https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112884
--- Comment #3 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
(In reply to Richard Biener from comment #2)
> I'd rather not do it in match.pd - well, maybe improve A / CST * CST
> folding to consider nonzero bits, that is.
I don't think there is anything to do in match.pd really.
Since `(A / CST) * CST` already simplifies to `A & ~(CST1)` for power of 2.
Rather VRP just needs to simplify `a_3(D) & 4294967294` to `a_3(D)` when it
knows that `a_3(D) & 1` is 0 already (which it knows at that point:
```
a_3(D) [irange] unsigned int [0, 0][2, 4294967294] MASK 0xfffffffe VALUE 0x0
```
).
Note we already handle:
```
unsigned src(unsigned a) {
if (a % 2 != 0) __builtin_unreachable();
return (a / 2 * 2);
}
```
Via match:
```
Folding statement: _3 = a_2(D) & 4294967294;
Applying pattern match.pd:1572, gimple-match-6.cc:16797
gimple_simplified to _3 = a_2(D);
Folded into: _3 = a_2(D);
```
This pattern uses get_nonzero_bits. get_nonzero_bits uses the global ranger to
find out the nonzero bits. We could add a similar optimization directly to the
folding part of VRP that uses the local ranger to find the nonzero bits.
