or and multiple whens are not equivalent, especially when there are exceptions in the game:

iex(1)> defmodule M do                                                          
...(1)> def guard_or(x) when x <> "foo" == "xfoo" or is_integer(x), do: true    
...(1)> def guard_or(_), do: false                                              
...(1)> 
...(1)> def guard_when(x) when x <> "foo" == "xfoo" when is_integer(x), do: true
...(1)> def guard_when(_), do: false                                            
...(1)> end
{:module, M,
 <<70, 79, 82, 49, 0, 0, 4, 172, 66, 69, 65, 77, 65, 116, 85, 56, 0, 0, 0, 105,
   0, 0, 0, 11, 8, 69, 108, 105, 120, 105, 114, 46, 77, 8, 95, 95, 105, 110,
   102, 111, 95, 95, 9, 102, 117, 110, 99, ...>>, {:guard_when, 1}}
iex(2)> M.guard_or(1)                                                           
false
iex(3)> M.guard_or("x")
true
iex(4)> M.guard_when(1)             
true
iex(5)> M.guard_when("x")
true

When using or an exception makes the whole guard false, while with multiple whens only the sub-expression is considered false.

I think you may be fixating on guards too much. It’s important to keep in mind that guards only allow a restricted set of expressions,

The reason for restricting the set of valid expressions is that evaluation of a guard expression must be guaranteed to be free of side effects.

Expressions — Erlang System Documentation v29.0.2

and that guards are a conditional extension to pattern matching - i.e. they can appear whenever a pattern match is possible. There are other alternatives for conditionals such as case/2 (which is based on pattern matching and therefore also supports guards) and cond/1.

To me cond seems most appropriate for your bmi_tell/1 example:

defmodule Demo do

  def bmi_tell_cond(bmi),
    do: (
      cond do
        bmi <= 18.5 -> "You're underweight!"
        bmi <= 25.0 -> "You're supposedly normal."
        bmi <= 30.0 -> "You're little too much!"
        true        -> "No way!"
      end)

  def bmi_tell_case(value),
    do: (
      case value do
        bmi when bmi <= 18.5 -> "You're underweight!"
        bmi when bmi <= 25.0 -> "You're supposedly normal."
        bmi when bmi <= 30.0 -> "You're little too much!"
        _                    -> "No way!"
      end)

end

IO.inspect(Demo.bmi_tell_cond(18.0))
IO.inspect(Demo.bmi_tell_cond(24.0))
IO.inspect(Demo.bmi_tell_cond(29.0))
IO.inspect(Demo.bmi_tell_cond(31.0))

IO.inspect(Demo.bmi_tell_case(18.0))
IO.inspect(Demo.bmi_tell_case(24.0))
IO.inspect(Demo.bmi_tell_case(29.0))
IO.inspect(Demo.bmi_tell_case(31.0))

Literally even!

Something like this:

def blah(a) when a<0, do: 0
def blah(1), do: 1
def blah(a) when a>0, do: 2

Quite literally compiles ‘almost’ identically to (‘almost’ meaning it is more lambda’ish internally):

def blah(a) do
  case a do
    a when a<0 -> 0
    1 -> 1
    a when a>0 -> 2
    _ -> raise %MatchError{...}
  end
end

Actually it compiles into a simple dispatch tree. It’s not like ‘way’ smart in that it can combine all parts possible, but it is pretty decent. In general you can think of it as ‘testing’ the first branch, if that fails it tests the second, if that fails it tests the third, and so on. It is a linear slowdown with the number of cases (though linear more in C terms so it is still blazing fast compared to anything you could write explicitly, this is also why guard calls are restricted).

Precisely. Multiple when’s compiles to different branches in the dispatch tree, where or’s get compiled into the same dispatch branch.