[llvm] [ThinLTO] optimize propagateAttributes performance (PR #132917)

[Nikita Popov via llvm-commits]

nikic wrote:

Inserting 50M values into a hash table shouldn't really take an appreciable amount of time. I wonder whether you're hitting some kind of catastrophic hash collision issue and complexity degenerates quadratically?

The implementation is
```
static unsigned getHashValue(ValueInfo I) { return (uintptr_t)I.getRef(); }
```
Which is already somewhat bad because the low three bits of I.getRef() are guaranteed to be zero, so you could try whether adding a `>> 3` improves things a bit. And then maybe try whether wrapping it in a hash_value() helps even more.

https://github.com/llvm/llvm-project/pull/132917