Reinstate printing of :line Exprs

[toivoh]

Line numbers in ASTs are nowadays often represented with LineNumberNodes, but this at least doesn't hold for macro arguments, which caused trouble in #15299, and is demonstrated by the example

macro q(ex)
    QuoteNode([ex])
end
@q begin x end

which prints

1-element Array{Expr,1}:
 quote 
    $(Expr(:line, 1, :none))
    x
end

This patch reinstates printing of :line expressions as it existed in Julia 0.3.
An alternative (and maybe preferable?) fix would be to consistently present LineNumberNodes to the user instead.

[toivoh]

@tkelman: Sure, I can add a few tests. Just wanted to know if this was at all desired.
But I guess that even if we would get rid of user-visible LineNumberNodes completely in 0.5, this could still bear back porting to 0.4.

[tkelman]

Not sure. Was it bisected to see if the removal was intentional?

[toivoh]

Seems to have been changed in 7bf3449, by @ihnorton. But the commit message doesn't say anything about :line Exprs.

[toivoh]

Added some tests.

[toivoh]

The Travis OSX build passed, while all the others failed.
They all say that there was an error while loading /tmp/julia/share/julia/test/show.jl, in expression starting on line 369.
Weird, as far as I can tell, the expression starting on line 369 is

@test replace(str_ex2, " ", "") == replace("""
begin  # none, line 1:
    x
end""", " ", "")

Anybody have any idea why that might trigger a loading error? I'd been slightly less surprised if the few lines just above it did, since they involve a macro definition.

[toivoh]

Sorry, silly mistake in the tests.
I was confused by the error message in the AppVeyor/Travis logs though:

LoadError: There was an error during testing

To me this would indicate that the test code caused an unexpected exception, not that a test failed. The actual failure message was out of sight one screen higher up, with output from a lot of other threads in between.

The test should hopefully pass this time. It seems that the reason that they passed on OSX was simply that the show.jl tests weren't run there.

[toivoh]

AppVeyor 64 bit failed in float16:

Expression: round(Int,Float16(x)) == round(Int,x)
    From worker 10:    Evaluated: 0 == 1

Seems totally unrelated?

[tkelman]

That is odd, and certainly unrelated.

[tkelman]

I'm a little amazed x = Float32(rand()); round(Int,Float16(x)) == round(Int,x) hasn't failed sooner. Maybe random seed issue, as more tests gradually get added? For a bunch of Float32 values just above 0.5, Float16(x) returns Float16(0.5) which rounds to 0, but round(Int, x) rounds to 1.

[rfourquet]

This rand test fails in approximately 0.025% of the runs.

[toivoh]

Thanks for the info, @rfourquet!
If there's no objections, I'll merge this in a couple of days.

[toivoh]

Btw am I right that this would be a candidate for back porting to Julia 0.4? The problem exists on 0.4 as well, and I'd consider this a straight bug fix.