[SR-7497] [stdlib] Fix KeyPaths hashing

[dmcyk]

• The cause of hashing regression was a shadowed hasher argument in component hashing, so the component type/offset wasn't taken into account
• Currently only path components are hashed, so paths with distinct types but properties of equal kind and offset result in same hash values. That differs from equatable implementation as type of two key paths must be the same to consider them equal.
  So I thought it'd make sense to include the type metadata pointer while hashing. As far as I understand it, the metadata pointer doesn't change during program execution.

Regarding code:
I'm not sure if type(of: self) as Any.Type cast is necessary.
Quickly running REPL trying to do unsafeBitCast(type(of: Int(0)), to: Int.self)crashes, casting the type(of:) result to Any.Type helps though.
Is this caused by value vs ref type difference? Doing so with a class object works.
I was looking at _getTypeName and other core functions, but I couldn't really find a place where Any.Type is guaranteed to be const Metadata* and how Any.Type differs from e.g. Int.Type.
Is it somehow implemented at SIL level? Would be great if someone could maybe give me a tip where to look for information/code regarding this.

Resolves SR-7497.

P.S. That's my first attempt at contribution to Swift. Hello everyone 👋

[huonw]

Quickly running REPL trying to do unsafeBitCast(type(of: Int(0)), to: Int.self)crashes, casting the type(of:) result to Any.Type helps though.
Is this caused by value vs ref type difference? Doing so with a class object works.

The metatype object for something that can't be subclassed (or otherwise store interesting information about the type) is zero-sized. The (meta)type Int.Type has just one value Int.self, meaning it is like ()/Void and has the same size. In contrast, given

class Super {}
class Sub1: Super {}
class Sub2: Super {}

The type Super.Type has (at least) three possible values Super.self, Sub1.self and Sub2.self. This is, I believe, represented by the values of Super.Type being a pointer to the actual metadata. I believe metatypes of protocol existentials and Any.Type are similar, and they are a pointer to the actual information about the type in the existential as you suspect. Someone like @jckarter will have to confirm this, though.

As I guess you might already know, unsafeBitCast only works between things of the same size, which is why it fails for Int.Type (size 0) -> Int (size 8).

(Hello and welcome!)

[dmcyk]

ah, I see. I misunderstood it. I thought Int.Type was representing the type metadata itself, rather than a meta type.
I found init_existential_metatype instruction in SIL reference, that says that existential meta types store the type metadata. There's also a note on Metatype types, saying indeed that concrete meta types require no storage. Thanks!

[jckarter]

This could be written ObjectIdentifier(type(of: self)).hash(into: &hasher)

[dmcyk]

Good to know, thanks 👌

[jckarter]

Nice catch!

[lorentey]

Thanks for spotting this!

[lorentey]

@swift-ci please smoke test and merge

[lorentey]

@swift-ci please smoke test and merge

[dmcyk]

@lorentey my pleasure 😁 I guess this could be merged into 4.2 too, right?

[lorentey]

@dmcyk Yes, we should get it into 4.2! I'll cherry pick it after it lands on master.