zygoloid wrote:
The semantics of `ArrayInitLoopExpr` are to first evaluate (once, up-front) the
common expression, and then evaluate the subexpression once for each array
element, where the subexpression can make repeated reference to the value of
the common expression. With this change, we will instead evaluate the common
expression once for each iteration of the loop. I would expect this change will
cause us to miscompile [https://godbolt.org/z/d4so4rfjh](this example):
```c++
struct X {
int arr[3];
};
constexpr X f(int &r) {
return {++r, ++r, ++r};
}
constexpr int g() {
int n = 0;
auto [a, b, c] = f(n).arr;
return a + b + c;
}
static_assert(g() == 6)
```
... because we'll interpret the initializer of `[a, b, c]` as `{f(n).arr[0],
f(n).arr[1], f(n).arr[2]}` instead of as `<common> = f(n).arr; {<common>[0],
<common>[1], <common>[2]}`.
https://github.com/llvm/llvm-project/pull/67716
_______________________________________________
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
