X86-64: SHR by one incorrectly marks OF as undefined

Given the following minimal example:

repro:
    xor ecx, ecx
    mov eax, 0DEADh
    mov edx, 0BEEFh
    mov r8, 8000000000000001h
    shr r8, 1
    cmovo rax, rdx
    ret

At MMAT_GENERATED, the SHR instruction produces:

mov    #1.1, t1.1 ; u=           d=t1.1
cfshr  r8.8, t1.1, cf.1 ; u=r8.8,t1.1  d=cf.1
shr    r8.8, t1.1, r8.8 ; u=r8.8,t1.1  d=r8.8
und    of.1 ; u=           d=of.1
setz   r8.8, #0.8, zf.1 ; u=r8.8       d=zf.1
setp   r8.8, #0.8, pf.1 ; u=r8.8       d=pf.1
sets   r8.8, sf.1 ; u=r8.8       d=sf.1

Notice the unconditional und of.1.

At MMAT_LVARS, the OF dependency remains unresolved:

mov    #0xDEAD.8, result.8 ; u=           d=rax.8
jcnd   lnot(of1.1{1}), @3 ; u=of.1

The resulting pseudocode contains an uninitialized flag variable:

__int64 result; // rax
char v1; // of

result = 0xDEAD;
if ( v1 )
  return 0xBEEF;
return result;

For SHR with a count of one, OF is defined as the original most significant bit of the destination operand.

Here, the original MSB is one, so OF must be one and the correct return value is 0xBEEF.

Unlike shifts with counts greater than one, OF is not undefined in this case. The initial lift should calculate OF from the original destination instead of marking it undefined.

Hi, thanks for taking the time to report these x86-64 issues (along with the related ones in other topics). We’ve opened tickets for them and will follow up once we know more.