Lancern wrote:
> @Lancern Have you seen any significant benefits from the invariant handling?
> In general, the LLVM optimizer is very good at detecting when alloca loads
> are redundant, and it seems to me that this might be introducing more
> fragility than benefit.
The initial motivation is to optimize for code like this:
```cpp
int g();
void use(int);
void h(const int *); // <-- The body of h is external.
void f() {
const int x = g();
h(&x);
use(x);
}
```
Since `h` is external, LLVM cannot take the risk to eliminate the reload from
alloca before calling `use`, despite that the reload is indeed unnecessary
according to the C++ standard. One hope of CIR was (and I believe it still is)
to enable language-specific optimizations like this.
To enable such optimizations we had two choices:
1. Implement some kind of const-propagation pass directly in CIR.
2. Teach LLVM to optimize this by emitting better LLVM IR or attaching more
metadata to the IR.
It was decided that 1 would probably be an overkill. Since LLVM already has
something like invariant group (although it's experimental and it has caveats
as @andykaylor pointed out), 2 would seem like a more reasonable approach back
then.
https://github.com/llvm/llvm-project/pull/175037
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
