what's up with return *splat ?

[Phlip]

Rubies:

Back in the Halcyon days (whatever that means) of Ruby 1.8, a function
could obey the contract "return nil, a scalar, or an array" with a
mere splat:

return * splat

The interpretation there is (roughly!) "splat behaves as if you had
written each argument in an array as scalars separated by commas."

So (roughly!), you would get one of these three results:

return nil
return scalar
return [ scalar1, scalar2, scalar3 ]

Now that I work in Ruby 1.9, the splat don't work like that. It just
passes through an array.

Am I using it wrong? Did splat change? If so, why? And can I use some
other 1.9-compliant trick?

 

[Robert Klemme]

Rubies:

Back in the Halcyon days (whatever that means) of Ruby 1.8, a function
could obey the contract "return nil, a scalar, or an array" with a
mere splat:

return * splat

The interpretation there is (roughly!) "splat behaves as if you had
written each argument in an array as scalars separated by commas."

So (roughly!), you would get one of these three results:

return nil
return scalar
return [ scalar1, scalar2, scalar3 ]

Now that I work in Ruby 1.9, the splat don't work like that. It just
passes through an array.

Am I using it wrong? Did splat change? If so, why? And can I use some
other 1.9-compliant trick?

Can you please show the code you used for testing? Thank you.

Kind regards

robert

 

[Phlip]

Can you please show the code you used for testing?  Thank you.

My assert_latest() stopped working. Here's an independent test case:

p RUBY_VERSION

def assert(x); x or raise 'broke!'; end

def splat(*wat); return *wat; end

assert nil == splat()
assert 'yo' == splat('yo')
assert ['yo', 'dude'] == splat('yo', 'dude')

assert [] == splat()
assert ['yo'] == splat('yo')
assert ['yo', 'dude'] == splat('yo', 'dude')

The top three assertions pass in 1.8.7, and the bottom three pass in
1.9.2.

 

[Robert Klemme]

Can you please show the code you used for testing? Thank you.

My assert_latest() stopped working. Here's an independent test case:

p RUBY_VERSION

def assert(x); x or raise 'broke!'; end

def splat(*wat); return *wat; end

assert nil == splat()
assert 'yo' == splat('yo')
assert ['yo', 'dude'] == splat('yo', 'dude')

assert [] == splat()
assert ['yo'] == splat('yo')
assert ['yo', 'dude'] == splat('yo', 'dude')

The top three assertions pass in 1.8.7, and the bottom three pass in
1.9.2.

I get different results:

$ ruby x.rb
"1.8.7"
broke! x.rb:12
broke! x.rb:13
$ ruby19 x.rb
"1.9.3"
broke! x.rb:8:in `<main>'
broke! x.rb:9:in `<main>'
$ cat -n x.rb
1
2 p RUBY_VERSION
3
4 def assert(x); x or warn "broke! #{caller[0]}"; end
5
6 def splat(*wat); return *wat; end
7
8 assert nil == splat()
9 assert 'yo' == splat('yo')
10 assert ['yo', 'dude'] == splat('yo', 'dude')
11
12 assert [] == splat()
13 assert ['yo'] == splat('yo')
14 assert ['yo', 'dude'] == splat('yo', 'dude')
15
16 a,b=splat(1,2)
17 assert a == 1
18 assert b == 2
19
20 a,b=splat(1)
21 assert a == 1
22 assert b == nil
23
24 a,b=splat
25 assert a == nil
26 assert b == nil

And the interesting bits (lines 16ff) are treated identical.

I think you use a function in one of two ways: either you expect one
result and that can be nil or not, or you expect multiple replies and
assign them to different variables which can either be nil or not.

What practical use case is impaired by the difference?

Kind regards

robert

 

[Phlip]

Tx for the experiment; it confirmed mine.

I think you use a function in one of two ways: either you expect one
result and that can be nil or not, or you expect multiple replies and
assign them to different variables which can either be nil or not.

What practical use case is impaired by the difference?

When I write an assertion, I know the ordinality of the expected
result.

note1, note2 = assert_latest User.notes do
User.create_notes()
end

assert{ user.notes == [note1, note2] }

frob = assert_latest Frob do
production_code_creating_frobs()
end

assert{ frob.member == 42 }

I want the assertions to break, with simple syntax errors, if the
ordinality is wrong. If assert_latest only returned arrays, that would
cause clutter like this:

frobs = assert_latest Frob do
production_code_creating_frobs()
end

assert{ frobs.count == 1 and frobs[0].member == 42 }

Instead of testing the count, I want to simply use the result as if
it's what I expect, and then fail as early as possible if it isn't.

When I first asked this question (on these very newsgroups IIRC), I
was directed to return *records. That solved the ton of crud required
to put the ordinality check _inside_ assert_latest().

Then return *splat's behavior changed in 1.9.2.

 

[Robert Klemme]

Tx for the experiment; it confirmed mine.

Strange. I said

I get different results:
...

I think you use a function in one of two ways: either you expect one
result and that can be nil or not, or you expect multiple replies and
assign them to different variables which can either be nil or not.

What practical use case is impaired by the difference?

When I write an assertion, I know the ordinality of the expected
result.

note1, note2 = assert_latest User.notes do
User.create_notes()
end

assert{ user.notes == [note1, note2] }

frob = assert_latest Frob do
production_code_creating_frobs()
end

assert{ frob.member == 42 }

I want the assertions to break, with simple syntax errors, if the
ordinality is wrong.

You cannot get syntax errors for this. This is checked at runtime and
not at parse time.

If assert_latest only returned arrays, that would
cause clutter like this:

frobs = assert_latest Frob do
production_code_creating_frobs()
end

assert{ frobs.count == 1 and frobs[0].member == 42 }

Instead of testing the count, I want to simply use the result as if
it's what I expect, and then fail as early as possible if it isn't.

Note that you can do this in 1.9.* and 1.8.7 (the comma):

irb(main):002:0> def f(*a) a end
=> nil
irb(main):003:0> f 1,2
=> [1, 2]
irb(main):004:0> x=f 1,2
=> [1, 2]
irb(main):005:0> x
=> [1, 2]
irb(main):006:0> x,=f 1,2
=> [1, 2]
irb(main):007:0> x
=> 1

Kind regards

robert

 

[Phlip]

Strange.  I said

Your experiment confirmed 1.9.2 changed the behavior, like mine did.

You cannot get syntax errors for this.  This is checked at runtime and
not at parse time.

I'm aware of the definition of syntax error.

irb(main):006:0> x,=f 1,2

That line does not die, with an-error-checked-at-runtime, nor would
x.member (given [] does not have .member) die with an-error-checked-at-
runtime.

When the author of a test writes assert_latest() they know whether
they expect a scalar or a list to return, and they should not waste
their time checking for a single-element list before using it in the
assertion that actually checks for the important stuff.

If I use record, = assert_latest(), and if a future bug added a second
record, the assertion would not break, that line would not break (with
a comma), and lines using record would not break.

Also I'm not sure why I need to justify the use case, if such a low-
level syntactical thing should not change in a language revision. It's
a bug in return *splat; it no longer behaves the same as = *splat.

 

[Robert Klemme]

Your experiment confirmed 1.9.2 changed the behavior, like mine did.

Yes, but I got different results than those you claimed. Which made me
wonder.

I'm aware of the definition of syntax error.

Why then do you use it in the wrong way?

irb(main):006:0> x,=f 1,2

That line does not die, with an-error-checked-at-runtime, nor would
x.member (given [] does not have .member) die with an-error-checked-at-
runtime.

When the author of a test writes assert_latest() they know whether
they expect a scalar or a list to return, and they should not waste
their time checking for a single-element list before using it in the
assertion that actually checks for the important stuff.

Well, if you use "x,=..." then you have the first element and no
additional checking is needed. But if you really want to ensure the
proper number of values is returned you need to test for the array
length anyway. At least you would need to to something like

irb(main):008:0> x,y,*remainder = f 1,2
=> [1, 2]
irb(main):009:0> remainder
=> []
irb(main):010:0> remainder.empty?
=> true
irb(main):011:0> x,y,*remainder = f 1,2,3
=> [1, 2, 3]
irb(main):012:0> remainder.empty?
=> false

For consistent checking of the returned data you can do

*x = f(...)

x will be an Array even if f returns a single value only.

If I use record, = assert_latest(), and if a future bug added a second
record, the assertion would not break, that line would not break (with
a comma), and lines using record would not break.

Also I'm not sure why I need to justify the use case, if such a low-
level syntactical thing should not change in a language revision.

Actually most people seem to cope pretty well with this as only corner
cases are affected of the change. If you want to ensure no additional
values are returned you need to splat assign any way to count values, i.e.

*x = f()

It's
a bug in return *splat; it no longer behaves the same as = *splat.

It's not syntax but runtime behavior. You can and will never get a
syntax error for this.

AFAIK Matz changed it deliberately - although I have to confess I don't
remember the reasoning. You'll probably find it in the archives.

Note though that also argument assignment changed quite a bit in 1.9
which has much more sophisticated options where 1.8 only allowed for the
last parameter to collect additional values. The 1.9 model is superior
to that:

irb(main):001:0> def f(a, *b, c) p a,b,c end
=> nil
irb(main):002:0> f 1
ArgumentError: wrong number of arguments (1 for 2)
from (irb):1:in `f'
from (irb):2
from /usr/local/bin/irb19:12:in `<main>'
irb(main):003:0> f 1,2
1
[]
2
=> [1, [], 2]
irb(main):004:0> f 1,2,3
1
[2]
3
=> [1, [2], 3]
irb(main):005:0> f 1,2,3,4,5
1
[2, 3, 4]
5
=> [1, [2, 3, 4], 5]

And you have pattern matching with block arguments

irb(main):007:0> f = lambda {|a, (b,c), d| p a,b,c,d}
=> #<Proc:0x202a96c4@(irb):7 (lambda)>
irb(main):008:0> f[1]
ArgumentError: wrong number of arguments (1 for 3)
from (irb):7:in `block in irb_binding'
from (irb):8:in `[]'
from (irb):8
from /usr/local/bin/irb19:12:in `<main>'
irb(main):009:0> f[1,2]
ArgumentError: wrong number of arguments (2 for 3)
from (irb):7:in `block in irb_binding'
from (irb):9:in `[]'
from (irb):9
from /usr/local/bin/irb19:12:in `<main>'
irb(main):010:0> f[1,2,3]
1
2
nil
3
=> [1, 2, nil, 3]
irb(main):011:0> f[1,[2],3]
1
2
nil
3
=> [1, 2, nil, 3]
irb(main):012:0> f[1,[2,3],4]
1
2
3
4
=> [1, 2, 3, 4]
irb(main):013:0> f[1,[2,3,4],5]
1
2
3
5
=> [1, 2, 3, 5]


Cheers

robert

 

[Phlip]

Actually most people seem to cope pretty well with this as only corner

cases are affected of the change.  If you want to ensure no additional
values are returned you need to splat assign any way to count values, i.e..

*x = f()

Sorry, I can't handle this level of talking-past each other.