[llvm] [APInt] Use a std::move() to avoid a copy in the loop in multiplicativeInverse. (PR #87655)

[Craig Topper via llvm-commits]

topperc wrote:

> This is still allocating every iteration. If you want to completely avoid that, you'd have to avoid freeing the allocation in T. For example:
> 
> ```
> while (1) {
>   T = *this;  // APInt has a special-case to avoid reallocating here.
>   T *= Factor;
>   if (T.isOne())
>     break;
>   T.negate();
>   T += 2;
>   Factor *= T;
> }
> ```
> 
> Not sure how much we want to trade off readability here, though. The current version might be good enough.
> 
> > I always wonder, why can't the compiler work this out for itself?
> 
> In C++, the compiler is required to emit the operations as written, barring a few special cases where copy-elision is specifically allowed. There are different tradeoffs in other rules (e.g. this example just works in Rust, but Rust doesn't allow overriding the move construction/assignment operators).

operator*= for APInt still does a new allocation doesn't it?

```
APInt &APInt::operator*=(const APInt &RHS) {
  *this = *this * RHS;
  return *this;
}
```

https://github.com/llvm/llvm-project/pull/87655