let a = someConst.unsafeAddr is very buggy and should be replaced

[timotheecour]

this recent PR #15466 enabled one to write let a = someConst.unsafeAddr; however this feature is very buggy, not documented, had 0 tests in the PR that introduced it and should be reverted; see upcoming proposal for a better and simpler way to solve nim-lang/RFCs#257 that has none of its drawbacks

Example 1

unsafeAddr doesn't work with most types

when true:
  template fn(a) =
    const b = a
    let c = b.unsafeAddr
  type A = object
    a0: int
  type B = ref A
  fn [1,2]
  fn @[1,2]
  fn A(a0:1)
  # these don't work: Error: expression has no address
  # fn "ab"
  # fn (1, 2)
  # fn 1
  # fn B(a0:1) # this one would work with const ref PR https://github.com/nim-lang/Nim/pull/15528

note that these would work with let c = b instead of let c = b.unsafeAddr

Example 2

when true:
  const a = [1,2]
  proc fn =
    let a1 = a.unsafeAddr
    a1[][0] = 7
    echo a1[]
    echo a
  static: fn()
  fn()

vm:
[7, 2] # modified from [1,2]
[1, 2] # unmodified from [1,2]: inconsistent
rt:
SIGBUS

Example 3

# sub.nim
when true:
  const a* = [1,2]
  proc fn1* =
    let a1 = a.unsafeAddr
    let a2 = a.unsafeAddr
    echo cast[int](a1)
    echo cast[int](a2)

# main.nim
when true:
  import sub
  proc fn2* =
    let a1 = a.unsafeAddr
    let a2 = a.unsafeAddr
    echo cast[int](a1)
    echo cast[int](a2)
  fn1()
  fn2()

prints:
4416151424
4416151424
4416151440 # address of a returned is different depending on which module it was called from
4416151440

and this is generated in each module:

static NIM_CONST tyArray__HU7qaqKu9czJLT84iCBJnsA TM__CLoBuXEHTlg1in6VU6Aqdg_2 = {((NI) 1),
((NI) 2)}

unlike what was required in nim-lang/RFCs#257

Example 4

unsafeAddr still doesn't help with types (including arrays) that (transitively) contain ref/ptr; it still needs #15528

Example 5

when defined case10:
  const a = @[1,2]
  proc fn* =
    let a1 = a.unsafeAddr
    let a2 = a.unsafeAddr
    echo cast[int](a1)
    echo cast[int](a2)
  static: fn()
  fn()

vm:
4378615304
4378615368 # different from previous line (inconsistent)
rt:
4382200032
4382200032 # same from previous line: inconsistent

Example 6

when true:
  const a1 = @[1,2]
  const a2 = [1,2]
  proc fn* =
    let b1 = a1.unsafeAddr
    echo cast[int](b1) # ok
    let b2 = a2.unsafeAddr
    echo cast[int](b2) # Error: unhandled exception: 'intVal' is not accessible using discriminant 
  static: fn()

Example 7

when true:
  const a = @[1,2]
  let b = a.unsafeAddr
  proc fn1 =
    b[][0] = 7
    echo b[]
  fn1()

nim r main: SIGBUS
nim r -b:cpp main: prints @[7, 2]

Example 8

when true:
  const a = [1,2]
  let b = a.unsafeAddr
  proc fn1 =
    b[][0] = 7
    echo b[]
  fn1()

nim r main: prints [1, 2] with -d:danger, or SIGBUG without
nim r -b:cpp main: prints [7, 2] with or without -d:danger

Additional Information

1.5.1 2b5841c

links

• const now works with ref types #15528
• implements https://github.com/nim-lang/RFCs/issues/257 #15466
• const should work with ref types RFCs#271
• for ai in a has quadratic complexity in VM when a is const #16790
• static vs N_LIB_PRIVATE #16738
• Noinline and/or Addressable consts RFCs#257

[Clyybber]

Example 2, 7 and 8 are invalid because you are using unsafeAddr to mutate an immutable location. (which is undefined behaviour, but you know that..)
Example 6 is not related and happens with let too, because it's a VM issue:

static:
  let a1 = @[1,2]
  let a2 = [1,2]
  proc fn =
    let b1 = a1.unsafeAddr
    echo cast[int](b1) # ok
    let b2 = a2.unsafeAddr
    echo cast[int](b2) # Error: unhandled exception: 'intVal' is not accessible using discriminant 
  fn()

Please remove them from the issue so that it doesn't distract from actual issues like as the one demonstrated in Example 3.

[timotheecour]

Example 2, 7 and 8 are invalid because you are using unsafeAddr to mutate an immutable location. (which is undefined behaviour, but you know that..)

that's the point, #15466 introduced unsafeAddr as way to implement nim-lang/RFCs#257 but as I'm showing with those examples, it creates an un-necessary unsafe loophole in the type system. My point is that nim-lang/RFCs#257 can be fixed in a simpler way without resorting to unsafeAddr and its associated issues; I haven't written the RFC yet but see the draft in timotheecour#553

[Clyybber]

There is no loophole. The "unsafe" is in the name. You must ensure not to modify the location, just as when you take the unsafeAddr of a let symbol.