[PATCH] D23937: [MemCpyOpt] Return value `memcpy` elision.

[bryant via llvm-commits]

bryant added a comment.

In https://reviews.llvm.org/D23937#526679, @efriedma wrote:

> This isn't safe.  Consider:
>
>   void f(bool b) {
>     S x, y;
>     g(&x, &y);
>     if (b) {
>       return x; // memcpy to sret
>     } else {
>       return y; // memcpy to sret
>     }
>   }
>
>
> With your transform, the two arguments to g become the same pointer.


Is this the only case it would be unsafe, i.e., in the general case, it's always safe as long as sret depends on a single alloca. Would that be the correct rule, or perhaps too restrictive?


Repository:
  rL LLVM

https://reviews.llvm.org/D23937