Getting a UTF8Span directly from a small String doesn't work

[glessard]

Description

UTF8Span instances obtained from a String in the small representation are not valid. They point to the wrong memory. In contrast, gettin a UTF8Span from a Span<UInt8> given out by UTF8View is fine. This is very surprising given that String.utf8Span is a thin wrapper over UTF8View.span.

Reproduction

let master = String(200) //"This string is not a small string."

func smallStringTest1() {
  let s = master
  let utf8 = s.utf8Span
  let span = utf8.span
  for i in span.indices {
      print(span[i], terminator: " ")
  }
  print()
}

smallStringTest1()

func smallStringTest2() {
  let s = master.utf8
  let view = s.span
  let utf8 = try! UTF8Span(validating: view)
  let span = utf8.span
  for i in span.indices {
      print(span[i], terminator: " ")
  }
  print()
}

smallStringTest2()

func smallStringTest3() {
  let s = master
  let utf8 = s.utf8Span
  var it = utf8.makeUnicodeScalarIterator()
  while let c = it.next() {
    print(c, terminator: " ")
  }
  print()
}

smallStringTest3()

func smallStringTest4() {
  let s = master.utf8
  let view = s.span
  let utf8 = try! UTF8Span(validating: view)
  var it = utf8.makeUnicodeScalarIterator()
  while let c = it.next() {
    print(c, terminator: " ")
  }
  print()
}

smallStringTest4()

(also here: https://swift.godbolt.org/z/zv15z94MG)

The 1st and 3rd functions print surprising output. The 3rd sometimes crashes. Sample output:

Program returned: 0
88 189 14 
50 48 48 
2 ½ � 
2 0 0

Expected behavior

Given this code, we'd expect the output of the 1st and 3rd to be identical to the 2nd and 4th, respectively:

Program returned: 0
50 48 48 
50 48 48 
2 0  0 
2 0  0 

Note that when master's value is changed to a "large" String, all the output is as expected.

Environment

swift-DEVELOPMENT-SNAPSHOT-2025-06-03-a
Observed on macOS and Linux.

Code example: https://swift.godbolt.org/z/zv15z94MG

Additional information

Related to: #81931

Also tracked as rdar://152615664

[jckarter]

From a quick glance, it looks like String isn't @_addressableForDependencies, which it would need to be in order for a Span to be able to depend on its internal representation.

[glessard]

The original implementation also had a sneaky dependence on a temporary, which is fixed in #82077.
The addressable attribute is fixed in #82013.