Simplification Comparison for (a | b) ? (a ^ b) : (a & b) etc. (Clang13 vs Clang trunk)

[k-arrows]

Consider the following six functions.
https://godbolt.org/z/cTsd43jhh

bool test1(bool a, bool b)
{
    return (a | b) ? (a ^ b) : (a & b);
}

bool test2(bool a, bool b)
{
    return (a | b) ? (a & b) : (a ^ b);
}

bool test3(bool a, bool b)
{
    return (a & b) ? (a | b) : (a ^ b);
}

bool test4(bool a, bool b)
{
    return (a & b) ? (a ^ b) : (a | b);
}

bool test5(bool a, bool b)
{
    return (a ^ b) ? (a | b) : (a & b);
}

bool test6(bool a, bool b)
{
    return (a ^ b) ? (a & b) : (a | b);
}

Clang13 can simplify four of these, test1, test2, test3 and test4, while Clang trunk cannot. Neither Clang13 nor Clang trunk can simplify test5 and test6.

[k-arrows]

Sorry, maybe I should have made a comparison with Clang14.
https://godbolt.org/z/ec1TbovGb

[ParkHanbum]

maybe this patch causes this.

f0071d4

But to me, the patch seems to have a clear purpose. There is a possibility that it can be solved by removing it, but it is questionable whether this is a desirable direction. if this issue not need to fix fast then I want to try.

[ParkHanbum]

Unlike the cause of the issue, the way to solve it has become more complicated. reason is that LLVM has decided to no longer allow casting after operation if operator has two casted operand.

maybe this can solved one of followed way.

1. Add WellKnown bit operation in gvn pass
2. Add to select (icmp eq), true, false in InstCombine pass

first, assume that we got the ir like below

  %a = zext i1 %x to i32
  %b = zext i1 %y to i32
  %or = or i32 %a, %b
  %or0 = icmp eq i32 %or, 0
  br i1 %or0, label %cond.true, label %cond.false

cond.true:                                        ; preds = %entry
  %xor7 = xor i32 %a, %b
  br label %cond.end

cond.false:                                       ; preds = %entry
  %and6 = and i32 %a, %b
  br label %cond.end

cond.end:                                         ; preds = %cond.false, %cond.true
  %cond.in = phi i32 [ %xor7, %cond.true ], [ %and6, %cond.false ]
  %r = icmp eq i32 %cond.in, 0
  ret i1 %r
}

It will change as follows

  %a = zext i1 %x to i32
  %b = zext i1 %y to i32
  %or = or i32 %b, %a
  %or0 = icmp eq i32 %or, 0
  %xor7 = xor i32 %b, %a
  %and6 = and i32 %b, %a
  %cond.in = select i1 %or0, i32 %xor7, i32 %and6
  %r = icmp eq i32 %cond.in, 0
  ret i1 %r

we already know (X|B=0) when (X^Y)=0
so, we can transform first ir like this:

  %cond.in = phi i32 [ 0, %cond.true ], [ %and6, %cond.false ]

even if you change just this, the results will change.

  %and61 = and i1 %x, %y
  %0 = xor i1 %and61, true
  ret i1 %0

and apply same logic to select also has same result

  %cond = select i1 %tobool4.not, i32 0, i32 %xor

  %tobool131 = xor i1 %a, %b
  ret i1 %tobool131

as this way. this issue could be solved. IMO