Simplify `sadd_overflow`+`assume` to `add nsw` [InstCombine]

[scottmcm]

The probably never comes up in C++, but in Rust people can easily end up there from things that boil down to

x.checked_add(y).unwrap_unchecked()

Proof: https://alive2.llvm.org/ce/z/m-n_PG

define i32 @src(i32 noundef %x, i32 noundef %y) noundef {
start:
  %#0 = sadd_overflow i32 noundef %x, noundef %y
  %_7.1 = extractvalue {i32, i1, i24} %#0, 1
  %_10 = xor i1 %_7.1, 1
  %_7.0 = extractvalue {i32, i1, i24} %#0, 0
  assume i1 %_10
  ret i32 %_7.0
}
=>
define i32 @tgt(i32 noundef %x, i32 noundef %y) noundef {
start:
  %r = add nsw i32 noundef %x, noundef %y
  ret i32 %r
}
Transformation seems to be correct!

And of course the uadd_overflow+assume → add nuw too.

[nikic]

Do you have an example with larger context here?

As written, this is going to lose information, so I'm not sure whether this makes sense. I assume that there is some follow-on optimization failure that you want to avoid?

[scottmcm]

I don't have a specific case, since I've always been able to use .unchecked_add instead of the checked+unwrap_unchecked dance.

That said, I don't think it loses information? Alive says the opposite transform would also be legal: https://alive2.llvm.org/ce/z/cLbTUj

So AFAICT the nsw/nuw preserves all the necessary information.

[nikic]

The opposite transform is only legal because you are passing the value to a noundef return value. Otherwise this downgrades from UB to poison.

[scottmcm]

Ah, makes sense! Thanks. I guess I'll close this, then, since I didn't feel particularly strongly about it and you found a reason not to do it.