, 6, 6, 6, 6, 6, 6 }
  %__n2 = zext <8 x i8> %__n1 to <8 x i32>
  ret <8 x i32> %__n2
}
=>
define <8 x i32> @cut.256(<32 x i8> %__n3, i16 %__n4, i32 %_reservedc_99) {
#0:
  %#1 = bitcast <32 x i8> %__n3 to <8 x i32>
  %intr = x86_avx2_psrli_d <8 x i32> %#1, i32 30
  ret <8 x i32> %intr
}

Alive2 provides a counterexample about poison values:

ERROR: Target is more poisonous than source

Example:
<32 x i8> %__n3 = < #x03 (3), #x03 (3), #x03 (3), #x03 (3), #x03 (3), #x03 (3), poison, #x03 (3), #x03 (3), #x03 (3), #x03 (3), #x03 (3), #x03 (3), #x03 (3), #x03 (3), #x03 (3), #x03 (3), #x03 (3), #x03 (3), #x03 (3), #x03 (3), #x03 (3), #x03 (3), #x03 (3), #x03 (3), #x03 (3), #x03 (3), #x03 (3), #x03 (3), #x03 (3), #x03 (3), #x03 (3) >
i16 %__n4 = poison
i32 %_reservedc_99 = poison

Source:
<8 x i8> %__n0 = < #x03 (3), #x03 (3), #x03 (3), #x03 (3), #x03 (3), #x03 (3), #x03 (3), #x03 (3) >
<8 x i8> %__n1 = < #x00 (0), #x00 (0), #x00 (0), #x00 (0), #x00 (0), #x00 (0), #x00 (0), #x00 (0) >
<8 x i32> %__n2 = < #x00000000 (0), #x00000000 (0), #x00000000 (0), #x00000000 (0), #x00000000 (0), #x00000000 (0), #x00000000 (0), #x00000000 (0) >

Target:
<8 x i32> %#1 = < #x03030303 (50529027), poison, #x03030303 (50529027), #x03030303 (50529027), #x03030303 (50529027), #x03030303 (50529027), #x03030303 (50529027), #x03030303 (50529027) >
<8 x i32> %intr = < #x00000000 (0), poison, #x00000000 (0), #x00000000 (0), #x00000000 (0), #x00000000 (0), #x00000000 (0), #x00000000 (0) >
Source value: < #x00000000 (0), #x00000000 (0), #x00000000 (0), #x00000000 (0), #x00000000 (0), #x00000000 (0), #x00000000 (0), #x00000000 (0) >
Target value: < #x00000000 (0), poison, #x00000000 (0), #x00000000 (0), #x00000000 (0), #x00000000 (0), #x00000000 (0), #x00000000 (0) >

I am a little confused with this inconsistency. Considering this example, I wonder whether this optimization should be treated as correct.