[PATCH] D33720: [PowerPC] Eliminate compares - add i64 sext/zext handling for SETNE

[Guozhi Wei via Phabricator via llvm-commits]

Carrot added a comment.

In https://reviews.llvm.org/D33720#788925, @nemanjai wrote:

> In https://reviews.llvm.org/D33720#788613, @Carrot wrote:
>
> >   SDValue Xor = IsRHSZero ? LHS :
> >        SDValue(CurDAG->getMachineNode(PPC::XOR8, dl, MVT::i64, LHS, RHS), 0);
> >
> > SDValue AC =
> >
> >     SDValue(CurDAG->getMachineNode(PPC::ADDIC8, dl, MVT::i64, MVT::Glue,
> >                                    Xor, getI32Imm(~0U, dl)), 0);
> >   return SDValue(CurDAG->getMachineNode(PPC::SUBFE8, dl, MVT::i64, AC,
> >                                         Xor, AC.getValue(1)), 0);
> >   
> >
> > Suppose we have RHS literal 0, LHS is a var, its runtime value is also 0, the expected value of the expression should be 0. But the runtime values are:
> >
> >   Xor  = LHS = 0
> >   AC = Xor + ~0 = ~0, no carry
> >   result = Xor - AC - carry = 0 - (~0) = 1
> >   
> >
> > So the code sequence computes out wrong result.
>
>
> That's actually not what the SUBFE instruction does. As per the ISA:
>
>   subfe RT,RA,RB
>   computes
>   RT = ~(RA) + (RB) + CA
>
>
> So in your example, it will be:
>
>   ~AC + XOR + CARRY
>   ~(~0) + 0 + 0
>
>
> Which of course will just be a zero.


Then the document is wrong.

Mathematically, subtraction is
RB - RA = RB + ~RA + 1
just like the description in instruction subf.

A subtraction when considering borrowing from lower bit is
RB - RA - CA = RB + ~RA + 1 - CA = RB + ~RA + !CA


Repository:
  rL LLVM

https://reviews.llvm.org/D33720